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.
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.
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]
22 * Copyright 2009 Sun Microsystems, Inc. All rights reserved.
23 * Use is subject to license terms.
26 /* Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T */
27 /* All Rights Reserved */
30 * University Copyright- Copyright (c) 1982, 1986, 1988
31 * The Regents of the University of California
34 * University Acknowledgment- Portions of this document are derived from
35 * software developed by the University of California, Berkeley, and its
40 * VM - shared or copy-on-write from a vnode/anonymous memory.
43 #include <sys/types.h>
44 #include <sys/param.h>
45 #include <sys/t_lock.h>
46 #include <sys/errno.h>
47 #include <sys/systm.h>
49 #include <sys/debug.h>
51 #include <sys/vmsystm.h>
52 #include <sys/tuneable.h>
53 #include <sys/bitmap.h>
56 #include <sys/sysmacros.h>
57 #include <sys/vtrace.h>
58 #include <sys/cmn_err.h>
59 #include <sys/callb.h>
61 #include <sys/dumphdr.h>
67 #include <vm/seg_vn.h>
74 #include <sys/project.h>
76 #include <sys/shm_impl.h>
78 * Private seg op routines.
80 static int segvn_dup(struct seg
*seg
, struct seg
*newseg
);
81 static int segvn_unmap(struct seg
*seg
, caddr_t addr
, size_t len
);
82 static void segvn_free(struct seg
*seg
);
83 static faultcode_t
segvn_fault(struct hat
*hat
, struct seg
*seg
,
84 caddr_t addr
, size_t len
, enum fault_type type
,
86 static faultcode_t
segvn_faulta(struct seg
*seg
, caddr_t addr
);
87 static int segvn_setprot(struct seg
*seg
, caddr_t addr
,
88 size_t len
, uint_t prot
);
89 static int segvn_checkprot(struct seg
*seg
, caddr_t addr
,
90 size_t len
, uint_t prot
);
91 static int segvn_kluster(struct seg
*seg
, caddr_t addr
, ssize_t delta
);
92 static size_t segvn_swapout(struct seg
*seg
);
93 static int segvn_sync(struct seg
*seg
, caddr_t addr
, size_t len
,
94 int attr
, uint_t flags
);
95 static size_t segvn_incore(struct seg
*seg
, caddr_t addr
, size_t len
,
97 static int segvn_lockop(struct seg
*seg
, caddr_t addr
, size_t len
,
98 int attr
, int op
, ulong_t
*lockmap
, size_t pos
);
99 static int segvn_getprot(struct seg
*seg
, caddr_t addr
, size_t len
,
101 static u_offset_t
segvn_getoffset(struct seg
*seg
, caddr_t addr
);
102 static int segvn_gettype(struct seg
*seg
, caddr_t addr
);
103 static int segvn_getvp(struct seg
*seg
, caddr_t addr
,
105 static int segvn_advise(struct seg
*seg
, caddr_t addr
, size_t len
,
107 static void segvn_dump(struct seg
*seg
);
108 static int segvn_pagelock(struct seg
*seg
, caddr_t addr
, size_t len
,
109 struct page
***ppp
, enum lock_type type
, enum seg_rw rw
);
110 static int segvn_setpagesize(struct seg
*seg
, caddr_t addr
, size_t len
,
112 static int segvn_getmemid(struct seg
*seg
, caddr_t addr
,
114 static lgrp_mem_policy_info_t
*segvn_getpolicy(struct seg
*, caddr_t
);
115 static int segvn_capable(struct seg
*seg
, segcapability_t capable
);
117 struct seg_ops segvn_ops
= {
144 * Common zfod structures, provided as a shorthand for others to use.
146 static segvn_crargs_t zfod_segvn_crargs
=
147 SEGVN_ZFOD_ARGS(PROT_ZFOD
, PROT_ALL
);
148 static segvn_crargs_t kzfod_segvn_crargs
=
149 SEGVN_ZFOD_ARGS(PROT_ZFOD
& ~PROT_USER
,
150 PROT_ALL
& ~PROT_USER
);
151 static segvn_crargs_t stack_noexec_crargs
=
152 SEGVN_ZFOD_ARGS(PROT_ZFOD
& ~PROT_EXEC
, PROT_ALL
);
154 caddr_t zfod_argsp
= (caddr_t
)&zfod_segvn_crargs
; /* user zfod argsp */
155 caddr_t kzfod_argsp
= (caddr_t
)&kzfod_segvn_crargs
; /* kernel zfod argsp */
156 caddr_t stack_exec_argsp
= (caddr_t
)&zfod_segvn_crargs
; /* executable stack */
157 caddr_t stack_noexec_argsp
= (caddr_t
)&stack_noexec_crargs
; /* noexec stack */
159 #define vpgtob(n) ((n) * sizeof (struct vpage)) /* For brevity */
161 size_t segvn_comb_thrshld
= UINT_MAX
; /* patchable -- see 1196681 */
163 size_t segvn_pglock_comb_thrshld
= (1UL << 16); /* 64K */
164 size_t segvn_pglock_comb_balign
= (1UL << 16); /* 64K */
165 uint_t segvn_pglock_comb_bshift
;
166 size_t segvn_pglock_comb_palign
;
168 static int segvn_concat(struct seg
*, struct seg
*, int);
169 static int segvn_extend_prev(struct seg
*, struct seg
*,
170 struct segvn_crargs
*, size_t);
171 static int segvn_extend_next(struct seg
*, struct seg
*,
172 struct segvn_crargs
*, size_t);
173 static void segvn_softunlock(struct seg
*, caddr_t
, size_t, enum seg_rw
);
174 static void segvn_pagelist_rele(page_t
**);
175 static void segvn_setvnode_mpss(vnode_t
*);
176 static void segvn_relocate_pages(page_t
**, page_t
*);
177 static int segvn_full_szcpages(page_t
**, uint_t
, int *, uint_t
*);
178 static int segvn_fill_vp_pages(struct segvn_data
*, vnode_t
*, u_offset_t
,
179 uint_t
, page_t
**, page_t
**, uint_t
*, int *);
180 static faultcode_t
segvn_fault_vnodepages(struct hat
*, struct seg
*, caddr_t
,
181 caddr_t
, enum fault_type
, enum seg_rw
, caddr_t
, caddr_t
, int);
182 static faultcode_t
segvn_fault_anonpages(struct hat
*, struct seg
*, caddr_t
,
183 caddr_t
, enum fault_type
, enum seg_rw
, caddr_t
, caddr_t
, int);
184 static faultcode_t
segvn_faultpage(struct hat
*, struct seg
*, caddr_t
,
185 u_offset_t
, struct vpage
*, page_t
**, uint_t
,
186 enum fault_type
, enum seg_rw
, int);
187 static void segvn_vpage(struct seg
*);
188 static size_t segvn_count_swap_by_vpages(struct seg
*);
190 static void segvn_purge(struct seg
*seg
);
191 static int segvn_reclaim(void *, caddr_t
, size_t, struct page
**,
193 static int shamp_reclaim(void *, caddr_t
, size_t, struct page
**,
196 static int sameprot(struct seg
*, caddr_t
, size_t);
198 static int segvn_demote_range(struct seg
*, caddr_t
, size_t, int, uint_t
);
199 static int segvn_clrszc(struct seg
*);
200 static struct seg
*segvn_split_seg(struct seg
*, caddr_t
);
201 static int segvn_claim_pages(struct seg
*, struct vpage
*, u_offset_t
,
204 static void segvn_hat_rgn_unload_callback(caddr_t
, caddr_t
, caddr_t
,
205 size_t, void *, u_offset_t
);
207 static struct kmem_cache
*segvn_cache
;
208 static struct kmem_cache
**segvn_szc_cache
;
211 static struct segvnvmstats_str
{
212 ulong_t fill_vp_pages
[31];
213 ulong_t fltvnpages
[49];
214 ulong_t fullszcpages
[10];
215 ulong_t relocatepages
[3];
216 ulong_t fltanpages
[17];
218 ulong_t demoterange
[3];
220 #endif /* VM_STATS */
222 #define SDR_RANGE 1 /* demote entire range */
223 #define SDR_END 2 /* demote non aligned ends only */
225 #define CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr) { \
227 lpgaddr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz); \
228 ASSERT(lpgaddr >= (seg)->s_base); \
229 lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)((addr) + \
231 ASSERT(lpgeaddr > lpgaddr); \
232 ASSERT(lpgeaddr <= (seg)->s_base + (seg)->s_size); \
234 lpgeaddr = lpgaddr = (addr); \
240 segvn_cache_constructor(void *buf
, void *cdrarg
, int kmflags
)
242 struct segvn_data
*svd
= buf
;
244 rw_init(&svd
->lock
, NULL
, RW_DEFAULT
, NULL
);
245 mutex_init(&svd
->segfree_syncmtx
, NULL
, MUTEX_DEFAULT
, NULL
);
246 svd
->svn_trnext
= svd
->svn_trprev
= NULL
;
252 segvn_cache_destructor(void *buf
, void *cdrarg
)
254 struct segvn_data
*svd
= buf
;
256 rw_destroy(&svd
->lock
);
257 mutex_destroy(&svd
->segfree_syncmtx
);
262 svntr_cache_constructor(void *buf
, void *cdrarg
, int kmflags
)
264 bzero(buf
, sizeof (svntr_t
));
269 * Patching this variable to non-zero allows the system to run with
270 * stacks marked as "not executable". It's a bit of a kludge, but is
271 * provided as a tweakable for platforms that export those ABIs
272 * (e.g. sparc V8) that have executable stacks enabled by default.
273 * There are also some restrictions for platforms that don't actually
274 * implement 'noexec' protections.
276 * Once enabled, the system is (therefore) unable to provide a fully
277 * ABI-compliant execution environment, though practically speaking,
278 * most everything works. The exceptions are generally some interpreters
279 * and debuggers that create executable code on the stack and jump
280 * into it (without explicitly mprotecting the address range to include
283 * One important class of applications that are disabled are those
284 * that have been transformed into malicious agents using one of the
285 * numerous "buffer overflow" attacks. See 4007890.
287 int noexec_user_stack
= 0;
288 int noexec_user_stack_log
= 1;
290 int segvn_lpg_disable
= 0;
291 uint_t segvn_maxpgszc
= 0;
293 ulong_t segvn_vmpss_clrszc_cnt
;
294 ulong_t segvn_vmpss_clrszc_err
;
295 ulong_t segvn_fltvnpages_clrszc_cnt
;
296 ulong_t segvn_fltvnpages_clrszc_err
;
297 ulong_t segvn_setpgsz_align_err
;
298 ulong_t segvn_setpgsz_anon_align_err
;
299 ulong_t segvn_setpgsz_getattr_err
;
300 ulong_t segvn_setpgsz_eof_err
;
301 ulong_t segvn_faultvnmpss_align_err1
;
302 ulong_t segvn_faultvnmpss_align_err2
;
303 ulong_t segvn_faultvnmpss_align_err3
;
304 ulong_t segvn_faultvnmpss_align_err4
;
305 ulong_t segvn_faultvnmpss_align_err5
;
306 ulong_t segvn_vmpss_pageio_deadlk_err
;
308 int segvn_use_regions
= 1;
311 * Segvn supports text replication optimization for NUMA platforms. Text
312 * replica's are represented by anon maps (amp). There's one amp per text file
313 * region per lgroup. A process chooses the amp for each of its text mappings
314 * based on the lgroup assignment of its main thread (t_tid = 1). All
315 * processes that want a replica on a particular lgroup for the same text file
316 * mapping share the same amp. amp's are looked up in svntr_hashtab hash table
317 * with vp,off,size,szc used as a key. Text replication segments are read only
318 * MAP_PRIVATE|MAP_TEXT segments that map vnode. Replication is achieved by
319 * forcing COW faults from vnode to amp and mapping amp pages instead of vnode
320 * pages. Replication amp is assigned to a segment when it gets its first
321 * pagefault. To handle main thread lgroup rehoming segvn_trasync_thread
322 * rechecks periodically if the process still maps an amp local to the main
323 * thread. If not async thread forces process to remap to an amp in the new
324 * home lgroup of the main thread. Current text replication implementation
325 * only provides the benefit to workloads that do most of their work in the
326 * main thread of a process or all the threads of a process run in the same
327 * lgroup. To extend text replication benefit to different types of
328 * multithreaded workloads further work would be needed in the hat layer to
329 * allow the same virtual address in the same hat to simultaneously map
330 * different physical addresses (i.e. page table replication would be needed
333 * amp pages are used instead of vnode pages as long as segment has a very
334 * simple life cycle. It's created via segvn_create(), handles S_EXEC
335 * (S_READ) pagefaults and is fully unmapped. If anything more complicated
336 * happens such as protection is changed, real COW fault happens, pagesize is
337 * changed, MC_LOCK is requested or segment is partially unmapped we turn off
338 * text replication by converting the segment back to vnode only segment
339 * (unmap segment's address range and set svd->amp to NULL).
341 * The original file can be changed after amp is inserted into
342 * svntr_hashtab. Processes that are launched after the file is already
343 * changed can't use the replica's created prior to the file change. To
344 * implement this functionality hash entries are timestamped. Replica's can
345 * only be used if current file modification time is the same as the timestamp
346 * saved when hash entry was created. However just timestamps alone are not
347 * sufficient to detect file modification via mmap(MAP_SHARED) mappings. We
348 * deal with file changes via MAP_SHARED mappings differently. When writable
349 * MAP_SHARED mappings are created to vnodes marked as executable we mark all
350 * existing replica's for this vnode as not usable for future text
351 * mappings. And we don't create new replica's for files that currently have
352 * potentially writable MAP_SHARED mappings (i.e. vn_is_mapped(V_WRITE) is
356 #define SEGVN_TEXTREPL_MAXBYTES_FACTOR (20)
357 size_t segvn_textrepl_max_bytes_factor
= SEGVN_TEXTREPL_MAXBYTES_FACTOR
;
359 static ulong_t svntr_hashtab_sz
= 512;
360 static svntr_bucket_t
*svntr_hashtab
= NULL
;
361 static struct kmem_cache
*svntr_cache
;
362 static svntr_stats_t
*segvn_textrepl_stats
;
363 static ksema_t segvn_trasync_sem
;
365 int segvn_disable_textrepl
= 1;
366 size_t textrepl_size_thresh
= (size_t)-1;
367 size_t segvn_textrepl_bytes
= 0;
368 size_t segvn_textrepl_max_bytes
= 0;
369 clock_t segvn_update_textrepl_interval
= 0;
370 int segvn_update_tr_time
= 10;
371 int segvn_disable_textrepl_update
= 0;
373 static void segvn_textrepl(struct seg
*);
374 static void segvn_textunrepl(struct seg
*, int);
375 static void segvn_inval_trcache(vnode_t
*);
376 static void segvn_trasync_thread(void);
377 static void segvn_trupdate_wakeup(void *);
378 static void segvn_trupdate(void);
379 static void segvn_trupdate_seg(struct seg
*, segvn_data_t
*, svntr_t
*,
383 * Initialize segvn data structures
392 segvn_cache
= kmem_cache_create("segvn_cache",
393 sizeof (struct segvn_data
), 0,
394 segvn_cache_constructor
, segvn_cache_destructor
, NULL
,
397 if (segvn_lpg_disable
== 0) {
398 szc
= maxszc
= page_num_pagesizes() - 1;
400 segvn_lpg_disable
= 1;
402 if (page_get_pagesize(0) != PAGESIZE
) {
403 panic("segvn_init: bad szc 0");
407 pgsz
= page_get_pagesize(szc
);
408 if (pgsz
<= PAGESIZE
|| !IS_P2ALIGNED(pgsz
, pgsz
)) {
409 panic("segvn_init: bad szc %d", szc
);
414 if (segvn_maxpgszc
== 0 || segvn_maxpgszc
> maxszc
)
415 segvn_maxpgszc
= maxszc
;
418 if (segvn_maxpgszc
) {
419 segvn_szc_cache
= (struct kmem_cache
**)kmem_alloc(
420 (segvn_maxpgszc
+ 1) * sizeof (struct kmem_cache
*),
424 for (szc
= 1; szc
<= segvn_maxpgszc
; szc
++) {
427 (void) sprintf(str
, "segvn_szc_cache%d", szc
);
428 segvn_szc_cache
[szc
] = kmem_cache_create(str
,
429 page_get_pagecnt(szc
) * sizeof (page_t
*), 0,
430 NULL
, NULL
, NULL
, NULL
, NULL
, KMC_NODEBUG
);
434 if (segvn_use_regions
&& !hat_supported(HAT_SHARED_REGIONS
, NULL
))
435 segvn_use_regions
= 0;
438 * For now shared regions and text replication segvn support
439 * are mutually exclusive. This is acceptable because
440 * currently significant benefit from text replication was
441 * only observed on AMD64 NUMA platforms (due to relatively
442 * small L2$ size) and currently we don't support shared
445 if (segvn_use_regions
&& !segvn_disable_textrepl
) {
446 segvn_disable_textrepl
= 1;
450 if (lgrp_optimizations() && textrepl_size_thresh
!= (size_t)-1 &&
451 !segvn_disable_textrepl
) {
453 size_t hsz
= svntr_hashtab_sz
* sizeof (svntr_bucket_t
);
455 svntr_cache
= kmem_cache_create("svntr_cache",
456 sizeof (svntr_t
), 0, svntr_cache_constructor
, NULL
,
457 NULL
, NULL
, NULL
, 0);
458 svntr_hashtab
= kmem_zalloc(hsz
, KM_SLEEP
);
459 for (i
= 0; i
< svntr_hashtab_sz
; i
++) {
460 mutex_init(&svntr_hashtab
[i
].tr_lock
, NULL
,
461 MUTEX_DEFAULT
, NULL
);
463 segvn_textrepl_max_bytes
= ptob(physmem
) /
464 segvn_textrepl_max_bytes_factor
;
465 segvn_textrepl_stats
= kmem_zalloc(NCPU
*
466 sizeof (svntr_stats_t
), KM_SLEEP
);
467 sema_init(&segvn_trasync_sem
, 0, NULL
, SEMA_DEFAULT
, NULL
);
468 (void) thread_create(NULL
, 0, segvn_trasync_thread
,
469 NULL
, 0, &p0
, TS_RUN
, minclsyspri
);
473 if (!ISP2(segvn_pglock_comb_balign
) ||
474 segvn_pglock_comb_balign
< PAGESIZE
) {
475 segvn_pglock_comb_balign
= 1UL << 16; /* 64K */
477 segvn_pglock_comb_bshift
= highbit(segvn_pglock_comb_balign
) - 1;
478 segvn_pglock_comb_palign
= btop(segvn_pglock_comb_balign
);
481 #define SEGVN_PAGEIO ((void *)0x1)
482 #define SEGVN_NOPAGEIO ((void *)0x2)
485 segvn_setvnode_mpss(vnode_t
*vp
)
489 ASSERT(vp
->v_mpssdata
== NULL
||
490 vp
->v_mpssdata
== SEGVN_PAGEIO
||
491 vp
->v_mpssdata
== SEGVN_NOPAGEIO
);
493 if (vp
->v_mpssdata
== NULL
) {
494 if (vn_vmpss_usepageio(vp
)) {
495 err
= VOP_PAGEIO(vp
, (page_t
*)NULL
,
496 (u_offset_t
)0, 0, 0, CRED(), NULL
);
501 * set v_mpssdata just once per vnode life
502 * so that it never changes.
504 mutex_enter(&vp
->v_lock
);
505 if (vp
->v_mpssdata
== NULL
) {
507 vp
->v_mpssdata
= SEGVN_PAGEIO
;
509 vp
->v_mpssdata
= SEGVN_NOPAGEIO
;
512 mutex_exit(&vp
->v_lock
);
517 segvn_create(struct seg
*seg
, void *argsp
)
519 struct segvn_crargs
*a
= (struct segvn_crargs
*)argsp
;
520 struct segvn_data
*svd
;
523 struct anon_map
*amp
;
526 lgrp_mem_policy_t mpolicy
= LGRP_MEM_POLICY_DEFAULT
;
530 ASSERT(seg
->s_as
&& AS_WRITE_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
532 if (a
->type
!= MAP_PRIVATE
&& a
->type
!= MAP_SHARED
) {
533 panic("segvn_create type");
538 * Check arguments. If a shared anon structure is given then
539 * it is illegal to also specify a vp.
541 if (a
->amp
!= NULL
&& a
->vp
!= NULL
) {
542 panic("segvn_create anon_map");
546 if (a
->type
== MAP_PRIVATE
&& (a
->flags
& MAP_TEXT
) &&
547 a
->vp
!= NULL
&& a
->prot
== (PROT_USER
| PROT_READ
| PROT_EXEC
) &&
552 /* MAP_NORESERVE on a MAP_SHARED segment is meaningless. */
553 if (a
->type
== MAP_SHARED
)
554 a
->flags
&= ~MAP_NORESERVE
;
557 if (segvn_lpg_disable
!= 0 || (a
->szc
== AS_MAP_NO_LPOOB
) ||
558 (a
->amp
!= NULL
&& a
->type
== MAP_PRIVATE
) ||
559 (a
->flags
& MAP_NORESERVE
) || seg
->s_as
== &kas
) {
562 if (a
->szc
> segvn_maxpgszc
)
563 a
->szc
= segvn_maxpgszc
;
564 pgsz
= page_get_pagesize(a
->szc
);
565 if (!IS_P2ALIGNED(seg
->s_base
, pgsz
) ||
566 !IS_P2ALIGNED(seg
->s_size
, pgsz
)) {
568 } else if (a
->vp
!= NULL
) {
569 extern struct vnode kvp
;
570 if (IS_SWAPFSVP(a
->vp
) || VN_ISKAS(a
->vp
)) {
573 * hat_page_demote() is not supported
577 } else if (map_addr_vacalign_check(seg
->s_base
,
578 a
->offset
& PAGEMASK
)) {
581 } else if (a
->amp
!= NULL
) {
582 pgcnt_t anum
= btopr(a
->offset
);
583 pgcnt_t pgcnt
= page_get_pagecnt(a
->szc
);
584 if (!IS_P2ALIGNED(anum
, pgcnt
)) {
592 * If segment may need private pages, reserve them now.
594 if (!(a
->flags
& MAP_NORESERVE
) && ((a
->vp
== NULL
&& a
->amp
== NULL
) ||
595 (a
->type
== MAP_PRIVATE
&& (a
->prot
& PROT_WRITE
)))) {
596 if (anon_resv_zone(seg
->s_size
,
597 seg
->s_as
->a_proc
->p_zone
) == 0)
599 swresv
= seg
->s_size
;
600 TRACE_3(TR_FAC_VM
, TR_ANON_PROC
, "anon proc:%p %lu %u",
605 * Reserve any mapping structures that may be required.
607 * Don't do it for segments that may use regions. It's currently a
608 * noop in the hat implementations anyway.
611 hat_map(seg
->s_as
->a_hat
, seg
->s_base
, seg
->s_size
, HAT_MAP
);
618 crhold(cred
= CRED());
621 /* Inform the vnode of the new mapping */
623 error
= VOP_ADDMAP(a
->vp
, a
->offset
& PAGEMASK
,
624 seg
->s_as
, seg
->s_base
, seg
->s_size
, a
->prot
,
625 a
->maxprot
, a
->type
, cred
, NULL
);
628 anon_unresv_zone(swresv
,
629 seg
->s_as
->a_proc
->p_zone
);
630 TRACE_3(TR_FAC_VM
, TR_ANON_PROC
,
631 "anon proc:%p %lu %u", seg
, swresv
, 0);
635 hat_unload(seg
->s_as
->a_hat
, seg
->s_base
,
636 seg
->s_size
, HAT_UNLOAD_UNMAP
);
641 * svntr_hashtab will be NULL if we support shared regions.
643 trok
= ((a
->flags
& MAP_TEXT
) &&
644 (seg
->s_size
> textrepl_size_thresh
||
645 (a
->flags
& _MAP_TEXTREPL
)) &&
646 lgrp_optimizations() && svntr_hashtab
!= NULL
&&
647 a
->type
== MAP_PRIVATE
&& swresv
== 0 &&
648 !(a
->flags
& MAP_NORESERVE
) &&
649 seg
->s_as
!= &kas
&& a
->vp
->v_type
== VREG
);
651 ASSERT(!trok
|| !use_rgn
);
655 * If more than one segment in the address space, and they're adjacent
656 * virtually, try to concatenate them. Don't concatenate if an
657 * explicit anon_map structure was supplied (e.g., SystemV shared
658 * memory) or if we'll use text replication for this segment.
660 if (a
->amp
== NULL
&& !use_rgn
&& !trok
) {
661 struct seg
*pseg
, *nseg
;
662 struct segvn_data
*psvd
, *nsvd
;
663 lgrp_mem_policy_t ppolicy
, npolicy
;
664 uint_t lgrp_mem_policy_flags
= 0;
665 extern lgrp_mem_policy_t lgrp_mem_default_policy
;
668 * Memory policy flags (lgrp_mem_policy_flags) is valid when
669 * extending stack/heap segments.
671 if ((a
->vp
== NULL
) && (a
->type
== MAP_PRIVATE
) &&
672 !(a
->flags
& MAP_NORESERVE
) && (seg
->s_as
!= &kas
)) {
673 lgrp_mem_policy_flags
= a
->lgrp_mem_policy_flags
;
676 * Get policy when not extending it from another segment
678 mpolicy
= lgrp_mem_policy_default(seg
->s_size
, a
->type
);
682 * First, try to concatenate the previous and new segments
684 pseg
= AS_SEGPREV(seg
->s_as
, seg
);
686 pseg
->s_base
+ pseg
->s_size
== seg
->s_base
&&
687 pseg
->s_ops
== &segvn_ops
) {
689 * Get memory allocation policy from previous segment.
690 * When extension is specified (e.g. for heap) apply
691 * this policy to the new segment regardless of the
692 * outcome of segment concatenation. Extension occurs
693 * for non-default policy otherwise default policy is
694 * used and is based on extended segment size.
696 psvd
= (struct segvn_data
*)pseg
->s_data
;
697 ppolicy
= psvd
->policy_info
.mem_policy
;
698 if (lgrp_mem_policy_flags
==
699 LGRP_MP_FLAG_EXTEND_UP
) {
700 if (ppolicy
!= lgrp_mem_default_policy
) {
703 mpolicy
= lgrp_mem_policy_default(
704 pseg
->s_size
+ seg
->s_size
,
709 if (mpolicy
== ppolicy
&&
710 (pseg
->s_size
+ seg
->s_size
<=
711 segvn_comb_thrshld
|| psvd
->amp
== NULL
) &&
712 segvn_extend_prev(pseg
, seg
, a
, swresv
) == 0) {
714 * success! now try to concatenate
718 nseg
= AS_SEGNEXT(pseg
->s_as
, pseg
);
721 nseg
->s_ops
== &segvn_ops
&&
722 pseg
->s_base
+ pseg
->s_size
==
724 (void) segvn_concat(pseg
, nseg
, 0);
725 ASSERT(pseg
->s_szc
== 0 ||
726 (a
->szc
== pseg
->s_szc
&&
727 IS_P2ALIGNED(pseg
->s_base
, pgsz
) &&
728 IS_P2ALIGNED(pseg
->s_size
, pgsz
)));
734 * Failed, so try to concatenate with following seg
736 nseg
= AS_SEGNEXT(seg
->s_as
, seg
);
738 seg
->s_base
+ seg
->s_size
== nseg
->s_base
&&
739 nseg
->s_ops
== &segvn_ops
) {
741 * Get memory allocation policy from next segment.
742 * When extension is specified (e.g. for stack) apply
743 * this policy to the new segment regardless of the
744 * outcome of segment concatenation. Extension occurs
745 * for non-default policy otherwise default policy is
746 * used and is based on extended segment size.
748 nsvd
= (struct segvn_data
*)nseg
->s_data
;
749 npolicy
= nsvd
->policy_info
.mem_policy
;
750 if (lgrp_mem_policy_flags
==
751 LGRP_MP_FLAG_EXTEND_DOWN
) {
752 if (npolicy
!= lgrp_mem_default_policy
) {
755 mpolicy
= lgrp_mem_policy_default(
756 nseg
->s_size
+ seg
->s_size
,
761 if (mpolicy
== npolicy
&&
762 segvn_extend_next(seg
, nseg
, a
, swresv
) == 0) {
764 ASSERT(nseg
->s_szc
== 0 ||
765 (a
->szc
== nseg
->s_szc
&&
766 IS_P2ALIGNED(nseg
->s_base
, pgsz
) &&
767 IS_P2ALIGNED(nseg
->s_size
, pgsz
)));
775 if (a
->type
== MAP_SHARED
)
776 lgrp_shm_policy_init(NULL
, a
->vp
);
778 svd
= kmem_cache_alloc(segvn_cache
, KM_SLEEP
);
780 seg
->s_ops
= &segvn_ops
;
781 seg
->s_data
= (void *)svd
;
787 * Anonymous mappings have no backing file so the offset is meaningless.
789 svd
->offset
= a
->vp
? (a
->offset
& PAGEMASK
) : 0;
791 svd
->maxprot
= a
->maxprot
;
796 svd
->advice
= MADV_NORMAL
;
798 svd
->flags
= (ushort_t
)a
->flags
;
799 svd
->softlockcnt
= 0;
800 svd
->softlockcnt_sbase
= 0;
801 svd
->softlockcnt_send
= 0;
802 svd
->rcookie
= HAT_INVALID_REGION_COOKIE
;
805 if (a
->szc
!= 0 && a
->vp
!= NULL
) {
806 segvn_setvnode_mpss(a
->vp
);
808 if (svd
->type
== MAP_SHARED
&& svd
->vp
!= NULL
&&
809 (svd
->vp
->v_flag
& VVMEXEC
) && (svd
->prot
& PROT_WRITE
)) {
810 ASSERT(vn_is_mapped(svd
->vp
, V_WRITE
));
811 segvn_inval_trcache(svd
->vp
);
815 if ((svd
->amp
= amp
) == NULL
) {
817 if (svd
->type
== MAP_SHARED
) {
820 * Shared mappings to a vp need no other setup.
821 * If we have a shared mapping to an anon_map object
822 * which hasn't been allocated yet, allocate the
823 * struct now so that it will be properly shared
824 * by remembering the swap reservation there.
827 svd
->amp
= anonmap_alloc(seg
->s_size
, swresv
,
829 svd
->amp
->a_szc
= seg
->s_szc
;
833 * Private mapping (with or without a vp).
834 * Allocate anon_map when needed.
836 svd
->swresv
= swresv
;
842 * Mapping to an existing anon_map structure without a vp.
843 * For now we will insure that the segment size isn't larger
844 * than the size - offset gives us. Later on we may wish to
845 * have the anon array dynamically allocated itself so that
846 * we don't always have to allocate all the anon pointer slots.
847 * This of course involves adding extra code to check that we
848 * aren't trying to use an anon pointer slot beyond the end
849 * of the currently allocated anon array.
851 if ((amp
->size
- a
->offset
) < seg
->s_size
) {
852 panic("segvn_create anon_map size");
856 anon_num
= btopr(a
->offset
);
858 if (a
->type
== MAP_SHARED
) {
860 * SHARED mapping to a given anon_map.
862 ANON_LOCK_ENTER(&
->a_rwlock
, RW_WRITER
);
864 if (a
->szc
> amp
->a_szc
) {
867 ANON_LOCK_EXIT(&
->a_rwlock
);
868 svd
->anon_index
= anon_num
;
872 * PRIVATE mapping to a given anon_map.
873 * Make sure that all the needed anon
874 * structures are created (so that we will
875 * share the underlying pages if nothing
876 * is written by this mapping) and then
877 * duplicate the anon array as is done
878 * when a privately mapped segment is dup'ed.
884 int hat_flag
= HAT_LOAD
;
886 if (svd
->flags
& MAP_TEXT
) {
887 hat_flag
|= HAT_LOAD_TEXT
;
890 svd
->amp
= anonmap_alloc(seg
->s_size
, 0, ANON_SLEEP
);
891 svd
->amp
->a_szc
= seg
->s_szc
;
893 svd
->swresv
= swresv
;
896 * Prevent 2 threads from allocating anon
897 * slots simultaneously.
899 ANON_LOCK_ENTER(&
->a_rwlock
, RW_WRITER
);
900 eaddr
= seg
->s_base
+ seg
->s_size
;
902 for (anon_idx
= anon_num
, addr
= seg
->s_base
;
903 addr
< eaddr
; addr
+= PAGESIZE
, anon_idx
++) {
906 if ((ap
= anon_get_ptr(amp
->ahp
,
911 * Allocate the anon struct now.
912 * Might as well load up translation
913 * to the page while we're at it...
915 pp
= anon_zero(seg
, addr
, &ap
, cred
);
916 if (ap
== NULL
|| pp
== NULL
) {
917 panic("segvn_create anon_zero");
922 * Re-acquire the anon_map lock and
923 * initialize the anon array entry.
925 ASSERT(anon_get_ptr(amp
->ahp
,
927 (void) anon_set_ptr(amp
->ahp
, anon_idx
, ap
,
930 ASSERT(seg
->s_szc
== 0);
931 ASSERT(!IS_VMODSORT(pp
->p_vnode
));
933 ASSERT(use_rgn
== 0);
934 hat_memload(seg
->s_as
->a_hat
, addr
, pp
,
935 svd
->prot
& ~PROT_WRITE
, hat_flag
);
939 ASSERT(seg
->s_szc
== 0);
940 anon_dup(amp
->ahp
, anon_num
, svd
->amp
->ahp
,
942 ANON_LOCK_EXIT(&
->a_rwlock
);
947 * Set default memory allocation policy for segment
949 * Always set policy for private memory at least for initialization
950 * even if this is a shared memory segment
952 (void) lgrp_privm_policy_set(mpolicy
, &svd
->policy_info
, seg
->s_size
);
954 if (svd
->type
== MAP_SHARED
)
955 (void) lgrp_shm_policy_set(mpolicy
, svd
->amp
, svd
->anon_index
,
956 svd
->vp
, svd
->offset
, seg
->s_size
);
960 ASSERT(svd
->amp
== NULL
);
961 svd
->rcookie
= hat_join_region(seg
->s_as
->a_hat
, seg
->s_base
,
962 seg
->s_size
, (void *)svd
->vp
, svd
->offset
, svd
->prot
,
963 (uchar_t
)seg
->s_szc
, segvn_hat_rgn_unload_callback
,
967 ASSERT(!trok
|| !(svd
->prot
& PROT_WRITE
));
968 svd
->tr_state
= trok
? SEGVN_TR_INIT
: SEGVN_TR_OFF
;
974 * Concatenate two existing segments, if possible.
975 * Return 0 on success, -1 if two segments are not compatible
976 * or -2 on memory allocation failure.
977 * If amp_cat == 1 then try and concat segments with anon maps
980 segvn_concat(struct seg
*seg1
, struct seg
*seg2
, int amp_cat
)
982 struct segvn_data
*svd1
= seg1
->s_data
;
983 struct segvn_data
*svd2
= seg2
->s_data
;
984 struct anon_map
*amp1
= svd1
->amp
;
985 struct anon_map
*amp2
= svd2
->amp
;
986 struct vpage
*vpage1
= svd1
->vpage
;
987 struct vpage
*vpage2
= svd2
->vpage
, *nvpage
= NULL
;
988 size_t size
, nvpsize
;
989 pgcnt_t npages1
, npages2
;
991 ASSERT(seg1
->s_as
&& seg2
->s_as
&& seg1
->s_as
== seg2
->s_as
);
992 ASSERT(AS_WRITE_HELD(seg1
->s_as
, &seg1
->s_as
->a_lock
));
993 ASSERT(seg1
->s_ops
== seg2
->s_ops
);
995 if (HAT_IS_REGION_COOKIE_VALID(svd1
->rcookie
) ||
996 HAT_IS_REGION_COOKIE_VALID(svd2
->rcookie
)) {
1000 /* both segments exist, try to merge them */
1001 #define incompat(x) (svd1->x != svd2->x)
1002 if (incompat(vp
) || incompat(maxprot
) ||
1003 (!svd1
->pageadvice
&& !svd2
->pageadvice
&& incompat(advice
)) ||
1004 (!svd1
->pageprot
&& !svd2
->pageprot
&& incompat(prot
)) ||
1005 incompat(type
) || incompat(cred
) || incompat(flags
) ||
1006 seg1
->s_szc
!= seg2
->s_szc
|| incompat(policy_info
.mem_policy
) ||
1007 (svd2
->softlockcnt
> 0) || svd1
->softlockcnt_send
> 0)
1012 * vp == NULL implies zfod, offset doesn't matter
1014 if (svd1
->vp
!= NULL
&&
1015 svd1
->offset
+ seg1
->s_size
!= svd2
->offset
) {
1020 * Don't concatenate if either segment uses text replication.
1022 if (svd1
->tr_state
!= SEGVN_TR_OFF
|| svd2
->tr_state
!= SEGVN_TR_OFF
) {
1027 * Fail early if we're not supposed to concatenate
1028 * segments with non NULL amp.
1030 if (amp_cat
== 0 && (amp1
!= NULL
|| amp2
!= NULL
)) {
1034 if (svd1
->vp
== NULL
&& svd1
->type
== MAP_SHARED
) {
1038 if (amp1
!= NULL
&& svd1
->anon_index
+ btop(seg1
->s_size
) !=
1042 ASSERT(amp1
== NULL
|| amp1
->refcnt
>= 2);
1046 * If either seg has vpages, create a new merged vpage array.
1048 if (vpage1
!= NULL
|| vpage2
!= NULL
) {
1049 struct vpage
*vp
, *evp
;
1051 npages1
= seg_pages(seg1
);
1052 npages2
= seg_pages(seg2
);
1053 nvpsize
= vpgtob(npages1
+ npages2
);
1055 if ((nvpage
= kmem_zalloc(nvpsize
, KM_NOSLEEP
)) == NULL
) {
1059 if (vpage1
!= NULL
) {
1060 bcopy(vpage1
, nvpage
, vpgtob(npages1
));
1062 evp
= nvpage
+ npages1
;
1063 for (vp
= nvpage
; vp
< evp
; vp
++) {
1064 VPP_SETPROT(vp
, svd1
->prot
);
1065 VPP_SETADVICE(vp
, svd1
->advice
);
1069 if (vpage2
!= NULL
) {
1070 bcopy(vpage2
, nvpage
+ npages1
, vpgtob(npages2
));
1072 evp
= nvpage
+ npages1
+ npages2
;
1073 for (vp
= nvpage
+ npages1
; vp
< evp
; vp
++) {
1074 VPP_SETPROT(vp
, svd2
->prot
);
1075 VPP_SETADVICE(vp
, svd2
->advice
);
1079 if (svd2
->pageswap
&& (!svd1
->pageswap
&& svd1
->swresv
)) {
1080 ASSERT(svd1
->swresv
== seg1
->s_size
);
1081 ASSERT(!(svd1
->flags
& MAP_NORESERVE
));
1082 ASSERT(!(svd2
->flags
& MAP_NORESERVE
));
1083 evp
= nvpage
+ npages1
;
1084 for (vp
= nvpage
; vp
< evp
; vp
++) {
1089 if (svd1
->pageswap
&& (!svd2
->pageswap
&& svd2
->swresv
)) {
1090 ASSERT(svd2
->swresv
== seg2
->s_size
);
1091 ASSERT(!(svd1
->flags
& MAP_NORESERVE
));
1092 ASSERT(!(svd2
->flags
& MAP_NORESERVE
));
1093 vp
= nvpage
+ npages1
;
1095 for (; vp
< evp
; vp
++) {
1100 ASSERT((vpage1
!= NULL
|| vpage2
!= NULL
) ||
1101 (svd1
->pageswap
== 0 && svd2
->pageswap
== 0));
1104 * If either segment has private pages, create a new merged anon
1105 * array. If mergeing shared anon segments just decrement anon map's
1108 if (amp1
!= NULL
&& svd1
->type
== MAP_SHARED
) {
1109 ASSERT(amp1
== amp2
&& svd1
->vp
== NULL
);
1110 ANON_LOCK_ENTER(&1
->a_rwlock
, RW_WRITER
);
1111 ASSERT(amp1
->refcnt
>= 2);
1113 ANON_LOCK_EXIT(&1
->a_rwlock
);
1115 } else if (amp1
!= NULL
|| amp2
!= NULL
) {
1116 struct anon_hdr
*nahp
;
1117 struct anon_map
*namp
= NULL
;
1120 ASSERT(svd1
->type
== MAP_PRIVATE
);
1122 asize
= seg1
->s_size
+ seg2
->s_size
;
1123 if ((nahp
= anon_create(btop(asize
), ANON_NOSLEEP
)) == NULL
) {
1124 if (nvpage
!= NULL
) {
1125 kmem_free(nvpage
, nvpsize
);
1131 * XXX anon rwlock is not really needed because
1132 * this is a private segment and we are writers.
1134 ANON_LOCK_ENTER(&1
->a_rwlock
, RW_WRITER
);
1135 ASSERT(amp1
->refcnt
== 1);
1136 if (anon_copy_ptr(amp1
->ahp
, svd1
->anon_index
,
1137 nahp
, 0, btop(seg1
->s_size
), ANON_NOSLEEP
)) {
1138 anon_release(nahp
, btop(asize
));
1139 ANON_LOCK_EXIT(&1
->a_rwlock
);
1140 if (nvpage
!= NULL
) {
1141 kmem_free(nvpage
, nvpsize
);
1147 ANON_LOCK_ENTER(&2
->a_rwlock
, RW_WRITER
);
1148 ASSERT(amp2
->refcnt
== 1);
1149 if (anon_copy_ptr(amp2
->ahp
, svd2
->anon_index
,
1150 nahp
, btop(seg1
->s_size
), btop(seg2
->s_size
),
1152 anon_release(nahp
, btop(asize
));
1153 ANON_LOCK_EXIT(&2
->a_rwlock
);
1155 ANON_LOCK_EXIT(&1
->a_rwlock
);
1157 if (nvpage
!= NULL
) {
1158 kmem_free(nvpage
, nvpsize
);
1165 anon_release(amp1
->ahp
, btop(amp1
->size
));
1169 ASSERT(amp1
== NULL
);
1171 anon_release(amp2
->ahp
, btop(amp2
->size
));
1174 ANON_LOCK_EXIT(&2
->a_rwlock
);
1177 svd2
->amp
= NULL
; /* needed for seg_free */
1182 svd1
->anon_index
= 0;
1183 ANON_LOCK_EXIT(&namp
->a_rwlock
);
1186 * Now free the old vpage structures.
1188 if (nvpage
!= NULL
) {
1189 if (vpage1
!= NULL
) {
1190 kmem_free(vpage1
, vpgtob(npages1
));
1192 if (vpage2
!= NULL
) {
1194 kmem_free(vpage2
, vpgtob(npages2
));
1196 if (svd2
->pageprot
) {
1199 if (svd2
->pageadvice
) {
1200 svd1
->pageadvice
= 1;
1202 if (svd2
->pageswap
) {
1205 svd1
->vpage
= nvpage
;
1208 /* all looks ok, merge segments */
1209 svd1
->swresv
+= svd2
->swresv
;
1210 svd2
->swresv
= 0; /* so seg_free doesn't release swap space */
1211 size
= seg2
->s_size
;
1213 seg1
->s_size
+= size
;
1218 * Extend the previous segment (seg1) to include the
1219 * new segment (seg2 + a), if possible.
1220 * Return 0 on success.
1223 segvn_extend_prev(seg1
, seg2
, a
, swresv
)
1224 struct seg
*seg1
, *seg2
;
1225 struct segvn_crargs
*a
;
1228 struct segvn_data
*svd1
= (struct segvn_data
*)seg1
->s_data
;
1230 struct anon_map
*amp1
;
1231 struct vpage
*new_vpage
;
1234 * We don't need any segment level locks for "segvn" data
1235 * since the address space is "write" locked.
1237 ASSERT(seg1
->s_as
&& AS_WRITE_HELD(seg1
->s_as
, &seg1
->s_as
->a_lock
));
1239 if (HAT_IS_REGION_COOKIE_VALID(svd1
->rcookie
)) {
1243 /* second segment is new, try to extend first */
1244 /* XXX - should also check cred */
1245 if (svd1
->vp
!= a
->vp
|| svd1
->maxprot
!= a
->maxprot
||
1246 (!svd1
->pageprot
&& (svd1
->prot
!= a
->prot
)) ||
1247 svd1
->type
!= a
->type
|| svd1
->flags
!= a
->flags
||
1248 seg1
->s_szc
!= a
->szc
|| svd1
->softlockcnt_send
> 0)
1251 /* vp == NULL implies zfod, offset doesn't matter */
1252 if (svd1
->vp
!= NULL
&&
1253 svd1
->offset
+ seg1
->s_size
!= (a
->offset
& PAGEMASK
))
1256 if (svd1
->tr_state
!= SEGVN_TR_OFF
) {
1265 * Segment has private pages, can data structures
1268 * Acquire the anon_map lock to prevent it from changing,
1269 * if it is shared. This ensures that the anon_map
1270 * will not change while a thread which has a read/write
1271 * lock on an address space references it.
1272 * XXX - Don't need the anon_map lock at all if "refcnt"
1275 * Can't grow a MAP_SHARED segment with an anonmap because
1276 * there may be existing anon slots where we want to extend
1277 * the segment and we wouldn't know what to do with them
1278 * (e.g., for tmpfs right thing is to just leave them there,
1279 * for /dev/zero they should be cleared out).
1281 if (svd1
->type
== MAP_SHARED
)
1284 ANON_LOCK_ENTER(&1
->a_rwlock
, RW_WRITER
);
1285 if (amp1
->refcnt
> 1) {
1286 ANON_LOCK_EXIT(&1
->a_rwlock
);
1289 newpgs
= anon_grow(amp1
->ahp
, &svd1
->anon_index
,
1290 btop(seg1
->s_size
), btop(seg2
->s_size
), ANON_NOSLEEP
);
1293 ANON_LOCK_EXIT(&1
->a_rwlock
);
1296 amp1
->size
= ptob(newpgs
);
1297 ANON_LOCK_EXIT(&1
->a_rwlock
);
1299 if (svd1
->vpage
!= NULL
) {
1300 struct vpage
*vp
, *evp
;
1302 kmem_zalloc(vpgtob(seg_pages(seg1
) + seg_pages(seg2
)),
1304 if (new_vpage
== NULL
)
1306 bcopy(svd1
->vpage
, new_vpage
, vpgtob(seg_pages(seg1
)));
1307 kmem_free(svd1
->vpage
, vpgtob(seg_pages(seg1
)));
1308 svd1
->vpage
= new_vpage
;
1310 vp
= new_vpage
+ seg_pages(seg1
);
1311 evp
= vp
+ seg_pages(seg2
);
1312 for (; vp
< evp
; vp
++)
1313 VPP_SETPROT(vp
, a
->prot
);
1314 if (svd1
->pageswap
&& swresv
) {
1315 ASSERT(!(svd1
->flags
& MAP_NORESERVE
));
1316 ASSERT(swresv
== seg2
->s_size
);
1317 vp
= new_vpage
+ seg_pages(seg1
);
1318 for (; vp
< evp
; vp
++) {
1323 ASSERT(svd1
->vpage
!= NULL
|| svd1
->pageswap
== 0);
1324 size
= seg2
->s_size
;
1326 seg1
->s_size
+= size
;
1327 svd1
->swresv
+= swresv
;
1328 if (svd1
->pageprot
&& (a
->prot
& PROT_WRITE
) &&
1329 svd1
->type
== MAP_SHARED
&& svd1
->vp
!= NULL
&&
1330 (svd1
->vp
->v_flag
& VVMEXEC
)) {
1331 ASSERT(vn_is_mapped(svd1
->vp
, V_WRITE
));
1332 segvn_inval_trcache(svd1
->vp
);
1338 * Extend the next segment (seg2) to include the
1339 * new segment (seg1 + a), if possible.
1340 * Return 0 on success.
1346 struct segvn_crargs
*a
,
1349 struct segvn_data
*svd2
= (struct segvn_data
*)seg2
->s_data
;
1351 struct anon_map
*amp2
;
1352 struct vpage
*new_vpage
;
1355 * We don't need any segment level locks for "segvn" data
1356 * since the address space is "write" locked.
1358 ASSERT(seg2
->s_as
&& AS_WRITE_HELD(seg2
->s_as
, &seg2
->s_as
->a_lock
));
1360 if (HAT_IS_REGION_COOKIE_VALID(svd2
->rcookie
)) {
1364 /* first segment is new, try to extend second */
1365 /* XXX - should also check cred */
1366 if (svd2
->vp
!= a
->vp
|| svd2
->maxprot
!= a
->maxprot
||
1367 (!svd2
->pageprot
&& (svd2
->prot
!= a
->prot
)) ||
1368 svd2
->type
!= a
->type
|| svd2
->flags
!= a
->flags
||
1369 seg2
->s_szc
!= a
->szc
|| svd2
->softlockcnt_sbase
> 0)
1371 /* vp == NULL implies zfod, offset doesn't matter */
1372 if (svd2
->vp
!= NULL
&&
1373 (a
->offset
& PAGEMASK
) + seg1
->s_size
!= svd2
->offset
)
1376 if (svd2
->tr_state
!= SEGVN_TR_OFF
) {
1385 * Segment has private pages, can data structures
1388 * Acquire the anon_map lock to prevent it from changing,
1389 * if it is shared. This ensures that the anon_map
1390 * will not change while a thread which has a read/write
1391 * lock on an address space references it.
1393 * XXX - Don't need the anon_map lock at all if "refcnt"
1396 if (svd2
->type
== MAP_SHARED
)
1399 ANON_LOCK_ENTER(&2
->a_rwlock
, RW_WRITER
);
1400 if (amp2
->refcnt
> 1) {
1401 ANON_LOCK_EXIT(&2
->a_rwlock
);
1404 newpgs
= anon_grow(amp2
->ahp
, &svd2
->anon_index
,
1405 btop(seg2
->s_size
), btop(seg1
->s_size
),
1406 ANON_NOSLEEP
| ANON_GROWDOWN
);
1409 ANON_LOCK_EXIT(&2
->a_rwlock
);
1412 amp2
->size
= ptob(newpgs
);
1413 ANON_LOCK_EXIT(&2
->a_rwlock
);
1415 if (svd2
->vpage
!= NULL
) {
1416 struct vpage
*vp
, *evp
;
1418 kmem_zalloc(vpgtob(seg_pages(seg1
) + seg_pages(seg2
)),
1420 if (new_vpage
== NULL
) {
1421 /* Not merging segments so adjust anon_index back */
1423 svd2
->anon_index
+= seg_pages(seg1
);
1426 bcopy(svd2
->vpage
, new_vpage
+ seg_pages(seg1
),
1427 vpgtob(seg_pages(seg2
)));
1428 kmem_free(svd2
->vpage
, vpgtob(seg_pages(seg2
)));
1429 svd2
->vpage
= new_vpage
;
1432 evp
= vp
+ seg_pages(seg1
);
1433 for (; vp
< evp
; vp
++)
1434 VPP_SETPROT(vp
, a
->prot
);
1435 if (svd2
->pageswap
&& swresv
) {
1436 ASSERT(!(svd2
->flags
& MAP_NORESERVE
));
1437 ASSERT(swresv
== seg1
->s_size
);
1439 for (; vp
< evp
; vp
++) {
1444 ASSERT(svd2
->vpage
!= NULL
|| svd2
->pageswap
== 0);
1445 size
= seg1
->s_size
;
1447 seg2
->s_size
+= size
;
1448 seg2
->s_base
-= size
;
1449 svd2
->offset
-= size
;
1450 svd2
->swresv
+= swresv
;
1451 if (svd2
->pageprot
&& (a
->prot
& PROT_WRITE
) &&
1452 svd2
->type
== MAP_SHARED
&& svd2
->vp
!= NULL
&&
1453 (svd2
->vp
->v_flag
& VVMEXEC
)) {
1454 ASSERT(vn_is_mapped(svd2
->vp
, V_WRITE
));
1455 segvn_inval_trcache(svd2
->vp
);
1461 segvn_dup(struct seg
*seg
, struct seg
*newseg
)
1463 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
1464 struct segvn_data
*newsvd
;
1465 pgcnt_t npages
= seg_pages(seg
);
1469 struct anon_map
*amp
;
1471 ASSERT(seg
->s_as
&& AS_WRITE_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
1472 ASSERT(newseg
->s_as
->a_proc
->p_parent
== curproc
);
1475 * If segment has anon reserved, reserve more for the new seg.
1476 * For a MAP_NORESERVE segment swresv will be a count of all the
1477 * allocated anon slots; thus we reserve for the child as many slots
1478 * as the parent has allocated. This semantic prevents the child or
1479 * parent from dieing during a copy-on-write fault caused by trying
1480 * to write a shared pre-existing anon page.
1482 if ((len
= svd
->swresv
) != 0) {
1483 if (anon_resv(svd
->swresv
) == 0)
1486 TRACE_3(TR_FAC_VM
, TR_ANON_PROC
, "anon proc:%p %lu %u",
1490 newsvd
= kmem_cache_alloc(segvn_cache
, KM_SLEEP
);
1492 newseg
->s_ops
= &segvn_ops
;
1493 newseg
->s_data
= (void *)newsvd
;
1494 newseg
->s_szc
= seg
->s_szc
;
1496 newsvd
->seg
= newseg
;
1497 if ((newsvd
->vp
= svd
->vp
) != NULL
) {
1499 if (svd
->type
== MAP_SHARED
)
1500 lgrp_shm_policy_init(NULL
, svd
->vp
);
1502 newsvd
->offset
= svd
->offset
;
1503 newsvd
->prot
= svd
->prot
;
1504 newsvd
->maxprot
= svd
->maxprot
;
1505 newsvd
->pageprot
= svd
->pageprot
;
1506 newsvd
->type
= svd
->type
;
1507 newsvd
->cred
= svd
->cred
;
1508 crhold(newsvd
->cred
);
1509 newsvd
->advice
= svd
->advice
;
1510 newsvd
->pageadvice
= svd
->pageadvice
;
1511 newsvd
->swresv
= svd
->swresv
;
1512 newsvd
->pageswap
= svd
->pageswap
;
1513 newsvd
->flags
= svd
->flags
;
1514 newsvd
->softlockcnt
= 0;
1515 newsvd
->softlockcnt_sbase
= 0;
1516 newsvd
->softlockcnt_send
= 0;
1517 newsvd
->policy_info
= svd
->policy_info
;
1518 newsvd
->rcookie
= HAT_INVALID_REGION_COOKIE
;
1520 if ((amp
= svd
->amp
) == NULL
|| svd
->tr_state
== SEGVN_TR_ON
) {
1522 * Not attaching to a shared anon object.
1524 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
) ||
1525 svd
->tr_state
== SEGVN_TR_OFF
);
1526 if (svd
->tr_state
== SEGVN_TR_ON
) {
1527 ASSERT(newsvd
->vp
!= NULL
&& amp
!= NULL
);
1528 newsvd
->tr_state
= SEGVN_TR_INIT
;
1530 newsvd
->tr_state
= svd
->tr_state
;
1533 newsvd
->anon_index
= 0;
1535 /* regions for now are only used on pure vnode segments */
1536 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
1537 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
1538 newsvd
->tr_state
= SEGVN_TR_OFF
;
1539 if (svd
->type
== MAP_SHARED
) {
1541 ANON_LOCK_ENTER(&
->a_rwlock
, RW_WRITER
);
1543 ANON_LOCK_EXIT(&
->a_rwlock
);
1544 newsvd
->anon_index
= svd
->anon_index
;
1549 * Allocate and initialize new anon_map structure.
1551 newsvd
->amp
= anonmap_alloc(newseg
->s_size
, 0,
1553 newsvd
->amp
->a_szc
= newseg
->s_szc
;
1554 newsvd
->anon_index
= 0;
1557 * We don't have to acquire the anon_map lock
1558 * for the new segment (since it belongs to an
1559 * address space that is still not associated
1560 * with any process), or the segment in the old
1561 * address space (since all threads in it
1562 * are stopped while duplicating the address space).
1566 * The goal of the following code is to make sure that
1567 * softlocked pages do not end up as copy on write
1568 * pages. This would cause problems where one
1569 * thread writes to a page that is COW and a different
1570 * thread in the same process has softlocked it. The
1571 * softlock lock would move away from this process
1572 * because the write would cause this process to get
1573 * a copy (without the softlock).
1575 * The strategy here is to just break the
1576 * sharing on pages that could possibly be
1580 if (svd
->softlockcnt
) {
1581 struct anon
*ap
, *newap
;
1584 page_t
*anon_pl
[1+1], *pp
;
1586 ulong_t old_idx
= svd
->anon_index
;
1587 ulong_t new_idx
= 0;
1590 * The softlock count might be non zero
1591 * because some pages are still stuck in the
1592 * cache for lazy reclaim. Flush the cache
1593 * now. This should drop the count to zero.
1594 * [or there is really I/O going on to these
1595 * pages]. Note, we have the writers lock so
1596 * nothing gets inserted during the flush.
1603 i
= btopr(seg
->s_size
);
1606 * XXX break cow sharing using PAGESIZE
1607 * pages. They will be relocated into larger
1608 * pages at fault time.
1611 if (ap
= anon_get_ptr(amp
->ahp
,
1613 error
= anon_getpage(&ap
,
1614 &vpprot
, anon_pl
, PAGESIZE
,
1618 newsvd
->vpage
= NULL
;
1622 * prot need not be computed
1623 * below 'cause anon_private is
1624 * going to ignore it anyway
1625 * as child doesn't inherit
1626 * pagelock from parent.
1628 prot
= svd
->pageprot
?
1631 seg_page(seg
, addr
)])
1633 pp
= anon_private(&newap
,
1639 newsvd
->vpage
= NULL
;
1643 (void) anon_set_ptr(
1644 newsvd
->amp
->ahp
, new_idx
,
1652 } else { /* common case */
1653 if (seg
->s_szc
!= 0) {
1655 * If at least one of anon slots of a
1656 * large page exists then make sure
1657 * all anon slots of a large page
1658 * exist to avoid partial cow sharing
1659 * of a large page in the future.
1661 anon_dup_fill_holes(amp
->ahp
,
1662 svd
->anon_index
, newsvd
->amp
->ahp
,
1663 0, seg
->s_size
, seg
->s_szc
,
1666 anon_dup(amp
->ahp
, svd
->anon_index
,
1667 newsvd
->amp
->ahp
, 0, seg
->s_size
);
1670 hat_clrattr(seg
->s_as
->a_hat
, seg
->s_base
,
1671 seg
->s_size
, PROT_WRITE
);
1676 * If necessary, create a vpage structure for the new segment.
1677 * Do not copy any page lock indications.
1679 if (svd
->vpage
!= NULL
) {
1681 struct vpage
*ovp
= svd
->vpage
;
1684 nvp
= newsvd
->vpage
=
1685 kmem_alloc(vpgtob(npages
), KM_SLEEP
);
1686 for (i
= 0; i
< npages
; i
++) {
1688 VPP_CLRPPLOCK(nvp
++);
1691 newsvd
->vpage
= NULL
;
1693 /* Inform the vnode of the new mapping */
1694 if (newsvd
->vp
!= NULL
) {
1695 error
= VOP_ADDMAP(newsvd
->vp
, (offset_t
)newsvd
->offset
,
1696 newseg
->s_as
, newseg
->s_base
, newseg
->s_size
, newsvd
->prot
,
1697 newsvd
->maxprot
, newsvd
->type
, newsvd
->cred
, NULL
);
1700 if (error
== 0 && HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
)) {
1701 ASSERT(newsvd
->amp
== NULL
);
1702 ASSERT(newsvd
->tr_state
== SEGVN_TR_OFF
);
1703 newsvd
->rcookie
= svd
->rcookie
;
1704 hat_dup_region(newseg
->s_as
->a_hat
, newsvd
->rcookie
);
1711 * callback function to invoke free_vp_pages() for only those pages actually
1712 * processed by the HAT when a shared region is destroyed.
1714 extern int free_pages
;
1717 segvn_hat_rgn_unload_callback(caddr_t saddr
, caddr_t eaddr
, caddr_t r_saddr
,
1718 size_t r_size
, void *r_obj
, u_offset_t r_objoff
)
1722 vnode_t
*vp
= (vnode_t
*)r_obj
;
1724 ASSERT(eaddr
> saddr
);
1725 ASSERT(saddr
>= r_saddr
);
1726 ASSERT(saddr
< r_saddr
+ r_size
);
1727 ASSERT(eaddr
> r_saddr
);
1728 ASSERT(eaddr
<= r_saddr
+ r_size
);
1735 len
= eaddr
- saddr
;
1736 off
= (saddr
- r_saddr
) + r_objoff
;
1737 free_vp_pages(vp
, off
, len
);
1741 * callback function used by segvn_unmap to invoke free_vp_pages() for only
1742 * those pages actually processed by the HAT
1745 segvn_hat_unload_callback(hat_callback_t
*cb
)
1747 struct seg
*seg
= cb
->hcb_data
;
1748 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
1752 ASSERT(svd
->vp
!= NULL
);
1753 ASSERT(cb
->hcb_end_addr
> cb
->hcb_start_addr
);
1754 ASSERT(cb
->hcb_start_addr
>= seg
->s_base
);
1756 len
= cb
->hcb_end_addr
- cb
->hcb_start_addr
;
1757 off
= cb
->hcb_start_addr
- seg
->s_base
;
1758 free_vp_pages(svd
->vp
, svd
->offset
+ off
, len
);
1762 * This function determines the number of bytes of swap reserved by
1763 * a segment for which per-page accounting is present. It is used to
1764 * calculate the correct value of a segvn_data's swresv.
1767 segvn_count_swap_by_vpages(struct seg
*seg
)
1769 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
1770 struct vpage
*vp
, *evp
;
1771 size_t nswappages
= 0;
1773 ASSERT(svd
->pageswap
);
1774 ASSERT(svd
->vpage
!= NULL
);
1776 evp
= &svd
->vpage
[seg_page(seg
, seg
->s_base
+ seg
->s_size
)];
1778 for (vp
= svd
->vpage
; vp
< evp
; vp
++) {
1779 if (VPP_ISSWAPRES(vp
))
1783 return (nswappages
<< PAGESHIFT
);
1787 segvn_unmap(struct seg
*seg
, caddr_t addr
, size_t len
)
1789 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
1790 struct segvn_data
*nsvd
;
1792 struct anon_map
*amp
;
1793 pgcnt_t opages
; /* old segment size in pages */
1794 pgcnt_t npages
; /* new segment size in pages */
1795 pgcnt_t dpages
; /* pages being deleted (unmapped) */
1796 hat_callback_t callback
; /* used for free_vp_pages() */
1797 hat_callback_t
*cbp
= NULL
;
1804 * We don't need any segment level locks for "segvn" data
1805 * since the address space is "write" locked.
1807 ASSERT(seg
->s_as
&& AS_WRITE_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
1810 * Fail the unmap if pages are SOFTLOCKed through this mapping.
1811 * softlockcnt is protected from change by the as write lock.
1814 if (svd
->softlockcnt
> 0) {
1815 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
1818 * If this is shared segment non 0 softlockcnt
1819 * means locked pages are still in use.
1821 if (svd
->type
== MAP_SHARED
) {
1826 * since we do have the writers lock nobody can fill
1827 * the cache during the purge. The flush either succeeds
1828 * or we still have pending I/Os.
1839 * Check for bad sizes
1841 if (addr
< seg
->s_base
|| addr
+ len
> seg
->s_base
+ seg
->s_size
||
1842 (len
& PAGEOFFSET
) || ((uintptr_t)addr
& PAGEOFFSET
)) {
1843 panic("segvn_unmap");
1847 if (seg
->s_szc
!= 0) {
1848 size_t pgsz
= page_get_pagesize(seg
->s_szc
);
1850 if (!IS_P2ALIGNED(addr
, pgsz
) || !IS_P2ALIGNED(len
, pgsz
)) {
1851 ASSERT(seg
->s_base
!= addr
|| seg
->s_size
!= len
);
1852 if (HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
)) {
1853 ASSERT(svd
->amp
== NULL
);
1854 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
1855 hat_leave_region(seg
->s_as
->a_hat
,
1856 svd
->rcookie
, HAT_REGION_TEXT
);
1857 svd
->rcookie
= HAT_INVALID_REGION_COOKIE
;
1859 * could pass a flag to segvn_demote_range()
1860 * below to tell it not to do any unloads but
1861 * this case is rare enough to not bother for
1864 } else if (svd
->tr_state
== SEGVN_TR_INIT
) {
1865 svd
->tr_state
= SEGVN_TR_OFF
;
1866 } else if (svd
->tr_state
== SEGVN_TR_ON
) {
1867 ASSERT(svd
->amp
!= NULL
);
1868 segvn_textunrepl(seg
, 1);
1869 ASSERT(svd
->amp
== NULL
);
1870 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
1872 VM_STAT_ADD(segvnvmstats
.demoterange
[0]);
1873 err
= segvn_demote_range(seg
, addr
, len
, SDR_END
, 0);
1881 /* Inform the vnode of the unmapping. */
1885 error
= VOP_DELMAP(svd
->vp
,
1886 (offset_t
)svd
->offset
+ (uintptr_t)(addr
- seg
->s_base
),
1887 seg
->s_as
, addr
, len
, svd
->prot
, svd
->maxprot
,
1888 svd
->type
, svd
->cred
, NULL
);
1890 if (error
== EAGAIN
)
1895 * Remove any page locks set through this mapping.
1896 * If text replication is not off no page locks could have been
1897 * established via this mapping.
1899 if (svd
->tr_state
== SEGVN_TR_OFF
) {
1900 (void) segvn_lockop(seg
, addr
, len
, 0, MC_UNLOCK
, NULL
, 0);
1903 if (HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
)) {
1904 ASSERT(svd
->amp
== NULL
);
1905 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
1906 ASSERT(svd
->type
== MAP_PRIVATE
);
1907 hat_leave_region(seg
->s_as
->a_hat
, svd
->rcookie
,
1909 svd
->rcookie
= HAT_INVALID_REGION_COOKIE
;
1910 } else if (svd
->tr_state
== SEGVN_TR_ON
) {
1911 ASSERT(svd
->amp
!= NULL
);
1912 ASSERT(svd
->pageprot
== 0 && !(svd
->prot
& PROT_WRITE
));
1913 segvn_textunrepl(seg
, 1);
1914 ASSERT(svd
->amp
== NULL
&& svd
->tr_state
== SEGVN_TR_OFF
);
1916 if (svd
->tr_state
!= SEGVN_TR_OFF
) {
1917 ASSERT(svd
->tr_state
== SEGVN_TR_INIT
);
1918 svd
->tr_state
= SEGVN_TR_OFF
;
1921 * Unload any hardware translations in the range to be taken
1922 * out. Use a callback to invoke free_vp_pages() effectively.
1924 if (svd
->vp
!= NULL
&& free_pages
!= 0) {
1925 callback
.hcb_data
= seg
;
1926 callback
.hcb_function
= segvn_hat_unload_callback
;
1929 hat_unload_callback(seg
->s_as
->a_hat
, addr
, len
,
1930 HAT_UNLOAD_UNMAP
, cbp
);
1932 if (svd
->type
== MAP_SHARED
&& svd
->vp
!= NULL
&&
1933 (svd
->vp
->v_flag
& VVMEXEC
) &&
1934 ((svd
->prot
& PROT_WRITE
) || svd
->pageprot
)) {
1935 segvn_inval_trcache(svd
->vp
);
1940 * Check for entire segment
1942 if (addr
== seg
->s_base
&& len
== seg
->s_size
) {
1947 opages
= seg_pages(seg
);
1949 npages
= opages
- dpages
;
1951 ASSERT(amp
== NULL
|| amp
->a_szc
>= seg
->s_szc
);
1954 * Check for beginning of segment
1956 if (addr
== seg
->s_base
) {
1957 if (svd
->vpage
!= NULL
) {
1959 struct vpage
*ovpage
;
1961 ovpage
= svd
->vpage
; /* keep pointer to vpage */
1963 nbytes
= vpgtob(npages
);
1964 svd
->vpage
= kmem_alloc(nbytes
, KM_SLEEP
);
1965 bcopy(&ovpage
[dpages
], svd
->vpage
, nbytes
);
1967 /* free up old vpage */
1968 kmem_free(ovpage
, vpgtob(opages
));
1971 ANON_LOCK_ENTER(&
->a_rwlock
, RW_WRITER
);
1972 if (amp
->refcnt
== 1 || svd
->type
== MAP_PRIVATE
) {
1974 * Shared anon map is no longer in use. Before
1975 * freeing its pages purge all entries from
1976 * pcache that belong to this amp.
1978 if (svd
->type
== MAP_SHARED
) {
1979 ASSERT(amp
->refcnt
== 1);
1980 ASSERT(svd
->softlockcnt
== 0);
1984 * Free up now unused parts of anon_map array.
1986 if (amp
->a_szc
== seg
->s_szc
) {
1987 if (seg
->s_szc
!= 0) {
1988 anon_free_pages(amp
->ahp
,
1989 svd
->anon_index
, len
,
1997 ASSERT(svd
->type
== MAP_SHARED
);
1998 ASSERT(amp
->a_szc
> seg
->s_szc
);
1999 anon_shmap_free_pages(amp
,
2000 svd
->anon_index
, len
);
2004 * Unreserve swap space for the
2005 * unmapped chunk of this segment in
2006 * case it's MAP_SHARED
2008 if (svd
->type
== MAP_SHARED
) {
2009 anon_unresv_zone(len
,
2010 seg
->s_as
->a_proc
->p_zone
);
2014 ANON_LOCK_EXIT(&
->a_rwlock
);
2015 svd
->anon_index
+= dpages
;
2017 if (svd
->vp
!= NULL
)
2024 if (svd
->flags
& MAP_NORESERVE
) {
2026 oswresv
= svd
->swresv
;
2028 svd
->swresv
= ptob(anon_pages(amp
->ahp
,
2029 svd
->anon_index
, npages
));
2030 anon_unresv_zone(oswresv
- svd
->swresv
,
2031 seg
->s_as
->a_proc
->p_zone
);
2032 if (SEG_IS_PARTIAL_RESV(seg
))
2033 seg
->s_as
->a_resvsize
-= oswresv
-
2038 if (svd
->pageswap
) {
2039 oswresv
= svd
->swresv
;
2041 segvn_count_swap_by_vpages(seg
);
2042 ASSERT(oswresv
>= svd
->swresv
);
2043 unlen
= oswresv
- svd
->swresv
;
2046 ASSERT(svd
->swresv
== seg
->s_size
);
2049 anon_unresv_zone(unlen
,
2050 seg
->s_as
->a_proc
->p_zone
);
2052 TRACE_3(TR_FAC_VM
, TR_ANON_PROC
, "anon proc:%p %lu %u",
2060 * Check for end of segment
2062 if (addr
+ len
== seg
->s_base
+ seg
->s_size
) {
2063 if (svd
->vpage
!= NULL
) {
2065 struct vpage
*ovpage
;
2067 ovpage
= svd
->vpage
; /* keep pointer to vpage */
2069 nbytes
= vpgtob(npages
);
2070 svd
->vpage
= kmem_alloc(nbytes
, KM_SLEEP
);
2071 bcopy(ovpage
, svd
->vpage
, nbytes
);
2073 /* free up old vpage */
2074 kmem_free(ovpage
, vpgtob(opages
));
2078 ANON_LOCK_ENTER(&
->a_rwlock
, RW_WRITER
);
2079 if (amp
->refcnt
== 1 || svd
->type
== MAP_PRIVATE
) {
2081 * Free up now unused parts of anon_map array.
2083 ulong_t an_idx
= svd
->anon_index
+ npages
;
2086 * Shared anon map is no longer in use. Before
2087 * freeing its pages purge all entries from
2088 * pcache that belong to this amp.
2090 if (svd
->type
== MAP_SHARED
) {
2091 ASSERT(amp
->refcnt
== 1);
2092 ASSERT(svd
->softlockcnt
== 0);
2096 if (amp
->a_szc
== seg
->s_szc
) {
2097 if (seg
->s_szc
!= 0) {
2098 anon_free_pages(amp
->ahp
,
2102 anon_free(amp
->ahp
, an_idx
,
2106 ASSERT(svd
->type
== MAP_SHARED
);
2107 ASSERT(amp
->a_szc
> seg
->s_szc
);
2108 anon_shmap_free_pages(amp
,
2113 * Unreserve swap space for the
2114 * unmapped chunk of this segment in
2115 * case it's MAP_SHARED
2117 if (svd
->type
== MAP_SHARED
) {
2118 anon_unresv_zone(len
,
2119 seg
->s_as
->a_proc
->p_zone
);
2123 ANON_LOCK_EXIT(&
->a_rwlock
);
2129 if (svd
->flags
& MAP_NORESERVE
) {
2131 oswresv
= svd
->swresv
;
2132 svd
->swresv
= ptob(anon_pages(amp
->ahp
,
2133 svd
->anon_index
, npages
));
2134 anon_unresv_zone(oswresv
- svd
->swresv
,
2135 seg
->s_as
->a_proc
->p_zone
);
2136 if (SEG_IS_PARTIAL_RESV(seg
))
2137 seg
->s_as
->a_resvsize
-= oswresv
-
2142 if (svd
->pageswap
) {
2143 oswresv
= svd
->swresv
;
2145 segvn_count_swap_by_vpages(seg
);
2146 ASSERT(oswresv
>= svd
->swresv
);
2147 unlen
= oswresv
- svd
->swresv
;
2150 ASSERT(svd
->swresv
== seg
->s_size
);
2153 anon_unresv_zone(unlen
,
2154 seg
->s_as
->a_proc
->p_zone
);
2156 TRACE_3(TR_FAC_VM
, TR_ANON_PROC
,
2157 "anon proc:%p %lu %u", seg
, len
, 0);
2164 * The section to go is in the middle of the segment,
2165 * have to make it into two segments. nseg is made for
2166 * the high end while seg is cut down at the low end.
2168 nbase
= addr
+ len
; /* new seg base */
2169 nsize
= (seg
->s_base
+ seg
->s_size
) - nbase
; /* new seg size */
2170 seg
->s_size
= addr
- seg
->s_base
; /* shrink old seg */
2171 nseg
= seg_alloc(seg
->s_as
, nbase
, nsize
);
2173 panic("segvn_unmap seg_alloc");
2176 nseg
->s_ops
= seg
->s_ops
;
2177 nsvd
= kmem_cache_alloc(segvn_cache
, KM_SLEEP
);
2178 nseg
->s_data
= (void *)nsvd
;
2179 nseg
->s_szc
= seg
->s_szc
;
2182 nsvd
->offset
= svd
->offset
+ (uintptr_t)(nseg
->s_base
- seg
->s_base
);
2184 nsvd
->softlockcnt
= 0;
2185 nsvd
->softlockcnt_sbase
= 0;
2186 nsvd
->softlockcnt_send
= 0;
2187 ASSERT(nsvd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
2189 if (svd
->vp
!= NULL
) {
2191 if (nsvd
->type
== MAP_SHARED
)
2192 lgrp_shm_policy_init(NULL
, nsvd
->vp
);
2196 if (svd
->vpage
== NULL
) {
2199 /* need to split vpage into two arrays */
2201 struct vpage
*ovpage
;
2203 ovpage
= svd
->vpage
; /* keep pointer to vpage */
2205 npages
= seg_pages(seg
); /* seg has shrunk */
2206 nbytes
= vpgtob(npages
);
2207 svd
->vpage
= kmem_alloc(nbytes
, KM_SLEEP
);
2209 bcopy(ovpage
, svd
->vpage
, nbytes
);
2211 npages
= seg_pages(nseg
);
2212 nbytes
= vpgtob(npages
);
2213 nsvd
->vpage
= kmem_alloc(nbytes
, KM_SLEEP
);
2215 bcopy(&ovpage
[opages
- npages
], nsvd
->vpage
, nbytes
);
2217 /* free up old vpage */
2218 kmem_free(ovpage
, vpgtob(opages
));
2223 nsvd
->anon_index
= 0;
2226 * Need to create a new anon map for the new segment.
2227 * We'll also allocate a new smaller array for the old
2228 * smaller segment to save space.
2230 opages
= btop((uintptr_t)(addr
- seg
->s_base
));
2231 ANON_LOCK_ENTER(&
->a_rwlock
, RW_WRITER
);
2232 if (amp
->refcnt
== 1 || svd
->type
== MAP_PRIVATE
) {
2234 * Free up now unused parts of anon_map array.
2236 ulong_t an_idx
= svd
->anon_index
+ opages
;
2239 * Shared anon map is no longer in use. Before
2240 * freeing its pages purge all entries from
2241 * pcache that belong to this amp.
2243 if (svd
->type
== MAP_SHARED
) {
2244 ASSERT(amp
->refcnt
== 1);
2245 ASSERT(svd
->softlockcnt
== 0);
2249 if (amp
->a_szc
== seg
->s_szc
) {
2250 if (seg
->s_szc
!= 0) {
2251 anon_free_pages(amp
->ahp
, an_idx
, len
,
2254 anon_free(amp
->ahp
, an_idx
,
2258 ASSERT(svd
->type
== MAP_SHARED
);
2259 ASSERT(amp
->a_szc
> seg
->s_szc
);
2260 anon_shmap_free_pages(amp
, an_idx
, len
);
2264 * Unreserve swap space for the
2265 * unmapped chunk of this segment in
2266 * case it's MAP_SHARED
2268 if (svd
->type
== MAP_SHARED
) {
2269 anon_unresv_zone(len
,
2270 seg
->s_as
->a_proc
->p_zone
);
2274 nsvd
->anon_index
= svd
->anon_index
+
2275 btop((uintptr_t)(nseg
->s_base
- seg
->s_base
));
2276 if (svd
->type
== MAP_SHARED
) {
2280 struct anon_map
*namp
;
2281 struct anon_hdr
*nahp
;
2283 ASSERT(svd
->type
== MAP_PRIVATE
);
2284 nahp
= anon_create(btop(seg
->s_size
), ANON_SLEEP
);
2285 namp
= anonmap_alloc(nseg
->s_size
, 0, ANON_SLEEP
);
2286 namp
->a_szc
= seg
->s_szc
;
2287 (void) anon_copy_ptr(amp
->ahp
, svd
->anon_index
, nahp
,
2288 0, btop(seg
->s_size
), ANON_SLEEP
);
2289 (void) anon_copy_ptr(amp
->ahp
, nsvd
->anon_index
,
2290 namp
->ahp
, 0, btop(nseg
->s_size
), ANON_SLEEP
);
2291 anon_release(amp
->ahp
, btop(amp
->size
));
2292 svd
->anon_index
= 0;
2293 nsvd
->anon_index
= 0;
2295 amp
->size
= seg
->s_size
;
2298 ANON_LOCK_EXIT(&
->a_rwlock
);
2301 if (svd
->flags
& MAP_NORESERVE
) {
2303 oswresv
= svd
->swresv
;
2304 svd
->swresv
= ptob(anon_pages(amp
->ahp
,
2305 svd
->anon_index
, btop(seg
->s_size
)));
2306 nsvd
->swresv
= ptob(anon_pages(nsvd
->amp
->ahp
,
2307 nsvd
->anon_index
, btop(nseg
->s_size
)));
2308 ASSERT(oswresv
>= (svd
->swresv
+ nsvd
->swresv
));
2309 anon_unresv_zone(oswresv
- (svd
->swresv
+ nsvd
->swresv
),
2310 seg
->s_as
->a_proc
->p_zone
);
2311 if (SEG_IS_PARTIAL_RESV(seg
))
2312 seg
->s_as
->a_resvsize
-= oswresv
-
2313 (svd
->swresv
+ nsvd
->swresv
);
2317 if (svd
->pageswap
) {
2318 oswresv
= svd
->swresv
;
2319 svd
->swresv
= segvn_count_swap_by_vpages(seg
);
2320 nsvd
->swresv
= segvn_count_swap_by_vpages(nseg
);
2321 ASSERT(oswresv
>= (svd
->swresv
+ nsvd
->swresv
));
2322 unlen
= oswresv
- (svd
->swresv
+ nsvd
->swresv
);
2324 if (seg
->s_size
+ nseg
->s_size
+ len
!=
2326 panic("segvn_unmap: cannot split "
2327 "swap reservation");
2330 svd
->swresv
= seg
->s_size
;
2331 nsvd
->swresv
= nseg
->s_size
;
2334 anon_unresv_zone(unlen
,
2335 seg
->s_as
->a_proc
->p_zone
);
2337 TRACE_3(TR_FAC_VM
, TR_ANON_PROC
, "anon proc:%p %lu %u",
2341 return (0); /* I'm glad that's all over with! */
2345 segvn_free(struct seg
*seg
)
2347 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
2348 pgcnt_t npages
= seg_pages(seg
);
2349 struct anon_map
*amp
;
2353 * We don't need any segment level locks for "segvn" data
2354 * since the address space is "write" locked.
2356 ASSERT(seg
->s_as
&& AS_WRITE_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
2357 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
2359 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
2362 * Be sure to unlock pages. XXX Why do things get free'ed instead
2365 (void) segvn_lockop(seg
, seg
->s_base
, seg
->s_size
,
2366 0, MC_UNLOCK
, NULL
, 0);
2369 * Deallocate the vpage and anon pointers if necessary and possible.
2371 if (svd
->vpage
!= NULL
) {
2372 kmem_free(svd
->vpage
, vpgtob(npages
));
2375 if ((amp
= svd
->amp
) != NULL
) {
2377 * If there are no more references to this anon_map
2378 * structure, then deallocate the structure after freeing
2379 * up all the anon slot pointers that we can.
2381 ANON_LOCK_ENTER(&
->a_rwlock
, RW_WRITER
);
2382 ASSERT(amp
->a_szc
>= seg
->s_szc
);
2383 if (--amp
->refcnt
== 0) {
2384 if (svd
->type
== MAP_PRIVATE
) {
2386 * Private - we only need to anon_free
2387 * the part that this segment refers to.
2389 if (seg
->s_szc
!= 0) {
2390 anon_free_pages(amp
->ahp
,
2391 svd
->anon_index
, seg
->s_size
,
2394 anon_free(amp
->ahp
, svd
->anon_index
,
2400 * Shared anon map is no longer in use. Before
2401 * freeing its pages purge all entries from
2402 * pcache that belong to this amp.
2404 ASSERT(svd
->softlockcnt
== 0);
2408 * Shared - anon_free the entire
2409 * anon_map's worth of stuff and
2410 * release any swap reservation.
2412 if (amp
->a_szc
!= 0) {
2413 anon_shmap_free_pages(amp
, 0,
2416 anon_free(amp
->ahp
, 0, amp
->size
);
2418 if ((len
= amp
->swresv
) != 0) {
2419 anon_unresv_zone(len
,
2420 seg
->s_as
->a_proc
->p_zone
);
2421 TRACE_3(TR_FAC_VM
, TR_ANON_PROC
,
2422 "anon proc:%p %lu %u", seg
, len
, 0);
2426 ANON_LOCK_EXIT(&
->a_rwlock
);
2428 } else if (svd
->type
== MAP_PRIVATE
) {
2430 * We had a private mapping which still has
2431 * a held anon_map so just free up all the
2432 * anon slot pointers that we were using.
2434 if (seg
->s_szc
!= 0) {
2435 anon_free_pages(amp
->ahp
, svd
->anon_index
,
2436 seg
->s_size
, seg
->s_szc
);
2438 anon_free(amp
->ahp
, svd
->anon_index
,
2441 ANON_LOCK_EXIT(&
->a_rwlock
);
2443 ANON_LOCK_EXIT(&
->a_rwlock
);
2448 * Release swap reservation.
2450 if ((len
= svd
->swresv
) != 0) {
2451 anon_unresv_zone(svd
->swresv
,
2452 seg
->s_as
->a_proc
->p_zone
);
2453 TRACE_3(TR_FAC_VM
, TR_ANON_PROC
, "anon proc:%p %lu %u",
2455 if (SEG_IS_PARTIAL_RESV(seg
))
2456 seg
->s_as
->a_resvsize
-= svd
->swresv
;
2460 * Release claim on vnode, credentials, and finally free the
2463 if (svd
->vp
!= NULL
) {
2464 if (svd
->type
== MAP_SHARED
)
2465 lgrp_shm_policy_fini(NULL
, svd
->vp
);
2471 svd
->pageadvice
= 0;
2476 * Take segfree_syncmtx lock to let segvn_reclaim() finish if it's
2477 * still working with this segment without holding as lock (in case
2478 * it's called by pcache async thread).
2480 ASSERT(svd
->softlockcnt
== 0);
2481 mutex_enter(&svd
->segfree_syncmtx
);
2482 mutex_exit(&svd
->segfree_syncmtx
);
2485 kmem_cache_free(segvn_cache
, svd
);
2489 * Do a F_SOFTUNLOCK call over the range requested. The range must have
2490 * already been F_SOFTLOCK'ed.
2491 * Caller must always match addr and len of a softunlock with a previous
2492 * softlock with exactly the same addr and len.
2495 segvn_softunlock(struct seg
*seg
, caddr_t addr
, size_t len
, enum seg_rw rw
)
2497 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
2503 struct anon_map
*amp
;
2504 struct anon
*ap
= NULL
;
2506 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
2507 ASSERT(SEGVN_LOCK_HELD(seg
->s_as
, &svd
->lock
));
2509 if ((amp
= svd
->amp
) != NULL
)
2510 anon_index
= svd
->anon_index
+ seg_page(seg
, addr
);
2512 if (HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
)) {
2513 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
2514 hat_unlock_region(seg
->s_as
->a_hat
, addr
, len
, svd
->rcookie
);
2516 hat_unlock(seg
->s_as
->a_hat
, addr
, len
);
2518 for (adr
= addr
; adr
< addr
+ len
; adr
+= PAGESIZE
) {
2520 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
2521 if ((ap
= anon_get_ptr(amp
->ahp
, anon_index
++))
2523 swap_xlate(ap
, &vp
, &offset
);
2526 offset
= svd
->offset
+
2527 (uintptr_t)(adr
- seg
->s_base
);
2529 ANON_LOCK_EXIT(&
->a_rwlock
);
2532 offset
= svd
->offset
+
2533 (uintptr_t)(adr
- seg
->s_base
);
2537 * Use page_find() instead of page_lookup() to
2538 * find the page since we know that it is locked.
2540 pp
= page_find(vp
, offset
);
2543 "segvn_softunlock: addr %p, ap %p, vp %p, off %llx",
2544 (void *)adr
, (void *)ap
, (void *)vp
, offset
);
2548 if (rw
== S_WRITE
) {
2550 if (seg
->s_as
->a_vbits
)
2551 hat_setstat(seg
->s_as
, adr
, PAGESIZE
,
2553 } else if (rw
!= S_OTHER
) {
2555 if (seg
->s_as
->a_vbits
)
2556 hat_setstat(seg
->s_as
, adr
, PAGESIZE
, P_REF
);
2558 TRACE_3(TR_FAC_VM
, TR_SEGVN_FAULT
,
2559 "segvn_fault:pp %p vp %p offset %llx", pp
, vp
, offset
);
2562 ASSERT(svd
->softlockcnt
>= btop(len
));
2563 if (!atomic_add_long_nv((ulong_t
*)&svd
->softlockcnt
, -btop(len
))) {
2565 * All SOFTLOCKS are gone. Wakeup any waiting
2566 * unmappers so they can try again to unmap.
2567 * Check for waiters first without the mutex
2568 * held so we don't always grab the mutex on
2571 if (AS_ISUNMAPWAIT(seg
->s_as
)) {
2572 mutex_enter(&seg
->s_as
->a_contents
);
2573 if (AS_ISUNMAPWAIT(seg
->s_as
)) {
2574 AS_CLRUNMAPWAIT(seg
->s_as
);
2575 cv_broadcast(&seg
->s_as
->a_cv
);
2577 mutex_exit(&seg
->s_as
->a_contents
);
2582 #define PAGE_HANDLED ((page_t *)-1)
2585 * Release all the pages in the NULL terminated ppp list
2586 * which haven't already been converted to PAGE_HANDLED.
2589 segvn_pagelist_rele(page_t
**ppp
)
2591 for (; *ppp
!= NULL
; ppp
++) {
2592 if (*ppp
!= PAGE_HANDLED
)
2597 static int stealcow
= 1;
2600 * Workaround for viking chip bug. See bug id 1220902.
2601 * To fix this down in pagefault() would require importing so
2602 * much as and segvn code as to be unmaintainable.
2604 int enable_mbit_wa
= 0;
2607 * Handles all the dirty work of getting the right
2608 * anonymous pages and loading up the translations.
2609 * This routine is called only from segvn_fault()
2610 * when looping over the range of addresses requested.
2612 * The basic algorithm here is:
2613 * If this is an anon_zero case
2614 * Call anon_zero to allocate page
2615 * Load up translation
2618 * If this is an anon page
2619 * Use anon_getpage to get the page
2621 * Find page in pl[] list passed in
2624 * Load up the translation to the page
2627 * Call anon_private to handle cow
2628 * Load up (writable) translation to new page
2632 struct hat
*hat
, /* the hat to use for mapping */
2633 struct seg
*seg
, /* seg_vn of interest */
2634 caddr_t addr
, /* address in as */
2635 u_offset_t off
, /* offset in vp */
2636 struct vpage
*vpage
, /* pointer to vpage for vp, off */
2637 page_t
*pl
[], /* object source page pointer */
2638 uint_t vpprot
, /* access allowed to object pages */
2639 enum fault_type type
, /* type of fault */
2640 enum seg_rw rw
, /* type of access at fault */
2641 int brkcow
) /* we may need to break cow */
2643 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
2645 uint_t pageflags
= 0;
2646 page_t
*anon_pl
[1 + 1];
2647 page_t
*opp
= NULL
; /* original page */
2654 struct anon
*ap
, *oldap
;
2655 struct anon_map
*amp
;
2656 int hat_flag
= (type
== F_SOFTLOCK
) ? HAT_LOAD_LOCK
: HAT_LOAD
;
2658 anon_sync_obj_t cookie
;
2660 if (svd
->flags
& MAP_TEXT
) {
2661 hat_flag
|= HAT_LOAD_TEXT
;
2664 ASSERT(SEGVN_READ_HELD(seg
->s_as
, &svd
->lock
));
2665 ASSERT(seg
->s_szc
== 0);
2666 ASSERT(svd
->tr_state
!= SEGVN_TR_INIT
);
2669 * Initialize protection value for this page.
2670 * If we have per page protection values check it now.
2672 if (svd
->pageprot
) {
2677 protchk
= PROT_READ
;
2680 protchk
= PROT_WRITE
;
2683 protchk
= PROT_EXEC
;
2687 protchk
= PROT_READ
| PROT_WRITE
| PROT_EXEC
;
2691 prot
= VPP_PROT(vpage
);
2692 if ((prot
& protchk
) == 0)
2693 return (FC_PROT
); /* illegal access type */
2698 if (type
== F_SOFTLOCK
) {
2699 atomic_add_long((ulong_t
*)&svd
->softlockcnt
, 1);
2703 * Always acquire the anon array lock to prevent 2 threads from
2704 * allocating separate anon slots for the same "addr".
2707 if ((amp
= svd
->amp
) != NULL
) {
2708 ASSERT(RW_READ_HELD(&
->a_rwlock
));
2709 anon_index
= svd
->anon_index
+ seg_page(seg
, addr
);
2710 anon_array_enter(amp
, anon_index
, &cookie
);
2714 if (svd
->vp
== NULL
&& amp
!= NULL
) {
2715 if ((ap
= anon_get_ptr(amp
->ahp
, anon_index
)) == NULL
) {
2717 * Allocate a (normally) writable anonymous page of
2718 * zeroes. If no advance reservations, reserve now.
2720 if (svd
->flags
& MAP_NORESERVE
) {
2721 if (anon_resv_zone(ptob(1),
2722 seg
->s_as
->a_proc
->p_zone
)) {
2723 atomic_add_long(&svd
->swresv
, ptob(1));
2724 atomic_add_long(&seg
->s_as
->a_resvsize
,
2731 if ((pp
= anon_zero(seg
, addr
, &ap
,
2732 svd
->cred
)) == NULL
) {
2734 goto out
; /* out of swap space */
2737 * Re-acquire the anon_map lock and
2738 * initialize the anon array entry.
2740 (void) anon_set_ptr(amp
->ahp
, anon_index
, ap
,
2743 ASSERT(pp
->p_szc
== 0);
2746 * Handle pages that have been marked for migration
2748 if (lgrp_optimizations())
2749 page_migrate(seg
, addr
, &pp
, 1);
2751 if (enable_mbit_wa
) {
2754 else if (!hat_ismod(pp
))
2755 prot
&= ~PROT_WRITE
;
2758 * If AS_PAGLCK is set in a_flags (via memcntl(2)
2759 * with MC_LOCKAS, MCL_FUTURE) and this is a
2760 * MAP_NORESERVE segment, we may need to
2761 * permanently lock the page as it is being faulted
2762 * for the first time. The following text applies
2763 * only to MAP_NORESERVE segments:
2765 * As per memcntl(2), if this segment was created
2766 * after MCL_FUTURE was applied (a "future"
2767 * segment), its pages must be locked. If this
2768 * segment existed at MCL_FUTURE application (a
2769 * "past" segment), the interface is unclear.
2771 * We decide to lock only if vpage is present:
2773 * - "future" segments will have a vpage array (see
2774 * as_map), and so will be locked as required
2776 * - "past" segments may not have a vpage array,
2777 * depending on whether events (such as
2778 * mprotect) have occurred. Locking if vpage
2779 * exists will preserve legacy behavior. Not
2780 * locking if vpage is absent, will not break
2781 * the interface or legacy behavior. Note that
2782 * allocating vpage here if it's absent requires
2783 * upgrading the segvn reader lock, the cost of
2784 * which does not seem worthwhile.
2786 * Usually testing and setting VPP_ISPPLOCK and
2787 * VPP_SETPPLOCK requires holding the segvn lock as
2788 * writer, but in this case all readers are
2789 * serializing on the anon array lock.
2791 if (AS_ISPGLCK(seg
->s_as
) && vpage
!= NULL
&&
2792 (svd
->flags
& MAP_NORESERVE
) &&
2793 !VPP_ISPPLOCK(vpage
)) {
2794 proc_t
*p
= seg
->s_as
->a_proc
;
2795 ASSERT(svd
->type
== MAP_PRIVATE
);
2796 mutex_enter(&p
->p_lock
);
2797 if (rctl_incr_locked_mem(p
, NULL
, PAGESIZE
,
2799 claim
= VPP_PROT(vpage
) & PROT_WRITE
;
2800 if (page_pp_lock(pp
, claim
, 0)) {
2801 VPP_SETPPLOCK(vpage
);
2803 rctl_decr_locked_mem(p
, NULL
,
2807 mutex_exit(&p
->p_lock
);
2810 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
2811 hat_memload(hat
, addr
, pp
, prot
, hat_flag
);
2813 if (!(hat_flag
& HAT_LOAD_LOCK
))
2816 anon_array_exit(&cookie
);
2822 * Obtain the page structure via anon_getpage() if it is
2823 * a private copy of an object (the result of a previous
2827 if ((ap
= anon_get_ptr(amp
->ahp
, anon_index
)) != NULL
) {
2828 err
= anon_getpage(&ap
, &vpprot
, anon_pl
, PAGESIZE
,
2829 seg
, addr
, rw
, svd
->cred
);
2833 if (svd
->type
== MAP_SHARED
) {
2835 * If this is a shared mapping to an
2836 * anon_map, then ignore the write
2837 * permissions returned by anon_getpage().
2838 * They apply to the private mappings
2841 vpprot
|= PROT_WRITE
;
2848 * Search the pl[] list passed in if it is from the
2849 * original object (i.e., not a private copy).
2853 * Find original page. We must be bringing it in
2854 * from the list in pl[].
2856 for (ppp
= pl
; (opp
= *ppp
) != NULL
; ppp
++) {
2857 if (opp
== PAGE_HANDLED
)
2859 ASSERT(opp
->p_vnode
== svd
->vp
); /* XXX */
2860 if (opp
->p_offset
== off
)
2864 panic("segvn_faultpage not found");
2867 *ppp
= PAGE_HANDLED
;
2871 ASSERT(PAGE_LOCKED(opp
));
2873 TRACE_3(TR_FAC_VM
, TR_SEGVN_FAULT
,
2874 "segvn_fault:pp %p vp %p offset %llx", opp
, NULL
, 0);
2877 * The fault is treated as a copy-on-write fault if a
2878 * write occurs on a private segment and the object
2879 * page (i.e., mapping) is write protected. We assume
2880 * that fatal protection checks have already been made.
2884 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
2885 cow
= !(vpprot
& PROT_WRITE
);
2886 } else if (svd
->tr_state
== SEGVN_TR_ON
) {
2888 * If we are doing text replication COW on first touch.
2890 ASSERT(amp
!= NULL
);
2891 ASSERT(svd
->vp
!= NULL
);
2892 ASSERT(rw
!= S_WRITE
);
2899 * If not a copy-on-write case load the translation
2905 * Handle pages that have been marked for migration
2907 if (lgrp_optimizations())
2908 page_migrate(seg
, addr
, &opp
, 1);
2910 if (IS_VMODSORT(opp
->p_vnode
) || enable_mbit_wa
) {
2913 else if (rw
!= S_OTHER
&& !hat_ismod(opp
))
2914 prot
&= ~PROT_WRITE
;
2917 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
||
2918 (!svd
->pageprot
&& svd
->prot
== (prot
& vpprot
)));
2919 ASSERT(amp
== NULL
||
2920 svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
2921 hat_memload_region(hat
, addr
, opp
, prot
& vpprot
, hat_flag
,
2924 if (!(hat_flag
& HAT_LOAD_LOCK
))
2928 anon_array_exit(&cookie
);
2933 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
2937 ASSERT(amp
!= NULL
&& anon_lock
);
2940 * Steal the page only if it isn't a private page
2941 * since stealing a private page is not worth the effort.
2943 if ((ap
= anon_get_ptr(amp
->ahp
, anon_index
)) == NULL
)
2947 * Steal the original page if the following conditions are true:
2949 * We are low on memory, the page is not private, page is not large,
2950 * not shared, not modified, not `locked' or if we have it `locked'
2951 * (i.e., p_cowcnt == 1 and p_lckcnt == 0, which also implies
2952 * that the page is not shared) and if it doesn't have any
2953 * translations. page_struct_lock isn't needed to look at p_cowcnt
2954 * and p_lckcnt because we first get exclusive lock on page.
2956 (void) hat_pagesync(opp
, HAT_SYNC_DONTZERO
| HAT_SYNC_STOPON_MOD
);
2958 if (stealcow
&& freemem
< minfree
&& steal
&& opp
->p_szc
== 0 &&
2959 page_tryupgrade(opp
) && !hat_ismod(opp
) &&
2960 ((opp
->p_lckcnt
== 0 && opp
->p_cowcnt
== 0) ||
2961 (opp
->p_lckcnt
== 0 && opp
->p_cowcnt
== 1 &&
2962 vpage
!= NULL
&& VPP_ISPPLOCK(vpage
)))) {
2964 * Check if this page has other translations
2965 * after unloading our translation.
2967 if (hat_page_is_mapped(opp
)) {
2968 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
2969 hat_unload(seg
->s_as
->a_hat
, addr
, PAGESIZE
,
2974 * hat_unload() might sync back someone else's recent
2975 * modification, so check again.
2977 if (!hat_ismod(opp
) && !hat_page_is_mapped(opp
))
2978 pageflags
|= STEAL_PAGE
;
2982 * If we have a vpage pointer, see if it indicates that we have
2983 * ``locked'' the page we map -- if so, tell anon_private to
2984 * transfer the locking resource to the new page.
2986 * See Statement at the beginning of segvn_lockop regarding
2987 * the way lockcnts/cowcnts are handled during COW.
2990 if (vpage
!= NULL
&& VPP_ISPPLOCK(vpage
))
2991 pageflags
|= LOCK_PAGE
;
2994 * Allocate a private page and perform the copy.
2995 * For MAP_NORESERVE reserve swap space now, unless this
2996 * is a cow fault on an existing anon page in which case
2997 * MAP_NORESERVE will have made advance reservations.
2999 if ((svd
->flags
& MAP_NORESERVE
) && (ap
== NULL
)) {
3000 if (anon_resv_zone(ptob(1), seg
->s_as
->a_proc
->p_zone
)) {
3001 atomic_add_long(&svd
->swresv
, ptob(1));
3002 atomic_add_long(&seg
->s_as
->a_resvsize
, ptob(1));
3010 pp
= anon_private(&ap
, seg
, addr
, prot
, opp
, pageflags
, svd
->cred
);
3012 err
= ENOMEM
; /* out of swap space */
3017 * If we copied away from an anonymous page, then
3018 * we are one step closer to freeing up an anon slot.
3020 * NOTE: The original anon slot must be released while
3021 * holding the "anon_map" lock. This is necessary to prevent
3022 * other threads from obtaining a pointer to the anon slot
3023 * which may be freed if its "refcnt" is 1.
3028 (void) anon_set_ptr(amp
->ahp
, anon_index
, ap
, ANON_SLEEP
);
3031 * Handle pages that have been marked for migration
3033 if (lgrp_optimizations())
3034 page_migrate(seg
, addr
, &pp
, 1);
3036 ASSERT(pp
->p_szc
== 0);
3038 ASSERT(!IS_VMODSORT(pp
->p_vnode
));
3039 if (enable_mbit_wa
) {
3042 else if (!hat_ismod(pp
))
3043 prot
&= ~PROT_WRITE
;
3046 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
3047 hat_memload(hat
, addr
, pp
, prot
, hat_flag
);
3049 if (!(hat_flag
& HAT_LOAD_LOCK
))
3053 anon_array_exit(&cookie
);
3057 anon_array_exit(&cookie
);
3059 if (type
== F_SOFTLOCK
) {
3060 atomic_add_long((ulong_t
*)&svd
->softlockcnt
, -1);
3062 return (FC_MAKE_ERR(err
));
3066 * relocate a bunch of smaller targ pages into one large repl page. all targ
3067 * pages must be complete pages smaller than replacement pages.
3068 * it's assumed that no page's szc can change since they are all PAGESIZE or
3069 * complete large pages locked SHARED.
3072 segvn_relocate_pages(page_t
**targ
, page_t
*replacement
)
3075 pgcnt_t repl_npgs
, curnpgs
;
3077 uint_t repl_szc
= replacement
->p_szc
;
3078 page_t
*first_repl
= replacement
;
3082 VM_STAT_ADD(segvnvmstats
.relocatepages
[0]);
3084 ASSERT(repl_szc
!= 0);
3085 npgs
= repl_npgs
= page_get_pagecnt(repl_szc
);
3091 ASSERT(replacement
!= NULL
);
3093 ASSERT(pp
->p_szc
< repl_szc
);
3094 ASSERT(PAGE_EXCL(pp
));
3095 ASSERT(!PP_ISFREE(pp
));
3096 curnpgs
= page_get_pagecnt(pp
->p_szc
);
3098 VM_STAT_ADD(segvnvmstats
.relocatepages
[1]);
3100 page_sub(&replacement
, repl
);
3101 ASSERT(PAGE_EXCL(repl
));
3102 ASSERT(!PP_ISFREE(repl
));
3103 ASSERT(repl
->p_szc
== repl_szc
);
3105 page_t
*repl_savepp
;
3107 VM_STAT_ADD(segvnvmstats
.relocatepages
[2]);
3108 repl_savepp
= replacement
;
3109 for (j
= 0; j
< curnpgs
; j
++) {
3111 page_sub(&replacement
, repl
);
3112 ASSERT(PAGE_EXCL(repl
));
3113 ASSERT(!PP_ISFREE(repl
));
3114 ASSERT(repl
->p_szc
== repl_szc
);
3115 ASSERT(page_pptonum(targ
[i
+ j
]) ==
3116 page_pptonum(targ
[i
]) + j
);
3119 ASSERT(IS_P2ALIGNED(page_pptonum(repl
), curnpgs
));
3121 err
= page_relocate(&pp
, &repl
, 0, 1, &nreloc
, NULL
);
3122 if (err
|| nreloc
!= curnpgs
) {
3123 panic("segvn_relocate_pages: "
3124 "page_relocate failed err=%d curnpgs=%ld "
3125 "nreloc=%ld", err
, curnpgs
, nreloc
);
3127 ASSERT(curnpgs
<= repl_npgs
);
3128 repl_npgs
-= curnpgs
;
3131 ASSERT(replacement
== NULL
);
3135 for (i
= 0; i
< repl_npgs
; i
++) {
3136 ASSERT(PAGE_EXCL(repl
));
3137 ASSERT(!PP_ISFREE(repl
));
3139 page_downgrade(targ
[i
]);
3145 * Check if all pages in ppa array are complete smaller than szc pages and
3146 * their roots will still be aligned relative to their current size if the
3147 * entire ppa array is relocated into one szc page. If these conditions are
3150 * If all pages are properly aligned attempt to upgrade their locks
3151 * to exclusive mode. If it fails set *upgrdfail to 1 and return 0.
3152 * upgrdfail was set to 0 by caller.
3154 * Return 1 if all pages are aligned and locked exclusively.
3156 * If all pages in ppa array happen to be physically contiguous to make one
3157 * szc page and all exclusive locks are successfully obtained promote the page
3158 * size to szc and set *pszc to szc. Return 1 with pages locked shared.
3161 segvn_full_szcpages(page_t
**ppa
, uint_t szc
, int *upgrdfail
, uint_t
*pszc
)
3165 pgcnt_t totnpgs
= page_get_pagecnt(szc
);
3176 VM_STAT_ADD(segvnvmstats
.fullszcpages
[0]);
3178 for (i
= 0; i
< totnpgs
; i
++) {
3180 ASSERT(PAGE_SHARED(pp
));
3181 ASSERT(!PP_ISFREE(pp
));
3182 pfn
= page_pptonum(pp
);
3184 if (!IS_P2ALIGNED(pfn
, totnpgs
)) {
3189 } else if (contig
&& pfn
!= first_pfn
+ i
) {
3192 if (pp
->p_szc
== 0) {
3194 VM_STAT_ADD(segvnvmstats
.fullszcpages
[1]);
3198 if ((curszc
= pp
->p_szc
) >= szc
) {
3199 VM_STAT_ADD(segvnvmstats
.fullszcpages
[2]);
3204 * p_szc changed means we don't have all pages
3205 * locked. return failure.
3207 VM_STAT_ADD(segvnvmstats
.fullszcpages
[3]);
3210 curnpgs
= page_get_pagecnt(curszc
);
3211 if (!IS_P2ALIGNED(pfn
, curnpgs
) ||
3212 !IS_P2ALIGNED(i
, curnpgs
)) {
3213 VM_STAT_ADD(segvnvmstats
.fullszcpages
[4]);
3219 VM_STAT_ADD(segvnvmstats
.fullszcpages
[5]);
3220 if (pp
->p_szc
!= curszc
) {
3221 VM_STAT_ADD(segvnvmstats
.fullszcpages
[6]);
3224 if (pfn
- 1 != page_pptonum(ppa
[i
- 1])) {
3225 panic("segvn_full_szcpages: "
3226 "large page not physically contiguous");
3228 if (P2PHASE(pfn
, curnpgs
) == curnpgs
- 1) {
3234 for (i
= 0; i
< totnpgs
; i
++) {
3235 ASSERT(ppa
[i
]->p_szc
< szc
);
3236 if (!page_tryupgrade(ppa
[i
])) {
3237 for (j
= 0; j
< i
; j
++) {
3238 page_downgrade(ppa
[j
]);
3240 *pszc
= ppa
[i
]->p_szc
;
3242 VM_STAT_ADD(segvnvmstats
.fullszcpages
[7]);
3248 * When a page is put a free cachelist its szc is set to 0. if file
3249 * system reclaimed pages from cachelist targ pages will be physically
3250 * contiguous with 0 p_szc. in this case just upgrade szc of targ
3251 * pages without any relocations.
3252 * To avoid any hat issues with previous small mappings
3253 * hat_pageunload() the target pages first.
3256 VM_STAT_ADD(segvnvmstats
.fullszcpages
[8]);
3257 for (i
= 0; i
< totnpgs
; i
++) {
3258 (void) hat_pageunload(ppa
[i
], HAT_FORCE_PGUNLOAD
);
3260 for (i
= 0; i
< totnpgs
; i
++) {
3261 ppa
[i
]->p_szc
= szc
;
3263 for (i
= 0; i
< totnpgs
; i
++) {
3264 ASSERT(PAGE_EXCL(ppa
[i
]));
3265 page_downgrade(ppa
[i
]);
3271 VM_STAT_ADD(segvnvmstats
.fullszcpages
[9]);
3276 * Create physically contiguous pages for [vp, off] - [vp, off +
3277 * page_size(szc)) range and for private segment return them in ppa array.
3278 * Pages are created either via IO or relocations.
3280 * Return 1 on success and 0 on failure.
3282 * If physically contiguous pages already exist for this range return 1 without
3283 * filling ppa array. Caller initializes ppa[0] as NULL to detect that ppa
3284 * array wasn't filled. In this case caller fills ppa array via VOP_GETPAGE().
3288 segvn_fill_vp_pages(struct segvn_data
*svd
, vnode_t
*vp
, u_offset_t off
,
3289 uint_t szc
, page_t
**ppa
, page_t
**ppplist
, uint_t
*ret_pszc
,
3293 page_t
*pplist
= *ppplist
;
3294 size_t pgsz
= page_get_pagesize(szc
);
3295 pgcnt_t pages
= btop(pgsz
);
3296 ulong_t start_off
= off
;
3297 u_offset_t eoff
= off
+ pgsz
;
3299 u_offset_t io_off
= off
;
3301 page_t
*io_pplist
= NULL
;
3302 page_t
*done_pplist
= NULL
;
3311 page_t
*targ_pplist
= NULL
;
3312 page_t
*repl_pplist
= NULL
;
3318 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[0]);
3321 ASSERT(pplist
->p_szc
== szc
);
3324 * downsize will be set to 1 only if we fail to lock pages. this will
3325 * allow subsequent faults to try to relocate the page again. If we
3326 * fail due to misalignment don't downsize and let the caller map the
3327 * whole region with small mappings to avoid more faults into the area
3328 * where we can't get large pages anyway.
3332 while (off
< eoff
) {
3334 ASSERT(newpp
!= NULL
);
3335 ASSERT(PAGE_EXCL(newpp
));
3336 ASSERT(!PP_ISFREE(newpp
));
3338 * we pass NULL for nrelocp to page_lookup_create()
3339 * so that it doesn't relocate. We relocate here
3340 * later only after we make sure we can lock all
3341 * pages in the range we handle and they are all
3344 pp
= page_lookup_create(vp
, off
, SE_SHARED
, newpp
, NULL
, 0);
3346 ASSERT(!PP_ISFREE(pp
));
3347 ASSERT(pp
->p_vnode
== vp
);
3348 ASSERT(pp
->p_offset
== off
);
3350 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[1]);
3351 page_sub(&pplist
, pp
);
3352 ASSERT(PAGE_EXCL(pp
));
3353 ASSERT(page_iolock_assert(pp
));
3354 page_list_concat(&io_pplist
, &pp
);
3358 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[2]);
3359 pfn
= page_pptonum(pp
);
3361 if (pszc
>= szc
&& targ_pplist
== NULL
&& io_pplist
== NULL
&&
3362 IS_P2ALIGNED(pfn
, pages
)) {
3363 ASSERT(repl_pplist
== NULL
);
3364 ASSERT(done_pplist
== NULL
);
3365 ASSERT(pplist
== *ppplist
);
3367 page_free_replacement_page(pplist
);
3368 page_create_putback(pages
);
3370 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[3]);
3375 segvn_faultvnmpss_align_err1
++;
3378 ppages
= page_get_pagecnt(pszc
);
3379 if (!IS_P2ALIGNED(pfn
, ppages
)) {
3382 * sizing down to pszc won't help.
3385 segvn_faultvnmpss_align_err2
++;
3388 pfn
= page_pptonum(newpp
);
3389 if (!IS_P2ALIGNED(pfn
, ppages
)) {
3392 * sizing down to pszc won't help.
3395 segvn_faultvnmpss_align_err3
++;
3398 if (!PAGE_EXCL(pp
)) {
3399 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[4]);
3402 *ret_pszc
= pp
->p_szc
;
3406 if (io_pplist
!= NULL
) {
3407 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[5]);
3408 io_len
= off
- io_off
;
3410 * Some file systems like NFS don't check EOF
3411 * conditions in VOP_PAGEIO(). Check it here
3412 * now that pages are locked SE_EXCL. Any file
3413 * truncation will wait until the pages are
3414 * unlocked so no need to worry that file will
3415 * be truncated after we check its size here.
3416 * XXX fix NFS to remove this check.
3418 va
.va_mask
= AT_SIZE
;
3419 if (VOP_GETATTR(vp
, &va
, ATTR_HINT
, svd
->cred
, NULL
)) {
3420 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[6]);
3421 page_unlock(targpp
);
3424 if (btopr(va
.va_size
) < btopr(io_off
+ io_len
)) {
3425 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[7]);
3428 page_unlock(targpp
);
3431 io_err
= VOP_PAGEIO(vp
, io_pplist
, io_off
, io_len
,
3432 B_READ
, svd
->cred
, NULL
);
3434 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[8]);
3435 page_unlock(targpp
);
3436 if (io_err
== EDEADLK
) {
3437 segvn_vmpss_pageio_deadlk_err
++;
3442 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[9]);
3443 while (io_pplist
!= NULL
) {
3445 page_sub(&io_pplist
, pp
);
3446 ASSERT(page_iolock_assert(pp
));
3448 pgidx
= (pp
->p_offset
- start_off
) >>
3450 ASSERT(pgidx
< pages
);
3452 page_list_concat(&done_pplist
, &pp
);
3456 ASSERT(PAGE_EXCL(pp
));
3457 ASSERT(pp
->p_szc
<= pszc
);
3458 if (pszc
!= 0 && !group_page_trylock(pp
, SE_EXCL
)) {
3459 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[10]);
3462 *ret_pszc
= pp
->p_szc
;
3465 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[11]);
3467 * page szc chould have changed before the entire group was
3468 * locked. reread page szc.
3471 ppages
= page_get_pagecnt(pszc
);
3473 /* link just the roots */
3474 page_list_concat(&targ_pplist
, &pp
);
3475 page_sub(&pplist
, newpp
);
3476 page_list_concat(&repl_pplist
, &newpp
);
3478 while (--ppages
!= 0) {
3480 page_sub(&pplist
, newpp
);
3485 if (io_pplist
!= NULL
) {
3486 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[12]);
3487 io_len
= eoff
- io_off
;
3488 va
.va_mask
= AT_SIZE
;
3489 if (VOP_GETATTR(vp
, &va
, ATTR_HINT
, svd
->cred
, NULL
) != 0) {
3490 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[13]);
3493 if (btopr(va
.va_size
) < btopr(io_off
+ io_len
)) {
3494 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[14]);
3499 io_err
= VOP_PAGEIO(vp
, io_pplist
, io_off
, io_len
,
3500 B_READ
, svd
->cred
, NULL
);
3502 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[15]);
3503 if (io_err
== EDEADLK
) {
3504 segvn_vmpss_pageio_deadlk_err
++;
3509 while (io_pplist
!= NULL
) {
3511 page_sub(&io_pplist
, pp
);
3512 ASSERT(page_iolock_assert(pp
));
3514 pgidx
= (pp
->p_offset
- start_off
) >> PAGESHIFT
;
3515 ASSERT(pgidx
< pages
);
3520 * we're now bound to succeed or panic.
3521 * remove pages from done_pplist. it's not needed anymore.
3523 while (done_pplist
!= NULL
) {
3525 page_sub(&done_pplist
, pp
);
3527 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[16]);
3528 ASSERT(pplist
== NULL
);
3530 while (targ_pplist
!= NULL
) {
3532 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[17]);
3533 ASSERT(repl_pplist
);
3535 page_sub(&targ_pplist
, pp
);
3536 pgidx
= (pp
->p_offset
- start_off
) >> PAGESHIFT
;
3537 newpp
= repl_pplist
;
3538 page_sub(&repl_pplist
, newpp
);
3540 pfn
= page_pptonum(pp
);
3542 ppages
= page_get_pagecnt(pszc
);
3543 ASSERT(IS_P2ALIGNED(pfn
, ppages
));
3544 pfn
= page_pptonum(newpp
);
3545 ASSERT(IS_P2ALIGNED(pfn
, ppages
));
3546 ASSERT(P2PHASE(pfn
, pages
) == pgidx
);
3549 ret
= page_relocate(&pp
, &newpp
, 0, 1, &nreloc
, NULL
);
3550 if (ret
!= 0 || nreloc
== 0) {
3551 panic("segvn_fill_vp_pages: "
3552 "page_relocate failed");
3555 while (nreloc
-- != 0) {
3556 ASSERT(PAGE_EXCL(pp
));
3557 ASSERT(pp
->p_vnode
== vp
);
3559 ((pp
->p_offset
- start_off
) >> PAGESHIFT
));
3565 if (svd
->type
== MAP_PRIVATE
) {
3566 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[18]);
3567 for (i
= 0; i
< pages
; i
++) {
3568 ASSERT(ppa
[i
] != NULL
);
3569 ASSERT(PAGE_EXCL(ppa
[i
]));
3570 ASSERT(ppa
[i
]->p_vnode
== vp
);
3571 ASSERT(ppa
[i
]->p_offset
==
3572 start_off
+ (i
<< PAGESHIFT
));
3573 page_downgrade(ppa
[i
]);
3577 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[19]);
3579 * the caller will still call VOP_GETPAGE() for shared segments
3580 * to check FS write permissions. For private segments we map
3581 * file read only anyway. so no VOP_GETPAGE is needed.
3583 for (i
= 0; i
< pages
; i
++) {
3584 ASSERT(ppa
[i
] != NULL
);
3585 ASSERT(PAGE_EXCL(ppa
[i
]));
3586 ASSERT(ppa
[i
]->p_vnode
== vp
);
3587 ASSERT(ppa
[i
]->p_offset
==
3588 start_off
+ (i
<< PAGESHIFT
));
3589 page_unlock(ppa
[i
]);
3597 * Do the cleanup. Unlock target pages we didn't relocate. They are
3598 * linked on targ_pplist by root pages. reassemble unused replacement
3599 * and io pages back to pplist.
3601 if (io_pplist
!= NULL
) {
3602 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[20]);
3605 ASSERT(pp
->p_vnode
== vp
);
3606 ASSERT(pp
->p_offset
== io_off
);
3607 ASSERT(page_iolock_assert(pp
));
3609 page_hashout(pp
, NULL
);
3611 } while ((pp
= pp
->p_next
) != io_pplist
);
3612 page_list_concat(&io_pplist
, &pplist
);
3616 while (targ_pplist
!= NULL
) {
3617 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[21]);
3619 ASSERT(PAGE_EXCL(pp
));
3620 page_sub(&targ_pplist
, pp
);
3623 ppages
= page_get_pagecnt(pszc
);
3624 ASSERT(IS_P2ALIGNED(page_pptonum(pp
), ppages
));
3627 group_page_unlock(pp
);
3633 ASSERT(PAGE_EXCL(pp
));
3634 ASSERT(pp
->p_szc
== szc
);
3635 page_sub(&repl_pplist
, pp
);
3637 ASSERT(IS_P2ALIGNED(page_pptonum(pp
), ppages
));
3639 /* relink replacement page */
3640 page_list_concat(&tmp_pplist
, &pp
);
3641 while (--ppages
!= 0) {
3642 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[22]);
3644 ASSERT(PAGE_EXCL(pp
));
3645 ASSERT(pp
->p_szc
== szc
);
3646 page_list_concat(&tmp_pplist
, &pp
);
3649 if (tmp_pplist
!= NULL
) {
3650 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[23]);
3651 page_list_concat(&tmp_pplist
, &pplist
);
3652 pplist
= tmp_pplist
;
3655 * at this point all pages are either on done_pplist or
3656 * pplist. They can't be all on done_pplist otherwise
3657 * we'd've been done.
3659 ASSERT(pplist
!= NULL
);
3661 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[24]);
3664 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[25]);
3665 ASSERT(pp
->p_szc
== szc
);
3666 ASSERT(PAGE_EXCL(pp
));
3667 ASSERT(pp
->p_vnode
!= vp
);
3669 } while ((pp
= pp
->p_next
) != pplist
);
3673 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[26]);
3674 ASSERT(pp
->p_szc
== szc
);
3675 ASSERT(PAGE_EXCL(pp
));
3676 ASSERT(pp
->p_vnode
== vp
);
3678 } while ((pp
= pp
->p_next
) != done_pplist
);
3680 while (pplist
!= NULL
) {
3681 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[27]);
3683 page_sub(&pplist
, pp
);
3687 while (done_pplist
!= NULL
) {
3688 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[28]);
3690 page_sub(&done_pplist
, pp
);
3696 ASSERT(pplist
== *ppplist
);
3698 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[29]);
3700 * don't downsize on io error.
3701 * see if vop_getpage succeeds.
3702 * pplist may still be used in this case
3707 VM_STAT_ADD(segvnvmstats
.fill_vp_pages
[30]);
3708 page_free_replacement_page(pplist
);
3709 page_create_putback(pages
);
3714 int segvn_anypgsz
= 0;
3716 #define SEGVN_RESTORE_SOFTLOCK_VP(type, pages) \
3717 if ((type) == F_SOFTLOCK) { \
3718 atomic_add_long((ulong_t *)&(svd)->softlockcnt, \
3722 #define SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot) \
3723 if (IS_VMODSORT((ppa)[0]->p_vnode)) { \
3724 if ((rw) == S_WRITE) { \
3725 for (i = 0; i < (pages); i++) { \
3726 ASSERT((ppa)[i]->p_vnode == \
3727 (ppa)[0]->p_vnode); \
3728 hat_setmod((ppa)[i]); \
3730 } else if ((rw) != S_OTHER && \
3731 ((prot) & (vpprot) & PROT_WRITE)) { \
3732 for (i = 0; i < (pages); i++) { \
3733 ASSERT((ppa)[i]->p_vnode == \
3734 (ppa)[0]->p_vnode); \
3735 if (!hat_ismod((ppa)[i])) { \
3736 prot &= ~PROT_WRITE; \
3745 #define SEGVN_VMSTAT_FLTVNPAGES(idx) \
3746 VM_STAT_ADD(segvnvmstats.fltvnpages[(idx)]);
3748 #else /* VM_STATS */
3750 #define SEGVN_VMSTAT_FLTVNPAGES(idx)
3755 segvn_fault_vnodepages(struct hat
*hat
, struct seg
*seg
, caddr_t lpgaddr
,
3756 caddr_t lpgeaddr
, enum fault_type type
, enum seg_rw rw
, caddr_t addr
,
3757 caddr_t eaddr
, int brkcow
)
3759 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
3760 struct anon_map
*amp
= svd
->amp
;
3761 uchar_t segtype
= svd
->type
;
3762 uint_t szc
= seg
->s_szc
;
3763 size_t pgsz
= page_get_pagesize(szc
);
3764 size_t maxpgsz
= pgsz
;
3765 pgcnt_t pages
= btop(pgsz
);
3766 pgcnt_t maxpages
= pages
;
3767 size_t ppasize
= (pages
+ 1) * sizeof (page_t
*);
3768 caddr_t a
= lpgaddr
;
3769 caddr_t maxlpgeaddr
= lpgeaddr
;
3770 u_offset_t off
= svd
->offset
+ (uintptr_t)(a
- seg
->s_base
);
3771 ulong_t aindx
= svd
->anon_index
+ seg_page(seg
, a
);
3772 struct vpage
*vpage
= (svd
->vpage
!= NULL
) ?
3773 &svd
->vpage
[seg_page(seg
, a
)] : NULL
;
3774 vnode_t
*vp
= svd
->vp
;
3779 faultcode_t err
= 0;
3781 int vop_size_err
= 0;
3782 uint_t protchk
, prot
, vpprot
;
3784 int hat_flag
= (type
== F_SOFTLOCK
) ? HAT_LOAD_LOCK
: HAT_LOAD
;
3785 anon_sync_obj_t an_cookie
;
3787 int alloc_failed
= 0;
3795 int segvn_anypgsz_vnode
= 0; /* for now map vnode with 2 page sizes */
3796 int tron
= (svd
->tr_state
== SEGVN_TR_ON
);
3800 ASSERT(brkcow
== 0 || amp
!= NULL
);
3801 ASSERT(tron
== 0 || amp
!= NULL
);
3802 ASSERT(enable_mbit_wa
== 0); /* no mbit simulations with large pages */
3803 ASSERT(!(svd
->flags
& MAP_NORESERVE
));
3804 ASSERT(type
!= F_SOFTUNLOCK
);
3805 ASSERT(IS_P2ALIGNED(a
, maxpgsz
));
3806 ASSERT(amp
== NULL
|| IS_P2ALIGNED(aindx
, maxpages
));
3807 ASSERT(SEGVN_LOCK_HELD(seg
->s_as
, &svd
->lock
));
3808 ASSERT(seg
->s_szc
< NBBY
* sizeof (int));
3809 ASSERT(type
!= F_SOFTLOCK
|| lpgeaddr
- a
== maxpgsz
);
3810 ASSERT(svd
->tr_state
!= SEGVN_TR_INIT
);
3812 VM_STAT_COND_ADD(type
== F_SOFTLOCK
, segvnvmstats
.fltvnpages
[0]);
3813 VM_STAT_COND_ADD(type
!= F_SOFTLOCK
, segvnvmstats
.fltvnpages
[1]);
3815 if (svd
->flags
& MAP_TEXT
) {
3816 hat_flag
|= HAT_LOAD_TEXT
;
3819 if (svd
->pageprot
) {
3822 protchk
= PROT_READ
;
3825 protchk
= PROT_WRITE
;
3828 protchk
= PROT_EXEC
;
3832 protchk
= PROT_READ
| PROT_WRITE
| PROT_EXEC
;
3837 /* caller has already done segment level protection check. */
3840 if (seg
->s_as
->a_hat
!= hat
) {
3844 if (rw
== S_WRITE
&& segtype
== MAP_PRIVATE
) {
3845 SEGVN_VMSTAT_FLTVNPAGES(2);
3851 ppa
= kmem_alloc(ppasize
, KM_SLEEP
);
3853 VM_STAT_COND_ADD(amp
!= NULL
, segvnvmstats
.fltvnpages
[3]);
3857 for (; a
< lpgeaddr
; a
+= pgsz
, off
+= pgsz
, aindx
+= pages
) {
3859 while (szc
< seg
->s_szc
) {
3862 tszc
= segvn_anypgsz_vnode
? szc
+ 1 :
3864 ppgsz
= page_get_pagesize(tszc
);
3865 if (!IS_P2ALIGNED(a
, ppgsz
) ||
3866 ((alloc_failed
>> tszc
) & 0x1)) {
3869 SEGVN_VMSTAT_FLTVNPAGES(4);
3873 e
= P2ROUNDUP((uintptr_t)eaddr
, pgsz
);
3874 lpgeaddr
= (caddr_t
)e
;
3879 if (IS_P2ALIGNED(a
, maxpgsz
) && amp
!= NULL
) {
3880 ASSERT(IS_P2ALIGNED(aindx
, maxpages
));
3881 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
3882 anon_array_enter(amp
, aindx
, &an_cookie
);
3883 if (anon_get_ptr(amp
->ahp
, aindx
) != NULL
) {
3884 SEGVN_VMSTAT_FLTVNPAGES(5);
3885 ASSERT(anon_pages(amp
->ahp
, aindx
,
3886 maxpages
) == maxpages
);
3887 anon_array_exit(&an_cookie
);
3888 ANON_LOCK_EXIT(&
->a_rwlock
);
3889 err
= segvn_fault_anonpages(hat
, seg
,
3890 a
, a
+ maxpgsz
, type
, rw
,
3892 MIN(a
+ maxpgsz
, eaddr
), brkcow
);
3894 SEGVN_VMSTAT_FLTVNPAGES(6);
3897 if (szc
< seg
->s_szc
) {
3901 lpgeaddr
= maxlpgeaddr
;
3905 ASSERT(anon_pages(amp
->ahp
, aindx
,
3907 SEGVN_VMSTAT_FLTVNPAGES(7);
3908 anon_array_exit(&an_cookie
);
3909 ANON_LOCK_EXIT(&
->a_rwlock
);
3912 ASSERT(!brkcow
|| IS_P2ALIGNED(a
, maxpgsz
));
3913 ASSERT(!tron
|| IS_P2ALIGNED(a
, maxpgsz
));
3915 if (svd
->pageprot
!= 0 && IS_P2ALIGNED(a
, maxpgsz
)) {
3916 ASSERT(vpage
!= NULL
);
3917 prot
= VPP_PROT(vpage
);
3918 ASSERT(sameprot(seg
, a
, maxpgsz
));
3919 if ((prot
& protchk
) == 0) {
3920 SEGVN_VMSTAT_FLTVNPAGES(8);
3925 if (type
== F_SOFTLOCK
) {
3926 atomic_add_long((ulong_t
*)&svd
->softlockcnt
,
3933 if (!brkcow
&& !tron
&& szc
&&
3934 !page_exists_physcontig(vp
, off
, szc
,
3935 segtype
== MAP_PRIVATE
? ppa
: NULL
)) {
3936 SEGVN_VMSTAT_FLTVNPAGES(9);
3937 if (page_alloc_pages(vp
, seg
, a
, &pplist
, NULL
,
3938 szc
, 0, 0) && type
!= F_SOFTLOCK
) {
3939 SEGVN_VMSTAT_FLTVNPAGES(10);
3942 alloc_failed
|= (1 << szc
);
3945 if (pplist
!= NULL
&&
3946 vp
->v_mpssdata
== SEGVN_PAGEIO
) {
3948 SEGVN_VMSTAT_FLTVNPAGES(11);
3949 physcontig
= segvn_fill_vp_pages(svd
,
3950 vp
, off
, szc
, ppa
, &pplist
,
3952 ASSERT(!physcontig
|| pplist
== NULL
);
3953 if (!physcontig
&& downsize
&&
3954 type
!= F_SOFTLOCK
) {
3955 ASSERT(pplist
== NULL
);
3956 SEGVN_VMSTAT_FLTVNPAGES(12);
3960 ASSERT(!physcontig
||
3961 segtype
== MAP_PRIVATE
||
3963 if (physcontig
&& ppa
[0] == NULL
) {
3967 } else if (!brkcow
&& !tron
&& szc
&& ppa
[0] != NULL
) {
3968 SEGVN_VMSTAT_FLTVNPAGES(13);
3969 ASSERT(segtype
== MAP_PRIVATE
);
3974 SEGVN_VMSTAT_FLTVNPAGES(14);
3976 ierr
= VOP_GETPAGE(vp
, (offset_t
)off
, pgsz
,
3977 &vpprot
, ppa
, pgsz
, seg
, a
, arw
,
3981 for (i
= 0; i
< pages
; i
++) {
3982 ASSERT(PAGE_LOCKED(ppa
[i
]));
3983 ASSERT(!PP_ISFREE(ppa
[i
]));
3984 ASSERT(ppa
[i
]->p_vnode
== vp
);
3985 ASSERT(ppa
[i
]->p_offset
==
3986 off
+ (i
<< PAGESHIFT
));
3990 if (segtype
== MAP_PRIVATE
) {
3991 SEGVN_VMSTAT_FLTVNPAGES(15);
3992 vpprot
&= ~PROT_WRITE
;
3995 ASSERT(segtype
== MAP_PRIVATE
);
3996 SEGVN_VMSTAT_FLTVNPAGES(16);
3997 vpprot
= PROT_ALL
& ~PROT_WRITE
;
4002 SEGVN_VMSTAT_FLTVNPAGES(17);
4003 if (pplist
!= NULL
) {
4004 SEGVN_VMSTAT_FLTVNPAGES(18);
4005 page_free_replacement_page(pplist
);
4006 page_create_putback(pages
);
4008 SEGVN_RESTORE_SOFTLOCK_VP(type
, pages
);
4009 if (a
+ pgsz
<= eaddr
) {
4010 SEGVN_VMSTAT_FLTVNPAGES(19);
4011 err
= FC_MAKE_ERR(ierr
);
4014 va
.va_mask
= AT_SIZE
;
4015 if (VOP_GETATTR(vp
, &va
, 0, svd
->cred
, NULL
)) {
4016 SEGVN_VMSTAT_FLTVNPAGES(20);
4017 err
= FC_MAKE_ERR(EIO
);
4020 if (btopr(va
.va_size
) >= btopr(off
+ pgsz
)) {
4021 SEGVN_VMSTAT_FLTVNPAGES(21);
4022 err
= FC_MAKE_ERR(ierr
);
4025 if (btopr(va
.va_size
) <
4026 btopr(off
+ (eaddr
- a
))) {
4027 SEGVN_VMSTAT_FLTVNPAGES(22);
4028 err
= FC_MAKE_ERR(ierr
);
4031 if (brkcow
|| tron
|| type
== F_SOFTLOCK
) {
4032 /* can't reduce map area */
4033 SEGVN_VMSTAT_FLTVNPAGES(23);
4037 SEGVN_VMSTAT_FLTVNPAGES(24);
4045 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
4046 anon_array_enter(amp
, aindx
, &an_cookie
);
4049 anon_get_ptr(amp
->ahp
, aindx
) != NULL
) {
4050 ulong_t taindx
= P2ALIGN(aindx
, maxpages
);
4052 SEGVN_VMSTAT_FLTVNPAGES(25);
4053 ASSERT(anon_pages(amp
->ahp
, taindx
,
4054 maxpages
) == maxpages
);
4055 for (i
= 0; i
< pages
; i
++) {
4056 page_unlock(ppa
[i
]);
4058 anon_array_exit(&an_cookie
);
4059 ANON_LOCK_EXIT(&
->a_rwlock
);
4060 if (pplist
!= NULL
) {
4061 page_free_replacement_page(pplist
);
4062 page_create_putback(pages
);
4064 SEGVN_RESTORE_SOFTLOCK_VP(type
, pages
);
4065 if (szc
< seg
->s_szc
) {
4066 SEGVN_VMSTAT_FLTVNPAGES(26);
4068 * For private segments SOFTLOCK
4069 * either always breaks cow (any rw
4070 * type except S_READ_NOCOW) or
4071 * address space is locked as writer
4072 * (S_READ_NOCOW case) and anon slots
4073 * can't show up on second check.
4074 * Therefore if we are here for
4075 * SOFTLOCK case it must be a cow
4076 * break but cow break never reduces
4077 * szc. text replication (tron) in
4078 * this case works as cow break.
4079 * Thus the assert below.
4081 ASSERT(!brkcow
&& !tron
&&
4082 type
!= F_SOFTLOCK
);
4087 ASSERT(IS_P2ALIGNED(a
, maxpgsz
));
4092 ulong_t taindx
= P2ALIGN(aindx
, maxpages
);
4093 ASSERT(!anon_pages(amp
->ahp
, taindx
, maxpages
));
4097 if (brkcow
|| tron
) {
4098 ASSERT(amp
!= NULL
);
4099 ASSERT(pplist
== NULL
);
4100 ASSERT(szc
== seg
->s_szc
);
4101 ASSERT(IS_P2ALIGNED(a
, maxpgsz
));
4102 ASSERT(IS_P2ALIGNED(aindx
, maxpages
));
4103 SEGVN_VMSTAT_FLTVNPAGES(27);
4104 ierr
= anon_map_privatepages(amp
, aindx
, szc
,
4105 seg
, a
, prot
, ppa
, vpage
, segvn_anypgsz
,
4106 tron
? PG_LOCAL
: 0, svd
->cred
);
4108 SEGVN_VMSTAT_FLTVNPAGES(28);
4109 anon_array_exit(&an_cookie
);
4110 ANON_LOCK_EXIT(&
->a_rwlock
);
4111 SEGVN_RESTORE_SOFTLOCK_VP(type
, pages
);
4112 err
= FC_MAKE_ERR(ierr
);
4116 ASSERT(!IS_VMODSORT(ppa
[0]->p_vnode
));
4118 * p_szc can't be changed for locked
4121 ASSERT(svd
->rcookie
==
4122 HAT_INVALID_REGION_COOKIE
);
4123 hat_memload_array(hat
, a
, pgsz
, ppa
, prot
,
4126 if (!(hat_flag
& HAT_LOAD_LOCK
)) {
4127 SEGVN_VMSTAT_FLTVNPAGES(29);
4128 for (i
= 0; i
< pages
; i
++) {
4129 page_unlock(ppa
[i
]);
4132 anon_array_exit(&an_cookie
);
4133 ANON_LOCK_EXIT(&
->a_rwlock
);
4137 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
||
4138 (!svd
->pageprot
&& svd
->prot
== (prot
& vpprot
)));
4140 pfn
= page_pptonum(ppa
[0]);
4142 * hat_page_demote() needs an SE_EXCL lock on one of
4143 * constituent page_t's and it decreases root's p_szc
4144 * last. This means if root's p_szc is equal szc and
4145 * all its constituent pages are locked
4146 * hat_page_demote() that could have changed p_szc to
4147 * szc is already done and no new have page_demote()
4148 * can start for this large page.
4152 * we need to make sure same mapping size is used for
4153 * the same address range if there's a possibility the
4154 * adddress is already mapped because hat layer panics
4155 * when translation is loaded for the range already
4156 * mapped with a different page size. We achieve it
4157 * by always using largest page size possible subject
4158 * to the constraints of page size, segment page size
4159 * and page alignment. Since mappings are invalidated
4160 * when those constraints change and make it
4161 * impossible to use previously used mapping size no
4162 * mapping size conflicts should happen.
4166 if ((pszc
= ppa
[0]->p_szc
) == szc
&&
4167 IS_P2ALIGNED(pfn
, pages
)) {
4169 SEGVN_VMSTAT_FLTVNPAGES(30);
4171 for (i
= 0; i
< pages
; i
++) {
4172 ASSERT(PAGE_LOCKED(ppa
[i
]));
4173 ASSERT(!PP_ISFREE(ppa
[i
]));
4174 ASSERT(page_pptonum(ppa
[i
]) ==
4176 ASSERT(ppa
[i
]->p_szc
== szc
);
4177 ASSERT(ppa
[i
]->p_vnode
== vp
);
4178 ASSERT(ppa
[i
]->p_offset
==
4179 off
+ (i
<< PAGESHIFT
));
4183 * All pages are of szc we need and they are
4184 * all locked so they can't change szc. load
4187 * if page got promoted since last check
4188 * we don't need pplist.
4190 if (pplist
!= NULL
) {
4191 page_free_replacement_page(pplist
);
4192 page_create_putback(pages
);
4194 if (PP_ISMIGRATE(ppa
[0])) {
4195 page_migrate(seg
, a
, ppa
, pages
);
4197 SEGVN_UPDATE_MODBITS(ppa
, pages
, rw
,
4200 hat_memload_array_region(hat
, a
, pgsz
,
4201 ppa
, prot
& vpprot
, hat_flag
,
4205 * avoid large xhat mappings to FS
4206 * pages so that hat_page_demote()
4207 * doesn't need to check for xhat
4209 * Don't use regions with xhats.
4211 for (i
= 0; i
< pages
; i
++) {
4213 a
+ (i
<< PAGESHIFT
),
4214 ppa
[i
], prot
& vpprot
,
4219 if (!(hat_flag
& HAT_LOAD_LOCK
)) {
4220 for (i
= 0; i
< pages
; i
++) {
4221 page_unlock(ppa
[i
]);
4225 anon_array_exit(&an_cookie
);
4226 ANON_LOCK_EXIT(&
->a_rwlock
);
4232 * See if upsize is possible.
4234 if (pszc
> szc
&& szc
< seg
->s_szc
&&
4235 (segvn_anypgsz_vnode
|| pszc
>= seg
->s_szc
)) {
4237 uint_t pszc1
= MIN(pszc
, seg
->s_szc
);
4238 ppgsz
= page_get_pagesize(pszc1
);
4239 ppages
= btop(ppgsz
);
4240 aphase
= btop(P2PHASE((uintptr_t)a
, ppgsz
));
4242 ASSERT(type
!= F_SOFTLOCK
);
4244 SEGVN_VMSTAT_FLTVNPAGES(31);
4245 if (aphase
!= P2PHASE(pfn
, ppages
)) {
4246 segvn_faultvnmpss_align_err4
++;
4248 SEGVN_VMSTAT_FLTVNPAGES(32);
4249 if (pplist
!= NULL
) {
4250 page_t
*pl
= pplist
;
4251 page_free_replacement_page(pl
);
4252 page_create_putback(pages
);
4254 for (i
= 0; i
< pages
; i
++) {
4255 page_unlock(ppa
[i
]);
4258 anon_array_exit(&an_cookie
);
4259 ANON_LOCK_EXIT(&
->a_rwlock
);
4268 * check if we should use smallest mapping size.
4271 if (szc
== 0 || xhat
||
4273 !IS_P2ALIGNED(pfn
, pages
)) ||
4275 !segvn_full_szcpages(ppa
, szc
, &upgrdfail
,
4278 if (upgrdfail
&& type
!= F_SOFTLOCK
) {
4280 * segvn_full_szcpages failed to lock
4281 * all pages EXCL. Size down.
4285 SEGVN_VMSTAT_FLTVNPAGES(33);
4287 if (pplist
!= NULL
) {
4288 page_t
*pl
= pplist
;
4289 page_free_replacement_page(pl
);
4290 page_create_putback(pages
);
4293 for (i
= 0; i
< pages
; i
++) {
4294 page_unlock(ppa
[i
]);
4297 anon_array_exit(&an_cookie
);
4298 ANON_LOCK_EXIT(&
->a_rwlock
);
4303 if (szc
!= 0 && !xhat
&& !upgrdfail
) {
4304 segvn_faultvnmpss_align_err5
++;
4306 SEGVN_VMSTAT_FLTVNPAGES(34);
4307 if (pplist
!= NULL
) {
4308 page_free_replacement_page(pplist
);
4309 page_create_putback(pages
);
4311 SEGVN_UPDATE_MODBITS(ppa
, pages
, rw
,
4313 if (upgrdfail
&& segvn_anypgsz_vnode
) {
4315 hat_memload_array_region(hat
, a
, pgsz
,
4316 ppa
, prot
& vpprot
, hat_flag
,
4319 for (i
= 0; i
< pages
; i
++) {
4320 hat_memload_region(hat
,
4321 a
+ (i
<< PAGESHIFT
),
4322 ppa
[i
], prot
& vpprot
,
4323 hat_flag
, svd
->rcookie
);
4326 if (!(hat_flag
& HAT_LOAD_LOCK
)) {
4327 for (i
= 0; i
< pages
; i
++) {
4328 page_unlock(ppa
[i
]);
4332 anon_array_exit(&an_cookie
);
4333 ANON_LOCK_EXIT(&
->a_rwlock
);
4340 * segvn_full_szcpages() upgraded pages szc.
4342 ASSERT(pszc
== ppa
[0]->p_szc
);
4343 ASSERT(IS_P2ALIGNED(pfn
, pages
));
4349 SEGVN_VMSTAT_FLTVNPAGES(35);
4351 * p_szc of ppa[0] can change since we haven't
4352 * locked all constituent pages. Call
4353 * page_lock_szc() to prevent szc changes.
4354 * This should be a rare case that happens when
4355 * multiple segments use a different page size
4356 * to map the same file offsets.
4358 szcmtx
= page_szc_lock(ppa
[0]);
4359 pszc
= ppa
[0]->p_szc
;
4360 ASSERT(szcmtx
!= NULL
|| pszc
== 0);
4361 ASSERT(ppa
[0]->p_szc
<= pszc
);
4363 SEGVN_VMSTAT_FLTVNPAGES(36);
4364 if (szcmtx
!= NULL
) {
4369 if (pplist
!= NULL
) {
4371 * page got promoted since last check.
4372 * we don't need preaalocated large
4375 SEGVN_VMSTAT_FLTVNPAGES(37);
4376 page_free_replacement_page(pplist
);
4377 page_create_putback(pages
);
4379 SEGVN_UPDATE_MODBITS(ppa
, pages
, rw
,
4381 hat_memload_array_region(hat
, a
, pgsz
, ppa
,
4382 prot
& vpprot
, hat_flag
, svd
->rcookie
);
4384 if (!(hat_flag
& HAT_LOAD_LOCK
)) {
4385 for (i
= 0; i
< pages
; i
++) {
4386 page_unlock(ppa
[i
]);
4390 anon_array_exit(&an_cookie
);
4391 ANON_LOCK_EXIT(&
->a_rwlock
);
4397 * if page got demoted since last check
4398 * we could have not allocated larger page.
4401 if (pplist
== NULL
&&
4402 page_alloc_pages(vp
, seg
, a
, &pplist
, NULL
,
4403 szc
, 0, 0) && type
!= F_SOFTLOCK
) {
4404 SEGVN_VMSTAT_FLTVNPAGES(38);
4405 for (i
= 0; i
< pages
; i
++) {
4406 page_unlock(ppa
[i
]);
4409 anon_array_exit(&an_cookie
);
4410 ANON_LOCK_EXIT(&
->a_rwlock
);
4413 alloc_failed
|= (1 << szc
);
4417 SEGVN_VMSTAT_FLTVNPAGES(39);
4419 if (pplist
!= NULL
) {
4420 segvn_relocate_pages(ppa
, pplist
);
4423 ASSERT(type
== F_SOFTLOCK
);
4424 SEGVN_VMSTAT_FLTVNPAGES(40);
4428 SEGVN_UPDATE_MODBITS(ppa
, pages
, rw
, prot
, vpprot
);
4430 if (pplist
== NULL
&& segvn_anypgsz_vnode
== 0) {
4431 ASSERT(type
== F_SOFTLOCK
);
4432 for (i
= 0; i
< pages
; i
++) {
4433 ASSERT(ppa
[i
]->p_szc
< szc
);
4434 hat_memload_region(hat
,
4435 a
+ (i
<< PAGESHIFT
),
4436 ppa
[i
], prot
& vpprot
, hat_flag
,
4440 ASSERT(pplist
!= NULL
|| type
== F_SOFTLOCK
);
4441 hat_memload_array_region(hat
, a
, pgsz
, ppa
,
4442 prot
& vpprot
, hat_flag
, svd
->rcookie
);
4444 if (!(hat_flag
& HAT_LOAD_LOCK
)) {
4445 for (i
= 0; i
< pages
; i
++) {
4446 ASSERT(PAGE_SHARED(ppa
[i
]));
4447 page_unlock(ppa
[i
]);
4451 anon_array_exit(&an_cookie
);
4452 ANON_LOCK_EXIT(&
->a_rwlock
);
4456 if (vpage
!= NULL
) {
4463 ASSERT(a
< lpgeaddr
);
4465 ASSERT(!brkcow
&& !tron
&& type
!= F_SOFTLOCK
);
4468 * ierr == -1 means we failed to map with a large page.
4469 * (either due to allocation/relocation failures or
4470 * misalignment with other mappings to this file.
4472 * ierr == -2 means some other thread allocated a large page
4473 * after we gave up tp map with a large page. retry with
4476 ASSERT(ierr
== -1 || ierr
== -2);
4477 ASSERT(ierr
== -2 || szc
!= 0);
4478 ASSERT(ierr
== -1 || szc
< seg
->s_szc
);
4480 SEGVN_VMSTAT_FLTVNPAGES(41);
4481 ASSERT(pszc
> szc
&& pszc
<= seg
->s_szc
);
4483 } else if (segvn_anypgsz_vnode
) {
4484 SEGVN_VMSTAT_FLTVNPAGES(42);
4487 SEGVN_VMSTAT_FLTVNPAGES(43);
4490 * other process created pszc large page.
4491 * but we still have to drop to 0 szc.
4496 pgsz
= page_get_pagesize(szc
);
4500 * Size up case. Note lpgaddr may only be needed for
4501 * softlock case so we don't adjust it here.
4503 a
= (caddr_t
)P2ALIGN((uintptr_t)a
, pgsz
);
4504 ASSERT(a
>= lpgaddr
);
4505 lpgeaddr
= (caddr_t
)P2ROUNDUP((uintptr_t)eaddr
, pgsz
);
4506 off
= svd
->offset
+ (uintptr_t)(a
- seg
->s_base
);
4507 aindx
= svd
->anon_index
+ seg_page(seg
, a
);
4508 vpage
= (svd
->vpage
!= NULL
) ?
4509 &svd
->vpage
[seg_page(seg
, a
)] : NULL
;
4512 * Size down case. Note lpgaddr may only be needed for
4513 * softlock case so we don't adjust it here.
4515 ASSERT(IS_P2ALIGNED(a
, pgsz
));
4516 ASSERT(IS_P2ALIGNED(lpgeaddr
, pgsz
));
4517 lpgeaddr
= (caddr_t
)P2ROUNDUP((uintptr_t)eaddr
, pgsz
);
4518 ASSERT(a
< lpgeaddr
);
4520 SEGVN_VMSTAT_FLTVNPAGES(44);
4522 * The beginning of the large page region can
4523 * be pulled to the right to make a smaller
4524 * region. We haven't yet faulted a single
4527 a
= (caddr_t
)P2ALIGN((uintptr_t)addr
, pgsz
);
4528 ASSERT(a
>= lpgaddr
);
4530 (uintptr_t)(a
- seg
->s_base
);
4531 aindx
= svd
->anon_index
+ seg_page(seg
, a
);
4532 vpage
= (svd
->vpage
!= NULL
) ?
4533 &svd
->vpage
[seg_page(seg
, a
)] : NULL
;
4538 kmem_free(ppa
, ppasize
);
4539 if (!err
&& !vop_size_err
) {
4540 SEGVN_VMSTAT_FLTVNPAGES(45);
4543 if (type
== F_SOFTLOCK
&& a
> lpgaddr
) {
4544 SEGVN_VMSTAT_FLTVNPAGES(46);
4545 segvn_softunlock(seg
, lpgaddr
, a
- lpgaddr
, S_OTHER
);
4547 if (!vop_size_err
) {
4548 SEGVN_VMSTAT_FLTVNPAGES(47);
4551 ASSERT(brkcow
|| tron
|| type
== F_SOFTLOCK
);
4553 * Large page end is mapped beyond the end of file and it's a cow
4554 * fault (can be a text replication induced cow) or softlock so we can't
4555 * reduce the map area. For now just demote the segment. This should
4556 * really only happen if the end of the file changed after the mapping
4557 * was established since when large page segments are created we make
4558 * sure they don't extend beyond the end of the file.
4560 SEGVN_VMSTAT_FLTVNPAGES(48);
4562 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
4563 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_WRITER
);
4565 if (seg
->s_szc
!= 0) {
4566 segvn_fltvnpages_clrszc_cnt
++;
4567 ASSERT(svd
->softlockcnt
== 0);
4568 err
= segvn_clrszc(seg
);
4570 segvn_fltvnpages_clrszc_err
++;
4573 ASSERT(err
|| seg
->s_szc
== 0);
4574 SEGVN_LOCK_DOWNGRADE(seg
->s_as
, &svd
->lock
);
4575 /* segvn_fault will do its job as if szc had been zero to begin with */
4576 return (err
== 0 ? IE_RETRY
: FC_MAKE_ERR(err
));
4580 * This routine will attempt to fault in one large page.
4581 * it will use smaller pages if that fails.
4582 * It should only be called for pure anonymous segments.
4585 segvn_fault_anonpages(struct hat
*hat
, struct seg
*seg
, caddr_t lpgaddr
,
4586 caddr_t lpgeaddr
, enum fault_type type
, enum seg_rw rw
, caddr_t addr
,
4587 caddr_t eaddr
, int brkcow
)
4589 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
4590 struct anon_map
*amp
= svd
->amp
;
4591 uchar_t segtype
= svd
->type
;
4592 uint_t szc
= seg
->s_szc
;
4593 size_t pgsz
= page_get_pagesize(szc
);
4594 size_t maxpgsz
= pgsz
;
4595 pgcnt_t pages
= btop(pgsz
);
4596 uint_t ppaszc
= szc
;
4597 caddr_t a
= lpgaddr
;
4598 ulong_t aindx
= svd
->anon_index
+ seg_page(seg
, a
);
4599 struct vpage
*vpage
= (svd
->vpage
!= NULL
) ?
4600 &svd
->vpage
[seg_page(seg
, a
)] : NULL
;
4605 uint_t protchk
, prot
, vpprot
;
4607 int hat_flag
= (type
== F_SOFTLOCK
) ? HAT_LOAD_LOCK
: HAT_LOAD
;
4608 anon_sync_obj_t cookie
;
4610 int pgflags
= (svd
->tr_state
== SEGVN_TR_ON
) ? PG_LOCAL
: 0;
4613 ASSERT(amp
!= NULL
);
4614 ASSERT(enable_mbit_wa
== 0); /* no mbit simulations with large pages */
4615 ASSERT(!(svd
->flags
& MAP_NORESERVE
));
4616 ASSERT(type
!= F_SOFTUNLOCK
);
4617 ASSERT(IS_P2ALIGNED(a
, maxpgsz
));
4618 ASSERT(!brkcow
|| svd
->tr_state
== SEGVN_TR_OFF
);
4619 ASSERT(svd
->tr_state
!= SEGVN_TR_INIT
);
4621 ASSERT(SEGVN_LOCK_HELD(seg
->s_as
, &svd
->lock
));
4623 VM_STAT_COND_ADD(type
== F_SOFTLOCK
, segvnvmstats
.fltanpages
[0]);
4624 VM_STAT_COND_ADD(type
!= F_SOFTLOCK
, segvnvmstats
.fltanpages
[1]);
4626 if (svd
->flags
& MAP_TEXT
) {
4627 hat_flag
|= HAT_LOAD_TEXT
;
4630 if (svd
->pageprot
) {
4633 protchk
= PROT_READ
;
4636 protchk
= PROT_WRITE
;
4639 protchk
= PROT_EXEC
;
4643 protchk
= PROT_READ
| PROT_WRITE
| PROT_EXEC
;
4646 VM_STAT_ADD(segvnvmstats
.fltanpages
[2]);
4649 /* caller has already done segment level protection check. */
4652 ppa
= kmem_cache_alloc(segvn_szc_cache
[ppaszc
], KM_SLEEP
);
4653 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
4656 for (; a
< lpgeaddr
; a
+= pgsz
, aindx
+= pages
) {
4657 if (svd
->pageprot
!= 0 && IS_P2ALIGNED(a
, maxpgsz
)) {
4658 VM_STAT_ADD(segvnvmstats
.fltanpages
[3]);
4659 ASSERT(vpage
!= NULL
);
4660 prot
= VPP_PROT(vpage
);
4661 ASSERT(sameprot(seg
, a
, maxpgsz
));
4662 if ((prot
& protchk
) == 0) {
4667 if (adjszc_chk
&& IS_P2ALIGNED(a
, maxpgsz
) &&
4669 ASSERT(a
> lpgaddr
);
4673 ASSERT(IS_P2ALIGNED(aindx
, pages
));
4674 lpgeaddr
= (caddr_t
)P2ROUNDUP((uintptr_t)eaddr
,
4677 if (type
== F_SOFTLOCK
) {
4678 atomic_add_long((ulong_t
*)&svd
->softlockcnt
,
4681 anon_array_enter(amp
, aindx
, &cookie
);
4682 ppa_szc
= (uint_t
)-1;
4683 ierr
= anon_map_getpages(amp
, aindx
, szc
, seg
, a
,
4684 prot
, &vpprot
, ppa
, &ppa_szc
, vpage
, rw
, brkcow
,
4685 segvn_anypgsz
, pgflags
, svd
->cred
);
4687 anon_array_exit(&cookie
);
4688 VM_STAT_ADD(segvnvmstats
.fltanpages
[4]);
4689 if (type
== F_SOFTLOCK
) {
4691 (ulong_t
*)&svd
->softlockcnt
,
4695 VM_STAT_ADD(segvnvmstats
.fltanpages
[6]);
4696 err
= FC_MAKE_ERR(ierr
);
4702 ASSERT(!IS_VMODSORT(ppa
[0]->p_vnode
));
4704 ASSERT(segtype
== MAP_SHARED
||
4705 ppa
[0]->p_szc
<= szc
);
4706 ASSERT(segtype
== MAP_PRIVATE
||
4707 ppa
[0]->p_szc
>= szc
);
4710 * Handle pages that have been marked for migration
4712 if (lgrp_optimizations())
4713 page_migrate(seg
, a
, ppa
, pages
);
4715 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
4717 if (segtype
== MAP_SHARED
) {
4718 vpprot
|= PROT_WRITE
;
4721 hat_memload_array(hat
, a
, pgsz
, ppa
,
4722 prot
& vpprot
, hat_flag
);
4724 if (hat_flag
& HAT_LOAD_LOCK
) {
4725 VM_STAT_ADD(segvnvmstats
.fltanpages
[7]);
4727 VM_STAT_ADD(segvnvmstats
.fltanpages
[8]);
4728 for (i
= 0; i
< pages
; i
++)
4729 page_unlock(ppa
[i
]);
4734 anon_array_exit(&cookie
);
4739 ASSERT(a
< lpgeaddr
);
4741 * ierr == -1 means we failed to allocate a large page.
4742 * so do a size down operation.
4744 * ierr == -2 means some other process that privately shares
4745 * pages with this process has allocated a larger page and we
4746 * need to retry with larger pages. So do a size up
4747 * operation. This relies on the fact that large pages are
4748 * never partially shared i.e. if we share any constituent
4749 * page of a large page with another process we must share the
4750 * entire large page. Note this cannot happen for SOFTLOCK
4751 * case, unless current address (a) is at the beginning of the
4752 * next page size boundary because the other process couldn't
4753 * have relocated locked pages.
4755 ASSERT(ierr
== -1 || ierr
== -2);
4757 if (segvn_anypgsz
) {
4758 ASSERT(ierr
== -2 || szc
!= 0);
4759 ASSERT(ierr
== -1 || szc
< seg
->s_szc
);
4760 szc
= (ierr
== -1) ? szc
- 1 : szc
+ 1;
4763 * For non COW faults and segvn_anypgsz == 0
4764 * we need to be careful not to loop forever
4765 * if existing page is found with szc other
4766 * than 0 or seg->s_szc. This could be due
4767 * to page relocations on behalf of DR or
4768 * more likely large page creation. For this
4769 * case simply re-size to existing page's szc
4770 * if returned by anon_map_getpages().
4772 if (ppa_szc
== (uint_t
)-1) {
4773 szc
= (ierr
== -1) ? 0 : seg
->s_szc
;
4775 ASSERT(ppa_szc
<= seg
->s_szc
);
4776 ASSERT(ierr
== -2 || ppa_szc
< szc
);
4777 ASSERT(ierr
== -1 || ppa_szc
> szc
);
4782 pgsz
= page_get_pagesize(szc
);
4784 ASSERT(type
!= F_SOFTLOCK
|| ierr
== -1 ||
4785 (IS_P2ALIGNED(a
, pgsz
) && IS_P2ALIGNED(lpgeaddr
, pgsz
)));
4786 if (type
== F_SOFTLOCK
) {
4788 * For softlocks we cannot reduce the fault area
4789 * (calculated based on the largest page size for this
4790 * segment) for size down and a is already next
4791 * page size aligned as assertted above for size
4792 * ups. Therefore just continue in case of softlock.
4794 VM_STAT_ADD(segvnvmstats
.fltanpages
[9]);
4795 continue; /* keep lint happy */
4796 } else if (ierr
== -2) {
4799 * Size up case. Note lpgaddr may only be needed for
4800 * softlock case so we don't adjust it here.
4802 VM_STAT_ADD(segvnvmstats
.fltanpages
[10]);
4803 a
= (caddr_t
)P2ALIGN((uintptr_t)a
, pgsz
);
4804 ASSERT(a
>= lpgaddr
);
4805 lpgeaddr
= (caddr_t
)P2ROUNDUP((uintptr_t)eaddr
, pgsz
);
4806 aindx
= svd
->anon_index
+ seg_page(seg
, a
);
4807 vpage
= (svd
->vpage
!= NULL
) ?
4808 &svd
->vpage
[seg_page(seg
, a
)] : NULL
;
4811 * Size down case. Note lpgaddr may only be needed for
4812 * softlock case so we don't adjust it here.
4814 VM_STAT_ADD(segvnvmstats
.fltanpages
[11]);
4815 ASSERT(IS_P2ALIGNED(a
, pgsz
));
4816 ASSERT(IS_P2ALIGNED(lpgeaddr
, pgsz
));
4817 lpgeaddr
= (caddr_t
)P2ROUNDUP((uintptr_t)eaddr
, pgsz
);
4818 ASSERT(a
< lpgeaddr
);
4821 * The beginning of the large page region can
4822 * be pulled to the right to make a smaller
4823 * region. We haven't yet faulted a single
4826 VM_STAT_ADD(segvnvmstats
.fltanpages
[12]);
4827 a
= (caddr_t
)P2ALIGN((uintptr_t)addr
, pgsz
);
4828 ASSERT(a
>= lpgaddr
);
4829 aindx
= svd
->anon_index
+ seg_page(seg
, a
);
4830 vpage
= (svd
->vpage
!= NULL
) ?
4831 &svd
->vpage
[seg_page(seg
, a
)] : NULL
;
4835 VM_STAT_ADD(segvnvmstats
.fltanpages
[13]);
4836 ANON_LOCK_EXIT(&
->a_rwlock
);
4837 kmem_cache_free(segvn_szc_cache
[ppaszc
], ppa
);
4840 VM_STAT_ADD(segvnvmstats
.fltanpages
[14]);
4841 ANON_LOCK_EXIT(&
->a_rwlock
);
4842 kmem_cache_free(segvn_szc_cache
[ppaszc
], ppa
);
4843 if (type
== F_SOFTLOCK
&& a
> lpgaddr
) {
4844 VM_STAT_ADD(segvnvmstats
.fltanpages
[15]);
4845 segvn_softunlock(seg
, lpgaddr
, a
- lpgaddr
, S_OTHER
);
4850 int fltadvice
= 1; /* set to free behind pages for sequential access */
4853 * This routine is called via a machine specific fault handling routine.
4854 * It is also called by software routines wishing to lock or unlock
4855 * a range of addresses.
4857 * Here is the basic algorithm:
4859 * Call segvn_softunlock
4862 * Checking and set up work
4863 * If we will need some non-anonymous pages
4864 * Call VOP_GETPAGE over the range of non-anonymous pages
4866 * Loop over all addresses requested
4867 * Call segvn_faultpage passing in page list
4868 * to load up translations and handle anonymous pages
4870 * Load up translation to any additional pages in page list not
4871 * already handled that fit into this segment
4874 segvn_fault(struct hat
*hat
, struct seg
*seg
, caddr_t addr
, size_t len
,
4875 enum fault_type type
, enum seg_rw rw
)
4877 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
4878 page_t
**plp
, **ppp
, *pp
;
4881 struct vpage
*vpage
;
4882 uint_t vpprot
, prot
;
4884 page_t
*pl
[PVN_GETPAGE_NUM
+ 1];
4885 size_t plsz
, pl_alloc_sz
;
4888 struct anon_map
*amp
;
4890 caddr_t lpgaddr
, lpgeaddr
;
4892 anon_sync_obj_t cookie
;
4893 int brkcow
= BREAK_COW_SHARE(rw
, type
, svd
->type
);
4895 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
4896 ASSERT(svd
->amp
== NULL
|| svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
4899 * First handle the easy stuff
4901 if (type
== F_SOFTUNLOCK
) {
4902 if (rw
== S_READ_NOCOW
) {
4904 ASSERT(AS_WRITE_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
4906 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_READER
);
4907 pgsz
= (seg
->s_szc
== 0) ? PAGESIZE
:
4908 page_get_pagesize(seg
->s_szc
);
4909 VM_STAT_COND_ADD(pgsz
> PAGESIZE
, segvnvmstats
.fltanpages
[16]);
4910 CALC_LPG_REGION(pgsz
, seg
, addr
, len
, lpgaddr
, lpgeaddr
);
4911 segvn_softunlock(seg
, lpgaddr
, lpgeaddr
- lpgaddr
, rw
);
4912 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
4916 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
||
4917 !HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
));
4919 if (svd
->tr_state
== SEGVN_TR_INIT
) {
4920 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_WRITER
);
4921 if (svd
->tr_state
== SEGVN_TR_INIT
) {
4922 ASSERT(svd
->vp
!= NULL
&& svd
->amp
== NULL
);
4923 ASSERT(svd
->flags
& MAP_TEXT
);
4924 ASSERT(svd
->type
== MAP_PRIVATE
);
4925 segvn_textrepl(seg
);
4926 ASSERT(svd
->tr_state
!= SEGVN_TR_INIT
);
4927 ASSERT(svd
->tr_state
!= SEGVN_TR_ON
||
4930 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
4932 } else if (svd
->tr_state
!= SEGVN_TR_OFF
) {
4933 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_WRITER
);
4935 if (rw
== S_WRITE
&& svd
->tr_state
!= SEGVN_TR_OFF
) {
4936 ASSERT(!svd
->pageprot
&& !(svd
->prot
& PROT_WRITE
));
4937 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
4941 if (svd
->tr_state
== SEGVN_TR_ON
) {
4942 ASSERT(svd
->vp
!= NULL
&& svd
->amp
!= NULL
);
4943 segvn_textunrepl(seg
, 0);
4944 ASSERT(svd
->amp
== NULL
&&
4945 svd
->tr_state
== SEGVN_TR_OFF
);
4946 } else if (svd
->tr_state
!= SEGVN_TR_OFF
) {
4947 svd
->tr_state
= SEGVN_TR_OFF
;
4949 ASSERT(svd
->amp
== NULL
&& svd
->tr_state
== SEGVN_TR_OFF
);
4950 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
4954 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_READER
);
4957 * If we have the same protections for the entire segment,
4958 * insure that the access being attempted is legitimate.
4961 if (svd
->pageprot
== 0) {
4967 protchk
= PROT_READ
;
4970 protchk
= PROT_WRITE
;
4973 protchk
= PROT_EXEC
;
4977 protchk
= PROT_READ
| PROT_WRITE
| PROT_EXEC
;
4981 if ((svd
->prot
& protchk
) == 0) {
4982 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
4983 return (FC_PROT
); /* illegal access type */
4987 if (brkcow
&& HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
)) {
4988 /* this must be SOFTLOCK S_READ fault */
4989 ASSERT(svd
->amp
== NULL
);
4990 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
4991 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
4992 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_WRITER
);
4993 if (HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
)) {
4995 * this must be the first ever non S_READ_NOCOW
4996 * softlock for this segment.
4998 ASSERT(svd
->softlockcnt
== 0);
4999 hat_leave_region(seg
->s_as
->a_hat
, svd
->rcookie
,
5001 svd
->rcookie
= HAT_INVALID_REGION_COOKIE
;
5003 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5008 * We can't allow the long term use of softlocks for vmpss segments,
5009 * because in some file truncation cases we should be able to demote
5010 * the segment, which requires that there are no softlocks. The
5011 * only case where it's ok to allow a SOFTLOCK fault against a vmpss
5012 * segment is S_READ_NOCOW, where the caller holds the address space
5013 * locked as writer and calls softunlock before dropping the as lock.
5014 * S_READ_NOCOW is used by /proc to read memory from another user.
5016 * Another deadlock between SOFTLOCK and file truncation can happen
5017 * because segvn_fault_vnodepages() calls the FS one pagesize at
5018 * a time. A second VOP_GETPAGE() call by segvn_fault_vnodepages()
5019 * can cause a deadlock because the first set of page_t's remain
5020 * locked SE_SHARED. To avoid this, we demote segments on a first
5021 * SOFTLOCK if they have a length greater than the segment's
5024 * So for now, we only avoid demoting a segment on a SOFTLOCK when
5025 * the access type is S_READ_NOCOW and the fault length is less than
5026 * or equal to the segment's page size. While this is quite restrictive,
5027 * it should be the most common case of SOFTLOCK against a vmpss
5030 * For S_READ_NOCOW, it's safe not to do a copy on write because the
5031 * caller makes sure no COW will be caused by another thread for a
5034 if (type
== F_SOFTLOCK
&& svd
->vp
!= NULL
&& seg
->s_szc
!= 0) {
5037 if (rw
!= S_READ_NOCOW
) {
5040 if (!demote
&& len
> PAGESIZE
) {
5041 pgsz
= page_get_pagesize(seg
->s_szc
);
5042 CALC_LPG_REGION(pgsz
, seg
, addr
, len
, lpgaddr
,
5044 if (lpgeaddr
- lpgaddr
> pgsz
) {
5049 ASSERT(demote
|| AS_WRITE_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
5052 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5053 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_WRITER
);
5054 if (seg
->s_szc
!= 0) {
5055 segvn_vmpss_clrszc_cnt
++;
5056 ASSERT(svd
->softlockcnt
== 0);
5057 err
= segvn_clrszc(seg
);
5059 segvn_vmpss_clrszc_err
++;
5060 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5061 return (FC_MAKE_ERR(err
));
5064 ASSERT(seg
->s_szc
== 0);
5065 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5071 * Check to see if we need to allocate an anon_map structure.
5073 if (svd
->amp
== NULL
&& (svd
->vp
== NULL
|| brkcow
)) {
5074 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
5076 * Drop the "read" lock on the segment and acquire
5077 * the "write" version since we have to allocate the
5080 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5081 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_WRITER
);
5083 if (svd
->amp
== NULL
) {
5084 svd
->amp
= anonmap_alloc(seg
->s_size
, 0, ANON_SLEEP
);
5085 svd
->amp
->a_szc
= seg
->s_szc
;
5087 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5090 * Start all over again since segment protections
5091 * may have changed after we dropped the "read" lock.
5097 * S_READ_NOCOW vs S_READ distinction was
5098 * only needed for the code above. After
5099 * that we treat it as S_READ.
5101 if (rw
== S_READ_NOCOW
) {
5102 ASSERT(type
== F_SOFTLOCK
);
5103 ASSERT(AS_WRITE_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
5110 * MADV_SEQUENTIAL work is ignored for large page segments.
5112 if (seg
->s_szc
!= 0) {
5113 pgsz
= page_get_pagesize(seg
->s_szc
);
5114 ASSERT(SEGVN_LOCK_HELD(seg
->s_as
, &svd
->lock
));
5115 CALC_LPG_REGION(pgsz
, seg
, addr
, len
, lpgaddr
, lpgeaddr
);
5116 if (svd
->vp
== NULL
) {
5117 err
= segvn_fault_anonpages(hat
, seg
, lpgaddr
,
5118 lpgeaddr
, type
, rw
, addr
, addr
+ len
, brkcow
);
5120 err
= segvn_fault_vnodepages(hat
, seg
, lpgaddr
,
5121 lpgeaddr
, type
, rw
, addr
, addr
+ len
, brkcow
);
5122 if (err
== IE_RETRY
) {
5123 ASSERT(seg
->s_szc
== 0);
5124 ASSERT(SEGVN_READ_HELD(seg
->s_as
, &svd
->lock
));
5125 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5129 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5133 page
= seg_page(seg
, addr
);
5135 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
5136 anon_index
= svd
->anon_index
+ page
;
5138 if (type
== F_PROT
&& rw
== S_READ
&&
5139 svd
->tr_state
== SEGVN_TR_OFF
&&
5140 svd
->type
== MAP_PRIVATE
&& svd
->pageprot
== 0) {
5141 size_t index
= anon_index
;
5144 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
5146 * The fast path could apply to S_WRITE also, except
5147 * that the protection fault could be caused by lazy
5148 * tlb flush when ro->rw. In this case, the pte is
5149 * RW already. But RO in the other cpu's tlb causes
5150 * the fault. Since hat_chgprot won't do anything if
5151 * pte doesn't change, we may end up faulting
5152 * indefinitely until the RO tlb entry gets replaced.
5154 for (a
= addr
; a
< addr
+ len
; a
+= PAGESIZE
, index
++) {
5155 anon_array_enter(amp
, index
, &cookie
);
5156 ap
= anon_get_ptr(amp
->ahp
, index
);
5157 anon_array_exit(&cookie
);
5158 if ((ap
== NULL
) || (ap
->an_refcnt
!= 1)) {
5159 ANON_LOCK_EXIT(&
->a_rwlock
);
5163 hat_chgprot(seg
->s_as
->a_hat
, addr
, len
, svd
->prot
);
5164 ANON_LOCK_EXIT(&
->a_rwlock
);
5165 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5171 if (svd
->vpage
== NULL
)
5174 vpage
= &svd
->vpage
[page
];
5176 off
= svd
->offset
+ (uintptr_t)(addr
- seg
->s_base
);
5179 * If MADV_SEQUENTIAL has been set for the particular page we
5180 * are faulting on, free behind all pages in the segment and put
5181 * them on the free list.
5184 if ((page
!= 0) && fltadvice
&& svd
->tr_state
!= SEGVN_TR_ON
) {
5186 ulong_t fanon_index
;
5188 u_offset_t pgoff
, fpgoff
;
5190 struct anon
*fap
= NULL
;
5192 if (svd
->advice
== MADV_SEQUENTIAL
||
5194 VPP_ADVICE(vpage
) == MADV_SEQUENTIAL
)) {
5195 pgoff
= off
- PAGESIZE
;
5198 vpp
= &svd
->vpage
[fpage
];
5200 fanon_index
= svd
->anon_index
+ fpage
;
5202 while (pgoff
> svd
->offset
) {
5203 if (svd
->advice
!= MADV_SEQUENTIAL
&&
5204 (!svd
->pageadvice
|| (vpage
&&
5205 VPP_ADVICE(vpp
) != MADV_SEQUENTIAL
)))
5209 * If this is an anon page, we must find the
5210 * correct <vp, offset> for it
5214 ANON_LOCK_ENTER(&
->a_rwlock
,
5216 anon_array_enter(amp
, fanon_index
,
5218 fap
= anon_get_ptr(amp
->ahp
,
5221 swap_xlate(fap
, &fvp
, &fpgoff
);
5226 anon_array_exit(&cookie
);
5227 ANON_LOCK_EXIT(&
->a_rwlock
);
5235 * Skip pages that are free or have an
5238 pp
= page_lookup_nowait(fvp
, fpgoff
, SE_SHARED
);
5242 * We don't need the page_struct_lock to test
5243 * as this is only advisory; even if we
5244 * acquire it someone might race in and lock
5245 * the page after we unlock and before the
5246 * PUTPAGE, then VOP_PUTPAGE will do nothing.
5248 if (pp
->p_lckcnt
== 0 && pp
->p_cowcnt
== 0) {
5250 * Hold the vnode before releasing
5251 * the page lock to prevent it from
5252 * being freed and re-used by some
5258 * We should build a page list
5259 * to kluster putpages XXX
5261 (void) VOP_PUTPAGE(fvp
,
5262 (offset_t
)fpgoff
, PAGESIZE
,
5263 (B_DONTNEED
|B_FREE
|B_ASYNC
),
5268 * XXX - Should the loop terminate if
5269 * the page is `locked'?
5285 * See if we need to call VOP_GETPAGE for
5286 * *any* of the range being faulted on.
5287 * We can skip all of this work if there
5288 * was no original vnode.
5290 if (svd
->vp
!= NULL
) {
5303 * Only acquire reader lock to prevent amp->ahp
5304 * from being changed. It's ok to miss pages,
5305 * hence we don't do anon_array_enter
5307 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
5308 ap
= anon_get_ptr(amp
->ahp
, anon_index
);
5310 if (len
<= PAGESIZE
)
5311 /* inline non_anon() */
5312 dogetpage
= (ap
== NULL
);
5314 dogetpage
= non_anon(amp
->ahp
, anon_index
,
5316 ANON_LOCK_EXIT(&
->a_rwlock
);
5321 struct as
*as
= seg
->s_as
;
5323 if (len
> ptob((sizeof (pl
) / sizeof (pl
[0])) - 1)) {
5325 * Page list won't fit in local array,
5326 * allocate one of the needed size.
5329 (btop(len
) + 1) * sizeof (page_t
*);
5330 plp
= kmem_alloc(pl_alloc_sz
, KM_SLEEP
);
5333 } else if (rw
== S_WRITE
&& svd
->type
== MAP_PRIVATE
||
5334 svd
->tr_state
== SEGVN_TR_ON
|| rw
== S_OTHER
||
5335 (((size_t)(addr
+ PAGESIZE
) <
5336 (size_t)(seg
->s_base
+ seg
->s_size
)) &&
5337 hat_probe(as
->a_hat
, addr
+ PAGESIZE
))) {
5339 * Ask VOP_GETPAGE to return the exact number
5341 * (a) this is a COW fault, or
5342 * (b) this is a software fault, or
5343 * (c) next page is already mapped.
5348 * Ask VOP_GETPAGE to return adjacent pages
5349 * within the segment.
5351 plsz
= MIN((size_t)PVN_GETPAGE_SZ
, (size_t)
5352 ((seg
->s_base
+ seg
->s_size
) - addr
));
5353 ASSERT((addr
+ plsz
) <=
5354 (seg
->s_base
+ seg
->s_size
));
5358 * Need to get some non-anonymous pages.
5359 * We need to make only one call to GETPAGE to do
5360 * this to prevent certain deadlocking conditions
5361 * when we are doing locking. In this case
5362 * non_anon() should have picked up the smallest
5363 * range which includes all the non-anonymous
5364 * pages in the requested range. We have to
5365 * be careful regarding which rw flag to pass in
5366 * because on a private mapping, the underlying
5367 * object is never allowed to be written.
5369 if (rw
== S_WRITE
&& svd
->type
== MAP_PRIVATE
) {
5375 TRACE_3(TR_FAC_VM
, TR_SEGVN_GETPAGE
,
5376 "segvn_getpage:seg %p addr %p vp %p",
5378 err
= VOP_GETPAGE(vp
, (offset_t
)vp_off
, vp_len
,
5379 &vpprot
, plp
, plsz
, seg
, addr
+ (vp_off
- off
), arw
,
5382 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5383 segvn_pagelist_rele(plp
);
5385 kmem_free(plp
, pl_alloc_sz
);
5386 return (FC_MAKE_ERR(err
));
5388 if (svd
->type
== MAP_PRIVATE
)
5389 vpprot
&= ~PROT_WRITE
;
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.
5399 * Always acquire the anon_array_lock to prevent
5400 * 2 threads from allocating separate anon slots for
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.
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
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.
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.
5426 * But really, the page lock should be downgraded after
5427 * the stolen page is page_rename'd.
5431 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
5434 * Ok, now loop over the address range and handle faults
5436 for (a
= addr
; a
< addr
+ len
; a
+= PAGESIZE
, off
+= PAGESIZE
) {
5437 err
= segvn_faultpage(hat
, seg
, a
, off
, vpage
, plp
, vpprot
,
5441 ANON_LOCK_EXIT(&
->a_rwlock
);
5442 if (type
== F_SOFTLOCK
&& a
> addr
) {
5443 segvn_softunlock(seg
, addr
, (a
- addr
),
5446 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5447 segvn_pagelist_rele(plp
);
5449 kmem_free(plp
, pl_alloc_sz
);
5454 } else if (svd
->vpage
) {
5455 page
= seg_page(seg
, addr
);
5456 vpage
= &svd
->vpage
[++page
];
5460 /* Didn't get pages from the underlying fs so we're done */
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 * VOP_GETPAGE() was called and vpprot has been initialized.
5471 if (svd
->pageprot
== 0)
5472 prot
= svd
->prot
& vpprot
;
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.
5482 for (ppp
= plp
; (pp
= *ppp
) != NULL
; ppp
++) {
5486 anon_sync_obj_t cookie
;
5487 int hat_flag
= HAT_LOAD_ADV
;
5489 if (svd
->flags
& MAP_TEXT
) {
5490 hat_flag
|= HAT_LOAD_TEXT
;
5493 if (pp
== PAGE_HANDLED
)
5496 if (svd
->tr_state
!= SEGVN_TR_ON
&&
5497 pp
->p_offset
>= svd
->offset
&&
5498 pp
->p_offset
< svd
->offset
+ seg
->s_size
) {
5500 diff
= pp
->p_offset
- svd
->offset
;
5503 * Large Files: Following is the assertion
5504 * validating the above cast.
5506 ASSERT(svd
->vp
== pp
->p_vnode
);
5510 prot
= VPP_PROT(&svd
->vpage
[page
]) & vpprot
;
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
5520 anon_index
= svd
->anon_index
+ page
;
5521 anon_array_enter(amp
, anon_index
, &cookie
);
5522 ap
= anon_get_ptr(amp
->ahp
, anon_index
);
5524 if ((amp
== NULL
) || (ap
== NULL
)) {
5525 if (IS_VMODSORT(pp
->p_vnode
) ||
5529 else if (rw
!= S_OTHER
&&
5531 prot
&= ~PROT_WRITE
;
5534 * Skip mapping read ahead pages marked
5535 * for migration, so they will get migrated
5538 ASSERT(amp
== NULL
||
5539 svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
5540 if ((prot
& PROT_READ
) && !PP_ISMIGRATE(pp
)) {
5541 hat_memload_region(hat
,
5548 anon_array_exit(&cookie
);
5554 ANON_LOCK_EXIT(&
->a_rwlock
);
5555 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5557 kmem_free(plp
, pl_alloc_sz
);
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
5567 segvn_faulta(struct seg
*seg
, caddr_t addr
)
5569 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
5571 struct anon_map
*amp
;
5574 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
5576 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_READER
);
5577 if ((amp
= svd
->amp
) != NULL
) {
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.
5585 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
5586 if ((ap
= anon_get_ptr(amp
->ahp
,
5587 svd
->anon_index
+ seg_page(seg
, addr
))) != NULL
) {
5589 err
= anon_getpage(&ap
, NULL
, NULL
,
5590 0, seg
, addr
, S_READ
, svd
->cred
);
5592 ANON_LOCK_EXIT(&
->a_rwlock
);
5593 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5595 return (FC_MAKE_ERR(err
));
5598 ANON_LOCK_EXIT(&
->a_rwlock
);
5601 if (svd
->vp
== NULL
) {
5602 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5603 return (0); /* zfod page - do nothing now */
5607 TRACE_3(TR_FAC_VM
, TR_SEGVN_GETPAGE
,
5608 "segvn_getpage:seg %p addr %p vp %p", seg
, addr
, vp
);
5609 err
= VOP_GETPAGE(vp
,
5610 (offset_t
)(svd
->offset
+ (uintptr_t)(addr
- seg
->s_base
)),
5611 PAGESIZE
, NULL
, NULL
, 0, seg
, addr
,
5612 S_OTHER
, svd
->cred
, NULL
);
5614 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5616 return (FC_MAKE_ERR(err
));
5621 segvn_setprot(struct seg
*seg
, caddr_t addr
, size_t len
, uint_t prot
)
5623 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
5624 struct vpage
*cvp
, *svp
, *evp
;
5628 anon_sync_obj_t cookie
;
5629 int unload_done
= 0;
5631 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
5633 if ((svd
->maxprot
& prot
) != prot
)
5634 return (EACCES
); /* violated maxprot */
5636 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_WRITER
);
5638 /* return if prot is the same */
5639 if (!svd
->pageprot
&& svd
->prot
== prot
) {
5640 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5645 * Since we change protections we first have to flush the cache.
5646 * This makes sure all the pagelock calls have to recheck
5649 if (svd
->softlockcnt
> 0) {
5650 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
5653 * If this is shared segment non 0 softlockcnt
5654 * means locked pages are still in use.
5656 if (svd
->type
== MAP_SHARED
) {
5657 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5662 * Since we do have the segvn writers lock nobody can fill
5663 * the cache with entries belonging to this seg during
5664 * the purge. The flush either succeeds or we still have
5668 if (svd
->softlockcnt
> 0) {
5669 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5674 if (HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
)) {
5675 ASSERT(svd
->amp
== NULL
);
5676 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
5677 hat_leave_region(seg
->s_as
->a_hat
, svd
->rcookie
,
5679 svd
->rcookie
= HAT_INVALID_REGION_COOKIE
;
5681 } else if (svd
->tr_state
== SEGVN_TR_INIT
) {
5682 svd
->tr_state
= SEGVN_TR_OFF
;
5683 } else if (svd
->tr_state
== SEGVN_TR_ON
) {
5684 ASSERT(svd
->amp
!= NULL
);
5685 segvn_textunrepl(seg
, 0);
5686 ASSERT(svd
->amp
== NULL
&& svd
->tr_state
== SEGVN_TR_OFF
);
5690 if ((prot
& PROT_WRITE
) && svd
->type
== MAP_SHARED
&&
5691 svd
->vp
!= NULL
&& (svd
->vp
->v_flag
& VVMEXEC
)) {
5692 ASSERT(vn_is_mapped(svd
->vp
, V_WRITE
));
5693 segvn_inval_trcache(svd
->vp
);
5695 if (seg
->s_szc
!= 0) {
5697 pgsz
= page_get_pagesize(seg
->s_szc
);
5698 pgcnt
= pgsz
>> PAGESHIFT
;
5699 ASSERT(IS_P2ALIGNED(pgcnt
, pgcnt
));
5700 if (!IS_P2ALIGNED(addr
, pgsz
) || !IS_P2ALIGNED(len
, pgsz
)) {
5701 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5702 ASSERT(seg
->s_base
!= addr
|| seg
->s_size
!= len
);
5704 * If we are holding the as lock as a reader then
5705 * we need to return IE_RETRY and let the as
5706 * layer drop and re-acquire the lock as a writer.
5708 if (AS_READ_HELD(seg
->s_as
, &seg
->s_as
->a_lock
))
5710 VM_STAT_ADD(segvnvmstats
.demoterange
[1]);
5711 if (svd
->type
== MAP_PRIVATE
|| svd
->vp
!= NULL
) {
5712 err
= segvn_demote_range(seg
, addr
, len
,
5715 uint_t szcvec
= map_pgszcvec(seg
->s_base
,
5716 pgsz
, (uintptr_t)seg
->s_base
,
5717 (svd
->flags
& MAP_TEXT
), MAPPGSZC_SHM
, 0);
5718 err
= segvn_demote_range(seg
, addr
, len
,
5731 * If it's a private mapping and we're making it writable then we
5732 * may have to reserve the additional swap space now. If we are
5733 * making writable only a part of the segment then we use its vpage
5734 * array to keep a record of the pages for which we have reserved
5735 * swap. In this case we set the pageswap field in the segment's
5736 * segvn structure to record this.
5738 * If it's a private mapping to a file (i.e., vp != NULL) and we're
5739 * removing write permission on the entire segment and we haven't
5740 * modified any pages, we can release the swap space.
5742 if (svd
->type
== MAP_PRIVATE
) {
5743 if (prot
& PROT_WRITE
) {
5744 if (!(svd
->flags
& MAP_NORESERVE
) &&
5745 !(svd
->swresv
&& svd
->pageswap
== 0)) {
5749 * Start by determining how much swap
5750 * space is required.
5752 if (addr
== seg
->s_base
&&
5753 len
== seg
->s_size
&&
5754 svd
->pageswap
== 0) {
5755 /* The whole segment */
5759 * Make sure that the vpage array
5760 * exists, and make a note of the
5761 * range of elements corresponding
5765 svp
= &svd
->vpage
[seg_page(seg
, addr
)];
5766 evp
= &svd
->vpage
[seg_page(seg
,
5769 if (svd
->pageswap
== 0) {
5771 * This is the first time we've
5772 * asked for a part of this
5773 * segment, so we need to
5774 * reserve everything we've
5780 * We have to count the number
5781 * of pages required.
5783 for (cvp
= svp
; cvp
< evp
;
5785 if (!VPP_ISSWAPRES(cvp
))
5792 /* Try to reserve the necessary swap. */
5793 if (anon_resv_zone(sz
,
5794 seg
->s_as
->a_proc
->p_zone
) == 0) {
5795 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5800 * Make a note of how much swap space
5803 if (svd
->pageswap
== 0 && sz
== seg
->s_size
) {
5806 ASSERT(svd
->vpage
!= NULL
);
5809 for (cvp
= svp
; cvp
< evp
; cvp
++) {
5810 if (!VPP_ISSWAPRES(cvp
))
5811 VPP_SETSWAPRES(cvp
);
5817 * Swap space is released only if this segment
5818 * does not map anonymous memory, since read faults
5819 * on such segments still need an anon slot to read
5822 if (svd
->swresv
!= 0 && svd
->vp
!= NULL
&&
5823 svd
->amp
== NULL
&& addr
== seg
->s_base
&&
5824 len
== seg
->s_size
&& svd
->pageprot
== 0) {
5825 ASSERT(svd
->pageswap
== 0);
5826 anon_unresv_zone(svd
->swresv
,
5827 seg
->s_as
->a_proc
->p_zone
);
5829 TRACE_3(TR_FAC_VM
, TR_ANON_PROC
,
5830 "anon proc:%p %lu %u", seg
, 0, 0);
5835 if (addr
== seg
->s_base
&& len
== seg
->s_size
&& svd
->vpage
== NULL
) {
5836 if (svd
->prot
== prot
) {
5837 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5838 return (0); /* all done */
5840 svd
->prot
= (uchar_t
)prot
;
5841 } else if (svd
->type
== MAP_PRIVATE
) {
5842 struct anon
*ap
= NULL
;
5844 u_offset_t offset
, off
;
5845 struct anon_map
*amp
;
5846 ulong_t anon_idx
= 0;
5849 * A vpage structure exists or else the change does not
5850 * involve the entire segment. Establish a vpage structure
5851 * if none is there. Then, for each page in the range,
5852 * adjust its individual permissions. Note that write-
5853 * enabling a MAP_PRIVATE page can affect the claims for
5854 * locked down memory. Overcommitting memory terminates
5859 if ((amp
= svd
->amp
) != NULL
) {
5860 anon_idx
= svd
->anon_index
+ seg_page(seg
, addr
);
5861 ASSERT(seg
->s_szc
== 0 ||
5862 IS_P2ALIGNED(anon_idx
, pgcnt
));
5863 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
5866 offset
= svd
->offset
+ (uintptr_t)(addr
- seg
->s_base
);
5867 evp
= &svd
->vpage
[seg_page(seg
, addr
+ len
)];
5870 * See Statement at the beginning of segvn_lockop regarding
5871 * the way cowcnts and lckcnts are handled.
5873 for (svp
= &svd
->vpage
[seg_page(seg
, addr
)]; svp
< evp
; svp
++) {
5875 if (seg
->s_szc
!= 0) {
5877 anon_array_enter(amp
, anon_idx
,
5880 if (IS_P2ALIGNED(anon_idx
, pgcnt
) &&
5881 !segvn_claim_pages(seg
, svp
, offset
,
5884 anon_array_exit(&cookie
);
5889 anon_array_exit(&cookie
);
5894 anon_array_enter(amp
, anon_idx
,
5896 ap
= anon_get_ptr(amp
->ahp
, anon_idx
++);
5899 if (VPP_ISPPLOCK(svp
) &&
5900 VPP_PROT(svp
) != prot
) {
5902 if (amp
== NULL
|| ap
== NULL
) {
5906 swap_xlate(ap
, &vp
, &off
);
5908 anon_array_exit(&cookie
);
5910 if ((pp
= page_lookup(vp
, off
,
5911 SE_SHARED
)) == NULL
) {
5912 panic("segvn_setprot: no page");
5915 ASSERT(seg
->s_szc
== 0);
5916 if ((VPP_PROT(svp
) ^ prot
) &
5918 if (prot
& PROT_WRITE
) {
5933 } else if (amp
!= NULL
)
5934 anon_array_exit(&cookie
);
5936 VPP_SETPROT(svp
, prot
);
5940 ANON_LOCK_EXIT(&
->a_rwlock
);
5943 * Did we terminate prematurely? If so, simply unload
5944 * the translations to the things we've updated so far.
5948 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5951 len
= (svp
- &svd
->vpage
[seg_page(seg
, addr
)]) *
5953 ASSERT(seg
->s_szc
== 0 || IS_P2ALIGNED(len
, pgsz
));
5955 hat_unload(seg
->s_as
->a_hat
, addr
,
5957 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5963 evp
= &svd
->vpage
[seg_page(seg
, addr
+ len
)];
5964 for (svp
= &svd
->vpage
[seg_page(seg
, addr
)]; svp
< evp
; svp
++) {
5965 VPP_SETPROT(svp
, prot
);
5970 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
5974 if (((prot
& PROT_WRITE
) != 0 &&
5975 (svd
->vp
!= NULL
|| svd
->type
== MAP_PRIVATE
)) ||
5976 (prot
& ~PROT_USER
) == PROT_NONE
) {
5978 * Either private or shared data with write access (in
5979 * which case we need to throw out all former translations
5980 * so that we get the right translations set up on fault
5981 * and we don't allow write access to any copy-on-write pages
5982 * that might be around or to prevent write access to pages
5983 * representing holes in a file), or we don't have permission
5984 * to access the memory at all (in which case we have to
5985 * unload any current translations that might exist).
5987 hat_unload(seg
->s_as
->a_hat
, addr
, len
, HAT_UNLOAD
);
5990 * A shared mapping or a private mapping in which write
5991 * protection is going to be denied - just change all the
5992 * protections over the range of addresses in question.
5993 * segvn does not support any other attributes other
5994 * than prot so we can use hat_chgattr.
5996 hat_chgattr(seg
->s_as
->a_hat
, addr
, len
, prot
);
5999 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
6005 * segvn_setpagesize is called via SEGOP_SETPAGESIZE from as_setpagesize,
6006 * to determine if the seg is capable of mapping the requested szc.
6009 segvn_setpagesize(struct seg
*seg
, caddr_t addr
, size_t len
, uint_t szc
)
6011 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
6012 struct segvn_data
*nsvd
;
6013 struct anon_map
*amp
= svd
->amp
;
6015 caddr_t eaddr
= addr
+ len
, a
;
6016 size_t pgsz
= page_get_pagesize(szc
);
6017 pgcnt_t pgcnt
= page_get_pagecnt(szc
);
6019 u_offset_t off
= svd
->offset
+ (uintptr_t)(addr
- seg
->s_base
);
6020 extern struct vnode kvp
;
6022 ASSERT(seg
->s_as
&& AS_WRITE_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
6023 ASSERT(addr
>= seg
->s_base
&& eaddr
<= seg
->s_base
+ seg
->s_size
);
6025 if (seg
->s_szc
== szc
|| segvn_lpg_disable
!= 0) {
6030 * addr should always be pgsz aligned but eaddr may be misaligned if
6031 * it's at the end of the segment.
6033 * XXX we should assert this condition since as_setpagesize() logic
6036 if (!IS_P2ALIGNED(addr
, pgsz
) ||
6037 (!IS_P2ALIGNED(eaddr
, pgsz
) &&
6038 eaddr
!= seg
->s_base
+ seg
->s_size
)) {
6040 segvn_setpgsz_align_err
++;
6044 if (amp
!= NULL
&& svd
->type
== MAP_SHARED
) {
6045 ulong_t an_idx
= svd
->anon_index
+ seg_page(seg
, addr
);
6046 if (!IS_P2ALIGNED(an_idx
, pgcnt
)) {
6048 segvn_setpgsz_anon_align_err
++;
6053 if ((svd
->flags
& MAP_NORESERVE
) || seg
->s_as
== &kas
||
6054 szc
> segvn_maxpgszc
) {
6058 /* paranoid check */
6059 if (svd
->vp
!= NULL
&&
6060 (IS_SWAPFSVP(svd
->vp
) || VN_ISKAS(svd
->vp
))) {
6064 if (seg
->s_szc
== 0 && svd
->vp
!= NULL
&&
6065 map_addr_vacalign_check(addr
, off
)) {
6070 * Check that protections are the same within new page
6073 if (svd
->pageprot
) {
6074 for (a
= addr
; a
< eaddr
; a
+= pgsz
) {
6075 if ((a
+ pgsz
) > eaddr
) {
6076 if (!sameprot(seg
, a
, eaddr
- a
)) {
6080 if (!sameprot(seg
, a
, pgsz
)) {
6088 * Since we are changing page size we first have to flush
6089 * the cache. This makes sure all the pagelock calls have
6090 * to recheck protections.
6092 if (svd
->softlockcnt
> 0) {
6093 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
6096 * If this is shared segment non 0 softlockcnt
6097 * means locked pages are still in use.
6099 if (svd
->type
== MAP_SHARED
) {
6104 * Since we do have the segvn writers lock nobody can fill
6105 * the cache with entries belonging to this seg during
6106 * the purge. The flush either succeeds or we still have
6110 if (svd
->softlockcnt
> 0) {
6115 if (HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
)) {
6116 ASSERT(svd
->amp
== NULL
);
6117 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
6118 hat_leave_region(seg
->s_as
->a_hat
, svd
->rcookie
,
6120 svd
->rcookie
= HAT_INVALID_REGION_COOKIE
;
6121 } else if (svd
->tr_state
== SEGVN_TR_INIT
) {
6122 svd
->tr_state
= SEGVN_TR_OFF
;
6123 } else if (svd
->tr_state
== SEGVN_TR_ON
) {
6124 ASSERT(svd
->amp
!= NULL
);
6125 segvn_textunrepl(seg
, 1);
6126 ASSERT(svd
->amp
== NULL
&& svd
->tr_state
== SEGVN_TR_OFF
);
6131 * Operation for sub range of existing segment.
6133 if (addr
!= seg
->s_base
|| eaddr
!= (seg
->s_base
+ seg
->s_size
)) {
6134 if (szc
< seg
->s_szc
) {
6135 VM_STAT_ADD(segvnvmstats
.demoterange
[2]);
6136 err
= segvn_demote_range(seg
, addr
, len
, SDR_RANGE
, 0);
6140 if (err
== ENOMEM
) {
6145 if (addr
!= seg
->s_base
) {
6146 nseg
= segvn_split_seg(seg
, addr
);
6147 if (eaddr
!= (nseg
->s_base
+ nseg
->s_size
)) {
6148 /* eaddr is szc aligned */
6149 (void) segvn_split_seg(nseg
, eaddr
);
6153 if (eaddr
!= (seg
->s_base
+ seg
->s_size
)) {
6154 /* eaddr is szc aligned */
6155 (void) segvn_split_seg(seg
, eaddr
);
6161 * Break any low level sharing and reset seg->s_szc to 0.
6163 if ((err
= segvn_clrszc(seg
)) != 0) {
6164 if (err
== ENOMEM
) {
6169 ASSERT(seg
->s_szc
== 0);
6172 * If the end of the current segment is not pgsz aligned
6173 * then attempt to concatenate with the next segment.
6175 if (!IS_P2ALIGNED(eaddr
, pgsz
)) {
6176 nseg
= AS_SEGNEXT(seg
->s_as
, seg
);
6177 if (nseg
== NULL
|| nseg
== seg
|| eaddr
!= nseg
->s_base
) {
6180 if (nseg
->s_ops
!= &segvn_ops
) {
6183 nsvd
= (struct segvn_data
*)nseg
->s_data
;
6184 if (nsvd
->softlockcnt
> 0) {
6186 * If this is shared segment non 0 softlockcnt
6187 * means locked pages are still in use.
6189 if (nsvd
->type
== MAP_SHARED
) {
6193 if (nsvd
->softlockcnt
> 0) {
6197 err
= segvn_clrszc(nseg
);
6198 if (err
== ENOMEM
) {
6204 ASSERT(nsvd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
6205 err
= segvn_concat(seg
, nseg
, 1);
6216 * May need to re-align anon array to
6220 if (!IS_P2ALIGNED(svd
->anon_index
, pgcnt
)) {
6221 struct anon_hdr
*nahp
;
6223 ASSERT(svd
->type
== MAP_PRIVATE
);
6225 ANON_LOCK_ENTER(&
->a_rwlock
, RW_WRITER
);
6226 ASSERT(amp
->refcnt
== 1);
6227 nahp
= anon_create(btop(amp
->size
), ANON_NOSLEEP
);
6229 ANON_LOCK_EXIT(&
->a_rwlock
);
6232 if (anon_copy_ptr(amp
->ahp
, svd
->anon_index
,
6233 nahp
, 0, btop(seg
->s_size
), ANON_NOSLEEP
)) {
6234 anon_release(nahp
, btop(amp
->size
));
6235 ANON_LOCK_EXIT(&
->a_rwlock
);
6238 anon_release(amp
->ahp
, btop(amp
->size
));
6240 svd
->anon_index
= 0;
6241 ANON_LOCK_EXIT(&
->a_rwlock
);
6244 if (svd
->vp
!= NULL
&& szc
!= 0) {
6246 u_offset_t eoffpage
= svd
->offset
;
6247 va
.va_mask
= AT_SIZE
;
6248 eoffpage
+= seg
->s_size
;
6249 eoffpage
= btopr(eoffpage
);
6250 if (VOP_GETATTR(svd
->vp
, &va
, 0, svd
->cred
, NULL
) != 0) {
6251 segvn_setpgsz_getattr_err
++;
6254 if (btopr(va
.va_size
) < eoffpage
) {
6255 segvn_setpgsz_eof_err
++;
6260 * anon_fill_cow_holes() may call VOP_GETPAGE().
6261 * don't take anon map lock here to avoid holding it
6262 * across VOP_GETPAGE() calls that may call back into
6263 * segvn for klsutering checks. We don't really need
6264 * anon map lock here since it's a private segment and
6265 * we hold as level lock as writers.
6267 if ((err
= anon_fill_cow_holes(seg
, seg
->s_base
,
6268 amp
->ahp
, svd
->anon_index
, svd
->vp
, svd
->offset
,
6269 seg
->s_size
, szc
, svd
->prot
, svd
->vpage
,
6274 segvn_setvnode_mpss(svd
->vp
);
6278 ANON_LOCK_ENTER(&
->a_rwlock
, RW_WRITER
);
6279 if (svd
->type
== MAP_PRIVATE
) {
6281 } else if (szc
> amp
->a_szc
) {
6284 ANON_LOCK_EXIT(&
->a_rwlock
);
6293 segvn_clrszc(struct seg
*seg
)
6295 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
6296 struct anon_map
*amp
= svd
->amp
;
6300 caddr_t a
= seg
->s_base
;
6301 caddr_t ea
= a
+ seg
->s_size
;
6302 ulong_t an_idx
= svd
->anon_index
;
6303 vnode_t
*vp
= svd
->vp
;
6304 struct vpage
*vpage
= svd
->vpage
;
6305 page_t
*anon_pl
[1 + 1], *pp
;
6306 struct anon
*ap
, *oldap
;
6307 uint_t prot
= svd
->prot
, vpprot
;
6310 ASSERT(AS_WRITE_HELD(seg
->s_as
, &seg
->s_as
->a_lock
) ||
6311 SEGVN_WRITE_HELD(seg
->s_as
, &svd
->lock
));
6312 ASSERT(svd
->softlockcnt
== 0);
6314 if (vp
== NULL
&& amp
== NULL
) {
6315 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
6320 if (HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
)) {
6321 ASSERT(svd
->amp
== NULL
);
6322 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
6323 hat_leave_region(seg
->s_as
->a_hat
, svd
->rcookie
,
6325 svd
->rcookie
= HAT_INVALID_REGION_COOKIE
;
6326 } else if (svd
->tr_state
== SEGVN_TR_ON
) {
6327 ASSERT(svd
->amp
!= NULL
);
6328 segvn_textunrepl(seg
, 1);
6329 ASSERT(svd
->amp
== NULL
&& svd
->tr_state
== SEGVN_TR_OFF
);
6332 if (svd
->tr_state
!= SEGVN_TR_OFF
) {
6333 ASSERT(svd
->tr_state
== SEGVN_TR_INIT
);
6334 svd
->tr_state
= SEGVN_TR_OFF
;
6338 * do HAT_UNLOAD_UNMAP since we are changing the pagesize.
6339 * unload argument is 0 when we are freeing the segment
6340 * and unload was already done.
6342 hat_unload(seg
->s_as
->a_hat
, seg
->s_base
, seg
->s_size
,
6346 if (amp
== NULL
|| svd
->type
== MAP_SHARED
) {
6351 pgsz
= page_get_pagesize(seg
->s_szc
);
6355 * XXX anon rwlock is not really needed because this is a
6356 * private segment and we are writers.
6358 ANON_LOCK_ENTER(&
->a_rwlock
, RW_WRITER
);
6360 for (; a
< ea
; a
+= pgsz
, an_idx
+= pages
) {
6361 if ((oldap
= anon_get_ptr(amp
->ahp
, an_idx
)) != NULL
) {
6362 ASSERT(vpage
!= NULL
|| svd
->pageprot
== 0);
6363 if (vpage
!= NULL
) {
6364 ASSERT(sameprot(seg
, a
, pgsz
));
6365 prot
= VPP_PROT(vpage
);
6366 pageflag
= VPP_ISPPLOCK(vpage
) ? LOCK_PAGE
: 0;
6368 if (seg
->s_szc
!= 0) {
6369 ASSERT(vp
== NULL
|| anon_pages(amp
->ahp
,
6370 an_idx
, pages
) == pages
);
6371 if ((err
= anon_map_demotepages(amp
, an_idx
,
6372 seg
, a
, prot
, vpage
, svd
->cred
)) != 0) {
6376 if (oldap
->an_refcnt
== 1) {
6379 if ((err
= anon_getpage(&oldap
, &vpprot
,
6380 anon_pl
, PAGESIZE
, seg
, a
, S_READ
,
6384 if ((pp
= anon_private(&ap
, seg
, a
, prot
,
6385 anon_pl
[0], pageflag
, svd
->cred
)) == NULL
) {
6390 (void) anon_set_ptr(amp
->ahp
, an_idx
, ap
,
6395 vpage
= (vpage
== NULL
) ? NULL
: vpage
+ pages
;
6401 ANON_LOCK_EXIT(&
->a_rwlock
);
6413 pgcnt_t pgcnt
= page_get_pagecnt(seg
->s_szc
);
6414 size_t ppasize
= (pgcnt
+ 1) * sizeof (page_t
*);
6416 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
6417 struct anon_map
*amp
= svd
->amp
;
6418 struct vpage
*evp
= svp
+ pgcnt
;
6419 caddr_t addr
= ((uintptr_t)(svp
- svd
->vpage
) << PAGESHIFT
)
6422 struct vnode
*vp
= svd
->vp
;
6427 int anon
= (amp
!= NULL
) ? 1 : 0;
6429 ASSERT(svd
->type
== MAP_PRIVATE
);
6430 ASSERT(svd
->vpage
!= NULL
);
6431 ASSERT(seg
->s_szc
!= 0);
6432 ASSERT(IS_P2ALIGNED(pgcnt
, pgcnt
));
6433 ASSERT(amp
== NULL
|| IS_P2ALIGNED(anon_idx
, pgcnt
));
6434 ASSERT(sameprot(seg
, addr
, pgcnt
<< PAGESHIFT
));
6436 if (VPP_PROT(svp
) == prot
)
6438 if (!((VPP_PROT(svp
) ^ prot
) & PROT_WRITE
))
6441 ppa
= kmem_alloc(ppasize
, KM_SLEEP
);
6442 if (anon
&& vp
!= NULL
) {
6443 if (anon_get_ptr(amp
->ahp
, anon_idx
) == NULL
) {
6445 ASSERT(!anon_pages(amp
->ahp
, anon_idx
, pgcnt
));
6448 anon_pages(amp
->ahp
, anon_idx
, pgcnt
) == pgcnt
);
6451 for (*ppa
= NULL
, pg_idx
= 0; svp
< evp
; svp
++, anon_idx
++) {
6452 if (!VPP_ISPPLOCK(svp
))
6455 ap
= anon_get_ptr(amp
->ahp
, anon_idx
);
6457 panic("segvn_claim_pages: no anon slot");
6459 swap_xlate(ap
, &vp
, &aoff
);
6460 off
= (u_offset_t
)aoff
;
6463 if ((pp
= page_lookup(vp
,
6464 (u_offset_t
)off
, SE_SHARED
)) == NULL
) {
6465 panic("segvn_claim_pages: no page");
6471 if (ppa
[0] == NULL
) {
6472 kmem_free(ppa
, ppasize
);
6476 ASSERT(pg_idx
<= pgcnt
);
6479 if (prot
& PROT_WRITE
)
6480 err
= page_addclaim_pages(ppa
);
6482 err
= page_subclaim_pages(ppa
);
6484 for (i
= 0; i
< pg_idx
; i
++) {
6485 ASSERT(ppa
[i
] != NULL
);
6486 page_unlock(ppa
[i
]);
6489 kmem_free(ppa
, ppasize
);
6494 * Returns right (upper address) segment if split occurred.
6495 * If the address is equal to the beginning or end of its segment it returns
6496 * the current segment.
6499 segvn_split_seg(struct seg
*seg
, caddr_t addr
)
6501 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
6504 struct segvn_data
*nsvd
;
6506 ASSERT(AS_WRITE_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
6507 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
6509 ASSERT(addr
>= seg
->s_base
);
6510 ASSERT(addr
<= seg
->s_base
+ seg
->s_size
);
6511 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
6513 if (addr
== seg
->s_base
|| addr
== seg
->s_base
+ seg
->s_size
)
6516 nsize
= seg
->s_base
+ seg
->s_size
- addr
;
6517 seg
->s_size
= addr
- seg
->s_base
;
6518 nseg
= seg_alloc(seg
->s_as
, addr
, nsize
);
6519 ASSERT(nseg
!= NULL
);
6520 nseg
->s_ops
= seg
->s_ops
;
6521 nsvd
= kmem_cache_alloc(segvn_cache
, KM_SLEEP
);
6522 nseg
->s_data
= (void *)nsvd
;
6523 nseg
->s_szc
= seg
->s_szc
;
6525 ASSERT(nsvd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
6527 rw_init(&nsvd
->lock
, NULL
, RW_DEFAULT
, NULL
);
6529 if (nsvd
->vp
!= NULL
) {
6531 nsvd
->offset
= svd
->offset
+
6532 (uintptr_t)(nseg
->s_base
- seg
->s_base
);
6533 if (nsvd
->type
== MAP_SHARED
)
6534 lgrp_shm_policy_init(NULL
, nsvd
->vp
);
6537 * The offset for an anonymous segment has no signifigance in
6538 * terms of an offset into a file. If we were to use the above
6539 * calculation instead, the structures read out of
6540 * /proc/<pid>/xmap would be more difficult to decipher since
6541 * it would be unclear whether two seemingly contiguous
6542 * prxmap_t structures represented different segments or a
6543 * single segment that had been split up into multiple prxmap_t
6544 * structures (e.g. if some part of the segment had not yet
6550 ASSERT(svd
->softlockcnt
== 0);
6551 ASSERT(svd
->softlockcnt_sbase
== 0);
6552 ASSERT(svd
->softlockcnt_send
== 0);
6555 if (svd
->vpage
!= NULL
) {
6556 size_t bytes
= vpgtob(seg_pages(seg
));
6557 size_t nbytes
= vpgtob(seg_pages(nseg
));
6558 struct vpage
*ovpage
= svd
->vpage
;
6560 svd
->vpage
= kmem_alloc(bytes
, KM_SLEEP
);
6561 bcopy(ovpage
, svd
->vpage
, bytes
);
6562 nsvd
->vpage
= kmem_alloc(nbytes
, KM_SLEEP
);
6563 bcopy(ovpage
+ seg_pages(seg
), nsvd
->vpage
, nbytes
);
6564 kmem_free(ovpage
, bytes
+ nbytes
);
6566 if (svd
->amp
!= NULL
&& svd
->type
== MAP_PRIVATE
) {
6567 struct anon_map
*oamp
= svd
->amp
, *namp
;
6568 struct anon_hdr
*nahp
;
6570 ANON_LOCK_ENTER(&oamp
->a_rwlock
, RW_WRITER
);
6571 ASSERT(oamp
->refcnt
== 1);
6572 nahp
= anon_create(btop(seg
->s_size
), ANON_SLEEP
);
6573 (void) anon_copy_ptr(oamp
->ahp
, svd
->anon_index
,
6574 nahp
, 0, btop(seg
->s_size
), ANON_SLEEP
);
6576 namp
= anonmap_alloc(nseg
->s_size
, 0, ANON_SLEEP
);
6577 namp
->a_szc
= nseg
->s_szc
;
6578 (void) anon_copy_ptr(oamp
->ahp
,
6579 svd
->anon_index
+ btop(seg
->s_size
),
6580 namp
->ahp
, 0, btop(nseg
->s_size
), ANON_SLEEP
);
6581 anon_release(oamp
->ahp
, btop(oamp
->size
));
6583 oamp
->size
= seg
->s_size
;
6584 svd
->anon_index
= 0;
6586 nsvd
->anon_index
= 0;
6587 ANON_LOCK_EXIT(&oamp
->a_rwlock
);
6588 } else if (svd
->amp
!= NULL
) {
6589 pgcnt_t pgcnt
= page_get_pagecnt(seg
->s_szc
);
6590 ASSERT(svd
->amp
== nsvd
->amp
);
6591 ASSERT(seg
->s_szc
<= svd
->amp
->a_szc
);
6592 nsvd
->anon_index
= svd
->anon_index
+ seg_pages(seg
);
6593 ASSERT(IS_P2ALIGNED(nsvd
->anon_index
, pgcnt
));
6594 ANON_LOCK_ENTER(&svd
->amp
->a_rwlock
, RW_WRITER
);
6596 ANON_LOCK_EXIT(&svd
->amp
->a_rwlock
);
6600 * Split the amount of swap reserved.
6604 * For MAP_NORESERVE, only allocate swap reserve for pages
6605 * being used. Other segments get enough to cover whole
6608 if (svd
->flags
& MAP_NORESERVE
) {
6612 oswresv
= svd
->swresv
;
6613 svd
->swresv
= ptob(anon_pages(svd
->amp
->ahp
,
6614 svd
->anon_index
, btop(seg
->s_size
)));
6615 nsvd
->swresv
= ptob(anon_pages(nsvd
->amp
->ahp
,
6616 nsvd
->anon_index
, btop(nseg
->s_size
)));
6617 ASSERT(oswresv
>= (svd
->swresv
+ nsvd
->swresv
));
6619 if (svd
->pageswap
) {
6620 svd
->swresv
= segvn_count_swap_by_vpages(seg
);
6621 ASSERT(nsvd
->swresv
>= svd
->swresv
);
6622 nsvd
->swresv
-= svd
->swresv
;
6624 ASSERT(svd
->swresv
== seg
->s_size
+
6626 svd
->swresv
= seg
->s_size
;
6627 nsvd
->swresv
= nseg
->s_size
;
6636 * called on memory operations (unmap, setprot, setpagesize) for a subset
6637 * of a large page segment to either demote the memory range (SDR_RANGE)
6638 * or the ends (SDR_END) by addr/len.
6640 * returns 0 on success. returns errno, including ENOMEM, on failure.
6650 caddr_t eaddr
= addr
+ len
;
6651 caddr_t lpgaddr
, lpgeaddr
;
6653 struct seg
*badseg1
= NULL
;
6654 struct seg
*badseg2
= NULL
;
6656 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
6658 uint_t szc
= seg
->s_szc
;
6661 ASSERT(AS_WRITE_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
6662 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
6664 pgsz
= page_get_pagesize(szc
);
6665 ASSERT(seg
->s_base
!= addr
|| seg
->s_size
!= len
);
6666 ASSERT(addr
>= seg
->s_base
&& eaddr
<= seg
->s_base
+ seg
->s_size
);
6667 ASSERT(svd
->softlockcnt
== 0);
6668 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
6669 ASSERT(szcvec
== 0 || (flag
== SDR_END
&& svd
->type
== MAP_SHARED
));
6671 CALC_LPG_REGION(pgsz
, seg
, addr
, len
, lpgaddr
, lpgeaddr
);
6672 ASSERT(flag
== SDR_RANGE
|| eaddr
< lpgeaddr
|| addr
> lpgaddr
);
6673 if (flag
== SDR_RANGE
) {
6674 /* demote entire range */
6675 badseg1
= nseg
= segvn_split_seg(seg
, lpgaddr
);
6676 (void) segvn_split_seg(nseg
, lpgeaddr
);
6677 ASSERT(badseg1
->s_base
== lpgaddr
);
6678 ASSERT(badseg1
->s_size
== lpgeaddr
- lpgaddr
);
6679 } else if (addr
!= lpgaddr
) {
6680 ASSERT(flag
== SDR_END
);
6681 badseg1
= nseg
= segvn_split_seg(seg
, lpgaddr
);
6682 if (eaddr
!= lpgeaddr
&& eaddr
> lpgaddr
+ pgsz
&&
6683 eaddr
< lpgaddr
+ 2 * pgsz
) {
6684 (void) segvn_split_seg(nseg
, lpgeaddr
);
6685 ASSERT(badseg1
->s_base
== lpgaddr
);
6686 ASSERT(badseg1
->s_size
== 2 * pgsz
);
6688 nseg
= segvn_split_seg(nseg
, lpgaddr
+ pgsz
);
6689 ASSERT(badseg1
->s_base
== lpgaddr
);
6690 ASSERT(badseg1
->s_size
== pgsz
);
6691 if (eaddr
!= lpgeaddr
&& eaddr
> lpgaddr
+ pgsz
) {
6692 ASSERT(lpgeaddr
- lpgaddr
> 2 * pgsz
);
6693 nseg
= segvn_split_seg(nseg
, lpgeaddr
- pgsz
);
6695 (void) segvn_split_seg(nseg
, lpgeaddr
);
6696 ASSERT(badseg2
->s_base
== lpgeaddr
- pgsz
);
6697 ASSERT(badseg2
->s_size
== pgsz
);
6701 ASSERT(flag
== SDR_END
);
6702 ASSERT(eaddr
< lpgeaddr
);
6703 badseg1
= nseg
= segvn_split_seg(seg
, lpgeaddr
- pgsz
);
6704 (void) segvn_split_seg(nseg
, lpgeaddr
);
6705 ASSERT(badseg1
->s_base
== lpgeaddr
- pgsz
);
6706 ASSERT(badseg1
->s_size
== pgsz
);
6709 ASSERT(badseg1
!= NULL
);
6710 ASSERT(badseg1
->s_szc
== szc
);
6711 ASSERT(flag
== SDR_RANGE
|| badseg1
->s_size
== pgsz
||
6712 badseg1
->s_size
== 2 * pgsz
);
6713 ASSERT(sameprot(badseg1
, badseg1
->s_base
, pgsz
));
6714 ASSERT(badseg1
->s_size
== pgsz
||
6715 sameprot(badseg1
, badseg1
->s_base
+ pgsz
, pgsz
));
6716 if (err
= segvn_clrszc(badseg1
)) {
6719 ASSERT(badseg1
->s_szc
== 0);
6721 if (szc
> 1 && (tszcvec
= P2PHASE(szcvec
, 1 << szc
)) > 1) {
6722 uint_t tszc
= highbit(tszcvec
) - 1;
6723 caddr_t ta
= MAX(addr
, badseg1
->s_base
);
6725 size_t tpgsz
= page_get_pagesize(tszc
);
6727 ASSERT(svd
->type
== MAP_SHARED
);
6728 ASSERT(flag
== SDR_END
);
6729 ASSERT(tszc
< szc
&& tszc
> 0);
6731 if (eaddr
> badseg1
->s_base
+ badseg1
->s_size
) {
6732 te
= badseg1
->s_base
+ badseg1
->s_size
;
6738 badseg1
->s_szc
= tszc
;
6739 if (!IS_P2ALIGNED(ta
, tpgsz
) || !IS_P2ALIGNED(te
, tpgsz
)) {
6740 if (badseg2
!= NULL
) {
6741 err
= segvn_demote_range(badseg1
, ta
, te
- ta
,
6747 return (segvn_demote_range(badseg1
, ta
,
6748 te
- ta
, SDR_END
, tszcvec
));
6753 if (badseg2
== NULL
)
6755 ASSERT(badseg2
->s_szc
== szc
);
6756 ASSERT(badseg2
->s_size
== pgsz
);
6757 ASSERT(sameprot(badseg2
, badseg2
->s_base
, badseg2
->s_size
));
6758 if (err
= segvn_clrszc(badseg2
)) {
6761 ASSERT(badseg2
->s_szc
== 0);
6763 if (szc
> 1 && (tszcvec
= P2PHASE(szcvec
, 1 << szc
)) > 1) {
6764 uint_t tszc
= highbit(tszcvec
) - 1;
6765 size_t tpgsz
= page_get_pagesize(tszc
);
6767 ASSERT(svd
->type
== MAP_SHARED
);
6768 ASSERT(flag
== SDR_END
);
6769 ASSERT(tszc
< szc
&& tszc
> 0);
6770 ASSERT(badseg2
->s_base
> addr
);
6771 ASSERT(eaddr
> badseg2
->s_base
);
6772 ASSERT(eaddr
< badseg2
->s_base
+ badseg2
->s_size
);
6774 badseg2
->s_szc
= tszc
;
6775 if (!IS_P2ALIGNED(eaddr
, tpgsz
)) {
6776 return (segvn_demote_range(badseg2
, badseg2
->s_base
,
6777 eaddr
- badseg2
->s_base
, SDR_END
, tszcvec
));
6785 segvn_checkprot(struct seg
*seg
, caddr_t addr
, size_t len
, uint_t prot
)
6787 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
6788 struct vpage
*vp
, *evp
;
6790 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
6792 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_READER
);
6794 * If segment protection can be used, simply check against them.
6796 if (svd
->pageprot
== 0) {
6799 err
= ((svd
->prot
& prot
) != prot
) ? EACCES
: 0;
6800 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
6805 * Have to check down to the vpage level.
6807 evp
= &svd
->vpage
[seg_page(seg
, addr
+ len
)];
6808 for (vp
= &svd
->vpage
[seg_page(seg
, addr
)]; vp
< evp
; vp
++) {
6809 if ((VPP_PROT(vp
) & prot
) != prot
) {
6810 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
6814 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
6819 segvn_getprot(struct seg
*seg
, caddr_t addr
, size_t len
, uint_t
*protv
)
6821 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
6822 size_t pgno
= seg_page(seg
, addr
+ len
) - seg_page(seg
, addr
) + 1;
6824 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
6827 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_READER
);
6828 if (svd
->pageprot
== 0) {
6830 protv
[--pgno
] = svd
->prot
;
6831 } while (pgno
!= 0);
6833 size_t pgoff
= seg_page(seg
, addr
);
6837 protv
[pgno
] = VPP_PROT(&svd
->vpage
[pgno
+pgoff
]);
6838 } while (pgno
!= 0);
6840 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
6846 segvn_getoffset(struct seg
*seg
, caddr_t addr
)
6848 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
6850 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
6852 return (svd
->offset
+ (uintptr_t)(addr
- seg
->s_base
));
6857 segvn_gettype(struct seg
*seg
, caddr_t addr
)
6859 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
6861 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
6863 return (svd
->type
| (svd
->flags
& (MAP_NORESERVE
| MAP_TEXT
|
6869 segvn_getvp(struct seg
*seg
, caddr_t addr
, struct vnode
**vpp
)
6871 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
6873 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
6880 * Check to see if it makes sense to do kluster/read ahead to
6881 * addr + delta relative to the mapping at addr. We assume here
6882 * that delta is a signed PAGESIZE'd multiple (which can be negative).
6884 * For segvn, we currently "approve" of the action if we are
6885 * still in the segment and it maps from the same vp/off,
6886 * or if the advice stored in segvn_data or vpages allows it.
6887 * Currently, klustering is not allowed only if MADV_RANDOM is set.
6890 segvn_kluster(struct seg
*seg
, caddr_t addr
, ssize_t delta
)
6892 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
6893 struct anon
*oap
, *ap
;
6896 struct vnode
*vp1
, *vp2
;
6897 u_offset_t off1
, off2
;
6898 struct anon_map
*amp
;
6900 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
6901 ASSERT(AS_WRITE_HELD(seg
->s_as
, &seg
->s_as
->a_lock
) ||
6902 SEGVN_LOCK_HELD(seg
->s_as
, &svd
->lock
));
6904 if (addr
+ delta
< seg
->s_base
||
6905 addr
+ delta
>= (seg
->s_base
+ seg
->s_size
))
6906 return (-1); /* exceeded segment bounds */
6908 pd
= delta
/ (ssize_t
)PAGESIZE
; /* divide to preserve sign bit */
6909 page
= seg_page(seg
, addr
);
6912 * Check to see if either of the pages addr or addr + delta
6913 * have advice set that prevents klustering (if MADV_RANDOM advice
6914 * is set for entire segment, or MADV_SEQUENTIAL is set and delta
6917 if (svd
->advice
== MADV_RANDOM
||
6918 svd
->advice
== MADV_SEQUENTIAL
&& delta
< 0)
6920 else if (svd
->pageadvice
&& svd
->vpage
) {
6921 struct vpage
*bvpp
, *evpp
;
6923 bvpp
= &svd
->vpage
[page
];
6924 evpp
= &svd
->vpage
[page
+ pd
];
6925 if (VPP_ADVICE(bvpp
) == MADV_RANDOM
||
6926 VPP_ADVICE(evpp
) == MADV_SEQUENTIAL
&& delta
< 0)
6928 if (VPP_ADVICE(bvpp
) != VPP_ADVICE(evpp
) &&
6929 VPP_ADVICE(evpp
) == MADV_RANDOM
)
6933 if (svd
->type
== MAP_SHARED
)
6934 return (0); /* shared mapping - all ok */
6936 if ((amp
= svd
->amp
) == NULL
)
6937 return (0); /* off original vnode */
6939 page
+= svd
->anon_index
;
6941 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
6943 oap
= anon_get_ptr(amp
->ahp
, page
);
6944 ap
= anon_get_ptr(amp
->ahp
, page
+ pd
);
6946 ANON_LOCK_EXIT(&
->a_rwlock
);
6948 if ((oap
== NULL
&& ap
!= NULL
) || (oap
!= NULL
&& ap
== NULL
)) {
6949 return (-1); /* one with and one without an anon */
6952 if (oap
== NULL
) { /* implies that ap == NULL */
6953 return (0); /* off original vnode */
6957 * Now we know we have two anon pointers - check to
6958 * see if they happen to be properly allocated.
6962 * XXX We cheat here and don't lock the anon slots. We can't because
6963 * we may have been called from the anon layer which might already
6964 * have locked them. We are holding a refcnt on the slots so they
6965 * can't disappear. The worst that will happen is we'll get the wrong
6966 * names (vp, off) for the slots and make a poor klustering decision.
6968 swap_xlate(ap
, &vp1
, &off1
);
6969 swap_xlate(oap
, &vp2
, &off2
);
6972 if (!VOP_CMP(vp1
, vp2
, NULL
) || off1
- off2
!= delta
)
6978 * Swap the pages of seg out to secondary storage, returning the
6979 * number of bytes of storage freed.
6981 * The basic idea is first to unload all translations and then to call
6982 * VOP_PUTPAGE() for all newly-unmapped pages, to push them out to the
6983 * swap device. Pages to which other segments have mappings will remain
6984 * mapped and won't be swapped. Our caller (as_swapout) has already
6985 * performed the unloading step.
6987 * The value returned is intended to correlate well with the process's
6988 * memory requirements. However, there are some caveats:
6989 * 1) When given a shared segment as argument, this routine will
6990 * only succeed in swapping out pages for the last sharer of the
6991 * segment. (Previous callers will only have decremented mapping
6992 * reference counts.)
6993 * 2) We assume that the hat layer maintains a large enough translation
6994 * cache to capture process reference patterns.
6997 segvn_swapout(struct seg
*seg
)
6999 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
7000 struct anon_map
*amp
;
7006 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
7008 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_READER
);
7010 * Find pages unmapped by our caller and force them
7011 * out to the virtual swap device.
7013 if ((amp
= svd
->amp
) != NULL
)
7014 anon_index
= svd
->anon_index
;
7015 npages
= seg
->s_size
>> PAGESHIFT
;
7016 for (page
= 0; page
< npages
; page
++) {
7021 anon_sync_obj_t cookie
;
7024 * Obtain <vp, off> pair for the page, then look it up.
7026 * Note that this code is willing to consider regular
7027 * pages as well as anon pages. Is this appropriate here?
7031 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
7032 if (anon_array_try_enter(amp
, anon_index
+ page
,
7034 ANON_LOCK_EXIT(&
->a_rwlock
);
7037 ap
= anon_get_ptr(amp
->ahp
, anon_index
+ page
);
7039 swap_xlate(ap
, &vp
, &off
);
7042 off
= svd
->offset
+ ptob(page
);
7044 anon_array_exit(&cookie
);
7045 ANON_LOCK_EXIT(&
->a_rwlock
);
7048 off
= svd
->offset
+ ptob(page
);
7050 if (vp
== NULL
) { /* untouched zfod page */
7055 pp
= page_lookup_nowait(vp
, off
, SE_SHARED
);
7061 * Examine the page to see whether it can be tossed out,
7062 * keeping track of how many we've found.
7064 if (!page_tryupgrade(pp
)) {
7066 * If the page has an i/o lock and no mappings,
7067 * it's very likely that the page is being
7068 * written out as a result of klustering.
7069 * Assume this is so and take credit for it here.
7071 if (!page_io_trylock(pp
)) {
7072 if (!hat_page_is_mapped(pp
))
7080 ASSERT(!page_iolock_assert(pp
));
7084 * Skip if page is locked or has mappings.
7085 * We don't need the page_struct_lock to look at lckcnt
7086 * and cowcnt because the page is exclusive locked.
7088 if (pp
->p_lckcnt
!= 0 || pp
->p_cowcnt
!= 0 ||
7089 hat_page_is_mapped(pp
)) {
7095 * dispose skips large pages so try to demote first.
7097 if (pp
->p_szc
!= 0 && !page_try_demote_pages(pp
)) {
7100 * XXX should skip the remaining page_t's of this
7106 ASSERT(pp
->p_szc
== 0);
7109 * No longer mapped -- we can toss it out. How
7110 * we do so depends on whether or not it's dirty.
7112 if (hat_ismod(pp
) && pp
->p_vnode
) {
7114 * We must clean the page before it can be
7115 * freed. Setting B_FREE will cause pvn_done
7116 * to free the page when the i/o completes.
7117 * XXX: This also causes it to be accounted
7118 * as a pageout instead of a swap: need
7119 * B_SWAPOUT bit to use instead of B_FREE.
7121 * Hold the vnode before releasing the page lock
7122 * to prevent it from being freed and re-used by
7123 * some other thread.
7129 * Queue all i/o requests for the pageout thread
7130 * to avoid saturating the pageout devices.
7132 if (!queue_io_request(vp
, off
))
7136 * The page was clean, free it.
7138 * XXX: Can we ever encounter modified pages
7139 * with no associated vnode here?
7141 ASSERT(pp
->p_vnode
!= NULL
);
7142 /*LINTED: constant in conditional context*/
7143 VN_DISPOSE(pp
, B_FREE
, 0, kcred
);
7147 * Credit now even if i/o is in progress.
7151 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7154 * Wakeup pageout to initiate i/o on all queued requests.
7156 cv_signal_pageout();
7157 return (ptob(pgcnt
));
7161 * Synchronize primary storage cache with real object in virtual memory.
7163 * XXX - Anonymous pages should not be sync'ed out at all.
7166 segvn_sync(struct seg
*seg
, caddr_t addr
, size_t len
, int attr
, uint_t flags
)
7168 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
7181 struct anon_map
*amp
;
7183 anon_sync_obj_t cookie
;
7185 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
7187 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_READER
);
7189 if (svd
->softlockcnt
> 0) {
7191 * If this is shared segment non 0 softlockcnt
7192 * means locked pages are still in use.
7194 if (svd
->type
== MAP_SHARED
) {
7195 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7200 * flush all pages from seg cache
7201 * otherwise we may deadlock in swap_putpage
7202 * for B_INVAL page (4175402).
7204 * Even if we grab segvn WRITER's lock
7205 * here, there might be another thread which could've
7206 * successfully performed lookup/insert just before
7207 * we acquired the lock here. So, grabbing either
7208 * lock here is of not much use. Until we devise
7209 * a strategy at upper layers to solve the
7210 * synchronization issues completely, we expect
7211 * applications to handle this appropriately.
7214 if (svd
->softlockcnt
> 0) {
7215 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7218 } else if (svd
->type
== MAP_SHARED
&& svd
->amp
!= NULL
&&
7219 svd
->amp
->a_softlockcnt
> 0) {
7221 * Try to purge this amp's entries from pcache. It will
7222 * succeed only if other segments that share the amp have no
7223 * outstanding softlock's.
7226 if (svd
->amp
->a_softlockcnt
> 0 || svd
->softlockcnt
> 0) {
7227 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7233 offset
= svd
->offset
+ (uintptr_t)(addr
- seg
->s_base
);
7234 bflags
= ((flags
& MS_ASYNC
) ? B_ASYNC
: 0) |
7235 ((flags
& MS_INVALIDATE
) ? B_INVAL
: 0);
7238 pageprot
= attr
& ~(SHARED
|PRIVATE
);
7239 segtype
= (attr
& SHARED
) ? MAP_SHARED
: MAP_PRIVATE
;
7242 * We are done if the segment types don't match
7243 * or if we have segment level protections and
7246 if (svd
->type
!= segtype
) {
7247 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7251 if (svd
->prot
!= pageprot
) {
7252 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7257 vpp
= &svd
->vpage
[seg_page(seg
, addr
)];
7259 } else if (svd
->vp
&& svd
->amp
== NULL
&&
7260 (flags
& MS_INVALIDATE
) == 0) {
7263 * No attributes, no anonymous pages and MS_INVALIDATE flag
7264 * is not on, just use one big request.
7266 err
= VOP_PUTPAGE(svd
->vp
, (offset_t
)offset
, len
,
7267 bflags
, svd
->cred
, NULL
);
7268 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7272 if ((amp
= svd
->amp
) != NULL
)
7273 anon_index
= svd
->anon_index
+ seg_page(seg
, addr
);
7275 for (eaddr
= addr
+ len
; addr
< eaddr
; addr
+= PAGESIZE
) {
7278 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
7279 anon_array_enter(amp
, anon_index
, &cookie
);
7280 ap
= anon_get_ptr(amp
->ahp
, anon_index
++);
7282 swap_xlate(ap
, &vp
, &off
);
7287 anon_array_exit(&cookie
);
7288 ANON_LOCK_EXIT(&
->a_rwlock
);
7295 if (vp
== NULL
) /* untouched zfod page */
7300 prot
= VPP_PROT(vpp
);
7303 if (prot
!= pageprot
) {
7309 * See if any of these pages are locked -- if so, then we
7310 * will have to truncate an invalidate request at the first
7311 * locked one. We don't need the page_struct_lock to test
7312 * as this is only advisory; even if we acquire it someone
7313 * might race in and lock the page after we unlock and before
7314 * we do the PUTPAGE, then PUTPAGE simply does nothing.
7316 if (flags
& MS_INVALIDATE
) {
7317 if ((pp
= page_lookup(vp
, off
, SE_SHARED
)) != NULL
) {
7318 if (pp
->p_lckcnt
!= 0 || pp
->p_cowcnt
!= 0) {
7320 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7323 if (ap
!= NULL
&& pp
->p_szc
!= 0 &&
7324 page_tryupgrade(pp
)) {
7325 if (pp
->p_lckcnt
== 0 &&
7326 pp
->p_cowcnt
== 0) {
7328 * swapfs VN_DISPOSE() won't
7329 * invalidate large pages.
7330 * Attempt to demote.
7331 * XXX can't help it if it
7332 * fails. But for swapfs
7333 * pages it is no big deal.
7335 (void) page_try_demote_pages(
7341 } else if (svd
->type
== MAP_SHARED
&& amp
!= NULL
) {
7343 * Avoid writing out to disk ISM's large pages
7344 * because segspt_free_pages() relies on NULL an_pvp
7345 * of anon slots of such pages.
7348 ASSERT(svd
->vp
== NULL
);
7350 * swapfs uses page_lookup_nowait if not freeing or
7351 * invalidating and skips a page if
7352 * page_lookup_nowait returns NULL.
7354 pp
= page_lookup_nowait(vp
, off
, SE_SHARED
);
7358 if (pp
->p_szc
!= 0) {
7364 * Note ISM pages are created large so (vp, off)'s
7365 * page cannot suddenly become large after we unlock
7371 * XXX - Should ultimately try to kluster
7372 * calls to VOP_PUTPAGE() for performance.
7375 err
= VOP_PUTPAGE(vp
, (offset_t
)off
, PAGESIZE
,
7376 (bflags
| (IS_SWAPFSVP(vp
) ? B_PAGE_NOWAIT
: 0)),
7383 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7388 * Determine if we have data corresponding to pages in the
7389 * primary storage virtual memory cache (i.e., "in core").
7392 segvn_incore(struct seg
*seg
, caddr_t addr
, size_t len
, char *vec
)
7394 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
7395 struct vnode
*vp
, *avp
;
7396 u_offset_t offset
, aoffset
;
7402 struct anon_map
*amp
; /* XXX - for locknest */
7405 anon_sync_obj_t cookie
;
7407 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
7409 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_READER
);
7410 if (svd
->amp
== NULL
&& svd
->vp
== NULL
) {
7411 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7412 bzero(vec
, btopr(len
));
7413 return (len
); /* no anonymous pages created yet */
7416 p
= seg_page(seg
, addr
);
7417 ep
= seg_page(seg
, addr
+ len
);
7418 start
= svd
->vp
? SEG_PAGE_VNODEBACKED
: 0;
7421 for (; p
< ep
; p
++, addr
+= PAGESIZE
) {
7422 vpp
= (svd
->vpage
) ? &svd
->vpage
[p
]: NULL
;
7426 /* Grab the vnode/offset for the anon slot */
7428 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
7429 anon_array_enter(amp
, svd
->anon_index
+ p
, &cookie
);
7430 ap
= anon_get_ptr(amp
->ahp
, svd
->anon_index
+ p
);
7432 swap_xlate(ap
, &avp
, &aoffset
);
7434 anon_array_exit(&cookie
);
7435 ANON_LOCK_EXIT(&
->a_rwlock
);
7437 if ((avp
!= NULL
) && page_exists(avp
, aoffset
)) {
7438 /* A page exists for the anon slot */
7439 ret
|= SEG_PAGE_INCORE
;
7442 * If page is mapped and writable
7445 if ((hat_getattr(seg
->s_as
->a_hat
, addr
,
7446 &attr
) != -1) && (attr
& PROT_WRITE
)) {
7447 ret
|= SEG_PAGE_ANON
;
7450 * Don't get page_struct lock for lckcnt and cowcnt,
7451 * since this is purely advisory.
7453 if ((pp
= page_lookup_nowait(avp
, aoffset
,
7454 SE_SHARED
)) != NULL
) {
7456 ret
|= SEG_PAGE_SOFTLOCK
;
7458 ret
|= SEG_PAGE_HASCOW
;
7463 /* Gather vnode statistics */
7465 offset
= svd
->offset
+ (uintptr_t)(addr
- seg
->s_base
);
7469 * Try to obtain a "shared" lock on the page
7470 * without blocking. If this fails, determine
7471 * if the page is in memory.
7473 pp
= page_lookup_nowait(vp
, offset
, SE_SHARED
);
7474 if ((pp
== NULL
) && (page_exists(vp
, offset
))) {
7475 /* Page is incore, and is named */
7476 ret
|= (SEG_PAGE_INCORE
| SEG_PAGE_VNODE
);
7479 * Don't get page_struct lock for lckcnt and cowcnt,
7480 * since this is purely advisory.
7483 ret
|= (SEG_PAGE_INCORE
| SEG_PAGE_VNODE
);
7485 ret
|= SEG_PAGE_SOFTLOCK
;
7487 ret
|= SEG_PAGE_HASCOW
;
7492 /* Gather virtual page information */
7494 if (VPP_ISPPLOCK(vpp
))
7495 ret
|= SEG_PAGE_LOCKED
;
7501 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7506 * Statement for p_cowcnts/p_lckcnts.
7508 * p_cowcnt is updated while mlock/munlocking MAP_PRIVATE and PROT_WRITE region
7509 * irrespective of the following factors or anything else:
7511 * (1) anon slots are populated or not
7512 * (2) cow is broken or not
7513 * (3) refcnt on ap is 1 or greater than 1
7515 * If it's not MAP_PRIVATE and PROT_WRITE, p_lckcnt is updated during mlock
7519 * Handling p_cowcnts/p_lckcnts during copy-on-write fault:
7521 * if vpage has PROT_WRITE
7522 * transfer cowcnt on the oldpage -> cowcnt on the newpage
7524 * transfer lckcnt on the oldpage -> lckcnt on the newpage
7526 * During copy-on-write, decrement p_cowcnt on the oldpage and increment
7527 * p_cowcnt on the newpage *if* the corresponding vpage has PROT_WRITE.
7529 * We may also break COW if softlocking on read access in the physio case.
7530 * In this case, vpage may not have PROT_WRITE. So, we need to decrement
7531 * p_lckcnt on the oldpage and increment p_lckcnt on the newpage *if* the
7532 * vpage doesn't have PROT_WRITE.
7535 * Handling p_cowcnts/p_lckcnts during mprotect on mlocked region:
7537 * If a MAP_PRIVATE region loses PROT_WRITE, we decrement p_cowcnt and
7538 * increment p_lckcnt by calling page_subclaim() which takes care of
7539 * availrmem accounting and p_lckcnt overflow.
7541 * If a MAP_PRIVATE region gains PROT_WRITE, we decrement p_lckcnt and
7542 * increment p_cowcnt by calling page_addclaim() which takes care of
7543 * availrmem availability and p_cowcnt overflow.
7547 * Lock down (or unlock) pages mapped by this segment.
7549 * XXX only creates PAGESIZE pages if anon slots are not initialized.
7550 * At fault time they will be relocated into larger pages.
7553 segvn_lockop(struct seg
*seg
, caddr_t addr
, size_t len
,
7554 int attr
, int op
, ulong_t
*lockmap
, size_t pos
)
7556 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
7567 struct anon_map
*amp
;
7570 anon_sync_obj_t cookie
;
7571 struct kshmid
*sp
= NULL
;
7572 struct proc
*p
= curproc
;
7573 kproject_t
*proj
= NULL
;
7575 size_t locked_bytes
= 0;
7576 size_t unlocked_bytes
= 0;
7580 * Hold write lock on address space because may split or concatenate
7583 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
7586 * If this is a shm, use shm's project and zone, else use
7587 * project and zone of calling process
7590 /* Determine if this segment backs a sysV shm */
7591 if (svd
->amp
!= NULL
&& svd
->amp
->a_sp
!= NULL
) {
7592 ASSERT(svd
->type
== MAP_SHARED
);
7593 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
7594 sp
= svd
->amp
->a_sp
;
7595 proj
= sp
->shm_perm
.ipc_proj
;
7599 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_WRITER
);
7601 pageprot
= attr
& ~(SHARED
|PRIVATE
);
7602 segtype
= attr
& SHARED
? MAP_SHARED
: MAP_PRIVATE
;
7605 * We are done if the segment types don't match
7606 * or if we have segment level protections and
7609 if (svd
->type
!= segtype
) {
7610 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7613 if (svd
->pageprot
== 0 && svd
->prot
!= pageprot
) {
7614 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7619 if (op
== MC_LOCK
) {
7620 if (svd
->tr_state
== SEGVN_TR_INIT
) {
7621 svd
->tr_state
= SEGVN_TR_OFF
;
7622 } else if (svd
->tr_state
== SEGVN_TR_ON
) {
7623 ASSERT(svd
->amp
!= NULL
);
7624 segvn_textunrepl(seg
, 0);
7625 ASSERT(svd
->amp
== NULL
&&
7626 svd
->tr_state
== SEGVN_TR_OFF
);
7631 * If we're locking, then we must create a vpage structure if
7632 * none exists. If we're unlocking, then check to see if there
7633 * is a vpage -- if not, then we could not have locked anything.
7636 if ((vpp
= svd
->vpage
) == NULL
) {
7640 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7646 * The anonymous data vector (i.e., previously
7647 * unreferenced mapping to swap space) can be allocated
7648 * by lazily testing for its existence.
7650 if (op
== MC_LOCK
&& svd
->amp
== NULL
&& svd
->vp
== NULL
) {
7651 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
7652 svd
->amp
= anonmap_alloc(seg
->s_size
, 0, ANON_SLEEP
);
7653 svd
->amp
->a_szc
= seg
->s_szc
;
7656 if ((amp
= svd
->amp
) != NULL
) {
7657 anon_index
= svd
->anon_index
+ seg_page(seg
, addr
);
7660 offset
= svd
->offset
+ (uintptr_t)(addr
- seg
->s_base
);
7661 evp
= &svd
->vpage
[seg_page(seg
, addr
+ len
)];
7664 mutex_enter(&sp
->shm_mlock
);
7666 /* determine number of unlocked bytes in range for lock operation */
7667 if (op
== MC_LOCK
) {
7670 for (vpp
= &svd
->vpage
[seg_page(seg
, addr
)]; vpp
< evp
;
7672 if (!VPP_ISPPLOCK(vpp
))
7673 unlocked_bytes
+= PAGESIZE
;
7676 ulong_t i_idx
, i_edx
;
7677 anon_sync_obj_t i_cookie
;
7682 /* Only count sysV pages once for locked memory */
7683 i_edx
= svd
->anon_index
+ seg_page(seg
, addr
+ len
);
7684 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
7685 for (i_idx
= anon_index
; i_idx
< i_edx
; i_idx
++) {
7686 anon_array_enter(amp
, i_idx
, &i_cookie
);
7687 i_ap
= anon_get_ptr(amp
->ahp
, i_idx
);
7689 unlocked_bytes
+= PAGESIZE
;
7690 anon_array_exit(&i_cookie
);
7693 swap_xlate(i_ap
, &i_vp
, &i_off
);
7694 anon_array_exit(&i_cookie
);
7695 pp
= page_lookup(i_vp
, i_off
, SE_SHARED
);
7697 unlocked_bytes
+= PAGESIZE
;
7699 } else if (pp
->p_lckcnt
== 0)
7700 unlocked_bytes
+= PAGESIZE
;
7703 ANON_LOCK_EXIT(&
->a_rwlock
);
7706 mutex_enter(&p
->p_lock
);
7707 err
= rctl_incr_locked_mem(p
, proj
, unlocked_bytes
,
7709 mutex_exit(&p
->p_lock
);
7713 mutex_exit(&sp
->shm_mlock
);
7714 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7719 * Loop over all pages in the range. Process if we're locking and
7720 * page has not already been locked in this mapping; or if we're
7721 * unlocking and the page has been locked.
7723 for (vpp
= &svd
->vpage
[seg_page(seg
, addr
)]; vpp
< evp
;
7724 vpp
++, pos
++, addr
+= PAGESIZE
, offset
+= PAGESIZE
, anon_index
++) {
7725 if ((attr
== 0 || VPP_PROT(vpp
) == pageprot
) &&
7726 ((op
== MC_LOCK
&& !VPP_ISPPLOCK(vpp
)) ||
7727 (op
== MC_UNLOCK
&& VPP_ISPPLOCK(vpp
)))) {
7730 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
7732 * If this isn't a MAP_NORESERVE segment and
7733 * we're locking, allocate anon slots if they
7734 * don't exist. The page is brought in later on.
7736 if (op
== MC_LOCK
&& svd
->vp
== NULL
&&
7737 ((svd
->flags
& MAP_NORESERVE
) == 0) &&
7739 ((ap
= anon_get_ptr(amp
->ahp
, anon_index
))
7741 anon_array_enter(amp
, anon_index
, &cookie
);
7743 if ((ap
= anon_get_ptr(amp
->ahp
,
7744 anon_index
)) == NULL
) {
7745 pp
= anon_zero(seg
, addr
, &ap
,
7748 anon_array_exit(&cookie
);
7749 ANON_LOCK_EXIT(&
->a_rwlock
);
7753 ASSERT(anon_get_ptr(amp
->ahp
,
7754 anon_index
) == NULL
);
7755 (void) anon_set_ptr(amp
->ahp
,
7756 anon_index
, ap
, ANON_SLEEP
);
7759 anon_array_exit(&cookie
);
7763 * Get name for page, accounting for
7764 * existence of private copy.
7768 anon_array_enter(amp
, anon_index
, &cookie
);
7769 ap
= anon_get_ptr(amp
->ahp
, anon_index
);
7771 swap_xlate(ap
, &vp
, &off
);
7773 if (svd
->vp
== NULL
&&
7774 (svd
->flags
& MAP_NORESERVE
)) {
7775 anon_array_exit(&cookie
);
7776 ANON_LOCK_EXIT(&
->a_rwlock
);
7782 if (op
!= MC_LOCK
|| ap
== NULL
) {
7783 anon_array_exit(&cookie
);
7784 ANON_LOCK_EXIT(&
->a_rwlock
);
7792 * Get page frame. It's ok if the page is
7793 * not available when we're unlocking, as this
7794 * may simply mean that a page we locked got
7795 * truncated out of existence after we locked it.
7797 * Invoke VOP_GETPAGE() to obtain the page struct
7798 * since we may need to read it from disk if its
7802 pp
= page_lookup(vp
, off
, SE_SHARED
);
7809 error
= VOP_GETPAGE(vp
, (offset_t
)off
, PAGESIZE
,
7810 (uint_t
*)NULL
, pl
, PAGESIZE
, seg
, addr
,
7811 S_OTHER
, svd
->cred
, NULL
);
7813 if (error
&& ap
!= NULL
) {
7814 anon_array_exit(&cookie
);
7815 ANON_LOCK_EXIT(&
->a_rwlock
);
7819 * If the error is EDEADLK then we must bounce
7820 * up and drop all vm subsystem locks and then
7821 * retry the operation later
7822 * This behavior is a temporary measure because
7823 * ufs/sds logging is badly designed and will
7824 * deadlock if we don't allow this bounce to
7825 * happen. The real solution is to re-design
7826 * the logging code to work properly. See bug
7827 * 4125102 for details of the problem.
7829 if (error
== EDEADLK
) {
7834 * Quit if we fail to fault in the page. Treat
7835 * the failure as an error, unless the addr
7836 * is mapped beyond the end of a file.
7838 if (error
&& svd
->vp
) {
7839 va
.va_mask
= AT_SIZE
;
7840 if (VOP_GETATTR(svd
->vp
, &va
, 0,
7841 svd
->cred
, NULL
) != 0) {
7845 if (btopr(va
.va_size
) >=
7861 * See Statement at the beginning of this routine.
7863 * claim is always set if MAP_PRIVATE and PROT_WRITE
7864 * irrespective of following factors:
7866 * (1) anon slots are populated or not
7867 * (2) cow is broken or not
7868 * (3) refcnt on ap is 1 or greater than 1
7870 * See 4140683 for details
7872 claim
= ((VPP_PROT(vpp
) & PROT_WRITE
) &&
7873 (svd
->type
== MAP_PRIVATE
));
7876 * Perform page-level operation appropriate to
7877 * operation. If locking, undo the SOFTLOCK
7878 * performed to bring the page into memory
7879 * after setting the lock. If unlocking,
7880 * and no page was found, account for the claim
7883 if (op
== MC_LOCK
) {
7884 int ret
= 1; /* Assume success */
7886 ASSERT(!VPP_ISPPLOCK(vpp
));
7888 ret
= page_pp_lock(pp
, claim
, 0);
7890 if (ap
->an_pvp
!= NULL
) {
7891 anon_swap_free(ap
, pp
);
7893 anon_array_exit(&cookie
);
7894 ANON_LOCK_EXIT(&
->a_rwlock
);
7897 /* locking page failed */
7904 if (pp
->p_lckcnt
== 1)
7905 locked_bytes
+= PAGESIZE
;
7907 locked_bytes
+= PAGESIZE
;
7909 if (lockmap
!= (ulong_t
*)NULL
)
7910 BT_SET(lockmap
, pos
);
7914 ASSERT(VPP_ISPPLOCK(vpp
));
7916 /* sysV pages should be locked */
7917 ASSERT(sp
== NULL
|| pp
->p_lckcnt
> 0);
7918 page_pp_unlock(pp
, claim
, 0);
7920 if (pp
->p_lckcnt
== 0)
7924 unlocked_bytes
+= PAGESIZE
;
7928 unlocked_bytes
+= PAGESIZE
;
7935 if (op
== MC_LOCK
) {
7936 /* Credit back bytes that did not get locked */
7937 if ((unlocked_bytes
- locked_bytes
) > 0) {
7939 mutex_enter(&p
->p_lock
);
7940 rctl_decr_locked_mem(p
, proj
,
7941 (unlocked_bytes
- locked_bytes
), chargeproc
);
7943 mutex_exit(&p
->p_lock
);
7947 /* Account bytes that were unlocked */
7948 if (unlocked_bytes
> 0) {
7950 mutex_enter(&p
->p_lock
);
7951 rctl_decr_locked_mem(p
, proj
, unlocked_bytes
,
7954 mutex_exit(&p
->p_lock
);
7958 mutex_exit(&sp
->shm_mlock
);
7959 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
7965 * Set advice from user for specified pages
7966 * There are 5 types of advice:
7967 * MADV_NORMAL - Normal (default) behavior (whatever that is)
7968 * MADV_RANDOM - Random page references
7969 * do not allow readahead or 'klustering'
7970 * MADV_SEQUENTIAL - Sequential page references
7971 * Pages previous to the one currently being
7972 * accessed (determined by fault) are 'not needed'
7973 * and are freed immediately
7974 * MADV_WILLNEED - Pages are likely to be used (fault ahead in mctl)
7975 * MADV_DONTNEED - Pages are not needed (synced out in mctl)
7976 * MADV_FREE - Contents can be discarded
7977 * MADV_ACCESS_DEFAULT- Default access
7978 * MADV_ACCESS_LWP - Next LWP will access heavily
7979 * MADV_ACCESS_MANY- Many LWPs or processes will access heavily
7982 segvn_advise(struct seg
*seg
, caddr_t addr
, size_t len
, uint_t behav
)
7984 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
7988 struct anon_map
*amp
;
7991 lgrp_mem_policy_t policy
;
7995 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
7998 * In case of MADV_FREE, we won't be modifying any segment private
7999 * data structures; so, we only need to grab READER's lock
8001 if (behav
!= MADV_FREE
) {
8002 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_WRITER
);
8003 if (svd
->tr_state
!= SEGVN_TR_OFF
) {
8004 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
8008 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_READER
);
8012 * Large pages are assumed to be only turned on when accesses to the
8013 * segment's address range have spatial and temporal locality. That
8014 * justifies ignoring MADV_SEQUENTIAL for large page segments.
8015 * Also, ignore advice affecting lgroup memory allocation
8016 * if don't need to do lgroup optimizations on this system
8019 if ((behav
== MADV_SEQUENTIAL
&&
8020 (seg
->s_szc
!= 0 || HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
))) ||
8021 (!lgrp_optimizations() && (behav
== MADV_ACCESS_DEFAULT
||
8022 behav
== MADV_ACCESS_LWP
|| behav
== MADV_ACCESS_MANY
))) {
8023 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
8027 if (behav
== MADV_SEQUENTIAL
|| behav
== MADV_ACCESS_DEFAULT
||
8028 behav
== MADV_ACCESS_LWP
|| behav
== MADV_ACCESS_MANY
) {
8030 * Since we are going to unload hat mappings
8031 * we first have to flush the cache. Otherwise
8032 * this might lead to system panic if another
8033 * thread is doing physio on the range whose
8034 * mappings are unloaded by madvise(3C).
8036 if (svd
->softlockcnt
> 0) {
8038 * If this is shared segment non 0 softlockcnt
8039 * means locked pages are still in use.
8041 if (svd
->type
== MAP_SHARED
) {
8042 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
8046 * Since we do have the segvn writers lock
8047 * nobody can fill the cache with entries
8048 * belonging to this seg during the purge.
8049 * The flush either succeeds or we still
8050 * have pending I/Os. In the later case,
8051 * madvise(3C) fails.
8054 if (svd
->softlockcnt
> 0) {
8056 * Since madvise(3C) is advisory and
8057 * it's not part of UNIX98, madvise(3C)
8058 * failure here doesn't cause any hardship.
8059 * Note that we don't block in "as" layer.
8061 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
8064 } else if (svd
->type
== MAP_SHARED
&& svd
->amp
!= NULL
&&
8065 svd
->amp
->a_softlockcnt
> 0) {
8067 * Try to purge this amp's entries from pcache. It
8068 * will succeed only if other segments that share the
8069 * amp have no outstanding softlock's.
8077 if (behav
== MADV_FREE
) {
8079 * MADV_FREE is not supported for segments with
8080 * underlying object; if anonmap is NULL, anon slots
8081 * are not yet populated and there is nothing for
8082 * us to do. As MADV_FREE is advisory, we don't
8083 * return error in either case.
8085 if (vp
!= NULL
|| amp
== NULL
) {
8086 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
8092 page
= seg_page(seg
, addr
);
8093 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
8094 anon_disclaim(amp
, svd
->anon_index
+ page
, len
);
8095 ANON_LOCK_EXIT(&
->a_rwlock
);
8096 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
8101 * If advice is to be applied to entire segment,
8102 * use advice field in seg_data structure
8103 * otherwise use appropriate vpage entry.
8105 if ((addr
== seg
->s_base
) && (len
== seg
->s_size
)) {
8107 case MADV_ACCESS_LWP
:
8108 case MADV_ACCESS_MANY
:
8109 case MADV_ACCESS_DEFAULT
:
8111 * Set memory allocation policy for this segment
8113 policy
= lgrp_madv_to_policy(behav
, len
, svd
->type
);
8114 if (svd
->type
== MAP_SHARED
)
8115 already_set
= lgrp_shm_policy_set(policy
, amp
,
8116 svd
->anon_index
, vp
, svd
->offset
, len
);
8119 * For private memory, need writers lock on
8120 * address space because the segment may be
8121 * split or concatenated when changing policy
8123 if (AS_READ_HELD(seg
->s_as
,
8124 &seg
->s_as
->a_lock
)) {
8125 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
8129 already_set
= lgrp_privm_policy_set(policy
,
8130 &svd
->policy_info
, len
);
8134 * If policy set already and it shouldn't be reapplied,
8135 * don't do anything.
8138 !LGRP_MEM_POLICY_REAPPLICABLE(policy
))
8142 * Mark any existing pages in given range for
8145 page_mark_migrate(seg
, addr
, len
, amp
, svd
->anon_index
,
8146 vp
, svd
->offset
, 1);
8149 * If same policy set already or this is a shared
8150 * memory segment, don't need to try to concatenate
8151 * segment with adjacent ones.
8153 if (already_set
|| svd
->type
== MAP_SHARED
)
8157 * Try to concatenate this segment with previous
8158 * one and next one, since we changed policy for
8159 * this one and it may be compatible with adjacent
8162 prev
= AS_SEGPREV(seg
->s_as
, seg
);
8163 next
= AS_SEGNEXT(seg
->s_as
, seg
);
8165 if (next
&& next
->s_ops
== &segvn_ops
&&
8166 addr
+ len
== next
->s_base
)
8167 (void) segvn_concat(seg
, next
, 1);
8169 if (prev
&& prev
->s_ops
== &segvn_ops
&&
8170 addr
== prev
->s_base
+ prev
->s_size
) {
8172 * Drop lock for private data of current
8173 * segment before concatenating (deleting) it
8174 * and return IE_REATTACH to tell as_ctl() that
8175 * current segment has changed
8177 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
8178 if (!segvn_concat(prev
, seg
, 1))
8185 case MADV_SEQUENTIAL
:
8187 * unloading mapping guarantees
8188 * detection in segvn_fault
8190 ASSERT(seg
->s_szc
== 0);
8191 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
8192 hat_unload(seg
->s_as
->a_hat
, addr
, len
,
8197 svd
->advice
= (uchar_t
)behav
;
8198 svd
->pageadvice
= 0;
8200 case MADV_WILLNEED
: /* handled in memcntl */
8201 case MADV_DONTNEED
: /* handled in memcntl */
8202 case MADV_FREE
: /* handled above */
8209 struct seg
*new_seg
;
8210 struct segvn_data
*new_svd
;
8214 page
= seg_page(seg
, addr
);
8219 struct vpage
*bvpp
, *evpp
;
8221 case MADV_ACCESS_LWP
:
8222 case MADV_ACCESS_MANY
:
8223 case MADV_ACCESS_DEFAULT
:
8225 * Set memory allocation policy for portion of this
8230 * Align address and length of advice to page
8231 * boundaries for large pages
8233 if (seg
->s_szc
!= 0) {
8236 pgsz
= page_get_pagesize(seg
->s_szc
);
8237 addr
= (caddr_t
)P2ALIGN((uintptr_t)addr
, pgsz
);
8238 len
= P2ROUNDUP(len
, pgsz
);
8242 * Check to see whether policy is set already
8244 policy
= lgrp_madv_to_policy(behav
, len
, svd
->type
);
8246 anon_index
= svd
->anon_index
+ page
;
8247 off
= svd
->offset
+ (uintptr_t)(addr
- seg
->s_base
);
8249 if (svd
->type
== MAP_SHARED
)
8250 already_set
= lgrp_shm_policy_set(policy
, amp
,
8251 anon_index
, vp
, off
, len
);
8254 (policy
== svd
->policy_info
.mem_policy
);
8257 * If policy set already and it shouldn't be reapplied,
8258 * don't do anything.
8261 !LGRP_MEM_POLICY_REAPPLICABLE(policy
))
8265 * For private memory, need writers lock on
8266 * address space because the segment may be
8267 * split or concatenated when changing policy
8269 if (svd
->type
== MAP_PRIVATE
&&
8270 AS_READ_HELD(seg
->s_as
, &seg
->s_as
->a_lock
)) {
8271 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
8276 * Mark any existing pages in given range for
8279 page_mark_migrate(seg
, addr
, len
, amp
, svd
->anon_index
,
8280 vp
, svd
->offset
, 1);
8283 * Don't need to try to split or concatenate
8284 * segments, since policy is same or this is a shared
8287 if (already_set
|| svd
->type
== MAP_SHARED
)
8290 if (HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
)) {
8291 ASSERT(svd
->amp
== NULL
);
8292 ASSERT(svd
->tr_state
== SEGVN_TR_OFF
);
8293 ASSERT(svd
->softlockcnt
== 0);
8294 hat_leave_region(seg
->s_as
->a_hat
, svd
->rcookie
,
8296 svd
->rcookie
= HAT_INVALID_REGION_COOKIE
;
8300 * Split off new segment if advice only applies to a
8301 * portion of existing segment starting in middle
8305 oldeaddr
= seg
->s_base
+ seg
->s_size
;
8306 if (addr
> seg
->s_base
) {
8308 * Must flush I/O page cache
8309 * before splitting segment
8311 if (svd
->softlockcnt
> 0)
8315 * Split segment and return IE_REATTACH to tell
8316 * as_ctl() that current segment changed
8318 new_seg
= segvn_split_seg(seg
, addr
);
8319 new_svd
= (struct segvn_data
*)new_seg
->s_data
;
8323 * If new segment ends where old one
8324 * did, try to concatenate the new
8325 * segment with next one.
8327 if (eaddr
== oldeaddr
) {
8329 * Set policy for new segment
8331 (void) lgrp_privm_policy_set(policy
,
8332 &new_svd
->policy_info
,
8335 next
= AS_SEGNEXT(new_seg
->s_as
,
8339 next
->s_ops
== &segvn_ops
&&
8340 eaddr
== next
->s_base
)
8341 (void) segvn_concat(new_seg
,
8347 * Split off end of existing segment if advice only
8348 * applies to a portion of segment ending before
8349 * end of the existing segment
8351 if (eaddr
< oldeaddr
) {
8353 * Must flush I/O page cache
8354 * before splitting segment
8356 if (svd
->softlockcnt
> 0)
8360 * If beginning of old segment was already
8361 * split off, use new segment to split end off
8364 if (new_seg
!= NULL
&& new_seg
!= seg
) {
8368 (void) segvn_split_seg(new_seg
, eaddr
);
8371 * Set policy for new segment
8373 (void) lgrp_privm_policy_set(policy
,
8374 &new_svd
->policy_info
,
8378 * Split segment and return IE_REATTACH
8379 * to tell as_ctl() that current
8382 (void) segvn_split_seg(seg
, eaddr
);
8385 (void) lgrp_privm_policy_set(policy
,
8386 &svd
->policy_info
, seg
->s_size
);
8389 * If new segment starts where old one
8390 * did, try to concatenate it with
8393 if (addr
== seg
->s_base
) {
8394 prev
= AS_SEGPREV(seg
->s_as
,
8398 * Drop lock for private data
8399 * of current segment before
8400 * concatenating (deleting) it
8405 addr
== prev
->s_base
+
8410 (void) segvn_concat(
8418 case MADV_SEQUENTIAL
:
8419 ASSERT(seg
->s_szc
== 0);
8420 ASSERT(svd
->rcookie
== HAT_INVALID_REGION_COOKIE
);
8421 hat_unload(seg
->s_as
->a_hat
, addr
, len
, HAT_UNLOAD
);
8425 bvpp
= &svd
->vpage
[page
];
8426 evpp
= &svd
->vpage
[page
+ (len
>> PAGESHIFT
)];
8427 for (; bvpp
< evpp
; bvpp
++)
8428 VPP_SETADVICE(bvpp
, behav
);
8429 svd
->advice
= MADV_NORMAL
;
8431 case MADV_WILLNEED
: /* handled in memcntl */
8432 case MADV_DONTNEED
: /* handled in memcntl */
8433 case MADV_FREE
: /* handled above */
8439 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
8444 * Create a vpage structure for this seg.
8447 segvn_vpage(struct seg
*seg
)
8449 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
8450 struct vpage
*vp
, *evp
;
8452 ASSERT(SEGVN_WRITE_HELD(seg
->s_as
, &svd
->lock
));
8455 * If no vpage structure exists, allocate one. Copy the protections
8456 * and the advice from the segment itself to the individual pages.
8458 if (svd
->vpage
== NULL
) {
8459 svd
->pageadvice
= 1;
8460 svd
->vpage
= kmem_zalloc(seg_pages(seg
) * sizeof (struct vpage
),
8462 evp
= &svd
->vpage
[seg_page(seg
, seg
->s_base
+ seg
->s_size
)];
8463 for (vp
= svd
->vpage
; vp
< evp
; vp
++) {
8464 VPP_SETPROT(vp
, svd
->prot
);
8465 VPP_SETADVICE(vp
, svd
->advice
);
8471 * Dump the pages belonging to this segvn segment.
8474 segvn_dump(struct seg
*seg
)
8476 struct segvn_data
*svd
;
8478 struct anon_map
*amp
;
8481 u_offset_t off
, offset
;
8483 pgcnt_t page
, npages
;
8486 npages
= seg_pages(seg
);
8487 svd
= (struct segvn_data
*)seg
->s_data
;
8489 off
= offset
= svd
->offset
;
8492 if ((amp
= svd
->amp
) != NULL
) {
8493 anon_index
= svd
->anon_index
;
8494 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
8497 for (page
= 0; page
< npages
; page
++, offset
+= PAGESIZE
) {
8501 if (amp
&& (ap
= anon_get_ptr(svd
->amp
->ahp
, anon_index
++))) {
8502 swap_xlate_nopanic(ap
, &vp
, &off
);
8509 * If pp == NULL, the page either does not exist
8510 * or is exclusively locked. So determine if it
8511 * exists before searching for it.
8514 if ((pp
= page_lookup_nowait(vp
, off
, SE_SHARED
)))
8517 pp
= page_exists(vp
, off
);
8520 pfn
= page_pptonum(pp
);
8521 dump_addpage(seg
->s_as
, addr
, pfn
);
8526 dump_timeleft
= dump_timeout
;
8530 ANON_LOCK_EXIT(&
->a_rwlock
);
8534 static uint32_t segvn_pglock_mtbf
= 0;
8537 #define PCACHE_SHWLIST ((page_t *)-2)
8538 #define NOPCACHE_SHWLIST ((page_t *)-1)
8541 * Lock/Unlock anon pages over a given range. Return shadow list. This routine
8542 * uses global segment pcache to cache shadow lists (i.e. pp arrays) of pages
8543 * to avoid the overhead of per page locking, unlocking for subsequent IOs to
8544 * the same parts of the segment. Currently shadow list creation is only
8545 * supported for pure anon segments. MAP_PRIVATE segment pcache entries are
8546 * tagged with segment pointer, starting virtual address and length. This
8547 * approach for MAP_SHARED segments may add many pcache entries for the same
8548 * set of pages and lead to long hash chains that decrease pcache lookup
8549 * performance. To avoid this issue for shared segments shared anon map and
8550 * starting anon index are used for pcache entry tagging. This allows all
8551 * segments to share pcache entries for the same anon range and reduces pcache
8552 * chain's length as well as memory overhead from duplicate shadow lists and
8555 * softlockcnt field in segvn_data structure counts the number of F_SOFTLOCK'd
8556 * pages via segvn_fault() and pagelock'd pages via this routine. But pagelock
8557 * part of softlockcnt accounting is done differently for private and shared
8558 * segments. In private segment case softlock is only incremented when a new
8559 * shadow list is created but not when an existing one is found via
8560 * seg_plookup(). pcache entries have reference count incremented/decremented
8561 * by each seg_plookup()/seg_pinactive() operation. Only entries that have 0
8562 * reference count can be purged (and purging is needed before segment can be
8563 * freed). When a private segment pcache entry is purged segvn_reclaim() will
8564 * decrement softlockcnt. Since in private segment case each of its pcache
8565 * entries only belongs to this segment we can expect that when
8566 * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8567 * segment purge will succeed and softlockcnt will drop to 0. In shared
8568 * segment case reference count in pcache entry counts active locks from many
8569 * different segments so we can't expect segment purging to succeed even when
8570 * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8571 * segment. To be able to determine when there're no pending pagelocks in
8572 * shared segment case we don't rely on purging to make softlockcnt drop to 0
8573 * but instead softlockcnt is incremented and decremented for every
8574 * segvn_pagelock(L_PAGELOCK/L_PAGEUNLOCK) call regardless if a new shadow
8575 * list was created or an existing one was found. When softlockcnt drops to 0
8576 * this segment no longer has any claims for pcached shadow lists and the
8577 * segment can be freed even if there're still active pcache entries
8578 * shared by this segment anon map. Shared segment pcache entries belong to
8579 * anon map and are typically removed when anon map is freed after all
8580 * processes destroy the segments that use this anon map.
8583 segvn_pagelock(struct seg
*seg
, caddr_t addr
, size_t len
, struct page
***ppp
,
8584 enum lock_type type
, enum seg_rw rw
)
8586 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
8588 pgcnt_t adjustpages
;
8591 uint_t protchk
= (rw
== S_READ
) ? PROT_READ
: PROT_WRITE
;
8593 struct anon_map
*amp
;
8595 struct page
**pplist
, **pl
, *pp
;
8598 caddr_t lpgaddr
, lpgeaddr
;
8599 anon_sync_obj_t cookie
;
8601 struct anon_map
*pamp
;
8603 seg_preclaim_cbfunc_t preclaim_callback
;
8608 int sftlck_sbase
= 0;
8609 int sftlck_send
= 0;
8612 if (type
== L_PAGELOCK
&& segvn_pglock_mtbf
) {
8613 hrtime_t ts
= gethrtime();
8614 if ((ts
% segvn_pglock_mtbf
) == 0) {
8617 if ((ts
% segvn_pglock_mtbf
) == 1) {
8623 TRACE_2(TR_FAC_PHYSIO
, TR_PHYSIO_SEGVN_START
,
8624 "segvn_pagelock: start seg %p addr %p", seg
, addr
);
8626 ASSERT(seg
->s_as
&& AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
8627 ASSERT(type
== L_PAGELOCK
|| type
== L_PAGEUNLOCK
);
8629 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_READER
);
8632 * for now we only support pagelock to anon memory. We would have to
8633 * check protections for vnode objects and call into the vnode driver.
8634 * That's too much for a fast path. Let the fault entry point handle
8637 if (svd
->vp
!= NULL
) {
8638 if (type
== L_PAGELOCK
) {
8642 panic("segvn_pagelock(L_PAGEUNLOCK): vp != NULL");
8644 if ((amp
= svd
->amp
) == NULL
) {
8645 if (type
== L_PAGELOCK
) {
8649 panic("segvn_pagelock(L_PAGEUNLOCK): amp == NULL");
8651 if (rw
!= S_READ
&& rw
!= S_WRITE
) {
8652 if (type
== L_PAGELOCK
) {
8656 panic("segvn_pagelock(L_PAGEUNLOCK): bad rw");
8659 if (seg
->s_szc
!= 0) {
8661 * We are adjusting the pagelock region to the large page size
8662 * boundary because the unlocked part of a large page cannot
8663 * be freed anyway unless all constituent pages of a large
8664 * page are locked. Bigger regions reduce pcache chain length
8665 * and improve lookup performance. The tradeoff is that the
8666 * very first segvn_pagelock() call for a given page is more
8667 * expensive if only 1 page_t is needed for IO. This is only
8668 * an issue if pcache entry doesn't get reused by several
8669 * subsequent calls. We optimize here for the case when pcache
8670 * is heavily used by repeated IOs to the same address range.
8672 * Note segment's page size cannot change while we are holding
8673 * as lock. And then it cannot change while softlockcnt is
8674 * not 0. This will allow us to correctly recalculate large
8675 * page size region for the matching pageunlock/reclaim call
8676 * since as_pageunlock() caller must always match
8677 * as_pagelock() call's addr and len.
8679 * For pageunlock *ppp points to the pointer of page_t that
8680 * corresponds to the real unadjusted start address. Similar
8681 * for pagelock *ppp must point to the pointer of page_t that
8682 * corresponds to the real unadjusted start address.
8684 pgsz
= page_get_pagesize(seg
->s_szc
);
8685 CALC_LPG_REGION(pgsz
, seg
, addr
, len
, lpgaddr
, lpgeaddr
);
8686 adjustpages
= btop((uintptr_t)(addr
- lpgaddr
));
8687 } else if (len
< segvn_pglock_comb_thrshld
) {
8689 lpgeaddr
= addr
+ len
;
8694 * Align the address range of large enough requests to allow
8695 * combining of different shadow lists into 1 to reduce memory
8696 * overhead from potentially overlapping large shadow lists
8697 * (worst case is we have a 1MB IO into buffers with start
8698 * addresses separated by 4K). Alignment is only possible if
8699 * padded chunks have sufficient access permissions. Note
8700 * permissions won't change between L_PAGELOCK and
8701 * L_PAGEUNLOCK calls since non 0 softlockcnt will force
8702 * segvn_setprot() to wait until softlockcnt drops to 0. This
8703 * allows us to determine in L_PAGEUNLOCK the same range we
8704 * computed in L_PAGELOCK.
8706 * If alignment is limited by segment ends set
8707 * sftlck_sbase/sftlck_send flags. In L_PAGELOCK case when
8708 * these flags are set bump softlockcnt_sbase/softlockcnt_send
8709 * per segment counters. In L_PAGEUNLOCK case decrease
8710 * softlockcnt_sbase/softlockcnt_send counters if
8711 * sftlck_sbase/sftlck_send flags are set. When
8712 * softlockcnt_sbase/softlockcnt_send are non 0
8713 * segvn_concat()/segvn_extend_prev()/segvn_extend_next()
8714 * won't merge the segments. This restriction combined with
8715 * restriction on segment unmapping and splitting for segments
8716 * that have non 0 softlockcnt allows L_PAGEUNLOCK to
8717 * correctly determine the same range that was previously
8718 * locked by matching L_PAGELOCK.
8720 pflags
= SEGP_PSHIFT
| (segvn_pglock_comb_bshift
<< 16);
8722 if (svd
->type
== MAP_PRIVATE
) {
8723 lpgaddr
= (caddr_t
)P2ALIGN((uintptr_t)addr
,
8724 segvn_pglock_comb_balign
);
8725 if (lpgaddr
< seg
->s_base
) {
8726 lpgaddr
= seg
->s_base
;
8730 ulong_t aix
= svd
->anon_index
+ seg_page(seg
, addr
);
8731 ulong_t aaix
= P2ALIGN(aix
, segvn_pglock_comb_palign
);
8732 if (aaix
< svd
->anon_index
) {
8733 lpgaddr
= seg
->s_base
;
8736 lpgaddr
= addr
- ptob(aix
- aaix
);
8737 ASSERT(lpgaddr
>= seg
->s_base
);
8740 if (svd
->pageprot
&& lpgaddr
!= addr
) {
8741 struct vpage
*vp
= &svd
->vpage
[seg_page(seg
, lpgaddr
)];
8742 struct vpage
*evp
= &svd
->vpage
[seg_page(seg
, addr
)];
8744 if ((VPP_PROT(vp
) & protchk
) == 0) {
8754 lpgeaddr
= addr
+ len
;
8756 if (svd
->type
== MAP_PRIVATE
) {
8757 lpgeaddr
= (caddr_t
)P2ROUNDUP(
8758 (uintptr_t)lpgeaddr
,
8759 segvn_pglock_comb_balign
);
8761 ulong_t aix
= svd
->anon_index
+
8762 seg_page(seg
, lpgeaddr
);
8763 ulong_t aaix
= P2ROUNDUP(aix
,
8764 segvn_pglock_comb_palign
);
8768 lpgeaddr
+= ptob(aaix
- aix
);
8771 if (lpgeaddr
== 0 ||
8772 lpgeaddr
> seg
->s_base
+ seg
->s_size
) {
8773 lpgeaddr
= seg
->s_base
+ seg
->s_size
;
8777 if (svd
->pageprot
&& lpgeaddr
!= addr
+ len
) {
8781 vp
= &svd
->vpage
[seg_page(seg
, addr
+ len
)];
8782 evp
= &svd
->vpage
[seg_page(seg
, lpgeaddr
)];
8785 if ((VPP_PROT(vp
) & protchk
) == 0) {
8791 lpgeaddr
= addr
+ len
;
8794 adjustpages
= btop((uintptr_t)(addr
- lpgaddr
));
8798 * For MAP_SHARED segments we create pcache entries tagged by amp and
8799 * anon index so that we can share pcache entries with other segments
8800 * that map this amp. For private segments pcache entries are tagged
8801 * with segment and virtual address.
8803 if (svd
->type
== MAP_SHARED
) {
8805 paddr
= (caddr_t
)((lpgaddr
- seg
->s_base
) +
8806 ptob(svd
->anon_index
));
8807 preclaim_callback
= shamp_reclaim
;
8811 preclaim_callback
= segvn_reclaim
;
8814 if (type
== L_PAGEUNLOCK
) {
8815 VM_STAT_ADD(segvnvmstats
.pagelock
[0]);
8818 * update hat ref bits for /proc. We need to make sure
8819 * that threads tracing the ref and mod bits of the
8820 * address space get the right data.
8821 * Note: page ref and mod bits are updated at reclaim time
8823 if (seg
->s_as
->a_vbits
) {
8824 for (a
= addr
; a
< addr
+ len
; a
+= PAGESIZE
) {
8825 if (rw
== S_WRITE
) {
8826 hat_setstat(seg
->s_as
, a
,
8827 PAGESIZE
, P_REF
| P_MOD
);
8829 hat_setstat(seg
->s_as
, a
,
8836 * Check the shadow list entry after the last page used in
8837 * this IO request. If it's NOPCACHE_SHWLIST the shadow list
8838 * was not inserted into pcache and is not large page
8839 * adjusted. In this case call reclaim callback directly and
8840 * don't adjust the shadow list start and size for large
8844 if ((*ppp
)[npages
] == NOPCACHE_SHWLIST
) {
8847 ASSERT(svd
->type
== MAP_SHARED
);
8848 ptag
= (void *)pamp
;
8849 paddr
= (caddr_t
)((addr
- seg
->s_base
) +
8850 ptob(svd
->anon_index
));
8855 (*preclaim_callback
)(ptag
, paddr
, len
, *ppp
, rw
, 0);
8857 ASSERT((*ppp
)[npages
] == PCACHE_SHWLIST
||
8858 IS_SWAPFSVP((*ppp
)[npages
]->p_vnode
));
8859 len
= lpgeaddr
- lpgaddr
;
8861 seg_pinactive(seg
, pamp
, paddr
, len
,
8862 *ppp
- adjustpages
, rw
, pflags
, preclaim_callback
);
8866 ASSERT(svd
->type
== MAP_SHARED
);
8867 ASSERT(svd
->softlockcnt
>= npages
);
8868 atomic_add_long((ulong_t
*)&svd
->softlockcnt
, -npages
);
8872 ASSERT(svd
->softlockcnt_sbase
> 0);
8873 atomic_add_long((ulong_t
*)&svd
->softlockcnt_sbase
, -1);
8876 ASSERT(svd
->softlockcnt_send
> 0);
8877 atomic_add_long((ulong_t
*)&svd
->softlockcnt_send
, -1);
8881 * If someone is blocked while unmapping, we purge
8882 * segment page cache and thus reclaim pplist synchronously
8883 * without waiting for seg_pasync_thread. This speeds up
8884 * unmapping in cases where munmap(2) is called, while
8885 * raw async i/o is still in progress or where a thread
8886 * exits on data fault in a multithreaded application.
8888 if (AS_ISUNMAPWAIT(seg
->s_as
)) {
8889 if (svd
->softlockcnt
== 0) {
8890 mutex_enter(&seg
->s_as
->a_contents
);
8891 if (AS_ISUNMAPWAIT(seg
->s_as
)) {
8892 AS_CLRUNMAPWAIT(seg
->s_as
);
8893 cv_broadcast(&seg
->s_as
->a_cv
);
8895 mutex_exit(&seg
->s_as
->a_contents
);
8896 } else if (pamp
== NULL
) {
8898 * softlockcnt is not 0 and this is a
8899 * MAP_PRIVATE segment. Try to purge its
8900 * pcache entries to reduce softlockcnt.
8901 * If it drops to 0 segvn_reclaim()
8902 * will wake up a thread waiting on
8905 * We don't purge MAP_SHARED segments with non
8906 * 0 softlockcnt since IO is still in progress
8907 * for such segments.
8909 ASSERT(svd
->type
== MAP_PRIVATE
);
8913 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
8914 TRACE_2(TR_FAC_PHYSIO
, TR_PHYSIO_SEGVN_UNLOCK_END
,
8915 "segvn_pagelock: unlock seg %p addr %p", seg
, addr
);
8919 /* The L_PAGELOCK case ... */
8921 VM_STAT_ADD(segvnvmstats
.pagelock
[1]);
8924 * For MAP_SHARED segments we have to check protections before
8925 * seg_plookup() since pcache entries may be shared by many segments
8926 * with potentially different page protections.
8929 ASSERT(svd
->type
== MAP_SHARED
);
8930 if (svd
->pageprot
== 0) {
8931 if ((svd
->prot
& protchk
) == 0) {
8937 * check page protections
8948 for (; a
< ea
; a
+= pgsz
) {
8951 ASSERT(seg
->s_szc
== 0 ||
8952 sameprot(seg
, a
, pgsz
));
8953 vp
= &svd
->vpage
[seg_page(seg
, a
)];
8954 if ((VPP_PROT(vp
) & protchk
) == 0) {
8963 * try to find pages in segment page cache
8965 pplist
= seg_plookup(seg
, pamp
, paddr
, lpgeaddr
- lpgaddr
, rw
, pflags
);
8966 if (pplist
!= NULL
) {
8968 npages
= btop((uintptr_t)(lpgeaddr
- lpgaddr
));
8969 ASSERT(svd
->type
== MAP_SHARED
);
8970 atomic_add_long((ulong_t
*)&svd
->softlockcnt
,
8974 atomic_add_long((ulong_t
*)&svd
->softlockcnt_sbase
, 1);
8977 atomic_add_long((ulong_t
*)&svd
->softlockcnt_send
, 1);
8979 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
8980 *ppp
= pplist
+ adjustpages
;
8981 TRACE_2(TR_FAC_PHYSIO
, TR_PHYSIO_SEGVN_HIT_END
,
8982 "segvn_pagelock: cache hit seg %p addr %p", seg
, addr
);
8987 * For MAP_SHARED segments we already verified above that segment
8988 * protections allow this pagelock operation.
8991 ASSERT(svd
->type
== MAP_PRIVATE
);
8992 if (svd
->pageprot
== 0) {
8993 if ((svd
->prot
& protchk
) == 0) {
8997 if (svd
->prot
& PROT_WRITE
) {
8998 wlen
= lpgeaddr
- lpgaddr
;
9001 ASSERT(rw
== S_READ
);
9006 * check page protections
9008 for (a
= lpgaddr
, wlen
= 0; a
< lpgeaddr
; a
+= pgsz
) {
9011 ASSERT(seg
->s_szc
== 0 ||
9012 sameprot(seg
, a
, pgsz
));
9013 vp
= &svd
->vpage
[seg_page(seg
, a
)];
9014 if ((VPP_PROT(vp
) & protchk
) == 0) {
9018 if (wcont
&& (VPP_PROT(vp
) & PROT_WRITE
)) {
9022 ASSERT(rw
== S_READ
);
9026 ASSERT(rw
== S_READ
|| wlen
== lpgeaddr
- lpgaddr
);
9027 ASSERT(rw
== S_WRITE
|| wlen
<= lpgeaddr
- lpgaddr
);
9031 * Only build large page adjusted shadow list if we expect to insert
9032 * it into pcache. For large enough pages it's a big overhead to
9033 * create a shadow list of the entire large page. But this overhead
9034 * should be amortized over repeated pcache hits on subsequent reuse
9035 * of this shadow list (IO into any range within this shadow list will
9036 * find it in pcache since we large page align the request for pcache
9037 * lookups). pcache performance is improved with bigger shadow lists
9038 * as it reduces the time to pcache the entire big segment and reduces
9039 * pcache chain length.
9041 if (seg_pinsert_check(seg
, pamp
, paddr
,
9042 lpgeaddr
- lpgaddr
, pflags
) == SEGP_SUCCESS
) {
9044 len
= lpgeaddr
- lpgaddr
;
9049 * Since this entry will not be inserted into the pcache, we
9050 * will not do any adjustments to the starting address or
9051 * size of the memory to be locked.
9057 pplist
= kmem_alloc(sizeof (page_t
*) * (npages
+ 1), KM_SLEEP
);
9059 *ppp
= pplist
+ adjustpages
;
9061 * If use_pcache is 0 this shadow list is not large page adjusted.
9062 * Record this info in the last entry of shadow array so that
9063 * L_PAGEUNLOCK can determine if it should large page adjust the
9064 * address range to find the real range that was locked.
9066 pl
[npages
] = use_pcache
? PCACHE_SHWLIST
: NOPCACHE_SHWLIST
;
9068 page
= seg_page(seg
, addr
);
9069 anon_index
= svd
->anon_index
+ page
;
9072 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
9073 ASSERT(amp
->a_szc
>= seg
->s_szc
);
9074 anpgcnt
= page_get_pagecnt(amp
->a_szc
);
9075 for (a
= addr
; a
< addr
+ len
; a
+= PAGESIZE
, anon_index
++) {
9081 * Lock and unlock anon array only once per large page.
9082 * anon_array_enter() locks the root anon slot according to
9083 * a_szc which can't change while anon map is locked. We lock
9084 * anon the first time through this loop and each time we
9085 * reach anon index that corresponds to a root of a large
9088 if (a
== addr
|| P2PHASE(anon_index
, anpgcnt
) == 0) {
9089 ASSERT(anlock
== 0);
9090 anon_array_enter(amp
, anon_index
, &cookie
);
9093 ap
= anon_get_ptr(amp
->ahp
, anon_index
);
9096 * We must never use seg_pcache for COW pages
9097 * because we might end up with original page still
9098 * lying in seg_pcache even after private page is
9099 * created. This leads to data corruption as
9100 * aio_write refers to the page still in cache
9101 * while all other accesses refer to the private
9104 if (ap
== NULL
|| ap
->an_refcnt
!= 1) {
9105 struct vpage
*vpage
;
9111 if (svd
->vpage
!= NULL
) {
9112 vpage
= &svd
->vpage
[seg_page(seg
, a
)];
9117 anon_array_exit(&cookie
);
9120 error
= segvn_faultpage(seg
->s_as
->a_hat
, seg
, a
, 0,
9121 vpage
, &pp
, 0, F_INVAL
, rw
, 1);
9123 error
= fc_decode(error
);
9126 anon_array_enter(amp
, anon_index
, &cookie
);
9128 ap
= anon_get_ptr(amp
->ahp
, anon_index
);
9129 if (ap
== NULL
|| ap
->an_refcnt
!= 1) {
9134 swap_xlate(ap
, &vp
, &off
);
9135 pp
= page_lookup_nowait(vp
, off
, SE_SHARED
);
9140 if (ap
->an_pvp
!= NULL
) {
9141 anon_swap_free(ap
, pp
);
9144 * Unlock anon if this is the last slot in a large page.
9146 if (P2PHASE(anon_index
, anpgcnt
) == anpgcnt
- 1) {
9148 anon_array_exit(&cookie
);
9153 if (anlock
) { /* Ensure the lock is dropped */
9154 anon_array_exit(&cookie
);
9156 ANON_LOCK_EXIT(&
->a_rwlock
);
9158 if (a
>= addr
+ len
) {
9159 atomic_add_long((ulong_t
*)&svd
->softlockcnt
, npages
);
9161 ASSERT(svd
->type
== MAP_SHARED
);
9162 atomic_add_long((ulong_t
*)&pamp
->a_softlockcnt
,
9167 atomic_add_long((ulong_t
*)&svd
->softlockcnt_sbase
, 1);
9170 atomic_add_long((ulong_t
*)&svd
->softlockcnt_send
, 1);
9173 (void) seg_pinsert(seg
, pamp
, paddr
, len
, wlen
, pl
,
9174 rw
, pflags
, preclaim_callback
);
9176 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
9177 TRACE_2(TR_FAC_PHYSIO
, TR_PHYSIO_SEGVN_FILL_END
,
9178 "segvn_pagelock: cache fill seg %p addr %p", seg
, addr
);
9183 np
= ((uintptr_t)(a
- addr
)) >> PAGESHIFT
;
9184 while (np
> (uint_t
)0) {
9185 ASSERT(PAGE_LOCKED(*pplist
));
9186 page_unlock(*pplist
);
9190 kmem_free(pl
, sizeof (page_t
*) * (npages
+ 1));
9192 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
9194 TRACE_2(TR_FAC_PHYSIO
, TR_PHYSIO_SEGVN_MISS_END
,
9195 "segvn_pagelock: cache miss seg %p addr %p", seg
, addr
);
9200 * purge any cached pages in the I/O page cache
9203 segvn_purge(struct seg
*seg
)
9205 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
9208 * pcache is only used by pure anon segments.
9210 if (svd
->amp
== NULL
|| svd
->vp
!= NULL
) {
9215 * For MAP_SHARED segments non 0 segment's softlockcnt means
9216 * active IO is still in progress via this segment. So we only
9217 * purge MAP_SHARED segments when their softlockcnt is 0.
9219 if (svd
->type
== MAP_PRIVATE
) {
9220 if (svd
->softlockcnt
) {
9221 seg_ppurge(seg
, NULL
, 0);
9223 } else if (svd
->softlockcnt
== 0 && svd
->amp
->a_softlockcnt
!= 0) {
9224 seg_ppurge(seg
, svd
->amp
, 0);
9229 * If async argument is not 0 we are called from pcache async thread and don't
9235 segvn_reclaim(void *ptag
, caddr_t addr
, size_t len
, struct page
**pplist
,
9236 enum seg_rw rw
, int async
)
9238 struct seg
*seg
= (struct seg
*)ptag
;
9239 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
9243 npages
= np
= btop(len
);
9246 ASSERT(svd
->vp
== NULL
&& svd
->amp
!= NULL
);
9247 ASSERT(svd
->softlockcnt
>= npages
);
9248 ASSERT(async
|| AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
9252 ASSERT(pl
[np
] == NOPCACHE_SHWLIST
|| pl
[np
] == PCACHE_SHWLIST
);
9253 ASSERT(!async
|| pl
[np
] == PCACHE_SHWLIST
);
9255 while (np
> (uint_t
)0) {
9256 if (rw
== S_WRITE
) {
9257 hat_setrefmod(*pplist
);
9259 hat_setref(*pplist
);
9261 page_unlock(*pplist
);
9266 kmem_free(pl
, sizeof (page_t
*) * (npages
+ 1));
9269 * If we are pcache async thread we don't hold AS lock. This means if
9270 * softlockcnt drops to 0 after the decrement below address space may
9271 * get freed. We can't allow it since after softlock derement to 0 we
9272 * still need to access as structure for possible wakeup of unmap
9273 * waiters. To prevent the disappearance of as we take this segment
9274 * segfree_syncmtx. segvn_free() also takes this mutex as a barrier to
9275 * make sure this routine completes before segment is freed.
9277 * The second complication we have to deal with in async case is a
9278 * possibility of missed wake up of unmap wait thread. When we don't
9279 * hold as lock here we may take a_contents lock before unmap wait
9280 * thread that was first to see softlockcnt was still not 0. As a
9281 * result we'll fail to wake up an unmap wait thread. To avoid this
9282 * race we set nounmapwait flag in as structure if we drop softlockcnt
9283 * to 0 when we were called by pcache async thread. unmapwait thread
9284 * will not block if this flag is set.
9287 mutex_enter(&svd
->segfree_syncmtx
);
9290 if (!atomic_add_long_nv((ulong_t
*)&svd
->softlockcnt
, -npages
)) {
9291 if (async
|| AS_ISUNMAPWAIT(seg
->s_as
)) {
9292 mutex_enter(&seg
->s_as
->a_contents
);
9294 AS_SETNOUNMAPWAIT(seg
->s_as
);
9296 if (AS_ISUNMAPWAIT(seg
->s_as
)) {
9297 AS_CLRUNMAPWAIT(seg
->s_as
);
9298 cv_broadcast(&seg
->s_as
->a_cv
);
9300 mutex_exit(&seg
->s_as
->a_contents
);
9305 mutex_exit(&svd
->segfree_syncmtx
);
9312 shamp_reclaim(void *ptag
, caddr_t addr
, size_t len
, struct page
**pplist
,
9313 enum seg_rw rw
, int async
)
9315 amp_t
*amp
= (amp_t
*)ptag
;
9319 npages
= np
= btop(len
);
9321 ASSERT(amp
->a_softlockcnt
>= npages
);
9325 ASSERT(pl
[np
] == NOPCACHE_SHWLIST
|| pl
[np
] == PCACHE_SHWLIST
);
9326 ASSERT(!async
|| pl
[np
] == PCACHE_SHWLIST
);
9328 while (np
> (uint_t
)0) {
9329 if (rw
== S_WRITE
) {
9330 hat_setrefmod(*pplist
);
9332 hat_setref(*pplist
);
9334 page_unlock(*pplist
);
9339 kmem_free(pl
, sizeof (page_t
*) * (npages
+ 1));
9342 * If somebody sleeps in anonmap_purge() wake them up if a_softlockcnt
9343 * drops to 0. anon map can't be freed until a_softlockcnt drops to 0
9344 * and anonmap_purge() acquires a_purgemtx.
9346 mutex_enter(&
->a_purgemtx
);
9347 if (!atomic_add_long_nv((ulong_t
*)&
->a_softlockcnt
, -npages
) &&
9349 amp
->a_purgewait
= 0;
9350 cv_broadcast(&
->a_purgecv
);
9352 mutex_exit(&
->a_purgemtx
);
9357 * get a memory ID for an addr in a given segment
9359 * XXX only creates PAGESIZE pages if anon slots are not initialized.
9360 * At fault time they will be relocated into larger pages.
9363 segvn_getmemid(struct seg
*seg
, caddr_t addr
, memid_t
*memidp
)
9365 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
9366 struct anon
*ap
= NULL
;
9368 struct anon_map
*amp
;
9369 anon_sync_obj_t cookie
;
9371 if (svd
->type
== MAP_PRIVATE
) {
9372 memidp
->val
[0] = (uintptr_t)seg
->s_as
;
9373 memidp
->val
[1] = (uintptr_t)addr
;
9377 if (svd
->type
== MAP_SHARED
) {
9379 memidp
->val
[0] = (uintptr_t)svd
->vp
;
9380 memidp
->val
[1] = (u_longlong_t
)svd
->offset
+
9381 (uintptr_t)(addr
- seg
->s_base
);
9385 SEGVN_LOCK_ENTER(seg
->s_as
, &svd
->lock
, RW_READER
);
9386 if ((amp
= svd
->amp
) != NULL
) {
9387 anon_index
= svd
->anon_index
+
9388 seg_page(seg
, addr
);
9390 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
9392 ASSERT(amp
!= NULL
);
9394 ANON_LOCK_ENTER(&
->a_rwlock
, RW_READER
);
9395 anon_array_enter(amp
, anon_index
, &cookie
);
9396 ap
= anon_get_ptr(amp
->ahp
, anon_index
);
9400 pp
= anon_zero(seg
, addr
, &ap
, svd
->cred
);
9402 anon_array_exit(&cookie
);
9403 ANON_LOCK_EXIT(&
->a_rwlock
);
9406 ASSERT(anon_get_ptr(amp
->ahp
, anon_index
)
9408 (void) anon_set_ptr(amp
->ahp
, anon_index
,
9413 anon_array_exit(&cookie
);
9414 ANON_LOCK_EXIT(&
->a_rwlock
);
9416 memidp
->val
[0] = (uintptr_t)ap
;
9417 memidp
->val
[1] = (uintptr_t)addr
& PAGEOFFSET
;
9425 sameprot(struct seg
*seg
, caddr_t a
, size_t len
)
9427 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
9428 struct vpage
*vpage
;
9429 spgcnt_t pages
= btop(len
);
9432 if (svd
->pageprot
== 0)
9435 ASSERT(svd
->vpage
!= NULL
);
9437 vpage
= &svd
->vpage
[seg_page(seg
, a
)];
9438 prot
= VPP_PROT(vpage
);
9441 while (pages
-- > 0) {
9442 if (prot
!= VPP_PROT(vpage
))
9450 * Get memory allocation policy info for specified address in given segment
9452 static lgrp_mem_policy_info_t
*
9453 segvn_getpolicy(struct seg
*seg
, caddr_t addr
)
9455 struct anon_map
*amp
;
9457 lgrp_mem_policy_info_t
*policy_info
;
9458 struct segvn_data
*svn_data
;
9462 ASSERT(seg
!= NULL
);
9464 svn_data
= (struct segvn_data
*)seg
->s_data
;
9465 if (svn_data
== NULL
)
9469 * Get policy info for private or shared memory
9471 if (svn_data
->type
!= MAP_SHARED
) {
9472 if (svn_data
->tr_state
!= SEGVN_TR_ON
) {
9473 policy_info
= &svn_data
->policy_info
;
9475 policy_info
= &svn_data
->tr_policy_info
;
9476 ASSERT(policy_info
->mem_policy
==
9477 LGRP_MEM_POLICY_NEXT_SEG
);
9480 amp
= svn_data
->amp
;
9481 anon_index
= svn_data
->anon_index
+ seg_page(seg
, addr
);
9483 vn_off
= svn_data
->offset
+ (uintptr_t)(addr
- seg
->s_base
);
9484 policy_info
= lgrp_shm_policy_get(amp
, anon_index
, vp
, vn_off
);
9487 return (policy_info
);
9492 segvn_capable(struct seg
*seg
, segcapability_t capability
)
9498 * Bind text vnode segment to an amp. If we bind successfully mappings will be
9499 * established to per vnode mapping per lgroup amp pages instead of to vnode
9500 * pages. There's one amp per vnode text mapping per lgroup. Many processes
9501 * may share the same text replication amp. If a suitable amp doesn't already
9502 * exist in svntr hash table create a new one. We may fail to bind to amp if
9503 * segment is not eligible for text replication. Code below first checks for
9504 * these conditions. If binding is successful segment tr_state is set to on
9505 * and svd->amp points to the amp to use. Otherwise tr_state is set to off and
9506 * svd->amp remains as NULL.
9509 segvn_textrepl(struct seg
*seg
)
9511 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
9512 vnode_t
*vp
= svd
->vp
;
9513 u_offset_t off
= svd
->offset
;
9514 size_t size
= seg
->s_size
;
9515 u_offset_t eoff
= off
+ size
;
9516 uint_t szc
= seg
->s_szc
;
9517 ulong_t hash
= SVNTR_HASH_FUNC(vp
);
9520 proc_t
*p
= seg
->s_as
->a_proc
;
9524 struct anon_map
*amp
;
9526 ASSERT(AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
9527 ASSERT(SEGVN_WRITE_HELD(seg
->s_as
, &svd
->lock
));
9529 ASSERT(svd
->tr_state
== SEGVN_TR_INIT
);
9530 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
));
9531 ASSERT(svd
->flags
& MAP_TEXT
);
9532 ASSERT(svd
->type
== MAP_PRIVATE
);
9533 ASSERT(vp
!= NULL
&& svd
->amp
== NULL
);
9534 ASSERT(!svd
->pageprot
&& !(svd
->prot
& PROT_WRITE
));
9535 ASSERT(!(svd
->flags
& MAP_NORESERVE
) && svd
->swresv
== 0);
9536 ASSERT(seg
->s_as
!= &kas
);
9538 ASSERT(svntr_hashtab
!= NULL
);
9541 * If numa optimizations are no longer desired bail out.
9543 if (!lgrp_optimizations()) {
9544 svd
->tr_state
= SEGVN_TR_OFF
;
9549 * Avoid creating anon maps with size bigger than the file size.
9550 * If VOP_GETATTR() call fails bail out.
9552 va
.va_mask
= AT_SIZE
| AT_MTIME
| AT_CTIME
;
9553 if (VOP_GETATTR(vp
, &va
, 0, svd
->cred
, NULL
) != 0) {
9554 svd
->tr_state
= SEGVN_TR_OFF
;
9555 SEGVN_TR_ADDSTAT(gaerr
);
9558 if (btopr(va
.va_size
) < btopr(eoff
)) {
9559 svd
->tr_state
= SEGVN_TR_OFF
;
9560 SEGVN_TR_ADDSTAT(overmap
);
9565 * VVMEXEC may not be set yet if exec() prefaults text segment. Set
9566 * this flag now before vn_is_mapped(V_WRITE) so that MAP_SHARED
9567 * mapping that checks if trcache for this vnode needs to be
9568 * invalidated can't miss us.
9570 if (!(vp
->v_flag
& VVMEXEC
)) {
9571 mutex_enter(&vp
->v_lock
);
9572 vp
->v_flag
|= VVMEXEC
;
9573 mutex_exit(&vp
->v_lock
);
9575 mutex_enter(&svntr_hashtab
[hash
].tr_lock
);
9577 * Bail out if potentially MAP_SHARED writable mappings exist to this
9578 * vnode. We don't want to use old file contents from existing
9579 * replicas if this mapping was established after the original file
9582 if (vn_is_mapped(vp
, V_WRITE
)) {
9583 mutex_exit(&svntr_hashtab
[hash
].tr_lock
);
9584 svd
->tr_state
= SEGVN_TR_OFF
;
9585 SEGVN_TR_ADDSTAT(wrcnt
);
9588 svntrp
= svntr_hashtab
[hash
].tr_head
;
9589 for (; svntrp
!= NULL
; svntrp
= svntrp
->tr_next
) {
9590 ASSERT(svntrp
->tr_refcnt
!= 0);
9591 if (svntrp
->tr_vp
!= vp
) {
9596 * Bail out if the file or its attributes were changed after
9597 * this replication entry was created since we need to use the
9598 * latest file contents. Note that mtime test alone is not
9599 * sufficient because a user can explicitly change mtime via
9600 * utimes(2) interfaces back to the old value after modifiying
9601 * the file contents. To detect this case we also have to test
9602 * ctime which among other things records the time of the last
9603 * mtime change by utimes(2). ctime is not changed when the file
9604 * is only read or executed so we expect that typically existing
9605 * replication amp's can be used most of the time.
9607 if (!svntrp
->tr_valid
||
9608 svntrp
->tr_mtime
.tv_sec
!= va
.va_mtime
.tv_sec
||
9609 svntrp
->tr_mtime
.tv_nsec
!= va
.va_mtime
.tv_nsec
||
9610 svntrp
->tr_ctime
.tv_sec
!= va
.va_ctime
.tv_sec
||
9611 svntrp
->tr_ctime
.tv_nsec
!= va
.va_ctime
.tv_nsec
) {
9612 mutex_exit(&svntr_hashtab
[hash
].tr_lock
);
9613 svd
->tr_state
= SEGVN_TR_OFF
;
9614 SEGVN_TR_ADDSTAT(stale
);
9618 * if off, eoff and szc match current segment we found the
9619 * existing entry we can use.
9621 if (svntrp
->tr_off
== off
&& svntrp
->tr_eoff
== eoff
&&
9622 svntrp
->tr_szc
== szc
) {
9626 * Don't create different but overlapping in file offsets
9627 * entries to avoid replication of the same file pages more
9628 * than once per lgroup.
9630 if ((off
>= svntrp
->tr_off
&& off
< svntrp
->tr_eoff
) ||
9631 (eoff
> svntrp
->tr_off
&& eoff
<= svntrp
->tr_eoff
)) {
9632 mutex_exit(&svntr_hashtab
[hash
].tr_lock
);
9633 svd
->tr_state
= SEGVN_TR_OFF
;
9634 SEGVN_TR_ADDSTAT(overlap
);
9639 * If we didn't find existing entry create a new one.
9641 if (svntrp
== NULL
) {
9642 svntrp
= kmem_cache_alloc(svntr_cache
, KM_NOSLEEP
);
9643 if (svntrp
== NULL
) {
9644 mutex_exit(&svntr_hashtab
[hash
].tr_lock
);
9645 svd
->tr_state
= SEGVN_TR_OFF
;
9646 SEGVN_TR_ADDSTAT(nokmem
);
9652 for (i
= 0; i
< NLGRPS_MAX
; i
++) {
9653 ASSERT(svntrp
->tr_amp
[i
] == NULL
);
9658 svntrp
->tr_off
= off
;
9659 svntrp
->tr_eoff
= eoff
;
9660 svntrp
->tr_szc
= szc
;
9661 svntrp
->tr_valid
= 1;
9662 svntrp
->tr_mtime
= va
.va_mtime
;
9663 svntrp
->tr_ctime
= va
.va_ctime
;
9664 svntrp
->tr_refcnt
= 0;
9665 svntrp
->tr_next
= svntr_hashtab
[hash
].tr_head
;
9666 svntr_hashtab
[hash
].tr_head
= svntrp
;
9671 * We want to pick a replica with pages on main thread's (t_tid = 1,
9672 * aka T1) lgrp. Currently text replication is only optimized for
9673 * workloads that either have all threads of a process on the same
9674 * lgrp or execute their large text primarily on main thread.
9676 lgrp_id
= p
->p_t1_lgrpid
;
9677 if (lgrp_id
== LGRP_NONE
) {
9679 * In case exec() prefaults text on non main thread use
9680 * current thread lgrpid. It will become main thread anyway
9683 lgrp_id
= lgrp_home_id(curthread
);
9686 * Set p_tr_lgrpid to lgrpid if it hasn't been set yet. Otherwise
9687 * just set it to NLGRPS_MAX if it's different from current process T1
9688 * home lgrp. p_tr_lgrpid is used to detect if process uses text
9689 * replication and T1 new home is different from lgrp used for text
9690 * replication. When this happens asyncronous segvn thread rechecks if
9691 * segments should change lgrps used for text replication. If we fail
9692 * to set p_tr_lgrpid with cas32 then set it to NLGRPS_MAX without cas
9693 * if it's not already NLGRPS_MAX and not equal lgrp_id we want to
9694 * use. We don't need to use cas in this case because another thread
9695 * that races in between our non atomic check and set may only change
9696 * p_tr_lgrpid to NLGRPS_MAX at this point.
9698 ASSERT(lgrp_id
!= LGRP_NONE
&& lgrp_id
< NLGRPS_MAX
);
9699 olid
= p
->p_tr_lgrpid
;
9700 if (lgrp_id
!= olid
&& olid
!= NLGRPS_MAX
) {
9701 lgrp_id_t nlid
= (olid
== LGRP_NONE
) ? lgrp_id
: NLGRPS_MAX
;
9702 if (cas32((uint32_t *)&p
->p_tr_lgrpid
, olid
, nlid
) != olid
) {
9703 olid
= p
->p_tr_lgrpid
;
9704 ASSERT(olid
!= LGRP_NONE
);
9705 if (olid
!= lgrp_id
&& olid
!= NLGRPS_MAX
) {
9706 p
->p_tr_lgrpid
= NLGRPS_MAX
;
9709 ASSERT(p
->p_tr_lgrpid
!= LGRP_NONE
);
9712 * lgrp_move_thread() won't schedule async recheck after
9713 * p->p_t1_lgrpid update unless p->p_tr_lgrpid is not
9714 * LGRP_NONE. Recheck p_t1_lgrpid once now that p->p_tr_lgrpid
9717 if (first
&& p
->p_t1_lgrpid
!= LGRP_NONE
&&
9718 p
->p_t1_lgrpid
!= lgrp_id
) {
9724 * If no amp was created yet for lgrp_id create a new one as long as
9725 * we have enough memory to afford it.
9727 if ((amp
= svntrp
->tr_amp
[lgrp_id
]) == NULL
) {
9728 size_t trmem
= atomic_add_long_nv(&segvn_textrepl_bytes
, size
);
9729 if (trmem
> segvn_textrepl_max_bytes
) {
9730 SEGVN_TR_ADDSTAT(normem
);
9733 if (anon_try_resv_zone(size
, NULL
) == 0) {
9734 SEGVN_TR_ADDSTAT(noanon
);
9737 amp
= anonmap_alloc(size
, size
, ANON_NOSLEEP
);
9739 anon_unresv_zone(size
, NULL
);
9740 SEGVN_TR_ADDSTAT(nokmem
);
9743 ASSERT(amp
->refcnt
== 1);
9745 svntrp
->tr_amp
[lgrp_id
] = amp
;
9746 SEGVN_TR_ADDSTAT(newamp
);
9748 svntrp
->tr_refcnt
++;
9749 ASSERT(svd
->svn_trnext
== NULL
);
9750 ASSERT(svd
->svn_trprev
== NULL
);
9751 svd
->svn_trnext
= svntrp
->tr_svnhead
;
9752 svd
->svn_trprev
= NULL
;
9753 if (svntrp
->tr_svnhead
!= NULL
) {
9754 svntrp
->tr_svnhead
->svn_trprev
= svd
;
9756 svntrp
->tr_svnhead
= svd
;
9757 ASSERT(amp
->a_szc
== szc
&& amp
->size
== size
&& amp
->swresv
== size
);
9758 ASSERT(amp
->refcnt
>= 1);
9760 svd
->anon_index
= 0;
9761 svd
->tr_policy_info
.mem_policy
= LGRP_MEM_POLICY_NEXT_SEG
;
9762 svd
->tr_policy_info
.mem_lgrpid
= lgrp_id
;
9763 svd
->tr_state
= SEGVN_TR_ON
;
9764 mutex_exit(&svntr_hashtab
[hash
].tr_lock
);
9765 SEGVN_TR_ADDSTAT(repl
);
9768 ASSERT(segvn_textrepl_bytes
>= size
);
9769 atomic_add_long(&segvn_textrepl_bytes
, -size
);
9770 ASSERT(svntrp
!= NULL
);
9771 ASSERT(svntrp
->tr_amp
[lgrp_id
] == NULL
);
9772 if (svntrp
->tr_refcnt
== 0) {
9773 ASSERT(svntrp
== svntr_hashtab
[hash
].tr_head
);
9774 svntr_hashtab
[hash
].tr_head
= svntrp
->tr_next
;
9775 mutex_exit(&svntr_hashtab
[hash
].tr_lock
);
9776 kmem_cache_free(svntr_cache
, svntrp
);
9778 mutex_exit(&svntr_hashtab
[hash
].tr_lock
);
9780 svd
->tr_state
= SEGVN_TR_OFF
;
9784 * Convert seg back to regular vnode mapping seg by unbinding it from its text
9785 * replication amp. This routine is most typically called when segment is
9786 * unmapped but can also be called when segment no longer qualifies for text
9787 * replication (e.g. due to protection changes). If unload_unmap is set use
9788 * HAT_UNLOAD_UNMAP flag in hat_unload_callback(). If we are the last user of
9789 * svntr free all its anon maps and remove it from the hash table.
9792 segvn_textunrepl(struct seg
*seg
, int unload_unmap
)
9794 struct segvn_data
*svd
= (struct segvn_data
*)seg
->s_data
;
9795 vnode_t
*vp
= svd
->vp
;
9796 u_offset_t off
= svd
->offset
;
9797 size_t size
= seg
->s_size
;
9798 u_offset_t eoff
= off
+ size
;
9799 uint_t szc
= seg
->s_szc
;
9800 ulong_t hash
= SVNTR_HASH_FUNC(vp
);
9802 svntr_t
**prv_svntrp
;
9803 lgrp_id_t lgrp_id
= svd
->tr_policy_info
.mem_lgrpid
;
9806 ASSERT(AS_LOCK_HELD(seg
->s_as
, &seg
->s_as
->a_lock
));
9807 ASSERT(AS_WRITE_HELD(seg
->s_as
, &seg
->s_as
->a_lock
) ||
9808 SEGVN_WRITE_HELD(seg
->s_as
, &svd
->lock
));
9809 ASSERT(svd
->tr_state
== SEGVN_TR_ON
);
9810 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
));
9811 ASSERT(svd
->amp
!= NULL
);
9812 ASSERT(svd
->amp
->refcnt
>= 1);
9813 ASSERT(svd
->anon_index
== 0);
9814 ASSERT(lgrp_id
!= LGRP_NONE
&& lgrp_id
< NLGRPS_MAX
);
9815 ASSERT(svntr_hashtab
!= NULL
);
9817 mutex_enter(&svntr_hashtab
[hash
].tr_lock
);
9818 prv_svntrp
= &svntr_hashtab
[hash
].tr_head
;
9819 for (; (svntrp
= *prv_svntrp
) != NULL
; prv_svntrp
= &svntrp
->tr_next
) {
9820 ASSERT(svntrp
->tr_refcnt
!= 0);
9821 if (svntrp
->tr_vp
== vp
&& svntrp
->tr_off
== off
&&
9822 svntrp
->tr_eoff
== eoff
&& svntrp
->tr_szc
== szc
) {
9826 if (svntrp
== NULL
) {
9827 panic("segvn_textunrepl: svntr record not found");
9829 if (svntrp
->tr_amp
[lgrp_id
] != svd
->amp
) {
9830 panic("segvn_textunrepl: amp mismatch");
9832 svd
->tr_state
= SEGVN_TR_OFF
;
9834 if (svd
->svn_trprev
== NULL
) {
9835 ASSERT(svntrp
->tr_svnhead
== svd
);
9836 svntrp
->tr_svnhead
= svd
->svn_trnext
;
9837 if (svntrp
->tr_svnhead
!= NULL
) {
9838 svntrp
->tr_svnhead
->svn_trprev
= NULL
;
9840 svd
->svn_trnext
= NULL
;
9842 svd
->svn_trprev
->svn_trnext
= svd
->svn_trnext
;
9843 if (svd
->svn_trnext
!= NULL
) {
9844 svd
->svn_trnext
->svn_trprev
= svd
->svn_trprev
;
9845 svd
->svn_trnext
= NULL
;
9847 svd
->svn_trprev
= NULL
;
9849 if (--svntrp
->tr_refcnt
) {
9850 mutex_exit(&svntr_hashtab
[hash
].tr_lock
);
9853 *prv_svntrp
= svntrp
->tr_next
;
9854 mutex_exit(&svntr_hashtab
[hash
].tr_lock
);
9855 for (i
= 0; i
< NLGRPS_MAX
; i
++) {
9856 struct anon_map
*amp
= svntrp
->tr_amp
[i
];
9860 ASSERT(amp
->refcnt
== 1);
9861 ASSERT(amp
->swresv
== size
);
9862 ASSERT(amp
->size
== size
);
9863 ASSERT(amp
->a_szc
== szc
);
9864 if (amp
->a_szc
!= 0) {
9865 anon_free_pages(amp
->ahp
, 0, size
, szc
);
9867 anon_free(amp
->ahp
, 0, size
);
9869 svntrp
->tr_amp
[i
] = NULL
;
9870 ASSERT(segvn_textrepl_bytes
>= size
);
9871 atomic_add_long(&segvn_textrepl_bytes
, -size
);
9872 anon_unresv_zone(amp
->swresv
, NULL
);
9876 kmem_cache_free(svntr_cache
, svntrp
);
9878 hat_unload_callback(seg
->s_as
->a_hat
, seg
->s_base
, size
,
9879 unload_unmap
? HAT_UNLOAD_UNMAP
: 0, NULL
);
9883 * This is called when a MAP_SHARED writable mapping is created to a vnode
9884 * that is currently used for execution (VVMEXEC flag is set). In this case we
9885 * need to prevent further use of existing replicas.
9888 segvn_inval_trcache(vnode_t
*vp
)
9890 ulong_t hash
= SVNTR_HASH_FUNC(vp
);
9893 ASSERT(vp
->v_flag
& VVMEXEC
);
9895 if (svntr_hashtab
== NULL
) {
9899 mutex_enter(&svntr_hashtab
[hash
].tr_lock
);
9900 svntrp
= svntr_hashtab
[hash
].tr_head
;
9901 for (; svntrp
!= NULL
; svntrp
= svntrp
->tr_next
) {
9902 ASSERT(svntrp
->tr_refcnt
!= 0);
9903 if (svntrp
->tr_vp
== vp
&& svntrp
->tr_valid
) {
9904 svntrp
->tr_valid
= 0;
9907 mutex_exit(&svntr_hashtab
[hash
].tr_lock
);
9911 segvn_trasync_thread(void)
9913 callb_cpr_t cpr_info
;
9914 kmutex_t cpr_lock
; /* just for CPR stuff */
9916 mutex_init(&cpr_lock
, NULL
, MUTEX_DEFAULT
, NULL
);
9918 CALLB_CPR_INIT(&cpr_info
, &cpr_lock
,
9919 callb_generic_cpr
, "segvn_async");
9921 if (segvn_update_textrepl_interval
== 0) {
9922 segvn_update_textrepl_interval
= segvn_update_tr_time
* hz
;
9924 segvn_update_textrepl_interval
*= hz
;
9926 (void) timeout(segvn_trupdate_wakeup
, NULL
,
9927 segvn_update_textrepl_interval
);
9930 mutex_enter(&cpr_lock
);
9931 CALLB_CPR_SAFE_BEGIN(&cpr_info
);
9932 mutex_exit(&cpr_lock
);
9933 sema_p(&segvn_trasync_sem
);
9934 mutex_enter(&cpr_lock
);
9935 CALLB_CPR_SAFE_END(&cpr_info
, &cpr_lock
);
9936 mutex_exit(&cpr_lock
);
9941 static uint64_t segvn_lgrp_trthr_migrs_snpsht
= 0;
9944 segvn_trupdate_wakeup(void *dummy
)
9946 uint64_t cur_lgrp_trthr_migrs
= lgrp_get_trthr_migrations();
9948 if (cur_lgrp_trthr_migrs
!= segvn_lgrp_trthr_migrs_snpsht
) {
9949 segvn_lgrp_trthr_migrs_snpsht
= cur_lgrp_trthr_migrs
;
9950 sema_v(&segvn_trasync_sem
);
9953 if (!segvn_disable_textrepl_update
&&
9954 segvn_update_textrepl_interval
!= 0) {
9955 (void) timeout(segvn_trupdate_wakeup
, dummy
,
9956 segvn_update_textrepl_interval
);
9961 segvn_trupdate(void)
9967 ASSERT(svntr_hashtab
!= NULL
);
9969 for (hash
= 0; hash
< svntr_hashtab_sz
; hash
++) {
9970 mutex_enter(&svntr_hashtab
[hash
].tr_lock
);
9971 svntrp
= svntr_hashtab
[hash
].tr_head
;
9972 for (; svntrp
!= NULL
; svntrp
= svntrp
->tr_next
) {
9973 ASSERT(svntrp
->tr_refcnt
!= 0);
9974 svd
= svntrp
->tr_svnhead
;
9975 for (; svd
!= NULL
; svd
= svd
->svn_trnext
) {
9976 segvn_trupdate_seg(svd
->seg
, svd
, svntrp
,
9980 mutex_exit(&svntr_hashtab
[hash
].tr_lock
);
9985 segvn_trupdate_seg(struct seg
*seg
,
9994 struct anon_map
*amp
;
9996 ASSERT(svd
->vp
!= NULL
);
9997 ASSERT(svd
->vp
== svntrp
->tr_vp
);
9998 ASSERT(svd
->offset
== svntrp
->tr_off
);
9999 ASSERT(svd
->offset
+ seg
->s_size
== svntrp
->tr_eoff
);
10000 ASSERT(seg
!= NULL
);
10001 ASSERT(svd
->seg
== seg
);
10002 ASSERT(seg
->s_data
== (void *)svd
);
10003 ASSERT(seg
->s_szc
== svntrp
->tr_szc
);
10004 ASSERT(svd
->tr_state
== SEGVN_TR_ON
);
10005 ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd
->rcookie
));
10006 ASSERT(svd
->amp
!= NULL
);
10007 ASSERT(svd
->tr_policy_info
.mem_policy
== LGRP_MEM_POLICY_NEXT_SEG
);
10008 ASSERT(svd
->tr_policy_info
.mem_lgrpid
!= LGRP_NONE
);
10009 ASSERT(svd
->tr_policy_info
.mem_lgrpid
< NLGRPS_MAX
);
10010 ASSERT(svntrp
->tr_amp
[svd
->tr_policy_info
.mem_lgrpid
] == svd
->amp
);
10011 ASSERT(svntrp
->tr_refcnt
!= 0);
10012 ASSERT(mutex_owned(&svntr_hashtab
[hash
].tr_lock
));
10015 ASSERT(as
!= NULL
&& as
!= &kas
);
10018 ASSERT(p
->p_tr_lgrpid
!= LGRP_NONE
);
10019 lgrp_id
= p
->p_t1_lgrpid
;
10020 if (lgrp_id
== LGRP_NONE
) {
10023 ASSERT(lgrp_id
< NLGRPS_MAX
);
10024 if (svd
->tr_policy_info
.mem_lgrpid
== lgrp_id
) {
10029 * Use tryenter locking since we are locking as/seg and svntr hash
10030 * lock in reverse from syncrounous thread order.
10032 if (!AS_LOCK_TRYENTER(as
, &as
->a_lock
, RW_READER
)) {
10033 SEGVN_TR_ADDSTAT(nolock
);
10034 if (segvn_lgrp_trthr_migrs_snpsht
) {
10035 segvn_lgrp_trthr_migrs_snpsht
= 0;
10039 if (!SEGVN_LOCK_TRYENTER(seg
->s_as
, &svd
->lock
, RW_WRITER
)) {
10040 AS_LOCK_EXIT(as
, &as
->a_lock
);
10041 SEGVN_TR_ADDSTAT(nolock
);
10042 if (segvn_lgrp_trthr_migrs_snpsht
) {
10043 segvn_lgrp_trthr_migrs_snpsht
= 0;
10047 size
= seg
->s_size
;
10048 if (svntrp
->tr_amp
[lgrp_id
] == NULL
) {
10049 size_t trmem
= atomic_add_long_nv(&segvn_textrepl_bytes
, size
);
10050 if (trmem
> segvn_textrepl_max_bytes
) {
10051 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
10052 AS_LOCK_EXIT(as
, &as
->a_lock
);
10053 atomic_add_long(&segvn_textrepl_bytes
, -size
);
10054 SEGVN_TR_ADDSTAT(normem
);
10057 if (anon_try_resv_zone(size
, NULL
) == 0) {
10058 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
10059 AS_LOCK_EXIT(as
, &as
->a_lock
);
10060 atomic_add_long(&segvn_textrepl_bytes
, -size
);
10061 SEGVN_TR_ADDSTAT(noanon
);
10064 amp
= anonmap_alloc(size
, size
, KM_NOSLEEP
);
10066 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
10067 AS_LOCK_EXIT(as
, &as
->a_lock
);
10068 atomic_add_long(&segvn_textrepl_bytes
, -size
);
10069 anon_unresv_zone(size
, NULL
);
10070 SEGVN_TR_ADDSTAT(nokmem
);
10073 ASSERT(amp
->refcnt
== 1);
10074 amp
->a_szc
= seg
->s_szc
;
10075 svntrp
->tr_amp
[lgrp_id
] = amp
;
10078 * We don't need to drop the bucket lock but here we give other
10079 * threads a chance. svntr and svd can't be unlinked as long as
10080 * segment lock is held as a writer and AS held as well. After we
10081 * retake bucket lock we'll continue from where we left. We'll be able
10082 * to reach the end of either list since new entries are always added
10083 * to the beginning of the lists.
10085 mutex_exit(&svntr_hashtab
[hash
].tr_lock
);
10086 hat_unload_callback(as
->a_hat
, seg
->s_base
, size
, 0, NULL
);
10087 mutex_enter(&svntr_hashtab
[hash
].tr_lock
);
10089 ASSERT(svd
->tr_state
== SEGVN_TR_ON
);
10090 ASSERT(svd
->amp
!= NULL
);
10091 ASSERT(svd
->tr_policy_info
.mem_policy
== LGRP_MEM_POLICY_NEXT_SEG
);
10092 ASSERT(svd
->tr_policy_info
.mem_lgrpid
!= lgrp_id
);
10093 ASSERT(svd
->amp
!= svntrp
->tr_amp
[lgrp_id
]);
10095 svd
->tr_policy_info
.mem_lgrpid
= lgrp_id
;
10096 svd
->amp
= svntrp
->tr_amp
[lgrp_id
];
10097 p
->p_tr_lgrpid
= NLGRPS_MAX
;
10098 SEGVN_LOCK_EXIT(seg
->s_as
, &svd
->lock
);
10099 AS_LOCK_EXIT(as
, &as
->a_lock
);
10101 ASSERT(svntrp
->tr_refcnt
!= 0);
10102 ASSERT(svd
->vp
== svntrp
->tr_vp
);
10103 ASSERT(svd
->tr_policy_info
.mem_lgrpid
== lgrp_id
);
10104 ASSERT(svd
->amp
!= NULL
&& svd
->amp
== svntrp
->tr_amp
[lgrp_id
]);
10105 ASSERT(svd
->seg
== seg
);
10106 ASSERT(svd
->tr_state
== SEGVN_TR_ON
);
10108 SEGVN_TR_ADDSTAT(asyncrepl
);