?  search:
summary | log | graphiclog1 | graphiclog2 | commit | commitdiff | tree | refs | edit | fork
blame | history | raw | HEAD
uts: make emu10k non-verbose
[unleashed.git] / kernel / vm / seg_vn.c
blobb924698a4efbb9c04fdf6286e0e31b1fa52c0440
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 2015, Joyent, Inc. All rights reserved.
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 *seg, void *argsp)
527 {
528 extern lgrp_mem_policy_t lgrp_mem_default_policy;
529 struct segvn_crargs *a = (struct segvn_crargs *)argsp;
530 struct segvn_data *svd;
531 size_t swresv = 0;
532 struct cred *cred;
533 struct anon_map *amp;
534 int error = 0;
535 size_t pgsz;
536 lgrp_mem_policy_t mpolicy = lgrp_mem_default_policy;
537 int use_rgn = 0;
538 int trok = 0;
539
540 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
541
542 if (a->type != MAP_PRIVATE && a->type != MAP_SHARED) {
543 panic("segvn_create type");
544 /*NOTREACHED*/
545 }
546
547 /*
548 * Check arguments. If a shared anon structure is given then
549 * it is illegal to also specify a vp.
550 */
551 if (a->amp != NULL && a->vp != NULL) {
552 panic("segvn_create anon_map");
553 /*NOTREACHED*/
554 }
555
556 if (a->type == MAP_PRIVATE && (a->flags & MAP_TEXT) &&
557 a->vp != NULL && a->prot == (PROT_USER | PROT_READ | PROT_EXEC) &&
558 segvn_use_regions) {
559 use_rgn = 1;
560 }
561
562 /* MAP_NORESERVE on a MAP_SHARED segment is meaningless. */
563 if (a->type == MAP_SHARED)
564 a->flags &= ~MAP_NORESERVE;
565
566 if (a->szc != 0) {
567 if (segvn_lpg_disable != 0 || (a->szc == AS_MAP_NO_LPOOB) ||
568 (a->amp != NULL && a->type == MAP_PRIVATE) ||
569 (a->flags & MAP_NORESERVE) || seg->s_as == &kas) {
570 a->szc = 0;
571 } else {
572 if (a->szc > segvn_maxpgszc)
573 a->szc = segvn_maxpgszc;
574 pgsz = page_get_pagesize(a->szc);
575 if (!IS_P2ALIGNED(seg->s_base, pgsz) ||
576 !IS_P2ALIGNED(seg->s_size, pgsz)) {
577 a->szc = 0;
578 } else if (a->vp != NULL) {
579 if (IS_SWAPFSVP(a->vp) || VN_ISKAS(a->vp)) {
580 /*
581 * paranoid check.
582 * hat_page_demote() is not supported
583 * on swapfs pages.
584 */
585 a->szc = 0;
586 } else if (map_addr_vacalign_check(seg->s_base,
587 a->offset & PAGEMASK)) {
588 a->szc = 0;
589 }
590 } else if (a->amp != NULL) {
591 pgcnt_t anum = btopr(a->offset);
592 pgcnt_t pgcnt = page_get_pagecnt(a->szc);
593 if (!IS_P2ALIGNED(anum, pgcnt)) {
594 a->szc = 0;
595 }
596 }
597 }
598 }
599
600 /*
601 * If segment may need private pages, reserve them now.
602 */
603 if (!(a->flags & MAP_NORESERVE) && ((a->vp == NULL && a->amp == NULL) ||
604 (a->type == MAP_PRIVATE && (a->prot & PROT_WRITE)))) {
605 if (anon_resv_zone(seg->s_size,
606 seg->s_as->a_proc->p_zone) == 0)
607 return (EAGAIN);
608 swresv = seg->s_size;
609 }
610
611 /*
612 * Reserve any mapping structures that may be required.
613 *
614 * Don't do it for segments that may use regions. It's currently a
615 * noop in the hat implementations anyway.
616 */
617 if (!use_rgn) {
618 hat_map(seg->s_as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
619 }
620
621 if (a->cred) {
622 cred = a->cred;
623 crhold(cred);
624 } else {
625 crhold(cred = CRED());
626 }
627
628 /* Inform the vnode of the new mapping */
629 if (a->vp != NULL) {
630 error = fop_addmap(a->vp, a->offset & PAGEMASK,
631 seg->s_as, seg->s_base, seg->s_size, a->prot,
632 a->maxprot, a->type, cred, NULL);
633 if (error) {
634 if (swresv != 0) {
635 anon_unresv_zone(swresv,
636 seg->s_as->a_proc->p_zone);
637 }
638 crfree(cred);
639 if (!use_rgn) {
640 hat_unload(seg->s_as->a_hat, seg->s_base,
641 seg->s_size, HAT_UNLOAD_UNMAP);
642 }
643 return (error);
644 }
645 /*
646 * svntr_hashtab will be NULL if we support shared regions.
647 */
648 trok = ((a->flags & MAP_TEXT) &&
649 (seg->s_size > textrepl_size_thresh ||
650 (a->flags & _MAP_TEXTREPL)) &&
651 lgrp_optimizations() && svntr_hashtab != NULL &&
652 a->type == MAP_PRIVATE && swresv == 0 &&
653 !(a->flags & MAP_NORESERVE) &&
654 seg->s_as != &kas && a->vp->v_type == VREG);
655
656 ASSERT(!trok || !use_rgn);
657 }
658
659 /*
660 * MAP_NORESERVE mappings don't count towards the VSZ of a process
661 * until we fault the pages in.
662 */
663 if ((a->vp == NULL || a->vp->v_type != VREG) &&
664 a->flags & MAP_NORESERVE) {
665 seg->s_as->a_resvsize -= seg->s_size;
666 }
667
668 /*
669 * If more than one segment in the address space, and they're adjacent
670 * virtually, try to concatenate them. Don't concatenate if an
671 * explicit anon_map structure was supplied (e.g., SystemV shared
672 * memory) or if we'll use text replication for this segment.
673 */
674 if (a->amp == NULL && !use_rgn && !trok) {
675 struct seg *pseg, *nseg;
676 struct segvn_data *psvd, *nsvd;
677 lgrp_mem_policy_t ppolicy, npolicy;
678 uint_t lgrp_mem_policy_flags = 0;
679
680 /*
681 * Memory policy flags (lgrp_mem_policy_flags) is valid when
682 * extending stack/heap segments.
683 */
684 if ((a->vp == NULL) && (a->type == MAP_PRIVATE) &&
685 !(a->flags & MAP_NORESERVE) && (seg->s_as != &kas)) {
686 lgrp_mem_policy_flags = a->lgrp_mem_policy_flags;
687 } else {
688 /*
689 * Get policy when not extending it from another segment
690 */
691 mpolicy = lgrp_mem_policy_default(seg->s_size, a->type);
692 }
693
694 /*
695 * First, try to concatenate the previous and new segments
696 */
697 pseg = AS_SEGPREV(seg->s_as, seg);
698 if (pseg != NULL &&
699 pseg->s_base + pseg->s_size == seg->s_base &&
700 pseg->s_ops == &segvn_ops) {
701 /*
702 * Get memory allocation policy from previous segment.
703 * When extension is specified (e.g. for heap) apply
704 * this policy to the new segment regardless of the
705 * outcome of segment concatenation. Extension occurs
706 * for non-default policy otherwise default policy is
707 * used and is based on extended segment size.
708 */
709 psvd = (struct segvn_data *)pseg->s_data;
710 ppolicy = psvd->policy_info.mem_policy;
711 if (lgrp_mem_policy_flags ==
712 LGRP_MP_FLAG_EXTEND_UP) {
713 if (ppolicy != lgrp_mem_default_policy) {
714 mpolicy = ppolicy;
715 } else {
716 mpolicy = lgrp_mem_policy_default(
717 pseg->s_size + seg->s_size,
718 a->type);
719 }
720 }
721
722 if (mpolicy == ppolicy &&
723 (pseg->s_size + seg->s_size <=
724 segvn_comb_thrshld || psvd->amp == NULL) &&
725 segvn_extend_prev(pseg, seg, a, swresv) == 0) {
726 /*
727 * success! now try to concatenate
728 * with following seg
729 */
730 crfree(cred);
731 nseg = AS_SEGNEXT(pseg->s_as, pseg);
732 if (nseg != NULL &&
733 nseg != pseg &&
734 nseg->s_ops == &segvn_ops &&
735 pseg->s_base + pseg->s_size ==
736 nseg->s_base)
737 (void) segvn_concat(pseg, nseg, 0);
738 ASSERT(pseg->s_szc == 0 ||
739 (a->szc == pseg->s_szc &&
740 IS_P2ALIGNED(pseg->s_base, pgsz) &&
741 IS_P2ALIGNED(pseg->s_size, pgsz)));
742 return (0);
743 }
744 }
745
746 /*
747 * Failed, so try to concatenate with following seg
748 */
749 nseg = AS_SEGNEXT(seg->s_as, seg);
750 if (nseg != NULL &&
751 seg->s_base + seg->s_size == nseg->s_base &&
752 nseg->s_ops == &segvn_ops) {
753 /*
754 * Get memory allocation policy from next segment.
755 * When extension is specified (e.g. for stack) apply
756 * this policy to the new segment regardless of the
757 * outcome of segment concatenation. Extension occurs
758 * for non-default policy otherwise default policy is
759 * used and is based on extended segment size.
760 */
761 nsvd = (struct segvn_data *)nseg->s_data;
762 npolicy = nsvd->policy_info.mem_policy;
763 if (lgrp_mem_policy_flags ==
764 LGRP_MP_FLAG_EXTEND_DOWN) {
765 if (npolicy != lgrp_mem_default_policy) {
766 mpolicy = npolicy;
767 } else {
768 mpolicy = lgrp_mem_policy_default(
769 nseg->s_size + seg->s_size,
770 a->type);
771 }
772 }
773
774 if (mpolicy == npolicy &&
775 segvn_extend_next(seg, nseg, a, swresv) == 0) {
776 crfree(cred);
777 ASSERT(nseg->s_szc == 0 ||
778 (a->szc == nseg->s_szc &&
779 IS_P2ALIGNED(nseg->s_base, pgsz) &&
780 IS_P2ALIGNED(nseg->s_size, pgsz)));
781 return (0);
782 }
783 }
784 }
785
786 if (a->vp != NULL) {
787 VN_HOLD(a->vp);
788 if (a->type == MAP_SHARED)
789 lgrp_shm_policy_init(NULL, a->vp);
790 }
791 svd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
792
793 seg->s_ops = &segvn_ops;
794 seg->s_data = (void *)svd;
795 seg->s_szc = a->szc;
796
797 svd->seg = seg;
798 svd->vp = a->vp;
799 /*
800 * Anonymous mappings have no backing file so the offset is meaningless.
801 */
802 svd->offset = a->vp ? (a->offset & PAGEMASK) : 0;
803 svd->prot = a->prot;
804 svd->maxprot = a->maxprot;
805 svd->pageprot = 0;
806 svd->type = a->type;
807 svd->vpage = NULL;
808 svd->cred = cred;
809 svd->advice = MADV_NORMAL;
810 svd->pageadvice = 0;
811 svd->flags = (ushort_t)a->flags;
812 svd->softlockcnt = 0;
813 svd->softlockcnt_sbase = 0;
814 svd->softlockcnt_send = 0;
815 svd->svn_inz = 0;
816 svd->rcookie = HAT_INVALID_REGION_COOKIE;
817 svd->pageswap = 0;
818
819 if (a->szc != 0 && a->vp != NULL) {
820 segvn_setvnode_mpss(a->vp);
821 }
822 if (svd->type == MAP_SHARED && svd->vp != NULL &&
823 (svd->vp->v_flag & VVMEXEC) && (svd->prot & PROT_WRITE)) {
824 ASSERT(vn_is_mapped(svd->vp, V_WRITE));
825 segvn_inval_trcache(svd->vp);
826 }
827
828 amp = a->amp;
829 if ((svd->amp = amp) == NULL) {
830 svd->anon_index = 0;
831 if (svd->type == MAP_SHARED) {
832 svd->swresv = 0;
833 /*
834 * Shared mappings to a vp need no other setup.
835 * If we have a shared mapping to an anon_map object
836 * which hasn't been allocated yet, allocate the
837 * struct now so that it will be properly shared
838 * by remembering the swap reservation there.
839 */
840 if (a->vp == NULL) {
841 svd->amp = anonmap_alloc(seg->s_size, swresv,
842 ANON_SLEEP);
843 svd->amp->a_szc = seg->s_szc;
844 }
845 } else {
846 /*
847 * Private mapping (with or without a vp).
848 * Allocate anon_map when needed.
849 */
850 svd->swresv = swresv;
851 }
852 } else {
853 pgcnt_t anon_num;
854
855 /*
856 * Mapping to an existing anon_map structure without a vp.
857 * For now we will insure that the segment size isn't larger
858 * than the size - offset gives us. Later on we may wish to
859 * have the anon array dynamically allocated itself so that
860 * we don't always have to allocate all the anon pointer slots.
861 * This of course involves adding extra code to check that we
862 * aren't trying to use an anon pointer slot beyond the end
863 * of the currently allocated anon array.
864 */
865 if ((amp->size - a->offset) < seg->s_size) {
866 panic("segvn_create anon_map size");
867 /*NOTREACHED*/
868 }
869
870 anon_num = btopr(a->offset);
871
872 if (a->type == MAP_SHARED) {
873 /*
874 * SHARED mapping to a given anon_map.
875 */
876 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
877 amp->refcnt++;
878 if (a->szc > amp->a_szc) {
879 amp->a_szc = a->szc;
880 }
881 ANON_LOCK_EXIT(&amp->a_rwlock);
882 svd->anon_index = anon_num;
883 svd->swresv = 0;
884 } else {
885 /*
886 * PRIVATE mapping to a given anon_map.
887 * Make sure that all the needed anon
888 * structures are created (so that we will
889 * share the underlying pages if nothing
890 * is written by this mapping) and then
891 * duplicate the anon array as is done
892 * when a privately mapped segment is dup'ed.
893 */
894 struct anon *ap;
895 caddr_t addr;
896 caddr_t eaddr;
897 ulong_t anon_idx;
898 int hat_flag = HAT_LOAD;
899
900 if (svd->flags & MAP_TEXT) {
901 hat_flag |= HAT_LOAD_TEXT;
902 }
903
904 svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
905 svd->amp->a_szc = seg->s_szc;
906 svd->anon_index = 0;
907 svd->swresv = swresv;
908
909 /*
910 * Prevent 2 threads from allocating anon
911 * slots simultaneously.
912 */
913 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
914 eaddr = seg->s_base + seg->s_size;
915
916 for (anon_idx = anon_num, addr = seg->s_base;
917 addr < eaddr; addr += PAGESIZE, anon_idx++) {
918 page_t *pp;
919
920 if ((ap = anon_get_ptr(amp->ahp,
921 anon_idx)) != NULL)
922 continue;
923
924 /*
925 * Allocate the anon struct now.
926 * Might as well load up translation
927 * to the page while we're at it...
928 */
929 pp = anon_zero(seg, addr, &ap, cred);
930 if (ap == NULL || pp == NULL) {
931 panic("segvn_create anon_zero");
932 /*NOTREACHED*/
933 }
934
935 /*
936 * Re-acquire the anon_map lock and
937 * initialize the anon array entry.
938 */
939 ASSERT(anon_get_ptr(amp->ahp,
940 anon_idx) == NULL);
941 (void) anon_set_ptr(amp->ahp, anon_idx, ap,
942 ANON_SLEEP);
943
944 ASSERT(seg->s_szc == 0);
945 ASSERT(!IS_VMODSORT(pp->p_vnode));
946
947 ASSERT(use_rgn == 0);
948 hat_memload(seg->s_as->a_hat, addr, pp,
949 svd->prot & ~PROT_WRITE, hat_flag);
950
951 page_unlock(pp);
952 }
953 ASSERT(seg->s_szc == 0);
954 anon_dup(amp->ahp, anon_num, svd->amp->ahp,
955 0, seg->s_size);
956 ANON_LOCK_EXIT(&amp->a_rwlock);
957 }
958 }
959
960 /*
961 * Set default memory allocation policy for segment
962 *
963 * Always set policy for private memory at least for initialization
964 * even if this is a shared memory segment
965 */
966 (void) lgrp_privm_policy_set(mpolicy, &svd->policy_info, seg->s_size);
967
968 if (svd->type == MAP_SHARED)
969 (void) lgrp_shm_policy_set(mpolicy, svd->amp, svd->anon_index,
970 svd->vp, svd->offset, seg->s_size);
971
972 if (use_rgn) {
973 ASSERT(!trok);
974 ASSERT(svd->amp == NULL);
975 svd->rcookie = hat_join_region(seg->s_as->a_hat, seg->s_base,
976 seg->s_size, (void *)svd->vp, svd->offset, svd->prot,
977 (uchar_t)seg->s_szc, segvn_hat_rgn_unload_callback,
978 HAT_REGION_TEXT);
979 }
980
981 ASSERT(!trok || !(svd->prot & PROT_WRITE));
982 svd->tr_state = trok ? SEGVN_TR_INIT : SEGVN_TR_OFF;
983
984 return (0);
985 }
986
987 /*
988 * Concatenate two existing segments, if possible.
989 * Return 0 on success, -1 if two segments are not compatible
990 * or -2 on memory allocation failure.
991 * If amp_cat == 1 then try and concat segments with anon maps
992 */
993 static int
994 segvn_concat(struct seg *seg1, struct seg *seg2, int amp_cat)
995 {
996 struct segvn_data *svd1 = seg1->s_data;
997 struct segvn_data *svd2 = seg2->s_data;
998 struct anon_map *amp1 = svd1->amp;
999 struct anon_map *amp2 = svd2->amp;
1000 struct vpage *vpage1 = svd1->vpage;
1001 struct vpage *vpage2 = svd2->vpage, *nvpage = NULL;
1002 size_t size, nvpsize;
1003 pgcnt_t npages1, npages2;
1004
1005 ASSERT(seg1->s_as && seg2->s_as && seg1->s_as == seg2->s_as);
1006 ASSERT(AS_WRITE_HELD(seg1->s_as));
1007 ASSERT(seg1->s_ops == seg2->s_ops);
1008
1009 if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie) ||
1010 HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1011 return (-1);
1012 }
1013
1014 /* both segments exist, try to merge them */
1015 #define incompat(x) (svd1->x != svd2->x)
1016 if (incompat(vp) || incompat(maxprot) ||
1017 (!svd1->pageadvice && !svd2->pageadvice && incompat(advice)) ||
1018 (!svd1->pageprot && !svd2->pageprot && incompat(prot)) ||
1019 incompat(type) || incompat(cred) || incompat(flags) ||
1020 seg1->s_szc != seg2->s_szc || incompat(policy_info.mem_policy) ||
1021 (svd2->softlockcnt > 0) || svd1->softlockcnt_send > 0)
1022 return (-1);
1023 #undef incompat
1024
1025 /*
1026 * vp == NULL implies zfod, offset doesn't matter
1027 */
1028 if (svd1->vp != NULL &&
1029 svd1->offset + seg1->s_size != svd2->offset) {
1030 return (-1);
1031 }
1032
1033 /*
1034 * Don't concatenate if either segment uses text replication.
1035 */
1036 if (svd1->tr_state != SEGVN_TR_OFF || svd2->tr_state != SEGVN_TR_OFF) {
1037 return (-1);
1038 }
1039
1040 /*
1041 * Fail early if we're not supposed to concatenate
1042 * segments with non NULL amp.
1043 */
1044 if (amp_cat == 0 && (amp1 != NULL || amp2 != NULL)) {
1045 return (-1);
1046 }
1047
1048 if (svd1->vp == NULL && svd1->type == MAP_SHARED) {
1049 if (amp1 != amp2) {
1050 return (-1);
1051 }
1052 if (amp1 != NULL && svd1->anon_index + btop(seg1->s_size) !=
1053 svd2->anon_index) {
1054 return (-1);
1055 }
1056 ASSERT(amp1 == NULL || amp1->refcnt >= 2);
1057 }
1058
1059 /*
1060 * If either seg has vpages, create a new merged vpage array.
1061 */
1062 if (vpage1 != NULL || vpage2 != NULL) {
1063 struct vpage *vp, *evp;
1064
1065 npages1 = seg_pages(seg1);
1066 npages2 = seg_pages(seg2);
1067 nvpsize = vpgtob(npages1 + npages2);
1068
1069 if ((nvpage = kmem_zalloc(nvpsize, KM_NOSLEEP)) == NULL) {
1070 return (-2);
1071 }
1072
1073 if (vpage1 != NULL) {
1074 bcopy(vpage1, nvpage, vpgtob(npages1));
1075 } else {
1076 evp = nvpage + npages1;
1077 for (vp = nvpage; vp < evp; vp++) {
1078 VPP_SETPROT(vp, svd1->prot);
1079 VPP_SETADVICE(vp, svd1->advice);
1080 }
1081 }
1082
1083 if (vpage2 != NULL) {
1084 bcopy(vpage2, nvpage + npages1, vpgtob(npages2));
1085 } else {
1086 evp = nvpage + npages1 + npages2;
1087 for (vp = nvpage + npages1; vp < evp; vp++) {
1088 VPP_SETPROT(vp, svd2->prot);
1089 VPP_SETADVICE(vp, svd2->advice);
1090 }
1091 }
1092
1093 if (svd2->pageswap && (!svd1->pageswap && svd1->swresv)) {
1094 ASSERT(svd1->swresv == seg1->s_size);
1095 ASSERT(!(svd1->flags & MAP_NORESERVE));
1096 ASSERT(!(svd2->flags & MAP_NORESERVE));
1097 evp = nvpage + npages1;
1098 for (vp = nvpage; vp < evp; vp++) {
1099 VPP_SETSWAPRES(vp);
1100 }
1101 }
1102
1103 if (svd1->pageswap && (!svd2->pageswap && svd2->swresv)) {
1104 ASSERT(svd2->swresv == seg2->s_size);
1105 ASSERT(!(svd1->flags & MAP_NORESERVE));
1106 ASSERT(!(svd2->flags & MAP_NORESERVE));
1107 vp = nvpage + npages1;
1108 evp = vp + npages2;
1109 for (; vp < evp; vp++) {
1110 VPP_SETSWAPRES(vp);
1111 }
1112 }
1113 }
1114 ASSERT((vpage1 != NULL || vpage2 != NULL) ||
1115 (svd1->pageswap == 0 && svd2->pageswap == 0));
1116
1117 /*
1118 * If either segment has private pages, create a new merged anon
1119 * array. If mergeing shared anon segments just decrement anon map's
1120 * refcnt.
1121 */
1122 if (amp1 != NULL && svd1->type == MAP_SHARED) {
1123 ASSERT(amp1 == amp2 && svd1->vp == NULL);
1124 ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1125 ASSERT(amp1->refcnt >= 2);
1126 amp1->refcnt--;
1127 ANON_LOCK_EXIT(&amp1->a_rwlock);
1128 svd2->amp = NULL;
1129 } else if (amp1 != NULL || amp2 != NULL) {
1130 struct anon_hdr *nahp;
1131 struct anon_map *namp = NULL;
1132 size_t asize;
1133
1134 ASSERT(svd1->type == MAP_PRIVATE);
1135
1136 asize = seg1->s_size + seg2->s_size;
1137 if ((nahp = anon_create(btop(asize), ANON_NOSLEEP)) == NULL) {
1138 if (nvpage != NULL) {
1139 kmem_free(nvpage, nvpsize);
1140 }
1141 return (-2);
1142 }
1143 if (amp1 != NULL) {
1144 /*
1145 * XXX anon rwlock is not really needed because
1146 * this is a private segment and we are writers.
1147 */
1148 ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1149 ASSERT(amp1->refcnt == 1);
1150 if (anon_copy_ptr(amp1->ahp, svd1->anon_index,
1151 nahp, 0, btop(seg1->s_size), ANON_NOSLEEP)) {
1152 anon_release(nahp, btop(asize));
1153 ANON_LOCK_EXIT(&amp1->a_rwlock);
1154 if (nvpage != NULL) {
1155 kmem_free(nvpage, nvpsize);
1156 }
1157 return (-2);
1158 }
1159 }
1160 if (amp2 != NULL) {
1161 ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1162 ASSERT(amp2->refcnt == 1);
1163 if (anon_copy_ptr(amp2->ahp, svd2->anon_index,
1164 nahp, btop(seg1->s_size), btop(seg2->s_size),
1165 ANON_NOSLEEP)) {
1166 anon_release(nahp, btop(asize));
1167 ANON_LOCK_EXIT(&amp2->a_rwlock);
1168 if (amp1 != NULL) {
1169 ANON_LOCK_EXIT(&amp1->a_rwlock);
1170 }
1171 if (nvpage != NULL) {
1172 kmem_free(nvpage, nvpsize);
1173 }
1174 return (-2);
1175 }
1176 }
1177 if (amp1 != NULL) {
1178 namp = amp1;
1179 anon_release(amp1->ahp, btop(amp1->size));
1180 }
1181 if (amp2 != NULL) {
1182 if (namp == NULL) {
1183 ASSERT(amp1 == NULL);
1184 namp = amp2;
1185 anon_release(amp2->ahp, btop(amp2->size));
1186 } else {
1187 amp2->refcnt--;
1188 ANON_LOCK_EXIT(&amp2->a_rwlock);
1189 anonmap_free(amp2);
1190 }
1191 svd2->amp = NULL; /* needed for seg_free */
1192 }
1193 namp->ahp = nahp;
1194 namp->size = asize;
1195 svd1->amp = namp;
1196 svd1->anon_index = 0;
1197 ANON_LOCK_EXIT(&namp->a_rwlock);
1198 }
1199 /*
1200 * Now free the old vpage structures.
1201 */
1202 if (nvpage != NULL) {
1203 if (vpage1 != NULL) {
1204 kmem_free(vpage1, vpgtob(npages1));
1205 }
1206 if (vpage2 != NULL) {
1207 svd2->vpage = NULL;
1208 kmem_free(vpage2, vpgtob(npages2));
1209 }
1210 if (svd2->pageprot) {
1211 svd1->pageprot = 1;
1212 }
1213 if (svd2->pageadvice) {
1214 svd1->pageadvice = 1;
1215 }
1216 if (svd2->pageswap) {
1217 svd1->pageswap = 1;
1218 }
1219 svd1->vpage = nvpage;
1220 }
1221
1222 /* all looks ok, merge segments */
1223 svd1->swresv += svd2->swresv;
1224 svd2->swresv = 0; /* so seg_free doesn't release swap space */
1225 size = seg2->s_size;
1226 seg_free(seg2);
1227 seg1->s_size += size;
1228 return (0);
1229 }
1230
1231 /*
1232 * Extend the previous segment (seg1) to include the
1233 * new segment (seg2 + a), if possible.
1234 * Return 0 on success.
1235 */
1236 static int
1237 segvn_extend_prev(seg1, seg2, a, swresv)
1238 struct seg *seg1, *seg2;
1239 struct segvn_crargs *a;
1240 size_t swresv;
1241 {
1242 struct segvn_data *svd1 = (struct segvn_data *)seg1->s_data;
1243 size_t size;
1244 struct anon_map *amp1;
1245 struct vpage *new_vpage;
1246
1247 /*
1248 * We don't need any segment level locks for "segvn" data
1249 * since the address space is "write" locked.
1250 */
1251 ASSERT(seg1->s_as && AS_WRITE_HELD(seg1->s_as));
1252
1253 if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie)) {
1254 return (-1);
1255 }
1256
1257 /* second segment is new, try to extend first */
1258 /* XXX - should also check cred */
1259 if (svd1->vp != a->vp || svd1->maxprot != a->maxprot ||
1260 (!svd1->pageprot && (svd1->prot != a->prot)) ||
1261 svd1->type != a->type || svd1->flags != a->flags ||
1262 seg1->s_szc != a->szc || svd1->softlockcnt_send > 0)
1263 return (-1);
1264
1265 /* vp == NULL implies zfod, offset doesn't matter */
1266 if (svd1->vp != NULL &&
1267 svd1->offset + seg1->s_size != (a->offset & PAGEMASK))
1268 return (-1);
1269
1270 if (svd1->tr_state != SEGVN_TR_OFF) {
1271 return (-1);
1272 }
1273
1274 amp1 = svd1->amp;
1275 if (amp1) {
1276 pgcnt_t newpgs;
1277
1278 /*
1279 * Segment has private pages, can data structures
1280 * be expanded?
1281 *
1282 * Acquire the anon_map lock to prevent it from changing,
1283 * if it is shared. This ensures that the anon_map
1284 * will not change while a thread which has a read/write
1285 * lock on an address space references it.
1286 * XXX - Don't need the anon_map lock at all if "refcnt"
1287 * is 1.
1288 *
1289 * Can't grow a MAP_SHARED segment with an anonmap because
1290 * there may be existing anon slots where we want to extend
1291 * the segment and we wouldn't know what to do with them
1292 * (e.g., for tmpfs right thing is to just leave them there,
1293 * for /dev/zero they should be cleared out).
1294 */
1295 if (svd1->type == MAP_SHARED)
1296 return (-1);
1297
1298 ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1299 if (amp1->refcnt > 1) {
1300 ANON_LOCK_EXIT(&amp1->a_rwlock);
1301 return (-1);
1302 }
1303 newpgs = anon_grow(amp1->ahp, &svd1->anon_index,
1304 btop(seg1->s_size), btop(seg2->s_size), ANON_NOSLEEP);
1305
1306 if (newpgs == 0) {
1307 ANON_LOCK_EXIT(&amp1->a_rwlock);
1308 return (-1);
1309 }
1310 amp1->size = ptob(newpgs);
1311 ANON_LOCK_EXIT(&amp1->a_rwlock);
1312 }
1313 if (svd1->vpage != NULL) {
1314 struct vpage *vp, *evp;
1315 new_vpage =
1316 kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1317 KM_NOSLEEP);
1318 if (new_vpage == NULL)
1319 return (-1);
1320 bcopy(svd1->vpage, new_vpage, vpgtob(seg_pages(seg1)));
1321 kmem_free(svd1->vpage, vpgtob(seg_pages(seg1)));
1322 svd1->vpage = new_vpage;
1323
1324 vp = new_vpage + seg_pages(seg1);
1325 evp = vp + seg_pages(seg2);
1326 for (; vp < evp; vp++)
1327 VPP_SETPROT(vp, a->prot);
1328 if (svd1->pageswap && swresv) {
1329 ASSERT(!(svd1->flags & MAP_NORESERVE));
1330 ASSERT(swresv == seg2->s_size);
1331 vp = new_vpage + seg_pages(seg1);
1332 for (; vp < evp; vp++) {
1333 VPP_SETSWAPRES(vp);
1334 }
1335 }
1336 }
1337 ASSERT(svd1->vpage != NULL || svd1->pageswap == 0);
1338 size = seg2->s_size;
1339 seg_free(seg2);
1340 seg1->s_size += size;
1341 svd1->swresv += swresv;
1342 if (svd1->pageprot && (a->prot & PROT_WRITE) &&
1343 svd1->type == MAP_SHARED && svd1->vp != NULL &&
1344 (svd1->vp->v_flag & VVMEXEC)) {
1345 ASSERT(vn_is_mapped(svd1->vp, V_WRITE));
1346 segvn_inval_trcache(svd1->vp);
1347 }
1348 return (0);
1349 }
1350
1351 /*
1352 * Extend the next segment (seg2) to include the
1353 * new segment (seg1 + a), if possible.
1354 * Return 0 on success.
1355 */
1356 static int
1357 segvn_extend_next(
1358 struct seg *seg1,
1359 struct seg *seg2,
1360 struct segvn_crargs *a,
1361 size_t swresv)
1362 {
1363 struct segvn_data *svd2 = (struct segvn_data *)seg2->s_data;
1364 size_t size;
1365 struct anon_map *amp2;
1366 struct vpage *new_vpage;
1367
1368 /*
1369 * We don't need any segment level locks for "segvn" data
1370 * since the address space is "write" locked.
1371 */
1372 ASSERT(seg2->s_as && AS_WRITE_HELD(seg2->s_as));
1373
1374 if (HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1375 return (-1);
1376 }
1377
1378 /* first segment is new, try to extend second */
1379 /* XXX - should also check cred */
1380 if (svd2->vp != a->vp || svd2->maxprot != a->maxprot ||
1381 (!svd2->pageprot && (svd2->prot != a->prot)) ||
1382 svd2->type != a->type || svd2->flags != a->flags ||
1383 seg2->s_szc != a->szc || svd2->softlockcnt_sbase > 0)
1384 return (-1);
1385 /* vp == NULL implies zfod, offset doesn't matter */
1386 if (svd2->vp != NULL &&
1387 (a->offset & PAGEMASK) + seg1->s_size != svd2->offset)
1388 return (-1);
1389
1390 if (svd2->tr_state != SEGVN_TR_OFF) {
1391 return (-1);
1392 }
1393
1394 amp2 = svd2->amp;
1395 if (amp2) {
1396 pgcnt_t newpgs;
1397
1398 /*
1399 * Segment has private pages, can data structures
1400 * be expanded?
1401 *
1402 * Acquire the anon_map lock to prevent it from changing,
1403 * if it is shared. This ensures that the anon_map
1404 * will not change while a thread which has a read/write
1405 * lock on an address space references it.
1406 *
1407 * XXX - Don't need the anon_map lock at all if "refcnt"
1408 * is 1.
1409 */
1410 if (svd2->type == MAP_SHARED)
1411 return (-1);
1412
1413 ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1414 if (amp2->refcnt > 1) {
1415 ANON_LOCK_EXIT(&amp2->a_rwlock);
1416 return (-1);
1417 }
1418 newpgs = anon_grow(amp2->ahp, &svd2->anon_index,
1419 btop(seg2->s_size), btop(seg1->s_size),
1420 ANON_NOSLEEP | ANON_GROWDOWN);
1421
1422 if (newpgs == 0) {
1423 ANON_LOCK_EXIT(&amp2->a_rwlock);
1424 return (-1);
1425 }
1426 amp2->size = ptob(newpgs);
1427 ANON_LOCK_EXIT(&amp2->a_rwlock);
1428 }
1429 if (svd2->vpage != NULL) {
1430 struct vpage *vp, *evp;
1431 new_vpage =
1432 kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1433 KM_NOSLEEP);
1434 if (new_vpage == NULL) {
1435 /* Not merging segments so adjust anon_index back */
1436 if (amp2)
1437 svd2->anon_index += seg_pages(seg1);
1438 return (-1);
1439 }
1440 bcopy(svd2->vpage, new_vpage + seg_pages(seg1),
1441 vpgtob(seg_pages(seg2)));
1442 kmem_free(svd2->vpage, vpgtob(seg_pages(seg2)));
1443 svd2->vpage = new_vpage;
1444
1445 vp = new_vpage;
1446 evp = vp + seg_pages(seg1);
1447 for (; vp < evp; vp++)
1448 VPP_SETPROT(vp, a->prot);
1449 if (svd2->pageswap && swresv) {
1450 ASSERT(!(svd2->flags & MAP_NORESERVE));
1451 ASSERT(swresv == seg1->s_size);
1452 vp = new_vpage;
1453 for (; vp < evp; vp++) {
1454 VPP_SETSWAPRES(vp);
1455 }
1456 }
1457 }
1458 ASSERT(svd2->vpage != NULL || svd2->pageswap == 0);
1459 size = seg1->s_size;
1460 seg_free(seg1);
1461 seg2->s_size += size;
1462 seg2->s_base -= size;
1463 svd2->offset -= size;
1464 svd2->swresv += swresv;
1465 if (svd2->pageprot && (a->prot & PROT_WRITE) &&
1466 svd2->type == MAP_SHARED && svd2->vp != NULL &&
1467 (svd2->vp->v_flag & VVMEXEC)) {
1468 ASSERT(vn_is_mapped(svd2->vp, V_WRITE));
1469 segvn_inval_trcache(svd2->vp);
1470 }
1471 return (0);
1472 }
1473
1474 /*
1475 * Duplicate all the pages in the segment. This may break COW sharing for a
1476 * given page. If the page is marked with inherit zero set, then instead of
1477 * duplicating the page, we zero the page.
1478 */
1479 static int
1480 segvn_dup_pages(struct seg *seg, struct seg *newseg)
1481 {
1482 int error;
1483 uint_t prot;
1484 page_t *pp;
1485 struct anon *ap, *newap;
1486 size_t i;
1487 caddr_t addr;
1488
1489 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1490 struct segvn_data *newsvd = (struct segvn_data *)newseg->s_data;
1491 ulong_t old_idx = svd->anon_index;
1492 ulong_t new_idx = 0;
1493
1494 i = btopr(seg->s_size);
1495 addr = seg->s_base;
1496
1497 /*
1498 * XXX break cow sharing using PAGESIZE
1499 * pages. They will be relocated into larger
1500 * pages at fault time.
1501 */
1502 while (i-- > 0) {
1503 if ((ap = anon_get_ptr(svd->amp->ahp, old_idx)) != NULL) {
1504 struct vpage *vpp;
1505
1506 vpp = &svd->vpage[seg_page(seg, addr)];
1507
1508 /*
1509 * prot need not be computed below 'cause anon_private
1510 * is going to ignore it anyway as child doesn't inherit
1511 * pagelock from parent.
1512 */
1513 prot = svd->pageprot ? VPP_PROT(vpp) : svd->prot;
1514
1515 /*
1516 * Check whether we should zero this or dup it.
1517 */
1518 if (svd->svn_inz == SEGVN_INZ_ALL ||
1519 (svd->svn_inz == SEGVN_INZ_VPP &&
1520 VPP_ISINHZERO(vpp))) {
1521 pp = anon_zero(newseg, addr, &newap,
1522 newsvd->cred);
1523 } else {
1524 page_t *anon_pl[1+1];
1525 uint_t vpprot;
1526 error = anon_getpage(&ap, &vpprot, anon_pl,
1527 PAGESIZE, seg, addr, S_READ, svd->cred);
1528 if (error != 0)
1529 return (error);
1530
1531 pp = anon_private(&newap, newseg, addr, prot,
1532 anon_pl[0], 0, newsvd->cred);
1533 }
1534 if (pp == NULL) {
1535 return (ENOMEM);
1536 }
1537 (void) anon_set_ptr(newsvd->amp->ahp, new_idx, newap,
1538 ANON_SLEEP);
1539 page_unlock(pp);
1540 }
1541 addr += PAGESIZE;
1542 old_idx++;
1543 new_idx++;
1544 }
1545
1546 return (0);
1547 }
1548
1549 static int
1550 segvn_dup(struct seg *seg, struct seg *newseg)
1551 {
1552 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1553 struct segvn_data *newsvd;
1554 pgcnt_t npages = seg_pages(seg);
1555 int error = 0;
1556 size_t len;
1557 struct anon_map *amp;
1558
1559 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1560 ASSERT(newseg->s_as->a_proc->p_parent == curproc);
1561
1562 /*
1563 * If segment has anon reserved, reserve more for the new seg.
1564 * For a MAP_NORESERVE segment swresv will be a count of all the
1565 * allocated anon slots; thus we reserve for the child as many slots
1566 * as the parent has allocated. This semantic prevents the child or
1567 * parent from dieing during a copy-on-write fault caused by trying
1568 * to write a shared pre-existing anon page.
1569 */
1570 if ((len = svd->swresv) != 0) {
1571 if (anon_resv(svd->swresv) == 0)
1572 return (ENOMEM);
1573 }
1574
1575 newsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
1576
1577 newseg->s_ops = &segvn_ops;
1578 newseg->s_data = (void *)newsvd;
1579 newseg->s_szc = seg->s_szc;
1580
1581 newsvd->seg = newseg;
1582 if ((newsvd->vp = svd->vp) != NULL) {
1583 VN_HOLD(svd->vp);
1584 if (svd->type == MAP_SHARED)
1585 lgrp_shm_policy_init(NULL, svd->vp);
1586 }
1587 newsvd->offset = svd->offset;
1588 newsvd->prot = svd->prot;
1589 newsvd->maxprot = svd->maxprot;
1590 newsvd->pageprot = svd->pageprot;
1591 newsvd->type = svd->type;
1592 newsvd->cred = svd->cred;
1593 crhold(newsvd->cred);
1594 newsvd->advice = svd->advice;
1595 newsvd->pageadvice = svd->pageadvice;
1596 newsvd->svn_inz = svd->svn_inz;
1597 newsvd->swresv = svd->swresv;
1598 newsvd->pageswap = svd->pageswap;
1599 newsvd->flags = svd->flags;
1600 newsvd->softlockcnt = 0;
1601 newsvd->softlockcnt_sbase = 0;
1602 newsvd->softlockcnt_send = 0;
1603 newsvd->policy_info = svd->policy_info;
1604 newsvd->rcookie = HAT_INVALID_REGION_COOKIE;
1605
1606 if ((amp = svd->amp) == NULL || svd->tr_state == SEGVN_TR_ON) {
1607 /*
1608 * Not attaching to a shared anon object.
1609 */
1610 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie) ||
1611 svd->tr_state == SEGVN_TR_OFF);
1612 if (svd->tr_state == SEGVN_TR_ON) {
1613 ASSERT(newsvd->vp != NULL && amp != NULL);
1614 newsvd->tr_state = SEGVN_TR_INIT;
1615 } else {
1616 newsvd->tr_state = svd->tr_state;
1617 }
1618 newsvd->amp = NULL;
1619 newsvd->anon_index = 0;
1620 } else {
1621 /* regions for now are only used on pure vnode segments */
1622 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
1623 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1624 newsvd->tr_state = SEGVN_TR_OFF;
1625 if (svd->type == MAP_SHARED) {
1626 ASSERT(svd->svn_inz == SEGVN_INZ_NONE);
1627 newsvd->amp = amp;
1628 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1629 amp->refcnt++;
1630 ANON_LOCK_EXIT(&amp->a_rwlock);
1631 newsvd->anon_index = svd->anon_index;
1632 } else {
1633 int reclaim = 1;
1634
1635 /*
1636 * Allocate and initialize new anon_map structure.
1637 */
1638 newsvd->amp = anonmap_alloc(newseg->s_size, 0,
1639 ANON_SLEEP);
1640 newsvd->amp->a_szc = newseg->s_szc;
1641 newsvd->anon_index = 0;
1642 ASSERT(svd->svn_inz == SEGVN_INZ_NONE ||
1643 svd->svn_inz == SEGVN_INZ_ALL ||
1644 svd->svn_inz == SEGVN_INZ_VPP);
1645
1646 /*
1647 * We don't have to acquire the anon_map lock
1648 * for the new segment (since it belongs to an
1649 * address space that is still not associated
1650 * with any process), or the segment in the old
1651 * address space (since all threads in it
1652 * are stopped while duplicating the address space).
1653 */
1654
1655 /*
1656 * The goal of the following code is to make sure that
1657 * softlocked pages do not end up as copy on write
1658 * pages. This would cause problems where one
1659 * thread writes to a page that is COW and a different
1660 * thread in the same process has softlocked it. The
1661 * softlock lock would move away from this process
1662 * because the write would cause this process to get
1663 * a copy (without the softlock).
1664 *
1665 * The strategy here is to just break the
1666 * sharing on pages that could possibly be
1667 * softlocked.
1668 *
1669 * In addition, if any pages have been marked that they
1670 * should be inherited as zero, then we immediately go
1671 * ahead and break COW and zero them. In the case of a
1672 * softlocked page that should be inherited zero, we
1673 * break COW and just get a zero page.
1674 */
1675 retry:
1676 if (svd->softlockcnt ||
1677 svd->svn_inz != SEGVN_INZ_NONE) {
1678 /*
1679 * The softlock count might be non zero
1680 * because some pages are still stuck in the
1681 * cache for lazy reclaim. Flush the cache
1682 * now. This should drop the count to zero.
1683 * [or there is really I/O going on to these
1684 * pages]. Note, we have the writers lock so
1685 * nothing gets inserted during the flush.
1686 */
1687 if (svd->softlockcnt && reclaim == 1) {
1688 segvn_purge(seg);
1689 reclaim = 0;
1690 goto retry;
1691 }
1692
1693 error = segvn_dup_pages(seg, newseg);
1694 if (error != 0) {
1695 newsvd->vpage = NULL;
1696 goto out;
1697 }
1698 } else { /* common case */
1699 if (seg->s_szc != 0) {
1700 /*
1701 * If at least one of anon slots of a
1702 * large page exists then make sure
1703 * all anon slots of a large page
1704 * exist to avoid partial cow sharing
1705 * of a large page in the future.
1706 */
1707 anon_dup_fill_holes(amp->ahp,
1708 svd->anon_index, newsvd->amp->ahp,
1709 0, seg->s_size, seg->s_szc,
1710 svd->vp != NULL);
1711 } else {
1712 anon_dup(amp->ahp, svd->anon_index,
1713 newsvd->amp->ahp, 0, seg->s_size);
1714 }
1715
1716 hat_clrattr(seg->s_as->a_hat, seg->s_base,
1717 seg->s_size, PROT_WRITE);
1718 }
1719 }
1720 }
1721 /*
1722 * If necessary, create a vpage structure for the new segment.
1723 * Do not copy any page lock indications.
1724 */
1725 if (svd->vpage != NULL) {
1726 uint_t i;
1727 struct vpage *ovp = svd->vpage;
1728 struct vpage *nvp;
1729
1730 nvp = newsvd->vpage =
1731 kmem_alloc(vpgtob(npages), KM_SLEEP);
1732 for (i = 0; i < npages; i++) {
1733 *nvp = *ovp++;
1734 VPP_CLRPPLOCK(nvp++);
1735 }
1736 } else
1737 newsvd->vpage = NULL;
1738
1739 /* Inform the vnode of the new mapping */
1740 if (newsvd->vp != NULL) {
1741 error = fop_addmap(newsvd->vp, (offset_t)newsvd->offset,
1742 newseg->s_as, newseg->s_base, newseg->s_size, newsvd->prot,
1743 newsvd->maxprot, newsvd->type, newsvd->cred, NULL);
1744 }
1745 out:
1746 if (error == 0 && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1747 ASSERT(newsvd->amp == NULL);
1748 ASSERT(newsvd->tr_state == SEGVN_TR_OFF);
1749 newsvd->rcookie = svd->rcookie;
1750 hat_dup_region(newseg->s_as->a_hat, newsvd->rcookie);
1751 }
1752 return (error);
1753 }
1754
1755
1756 /*
1757 * callback function to invoke free_vp_pages() for only those pages actually
1758 * processed by the HAT when a shared region is destroyed.
1759 */
1760 extern int free_pages;
1761
1762 static void
1763 segvn_hat_rgn_unload_callback(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
1764 size_t r_size, void *r_obj, uoff_t r_objoff)
1765 {
1766 uoff_t off;
1767 size_t len;
1768 vnode_t *vp = (vnode_t *)r_obj;
1769
1770 ASSERT(eaddr > saddr);
1771 ASSERT(saddr >= r_saddr);
1772 ASSERT(saddr < r_saddr + r_size);
1773 ASSERT(eaddr > r_saddr);
1774 ASSERT(eaddr <= r_saddr + r_size);
1775 ASSERT(vp != NULL);
1776
1777 if (!free_pages) {
1778 return;
1779 }
1780
1781 len = eaddr - saddr;
1782 off = (saddr - r_saddr) + r_objoff;
1783 free_vp_pages(vp, off, len);
1784 }
1785
1786 /*
1787 * callback function used by segvn_unmap to invoke free_vp_pages() for only
1788 * those pages actually processed by the HAT
1789 */
1790 static void
1791 segvn_hat_unload_callback(hat_callback_t *cb)
1792 {
1793 struct seg *seg = cb->hcb_data;
1794 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1795 size_t len;
1796 uoff_t off;
1797
1798 ASSERT(svd->vp != NULL);
1799 ASSERT(cb->hcb_end_addr > cb->hcb_start_addr);
1800 ASSERT(cb->hcb_start_addr >= seg->s_base);
1801
1802 len = cb->hcb_end_addr - cb->hcb_start_addr;
1803 off = cb->hcb_start_addr - seg->s_base;
1804 free_vp_pages(svd->vp, svd->offset + off, len);
1805 }
1806
1807 /*
1808 * This function determines the number of bytes of swap reserved by
1809 * a segment for which per-page accounting is present. It is used to
1810 * calculate the correct value of a segvn_data's swresv.
1811 */
1812 static size_t
1813 segvn_count_swap_by_vpages(struct seg *seg)
1814 {
1815 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1816 struct vpage *vp, *evp;
1817 size_t nswappages = 0;
1818
1819 ASSERT(svd->pageswap);
1820 ASSERT(svd->vpage != NULL);
1821
1822 evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
1823
1824 for (vp = svd->vpage; vp < evp; vp++) {
1825 if (VPP_ISSWAPRES(vp))
1826 nswappages++;
1827 }
1828
1829 return (nswappages << PAGESHIFT);
1830 }
1831
1832 static int
1833 segvn_unmap(struct seg *seg, caddr_t addr, size_t len)
1834 {
1835 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1836 struct segvn_data *nsvd;
1837 struct seg *nseg;
1838 struct anon_map *amp;
1839 pgcnt_t opages; /* old segment size in pages */
1840 pgcnt_t npages; /* new segment size in pages */
1841 pgcnt_t dpages; /* pages being deleted (unmapped) */
1842 hat_callback_t callback; /* used for free_vp_pages() */
1843 hat_callback_t *cbp = NULL;
1844 caddr_t nbase;
1845 size_t nsize;
1846 size_t oswresv;
1847 int reclaim = 1;
1848
1849 /*
1850 * We don't need any segment level locks for "segvn" data
1851 * since the address space is "write" locked.
1852 */
1853 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
1854
1855 /*
1856 * Fail the unmap if pages are SOFTLOCKed through this mapping.
1857 * softlockcnt is protected from change by the as write lock.
1858 */
1859 retry:
1860 if (svd->softlockcnt > 0) {
1861 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1862
1863 /*
1864 * If this is shared segment non 0 softlockcnt
1865 * means locked pages are still in use.
1866 */
1867 if (svd->type == MAP_SHARED) {
1868 return (EAGAIN);
1869 }
1870
1871 /*
1872 * since we do have the writers lock nobody can fill
1873 * the cache during the purge. The flush either succeeds
1874 * or we still have pending I/Os.
1875 */
1876 if (reclaim == 1) {
1877 segvn_purge(seg);
1878 reclaim = 0;
1879 goto retry;
1880 }
1881 return (EAGAIN);
1882 }
1883
1884 /*
1885 * Check for bad sizes
1886 */
1887 if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
1888 (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
1889 panic("segvn_unmap");
1890 /*NOTREACHED*/
1891 }
1892
1893 if (seg->s_szc != 0) {
1894 size_t pgsz = page_get_pagesize(seg->s_szc);
1895 int err;
1896 if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
1897 ASSERT(seg->s_base != addr || seg->s_size != len);
1898 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1899 ASSERT(svd->amp == NULL);
1900 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1901 hat_leave_region(seg->s_as->a_hat,
1902 svd->rcookie, HAT_REGION_TEXT);
1903 svd->rcookie = HAT_INVALID_REGION_COOKIE;
1904 /*
1905 * could pass a flag to segvn_demote_range()
1906 * below to tell it not to do any unloads but
1907 * this case is rare enough to not bother for
1908 * now.
1909 */
1910 } else if (svd->tr_state == SEGVN_TR_INIT) {
1911 svd->tr_state = SEGVN_TR_OFF;
1912 } else if (svd->tr_state == SEGVN_TR_ON) {
1913 ASSERT(svd->amp != NULL);
1914 segvn_textunrepl(seg, 1);
1915 ASSERT(svd->amp == NULL);
1916 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1917 }
1918 VM_STAT_ADD(segvnvmstats.demoterange[0]);
1919 err = segvn_demote_range(seg, addr, len, SDR_END, 0);
1920 if (err == 0) {
1921 return (IE_RETRY);
1922 }
1923 return (err);
1924 }
1925 }
1926
1927 /* Inform the vnode of the unmapping. */
1928 if (svd->vp) {
1929 int error;
1930
1931 error = fop_delmap(svd->vp,
1932 (offset_t)svd->offset + (uintptr_t)(addr - seg->s_base),
1933 seg->s_as, addr, len, svd->prot, svd->maxprot,
1934 svd->type, svd->cred, NULL);
1935
1936 if (error == EAGAIN)
1937 return (error);
1938 }
1939
1940 /*
1941 * Remove any page locks set through this mapping.
1942 * If text replication is not off no page locks could have been
1943 * established via this mapping.
1944 */
1945 if (svd->tr_state == SEGVN_TR_OFF) {
1946 (void) segvn_lockop(seg, addr, len, 0, MC_UNLOCK, NULL, 0);
1947 }
1948
1949 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1950 ASSERT(svd->amp == NULL);
1951 ASSERT(svd->tr_state == SEGVN_TR_OFF);
1952 ASSERT(svd->type == MAP_PRIVATE);
1953 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
1954 HAT_REGION_TEXT);
1955 svd->rcookie = HAT_INVALID_REGION_COOKIE;
1956 } else if (svd->tr_state == SEGVN_TR_ON) {
1957 ASSERT(svd->amp != NULL);
1958 ASSERT(svd->pageprot == 0 && !(svd->prot & PROT_WRITE));
1959 segvn_textunrepl(seg, 1);
1960 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
1961 } else {
1962 if (svd->tr_state != SEGVN_TR_OFF) {
1963 ASSERT(svd->tr_state == SEGVN_TR_INIT);
1964 svd->tr_state = SEGVN_TR_OFF;
1965 }
1966 /*
1967 * Unload any hardware translations in the range to be taken
1968 * out. Use a callback to invoke free_vp_pages() effectively.
1969 */
1970 if (svd->vp != NULL && free_pages != 0) {
1971 callback.hcb_data = seg;
1972 callback.hcb_function = segvn_hat_unload_callback;
1973 cbp = &callback;
1974 }
1975 hat_unload_callback(seg->s_as->a_hat, addr, len,
1976 HAT_UNLOAD_UNMAP, cbp);
1977
1978 if (svd->type == MAP_SHARED && svd->vp != NULL &&
1979 (svd->vp->v_flag & VVMEXEC) &&
1980 ((svd->prot & PROT_WRITE) || svd->pageprot)) {
1981 segvn_inval_trcache(svd->vp);
1982 }
1983 }
1984
1985 /*
1986 * Check for entire segment
1987 */
1988 if (addr == seg->s_base && len == seg->s_size) {
1989 seg_free(seg);
1990 return (0);
1991 }
1992
1993 opages = seg_pages(seg);
1994 dpages = btop(len);
1995 npages = opages - dpages;
1996 amp = svd->amp;
1997 ASSERT(amp == NULL || amp->a_szc >= seg->s_szc);
1998
1999 /*
2000 * Check for beginning of segment
2001 */
2002 if (addr == seg->s_base) {
2003 if (svd->vpage != NULL) {
2004 size_t nbytes;
2005 struct vpage *ovpage;
2006
2007 ovpage = svd->vpage; /* keep pointer to vpage */
2008
2009 nbytes = vpgtob(npages);
2010 svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2011 bcopy(&ovpage[dpages], svd->vpage, nbytes);
2012
2013 /* free up old vpage */
2014 kmem_free(ovpage, vpgtob(opages));
2015 }
2016 if (amp != NULL) {
2017 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2018 if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2019 /*
2020 * Shared anon map is no longer in use. Before
2021 * freeing its pages purge all entries from
2022 * pcache that belong to this amp.
2023 */
2024 if (svd->type == MAP_SHARED) {
2025 ASSERT(amp->refcnt == 1);
2026 ASSERT(svd->softlockcnt == 0);
2027 anonmap_purge(amp);
2028 }
2029 /*
2030 * Free up now unused parts of anon_map array.
2031 */
2032 if (amp->a_szc == seg->s_szc) {
2033 if (seg->s_szc != 0) {
2034 anon_free_pages(amp->ahp,
2035 svd->anon_index, len,
2036 seg->s_szc);
2037 } else {
2038 anon_free(amp->ahp,
2039 svd->anon_index,
2040 len);
2041 }
2042 } else {
2043 ASSERT(svd->type == MAP_SHARED);
2044 ASSERT(amp->a_szc > seg->s_szc);
2045 anon_shmap_free_pages(amp,
2046 svd->anon_index, len);
2047 }
2048
2049 /*
2050 * Unreserve swap space for the
2051 * unmapped chunk of this segment in
2052 * case it's MAP_SHARED
2053 */
2054 if (svd->type == MAP_SHARED) {
2055 anon_unresv_zone(len,
2056 seg->s_as->a_proc->p_zone);
2057 amp->swresv -= len;
2058 }
2059 }
2060 ANON_LOCK_EXIT(&amp->a_rwlock);
2061 svd->anon_index += dpages;
2062 }
2063 if (svd->vp != NULL)
2064 svd->offset += len;
2065
2066 seg->s_base += len;
2067 seg->s_size -= len;
2068
2069 if (svd->swresv) {
2070 if (svd->flags & MAP_NORESERVE) {
2071 ASSERT(amp);
2072 oswresv = svd->swresv;
2073
2074 svd->swresv = ptob(anon_pages(amp->ahp,
2075 svd->anon_index, npages));
2076 anon_unresv_zone(oswresv - svd->swresv,
2077 seg->s_as->a_proc->p_zone);
2078 if (SEG_IS_PARTIAL_RESV(seg))
2079 seg->s_as->a_resvsize -= oswresv -
2080 svd->swresv;
2081 } else {
2082 size_t unlen;
2083
2084 if (svd->pageswap) {
2085 oswresv = svd->swresv;
2086 svd->swresv =
2087 segvn_count_swap_by_vpages(seg);
2088 ASSERT(oswresv >= svd->swresv);
2089 unlen = oswresv - svd->swresv;
2090 } else {
2091 svd->swresv -= len;
2092 ASSERT(svd->swresv == seg->s_size);
2093 unlen = len;
2094 }
2095 anon_unresv_zone(unlen,
2096 seg->s_as->a_proc->p_zone);
2097 }
2098 }
2099
2100 return (0);
2101 }
2102
2103 /*
2104 * Check for end of segment
2105 */
2106 if (addr + len == seg->s_base + seg->s_size) {
2107 if (svd->vpage != NULL) {
2108 size_t nbytes;
2109 struct vpage *ovpage;
2110
2111 ovpage = svd->vpage; /* keep pointer to vpage */
2112
2113 nbytes = vpgtob(npages);
2114 svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2115 bcopy(ovpage, svd->vpage, nbytes);
2116
2117 /* free up old vpage */
2118 kmem_free(ovpage, vpgtob(opages));
2119
2120 }
2121 if (amp != NULL) {
2122 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2123 if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2124 /*
2125 * Free up now unused parts of anon_map array.
2126 */
2127 ulong_t an_idx = svd->anon_index + npages;
2128
2129 /*
2130 * Shared anon map is no longer in use. Before
2131 * freeing its pages purge all entries from
2132 * pcache that belong to this amp.
2133 */
2134 if (svd->type == MAP_SHARED) {
2135 ASSERT(amp->refcnt == 1);
2136 ASSERT(svd->softlockcnt == 0);
2137 anonmap_purge(amp);
2138 }
2139
2140 if (amp->a_szc == seg->s_szc) {
2141 if (seg->s_szc != 0) {
2142 anon_free_pages(amp->ahp,
2143 an_idx, len,
2144 seg->s_szc);
2145 } else {
2146 anon_free(amp->ahp, an_idx,
2147 len);
2148 }
2149 } else {
2150 ASSERT(svd->type == MAP_SHARED);
2151 ASSERT(amp->a_szc > seg->s_szc);
2152 anon_shmap_free_pages(amp,
2153 an_idx, len);
2154 }
2155
2156 /*
2157 * Unreserve swap space for the
2158 * unmapped chunk of this segment in
2159 * case it's MAP_SHARED
2160 */
2161 if (svd->type == MAP_SHARED) {
2162 anon_unresv_zone(len,
2163 seg->s_as->a_proc->p_zone);
2164 amp->swresv -= len;
2165 }
2166 }
2167 ANON_LOCK_EXIT(&amp->a_rwlock);
2168 }
2169
2170 seg->s_size -= len;
2171
2172 if (svd->swresv) {
2173 if (svd->flags & MAP_NORESERVE) {
2174 ASSERT(amp);
2175 oswresv = svd->swresv;
2176 svd->swresv = ptob(anon_pages(amp->ahp,
2177 svd->anon_index, npages));
2178 anon_unresv_zone(oswresv - svd->swresv,
2179 seg->s_as->a_proc->p_zone);
2180 if (SEG_IS_PARTIAL_RESV(seg))
2181 seg->s_as->a_resvsize -= oswresv -
2182 svd->swresv;
2183 } else {
2184 size_t unlen;
2185
2186 if (svd->pageswap) {
2187 oswresv = svd->swresv;
2188 svd->swresv =
2189 segvn_count_swap_by_vpages(seg);
2190 ASSERT(oswresv >= svd->swresv);
2191 unlen = oswresv - svd->swresv;
2192 } else {
2193 svd->swresv -= len;
2194 ASSERT(svd->swresv == seg->s_size);
2195 unlen = len;
2196 }
2197 anon_unresv_zone(unlen,
2198 seg->s_as->a_proc->p_zone);
2199 }
2200 }
2201
2202 return (0);
2203 }
2204
2205 /*
2206 * The section to go is in the middle of the segment,
2207 * have to make it into two segments. nseg is made for
2208 * the high end while seg is cut down at the low end.
2209 */
2210 nbase = addr + len; /* new seg base */
2211 nsize = (seg->s_base + seg->s_size) - nbase; /* new seg size */
2212 seg->s_size = addr - seg->s_base; /* shrink old seg */
2213 nseg = seg_alloc(seg->s_as, nbase, nsize);
2214 if (nseg == NULL) {
2215 panic("segvn_unmap seg_alloc");
2216 /*NOTREACHED*/
2217 }
2218 nseg->s_ops = seg->s_ops;
2219 nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
2220 nseg->s_data = (void *)nsvd;
2221 nseg->s_szc = seg->s_szc;
2222 *nsvd = *svd;
2223 nsvd->seg = nseg;
2224 nsvd->offset = svd->offset + (uintptr_t)(nseg->s_base - seg->s_base);
2225 nsvd->swresv = 0;
2226 nsvd->softlockcnt = 0;
2227 nsvd->softlockcnt_sbase = 0;
2228 nsvd->softlockcnt_send = 0;
2229 nsvd->svn_inz = svd->svn_inz;
2230 ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
2231
2232 if (svd->vp != NULL) {
2233 VN_HOLD(nsvd->vp);
2234 if (nsvd->type == MAP_SHARED)
2235 lgrp_shm_policy_init(NULL, nsvd->vp);
2236 }
2237 crhold(svd->cred);
2238
2239 if (svd->vpage == NULL) {
2240 nsvd->vpage = NULL;
2241 } else {
2242 /* need to split vpage into two arrays */
2243 size_t nbytes;
2244 struct vpage *ovpage;
2245
2246 ovpage = svd->vpage; /* keep pointer to vpage */
2247
2248 npages = seg_pages(seg); /* seg has shrunk */
2249 nbytes = vpgtob(npages);
2250 svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2251
2252 bcopy(ovpage, svd->vpage, nbytes);
2253
2254 npages = seg_pages(nseg);
2255 nbytes = vpgtob(npages);
2256 nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2257
2258 bcopy(&ovpage[opages - npages], nsvd->vpage, nbytes);
2259
2260 /* free up old vpage */
2261 kmem_free(ovpage, vpgtob(opages));
2262 }
2263
2264 if (amp == NULL) {
2265 nsvd->amp = NULL;
2266 nsvd->anon_index = 0;
2267 } else {
2268 /*
2269 * Need to create a new anon map for the new segment.
2270 * We'll also allocate a new smaller array for the old
2271 * smaller segment to save space.
2272 */
2273 opages = btop((uintptr_t)(addr - seg->s_base));
2274 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2275 if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2276 /*
2277 * Free up now unused parts of anon_map array.
2278 */
2279 ulong_t an_idx = svd->anon_index + opages;
2280
2281 /*
2282 * Shared anon map is no longer in use. Before
2283 * freeing its pages purge all entries from
2284 * pcache that belong to this amp.
2285 */
2286 if (svd->type == MAP_SHARED) {
2287 ASSERT(amp->refcnt == 1);
2288 ASSERT(svd->softlockcnt == 0);
2289 anonmap_purge(amp);
2290 }
2291
2292 if (amp->a_szc == seg->s_szc) {
2293 if (seg->s_szc != 0) {
2294 anon_free_pages(amp->ahp, an_idx, len,
2295 seg->s_szc);
2296 } else {
2297 anon_free(amp->ahp, an_idx,
2298 len);
2299 }
2300 } else {
2301 ASSERT(svd->type == MAP_SHARED);
2302 ASSERT(amp->a_szc > seg->s_szc);
2303 anon_shmap_free_pages(amp, an_idx, len);
2304 }
2305
2306 /*
2307 * Unreserve swap space for the
2308 * unmapped chunk of this segment in
2309 * case it's MAP_SHARED
2310 */
2311 if (svd->type == MAP_SHARED) {
2312 anon_unresv_zone(len,
2313 seg->s_as->a_proc->p_zone);
2314 amp->swresv -= len;
2315 }
2316 }
2317 nsvd->anon_index = svd->anon_index +
2318 btop((uintptr_t)(nseg->s_base - seg->s_base));
2319 if (svd->type == MAP_SHARED) {
2320 amp->refcnt++;
2321 nsvd->amp = amp;
2322 } else {
2323 struct anon_map *namp;
2324 struct anon_hdr *nahp;
2325
2326 ASSERT(svd->type == MAP_PRIVATE);
2327 nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
2328 namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
2329 namp->a_szc = seg->s_szc;
2330 (void) anon_copy_ptr(amp->ahp, svd->anon_index, nahp,
2331 0, btop(seg->s_size), ANON_SLEEP);
2332 (void) anon_copy_ptr(amp->ahp, nsvd->anon_index,
2333 namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
2334 anon_release(amp->ahp, btop(amp->size));
2335 svd->anon_index = 0;
2336 nsvd->anon_index = 0;
2337 amp->ahp = nahp;
2338 amp->size = seg->s_size;
2339 nsvd->amp = namp;
2340 }
2341 ANON_LOCK_EXIT(&amp->a_rwlock);
2342 }
2343 if (svd->swresv) {
2344 if (svd->flags & MAP_NORESERVE) {
2345 ASSERT(amp);
2346 oswresv = svd->swresv;
2347 svd->swresv = ptob(anon_pages(amp->ahp,
2348 svd->anon_index, btop(seg->s_size)));
2349 nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
2350 nsvd->anon_index, btop(nseg->s_size)));
2351 ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2352 anon_unresv_zone(oswresv - (svd->swresv + nsvd->swresv),
2353 seg->s_as->a_proc->p_zone);
2354 if (SEG_IS_PARTIAL_RESV(seg))
2355 seg->s_as->a_resvsize -= oswresv -
2356 (svd->swresv + nsvd->swresv);
2357 } else {
2358 size_t unlen;
2359
2360 if (svd->pageswap) {
2361 oswresv = svd->swresv;
2362 svd->swresv = segvn_count_swap_by_vpages(seg);
2363 nsvd->swresv = segvn_count_swap_by_vpages(nseg);
2364 ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2365 unlen = oswresv - (svd->swresv + nsvd->swresv);
2366 } else {
2367 if (seg->s_size + nseg->s_size + len !=
2368 svd->swresv) {
2369 panic("segvn_unmap: cannot split "
2370 "swap reservation");
2371 /*NOTREACHED*/
2372 }
2373 svd->swresv = seg->s_size;
2374 nsvd->swresv = nseg->s_size;
2375 unlen = len;
2376 }
2377 anon_unresv_zone(unlen,
2378 seg->s_as->a_proc->p_zone);
2379 }
2380 }
2381
2382 return (0); /* I'm glad that's all over with! */
2383 }
2384
2385 static void
2386 segvn_free(struct seg *seg)
2387 {
2388 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2389 pgcnt_t npages = seg_pages(seg);
2390 struct anon_map *amp;
2391 size_t len;
2392
2393 /*
2394 * We don't need any segment level locks for "segvn" data
2395 * since the address space is "write" locked.
2396 */
2397 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
2398 ASSERT(svd->tr_state == SEGVN_TR_OFF);
2399
2400 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2401
2402 /*
2403 * Be sure to unlock pages. XXX Why do things get free'ed instead
2404 * of unmapped? XXX
2405 */
2406 (void) segvn_lockop(seg, seg->s_base, seg->s_size,
2407 0, MC_UNLOCK, NULL, 0);
2408
2409 /*
2410 * Deallocate the vpage and anon pointers if necessary and possible.
2411 */
2412 if (svd->vpage != NULL) {
2413 kmem_free(svd->vpage, vpgtob(npages));
2414 svd->vpage = NULL;
2415 }
2416 if ((amp = svd->amp) != NULL) {
2417 /*
2418 * If there are no more references to this anon_map
2419 * structure, then deallocate the structure after freeing
2420 * up all the anon slot pointers that we can.
2421 */
2422 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2423 ASSERT(amp->a_szc >= seg->s_szc);
2424 if (--amp->refcnt == 0) {
2425 if (svd->type == MAP_PRIVATE) {
2426 /*
2427 * Private - we only need to anon_free
2428 * the part that this segment refers to.
2429 */
2430 if (seg->s_szc != 0) {
2431 anon_free_pages(amp->ahp,
2432 svd->anon_index, seg->s_size,
2433 seg->s_szc);
2434 } else {
2435 anon_free(amp->ahp, svd->anon_index,
2436 seg->s_size);
2437 }
2438 } else {
2439
2440 /*
2441 * Shared anon map is no longer in use. Before
2442 * freeing its pages purge all entries from
2443 * pcache that belong to this amp.
2444 */
2445 ASSERT(svd->softlockcnt == 0);
2446 anonmap_purge(amp);
2447
2448 /*
2449 * Shared - anon_free the entire
2450 * anon_map's worth of stuff and
2451 * release any swap reservation.
2452 */
2453 if (amp->a_szc != 0) {
2454 anon_shmap_free_pages(amp, 0,
2455 amp->size);
2456 } else {
2457 anon_free(amp->ahp, 0, amp->size);
2458 }
2459 if ((len = amp->swresv) != 0) {
2460 anon_unresv_zone(len,
2461 seg->s_as->a_proc->p_zone);
2462 }
2463 }
2464 svd->amp = NULL;
2465 ANON_LOCK_EXIT(&amp->a_rwlock);
2466 anonmap_free(amp);
2467 } else if (svd->type == MAP_PRIVATE) {
2468 /*
2469 * We had a private mapping which still has
2470 * a held anon_map so just free up all the
2471 * anon slot pointers that we were using.
2472 */
2473 if (seg->s_szc != 0) {
2474 anon_free_pages(amp->ahp, svd->anon_index,
2475 seg->s_size, seg->s_szc);
2476 } else {
2477 anon_free(amp->ahp, svd->anon_index,
2478 seg->s_size);
2479 }
2480 ANON_LOCK_EXIT(&amp->a_rwlock);
2481 } else {
2482 ANON_LOCK_EXIT(&amp->a_rwlock);
2483 }
2484 }
2485
2486 /*
2487 * Release swap reservation.
2488 */
2489 if ((len = svd->swresv) != 0) {
2490 anon_unresv_zone(svd->swresv,
2491 seg->s_as->a_proc->p_zone);
2492 if (SEG_IS_PARTIAL_RESV(seg))
2493 seg->s_as->a_resvsize -= svd->swresv;
2494 svd->swresv = 0;
2495 }
2496 /*
2497 * Release claim on vnode, credentials, and finally free the
2498 * private data.
2499 */
2500 if (svd->vp != NULL) {
2501 if (svd->type == MAP_SHARED)
2502 lgrp_shm_policy_fini(NULL, svd->vp);
2503 VN_RELE(svd->vp);
2504 svd->vp = NULL;
2505 }
2506 crfree(svd->cred);
2507 svd->pageprot = 0;
2508 svd->pageadvice = 0;
2509 svd->pageswap = 0;
2510 svd->cred = NULL;
2511
2512 /*
2513 * Take segfree_syncmtx lock to let segvn_reclaim() finish if it's
2514 * still working with this segment without holding as lock (in case
2515 * it's called by pcache async thread).
2516 */
2517 ASSERT(svd->softlockcnt == 0);
2518 mutex_enter(&svd->segfree_syncmtx);
2519 mutex_exit(&svd->segfree_syncmtx);
2520
2521 seg->s_data = NULL;
2522 kmem_cache_free(segvn_cache, svd);
2523 }
2524
2525 /*
2526 * Do a F_SOFTUNLOCK call over the range requested. The range must have
2527 * already been F_SOFTLOCK'ed.
2528 * Caller must always match addr and len of a softunlock with a previous
2529 * softlock with exactly the same addr and len.
2530 */
2531 static void
2532 segvn_softunlock(struct seg *seg, caddr_t addr, size_t len, enum seg_rw rw)
2533 {
2534 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2535 page_t *pp;
2536 caddr_t adr;
2537 struct vnode *vp;
2538 uoff_t offset;
2539 ulong_t anon_index;
2540 struct anon_map *amp;
2541 struct anon *ap = NULL;
2542
2543 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
2544 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
2545
2546 if ((amp = svd->amp) != NULL)
2547 anon_index = svd->anon_index + seg_page(seg, addr);
2548
2549 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
2550 ASSERT(svd->tr_state == SEGVN_TR_OFF);
2551 hat_unlock_region(seg->s_as->a_hat, addr, len, svd->rcookie);
2552 } else {
2553 hat_unlock(seg->s_as->a_hat, addr, len);
2554 }
2555 for (adr = addr; adr < addr + len; adr += PAGESIZE) {
2556 if (amp != NULL) {
2557 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2558 if ((ap = anon_get_ptr(amp->ahp, anon_index++))
2559 != NULL) {
2560 swap_xlate(ap, &vp, &offset);
2561 } else {
2562 vp = svd->vp;
2563 offset = svd->offset +
2564 (uintptr_t)(adr - seg->s_base);
2565 }
2566 ANON_LOCK_EXIT(&amp->a_rwlock);
2567 } else {
2568 vp = svd->vp;
2569 offset = svd->offset +
2570 (uintptr_t)(adr - seg->s_base);
2571 }
2572
2573 /*
2574 * Use page_find() instead of page_lookup() to
2575 * find the page since we know that it is locked.
2576 */
2577 pp = page_find(vp, offset);
2578 if (pp == NULL) {
2579 panic(
2580 "segvn_softunlock: addr %p, ap %p, vp %p, off %llx",
2581 (void *)adr, (void *)ap, (void *)vp, offset);
2582 /*NOTREACHED*/
2583 }
2584
2585 if (rw == S_WRITE) {
2586 hat_setrefmod(pp);
2587 if (seg->s_as->a_vbits)
2588 hat_setstat(seg->s_as, adr, PAGESIZE,
2589 P_REF | P_MOD);
2590 } else if (rw != S_OTHER) {
2591 hat_setref(pp);
2592 if (seg->s_as->a_vbits)
2593 hat_setstat(seg->s_as, adr, PAGESIZE, P_REF);
2594 }
2595 page_unlock(pp);
2596 }
2597 ASSERT(svd->softlockcnt >= btop(len));
2598 if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -btop(len))) {
2599 /*
2600 * All SOFTLOCKS are gone. Wakeup any waiting
2601 * unmappers so they can try again to unmap.
2602 * Check for waiters first without the mutex
2603 * held so we don't always grab the mutex on
2604 * softunlocks.
2605 */
2606 if (AS_ISUNMAPWAIT(seg->s_as)) {
2607 mutex_enter(&seg->s_as->a_contents);
2608 if (AS_ISUNMAPWAIT(seg->s_as)) {
2609 AS_CLRUNMAPWAIT(seg->s_as);
2610 cv_broadcast(&seg->s_as->a_cv);
2611 }
2612 mutex_exit(&seg->s_as->a_contents);
2613 }
2614 }
2615 }
2616
2617 #define PAGE_HANDLED ((page_t *)-1)
2618
2619 /*
2620 * Release all the pages in the NULL terminated ppp list
2621 * which haven't already been converted to PAGE_HANDLED.
2622 */
2623 static void
2624 segvn_pagelist_rele(page_t **ppp)
2625 {
2626 for (; *ppp != NULL; ppp++) {
2627 if (*ppp != PAGE_HANDLED)
2628 page_unlock(*ppp);
2629 }
2630 }
2631
2632 static int stealcow = 1;
2633
2634 /*
2635 * Workaround for viking chip bug. See bug id 1220902.
2636 * To fix this down in pagefault() would require importing so
2637 * much as and segvn code as to be unmaintainable.
2638 */
2639 int enable_mbit_wa = 0;
2640
2641 /*
2642 * Handles all the dirty work of getting the right
2643 * anonymous pages and loading up the translations.
2644 * This routine is called only from segvn_fault()
2645 * when looping over the range of addresses requested.
2646 *
2647 * The basic algorithm here is:
2648 * If this is an anon_zero case
2649 * Call anon_zero to allocate page
2650 * Load up translation
2651 * Return
2652 * endif
2653 * If this is an anon page
2654 * Use anon_getpage to get the page
2655 * else
2656 * Find page in pl[] list passed in
2657 * endif
2658 * If not a cow
2659 * Load up the translation to the page
2660 * return
2661 * endif
2662 * Call anon_private to handle cow
2663 * Load up (writable) translation to new page
2664 */
2665 static faultcode_t
2666 segvn_faultpage(
2667 struct hat *hat, /* the hat to use for mapping */
2668 struct seg *seg, /* seg_vn of interest */
2669 caddr_t addr, /* address in as */
2670 uoff_t off, /* offset in vp */
2671 struct vpage *vpage, /* pointer to vpage for vp, off */
2672 page_t *pl[], /* object source page pointer */
2673 uint_t vpprot, /* access allowed to object pages */
2674 enum fault_type type, /* type of fault */
2675 enum seg_rw rw, /* type of access at fault */
2676 int brkcow) /* we may need to break cow */
2677 {
2678 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2679 page_t *pp, **ppp;
2680 uint_t pageflags = 0;
2681 page_t *anon_pl[1 + 1];
2682 page_t *opp = NULL; /* original page */
2683 uint_t prot;
2684 int err;
2685 int cow;
2686 int claim;
2687 int steal = 0;
2688 ulong_t anon_index;
2689 struct anon *ap, *oldap;
2690 struct anon_map *amp;
2691 int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
2692 int anon_lock = 0;
2693 anon_sync_obj_t cookie;
2694
2695 if (svd->flags & MAP_TEXT) {
2696 hat_flag |= HAT_LOAD_TEXT;
2697 }
2698
2699 ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
2700 ASSERT(seg->s_szc == 0);
2701 ASSERT(svd->tr_state != SEGVN_TR_INIT);
2702
2703 /*
2704 * Initialize protection value for this page.
2705 * If we have per page protection values check it now.
2706 */
2707 if (svd->pageprot) {
2708 uint_t protchk;
2709
2710 switch (rw) {
2711 case S_READ:
2712 protchk = PROT_READ;
2713 break;
2714 case S_WRITE:
2715 protchk = PROT_WRITE;
2716 break;
2717 case S_EXEC:
2718 protchk = PROT_EXEC;
2719 break;
2720 case S_OTHER:
2721 default:
2722 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
2723 break;
2724 }
2725
2726 prot = VPP_PROT(vpage);
2727 if ((prot & protchk) == 0)
2728 return (FC_PROT); /* illegal access type */
2729 } else {
2730 prot = svd->prot;
2731 }
2732
2733 if (type == F_SOFTLOCK) {
2734 atomic_inc_ulong((ulong_t *)&svd->softlockcnt);
2735 }
2736
2737 /*
2738 * Always acquire the anon array lock to prevent 2 threads from
2739 * allocating separate anon slots for the same "addr".
2740 */
2741
2742 if ((amp = svd->amp) != NULL) {
2743 ASSERT(RW_READ_HELD(&amp->a_rwlock));
2744 anon_index = svd->anon_index + seg_page(seg, addr);
2745 anon_array_enter(amp, anon_index, &cookie);
2746 anon_lock = 1;
2747 }
2748
2749 if (svd->vp == NULL && amp != NULL) {
2750 if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL) {
2751 /*
2752 * Allocate a (normally) writable anonymous page of
2753 * zeroes. If no advance reservations, reserve now.
2754 */
2755 if (svd->flags & MAP_NORESERVE) {
2756 if (anon_resv_zone(ptob(1),
2757 seg->s_as->a_proc->p_zone)) {
2758 atomic_add_long(&svd->swresv, ptob(1));
2759 atomic_add_long(&seg->s_as->a_resvsize,
2760 ptob(1));
2761 } else {
2762 err = ENOMEM;
2763 goto out;
2764 }
2765 }
2766 if ((pp = anon_zero(seg, addr, &ap,
2767 svd->cred)) == NULL) {
2768 err = ENOMEM;
2769 goto out; /* out of swap space */
2770 }
2771 /*
2772 * Re-acquire the anon_map lock and
2773 * initialize the anon array entry.
2774 */
2775 (void) anon_set_ptr(amp->ahp, anon_index, ap,
2776 ANON_SLEEP);
2777
2778 ASSERT(pp->p_szc == 0);
2779
2780 /*
2781 * Handle pages that have been marked for migration
2782 */
2783 if (lgrp_optimizations())
2784 page_migrate(seg, addr, &pp, 1);
2785
2786 if (enable_mbit_wa) {
2787 if (rw == S_WRITE)
2788 hat_setmod(pp);
2789 else if (!hat_ismod(pp))
2790 prot &= ~PROT_WRITE;
2791 }
2792 /*
2793 * If AS_PAGLCK is set in a_flags (via memcntl(2)
2794 * with MC_LOCKAS, MCL_FUTURE) and this is a
2795 * MAP_NORESERVE segment, we may need to
2796 * permanently lock the page as it is being faulted
2797 * for the first time. The following text applies
2798 * only to MAP_NORESERVE segments:
2799 *
2800 * As per memcntl(2), if this segment was created
2801 * after MCL_FUTURE was applied (a "future"
2802 * segment), its pages must be locked. If this
2803 * segment existed at MCL_FUTURE application (a
2804 * "past" segment), the interface is unclear.
2805 *
2806 * We decide to lock only if vpage is present:
2807 *
2808 * - "future" segments will have a vpage array (see
2809 * as_map), and so will be locked as required
2810 *
2811 * - "past" segments may not have a vpage array,
2812 * depending on whether events (such as
2813 * mprotect) have occurred. Locking if vpage
2814 * exists will preserve legacy behavior. Not
2815 * locking if vpage is absent, will not break
2816 * the interface or legacy behavior. Note that
2817 * allocating vpage here if it's absent requires
2818 * upgrading the segvn reader lock, the cost of
2819 * which does not seem worthwhile.
2820 *
2821 * Usually testing and setting VPP_ISPPLOCK and
2822 * VPP_SETPPLOCK requires holding the segvn lock as
2823 * writer, but in this case all readers are
2824 * serializing on the anon array lock.
2825 */
2826 if (AS_ISPGLCK(seg->s_as) && vpage != NULL &&
2827 (svd->flags & MAP_NORESERVE) &&
2828 !VPP_ISPPLOCK(vpage)) {
2829 proc_t *p = seg->s_as->a_proc;
2830 ASSERT(svd->type == MAP_PRIVATE);
2831 mutex_enter(&p->p_lock);
2832 if (rctl_incr_locked_mem(p, NULL, PAGESIZE,
2833 1) == 0) {
2834 claim = VPP_PROT(vpage) & PROT_WRITE;
2835 if (page_pp_lock(pp, claim, 0)) {
2836 VPP_SETPPLOCK(vpage);
2837 } else {
2838 rctl_decr_locked_mem(p, NULL,
2839 PAGESIZE, 1);
2840 }
2841 }
2842 mutex_exit(&p->p_lock);
2843 }
2844
2845 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2846 hat_memload(hat, addr, pp, prot, hat_flag);
2847
2848 if (!(hat_flag & HAT_LOAD_LOCK))
2849 page_unlock(pp);
2850
2851 anon_array_exit(&cookie);
2852 return (0);
2853 }
2854 }
2855
2856 /*
2857 * Obtain the page structure via anon_getpage() if it is
2858 * a private copy of an object (the result of a previous
2859 * copy-on-write).
2860 */
2861 if (amp != NULL) {
2862 if ((ap = anon_get_ptr(amp->ahp, anon_index)) != NULL) {
2863 err = anon_getpage(&ap, &vpprot, anon_pl, PAGESIZE,
2864 seg, addr, rw, svd->cred);
2865 if (err)
2866 goto out;
2867
2868 if (svd->type == MAP_SHARED) {
2869 /*
2870 * If this is a shared mapping to an
2871 * anon_map, then ignore the write
2872 * permissions returned by anon_getpage().
2873 * They apply to the private mappings
2874 * of this anon_map.
2875 */
2876 vpprot |= PROT_WRITE;
2877 }
2878 opp = anon_pl[0];
2879 }
2880 }
2881
2882 /*
2883 * Search the pl[] list passed in if it is from the
2884 * original object (i.e., not a private copy).
2885 */
2886 if (opp == NULL) {
2887 /*
2888 * Find original page. We must be bringing it in
2889 * from the list in pl[].
2890 */
2891 for (ppp = pl; (opp = *ppp) != NULL; ppp++) {
2892 if (opp == PAGE_HANDLED)
2893 continue;
2894 ASSERT(opp->p_vnode == svd->vp); /* XXX */
2895 if (opp->p_offset == off)
2896 break;
2897 }
2898 if (opp == NULL) {
2899 panic("segvn_faultpage not found");
2900 /*NOTREACHED*/
2901 }
2902 *ppp = PAGE_HANDLED;
2903
2904 }
2905
2906 ASSERT(PAGE_LOCKED(opp));
2907
2908 /*
2909 * The fault is treated as a copy-on-write fault if a
2910 * write occurs on a private segment and the object
2911 * page (i.e., mapping) is write protected. We assume
2912 * that fatal protection checks have already been made.
2913 */
2914
2915 if (brkcow) {
2916 ASSERT(svd->tr_state == SEGVN_TR_OFF);
2917 cow = !(vpprot & PROT_WRITE);
2918 } else if (svd->tr_state == SEGVN_TR_ON) {
2919 /*
2920 * If we are doing text replication COW on first touch.
2921 */
2922 ASSERT(amp != NULL);
2923 ASSERT(svd->vp != NULL);
2924 ASSERT(rw != S_WRITE);
2925 cow = (ap == NULL);
2926 } else {
2927 cow = 0;
2928 }
2929
2930 /*
2931 * If not a copy-on-write case load the translation
2932 * and return.
2933 */
2934 if (cow == 0) {
2935
2936 /*
2937 * Handle pages that have been marked for migration
2938 */
2939 if (lgrp_optimizations())
2940 page_migrate(seg, addr, &opp, 1);
2941
2942 if (IS_VMODSORT(opp->p_vnode) || enable_mbit_wa) {
2943 if (rw == S_WRITE)
2944 hat_setmod(opp);
2945 else if (rw != S_OTHER && !hat_ismod(opp))
2946 prot &= ~PROT_WRITE;
2947 }
2948
2949 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
2950 (!svd->pageprot && svd->prot == (prot & vpprot)));
2951 ASSERT(amp == NULL ||
2952 svd->rcookie == HAT_INVALID_REGION_COOKIE);
2953 hat_memload_region(hat, addr, opp, prot & vpprot, hat_flag,
2954 svd->rcookie);
2955
2956 if (!(hat_flag & HAT_LOAD_LOCK))
2957 page_unlock(opp);
2958
2959 if (anon_lock) {
2960 anon_array_exit(&cookie);
2961 }
2962 return (0);
2963 }
2964
2965 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2966
2967 hat_setref(opp);
2968
2969 ASSERT(amp != NULL && anon_lock);
2970
2971 /*
2972 * Steal the page only if it isn't a private page
2973 * since stealing a private page is not worth the effort.
2974 */
2975 if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL)
2976 steal = 1;
2977
2978 /*
2979 * Steal the original page if the following conditions are true:
2980 *
2981 * We are low on memory, the page is not private, page is not large,
2982 * not shared, not modified, not `locked' or if we have it `locked'
2983 * (i.e., p_cowcnt == 1 and p_lckcnt == 0, which also implies
2984 * that the page is not shared) and if it doesn't have any
2985 * translations. page_struct_lock isn't needed to look at p_cowcnt
2986 * and p_lckcnt because we first get exclusive lock on page.
2987 */
2988 (void) hat_pagesync(opp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD);
2989
2990 if (stealcow && freemem < minfree && steal && opp->p_szc == 0 &&
2991 page_tryupgrade(opp) && !hat_ismod(opp) &&
2992 ((opp->p_lckcnt == 0 && opp->p_cowcnt == 0) ||
2993 (opp->p_lckcnt == 0 && opp->p_cowcnt == 1 &&
2994 vpage != NULL && VPP_ISPPLOCK(vpage)))) {
2995 /*
2996 * Check if this page has other translations
2997 * after unloading our translation.
2998 */
2999 if (hat_page_is_mapped(opp)) {
3000 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3001 hat_unload(seg->s_as->a_hat, addr, PAGESIZE,
3002 HAT_UNLOAD);
3003 }
3004
3005 /*
3006 * hat_unload() might sync back someone else's recent
3007 * modification, so check again.
3008 */
3009 if (!hat_ismod(opp) && !hat_page_is_mapped(opp))
3010 pageflags |= STEAL_PAGE;
3011 }
3012
3013 /*
3014 * If we have a vpage pointer, see if it indicates that we have
3015 * ``locked'' the page we map -- if so, tell anon_private to
3016 * transfer the locking resource to the new page.
3017 *
3018 * See Statement at the beginning of segvn_lockop regarding
3019 * the way lockcnts/cowcnts are handled during COW.
3020 *
3021 */
3022 if (vpage != NULL && VPP_ISPPLOCK(vpage))
3023 pageflags |= LOCK_PAGE;
3024
3025 /*
3026 * Allocate a private page and perform the copy.
3027 * For MAP_NORESERVE reserve swap space now, unless this
3028 * is a cow fault on an existing anon page in which case
3029 * MAP_NORESERVE will have made advance reservations.
3030 */
3031 if ((svd->flags & MAP_NORESERVE) && (ap == NULL)) {
3032 if (anon_resv_zone(ptob(1), seg->s_as->a_proc->p_zone)) {
3033 atomic_add_long(&svd->swresv, ptob(1));
3034 atomic_add_long(&seg->s_as->a_resvsize, ptob(1));
3035 } else {
3036 page_unlock(opp);
3037 err = ENOMEM;
3038 goto out;
3039 }
3040 }
3041 oldap = ap;
3042 pp = anon_private(&ap, seg, addr, prot, opp, pageflags, svd->cred);
3043 if (pp == NULL) {
3044 err = ENOMEM; /* out of swap space */
3045 goto out;
3046 }
3047
3048 /*
3049 * If we copied away from an anonymous page, then
3050 * we are one step closer to freeing up an anon slot.
3051 *
3052 * NOTE: The original anon slot must be released while
3053 * holding the "anon_map" lock. This is necessary to prevent
3054 * other threads from obtaining a pointer to the anon slot
3055 * which may be freed if its "refcnt" is 1.
3056 */
3057 if (oldap != NULL)
3058 anon_decref(oldap);
3059
3060 (void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
3061
3062 /*
3063 * Handle pages that have been marked for migration
3064 */
3065 if (lgrp_optimizations())
3066 page_migrate(seg, addr, &pp, 1);
3067
3068 ASSERT(pp->p_szc == 0);
3069
3070 ASSERT(!IS_VMODSORT(pp->p_vnode));
3071 if (enable_mbit_wa) {
3072 if (rw == S_WRITE)
3073 hat_setmod(pp);
3074 else if (!hat_ismod(pp))
3075 prot &= ~PROT_WRITE;
3076 }
3077
3078 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3079 hat_memload(hat, addr, pp, prot, hat_flag);
3080
3081 if (!(hat_flag & HAT_LOAD_LOCK))
3082 page_unlock(pp);
3083
3084 ASSERT(anon_lock);
3085 anon_array_exit(&cookie);
3086 return (0);
3087 out:
3088 if (anon_lock)
3089 anon_array_exit(&cookie);
3090
3091 if (type == F_SOFTLOCK) {
3092 atomic_dec_ulong((ulong_t *)&svd->softlockcnt);
3093 }
3094 return (FC_MAKE_ERR(err));
3095 }
3096
3097 /*
3098 * relocate a bunch of smaller targ pages into one large repl page. all targ
3099 * pages must be complete pages smaller than replacement pages.
3100 * it's assumed that no page's szc can change since they are all PAGESIZE or
3101 * complete large pages locked SHARED.
3102 */
3103 static void
3104 segvn_relocate_pages(page_t **targ, page_t *replacement)
3105 {
3106 page_t *pp;
3107 pgcnt_t repl_npgs, curnpgs;
3108 pgcnt_t i;
3109 uint_t repl_szc = replacement->p_szc;
3110 page_t *first_repl = replacement;
3111 page_t *repl;
3112 spgcnt_t npgs;
3113
3114 VM_STAT_ADD(segvnvmstats.relocatepages[0]);
3115
3116 ASSERT(repl_szc != 0);
3117 npgs = repl_npgs = page_get_pagecnt(repl_szc);
3118
3119 i = 0;
3120 while (repl_npgs) {
3121 spgcnt_t nreloc;
3122 int err;
3123 ASSERT(replacement != NULL);
3124 pp = targ[i];
3125 ASSERT(pp->p_szc < repl_szc);
3126 ASSERT(PAGE_EXCL(pp));
3127 ASSERT(!PP_ISFREE(pp));
3128 curnpgs = page_get_pagecnt(pp->p_szc);
3129 if (curnpgs == 1) {
3130 VM_STAT_ADD(segvnvmstats.relocatepages[1]);
3131 repl = replacement;
3132 page_sub(&replacement, repl);
3133 ASSERT(PAGE_EXCL(repl));
3134 ASSERT(!PP_ISFREE(repl));
3135 ASSERT(repl->p_szc == repl_szc);
3136 } else {
3137 page_t *repl_savepp;
3138 int j;
3139 VM_STAT_ADD(segvnvmstats.relocatepages[2]);
3140 repl_savepp = replacement;
3141 for (j = 0; j < curnpgs; j++) {
3142 repl = replacement;
3143 page_sub(&replacement, repl);
3144 ASSERT(PAGE_EXCL(repl));
3145 ASSERT(!PP_ISFREE(repl));
3146 ASSERT(repl->p_szc == repl_szc);
3147 ASSERT(page_pptonum(targ[i + j]) ==
3148 page_pptonum(targ[i]) + j);
3149 }
3150 repl = repl_savepp;
3151 ASSERT(IS_P2ALIGNED(page_pptonum(repl), curnpgs));
3152 }
3153 err = page_relocate(&pp, &repl, 0, 1, &nreloc, NULL);
3154 if (err || nreloc != curnpgs) {
3155 panic("segvn_relocate_pages: "
3156 "page_relocate failed err=%d curnpgs=%ld "
3157 "nreloc=%ld", err, curnpgs, nreloc);
3158 }
3159 ASSERT(curnpgs <= repl_npgs);
3160 repl_npgs -= curnpgs;
3161 i += curnpgs;
3162 }
3163 ASSERT(replacement == NULL);
3164
3165 repl = first_repl;
3166 repl_npgs = npgs;
3167 for (i = 0; i < repl_npgs; i++) {
3168 ASSERT(PAGE_EXCL(repl));
3169 ASSERT(!PP_ISFREE(repl));
3170 targ[i] = repl;
3171 page_downgrade(targ[i]);
3172 repl++;
3173 }
3174 }
3175
3176 /*
3177 * Check if all pages in ppa array are complete smaller than szc pages and
3178 * their roots will still be aligned relative to their current size if the
3179 * entire ppa array is relocated into one szc page. If these conditions are
3180 * not met return 0.
3181 *
3182 * If all pages are properly aligned attempt to upgrade their locks
3183 * to exclusive mode. If it fails set *upgrdfail to 1 and return 0.
3184 * upgrdfail was set to 0 by caller.
3185 *
3186 * Return 1 if all pages are aligned and locked exclusively.
3187 *
3188 * If all pages in ppa array happen to be physically contiguous to make one
3189 * szc page and all exclusive locks are successfully obtained promote the page
3190 * size to szc and set *pszc to szc. Return 1 with pages locked shared.
3191 */
3192 static int
3193 segvn_full_szcpages(page_t **ppa, uint_t szc, int *upgrdfail, uint_t *pszc)
3194 {
3195 page_t *pp;
3196 pfn_t pfn;
3197 pgcnt_t totnpgs = page_get_pagecnt(szc);
3198 pfn_t first_pfn;
3199 int contig = 1;
3200 pgcnt_t i;
3201 pgcnt_t j;
3202 uint_t curszc;
3203 pgcnt_t curnpgs;
3204 int root = 0;
3205
3206 ASSERT(szc > 0);
3207
3208 VM_STAT_ADD(segvnvmstats.fullszcpages[0]);
3209
3210 for (i = 0; i < totnpgs; i++) {
3211 pp = ppa[i];
3212 ASSERT(PAGE_SHARED(pp));
3213 ASSERT(!PP_ISFREE(pp));
3214 pfn = page_pptonum(pp);
3215 if (i == 0) {
3216 if (!IS_P2ALIGNED(pfn, totnpgs)) {
3217 contig = 0;
3218 } else {
3219 first_pfn = pfn;
3220 }
3221 } else if (contig && pfn != first_pfn + i) {
3222 contig = 0;
3223 }
3224 if (pp->p_szc == 0) {
3225 if (root) {
3226 VM_STAT_ADD(segvnvmstats.fullszcpages[1]);
3227 return (0);
3228 }
3229 } else if (!root) {
3230 if ((curszc = pp->p_szc) >= szc) {
3231 VM_STAT_ADD(segvnvmstats.fullszcpages[2]);
3232 return (0);
3233 }
3234 if (curszc == 0) {
3235 /*
3236 * p_szc changed means we don't have all pages
3237 * locked. return failure.
3238 */
3239 VM_STAT_ADD(segvnvmstats.fullszcpages[3]);
3240 return (0);
3241 }
3242 curnpgs = page_get_pagecnt(curszc);
3243 if (!IS_P2ALIGNED(pfn, curnpgs) ||
3244 !IS_P2ALIGNED(i, curnpgs)) {
3245 VM_STAT_ADD(segvnvmstats.fullszcpages[4]);
3246 return (0);
3247 }
3248 root = 1;
3249 } else {
3250 ASSERT(i > 0);
3251 VM_STAT_ADD(segvnvmstats.fullszcpages[5]);
3252 if (pp->p_szc != curszc) {
3253 VM_STAT_ADD(segvnvmstats.fullszcpages[6]);
3254 return (0);
3255 }
3256 if (pfn - 1 != page_pptonum(ppa[i - 1])) {
3257 panic("segvn_full_szcpages: "
3258 "large page not physically contiguous");
3259 }
3260 if (P2PHASE(pfn, curnpgs) == curnpgs - 1) {
3261 root = 0;
3262 }
3263 }
3264 }
3265
3266 for (i = 0; i < totnpgs; i++) {
3267 ASSERT(ppa[i]->p_szc < szc);
3268 if (!page_tryupgrade(ppa[i])) {
3269 for (j = 0; j < i; j++) {
3270 page_downgrade(ppa[j]);
3271 }
3272 *pszc = ppa[i]->p_szc;
3273 *upgrdfail = 1;
3274 VM_STAT_ADD(segvnvmstats.fullszcpages[7]);
3275 return (0);
3276 }
3277 }
3278
3279 /*
3280 * When a page is put a free cachelist its szc is set to 0. if file
3281 * system reclaimed pages from cachelist targ pages will be physically
3282 * contiguous with 0 p_szc. in this case just upgrade szc of targ
3283 * pages without any relocations.
3284 * To avoid any hat issues with previous small mappings
3285 * hat_pageunload() the target pages first.
3286 */
3287 if (contig) {
3288 VM_STAT_ADD(segvnvmstats.fullszcpages[8]);
3289 for (i = 0; i < totnpgs; i++) {
3290 (void) hat_pageunload(ppa[i], HAT_FORCE_PGUNLOAD);
3291 }
3292 for (i = 0; i < totnpgs; i++) {
3293 ppa[i]->p_szc = szc;
3294 }
3295 for (i = 0; i < totnpgs; i++) {
3296 ASSERT(PAGE_EXCL(ppa[i]));
3297 page_downgrade(ppa[i]);
3298 }
3299 if (pszc != NULL) {
3300 *pszc = szc;
3301 }
3302 }
3303 VM_STAT_ADD(segvnvmstats.fullszcpages[9]);
3304 return (1);
3305 }
3306
3307 /*
3308 * Create physically contiguous pages for [vp, off] - [vp, off +
3309 * page_size(szc)) range and for private segment return them in ppa array.
3310 * Pages are created either via IO or relocations.
3311 *
3312 * Return 1 on success and 0 on failure.
3313 *
3314 * If physically contiguous pages already exist for this range return 1 without
3315 * filling ppa array. Caller initializes ppa[0] as NULL to detect that ppa
3316 * array wasn't filled. In this case caller fills ppa array via fop_getpage().
3317 */
3318
3319 static int
3320 segvn_fill_vp_pages(struct segvn_data *svd, vnode_t *vp, uoff_t off,
3321 uint_t szc, page_t **ppa, page_t **ppplist, uint_t *ret_pszc,
3322 int *downsize)
3323
3324 {
3325 page_t *pplist = *ppplist;
3326 size_t pgsz = page_get_pagesize(szc);
3327 pgcnt_t pages = btop(pgsz);
3328 ulong_t start_off = off;
3329 uoff_t eoff = off + pgsz;
3330 spgcnt_t nreloc;
3331 uoff_t io_off = off;
3332 size_t io_len;
3333 page_t *io_pplist = NULL;
3334 page_t *done_pplist = NULL;
3335 pgcnt_t pgidx = 0;
3336 page_t *pp;
3337 page_t *newpp;
3338 page_t *targpp;
3339 int io_err = 0;
3340 int i;
3341 pfn_t pfn;
3342 ulong_t ppages;
3343 page_t *targ_pplist = NULL;
3344 page_t *repl_pplist = NULL;
3345 page_t *tmp_pplist;
3346 int nios = 0;
3347 uint_t pszc;
3348 struct vattr va;
3349
3350 VM_STAT_ADD(segvnvmstats.fill_vp_pages[0]);
3351
3352 ASSERT(szc != 0);
3353 ASSERT(pplist->p_szc == szc);
3354
3355 /*
3356 * downsize will be set to 1 only if we fail to lock pages. this will
3357 * allow subsequent faults to try to relocate the page again. If we
3358 * fail due to misalignment don't downsize and let the caller map the
3359 * whole region with small mappings to avoid more faults into the area
3360 * where we can't get large pages anyway.
3361 */
3362 *downsize = 0;
3363
3364 while (off < eoff) {
3365 newpp = pplist;
3366 ASSERT(newpp != NULL);
3367 ASSERT(PAGE_EXCL(newpp));
3368 ASSERT(!PP_ISFREE(newpp));
3369 /*
3370 * we pass NULL for nrelocp to page_lookup_create()
3371 * so that it doesn't relocate. We relocate here
3372 * later only after we make sure we can lock all
3373 * pages in the range we handle and they are all
3374 * aligned.
3375 */
3376 pp = page_lookup_create(vp, off, SE_SHARED, newpp, NULL, 0);
3377 ASSERT(pp != NULL);
3378 ASSERT(!PP_ISFREE(pp));
3379 ASSERT(pp->p_vnode == vp);
3380 ASSERT(pp->p_offset == off);
3381 if (pp == newpp) {
3382 VM_STAT_ADD(segvnvmstats.fill_vp_pages[1]);
3383 page_sub(&pplist, pp);
3384 ASSERT(PAGE_EXCL(pp));
3385 ASSERT(page_iolock_assert(pp));
3386 page_list_concat(&io_pplist, &pp);
3387 off += PAGESIZE;
3388 continue;
3389 }
3390 VM_STAT_ADD(segvnvmstats.fill_vp_pages[2]);
3391 pfn = page_pptonum(pp);
3392 pszc = pp->p_szc;
3393 if (pszc >= szc && targ_pplist == NULL && io_pplist == NULL &&
3394 IS_P2ALIGNED(pfn, pages)) {
3395 ASSERT(repl_pplist == NULL);
3396 ASSERT(done_pplist == NULL);
3397 ASSERT(pplist == *ppplist);
3398 page_unlock(pp);
3399 page_free_replacement_page(pplist);
3400 page_create_putback(pages);
3401 *ppplist = NULL;
3402 VM_STAT_ADD(segvnvmstats.fill_vp_pages[3]);
3403 return (1);
3404 }
3405 if (pszc >= szc) {
3406 page_unlock(pp);
3407 segvn_faultvnmpss_align_err1++;
3408 goto out;
3409 }
3410 ppages = page_get_pagecnt(pszc);
3411 if (!IS_P2ALIGNED(pfn, ppages)) {
3412 ASSERT(pszc > 0);
3413 /*
3414 * sizing down to pszc won't help.
3415 */
3416 page_unlock(pp);
3417 segvn_faultvnmpss_align_err2++;
3418 goto out;
3419 }
3420 pfn = page_pptonum(newpp);
3421 if (!IS_P2ALIGNED(pfn, ppages)) {
3422 ASSERT(pszc > 0);
3423 /*
3424 * sizing down to pszc won't help.
3425 */
3426 page_unlock(pp);
3427 segvn_faultvnmpss_align_err3++;
3428 goto out;
3429 }
3430 if (!PAGE_EXCL(pp)) {
3431 VM_STAT_ADD(segvnvmstats.fill_vp_pages[4]);
3432 page_unlock(pp);
3433 *downsize = 1;
3434 *ret_pszc = pp->p_szc;
3435 goto out;
3436 }
3437 targpp = pp;
3438 if (io_pplist != NULL) {
3439 VM_STAT_ADD(segvnvmstats.fill_vp_pages[5]);
3440 io_len = off - io_off;
3441 /*
3442 * Some file systems like NFS don't check EOF
3443 * conditions in fop_pageio(). Check it here
3444 * now that pages are locked SE_EXCL. Any file
3445 * truncation will wait until the pages are
3446 * unlocked so no need to worry that file will
3447 * be truncated after we check its size here.
3448 * XXX fix NFS to remove this check.
3449 */
3450 va.va_mask = AT_SIZE;
3451 if (fop_getattr(vp, &va, ATTR_HINT, svd->cred, NULL)) {
3452 VM_STAT_ADD(segvnvmstats.fill_vp_pages[6]);
3453 page_unlock(targpp);
3454 goto out;
3455 }
3456 if (btopr(va.va_size) < btopr(io_off + io_len)) {
3457 VM_STAT_ADD(segvnvmstats.fill_vp_pages[7]);
3458 *downsize = 1;
3459 *ret_pszc = 0;
3460 page_unlock(targpp);
3461 goto out;
3462 }
3463 io_err = fop_pageio(vp, io_pplist, io_off, io_len,
3464 B_READ, svd->cred, NULL);
3465 if (io_err) {
3466 VM_STAT_ADD(segvnvmstats.fill_vp_pages[8]);
3467 page_unlock(targpp);
3468 if (io_err == EDEADLK) {
3469 segvn_vmpss_pageio_deadlk_err++;
3470 }
3471 goto out;
3472 }
3473 nios++;
3474 VM_STAT_ADD(segvnvmstats.fill_vp_pages[9]);
3475 while (io_pplist != NULL) {
3476 pp = io_pplist;
3477 page_sub(&io_pplist, pp);
3478 ASSERT(page_iolock_assert(pp));
3479 page_io_unlock(pp);
3480 pgidx = (pp->p_offset - start_off) >>
3481 PAGESHIFT;
3482 ASSERT(pgidx < pages);
3483 ppa[pgidx] = pp;
3484 page_list_concat(&done_pplist, &pp);
3485 }
3486 }
3487 pp = targpp;
3488 ASSERT(PAGE_EXCL(pp));
3489 ASSERT(pp->p_szc <= pszc);
3490 if (pszc != 0 && !group_page_trylock(pp, SE_EXCL)) {
3491 VM_STAT_ADD(segvnvmstats.fill_vp_pages[10]);
3492 page_unlock(pp);
3493 *downsize = 1;
3494 *ret_pszc = pp->p_szc;
3495 goto out;
3496 }
3497 VM_STAT_ADD(segvnvmstats.fill_vp_pages[11]);
3498 /*
3499 * page szc chould have changed before the entire group was
3500 * locked. reread page szc.
3501 */
3502 pszc = pp->p_szc;
3503 ppages = page_get_pagecnt(pszc);
3504
3505 /* link just the roots */
3506 page_list_concat(&targ_pplist, &pp);
3507 page_sub(&pplist, newpp);
3508 page_list_concat(&repl_pplist, &newpp);
3509 off += PAGESIZE;
3510 while (--ppages != 0) {
3511 newpp = pplist;
3512 page_sub(&pplist, newpp);
3513 off += PAGESIZE;
3514 }
3515 io_off = off;
3516 }
3517 if (io_pplist != NULL) {
3518 VM_STAT_ADD(segvnvmstats.fill_vp_pages[12]);
3519 io_len = eoff - io_off;
3520 va.va_mask = AT_SIZE;
3521 if (fop_getattr(vp, &va, ATTR_HINT, svd->cred, NULL) != 0) {
3522 VM_STAT_ADD(segvnvmstats.fill_vp_pages[13]);
3523 goto out;
3524 }
3525 if (btopr(va.va_size) < btopr(io_off + io_len)) {
3526 VM_STAT_ADD(segvnvmstats.fill_vp_pages[14]);
3527 *downsize = 1;
3528 *ret_pszc = 0;
3529 goto out;
3530 }
3531 io_err = fop_pageio(vp, io_pplist, io_off, io_len,
3532 B_READ, svd->cred, NULL);
3533 if (io_err) {
3534 VM_STAT_ADD(segvnvmstats.fill_vp_pages[15]);
3535 if (io_err == EDEADLK) {
3536 segvn_vmpss_pageio_deadlk_err++;
3537 }
3538 goto out;
3539 }
3540 nios++;
3541 while (io_pplist != NULL) {
3542 pp = io_pplist;
3543 page_sub(&io_pplist, pp);
3544 ASSERT(page_iolock_assert(pp));
3545 page_io_unlock(pp);
3546 pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3547 ASSERT(pgidx < pages);
3548 ppa[pgidx] = pp;
3549 }
3550 }
3551 /*
3552 * we're now bound to succeed or panic.
3553 * remove pages from done_pplist. it's not needed anymore.
3554 */
3555 while (done_pplist != NULL) {
3556 pp = done_pplist;
3557 page_sub(&done_pplist, pp);
3558 }
3559 VM_STAT_ADD(segvnvmstats.fill_vp_pages[16]);
3560 ASSERT(pplist == NULL);
3561 *ppplist = NULL;
3562 while (targ_pplist != NULL) {
3563 int ret;
3564 VM_STAT_ADD(segvnvmstats.fill_vp_pages[17]);
3565 ASSERT(repl_pplist);
3566 pp = targ_pplist;
3567 page_sub(&targ_pplist, pp);
3568 pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3569 newpp = repl_pplist;
3570 page_sub(&repl_pplist, newpp);
3571 #ifdef DEBUG
3572 pfn = page_pptonum(pp);
3573 pszc = pp->p_szc;
3574 ppages = page_get_pagecnt(pszc);
3575 ASSERT(IS_P2ALIGNED(pfn, ppages));
3576 pfn = page_pptonum(newpp);
3577 ASSERT(IS_P2ALIGNED(pfn, ppages));
3578 ASSERT(P2PHASE(pfn, pages) == pgidx);
3579 #endif
3580 nreloc = 0;
3581 ret = page_relocate(&pp, &newpp, 0, 1, &nreloc, NULL);
3582 if (ret != 0 || nreloc == 0) {
3583 panic("segvn_fill_vp_pages: "
3584 "page_relocate failed");
3585 }
3586 pp = newpp;
3587 while (nreloc-- != 0) {
3588 ASSERT(PAGE_EXCL(pp));
3589 ASSERT(pp->p_vnode == vp);
3590 ASSERT(pgidx ==
3591 ((pp->p_offset - start_off) >> PAGESHIFT));
3592 ppa[pgidx++] = pp;
3593 pp++;
3594 }
3595 }
3596
3597 if (svd->type == MAP_PRIVATE) {
3598 VM_STAT_ADD(segvnvmstats.fill_vp_pages[18]);
3599 for (i = 0; i < pages; i++) {
3600 ASSERT(ppa[i] != NULL);
3601 ASSERT(PAGE_EXCL(ppa[i]));
3602 ASSERT(ppa[i]->p_vnode == vp);
3603 ASSERT(ppa[i]->p_offset ==
3604 start_off + (i << PAGESHIFT));
3605 page_downgrade(ppa[i]);
3606 }
3607 ppa[pages] = NULL;
3608 } else {
3609 VM_STAT_ADD(segvnvmstats.fill_vp_pages[19]);
3610 /*
3611 * the caller will still call fop_getpage() for shared segments
3612 * to check FS write permissions. For private segments we map
3613 * file read only anyway. so no fop_getpage is needed.
3614 */
3615 for (i = 0; i < pages; i++) {
3616 ASSERT(ppa[i] != NULL);
3617 ASSERT(PAGE_EXCL(ppa[i]));
3618 ASSERT(ppa[i]->p_vnode == vp);
3619 ASSERT(ppa[i]->p_offset ==
3620 start_off + (i << PAGESHIFT));
3621 page_unlock(ppa[i]);
3622 }
3623 ppa[0] = NULL;
3624 }
3625
3626 return (1);
3627 out:
3628 /*
3629 * Do the cleanup. Unlock target pages we didn't relocate. They are
3630 * linked on targ_pplist by root pages. reassemble unused replacement
3631 * and io pages back to pplist.
3632 */
3633 if (io_pplist != NULL) {
3634 VM_STAT_ADD(segvnvmstats.fill_vp_pages[20]);
3635 pp = io_pplist;
3636 do {
3637 ASSERT(pp->p_vnode == vp);
3638 ASSERT(pp->p_offset == io_off);
3639 ASSERT(page_iolock_assert(pp));
3640 page_io_unlock(pp);
3641 page_hashout(pp, false);
3642 io_off += PAGESIZE;
3643 } while ((pp = pp->p_next) != io_pplist);
3644 page_list_concat(&io_pplist, &pplist);
3645 pplist = io_pplist;
3646 }
3647 tmp_pplist = NULL;
3648 while (targ_pplist != NULL) {
3649 VM_STAT_ADD(segvnvmstats.fill_vp_pages[21]);
3650 pp = targ_pplist;
3651 ASSERT(PAGE_EXCL(pp));
3652 page_sub(&targ_pplist, pp);
3653
3654 pszc = pp->p_szc;
3655 ppages = page_get_pagecnt(pszc);
3656 ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3657
3658 if (pszc != 0) {
3659 group_page_unlock(pp);
3660 }
3661 page_unlock(pp);
3662
3663 pp = repl_pplist;
3664 ASSERT(pp != NULL);
3665 ASSERT(PAGE_EXCL(pp));
3666 ASSERT(pp->p_szc == szc);
3667 page_sub(&repl_pplist, pp);
3668
3669 ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3670
3671 /* relink replacement page */
3672 page_list_concat(&tmp_pplist, &pp);
3673 while (--ppages != 0) {
3674 VM_STAT_ADD(segvnvmstats.fill_vp_pages[22]);
3675 pp++;
3676 ASSERT(PAGE_EXCL(pp));
3677 ASSERT(pp->p_szc == szc);
3678 page_list_concat(&tmp_pplist, &pp);
3679 }
3680 }
3681 if (tmp_pplist != NULL) {
3682 VM_STAT_ADD(segvnvmstats.fill_vp_pages[23]);
3683 page_list_concat(&tmp_pplist, &pplist);
3684 pplist = tmp_pplist;
3685 }
3686 /*
3687 * at this point all pages are either on done_pplist or
3688 * pplist. They can't be all on done_pplist otherwise
3689 * we'd've been done.
3690 */
3691 ASSERT(pplist != NULL);
3692 if (nios != 0) {
3693 VM_STAT_ADD(segvnvmstats.fill_vp_pages[24]);
3694 pp = pplist;
3695 do {
3696 VM_STAT_ADD(segvnvmstats.fill_vp_pages[25]);
3697 ASSERT(pp->p_szc == szc);
3698 ASSERT(PAGE_EXCL(pp));
3699 ASSERT(pp->p_vnode != vp);
3700 pp->p_szc = 0;
3701 } while ((pp = pp->p_next) != pplist);
3702
3703 pp = done_pplist;
3704 do {
3705 VM_STAT_ADD(segvnvmstats.fill_vp_pages[26]);
3706 ASSERT(pp->p_szc == szc);
3707 ASSERT(PAGE_EXCL(pp));
3708 ASSERT(pp->p_vnode == vp);
3709 pp->p_szc = 0;
3710 } while ((pp = pp->p_next) != done_pplist);
3711
3712 while (pplist != NULL) {
3713 VM_STAT_ADD(segvnvmstats.fill_vp_pages[27]);
3714 pp = pplist;
3715 page_sub(&pplist, pp);
3716 page_free(pp, 0);
3717 }
3718
3719 while (done_pplist != NULL) {
3720 VM_STAT_ADD(segvnvmstats.fill_vp_pages[28]);
3721 pp = done_pplist;
3722 page_sub(&done_pplist, pp);
3723 page_unlock(pp);
3724 }
3725 *ppplist = NULL;
3726 return (0);
3727 }
3728 ASSERT(pplist == *ppplist);
3729 if (io_err) {
3730 VM_STAT_ADD(segvnvmstats.fill_vp_pages[29]);
3731 /*
3732 * don't downsize on io error.
3733 * see if vop_getpage succeeds.
3734 * pplist may still be used in this case
3735 * for relocations.
3736 */
3737 return (0);
3738 }
3739 VM_STAT_ADD(segvnvmstats.fill_vp_pages[30]);
3740 page_free_replacement_page(pplist);
3741 page_create_putback(pages);
3742 *ppplist = NULL;
3743 return (0);
3744 }
3745
3746 int segvn_anypgsz = 0;
3747
3748 #define SEGVN_RESTORE_SOFTLOCK_VP(type, pages) \
3749 if ((type) == F_SOFTLOCK) { \
3750 atomic_add_long((ulong_t *)&(svd)->softlockcnt, \
3751 -(pages)); \
3752 }
3753
3754 #define SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot) \
3755 if (IS_VMODSORT((ppa)[0]->p_vnode)) { \
3756 if ((rw) == S_WRITE) { \
3757 for (i = 0; i < (pages); i++) { \
3758 ASSERT((ppa)[i]->p_vnode == \
3759 (ppa)[0]->p_vnode); \
3760 hat_setmod((ppa)[i]); \
3761 } \
3762 } else if ((rw) != S_OTHER && \
3763 ((prot) & (vpprot) & PROT_WRITE)) { \
3764 for (i = 0; i < (pages); i++) { \
3765 ASSERT((ppa)[i]->p_vnode == \
3766 (ppa)[0]->p_vnode); \
3767 if (!hat_ismod((ppa)[i])) { \
3768 prot &= ~PROT_WRITE; \
3769 break; \
3770 } \
3771 } \
3772 } \
3773 }
3774
3775 #define SEGVN_VMSTAT_FLTVNPAGES(idx) \
3776 VM_STAT_ADD(segvnvmstats.fltvnpages[(idx)]);
3777
3778 static faultcode_t
3779 segvn_fault_vnodepages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
3780 caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
3781 caddr_t eaddr, int brkcow)
3782 {
3783 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
3784 struct anon_map *amp = svd->amp;
3785 uchar_t segtype = svd->type;
3786 uint_t szc = seg->s_szc;
3787 size_t pgsz = page_get_pagesize(szc);
3788 size_t maxpgsz = pgsz;
3789 pgcnt_t pages = btop(pgsz);
3790 pgcnt_t maxpages = pages;
3791 size_t ppasize = (pages + 1) * sizeof (page_t *);
3792 caddr_t a = lpgaddr;
3793 caddr_t maxlpgeaddr = lpgeaddr;
3794 uoff_t off = svd->offset + (uintptr_t)(a - seg->s_base);
3795 ulong_t aindx = svd->anon_index + seg_page(seg, a);
3796 struct vpage *vpage = (svd->vpage != NULL) ?
3797 &svd->vpage[seg_page(seg, a)] : NULL;
3798 vnode_t *vp = svd->vp;
3799 page_t **ppa;
3800 uint_t pszc;
3801 size_t ppgsz;
3802 pgcnt_t ppages;
3803 faultcode_t err = 0;
3804 int ierr;
3805 int vop_size_err = 0;
3806 uint_t protchk, prot, vpprot;
3807 ulong_t i;
3808 int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
3809 anon_sync_obj_t an_cookie;
3810 enum seg_rw arw;
3811 int alloc_failed = 0;
3812 int adjszc_chk;
3813 struct vattr va;
3814 page_t *pplist;
3815 pfn_t pfn;
3816 int physcontig;
3817 int upgrdfail;
3818 int segvn_anypgsz_vnode = 0; /* for now map vnode with 2 page sizes */
3819 int tron = (svd->tr_state == SEGVN_TR_ON);
3820
3821 ASSERT(szc != 0);
3822 ASSERT(vp != NULL);
3823 ASSERT(brkcow == 0 || amp != NULL);
3824 ASSERT(tron == 0 || amp != NULL);
3825 ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
3826 ASSERT(!(svd->flags & MAP_NORESERVE));
3827 ASSERT(type != F_SOFTUNLOCK);
3828 ASSERT(IS_P2ALIGNED(a, maxpgsz));
3829 ASSERT(amp == NULL || IS_P2ALIGNED(aindx, maxpages));
3830 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
3831 ASSERT(seg->s_szc < NBBY * sizeof (int));
3832 ASSERT(type != F_SOFTLOCK || lpgeaddr - a == maxpgsz);
3833 ASSERT(svd->tr_state != SEGVN_TR_INIT);
3834
3835 VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltvnpages[0]);
3836 VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltvnpages[1]);
3837
3838 if (svd->flags & MAP_TEXT) {
3839 hat_flag |= HAT_LOAD_TEXT;
3840 }
3841
3842 if (svd->pageprot) {
3843 switch (rw) {
3844 case S_READ:
3845 protchk = PROT_READ;
3846 break;
3847 case S_WRITE:
3848 protchk = PROT_WRITE;
3849 break;
3850 case S_EXEC:
3851 protchk = PROT_EXEC;
3852 break;
3853 case S_OTHER:
3854 default:
3855 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
3856 break;
3857 }
3858 } else {
3859 prot = svd->prot;
3860 /* caller has already done segment level protection check. */
3861 }
3862
3863 if (rw == S_WRITE && segtype == MAP_PRIVATE) {
3864 SEGVN_VMSTAT_FLTVNPAGES(2);
3865 arw = S_READ;
3866 } else {
3867 arw = rw;
3868 }
3869
3870 ppa = kmem_alloc(ppasize, KM_SLEEP);
3871
3872 VM_STAT_COND_ADD(amp != NULL, segvnvmstats.fltvnpages[3]);
3873
3874 for (;;) {
3875 adjszc_chk = 0;
3876 for (; a < lpgeaddr; a += pgsz, off += pgsz, aindx += pages) {
3877 if (adjszc_chk) {
3878 while (szc < seg->s_szc) {
3879 uintptr_t e;
3880 uint_t tszc;
3881 tszc = segvn_anypgsz_vnode ? szc + 1 :
3882 seg->s_szc;
3883 ppgsz = page_get_pagesize(tszc);
3884 if (!IS_P2ALIGNED(a, ppgsz) ||
3885 ((alloc_failed >> tszc) & 0x1)) {
3886 break;
3887 }
3888 SEGVN_VMSTAT_FLTVNPAGES(4);
3889 szc = tszc;
3890 pgsz = ppgsz;
3891 pages = btop(pgsz);
3892 e = P2ROUNDUP((uintptr_t)eaddr, pgsz);
3893 lpgeaddr = (caddr_t)e;
3894 }
3895 }
3896
3897 again:
3898 if (IS_P2ALIGNED(a, maxpgsz) && amp != NULL) {
3899 ASSERT(IS_P2ALIGNED(aindx, maxpages));
3900 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
3901 anon_array_enter(amp, aindx, &an_cookie);
3902 if (anon_get_ptr(amp->ahp, aindx) != NULL) {
3903 SEGVN_VMSTAT_FLTVNPAGES(5);
3904 ASSERT(anon_pages(amp->ahp, aindx,
3905 maxpages) == maxpages);
3906 anon_array_exit(&an_cookie);
3907 ANON_LOCK_EXIT(&amp->a_rwlock);
3908 err = segvn_fault_anonpages(hat, seg,
3909 a, a + maxpgsz, type, rw,
3910 MAX(a, addr),
3911 MIN(a + maxpgsz, eaddr), brkcow);
3912 if (err != 0) {
3913 SEGVN_VMSTAT_FLTVNPAGES(6);
3914 goto out;
3915 }
3916 if (szc < seg->s_szc) {
3917 szc = seg->s_szc;
3918 pgsz = maxpgsz;
3919 pages = maxpages;
3920 lpgeaddr = maxlpgeaddr;
3921 }
3922 goto next;
3923 } else {
3924 ASSERT(anon_pages(amp->ahp, aindx,
3925 maxpages) == 0);
3926 SEGVN_VMSTAT_FLTVNPAGES(7);
3927 anon_array_exit(&an_cookie);
3928 ANON_LOCK_EXIT(&amp->a_rwlock);
3929 }
3930 }
3931 ASSERT(!brkcow || IS_P2ALIGNED(a, maxpgsz));
3932 ASSERT(!tron || IS_P2ALIGNED(a, maxpgsz));
3933
3934 if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
3935 ASSERT(vpage != NULL);
3936 prot = VPP_PROT(vpage);
3937 ASSERT(sameprot(seg, a, maxpgsz));
3938 if ((prot & protchk) == 0) {
3939 SEGVN_VMSTAT_FLTVNPAGES(8);
3940 err = FC_PROT;
3941 goto out;
3942 }
3943 }
3944 if (type == F_SOFTLOCK) {
3945 atomic_add_long((ulong_t *)&svd->softlockcnt,
3946 pages);
3947 }
3948
3949 pplist = NULL;
3950 physcontig = 0;
3951 ppa[0] = NULL;
3952 if (!brkcow && !tron && szc &&
3953 !page_exists_physcontig(vp, off, szc,
3954 segtype == MAP_PRIVATE ? ppa : NULL)) {
3955 SEGVN_VMSTAT_FLTVNPAGES(9);
3956 if (page_alloc_pages(vp, seg, a, &pplist, NULL,
3957 szc, 0, 0) && type != F_SOFTLOCK) {
3958 SEGVN_VMSTAT_FLTVNPAGES(10);
3959 pszc = 0;
3960 ierr = -1;
3961 alloc_failed |= (1 << szc);
3962 break;
3963 }
3964 if (pplist != NULL &&
3965 vp->v_mpssdata == SEGVN_PAGEIO) {
3966 int downsize;
3967 SEGVN_VMSTAT_FLTVNPAGES(11);
3968 physcontig = segvn_fill_vp_pages(svd,
3969 vp, off, szc, ppa, &pplist,
3970 &pszc, &downsize);
3971 ASSERT(!physcontig || pplist == NULL);
3972 if (!physcontig && downsize &&
3973 type != F_SOFTLOCK) {
3974 ASSERT(pplist == NULL);
3975 SEGVN_VMSTAT_FLTVNPAGES(12);
3976 ierr = -1;
3977 break;
3978 }
3979 ASSERT(!physcontig ||
3980 segtype == MAP_PRIVATE ||
3981 ppa[0] == NULL);
3982 if (physcontig && ppa[0] == NULL) {
3983 physcontig = 0;
3984 }
3985 }
3986 } else if (!brkcow && !tron && szc && ppa[0] != NULL) {
3987 SEGVN_VMSTAT_FLTVNPAGES(13);
3988 ASSERT(segtype == MAP_PRIVATE);
3989 physcontig = 1;
3990 }
3991
3992 if (!physcontig) {
3993 SEGVN_VMSTAT_FLTVNPAGES(14);
3994 ppa[0] = NULL;
3995 ierr = fop_getpage(vp, (offset_t)off, pgsz,
3996 &vpprot, ppa, pgsz, seg, a, arw,
3997 svd->cred, NULL);
3998 #ifdef DEBUG
3999 if (ierr == 0) {
4000 for (i = 0; i < pages; i++) {
4001 ASSERT(PAGE_LOCKED(ppa[i]));
4002 ASSERT(!PP_ISFREE(ppa[i]));
4003 ASSERT(ppa[i]->p_vnode == vp);
4004 ASSERT(ppa[i]->p_offset ==
4005 off + (i << PAGESHIFT));
4006 }
4007 }
4008 #endif /* DEBUG */
4009 if (segtype == MAP_PRIVATE) {
4010 SEGVN_VMSTAT_FLTVNPAGES(15);
4011 vpprot &= ~PROT_WRITE;
4012 }
4013 } else {
4014 ASSERT(segtype == MAP_PRIVATE);
4015 SEGVN_VMSTAT_FLTVNPAGES(16);
4016 vpprot = PROT_ALL & ~PROT_WRITE;
4017 ierr = 0;
4018 }
4019
4020 if (ierr != 0) {
4021 SEGVN_VMSTAT_FLTVNPAGES(17);
4022 if (pplist != NULL) {
4023 SEGVN_VMSTAT_FLTVNPAGES(18);
4024 page_free_replacement_page(pplist);
4025 page_create_putback(pages);
4026 }
4027 SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4028 if (a + pgsz <= eaddr) {
4029 SEGVN_VMSTAT_FLTVNPAGES(19);
4030 err = FC_MAKE_ERR(ierr);
4031 goto out;
4032 }
4033 va.va_mask = AT_SIZE;
4034 if (fop_getattr(vp, &va, 0, svd->cred, NULL)) {
4035 SEGVN_VMSTAT_FLTVNPAGES(20);
4036 err = FC_MAKE_ERR(EIO);
4037 goto out;
4038 }
4039 if (btopr(va.va_size) >= btopr(off + pgsz)) {
4040 SEGVN_VMSTAT_FLTVNPAGES(21);
4041 err = FC_MAKE_ERR(ierr);
4042 goto out;
4043 }
4044 if (btopr(va.va_size) <
4045 btopr(off + (eaddr - a))) {
4046 SEGVN_VMSTAT_FLTVNPAGES(22);
4047 err = FC_MAKE_ERR(ierr);
4048 goto out;
4049 }
4050 if (brkcow || tron || type == F_SOFTLOCK) {
4051 /* can't reduce map area */
4052 SEGVN_VMSTAT_FLTVNPAGES(23);
4053 vop_size_err = 1;
4054 goto out;
4055 }
4056 SEGVN_VMSTAT_FLTVNPAGES(24);
4057 ASSERT(szc != 0);
4058 pszc = 0;
4059 ierr = -1;
4060 break;
4061 }
4062
4063 if (amp != NULL) {
4064 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4065 anon_array_enter(amp, aindx, &an_cookie);
4066 }
4067 if (amp != NULL &&
4068 anon_get_ptr(amp->ahp, aindx) != NULL) {
4069 ulong_t taindx = P2ALIGN(aindx, maxpages);
4070
4071 SEGVN_VMSTAT_FLTVNPAGES(25);
4072 ASSERT(anon_pages(amp->ahp, taindx,
4073 maxpages) == maxpages);
4074 for (i = 0; i < pages; i++) {
4075 page_unlock(ppa[i]);
4076 }
4077 anon_array_exit(&an_cookie);
4078 ANON_LOCK_EXIT(&amp->a_rwlock);
4079 if (pplist != NULL) {
4080 page_free_replacement_page(pplist);
4081 page_create_putback(pages);
4082 }
4083 SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4084 if (szc < seg->s_szc) {
4085 SEGVN_VMSTAT_FLTVNPAGES(26);
4086 /*
4087 * For private segments SOFTLOCK
4088 * either always breaks cow (any rw
4089 * type except S_READ_NOCOW) or
4090 * address space is locked as writer
4091 * (S_READ_NOCOW case) and anon slots
4092 * can't show up on second check.
4093 * Therefore if we are here for
4094 * SOFTLOCK case it must be a cow
4095 * break but cow break never reduces
4096 * szc. text replication (tron) in
4097 * this case works as cow break.
4098 * Thus the assert below.
4099 */
4100 ASSERT(!brkcow && !tron &&
4101 type != F_SOFTLOCK);
4102 pszc = seg->s_szc;
4103 ierr = -2;
4104 break;
4105 }
4106 ASSERT(IS_P2ALIGNED(a, maxpgsz));
4107 goto again;
4108 }
4109 #ifdef DEBUG
4110 if (amp != NULL) {
4111 ulong_t taindx = P2ALIGN(aindx, maxpages);
4112 ASSERT(!anon_pages(amp->ahp, taindx, maxpages));
4113 }
4114 #endif /* DEBUG */
4115
4116 if (brkcow || tron) {
4117 ASSERT(amp != NULL);
4118 ASSERT(pplist == NULL);
4119 ASSERT(szc == seg->s_szc);
4120 ASSERT(IS_P2ALIGNED(a, maxpgsz));
4121 ASSERT(IS_P2ALIGNED(aindx, maxpages));
4122 SEGVN_VMSTAT_FLTVNPAGES(27);
4123 ierr = anon_map_privatepages(amp, aindx, szc,
4124 seg, a, prot, ppa, vpage, segvn_anypgsz,
4125 tron ? PG_LOCAL : 0, svd->cred);
4126 if (ierr != 0) {
4127 SEGVN_VMSTAT_FLTVNPAGES(28);
4128 anon_array_exit(&an_cookie);
4129 ANON_LOCK_EXIT(&amp->a_rwlock);
4130 SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4131 err = FC_MAKE_ERR(ierr);
4132 goto out;
4133 }
4134
4135 ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4136 /*
4137 * p_szc can't be changed for locked
4138 * swapfs pages.
4139 */
4140 ASSERT(svd->rcookie ==
4141 HAT_INVALID_REGION_COOKIE);
4142 hat_memload_array(hat, a, pgsz, ppa, prot,
4143 hat_flag);
4144
4145 if (!(hat_flag & HAT_LOAD_LOCK)) {
4146 SEGVN_VMSTAT_FLTVNPAGES(29);
4147 for (i = 0; i < pages; i++) {
4148 page_unlock(ppa[i]);
4149 }
4150 }
4151 anon_array_exit(&an_cookie);
4152 ANON_LOCK_EXIT(&amp->a_rwlock);
4153 goto next;
4154 }
4155
4156 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
4157 (!svd->pageprot && svd->prot == (prot & vpprot)));
4158
4159 pfn = page_pptonum(ppa[0]);
4160 /*
4161 * hat_page_demote() needs an SE_EXCL lock on one of
4162 * constituent page_t's and it decreases root's p_szc
4163 * last. This means if root's p_szc is equal szc and
4164 * all its constituent pages are locked
4165 * hat_page_demote() that could have changed p_szc to
4166 * szc is already done and no new have page_demote()
4167 * can start for this large page.
4168 */
4169
4170 /*
4171 * we need to make sure same mapping size is used for
4172 * the same address range if there's a possibility the
4173 * adddress is already mapped because hat layer panics
4174 * when translation is loaded for the range already
4175 * mapped with a different page size. We achieve it
4176 * by always using largest page size possible subject
4177 * to the constraints of page size, segment page size
4178 * and page alignment. Since mappings are invalidated
4179 * when those constraints change and make it
4180 * impossible to use previously used mapping size no
4181 * mapping size conflicts should happen.
4182 */
4183
4184 chkszc:
4185 if ((pszc = ppa[0]->p_szc) == szc &&
4186 IS_P2ALIGNED(pfn, pages)) {
4187
4188 SEGVN_VMSTAT_FLTVNPAGES(30);
4189 #ifdef DEBUG
4190 for (i = 0; i < pages; i++) {
4191 ASSERT(PAGE_LOCKED(ppa[i]));
4192 ASSERT(!PP_ISFREE(ppa[i]));
4193 ASSERT(page_pptonum(ppa[i]) ==
4194 pfn + i);
4195 ASSERT(ppa[i]->p_szc == szc);
4196 ASSERT(ppa[i]->p_vnode == vp);
4197 ASSERT(ppa[i]->p_offset ==
4198 off + (i << PAGESHIFT));
4199 }
4200 #endif /* DEBUG */
4201 /*
4202 * All pages are of szc we need and they are
4203 * all locked so they can't change szc. load
4204 * translations.
4205 *
4206 * if page got promoted since last check
4207 * we don't need pplist.
4208 */
4209 if (pplist != NULL) {
4210 page_free_replacement_page(pplist);
4211 page_create_putback(pages);
4212 }
4213 if (PP_ISMIGRATE(ppa[0])) {
4214 page_migrate(seg, a, ppa, pages);
4215 }
4216 SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4217 prot, vpprot);
4218 hat_memload_array_region(hat, a, pgsz,
4219 ppa, prot & vpprot, hat_flag,
4220 svd->rcookie);
4221
4222 if (!(hat_flag & HAT_LOAD_LOCK)) {
4223 for (i = 0; i < pages; i++) {
4224 page_unlock(ppa[i]);
4225 }
4226 }
4227 if (amp != NULL) {
4228 anon_array_exit(&an_cookie);
4229 ANON_LOCK_EXIT(&amp->a_rwlock);
4230 }
4231 goto next;
4232 }
4233
4234 /*
4235 * See if upsize is possible.
4236 */
4237 if (pszc > szc && szc < seg->s_szc &&
4238 (segvn_anypgsz_vnode || pszc >= seg->s_szc)) {
4239 pgcnt_t aphase;
4240 uint_t pszc1 = MIN(pszc, seg->s_szc);
4241 ppgsz = page_get_pagesize(pszc1);
4242 ppages = btop(ppgsz);
4243 aphase = btop(P2PHASE((uintptr_t)a, ppgsz));
4244
4245 ASSERT(type != F_SOFTLOCK);
4246
4247 SEGVN_VMSTAT_FLTVNPAGES(31);
4248 if (aphase != P2PHASE(pfn, ppages)) {
4249 segvn_faultvnmpss_align_err4++;
4250 } else {
4251 SEGVN_VMSTAT_FLTVNPAGES(32);
4252 if (pplist != NULL) {
4253 page_t *pl = pplist;
4254 page_free_replacement_page(pl);
4255 page_create_putback(pages);
4256 }
4257 for (i = 0; i < pages; i++) {
4258 page_unlock(ppa[i]);
4259 }
4260 if (amp != NULL) {
4261 anon_array_exit(&an_cookie);
4262 ANON_LOCK_EXIT(&amp->a_rwlock);
4263 }
4264 pszc = pszc1;
4265 ierr = -2;
4266 break;
4267 }
4268 }
4269
4270 /*
4271 * check if we should use smallest mapping size.
4272 */
4273 upgrdfail = 0;
4274 if (szc == 0 ||
4275 (pszc >= szc &&
4276 !IS_P2ALIGNED(pfn, pages)) ||
4277 (pszc < szc &&
4278 !segvn_full_szcpages(ppa, szc, &upgrdfail,
4279 &pszc))) {
4280
4281 if (upgrdfail && type != F_SOFTLOCK) {
4282 /*
4283 * segvn_full_szcpages failed to lock
4284 * all pages EXCL. Size down.
4285 */
4286 ASSERT(pszc < szc);
4287
4288 SEGVN_VMSTAT_FLTVNPAGES(33);
4289
4290 if (pplist != NULL) {
4291 page_t *pl = pplist;
4292 page_free_replacement_page(pl);
4293 page_create_putback(pages);
4294 }
4295
4296 for (i = 0; i < pages; i++) {
4297 page_unlock(ppa[i]);
4298 }
4299 if (amp != NULL) {
4300 anon_array_exit(&an_cookie);
4301 ANON_LOCK_EXIT(&amp->a_rwlock);
4302 }
4303 ierr = -1;
4304 break;
4305 }
4306 if (szc != 0 && !upgrdfail) {
4307 segvn_faultvnmpss_align_err5++;
4308 }
4309 SEGVN_VMSTAT_FLTVNPAGES(34);
4310 if (pplist != NULL) {
4311 page_free_replacement_page(pplist);
4312 page_create_putback(pages);
4313 }
4314 SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4315 prot, vpprot);
4316 if (upgrdfail && segvn_anypgsz_vnode) {
4317 /* SOFTLOCK case */
4318 hat_memload_array_region(hat, a, pgsz,
4319 ppa, prot & vpprot, hat_flag,
4320 svd->rcookie);
4321 } else {
4322 for (i = 0; i < pages; i++) {
4323 hat_memload_region(hat,
4324 a + (i << PAGESHIFT),
4325 ppa[i], prot & vpprot,
4326 hat_flag, svd->rcookie);
4327 }
4328 }
4329 if (!(hat_flag & HAT_LOAD_LOCK)) {
4330 for (i = 0; i < pages; i++) {
4331 page_unlock(ppa[i]);
4332 }
4333 }
4334 if (amp != NULL) {
4335 anon_array_exit(&an_cookie);
4336 ANON_LOCK_EXIT(&amp->a_rwlock);
4337 }
4338 goto next;
4339 }
4340
4341 if (pszc == szc) {
4342 /*
4343 * segvn_full_szcpages() upgraded pages szc.
4344 */
4345 ASSERT(pszc == ppa[0]->p_szc);
4346 ASSERT(IS_P2ALIGNED(pfn, pages));
4347 goto chkszc;
4348 }
4349
4350 if (pszc > szc) {
4351 kmutex_t *szcmtx;
4352 SEGVN_VMSTAT_FLTVNPAGES(35);
4353 /*
4354 * p_szc of ppa[0] can change since we haven't
4355 * locked all constituent pages. Call
4356 * page_lock_szc() to prevent szc changes.
4357 * This should be a rare case that happens when
4358 * multiple segments use a different page size
4359 * to map the same file offsets.
4360 */
4361 szcmtx = page_szc_lock(ppa[0]);
4362 pszc = ppa[0]->p_szc;
4363 ASSERT(szcmtx != NULL || pszc == 0);
4364 ASSERT(ppa[0]->p_szc <= pszc);
4365 if (pszc <= szc) {
4366 SEGVN_VMSTAT_FLTVNPAGES(36);
4367 if (szcmtx != NULL) {
4368 mutex_exit(szcmtx);
4369 }
4370 goto chkszc;
4371 }
4372 if (pplist != NULL) {
4373 /*
4374 * page got promoted since last check.
4375 * we don't need preaalocated large
4376 * page.
4377 */
4378 SEGVN_VMSTAT_FLTVNPAGES(37);
4379 page_free_replacement_page(pplist);
4380 page_create_putback(pages);
4381 }
4382 SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4383 prot, vpprot);
4384 hat_memload_array_region(hat, a, pgsz, ppa,
4385 prot & vpprot, hat_flag, svd->rcookie);
4386 mutex_exit(szcmtx);
4387 if (!(hat_flag & HAT_LOAD_LOCK)) {
4388 for (i = 0; i < pages; i++) {
4389 page_unlock(ppa[i]);
4390 }
4391 }
4392 if (amp != NULL) {
4393 anon_array_exit(&an_cookie);
4394 ANON_LOCK_EXIT(&amp->a_rwlock);
4395 }
4396 goto next;
4397 }
4398
4399 /*
4400 * if page got demoted since last check
4401 * we could have not allocated larger page.
4402 * allocate now.
4403 */
4404 if (pplist == NULL &&
4405 page_alloc_pages(vp, seg, a, &pplist, NULL,
4406 szc, 0, 0) && type != F_SOFTLOCK) {
4407 SEGVN_VMSTAT_FLTVNPAGES(38);
4408 for (i = 0; i < pages; i++) {
4409 page_unlock(ppa[i]);
4410 }
4411 if (amp != NULL) {
4412 anon_array_exit(&an_cookie);
4413 ANON_LOCK_EXIT(&amp->a_rwlock);
4414 }
4415 ierr = -1;
4416 alloc_failed |= (1 << szc);
4417 break;
4418 }
4419
4420 SEGVN_VMSTAT_FLTVNPAGES(39);
4421
4422 if (pplist != NULL) {
4423 segvn_relocate_pages(ppa, pplist);
4424 #ifdef DEBUG
4425 } else {
4426 ASSERT(type == F_SOFTLOCK);
4427 SEGVN_VMSTAT_FLTVNPAGES(40);
4428 #endif /* DEBUG */
4429 }
4430
4431 SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot);
4432
4433 if (pplist == NULL && segvn_anypgsz_vnode == 0) {
4434 ASSERT(type == F_SOFTLOCK);
4435 for (i = 0; i < pages; i++) {
4436 ASSERT(ppa[i]->p_szc < szc);
4437 hat_memload_region(hat,
4438 a + (i << PAGESHIFT),
4439 ppa[i], prot & vpprot, hat_flag,
4440 svd->rcookie);
4441 }
4442 } else {
4443 ASSERT(pplist != NULL || type == F_SOFTLOCK);
4444 hat_memload_array_region(hat, a, pgsz, ppa,
4445 prot & vpprot, hat_flag, svd->rcookie);
4446 }
4447 if (!(hat_flag & HAT_LOAD_LOCK)) {
4448 for (i = 0; i < pages; i++) {
4449 ASSERT(PAGE_SHARED(ppa[i]));
4450 page_unlock(ppa[i]);
4451 }
4452 }
4453 if (amp != NULL) {
4454 anon_array_exit(&an_cookie);
4455 ANON_LOCK_EXIT(&amp->a_rwlock);
4456 }
4457
4458 next:
4459 if (vpage != NULL) {
4460 vpage += pages;
4461 }
4462 adjszc_chk = 1;
4463 }
4464 if (a == lpgeaddr)
4465 break;
4466 ASSERT(a < lpgeaddr);
4467
4468 ASSERT(!brkcow && !tron && type != F_SOFTLOCK);
4469
4470 /*
4471 * ierr == -1 means we failed to map with a large page.
4472 * (either due to allocation/relocation failures or
4473 * misalignment with other mappings to this file.
4474 *
4475 * ierr == -2 means some other thread allocated a large page
4476 * after we gave up tp map with a large page. retry with
4477 * larger mapping.
4478 */
4479 ASSERT(ierr == -1 || ierr == -2);
4480 ASSERT(ierr == -2 || szc != 0);
4481 ASSERT(ierr == -1 || szc < seg->s_szc);
4482 if (ierr == -2) {
4483 SEGVN_VMSTAT_FLTVNPAGES(41);
4484 ASSERT(pszc > szc && pszc <= seg->s_szc);
4485 szc = pszc;
4486 } else if (segvn_anypgsz_vnode) {
4487 SEGVN_VMSTAT_FLTVNPAGES(42);
4488 szc--;
4489 } else {
4490 SEGVN_VMSTAT_FLTVNPAGES(43);
4491 ASSERT(pszc < szc);
4492 /*
4493 * other process created pszc large page.
4494 * but we still have to drop to 0 szc.
4495 */
4496 szc = 0;
4497 }
4498
4499 pgsz = page_get_pagesize(szc);
4500 pages = btop(pgsz);
4501 if (ierr == -2) {
4502 /*
4503 * Size up case. Note lpgaddr may only be needed for
4504 * softlock case so we don't adjust it here.
4505 */
4506 a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4507 ASSERT(a >= lpgaddr);
4508 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4509 off = svd->offset + (uintptr_t)(a - seg->s_base);
4510 aindx = svd->anon_index + seg_page(seg, a);
4511 vpage = (svd->vpage != NULL) ?
4512 &svd->vpage[seg_page(seg, a)] : NULL;
4513 } else {
4514 /*
4515 * Size down case. Note lpgaddr may only be needed for
4516 * softlock case so we don't adjust it here.
4517 */
4518 ASSERT(IS_P2ALIGNED(a, pgsz));
4519 ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4520 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4521 ASSERT(a < lpgeaddr);
4522 if (a < addr) {
4523 SEGVN_VMSTAT_FLTVNPAGES(44);
4524 /*
4525 * The beginning of the large page region can
4526 * be pulled to the right to make a smaller
4527 * region. We haven't yet faulted a single
4528 * page.
4529 */
4530 a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4531 ASSERT(a >= lpgaddr);
4532 off = svd->offset +
4533 (uintptr_t)(a - seg->s_base);
4534 aindx = svd->anon_index + seg_page(seg, a);
4535 vpage = (svd->vpage != NULL) ?
4536 &svd->vpage[seg_page(seg, a)] : NULL;
4537 }
4538 }
4539 }
4540 out:
4541 kmem_free(ppa, ppasize);
4542 if (!err && !vop_size_err) {
4543 SEGVN_VMSTAT_FLTVNPAGES(45);
4544 return (0);
4545 }
4546 if (type == F_SOFTLOCK && a > lpgaddr) {
4547 SEGVN_VMSTAT_FLTVNPAGES(46);
4548 segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4549 }
4550 if (!vop_size_err) {
4551 SEGVN_VMSTAT_FLTVNPAGES(47);
4552 return (err);
4553 }
4554 ASSERT(brkcow || tron || type == F_SOFTLOCK);
4555 /*
4556 * Large page end is mapped beyond the end of file and it's a cow
4557 * fault (can be a text replication induced cow) or softlock so we can't
4558 * reduce the map area. For now just demote the segment. This should
4559 * really only happen if the end of the file changed after the mapping
4560 * was established since when large page segments are created we make
4561 * sure they don't extend beyond the end of the file.
4562 */
4563 SEGVN_VMSTAT_FLTVNPAGES(48);
4564
4565 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4566 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4567 err = 0;
4568 if (seg->s_szc != 0) {
4569 segvn_fltvnpages_clrszc_cnt++;
4570 ASSERT(svd->softlockcnt == 0);
4571 err = segvn_clrszc(seg);
4572 if (err != 0) {
4573 segvn_fltvnpages_clrszc_err++;
4574 }
4575 }
4576 ASSERT(err || seg->s_szc == 0);
4577 SEGVN_LOCK_DOWNGRADE(seg->s_as, &svd->lock);
4578 /* segvn_fault will do its job as if szc had been zero to begin with */
4579 return (err == 0 ? IE_RETRY : FC_MAKE_ERR(err));
4580 }
4581
4582 /*
4583 * This routine will attempt to fault in one large page.
4584 * it will use smaller pages if that fails.
4585 * It should only be called for pure anonymous segments.
4586 */
4587 static faultcode_t
4588 segvn_fault_anonpages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
4589 caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
4590 caddr_t eaddr, int brkcow)
4591 {
4592 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4593 struct anon_map *amp = svd->amp;
4594 uchar_t segtype = svd->type;
4595 uint_t szc = seg->s_szc;
4596 size_t pgsz = page_get_pagesize(szc);
4597 size_t maxpgsz = pgsz;
4598 pgcnt_t pages = btop(pgsz);
4599 uint_t ppaszc = szc;
4600 caddr_t a = lpgaddr;
4601 ulong_t aindx = svd->anon_index + seg_page(seg, a);
4602 struct vpage *vpage = (svd->vpage != NULL) ?
4603 &svd->vpage[seg_page(seg, a)] : NULL;
4604 page_t **ppa;
4605 uint_t ppa_szc;
4606 faultcode_t err;
4607 int ierr;
4608 uint_t protchk, prot, vpprot;
4609 ulong_t i;
4610 int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
4611 anon_sync_obj_t cookie;
4612 int adjszc_chk;
4613 int pgflags = (svd->tr_state == SEGVN_TR_ON) ? PG_LOCAL : 0;
4614
4615 ASSERT(szc != 0);
4616 ASSERT(amp != NULL);
4617 ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
4618 ASSERT(!(svd->flags & MAP_NORESERVE));
4619 ASSERT(type != F_SOFTUNLOCK);
4620 ASSERT(IS_P2ALIGNED(a, maxpgsz));
4621 ASSERT(!brkcow || svd->tr_state == SEGVN_TR_OFF);
4622 ASSERT(svd->tr_state != SEGVN_TR_INIT);
4623
4624 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
4625
4626 VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltanpages[0]);
4627 VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltanpages[1]);
4628
4629 if (svd->flags & MAP_TEXT) {
4630 hat_flag |= HAT_LOAD_TEXT;
4631 }
4632
4633 if (svd->pageprot) {
4634 switch (rw) {
4635 case S_READ:
4636 protchk = PROT_READ;
4637 break;
4638 case S_WRITE:
4639 protchk = PROT_WRITE;
4640 break;
4641 case S_EXEC:
4642 protchk = PROT_EXEC;
4643 break;
4644 case S_OTHER:
4645 default:
4646 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4647 break;
4648 }
4649 VM_STAT_ADD(segvnvmstats.fltanpages[2]);
4650 } else {
4651 prot = svd->prot;
4652 /* caller has already done segment level protection check. */
4653 }
4654
4655 ppa = kmem_cache_alloc(segvn_szc_cache[ppaszc], KM_SLEEP);
4656 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4657 for (;;) {
4658 adjszc_chk = 0;
4659 for (; a < lpgeaddr; a += pgsz, aindx += pages) {
4660 if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
4661 VM_STAT_ADD(segvnvmstats.fltanpages[3]);
4662 ASSERT(vpage != NULL);
4663 prot = VPP_PROT(vpage);
4664 ASSERT(sameprot(seg, a, maxpgsz));
4665 if ((prot & protchk) == 0) {
4666 err = FC_PROT;
4667 goto error;
4668 }
4669 }
4670 if (adjszc_chk && IS_P2ALIGNED(a, maxpgsz) &&
4671 pgsz < maxpgsz) {
4672 ASSERT(a > lpgaddr);
4673 szc = seg->s_szc;
4674 pgsz = maxpgsz;
4675 pages = btop(pgsz);
4676 ASSERT(IS_P2ALIGNED(aindx, pages));
4677 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr,
4678 pgsz);
4679 }
4680 if (type == F_SOFTLOCK) {
4681 atomic_add_long((ulong_t *)&svd->softlockcnt,
4682 pages);
4683 }
4684 anon_array_enter(amp, aindx, &cookie);
4685 ppa_szc = (uint_t)-1;
4686 ierr = anon_map_getpages(amp, aindx, szc, seg, a,
4687 prot, &vpprot, ppa, &ppa_szc, vpage, rw, brkcow,
4688 segvn_anypgsz, pgflags, svd->cred);
4689 if (ierr != 0) {
4690 anon_array_exit(&cookie);
4691 VM_STAT_ADD(segvnvmstats.fltanpages[4]);
4692 if (type == F_SOFTLOCK) {
4693 atomic_add_long(
4694 (ulong_t *)&svd->softlockcnt,
4695 -pages);
4696 }
4697 if (ierr > 0) {
4698 VM_STAT_ADD(segvnvmstats.fltanpages[6]);
4699 err = FC_MAKE_ERR(ierr);
4700 goto error;
4701 }
4702 break;
4703 }
4704
4705 ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4706
4707 ASSERT(segtype == MAP_SHARED ||
4708 ppa[0]->p_szc <= szc);
4709 ASSERT(segtype == MAP_PRIVATE ||
4710 ppa[0]->p_szc >= szc);
4711
4712 /*
4713 * Handle pages that have been marked for migration
4714 */
4715 if (lgrp_optimizations())
4716 page_migrate(seg, a, ppa, pages);
4717
4718 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
4719
4720 if (segtype == MAP_SHARED) {
4721 vpprot |= PROT_WRITE;
4722 }
4723
4724 hat_memload_array(hat, a, pgsz, ppa,
4725 prot & vpprot, hat_flag);
4726
4727 if (hat_flag & HAT_LOAD_LOCK) {
4728 VM_STAT_ADD(segvnvmstats.fltanpages[7]);
4729 } else {
4730 VM_STAT_ADD(segvnvmstats.fltanpages[8]);
4731 for (i = 0; i < pages; i++)
4732 page_unlock(ppa[i]);
4733 }
4734 if (vpage != NULL)
4735 vpage += pages;
4736
4737 anon_array_exit(&cookie);
4738 adjszc_chk = 1;
4739 }
4740 if (a == lpgeaddr)
4741 break;
4742 ASSERT(a < lpgeaddr);
4743 /*
4744 * ierr == -1 means we failed to allocate a large page.
4745 * so do a size down operation.
4746 *
4747 * ierr == -2 means some other process that privately shares
4748 * pages with this process has allocated a larger page and we
4749 * need to retry with larger pages. So do a size up
4750 * operation. This relies on the fact that large pages are
4751 * never partially shared i.e. if we share any constituent
4752 * page of a large page with another process we must share the
4753 * entire large page. Note this cannot happen for SOFTLOCK
4754 * case, unless current address (a) is at the beginning of the
4755 * next page size boundary because the other process couldn't
4756 * have relocated locked pages.
4757 */
4758 ASSERT(ierr == -1 || ierr == -2);
4759
4760 if (segvn_anypgsz) {
4761 ASSERT(ierr == -2 || szc != 0);
4762 ASSERT(ierr == -1 || szc < seg->s_szc);
4763 szc = (ierr == -1) ? szc - 1 : szc + 1;
4764 } else {
4765 /*
4766 * For non COW faults and segvn_anypgsz == 0
4767 * we need to be careful not to loop forever
4768 * if existing page is found with szc other
4769 * than 0 or seg->s_szc. This could be due
4770 * to page relocations on behalf of DR or
4771 * more likely large page creation. For this
4772 * case simply re-size to existing page's szc
4773 * if returned by anon_map_getpages().
4774 */
4775 if (ppa_szc == (uint_t)-1) {
4776 szc = (ierr == -1) ? 0 : seg->s_szc;
4777 } else {
4778 ASSERT(ppa_szc <= seg->s_szc);
4779 ASSERT(ierr == -2 || ppa_szc < szc);
4780 ASSERT(ierr == -1 || ppa_szc > szc);
4781 szc = ppa_szc;
4782 }
4783 }
4784
4785 pgsz = page_get_pagesize(szc);
4786 pages = btop(pgsz);
4787 ASSERT(type != F_SOFTLOCK || ierr == -1 ||
4788 (IS_P2ALIGNED(a, pgsz) && IS_P2ALIGNED(lpgeaddr, pgsz)));
4789 if (type == F_SOFTLOCK) {
4790 /*
4791 * For softlocks we cannot reduce the fault area
4792 * (calculated based on the largest page size for this
4793 * segment) for size down and a is already next
4794 * page size aligned as assertted above for size
4795 * ups. Therefore just continue in case of softlock.
4796 */
4797 VM_STAT_ADD(segvnvmstats.fltanpages[9]);
4798 continue; /* keep lint happy */
4799 } else if (ierr == -2) {
4800
4801 /*
4802 * Size up case. Note lpgaddr may only be needed for
4803 * softlock case so we don't adjust it here.
4804 */
4805 VM_STAT_ADD(segvnvmstats.fltanpages[10]);
4806 a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4807 ASSERT(a >= lpgaddr);
4808 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4809 aindx = svd->anon_index + seg_page(seg, a);
4810 vpage = (svd->vpage != NULL) ?
4811 &svd->vpage[seg_page(seg, a)] : NULL;
4812 } else {
4813 /*
4814 * Size down case. Note lpgaddr may only be needed for
4815 * softlock case so we don't adjust it here.
4816 */
4817 VM_STAT_ADD(segvnvmstats.fltanpages[11]);
4818 ASSERT(IS_P2ALIGNED(a, pgsz));
4819 ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4820 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4821 ASSERT(a < lpgeaddr);
4822 if (a < addr) {
4823 /*
4824 * The beginning of the large page region can
4825 * be pulled to the right to make a smaller
4826 * region. We haven't yet faulted a single
4827 * page.
4828 */
4829 VM_STAT_ADD(segvnvmstats.fltanpages[12]);
4830 a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4831 ASSERT(a >= lpgaddr);
4832 aindx = svd->anon_index + seg_page(seg, a);
4833 vpage = (svd->vpage != NULL) ?
4834 &svd->vpage[seg_page(seg, a)] : NULL;
4835 }
4836 }
4837 }
4838 VM_STAT_ADD(segvnvmstats.fltanpages[13]);
4839 ANON_LOCK_EXIT(&amp->a_rwlock);
4840 kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4841 return (0);
4842 error:
4843 VM_STAT_ADD(segvnvmstats.fltanpages[14]);
4844 ANON_LOCK_EXIT(&amp->a_rwlock);
4845 kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4846 if (type == F_SOFTLOCK && a > lpgaddr) {
4847 VM_STAT_ADD(segvnvmstats.fltanpages[15]);
4848 segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4849 }
4850 return (err);
4851 }
4852
4853 int fltadvice = 1; /* set to free behind pages for sequential access */
4854
4855 /*
4856 * This routine is called via a machine specific fault handling routine.
4857 * It is also called by software routines wishing to lock or unlock
4858 * a range of addresses.
4859 *
4860 * Here is the basic algorithm:
4861 * If unlocking
4862 * Call segvn_softunlock
4863 * Return
4864 * endif
4865 * Checking and set up work
4866 * If we will need some non-anonymous pages
4867 * Call fop_getpage over the range of non-anonymous pages
4868 * endif
4869 * Loop over all addresses requested
4870 * Call segvn_faultpage passing in page list
4871 * to load up translations and handle anonymous pages
4872 * endloop
4873 * Load up translation to any additional pages in page list not
4874 * already handled that fit into this segment
4875 */
4876 static faultcode_t
4877 segvn_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
4878 enum fault_type type, enum seg_rw rw)
4879 {
4880 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4881 page_t **plp, **ppp, *pp;
4882 uoff_t off;
4883 caddr_t a;
4884 struct vpage *vpage;
4885 uint_t vpprot, prot;
4886 int err;
4887 page_t *pl[FAULT_TMP_PAGES_NUM + 1];
4888 size_t plsz, pl_alloc_sz;
4889 size_t page;
4890 ulong_t anon_index;
4891 struct anon_map *amp;
4892 int dogetpage = 0;
4893 caddr_t lpgaddr, lpgeaddr;
4894 size_t pgsz;
4895 anon_sync_obj_t cookie;
4896 int brkcow = BREAK_COW_SHARE(rw, type, svd->type);
4897
4898 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
4899 ASSERT(svd->amp == NULL || svd->rcookie == HAT_INVALID_REGION_COOKIE);
4900
4901 /*
4902 * First handle the easy stuff
4903 */
4904 if (type == F_SOFTUNLOCK) {
4905 if (rw == S_READ_NOCOW) {
4906 rw = S_READ;
4907 ASSERT(AS_WRITE_HELD(seg->s_as));
4908 }
4909 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4910 pgsz = (seg->s_szc == 0) ? PAGESIZE :
4911 page_get_pagesize(seg->s_szc);
4912 VM_STAT_COND_ADD(pgsz > PAGESIZE, segvnvmstats.fltanpages[16]);
4913 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
4914 segvn_softunlock(seg, lpgaddr, lpgeaddr - lpgaddr, rw);
4915 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4916 return (0);
4917 }
4918
4919 ASSERT(svd->tr_state == SEGVN_TR_OFF ||
4920 !HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
4921 if (brkcow == 0) {
4922 if (svd->tr_state == SEGVN_TR_INIT) {
4923 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4924 if (svd->tr_state == SEGVN_TR_INIT) {
4925 ASSERT(svd->vp != NULL && svd->amp == NULL);
4926 ASSERT(svd->flags & MAP_TEXT);
4927 ASSERT(svd->type == MAP_PRIVATE);
4928 segvn_textrepl(seg);
4929 ASSERT(svd->tr_state != SEGVN_TR_INIT);
4930 ASSERT(svd->tr_state != SEGVN_TR_ON ||
4931 svd->amp != NULL);
4932 }
4933 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4934 }
4935 } else if (svd->tr_state != SEGVN_TR_OFF) {
4936 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4937
4938 if (rw == S_WRITE && svd->tr_state != SEGVN_TR_OFF) {
4939 ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
4940 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4941 return (FC_PROT);
4942 }
4943
4944 if (svd->tr_state == SEGVN_TR_ON) {
4945 ASSERT(svd->vp != NULL && svd->amp != NULL);
4946 segvn_textunrepl(seg, 0);
4947 ASSERT(svd->amp == NULL &&
4948 svd->tr_state == SEGVN_TR_OFF);
4949 } else if (svd->tr_state != SEGVN_TR_OFF) {
4950 svd->tr_state = SEGVN_TR_OFF;
4951 }
4952 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
4953 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4954 }
4955
4956 top:
4957 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4958
4959 /*
4960 * If we have the same protections for the entire segment,
4961 * insure that the access being attempted is legitimate.
4962 */
4963
4964 if (svd->pageprot == 0) {
4965 uint_t protchk;
4966
4967 switch (rw) {
4968 case S_READ:
4969 case S_READ_NOCOW:
4970 protchk = PROT_READ;
4971 break;
4972 case S_WRITE:
4973 protchk = PROT_WRITE;
4974 break;
4975 case S_EXEC:
4976 protchk = PROT_EXEC;
4977 break;
4978 case S_OTHER:
4979 default:
4980 protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4981 break;
4982 }
4983
4984 if ((svd->prot & protchk) == 0) {
4985 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4986 return (FC_PROT); /* illegal access type */
4987 }
4988 }
4989
4990 if (brkcow && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
4991 /* this must be SOFTLOCK S_READ fault */
4992 ASSERT(svd->amp == NULL);
4993 ASSERT(svd->tr_state == SEGVN_TR_OFF);
4994 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4995 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4996 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
4997 /*
4998 * this must be the first ever non S_READ_NOCOW
4999 * softlock for this segment.
5000 */
5001 ASSERT(svd->softlockcnt == 0);
5002 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5003 HAT_REGION_TEXT);
5004 svd->rcookie = HAT_INVALID_REGION_COOKIE;
5005 }
5006 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5007 goto top;
5008 }
5009
5010 /*
5011 * We can't allow the long term use of softlocks for vmpss segments,
5012 * because in some file truncation cases we should be able to demote
5013 * the segment, which requires that there are no softlocks. The
5014 * only case where it's ok to allow a SOFTLOCK fault against a vmpss
5015 * segment is S_READ_NOCOW, where the caller holds the address space
5016 * locked as writer and calls softunlock before dropping the as lock.
5017 * S_READ_NOCOW is used by /proc to read memory from another user.
5018 *
5019 * Another deadlock between SOFTLOCK and file truncation can happen
5020 * because segvn_fault_vnodepages() calls the FS one pagesize at
5021 * a time. A second fop_getpage() call by segvn_fault_vnodepages()
5022 * can cause a deadlock because the first set of page_t's remain
5023 * locked SE_SHARED. To avoid this, we demote segments on a first
5024 * SOFTLOCK if they have a length greater than the segment's
5025 * page size.
5026 *
5027 * So for now, we only avoid demoting a segment on a SOFTLOCK when
5028 * the access type is S_READ_NOCOW and the fault length is less than
5029 * or equal to the segment's page size. While this is quite restrictive,
5030 * it should be the most common case of SOFTLOCK against a vmpss
5031 * segment.
5032 *
5033 * For S_READ_NOCOW, it's safe not to do a copy on write because the
5034 * caller makes sure no COW will be caused by another thread for a
5035 * softlocked page.
5036 */
5037 if (type == F_SOFTLOCK && svd->vp != NULL && seg->s_szc != 0) {
5038 int demote = 0;
5039
5040 if (rw != S_READ_NOCOW) {
5041 demote = 1;
5042 }
5043 if (!demote && len > PAGESIZE) {
5044 pgsz = page_get_pagesize(seg->s_szc);
5045 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr,
5046 lpgeaddr);
5047 if (lpgeaddr - lpgaddr > pgsz) {
5048 demote = 1;
5049 }
5050 }
5051
5052 ASSERT(demote || AS_WRITE_HELD(seg->s_as));
5053
5054 if (demote) {
5055 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5056 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5057 if (seg->s_szc != 0) {
5058 segvn_vmpss_clrszc_cnt++;
5059 ASSERT(svd->softlockcnt == 0);
5060 err = segvn_clrszc(seg);
5061 if (err) {
5062 segvn_vmpss_clrszc_err++;
5063 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5064 return (FC_MAKE_ERR(err));
5065 }
5066 }
5067 ASSERT(seg->s_szc == 0);
5068 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5069 goto top;
5070 }
5071 }
5072
5073 /*
5074 * Check to see if we need to allocate an anon_map structure.
5075 */
5076 if (svd->amp == NULL && (svd->vp == NULL || brkcow)) {
5077 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5078 /*
5079 * Drop the "read" lock on the segment and acquire
5080 * the "write" version since we have to allocate the
5081 * anon_map.
5082 */
5083 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5084 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5085
5086 if (svd->amp == NULL) {
5087 svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
5088 svd->amp->a_szc = seg->s_szc;
5089 }
5090 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5091
5092 /*
5093 * Start all over again since segment protections
5094 * may have changed after we dropped the "read" lock.
5095 */
5096 goto top;
5097 }
5098
5099 /*
5100 * S_READ_NOCOW vs S_READ distinction was
5101 * only needed for the code above. After
5102 * that we treat it as S_READ.
5103 */
5104 if (rw == S_READ_NOCOW) {
5105 ASSERT(type == F_SOFTLOCK);
5106 ASSERT(AS_WRITE_HELD(seg->s_as));
5107 rw = S_READ;
5108 }
5109
5110 amp = svd->amp;
5111
5112 /*
5113 * MADV_SEQUENTIAL work is ignored for large page segments.
5114 */
5115 if (seg->s_szc != 0) {
5116 pgsz = page_get_pagesize(seg->s_szc);
5117 ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
5118 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
5119 if (svd->vp == NULL) {
5120 err = segvn_fault_anonpages(hat, seg, lpgaddr,
5121 lpgeaddr, type, rw, addr, addr + len, brkcow);
5122 } else {
5123 err = segvn_fault_vnodepages(hat, seg, lpgaddr,
5124 lpgeaddr, type, rw, addr, addr + len, brkcow);
5125 if (err == IE_RETRY) {
5126 ASSERT(seg->s_szc == 0);
5127 ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
5128 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5129 goto top;
5130 }
5131 }
5132 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5133 return (err);
5134 }
5135
5136 page = seg_page(seg, addr);
5137 if (amp != NULL) {
5138 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5139 anon_index = svd->anon_index + page;
5140
5141 if (type == F_PROT && rw == S_READ &&
5142 svd->tr_state == SEGVN_TR_OFF &&
5143 svd->type == MAP_PRIVATE && svd->pageprot == 0) {
5144 size_t index = anon_index;
5145 struct anon *ap;
5146
5147 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5148 /*
5149 * The fast path could apply to S_WRITE also, except
5150 * that the protection fault could be caused by lazy
5151 * tlb flush when ro->rw. In this case, the pte is
5152 * RW already. But RO in the other cpu's tlb causes
5153 * the fault. Since hat_chgprot won't do anything if
5154 * pte doesn't change, we may end up faulting
5155 * indefinitely until the RO tlb entry gets replaced.
5156 */
5157 for (a = addr; a < addr + len; a += PAGESIZE, index++) {
5158 anon_array_enter(amp, index, &cookie);
5159 ap = anon_get_ptr(amp->ahp, index);
5160 anon_array_exit(&cookie);
5161 if ((ap == NULL) || (ap->an_refcnt != 1)) {
5162 ANON_LOCK_EXIT(&amp->a_rwlock);
5163 goto slow;
5164 }
5165 }
5166 hat_chgprot(seg->s_as->a_hat, addr, len, svd->prot);
5167 ANON_LOCK_EXIT(&amp->a_rwlock);
5168 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5169 return (0);
5170 }
5171 }
5172 slow:
5173
5174 if (svd->vpage == NULL)
5175 vpage = NULL;
5176 else
5177 vpage = &svd->vpage[page];
5178
5179 off = svd->offset + (uintptr_t)(addr - seg->s_base);
5180
5181 /*
5182 * If MADV_SEQUENTIAL has been set for the particular page we
5183 * are faulting on, free behind all pages in the segment and put
5184 * them on the free list.
5185 */
5186
5187 if ((page != 0) && fltadvice && svd->tr_state != SEGVN_TR_ON) {
5188 struct vpage *vpp;
5189 ulong_t fanon_index;
5190 size_t fpage;
5191 uoff_t pgoff, fpgoff;
5192 struct vnode *fvp;
5193 struct anon *fap = NULL;
5194
5195 if (svd->advice == MADV_SEQUENTIAL ||
5196 (svd->pageadvice &&
5197 VPP_ADVICE(vpage) == MADV_SEQUENTIAL)) {
5198 pgoff = off - PAGESIZE;
5199 fpage = page - 1;
5200 if (vpage != NULL)
5201 vpp = &svd->vpage[fpage];
5202 if (amp != NULL)
5203 fanon_index = svd->anon_index + fpage;
5204
5205 while (pgoff > svd->offset) {
5206 if (svd->advice != MADV_SEQUENTIAL &&
5207 (!svd->pageadvice || (vpage &&
5208 VPP_ADVICE(vpp) != MADV_SEQUENTIAL)))
5209 break;
5210
5211 /*
5212 * If this is an anon page, we must find the
5213 * correct <vp, offset> for it
5214 */
5215 fap = NULL;
5216 if (amp != NULL) {
5217 ANON_LOCK_ENTER(&amp->a_rwlock,
5218 RW_READER);
5219 anon_array_enter(amp, fanon_index,
5220 &cookie);
5221 fap = anon_get_ptr(amp->ahp,
5222 fanon_index);
5223 if (fap != NULL) {
5224 swap_xlate(fap, &fvp, &fpgoff);
5225 } else {
5226 fpgoff = pgoff;
5227 fvp = svd->vp;
5228 }
5229 anon_array_exit(&cookie);
5230 ANON_LOCK_EXIT(&amp->a_rwlock);
5231 } else {
5232 fpgoff = pgoff;
5233 fvp = svd->vp;
5234 }
5235 if (fvp == NULL)
5236 break; /* XXX */
5237 /*
5238 * Skip pages that are free or have an
5239 * "exclusive" lock.
5240 */
5241 pp = page_lookup_nowait(fvp, fpgoff, SE_SHARED);
5242 if (pp == NULL)
5243 break;
5244 /*
5245 * We don't need the page_struct_lock to test
5246 * as this is only advisory; even if we
5247 * acquire it someone might race in and lock
5248 * the page after we unlock and before the
5249 * PUTPAGE, then fop_putpage will do nothing.
5250 */
5251 if (pp->p_lckcnt == 0 && pp->p_cowcnt == 0) {
5252 /*
5253 * Hold the vnode before releasing
5254 * the page lock to prevent it from
5255 * being freed and re-used by some
5256 * other thread.
5257 */
5258 VN_HOLD(fvp);
5259 page_unlock(pp);
5260 /*
5261 * We should build a page list
5262 * to kluster putpages XXX
5263 */
5264 (void) fop_putpage(fvp,
5265 (offset_t)fpgoff, PAGESIZE,
5266 (B_DONTNEED|B_FREE|B_ASYNC),
5267 svd->cred, NULL);
5268 VN_RELE(fvp);
5269 } else {
5270 /*
5271 * XXX - Should the loop terminate if
5272 * the page is `locked'?
5273 */
5274 page_unlock(pp);
5275 }
5276 --vpp;
5277 --fanon_index;
5278 pgoff -= PAGESIZE;
5279 }
5280 }
5281 }
5282
5283 plp = pl;
5284 *plp = NULL;
5285 pl_alloc_sz = 0;
5286
5287 /*
5288 * See if we need to call fop_getpage for
5289 * *any* of the range being faulted on.
5290 * We can skip all of this work if there
5291 * was no original vnode.
5292 */
5293 if (svd->vp != NULL) {
5294 uoff_t vp_off;
5295 size_t vp_len;
5296 struct anon *ap;
5297 vnode_t *vp;
5298
5299 vp_off = off;
5300 vp_len = len;
5301
5302 if (amp == NULL)
5303 dogetpage = 1;
5304 else {
5305 /*
5306 * Only acquire reader lock to prevent amp->ahp
5307 * from being changed. It's ok to miss pages,
5308 * hence we don't do anon_array_enter
5309 */
5310 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5311 ap = anon_get_ptr(amp->ahp, anon_index);
5312
5313 if (len <= PAGESIZE)
5314 /* inline non_anon() */
5315 dogetpage = (ap == NULL);
5316 else
5317 dogetpage = non_anon(amp->ahp, anon_index,
5318 &vp_off, &vp_len);
5319 ANON_LOCK_EXIT(&amp->a_rwlock);
5320 }
5321
5322 if (dogetpage) {
5323 enum seg_rw arw;
5324 struct as *as = seg->s_as;
5325
5326 if (len > FAULT_TMP_PAGES_SZ) {
5327 /*
5328 * Page list won't fit in local array,
5329 * allocate one of the needed size.
5330 */
5331 pl_alloc_sz =
5332 (btop(len) + 1) * sizeof (page_t *);
5333 plp = kmem_alloc(pl_alloc_sz, KM_SLEEP);
5334 plp[0] = NULL;
5335 plsz = len;
5336 } else if (rw == S_WRITE && svd->type == MAP_PRIVATE ||
5337 svd->tr_state == SEGVN_TR_ON || rw == S_OTHER ||
5338 (((size_t)(addr + PAGESIZE) <
5339 (size_t)(seg->s_base + seg->s_size)) &&
5340 hat_probe(as->a_hat, addr + PAGESIZE))) {
5341 /*
5342 * Ask fop_getpage to return the exact number
5343 * of pages if
5344 * (a) this is a COW fault, or
5345 * (b) this is a software fault, or
5346 * (c) next page is already mapped.
5347 */
5348 plsz = len;
5349 } else {
5350 /*
5351 * Ask fop_getpage to return adjacent pages
5352 * within the segment.
5353 */
5354 plsz = MIN((size_t)FAULT_TMP_PAGES_SZ, (size_t)
5355 ((seg->s_base + seg->s_size) - addr));
5356 ASSERT((addr + plsz) <=
5357 (seg->s_base + seg->s_size));
5358 }
5359
5360 /*
5361 * Need to get some non-anonymous pages.
5362 * We need to make only one call to GETPAGE to do
5363 * this to prevent certain deadlocking conditions
5364 * when we are doing locking. In this case
5365 * non_anon() should have picked up the smallest
5366 * range which includes all the non-anonymous
5367 * pages in the requested range. We have to
5368 * be careful regarding which rw flag to pass in
5369 * because on a private mapping, the underlying
5370 * object is never allowed to be written.
5371 */
5372 if (rw == S_WRITE && svd->type == MAP_PRIVATE) {
5373 arw = S_READ;
5374 } else {
5375 arw = rw;
5376 }
5377 vp = svd->vp;
5378 err = fop_getpage(vp, (offset_t)vp_off, vp_len,
5379 &vpprot, plp, plsz, seg, addr + (vp_off - off), arw,
5380 svd->cred, NULL);
5381 if (err) {
5382 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5383 segvn_pagelist_rele(plp);
5384 if (pl_alloc_sz)
5385 kmem_free(plp, pl_alloc_sz);
5386 return (FC_MAKE_ERR(err));
5387 }
5388 if (svd->type == MAP_PRIVATE)
5389 vpprot &= ~PROT_WRITE;
5390 }
5391 }
5392
5393 /*
5394 * N.B. at this time the plp array has all the needed non-anon
5395 * pages in addition to (possibly) having some adjacent pages.
5396 */
5397
5398 /*
5399 * Always acquire the anon_array_lock to prevent
5400 * 2 threads from allocating separate anon slots for
5401 * the same "addr".
5402 *
5403 * If this is a copy-on-write fault and we don't already
5404 * have the anon_array_lock, acquire it to prevent the
5405 * fault routine from handling multiple copy-on-write faults
5406 * on the same "addr" in the same address space.
5407 *
5408 * Only one thread should deal with the fault since after
5409 * it is handled, the other threads can acquire a translation
5410 * to the newly created private page. This prevents two or
5411 * more threads from creating different private pages for the
5412 * same fault.
5413 *
5414 * We grab "serialization" lock here if this is a MAP_PRIVATE segment
5415 * to prevent deadlock between this thread and another thread
5416 * which has soft-locked this page and wants to acquire serial_lock.
5417 * ( bug 4026339 )
5418 *
5419 * The fix for bug 4026339 becomes unnecessary when using the
5420 * locking scheme with per amp rwlock and a global set of hash
5421 * lock, anon_array_lock. If we steal a vnode page when low
5422 * on memory and upgrad the page lock through page_rename,
5423 * then the page is PAGE_HANDLED, nothing needs to be done
5424 * for this page after returning from segvn_faultpage.
5425 *
5426 * But really, the page lock should be downgraded after
5427 * the stolen page is page_rename'd.
5428 */
5429
5430 if (amp != NULL)
5431 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5432
5433 /*
5434 * Ok, now loop over the address range and handle faults
5435 */
5436 for (a = addr; a < addr + len; a += PAGESIZE, off += PAGESIZE) {
5437 err = segvn_faultpage(hat, seg, a, off, vpage, plp, vpprot,
5438 type, rw, brkcow);
5439 if (err) {
5440 if (amp != NULL)
5441 ANON_LOCK_EXIT(&amp->a_rwlock);
5442 if (type == F_SOFTLOCK && a > addr) {
5443 segvn_softunlock(seg, addr, (a - addr),
5444 S_OTHER);
5445 }
5446 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5447 segvn_pagelist_rele(plp);
5448 if (pl_alloc_sz)
5449 kmem_free(plp, pl_alloc_sz);
5450 return (err);
5451 }
5452 if (vpage) {
5453 vpage++;
5454 } else if (svd->vpage) {
5455 page = seg_page(seg, addr);
5456 vpage = &svd->vpage[++page];
5457 }
5458 }
5459
5460 /* Didn't get pages from the underlying fs so we're done */
5461 if (!dogetpage)
5462 goto done;
5463
5464 /*
5465 * Now handle any other pages in the list returned.
5466 * If the page can be used, load up the translations now.
5467 * Note that the for loop will only be entered if "plp"
5468 * is pointing to a non-NULL page pointer which means that
5469 * fop_getpage() was called and vpprot has been initialized.
5470 */
5471 if (svd->pageprot == 0)
5472 prot = svd->prot & vpprot;
5473
5474
5475 /*
5476 * Large Files: diff should be unsigned value because we started
5477 * supporting > 2GB segment sizes from 2.5.1 and when a
5478 * large file of size > 2GB gets mapped to address space
5479 * the diff value can be > 2GB.
5480 */
5481
5482 for (ppp = plp; (pp = *ppp) != NULL; ppp++) {
5483 size_t diff;
5484 struct anon *ap;
5485 int anon_index;
5486 anon_sync_obj_t cookie;
5487 int hat_flag = HAT_LOAD_ADV;
5488
5489 if (svd->flags & MAP_TEXT) {
5490 hat_flag |= HAT_LOAD_TEXT;
5491 }
5492
5493 if (pp == PAGE_HANDLED)
5494 continue;
5495
5496 if (svd->tr_state != SEGVN_TR_ON &&
5497 pp->p_offset >= svd->offset &&
5498 pp->p_offset < svd->offset + seg->s_size) {
5499
5500 diff = pp->p_offset - svd->offset;
5501
5502 /*
5503 * Large Files: Following is the assertion
5504 * validating the above cast.
5505 */
5506 ASSERT(svd->vp == pp->p_vnode);
5507
5508 page = btop(diff);
5509 if (svd->pageprot)
5510 prot = VPP_PROT(&svd->vpage[page]) & vpprot;
5511
5512 /*
5513 * Prevent other threads in the address space from
5514 * creating private pages (i.e., allocating anon slots)
5515 * while we are in the process of loading translations
5516 * to additional pages returned by the underlying
5517 * object.
5518 */
5519 if (amp != NULL) {
5520 anon_index = svd->anon_index + page;
5521 anon_array_enter(amp, anon_index, &cookie);
5522 ap = anon_get_ptr(amp->ahp, anon_index);
5523 }
5524 if ((amp == NULL) || (ap == NULL)) {
5525 if (IS_VMODSORT(pp->p_vnode) ||
5526 enable_mbit_wa) {
5527 if (rw == S_WRITE)
5528 hat_setmod(pp);
5529 else if (rw != S_OTHER &&
5530 !hat_ismod(pp))
5531 prot &= ~PROT_WRITE;
5532 }
5533 /*
5534 * Skip mapping read ahead pages marked
5535 * for migration, so they will get migrated
5536 * properly on fault
5537 */
5538 ASSERT(amp == NULL ||
5539 svd->rcookie == HAT_INVALID_REGION_COOKIE);
5540 if ((prot & PROT_READ) && !PP_ISMIGRATE(pp)) {
5541 hat_memload_region(hat,
5542 seg->s_base + diff,
5543 pp, prot, hat_flag,
5544 svd->rcookie);
5545 }
5546 }
5547 if (amp != NULL)
5548 anon_array_exit(&cookie);
5549 }
5550 page_unlock(pp);
5551 }
5552 done:
5553 if (amp != NULL)
5554 ANON_LOCK_EXIT(&amp->a_rwlock);
5555 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5556 if (pl_alloc_sz)
5557 kmem_free(plp, pl_alloc_sz);
5558 return (0);
5559 }
5560
5561 /*
5562 * This routine is used to start I/O on pages asynchronously. XXX it will
5563 * only create PAGESIZE pages. At fault time they will be relocated into
5564 * larger pages.
5565 */
5566 static faultcode_t
5567 segvn_faulta(struct seg *seg, caddr_t addr)
5568 {
5569 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5570 int err;
5571 struct anon_map *amp;
5572 vnode_t *vp;
5573
5574 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
5575
5576 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5577 if ((amp = svd->amp) != NULL) {
5578 struct anon *ap;
5579
5580 /*
5581 * Reader lock to prevent amp->ahp from being changed.
5582 * This is advisory, it's ok to miss a page, so
5583 * we don't do anon_array_enter lock.
5584 */
5585 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5586 if ((ap = anon_get_ptr(amp->ahp,
5587 svd->anon_index + seg_page(seg, addr))) != NULL) {
5588
5589 err = anon_getpage(&ap, NULL, NULL,
5590 0, seg, addr, S_READ, svd->cred);
5591
5592 ANON_LOCK_EXIT(&amp->a_rwlock);
5593 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5594 if (err)
5595 return (FC_MAKE_ERR(err));
5596 return (0);
5597 }
5598 ANON_LOCK_EXIT(&amp->a_rwlock);
5599 }
5600
5601 if (svd->vp == NULL) {
5602 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5603 return (0); /* zfod page - do nothing now */
5604 }
5605
5606 vp = svd->vp;
5607 err = fop_getpage(vp,
5608 (offset_t)(svd->offset + (uintptr_t)(addr - seg->s_base)),
5609 PAGESIZE, NULL, NULL, 0, seg, addr,
5610 S_OTHER, svd->cred, NULL);
5611
5612 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5613 if (err)
5614 return (FC_MAKE_ERR(err));
5615 return (0);
5616 }
5617
5618 static int
5619 segvn_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
5620 {
5621 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5622 struct vpage *cvp, *svp, *evp;
5623 struct vnode *vp;
5624 size_t pgsz;
5625 pgcnt_t pgcnt;
5626 anon_sync_obj_t cookie;
5627 int unload_done = 0;
5628
5629 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
5630
5631 if ((svd->maxprot & prot) != prot)
5632 return (EACCES); /* violated maxprot */
5633
5634 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5635
5636 /* return if prot is the same */
5637 if (!svd->pageprot && svd->prot == prot) {
5638 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5639 return (0);
5640 }
5641
5642 /*
5643 * Since we change protections we first have to flush the cache.
5644 * This makes sure all the pagelock calls have to recheck
5645 * protections.
5646 */
5647 if (svd->softlockcnt > 0) {
5648 ASSERT(svd->tr_state == SEGVN_TR_OFF);
5649
5650 /*
5651 * If this is shared segment non 0 softlockcnt
5652 * means locked pages are still in use.
5653 */
5654 if (svd->type == MAP_SHARED) {
5655 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5656 return (EAGAIN);
5657 }
5658
5659 /*
5660 * Since we do have the segvn writers lock nobody can fill
5661 * the cache with entries belonging to this seg during
5662 * the purge. The flush either succeeds or we still have
5663 * pending I/Os.
5664 */
5665 segvn_purge(seg);
5666 if (svd->softlockcnt > 0) {
5667 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5668 return (EAGAIN);
5669 }
5670 }
5671
5672 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5673 ASSERT(svd->amp == NULL);
5674 ASSERT(svd->tr_state == SEGVN_TR_OFF);
5675 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5676 HAT_REGION_TEXT);
5677 svd->rcookie = HAT_INVALID_REGION_COOKIE;
5678 unload_done = 1;
5679 } else if (svd->tr_state == SEGVN_TR_INIT) {
5680 svd->tr_state = SEGVN_TR_OFF;
5681 } else if (svd->tr_state == SEGVN_TR_ON) {
5682 ASSERT(svd->amp != NULL);
5683 segvn_textunrepl(seg, 0);
5684 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5685 unload_done = 1;
5686 }
5687
5688 if ((prot & PROT_WRITE) && svd->type == MAP_SHARED &&
5689 svd->vp != NULL && (svd->vp->v_flag & VVMEXEC)) {
5690 ASSERT(vn_is_mapped(svd->vp, V_WRITE));
5691 segvn_inval_trcache(svd->vp);
5692 }
5693 if (seg->s_szc != 0) {
5694 int err;
5695 pgsz = page_get_pagesize(seg->s_szc);
5696 pgcnt = pgsz >> PAGESHIFT;
5697 ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
5698 if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
5699 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5700 ASSERT(seg->s_base != addr || seg->s_size != len);
5701 /*
5702 * If we are holding the as lock as a reader then
5703 * we need to return IE_RETRY and let the as
5704 * layer drop and re-acquire the lock as a writer.
5705 */
5706 if (AS_READ_HELD(seg->s_as))
5707 return (IE_RETRY);
5708 VM_STAT_ADD(segvnvmstats.demoterange[1]);
5709 if (svd->type == MAP_PRIVATE || svd->vp != NULL) {
5710 err = segvn_demote_range(seg, addr, len,
5711 SDR_END, 0);
5712 } else {
5713 uint_t szcvec = map_pgszcvec(seg->s_base,
5714 pgsz, (uintptr_t)seg->s_base,
5715 (svd->flags & MAP_TEXT), MAPPGSZC_SHM, 0);
5716 err = segvn_demote_range(seg, addr, len,
5717 SDR_END, szcvec);
5718 }
5719 if (err == 0)
5720 return (IE_RETRY);
5721 if (err == ENOMEM)
5722 return (IE_NOMEM);
5723 return (err);
5724 }
5725 }
5726
5727
5728 /*
5729 * If it's a private mapping and we're making it writable then we
5730 * may have to reserve the additional swap space now. If we are
5731 * making writable only a part of the segment then we use its vpage
5732 * array to keep a record of the pages for which we have reserved
5733 * swap. In this case we set the pageswap field in the segment's
5734 * segvn structure to record this.
5735 *
5736 * If it's a private mapping to a file (i.e., vp != NULL) and we're
5737 * removing write permission on the entire segment and we haven't
5738 * modified any pages, we can release the swap space.
5739 */
5740 if (svd->type == MAP_PRIVATE) {
5741 if (prot & PROT_WRITE) {
5742 if (!(svd->flags & MAP_NORESERVE) &&
5743 !(svd->swresv && svd->pageswap == 0)) {
5744 size_t sz = 0;
5745
5746 /*
5747 * Start by determining how much swap
5748 * space is required.
5749 */
5750 if (addr == seg->s_base &&
5751 len == seg->s_size &&
5752 svd->pageswap == 0) {
5753 /* The whole segment */
5754 sz = seg->s_size;
5755 } else {
5756 /*
5757 * Make sure that the vpage array
5758 * exists, and make a note of the
5759 * range of elements corresponding
5760 * to len.
5761 */
5762 segvn_vpage(seg);
5763 if (svd->vpage == NULL) {
5764 SEGVN_LOCK_EXIT(seg->s_as,
5765 &svd->lock);
5766 return (ENOMEM);
5767 }
5768 svp = &svd->vpage[seg_page(seg, addr)];
5769 evp = &svd->vpage[seg_page(seg,
5770 addr + len)];
5771
5772 if (svd->pageswap == 0) {
5773 /*
5774 * This is the first time we've
5775 * asked for a part of this
5776 * segment, so we need to
5777 * reserve everything we've
5778 * been asked for.
5779 */
5780 sz = len;
5781 } else {
5782 /*
5783 * We have to count the number
5784 * of pages required.
5785 */
5786 for (cvp = svp; cvp < evp;
5787 cvp++) {
5788 if (!VPP_ISSWAPRES(cvp))
5789 sz++;
5790 }
5791 sz <<= PAGESHIFT;
5792 }
5793 }
5794
5795 /* Try to reserve the necessary swap. */
5796 if (anon_resv_zone(sz,
5797 seg->s_as->a_proc->p_zone) == 0) {
5798 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5799 return (IE_NOMEM);
5800 }
5801
5802 /*
5803 * Make a note of how much swap space
5804 * we've reserved.
5805 */
5806 if (svd->pageswap == 0 && sz == seg->s_size) {
5807 svd->swresv = sz;
5808 } else {
5809 ASSERT(svd->vpage != NULL);
5810 svd->swresv += sz;
5811 svd->pageswap = 1;
5812 for (cvp = svp; cvp < evp; cvp++) {
5813 if (!VPP_ISSWAPRES(cvp))
5814 VPP_SETSWAPRES(cvp);
5815 }
5816 }
5817 }
5818 } else {
5819 /*
5820 * Swap space is released only if this segment
5821 * does not map anonymous memory, since read faults
5822 * on such segments still need an anon slot to read
5823 * in the data.
5824 */
5825 if (svd->swresv != 0 && svd->vp != NULL &&
5826 svd->amp == NULL && addr == seg->s_base &&
5827 len == seg->s_size && svd->pageprot == 0) {
5828 ASSERT(svd->pageswap == 0);
5829 anon_unresv_zone(svd->swresv,
5830 seg->s_as->a_proc->p_zone);
5831 svd->swresv = 0;
5832 }
5833 }
5834 }
5835
5836 if (addr == seg->s_base && len == seg->s_size && svd->vpage == NULL) {
5837 if (svd->prot == prot) {
5838 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5839 return (0); /* all done */
5840 }
5841 svd->prot = (uchar_t)prot;
5842 } else if (svd->type == MAP_PRIVATE) {
5843 struct anon *ap = NULL;
5844 page_t *pp;
5845 uoff_t offset, off;
5846 struct anon_map *amp;
5847 ulong_t anon_idx = 0;
5848
5849 /*
5850 * A vpage structure exists or else the change does not
5851 * involve the entire segment. Establish a vpage structure
5852 * if none is there. Then, for each page in the range,
5853 * adjust its individual permissions. Note that write-
5854 * enabling a MAP_PRIVATE page can affect the claims for
5855 * locked down memory. Overcommitting memory terminates
5856 * the operation.
5857 */
5858 segvn_vpage(seg);
5859 if (svd->vpage == NULL) {
5860 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5861 return (ENOMEM);
5862 }
5863 svd->pageprot = 1;
5864 if ((amp = svd->amp) != NULL) {
5865 anon_idx = svd->anon_index + seg_page(seg, addr);
5866 ASSERT(seg->s_szc == 0 ||
5867 IS_P2ALIGNED(anon_idx, pgcnt));
5868 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5869 }
5870
5871 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
5872 evp = &svd->vpage[seg_page(seg, addr + len)];
5873
5874 /*
5875 * See Statement at the beginning of segvn_lockop regarding
5876 * the way cowcnts and lckcnts are handled.
5877 */
5878 for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5879
5880 if (seg->s_szc != 0) {
5881 if (amp != NULL) {
5882 anon_array_enter(amp, anon_idx,
5883 &cookie);
5884 }
5885 if (IS_P2ALIGNED(anon_idx, pgcnt) &&
5886 !segvn_claim_pages(seg, svp, offset,
5887 anon_idx, prot)) {
5888 if (amp != NULL) {
5889 anon_array_exit(&cookie);
5890 }
5891 break;
5892 }
5893 if (amp != NULL) {
5894 anon_array_exit(&cookie);
5895 }
5896 anon_idx++;
5897 } else {
5898 if (amp != NULL) {
5899 anon_array_enter(amp, anon_idx,
5900 &cookie);
5901 ap = anon_get_ptr(amp->ahp, anon_idx++);
5902 }
5903
5904 if (VPP_ISPPLOCK(svp) &&
5905 VPP_PROT(svp) != prot) {
5906
5907 if (amp == NULL || ap == NULL) {
5908 vp = svd->vp;
5909 off = offset;
5910 } else
5911 swap_xlate(ap, &vp, &off);
5912 if (amp != NULL)
5913 anon_array_exit(&cookie);
5914
5915 if ((pp = page_lookup(vp, off,
5916 SE_SHARED)) == NULL) {
5917 panic("segvn_setprot: no page");
5918 /*NOTREACHED*/
5919 }
5920 ASSERT(seg->s_szc == 0);
5921 if ((VPP_PROT(svp) ^ prot) &
5922 PROT_WRITE) {
5923 if (prot & PROT_WRITE) {
5924 if (!page_addclaim(
5925 pp)) {
5926 page_unlock(pp);
5927 break;
5928 }
5929 } else {
5930 if (!page_subclaim(
5931 pp)) {
5932 page_unlock(pp);
5933 break;
5934 }
5935 }
5936 }
5937 page_unlock(pp);
5938 } else if (amp != NULL)
5939 anon_array_exit(&cookie);
5940 }
5941 VPP_SETPROT(svp, prot);
5942 offset += PAGESIZE;
5943 }
5944 if (amp != NULL)
5945 ANON_LOCK_EXIT(&amp->a_rwlock);
5946
5947 /*
5948 * Did we terminate prematurely? If so, simply unload
5949 * the translations to the things we've updated so far.
5950 */
5951 if (svp != evp) {
5952 if (unload_done) {
5953 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5954 return (IE_NOMEM);
5955 }
5956 len = (svp - &svd->vpage[seg_page(seg, addr)]) *
5957 PAGESIZE;
5958 ASSERT(seg->s_szc == 0 || IS_P2ALIGNED(len, pgsz));
5959 if (len != 0)
5960 hat_unload(seg->s_as->a_hat, addr,
5961 len, HAT_UNLOAD);
5962 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5963 return (IE_NOMEM);
5964 }
5965 } else {
5966 segvn_vpage(seg);
5967 if (svd->vpage == NULL) {
5968 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5969 return (ENOMEM);
5970 }
5971 svd->pageprot = 1;
5972 evp = &svd->vpage[seg_page(seg, addr + len)];
5973 for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5974 VPP_SETPROT(svp, prot);
5975 }
5976 }
5977
5978 if (unload_done) {
5979 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5980 return (0);
5981 }
5982
5983 if (((prot & PROT_WRITE) != 0 &&
5984 (svd->vp != NULL || svd->type == MAP_PRIVATE)) ||
5985 (prot & ~PROT_USER) == PROT_NONE) {
5986 /*
5987 * Either private or shared data with write access (in
5988 * which case we need to throw out all former translations
5989 * so that we get the right translations set up on fault
5990 * and we don't allow write access to any copy-on-write pages
5991 * that might be around or to prevent write access to pages
5992 * representing holes in a file), or we don't have permission
5993 * to access the memory at all (in which case we have to
5994 * unload any current translations that might exist).
5995 */
5996 hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
5997 } else {
5998 /*
5999 * A shared mapping or a private mapping in which write
6000 * protection is going to be denied - just change all the
6001 * protections over the range of addresses in question.
6002 * segvn does not support any other attributes other
6003 * than prot so we can use hat_chgattr.
6004 */
6005 hat_chgattr(seg->s_as->a_hat, addr, len, prot);
6006 }
6007
6008 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6009
6010 return (0);
6011 }
6012
6013 /*
6014 * segvn_setpagesize is called via segop_setpagesize from as_setpagesize,
6015 * to determine if the seg is capable of mapping the requested szc.
6016 */
6017 static int
6018 segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
6019 {
6020 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6021 struct segvn_data *nsvd;
6022 struct anon_map *amp = svd->amp;
6023 struct seg *nseg;
6024 caddr_t eaddr = addr + len, a;
6025 size_t pgsz = page_get_pagesize(szc);
6026 pgcnt_t pgcnt = page_get_pagecnt(szc);
6027 int err;
6028 uoff_t off = svd->offset + (uintptr_t)(addr - seg->s_base);
6029
6030 ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as));
6031 ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6032
6033 if (seg->s_szc == szc || segvn_lpg_disable != 0) {
6034 return (0);
6035 }
6036
6037 /*
6038 * addr should always be pgsz aligned but eaddr may be misaligned if
6039 * it's at the end of the segment.
6040 *
6041 * XXX we should assert this condition since as_setpagesize() logic
6042 * guarantees it.
6043 */
6044 if (!IS_P2ALIGNED(addr, pgsz) ||
6045 (!IS_P2ALIGNED(eaddr, pgsz) &&
6046 eaddr != seg->s_base + seg->s_size)) {
6047
6048 segvn_setpgsz_align_err++;
6049 return (EINVAL);
6050 }
6051
6052 if (amp != NULL && svd->type == MAP_SHARED) {
6053 ulong_t an_idx = svd->anon_index + seg_page(seg, addr);
6054 if (!IS_P2ALIGNED(an_idx, pgcnt)) {
6055
6056 segvn_setpgsz_anon_align_err++;
6057 return (EINVAL);
6058 }
6059 }
6060
6061 if ((svd->flags & MAP_NORESERVE) || seg->s_as == &kas ||
6062 szc > segvn_maxpgszc) {
6063 return (EINVAL);
6064 }
6065
6066 /* paranoid check */
6067 if (svd->vp != NULL &&
6068 (IS_SWAPFSVP(svd->vp) || VN_ISKAS(svd->vp))) {
6069 return (EINVAL);
6070 }
6071
6072 if (seg->s_szc == 0 && svd->vp != NULL &&
6073 map_addr_vacalign_check(addr, off)) {
6074 return (EINVAL);
6075 }
6076
6077 /*
6078 * Check that protections are the same within new page
6079 * size boundaries.
6080 */
6081 if (svd->pageprot) {
6082 for (a = addr; a < eaddr; a += pgsz) {
6083 if ((a + pgsz) > eaddr) {
6084 if (!sameprot(seg, a, eaddr - a)) {
6085 return (EINVAL);
6086 }
6087 } else {
6088 if (!sameprot(seg, a, pgsz)) {
6089 return (EINVAL);
6090 }
6091 }
6092 }
6093 }
6094
6095 /*
6096 * Since we are changing page size we first have to flush
6097 * the cache. This makes sure all the pagelock calls have
6098 * to recheck protections.
6099 */
6100 if (svd->softlockcnt > 0) {
6101 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6102
6103 /*
6104 * If this is shared segment non 0 softlockcnt
6105 * means locked pages are still in use.
6106 */
6107 if (svd->type == MAP_SHARED) {
6108 return (EAGAIN);
6109 }
6110
6111 /*
6112 * Since we do have the segvn writers lock nobody can fill
6113 * the cache with entries belonging to this seg during
6114 * the purge. The flush either succeeds or we still have
6115 * pending I/Os.
6116 */
6117 segvn_purge(seg);
6118 if (svd->softlockcnt > 0) {
6119 return (EAGAIN);
6120 }
6121 }
6122
6123 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6124 ASSERT(svd->amp == NULL);
6125 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6126 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6127 HAT_REGION_TEXT);
6128 svd->rcookie = HAT_INVALID_REGION_COOKIE;
6129 } else if (svd->tr_state == SEGVN_TR_INIT) {
6130 svd->tr_state = SEGVN_TR_OFF;
6131 } else if (svd->tr_state == SEGVN_TR_ON) {
6132 ASSERT(svd->amp != NULL);
6133 segvn_textunrepl(seg, 1);
6134 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6135 amp = NULL;
6136 }
6137
6138 /*
6139 * Operation for sub range of existing segment.
6140 */
6141 if (addr != seg->s_base || eaddr != (seg->s_base + seg->s_size)) {
6142 if (szc < seg->s_szc) {
6143 VM_STAT_ADD(segvnvmstats.demoterange[2]);
6144 err = segvn_demote_range(seg, addr, len, SDR_RANGE, 0);
6145 if (err == 0) {
6146 return (IE_RETRY);
6147 }
6148 if (err == ENOMEM) {
6149 return (IE_NOMEM);
6150 }
6151 return (err);
6152 }
6153 if (addr != seg->s_base) {
6154 nseg = segvn_split_seg(seg, addr);
6155 if (eaddr != (nseg->s_base + nseg->s_size)) {
6156 /* eaddr is szc aligned */
6157 (void) segvn_split_seg(nseg, eaddr);
6158 }
6159 return (IE_RETRY);
6160 }
6161 if (eaddr != (seg->s_base + seg->s_size)) {
6162 /* eaddr is szc aligned */
6163 (void) segvn_split_seg(seg, eaddr);
6164 }
6165 return (IE_RETRY);
6166 }
6167
6168 /*
6169 * Break any low level sharing and reset seg->s_szc to 0.
6170 */
6171 if ((err = segvn_clrszc(seg)) != 0) {
6172 if (err == ENOMEM) {
6173 err = IE_NOMEM;
6174 }
6175 return (err);
6176 }
6177 ASSERT(seg->s_szc == 0);
6178
6179 /*
6180 * If the end of the current segment is not pgsz aligned
6181 * then attempt to concatenate with the next segment.
6182 */
6183 if (!IS_P2ALIGNED(eaddr, pgsz)) {
6184 nseg = AS_SEGNEXT(seg->s_as, seg);
6185 if (nseg == NULL || nseg == seg || eaddr != nseg->s_base) {
6186 return (ENOMEM);
6187 }
6188 if (nseg->s_ops != &segvn_ops) {
6189 return (EINVAL);
6190 }
6191 nsvd = (struct segvn_data *)nseg->s_data;
6192 if (nsvd->softlockcnt > 0) {
6193 /*
6194 * If this is shared segment non 0 softlockcnt
6195 * means locked pages are still in use.
6196 */
6197 if (nsvd->type == MAP_SHARED) {
6198 return (EAGAIN);
6199 }
6200 segvn_purge(nseg);
6201 if (nsvd->softlockcnt > 0) {
6202 return (EAGAIN);
6203 }
6204 }
6205 err = segvn_clrszc(nseg);
6206 if (err == ENOMEM) {
6207 err = IE_NOMEM;
6208 }
6209 if (err != 0) {
6210 return (err);
6211 }
6212 ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6213 err = segvn_concat(seg, nseg, 1);
6214 if (err == -1) {
6215 return (EINVAL);
6216 }
6217 if (err == -2) {
6218 return (IE_NOMEM);
6219 }
6220 return (IE_RETRY);
6221 }
6222
6223 /*
6224 * May need to re-align anon array to
6225 * new szc.
6226 */
6227 if (amp != NULL) {
6228 if (!IS_P2ALIGNED(svd->anon_index, pgcnt)) {
6229 struct anon_hdr *nahp;
6230
6231 ASSERT(svd->type == MAP_PRIVATE);
6232
6233 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6234 ASSERT(amp->refcnt == 1);
6235 nahp = anon_create(btop(amp->size), ANON_NOSLEEP);
6236 if (nahp == NULL) {
6237 ANON_LOCK_EXIT(&amp->a_rwlock);
6238 return (IE_NOMEM);
6239 }
6240 if (anon_copy_ptr(amp->ahp, svd->anon_index,
6241 nahp, 0, btop(seg->s_size), ANON_NOSLEEP)) {
6242 anon_release(nahp, btop(amp->size));
6243 ANON_LOCK_EXIT(&amp->a_rwlock);
6244 return (IE_NOMEM);
6245 }
6246 anon_release(amp->ahp, btop(amp->size));
6247 amp->ahp = nahp;
6248 svd->anon_index = 0;
6249 ANON_LOCK_EXIT(&amp->a_rwlock);
6250 }
6251 }
6252 if (svd->vp != NULL && szc != 0) {
6253 struct vattr va;
6254 uoff_t eoffpage = svd->offset;
6255 va.va_mask = AT_SIZE;
6256 eoffpage += seg->s_size;
6257 eoffpage = btopr(eoffpage);
6258 if (fop_getattr(svd->vp, &va, 0, svd->cred, NULL) != 0) {
6259 segvn_setpgsz_getattr_err++;
6260 return (EINVAL);
6261 }
6262 if (btopr(va.va_size) < eoffpage) {
6263 segvn_setpgsz_eof_err++;
6264 return (EINVAL);
6265 }
6266 if (amp != NULL) {
6267 /*
6268 * anon_fill_cow_holes() may call fop_getpage().
6269 * don't take anon map lock here to avoid holding it
6270 * across fop_getpage() calls that may call back into
6271 * segvn for klsutering checks. We don't really need
6272 * anon map lock here since it's a private segment and
6273 * we hold as level lock as writers.
6274 */
6275 if ((err = anon_fill_cow_holes(seg, seg->s_base,
6276 amp->ahp, svd->anon_index, svd->vp, svd->offset,
6277 seg->s_size, szc, svd->prot, svd->vpage,
6278 svd->cred)) != 0) {
6279 return (EINVAL);
6280 }
6281 }
6282 segvn_setvnode_mpss(svd->vp);
6283 }
6284
6285 if (amp != NULL) {
6286 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6287 if (svd->type == MAP_PRIVATE) {
6288 amp->a_szc = szc;
6289 } else if (szc > amp->a_szc) {
6290 amp->a_szc = szc;
6291 }
6292 ANON_LOCK_EXIT(&amp->a_rwlock);
6293 }
6294
6295 seg->s_szc = szc;
6296
6297 return (0);
6298 }
6299
6300 static int
6301 segvn_clrszc(struct seg *seg)
6302 {
6303 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6304 struct anon_map *amp = svd->amp;
6305 size_t pgsz;
6306 pgcnt_t pages;
6307 int err = 0;
6308 caddr_t a = seg->s_base;
6309 caddr_t ea = a + seg->s_size;
6310 ulong_t an_idx = svd->anon_index;
6311 vnode_t *vp = svd->vp;
6312 struct vpage *vpage = svd->vpage;
6313 page_t *anon_pl[1 + 1], *pp;
6314 struct anon *ap, *oldap;
6315 uint_t prot = svd->prot, vpprot;
6316 int pageflag = 0;
6317
6318 ASSERT(AS_WRITE_HELD(seg->s_as) ||
6319 SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
6320 ASSERT(svd->softlockcnt == 0);
6321
6322 if (vp == NULL && amp == NULL) {
6323 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6324 seg->s_szc = 0;
6325 return (0);
6326 }
6327
6328 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6329 ASSERT(svd->amp == NULL);
6330 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6331 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6332 HAT_REGION_TEXT);
6333 svd->rcookie = HAT_INVALID_REGION_COOKIE;
6334 } else if (svd->tr_state == SEGVN_TR_ON) {
6335 ASSERT(svd->amp != NULL);
6336 segvn_textunrepl(seg, 1);
6337 ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6338 amp = NULL;
6339 } else {
6340 if (svd->tr_state != SEGVN_TR_OFF) {
6341 ASSERT(svd->tr_state == SEGVN_TR_INIT);
6342 svd->tr_state = SEGVN_TR_OFF;
6343 }
6344
6345 /*
6346 * do HAT_UNLOAD_UNMAP since we are changing the pagesize.
6347 * unload argument is 0 when we are freeing the segment
6348 * and unload was already done.
6349 */
6350 hat_unload(seg->s_as->a_hat, seg->s_base, seg->s_size,
6351 HAT_UNLOAD_UNMAP);
6352 }
6353
6354 if (amp == NULL || svd->type == MAP_SHARED) {
6355 seg->s_szc = 0;
6356 return (0);
6357 }
6358
6359 pgsz = page_get_pagesize(seg->s_szc);
6360 pages = btop(pgsz);
6361
6362 /*
6363 * XXX anon rwlock is not really needed because this is a
6364 * private segment and we are writers.
6365 */
6366 ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6367
6368 for (; a < ea; a += pgsz, an_idx += pages) {
6369 if ((oldap = anon_get_ptr(amp->ahp, an_idx)) != NULL) {
6370 ASSERT(vpage != NULL || svd->pageprot == 0);
6371 if (vpage != NULL) {
6372 ASSERT(sameprot(seg, a, pgsz));
6373 prot = VPP_PROT(vpage);
6374 pageflag = VPP_ISPPLOCK(vpage) ? LOCK_PAGE : 0;
6375 }
6376 if (seg->s_szc != 0) {
6377 ASSERT(vp == NULL || anon_pages(amp->ahp,
6378 an_idx, pages) == pages);
6379 if ((err = anon_map_demotepages(amp, an_idx,
6380 seg, a, prot, vpage, svd->cred)) != 0) {
6381 goto out;
6382 }
6383 } else {
6384 if (oldap->an_refcnt == 1) {
6385 continue;
6386 }
6387 if ((err = anon_getpage(&oldap, &vpprot,
6388 anon_pl, PAGESIZE, seg, a, S_READ,
6389 svd->cred))) {
6390 goto out;
6391 }
6392 if ((pp = anon_private(&ap, seg, a, prot,
6393 anon_pl[0], pageflag, svd->cred)) == NULL) {
6394 err = ENOMEM;
6395 goto out;
6396 }
6397 anon_decref(oldap);
6398 (void) anon_set_ptr(amp->ahp, an_idx, ap,
6399 ANON_SLEEP);
6400 page_unlock(pp);
6401 }
6402 }
6403 vpage = (vpage == NULL) ? NULL : vpage + pages;
6404 }
6405
6406 amp->a_szc = 0;
6407 seg->s_szc = 0;
6408 out:
6409 ANON_LOCK_EXIT(&amp->a_rwlock);
6410 return (err);
6411 }
6412
6413 static int
6414 segvn_claim_pages(
6415 struct seg *seg,
6416 struct vpage *svp,
6417 uoff_t off,
6418 ulong_t anon_idx,
6419 uint_t prot)
6420 {
6421 pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6422 size_t ppasize = (pgcnt + 1) * sizeof (page_t *);
6423 page_t **ppa;
6424 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6425 struct anon_map *amp = svd->amp;
6426 struct vpage *evp = svp + pgcnt;
6427 caddr_t addr = ((uintptr_t)(svp - svd->vpage) << PAGESHIFT)
6428 + seg->s_base;
6429 struct anon *ap;
6430 struct vnode *vp = svd->vp;
6431 page_t *pp;
6432 pgcnt_t pg_idx, i;
6433 int err = 0;
6434 anoff_t aoff;
6435 int anon = (amp != NULL) ? 1 : 0;
6436
6437 ASSERT(svd->type == MAP_PRIVATE);
6438 ASSERT(svd->vpage != NULL);
6439 ASSERT(seg->s_szc != 0);
6440 ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
6441 ASSERT(amp == NULL || IS_P2ALIGNED(anon_idx, pgcnt));
6442 ASSERT(sameprot(seg, addr, pgcnt << PAGESHIFT));
6443
6444 if (VPP_PROT(svp) == prot)
6445 return (1);
6446 if (!((VPP_PROT(svp) ^ prot) & PROT_WRITE))
6447 return (1);
6448
6449 ppa = kmem_alloc(ppasize, KM_SLEEP);
6450 if (anon && vp != NULL) {
6451 if (anon_get_ptr(amp->ahp, anon_idx) == NULL) {
6452 anon = 0;
6453 ASSERT(!anon_pages(amp->ahp, anon_idx, pgcnt));
6454 }
6455 ASSERT(!anon ||
6456 anon_pages(amp->ahp, anon_idx, pgcnt) == pgcnt);
6457 }
6458
6459 for (*ppa = NULL, pg_idx = 0; svp < evp; svp++, anon_idx++) {
6460 if (!VPP_ISPPLOCK(svp))
6461 continue;
6462 if (anon) {
6463 ap = anon_get_ptr(amp->ahp, anon_idx);
6464 if (ap == NULL) {
6465 panic("segvn_claim_pages: no anon slot");
6466 }
6467 swap_xlate(ap, &vp, &aoff);
6468 off = (uoff_t)aoff;
6469 }
6470 ASSERT(vp != NULL);
6471 if ((pp = page_lookup(vp,
6472 (uoff_t)off, SE_SHARED)) == NULL) {
6473 panic("segvn_claim_pages: no page");
6474 }
6475 ppa[pg_idx++] = pp;
6476 off += PAGESIZE;
6477 }
6478
6479 if (ppa[0] == NULL) {
6480 kmem_free(ppa, ppasize);
6481 return (1);
6482 }
6483
6484 ASSERT(pg_idx <= pgcnt);
6485 ppa[pg_idx] = NULL;
6486
6487
6488 /* Find each large page within ppa, and adjust its claim */
6489
6490 /* Does ppa cover a single large page? */
6491 if (ppa[0]->p_szc == seg->s_szc) {
6492 if (prot & PROT_WRITE)
6493 err = page_addclaim_pages(ppa);
6494 else
6495 err = page_subclaim_pages(ppa);
6496 } else {
6497 for (i = 0; ppa[i]; i += pgcnt) {
6498 ASSERT(IS_P2ALIGNED(page_pptonum(ppa[i]), pgcnt));
6499 if (prot & PROT_WRITE)
6500 err = page_addclaim_pages(&ppa[i]);
6501 else
6502 err = page_subclaim_pages(&ppa[i]);
6503 if (err == 0)
6504 break;
6505 }
6506 }
6507
6508 for (i = 0; i < pg_idx; i++) {
6509 ASSERT(ppa[i] != NULL);
6510 page_unlock(ppa[i]);
6511 }
6512
6513 kmem_free(ppa, ppasize);
6514 return (err);
6515 }
6516
6517 /*
6518 * Returns right (upper address) segment if split occurred.
6519 * If the address is equal to the beginning or end of its segment it returns
6520 * the current segment.
6521 */
6522 static struct seg *
6523 segvn_split_seg(struct seg *seg, caddr_t addr)
6524 {
6525 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6526 struct seg *nseg;
6527 size_t nsize;
6528 struct segvn_data *nsvd;
6529
6530 ASSERT(AS_WRITE_HELD(seg->s_as));
6531 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6532
6533 ASSERT(addr >= seg->s_base);
6534 ASSERT(addr <= seg->s_base + seg->s_size);
6535 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6536
6537 if (addr == seg->s_base || addr == seg->s_base + seg->s_size)
6538 return (seg);
6539
6540 nsize = seg->s_base + seg->s_size - addr;
6541 seg->s_size = addr - seg->s_base;
6542 nseg = seg_alloc(seg->s_as, addr, nsize);
6543 ASSERT(nseg != NULL);
6544 nseg->s_ops = seg->s_ops;
6545 nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
6546 nseg->s_data = (void *)nsvd;
6547 nseg->s_szc = seg->s_szc;
6548 *nsvd = *svd;
6549 ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6550 nsvd->seg = nseg;
6551 rw_init(&nsvd->lock, NULL, RW_DEFAULT, NULL);
6552
6553 if (nsvd->vp != NULL) {
6554 VN_HOLD(nsvd->vp);
6555 nsvd->offset = svd->offset +
6556 (uintptr_t)(nseg->s_base - seg->s_base);
6557 if (nsvd->type == MAP_SHARED)
6558 lgrp_shm_policy_init(NULL, nsvd->vp);
6559 } else {
6560 /*
6561 * The offset for an anonymous segment has no signifigance in
6562 * terms of an offset into a file. If we were to use the above
6563 * calculation instead, the structures read out of
6564 * /proc/<pid>/xmap would be more difficult to decipher since
6565 * it would be unclear whether two seemingly contiguous
6566 * prxmap_t structures represented different segments or a
6567 * single segment that had been split up into multiple prxmap_t
6568 * structures (e.g. if some part of the segment had not yet
6569 * been faulted in).
6570 */
6571 nsvd->offset = 0;
6572 }
6573
6574 ASSERT(svd->softlockcnt == 0);
6575 ASSERT(svd->softlockcnt_sbase == 0);
6576 ASSERT(svd->softlockcnt_send == 0);
6577 crhold(svd->cred);
6578
6579 if (svd->vpage != NULL) {
6580 size_t bytes = vpgtob(seg_pages(seg));
6581 size_t nbytes = vpgtob(seg_pages(nseg));
6582 struct vpage *ovpage = svd->vpage;
6583
6584 svd->vpage = kmem_alloc(bytes, KM_SLEEP);
6585 bcopy(ovpage, svd->vpage, bytes);
6586 nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
6587 bcopy(ovpage + seg_pages(seg), nsvd->vpage, nbytes);
6588 kmem_free(ovpage, bytes + nbytes);
6589 }
6590 if (svd->amp != NULL && svd->type == MAP_PRIVATE) {
6591 struct anon_map *oamp = svd->amp, *namp;
6592 struct anon_hdr *nahp;
6593
6594 ANON_LOCK_ENTER(&oamp->a_rwlock, RW_WRITER);
6595 ASSERT(oamp->refcnt == 1);
6596 nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
6597 (void) anon_copy_ptr(oamp->ahp, svd->anon_index,
6598 nahp, 0, btop(seg->s_size), ANON_SLEEP);
6599
6600 namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
6601 namp->a_szc = nseg->s_szc;
6602 (void) anon_copy_ptr(oamp->ahp,
6603 svd->anon_index + btop(seg->s_size),
6604 namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
6605 anon_release(oamp->ahp, btop(oamp->size));
6606 oamp->ahp = nahp;
6607 oamp->size = seg->s_size;
6608 svd->anon_index = 0;
6609 nsvd->amp = namp;
6610 nsvd->anon_index = 0;
6611 ANON_LOCK_EXIT(&oamp->a_rwlock);
6612 } else if (svd->amp != NULL) {
6613 pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6614 ASSERT(svd->amp == nsvd->amp);
6615 ASSERT(seg->s_szc <= svd->amp->a_szc);
6616 nsvd->anon_index = svd->anon_index + seg_pages(seg);
6617 ASSERT(IS_P2ALIGNED(nsvd->anon_index, pgcnt));
6618 ANON_LOCK_ENTER(&svd->amp->a_rwlock, RW_WRITER);
6619 svd->amp->refcnt++;
6620 ANON_LOCK_EXIT(&svd->amp->a_rwlock);
6621 }
6622
6623 /*
6624 * Split the amount of swap reserved.
6625 */
6626 if (svd->swresv) {
6627 /*
6628 * For MAP_NORESERVE, only allocate swap reserve for pages
6629 * being used. Other segments get enough to cover whole
6630 * segment.
6631 */
6632 if (svd->flags & MAP_NORESERVE) {
6633 size_t oswresv;
6634
6635 ASSERT(svd->amp);
6636 oswresv = svd->swresv;
6637 svd->swresv = ptob(anon_pages(svd->amp->ahp,
6638 svd->anon_index, btop(seg->s_size)));
6639 nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
6640 nsvd->anon_index, btop(nseg->s_size)));
6641 ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
6642 } else {
6643 if (svd->pageswap) {
6644 svd->swresv = segvn_count_swap_by_vpages(seg);
6645 ASSERT(nsvd->swresv >= svd->swresv);
6646 nsvd->swresv -= svd->swresv;
6647 } else {
6648 ASSERT(svd->swresv == seg->s_size +
6649 nseg->s_size);
6650 svd->swresv = seg->s_size;
6651 nsvd->swresv = nseg->s_size;
6652 }
6653 }
6654 }
6655
6656 return (nseg);
6657 }
6658
6659 /*
6660 * called on memory operations (unmap, setprot, setpagesize) for a subset
6661 * of a large page segment to either demote the memory range (SDR_RANGE)
6662 * or the ends (SDR_END) by addr/len.
6663 *
6664 * returns 0 on success. returns errno, including ENOMEM, on failure.
6665 */
6666 static int
6667 segvn_demote_range(
6668 struct seg *seg,
6669 caddr_t addr,
6670 size_t len,
6671 int flag,
6672 uint_t szcvec)
6673 {
6674 caddr_t eaddr = addr + len;
6675 caddr_t lpgaddr, lpgeaddr;
6676 struct seg *nseg;
6677 struct seg *badseg1 = NULL;
6678 struct seg *badseg2 = NULL;
6679 size_t pgsz;
6680 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6681 int err;
6682 uint_t szc = seg->s_szc;
6683 uint_t tszcvec;
6684
6685 ASSERT(AS_WRITE_HELD(seg->s_as));
6686 ASSERT(svd->tr_state == SEGVN_TR_OFF);
6687 ASSERT(szc != 0);
6688 pgsz = page_get_pagesize(szc);
6689 ASSERT(seg->s_base != addr || seg->s_size != len);
6690 ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6691 ASSERT(svd->softlockcnt == 0);
6692 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6693 ASSERT(szcvec == 0 || (flag == SDR_END && svd->type == MAP_SHARED));
6694
6695 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
6696 ASSERT(flag == SDR_RANGE || eaddr < lpgeaddr || addr > lpgaddr);
6697 if (flag == SDR_RANGE) {
6698 /* demote entire range */
6699 badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6700 (void) segvn_split_seg(nseg, lpgeaddr);
6701 ASSERT(badseg1->s_base == lpgaddr);
6702 ASSERT(badseg1->s_size == lpgeaddr - lpgaddr);
6703 } else if (addr != lpgaddr) {
6704 ASSERT(flag == SDR_END);
6705 badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6706 if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz &&
6707 eaddr < lpgaddr + 2 * pgsz) {
6708 (void) segvn_split_seg(nseg, lpgeaddr);
6709 ASSERT(badseg1->s_base == lpgaddr);
6710 ASSERT(badseg1->s_size == 2 * pgsz);
6711 } else {
6712 nseg = segvn_split_seg(nseg, lpgaddr + pgsz);
6713 ASSERT(badseg1->s_base == lpgaddr);
6714 ASSERT(badseg1->s_size == pgsz);
6715 if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz) {
6716 ASSERT(lpgeaddr - lpgaddr > 2 * pgsz);
6717 nseg = segvn_split_seg(nseg, lpgeaddr - pgsz);
6718 badseg2 = nseg;
6719 (void) segvn_split_seg(nseg, lpgeaddr);
6720 ASSERT(badseg2->s_base == lpgeaddr - pgsz);
6721 ASSERT(badseg2->s_size == pgsz);
6722 }
6723 }
6724 } else {
6725 ASSERT(flag == SDR_END);
6726 ASSERT(eaddr < lpgeaddr);
6727 badseg1 = nseg = segvn_split_seg(seg, lpgeaddr - pgsz);
6728 (void) segvn_split_seg(nseg, lpgeaddr);
6729 ASSERT(badseg1->s_base == lpgeaddr - pgsz);
6730 ASSERT(badseg1->s_size == pgsz);
6731 }
6732
6733 ASSERT(badseg1 != NULL);
6734 ASSERT(badseg1->s_szc == szc);
6735 ASSERT(flag == SDR_RANGE || badseg1->s_size == pgsz ||
6736 badseg1->s_size == 2 * pgsz);
6737 ASSERT(sameprot(badseg1, badseg1->s_base, pgsz));
6738 ASSERT(badseg1->s_size == pgsz ||
6739 sameprot(badseg1, badseg1->s_base + pgsz, pgsz));
6740 if (err = segvn_clrszc(badseg1)) {
6741 return (err);
6742 }
6743 ASSERT(badseg1->s_szc == 0);
6744
6745 if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6746 uint_t tszc = highbit(tszcvec) - 1;
6747 caddr_t ta = MAX(addr, badseg1->s_base);
6748 caddr_t te;
6749 size_t tpgsz = page_get_pagesize(tszc);
6750
6751 ASSERT(svd->type == MAP_SHARED);
6752 ASSERT(flag == SDR_END);
6753 ASSERT(tszc < szc && tszc > 0);
6754
6755 if (eaddr > badseg1->s_base + badseg1->s_size) {
6756 te = badseg1->s_base + badseg1->s_size;
6757 } else {
6758 te = eaddr;
6759 }
6760
6761 ASSERT(ta <= te);
6762 badseg1->s_szc = tszc;
6763 if (!IS_P2ALIGNED(ta, tpgsz) || !IS_P2ALIGNED(te, tpgsz)) {
6764 if (badseg2 != NULL) {
6765 err = segvn_demote_range(badseg1, ta, te - ta,
6766 SDR_END, tszcvec);
6767 if (err != 0) {
6768 return (err);
6769 }
6770 } else {
6771 return (segvn_demote_range(badseg1, ta,
6772 te - ta, SDR_END, tszcvec));
6773 }
6774 }
6775 }
6776
6777 if (badseg2 == NULL)
6778 return (0);
6779 ASSERT(badseg2->s_szc == szc);
6780 ASSERT(badseg2->s_size == pgsz);
6781 ASSERT(sameprot(badseg2, badseg2->s_base, badseg2->s_size));
6782 if (err = segvn_clrszc(badseg2)) {
6783 return (err);
6784 }
6785 ASSERT(badseg2->s_szc == 0);
6786
6787 if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6788 uint_t tszc = highbit(tszcvec) - 1;
6789 size_t tpgsz = page_get_pagesize(tszc);
6790
6791 ASSERT(svd->type == MAP_SHARED);
6792 ASSERT(flag == SDR_END);
6793 ASSERT(tszc < szc && tszc > 0);
6794 ASSERT(badseg2->s_base > addr);
6795 ASSERT(eaddr > badseg2->s_base);
6796 ASSERT(eaddr < badseg2->s_base + badseg2->s_size);
6797
6798 badseg2->s_szc = tszc;
6799 if (!IS_P2ALIGNED(eaddr, tpgsz)) {
6800 return (segvn_demote_range(badseg2, badseg2->s_base,
6801 eaddr - badseg2->s_base, SDR_END, tszcvec));
6802 }
6803 }
6804
6805 return (0);
6806 }
6807
6808 static int
6809 segvn_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
6810 {
6811 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6812 struct vpage *vp, *evp;
6813
6814 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6815
6816 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6817 /*
6818 * If segment protection can be used, simply check against them.
6819 */
6820 if (svd->pageprot == 0) {
6821 int err;
6822
6823 err = ((svd->prot & prot) != prot) ? EACCES : 0;
6824 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6825 return (err);
6826 }
6827
6828 /*
6829 * Have to check down to the vpage level.
6830 */
6831 evp = &svd->vpage[seg_page(seg, addr + len)];
6832 for (vp = &svd->vpage[seg_page(seg, addr)]; vp < evp; vp++) {
6833 if ((VPP_PROT(vp) & prot) != prot) {
6834 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6835 return (EACCES);
6836 }
6837 }
6838 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6839 return (0);
6840 }
6841
6842 static int
6843 segvn_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
6844 {
6845 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6846 size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
6847
6848 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6849
6850 if (pgno != 0) {
6851 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6852 if (svd->pageprot == 0) {
6853 do {
6854 protv[--pgno] = svd->prot;
6855 } while (pgno != 0);
6856 } else {
6857 size_t pgoff = seg_page(seg, addr);
6858
6859 do {
6860 pgno--;
6861 protv[pgno] = VPP_PROT(&svd->vpage[pgno+pgoff]);
6862 } while (pgno != 0);
6863 }
6864 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6865 }
6866 return (0);
6867 }
6868
6869 static uoff_t
6870 segvn_getoffset(struct seg *seg, caddr_t addr)
6871 {
6872 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6873
6874 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6875
6876 return (svd->offset + (uintptr_t)(addr - seg->s_base));
6877 }
6878
6879 /*ARGSUSED*/
6880 static int
6881 segvn_gettype(struct seg *seg, caddr_t addr)
6882 {
6883 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6884
6885 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6886
6887 return (svd->type | (svd->flags & (MAP_NORESERVE | MAP_TEXT |
6888 MAP_INITDATA)));
6889 }
6890
6891 /*ARGSUSED*/
6892 static int
6893 segvn_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
6894 {
6895 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6896
6897 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6898
6899 *vpp = svd->vp;
6900 return (0);
6901 }
6902
6903 /*
6904 * Check to see if it makes sense to do kluster/read ahead to
6905 * addr + delta relative to the mapping at addr. We assume here
6906 * that delta is a signed PAGESIZE'd multiple (which can be negative).
6907 *
6908 * For segvn, we currently "approve" of the action if we are
6909 * still in the segment and it maps from the same vp/off,
6910 * or if the advice stored in segvn_data or vpages allows it.
6911 * Currently, klustering is not allowed only if MADV_RANDOM is set.
6912 */
6913 static int
6914 segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
6915 {
6916 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6917 struct anon *oap, *ap;
6918 ssize_t pd;
6919 size_t page;
6920 struct vnode *vp1, *vp2;
6921 uoff_t off1, off2;
6922 struct anon_map *amp;
6923
6924 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
6925 ASSERT(AS_WRITE_HELD(seg->s_as) ||
6926 SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
6927
6928 if (addr + delta < seg->s_base ||
6929 addr + delta >= (seg->s_base + seg->s_size))
6930 return (-1); /* exceeded segment bounds */
6931
6932 pd = delta / (ssize_t)PAGESIZE; /* divide to preserve sign bit */
6933 page = seg_page(seg, addr);
6934
6935 /*
6936 * Check to see if either of the pages addr or addr + delta
6937 * have advice set that prevents klustering (if MADV_RANDOM advice
6938 * is set for entire segment, or MADV_SEQUENTIAL is set and delta
6939 * is negative).
6940 */
6941 if (svd->advice == MADV_RANDOM ||
6942 svd->advice == MADV_SEQUENTIAL && delta < 0)
6943 return (-1);
6944 else if (svd->pageadvice && svd->vpage) {
6945 struct vpage *bvpp, *evpp;
6946
6947 bvpp = &svd->vpage[page];
6948 evpp = &svd->vpage[page + pd];
6949 if (VPP_ADVICE(bvpp) == MADV_RANDOM ||
6950 VPP_ADVICE(evpp) == MADV_SEQUENTIAL && delta < 0)
6951 return (-1);
6952 if (VPP_ADVICE(bvpp) != VPP_ADVICE(evpp) &&
6953 VPP_ADVICE(evpp) == MADV_RANDOM)
6954 return (-1);
6955 }
6956
6957 if (svd->type == MAP_SHARED)
6958 return (0); /* shared mapping - all ok */
6959
6960 if ((amp = svd->amp) == NULL)
6961 return (0); /* off original vnode */
6962
6963 page += svd->anon_index;
6964
6965 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
6966
6967 oap = anon_get_ptr(amp->ahp, page);
6968 ap = anon_get_ptr(amp->ahp, page + pd);
6969
6970 ANON_LOCK_EXIT(&amp->a_rwlock);
6971
6972 if ((oap == NULL && ap != NULL) || (oap != NULL && ap == NULL)) {
6973 return (-1); /* one with and one without an anon */
6974 }
6975
6976 if (oap == NULL) { /* implies that ap == NULL */
6977 return (0); /* off original vnode */
6978 }
6979
6980 /*
6981 * Now we know we have two anon pointers - check to
6982 * see if they happen to be properly allocated.
6983 */
6984
6985 /*
6986 * XXX We cheat here and don't lock the anon slots. We can't because
6987 * we may have been called from the anon layer which might already
6988 * have locked them. We are holding a refcnt on the slots so they
6989 * can't disappear. The worst that will happen is we'll get the wrong
6990 * names (vp, off) for the slots and make a poor klustering decision.
6991 */
6992 swap_xlate(ap, &vp1, &off1);
6993 swap_xlate(oap, &vp2, &off2);
6994
6995
6996 if (!fop_cmp(vp1, vp2, NULL) || off1 - off2 != delta)
6997 return (-1);
6998 return (0);
6999 }
7000
7001 /*
7002 * Synchronize primary storage cache with real object in virtual memory.
7003 *
7004 * XXX - Anonymous pages should not be sync'ed out at all.
7005 */
7006 static int
7007 segvn_sync(struct seg *seg, caddr_t addr, size_t len, int attr, uint_t flags)
7008 {
7009 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7010 struct vpage *vpp;
7011 page_t *pp;
7012 uoff_t offset;
7013 struct vnode *vp;
7014 uoff_t off;
7015 caddr_t eaddr;
7016 int bflags;
7017 int err = 0;
7018 int segtype;
7019 int pageprot;
7020 int prot;
7021 ulong_t anon_index;
7022 struct anon_map *amp;
7023 struct anon *ap;
7024 anon_sync_obj_t cookie;
7025
7026 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7027
7028 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7029
7030 if (svd->softlockcnt > 0) {
7031 /*
7032 * If this is shared segment non 0 softlockcnt
7033 * means locked pages are still in use.
7034 */
7035 if (svd->type == MAP_SHARED) {
7036 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7037 return (EAGAIN);
7038 }
7039
7040 /*
7041 * flush all pages from seg cache
7042 * otherwise we may deadlock in swap_putpage
7043 * for B_INVAL page (4175402).
7044 *
7045 * Even if we grab segvn WRITER's lock
7046 * here, there might be another thread which could've
7047 * successfully performed lookup/insert just before
7048 * we acquired the lock here. So, grabbing either
7049 * lock here is of not much use. Until we devise
7050 * a strategy at upper layers to solve the
7051 * synchronization issues completely, we expect
7052 * applications to handle this appropriately.
7053 */
7054 segvn_purge(seg);
7055 if (svd->softlockcnt > 0) {
7056 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7057 return (EAGAIN);
7058 }
7059 } else if (svd->type == MAP_SHARED && svd->amp != NULL &&
7060 svd->amp->a_softlockcnt > 0) {
7061 /*
7062 * Try to purge this amp's entries from pcache. It will
7063 * succeed only if other segments that share the amp have no
7064 * outstanding softlock's.
7065 */
7066 segvn_purge(seg);
7067 if (svd->amp->a_softlockcnt > 0 || svd->softlockcnt > 0) {
7068 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7069 return (EAGAIN);
7070 }
7071 }
7072
7073 vpp = svd->vpage;
7074 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7075 bflags = ((flags & MS_ASYNC) ? B_ASYNC : 0) |
7076 ((flags & MS_INVALIDATE) ? B_INVAL : 0);
7077
7078 if (attr) {
7079 pageprot = attr & ~(SHARED|PRIVATE);
7080 segtype = (attr & SHARED) ? MAP_SHARED : MAP_PRIVATE;
7081
7082 /*
7083 * We are done if the segment types don't match
7084 * or if we have segment level protections and
7085 * they don't match.
7086 */
7087 if (svd->type != segtype) {
7088 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7089 return (0);
7090 }
7091 if (vpp == NULL) {
7092 if (svd->prot != pageprot) {
7093 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7094 return (0);
7095 }
7096 prot = svd->prot;
7097 } else
7098 vpp = &svd->vpage[seg_page(seg, addr)];
7099
7100 } else if (svd->vp && svd->amp == NULL &&
7101 (flags & MS_INVALIDATE) == 0) {
7102
7103 /*
7104 * No attributes, no anonymous pages and MS_INVALIDATE flag
7105 * is not on, just use one big request.
7106 */
7107 err = fop_putpage(svd->vp, (offset_t)offset, len,
7108 bflags, svd->cred, NULL);
7109 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7110 return (err);
7111 }
7112
7113 if ((amp = svd->amp) != NULL)
7114 anon_index = svd->anon_index + seg_page(seg, addr);
7115
7116 for (eaddr = addr + len; addr < eaddr; addr += PAGESIZE) {
7117 ap = NULL;
7118 if (amp != NULL) {
7119 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7120 anon_array_enter(amp, anon_index, &cookie);
7121 ap = anon_get_ptr(amp->ahp, anon_index++);
7122 if (ap != NULL) {
7123 swap_xlate(ap, &vp, &off);
7124 } else {
7125 vp = svd->vp;
7126 off = offset;
7127 }
7128 anon_array_exit(&cookie);
7129 ANON_LOCK_EXIT(&amp->a_rwlock);
7130 } else {
7131 vp = svd->vp;
7132 off = offset;
7133 }
7134 offset += PAGESIZE;
7135
7136 if (vp == NULL) /* untouched zfod page */
7137 continue;
7138
7139 if (attr) {
7140 if (vpp) {
7141 prot = VPP_PROT(vpp);
7142 vpp++;
7143 }
7144 if (prot != pageprot) {
7145 continue;
7146 }
7147 }
7148
7149 /*
7150 * See if any of these pages are locked -- if so, then we
7151 * will have to truncate an invalidate request at the first
7152 * locked one. We don't need the page_struct_lock to test
7153 * as this is only advisory; even if we acquire it someone
7154 * might race in and lock the page after we unlock and before
7155 * we do the PUTPAGE, then PUTPAGE simply does nothing.
7156 */
7157 if (flags & MS_INVALIDATE) {
7158 if ((pp = page_lookup(vp, off, SE_SHARED)) != NULL) {
7159 if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
7160 page_unlock(pp);
7161 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7162 return (EBUSY);
7163 }
7164 if (ap != NULL && pp->p_szc != 0 &&
7165 page_tryupgrade(pp)) {
7166 if (pp->p_lckcnt == 0 &&
7167 pp->p_cowcnt == 0) {
7168 /*
7169 * swapfs VN_DISPOSE() won't
7170 * invalidate large pages.
7171 * Attempt to demote.
7172 * XXX can't help it if it
7173 * fails. But for swapfs
7174 * pages it is no big deal.
7175 */
7176 (void) page_try_demote_pages(
7177 pp);
7178 }
7179 }
7180 page_unlock(pp);
7181 }
7182 } else if (svd->type == MAP_SHARED && amp != NULL) {
7183 /*
7184 * Avoid writing out to disk ISM's large pages
7185 * because segspt_free_pages() relies on NULL an_pvp
7186 * of anon slots of such pages.
7187 */
7188
7189 ASSERT(svd->vp == NULL);
7190 /*
7191 * swapfs uses page_lookup_nowait if not freeing or
7192 * invalidating and skips a page if
7193 * page_lookup_nowait returns NULL.
7194 */
7195 pp = page_lookup_nowait(vp, off, SE_SHARED);
7196 if (pp == NULL) {
7197 continue;
7198 }
7199 if (pp->p_szc != 0) {
7200 page_unlock(pp);
7201 continue;
7202 }
7203
7204 /*
7205 * Note ISM pages are created large so (vp, off)'s
7206 * page cannot suddenly become large after we unlock
7207 * pp.
7208 */
7209 page_unlock(pp);
7210 }
7211 /*
7212 * XXX - Should ultimately try to kluster
7213 * calls to fop_putpage() for performance.
7214 */
7215 VN_HOLD(vp);
7216 err = fop_putpage(vp, (offset_t)off, PAGESIZE,
7217 (bflags | (IS_SWAPFSVP(vp) ? B_PAGE_NOWAIT : 0)),
7218 svd->cred, NULL);
7219
7220 VN_RELE(vp);
7221 if (err)
7222 break;
7223 }
7224 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7225 return (err);
7226 }
7227
7228 /*
7229 * Determine if we have data corresponding to pages in the
7230 * primary storage virtual memory cache (i.e., "in core").
7231 */
7232 static size_t
7233 segvn_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
7234 {
7235 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7236 struct vnode *vp, *avp;
7237 uoff_t offset, aoffset;
7238 size_t p, ep;
7239 int ret;
7240 struct vpage *vpp;
7241 page_t *pp;
7242 uint_t start;
7243 struct anon_map *amp; /* XXX - for locknest */
7244 struct anon *ap;
7245 uint_t attr;
7246 anon_sync_obj_t cookie;
7247
7248 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7249
7250 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7251 if (svd->amp == NULL && svd->vp == NULL) {
7252 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7253 bzero(vec, btopr(len));
7254 return (len); /* no anonymous pages created yet */
7255 }
7256
7257 p = seg_page(seg, addr);
7258 ep = seg_page(seg, addr + len);
7259 start = svd->vp ? SEG_PAGE_VNODEBACKED : 0;
7260
7261 amp = svd->amp;
7262 for (; p < ep; p++, addr += PAGESIZE) {
7263 vpp = (svd->vpage) ? &svd->vpage[p]: NULL;
7264 ret = start;
7265 ap = NULL;
7266 avp = NULL;
7267 /* Grab the vnode/offset for the anon slot */
7268 if (amp != NULL) {
7269 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7270 anon_array_enter(amp, svd->anon_index + p, &cookie);
7271 ap = anon_get_ptr(amp->ahp, svd->anon_index + p);
7272 if (ap != NULL) {
7273 swap_xlate(ap, &avp, &aoffset);
7274 }
7275 anon_array_exit(&cookie);
7276 ANON_LOCK_EXIT(&amp->a_rwlock);
7277 }
7278 if ((avp != NULL) && page_exists(avp, aoffset)) {
7279 /* A page exists for the anon slot */
7280 ret |= SEG_PAGE_INCORE;
7281
7282 /*
7283 * If page is mapped and writable
7284 */
7285 attr = (uint_t)0;
7286 if ((hat_getattr(seg->s_as->a_hat, addr,
7287 &attr) != -1) && (attr & PROT_WRITE)) {
7288 ret |= SEG_PAGE_ANON;
7289 }
7290 /*
7291 * Don't get page_struct lock for lckcnt and cowcnt,
7292 * since this is purely advisory.
7293 */
7294 if ((pp = page_lookup_nowait(avp, aoffset,
7295 SE_SHARED)) != NULL) {
7296 if (pp->p_lckcnt)
7297 ret |= SEG_PAGE_SOFTLOCK;
7298 if (pp->p_cowcnt)
7299 ret |= SEG_PAGE_HASCOW;
7300 page_unlock(pp);
7301 }
7302 }
7303
7304 /* Gather vnode statistics */
7305 vp = svd->vp;
7306 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7307
7308 if (vp != NULL) {
7309 /*
7310 * Try to obtain a "shared" lock on the page
7311 * without blocking. If this fails, determine
7312 * if the page is in memory.
7313 */
7314 pp = page_lookup_nowait(vp, offset, SE_SHARED);
7315 if ((pp == NULL) && (page_exists(vp, offset))) {
7316 /* Page is incore, and is named */
7317 ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7318 }
7319 /*
7320 * Don't get page_struct lock for lckcnt and cowcnt,
7321 * since this is purely advisory.
7322 */
7323 if (pp != NULL) {
7324 ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7325 if (pp->p_lckcnt)
7326 ret |= SEG_PAGE_SOFTLOCK;
7327 if (pp->p_cowcnt)
7328 ret |= SEG_PAGE_HASCOW;
7329 page_unlock(pp);
7330 }
7331 }
7332
7333 /* Gather virtual page information */
7334 if (vpp) {
7335 if (VPP_ISPPLOCK(vpp))
7336 ret |= SEG_PAGE_LOCKED;
7337 vpp++;
7338 }
7339
7340 *vec++ = (char)ret;
7341 }
7342 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7343 return (len);
7344 }
7345
7346 /*
7347 * Statement for p_cowcnts/p_lckcnts.
7348 *
7349 * p_cowcnt is updated while mlock/munlocking MAP_PRIVATE and PROT_WRITE region
7350 * irrespective of the following factors or anything else:
7351 *
7352 * (1) anon slots are populated or not
7353 * (2) cow is broken or not
7354 * (3) refcnt on ap is 1 or greater than 1
7355 *
7356 * If it's not MAP_PRIVATE and PROT_WRITE, p_lckcnt is updated during mlock
7357 * and munlock.
7358 *
7359 *
7360 * Handling p_cowcnts/p_lckcnts during copy-on-write fault:
7361 *
7362 * if vpage has PROT_WRITE
7363 * transfer cowcnt on the oldpage -> cowcnt on the newpage
7364 * else
7365 * transfer lckcnt on the oldpage -> lckcnt on the newpage
7366 *
7367 * During copy-on-write, decrement p_cowcnt on the oldpage and increment
7368 * p_cowcnt on the newpage *if* the corresponding vpage has PROT_WRITE.
7369 *
7370 * We may also break COW if softlocking on read access in the physio case.
7371 * In this case, vpage may not have PROT_WRITE. So, we need to decrement
7372 * p_lckcnt on the oldpage and increment p_lckcnt on the newpage *if* the
7373 * vpage doesn't have PROT_WRITE.
7374 *
7375 *
7376 * Handling p_cowcnts/p_lckcnts during mprotect on mlocked region:
7377 *
7378 * If a MAP_PRIVATE region loses PROT_WRITE, we decrement p_cowcnt and
7379 * increment p_lckcnt by calling page_subclaim() which takes care of
7380 * availrmem accounting and p_lckcnt overflow.
7381 *
7382 * If a MAP_PRIVATE region gains PROT_WRITE, we decrement p_lckcnt and
7383 * increment p_cowcnt by calling page_addclaim() which takes care of
7384 * availrmem availability and p_cowcnt overflow.
7385 */
7386
7387 /*
7388 * Lock down (or unlock) pages mapped by this segment.
7389 *
7390 * XXX only creates PAGESIZE pages if anon slots are not initialized.
7391 * At fault time they will be relocated into larger pages.
7392 */
7393 static int
7394 segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
7395 int attr, int op, ulong_t *lockmap, size_t pos)
7396 {
7397 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7398 struct vpage *vpp;
7399 struct vpage *evp;
7400 page_t *pp;
7401 uoff_t offset;
7402 uoff_t off;
7403 int segtype;
7404 int pageprot;
7405 int claim;
7406 struct vnode *vp;
7407 ulong_t anon_index;
7408 struct anon_map *amp;
7409 struct anon *ap;
7410 struct vattr va;
7411 anon_sync_obj_t cookie;
7412 struct kshmid *sp = NULL;
7413 struct proc *p = curproc;
7414 kproject_t *proj = NULL;
7415 int chargeproc = 1;
7416 size_t locked_bytes = 0;
7417 size_t unlocked_bytes = 0;
7418 int err = 0;
7419
7420 /*
7421 * Hold write lock on address space because may split or concatenate
7422 * segments
7423 */
7424 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7425
7426 /*
7427 * If this is a shm, use shm's project and zone, else use
7428 * project and zone of calling process
7429 */
7430
7431 /* Determine if this segment backs a sysV shm */
7432 if (svd->amp != NULL && svd->amp->a_sp != NULL) {
7433 ASSERT(svd->type == MAP_SHARED);
7434 ASSERT(svd->tr_state == SEGVN_TR_OFF);
7435 sp = svd->amp->a_sp;
7436 proj = sp->shm_perm.ipc_proj;
7437 chargeproc = 0;
7438 }
7439
7440 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7441 if (attr) {
7442 pageprot = attr & ~(SHARED|PRIVATE);
7443 segtype = attr & SHARED ? MAP_SHARED : MAP_PRIVATE;
7444
7445 /*
7446 * We are done if the segment types don't match
7447 * or if we have segment level protections and
7448 * they don't match.
7449 */
7450 if (svd->type != segtype) {
7451 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7452 return (0);
7453 }
7454 if (svd->pageprot == 0 && svd->prot != pageprot) {
7455 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7456 return (0);
7457 }
7458 }
7459
7460 if (op == MC_LOCK) {
7461 if (svd->tr_state == SEGVN_TR_INIT) {
7462 svd->tr_state = SEGVN_TR_OFF;
7463 } else if (svd->tr_state == SEGVN_TR_ON) {
7464 ASSERT(svd->amp != NULL);
7465 segvn_textunrepl(seg, 0);
7466 ASSERT(svd->amp == NULL &&
7467 svd->tr_state == SEGVN_TR_OFF);
7468 }
7469 }
7470
7471 /*
7472 * If we're locking, then we must create a vpage structure if
7473 * none exists. If we're unlocking, then check to see if there
7474 * is a vpage -- if not, then we could not have locked anything.
7475 */
7476
7477 if ((vpp = svd->vpage) == NULL) {
7478 if (op == MC_LOCK) {
7479 segvn_vpage(seg);
7480 if (svd->vpage == NULL) {
7481 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7482 return (ENOMEM);
7483 }
7484 } else {
7485 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7486 return (0);
7487 }
7488 }
7489
7490 /*
7491 * The anonymous data vector (i.e., previously
7492 * unreferenced mapping to swap space) can be allocated
7493 * by lazily testing for its existence.
7494 */
7495 if (op == MC_LOCK && svd->amp == NULL && svd->vp == NULL) {
7496 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
7497 svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
7498 svd->amp->a_szc = seg->s_szc;
7499 }
7500
7501 if ((amp = svd->amp) != NULL) {
7502 anon_index = svd->anon_index + seg_page(seg, addr);
7503 }
7504
7505 offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7506 evp = &svd->vpage[seg_page(seg, addr + len)];
7507
7508 if (sp != NULL)
7509 mutex_enter(&sp->shm_mlock);
7510
7511 /* determine number of unlocked bytes in range for lock operation */
7512 if (op == MC_LOCK) {
7513
7514 if (sp == NULL) {
7515 for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7516 vpp++) {
7517 if (!VPP_ISPPLOCK(vpp))
7518 unlocked_bytes += PAGESIZE;
7519 }
7520 } else {
7521 ulong_t i_idx, i_edx;
7522 anon_sync_obj_t i_cookie;
7523 struct anon *i_ap;
7524 struct vnode *i_vp;
7525 uoff_t i_off;
7526
7527 /* Only count sysV pages once for locked memory */
7528 i_edx = svd->anon_index + seg_page(seg, addr + len);
7529 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7530 for (i_idx = anon_index; i_idx < i_edx; i_idx++) {
7531 anon_array_enter(amp, i_idx, &i_cookie);
7532 i_ap = anon_get_ptr(amp->ahp, i_idx);
7533 if (i_ap == NULL) {
7534 unlocked_bytes += PAGESIZE;
7535 anon_array_exit(&i_cookie);
7536 continue;
7537 }
7538 swap_xlate(i_ap, &i_vp, &i_off);
7539 anon_array_exit(&i_cookie);
7540 pp = page_lookup(i_vp, i_off, SE_SHARED);
7541 if (pp == NULL) {
7542 unlocked_bytes += PAGESIZE;
7543 continue;
7544 } else if (pp->p_lckcnt == 0)
7545 unlocked_bytes += PAGESIZE;
7546 page_unlock(pp);
7547 }
7548 ANON_LOCK_EXIT(&amp->a_rwlock);
7549 }
7550
7551 mutex_enter(&p->p_lock);
7552 err = rctl_incr_locked_mem(p, proj, unlocked_bytes,
7553 chargeproc);
7554 mutex_exit(&p->p_lock);
7555
7556 if (err) {
7557 if (sp != NULL)
7558 mutex_exit(&sp->shm_mlock);
7559 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7560 return (err);
7561 }
7562 }
7563 /*
7564 * Loop over all pages in the range. Process if we're locking and
7565 * page has not already been locked in this mapping; or if we're
7566 * unlocking and the page has been locked.
7567 */
7568 for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7569 vpp++, pos++, addr += PAGESIZE, offset += PAGESIZE, anon_index++) {
7570 if ((attr == 0 || VPP_PROT(vpp) == pageprot) &&
7571 ((op == MC_LOCK && !VPP_ISPPLOCK(vpp)) ||
7572 (op == MC_UNLOCK && VPP_ISPPLOCK(vpp)))) {
7573
7574 if (amp != NULL)
7575 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7576 /*
7577 * If this isn't a MAP_NORESERVE segment and
7578 * we're locking, allocate anon slots if they
7579 * don't exist. The page is brought in later on.
7580 */
7581 if (op == MC_LOCK && svd->vp == NULL &&
7582 ((svd->flags & MAP_NORESERVE) == 0) &&
7583 amp != NULL &&
7584 ((ap = anon_get_ptr(amp->ahp, anon_index))
7585 == NULL)) {
7586 anon_array_enter(amp, anon_index, &cookie);
7587
7588 if ((ap = anon_get_ptr(amp->ahp,
7589 anon_index)) == NULL) {
7590 pp = anon_zero(seg, addr, &ap,
7591 svd->cred);
7592 if (pp == NULL) {
7593 anon_array_exit(&cookie);
7594 ANON_LOCK_EXIT(&amp->a_rwlock);
7595 err = ENOMEM;
7596 goto out;
7597 }
7598 ASSERT(anon_get_ptr(amp->ahp,
7599 anon_index) == NULL);
7600 (void) anon_set_ptr(amp->ahp,
7601 anon_index, ap, ANON_SLEEP);
7602 page_unlock(pp);
7603 }
7604 anon_array_exit(&cookie);
7605 }
7606
7607 /*
7608 * Get name for page, accounting for
7609 * existence of private copy.
7610 */
7611 ap = NULL;
7612 if (amp != NULL) {
7613 anon_array_enter(amp, anon_index, &cookie);
7614 ap = anon_get_ptr(amp->ahp, anon_index);
7615 if (ap != NULL) {
7616 swap_xlate(ap, &vp, &off);
7617 } else {
7618 if (svd->vp == NULL &&
7619 (svd->flags & MAP_NORESERVE)) {
7620 anon_array_exit(&cookie);
7621 ANON_LOCK_EXIT(&amp->a_rwlock);
7622 continue;
7623 }
7624 vp = svd->vp;
7625 off = offset;
7626 }
7627 if (op != MC_LOCK || ap == NULL) {
7628 anon_array_exit(&cookie);
7629 ANON_LOCK_EXIT(&amp->a_rwlock);
7630 }
7631 } else {
7632 vp = svd->vp;
7633 off = offset;
7634 }
7635
7636 /*
7637 * Get page frame. It's ok if the page is
7638 * not available when we're unlocking, as this
7639 * may simply mean that a page we locked got
7640 * truncated out of existence after we locked it.
7641 *
7642 * Invoke fop_getpage() to obtain the page struct
7643 * since we may need to read it from disk if its
7644 * been paged out.
7645 */
7646 if (op != MC_LOCK)
7647 pp = page_lookup(vp, off, SE_SHARED);
7648 else {
7649 page_t *pl[1 + 1];
7650 int error;
7651
7652 ASSERT(vp != NULL);
7653
7654 error = fop_getpage(vp, (offset_t)off, PAGESIZE,
7655 (uint_t *)NULL, pl, PAGESIZE, seg, addr,
7656 S_OTHER, svd->cred, NULL);
7657
7658 if (error && ap != NULL) {
7659 anon_array_exit(&cookie);
7660 ANON_LOCK_EXIT(&amp->a_rwlock);
7661 }
7662
7663 /*
7664 * If the error is EDEADLK then we must bounce
7665 * up and drop all vm subsystem locks and then
7666 * retry the operation later
7667 * This behavior is a temporary measure because
7668 * ufs/sds logging is badly designed and will
7669 * deadlock if we don't allow this bounce to
7670 * happen. The real solution is to re-design
7671 * the logging code to work properly. See bug
7672 * 4125102 for details of the problem.
7673 */
7674 if (error == EDEADLK) {
7675 err = error;
7676 goto out;
7677 }
7678 /*
7679 * Quit if we fail to fault in the page. Treat
7680 * the failure as an error, unless the addr
7681 * is mapped beyond the end of a file.
7682 */
7683 if (error && svd->vp) {
7684 va.va_mask = AT_SIZE;
7685 if (fop_getattr(svd->vp, &va, 0,
7686 svd->cred, NULL) != 0) {
7687 err = EIO;
7688 goto out;
7689 }
7690 if (btopr(va.va_size) >=
7691 btopr(off + 1)) {
7692 err = EIO;
7693 goto out;
7694 }
7695 goto out;
7696
7697 } else if (error) {
7698 err = EIO;
7699 goto out;
7700 }
7701 pp = pl[0];
7702 ASSERT(pp != NULL);
7703 }
7704
7705 /*
7706 * See Statement at the beginning of this routine.
7707 *
7708 * claim is always set if MAP_PRIVATE and PROT_WRITE
7709 * irrespective of following factors:
7710 *
7711 * (1) anon slots are populated or not
7712 * (2) cow is broken or not
7713 * (3) refcnt on ap is 1 or greater than 1
7714 *
7715 * See 4140683 for details
7716 */
7717 claim = ((VPP_PROT(vpp) & PROT_WRITE) &&
7718 (svd->type == MAP_PRIVATE));
7719
7720 /*
7721 * Perform page-level operation appropriate to
7722 * operation. If locking, undo the SOFTLOCK
7723 * performed to bring the page into memory
7724 * after setting the lock. If unlocking,
7725 * and no page was found, account for the claim
7726 * separately.
7727 */
7728 if (op == MC_LOCK) {
7729 int ret = 1; /* Assume success */
7730
7731 ASSERT(!VPP_ISPPLOCK(vpp));
7732
7733 ret = page_pp_lock(pp, claim, 0);
7734 if (ap != NULL) {
7735 if (ap->an_pvp != NULL) {
7736 anon_swap_free(ap, pp);
7737 }
7738 anon_array_exit(&cookie);
7739 ANON_LOCK_EXIT(&amp->a_rwlock);
7740 }
7741 if (ret == 0) {
7742 /* locking page failed */
7743 page_unlock(pp);
7744 err = EAGAIN;
7745 goto out;
7746 }
7747 VPP_SETPPLOCK(vpp);
7748 if (sp != NULL) {
7749 if (pp->p_lckcnt == 1)
7750 locked_bytes += PAGESIZE;
7751 } else
7752 locked_bytes += PAGESIZE;
7753
7754 if (lockmap != NULL)
7755 BT_SET(lockmap, pos);
7756
7757 page_unlock(pp);
7758 } else {
7759 ASSERT(VPP_ISPPLOCK(vpp));
7760 if (pp != NULL) {
7761 /* sysV pages should be locked */
7762 ASSERT(sp == NULL || pp->p_lckcnt > 0);
7763 page_pp_unlock(pp, claim, 0);
7764 if (sp != NULL) {
7765 if (pp->p_lckcnt == 0)
7766 unlocked_bytes
7767 += PAGESIZE;
7768 } else
7769 unlocked_bytes += PAGESIZE;
7770 page_unlock(pp);
7771 } else {
7772 ASSERT(sp == NULL);
7773 unlocked_bytes += PAGESIZE;
7774 }
7775 VPP_CLRPPLOCK(vpp);
7776 }
7777 }
7778 }
7779 out:
7780 if (op == MC_LOCK) {
7781 /* Credit back bytes that did not get locked */
7782 if ((unlocked_bytes - locked_bytes) > 0) {
7783 if (proj == NULL)
7784 mutex_enter(&p->p_lock);
7785 rctl_decr_locked_mem(p, proj,
7786 (unlocked_bytes - locked_bytes), chargeproc);
7787 if (proj == NULL)
7788 mutex_exit(&p->p_lock);
7789 }
7790
7791 } else {
7792 /* Account bytes that were unlocked */
7793 if (unlocked_bytes > 0) {
7794 if (proj == NULL)
7795 mutex_enter(&p->p_lock);
7796 rctl_decr_locked_mem(p, proj, unlocked_bytes,
7797 chargeproc);
7798 if (proj == NULL)
7799 mutex_exit(&p->p_lock);
7800 }
7801 }
7802 if (sp != NULL)
7803 mutex_exit(&sp->shm_mlock);
7804 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7805
7806 return (err);
7807 }
7808
7809 /*
7810 * Set advice from user for specified pages
7811 * There are 10 types of advice:
7812 * MADV_NORMAL - Normal (default) behavior (whatever that is)
7813 * MADV_RANDOM - Random page references
7814 * do not allow readahead or 'klustering'
7815 * MADV_SEQUENTIAL - Sequential page references
7816 * Pages previous to the one currently being
7817 * accessed (determined by fault) are 'not needed'
7818 * and are freed immediately
7819 * MADV_WILLNEED - Pages are likely to be used (fault ahead in mctl)
7820 * MADV_DONTNEED - Pages are not needed (synced out in mctl)
7821 * MADV_FREE - Contents can be discarded
7822 * MADV_ACCESS_DEFAULT- Default access
7823 * MADV_ACCESS_LWP - Next LWP will access heavily
7824 * MADV_ACCESS_MANY- Many LWPs or processes will access heavily
7825 * MADV_PURGE - Contents will be immediately discarded
7826 */
7827 static int
7828 segvn_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
7829 {
7830 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7831 size_t page;
7832 int err = 0;
7833 int already_set;
7834 struct anon_map *amp;
7835 ulong_t anon_index;
7836 struct seg *next;
7837 lgrp_mem_policy_t policy;
7838 struct seg *prev;
7839 struct vnode *vp;
7840
7841 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
7842
7843 /*
7844 * In case of MADV_FREE/MADV_PURGE, we won't be modifying any segment
7845 * private data structures; so, we only need to grab READER's lock
7846 */
7847 if (behav != MADV_FREE && behav != MADV_PURGE) {
7848 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7849 if (svd->tr_state != SEGVN_TR_OFF) {
7850 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7851 return (0);
7852 }
7853 } else {
7854 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7855 }
7856
7857 /*
7858 * Large pages are assumed to be only turned on when accesses to the
7859 * segment's address range have spatial and temporal locality. That
7860 * justifies ignoring MADV_SEQUENTIAL for large page segments.
7861 * Also, ignore advice affecting lgroup memory allocation
7862 * if don't need to do lgroup optimizations on this system
7863 */
7864
7865 if ((behav == MADV_SEQUENTIAL &&
7866 (seg->s_szc != 0 || HAT_IS_REGION_COOKIE_VALID(svd->rcookie))) ||
7867 (!lgrp_optimizations() && (behav == MADV_ACCESS_DEFAULT ||
7868 behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY))) {
7869 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7870 return (0);
7871 }
7872
7873 if (behav == MADV_SEQUENTIAL || behav == MADV_ACCESS_DEFAULT ||
7874 behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY) {
7875 /*
7876 * Since we are going to unload hat mappings
7877 * we first have to flush the cache. Otherwise
7878 * this might lead to system panic if another
7879 * thread is doing physio on the range whose
7880 * mappings are unloaded by madvise(3C).
7881 */
7882 if (svd->softlockcnt > 0) {
7883 /*
7884 * If this is shared segment non 0 softlockcnt
7885 * means locked pages are still in use.
7886 */
7887 if (svd->type == MAP_SHARED) {
7888 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7889 return (EAGAIN);
7890 }
7891 /*
7892 * Since we do have the segvn writers lock
7893 * nobody can fill the cache with entries
7894 * belonging to this seg during the purge.
7895 * The flush either succeeds or we still
7896 * have pending I/Os. In the later case,
7897 * madvise(3C) fails.
7898 */
7899 segvn_purge(seg);
7900 if (svd->softlockcnt > 0) {
7901 /*
7902 * Since madvise(3C) is advisory and
7903 * it's not part of UNIX98, madvise(3C)
7904 * failure here doesn't cause any hardship.
7905 * Note that we don't block in "as" layer.
7906 */
7907 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7908 return (EAGAIN);
7909 }
7910 } else if (svd->type == MAP_SHARED && svd->amp != NULL &&
7911 svd->amp->a_softlockcnt > 0) {
7912 /*
7913 * Try to purge this amp's entries from pcache. It
7914 * will succeed only if other segments that share the
7915 * amp have no outstanding softlock's.
7916 */
7917 segvn_purge(seg);
7918 }
7919 }
7920
7921 amp = svd->amp;
7922 vp = svd->vp;
7923 if (behav == MADV_FREE || behav == MADV_PURGE) {
7924 pgcnt_t purged;
7925
7926 if (behav == MADV_FREE && (vp != NULL || amp == NULL)) {
7927 /*
7928 * MADV_FREE is not supported for segments with an
7929 * underlying object; if anonmap is NULL, anon slots
7930 * are not yet populated and there is nothing for us
7931 * to do. As MADV_FREE is advisory, we don't return an
7932 * error in either case.
7933 */
7934 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7935 return (0);
7936 }
7937
7938 if (amp == NULL) {
7939 /*
7940 * If we're here with a NULL anonmap, it's because we
7941 * are doing a MADV_PURGE. We have nothing to do, but
7942 * because MADV_PURGE isn't merely advisory, we return
7943 * an error in this case.
7944 */
7945 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7946 return (EBUSY);
7947 }
7948
7949 segvn_purge(seg);
7950
7951 page = seg_page(seg, addr);
7952 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7953 err = anon_disclaim(amp,
7954 svd->anon_index + page, len, behav, &purged);
7955
7956 if (purged != 0 && (svd->flags & MAP_NORESERVE)) {
7957 /*
7958 * If we purged pages on a MAP_NORESERVE mapping, we
7959 * need to be sure to now unreserve our reserved swap.
7960 * (We use the atomic operations to manipulate our
7961 * segment and address space counters because we only
7962 * have the corresponding locks held as reader, not
7963 * writer.)
7964 */
7965 ssize_t bytes = ptob(purged);
7966
7967 anon_unresv_zone(bytes, seg->s_as->a_proc->p_zone);
7968 atomic_add_long(&svd->swresv, -bytes);
7969 atomic_add_long(&seg->s_as->a_resvsize, -bytes);
7970 }
7971
7972 ANON_LOCK_EXIT(&amp->a_rwlock);
7973 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7974
7975 /*
7976 * MADV_PURGE and MADV_FREE differ in their return semantics:
7977 * because MADV_PURGE is designed to be bug-for-bug compatible
7978 * with its clumsy Linux forebear, it will fail where MADV_FREE
7979 * does not.
7980 */
7981 return (behav == MADV_PURGE ? err : 0);
7982 }
7983
7984 /*
7985 * If advice is to be applied to entire segment,
7986 * use advice field in seg_data structure
7987 * otherwise use appropriate vpage entry.
7988 */
7989 if ((addr == seg->s_base) && (len == seg->s_size)) {
7990 switch (behav) {
7991 case MADV_ACCESS_LWP:
7992 case MADV_ACCESS_MANY:
7993 case MADV_ACCESS_DEFAULT:
7994 /*
7995 * Set memory allocation policy for this segment
7996 */
7997 policy = lgrp_madv_to_policy(behav, len, svd->type);
7998 if (svd->type == MAP_SHARED)
7999 already_set = lgrp_shm_policy_set(policy, amp,
8000 svd->anon_index, vp, svd->offset, len);
8001 else {
8002 /*
8003 * For private memory, need writers lock on
8004 * address space because the segment may be
8005 * split or concatenated when changing policy
8006 */
8007 if (AS_READ_HELD(seg->s_as)) {
8008 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8009 return (IE_RETRY);
8010 }
8011
8012 already_set = lgrp_privm_policy_set(policy,
8013 &svd->policy_info, len);
8014 }
8015
8016 /*
8017 * If policy set already and it shouldn't be reapplied,
8018 * don't do anything.
8019 */
8020 if (already_set &&
8021 !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8022 break;
8023
8024 /*
8025 * Mark any existing pages in given range for
8026 * migration
8027 */
8028 page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8029 vp, svd->offset, 1);
8030
8031 /*
8032 * If same policy set already or this is a shared
8033 * memory segment, don't need to try to concatenate
8034 * segment with adjacent ones.
8035 */
8036 if (already_set || svd->type == MAP_SHARED)
8037 break;
8038
8039 /*
8040 * Try to concatenate this segment with previous
8041 * one and next one, since we changed policy for
8042 * this one and it may be compatible with adjacent
8043 * ones now.
8044 */
8045 prev = AS_SEGPREV(seg->s_as, seg);
8046 next = AS_SEGNEXT(seg->s_as, seg);
8047
8048 if (next && next->s_ops == &segvn_ops &&
8049 addr + len == next->s_base)
8050 (void) segvn_concat(seg, next, 1);
8051
8052 if (prev && prev->s_ops == &segvn_ops &&
8053 addr == prev->s_base + prev->s_size) {
8054 /*
8055 * Drop lock for private data of current
8056 * segment before concatenating (deleting) it
8057 * and return IE_REATTACH to tell as_ctl() that
8058 * current segment has changed
8059 */
8060 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8061 if (!segvn_concat(prev, seg, 1))
8062 err = IE_REATTACH;
8063
8064 return (err);
8065 }
8066 break;
8067
8068 case MADV_SEQUENTIAL:
8069 /*
8070 * unloading mapping guarantees
8071 * detection in segvn_fault
8072 */
8073 ASSERT(seg->s_szc == 0);
8074 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8075 hat_unload(seg->s_as->a_hat, addr, len,
8076 HAT_UNLOAD);
8077 /* FALLTHROUGH */
8078 case MADV_NORMAL:
8079 case MADV_RANDOM:
8080 svd->advice = (uchar_t)behav;
8081 svd->pageadvice = 0;
8082 break;
8083 case MADV_WILLNEED: /* handled in memcntl */
8084 case MADV_DONTNEED: /* handled in memcntl */
8085 case MADV_FREE: /* handled above */
8086 case MADV_PURGE: /* handled above */
8087 break;
8088 default:
8089 err = EINVAL;
8090 }
8091 } else {
8092 caddr_t eaddr;
8093 struct seg *new_seg;
8094 struct segvn_data *new_svd;
8095 uoff_t off;
8096 caddr_t oldeaddr;
8097
8098 page = seg_page(seg, addr);
8099
8100 segvn_vpage(seg);
8101 if (svd->vpage == NULL) {
8102 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8103 return (ENOMEM);
8104 }
8105
8106 switch (behav) {
8107 struct vpage *bvpp, *evpp;
8108
8109 case MADV_ACCESS_LWP:
8110 case MADV_ACCESS_MANY:
8111 case MADV_ACCESS_DEFAULT:
8112 /*
8113 * Set memory allocation policy for portion of this
8114 * segment
8115 */
8116
8117 /*
8118 * Align address and length of advice to page
8119 * boundaries for large pages
8120 */
8121 if (seg->s_szc != 0) {
8122 size_t pgsz;
8123
8124 pgsz = page_get_pagesize(seg->s_szc);
8125 addr = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
8126 len = P2ROUNDUP(len, pgsz);
8127 }
8128
8129 /*
8130 * Check to see whether policy is set already
8131 */
8132 policy = lgrp_madv_to_policy(behav, len, svd->type);
8133
8134 anon_index = svd->anon_index + page;
8135 off = svd->offset + (uintptr_t)(addr - seg->s_base);
8136
8137 if (svd->type == MAP_SHARED)
8138 already_set = lgrp_shm_policy_set(policy, amp,
8139 anon_index, vp, off, len);
8140 else
8141 already_set =
8142 (policy == svd->policy_info.mem_policy);
8143
8144 /*
8145 * If policy set already and it shouldn't be reapplied,
8146 * don't do anything.
8147 */
8148 if (already_set &&
8149 !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8150 break;
8151
8152 /*
8153 * For private memory, need writers lock on
8154 * address space because the segment may be
8155 * split or concatenated when changing policy
8156 */
8157 if (svd->type == MAP_PRIVATE &&
8158 AS_READ_HELD(seg->s_as)) {
8159 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8160 return (IE_RETRY);
8161 }
8162
8163 /*
8164 * Mark any existing pages in given range for
8165 * migration
8166 */
8167 page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8168 vp, svd->offset, 1);
8169
8170 /*
8171 * Don't need to try to split or concatenate
8172 * segments, since policy is same or this is a shared
8173 * memory segment
8174 */
8175 if (already_set || svd->type == MAP_SHARED)
8176 break;
8177
8178 if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
8179 ASSERT(svd->amp == NULL);
8180 ASSERT(svd->tr_state == SEGVN_TR_OFF);
8181 ASSERT(svd->softlockcnt == 0);
8182 hat_leave_region(seg->s_as->a_hat, svd->rcookie,
8183 HAT_REGION_TEXT);
8184 svd->rcookie = HAT_INVALID_REGION_COOKIE;
8185 }
8186
8187 /*
8188 * Split off new segment if advice only applies to a
8189 * portion of existing segment starting in middle
8190 */
8191 new_seg = NULL;
8192 eaddr = addr + len;
8193 oldeaddr = seg->s_base + seg->s_size;
8194 if (addr > seg->s_base) {
8195 /*
8196 * Must flush I/O page cache
8197 * before splitting segment
8198 */
8199 if (svd->softlockcnt > 0)
8200 segvn_purge(seg);
8201
8202 /*
8203 * Split segment and return IE_REATTACH to tell
8204 * as_ctl() that current segment changed
8205 */
8206 new_seg = segvn_split_seg(seg, addr);
8207 new_svd = (struct segvn_data *)new_seg->s_data;
8208 err = IE_REATTACH;
8209
8210 /*
8211 * If new segment ends where old one
8212 * did, try to concatenate the new
8213 * segment with next one.
8214 */
8215 if (eaddr == oldeaddr) {
8216 /*
8217 * Set policy for new segment
8218 */
8219 (void) lgrp_privm_policy_set(policy,
8220 &new_svd->policy_info,
8221 new_seg->s_size);
8222
8223 next = AS_SEGNEXT(new_seg->s_as,
8224 new_seg);
8225
8226 if (next &&
8227 next->s_ops == &segvn_ops &&
8228 eaddr == next->s_base)
8229 (void) segvn_concat(new_seg,
8230 next, 1);
8231 }
8232 }
8233
8234 /*
8235 * Split off end of existing segment if advice only
8236 * applies to a portion of segment ending before
8237 * end of the existing segment
8238 */
8239 if (eaddr < oldeaddr) {
8240 /*
8241 * Must flush I/O page cache
8242 * before splitting segment
8243 */
8244 if (svd->softlockcnt > 0)
8245 segvn_purge(seg);
8246
8247 /*
8248 * If beginning of old segment was already
8249 * split off, use new segment to split end off
8250 * from.
8251 */
8252 if (new_seg != NULL && new_seg != seg) {
8253 /*
8254 * Split segment
8255 */
8256 (void) segvn_split_seg(new_seg, eaddr);
8257
8258 /*
8259 * Set policy for new segment
8260 */
8261 (void) lgrp_privm_policy_set(policy,
8262 &new_svd->policy_info,
8263 new_seg->s_size);
8264 } else {
8265 /*
8266 * Split segment and return IE_REATTACH
8267 * to tell as_ctl() that current
8268 * segment changed
8269 */
8270 (void) segvn_split_seg(seg, eaddr);
8271 err = IE_REATTACH;
8272
8273 (void) lgrp_privm_policy_set(policy,
8274 &svd->policy_info, seg->s_size);
8275
8276 /*
8277 * If new segment starts where old one
8278 * did, try to concatenate it with
8279 * previous segment.
8280 */
8281 if (addr == seg->s_base) {
8282 prev = AS_SEGPREV(seg->s_as,
8283 seg);
8284
8285 /*
8286 * Drop lock for private data
8287 * of current segment before
8288 * concatenating (deleting) it
8289 */
8290 if (prev &&
8291 prev->s_ops ==
8292 &segvn_ops &&
8293 addr == prev->s_base +
8294 prev->s_size) {
8295 SEGVN_LOCK_EXIT(
8296 seg->s_as,
8297 &svd->lock);
8298 (void) segvn_concat(
8299 prev, seg, 1);
8300 return (err);
8301 }
8302 }
8303 }
8304 }
8305 break;
8306 case MADV_SEQUENTIAL:
8307 ASSERT(seg->s_szc == 0);
8308 ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8309 hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
8310 /* FALLTHROUGH */
8311 case MADV_NORMAL:
8312 case MADV_RANDOM:
8313 bvpp = &svd->vpage[page];
8314 evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8315 for (; bvpp < evpp; bvpp++)
8316 VPP_SETADVICE(bvpp, behav);
8317 svd->advice = MADV_NORMAL;
8318 break;
8319 case MADV_WILLNEED: /* handled in memcntl */
8320 case MADV_DONTNEED: /* handled in memcntl */
8321 case MADV_FREE: /* handled above */
8322 case MADV_PURGE: /* handled above */
8323 break;
8324 default:
8325 err = EINVAL;
8326 }
8327 }
8328 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8329 return (err);
8330 }
8331
8332 /*
8333 * There is one kind of inheritance that can be specified for pages:
8334 *
8335 * SEGP_INH_ZERO - Pages should be zeroed in the child
8336 */
8337 static int
8338 segvn_inherit(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
8339 {
8340 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8341 struct vpage *bvpp, *evpp;
8342 size_t page;
8343 int ret = 0;
8344
8345 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
8346
8347 /* Can't support something we don't know about */
8348 if (behav != SEGP_INH_ZERO)
8349 return (ENOTSUP);
8350
8351 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
8352
8353 /*
8354 * This must be a straightforward anonymous segment that is mapped
8355 * privately and is not backed by a vnode.
8356 */
8357 if (svd->tr_state != SEGVN_TR_OFF ||
8358 svd->type != MAP_PRIVATE ||
8359 svd->vp != NULL) {
8360 ret = EINVAL;
8361 goto out;
8362 }
8363
8364 /*
8365 * If the entire segment has been marked as inherit zero, then no reason
8366 * to do anything else.
8367 */
8368 if (svd->svn_inz == SEGVN_INZ_ALL) {
8369 ret = 0;
8370 goto out;
8371 }
8372
8373 /*
8374 * If this applies to the entire segment, simply mark it and we're done.
8375 */
8376 if ((addr == seg->s_base) && (len == seg->s_size)) {
8377 svd->svn_inz = SEGVN_INZ_ALL;
8378 ret = 0;
8379 goto out;
8380 }
8381
8382 /*
8383 * We've been asked to mark a subset of this segment as inherit zero,
8384 * therefore we need to mainpulate its vpages.
8385 */
8386 if (svd->vpage == NULL) {
8387 segvn_vpage(seg);
8388 if (svd->vpage == NULL) {
8389 ret = ENOMEM;
8390 goto out;
8391 }
8392 }
8393
8394 svd->svn_inz = SEGVN_INZ_VPP;
8395 page = seg_page(seg, addr);
8396 bvpp = &svd->vpage[page];
8397 evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8398 for (; bvpp < evpp; bvpp++)
8399 VPP_SETINHZERO(bvpp);
8400 ret = 0;
8401
8402 out:
8403 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8404 return (ret);
8405 }
8406
8407 /*
8408 * Create a vpage structure for this seg.
8409 */
8410 static void
8411 segvn_vpage(struct seg *seg)
8412 {
8413 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8414 struct vpage *vp, *evp;
8415 static pgcnt_t page_limit = 0;
8416
8417 ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8418
8419 /*
8420 * If no vpage structure exists, allocate one. Copy the protections
8421 * and the advice from the segment itself to the individual pages.
8422 */
8423 if (svd->vpage == NULL) {
8424 /*
8425 * Start by calculating the number of pages we must allocate to
8426 * track the per-page vpage structs needs for this entire
8427 * segment. If we know now that it will require more than our
8428 * heuristic for the maximum amount of kmem we can consume then
8429 * fail. We do this here, instead of trying to detect this deep
8430 * in page_resv and propagating the error up, since the entire
8431 * memory allocation stack is not amenable to passing this
8432 * back. Instead, it wants to keep trying.
8433 *
8434 * As a heuristic we set a page limit of 5/8s of total_pages
8435 * for this allocation. We use shifts so that no floating
8436 * point conversion takes place and only need to do the
8437 * calculation once.
8438 */
8439 ulong_t mem_needed = seg_pages(seg) * sizeof (struct vpage);
8440 pgcnt_t npages = mem_needed >> PAGESHIFT;
8441
8442 if (page_limit == 0)
8443 page_limit = (total_pages >> 1) + (total_pages >> 3);
8444
8445 if (npages > page_limit)
8446 return;
8447
8448 svd->pageadvice = 1;
8449 svd->vpage = kmem_zalloc(mem_needed, KM_SLEEP);
8450 evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
8451 for (vp = svd->vpage; vp < evp; vp++) {
8452 VPP_SETPROT(vp, svd->prot);
8453 VPP_SETADVICE(vp, svd->advice);
8454 }
8455 }
8456 }
8457
8458 /*
8459 * Dump the pages belonging to this segvn segment.
8460 */
8461 static void
8462 segvn_dump(struct seg *seg)
8463 {
8464 struct segvn_data *svd;
8465 page_t *pp;
8466 struct anon_map *amp;
8467 ulong_t anon_index;
8468 struct vnode *vp;
8469 uoff_t off, offset;
8470 pfn_t pfn;
8471 pgcnt_t page, npages;
8472 caddr_t addr;
8473
8474 npages = seg_pages(seg);
8475 svd = (struct segvn_data *)seg->s_data;
8476 vp = svd->vp;
8477 off = offset = svd->offset;
8478 addr = seg->s_base;
8479
8480 if ((amp = svd->amp) != NULL) {
8481 anon_index = svd->anon_index;
8482 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8483 }
8484
8485 for (page = 0; page < npages; page++, offset += PAGESIZE) {
8486 struct anon *ap;
8487 int we_own_it = 0;
8488
8489 if (amp && (ap = anon_get_ptr(svd->amp->ahp, anon_index++))) {
8490 swap_xlate_nopanic(ap, &vp, &off);
8491 } else {
8492 vp = svd->vp;
8493 off = offset;
8494 }
8495
8496 /*
8497 * If pp == NULL, the page either does not exist
8498 * or is exclusively locked. So determine if it
8499 * exists before searching for it.
8500 */
8501
8502 if ((pp = page_lookup_nowait(vp, off, SE_SHARED)))
8503 we_own_it = 1;
8504 else
8505 pp = page_exists(vp, off);
8506
8507 if (pp) {
8508 pfn = page_pptonum(pp);
8509 dump_addpage(seg->s_as, addr, pfn);
8510 if (we_own_it)
8511 page_unlock(pp);
8512 }
8513 addr += PAGESIZE;
8514 dump_timeleft = dump_timeout;
8515 }
8516
8517 if (amp != NULL)
8518 ANON_LOCK_EXIT(&amp->a_rwlock);
8519 }
8520
8521 #ifdef DEBUG
8522 static uint32_t segvn_pglock_mtbf = 0;
8523 #endif
8524
8525 #define PCACHE_SHWLIST ((page_t *)-2)
8526 #define NOPCACHE_SHWLIST ((page_t *)-1)
8527
8528 /*
8529 * Lock/Unlock anon pages over a given range. Return shadow list. This routine
8530 * uses global segment pcache to cache shadow lists (i.e. pp arrays) of pages
8531 * to avoid the overhead of per page locking, unlocking for subsequent IOs to
8532 * the same parts of the segment. Currently shadow list creation is only
8533 * supported for pure anon segments. MAP_PRIVATE segment pcache entries are
8534 * tagged with segment pointer, starting virtual address and length. This
8535 * approach for MAP_SHARED segments may add many pcache entries for the same
8536 * set of pages and lead to long hash chains that decrease pcache lookup
8537 * performance. To avoid this issue for shared segments shared anon map and
8538 * starting anon index are used for pcache entry tagging. This allows all
8539 * segments to share pcache entries for the same anon range and reduces pcache
8540 * chain's length as well as memory overhead from duplicate shadow lists and
8541 * pcache entries.
8542 *
8543 * softlockcnt field in segvn_data structure counts the number of F_SOFTLOCK'd
8544 * pages via segvn_fault() and pagelock'd pages via this routine. But pagelock
8545 * part of softlockcnt accounting is done differently for private and shared
8546 * segments. In private segment case softlock is only incremented when a new
8547 * shadow list is created but not when an existing one is found via
8548 * seg_plookup(). pcache entries have reference count incremented/decremented
8549 * by each seg_plookup()/seg_pinactive() operation. Only entries that have 0
8550 * reference count can be purged (and purging is needed before segment can be
8551 * freed). When a private segment pcache entry is purged segvn_reclaim() will
8552 * decrement softlockcnt. Since in private segment case each of its pcache
8553 * entries only belongs to this segment we can expect that when
8554 * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8555 * segment purge will succeed and softlockcnt will drop to 0. In shared
8556 * segment case reference count in pcache entry counts active locks from many
8557 * different segments so we can't expect segment purging to succeed even when
8558 * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8559 * segment. To be able to determine when there're no pending pagelocks in
8560 * shared segment case we don't rely on purging to make softlockcnt drop to 0
8561 * but instead softlockcnt is incremented and decremented for every
8562 * segvn_pagelock(L_PAGELOCK/L_PAGEUNLOCK) call regardless if a new shadow
8563 * list was created or an existing one was found. When softlockcnt drops to 0
8564 * this segment no longer has any claims for pcached shadow lists and the
8565 * segment can be freed even if there're still active pcache entries
8566 * shared by this segment anon map. Shared segment pcache entries belong to
8567 * anon map and are typically removed when anon map is freed after all
8568 * processes destroy the segments that use this anon map.
8569 */
8570 static int
8571 segvn_pagelock(struct seg *seg, caddr_t addr, size_t len, struct page ***ppp,
8572 enum lock_type type, enum seg_rw rw)
8573 {
8574 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8575 size_t np;
8576 pgcnt_t adjustpages;
8577 pgcnt_t npages;
8578 ulong_t anon_index;
8579 uint_t protchk = (rw == S_READ) ? PROT_READ : PROT_WRITE;
8580 uint_t error;
8581 struct anon_map *amp;
8582 pgcnt_t anpgcnt;
8583 struct page **pplist, **pl, *pp;
8584 caddr_t a;
8585 size_t page;
8586 caddr_t lpgaddr, lpgeaddr;
8587 anon_sync_obj_t cookie;
8588 int anlock;
8589 struct anon_map *pamp;
8590 caddr_t paddr;
8591 seg_preclaim_cbfunc_t preclaim_callback;
8592 size_t pgsz;
8593 int use_pcache;
8594 size_t wlen;
8595 uint_t pflags = 0;
8596 int sftlck_sbase = 0;
8597 int sftlck_send = 0;
8598
8599 #ifdef DEBUG
8600 if (type == L_PAGELOCK && segvn_pglock_mtbf) {
8601 hrtime_t ts = gethrtime();
8602 if ((ts % segvn_pglock_mtbf) == 0) {
8603 return (ENOTSUP);
8604 }
8605 if ((ts % segvn_pglock_mtbf) == 1) {
8606 return (EFAULT);
8607 }
8608 }
8609 #endif
8610
8611 ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as));
8612 ASSERT(type == L_PAGELOCK || type == L_PAGEUNLOCK);
8613
8614 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8615
8616 /*
8617 * for now we only support pagelock to anon memory. We would have to
8618 * check protections for vnode objects and call into the vnode driver.
8619 * That's too much for a fast path. Let the fault entry point handle
8620 * it.
8621 */
8622 if (svd->vp != NULL) {
8623 if (type == L_PAGELOCK) {
8624 error = ENOTSUP;
8625 goto out;
8626 }
8627 panic("segvn_pagelock(L_PAGEUNLOCK): vp != NULL");
8628 }
8629 if ((amp = svd->amp) == NULL) {
8630 if (type == L_PAGELOCK) {
8631 error = EFAULT;
8632 goto out;
8633 }
8634 panic("segvn_pagelock(L_PAGEUNLOCK): amp == NULL");
8635 }
8636 if (rw != S_READ && rw != S_WRITE) {
8637 if (type == L_PAGELOCK) {
8638 error = ENOTSUP;
8639 goto out;
8640 }
8641 panic("segvn_pagelock(L_PAGEUNLOCK): bad rw");
8642 }
8643
8644 if (seg->s_szc != 0) {
8645 /*
8646 * We are adjusting the pagelock region to the large page size
8647 * boundary because the unlocked part of a large page cannot
8648 * be freed anyway unless all constituent pages of a large
8649 * page are locked. Bigger regions reduce pcache chain length
8650 * and improve lookup performance. The tradeoff is that the
8651 * very first segvn_pagelock() call for a given page is more
8652 * expensive if only 1 page_t is needed for IO. This is only
8653 * an issue if pcache entry doesn't get reused by several
8654 * subsequent calls. We optimize here for the case when pcache
8655 * is heavily used by repeated IOs to the same address range.
8656 *
8657 * Note segment's page size cannot change while we are holding
8658 * as lock. And then it cannot change while softlockcnt is
8659 * not 0. This will allow us to correctly recalculate large
8660 * page size region for the matching pageunlock/reclaim call
8661 * since as_pageunlock() caller must always match
8662 * as_pagelock() call's addr and len.
8663 *
8664 * For pageunlock *ppp points to the pointer of page_t that
8665 * corresponds to the real unadjusted start address. Similar
8666 * for pagelock *ppp must point to the pointer of page_t that
8667 * corresponds to the real unadjusted start address.
8668 */
8669 pgsz = page_get_pagesize(seg->s_szc);
8670 CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
8671 adjustpages = btop((uintptr_t)(addr - lpgaddr));
8672 } else if (len < segvn_pglock_comb_thrshld) {
8673 lpgaddr = addr;
8674 lpgeaddr = addr + len;
8675 adjustpages = 0;
8676 pgsz = PAGESIZE;
8677 } else {
8678 /*
8679 * Align the address range of large enough requests to allow
8680 * combining of different shadow lists into 1 to reduce memory
8681 * overhead from potentially overlapping large shadow lists
8682 * (worst case is we have a 1MB IO into buffers with start
8683 * addresses separated by 4K). Alignment is only possible if
8684 * padded chunks have sufficient access permissions. Note
8685 * permissions won't change between L_PAGELOCK and
8686 * L_PAGEUNLOCK calls since non 0 softlockcnt will force
8687 * segvn_setprot() to wait until softlockcnt drops to 0. This
8688 * allows us to determine in L_PAGEUNLOCK the same range we
8689 * computed in L_PAGELOCK.
8690 *
8691 * If alignment is limited by segment ends set
8692 * sftlck_sbase/sftlck_send flags. In L_PAGELOCK case when
8693 * these flags are set bump softlockcnt_sbase/softlockcnt_send
8694 * per segment counters. In L_PAGEUNLOCK case decrease
8695 * softlockcnt_sbase/softlockcnt_send counters if
8696 * sftlck_sbase/sftlck_send flags are set. When
8697 * softlockcnt_sbase/softlockcnt_send are non 0
8698 * segvn_concat()/segvn_extend_prev()/segvn_extend_next()
8699 * won't merge the segments. This restriction combined with
8700 * restriction on segment unmapping and splitting for segments
8701 * that have non 0 softlockcnt allows L_PAGEUNLOCK to
8702 * correctly determine the same range that was previously
8703 * locked by matching L_PAGELOCK.
8704 */
8705 pflags = SEGP_PSHIFT | (segvn_pglock_comb_bshift << 16);
8706 pgsz = PAGESIZE;
8707 if (svd->type == MAP_PRIVATE) {
8708 lpgaddr = (caddr_t)P2ALIGN((uintptr_t)addr,
8709 segvn_pglock_comb_balign);
8710 if (lpgaddr < seg->s_base) {
8711 lpgaddr = seg->s_base;
8712 sftlck_sbase = 1;
8713 }
8714 } else {
8715 ulong_t aix = svd->anon_index + seg_page(seg, addr);
8716 ulong_t aaix = P2ALIGN(aix, segvn_pglock_comb_palign);
8717 if (aaix < svd->anon_index) {
8718 lpgaddr = seg->s_base;
8719 sftlck_sbase = 1;
8720 } else {
8721 lpgaddr = addr - ptob(aix - aaix);
8722 ASSERT(lpgaddr >= seg->s_base);
8723 }
8724 }
8725 if (svd->pageprot && lpgaddr != addr) {
8726 struct vpage *vp = &svd->vpage[seg_page(seg, lpgaddr)];
8727 struct vpage *evp = &svd->vpage[seg_page(seg, addr)];
8728 while (vp < evp) {
8729 if ((VPP_PROT(vp) & protchk) == 0) {
8730 break;
8731 }
8732 vp++;
8733 }
8734 if (vp < evp) {
8735 lpgaddr = addr;
8736 pflags = 0;
8737 }
8738 }
8739 lpgeaddr = addr + len;
8740 if (pflags) {
8741 if (svd->type == MAP_PRIVATE) {
8742 lpgeaddr = (caddr_t)P2ROUNDUP(
8743 (uintptr_t)lpgeaddr,
8744 segvn_pglock_comb_balign);
8745 } else {
8746 ulong_t aix = svd->anon_index +
8747 seg_page(seg, lpgeaddr);
8748 ulong_t aaix = P2ROUNDUP(aix,
8749 segvn_pglock_comb_palign);
8750 if (aaix < aix) {
8751 lpgeaddr = 0;
8752 } else {
8753 lpgeaddr += ptob(aaix - aix);
8754 }
8755 }
8756 if (lpgeaddr == 0 ||
8757 lpgeaddr > seg->s_base + seg->s_size) {
8758 lpgeaddr = seg->s_base + seg->s_size;
8759 sftlck_send = 1;
8760 }
8761 }
8762 if (svd->pageprot && lpgeaddr != addr + len) {
8763 struct vpage *vp;
8764 struct vpage *evp;
8765
8766 vp = &svd->vpage[seg_page(seg, addr + len)];
8767 evp = &svd->vpage[seg_page(seg, lpgeaddr)];
8768
8769 while (vp < evp) {
8770 if ((VPP_PROT(vp) & protchk) == 0) {
8771 break;
8772 }
8773 vp++;
8774 }
8775 if (vp < evp) {
8776 lpgeaddr = addr + len;
8777 }
8778 }
8779 adjustpages = btop((uintptr_t)(addr - lpgaddr));
8780 }
8781
8782 /*
8783 * For MAP_SHARED segments we create pcache entries tagged by amp and
8784 * anon index so that we can share pcache entries with other segments
8785 * that map this amp. For private segments pcache entries are tagged
8786 * with segment and virtual address.
8787 */
8788 if (svd->type == MAP_SHARED) {
8789 pamp = amp;
8790 paddr = (caddr_t)((lpgaddr - seg->s_base) +
8791 ptob(svd->anon_index));
8792 preclaim_callback = shamp_reclaim;
8793 } else {
8794 pamp = NULL;
8795 paddr = lpgaddr;
8796 preclaim_callback = segvn_reclaim;
8797 }
8798
8799 if (type == L_PAGEUNLOCK) {
8800 VM_STAT_ADD(segvnvmstats.pagelock[0]);
8801
8802 /*
8803 * update hat ref bits for /proc. We need to make sure
8804 * that threads tracing the ref and mod bits of the
8805 * address space get the right data.
8806 * Note: page ref and mod bits are updated at reclaim time
8807 */
8808 if (seg->s_as->a_vbits) {
8809 for (a = addr; a < addr + len; a += PAGESIZE) {
8810 if (rw == S_WRITE) {
8811 hat_setstat(seg->s_as, a,
8812 PAGESIZE, P_REF | P_MOD);
8813 } else {
8814 hat_setstat(seg->s_as, a,
8815 PAGESIZE, P_REF);
8816 }
8817 }
8818 }
8819
8820 /*
8821 * Check the shadow list entry after the last page used in
8822 * this IO request. If it's NOPCACHE_SHWLIST the shadow list
8823 * was not inserted into pcache and is not large page
8824 * adjusted. In this case call reclaim callback directly and
8825 * don't adjust the shadow list start and size for large
8826 * pages.
8827 */
8828 npages = btop(len);
8829 if ((*ppp)[npages] == NOPCACHE_SHWLIST) {
8830 void *ptag;
8831 if (pamp != NULL) {
8832 ASSERT(svd->type == MAP_SHARED);
8833 ptag = (void *)pamp;
8834 paddr = (caddr_t)((addr - seg->s_base) +
8835 ptob(svd->anon_index));
8836 } else {
8837 ptag = (void *)seg;
8838 paddr = addr;
8839 }
8840 (*preclaim_callback)(ptag, paddr, len, *ppp, rw, 0);
8841 } else {
8842 ASSERT((*ppp)[npages] == PCACHE_SHWLIST ||
8843 IS_SWAPFSVP((*ppp)[npages]->p_vnode));
8844 len = lpgeaddr - lpgaddr;
8845 npages = btop(len);
8846 seg_pinactive(seg, pamp, paddr, len,
8847 *ppp - adjustpages, rw, pflags, preclaim_callback);
8848 }
8849
8850 if (pamp != NULL) {
8851 ASSERT(svd->type == MAP_SHARED);
8852 ASSERT(svd->softlockcnt >= npages);
8853 atomic_add_long((ulong_t *)&svd->softlockcnt, -npages);
8854 }
8855
8856 if (sftlck_sbase) {
8857 ASSERT(svd->softlockcnt_sbase > 0);
8858 atomic_dec_ulong((ulong_t *)&svd->softlockcnt_sbase);
8859 }
8860 if (sftlck_send) {
8861 ASSERT(svd->softlockcnt_send > 0);
8862 atomic_dec_ulong((ulong_t *)&svd->softlockcnt_send);
8863 }
8864
8865 /*
8866 * If someone is blocked while unmapping, we purge
8867 * segment page cache and thus reclaim pplist synchronously
8868 * without waiting for seg_pasync_thread. This speeds up
8869 * unmapping in cases where munmap(2) is called, while
8870 * raw async i/o is still in progress or where a thread
8871 * exits on data fault in a multithreaded application.
8872 */
8873 if (AS_ISUNMAPWAIT(seg->s_as)) {
8874 if (svd->softlockcnt == 0) {
8875 mutex_enter(&seg->s_as->a_contents);
8876 if (AS_ISUNMAPWAIT(seg->s_as)) {
8877 AS_CLRUNMAPWAIT(seg->s_as);
8878 cv_broadcast(&seg->s_as->a_cv);
8879 }
8880 mutex_exit(&seg->s_as->a_contents);
8881 } else if (pamp == NULL) {
8882 /*
8883 * softlockcnt is not 0 and this is a
8884 * MAP_PRIVATE segment. Try to purge its
8885 * pcache entries to reduce softlockcnt.
8886 * If it drops to 0 segvn_reclaim()
8887 * will wake up a thread waiting on
8888 * unmapwait flag.
8889 *
8890 * We don't purge MAP_SHARED segments with non
8891 * 0 softlockcnt since IO is still in progress
8892 * for such segments.
8893 */
8894 ASSERT(svd->type == MAP_PRIVATE);
8895 segvn_purge(seg);
8896 }
8897 }
8898 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8899 return (0);
8900 }
8901
8902 /* The L_PAGELOCK case ... */
8903
8904 VM_STAT_ADD(segvnvmstats.pagelock[1]);
8905
8906 /*
8907 * For MAP_SHARED segments we have to check protections before
8908 * seg_plookup() since pcache entries may be shared by many segments
8909 * with potentially different page protections.
8910 */
8911 if (pamp != NULL) {
8912 ASSERT(svd->type == MAP_SHARED);
8913 if (svd->pageprot == 0) {
8914 if ((svd->prot & protchk) == 0) {
8915 error = EACCES;
8916 goto out;
8917 }
8918 } else {
8919 /*
8920 * check page protections
8921 */
8922 caddr_t ea;
8923
8924 if (seg->s_szc) {
8925 a = lpgaddr;
8926 ea = lpgeaddr;
8927 } else {
8928 a = addr;
8929 ea = addr + len;
8930 }
8931 for (; a < ea; a += pgsz) {
8932 struct vpage *vp;
8933
8934 ASSERT(seg->s_szc == 0 ||
8935 sameprot(seg, a, pgsz));
8936 vp = &svd->vpage[seg_page(seg, a)];
8937 if ((VPP_PROT(vp) & protchk) == 0) {
8938 error = EACCES;
8939 goto out;
8940 }
8941 }
8942 }
8943 }
8944
8945 /*
8946 * try to find pages in segment page cache
8947 */
8948 pplist = seg_plookup(seg, pamp, paddr, lpgeaddr - lpgaddr, rw, pflags);
8949 if (pplist != NULL) {
8950 if (pamp != NULL) {
8951 npages = btop((uintptr_t)(lpgeaddr - lpgaddr));
8952 ASSERT(svd->type == MAP_SHARED);
8953 atomic_add_long((ulong_t *)&svd->softlockcnt,
8954 npages);
8955 }
8956 if (sftlck_sbase) {
8957 atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
8958 }
8959 if (sftlck_send) {
8960 atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
8961 }
8962 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8963 *ppp = pplist + adjustpages;
8964 return (0);
8965 }
8966
8967 /*
8968 * For MAP_SHARED segments we already verified above that segment
8969 * protections allow this pagelock operation.
8970 */
8971 if (pamp == NULL) {
8972 ASSERT(svd->type == MAP_PRIVATE);
8973 if (svd->pageprot == 0) {
8974 if ((svd->prot & protchk) == 0) {
8975 error = EACCES;
8976 goto out;
8977 }
8978 if (svd->prot & PROT_WRITE) {
8979 wlen = lpgeaddr - lpgaddr;
8980 } else {
8981 wlen = 0;
8982 ASSERT(rw == S_READ);
8983 }
8984 } else {
8985 int wcont = 1;
8986 /*
8987 * check page protections
8988 */
8989 for (a = lpgaddr, wlen = 0; a < lpgeaddr; a += pgsz) {
8990 struct vpage *vp;
8991
8992 ASSERT(seg->s_szc == 0 ||
8993 sameprot(seg, a, pgsz));
8994 vp = &svd->vpage[seg_page(seg, a)];
8995 if ((VPP_PROT(vp) & protchk) == 0) {
8996 error = EACCES;
8997 goto out;
8998 }
8999 if (wcont && (VPP_PROT(vp) & PROT_WRITE)) {
9000 wlen += pgsz;
9001 } else {
9002 wcont = 0;
9003 ASSERT(rw == S_READ);
9004 }
9005 }
9006 }
9007 ASSERT(rw == S_READ || wlen == lpgeaddr - lpgaddr);
9008 ASSERT(rw == S_WRITE || wlen <= lpgeaddr - lpgaddr);
9009 }
9010
9011 /*
9012 * Only build large page adjusted shadow list if we expect to insert
9013 * it into pcache. For large enough pages it's a big overhead to
9014 * create a shadow list of the entire large page. But this overhead
9015 * should be amortized over repeated pcache hits on subsequent reuse
9016 * of this shadow list (IO into any range within this shadow list will
9017 * find it in pcache since we large page align the request for pcache
9018 * lookups). pcache performance is improved with bigger shadow lists
9019 * as it reduces the time to pcache the entire big segment and reduces
9020 * pcache chain length.
9021 */
9022 if (seg_pinsert_check(seg, pamp, paddr,
9023 lpgeaddr - lpgaddr, pflags) == SEGP_SUCCESS) {
9024 addr = lpgaddr;
9025 len = lpgeaddr - lpgaddr;
9026 use_pcache = 1;
9027 } else {
9028 use_pcache = 0;
9029 /*
9030 * Since this entry will not be inserted into the pcache, we
9031 * will not do any adjustments to the starting address or
9032 * size of the memory to be locked.
9033 */
9034 adjustpages = 0;
9035 }
9036 npages = btop(len);
9037
9038 pplist = kmem_alloc(sizeof (page_t *) * (npages + 1), KM_SLEEP);
9039 pl = pplist;
9040 *ppp = pplist + adjustpages;
9041 /*
9042 * If use_pcache is 0 this shadow list is not large page adjusted.
9043 * Record this info in the last entry of shadow array so that
9044 * L_PAGEUNLOCK can determine if it should large page adjust the
9045 * address range to find the real range that was locked.
9046 */
9047 pl[npages] = use_pcache ? PCACHE_SHWLIST : NOPCACHE_SHWLIST;
9048
9049 page = seg_page(seg, addr);
9050 anon_index = svd->anon_index + page;
9051
9052 anlock = 0;
9053 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9054 ASSERT(amp->a_szc >= seg->s_szc);
9055 anpgcnt = page_get_pagecnt(amp->a_szc);
9056 for (a = addr; a < addr + len; a += PAGESIZE, anon_index++) {
9057 struct anon *ap;
9058 struct vnode *vp;
9059 uoff_t off;
9060
9061 /*
9062 * Lock and unlock anon array only once per large page.
9063 * anon_array_enter() locks the root anon slot according to
9064 * a_szc which can't change while anon map is locked. We lock
9065 * anon the first time through this loop and each time we
9066 * reach anon index that corresponds to a root of a large
9067 * page.
9068 */
9069 if (a == addr || P2PHASE(anon_index, anpgcnt) == 0) {
9070 ASSERT(anlock == 0);
9071 anon_array_enter(amp, anon_index, &cookie);
9072 anlock = 1;
9073 }
9074 ap = anon_get_ptr(amp->ahp, anon_index);
9075
9076 /*
9077 * We must never use seg_pcache for COW pages
9078 * because we might end up with original page still
9079 * lying in seg_pcache even after private page is
9080 * created. This leads to data corruption as
9081 * aio_write refers to the page still in cache
9082 * while all other accesses refer to the private
9083 * page.
9084 */
9085 if (ap == NULL || ap->an_refcnt != 1) {
9086 struct vpage *vpage;
9087
9088 if (seg->s_szc) {
9089 error = EFAULT;
9090 break;
9091 }
9092 if (svd->vpage != NULL) {
9093 vpage = &svd->vpage[seg_page(seg, a)];
9094 } else {
9095 vpage = NULL;
9096 }
9097 ASSERT(anlock);
9098 anon_array_exit(&cookie);
9099 anlock = 0;
9100 pp = NULL;
9101 error = segvn_faultpage(seg->s_as->a_hat, seg, a, 0,
9102 vpage, &pp, 0, F_INVAL, rw, 1);
9103 if (error) {
9104 error = fc_decode(error);
9105 break;
9106 }
9107 anon_array_enter(amp, anon_index, &cookie);
9108 anlock = 1;
9109 ap = anon_get_ptr(amp->ahp, anon_index);
9110 if (ap == NULL || ap->an_refcnt != 1) {
9111 error = EFAULT;
9112 break;
9113 }
9114 }
9115 swap_xlate(ap, &vp, &off);
9116 pp = page_lookup_nowait(vp, off, SE_SHARED);
9117 if (pp == NULL) {
9118 error = EFAULT;
9119 break;
9120 }
9121 if (ap->an_pvp != NULL) {
9122 anon_swap_free(ap, pp);
9123 }
9124 /*
9125 * Unlock anon if this is the last slot in a large page.
9126 */
9127 if (P2PHASE(anon_index, anpgcnt) == anpgcnt - 1) {
9128 ASSERT(anlock);
9129 anon_array_exit(&cookie);
9130 anlock = 0;
9131 }
9132 *pplist++ = pp;
9133 }
9134 if (anlock) { /* Ensure the lock is dropped */
9135 anon_array_exit(&cookie);
9136 }
9137 ANON_LOCK_EXIT(&amp->a_rwlock);
9138
9139 if (a >= addr + len) {
9140 atomic_add_long((ulong_t *)&svd->softlockcnt, npages);
9141 if (pamp != NULL) {
9142 ASSERT(svd->type == MAP_SHARED);
9143 atomic_add_long((ulong_t *)&pamp->a_softlockcnt,
9144 npages);
9145 wlen = len;
9146 }
9147 if (sftlck_sbase) {
9148 atomic_inc_ulong((ulong_t *)&svd->softlockcnt_sbase);
9149 }
9150 if (sftlck_send) {
9151 atomic_inc_ulong((ulong_t *)&svd->softlockcnt_send);
9152 }
9153 if (use_pcache) {
9154 (void) seg_pinsert(seg, pamp, paddr, len, wlen, pl,
9155 rw, pflags, preclaim_callback);
9156 }
9157 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9158 return (0);
9159 }
9160
9161 pplist = pl;
9162 np = ((uintptr_t)(a - addr)) >> PAGESHIFT;
9163 while (np > (uint_t)0) {
9164 ASSERT(PAGE_LOCKED(*pplist));
9165 page_unlock(*pplist);
9166 np--;
9167 pplist++;
9168 }
9169 kmem_free(pl, sizeof (page_t *) * (npages + 1));
9170 out:
9171 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9172 *ppp = NULL;
9173 return (error);
9174 }
9175
9176 /*
9177 * purge any cached pages in the I/O page cache
9178 */
9179 static void
9180 segvn_purge(struct seg *seg)
9181 {
9182 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9183
9184 /*
9185 * pcache is only used by pure anon segments.
9186 */
9187 if (svd->amp == NULL || svd->vp != NULL) {
9188 return;
9189 }
9190
9191 /*
9192 * For MAP_SHARED segments non 0 segment's softlockcnt means
9193 * active IO is still in progress via this segment. So we only
9194 * purge MAP_SHARED segments when their softlockcnt is 0.
9195 */
9196 if (svd->type == MAP_PRIVATE) {
9197 if (svd->softlockcnt) {
9198 seg_ppurge(seg, NULL, 0);
9199 }
9200 } else if (svd->softlockcnt == 0 && svd->amp->a_softlockcnt != 0) {
9201 seg_ppurge(seg, svd->amp, 0);
9202 }
9203 }
9204
9205 /*
9206 * If async argument is not 0 we are called from pcache async thread and don't
9207 * hold AS lock.
9208 */
9209
9210 /*ARGSUSED*/
9211 static int
9212 segvn_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9213 enum seg_rw rw, int async)
9214 {
9215 struct seg *seg = (struct seg *)ptag;
9216 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9217 pgcnt_t np, npages;
9218 struct page **pl;
9219
9220 npages = np = btop(len);
9221 ASSERT(npages);
9222
9223 ASSERT(svd->vp == NULL && svd->amp != NULL);
9224 ASSERT(svd->softlockcnt >= npages);
9225 ASSERT(async || AS_LOCK_HELD(seg->s_as));
9226
9227 pl = pplist;
9228
9229 ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9230 ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9231
9232 while (np > (uint_t)0) {
9233 if (rw == S_WRITE) {
9234 hat_setrefmod(*pplist);
9235 } else {
9236 hat_setref(*pplist);
9237 }
9238 page_unlock(*pplist);
9239 np--;
9240 pplist++;
9241 }
9242
9243 kmem_free(pl, sizeof (page_t *) * (npages + 1));
9244
9245 /*
9246 * If we are pcache async thread we don't hold AS lock. This means if
9247 * softlockcnt drops to 0 after the decrement below address space may
9248 * get freed. We can't allow it since after softlock derement to 0 we
9249 * still need to access as structure for possible wakeup of unmap
9250 * waiters. To prevent the disappearance of as we take this segment
9251 * segfree_syncmtx. segvn_free() also takes this mutex as a barrier to
9252 * make sure this routine completes before segment is freed.
9253 *
9254 * The second complication we have to deal with in async case is a
9255 * possibility of missed wake up of unmap wait thread. When we don't
9256 * hold as lock here we may take a_contents lock before unmap wait
9257 * thread that was first to see softlockcnt was still not 0. As a
9258 * result we'll fail to wake up an unmap wait thread. To avoid this
9259 * race we set nounmapwait flag in as structure if we drop softlockcnt
9260 * to 0 when we were called by pcache async thread. unmapwait thread
9261 * will not block if this flag is set.
9262 */
9263 if (async) {
9264 mutex_enter(&svd->segfree_syncmtx);
9265 }
9266
9267 if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -npages)) {
9268 if (async || AS_ISUNMAPWAIT(seg->s_as)) {
9269 mutex_enter(&seg->s_as->a_contents);
9270 if (async) {
9271 AS_SETNOUNMAPWAIT(seg->s_as);
9272 }
9273 if (AS_ISUNMAPWAIT(seg->s_as)) {
9274 AS_CLRUNMAPWAIT(seg->s_as);
9275 cv_broadcast(&seg->s_as->a_cv);
9276 }
9277 mutex_exit(&seg->s_as->a_contents);
9278 }
9279 }
9280
9281 if (async) {
9282 mutex_exit(&svd->segfree_syncmtx);
9283 }
9284 return (0);
9285 }
9286
9287 /*ARGSUSED*/
9288 static int
9289 shamp_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9290 enum seg_rw rw, int async)
9291 {
9292 amp_t *amp = (amp_t *)ptag;
9293 pgcnt_t np, npages;
9294 struct page **pl;
9295
9296 npages = np = btop(len);
9297 ASSERT(npages);
9298 ASSERT(amp->a_softlockcnt >= npages);
9299
9300 pl = pplist;
9301
9302 ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9303 ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9304
9305 while (np > (uint_t)0) {
9306 if (rw == S_WRITE) {
9307 hat_setrefmod(*pplist);
9308 } else {
9309 hat_setref(*pplist);
9310 }
9311 page_unlock(*pplist);
9312 np--;
9313 pplist++;
9314 }
9315
9316 kmem_free(pl, sizeof (page_t *) * (npages + 1));
9317
9318 /*
9319 * If somebody sleeps in anonmap_purge() wake them up if a_softlockcnt
9320 * drops to 0. anon map can't be freed until a_softlockcnt drops to 0
9321 * and anonmap_purge() acquires a_purgemtx.
9322 */
9323 mutex_enter(&amp->a_purgemtx);
9324 if (!atomic_add_long_nv((ulong_t *)&amp->a_softlockcnt, -npages) &&
9325 amp->a_purgewait) {
9326 amp->a_purgewait = 0;
9327 cv_broadcast(&amp->a_purgecv);
9328 }
9329 mutex_exit(&amp->a_purgemtx);
9330 return (0);
9331 }
9332
9333 /*
9334 * get a memory ID for an addr in a given segment
9335 *
9336 * XXX only creates PAGESIZE pages if anon slots are not initialized.
9337 * At fault time they will be relocated into larger pages.
9338 */
9339 static int
9340 segvn_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
9341 {
9342 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9343 struct anon *ap = NULL;
9344 ulong_t anon_index;
9345 struct anon_map *amp;
9346 anon_sync_obj_t cookie;
9347
9348 if (svd->type == MAP_PRIVATE) {
9349 memidp->val[0] = (uintptr_t)seg->s_as;
9350 memidp->val[1] = (uintptr_t)addr;
9351 return (0);
9352 }
9353
9354 if (svd->type == MAP_SHARED) {
9355 if (svd->vp) {
9356 memidp->val[0] = (uintptr_t)svd->vp;
9357 memidp->val[1] = (u_longlong_t)svd->offset +
9358 (uintptr_t)(addr - seg->s_base);
9359 return (0);
9360 } else {
9361
9362 SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
9363 if ((amp = svd->amp) != NULL) {
9364 anon_index = svd->anon_index +
9365 seg_page(seg, addr);
9366 }
9367 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9368
9369 ASSERT(amp != NULL);
9370
9371 ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9372 anon_array_enter(amp, anon_index, &cookie);
9373 ap = anon_get_ptr(amp->ahp, anon_index);
9374 if (ap == NULL) {
9375 page_t *pp;
9376
9377 pp = anon_zero(seg, addr, &ap, svd->cred);
9378 if (pp == NULL) {
9379 anon_array_exit(&cookie);
9380 ANON_LOCK_EXIT(&amp->a_rwlock);
9381 return (ENOMEM);
9382 }
9383 ASSERT(anon_get_ptr(amp->ahp, anon_index)
9384 == NULL);
9385 (void) anon_set_ptr(amp->ahp, anon_index,
9386 ap, ANON_SLEEP);
9387 page_unlock(pp);
9388 }
9389
9390 anon_array_exit(&cookie);
9391 ANON_LOCK_EXIT(&amp->a_rwlock);
9392
9393 memidp->val[0] = (uintptr_t)ap;
9394 memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
9395 return (0);
9396 }
9397 }
9398 return (EINVAL);
9399 }
9400
9401 static int
9402 sameprot(struct seg *seg, caddr_t a, size_t len)
9403 {
9404 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9405 struct vpage *vpage;
9406 spgcnt_t pages = btop(len);
9407 uint_t prot;
9408
9409 if (svd->pageprot == 0)
9410 return (1);
9411
9412 ASSERT(svd->vpage != NULL);
9413
9414 vpage = &svd->vpage[seg_page(seg, a)];
9415 prot = VPP_PROT(vpage);
9416 vpage++;
9417 pages--;
9418 while (pages-- > 0) {
9419 if (prot != VPP_PROT(vpage))
9420 return (0);
9421 vpage++;
9422 }
9423 return (1);
9424 }
9425
9426 /*
9427 * Get memory allocation policy info for specified address in given segment
9428 */
9429 static lgrp_mem_policy_info_t *
9430 segvn_getpolicy(struct seg *seg, caddr_t addr)
9431 {
9432 struct anon_map *amp;
9433 ulong_t anon_index;
9434 lgrp_mem_policy_info_t *policy_info;
9435 struct segvn_data *svn_data;
9436 uoff_t vn_off;
9437 vnode_t *vp;
9438
9439 ASSERT(seg != NULL);
9440
9441 svn_data = (struct segvn_data *)seg->s_data;
9442 if (svn_data == NULL)
9443 return (NULL);
9444
9445 /*
9446 * Get policy info for private or shared memory
9447 */
9448 if (svn_data->type != MAP_SHARED) {
9449 if (svn_data->tr_state != SEGVN_TR_ON) {
9450 policy_info = &svn_data->policy_info;
9451 } else {
9452 policy_info = &svn_data->tr_policy_info;
9453 ASSERT(policy_info->mem_policy ==
9454 LGRP_MEM_POLICY_NEXT_SEG);
9455 }
9456 } else {
9457 amp = svn_data->amp;
9458 anon_index = svn_data->anon_index + seg_page(seg, addr);
9459 vp = svn_data->vp;
9460 vn_off = svn_data->offset + (uintptr_t)(addr - seg->s_base);
9461 policy_info = lgrp_shm_policy_get(amp, anon_index, vp, vn_off);
9462 }
9463
9464 return (policy_info);
9465 }
9466
9467 /*
9468 * Bind text vnode segment to an amp. If we bind successfully mappings will be
9469 * established to per vnode mapping per lgroup amp pages instead of to vnode
9470 * pages. There's one amp per vnode text mapping per lgroup. Many processes
9471 * may share the same text replication amp. If a suitable amp doesn't already
9472 * exist in svntr hash table create a new one. We may fail to bind to amp if
9473 * segment is not eligible for text replication. Code below first checks for
9474 * these conditions. If binding is successful segment tr_state is set to on
9475 * and svd->amp points to the amp to use. Otherwise tr_state is set to off and
9476 * svd->amp remains as NULL.
9477 */
9478 static void
9479 segvn_textrepl(struct seg *seg)
9480 {
9481 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9482 vnode_t *vp = svd->vp;
9483 uoff_t off = svd->offset;
9484 size_t size = seg->s_size;
9485 uoff_t eoff = off + size;
9486 uint_t szc = seg->s_szc;
9487 ulong_t hash = SVNTR_HASH_FUNC(vp);
9488 svntr_t *svntrp;
9489 struct vattr va;
9490 proc_t *p = seg->s_as->a_proc;
9491 lgrp_id_t lgrp_id;
9492 lgrp_id_t olid;
9493 int first;
9494 struct anon_map *amp;
9495
9496 ASSERT(AS_LOCK_HELD(seg->s_as));
9497 ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9498 ASSERT(p != NULL);
9499 ASSERT(svd->tr_state == SEGVN_TR_INIT);
9500 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9501 ASSERT(svd->flags & MAP_TEXT);
9502 ASSERT(svd->type == MAP_PRIVATE);
9503 ASSERT(vp != NULL && svd->amp == NULL);
9504 ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
9505 ASSERT(!(svd->flags & MAP_NORESERVE) && svd->swresv == 0);
9506 ASSERT(seg->s_as != &kas);
9507 ASSERT(off < eoff);
9508 ASSERT(svntr_hashtab != NULL);
9509
9510 /*
9511 * If numa optimizations are no longer desired bail out.
9512 */
9513 if (!lgrp_optimizations()) {
9514 svd->tr_state = SEGVN_TR_OFF;
9515 return;
9516 }
9517
9518 /*
9519 * Avoid creating anon maps with size bigger than the file size.
9520 * If fop_getattr() call fails bail out.
9521 */
9522 va.va_mask = AT_SIZE | AT_MTIME | AT_CTIME;
9523 if (fop_getattr(vp, &va, 0, svd->cred, NULL) != 0) {
9524 svd->tr_state = SEGVN_TR_OFF;
9525 SEGVN_TR_ADDSTAT(gaerr);
9526 return;
9527 }
9528 if (btopr(va.va_size) < btopr(eoff)) {
9529 svd->tr_state = SEGVN_TR_OFF;
9530 SEGVN_TR_ADDSTAT(overmap);
9531 return;
9532 }
9533
9534 /*
9535 * VVMEXEC may not be set yet if exec() prefaults text segment. Set
9536 * this flag now before vn_is_mapped(V_WRITE) so that MAP_SHARED
9537 * mapping that checks if trcache for this vnode needs to be
9538 * invalidated can't miss us.
9539 */
9540 if (!(vp->v_flag & VVMEXEC)) {
9541 mutex_enter(&vp->v_lock);
9542 vp->v_flag |= VVMEXEC;
9543 mutex_exit(&vp->v_lock);
9544 }
9545 mutex_enter(&svntr_hashtab[hash].tr_lock);
9546 /*
9547 * Bail out if potentially MAP_SHARED writable mappings exist to this
9548 * vnode. We don't want to use old file contents from existing
9549 * replicas if this mapping was established after the original file
9550 * was changed.
9551 */
9552 if (vn_is_mapped(vp, V_WRITE)) {
9553 mutex_exit(&svntr_hashtab[hash].tr_lock);
9554 svd->tr_state = SEGVN_TR_OFF;
9555 SEGVN_TR_ADDSTAT(wrcnt);
9556 return;
9557 }
9558 svntrp = svntr_hashtab[hash].tr_head;
9559 for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9560 ASSERT(svntrp->tr_refcnt != 0);
9561 if (svntrp->tr_vp != vp) {
9562 continue;
9563 }
9564
9565 /*
9566 * Bail out if the file or its attributes were changed after
9567 * this replication entry was created since we need to use the
9568 * latest file contents. Note that mtime test alone is not
9569 * sufficient because a user can explicitly change mtime via
9570 * utimes(2) interfaces back to the old value after modifiying
9571 * the file contents. To detect this case we also have to test
9572 * ctime which among other things records the time of the last
9573 * mtime change by utimes(2). ctime is not changed when the file
9574 * is only read or executed so we expect that typically existing
9575 * replication amp's can be used most of the time.
9576 */
9577 if (!svntrp->tr_valid ||
9578 svntrp->tr_mtime.tv_sec != va.va_mtime.tv_sec ||
9579 svntrp->tr_mtime.tv_nsec != va.va_mtime.tv_nsec ||
9580 svntrp->tr_ctime.tv_sec != va.va_ctime.tv_sec ||
9581 svntrp->tr_ctime.tv_nsec != va.va_ctime.tv_nsec) {
9582 mutex_exit(&svntr_hashtab[hash].tr_lock);
9583 svd->tr_state = SEGVN_TR_OFF;
9584 SEGVN_TR_ADDSTAT(stale);
9585 return;
9586 }
9587 /*
9588 * if off, eoff and szc match current segment we found the
9589 * existing entry we can use.
9590 */
9591 if (svntrp->tr_off == off && svntrp->tr_eoff == eoff &&
9592 svntrp->tr_szc == szc) {
9593 break;
9594 }
9595 /*
9596 * Don't create different but overlapping in file offsets
9597 * entries to avoid replication of the same file pages more
9598 * than once per lgroup.
9599 */
9600 if ((off >= svntrp->tr_off && off < svntrp->tr_eoff) ||
9601 (eoff > svntrp->tr_off && eoff <= svntrp->tr_eoff)) {
9602 mutex_exit(&svntr_hashtab[hash].tr_lock);
9603 svd->tr_state = SEGVN_TR_OFF;
9604 SEGVN_TR_ADDSTAT(overlap);
9605 return;
9606 }
9607 }
9608 /*
9609 * If we didn't find existing entry create a new one.
9610 */
9611 if (svntrp == NULL) {
9612 svntrp = kmem_cache_alloc(svntr_cache, KM_NOSLEEP);
9613 if (svntrp == NULL) {
9614 mutex_exit(&svntr_hashtab[hash].tr_lock);
9615 svd->tr_state = SEGVN_TR_OFF;
9616 SEGVN_TR_ADDSTAT(nokmem);
9617 return;
9618 }
9619 #ifdef DEBUG
9620 {
9621 lgrp_id_t i;
9622 for (i = 0; i < NLGRPS_MAX; i++) {
9623 ASSERT(svntrp->tr_amp[i] == NULL);
9624 }
9625 }
9626 #endif /* DEBUG */
9627 svntrp->tr_vp = vp;
9628 svntrp->tr_off = off;
9629 svntrp->tr_eoff = eoff;
9630 svntrp->tr_szc = szc;
9631 svntrp->tr_valid = 1;
9632 svntrp->tr_mtime = va.va_mtime;
9633 svntrp->tr_ctime = va.va_ctime;
9634 svntrp->tr_refcnt = 0;
9635 svntrp->tr_next = svntr_hashtab[hash].tr_head;
9636 svntr_hashtab[hash].tr_head = svntrp;
9637 }
9638 first = 1;
9639 again:
9640 /*
9641 * We want to pick a replica with pages on main thread's (t_tid = 1,
9642 * aka T1) lgrp. Currently text replication is only optimized for
9643 * workloads that either have all threads of a process on the same
9644 * lgrp or execute their large text primarily on main thread.
9645 */
9646 lgrp_id = p->p_t1_lgrpid;
9647 if (lgrp_id == LGRP_NONE) {
9648 /*
9649 * In case exec() prefaults text on non main thread use
9650 * current thread lgrpid. It will become main thread anyway
9651 * soon.
9652 */
9653 lgrp_id = lgrp_home_id(curthread);
9654 }
9655 /*
9656 * Set p_tr_lgrpid to lgrpid if it hasn't been set yet. Otherwise
9657 * just set it to NLGRPS_MAX if it's different from current process T1
9658 * home lgrp. p_tr_lgrpid is used to detect if process uses text
9659 * replication and T1 new home is different from lgrp used for text
9660 * replication. When this happens asyncronous segvn thread rechecks if
9661 * segments should change lgrps used for text replication. If we fail
9662 * to set p_tr_lgrpid with atomic_cas_32 then set it to NLGRPS_MAX
9663 * without cas if it's not already NLGRPS_MAX and not equal lgrp_id
9664 * we want to use. We don't need to use cas in this case because
9665 * another thread that races in between our non atomic check and set
9666 * may only change p_tr_lgrpid to NLGRPS_MAX at this point.
9667 */
9668 ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9669 olid = p->p_tr_lgrpid;
9670 if (lgrp_id != olid && olid != NLGRPS_MAX) {
9671 lgrp_id_t nlid = (olid == LGRP_NONE) ? lgrp_id : NLGRPS_MAX;
9672 if (atomic_cas_32((uint32_t *)&p->p_tr_lgrpid, olid, nlid) !=
9673 olid) {
9674 olid = p->p_tr_lgrpid;
9675 ASSERT(olid != LGRP_NONE);
9676 if (olid != lgrp_id && olid != NLGRPS_MAX) {
9677 p->p_tr_lgrpid = NLGRPS_MAX;
9678 }
9679 }
9680 ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9681 membar_producer();
9682 /*
9683 * lgrp_move_thread() won't schedule async recheck after
9684 * p->p_t1_lgrpid update unless p->p_tr_lgrpid is not
9685 * LGRP_NONE. Recheck p_t1_lgrpid once now that p->p_tr_lgrpid
9686 * is not LGRP_NONE.
9687 */
9688 if (first && p->p_t1_lgrpid != LGRP_NONE &&
9689 p->p_t1_lgrpid != lgrp_id) {
9690 first = 0;
9691 goto again;
9692 }
9693 }
9694 /*
9695 * If no amp was created yet for lgrp_id create a new one as long as
9696 * we have enough memory to afford it.
9697 */
9698 if ((amp = svntrp->tr_amp[lgrp_id]) == NULL) {
9699 size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9700 if (trmem > segvn_textrepl_max_bytes) {
9701 SEGVN_TR_ADDSTAT(normem);
9702 goto fail;
9703 }
9704 if (anon_try_resv_zone(size, NULL) == 0) {
9705 SEGVN_TR_ADDSTAT(noanon);
9706 goto fail;
9707 }
9708 amp = anonmap_alloc(size, size, ANON_NOSLEEP);
9709 if (amp == NULL) {
9710 anon_unresv_zone(size, NULL);
9711 SEGVN_TR_ADDSTAT(nokmem);
9712 goto fail;
9713 }
9714 ASSERT(amp->refcnt == 1);
9715 amp->a_szc = szc;
9716 svntrp->tr_amp[lgrp_id] = amp;
9717 SEGVN_TR_ADDSTAT(newamp);
9718 }
9719 svntrp->tr_refcnt++;
9720 ASSERT(svd->svn_trnext == NULL);
9721 ASSERT(svd->svn_trprev == NULL);
9722 svd->svn_trnext = svntrp->tr_svnhead;
9723 svd->svn_trprev = NULL;
9724 if (svntrp->tr_svnhead != NULL) {
9725 svntrp->tr_svnhead->svn_trprev = svd;
9726 }
9727 svntrp->tr_svnhead = svd;
9728 ASSERT(amp->a_szc == szc && amp->size == size && amp->swresv == size);
9729 ASSERT(amp->refcnt >= 1);
9730 svd->amp = amp;
9731 svd->anon_index = 0;
9732 svd->tr_policy_info.mem_policy = LGRP_MEM_POLICY_NEXT_SEG;
9733 svd->tr_policy_info.mem_lgrpid = lgrp_id;
9734 svd->tr_state = SEGVN_TR_ON;
9735 mutex_exit(&svntr_hashtab[hash].tr_lock);
9736 SEGVN_TR_ADDSTAT(repl);
9737 return;
9738 fail:
9739 ASSERT(segvn_textrepl_bytes >= size);
9740 atomic_add_long(&segvn_textrepl_bytes, -size);
9741 ASSERT(svntrp != NULL);
9742 ASSERT(svntrp->tr_amp[lgrp_id] == NULL);
9743 if (svntrp->tr_refcnt == 0) {
9744 ASSERT(svntrp == svntr_hashtab[hash].tr_head);
9745 svntr_hashtab[hash].tr_head = svntrp->tr_next;
9746 mutex_exit(&svntr_hashtab[hash].tr_lock);
9747 kmem_cache_free(svntr_cache, svntrp);
9748 } else {
9749 mutex_exit(&svntr_hashtab[hash].tr_lock);
9750 }
9751 svd->tr_state = SEGVN_TR_OFF;
9752 }
9753
9754 /*
9755 * Convert seg back to regular vnode mapping seg by unbinding it from its text
9756 * replication amp. This routine is most typically called when segment is
9757 * unmapped but can also be called when segment no longer qualifies for text
9758 * replication (e.g. due to protection changes). If unload_unmap is set use
9759 * HAT_UNLOAD_UNMAP flag in hat_unload_callback(). If we are the last user of
9760 * svntr free all its anon maps and remove it from the hash table.
9761 */
9762 static void
9763 segvn_textunrepl(struct seg *seg, int unload_unmap)
9764 {
9765 struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9766 vnode_t *vp = svd->vp;
9767 uoff_t off = svd->offset;
9768 size_t size = seg->s_size;
9769 uoff_t eoff = off + size;
9770 uint_t szc = seg->s_szc;
9771 ulong_t hash = SVNTR_HASH_FUNC(vp);
9772 svntr_t *svntrp;
9773 svntr_t **prv_svntrp;
9774 lgrp_id_t lgrp_id = svd->tr_policy_info.mem_lgrpid;
9775 lgrp_id_t i;
9776
9777 ASSERT(AS_LOCK_HELD(seg->s_as));
9778 ASSERT(AS_WRITE_HELD(seg->s_as) ||
9779 SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9780 ASSERT(svd->tr_state == SEGVN_TR_ON);
9781 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9782 ASSERT(svd->amp != NULL);
9783 ASSERT(svd->amp->refcnt >= 1);
9784 ASSERT(svd->anon_index == 0);
9785 ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9786 ASSERT(svntr_hashtab != NULL);
9787
9788 mutex_enter(&svntr_hashtab[hash].tr_lock);
9789 prv_svntrp = &svntr_hashtab[hash].tr_head;
9790 for (; (svntrp = *prv_svntrp) != NULL; prv_svntrp = &svntrp->tr_next) {
9791 ASSERT(svntrp->tr_refcnt != 0);
9792 if (svntrp->tr_vp == vp && svntrp->tr_off == off &&
9793 svntrp->tr_eoff == eoff && svntrp->tr_szc == szc) {
9794 break;
9795 }
9796 }
9797 if (svntrp == NULL) {
9798 panic("segvn_textunrepl: svntr record not found");
9799 }
9800 if (svntrp->tr_amp[lgrp_id] != svd->amp) {
9801 panic("segvn_textunrepl: amp mismatch");
9802 }
9803 svd->tr_state = SEGVN_TR_OFF;
9804 svd->amp = NULL;
9805 if (svd->svn_trprev == NULL) {
9806 ASSERT(svntrp->tr_svnhead == svd);
9807 svntrp->tr_svnhead = svd->svn_trnext;
9808 if (svntrp->tr_svnhead != NULL) {
9809 svntrp->tr_svnhead->svn_trprev = NULL;
9810 }
9811 svd->svn_trnext = NULL;
9812 } else {
9813 svd->svn_trprev->svn_trnext = svd->svn_trnext;
9814 if (svd->svn_trnext != NULL) {
9815 svd->svn_trnext->svn_trprev = svd->svn_trprev;
9816 svd->svn_trnext = NULL;
9817 }
9818 svd->svn_trprev = NULL;
9819 }
9820 if (--svntrp->tr_refcnt) {
9821 mutex_exit(&svntr_hashtab[hash].tr_lock);
9822 goto done;
9823 }
9824 *prv_svntrp = svntrp->tr_next;
9825 mutex_exit(&svntr_hashtab[hash].tr_lock);
9826 for (i = 0; i < NLGRPS_MAX; i++) {
9827 struct anon_map *amp = svntrp->tr_amp[i];
9828 if (amp == NULL) {
9829 continue;
9830 }
9831 ASSERT(amp->refcnt == 1);
9832 ASSERT(amp->swresv == size);
9833 ASSERT(amp->size == size);
9834 ASSERT(amp->a_szc == szc);
9835 if (amp->a_szc != 0) {
9836 anon_free_pages(amp->ahp, 0, size, szc);
9837 } else {
9838 anon_free(amp->ahp, 0, size);
9839 }
9840 svntrp->tr_amp[i] = NULL;
9841 ASSERT(segvn_textrepl_bytes >= size);
9842 atomic_add_long(&segvn_textrepl_bytes, -size);
9843 anon_unresv_zone(amp->swresv, NULL);
9844 amp->refcnt = 0;
9845 anonmap_free(amp);
9846 }
9847 kmem_cache_free(svntr_cache, svntrp);
9848 done:
9849 hat_unload_callback(seg->s_as->a_hat, seg->s_base, size,
9850 unload_unmap ? HAT_UNLOAD_UNMAP : 0, NULL);
9851 }
9852
9853 /*
9854 * This is called when a MAP_SHARED writable mapping is created to a vnode
9855 * that is currently used for execution (VVMEXEC flag is set). In this case we
9856 * need to prevent further use of existing replicas.
9857 */
9858 static void
9859 segvn_inval_trcache(vnode_t *vp)
9860 {
9861 ulong_t hash = SVNTR_HASH_FUNC(vp);
9862 svntr_t *svntrp;
9863
9864 ASSERT(vp->v_flag & VVMEXEC);
9865
9866 if (svntr_hashtab == NULL) {
9867 return;
9868 }
9869
9870 mutex_enter(&svntr_hashtab[hash].tr_lock);
9871 svntrp = svntr_hashtab[hash].tr_head;
9872 for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9873 ASSERT(svntrp->tr_refcnt != 0);
9874 if (svntrp->tr_vp == vp && svntrp->tr_valid) {
9875 svntrp->tr_valid = 0;
9876 }
9877 }
9878 mutex_exit(&svntr_hashtab[hash].tr_lock);
9879 }
9880
9881 static void
9882 segvn_trasync_thread(void)
9883 {
9884 callb_cpr_t cpr_info;
9885 kmutex_t cpr_lock; /* just for CPR stuff */
9886
9887 mutex_init(&cpr_lock, NULL, MUTEX_DEFAULT, NULL);
9888
9889 CALLB_CPR_INIT(&cpr_info, &cpr_lock,
9890 callb_generic_cpr, "segvn_async");
9891
9892 if (segvn_update_textrepl_interval == 0) {
9893 segvn_update_textrepl_interval = segvn_update_tr_time * hz;
9894 } else {
9895 segvn_update_textrepl_interval *= hz;
9896 }
9897 (void) timeout(segvn_trupdate_wakeup, NULL,
9898 segvn_update_textrepl_interval);
9899
9900 for (;;) {
9901 mutex_enter(&cpr_lock);
9902 CALLB_CPR_SAFE_BEGIN(&cpr_info);
9903 mutex_exit(&cpr_lock);
9904 sema_p(&segvn_trasync_sem);
9905 mutex_enter(&cpr_lock);
9906 CALLB_CPR_SAFE_END(&cpr_info, &cpr_lock);
9907 mutex_exit(&cpr_lock);
9908 segvn_trupdate();
9909 }
9910 }
9911
9912 static uint64_t segvn_lgrp_trthr_migrs_snpsht = 0;
9913
9914 static void
9915 segvn_trupdate_wakeup(void *dummy)
9916 {
9917 uint64_t cur_lgrp_trthr_migrs = lgrp_get_trthr_migrations();
9918
9919 if (cur_lgrp_trthr_migrs != segvn_lgrp_trthr_migrs_snpsht) {
9920 segvn_lgrp_trthr_migrs_snpsht = cur_lgrp_trthr_migrs;
9921 sema_v(&segvn_trasync_sem);
9922 }
9923
9924 if (!segvn_disable_textrepl_update &&
9925 segvn_update_textrepl_interval != 0) {
9926 (void) timeout(segvn_trupdate_wakeup, dummy,
9927 segvn_update_textrepl_interval);
9928 }
9929 }
9930
9931 static void
9932 segvn_trupdate(void)
9933 {
9934 ulong_t hash;
9935 svntr_t *svntrp;
9936 segvn_data_t *svd;
9937
9938 ASSERT(svntr_hashtab != NULL);
9939
9940 for (hash = 0; hash < svntr_hashtab_sz; hash++) {
9941 mutex_enter(&svntr_hashtab[hash].tr_lock);
9942 svntrp = svntr_hashtab[hash].tr_head;
9943 for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9944 ASSERT(svntrp->tr_refcnt != 0);
9945 svd = svntrp->tr_svnhead;
9946 for (; svd != NULL; svd = svd->svn_trnext) {
9947 segvn_trupdate_seg(svd->seg, svd, svntrp,
9948 hash);
9949 }
9950 }
9951 mutex_exit(&svntr_hashtab[hash].tr_lock);
9952 }
9953 }
9954
9955 static void
9956 segvn_trupdate_seg(struct seg *seg,
9957 segvn_data_t *svd,
9958 svntr_t *svntrp,
9959 ulong_t hash)
9960 {
9961 proc_t *p;
9962 lgrp_id_t lgrp_id;
9963 struct as *as;
9964 size_t size;
9965 struct anon_map *amp;
9966
9967 ASSERT(svd->vp != NULL);
9968 ASSERT(svd->vp == svntrp->tr_vp);
9969 ASSERT(svd->offset == svntrp->tr_off);
9970 ASSERT(svd->offset + seg->s_size == svntrp->tr_eoff);
9971 ASSERT(seg != NULL);
9972 ASSERT(svd->seg == seg);
9973 ASSERT(seg->s_data == (void *)svd);
9974 ASSERT(seg->s_szc == svntrp->tr_szc);
9975 ASSERT(svd->tr_state == SEGVN_TR_ON);
9976 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9977 ASSERT(svd->amp != NULL);
9978 ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
9979 ASSERT(svd->tr_policy_info.mem_lgrpid != LGRP_NONE);
9980 ASSERT(svd->tr_policy_info.mem_lgrpid < NLGRPS_MAX);
9981 ASSERT(svntrp->tr_amp[svd->tr_policy_info.mem_lgrpid] == svd->amp);
9982 ASSERT(svntrp->tr_refcnt != 0);
9983 ASSERT(mutex_owned(&svntr_hashtab[hash].tr_lock));
9984
9985 as = seg->s_as;
9986 ASSERT(as != NULL && as != &kas);
9987 p = as->a_proc;
9988 ASSERT(p != NULL);
9989 ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9990 lgrp_id = p->p_t1_lgrpid;
9991 if (lgrp_id == LGRP_NONE) {
9992 return;
9993 }
9994 ASSERT(lgrp_id < NLGRPS_MAX);
9995 if (svd->tr_policy_info.mem_lgrpid == lgrp_id) {
9996 return;
9997 }
9998
9999 /*
10000 * Use tryenter locking since we are locking as/seg and svntr hash
10001 * lock in reverse from syncrounous thread order.
10002 */
10003 if (!AS_LOCK_TRYENTER(as, RW_READER)) {
10004 SEGVN_TR_ADDSTAT(nolock);
10005 if (segvn_lgrp_trthr_migrs_snpsht) {
10006 segvn_lgrp_trthr_migrs_snpsht = 0;
10007 }
10008 return;
10009 }
10010 if (!SEGVN_LOCK_TRYENTER(seg->s_as, &svd->lock, RW_WRITER)) {
10011 AS_LOCK_EXIT(as);
10012 SEGVN_TR_ADDSTAT(nolock);
10013 if (segvn_lgrp_trthr_migrs_snpsht) {
10014 segvn_lgrp_trthr_migrs_snpsht = 0;
10015 }
10016 return;
10017 }
10018 size = seg->s_size;
10019 if (svntrp->tr_amp[lgrp_id] == NULL) {
10020 size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
10021 if (trmem > segvn_textrepl_max_bytes) {
10022 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10023 AS_LOCK_EXIT(as);
10024 atomic_add_long(&segvn_textrepl_bytes, -size);
10025 SEGVN_TR_ADDSTAT(normem);
10026 return;
10027 }
10028 if (anon_try_resv_zone(size, NULL) == 0) {
10029 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10030 AS_LOCK_EXIT(as);
10031 atomic_add_long(&segvn_textrepl_bytes, -size);
10032 SEGVN_TR_ADDSTAT(noanon);
10033 return;
10034 }
10035 amp = anonmap_alloc(size, size, KM_NOSLEEP);
10036 if (amp == NULL) {
10037 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10038 AS_LOCK_EXIT(as);
10039 atomic_add_long(&segvn_textrepl_bytes, -size);
10040 anon_unresv_zone(size, NULL);
10041 SEGVN_TR_ADDSTAT(nokmem);
10042 return;
10043 }
10044 ASSERT(amp->refcnt == 1);
10045 amp->a_szc = seg->s_szc;
10046 svntrp->tr_amp[lgrp_id] = amp;
10047 }
10048 /*
10049 * We don't need to drop the bucket lock but here we give other
10050 * threads a chance. svntr and svd can't be unlinked as long as
10051 * segment lock is held as a writer and AS held as well. After we
10052 * retake bucket lock we'll continue from where we left. We'll be able
10053 * to reach the end of either list since new entries are always added
10054 * to the beginning of the lists.
10055 */
10056 mutex_exit(&svntr_hashtab[hash].tr_lock);
10057 hat_unload_callback(as->a_hat, seg->s_base, size, 0, NULL);
10058 mutex_enter(&svntr_hashtab[hash].tr_lock);
10059
10060 ASSERT(svd->tr_state == SEGVN_TR_ON);
10061 ASSERT(svd->amp != NULL);
10062 ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10063 ASSERT(svd->tr_policy_info.mem_lgrpid != lgrp_id);
10064 ASSERT(svd->amp != svntrp->tr_amp[lgrp_id]);
10065
10066 svd->tr_policy_info.mem_lgrpid = lgrp_id;
10067 svd->amp = svntrp->tr_amp[lgrp_id];
10068 p->p_tr_lgrpid = NLGRPS_MAX;
10069 SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10070 AS_LOCK_EXIT(as);
10071
10072 ASSERT(svntrp->tr_refcnt != 0);
10073 ASSERT(svd->vp == svntrp->tr_vp);
10074 ASSERT(svd->tr_policy_info.mem_lgrpid == lgrp_id);
10075 ASSERT(svd->amp != NULL && svd->amp == svntrp->tr_amp[lgrp_id]);
10076 ASSERT(svd->seg == seg);
10077 ASSERT(svd->tr_state == SEGVN_TR_ON);
10078
10079 SEGVN_TR_ADDSTAT(asyncrepl);
10080 }
Unleashed OS