xref: /titanic_41/usr/src/uts/common/vm/seg_vn.c (revision 7800901e60d340b6af88e94a2149805dcfcaaf56)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
27 /*	  All Rights Reserved  	*/
28 
29 /*
30  * University Copyright- Copyright (c) 1982, 1986, 1988
31  * The Regents of the University of California
32  * All Rights Reserved
33  *
34  * University Acknowledgment- Portions of this document are derived from
35  * software developed by the University of California, Berkeley, and its
36  * contributors.
37  */
38 
39 #pragma ident	"%Z%%M%	%I%	%E% SMI"
40 
41 /*
42  * VM - shared or copy-on-write from a vnode/anonymous memory.
43  */
44 
45 #include <sys/types.h>
46 #include <sys/param.h>
47 #include <sys/t_lock.h>
48 #include <sys/errno.h>
49 #include <sys/systm.h>
50 #include <sys/mman.h>
51 #include <sys/debug.h>
52 #include <sys/cred.h>
53 #include <sys/vmsystm.h>
54 #include <sys/tuneable.h>
55 #include <sys/bitmap.h>
56 #include <sys/swap.h>
57 #include <sys/kmem.h>
58 #include <sys/sysmacros.h>
59 #include <sys/vtrace.h>
60 #include <sys/cmn_err.h>
61 #include <sys/callb.h>
62 #include <sys/vm.h>
63 #include <sys/dumphdr.h>
64 #include <sys/lgrp.h>
65 
66 #include <vm/hat.h>
67 #include <vm/as.h>
68 #include <vm/seg.h>
69 #include <vm/seg_vn.h>
70 #include <vm/pvn.h>
71 #include <vm/anon.h>
72 #include <vm/page.h>
73 #include <vm/vpage.h>
74 #include <sys/proc.h>
75 #include <sys/task.h>
76 #include <sys/project.h>
77 #include <sys/zone.h>
78 #include <sys/shm_impl.h>
79 /*
80  * Private seg op routines.
81  */
82 static int	segvn_dup(struct seg *seg, struct seg *newseg);
83 static int	segvn_unmap(struct seg *seg, caddr_t addr, size_t len);
84 static void	segvn_free(struct seg *seg);
85 static faultcode_t segvn_fault(struct hat *hat, struct seg *seg,
86 		    caddr_t addr, size_t len, enum fault_type type,
87 		    enum seg_rw rw);
88 static faultcode_t segvn_faulta(struct seg *seg, caddr_t addr);
89 static int	segvn_setprot(struct seg *seg, caddr_t addr,
90 		    size_t len, uint_t prot);
91 static int	segvn_checkprot(struct seg *seg, caddr_t addr,
92 		    size_t len, uint_t prot);
93 static int	segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta);
94 static size_t	segvn_swapout(struct seg *seg);
95 static int	segvn_sync(struct seg *seg, caddr_t addr, size_t len,
96 		    int attr, uint_t flags);
97 static size_t	segvn_incore(struct seg *seg, caddr_t addr, size_t len,
98 		    char *vec);
99 static int	segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
100 		    int attr, int op, ulong_t *lockmap, size_t pos);
101 static int	segvn_getprot(struct seg *seg, caddr_t addr, size_t len,
102 		    uint_t *protv);
103 static u_offset_t	segvn_getoffset(struct seg *seg, caddr_t addr);
104 static int	segvn_gettype(struct seg *seg, caddr_t addr);
105 static int	segvn_getvp(struct seg *seg, caddr_t addr,
106 		    struct vnode **vpp);
107 static int	segvn_advise(struct seg *seg, caddr_t addr, size_t len,
108 		    uint_t behav);
109 static void	segvn_dump(struct seg *seg);
110 static int	segvn_pagelock(struct seg *seg, caddr_t addr, size_t len,
111 		    struct page ***ppp, enum lock_type type, enum seg_rw rw);
112 static int	segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len,
113 		    uint_t szc);
114 static int	segvn_getmemid(struct seg *seg, caddr_t addr,
115 		    memid_t *memidp);
116 static lgrp_mem_policy_info_t	*segvn_getpolicy(struct seg *, caddr_t);
117 static int	segvn_capable(struct seg *seg, segcapability_t capable);
118 
119 struct	seg_ops segvn_ops = {
120 	segvn_dup,
121 	segvn_unmap,
122 	segvn_free,
123 	segvn_fault,
124 	segvn_faulta,
125 	segvn_setprot,
126 	segvn_checkprot,
127 	segvn_kluster,
128 	segvn_swapout,
129 	segvn_sync,
130 	segvn_incore,
131 	segvn_lockop,
132 	segvn_getprot,
133 	segvn_getoffset,
134 	segvn_gettype,
135 	segvn_getvp,
136 	segvn_advise,
137 	segvn_dump,
138 	segvn_pagelock,
139 	segvn_setpagesize,
140 	segvn_getmemid,
141 	segvn_getpolicy,
142 	segvn_capable,
143 };
144 
145 /*
146  * Common zfod structures, provided as a shorthand for others to use.
147  */
148 static segvn_crargs_t zfod_segvn_crargs =
149 	SEGVN_ZFOD_ARGS(PROT_ZFOD, PROT_ALL);
150 static segvn_crargs_t kzfod_segvn_crargs =
151 	SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_USER,
152 	PROT_ALL & ~PROT_USER);
153 static segvn_crargs_t stack_noexec_crargs =
154 	SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_EXEC, PROT_ALL);
155 
156 caddr_t	zfod_argsp = (caddr_t)&zfod_segvn_crargs;	/* user zfod argsp */
157 caddr_t	kzfod_argsp = (caddr_t)&kzfod_segvn_crargs;	/* kernel zfod argsp */
158 caddr_t	stack_exec_argsp = (caddr_t)&zfod_segvn_crargs;	/* executable stack */
159 caddr_t	stack_noexec_argsp = (caddr_t)&stack_noexec_crargs; /* noexec stack */
160 
161 #define	vpgtob(n)	((n) * sizeof (struct vpage))	/* For brevity */
162 
163 size_t	segvn_comb_thrshld = UINT_MAX;	/* patchable -- see 1196681 */
164 
165 static int	segvn_concat(struct seg *, struct seg *, int);
166 static int	segvn_extend_prev(struct seg *, struct seg *,
167 		    struct segvn_crargs *, size_t);
168 static int	segvn_extend_next(struct seg *, struct seg *,
169 		    struct segvn_crargs *, size_t);
170 static void	segvn_softunlock(struct seg *, caddr_t, size_t, enum seg_rw);
171 static void	segvn_pagelist_rele(page_t **);
172 static void	segvn_setvnode_mpss(vnode_t *);
173 static void	segvn_relocate_pages(page_t **, page_t *);
174 static int	segvn_full_szcpages(page_t **, uint_t, int *, uint_t *);
175 static int	segvn_fill_vp_pages(struct segvn_data *, vnode_t *, u_offset_t,
176     uint_t, page_t **, page_t **, uint_t *, int *);
177 static faultcode_t segvn_fault_vnodepages(struct hat *, struct seg *, caddr_t,
178     caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
179 static faultcode_t segvn_fault_anonpages(struct hat *, struct seg *, caddr_t,
180     caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
181 static faultcode_t segvn_faultpage(struct hat *, struct seg *, caddr_t,
182     u_offset_t, struct vpage *, page_t **, uint_t,
183     enum fault_type, enum seg_rw, int, int);
184 static void	segvn_vpage(struct seg *);
185 
186 static void segvn_purge(struct seg *seg);
187 static int segvn_reclaim(struct seg *, caddr_t, size_t, struct page **,
188     enum seg_rw);
189 
190 static int sameprot(struct seg *, caddr_t, size_t);
191 
192 static int segvn_demote_range(struct seg *, caddr_t, size_t, int, uint_t);
193 static int segvn_clrszc(struct seg *);
194 static struct seg *segvn_split_seg(struct seg *, caddr_t);
195 static int segvn_claim_pages(struct seg *, struct vpage *, u_offset_t,
196     ulong_t, uint_t);
197 
198 static void segvn_hat_rgn_unload_callback(caddr_t, caddr_t, caddr_t,
199     size_t, void *, u_offset_t);
200 
201 static int segvn_slock_anonpages(page_t *, int);
202 static void segvn_sunlock_anonpages(page_t *, int);
203 
204 static struct kmem_cache *segvn_cache;
205 
206 #ifdef VM_STATS
207 static struct segvnvmstats_str {
208 	ulong_t	fill_vp_pages[31];
209 	ulong_t fltvnpages[49];
210 	ulong_t	fullszcpages[10];
211 	ulong_t	relocatepages[3];
212 	ulong_t	fltanpages[17];
213 	ulong_t pagelock[3];
214 	ulong_t	demoterange[3];
215 } segvnvmstats;
216 #endif /* VM_STATS */
217 
218 #define	SDR_RANGE	1		/* demote entire range */
219 #define	SDR_END		2		/* demote non aligned ends only */
220 
221 #define	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr) {	    \
222 		if ((len) != 0) { 		      	      		      \
223 			lpgaddr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz);  \
224 			ASSERT(lpgaddr >= (seg)->s_base);	      	      \
225 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)((addr) +    \
226 			    (len)), pgsz);				      \
227 			ASSERT(lpgeaddr > lpgaddr);		      	      \
228 			ASSERT(lpgeaddr <= (seg)->s_base + (seg)->s_size);    \
229 		} else {					      	      \
230 			lpgeaddr = lpgaddr = (addr);	      		      \
231 		}							      \
232 	}
233 
234 /*ARGSUSED*/
235 static int
236 segvn_cache_constructor(void *buf, void *cdrarg, int kmflags)
237 {
238 	struct segvn_data *svd = buf;
239 
240 	rw_init(&svd->lock, NULL, RW_DEFAULT, NULL);
241 	mutex_init(&svd->segp_slock, NULL, MUTEX_DEFAULT, NULL);
242 	svd->svn_trnext = svd->svn_trprev = NULL;
243 	return (0);
244 }
245 
246 /*ARGSUSED1*/
247 static void
248 segvn_cache_destructor(void *buf, void *cdrarg)
249 {
250 	struct segvn_data *svd = buf;
251 
252 	rw_destroy(&svd->lock);
253 	mutex_destroy(&svd->segp_slock);
254 }
255 
256 /*ARGSUSED*/
257 static int
258 svntr_cache_constructor(void *buf, void *cdrarg, int kmflags)
259 {
260 	bzero(buf, sizeof (svntr_t));
261 	return (0);
262 }
263 
264 /*
265  * Patching this variable to non-zero allows the system to run with
266  * stacks marked as "not executable".  It's a bit of a kludge, but is
267  * provided as a tweakable for platforms that export those ABIs
268  * (e.g. sparc V8) that have executable stacks enabled by default.
269  * There are also some restrictions for platforms that don't actually
270  * implement 'noexec' protections.
271  *
272  * Once enabled, the system is (therefore) unable to provide a fully
273  * ABI-compliant execution environment, though practically speaking,
274  * most everything works.  The exceptions are generally some interpreters
275  * and debuggers that create executable code on the stack and jump
276  * into it (without explicitly mprotecting the address range to include
277  * PROT_EXEC).
278  *
279  * One important class of applications that are disabled are those
280  * that have been transformed into malicious agents using one of the
281  * numerous "buffer overflow" attacks.  See 4007890.
282  */
283 int noexec_user_stack = 0;
284 int noexec_user_stack_log = 1;
285 
286 int segvn_lpg_disable = 0;
287 uint_t segvn_maxpgszc = 0;
288 
289 ulong_t segvn_vmpss_clrszc_cnt;
290 ulong_t segvn_vmpss_clrszc_err;
291 ulong_t segvn_fltvnpages_clrszc_cnt;
292 ulong_t segvn_fltvnpages_clrszc_err;
293 ulong_t segvn_setpgsz_align_err;
294 ulong_t segvn_setpgsz_anon_align_err;
295 ulong_t segvn_setpgsz_getattr_err;
296 ulong_t segvn_setpgsz_eof_err;
297 ulong_t segvn_faultvnmpss_align_err1;
298 ulong_t segvn_faultvnmpss_align_err2;
299 ulong_t segvn_faultvnmpss_align_err3;
300 ulong_t segvn_faultvnmpss_align_err4;
301 ulong_t segvn_faultvnmpss_align_err5;
302 ulong_t	segvn_vmpss_pageio_deadlk_err;
303 
304 int segvn_use_regions = 1;
305 
306 /*
307  * Segvn supports text replication optimization for NUMA platforms. Text
308  * replica's are represented by anon maps (amp). There's one amp per text file
309  * region per lgroup. A process chooses the amp for each of its text mappings
310  * based on the lgroup assignment of its main thread (t_tid = 1). All
311  * processes that want a replica on a particular lgroup for the same text file
312  * mapping share the same amp. amp's are looked up in svntr_hashtab hash table
313  * with vp,off,size,szc used as a key. Text replication segments are read only
314  * MAP_PRIVATE|MAP_TEXT segments that map vnode. Replication is achieved by
315  * forcing COW faults from vnode to amp and mapping amp pages instead of vnode
316  * pages. Replication amp is assigned to a segment when it gets its first
317  * pagefault. To handle main thread lgroup rehoming segvn_trasync_thread
318  * rechecks periodically if the process still maps an amp local to the main
319  * thread. If not async thread forces process to remap to an amp in the new
320  * home lgroup of the main thread. Current text replication implementation
321  * only provides the benefit to workloads that do most of their work in the
322  * main thread of a process or all the threads of a process run in the same
323  * lgroup. To extend text replication benefit to different types of
324  * multithreaded workloads further work would be needed in the hat layer to
325  * allow the same virtual address in the same hat to simultaneously map
326  * different physical addresses (i.e. page table replication would be needed
327  * for x86).
328  *
329  * amp pages are used instead of vnode pages as long as segment has a very
330  * simple life cycle.  It's created via segvn_create(), handles S_EXEC
331  * (S_READ) pagefaults and is fully unmapped.  If anything more complicated
332  * happens such as protection is changed, real COW fault happens, pagesize is
333  * changed, MC_LOCK is requested or segment is partially unmapped we turn off
334  * text replication by converting the segment back to vnode only segment
335  * (unmap segment's address range and set svd->amp to NULL).
336  *
337  * The original file can be changed after amp is inserted into
338  * svntr_hashtab. Processes that are launched after the file is already
339  * changed can't use the replica's created prior to the file change. To
340  * implement this functionality hash entries are timestamped. Replica's can
341  * only be used if current file modification time is the same as the timestamp
342  * saved when hash entry was created. However just timestamps alone are not
343  * sufficient to detect file modification via mmap(MAP_SHARED) mappings. We
344  * deal with file changes via MAP_SHARED mappings differently. When writable
345  * MAP_SHARED mappings are created to vnodes marked as executable we mark all
346  * existing replica's for this vnode as not usable for future text
347  * mappings. And we don't create new replica's for files that currently have
348  * potentially writable MAP_SHARED mappings (i.e. vn_is_mapped(V_WRITE) is
349  * true).
350  */
351 
352 #define	SEGVN_TEXTREPL_MAXBYTES_FACTOR	(20)
353 size_t	segvn_textrepl_max_bytes_factor = SEGVN_TEXTREPL_MAXBYTES_FACTOR;
354 
355 static ulong_t			svntr_hashtab_sz = 512;
356 static svntr_bucket_t		*svntr_hashtab = NULL;
357 static struct kmem_cache	*svntr_cache;
358 static svntr_stats_t		*segvn_textrepl_stats;
359 static ksema_t 			segvn_trasync_sem;
360 
361 int				segvn_disable_textrepl = 1;
362 size_t				textrepl_size_thresh = (size_t)-1;
363 size_t				segvn_textrepl_bytes = 0;
364 size_t				segvn_textrepl_max_bytes = 0;
365 clock_t				segvn_update_textrepl_interval = 0;
366 int				segvn_update_tr_time = 10;
367 int				segvn_disable_textrepl_update = 0;
368 
369 static void segvn_textrepl(struct seg *);
370 static void segvn_textunrepl(struct seg *, int);
371 static void segvn_inval_trcache(vnode_t *);
372 static void segvn_trasync_thread(void);
373 static void segvn_trupdate_wakeup(void *);
374 static void segvn_trupdate(void);
375 static void segvn_trupdate_seg(struct seg *, segvn_data_t *, svntr_t *,
376     ulong_t);
377 
378 /*
379  * Initialize segvn data structures
380  */
381 void
382 segvn_init(void)
383 {
384 	uint_t maxszc;
385 	uint_t szc;
386 	size_t pgsz;
387 
388 	segvn_cache = kmem_cache_create("segvn_cache",
389 		sizeof (struct segvn_data), 0,
390 		segvn_cache_constructor, segvn_cache_destructor, NULL,
391 		NULL, NULL, 0);
392 
393 	if (segvn_lpg_disable == 0) {
394 		szc = maxszc = page_num_pagesizes() - 1;
395 		if (szc == 0) {
396 			segvn_lpg_disable = 1;
397 		}
398 		if (page_get_pagesize(0) != PAGESIZE) {
399 			panic("segvn_init: bad szc 0");
400 			/*NOTREACHED*/
401 		}
402 		while (szc != 0) {
403 			pgsz = page_get_pagesize(szc);
404 			if (pgsz <= PAGESIZE || !IS_P2ALIGNED(pgsz, pgsz)) {
405 				panic("segvn_init: bad szc %d", szc);
406 				/*NOTREACHED*/
407 			}
408 			szc--;
409 		}
410 		if (segvn_maxpgszc == 0 || segvn_maxpgszc > maxszc)
411 			segvn_maxpgszc = maxszc;
412 	}
413 
414 	if (segvn_use_regions && !hat_supported(HAT_SHARED_REGIONS, NULL))
415 		segvn_use_regions = 0;
416 
417 	/*
418 	 * For now shared regions and text replication segvn support
419 	 * are mutually exclusive. This is acceptable because
420 	 * currently significant benefit from text replication was
421 	 * only observed on AMD64 NUMA platforms (due to relatively
422 	 * small L2$ size) and currently we don't support shared
423 	 * regions on x86.
424 	 */
425 	if (segvn_use_regions && !segvn_disable_textrepl) {
426 		segvn_disable_textrepl = 1;
427 	}
428 
429 #if defined(_LP64)
430 	if (lgrp_optimizations() && textrepl_size_thresh != (size_t)-1 &&
431 	    !segvn_disable_textrepl) {
432 		ulong_t i;
433 		size_t hsz = svntr_hashtab_sz * sizeof (svntr_bucket_t);
434 
435 		svntr_cache = kmem_cache_create("svntr_cache",
436 		    sizeof (svntr_t), 0, svntr_cache_constructor, NULL,
437 		    NULL, NULL, NULL, 0);
438 		svntr_hashtab = kmem_zalloc(hsz, KM_SLEEP);
439 		for (i = 0; i < svntr_hashtab_sz; i++) {
440 			mutex_init(&svntr_hashtab[i].tr_lock,  NULL,
441 			    MUTEX_DEFAULT, NULL);
442 		}
443 		segvn_textrepl_max_bytes = ptob(physmem) /
444 		    segvn_textrepl_max_bytes_factor;
445 		segvn_textrepl_stats = kmem_zalloc(NCPU *
446 		    sizeof (svntr_stats_t), KM_SLEEP);
447 		sema_init(&segvn_trasync_sem, 0, NULL, SEMA_DEFAULT, NULL);
448 		(void) thread_create(NULL, 0, segvn_trasync_thread,
449 		    NULL, 0, &p0, TS_RUN, minclsyspri);
450 	}
451 #endif
452 }
453 
454 #define	SEGVN_PAGEIO	((void *)0x1)
455 #define	SEGVN_NOPAGEIO	((void *)0x2)
456 
457 static void
458 segvn_setvnode_mpss(vnode_t *vp)
459 {
460 	int err;
461 
462 	ASSERT(vp->v_mpssdata == NULL ||
463 	    vp->v_mpssdata == SEGVN_PAGEIO ||
464 	    vp->v_mpssdata == SEGVN_NOPAGEIO);
465 
466 	if (vp->v_mpssdata == NULL) {
467 		if (vn_vmpss_usepageio(vp)) {
468 			err = VOP_PAGEIO(vp, (page_t *)NULL,
469 			    (u_offset_t)0, 0, 0, CRED(), NULL);
470 		} else {
471 			err = ENOSYS;
472 		}
473 		/*
474 		 * set v_mpssdata just once per vnode life
475 		 * so that it never changes.
476 		 */
477 		mutex_enter(&vp->v_lock);
478 		if (vp->v_mpssdata == NULL) {
479 			if (err == EINVAL) {
480 				vp->v_mpssdata = SEGVN_PAGEIO;
481 			} else {
482 				vp->v_mpssdata = SEGVN_NOPAGEIO;
483 			}
484 		}
485 		mutex_exit(&vp->v_lock);
486 	}
487 }
488 
489 int
490 segvn_create(struct seg *seg, void *argsp)
491 {
492 	struct segvn_crargs *a = (struct segvn_crargs *)argsp;
493 	struct segvn_data *svd;
494 	size_t swresv = 0;
495 	struct cred *cred;
496 	struct anon_map *amp;
497 	int error = 0;
498 	size_t pgsz;
499 	lgrp_mem_policy_t mpolicy = LGRP_MEM_POLICY_DEFAULT;
500 	int use_rgn = 0;
501 	int trok = 0;
502 
503 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
504 
505 	if (a->type != MAP_PRIVATE && a->type != MAP_SHARED) {
506 		panic("segvn_create type");
507 		/*NOTREACHED*/
508 	}
509 
510 	/*
511 	 * Check arguments.  If a shared anon structure is given then
512 	 * it is illegal to also specify a vp.
513 	 */
514 	if (a->amp != NULL && a->vp != NULL) {
515 		panic("segvn_create anon_map");
516 		/*NOTREACHED*/
517 	}
518 
519 	if (a->type == MAP_PRIVATE && (a->flags & MAP_TEXT) &&
520 	    a->vp != NULL && a->prot == (PROT_USER | PROT_READ | PROT_EXEC) &&
521 	    segvn_use_regions) {
522 		use_rgn = 1;
523 	}
524 
525 	/* MAP_NORESERVE on a MAP_SHARED segment is meaningless. */
526 	if (a->type == MAP_SHARED)
527 		a->flags &= ~MAP_NORESERVE;
528 
529 	if (a->szc != 0) {
530 		if (segvn_lpg_disable != 0 || (a->szc == AS_MAP_NO_LPOOB) ||
531 		    (a->amp != NULL && a->type == MAP_PRIVATE) ||
532 		    (a->flags & MAP_NORESERVE) || seg->s_as == &kas) {
533 			a->szc = 0;
534 		} else {
535 			if (a->szc > segvn_maxpgszc)
536 				a->szc = segvn_maxpgszc;
537 			pgsz = page_get_pagesize(a->szc);
538 			if (!IS_P2ALIGNED(seg->s_base, pgsz) ||
539 			    !IS_P2ALIGNED(seg->s_size, pgsz)) {
540 				a->szc = 0;
541 			} else if (a->vp != NULL) {
542 				extern struct vnode kvp;
543 				if (IS_SWAPFSVP(a->vp) || VN_ISKAS(a->vp)) {
544 					/*
545 					 * paranoid check.
546 					 * hat_page_demote() is not supported
547 					 * on swapfs pages.
548 					 */
549 					a->szc = 0;
550 				} else if (map_addr_vacalign_check(seg->s_base,
551 				    a->offset & PAGEMASK)) {
552 					a->szc = 0;
553 				}
554 			} else if (a->amp != NULL) {
555 				pgcnt_t anum = btopr(a->offset);
556 				pgcnt_t pgcnt = page_get_pagecnt(a->szc);
557 				if (!IS_P2ALIGNED(anum, pgcnt)) {
558 					a->szc = 0;
559 				}
560 			}
561 		}
562 	}
563 
564 	/*
565 	 * If segment may need private pages, reserve them now.
566 	 */
567 	if (!(a->flags & MAP_NORESERVE) && ((a->vp == NULL && a->amp == NULL) ||
568 	    (a->type == MAP_PRIVATE && (a->prot & PROT_WRITE)))) {
569 		if (anon_resv(seg->s_size) == 0)
570 			return (EAGAIN);
571 		swresv = seg->s_size;
572 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
573 			seg, swresv, 1);
574 	}
575 
576 	/*
577 	 * Reserve any mapping structures that may be required.
578 	 *
579 	 * Don't do it for segments that may use regions. It's currently a
580 	 * noop in the hat implementations anyway.
581 	 */
582 	if (!use_rgn) {
583 		hat_map(seg->s_as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
584 	}
585 
586 	if (a->cred) {
587 		cred = a->cred;
588 		crhold(cred);
589 	} else {
590 		crhold(cred = CRED());
591 	}
592 
593 	/* Inform the vnode of the new mapping */
594 	if (a->vp != NULL) {
595 		error = VOP_ADDMAP(a->vp, a->offset & PAGEMASK,
596 		    seg->s_as, seg->s_base, seg->s_size, a->prot,
597 		    a->maxprot, a->type, cred, NULL);
598 		if (error) {
599 			if (swresv != 0) {
600 				anon_unresv(swresv);
601 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
602 					"anon proc:%p %lu %u",
603 					seg, swresv, 0);
604 			}
605 			crfree(cred);
606 			if (!use_rgn) {
607 				hat_unload(seg->s_as->a_hat, seg->s_base,
608 				    seg->s_size, HAT_UNLOAD_UNMAP);
609 			}
610 			return (error);
611 		}
612 		/*
613 		 * svntr_hashtab will be NULL if we support shared regions.
614 		 */
615 		trok = ((a->flags & MAP_TEXT) &&
616 		    (seg->s_size > textrepl_size_thresh ||
617 			(a->flags & _MAP_TEXTREPL)) &&
618 		    lgrp_optimizations() && svntr_hashtab != NULL &&
619 		    a->type == MAP_PRIVATE && swresv == 0 &&
620 		    !(a->flags & MAP_NORESERVE) &&
621 		    seg->s_as != &kas && a->vp->v_type == VREG);
622 
623 		ASSERT(!trok || !use_rgn);
624 	}
625 
626 	/*
627 	 * If more than one segment in the address space, and they're adjacent
628 	 * virtually, try to concatenate them.  Don't concatenate if an
629 	 * explicit anon_map structure was supplied (e.g., SystemV shared
630 	 * memory) or if we'll use text replication for this segment.
631 	 */
632 	if (a->amp == NULL && !use_rgn && !trok) {
633 		struct seg *pseg, *nseg;
634 		struct segvn_data *psvd, *nsvd;
635 		lgrp_mem_policy_t ppolicy, npolicy;
636 		uint_t	lgrp_mem_policy_flags = 0;
637 		extern lgrp_mem_policy_t lgrp_mem_default_policy;
638 
639 		/*
640 		 * Memory policy flags (lgrp_mem_policy_flags) is valid when
641 		 * extending stack/heap segments.
642 		 */
643 		if ((a->vp == NULL) && (a->type == MAP_PRIVATE) &&
644 		    !(a->flags & MAP_NORESERVE) && (seg->s_as != &kas)) {
645 			lgrp_mem_policy_flags = a->lgrp_mem_policy_flags;
646 		} else {
647 			/*
648 			 * Get policy when not extending it from another segment
649 			 */
650 			mpolicy = lgrp_mem_policy_default(seg->s_size, a->type);
651 		}
652 
653 		/*
654 		 * First, try to concatenate the previous and new segments
655 		 */
656 		pseg = AS_SEGPREV(seg->s_as, seg);
657 		if (pseg != NULL &&
658 		    pseg->s_base + pseg->s_size == seg->s_base &&
659 		    pseg->s_ops == &segvn_ops) {
660 			/*
661 			 * Get memory allocation policy from previous segment.
662 			 * When extension is specified (e.g. for heap) apply
663 			 * this policy to the new segment regardless of the
664 			 * outcome of segment concatenation.  Extension occurs
665 			 * for non-default policy otherwise default policy is
666 			 * used and is based on extended segment size.
667 			 */
668 			psvd = (struct segvn_data *)pseg->s_data;
669 			ppolicy = psvd->policy_info.mem_policy;
670 			if (lgrp_mem_policy_flags ==
671 			    LGRP_MP_FLAG_EXTEND_UP) {
672 				if (ppolicy != lgrp_mem_default_policy) {
673 					mpolicy = ppolicy;
674 				} else {
675 					mpolicy = lgrp_mem_policy_default(
676 					    pseg->s_size + seg->s_size,
677 					    a->type);
678 				}
679 			}
680 
681 			if (mpolicy == ppolicy &&
682 			    (pseg->s_size + seg->s_size <=
683 			    segvn_comb_thrshld || psvd->amp == NULL) &&
684 			    segvn_extend_prev(pseg, seg, a, swresv) == 0) {
685 				/*
686 				 * success! now try to concatenate
687 				 * with following seg
688 				 */
689 				crfree(cred);
690 				nseg = AS_SEGNEXT(pseg->s_as, pseg);
691 				if (nseg != NULL &&
692 				    nseg != pseg &&
693 				    nseg->s_ops == &segvn_ops &&
694 				    pseg->s_base + pseg->s_size ==
695 				    nseg->s_base)
696 					(void) segvn_concat(pseg, nseg, 0);
697 				ASSERT(pseg->s_szc == 0 ||
698 				    (a->szc == pseg->s_szc &&
699 				    IS_P2ALIGNED(pseg->s_base, pgsz) &&
700 				    IS_P2ALIGNED(pseg->s_size, pgsz)));
701 				return (0);
702 			}
703 		}
704 
705 		/*
706 		 * Failed, so try to concatenate with following seg
707 		 */
708 		nseg = AS_SEGNEXT(seg->s_as, seg);
709 		if (nseg != NULL &&
710 		    seg->s_base + seg->s_size == nseg->s_base &&
711 		    nseg->s_ops == &segvn_ops) {
712 			/*
713 			 * Get memory allocation policy from next segment.
714 			 * When extension is specified (e.g. for stack) apply
715 			 * this policy to the new segment regardless of the
716 			 * outcome of segment concatenation.  Extension occurs
717 			 * for non-default policy otherwise default policy is
718 			 * used and is based on extended segment size.
719 			 */
720 			nsvd = (struct segvn_data *)nseg->s_data;
721 			npolicy = nsvd->policy_info.mem_policy;
722 			if (lgrp_mem_policy_flags ==
723 			    LGRP_MP_FLAG_EXTEND_DOWN) {
724 				if (npolicy != lgrp_mem_default_policy) {
725 					mpolicy = npolicy;
726 				} else {
727 					mpolicy = lgrp_mem_policy_default(
728 					    nseg->s_size + seg->s_size,
729 					    a->type);
730 				}
731 			}
732 
733 			if (mpolicy == npolicy &&
734 			    segvn_extend_next(seg, nseg, a, swresv) == 0) {
735 				crfree(cred);
736 				ASSERT(nseg->s_szc == 0 ||
737 				    (a->szc == nseg->s_szc &&
738 				    IS_P2ALIGNED(nseg->s_base, pgsz) &&
739 				    IS_P2ALIGNED(nseg->s_size, pgsz)));
740 				return (0);
741 			}
742 		}
743 	}
744 
745 	if (a->vp != NULL) {
746 		VN_HOLD(a->vp);
747 		if (a->type == MAP_SHARED)
748 			lgrp_shm_policy_init(NULL, a->vp);
749 	}
750 	svd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
751 
752 	seg->s_ops = &segvn_ops;
753 	seg->s_data = (void *)svd;
754 	seg->s_szc = a->szc;
755 
756 	svd->seg = seg;
757 	svd->vp = a->vp;
758 	/*
759 	 * Anonymous mappings have no backing file so the offset is meaningless.
760 	 */
761 	svd->offset = a->vp ? (a->offset & PAGEMASK) : 0;
762 	svd->prot = a->prot;
763 	svd->maxprot = a->maxprot;
764 	svd->pageprot = 0;
765 	svd->type = a->type;
766 	svd->vpage = NULL;
767 	svd->cred = cred;
768 	svd->advice = MADV_NORMAL;
769 	svd->pageadvice = 0;
770 	svd->flags = (ushort_t)a->flags;
771 	svd->softlockcnt = 0;
772 	svd->rcookie = HAT_INVALID_REGION_COOKIE;
773 
774 	if (a->szc != 0 && a->vp != NULL) {
775 		segvn_setvnode_mpss(a->vp);
776 	}
777 	if (svd->type == MAP_SHARED && svd->vp != NULL &&
778 	    (svd->vp->v_flag & VVMEXEC) && (svd->prot & PROT_WRITE)) {
779 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
780 		segvn_inval_trcache(svd->vp);
781 	}
782 
783 	amp = a->amp;
784 	if ((svd->amp = amp) == NULL) {
785 		svd->anon_index = 0;
786 		if (svd->type == MAP_SHARED) {
787 			svd->swresv = 0;
788 			/*
789 			 * Shared mappings to a vp need no other setup.
790 			 * If we have a shared mapping to an anon_map object
791 			 * which hasn't been allocated yet,  allocate the
792 			 * struct now so that it will be properly shared
793 			 * by remembering the swap reservation there.
794 			 */
795 			if (a->vp == NULL) {
796 				svd->amp = anonmap_alloc(seg->s_size, swresv,
797 				    ANON_SLEEP);
798 				svd->amp->a_szc = seg->s_szc;
799 			}
800 		} else {
801 			/*
802 			 * Private mapping (with or without a vp).
803 			 * Allocate anon_map when needed.
804 			 */
805 			svd->swresv = swresv;
806 		}
807 	} else {
808 		pgcnt_t anon_num;
809 
810 		/*
811 		 * Mapping to an existing anon_map structure without a vp.
812 		 * For now we will insure that the segment size isn't larger
813 		 * than the size - offset gives us.  Later on we may wish to
814 		 * have the anon array dynamically allocated itself so that
815 		 * we don't always have to allocate all the anon pointer slots.
816 		 * This of course involves adding extra code to check that we
817 		 * aren't trying to use an anon pointer slot beyond the end
818 		 * of the currently allocated anon array.
819 		 */
820 		if ((amp->size - a->offset) < seg->s_size) {
821 			panic("segvn_create anon_map size");
822 			/*NOTREACHED*/
823 		}
824 
825 		anon_num = btopr(a->offset);
826 
827 		if (a->type == MAP_SHARED) {
828 			/*
829 			 * SHARED mapping to a given anon_map.
830 			 */
831 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
832 			amp->refcnt++;
833 			if (a->szc > amp->a_szc) {
834 				amp->a_szc = a->szc;
835 			}
836 			ANON_LOCK_EXIT(&amp->a_rwlock);
837 			svd->anon_index = anon_num;
838 			svd->swresv = 0;
839 		} else {
840 			/*
841 			 * PRIVATE mapping to a given anon_map.
842 			 * Make sure that all the needed anon
843 			 * structures are created (so that we will
844 			 * share the underlying pages if nothing
845 			 * is written by this mapping) and then
846 			 * duplicate the anon array as is done
847 			 * when a privately mapped segment is dup'ed.
848 			 */
849 			struct anon *ap;
850 			caddr_t addr;
851 			caddr_t eaddr;
852 			ulong_t	anon_idx;
853 			int hat_flag = HAT_LOAD;
854 
855 			if (svd->flags & MAP_TEXT) {
856 				hat_flag |= HAT_LOAD_TEXT;
857 			}
858 
859 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
860 			svd->amp->a_szc = seg->s_szc;
861 			svd->anon_index = 0;
862 			svd->swresv = swresv;
863 
864 			/*
865 			 * Prevent 2 threads from allocating anon
866 			 * slots simultaneously.
867 			 */
868 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
869 			eaddr = seg->s_base + seg->s_size;
870 
871 			for (anon_idx = anon_num, addr = seg->s_base;
872 			    addr < eaddr; addr += PAGESIZE, anon_idx++) {
873 				page_t *pp;
874 
875 				if ((ap = anon_get_ptr(amp->ahp,
876 				    anon_idx)) != NULL)
877 					continue;
878 
879 				/*
880 				 * Allocate the anon struct now.
881 				 * Might as well load up translation
882 				 * to the page while we're at it...
883 				 */
884 				pp = anon_zero(seg, addr, &ap, cred);
885 				if (ap == NULL || pp == NULL) {
886 					panic("segvn_create anon_zero");
887 					/*NOTREACHED*/
888 				}
889 
890 				/*
891 				 * Re-acquire the anon_map lock and
892 				 * initialize the anon array entry.
893 				 */
894 				ASSERT(anon_get_ptr(amp->ahp,
895 				    anon_idx) == NULL);
896 				(void) anon_set_ptr(amp->ahp, anon_idx, ap,
897 				    ANON_SLEEP);
898 
899 				ASSERT(seg->s_szc == 0);
900 				ASSERT(!IS_VMODSORT(pp->p_vnode));
901 
902 				ASSERT(use_rgn == 0);
903 				hat_memload(seg->s_as->a_hat, addr, pp,
904 					svd->prot & ~PROT_WRITE, hat_flag);
905 
906 				page_unlock(pp);
907 			}
908 			ASSERT(seg->s_szc == 0);
909 			anon_dup(amp->ahp, anon_num, svd->amp->ahp,
910 			    0, seg->s_size);
911 			ANON_LOCK_EXIT(&amp->a_rwlock);
912 		}
913 	}
914 
915 	/*
916 	 * Set default memory allocation policy for segment
917 	 *
918 	 * Always set policy for private memory at least for initialization
919 	 * even if this is a shared memory segment
920 	 */
921 	(void) lgrp_privm_policy_set(mpolicy, &svd->policy_info, seg->s_size);
922 
923 	if (svd->type == MAP_SHARED)
924 		(void) lgrp_shm_policy_set(mpolicy, svd->amp, svd->anon_index,
925 		    svd->vp, svd->offset, seg->s_size);
926 
927 	if (use_rgn) {
928 		ASSERT(!trok);
929 		ASSERT(svd->amp == NULL);
930 		svd->rcookie = hat_join_region(seg->s_as->a_hat, seg->s_base,
931 		    seg->s_size, (void *)svd->vp, svd->offset, svd->prot,
932 		    (uchar_t)seg->s_szc, segvn_hat_rgn_unload_callback,
933 		    HAT_REGION_TEXT);
934 	}
935 
936 	ASSERT(!trok || !(svd->prot & PROT_WRITE));
937 	svd->tr_state = trok ? SEGVN_TR_INIT : SEGVN_TR_OFF;
938 
939 	return (0);
940 }
941 
942 /*
943  * Concatenate two existing segments, if possible.
944  * Return 0 on success, -1 if two segments are not compatible
945  * or -2 on memory allocation failure.
946  * If amp_cat == 1 then try and concat segments with anon maps
947  */
948 static int
949 segvn_concat(struct seg *seg1, struct seg *seg2, int amp_cat)
950 {
951 	struct segvn_data *svd1 = seg1->s_data;
952 	struct segvn_data *svd2 = seg2->s_data;
953 	struct anon_map *amp1 = svd1->amp;
954 	struct anon_map *amp2 = svd2->amp;
955 	struct vpage *vpage1 = svd1->vpage;
956 	struct vpage *vpage2 = svd2->vpage, *nvpage = NULL;
957 	size_t size, nvpsize;
958 	pgcnt_t npages1, npages2;
959 
960 	ASSERT(seg1->s_as && seg2->s_as && seg1->s_as == seg2->s_as);
961 	ASSERT(AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
962 	ASSERT(seg1->s_ops == seg2->s_ops);
963 
964 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie) ||
965 	    HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
966 		return (-1);
967 	}
968 
969 	/* both segments exist, try to merge them */
970 #define	incompat(x)	(svd1->x != svd2->x)
971 	if (incompat(vp) || incompat(maxprot) ||
972 	    (!svd1->pageadvice && !svd2->pageadvice && incompat(advice)) ||
973 	    (!svd1->pageprot && !svd2->pageprot && incompat(prot)) ||
974 	    incompat(type) || incompat(cred) || incompat(flags) ||
975 	    seg1->s_szc != seg2->s_szc || incompat(policy_info.mem_policy) ||
976 	    (svd2->softlockcnt > 0))
977 		return (-1);
978 #undef incompat
979 
980 	/*
981 	 * vp == NULL implies zfod, offset doesn't matter
982 	 */
983 	if (svd1->vp != NULL &&
984 	    svd1->offset + seg1->s_size != svd2->offset) {
985 		return (-1);
986 	}
987 
988 	/*
989 	 * Don't concatenate if either segment uses text replication.
990 	 */
991 	if (svd1->tr_state != SEGVN_TR_OFF || svd2->tr_state != SEGVN_TR_OFF) {
992 		return (-1);
993 	}
994 
995 	/*
996 	 * Fail early if we're not supposed to concatenate
997 	 * segments with non NULL amp.
998 	 */
999 	if (amp_cat == 0 && (amp1 != NULL || amp2 != NULL)) {
1000 		return (-1);
1001 	}
1002 
1003 	if (svd1->vp == NULL && svd1->type == MAP_SHARED) {
1004 		if (amp1 != amp2) {
1005 			return (-1);
1006 		}
1007 		if (amp1 != NULL && svd1->anon_index + btop(seg1->s_size) !=
1008 		    svd2->anon_index) {
1009 			return (-1);
1010 		}
1011 		ASSERT(amp1 == NULL || amp1->refcnt >= 2);
1012 	}
1013 
1014 	/*
1015 	 * If either seg has vpages, create a new merged vpage array.
1016 	 */
1017 	if (vpage1 != NULL || vpage2 != NULL) {
1018 		struct vpage *vp;
1019 
1020 		npages1 = seg_pages(seg1);
1021 		npages2 = seg_pages(seg2);
1022 		nvpsize = vpgtob(npages1 + npages2);
1023 
1024 		if ((nvpage = kmem_zalloc(nvpsize, KM_NOSLEEP)) == NULL) {
1025 			return (-2);
1026 		}
1027 
1028 		if (vpage1 != NULL) {
1029 			bcopy(vpage1, nvpage, vpgtob(npages1));
1030 		} else {
1031 			for (vp = nvpage; vp < nvpage + npages1; vp++) {
1032 				VPP_SETPROT(vp, svd1->prot);
1033 				VPP_SETADVICE(vp, svd1->advice);
1034 			}
1035 		}
1036 
1037 		if (vpage2 != NULL) {
1038 			bcopy(vpage2, nvpage + npages1, vpgtob(npages2));
1039 		} else {
1040 			for (vp = nvpage + npages1;
1041 			    vp < nvpage + npages1 + npages2; vp++) {
1042 				VPP_SETPROT(vp, svd2->prot);
1043 				VPP_SETADVICE(vp, svd2->advice);
1044 			}
1045 		}
1046 	}
1047 
1048 	/*
1049 	 * If either segment has private pages, create a new merged anon
1050 	 * array. If mergeing shared anon segments just decrement anon map's
1051 	 * refcnt.
1052 	 */
1053 	if (amp1 != NULL && svd1->type == MAP_SHARED) {
1054 		ASSERT(amp1 == amp2 && svd1->vp == NULL);
1055 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1056 		ASSERT(amp1->refcnt >= 2);
1057 		amp1->refcnt--;
1058 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1059 		svd2->amp = NULL;
1060 	} else if (amp1 != NULL || amp2 != NULL) {
1061 		struct anon_hdr *nahp;
1062 		struct anon_map *namp = NULL;
1063 		size_t asize;
1064 
1065 		ASSERT(svd1->type == MAP_PRIVATE);
1066 
1067 		asize = seg1->s_size + seg2->s_size;
1068 		if ((nahp = anon_create(btop(asize), ANON_NOSLEEP)) == NULL) {
1069 			if (nvpage != NULL) {
1070 				kmem_free(nvpage, nvpsize);
1071 			}
1072 			return (-2);
1073 		}
1074 		if (amp1 != NULL) {
1075 			/*
1076 			 * XXX anon rwlock is not really needed because
1077 			 * this is a private segment and we are writers.
1078 			 */
1079 			ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1080 			ASSERT(amp1->refcnt == 1);
1081 			if (anon_copy_ptr(amp1->ahp, svd1->anon_index,
1082 			    nahp, 0, btop(seg1->s_size), ANON_NOSLEEP)) {
1083 				anon_release(nahp, btop(asize));
1084 				ANON_LOCK_EXIT(&amp1->a_rwlock);
1085 				if (nvpage != NULL) {
1086 					kmem_free(nvpage, nvpsize);
1087 				}
1088 				return (-2);
1089 			}
1090 		}
1091 		if (amp2 != NULL) {
1092 			ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1093 			ASSERT(amp2->refcnt == 1);
1094 			if (anon_copy_ptr(amp2->ahp, svd2->anon_index,
1095 			    nahp, btop(seg1->s_size), btop(seg2->s_size),
1096 			    ANON_NOSLEEP)) {
1097 				anon_release(nahp, btop(asize));
1098 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1099 				if (amp1 != NULL) {
1100 					ANON_LOCK_EXIT(&amp1->a_rwlock);
1101 				}
1102 				if (nvpage != NULL) {
1103 					kmem_free(nvpage, nvpsize);
1104 				}
1105 				return (-2);
1106 			}
1107 		}
1108 		if (amp1 != NULL) {
1109 			namp = amp1;
1110 			anon_release(amp1->ahp, btop(amp1->size));
1111 		}
1112 		if (amp2 != NULL) {
1113 			if (namp == NULL) {
1114 				ASSERT(amp1 == NULL);
1115 				namp = amp2;
1116 				anon_release(amp2->ahp, btop(amp2->size));
1117 			} else {
1118 				amp2->refcnt--;
1119 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1120 				anonmap_free(amp2);
1121 			}
1122 			svd2->amp = NULL; /* needed for seg_free */
1123 		}
1124 		namp->ahp = nahp;
1125 		namp->size = asize;
1126 		svd1->amp = namp;
1127 		svd1->anon_index = 0;
1128 		ANON_LOCK_EXIT(&namp->a_rwlock);
1129 	}
1130 	/*
1131 	 * Now free the old vpage structures.
1132 	 */
1133 	if (nvpage != NULL) {
1134 		if (vpage1 != NULL) {
1135 			kmem_free(vpage1, vpgtob(npages1));
1136 		}
1137 		if (vpage2 != NULL) {
1138 			svd2->vpage = NULL;
1139 			kmem_free(vpage2, vpgtob(npages2));
1140 		}
1141 		if (svd2->pageprot) {
1142 			svd1->pageprot = 1;
1143 		}
1144 		if (svd2->pageadvice) {
1145 			svd1->pageadvice = 1;
1146 		}
1147 		svd1->vpage = nvpage;
1148 	}
1149 
1150 	/* all looks ok, merge segments */
1151 	svd1->swresv += svd2->swresv;
1152 	svd2->swresv = 0;  /* so seg_free doesn't release swap space */
1153 	size = seg2->s_size;
1154 	seg_free(seg2);
1155 	seg1->s_size += size;
1156 	return (0);
1157 }
1158 
1159 /*
1160  * Extend the previous segment (seg1) to include the
1161  * new segment (seg2 + a), if possible.
1162  * Return 0 on success.
1163  */
1164 static int
1165 segvn_extend_prev(seg1, seg2, a, swresv)
1166 	struct seg *seg1, *seg2;
1167 	struct segvn_crargs *a;
1168 	size_t swresv;
1169 {
1170 	struct segvn_data *svd1 = (struct segvn_data *)seg1->s_data;
1171 	size_t size;
1172 	struct anon_map *amp1;
1173 	struct vpage *new_vpage;
1174 
1175 	/*
1176 	 * We don't need any segment level locks for "segvn" data
1177 	 * since the address space is "write" locked.
1178 	 */
1179 	ASSERT(seg1->s_as && AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
1180 
1181 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie)) {
1182 		return (-1);
1183 	}
1184 
1185 	/* second segment is new, try to extend first */
1186 	/* XXX - should also check cred */
1187 	if (svd1->vp != a->vp || svd1->maxprot != a->maxprot ||
1188 	    (!svd1->pageprot && (svd1->prot != a->prot)) ||
1189 	    svd1->type != a->type || svd1->flags != a->flags ||
1190 	    seg1->s_szc != a->szc)
1191 		return (-1);
1192 
1193 	/* vp == NULL implies zfod, offset doesn't matter */
1194 	if (svd1->vp != NULL &&
1195 	    svd1->offset + seg1->s_size != (a->offset & PAGEMASK))
1196 		return (-1);
1197 
1198 	if (svd1->tr_state != SEGVN_TR_OFF) {
1199 		return (-1);
1200 	}
1201 
1202 	amp1 = svd1->amp;
1203 	if (amp1) {
1204 		pgcnt_t newpgs;
1205 
1206 		/*
1207 		 * Segment has private pages, can data structures
1208 		 * be expanded?
1209 		 *
1210 		 * Acquire the anon_map lock to prevent it from changing,
1211 		 * if it is shared.  This ensures that the anon_map
1212 		 * will not change while a thread which has a read/write
1213 		 * lock on an address space references it.
1214 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1215 		 * is 1.
1216 		 *
1217 		 * Can't grow a MAP_SHARED segment with an anonmap because
1218 		 * there may be existing anon slots where we want to extend
1219 		 * the segment and we wouldn't know what to do with them
1220 		 * (e.g., for tmpfs right thing is to just leave them there,
1221 		 * for /dev/zero they should be cleared out).
1222 		 */
1223 		if (svd1->type == MAP_SHARED)
1224 			return (-1);
1225 
1226 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1227 		if (amp1->refcnt > 1) {
1228 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1229 			return (-1);
1230 		}
1231 		newpgs = anon_grow(amp1->ahp, &svd1->anon_index,
1232 		    btop(seg1->s_size), btop(seg2->s_size), ANON_NOSLEEP);
1233 
1234 		if (newpgs == 0) {
1235 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1236 			return (-1);
1237 		}
1238 		amp1->size = ptob(newpgs);
1239 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1240 	}
1241 	if (svd1->vpage != NULL) {
1242 		struct vpage *vp, *evp;
1243 		new_vpage =
1244 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1245 			KM_NOSLEEP);
1246 		if (new_vpage == NULL)
1247 			return (-1);
1248 		bcopy(svd1->vpage, new_vpage, vpgtob(seg_pages(seg1)));
1249 		kmem_free(svd1->vpage, vpgtob(seg_pages(seg1)));
1250 		svd1->vpage = new_vpage;
1251 
1252 		vp = new_vpage + seg_pages(seg1);
1253 		evp = vp + seg_pages(seg2);
1254 		for (; vp < evp; vp++)
1255 			VPP_SETPROT(vp, a->prot);
1256 	}
1257 	size = seg2->s_size;
1258 	seg_free(seg2);
1259 	seg1->s_size += size;
1260 	svd1->swresv += swresv;
1261 	if (svd1->pageprot && (a->prot & PROT_WRITE) &&
1262 	    svd1->type == MAP_SHARED && svd1->vp != NULL &&
1263 	    (svd1->vp->v_flag & VVMEXEC)) {
1264 		ASSERT(vn_is_mapped(svd1->vp, V_WRITE));
1265 		segvn_inval_trcache(svd1->vp);
1266 	}
1267 	return (0);
1268 }
1269 
1270 /*
1271  * Extend the next segment (seg2) to include the
1272  * new segment (seg1 + a), if possible.
1273  * Return 0 on success.
1274  */
1275 static int
1276 segvn_extend_next(
1277 	struct seg *seg1,
1278 	struct seg *seg2,
1279 	struct segvn_crargs *a,
1280 	size_t swresv)
1281 {
1282 	struct segvn_data *svd2 = (struct segvn_data *)seg2->s_data;
1283 	size_t size;
1284 	struct anon_map *amp2;
1285 	struct vpage *new_vpage;
1286 
1287 	/*
1288 	 * We don't need any segment level locks for "segvn" data
1289 	 * since the address space is "write" locked.
1290 	 */
1291 	ASSERT(seg2->s_as && AS_WRITE_HELD(seg2->s_as, &seg2->s_as->a_lock));
1292 
1293 	if (HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1294 		return (-1);
1295 	}
1296 
1297 	/* first segment is new, try to extend second */
1298 	/* XXX - should also check cred */
1299 	if (svd2->vp != a->vp || svd2->maxprot != a->maxprot ||
1300 	    (!svd2->pageprot && (svd2->prot != a->prot)) ||
1301 	    svd2->type != a->type || svd2->flags != a->flags ||
1302 	    seg2->s_szc != a->szc)
1303 		return (-1);
1304 	/* vp == NULL implies zfod, offset doesn't matter */
1305 	if (svd2->vp != NULL &&
1306 	    (a->offset & PAGEMASK) + seg1->s_size != svd2->offset)
1307 		return (-1);
1308 
1309 	if (svd2->tr_state != SEGVN_TR_OFF) {
1310 		return (-1);
1311 	}
1312 
1313 	amp2 = svd2->amp;
1314 	if (amp2) {
1315 		pgcnt_t newpgs;
1316 
1317 		/*
1318 		 * Segment has private pages, can data structures
1319 		 * be expanded?
1320 		 *
1321 		 * Acquire the anon_map lock to prevent it from changing,
1322 		 * if it is shared.  This ensures that the anon_map
1323 		 * will not change while a thread which has a read/write
1324 		 * lock on an address space references it.
1325 		 *
1326 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1327 		 * is 1.
1328 		 */
1329 		if (svd2->type == MAP_SHARED)
1330 			return (-1);
1331 
1332 		ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1333 		if (amp2->refcnt > 1) {
1334 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1335 			return (-1);
1336 		}
1337 		newpgs = anon_grow(amp2->ahp, &svd2->anon_index,
1338 		    btop(seg2->s_size), btop(seg1->s_size),
1339 		    ANON_NOSLEEP | ANON_GROWDOWN);
1340 
1341 		if (newpgs == 0) {
1342 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1343 			return (-1);
1344 		}
1345 		amp2->size = ptob(newpgs);
1346 		ANON_LOCK_EXIT(&amp2->a_rwlock);
1347 	}
1348 	if (svd2->vpage != NULL) {
1349 		struct vpage *vp, *evp;
1350 		new_vpage =
1351 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1352 			KM_NOSLEEP);
1353 		if (new_vpage == NULL) {
1354 			/* Not merging segments so adjust anon_index back */
1355 			if (amp2)
1356 				svd2->anon_index += seg_pages(seg1);
1357 			return (-1);
1358 		}
1359 		bcopy(svd2->vpage, new_vpage + seg_pages(seg1),
1360 		    vpgtob(seg_pages(seg2)));
1361 		kmem_free(svd2->vpage, vpgtob(seg_pages(seg2)));
1362 		svd2->vpage = new_vpage;
1363 
1364 		vp = new_vpage;
1365 		evp = vp + seg_pages(seg1);
1366 		for (; vp < evp; vp++)
1367 			VPP_SETPROT(vp, a->prot);
1368 	}
1369 	size = seg1->s_size;
1370 	seg_free(seg1);
1371 	seg2->s_size += size;
1372 	seg2->s_base -= size;
1373 	svd2->offset -= size;
1374 	svd2->swresv += swresv;
1375 	if (svd2->pageprot && (a->prot & PROT_WRITE) &&
1376 	    svd2->type == MAP_SHARED && svd2->vp != NULL &&
1377 	    (svd2->vp->v_flag & VVMEXEC)) {
1378 		ASSERT(vn_is_mapped(svd2->vp, V_WRITE));
1379 		segvn_inval_trcache(svd2->vp);
1380 	}
1381 	return (0);
1382 }
1383 
1384 static int
1385 segvn_dup(struct seg *seg, struct seg *newseg)
1386 {
1387 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1388 	struct segvn_data *newsvd;
1389 	pgcnt_t npages = seg_pages(seg);
1390 	int error = 0;
1391 	uint_t prot;
1392 	size_t len;
1393 	struct anon_map *amp;
1394 
1395 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1396 
1397 	/*
1398 	 * If segment has anon reserved, reserve more for the new seg.
1399 	 * For a MAP_NORESERVE segment swresv will be a count of all the
1400 	 * allocated anon slots; thus we reserve for the child as many slots
1401 	 * as the parent has allocated. This semantic prevents the child or
1402 	 * parent from dieing during a copy-on-write fault caused by trying
1403 	 * to write a shared pre-existing anon page.
1404 	 */
1405 	if ((len = svd->swresv) != 0) {
1406 		if (anon_resv(svd->swresv) == 0)
1407 			return (ENOMEM);
1408 
1409 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
1410 			seg, len, 0);
1411 	}
1412 
1413 	newsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
1414 
1415 	newseg->s_ops = &segvn_ops;
1416 	newseg->s_data = (void *)newsvd;
1417 	newseg->s_szc = seg->s_szc;
1418 
1419 	newsvd->seg = newseg;
1420 	if ((newsvd->vp = svd->vp) != NULL) {
1421 		VN_HOLD(svd->vp);
1422 		if (svd->type == MAP_SHARED)
1423 			lgrp_shm_policy_init(NULL, svd->vp);
1424 	}
1425 	newsvd->offset = svd->offset;
1426 	newsvd->prot = svd->prot;
1427 	newsvd->maxprot = svd->maxprot;
1428 	newsvd->pageprot = svd->pageprot;
1429 	newsvd->type = svd->type;
1430 	newsvd->cred = svd->cred;
1431 	crhold(newsvd->cred);
1432 	newsvd->advice = svd->advice;
1433 	newsvd->pageadvice = svd->pageadvice;
1434 	newsvd->swresv = svd->swresv;
1435 	newsvd->flags = svd->flags;
1436 	newsvd->softlockcnt = 0;
1437 	newsvd->policy_info = svd->policy_info;
1438 	newsvd->rcookie = HAT_INVALID_REGION_COOKIE;
1439 
1440 	if ((amp = svd->amp) == NULL || svd->tr_state == SEGVN_TR_ON) {
1441 		/*
1442 		 * Not attaching to a shared anon object.
1443 		 */
1444 		ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie) ||
1445 		    svd->tr_state == SEGVN_TR_OFF);
1446 		if (svd->tr_state == SEGVN_TR_ON) {
1447 			ASSERT(newsvd->vp != NULL && amp != NULL);
1448 			newsvd->tr_state = SEGVN_TR_INIT;
1449 		} else {
1450 			newsvd->tr_state = svd->tr_state;
1451 		}
1452 		newsvd->amp = NULL;
1453 		newsvd->anon_index = 0;
1454 	} else {
1455 		/* regions for now are only used on pure vnode segments */
1456 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
1457 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1458 		newsvd->tr_state = SEGVN_TR_OFF;
1459 		if (svd->type == MAP_SHARED) {
1460 			newsvd->amp = amp;
1461 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1462 			amp->refcnt++;
1463 			ANON_LOCK_EXIT(&amp->a_rwlock);
1464 			newsvd->anon_index = svd->anon_index;
1465 		} else {
1466 			int reclaim = 1;
1467 
1468 			/*
1469 			 * Allocate and initialize new anon_map structure.
1470 			 */
1471 			newsvd->amp = anonmap_alloc(newseg->s_size, 0,
1472 			    ANON_SLEEP);
1473 			newsvd->amp->a_szc = newseg->s_szc;
1474 			newsvd->anon_index = 0;
1475 
1476 			/*
1477 			 * We don't have to acquire the anon_map lock
1478 			 * for the new segment (since it belongs to an
1479 			 * address space that is still not associated
1480 			 * with any process), or the segment in the old
1481 			 * address space (since all threads in it
1482 			 * are stopped while duplicating the address space).
1483 			 */
1484 
1485 			/*
1486 			 * The goal of the following code is to make sure that
1487 			 * softlocked pages do not end up as copy on write
1488 			 * pages.  This would cause problems where one
1489 			 * thread writes to a page that is COW and a different
1490 			 * thread in the same process has softlocked it.  The
1491 			 * softlock lock would move away from this process
1492 			 * because the write would cause this process to get
1493 			 * a copy (without the softlock).
1494 			 *
1495 			 * The strategy here is to just break the
1496 			 * sharing on pages that could possibly be
1497 			 * softlocked.
1498 			 */
1499 retry:
1500 			if (svd->softlockcnt) {
1501 				struct anon *ap, *newap;
1502 				size_t i;
1503 				uint_t vpprot;
1504 				page_t *anon_pl[1+1], *pp;
1505 				caddr_t addr;
1506 				ulong_t old_idx = svd->anon_index;
1507 				ulong_t new_idx = 0;
1508 
1509 				/*
1510 				 * The softlock count might be non zero
1511 				 * because some pages are still stuck in the
1512 				 * cache for lazy reclaim. Flush the cache
1513 				 * now. This should drop the count to zero.
1514 				 * [or there is really I/O going on to these
1515 				 * pages]. Note, we have the writers lock so
1516 				 * nothing gets inserted during the flush.
1517 				 */
1518 				if (reclaim == 1) {
1519 					segvn_purge(seg);
1520 					reclaim = 0;
1521 					goto retry;
1522 				}
1523 				i = btopr(seg->s_size);
1524 				addr = seg->s_base;
1525 				/*
1526 				 * XXX break cow sharing using PAGESIZE
1527 				 * pages. They will be relocated into larger
1528 				 * pages at fault time.
1529 				 */
1530 				while (i-- > 0) {
1531 					if (ap = anon_get_ptr(amp->ahp,
1532 					    old_idx)) {
1533 						error = anon_getpage(&ap,
1534 						    &vpprot, anon_pl, PAGESIZE,
1535 						    seg, addr, S_READ,
1536 						    svd->cred);
1537 						if (error) {
1538 							newsvd->vpage = NULL;
1539 							goto out;
1540 						}
1541 						/*
1542 						 * prot need not be computed
1543 						 * below 'cause anon_private is
1544 						 * going to ignore it anyway
1545 						 * as child doesn't inherit
1546 						 * pagelock from parent.
1547 						 */
1548 						prot = svd->pageprot ?
1549 						    VPP_PROT(
1550 						    &svd->vpage[
1551 						    seg_page(seg, addr)])
1552 						    : svd->prot;
1553 						pp = anon_private(&newap,
1554 						    newseg, addr, prot,
1555 						    anon_pl[0],	0,
1556 						    newsvd->cred);
1557 						if (pp == NULL) {
1558 							/* no mem abort */
1559 							newsvd->vpage = NULL;
1560 							error = ENOMEM;
1561 							goto out;
1562 						}
1563 						(void) anon_set_ptr(
1564 						    newsvd->amp->ahp, new_idx,
1565 						    newap, ANON_SLEEP);
1566 						page_unlock(pp);
1567 					}
1568 					addr += PAGESIZE;
1569 					old_idx++;
1570 					new_idx++;
1571 				}
1572 			} else {	/* common case */
1573 				if (seg->s_szc != 0) {
1574 					/*
1575 					 * If at least one of anon slots of a
1576 					 * large page exists then make sure
1577 					 * all anon slots of a large page
1578 					 * exist to avoid partial cow sharing
1579 					 * of a large page in the future.
1580 					 */
1581 					anon_dup_fill_holes(amp->ahp,
1582 					    svd->anon_index, newsvd->amp->ahp,
1583 					    0, seg->s_size, seg->s_szc,
1584 					    svd->vp != NULL);
1585 				} else {
1586 					anon_dup(amp->ahp, svd->anon_index,
1587 					    newsvd->amp->ahp, 0, seg->s_size);
1588 				}
1589 
1590 				hat_clrattr(seg->s_as->a_hat, seg->s_base,
1591 				    seg->s_size, PROT_WRITE);
1592 			}
1593 		}
1594 	}
1595 	/*
1596 	 * If necessary, create a vpage structure for the new segment.
1597 	 * Do not copy any page lock indications.
1598 	 */
1599 	if (svd->vpage != NULL) {
1600 		uint_t i;
1601 		struct vpage *ovp = svd->vpage;
1602 		struct vpage *nvp;
1603 
1604 		nvp = newsvd->vpage =
1605 		    kmem_alloc(vpgtob(npages), KM_SLEEP);
1606 		for (i = 0; i < npages; i++) {
1607 			*nvp = *ovp++;
1608 			VPP_CLRPPLOCK(nvp++);
1609 		}
1610 	} else
1611 		newsvd->vpage = NULL;
1612 
1613 	/* Inform the vnode of the new mapping */
1614 	if (newsvd->vp != NULL) {
1615 		error = VOP_ADDMAP(newsvd->vp, (offset_t)newsvd->offset,
1616 		    newseg->s_as, newseg->s_base, newseg->s_size, newsvd->prot,
1617 		    newsvd->maxprot, newsvd->type, newsvd->cred, NULL);
1618 	}
1619 out:
1620 	if (error == 0 && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1621 		ASSERT(newsvd->amp == NULL);
1622 		ASSERT(newsvd->tr_state == SEGVN_TR_OFF);
1623 		newsvd->rcookie = svd->rcookie;
1624 		hat_dup_region(newseg->s_as->a_hat, newsvd->rcookie);
1625 	}
1626 	return (error);
1627 }
1628 
1629 
1630 /*
1631  * callback function to invoke free_vp_pages() for only those pages actually
1632  * processed by the HAT when a shared region is destroyed.
1633  */
1634 extern int free_pages;
1635 
1636 static void
1637 segvn_hat_rgn_unload_callback(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
1638     size_t r_size, void *r_obj, u_offset_t r_objoff)
1639 {
1640 	u_offset_t off;
1641 	size_t len;
1642 	vnode_t *vp = (vnode_t *)r_obj;
1643 
1644 	ASSERT(eaddr > saddr);
1645 	ASSERT(saddr >= r_saddr);
1646 	ASSERT(saddr < r_saddr + r_size);
1647 	ASSERT(eaddr > r_saddr);
1648 	ASSERT(eaddr <= r_saddr + r_size);
1649 	ASSERT(vp != NULL);
1650 
1651 	if (!free_pages) {
1652 		return;
1653 	}
1654 
1655 	len = eaddr - saddr;
1656 	off = (saddr - r_saddr) + r_objoff;
1657 	free_vp_pages(vp, off, len);
1658 }
1659 
1660 /*
1661  * callback function used by segvn_unmap to invoke free_vp_pages() for only
1662  * those pages actually processed by the HAT
1663  */
1664 static void
1665 segvn_hat_unload_callback(hat_callback_t *cb)
1666 {
1667 	struct seg		*seg = cb->hcb_data;
1668 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
1669 	size_t			len;
1670 	u_offset_t		off;
1671 
1672 	ASSERT(svd->vp != NULL);
1673 	ASSERT(cb->hcb_end_addr > cb->hcb_start_addr);
1674 	ASSERT(cb->hcb_start_addr >= seg->s_base);
1675 
1676 	len = cb->hcb_end_addr - cb->hcb_start_addr;
1677 	off = cb->hcb_start_addr - seg->s_base;
1678 	free_vp_pages(svd->vp, svd->offset + off, len);
1679 }
1680 
1681 static int
1682 segvn_unmap(struct seg *seg, caddr_t addr, size_t len)
1683 {
1684 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1685 	struct segvn_data *nsvd;
1686 	struct seg *nseg;
1687 	struct anon_map *amp;
1688 	pgcnt_t	opages;		/* old segment size in pages */
1689 	pgcnt_t	npages;		/* new segment size in pages */
1690 	pgcnt_t	dpages;		/* pages being deleted (unmapped) */
1691 	hat_callback_t callback;	/* used for free_vp_pages() */
1692 	hat_callback_t *cbp = NULL;
1693 	caddr_t nbase;
1694 	size_t nsize;
1695 	size_t oswresv;
1696 	int reclaim = 1;
1697 
1698 	/*
1699 	 * We don't need any segment level locks for "segvn" data
1700 	 * since the address space is "write" locked.
1701 	 */
1702 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1703 
1704 	/*
1705 	 * Fail the unmap if pages are SOFTLOCKed through this mapping.
1706 	 * softlockcnt is protected from change by the as write lock.
1707 	 */
1708 retry:
1709 	if (svd->softlockcnt > 0) {
1710 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1711 		/*
1712 		 * since we do have the writers lock nobody can fill
1713 		 * the cache during the purge. The flush either succeeds
1714 		 * or we still have pending I/Os.
1715 		 */
1716 		if (reclaim == 1) {
1717 			segvn_purge(seg);
1718 			reclaim = 0;
1719 			goto retry;
1720 		}
1721 		return (EAGAIN);
1722 	}
1723 
1724 	/*
1725 	 * Check for bad sizes
1726 	 */
1727 	if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
1728 	    (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
1729 		panic("segvn_unmap");
1730 		/*NOTREACHED*/
1731 	}
1732 
1733 	if (seg->s_szc != 0) {
1734 		size_t pgsz = page_get_pagesize(seg->s_szc);
1735 		int err;
1736 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
1737 			ASSERT(seg->s_base != addr || seg->s_size != len);
1738 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1739 				ASSERT(svd->amp == NULL);
1740 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1741 				hat_leave_region(seg->s_as->a_hat,
1742 				    svd->rcookie, HAT_REGION_TEXT);
1743 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
1744 				/*
1745 				 * could pass a flag to segvn_demote_range()
1746 				 * below to tell it not to do any unloads but
1747 				 * this case is rare enough to not bother for
1748 				 * now.
1749 				 */
1750 			} else if (svd->tr_state == SEGVN_TR_INIT) {
1751 				svd->tr_state = SEGVN_TR_OFF;
1752 			} else if (svd->tr_state == SEGVN_TR_ON) {
1753 				ASSERT(svd->amp != NULL);
1754 				segvn_textunrepl(seg, 1);
1755 				ASSERT(svd->amp == NULL);
1756 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1757 			}
1758 			VM_STAT_ADD(segvnvmstats.demoterange[0]);
1759 			err = segvn_demote_range(seg, addr, len, SDR_END, 0);
1760 			if (err == 0) {
1761 				return (IE_RETRY);
1762 			}
1763 			return (err);
1764 		}
1765 	}
1766 
1767 	/* Inform the vnode of the unmapping. */
1768 	if (svd->vp) {
1769 		int error;
1770 
1771 		error = VOP_DELMAP(svd->vp,
1772 			(offset_t)svd->offset + (uintptr_t)(addr - seg->s_base),
1773 			seg->s_as, addr, len, svd->prot, svd->maxprot,
1774 			svd->type, svd->cred, NULL);
1775 
1776 		if (error == EAGAIN)
1777 			return (error);
1778 	}
1779 
1780 	/*
1781 	 * Remove any page locks set through this mapping.
1782 	 * If text replication is not off no page locks could have been
1783 	 * established via this mapping.
1784 	 */
1785 	if (svd->tr_state == SEGVN_TR_OFF) {
1786 		(void) segvn_lockop(seg, addr, len, 0, MC_UNLOCK, NULL, 0);
1787 	}
1788 
1789 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1790 		ASSERT(svd->amp == NULL);
1791 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1792 		ASSERT(svd->type == MAP_PRIVATE);
1793 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
1794 		    HAT_REGION_TEXT);
1795 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
1796 	} else if (svd->tr_state == SEGVN_TR_ON) {
1797 		ASSERT(svd->amp != NULL);
1798 		ASSERT(svd->pageprot == 0 && !(svd->prot & PROT_WRITE));
1799 		segvn_textunrepl(seg, 1);
1800 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
1801 	} else {
1802 		if (svd->tr_state != SEGVN_TR_OFF) {
1803 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
1804 			svd->tr_state = SEGVN_TR_OFF;
1805 		}
1806 		/*
1807 		 * Unload any hardware translations in the range to be taken
1808 		 * out. Use a callback to invoke free_vp_pages() effectively.
1809 		 */
1810 		if (svd->vp != NULL && free_pages != 0) {
1811 			callback.hcb_data = seg;
1812 			callback.hcb_function = segvn_hat_unload_callback;
1813 			cbp = &callback;
1814 		}
1815 		hat_unload_callback(seg->s_as->a_hat, addr, len,
1816 		    HAT_UNLOAD_UNMAP, cbp);
1817 
1818 		if (svd->type == MAP_SHARED && svd->vp != NULL &&
1819 		    (svd->vp->v_flag & VVMEXEC) &&
1820 		    ((svd->prot & PROT_WRITE) || svd->pageprot)) {
1821 			segvn_inval_trcache(svd->vp);
1822 		}
1823 	}
1824 
1825 	/*
1826 	 * Check for entire segment
1827 	 */
1828 	if (addr == seg->s_base && len == seg->s_size) {
1829 		seg_free(seg);
1830 		return (0);
1831 	}
1832 
1833 	opages = seg_pages(seg);
1834 	dpages = btop(len);
1835 	npages = opages - dpages;
1836 	amp = svd->amp;
1837 	ASSERT(amp == NULL || amp->a_szc >= seg->s_szc);
1838 
1839 	/*
1840 	 * Check for beginning of segment
1841 	 */
1842 	if (addr == seg->s_base) {
1843 		if (svd->vpage != NULL) {
1844 			size_t nbytes;
1845 			struct vpage *ovpage;
1846 
1847 			ovpage = svd->vpage;	/* keep pointer to vpage */
1848 
1849 			nbytes = vpgtob(npages);
1850 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
1851 			bcopy(&ovpage[dpages], svd->vpage, nbytes);
1852 
1853 			/* free up old vpage */
1854 			kmem_free(ovpage, vpgtob(opages));
1855 		}
1856 		if (amp != NULL) {
1857 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1858 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
1859 				/*
1860 				 * Free up now unused parts of anon_map array.
1861 				 */
1862 				if (amp->a_szc == seg->s_szc) {
1863 					if (seg->s_szc != 0) {
1864 						anon_free_pages(amp->ahp,
1865 						    svd->anon_index, len,
1866 						    seg->s_szc);
1867 					} else {
1868 						anon_free(amp->ahp,
1869 						    svd->anon_index,
1870 						    len);
1871 					}
1872 				} else {
1873 					ASSERT(svd->type == MAP_SHARED);
1874 					ASSERT(amp->a_szc > seg->s_szc);
1875 					anon_shmap_free_pages(amp,
1876 					    svd->anon_index, len);
1877 				}
1878 
1879 				/*
1880 				 * Unreserve swap space for the
1881 				 * unmapped chunk of this segment in
1882 				 * case it's MAP_SHARED
1883 				 */
1884 				if (svd->type == MAP_SHARED) {
1885 					anon_unresv(len);
1886 					amp->swresv -= len;
1887 				}
1888 			}
1889 			ANON_LOCK_EXIT(&amp->a_rwlock);
1890 			svd->anon_index += dpages;
1891 		}
1892 		if (svd->vp != NULL)
1893 			svd->offset += len;
1894 
1895 		if (svd->swresv) {
1896 			if (svd->flags & MAP_NORESERVE) {
1897 				ASSERT(amp);
1898 				oswresv = svd->swresv;
1899 
1900 				svd->swresv = ptob(anon_pages(amp->ahp,
1901 				    svd->anon_index, npages));
1902 				anon_unresv(oswresv - svd->swresv);
1903 			} else {
1904 				anon_unresv(len);
1905 				svd->swresv -= len;
1906 			}
1907 			TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
1908 				seg, len, 0);
1909 		}
1910 
1911 		seg->s_base += len;
1912 		seg->s_size -= len;
1913 		return (0);
1914 	}
1915 
1916 	/*
1917 	 * Check for end of segment
1918 	 */
1919 	if (addr + len == seg->s_base + seg->s_size) {
1920 		if (svd->vpage != NULL) {
1921 			size_t nbytes;
1922 			struct vpage *ovpage;
1923 
1924 			ovpage = svd->vpage;	/* keep pointer to vpage */
1925 
1926 			nbytes = vpgtob(npages);
1927 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
1928 			bcopy(ovpage, svd->vpage, nbytes);
1929 
1930 			/* free up old vpage */
1931 			kmem_free(ovpage, vpgtob(opages));
1932 
1933 		}
1934 		if (amp != NULL) {
1935 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1936 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
1937 				/*
1938 				 * Free up now unused parts of anon_map array.
1939 				 */
1940 				ulong_t an_idx = svd->anon_index + npages;
1941 				if (amp->a_szc == seg->s_szc) {
1942 					if (seg->s_szc != 0) {
1943 						anon_free_pages(amp->ahp,
1944 						    an_idx, len,
1945 						    seg->s_szc);
1946 					} else {
1947 						anon_free(amp->ahp, an_idx,
1948 						    len);
1949 					}
1950 				} else {
1951 					ASSERT(svd->type == MAP_SHARED);
1952 					ASSERT(amp->a_szc > seg->s_szc);
1953 					anon_shmap_free_pages(amp,
1954 					    an_idx, len);
1955 				}
1956 
1957 				/*
1958 				 * Unreserve swap space for the
1959 				 * unmapped chunk of this segment in
1960 				 * case it's MAP_SHARED
1961 				 */
1962 				if (svd->type == MAP_SHARED) {
1963 					anon_unresv(len);
1964 					amp->swresv -= len;
1965 				}
1966 			}
1967 			ANON_LOCK_EXIT(&amp->a_rwlock);
1968 		}
1969 
1970 		if (svd->swresv) {
1971 			if (svd->flags & MAP_NORESERVE) {
1972 				ASSERT(amp);
1973 				oswresv = svd->swresv;
1974 				svd->swresv = ptob(anon_pages(amp->ahp,
1975 					svd->anon_index, npages));
1976 				anon_unresv(oswresv - svd->swresv);
1977 			} else {
1978 				anon_unresv(len);
1979 				svd->swresv -= len;
1980 			}
1981 			TRACE_3(TR_FAC_VM, TR_ANON_PROC,
1982 				"anon proc:%p %lu %u", seg, len, 0);
1983 		}
1984 
1985 		seg->s_size -= len;
1986 		return (0);
1987 	}
1988 
1989 	/*
1990 	 * The section to go is in the middle of the segment,
1991 	 * have to make it into two segments.  nseg is made for
1992 	 * the high end while seg is cut down at the low end.
1993 	 */
1994 	nbase = addr + len;				/* new seg base */
1995 	nsize = (seg->s_base + seg->s_size) - nbase;	/* new seg size */
1996 	seg->s_size = addr - seg->s_base;		/* shrink old seg */
1997 	nseg = seg_alloc(seg->s_as, nbase, nsize);
1998 	if (nseg == NULL) {
1999 		panic("segvn_unmap seg_alloc");
2000 		/*NOTREACHED*/
2001 	}
2002 	nseg->s_ops = seg->s_ops;
2003 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
2004 	nseg->s_data = (void *)nsvd;
2005 	nseg->s_szc = seg->s_szc;
2006 	*nsvd = *svd;
2007 	nsvd->seg = nseg;
2008 	nsvd->offset = svd->offset + (uintptr_t)(nseg->s_base - seg->s_base);
2009 	nsvd->swresv = 0;
2010 	nsvd->softlockcnt = 0;
2011 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
2012 
2013 	if (svd->vp != NULL) {
2014 		VN_HOLD(nsvd->vp);
2015 		if (nsvd->type == MAP_SHARED)
2016 			lgrp_shm_policy_init(NULL, nsvd->vp);
2017 	}
2018 	crhold(svd->cred);
2019 
2020 	if (svd->vpage == NULL) {
2021 		nsvd->vpage = NULL;
2022 	} else {
2023 		/* need to split vpage into two arrays */
2024 		size_t nbytes;
2025 		struct vpage *ovpage;
2026 
2027 		ovpage = svd->vpage;		/* keep pointer to vpage */
2028 
2029 		npages = seg_pages(seg);	/* seg has shrunk */
2030 		nbytes = vpgtob(npages);
2031 		svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2032 
2033 		bcopy(ovpage, svd->vpage, nbytes);
2034 
2035 		npages = seg_pages(nseg);
2036 		nbytes = vpgtob(npages);
2037 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2038 
2039 		bcopy(&ovpage[opages - npages], nsvd->vpage, nbytes);
2040 
2041 		/* free up old vpage */
2042 		kmem_free(ovpage, vpgtob(opages));
2043 	}
2044 
2045 	if (amp == NULL) {
2046 		nsvd->amp = NULL;
2047 		nsvd->anon_index = 0;
2048 	} else {
2049 		/*
2050 		 * Need to create a new anon map for the new segment.
2051 		 * We'll also allocate a new smaller array for the old
2052 		 * smaller segment to save space.
2053 		 */
2054 		opages = btop((uintptr_t)(addr - seg->s_base));
2055 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2056 		if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2057 			/*
2058 			 * Free up now unused parts of anon_map array.
2059 			 */
2060 			ulong_t an_idx = svd->anon_index + opages;
2061 			if (amp->a_szc == seg->s_szc) {
2062 				if (seg->s_szc != 0) {
2063 					anon_free_pages(amp->ahp, an_idx, len,
2064 					    seg->s_szc);
2065 				} else {
2066 					anon_free(amp->ahp, an_idx,
2067 					    len);
2068 				}
2069 			} else {
2070 				ASSERT(svd->type == MAP_SHARED);
2071 				ASSERT(amp->a_szc > seg->s_szc);
2072 				anon_shmap_free_pages(amp, an_idx, len);
2073 			}
2074 
2075 			/*
2076 			 * Unreserve swap space for the
2077 			 * unmapped chunk of this segment in
2078 			 * case it's MAP_SHARED
2079 			 */
2080 			if (svd->type == MAP_SHARED) {
2081 				anon_unresv(len);
2082 				amp->swresv -= len;
2083 			}
2084 		}
2085 		nsvd->anon_index = svd->anon_index +
2086 		    btop((uintptr_t)(nseg->s_base - seg->s_base));
2087 		if (svd->type == MAP_SHARED) {
2088 			amp->refcnt++;
2089 			nsvd->amp = amp;
2090 		} else {
2091 			struct anon_map *namp;
2092 			struct anon_hdr *nahp;
2093 
2094 			ASSERT(svd->type == MAP_PRIVATE);
2095 			nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
2096 			namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
2097 			namp->a_szc = seg->s_szc;
2098 			(void) anon_copy_ptr(amp->ahp, svd->anon_index, nahp,
2099 			    0, btop(seg->s_size), ANON_SLEEP);
2100 			(void) anon_copy_ptr(amp->ahp, nsvd->anon_index,
2101 			    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
2102 			anon_release(amp->ahp, btop(amp->size));
2103 			svd->anon_index = 0;
2104 			nsvd->anon_index = 0;
2105 			amp->ahp = nahp;
2106 			amp->size = seg->s_size;
2107 			nsvd->amp = namp;
2108 		}
2109 		ANON_LOCK_EXIT(&amp->a_rwlock);
2110 	}
2111 	if (svd->swresv) {
2112 		if (svd->flags & MAP_NORESERVE) {
2113 			ASSERT(amp);
2114 			oswresv = svd->swresv;
2115 			svd->swresv = ptob(anon_pages(amp->ahp,
2116 				svd->anon_index, btop(seg->s_size)));
2117 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
2118 				nsvd->anon_index, btop(nseg->s_size)));
2119 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2120 			anon_unresv(oswresv - (svd->swresv + nsvd->swresv));
2121 		} else {
2122 			if (seg->s_size + nseg->s_size + len != svd->swresv) {
2123 				panic("segvn_unmap: "
2124 				    "cannot split swap reservation");
2125 				/*NOTREACHED*/
2126 			}
2127 			anon_unresv(len);
2128 			svd->swresv = seg->s_size;
2129 			nsvd->swresv = nseg->s_size;
2130 		}
2131 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2132 			seg, len, 0);
2133 	}
2134 
2135 	return (0);			/* I'm glad that's all over with! */
2136 }
2137 
2138 static void
2139 segvn_free(struct seg *seg)
2140 {
2141 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2142 	pgcnt_t npages = seg_pages(seg);
2143 	struct anon_map *amp;
2144 	size_t len;
2145 
2146 	/*
2147 	 * We don't need any segment level locks for "segvn" data
2148 	 * since the address space is "write" locked.
2149 	 */
2150 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
2151 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
2152 
2153 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2154 
2155 	/*
2156 	 * Be sure to unlock pages. XXX Why do things get free'ed instead
2157 	 * of unmapped? XXX
2158 	 */
2159 	(void) segvn_lockop(seg, seg->s_base, seg->s_size,
2160 	    0, MC_UNLOCK, NULL, 0);
2161 
2162 	/*
2163 	 * Deallocate the vpage and anon pointers if necessary and possible.
2164 	 */
2165 	if (svd->vpage != NULL) {
2166 		kmem_free(svd->vpage, vpgtob(npages));
2167 		svd->vpage = NULL;
2168 	}
2169 	if ((amp = svd->amp) != NULL) {
2170 		/*
2171 		 * If there are no more references to this anon_map
2172 		 * structure, then deallocate the structure after freeing
2173 		 * up all the anon slot pointers that we can.
2174 		 */
2175 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2176 		ASSERT(amp->a_szc >= seg->s_szc);
2177 		if (--amp->refcnt == 0) {
2178 			if (svd->type == MAP_PRIVATE) {
2179 				/*
2180 				 * Private - we only need to anon_free
2181 				 * the part that this segment refers to.
2182 				 */
2183 				if (seg->s_szc != 0) {
2184 					anon_free_pages(amp->ahp,
2185 					    svd->anon_index, seg->s_size,
2186 					    seg->s_szc);
2187 				} else {
2188 					anon_free(amp->ahp, svd->anon_index,
2189 					    seg->s_size);
2190 				}
2191 			} else {
2192 				/*
2193 				 * Shared - anon_free the entire
2194 				 * anon_map's worth of stuff and
2195 				 * release any swap reservation.
2196 				 */
2197 				if (amp->a_szc != 0) {
2198 					anon_shmap_free_pages(amp, 0,
2199 					    amp->size);
2200 				} else {
2201 					anon_free(amp->ahp, 0, amp->size);
2202 				}
2203 				if ((len = amp->swresv) != 0) {
2204 					anon_unresv(len);
2205 					TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2206 						"anon proc:%p %lu %u",
2207 						seg, len, 0);
2208 				}
2209 			}
2210 			svd->amp = NULL;
2211 			ANON_LOCK_EXIT(&amp->a_rwlock);
2212 			anonmap_free(amp);
2213 		} else if (svd->type == MAP_PRIVATE) {
2214 			/*
2215 			 * We had a private mapping which still has
2216 			 * a held anon_map so just free up all the
2217 			 * anon slot pointers that we were using.
2218 			 */
2219 			if (seg->s_szc != 0) {
2220 				anon_free_pages(amp->ahp, svd->anon_index,
2221 				    seg->s_size, seg->s_szc);
2222 			} else {
2223 				anon_free(amp->ahp, svd->anon_index,
2224 				    seg->s_size);
2225 			}
2226 			ANON_LOCK_EXIT(&amp->a_rwlock);
2227 		} else {
2228 			ANON_LOCK_EXIT(&amp->a_rwlock);
2229 		}
2230 	}
2231 
2232 	/*
2233 	 * Release swap reservation.
2234 	 */
2235 	if ((len = svd->swresv) != 0) {
2236 		anon_unresv(svd->swresv);
2237 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2238 			seg, len, 0);
2239 		svd->swresv = 0;
2240 	}
2241 	/*
2242 	 * Release claim on vnode, credentials, and finally free the
2243 	 * private data.
2244 	 */
2245 	if (svd->vp != NULL) {
2246 		if (svd->type == MAP_SHARED)
2247 			lgrp_shm_policy_fini(NULL, svd->vp);
2248 		VN_RELE(svd->vp);
2249 		svd->vp = NULL;
2250 	}
2251 	crfree(svd->cred);
2252 	svd->cred = NULL;
2253 
2254 	seg->s_data = NULL;
2255 	kmem_cache_free(segvn_cache, svd);
2256 }
2257 
2258 #ifdef DEBUG
2259 uint32_t segvn_slock_mtbf = 0;
2260 #endif
2261 
2262 ulong_t segvn_lpglck_limit = 0;
2263 
2264 /*
2265  * Support routines used by segvn_pagelock() and softlock faults for anonymous
2266  * pages to implement availrmem accounting in a way that makes sure the
2267  * same memory is accounted just once for all softlock/pagelock purposes.
2268  * This prevents a bug when availrmem is quickly incorrectly exhausted from
2269  * several pagelocks to different parts of the same large page since each
2270  * pagelock has to decrement availrmem by the size of the entire large
2271  * page. Note those pages are not COW shared until softunlock/pageunlock so
2272  * we don't need to use cow style accounting here.  We also need to make sure
2273  * the entire large page is accounted even if softlock range is less than the
2274  * entire large page because large anon pages can't be demoted when any of
2275  * constituent pages is locked. The caller calls this routine for every page_t
2276  * it locks. The very first page in the range may not be the root page of a
2277  * large page. For all other pages it's guaranteed we are going to visit the
2278  * root of a particular large page before any other constituent page as we are
2279  * locking sequential pages belonging to the same anon map. So we do all the
2280  * locking when the root is encountered except for the very first page.  Since
2281  * softlocking is not supported (except S_READ_NOCOW special case) for vmpss
2282  * segments and since vnode pages can be demoted without locking all
2283  * constituent pages vnode pages don't come here.  Unlocking relies on the
2284  * fact that pagesize can't change whenever any of constituent large pages is
2285  * locked at least SE_SHARED. This allows unlocking code to find the right
2286  * root and decrement availrmem by the same amount it was incremented when the
2287  * page was locked.
2288  */
2289 static int
2290 segvn_slock_anonpages(page_t *pp, int first)
2291 {
2292 	pgcnt_t		pages;
2293 	pfn_t		pfn;
2294 	uchar_t		szc = pp->p_szc;
2295 
2296 	ASSERT(PAGE_LOCKED(pp));
2297 	ASSERT(pp->p_vnode != NULL);
2298 	ASSERT(IS_SWAPFSVP(pp->p_vnode));
2299 
2300 	/*
2301 	 * pagesize won't change as long as any constituent page is locked.
2302 	 */
2303 	pages = page_get_pagecnt(pp->p_szc);
2304 	pfn = page_pptonum(pp);
2305 
2306 	if (!first) {
2307 		if (!IS_P2ALIGNED(pfn, pages)) {
2308 #ifdef DEBUG
2309 			pp = &pp[-(spgcnt_t)(pfn & (pages - 1))];
2310 			pfn = page_pptonum(pp);
2311 			ASSERT(IS_P2ALIGNED(pfn, pages));
2312 			ASSERT(pp->p_szc == szc);
2313 			ASSERT(pp->p_vnode != NULL);
2314 			ASSERT(IS_SWAPFSVP(pp->p_vnode));
2315 			ASSERT(pp->p_slckcnt != 0);
2316 #endif /* DEBUG */
2317 			return (1);
2318 		}
2319 	} else if (!IS_P2ALIGNED(pfn, pages)) {
2320 		pp = &pp[-(spgcnt_t)(pfn & (pages - 1))];
2321 #ifdef DEBUG
2322 		pfn = page_pptonum(pp);
2323 		ASSERT(IS_P2ALIGNED(pfn, pages));
2324 		ASSERT(pp->p_szc == szc);
2325 		ASSERT(pp->p_vnode != NULL);
2326 		ASSERT(IS_SWAPFSVP(pp->p_vnode));
2327 #endif /* DEBUG */
2328 	}
2329 
2330 #ifdef DEBUG
2331 	if (segvn_slock_mtbf && !(gethrtime() % segvn_slock_mtbf)) {
2332 		return (0);
2333 	}
2334 #endif /* DEBUG */
2335 
2336 	/*
2337 	 * pp is a root page.
2338 	 * We haven't locked this large page yet.
2339 	 */
2340 	page_struct_lock(pp);
2341 	if (pp->p_slckcnt != 0) {
2342 		if (pp->p_slckcnt < PAGE_SLOCK_MAXIMUM) {
2343 			pp->p_slckcnt++;
2344 			page_struct_unlock(pp);
2345 			return (1);
2346 		}
2347 		page_struct_unlock(pp);
2348 		segvn_lpglck_limit++;
2349 		return (0);
2350 	}
2351 	mutex_enter(&freemem_lock);
2352 	if (availrmem < tune.t_minarmem + pages) {
2353 		mutex_exit(&freemem_lock);
2354 		page_struct_unlock(pp);
2355 		return (0);
2356 	}
2357 	pp->p_slckcnt++;
2358 	availrmem -= pages;
2359 	mutex_exit(&freemem_lock);
2360 	page_struct_unlock(pp);
2361 	return (1);
2362 }
2363 
2364 static void
2365 segvn_sunlock_anonpages(page_t *pp, int first)
2366 {
2367 	pgcnt_t		pages;
2368 	pfn_t		pfn;
2369 
2370 	ASSERT(PAGE_LOCKED(pp));
2371 	ASSERT(pp->p_vnode != NULL);
2372 	ASSERT(IS_SWAPFSVP(pp->p_vnode));
2373 
2374 	/*
2375 	 * pagesize won't change as long as any constituent page is locked.
2376 	 */
2377 	pages = page_get_pagecnt(pp->p_szc);
2378 	pfn = page_pptonum(pp);
2379 
2380 	if (!first) {
2381 		if (!IS_P2ALIGNED(pfn, pages)) {
2382 			return;
2383 		}
2384 	} else if (!IS_P2ALIGNED(pfn, pages)) {
2385 		pp = &pp[-(spgcnt_t)(pfn & (pages - 1))];
2386 #ifdef DEBUG
2387 		pfn = page_pptonum(pp);
2388 		ASSERT(IS_P2ALIGNED(pfn, pages));
2389 #endif /* DEBUG */
2390 	}
2391 	ASSERT(pp->p_vnode != NULL);
2392 	ASSERT(IS_SWAPFSVP(pp->p_vnode));
2393 	ASSERT(pp->p_slckcnt != 0);
2394 	page_struct_lock(pp);
2395 	if (--pp->p_slckcnt == 0) {
2396 		mutex_enter(&freemem_lock);
2397 		availrmem += pages;
2398 		mutex_exit(&freemem_lock);
2399 	}
2400 	page_struct_unlock(pp);
2401 }
2402 
2403 /*
2404  * Do a F_SOFTUNLOCK call over the range requested.  The range must have
2405  * already been F_SOFTLOCK'ed.
2406  * Caller must always match addr and len of a softunlock with a previous
2407  * softlock with exactly the same addr and len.
2408  */
2409 static void
2410 segvn_softunlock(struct seg *seg, caddr_t addr, size_t len, enum seg_rw rw)
2411 {
2412 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2413 	page_t *pp;
2414 	caddr_t adr;
2415 	struct vnode *vp;
2416 	u_offset_t offset;
2417 	ulong_t anon_index;
2418 	struct anon_map *amp;
2419 	struct anon *ap = NULL;
2420 
2421 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
2422 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
2423 
2424 	if ((amp = svd->amp) != NULL)
2425 		anon_index = svd->anon_index + seg_page(seg, addr);
2426 
2427 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
2428 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2429 		hat_unlock_region(seg->s_as->a_hat, addr, len, svd->rcookie);
2430 	} else {
2431 		hat_unlock(seg->s_as->a_hat, addr, len);
2432 	}
2433 	for (adr = addr; adr < addr + len; adr += PAGESIZE) {
2434 		if (amp != NULL) {
2435 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2436 			if ((ap = anon_get_ptr(amp->ahp, anon_index++))
2437 								!= NULL) {
2438 				swap_xlate(ap, &vp, &offset);
2439 			} else {
2440 				vp = svd->vp;
2441 				offset = svd->offset +
2442 				    (uintptr_t)(adr - seg->s_base);
2443 			}
2444 			ANON_LOCK_EXIT(&amp->a_rwlock);
2445 		} else {
2446 			vp = svd->vp;
2447 			offset = svd->offset +
2448 			    (uintptr_t)(adr - seg->s_base);
2449 		}
2450 
2451 		/*
2452 		 * Use page_find() instead of page_lookup() to
2453 		 * find the page since we know that it is locked.
2454 		 */
2455 		pp = page_find(vp, offset);
2456 		if (pp == NULL) {
2457 			panic(
2458 			    "segvn_softunlock: addr %p, ap %p, vp %p, off %llx",
2459 			    (void *)adr, (void *)ap, (void *)vp, offset);
2460 			/*NOTREACHED*/
2461 		}
2462 
2463 		if (rw == S_WRITE) {
2464 			hat_setrefmod(pp);
2465 			if (seg->s_as->a_vbits)
2466 				hat_setstat(seg->s_as, adr, PAGESIZE,
2467 				    P_REF | P_MOD);
2468 		} else if (rw != S_OTHER) {
2469 			hat_setref(pp);
2470 			if (seg->s_as->a_vbits)
2471 				hat_setstat(seg->s_as, adr, PAGESIZE, P_REF);
2472 		}
2473 		TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2474 			"segvn_fault:pp %p vp %p offset %llx", pp, vp, offset);
2475 		if (svd->vp == NULL) {
2476 			segvn_sunlock_anonpages(pp, adr == addr);
2477 		}
2478 		page_unlock(pp);
2479 	}
2480 	mutex_enter(&freemem_lock); /* for availrmem */
2481 	if (svd->vp != NULL) {
2482 		availrmem += btop(len);
2483 	}
2484 	segvn_pages_locked -= btop(len);
2485 	svd->softlockcnt -= btop(len);
2486 	mutex_exit(&freemem_lock);
2487 	if (svd->softlockcnt == 0) {
2488 		/*
2489 		 * All SOFTLOCKS are gone. Wakeup any waiting
2490 		 * unmappers so they can try again to unmap.
2491 		 * Check for waiters first without the mutex
2492 		 * held so we don't always grab the mutex on
2493 		 * softunlocks.
2494 		 */
2495 		if (AS_ISUNMAPWAIT(seg->s_as)) {
2496 			mutex_enter(&seg->s_as->a_contents);
2497 			if (AS_ISUNMAPWAIT(seg->s_as)) {
2498 				AS_CLRUNMAPWAIT(seg->s_as);
2499 				cv_broadcast(&seg->s_as->a_cv);
2500 			}
2501 			mutex_exit(&seg->s_as->a_contents);
2502 		}
2503 	}
2504 }
2505 
2506 #define	PAGE_HANDLED	((page_t *)-1)
2507 
2508 /*
2509  * Release all the pages in the NULL terminated ppp list
2510  * which haven't already been converted to PAGE_HANDLED.
2511  */
2512 static void
2513 segvn_pagelist_rele(page_t **ppp)
2514 {
2515 	for (; *ppp != NULL; ppp++) {
2516 		if (*ppp != PAGE_HANDLED)
2517 			page_unlock(*ppp);
2518 	}
2519 }
2520 
2521 static int stealcow = 1;
2522 
2523 /*
2524  * Workaround for viking chip bug.  See bug id 1220902.
2525  * To fix this down in pagefault() would require importing so
2526  * much as and segvn code as to be unmaintainable.
2527  */
2528 int enable_mbit_wa = 0;
2529 
2530 /*
2531  * Handles all the dirty work of getting the right
2532  * anonymous pages and loading up the translations.
2533  * This routine is called only from segvn_fault()
2534  * when looping over the range of addresses requested.
2535  *
2536  * The basic algorithm here is:
2537  * 	If this is an anon_zero case
2538  *		Call anon_zero to allocate page
2539  *		Load up translation
2540  *		Return
2541  *	endif
2542  *	If this is an anon page
2543  *		Use anon_getpage to get the page
2544  *	else
2545  *		Find page in pl[] list passed in
2546  *	endif
2547  *	If not a cow
2548  *		Load up the translation to the page
2549  *		return
2550  *	endif
2551  *	Call anon_private to handle cow
2552  *	Load up (writable) translation to new page
2553  */
2554 static faultcode_t
2555 segvn_faultpage(
2556 	struct hat *hat,		/* the hat to use for mapping */
2557 	struct seg *seg,		/* seg_vn of interest */
2558 	caddr_t addr,			/* address in as */
2559 	u_offset_t off,			/* offset in vp */
2560 	struct vpage *vpage,		/* pointer to vpage for vp, off */
2561 	page_t *pl[],			/* object source page pointer */
2562 	uint_t vpprot,			/* access allowed to object pages */
2563 	enum fault_type type,		/* type of fault */
2564 	enum seg_rw rw,			/* type of access at fault */
2565 	int brkcow,			/* we may need to break cow */
2566 	int first)			/* first page for this fault if 1 */
2567 {
2568 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2569 	page_t *pp, **ppp;
2570 	uint_t pageflags = 0;
2571 	page_t *anon_pl[1 + 1];
2572 	page_t *opp = NULL;		/* original page */
2573 	uint_t prot;
2574 	int err;
2575 	int cow;
2576 	int claim;
2577 	int steal = 0;
2578 	ulong_t anon_index;
2579 	struct anon *ap, *oldap;
2580 	struct anon_map *amp;
2581 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
2582 	int anon_lock = 0;
2583 	anon_sync_obj_t cookie;
2584 
2585 	if (svd->flags & MAP_TEXT) {
2586 		hat_flag |= HAT_LOAD_TEXT;
2587 	}
2588 
2589 	ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
2590 	ASSERT(seg->s_szc == 0);
2591 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
2592 
2593 	/*
2594 	 * Initialize protection value for this page.
2595 	 * If we have per page protection values check it now.
2596 	 */
2597 	if (svd->pageprot) {
2598 		uint_t protchk;
2599 
2600 		switch (rw) {
2601 		case S_READ:
2602 			protchk = PROT_READ;
2603 			break;
2604 		case S_WRITE:
2605 			protchk = PROT_WRITE;
2606 			break;
2607 		case S_EXEC:
2608 			protchk = PROT_EXEC;
2609 			break;
2610 		case S_OTHER:
2611 		default:
2612 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
2613 			break;
2614 		}
2615 
2616 		prot = VPP_PROT(vpage);
2617 		if ((prot & protchk) == 0)
2618 			return (FC_PROT);	/* illegal access type */
2619 	} else {
2620 		prot = svd->prot;
2621 	}
2622 
2623 	if (type == F_SOFTLOCK && svd->vp != NULL) {
2624 		mutex_enter(&freemem_lock);
2625 		if (availrmem <= tune.t_minarmem) {
2626 			mutex_exit(&freemem_lock);
2627 			return (FC_MAKE_ERR(ENOMEM));	/* out of real memory */
2628 		} else {
2629 			availrmem--;
2630 			svd->softlockcnt++;
2631 			segvn_pages_locked++;
2632 		}
2633 		mutex_exit(&freemem_lock);
2634 	}
2635 
2636 	/*
2637 	 * Always acquire the anon array lock to prevent 2 threads from
2638 	 * allocating separate anon slots for the same "addr".
2639 	 */
2640 
2641 	if ((amp = svd->amp) != NULL) {
2642 		ASSERT(RW_READ_HELD(&amp->a_rwlock));
2643 		anon_index = svd->anon_index + seg_page(seg, addr);
2644 		anon_array_enter(amp, anon_index, &cookie);
2645 		anon_lock = 1;
2646 	}
2647 
2648 	if (svd->vp == NULL && amp != NULL) {
2649 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL) {
2650 			/*
2651 			 * Allocate a (normally) writable anonymous page of
2652 			 * zeroes. If no advance reservations, reserve now.
2653 			 */
2654 			if (svd->flags & MAP_NORESERVE) {
2655 				if (anon_resv_zone(ptob(1),
2656 				    seg->s_as->a_proc->p_zone)) {
2657 					atomic_add_long(&svd->swresv, ptob(1));
2658 				} else {
2659 					err = ENOMEM;
2660 					goto out;
2661 				}
2662 			}
2663 			if ((pp = anon_zero(seg, addr, &ap,
2664 			    svd->cred)) == NULL) {
2665 				err = ENOMEM;
2666 				goto out;	/* out of swap space */
2667 			}
2668 			/*
2669 			 * Re-acquire the anon_map lock and
2670 			 * initialize the anon array entry.
2671 			 */
2672 			(void) anon_set_ptr(amp->ahp, anon_index, ap,
2673 				ANON_SLEEP);
2674 
2675 			ASSERT(pp->p_szc == 0);
2676 
2677 			/*
2678 			 * Handle pages that have been marked for migration
2679 			 */
2680 			if (lgrp_optimizations())
2681 				page_migrate(seg, addr, &pp, 1);
2682 
2683 			if (type == F_SOFTLOCK) {
2684 				if (!segvn_slock_anonpages(pp, first)) {
2685 					page_unlock(pp);
2686 					err = ENOMEM;
2687 					goto out;
2688 				} else {
2689 					mutex_enter(&freemem_lock);
2690 					svd->softlockcnt++;
2691 					segvn_pages_locked++;
2692 					mutex_exit(&freemem_lock);
2693 				}
2694 			}
2695 
2696 			if (enable_mbit_wa) {
2697 				if (rw == S_WRITE)
2698 					hat_setmod(pp);
2699 				else if (!hat_ismod(pp))
2700 					prot &= ~PROT_WRITE;
2701 			}
2702 			/*
2703 			 * If AS_PAGLCK is set in a_flags (via memcntl(2)
2704 			 * with MC_LOCKAS, MCL_FUTURE) and this is a
2705 			 * MAP_NORESERVE segment, we may need to
2706 			 * permanently lock the page as it is being faulted
2707 			 * for the first time. The following text applies
2708 			 * only to MAP_NORESERVE segments:
2709 			 *
2710 			 * As per memcntl(2), if this segment was created
2711 			 * after MCL_FUTURE was applied (a "future"
2712 			 * segment), its pages must be locked.  If this
2713 			 * segment existed at MCL_FUTURE application (a
2714 			 * "past" segment), the interface is unclear.
2715 			 *
2716 			 * We decide to lock only if vpage is present:
2717 			 *
2718 			 * - "future" segments will have a vpage array (see
2719 			 *    as_map), and so will be locked as required
2720 			 *
2721 			 * - "past" segments may not have a vpage array,
2722 			 *    depending on whether events (such as
2723 			 *    mprotect) have occurred. Locking if vpage
2724 			 *    exists will preserve legacy behavior.  Not
2725 			 *    locking if vpage is absent, will not break
2726 			 *    the interface or legacy behavior.  Note that
2727 			 *    allocating vpage here if it's absent requires
2728 			 *    upgrading the segvn reader lock, the cost of
2729 			 *    which does not seem worthwhile.
2730 			 *
2731 			 * Usually testing and setting VPP_ISPPLOCK and
2732 			 * VPP_SETPPLOCK requires holding the segvn lock as
2733 			 * writer, but in this case all readers are
2734 			 * serializing on the anon array lock.
2735 			 */
2736 			if (AS_ISPGLCK(seg->s_as) && vpage != NULL &&
2737 			    (svd->flags & MAP_NORESERVE) &&
2738 			    !VPP_ISPPLOCK(vpage)) {
2739 				proc_t *p = seg->s_as->a_proc;
2740 				ASSERT(svd->type == MAP_PRIVATE);
2741 				mutex_enter(&p->p_lock);
2742 				if (rctl_incr_locked_mem(p, NULL, PAGESIZE,
2743 				    1) == 0) {
2744 					claim = VPP_PROT(vpage) & PROT_WRITE;
2745 					if (page_pp_lock(pp, claim, 0)) {
2746 						VPP_SETPPLOCK(vpage);
2747 					} else {
2748 						rctl_decr_locked_mem(p, NULL,
2749 						    PAGESIZE, 1);
2750 					}
2751 				}
2752 				mutex_exit(&p->p_lock);
2753 			}
2754 
2755 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2756 			hat_memload(hat, addr, pp, prot, hat_flag);
2757 
2758 			if (!(hat_flag & HAT_LOAD_LOCK))
2759 				page_unlock(pp);
2760 
2761 			anon_array_exit(&cookie);
2762 			return (0);
2763 		}
2764 	}
2765 
2766 	/*
2767 	 * Obtain the page structure via anon_getpage() if it is
2768 	 * a private copy of an object (the result of a previous
2769 	 * copy-on-write).
2770 	 */
2771 	if (amp != NULL) {
2772 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) != NULL) {
2773 			err = anon_getpage(&ap, &vpprot, anon_pl, PAGESIZE,
2774 			    seg, addr, rw, svd->cred);
2775 			if (err)
2776 				goto out;
2777 
2778 			if (svd->type == MAP_SHARED) {
2779 				/*
2780 				 * If this is a shared mapping to an
2781 				 * anon_map, then ignore the write
2782 				 * permissions returned by anon_getpage().
2783 				 * They apply to the private mappings
2784 				 * of this anon_map.
2785 				 */
2786 				vpprot |= PROT_WRITE;
2787 			}
2788 			opp = anon_pl[0];
2789 		}
2790 	}
2791 
2792 	/*
2793 	 * Search the pl[] list passed in if it is from the
2794 	 * original object (i.e., not a private copy).
2795 	 */
2796 	if (opp == NULL) {
2797 		/*
2798 		 * Find original page.  We must be bringing it in
2799 		 * from the list in pl[].
2800 		 */
2801 		for (ppp = pl; (opp = *ppp) != NULL; ppp++) {
2802 			if (opp == PAGE_HANDLED)
2803 				continue;
2804 			ASSERT(opp->p_vnode == svd->vp); /* XXX */
2805 			if (opp->p_offset == off)
2806 				break;
2807 		}
2808 		if (opp == NULL) {
2809 			panic("segvn_faultpage not found");
2810 			/*NOTREACHED*/
2811 		}
2812 		*ppp = PAGE_HANDLED;
2813 
2814 	}
2815 
2816 	ASSERT(PAGE_LOCKED(opp));
2817 
2818 	TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2819 		"segvn_fault:pp %p vp %p offset %llx",
2820 		opp, NULL, 0);
2821 
2822 	/*
2823 	 * The fault is treated as a copy-on-write fault if a
2824 	 * write occurs on a private segment and the object
2825 	 * page (i.e., mapping) is write protected.  We assume
2826 	 * that fatal protection checks have already been made.
2827 	 */
2828 
2829 	if (brkcow) {
2830 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2831 		cow = !(vpprot & PROT_WRITE);
2832 	} else if (svd->tr_state == SEGVN_TR_ON) {
2833 		/*
2834 		 * If we are doing text replication COW on first touch.
2835 		 */
2836 		ASSERT(amp != NULL);
2837 		ASSERT(svd->vp != NULL);
2838 		ASSERT(rw != S_WRITE);
2839 		cow = (ap == NULL);
2840 	} else {
2841 		cow = 0;
2842 	}
2843 
2844 	/*
2845 	 * If not a copy-on-write case load the translation
2846 	 * and return.
2847 	 */
2848 	if (cow == 0) {
2849 
2850 		/*
2851 		 * Handle pages that have been marked for migration
2852 		 */
2853 		if (lgrp_optimizations())
2854 			page_migrate(seg, addr, &opp, 1);
2855 
2856 		if (type == F_SOFTLOCK && svd->vp == NULL) {
2857 
2858 			ASSERT(opp->p_szc == 0 ||
2859 			    (svd->type == MAP_SHARED &&
2860 				amp != NULL && amp->a_szc != 0));
2861 
2862 			if (!segvn_slock_anonpages(opp, first)) {
2863 				page_unlock(opp);
2864 				err = ENOMEM;
2865 				goto out;
2866 			} else {
2867 				mutex_enter(&freemem_lock);
2868 				svd->softlockcnt++;
2869 				segvn_pages_locked++;
2870 				mutex_exit(&freemem_lock);
2871 			}
2872 		}
2873 		if (IS_VMODSORT(opp->p_vnode) || enable_mbit_wa) {
2874 			if (rw == S_WRITE)
2875 				hat_setmod(opp);
2876 			else if (rw != S_OTHER && !hat_ismod(opp))
2877 				prot &= ~PROT_WRITE;
2878 		}
2879 
2880 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
2881 		    (!svd->pageprot && svd->prot == (prot & vpprot)));
2882 		ASSERT(amp == NULL ||
2883 		    svd->rcookie == HAT_INVALID_REGION_COOKIE);
2884 		hat_memload_region(hat, addr, opp, prot & vpprot, hat_flag,
2885 		    svd->rcookie);
2886 
2887 		if (!(hat_flag & HAT_LOAD_LOCK))
2888 			page_unlock(opp);
2889 
2890 		if (anon_lock) {
2891 			anon_array_exit(&cookie);
2892 		}
2893 		return (0);
2894 	}
2895 
2896 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2897 
2898 	hat_setref(opp);
2899 
2900 	ASSERT(amp != NULL && anon_lock);
2901 
2902 	/*
2903 	 * Steal the page only if it isn't a private page
2904 	 * since stealing a private page is not worth the effort.
2905 	 */
2906 	if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL)
2907 		steal = 1;
2908 
2909 	/*
2910 	 * Steal the original page if the following conditions are true:
2911 	 *
2912 	 * We are low on memory, the page is not private, page is not large,
2913 	 * not shared, not modified, not `locked' or if we have it `locked'
2914 	 * (i.e., p_cowcnt == 1 and p_lckcnt == 0, which also implies
2915 	 * that the page is not shared) and if it doesn't have any
2916 	 * translations. page_struct_lock isn't needed to look at p_cowcnt
2917 	 * and p_lckcnt because we first get exclusive lock on page.
2918 	 */
2919 	(void) hat_pagesync(opp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD);
2920 
2921 	if (stealcow && freemem < minfree && steal && opp->p_szc == 0 &&
2922 	    page_tryupgrade(opp) && !hat_ismod(opp) &&
2923 	    ((opp->p_lckcnt == 0 && opp->p_cowcnt == 0) ||
2924 	    (opp->p_lckcnt == 0 && opp->p_cowcnt == 1 &&
2925 	    vpage != NULL && VPP_ISPPLOCK(vpage)))) {
2926 		/*
2927 		 * Check if this page has other translations
2928 		 * after unloading our translation.
2929 		 */
2930 		if (hat_page_is_mapped(opp)) {
2931 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2932 			hat_unload(seg->s_as->a_hat, addr, PAGESIZE,
2933 				HAT_UNLOAD);
2934 		}
2935 
2936 		/*
2937 		 * hat_unload() might sync back someone else's recent
2938 		 * modification, so check again.
2939 		 */
2940 		if (!hat_ismod(opp) && !hat_page_is_mapped(opp))
2941 			pageflags |= STEAL_PAGE;
2942 	}
2943 
2944 	/*
2945 	 * If we have a vpage pointer, see if it indicates that we have
2946 	 * ``locked'' the page we map -- if so, tell anon_private to
2947 	 * transfer the locking resource to the new page.
2948 	 *
2949 	 * See Statement at the beginning of segvn_lockop regarding
2950 	 * the way lockcnts/cowcnts are handled during COW.
2951 	 *
2952 	 */
2953 	if (vpage != NULL && VPP_ISPPLOCK(vpage))
2954 		pageflags |= LOCK_PAGE;
2955 
2956 	/*
2957 	 * Allocate a private page and perform the copy.
2958 	 * For MAP_NORESERVE reserve swap space now, unless this
2959 	 * is a cow fault on an existing anon page in which case
2960 	 * MAP_NORESERVE will have made advance reservations.
2961 	 */
2962 	if ((svd->flags & MAP_NORESERVE) && (ap == NULL)) {
2963 		if (anon_resv_zone(ptob(1), seg->s_as->a_proc->p_zone)) {
2964 			atomic_add_long(&svd->swresv, ptob(1));
2965 		} else {
2966 			page_unlock(opp);
2967 			err = ENOMEM;
2968 			goto out;
2969 		}
2970 	}
2971 	oldap = ap;
2972 	pp = anon_private(&ap, seg, addr, prot, opp, pageflags, svd->cred);
2973 	if (pp == NULL) {
2974 		err = ENOMEM;	/* out of swap space */
2975 		goto out;
2976 	}
2977 
2978 	/*
2979 	 * If we copied away from an anonymous page, then
2980 	 * we are one step closer to freeing up an anon slot.
2981 	 *
2982 	 * NOTE:  The original anon slot must be released while
2983 	 * holding the "anon_map" lock.  This is necessary to prevent
2984 	 * other threads from obtaining a pointer to the anon slot
2985 	 * which may be freed if its "refcnt" is 1.
2986 	 */
2987 	if (oldap != NULL)
2988 		anon_decref(oldap);
2989 
2990 	(void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
2991 
2992 	/*
2993 	 * Handle pages that have been marked for migration
2994 	 */
2995 	if (lgrp_optimizations())
2996 		page_migrate(seg, addr, &pp, 1);
2997 
2998 	ASSERT(pp->p_szc == 0);
2999 	if (type == F_SOFTLOCK && svd->vp == NULL) {
3000 		if (!segvn_slock_anonpages(pp, first)) {
3001 			page_unlock(pp);
3002 			err = ENOMEM;
3003 			goto out;
3004 		} else {
3005 			mutex_enter(&freemem_lock);
3006 			svd->softlockcnt++;
3007 			segvn_pages_locked++;
3008 			mutex_exit(&freemem_lock);
3009 		}
3010 	}
3011 
3012 	ASSERT(!IS_VMODSORT(pp->p_vnode));
3013 	if (enable_mbit_wa) {
3014 		if (rw == S_WRITE)
3015 			hat_setmod(pp);
3016 		else if (!hat_ismod(pp))
3017 			prot &= ~PROT_WRITE;
3018 	}
3019 
3020 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3021 	hat_memload(hat, addr, pp, prot, hat_flag);
3022 
3023 	if (!(hat_flag & HAT_LOAD_LOCK))
3024 		page_unlock(pp);
3025 
3026 	ASSERT(anon_lock);
3027 	anon_array_exit(&cookie);
3028 	return (0);
3029 out:
3030 	if (anon_lock)
3031 		anon_array_exit(&cookie);
3032 
3033 	if (type == F_SOFTLOCK && svd->vp != NULL) {
3034 		mutex_enter(&freemem_lock);
3035 		availrmem++;
3036 		segvn_pages_locked--;
3037 		svd->softlockcnt--;
3038 		mutex_exit(&freemem_lock);
3039 	}
3040 	return (FC_MAKE_ERR(err));
3041 }
3042 
3043 /*
3044  * relocate a bunch of smaller targ pages into one large repl page. all targ
3045  * pages must be complete pages smaller than replacement pages.
3046  * it's assumed that no page's szc can change since they are all PAGESIZE or
3047  * complete large pages locked SHARED.
3048  */
3049 static void
3050 segvn_relocate_pages(page_t **targ, page_t *replacement)
3051 {
3052 	page_t *pp;
3053 	pgcnt_t repl_npgs, curnpgs;
3054 	pgcnt_t i;
3055 	uint_t repl_szc = replacement->p_szc;
3056 	page_t *first_repl = replacement;
3057 	page_t *repl;
3058 	spgcnt_t npgs;
3059 
3060 	VM_STAT_ADD(segvnvmstats.relocatepages[0]);
3061 
3062 	ASSERT(repl_szc != 0);
3063 	npgs = repl_npgs = page_get_pagecnt(repl_szc);
3064 
3065 	i = 0;
3066 	while (repl_npgs) {
3067 		spgcnt_t nreloc;
3068 		int err;
3069 		ASSERT(replacement != NULL);
3070 		pp = targ[i];
3071 		ASSERT(pp->p_szc < repl_szc);
3072 		ASSERT(PAGE_EXCL(pp));
3073 		ASSERT(!PP_ISFREE(pp));
3074 		curnpgs = page_get_pagecnt(pp->p_szc);
3075 		if (curnpgs == 1) {
3076 			VM_STAT_ADD(segvnvmstats.relocatepages[1]);
3077 			repl = replacement;
3078 			page_sub(&replacement, repl);
3079 			ASSERT(PAGE_EXCL(repl));
3080 			ASSERT(!PP_ISFREE(repl));
3081 			ASSERT(repl->p_szc == repl_szc);
3082 		} else {
3083 			page_t *repl_savepp;
3084 			int j;
3085 			VM_STAT_ADD(segvnvmstats.relocatepages[2]);
3086 			repl_savepp = replacement;
3087 			for (j = 0; j < curnpgs; j++) {
3088 				repl = replacement;
3089 				page_sub(&replacement, repl);
3090 				ASSERT(PAGE_EXCL(repl));
3091 				ASSERT(!PP_ISFREE(repl));
3092 				ASSERT(repl->p_szc == repl_szc);
3093 				ASSERT(page_pptonum(targ[i + j]) ==
3094 				    page_pptonum(targ[i]) + j);
3095 			}
3096 			repl = repl_savepp;
3097 			ASSERT(IS_P2ALIGNED(page_pptonum(repl), curnpgs));
3098 		}
3099 		err = page_relocate(&pp, &repl, 0, 1, &nreloc, NULL);
3100 		if (err || nreloc != curnpgs) {
3101 			panic("segvn_relocate_pages: "
3102 			    "page_relocate failed err=%d curnpgs=%ld "
3103 			    "nreloc=%ld", err, curnpgs, nreloc);
3104 		}
3105 		ASSERT(curnpgs <= repl_npgs);
3106 		repl_npgs -= curnpgs;
3107 		i += curnpgs;
3108 	}
3109 	ASSERT(replacement == NULL);
3110 
3111 	repl = first_repl;
3112 	repl_npgs = npgs;
3113 	for (i = 0; i < repl_npgs; i++) {
3114 		ASSERT(PAGE_EXCL(repl));
3115 		ASSERT(!PP_ISFREE(repl));
3116 		targ[i] = repl;
3117 		page_downgrade(targ[i]);
3118 		repl++;
3119 	}
3120 }
3121 
3122 /*
3123  * Check if all pages in ppa array are complete smaller than szc pages and
3124  * their roots will still be aligned relative to their current size if the
3125  * entire ppa array is relocated into one szc page. If these conditions are
3126  * not met return 0.
3127  *
3128  * If all pages are properly aligned attempt to upgrade their locks
3129  * to exclusive mode. If it fails set *upgrdfail to 1 and return 0.
3130  * upgrdfail was set to 0 by caller.
3131  *
3132  * Return 1 if all pages are aligned and locked exclusively.
3133  *
3134  * If all pages in ppa array happen to be physically contiguous to make one
3135  * szc page and all exclusive locks are successfully obtained promote the page
3136  * size to szc and set *pszc to szc. Return 1 with pages locked shared.
3137  */
3138 static int
3139 segvn_full_szcpages(page_t **ppa, uint_t szc, int *upgrdfail, uint_t *pszc)
3140 {
3141 	page_t *pp;
3142 	pfn_t pfn;
3143 	pgcnt_t totnpgs = page_get_pagecnt(szc);
3144 	pfn_t first_pfn;
3145 	int contig = 1;
3146 	pgcnt_t i;
3147 	pgcnt_t j;
3148 	uint_t curszc;
3149 	pgcnt_t curnpgs;
3150 	int root = 0;
3151 
3152 	ASSERT(szc > 0);
3153 
3154 	VM_STAT_ADD(segvnvmstats.fullszcpages[0]);
3155 
3156 	for (i = 0; i < totnpgs; i++) {
3157 		pp = ppa[i];
3158 		ASSERT(PAGE_SHARED(pp));
3159 		ASSERT(!PP_ISFREE(pp));
3160 		pfn = page_pptonum(pp);
3161 		if (i == 0) {
3162 			if (!IS_P2ALIGNED(pfn, totnpgs)) {
3163 				contig = 0;
3164 			} else {
3165 				first_pfn = pfn;
3166 			}
3167 		} else if (contig && pfn != first_pfn + i) {
3168 			contig = 0;
3169 		}
3170 		if (pp->p_szc == 0) {
3171 			if (root) {
3172 				VM_STAT_ADD(segvnvmstats.fullszcpages[1]);
3173 				return (0);
3174 			}
3175 		} else if (!root) {
3176 			if ((curszc = pp->p_szc) >= szc) {
3177 				VM_STAT_ADD(segvnvmstats.fullszcpages[2]);
3178 				return (0);
3179 			}
3180 			if (curszc == 0) {
3181 				/*
3182 				 * p_szc changed means we don't have all pages
3183 				 * locked. return failure.
3184 				 */
3185 				VM_STAT_ADD(segvnvmstats.fullszcpages[3]);
3186 				return (0);
3187 			}
3188 			curnpgs = page_get_pagecnt(curszc);
3189 			if (!IS_P2ALIGNED(pfn, curnpgs) ||
3190 			    !IS_P2ALIGNED(i, curnpgs)) {
3191 				VM_STAT_ADD(segvnvmstats.fullszcpages[4]);
3192 				return (0);
3193 			}
3194 			root = 1;
3195 		} else {
3196 			ASSERT(i > 0);
3197 			VM_STAT_ADD(segvnvmstats.fullszcpages[5]);
3198 			if (pp->p_szc != curszc) {
3199 				VM_STAT_ADD(segvnvmstats.fullszcpages[6]);
3200 				return (0);
3201 			}
3202 			if (pfn - 1 != page_pptonum(ppa[i - 1])) {
3203 				panic("segvn_full_szcpages: "
3204 				    "large page not physically contiguous");
3205 			}
3206 			if (P2PHASE(pfn, curnpgs) == curnpgs - 1) {
3207 				root = 0;
3208 			}
3209 		}
3210 	}
3211 
3212 	for (i = 0; i < totnpgs; i++) {
3213 		ASSERT(ppa[i]->p_szc < szc);
3214 		if (!page_tryupgrade(ppa[i])) {
3215 			for (j = 0; j < i; j++) {
3216 				page_downgrade(ppa[j]);
3217 			}
3218 			*pszc = ppa[i]->p_szc;
3219 			*upgrdfail = 1;
3220 			VM_STAT_ADD(segvnvmstats.fullszcpages[7]);
3221 			return (0);
3222 		}
3223 	}
3224 
3225 	/*
3226 	 * When a page is put a free cachelist its szc is set to 0.  if file
3227 	 * system reclaimed pages from cachelist targ pages will be physically
3228 	 * contiguous with 0 p_szc.  in this case just upgrade szc of targ
3229 	 * pages without any relocations.
3230 	 * To avoid any hat issues with previous small mappings
3231 	 * hat_pageunload() the target pages first.
3232 	 */
3233 	if (contig) {
3234 		VM_STAT_ADD(segvnvmstats.fullszcpages[8]);
3235 		for (i = 0; i < totnpgs; i++) {
3236 			(void) hat_pageunload(ppa[i], HAT_FORCE_PGUNLOAD);
3237 		}
3238 		for (i = 0; i < totnpgs; i++) {
3239 			ppa[i]->p_szc = szc;
3240 		}
3241 		for (i = 0; i < totnpgs; i++) {
3242 			ASSERT(PAGE_EXCL(ppa[i]));
3243 			page_downgrade(ppa[i]);
3244 		}
3245 		if (pszc != NULL) {
3246 			*pszc = szc;
3247 		}
3248 	}
3249 	VM_STAT_ADD(segvnvmstats.fullszcpages[9]);
3250 	return (1);
3251 }
3252 
3253 /*
3254  * Create physically contiguous pages for [vp, off] - [vp, off +
3255  * page_size(szc)) range and for private segment return them in ppa array.
3256  * Pages are created either via IO or relocations.
3257  *
3258  * Return 1 on success and 0 on failure.
3259  *
3260  * If physically contiguous pages already exist for this range return 1 without
3261  * filling ppa array. Caller initializes ppa[0] as NULL to detect that ppa
3262  * array wasn't filled. In this case caller fills ppa array via VOP_GETPAGE().
3263  */
3264 
3265 static int
3266 segvn_fill_vp_pages(struct segvn_data *svd, vnode_t *vp, u_offset_t off,
3267     uint_t szc, page_t **ppa, page_t **ppplist, uint_t *ret_pszc,
3268     int *downsize)
3269 
3270 {
3271 	page_t *pplist = *ppplist;
3272 	size_t pgsz = page_get_pagesize(szc);
3273 	pgcnt_t pages = btop(pgsz);
3274 	ulong_t start_off = off;
3275 	u_offset_t eoff = off + pgsz;
3276 	spgcnt_t nreloc;
3277 	u_offset_t io_off = off;
3278 	size_t io_len;
3279 	page_t *io_pplist = NULL;
3280 	page_t *done_pplist = NULL;
3281 	pgcnt_t pgidx = 0;
3282 	page_t *pp;
3283 	page_t *newpp;
3284 	page_t *targpp;
3285 	int io_err = 0;
3286 	int i;
3287 	pfn_t pfn;
3288 	ulong_t ppages;
3289 	page_t *targ_pplist = NULL;
3290 	page_t *repl_pplist = NULL;
3291 	page_t *tmp_pplist;
3292 	int nios = 0;
3293 	uint_t pszc;
3294 	struct vattr va;
3295 
3296 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[0]);
3297 
3298 	ASSERT(szc != 0);
3299 	ASSERT(pplist->p_szc == szc);
3300 
3301 	/*
3302 	 * downsize will be set to 1 only if we fail to lock pages. this will
3303 	 * allow subsequent faults to try to relocate the page again. If we
3304 	 * fail due to misalignment don't downsize and let the caller map the
3305 	 * whole region with small mappings to avoid more faults into the area
3306 	 * where we can't get large pages anyway.
3307 	 */
3308 	*downsize = 0;
3309 
3310 	while (off < eoff) {
3311 		newpp = pplist;
3312 		ASSERT(newpp != NULL);
3313 		ASSERT(PAGE_EXCL(newpp));
3314 		ASSERT(!PP_ISFREE(newpp));
3315 		/*
3316 		 * we pass NULL for nrelocp to page_lookup_create()
3317 		 * so that it doesn't relocate. We relocate here
3318 		 * later only after we make sure we can lock all
3319 		 * pages in the range we handle and they are all
3320 		 * aligned.
3321 		 */
3322 		pp = page_lookup_create(vp, off, SE_SHARED, newpp, NULL, 0);
3323 		ASSERT(pp != NULL);
3324 		ASSERT(!PP_ISFREE(pp));
3325 		ASSERT(pp->p_vnode == vp);
3326 		ASSERT(pp->p_offset == off);
3327 		if (pp == newpp) {
3328 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[1]);
3329 			page_sub(&pplist, pp);
3330 			ASSERT(PAGE_EXCL(pp));
3331 			ASSERT(page_iolock_assert(pp));
3332 			page_list_concat(&io_pplist, &pp);
3333 			off += PAGESIZE;
3334 			continue;
3335 		}
3336 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[2]);
3337 		pfn = page_pptonum(pp);
3338 		pszc = pp->p_szc;
3339 		if (pszc >= szc && targ_pplist == NULL && io_pplist == NULL &&
3340 		    IS_P2ALIGNED(pfn, pages)) {
3341 			ASSERT(repl_pplist == NULL);
3342 			ASSERT(done_pplist == NULL);
3343 			ASSERT(pplist == *ppplist);
3344 			page_unlock(pp);
3345 			page_free_replacement_page(pplist);
3346 			page_create_putback(pages);
3347 			*ppplist = NULL;
3348 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[3]);
3349 			return (1);
3350 		}
3351 		if (pszc >= szc) {
3352 			page_unlock(pp);
3353 			segvn_faultvnmpss_align_err1++;
3354 			goto out;
3355 		}
3356 		ppages = page_get_pagecnt(pszc);
3357 		if (!IS_P2ALIGNED(pfn, ppages)) {
3358 			ASSERT(pszc > 0);
3359 			/*
3360 			 * sizing down to pszc won't help.
3361 			 */
3362 			page_unlock(pp);
3363 			segvn_faultvnmpss_align_err2++;
3364 			goto out;
3365 		}
3366 		pfn = page_pptonum(newpp);
3367 		if (!IS_P2ALIGNED(pfn, ppages)) {
3368 			ASSERT(pszc > 0);
3369 			/*
3370 			 * sizing down to pszc won't help.
3371 			 */
3372 			page_unlock(pp);
3373 			segvn_faultvnmpss_align_err3++;
3374 			goto out;
3375 		}
3376 		if (!PAGE_EXCL(pp)) {
3377 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[4]);
3378 			page_unlock(pp);
3379 			*downsize = 1;
3380 			*ret_pszc = pp->p_szc;
3381 			goto out;
3382 		}
3383 		targpp = pp;
3384 		if (io_pplist != NULL) {
3385 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[5]);
3386 			io_len = off - io_off;
3387 			/*
3388 			 * Some file systems like NFS don't check EOF
3389 			 * conditions in VOP_PAGEIO(). Check it here
3390 			 * now that pages are locked SE_EXCL. Any file
3391 			 * truncation will wait until the pages are
3392 			 * unlocked so no need to worry that file will
3393 			 * be truncated after we check its size here.
3394 			 * XXX fix NFS to remove this check.
3395 			 */
3396 			va.va_mask = AT_SIZE;
3397 			if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL)) {
3398 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[6]);
3399 				page_unlock(targpp);
3400 				goto out;
3401 			}
3402 			if (btopr(va.va_size) < btopr(io_off + io_len)) {
3403 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[7]);
3404 				*downsize = 1;
3405 				*ret_pszc = 0;
3406 				page_unlock(targpp);
3407 				goto out;
3408 			}
3409 			io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3410 				B_READ, svd->cred, NULL);
3411 			if (io_err) {
3412 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[8]);
3413 				page_unlock(targpp);
3414 				if (io_err == EDEADLK) {
3415 					segvn_vmpss_pageio_deadlk_err++;
3416 				}
3417 				goto out;
3418 			}
3419 			nios++;
3420 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[9]);
3421 			while (io_pplist != NULL) {
3422 				pp = io_pplist;
3423 				page_sub(&io_pplist, pp);
3424 				ASSERT(page_iolock_assert(pp));
3425 				page_io_unlock(pp);
3426 				pgidx = (pp->p_offset - start_off) >>
3427 				    PAGESHIFT;
3428 				ASSERT(pgidx < pages);
3429 				ppa[pgidx] = pp;
3430 				page_list_concat(&done_pplist, &pp);
3431 			}
3432 		}
3433 		pp = targpp;
3434 		ASSERT(PAGE_EXCL(pp));
3435 		ASSERT(pp->p_szc <= pszc);
3436 		if (pszc != 0 && !group_page_trylock(pp, SE_EXCL)) {
3437 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[10]);
3438 			page_unlock(pp);
3439 			*downsize = 1;
3440 			*ret_pszc = pp->p_szc;
3441 			goto out;
3442 		}
3443 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[11]);
3444 		/*
3445 		 * page szc chould have changed before the entire group was
3446 		 * locked. reread page szc.
3447 		 */
3448 		pszc = pp->p_szc;
3449 		ppages = page_get_pagecnt(pszc);
3450 
3451 		/* link just the roots */
3452 		page_list_concat(&targ_pplist, &pp);
3453 		page_sub(&pplist, newpp);
3454 		page_list_concat(&repl_pplist, &newpp);
3455 		off += PAGESIZE;
3456 		while (--ppages != 0) {
3457 			newpp = pplist;
3458 			page_sub(&pplist, newpp);
3459 			off += PAGESIZE;
3460 		}
3461 		io_off = off;
3462 	}
3463 	if (io_pplist != NULL) {
3464 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[12]);
3465 		io_len = eoff - io_off;
3466 		va.va_mask = AT_SIZE;
3467 		if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL) != 0) {
3468 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[13]);
3469 			goto out;
3470 		}
3471 		if (btopr(va.va_size) < btopr(io_off + io_len)) {
3472 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[14]);
3473 			*downsize = 1;
3474 			*ret_pszc = 0;
3475 			goto out;
3476 		}
3477 		io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3478 		    B_READ, svd->cred, NULL);
3479 		if (io_err) {
3480 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[15]);
3481 			if (io_err == EDEADLK) {
3482 				segvn_vmpss_pageio_deadlk_err++;
3483 			}
3484 			goto out;
3485 		}
3486 		nios++;
3487 		while (io_pplist != NULL) {
3488 			pp = io_pplist;
3489 			page_sub(&io_pplist, pp);
3490 			ASSERT(page_iolock_assert(pp));
3491 			page_io_unlock(pp);
3492 			pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3493 			ASSERT(pgidx < pages);
3494 			ppa[pgidx] = pp;
3495 		}
3496 	}
3497 	/*
3498 	 * we're now bound to succeed or panic.
3499 	 * remove pages from done_pplist. it's not needed anymore.
3500 	 */
3501 	while (done_pplist != NULL) {
3502 		pp = done_pplist;
3503 		page_sub(&done_pplist, pp);
3504 	}
3505 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[16]);
3506 	ASSERT(pplist == NULL);
3507 	*ppplist = NULL;
3508 	while (targ_pplist != NULL) {
3509 		int ret;
3510 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[17]);
3511 		ASSERT(repl_pplist);
3512 		pp = targ_pplist;
3513 		page_sub(&targ_pplist, pp);
3514 		pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3515 		newpp = repl_pplist;
3516 		page_sub(&repl_pplist, newpp);
3517 #ifdef DEBUG
3518 		pfn = page_pptonum(pp);
3519 		pszc = pp->p_szc;
3520 		ppages = page_get_pagecnt(pszc);
3521 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3522 		pfn = page_pptonum(newpp);
3523 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3524 		ASSERT(P2PHASE(pfn, pages) == pgidx);
3525 #endif
3526 		nreloc = 0;
3527 		ret = page_relocate(&pp, &newpp, 0, 1, &nreloc, NULL);
3528 		if (ret != 0 || nreloc == 0) {
3529 			panic("segvn_fill_vp_pages: "
3530 			    "page_relocate failed");
3531 		}
3532 		pp = newpp;
3533 		while (nreloc-- != 0) {
3534 			ASSERT(PAGE_EXCL(pp));
3535 			ASSERT(pp->p_vnode == vp);
3536 			ASSERT(pgidx ==
3537 			    ((pp->p_offset - start_off) >> PAGESHIFT));
3538 			ppa[pgidx++] = pp;
3539 			pp++;
3540 		}
3541 	}
3542 
3543 	if (svd->type == MAP_PRIVATE) {
3544 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[18]);
3545 		for (i = 0; i < pages; i++) {
3546 			ASSERT(ppa[i] != NULL);
3547 			ASSERT(PAGE_EXCL(ppa[i]));
3548 			ASSERT(ppa[i]->p_vnode == vp);
3549 			ASSERT(ppa[i]->p_offset ==
3550 			    start_off + (i << PAGESHIFT));
3551 			page_downgrade(ppa[i]);
3552 		}
3553 		ppa[pages] = NULL;
3554 	} else {
3555 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[19]);
3556 		/*
3557 		 * the caller will still call VOP_GETPAGE() for shared segments
3558 		 * to check FS write permissions. For private segments we map
3559 		 * file read only anyway.  so no VOP_GETPAGE is needed.
3560 		 */
3561 		for (i = 0; i < pages; i++) {
3562 			ASSERT(ppa[i] != NULL);
3563 			ASSERT(PAGE_EXCL(ppa[i]));
3564 			ASSERT(ppa[i]->p_vnode == vp);
3565 			ASSERT(ppa[i]->p_offset ==
3566 			    start_off + (i << PAGESHIFT));
3567 			page_unlock(ppa[i]);
3568 		}
3569 		ppa[0] = NULL;
3570 	}
3571 
3572 	return (1);
3573 out:
3574 	/*
3575 	 * Do the cleanup. Unlock target pages we didn't relocate. They are
3576 	 * linked on targ_pplist by root pages. reassemble unused replacement
3577 	 * and io pages back to pplist.
3578 	 */
3579 	if (io_pplist != NULL) {
3580 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[20]);
3581 		pp = io_pplist;
3582 		do {
3583 			ASSERT(pp->p_vnode == vp);
3584 			ASSERT(pp->p_offset == io_off);
3585 			ASSERT(page_iolock_assert(pp));
3586 			page_io_unlock(pp);
3587 			page_hashout(pp, NULL);
3588 			io_off += PAGESIZE;
3589 		} while ((pp = pp->p_next) != io_pplist);
3590 		page_list_concat(&io_pplist, &pplist);
3591 		pplist = io_pplist;
3592 	}
3593 	tmp_pplist = NULL;
3594 	while (targ_pplist != NULL) {
3595 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[21]);
3596 		pp = targ_pplist;
3597 		ASSERT(PAGE_EXCL(pp));
3598 		page_sub(&targ_pplist, pp);
3599 
3600 		pszc = pp->p_szc;
3601 		ppages = page_get_pagecnt(pszc);
3602 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3603 
3604 		if (pszc != 0) {
3605 			group_page_unlock(pp);
3606 		}
3607 		page_unlock(pp);
3608 
3609 		pp = repl_pplist;
3610 		ASSERT(pp != NULL);
3611 		ASSERT(PAGE_EXCL(pp));
3612 		ASSERT(pp->p_szc == szc);
3613 		page_sub(&repl_pplist, pp);
3614 
3615 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3616 
3617 		/* relink replacement page */
3618 		page_list_concat(&tmp_pplist, &pp);
3619 		while (--ppages != 0) {
3620 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[22]);
3621 			pp++;
3622 			ASSERT(PAGE_EXCL(pp));
3623 			ASSERT(pp->p_szc == szc);
3624 			page_list_concat(&tmp_pplist, &pp);
3625 		}
3626 	}
3627 	if (tmp_pplist != NULL) {
3628 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[23]);
3629 		page_list_concat(&tmp_pplist, &pplist);
3630 		pplist = tmp_pplist;
3631 	}
3632 	/*
3633 	 * at this point all pages are either on done_pplist or
3634 	 * pplist. They can't be all on done_pplist otherwise
3635 	 * we'd've been done.
3636 	 */
3637 	ASSERT(pplist != NULL);
3638 	if (nios != 0) {
3639 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[24]);
3640 		pp = pplist;
3641 		do {
3642 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[25]);
3643 			ASSERT(pp->p_szc == szc);
3644 			ASSERT(PAGE_EXCL(pp));
3645 			ASSERT(pp->p_vnode != vp);
3646 			pp->p_szc = 0;
3647 		} while ((pp = pp->p_next) != pplist);
3648 
3649 		pp = done_pplist;
3650 		do {
3651 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[26]);
3652 			ASSERT(pp->p_szc == szc);
3653 			ASSERT(PAGE_EXCL(pp));
3654 			ASSERT(pp->p_vnode == vp);
3655 			pp->p_szc = 0;
3656 		} while ((pp = pp->p_next) != done_pplist);
3657 
3658 		while (pplist != NULL) {
3659 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[27]);
3660 			pp = pplist;
3661 			page_sub(&pplist, pp);
3662 			page_free(pp, 0);
3663 		}
3664 
3665 		while (done_pplist != NULL) {
3666 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[28]);
3667 			pp = done_pplist;
3668 			page_sub(&done_pplist, pp);
3669 			page_unlock(pp);
3670 		}
3671 		*ppplist = NULL;
3672 		return (0);
3673 	}
3674 	ASSERT(pplist == *ppplist);
3675 	if (io_err) {
3676 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[29]);
3677 		/*
3678 		 * don't downsize on io error.
3679 		 * see if vop_getpage succeeds.
3680 		 * pplist may still be used in this case
3681 		 * for relocations.
3682 		 */
3683 		return (0);
3684 	}
3685 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[30]);
3686 	page_free_replacement_page(pplist);
3687 	page_create_putback(pages);
3688 	*ppplist = NULL;
3689 	return (0);
3690 }
3691 
3692 int segvn_anypgsz = 0;
3693 
3694 #define	SEGVN_RESTORE_SOFTLOCK(type, pages) 		\
3695 		if ((type) == F_SOFTLOCK) {		\
3696 			mutex_enter(&freemem_lock);	\
3697 			availrmem += (pages);		\
3698 			segvn_pages_locked -= (pages);	\
3699 			svd->softlockcnt -= (pages);	\
3700 			mutex_exit(&freemem_lock);	\
3701 		}
3702 
3703 #define	SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot)		\
3704 		if (IS_VMODSORT((ppa)[0]->p_vnode)) {			\
3705 			if ((rw) == S_WRITE) {				\
3706 				for (i = 0; i < (pages); i++) {		\
3707 					ASSERT((ppa)[i]->p_vnode ==	\
3708 					    (ppa)[0]->p_vnode);		\
3709 					hat_setmod((ppa)[i]);		\
3710 				}					\
3711 			} else if ((rw) != S_OTHER &&			\
3712 			    ((prot) & (vpprot) & PROT_WRITE)) {		\
3713 				for (i = 0; i < (pages); i++) {		\
3714 					ASSERT((ppa)[i]->p_vnode ==	\
3715 					    (ppa)[0]->p_vnode);		\
3716 					if (!hat_ismod((ppa)[i])) {	\
3717 						prot &= ~PROT_WRITE;	\
3718 						break;			\
3719 					}				\
3720 				}					\
3721 			}						\
3722 		}
3723 
3724 #ifdef  VM_STATS
3725 
3726 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)					\
3727 		VM_STAT_ADD(segvnvmstats.fltvnpages[(idx)]);
3728 
3729 #else /* VM_STATS */
3730 
3731 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)
3732 
3733 #endif
3734 
3735 static faultcode_t
3736 segvn_fault_vnodepages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
3737     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
3738     caddr_t eaddr, int brkcow)
3739 {
3740 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
3741 	struct anon_map *amp = svd->amp;
3742 	uchar_t segtype = svd->type;
3743 	uint_t szc = seg->s_szc;
3744 	size_t pgsz = page_get_pagesize(szc);
3745 	size_t maxpgsz = pgsz;
3746 	pgcnt_t pages = btop(pgsz);
3747 	pgcnt_t maxpages = pages;
3748 	size_t ppasize = (pages + 1) * sizeof (page_t *);
3749 	caddr_t a = lpgaddr;
3750 	caddr_t	maxlpgeaddr = lpgeaddr;
3751 	u_offset_t off = svd->offset + (uintptr_t)(a - seg->s_base);
3752 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
3753 	struct vpage *vpage = (svd->vpage != NULL) ?
3754 	    &svd->vpage[seg_page(seg, a)] : NULL;
3755 	vnode_t *vp = svd->vp;
3756 	page_t **ppa;
3757 	uint_t	pszc;
3758 	size_t	ppgsz;
3759 	pgcnt_t	ppages;
3760 	faultcode_t err = 0;
3761 	int ierr;
3762 	int vop_size_err = 0;
3763 	uint_t protchk, prot, vpprot;
3764 	ulong_t i;
3765 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
3766 	anon_sync_obj_t an_cookie;
3767 	enum seg_rw arw;
3768 	int alloc_failed = 0;
3769 	int adjszc_chk;
3770 	struct vattr va;
3771 	int xhat = 0;
3772 	page_t *pplist;
3773 	pfn_t pfn;
3774 	int physcontig;
3775 	int upgrdfail;
3776 	int segvn_anypgsz_vnode = 0; /* for now map vnode with 2 page sizes */
3777 	int tron = (svd->tr_state == SEGVN_TR_ON);
3778 
3779 	ASSERT(szc != 0);
3780 	ASSERT(vp != NULL);
3781 	ASSERT(brkcow == 0 || amp != NULL);
3782 	ASSERT(tron == 0 || amp != NULL);
3783 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
3784 	ASSERT(!(svd->flags & MAP_NORESERVE));
3785 	ASSERT(type != F_SOFTUNLOCK);
3786 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
3787 	ASSERT(amp == NULL || IS_P2ALIGNED(aindx, maxpages));
3788 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
3789 	ASSERT(seg->s_szc < NBBY * sizeof (int));
3790 	ASSERT(type != F_SOFTLOCK || lpgeaddr - a == maxpgsz);
3791 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
3792 
3793 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltvnpages[0]);
3794 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltvnpages[1]);
3795 
3796 	if (svd->flags & MAP_TEXT) {
3797 		hat_flag |= HAT_LOAD_TEXT;
3798 	}
3799 
3800 	if (svd->pageprot) {
3801 		switch (rw) {
3802 		case S_READ:
3803 			protchk = PROT_READ;
3804 			break;
3805 		case S_WRITE:
3806 			protchk = PROT_WRITE;
3807 			break;
3808 		case S_EXEC:
3809 			protchk = PROT_EXEC;
3810 			break;
3811 		case S_OTHER:
3812 		default:
3813 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
3814 			break;
3815 		}
3816 	} else {
3817 		prot = svd->prot;
3818 		/* caller has already done segment level protection check. */
3819 	}
3820 
3821 	if (seg->s_as->a_hat != hat) {
3822 		xhat = 1;
3823 	}
3824 
3825 	if (rw == S_WRITE && segtype == MAP_PRIVATE) {
3826 		SEGVN_VMSTAT_FLTVNPAGES(2);
3827 		arw = S_READ;
3828 	} else {
3829 		arw = rw;
3830 	}
3831 
3832 	ppa = kmem_alloc(ppasize, KM_SLEEP);
3833 
3834 	VM_STAT_COND_ADD(amp != NULL, segvnvmstats.fltvnpages[3]);
3835 
3836 	for (;;) {
3837 		adjszc_chk = 0;
3838 		for (; a < lpgeaddr; a += pgsz, off += pgsz, aindx += pages) {
3839 			if (adjszc_chk) {
3840 				while (szc < seg->s_szc) {
3841 					uintptr_t e;
3842 					uint_t tszc;
3843 					tszc = segvn_anypgsz_vnode ? szc + 1 :
3844 					    seg->s_szc;
3845 					ppgsz = page_get_pagesize(tszc);
3846 					if (!IS_P2ALIGNED(a, ppgsz) ||
3847 					    ((alloc_failed >> tszc) &
3848 						0x1)) {
3849 						break;
3850 					}
3851 					SEGVN_VMSTAT_FLTVNPAGES(4);
3852 					szc = tszc;
3853 					pgsz = ppgsz;
3854 					pages = btop(pgsz);
3855 					e = P2ROUNDUP((uintptr_t)eaddr, pgsz);
3856 					lpgeaddr = (caddr_t)e;
3857 				}
3858 			}
3859 
3860 		again:
3861 			if (IS_P2ALIGNED(a, maxpgsz) && amp != NULL) {
3862 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
3863 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
3864 				anon_array_enter(amp, aindx, &an_cookie);
3865 				if (anon_get_ptr(amp->ahp, aindx) != NULL) {
3866 					SEGVN_VMSTAT_FLTVNPAGES(5);
3867 					ASSERT(anon_pages(amp->ahp, aindx,
3868 					    maxpages) == maxpages);
3869 					anon_array_exit(&an_cookie);
3870 					ANON_LOCK_EXIT(&amp->a_rwlock);
3871 					err = segvn_fault_anonpages(hat, seg,
3872 					    a, a + maxpgsz, type, rw,
3873 					    MAX(a, addr),
3874 					    MIN(a + maxpgsz, eaddr), brkcow);
3875 					if (err != 0) {
3876 						SEGVN_VMSTAT_FLTVNPAGES(6);
3877 						goto out;
3878 					}
3879 					if (szc < seg->s_szc) {
3880 						szc = seg->s_szc;
3881 						pgsz = maxpgsz;
3882 						pages = maxpages;
3883 						lpgeaddr = maxlpgeaddr;
3884 					}
3885 					goto next;
3886 				} else {
3887 					ASSERT(anon_pages(amp->ahp, aindx,
3888 					    maxpages) == 0);
3889 					SEGVN_VMSTAT_FLTVNPAGES(7);
3890 					anon_array_exit(&an_cookie);
3891 					ANON_LOCK_EXIT(&amp->a_rwlock);
3892 				}
3893 			}
3894 			ASSERT(!brkcow || IS_P2ALIGNED(a, maxpgsz));
3895 			ASSERT(!tron || IS_P2ALIGNED(a, maxpgsz));
3896 
3897 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
3898 				ASSERT(vpage != NULL);
3899 				prot = VPP_PROT(vpage);
3900 				ASSERT(sameprot(seg, a, maxpgsz));
3901 				if ((prot & protchk) == 0) {
3902 					SEGVN_VMSTAT_FLTVNPAGES(8);
3903 					err = FC_PROT;
3904 					goto out;
3905 				}
3906 			}
3907 			if (type == F_SOFTLOCK) {
3908 				mutex_enter(&freemem_lock);
3909 				if (availrmem < tune.t_minarmem + pages) {
3910 					mutex_exit(&freemem_lock);
3911 					err = FC_MAKE_ERR(ENOMEM);
3912 					goto out;
3913 				} else {
3914 					availrmem -= pages;
3915 					segvn_pages_locked += pages;
3916 					svd->softlockcnt += pages;
3917 				}
3918 				mutex_exit(&freemem_lock);
3919 			}
3920 
3921 			pplist = NULL;
3922 			physcontig = 0;
3923 			ppa[0] = NULL;
3924 			if (!brkcow && !tron && szc &&
3925 			    !page_exists_physcontig(vp, off, szc,
3926 				segtype == MAP_PRIVATE ? ppa : NULL)) {
3927 				SEGVN_VMSTAT_FLTVNPAGES(9);
3928 				if (page_alloc_pages(vp, seg, a, &pplist, NULL,
3929 				    szc, 0, 0) && type != F_SOFTLOCK) {
3930 					SEGVN_VMSTAT_FLTVNPAGES(10);
3931 					pszc = 0;
3932 					ierr = -1;
3933 					alloc_failed |= (1 << szc);
3934 					break;
3935 				}
3936 				if (pplist != NULL &&
3937 				    vp->v_mpssdata == SEGVN_PAGEIO) {
3938 					int downsize;
3939 					SEGVN_VMSTAT_FLTVNPAGES(11);
3940 					physcontig = segvn_fill_vp_pages(svd,
3941 					    vp, off, szc, ppa, &pplist,
3942 					    &pszc, &downsize);
3943 					ASSERT(!physcontig || pplist == NULL);
3944 					if (!physcontig && downsize &&
3945 					    type != F_SOFTLOCK) {
3946 						ASSERT(pplist == NULL);
3947 						SEGVN_VMSTAT_FLTVNPAGES(12);
3948 						ierr = -1;
3949 						break;
3950 					}
3951 					ASSERT(!physcontig ||
3952 					    segtype == MAP_PRIVATE ||
3953 					    ppa[0] == NULL);
3954 					if (physcontig && ppa[0] == NULL) {
3955 						physcontig = 0;
3956 					}
3957 				}
3958 			} else if (!brkcow && !tron && szc && ppa[0] != NULL) {
3959 				SEGVN_VMSTAT_FLTVNPAGES(13);
3960 				ASSERT(segtype == MAP_PRIVATE);
3961 				physcontig = 1;
3962 			}
3963 
3964 			if (!physcontig) {
3965 				SEGVN_VMSTAT_FLTVNPAGES(14);
3966 				ppa[0] = NULL;
3967 				ierr = VOP_GETPAGE(vp, (offset_t)off, pgsz,
3968 				    &vpprot, ppa, pgsz, seg, a, arw,
3969 				    svd->cred, NULL);
3970 #ifdef DEBUG
3971 				if (ierr == 0) {
3972 					for (i = 0; i < pages; i++) {
3973 						ASSERT(PAGE_LOCKED(ppa[i]));
3974 						ASSERT(!PP_ISFREE(ppa[i]));
3975 						ASSERT(ppa[i]->p_vnode == vp);
3976 						ASSERT(ppa[i]->p_offset ==
3977 						    off + (i << PAGESHIFT));
3978 					}
3979 				}
3980 #endif /* DEBUG */
3981 				if (segtype == MAP_PRIVATE) {
3982 					SEGVN_VMSTAT_FLTVNPAGES(15);
3983 					vpprot &= ~PROT_WRITE;
3984 				}
3985 			} else {
3986 				ASSERT(segtype == MAP_PRIVATE);
3987 				SEGVN_VMSTAT_FLTVNPAGES(16);
3988 				vpprot = PROT_ALL & ~PROT_WRITE;
3989 				ierr = 0;
3990 			}
3991 
3992 			if (ierr != 0) {
3993 				SEGVN_VMSTAT_FLTVNPAGES(17);
3994 				if (pplist != NULL) {
3995 					SEGVN_VMSTAT_FLTVNPAGES(18);
3996 					page_free_replacement_page(pplist);
3997 					page_create_putback(pages);
3998 				}
3999 				SEGVN_RESTORE_SOFTLOCK(type, pages);
4000 				if (a + pgsz <= eaddr) {
4001 					SEGVN_VMSTAT_FLTVNPAGES(19);
4002 					err = FC_MAKE_ERR(ierr);
4003 					goto out;
4004 				}
4005 				va.va_mask = AT_SIZE;
4006 				if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL)) {
4007 					SEGVN_VMSTAT_FLTVNPAGES(20);
4008 					err = FC_MAKE_ERR(EIO);
4009 					goto out;
4010 				}
4011 				if (btopr(va.va_size) >= btopr(off + pgsz)) {
4012 					SEGVN_VMSTAT_FLTVNPAGES(21);
4013 					err = FC_MAKE_ERR(ierr);
4014 					goto out;
4015 				}
4016 				if (btopr(va.va_size) <
4017 				    btopr(off + (eaddr - a))) {
4018 					SEGVN_VMSTAT_FLTVNPAGES(22);
4019 					err = FC_MAKE_ERR(ierr);
4020 					goto out;
4021 				}
4022 				if (brkcow || tron || type == F_SOFTLOCK) {
4023 					/* can't reduce map area */
4024 					SEGVN_VMSTAT_FLTVNPAGES(23);
4025 					vop_size_err = 1;
4026 					goto out;
4027 				}
4028 				SEGVN_VMSTAT_FLTVNPAGES(24);
4029 				ASSERT(szc != 0);
4030 				pszc = 0;
4031 				ierr = -1;
4032 				break;
4033 			}
4034 
4035 			if (amp != NULL) {
4036 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4037 				anon_array_enter(amp, aindx, &an_cookie);
4038 			}
4039 			if (amp != NULL &&
4040 			    anon_get_ptr(amp->ahp, aindx) != NULL) {
4041 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4042 
4043 				SEGVN_VMSTAT_FLTVNPAGES(25);
4044 				ASSERT(anon_pages(amp->ahp, taindx,
4045 				    maxpages) == maxpages);
4046 				for (i = 0; i < pages; i++) {
4047 					page_unlock(ppa[i]);
4048 				}
4049 				anon_array_exit(&an_cookie);
4050 				ANON_LOCK_EXIT(&amp->a_rwlock);
4051 				if (pplist != NULL) {
4052 					page_free_replacement_page(pplist);
4053 					page_create_putback(pages);
4054 				}
4055 				SEGVN_RESTORE_SOFTLOCK(type, pages);
4056 				if (szc < seg->s_szc) {
4057 					SEGVN_VMSTAT_FLTVNPAGES(26);
4058 					/*
4059 					 * For private segments SOFTLOCK
4060 					 * either always breaks cow (any rw
4061 					 * type except S_READ_NOCOW) or
4062 					 * address space is locked as writer
4063 					 * (S_READ_NOCOW case) and anon slots
4064 					 * can't show up on second check.
4065 					 * Therefore if we are here for
4066 					 * SOFTLOCK case it must be a cow
4067 					 * break but cow break never reduces
4068 					 * szc. text replication (tron) in
4069 					 * this case works as cow break.
4070 					 * Thus the assert below.
4071 					 */
4072 					ASSERT(!brkcow && !tron &&
4073 					    type != F_SOFTLOCK);
4074 					pszc = seg->s_szc;
4075 					ierr = -2;
4076 					break;
4077 				}
4078 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4079 				goto again;
4080 			}
4081 #ifdef DEBUG
4082 			if (amp != NULL) {
4083 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4084 				ASSERT(!anon_pages(amp->ahp, taindx, maxpages));
4085 			}
4086 #endif /* DEBUG */
4087 
4088 			if (brkcow || tron) {
4089 				ASSERT(amp != NULL);
4090 				ASSERT(pplist == NULL);
4091 				ASSERT(szc == seg->s_szc);
4092 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4093 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
4094 				SEGVN_VMSTAT_FLTVNPAGES(27);
4095 				ierr = anon_map_privatepages(amp, aindx, szc,
4096 				    seg, a, prot, ppa, vpage, segvn_anypgsz,
4097 				    tron ? PG_LOCAL : 0, svd->cred);
4098 				if (ierr != 0) {
4099 					SEGVN_VMSTAT_FLTVNPAGES(28);
4100 					anon_array_exit(&an_cookie);
4101 					ANON_LOCK_EXIT(&amp->a_rwlock);
4102 					SEGVN_RESTORE_SOFTLOCK(type, pages);
4103 					err = FC_MAKE_ERR(ierr);
4104 					goto out;
4105 				}
4106 
4107 				ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4108 				/*
4109 				 * p_szc can't be changed for locked
4110 				 * swapfs pages.
4111 				 */
4112 				ASSERT(svd->rcookie ==
4113 				    HAT_INVALID_REGION_COOKIE);
4114 				hat_memload_array(hat, a, pgsz, ppa, prot,
4115 				    hat_flag);
4116 
4117 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4118 					SEGVN_VMSTAT_FLTVNPAGES(29);
4119 					for (i = 0; i < pages; i++) {
4120 						page_unlock(ppa[i]);
4121 					}
4122 				}
4123 				anon_array_exit(&an_cookie);
4124 				ANON_LOCK_EXIT(&amp->a_rwlock);
4125 				goto next;
4126 			}
4127 
4128 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
4129 			    (!svd->pageprot && svd->prot == (prot & vpprot)));
4130 
4131 			pfn = page_pptonum(ppa[0]);
4132 			/*
4133 			 * hat_page_demote() needs an SE_EXCL lock on one of
4134 			 * constituent page_t's and it decreases root's p_szc
4135 			 * last. This means if root's p_szc is equal szc and
4136 			 * all its constituent pages are locked
4137 			 * hat_page_demote() that could have changed p_szc to
4138 			 * szc is already done and no new have page_demote()
4139 			 * can start for this large page.
4140 			 */
4141 
4142 			/*
4143 			 * we need to make sure same mapping size is used for
4144 			 * the same address range if there's a possibility the
4145 			 * adddress is already mapped because hat layer panics
4146 			 * when translation is loaded for the range already
4147 			 * mapped with a different page size.  We achieve it
4148 			 * by always using largest page size possible subject
4149 			 * to the constraints of page size, segment page size
4150 			 * and page alignment.  Since mappings are invalidated
4151 			 * when those constraints change and make it
4152 			 * impossible to use previously used mapping size no
4153 			 * mapping size conflicts should happen.
4154 			 */
4155 
4156 		chkszc:
4157 			if ((pszc = ppa[0]->p_szc) == szc &&
4158 			    IS_P2ALIGNED(pfn, pages)) {
4159 
4160 				SEGVN_VMSTAT_FLTVNPAGES(30);
4161 #ifdef DEBUG
4162 				for (i = 0; i < pages; i++) {
4163 					ASSERT(PAGE_LOCKED(ppa[i]));
4164 					ASSERT(!PP_ISFREE(ppa[i]));
4165 					ASSERT(page_pptonum(ppa[i]) ==
4166 					    pfn + i);
4167 					ASSERT(ppa[i]->p_szc == szc);
4168 					ASSERT(ppa[i]->p_vnode == vp);
4169 					ASSERT(ppa[i]->p_offset ==
4170 					    off + (i << PAGESHIFT));
4171 				}
4172 #endif /* DEBUG */
4173 				/*
4174 				 * All pages are of szc we need and they are
4175 				 * all locked so they can't change szc. load
4176 				 * translations.
4177 				 *
4178 				 * if page got promoted since last check
4179 				 * we don't need pplist.
4180 				 */
4181 				if (pplist != NULL) {
4182 					page_free_replacement_page(pplist);
4183 					page_create_putback(pages);
4184 				}
4185 				if (PP_ISMIGRATE(ppa[0])) {
4186 					page_migrate(seg, a, ppa, pages);
4187 				}
4188 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4189 				    prot, vpprot);
4190 				if (!xhat) {
4191 					hat_memload_array_region(hat, a, pgsz,
4192 					    ppa, prot & vpprot, hat_flag,
4193 					    svd->rcookie);
4194 				} else {
4195 					/*
4196 					 * avoid large xhat mappings to FS
4197 					 * pages so that hat_page_demote()
4198 					 * doesn't need to check for xhat
4199 					 * large mappings.
4200 					 * Don't use regions with xhats.
4201 					 */
4202 					for (i = 0; i < pages; i++) {
4203 						hat_memload(hat,
4204 						    a + (i << PAGESHIFT),
4205 						    ppa[i], prot & vpprot,
4206 						    hat_flag);
4207 					}
4208 				}
4209 
4210 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4211 					for (i = 0; i < pages; i++) {
4212 						page_unlock(ppa[i]);
4213 					}
4214 				}
4215 				if (amp != NULL) {
4216 					anon_array_exit(&an_cookie);
4217 					ANON_LOCK_EXIT(&amp->a_rwlock);
4218 				}
4219 				goto next;
4220 			}
4221 
4222 			/*
4223 			 * See if upsize is possible.
4224 			 */
4225 			if (pszc > szc && szc < seg->s_szc &&
4226 			    (segvn_anypgsz_vnode || pszc >= seg->s_szc)) {
4227 				pgcnt_t aphase;
4228 				uint_t pszc1 = MIN(pszc, seg->s_szc);
4229 				ppgsz = page_get_pagesize(pszc1);
4230 				ppages = btop(ppgsz);
4231 				aphase = btop(P2PHASE((uintptr_t)a, ppgsz));
4232 
4233 				ASSERT(type != F_SOFTLOCK);
4234 
4235 				SEGVN_VMSTAT_FLTVNPAGES(31);
4236 				if (aphase != P2PHASE(pfn, ppages)) {
4237 					segvn_faultvnmpss_align_err4++;
4238 				} else {
4239 					SEGVN_VMSTAT_FLTVNPAGES(32);
4240 					if (pplist != NULL) {
4241 						page_t *pl = pplist;
4242 						page_free_replacement_page(pl);
4243 						page_create_putback(pages);
4244 					}
4245 					for (i = 0; i < pages; i++) {
4246 						page_unlock(ppa[i]);
4247 					}
4248 					if (amp != NULL) {
4249 						anon_array_exit(&an_cookie);
4250 						ANON_LOCK_EXIT(&amp->a_rwlock);
4251 					}
4252 					pszc = pszc1;
4253 					ierr = -2;
4254 					break;
4255 				}
4256 			}
4257 
4258 			/*
4259 			 * check if we should use smallest mapping size.
4260 			 */
4261 			upgrdfail = 0;
4262 			if (szc == 0 || xhat ||
4263 			    (pszc >= szc &&
4264 			    !IS_P2ALIGNED(pfn, pages)) ||
4265 			    (pszc < szc &&
4266 			    !segvn_full_szcpages(ppa, szc, &upgrdfail,
4267 				&pszc))) {
4268 
4269 				if (upgrdfail && type != F_SOFTLOCK) {
4270 					/*
4271 					 * segvn_full_szcpages failed to lock
4272 					 * all pages EXCL. Size down.
4273 					 */
4274 					ASSERT(pszc < szc);
4275 
4276 					SEGVN_VMSTAT_FLTVNPAGES(33);
4277 
4278 					if (pplist != NULL) {
4279 						page_t *pl = pplist;
4280 						page_free_replacement_page(pl);
4281 						page_create_putback(pages);
4282 					}
4283 
4284 					for (i = 0; i < pages; i++) {
4285 						page_unlock(ppa[i]);
4286 					}
4287 					if (amp != NULL) {
4288 						anon_array_exit(&an_cookie);
4289 						ANON_LOCK_EXIT(&amp->a_rwlock);
4290 					}
4291 					ierr = -1;
4292 					break;
4293 				}
4294 				if (szc != 0 && !xhat && !upgrdfail) {
4295 					segvn_faultvnmpss_align_err5++;
4296 				}
4297 				SEGVN_VMSTAT_FLTVNPAGES(34);
4298 				if (pplist != NULL) {
4299 					page_free_replacement_page(pplist);
4300 					page_create_putback(pages);
4301 				}
4302 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4303 				    prot, vpprot);
4304 				if (upgrdfail && segvn_anypgsz_vnode) {
4305 					/* SOFTLOCK case */
4306 					hat_memload_array_region(hat, a, pgsz,
4307 					    ppa, prot & vpprot, hat_flag,
4308 					    svd->rcookie);
4309 				} else {
4310 					for (i = 0; i < pages; i++) {
4311 						hat_memload_region(hat,
4312 						    a + (i << PAGESHIFT),
4313 						    ppa[i], prot & vpprot,
4314 						    hat_flag, svd->rcookie);
4315 					}
4316 				}
4317 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4318 					for (i = 0; i < pages; i++) {
4319 						page_unlock(ppa[i]);
4320 					}
4321 				}
4322 				if (amp != NULL) {
4323 					anon_array_exit(&an_cookie);
4324 					ANON_LOCK_EXIT(&amp->a_rwlock);
4325 				}
4326 				goto next;
4327 			}
4328 
4329 			if (pszc == szc) {
4330 				/*
4331 				 * segvn_full_szcpages() upgraded pages szc.
4332 				 */
4333 				ASSERT(pszc == ppa[0]->p_szc);
4334 				ASSERT(IS_P2ALIGNED(pfn, pages));
4335 				goto chkszc;
4336 			}
4337 
4338 			if (pszc > szc) {
4339 				kmutex_t *szcmtx;
4340 				SEGVN_VMSTAT_FLTVNPAGES(35);
4341 				/*
4342 				 * p_szc of ppa[0] can change since we haven't
4343 				 * locked all constituent pages. Call
4344 				 * page_lock_szc() to prevent szc changes.
4345 				 * This should be a rare case that happens when
4346 				 * multiple segments use a different page size
4347 				 * to map the same file offsets.
4348 				 */
4349 				szcmtx = page_szc_lock(ppa[0]);
4350 				pszc = ppa[0]->p_szc;
4351 				ASSERT(szcmtx != NULL || pszc == 0);
4352 				ASSERT(ppa[0]->p_szc <= pszc);
4353 				if (pszc <= szc) {
4354 					SEGVN_VMSTAT_FLTVNPAGES(36);
4355 					if (szcmtx != NULL) {
4356 						mutex_exit(szcmtx);
4357 					}
4358 					goto chkszc;
4359 				}
4360 				if (pplist != NULL) {
4361 					/*
4362 					 * page got promoted since last check.
4363 					 * we don't need preaalocated large
4364 					 * page.
4365 					 */
4366 					SEGVN_VMSTAT_FLTVNPAGES(37);
4367 					page_free_replacement_page(pplist);
4368 					page_create_putback(pages);
4369 				}
4370 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4371 				    prot, vpprot);
4372 				hat_memload_array_region(hat, a, pgsz, ppa,
4373 				    prot & vpprot, hat_flag, svd->rcookie);
4374 				mutex_exit(szcmtx);
4375 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4376 					for (i = 0; i < pages; i++) {
4377 						page_unlock(ppa[i]);
4378 					}
4379 				}
4380 				if (amp != NULL) {
4381 					anon_array_exit(&an_cookie);
4382 					ANON_LOCK_EXIT(&amp->a_rwlock);
4383 				}
4384 				goto next;
4385 			}
4386 
4387 			/*
4388 			 * if page got demoted since last check
4389 			 * we could have not allocated larger page.
4390 			 * allocate now.
4391 			 */
4392 			if (pplist == NULL &&
4393 			    page_alloc_pages(vp, seg, a, &pplist, NULL,
4394 				szc, 0, 0) && type != F_SOFTLOCK) {
4395 				SEGVN_VMSTAT_FLTVNPAGES(38);
4396 				for (i = 0; i < pages; i++) {
4397 					page_unlock(ppa[i]);
4398 				}
4399 				if (amp != NULL) {
4400 					anon_array_exit(&an_cookie);
4401 					ANON_LOCK_EXIT(&amp->a_rwlock);
4402 				}
4403 				ierr = -1;
4404 				alloc_failed |= (1 << szc);
4405 				break;
4406 			}
4407 
4408 			SEGVN_VMSTAT_FLTVNPAGES(39);
4409 
4410 			if (pplist != NULL) {
4411 				segvn_relocate_pages(ppa, pplist);
4412 #ifdef DEBUG
4413 			} else {
4414 				ASSERT(type == F_SOFTLOCK);
4415 				SEGVN_VMSTAT_FLTVNPAGES(40);
4416 #endif /* DEBUG */
4417 			}
4418 
4419 			SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot);
4420 
4421 			if (pplist == NULL && segvn_anypgsz_vnode == 0) {
4422 				ASSERT(type == F_SOFTLOCK);
4423 				for (i = 0; i < pages; i++) {
4424 					ASSERT(ppa[i]->p_szc < szc);
4425 					hat_memload_region(hat,
4426 					    a + (i << PAGESHIFT),
4427 					    ppa[i], prot & vpprot, hat_flag,
4428 					    svd->rcookie);
4429 				}
4430 			} else {
4431 				ASSERT(pplist != NULL || type == F_SOFTLOCK);
4432 				hat_memload_array_region(hat, a, pgsz, ppa,
4433 				    prot & vpprot, hat_flag, svd->rcookie);
4434 			}
4435 			if (!(hat_flag & HAT_LOAD_LOCK)) {
4436 				for (i = 0; i < pages; i++) {
4437 					ASSERT(PAGE_SHARED(ppa[i]));
4438 					page_unlock(ppa[i]);
4439 				}
4440 			}
4441 			if (amp != NULL) {
4442 				anon_array_exit(&an_cookie);
4443 				ANON_LOCK_EXIT(&amp->a_rwlock);
4444 			}
4445 
4446 		next:
4447 			if (vpage != NULL) {
4448 				vpage += pages;
4449 			}
4450 			adjszc_chk = 1;
4451 		}
4452 		if (a == lpgeaddr)
4453 			break;
4454 		ASSERT(a < lpgeaddr);
4455 
4456 		ASSERT(!brkcow && !tron && type != F_SOFTLOCK);
4457 
4458 		/*
4459 		 * ierr == -1 means we failed to map with a large page.
4460 		 * (either due to allocation/relocation failures or
4461 		 * misalignment with other mappings to this file.
4462 		 *
4463 		 * ierr == -2 means some other thread allocated a large page
4464 		 * after we gave up tp map with a large page.  retry with
4465 		 * larger mapping.
4466 		 */
4467 		ASSERT(ierr == -1 || ierr == -2);
4468 		ASSERT(ierr == -2 || szc != 0);
4469 		ASSERT(ierr == -1 || szc < seg->s_szc);
4470 		if (ierr == -2) {
4471 			SEGVN_VMSTAT_FLTVNPAGES(41);
4472 			ASSERT(pszc > szc && pszc <= seg->s_szc);
4473 			szc = pszc;
4474 		} else if (segvn_anypgsz_vnode) {
4475 			SEGVN_VMSTAT_FLTVNPAGES(42);
4476 			szc--;
4477 		} else {
4478 			SEGVN_VMSTAT_FLTVNPAGES(43);
4479 			ASSERT(pszc < szc);
4480 			/*
4481 			 * other process created pszc large page.
4482 			 * but we still have to drop to 0 szc.
4483 			 */
4484 			szc = 0;
4485 		}
4486 
4487 		pgsz = page_get_pagesize(szc);
4488 		pages = btop(pgsz);
4489 		if (ierr == -2) {
4490 			/*
4491 			 * Size up case. Note lpgaddr may only be needed for
4492 			 * softlock case so we don't adjust it here.
4493 			 */
4494 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4495 			ASSERT(a >= lpgaddr);
4496 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4497 			off = svd->offset + (uintptr_t)(a - seg->s_base);
4498 			aindx = svd->anon_index + seg_page(seg, a);
4499 			vpage = (svd->vpage != NULL) ?
4500 			    &svd->vpage[seg_page(seg, a)] : NULL;
4501 		} else {
4502 			/*
4503 			 * Size down case. Note lpgaddr may only be needed for
4504 			 * softlock case so we don't adjust it here.
4505 			 */
4506 			ASSERT(IS_P2ALIGNED(a, pgsz));
4507 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4508 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4509 			ASSERT(a < lpgeaddr);
4510 			if (a < addr) {
4511 				SEGVN_VMSTAT_FLTVNPAGES(44);
4512 				/*
4513 				 * The beginning of the large page region can
4514 				 * be pulled to the right to make a smaller
4515 				 * region. We haven't yet faulted a single
4516 				 * page.
4517 				 */
4518 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4519 				ASSERT(a >= lpgaddr);
4520 				off = svd->offset +
4521 				    (uintptr_t)(a - seg->s_base);
4522 				aindx = svd->anon_index + seg_page(seg, a);
4523 				vpage = (svd->vpage != NULL) ?
4524 				    &svd->vpage[seg_page(seg, a)] : NULL;
4525 			}
4526 		}
4527 	}
4528 out:
4529 	kmem_free(ppa, ppasize);
4530 	if (!err && !vop_size_err) {
4531 		SEGVN_VMSTAT_FLTVNPAGES(45);
4532 		return (0);
4533 	}
4534 	if (type == F_SOFTLOCK && a > lpgaddr) {
4535 		SEGVN_VMSTAT_FLTVNPAGES(46);
4536 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4537 	}
4538 	if (!vop_size_err) {
4539 		SEGVN_VMSTAT_FLTVNPAGES(47);
4540 		return (err);
4541 	}
4542 	ASSERT(brkcow || tron || type == F_SOFTLOCK);
4543 	/*
4544 	 * Large page end is mapped beyond the end of file and it's a cow
4545 	 * fault (can be a text replication induced cow) or softlock so we can't
4546 	 * reduce the map area.  For now just demote the segment. This should
4547 	 * really only happen if the end of the file changed after the mapping
4548 	 * was established since when large page segments are created we make
4549 	 * sure they don't extend beyond the end of the file.
4550 	 */
4551 	SEGVN_VMSTAT_FLTVNPAGES(48);
4552 
4553 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4554 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4555 	err = 0;
4556 	if (seg->s_szc != 0) {
4557 		segvn_fltvnpages_clrszc_cnt++;
4558 		ASSERT(svd->softlockcnt == 0);
4559 		err = segvn_clrszc(seg);
4560 		if (err != 0) {
4561 			segvn_fltvnpages_clrszc_err++;
4562 		}
4563 	}
4564 	ASSERT(err || seg->s_szc == 0);
4565 	SEGVN_LOCK_DOWNGRADE(seg->s_as, &svd->lock);
4566 	/* segvn_fault will do its job as if szc had been zero to begin with */
4567 	return (err == 0 ? IE_RETRY : FC_MAKE_ERR(err));
4568 }
4569 
4570 /*
4571  * This routine will attempt to fault in one large page.
4572  * it will use smaller pages if that fails.
4573  * It should only be called for pure anonymous segments.
4574  */
4575 static faultcode_t
4576 segvn_fault_anonpages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
4577     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
4578     caddr_t eaddr, int brkcow)
4579 {
4580 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4581 	struct anon_map *amp = svd->amp;
4582 	uchar_t segtype = svd->type;
4583 	uint_t szc = seg->s_szc;
4584 	size_t pgsz = page_get_pagesize(szc);
4585 	size_t maxpgsz = pgsz;
4586 	pgcnt_t pages = btop(pgsz);
4587 	size_t ppasize = pages * sizeof (page_t *);
4588 	caddr_t a = lpgaddr;
4589 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
4590 	struct vpage *vpage = (svd->vpage != NULL) ?
4591 	    &svd->vpage[seg_page(seg, a)] : NULL;
4592 	page_t **ppa;
4593 	uint_t	ppa_szc;
4594 	faultcode_t err;
4595 	int ierr;
4596 	uint_t protchk, prot, vpprot;
4597 	ulong_t i;
4598 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
4599 	anon_sync_obj_t cookie;
4600 	int first = 1;
4601 	int adjszc_chk;
4602 	int purged = 0;
4603 	int pgflags = (svd->tr_state == SEGVN_TR_ON) ? PG_LOCAL : 0;
4604 
4605 	ASSERT(szc != 0);
4606 	ASSERT(amp != NULL);
4607 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
4608 	ASSERT(!(svd->flags & MAP_NORESERVE));
4609 	ASSERT(type != F_SOFTUNLOCK);
4610 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
4611 	ASSERT(!brkcow || svd->tr_state == SEGVN_TR_OFF);
4612 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
4613 
4614 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
4615 
4616 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltanpages[0]);
4617 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltanpages[1]);
4618 
4619 	if (svd->flags & MAP_TEXT) {
4620 		hat_flag |= HAT_LOAD_TEXT;
4621 	}
4622 
4623 	if (svd->pageprot) {
4624 		switch (rw) {
4625 		case S_READ:
4626 			protchk = PROT_READ;
4627 			break;
4628 		case S_WRITE:
4629 			protchk = PROT_WRITE;
4630 			break;
4631 		case S_EXEC:
4632 			protchk = PROT_EXEC;
4633 			break;
4634 		case S_OTHER:
4635 		default:
4636 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4637 			break;
4638 		}
4639 		VM_STAT_ADD(segvnvmstats.fltanpages[2]);
4640 	} else {
4641 		prot = svd->prot;
4642 		/* caller has already done segment level protection check. */
4643 	}
4644 
4645 	ppa = kmem_alloc(ppasize, KM_SLEEP);
4646 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4647 	for (;;) {
4648 		adjszc_chk = 0;
4649 		for (; a < lpgeaddr; a += pgsz, aindx += pages) {
4650 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
4651 				VM_STAT_ADD(segvnvmstats.fltanpages[3]);
4652 				ASSERT(vpage != NULL);
4653 				prot = VPP_PROT(vpage);
4654 				ASSERT(sameprot(seg, a, maxpgsz));
4655 				if ((prot & protchk) == 0) {
4656 					err = FC_PROT;
4657 					goto error;
4658 				}
4659 			}
4660 			if (adjszc_chk && IS_P2ALIGNED(a, maxpgsz) &&
4661 			    pgsz < maxpgsz) {
4662 				ASSERT(a > lpgaddr);
4663 				szc = seg->s_szc;
4664 				pgsz = maxpgsz;
4665 				pages = btop(pgsz);
4666 				ASSERT(IS_P2ALIGNED(aindx, pages));
4667 				lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr,
4668 				    pgsz);
4669 			}
4670 			if (type == F_SOFTLOCK && svd->vp != NULL) {
4671 				mutex_enter(&freemem_lock);
4672 				if (availrmem < tune.t_minarmem + pages) {
4673 					mutex_exit(&freemem_lock);
4674 					err = FC_MAKE_ERR(ENOMEM);
4675 					goto error;
4676 				} else {
4677 					availrmem -= pages;
4678 					segvn_pages_locked += pages;
4679 					svd->softlockcnt += pages;
4680 				}
4681 				mutex_exit(&freemem_lock);
4682 			}
4683 			anon_array_enter(amp, aindx, &cookie);
4684 			ppa_szc = (uint_t)-1;
4685 			ierr = anon_map_getpages(amp, aindx, szc, seg, a,
4686 				prot, &vpprot, ppa, &ppa_szc, vpage, rw, brkcow,
4687 				segvn_anypgsz, pgflags, svd->cred);
4688 			if (ierr != 0) {
4689 				anon_array_exit(&cookie);
4690 				VM_STAT_ADD(segvnvmstats.fltanpages[4]);
4691 				if (type == F_SOFTLOCK && svd->vp != NULL) {
4692 					VM_STAT_ADD(segvnvmstats.fltanpages[5]);
4693 					mutex_enter(&freemem_lock);
4694 					availrmem += pages;
4695 					segvn_pages_locked -= pages;
4696 					svd->softlockcnt -= pages;
4697 					mutex_exit(&freemem_lock);
4698 				}
4699 				if (ierr > 0) {
4700 					VM_STAT_ADD(segvnvmstats.fltanpages[6]);
4701 					err = FC_MAKE_ERR(ierr);
4702 					goto error;
4703 				}
4704 				break;
4705 			}
4706 
4707 			ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4708 
4709 			ASSERT(segtype == MAP_SHARED ||
4710 			    ppa[0]->p_szc <= szc);
4711 			ASSERT(segtype == MAP_PRIVATE ||
4712 			    ppa[0]->p_szc >= szc);
4713 
4714 			/*
4715 			 * Handle pages that have been marked for migration
4716 			 */
4717 			if (lgrp_optimizations())
4718 				page_migrate(seg, a, ppa, pages);
4719 
4720 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
4721 			if (type == F_SOFTLOCK && svd->vp == NULL) {
4722 				/*
4723 				 * If all pages in ppa array belong to the same
4724 				 * large page call segvn_slock_anonpages()
4725 				 * just for ppa[0].
4726 				 */
4727 				for (i = 0; i < pages; i++) {
4728 					if (!segvn_slock_anonpages(ppa[i],
4729 					    i == 0 && first)) {
4730 						ulong_t j;
4731 						for (j = 0; j < i; j++) {
4732 							segvn_sunlock_anonpages(
4733 								ppa[j],
4734 								j == 0 &&
4735 								first);
4736 							page_unlock(ppa[j]);
4737 						}
4738 						for (j = i; j < pages; j++) {
4739 							page_unlock(ppa[j]);
4740 						}
4741 						anon_array_exit(&cookie);
4742 						err = FC_MAKE_ERR(ENOMEM);
4743 						goto error;
4744 					}
4745 					if (i == 0 && ppa[0]->p_szc >= szc) {
4746 						ASSERT(!(page_pptonum(ppa[0]) &
4747 						    (pages - 1)));
4748 						break;
4749 					}
4750 				}
4751 				first = 0;
4752 				mutex_enter(&freemem_lock);
4753 				svd->softlockcnt += pages;
4754 				segvn_pages_locked += pages;
4755 				mutex_exit(&freemem_lock);
4756 			}
4757 
4758 			if (segtype == MAP_SHARED) {
4759 				vpprot |= PROT_WRITE;
4760 			}
4761 
4762 			hat_memload_array(hat, a, pgsz, ppa,
4763 			    prot & vpprot, hat_flag);
4764 
4765 			if (hat_flag & HAT_LOAD_LOCK) {
4766 				VM_STAT_ADD(segvnvmstats.fltanpages[7]);
4767 			} else {
4768 				VM_STAT_ADD(segvnvmstats.fltanpages[8]);
4769 				for (i = 0; i < pages; i++)
4770 					page_unlock(ppa[i]);
4771 			}
4772 			if (vpage != NULL)
4773 				vpage += pages;
4774 
4775 			anon_array_exit(&cookie);
4776 			adjszc_chk = 1;
4777 		}
4778 		if (a == lpgeaddr)
4779 			break;
4780 		ASSERT(a < lpgeaddr);
4781 		/*
4782 		 * ierr == -1 means we failed to allocate a large page.
4783 		 * so do a size down operation.
4784 		 *
4785 		 * ierr == -2 means some other process that privately shares
4786 		 * pages with this process has allocated a larger page and we
4787 		 * need to retry with larger pages. So do a size up
4788 		 * operation. This relies on the fact that large pages are
4789 		 * never partially shared i.e. if we share any constituent
4790 		 * page of a large page with another process we must share the
4791 		 * entire large page. Note this cannot happen for SOFTLOCK
4792 		 * case, unless current address (a) is at the beginning of the
4793 		 * next page size boundary because the other process couldn't
4794 		 * have relocated locked pages.
4795 		 */
4796 		ASSERT(ierr == -1 || ierr == -2);
4797 		/*
4798 		 * For the very first relocation failure try to purge this
4799 		 * segment's cache so that the relocator can obtain an
4800 		 * exclusive lock on pages we want to relocate.
4801 		 */
4802 		if (!purged && ierr == -1 && ppa_szc != (uint_t)-1 &&
4803 		    svd->softlockcnt != 0) {
4804 			purged = 1;
4805 			segvn_purge(seg);
4806 			continue;
4807 		}
4808 
4809 		if (segvn_anypgsz) {
4810 			ASSERT(ierr == -2 || szc != 0);
4811 			ASSERT(ierr == -1 || szc < seg->s_szc);
4812 			szc = (ierr == -1) ? szc - 1 : szc + 1;
4813 		} else {
4814 			/*
4815 			 * For non COW faults and segvn_anypgsz == 0
4816 			 * we need to be careful not to loop forever
4817 			 * if existing page is found with szc other
4818 			 * than 0 or seg->s_szc. This could be due
4819 			 * to page relocations on behalf of DR or
4820 			 * more likely large page creation. For this
4821 			 * case simply re-size to existing page's szc
4822 			 * if returned by anon_map_getpages().
4823 			 */
4824 			if (ppa_szc == (uint_t)-1) {
4825 				szc = (ierr == -1) ? 0 : seg->s_szc;
4826 			} else {
4827 				ASSERT(ppa_szc <= seg->s_szc);
4828 				ASSERT(ierr == -2 || ppa_szc < szc);
4829 				ASSERT(ierr == -1 || ppa_szc > szc);
4830 				szc = ppa_szc;
4831 			}
4832 		}
4833 
4834 		pgsz = page_get_pagesize(szc);
4835 		pages = btop(pgsz);
4836 		ASSERT(type != F_SOFTLOCK || ierr == -1 ||
4837 		    (IS_P2ALIGNED(a, pgsz) && IS_P2ALIGNED(lpgeaddr, pgsz)));
4838 		if (type == F_SOFTLOCK) {
4839 			/*
4840 			 * For softlocks we cannot reduce the fault area
4841 			 * (calculated based on the largest page size for this
4842 			 * segment) for size down and a is already next
4843 			 * page size aligned as assertted above for size
4844 			 * ups. Therefore just continue in case of softlock.
4845 			 */
4846 			VM_STAT_ADD(segvnvmstats.fltanpages[9]);
4847 			continue; /* keep lint happy */
4848 		} else if (ierr == -2) {
4849 
4850 			/*
4851 			 * Size up case. Note lpgaddr may only be needed for
4852 			 * softlock case so we don't adjust it here.
4853 			 */
4854 			VM_STAT_ADD(segvnvmstats.fltanpages[10]);
4855 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4856 			ASSERT(a >= lpgaddr);
4857 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4858 			aindx = svd->anon_index + seg_page(seg, a);
4859 			vpage = (svd->vpage != NULL) ?
4860 			    &svd->vpage[seg_page(seg, a)] : NULL;
4861 		} else {
4862 			/*
4863 			 * Size down case. Note lpgaddr may only be needed for
4864 			 * softlock case so we don't adjust it here.
4865 			 */
4866 			VM_STAT_ADD(segvnvmstats.fltanpages[11]);
4867 			ASSERT(IS_P2ALIGNED(a, pgsz));
4868 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4869 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4870 			ASSERT(a < lpgeaddr);
4871 			if (a < addr) {
4872 				/*
4873 				 * The beginning of the large page region can
4874 				 * be pulled to the right to make a smaller
4875 				 * region. We haven't yet faulted a single
4876 				 * page.
4877 				 */
4878 				VM_STAT_ADD(segvnvmstats.fltanpages[12]);
4879 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4880 				ASSERT(a >= lpgaddr);
4881 				aindx = svd->anon_index + seg_page(seg, a);
4882 				vpage = (svd->vpage != NULL) ?
4883 				    &svd->vpage[seg_page(seg, a)] : NULL;
4884 			}
4885 		}
4886 	}
4887 	VM_STAT_ADD(segvnvmstats.fltanpages[13]);
4888 	ANON_LOCK_EXIT(&amp->a_rwlock);
4889 	kmem_free(ppa, ppasize);
4890 	return (0);
4891 error:
4892 	VM_STAT_ADD(segvnvmstats.fltanpages[14]);
4893 	ANON_LOCK_EXIT(&amp->a_rwlock);
4894 	kmem_free(ppa, ppasize);
4895 	if (type == F_SOFTLOCK && a > lpgaddr) {
4896 		VM_STAT_ADD(segvnvmstats.fltanpages[15]);
4897 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4898 	}
4899 	return (err);
4900 }
4901 
4902 int fltadvice = 1;	/* set to free behind pages for sequential access */
4903 
4904 /*
4905  * This routine is called via a machine specific fault handling routine.
4906  * It is also called by software routines wishing to lock or unlock
4907  * a range of addresses.
4908  *
4909  * Here is the basic algorithm:
4910  *	If unlocking
4911  *		Call segvn_softunlock
4912  *		Return
4913  *	endif
4914  *	Checking and set up work
4915  *	If we will need some non-anonymous pages
4916  *		Call VOP_GETPAGE over the range of non-anonymous pages
4917  *	endif
4918  *	Loop over all addresses requested
4919  *		Call segvn_faultpage passing in page list
4920  *		    to load up translations and handle anonymous pages
4921  *	endloop
4922  *	Load up translation to any additional pages in page list not
4923  *	    already handled that fit into this segment
4924  */
4925 static faultcode_t
4926 segvn_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
4927     enum fault_type type, enum seg_rw rw)
4928 {
4929 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4930 	page_t **plp, **ppp, *pp;
4931 	u_offset_t off;
4932 	caddr_t a;
4933 	struct vpage *vpage;
4934 	uint_t vpprot, prot;
4935 	int err;
4936 	page_t *pl[PVN_GETPAGE_NUM + 1];
4937 	size_t plsz, pl_alloc_sz;
4938 	size_t page;
4939 	ulong_t anon_index;
4940 	struct anon_map *amp;
4941 	int dogetpage = 0;
4942 	caddr_t	lpgaddr, lpgeaddr;
4943 	size_t pgsz;
4944 	anon_sync_obj_t cookie;
4945 	int brkcow = BREAK_COW_SHARE(rw, type, svd->type);
4946 
4947 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
4948 	ASSERT(svd->amp == NULL || svd->rcookie == HAT_INVALID_REGION_COOKIE);
4949 
4950 	/*
4951 	 * First handle the easy stuff
4952 	 */
4953 	if (type == F_SOFTUNLOCK) {
4954 		if (rw == S_READ_NOCOW) {
4955 			rw = S_READ;
4956 			ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
4957 		}
4958 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4959 		pgsz = (seg->s_szc == 0) ? PAGESIZE :
4960 		    page_get_pagesize(seg->s_szc);
4961 		VM_STAT_COND_ADD(pgsz > PAGESIZE, segvnvmstats.fltanpages[16]);
4962 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
4963 		segvn_softunlock(seg, lpgaddr, lpgeaddr - lpgaddr, rw);
4964 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4965 		return (0);
4966 	}
4967 
4968 	ASSERT(svd->tr_state == SEGVN_TR_OFF ||
4969 	    !HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
4970 	if (brkcow == 0) {
4971 		if (svd->tr_state == SEGVN_TR_INIT) {
4972 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4973 			if (svd->tr_state == SEGVN_TR_INIT) {
4974 				ASSERT(svd->vp != NULL && svd->amp == NULL);
4975 				ASSERT(svd->flags & MAP_TEXT);
4976 				ASSERT(svd->type == MAP_PRIVATE);
4977 				segvn_textrepl(seg);
4978 				ASSERT(svd->tr_state != SEGVN_TR_INIT);
4979 				ASSERT(svd->tr_state != SEGVN_TR_ON ||
4980 				    svd->amp != NULL);
4981 			}
4982 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4983 		}
4984 	} else if (svd->tr_state != SEGVN_TR_OFF) {
4985 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4986 
4987 		if (rw == S_WRITE && svd->tr_state != SEGVN_TR_OFF) {
4988 			ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
4989 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4990 			return (FC_PROT);
4991 		}
4992 
4993 		if (svd->tr_state == SEGVN_TR_ON) {
4994 			ASSERT(svd->vp != NULL && svd->amp != NULL);
4995 			segvn_textunrepl(seg, 0);
4996 			ASSERT(svd->amp == NULL &&
4997 			    svd->tr_state == SEGVN_TR_OFF);
4998 		} else if (svd->tr_state != SEGVN_TR_OFF) {
4999 			svd->tr_state = SEGVN_TR_OFF;
5000 		}
5001 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5002 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5003 	}
5004 
5005 top:
5006 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5007 
5008 	/*
5009 	 * If we have the same protections for the entire segment,
5010 	 * insure that the access being attempted is legitimate.
5011 	 */
5012 
5013 	if (svd->pageprot == 0) {
5014 		uint_t protchk;
5015 
5016 		switch (rw) {
5017 		case S_READ:
5018 		case S_READ_NOCOW:
5019 			protchk = PROT_READ;
5020 			break;
5021 		case S_WRITE:
5022 			protchk = PROT_WRITE;
5023 			break;
5024 		case S_EXEC:
5025 			protchk = PROT_EXEC;
5026 			break;
5027 		case S_OTHER:
5028 		default:
5029 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
5030 			break;
5031 		}
5032 
5033 		if ((svd->prot & protchk) == 0) {
5034 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5035 			return (FC_PROT);	/* illegal access type */
5036 		}
5037 	}
5038 
5039 	if (brkcow && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5040 		/* this must be SOFTLOCK S_READ fault */
5041 		ASSERT(svd->amp == NULL);
5042 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5043 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5044 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5045 		if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5046 			/*
5047 			 * this must be the first ever non S_READ_NOCOW
5048 			 * softlock for this segment.
5049 			 */
5050 			ASSERT(svd->softlockcnt == 0);
5051 			hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5052 			    HAT_REGION_TEXT);
5053 			svd->rcookie = HAT_INVALID_REGION_COOKIE;
5054 		}
5055 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5056 		goto top;
5057 	}
5058 
5059 	/*
5060 	 * We can't allow the long term use of softlocks for vmpss segments,
5061 	 * because in some file truncation cases we should be able to demote
5062 	 * the segment, which requires that there are no softlocks.  The
5063 	 * only case where it's ok to allow a SOFTLOCK fault against a vmpss
5064 	 * segment is S_READ_NOCOW, where the caller holds the address space
5065 	 * locked as writer and calls softunlock before dropping the as lock.
5066 	 * S_READ_NOCOW is used by /proc to read memory from another user.
5067 	 *
5068 	 * Another deadlock between SOFTLOCK and file truncation can happen
5069 	 * because segvn_fault_vnodepages() calls the FS one pagesize at
5070 	 * a time. A second VOP_GETPAGE() call by segvn_fault_vnodepages()
5071 	 * can cause a deadlock because the first set of page_t's remain
5072 	 * locked SE_SHARED.  To avoid this, we demote segments on a first
5073 	 * SOFTLOCK if they have a length greater than the segment's
5074 	 * page size.
5075 	 *
5076 	 * So for now, we only avoid demoting a segment on a SOFTLOCK when
5077 	 * the access type is S_READ_NOCOW and the fault length is less than
5078 	 * or equal to the segment's page size. While this is quite restrictive,
5079 	 * it should be the most common case of SOFTLOCK against a vmpss
5080 	 * segment.
5081 	 *
5082 	 * For S_READ_NOCOW, it's safe not to do a copy on write because the
5083 	 * caller makes sure no COW will be caused by another thread for a
5084 	 * softlocked page.
5085 	 */
5086 	if (type == F_SOFTLOCK && svd->vp != NULL && seg->s_szc != 0) {
5087 		int demote = 0;
5088 
5089 		if (rw != S_READ_NOCOW) {
5090 			demote = 1;
5091 		}
5092 		if (!demote && len > PAGESIZE) {
5093 			pgsz = page_get_pagesize(seg->s_szc);
5094 			CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr,
5095 			    lpgeaddr);
5096 			if (lpgeaddr - lpgaddr > pgsz) {
5097 				demote = 1;
5098 			}
5099 		}
5100 
5101 		ASSERT(demote || AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5102 
5103 		if (demote) {
5104 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5105 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5106 			if (seg->s_szc != 0) {
5107 				segvn_vmpss_clrszc_cnt++;
5108 				ASSERT(svd->softlockcnt == 0);
5109 				err = segvn_clrszc(seg);
5110 				if (err) {
5111 					segvn_vmpss_clrszc_err++;
5112 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5113 					return (FC_MAKE_ERR(err));
5114 				}
5115 			}
5116 			ASSERT(seg->s_szc == 0);
5117 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5118 			goto top;
5119 		}
5120 	}
5121 
5122 	/*
5123 	 * Check to see if we need to allocate an anon_map structure.
5124 	 */
5125 	if (svd->amp == NULL && (svd->vp == NULL || brkcow)) {
5126 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5127 		/*
5128 		 * Drop the "read" lock on the segment and acquire
5129 		 * the "write" version since we have to allocate the
5130 		 * anon_map.
5131 		 */
5132 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5133 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5134 
5135 		if (svd->amp == NULL) {
5136 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
5137 			svd->amp->a_szc = seg->s_szc;
5138 		}
5139 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5140 
5141 		/*
5142 		 * Start all over again since segment protections
5143 		 * may have changed after we dropped the "read" lock.
5144 		 */
5145 		goto top;
5146 	}
5147 
5148 	/*
5149 	 * S_READ_NOCOW vs S_READ distinction was
5150 	 * only needed for the code above. After
5151 	 * that we treat it as S_READ.
5152 	 */
5153 	if (rw == S_READ_NOCOW) {
5154 		ASSERT(type == F_SOFTLOCK);
5155 		ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5156 		rw = S_READ;
5157 	}
5158 
5159 	amp = svd->amp;
5160 
5161 	/*
5162 	 * MADV_SEQUENTIAL work is ignored for large page segments.
5163 	 */
5164 	if (seg->s_szc != 0) {
5165 		pgsz = page_get_pagesize(seg->s_szc);
5166 		ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
5167 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
5168 		if (svd->vp == NULL) {
5169 			err = segvn_fault_anonpages(hat, seg, lpgaddr,
5170 			    lpgeaddr, type, rw, addr, addr + len, brkcow);
5171 		} else {
5172 			err = segvn_fault_vnodepages(hat, seg, lpgaddr,
5173 				lpgeaddr, type, rw, addr, addr + len, brkcow);
5174 			if (err == IE_RETRY) {
5175 				ASSERT(seg->s_szc == 0);
5176 				ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
5177 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5178 				goto top;
5179 			}
5180 		}
5181 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5182 		return (err);
5183 	}
5184 
5185 	page = seg_page(seg, addr);
5186 	if (amp != NULL) {
5187 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5188 		anon_index = svd->anon_index + page;
5189 
5190 		if (type == F_PROT && rw == S_READ &&
5191 		    svd->tr_state == SEGVN_TR_OFF &&
5192 		    svd->type == MAP_PRIVATE && svd->pageprot == 0) {
5193 			size_t index = anon_index;
5194 			struct anon *ap;
5195 
5196 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5197 			/*
5198 			 * The fast path could apply to S_WRITE also, except
5199 			 * that the protection fault could be caused by lazy
5200 			 * tlb flush when ro->rw. In this case, the pte is
5201 			 * RW already. But RO in the other cpu's tlb causes
5202 			 * the fault. Since hat_chgprot won't do anything if
5203 			 * pte doesn't change, we may end up faulting
5204 			 * indefinitely until the RO tlb entry gets replaced.
5205 			 */
5206 			for (a = addr; a < addr + len; a += PAGESIZE, index++) {
5207 				anon_array_enter(amp, index, &cookie);
5208 				ap = anon_get_ptr(amp->ahp, index);
5209 				anon_array_exit(&cookie);
5210 				if ((ap == NULL) || (ap->an_refcnt != 1)) {
5211 					ANON_LOCK_EXIT(&amp->a_rwlock);
5212 					goto slow;
5213 				}
5214 			}
5215 			hat_chgprot(seg->s_as->a_hat, addr, len, svd->prot);
5216 			ANON_LOCK_EXIT(&amp->a_rwlock);
5217 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5218 			return (0);
5219 		}
5220 	}
5221 slow:
5222 
5223 	if (svd->vpage == NULL)
5224 		vpage = NULL;
5225 	else
5226 		vpage = &svd->vpage[page];
5227 
5228 	off = svd->offset + (uintptr_t)(addr - seg->s_base);
5229 
5230 	/*
5231 	 * If MADV_SEQUENTIAL has been set for the particular page we
5232 	 * are faulting on, free behind all pages in the segment and put
5233 	 * them on the free list.
5234 	 */
5235 
5236 	if ((page != 0) && fltadvice && svd->tr_state != SEGVN_TR_ON) {
5237 		struct vpage *vpp;
5238 		ulong_t fanon_index;
5239 		size_t fpage;
5240 		u_offset_t pgoff, fpgoff;
5241 		struct vnode *fvp;
5242 		struct anon *fap = NULL;
5243 
5244 		if (svd->advice == MADV_SEQUENTIAL ||
5245 		    (svd->pageadvice &&
5246 		    VPP_ADVICE(vpage) == MADV_SEQUENTIAL)) {
5247 			pgoff = off - PAGESIZE;
5248 			fpage = page - 1;
5249 			if (vpage != NULL)
5250 				vpp = &svd->vpage[fpage];
5251 			if (amp != NULL)
5252 				fanon_index = svd->anon_index + fpage;
5253 
5254 			while (pgoff > svd->offset) {
5255 				if (svd->advice != MADV_SEQUENTIAL &&
5256 				    (!svd->pageadvice || (vpage &&
5257 				    VPP_ADVICE(vpp) != MADV_SEQUENTIAL)))
5258 					break;
5259 
5260 				/*
5261 				 * If this is an anon page, we must find the
5262 				 * correct <vp, offset> for it
5263 				 */
5264 				fap = NULL;
5265 				if (amp != NULL) {
5266 					ANON_LOCK_ENTER(&amp->a_rwlock,
5267 						RW_READER);
5268 					anon_array_enter(amp, fanon_index,
5269 						&cookie);
5270 					fap = anon_get_ptr(amp->ahp,
5271 					    fanon_index);
5272 					if (fap != NULL) {
5273 						swap_xlate(fap, &fvp, &fpgoff);
5274 					} else {
5275 						fpgoff = pgoff;
5276 						fvp = svd->vp;
5277 					}
5278 					anon_array_exit(&cookie);
5279 					ANON_LOCK_EXIT(&amp->a_rwlock);
5280 				} else {
5281 					fpgoff = pgoff;
5282 					fvp = svd->vp;
5283 				}
5284 				if (fvp == NULL)
5285 					break;	/* XXX */
5286 				/*
5287 				 * Skip pages that are free or have an
5288 				 * "exclusive" lock.
5289 				 */
5290 				pp = page_lookup_nowait(fvp, fpgoff, SE_SHARED);
5291 				if (pp == NULL)
5292 					break;
5293 				/*
5294 				 * We don't need the page_struct_lock to test
5295 				 * as this is only advisory; even if we
5296 				 * acquire it someone might race in and lock
5297 				 * the page after we unlock and before the
5298 				 * PUTPAGE, then VOP_PUTPAGE will do nothing.
5299 				 */
5300 				if (pp->p_lckcnt == 0 && pp->p_cowcnt == 0) {
5301 					/*
5302 					 * Hold the vnode before releasing
5303 					 * the page lock to prevent it from
5304 					 * being freed and re-used by some
5305 					 * other thread.
5306 					 */
5307 					VN_HOLD(fvp);
5308 					page_unlock(pp);
5309 					/*
5310 					 * We should build a page list
5311 					 * to kluster putpages XXX
5312 					 */
5313 					(void) VOP_PUTPAGE(fvp,
5314 					    (offset_t)fpgoff, PAGESIZE,
5315 					    (B_DONTNEED|B_FREE|B_ASYNC),
5316 					    svd->cred, NULL);
5317 					VN_RELE(fvp);
5318 				} else {
5319 					/*
5320 					 * XXX - Should the loop terminate if
5321 					 * the page is `locked'?
5322 					 */
5323 					page_unlock(pp);
5324 				}
5325 				--vpp;
5326 				--fanon_index;
5327 				pgoff -= PAGESIZE;
5328 			}
5329 		}
5330 	}
5331 
5332 	plp = pl;
5333 	*plp = NULL;
5334 	pl_alloc_sz = 0;
5335 
5336 	/*
5337 	 * See if we need to call VOP_GETPAGE for
5338 	 * *any* of the range being faulted on.
5339 	 * We can skip all of this work if there
5340 	 * was no original vnode.
5341 	 */
5342 	if (svd->vp != NULL) {
5343 		u_offset_t vp_off;
5344 		size_t vp_len;
5345 		struct anon *ap;
5346 		vnode_t *vp;
5347 
5348 		vp_off = off;
5349 		vp_len = len;
5350 
5351 		if (amp == NULL)
5352 			dogetpage = 1;
5353 		else {
5354 			/*
5355 			 * Only acquire reader lock to prevent amp->ahp
5356 			 * from being changed.  It's ok to miss pages,
5357 			 * hence we don't do anon_array_enter
5358 			 */
5359 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5360 			ap = anon_get_ptr(amp->ahp, anon_index);
5361 
5362 			if (len <= PAGESIZE)
5363 				/* inline non_anon() */
5364 				dogetpage = (ap == NULL);
5365 			else
5366 				dogetpage = non_anon(amp->ahp, anon_index,
5367 				    &vp_off, &vp_len);
5368 			ANON_LOCK_EXIT(&amp->a_rwlock);
5369 		}
5370 
5371 		if (dogetpage) {
5372 			enum seg_rw arw;
5373 			struct as *as = seg->s_as;
5374 
5375 			if (len > ptob((sizeof (pl) / sizeof (pl[0])) - 1)) {
5376 				/*
5377 				 * Page list won't fit in local array,
5378 				 * allocate one of the needed size.
5379 				 */
5380 				pl_alloc_sz =
5381 				    (btop(len) + 1) * sizeof (page_t *);
5382 				plp = kmem_alloc(pl_alloc_sz, KM_SLEEP);
5383 				plp[0] = NULL;
5384 				plsz = len;
5385 			} else if (rw == S_WRITE && svd->type == MAP_PRIVATE ||
5386 			    svd->tr_state == SEGVN_TR_ON || rw == S_OTHER ||
5387 			    (((size_t)(addr + PAGESIZE) <
5388 			    (size_t)(seg->s_base + seg->s_size)) &&
5389 			    hat_probe(as->a_hat, addr + PAGESIZE))) {
5390 				/*
5391 				 * Ask VOP_GETPAGE to return the exact number
5392 				 * of pages if
5393 				 * (a) this is a COW fault, or
5394 				 * (b) this is a software fault, or
5395 				 * (c) next page is already mapped.
5396 				 */
5397 				plsz = len;
5398 			} else {
5399 				/*
5400 				 * Ask VOP_GETPAGE to return adjacent pages
5401 				 * within the segment.
5402 				 */
5403 				plsz = MIN((size_t)PVN_GETPAGE_SZ, (size_t)
5404 					((seg->s_base + seg->s_size) - addr));
5405 				ASSERT((addr + plsz) <=
5406 				    (seg->s_base + seg->s_size));
5407 			}
5408 
5409 			/*
5410 			 * Need to get some non-anonymous pages.
5411 			 * We need to make only one call to GETPAGE to do
5412 			 * this to prevent certain deadlocking conditions
5413 			 * when we are doing locking.  In this case
5414 			 * non_anon() should have picked up the smallest
5415 			 * range which includes all the non-anonymous
5416 			 * pages in the requested range.  We have to
5417 			 * be careful regarding which rw flag to pass in
5418 			 * because on a private mapping, the underlying
5419 			 * object is never allowed to be written.
5420 			 */
5421 			if (rw == S_WRITE && svd->type == MAP_PRIVATE) {
5422 				arw = S_READ;
5423 			} else {
5424 				arw = rw;
5425 			}
5426 			vp = svd->vp;
5427 			TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5428 				"segvn_getpage:seg %p addr %p vp %p",
5429 				seg, addr, vp);
5430 			err = VOP_GETPAGE(vp, (offset_t)vp_off, vp_len,
5431 			    &vpprot, plp, plsz, seg, addr + (vp_off - off), arw,
5432 			    svd->cred, NULL);
5433 			if (err) {
5434 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5435 				segvn_pagelist_rele(plp);
5436 				if (pl_alloc_sz)
5437 					kmem_free(plp, pl_alloc_sz);
5438 				return (FC_MAKE_ERR(err));
5439 			}
5440 			if (svd->type == MAP_PRIVATE)
5441 				vpprot &= ~PROT_WRITE;
5442 		}
5443 	}
5444 
5445 	/*
5446 	 * N.B. at this time the plp array has all the needed non-anon
5447 	 * pages in addition to (possibly) having some adjacent pages.
5448 	 */
5449 
5450 	/*
5451 	 * Always acquire the anon_array_lock to prevent
5452 	 * 2 threads from allocating separate anon slots for
5453 	 * the same "addr".
5454 	 *
5455 	 * If this is a copy-on-write fault and we don't already
5456 	 * have the anon_array_lock, acquire it to prevent the
5457 	 * fault routine from handling multiple copy-on-write faults
5458 	 * on the same "addr" in the same address space.
5459 	 *
5460 	 * Only one thread should deal with the fault since after
5461 	 * it is handled, the other threads can acquire a translation
5462 	 * to the newly created private page.  This prevents two or
5463 	 * more threads from creating different private pages for the
5464 	 * same fault.
5465 	 *
5466 	 * We grab "serialization" lock here if this is a MAP_PRIVATE segment
5467 	 * to prevent deadlock between this thread and another thread
5468 	 * which has soft-locked this page and wants to acquire serial_lock.
5469 	 * ( bug 4026339 )
5470 	 *
5471 	 * The fix for bug 4026339 becomes unnecessary when using the
5472 	 * locking scheme with per amp rwlock and a global set of hash
5473 	 * lock, anon_array_lock.  If we steal a vnode page when low
5474 	 * on memory and upgrad the page lock through page_rename,
5475 	 * then the page is PAGE_HANDLED, nothing needs to be done
5476 	 * for this page after returning from segvn_faultpage.
5477 	 *
5478 	 * But really, the page lock should be downgraded after
5479 	 * the stolen page is page_rename'd.
5480 	 */
5481 
5482 	if (amp != NULL)
5483 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5484 
5485 	/*
5486 	 * Ok, now loop over the address range and handle faults
5487 	 */
5488 	for (a = addr; a < addr + len; a += PAGESIZE, off += PAGESIZE) {
5489 		err = segvn_faultpage(hat, seg, a, off, vpage, plp, vpprot,
5490 		    type, rw, brkcow, a == addr);
5491 		if (err) {
5492 			if (amp != NULL)
5493 				ANON_LOCK_EXIT(&amp->a_rwlock);
5494 			if (type == F_SOFTLOCK && a > addr) {
5495 				segvn_softunlock(seg, addr, (a - addr),
5496 				    S_OTHER);
5497 			}
5498 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5499 			segvn_pagelist_rele(plp);
5500 			if (pl_alloc_sz)
5501 				kmem_free(plp, pl_alloc_sz);
5502 			return (err);
5503 		}
5504 		if (vpage) {
5505 			vpage++;
5506 		} else if (svd->vpage) {
5507 			page = seg_page(seg, addr);
5508 			vpage = &svd->vpage[++page];
5509 		}
5510 	}
5511 
5512 	/* Didn't get pages from the underlying fs so we're done */
5513 	if (!dogetpage)
5514 		goto done;
5515 
5516 	/*
5517 	 * Now handle any other pages in the list returned.
5518 	 * If the page can be used, load up the translations now.
5519 	 * Note that the for loop will only be entered if "plp"
5520 	 * is pointing to a non-NULL page pointer which means that
5521 	 * VOP_GETPAGE() was called and vpprot has been initialized.
5522 	 */
5523 	if (svd->pageprot == 0)
5524 		prot = svd->prot & vpprot;
5525 
5526 
5527 	/*
5528 	 * Large Files: diff should be unsigned value because we started
5529 	 * supporting > 2GB segment sizes from 2.5.1 and when a
5530 	 * large file of size > 2GB gets mapped to address space
5531 	 * the diff value can be > 2GB.
5532 	 */
5533 
5534 	for (ppp = plp; (pp = *ppp) != NULL; ppp++) {
5535 		size_t diff;
5536 		struct anon *ap;
5537 		int anon_index;
5538 		anon_sync_obj_t cookie;
5539 		int hat_flag = HAT_LOAD_ADV;
5540 
5541 		if (svd->flags & MAP_TEXT) {
5542 			hat_flag |= HAT_LOAD_TEXT;
5543 		}
5544 
5545 		if (pp == PAGE_HANDLED)
5546 			continue;
5547 
5548 		if (svd->tr_state != SEGVN_TR_ON &&
5549 		    pp->p_offset >=  svd->offset &&
5550 		    pp->p_offset < svd->offset + seg->s_size) {
5551 
5552 			diff = pp->p_offset - svd->offset;
5553 
5554 			/*
5555 			 * Large Files: Following is the assertion
5556 			 * validating the above cast.
5557 			 */
5558 			ASSERT(svd->vp == pp->p_vnode);
5559 
5560 			page = btop(diff);
5561 			if (svd->pageprot)
5562 				prot = VPP_PROT(&svd->vpage[page]) & vpprot;
5563 
5564 			/*
5565 			 * Prevent other threads in the address space from
5566 			 * creating private pages (i.e., allocating anon slots)
5567 			 * while we are in the process of loading translations
5568 			 * to additional pages returned by the underlying
5569 			 * object.
5570 			 */
5571 			if (amp != NULL) {
5572 				anon_index = svd->anon_index + page;
5573 				anon_array_enter(amp, anon_index, &cookie);
5574 				ap = anon_get_ptr(amp->ahp, anon_index);
5575 			}
5576 			if ((amp == NULL) || (ap == NULL)) {
5577 				if (IS_VMODSORT(pp->p_vnode) ||
5578 				    enable_mbit_wa) {
5579 					if (rw == S_WRITE)
5580 						hat_setmod(pp);
5581 					else if (rw != S_OTHER &&
5582 					    !hat_ismod(pp))
5583 						prot &= ~PROT_WRITE;
5584 				}
5585 				/*
5586 				 * Skip mapping read ahead pages marked
5587 				 * for migration, so they will get migrated
5588 				 * properly on fault
5589 				 */
5590 				ASSERT(amp == NULL ||
5591 				    svd->rcookie == HAT_INVALID_REGION_COOKIE);
5592 				if ((prot & PROT_READ) && !PP_ISMIGRATE(pp)) {
5593 					hat_memload_region(hat,
5594 					    seg->s_base + diff,
5595 					    pp, prot, hat_flag,
5596 					    svd->rcookie);
5597 				}
5598 			}
5599 			if (amp != NULL)
5600 				anon_array_exit(&cookie);
5601 		}
5602 		page_unlock(pp);
5603 	}
5604 done:
5605 	if (amp != NULL)
5606 		ANON_LOCK_EXIT(&amp->a_rwlock);
5607 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5608 	if (pl_alloc_sz)
5609 		kmem_free(plp, pl_alloc_sz);
5610 	return (0);
5611 }
5612 
5613 /*
5614  * This routine is used to start I/O on pages asynchronously.  XXX it will
5615  * only create PAGESIZE pages. At fault time they will be relocated into
5616  * larger pages.
5617  */
5618 static faultcode_t
5619 segvn_faulta(struct seg *seg, caddr_t addr)
5620 {
5621 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5622 	int err;
5623 	struct anon_map *amp;
5624 	vnode_t *vp;
5625 
5626 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5627 
5628 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5629 	if ((amp = svd->amp) != NULL) {
5630 		struct anon *ap;
5631 
5632 		/*
5633 		 * Reader lock to prevent amp->ahp from being changed.
5634 		 * This is advisory, it's ok to miss a page, so
5635 		 * we don't do anon_array_enter lock.
5636 		 */
5637 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5638 		if ((ap = anon_get_ptr(amp->ahp,
5639 			svd->anon_index + seg_page(seg, addr))) != NULL) {
5640 
5641 			err = anon_getpage(&ap, NULL, NULL,
5642 			    0, seg, addr, S_READ, svd->cred);
5643 
5644 			ANON_LOCK_EXIT(&amp->a_rwlock);
5645 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5646 			if (err)
5647 				return (FC_MAKE_ERR(err));
5648 			return (0);
5649 		}
5650 		ANON_LOCK_EXIT(&amp->a_rwlock);
5651 	}
5652 
5653 	if (svd->vp == NULL) {
5654 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5655 		return (0);			/* zfod page - do nothing now */
5656 	}
5657 
5658 	vp = svd->vp;
5659 	TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5660 		"segvn_getpage:seg %p addr %p vp %p", seg, addr, vp);
5661 	err = VOP_GETPAGE(vp,
5662 	    (offset_t)(svd->offset + (uintptr_t)(addr - seg->s_base)),
5663 	    PAGESIZE, NULL, NULL, 0, seg, addr,
5664 	    S_OTHER, svd->cred, NULL);
5665 
5666 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5667 	if (err)
5668 		return (FC_MAKE_ERR(err));
5669 	return (0);
5670 }
5671 
5672 static int
5673 segvn_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
5674 {
5675 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5676 	struct vpage *svp, *evp;
5677 	struct vnode *vp;
5678 	size_t pgsz;
5679 	pgcnt_t pgcnt;
5680 	anon_sync_obj_t cookie;
5681 	int unload_done = 0;
5682 
5683 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5684 
5685 	if ((svd->maxprot & prot) != prot)
5686 		return (EACCES);			/* violated maxprot */
5687 
5688 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5689 
5690 	/* return if prot is the same */
5691 	if (!svd->pageprot && svd->prot == prot) {
5692 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5693 		return (0);
5694 	}
5695 
5696 	/*
5697 	 * Since we change protections we first have to flush the cache.
5698 	 * This makes sure all the pagelock calls have to recheck
5699 	 * protections.
5700 	 */
5701 	if (svd->softlockcnt > 0) {
5702 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5703 		/*
5704 		 * Since we do have the segvn writers lock nobody can fill
5705 		 * the cache with entries belonging to this seg during
5706 		 * the purge. The flush either succeeds or we still have
5707 		 * pending I/Os.
5708 		 */
5709 		segvn_purge(seg);
5710 		if (svd->softlockcnt > 0) {
5711 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5712 			return (EAGAIN);
5713 		}
5714 	}
5715 
5716 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5717 		ASSERT(svd->amp == NULL);
5718 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5719 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5720 		    HAT_REGION_TEXT);
5721 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
5722 		unload_done = 1;
5723 	} else if (svd->tr_state == SEGVN_TR_INIT) {
5724 		svd->tr_state = SEGVN_TR_OFF;
5725 	} else if (svd->tr_state == SEGVN_TR_ON) {
5726 		ASSERT(svd->amp != NULL);
5727 		segvn_textunrepl(seg, 0);
5728 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5729 		unload_done = 1;
5730 	}
5731 
5732 	if ((prot & PROT_WRITE) && svd->type == MAP_SHARED &&
5733 	    svd->vp != NULL && (svd->vp->v_flag & VVMEXEC)) {
5734 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
5735 		segvn_inval_trcache(svd->vp);
5736 	}
5737 	if (seg->s_szc != 0) {
5738 		int err;
5739 		pgsz = page_get_pagesize(seg->s_szc);
5740 		pgcnt = pgsz >> PAGESHIFT;
5741 		ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
5742 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
5743 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5744 			ASSERT(seg->s_base != addr || seg->s_size != len);
5745 			/*
5746 			 * If we are holding the as lock as a reader then
5747 			 * we need to return IE_RETRY and let the as
5748 			 * layer drop and re-acquire the lock as a writer.
5749 			 */
5750 			if (AS_READ_HELD(seg->s_as, &seg->s_as->a_lock))
5751 				return (IE_RETRY);
5752 			VM_STAT_ADD(segvnvmstats.demoterange[1]);
5753 			if (svd->type == MAP_PRIVATE || svd->vp != NULL) {
5754 				err = segvn_demote_range(seg, addr, len,
5755 				    SDR_END, 0);
5756 			} else {
5757 				uint_t szcvec = map_pgszcvec(seg->s_base,
5758 				    pgsz, (uintptr_t)seg->s_base,
5759 				    (svd->flags & MAP_TEXT), MAPPGSZC_SHM, 0);
5760 				err = segvn_demote_range(seg, addr, len,
5761 				    SDR_END, szcvec);
5762 			}
5763 			if (err == 0)
5764 				return (IE_RETRY);
5765 			if (err == ENOMEM)
5766 				return (IE_NOMEM);
5767 			return (err);
5768 		}
5769 	}
5770 
5771 
5772 	/*
5773 	 * If it's a private mapping and we're making it writable
5774 	 * and no swap space has been reserved, have to reserve
5775 	 * it all now.  If it's a private mapping to a file (i.e., vp != NULL)
5776 	 * and we're removing write permission on the entire segment and
5777 	 * we haven't modified any pages, we can release the swap space.
5778 	 */
5779 	if (svd->type == MAP_PRIVATE) {
5780 		if (prot & PROT_WRITE) {
5781 			size_t sz;
5782 			if (svd->swresv == 0 && !(svd->flags & MAP_NORESERVE)) {
5783 				if (anon_resv_zone(seg->s_size,
5784 				    seg->s_as->a_proc->p_zone) == 0) {
5785 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5786 					return (IE_NOMEM);
5787 				}
5788 				sz = svd->swresv = seg->s_size;
5789 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
5790 					"anon proc:%p %lu %u",
5791 					seg, sz, 1);
5792 			}
5793 		} else {
5794 			/*
5795 			 * Swap space is released only if this segment
5796 			 * does not map anonymous memory, since read faults
5797 			 * on such segments still need an anon slot to read
5798 			 * in the data.
5799 			 */
5800 			if (svd->swresv != 0 && svd->vp != NULL &&
5801 			    svd->amp == NULL && addr == seg->s_base &&
5802 			    len == seg->s_size && svd->pageprot == 0) {
5803 				anon_unresv_zone(svd->swresv,
5804 				    seg->s_as->a_proc->p_zone);
5805 				svd->swresv = 0;
5806 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
5807 					"anon proc:%p %lu %u",
5808 					seg, 0, 0);
5809 			}
5810 		}
5811 	}
5812 
5813 	if (addr == seg->s_base && len == seg->s_size && svd->vpage == NULL) {
5814 		if (svd->prot == prot) {
5815 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5816 			return (0);			/* all done */
5817 		}
5818 		svd->prot = (uchar_t)prot;
5819 	} else if (svd->type == MAP_PRIVATE) {
5820 		struct anon *ap = NULL;
5821 		page_t *pp;
5822 		u_offset_t offset, off;
5823 		struct anon_map *amp;
5824 		ulong_t anon_idx = 0;
5825 
5826 		/*
5827 		 * A vpage structure exists or else the change does not
5828 		 * involve the entire segment.  Establish a vpage structure
5829 		 * if none is there.  Then, for each page in the range,
5830 		 * adjust its individual permissions.  Note that write-
5831 		 * enabling a MAP_PRIVATE page can affect the claims for
5832 		 * locked down memory.  Overcommitting memory terminates
5833 		 * the operation.
5834 		 */
5835 		segvn_vpage(seg);
5836 		svd->pageprot = 1;
5837 		if ((amp = svd->amp) != NULL) {
5838 			anon_idx = svd->anon_index + seg_page(seg, addr);
5839 			ASSERT(seg->s_szc == 0 ||
5840 			    IS_P2ALIGNED(anon_idx, pgcnt));
5841 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5842 		}
5843 
5844 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
5845 		evp = &svd->vpage[seg_page(seg, addr + len)];
5846 
5847 		/*
5848 		 * See Statement at the beginning of segvn_lockop regarding
5849 		 * the way cowcnts and lckcnts are handled.
5850 		 */
5851 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5852 
5853 			if (seg->s_szc != 0) {
5854 				if (amp != NULL) {
5855 					anon_array_enter(amp, anon_idx,
5856 					    &cookie);
5857 				}
5858 				if (IS_P2ALIGNED(anon_idx, pgcnt) &&
5859 				    !segvn_claim_pages(seg, svp, offset,
5860 					anon_idx, prot)) {
5861 					if (amp != NULL) {
5862 						anon_array_exit(&cookie);
5863 					}
5864 					break;
5865 				}
5866 				if (amp != NULL) {
5867 					anon_array_exit(&cookie);
5868 				}
5869 				anon_idx++;
5870 			} else {
5871 				if (amp != NULL) {
5872 					anon_array_enter(amp, anon_idx,
5873 						&cookie);
5874 					ap = anon_get_ptr(amp->ahp, anon_idx++);
5875 				}
5876 
5877 				if (VPP_ISPPLOCK(svp) &&
5878 				    VPP_PROT(svp) != prot) {
5879 
5880 					if (amp == NULL || ap == NULL) {
5881 						vp = svd->vp;
5882 						off = offset;
5883 					} else
5884 						swap_xlate(ap, &vp, &off);
5885 					if (amp != NULL)
5886 						anon_array_exit(&cookie);
5887 
5888 					if ((pp = page_lookup(vp, off,
5889 					    SE_SHARED)) == NULL) {
5890 						panic("segvn_setprot: no page");
5891 						/*NOTREACHED*/
5892 					}
5893 					ASSERT(seg->s_szc == 0);
5894 					if ((VPP_PROT(svp) ^ prot) &
5895 					    PROT_WRITE) {
5896 						if (prot & PROT_WRITE) {
5897 						    if (!page_addclaim(pp)) {
5898 							page_unlock(pp);
5899 							break;
5900 						    }
5901 						} else {
5902 						    if (!page_subclaim(pp)) {
5903 							page_unlock(pp);
5904 							break;
5905 						    }
5906 						}
5907 					}
5908 					page_unlock(pp);
5909 				} else if (amp != NULL)
5910 					anon_array_exit(&cookie);
5911 			}
5912 			VPP_SETPROT(svp, prot);
5913 			offset += PAGESIZE;
5914 		}
5915 		if (amp != NULL)
5916 			ANON_LOCK_EXIT(&amp->a_rwlock);
5917 
5918 		/*
5919 		 * Did we terminate prematurely?  If so, simply unload
5920 		 * the translations to the things we've updated so far.
5921 		 */
5922 		if (svp != evp) {
5923 			if (unload_done) {
5924 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5925 				return (IE_NOMEM);
5926 			}
5927 			len = (svp - &svd->vpage[seg_page(seg, addr)]) *
5928 			    PAGESIZE;
5929 			ASSERT(seg->s_szc == 0 || IS_P2ALIGNED(len, pgsz));
5930 			if (len != 0)
5931 				hat_unload(seg->s_as->a_hat, addr,
5932 				    len, HAT_UNLOAD);
5933 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5934 			return (IE_NOMEM);
5935 		}
5936 	} else {
5937 		segvn_vpage(seg);
5938 		svd->pageprot = 1;
5939 		evp = &svd->vpage[seg_page(seg, addr + len)];
5940 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5941 			VPP_SETPROT(svp, prot);
5942 		}
5943 	}
5944 
5945 	if (unload_done) {
5946 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5947 		return (0);
5948 	}
5949 
5950 	if (((prot & PROT_WRITE) != 0 &&
5951 	    (svd->vp != NULL || svd->type == MAP_PRIVATE)) ||
5952 	    (prot & ~PROT_USER) == PROT_NONE) {
5953 		/*
5954 		 * Either private or shared data with write access (in
5955 		 * which case we need to throw out all former translations
5956 		 * so that we get the right translations set up on fault
5957 		 * and we don't allow write access to any copy-on-write pages
5958 		 * that might be around or to prevent write access to pages
5959 		 * representing holes in a file), or we don't have permission
5960 		 * to access the memory at all (in which case we have to
5961 		 * unload any current translations that might exist).
5962 		 */
5963 		hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
5964 	} else {
5965 		/*
5966 		 * A shared mapping or a private mapping in which write
5967 		 * protection is going to be denied - just change all the
5968 		 * protections over the range of addresses in question.
5969 		 * segvn does not support any other attributes other
5970 		 * than prot so we can use hat_chgattr.
5971 		 */
5972 		hat_chgattr(seg->s_as->a_hat, addr, len, prot);
5973 	}
5974 
5975 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5976 
5977 	return (0);
5978 }
5979 
5980 /*
5981  * segvn_setpagesize is called via SEGOP_SETPAGESIZE from as_setpagesize,
5982  * to determine if the seg is capable of mapping the requested szc.
5983  */
5984 static int
5985 segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
5986 {
5987 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5988 	struct segvn_data *nsvd;
5989 	struct anon_map *amp = svd->amp;
5990 	struct seg *nseg;
5991 	caddr_t eaddr = addr + len, a;
5992 	size_t pgsz = page_get_pagesize(szc);
5993 	pgcnt_t pgcnt = page_get_pagecnt(szc);
5994 	int err;
5995 	u_offset_t off = svd->offset + (uintptr_t)(addr - seg->s_base);
5996 	extern struct vnode kvp;
5997 
5998 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5999 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6000 
6001 	if (seg->s_szc == szc || segvn_lpg_disable != 0) {
6002 		return (0);
6003 	}
6004 
6005 	/*
6006 	 * addr should always be pgsz aligned but eaddr may be misaligned if
6007 	 * it's at the end of the segment.
6008 	 *
6009 	 * XXX we should assert this condition since as_setpagesize() logic
6010 	 * guarantees it.
6011 	 */
6012 	if (!IS_P2ALIGNED(addr, pgsz) ||
6013 	    (!IS_P2ALIGNED(eaddr, pgsz) &&
6014 		eaddr != seg->s_base + seg->s_size)) {
6015 
6016 		segvn_setpgsz_align_err++;
6017 		return (EINVAL);
6018 	}
6019 
6020 	if (amp != NULL && svd->type == MAP_SHARED) {
6021 		ulong_t an_idx = svd->anon_index + seg_page(seg, addr);
6022 		if (!IS_P2ALIGNED(an_idx, pgcnt)) {
6023 
6024 			segvn_setpgsz_anon_align_err++;
6025 			return (EINVAL);
6026 		}
6027 	}
6028 
6029 	if ((svd->flags & MAP_NORESERVE) || seg->s_as == &kas ||
6030 	    szc > segvn_maxpgszc) {
6031 		return (EINVAL);
6032 	}
6033 
6034 	/* paranoid check */
6035 	if (svd->vp != NULL &&
6036 	    (IS_SWAPFSVP(svd->vp) || VN_ISKAS(svd->vp))) {
6037 		    return (EINVAL);
6038 	}
6039 
6040 	if (seg->s_szc == 0 && svd->vp != NULL &&
6041 	    map_addr_vacalign_check(addr, off)) {
6042 		return (EINVAL);
6043 	}
6044 
6045 	/*
6046 	 * Check that protections are the same within new page
6047 	 * size boundaries.
6048 	 */
6049 	if (svd->pageprot) {
6050 		for (a = addr; a < eaddr; a += pgsz) {
6051 			if ((a + pgsz) > eaddr) {
6052 				if (!sameprot(seg, a, eaddr - a)) {
6053 					return (EINVAL);
6054 				}
6055 			} else {
6056 				if (!sameprot(seg, a, pgsz)) {
6057 					return (EINVAL);
6058 				}
6059 			}
6060 		}
6061 	}
6062 
6063 	/*
6064 	 * Since we are changing page size we first have to flush
6065 	 * the cache. This makes sure all the pagelock calls have
6066 	 * to recheck protections.
6067 	 */
6068 	if (svd->softlockcnt > 0) {
6069 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6070 		/*
6071 		 * Since we do have the segvn writers lock nobody can fill
6072 		 * the cache with entries belonging to this seg during
6073 		 * the purge. The flush either succeeds or we still have
6074 		 * pending I/Os.
6075 		 */
6076 		segvn_purge(seg);
6077 		if (svd->softlockcnt > 0) {
6078 			return (EAGAIN);
6079 		}
6080 	}
6081 
6082 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6083 		ASSERT(svd->amp == NULL);
6084 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6085 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6086 		    HAT_REGION_TEXT);
6087 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6088 	} else if (svd->tr_state == SEGVN_TR_INIT) {
6089 		svd->tr_state = SEGVN_TR_OFF;
6090 	} else if (svd->tr_state == SEGVN_TR_ON) {
6091 		ASSERT(svd->amp != NULL);
6092 		segvn_textunrepl(seg, 1);
6093 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6094 		amp = NULL;
6095 	}
6096 
6097 	/*
6098 	 * Operation for sub range of existing segment.
6099 	 */
6100 	if (addr != seg->s_base || eaddr != (seg->s_base + seg->s_size)) {
6101 		if (szc < seg->s_szc) {
6102 			VM_STAT_ADD(segvnvmstats.demoterange[2]);
6103 			err = segvn_demote_range(seg, addr, len, SDR_RANGE, 0);
6104 			if (err == 0) {
6105 				return (IE_RETRY);
6106 			}
6107 			if (err == ENOMEM) {
6108 				return (IE_NOMEM);
6109 			}
6110 			return (err);
6111 		}
6112 		if (addr != seg->s_base) {
6113 			nseg = segvn_split_seg(seg, addr);
6114 			if (eaddr != (nseg->s_base + nseg->s_size)) {
6115 				/* eaddr is szc aligned */
6116 				(void) segvn_split_seg(nseg, eaddr);
6117 			}
6118 			return (IE_RETRY);
6119 		}
6120 		if (eaddr != (seg->s_base + seg->s_size)) {
6121 			/* eaddr is szc aligned */
6122 			(void) segvn_split_seg(seg, eaddr);
6123 		}
6124 		return (IE_RETRY);
6125 	}
6126 
6127 	/*
6128 	 * Break any low level sharing and reset seg->s_szc to 0.
6129 	 */
6130 	if ((err = segvn_clrszc(seg)) != 0) {
6131 		if (err == ENOMEM) {
6132 			err = IE_NOMEM;
6133 		}
6134 		return (err);
6135 	}
6136 	ASSERT(seg->s_szc == 0);
6137 
6138 	/*
6139 	 * If the end of the current segment is not pgsz aligned
6140 	 * then attempt to concatenate with the next segment.
6141 	 */
6142 	if (!IS_P2ALIGNED(eaddr, pgsz)) {
6143 		nseg = AS_SEGNEXT(seg->s_as, seg);
6144 		if (nseg == NULL || nseg == seg || eaddr != nseg->s_base) {
6145 			return (ENOMEM);
6146 		}
6147 		if (nseg->s_ops != &segvn_ops) {
6148 			return (EINVAL);
6149 		}
6150 		nsvd = (struct segvn_data *)nseg->s_data;
6151 		if (nsvd->softlockcnt > 0) {
6152 			segvn_purge(nseg);
6153 			if (nsvd->softlockcnt > 0) {
6154 				return (EAGAIN);
6155 			}
6156 		}
6157 		err = segvn_clrszc(nseg);
6158 		if (err == ENOMEM) {
6159 			err = IE_NOMEM;
6160 		}
6161 		if (err != 0) {
6162 			return (err);
6163 		}
6164 		ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6165 		err = segvn_concat(seg, nseg, 1);
6166 		if (err == -1) {
6167 			return (EINVAL);
6168 		}
6169 		if (err == -2) {
6170 			return (IE_NOMEM);
6171 		}
6172 		return (IE_RETRY);
6173 	}
6174 
6175 	/*
6176 	 * May need to re-align anon array to
6177 	 * new szc.
6178 	 */
6179 	if (amp != NULL) {
6180 		if (!IS_P2ALIGNED(svd->anon_index, pgcnt)) {
6181 			struct anon_hdr *nahp;
6182 
6183 			ASSERT(svd->type == MAP_PRIVATE);
6184 
6185 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6186 			ASSERT(amp->refcnt == 1);
6187 			nahp = anon_create(btop(amp->size), ANON_NOSLEEP);
6188 			if (nahp == NULL) {
6189 				ANON_LOCK_EXIT(&amp->a_rwlock);
6190 				return (IE_NOMEM);
6191 			}
6192 			if (anon_copy_ptr(amp->ahp, svd->anon_index,
6193 				nahp, 0, btop(seg->s_size), ANON_NOSLEEP)) {
6194 				anon_release(nahp, btop(amp->size));
6195 				ANON_LOCK_EXIT(&amp->a_rwlock);
6196 				return (IE_NOMEM);
6197 			}
6198 			anon_release(amp->ahp, btop(amp->size));
6199 			amp->ahp = nahp;
6200 			svd->anon_index = 0;
6201 			ANON_LOCK_EXIT(&amp->a_rwlock);
6202 		}
6203 	}
6204 	if (svd->vp != NULL && szc != 0) {
6205 		struct vattr va;
6206 		u_offset_t eoffpage = svd->offset;
6207 		va.va_mask = AT_SIZE;
6208 		eoffpage += seg->s_size;
6209 		eoffpage = btopr(eoffpage);
6210 		if (VOP_GETATTR(svd->vp, &va, 0, svd->cred, NULL) != 0) {
6211 			segvn_setpgsz_getattr_err++;
6212 			return (EINVAL);
6213 		}
6214 		if (btopr(va.va_size) < eoffpage) {
6215 			segvn_setpgsz_eof_err++;
6216 			return (EINVAL);
6217 		}
6218 		if (amp != NULL) {
6219 			/*
6220 			 * anon_fill_cow_holes() may call VOP_GETPAGE().
6221 			 * don't take anon map lock here to avoid holding it
6222 			 * across VOP_GETPAGE() calls that may call back into
6223 			 * segvn for klsutering checks. We don't really need
6224 			 * anon map lock here since it's a private segment and
6225 			 * we hold as level lock as writers.
6226 			 */
6227 			if ((err = anon_fill_cow_holes(seg, seg->s_base,
6228 			    amp->ahp, svd->anon_index, svd->vp, svd->offset,
6229 			    seg->s_size, szc, svd->prot, svd->vpage,
6230 			    svd->cred)) != 0) {
6231 				return (EINVAL);
6232 			}
6233 		}
6234 		segvn_setvnode_mpss(svd->vp);
6235 	}
6236 
6237 	if (amp != NULL) {
6238 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6239 		if (svd->type == MAP_PRIVATE) {
6240 			amp->a_szc = szc;
6241 		} else if (szc > amp->a_szc) {
6242 			amp->a_szc = szc;
6243 		}
6244 		ANON_LOCK_EXIT(&amp->a_rwlock);
6245 	}
6246 
6247 	seg->s_szc = szc;
6248 
6249 	return (0);
6250 }
6251 
6252 static int
6253 segvn_clrszc(struct seg *seg)
6254 {
6255 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6256 	struct anon_map *amp = svd->amp;
6257 	size_t pgsz;
6258 	pgcnt_t pages;
6259 	int err = 0;
6260 	caddr_t a = seg->s_base;
6261 	caddr_t ea = a + seg->s_size;
6262 	ulong_t an_idx = svd->anon_index;
6263 	vnode_t *vp = svd->vp;
6264 	struct vpage *vpage = svd->vpage;
6265 	page_t *anon_pl[1 + 1], *pp;
6266 	struct anon *ap, *oldap;
6267 	uint_t prot = svd->prot, vpprot;
6268 	int pageflag = 0;
6269 
6270 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6271 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
6272 	ASSERT(svd->softlockcnt == 0);
6273 
6274 	if (vp == NULL && amp == NULL) {
6275 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6276 		seg->s_szc = 0;
6277 		return (0);
6278 	}
6279 
6280 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6281 		ASSERT(svd->amp == NULL);
6282 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6283 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6284 		    HAT_REGION_TEXT);
6285 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6286 	} else if (svd->tr_state == SEGVN_TR_ON) {
6287 		ASSERT(svd->amp != NULL);
6288 		segvn_textunrepl(seg, 1);
6289 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6290 		amp = NULL;
6291 	} else {
6292 		if (svd->tr_state != SEGVN_TR_OFF) {
6293 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
6294 			svd->tr_state = SEGVN_TR_OFF;
6295 		}
6296 
6297 		/*
6298 		 * do HAT_UNLOAD_UNMAP since we are changing the pagesize.
6299 		 * unload argument is 0 when we are freeing the segment
6300 		 * and unload was already done.
6301 		 */
6302 		hat_unload(seg->s_as->a_hat, seg->s_base, seg->s_size,
6303 		    HAT_UNLOAD_UNMAP);
6304 	}
6305 
6306 	if (amp == NULL || svd->type == MAP_SHARED) {
6307 		seg->s_szc = 0;
6308 		return (0);
6309 	}
6310 
6311 	pgsz = page_get_pagesize(seg->s_szc);
6312 	pages = btop(pgsz);
6313 
6314 	/*
6315 	 * XXX anon rwlock is not really needed because this is a
6316 	 * private segment and we are writers.
6317 	 */
6318 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6319 
6320 	for (; a < ea; a += pgsz, an_idx += pages) {
6321 		if ((oldap = anon_get_ptr(amp->ahp, an_idx)) != NULL) {
6322 			ASSERT(vpage != NULL || svd->pageprot == 0);
6323 			if (vpage != NULL) {
6324 				ASSERT(sameprot(seg, a, pgsz));
6325 				prot = VPP_PROT(vpage);
6326 				pageflag = VPP_ISPPLOCK(vpage) ? LOCK_PAGE : 0;
6327 			}
6328 			if (seg->s_szc != 0) {
6329 				ASSERT(vp == NULL || anon_pages(amp->ahp,
6330 				    an_idx, pages) == pages);
6331 				if ((err = anon_map_demotepages(amp, an_idx,
6332 				    seg, a, prot, vpage, svd->cred)) != 0) {
6333 					goto out;
6334 				}
6335 			} else {
6336 				if (oldap->an_refcnt == 1) {
6337 					continue;
6338 				}
6339 				if ((err = anon_getpage(&oldap, &vpprot,
6340 				    anon_pl, PAGESIZE, seg, a, S_READ,
6341 				    svd->cred))) {
6342 					goto out;
6343 				}
6344 				if ((pp = anon_private(&ap, seg, a, prot,
6345 				    anon_pl[0], pageflag, svd->cred)) == NULL) {
6346 					err = ENOMEM;
6347 					goto out;
6348 				}
6349 				anon_decref(oldap);
6350 				(void) anon_set_ptr(amp->ahp, an_idx, ap,
6351 				    ANON_SLEEP);
6352 				page_unlock(pp);
6353 			}
6354 		}
6355 		vpage = (vpage == NULL) ? NULL : vpage + pages;
6356 	}
6357 
6358 	amp->a_szc = 0;
6359 	seg->s_szc = 0;
6360 out:
6361 	ANON_LOCK_EXIT(&amp->a_rwlock);
6362 	return (err);
6363 }
6364 
6365 static int
6366 segvn_claim_pages(
6367 	struct seg *seg,
6368 	struct vpage *svp,
6369 	u_offset_t off,
6370 	ulong_t anon_idx,
6371 	uint_t prot)
6372 {
6373 	pgcnt_t	pgcnt = page_get_pagecnt(seg->s_szc);
6374 	size_t ppasize = (pgcnt + 1) * sizeof (page_t *);
6375 	page_t	**ppa;
6376 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6377 	struct anon_map *amp = svd->amp;
6378 	struct vpage *evp = svp + pgcnt;
6379 	caddr_t addr = ((uintptr_t)(svp - svd->vpage) << PAGESHIFT)
6380 	    + seg->s_base;
6381 	struct anon *ap;
6382 	struct vnode *vp = svd->vp;
6383 	page_t *pp;
6384 	pgcnt_t pg_idx, i;
6385 	int err = 0;
6386 	anoff_t aoff;
6387 	int anon = (amp != NULL) ? 1 : 0;
6388 
6389 	ASSERT(svd->type == MAP_PRIVATE);
6390 	ASSERT(svd->vpage != NULL);
6391 	ASSERT(seg->s_szc != 0);
6392 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
6393 	ASSERT(amp == NULL || IS_P2ALIGNED(anon_idx, pgcnt));
6394 	ASSERT(sameprot(seg, addr, pgcnt << PAGESHIFT));
6395 
6396 	if (VPP_PROT(svp) == prot)
6397 		return (1);
6398 	if (!((VPP_PROT(svp) ^ prot) & PROT_WRITE))
6399 		return (1);
6400 
6401 	ppa = kmem_alloc(ppasize, KM_SLEEP);
6402 	if (anon && vp != NULL) {
6403 		if (anon_get_ptr(amp->ahp, anon_idx) == NULL) {
6404 			anon = 0;
6405 			ASSERT(!anon_pages(amp->ahp, anon_idx, pgcnt));
6406 		}
6407 		ASSERT(!anon ||
6408 		    anon_pages(amp->ahp, anon_idx, pgcnt) == pgcnt);
6409 	}
6410 
6411 	for (*ppa = NULL, pg_idx = 0; svp < evp; svp++, anon_idx++) {
6412 		if (!VPP_ISPPLOCK(svp))
6413 			continue;
6414 		if (anon) {
6415 			ap = anon_get_ptr(amp->ahp, anon_idx);
6416 			if (ap == NULL) {
6417 				panic("segvn_claim_pages: no anon slot");
6418 			}
6419 			swap_xlate(ap, &vp, &aoff);
6420 			off = (u_offset_t)aoff;
6421 		}
6422 		ASSERT(vp != NULL);
6423 		if ((pp = page_lookup(vp,
6424 		    (u_offset_t)off, SE_SHARED)) == NULL) {
6425 			panic("segvn_claim_pages: no page");
6426 		}
6427 		ppa[pg_idx++] = pp;
6428 		off += PAGESIZE;
6429 	}
6430 
6431 	if (ppa[0] == NULL) {
6432 		kmem_free(ppa, ppasize);
6433 		return (1);
6434 	}
6435 
6436 	ASSERT(pg_idx <= pgcnt);
6437 	ppa[pg_idx] = NULL;
6438 
6439 	if (prot & PROT_WRITE)
6440 		err = page_addclaim_pages(ppa);
6441 	else
6442 		err = page_subclaim_pages(ppa);
6443 
6444 	for (i = 0; i < pg_idx; i++) {
6445 		ASSERT(ppa[i] != NULL);
6446 		page_unlock(ppa[i]);
6447 	}
6448 
6449 	kmem_free(ppa, ppasize);
6450 	return (err);
6451 }
6452 
6453 /*
6454  * Returns right (upper address) segment if split occurred.
6455  * If the address is equal to the beginning or end of its segment it returns
6456  * the current segment.
6457  */
6458 static struct seg *
6459 segvn_split_seg(struct seg *seg, caddr_t addr)
6460 {
6461 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6462 	struct seg *nseg;
6463 	size_t nsize;
6464 	struct segvn_data *nsvd;
6465 
6466 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6467 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6468 
6469 	ASSERT(addr >= seg->s_base);
6470 	ASSERT(addr <= seg->s_base + seg->s_size);
6471 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6472 
6473 	if (addr == seg->s_base || addr == seg->s_base + seg->s_size)
6474 		return (seg);
6475 
6476 	nsize = seg->s_base + seg->s_size - addr;
6477 	seg->s_size = addr - seg->s_base;
6478 	nseg = seg_alloc(seg->s_as, addr, nsize);
6479 	ASSERT(nseg != NULL);
6480 	nseg->s_ops = seg->s_ops;
6481 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
6482 	nseg->s_data = (void *)nsvd;
6483 	nseg->s_szc = seg->s_szc;
6484 	*nsvd = *svd;
6485 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6486 	nsvd->seg = nseg;
6487 	rw_init(&nsvd->lock, NULL, RW_DEFAULT, NULL);
6488 
6489 	if (nsvd->vp != NULL) {
6490 		VN_HOLD(nsvd->vp);
6491 		nsvd->offset = svd->offset +
6492 		    (uintptr_t)(nseg->s_base - seg->s_base);
6493 		if (nsvd->type == MAP_SHARED)
6494 			lgrp_shm_policy_init(NULL, nsvd->vp);
6495 	} else {
6496 		/*
6497 		 * The offset for an anonymous segment has no signifigance in
6498 		 * terms of an offset into a file. If we were to use the above
6499 		 * calculation instead, the structures read out of
6500 		 * /proc/<pid>/xmap would be more difficult to decipher since
6501 		 * it would be unclear whether two seemingly contiguous
6502 		 * prxmap_t structures represented different segments or a
6503 		 * single segment that had been split up into multiple prxmap_t
6504 		 * structures (e.g. if some part of the segment had not yet
6505 		 * been faulted in).
6506 		 */
6507 		nsvd->offset = 0;
6508 	}
6509 
6510 	ASSERT(svd->softlockcnt == 0);
6511 	crhold(svd->cred);
6512 
6513 	if (svd->vpage != NULL) {
6514 		size_t bytes = vpgtob(seg_pages(seg));
6515 		size_t nbytes = vpgtob(seg_pages(nseg));
6516 		struct vpage *ovpage = svd->vpage;
6517 
6518 		svd->vpage = kmem_alloc(bytes, KM_SLEEP);
6519 		bcopy(ovpage, svd->vpage, bytes);
6520 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
6521 		bcopy(ovpage + seg_pages(seg), nsvd->vpage, nbytes);
6522 		kmem_free(ovpage, bytes + nbytes);
6523 	}
6524 	if (svd->amp != NULL && svd->type == MAP_PRIVATE) {
6525 		struct anon_map *oamp = svd->amp, *namp;
6526 		struct anon_hdr *nahp;
6527 
6528 		ANON_LOCK_ENTER(&oamp->a_rwlock, RW_WRITER);
6529 		ASSERT(oamp->refcnt == 1);
6530 		nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
6531 		(void) anon_copy_ptr(oamp->ahp, svd->anon_index,
6532 		    nahp, 0, btop(seg->s_size), ANON_SLEEP);
6533 
6534 		namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
6535 		namp->a_szc = nseg->s_szc;
6536 		(void) anon_copy_ptr(oamp->ahp,
6537 		    svd->anon_index + btop(seg->s_size),
6538 		    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
6539 		anon_release(oamp->ahp, btop(oamp->size));
6540 		oamp->ahp = nahp;
6541 		oamp->size = seg->s_size;
6542 		svd->anon_index = 0;
6543 		nsvd->amp = namp;
6544 		nsvd->anon_index = 0;
6545 		ANON_LOCK_EXIT(&oamp->a_rwlock);
6546 	} else if (svd->amp != NULL) {
6547 		pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6548 		ASSERT(svd->amp == nsvd->amp);
6549 		ASSERT(seg->s_szc <= svd->amp->a_szc);
6550 		nsvd->anon_index = svd->anon_index + seg_pages(seg);
6551 		ASSERT(IS_P2ALIGNED(nsvd->anon_index, pgcnt));
6552 		ANON_LOCK_ENTER(&svd->amp->a_rwlock, RW_WRITER);
6553 		svd->amp->refcnt++;
6554 		ANON_LOCK_EXIT(&svd->amp->a_rwlock);
6555 	}
6556 
6557 	/*
6558 	 * Split amount of swap reserve
6559 	 */
6560 	if (svd->swresv) {
6561 		/*
6562 		 * For MAP_NORESERVE, only allocate swap reserve for pages
6563 		 * being used.  Other segments get enough to cover whole
6564 		 * segment.
6565 		 */
6566 		if (svd->flags & MAP_NORESERVE) {
6567 			size_t	oswresv;
6568 
6569 			ASSERT(svd->amp);
6570 			oswresv = svd->swresv;
6571 			svd->swresv = ptob(anon_pages(svd->amp->ahp,
6572 				svd->anon_index, btop(seg->s_size)));
6573 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
6574 				nsvd->anon_index, btop(nseg->s_size)));
6575 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
6576 		} else {
6577 			ASSERT(svd->swresv == seg->s_size + nseg->s_size);
6578 			svd->swresv = seg->s_size;
6579 			nsvd->swresv = nseg->s_size;
6580 		}
6581 	}
6582 
6583 	return (nseg);
6584 }
6585 
6586 /*
6587  * called on memory operations (unmap, setprot, setpagesize) for a subset
6588  * of a large page segment to either demote the memory range (SDR_RANGE)
6589  * or the ends (SDR_END) by addr/len.
6590  *
6591  * returns 0 on success. returns errno, including ENOMEM, on failure.
6592  */
6593 static int
6594 segvn_demote_range(
6595 	struct seg *seg,
6596 	caddr_t addr,
6597 	size_t len,
6598 	int flag,
6599 	uint_t szcvec)
6600 {
6601 	caddr_t eaddr = addr + len;
6602 	caddr_t lpgaddr, lpgeaddr;
6603 	struct seg *nseg;
6604 	struct seg *badseg1 = NULL;
6605 	struct seg *badseg2 = NULL;
6606 	size_t pgsz;
6607 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6608 	int err;
6609 	uint_t szc = seg->s_szc;
6610 	uint_t tszcvec;
6611 
6612 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6613 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6614 	ASSERT(szc != 0);
6615 	pgsz = page_get_pagesize(szc);
6616 	ASSERT(seg->s_base != addr || seg->s_size != len);
6617 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6618 	ASSERT(svd->softlockcnt == 0);
6619 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6620 	ASSERT(szcvec == 0 || (flag == SDR_END && svd->type == MAP_SHARED));
6621 
6622 	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
6623 	ASSERT(flag == SDR_RANGE || eaddr < lpgeaddr || addr > lpgaddr);
6624 	if (flag == SDR_RANGE) {
6625 		/* demote entire range */
6626 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6627 		(void) segvn_split_seg(nseg, lpgeaddr);
6628 		ASSERT(badseg1->s_base == lpgaddr);
6629 		ASSERT(badseg1->s_size == lpgeaddr - lpgaddr);
6630 	} else if (addr != lpgaddr) {
6631 		ASSERT(flag == SDR_END);
6632 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6633 		if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz &&
6634 		    eaddr < lpgaddr + 2 * pgsz) {
6635 			(void) segvn_split_seg(nseg, lpgeaddr);
6636 			ASSERT(badseg1->s_base == lpgaddr);
6637 			ASSERT(badseg1->s_size == 2 * pgsz);
6638 		} else {
6639 			nseg = segvn_split_seg(nseg, lpgaddr + pgsz);
6640 			ASSERT(badseg1->s_base == lpgaddr);
6641 			ASSERT(badseg1->s_size == pgsz);
6642 			if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz) {
6643 				ASSERT(lpgeaddr - lpgaddr > 2 * pgsz);
6644 				nseg = segvn_split_seg(nseg, lpgeaddr - pgsz);
6645 				badseg2 = nseg;
6646 				(void) segvn_split_seg(nseg, lpgeaddr);
6647 				ASSERT(badseg2->s_base == lpgeaddr - pgsz);
6648 				ASSERT(badseg2->s_size == pgsz);
6649 			}
6650 		}
6651 	} else {
6652 		ASSERT(flag == SDR_END);
6653 		ASSERT(eaddr < lpgeaddr);
6654 		badseg1 = nseg = segvn_split_seg(seg, lpgeaddr - pgsz);
6655 		(void) segvn_split_seg(nseg, lpgeaddr);
6656 		ASSERT(badseg1->s_base == lpgeaddr - pgsz);
6657 		ASSERT(badseg1->s_size == pgsz);
6658 	}
6659 
6660 	ASSERT(badseg1 != NULL);
6661 	ASSERT(badseg1->s_szc == szc);
6662 	ASSERT(flag == SDR_RANGE || badseg1->s_size == pgsz ||
6663 	    badseg1->s_size == 2 * pgsz);
6664 	ASSERT(sameprot(badseg1, badseg1->s_base, pgsz));
6665 	ASSERT(badseg1->s_size == pgsz ||
6666 	    sameprot(badseg1, badseg1->s_base + pgsz, pgsz));
6667 	if (err = segvn_clrszc(badseg1)) {
6668 		return (err);
6669 	}
6670 	ASSERT(badseg1->s_szc == 0);
6671 
6672 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6673 		uint_t tszc = highbit(tszcvec) - 1;
6674 		caddr_t ta = MAX(addr, badseg1->s_base);
6675 		caddr_t te;
6676 		size_t tpgsz = page_get_pagesize(tszc);
6677 
6678 		ASSERT(svd->type == MAP_SHARED);
6679 		ASSERT(flag == SDR_END);
6680 		ASSERT(tszc < szc && tszc > 0);
6681 
6682 		if (eaddr > badseg1->s_base + badseg1->s_size) {
6683 			te = badseg1->s_base + badseg1->s_size;
6684 		} else {
6685 			te = eaddr;
6686 		}
6687 
6688 		ASSERT(ta <= te);
6689 		badseg1->s_szc = tszc;
6690 		if (!IS_P2ALIGNED(ta, tpgsz) || !IS_P2ALIGNED(te, tpgsz)) {
6691 			if (badseg2 != NULL) {
6692 				err = segvn_demote_range(badseg1, ta, te - ta,
6693 				    SDR_END, tszcvec);
6694 				if (err != 0) {
6695 					return (err);
6696 				}
6697 			} else {
6698 				return (segvn_demote_range(badseg1, ta,
6699 				    te - ta, SDR_END, tszcvec));
6700 			}
6701 		}
6702 	}
6703 
6704 	if (badseg2 == NULL)
6705 		return (0);
6706 	ASSERT(badseg2->s_szc == szc);
6707 	ASSERT(badseg2->s_size == pgsz);
6708 	ASSERT(sameprot(badseg2, badseg2->s_base, badseg2->s_size));
6709 	if (err = segvn_clrszc(badseg2)) {
6710 		return (err);
6711 	}
6712 	ASSERT(badseg2->s_szc == 0);
6713 
6714 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6715 		uint_t tszc = highbit(tszcvec) - 1;
6716 		size_t tpgsz = page_get_pagesize(tszc);
6717 
6718 		ASSERT(svd->type == MAP_SHARED);
6719 		ASSERT(flag == SDR_END);
6720 		ASSERT(tszc < szc && tszc > 0);
6721 		ASSERT(badseg2->s_base > addr);
6722 		ASSERT(eaddr > badseg2->s_base);
6723 		ASSERT(eaddr < badseg2->s_base + badseg2->s_size);
6724 
6725 		badseg2->s_szc = tszc;
6726 		if (!IS_P2ALIGNED(eaddr, tpgsz)) {
6727 			return (segvn_demote_range(badseg2, badseg2->s_base,
6728 			    eaddr - badseg2->s_base, SDR_END, tszcvec));
6729 		}
6730 	}
6731 
6732 	return (0);
6733 }
6734 
6735 static int
6736 segvn_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
6737 {
6738 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6739 	struct vpage *vp, *evp;
6740 
6741 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6742 
6743 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6744 	/*
6745 	 * If segment protection can be used, simply check against them.
6746 	 */
6747 	if (svd->pageprot == 0) {
6748 		int err;
6749 
6750 		err = ((svd->prot & prot) != prot) ? EACCES : 0;
6751 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6752 		return (err);
6753 	}
6754 
6755 	/*
6756 	 * Have to check down to the vpage level.
6757 	 */
6758 	evp = &svd->vpage[seg_page(seg, addr + len)];
6759 	for (vp = &svd->vpage[seg_page(seg, addr)]; vp < evp; vp++) {
6760 		if ((VPP_PROT(vp) & prot) != prot) {
6761 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6762 			return (EACCES);
6763 		}
6764 	}
6765 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6766 	return (0);
6767 }
6768 
6769 static int
6770 segvn_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
6771 {
6772 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6773 	size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
6774 
6775 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6776 
6777 	if (pgno != 0) {
6778 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6779 		if (svd->pageprot == 0) {
6780 			do {
6781 				protv[--pgno] = svd->prot;
6782 			} while (pgno != 0);
6783 		} else {
6784 			size_t pgoff = seg_page(seg, addr);
6785 
6786 			do {
6787 				pgno--;
6788 				protv[pgno] = VPP_PROT(&svd->vpage[pgno+pgoff]);
6789 			} while (pgno != 0);
6790 		}
6791 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6792 	}
6793 	return (0);
6794 }
6795 
6796 static u_offset_t
6797 segvn_getoffset(struct seg *seg, caddr_t addr)
6798 {
6799 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6800 
6801 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6802 
6803 	return (svd->offset + (uintptr_t)(addr - seg->s_base));
6804 }
6805 
6806 /*ARGSUSED*/
6807 static int
6808 segvn_gettype(struct seg *seg, caddr_t addr)
6809 {
6810 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6811 
6812 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6813 
6814 	return (svd->type | (svd->flags & (MAP_NORESERVE | MAP_TEXT |
6815 	    MAP_INITDATA)));
6816 }
6817 
6818 /*ARGSUSED*/
6819 static int
6820 segvn_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
6821 {
6822 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6823 
6824 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6825 
6826 	*vpp = svd->vp;
6827 	return (0);
6828 }
6829 
6830 /*
6831  * Check to see if it makes sense to do kluster/read ahead to
6832  * addr + delta relative to the mapping at addr.  We assume here
6833  * that delta is a signed PAGESIZE'd multiple (which can be negative).
6834  *
6835  * For segvn, we currently "approve" of the action if we are
6836  * still in the segment and it maps from the same vp/off,
6837  * or if the advice stored in segvn_data or vpages allows it.
6838  * Currently, klustering is not allowed only if MADV_RANDOM is set.
6839  */
6840 static int
6841 segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
6842 {
6843 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6844 	struct anon *oap, *ap;
6845 	ssize_t pd;
6846 	size_t page;
6847 	struct vnode *vp1, *vp2;
6848 	u_offset_t off1, off2;
6849 	struct anon_map *amp;
6850 
6851 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6852 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6853 	    SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
6854 
6855 	if (addr + delta < seg->s_base ||
6856 	    addr + delta >= (seg->s_base + seg->s_size))
6857 		return (-1);		/* exceeded segment bounds */
6858 
6859 	pd = delta / (ssize_t)PAGESIZE;	/* divide to preserve sign bit */
6860 	page = seg_page(seg, addr);
6861 
6862 	/*
6863 	 * Check to see if either of the pages addr or addr + delta
6864 	 * have advice set that prevents klustering (if MADV_RANDOM advice
6865 	 * is set for entire segment, or MADV_SEQUENTIAL is set and delta
6866 	 * is negative).
6867 	 */
6868 	if (svd->advice == MADV_RANDOM ||
6869 	    svd->advice == MADV_SEQUENTIAL && delta < 0)
6870 		return (-1);
6871 	else if (svd->pageadvice && svd->vpage) {
6872 		struct vpage *bvpp, *evpp;
6873 
6874 		bvpp = &svd->vpage[page];
6875 		evpp = &svd->vpage[page + pd];
6876 		if (VPP_ADVICE(bvpp) == MADV_RANDOM ||
6877 		    VPP_ADVICE(evpp) == MADV_SEQUENTIAL && delta < 0)
6878 			return (-1);
6879 		if (VPP_ADVICE(bvpp) != VPP_ADVICE(evpp) &&
6880 		    VPP_ADVICE(evpp) == MADV_RANDOM)
6881 			return (-1);
6882 	}
6883 
6884 	if (svd->type == MAP_SHARED)
6885 		return (0);		/* shared mapping - all ok */
6886 
6887 	if ((amp = svd->amp) == NULL)
6888 		return (0);		/* off original vnode */
6889 
6890 	page += svd->anon_index;
6891 
6892 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
6893 
6894 	oap = anon_get_ptr(amp->ahp, page);
6895 	ap = anon_get_ptr(amp->ahp, page + pd);
6896 
6897 	ANON_LOCK_EXIT(&amp->a_rwlock);
6898 
6899 	if ((oap == NULL && ap != NULL) || (oap != NULL && ap == NULL)) {
6900 		return (-1);		/* one with and one without an anon */
6901 	}
6902 
6903 	if (oap == NULL) {		/* implies that ap == NULL */
6904 		return (0);		/* off original vnode */
6905 	}
6906 
6907 	/*
6908 	 * Now we know we have two anon pointers - check to
6909 	 * see if they happen to be properly allocated.
6910 	 */
6911 
6912 	/*
6913 	 * XXX We cheat here and don't lock the anon slots. We can't because
6914 	 * we may have been called from the anon layer which might already
6915 	 * have locked them. We are holding a refcnt on the slots so they
6916 	 * can't disappear. The worst that will happen is we'll get the wrong
6917 	 * names (vp, off) for the slots and make a poor klustering decision.
6918 	 */
6919 	swap_xlate(ap, &vp1, &off1);
6920 	swap_xlate(oap, &vp2, &off2);
6921 
6922 
6923 	if (!VOP_CMP(vp1, vp2, NULL) || off1 - off2 != delta)
6924 		return (-1);
6925 	return (0);
6926 }
6927 
6928 /*
6929  * Swap the pages of seg out to secondary storage, returning the
6930  * number of bytes of storage freed.
6931  *
6932  * The basic idea is first to unload all translations and then to call
6933  * VOP_PUTPAGE() for all newly-unmapped pages, to push them out to the
6934  * swap device.  Pages to which other segments have mappings will remain
6935  * mapped and won't be swapped.  Our caller (as_swapout) has already
6936  * performed the unloading step.
6937  *
6938  * The value returned is intended to correlate well with the process's
6939  * memory requirements.  However, there are some caveats:
6940  * 1)	When given a shared segment as argument, this routine will
6941  *	only succeed in swapping out pages for the last sharer of the
6942  *	segment.  (Previous callers will only have decremented mapping
6943  *	reference counts.)
6944  * 2)	We assume that the hat layer maintains a large enough translation
6945  *	cache to capture process reference patterns.
6946  */
6947 static size_t
6948 segvn_swapout(struct seg *seg)
6949 {
6950 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6951 	struct anon_map *amp;
6952 	pgcnt_t pgcnt = 0;
6953 	pgcnt_t npages;
6954 	pgcnt_t page;
6955 	ulong_t anon_index;
6956 
6957 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6958 
6959 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6960 	/*
6961 	 * Find pages unmapped by our caller and force them
6962 	 * out to the virtual swap device.
6963 	 */
6964 	if ((amp = svd->amp) != NULL)
6965 		anon_index = svd->anon_index;
6966 	npages = seg->s_size >> PAGESHIFT;
6967 	for (page = 0; page < npages; page++) {
6968 		page_t *pp;
6969 		struct anon *ap;
6970 		struct vnode *vp;
6971 		u_offset_t off;
6972 		anon_sync_obj_t cookie;
6973 
6974 		/*
6975 		 * Obtain <vp, off> pair for the page, then look it up.
6976 		 *
6977 		 * Note that this code is willing to consider regular
6978 		 * pages as well as anon pages.  Is this appropriate here?
6979 		 */
6980 		ap = NULL;
6981 		if (amp != NULL) {
6982 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
6983 			if (anon_array_try_enter(amp, anon_index + page,
6984 						&cookie)) {
6985 				ANON_LOCK_EXIT(&amp->a_rwlock);
6986 				continue;
6987 			}
6988 			ap = anon_get_ptr(amp->ahp, anon_index + page);
6989 			if (ap != NULL) {
6990 				swap_xlate(ap, &vp, &off);
6991 			} else {
6992 				vp = svd->vp;
6993 				off = svd->offset + ptob(page);
6994 			}
6995 			anon_array_exit(&cookie);
6996 			ANON_LOCK_EXIT(&amp->a_rwlock);
6997 		} else {
6998 			vp = svd->vp;
6999 			off = svd->offset + ptob(page);
7000 		}
7001 		if (vp == NULL) {		/* untouched zfod page */
7002 			ASSERT(ap == NULL);
7003 			continue;
7004 		}
7005 
7006 		pp = page_lookup_nowait(vp, off, SE_SHARED);
7007 		if (pp == NULL)
7008 			continue;
7009 
7010 
7011 		/*
7012 		 * Examine the page to see whether it can be tossed out,
7013 		 * keeping track of how many we've found.
7014 		 */
7015 		if (!page_tryupgrade(pp)) {
7016 			/*
7017 			 * If the page has an i/o lock and no mappings,
7018 			 * it's very likely that the page is being
7019 			 * written out as a result of klustering.
7020 			 * Assume this is so and take credit for it here.
7021 			 */
7022 			if (!page_io_trylock(pp)) {
7023 				if (!hat_page_is_mapped(pp))
7024 					pgcnt++;
7025 			} else {
7026 				page_io_unlock(pp);
7027 			}
7028 			page_unlock(pp);
7029 			continue;
7030 		}
7031 		ASSERT(!page_iolock_assert(pp));
7032 
7033 
7034 		/*
7035 		 * Skip if page is locked or has mappings.
7036 		 * We don't need the page_struct_lock to look at lckcnt
7037 		 * and cowcnt because the page is exclusive locked.
7038 		 */
7039 		if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
7040 		    hat_page_is_mapped(pp)) {
7041 			page_unlock(pp);
7042 			continue;
7043 		}
7044 
7045 		/*
7046 		 * dispose skips large pages so try to demote first.
7047 		 */
7048 		if (pp->p_szc != 0 && !page_try_demote_pages(pp)) {
7049 			page_unlock(pp);
7050 			/*
7051 			 * XXX should skip the remaining page_t's of this
7052 			 * large page.
7053 			 */
7054 			continue;
7055 		}
7056 
7057 		ASSERT(pp->p_szc == 0);
7058 
7059 		/*
7060 		 * No longer mapped -- we can toss it out.  How
7061 		 * we do so depends on whether or not it's dirty.
7062 		 */
7063 		if (hat_ismod(pp) && pp->p_vnode) {
7064 			/*
7065 			 * We must clean the page before it can be
7066 			 * freed.  Setting B_FREE will cause pvn_done
7067 			 * to free the page when the i/o completes.
7068 			 * XXX:	This also causes it to be accounted
7069 			 *	as a pageout instead of a swap: need
7070 			 *	B_SWAPOUT bit to use instead of B_FREE.
7071 			 *
7072 			 * Hold the vnode before releasing the page lock
7073 			 * to prevent it from being freed and re-used by
7074 			 * some other thread.
7075 			 */
7076 			VN_HOLD(vp);
7077 			page_unlock(pp);
7078 
7079 			/*
7080 			 * Queue all i/o requests for the pageout thread
7081 			 * to avoid saturating the pageout devices.
7082 			 */
7083 			if (!queue_io_request(vp, off))
7084 				VN_RELE(vp);
7085 		} else {
7086 			/*
7087 			 * The page was clean, free it.
7088 			 *
7089 			 * XXX:	Can we ever encounter modified pages
7090 			 *	with no associated vnode here?
7091 			 */
7092 			ASSERT(pp->p_vnode != NULL);
7093 			/*LINTED: constant in conditional context*/
7094 			VN_DISPOSE(pp, B_FREE, 0, kcred);
7095 		}
7096 
7097 		/*
7098 		 * Credit now even if i/o is in progress.
7099 		 */
7100 		pgcnt++;
7101 	}
7102 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7103 
7104 	/*
7105 	 * Wakeup pageout to initiate i/o on all queued requests.
7106 	 */
7107 	cv_signal_pageout();
7108 	return (ptob(pgcnt));
7109 }
7110 
7111 /*
7112  * Synchronize primary storage cache with real object in virtual memory.
7113  *
7114  * XXX - Anonymous pages should not be sync'ed out at all.
7115  */
7116 static int
7117 segvn_sync(struct seg *seg, caddr_t addr, size_t len, int attr, uint_t flags)
7118 {
7119 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7120 	struct vpage *vpp;
7121 	page_t *pp;
7122 	u_offset_t offset;
7123 	struct vnode *vp;
7124 	u_offset_t off;
7125 	caddr_t eaddr;
7126 	int bflags;
7127 	int err = 0;
7128 	int segtype;
7129 	int pageprot;
7130 	int prot;
7131 	ulong_t anon_index;
7132 	struct anon_map *amp;
7133 	struct anon *ap;
7134 	anon_sync_obj_t cookie;
7135 
7136 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7137 
7138 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7139 
7140 	if (svd->softlockcnt > 0) {
7141 		/*
7142 		 * flush all pages from seg cache
7143 		 * otherwise we may deadlock in swap_putpage
7144 		 * for B_INVAL page (4175402).
7145 		 *
7146 		 * Even if we grab segvn WRITER's lock or segp_slock
7147 		 * here, there might be another thread which could've
7148 		 * successfully performed lookup/insert just before
7149 		 * we acquired the lock here.  So, grabbing either
7150 		 * lock here is of not much use.  Until we devise
7151 		 * a strategy at upper layers to solve the
7152 		 * synchronization issues completely, we expect
7153 		 * applications to handle this appropriately.
7154 		 */
7155 		segvn_purge(seg);
7156 		if (svd->softlockcnt > 0) {
7157 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7158 			return (EAGAIN);
7159 		}
7160 	}
7161 
7162 	vpp = svd->vpage;
7163 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7164 	bflags = ((flags & MS_ASYNC) ? B_ASYNC : 0) |
7165 	    ((flags & MS_INVALIDATE) ? B_INVAL : 0);
7166 
7167 	if (attr) {
7168 		pageprot = attr & ~(SHARED|PRIVATE);
7169 		segtype = (attr & SHARED) ? MAP_SHARED : MAP_PRIVATE;
7170 
7171 		/*
7172 		 * We are done if the segment types don't match
7173 		 * or if we have segment level protections and
7174 		 * they don't match.
7175 		 */
7176 		if (svd->type != segtype) {
7177 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7178 			return (0);
7179 		}
7180 		if (vpp == NULL) {
7181 			if (svd->prot != pageprot) {
7182 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7183 				return (0);
7184 			}
7185 			prot = svd->prot;
7186 		} else
7187 			vpp = &svd->vpage[seg_page(seg, addr)];
7188 
7189 	} else if (svd->vp && svd->amp == NULL &&
7190 	    (flags & MS_INVALIDATE) == 0) {
7191 
7192 		/*
7193 		 * No attributes, no anonymous pages and MS_INVALIDATE flag
7194 		 * is not on, just use one big request.
7195 		 */
7196 		err = VOP_PUTPAGE(svd->vp, (offset_t)offset, len,
7197 		    bflags, svd->cred, NULL);
7198 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7199 		return (err);
7200 	}
7201 
7202 	if ((amp = svd->amp) != NULL)
7203 		anon_index = svd->anon_index + seg_page(seg, addr);
7204 
7205 	for (eaddr = addr + len; addr < eaddr; addr += PAGESIZE) {
7206 		ap = NULL;
7207 		if (amp != NULL) {
7208 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7209 			anon_array_enter(amp, anon_index, &cookie);
7210 			ap = anon_get_ptr(amp->ahp, anon_index++);
7211 			if (ap != NULL) {
7212 				swap_xlate(ap, &vp, &off);
7213 			} else {
7214 				vp = svd->vp;
7215 				off = offset;
7216 			}
7217 			anon_array_exit(&cookie);
7218 			ANON_LOCK_EXIT(&amp->a_rwlock);
7219 		} else {
7220 			vp = svd->vp;
7221 			off = offset;
7222 		}
7223 		offset += PAGESIZE;
7224 
7225 		if (vp == NULL)		/* untouched zfod page */
7226 			continue;
7227 
7228 		if (attr) {
7229 			if (vpp) {
7230 				prot = VPP_PROT(vpp);
7231 				vpp++;
7232 			}
7233 			if (prot != pageprot) {
7234 				continue;
7235 			}
7236 		}
7237 
7238 		/*
7239 		 * See if any of these pages are locked --  if so, then we
7240 		 * will have to truncate an invalidate request at the first
7241 		 * locked one. We don't need the page_struct_lock to test
7242 		 * as this is only advisory; even if we acquire it someone
7243 		 * might race in and lock the page after we unlock and before
7244 		 * we do the PUTPAGE, then PUTPAGE simply does nothing.
7245 		 */
7246 		if (flags & MS_INVALIDATE) {
7247 			if ((pp = page_lookup(vp, off, SE_SHARED)) != NULL) {
7248 				if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
7249 					page_unlock(pp);
7250 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7251 					return (EBUSY);
7252 				}
7253 				if (ap != NULL && pp->p_szc != 0 &&
7254 				    page_tryupgrade(pp)) {
7255 					if (pp->p_lckcnt == 0 &&
7256 					    pp->p_cowcnt == 0) {
7257 						/*
7258 						 * swapfs VN_DISPOSE() won't
7259 						 * invalidate large pages.
7260 						 * Attempt to demote.
7261 						 * XXX can't help it if it
7262 						 * fails. But for swapfs
7263 						 * pages it is no big deal.
7264 						 */
7265 						(void) page_try_demote_pages(
7266 						    pp);
7267 				    }
7268 				}
7269 				page_unlock(pp);
7270 			}
7271 		} else if (svd->type == MAP_SHARED && amp != NULL) {
7272 			/*
7273 			 * Avoid writing out to disk ISM's large pages
7274 			 * because segspt_free_pages() relies on NULL an_pvp
7275 			 * of anon slots of such pages.
7276 			 */
7277 
7278 			ASSERT(svd->vp == NULL);
7279 			/*
7280 			 * swapfs uses page_lookup_nowait if not freeing or
7281 			 * invalidating and skips a page if
7282 			 * page_lookup_nowait returns NULL.
7283 			 */
7284 			pp = page_lookup_nowait(vp, off, SE_SHARED);
7285 			if (pp == NULL) {
7286 				continue;
7287 			}
7288 			if (pp->p_szc != 0) {
7289 				page_unlock(pp);
7290 				continue;
7291 			}
7292 
7293 			/*
7294 			 * Note ISM pages are created large so (vp, off)'s
7295 			 * page cannot suddenly become large after we unlock
7296 			 * pp.
7297 			 */
7298 			page_unlock(pp);
7299 		}
7300 		/*
7301 		 * XXX - Should ultimately try to kluster
7302 		 * calls to VOP_PUTPAGE() for performance.
7303 		 */
7304 		VN_HOLD(vp);
7305 		err = VOP_PUTPAGE(vp, (offset_t)off, PAGESIZE,
7306 		    bflags, svd->cred, NULL);
7307 		VN_RELE(vp);
7308 		if (err)
7309 			break;
7310 	}
7311 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7312 	return (err);
7313 }
7314 
7315 /*
7316  * Determine if we have data corresponding to pages in the
7317  * primary storage virtual memory cache (i.e., "in core").
7318  */
7319 static size_t
7320 segvn_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
7321 {
7322 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7323 	struct vnode *vp, *avp;
7324 	u_offset_t offset, aoffset;
7325 	size_t p, ep;
7326 	int ret;
7327 	struct vpage *vpp;
7328 	page_t *pp;
7329 	uint_t start;
7330 	struct anon_map *amp;		/* XXX - for locknest */
7331 	struct anon *ap;
7332 	uint_t attr;
7333 	anon_sync_obj_t cookie;
7334 
7335 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7336 
7337 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7338 	if (svd->amp == NULL && svd->vp == NULL) {
7339 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7340 		bzero(vec, btopr(len));
7341 		return (len);	/* no anonymous pages created yet */
7342 	}
7343 
7344 	p = seg_page(seg, addr);
7345 	ep = seg_page(seg, addr + len);
7346 	start = svd->vp ? SEG_PAGE_VNODEBACKED : 0;
7347 
7348 	amp = svd->amp;
7349 	for (; p < ep; p++, addr += PAGESIZE) {
7350 		vpp = (svd->vpage) ? &svd->vpage[p]: NULL;
7351 		ret = start;
7352 		ap = NULL;
7353 		avp = NULL;
7354 		/* Grab the vnode/offset for the anon slot */
7355 		if (amp != NULL) {
7356 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7357 			anon_array_enter(amp, svd->anon_index + p, &cookie);
7358 			ap = anon_get_ptr(amp->ahp, svd->anon_index + p);
7359 			if (ap != NULL) {
7360 				swap_xlate(ap, &avp, &aoffset);
7361 			}
7362 			anon_array_exit(&cookie);
7363 			ANON_LOCK_EXIT(&amp->a_rwlock);
7364 		}
7365 		if ((avp != NULL) && page_exists(avp, aoffset)) {
7366 			/* A page exists for the anon slot */
7367 			ret |= SEG_PAGE_INCORE;
7368 
7369 			/*
7370 			 * If page is mapped and writable
7371 			 */
7372 			attr = (uint_t)0;
7373 			if ((hat_getattr(seg->s_as->a_hat, addr,
7374 			    &attr) != -1) && (attr & PROT_WRITE)) {
7375 				ret |= SEG_PAGE_ANON;
7376 			}
7377 			/*
7378 			 * Don't get page_struct lock for lckcnt and cowcnt,
7379 			 * since this is purely advisory.
7380 			 */
7381 			if ((pp = page_lookup_nowait(avp, aoffset,
7382 			    SE_SHARED)) != NULL) {
7383 				if (pp->p_lckcnt)
7384 					ret |= SEG_PAGE_SOFTLOCK;
7385 				if (pp->p_cowcnt)
7386 					ret |= SEG_PAGE_HASCOW;
7387 				page_unlock(pp);
7388 			}
7389 		}
7390 
7391 		/* Gather vnode statistics */
7392 		vp = svd->vp;
7393 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7394 
7395 		if (vp != NULL) {
7396 			/*
7397 			 * Try to obtain a "shared" lock on the page
7398 			 * without blocking.  If this fails, determine
7399 			 * if the page is in memory.
7400 			 */
7401 			pp = page_lookup_nowait(vp, offset, SE_SHARED);
7402 			if ((pp == NULL) && (page_exists(vp, offset))) {
7403 				/* Page is incore, and is named */
7404 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7405 			}
7406 			/*
7407 			 * Don't get page_struct lock for lckcnt and cowcnt,
7408 			 * since this is purely advisory.
7409 			 */
7410 			if (pp != NULL) {
7411 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7412 				if (pp->p_lckcnt)
7413 					ret |= SEG_PAGE_SOFTLOCK;
7414 				if (pp->p_cowcnt)
7415 					ret |= SEG_PAGE_HASCOW;
7416 				page_unlock(pp);
7417 			}
7418 		}
7419 
7420 		/* Gather virtual page information */
7421 		if (vpp) {
7422 			if (VPP_ISPPLOCK(vpp))
7423 				ret |= SEG_PAGE_LOCKED;
7424 			vpp++;
7425 		}
7426 
7427 		*vec++ = (char)ret;
7428 	}
7429 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7430 	return (len);
7431 }
7432 
7433 /*
7434  * Statement for p_cowcnts/p_lckcnts.
7435  *
7436  * p_cowcnt is updated while mlock/munlocking MAP_PRIVATE and PROT_WRITE region
7437  * irrespective of the following factors or anything else:
7438  *
7439  *	(1) anon slots are populated or not
7440  *	(2) cow is broken or not
7441  *	(3) refcnt on ap is 1 or greater than 1
7442  *
7443  * If it's not MAP_PRIVATE and PROT_WRITE, p_lckcnt is updated during mlock
7444  * and munlock.
7445  *
7446  *
7447  * Handling p_cowcnts/p_lckcnts during copy-on-write fault:
7448  *
7449  *	if vpage has PROT_WRITE
7450  *		transfer cowcnt on the oldpage -> cowcnt on the newpage
7451  *	else
7452  *		transfer lckcnt on the oldpage -> lckcnt on the newpage
7453  *
7454  *	During copy-on-write, decrement p_cowcnt on the oldpage and increment
7455  *	p_cowcnt on the newpage *if* the corresponding vpage has PROT_WRITE.
7456  *
7457  *	We may also break COW if softlocking on read access in the physio case.
7458  *	In this case, vpage may not have PROT_WRITE. So, we need to decrement
7459  *	p_lckcnt on the oldpage and increment p_lckcnt on the newpage *if* the
7460  *	vpage doesn't have PROT_WRITE.
7461  *
7462  *
7463  * Handling p_cowcnts/p_lckcnts during mprotect on mlocked region:
7464  *
7465  * 	If a MAP_PRIVATE region loses PROT_WRITE, we decrement p_cowcnt and
7466  *	increment p_lckcnt by calling page_subclaim() which takes care of
7467  * 	availrmem accounting and p_lckcnt overflow.
7468  *
7469  *	If a MAP_PRIVATE region gains PROT_WRITE, we decrement p_lckcnt and
7470  *	increment p_cowcnt by calling page_addclaim() which takes care of
7471  *	availrmem availability and p_cowcnt overflow.
7472  */
7473 
7474 /*
7475  * Lock down (or unlock) pages mapped by this segment.
7476  *
7477  * XXX only creates PAGESIZE pages if anon slots are not initialized.
7478  * At fault time they will be relocated into larger pages.
7479  */
7480 static int
7481 segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
7482     int attr, int op, ulong_t *lockmap, size_t pos)
7483 {
7484 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7485 	struct vpage *vpp;
7486 	struct vpage *evp;
7487 	page_t *pp;
7488 	u_offset_t offset;
7489 	u_offset_t off;
7490 	int segtype;
7491 	int pageprot;
7492 	int claim;
7493 	struct vnode *vp;
7494 	ulong_t anon_index;
7495 	struct anon_map *amp;
7496 	struct anon *ap;
7497 	struct vattr va;
7498 	anon_sync_obj_t cookie;
7499 	struct kshmid *sp = NULL;
7500 	struct proc	*p = curproc;
7501 	kproject_t	*proj = NULL;
7502 	int chargeproc = 1;
7503 	size_t locked_bytes = 0;
7504 	size_t unlocked_bytes = 0;
7505 	int err = 0;
7506 
7507 	/*
7508 	 * Hold write lock on address space because may split or concatenate
7509 	 * segments
7510 	 */
7511 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7512 
7513 	/*
7514 	 * If this is a shm, use shm's project and zone, else use
7515 	 * project and zone of calling process
7516 	 */
7517 
7518 	/* Determine if this segment backs a sysV shm */
7519 	if (svd->amp != NULL && svd->amp->a_sp != NULL) {
7520 		ASSERT(svd->type == MAP_SHARED);
7521 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
7522 		sp = svd->amp->a_sp;
7523 		proj = sp->shm_perm.ipc_proj;
7524 		chargeproc = 0;
7525 	}
7526 
7527 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7528 	if (attr) {
7529 		pageprot = attr & ~(SHARED|PRIVATE);
7530 		segtype = attr & SHARED ? MAP_SHARED : MAP_PRIVATE;
7531 
7532 		/*
7533 		 * We are done if the segment types don't match
7534 		 * or if we have segment level protections and
7535 		 * they don't match.
7536 		 */
7537 		if (svd->type != segtype) {
7538 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7539 			return (0);
7540 		}
7541 		if (svd->pageprot == 0 && svd->prot != pageprot) {
7542 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7543 			return (0);
7544 		}
7545 	}
7546 
7547 	if (op == MC_LOCK) {
7548 		if (svd->tr_state == SEGVN_TR_INIT) {
7549 			svd->tr_state = SEGVN_TR_OFF;
7550 		} else if (svd->tr_state == SEGVN_TR_ON) {
7551 			ASSERT(svd->amp != NULL);
7552 			segvn_textunrepl(seg, 0);
7553 			ASSERT(svd->amp == NULL &&
7554 			    svd->tr_state == SEGVN_TR_OFF);
7555 		}
7556 	}
7557 
7558 	/*
7559 	 * If we're locking, then we must create a vpage structure if
7560 	 * none exists.  If we're unlocking, then check to see if there
7561 	 * is a vpage --  if not, then we could not have locked anything.
7562 	 */
7563 
7564 	if ((vpp = svd->vpage) == NULL) {
7565 		if (op == MC_LOCK)
7566 			segvn_vpage(seg);
7567 		else {
7568 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7569 			return (0);
7570 		}
7571 	}
7572 
7573 	/*
7574 	 * The anonymous data vector (i.e., previously
7575 	 * unreferenced mapping to swap space) can be allocated
7576 	 * by lazily testing for its existence.
7577 	 */
7578 	if (op == MC_LOCK && svd->amp == NULL && svd->vp == NULL) {
7579 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
7580 		svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
7581 		svd->amp->a_szc = seg->s_szc;
7582 	}
7583 
7584 	if ((amp = svd->amp) != NULL) {
7585 		anon_index = svd->anon_index + seg_page(seg, addr);
7586 	}
7587 
7588 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7589 	evp = &svd->vpage[seg_page(seg, addr + len)];
7590 
7591 	if (sp != NULL)
7592 		mutex_enter(&sp->shm_mlock);
7593 
7594 	/* determine number of unlocked bytes in range for lock operation */
7595 	if (op == MC_LOCK) {
7596 
7597 		if (sp == NULL) {
7598 			for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7599 			    vpp++) {
7600 				if (!VPP_ISPPLOCK(vpp))
7601 					unlocked_bytes += PAGESIZE;
7602 			}
7603 		} else {
7604 			ulong_t		i_idx, i_edx;
7605 			anon_sync_obj_t	i_cookie;
7606 			struct anon	*i_ap;
7607 			struct vnode	*i_vp;
7608 			u_offset_t	i_off;
7609 
7610 			/* Only count sysV pages once for locked memory */
7611 			i_edx = svd->anon_index + seg_page(seg, addr + len);
7612 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7613 			for (i_idx = anon_index; i_idx < i_edx; i_idx++) {
7614 				anon_array_enter(amp, i_idx, &i_cookie);
7615 				i_ap = anon_get_ptr(amp->ahp, i_idx);
7616 				if (i_ap == NULL) {
7617 					unlocked_bytes += PAGESIZE;
7618 					anon_array_exit(&i_cookie);
7619 					continue;
7620 				}
7621 				swap_xlate(i_ap, &i_vp, &i_off);
7622 				anon_array_exit(&i_cookie);
7623 				pp = page_lookup(i_vp, i_off, SE_SHARED);
7624 				if (pp == NULL) {
7625 					unlocked_bytes += PAGESIZE;
7626 					continue;
7627 				} else if (pp->p_lckcnt == 0)
7628 					unlocked_bytes += PAGESIZE;
7629 				page_unlock(pp);
7630 			}
7631 			ANON_LOCK_EXIT(&amp->a_rwlock);
7632 		}
7633 
7634 		mutex_enter(&p->p_lock);
7635 		err = rctl_incr_locked_mem(p, proj, unlocked_bytes,
7636 		    chargeproc);
7637 		mutex_exit(&p->p_lock);
7638 
7639 		if (err) {
7640 			if (sp != NULL)
7641 				mutex_exit(&sp->shm_mlock);
7642 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7643 			return (err);
7644 		}
7645 	}
7646 	/*
7647 	 * Loop over all pages in the range.  Process if we're locking and
7648 	 * page has not already been locked in this mapping; or if we're
7649 	 * unlocking and the page has been locked.
7650 	 */
7651 	for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7652 	    vpp++, pos++, addr += PAGESIZE, offset += PAGESIZE, anon_index++) {
7653 		if ((attr == 0 || VPP_PROT(vpp) == pageprot) &&
7654 		    ((op == MC_LOCK && !VPP_ISPPLOCK(vpp)) ||
7655 		    (op == MC_UNLOCK && VPP_ISPPLOCK(vpp)))) {
7656 
7657 			if (amp != NULL)
7658 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7659 			/*
7660 			 * If this isn't a MAP_NORESERVE segment and
7661 			 * we're locking, allocate anon slots if they
7662 			 * don't exist.  The page is brought in later on.
7663 			 */
7664 			if (op == MC_LOCK && svd->vp == NULL &&
7665 			    ((svd->flags & MAP_NORESERVE) == 0) &&
7666 			    amp != NULL &&
7667 			    ((ap = anon_get_ptr(amp->ahp, anon_index))
7668 								== NULL)) {
7669 				anon_array_enter(amp, anon_index, &cookie);
7670 
7671 				if ((ap = anon_get_ptr(amp->ahp,
7672 						anon_index)) == NULL) {
7673 					pp = anon_zero(seg, addr, &ap,
7674 					    svd->cred);
7675 					if (pp == NULL) {
7676 						anon_array_exit(&cookie);
7677 						ANON_LOCK_EXIT(&amp->a_rwlock);
7678 						err = ENOMEM;
7679 						goto out;
7680 					}
7681 					ASSERT(anon_get_ptr(amp->ahp,
7682 						anon_index) == NULL);
7683 					(void) anon_set_ptr(amp->ahp,
7684 						anon_index, ap, ANON_SLEEP);
7685 					page_unlock(pp);
7686 				}
7687 				anon_array_exit(&cookie);
7688 			}
7689 
7690 			/*
7691 			 * Get name for page, accounting for
7692 			 * existence of private copy.
7693 			 */
7694 			ap = NULL;
7695 			if (amp != NULL) {
7696 				anon_array_enter(amp, anon_index, &cookie);
7697 				ap = anon_get_ptr(amp->ahp, anon_index);
7698 				if (ap != NULL) {
7699 					swap_xlate(ap, &vp, &off);
7700 				} else {
7701 					if (svd->vp == NULL &&
7702 					    (svd->flags & MAP_NORESERVE)) {
7703 						anon_array_exit(&cookie);
7704 						ANON_LOCK_EXIT(&amp->a_rwlock);
7705 						continue;
7706 					}
7707 					vp = svd->vp;
7708 					off = offset;
7709 				}
7710 				anon_array_exit(&cookie);
7711 				ANON_LOCK_EXIT(&amp->a_rwlock);
7712 			} else {
7713 				vp = svd->vp;
7714 				off = offset;
7715 			}
7716 
7717 			/*
7718 			 * Get page frame.  It's ok if the page is
7719 			 * not available when we're unlocking, as this
7720 			 * may simply mean that a page we locked got
7721 			 * truncated out of existence after we locked it.
7722 			 *
7723 			 * Invoke VOP_GETPAGE() to obtain the page struct
7724 			 * since we may need to read it from disk if its
7725 			 * been paged out.
7726 			 */
7727 			if (op != MC_LOCK)
7728 				pp = page_lookup(vp, off, SE_SHARED);
7729 			else {
7730 				page_t *pl[1 + 1];
7731 				int error;
7732 
7733 				ASSERT(vp != NULL);
7734 
7735 				error = VOP_GETPAGE(vp, (offset_t)off, PAGESIZE,
7736 				    (uint_t *)NULL, pl, PAGESIZE, seg, addr,
7737 				    S_OTHER, svd->cred, NULL);
7738 
7739 				/*
7740 				 * If the error is EDEADLK then we must bounce
7741 				 * up and drop all vm subsystem locks and then
7742 				 * retry the operation later
7743 				 * This behavior is a temporary measure because
7744 				 * ufs/sds logging is badly designed and will
7745 				 * deadlock if we don't allow this bounce to
7746 				 * happen.  The real solution is to re-design
7747 				 * the logging code to work properly.  See bug
7748 				 * 4125102 for details of the problem.
7749 				 */
7750 				if (error == EDEADLK) {
7751 					err = error;
7752 					goto out;
7753 				}
7754 				/*
7755 				 * Quit if we fail to fault in the page.  Treat
7756 				 * the failure as an error, unless the addr
7757 				 * is mapped beyond the end of a file.
7758 				 */
7759 				if (error && svd->vp) {
7760 					va.va_mask = AT_SIZE;
7761 					if (VOP_GETATTR(svd->vp, &va, 0,
7762 					    svd->cred, NULL) != 0) {
7763 						err = EIO;
7764 						goto out;
7765 					}
7766 					if (btopr(va.va_size) >=
7767 					    btopr(off + 1)) {
7768 						err = EIO;
7769 						goto out;
7770 					}
7771 					goto out;
7772 
7773 				} else if (error) {
7774 					err = EIO;
7775 					goto out;
7776 				}
7777 				pp = pl[0];
7778 				ASSERT(pp != NULL);
7779 			}
7780 
7781 			/*
7782 			 * See Statement at the beginning of this routine.
7783 			 *
7784 			 * claim is always set if MAP_PRIVATE and PROT_WRITE
7785 			 * irrespective of following factors:
7786 			 *
7787 			 * (1) anon slots are populated or not
7788 			 * (2) cow is broken or not
7789 			 * (3) refcnt on ap is 1 or greater than 1
7790 			 *
7791 			 * See 4140683 for details
7792 			 */
7793 			claim = ((VPP_PROT(vpp) & PROT_WRITE) &&
7794 				(svd->type == MAP_PRIVATE));
7795 
7796 			/*
7797 			 * Perform page-level operation appropriate to
7798 			 * operation.  If locking, undo the SOFTLOCK
7799 			 * performed to bring the page into memory
7800 			 * after setting the lock.  If unlocking,
7801 			 * and no page was found, account for the claim
7802 			 * separately.
7803 			 */
7804 			if (op == MC_LOCK) {
7805 				int ret = 1;	/* Assume success */
7806 
7807 				ASSERT(!VPP_ISPPLOCK(vpp));
7808 
7809 				ret = page_pp_lock(pp, claim, 0);
7810 				if (ret == 0) {
7811 					/* locking page failed */
7812 					page_unlock(pp);
7813 					err = EAGAIN;
7814 					goto out;
7815 				}
7816 				VPP_SETPPLOCK(vpp);
7817 				if (sp != NULL) {
7818 					if (pp->p_lckcnt == 1)
7819 						locked_bytes += PAGESIZE;
7820 				} else
7821 					locked_bytes += PAGESIZE;
7822 
7823 				if (lockmap != (ulong_t *)NULL)
7824 					BT_SET(lockmap, pos);
7825 
7826 				page_unlock(pp);
7827 			} else {
7828 				ASSERT(VPP_ISPPLOCK(vpp));
7829 				if (pp != NULL) {
7830 					/* sysV pages should be locked */
7831 					ASSERT(sp == NULL || pp->p_lckcnt > 0);
7832 					page_pp_unlock(pp, claim, 0);
7833 					if (sp != NULL) {
7834 						if (pp->p_lckcnt == 0)
7835 							unlocked_bytes
7836 							    += PAGESIZE;
7837 					} else
7838 						unlocked_bytes += PAGESIZE;
7839 					page_unlock(pp);
7840 				} else {
7841 					ASSERT(sp == NULL);
7842 					unlocked_bytes += PAGESIZE;
7843 				}
7844 				VPP_CLRPPLOCK(vpp);
7845 			}
7846 		}
7847 	}
7848 out:
7849 	if (op == MC_LOCK) {
7850 		/* Credit back bytes that did not get locked */
7851 		if ((unlocked_bytes - locked_bytes) > 0) {
7852 			if (proj == NULL)
7853 				mutex_enter(&p->p_lock);
7854 			rctl_decr_locked_mem(p, proj,
7855 			    (unlocked_bytes - locked_bytes), chargeproc);
7856 			if (proj == NULL)
7857 				mutex_exit(&p->p_lock);
7858 		}
7859 
7860 	} else {
7861 		/* Account bytes that were unlocked */
7862 		if (unlocked_bytes > 0) {
7863 			if (proj == NULL)
7864 				mutex_enter(&p->p_lock);
7865 			rctl_decr_locked_mem(p, proj, unlocked_bytes,
7866 			    chargeproc);
7867 			if (proj == NULL)
7868 				mutex_exit(&p->p_lock);
7869 		}
7870 	}
7871 	if (sp != NULL)
7872 		mutex_exit(&sp->shm_mlock);
7873 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7874 
7875 	return (err);
7876 }
7877 
7878 /*
7879  * Set advice from user for specified pages
7880  * There are 5 types of advice:
7881  *	MADV_NORMAL	- Normal (default) behavior (whatever that is)
7882  *	MADV_RANDOM	- Random page references
7883  *				do not allow readahead or 'klustering'
7884  *	MADV_SEQUENTIAL	- Sequential page references
7885  *				Pages previous to the one currently being
7886  *				accessed (determined by fault) are 'not needed'
7887  *				and are freed immediately
7888  *	MADV_WILLNEED	- Pages are likely to be used (fault ahead in mctl)
7889  *	MADV_DONTNEED	- Pages are not needed (synced out in mctl)
7890  *	MADV_FREE	- Contents can be discarded
7891  *	MADV_ACCESS_DEFAULT- Default access
7892  *	MADV_ACCESS_LWP	- Next LWP will access heavily
7893  *	MADV_ACCESS_MANY- Many LWPs or processes will access heavily
7894  */
7895 static int
7896 segvn_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
7897 {
7898 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7899 	size_t page;
7900 	int err = 0;
7901 	int already_set;
7902 	struct anon_map *amp;
7903 	ulong_t anon_index;
7904 	struct seg *next;
7905 	lgrp_mem_policy_t policy;
7906 	struct seg *prev;
7907 	struct vnode *vp;
7908 
7909 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7910 
7911 	/*
7912 	 * In case of MADV_FREE, we won't be modifying any segment private
7913 	 * data structures; so, we only need to grab READER's lock
7914 	 */
7915 	if (behav != MADV_FREE) {
7916 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7917 		if (svd->tr_state != SEGVN_TR_OFF) {
7918 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7919 			return (0);
7920 		}
7921 	} else {
7922 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7923 	}
7924 
7925 	/*
7926 	 * Large pages are assumed to be only turned on when accesses to the
7927 	 * segment's address range have spatial and temporal locality. That
7928 	 * justifies ignoring MADV_SEQUENTIAL for large page segments.
7929 	 * Also, ignore advice affecting lgroup memory allocation
7930 	 * if don't need to do lgroup optimizations on this system
7931 	 */
7932 
7933 	if ((behav == MADV_SEQUENTIAL &&
7934 	    (seg->s_szc != 0 || HAT_IS_REGION_COOKIE_VALID(svd->rcookie))) ||
7935 	    (!lgrp_optimizations() && (behav == MADV_ACCESS_DEFAULT ||
7936 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY))) {
7937 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7938 		return (0);
7939 	}
7940 
7941 	if (behav == MADV_SEQUENTIAL || behav == MADV_ACCESS_DEFAULT ||
7942 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY) {
7943 		/*
7944 		 * Since we are going to unload hat mappings
7945 		 * we first have to flush the cache. Otherwise
7946 		 * this might lead to system panic if another
7947 		 * thread is doing physio on the range whose
7948 		 * mappings are unloaded by madvise(3C).
7949 		 */
7950 		if (svd->softlockcnt > 0) {
7951 			/*
7952 			 * Since we do have the segvn writers lock
7953 			 * nobody can fill the cache with entries
7954 			 * belonging to this seg during the purge.
7955 			 * The flush either succeeds or we still
7956 			 * have pending I/Os. In the later case,
7957 			 * madvise(3C) fails.
7958 			 */
7959 			segvn_purge(seg);
7960 			if (svd->softlockcnt > 0) {
7961 				/*
7962 				 * Since madvise(3C) is advisory and
7963 				 * it's not part of UNIX98, madvise(3C)
7964 				 * failure here doesn't cause any hardship.
7965 				 * Note that we don't block in "as" layer.
7966 				 */
7967 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7968 				return (EAGAIN);
7969 			}
7970 		}
7971 	}
7972 
7973 	amp = svd->amp;
7974 	vp = svd->vp;
7975 	if (behav == MADV_FREE) {
7976 		/*
7977 		 * MADV_FREE is not supported for segments with
7978 		 * underlying object; if anonmap is NULL, anon slots
7979 		 * are not yet populated and there is nothing for
7980 		 * us to do. As MADV_FREE is advisory, we don't
7981 		 * return error in either case.
7982 		 */
7983 		if (vp != NULL || amp == NULL) {
7984 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7985 			return (0);
7986 		}
7987 
7988 		page = seg_page(seg, addr);
7989 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7990 		anon_disclaim(amp, svd->anon_index + page, len);
7991 		ANON_LOCK_EXIT(&amp->a_rwlock);
7992 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7993 		return (0);
7994 	}
7995 
7996 	/*
7997 	 * If advice is to be applied to entire segment,
7998 	 * use advice field in seg_data structure
7999 	 * otherwise use appropriate vpage entry.
8000 	 */
8001 	if ((addr == seg->s_base) && (len == seg->s_size)) {
8002 		switch (behav) {
8003 		case MADV_ACCESS_LWP:
8004 		case MADV_ACCESS_MANY:
8005 		case MADV_ACCESS_DEFAULT:
8006 			/*
8007 			 * Set memory allocation policy for this segment
8008 			 */
8009 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8010 			if (svd->type == MAP_SHARED)
8011 				already_set = lgrp_shm_policy_set(policy, amp,
8012 				    svd->anon_index, vp, svd->offset, len);
8013 			else {
8014 				/*
8015 				 * For private memory, need writers lock on
8016 				 * address space because the segment may be
8017 				 * split or concatenated when changing policy
8018 				 */
8019 				if (AS_READ_HELD(seg->s_as,
8020 				    &seg->s_as->a_lock)) {
8021 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8022 					return (IE_RETRY);
8023 				}
8024 
8025 				already_set = lgrp_privm_policy_set(policy,
8026 				    &svd->policy_info, len);
8027 			}
8028 
8029 			/*
8030 			 * If policy set already and it shouldn't be reapplied,
8031 			 * don't do anything.
8032 			 */
8033 			if (already_set &&
8034 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8035 				break;
8036 
8037 			/*
8038 			 * Mark any existing pages in given range for
8039 			 * migration
8040 			 */
8041 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8042 			    vp, svd->offset, 1);
8043 
8044 			/*
8045 			 * If same policy set already or this is a shared
8046 			 * memory segment, don't need to try to concatenate
8047 			 * segment with adjacent ones.
8048 			 */
8049 			if (already_set || svd->type == MAP_SHARED)
8050 				break;
8051 
8052 			/*
8053 			 * Try to concatenate this segment with previous
8054 			 * one and next one, since we changed policy for
8055 			 * this one and it may be compatible with adjacent
8056 			 * ones now.
8057 			 */
8058 			prev = AS_SEGPREV(seg->s_as, seg);
8059 			next = AS_SEGNEXT(seg->s_as, seg);
8060 
8061 			if (next && next->s_ops == &segvn_ops &&
8062 			    addr + len == next->s_base)
8063 				(void) segvn_concat(seg, next, 1);
8064 
8065 			if (prev && prev->s_ops == &segvn_ops &&
8066 			    addr == prev->s_base + prev->s_size) {
8067 				/*
8068 				 * Drop lock for private data of current
8069 				 * segment before concatenating (deleting) it
8070 				 * and return IE_REATTACH to tell as_ctl() that
8071 				 * current segment has changed
8072 				 */
8073 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8074 				if (!segvn_concat(prev, seg, 1))
8075 					err = IE_REATTACH;
8076 
8077 				return (err);
8078 			}
8079 			break;
8080 
8081 		case MADV_SEQUENTIAL:
8082 			/*
8083 			 * unloading mapping guarantees
8084 			 * detection in segvn_fault
8085 			 */
8086 			ASSERT(seg->s_szc == 0);
8087 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8088 			hat_unload(seg->s_as->a_hat, addr, len,
8089 				HAT_UNLOAD);
8090 			/* FALLTHROUGH */
8091 		case MADV_NORMAL:
8092 		case MADV_RANDOM:
8093 			svd->advice = (uchar_t)behav;
8094 			svd->pageadvice = 0;
8095 			break;
8096 		case MADV_WILLNEED:	/* handled in memcntl */
8097 		case MADV_DONTNEED:	/* handled in memcntl */
8098 		case MADV_FREE:		/* handled above */
8099 			break;
8100 		default:
8101 			err = EINVAL;
8102 		}
8103 	} else {
8104 		caddr_t			eaddr;
8105 		struct seg		*new_seg;
8106 		struct segvn_data	*new_svd;
8107 		u_offset_t		off;
8108 		caddr_t			oldeaddr;
8109 
8110 		page = seg_page(seg, addr);
8111 
8112 		segvn_vpage(seg);
8113 
8114 		switch (behav) {
8115 			struct vpage *bvpp, *evpp;
8116 
8117 		case MADV_ACCESS_LWP:
8118 		case MADV_ACCESS_MANY:
8119 		case MADV_ACCESS_DEFAULT:
8120 			/*
8121 			 * Set memory allocation policy for portion of this
8122 			 * segment
8123 			 */
8124 
8125 			/*
8126 			 * Align address and length of advice to page
8127 			 * boundaries for large pages
8128 			 */
8129 			if (seg->s_szc != 0) {
8130 				size_t	pgsz;
8131 
8132 				pgsz = page_get_pagesize(seg->s_szc);
8133 				addr = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
8134 				len = P2ROUNDUP(len, pgsz);
8135 			}
8136 
8137 			/*
8138 			 * Check to see whether policy is set already
8139 			 */
8140 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8141 
8142 			anon_index = svd->anon_index + page;
8143 			off = svd->offset + (uintptr_t)(addr - seg->s_base);
8144 
8145 			if (svd->type == MAP_SHARED)
8146 				already_set = lgrp_shm_policy_set(policy, amp,
8147 				    anon_index, vp, off, len);
8148 			else
8149 				already_set =
8150 				    (policy == svd->policy_info.mem_policy);
8151 
8152 			/*
8153 			 * If policy set already and it shouldn't be reapplied,
8154 			 * don't do anything.
8155 			 */
8156 			if (already_set &&
8157 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8158 				break;
8159 
8160 			/*
8161 			 * For private memory, need writers lock on
8162 			 * address space because the segment may be
8163 			 * split or concatenated when changing policy
8164 			 */
8165 			if (svd->type == MAP_PRIVATE &&
8166 			    AS_READ_HELD(seg->s_as, &seg->s_as->a_lock)) {
8167 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8168 				return (IE_RETRY);
8169 			}
8170 
8171 			/*
8172 			 * Mark any existing pages in given range for
8173 			 * migration
8174 			 */
8175 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8176 			    vp, svd->offset, 1);
8177 
8178 			/*
8179 			 * Don't need to try to split or concatenate
8180 			 * segments, since policy is same or this is a shared
8181 			 * memory segment
8182 			 */
8183 			if (already_set || svd->type == MAP_SHARED)
8184 				break;
8185 
8186 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
8187 				ASSERT(svd->amp == NULL);
8188 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
8189 				ASSERT(svd->softlockcnt == 0);
8190 				hat_leave_region(seg->s_as->a_hat, svd->rcookie,
8191 				    HAT_REGION_TEXT);
8192 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
8193 			}
8194 
8195 			/*
8196 			 * Split off new segment if advice only applies to a
8197 			 * portion of existing segment starting in middle
8198 			 */
8199 			new_seg = NULL;
8200 			eaddr = addr + len;
8201 			oldeaddr = seg->s_base + seg->s_size;
8202 			if (addr > seg->s_base) {
8203 				/*
8204 				 * Must flush I/O page cache
8205 				 * before splitting segment
8206 				 */
8207 				if (svd->softlockcnt > 0)
8208 					segvn_purge(seg);
8209 
8210 				/*
8211 				 * Split segment and return IE_REATTACH to tell
8212 				 * as_ctl() that current segment changed
8213 				 */
8214 				new_seg = segvn_split_seg(seg, addr);
8215 				new_svd = (struct segvn_data *)new_seg->s_data;
8216 				err = IE_REATTACH;
8217 
8218 				/*
8219 				 * If new segment ends where old one
8220 				 * did, try to concatenate the new
8221 				 * segment with next one.
8222 				 */
8223 				if (eaddr == oldeaddr) {
8224 					/*
8225 					 * Set policy for new segment
8226 					 */
8227 					(void) lgrp_privm_policy_set(policy,
8228 					    &new_svd->policy_info,
8229 					    new_seg->s_size);
8230 
8231 					next = AS_SEGNEXT(new_seg->s_as,
8232 					    new_seg);
8233 
8234 					if (next &&
8235 					    next->s_ops == &segvn_ops &&
8236 					    eaddr == next->s_base)
8237 						(void) segvn_concat(new_seg,
8238 						    next, 1);
8239 				}
8240 			}
8241 
8242 			/*
8243 			 * Split off end of existing segment if advice only
8244 			 * applies to a portion of segment ending before
8245 			 * end of the existing segment
8246 			 */
8247 			if (eaddr < oldeaddr) {
8248 				/*
8249 				 * Must flush I/O page cache
8250 				 * before splitting segment
8251 				 */
8252 				if (svd->softlockcnt > 0)
8253 					segvn_purge(seg);
8254 
8255 				/*
8256 				 * If beginning of old segment was already
8257 				 * split off, use new segment to split end off
8258 				 * from.
8259 				 */
8260 				if (new_seg != NULL && new_seg != seg) {
8261 					/*
8262 					 * Split segment
8263 					 */
8264 					(void) segvn_split_seg(new_seg, eaddr);
8265 
8266 					/*
8267 					 * Set policy for new segment
8268 					 */
8269 					(void) lgrp_privm_policy_set(policy,
8270 					    &new_svd->policy_info,
8271 					    new_seg->s_size);
8272 				} else {
8273 					/*
8274 					 * Split segment and return IE_REATTACH
8275 					 * to tell as_ctl() that current
8276 					 * segment changed
8277 					 */
8278 					(void) segvn_split_seg(seg, eaddr);
8279 					err = IE_REATTACH;
8280 
8281 					(void) lgrp_privm_policy_set(policy,
8282 					    &svd->policy_info, seg->s_size);
8283 
8284 					/*
8285 					 * If new segment starts where old one
8286 					 * did, try to concatenate it with
8287 					 * previous segment.
8288 					 */
8289 					if (addr == seg->s_base) {
8290 						prev = AS_SEGPREV(seg->s_as,
8291 						    seg);
8292 
8293 						/*
8294 						 * Drop lock for private data
8295 						 * of current segment before
8296 						 * concatenating (deleting) it
8297 						 */
8298 						if (prev &&
8299 						    prev->s_ops ==
8300 						    &segvn_ops &&
8301 						    addr == prev->s_base +
8302 						    prev->s_size) {
8303 							SEGVN_LOCK_EXIT(
8304 							    seg->s_as,
8305 							    &svd->lock);
8306 							(void) segvn_concat(
8307 							    prev, seg, 1);
8308 							return (err);
8309 						}
8310 					}
8311 				}
8312 			}
8313 			break;
8314 		case MADV_SEQUENTIAL:
8315 			ASSERT(seg->s_szc == 0);
8316 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8317 			hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
8318 			/* FALLTHROUGH */
8319 		case MADV_NORMAL:
8320 		case MADV_RANDOM:
8321 			bvpp = &svd->vpage[page];
8322 			evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8323 			for (; bvpp < evpp; bvpp++)
8324 				VPP_SETADVICE(bvpp, behav);
8325 			svd->advice = MADV_NORMAL;
8326 			break;
8327 		case MADV_WILLNEED:	/* handled in memcntl */
8328 		case MADV_DONTNEED:	/* handled in memcntl */
8329 		case MADV_FREE:		/* handled above */
8330 			break;
8331 		default:
8332 			err = EINVAL;
8333 		}
8334 	}
8335 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8336 	return (err);
8337 }
8338 
8339 /*
8340  * Create a vpage structure for this seg.
8341  */
8342 static void
8343 segvn_vpage(struct seg *seg)
8344 {
8345 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8346 	struct vpage *vp, *evp;
8347 
8348 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8349 
8350 	/*
8351 	 * If no vpage structure exists, allocate one.  Copy the protections
8352 	 * and the advice from the segment itself to the individual pages.
8353 	 */
8354 	if (svd->vpage == NULL) {
8355 		svd->pageadvice = 1;
8356 		svd->vpage = kmem_zalloc(seg_pages(seg) * sizeof (struct vpage),
8357 		    KM_SLEEP);
8358 		evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
8359 		for (vp = svd->vpage; vp < evp; vp++) {
8360 			VPP_SETPROT(vp, svd->prot);
8361 			VPP_SETADVICE(vp, svd->advice);
8362 		}
8363 	}
8364 }
8365 
8366 /*
8367  * Dump the pages belonging to this segvn segment.
8368  */
8369 static void
8370 segvn_dump(struct seg *seg)
8371 {
8372 	struct segvn_data *svd;
8373 	page_t *pp;
8374 	struct anon_map *amp;
8375 	ulong_t	anon_index;
8376 	struct vnode *vp;
8377 	u_offset_t off, offset;
8378 	pfn_t pfn;
8379 	pgcnt_t page, npages;
8380 	caddr_t addr;
8381 
8382 	npages = seg_pages(seg);
8383 	svd = (struct segvn_data *)seg->s_data;
8384 	vp = svd->vp;
8385 	off = offset = svd->offset;
8386 	addr = seg->s_base;
8387 
8388 	if ((amp = svd->amp) != NULL) {
8389 		anon_index = svd->anon_index;
8390 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8391 	}
8392 
8393 	for (page = 0; page < npages; page++, offset += PAGESIZE) {
8394 		struct anon *ap;
8395 		int we_own_it = 0;
8396 
8397 		if (amp && (ap = anon_get_ptr(svd->amp->ahp, anon_index++))) {
8398 			swap_xlate_nopanic(ap, &vp, &off);
8399 		} else {
8400 			vp = svd->vp;
8401 			off = offset;
8402 		}
8403 
8404 		/*
8405 		 * If pp == NULL, the page either does not exist
8406 		 * or is exclusively locked.  So determine if it
8407 		 * exists before searching for it.
8408 		 */
8409 
8410 		if ((pp = page_lookup_nowait(vp, off, SE_SHARED)))
8411 			we_own_it = 1;
8412 		else
8413 			pp = page_exists(vp, off);
8414 
8415 		if (pp) {
8416 			pfn = page_pptonum(pp);
8417 			dump_addpage(seg->s_as, addr, pfn);
8418 			if (we_own_it)
8419 				page_unlock(pp);
8420 		}
8421 		addr += PAGESIZE;
8422 		dump_timeleft = dump_timeout;
8423 	}
8424 
8425 	if (amp != NULL)
8426 		ANON_LOCK_EXIT(&amp->a_rwlock);
8427 }
8428 
8429 /*
8430  * lock/unlock anon pages over a given range. Return shadow list
8431  */
8432 static int
8433 segvn_pagelock(struct seg *seg, caddr_t addr, size_t len, struct page ***ppp,
8434     enum lock_type type, enum seg_rw rw)
8435 {
8436 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8437 	size_t np, adjustpages = 0, npages = (len >> PAGESHIFT);
8438 	ulong_t anon_index;
8439 	uint_t protchk;
8440 	uint_t error;
8441 	struct anon_map *amp;
8442 	struct page **pplist, **pl, *pp;
8443 	caddr_t a;
8444 	size_t page;
8445 	caddr_t lpgaddr, lpgeaddr;
8446 	pgcnt_t szc0_npages = 0;
8447 
8448 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_START,
8449 		"segvn_pagelock: start seg %p addr %p", seg, addr);
8450 
8451 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
8452 	if (seg->s_szc != 0 && (type == L_PAGELOCK || type == L_PAGEUNLOCK)) {
8453 		/*
8454 		 * We are adjusting the pagelock region to the large page size
8455 		 * boundary because the unlocked part of a large page cannot
8456 		 * be freed anyway unless all constituent pages of a large
8457 		 * page are locked. Therefore this adjustment allows us to
8458 		 * decrement availrmem by the right value (note we don't want
8459 		 * to just decrement availrem by the large page size without
8460 		 * adjusting addr and len because then we may end up
8461 		 * decrementing availrmem by large page size for every
8462 		 * constituent page locked by a new as_pagelock call).
8463 		 * as_pageunlock caller must always match as_pagelock call's
8464 		 * addr and len.
8465 		 *
8466 		 * Note segment's page size cannot change while we are holding
8467 		 * as lock.  And then it cannot change while softlockcnt is
8468 		 * not 0. This will allow us to correctly recalculate large
8469 		 * page size region for the matching pageunlock/reclaim call.
8470 		 *
8471 		 * for pageunlock *ppp points to the pointer of page_t that
8472 		 * corresponds to the real unadjusted start address. Similar
8473 		 * for pagelock *ppp must point to the pointer of page_t that
8474 		 * corresponds to the real unadjusted start address.
8475 		 */
8476 		size_t pgsz = page_get_pagesize(seg->s_szc);
8477 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
8478 		adjustpages = ((uintptr_t)(addr - lpgaddr)) >> PAGESHIFT;
8479 	}
8480 
8481 	if (type == L_PAGEUNLOCK) {
8482 
8483 		/*
8484 		 * update hat ref bits for /proc. We need to make sure
8485 		 * that threads tracing the ref and mod bits of the
8486 		 * address space get the right data.
8487 		 * Note: page ref and mod bits are updated at reclaim time
8488 		 */
8489 		if (seg->s_as->a_vbits) {
8490 			for (a = addr; a < addr + len; a += PAGESIZE) {
8491 				if (rw == S_WRITE) {
8492 					hat_setstat(seg->s_as, a,
8493 					    PAGESIZE, P_REF | P_MOD);
8494 				} else {
8495 					hat_setstat(seg->s_as, a,
8496 					    PAGESIZE, P_REF);
8497 				}
8498 			}
8499 		}
8500 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8501 		if (seg->s_szc != 0) {
8502 			VM_STAT_ADD(segvnvmstats.pagelock[0]);
8503 			seg_pinactive(seg, lpgaddr, lpgeaddr - lpgaddr,
8504 			    *ppp - adjustpages, rw, segvn_reclaim);
8505 		} else {
8506 			seg_pinactive(seg, addr, len, *ppp, rw, segvn_reclaim);
8507 		}
8508 
8509 		/*
8510 		 * If someone is blocked while unmapping, we purge
8511 		 * segment page cache and thus reclaim pplist synchronously
8512 		 * without waiting for seg_pasync_thread. This speeds up
8513 		 * unmapping in cases where munmap(2) is called, while
8514 		 * raw async i/o is still in progress or where a thread
8515 		 * exits on data fault in a multithreaded application.
8516 		 */
8517 		if (AS_ISUNMAPWAIT(seg->s_as) && (svd->softlockcnt > 0)) {
8518 			/*
8519 			 * Even if we grab segvn WRITER's lock or segp_slock
8520 			 * here, there might be another thread which could've
8521 			 * successfully performed lookup/insert just before
8522 			 * we acquired the lock here.  So, grabbing either
8523 			 * lock here is of not much use.  Until we devise
8524 			 * a strategy at upper layers to solve the
8525 			 * synchronization issues completely, we expect
8526 			 * applications to handle this appropriately.
8527 			 */
8528 			segvn_purge(seg);
8529 		}
8530 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8531 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_UNLOCK_END,
8532 			"segvn_pagelock: unlock seg %p addr %p", seg, addr);
8533 		return (0);
8534 	} else if (type == L_PAGERECLAIM) {
8535 		VM_STAT_COND_ADD(seg->s_szc != 0, segvnvmstats.pagelock[1]);
8536 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8537 		(void) segvn_reclaim(seg, addr, len, *ppp, rw);
8538 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8539 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_UNLOCK_END,
8540 			"segvn_pagelock: reclaim seg %p addr %p", seg, addr);
8541 		return (0);
8542 	}
8543 
8544 	if (seg->s_szc != 0) {
8545 		VM_STAT_ADD(segvnvmstats.pagelock[2]);
8546 		addr = lpgaddr;
8547 		len = lpgeaddr - lpgaddr;
8548 		npages = (len >> PAGESHIFT);
8549 	}
8550 
8551 	/*
8552 	 * for now we only support pagelock to anon memory. We've to check
8553 	 * protections for vnode objects and call into the vnode driver.
8554 	 * That's too much for a fast path. Let the fault entry point handle it.
8555 	 */
8556 	if (svd->vp != NULL) {
8557 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
8558 		    "segvn_pagelock: mapped vnode seg %p addr %p", seg, addr);
8559 		*ppp = NULL;
8560 		return (ENOTSUP);
8561 	}
8562 
8563 	/*
8564 	 * if anonmap is not yet created, let the fault entry point populate it
8565 	 * with anon ptrs.
8566 	 */
8567 	if ((amp = svd->amp) == NULL) {
8568 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
8569 		    "segvn_pagelock: anonmap null seg %p addr %p", seg, addr);
8570 		*ppp = NULL;
8571 		return (EFAULT);
8572 	}
8573 
8574 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8575 
8576 	/*
8577 	 * we acquire segp_slock to prevent duplicate entries
8578 	 * in seg_pcache
8579 	 */
8580 	mutex_enter(&svd->segp_slock);
8581 
8582 	/*
8583 	 * try to find pages in segment page cache
8584 	 */
8585 	pplist = seg_plookup(seg, addr, len, rw);
8586 	if (pplist != NULL) {
8587 		mutex_exit(&svd->segp_slock);
8588 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8589 		*ppp = pplist + adjustpages;
8590 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_HIT_END,
8591 			"segvn_pagelock: cache hit seg %p addr %p", seg, addr);
8592 		return (0);
8593 	}
8594 
8595 	if (rw == S_READ) {
8596 		protchk = PROT_READ;
8597 	} else {
8598 		protchk = PROT_WRITE;
8599 	}
8600 
8601 	if (svd->pageprot == 0) {
8602 		if ((svd->prot & protchk) == 0) {
8603 			mutex_exit(&svd->segp_slock);
8604 			error = EFAULT;
8605 			goto out;
8606 		}
8607 	} else {
8608 		/*
8609 		 * check page protections
8610 		 */
8611 		for (a = addr; a < addr + len; a += PAGESIZE) {
8612 			struct vpage *vp;
8613 
8614 			vp = &svd->vpage[seg_page(seg, a)];
8615 			if ((VPP_PROT(vp) & protchk) == 0) {
8616 				mutex_exit(&svd->segp_slock);
8617 				error = EFAULT;
8618 				goto out;
8619 			}
8620 		}
8621 	}
8622 
8623 	/*
8624 	 * Avoid per page overhead of segvn_slock_anonpages() for small
8625 	 * pages. For large pages segvn_slock_anonpages() only does real
8626 	 * work once per large page.  The tradeoff is that we may decrement
8627 	 * availrmem more than once for the same page but this is ok
8628 	 * for small pages.
8629 	 */
8630 	if (seg->s_szc == 0) {
8631 		mutex_enter(&freemem_lock);
8632 		if (availrmem < tune.t_minarmem + npages) {
8633 			mutex_exit(&freemem_lock);
8634 			mutex_exit(&svd->segp_slock);
8635 			error = ENOMEM;
8636 			goto out;
8637 		}
8638 		availrmem -= npages;
8639 		mutex_exit(&freemem_lock);
8640 	}
8641 
8642 	pplist = kmem_alloc(sizeof (page_t *) * npages, KM_SLEEP);
8643 	pl = pplist;
8644 	*ppp = pplist + adjustpages;
8645 
8646 	page = seg_page(seg, addr);
8647 	anon_index = svd->anon_index + page;
8648 
8649 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8650 	for (a = addr; a < addr + len; a += PAGESIZE, anon_index++) {
8651 		struct anon *ap;
8652 		struct vnode *vp;
8653 		u_offset_t off;
8654 		anon_sync_obj_t cookie;
8655 
8656 		anon_array_enter(amp, anon_index, &cookie);
8657 		ap = anon_get_ptr(amp->ahp, anon_index);
8658 		if (ap == NULL) {
8659 			anon_array_exit(&cookie);
8660 			break;
8661 		} else {
8662 			/*
8663 			 * We must never use seg_pcache for COW pages
8664 			 * because we might end up with original page still
8665 			 * lying in seg_pcache even after private page is
8666 			 * created. This leads to data corruption as
8667 			 * aio_write refers to the page still in cache
8668 			 * while all other accesses refer to the private
8669 			 * page.
8670 			 */
8671 			if (ap->an_refcnt != 1) {
8672 				anon_array_exit(&cookie);
8673 				break;
8674 			}
8675 		}
8676 		swap_xlate(ap, &vp, &off);
8677 		anon_array_exit(&cookie);
8678 
8679 		pp = page_lookup_nowait(vp, off, SE_SHARED);
8680 		if (pp == NULL) {
8681 			break;
8682 		}
8683 		if (seg->s_szc != 0 || pp->p_szc != 0) {
8684 			if (!segvn_slock_anonpages(pp, a == addr)) {
8685 				page_unlock(pp);
8686 				break;
8687 			}
8688 		} else {
8689 			szc0_npages++;
8690 		}
8691 		*pplist++ = pp;
8692 	}
8693 	ANON_LOCK_EXIT(&amp->a_rwlock);
8694 
8695 	ASSERT(npages >= szc0_npages);
8696 
8697 	if (a >= addr + len) {
8698 		mutex_enter(&freemem_lock);
8699 		if (seg->s_szc == 0 && npages != szc0_npages) {
8700 			ASSERT(svd->type == MAP_SHARED && amp->a_szc > 0);
8701 			availrmem += (npages - szc0_npages);
8702 		}
8703 		svd->softlockcnt += npages;
8704 		segvn_pages_locked += npages;
8705 		mutex_exit(&freemem_lock);
8706 		(void) seg_pinsert(seg, addr, len, pl, rw, SEGP_ASYNC_FLUSH,
8707 			segvn_reclaim);
8708 		mutex_exit(&svd->segp_slock);
8709 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8710 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_FILL_END,
8711 		    "segvn_pagelock: cache fill seg %p addr %p", seg, addr);
8712 		return (0);
8713 	}
8714 
8715 	mutex_exit(&svd->segp_slock);
8716 	if (seg->s_szc == 0) {
8717 		mutex_enter(&freemem_lock);
8718 		availrmem += npages;
8719 		mutex_exit(&freemem_lock);
8720 	}
8721 	error = EFAULT;
8722 	pplist = pl;
8723 	np = ((uintptr_t)(a - addr)) >> PAGESHIFT;
8724 	while (np > (uint_t)0) {
8725 		ASSERT(PAGE_LOCKED(*pplist));
8726 		if (seg->s_szc != 0 || (*pplist)->p_szc != 0) {
8727 			segvn_sunlock_anonpages(*pplist, pplist == pl);
8728 		}
8729 		page_unlock(*pplist);
8730 		np--;
8731 		pplist++;
8732 	}
8733 	kmem_free(pl, sizeof (page_t *) * npages);
8734 out:
8735 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8736 	*ppp = NULL;
8737 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
8738 		"segvn_pagelock: cache miss seg %p addr %p", seg, addr);
8739 	return (error);
8740 }
8741 
8742 /*
8743  * purge any cached pages in the I/O page cache
8744  */
8745 static void
8746 segvn_purge(struct seg *seg)
8747 {
8748 	seg_ppurge(seg);
8749 }
8750 
8751 static int
8752 segvn_reclaim(struct seg *seg, caddr_t addr, size_t len, struct page **pplist,
8753 	enum seg_rw rw)
8754 {
8755 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8756 	pgcnt_t np, npages;
8757 	struct page **pl;
8758 	pgcnt_t szc0_npages = 0;
8759 
8760 #ifdef lint
8761 	addr = addr;
8762 #endif
8763 
8764 	npages = np = (len >> PAGESHIFT);
8765 	ASSERT(npages);
8766 	pl = pplist;
8767 	if (seg->s_szc != 0) {
8768 		size_t pgsz = page_get_pagesize(seg->s_szc);
8769 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
8770 			panic("segvn_reclaim: unaligned addr or len");
8771 			/*NOTREACHED*/
8772 		}
8773 	}
8774 
8775 	ASSERT(svd->vp == NULL && svd->amp != NULL);
8776 
8777 	while (np > (uint_t)0) {
8778 		if (rw == S_WRITE) {
8779 			hat_setrefmod(*pplist);
8780 		} else {
8781 			hat_setref(*pplist);
8782 		}
8783 		if (seg->s_szc != 0 || (*pplist)->p_szc != 0) {
8784 			segvn_sunlock_anonpages(*pplist, pplist == pl);
8785 		} else {
8786 			szc0_npages++;
8787 		}
8788 		page_unlock(*pplist);
8789 		np--;
8790 		pplist++;
8791 	}
8792 	kmem_free(pl, sizeof (page_t *) * npages);
8793 
8794 	mutex_enter(&freemem_lock);
8795 	segvn_pages_locked -= npages;
8796 	svd->softlockcnt -= npages;
8797 	if (szc0_npages != 0) {
8798 		availrmem += szc0_npages;
8799 	}
8800 	mutex_exit(&freemem_lock);
8801 	if (svd->softlockcnt <= 0) {
8802 		if (AS_ISUNMAPWAIT(seg->s_as)) {
8803 			mutex_enter(&seg->s_as->a_contents);
8804 			if (AS_ISUNMAPWAIT(seg->s_as)) {
8805 				AS_CLRUNMAPWAIT(seg->s_as);
8806 				cv_broadcast(&seg->s_as->a_cv);
8807 			}
8808 			mutex_exit(&seg->s_as->a_contents);
8809 		}
8810 	}
8811 	return (0);
8812 }
8813 /*
8814  * get a memory ID for an addr in a given segment
8815  *
8816  * XXX only creates PAGESIZE pages if anon slots are not initialized.
8817  * At fault time they will be relocated into larger pages.
8818  */
8819 static int
8820 segvn_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
8821 {
8822 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8823 	struct anon 	*ap = NULL;
8824 	ulong_t		anon_index;
8825 	struct anon_map	*amp;
8826 	anon_sync_obj_t cookie;
8827 
8828 	if (svd->type == MAP_PRIVATE) {
8829 		memidp->val[0] = (uintptr_t)seg->s_as;
8830 		memidp->val[1] = (uintptr_t)addr;
8831 		return (0);
8832 	}
8833 
8834 	if (svd->type == MAP_SHARED) {
8835 		if (svd->vp) {
8836 			memidp->val[0] = (uintptr_t)svd->vp;
8837 			memidp->val[1] = (u_longlong_t)svd->offset +
8838 			    (uintptr_t)(addr - seg->s_base);
8839 			return (0);
8840 		} else {
8841 
8842 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8843 			if ((amp = svd->amp) != NULL) {
8844 				anon_index = svd->anon_index +
8845 				    seg_page(seg, addr);
8846 			}
8847 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8848 
8849 			ASSERT(amp != NULL);
8850 
8851 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8852 			anon_array_enter(amp, anon_index, &cookie);
8853 			ap = anon_get_ptr(amp->ahp, anon_index);
8854 			if (ap == NULL) {
8855 				page_t		*pp;
8856 
8857 				pp = anon_zero(seg, addr, &ap, svd->cred);
8858 				if (pp == NULL) {
8859 					anon_array_exit(&cookie);
8860 					ANON_LOCK_EXIT(&amp->a_rwlock);
8861 					return (ENOMEM);
8862 				}
8863 				ASSERT(anon_get_ptr(amp->ahp, anon_index)
8864 								== NULL);
8865 				(void) anon_set_ptr(amp->ahp, anon_index,
8866 				    ap, ANON_SLEEP);
8867 				page_unlock(pp);
8868 			}
8869 
8870 			anon_array_exit(&cookie);
8871 			ANON_LOCK_EXIT(&amp->a_rwlock);
8872 
8873 			memidp->val[0] = (uintptr_t)ap;
8874 			memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
8875 			return (0);
8876 		}
8877 	}
8878 	return (EINVAL);
8879 }
8880 
8881 static int
8882 sameprot(struct seg *seg, caddr_t a, size_t len)
8883 {
8884 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8885 	struct vpage *vpage;
8886 	spgcnt_t pages = btop(len);
8887 	uint_t prot;
8888 
8889 	if (svd->pageprot == 0)
8890 		return (1);
8891 
8892 	ASSERT(svd->vpage != NULL);
8893 
8894 	vpage = &svd->vpage[seg_page(seg, a)];
8895 	prot = VPP_PROT(vpage);
8896 	vpage++;
8897 	pages--;
8898 	while (pages-- > 0) {
8899 		if (prot != VPP_PROT(vpage))
8900 			return (0);
8901 		vpage++;
8902 	}
8903 	return (1);
8904 }
8905 
8906 /*
8907  * Get memory allocation policy info for specified address in given segment
8908  */
8909 static lgrp_mem_policy_info_t *
8910 segvn_getpolicy(struct seg *seg, caddr_t addr)
8911 {
8912 	struct anon_map		*amp;
8913 	ulong_t			anon_index;
8914 	lgrp_mem_policy_info_t	*policy_info;
8915 	struct segvn_data	*svn_data;
8916 	u_offset_t		vn_off;
8917 	vnode_t			*vp;
8918 
8919 	ASSERT(seg != NULL);
8920 
8921 	svn_data = (struct segvn_data *)seg->s_data;
8922 	if (svn_data == NULL)
8923 		return (NULL);
8924 
8925 	/*
8926 	 * Get policy info for private or shared memory
8927 	 */
8928 	if (svn_data->type != MAP_SHARED) {
8929 		if (svn_data->tr_state != SEGVN_TR_ON) {
8930 			policy_info = &svn_data->policy_info;
8931 		} else {
8932 			policy_info = &svn_data->tr_policy_info;
8933 			ASSERT(policy_info->mem_policy ==
8934 			    LGRP_MEM_POLICY_NEXT_SEG);
8935 		}
8936 	} else {
8937 		amp = svn_data->amp;
8938 		anon_index = svn_data->anon_index + seg_page(seg, addr);
8939 		vp = svn_data->vp;
8940 		vn_off = svn_data->offset + (uintptr_t)(addr - seg->s_base);
8941 		policy_info = lgrp_shm_policy_get(amp, anon_index, vp, vn_off);
8942 	}
8943 
8944 	return (policy_info);
8945 }
8946 
8947 /*ARGSUSED*/
8948 static int
8949 segvn_capable(struct seg *seg, segcapability_t capability)
8950 {
8951 	return (0);
8952 }
8953 
8954 /*
8955  * Bind text vnode segment to an amp. If we bind successfully mappings will be
8956  * established to per vnode mapping per lgroup amp pages instead of to vnode
8957  * pages. There's one amp per vnode text mapping per lgroup. Many processes
8958  * may share the same text replication amp. If a suitable amp doesn't already
8959  * exist in svntr hash table create a new one.  We may fail to bind to amp if
8960  * segment is not eligible for text replication.  Code below first checks for
8961  * these conditions. If binding is successful segment tr_state is set to on
8962  * and svd->amp points to the amp to use. Otherwise tr_state is set to off and
8963  * svd->amp remains as NULL.
8964  */
8965 static void
8966 segvn_textrepl(struct seg *seg)
8967 {
8968 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
8969 	vnode_t			*vp = svd->vp;
8970 	u_offset_t		off = svd->offset;
8971 	size_t			size = seg->s_size;
8972 	u_offset_t		eoff = off + size;
8973 	uint_t			szc = seg->s_szc;
8974 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
8975 	svntr_t			*svntrp;
8976 	struct vattr		va;
8977 	proc_t			*p = seg->s_as->a_proc;
8978 	lgrp_id_t		lgrp_id;
8979 	lgrp_id_t		olid;
8980 	int			first;
8981 	struct anon_map		*amp;
8982 
8983 	ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
8984 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8985 	ASSERT(p != NULL);
8986 	ASSERT(svd->tr_state == SEGVN_TR_INIT);
8987 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
8988 	ASSERT(svd->flags & MAP_TEXT);
8989 	ASSERT(svd->type == MAP_PRIVATE);
8990 	ASSERT(vp != NULL && svd->amp == NULL);
8991 	ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
8992 	ASSERT(!(svd->flags & MAP_NORESERVE) && svd->swresv == 0);
8993 	ASSERT(seg->s_as != &kas);
8994 	ASSERT(off < eoff);
8995 	ASSERT(svntr_hashtab != NULL);
8996 
8997 	/*
8998 	 * If numa optimizations are no longer desired bail out.
8999 	 */
9000 	if (!lgrp_optimizations()) {
9001 		svd->tr_state = SEGVN_TR_OFF;
9002 		return;
9003 	}
9004 
9005 	/*
9006 	 * Avoid creating anon maps with size bigger than the file size.
9007 	 * If VOP_GETATTR() call fails bail out.
9008 	 */
9009 	va.va_mask = AT_SIZE | AT_MTIME | AT_CTIME;
9010 	if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL) != 0) {
9011 		svd->tr_state = SEGVN_TR_OFF;
9012 		SEGVN_TR_ADDSTAT(gaerr);
9013 		return;
9014 	}
9015 	if (btopr(va.va_size) < btopr(eoff)) {
9016 		svd->tr_state = SEGVN_TR_OFF;
9017 		SEGVN_TR_ADDSTAT(overmap);
9018 		return;
9019 	}
9020 
9021 	/*
9022 	 * VVMEXEC may not be set yet if exec() prefaults text segment. Set
9023 	 * this flag now before vn_is_mapped(V_WRITE) so that MAP_SHARED
9024 	 * mapping that checks if trcache for this vnode needs to be
9025 	 * invalidated can't miss us.
9026 	 */
9027 	if (!(vp->v_flag & VVMEXEC)) {
9028 		mutex_enter(&vp->v_lock);
9029 		vp->v_flag |= VVMEXEC;
9030 		mutex_exit(&vp->v_lock);
9031 	}
9032 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9033 	/*
9034 	 * Bail out if potentially MAP_SHARED writable mappings exist to this
9035 	 * vnode.  We don't want to use old file contents from existing
9036 	 * replicas if this mapping was established after the original file
9037 	 * was changed.
9038 	 */
9039 	if (vn_is_mapped(vp, V_WRITE)) {
9040 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9041 		svd->tr_state = SEGVN_TR_OFF;
9042 		SEGVN_TR_ADDSTAT(wrcnt);
9043 		return;
9044 	}
9045 	svntrp = svntr_hashtab[hash].tr_head;
9046 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9047 		ASSERT(svntrp->tr_refcnt != 0);
9048 		if (svntrp->tr_vp != vp) {
9049 			continue;
9050 		}
9051 
9052 		/*
9053 		 * Bail out if the file or its attributes were changed after
9054 		 * this replication entry was created since we need to use the
9055 		 * latest file contents. Note that mtime test alone is not
9056 		 * sufficient because a user can explicitly change mtime via
9057 		 * utimes(2) interfaces back to the old value after modifiying
9058 		 * the file contents. To detect this case we also have to test
9059 		 * ctime which among other things records the time of the last
9060 		 * mtime change by utimes(2). ctime is not changed when the file
9061 		 * is only read or executed so we expect that typically existing
9062 		 * replication amp's can be used most of the time.
9063 		 */
9064 		if (!svntrp->tr_valid ||
9065 		    svntrp->tr_mtime.tv_sec != va.va_mtime.tv_sec ||
9066 		    svntrp->tr_mtime.tv_nsec != va.va_mtime.tv_nsec ||
9067 		    svntrp->tr_ctime.tv_sec != va.va_ctime.tv_sec ||
9068 		    svntrp->tr_ctime.tv_nsec != va.va_ctime.tv_nsec) {
9069 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9070 			svd->tr_state = SEGVN_TR_OFF;
9071 			SEGVN_TR_ADDSTAT(stale);
9072 			return;
9073 		}
9074 		/*
9075 		 * if off, eoff and szc match current segment we found the
9076 		 * existing entry we can use.
9077 		 */
9078 		if (svntrp->tr_off == off && svntrp->tr_eoff == eoff &&
9079 		    svntrp->tr_szc == szc) {
9080 			break;
9081 		}
9082 		/*
9083 		 * Don't create different but overlapping in file offsets
9084 		 * entries to avoid replication of the same file pages more
9085 		 * than once per lgroup.
9086 		 */
9087 		if ((off >= svntrp->tr_off && off < svntrp->tr_eoff) ||
9088 		    (eoff > svntrp->tr_off && eoff <= svntrp->tr_eoff)) {
9089 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9090 			svd->tr_state = SEGVN_TR_OFF;
9091 			SEGVN_TR_ADDSTAT(overlap);
9092 			return;
9093 		}
9094 	}
9095 	/*
9096 	 * If we didn't find existing entry create a new one.
9097 	 */
9098 	if (svntrp == NULL) {
9099 		svntrp = kmem_cache_alloc(svntr_cache, KM_NOSLEEP);
9100 		if (svntrp == NULL) {
9101 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9102 			svd->tr_state = SEGVN_TR_OFF;
9103 			SEGVN_TR_ADDSTAT(nokmem);
9104 			return;
9105 		}
9106 #ifdef DEBUG
9107 		{
9108 			lgrp_id_t i;
9109 			for (i = 0; i < NLGRPS_MAX; i++) {
9110 				ASSERT(svntrp->tr_amp[i] == NULL);
9111 			}
9112 		}
9113 #endif /* DEBUG */
9114 		svntrp->tr_vp = vp;
9115 		svntrp->tr_off = off;
9116 		svntrp->tr_eoff = eoff;
9117 		svntrp->tr_szc = szc;
9118 		svntrp->tr_valid = 1;
9119 		svntrp->tr_mtime = va.va_mtime;
9120 		svntrp->tr_ctime = va.va_ctime;
9121 		svntrp->tr_refcnt = 0;
9122 		svntrp->tr_next = svntr_hashtab[hash].tr_head;
9123 		svntr_hashtab[hash].tr_head = svntrp;
9124 	}
9125 	first = 1;
9126 again:
9127 	/*
9128 	 * We want to pick a replica with pages on main thread's (t_tid = 1,
9129 	 * aka T1) lgrp. Currently text replication is only optimized for
9130 	 * workloads that either have all threads of a process on the same
9131 	 * lgrp or execute their large text primarily on main thread.
9132 	 */
9133 	lgrp_id = p->p_t1_lgrpid;
9134 	if (lgrp_id == LGRP_NONE) {
9135 		/*
9136 		 * In case exec() prefaults text on non main thread use
9137 		 * current thread lgrpid.  It will become main thread anyway
9138 		 * soon.
9139 		 */
9140 		lgrp_id = lgrp_home_id(curthread);
9141 	}
9142 	/*
9143 	 * Set p_tr_lgrpid to lgrpid if it hasn't been set yet.  Otherwise
9144 	 * just set it to NLGRPS_MAX if it's different from current process T1
9145 	 * home lgrp.  p_tr_lgrpid is used to detect if process uses text
9146 	 * replication and T1 new home is different from lgrp used for text
9147 	 * replication. When this happens asyncronous segvn thread rechecks if
9148 	 * segments should change lgrps used for text replication.  If we fail
9149 	 * to set p_tr_lgrpid with cas32 then set it to NLGRPS_MAX without cas
9150 	 * if it's not already NLGRPS_MAX and not equal lgrp_id we want to
9151 	 * use.  We don't need to use cas in this case because another thread
9152 	 * that races in between our non atomic check and set may only change
9153 	 * p_tr_lgrpid to NLGRPS_MAX at this point.
9154 	 */
9155 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9156 	olid = p->p_tr_lgrpid;
9157 	if (lgrp_id != olid && olid != NLGRPS_MAX) {
9158 		lgrp_id_t nlid = (olid == LGRP_NONE) ? lgrp_id : NLGRPS_MAX;
9159 		if (cas32((uint32_t *)&p->p_tr_lgrpid, olid, nlid) != olid) {
9160 			olid = p->p_tr_lgrpid;
9161 			ASSERT(olid != LGRP_NONE);
9162 			if (olid != lgrp_id && olid != NLGRPS_MAX) {
9163 				p->p_tr_lgrpid = NLGRPS_MAX;
9164 			}
9165 		}
9166 		ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9167 		membar_producer();
9168 		/*
9169 		 * lgrp_move_thread() won't schedule async recheck after
9170 		 * p->p_t1_lgrpid update unless p->p_tr_lgrpid is not
9171 		 * LGRP_NONE. Recheck p_t1_lgrpid once now that p->p_tr_lgrpid
9172 		 * is not LGRP_NONE.
9173 		 */
9174 		if (first && p->p_t1_lgrpid != LGRP_NONE &&
9175 		    p->p_t1_lgrpid != lgrp_id) {
9176 			first = 0;
9177 			goto again;
9178 		}
9179 	}
9180 	/*
9181 	 * If no amp was created yet for lgrp_id create a new one as long as
9182 	 * we have enough memory to afford it.
9183 	 */
9184 	if ((amp = svntrp->tr_amp[lgrp_id]) == NULL) {
9185 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9186 		if (trmem > segvn_textrepl_max_bytes) {
9187 			SEGVN_TR_ADDSTAT(normem);
9188 			goto fail;
9189 		}
9190 		if (anon_try_resv_zone(size, NULL) == 0) {
9191 			SEGVN_TR_ADDSTAT(noanon);
9192 			goto fail;
9193 		}
9194 		amp = anonmap_alloc(size, size, ANON_NOSLEEP);
9195 		if (amp == NULL) {
9196 			anon_unresv_zone(size, NULL);
9197 			SEGVN_TR_ADDSTAT(nokmem);
9198 			goto fail;
9199 		}
9200 		ASSERT(amp->refcnt == 1);
9201 		amp->a_szc = szc;
9202 		svntrp->tr_amp[lgrp_id] = amp;
9203 		SEGVN_TR_ADDSTAT(newamp);
9204 	}
9205 	svntrp->tr_refcnt++;
9206 	ASSERT(svd->svn_trnext == NULL);
9207 	ASSERT(svd->svn_trprev == NULL);
9208 	svd->svn_trnext = svntrp->tr_svnhead;
9209 	svd->svn_trprev = NULL;
9210 	if (svntrp->tr_svnhead != NULL) {
9211 		svntrp->tr_svnhead->svn_trprev = svd;
9212 	}
9213 	svntrp->tr_svnhead = svd;
9214 	ASSERT(amp->a_szc == szc && amp->size == size && amp->swresv == size);
9215 	ASSERT(amp->refcnt >= 1);
9216 	svd->amp = amp;
9217 	svd->anon_index = 0;
9218 	svd->tr_policy_info.mem_policy = LGRP_MEM_POLICY_NEXT_SEG;
9219 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
9220 	svd->tr_state = SEGVN_TR_ON;
9221 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9222 	SEGVN_TR_ADDSTAT(repl);
9223 	return;
9224 fail:
9225 	ASSERT(segvn_textrepl_bytes >= size);
9226 	atomic_add_long(&segvn_textrepl_bytes, -size);
9227 	ASSERT(svntrp != NULL);
9228 	ASSERT(svntrp->tr_amp[lgrp_id] == NULL);
9229 	if (svntrp->tr_refcnt == 0) {
9230 		ASSERT(svntrp == svntr_hashtab[hash].tr_head);
9231 		svntr_hashtab[hash].tr_head = svntrp->tr_next;
9232 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9233 		kmem_cache_free(svntr_cache, svntrp);
9234 	} else {
9235 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9236 	}
9237 	svd->tr_state = SEGVN_TR_OFF;
9238 }
9239 
9240 /*
9241  * Convert seg back to regular vnode mapping seg by unbinding it from its text
9242  * replication amp.  This routine is most typically called when segment is
9243  * unmapped but can also be called when segment no longer qualifies for text
9244  * replication (e.g. due to protection changes). If unload_unmap is set use
9245  * HAT_UNLOAD_UNMAP flag in hat_unload_callback().  If we are the last user of
9246  * svntr free all its anon maps and remove it from the hash table.
9247  */
9248 static void
9249 segvn_textunrepl(struct seg *seg, int unload_unmap)
9250 {
9251 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
9252 	vnode_t			*vp = svd->vp;
9253 	u_offset_t		off = svd->offset;
9254 	size_t			size = seg->s_size;
9255 	u_offset_t		eoff = off + size;
9256 	uint_t			szc = seg->s_szc;
9257 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9258 	svntr_t			*svntrp;
9259 	svntr_t			**prv_svntrp;
9260 	lgrp_id_t		lgrp_id = svd->tr_policy_info.mem_lgrpid;
9261 	lgrp_id_t		i;
9262 
9263 	ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9264 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
9265 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9266 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9267 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9268 	ASSERT(svd->amp != NULL);
9269 	ASSERT(svd->amp->refcnt >= 1);
9270 	ASSERT(svd->anon_index == 0);
9271 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9272 	ASSERT(svntr_hashtab != NULL);
9273 
9274 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9275 	prv_svntrp = &svntr_hashtab[hash].tr_head;
9276 	for (; (svntrp = *prv_svntrp) != NULL; prv_svntrp = &svntrp->tr_next) {
9277 		ASSERT(svntrp->tr_refcnt != 0);
9278 		if (svntrp->tr_vp == vp && svntrp->tr_off == off &&
9279 		    svntrp->tr_eoff == eoff && svntrp->tr_szc == szc) {
9280 			break;
9281 		}
9282 	}
9283 	if (svntrp == NULL) {
9284 		panic("segvn_textunrepl: svntr record not found");
9285 	}
9286 	if (svntrp->tr_amp[lgrp_id] != svd->amp) {
9287 		panic("segvn_textunrepl: amp mismatch");
9288 	}
9289 	svd->tr_state = SEGVN_TR_OFF;
9290 	svd->amp = NULL;
9291 	if (svd->svn_trprev == NULL) {
9292 		ASSERT(svntrp->tr_svnhead == svd);
9293 		svntrp->tr_svnhead = svd->svn_trnext;
9294 		if (svntrp->tr_svnhead != NULL) {
9295 			svntrp->tr_svnhead->svn_trprev = NULL;
9296 		}
9297 		svd->svn_trnext = NULL;
9298 	} else {
9299 		svd->svn_trprev->svn_trnext = svd->svn_trnext;
9300 		if (svd->svn_trnext != NULL) {
9301 			svd->svn_trnext->svn_trprev = svd->svn_trprev;
9302 			svd->svn_trnext = NULL;
9303 		}
9304 		svd->svn_trprev = NULL;
9305 	}
9306 	if (--svntrp->tr_refcnt) {
9307 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9308 		goto done;
9309 	}
9310 	*prv_svntrp = svntrp->tr_next;
9311 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9312 	for (i = 0; i < NLGRPS_MAX; i++) {
9313 		struct anon_map *amp = svntrp->tr_amp[i];
9314 		if (amp == NULL) {
9315 			continue;
9316 		}
9317 		ASSERT(amp->refcnt == 1);
9318 		ASSERT(amp->swresv == size);
9319 		ASSERT(amp->size == size);
9320 		ASSERT(amp->a_szc == szc);
9321 		if (amp->a_szc != 0) {
9322 			anon_free_pages(amp->ahp, 0, size, szc);
9323 		} else {
9324 			anon_free(amp->ahp, 0, size);
9325 		}
9326 		svntrp->tr_amp[i] = NULL;
9327 		ASSERT(segvn_textrepl_bytes >= size);
9328 		atomic_add_long(&segvn_textrepl_bytes, -size);
9329 		anon_unresv_zone(amp->swresv, NULL);
9330 		amp->refcnt = 0;
9331 		anonmap_free(amp);
9332 	}
9333 	kmem_cache_free(svntr_cache, svntrp);
9334 done:
9335 	hat_unload_callback(seg->s_as->a_hat, seg->s_base, size,
9336 	    unload_unmap ? HAT_UNLOAD_UNMAP : 0, NULL);
9337 }
9338 
9339 /*
9340  * This is called when a MAP_SHARED writable mapping is created to a vnode
9341  * that is currently used for execution (VVMEXEC flag is set). In this case we
9342  * need to prevent further use of existing replicas.
9343  */
9344 static void
9345 segvn_inval_trcache(vnode_t *vp)
9346 {
9347 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9348 	svntr_t			*svntrp;
9349 
9350 	ASSERT(vp->v_flag & VVMEXEC);
9351 
9352 	if (svntr_hashtab == NULL) {
9353 		return;
9354 	}
9355 
9356 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9357 	svntrp = svntr_hashtab[hash].tr_head;
9358 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9359 		ASSERT(svntrp->tr_refcnt != 0);
9360 		if (svntrp->tr_vp == vp && svntrp->tr_valid) {
9361 			svntrp->tr_valid = 0;
9362 		}
9363 	}
9364 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9365 }
9366 
9367 static void
9368 segvn_trasync_thread(void)
9369 {
9370 	callb_cpr_t cpr_info;
9371 	kmutex_t cpr_lock;	/* just for CPR stuff */
9372 
9373 	mutex_init(&cpr_lock, NULL, MUTEX_DEFAULT, NULL);
9374 
9375 	CALLB_CPR_INIT(&cpr_info, &cpr_lock,
9376 	    callb_generic_cpr, "segvn_async");
9377 
9378 	if (segvn_update_textrepl_interval == 0) {
9379 		segvn_update_textrepl_interval = segvn_update_tr_time * hz;
9380 	} else {
9381 		segvn_update_textrepl_interval *= hz;
9382 	}
9383 	(void) timeout(segvn_trupdate_wakeup, NULL,
9384 	    segvn_update_textrepl_interval);
9385 
9386 	for (;;) {
9387 		mutex_enter(&cpr_lock);
9388 		CALLB_CPR_SAFE_BEGIN(&cpr_info);
9389 		mutex_exit(&cpr_lock);
9390 		sema_p(&segvn_trasync_sem);
9391 		mutex_enter(&cpr_lock);
9392 		CALLB_CPR_SAFE_END(&cpr_info, &cpr_lock);
9393 		mutex_exit(&cpr_lock);
9394 		segvn_trupdate();
9395 	}
9396 }
9397 
9398 static uint64_t segvn_lgrp_trthr_migrs_snpsht = 0;
9399 
9400 static void
9401 segvn_trupdate_wakeup(void *dummy)
9402 {
9403 	uint64_t cur_lgrp_trthr_migrs = lgrp_get_trthr_migrations();
9404 
9405 	if (cur_lgrp_trthr_migrs != segvn_lgrp_trthr_migrs_snpsht) {
9406 		segvn_lgrp_trthr_migrs_snpsht = cur_lgrp_trthr_migrs;
9407 		sema_v(&segvn_trasync_sem);
9408 	}
9409 
9410 	if (!segvn_disable_textrepl_update &&
9411 	    segvn_update_textrepl_interval != 0) {
9412 		(void) timeout(segvn_trupdate_wakeup, dummy,
9413 		    segvn_update_textrepl_interval);
9414 	}
9415 }
9416 
9417 static void
9418 segvn_trupdate(void)
9419 {
9420 	ulong_t		hash;
9421 	svntr_t		*svntrp;
9422 	segvn_data_t	*svd;
9423 
9424 	ASSERT(svntr_hashtab != NULL);
9425 
9426 	for (hash = 0; hash < svntr_hashtab_sz; hash++) {
9427 		mutex_enter(&svntr_hashtab[hash].tr_lock);
9428 		svntrp = svntr_hashtab[hash].tr_head;
9429 		for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9430 			ASSERT(svntrp->tr_refcnt != 0);
9431 			svd = svntrp->tr_svnhead;
9432 			for (; svd != NULL; svd = svd->svn_trnext) {
9433 				segvn_trupdate_seg(svd->seg, svd, svntrp,
9434 				    hash);
9435 			}
9436 		}
9437 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9438 	}
9439 }
9440 
9441 static void
9442 segvn_trupdate_seg(struct seg *seg,
9443 	segvn_data_t *svd,
9444 	svntr_t *svntrp,
9445 	ulong_t hash)
9446 {
9447 	proc_t			*p;
9448 	lgrp_id_t		lgrp_id;
9449 	struct as		*as;
9450 	size_t			size;
9451 	struct anon_map		*amp;
9452 
9453 	ASSERT(svd->vp != NULL);
9454 	ASSERT(svd->vp == svntrp->tr_vp);
9455 	ASSERT(svd->offset == svntrp->tr_off);
9456 	ASSERT(svd->offset + seg->s_size == svntrp->tr_eoff);
9457 	ASSERT(seg != NULL);
9458 	ASSERT(svd->seg == seg);
9459 	ASSERT(seg->s_data == (void *)svd);
9460 	ASSERT(seg->s_szc == svntrp->tr_szc);
9461 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9462 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9463 	ASSERT(svd->amp != NULL);
9464 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
9465 	ASSERT(svd->tr_policy_info.mem_lgrpid != LGRP_NONE);
9466 	ASSERT(svd->tr_policy_info.mem_lgrpid < NLGRPS_MAX);
9467 	ASSERT(svntrp->tr_amp[svd->tr_policy_info.mem_lgrpid] == svd->amp);
9468 	ASSERT(svntrp->tr_refcnt != 0);
9469 	ASSERT(mutex_owned(&svntr_hashtab[hash].tr_lock));
9470 
9471 	as = seg->s_as;
9472 	ASSERT(as != NULL && as != &kas);
9473 	p = as->a_proc;
9474 	ASSERT(p != NULL);
9475 	ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9476 	lgrp_id = p->p_t1_lgrpid;
9477 	if (lgrp_id == LGRP_NONE) {
9478 		return;
9479 	}
9480 	ASSERT(lgrp_id < NLGRPS_MAX);
9481 	if (svd->tr_policy_info.mem_lgrpid == lgrp_id) {
9482 		return;
9483 	}
9484 
9485 	/*
9486 	 * Use tryenter locking since we are locking as/seg and svntr hash
9487 	 * lock in reverse from syncrounous thread order.
9488 	 */
9489 	if (!AS_LOCK_TRYENTER(as, &as->a_lock, RW_READER)) {
9490 		SEGVN_TR_ADDSTAT(nolock);
9491 		if (segvn_lgrp_trthr_migrs_snpsht) {
9492 			segvn_lgrp_trthr_migrs_snpsht = 0;
9493 		}
9494 		return;
9495 	}
9496 	if (!SEGVN_LOCK_TRYENTER(seg->s_as, &svd->lock, RW_WRITER)) {
9497 		AS_LOCK_EXIT(as, &as->a_lock);
9498 		SEGVN_TR_ADDSTAT(nolock);
9499 		if (segvn_lgrp_trthr_migrs_snpsht) {
9500 			segvn_lgrp_trthr_migrs_snpsht = 0;
9501 		}
9502 		return;
9503 	}
9504 	size = seg->s_size;
9505 	if (svntrp->tr_amp[lgrp_id] == NULL) {
9506 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9507 		if (trmem > segvn_textrepl_max_bytes) {
9508 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9509 			AS_LOCK_EXIT(as, &as->a_lock);
9510 			atomic_add_long(&segvn_textrepl_bytes, -size);
9511 			SEGVN_TR_ADDSTAT(normem);
9512 			return;
9513 		}
9514 		if (anon_try_resv_zone(size, NULL) == 0) {
9515 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9516 			AS_LOCK_EXIT(as, &as->a_lock);
9517 			atomic_add_long(&segvn_textrepl_bytes, -size);
9518 			SEGVN_TR_ADDSTAT(noanon);
9519 			return;
9520 		}
9521 		amp = anonmap_alloc(size, size, KM_NOSLEEP);
9522 		if (amp == NULL) {
9523 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9524 			AS_LOCK_EXIT(as, &as->a_lock);
9525 			atomic_add_long(&segvn_textrepl_bytes, -size);
9526 			anon_unresv_zone(size, NULL);
9527 			SEGVN_TR_ADDSTAT(nokmem);
9528 			return;
9529 		}
9530 		ASSERT(amp->refcnt == 1);
9531 		amp->a_szc = seg->s_szc;
9532 		svntrp->tr_amp[lgrp_id] = amp;
9533 	}
9534 	/*
9535 	 * We don't need to drop the bucket lock but here we give other
9536 	 * threads a chance.  svntr and svd can't be unlinked as long as
9537 	 * segment lock is held as a writer and AS held as well.  After we
9538 	 * retake bucket lock we'll continue from where we left. We'll be able
9539 	 * to reach the end of either list since new entries are always added
9540 	 * to the beginning of the lists.
9541 	 */
9542 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9543 	hat_unload_callback(as->a_hat, seg->s_base, size, 0, NULL);
9544 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9545 
9546 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9547 	ASSERT(svd->amp != NULL);
9548 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
9549 	ASSERT(svd->tr_policy_info.mem_lgrpid != lgrp_id);
9550 	ASSERT(svd->amp != svntrp->tr_amp[lgrp_id]);
9551 
9552 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
9553 	svd->amp = svntrp->tr_amp[lgrp_id];
9554 	p->p_tr_lgrpid = NLGRPS_MAX;
9555 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9556 	AS_LOCK_EXIT(as, &as->a_lock);
9557 
9558 	ASSERT(svntrp->tr_refcnt != 0);
9559 	ASSERT(svd->vp == svntrp->tr_vp);
9560 	ASSERT(svd->tr_policy_info.mem_lgrpid == lgrp_id);
9561 	ASSERT(svd->amp != NULL && svd->amp == svntrp->tr_amp[lgrp_id]);
9562 	ASSERT(svd->seg == seg);
9563 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9564 
9565 	SEGVN_TR_ADDSTAT(asyncrepl);
9566 }
9567