?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
Unleashed v1.4
[unleashed.git] / kernel / vm / seg_vn.c
blob93629d234c14d409f556e73029219fba655a4c61
1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 1986, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright 2018 Joyent, Inc.
24 * Copyright 2015 Nexenta Systems, Inc. All rights reserved.
25 */
26
27 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
28 /* All Rights Reserved */
29
30 /*
31 * University Copyright- Copyright (c) 1982, 1986, 1988
32 * The Regents of the University of California
33 * All Rights Reserved
34 *
35 * University Acknowledgment- Portions of this document are derived from
36 * software developed by the University of California, Berkeley, and its
37 * contributors.
38 */
39
40 /*
41 * VM - shared or copy-on-write from a vnode/anonymous memory.
42 */
43
44 #include <sys/types.h>
45 #include <sys/param.h>
46 #include <sys/t_lock.h>
47 #include <sys/errno.h>
48 #include <sys/systm.h>
49 #include <sys/mman.h>
50 #include <sys/debug.h>
51 #include <sys/cred.h>
52 #include <sys/vmsystm.h>
53 #include <sys/tuneable.h>
54 #include <sys/bitmap.h>
55 #include <sys/swap.h>
56 #include <sys/kmem.h>
57 #include <sys/sysmacros.h>
58 #include <sys/vtrace.h>
59 #include <sys/cmn_err.h>
60 #include <sys/callb.h>
61 #include <sys/vm.h>
62 #include <sys/dumphdr.h>
63 #include <sys/lgrp.h>
64
65 #include <vm/hat.h>
66 #include <vm/as.h>
67 #include <vm/seg.h>
68 #include <vm/seg_vn.h>
69 #include <vm/pvn.h>
70 #include <vm/anon.h>
71 #include <vm/page.h>
72 #include <vm/vpage.h>
73 #include <sys/proc.h>
74 #include <sys/task.h>
75 #include <sys/project.h>
76 #include <sys/zone.h>
77 #include <sys/shm_impl.h>
78
79 /*
80 * segvn_fault needs a temporary page list array. To avoid calling kmem all
81 * the time, it creates a small (FAULT_TMP_PAGES_NUM entry) array and uses
82 * it if it can. In the rare case when this page list is not large enough,
83 * it goes and gets a large enough array from kmem.
84 */
85 #define FAULT_TMP_PAGES_NUM 0x8
86 #define FAULT_TMP_PAGES_SZ ptob(FAULT_TMP_PAGES_NUM)
87
88 /*
89 * Private seg op routines.
90 */
91 static int segvn_dup(struct seg *seg, struct seg *newseg);
92 static int segvn_unmap(struct seg *seg, caddr_t addr, size_t len);
93 static void segvn_free(struct seg *seg);
94 static faultcode_t segvn_fault(struct hat *hat, struct seg *seg,
95 caddr_t addr, size_t len, enum fault_type type,
96 enum seg_rw rw);
97 static faultcode_t segvn_faulta(struct seg *seg, caddr_t addr);
98 static int segvn_setprot(struct seg *seg, caddr_t addr,
99 size_t len, uint_t prot);
100 static int segvn_checkprot(struct seg *seg, caddr_t addr,
101 size_t len, uint_t prot);
102 static int segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta);
103 static int segvn_sync(struct seg *seg, caddr_t addr, size_t len,
104 int attr, uint_t flags);
105 static size_t segvn_incore(struct seg *seg, caddr_t addr, size_t len,
106 char *vec);
107 static int segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
108 int attr, int op, ulong_t *lockmap, size_t pos);
109 static int segvn_getprot(struct seg *seg, caddr_t addr, size_t len,
110 uint_t *protv);
111 static uoff_t segvn_getoffset(struct seg *seg, caddr_t addr);
112 static int segvn_gettype(struct seg *seg, caddr_t addr);
113 static int segvn_getvp(struct seg *seg, caddr_t addr,
114 struct vnode **vpp);
115 static int segvn_advise(struct seg *seg, caddr_t addr, size_t len,
116 uint_t behav);
117 static void segvn_dump(struct seg *seg);
118 static int segvn_pagelock(struct seg *seg, caddr_t addr, size_t len,
119 struct page ***ppp, enum lock_type type, enum seg_rw rw);
120 static int segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len,
121 uint_t szc);
122 static int segvn_getmemid(struct seg *seg, caddr_t addr,
123 memid_t *memidp);
124 static lgrp_mem_policy_info_t *segvn_getpolicy(struct seg *, caddr_t);
125 static int segvn_inherit(struct seg *, caddr_t, size_t, uint_t);
126
127 const struct seg_ops segvn_ops = {
128 .dup = segvn_dup,
129 .unmap = segvn_unmap,
130 .free = segvn_free,
131 .fault = segvn_fault,
132 .faulta = segvn_faulta,
133 .setprot = segvn_setprot,
134 .checkprot = segvn_checkprot,
135 .kluster = segvn_kluster,
136 .sync = segvn_sync,
137 .incore = segvn_incore,
138 .lockop = segvn_lockop,
139 .getprot = segvn_getprot,
140 .getoffset = segvn_getoffset,
141 .gettype = segvn_gettype,
142 .getvp = segvn_getvp,
143 .advise = segvn_advise,
144 .dump = segvn_dump,
145 .pagelock = segvn_pagelock,
146 .setpagesize = segvn_setpagesize,
147 .getmemid = segvn_getmemid,
148 .getpolicy = segvn_getpolicy,
149 .inherit = segvn_inherit,
150 };
151
152 /*
153 * Common zfod structures, provided as a shorthand for others to use.
154 */
155 static segvn_crargs_t zfod_segvn_crargs =
156 SEGVN_ZFOD_ARGS(PROT_ZFOD, PROT_ALL);
157 static segvn_crargs_t kzfod_segvn_crargs =
158 SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_USER,
159 PROT_ALL & ~PROT_USER);
160 static segvn_crargs_t stack_noexec_crargs =
161 SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_EXEC, PROT_ALL);
162
163 caddr_t zfod_argsp = (caddr_t)&zfod_segvn_crargs; /* user zfod argsp */
164 caddr_t kzfod_argsp = (caddr_t)&kzfod_segvn_crargs; /* kernel zfod argsp */
165 caddr_t stack_exec_argsp = (caddr_t)&zfod_segvn_crargs; /* executable stack */
166 caddr_t stack_noexec_argsp = (caddr_t)&stack_noexec_crargs; /* noexec stack */
167
168 #define vpgtob(n) ((n) * sizeof (struct vpage)) /* For brevity */
169
170 size_t segvn_comb_thrshld = UINT_MAX; /* patchable -- see 1196681 */
171
172 size_t segvn_pglock_comb_thrshld = (1UL << 16); /* 64K */
173 size_t segvn_pglock_comb_balign = (1UL << 16); /* 64K */
174 uint_t segvn_pglock_comb_bshift;
175 size_t segvn_pglock_comb_palign;
176
177 static int segvn_concat(struct seg *, struct seg *, int);
178 static int segvn_extend_prev(struct seg *, struct seg *,
179 struct segvn_crargs *, size_t);
180 static int segvn_extend_next(struct seg *, struct seg *,
181 struct segvn_crargs *, size_t);
182 static void segvn_softunlock(struct seg *, caddr_t, size_t, enum seg_rw);
183 static void segvn_pagelist_rele(page_t **);
184 static void segvn_setvnode_mpss(vnode_t *);
185 static void segvn_relocate_pages(page_t **, page_t *);
186 static int segvn_full_szcpages(page_t **, uint_t, int *, uint_t *);
187 static int segvn_fill_vp_pages(struct segvn_data *, vnode_t *, uoff_t,
188 uint_t, page_t **, page_t **, uint_t *, int *);
189 static faultcode_t segvn_fault_vnodepages(struct hat *, struct seg *, caddr_t,
190 caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
191 static faultcode_t segvn_fault_anonpages(struct hat *, struct seg *, caddr_t,
192 caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
193 static faultcode_t segvn_faultpage(struct hat *, struct seg *, caddr_t,
194 uoff_t, struct vpage *, page_t **, uint_t,
195 enum fault_type, enum seg_rw, int);
196 static void segvn_vpage(struct seg *);
197 static size_t segvn_count_swap_by_vpages(struct seg *);
198
199 static void segvn_purge(struct seg *seg);
200 static int segvn_reclaim(void *, caddr_t, size_t, struct page **,
201 enum seg_rw, int);
202 static int shamp_reclaim(void *, caddr_t, size_t, struct page **,
203 enum seg_rw, int);
204
205 static int sameprot(struct seg *, caddr_t, size_t);
206
207 static int segvn_demote_range(struct seg *, caddr_t, size_t, int, uint_t);
208 static int segvn_clrszc(struct seg *);
209 static struct seg *segvn_split_seg(struct seg *, caddr_t);
210 static int segvn_claim_pages(struct seg *, struct vpage *, uoff_t,
211 ulong_t, uint_t);
212
213 static void segvn_hat_rgn_unload_callback(caddr_t, caddr_t, caddr_t,
214 size_t, void *, uoff_t);
215
216 static struct kmem_cache *segvn_cache;
217 static struct kmem_cache **segvn_szc_cache;
218
219 #ifdef VM_STATS
220 static struct segvnvmstats_str {
221 ulong_t fill_vp_pages[31];
222 ulong_t fltvnpages[49];
223 ulong_t fullszcpages[10];
224 ulong_t relocatepages[3];
225 ulong_t fltanpages[17];
226 ulong_t pagelock[2];
227 ulong_t demoterange[3];
228 } segvnvmstats;
229 #endif /* VM_STATS */
230
231 #define SDR_RANGE 1 /* demote entire range */
232 #define SDR_END 2 /* demote non aligned ends only */
233
234 #define CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr) { \
235 if ((len) != 0) { \
236 lpgaddr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz); \
237 ASSERT(lpgaddr >= (seg)->s_base); \
238 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)((addr) + \
239 (len)), pgsz); \
240 ASSERT(lpgeaddr > lpgaddr); \
241 ASSERT(lpgeaddr <= (seg)->s_base + (seg)->s_size); \
242 } else { \
243 lpgeaddr = lpgaddr = (addr); \
244 } \
245 }
246
247 /*ARGSUSED*/
248 static int
249 segvn_cache_constructor(void *buf, void *cdrarg, int kmflags)
250 {
251 struct segvn_data *svd = buf;
252
253 rw_init(&svd->lock, NULL, RW_DEFAULT, NULL);
254 mutex_init(&svd->segfree_syncmtx, NULL, MUTEX_DEFAULT, NULL);
255 svd->svn_trnext = svd->svn_trprev = NULL;
256 return (0);
257 }
258
259 /*ARGSUSED1*/
260 static void
261 segvn_cache_destructor(void *buf, void *cdrarg)
262 {
263 struct segvn_data *svd = buf;
264
265 rw_destroy(&svd->lock);
266 mutex_destroy(&svd->segfree_syncmtx);
267 }
268
269 /*ARGSUSED*/
270 static int
271 svntr_cache_constructor(void *buf, void *cdrarg, int kmflags)
272 {
273 bzero(buf, sizeof (svntr_t));
274 return (0);
275 }
276
277 /*
278 * Patching this variable to non-zero allows the system to run with
279 * stacks marked as "not executable". It's a bit of a kludge, but is
280 * provided as a tweakable for platforms that export those ABIs
281 * (e.g. sparc V8) that have executable stacks enabled by default.
282 * There are also some restrictions for platforms that don't actually
283 * implement 'noexec' protections.
284 *
285 * Once enabled, the system is (therefore) unable to provide a fully
286 * ABI-compliant execution environment, though practically speaking,
287 * most everything works. The exceptions are generally some interpreters
288 * and debuggers that create executable code on the stack and jump
289 * into it (without explicitly mprotecting the address range to include
290 * PROT_EXEC).
291 *
292 * One important class of applications that are disabled are those
293 * that have been transformed into malicious agents using one of the
294 * numerous "buffer overflow" attacks. See 4007890.
295 */
296 int noexec_user_stack = 0;
297 int noexec_user_stack_log = 1;
298
299 int segvn_lpg_disable = 0;
300 uint_t segvn_maxpgszc = 0;
301
302 ulong_t segvn_vmpss_clrszc_cnt;
303 ulong_t segvn_vmpss_clrszc_err;
304 ulong_t segvn_fltvnpages_clrszc_cnt;
305 ulong_t segvn_fltvnpages_clrszc_err;
306 ulong_t segvn_setpgsz_align_err;
307 ulong_t segvn_setpgsz_anon_align_err;
308 ulong_t segvn_setpgsz_getattr_err;
309 ulong_t segvn_setpgsz_eof_err;
310 ulong_t segvn_faultvnmpss_align_err1;
311 ulong_t segvn_faultvnmpss_align_err2;
312 ulong_t segvn_faultvnmpss_align_err3;
313 ulong_t segvn_faultvnmpss_align_err4;
314 ulong_t segvn_faultvnmpss_align_err5;
315 ulong_t segvn_vmpss_pageio_deadlk_err;
316
317 int segvn_use_regions = 1;
318
319 /*
320 * Segvn supports text replication optimization for NUMA platforms. Text
321 * replica's are represented by anon maps (amp). There's one amp per text file
322 * region per lgroup. A process chooses the amp for each of its text mappings
323 * based on the lgroup assignment of its main thread (t_tid = 1). All
324 * processes that want a replica on a particular lgroup for the same text file
325 * mapping share the same amp. amp's are looked up in svntr_hashtab hash table
326 * with vp,off,size,szc used as a key. Text replication segments are read only
327 * MAP_PRIVATE|MAP_TEXT segments that map vnode. Replication is achieved by
328 * forcing COW faults from vnode to amp and mapping amp pages instead of vnode
329 * pages. Replication amp is assigned to a segment when it gets its first
330 * pagefault. To handle main thread lgroup rehoming segvn_trasync_thread
331 * rechecks periodically if the process still maps an amp local to the main
332 * thread. If not async thread forces process to remap to an amp in the new
333 * home lgroup of the main thread. Current text replication implementation
334 * only provides the benefit to workloads that do most of their work in the
335 * main thread of a process or all the threads of a process run in the same
336 * lgroup. To extend text replication benefit to different types of
337 * multithreaded workloads further work would be needed in the hat layer to
338 * allow the same virtual address in the same hat to simultaneously map
339 * different physical addresses (i.e. page table replication would be needed
340 * for x86).
341 *
342 * amp pages are used instead of vnode pages as long as segment has a very
343 * simple life cycle. It's created via segvn_create(), handles S_EXEC
344 * (S_READ) pagefaults and is fully unmapped. If anything more complicated
345 * happens such as protection is changed, real COW fault happens, pagesize is
346 * changed, MC_LOCK is requested or segment is partially unmapped we turn off
347 * text replication by converting the segment back to vnode only segment
348 * (unmap segment's address range and set svd->amp to NULL).
349 *
350 * The original file can be changed after amp is inserted into
351 * svntr_hashtab. Processes that are launched after the file is already
352 * changed can't use the replica's created prior to the file change. To
353 * implement this functionality hash entries are timestamped. Replica's can
354 * only be used if current file modification time is the same as the timestamp
355 * saved when hash entry was created. However just timestamps alone are not
356 * sufficient to detect file modification via mmap(MAP_SHARED) mappings. We
357 * deal with file changes via MAP_SHARED mappings differently. When writable
358 * MAP_SHARED mappings are created to vnodes marked as executable we mark all
359 * existing replica's for this vnode as not usable for future text
360 * mappings. And we don't create new replica's for files that currently have
361 * potentially writable MAP_SHARED mappings (i.e. vn_is_mapped(V_WRITE) is
362 * true).
363 */
364
365 #define SEGVN_TEXTREPL_MAXBYTES_FACTOR (20)
366 size_t segvn_textrepl_max_bytes_factor = SEGVN_TEXTREPL_MAXBYTES_FACTOR;
367
368 static ulong_t svntr_hashtab_sz = 512;
369 static svntr_bucket_t *svntr_hashtab = NULL;
370 static struct kmem_cache *svntr_cache;
371 static svntr_stats_t *segvn_textrepl_stats;
372 static ksema_t segvn_trasync_sem;
373
374 int segvn_disable_textrepl = 1;
375 size_t textrepl_size_thresh = (size_t)-1;
376 size_t segvn_textrepl_bytes = 0;
377 size_t segvn_textrepl_max_bytes = 0;
378 clock_t segvn_update_textrepl_interval = 0;
379 int segvn_update_tr_time = 10;
380 int segvn_disable_textrepl_update = 0;
381
382 static void segvn_textrepl(struct seg *);
383 static void segvn_textunrepl(struct seg *, int);
384 static void segvn_inval_trcache(vnode_t *);
385 static void segvn_trasync_thread(void);
386 static void segvn_trupdate_wakeup(void *);
387 static void segvn_trupdate(void);
388 static void segvn_trupdate_seg(struct seg *, segvn_data_t *, svntr_t *,
389 ulong_t);
390
391 /*
392 * Initialize segvn data structures
393 */
394 void
395 segvn_init(void)
396 {
397 uint_t maxszc;
398 uint_t szc;
399 size_t pgsz;
400
401 segvn_cache = kmem_cache_create("segvn_cache",
402 sizeof (struct segvn_data), 0,
403 segvn_cache_constructor, segvn_cache_destructor, NULL,
404 NULL, NULL, 0);
405
406 if (segvn_lpg_disable == 0) {
407 szc = maxszc = page_num_pagesizes() - 1;
408 if (szc == 0) {
409 segvn_lpg_disable = 1;
410 }
411 if (page_get_pagesize(0) != PAGESIZE) {
412 panic("segvn_init: bad szc 0");
413 /*NOTREACHED*/
414 }
415 while (szc != 0) {
416 pgsz = page_get_pagesize(szc);
417 if (pgsz <= PAGESIZE || !IS_P2ALIGNED(pgsz, pgsz)) {
418 panic("segvn_init: bad szc %d", szc);
419 /*NOTREACHED*/
420 }
421 szc--;
422 }
423 if (segvn_maxpgszc == 0 || segvn_maxpgszc > maxszc)
424 segvn_maxpgszc = maxszc;
425 }
426
427 if (segvn_maxpgszc) {
428 segvn_szc_cache = (struct kmem_cache **)kmem_alloc(
429 (segvn_maxpgszc + 1) * sizeof (struct kmem_cache *),
430 KM_SLEEP);
431 }
432
433 for (szc = 1; szc <= segvn_maxpgszc; szc++) {
434 char str[32];
435
436 (void) sprintf(str, "segvn_szc_cache%d", szc);
437 segvn_szc_cache[szc] = kmem_cache_create(str,
438 page_get_pagecnt(szc) * sizeof (page_t *), 0,
439 NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG);
440 }
441
442
443 if (segvn_use_regions && !hat_supported(HAT_SHARED_REGIONS, NULL))
444 segvn_use_regions = 0;
445
446 /*
447 * For now shared regions and text replication segvn support
448 * are mutually exclusive. This is acceptable because
449 * currently significant benefit from text replication was
450 * only observed on AMD64 NUMA platforms (due to relatively
451 * small L2$ size) and currently we don't support shared
452 * regions on x86.
453 */
454 if (segvn_use_regions && !segvn_disable_textrepl) {
455 segvn_disable_textrepl = 1;
456 }
457
458 #if defined(_LP64)
459 if (lgrp_optimizations() && textrepl_size_thresh != (size_t)-1 &&
460 !segvn_disable_textrepl) {
461 ulong_t i;
462 size_t hsz = svntr_hashtab_sz * sizeof (svntr_bucket_t);
463
464 svntr_cache = kmem_cache_create("svntr_cache",
465 sizeof (svntr_t), 0, svntr_cache_constructor, NULL,
466 NULL, NULL, NULL, 0);
467 svntr_hashtab = kmem_zalloc(hsz, KM_SLEEP);
468 for (i = 0; i < svntr_hashtab_sz; i++) {
469 mutex_init(&svntr_hashtab[i].tr_lock, NULL,
470 MUTEX_DEFAULT, NULL);
471 }
472 segvn_textrepl_max_bytes = ptob(physmem) /
473 segvn_textrepl_max_bytes_factor;
474 segvn_textrepl_stats = kmem_zalloc(NCPU *
475 sizeof (svntr_stats_t), KM_SLEEP);
476 sema_init(&segvn_trasync_sem, 0, NULL, SEMA_DEFAULT, NULL);
477 (void) thread_create(NULL, 0, segvn_trasync_thread,
478 NULL, 0, &p0, TS_RUN, minclsyspri);
479 }
480 #endif
481
482 if (!ISP2(segvn_pglock_comb_balign) ||
483 segvn_pglock_comb_balign < PAGESIZE) {
484 segvn_pglock_comb_balign = 1UL << 16; /* 64K */
485 }
486 segvn_pglock_comb_bshift = highbit(segvn_pglock_comb_balign) - 1;
487 segvn_pglock_comb_palign = btop(segvn_pglock_comb_balign);
488 }
489
490 #define SEGVN_PAGEIO ((void *)0x1)
491 #define SEGVN_NOPAGEIO ((void *)0x2)
492
493 static void
494 segvn_setvnode_mpss(vnode_t *vp)
495 {
496 int err;
497
498 ASSERT(vp->v_mpssdata == NULL ||
499 vp->v_mpssdata == SEGVN_PAGEIO ||
500 vp->v_mpssdata == SEGVN_NOPAGEIO);
501
502 if (vp->v_mpssdata == NULL) {
503 if (vn_vmpss_usepageio(vp)) {
504 err = fop_pageio(vp, NULL,
505 0, 0, 0, CRED(), NULL);
506 } else {
507 err = ENOSYS;
508 }
509 /*
510 * set v_mpssdata just once per vnode life
511 * so that it never changes.
512 */
513 mutex_enter(&vp->v_lock);
514 if (vp->v_mpssdata == NULL) {
515 if (err == EINVAL) {
516 vp->v_mpssdata = SEGVN_PAGEIO;
517 } else {
518 vp->v_mpssdata = SEGVN_NOPAGEIO;
519 }
520 }
521 mutex_exit(&vp->v_lock);
522 }
523 }
524
525 int
526 segvn_create(struct seg **segpp, void *argsp)
527 {
528 struct seg *seg = *segpp;
529 extern lgrp_mem_policy_t lgrp_mem_default_policy;
530 struct segvn_crargs *a = (struct segvn_crargs *)argsp;
531 struct segvn_data *svd;
532 size_t swresv = 0;
533 struct cred *cred;
534 struct anon_map *amp;
535 int error = 0;
536 size_t pgsz;
537 lgrp_mem_policy_t mpolicy = lgrp_mem_default_policy;
538 int use_rgn = 0;
539 int trok = 0;
540
541 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
542
543 if (a->type != MAP_PRIVATE && a->type != MAP_SHARED) {
544 panic("segvn_create type");
545 /*NOTREACHED*/
546 }
547
548 /*
549 * Check arguments. If a shared anon structure is given then
550 * it is illegal to also specify a vp.
551 */
552 if (a->amp != NULL && a->vp != NULL) {
553 panic("segvn_create anon_map");
554 /*NOTREACHED*/
555 }
556
557 if (a->type == MAP_PRIVATE && (a->flags & MAP_TEXT) &&
558 a->vp != NULL && a->prot == (PROT_USER | PROT_READ | PROT_EXEC) &&
559 segvn_use_regions) {
560 use_rgn = 1;
561 }
562
563 /* MAP_NORESERVE on a MAP_SHARED segment is meaningless. */
564 if (a->type == MAP_SHARED)
565 a->flags &= ~MAP_NORESERVE;
566
567 if (a->szc != 0) {
568 if (segvn_lpg_disable != 0 || (a->szc == AS_MAP_NO_LPOOB) ||
569 (a->amp != NULL && a->type == MAP_PRIVATE) ||
570 (a->flags & MAP_NORESERVE) || seg->s_as == &kas) {
571 a->szc = 0;
572 } else {
573 if (a->szc > segvn_maxpgszc)
574 a->szc = segvn_maxpgszc;
575 pgsz = page_get_pagesize(a->szc);
576 if (!IS_P2ALIGNED(seg->s_base, pgsz) ||
577 !IS_P2ALIGNED(seg->s_size, pgsz)) {
578 a->szc = 0;
579 } else if (a->vp != NULL) {
580 if (IS_SWAPFSVP(a->vp) || VN_ISKAS(a->vp)) {
581 /*
582 * paranoid check.
583 * hat_page_demote() is not supported
584 * on swapfs pages.
585 */
586 a->szc = 0;
587 } else if (map_addr_vacalign_check(seg->s_base,
588 a->offset & PAGEMASK)) {
589 a->szc = 0;
590 }
591 } else if (a->amp != NULL) {
592 pgcnt_t anum = btopr(a->offset);
593 pgcnt_t pgcnt = page_get_pagecnt(a->szc);
594 if (!IS_P2ALIGNED(anum, pgcnt)) {
595 a->szc = 0;
596 }
597 }
598 }
599 }
600
601 /*
602 * If segment may need private pages, reserve them now.
603 */
604 if (!(a->flags & MAP_NORESERVE) && ((a->vp == NULL && a->amp == NULL) ||
605 (a->type == MAP_PRIVATE && (a->prot & PROT_WRITE)))) {
606 if (anon_resv_zone(seg->s_size,
607 seg->s_as->a_proc->p_zone) == 0)
608 return (EAGAIN);
609 swresv = seg->s_size;
610 }
611
612 /*
613 * Reserve any mapping structures that may be required.
614 *
615 * Don't do it for segments that may use regions. It's currently a
616 * noop in the hat implementations anyway.
617 */
618 if (!use_rgn) {
619 hat_map(seg->s_as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
620 }
621
622 if (a->cred) {
623 cred = a->cred;
624 crhold(cred);
625 } else {
626 crhold(cred = CRED());
627 }
628
629 /* Inform the vnode of the new mapping */
630 if (a->vp != NULL) {
631 error = fop_addmap(a->vp, a->offset & PAGEMASK,
632 seg->s_as, seg->s_base, seg->s_size, a->prot,
633 a->maxprot, a->type, cred, NULL);
634 if (error) {
635 if (swresv != 0) {
636 anon_unresv_zone(swresv,
637 seg->s_as->a_proc->p_zone);
638 }
639 crfree(cred);
640 if (!use_rgn) {
641 hat_unload(seg->s_as->a_hat, seg->s_base,
642 seg->s_size, HAT_UNLOAD_UNMAP);
643 }
644 return (error);
645 }
646 /*
647 * svntr_hashtab will be NULL if we support shared regions.
648 */
649 trok = ((a->flags & MAP_TEXT) &&
650 (seg->s_size > textrepl_size_thresh ||
651 (a->flags & _MAP_TEXTREPL)) &&
652 lgrp_optimizations() && svntr_hashtab != NULL &&
653 a->type == MAP_PRIVATE && swresv == 0 &&
654 !(a->flags & MAP_NORESERVE) &&
655 seg->s_as != &kas && a->vp->v_type == VREG);
656
657 ASSERT(!trok || !use_rgn);
658 }
659
660 /*
661 * MAP_NORESERVE mappings don't count towards the VSZ of a process
662 * until we fault the pages in.
663 */
664 if ((a->vp == NULL || a->vp->v_type != VREG) &&
665 a->flags & MAP_NORESERVE) {
666 seg->s_as->a_resvsize -= seg->s_size;
667 }
668
669 /*
670 * If more than one segment in the address space, and they're adjacent
671 * virtually, try to concatenate them. Don't concatenate if an
672 * explicit anon_map structure was supplied (e.g., SystemV shared
673 * memory) or if we'll use text replication for this segment.
674 */
675 if (a->amp == NULL && !use_rgn && !trok) {
676 struct seg *pseg, *nseg;
677 struct segvn_data *psvd, *nsvd;
678 lgrp_mem_policy_t ppolicy, npolicy;
679 uint_t lgrp_mem_policy_flags = 0;
680
681 /*
682 * Memory policy flags (lgrp_mem_policy_flags) is valid when
683 * extending stack/heap segments.
684 */
685 if ((a->vp == NULL) && (a->type == MAP_PRIVATE) &&
686 !(a->flags & MAP_NORESERVE) && (seg->s_as != &kas)) {
687 lgrp_mem_policy_flags = a->lgrp_mem_policy_flags;
688 } else {
689 /*
690 * Get policy when not extending it from another segment
691 */
692 mpolicy = lgrp_mem_policy_default(seg->s_size, a->type);
693 }
694
695 /*
696 * First, try to concatenate the previous and new segments
697 */
698 pseg = AS_SEGPREV(seg->s_as, seg);
699 if (pseg != NULL &&
700 pseg->s_base + pseg->s_size == seg->s_base &&
701 pseg->s_ops == &segvn_ops) {
702 /*
703 * Get memory allocation policy from previous segment.
704 * When extension is specified (e.g. for heap) apply
705 * this policy to the new segment regardless of the
706 * outcome of segment concatenation. Extension occurs
707 * for non-default policy otherwise default policy is
708 * used and is based on extended segment size.
709 */
710 psvd = (struct segvn_data *)pseg->s_data;
711 ppolicy = psvd->policy_info.mem_policy;
712 if (lgrp_mem_policy_flags ==
713 LGRP_MP_FLAG_EXTEND_UP) {
714 if (ppolicy != lgrp_mem_default_policy) {
715 mpolicy = ppolicy;
716 } else {
717 mpolicy = lgrp_mem_policy_default(
718 pseg->s_size + seg->s_size,
719 a->type);
720 }
721 }
722
723 if (mpolicy == ppolicy &&
724 (pseg->s_size + seg->s_size <=
725 segvn_comb_thrshld || psvd->amp == NULL) &&
726 segvn_extend_prev(pseg, seg, a, swresv) == 0) {
727 /*
728 * success! now try to concatenate
729 * with following seg
730 */
731 crfree(cred);
732 nseg = AS_SEGNEXT(pseg->s_as, pseg);
733 if (nseg != NULL &&
734 nseg != pseg &&
735 nseg->s_ops == &segvn_ops &&
736 pseg->s_base + pseg->s_size ==
737 nseg->s_base)
738 (void) segvn_concat(pseg, nseg, 0);
739 ASSERT(pseg->s_szc == 0 ||
740 (a->szc == pseg->s_szc &&
741 IS_P2ALIGNED(pseg->s_base, pgsz) &&
742 IS_P2ALIGNED(pseg->s_size, pgsz)));
743 /*
744 * Communicate out the newly concatenated
745 * segment as part of the result.
746 */
747 *segpp = pseg;
748 return (0);
749 }
750 }
751
752 /*
753 * Failed, so try to concatenate with following seg
754 */
755 nseg = AS_SEGNEXT(seg->s_as, seg);
756 if (nseg != NULL &&
757 seg->s_base + seg->s_size == nseg->s_base &&
758 nseg->s_ops == &segvn_ops) {
759 /*
760 * Get memory allocation policy from next segment.
761 * When extension is specified (e.g. for stack) apply
762 * this policy to the new segment regardless of the
763 * outcome of segment concatenation. Extension occurs
764 * for non-default policy otherwise default policy is
765 * used and is based on extended segment size.
766 */
767 nsvd = (struct segvn_data *)nseg->s_data;
768 npolicy = nsvd->policy_info.mem_policy;
769 if (lgrp_mem_policy_flags ==
770 LGRP_MP_FLAG_EXTEND_DOWN) {
771 if (npolicy != lgrp_mem_default_policy) {
772 mpolicy = npolicy;
773 } else {
774 mpolicy = lgrp_mem_policy_default(
775 nseg->s_size + seg->s_size,
776 a->type);
777 }
778 }
779
780 if (mpolicy == npolicy &&
781 segvn_extend_next(seg, nseg, a, swresv) == 0) {
782 crfree(cred);
783 ASSERT(nseg->s_szc == 0 ||
784 (a->szc == nseg->s_szc &&
785 IS_P2ALIGNED(nseg->s_base, pgsz) &&
786 IS_P2ALIGNED(nseg->s_size, pgsz)));
787 /*
788 * Communicate out the newly concatenated
789 * segment as part of the result.
790 */
791 *segpp = nseg;
792 return (0);
793 }
794 }
795 }
796
797 if (a->vp != NULL) {
798 VN_HOLD(a->vp);
799 if (a->type == MAP_SHARED)
800 lgrp_shm_policy_init(NULL, a->vp);
801 }
802 svd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
803
804 seg->s_ops = &segvn_ops;
805 seg->s_data = (void *)svd;
806 seg->s_szc = a->szc;
807
808 svd->seg = seg;
809 svd->vp = a->vp;
810 /*
811 * Anonymous mappings have no backing file so the offset is meaningless.
812 */
813 svd->offset = a->vp ? (a->offset & PAGEMASK) : 0;
814 svd->prot = a->prot;
815 svd->maxprot = a->maxprot;
816 svd->pageprot = 0;
817 svd->type = a->type;
818 svd->vpage = NULL;
819 svd->cred = cred;
820 svd->advice = MADV_NORMAL;
821 svd->pageadvice = 0;
822 svd->flags = (ushort_t)a->flags;
823 svd->softlockcnt = 0;
824 svd->softlockcnt_sbase = 0;
825 svd->softlockcnt_send = 0;
826 svd->svn_inz = 0;
827 svd->rcookie = HAT_INVALID_REGION_COOKIE;
828 svd->pageswap = 0;
829
830 if (a->szc != 0 && a->vp != NULL) {
831 segvn_setvnode_mpss(a->vp);
832 }
833 if (svd->type == MAP_SHARED && svd->vp != NULL &&
834 (svd->vp->v_flag & VVMEXEC) && (svd->prot & PROT_WRITE)) {
835 ASSERT(vn_is_mapped(svd->vp, V_WRITE));
836 segvn_inval_trcache(svd->vp);
837 }
838
839 amp = a->amp;
840 if ((svd->amp = amp) == NULL) {
841 svd->anon_index = 0;
842 if (svd->type == MAP_SHARED) {
843 svd->swresv = 0;
844 /*
845 * Shared mappings to a vp need no other setup.
846 * If we have a shared mapping to an anon_map object
847 * which hasn't been allocated yet, allocate the
848 * struct now so that it will be properly shared
849 * by remembering the swap reservation there.
850 */
851 if (a->vp == NULL) {
852 svd->amp = anonmap_alloc(seg->s_size, swresv,
853 ANON_SLEEP);
854 svd->amp->a_szc = seg->s_szc;
855 }
856 } else {
857 /*
858 * Private mapping (with or without a vp).
859 * Allocate anon_map when needed.
860 */
861 svd->swresv = swresv;
862 }
863 } else {
864 pgcnt_t anon_num;
865
866 /*
867 * Mapping to an existing anon_map structure without a vp.
868 * For now we will insure that the segment size isn't larger
869 * than the size - offset gives us. Later on we may wish to
870 * have the anon array dynamically allocated itself so that
871 * we don't always have to allocate all the anon pointer slots.
872 * This of course involves adding extra code to check that we
873 * aren't trying to use an anon pointer slot beyond the end
874 * of the currently allocated anon array.
875 */
876 if ((amp->size - a->offset) < seg->s_size) {
877 panic("segvn_create anon_map size");
878 /*NOTREACHED*/
879 }
880
881 anon_num = btopr(a->offset);
882
883 if (a->type == MAP_SHARED) {
884 /*
885 * SHARED mapping to a given anon_map.
886 */
887 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
888 amp->refcnt++;
889 if (a->szc > amp->a_szc) {
890 amp->a_szc = a->szc;
891 }
892 ANON_LOCK_EXIT(&amp->a_rwlock);
893 svd->anon_index = anon_num;
894 svd->swresv = 0;
895 } else {
896 /*
897 * PRIVATE mapping to a given anon_map.
898 * Make sure that all the needed anon
899 * structures are created (so that we will
900 * share the underlying pages if nothing
901 * is written by this mapping) and then
902 * duplicate the anon array as is done
903 * when a privately mapped segment is dup'ed.
904 */
905 struct anon *ap;
906 caddr_t addr;
907 caddr_t eaddr;
908 ulong_t anon_idx;
909 int hat_flag = HAT_LOAD;
910
911 if (svd->flags & MAP_TEXT) {
912 hat_flag |= HAT_LOAD_TEXT;
913 }
914
915 svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
916 svd->amp->a_szc = seg->s_szc;
917 svd->anon_index = 0;
918 svd->swresv = swresv;
919
920 /*
921 * Prevent 2 threads from allocating anon
922 * slots simultaneously.
923 */
924 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
925 eaddr = seg->s_base + seg->s_size;
926
927 for (anon_idx = anon_num, addr = seg->s_base;
928 addr < eaddr; addr += PAGESIZE, anon_idx++) {
929 page_t *pp;
930
931 if ((ap = anon_get_ptr(amp->ahp,
932 anon_idx)) != NULL)
933 continue;
934
935 /*
936 * Allocate the anon struct now.
937 * Might as well load up translation
938 * to the page while we're at it...
939 */
940 pp = anon_zero(seg, addr, &ap, cred);
941 if (ap == NULL || pp == NULL) {
942 panic("segvn_create anon_zero");
943 /*NOTREACHED*/
944 }
945
946 /*
947 * Re-acquire the anon_map lock and
948 * initialize the anon array entry.
949 */
950 ASSERT(anon_get_ptr(amp->ahp,
951 anon_idx) == NULL);
952 (void) anon_set_ptr(amp->ahp, anon_idx, ap,
953 ANON_SLEEP);
954
955 ASSERT(seg->s_szc == 0);
956 ASSERT(!IS_VMODSORT(pp->p_vnode));
957
958 ASSERT(use_rgn == 0);
959 hat_memload(seg->s_as->a_hat, addr, pp,
960 svd->prot & ~PROT_WRITE, hat_flag);
961
962 page_unlock(pp);
963 }
964 ASSERT(seg->s_szc == 0);
965 anon_dup(amp->ahp, anon_num, svd->amp->ahp,
966 0, seg->s_size);
967 ANON_LOCK_EXIT(&amp->a_rwlock);
968 }
969 }
970
971 /*
972 * Set default memory allocation policy for segment
973 *
974 * Always set policy for private memory at least for initialization
975 * even if this is a shared memory segment
976 */
977 (void) lgrp_privm_policy_set(mpolicy, &svd->policy_info, seg->s_size);
978
979 if (svd->type == MAP_SHARED)
980 (void) lgrp_shm_policy_set(mpolicy, svd->amp, svd->anon_index,
981 svd->vp, svd->offset, seg->s_size);
982
983 if (use_rgn) {
984 ASSERT(!trok);
985 ASSERT(svd->amp == NULL);
986 svd->rcookie = hat_join_region(seg->s_as->a_hat, seg->s_base,
987 seg->s_size, (void *)svd->vp, svd->offset, svd->prot,
988 (uchar_t)seg->s_szc, segvn_hat_rgn_unload_callback,
989 HAT_REGION_TEXT);
990 }
991
992 ASSERT(!trok || !(svd->prot & PROT_WRITE));
993 svd->tr_state = trok ? SEGVN_TR_INIT : SEGVN_TR_OFF;
994
995 return (0);
996 }
997
998 /*
999 * Concatenate two existing segments, if possible.
1000 * Return 0 on success, -1 if two segments are not compatible
1001 * or -2 on memory allocation failure.
1002 * If amp_cat == 1 then try and concat segments with anon maps
1003 */
1004 static int
1005 segvn_concat(struct seg *seg1, struct seg *seg2, int amp_cat)
1006 {
1007 struct segvn_data *svd1 = seg1->s_data;
1008 struct segvn_data *svd2 = seg2->s_data;
1009 struct anon_map *amp1 = svd1->amp;
1010 struct anon_map *amp2 = svd2->amp;
1011 struct vpage *vpage1 = svd1->vpage;
1012 struct vpage *vpage2 = svd2->vpage, *nvpage = NULL;
1013 size_t size, nvpsize;
1014 pgcnt_t npages1, npages2;
1015
1016 ASSERT(seg1->s_as && seg2->s_as && seg1->s_as == seg2->s_as);
1017 ASSERT(AS_WRITE_HELD(seg1->s_as));
1018 ASSERT(seg1->s_ops == seg2->s_ops);
1019
1020 if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie) ||
1021 HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1022 return (-1);
1023 }
1024
1025 /* both segments exist, try to merge them */
1026 #define incompat(x) (svd1->x != svd2->x)
1027 if (incompat(vp) || incompat(maxprot) ||
1028 (!svd1->pageadvice && !svd2->pageadvice && incompat(advice)) ||
1029 (!svd1->pageprot && !svd2->pageprot && incompat(prot)) ||
1030 incompat(type) || incompat(cred) || incompat(flags) ||
1031 seg1->s_szc != seg2->s_szc || incompat(policy_info.mem_policy) ||
1032 (svd2->softlockcnt > 0) || svd1->softlockcnt_send > 0)
1033 return (-1);
1034 #undef incompat
1035
1036 /*
1037 * vp == NULL implies zfod, offset doesn't matter
1038 */
1039 if (svd1->vp != NULL &&
1040 svd1->offset + seg1->s_size != svd2->offset) {
1041 return (-1);
1042 }
1043
1044 /*
1045 * Don't concatenate if either segment uses text replication.
1046 */
1047 if (svd1->tr_state != SEGVN_TR_OFF || svd2->tr_state != SEGVN_TR_OFF) {
1048 return (-1);
1049 }
1050
1051 /*
1052 * Fail early if we're not supposed to concatenate
1053 * segments with non NULL amp.
1054 */
1055 if (amp_cat == 0 && (amp1 != NULL || amp2 != NULL)) {
1056 return (-1);
1057 }
1058
1059 if (svd1->vp == NULL && svd1->type == MAP_SHARED) {
1060 if (amp1 != amp2) {
1061 return (-1);
1062 }
1063 if (amp1 != NULL && svd1->anon_index + btop(seg1->s_size) !=
1064 svd2->anon_index) {
1065 return (-1);
1066 }
1067 ASSERT(amp1 == NULL || amp1->refcnt >= 2);
1068 }
1069
1070 /*
1071 * If either seg has vpages, create a new merged vpage array.
1072 */
1073 if (vpage1 != NULL || vpage2 != NULL) {
1074 struct vpage *vp, *evp;
1075
1076 npages1 = seg_pages(seg1);
1077 npages2 = seg_pages(seg2);
1078 nvpsize = vpgtob(npages1 + npages2);
1079
1080 if ((nvpage = kmem_zalloc(nvpsize, KM_NOSLEEP)) == NULL) {
1081 return (-2);
1082 }
1083
1084 if (vpage1 != NULL) {
1085 bcopy(vpage1, nvpage, vpgtob(npages1));
1086 } else {
1087 evp = nvpage + npages1;
1088 for (vp = nvpage; vp < evp; vp++) {
1089 VPP_SETPROT(vp, svd1->prot);
1090 VPP_SETADVICE(vp, svd1->advice);
1091 }
1092 }
1093
1094 if (vpage2 != NULL) {
1095 bcopy(vpage2, nvpage + npages1, vpgtob(npages2));
1096 } else {
1097 evp = nvpage + npages1 + npages2;
1098 for (vp = nvpage + npages1; vp < evp; vp++) {
1099 VPP_SETPROT(vp, svd2->prot);
1100 VPP_SETADVICE(vp, svd2->advice);
1101 }
1102 }
1103
1104 if (svd2->pageswap && (!svd1->pageswap && svd1->swresv)) {
1105 ASSERT(svd1->swresv == seg1->s_size);
1106 ASSERT(!(svd1->flags & MAP_NORESERVE));
1107 ASSERT(!(svd2->flags & MAP_NORESERVE));
1108 evp = nvpage + npages1;
1109 for (vp = nvpage; vp < evp; vp++) {
1110 VPP_SETSWAPRES(vp);
1111 }
1112 }
1113
1114 if (svd1->pageswap && (!svd2->pageswap && svd2->swresv)) {
1115 ASSERT(svd2->swresv == seg2->s_size);
1116 ASSERT(!(svd1->flags & MAP_NORESERVE));
1117 ASSERT(!(svd2->flags & MAP_NORESERVE));
1118 vp = nvpage + npages1;
1119 evp = vp + npages2;
1120 for (; vp < evp; vp++) {
1121 VPP_SETSWAPRES(vp);
1122 }
1123 }
1124 }
1125 ASSERT((vpage1 != NULL || vpage2 != NULL) ||
1126 (svd1->pageswap == 0 && svd2->pageswap == 0));
1127
1128 /*
1129 * If either segment has private pages, create a new merged anon
1130 * array. If mergeing shared anon segments just decrement anon map's
1131 * refcnt.
1132 */
1133 if (amp1 != NULL && svd1->type == MAP_SHARED) {
1134 ASSERT(amp1 == amp2 && svd1->vp == NULL);
1135 ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1136 ASSERT(amp1->refcnt >= 2);
1137 amp1->refcnt--;
1138 ANON_LOCK_EXIT(&amp1->a_rwlock);
1139 svd2->amp = NULL;
1140 } else if (amp1 != NULL || amp2 != NULL) {
1141 struct anon_hdr *nahp;
1142 struct anon_map *namp = NULL;
1143 size_t asize;
1144
1145 ASSERT(svd1->type == MAP_PRIVATE);
1146
1147 asize = seg1->s_size + seg2->s_size;
1148 if ((nahp = anon_create(btop(asize), ANON_NOSLEEP)) == NULL) {
1149 if (nvpage != NULL) {
1150 kmem_free(nvpage, nvpsize);
1151 }
1152 return (-2);
1153 }
1154 if (amp1 != NULL) {
1155 /*
1156 * XXX anon rwlock is not really needed because
1157 * this is a private segment and we are writers.
1158 */
1159 ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1160 ASSERT(amp1->refcnt == 1);
1161 if (anon_copy_ptr(amp1->ahp, svd1->anon_index,
1162 nahp, 0, btop(seg1->s_size), ANON_NOSLEEP)) {
1163 anon_release(nahp, btop(asize));
1164 ANON_LOCK_EXIT(&amp1->a_rwlock);
1165 if (nvpage != NULL) {
1166 kmem_free(nvpage, nvpsize);
1167 }
1168 return (-2);
1169 }
1170 }
1171 if (amp2 != NULL) {
1172 ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1173 ASSERT(amp2->refcnt == 1);
1174 if (anon_copy_ptr(amp2->ahp, svd2->anon_index,
1175 nahp, btop(seg1->s_size), btop(seg2->s_size),
1176 ANON_NOSLEEP)) {
1177 anon_release(nahp, btop(asize));
1178 ANON_LOCK_EXIT(&amp2->a_rwlock);
1179 if (amp1 != NULL) {
1180 ANON_LOCK_EXIT(&amp1->a_rwlock);
1181 }
1182 if (nvpage != NULL) {
1183 kmem_free(nvpage, nvpsize);
1184 }
1185 return (-2);
1186 }
1187 }
1188 if (amp1 != NULL) {
1189 namp = amp1;
1190 anon_release(amp1->ahp, btop(amp1->size));
1191 }
1192 if (amp2 != NULL) {
1193 if (namp == NULL) {
1194 ASSERT(amp1 == NULL);
1195 namp = amp2;
1196 anon_release(amp2->ahp, btop(amp2->size));
1197 } else {
1198 amp2->refcnt--;
1199 ANON_LOCK_EXIT(&amp2->a_rwlock);
1200 anonmap_free(amp2);
1201 }
1202 svd2->amp = NULL; /* needed for seg_free */
1203 }
1204 namp->ahp = nahp;
1205 namp->size = asize;
1206 svd1->amp = namp;
1207 svd1->anon_index = 0;
1208 ANON_LOCK_EXIT(&namp->a_rwlock);
1209 }
1210 /*
1211 * Now free the old vpage structures.
1212 */
1213 if (nvpage != NULL) {
1214 if (vpage1 != NULL) {
1215 kmem_free(vpage1, vpgtob(npages1));
1216 }
1217 if (vpage2 != NULL) {
1218 svd2->vpage = NULL;
1219 kmem_free(vpage2, vpgtob(npages2));
1220 }
1221 if (svd2->pageprot) {
1222 svd1->pageprot = 1;
1223 }
1224 if (svd2->pageadvice) {
1225 svd1->pageadvice = 1;
1226 }
1227 if (svd2->pageswap) {
1228 svd1->pageswap = 1;
1229 }
1230 svd1->vpage = nvpage;
1231 }
1232
1233 /* all looks ok, merge segments */
1234 svd1->swresv += svd2->swresv;
1235 svd2->swresv = 0; /* so seg_free doesn't release swap space */
1236 size = seg2->s_size;
1237 seg_free(seg2);
1238 seg1->s_size += size;
1239 return (0);
1240 }
1241
1242 /*
1243 * Extend the previous segment (seg1) to include the
1244 * new segment (seg2 + a), if possible.
1245 * Return 0 on success.
1246 */
1247 static int
1248 segvn_extend_prev(struct seg *seg1, struct seg *seg2, struct segvn_crargs *a,
1249 size_t swresv)
1250 {
1251 struct segvn_data *svd1 = (struct segvn_data *)seg1->s_data;
1252 size_t size;
1253 struct anon_map *amp1;
1254 struct vpage *new_vpage;
1255
1256 /*
1257 * We don't need any segment level locks for "segvn" data
1258 * since the address space is "write" locked.
1259 */
1260 ASSERT(seg1->s_as && AS_WRITE_HELD(seg1->s_as));
1261
1262 if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie)) {
1263 return (-1);
1264 }
1265
1266 /* second segment is new, try to extend first */
1267 /* XXX - should also check cred */
1268 if (svd1->vp != a->vp || svd1->maxprot != a->maxprot ||
1269 (!svd1->pageprot && (svd1->prot != a->prot)) ||
1270 svd1->type != a->type || svd1->flags != a->flags ||
1271 seg1->s_szc != a->szc || svd1->softlockcnt_send > 0)
1272 return (-1);
1273
1274 /* vp == NULL implies zfod, offset doesn't matter */
1275 if (svd1->vp != NULL &&
1276 svd1->offset + seg1->s_size != (a->offset & PAGEMASK))
1277 return (-1);
1278
1279 if (svd1->tr_state != SEGVN_TR_OFF) {
1280 return (-1);
1281 }
1282
1283 amp1 = svd1->amp;
1284 if (amp1) {
1285 pgcnt_t newpgs;
1286
1287 /*
1288 * Segment has private pages, can data structures
1289 * be expanded?
1290 *
1291 * Acquire the anon_map lock to prevent it from changing,
1292 * if it is shared. This ensures that the anon_map
1293 * will not change while a thread which has a read/write
1294 * lock on an address space references it.
1295 * XXX - Don't need the anon_map lock at all if "refcnt"
1296 * is 1.
1297 *
1298 * Can't grow a MAP_SHARED segment with an anonmap because
1299 * there may be existing anon slots where we want to extend
1300 * the segment and we wouldn't know what to do with them
1301 * (e.g., for tmpfs right thing is to just leave them there,
1302 * for /dev/zero they should be cleared out).
1303 */
1304 if (svd1->type == MAP_SHARED)
1305 return (-1);
1306
1307 ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1308 if (amp1->refcnt > 1) {
1309 ANON_LOCK_EXIT(&amp1->a_rwlock);
1310 return (-1);
1311 }
1312 newpgs = anon_grow(amp1->ahp, &svd1->anon_index,
1313 btop(seg1->s_size), btop(seg2->s_size), ANON_NOSLEEP);
1314
1315 if (newpgs == 0) {
1316 ANON_LOCK_EXIT(&amp1->a_rwlock);
1317 return (-1);
1318 }
1319 amp1->size = ptob(newpgs);
1320 ANON_LOCK_EXIT(&amp1->a_rwlock);
1321 }
1322 if (svd1->vpage != NULL) {
1323 struct vpage *vp, *evp;
1324 new_vpage =
1325 kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1326 KM_NOSLEEP);
1327 if (new_vpage == NULL)
1328 return (-1);
1329 bcopy(svd1->vpage, new_vpage, vpgtob(seg_pages(seg1)));
1330 kmem_free(svd1->vpage, vpgtob(seg_pages(seg1)));
1331 svd1->vpage = new_vpage;
1332
1333 vp = new_vpage + seg_pages(seg1);
1334 evp = vp + seg_pages(seg2);
1335 for (; vp < evp; vp++)
1336 VPP_SETPROT(vp, a->prot);
1337 if (svd1->pageswap && swresv) {
1338 ASSERT(!(svd1->flags & MAP_NORESERVE));
1339 ASSERT(swresv == seg2->s_size);
1340 vp = new_vpage + seg_pages(seg1);
1341 for (; vp < evp; vp++) {
1342 VPP_SETSWAPRES(vp);
1343 }
1344 }
1345 }
1346 ASSERT(svd1->vpage != NULL || svd1->pageswap == 0);
1347 size = seg2->s_size;
1348 seg_free(seg2);
1349 seg1->s_size += size;
1350 svd1->swresv += swresv;
1351 if (svd1->pageprot && (a->prot & PROT_WRITE) &&
1352 svd1->type == MAP_SHARED && svd1->vp != NULL &&
1353 (svd1->vp->v_flag & VVMEXEC)) {
1354 ASSERT(vn_is_mapped(svd1->vp, V_WRITE));
1355 segvn_inval_trcache(svd1->vp);
1356 }
1357 return (0);
1358 }
1359
1360 /*
1361 * Extend the next segment (seg2) to include the
1362 * new segment (seg1 + a), if possible.
1363 * Return 0 on success.
1364 */
1365 static int
1366 segvn_extend_next(struct seg *seg1, struct seg *seg2, struct segvn_crargs *a,
1367 size_t swresv)
1368 {
1369 struct segvn_data *svd2 = (struct segvn_data *)seg2->s_data;
1370 size_t size;
1371 struct anon_map *amp2;
1372 struct vpage *new_vpage;
1373
1374 /*
1375 * We don't need any segment level locks for "segvn" data
1376 * since the address space is "write" locked.
1377 */
1378 ASSERT(seg2->s_as && AS_WRITE_HELD(seg2->s_as));
1379
1380 if (HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1381 return (-1);
1382 }
1383
1384 /* first segment is new, try to extend second */
1385 /* XXX - should also check cred */
1386 if (svd2->vp != a->vp || svd2->maxprot != a->maxprot ||
1387 (!svd2->pageprot && (svd2->prot != a->prot)) ||
1388 svd2->type != a->type || svd2->flags != a->flags ||
1389 seg2->s_szc != a->szc || svd2->softlockcnt_sbase > 0)
1390 return (-1);
1391 /* vp == NULL implies zfod, offset doesn't matter */
1392 if (svd2->vp != NULL &&
1393 (a->offset & PAGEMASK) + seg1->s_size != svd2->offset)
1394 return (-1);
1395
1396 if (svd2->tr_state != SEGVN_TR_OFF) {
1397 return (-1);
1398 }
1399
1400 amp2 = svd2->amp;
1401 if (amp2) {
1402 pgcnt_t newpgs;
1403
1404 /*
1405 * Segment has private pages, can data structures
1406 * be expanded?
1407 *
1408 * Acquire the anon_map lock to prevent it from changing,
1409 * if it is shared. This ensures that the anon_map
1410 * will not change while a thread which has a read/write
1411 * lock on an address space references it.
1412 *
1413 * XXX - Don't need the anon_map lock at all if "refcnt"
1414 * is 1.
1415 */
1416 if (svd2->type == MAP_SHARED)
1417 return (-1);
1418
1419 ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1420 if (amp2->refcnt > 1) {
1421 ANON_LOCK_EXIT(&amp2->a_rwlock);
1422 return (-1);
1423 }
1424 newpgs = anon_grow(amp2->ahp, &svd2->anon_index,
1425 btop(seg2->s_size), btop(seg1->s_size),
1426 ANON_NOSLEEP | ANON_GROWDOWN);
1427
1428 if (newpgs == 0) {
1429 ANON_LOCK_EXIT(&amp2->a_rwlock);
1430 return (-1);
1431 }
1432 amp2->size = ptob(newpgs);
1433 ANON_LOCK_EXIT(&amp2->a_rwlock);
1434 }
1435 if (svd2->vpage != NULL) {
1436 struct vpage *vp, *evp;
1437 new_vpage =
1438 kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1439 KM_NOSLEEP);
1440 if (new_vpage == NULL) {
1441 /* Not merging segments so adjust anon_index back */
1442 if (amp2)
1443 svd2->anon_index += seg_pages(seg1);
1444 return (-1);
1445 }
1446 bcopy(svd2->vpage, new_vpage + seg_pages(seg1),
1447 vpgtob(seg_pages(seg2)));
1448 kmem_free(svd2->vpage, vpgtob(seg_pages(seg2)));
1449 svd2->vpage = new_vpage;
1450
1451 vp = new_vpage;
1452 evp = vp + seg_pages(seg1);
1453 for (; vp < evp; vp++)
1454 VPP_SETPROT(vp, a->prot);
1455 if (svd2->pageswap && swresv) {
1456 ASSERT(!(svd2->flags & MAP_NORESERVE));
1457 ASSERT(swresv == seg1->s_size);
1458 vp = new_vpage;
1459 for (; vp < evp; vp++) {
1460 VPP_SETSWAPRES(vp);
1461 }
1462 }
1463 }
1464 ASSERT(svd2->vpage != NULL || svd2->pageswap == 0);
1465 size = seg1->s_size;
1466 seg_free(seg1);
1467 seg2->s_size += size;
1468 seg2->s_base -= size;
1469 svd2->offset -= size;
1470 svd2->swresv += swresv;
1471 if (svd2->pageprot && (a->prot & PROT_WRITE) &&
1472 svd2->type == MAP_SHARED && svd2->vp != NULL &&
1473 (svd2->vp->v_flag & VVMEXEC)) {
1474 ASSERT(vn_is_mapped(svd2->vp, V_WRITE));
1475 segvn_inval_trcache(svd2->vp);
1476 }
1477 return (0);
1478 }
1479
1480 /*
1481 * Duplicate all the pages in the segment. This may break COW sharing for a
1482 * given page. If the page is marked with inherit zero set, then instead of
1483 * duplicating the page, we zero the page.
1484 */
1485 static int
1486 segvn_dup_pages(struct seg *seg, struct seg *newseg)
1487 {
1488 int error;
1489 uint_t prot;
1490 page_t *pp;
1491 struct anon *ap, *newap;
1492 size_t i;
1493 caddr_t addr;
1494
1495 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1496 struct segvn_data *newsvd = (struct segvn_data *)newseg->s_data;
1497 ulong_t old_idx = svd->anon_index;
1498 ulong_t new_idx = 0;
1499
1500 i = btopr(seg->s_size);
1501 addr = seg->s_base;
1502
1503 /*
1504 * XXX break cow sharing using PAGESIZE
1505 * pages. They will be relocated into larger
1506 * pages at fault time.
1507 */
1508 while (i-- > 0) {
1509 if ((ap = anon_get_ptr(svd->amp->ahp, old_idx)) != NULL) {
1510 struct vpage *vpp;
1511
1512 vpp = &svd->vpage[seg_page(seg, addr)];
1513
1514 /*
1515 * prot need not be computed below 'cause anon_private
1516 * is going to ignore it anyway as child doesn't inherit
1517 * pagelock from parent.
1518 */
1519 prot = svd->pageprot ? VPP_PROT(vpp) : svd->prot;
1520
1521 /*
1522 * Check whether we should zero this or dup it.
1523 */
1524 if (svd->svn_inz == SEGVN_INZ_ALL ||
1525 (svd->svn_inz == SEGVN_INZ_VPP &&
1526 VPP_ISINHZERO(vpp))) {
1527 pp = anon_zero(newseg, addr, &newap,
1528 newsvd->cred);
1529 } else {
1530 page_t *anon_pl[1+1];
1531 uint_t vpprot;
1532 error = anon_getpage(&ap, &vpprot, anon_pl,
1533 PAGESIZE, seg, addr, S_READ, svd->cred);
1534 if (error != 0)
1535 return (error);
1536
1537 pp = anon_private(&newap, newseg, addr, prot,
1538 anon_pl[0], 0, newsvd->cred);
1539 }
1540 if (pp == NULL) {
1541 return (ENOMEM);
1542 }
1543 (void) anon_set_ptr(newsvd->amp->ahp, new_idx, newap,
1544 ANON_SLEEP);
1545 page_unlock(pp);
1546 }
1547 addr += PAGESIZE;
1548 old_idx++;
1549 new_idx++;
1550 }
1551
1552 return (0);
1553 }
1554
1555 static int
1556 segvn_dup(struct seg *seg, struct seg *newseg)
1557 {
1558 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1559 struct segvn_data *newsvd;
1560 pgcnt_t npages = seg_pages(seg);
1561 int error = 0;
1562 size_t len;
1563 struct anon_map *amp;
1564
1565 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1566 ASSERT(newseg->s_as->a_proc->p_parent == curproc);
1567
1568 /*
1569 * If segment has anon reserved, reserve more for the new seg.
1570 * For a MAP_NORESERVE segment swresv will be a count of all the
1571 * allocated anon slots; thus we reserve for the child as many slots
1572 * as the parent has allocated. This semantic prevents the child or
1573 * parent from dieing during a copy-on-write fault caused by trying
1574 * to write a shared pre-existing anon page.
1575 */
1576 if ((len = svd->swresv) != 0) {
1577 if (anon_resv(svd->swresv) == 0)
1578 return (ENOMEM);
1579 }
1580
1581 newsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
1582
1583 newseg->s_ops = &segvn_ops;
1584 newseg->s_data = (void *)newsvd;
1585 newseg->s_szc = seg->s_szc;
1586
1587 newsvd->seg = newseg;
1588 if ((newsvd->vp = svd->vp) != NULL) {
1589 VN_HOLD(svd->vp);
1590 if (svd->type == MAP_SHARED)
1591 lgrp_shm_policy_init(NULL, svd->vp);
1592 }
1593 newsvd->offset = svd->offset;
1594 newsvd->prot = svd->prot;
1595 newsvd->maxprot = svd->maxprot;
1596 newsvd->pageprot = svd->pageprot;
1597 newsvd->type = svd->type;
1598 newsvd->cred = svd->cred;
1599 crhold(newsvd->cred);
1600 newsvd->advice = svd->advice;
1601 newsvd->pageadvice = svd->pageadvice;
1602 newsvd->svn_inz = svd->svn_inz;
1603 newsvd->swresv = svd->swresv;
1604 newsvd->pageswap = svd->pageswap;
1605 newsvd->flags = svd->flags;
1606 newsvd->softlockcnt = 0;
1607 newsvd->softlockcnt_sbase = 0;
1608 newsvd->softlockcnt_send = 0;
1609 newsvd->policy_info = svd->policy_info;
1610 newsvd->rcookie = HAT_INVALID_REGION_COOKIE;
1611
1612 if ((amp = svd->amp) == NULL || svd->tr_state == SEGVN_TR_ON) {
1613 /*
1614 * Not attaching to a shared anon object.
1615 */
1616 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie) ||
1617 svd->tr_state == SEGVN_TR_OFF);
1618 if (svd->tr_state == SEGVN_TR_ON) {
1619 ASSERT(newsvd->vp != NULL && amp != NULL);
1620 newsvd->tr_state = SEGVN_TR_INIT;
1621 } else {
1622 newsvd->tr_state = svd->tr_state;
1623 }
1624 newsvd->amp = NULL;
1625 newsvd->anon_index = 0;
1626 } else {
1627 /* regions for now are only used on pure vnode segments */
1628 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
1629 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1630 newsvd->tr_state = SEGVN_TR_OFF;
1631 if (svd->type == MAP_SHARED) {
1632 ASSERT(svd->svn_inz == SEGVN_INZ_NONE);
1633 newsvd->amp = amp;
1634 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1635 amp->refcnt++;
1636 ANON_LOCK_EXIT(&amp->a_rwlock);
1637 newsvd->anon_index = svd->anon_index;
1638 } else {
1639 int reclaim = 1;
1640
1641 /*
1642 * Allocate and initialize new anon_map structure.
1643 */
1644 newsvd->amp = anonmap_alloc(newseg->s_size, 0,
1645 ANON_SLEEP);
1646 newsvd->amp->a_szc = newseg->s_szc;
1647 newsvd->anon_index = 0;
1648 ASSERT(svd->svn_inz == SEGVN_INZ_NONE ||
1649 svd->svn_inz == SEGVN_INZ_ALL ||
1650 svd->svn_inz == SEGVN_INZ_VPP);
1651
1652 /*
1653 * We don't have to acquire the anon_map lock
1654 * for the new segment (since it belongs to an
1655 * address space that is still not associated
1656 * with any process), or the segment in the old
1657 * address space (since all threads in it
1658 * are stopped while duplicating the address space).
1659 */
1660
1661 /*
1662 * The goal of the following code is to make sure that
1663 * softlocked pages do not end up as copy on write
1664 * pages. This would cause problems where one
1665 * thread writes to a page that is COW and a different
1666 * thread in the same process has softlocked it. The
1667 * softlock lock would move away from this process
1668 * because the write would cause this process to get
1669 * a copy (without the softlock).
1670 *
1671 * The strategy here is to just break the
1672 * sharing on pages that could possibly be
1673 * softlocked.
1674 *
1675 * In addition, if any pages have been marked that they
1676 * should be inherited as zero, then we immediately go
1677 * ahead and break COW and zero them. In the case of a
1678 * softlocked page that should be inherited zero, we
1679 * break COW and just get a zero page.
1680 */
1681 retry:
1682 if (svd->softlockcnt ||
1683 svd->svn_inz != SEGVN_INZ_NONE) {
1684 /*
1685 * The softlock count might be non zero
1686 * because some pages are still stuck in the
1687 * cache for lazy reclaim. Flush the cache
1688 * now. This should drop the count to zero.
1689 * [or there is really I/O going on to these
1690 * pages]. Note, we have the writers lock so
1691 * nothing gets inserted during the flush.
1692 */
1693 if (svd->softlockcnt && reclaim == 1) {
1694 segvn_purge(seg);
1695 reclaim = 0;
1696 goto retry;
1697 }
1698
1699 error = segvn_dup_pages(seg, newseg);
1700 if (error != 0) {
1701 newsvd->vpage = NULL;
1702 goto out;
1703 }
1704 } else { /* common case */
1705 if (seg->s_szc != 0) {
1706 /*
1707 * If at least one of anon slots of a
1708 * large page exists then make sure
1709 * all anon slots of a large page
1710 * exist to avoid partial cow sharing
1711 * of a large page in the future.
1712 */
1713 anon_dup_fill_holes(amp->ahp,
1714 svd->anon_index, newsvd->amp->ahp,
1715 0, seg->s_size, seg->s_szc,
1716 svd->vp != NULL);
1717 } else {
1718 anon_dup(amp->ahp, svd->anon_index,
1719 newsvd->amp->ahp, 0, seg->s_size);
1720 }
1721
1722 hat_clrattr(seg->s_as->a_hat, seg->s_base,
1723 seg->s_size, PROT_WRITE);
1724 }
1725 }
1726 }
1727 /*
1728 * If necessary, create a vpage structure for the new segment.
1729 * Do not copy any page lock indications.
1730 */
1731 if (svd->vpage != NULL) {
1732 uint_t i;
1733 struct vpage *ovp = svd->vpage;
1734 struct vpage *nvp;
1735
1736 nvp = newsvd->vpage =
1737 kmem_alloc(vpgtob(npages), KM_SLEEP);
1738 for (i = 0; i < npages; i++) {
1739 *nvp = *ovp++;
1740 VPP_CLRPPLOCK(nvp++);
1741 }
1742 } else
1743 newsvd->vpage = NULL;
1744
1745 /* Inform the vnode of the new mapping */
1746 if (newsvd->vp != NULL) {
1747 error = fop_addmap(newsvd->vp, (offset_t)newsvd->offset,
1748 newseg->s_as, newseg->s_base, newseg->s_size, newsvd->prot,
1749 newsvd->maxprot, newsvd->type, newsvd->cred, NULL);
1750 }
1751 out:
1752 if (error == 0 && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1753 ASSERT(newsvd->amp == NULL);
1754 ASSERT(newsvd->tr_state == SEGVN_TR_OFF);
1755 newsvd->rcookie = svd->rcookie;
1756 hat_dup_region(newseg->s_as->a_hat, newsvd->rcookie);
1757 }
1758 return (error);
1759 }
1760
1761
1762 /*
1763 * callback function to invoke free_vp_pages() for only those pages actually
1764 * processed by the HAT when a shared region is destroyed.
1765 */
1766 extern int free_pages;
1767
1768 static void
1769 segvn_hat_rgn_unload_callback(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
1770 size_t r_size, void *r_obj, uoff_t r_objoff)
1771 {
1772 uoff_t off;
1773 size_t len;
1774 vnode_t *vp = (vnode_t *)r_obj;
1775
1776 ASSERT(eaddr > saddr);
1777 ASSERT(saddr >= r_saddr);
1778 ASSERT(saddr < r_saddr + r_size);
1779 ASSERT(eaddr > r_saddr);
1780 ASSERT(eaddr <= r_saddr + r_size);
1781 ASSERT(vp != NULL);
1782
1783 if (!free_pages) {
1784 return;
1785 }
1786
1787 len = eaddr - saddr;
1788 off = (saddr - r_saddr) + r_objoff;
1789 free_vp_pages(&vp->v_object, off, len);
1790 }
1791
1792 /*
1793 * callback function used by segvn_unmap to invoke free_vp_pages() for only
1794 * those pages actually processed by the HAT
1795 */
1796 static void
1797 segvn_hat_unload_callback(hat_callback_t *cb)
1798 {
1799 struct seg *seg = cb->hcb_data;
1800 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1801 size_t len;
1802 uoff_t off;
1803
1804 ASSERT(svd->vp != NULL);
1805 ASSERT(cb->hcb_end_addr > cb->hcb_start_addr);
1806 ASSERT(cb->hcb_start_addr >= seg->s_base);
1807
1808 len = cb->hcb_end_addr - cb->hcb_start_addr;
1809 off = cb->hcb_start_addr - seg->s_base;
1810 free_vp_pages(&svd->vp->v_object, svd->offset + off, len);
1811 }
1812
1813 /*
1814 * This function determines the number of bytes of swap reserved by
1815 * a segment for which per-page accounting is present. It is used to
1816 * calculate the correct value of a segvn_data's swresv.
1817 */
1818 static size_t
1819 segvn_count_swap_by_vpages(struct seg *seg)
1820 {
1821 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1822 struct vpage *vp, *evp;
1823 size_t nswappages = 0;
1824
1825 ASSERT(svd->pageswap);
1826 ASSERT(svd->vpage != NULL);
1827
1828 evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
1829
1830 for (vp = svd->vpage; vp < evp; vp++) {
1831 if (VPP_ISSWAPRES(vp))
1832 nswappages++;
1833 }
1834
1835 return (nswappages << PAGESHIFT);
1836 }
1837
1838 static int
1839 segvn_unmap(struct seg *seg, caddr_t addr, size_t len)
1840 {
1841 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1842 struct segvn_data *nsvd;
1843 struct seg *nseg;
1844 struct anon_map *amp;
1845 pgcnt_t opages; /* old segment size in pages */
1846 pgcnt_t npages; /* new segment size in pages */
1847 pgcnt_t dpages; /* pages being deleted (unmapped) */
1848 hat_callback_t callback; /* used for free_vp_pages() */
1849 hat_callback_t *cbp = NULL;
1850 caddr_t nbase;
1851 size_t nsize;
1852 size_t oswresv;
1853 int reclaim = 1;
1854
1855 /*
1856 * We don't need any segment level locks for "segvn" data
1857 * since the address space is "write" locked.
1858 */
1859 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1860
1861 /*
1862 * Fail the unmap if pages are SOFTLOCKed through this mapping.
1863 * softlockcnt is protected from change by the as write lock.
1864 */
1865 retry:
1866 if (svd->softlockcnt > 0) {
1867 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1868
1869 /*
1870 * If this is shared segment non 0 softlockcnt
1871 * means locked pages are still in use.
1872 */
1873 if (svd->type == MAP_SHARED) {
1874 return (EAGAIN);
1875 }
1876
1877 /*
1878 * since we do have the writers lock nobody can fill
1879 * the cache during the purge. The flush either succeeds
1880 * or we still have pending I/Os.
1881 */
1882 if (reclaim == 1) {
1883 segvn_purge(seg);
1884 reclaim = 0;
1885 goto retry;
1886 }
1887 return (EAGAIN);
1888 }
1889
1890 /*
1891 * Check for bad sizes
1892 */
1893 if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
1894 (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
1895 panic("segvn_unmap");
1896 /*NOTREACHED*/
1897 }
1898
1899 if (seg->s_szc != 0) {
1900 size_t pgsz = page_get_pagesize(seg->s_szc);
1901 int err;
1902 if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
1903 ASSERT(seg->s_base != addr || seg->s_size != len);
1904 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1905 ASSERT(svd->amp == NULL);
1906 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1907 hat_leave_region(seg->s_as->a_hat,
1908 svd->rcookie, HAT_REGION_TEXT);
1909 svd->rcookie = HAT_INVALID_REGION_COOKIE;
1910 /*
1911 * could pass a flag to segvn_demote_range()
1912 * below to tell it not to do any unloads but
1913 * this case is rare enough to not bother for
1914 * now.
1915 */
1916 } else if (svd->tr_state == SEGVN_TR_INIT) {
1917 svd->tr_state = SEGVN_TR_OFF;
1918 } else if (svd->tr_state == SEGVN_TR_ON) {
1919 ASSERT(svd->amp != NULL);
1920 segvn_textunrepl(seg, 1);
1921 ASSERT(svd->amp == NULL);
1922 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1923 }
1924 VM_STAT_ADD(segvnvmstats.demoterange[0]);
1925 err = segvn_demote_range(seg, addr, len, SDR_END, 0);
1926 if (err == 0) {
1927 return (IE_RETRY);
1928 }
1929 return (err);
1930 }
1931 }
1932
1933 /* Inform the vnode of the unmapping. */
1934 if (svd->vp) {
1935 int error;
1936
1937 error = fop_delmap(svd->vp,
1938 (offset_t)svd->offset + (uintptr_t)(addr - seg->s_base),
1939 seg->s_as, addr, len, svd->prot, svd->maxprot,
1940 svd->type, svd->cred, NULL);
1941
1942 if (error == EAGAIN)
1943 return (error);
1944 }
1945
1946 /*
1947 * Remove any page locks set through this mapping.
1948 * If text replication is not off no page locks could have been
1949 * established via this mapping.
1950 */
1951 if (svd->tr_state == SEGVN_TR_OFF) {
1952 (void) segvn_lockop(seg, addr, len, 0, MC_UNLOCK, NULL, 0);
1953 }
1954
1955 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1956 ASSERT(svd->amp == NULL);
1957 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1958 ASSERT(svd->type == MAP_PRIVATE);
1959 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
1960 HAT_REGION_TEXT);
1961 svd->rcookie = HAT_INVALID_REGION_COOKIE;
1962 } else if (svd->tr_state == SEGVN_TR_ON) {
1963 ASSERT(svd->amp != NULL);
1964 ASSERT(svd->pageprot == 0 && !(svd->prot & PROT_WRITE));
1965 segvn_textunrepl(seg, 1);
1966 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
1967 } else {
1968 if (svd->tr_state != SEGVN_TR_OFF) {
1969 ASSERT(svd->tr_state == SEGVN_TR_INIT);
1970 svd->tr_state = SEGVN_TR_OFF;
1971 }
1972 /*
1973 * Unload any hardware translations in the range to be taken
1974 * out. Use a callback to invoke free_vp_pages() effectively.
1975 */
1976 if (svd->vp != NULL && free_pages != 0) {
1977 callback.hcb_data = seg;
1978 callback.hcb_function = segvn_hat_unload_callback;
1979 cbp = &callback;
1980 }
1981 hat_unload_callback(seg->s_as->a_hat, addr, len,
1982 HAT_UNLOAD_UNMAP, cbp);
1983
1984 if (svd->type == MAP_SHARED && svd->vp != NULL &&
1985 (svd->vp->v_flag & VVMEXEC) &&
1986 ((svd->prot & PROT_WRITE) || svd->pageprot)) {
1987 segvn_inval_trcache(svd->vp);
1988 }
1989 }
1990
1991 /*
1992 * Check for entire segment
1993 */
1994 if (addr == seg->s_base && len == seg->s_size) {
1995 seg_free(seg);
1996 return (0);
1997 }
1998
1999 opages = seg_pages(seg);
2000 dpages = btop(len);
2001 npages = opages - dpages;
2002 amp = svd->amp;
2003 ASSERT(amp == NULL || amp->a_szc >= seg->s_szc);
2004
2005 /*
2006 * Check for beginning of segment
2007 */
2008 if (addr == seg->s_base) {
2009 if (svd->vpage != NULL) {
2010 size_t nbytes;
2011 struct vpage *ovpage;
2012
2013 ovpage = svd->vpage; /* keep pointer to vpage */
2014
2015 nbytes = vpgtob(npages);
2016 svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2017 bcopy(&ovpage[dpages], svd->vpage, nbytes);
2018
2019 /* free up old vpage */
2020 kmem_free(ovpage, vpgtob(opages));
2021 }
2022 if (amp != NULL) {
2023 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2024 if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2025 /*
2026 * Shared anon map is no longer in use. Before
2027 * freeing its pages purge all entries from
2028 * pcache that belong to this amp.
2029 */
2030 if (svd->type == MAP_SHARED) {
2031 ASSERT(amp->refcnt == 1);
2032 ASSERT(svd->softlockcnt == 0);
2033 anonmap_purge(amp);
2034 }
2035 /*
2036 * Free up now unused parts of anon_map array.
2037 */
2038 if (amp->a_szc == seg->s_szc) {
2039 if (seg->s_szc != 0) {
2040 anon_free_pages(amp->ahp,
2041 svd->anon_index, len,
2042 seg->s_szc);
2043 } else {
2044 anon_free(amp->ahp,
2045 svd->anon_index,
2046 len);
2047 }
2048 } else {
2049 ASSERT(svd->type == MAP_SHARED);
2050 ASSERT(amp->a_szc > seg->s_szc);
2051 anon_shmap_free_pages(amp,
2052 svd->anon_index, len);
2053 }
2054
2055 /*
2056 * Unreserve swap space for the
2057 * unmapped chunk of this segment in
2058 * case it's MAP_SHARED
2059 */
2060 if (svd->type == MAP_SHARED) {
2061 anon_unresv_zone(len,
2062 seg->s_as->a_proc->p_zone);
2063 amp->swresv -= len;
2064 }
2065 }
2066 ANON_LOCK_EXIT(&amp->a_rwlock);
2067 svd->anon_index += dpages;
2068 }
2069 if (svd->vp != NULL)
2070 svd->offset += len;
2071
2072 seg->s_base += len;
2073 seg->s_size -= len;
2074
2075 if (svd->swresv) {
2076 if (svd->flags & MAP_NORESERVE) {
2077 ASSERT(amp);
2078 oswresv = svd->swresv;
2079
2080 svd->swresv = ptob(anon_pages(amp->ahp,
2081 svd->anon_index, npages));
2082 anon_unresv_zone(oswresv - svd->swresv,
2083 seg->s_as->a_proc->p_zone);
2084 if (SEG_IS_PARTIAL_RESV(seg))
2085 seg->s_as->a_resvsize -= oswresv -
2086 svd->swresv;
2087 } else {
2088 size_t unlen;
2089
2090 if (svd->pageswap) {
2091 oswresv = svd->swresv;
2092 svd->swresv =
2093 segvn_count_swap_by_vpages(seg);
2094 ASSERT(oswresv >= svd->swresv);
2095 unlen = oswresv - svd->swresv;
2096 } else {
2097 svd->swresv -= len;
2098 ASSERT(svd->swresv == seg->s_size);
2099 unlen = len;
2100 }
2101 anon_unresv_zone(unlen,
2102 seg->s_as->a_proc->p_zone);
2103 }
2104 }
2105
2106 return (0);
2107 }
2108
2109 /*
2110 * Check for end of segment
2111 */
2112 if (addr + len == seg->s_base + seg->s_size) {
2113 if (svd->vpage != NULL) {
2114 size_t nbytes;
2115 struct vpage *ovpage;
2116
2117 ovpage = svd->vpage; /* keep pointer to vpage */
2118
2119 nbytes = vpgtob(npages);
2120 svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2121 bcopy(ovpage, svd->vpage, nbytes);
2122
2123 /* free up old vpage */
2124 kmem_free(ovpage, vpgtob(opages));
2125
2126 }
2127 if (amp != NULL) {
2128 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2129 if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2130 /*
2131 * Free up now unused parts of anon_map array.
2132 */
2133 ulong_t an_idx = svd->anon_index + npages;
2134
2135 /*
2136 * Shared anon map is no longer in use. Before
2137 * freeing its pages purge all entries from
2138 * pcache that belong to this amp.
2139 */
2140 if (svd->type == MAP_SHARED) {
2141 ASSERT(amp->refcnt == 1);
2142 ASSERT(svd->softlockcnt == 0);
2143 anonmap_purge(amp);
2144 }
2145
2146 if (amp->a_szc == seg->s_szc) {
2147 if (seg->s_szc != 0) {
2148 anon_free_pages(amp->ahp,
2149 an_idx, len,
2150 seg->s_szc);
2151 } else {
2152 anon_free(amp->ahp, an_idx,
2153 len);
2154 }
2155 } else {
2156 ASSERT(svd->type == MAP_SHARED);
2157 ASSERT(amp->a_szc > seg->s_szc);
2158 anon_shmap_free_pages(amp,
2159 an_idx, len);
2160 }
2161
2162 /*
2163 * Unreserve swap space for the
2164 * unmapped chunk of this segment in
2165 * case it's MAP_SHARED
2166 */
2167 if (svd->type == MAP_SHARED) {
2168 anon_unresv_zone(len,
2169 seg->s_as->a_proc->p_zone);
2170 amp->swresv -= len;
2171 }
2172 }
2173 ANON_LOCK_EXIT(&amp->a_rwlock);
2174 }
2175
2176 seg->s_size -= len;
2177
2178 if (svd->swresv) {
2179 if (svd->flags & MAP_NORESERVE) {
2180 ASSERT(amp);
2181 oswresv = svd->swresv;
2182 svd->swresv = ptob(anon_pages(amp->ahp,
2183 svd->anon_index, npages));
2184 anon_unresv_zone(oswresv - svd->swresv,
2185 seg->s_as->a_proc->p_zone);
2186 if (SEG_IS_PARTIAL_RESV(seg))
2187 seg->s_as->a_resvsize -= oswresv -
2188 svd->swresv;
2189 } else {
2190 size_t unlen;
2191
2192 if (svd->pageswap) {
2193 oswresv = svd->swresv;
2194 svd->swresv =
2195 segvn_count_swap_by_vpages(seg);
2196 ASSERT(oswresv >= svd->swresv);
2197 unlen = oswresv - svd->swresv;
2198 } else {
2199 svd->swresv -= len;
2200 ASSERT(svd->swresv == seg->s_size);
2201 unlen = len;
2202 }
2203 anon_unresv_zone(unlen,
2204 seg->s_as->a_proc->p_zone);
2205 }
2206 }
2207
2208 return (0);
2209 }
2210
2211 /*
2212 * The section to go is in the middle of the segment,
2213 * have to make it into two segments. nseg is made for
2214 * the high end while seg is cut down at the low end.
2215 */
2216 nbase = addr + len; /* new seg base */
2217 nsize = (seg->s_base + seg->s_size) - nbase; /* new seg size */
2218 seg->s_size = addr - seg->s_base; /* shrink old seg */
2219 nseg = seg_alloc(seg->s_as, nbase, nsize);
2220 if (nseg == NULL) {
2221 panic("segvn_unmap seg_alloc");
2222 /*NOTREACHED*/
2223 }
2224 nseg->s_ops = seg->s_ops;
2225 nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
2226 nseg->s_data = (void *)nsvd;
2227 nseg->s_szc = seg->s_szc;
2228 *nsvd = *svd;
2229 nsvd->seg = nseg;
2230 nsvd->offset = svd->offset + (uintptr_t)(nseg->s_base - seg->s_base);
2231 nsvd->swresv = 0;
2232 nsvd->softlockcnt = 0;
2233 nsvd->softlockcnt_sbase = 0;
2234 nsvd->softlockcnt_send = 0;
2235 nsvd->svn_inz = svd->svn_inz;
2236 ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
2237
2238 if (svd->vp != NULL) {
2239 VN_HOLD(nsvd->vp);
2240 if (nsvd->type == MAP_SHARED)
2241 lgrp_shm_policy_init(NULL, nsvd->vp);
2242 }
2243 crhold(svd->cred);
2244
2245 if (svd->vpage == NULL) {
2246 nsvd->vpage = NULL;
2247 } else {
2248 /* need to split vpage into two arrays */
2249 size_t nbytes;
2250 struct vpage *ovpage;
2251
2252 ovpage = svd->vpage; /* keep pointer to vpage */
2253
2254 npages = seg_pages(seg); /* seg has shrunk */
2255 nbytes = vpgtob(npages);
2256 svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2257
2258 bcopy(ovpage, svd->vpage, nbytes);
2259
2260 npages = seg_pages(nseg);
2261 nbytes = vpgtob(npages);
2262 nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2263
2264 bcopy(&ovpage[opages - npages], nsvd->vpage, nbytes);
2265
2266 /* free up old vpage */
2267 kmem_free(ovpage, vpgtob(opages));
2268 }
2269
2270 if (amp == NULL) {
2271 nsvd->amp = NULL;
2272 nsvd->anon_index = 0;
2273 } else {
2274 /*
2275 * Need to create a new anon map for the new segment.
2276 * We'll also allocate a new smaller array for the old
2277 * smaller segment to save space.
2278 */
2279 opages = btop((uintptr_t)(addr - seg->s_base));
2280 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2281 if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2282 /*
2283 * Free up now unused parts of anon_map array.
2284 */
2285 ulong_t an_idx = svd->anon_index + opages;
2286
2287 /*
2288 * Shared anon map is no longer in use. Before
2289 * freeing its pages purge all entries from
2290 * pcache that belong to this amp.
2291 */
2292 if (svd->type == MAP_SHARED) {
2293 ASSERT(amp->refcnt == 1);
2294 ASSERT(svd->softlockcnt == 0);
2295 anonmap_purge(amp);
2296 }
2297
2298 if (amp->a_szc == seg->s_szc) {
2299 if (seg->s_szc != 0) {
2300 anon_free_pages(amp->ahp, an_idx, len,
2301 seg->s_szc);
2302 } else {
2303 anon_free(amp->ahp, an_idx,
2304 len);
2305 }
2306 } else {
2307 ASSERT(svd->type == MAP_SHARED);
2308 ASSERT(amp->a_szc > seg->s_szc);
2309 anon_shmap_free_pages(amp, an_idx, len);
2310 }
2311
2312 /*
2313 * Unreserve swap space for the
2314 * unmapped chunk of this segment in
2315 * case it's MAP_SHARED
2316 */
2317 if (svd->type == MAP_SHARED) {
2318 anon_unresv_zone(len,
2319 seg->s_as->a_proc->p_zone);
2320 amp->swresv -= len;
2321 }
2322 }
2323 nsvd->anon_index = svd->anon_index +
2324 btop((uintptr_t)(nseg->s_base - seg->s_base));
2325 if (svd->type == MAP_SHARED) {
2326 amp->refcnt++;
2327 nsvd->amp = amp;
2328 } else {
2329 struct anon_map *namp;
2330 struct anon_hdr *nahp;
2331
2332 ASSERT(svd->type == MAP_PRIVATE);
2333 nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
2334 namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
2335 namp->a_szc = seg->s_szc;
2336 (void) anon_copy_ptr(amp->ahp, svd->anon_index, nahp,
2337 0, btop(seg->s_size), ANON_SLEEP);
2338 (void) anon_copy_ptr(amp->ahp, nsvd->anon_index,
2339 namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
2340 anon_release(amp->ahp, btop(amp->size));
2341 svd->anon_index = 0;
2342 nsvd->anon_index = 0;
2343 amp->ahp = nahp;
2344 amp->size = seg->s_size;
2345 nsvd->amp = namp;
2346 }
2347 ANON_LOCK_EXIT(&amp->a_rwlock);
2348 }
2349 if (svd->swresv) {
2350 if (svd->flags & MAP_NORESERVE) {
2351 ASSERT(amp);
2352 oswresv = svd->swresv;
2353 svd->swresv = ptob(anon_pages(amp->ahp,
2354 svd->anon_index, btop(seg->s_size)));
2355 nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
2356 nsvd->anon_index, btop(nseg->s_size)));
2357 ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2358 anon_unresv_zone(oswresv - (svd->swresv + nsvd->swresv),
2359 seg->s_as->a_proc->p_zone);
2360 if (SEG_IS_PARTIAL_RESV(seg))
2361 seg->s_as->a_resvsize -= oswresv -
2362 (svd->swresv + nsvd->swresv);
2363 } else {
2364 size_t unlen;
2365
2366 if (svd->pageswap) {
2367 oswresv = svd->swresv;
2368 svd->swresv = segvn_count_swap_by_vpages(seg);
2369 nsvd->swresv = segvn_count_swap_by_vpages(nseg);
2370 ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2371 unlen = oswresv - (svd->swresv + nsvd->swresv);
2372 } else {
2373 if (seg->s_size + nseg->s_size + len !=
2374 svd->swresv) {
2375 panic("segvn_unmap: cannot split "
2376 "swap reservation");
2377 /*NOTREACHED*/
2378 }
2379 svd->swresv = seg->s_size;
2380 nsvd->swresv = nseg->s_size;
2381 unlen = len;
2382 }
2383 anon_unresv_zone(unlen,
2384 seg->s_as->a_proc->p_zone);
2385 }
2386 }
2387
2388 return (0); /* I'm glad that's all over with! */
2389 }
2390
2391 static void
2392 segvn_free(struct seg *seg)
2393 {
2394 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2395 pgcnt_t npages = seg_pages(seg);
2396 struct anon_map *amp;
2397 size_t len;
2398
2399 /*
2400 * We don't need any segment level locks for "segvn" data
2401 * since the address space is "write" locked.
2402 */
2403 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
2404 ASSERT(svd->tr_state == SEGVN_TR_OFF);
2405
2406 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2407
2408 /*
2409 * Be sure to unlock pages. XXX Why do things get free'ed instead
2410 * of unmapped? XXX
2411 */
2412 (void) segvn_lockop(seg, seg->s_base, seg->s_size,
2413 0, MC_UNLOCK, NULL, 0);
2414
2415 /*
2416 * Deallocate the vpage and anon pointers if necessary and possible.
2417 */
2418 if (svd->vpage != NULL) {
2419 kmem_free(svd->vpage, vpgtob(npages));
2420 svd->vpage = NULL;
2421 }
2422 if ((amp = svd->amp) != NULL) {
2423 /*
2424 * If there are no more references to this anon_map
2425 * structure, then deallocate the structure after freeing
2426 * up all the anon slot pointers that we can.
2427 */
2428 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2429 ASSERT(amp->a_szc >= seg->s_szc);
2430 if (--amp->refcnt == 0) {
2431 if (svd->type == MAP_PRIVATE) {
2432 /*
2433 * Private - we only need to anon_free
2434 * the part that this segment refers to.
2435 */
2436 if (seg->s_szc != 0) {
2437 anon_free_pages(amp->ahp,
2438 svd->anon_index, seg->s_size,
2439 seg->s_szc);
2440 } else {
2441 anon_free(amp->ahp, svd->anon_index,
2442 seg->s_size);
2443 }
2444 } else {
2445
2446 /*
2447 * Shared anon map is no longer in use. Before
2448 * freeing its pages purge all entries from
2449 * pcache that belong to this amp.
2450 */
2451 ASSERT(svd->softlockcnt == 0);
2452 anonmap_purge(amp);
2453
2454 /*
2455 * Shared - anon_free the entire
2456 * anon_map's worth of stuff and
2457 * release any swap reservation.
2458 */
2459 if (amp->a_szc != 0) {
2460 anon_shmap_free_pages(amp, 0,
2461 amp->size);
2462 } else {
2463 anon_free(amp->ahp, 0, amp->size);
2464 }
2465 if ((len = amp->swresv) != 0) {
2466 anon_unresv_zone(len,
2467 seg->s_as->a_proc->p_zone);
2468 }
2469 }
2470 svd->amp = NULL;
2471 ANON_LOCK_EXIT(&amp->a_rwlock);
2472 anonmap_free(amp);
2473 } else if (svd->type == MAP_PRIVATE) {
2474 /*
2475 * We had a private mapping which still has
2476 * a held anon_map so just free up all the
2477 * anon slot pointers that we were using.
2478 */
2479 if (seg->s_szc != 0) {
2480 anon_free_pages(amp->ahp, svd->anon_index,
2481 seg->s_size, seg->s_szc);
2482 } else {
2483 anon_free(amp->ahp, svd->anon_index,
2484 seg->s_size);
2485 }
2486 ANON_LOCK_EXIT(&amp->a_rwlock);
2487 } else {
2488 ANON_LOCK_EXIT(&amp->a_rwlock);
2489 }
2490 }
2491
2492 /*
2493 * Release swap reservation.
2494 */
2495 if ((len = svd->swresv) != 0) {
2496 anon_unresv_zone(svd->swresv,
2497 seg->s_as->a_proc->p_zone);
2498 if (SEG_IS_PARTIAL_RESV(seg))
2499 seg->s_as->a_resvsize -= svd->swresv;
2500 svd->swresv = 0;
2501 }
2502 /*
2503 * Release claim on vnode, credentials, and finally free the
2504 * private data.
2505 */
2506 if (svd->vp != NULL) {
2507 if (svd->type == MAP_SHARED)
2508 lgrp_shm_policy_fini(NULL, svd->vp);
2509 VN_RELE(svd->vp);
2510 svd->vp = NULL;
2511 }
2512 crfree(svd->cred);
2513 svd->pageprot = 0;
2514 svd->pageadvice = 0;
2515 svd->pageswap = 0;
2516 svd->cred = NULL;
2517
2518 /*
2519 * Take segfree_syncmtx lock to let segvn_reclaim() finish if it's
2520 * still working with this segment without holding as lock (in case
2521 * it's called by pcache async thread).
2522 */
2523 ASSERT(svd->softlockcnt == 0);
2524 mutex_enter(&svd->segfree_syncmtx);
2525 mutex_exit(&svd->segfree_syncmtx);
2526
2527 seg->s_data = NULL;
2528 kmem_cache_free(segvn_cache, svd);
2529 }
2530
2531 /*
2532 * Do a F_SOFTUNLOCK call over the range requested. The range must have
2533 * already been F_SOFTLOCK'ed.
2534 * Caller must always match addr and len of a softunlock with a previous
2535 * softlock with exactly the same addr and len.
2536 */
2537 static void
2538 segvn_softunlock(struct seg *seg, caddr_t addr, size_t len, enum seg_rw rw)
2539 {
2540 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2541 page_t *pp;
2542 caddr_t adr;
2543 struct vnode *vp;
2544 uoff_t offset;
2545 ulong_t anon_index;
2546 struct anon_map *amp;
2547 struct anon *ap = NULL;
2548
2549 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2550 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
2551
2552 if ((amp = svd->amp) != NULL)
2553 anon_index = svd->anon_index + seg_page(seg, addr);
2554
2555 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
2556 ASSERT(svd->tr_state == SEGVN_TR_OFF);
2557 hat_unlock_region(seg->s_as->a_hat, addr, len, svd->rcookie);
2558 } else {
2559 hat_unlock(seg->s_as->a_hat, addr, len);
2560 }
2561 for (adr = addr; adr < addr + len; adr += PAGESIZE) {
2562 if (amp != NULL) {
2563 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2564 if ((ap = anon_get_ptr(amp->ahp, anon_index++))
2565 != NULL) {
2566 swap_xlate(ap, &vp, &offset);
2567 } else {
2568 vp = svd->vp;
2569 offset = svd->offset +
2570 (uintptr_t)(adr - seg->s_base);
2571 }
2572 ANON_LOCK_EXIT(&amp->a_rwlock);
2573 } else {
2574 vp = svd->vp;
2575 offset = svd->offset +
2576 (uintptr_t)(adr - seg->s_base);
2577 }
2578
2579 /*
2580 * Use page_find() instead of page_lookup() to
2581 * find the page since we know that it is locked.
2582 */
2583 pp = page_find(&vp->v_object, offset);
2584 if (pp == NULL) {
2585 panic(
2586 "segvn_softunlock: addr %p, ap %p, vp %p, off %llx",
2587 (void *)adr, (void *)ap, (void *)vp, offset);
2588 /*NOTREACHED*/
2589 }
2590
2591 if (rw == S_WRITE) {
2592 hat_setrefmod(pp);
2593 if (seg->s_as->a_vbits)
2594 hat_setstat(seg->s_as, adr, PAGESIZE,
2595 P_REF | P_MOD);
2596 } else if (rw != S_OTHER) {
2597 hat_setref(pp);
2598 if (seg->s_as->a_vbits)
2599 hat_setstat(seg->s_as, adr, PAGESIZE, P_REF);
2600 }
2601 page_unlock(pp);
2602 }
2603 ASSERT(svd->softlockcnt >= btop(len));
2604 if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -btop(len))) {
2605 /*
2606 * All SOFTLOCKS are gone. Wakeup any waiting
2607 * unmappers so they can try again to unmap.
2608 * Check for waiters first without the mutex
2609 * held so we don't always grab the mutex on
2610 * softunlocks.
2611 */
2612 if (AS_ISUNMAPWAIT(seg->s_as)) {
2613 mutex_enter(&seg->s_as->a_contents);
2614 if (AS_ISUNMAPWAIT(seg->s_as)) {
2615 AS_CLRUNMAPWAIT(seg->s_as);
2616 cv_broadcast(&seg->s_as->a_cv);
2617 }
2618 mutex_exit(&seg->s_as->a_contents);
2619 }
2620 }
2621 }
2622
2623 #define PAGE_HANDLED ((page_t *)-1)
2624
2625 /*
2626 * Release all the pages in the NULL terminated ppp list
2627 * which haven't already been converted to PAGE_HANDLED.
2628 */
2629 static void
2630 segvn_pagelist_rele(page_t **ppp)
2631 {
2632 for (; *ppp != NULL; ppp++) {
2633 if (*ppp != PAGE_HANDLED)
2634 page_unlock(*ppp);
2635 }
2636 }
2637
2638 static int stealcow = 1;
2639
2640 /*
2641 * Workaround for viking chip bug. See bug id 1220902.
2642 * To fix this down in pagefault() would require importing so
2643 * much as and segvn code as to be unmaintainable.
2644 */
2645 int enable_mbit_wa = 0;
2646
2647 /*
2648 * Handles all the dirty work of getting the right
2649 * anonymous pages and loading up the translations.
2650 * This routine is called only from segvn_fault()
2651 * when looping over the range of addresses requested.
2652 *
2653 * The basic algorithm here is:
2654 * If this is an anon_zero case
2655 * Call anon_zero to allocate page
2656 * Load up translation
2657 * Return
2658 * endif
2659 * If this is an anon page
2660 * Use anon_getpage to get the page
2661 * else
2662 * Find page in pl[] list passed in
2663 * endif
2664 * If not a cow
2665 * Load up the translation to the page
2666 * return
2667 * endif
2668 * Call anon_private to handle cow
2669 * Load up (writable) translation to new page
2670 */
2671 static faultcode_t
2672 segvn_faultpage(
2673 struct hat *hat, /* the hat to use for mapping */
2674 struct seg *seg, /* seg_vn of interest */
2675 caddr_t addr, /* address in as */
2676 uoff_t off, /* offset in vp */
2677 struct vpage *vpage, /* pointer to vpage for vp, off */
2678 page_t *pl[], /* object source page pointer */
2679 uint_t vpprot, /* access allowed to object pages */
2680 enum fault_type type, /* type of fault */
2681 enum seg_rw rw, /* type of access at fault */
2682 int brkcow) /* we may need to break cow */
2683 {
2684 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2685 page_t *pp, **ppp;
2686 uint_t pageflags = 0;
2687 page_t *anon_pl[1 + 1];
2688 page_t *opp = NULL; /* original page */
2689 uint_t prot;
2690 int err;
2691 int cow;
2692 int claim;
2693 int steal = 0;
2694 ulong_t anon_index;
2695 struct anon *ap, *oldap;
2696 struct anon_map *amp;
2697 int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
2698 int anon_lock = 0;
2699 anon_sync_obj_t cookie;
2700
2701 if (svd->flags & MAP_TEXT) {
2702 hat_flag |= HAT_LOAD_TEXT;
2703 }
2704
2705 ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
2706 ASSERT(seg->s_szc == 0);
2707 ASSERT(svd->tr_state != SEGVN_TR_INIT);
2708
2709 /*
2710 * Initialize protection value for this page.
2711 * If we have per page protection values check it now.
2712 */
2713 if (svd->pageprot) {
2714 uint_t protchk;
2715
2716 switch (rw) {
2717 case S_READ:
2718 protchk = PROT_READ;
2719 break;
2720 case S_WRITE:
2721 protchk = PROT_WRITE;
2722 break;
2723 case S_EXEC:
2724 protchk = PROT_EXEC;
2725 break;
2726 case S_OTHER:
2727 default:
2728 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
2729 break;
2730 }
2731
2732 prot = VPP_PROT(vpage);
2733 if ((prot & protchk) == 0)
2734 return (FC_PROT); /* illegal access type */
2735 } else {
2736 prot = svd->prot;
2737 }
2738
2739 if (type == F_SOFTLOCK) {
2740 atomic_inc_ulong((ulong_t *)&svd->softlockcnt);
2741 }
2742
2743 /*
2744 * Always acquire the anon array lock to prevent 2 threads from
2745 * allocating separate anon slots for the same "addr".
2746 */
2747
2748 if ((amp = svd->amp) != NULL) {
2749 ASSERT(RW_READ_HELD(&amp->a_rwlock));
2750 anon_index = svd->anon_index + seg_page(seg, addr);
2751 anon_array_enter(amp, anon_index, &cookie);
2752 anon_lock = 1;
2753 }
2754
2755 if (svd->vp == NULL && amp != NULL) {
2756 if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL) {
2757 /*
2758 * Allocate a (normally) writable anonymous page of
2759 * zeroes. If no advance reservations, reserve now.
2760 */
2761 if (svd->flags & MAP_NORESERVE) {
2762 if (anon_resv_zone(ptob(1),
2763 seg->s_as->a_proc->p_zone)) {
2764 atomic_add_long(&svd->swresv, ptob(1));
2765 atomic_add_long(&seg->s_as->a_resvsize,
2766 ptob(1));
2767 } else {
2768 err = ENOMEM;
2769 goto out;
2770 }
2771 }
2772 if ((pp = anon_zero(seg, addr, &ap,
2773 svd->cred)) == NULL) {
2774 err = ENOMEM;
2775 goto out; /* out of swap space */
2776 }
2777 /*
2778 * Re-acquire the anon_map lock and
2779 * initialize the anon array entry.
2780 */
2781 (void) anon_set_ptr(amp->ahp, anon_index, ap,
2782 ANON_SLEEP);
2783
2784 ASSERT(pp->p_szc == 0);
2785
2786 /*
2787 * Handle pages that have been marked for migration
2788 */
2789 if (lgrp_optimizations())
2790 page_migrate(seg, addr, &pp, 1);
2791
2792 if (enable_mbit_wa) {
2793 if (rw == S_WRITE)
2794 hat_setmod(pp);
2795 else if (!hat_ismod(pp))
2796 prot &= ~PROT_WRITE;
2797 }
2798 /*
2799 * If AS_PAGLCK is set in a_flags (via memcntl(2)
2800 * with MC_LOCKAS, MCL_FUTURE) and this is a
2801 * MAP_NORESERVE segment, we may need to
2802 * permanently lock the page as it is being faulted
2803 * for the first time. The following text applies
2804 * only to MAP_NORESERVE segments:
2805 *
2806 * As per memcntl(2), if this segment was created
2807 * after MCL_FUTURE was applied (a "future"
2808 * segment), its pages must be locked. If this
2809 * segment existed at MCL_FUTURE application (a
2810 * "past" segment), the interface is unclear.
2811 *
2812 * We decide to lock only if vpage is present:
2813 *
2814 * - "future" segments will have a vpage array (see
2815 * as_map), and so will be locked as required
2816 *
2817 * - "past" segments may not have a vpage array,
2818 * depending on whether events (such as
2819 * mprotect) have occurred. Locking if vpage
2820 * exists will preserve legacy behavior. Not
2821 * locking if vpage is absent, will not break
2822 * the interface or legacy behavior. Note that
2823 * allocating vpage here if it's absent requires
2824 * upgrading the segvn reader lock, the cost of
2825 * which does not seem worthwhile.
2826 *
2827 * Usually testing and setting VPP_ISPPLOCK and
2828 * VPP_SETPPLOCK requires holding the segvn lock as
2829 * writer, but in this case all readers are
2830 * serializing on the anon array lock.
2831 */
2832 if (AS_ISPGLCK(seg->s_as) && vpage != NULL &&
2833 (svd->flags & MAP_NORESERVE) &&
2834 !VPP_ISPPLOCK(vpage)) {
2835 proc_t *p = seg->s_as->a_proc;
2836 ASSERT(svd->type == MAP_PRIVATE);
2837 mutex_enter(&p->p_lock);
2838 if (rctl_incr_locked_mem(p, NULL, PAGESIZE,
2839 1) == 0) {
2840 claim = VPP_PROT(vpage) & PROT_WRITE;
2841 if (page_pp_lock(pp, claim, 0)) {
2842 VPP_SETPPLOCK(vpage);
2843 } else {
2844 rctl_decr_locked_mem(p, NULL,
2845 PAGESIZE, 1);
2846 }
2847 }
2848 mutex_exit(&p->p_lock);
2849 }
2850
2851 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2852 hat_memload(hat, addr, pp, prot, hat_flag);
2853
2854 if (!(hat_flag & HAT_LOAD_LOCK))
2855 page_unlock(pp);
2856
2857 anon_array_exit(&cookie);
2858 return (0);
2859 }
2860 }
2861
2862 /*
2863 * Obtain the page structure via anon_getpage() if it is
2864 * a private copy of an object (the result of a previous
2865 * copy-on-write).
2866 */
2867 if (amp != NULL) {
2868 if ((ap = anon_get_ptr(amp->ahp, anon_index)) != NULL) {
2869 err = anon_getpage(&ap, &vpprot, anon_pl, PAGESIZE,
2870 seg, addr, rw, svd->cred);
2871 if (err)
2872 goto out;
2873
2874 if (svd->type == MAP_SHARED) {
2875 /*
2876 * If this is a shared mapping to an
2877 * anon_map, then ignore the write
2878 * permissions returned by anon_getpage().
2879 * They apply to the private mappings
2880 * of this anon_map.
2881 */
2882 vpprot |= PROT_WRITE;
2883 }
2884 opp = anon_pl[0];
2885 }
2886 }
2887
2888 /*
2889 * Search the pl[] list passed in if it is from the
2890 * original object (i.e., not a private copy).
2891 */
2892 if (opp == NULL) {
2893 /*
2894 * Find original page. We must be bringing it in
2895 * from the list in pl[].
2896 */
2897 for (ppp = pl; (opp = *ppp) != NULL; ppp++) {
2898 if (opp == PAGE_HANDLED)
2899 continue;
2900 VERIFY(opp->p_object == &svd->vp->v_object); /* XXX */
2901 ASSERT(opp->p_vnode == svd->vp); /* XXX */
2902 if (opp->p_offset == off)
2903 break;
2904 }
2905 if (opp == NULL) {
2906 panic("segvn_faultpage not found");
2907 /*NOTREACHED*/
2908 }
2909 *ppp = PAGE_HANDLED;
2910
2911 }
2912
2913 ASSERT(PAGE_LOCKED(opp));
2914
2915 /*
2916 * The fault is treated as a copy-on-write fault if a
2917 * write occurs on a private segment and the object
2918 * page (i.e., mapping) is write protected. We assume
2919 * that fatal protection checks have already been made.
2920 */
2921
2922 if (brkcow) {
2923 ASSERT(svd->tr_state == SEGVN_TR_OFF);
2924 cow = !(vpprot & PROT_WRITE);
2925 } else if (svd->tr_state == SEGVN_TR_ON) {
2926 /*
2927 * If we are doing text replication COW on first touch.
2928 */
2929 ASSERT(amp != NULL);
2930 ASSERT(svd->vp != NULL);
2931 ASSERT(rw != S_WRITE);
2932 cow = (ap == NULL);
2933 } else {
2934 cow = 0;
2935 }
2936
2937 /*
2938 * If not a copy-on-write case load the translation
2939 * and return.
2940 */
2941 if (cow == 0) {
2942
2943 /*
2944 * Handle pages that have been marked for migration
2945 */
2946 if (lgrp_optimizations())
2947 page_migrate(seg, addr, &opp, 1);
2948
2949 if (IS_VMODSORT(opp->p_vnode) || enable_mbit_wa) {
2950 if (rw == S_WRITE)
2951 hat_setmod(opp);
2952 else if (rw != S_OTHER && !hat_ismod(opp))
2953 prot &= ~PROT_WRITE;
2954 }
2955
2956 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
2957 (!svd->pageprot && svd->prot == (prot & vpprot)));
2958 ASSERT(amp == NULL ||
2959 svd->rcookie == HAT_INVALID_REGION_COOKIE);
2960 hat_memload_region(hat, addr, opp, prot & vpprot, hat_flag,
2961 svd->rcookie);
2962
2963 if (!(hat_flag & HAT_LOAD_LOCK))
2964 page_unlock(opp);
2965
2966 if (anon_lock) {
2967 anon_array_exit(&cookie);
2968 }
2969 return (0);
2970 }
2971
2972 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2973
2974 hat_setref(opp);
2975
2976 ASSERT(amp != NULL && anon_lock);
2977
2978 /*
2979 * Steal the page only if it isn't a private page
2980 * since stealing a private page is not worth the effort.
2981 */
2982 if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL)
2983 steal = 1;
2984
2985 /*
2986 * Steal the original page if the following conditions are true:
2987 *
2988 * We are low on memory, the page is not private, page is not large,
2989 * not shared, not modified, not `locked' or if we have it `locked'
2990 * (i.e., p_cowcnt == 1 and p_lckcnt == 0, which also implies
2991 * that the page is not shared) and if it doesn't have any
2992 * translations. page_struct_lock isn't needed to look at p_cowcnt
2993 * and p_lckcnt because we first get exclusive lock on page.
2994 */
2995 (void) hat_pagesync(opp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD);
2996
2997 if (stealcow && freemem < minfree && steal && opp->p_szc == 0 &&
2998 page_tryupgrade(opp) && !hat_ismod(opp) &&
2999 ((opp->p_lckcnt == 0 && opp->p_cowcnt == 0) ||
3000 (opp->p_lckcnt == 0 && opp->p_cowcnt == 1 &&
3001 vpage != NULL && VPP_ISPPLOCK(vpage)))) {
3002 /*
3003 * Check if this page has other translations
3004 * after unloading our translation.
3005 */
3006 if (hat_page_is_mapped(opp)) {
3007 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3008 hat_unload(seg->s_as->a_hat, addr, PAGESIZE,
3009 HAT_UNLOAD);
3010 }
3011
3012 /*
3013 * hat_unload() might sync back someone else's recent
3014 * modification, so check again.
3015 */
3016 if (!hat_ismod(opp) && !hat_page_is_mapped(opp))
3017 pageflags |= STEAL_PAGE;
3018 }
3019
3020 /*
3021 * If we have a vpage pointer, see if it indicates that we have
3022 * ``locked'' the page we map -- if so, tell anon_private to
3023 * transfer the locking resource to the new page.
3024 *
3025 * See Statement at the beginning of segvn_lockop regarding
3026 * the way lockcnts/cowcnts are handled during COW.
3027 *
3028 */
3029 if (vpage != NULL && VPP_ISPPLOCK(vpage))
3030 pageflags |= LOCK_PAGE;
3031
3032 /*
3033 * Allocate a private page and perform the copy.
3034 * For MAP_NORESERVE reserve swap space now, unless this
3035 * is a cow fault on an existing anon page in which case
3036 * MAP_NORESERVE will have made advance reservations.
3037 */
3038 if ((svd->flags & MAP_NORESERVE) && (ap == NULL)) {
3039 if (anon_resv_zone(ptob(1), seg->s_as->a_proc->p_zone)) {
3040 atomic_add_long(&svd->swresv, ptob(1));
3041 atomic_add_long(&seg->s_as->a_resvsize, ptob(1));
3042 } else {
3043 page_unlock(opp);
3044 err = ENOMEM;
3045 goto out;
3046 }
3047 }
3048 oldap = ap;
3049 pp = anon_private(&ap, seg, addr, prot, opp, pageflags, svd->cred);
3050 if (pp == NULL) {
3051 err = ENOMEM; /* out of swap space */
3052 goto out;
3053 }
3054
3055 /*
3056 * If we copied away from an anonymous page, then
3057 * we are one step closer to freeing up an anon slot.
3058 *
3059 * NOTE: The original anon slot must be released while
3060 * holding the "anon_map" lock. This is necessary to prevent
3061 * other threads from obtaining a pointer to the anon slot
3062 * which may be freed if its "refcnt" is 1.
3063 */
3064 if (oldap != NULL)
3065 anon_decref(oldap);
3066
3067 (void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
3068
3069 /*
3070 * Handle pages that have been marked for migration
3071 */
3072 if (lgrp_optimizations())
3073 page_migrate(seg, addr, &pp, 1);
3074
3075 ASSERT(pp->p_szc == 0);
3076
3077 ASSERT(!IS_VMODSORT(pp->p_vnode));
3078 if (enable_mbit_wa) {
3079 if (rw == S_WRITE)
3080 hat_setmod(pp);
3081 else if (!hat_ismod(pp))
3082 prot &= ~PROT_WRITE;
3083 }
3084
3085 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3086 hat_memload(hat, addr, pp, prot, hat_flag);
3087
3088 if (!(hat_flag & HAT_LOAD_LOCK))
3089 page_unlock(pp);
3090
3091 ASSERT(anon_lock);
3092 anon_array_exit(&cookie);
3093 return (0);
3094 out:
3095 if (anon_lock)
3096 anon_array_exit(&cookie);
3097
3098 if (type == F_SOFTLOCK) {
3099 atomic_dec_ulong((ulong_t *)&svd->softlockcnt);
3100 }
3101 return (FC_MAKE_ERR(err));
3102 }
3103
3104 /*
3105 * relocate a bunch of smaller targ pages into one large repl page. all targ
3106 * pages must be complete pages smaller than replacement pages.
3107 * it's assumed that no page's szc can change since they are all PAGESIZE or
3108 * complete large pages locked SHARED.
3109 */
3110 static void
3111 segvn_relocate_pages(page_t **targ, page_t *replacement)
3112 {
3113 page_t *pp;
3114 pgcnt_t repl_npgs, curnpgs;
3115 pgcnt_t i;
3116 uint_t repl_szc = replacement->p_szc;
3117 page_t *first_repl = replacement;
3118 page_t *repl;
3119 spgcnt_t npgs;
3120
3121 VM_STAT_ADD(segvnvmstats.relocatepages[0]);
3122
3123 ASSERT(repl_szc != 0);
3124 npgs = repl_npgs = page_get_pagecnt(repl_szc);
3125
3126 i = 0;
3127 while (repl_npgs) {
3128 spgcnt_t nreloc;
3129 int err;
3130 ASSERT(replacement != NULL);
3131 pp = targ[i];
3132 ASSERT(pp->p_szc < repl_szc);
3133 ASSERT(PAGE_EXCL(pp));
3134 ASSERT(!PP_ISFREE(pp));
3135 curnpgs = page_get_pagecnt(pp->p_szc);
3136 if (curnpgs == 1) {
3137 VM_STAT_ADD(segvnvmstats.relocatepages[1]);
3138 repl = replacement;
3139 page_sub(&replacement, repl);
3140 ASSERT(PAGE_EXCL(repl));
3141 ASSERT(!PP_ISFREE(repl));
3142 ASSERT(repl->p_szc == repl_szc);
3143 } else {
3144 page_t *repl_savepp;
3145 int j;
3146 VM_STAT_ADD(segvnvmstats.relocatepages[2]);
3147 repl_savepp = replacement;
3148 for (j = 0; j < curnpgs; j++) {
3149 repl = replacement;
3150 page_sub(&replacement, repl);
3151 ASSERT(PAGE_EXCL(repl));
3152 ASSERT(!PP_ISFREE(repl));
3153 ASSERT(repl->p_szc == repl_szc);
3154 ASSERT(page_pptonum(targ[i + j]) ==
3155 page_pptonum(targ[i]) + j);
3156 }
3157 repl = repl_savepp;
3158 ASSERT(IS_P2ALIGNED(page_pptonum(repl), curnpgs));
3159 }
3160 err = page_relocate(&pp, &repl, 0, 1, &nreloc, NULL);
3161 if (err || nreloc != curnpgs) {
3162 panic("segvn_relocate_pages: "
3163 "page_relocate failed err=%d curnpgs=%ld "
3164 "nreloc=%ld", err, curnpgs, nreloc);
3165 }
3166 ASSERT(curnpgs <= repl_npgs);
3167 repl_npgs -= curnpgs;
3168 i += curnpgs;
3169 }
3170 ASSERT(replacement == NULL);
3171
3172 repl = first_repl;
3173 repl_npgs = npgs;
3174 for (i = 0; i < repl_npgs; i++) {
3175 ASSERT(PAGE_EXCL(repl));
3176 ASSERT(!PP_ISFREE(repl));
3177 targ[i] = repl;
3178 page_downgrade(targ[i]);
3179 repl++;
3180 }
3181 }
3182
3183 /*
3184 * Check if all pages in ppa array are complete smaller than szc pages and
3185 * their roots will still be aligned relative to their current size if the
3186 * entire ppa array is relocated into one szc page. If these conditions are
3187 * not met return 0.
3188 *
3189 * If all pages are properly aligned attempt to upgrade their locks
3190 * to exclusive mode. If it fails set *upgrdfail to 1 and return 0.
3191 * upgrdfail was set to 0 by caller.
3192 *
3193 * Return 1 if all pages are aligned and locked exclusively.
3194 *
3195 * If all pages in ppa array happen to be physically contiguous to make one
3196 * szc page and all exclusive locks are successfully obtained promote the page
3197 * size to szc and set *pszc to szc. Return 1 with pages locked shared.
3198 */
3199 static int
3200 segvn_full_szcpages(page_t **ppa, uint_t szc, int *upgrdfail, uint_t *pszc)
3201 {
3202 page_t *pp;
3203 pfn_t pfn;
3204 pgcnt_t totnpgs = page_get_pagecnt(szc);
3205 pfn_t first_pfn;
3206 int contig = 1;
3207 pgcnt_t i;
3208 pgcnt_t j;
3209 uint_t curszc;
3210 pgcnt_t curnpgs;
3211 int root = 0;
3212
3213 ASSERT(szc > 0);
3214
3215 VM_STAT_ADD(segvnvmstats.fullszcpages[0]);
3216
3217 for (i = 0; i < totnpgs; i++) {
3218 pp = ppa[i];
3219 ASSERT(PAGE_SHARED(pp));
3220 ASSERT(!PP_ISFREE(pp));
3221 pfn = page_pptonum(pp);
3222 if (i == 0) {
3223 if (!IS_P2ALIGNED(pfn, totnpgs)) {
3224 contig = 0;
3225 } else {
3226 first_pfn = pfn;
3227 }
3228 } else if (contig && pfn != first_pfn + i) {
3229 contig = 0;
3230 }
3231 if (pp->p_szc == 0) {
3232 if (root) {
3233 VM_STAT_ADD(segvnvmstats.fullszcpages[1]);
3234 return (0);
3235 }
3236 } else if (!root) {
3237 if ((curszc = pp->p_szc) >= szc) {
3238 VM_STAT_ADD(segvnvmstats.fullszcpages[2]);
3239 return (0);
3240 }
3241 if (curszc == 0) {
3242 /*
3243 * p_szc changed means we don't have all pages
3244 * locked. return failure.
3245 */
3246 VM_STAT_ADD(segvnvmstats.fullszcpages[3]);
3247 return (0);
3248 }
3249 curnpgs = page_get_pagecnt(curszc);
3250 if (!IS_P2ALIGNED(pfn, curnpgs) ||
3251 !IS_P2ALIGNED(i, curnpgs)) {
3252 VM_STAT_ADD(segvnvmstats.fullszcpages[4]);
3253 return (0);
3254 }
3255 root = 1;
3256 } else {
3257 ASSERT(i > 0);
3258 VM_STAT_ADD(segvnvmstats.fullszcpages[5]);
3259 if (pp->p_szc != curszc) {
3260 VM_STAT_ADD(segvnvmstats.fullszcpages[6]);
3261 return (0);
3262 }
3263 if (pfn - 1 != page_pptonum(ppa[i - 1])) {
3264 panic("segvn_full_szcpages: "
3265 "large page not physically contiguous");
3266 }
3267 if (P2PHASE(pfn, curnpgs) == curnpgs - 1) {
3268 root = 0;
3269 }
3270 }
3271 }
3272
3273 for (i = 0; i < totnpgs; i++) {
3274 ASSERT(ppa[i]->p_szc < szc);
3275 if (!page_tryupgrade(ppa[i])) {
3276 for (j = 0; j < i; j++) {
3277 page_downgrade(ppa[j]);
3278 }
3279 *pszc = ppa[i]->p_szc;
3280 *upgrdfail = 1;
3281 VM_STAT_ADD(segvnvmstats.fullszcpages[7]);
3282 return (0);
3283 }
3284 }
3285
3286 /*
3287 * When a page is put a free cachelist its szc is set to 0. if file
3288 * system reclaimed pages from cachelist targ pages will be physically
3289 * contiguous with 0 p_szc. in this case just upgrade szc of targ
3290 * pages without any relocations.
3291 * To avoid any hat issues with previous small mappings
3292 * hat_pageunload() the target pages first.
3293 */
3294 if (contig) {
3295 VM_STAT_ADD(segvnvmstats.fullszcpages[8]);
3296 for (i = 0; i < totnpgs; i++) {
3297 (void) hat_pageunload(ppa[i], HAT_FORCE_PGUNLOAD);
3298 }
3299 for (i = 0; i < totnpgs; i++) {
3300 ppa[i]->p_szc = szc;
3301 }
3302 for (i = 0; i < totnpgs; i++) {
3303 ASSERT(PAGE_EXCL(ppa[i]));
3304 page_downgrade(ppa[i]);
3305 }
3306 if (pszc != NULL) {
3307 *pszc = szc;
3308 }
3309 }
3310 VM_STAT_ADD(segvnvmstats.fullszcpages[9]);
3311 return (1);
3312 }
3313
3314 /*
3315 * Create physically contiguous pages for [vp, off] - [vp, off +
3316 * page_size(szc)) range and for private segment return them in ppa array.
3317 * Pages are created either via IO or relocations.
3318 *
3319 * Return 1 on success and 0 on failure.
3320 *
3321 * If physically contiguous pages already exist for this range return 1 without
3322 * filling ppa array. Caller initializes ppa[0] as NULL to detect that ppa
3323 * array wasn't filled. In this case caller fills ppa array via fop_getpage().
3324 */
3325
3326 static int
3327 segvn_fill_vp_pages(struct segvn_data *svd, vnode_t *vp, uoff_t off,
3328 uint_t szc, page_t **ppa, page_t **ppplist, uint_t *ret_pszc,
3329 int *downsize)
3330 {
3331 page_t *pplist = *ppplist;
3332 size_t pgsz = page_get_pagesize(szc);
3333 pgcnt_t pages = btop(pgsz);
3334 ulong_t start_off = off;
3335 uoff_t eoff = off + pgsz;
3336 spgcnt_t nreloc;
3337 uoff_t io_off = off;
3338 size_t io_len;
3339 page_t *io_pplist = NULL;
3340 page_t *done_pplist = NULL;
3341 pgcnt_t pgidx = 0;
3342 page_t *pp;
3343 page_t *newpp;
3344 page_t *targpp;
3345 int io_err = 0;
3346 int i;
3347 pfn_t pfn;
3348 ulong_t ppages;
3349 page_t *targ_pplist = NULL;
3350 page_t *repl_pplist = NULL;
3351 page_t *tmp_pplist;
3352 int nios = 0;
3353 uint_t pszc;
3354 struct vattr va;
3355
3356 VM_STAT_ADD(segvnvmstats.fill_vp_pages[0]);
3357
3358 ASSERT(szc != 0);
3359 ASSERT(pplist->p_szc == szc);
3360
3361 /*
3362 * downsize will be set to 1 only if we fail to lock pages. this will
3363 * allow subsequent faults to try to relocate the page again. If we
3364 * fail due to misalignment don't downsize and let the caller map the
3365 * whole region with small mappings to avoid more faults into the area
3366 * where we can't get large pages anyway.
3367 */
3368 *downsize = 0;
3369
3370 while (off < eoff) {
3371 newpp = pplist;
3372 ASSERT(newpp != NULL);
3373 ASSERT(PAGE_EXCL(newpp));
3374 ASSERT(!PP_ISFREE(newpp));
3375 /*
3376 * we pass NULL for nrelocp to page_lookup_create()
3377 * so that it doesn't relocate. We relocate here
3378 * later only after we make sure we can lock all
3379 * pages in the range we handle and they are all
3380 * aligned.
3381 */
3382 pp = page_lookup_create(&vp->v_object, off, SE_SHARED, newpp,
3383 NULL, 0);
3384 ASSERT(pp != NULL);
3385 ASSERT(!PP_ISFREE(pp));
3386 VERIFY(pp->p_object == &vp->v_object);
3387 ASSERT(pp->p_vnode == vp);
3388 ASSERT(pp->p_offset == off);
3389 if (pp == newpp) {
3390 VM_STAT_ADD(segvnvmstats.fill_vp_pages[1]);
3391 page_sub(&pplist, pp);
3392 ASSERT(PAGE_EXCL(pp));
3393 ASSERT(page_iolock_assert(pp));
3394 page_list_concat(&io_pplist, &pp);
3395 off += PAGESIZE;
3396 continue;
3397 }
3398 VM_STAT_ADD(segvnvmstats.fill_vp_pages[2]);
3399 pfn = page_pptonum(pp);
3400 pszc = pp->p_szc;
3401 if (pszc >= szc && targ_pplist == NULL && io_pplist == NULL &&
3402 IS_P2ALIGNED(pfn, pages)) {
3403 ASSERT(repl_pplist == NULL);
3404 ASSERT(done_pplist == NULL);
3405 ASSERT(pplist == *ppplist);
3406 page_unlock(pp);
3407 page_free_replacement_page(pplist);
3408 page_create_putback(pages);
3409 *ppplist = NULL;
3410 VM_STAT_ADD(segvnvmstats.fill_vp_pages[3]);
3411 return (1);
3412 }
3413 if (pszc >= szc) {
3414 page_unlock(pp);
3415 segvn_faultvnmpss_align_err1++;
3416 goto out;
3417 }
3418 ppages = page_get_pagecnt(pszc);
3419 if (!IS_P2ALIGNED(pfn, ppages)) {
3420 ASSERT(pszc > 0);
3421 /*
3422 * sizing down to pszc won't help.
3423 */
3424 page_unlock(pp);
3425 segvn_faultvnmpss_align_err2++;
3426 goto out;
3427 }
3428 pfn = page_pptonum(newpp);
3429 if (!IS_P2ALIGNED(pfn, ppages)) {
3430 ASSERT(pszc > 0);
3431 /*
3432 * sizing down to pszc won't help.
3433 */
3434 page_unlock(pp);
3435 segvn_faultvnmpss_align_err3++;
3436 goto out;
3437 }
3438 if (!PAGE_EXCL(pp)) {
3439 VM_STAT_ADD(segvnvmstats.fill_vp_pages[4]);
3440 page_unlock(pp);
3441 *downsize = 1;
3442 *ret_pszc = pp->p_szc;
3443 goto out;
3444 }
3445 targpp = pp;
3446 if (io_pplist != NULL) {
3447 VM_STAT_ADD(segvnvmstats.fill_vp_pages[5]);
3448 io_len = off - io_off;
3449 /*
3450 * Some file systems like NFS don't check EOF
3451 * conditions in fop_pageio(). Check it here
3452 * now that pages are locked SE_EXCL. Any file
3453 * truncation will wait until the pages are
3454 * unlocked so no need to worry that file will
3455 * be truncated after we check its size here.
3456 * XXX fix NFS to remove this check.
3457 */
3458 va.va_mask = VATTR_SIZE;
3459 if (fop_getattr(vp, &va, ATTR_HINT, svd->cred, NULL)) {
3460 VM_STAT_ADD(segvnvmstats.fill_vp_pages[6]);
3461 page_unlock(targpp);
3462 goto out;
3463 }
3464 if (btopr(va.va_size) < btopr(io_off + io_len)) {
3465 VM_STAT_ADD(segvnvmstats.fill_vp_pages[7]);
3466 *downsize = 1;
3467 *ret_pszc = 0;
3468 page_unlock(targpp);
3469 goto out;
3470 }
3471 io_err = fop_pageio(vp, io_pplist, io_off, io_len,
3472 B_READ, svd->cred, NULL);
3473 if (io_err) {
3474 VM_STAT_ADD(segvnvmstats.fill_vp_pages[8]);
3475 page_unlock(targpp);
3476 if (io_err == EDEADLK) {
3477 segvn_vmpss_pageio_deadlk_err++;
3478 }
3479 goto out;
3480 }
3481 nios++;
3482 VM_STAT_ADD(segvnvmstats.fill_vp_pages[9]);
3483 while (io_pplist != NULL) {
3484 pp = io_pplist;
3485 page_sub(&io_pplist, pp);
3486 ASSERT(page_iolock_assert(pp));
3487 page_io_unlock(pp);
3488 pgidx = (pp->p_offset - start_off) >>
3489 PAGESHIFT;
3490 ASSERT(pgidx < pages);
3491 ppa[pgidx] = pp;
3492 page_list_concat(&done_pplist, &pp);
3493 }
3494 }
3495 pp = targpp;
3496 ASSERT(PAGE_EXCL(pp));
3497 ASSERT(pp->p_szc <= pszc);
3498 if (pszc != 0 && !group_page_trylock(pp, SE_EXCL)) {
3499 VM_STAT_ADD(segvnvmstats.fill_vp_pages[10]);
3500 page_unlock(pp);
3501 *downsize = 1;
3502 *ret_pszc = pp->p_szc;
3503 goto out;
3504 }
3505 VM_STAT_ADD(segvnvmstats.fill_vp_pages[11]);
3506 /*
3507 * page szc chould have changed before the entire group was
3508 * locked. reread page szc.
3509 */
3510 pszc = pp->p_szc;
3511 ppages = page_get_pagecnt(pszc);
3512
3513 /* link just the roots */
3514 page_list_concat(&targ_pplist, &pp);
3515 page_sub(&pplist, newpp);
3516 page_list_concat(&repl_pplist, &newpp);
3517 off += PAGESIZE;
3518 while (--ppages != 0) {
3519 newpp = pplist;
3520 page_sub(&pplist, newpp);
3521 off += PAGESIZE;
3522 }
3523 io_off = off;
3524 }
3525 if (io_pplist != NULL) {
3526 VM_STAT_ADD(segvnvmstats.fill_vp_pages[12]);
3527 io_len = eoff - io_off;
3528 va.va_mask = VATTR_SIZE;
3529 if (fop_getattr(vp, &va, ATTR_HINT, svd->cred, NULL) != 0) {
3530 VM_STAT_ADD(segvnvmstats.fill_vp_pages[13]);
3531 goto out;
3532 }
3533 if (btopr(va.va_size) < btopr(io_off + io_len)) {
3534 VM_STAT_ADD(segvnvmstats.fill_vp_pages[14]);
3535 *downsize = 1;
3536 *ret_pszc = 0;
3537 goto out;
3538 }
3539 io_err = fop_pageio(vp, io_pplist, io_off, io_len,
3540 B_READ, svd->cred, NULL);
3541 if (io_err) {
3542 VM_STAT_ADD(segvnvmstats.fill_vp_pages[15]);
3543 if (io_err == EDEADLK) {
3544 segvn_vmpss_pageio_deadlk_err++;
3545 }
3546 goto out;
3547 }
3548 nios++;
3549 while (io_pplist != NULL) {
3550 pp = io_pplist;
3551 page_sub(&io_pplist, pp);
3552 ASSERT(page_iolock_assert(pp));
3553 page_io_unlock(pp);
3554 pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3555 ASSERT(pgidx < pages);
3556 ppa[pgidx] = pp;
3557 }
3558 }
3559 /*
3560 * we're now bound to succeed or panic.
3561 * remove pages from done_pplist. it's not needed anymore.
3562 */
3563 while (done_pplist != NULL) {
3564 pp = done_pplist;
3565 page_sub(&done_pplist, pp);
3566 }
3567 VM_STAT_ADD(segvnvmstats.fill_vp_pages[16]);
3568 ASSERT(pplist == NULL);
3569 *ppplist = NULL;
3570 while (targ_pplist != NULL) {
3571 int ret;
3572 VM_STAT_ADD(segvnvmstats.fill_vp_pages[17]);
3573 ASSERT(repl_pplist);
3574 pp = targ_pplist;
3575 page_sub(&targ_pplist, pp);
3576 pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3577 newpp = repl_pplist;
3578 page_sub(&repl_pplist, newpp);
3579 #ifdef DEBUG
3580 pfn = page_pptonum(pp);
3581 pszc = pp->p_szc;
3582 ppages = page_get_pagecnt(pszc);
3583 ASSERT(IS_P2ALIGNED(pfn, ppages));
3584 pfn = page_pptonum(newpp);
3585 ASSERT(IS_P2ALIGNED(pfn, ppages));
3586 ASSERT(P2PHASE(pfn, pages) == pgidx);
3587 #endif
3588 nreloc = 0;
3589 ret = page_relocate(&pp, &newpp, 0, 1, &nreloc, NULL);
3590 if (ret != 0 || nreloc == 0) {
3591 panic("segvn_fill_vp_pages: "
3592 "page_relocate failed");
3593 }
3594 pp = newpp;
3595 while (nreloc-- != 0) {
3596 ASSERT(PAGE_EXCL(pp));
3597 VERIFY(pp->p_object == &vp->v_object);
3598 ASSERT(pp->p_vnode == vp);
3599 ASSERT(pgidx ==
3600 ((pp->p_offset - start_off) >> PAGESHIFT));
3601 ppa[pgidx++] = pp;
3602 pp++;
3603 }
3604 }
3605
3606 if (svd->type == MAP_PRIVATE) {
3607 VM_STAT_ADD(segvnvmstats.fill_vp_pages[18]);
3608 for (i = 0; i < pages; i++) {
3609 ASSERT(ppa[i] != NULL);
3610 ASSERT(PAGE_EXCL(ppa[i]));
3611 VERIFY(ppa[i]->p_object == &vp->v_object);
3612 ASSERT(ppa[i]->p_vnode == vp);
3613 ASSERT(ppa[i]->p_offset ==
3614 start_off + (i << PAGESHIFT));
3615 page_downgrade(ppa[i]);
3616 }
3617 ppa[pages] = NULL;
3618 } else {
3619 VM_STAT_ADD(segvnvmstats.fill_vp_pages[19]);
3620 /*
3621 * the caller will still call fop_getpage() for shared segments
3622 * to check FS write permissions. For private segments we map
3623 * file read only anyway. so no fop_getpage is needed.
3624 */
3625 for (i = 0; i < pages; i++) {
3626 ASSERT(ppa[i] != NULL);
3627 ASSERT(PAGE_EXCL(ppa[i]));
3628 VERIFY(ppa[i]->p_object == &vp->v_object);
3629 ASSERT(ppa[i]->p_vnode == vp);
3630 ASSERT(ppa[i]->p_offset ==
3631 start_off + (i << PAGESHIFT));
3632 page_unlock(ppa[i]);
3633 }
3634 ppa[0] = NULL;
3635 }
3636
3637 return (1);
3638 out:
3639 /*
3640 * Do the cleanup. Unlock target pages we didn't relocate. They are
3641 * linked on targ_pplist by root pages. reassemble unused replacement
3642 * and io pages back to pplist.
3643 */
3644 if (io_pplist != NULL) {
3645 VM_STAT_ADD(segvnvmstats.fill_vp_pages[20]);
3646 pp = io_pplist;
3647 do {
3648 VERIFY(pp->p_object == &vp->v_object);
3649 ASSERT(pp->p_vnode == vp);
3650 ASSERT(pp->p_offset == io_off);
3651 ASSERT(page_iolock_assert(pp));
3652 page_io_unlock(pp);
3653 page_hashout(pp, false);
3654 io_off += PAGESIZE;
3655 } while ((pp = pp->p_next) != io_pplist);
3656 page_list_concat(&io_pplist, &pplist);
3657 pplist = io_pplist;
3658 }
3659 tmp_pplist = NULL;
3660 while (targ_pplist != NULL) {
3661 VM_STAT_ADD(segvnvmstats.fill_vp_pages[21]);
3662 pp = targ_pplist;
3663 ASSERT(PAGE_EXCL(pp));
3664 page_sub(&targ_pplist, pp);
3665
3666 pszc = pp->p_szc;
3667 ppages = page_get_pagecnt(pszc);
3668 ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3669
3670 if (pszc != 0) {
3671 group_page_unlock(pp);
3672 }
3673 page_unlock(pp);
3674
3675 pp = repl_pplist;
3676 ASSERT(pp != NULL);
3677 ASSERT(PAGE_EXCL(pp));
3678 ASSERT(pp->p_szc == szc);
3679 page_sub(&repl_pplist, pp);
3680
3681 ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3682
3683 /* relink replacement page */
3684 page_list_concat(&tmp_pplist, &pp);
3685 while (--ppages != 0) {
3686 VM_STAT_ADD(segvnvmstats.fill_vp_pages[22]);
3687 pp++;
3688 ASSERT(PAGE_EXCL(pp));
3689 ASSERT(pp->p_szc == szc);
3690 page_list_concat(&tmp_pplist, &pp);
3691 }
3692 }
3693 if (tmp_pplist != NULL) {
3694 VM_STAT_ADD(segvnvmstats.fill_vp_pages[23]);
3695 page_list_concat(&tmp_pplist, &pplist);
3696 pplist = tmp_pplist;
3697 }
3698 /*
3699 * at this point all pages are either on done_pplist or
3700 * pplist. They can't be all on done_pplist otherwise
3701 * we'd've been done.
3702 */
3703 ASSERT(pplist != NULL);
3704 if (nios != 0) {
3705 VM_STAT_ADD(segvnvmstats.fill_vp_pages[24]);
3706 pp = pplist;
3707 do {
3708 VM_STAT_ADD(segvnvmstats.fill_vp_pages[25]);
3709 ASSERT(pp->p_szc == szc);
3710 ASSERT(PAGE_EXCL(pp));
3711 VERIFY(pp->p_object != &vp->v_object);
3712 ASSERT(pp->p_vnode != vp);
3713 pp->p_szc = 0;
3714 } while ((pp = pp->p_next) != pplist);
3715
3716 pp = done_pplist;
3717 do {
3718 VM_STAT_ADD(segvnvmstats.fill_vp_pages[26]);
3719 ASSERT(pp->p_szc == szc);
3720 ASSERT(PAGE_EXCL(pp));
3721 VERIFY(pp->p_object == &vp->v_object);
3722 ASSERT(pp->p_vnode == vp);
3723 pp->p_szc = 0;
3724 } while ((pp = pp->p_next) != done_pplist);
3725
3726 while (pplist != NULL) {
3727 VM_STAT_ADD(segvnvmstats.fill_vp_pages[27]);
3728 pp = pplist;
3729 page_sub(&pplist, pp);
3730 page_free(pp, 0);
3731 }
3732
3733 while (done_pplist != NULL) {
3734 VM_STAT_ADD(segvnvmstats.fill_vp_pages[28]);
3735 pp = done_pplist;
3736 page_sub(&done_pplist, pp);
3737 page_unlock(pp);
3738 }
3739 *ppplist = NULL;
3740 return (0);
3741 }
3742 ASSERT(pplist == *ppplist);
3743 if (io_err) {
3744 VM_STAT_ADD(segvnvmstats.fill_vp_pages[29]);
3745 /*
3746 * don't downsize on io error.
3747 * see if vop_getpage succeeds.
3748 * pplist may still be used in this case
3749 * for relocations.
3750 */
3751 return (0);
3752 }
3753 VM_STAT_ADD(segvnvmstats.fill_vp_pages[30]);
3754 page_free_replacement_page(pplist);
3755 page_create_putback(pages);
3756 *ppplist = NULL;
3757 return (0);
3758 }
3759
3760 int segvn_anypgsz = 0;
3761
3762 #define SEGVN_RESTORE_SOFTLOCK_VP(type, pages) \
3763 if ((type) == F_SOFTLOCK) { \
3764 atomic_add_long((ulong_t *)&(svd)->softlockcnt, \
3765 -(pages)); \
3766 }
3767
3768 #define SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot) \
3769 if (IS_VMODSORT((ppa)[0]->p_vnode)) { \
3770 if ((rw) == S_WRITE) { \
3771 for (i = 0; i < (pages); i++) { \
3772 VERIFY((ppa)[i]->p_object == \
3773 (ppa)[i]->p_object); \
3774 ASSERT((ppa)[i]->p_vnode == \
3775 (ppa)[0]->p_vnode); \
3776 hat_setmod((ppa)[i]); \
3777 } \
3778 } else if ((rw) != S_OTHER && \
3779 ((prot) & (vpprot) & PROT_WRITE)) { \
3780 for (i = 0; i < (pages); i++) { \
3781 VERIFY((ppa)[i]->p_object == \
3782 (ppa)[i]->p_object); \
3783 ASSERT((ppa)[i]->p_vnode == \
3784 (ppa)[0]->p_vnode); \
3785 if (!hat_ismod((ppa)[i])) { \
3786 prot &= ~PROT_WRITE; \
3787 break; \
3788 } \
3789 } \
3790 } \
3791 }
3792
3793 #define SEGVN_VMSTAT_FLTVNPAGES(idx) \
3794 VM_STAT_ADD(segvnvmstats.fltvnpages[(idx)]);
3795
3796 static faultcode_t
3797 segvn_fault_vnodepages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
3798 caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
3799 caddr_t eaddr, int brkcow)
3800 {
3801 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
3802 struct anon_map *amp = svd->amp;
3803 uchar_t segtype = svd->type;
3804 uint_t szc = seg->s_szc;
3805 size_t pgsz = page_get_pagesize(szc);
3806 size_t maxpgsz = pgsz;
3807 pgcnt_t pages = btop(pgsz);
3808 pgcnt_t maxpages = pages;
3809 size_t ppasize = (pages + 1) * sizeof (page_t *);
3810 caddr_t a = lpgaddr;
3811 caddr_t maxlpgeaddr = lpgeaddr;
3812 uoff_t off = svd->offset + (uintptr_t)(a - seg->s_base);
3813 ulong_t aindx = svd->anon_index + seg_page(seg, a);
3814 struct vpage *vpage = (svd->vpage != NULL) ?
3815 &svd->vpage[seg_page(seg, a)] : NULL;
3816 vnode_t *vp = svd->vp;
3817 page_t **ppa;
3818 uint_t pszc;
3819 size_t ppgsz;
3820 pgcnt_t ppages;
3821 faultcode_t err = 0;
3822 int ierr;
3823 int vop_size_err = 0;
3824 uint_t protchk, prot, vpprot;
3825 ulong_t i;
3826 int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
3827 anon_sync_obj_t an_cookie;
3828 enum seg_rw arw;
3829 int alloc_failed = 0;
3830 int adjszc_chk;
3831 struct vattr va;
3832 page_t *pplist;
3833 pfn_t pfn;
3834 int physcontig;
3835 int upgrdfail;
3836 int segvn_anypgsz_vnode = 0; /* for now map vnode with 2 page sizes */
3837 int tron = (svd->tr_state == SEGVN_TR_ON);
3838
3839 ASSERT(szc != 0);
3840 ASSERT(vp != NULL);
3841 ASSERT(brkcow == 0 || amp != NULL);
3842 ASSERT(tron == 0 || amp != NULL);
3843 ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
3844 ASSERT(!(svd->flags & MAP_NORESERVE));
3845 ASSERT(type != F_SOFTUNLOCK);
3846 ASSERT(IS_P2ALIGNED(a, maxpgsz));
3847 ASSERT(amp == NULL || IS_P2ALIGNED(aindx, maxpages));
3848 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
3849 ASSERT(seg->s_szc < NBBY * sizeof (int));
3850 ASSERT(type != F_SOFTLOCK || lpgeaddr - a == maxpgsz);
3851 ASSERT(svd->tr_state != SEGVN_TR_INIT);
3852
3853 VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltvnpages[0]);
3854 VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltvnpages[1]);
3855
3856 if (svd->flags & MAP_TEXT) {
3857 hat_flag |= HAT_LOAD_TEXT;
3858 }
3859
3860 if (svd->pageprot) {
3861 switch (rw) {
3862 case S_READ:
3863 protchk = PROT_READ;
3864 break;
3865 case S_WRITE:
3866 protchk = PROT_WRITE;
3867 break;
3868 case S_EXEC:
3869 protchk = PROT_EXEC;
3870 break;
3871 case S_OTHER:
3872 default:
3873 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
3874 break;
3875 }
3876 } else {
3877 prot = svd->prot;
3878 /* caller has already done segment level protection check. */
3879 }
3880
3881 if (rw == S_WRITE && segtype == MAP_PRIVATE) {
3882 SEGVN_VMSTAT_FLTVNPAGES(2);
3883 arw = S_READ;
3884 } else {
3885 arw = rw;
3886 }
3887
3888 ppa = kmem_alloc(ppasize, KM_SLEEP);
3889
3890 VM_STAT_COND_ADD(amp != NULL, segvnvmstats.fltvnpages[3]);
3891
3892 for (;;) {
3893 adjszc_chk = 0;
3894 for (; a < lpgeaddr; a += pgsz, off += pgsz, aindx += pages) {
3895 if (adjszc_chk) {
3896 while (szc < seg->s_szc) {
3897 uintptr_t e;
3898 uint_t tszc;
3899 tszc = segvn_anypgsz_vnode ? szc + 1 :
3900 seg->s_szc;
3901 ppgsz = page_get_pagesize(tszc);
3902 if (!IS_P2ALIGNED(a, ppgsz) ||
3903 ((alloc_failed >> tszc) & 0x1)) {
3904 break;
3905 }
3906 SEGVN_VMSTAT_FLTVNPAGES(4);
3907 szc = tszc;
3908 pgsz = ppgsz;
3909 pages = btop(pgsz);
3910 e = P2ROUNDUP((uintptr_t)eaddr, pgsz);
3911 lpgeaddr = (caddr_t)e;
3912 }
3913 }
3914
3915 again:
3916 if (IS_P2ALIGNED(a, maxpgsz) && amp != NULL) {
3917 ASSERT(IS_P2ALIGNED(aindx, maxpages));
3918 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
3919 anon_array_enter(amp, aindx, &an_cookie);
3920 if (anon_get_ptr(amp->ahp, aindx) != NULL) {
3921 SEGVN_VMSTAT_FLTVNPAGES(5);
3922 ASSERT(anon_pages(amp->ahp, aindx,
3923 maxpages) == maxpages);
3924 anon_array_exit(&an_cookie);
3925 ANON_LOCK_EXIT(&amp->a_rwlock);
3926 err = segvn_fault_anonpages(hat, seg,
3927 a, a + maxpgsz, type, rw,
3928 MAX(a, addr),
3929 MIN(a + maxpgsz, eaddr), brkcow);
3930 if (err != 0) {
3931 SEGVN_VMSTAT_FLTVNPAGES(6);
3932 goto out;
3933 }
3934 if (szc < seg->s_szc) {
3935 szc = seg->s_szc;
3936 pgsz = maxpgsz;
3937 pages = maxpages;
3938 lpgeaddr = maxlpgeaddr;
3939 }
3940 goto next;
3941 } else {
3942 ASSERT(anon_pages(amp->ahp, aindx,
3943 maxpages) == 0);
3944 SEGVN_VMSTAT_FLTVNPAGES(7);
3945 anon_array_exit(&an_cookie);
3946 ANON_LOCK_EXIT(&amp->a_rwlock);
3947 }
3948 }
3949 ASSERT(!brkcow || IS_P2ALIGNED(a, maxpgsz));
3950 ASSERT(!tron || IS_P2ALIGNED(a, maxpgsz));
3951
3952 if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
3953 ASSERT(vpage != NULL);
3954 prot = VPP_PROT(vpage);
3955 ASSERT(sameprot(seg, a, maxpgsz));
3956 if ((prot & protchk) == 0) {
3957 SEGVN_VMSTAT_FLTVNPAGES(8);
3958 err = FC_PROT;
3959 goto out;
3960 }
3961 }
3962 if (type == F_SOFTLOCK) {
3963 atomic_add_long((ulong_t *)&svd->softlockcnt,
3964 pages);
3965 }
3966
3967 pplist = NULL;
3968 physcontig = 0;
3969 ppa[0] = NULL;
3970 if (!brkcow && !tron && szc &&
3971 !page_exists_physcontig(&vp->v_object, off, szc,
3972 segtype == MAP_PRIVATE ? ppa : NULL)) {
3973 SEGVN_VMSTAT_FLTVNPAGES(9);
3974 if (page_alloc_pages(&vp->v_object, seg, a,
3975 &pplist, NULL, szc, 0, 0) &&
3976 type != F_SOFTLOCK) {
3977 SEGVN_VMSTAT_FLTVNPAGES(10);
3978 pszc = 0;
3979 ierr = -1;
3980 alloc_failed |= (1 << szc);
3981 break;
3982 }
3983 if (pplist != NULL &&
3984 vp->v_mpssdata == SEGVN_PAGEIO) {
3985 int downsize;
3986 SEGVN_VMSTAT_FLTVNPAGES(11);
3987 physcontig = segvn_fill_vp_pages(svd,
3988 vp, off, szc, ppa, &pplist,
3989 &pszc, &downsize);
3990 ASSERT(!physcontig || pplist == NULL);
3991 if (!physcontig && downsize &&
3992 type != F_SOFTLOCK) {
3993 ASSERT(pplist == NULL);
3994 SEGVN_VMSTAT_FLTVNPAGES(12);
3995 ierr = -1;
3996 break;
3997 }
3998 ASSERT(!physcontig ||
3999 segtype == MAP_PRIVATE ||
4000 ppa[0] == NULL);
4001 if (physcontig && ppa[0] == NULL) {
4002 physcontig = 0;
4003 }
4004 }
4005 } else if (!brkcow && !tron && szc && ppa[0] != NULL) {
4006 SEGVN_VMSTAT_FLTVNPAGES(13);
4007 ASSERT(segtype == MAP_PRIVATE);
4008 physcontig = 1;
4009 }
4010
4011 if (!physcontig) {
4012 SEGVN_VMSTAT_FLTVNPAGES(14);
4013 ppa[0] = NULL;
4014 ierr = fop_getpage(vp, (offset_t)off, pgsz,
4015 &vpprot, ppa, pgsz, seg, a, arw,
4016 svd->cred, NULL);
4017 #ifdef DEBUG
4018 if (ierr == 0) {
4019 for (i = 0; i < pages; i++) {
4020 ASSERT(PAGE_LOCKED(ppa[i]));
4021 ASSERT(!PP_ISFREE(ppa[i]));
4022 VERIFY(ppa[i]->p_object ==
4023 &vp->v_object);
4024 ASSERT(ppa[i]->p_vnode == vp);
4025 ASSERT(ppa[i]->p_offset ==
4026 off + (i << PAGESHIFT));
4027 }
4028 }
4029 #endif /* DEBUG */
4030 if (segtype == MAP_PRIVATE) {
4031 SEGVN_VMSTAT_FLTVNPAGES(15);
4032 vpprot &= ~PROT_WRITE;
4033 }
4034 } else {
4035 ASSERT(segtype == MAP_PRIVATE);
4036 SEGVN_VMSTAT_FLTVNPAGES(16);
4037 vpprot = PROT_ALL & ~PROT_WRITE;
4038 ierr = 0;
4039 }
4040
4041 if (ierr != 0) {
4042 SEGVN_VMSTAT_FLTVNPAGES(17);
4043 if (pplist != NULL) {
4044 SEGVN_VMSTAT_FLTVNPAGES(18);
4045 page_free_replacement_page(pplist);
4046 page_create_putback(pages);
4047 }
4048 SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4049 if (a + pgsz <= eaddr) {
4050 SEGVN_VMSTAT_FLTVNPAGES(19);
4051 err = FC_MAKE_ERR(ierr);
4052 goto out;
4053 }
4054 va.va_mask = VATTR_SIZE;
4055 if (fop_getattr(vp, &va, 0, svd->cred, NULL)) {
4056 SEGVN_VMSTAT_FLTVNPAGES(20);
4057 err = FC_MAKE_ERR(EIO);
4058 goto out;
4059 }
4060 if (btopr(va.va_size) >= btopr(off + pgsz)) {
4061 SEGVN_VMSTAT_FLTVNPAGES(21);
4062 err = FC_MAKE_ERR(ierr);
4063 goto out;
4064 }
4065 if (btopr(va.va_size) <
4066 btopr(off + (eaddr - a))) {
4067 SEGVN_VMSTAT_FLTVNPAGES(22);
4068 err = FC_MAKE_ERR(ierr);
4069 goto out;
4070 }
4071 if (brkcow || tron || type == F_SOFTLOCK) {
4072 /* can't reduce map area */
4073 SEGVN_VMSTAT_FLTVNPAGES(23);
4074 vop_size_err = 1;
4075 goto out;
4076 }
4077 SEGVN_VMSTAT_FLTVNPAGES(24);
4078 ASSERT(szc != 0);
4079 pszc = 0;
4080 ierr = -1;
4081 break;
4082 }
4083
4084 if (amp != NULL) {
4085 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4086 anon_array_enter(amp, aindx, &an_cookie);
4087 }
4088 if (amp != NULL &&
4089 anon_get_ptr(amp->ahp, aindx) != NULL) {
4090 ulong_t taindx = P2ALIGN(aindx, maxpages);
4091
4092 SEGVN_VMSTAT_FLTVNPAGES(25);
4093 ASSERT(anon_pages(amp->ahp, taindx,
4094 maxpages) == maxpages);
4095 for (i = 0; i < pages; i++) {
4096 page_unlock(ppa[i]);
4097 }
4098 anon_array_exit(&an_cookie);
4099 ANON_LOCK_EXIT(&amp->a_rwlock);
4100 if (pplist != NULL) {
4101 page_free_replacement_page(pplist);
4102 page_create_putback(pages);
4103 }
4104 SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4105 if (szc < seg->s_szc) {
4106 SEGVN_VMSTAT_FLTVNPAGES(26);
4107 /*
4108 * For private segments SOFTLOCK
4109 * either always breaks cow (any rw
4110 * type except S_READ_NOCOW) or
4111 * address space is locked as writer
4112 * (S_READ_NOCOW case) and anon slots
4113 * can't show up on second check.
4114 * Therefore if we are here for
4115 * SOFTLOCK case it must be a cow
4116 * break but cow break never reduces
4117 * szc. text replication (tron) in
4118 * this case works as cow break.
4119 * Thus the assert below.
4120 */
4121 ASSERT(!brkcow && !tron &&
4122 type != F_SOFTLOCK);
4123 pszc = seg->s_szc;
4124 ierr = -2;
4125 break;
4126 }
4127 ASSERT(IS_P2ALIGNED(a, maxpgsz));
4128 goto again;
4129 }
4130 #ifdef DEBUG
4131 if (amp != NULL) {
4132 ulong_t taindx = P2ALIGN(aindx, maxpages);
4133 ASSERT(!anon_pages(amp->ahp, taindx, maxpages));
4134 }
4135 #endif /* DEBUG */
4136
4137 if (brkcow || tron) {
4138 ASSERT(amp != NULL);
4139 ASSERT(pplist == NULL);
4140 ASSERT(szc == seg->s_szc);
4141 ASSERT(IS_P2ALIGNED(a, maxpgsz));
4142 ASSERT(IS_P2ALIGNED(aindx, maxpages));
4143 SEGVN_VMSTAT_FLTVNPAGES(27);
4144 ierr = anon_map_privatepages(amp, aindx, szc,
4145 seg, a, prot, ppa, vpage, segvn_anypgsz,
4146 tron ? PG_LOCAL : 0, svd->cred);
4147 if (ierr != 0) {
4148 SEGVN_VMSTAT_FLTVNPAGES(28);
4149 anon_array_exit(&an_cookie);
4150 ANON_LOCK_EXIT(&amp->a_rwlock);
4151 SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4152 err = FC_MAKE_ERR(ierr);
4153 goto out;
4154 }
4155
4156 ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4157 /*
4158 * p_szc can't be changed for locked
4159 * swapfs pages.
4160 */
4161 ASSERT(svd->rcookie ==
4162 HAT_INVALID_REGION_COOKIE);
4163 hat_memload_array(hat, a, pgsz, ppa, prot,
4164 hat_flag);
4165
4166 if (!(hat_flag & HAT_LOAD_LOCK)) {
4167 SEGVN_VMSTAT_FLTVNPAGES(29);
4168 for (i = 0; i < pages; i++) {
4169 page_unlock(ppa[i]);
4170 }
4171 }
4172 anon_array_exit(&an_cookie);
4173 ANON_LOCK_EXIT(&amp->a_rwlock);
4174 goto next;
4175 }
4176
4177 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
4178 (!svd->pageprot && svd->prot == (prot & vpprot)));
4179
4180 pfn = page_pptonum(ppa[0]);
4181 /*
4182 * hat_page_demote() needs an SE_EXCL lock on one of
4183 * constituent page_t's and it decreases root's p_szc
4184 * last. This means if root's p_szc is equal szc and
4185 * all its constituent pages are locked
4186 * hat_page_demote() that could have changed p_szc to
4187 * szc is already done and no new have page_demote()
4188 * can start for this large page.
4189 */
4190
4191 /*
4192 * we need to make sure same mapping size is used for
4193 * the same address range if there's a possibility the
4194 * adddress is already mapped because hat layer panics
4195 * when translation is loaded for the range already
4196 * mapped with a different page size. We achieve it
4197 * by always using largest page size possible subject
4198 * to the constraints of page size, segment page size
4199 * and page alignment. Since mappings are invalidated
4200 * when those constraints change and make it
4201 * impossible to use previously used mapping size no
4202 * mapping size conflicts should happen.
4203 */
4204
4205 chkszc:
4206 if ((pszc = ppa[0]->p_szc) == szc &&
4207 IS_P2ALIGNED(pfn, pages)) {
4208
4209 SEGVN_VMSTAT_FLTVNPAGES(30);
4210 #ifdef DEBUG
4211 for (i = 0; i < pages; i++) {
4212 ASSERT(PAGE_LOCKED(ppa[i]));
4213 ASSERT(!PP_ISFREE(ppa[i]));
4214 ASSERT(page_pptonum(ppa[i]) ==
4215 pfn + i);
4216 ASSERT(ppa[i]->p_szc == szc);
4217 VERIFY(ppa[i]->p_object == &vp->v_object);
4218 ASSERT(ppa[i]->p_vnode == vp);
4219 ASSERT(ppa[i]->p_offset ==
4220 off + (i << PAGESHIFT));
4221 }
4222 #endif /* DEBUG */
4223 /*
4224 * All pages are of szc we need and they are
4225 * all locked so they can't change szc. load
4226 * translations.
4227 *
4228 * if page got promoted since last check
4229 * we don't need pplist.
4230 */
4231 if (pplist != NULL) {
4232 page_free_replacement_page(pplist);
4233 page_create_putback(pages);
4234 }
4235 if (PP_ISMIGRATE(ppa[0])) {
4236 page_migrate(seg, a, ppa, pages);
4237 }
4238 SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4239 prot, vpprot);
4240 hat_memload_array_region(hat, a, pgsz,
4241 ppa, prot & vpprot, hat_flag,
4242 svd->rcookie);
4243
4244 if (!(hat_flag & HAT_LOAD_LOCK)) {
4245 for (i = 0; i < pages; i++) {
4246 page_unlock(ppa[i]);
4247 }
4248 }
4249 if (amp != NULL) {
4250 anon_array_exit(&an_cookie);
4251 ANON_LOCK_EXIT(&amp->a_rwlock);
4252 }
4253 goto next;
4254 }
4255
4256 /*
4257 * See if upsize is possible.
4258 */
4259 if (pszc > szc && szc < seg->s_szc &&
4260 (segvn_anypgsz_vnode || pszc >= seg->s_szc)) {
4261 pgcnt_t aphase;
4262 uint_t pszc1 = MIN(pszc, seg->s_szc);
4263 ppgsz = page_get_pagesize(pszc1);
4264 ppages = btop(ppgsz);
4265 aphase = btop(P2PHASE((uintptr_t)a, ppgsz));
4266
4267 ASSERT(type != F_SOFTLOCK);
4268
4269 SEGVN_VMSTAT_FLTVNPAGES(31);
4270 if (aphase != P2PHASE(pfn, ppages)) {
4271 segvn_faultvnmpss_align_err4++;
4272 } else {
4273 SEGVN_VMSTAT_FLTVNPAGES(32);
4274 if (pplist != NULL) {
4275 page_t *pl = pplist;
4276 page_free_replacement_page(pl);
4277 page_create_putback(pages);
4278 }
4279 for (i = 0; i < pages; i++) {
4280 page_unlock(ppa[i]);
4281 }
4282 if (amp != NULL) {
4283 anon_array_exit(&an_cookie);
4284 ANON_LOCK_EXIT(&amp->a_rwlock);
4285 }
4286 pszc = pszc1;
4287 ierr = -2;
4288 break;
4289 }
4290 }
4291
4292 /*
4293 * check if we should use smallest mapping size.
4294 */
4295 upgrdfail = 0;
4296 if (szc == 0 ||
4297 (pszc >= szc &&
4298 !IS_P2ALIGNED(pfn, pages)) ||
4299 (pszc < szc &&
4300 !segvn_full_szcpages(ppa, szc, &upgrdfail,
4301 &pszc))) {
4302
4303 if (upgrdfail && type != F_SOFTLOCK) {
4304 /*
4305 * segvn_full_szcpages failed to lock
4306 * all pages EXCL. Size down.
4307 */
4308 ASSERT(pszc < szc);
4309
4310 SEGVN_VMSTAT_FLTVNPAGES(33);
4311
4312 if (pplist != NULL) {
4313 page_t *pl = pplist;
4314 page_free_replacement_page(pl);
4315 page_create_putback(pages);
4316 }
4317
4318 for (i = 0; i < pages; i++) {
4319 page_unlock(ppa[i]);
4320 }
4321 if (amp != NULL) {
4322 anon_array_exit(&an_cookie);
4323 ANON_LOCK_EXIT(&amp->a_rwlock);
4324 }
4325 ierr = -1;
4326 break;
4327 }
4328 if (szc != 0 && !upgrdfail) {
4329 segvn_faultvnmpss_align_err5++;
4330 }
4331 SEGVN_VMSTAT_FLTVNPAGES(34);
4332 if (pplist != NULL) {
4333 page_free_replacement_page(pplist);
4334 page_create_putback(pages);
4335 }
4336 SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4337 prot, vpprot);
4338 if (upgrdfail && segvn_anypgsz_vnode) {
4339 /* SOFTLOCK case */
4340 hat_memload_array_region(hat, a, pgsz,
4341 ppa, prot & vpprot, hat_flag,
4342 svd->rcookie);
4343 } else {
4344 for (i = 0; i < pages; i++) {
4345 hat_memload_region(hat,
4346 a + (i << PAGESHIFT),
4347 ppa[i], prot & vpprot,
4348 hat_flag, svd->rcookie);
4349 }
4350 }
4351 if (!(hat_flag & HAT_LOAD_LOCK)) {
4352 for (i = 0; i < pages; i++) {
4353 page_unlock(ppa[i]);
4354 }
4355 }
4356 if (amp != NULL) {
4357 anon_array_exit(&an_cookie);
4358 ANON_LOCK_EXIT(&amp->a_rwlock);
4359 }
4360 goto next;
4361 }
4362
4363 if (pszc == szc) {
4364 /*
4365 * segvn_full_szcpages() upgraded pages szc.
4366 */
4367 ASSERT(pszc == ppa[0]->p_szc);
4368 ASSERT(IS_P2ALIGNED(pfn, pages));
4369 goto chkszc;
4370 }
4371
4372 if (pszc > szc) {
4373 kmutex_t *szcmtx;
4374 SEGVN_VMSTAT_FLTVNPAGES(35);
4375 /*
4376 * p_szc of ppa[0] can change since we haven't
4377 * locked all constituent pages. Call
4378 * page_lock_szc() to prevent szc changes.
4379 * This should be a rare case that happens when
4380 * multiple segments use a different page size
4381 * to map the same file offsets.
4382 */
4383 szcmtx = page_szc_lock(ppa[0]);
4384 pszc = ppa[0]->p_szc;
4385 ASSERT(szcmtx != NULL || pszc == 0);
4386 ASSERT(ppa[0]->p_szc <= pszc);
4387 if (pszc <= szc) {
4388 SEGVN_VMSTAT_FLTVNPAGES(36);
4389 if (szcmtx != NULL) {
4390 mutex_exit(szcmtx);
4391 }
4392 goto chkszc;
4393 }
4394 if (pplist != NULL) {
4395 /*
4396 * page got promoted since last check.
4397 * we don't need preaalocated large
4398 * page.
4399 */
4400 SEGVN_VMSTAT_FLTVNPAGES(37);
4401 page_free_replacement_page(pplist);
4402 page_create_putback(pages);
4403 }
4404 SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4405 prot, vpprot);
4406 hat_memload_array_region(hat, a, pgsz, ppa,
4407 prot & vpprot, hat_flag, svd->rcookie);
4408 mutex_exit(szcmtx);
4409 if (!(hat_flag & HAT_LOAD_LOCK)) {
4410 for (i = 0; i < pages; i++) {
4411 page_unlock(ppa[i]);
4412 }
4413 }
4414 if (amp != NULL) {
4415 anon_array_exit(&an_cookie);
4416 ANON_LOCK_EXIT(&amp->a_rwlock);
4417 }
4418 goto next;
4419 }
4420
4421 /*
4422 * if page got demoted since last check
4423 * we could have not allocated larger page.
4424 * allocate now.
4425 */
4426 if (pplist == NULL &&
4427 page_alloc_pages(&vp->v_object, seg, a, &pplist,
4428 NULL, szc, 0, 0) && type != F_SOFTLOCK) {
4429 SEGVN_VMSTAT_FLTVNPAGES(38);
4430 for (i = 0; i < pages; i++) {
4431 page_unlock(ppa[i]);
4432 }
4433 if (amp != NULL) {
4434 anon_array_exit(&an_cookie);
4435 ANON_LOCK_EXIT(&amp->a_rwlock);
4436 }
4437 ierr = -1;
4438 alloc_failed |= (1 << szc);
4439 break;
4440 }
4441
4442 SEGVN_VMSTAT_FLTVNPAGES(39);
4443
4444 if (pplist != NULL) {
4445 segvn_relocate_pages(ppa, pplist);
4446 #ifdef DEBUG
4447 } else {
4448 ASSERT(type == F_SOFTLOCK);
4449 SEGVN_VMSTAT_FLTVNPAGES(40);
4450 #endif /* DEBUG */
4451 }
4452
4453 SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot);
4454
4455 if (pplist == NULL && segvn_anypgsz_vnode == 0) {
4456 ASSERT(type == F_SOFTLOCK);
4457 for (i = 0; i < pages; i++) {
4458 ASSERT(ppa[i]->p_szc < szc);
4459 hat_memload_region(hat,
4460 a + (i << PAGESHIFT),
4461 ppa[i], prot & vpprot, hat_flag,
4462 svd->rcookie);
4463 }
4464 } else {
4465 ASSERT(pplist != NULL || type == F_SOFTLOCK);
4466 hat_memload_array_region(hat, a, pgsz, ppa,
4467 prot & vpprot, hat_flag, svd->rcookie);
4468 }
4469 if (!(hat_flag & HAT_LOAD_LOCK)) {
4470 for (i = 0; i < pages; i++) {
4471 ASSERT(PAGE_SHARED(ppa[i]));
4472 page_unlock(ppa[i]);
4473 }
4474 }
4475 if (amp != NULL) {
4476 anon_array_exit(&an_cookie);
4477 ANON_LOCK_EXIT(&amp->a_rwlock);
4478 }
4479
4480 next:
4481 if (vpage != NULL) {
4482 vpage += pages;
4483 }
4484 adjszc_chk = 1;
4485 }
4486 if (a == lpgeaddr)
4487 break;
4488 ASSERT(a < lpgeaddr);
4489
4490 ASSERT(!brkcow && !tron && type != F_SOFTLOCK);
4491
4492 /*
4493 * ierr == -1 means we failed to map with a large page.
4494 * (either due to allocation/relocation failures or
4495 * misalignment with other mappings to this file.
4496 *
4497 * ierr == -2 means some other thread allocated a large page
4498 * after we gave up tp map with a large page. retry with
4499 * larger mapping.
4500 */
4501 ASSERT(ierr == -1 || ierr == -2);
4502 ASSERT(ierr == -2 || szc != 0);
4503 ASSERT(ierr == -1 || szc < seg->s_szc);
4504 if (ierr == -2) {
4505 SEGVN_VMSTAT_FLTVNPAGES(41);
4506 ASSERT(pszc > szc && pszc <= seg->s_szc);
4507 szc = pszc;
4508 } else if (segvn_anypgsz_vnode) {
4509 SEGVN_VMSTAT_FLTVNPAGES(42);
4510 szc--;
4511 } else {
4512 SEGVN_VMSTAT_FLTVNPAGES(43);
4513 ASSERT(pszc < szc);
4514 /*
4515 * other process created pszc large page.
4516 * but we still have to drop to 0 szc.
4517 */
4518 szc = 0;
4519 }
4520
4521 pgsz = page_get_pagesize(szc);
4522 pages = btop(pgsz);
4523 if (ierr == -2) {
4524 /*
4525 * Size up case. Note lpgaddr may only be needed for
4526 * softlock case so we don't adjust it here.
4527 */
4528 a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4529 ASSERT(a >= lpgaddr);
4530 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4531 off = svd->offset + (uintptr_t)(a - seg->s_base);
4532 aindx = svd->anon_index + seg_page(seg, a);
4533 vpage = (svd->vpage != NULL) ?
4534 &svd->vpage[seg_page(seg, a)] : NULL;
4535 } else {
4536 /*
4537 * Size down case. Note lpgaddr may only be needed for
4538 * softlock case so we don't adjust it here.
4539 */
4540 ASSERT(IS_P2ALIGNED(a, pgsz));
4541 ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4542 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4543 ASSERT(a < lpgeaddr);
4544 if (a < addr) {
4545 SEGVN_VMSTAT_FLTVNPAGES(44);
4546 /*
4547 * The beginning of the large page region can
4548 * be pulled to the right to make a smaller
4549 * region. We haven't yet faulted a single
4550 * page.
4551 */
4552 a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4553 ASSERT(a >= lpgaddr);
4554 off = svd->offset +
4555 (uintptr_t)(a - seg->s_base);
4556 aindx = svd->anon_index + seg_page(seg, a);
4557 vpage = (svd->vpage != NULL) ?
4558 &svd->vpage[seg_page(seg, a)] : NULL;
4559 }
4560 }
4561 }
4562 out:
4563 kmem_free(ppa, ppasize);
4564 if (!err && !vop_size_err) {
4565 SEGVN_VMSTAT_FLTVNPAGES(45);
4566 return (0);
4567 }
4568 if (type == F_SOFTLOCK && a > lpgaddr) {
4569 SEGVN_VMSTAT_FLTVNPAGES(46);
4570 segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4571 }
4572 if (!vop_size_err) {
4573 SEGVN_VMSTAT_FLTVNPAGES(47);
4574 return (err);
4575 }
4576 ASSERT(brkcow || tron || type == F_SOFTLOCK);
4577 /*
4578 * Large page end is mapped beyond the end of file and it's a cow
4579 * fault (can be a text replication induced cow) or softlock so we can't
4580 * reduce the map area. For now just demote the segment. This should
4581 * really only happen if the end of the file changed after the mapping
4582 * was established since when large page segments are created we make
4583 * sure they don't extend beyond the end of the file.
4584 */
4585 SEGVN_VMSTAT_FLTVNPAGES(48);
4586
4587 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4588 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4589 err = 0;
4590 if (seg->s_szc != 0) {
4591 segvn_fltvnpages_clrszc_cnt++;
4592 ASSERT(svd->softlockcnt == 0);
4593 err = segvn_clrszc(seg);
4594 if (err != 0) {
4595 segvn_fltvnpages_clrszc_err++;
4596 }
4597 }
4598 ASSERT(err || seg->s_szc == 0);
4599 SEGVN_LOCK_DOWNGRADE(seg->s_as, &svd->lock);
4600 /* segvn_fault will do its job as if szc had been zero to begin with */
4601 return (err == 0 ? IE_RETRY : FC_MAKE_ERR(err));
4602 }
4603
4604 /*
4605 * This routine will attempt to fault in one large page.
4606 * it will use smaller pages if that fails.
4607 * It should only be called for pure anonymous segments.
4608 */
4609 static faultcode_t
4610 segvn_fault_anonpages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
4611 caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
4612 caddr_t eaddr, int brkcow)
4613 {
4614 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4615 struct anon_map *amp = svd->amp;
4616 uchar_t segtype = svd->type;
4617 uint_t szc = seg->s_szc;
4618 size_t pgsz = page_get_pagesize(szc);
4619 size_t maxpgsz = pgsz;
4620 pgcnt_t pages = btop(pgsz);
4621 uint_t ppaszc = szc;
4622 caddr_t a = lpgaddr;
4623 ulong_t aindx = svd->anon_index + seg_page(seg, a);
4624 struct vpage *vpage = (svd->vpage != NULL) ?
4625 &svd->vpage[seg_page(seg, a)] : NULL;
4626 page_t **ppa;
4627 uint_t ppa_szc;
4628 faultcode_t err;
4629 int ierr;
4630 uint_t protchk, prot, vpprot;
4631 ulong_t i;
4632 int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
4633 anon_sync_obj_t cookie;
4634 int adjszc_chk;
4635 int pgflags = (svd->tr_state == SEGVN_TR_ON) ? PG_LOCAL : 0;
4636
4637 ASSERT(szc != 0);
4638 ASSERT(amp != NULL);
4639 ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
4640 ASSERT(!(svd->flags & MAP_NORESERVE));
4641 ASSERT(type != F_SOFTUNLOCK);
4642 ASSERT(IS_P2ALIGNED(a, maxpgsz));
4643 ASSERT(!brkcow || svd->tr_state == SEGVN_TR_OFF);
4644 ASSERT(svd->tr_state != SEGVN_TR_INIT);
4645
4646 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
4647
4648 VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltanpages[0]);
4649 VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltanpages[1]);
4650
4651 if (svd->flags & MAP_TEXT) {
4652 hat_flag |= HAT_LOAD_TEXT;
4653 }
4654
4655 if (svd->pageprot) {
4656 switch (rw) {
4657 case S_READ:
4658 protchk = PROT_READ;
4659 break;
4660 case S_WRITE:
4661 protchk = PROT_WRITE;
4662 break;
4663 case S_EXEC:
4664 protchk = PROT_EXEC;
4665 break;
4666 case S_OTHER:
4667 default:
4668 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4669 break;
4670 }
4671 VM_STAT_ADD(segvnvmstats.fltanpages[2]);
4672 } else {
4673 prot = svd->prot;
4674 /* caller has already done segment level protection check. */
4675 }
4676
4677 ppa = kmem_cache_alloc(segvn_szc_cache[ppaszc], KM_SLEEP);
4678 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4679 for (;;) {
4680 adjszc_chk = 0;
4681 for (; a < lpgeaddr; a += pgsz, aindx += pages) {
4682 if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
4683 VM_STAT_ADD(segvnvmstats.fltanpages[3]);
4684 ASSERT(vpage != NULL);
4685 prot = VPP_PROT(vpage);
4686 ASSERT(sameprot(seg, a, maxpgsz));
4687 if ((prot & protchk) == 0) {
4688 err = FC_PROT;
4689 goto error;
4690 }
4691 }
4692 if (adjszc_chk && IS_P2ALIGNED(a, maxpgsz) &&
4693 pgsz < maxpgsz) {
4694 ASSERT(a > lpgaddr);
4695 szc = seg->s_szc;
4696 pgsz = maxpgsz;
4697 pages = btop(pgsz);
4698 ASSERT(IS_P2ALIGNED(aindx, pages));
4699 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr,
4700 pgsz);
4701 }
4702 if (type == F_SOFTLOCK) {
4703 atomic_add_long((ulong_t *)&svd->softlockcnt,
4704 pages);
4705 }
4706 anon_array_enter(amp, aindx, &cookie);
4707 ppa_szc = (uint_t)-1;
4708 ierr = anon_map_getpages(amp, aindx, szc, seg, a,
4709 prot, &vpprot, ppa, &ppa_szc, vpage, rw, brkcow,
4710 segvn_anypgsz, pgflags, svd->cred);
4711 if (ierr != 0) {
4712 anon_array_exit(&cookie);
4713 VM_STAT_ADD(segvnvmstats.fltanpages[4]);
4714 if (type == F_SOFTLOCK) {
4715 atomic_add_long(
4716 (ulong_t *)&svd->softlockcnt,
4717 -pages);
4718 }
4719 if (ierr > 0) {
4720 VM_STAT_ADD(segvnvmstats.fltanpages[6]);
4721 err = FC_MAKE_ERR(ierr);
4722 goto error;
4723 }
4724 break;
4725 }
4726
4727 ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4728
4729 ASSERT(segtype == MAP_SHARED ||
4730 ppa[0]->p_szc <= szc);
4731 ASSERT(segtype == MAP_PRIVATE ||
4732 ppa[0]->p_szc >= szc);
4733
4734 /*
4735 * Handle pages that have been marked for migration
4736 */
4737 if (lgrp_optimizations())
4738 page_migrate(seg, a, ppa, pages);
4739
4740 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
4741
4742 if (segtype == MAP_SHARED) {
4743 vpprot |= PROT_WRITE;
4744 }
4745
4746 hat_memload_array(hat, a, pgsz, ppa,
4747 prot & vpprot, hat_flag);
4748
4749 if (hat_flag & HAT_LOAD_LOCK) {
4750 VM_STAT_ADD(segvnvmstats.fltanpages[7]);
4751 } else {
4752 VM_STAT_ADD(segvnvmstats.fltanpages[8]);
4753 for (i = 0; i < pages; i++)
4754 page_unlock(ppa[i]);
4755 }
4756 if (vpage != NULL)
4757 vpage += pages;
4758
4759 anon_array_exit(&cookie);
4760 adjszc_chk = 1;
4761 }
4762 if (a == lpgeaddr)
4763 break;
4764 ASSERT(a < lpgeaddr);
4765 /*
4766 * ierr == -1 means we failed to allocate a large page.
4767 * so do a size down operation.
4768 *
4769 * ierr == -2 means some other process that privately shares
4770 * pages with this process has allocated a larger page and we
4771 * need to retry with larger pages. So do a size up
4772 * operation. This relies on the fact that large pages are
4773 * never partially shared i.e. if we share any constituent
4774 * page of a large page with another process we must share the
4775 * entire large page. Note this cannot happen for SOFTLOCK
4776 * case, unless current address (a) is at the beginning of the
4777 * next page size boundary because the other process couldn't
4778 * have relocated locked pages.
4779 */
4780 ASSERT(ierr == -1 || ierr == -2);
4781
4782 if (segvn_anypgsz) {
4783 ASSERT(ierr == -2 || szc != 0);
4784 ASSERT(ierr == -1 || szc < seg->s_szc);
4785 szc = (ierr == -1) ? szc - 1 : szc + 1;
4786 } else {
4787 /*
4788 * For non COW faults and segvn_anypgsz == 0
4789 * we need to be careful not to loop forever
4790 * if existing page is found with szc other
4791 * than 0 or seg->s_szc. This could be due
4792 * to page relocations on behalf of DR or
4793 * more likely large page creation. For this
4794 * case simply re-size to existing page's szc
4795 * if returned by anon_map_getpages().
4796 */
4797 if (ppa_szc == (uint_t)-1) {
4798 szc = (ierr == -1) ? 0 : seg->s_szc;
4799 } else {
4800 ASSERT(ppa_szc <= seg->s_szc);
4801 ASSERT(ierr == -2 || ppa_szc < szc);
4802 ASSERT(ierr == -1 || ppa_szc > szc);
4803 szc = ppa_szc;
4804 }
4805 }
4806
4807 pgsz = page_get_pagesize(szc);
4808 pages = btop(pgsz);
4809 ASSERT(type != F_SOFTLOCK || ierr == -1 ||
4810 (IS_P2ALIGNED(a, pgsz) && IS_P2ALIGNED(lpgeaddr, pgsz)));
4811 if (type == F_SOFTLOCK) {
4812 /*
4813 * For softlocks we cannot reduce the fault area
4814 * (calculated based on the largest page size for this
4815 * segment) for size down and a is already next
4816 * page size aligned as assertted above for size
4817 * ups. Therefore just continue in case of softlock.
4818 */
4819 VM_STAT_ADD(segvnvmstats.fltanpages[9]);
4820 continue; /* keep lint happy */
4821 } else if (ierr == -2) {
4822
4823 /*
4824 * Size up case. Note lpgaddr may only be needed for
4825 * softlock case so we don't adjust it here.
4826 */
4827 VM_STAT_ADD(segvnvmstats.fltanpages[10]);
4828 a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4829 ASSERT(a >= lpgaddr);
4830 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4831 aindx = svd->anon_index + seg_page(seg, a);
4832 vpage = (svd->vpage != NULL) ?
4833 &svd->vpage[seg_page(seg, a)] : NULL;
4834 } else {
4835 /*
4836 * Size down case. Note lpgaddr may only be needed for
4837 * softlock case so we don't adjust it here.
4838 */
4839 VM_STAT_ADD(segvnvmstats.fltanpages[11]);
4840 ASSERT(IS_P2ALIGNED(a, pgsz));
4841 ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4842 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4843 ASSERT(a < lpgeaddr);
4844 if (a < addr) {
4845 /*
4846 * The beginning of the large page region can
4847 * be pulled to the right to make a smaller
4848 * region. We haven't yet faulted a single
4849 * page.
4850 */
4851 VM_STAT_ADD(segvnvmstats.fltanpages[12]);
4852 a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4853 ASSERT(a >= lpgaddr);
4854 aindx = svd->anon_index + seg_page(seg, a);
4855 vpage = (svd->vpage != NULL) ?
4856 &svd->vpage[seg_page(seg, a)] : NULL;
4857 }
4858 }
4859 }
4860 VM_STAT_ADD(segvnvmstats.fltanpages[13]);
4861 ANON_LOCK_EXIT(&amp->a_rwlock);
4862 kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4863 return (0);
4864 error:
4865 VM_STAT_ADD(segvnvmstats.fltanpages[14]);
4866 ANON_LOCK_EXIT(&amp->a_rwlock);
4867 kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4868 if (type == F_SOFTLOCK && a > lpgaddr) {
4869 VM_STAT_ADD(segvnvmstats.fltanpages[15]);
4870 segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4871 }
4872 return (err);
4873 }
4874
4875 int fltadvice = 1; /* set to free behind pages for sequential access */
4876
4877 /*
4878 * This routine is called via a machine specific fault handling routine.
4879 * It is also called by software routines wishing to lock or unlock
4880 * a range of addresses.
4881 *
4882 * Here is the basic algorithm:
4883 * If unlocking
4884 * Call segvn_softunlock
4885 * Return
4886 * endif
4887 * Checking and set up work
4888 * If we will need some non-anonymous pages
4889 * Call fop_getpage over the range of non-anonymous pages
4890 * endif
4891 * Loop over all addresses requested
4892 * Call segvn_faultpage passing in page list
4893 * to load up translations and handle anonymous pages
4894 * endloop
4895 * Load up translation to any additional pages in page list not
4896 * already handled that fit into this segment
4897 */
4898 static faultcode_t
4899 segvn_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
4900 enum fault_type type, enum seg_rw rw)
4901 {
4902 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4903 page_t **plp, **ppp, *pp;
4904 uoff_t off;
4905 caddr_t a;
4906 struct vpage *vpage;
4907 uint_t vpprot, prot;
4908 int err;
4909 page_t *pl[FAULT_TMP_PAGES_NUM + 1];
4910 size_t plsz, pl_alloc_sz;
4911 size_t page;
4912 ulong_t anon_index;
4913 struct anon_map *amp;
4914 int dogetpage = 0;
4915 caddr_t lpgaddr, lpgeaddr;
4916 size_t pgsz;
4917 anon_sync_obj_t cookie;
4918 int brkcow = BREAK_COW_SHARE(rw, type, svd->type);
4919
4920 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
4921 ASSERT(svd->amp == NULL || svd->rcookie == HAT_INVALID_REGION_COOKIE);
4922
4923 /*
4924 * First handle the easy stuff
4925 */
4926 if (type == F_SOFTUNLOCK) {
4927 if (rw == S_READ_NOCOW) {
4928 rw = S_READ;
4929 ASSERT(AS_WRITE_HELD(seg->s_as));
4930 }
4931 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4932 pgsz = (seg->s_szc == 0) ? PAGESIZE :
4933 page_get_pagesize(seg->s_szc);
4934 VM_STAT_COND_ADD(pgsz > PAGESIZE, segvnvmstats.fltanpages[16]);
4935 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
4936 segvn_softunlock(seg, lpgaddr, lpgeaddr - lpgaddr, rw);
4937 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4938 return (0);
4939 }
4940
4941 ASSERT(svd->tr_state == SEGVN_TR_OFF ||
4942 !HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
4943 if (brkcow == 0) {
4944 if (svd->tr_state == SEGVN_TR_INIT) {
4945 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4946 if (svd->tr_state == SEGVN_TR_INIT) {
4947 ASSERT(svd->vp != NULL && svd->amp == NULL);
4948 ASSERT(svd->flags & MAP_TEXT);
4949 ASSERT(svd->type == MAP_PRIVATE);
4950 segvn_textrepl(seg);
4951 ASSERT(svd->tr_state != SEGVN_TR_INIT);
4952 ASSERT(svd->tr_state != SEGVN_TR_ON ||
4953 svd->amp != NULL);
4954 }
4955 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4956 }
4957 } else if (svd->tr_state != SEGVN_TR_OFF) {
4958 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4959
4960 if (rw == S_WRITE && svd->tr_state != SEGVN_TR_OFF) {
4961 ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
4962 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4963 return (FC_PROT);
4964 }
4965
4966 if (svd->tr_state == SEGVN_TR_ON) {
4967 ASSERT(svd->vp != NULL && svd->amp != NULL);
4968 segvn_textunrepl(seg, 0);
4969 ASSERT(svd->amp == NULL &&
4970 svd->tr_state == SEGVN_TR_OFF);
4971 } else if (svd->tr_state != SEGVN_TR_OFF) {
4972 svd->tr_state = SEGVN_TR_OFF;
4973 }
4974 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
4975 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4976 }
4977
4978 top:
4979 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4980
4981 /*
4982 * If we have the same protections for the entire segment,
4983 * insure that the access being attempted is legitimate.
4984 */
4985
4986 if (svd->pageprot == 0) {
4987 uint_t protchk;
4988
4989 switch (rw) {
4990 case S_READ:
4991 case S_READ_NOCOW:
4992 protchk = PROT_READ;
4993 break;
4994 case S_WRITE:
4995 protchk = PROT_WRITE;
4996 break;
4997 case S_EXEC:
4998 protchk = PROT_EXEC;
4999 break;
5000 case S_OTHER:
5001 default:
5002 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
5003 break;
5004 }
5005
5006 if ((svd->prot & protchk) == 0) {
5007 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5008 return (FC_PROT); /* illegal access type */
5009 }
5010 }
5011
5012 if (brkcow && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5013 /* this must be SOFTLOCK S_READ fault */
5014 ASSERT(svd->amp == NULL);
5015 ASSERT(svd->tr_state == SEGVN_TR_OFF);
5016 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5017 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5018 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5019 /*
5020 * this must be the first ever non S_READ_NOCOW
5021 * softlock for this segment.
5022 */
5023 ASSERT(svd->softlockcnt == 0);
5024 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5025 HAT_REGION_TEXT);
5026 svd->rcookie = HAT_INVALID_REGION_COOKIE;
5027 }
5028 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5029 goto top;
5030 }
5031
5032 /*
5033 * We can't allow the long term use of softlocks for vmpss segments,
5034 * because in some file truncation cases we should be able to demote
5035 * the segment, which requires that there are no softlocks. The
5036 * only case where it's ok to allow a SOFTLOCK fault against a vmpss
5037 * segment is S_READ_NOCOW, where the caller holds the address space
5038 * locked as writer and calls softunlock before dropping the as lock.
5039 * S_READ_NOCOW is used by /proc to read memory from another user.
5040 *
5041 * Another deadlock between SOFTLOCK and file truncation can happen
5042 * because segvn_fault_vnodepages() calls the FS one pagesize at
5043 * a time. A second fop_getpage() call by segvn_fault_vnodepages()
5044 * can cause a deadlock because the first set of page_t's remain
5045 * locked SE_SHARED. To avoid this, we demote segments on a first
5046 * SOFTLOCK if they have a length greater than the segment's
5047 * page size.
5048 *
5049 * So for now, we only avoid demoting a segment on a SOFTLOCK when
5050 * the access type is S_READ_NOCOW and the fault length is less than
5051 * or equal to the segment's page size. While this is quite restrictive,
5052 * it should be the most common case of SOFTLOCK against a vmpss
5053 * segment.
5054 *
5055 * For S_READ_NOCOW, it's safe not to do a copy on write because the
5056 * caller makes sure no COW will be caused by another thread for a
5057 * softlocked page.
5058 */
5059 if (type == F_SOFTLOCK && svd->vp != NULL && seg->s_szc != 0) {
5060 int demote = 0;
5061
5062 if (rw != S_READ_NOCOW) {
5063 demote = 1;
5064 }
5065 if (!demote && len > PAGESIZE) {
5066 pgsz = page_get_pagesize(seg->s_szc);
5067 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr,
5068 lpgeaddr);
5069 if (lpgeaddr - lpgaddr > pgsz) {
5070 demote = 1;
5071 }
5072 }
5073
5074 ASSERT(demote || AS_WRITE_HELD(seg->s_as));
5075
5076 if (demote) {
5077 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5078 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5079 if (seg->s_szc != 0) {
5080 segvn_vmpss_clrszc_cnt++;
5081 ASSERT(svd->softlockcnt == 0);
5082 err = segvn_clrszc(seg);
5083 if (err) {
5084 segvn_vmpss_clrszc_err++;
5085 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5086 return (FC_MAKE_ERR(err));
5087 }
5088 }
5089 ASSERT(seg->s_szc == 0);
5090 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5091 goto top;
5092 }
5093 }
5094
5095 /*
5096 * Check to see if we need to allocate an anon_map structure.
5097 */
5098 if (svd->amp == NULL && (svd->vp == NULL || brkcow)) {
5099 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5100 /*
5101 * Drop the "read" lock on the segment and acquire
5102 * the "write" version since we have to allocate the
5103 * anon_map.
5104 */
5105 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5106 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5107
5108 if (svd->amp == NULL) {
5109 svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
5110 svd->amp->a_szc = seg->s_szc;
5111 }
5112 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5113
5114 /*
5115 * Start all over again since segment protections
5116 * may have changed after we dropped the "read" lock.
5117 */
5118 goto top;
5119 }
5120
5121 /*
5122 * S_READ_NOCOW vs S_READ distinction was
5123 * only needed for the code above. After
5124 * that we treat it as S_READ.
5125 */
5126 if (rw == S_READ_NOCOW) {
5127 ASSERT(type == F_SOFTLOCK);
5128 ASSERT(AS_WRITE_HELD(seg->s_as));
5129 rw = S_READ;
5130 }
5131
5132 amp = svd->amp;
5133
5134 /*
5135 * MADV_SEQUENTIAL work is ignored for large page segments.
5136 */
5137 if (seg->s_szc != 0) {
5138 pgsz = page_get_pagesize(seg->s_szc);
5139 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
5140 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
5141 if (svd->vp == NULL) {
5142 err = segvn_fault_anonpages(hat, seg, lpgaddr,
5143 lpgeaddr, type, rw, addr, addr + len, brkcow);
5144 } else {
5145 err = segvn_fault_vnodepages(hat, seg, lpgaddr,
5146 lpgeaddr, type, rw, addr, addr + len, brkcow);
5147 if (err == IE_RETRY) {
5148 ASSERT(seg->s_szc == 0);
5149 ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
5150 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5151 goto top;
5152 }
5153 }
5154 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5155 return (err);
5156 }
5157
5158 page = seg_page(seg, addr);
5159 if (amp != NULL) {
5160 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5161 anon_index = svd->anon_index + page;
5162
5163 if (type == F_PROT && rw == S_READ &&
5164 svd->tr_state == SEGVN_TR_OFF &&
5165 svd->type == MAP_PRIVATE && svd->pageprot == 0) {
5166 size_t index = anon_index;
5167 struct anon *ap;
5168
5169 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5170 /*
5171 * The fast path could apply to S_WRITE also, except
5172 * that the protection fault could be caused by lazy
5173 * tlb flush when ro->rw. In this case, the pte is
5174 * RW already. But RO in the other cpu's tlb causes
5175 * the fault. Since hat_chgprot won't do anything if
5176 * pte doesn't change, we may end up faulting
5177 * indefinitely until the RO tlb entry gets replaced.
5178 */
5179 for (a = addr; a < addr + len; a += PAGESIZE, index++) {
5180 anon_array_enter(amp, index, &cookie);
5181 ap = anon_get_ptr(amp->ahp, index);
5182 anon_array_exit(&cookie);
5183 if ((ap == NULL) || (ap->an_refcnt != 1)) {
5184 ANON_LOCK_EXIT(&amp->a_rwlock);
5185 goto slow;
5186 }
5187 }
5188 hat_chgprot(seg->s_as->a_hat, addr, len, svd->prot);
5189 ANON_LOCK_EXIT(&amp->a_rwlock);
5190 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5191 return (0);
5192 }
5193 }
5194 slow:
5195
5196 if (svd->vpage == NULL)
5197 vpage = NULL;
5198 else
5199 vpage = &svd->vpage[page];
5200
5201 off = svd->offset + (uintptr_t)(addr - seg->s_base);
5202
5203 /*
5204 * If MADV_SEQUENTIAL has been set for the particular page we
5205 * are faulting on, free behind all pages in the segment and put
5206 * them on the free list.
5207 */
5208
5209 if ((page != 0) && fltadvice && svd->tr_state != SEGVN_TR_ON) {
5210 struct vpage *vpp;
5211 ulong_t fanon_index;
5212 size_t fpage;
5213 uoff_t pgoff, fpgoff;
5214 struct vnode *fvp;
5215 struct anon *fap = NULL;
5216
5217 if (svd->advice == MADV_SEQUENTIAL ||
5218 (svd->pageadvice &&
5219 VPP_ADVICE(vpage) == MADV_SEQUENTIAL)) {
5220 pgoff = off - PAGESIZE;
5221 fpage = page - 1;
5222 if (vpage != NULL)
5223 vpp = &svd->vpage[fpage];
5224 if (amp != NULL)
5225 fanon_index = svd->anon_index + fpage;
5226
5227 while (pgoff > svd->offset) {
5228 if (svd->advice != MADV_SEQUENTIAL &&
5229 (!svd->pageadvice || (vpage &&
5230 VPP_ADVICE(vpp) != MADV_SEQUENTIAL)))
5231 break;
5232
5233 /*
5234 * If this is an anon page, we must find the
5235 * correct <vp, offset> for it
5236 */
5237 fap = NULL;
5238 if (amp != NULL) {
5239 ANON_LOCK_ENTER(&amp->a_rwlock,
5240 RW_READER);
5241 anon_array_enter(amp, fanon_index,
5242 &cookie);
5243 fap = anon_get_ptr(amp->ahp,
5244 fanon_index);
5245 if (fap != NULL) {
5246 swap_xlate(fap, &fvp, &fpgoff);
5247 } else {
5248 fpgoff = pgoff;
5249 fvp = svd->vp;
5250 }
5251 anon_array_exit(&cookie);
5252 ANON_LOCK_EXIT(&amp->a_rwlock);
5253 } else {
5254 fpgoff = pgoff;
5255 fvp = svd->vp;
5256 }
5257 if (fvp == NULL)
5258 break; /* XXX */
5259 /*
5260 * Skip pages that are free or have an
5261 * "exclusive" lock.
5262 */
5263 pp = page_lookup_nowait(&fvp->v_object,
5264 fpgoff, SE_SHARED);
5265 if (pp == NULL)
5266 break;
5267 /*
5268 * We don't need the page_struct_lock to test
5269 * as this is only advisory; even if we
5270 * acquire it someone might race in and lock
5271 * the page after we unlock and before the
5272 * PUTPAGE, then fop_putpage will do nothing.
5273 */
5274 if (pp->p_lckcnt == 0 && pp->p_cowcnt == 0) {
5275 /*
5276 * Hold the vnode before releasing
5277 * the page lock to prevent it from
5278 * being freed and re-used by some
5279 * other thread.
5280 */
5281 VN_HOLD(fvp);
5282 page_unlock(pp);
5283 /*
5284 * We should build a page list
5285 * to kluster putpages XXX
5286 */
5287 (void) fop_putpage(fvp,
5288 (offset_t)fpgoff, PAGESIZE,
5289 (B_DONTNEED|B_FREE|B_ASYNC),
5290 svd->cred, NULL);
5291 VN_RELE(fvp);
5292 } else {
5293 /*
5294 * XXX - Should the loop terminate if
5295 * the page is `locked'?
5296 */
5297 page_unlock(pp);
5298 }
5299 --vpp;
5300 --fanon_index;
5301 pgoff -= PAGESIZE;
5302 }
5303 }
5304 }
5305
5306 plp = pl;
5307 *plp = NULL;
5308 pl_alloc_sz = 0;
5309
5310 /*
5311 * See if we need to call fop_getpage for
5312 * *any* of the range being faulted on.
5313 * We can skip all of this work if there
5314 * was no original vnode.
5315 */
5316 if (svd->vp != NULL) {
5317 uoff_t vp_off;
5318 size_t vp_len;
5319 struct anon *ap;
5320 vnode_t *vp;
5321
5322 vp_off = off;
5323 vp_len = len;
5324
5325 if (amp == NULL)
5326 dogetpage = 1;
5327 else {
5328 /*
5329 * Only acquire reader lock to prevent amp->ahp
5330 * from being changed. It's ok to miss pages,
5331 * hence we don't do anon_array_enter
5332 */
5333 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5334 ap = anon_get_ptr(amp->ahp, anon_index);
5335
5336 if (len <= PAGESIZE)
5337 /* inline non_anon() */
5338 dogetpage = (ap == NULL);
5339 else
5340 dogetpage = non_anon(amp->ahp, anon_index,
5341 &vp_off, &vp_len);
5342 ANON_LOCK_EXIT(&amp->a_rwlock);
5343 }
5344
5345 if (dogetpage) {
5346 enum seg_rw arw;
5347 struct as *as = seg->s_as;
5348
5349 if (len > FAULT_TMP_PAGES_SZ) {
5350 /*
5351 * Page list won't fit in local array,
5352 * allocate one of the needed size.
5353 */
5354 pl_alloc_sz =
5355 (btop(len) + 1) * sizeof (page_t *);
5356 plp = kmem_alloc(pl_alloc_sz, KM_SLEEP);
5357 plp[0] = NULL;
5358 plsz = len;
5359 } else if (rw == S_WRITE && svd->type == MAP_PRIVATE ||
5360 svd->tr_state == SEGVN_TR_ON || rw == S_OTHER ||
5361 (((size_t)(addr + PAGESIZE) <
5362 (size_t)(seg->s_base + seg->s_size)) &&
5363 hat_probe(as->a_hat, addr + PAGESIZE))) {
5364 /*
5365 * Ask fop_getpage to return the exact number
5366 * of pages if
5367 * (a) this is a COW fault, or
5368 * (b) this is a software fault, or
5369 * (c) next page is already mapped.
5370 */
5371 plsz = len;
5372 } else {
5373 /*
5374 * Ask fop_getpage to return adjacent pages
5375 * within the segment.
5376 */
5377 plsz = MIN((size_t)FAULT_TMP_PAGES_SZ, (size_t)
5378 ((seg->s_base + seg->s_size) - addr));
5379 ASSERT((addr + plsz) <=
5380 (seg->s_base + seg->s_size));
5381 }
5382
5383 /*
5384 * Need to get some non-anonymous pages.
5385 * We need to make only one call to GETPAGE to do
5386 * this to prevent certain deadlocking conditions
5387 * when we are doing locking. In this case
5388 * non_anon() should have picked up the smallest
5389 * range which includes all the non-anonymous
5390 * pages in the requested range. We have to
5391 * be careful regarding which rw flag to pass in
5392 * because on a private mapping, the underlying
5393 * object is never allowed to be written.
5394 */
5395 if (rw == S_WRITE && svd->type == MAP_PRIVATE) {
5396 arw = S_READ;
5397 } else {
5398 arw = rw;
5399 }
5400 vp = svd->vp;
5401 err = fop_getpage(vp, (offset_t)vp_off, vp_len,
5402 &vpprot, plp, plsz, seg, addr + (vp_off - off), arw,
5403 svd->cred, NULL);
5404 if (err) {
5405 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5406 segvn_pagelist_rele(plp);
5407 if (pl_alloc_sz)
5408 kmem_free(plp, pl_alloc_sz);
5409 return (FC_MAKE_ERR(err));
5410 }
5411 if (svd->type == MAP_PRIVATE)
5412 vpprot &= ~PROT_WRITE;
5413 }
5414 }
5415
5416 /*
5417 * N.B. at this time the plp array has all the needed non-anon
5418 * pages in addition to (possibly) having some adjacent pages.
5419 */
5420
5421 /*
5422 * Always acquire the anon_array_lock to prevent
5423 * 2 threads from allocating separate anon slots for
5424 * the same "addr".
5425 *
5426 * If this is a copy-on-write fault and we don't already
5427 * have the anon_array_lock, acquire it to prevent the
5428 * fault routine from handling multiple copy-on-write faults
5429 * on the same "addr" in the same address space.
5430 *
5431 * Only one thread should deal with the fault since after
5432 * it is handled, the other threads can acquire a translation
5433 * to the newly created private page. This prevents two or
5434 * more threads from creating different private pages for the
5435 * same fault.
5436 *
5437 * We grab "serialization" lock here if this is a MAP_PRIVATE segment
5438 * to prevent deadlock between this thread and another thread
5439 * which has soft-locked this page and wants to acquire serial_lock.
5440 * ( bug 4026339 )
5441 *
5442 * The fix for bug 4026339 becomes unnecessary when using the
5443 * locking scheme with per amp rwlock and a global set of hash
5444 * lock, anon_array_lock. If we steal a vnode page when low
5445 * on memory and upgrad the page lock through page_rename,
5446 * then the page is PAGE_HANDLED, nothing needs to be done
5447 * for this page after returning from segvn_faultpage.
5448 *
5449 * But really, the page lock should be downgraded after
5450 * the stolen page is page_rename'd.
5451 */
5452
5453 if (amp != NULL)
5454 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5455
5456 /*
5457 * Ok, now loop over the address range and handle faults
5458 */
5459 for (a = addr; a < addr + len; a += PAGESIZE, off += PAGESIZE) {
5460 err = segvn_faultpage(hat, seg, a, off, vpage, plp, vpprot,
5461 type, rw, brkcow);
5462 if (err) {
5463 if (amp != NULL)
5464 ANON_LOCK_EXIT(&amp->a_rwlock);
5465 if (type == F_SOFTLOCK && a > addr) {
5466 segvn_softunlock(seg, addr, (a - addr),
5467 S_OTHER);
5468 }
5469 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5470 segvn_pagelist_rele(plp);
5471 if (pl_alloc_sz)
5472 kmem_free(plp, pl_alloc_sz);
5473 return (err);
5474 }
5475 if (vpage) {
5476 vpage++;
5477 } else if (svd->vpage) {
5478 page = seg_page(seg, addr);
5479 vpage = &svd->vpage[++page];
5480 }
5481 }
5482
5483 /* Didn't get pages from the underlying fs so we're done */
5484 if (!dogetpage)
5485 goto done;
5486
5487 /*
5488 * Now handle any other pages in the list returned.
5489 * If the page can be used, load up the translations now.
5490 * Note that the for loop will only be entered if "plp"
5491 * is pointing to a non-NULL page pointer which means that
5492 * fop_getpage() was called and vpprot has been initialized.
5493 */
5494 if (svd->pageprot == 0)
5495 prot = svd->prot & vpprot;
5496
5497
5498 /*
5499 * Large Files: diff should be unsigned value because we started
5500 * supporting > 2GB segment sizes from 2.5.1 and when a
5501 * large file of size > 2GB gets mapped to address space
5502 * the diff value can be > 2GB.
5503 */
5504
5505 for (ppp = plp; (pp = *ppp) != NULL; ppp++) {
5506 size_t diff;
5507 struct anon *ap;
5508 int anon_index;
5509 anon_sync_obj_t cookie;
5510 int hat_flag = HAT_LOAD_ADV;
5511
5512 if (svd->flags & MAP_TEXT) {
5513 hat_flag |= HAT_LOAD_TEXT;
5514 }
5515
5516 if (pp == PAGE_HANDLED)
5517 continue;
5518
5519 if (svd->tr_state != SEGVN_TR_ON &&
5520 pp->p_offset >= svd->offset &&
5521 pp->p_offset < svd->offset + seg->s_size) {
5522
5523 diff = pp->p_offset - svd->offset;
5524
5525 /*
5526 * Large Files: Following is the assertion
5527 * validating the above cast.
5528 */
5529 VERIFY(&svd->vp->v_object == pp->p_object);
5530 ASSERT(svd->vp == pp->p_vnode);
5531
5532 page = btop(diff);
5533 if (svd->pageprot)
5534 prot = VPP_PROT(&svd->vpage[page]) & vpprot;
5535
5536 /*
5537 * Prevent other threads in the address space from
5538 * creating private pages (i.e., allocating anon slots)
5539 * while we are in the process of loading translations
5540 * to additional pages returned by the underlying
5541 * object.
5542 */
5543 if (amp != NULL) {
5544 anon_index = svd->anon_index + page;
5545 anon_array_enter(amp, anon_index, &cookie);
5546 ap = anon_get_ptr(amp->ahp, anon_index);
5547 }
5548 if ((amp == NULL) || (ap == NULL)) {
5549 if (IS_VMODSORT(pp->p_vnode) ||
5550 enable_mbit_wa) {
5551 if (rw == S_WRITE)
5552 hat_setmod(pp);
5553 else if (rw != S_OTHER &&
5554 !hat_ismod(pp))
5555 prot &= ~PROT_WRITE;
5556 }
5557 /*
5558 * Skip mapping read ahead pages marked
5559 * for migration, so they will get migrated
5560 * properly on fault
5561 */
5562 ASSERT(amp == NULL ||
5563 svd->rcookie == HAT_INVALID_REGION_COOKIE);
5564 if ((prot & PROT_READ) && !PP_ISMIGRATE(pp)) {
5565 hat_memload_region(hat,
5566 seg->s_base + diff,
5567 pp, prot, hat_flag,
5568 svd->rcookie);
5569 }
5570 }
5571 if (amp != NULL)
5572 anon_array_exit(&cookie);
5573 }
5574 page_unlock(pp);
5575 }
5576 done:
5577 if (amp != NULL)
5578 ANON_LOCK_EXIT(&amp->a_rwlock);
5579 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5580 if (pl_alloc_sz)
5581 kmem_free(plp, pl_alloc_sz);
5582 return (0);
5583 }
5584
5585 /*
5586 * This routine is used to start I/O on pages asynchronously. XXX it will
5587 * only create PAGESIZE pages. At fault time they will be relocated into
5588 * larger pages.
5589 */
5590 static faultcode_t
5591 segvn_faulta(struct seg *seg, caddr_t addr)
5592 {
5593 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5594 int err;
5595 struct anon_map *amp;
5596 vnode_t *vp;
5597
5598 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
5599
5600 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5601 if ((amp = svd->amp) != NULL) {
5602 struct anon *ap;
5603
5604 /*
5605 * Reader lock to prevent amp->ahp from being changed.
5606 * This is advisory, it's ok to miss a page, so
5607 * we don't do anon_array_enter lock.
5608 */
5609 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5610 if ((ap = anon_get_ptr(amp->ahp,
5611 svd->anon_index + seg_page(seg, addr))) != NULL) {
5612
5613 err = anon_getpage(&ap, NULL, NULL,
5614 0, seg, addr, S_READ, svd->cred);
5615
5616 ANON_LOCK_EXIT(&amp->a_rwlock);
5617 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5618 if (err)
5619 return (FC_MAKE_ERR(err));
5620 return (0);
5621 }
5622 ANON_LOCK_EXIT(&amp->a_rwlock);
5623 }
5624
5625 if (svd->vp == NULL) {
5626 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5627 return (0); /* zfod page - do nothing now */
5628 }
5629
5630 vp = svd->vp;
5631 err = fop_getpage(vp,
5632 (offset_t)(svd->offset + (uintptr_t)(addr - seg->s_base)),
5633 PAGESIZE, NULL, NULL, 0, seg, addr,
5634 S_OTHER, svd->cred, NULL);
5635
5636 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5637 if (err)
5638 return (FC_MAKE_ERR(err));
5639 return (0);
5640 }
5641
5642 static int
5643 segvn_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
5644 {
5645 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5646 struct vpage *cvp, *svp, *evp;
5647 struct vnode *vp;
5648 size_t pgsz;
5649 pgcnt_t pgcnt;
5650 anon_sync_obj_t cookie;
5651 int unload_done = 0;
5652
5653 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
5654
5655 if ((svd->maxprot & prot) != prot)
5656 return (EACCES); /* violated maxprot */
5657
5658 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5659
5660 /* return if prot is the same */
5661 if (!svd->pageprot && svd->prot == prot) {
5662 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5663 return (0);
5664 }
5665
5666 /*
5667 * Since we change protections we first have to flush the cache.
5668 * This makes sure all the pagelock calls have to recheck
5669 * protections.
5670 */
5671 if (svd->softlockcnt > 0) {
5672 ASSERT(svd->tr_state == SEGVN_TR_OFF);
5673
5674 /*
5675 * If this is shared segment non 0 softlockcnt
5676 * means locked pages are still in use.
5677 */
5678 if (svd->type == MAP_SHARED) {
5679 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5680 return (EAGAIN);
5681 }
5682
5683 /*
5684 * Since we do have the segvn writers lock nobody can fill
5685 * the cache with entries belonging to this seg during
5686 * the purge. The flush either succeeds or we still have
5687 * pending I/Os.
5688 */
5689 segvn_purge(seg);
5690 if (svd->softlockcnt > 0) {
5691 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5692 return (EAGAIN);
5693 }
5694 }
5695
5696 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5697 ASSERT(svd->amp == NULL);
5698 ASSERT(svd->tr_state == SEGVN_TR_OFF);
5699 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5700 HAT_REGION_TEXT);
5701 svd->rcookie = HAT_INVALID_REGION_COOKIE;
5702 unload_done = 1;
5703 } else if (svd->tr_state == SEGVN_TR_INIT) {
5704 svd->tr_state = SEGVN_TR_OFF;
5705 } else if (svd->tr_state == SEGVN_TR_ON) {
5706 ASSERT(svd->amp != NULL);
5707 segvn_textunrepl(seg, 0);
5708 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5709 unload_done = 1;
5710 }
5711
5712 if ((prot & PROT_WRITE) && svd->type == MAP_SHARED &&
5713 svd->vp != NULL && (svd->vp->v_flag & VVMEXEC)) {
5714 ASSERT(vn_is_mapped(svd->vp, V_WRITE));
5715 segvn_inval_trcache(svd->vp);
5716 }
5717 if (seg->s_szc != 0) {
5718 int err;
5719 pgsz = page_get_pagesize(seg->s_szc);
5720 pgcnt = pgsz >> PAGESHIFT;
5721 ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
5722 if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
5723 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5724 ASSERT(seg->s_base != addr || seg->s_size != len);
5725 /*
5726 * If we are holding the as lock as a reader then
5727 * we need to return IE_RETRY and let the as
5728 * layer drop and re-acquire the lock as a writer.
5729 */
5730 if (AS_READ_HELD(seg->s_as))
5731 return (IE_RETRY);
5732 VM_STAT_ADD(segvnvmstats.demoterange[1]);
5733 if (svd->type == MAP_PRIVATE || svd->vp != NULL) {
5734 err = segvn_demote_range(seg, addr, len,
5735 SDR_END, 0);
5736 } else {
5737 uint_t szcvec = map_pgszcvec(seg->s_base,
5738 pgsz, (uintptr_t)seg->s_base,
5739 (svd->flags & MAP_TEXT), MAPPGSZC_SHM, 0);
5740 err = segvn_demote_range(seg, addr, len,
5741 SDR_END, szcvec);
5742 }
5743 if (err == 0)
5744 return (IE_RETRY);
5745 if (err == ENOMEM)
5746 return (IE_NOMEM);
5747 return (err);
5748 }
5749 }
5750
5751
5752 /*
5753 * If it's a private mapping and we're making it writable then we
5754 * may have to reserve the additional swap space now. If we are
5755 * making writable only a part of the segment then we use its vpage
5756 * array to keep a record of the pages for which we have reserved
5757 * swap. In this case we set the pageswap field in the segment's
5758 * segvn structure to record this.
5759 *
5760 * If it's a private mapping to a file (i.e., vp != NULL) and we're
5761 * removing write permission on the entire segment and we haven't
5762 * modified any pages, we can release the swap space.
5763 */
5764 if (svd->type == MAP_PRIVATE) {
5765 if (prot & PROT_WRITE) {
5766 if (!(svd->flags & MAP_NORESERVE) &&
5767 !(svd->swresv && svd->pageswap == 0)) {
5768 size_t sz = 0;
5769
5770 /*
5771 * Start by determining how much swap
5772 * space is required.
5773 */
5774 if (addr == seg->s_base &&
5775 len == seg->s_size &&
5776 svd->pageswap == 0) {
5777 /* The whole segment */
5778 sz = seg->s_size;
5779 } else {
5780 /*
5781 * Make sure that the vpage array
5782 * exists, and make a note of the
5783 * range of elements corresponding
5784 * to len.
5785 */
5786 segvn_vpage(seg);
5787 if (svd->vpage == NULL) {
5788 SEGVN_LOCK_EXIT(seg->s_as,
5789 &svd->lock);
5790 return (ENOMEM);
5791 }
5792 svp = &svd->vpage[seg_page(seg, addr)];
5793 evp = &svd->vpage[seg_page(seg,
5794 addr + len)];
5795
5796 if (svd->pageswap == 0) {
5797 /*
5798 * This is the first time we've
5799 * asked for a part of this
5800 * segment, so we need to
5801 * reserve everything we've
5802 * been asked for.
5803 */
5804 sz = len;
5805 } else {
5806 /*
5807 * We have to count the number
5808 * of pages required.
5809 */
5810 for (cvp = svp; cvp < evp;
5811 cvp++) {
5812 if (!VPP_ISSWAPRES(cvp))
5813 sz++;
5814 }
5815 sz <<= PAGESHIFT;
5816 }
5817 }
5818
5819 /* Try to reserve the necessary swap. */
5820 if (anon_resv_zone(sz,
5821 seg->s_as->a_proc->p_zone) == 0) {
5822 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5823 return (IE_NOMEM);
5824 }
5825
5826 /*
5827 * Make a note of how much swap space
5828 * we've reserved.
5829 */
5830 if (svd->pageswap == 0 && sz == seg->s_size) {
5831 svd->swresv = sz;
5832 } else {
5833 ASSERT(svd->vpage != NULL);
5834 svd->swresv += sz;
5835 svd->pageswap = 1;
5836 for (cvp = svp; cvp < evp; cvp++) {
5837 if (!VPP_ISSWAPRES(cvp))
5838 VPP_SETSWAPRES(cvp);
5839 }
5840 }
5841 }
5842 } else {
5843 /*
5844 * Swap space is released only if this segment
5845 * does not map anonymous memory, since read faults
5846 * on such segments still need an anon slot to read
5847 * in the data.
5848 */
5849 if (svd->swresv != 0 && svd->vp != NULL &&
5850 svd->amp == NULL && addr == seg->s_base &&
5851 len == seg->s_size && svd->pageprot == 0) {
5852 ASSERT(svd->pageswap == 0);
5853 anon_unresv_zone(svd->swresv,
5854 seg->s_as->a_proc->p_zone);
5855 svd->swresv = 0;
5856 }
5857 }
5858 }
5859
5860 if (addr == seg->s_base && len == seg->s_size && svd->vpage == NULL) {
5861 if (svd->prot == prot) {
5862 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5863 return (0); /* all done */
5864 }
5865 svd->prot = (uchar_t)prot;
5866 } else if (svd->type == MAP_PRIVATE) {
5867 struct anon *ap = NULL;
5868 page_t *pp;
5869 uoff_t offset, off;
5870 struct anon_map *amp;
5871 ulong_t anon_idx = 0;
5872
5873 /*
5874 * A vpage structure exists or else the change does not
5875 * involve the entire segment. Establish a vpage structure
5876 * if none is there. Then, for each page in the range,
5877 * adjust its individual permissions. Note that write-
5878 * enabling a MAP_PRIVATE page can affect the claims for
5879 * locked down memory. Overcommitting memory terminates
5880 * the operation.
5881 */
5882 segvn_vpage(seg);
5883 if (svd->vpage == NULL) {
5884 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5885 return (ENOMEM);
5886 }
5887 svd->pageprot = 1;
5888 if ((amp = svd->amp) != NULL) {
5889 anon_idx = svd->anon_index + seg_page(seg, addr);
5890 ASSERT(seg->s_szc == 0 ||
5891 IS_P2ALIGNED(anon_idx, pgcnt));
5892 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5893 }
5894
5895 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
5896 evp = &svd->vpage[seg_page(seg, addr + len)];
5897
5898 /*
5899 * See Statement at the beginning of segvn_lockop regarding
5900 * the way cowcnts and lckcnts are handled.
5901 */
5902 for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5903
5904 if (seg->s_szc != 0) {
5905 if (amp != NULL) {
5906 anon_array_enter(amp, anon_idx,
5907 &cookie);
5908 }
5909 if (IS_P2ALIGNED(anon_idx, pgcnt) &&
5910 !segvn_claim_pages(seg, svp, offset,
5911 anon_idx, prot)) {
5912 if (amp != NULL) {
5913 anon_array_exit(&cookie);
5914 }
5915 break;
5916 }
5917 if (amp != NULL) {
5918 anon_array_exit(&cookie);
5919 }
5920 anon_idx++;
5921 } else {
5922 if (amp != NULL) {
5923 anon_array_enter(amp, anon_idx,
5924 &cookie);
5925 ap = anon_get_ptr(amp->ahp, anon_idx++);
5926 }
5927
5928 if (VPP_ISPPLOCK(svp) &&
5929 VPP_PROT(svp) != prot) {
5930
5931 if (amp == NULL || ap == NULL) {
5932 vp = svd->vp;
5933 off = offset;
5934 } else
5935 swap_xlate(ap, &vp, &off);
5936 if (amp != NULL)
5937 anon_array_exit(&cookie);
5938
5939 if ((pp = page_lookup(&vp->v_object, off, SE_SHARED)) == NULL) {
5940 panic("segvn_setprot: no page");
5941 /*NOTREACHED*/
5942 }
5943 ASSERT(seg->s_szc == 0);
5944 if ((VPP_PROT(svp) ^ prot) &
5945 PROT_WRITE) {
5946 if (prot & PROT_WRITE) {
5947 if (!page_addclaim(
5948 pp)) {
5949 page_unlock(pp);
5950 break;
5951 }
5952 } else {
5953 if (!page_subclaim(
5954 pp)) {
5955 page_unlock(pp);
5956 break;
5957 }
5958 }
5959 }
5960 page_unlock(pp);
5961 } else if (amp != NULL)
5962 anon_array_exit(&cookie);
5963 }
5964 VPP_SETPROT(svp, prot);
5965 offset += PAGESIZE;
5966 }
5967 if (amp != NULL)
5968 ANON_LOCK_EXIT(&amp->a_rwlock);
5969
5970 /*
5971 * Did we terminate prematurely? If so, simply unload
5972 * the translations to the things we've updated so far.
5973 */
5974 if (svp != evp) {
5975 if (unload_done) {
5976 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5977 return (IE_NOMEM);
5978 }
5979 len = (svp - &svd->vpage[seg_page(seg, addr)]) *
5980 PAGESIZE;
5981 ASSERT(seg->s_szc == 0 || IS_P2ALIGNED(len, pgsz));
5982 if (len != 0)
5983 hat_unload(seg->s_as->a_hat, addr,
5984 len, HAT_UNLOAD);
5985 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5986 return (IE_NOMEM);
5987 }
5988 } else {
5989 segvn_vpage(seg);
5990 if (svd->vpage == NULL) {
5991 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5992 return (ENOMEM);
5993 }
5994 svd->pageprot = 1;
5995 evp = &svd->vpage[seg_page(seg, addr + len)];
5996 for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5997 VPP_SETPROT(svp, prot);
5998 }
5999 }
6000
6001 if (unload_done) {
6002 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6003 return (0);
6004 }
6005
6006 if (((prot & PROT_WRITE) != 0 &&
6007 (svd->vp != NULL || svd->type == MAP_PRIVATE)) ||
6008 (prot & ~PROT_USER) == PROT_NONE) {
6009 /*
6010 * Either private or shared data with write access (in
6011 * which case we need to throw out all former translations
6012 * so that we get the right translations set up on fault
6013 * and we don't allow write access to any copy-on-write pages
6014 * that might be around or to prevent write access to pages
6015 * representing holes in a file), or we don't have permission
6016 * to access the memory at all (in which case we have to
6017 * unload any current translations that might exist).
6018 */
6019 hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
6020 } else {
6021 /*
6022 * A shared mapping or a private mapping in which write
6023 * protection is going to be denied - just change all the
6024 * protections over the range of addresses in question.
6025 * segvn does not support any other attributes other
6026 * than prot so we can use hat_chgattr.
6027 */
6028 hat_chgattr(seg->s_as->a_hat, addr, len, prot);
6029 }
6030
6031 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6032
6033 return (0);
6034 }
6035
6036 /*
6037 * segvn_setpagesize is called via segop_setpagesize from as_setpagesize,
6038 * to determine if the seg is capable of mapping the requested szc.
6039 */
6040 static int
6041 segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
6042 {
6043 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6044 struct segvn_data *nsvd;
6045 struct anon_map *amp = svd->amp;
6046 struct seg *nseg;
6047 caddr_t eaddr = addr + len, a;
6048 size_t pgsz = page_get_pagesize(szc);
6049 pgcnt_t pgcnt = page_get_pagecnt(szc);
6050 int err;
6051 uoff_t off = svd->offset + (uintptr_t)(addr - seg->s_base);
6052
6053 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
6054 ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6055
6056 if (seg->s_szc == szc || segvn_lpg_disable != 0) {
6057 return (0);
6058 }
6059
6060 /*
6061 * addr should always be pgsz aligned but eaddr may be misaligned if
6062 * it's at the end of the segment.
6063 *
6064 * XXX we should assert this condition since as_setpagesize() logic
6065 * guarantees it.
6066 */
6067 if (!IS_P2ALIGNED(addr, pgsz) ||
6068 (!IS_P2ALIGNED(eaddr, pgsz) &&
6069 eaddr != seg->s_base + seg->s_size)) {
6070
6071 segvn_setpgsz_align_err++;
6072 return (EINVAL);
6073 }
6074
6075 if (amp != NULL && svd->type == MAP_SHARED) {
6076 ulong_t an_idx = svd->anon_index + seg_page(seg, addr);
6077 if (!IS_P2ALIGNED(an_idx, pgcnt)) {
6078
6079 segvn_setpgsz_anon_align_err++;
6080 return (EINVAL);
6081 }
6082 }
6083
6084 if ((svd->flags & MAP_NORESERVE) || seg->s_as == &kas ||
6085 szc > segvn_maxpgszc) {
6086 return (EINVAL);
6087 }
6088
6089 /* paranoid check */
6090 if (svd->vp != NULL &&
6091 (IS_SWAPFSVP(svd->vp) || VN_ISKAS(svd->vp))) {
6092 return (EINVAL);
6093 }
6094
6095 if (seg->s_szc == 0 && svd->vp != NULL &&
6096 map_addr_vacalign_check(addr, off)) {
6097 return (EINVAL);
6098 }
6099
6100 /*
6101 * Check that protections are the same within new page
6102 * size boundaries.
6103 */
6104 if (svd->pageprot) {
6105 for (a = addr; a < eaddr; a += pgsz) {
6106 if ((a + pgsz) > eaddr) {
6107 if (!sameprot(seg, a, eaddr - a)) {
6108 return (EINVAL);
6109 }
6110 } else {
6111 if (!sameprot(seg, a, pgsz)) {
6112 return (EINVAL);
6113 }
6114 }
6115 }
6116 }
6117
6118 /*
6119 * Since we are changing page size we first have to flush
6120 * the cache. This makes sure all the pagelock calls have
6121 * to recheck protections.
6122 */
6123 if (svd->softlockcnt > 0) {
6124 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6125
6126 /*
6127 * If this is shared segment non 0 softlockcnt
6128 * means locked pages are still in use.
6129 */
6130 if (svd->type == MAP_SHARED) {
6131 return (EAGAIN);
6132 }
6133
6134 /*
6135 * Since we do have the segvn writers lock nobody can fill
6136 * the cache with entries belonging to this seg during
6137 * the purge. The flush either succeeds or we still have
6138 * pending I/Os.
6139 */
6140 segvn_purge(seg);
6141 if (svd->softlockcnt > 0) {
6142 return (EAGAIN);
6143 }
6144 }
6145
6146 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6147 ASSERT(svd->amp == NULL);
6148 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6149 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6150 HAT_REGION_TEXT);
6151 svd->rcookie = HAT_INVALID_REGION_COOKIE;
6152 } else if (svd->tr_state == SEGVN_TR_INIT) {
6153 svd->tr_state = SEGVN_TR_OFF;
6154 } else if (svd->tr_state == SEGVN_TR_ON) {
6155 ASSERT(svd->amp != NULL);
6156 segvn_textunrepl(seg, 1);
6157 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6158 amp = NULL;
6159 }
6160
6161 /*
6162 * Operation for sub range of existing segment.
6163 */
6164 if (addr != seg->s_base || eaddr != (seg->s_base + seg->s_size)) {
6165 if (szc < seg->s_szc) {
6166 VM_STAT_ADD(segvnvmstats.demoterange[2]);
6167 err = segvn_demote_range(seg, addr, len, SDR_RANGE, 0);
6168 if (err == 0) {
6169 return (IE_RETRY);
6170 }
6171 if (err == ENOMEM) {
6172 return (IE_NOMEM);
6173 }
6174 return (err);
6175 }
6176 if (addr != seg->s_base) {
6177 nseg = segvn_split_seg(seg, addr);
6178 if (eaddr != (nseg->s_base + nseg->s_size)) {
6179 /* eaddr is szc aligned */
6180 (void) segvn_split_seg(nseg, eaddr);
6181 }
6182 return (IE_RETRY);
6183 }
6184 if (eaddr != (seg->s_base + seg->s_size)) {
6185 /* eaddr is szc aligned */
6186 (void) segvn_split_seg(seg, eaddr);
6187 }
6188 return (IE_RETRY);
6189 }
6190
6191 /*
6192 * Break any low level sharing and reset seg->s_szc to 0.
6193 */
6194 if ((err = segvn_clrszc(seg)) != 0) {
6195 if (err == ENOMEM) {
6196 err = IE_NOMEM;
6197 }
6198 return (err);
6199 }
6200 ASSERT(seg->s_szc == 0);
6201
6202 /*
6203 * If the end of the current segment is not pgsz aligned
6204 * then attempt to concatenate with the next segment.
6205 */
6206 if (!IS_P2ALIGNED(eaddr, pgsz)) {
6207 nseg = AS_SEGNEXT(seg->s_as, seg);
6208 if (nseg == NULL || nseg == seg || eaddr != nseg->s_base) {
6209 return (ENOMEM);
6210 }
6211 if (nseg->s_ops != &segvn_ops) {
6212 return (EINVAL);
6213 }
6214 nsvd = (struct segvn_data *)nseg->s_data;
6215 if (nsvd->softlockcnt > 0) {
6216 /*
6217 * If this is shared segment non 0 softlockcnt
6218 * means locked pages are still in use.
6219 */
6220 if (nsvd->type == MAP_SHARED) {
6221 return (EAGAIN);
6222 }
6223 segvn_purge(nseg);
6224 if (nsvd->softlockcnt > 0) {
6225 return (EAGAIN);
6226 }
6227 }
6228 err = segvn_clrszc(nseg);
6229 if (err == ENOMEM) {
6230 err = IE_NOMEM;
6231 }
6232 if (err != 0) {
6233 return (err);
6234 }
6235 ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6236 err = segvn_concat(seg, nseg, 1);
6237 if (err == -1) {
6238 return (EINVAL);
6239 }
6240 if (err == -2) {
6241 return (IE_NOMEM);
6242 }
6243 return (IE_RETRY);
6244 }
6245
6246 /*
6247 * May need to re-align anon array to
6248 * new szc.
6249 */
6250 if (amp != NULL) {
6251 if (!IS_P2ALIGNED(svd->anon_index, pgcnt)) {
6252 struct anon_hdr *nahp;
6253
6254 ASSERT(svd->type == MAP_PRIVATE);
6255
6256 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6257 ASSERT(amp->refcnt == 1);
6258 nahp = anon_create(btop(amp->size), ANON_NOSLEEP);
6259 if (nahp == NULL) {
6260 ANON_LOCK_EXIT(&amp->a_rwlock);
6261 return (IE_NOMEM);
6262 }
6263 if (anon_copy_ptr(amp->ahp, svd->anon_index,
6264 nahp, 0, btop(seg->s_size), ANON_NOSLEEP)) {
6265 anon_release(nahp, btop(amp->size));
6266 ANON_LOCK_EXIT(&amp->a_rwlock);
6267 return (IE_NOMEM);
6268 }
6269 anon_release(amp->ahp, btop(amp->size));
6270 amp->ahp = nahp;
6271 svd->anon_index = 0;
6272 ANON_LOCK_EXIT(&amp->a_rwlock);
6273 }
6274 }
6275 if (svd->vp != NULL && szc != 0) {
6276 struct vattr va;
6277 uoff_t eoffpage = svd->offset;
6278 va.va_mask = VATTR_SIZE;
6279 eoffpage += seg->s_size;
6280 eoffpage = btopr(eoffpage);
6281 if (fop_getattr(svd->vp, &va, 0, svd->cred, NULL) != 0) {
6282 segvn_setpgsz_getattr_err++;
6283 return (EINVAL);
6284 }
6285 if (btopr(va.va_size) < eoffpage) {
6286 segvn_setpgsz_eof_err++;
6287 return (EINVAL);
6288 }
6289 if (amp != NULL) {
6290 /*
6291 * anon_fill_cow_holes() may call fop_getpage().
6292 * don't take anon map lock here to avoid holding it
6293 * across fop_getpage() calls that may call back into
6294 * segvn for klsutering checks. We don't really need
6295 * anon map lock here since it's a private segment and
6296 * we hold as level lock as writers.
6297 */
6298 if ((err = anon_fill_cow_holes(seg, seg->s_base,
6299 amp->ahp, svd->anon_index, svd->vp, svd->offset,
6300 seg->s_size, szc, svd->prot, svd->vpage,
6301 svd->cred)) != 0) {
6302 return (EINVAL);
6303 }
6304 }
6305 segvn_setvnode_mpss(svd->vp);
6306 }
6307
6308 if (amp != NULL) {
6309 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6310 if (svd->type == MAP_PRIVATE) {
6311 amp->a_szc = szc;
6312 } else if (szc > amp->a_szc) {
6313 amp->a_szc = szc;
6314 }
6315 ANON_LOCK_EXIT(&amp->a_rwlock);
6316 }
6317
6318 seg->s_szc = szc;
6319
6320 return (0);
6321 }
6322
6323 static int
6324 segvn_clrszc(struct seg *seg)
6325 {
6326 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6327 struct anon_map *amp = svd->amp;
6328 size_t pgsz;
6329 pgcnt_t pages;
6330 int err = 0;
6331 caddr_t a = seg->s_base;
6332 caddr_t ea = a + seg->s_size;
6333 ulong_t an_idx = svd->anon_index;
6334 vnode_t *vp = svd->vp;
6335 struct vpage *vpage = svd->vpage;
6336 page_t *anon_pl[1 + 1], *pp;
6337 struct anon *ap, *oldap;
6338 uint_t prot = svd->prot, vpprot;
6339 int pageflag = 0;
6340
6341 ASSERT(AS_WRITE_HELD(seg->s_as) ||
6342 SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
6343 ASSERT(svd->softlockcnt == 0);
6344
6345 if (vp == NULL && amp == NULL) {
6346 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6347 seg->s_szc = 0;
6348 return (0);
6349 }
6350
6351 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6352 ASSERT(svd->amp == NULL);
6353 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6354 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6355 HAT_REGION_TEXT);
6356 svd->rcookie = HAT_INVALID_REGION_COOKIE;
6357 } else if (svd->tr_state == SEGVN_TR_ON) {
6358 ASSERT(svd->amp != NULL);
6359 segvn_textunrepl(seg, 1);
6360 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6361 amp = NULL;
6362 } else {
6363 if (svd->tr_state != SEGVN_TR_OFF) {
6364 ASSERT(svd->tr_state == SEGVN_TR_INIT);
6365 svd->tr_state = SEGVN_TR_OFF;
6366 }
6367
6368 /*
6369 * do HAT_UNLOAD_UNMAP since we are changing the pagesize.
6370 * unload argument is 0 when we are freeing the segment
6371 * and unload was already done.
6372 */
6373 hat_unload(seg->s_as->a_hat, seg->s_base, seg->s_size,
6374 HAT_UNLOAD_UNMAP);
6375 }
6376
6377 if (amp == NULL || svd->type == MAP_SHARED) {
6378 seg->s_szc = 0;
6379 return (0);
6380 }
6381
6382 pgsz = page_get_pagesize(seg->s_szc);
6383 pages = btop(pgsz);
6384
6385 /*
6386 * XXX anon rwlock is not really needed because this is a
6387 * private segment and we are writers.
6388 */
6389 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6390
6391 for (; a < ea; a += pgsz, an_idx += pages) {
6392 if ((oldap = anon_get_ptr(amp->ahp, an_idx)) != NULL) {
6393 ASSERT(vpage != NULL || svd->pageprot == 0);
6394 if (vpage != NULL) {
6395 ASSERT(sameprot(seg, a, pgsz));
6396 prot = VPP_PROT(vpage);
6397 pageflag = VPP_ISPPLOCK(vpage) ? LOCK_PAGE : 0;
6398 }
6399 if (seg->s_szc != 0) {
6400 ASSERT(vp == NULL || anon_pages(amp->ahp,
6401 an_idx, pages) == pages);
6402 if ((err = anon_map_demotepages(amp, an_idx,
6403 seg, a, prot, vpage, svd->cred)) != 0) {
6404 goto out;
6405 }
6406 } else {
6407 if (oldap->an_refcnt == 1) {
6408 continue;
6409 }
6410 if ((err = anon_getpage(&oldap, &vpprot,
6411 anon_pl, PAGESIZE, seg, a, S_READ,
6412 svd->cred))) {
6413 goto out;
6414 }
6415 if ((pp = anon_private(&ap, seg, a, prot,
6416 anon_pl[0], pageflag, svd->cred)) == NULL) {
6417 err = ENOMEM;
6418 goto out;
6419 }
6420 anon_decref(oldap);
6421 (void) anon_set_ptr(amp->ahp, an_idx, ap,
6422 ANON_SLEEP);
6423 page_unlock(pp);
6424 }
6425 }
6426 vpage = (vpage == NULL) ? NULL : vpage + pages;
6427 }
6428
6429 amp->a_szc = 0;
6430 seg->s_szc = 0;
6431 out:
6432 ANON_LOCK_EXIT(&amp->a_rwlock);
6433 return (err);
6434 }
6435
6436 static int
6437 segvn_claim_pages(
6438 struct seg *seg,
6439 struct vpage *svp,
6440 uoff_t off,
6441 ulong_t anon_idx,
6442 uint_t prot)
6443 {
6444 pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6445 size_t ppasize = (pgcnt + 1) * sizeof (page_t *);
6446 page_t **ppa;
6447 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6448 struct anon_map *amp = svd->amp;
6449 struct vpage *evp = svp + pgcnt;
6450 caddr_t addr = ((uintptr_t)(svp - svd->vpage) << PAGESHIFT)
6451 + seg->s_base;
6452 struct anon *ap;
6453 struct vnode *vp = svd->vp;
6454 page_t *pp;
6455 pgcnt_t pg_idx, i;
6456 int err = 0;
6457 anoff_t aoff;
6458 int anon = (amp != NULL) ? 1 : 0;
6459
6460 ASSERT(svd->type == MAP_PRIVATE);
6461 ASSERT(svd->vpage != NULL);
6462 ASSERT(seg->s_szc != 0);
6463 ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
6464 ASSERT(amp == NULL || IS_P2ALIGNED(anon_idx, pgcnt));
6465 ASSERT(sameprot(seg, addr, pgcnt << PAGESHIFT));
6466
6467 if (VPP_PROT(svp) == prot)
6468 return (1);
6469 if (!((VPP_PROT(svp) ^ prot) & PROT_WRITE))
6470 return (1);
6471
6472 ppa = kmem_alloc(ppasize, KM_SLEEP);
6473 if (anon && vp != NULL) {
6474 if (anon_get_ptr(amp->ahp, anon_idx) == NULL) {
6475 anon = 0;
6476 ASSERT(!anon_pages(amp->ahp, anon_idx, pgcnt));
6477 }
6478 ASSERT(!anon ||
6479 anon_pages(amp->ahp, anon_idx, pgcnt) == pgcnt);
6480 }
6481
6482 for (*ppa = NULL, pg_idx = 0; svp < evp; svp++, anon_idx++) {
6483 if (!VPP_ISPPLOCK(svp))
6484 continue;
6485 if (anon) {
6486 ap = anon_get_ptr(amp->ahp, anon_idx);
6487 if (ap == NULL) {
6488 panic("segvn_claim_pages: no anon slot");
6489 }
6490 swap_xlate(ap, &vp, &aoff);
6491 off = (uoff_t)aoff;
6492 }
6493 ASSERT(vp != NULL);
6494 if ((pp = page_lookup(&vp->v_object, (uoff_t)off, SE_SHARED)) == NULL) {
6495 panic("segvn_claim_pages: no page");
6496 }
6497 ppa[pg_idx++] = pp;
6498 off += PAGESIZE;
6499 }
6500
6501 if (ppa[0] == NULL) {
6502 kmem_free(ppa, ppasize);
6503 return (1);
6504 }
6505
6506 ASSERT(pg_idx <= pgcnt);
6507 ppa[pg_idx] = NULL;
6508
6509
6510 /* Find each large page within ppa, and adjust its claim */
6511
6512 /* Does ppa cover a single large page? */
6513 if (ppa[0]->p_szc == seg->s_szc) {
6514 if (prot & PROT_WRITE)
6515 err = page_addclaim_pages(ppa);
6516 else
6517 err = page_subclaim_pages(ppa);
6518 } else {
6519 for (i = 0; ppa[i]; i += pgcnt) {
6520 ASSERT(IS_P2ALIGNED(page_pptonum(ppa[i]), pgcnt));
6521 if (prot & PROT_WRITE)
6522 err = page_addclaim_pages(&ppa[i]);
6523 else
6524 err = page_subclaim_pages(&ppa[i]);
6525 if (err == 0)
6526 break;
6527 }
6528 }
6529
6530 for (i = 0; i < pg_idx; i++) {
6531 ASSERT(ppa[i] != NULL);
6532 page_unlock(ppa[i]);
6533 }
6534
6535 kmem_free(ppa, ppasize);
6536 return (err);
6537 }
6538
6539 /*
6540 * Returns right (upper address) segment if split occurred.
6541 * If the address is equal to the beginning or end of its segment it returns
6542 * the current segment.
6543 */
6544 static struct seg *
6545 segvn_split_seg(struct seg *seg, caddr_t addr)
6546 {
6547 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6548 struct seg *nseg;
6549 size_t nsize;
6550 struct segvn_data *nsvd;
6551
6552 ASSERT(AS_WRITE_HELD(seg->s_as));
6553 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6554
6555 ASSERT(addr >= seg->s_base);
6556 ASSERT(addr <= seg->s_base + seg->s_size);
6557 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6558
6559 if (addr == seg->s_base || addr == seg->s_base + seg->s_size)
6560 return (seg);
6561
6562 nsize = seg->s_base + seg->s_size - addr;
6563 seg->s_size = addr - seg->s_base;
6564 nseg = seg_alloc(seg->s_as, addr, nsize);
6565 ASSERT(nseg != NULL);
6566 nseg->s_ops = seg->s_ops;
6567 nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
6568 nseg->s_data = (void *)nsvd;
6569 nseg->s_szc = seg->s_szc;
6570 *nsvd = *svd;
6571 ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6572 nsvd->seg = nseg;
6573 rw_init(&nsvd->lock, NULL, RW_DEFAULT, NULL);
6574
6575 if (nsvd->vp != NULL) {
6576 VN_HOLD(nsvd->vp);
6577 nsvd->offset = svd->offset +
6578 (uintptr_t)(nseg->s_base - seg->s_base);
6579 if (nsvd->type == MAP_SHARED)
6580 lgrp_shm_policy_init(NULL, nsvd->vp);
6581 } else {
6582 /*
6583 * The offset for an anonymous segment has no signifigance in
6584 * terms of an offset into a file. If we were to use the above
6585 * calculation instead, the structures read out of
6586 * /proc/<pid>/xmap would be more difficult to decipher since
6587 * it would be unclear whether two seemingly contiguous
6588 * prxmap_t structures represented different segments or a
6589 * single segment that had been split up into multiple prxmap_t
6590 * structures (e.g. if some part of the segment had not yet
6591 * been faulted in).
6592 */
6593 nsvd->offset = 0;
6594 }
6595
6596 ASSERT(svd->softlockcnt == 0);
6597 ASSERT(svd->softlockcnt_sbase == 0);
6598 ASSERT(svd->softlockcnt_send == 0);
6599 crhold(svd->cred);
6600
6601 if (svd->vpage != NULL) {
6602 size_t bytes = vpgtob(seg_pages(seg));
6603 size_t nbytes = vpgtob(seg_pages(nseg));
6604 struct vpage *ovpage = svd->vpage;
6605
6606 svd->vpage = kmem_alloc(bytes, KM_SLEEP);
6607 bcopy(ovpage, svd->vpage, bytes);
6608 nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
6609 bcopy(ovpage + seg_pages(seg), nsvd->vpage, nbytes);
6610 kmem_free(ovpage, bytes + nbytes);
6611 }
6612 if (svd->amp != NULL && svd->type == MAP_PRIVATE) {
6613 struct anon_map *oamp = svd->amp, *namp;
6614 struct anon_hdr *nahp;
6615
6616 ANON_LOCK_ENTER(&oamp->a_rwlock, RW_WRITER);
6617 ASSERT(oamp->refcnt == 1);
6618 nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
6619 (void) anon_copy_ptr(oamp->ahp, svd->anon_index,
6620 nahp, 0, btop(seg->s_size), ANON_SLEEP);
6621
6622 namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
6623 namp->a_szc = nseg->s_szc;
6624 (void) anon_copy_ptr(oamp->ahp,
6625 svd->anon_index + btop(seg->s_size),
6626 namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
6627 anon_release(oamp->ahp, btop(oamp->size));
6628 oamp->ahp = nahp;
6629 oamp->size = seg->s_size;
6630 svd->anon_index = 0;
6631 nsvd->amp = namp;
6632 nsvd->anon_index = 0;
6633 ANON_LOCK_EXIT(&oamp->a_rwlock);
6634 } else if (svd->amp != NULL) {
6635 pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6636 ASSERT(svd->amp == nsvd->amp);
6637 ASSERT(seg->s_szc <= svd->amp->a_szc);
6638 nsvd->anon_index = svd->anon_index + seg_pages(seg);
6639 ASSERT(IS_P2ALIGNED(nsvd->anon_index, pgcnt));
6640 ANON_LOCK_ENTER(&svd->amp->a_rwlock, RW_WRITER);
6641 svd->amp->refcnt++;
6642 ANON_LOCK_EXIT(&svd->amp->a_rwlock);
6643 }
6644
6645 /*
6646 * Split the amount of swap reserved.
6647 */
6648 if (svd->swresv) {
6649 /*
6650 * For MAP_NORESERVE, only allocate swap reserve for pages
6651 * being used. Other segments get enough to cover whole
6652 * segment.
6653 */
6654 if (svd->flags & MAP_NORESERVE) {
6655 size_t oswresv;
6656
6657 ASSERT(svd->amp);
6658 oswresv = svd->swresv;
6659 svd->swresv = ptob(anon_pages(svd->amp->ahp,
6660 svd->anon_index, btop(seg->s_size)));
6661 nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
6662 nsvd->anon_index, btop(nseg->s_size)));
6663 ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
6664 } else {
6665 if (svd->pageswap) {
6666 svd->swresv = segvn_count_swap_by_vpages(seg);
6667 ASSERT(nsvd->swresv >= svd->swresv);
6668 nsvd->swresv -= svd->swresv;
6669 } else {
6670 ASSERT(svd->swresv == seg->s_size +
6671 nseg->s_size);
6672 svd->swresv = seg->s_size;
6673 nsvd->swresv = nseg->s_size;
6674 }
6675 }
6676 }
6677
6678 return (nseg);
6679 }
6680
6681 /*
6682 * called on memory operations (unmap, setprot, setpagesize) for a subset
6683 * of a large page segment to either demote the memory range (SDR_RANGE)
6684 * or the ends (SDR_END) by addr/len.
6685 *
6686 * returns 0 on success. returns errno, including ENOMEM, on failure.
6687 */
6688 static int
6689 segvn_demote_range(
6690 struct seg *seg,
6691 caddr_t addr,
6692 size_t len,
6693 int flag,
6694 uint_t szcvec)
6695 {
6696 caddr_t eaddr = addr + len;
6697 caddr_t lpgaddr, lpgeaddr;
6698 struct seg *nseg;
6699 struct seg *badseg1 = NULL;
6700 struct seg *badseg2 = NULL;
6701 size_t pgsz;
6702 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6703 int err;
6704 uint_t szc = seg->s_szc;
6705 uint_t tszcvec;
6706
6707 ASSERT(AS_WRITE_HELD(seg->s_as));
6708 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6709 ASSERT(szc != 0);
6710 pgsz = page_get_pagesize(szc);
6711 ASSERT(seg->s_base != addr || seg->s_size != len);
6712 ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6713 ASSERT(svd->softlockcnt == 0);
6714 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6715 ASSERT(szcvec == 0 || (flag == SDR_END && svd->type == MAP_SHARED));
6716
6717 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
6718 ASSERT(flag == SDR_RANGE || eaddr < lpgeaddr || addr > lpgaddr);
6719 if (flag == SDR_RANGE) {
6720 /* demote entire range */
6721 badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6722 (void) segvn_split_seg(nseg, lpgeaddr);
6723 ASSERT(badseg1->s_base == lpgaddr);
6724 ASSERT(badseg1->s_size == lpgeaddr - lpgaddr);
6725 } else if (addr != lpgaddr) {
6726 ASSERT(flag == SDR_END);
6727 badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6728 if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz &&
6729 eaddr < lpgaddr + 2 * pgsz) {
6730 (void) segvn_split_seg(nseg, lpgeaddr);
6731 ASSERT(badseg1->s_base == lpgaddr);
6732 ASSERT(badseg1->s_size == 2 * pgsz);
6733 } else {
6734 nseg = segvn_split_seg(nseg, lpgaddr + pgsz);
6735 ASSERT(badseg1->s_base == lpgaddr);
6736 ASSERT(badseg1->s_size == pgsz);
6737 if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz) {
6738 ASSERT(lpgeaddr - lpgaddr > 2 * pgsz);
6739 nseg = segvn_split_seg(nseg, lpgeaddr - pgsz);
6740 badseg2 = nseg;
6741 (void) segvn_split_seg(nseg, lpgeaddr);
6742 ASSERT(badseg2->s_base == lpgeaddr - pgsz);
6743 ASSERT(badseg2->s_size == pgsz);
6744 }
6745 }
6746 } else {
6747 ASSERT(flag == SDR_END);
6748 ASSERT(eaddr < lpgeaddr);
6749 badseg1 = nseg = segvn_split_seg(seg, lpgeaddr - pgsz);
6750 (void) segvn_split_seg(nseg, lpgeaddr);
6751 ASSERT(badseg1->s_base == lpgeaddr - pgsz);
6752 ASSERT(badseg1->s_size == pgsz);
6753 }
6754
6755 ASSERT(badseg1 != NULL);
6756 ASSERT(badseg1->s_szc == szc);
6757 ASSERT(flag == SDR_RANGE || badseg1->s_size == pgsz ||
6758 badseg1->s_size == 2 * pgsz);
6759 ASSERT(sameprot(badseg1, badseg1->s_base, pgsz));
6760 ASSERT(badseg1->s_size == pgsz ||
6761 sameprot(badseg1, badseg1->s_base + pgsz, pgsz));
6762 if (err = segvn_clrszc(badseg1)) {
6763 return (err);
6764 }
6765 ASSERT(badseg1->s_szc == 0);
6766
6767 if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6768 uint_t tszc = highbit(tszcvec) - 1;
6769 caddr_t ta = MAX(addr, badseg1->s_base);
6770 caddr_t te;
6771 size_t tpgsz = page_get_pagesize(tszc);
6772
6773 ASSERT(svd->type == MAP_SHARED);
6774 ASSERT(flag == SDR_END);
6775 ASSERT(tszc < szc && tszc > 0);
6776
6777 if (eaddr > badseg1->s_base + badseg1->s_size) {
6778 te = badseg1->s_base + badseg1->s_size;
6779 } else {
6780 te = eaddr;
6781 }
6782
6783 ASSERT(ta <= te);
6784 badseg1->s_szc = tszc;
6785 if (!IS_P2ALIGNED(ta, tpgsz) || !IS_P2ALIGNED(te, tpgsz)) {
6786 if (badseg2 != NULL) {
6787 err = segvn_demote_range(badseg1, ta, te - ta,
6788 SDR_END, tszcvec);
6789 if (err != 0) {
6790 return (err);
6791 }
6792 } else {
6793 return (segvn_demote_range(badseg1, ta,
6794 te - ta, SDR_END, tszcvec));
6795 }
6796 }
6797 }
6798
6799 if (badseg2 == NULL)
6800 return (0);
6801 ASSERT(badseg2->s_szc == szc);
6802 ASSERT(badseg2->s_size == pgsz);
6803 ASSERT(sameprot(badseg2, badseg2->s_base, badseg2->s_size));
6804 if (err = segvn_clrszc(badseg2)) {
6805 return (err);
6806 }
6807 ASSERT(badseg2->s_szc == 0);
6808
6809 if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6810 uint_t tszc = highbit(tszcvec) - 1;
6811 size_t tpgsz = page_get_pagesize(tszc);
6812
6813 ASSERT(svd->type == MAP_SHARED);
6814 ASSERT(flag == SDR_END);
6815 ASSERT(tszc < szc && tszc > 0);
6816 ASSERT(badseg2->s_base > addr);
6817 ASSERT(eaddr > badseg2->s_base);
6818 ASSERT(eaddr < badseg2->s_base + badseg2->s_size);
6819
6820 badseg2->s_szc = tszc;
6821 if (!IS_P2ALIGNED(eaddr, tpgsz)) {
6822 return (segvn_demote_range(badseg2, badseg2->s_base,
6823 eaddr - badseg2->s_base, SDR_END, tszcvec));
6824 }
6825 }
6826
6827 return (0);
6828 }
6829
6830 static int
6831 segvn_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
6832 {
6833 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6834 struct vpage *vp, *evp;
6835
6836 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6837
6838 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6839 /*
6840 * If segment protection can be used, simply check against them.
6841 */
6842 if (svd->pageprot == 0) {
6843 int err;
6844
6845 err = ((svd->prot & prot) != prot) ? EACCES : 0;
6846 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6847 return (err);
6848 }
6849
6850 /*
6851 * Have to check down to the vpage level.
6852 */
6853 evp = &svd->vpage[seg_page(seg, addr + len)];
6854 for (vp = &svd->vpage[seg_page(seg, addr)]; vp < evp; vp++) {
6855 if ((VPP_PROT(vp) & prot) != prot) {
6856 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6857 return (EACCES);
6858 }
6859 }
6860 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6861 return (0);
6862 }
6863
6864 static int
6865 segvn_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
6866 {
6867 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6868 size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
6869
6870 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6871
6872 if (pgno != 0) {
6873 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6874 if (svd->pageprot == 0) {
6875 do {
6876 protv[--pgno] = svd->prot;
6877 } while (pgno != 0);
6878 } else {
6879 size_t pgoff = seg_page(seg, addr);
6880
6881 do {
6882 pgno--;
6883 protv[pgno] = VPP_PROT(&svd->vpage[pgno+pgoff]);
6884 } while (pgno != 0);
6885 }
6886 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6887 }
6888 return (0);
6889 }
6890
6891 static uoff_t
6892 segvn_getoffset(struct seg *seg, caddr_t addr)
6893 {
6894 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6895
6896 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6897
6898 return (svd->offset + (uintptr_t)(addr - seg->s_base));
6899 }
6900
6901 /*ARGSUSED*/
6902 static int
6903 segvn_gettype(struct seg *seg, caddr_t addr)
6904 {
6905 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6906
6907 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6908
6909 return (svd->type | (svd->flags & (MAP_NORESERVE | MAP_TEXT |
6910 MAP_INITDATA)));
6911 }
6912
6913 /*ARGSUSED*/
6914 static int
6915 segvn_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
6916 {
6917 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6918
6919 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6920
6921 *vpp = svd->vp;
6922 return (0);
6923 }
6924
6925 /*
6926 * Check to see if it makes sense to do kluster/read ahead to
6927 * addr + delta relative to the mapping at addr. We assume here
6928 * that delta is a signed PAGESIZE'd multiple (which can be negative).
6929 *
6930 * For segvn, we currently "approve" of the action if we are
6931 * still in the segment and it maps from the same vp/off,
6932 * or if the advice stored in segvn_data or vpages allows it.
6933 * Currently, klustering is not allowed only if MADV_RANDOM is set.
6934 */
6935 static int
6936 segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
6937 {
6938 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6939 struct anon *oap, *ap;
6940 ssize_t pd;
6941 size_t page;
6942 struct vnode *vp1, *vp2;
6943 uoff_t off1, off2;
6944 struct anon_map *amp;
6945
6946 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6947 ASSERT(AS_WRITE_HELD(seg->s_as) ||
6948 SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
6949
6950 if (addr + delta < seg->s_base ||
6951 addr + delta >= (seg->s_base + seg->s_size))
6952 return (-1); /* exceeded segment bounds */
6953
6954 pd = delta / (ssize_t)PAGESIZE; /* divide to preserve sign bit */
6955 page = seg_page(seg, addr);
6956
6957 /*
6958 * Check to see if either of the pages addr or addr + delta
6959 * have advice set that prevents klustering (if MADV_RANDOM advice
6960 * is set for entire segment, or MADV_SEQUENTIAL is set and delta
6961 * is negative).
6962 */
6963 if (svd->advice == MADV_RANDOM ||
6964 svd->advice == MADV_SEQUENTIAL && delta < 0)
6965 return (-1);
6966 else if (svd->pageadvice && svd->vpage) {
6967 struct vpage *bvpp, *evpp;
6968
6969 bvpp = &svd->vpage[page];
6970 evpp = &svd->vpage[page + pd];
6971 if (VPP_ADVICE(bvpp) == MADV_RANDOM ||
6972 VPP_ADVICE(evpp) == MADV_SEQUENTIAL && delta < 0)
6973 return (-1);
6974 if (VPP_ADVICE(bvpp) != VPP_ADVICE(evpp) &&
6975 VPP_ADVICE(evpp) == MADV_RANDOM)
6976 return (-1);
6977 }
6978
6979 if (svd->type == MAP_SHARED)
6980 return (0); /* shared mapping - all ok */
6981
6982 if ((amp = svd->amp) == NULL)
6983 return (0); /* off original vnode */
6984
6985 page += svd->anon_index;
6986
6987 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
6988
6989 oap = anon_get_ptr(amp->ahp, page);
6990 ap = anon_get_ptr(amp->ahp, page + pd);
6991
6992 ANON_LOCK_EXIT(&amp->a_rwlock);
6993
6994 if ((oap == NULL && ap != NULL) || (oap != NULL && ap == NULL)) {
6995 return (-1); /* one with and one without an anon */
6996 }
6997
6998 if (oap == NULL) { /* implies that ap == NULL */
6999 return (0); /* off original vnode */
7000 }
7001
7002 /*
7003 * Now we know we have two anon pointers - check to
7004 * see if they happen to be properly allocated.
7005 */
7006
7007 /*
7008 * XXX We cheat here and don't lock the anon slots. We can't because
7009 * we may have been called from the anon layer which might already
7010 * have locked them. We are holding a refcnt on the slots so they
7011 * can't disappear. The worst that will happen is we'll get the wrong
7012 * names (vp, off) for the slots and make a poor klustering decision.
7013 */
7014 swap_xlate(ap, &vp1, &off1);
7015 swap_xlate(oap, &vp2, &off2);
7016
7017
7018 if (!fop_cmp(vp1, vp2, NULL) || off1 - off2 != delta)
7019 return (-1);
7020 return (0);
7021 }
7022
7023 /*
7024 * Synchronize primary storage cache with real object in virtual memory.
7025 *
7026 * XXX - Anonymous pages should not be sync'ed out at all.
7027 */
7028 static int
7029 segvn_sync(struct seg *seg, caddr_t addr, size_t len, int attr, uint_t flags)
7030 {
7031 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7032 struct vpage *vpp;
7033 page_t *pp;
7034 uoff_t offset;
7035 struct vnode *vp;
7036 uoff_t off;
7037 caddr_t eaddr;
7038 int bflags;
7039 int err = 0;
7040 int segtype;
7041 int pageprot;
7042 int prot;
7043 ulong_t anon_index;
7044 struct anon_map *amp;
7045 struct anon *ap;
7046 anon_sync_obj_t cookie;
7047
7048 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7049
7050 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7051
7052 if (svd->softlockcnt > 0) {
7053 /*
7054 * If this is shared segment non 0 softlockcnt
7055 * means locked pages are still in use.
7056 */
7057 if (svd->type == MAP_SHARED) {
7058 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7059 return (EAGAIN);
7060 }
7061
7062 /*
7063 * flush all pages from seg cache
7064 * otherwise we may deadlock in swap_putpage
7065 * for B_INVAL page (4175402).
7066 *
7067 * Even if we grab segvn WRITER's lock
7068 * here, there might be another thread which could've
7069 * successfully performed lookup/insert just before
7070 * we acquired the lock here. So, grabbing either
7071 * lock here is of not much use. Until we devise
7072 * a strategy at upper layers to solve the
7073 * synchronization issues completely, we expect
7074 * applications to handle this appropriately.
7075 */
7076 segvn_purge(seg);
7077 if (svd->softlockcnt > 0) {
7078 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7079 return (EAGAIN);
7080 }
7081 } else if (svd->type == MAP_SHARED && svd->amp != NULL &&
7082 svd->amp->a_softlockcnt > 0) {
7083 /*
7084 * Try to purge this amp's entries from pcache. It will
7085 * succeed only if other segments that share the amp have no
7086 * outstanding softlock's.
7087 */
7088 segvn_purge(seg);
7089 if (svd->amp->a_softlockcnt > 0 || svd->softlockcnt > 0) {
7090 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7091 return (EAGAIN);
7092 }
7093 }
7094
7095 vpp = svd->vpage;
7096 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7097 bflags = ((flags & MS_ASYNC) ? B_ASYNC : 0) |
7098 ((flags & MS_INVALIDATE) ? B_INVAL : 0);
7099
7100 if (attr) {
7101 pageprot = attr & ~(SHARED|PRIVATE);
7102 segtype = (attr & SHARED) ? MAP_SHARED : MAP_PRIVATE;
7103
7104 /*
7105 * We are done if the segment types don't match
7106 * or if we have segment level protections and
7107 * they don't match.
7108 */
7109 if (svd->type != segtype) {
7110 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7111 return (0);
7112 }
7113 if (vpp == NULL) {
7114 if (svd->prot != pageprot) {
7115 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7116 return (0);
7117 }
7118 prot = svd->prot;
7119 } else
7120 vpp = &svd->vpage[seg_page(seg, addr)];
7121
7122 } else if (svd->vp && svd->amp == NULL &&
7123 (flags & MS_INVALIDATE) == 0) {
7124
7125 /*
7126 * No attributes, no anonymous pages and MS_INVALIDATE flag
7127 * is not on, just use one big request.
7128 */
7129 err = fop_putpage(svd->vp, (offset_t)offset, len,
7130 bflags, svd->cred, NULL);
7131 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7132 return (err);
7133 }
7134
7135 if ((amp = svd->amp) != NULL)
7136 anon_index = svd->anon_index + seg_page(seg, addr);
7137
7138 for (eaddr = addr + len; addr < eaddr; addr += PAGESIZE) {
7139 ap = NULL;
7140 if (amp != NULL) {
7141 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7142 anon_array_enter(amp, anon_index, &cookie);
7143 ap = anon_get_ptr(amp->ahp, anon_index++);
7144 if (ap != NULL) {
7145 swap_xlate(ap, &vp, &off);
7146 } else {
7147 vp = svd->vp;
7148 off = offset;
7149 }
7150 anon_array_exit(&cookie);
7151 ANON_LOCK_EXIT(&amp->a_rwlock);
7152 } else {
7153 vp = svd->vp;
7154 off = offset;
7155 }
7156 offset += PAGESIZE;
7157
7158 if (vp == NULL) /* untouched zfod page */
7159 continue;
7160
7161 if (attr) {
7162 if (vpp) {
7163 prot = VPP_PROT(vpp);
7164 vpp++;
7165 }
7166 if (prot != pageprot) {
7167 continue;
7168 }
7169 }
7170
7171 /*
7172 * See if any of these pages are locked -- if so, then we
7173 * will have to truncate an invalidate request at the first
7174 * locked one. We don't need the page_struct_lock to test
7175 * as this is only advisory; even if we acquire it someone
7176 * might race in and lock the page after we unlock and before
7177 * we do the PUTPAGE, then PUTPAGE simply does nothing.
7178 */
7179 if (flags & MS_INVALIDATE) {
7180 if ((pp = page_lookup(&vp->v_object, off, SE_SHARED)) != NULL) {
7181 if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
7182 page_unlock(pp);
7183 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7184 return (EBUSY);
7185 }
7186 if (ap != NULL && pp->p_szc != 0 &&
7187 page_tryupgrade(pp)) {
7188 if (pp->p_lckcnt == 0 &&
7189 pp->p_cowcnt == 0) {
7190 /*
7191 * swapfs VN_DISPOSE() won't
7192 * invalidate large pages.
7193 * Attempt to demote.
7194 * XXX can't help it if it
7195 * fails. But for swapfs
7196 * pages it is no big deal.
7197 */
7198 (void) page_try_demote_pages(
7199 pp);
7200 }
7201 }
7202 page_unlock(pp);
7203 }
7204 } else if (svd->type == MAP_SHARED && amp != NULL) {
7205 /*
7206 * Avoid writing out to disk ISM's large pages
7207 * because segspt_free_pages() relies on NULL an_pvp
7208 * of anon slots of such pages.
7209 */
7210
7211 ASSERT(svd->vp == NULL);
7212 /*
7213 * swapfs uses page_lookup_nowait if not freeing or
7214 * invalidating and skips a page if
7215 * page_lookup_nowait returns NULL.
7216 */
7217 pp = page_lookup_nowait(&vp->v_object, off, SE_SHARED);
7218 if (pp == NULL) {
7219 continue;
7220 }
7221 if (pp->p_szc != 0) {
7222 page_unlock(pp);
7223 continue;
7224 }
7225
7226 /*
7227 * Note ISM pages are created large so (vp, off)'s
7228 * page cannot suddenly become large after we unlock
7229 * pp.
7230 */
7231 page_unlock(pp);
7232 }
7233 /*
7234 * XXX - Should ultimately try to kluster
7235 * calls to fop_putpage() for performance.
7236 */
7237 VN_HOLD(vp);
7238 err = fop_putpage(vp, (offset_t)off, PAGESIZE,
7239 (bflags | (IS_SWAPFSVP(vp) ? B_PAGE_NOWAIT : 0)),
7240 svd->cred, NULL);
7241
7242 VN_RELE(vp);
7243 if (err)
7244 break;
7245 }
7246 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7247 return (err);
7248 }
7249
7250 /*
7251 * Determine if we have data corresponding to pages in the
7252 * primary storage virtual memory cache (i.e., "in core").
7253 */
7254 static size_t
7255 segvn_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
7256 {
7257 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7258 struct vnode *vp, *avp;
7259 uoff_t offset, aoffset;
7260 size_t p, ep;
7261 int ret;
7262 struct vpage *vpp;
7263 page_t *pp;
7264 uint_t start;
7265 struct anon_map *amp; /* XXX - for locknest */
7266 struct anon *ap;
7267 uint_t attr;
7268 anon_sync_obj_t cookie;
7269
7270 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7271
7272 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7273 if (svd->amp == NULL && svd->vp == NULL) {
7274 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7275 bzero(vec, btopr(len));
7276 return (len); /* no anonymous pages created yet */
7277 }
7278
7279 p = seg_page(seg, addr);
7280 ep = seg_page(seg, addr + len);
7281 start = svd->vp ? SEG_PAGE_VNODEBACKED : 0;
7282
7283 amp = svd->amp;
7284 for (; p < ep; p++, addr += PAGESIZE) {
7285 vpp = (svd->vpage) ? &svd->vpage[p]: NULL;
7286 ret = start;
7287 ap = NULL;
7288 avp = NULL;
7289 /* Grab the vnode/offset for the anon slot */
7290 if (amp != NULL) {
7291 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7292 anon_array_enter(amp, svd->anon_index + p, &cookie);
7293 ap = anon_get_ptr(amp->ahp, svd->anon_index + p);
7294 if (ap != NULL) {
7295 swap_xlate(ap, &avp, &aoffset);
7296 }
7297 anon_array_exit(&cookie);
7298 ANON_LOCK_EXIT(&amp->a_rwlock);
7299 }
7300 if ((avp != NULL) && page_exists(&avp->v_object, aoffset)) {
7301 /* A page exists for the anon slot */
7302 ret |= SEG_PAGE_INCORE;
7303
7304 /*
7305 * If page is mapped and writable
7306 */
7307 attr = 0;
7308 if ((hat_getattr(seg->s_as->a_hat, addr,
7309 &attr) != -1) && (attr & PROT_WRITE)) {
7310 ret |= SEG_PAGE_ANON;
7311 }
7312 /*
7313 * Don't get page_struct lock for lckcnt and cowcnt,
7314 * since this is purely advisory.
7315 */
7316 if ((pp = page_lookup_nowait(&avp->v_object,
7317 aoffset,
7318 SE_SHARED)) != NULL) {
7319 if (pp->p_lckcnt)
7320 ret |= SEG_PAGE_SOFTLOCK;
7321 if (pp->p_cowcnt)
7322 ret |= SEG_PAGE_HASCOW;
7323 page_unlock(pp);
7324 }
7325 }
7326
7327 /* Gather vnode statistics */
7328 vp = svd->vp;
7329 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7330
7331 if (vp != NULL) {
7332 /*
7333 * Try to obtain a "shared" lock on the page
7334 * without blocking. If this fails, determine
7335 * if the page is in memory.
7336 */
7337 pp = page_lookup_nowait(&vp->v_object, offset,
7338 SE_SHARED);
7339 if ((pp == NULL) && (page_exists(&vp->v_object, offset))) {
7340 /* Page is incore, and is named */
7341 ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7342 }
7343 /*
7344 * Don't get page_struct lock for lckcnt and cowcnt,
7345 * since this is purely advisory.
7346 */
7347 if (pp != NULL) {
7348 ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7349 if (pp->p_lckcnt)
7350 ret |= SEG_PAGE_SOFTLOCK;
7351 if (pp->p_cowcnt)
7352 ret |= SEG_PAGE_HASCOW;
7353 page_unlock(pp);
7354 }
7355 }
7356
7357 /* Gather virtual page information */
7358 if (vpp) {
7359 if (VPP_ISPPLOCK(vpp))
7360 ret |= SEG_PAGE_LOCKED;
7361 vpp++;
7362 }
7363
7364 *vec++ = (char)ret;
7365 }
7366 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7367 return (len);
7368 }
7369
7370 /*
7371 * Statement for p_cowcnts/p_lckcnts.
7372 *
7373 * p_cowcnt is updated while mlock/munlocking MAP_PRIVATE and PROT_WRITE region
7374 * irrespective of the following factors or anything else:
7375 *
7376 * (1) anon slots are populated or not
7377 * (2) cow is broken or not
7378 * (3) refcnt on ap is 1 or greater than 1
7379 *
7380 * If it's not MAP_PRIVATE and PROT_WRITE, p_lckcnt is updated during mlock
7381 * and munlock.
7382 *
7383 *
7384 * Handling p_cowcnts/p_lckcnts during copy-on-write fault:
7385 *
7386 * if vpage has PROT_WRITE
7387 * transfer cowcnt on the oldpage -> cowcnt on the newpage
7388 * else
7389 * transfer lckcnt on the oldpage -> lckcnt on the newpage
7390 *
7391 * During copy-on-write, decrement p_cowcnt on the oldpage and increment
7392 * p_cowcnt on the newpage *if* the corresponding vpage has PROT_WRITE.
7393 *
7394 * We may also break COW if softlocking on read access in the physio case.
7395 * In this case, vpage may not have PROT_WRITE. So, we need to decrement
7396 * p_lckcnt on the oldpage and increment p_lckcnt on the newpage *if* the
7397 * vpage doesn't have PROT_WRITE.
7398 *
7399 *
7400 * Handling p_cowcnts/p_lckcnts during mprotect on mlocked region:
7401 *
7402 * If a MAP_PRIVATE region loses PROT_WRITE, we decrement p_cowcnt and
7403 * increment p_lckcnt by calling page_subclaim() which takes care of
7404 * availrmem accounting and p_lckcnt overflow.
7405 *
7406 * If a MAP_PRIVATE region gains PROT_WRITE, we decrement p_lckcnt and
7407 * increment p_cowcnt by calling page_addclaim() which takes care of
7408 * availrmem availability and p_cowcnt overflow.
7409 */
7410
7411 /*
7412 * Lock down (or unlock) pages mapped by this segment.
7413 *
7414 * XXX only creates PAGESIZE pages if anon slots are not initialized.
7415 * At fault time they will be relocated into larger pages.
7416 */
7417 static int
7418 segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
7419 int attr, int op, ulong_t *lockmap, size_t pos)
7420 {
7421 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7422 struct vpage *vpp;
7423 struct vpage *evp;
7424 page_t *pp;
7425 uoff_t offset;
7426 uoff_t off;
7427 int segtype;
7428 int pageprot;
7429 int claim;
7430 struct vnode *vp;
7431 ulong_t anon_index;
7432 struct anon_map *amp;
7433 struct anon *ap;
7434 struct vattr va;
7435 anon_sync_obj_t cookie;
7436 struct kshmid *sp = NULL;
7437 struct proc *p = curproc;
7438 kproject_t *proj = NULL;
7439 int chargeproc = 1;
7440 size_t locked_bytes = 0;
7441 size_t unlocked_bytes = 0;
7442 int err = 0;
7443
7444 /*
7445 * Hold write lock on address space because may split or concatenate
7446 * segments
7447 */
7448 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7449
7450 /*
7451 * If this is a shm, use shm's project and zone, else use
7452 * project and zone of calling process
7453 */
7454
7455 /* Determine if this segment backs a sysV shm */
7456 if (svd->amp != NULL && svd->amp->a_sp != NULL) {
7457 ASSERT(svd->type == MAP_SHARED);
7458 ASSERT(svd->tr_state == SEGVN_TR_OFF);
7459 sp = svd->amp->a_sp;
7460 proj = sp->shm_perm.ipc_proj;
7461 chargeproc = 0;
7462 }
7463
7464 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7465 if (attr) {
7466 pageprot = attr & ~(SHARED|PRIVATE);
7467 segtype = attr & SHARED ? MAP_SHARED : MAP_PRIVATE;
7468
7469 /*
7470 * We are done if the segment types don't match
7471 * or if we have segment level protections and
7472 * they don't match.
7473 */
7474 if (svd->type != segtype) {
7475 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7476 return (0);
7477 }
7478 if (svd->pageprot == 0 && svd->prot != pageprot) {
7479 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7480 return (0);
7481 }
7482 }
7483
7484 if (op == MC_LOCK) {
7485 if (svd->tr_state == SEGVN_TR_INIT) {
7486 svd->tr_state = SEGVN_TR_OFF;
7487 } else if (svd->tr_state == SEGVN_TR_ON) {
7488 ASSERT(svd->amp != NULL);
7489 segvn_textunrepl(seg, 0);
7490 ASSERT(svd->amp == NULL &&
7491 svd->tr_state == SEGVN_TR_OFF);
7492 }
7493 }
7494
7495 /*
7496 * If we're locking, then we must create a vpage structure if
7497 * none exists. If we're unlocking, then check to see if there
7498 * is a vpage -- if not, then we could not have locked anything.
7499 */
7500
7501 if ((vpp = svd->vpage) == NULL) {
7502 if (op == MC_LOCK) {
7503 segvn_vpage(seg);
7504 if (svd->vpage == NULL) {
7505 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7506 return (ENOMEM);
7507 }
7508 } else {
7509 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7510 return (0);
7511 }
7512 }
7513
7514 /*
7515 * The anonymous data vector (i.e., previously
7516 * unreferenced mapping to swap space) can be allocated
7517 * by lazily testing for its existence.
7518 */
7519 if (op == MC_LOCK && svd->amp == NULL && svd->vp == NULL) {
7520 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
7521 svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
7522 svd->amp->a_szc = seg->s_szc;
7523 }
7524
7525 if ((amp = svd->amp) != NULL) {
7526 anon_index = svd->anon_index + seg_page(seg, addr);
7527 }
7528
7529 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7530 evp = &svd->vpage[seg_page(seg, addr + len)];
7531
7532 if (sp != NULL)
7533 mutex_enter(&sp->shm_mlock);
7534
7535 /* determine number of unlocked bytes in range for lock operation */
7536 if (op == MC_LOCK) {
7537
7538 if (sp == NULL) {
7539 for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7540 vpp++) {
7541 if (!VPP_ISPPLOCK(vpp))
7542 unlocked_bytes += PAGESIZE;
7543 }
7544 } else {
7545 ulong_t i_idx, i_edx;
7546 anon_sync_obj_t i_cookie;
7547 struct anon *i_ap;
7548 struct vnode *i_vp;
7549 uoff_t i_off;
7550
7551 /* Only count sysV pages once for locked memory */
7552 i_edx = svd->anon_index + seg_page(seg, addr + len);
7553 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7554 for (i_idx = anon_index; i_idx < i_edx; i_idx++) {
7555 anon_array_enter(amp, i_idx, &i_cookie);
7556 i_ap = anon_get_ptr(amp->ahp, i_idx);
7557 if (i_ap == NULL) {
7558 unlocked_bytes += PAGESIZE;
7559 anon_array_exit(&i_cookie);
7560 continue;
7561 }
7562 swap_xlate(i_ap, &i_vp, &i_off);
7563 anon_array_exit(&i_cookie);
7564 pp = page_lookup(&i_vp->v_object, i_off,
7565 SE_SHARED);
7566 if (pp == NULL) {
7567 unlocked_bytes += PAGESIZE;
7568 continue;
7569 } else if (pp->p_lckcnt == 0)
7570 unlocked_bytes += PAGESIZE;
7571 page_unlock(pp);
7572 }
7573 ANON_LOCK_EXIT(&amp->a_rwlock);
7574 }
7575
7576 mutex_enter(&p->p_lock);
7577 err = rctl_incr_locked_mem(p, proj, unlocked_bytes,
7578 chargeproc);
7579 mutex_exit(&p->p_lock);
7580
7581 if (err) {
7582 if (sp != NULL)
7583 mutex_exit(&sp->shm_mlock);
7584 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7585 return (err);
7586 }
7587 }
7588 /*
7589 * Loop over all pages in the range. Process if we're locking and
7590 * page has not already been locked in this mapping; or if we're
7591 * unlocking and the page has been locked.
7592 */
7593 for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7594 vpp++, pos++, addr += PAGESIZE, offset += PAGESIZE, anon_index++) {
7595 if ((attr == 0 || VPP_PROT(vpp) == pageprot) &&
7596 ((op == MC_LOCK && !VPP_ISPPLOCK(vpp)) ||
7597 (op == MC_UNLOCK && VPP_ISPPLOCK(vpp)))) {
7598
7599 if (amp != NULL)
7600 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7601 /*
7602 * If this isn't a MAP_NORESERVE segment and
7603 * we're locking, allocate anon slots if they
7604 * don't exist. The page is brought in later on.
7605 */
7606 if (op == MC_LOCK && svd->vp == NULL &&
7607 ((svd->flags & MAP_NORESERVE) == 0) &&
7608 amp != NULL &&
7609 ((ap = anon_get_ptr(amp->ahp, anon_index))
7610 == NULL)) {
7611 anon_array_enter(amp, anon_index, &cookie);
7612
7613 if ((ap = anon_get_ptr(amp->ahp,
7614 anon_index)) == NULL) {
7615 pp = anon_zero(seg, addr, &ap,
7616 svd->cred);
7617 if (pp == NULL) {
7618 anon_array_exit(&cookie);
7619 ANON_LOCK_EXIT(&amp->a_rwlock);
7620 err = ENOMEM;
7621 goto out;
7622 }
7623 ASSERT(anon_get_ptr(amp->ahp,
7624 anon_index) == NULL);
7625 (void) anon_set_ptr(amp->ahp,
7626 anon_index, ap, ANON_SLEEP);
7627 page_unlock(pp);
7628 }
7629 anon_array_exit(&cookie);
7630 }
7631
7632 /*
7633 * Get name for page, accounting for
7634 * existence of private copy.
7635 */
7636 ap = NULL;
7637 if (amp != NULL) {
7638 anon_array_enter(amp, anon_index, &cookie);
7639 ap = anon_get_ptr(amp->ahp, anon_index);
7640 if (ap != NULL) {
7641 swap_xlate(ap, &vp, &off);
7642 } else {
7643 if (svd->vp == NULL &&
7644 (svd->flags & MAP_NORESERVE)) {
7645 anon_array_exit(&cookie);
7646 ANON_LOCK_EXIT(&amp->a_rwlock);
7647 continue;
7648 }
7649 vp = svd->vp;
7650 off = offset;
7651 }
7652 if (op != MC_LOCK || ap == NULL) {
7653 anon_array_exit(&cookie);
7654 ANON_LOCK_EXIT(&amp->a_rwlock);
7655 }
7656 } else {
7657 vp = svd->vp;
7658 off = offset;
7659 }
7660
7661 /*
7662 * Get page frame. It's ok if the page is
7663 * not available when we're unlocking, as this
7664 * may simply mean that a page we locked got
7665 * truncated out of existence after we locked it.
7666 *
7667 * Invoke fop_getpage() to obtain the page struct
7668 * since we may need to read it from disk if its
7669 * been paged out.
7670 */
7671 if (op != MC_LOCK)
7672 pp = page_lookup(&vp->v_object, off,
7673 SE_SHARED);
7674 else {
7675 page_t *pl[1 + 1];
7676 int error;
7677
7678 ASSERT(vp != NULL);
7679
7680 error = fop_getpage(vp, (offset_t)off, PAGESIZE,
7681 (uint_t *)NULL, pl, PAGESIZE, seg, addr,
7682 S_OTHER, svd->cred, NULL);
7683
7684 if (error && ap != NULL) {
7685 anon_array_exit(&cookie);
7686 ANON_LOCK_EXIT(&amp->a_rwlock);
7687 }
7688
7689 /*
7690 * If the error is EDEADLK then we must bounce
7691 * up and drop all vm subsystem locks and then
7692 * retry the operation later
7693 * This behavior is a temporary measure because
7694 * ufs/sds logging is badly designed and will
7695 * deadlock if we don't allow this bounce to
7696 * happen. The real solution is to re-design
7697 * the logging code to work properly. See bug
7698 * 4125102 for details of the problem.
7699 */
7700 if (error == EDEADLK) {
7701 err = error;
7702 goto out;
7703 }
7704 /*
7705 * Quit if we fail to fault in the page. Treat
7706 * the failure as an error, unless the addr
7707 * is mapped beyond the end of a file.
7708 */
7709 if (error && svd->vp) {
7710 va.va_mask = VATTR_SIZE;
7711 if (fop_getattr(svd->vp, &va, 0,
7712 svd->cred, NULL) != 0) {
7713 err = EIO;
7714 goto out;
7715 }
7716 if (btopr(va.va_size) >=
7717 btopr(off + 1)) {
7718 err = EIO;
7719 goto out;
7720 }
7721 goto out;
7722
7723 } else if (error) {
7724 err = EIO;
7725 goto out;
7726 }
7727 pp = pl[0];
7728 ASSERT(pp != NULL);
7729 }
7730
7731 /*
7732 * See Statement at the beginning of this routine.
7733 *
7734 * claim is always set if MAP_PRIVATE and PROT_WRITE
7735 * irrespective of following factors:
7736 *
7737 * (1) anon slots are populated or not
7738 * (2) cow is broken or not
7739 * (3) refcnt on ap is 1 or greater than 1
7740 *
7741 * See 4140683 for details
7742 */
7743 claim = ((VPP_PROT(vpp) & PROT_WRITE) &&
7744 (svd->type == MAP_PRIVATE));
7745
7746 /*
7747 * Perform page-level operation appropriate to
7748 * operation. If locking, undo the SOFTLOCK
7749 * performed to bring the page into memory
7750 * after setting the lock. If unlocking,
7751 * and no page was found, account for the claim
7752 * separately.
7753 */
7754 if (op == MC_LOCK) {
7755 int ret = 1; /* Assume success */
7756
7757 ASSERT(!VPP_ISPPLOCK(vpp));
7758
7759 ret = page_pp_lock(pp, claim, 0);
7760 if (ap != NULL) {
7761 if (ap->an_pvp != NULL) {
7762 anon_swap_free(ap, pp);
7763 }
7764 anon_array_exit(&cookie);
7765 ANON_LOCK_EXIT(&amp->a_rwlock);
7766 }
7767 if (ret == 0) {
7768 /* locking page failed */
7769 page_unlock(pp);
7770 err = EAGAIN;
7771 goto out;
7772 }
7773 VPP_SETPPLOCK(vpp);
7774 if (sp != NULL) {
7775 if (pp->p_lckcnt == 1)
7776 locked_bytes += PAGESIZE;
7777 } else
7778 locked_bytes += PAGESIZE;
7779
7780 if (lockmap != NULL)
7781 BT_SET(lockmap, pos);
7782
7783 page_unlock(pp);
7784 } else {
7785 ASSERT(VPP_ISPPLOCK(vpp));
7786 if (pp != NULL) {
7787 /* sysV pages should be locked */
7788 ASSERT(sp == NULL || pp->p_lckcnt > 0);
7789 page_pp_unlock(pp, claim, 0);
7790 if (sp != NULL) {
7791 if (pp->p_lckcnt == 0)
7792 unlocked_bytes
7793 += PAGESIZE;
7794 } else
7795 unlocked_bytes += PAGESIZE;
7796 page_unlock(pp);
7797 } else {
7798 ASSERT(sp == NULL);
7799 unlocked_bytes += PAGESIZE;
7800 }
7801 VPP_CLRPPLOCK(vpp);
7802 }
7803 }
7804 }
7805 out:
7806 if (op == MC_LOCK) {
7807 /* Credit back bytes that did not get locked */
7808 if ((unlocked_bytes - locked_bytes) > 0) {
7809 if (proj == NULL)
7810 mutex_enter(&p->p_lock);
7811 rctl_decr_locked_mem(p, proj,
7812 (unlocked_bytes - locked_bytes), chargeproc);
7813 if (proj == NULL)
7814 mutex_exit(&p->p_lock);
7815 }
7816
7817 } else {
7818 /* Account bytes that were unlocked */
7819 if (unlocked_bytes > 0) {
7820 if (proj == NULL)
7821 mutex_enter(&p->p_lock);
7822 rctl_decr_locked_mem(p, proj, unlocked_bytes,
7823 chargeproc);
7824 if (proj == NULL)
7825 mutex_exit(&p->p_lock);
7826 }
7827 }
7828 if (sp != NULL)
7829 mutex_exit(&sp->shm_mlock);
7830 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7831
7832 return (err);
7833 }
7834
7835 /*
7836 * Set advice from user for specified pages
7837 * There are 10 types of advice:
7838 * MADV_NORMAL - Normal (default) behavior (whatever that is)
7839 * MADV_RANDOM - Random page references
7840 * do not allow readahead or 'klustering'
7841 * MADV_SEQUENTIAL - Sequential page references
7842 * Pages previous to the one currently being
7843 * accessed (determined by fault) are 'not needed'
7844 * and are freed immediately
7845 * MADV_WILLNEED - Pages are likely to be used (fault ahead in mctl)
7846 * MADV_DONTNEED - Pages are not needed (synced out in mctl)
7847 * MADV_FREE - Contents can be discarded
7848 * MADV_ACCESS_DEFAULT- Default access
7849 * MADV_ACCESS_LWP - Next LWP will access heavily
7850 * MADV_ACCESS_MANY- Many LWPs or processes will access heavily
7851 * MADV_PURGE - Contents will be immediately discarded
7852 */
7853 static int
7854 segvn_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
7855 {
7856 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7857 size_t page;
7858 int err = 0;
7859 int already_set;
7860 struct anon_map *amp;
7861 ulong_t anon_index;
7862 struct seg *next;
7863 lgrp_mem_policy_t policy;
7864 struct seg *prev;
7865 struct vnode *vp;
7866
7867 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7868
7869 /*
7870 * In case of MADV_FREE/MADV_PURGE, we won't be modifying any segment
7871 * private data structures; so, we only need to grab READER's lock
7872 */
7873 if (behav != MADV_FREE && behav != MADV_PURGE) {
7874 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7875 if (svd->tr_state != SEGVN_TR_OFF) {
7876 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7877 return (0);
7878 }
7879 } else {
7880 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7881 }
7882
7883 /*
7884 * Large pages are assumed to be only turned on when accesses to the
7885 * segment's address range have spatial and temporal locality. That
7886 * justifies ignoring MADV_SEQUENTIAL for large page segments.
7887 * Also, ignore advice affecting lgroup memory allocation
7888 * if don't need to do lgroup optimizations on this system
7889 */
7890
7891 if ((behav == MADV_SEQUENTIAL &&
7892 (seg->s_szc != 0 || HAT_IS_REGION_COOKIE_VALID(svd->rcookie))) ||
7893 (!lgrp_optimizations() && (behav == MADV_ACCESS_DEFAULT ||
7894 behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY))) {
7895 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7896 return (0);
7897 }
7898
7899 if (behav == MADV_SEQUENTIAL || behav == MADV_ACCESS_DEFAULT ||
7900 behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY) {
7901 /*
7902 * Since we are going to unload hat mappings
7903 * we first have to flush the cache. Otherwise
7904 * this might lead to system panic if another
7905 * thread is doing physio on the range whose
7906 * mappings are unloaded by madvise(3C).
7907 */
7908 if (svd->softlockcnt > 0) {
7909 /*
7910 * If this is shared segment non 0 softlockcnt
7911 * means locked pages are still in use.
7912 */
7913 if (svd->type == MAP_SHARED) {
7914 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7915 return (EAGAIN);
7916 }
7917 /*
7918 * Since we do have the segvn writers lock
7919 * nobody can fill the cache with entries
7920 * belonging to this seg during the purge.
7921 * The flush either succeeds or we still
7922 * have pending I/Os. In the later case,
7923 * madvise(3C) fails.
7924 */
7925 segvn_purge(seg);
7926 if (svd->softlockcnt > 0) {
7927 /*
7928 * Since madvise(3C) is advisory and
7929 * it's not part of UNIX98, madvise(3C)
7930 * failure here doesn't cause any hardship.
7931 * Note that we don't block in "as" layer.
7932 */
7933 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7934 return (EAGAIN);
7935 }
7936 } else if (svd->type == MAP_SHARED && svd->amp != NULL &&
7937 svd->amp->a_softlockcnt > 0) {
7938 /*
7939 * Try to purge this amp's entries from pcache. It
7940 * will succeed only if other segments that share the
7941 * amp have no outstanding softlock's.
7942 */
7943 segvn_purge(seg);
7944 }
7945 }
7946
7947 amp = svd->amp;
7948 vp = svd->vp;
7949 if (behav == MADV_FREE || behav == MADV_PURGE) {
7950 pgcnt_t purged;
7951
7952 if (behav == MADV_FREE && (vp != NULL || amp == NULL)) {
7953 /*
7954 * MADV_FREE is not supported for segments with an
7955 * underlying object; if anonmap is NULL, anon slots
7956 * are not yet populated and there is nothing for us
7957 * to do. As MADV_FREE is advisory, we don't return an
7958 * error in either case.
7959 */
7960 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7961 return (0);
7962 }
7963
7964 if (amp == NULL) {
7965 /*
7966 * If we're here with a NULL anonmap, it's because we
7967 * are doing a MADV_PURGE. We have nothing to do, but
7968 * because MADV_PURGE isn't merely advisory, we return
7969 * an error in this case.
7970 */
7971 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7972 return (EBUSY);
7973 }
7974
7975 segvn_purge(seg);
7976
7977 page = seg_page(seg, addr);
7978 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7979 err = anon_disclaim(amp,
7980 svd->anon_index + page, len, behav, &purged);
7981
7982 if (purged != 0 && (svd->flags & MAP_NORESERVE)) {
7983 /*
7984 * If we purged pages on a MAP_NORESERVE mapping, we
7985 * need to be sure to now unreserve our reserved swap.
7986 * (We use the atomic operations to manipulate our
7987 * segment and address space counters because we only
7988 * have the corresponding locks held as reader, not
7989 * writer.)
7990 */
7991 ssize_t bytes = ptob(purged);
7992
7993 anon_unresv_zone(bytes, seg->s_as->a_proc->p_zone);
7994 atomic_add_long(&svd->swresv, -bytes);
7995 atomic_add_long(&seg->s_as->a_resvsize, -bytes);
7996 }
7997
7998 ANON_LOCK_EXIT(&amp->a_rwlock);
7999 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8000
8001 /*
8002 * MADV_PURGE and MADV_FREE differ in their return semantics:
8003 * because MADV_PURGE is designed to be bug-for-bug compatible
8004 * with its clumsy Linux forebear, it will fail where MADV_FREE
8005 * does not.
8006 */
8007 return (behav == MADV_PURGE ? err : 0);
8008 }
8009
8010 /*
8011 * If advice is to be applied to entire segment,
8012 * use advice field in seg_data structure
8013 * otherwise use appropriate vpage entry.
8014 */
8015 if ((addr == seg->s_base) && (len == seg->s_size)) {
8016 switch (behav) {
8017 case MADV_ACCESS_LWP:
8018 case MADV_ACCESS_MANY:
8019 case MADV_ACCESS_DEFAULT:
8020 /*
8021 * Set memory allocation policy for this segment
8022 */
8023 policy = lgrp_madv_to_policy(behav, len, svd->type);
8024 if (svd->type == MAP_SHARED)
8025 already_set = lgrp_shm_policy_set(policy, amp,
8026 svd->anon_index, vp, svd->offset, len);
8027 else {
8028 /*
8029 * For private memory, need writers lock on
8030 * address space because the segment may be
8031 * split or concatenated when changing policy
8032 */
8033 if (AS_READ_HELD(seg->s_as)) {
8034 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8035 return (IE_RETRY);
8036 }
8037
8038 already_set = lgrp_privm_policy_set(policy,
8039 &svd->policy_info, len);
8040 }
8041
8042 /*
8043 * If policy set already and it shouldn't be reapplied,
8044 * don't do anything.
8045 */
8046 if (already_set &&
8047 !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8048 break;
8049
8050 /*
8051 * Mark any existing pages in given range for
8052 * migration
8053 */
8054 page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8055 &vp->v_object, svd->offset, 1);
8056
8057 /*
8058 * If same policy set already or this is a shared
8059 * memory segment, don't need to try to concatenate
8060 * segment with adjacent ones.
8061 */
8062 if (already_set || svd->type == MAP_SHARED)
8063 break;
8064
8065 /*
8066 * Try to concatenate this segment with previous
8067 * one and next one, since we changed policy for
8068 * this one and it may be compatible with adjacent
8069 * ones now.
8070 */
8071 prev = AS_SEGPREV(seg->s_as, seg);
8072 next = AS_SEGNEXT(seg->s_as, seg);
8073
8074 if (next && next->s_ops == &segvn_ops &&
8075 addr + len == next->s_base)
8076 (void) segvn_concat(seg, next, 1);
8077
8078 if (prev && prev->s_ops == &segvn_ops &&
8079 addr == prev->s_base + prev->s_size) {
8080 /*
8081 * Drop lock for private data of current
8082 * segment before concatenating (deleting) it
8083 * and return IE_REATTACH to tell as_ctl() that
8084 * current segment has changed
8085 */
8086 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8087 if (!segvn_concat(prev, seg, 1))
8088 err = IE_REATTACH;
8089
8090 return (err);
8091 }
8092 break;
8093
8094 case MADV_SEQUENTIAL:
8095 /*
8096 * unloading mapping guarantees
8097 * detection in segvn_fault
8098 */
8099 ASSERT(seg->s_szc == 0);
8100 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8101 hat_unload(seg->s_as->a_hat, addr, len,
8102 HAT_UNLOAD);
8103 /* FALLTHROUGH */
8104 case MADV_NORMAL:
8105 case MADV_RANDOM:
8106 svd->advice = (uchar_t)behav;
8107 svd->pageadvice = 0;
8108 break;
8109 case MADV_WILLNEED: /* handled in memcntl */
8110 case MADV_DONTNEED: /* handled in memcntl */
8111 case MADV_FREE: /* handled above */
8112 case MADV_PURGE: /* handled above */
8113 break;
8114 default:
8115 err = EINVAL;
8116 }
8117 } else {
8118 caddr_t eaddr;
8119 struct seg *new_seg;
8120 struct segvn_data *new_svd;
8121 uoff_t off;
8122 caddr_t oldeaddr;
8123
8124 page = seg_page(seg, addr);
8125
8126 segvn_vpage(seg);
8127 if (svd->vpage == NULL) {
8128 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8129 return (ENOMEM);
8130 }
8131
8132 switch (behav) {
8133 struct vpage *bvpp, *evpp;
8134
8135 case MADV_ACCESS_LWP:
8136 case MADV_ACCESS_MANY:
8137 case MADV_ACCESS_DEFAULT:
8138 /*
8139 * Set memory allocation policy for portion of this
8140 * segment
8141 */
8142
8143 /*
8144 * Align address and length of advice to page
8145 * boundaries for large pages
8146 */
8147 if (seg->s_szc != 0) {
8148 size_t pgsz;
8149
8150 pgsz = page_get_pagesize(seg->s_szc);
8151 addr = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
8152 len = P2ROUNDUP(len, pgsz);
8153 }
8154
8155 /*
8156 * Check to see whether policy is set already
8157 */
8158 policy = lgrp_madv_to_policy(behav, len, svd->type);
8159
8160 anon_index = svd->anon_index + page;
8161 off = svd->offset + (uintptr_t)(addr - seg->s_base);
8162
8163 if (svd->type == MAP_SHARED)
8164 already_set = lgrp_shm_policy_set(policy, amp,
8165 anon_index, vp, off, len);
8166 else
8167 already_set =
8168 (policy == svd->policy_info.mem_policy);
8169
8170 /*
8171 * If policy set already and it shouldn't be reapplied,
8172 * don't do anything.
8173 */
8174 if (already_set &&
8175 !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8176 break;
8177
8178 /*
8179 * For private memory, need writers lock on
8180 * address space because the segment may be
8181 * split or concatenated when changing policy
8182 */
8183 if (svd->type == MAP_PRIVATE &&
8184 AS_READ_HELD(seg->s_as)) {
8185 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8186 return (IE_RETRY);
8187 }
8188
8189 /*
8190 * Mark any existing pages in given range for
8191 * migration
8192 */
8193 page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8194 &vp->v_object, svd->offset, 1);
8195
8196 /*
8197 * Don't need to try to split or concatenate
8198 * segments, since policy is same or this is a shared
8199 * memory segment
8200 */
8201 if (already_set || svd->type == MAP_SHARED)
8202 break;
8203
8204 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
8205 ASSERT(svd->amp == NULL);
8206 ASSERT(svd->tr_state == SEGVN_TR_OFF);
8207 ASSERT(svd->softlockcnt == 0);
8208 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
8209 HAT_REGION_TEXT);
8210 svd->rcookie = HAT_INVALID_REGION_COOKIE;
8211 }
8212
8213 /*
8214 * Split off new segment if advice only applies to a
8215 * portion of existing segment starting in middle
8216 */
8217 new_seg = NULL;
8218 eaddr = addr + len;
8219 oldeaddr = seg->s_base + seg->s_size;
8220 if (addr > seg->s_base) {
8221 /*
8222 * Must flush I/O page cache
8223 * before splitting segment
8224 */
8225 if (svd->softlockcnt > 0)
8226 segvn_purge(seg);
8227
8228 /*
8229 * Split segment and return IE_REATTACH to tell
8230 * as_ctl() that current segment changed
8231 */
8232 new_seg = segvn_split_seg(seg, addr);
8233 new_svd = (struct segvn_data *)new_seg->s_data;
8234 err = IE_REATTACH;
8235
8236 /*
8237 * If new segment ends where old one
8238 * did, try to concatenate the new
8239 * segment with next one.
8240 */
8241 if (eaddr == oldeaddr) {
8242 /*
8243 * Set policy for new segment
8244 */
8245 (void) lgrp_privm_policy_set(policy,
8246 &new_svd->policy_info,
8247 new_seg->s_size);
8248
8249 next = AS_SEGNEXT(new_seg->s_as,
8250 new_seg);
8251
8252 if (next &&
8253 next->s_ops == &segvn_ops &&
8254 eaddr == next->s_base)
8255 (void) segvn_concat(new_seg,
8256 next, 1);
8257 }
8258 }
8259
8260 /*
8261 * Split off end of existing segment if advice only
8262 * applies to a portion of segment ending before
8263 * end of the existing segment
8264 */
8265 if (eaddr < oldeaddr) {
8266 /*
8267 * Must flush I/O page cache
8268 * before splitting segment
8269 */
8270 if (svd->softlockcnt > 0)
8271 segvn_purge(seg);
8272
8273 /*
8274 * If beginning of old segment was already
8275 * split off, use new segment to split end off
8276 * from.
8277 */
8278 if (new_seg != NULL && new_seg != seg) {
8279 /*
8280 * Split segment
8281 */
8282 (void) segvn_split_seg(new_seg, eaddr);
8283
8284 /*
8285 * Set policy for new segment
8286 */
8287 (void) lgrp_privm_policy_set(policy,
8288 &new_svd->policy_info,
8289 new_seg->s_size);
8290 } else {
8291 /*
8292 * Split segment and return IE_REATTACH
8293 * to tell as_ctl() that current
8294 * segment changed
8295 */
8296 (void) segvn_split_seg(seg, eaddr);
8297 err = IE_REATTACH;
8298
8299 (void) lgrp_privm_policy_set(policy,
8300 &svd->policy_info, seg->s_size);
8301
8302 /*
8303 * If new segment starts where old one
8304 * did, try to concatenate it with
8305 * previous segment.
8306 */
8307 if (addr == seg->s_base) {
8308 prev = AS_SEGPREV(seg->s_as,
8309 seg);
8310
8311 /*
8312 * Drop lock for private data
8313 * of current segment before
8314 * concatenating (deleting) it
8315 */
8316 if (prev &&
8317 prev->s_ops ==
8318 &segvn_ops &&
8319 addr == prev->s_base +
8320 prev->s_size) {
8321 SEGVN_LOCK_EXIT(
8322 seg->s_as,
8323 &svd->lock);
8324 (void) segvn_concat(
8325 prev, seg, 1);
8326 return (err);
8327 }
8328 }
8329 }
8330 }
8331 break;
8332 case MADV_SEQUENTIAL:
8333 ASSERT(seg->s_szc == 0);
8334 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8335 hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
8336 /* FALLTHROUGH */
8337 case MADV_NORMAL:
8338 case MADV_RANDOM:
8339 bvpp = &svd->vpage[page];
8340 evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8341 for (; bvpp < evpp; bvpp++)
8342 VPP_SETADVICE(bvpp, behav);
8343 svd->advice = MADV_NORMAL;
8344 break;
8345 case MADV_WILLNEED: /* handled in memcntl */
8346 case MADV_DONTNEED: /* handled in memcntl */
8347 case MADV_FREE: /* handled above */
8348 case MADV_PURGE: /* handled above */
8349 break;
8350 default:
8351 err = EINVAL;
8352 }
8353 }
8354 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8355 return (err);
8356 }
8357
8358 /*
8359 * There is one kind of inheritance that can be specified for pages:
8360 *
8361 * SEGP_INH_ZERO - Pages should be zeroed in the child
8362 */
8363 static int
8364 segvn_inherit(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
8365 {
8366 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8367 struct vpage *bvpp, *evpp;
8368 size_t page;
8369 int ret = 0;
8370
8371 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
8372
8373 /* Can't support something we don't know about */
8374 if (behav != SEGP_INH_ZERO)
8375 return (ENOTSUP);
8376
8377 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
8378
8379 /*
8380 * This must be a straightforward anonymous segment that is mapped
8381 * privately and is not backed by a vnode.
8382 */
8383 if (svd->tr_state != SEGVN_TR_OFF ||
8384 svd->type != MAP_PRIVATE ||
8385 svd->vp != NULL) {
8386 ret = EINVAL;
8387 goto out;
8388 }
8389
8390 /*
8391 * If the entire segment has been marked as inherit zero, then no reason
8392 * to do anything else.
8393 */
8394 if (svd->svn_inz == SEGVN_INZ_ALL) {
8395 ret = 0;
8396 goto out;
8397 }
8398
8399 /*
8400 * If this applies to the entire segment, simply mark it and we're done.
8401 */
8402 if ((addr == seg->s_base) && (len == seg->s_size)) {
8403 svd->svn_inz = SEGVN_INZ_ALL;
8404 ret = 0;
8405 goto out;
8406 }
8407
8408 /*
8409 * We've been asked to mark a subset of this segment as inherit zero,
8410 * therefore we need to mainpulate its vpages.
8411 */
8412 if (svd->vpage == NULL) {
8413 segvn_vpage(seg);
8414 if (svd->vpage == NULL) {
8415 ret = ENOMEM;
8416 goto out;
8417 }
8418 }
8419
8420 svd->svn_inz = SEGVN_INZ_VPP;
8421 page = seg_page(seg, addr);
8422 bvpp = &svd->vpage[page];
8423 evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8424 for (; bvpp < evpp; bvpp++)
8425 VPP_SETINHZERO(bvpp);
8426 ret = 0;
8427
8428 out:
8429 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8430 return (ret);
8431 }
8432
8433 /*
8434 * Create a vpage structure for this seg.
8435 */
8436 static void
8437 segvn_vpage(struct seg *seg)
8438 {
8439 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8440 struct vpage *vp, *evp;
8441 static pgcnt_t page_limit = 0;
8442
8443 ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8444
8445 /*
8446 * If no vpage structure exists, allocate one. Copy the protections
8447 * and the advice from the segment itself to the individual pages.
8448 */
8449 if (svd->vpage == NULL) {
8450 /*
8451 * Start by calculating the number of pages we must allocate to
8452 * track the per-page vpage structs needs for this entire
8453 * segment. If we know now that it will require more than our
8454 * heuristic for the maximum amount of kmem we can consume then
8455 * fail. We do this here, instead of trying to detect this deep
8456 * in page_resv and propagating the error up, since the entire
8457 * memory allocation stack is not amenable to passing this
8458 * back. Instead, it wants to keep trying.
8459 *
8460 * As a heuristic we set a page limit of 5/8s of total_pages
8461 * for this allocation. We use shifts so that no floating
8462 * point conversion takes place and only need to do the
8463 * calculation once.
8464 */
8465 ulong_t mem_needed = seg_pages(seg) * sizeof (struct vpage);
8466 pgcnt_t npages = mem_needed >> PAGESHIFT;
8467
8468 if (page_limit == 0)
8469 page_limit = (total_pages >> 1) + (total_pages >> 3);
8470
8471 if (npages > page_limit)
8472 return;
8473
8474 svd->pageadvice = 1;
8475 svd->vpage = kmem_zalloc(mem_needed, KM_SLEEP);
8476 evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
8477 for (vp = svd->vpage; vp < evp; vp++) {
8478 VPP_SETPROT(vp, svd->prot);
8479 VPP_SETADVICE(vp, svd->advice);
8480 }
8481 }
8482 }
8483
8484 /*
8485 * Dump the pages belonging to this segvn segment.
8486 */
8487 static void
8488 segvn_dump(struct seg *seg)
8489 {
8490 struct segvn_data *svd;
8491 page_t *pp;
8492 struct anon_map *amp;
8493 ulong_t anon_index;
8494 struct vnode *vp;
8495 uoff_t off, offset;
8496 pfn_t pfn;
8497 pgcnt_t page, npages;
8498 caddr_t addr;
8499
8500 npages = seg_pages(seg);
8501 svd = (struct segvn_data *)seg->s_data;
8502 vp = svd->vp;
8503 off = offset = svd->offset;
8504 addr = seg->s_base;
8505
8506 if ((amp = svd->amp) != NULL) {
8507 anon_index = svd->anon_index;
8508 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8509 }
8510
8511 for (page = 0; page < npages; page++, offset += PAGESIZE) {
8512 struct anon *ap;
8513 int we_own_it = 0;
8514
8515 if (amp && (ap = anon_get_ptr(svd->amp->ahp, anon_index++))) {
8516 swap_xlate_nopanic(ap, &vp, &off);
8517 } else {
8518 vp = svd->vp;
8519 off = offset;
8520 }
8521
8522 /*
8523 * If pp == NULL, the page either does not exist
8524 * or is exclusively locked. So determine if it
8525 * exists before searching for it.
8526 */
8527
8528 if ((pp = page_lookup_nowait(&vp->v_object, off, SE_SHARED)))
8529 we_own_it = 1;
8530 else
8531 pp = page_exists(&vp->v_object, off);
8532
8533 if (pp) {
8534 pfn = page_pptonum(pp);
8535 dump_addpage(seg->s_as, addr, pfn);
8536 if (we_own_it)
8537 page_unlock(pp);
8538 }
8539 addr += PAGESIZE;
8540 dump_timeleft = dump_timeout;
8541 }
8542
8543 if (amp != NULL)
8544 ANON_LOCK_EXIT(&amp->a_rwlock);
8545 }
8546
8547 #ifdef DEBUG
8548 static uint32_t segvn_pglock_mtbf = 0;
8549 #endif
8550
8551 #define PCACHE_SHWLIST ((page_t *)-2)
8552 #define NOPCACHE_SHWLIST ((page_t *)-1)
8553
8554 /*
8555 * Lock/Unlock anon pages over a given range. Return shadow list. This routine
8556 * uses global segment pcache to cache shadow lists (i.e. pp arrays) of pages
8557 * to avoid the overhead of per page locking, unlocking for subsequent IOs to
8558 * the same parts of the segment. Currently shadow list creation is only
8559 * supported for pure anon segments. MAP_PRIVATE segment pcache entries are
8560 * tagged with segment pointer, starting virtual address and length. This
8561 * approach for MAP_SHARED segments may add many pcache entries for the same
8562 * set of pages and lead to long hash chains that decrease pcache lookup
8563 * performance. To avoid this issue for shared segments shared anon map and
8564 * starting anon index are used for pcache entry tagging. This allows all
8565 * segments to share pcache entries for the same anon range and reduces pcache
8566 * chain's length as well as memory overhead from duplicate shadow lists and
8567 * pcache entries.
8568 *
8569 * softlockcnt field in segvn_data structure counts the number of F_SOFTLOCK'd
8570 * pages via segvn_fault() and pagelock'd pages via this routine. But pagelock
8571 * part of softlockcnt accounting is done differently for private and shared
8572 * segments. In private segment case softlock is only incremented when a new
8573 * shadow list is created but not when an existing one is found via
8574 * seg_plookup(). pcache entries have reference count incremented/decremented
8575 * by each seg_plookup()/seg_pinactive() operation. Only entries that have 0
8576 * reference count can be purged (and purging is needed before segment can be
8577 * freed). When a private segment pcache entry is purged segvn_reclaim() will
8578 * decrement softlockcnt. Since in private segment case each of its pcache
8579 * entries only belongs to this segment we can expect that when
8580 * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8581 * segment purge will succeed and softlockcnt will drop to 0. In shared
8582 * segment case reference count in pcache entry counts active locks from many
8583 * different segments so we can't expect segment purging to succeed even when
8584 * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8585 * segment. To be able to determine when there're no pending pagelocks in
8586 * shared segment case we don't rely on purging to make softlockcnt drop to 0
8587 * but instead softlockcnt is incremented and decremented for every
8588 * segvn_pagelock(L_PAGELOCK/L_PAGEUNLOCK) call regardless if a new shadow
8589 * list was created or an existing one was found. When softlockcnt drops to 0
8590 * this segment no longer has any claims for pcached shadow lists and the
8591 * segment can be freed even if there're still active pcache entries
8592 * shared by this segment anon map. Shared segment pcache entries belong to
8593 * anon map and are typically removed when anon map is freed after all
8594 * processes destroy the segments that use this anon map.
8595 */
8596 static int
8597 segvn_pagelock(struct seg *seg, caddr_t addr, size_t len, struct page ***ppp,
8598 enum lock_type type, enum seg_rw rw)
8599 {
8600 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8601 size_t np;
8602 pgcnt_t adjustpages;
8603 pgcnt_t npages;
8604 ulong_t anon_index;
8605 uint_t protchk = (rw == S_READ) ? PROT_READ : PROT_WRITE;
8606 uint_t error;
8607 struct anon_map *amp;
8608 pgcnt_t anpgcnt;
8609 struct page **pplist, **pl, *pp;
8610 caddr_t a;
8611 size_t page;
8612 caddr_t lpgaddr, lpgeaddr;
8613 anon_sync_obj_t cookie;
8614 int anlock;
8615 struct anon_map *pamp;
8616 caddr_t paddr;
8617 seg_preclaim_cbfunc_t preclaim_callback;
8618 size_t pgsz;
8619 int use_pcache;
8620 size_t wlen;
8621 uint_t pflags = 0;
8622 int sftlck_sbase = 0;
8623 int sftlck_send = 0;
8624
8625 #ifdef DEBUG
8626 if (type == L_PAGELOCK && segvn_pglock_mtbf) {
8627 hrtime_t ts = gethrtime();
8628 if ((ts % segvn_pglock_mtbf) == 0) {
8629 return (ENOTSUP);
8630 }
8631 if ((ts % segvn_pglock_mtbf) == 1) {
8632 return (EFAULT);
8633 }
8634 }
8635 #endif
8636
8637 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
8638 ASSERT(type == L_PAGELOCK || type == L_PAGEUNLOCK);
8639
8640 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8641
8642 /*
8643 * for now we only support pagelock to anon memory. We would have to
8644 * check protections for vnode objects and call into the vnode driver.
8645 * That's too much for a fast path. Let the fault entry point handle
8646 * it.
8647 */
8648 if (svd->vp != NULL) {
8649 if (type == L_PAGELOCK) {
8650 error = ENOTSUP;
8651 goto out;
8652 }
8653 panic("segvn_pagelock(L_PAGEUNLOCK): vp != NULL");
8654 }
8655 if ((amp = svd->amp) == NULL) {
8656 if (type == L_PAGELOCK) {
8657 error = EFAULT;
8658 goto out;
8659 }
8660 panic("segvn_pagelock(L_PAGEUNLOCK): amp == NULL");
8661 }
8662 if (rw != S_READ && rw != S_WRITE) {
8663 if (type == L_PAGELOCK) {
8664 error = ENOTSUP;
8665 goto out;
8666 }
8667 panic("segvn_pagelock(L_PAGEUNLOCK): bad rw");
8668 }
8669
8670 if (seg->s_szc != 0) {
8671 /*
8672 * We are adjusting the pagelock region to the large page size
8673 * boundary because the unlocked part of a large page cannot
8674 * be freed anyway unless all constituent pages of a large
8675 * page are locked. Bigger regions reduce pcache chain length
8676 * and improve lookup performance. The tradeoff is that the
8677 * very first segvn_pagelock() call for a given page is more
8678 * expensive if only 1 page_t is needed for IO. This is only
8679 * an issue if pcache entry doesn't get reused by several
8680 * subsequent calls. We optimize here for the case when pcache
8681 * is heavily used by repeated IOs to the same address range.
8682 *
8683 * Note segment's page size cannot change while we are holding
8684 * as lock. And then it cannot change while softlockcnt is
8685 * not 0. This will allow us to correctly recalculate large
8686 * page size region for the matching pageunlock/reclaim call
8687 * since as_pageunlock() caller must always match
8688 * as_pagelock() call's addr and len.
8689 *
8690 * For pageunlock *ppp points to the pointer of page_t that
8691 * corresponds to the real unadjusted start address. Similar
8692 * for pagelock *ppp must point to the pointer of page_t that
8693 * corresponds to the real unadjusted start address.
8694 */
8695 pgsz = page_get_pagesize(seg->s_szc);
8696 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
8697 adjustpages = btop((uintptr_t)(addr - lpgaddr));
8698 } else if (len < segvn_pglock_comb_thrshld) {
8699 lpgaddr = addr;
8700 lpgeaddr = addr + len;
8701 adjustpages = 0;
8702 pgsz = PAGESIZE;
8703 } else {
8704 /*
8705 * Align the address range of large enough requests to allow
8706 * combining of different shadow lists into 1 to reduce memory
8707 * overhead from potentially overlapping large shadow lists
8708 * (worst case is we have a 1MB IO into buffers with start
8709 * addresses separated by 4K). Alignment is only possible if
8710 * padded chunks have sufficient access permissions. Note
8711 * permissions won't change between L_PAGELOCK and
8712 * L_PAGEUNLOCK calls since non 0 softlockcnt will force
8713 * segvn_setprot() to wait until softlockcnt drops to 0. This
8714 * allows us to determine in L_PAGEUNLOCK the same range we
8715 * computed in L_PAGELOCK.
8716 *
8717 * If alignment is limited by segment ends set
8718 * sftlck_sbase/sftlck_send flags. In L_PAGELOCK case when
8719 * these flags are set bump softlockcnt_sbase/softlockcnt_send
8720 * per segment counters. In L_PAGEUNLOCK case decrease
8721 * softlockcnt_sbase/softlockcnt_send counters if
8722 * sftlck_sbase/sftlck_send flags are set. When
8723 * softlockcnt_sbase/softlockcnt_send are non 0
8724 * segvn_concat()/segvn_extend_prev()/segvn_extend_next()
8725 * won't merge the segments. This restriction combined with
8726 * restriction on segment unmapping and splitting for segments
8727 * that have non 0 softlockcnt allows L_PAGEUNLOCK to
8728 * correctly determine the same range that was previously
8729 * locked by matching L_PAGELOCK.
8730 */
8731 pflags = SEGP_PSHIFT | (segvn_pglock_comb_bshift << 16);
8732 pgsz = PAGESIZE;
8733 if (svd->type == MAP_PRIVATE) {
8734 lpgaddr = (caddr_t)P2ALIGN((uintptr_t)addr,
8735 segvn_pglock_comb_balign);
8736 if (lpgaddr < seg->s_base) {
8737 lpgaddr = seg->s_base;
8738 sftlck_sbase = 1;
8739 }
8740 } else {
8741 ulong_t aix = svd->anon_index + seg_page(seg, addr);
8742 ulong_t aaix = P2ALIGN(aix, segvn_pglock_comb_palign);
8743 if (aaix < svd->anon_index) {
8744 lpgaddr = seg->s_base;
8745 sftlck_sbase = 1;
8746 } else {
8747 lpgaddr = addr - ptob(aix - aaix);
8748 ASSERT(lpgaddr >= seg->s_base);
8749 }
8750 }
8751 if (svd->pageprot && lpgaddr != addr) {
8752 struct vpage *vp = &svd->vpage[seg_page(seg, lpgaddr)];
8753 struct vpage *evp = &svd->vpage[seg_page(seg, addr)];
8754 while (vp < evp) {
8755 if ((VPP_PROT(vp) & protchk) == 0) {
8756 break;
8757 }
8758 vp++;
8759 }
8760 if (vp < evp) {
8761 lpgaddr = addr;
8762 pflags = 0;
8763 }
8764 }
8765 lpgeaddr = addr + len;
8766 if (pflags) {
8767 if (svd->type == MAP_PRIVATE) {
8768 lpgeaddr = (caddr_t)P2ROUNDUP(
8769 (uintptr_t)lpgeaddr,
8770 segvn_pglock_comb_balign);
8771 } else {
8772 ulong_t aix = svd->anon_index +
8773 seg_page(seg, lpgeaddr);
8774 ulong_t aaix = P2ROUNDUP(aix,
8775 segvn_pglock_comb_palign);
8776 if (aaix < aix) {
8777 lpgeaddr = 0;
8778 } else {
8779 lpgeaddr += ptob(aaix - aix);
8780 }
8781 }
8782 if (lpgeaddr == 0 ||
8783 lpgeaddr > seg->s_base + seg->s_size) {
8784 lpgeaddr = seg->s_base + seg->s_size;
8785 sftlck_send = 1;
8786 }
8787 }
8788 if (svd->pageprot && lpgeaddr != addr + len) {
8789 struct vpage *vp;
8790 struct vpage *evp;
8791
8792 vp = &svd->vpage[seg_page(seg, addr + len)];
8793 evp = &svd->vpage[seg_page(seg, lpgeaddr)];
8794
8795 while (vp < evp) {
8796 if ((VPP_PROT(vp) & protchk) == 0) {
8797 break;
8798 }
8799 vp++;
8800 }
8801 if (vp < evp) {
8802 lpgeaddr = addr + len;
8803 }
8804 }
8805 adjustpages = btop((uintptr_t)(addr - lpgaddr));
8806 }
8807
8808 /*
8809 * For MAP_SHARED segments we create pcache entries tagged by amp and
8810 * anon index so that we can share pcache entries with other segments
8811 * that map this amp. For private segments pcache entries are tagged
8812 * with segment and virtual address.
8813 */
8814 if (svd->type == MAP_SHARED) {
8815 pamp = amp;
8816 paddr = (caddr_t)((lpgaddr - seg->s_base) +
8817 ptob(svd->anon_index));
8818 preclaim_callback = shamp_reclaim;
8819 } else {
8820 pamp = NULL;
8821 paddr = lpgaddr;
8822 preclaim_callback = segvn_reclaim;
8823 }
8824
8825 if (type == L_PAGEUNLOCK) {
8826 VM_STAT_ADD(segvnvmstats.pagelock[0]);
8827
8828 /*
8829 * update hat ref bits for /proc. We need to make sure
8830 * that threads tracing the ref and mod bits of the
8831 * address space get the right data.
8832 * Note: page ref and mod bits are updated at reclaim time
8833 */
8834 if (seg->s_as->a_vbits) {
8835 for (a = addr; a < addr + len; a += PAGESIZE) {
8836 if (rw == S_WRITE) {
8837 hat_setstat(seg->s_as, a,
8838 PAGESIZE, P_REF | P_MOD);
8839 } else {
8840 hat_setstat(seg->s_as, a,
8841 PAGESIZE, P_REF);
8842 }
8843 }
8844 }
8845
8846 /*
8847 * Check the shadow list entry after the last page used in
8848 * this IO request. If it's NOPCACHE_SHWLIST the shadow list
8849 * was not inserted into pcache and is not large page
8850 * adjusted. In this case call reclaim callback directly and
8851 * don't adjust the shadow list start and size for large
8852 * pages.
8853 */
8854 npages = btop(len);
8855 if ((*ppp)[npages] == NOPCACHE_SHWLIST) {
8856 void *ptag;
8857 if (pamp != NULL) {
8858 ASSERT(svd->type == MAP_SHARED);
8859 ptag = (void *)pamp;
8860 paddr = (caddr_t)((addr - seg->s_base) +
8861 ptob(svd->anon_index));
8862 } else {
8863 ptag = (void *)seg;
8864 paddr = addr;
8865 }
8866 (*preclaim_callback)(ptag, paddr, len, *ppp, rw, 0);
8867 } else {
8868 ASSERT((*ppp)[npages] == PCACHE_SHWLIST ||
8869 IS_SWAPFSVP((*ppp)[npages]->p_vnode));
8870 len = lpgeaddr - lpgaddr;
8871 npages = btop(len);
8872 seg_pinactive(seg, pamp, paddr, len,
8873 *ppp - adjustpages, rw, pflags, preclaim_callback);
8874 }
8875
8876 if (pamp != NULL) {
8877 ASSERT(svd->type == MAP_SHARED);
8878 ASSERT(svd->softlockcnt >= npages);
8879 atomic_add_long((ulong_t *)&svd->softlockcnt, -npages);
8880 }
8881
8882 if (sftlck_sbase) {
8883 ASSERT(svd->softlockcnt_sbase > 0);
8884 atomic_dec_ulong((ulong_t *)&svd->softlockcnt_sbase);
8885 }
8886 if (sftlck_send) {
8887 ASSERT(svd->softlockcnt_send > 0);
8888 atomic_dec_ulong((ulong_t *)&svd->softlockcnt_send);
8889 }
8890
8891 /*
8892 * If someone is blocked while unmapping, we purge
8893 * segment page cache and thus reclaim pplist synchronously
8894 * without waiting for seg_pasync_thread. This speeds up
8895 * unmapping in cases where munmap(2) is called, while
8896 * raw async i/o is still in progress or where a thread
8897 * exits on data fault in a multithreaded application.
8898 */
8899 if (AS_ISUNMAPWAIT(seg->s_as)) {
8900 if (svd->softlockcnt == 0) {
8901 mutex_enter(&seg->s_as->a_contents);
8902 if (AS_ISUNMAPWAIT(seg->s_as)) {
8903 AS_CLRUNMAPWAIT(seg->s_as);
8904 cv_broadcast(&seg->s_as->a_cv);
8905 }
8906 mutex_exit(&seg->s_as->a_contents);
8907 } else if (pamp == NULL) {
8908 /*
8909 * softlockcnt is not 0 and this is a
8910 * MAP_PRIVATE segment. Try to purge its
8911 * pcache entries to reduce softlockcnt.
8912 * If it drops to 0 segvn_reclaim()
8913 * will wake up a thread waiting on
8914 * unmapwait flag.
8915 *
8916 * We don't purge MAP_SHARED segments with non
8917 * 0 softlockcnt since IO is still in progress
8918 * for such segments.
8919 */
8920 ASSERT(svd->type == MAP_PRIVATE);
8921 segvn_purge(seg);
8922 }
8923 }
8924 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8925 return (0);
8926 }
8927
8928 /* The L_PAGELOCK case ... */
8929
8930 VM_STAT_ADD(segvnvmstats.pagelock[1]);
8931
8932 /*
8933 * For MAP_SHARED segments we have to check protections before
8934 * seg_plookup() since pcache entries may be shared by many segments
8935 * with potentially different page protections.
8936 */
8937 if (pamp != NULL) {
8938 ASSERT(svd->type == MAP_SHARED);
8939 if (svd->pageprot == 0) {
8940 if ((svd->prot & protchk) == 0) {
8941 error = EACCES;
8942 goto out;
8943 }
8944 } else {
8945 /*
8946 * check page protections
8947 */
8948 caddr_t ea;
8949
8950 if (seg->s_szc) {
8951 a = lpgaddr;
8952 ea = lpgeaddr;
8953 } else {
8954 a = addr;
8955 ea = addr + len;
8956 }
8957 for (; a < ea; a += pgsz) {
8958 struct vpage *vp;
8959
8960 ASSERT(seg->s_szc == 0 ||
8961 sameprot(seg, a, pgsz));
8962 vp = &svd->vpage[seg_page(seg, a)];
8963 if ((VPP_PROT(vp) & protchk) == 0) {
8964 error = EACCES;
8965 goto out;
8966 }
8967 }
8968 }
8969 }
8970
8971 /*
8972 * try to find pages in segment page cache
8973 */
8974 pplist = seg_plookup(seg, pamp, paddr, lpgeaddr - lpgaddr, rw, pflags);
8975 if (pplist != NULL) {
8976 if (pamp != NULL) {
8977 npages = btop((uintptr_t)(lpgeaddr - lpgaddr));
8978 ASSERT(svd->type == MAP_SHARED);
8979 atomic_add_long((ulong_t *)&svd->softlockcnt,
8980 npages);
8981 }
8982 if (sftlck_sbase) {
8983 atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
8984 }
8985 if (sftlck_send) {
8986 atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
8987 }
8988 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8989 *ppp = pplist + adjustpages;
8990 return (0);
8991 }
8992
8993 /*
8994 * For MAP_SHARED segments we already verified above that segment
8995 * protections allow this pagelock operation.
8996 */
8997 if (pamp == NULL) {
8998 ASSERT(svd->type == MAP_PRIVATE);
8999 if (svd->pageprot == 0) {
9000 if ((svd->prot & protchk) == 0) {
9001 error = EACCES;
9002 goto out;
9003 }
9004 if (svd->prot & PROT_WRITE) {
9005 wlen = lpgeaddr - lpgaddr;
9006 } else {
9007 wlen = 0;
9008 ASSERT(rw == S_READ);
9009 }
9010 } else {
9011 int wcont = 1;
9012 /*
9013 * check page protections
9014 */
9015 for (a = lpgaddr, wlen = 0; a < lpgeaddr; a += pgsz) {
9016 struct vpage *vp;
9017
9018 ASSERT(seg->s_szc == 0 ||
9019 sameprot(seg, a, pgsz));
9020 vp = &svd->vpage[seg_page(seg, a)];
9021 if ((VPP_PROT(vp) & protchk) == 0) {
9022 error = EACCES;
9023 goto out;
9024 }
9025 if (wcont && (VPP_PROT(vp) & PROT_WRITE)) {
9026 wlen += pgsz;
9027 } else {
9028 wcont = 0;
9029 ASSERT(rw == S_READ);
9030 }
9031 }
9032 }
9033 ASSERT(rw == S_READ || wlen == lpgeaddr - lpgaddr);
9034 ASSERT(rw == S_WRITE || wlen <= lpgeaddr - lpgaddr);
9035 }
9036
9037 /*
9038 * Only build large page adjusted shadow list if we expect to insert
9039 * it into pcache. For large enough pages it's a big overhead to
9040 * create a shadow list of the entire large page. But this overhead
9041 * should be amortized over repeated pcache hits on subsequent reuse
9042 * of this shadow list (IO into any range within this shadow list will
9043 * find it in pcache since we large page align the request for pcache
9044 * lookups). pcache performance is improved with bigger shadow lists
9045 * as it reduces the time to pcache the entire big segment and reduces
9046 * pcache chain length.
9047 */
9048 if (seg_pinsert_check(seg, pamp, paddr,
9049 lpgeaddr - lpgaddr, pflags) == SEGP_SUCCESS) {
9050 addr = lpgaddr;
9051 len = lpgeaddr - lpgaddr;
9052 use_pcache = 1;
9053 } else {
9054 use_pcache = 0;
9055 /*
9056 * Since this entry will not be inserted into the pcache, we
9057 * will not do any adjustments to the starting address or
9058 * size of the memory to be locked.
9059 */
9060 adjustpages = 0;
9061 }
9062 npages = btop(len);
9063
9064 pplist = kmem_alloc(sizeof (page_t *) * (npages + 1), KM_SLEEP);
9065 pl = pplist;
9066 *ppp = pplist + adjustpages;
9067 /*
9068 * If use_pcache is 0 this shadow list is not large page adjusted.
9069 * Record this info in the last entry of shadow array so that
9070 * L_PAGEUNLOCK can determine if it should large page adjust the
9071 * address range to find the real range that was locked.
9072 */
9073 pl[npages] = use_pcache ? PCACHE_SHWLIST : NOPCACHE_SHWLIST;
9074
9075 page = seg_page(seg, addr);
9076 anon_index = svd->anon_index + page;
9077
9078 anlock = 0;
9079 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9080 ASSERT(amp->a_szc >= seg->s_szc);
9081 anpgcnt = page_get_pagecnt(amp->a_szc);
9082 for (a = addr; a < addr + len; a += PAGESIZE, anon_index++) {
9083 struct anon *ap;
9084 struct vnode *vp;
9085 uoff_t off;
9086
9087 /*
9088 * Lock and unlock anon array only once per large page.
9089 * anon_array_enter() locks the root anon slot according to
9090 * a_szc which can't change while anon map is locked. We lock
9091 * anon the first time through this loop and each time we
9092 * reach anon index that corresponds to a root of a large
9093 * page.
9094 */
9095 if (a == addr || P2PHASE(anon_index, anpgcnt) == 0) {
9096 ASSERT(anlock == 0);
9097 anon_array_enter(amp, anon_index, &cookie);
9098 anlock = 1;
9099 }
9100 ap = anon_get_ptr(amp->ahp, anon_index);
9101
9102 /*
9103 * We must never use seg_pcache for COW pages
9104 * because we might end up with original page still
9105 * lying in seg_pcache even after private page is
9106 * created. This leads to data corruption as
9107 * aio_write refers to the page still in cache
9108 * while all other accesses refer to the private
9109 * page.
9110 */
9111 if (ap == NULL || ap->an_refcnt != 1) {
9112 struct vpage *vpage;
9113
9114 if (seg->s_szc) {
9115 error = EFAULT;
9116 break;
9117 }
9118 if (svd->vpage != NULL) {
9119 vpage = &svd->vpage[seg_page(seg, a)];
9120 } else {
9121 vpage = NULL;
9122 }
9123 ASSERT(anlock);
9124 anon_array_exit(&cookie);
9125 anlock = 0;
9126 pp = NULL;
9127 error = segvn_faultpage(seg->s_as->a_hat, seg, a, 0,
9128 vpage, &pp, 0, F_INVAL, rw, 1);
9129 if (error) {
9130 error = fc_decode(error);
9131 break;
9132 }
9133 anon_array_enter(amp, anon_index, &cookie);
9134 anlock = 1;
9135 ap = anon_get_ptr(amp->ahp, anon_index);
9136 if (ap == NULL || ap->an_refcnt != 1) {
9137 error = EFAULT;
9138 break;
9139 }
9140 }
9141 swap_xlate(ap, &vp, &off);
9142 pp = page_lookup_nowait(&vp->v_object, off, SE_SHARED);
9143 if (pp == NULL) {
9144 error = EFAULT;
9145 break;
9146 }
9147 if (ap->an_pvp != NULL) {
9148 anon_swap_free(ap, pp);
9149 }
9150 /*
9151 * Unlock anon if this is the last slot in a large page.
9152 */
9153 if (P2PHASE(anon_index, anpgcnt) == anpgcnt - 1) {
9154 ASSERT(anlock);
9155 anon_array_exit(&cookie);
9156 anlock = 0;
9157 }
9158 *pplist++ = pp;
9159 }
9160 if (anlock) { /* Ensure the lock is dropped */
9161 anon_array_exit(&cookie);
9162 }
9163 ANON_LOCK_EXIT(&amp->a_rwlock);
9164
9165 if (a >= addr + len) {
9166 atomic_add_long((ulong_t *)&svd->softlockcnt, npages);
9167 if (pamp != NULL) {
9168 ASSERT(svd->type == MAP_SHARED);
9169 atomic_add_long((ulong_t *)&pamp->a_softlockcnt,
9170 npages);
9171 wlen = len;
9172 }
9173 if (sftlck_sbase) {
9174 atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
9175 }
9176 if (sftlck_send) {
9177 atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
9178 }
9179 if (use_pcache) {
9180 (void) seg_pinsert(seg, pamp, paddr, len, wlen, pl,
9181 rw, pflags, preclaim_callback);
9182 }
9183 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9184 return (0);
9185 }
9186
9187 pplist = pl;
9188 np = ((uintptr_t)(a - addr)) >> PAGESHIFT;
9189 while (np > 0) {
9190 ASSERT(PAGE_LOCKED(*pplist));
9191 page_unlock(*pplist);
9192 np--;
9193 pplist++;
9194 }
9195 kmem_free(pl, sizeof (page_t *) * (npages + 1));
9196 out:
9197 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9198 *ppp = NULL;
9199 return (error);
9200 }
9201
9202 /*
9203 * purge any cached pages in the I/O page cache
9204 */
9205 static void
9206 segvn_purge(struct seg *seg)
9207 {
9208 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9209
9210 /*
9211 * pcache is only used by pure anon segments.
9212 */
9213 if (svd->amp == NULL || svd->vp != NULL) {
9214 return;
9215 }
9216
9217 /*
9218 * For MAP_SHARED segments non 0 segment's softlockcnt means
9219 * active IO is still in progress via this segment. So we only
9220 * purge MAP_SHARED segments when their softlockcnt is 0.
9221 */
9222 if (svd->type == MAP_PRIVATE) {
9223 if (svd->softlockcnt) {
9224 seg_ppurge(seg, NULL, 0);
9225 }
9226 } else if (svd->softlockcnt == 0 && svd->amp->a_softlockcnt != 0) {
9227 seg_ppurge(seg, svd->amp, 0);
9228 }
9229 }
9230
9231 /*
9232 * If async argument is not 0 we are called from pcache async thread and don't
9233 * hold AS lock.
9234 */
9235
9236 /*ARGSUSED*/
9237 static int
9238 segvn_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9239 enum seg_rw rw, int async)
9240 {
9241 struct seg *seg = (struct seg *)ptag;
9242 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9243 pgcnt_t np, npages;
9244 struct page **pl;
9245
9246 npages = np = btop(len);
9247 ASSERT(npages);
9248
9249 ASSERT(svd->vp == NULL && svd->amp != NULL);
9250 ASSERT(svd->softlockcnt >= npages);
9251 ASSERT(async || AS_LOCK_HELD(seg->s_as));
9252
9253 pl = pplist;
9254
9255 ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9256 ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9257
9258 while (np > 0) {
9259 if (rw == S_WRITE) {
9260 hat_setrefmod(*pplist);
9261 } else {
9262 hat_setref(*pplist);
9263 }
9264 page_unlock(*pplist);
9265 np--;
9266 pplist++;
9267 }
9268
9269 kmem_free(pl, sizeof (page_t *) * (npages + 1));
9270
9271 /*
9272 * If we are pcache async thread we don't hold AS lock. This means if
9273 * softlockcnt drops to 0 after the decrement below address space may
9274 * get freed. We can't allow it since after softlock derement to 0 we
9275 * still need to access as structure for possible wakeup of unmap
9276 * waiters. To prevent the disappearance of as we take this segment
9277 * segfree_syncmtx. segvn_free() also takes this mutex as a barrier to
9278 * make sure this routine completes before segment is freed.
9279 *
9280 * The second complication we have to deal with in async case is a
9281 * possibility of missed wake up of unmap wait thread. When we don't
9282 * hold as lock here we may take a_contents lock before unmap wait
9283 * thread that was first to see softlockcnt was still not 0. As a
9284 * result we'll fail to wake up an unmap wait thread. To avoid this
9285 * race we set nounmapwait flag in as structure if we drop softlockcnt
9286 * to 0 when we were called by pcache async thread. unmapwait thread
9287 * will not block if this flag is set.
9288 */
9289 if (async) {
9290 mutex_enter(&svd->segfree_syncmtx);
9291 }
9292
9293 if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -npages)) {
9294 if (async || AS_ISUNMAPWAIT(seg->s_as)) {
9295 mutex_enter(&seg->s_as->a_contents);
9296 if (async) {
9297 AS_SETNOUNMAPWAIT(seg->s_as);
9298 }
9299 if (AS_ISUNMAPWAIT(seg->s_as)) {
9300 AS_CLRUNMAPWAIT(seg->s_as);
9301 cv_broadcast(&seg->s_as->a_cv);
9302 }
9303 mutex_exit(&seg->s_as->a_contents);
9304 }
9305 }
9306
9307 if (async) {
9308 mutex_exit(&svd->segfree_syncmtx);
9309 }
9310 return (0);
9311 }
9312
9313 /*ARGSUSED*/
9314 static int
9315 shamp_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9316 enum seg_rw rw, int async)
9317 {
9318 amp_t *amp = (amp_t *)ptag;
9319 pgcnt_t np, npages;
9320 struct page **pl;
9321
9322 npages = np = btop(len);
9323 ASSERT(npages);
9324 ASSERT(amp->a_softlockcnt >= npages);
9325
9326 pl = pplist;
9327
9328 ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9329 ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9330
9331 while (np > 0) {
9332 if (rw == S_WRITE) {
9333 hat_setrefmod(*pplist);
9334 } else {
9335 hat_setref(*pplist);
9336 }
9337 page_unlock(*pplist);
9338 np--;
9339 pplist++;
9340 }
9341
9342 kmem_free(pl, sizeof (page_t *) * (npages + 1));
9343
9344 /*
9345 * If somebody sleeps in anonmap_purge() wake them up if a_softlockcnt
9346 * drops to 0. anon map can't be freed until a_softlockcnt drops to 0
9347 * and anonmap_purge() acquires a_purgemtx.
9348 */
9349 mutex_enter(&amp->a_purgemtx);
9350 if (!atomic_add_long_nv((ulong_t *)&amp->a_softlockcnt, -npages) &&
9351 amp->a_purgewait) {
9352 amp->a_purgewait = 0;
9353 cv_broadcast(&amp->a_purgecv);
9354 }
9355 mutex_exit(&amp->a_purgemtx);
9356 return (0);
9357 }
9358
9359 /*
9360 * get a memory ID for an addr in a given segment
9361 *
9362 * XXX only creates PAGESIZE pages if anon slots are not initialized.
9363 * At fault time they will be relocated into larger pages.
9364 */
9365 static int
9366 segvn_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
9367 {
9368 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9369 struct anon *ap = NULL;
9370 ulong_t anon_index;
9371 struct anon_map *amp;
9372 anon_sync_obj_t cookie;
9373
9374 if (svd->type == MAP_PRIVATE) {
9375 memidp->val[0] = (uintptr_t)seg->s_as;
9376 memidp->val[1] = (uintptr_t)addr;
9377 return (0);
9378 }
9379
9380 if (svd->type == MAP_SHARED) {
9381 if (svd->vp) {
9382 memidp->val[0] = (uintptr_t)svd->vp;
9383 memidp->val[1] = (u_longlong_t)svd->offset +
9384 (uintptr_t)(addr - seg->s_base);
9385 return (0);
9386 } else {
9387
9388 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
9389 if ((amp = svd->amp) != NULL) {
9390 anon_index = svd->anon_index +
9391 seg_page(seg, addr);
9392 }
9393 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9394
9395 ASSERT(amp != NULL);
9396
9397 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9398 anon_array_enter(amp, anon_index, &cookie);
9399 ap = anon_get_ptr(amp->ahp, anon_index);
9400 if (ap == NULL) {
9401 page_t *pp;
9402
9403 pp = anon_zero(seg, addr, &ap, svd->cred);
9404 if (pp == NULL) {
9405 anon_array_exit(&cookie);
9406 ANON_LOCK_EXIT(&amp->a_rwlock);
9407 return (ENOMEM);
9408 }
9409 ASSERT(anon_get_ptr(amp->ahp, anon_index)
9410 == NULL);
9411 (void) anon_set_ptr(amp->ahp, anon_index,
9412 ap, ANON_SLEEP);
9413 page_unlock(pp);
9414 }
9415
9416 anon_array_exit(&cookie);
9417 ANON_LOCK_EXIT(&amp->a_rwlock);
9418
9419 memidp->val[0] = (uintptr_t)ap;
9420 memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
9421 return (0);
9422 }
9423 }
9424 return (EINVAL);
9425 }
9426
9427 static int
9428 sameprot(struct seg *seg, caddr_t a, size_t len)
9429 {
9430 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9431 struct vpage *vpage;
9432 spgcnt_t pages = btop(len);
9433 uint_t prot;
9434
9435 if (svd->pageprot == 0)
9436 return (1);
9437
9438 ASSERT(svd->vpage != NULL);
9439
9440 vpage = &svd->vpage[seg_page(seg, a)];
9441 prot = VPP_PROT(vpage);
9442 vpage++;
9443 pages--;
9444 while (pages-- > 0) {
9445 if (prot != VPP_PROT(vpage))
9446 return (0);
9447 vpage++;
9448 }
9449 return (1);
9450 }
9451
9452 /*
9453 * Get memory allocation policy info for specified address in given segment
9454 */
9455 static lgrp_mem_policy_info_t *
9456 segvn_getpolicy(struct seg *seg, caddr_t addr)
9457 {
9458 struct anon_map *amp;
9459 ulong_t anon_index;
9460 lgrp_mem_policy_info_t *policy_info;
9461 struct segvn_data *svn_data;
9462 uoff_t vn_off;
9463 vnode_t *vp;
9464
9465 ASSERT(seg != NULL);
9466
9467 svn_data = (struct segvn_data *)seg->s_data;
9468 if (svn_data == NULL)
9469 return (NULL);
9470
9471 /*
9472 * Get policy info for private or shared memory
9473 */
9474 if (svn_data->type != MAP_SHARED) {
9475 if (svn_data->tr_state != SEGVN_TR_ON) {
9476 policy_info = &svn_data->policy_info;
9477 } else {
9478 policy_info = &svn_data->tr_policy_info;
9479 ASSERT(policy_info->mem_policy ==
9480 LGRP_MEM_POLICY_NEXT_SEG);
9481 }
9482 } else {
9483 amp = svn_data->amp;
9484 anon_index = svn_data->anon_index + seg_page(seg, addr);
9485 vp = svn_data->vp;
9486 vn_off = svn_data->offset + (uintptr_t)(addr - seg->s_base);
9487 policy_info = lgrp_shm_policy_get(amp, anon_index, vp, vn_off);
9488 }
9489
9490 return (policy_info);
9491 }
9492
9493 /*
9494 * Bind text vnode segment to an amp. If we bind successfully mappings will be
9495 * established to per vnode mapping per lgroup amp pages instead of to vnode
9496 * pages. There's one amp per vnode text mapping per lgroup. Many processes
9497 * may share the same text replication amp. If a suitable amp doesn't already
9498 * exist in svntr hash table create a new one. We may fail to bind to amp if
9499 * segment is not eligible for text replication. Code below first checks for
9500 * these conditions. If binding is successful segment tr_state is set to on
9501 * and svd->amp points to the amp to use. Otherwise tr_state is set to off and
9502 * svd->amp remains as NULL.
9503 */
9504 static void
9505 segvn_textrepl(struct seg *seg)
9506 {
9507 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9508 vnode_t *vp = svd->vp;
9509 uoff_t off = svd->offset;
9510 size_t size = seg->s_size;
9511 uoff_t eoff = off + size;
9512 uint_t szc = seg->s_szc;
9513 ulong_t hash = SVNTR_HASH_FUNC(vp);
9514 svntr_t *svntrp;
9515 struct vattr va;
9516 proc_t *p = seg->s_as->a_proc;
9517 lgrp_id_t lgrp_id;
9518 lgrp_id_t olid;
9519 int first;
9520 struct anon_map *amp;
9521
9522 ASSERT(AS_LOCK_HELD(seg->s_as));
9523 ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9524 ASSERT(p != NULL);
9525 ASSERT(svd->tr_state == SEGVN_TR_INIT);
9526 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9527 ASSERT(svd->flags & MAP_TEXT);
9528 ASSERT(svd->type == MAP_PRIVATE);
9529 ASSERT(vp != NULL && svd->amp == NULL);
9530 ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
9531 ASSERT(!(svd->flags & MAP_NORESERVE) && svd->swresv == 0);
9532 ASSERT(seg->s_as != &kas);
9533 ASSERT(off < eoff);
9534 ASSERT(svntr_hashtab != NULL);
9535
9536 /*
9537 * If numa optimizations are no longer desired bail out.
9538 */
9539 if (!lgrp_optimizations()) {
9540 svd->tr_state = SEGVN_TR_OFF;
9541 return;
9542 }
9543
9544 /*
9545 * Avoid creating anon maps with size bigger than the file size.
9546 * If fop_getattr() call fails bail out.
9547 */
9548 va.va_mask = VATTR_SIZE | VATTR_MTIME | VATTR_CTIME;
9549 if (fop_getattr(vp, &va, 0, svd->cred, NULL) != 0) {
9550 svd->tr_state = SEGVN_TR_OFF;
9551 SEGVN_TR_ADDSTAT(gaerr);
9552 return;
9553 }
9554 if (btopr(va.va_size) < btopr(eoff)) {
9555 svd->tr_state = SEGVN_TR_OFF;
9556 SEGVN_TR_ADDSTAT(overmap);
9557 return;
9558 }
9559
9560 /*
9561 * VVMEXEC may not be set yet if exec() prefaults text segment. Set
9562 * this flag now before vn_is_mapped(V_WRITE) so that MAP_SHARED
9563 * mapping that checks if trcache for this vnode needs to be
9564 * invalidated can't miss us.
9565 */
9566 if (!(vp->v_flag & VVMEXEC)) {
9567 mutex_enter(&vp->v_lock);
9568 vp->v_flag |= VVMEXEC;
9569 mutex_exit(&vp->v_lock);
9570 }
9571 mutex_enter(&svntr_hashtab[hash].tr_lock);
9572 /*
9573 * Bail out if potentially MAP_SHARED writable mappings exist to this
9574 * vnode. We don't want to use old file contents from existing
9575 * replicas if this mapping was established after the original file
9576 * was changed.
9577 */
9578 if (vn_is_mapped(vp, V_WRITE)) {
9579 mutex_exit(&svntr_hashtab[hash].tr_lock);
9580 svd->tr_state = SEGVN_TR_OFF;
9581 SEGVN_TR_ADDSTAT(wrcnt);
9582 return;
9583 }
9584 svntrp = svntr_hashtab[hash].tr_head;
9585 for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9586 ASSERT(svntrp->tr_refcnt != 0);
9587 if (svntrp->tr_vp != vp) {
9588 continue;
9589 }
9590
9591 /*
9592 * Bail out if the file or its attributes were changed after
9593 * this replication entry was created since we need to use the
9594 * latest file contents. Note that mtime test alone is not
9595 * sufficient because a user can explicitly change mtime via
9596 * utimes(2) interfaces back to the old value after modifiying
9597 * the file contents. To detect this case we also have to test
9598 * ctime which among other things records the time of the last
9599 * mtime change by utimes(2). ctime is not changed when the file
9600 * is only read or executed so we expect that typically existing
9601 * replication amp's can be used most of the time.
9602 */
9603 if (!svntrp->tr_valid ||
9604 svntrp->tr_mtime.tv_sec != va.va_mtime.tv_sec ||
9605 svntrp->tr_mtime.tv_nsec != va.va_mtime.tv_nsec ||
9606 svntrp->tr_ctime.tv_sec != va.va_ctime.tv_sec ||
9607 svntrp->tr_ctime.tv_nsec != va.va_ctime.tv_nsec) {
9608 mutex_exit(&svntr_hashtab[hash].tr_lock);
9609 svd->tr_state = SEGVN_TR_OFF;
9610 SEGVN_TR_ADDSTAT(stale);
9611 return;
9612 }
9613 /*
9614 * if off, eoff and szc match current segment we found the
9615 * existing entry we can use.
9616 */
9617 if (svntrp->tr_off == off && svntrp->tr_eoff == eoff &&
9618 svntrp->tr_szc == szc) {
9619 break;
9620 }
9621 /*
9622 * Don't create different but overlapping in file offsets
9623 * entries to avoid replication of the same file pages more
9624 * than once per lgroup.
9625 */
9626 if ((off >= svntrp->tr_off && off < svntrp->tr_eoff) ||
9627 (eoff > svntrp->tr_off && eoff <= svntrp->tr_eoff)) {
9628 mutex_exit(&svntr_hashtab[hash].tr_lock);
9629 svd->tr_state = SEGVN_TR_OFF;
9630 SEGVN_TR_ADDSTAT(overlap);
9631 return;
9632 }
9633 }
9634 /*
9635 * If we didn't find existing entry create a new one.
9636 */
9637 if (svntrp == NULL) {
9638 svntrp = kmem_cache_alloc(svntr_cache, KM_NOSLEEP);
9639 if (svntrp == NULL) {
9640 mutex_exit(&svntr_hashtab[hash].tr_lock);
9641 svd->tr_state = SEGVN_TR_OFF;
9642 SEGVN_TR_ADDSTAT(nokmem);
9643 return;
9644 }
9645 #ifdef DEBUG
9646 {
9647 lgrp_id_t i;
9648 for (i = 0; i < NLGRPS_MAX; i++) {
9649 ASSERT(svntrp->tr_amp[i] == NULL);
9650 }
9651 }
9652 #endif /* DEBUG */
9653 svntrp->tr_vp = vp;
9654 svntrp->tr_off = off;
9655 svntrp->tr_eoff = eoff;
9656 svntrp->tr_szc = szc;
9657 svntrp->tr_valid = 1;
9658 svntrp->tr_mtime = va.va_mtime;
9659 svntrp->tr_ctime = va.va_ctime;
9660 svntrp->tr_refcnt = 0;
9661 svntrp->tr_next = svntr_hashtab[hash].tr_head;
9662 svntr_hashtab[hash].tr_head = svntrp;
9663 }
9664 first = 1;
9665 again:
9666 /*
9667 * We want to pick a replica with pages on main thread's (t_tid = 1,
9668 * aka T1) lgrp. Currently text replication is only optimized for
9669 * workloads that either have all threads of a process on the same
9670 * lgrp or execute their large text primarily on main thread.
9671 */
9672 lgrp_id = p->p_t1_lgrpid;
9673 if (lgrp_id == LGRP_NONE) {
9674 /*
9675 * In case exec() prefaults text on non main thread use
9676 * current thread lgrpid. It will become main thread anyway
9677 * soon.
9678 */
9679 lgrp_id = lgrp_home_id(curthread);
9680 }
9681 /*
9682 * Set p_tr_lgrpid to lgrpid if it hasn't been set yet. Otherwise
9683 * just set it to NLGRPS_MAX if it's different from current process T1
9684 * home lgrp. p_tr_lgrpid is used to detect if process uses text
9685 * replication and T1 new home is different from lgrp used for text
9686 * replication. When this happens asyncronous segvn thread rechecks if
9687 * segments should change lgrps used for text replication. If we fail
9688 * to set p_tr_lgrpid with atomic_cas_32 then set it to NLGRPS_MAX
9689 * without cas if it's not already NLGRPS_MAX and not equal lgrp_id
9690 * we want to use. We don't need to use cas in this case because
9691 * another thread that races in between our non atomic check and set
9692 * may only change p_tr_lgrpid to NLGRPS_MAX at this point.
9693 */
9694 ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9695 olid = p->p_tr_lgrpid;
9696 if (lgrp_id != olid && olid != NLGRPS_MAX) {
9697 lgrp_id_t nlid = (olid == LGRP_NONE) ? lgrp_id : NLGRPS_MAX;
9698 if (atomic_cas_32((uint32_t *)&p->p_tr_lgrpid, olid, nlid) !=
9699 olid) {
9700 olid = p->p_tr_lgrpid;
9701 ASSERT(olid != LGRP_NONE);
9702 if (olid != lgrp_id && olid != NLGRPS_MAX) {
9703 p->p_tr_lgrpid = NLGRPS_MAX;
9704 }
9705 }
9706 ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9707 membar_producer();
9708 /*
9709 * lgrp_move_thread() won't schedule async recheck after
9710 * p->p_t1_lgrpid update unless p->p_tr_lgrpid is not
9711 * LGRP_NONE. Recheck p_t1_lgrpid once now that p->p_tr_lgrpid
9712 * is not LGRP_NONE.
9713 */
9714 if (first && p->p_t1_lgrpid != LGRP_NONE &&
9715 p->p_t1_lgrpid != lgrp_id) {
9716 first = 0;
9717 goto again;
9718 }
9719 }
9720 /*
9721 * If no amp was created yet for lgrp_id create a new one as long as
9722 * we have enough memory to afford it.
9723 */
9724 if ((amp = svntrp->tr_amp[lgrp_id]) == NULL) {
9725 size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9726 if (trmem > segvn_textrepl_max_bytes) {
9727 SEGVN_TR_ADDSTAT(normem);
9728 goto fail;
9729 }
9730 if (anon_try_resv_zone(size, NULL) == 0) {
9731 SEGVN_TR_ADDSTAT(noanon);
9732 goto fail;
9733 }
9734 amp = anonmap_alloc(size, size, ANON_NOSLEEP);
9735 if (amp == NULL) {
9736 anon_unresv_zone(size, NULL);
9737 SEGVN_TR_ADDSTAT(nokmem);
9738 goto fail;
9739 }
9740 ASSERT(amp->refcnt == 1);
9741 amp->a_szc = szc;
9742 svntrp->tr_amp[lgrp_id] = amp;
9743 SEGVN_TR_ADDSTAT(newamp);
9744 }
9745 svntrp->tr_refcnt++;
9746 ASSERT(svd->svn_trnext == NULL);
9747 ASSERT(svd->svn_trprev == NULL);
9748 svd->svn_trnext = svntrp->tr_svnhead;
9749 svd->svn_trprev = NULL;
9750 if (svntrp->tr_svnhead != NULL) {
9751 svntrp->tr_svnhead->svn_trprev = svd;
9752 }
9753 svntrp->tr_svnhead = svd;
9754 ASSERT(amp->a_szc == szc && amp->size == size && amp->swresv == size);
9755 ASSERT(amp->refcnt >= 1);
9756 svd->amp = amp;
9757 svd->anon_index = 0;
9758 svd->tr_policy_info.mem_policy = LGRP_MEM_POLICY_NEXT_SEG;
9759 svd->tr_policy_info.mem_lgrpid = lgrp_id;
9760 svd->tr_state = SEGVN_TR_ON;
9761 mutex_exit(&svntr_hashtab[hash].tr_lock);
9762 SEGVN_TR_ADDSTAT(repl);
9763 return;
9764 fail:
9765 ASSERT(segvn_textrepl_bytes >= size);
9766 atomic_add_long(&segvn_textrepl_bytes, -size);
9767 ASSERT(svntrp != NULL);
9768 ASSERT(svntrp->tr_amp[lgrp_id] == NULL);
9769 if (svntrp->tr_refcnt == 0) {
9770 ASSERT(svntrp == svntr_hashtab[hash].tr_head);
9771 svntr_hashtab[hash].tr_head = svntrp->tr_next;
9772 mutex_exit(&svntr_hashtab[hash].tr_lock);
9773 kmem_cache_free(svntr_cache, svntrp);
9774 } else {
9775 mutex_exit(&svntr_hashtab[hash].tr_lock);
9776 }
9777 svd->tr_state = SEGVN_TR_OFF;
9778 }
9779
9780 /*
9781 * Convert seg back to regular vnode mapping seg by unbinding it from its text
9782 * replication amp. This routine is most typically called when segment is
9783 * unmapped but can also be called when segment no longer qualifies for text
9784 * replication (e.g. due to protection changes). If unload_unmap is set use
9785 * HAT_UNLOAD_UNMAP flag in hat_unload_callback(). If we are the last user of
9786 * svntr free all its anon maps and remove it from the hash table.
9787 */
9788 static void
9789 segvn_textunrepl(struct seg *seg, int unload_unmap)
9790 {
9791 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9792 vnode_t *vp = svd->vp;
9793 uoff_t off = svd->offset;
9794 size_t size = seg->s_size;
9795 uoff_t eoff = off + size;
9796 uint_t szc = seg->s_szc;
9797 ulong_t hash = SVNTR_HASH_FUNC(vp);
9798 svntr_t *svntrp;
9799 svntr_t **prv_svntrp;
9800 lgrp_id_t lgrp_id = svd->tr_policy_info.mem_lgrpid;
9801 lgrp_id_t i;
9802
9803 ASSERT(AS_LOCK_HELD(seg->s_as));
9804 ASSERT(AS_WRITE_HELD(seg->s_as) ||
9805 SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9806 ASSERT(svd->tr_state == SEGVN_TR_ON);
9807 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9808 ASSERT(svd->amp != NULL);
9809 ASSERT(svd->amp->refcnt >= 1);
9810 ASSERT(svd->anon_index == 0);
9811 ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9812 ASSERT(svntr_hashtab != NULL);
9813
9814 mutex_enter(&svntr_hashtab[hash].tr_lock);
9815 prv_svntrp = &svntr_hashtab[hash].tr_head;
9816 for (; (svntrp = *prv_svntrp) != NULL; prv_svntrp = &svntrp->tr_next) {
9817 ASSERT(svntrp->tr_refcnt != 0);
9818 if (svntrp->tr_vp == vp && svntrp->tr_off == off &&
9819 svntrp->tr_eoff == eoff && svntrp->tr_szc == szc) {
9820 break;
9821 }
9822 }
9823 if (svntrp == NULL) {
9824 panic("segvn_textunrepl: svntr record not found");
9825 }
9826 if (svntrp->tr_amp[lgrp_id] != svd->amp) {
9827 panic("segvn_textunrepl: amp mismatch");
9828 }
9829 svd->tr_state = SEGVN_TR_OFF;
9830 svd->amp = NULL;
9831 if (svd->svn_trprev == NULL) {
9832 ASSERT(svntrp->tr_svnhead == svd);
9833 svntrp->tr_svnhead = svd->svn_trnext;
9834 if (svntrp->tr_svnhead != NULL) {
9835 svntrp->tr_svnhead->svn_trprev = NULL;
9836 }
9837 svd->svn_trnext = NULL;
9838 } else {
9839 svd->svn_trprev->svn_trnext = svd->svn_trnext;
9840 if (svd->svn_trnext != NULL) {
9841 svd->svn_trnext->svn_trprev = svd->svn_trprev;
9842 svd->svn_trnext = NULL;
9843 }
9844 svd->svn_trprev = NULL;
9845 }
9846 if (--svntrp->tr_refcnt) {
9847 mutex_exit(&svntr_hashtab[hash].tr_lock);
9848 goto done;
9849 }
9850 *prv_svntrp = svntrp->tr_next;
9851 mutex_exit(&svntr_hashtab[hash].tr_lock);
9852 for (i = 0; i < NLGRPS_MAX; i++) {
9853 struct anon_map *amp = svntrp->tr_amp[i];
9854 if (amp == NULL) {
9855 continue;
9856 }
9857 ASSERT(amp->refcnt == 1);
9858 ASSERT(amp->swresv == size);
9859 ASSERT(amp->size == size);
9860 ASSERT(amp->a_szc == szc);
9861 if (amp->a_szc != 0) {
9862 anon_free_pages(amp->ahp, 0, size, szc);
9863 } else {
9864 anon_free(amp->ahp, 0, size);
9865 }
9866 svntrp->tr_amp[i] = NULL;
9867 ASSERT(segvn_textrepl_bytes >= size);
9868 atomic_add_long(&segvn_textrepl_bytes, -size);
9869 anon_unresv_zone(amp->swresv, NULL);
9870 amp->refcnt = 0;
9871 anonmap_free(amp);
9872 }
9873 kmem_cache_free(svntr_cache, svntrp);
9874 done:
9875 hat_unload_callback(seg->s_as->a_hat, seg->s_base, size,
9876 unload_unmap ? HAT_UNLOAD_UNMAP : 0, NULL);
9877 }
9878
9879 /*
9880 * This is called when a MAP_SHARED writable mapping is created to a vnode
9881 * that is currently used for execution (VVMEXEC flag is set). In this case we
9882 * need to prevent further use of existing replicas.
9883 */
9884 static void
9885 segvn_inval_trcache(vnode_t *vp)
9886 {
9887 ulong_t hash = SVNTR_HASH_FUNC(vp);
9888 svntr_t *svntrp;
9889
9890 ASSERT(vp->v_flag & VVMEXEC);
9891
9892 if (svntr_hashtab == NULL) {
9893 return;
9894 }
9895
9896 mutex_enter(&svntr_hashtab[hash].tr_lock);
9897 svntrp = svntr_hashtab[hash].tr_head;
9898 for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9899 ASSERT(svntrp->tr_refcnt != 0);
9900 if (svntrp->tr_vp == vp && svntrp->tr_valid) {
9901 svntrp->tr_valid = 0;
9902 }
9903 }
9904 mutex_exit(&svntr_hashtab[hash].tr_lock);
9905 }
9906
9907 static void
9908 segvn_trasync_thread(void)
9909 {
9910 callb_cpr_t cpr_info;
9911 kmutex_t cpr_lock; /* just for CPR stuff */
9912
9913 mutex_init(&cpr_lock, NULL, MUTEX_DEFAULT, NULL);
9914
9915 CALLB_CPR_INIT(&cpr_info, &cpr_lock,
9916 callb_generic_cpr, "segvn_async");
9917
9918 if (segvn_update_textrepl_interval == 0) {
9919 segvn_update_textrepl_interval = segvn_update_tr_time * hz;
9920 } else {
9921 segvn_update_textrepl_interval *= hz;
9922 }
9923 (void) timeout(segvn_trupdate_wakeup, NULL,
9924 segvn_update_textrepl_interval);
9925
9926 for (;;) {
9927 mutex_enter(&cpr_lock);
9928 CALLB_CPR_SAFE_BEGIN(&cpr_info);
9929 mutex_exit(&cpr_lock);
9930 sema_p(&segvn_trasync_sem);
9931 mutex_enter(&cpr_lock);
9932 CALLB_CPR_SAFE_END(&cpr_info, &cpr_lock);
9933 mutex_exit(&cpr_lock);
9934 segvn_trupdate();
9935 }
9936 }
9937
9938 static uint64_t segvn_lgrp_trthr_migrs_snpsht = 0;
9939
9940 static void
9941 segvn_trupdate_wakeup(void *dummy)
9942 {
9943 uint64_t cur_lgrp_trthr_migrs = lgrp_get_trthr_migrations();
9944
9945 if (cur_lgrp_trthr_migrs != segvn_lgrp_trthr_migrs_snpsht) {
9946 segvn_lgrp_trthr_migrs_snpsht = cur_lgrp_trthr_migrs;
9947 sema_v(&segvn_trasync_sem);
9948 }
9949
9950 if (!segvn_disable_textrepl_update &&
9951 segvn_update_textrepl_interval != 0) {
9952 (void) timeout(segvn_trupdate_wakeup, dummy,
9953 segvn_update_textrepl_interval);
9954 }
9955 }
9956
9957 static void
9958 segvn_trupdate(void)
9959 {
9960 ulong_t hash;
9961 svntr_t *svntrp;
9962 segvn_data_t *svd;
9963
9964 ASSERT(svntr_hashtab != NULL);
9965
9966 for (hash = 0; hash < svntr_hashtab_sz; hash++) {
9967 mutex_enter(&svntr_hashtab[hash].tr_lock);
9968 svntrp = svntr_hashtab[hash].tr_head;
9969 for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9970 ASSERT(svntrp->tr_refcnt != 0);
9971 svd = svntrp->tr_svnhead;
9972 for (; svd != NULL; svd = svd->svn_trnext) {
9973 segvn_trupdate_seg(svd->seg, svd, svntrp,
9974 hash);
9975 }
9976 }
9977 mutex_exit(&svntr_hashtab[hash].tr_lock);
9978 }
9979 }
9980
9981 static void
9982 segvn_trupdate_seg(struct seg *seg, segvn_data_t *svd, svntr_t *svntrp,
9983 ulong_t hash)
9984 {
9985 proc_t *p;
9986 lgrp_id_t lgrp_id;
9987 struct as *as;
9988 size_t size;
9989 struct anon_map *amp;
9990
9991 ASSERT(svd->vp != NULL);
9992 ASSERT(svd->vp == svntrp->tr_vp);
9993 ASSERT(svd->offset == svntrp->tr_off);
9994 ASSERT(svd->offset + seg->s_size == svntrp->tr_eoff);
9995 ASSERT(seg != NULL);
9996 ASSERT(svd->seg == seg);
9997 ASSERT(seg->s_data == (void *)svd);
9998 ASSERT(seg->s_szc == svntrp->tr_szc);
9999 ASSERT(svd->tr_state == SEGVN_TR_ON);
10000 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
10001 ASSERT(svd->amp != NULL);
10002 ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10003 ASSERT(svd->tr_policy_info.mem_lgrpid != LGRP_NONE);
10004 ASSERT(svd->tr_policy_info.mem_lgrpid < NLGRPS_MAX);
10005 ASSERT(svntrp->tr_amp[svd->tr_policy_info.mem_lgrpid] == svd->amp);
10006 ASSERT(svntrp->tr_refcnt != 0);
10007 ASSERT(mutex_owned(&svntr_hashtab[hash].tr_lock));
10008
10009 as = seg->s_as;
10010 ASSERT(as != NULL && as != &kas);
10011 p = as->a_proc;
10012 ASSERT(p != NULL);
10013 ASSERT(p->p_tr_lgrpid != LGRP_NONE);
10014 lgrp_id = p->p_t1_lgrpid;
10015 if (lgrp_id == LGRP_NONE) {
10016 return;
10017 }
10018 ASSERT(lgrp_id < NLGRPS_MAX);
10019 if (svd->tr_policy_info.mem_lgrpid == lgrp_id) {
10020 return;
10021 }
10022
10023 /*
10024 * Use tryenter locking since we are locking as/seg and svntr hash
10025 * lock in reverse from syncrounous thread order.
10026 */
10027 if (!AS_LOCK_TRYENTER(as, RW_READER)) {
10028 SEGVN_TR_ADDSTAT(nolock);
10029 if (segvn_lgrp_trthr_migrs_snpsht) {
10030 segvn_lgrp_trthr_migrs_snpsht = 0;
10031 }
10032 return;
10033 }
10034 if (!SEGVN_LOCK_TRYENTER(seg->s_as, &svd->lock, RW_WRITER)) {
10035 AS_LOCK_EXIT(as);
10036 SEGVN_TR_ADDSTAT(nolock);
10037 if (segvn_lgrp_trthr_migrs_snpsht) {
10038 segvn_lgrp_trthr_migrs_snpsht = 0;
10039 }
10040 return;
10041 }
10042 size = seg->s_size;
10043 if (svntrp->tr_amp[lgrp_id] == NULL) {
10044 size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
10045 if (trmem > segvn_textrepl_max_bytes) {
10046 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10047 AS_LOCK_EXIT(as);
10048 atomic_add_long(&segvn_textrepl_bytes, -size);
10049 SEGVN_TR_ADDSTAT(normem);
10050 return;
10051 }
10052 if (anon_try_resv_zone(size, NULL) == 0) {
10053 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10054 AS_LOCK_EXIT(as);
10055 atomic_add_long(&segvn_textrepl_bytes, -size);
10056 SEGVN_TR_ADDSTAT(noanon);
10057 return;
10058 }
10059 amp = anonmap_alloc(size, size, KM_NOSLEEP);
10060 if (amp == NULL) {
10061 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10062 AS_LOCK_EXIT(as);
10063 atomic_add_long(&segvn_textrepl_bytes, -size);
10064 anon_unresv_zone(size, NULL);
10065 SEGVN_TR_ADDSTAT(nokmem);
10066 return;
10067 }
10068 ASSERT(amp->refcnt == 1);
10069 amp->a_szc = seg->s_szc;
10070 svntrp->tr_amp[lgrp_id] = amp;
10071 }
10072 /*
10073 * We don't need to drop the bucket lock but here we give other
10074 * threads a chance. svntr and svd can't be unlinked as long as
10075 * segment lock is held as a writer and AS held as well. After we
10076 * retake bucket lock we'll continue from where we left. We'll be able
10077 * to reach the end of either list since new entries are always added
10078 * to the beginning of the lists.
10079 */
10080 mutex_exit(&svntr_hashtab[hash].tr_lock);
10081 hat_unload_callback(as->a_hat, seg->s_base, size, 0, NULL);
10082 mutex_enter(&svntr_hashtab[hash].tr_lock);
10083
10084 ASSERT(svd->tr_state == SEGVN_TR_ON);
10085 ASSERT(svd->amp != NULL);
10086 ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10087 ASSERT(svd->tr_policy_info.mem_lgrpid != lgrp_id);
10088 ASSERT(svd->amp != svntrp->tr_amp[lgrp_id]);
10089
10090 svd->tr_policy_info.mem_lgrpid = lgrp_id;
10091 svd->amp = svntrp->tr_amp[lgrp_id];
10092 p->p_tr_lgrpid = NLGRPS_MAX;
10093 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10094 AS_LOCK_EXIT(as);
10095
10096 ASSERT(svntrp->tr_refcnt != 0);
10097 ASSERT(svd->vp == svntrp->tr_vp);
10098 ASSERT(svd->tr_policy_info.mem_lgrpid == lgrp_id);
10099 ASSERT(svd->amp != NULL && svd->amp == svntrp->tr_amp[lgrp_id]);
10100 ASSERT(svd->seg == seg);
10101 ASSERT(svd->tr_state == SEGVN_TR_ON);
10102
10103 SEGVN_TR_ADDSTAT(asyncrepl);
10104 }
Unleashed OS