xref: /illumos-gate/usr/src/uts/common/vm/seg_vn.c (revision 9ec394dbf343c1f23c6e13c39df427f238e5a369)
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 2008 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 size_t	segvn_pglock_comb_thrshld = (1UL << 16);	/* 64K */
166 size_t	segvn_pglock_comb_balign = (1UL << 16);		/* 64K */
167 uint_t	segvn_pglock_comb_bshift;
168 size_t	segvn_pglock_comb_palign;
169 
170 static int	segvn_concat(struct seg *, struct seg *, int);
171 static int	segvn_extend_prev(struct seg *, struct seg *,
172 		    struct segvn_crargs *, size_t);
173 static int	segvn_extend_next(struct seg *, struct seg *,
174 		    struct segvn_crargs *, size_t);
175 static void	segvn_softunlock(struct seg *, caddr_t, size_t, enum seg_rw);
176 static void	segvn_pagelist_rele(page_t **);
177 static void	segvn_setvnode_mpss(vnode_t *);
178 static void	segvn_relocate_pages(page_t **, page_t *);
179 static int	segvn_full_szcpages(page_t **, uint_t, int *, uint_t *);
180 static int	segvn_fill_vp_pages(struct segvn_data *, vnode_t *, u_offset_t,
181     uint_t, page_t **, page_t **, uint_t *, int *);
182 static faultcode_t segvn_fault_vnodepages(struct hat *, struct seg *, caddr_t,
183     caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
184 static faultcode_t segvn_fault_anonpages(struct hat *, struct seg *, caddr_t,
185     caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
186 static faultcode_t segvn_faultpage(struct hat *, struct seg *, caddr_t,
187     u_offset_t, struct vpage *, page_t **, uint_t,
188     enum fault_type, enum seg_rw, int);
189 static void	segvn_vpage(struct seg *);
190 static size_t	segvn_count_swap_by_vpages(struct seg *);
191 
192 static void segvn_purge(struct seg *seg);
193 static int segvn_reclaim(void *, caddr_t, size_t, struct page **,
194     enum seg_rw, int);
195 static int shamp_reclaim(void *, caddr_t, size_t, struct page **,
196     enum seg_rw, int);
197 
198 static int sameprot(struct seg *, caddr_t, size_t);
199 
200 static int segvn_demote_range(struct seg *, caddr_t, size_t, int, uint_t);
201 static int segvn_clrszc(struct seg *);
202 static struct seg *segvn_split_seg(struct seg *, caddr_t);
203 static int segvn_claim_pages(struct seg *, struct vpage *, u_offset_t,
204     ulong_t, uint_t);
205 
206 static void segvn_hat_rgn_unload_callback(caddr_t, caddr_t, caddr_t,
207     size_t, void *, u_offset_t);
208 
209 static struct kmem_cache *segvn_cache;
210 static struct kmem_cache **segvn_szc_cache;
211 
212 #ifdef VM_STATS
213 static struct segvnvmstats_str {
214 	ulong_t	fill_vp_pages[31];
215 	ulong_t fltvnpages[49];
216 	ulong_t	fullszcpages[10];
217 	ulong_t	relocatepages[3];
218 	ulong_t	fltanpages[17];
219 	ulong_t pagelock[2];
220 	ulong_t	demoterange[3];
221 } segvnvmstats;
222 #endif /* VM_STATS */
223 
224 #define	SDR_RANGE	1		/* demote entire range */
225 #define	SDR_END		2		/* demote non aligned ends only */
226 
227 #define	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr) {	    \
228 		if ((len) != 0) { 		      	      		      \
229 			lpgaddr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz);  \
230 			ASSERT(lpgaddr >= (seg)->s_base);	      	      \
231 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)((addr) +    \
232 			    (len)), pgsz);				      \
233 			ASSERT(lpgeaddr > lpgaddr);		      	      \
234 			ASSERT(lpgeaddr <= (seg)->s_base + (seg)->s_size);    \
235 		} else {					      	      \
236 			lpgeaddr = lpgaddr = (addr);	      		      \
237 		}							      \
238 	}
239 
240 /*ARGSUSED*/
241 static int
242 segvn_cache_constructor(void *buf, void *cdrarg, int kmflags)
243 {
244 	struct segvn_data *svd = buf;
245 
246 	rw_init(&svd->lock, NULL, RW_DEFAULT, NULL);
247 	mutex_init(&svd->segfree_syncmtx, NULL, MUTEX_DEFAULT, NULL);
248 	svd->svn_trnext = svd->svn_trprev = NULL;
249 	return (0);
250 }
251 
252 /*ARGSUSED1*/
253 static void
254 segvn_cache_destructor(void *buf, void *cdrarg)
255 {
256 	struct segvn_data *svd = buf;
257 
258 	rw_destroy(&svd->lock);
259 	mutex_destroy(&svd->segfree_syncmtx);
260 }
261 
262 /*ARGSUSED*/
263 static int
264 svntr_cache_constructor(void *buf, void *cdrarg, int kmflags)
265 {
266 	bzero(buf, sizeof (svntr_t));
267 	return (0);
268 }
269 
270 /*
271  * Patching this variable to non-zero allows the system to run with
272  * stacks marked as "not executable".  It's a bit of a kludge, but is
273  * provided as a tweakable for platforms that export those ABIs
274  * (e.g. sparc V8) that have executable stacks enabled by default.
275  * There are also some restrictions for platforms that don't actually
276  * implement 'noexec' protections.
277  *
278  * Once enabled, the system is (therefore) unable to provide a fully
279  * ABI-compliant execution environment, though practically speaking,
280  * most everything works.  The exceptions are generally some interpreters
281  * and debuggers that create executable code on the stack and jump
282  * into it (without explicitly mprotecting the address range to include
283  * PROT_EXEC).
284  *
285  * One important class of applications that are disabled are those
286  * that have been transformed into malicious agents using one of the
287  * numerous "buffer overflow" attacks.  See 4007890.
288  */
289 int noexec_user_stack = 0;
290 int noexec_user_stack_log = 1;
291 
292 int segvn_lpg_disable = 0;
293 uint_t segvn_maxpgszc = 0;
294 
295 ulong_t segvn_vmpss_clrszc_cnt;
296 ulong_t segvn_vmpss_clrszc_err;
297 ulong_t segvn_fltvnpages_clrszc_cnt;
298 ulong_t segvn_fltvnpages_clrszc_err;
299 ulong_t segvn_setpgsz_align_err;
300 ulong_t segvn_setpgsz_anon_align_err;
301 ulong_t segvn_setpgsz_getattr_err;
302 ulong_t segvn_setpgsz_eof_err;
303 ulong_t segvn_faultvnmpss_align_err1;
304 ulong_t segvn_faultvnmpss_align_err2;
305 ulong_t segvn_faultvnmpss_align_err3;
306 ulong_t segvn_faultvnmpss_align_err4;
307 ulong_t segvn_faultvnmpss_align_err5;
308 ulong_t	segvn_vmpss_pageio_deadlk_err;
309 
310 int segvn_use_regions = 1;
311 
312 /*
313  * Segvn supports text replication optimization for NUMA platforms. Text
314  * replica's are represented by anon maps (amp). There's one amp per text file
315  * region per lgroup. A process chooses the amp for each of its text mappings
316  * based on the lgroup assignment of its main thread (t_tid = 1). All
317  * processes that want a replica on a particular lgroup for the same text file
318  * mapping share the same amp. amp's are looked up in svntr_hashtab hash table
319  * with vp,off,size,szc used as a key. Text replication segments are read only
320  * MAP_PRIVATE|MAP_TEXT segments that map vnode. Replication is achieved by
321  * forcing COW faults from vnode to amp and mapping amp pages instead of vnode
322  * pages. Replication amp is assigned to a segment when it gets its first
323  * pagefault. To handle main thread lgroup rehoming segvn_trasync_thread
324  * rechecks periodically if the process still maps an amp local to the main
325  * thread. If not async thread forces process to remap to an amp in the new
326  * home lgroup of the main thread. Current text replication implementation
327  * only provides the benefit to workloads that do most of their work in the
328  * main thread of a process or all the threads of a process run in the same
329  * lgroup. To extend text replication benefit to different types of
330  * multithreaded workloads further work would be needed in the hat layer to
331  * allow the same virtual address in the same hat to simultaneously map
332  * different physical addresses (i.e. page table replication would be needed
333  * for x86).
334  *
335  * amp pages are used instead of vnode pages as long as segment has a very
336  * simple life cycle.  It's created via segvn_create(), handles S_EXEC
337  * (S_READ) pagefaults and is fully unmapped.  If anything more complicated
338  * happens such as protection is changed, real COW fault happens, pagesize is
339  * changed, MC_LOCK is requested or segment is partially unmapped we turn off
340  * text replication by converting the segment back to vnode only segment
341  * (unmap segment's address range and set svd->amp to NULL).
342  *
343  * The original file can be changed after amp is inserted into
344  * svntr_hashtab. Processes that are launched after the file is already
345  * changed can't use the replica's created prior to the file change. To
346  * implement this functionality hash entries are timestamped. Replica's can
347  * only be used if current file modification time is the same as the timestamp
348  * saved when hash entry was created. However just timestamps alone are not
349  * sufficient to detect file modification via mmap(MAP_SHARED) mappings. We
350  * deal with file changes via MAP_SHARED mappings differently. When writable
351  * MAP_SHARED mappings are created to vnodes marked as executable we mark all
352  * existing replica's for this vnode as not usable for future text
353  * mappings. And we don't create new replica's for files that currently have
354  * potentially writable MAP_SHARED mappings (i.e. vn_is_mapped(V_WRITE) is
355  * true).
356  */
357 
358 #define	SEGVN_TEXTREPL_MAXBYTES_FACTOR	(20)
359 size_t	segvn_textrepl_max_bytes_factor = SEGVN_TEXTREPL_MAXBYTES_FACTOR;
360 
361 static ulong_t			svntr_hashtab_sz = 512;
362 static svntr_bucket_t		*svntr_hashtab = NULL;
363 static struct kmem_cache	*svntr_cache;
364 static svntr_stats_t		*segvn_textrepl_stats;
365 static ksema_t 			segvn_trasync_sem;
366 
367 int				segvn_disable_textrepl = 1;
368 size_t				textrepl_size_thresh = (size_t)-1;
369 size_t				segvn_textrepl_bytes = 0;
370 size_t				segvn_textrepl_max_bytes = 0;
371 clock_t				segvn_update_textrepl_interval = 0;
372 int				segvn_update_tr_time = 10;
373 int				segvn_disable_textrepl_update = 0;
374 
375 static void segvn_textrepl(struct seg *);
376 static void segvn_textunrepl(struct seg *, int);
377 static void segvn_inval_trcache(vnode_t *);
378 static void segvn_trasync_thread(void);
379 static void segvn_trupdate_wakeup(void *);
380 static void segvn_trupdate(void);
381 static void segvn_trupdate_seg(struct seg *, segvn_data_t *, svntr_t *,
382     ulong_t);
383 
384 /*
385  * Initialize segvn data structures
386  */
387 void
388 segvn_init(void)
389 {
390 	uint_t maxszc;
391 	uint_t szc;
392 	size_t pgsz;
393 
394 	segvn_cache = kmem_cache_create("segvn_cache",
395 	    sizeof (struct segvn_data), 0,
396 	    segvn_cache_constructor, segvn_cache_destructor, NULL,
397 	    NULL, NULL, 0);
398 
399 	if (segvn_lpg_disable == 0) {
400 		szc = maxszc = page_num_pagesizes() - 1;
401 		if (szc == 0) {
402 			segvn_lpg_disable = 1;
403 		}
404 		if (page_get_pagesize(0) != PAGESIZE) {
405 			panic("segvn_init: bad szc 0");
406 			/*NOTREACHED*/
407 		}
408 		while (szc != 0) {
409 			pgsz = page_get_pagesize(szc);
410 			if (pgsz <= PAGESIZE || !IS_P2ALIGNED(pgsz, pgsz)) {
411 				panic("segvn_init: bad szc %d", szc);
412 				/*NOTREACHED*/
413 			}
414 			szc--;
415 		}
416 		if (segvn_maxpgszc == 0 || segvn_maxpgszc > maxszc)
417 			segvn_maxpgszc = maxszc;
418 	}
419 
420 	if (segvn_maxpgszc) {
421 		segvn_szc_cache = (struct kmem_cache **)kmem_alloc(
422 		    (segvn_maxpgszc + 1) * sizeof (struct kmem_cache *),
423 		    KM_SLEEP);
424 	}
425 
426 	for (szc = 1; szc <= segvn_maxpgszc; szc++) {
427 		char	str[32];
428 
429 		(void) sprintf(str, "segvn_szc_cache%d", szc);
430 		segvn_szc_cache[szc] = kmem_cache_create(str,
431 		    page_get_pagecnt(szc) * sizeof (page_t *), 0,
432 		    NULL, NULL, NULL, NULL, NULL, KMC_NODEBUG);
433 	}
434 
435 
436 	if (segvn_use_regions && !hat_supported(HAT_SHARED_REGIONS, NULL))
437 		segvn_use_regions = 0;
438 
439 	/*
440 	 * For now shared regions and text replication segvn support
441 	 * are mutually exclusive. This is acceptable because
442 	 * currently significant benefit from text replication was
443 	 * only observed on AMD64 NUMA platforms (due to relatively
444 	 * small L2$ size) and currently we don't support shared
445 	 * regions on x86.
446 	 */
447 	if (segvn_use_regions && !segvn_disable_textrepl) {
448 		segvn_disable_textrepl = 1;
449 	}
450 
451 #if defined(_LP64)
452 	if (lgrp_optimizations() && textrepl_size_thresh != (size_t)-1 &&
453 	    !segvn_disable_textrepl) {
454 		ulong_t i;
455 		size_t hsz = svntr_hashtab_sz * sizeof (svntr_bucket_t);
456 
457 		svntr_cache = kmem_cache_create("svntr_cache",
458 		    sizeof (svntr_t), 0, svntr_cache_constructor, NULL,
459 		    NULL, NULL, NULL, 0);
460 		svntr_hashtab = kmem_zalloc(hsz, KM_SLEEP);
461 		for (i = 0; i < svntr_hashtab_sz; i++) {
462 			mutex_init(&svntr_hashtab[i].tr_lock,  NULL,
463 			    MUTEX_DEFAULT, NULL);
464 		}
465 		segvn_textrepl_max_bytes = ptob(physmem) /
466 		    segvn_textrepl_max_bytes_factor;
467 		segvn_textrepl_stats = kmem_zalloc(NCPU *
468 		    sizeof (svntr_stats_t), KM_SLEEP);
469 		sema_init(&segvn_trasync_sem, 0, NULL, SEMA_DEFAULT, NULL);
470 		(void) thread_create(NULL, 0, segvn_trasync_thread,
471 		    NULL, 0, &p0, TS_RUN, minclsyspri);
472 	}
473 #endif
474 
475 	if (!ISP2(segvn_pglock_comb_balign) ||
476 	    segvn_pglock_comb_balign < PAGESIZE) {
477 		segvn_pglock_comb_balign = 1UL << 16; /* 64K */
478 	}
479 	segvn_pglock_comb_bshift = highbit(segvn_pglock_comb_balign) - 1;
480 	segvn_pglock_comb_palign = btop(segvn_pglock_comb_balign);
481 }
482 
483 #define	SEGVN_PAGEIO	((void *)0x1)
484 #define	SEGVN_NOPAGEIO	((void *)0x2)
485 
486 static void
487 segvn_setvnode_mpss(vnode_t *vp)
488 {
489 	int err;
490 
491 	ASSERT(vp->v_mpssdata == NULL ||
492 	    vp->v_mpssdata == SEGVN_PAGEIO ||
493 	    vp->v_mpssdata == SEGVN_NOPAGEIO);
494 
495 	if (vp->v_mpssdata == NULL) {
496 		if (vn_vmpss_usepageio(vp)) {
497 			err = VOP_PAGEIO(vp, (page_t *)NULL,
498 			    (u_offset_t)0, 0, 0, CRED(), NULL);
499 		} else {
500 			err = ENOSYS;
501 		}
502 		/*
503 		 * set v_mpssdata just once per vnode life
504 		 * so that it never changes.
505 		 */
506 		mutex_enter(&vp->v_lock);
507 		if (vp->v_mpssdata == NULL) {
508 			if (err == EINVAL) {
509 				vp->v_mpssdata = SEGVN_PAGEIO;
510 			} else {
511 				vp->v_mpssdata = SEGVN_NOPAGEIO;
512 			}
513 		}
514 		mutex_exit(&vp->v_lock);
515 	}
516 }
517 
518 int
519 segvn_create(struct seg *seg, void *argsp)
520 {
521 	struct segvn_crargs *a = (struct segvn_crargs *)argsp;
522 	struct segvn_data *svd;
523 	size_t swresv = 0;
524 	struct cred *cred;
525 	struct anon_map *amp;
526 	int error = 0;
527 	size_t pgsz;
528 	lgrp_mem_policy_t mpolicy = LGRP_MEM_POLICY_DEFAULT;
529 	int use_rgn = 0;
530 	int trok = 0;
531 
532 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
533 
534 	if (a->type != MAP_PRIVATE && a->type != MAP_SHARED) {
535 		panic("segvn_create type");
536 		/*NOTREACHED*/
537 	}
538 
539 	/*
540 	 * Check arguments.  If a shared anon structure is given then
541 	 * it is illegal to also specify a vp.
542 	 */
543 	if (a->amp != NULL && a->vp != NULL) {
544 		panic("segvn_create anon_map");
545 		/*NOTREACHED*/
546 	}
547 
548 	if (a->type == MAP_PRIVATE && (a->flags & MAP_TEXT) &&
549 	    a->vp != NULL && a->prot == (PROT_USER | PROT_READ | PROT_EXEC) &&
550 	    segvn_use_regions) {
551 		use_rgn = 1;
552 	}
553 
554 	/* MAP_NORESERVE on a MAP_SHARED segment is meaningless. */
555 	if (a->type == MAP_SHARED)
556 		a->flags &= ~MAP_NORESERVE;
557 
558 	if (a->szc != 0) {
559 		if (segvn_lpg_disable != 0 || (a->szc == AS_MAP_NO_LPOOB) ||
560 		    (a->amp != NULL && a->type == MAP_PRIVATE) ||
561 		    (a->flags & MAP_NORESERVE) || seg->s_as == &kas) {
562 			a->szc = 0;
563 		} else {
564 			if (a->szc > segvn_maxpgszc)
565 				a->szc = segvn_maxpgszc;
566 			pgsz = page_get_pagesize(a->szc);
567 			if (!IS_P2ALIGNED(seg->s_base, pgsz) ||
568 			    !IS_P2ALIGNED(seg->s_size, pgsz)) {
569 				a->szc = 0;
570 			} else if (a->vp != NULL) {
571 				extern struct vnode kvp;
572 				if (IS_SWAPFSVP(a->vp) || VN_ISKAS(a->vp)) {
573 					/*
574 					 * paranoid check.
575 					 * hat_page_demote() is not supported
576 					 * on swapfs pages.
577 					 */
578 					a->szc = 0;
579 				} else if (map_addr_vacalign_check(seg->s_base,
580 				    a->offset & PAGEMASK)) {
581 					a->szc = 0;
582 				}
583 			} else if (a->amp != NULL) {
584 				pgcnt_t anum = btopr(a->offset);
585 				pgcnt_t pgcnt = page_get_pagecnt(a->szc);
586 				if (!IS_P2ALIGNED(anum, pgcnt)) {
587 					a->szc = 0;
588 				}
589 			}
590 		}
591 	}
592 
593 	/*
594 	 * If segment may need private pages, reserve them now.
595 	 */
596 	if (!(a->flags & MAP_NORESERVE) && ((a->vp == NULL && a->amp == NULL) ||
597 	    (a->type == MAP_PRIVATE && (a->prot & PROT_WRITE)))) {
598 		if (anon_resv(seg->s_size) == 0)
599 			return (EAGAIN);
600 		swresv = seg->s_size;
601 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
602 		    seg, swresv, 1);
603 	}
604 
605 	/*
606 	 * Reserve any mapping structures that may be required.
607 	 *
608 	 * Don't do it for segments that may use regions. It's currently a
609 	 * noop in the hat implementations anyway.
610 	 */
611 	if (!use_rgn) {
612 		hat_map(seg->s_as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
613 	}
614 
615 	if (a->cred) {
616 		cred = a->cred;
617 		crhold(cred);
618 	} else {
619 		crhold(cred = CRED());
620 	}
621 
622 	/* Inform the vnode of the new mapping */
623 	if (a->vp != NULL) {
624 		error = VOP_ADDMAP(a->vp, a->offset & PAGEMASK,
625 		    seg->s_as, seg->s_base, seg->s_size, a->prot,
626 		    a->maxprot, a->type, cred, NULL);
627 		if (error) {
628 			if (swresv != 0) {
629 				anon_unresv(swresv);
630 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
631 				    "anon proc:%p %lu %u", seg, swresv, 0);
632 			}
633 			crfree(cred);
634 			if (!use_rgn) {
635 				hat_unload(seg->s_as->a_hat, seg->s_base,
636 				    seg->s_size, HAT_UNLOAD_UNMAP);
637 			}
638 			return (error);
639 		}
640 		/*
641 		 * svntr_hashtab will be NULL if we support shared regions.
642 		 */
643 		trok = ((a->flags & MAP_TEXT) &&
644 		    (seg->s_size > textrepl_size_thresh ||
645 		    (a->flags & _MAP_TEXTREPL)) &&
646 		    lgrp_optimizations() && svntr_hashtab != NULL &&
647 		    a->type == MAP_PRIVATE && swresv == 0 &&
648 		    !(a->flags & MAP_NORESERVE) &&
649 		    seg->s_as != &kas && a->vp->v_type == VREG);
650 
651 		ASSERT(!trok || !use_rgn);
652 	}
653 
654 	/*
655 	 * If more than one segment in the address space, and they're adjacent
656 	 * virtually, try to concatenate them.  Don't concatenate if an
657 	 * explicit anon_map structure was supplied (e.g., SystemV shared
658 	 * memory) or if we'll use text replication for this segment.
659 	 */
660 	if (a->amp == NULL && !use_rgn && !trok) {
661 		struct seg *pseg, *nseg;
662 		struct segvn_data *psvd, *nsvd;
663 		lgrp_mem_policy_t ppolicy, npolicy;
664 		uint_t	lgrp_mem_policy_flags = 0;
665 		extern lgrp_mem_policy_t lgrp_mem_default_policy;
666 
667 		/*
668 		 * Memory policy flags (lgrp_mem_policy_flags) is valid when
669 		 * extending stack/heap segments.
670 		 */
671 		if ((a->vp == NULL) && (a->type == MAP_PRIVATE) &&
672 		    !(a->flags & MAP_NORESERVE) && (seg->s_as != &kas)) {
673 			lgrp_mem_policy_flags = a->lgrp_mem_policy_flags;
674 		} else {
675 			/*
676 			 * Get policy when not extending it from another segment
677 			 */
678 			mpolicy = lgrp_mem_policy_default(seg->s_size, a->type);
679 		}
680 
681 		/*
682 		 * First, try to concatenate the previous and new segments
683 		 */
684 		pseg = AS_SEGPREV(seg->s_as, seg);
685 		if (pseg != NULL &&
686 		    pseg->s_base + pseg->s_size == seg->s_base &&
687 		    pseg->s_ops == &segvn_ops) {
688 			/*
689 			 * Get memory allocation policy from previous segment.
690 			 * When extension is specified (e.g. for heap) apply
691 			 * this policy to the new segment regardless of the
692 			 * outcome of segment concatenation.  Extension occurs
693 			 * for non-default policy otherwise default policy is
694 			 * used and is based on extended segment size.
695 			 */
696 			psvd = (struct segvn_data *)pseg->s_data;
697 			ppolicy = psvd->policy_info.mem_policy;
698 			if (lgrp_mem_policy_flags ==
699 			    LGRP_MP_FLAG_EXTEND_UP) {
700 				if (ppolicy != lgrp_mem_default_policy) {
701 					mpolicy = ppolicy;
702 				} else {
703 					mpolicy = lgrp_mem_policy_default(
704 					    pseg->s_size + seg->s_size,
705 					    a->type);
706 				}
707 			}
708 
709 			if (mpolicy == ppolicy &&
710 			    (pseg->s_size + seg->s_size <=
711 			    segvn_comb_thrshld || psvd->amp == NULL) &&
712 			    segvn_extend_prev(pseg, seg, a, swresv) == 0) {
713 				/*
714 				 * success! now try to concatenate
715 				 * with following seg
716 				 */
717 				crfree(cred);
718 				nseg = AS_SEGNEXT(pseg->s_as, pseg);
719 				if (nseg != NULL &&
720 				    nseg != pseg &&
721 				    nseg->s_ops == &segvn_ops &&
722 				    pseg->s_base + pseg->s_size ==
723 				    nseg->s_base)
724 					(void) segvn_concat(pseg, nseg, 0);
725 				ASSERT(pseg->s_szc == 0 ||
726 				    (a->szc == pseg->s_szc &&
727 				    IS_P2ALIGNED(pseg->s_base, pgsz) &&
728 				    IS_P2ALIGNED(pseg->s_size, pgsz)));
729 				return (0);
730 			}
731 		}
732 
733 		/*
734 		 * Failed, so try to concatenate with following seg
735 		 */
736 		nseg = AS_SEGNEXT(seg->s_as, seg);
737 		if (nseg != NULL &&
738 		    seg->s_base + seg->s_size == nseg->s_base &&
739 		    nseg->s_ops == &segvn_ops) {
740 			/*
741 			 * Get memory allocation policy from next segment.
742 			 * When extension is specified (e.g. for stack) apply
743 			 * this policy to the new segment regardless of the
744 			 * outcome of segment concatenation.  Extension occurs
745 			 * for non-default policy otherwise default policy is
746 			 * used and is based on extended segment size.
747 			 */
748 			nsvd = (struct segvn_data *)nseg->s_data;
749 			npolicy = nsvd->policy_info.mem_policy;
750 			if (lgrp_mem_policy_flags ==
751 			    LGRP_MP_FLAG_EXTEND_DOWN) {
752 				if (npolicy != lgrp_mem_default_policy) {
753 					mpolicy = npolicy;
754 				} else {
755 					mpolicy = lgrp_mem_policy_default(
756 					    nseg->s_size + seg->s_size,
757 					    a->type);
758 				}
759 			}
760 
761 			if (mpolicy == npolicy &&
762 			    segvn_extend_next(seg, nseg, a, swresv) == 0) {
763 				crfree(cred);
764 				ASSERT(nseg->s_szc == 0 ||
765 				    (a->szc == nseg->s_szc &&
766 				    IS_P2ALIGNED(nseg->s_base, pgsz) &&
767 				    IS_P2ALIGNED(nseg->s_size, pgsz)));
768 				return (0);
769 			}
770 		}
771 	}
772 
773 	if (a->vp != NULL) {
774 		VN_HOLD(a->vp);
775 		if (a->type == MAP_SHARED)
776 			lgrp_shm_policy_init(NULL, a->vp);
777 	}
778 	svd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
779 
780 	seg->s_ops = &segvn_ops;
781 	seg->s_data = (void *)svd;
782 	seg->s_szc = a->szc;
783 
784 	svd->seg = seg;
785 	svd->vp = a->vp;
786 	/*
787 	 * Anonymous mappings have no backing file so the offset is meaningless.
788 	 */
789 	svd->offset = a->vp ? (a->offset & PAGEMASK) : 0;
790 	svd->prot = a->prot;
791 	svd->maxprot = a->maxprot;
792 	svd->pageprot = 0;
793 	svd->type = a->type;
794 	svd->vpage = NULL;
795 	svd->cred = cred;
796 	svd->advice = MADV_NORMAL;
797 	svd->pageadvice = 0;
798 	svd->flags = (ushort_t)a->flags;
799 	svd->softlockcnt = 0;
800 	svd->softlockcnt_sbase = 0;
801 	svd->softlockcnt_send = 0;
802 	svd->rcookie = HAT_INVALID_REGION_COOKIE;
803 	svd->pageswap = 0;
804 
805 	if (a->szc != 0 && a->vp != NULL) {
806 		segvn_setvnode_mpss(a->vp);
807 	}
808 	if (svd->type == MAP_SHARED && svd->vp != NULL &&
809 	    (svd->vp->v_flag & VVMEXEC) && (svd->prot & PROT_WRITE)) {
810 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
811 		segvn_inval_trcache(svd->vp);
812 	}
813 
814 	amp = a->amp;
815 	if ((svd->amp = amp) == NULL) {
816 		svd->anon_index = 0;
817 		if (svd->type == MAP_SHARED) {
818 			svd->swresv = 0;
819 			/*
820 			 * Shared mappings to a vp need no other setup.
821 			 * If we have a shared mapping to an anon_map object
822 			 * which hasn't been allocated yet,  allocate the
823 			 * struct now so that it will be properly shared
824 			 * by remembering the swap reservation there.
825 			 */
826 			if (a->vp == NULL) {
827 				svd->amp = anonmap_alloc(seg->s_size, swresv,
828 				    ANON_SLEEP);
829 				svd->amp->a_szc = seg->s_szc;
830 			}
831 		} else {
832 			/*
833 			 * Private mapping (with or without a vp).
834 			 * Allocate anon_map when needed.
835 			 */
836 			svd->swresv = swresv;
837 		}
838 	} else {
839 		pgcnt_t anon_num;
840 
841 		/*
842 		 * Mapping to an existing anon_map structure without a vp.
843 		 * For now we will insure that the segment size isn't larger
844 		 * than the size - offset gives us.  Later on we may wish to
845 		 * have the anon array dynamically allocated itself so that
846 		 * we don't always have to allocate all the anon pointer slots.
847 		 * This of course involves adding extra code to check that we
848 		 * aren't trying to use an anon pointer slot beyond the end
849 		 * of the currently allocated anon array.
850 		 */
851 		if ((amp->size - a->offset) < seg->s_size) {
852 			panic("segvn_create anon_map size");
853 			/*NOTREACHED*/
854 		}
855 
856 		anon_num = btopr(a->offset);
857 
858 		if (a->type == MAP_SHARED) {
859 			/*
860 			 * SHARED mapping to a given anon_map.
861 			 */
862 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
863 			amp->refcnt++;
864 			if (a->szc > amp->a_szc) {
865 				amp->a_szc = a->szc;
866 			}
867 			ANON_LOCK_EXIT(&amp->a_rwlock);
868 			svd->anon_index = anon_num;
869 			svd->swresv = 0;
870 		} else {
871 			/*
872 			 * PRIVATE mapping to a given anon_map.
873 			 * Make sure that all the needed anon
874 			 * structures are created (so that we will
875 			 * share the underlying pages if nothing
876 			 * is written by this mapping) and then
877 			 * duplicate the anon array as is done
878 			 * when a privately mapped segment is dup'ed.
879 			 */
880 			struct anon *ap;
881 			caddr_t addr;
882 			caddr_t eaddr;
883 			ulong_t	anon_idx;
884 			int hat_flag = HAT_LOAD;
885 
886 			if (svd->flags & MAP_TEXT) {
887 				hat_flag |= HAT_LOAD_TEXT;
888 			}
889 
890 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
891 			svd->amp->a_szc = seg->s_szc;
892 			svd->anon_index = 0;
893 			svd->swresv = swresv;
894 
895 			/*
896 			 * Prevent 2 threads from allocating anon
897 			 * slots simultaneously.
898 			 */
899 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
900 			eaddr = seg->s_base + seg->s_size;
901 
902 			for (anon_idx = anon_num, addr = seg->s_base;
903 			    addr < eaddr; addr += PAGESIZE, anon_idx++) {
904 				page_t *pp;
905 
906 				if ((ap = anon_get_ptr(amp->ahp,
907 				    anon_idx)) != NULL)
908 					continue;
909 
910 				/*
911 				 * Allocate the anon struct now.
912 				 * Might as well load up translation
913 				 * to the page while we're at it...
914 				 */
915 				pp = anon_zero(seg, addr, &ap, cred);
916 				if (ap == NULL || pp == NULL) {
917 					panic("segvn_create anon_zero");
918 					/*NOTREACHED*/
919 				}
920 
921 				/*
922 				 * Re-acquire the anon_map lock and
923 				 * initialize the anon array entry.
924 				 */
925 				ASSERT(anon_get_ptr(amp->ahp,
926 				    anon_idx) == NULL);
927 				(void) anon_set_ptr(amp->ahp, anon_idx, ap,
928 				    ANON_SLEEP);
929 
930 				ASSERT(seg->s_szc == 0);
931 				ASSERT(!IS_VMODSORT(pp->p_vnode));
932 
933 				ASSERT(use_rgn == 0);
934 				hat_memload(seg->s_as->a_hat, addr, pp,
935 				    svd->prot & ~PROT_WRITE, hat_flag);
936 
937 				page_unlock(pp);
938 			}
939 			ASSERT(seg->s_szc == 0);
940 			anon_dup(amp->ahp, anon_num, svd->amp->ahp,
941 			    0, seg->s_size);
942 			ANON_LOCK_EXIT(&amp->a_rwlock);
943 		}
944 	}
945 
946 	/*
947 	 * Set default memory allocation policy for segment
948 	 *
949 	 * Always set policy for private memory at least for initialization
950 	 * even if this is a shared memory segment
951 	 */
952 	(void) lgrp_privm_policy_set(mpolicy, &svd->policy_info, seg->s_size);
953 
954 	if (svd->type == MAP_SHARED)
955 		(void) lgrp_shm_policy_set(mpolicy, svd->amp, svd->anon_index,
956 		    svd->vp, svd->offset, seg->s_size);
957 
958 	if (use_rgn) {
959 		ASSERT(!trok);
960 		ASSERT(svd->amp == NULL);
961 		svd->rcookie = hat_join_region(seg->s_as->a_hat, seg->s_base,
962 		    seg->s_size, (void *)svd->vp, svd->offset, svd->prot,
963 		    (uchar_t)seg->s_szc, segvn_hat_rgn_unload_callback,
964 		    HAT_REGION_TEXT);
965 	}
966 
967 	ASSERT(!trok || !(svd->prot & PROT_WRITE));
968 	svd->tr_state = trok ? SEGVN_TR_INIT : SEGVN_TR_OFF;
969 
970 	return (0);
971 }
972 
973 /*
974  * Concatenate two existing segments, if possible.
975  * Return 0 on success, -1 if two segments are not compatible
976  * or -2 on memory allocation failure.
977  * If amp_cat == 1 then try and concat segments with anon maps
978  */
979 static int
980 segvn_concat(struct seg *seg1, struct seg *seg2, int amp_cat)
981 {
982 	struct segvn_data *svd1 = seg1->s_data;
983 	struct segvn_data *svd2 = seg2->s_data;
984 	struct anon_map *amp1 = svd1->amp;
985 	struct anon_map *amp2 = svd2->amp;
986 	struct vpage *vpage1 = svd1->vpage;
987 	struct vpage *vpage2 = svd2->vpage, *nvpage = NULL;
988 	size_t size, nvpsize;
989 	pgcnt_t npages1, npages2;
990 
991 	ASSERT(seg1->s_as && seg2->s_as && seg1->s_as == seg2->s_as);
992 	ASSERT(AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
993 	ASSERT(seg1->s_ops == seg2->s_ops);
994 
995 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie) ||
996 	    HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
997 		return (-1);
998 	}
999 
1000 	/* both segments exist, try to merge them */
1001 #define	incompat(x)	(svd1->x != svd2->x)
1002 	if (incompat(vp) || incompat(maxprot) ||
1003 	    (!svd1->pageadvice && !svd2->pageadvice && incompat(advice)) ||
1004 	    (!svd1->pageprot && !svd2->pageprot && incompat(prot)) ||
1005 	    incompat(type) || incompat(cred) || incompat(flags) ||
1006 	    seg1->s_szc != seg2->s_szc || incompat(policy_info.mem_policy) ||
1007 	    (svd2->softlockcnt > 0) || svd1->softlockcnt_send > 0)
1008 		return (-1);
1009 #undef incompat
1010 
1011 	/*
1012 	 * vp == NULL implies zfod, offset doesn't matter
1013 	 */
1014 	if (svd1->vp != NULL &&
1015 	    svd1->offset + seg1->s_size != svd2->offset) {
1016 		return (-1);
1017 	}
1018 
1019 	/*
1020 	 * Don't concatenate if either segment uses text replication.
1021 	 */
1022 	if (svd1->tr_state != SEGVN_TR_OFF || svd2->tr_state != SEGVN_TR_OFF) {
1023 		return (-1);
1024 	}
1025 
1026 	/*
1027 	 * Fail early if we're not supposed to concatenate
1028 	 * segments with non NULL amp.
1029 	 */
1030 	if (amp_cat == 0 && (amp1 != NULL || amp2 != NULL)) {
1031 		return (-1);
1032 	}
1033 
1034 	if (svd1->vp == NULL && svd1->type == MAP_SHARED) {
1035 		if (amp1 != amp2) {
1036 			return (-1);
1037 		}
1038 		if (amp1 != NULL && svd1->anon_index + btop(seg1->s_size) !=
1039 		    svd2->anon_index) {
1040 			return (-1);
1041 		}
1042 		ASSERT(amp1 == NULL || amp1->refcnt >= 2);
1043 	}
1044 
1045 	/*
1046 	 * If either seg has vpages, create a new merged vpage array.
1047 	 */
1048 	if (vpage1 != NULL || vpage2 != NULL) {
1049 		struct vpage *vp, *evp;
1050 
1051 		npages1 = seg_pages(seg1);
1052 		npages2 = seg_pages(seg2);
1053 		nvpsize = vpgtob(npages1 + npages2);
1054 
1055 		if ((nvpage = kmem_zalloc(nvpsize, KM_NOSLEEP)) == NULL) {
1056 			return (-2);
1057 		}
1058 
1059 		if (vpage1 != NULL) {
1060 			bcopy(vpage1, nvpage, vpgtob(npages1));
1061 		} else {
1062 			evp = nvpage + npages1;
1063 			for (vp = nvpage; vp < evp; vp++) {
1064 				VPP_SETPROT(vp, svd1->prot);
1065 				VPP_SETADVICE(vp, svd1->advice);
1066 			}
1067 		}
1068 
1069 		if (vpage2 != NULL) {
1070 			bcopy(vpage2, nvpage + npages1, vpgtob(npages2));
1071 		} else {
1072 			evp = nvpage + npages1 + npages2;
1073 			for (vp = nvpage + npages1; vp < evp; vp++) {
1074 				VPP_SETPROT(vp, svd2->prot);
1075 				VPP_SETADVICE(vp, svd2->advice);
1076 			}
1077 		}
1078 
1079 		if (svd2->pageswap && (!svd1->pageswap && svd1->swresv)) {
1080 			ASSERT(svd1->swresv == seg1->s_size);
1081 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1082 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1083 			evp = nvpage + npages1;
1084 			for (vp = nvpage; vp < evp; vp++) {
1085 				VPP_SETSWAPRES(vp);
1086 			}
1087 		}
1088 
1089 		if (svd1->pageswap && (!svd2->pageswap && svd2->swresv)) {
1090 			ASSERT(svd2->swresv == seg2->s_size);
1091 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1092 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1093 			vp = nvpage + npages1;
1094 			evp = vp + npages2;
1095 			for (; vp < evp; vp++) {
1096 				VPP_SETSWAPRES(vp);
1097 			}
1098 		}
1099 	}
1100 	ASSERT((vpage1 != NULL || vpage2 != NULL) ||
1101 	    (svd1->pageswap == 0 && svd2->pageswap == 0));
1102 
1103 	/*
1104 	 * If either segment has private pages, create a new merged anon
1105 	 * array. If mergeing shared anon segments just decrement anon map's
1106 	 * refcnt.
1107 	 */
1108 	if (amp1 != NULL && svd1->type == MAP_SHARED) {
1109 		ASSERT(amp1 == amp2 && svd1->vp == NULL);
1110 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1111 		ASSERT(amp1->refcnt >= 2);
1112 		amp1->refcnt--;
1113 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1114 		svd2->amp = NULL;
1115 	} else if (amp1 != NULL || amp2 != NULL) {
1116 		struct anon_hdr *nahp;
1117 		struct anon_map *namp = NULL;
1118 		size_t asize;
1119 
1120 		ASSERT(svd1->type == MAP_PRIVATE);
1121 
1122 		asize = seg1->s_size + seg2->s_size;
1123 		if ((nahp = anon_create(btop(asize), ANON_NOSLEEP)) == NULL) {
1124 			if (nvpage != NULL) {
1125 				kmem_free(nvpage, nvpsize);
1126 			}
1127 			return (-2);
1128 		}
1129 		if (amp1 != NULL) {
1130 			/*
1131 			 * XXX anon rwlock is not really needed because
1132 			 * this is a private segment and we are writers.
1133 			 */
1134 			ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1135 			ASSERT(amp1->refcnt == 1);
1136 			if (anon_copy_ptr(amp1->ahp, svd1->anon_index,
1137 			    nahp, 0, btop(seg1->s_size), ANON_NOSLEEP)) {
1138 				anon_release(nahp, btop(asize));
1139 				ANON_LOCK_EXIT(&amp1->a_rwlock);
1140 				if (nvpage != NULL) {
1141 					kmem_free(nvpage, nvpsize);
1142 				}
1143 				return (-2);
1144 			}
1145 		}
1146 		if (amp2 != NULL) {
1147 			ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1148 			ASSERT(amp2->refcnt == 1);
1149 			if (anon_copy_ptr(amp2->ahp, svd2->anon_index,
1150 			    nahp, btop(seg1->s_size), btop(seg2->s_size),
1151 			    ANON_NOSLEEP)) {
1152 				anon_release(nahp, btop(asize));
1153 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1154 				if (amp1 != NULL) {
1155 					ANON_LOCK_EXIT(&amp1->a_rwlock);
1156 				}
1157 				if (nvpage != NULL) {
1158 					kmem_free(nvpage, nvpsize);
1159 				}
1160 				return (-2);
1161 			}
1162 		}
1163 		if (amp1 != NULL) {
1164 			namp = amp1;
1165 			anon_release(amp1->ahp, btop(amp1->size));
1166 		}
1167 		if (amp2 != NULL) {
1168 			if (namp == NULL) {
1169 				ASSERT(amp1 == NULL);
1170 				namp = amp2;
1171 				anon_release(amp2->ahp, btop(amp2->size));
1172 			} else {
1173 				amp2->refcnt--;
1174 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1175 				anonmap_free(amp2);
1176 			}
1177 			svd2->amp = NULL; /* needed for seg_free */
1178 		}
1179 		namp->ahp = nahp;
1180 		namp->size = asize;
1181 		svd1->amp = namp;
1182 		svd1->anon_index = 0;
1183 		ANON_LOCK_EXIT(&namp->a_rwlock);
1184 	}
1185 	/*
1186 	 * Now free the old vpage structures.
1187 	 */
1188 	if (nvpage != NULL) {
1189 		if (vpage1 != NULL) {
1190 			kmem_free(vpage1, vpgtob(npages1));
1191 		}
1192 		if (vpage2 != NULL) {
1193 			svd2->vpage = NULL;
1194 			kmem_free(vpage2, vpgtob(npages2));
1195 		}
1196 		if (svd2->pageprot) {
1197 			svd1->pageprot = 1;
1198 		}
1199 		if (svd2->pageadvice) {
1200 			svd1->pageadvice = 1;
1201 		}
1202 		if (svd2->pageswap) {
1203 			svd1->pageswap = 1;
1204 		}
1205 		svd1->vpage = nvpage;
1206 	}
1207 
1208 	/* all looks ok, merge segments */
1209 	svd1->swresv += svd2->swresv;
1210 	svd2->swresv = 0;  /* so seg_free doesn't release swap space */
1211 	size = seg2->s_size;
1212 	seg_free(seg2);
1213 	seg1->s_size += size;
1214 	return (0);
1215 }
1216 
1217 /*
1218  * Extend the previous segment (seg1) to include the
1219  * new segment (seg2 + a), if possible.
1220  * Return 0 on success.
1221  */
1222 static int
1223 segvn_extend_prev(seg1, seg2, a, swresv)
1224 	struct seg *seg1, *seg2;
1225 	struct segvn_crargs *a;
1226 	size_t swresv;
1227 {
1228 	struct segvn_data *svd1 = (struct segvn_data *)seg1->s_data;
1229 	size_t size;
1230 	struct anon_map *amp1;
1231 	struct vpage *new_vpage;
1232 
1233 	/*
1234 	 * We don't need any segment level locks for "segvn" data
1235 	 * since the address space is "write" locked.
1236 	 */
1237 	ASSERT(seg1->s_as && AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
1238 
1239 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie)) {
1240 		return (-1);
1241 	}
1242 
1243 	/* second segment is new, try to extend first */
1244 	/* XXX - should also check cred */
1245 	if (svd1->vp != a->vp || svd1->maxprot != a->maxprot ||
1246 	    (!svd1->pageprot && (svd1->prot != a->prot)) ||
1247 	    svd1->type != a->type || svd1->flags != a->flags ||
1248 	    seg1->s_szc != a->szc || svd1->softlockcnt_send > 0)
1249 		return (-1);
1250 
1251 	/* vp == NULL implies zfod, offset doesn't matter */
1252 	if (svd1->vp != NULL &&
1253 	    svd1->offset + seg1->s_size != (a->offset & PAGEMASK))
1254 		return (-1);
1255 
1256 	if (svd1->tr_state != SEGVN_TR_OFF) {
1257 		return (-1);
1258 	}
1259 
1260 	amp1 = svd1->amp;
1261 	if (amp1) {
1262 		pgcnt_t newpgs;
1263 
1264 		/*
1265 		 * Segment has private pages, can data structures
1266 		 * be expanded?
1267 		 *
1268 		 * Acquire the anon_map lock to prevent it from changing,
1269 		 * if it is shared.  This ensures that the anon_map
1270 		 * will not change while a thread which has a read/write
1271 		 * lock on an address space references it.
1272 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1273 		 * is 1.
1274 		 *
1275 		 * Can't grow a MAP_SHARED segment with an anonmap because
1276 		 * there may be existing anon slots where we want to extend
1277 		 * the segment and we wouldn't know what to do with them
1278 		 * (e.g., for tmpfs right thing is to just leave them there,
1279 		 * for /dev/zero they should be cleared out).
1280 		 */
1281 		if (svd1->type == MAP_SHARED)
1282 			return (-1);
1283 
1284 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1285 		if (amp1->refcnt > 1) {
1286 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1287 			return (-1);
1288 		}
1289 		newpgs = anon_grow(amp1->ahp, &svd1->anon_index,
1290 		    btop(seg1->s_size), btop(seg2->s_size), ANON_NOSLEEP);
1291 
1292 		if (newpgs == 0) {
1293 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1294 			return (-1);
1295 		}
1296 		amp1->size = ptob(newpgs);
1297 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1298 	}
1299 	if (svd1->vpage != NULL) {
1300 		struct vpage *vp, *evp;
1301 		new_vpage =
1302 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1303 			KM_NOSLEEP);
1304 		if (new_vpage == NULL)
1305 			return (-1);
1306 		bcopy(svd1->vpage, new_vpage, vpgtob(seg_pages(seg1)));
1307 		kmem_free(svd1->vpage, vpgtob(seg_pages(seg1)));
1308 		svd1->vpage = new_vpage;
1309 
1310 		vp = new_vpage + seg_pages(seg1);
1311 		evp = vp + seg_pages(seg2);
1312 		for (; vp < evp; vp++)
1313 			VPP_SETPROT(vp, a->prot);
1314 		if (svd1->pageswap && swresv) {
1315 			ASSERT(!(svd1->flags & MAP_NORESERVE));
1316 			ASSERT(swresv == seg2->s_size);
1317 			vp = new_vpage + seg_pages(seg1);
1318 			for (; vp < evp; vp++) {
1319 				VPP_SETSWAPRES(vp);
1320 			}
1321 		}
1322 	}
1323 	ASSERT(svd1->vpage != NULL || svd1->pageswap == 0);
1324 	size = seg2->s_size;
1325 	seg_free(seg2);
1326 	seg1->s_size += size;
1327 	svd1->swresv += swresv;
1328 	if (svd1->pageprot && (a->prot & PROT_WRITE) &&
1329 	    svd1->type == MAP_SHARED && svd1->vp != NULL &&
1330 	    (svd1->vp->v_flag & VVMEXEC)) {
1331 		ASSERT(vn_is_mapped(svd1->vp, V_WRITE));
1332 		segvn_inval_trcache(svd1->vp);
1333 	}
1334 	return (0);
1335 }
1336 
1337 /*
1338  * Extend the next segment (seg2) to include the
1339  * new segment (seg1 + a), if possible.
1340  * Return 0 on success.
1341  */
1342 static int
1343 segvn_extend_next(
1344 	struct seg *seg1,
1345 	struct seg *seg2,
1346 	struct segvn_crargs *a,
1347 	size_t swresv)
1348 {
1349 	struct segvn_data *svd2 = (struct segvn_data *)seg2->s_data;
1350 	size_t size;
1351 	struct anon_map *amp2;
1352 	struct vpage *new_vpage;
1353 
1354 	/*
1355 	 * We don't need any segment level locks for "segvn" data
1356 	 * since the address space is "write" locked.
1357 	 */
1358 	ASSERT(seg2->s_as && AS_WRITE_HELD(seg2->s_as, &seg2->s_as->a_lock));
1359 
1360 	if (HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1361 		return (-1);
1362 	}
1363 
1364 	/* first segment is new, try to extend second */
1365 	/* XXX - should also check cred */
1366 	if (svd2->vp != a->vp || svd2->maxprot != a->maxprot ||
1367 	    (!svd2->pageprot && (svd2->prot != a->prot)) ||
1368 	    svd2->type != a->type || svd2->flags != a->flags ||
1369 	    seg2->s_szc != a->szc || svd2->softlockcnt_sbase > 0)
1370 		return (-1);
1371 	/* vp == NULL implies zfod, offset doesn't matter */
1372 	if (svd2->vp != NULL &&
1373 	    (a->offset & PAGEMASK) + seg1->s_size != svd2->offset)
1374 		return (-1);
1375 
1376 	if (svd2->tr_state != SEGVN_TR_OFF) {
1377 		return (-1);
1378 	}
1379 
1380 	amp2 = svd2->amp;
1381 	if (amp2) {
1382 		pgcnt_t newpgs;
1383 
1384 		/*
1385 		 * Segment has private pages, can data structures
1386 		 * be expanded?
1387 		 *
1388 		 * Acquire the anon_map lock to prevent it from changing,
1389 		 * if it is shared.  This ensures that the anon_map
1390 		 * will not change while a thread which has a read/write
1391 		 * lock on an address space references it.
1392 		 *
1393 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1394 		 * is 1.
1395 		 */
1396 		if (svd2->type == MAP_SHARED)
1397 			return (-1);
1398 
1399 		ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1400 		if (amp2->refcnt > 1) {
1401 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1402 			return (-1);
1403 		}
1404 		newpgs = anon_grow(amp2->ahp, &svd2->anon_index,
1405 		    btop(seg2->s_size), btop(seg1->s_size),
1406 		    ANON_NOSLEEP | ANON_GROWDOWN);
1407 
1408 		if (newpgs == 0) {
1409 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1410 			return (-1);
1411 		}
1412 		amp2->size = ptob(newpgs);
1413 		ANON_LOCK_EXIT(&amp2->a_rwlock);
1414 	}
1415 	if (svd2->vpage != NULL) {
1416 		struct vpage *vp, *evp;
1417 		new_vpage =
1418 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1419 		    KM_NOSLEEP);
1420 		if (new_vpage == NULL) {
1421 			/* Not merging segments so adjust anon_index back */
1422 			if (amp2)
1423 				svd2->anon_index += seg_pages(seg1);
1424 			return (-1);
1425 		}
1426 		bcopy(svd2->vpage, new_vpage + seg_pages(seg1),
1427 		    vpgtob(seg_pages(seg2)));
1428 		kmem_free(svd2->vpage, vpgtob(seg_pages(seg2)));
1429 		svd2->vpage = new_vpage;
1430 
1431 		vp = new_vpage;
1432 		evp = vp + seg_pages(seg1);
1433 		for (; vp < evp; vp++)
1434 			VPP_SETPROT(vp, a->prot);
1435 		if (svd2->pageswap && swresv) {
1436 			ASSERT(!(svd2->flags & MAP_NORESERVE));
1437 			ASSERT(swresv == seg1->s_size);
1438 			vp = new_vpage;
1439 			for (; vp < evp; vp++) {
1440 				VPP_SETSWAPRES(vp);
1441 			}
1442 		}
1443 	}
1444 	ASSERT(svd2->vpage != NULL || svd2->pageswap == 0);
1445 	size = seg1->s_size;
1446 	seg_free(seg1);
1447 	seg2->s_size += size;
1448 	seg2->s_base -= size;
1449 	svd2->offset -= size;
1450 	svd2->swresv += swresv;
1451 	if (svd2->pageprot && (a->prot & PROT_WRITE) &&
1452 	    svd2->type == MAP_SHARED && svd2->vp != NULL &&
1453 	    (svd2->vp->v_flag & VVMEXEC)) {
1454 		ASSERT(vn_is_mapped(svd2->vp, V_WRITE));
1455 		segvn_inval_trcache(svd2->vp);
1456 	}
1457 	return (0);
1458 }
1459 
1460 static int
1461 segvn_dup(struct seg *seg, struct seg *newseg)
1462 {
1463 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1464 	struct segvn_data *newsvd;
1465 	pgcnt_t npages = seg_pages(seg);
1466 	int error = 0;
1467 	uint_t prot;
1468 	size_t len;
1469 	struct anon_map *amp;
1470 
1471 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1472 
1473 	/*
1474 	 * If segment has anon reserved, reserve more for the new seg.
1475 	 * For a MAP_NORESERVE segment swresv will be a count of all the
1476 	 * allocated anon slots; thus we reserve for the child as many slots
1477 	 * as the parent has allocated. This semantic prevents the child or
1478 	 * parent from dieing during a copy-on-write fault caused by trying
1479 	 * to write a shared pre-existing anon page.
1480 	 */
1481 	if ((len = svd->swresv) != 0) {
1482 		if (anon_resv(svd->swresv) == 0)
1483 			return (ENOMEM);
1484 
1485 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
1486 		    seg, len, 0);
1487 	}
1488 
1489 	newsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
1490 
1491 	newseg->s_ops = &segvn_ops;
1492 	newseg->s_data = (void *)newsvd;
1493 	newseg->s_szc = seg->s_szc;
1494 
1495 	newsvd->seg = newseg;
1496 	if ((newsvd->vp = svd->vp) != NULL) {
1497 		VN_HOLD(svd->vp);
1498 		if (svd->type == MAP_SHARED)
1499 			lgrp_shm_policy_init(NULL, svd->vp);
1500 	}
1501 	newsvd->offset = svd->offset;
1502 	newsvd->prot = svd->prot;
1503 	newsvd->maxprot = svd->maxprot;
1504 	newsvd->pageprot = svd->pageprot;
1505 	newsvd->type = svd->type;
1506 	newsvd->cred = svd->cred;
1507 	crhold(newsvd->cred);
1508 	newsvd->advice = svd->advice;
1509 	newsvd->pageadvice = svd->pageadvice;
1510 	newsvd->swresv = svd->swresv;
1511 	newsvd->pageswap = svd->pageswap;
1512 	newsvd->flags = svd->flags;
1513 	newsvd->softlockcnt = 0;
1514 	newsvd->softlockcnt_sbase = 0;
1515 	newsvd->softlockcnt_send = 0;
1516 	newsvd->policy_info = svd->policy_info;
1517 	newsvd->rcookie = HAT_INVALID_REGION_COOKIE;
1518 
1519 	if ((amp = svd->amp) == NULL || svd->tr_state == SEGVN_TR_ON) {
1520 		/*
1521 		 * Not attaching to a shared anon object.
1522 		 */
1523 		ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie) ||
1524 		    svd->tr_state == SEGVN_TR_OFF);
1525 		if (svd->tr_state == SEGVN_TR_ON) {
1526 			ASSERT(newsvd->vp != NULL && amp != NULL);
1527 			newsvd->tr_state = SEGVN_TR_INIT;
1528 		} else {
1529 			newsvd->tr_state = svd->tr_state;
1530 		}
1531 		newsvd->amp = NULL;
1532 		newsvd->anon_index = 0;
1533 	} else {
1534 		/* regions for now are only used on pure vnode segments */
1535 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
1536 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1537 		newsvd->tr_state = SEGVN_TR_OFF;
1538 		if (svd->type == MAP_SHARED) {
1539 			newsvd->amp = amp;
1540 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1541 			amp->refcnt++;
1542 			ANON_LOCK_EXIT(&amp->a_rwlock);
1543 			newsvd->anon_index = svd->anon_index;
1544 		} else {
1545 			int reclaim = 1;
1546 
1547 			/*
1548 			 * Allocate and initialize new anon_map structure.
1549 			 */
1550 			newsvd->amp = anonmap_alloc(newseg->s_size, 0,
1551 			    ANON_SLEEP);
1552 			newsvd->amp->a_szc = newseg->s_szc;
1553 			newsvd->anon_index = 0;
1554 
1555 			/*
1556 			 * We don't have to acquire the anon_map lock
1557 			 * for the new segment (since it belongs to an
1558 			 * address space that is still not associated
1559 			 * with any process), or the segment in the old
1560 			 * address space (since all threads in it
1561 			 * are stopped while duplicating the address space).
1562 			 */
1563 
1564 			/*
1565 			 * The goal of the following code is to make sure that
1566 			 * softlocked pages do not end up as copy on write
1567 			 * pages.  This would cause problems where one
1568 			 * thread writes to a page that is COW and a different
1569 			 * thread in the same process has softlocked it.  The
1570 			 * softlock lock would move away from this process
1571 			 * because the write would cause this process to get
1572 			 * a copy (without the softlock).
1573 			 *
1574 			 * The strategy here is to just break the
1575 			 * sharing on pages that could possibly be
1576 			 * softlocked.
1577 			 */
1578 retry:
1579 			if (svd->softlockcnt) {
1580 				struct anon *ap, *newap;
1581 				size_t i;
1582 				uint_t vpprot;
1583 				page_t *anon_pl[1+1], *pp;
1584 				caddr_t addr;
1585 				ulong_t old_idx = svd->anon_index;
1586 				ulong_t new_idx = 0;
1587 
1588 				/*
1589 				 * The softlock count might be non zero
1590 				 * because some pages are still stuck in the
1591 				 * cache for lazy reclaim. Flush the cache
1592 				 * now. This should drop the count to zero.
1593 				 * [or there is really I/O going on to these
1594 				 * pages]. Note, we have the writers lock so
1595 				 * nothing gets inserted during the flush.
1596 				 */
1597 				if (reclaim == 1) {
1598 					segvn_purge(seg);
1599 					reclaim = 0;
1600 					goto retry;
1601 				}
1602 				i = btopr(seg->s_size);
1603 				addr = seg->s_base;
1604 				/*
1605 				 * XXX break cow sharing using PAGESIZE
1606 				 * pages. They will be relocated into larger
1607 				 * pages at fault time.
1608 				 */
1609 				while (i-- > 0) {
1610 					if (ap = anon_get_ptr(amp->ahp,
1611 					    old_idx)) {
1612 						error = anon_getpage(&ap,
1613 						    &vpprot, anon_pl, PAGESIZE,
1614 						    seg, addr, S_READ,
1615 						    svd->cred);
1616 						if (error) {
1617 							newsvd->vpage = NULL;
1618 							goto out;
1619 						}
1620 						/*
1621 						 * prot need not be computed
1622 						 * below 'cause anon_private is
1623 						 * going to ignore it anyway
1624 						 * as child doesn't inherit
1625 						 * pagelock from parent.
1626 						 */
1627 						prot = svd->pageprot ?
1628 						    VPP_PROT(
1629 						    &svd->vpage[
1630 						    seg_page(seg, addr)])
1631 						    : svd->prot;
1632 						pp = anon_private(&newap,
1633 						    newseg, addr, prot,
1634 						    anon_pl[0],	0,
1635 						    newsvd->cred);
1636 						if (pp == NULL) {
1637 							/* no mem abort */
1638 							newsvd->vpage = NULL;
1639 							error = ENOMEM;
1640 							goto out;
1641 						}
1642 						(void) anon_set_ptr(
1643 						    newsvd->amp->ahp, new_idx,
1644 						    newap, ANON_SLEEP);
1645 						page_unlock(pp);
1646 					}
1647 					addr += PAGESIZE;
1648 					old_idx++;
1649 					new_idx++;
1650 				}
1651 			} else {	/* common case */
1652 				if (seg->s_szc != 0) {
1653 					/*
1654 					 * If at least one of anon slots of a
1655 					 * large page exists then make sure
1656 					 * all anon slots of a large page
1657 					 * exist to avoid partial cow sharing
1658 					 * of a large page in the future.
1659 					 */
1660 					anon_dup_fill_holes(amp->ahp,
1661 					    svd->anon_index, newsvd->amp->ahp,
1662 					    0, seg->s_size, seg->s_szc,
1663 					    svd->vp != NULL);
1664 				} else {
1665 					anon_dup(amp->ahp, svd->anon_index,
1666 					    newsvd->amp->ahp, 0, seg->s_size);
1667 				}
1668 
1669 				hat_clrattr(seg->s_as->a_hat, seg->s_base,
1670 				    seg->s_size, PROT_WRITE);
1671 			}
1672 		}
1673 	}
1674 	/*
1675 	 * If necessary, create a vpage structure for the new segment.
1676 	 * Do not copy any page lock indications.
1677 	 */
1678 	if (svd->vpage != NULL) {
1679 		uint_t i;
1680 		struct vpage *ovp = svd->vpage;
1681 		struct vpage *nvp;
1682 
1683 		nvp = newsvd->vpage =
1684 		    kmem_alloc(vpgtob(npages), KM_SLEEP);
1685 		for (i = 0; i < npages; i++) {
1686 			*nvp = *ovp++;
1687 			VPP_CLRPPLOCK(nvp++);
1688 		}
1689 	} else
1690 		newsvd->vpage = NULL;
1691 
1692 	/* Inform the vnode of the new mapping */
1693 	if (newsvd->vp != NULL) {
1694 		error = VOP_ADDMAP(newsvd->vp, (offset_t)newsvd->offset,
1695 		    newseg->s_as, newseg->s_base, newseg->s_size, newsvd->prot,
1696 		    newsvd->maxprot, newsvd->type, newsvd->cred, NULL);
1697 	}
1698 out:
1699 	if (error == 0 && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1700 		ASSERT(newsvd->amp == NULL);
1701 		ASSERT(newsvd->tr_state == SEGVN_TR_OFF);
1702 		newsvd->rcookie = svd->rcookie;
1703 		hat_dup_region(newseg->s_as->a_hat, newsvd->rcookie);
1704 	}
1705 	return (error);
1706 }
1707 
1708 
1709 /*
1710  * callback function to invoke free_vp_pages() for only those pages actually
1711  * processed by the HAT when a shared region is destroyed.
1712  */
1713 extern int free_pages;
1714 
1715 static void
1716 segvn_hat_rgn_unload_callback(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
1717     size_t r_size, void *r_obj, u_offset_t r_objoff)
1718 {
1719 	u_offset_t off;
1720 	size_t len;
1721 	vnode_t *vp = (vnode_t *)r_obj;
1722 
1723 	ASSERT(eaddr > saddr);
1724 	ASSERT(saddr >= r_saddr);
1725 	ASSERT(saddr < r_saddr + r_size);
1726 	ASSERT(eaddr > r_saddr);
1727 	ASSERT(eaddr <= r_saddr + r_size);
1728 	ASSERT(vp != NULL);
1729 
1730 	if (!free_pages) {
1731 		return;
1732 	}
1733 
1734 	len = eaddr - saddr;
1735 	off = (saddr - r_saddr) + r_objoff;
1736 	free_vp_pages(vp, off, len);
1737 }
1738 
1739 /*
1740  * callback function used by segvn_unmap to invoke free_vp_pages() for only
1741  * those pages actually processed by the HAT
1742  */
1743 static void
1744 segvn_hat_unload_callback(hat_callback_t *cb)
1745 {
1746 	struct seg		*seg = cb->hcb_data;
1747 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
1748 	size_t			len;
1749 	u_offset_t		off;
1750 
1751 	ASSERT(svd->vp != NULL);
1752 	ASSERT(cb->hcb_end_addr > cb->hcb_start_addr);
1753 	ASSERT(cb->hcb_start_addr >= seg->s_base);
1754 
1755 	len = cb->hcb_end_addr - cb->hcb_start_addr;
1756 	off = cb->hcb_start_addr - seg->s_base;
1757 	free_vp_pages(svd->vp, svd->offset + off, len);
1758 }
1759 
1760 /*
1761  * This function determines the number of bytes of swap reserved by
1762  * a segment for which per-page accounting is present. It is used to
1763  * calculate the correct value of a segvn_data's swresv.
1764  */
1765 static size_t
1766 segvn_count_swap_by_vpages(struct seg *seg)
1767 {
1768 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1769 	struct vpage *vp, *evp;
1770 	size_t nswappages = 0;
1771 
1772 	ASSERT(svd->pageswap);
1773 	ASSERT(svd->vpage != NULL);
1774 
1775 	evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
1776 
1777 	for (vp = svd->vpage; vp < evp; vp++) {
1778 		if (VPP_ISSWAPRES(vp))
1779 			nswappages++;
1780 	}
1781 
1782 	return (nswappages << PAGESHIFT);
1783 }
1784 
1785 static int
1786 segvn_unmap(struct seg *seg, caddr_t addr, size_t len)
1787 {
1788 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1789 	struct segvn_data *nsvd;
1790 	struct seg *nseg;
1791 	struct anon_map *amp;
1792 	pgcnt_t	opages;		/* old segment size in pages */
1793 	pgcnt_t	npages;		/* new segment size in pages */
1794 	pgcnt_t	dpages;		/* pages being deleted (unmapped) */
1795 	hat_callback_t callback;	/* used for free_vp_pages() */
1796 	hat_callback_t *cbp = NULL;
1797 	caddr_t nbase;
1798 	size_t nsize;
1799 	size_t oswresv;
1800 	int reclaim = 1;
1801 
1802 	/*
1803 	 * We don't need any segment level locks for "segvn" data
1804 	 * since the address space is "write" locked.
1805 	 */
1806 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1807 
1808 	/*
1809 	 * Fail the unmap if pages are SOFTLOCKed through this mapping.
1810 	 * softlockcnt is protected from change by the as write lock.
1811 	 */
1812 retry:
1813 	if (svd->softlockcnt > 0) {
1814 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1815 
1816 		/*
1817 		 * If this is shared segment non 0 softlockcnt
1818 		 * means locked pages are still in use.
1819 		 */
1820 		if (svd->type == MAP_SHARED) {
1821 			return (EAGAIN);
1822 		}
1823 
1824 		/*
1825 		 * since we do have the writers lock nobody can fill
1826 		 * the cache during the purge. The flush either succeeds
1827 		 * or we still have pending I/Os.
1828 		 */
1829 		if (reclaim == 1) {
1830 			segvn_purge(seg);
1831 			reclaim = 0;
1832 			goto retry;
1833 		}
1834 		return (EAGAIN);
1835 	}
1836 
1837 	/*
1838 	 * Check for bad sizes
1839 	 */
1840 	if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
1841 	    (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
1842 		panic("segvn_unmap");
1843 		/*NOTREACHED*/
1844 	}
1845 
1846 	if (seg->s_szc != 0) {
1847 		size_t pgsz = page_get_pagesize(seg->s_szc);
1848 		int err;
1849 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
1850 			ASSERT(seg->s_base != addr || seg->s_size != len);
1851 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1852 				ASSERT(svd->amp == NULL);
1853 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1854 				hat_leave_region(seg->s_as->a_hat,
1855 				    svd->rcookie, HAT_REGION_TEXT);
1856 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
1857 				/*
1858 				 * could pass a flag to segvn_demote_range()
1859 				 * below to tell it not to do any unloads but
1860 				 * this case is rare enough to not bother for
1861 				 * now.
1862 				 */
1863 			} else if (svd->tr_state == SEGVN_TR_INIT) {
1864 				svd->tr_state = SEGVN_TR_OFF;
1865 			} else if (svd->tr_state == SEGVN_TR_ON) {
1866 				ASSERT(svd->amp != NULL);
1867 				segvn_textunrepl(seg, 1);
1868 				ASSERT(svd->amp == NULL);
1869 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1870 			}
1871 			VM_STAT_ADD(segvnvmstats.demoterange[0]);
1872 			err = segvn_demote_range(seg, addr, len, SDR_END, 0);
1873 			if (err == 0) {
1874 				return (IE_RETRY);
1875 			}
1876 			return (err);
1877 		}
1878 	}
1879 
1880 	/* Inform the vnode of the unmapping. */
1881 	if (svd->vp) {
1882 		int error;
1883 
1884 		error = VOP_DELMAP(svd->vp,
1885 		    (offset_t)svd->offset + (uintptr_t)(addr - seg->s_base),
1886 		    seg->s_as, addr, len, svd->prot, svd->maxprot,
1887 		    svd->type, svd->cred, NULL);
1888 
1889 		if (error == EAGAIN)
1890 			return (error);
1891 	}
1892 
1893 	/*
1894 	 * Remove any page locks set through this mapping.
1895 	 * If text replication is not off no page locks could have been
1896 	 * established via this mapping.
1897 	 */
1898 	if (svd->tr_state == SEGVN_TR_OFF) {
1899 		(void) segvn_lockop(seg, addr, len, 0, MC_UNLOCK, NULL, 0);
1900 	}
1901 
1902 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1903 		ASSERT(svd->amp == NULL);
1904 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1905 		ASSERT(svd->type == MAP_PRIVATE);
1906 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
1907 		    HAT_REGION_TEXT);
1908 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
1909 	} else if (svd->tr_state == SEGVN_TR_ON) {
1910 		ASSERT(svd->amp != NULL);
1911 		ASSERT(svd->pageprot == 0 && !(svd->prot & PROT_WRITE));
1912 		segvn_textunrepl(seg, 1);
1913 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
1914 	} else {
1915 		if (svd->tr_state != SEGVN_TR_OFF) {
1916 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
1917 			svd->tr_state = SEGVN_TR_OFF;
1918 		}
1919 		/*
1920 		 * Unload any hardware translations in the range to be taken
1921 		 * out. Use a callback to invoke free_vp_pages() effectively.
1922 		 */
1923 		if (svd->vp != NULL && free_pages != 0) {
1924 			callback.hcb_data = seg;
1925 			callback.hcb_function = segvn_hat_unload_callback;
1926 			cbp = &callback;
1927 		}
1928 		hat_unload_callback(seg->s_as->a_hat, addr, len,
1929 		    HAT_UNLOAD_UNMAP, cbp);
1930 
1931 		if (svd->type == MAP_SHARED && svd->vp != NULL &&
1932 		    (svd->vp->v_flag & VVMEXEC) &&
1933 		    ((svd->prot & PROT_WRITE) || svd->pageprot)) {
1934 			segvn_inval_trcache(svd->vp);
1935 		}
1936 	}
1937 
1938 	/*
1939 	 * Check for entire segment
1940 	 */
1941 	if (addr == seg->s_base && len == seg->s_size) {
1942 		seg_free(seg);
1943 		return (0);
1944 	}
1945 
1946 	opages = seg_pages(seg);
1947 	dpages = btop(len);
1948 	npages = opages - dpages;
1949 	amp = svd->amp;
1950 	ASSERT(amp == NULL || amp->a_szc >= seg->s_szc);
1951 
1952 	/*
1953 	 * Check for beginning of segment
1954 	 */
1955 	if (addr == seg->s_base) {
1956 		if (svd->vpage != NULL) {
1957 			size_t nbytes;
1958 			struct vpage *ovpage;
1959 
1960 			ovpage = svd->vpage;	/* keep pointer to vpage */
1961 
1962 			nbytes = vpgtob(npages);
1963 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
1964 			bcopy(&ovpage[dpages], svd->vpage, nbytes);
1965 
1966 			/* free up old vpage */
1967 			kmem_free(ovpage, vpgtob(opages));
1968 		}
1969 		if (amp != NULL) {
1970 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1971 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
1972 				/*
1973 				 * Shared anon map is no longer in use. Before
1974 				 * freeing its pages purge all entries from
1975 				 * pcache that belong to this amp.
1976 				 */
1977 				if (svd->type == MAP_SHARED) {
1978 					ASSERT(amp->refcnt == 1);
1979 					ASSERT(svd->softlockcnt == 0);
1980 					anonmap_purge(amp);
1981 				}
1982 				/*
1983 				 * Free up now unused parts of anon_map array.
1984 				 */
1985 				if (amp->a_szc == seg->s_szc) {
1986 					if (seg->s_szc != 0) {
1987 						anon_free_pages(amp->ahp,
1988 						    svd->anon_index, len,
1989 						    seg->s_szc);
1990 					} else {
1991 						anon_free(amp->ahp,
1992 						    svd->anon_index,
1993 						    len);
1994 					}
1995 				} else {
1996 					ASSERT(svd->type == MAP_SHARED);
1997 					ASSERT(amp->a_szc > seg->s_szc);
1998 					anon_shmap_free_pages(amp,
1999 					    svd->anon_index, len);
2000 				}
2001 
2002 				/*
2003 				 * Unreserve swap space for the
2004 				 * unmapped chunk of this segment in
2005 				 * case it's MAP_SHARED
2006 				 */
2007 				if (svd->type == MAP_SHARED) {
2008 					anon_unresv(len);
2009 					amp->swresv -= len;
2010 				}
2011 			}
2012 			ANON_LOCK_EXIT(&amp->a_rwlock);
2013 			svd->anon_index += dpages;
2014 		}
2015 		if (svd->vp != NULL)
2016 			svd->offset += len;
2017 
2018 		seg->s_base += len;
2019 		seg->s_size -= len;
2020 
2021 		if (svd->swresv) {
2022 			if (svd->flags & MAP_NORESERVE) {
2023 				ASSERT(amp);
2024 				oswresv = svd->swresv;
2025 
2026 				svd->swresv = ptob(anon_pages(amp->ahp,
2027 				    svd->anon_index, npages));
2028 				anon_unresv(oswresv - svd->swresv);
2029 			} else {
2030 				size_t unlen;
2031 
2032 				if (svd->pageswap) {
2033 					oswresv = svd->swresv;
2034 					svd->swresv =
2035 					    segvn_count_swap_by_vpages(seg);
2036 					ASSERT(oswresv >= svd->swresv);
2037 					unlen = oswresv - svd->swresv;
2038 				} else {
2039 					svd->swresv -= len;
2040 					ASSERT(svd->swresv == seg->s_size);
2041 					unlen = len;
2042 				}
2043 				anon_unresv(unlen);
2044 			}
2045 			TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2046 			    seg, len, 0);
2047 		}
2048 
2049 		return (0);
2050 	}
2051 
2052 	/*
2053 	 * Check for end of segment
2054 	 */
2055 	if (addr + len == seg->s_base + seg->s_size) {
2056 		if (svd->vpage != NULL) {
2057 			size_t nbytes;
2058 			struct vpage *ovpage;
2059 
2060 			ovpage = svd->vpage;	/* keep pointer to vpage */
2061 
2062 			nbytes = vpgtob(npages);
2063 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2064 			bcopy(ovpage, svd->vpage, nbytes);
2065 
2066 			/* free up old vpage */
2067 			kmem_free(ovpage, vpgtob(opages));
2068 
2069 		}
2070 		if (amp != NULL) {
2071 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2072 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2073 				/*
2074 				 * Free up now unused parts of anon_map array.
2075 				 */
2076 				ulong_t an_idx = svd->anon_index + npages;
2077 
2078 				/*
2079 				 * Shared anon map is no longer in use. Before
2080 				 * freeing its pages purge all entries from
2081 				 * pcache that belong to this amp.
2082 				 */
2083 				if (svd->type == MAP_SHARED) {
2084 					ASSERT(amp->refcnt == 1);
2085 					ASSERT(svd->softlockcnt == 0);
2086 					anonmap_purge(amp);
2087 				}
2088 
2089 				if (amp->a_szc == seg->s_szc) {
2090 					if (seg->s_szc != 0) {
2091 						anon_free_pages(amp->ahp,
2092 						    an_idx, len,
2093 						    seg->s_szc);
2094 					} else {
2095 						anon_free(amp->ahp, an_idx,
2096 						    len);
2097 					}
2098 				} else {
2099 					ASSERT(svd->type == MAP_SHARED);
2100 					ASSERT(amp->a_szc > seg->s_szc);
2101 					anon_shmap_free_pages(amp,
2102 					    an_idx, len);
2103 				}
2104 
2105 				/*
2106 				 * Unreserve swap space for the
2107 				 * unmapped chunk of this segment in
2108 				 * case it's MAP_SHARED
2109 				 */
2110 				if (svd->type == MAP_SHARED) {
2111 					anon_unresv(len);
2112 					amp->swresv -= len;
2113 				}
2114 			}
2115 			ANON_LOCK_EXIT(&amp->a_rwlock);
2116 		}
2117 
2118 		seg->s_size -= len;
2119 
2120 		if (svd->swresv) {
2121 			if (svd->flags & MAP_NORESERVE) {
2122 				ASSERT(amp);
2123 				oswresv = svd->swresv;
2124 				svd->swresv = ptob(anon_pages(amp->ahp,
2125 				    svd->anon_index, npages));
2126 				anon_unresv(oswresv - svd->swresv);
2127 			} else {
2128 				size_t unlen;
2129 
2130 				if (svd->pageswap) {
2131 					oswresv = svd->swresv;
2132 					svd->swresv =
2133 					    segvn_count_swap_by_vpages(seg);
2134 					ASSERT(oswresv >= svd->swresv);
2135 					unlen = oswresv - svd->swresv;
2136 				} else {
2137 					svd->swresv -= len;
2138 					ASSERT(svd->swresv == seg->s_size);
2139 					unlen = len;
2140 				}
2141 				anon_unresv(unlen);
2142 			}
2143 			TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2144 			    "anon proc:%p %lu %u", seg, len, 0);
2145 		}
2146 
2147 		return (0);
2148 	}
2149 
2150 	/*
2151 	 * The section to go is in the middle of the segment,
2152 	 * have to make it into two segments.  nseg is made for
2153 	 * the high end while seg is cut down at the low end.
2154 	 */
2155 	nbase = addr + len;				/* new seg base */
2156 	nsize = (seg->s_base + seg->s_size) - nbase;	/* new seg size */
2157 	seg->s_size = addr - seg->s_base;		/* shrink old seg */
2158 	nseg = seg_alloc(seg->s_as, nbase, nsize);
2159 	if (nseg == NULL) {
2160 		panic("segvn_unmap seg_alloc");
2161 		/*NOTREACHED*/
2162 	}
2163 	nseg->s_ops = seg->s_ops;
2164 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
2165 	nseg->s_data = (void *)nsvd;
2166 	nseg->s_szc = seg->s_szc;
2167 	*nsvd = *svd;
2168 	nsvd->seg = nseg;
2169 	nsvd->offset = svd->offset + (uintptr_t)(nseg->s_base - seg->s_base);
2170 	nsvd->swresv = 0;
2171 	nsvd->softlockcnt = 0;
2172 	nsvd->softlockcnt_sbase = 0;
2173 	nsvd->softlockcnt_send = 0;
2174 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
2175 
2176 	if (svd->vp != NULL) {
2177 		VN_HOLD(nsvd->vp);
2178 		if (nsvd->type == MAP_SHARED)
2179 			lgrp_shm_policy_init(NULL, nsvd->vp);
2180 	}
2181 	crhold(svd->cred);
2182 
2183 	if (svd->vpage == NULL) {
2184 		nsvd->vpage = NULL;
2185 	} else {
2186 		/* need to split vpage into two arrays */
2187 		size_t nbytes;
2188 		struct vpage *ovpage;
2189 
2190 		ovpage = svd->vpage;		/* keep pointer to vpage */
2191 
2192 		npages = seg_pages(seg);	/* seg has shrunk */
2193 		nbytes = vpgtob(npages);
2194 		svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2195 
2196 		bcopy(ovpage, svd->vpage, nbytes);
2197 
2198 		npages = seg_pages(nseg);
2199 		nbytes = vpgtob(npages);
2200 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2201 
2202 		bcopy(&ovpage[opages - npages], nsvd->vpage, nbytes);
2203 
2204 		/* free up old vpage */
2205 		kmem_free(ovpage, vpgtob(opages));
2206 	}
2207 
2208 	if (amp == NULL) {
2209 		nsvd->amp = NULL;
2210 		nsvd->anon_index = 0;
2211 	} else {
2212 		/*
2213 		 * Need to create a new anon map for the new segment.
2214 		 * We'll also allocate a new smaller array for the old
2215 		 * smaller segment to save space.
2216 		 */
2217 		opages = btop((uintptr_t)(addr - seg->s_base));
2218 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2219 		if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2220 			/*
2221 			 * Free up now unused parts of anon_map array.
2222 			 */
2223 			ulong_t an_idx = svd->anon_index + opages;
2224 
2225 			/*
2226 			 * Shared anon map is no longer in use. Before
2227 			 * freeing its pages purge all entries from
2228 			 * pcache that belong to this amp.
2229 			 */
2230 			if (svd->type == MAP_SHARED) {
2231 				ASSERT(amp->refcnt == 1);
2232 				ASSERT(svd->softlockcnt == 0);
2233 				anonmap_purge(amp);
2234 			}
2235 
2236 			if (amp->a_szc == seg->s_szc) {
2237 				if (seg->s_szc != 0) {
2238 					anon_free_pages(amp->ahp, an_idx, len,
2239 					    seg->s_szc);
2240 				} else {
2241 					anon_free(amp->ahp, an_idx,
2242 					    len);
2243 				}
2244 			} else {
2245 				ASSERT(svd->type == MAP_SHARED);
2246 				ASSERT(amp->a_szc > seg->s_szc);
2247 				anon_shmap_free_pages(amp, an_idx, len);
2248 			}
2249 
2250 			/*
2251 			 * Unreserve swap space for the
2252 			 * unmapped chunk of this segment in
2253 			 * case it's MAP_SHARED
2254 			 */
2255 			if (svd->type == MAP_SHARED) {
2256 				anon_unresv(len);
2257 				amp->swresv -= len;
2258 			}
2259 		}
2260 		nsvd->anon_index = svd->anon_index +
2261 		    btop((uintptr_t)(nseg->s_base - seg->s_base));
2262 		if (svd->type == MAP_SHARED) {
2263 			amp->refcnt++;
2264 			nsvd->amp = amp;
2265 		} else {
2266 			struct anon_map *namp;
2267 			struct anon_hdr *nahp;
2268 
2269 			ASSERT(svd->type == MAP_PRIVATE);
2270 			nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
2271 			namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
2272 			namp->a_szc = seg->s_szc;
2273 			(void) anon_copy_ptr(amp->ahp, svd->anon_index, nahp,
2274 			    0, btop(seg->s_size), ANON_SLEEP);
2275 			(void) anon_copy_ptr(amp->ahp, nsvd->anon_index,
2276 			    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
2277 			anon_release(amp->ahp, btop(amp->size));
2278 			svd->anon_index = 0;
2279 			nsvd->anon_index = 0;
2280 			amp->ahp = nahp;
2281 			amp->size = seg->s_size;
2282 			nsvd->amp = namp;
2283 		}
2284 		ANON_LOCK_EXIT(&amp->a_rwlock);
2285 	}
2286 	if (svd->swresv) {
2287 		if (svd->flags & MAP_NORESERVE) {
2288 			ASSERT(amp);
2289 			oswresv = svd->swresv;
2290 			svd->swresv = ptob(anon_pages(amp->ahp,
2291 			    svd->anon_index, btop(seg->s_size)));
2292 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
2293 			    nsvd->anon_index, btop(nseg->s_size)));
2294 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2295 			anon_unresv(oswresv - (svd->swresv + nsvd->swresv));
2296 		} else {
2297 			size_t unlen;
2298 
2299 			if (svd->pageswap) {
2300 				oswresv = svd->swresv;
2301 				svd->swresv = segvn_count_swap_by_vpages(seg);
2302 				nsvd->swresv = segvn_count_swap_by_vpages(nseg);
2303 				ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2304 				unlen = oswresv - (svd->swresv + nsvd->swresv);
2305 			} else {
2306 				if (seg->s_size + nseg->s_size + len !=
2307 				    svd->swresv) {
2308 					panic("segvn_unmap: cannot split "
2309 					    "swap reservation");
2310 					/*NOTREACHED*/
2311 				}
2312 				svd->swresv = seg->s_size;
2313 				nsvd->swresv = nseg->s_size;
2314 				unlen = len;
2315 			}
2316 			anon_unresv(unlen);
2317 		}
2318 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2319 		    seg, len, 0);
2320 	}
2321 
2322 	return (0);			/* I'm glad that's all over with! */
2323 }
2324 
2325 static void
2326 segvn_free(struct seg *seg)
2327 {
2328 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2329 	pgcnt_t npages = seg_pages(seg);
2330 	struct anon_map *amp;
2331 	size_t len;
2332 
2333 	/*
2334 	 * We don't need any segment level locks for "segvn" data
2335 	 * since the address space is "write" locked.
2336 	 */
2337 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
2338 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
2339 
2340 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2341 
2342 	/*
2343 	 * Be sure to unlock pages. XXX Why do things get free'ed instead
2344 	 * of unmapped? XXX
2345 	 */
2346 	(void) segvn_lockop(seg, seg->s_base, seg->s_size,
2347 	    0, MC_UNLOCK, NULL, 0);
2348 
2349 	/*
2350 	 * Deallocate the vpage and anon pointers if necessary and possible.
2351 	 */
2352 	if (svd->vpage != NULL) {
2353 		kmem_free(svd->vpage, vpgtob(npages));
2354 		svd->vpage = NULL;
2355 	}
2356 	if ((amp = svd->amp) != NULL) {
2357 		/*
2358 		 * If there are no more references to this anon_map
2359 		 * structure, then deallocate the structure after freeing
2360 		 * up all the anon slot pointers that we can.
2361 		 */
2362 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2363 		ASSERT(amp->a_szc >= seg->s_szc);
2364 		if (--amp->refcnt == 0) {
2365 			if (svd->type == MAP_PRIVATE) {
2366 				/*
2367 				 * Private - we only need to anon_free
2368 				 * the part that this segment refers to.
2369 				 */
2370 				if (seg->s_szc != 0) {
2371 					anon_free_pages(amp->ahp,
2372 					    svd->anon_index, seg->s_size,
2373 					    seg->s_szc);
2374 				} else {
2375 					anon_free(amp->ahp, svd->anon_index,
2376 					    seg->s_size);
2377 				}
2378 			} else {
2379 
2380 				/*
2381 				 * Shared anon map is no longer in use. Before
2382 				 * freeing its pages purge all entries from
2383 				 * pcache that belong to this amp.
2384 				 */
2385 				ASSERT(svd->softlockcnt == 0);
2386 				anonmap_purge(amp);
2387 
2388 				/*
2389 				 * Shared - anon_free the entire
2390 				 * anon_map's worth of stuff and
2391 				 * release any swap reservation.
2392 				 */
2393 				if (amp->a_szc != 0) {
2394 					anon_shmap_free_pages(amp, 0,
2395 					    amp->size);
2396 				} else {
2397 					anon_free(amp->ahp, 0, amp->size);
2398 				}
2399 				if ((len = amp->swresv) != 0) {
2400 					anon_unresv(len);
2401 					TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2402 					    "anon proc:%p %lu %u", seg, len, 0);
2403 				}
2404 			}
2405 			svd->amp = NULL;
2406 			ANON_LOCK_EXIT(&amp->a_rwlock);
2407 			anonmap_free(amp);
2408 		} else if (svd->type == MAP_PRIVATE) {
2409 			/*
2410 			 * We had a private mapping which still has
2411 			 * a held anon_map so just free up all the
2412 			 * anon slot pointers that we were using.
2413 			 */
2414 			if (seg->s_szc != 0) {
2415 				anon_free_pages(amp->ahp, svd->anon_index,
2416 				    seg->s_size, seg->s_szc);
2417 			} else {
2418 				anon_free(amp->ahp, svd->anon_index,
2419 				    seg->s_size);
2420 			}
2421 			ANON_LOCK_EXIT(&amp->a_rwlock);
2422 		} else {
2423 			ANON_LOCK_EXIT(&amp->a_rwlock);
2424 		}
2425 	}
2426 
2427 	/*
2428 	 * Release swap reservation.
2429 	 */
2430 	if ((len = svd->swresv) != 0) {
2431 		anon_unresv(svd->swresv);
2432 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2433 		    seg, len, 0);
2434 		svd->swresv = 0;
2435 	}
2436 	/*
2437 	 * Release claim on vnode, credentials, and finally free the
2438 	 * private data.
2439 	 */
2440 	if (svd->vp != NULL) {
2441 		if (svd->type == MAP_SHARED)
2442 			lgrp_shm_policy_fini(NULL, svd->vp);
2443 		VN_RELE(svd->vp);
2444 		svd->vp = NULL;
2445 	}
2446 	crfree(svd->cred);
2447 	svd->pageprot = 0;
2448 	svd->pageadvice = 0;
2449 	svd->pageswap = 0;
2450 	svd->cred = NULL;
2451 
2452 	/*
2453 	 * Take segfree_syncmtx lock to let segvn_reclaim() finish if it's
2454 	 * still working with this segment without holding as lock (in case
2455 	 * it's called by pcache async thread).
2456 	 */
2457 	ASSERT(svd->softlockcnt == 0);
2458 	mutex_enter(&svd->segfree_syncmtx);
2459 	mutex_exit(&svd->segfree_syncmtx);
2460 
2461 	seg->s_data = NULL;
2462 	kmem_cache_free(segvn_cache, svd);
2463 }
2464 
2465 /*
2466  * Do a F_SOFTUNLOCK call over the range requested.  The range must have
2467  * already been F_SOFTLOCK'ed.
2468  * Caller must always match addr and len of a softunlock with a previous
2469  * softlock with exactly the same addr and len.
2470  */
2471 static void
2472 segvn_softunlock(struct seg *seg, caddr_t addr, size_t len, enum seg_rw rw)
2473 {
2474 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2475 	page_t *pp;
2476 	caddr_t adr;
2477 	struct vnode *vp;
2478 	u_offset_t offset;
2479 	ulong_t anon_index;
2480 	struct anon_map *amp;
2481 	struct anon *ap = NULL;
2482 
2483 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
2484 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
2485 
2486 	if ((amp = svd->amp) != NULL)
2487 		anon_index = svd->anon_index + seg_page(seg, addr);
2488 
2489 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
2490 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2491 		hat_unlock_region(seg->s_as->a_hat, addr, len, svd->rcookie);
2492 	} else {
2493 		hat_unlock(seg->s_as->a_hat, addr, len);
2494 	}
2495 	for (adr = addr; adr < addr + len; adr += PAGESIZE) {
2496 		if (amp != NULL) {
2497 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2498 			if ((ap = anon_get_ptr(amp->ahp, anon_index++))
2499 			    != NULL) {
2500 				swap_xlate(ap, &vp, &offset);
2501 			} else {
2502 				vp = svd->vp;
2503 				offset = svd->offset +
2504 				    (uintptr_t)(adr - seg->s_base);
2505 			}
2506 			ANON_LOCK_EXIT(&amp->a_rwlock);
2507 		} else {
2508 			vp = svd->vp;
2509 			offset = svd->offset +
2510 			    (uintptr_t)(adr - seg->s_base);
2511 		}
2512 
2513 		/*
2514 		 * Use page_find() instead of page_lookup() to
2515 		 * find the page since we know that it is locked.
2516 		 */
2517 		pp = page_find(vp, offset);
2518 		if (pp == NULL) {
2519 			panic(
2520 			    "segvn_softunlock: addr %p, ap %p, vp %p, off %llx",
2521 			    (void *)adr, (void *)ap, (void *)vp, offset);
2522 			/*NOTREACHED*/
2523 		}
2524 
2525 		if (rw == S_WRITE) {
2526 			hat_setrefmod(pp);
2527 			if (seg->s_as->a_vbits)
2528 				hat_setstat(seg->s_as, adr, PAGESIZE,
2529 				    P_REF | P_MOD);
2530 		} else if (rw != S_OTHER) {
2531 			hat_setref(pp);
2532 			if (seg->s_as->a_vbits)
2533 				hat_setstat(seg->s_as, adr, PAGESIZE, P_REF);
2534 		}
2535 		TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2536 		    "segvn_fault:pp %p vp %p offset %llx", pp, vp, offset);
2537 		page_unlock(pp);
2538 	}
2539 	ASSERT(svd->softlockcnt >= btop(len));
2540 	if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -btop(len))) {
2541 		/*
2542 		 * All SOFTLOCKS are gone. Wakeup any waiting
2543 		 * unmappers so they can try again to unmap.
2544 		 * Check for waiters first without the mutex
2545 		 * held so we don't always grab the mutex on
2546 		 * softunlocks.
2547 		 */
2548 		if (AS_ISUNMAPWAIT(seg->s_as)) {
2549 			mutex_enter(&seg->s_as->a_contents);
2550 			if (AS_ISUNMAPWAIT(seg->s_as)) {
2551 				AS_CLRUNMAPWAIT(seg->s_as);
2552 				cv_broadcast(&seg->s_as->a_cv);
2553 			}
2554 			mutex_exit(&seg->s_as->a_contents);
2555 		}
2556 	}
2557 }
2558 
2559 #define	PAGE_HANDLED	((page_t *)-1)
2560 
2561 /*
2562  * Release all the pages in the NULL terminated ppp list
2563  * which haven't already been converted to PAGE_HANDLED.
2564  */
2565 static void
2566 segvn_pagelist_rele(page_t **ppp)
2567 {
2568 	for (; *ppp != NULL; ppp++) {
2569 		if (*ppp != PAGE_HANDLED)
2570 			page_unlock(*ppp);
2571 	}
2572 }
2573 
2574 static int stealcow = 1;
2575 
2576 /*
2577  * Workaround for viking chip bug.  See bug id 1220902.
2578  * To fix this down in pagefault() would require importing so
2579  * much as and segvn code as to be unmaintainable.
2580  */
2581 int enable_mbit_wa = 0;
2582 
2583 /*
2584  * Handles all the dirty work of getting the right
2585  * anonymous pages and loading up the translations.
2586  * This routine is called only from segvn_fault()
2587  * when looping over the range of addresses requested.
2588  *
2589  * The basic algorithm here is:
2590  * 	If this is an anon_zero case
2591  *		Call anon_zero to allocate page
2592  *		Load up translation
2593  *		Return
2594  *	endif
2595  *	If this is an anon page
2596  *		Use anon_getpage to get the page
2597  *	else
2598  *		Find page in pl[] list passed in
2599  *	endif
2600  *	If not a cow
2601  *		Load up the translation to the page
2602  *		return
2603  *	endif
2604  *	Call anon_private to handle cow
2605  *	Load up (writable) translation to new page
2606  */
2607 static faultcode_t
2608 segvn_faultpage(
2609 	struct hat *hat,		/* the hat to use for mapping */
2610 	struct seg *seg,		/* seg_vn of interest */
2611 	caddr_t addr,			/* address in as */
2612 	u_offset_t off,			/* offset in vp */
2613 	struct vpage *vpage,		/* pointer to vpage for vp, off */
2614 	page_t *pl[],			/* object source page pointer */
2615 	uint_t vpprot,			/* access allowed to object pages */
2616 	enum fault_type type,		/* type of fault */
2617 	enum seg_rw rw,			/* type of access at fault */
2618 	int brkcow)			/* we may need to break cow */
2619 {
2620 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2621 	page_t *pp, **ppp;
2622 	uint_t pageflags = 0;
2623 	page_t *anon_pl[1 + 1];
2624 	page_t *opp = NULL;		/* original page */
2625 	uint_t prot;
2626 	int err;
2627 	int cow;
2628 	int claim;
2629 	int steal = 0;
2630 	ulong_t anon_index;
2631 	struct anon *ap, *oldap;
2632 	struct anon_map *amp;
2633 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
2634 	int anon_lock = 0;
2635 	anon_sync_obj_t cookie;
2636 
2637 	if (svd->flags & MAP_TEXT) {
2638 		hat_flag |= HAT_LOAD_TEXT;
2639 	}
2640 
2641 	ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
2642 	ASSERT(seg->s_szc == 0);
2643 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
2644 
2645 	/*
2646 	 * Initialize protection value for this page.
2647 	 * If we have per page protection values check it now.
2648 	 */
2649 	if (svd->pageprot) {
2650 		uint_t protchk;
2651 
2652 		switch (rw) {
2653 		case S_READ:
2654 			protchk = PROT_READ;
2655 			break;
2656 		case S_WRITE:
2657 			protchk = PROT_WRITE;
2658 			break;
2659 		case S_EXEC:
2660 			protchk = PROT_EXEC;
2661 			break;
2662 		case S_OTHER:
2663 		default:
2664 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
2665 			break;
2666 		}
2667 
2668 		prot = VPP_PROT(vpage);
2669 		if ((prot & protchk) == 0)
2670 			return (FC_PROT);	/* illegal access type */
2671 	} else {
2672 		prot = svd->prot;
2673 	}
2674 
2675 	if (type == F_SOFTLOCK) {
2676 		atomic_add_long((ulong_t *)&svd->softlockcnt, 1);
2677 	}
2678 
2679 	/*
2680 	 * Always acquire the anon array lock to prevent 2 threads from
2681 	 * allocating separate anon slots for the same "addr".
2682 	 */
2683 
2684 	if ((amp = svd->amp) != NULL) {
2685 		ASSERT(RW_READ_HELD(&amp->a_rwlock));
2686 		anon_index = svd->anon_index + seg_page(seg, addr);
2687 		anon_array_enter(amp, anon_index, &cookie);
2688 		anon_lock = 1;
2689 	}
2690 
2691 	if (svd->vp == NULL && amp != NULL) {
2692 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL) {
2693 			/*
2694 			 * Allocate a (normally) writable anonymous page of
2695 			 * zeroes. If no advance reservations, reserve now.
2696 			 */
2697 			if (svd->flags & MAP_NORESERVE) {
2698 				if (anon_resv_zone(ptob(1),
2699 				    seg->s_as->a_proc->p_zone)) {
2700 					atomic_add_long(&svd->swresv, ptob(1));
2701 				} else {
2702 					err = ENOMEM;
2703 					goto out;
2704 				}
2705 			}
2706 			if ((pp = anon_zero(seg, addr, &ap,
2707 			    svd->cred)) == NULL) {
2708 				err = ENOMEM;
2709 				goto out;	/* out of swap space */
2710 			}
2711 			/*
2712 			 * Re-acquire the anon_map lock and
2713 			 * initialize the anon array entry.
2714 			 */
2715 			(void) anon_set_ptr(amp->ahp, anon_index, ap,
2716 			    ANON_SLEEP);
2717 
2718 			ASSERT(pp->p_szc == 0);
2719 
2720 			/*
2721 			 * Handle pages that have been marked for migration
2722 			 */
2723 			if (lgrp_optimizations())
2724 				page_migrate(seg, addr, &pp, 1);
2725 
2726 			if (enable_mbit_wa) {
2727 				if (rw == S_WRITE)
2728 					hat_setmod(pp);
2729 				else if (!hat_ismod(pp))
2730 					prot &= ~PROT_WRITE;
2731 			}
2732 			/*
2733 			 * If AS_PAGLCK is set in a_flags (via memcntl(2)
2734 			 * with MC_LOCKAS, MCL_FUTURE) and this is a
2735 			 * MAP_NORESERVE segment, we may need to
2736 			 * permanently lock the page as it is being faulted
2737 			 * for the first time. The following text applies
2738 			 * only to MAP_NORESERVE segments:
2739 			 *
2740 			 * As per memcntl(2), if this segment was created
2741 			 * after MCL_FUTURE was applied (a "future"
2742 			 * segment), its pages must be locked.  If this
2743 			 * segment existed at MCL_FUTURE application (a
2744 			 * "past" segment), the interface is unclear.
2745 			 *
2746 			 * We decide to lock only if vpage is present:
2747 			 *
2748 			 * - "future" segments will have a vpage array (see
2749 			 *    as_map), and so will be locked as required
2750 			 *
2751 			 * - "past" segments may not have a vpage array,
2752 			 *    depending on whether events (such as
2753 			 *    mprotect) have occurred. Locking if vpage
2754 			 *    exists will preserve legacy behavior.  Not
2755 			 *    locking if vpage is absent, will not break
2756 			 *    the interface or legacy behavior.  Note that
2757 			 *    allocating vpage here if it's absent requires
2758 			 *    upgrading the segvn reader lock, the cost of
2759 			 *    which does not seem worthwhile.
2760 			 *
2761 			 * Usually testing and setting VPP_ISPPLOCK and
2762 			 * VPP_SETPPLOCK requires holding the segvn lock as
2763 			 * writer, but in this case all readers are
2764 			 * serializing on the anon array lock.
2765 			 */
2766 			if (AS_ISPGLCK(seg->s_as) && vpage != NULL &&
2767 			    (svd->flags & MAP_NORESERVE) &&
2768 			    !VPP_ISPPLOCK(vpage)) {
2769 				proc_t *p = seg->s_as->a_proc;
2770 				ASSERT(svd->type == MAP_PRIVATE);
2771 				mutex_enter(&p->p_lock);
2772 				if (rctl_incr_locked_mem(p, NULL, PAGESIZE,
2773 				    1) == 0) {
2774 					claim = VPP_PROT(vpage) & PROT_WRITE;
2775 					if (page_pp_lock(pp, claim, 0)) {
2776 						VPP_SETPPLOCK(vpage);
2777 					} else {
2778 						rctl_decr_locked_mem(p, NULL,
2779 						    PAGESIZE, 1);
2780 					}
2781 				}
2782 				mutex_exit(&p->p_lock);
2783 			}
2784 
2785 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2786 			hat_memload(hat, addr, pp, prot, hat_flag);
2787 
2788 			if (!(hat_flag & HAT_LOAD_LOCK))
2789 				page_unlock(pp);
2790 
2791 			anon_array_exit(&cookie);
2792 			return (0);
2793 		}
2794 	}
2795 
2796 	/*
2797 	 * Obtain the page structure via anon_getpage() if it is
2798 	 * a private copy of an object (the result of a previous
2799 	 * copy-on-write).
2800 	 */
2801 	if (amp != NULL) {
2802 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) != NULL) {
2803 			err = anon_getpage(&ap, &vpprot, anon_pl, PAGESIZE,
2804 			    seg, addr, rw, svd->cred);
2805 			if (err)
2806 				goto out;
2807 
2808 			if (svd->type == MAP_SHARED) {
2809 				/*
2810 				 * If this is a shared mapping to an
2811 				 * anon_map, then ignore the write
2812 				 * permissions returned by anon_getpage().
2813 				 * They apply to the private mappings
2814 				 * of this anon_map.
2815 				 */
2816 				vpprot |= PROT_WRITE;
2817 			}
2818 			opp = anon_pl[0];
2819 		}
2820 	}
2821 
2822 	/*
2823 	 * Search the pl[] list passed in if it is from the
2824 	 * original object (i.e., not a private copy).
2825 	 */
2826 	if (opp == NULL) {
2827 		/*
2828 		 * Find original page.  We must be bringing it in
2829 		 * from the list in pl[].
2830 		 */
2831 		for (ppp = pl; (opp = *ppp) != NULL; ppp++) {
2832 			if (opp == PAGE_HANDLED)
2833 				continue;
2834 			ASSERT(opp->p_vnode == svd->vp); /* XXX */
2835 			if (opp->p_offset == off)
2836 				break;
2837 		}
2838 		if (opp == NULL) {
2839 			panic("segvn_faultpage not found");
2840 			/*NOTREACHED*/
2841 		}
2842 		*ppp = PAGE_HANDLED;
2843 
2844 	}
2845 
2846 	ASSERT(PAGE_LOCKED(opp));
2847 
2848 	TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2849 	    "segvn_fault:pp %p vp %p offset %llx", opp, NULL, 0);
2850 
2851 	/*
2852 	 * The fault is treated as a copy-on-write fault if a
2853 	 * write occurs on a private segment and the object
2854 	 * page (i.e., mapping) is write protected.  We assume
2855 	 * that fatal protection checks have already been made.
2856 	 */
2857 
2858 	if (brkcow) {
2859 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2860 		cow = !(vpprot & PROT_WRITE);
2861 	} else if (svd->tr_state == SEGVN_TR_ON) {
2862 		/*
2863 		 * If we are doing text replication COW on first touch.
2864 		 */
2865 		ASSERT(amp != NULL);
2866 		ASSERT(svd->vp != NULL);
2867 		ASSERT(rw != S_WRITE);
2868 		cow = (ap == NULL);
2869 	} else {
2870 		cow = 0;
2871 	}
2872 
2873 	/*
2874 	 * If not a copy-on-write case load the translation
2875 	 * and return.
2876 	 */
2877 	if (cow == 0) {
2878 
2879 		/*
2880 		 * Handle pages that have been marked for migration
2881 		 */
2882 		if (lgrp_optimizations())
2883 			page_migrate(seg, addr, &opp, 1);
2884 
2885 		if (IS_VMODSORT(opp->p_vnode) || enable_mbit_wa) {
2886 			if (rw == S_WRITE)
2887 				hat_setmod(opp);
2888 			else if (rw != S_OTHER && !hat_ismod(opp))
2889 				prot &= ~PROT_WRITE;
2890 		}
2891 
2892 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
2893 		    (!svd->pageprot && svd->prot == (prot & vpprot)));
2894 		ASSERT(amp == NULL ||
2895 		    svd->rcookie == HAT_INVALID_REGION_COOKIE);
2896 		hat_memload_region(hat, addr, opp, prot & vpprot, hat_flag,
2897 		    svd->rcookie);
2898 
2899 		if (!(hat_flag & HAT_LOAD_LOCK))
2900 			page_unlock(opp);
2901 
2902 		if (anon_lock) {
2903 			anon_array_exit(&cookie);
2904 		}
2905 		return (0);
2906 	}
2907 
2908 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2909 
2910 	hat_setref(opp);
2911 
2912 	ASSERT(amp != NULL && anon_lock);
2913 
2914 	/*
2915 	 * Steal the page only if it isn't a private page
2916 	 * since stealing a private page is not worth the effort.
2917 	 */
2918 	if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL)
2919 		steal = 1;
2920 
2921 	/*
2922 	 * Steal the original page if the following conditions are true:
2923 	 *
2924 	 * We are low on memory, the page is not private, page is not large,
2925 	 * not shared, not modified, not `locked' or if we have it `locked'
2926 	 * (i.e., p_cowcnt == 1 and p_lckcnt == 0, which also implies
2927 	 * that the page is not shared) and if it doesn't have any
2928 	 * translations. page_struct_lock isn't needed to look at p_cowcnt
2929 	 * and p_lckcnt because we first get exclusive lock on page.
2930 	 */
2931 	(void) hat_pagesync(opp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD);
2932 
2933 	if (stealcow && freemem < minfree && steal && opp->p_szc == 0 &&
2934 	    page_tryupgrade(opp) && !hat_ismod(opp) &&
2935 	    ((opp->p_lckcnt == 0 && opp->p_cowcnt == 0) ||
2936 	    (opp->p_lckcnt == 0 && opp->p_cowcnt == 1 &&
2937 	    vpage != NULL && VPP_ISPPLOCK(vpage)))) {
2938 		/*
2939 		 * Check if this page has other translations
2940 		 * after unloading our translation.
2941 		 */
2942 		if (hat_page_is_mapped(opp)) {
2943 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2944 			hat_unload(seg->s_as->a_hat, addr, PAGESIZE,
2945 			    HAT_UNLOAD);
2946 		}
2947 
2948 		/*
2949 		 * hat_unload() might sync back someone else's recent
2950 		 * modification, so check again.
2951 		 */
2952 		if (!hat_ismod(opp) && !hat_page_is_mapped(opp))
2953 			pageflags |= STEAL_PAGE;
2954 	}
2955 
2956 	/*
2957 	 * If we have a vpage pointer, see if it indicates that we have
2958 	 * ``locked'' the page we map -- if so, tell anon_private to
2959 	 * transfer the locking resource to the new page.
2960 	 *
2961 	 * See Statement at the beginning of segvn_lockop regarding
2962 	 * the way lockcnts/cowcnts are handled during COW.
2963 	 *
2964 	 */
2965 	if (vpage != NULL && VPP_ISPPLOCK(vpage))
2966 		pageflags |= LOCK_PAGE;
2967 
2968 	/*
2969 	 * Allocate a private page and perform the copy.
2970 	 * For MAP_NORESERVE reserve swap space now, unless this
2971 	 * is a cow fault on an existing anon page in which case
2972 	 * MAP_NORESERVE will have made advance reservations.
2973 	 */
2974 	if ((svd->flags & MAP_NORESERVE) && (ap == NULL)) {
2975 		if (anon_resv_zone(ptob(1), seg->s_as->a_proc->p_zone)) {
2976 			atomic_add_long(&svd->swresv, ptob(1));
2977 		} else {
2978 			page_unlock(opp);
2979 			err = ENOMEM;
2980 			goto out;
2981 		}
2982 	}
2983 	oldap = ap;
2984 	pp = anon_private(&ap, seg, addr, prot, opp, pageflags, svd->cred);
2985 	if (pp == NULL) {
2986 		err = ENOMEM;	/* out of swap space */
2987 		goto out;
2988 	}
2989 
2990 	/*
2991 	 * If we copied away from an anonymous page, then
2992 	 * we are one step closer to freeing up an anon slot.
2993 	 *
2994 	 * NOTE:  The original anon slot must be released while
2995 	 * holding the "anon_map" lock.  This is necessary to prevent
2996 	 * other threads from obtaining a pointer to the anon slot
2997 	 * which may be freed if its "refcnt" is 1.
2998 	 */
2999 	if (oldap != NULL)
3000 		anon_decref(oldap);
3001 
3002 	(void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
3003 
3004 	/*
3005 	 * Handle pages that have been marked for migration
3006 	 */
3007 	if (lgrp_optimizations())
3008 		page_migrate(seg, addr, &pp, 1);
3009 
3010 	ASSERT(pp->p_szc == 0);
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) {
3034 		atomic_add_long((ulong_t *)&svd->softlockcnt, -1);
3035 	}
3036 	return (FC_MAKE_ERR(err));
3037 }
3038 
3039 /*
3040  * relocate a bunch of smaller targ pages into one large repl page. all targ
3041  * pages must be complete pages smaller than replacement pages.
3042  * it's assumed that no page's szc can change since they are all PAGESIZE or
3043  * complete large pages locked SHARED.
3044  */
3045 static void
3046 segvn_relocate_pages(page_t **targ, page_t *replacement)
3047 {
3048 	page_t *pp;
3049 	pgcnt_t repl_npgs, curnpgs;
3050 	pgcnt_t i;
3051 	uint_t repl_szc = replacement->p_szc;
3052 	page_t *first_repl = replacement;
3053 	page_t *repl;
3054 	spgcnt_t npgs;
3055 
3056 	VM_STAT_ADD(segvnvmstats.relocatepages[0]);
3057 
3058 	ASSERT(repl_szc != 0);
3059 	npgs = repl_npgs = page_get_pagecnt(repl_szc);
3060 
3061 	i = 0;
3062 	while (repl_npgs) {
3063 		spgcnt_t nreloc;
3064 		int err;
3065 		ASSERT(replacement != NULL);
3066 		pp = targ[i];
3067 		ASSERT(pp->p_szc < repl_szc);
3068 		ASSERT(PAGE_EXCL(pp));
3069 		ASSERT(!PP_ISFREE(pp));
3070 		curnpgs = page_get_pagecnt(pp->p_szc);
3071 		if (curnpgs == 1) {
3072 			VM_STAT_ADD(segvnvmstats.relocatepages[1]);
3073 			repl = replacement;
3074 			page_sub(&replacement, repl);
3075 			ASSERT(PAGE_EXCL(repl));
3076 			ASSERT(!PP_ISFREE(repl));
3077 			ASSERT(repl->p_szc == repl_szc);
3078 		} else {
3079 			page_t *repl_savepp;
3080 			int j;
3081 			VM_STAT_ADD(segvnvmstats.relocatepages[2]);
3082 			repl_savepp = replacement;
3083 			for (j = 0; j < curnpgs; j++) {
3084 				repl = replacement;
3085 				page_sub(&replacement, repl);
3086 				ASSERT(PAGE_EXCL(repl));
3087 				ASSERT(!PP_ISFREE(repl));
3088 				ASSERT(repl->p_szc == repl_szc);
3089 				ASSERT(page_pptonum(targ[i + j]) ==
3090 				    page_pptonum(targ[i]) + j);
3091 			}
3092 			repl = repl_savepp;
3093 			ASSERT(IS_P2ALIGNED(page_pptonum(repl), curnpgs));
3094 		}
3095 		err = page_relocate(&pp, &repl, 0, 1, &nreloc, NULL);
3096 		if (err || nreloc != curnpgs) {
3097 			panic("segvn_relocate_pages: "
3098 			    "page_relocate failed err=%d curnpgs=%ld "
3099 			    "nreloc=%ld", err, curnpgs, nreloc);
3100 		}
3101 		ASSERT(curnpgs <= repl_npgs);
3102 		repl_npgs -= curnpgs;
3103 		i += curnpgs;
3104 	}
3105 	ASSERT(replacement == NULL);
3106 
3107 	repl = first_repl;
3108 	repl_npgs = npgs;
3109 	for (i = 0; i < repl_npgs; i++) {
3110 		ASSERT(PAGE_EXCL(repl));
3111 		ASSERT(!PP_ISFREE(repl));
3112 		targ[i] = repl;
3113 		page_downgrade(targ[i]);
3114 		repl++;
3115 	}
3116 }
3117 
3118 /*
3119  * Check if all pages in ppa array are complete smaller than szc pages and
3120  * their roots will still be aligned relative to their current size if the
3121  * entire ppa array is relocated into one szc page. If these conditions are
3122  * not met return 0.
3123  *
3124  * If all pages are properly aligned attempt to upgrade their locks
3125  * to exclusive mode. If it fails set *upgrdfail to 1 and return 0.
3126  * upgrdfail was set to 0 by caller.
3127  *
3128  * Return 1 if all pages are aligned and locked exclusively.
3129  *
3130  * If all pages in ppa array happen to be physically contiguous to make one
3131  * szc page and all exclusive locks are successfully obtained promote the page
3132  * size to szc and set *pszc to szc. Return 1 with pages locked shared.
3133  */
3134 static int
3135 segvn_full_szcpages(page_t **ppa, uint_t szc, int *upgrdfail, uint_t *pszc)
3136 {
3137 	page_t *pp;
3138 	pfn_t pfn;
3139 	pgcnt_t totnpgs = page_get_pagecnt(szc);
3140 	pfn_t first_pfn;
3141 	int contig = 1;
3142 	pgcnt_t i;
3143 	pgcnt_t j;
3144 	uint_t curszc;
3145 	pgcnt_t curnpgs;
3146 	int root = 0;
3147 
3148 	ASSERT(szc > 0);
3149 
3150 	VM_STAT_ADD(segvnvmstats.fullszcpages[0]);
3151 
3152 	for (i = 0; i < totnpgs; i++) {
3153 		pp = ppa[i];
3154 		ASSERT(PAGE_SHARED(pp));
3155 		ASSERT(!PP_ISFREE(pp));
3156 		pfn = page_pptonum(pp);
3157 		if (i == 0) {
3158 			if (!IS_P2ALIGNED(pfn, totnpgs)) {
3159 				contig = 0;
3160 			} else {
3161 				first_pfn = pfn;
3162 			}
3163 		} else if (contig && pfn != first_pfn + i) {
3164 			contig = 0;
3165 		}
3166 		if (pp->p_szc == 0) {
3167 			if (root) {
3168 				VM_STAT_ADD(segvnvmstats.fullszcpages[1]);
3169 				return (0);
3170 			}
3171 		} else if (!root) {
3172 			if ((curszc = pp->p_szc) >= szc) {
3173 				VM_STAT_ADD(segvnvmstats.fullszcpages[2]);
3174 				return (0);
3175 			}
3176 			if (curszc == 0) {
3177 				/*
3178 				 * p_szc changed means we don't have all pages
3179 				 * locked. return failure.
3180 				 */
3181 				VM_STAT_ADD(segvnvmstats.fullszcpages[3]);
3182 				return (0);
3183 			}
3184 			curnpgs = page_get_pagecnt(curszc);
3185 			if (!IS_P2ALIGNED(pfn, curnpgs) ||
3186 			    !IS_P2ALIGNED(i, curnpgs)) {
3187 				VM_STAT_ADD(segvnvmstats.fullszcpages[4]);
3188 				return (0);
3189 			}
3190 			root = 1;
3191 		} else {
3192 			ASSERT(i > 0);
3193 			VM_STAT_ADD(segvnvmstats.fullszcpages[5]);
3194 			if (pp->p_szc != curszc) {
3195 				VM_STAT_ADD(segvnvmstats.fullszcpages[6]);
3196 				return (0);
3197 			}
3198 			if (pfn - 1 != page_pptonum(ppa[i - 1])) {
3199 				panic("segvn_full_szcpages: "
3200 				    "large page not physically contiguous");
3201 			}
3202 			if (P2PHASE(pfn, curnpgs) == curnpgs - 1) {
3203 				root = 0;
3204 			}
3205 		}
3206 	}
3207 
3208 	for (i = 0; i < totnpgs; i++) {
3209 		ASSERT(ppa[i]->p_szc < szc);
3210 		if (!page_tryupgrade(ppa[i])) {
3211 			for (j = 0; j < i; j++) {
3212 				page_downgrade(ppa[j]);
3213 			}
3214 			*pszc = ppa[i]->p_szc;
3215 			*upgrdfail = 1;
3216 			VM_STAT_ADD(segvnvmstats.fullszcpages[7]);
3217 			return (0);
3218 		}
3219 	}
3220 
3221 	/*
3222 	 * When a page is put a free cachelist its szc is set to 0.  if file
3223 	 * system reclaimed pages from cachelist targ pages will be physically
3224 	 * contiguous with 0 p_szc.  in this case just upgrade szc of targ
3225 	 * pages without any relocations.
3226 	 * To avoid any hat issues with previous small mappings
3227 	 * hat_pageunload() the target pages first.
3228 	 */
3229 	if (contig) {
3230 		VM_STAT_ADD(segvnvmstats.fullszcpages[8]);
3231 		for (i = 0; i < totnpgs; i++) {
3232 			(void) hat_pageunload(ppa[i], HAT_FORCE_PGUNLOAD);
3233 		}
3234 		for (i = 0; i < totnpgs; i++) {
3235 			ppa[i]->p_szc = szc;
3236 		}
3237 		for (i = 0; i < totnpgs; i++) {
3238 			ASSERT(PAGE_EXCL(ppa[i]));
3239 			page_downgrade(ppa[i]);
3240 		}
3241 		if (pszc != NULL) {
3242 			*pszc = szc;
3243 		}
3244 	}
3245 	VM_STAT_ADD(segvnvmstats.fullszcpages[9]);
3246 	return (1);
3247 }
3248 
3249 /*
3250  * Create physically contiguous pages for [vp, off] - [vp, off +
3251  * page_size(szc)) range and for private segment return them in ppa array.
3252  * Pages are created either via IO or relocations.
3253  *
3254  * Return 1 on success and 0 on failure.
3255  *
3256  * If physically contiguous pages already exist for this range return 1 without
3257  * filling ppa array. Caller initializes ppa[0] as NULL to detect that ppa
3258  * array wasn't filled. In this case caller fills ppa array via VOP_GETPAGE().
3259  */
3260 
3261 static int
3262 segvn_fill_vp_pages(struct segvn_data *svd, vnode_t *vp, u_offset_t off,
3263     uint_t szc, page_t **ppa, page_t **ppplist, uint_t *ret_pszc,
3264     int *downsize)
3265 
3266 {
3267 	page_t *pplist = *ppplist;
3268 	size_t pgsz = page_get_pagesize(szc);
3269 	pgcnt_t pages = btop(pgsz);
3270 	ulong_t start_off = off;
3271 	u_offset_t eoff = off + pgsz;
3272 	spgcnt_t nreloc;
3273 	u_offset_t io_off = off;
3274 	size_t io_len;
3275 	page_t *io_pplist = NULL;
3276 	page_t *done_pplist = NULL;
3277 	pgcnt_t pgidx = 0;
3278 	page_t *pp;
3279 	page_t *newpp;
3280 	page_t *targpp;
3281 	int io_err = 0;
3282 	int i;
3283 	pfn_t pfn;
3284 	ulong_t ppages;
3285 	page_t *targ_pplist = NULL;
3286 	page_t *repl_pplist = NULL;
3287 	page_t *tmp_pplist;
3288 	int nios = 0;
3289 	uint_t pszc;
3290 	struct vattr va;
3291 
3292 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[0]);
3293 
3294 	ASSERT(szc != 0);
3295 	ASSERT(pplist->p_szc == szc);
3296 
3297 	/*
3298 	 * downsize will be set to 1 only if we fail to lock pages. this will
3299 	 * allow subsequent faults to try to relocate the page again. If we
3300 	 * fail due to misalignment don't downsize and let the caller map the
3301 	 * whole region with small mappings to avoid more faults into the area
3302 	 * where we can't get large pages anyway.
3303 	 */
3304 	*downsize = 0;
3305 
3306 	while (off < eoff) {
3307 		newpp = pplist;
3308 		ASSERT(newpp != NULL);
3309 		ASSERT(PAGE_EXCL(newpp));
3310 		ASSERT(!PP_ISFREE(newpp));
3311 		/*
3312 		 * we pass NULL for nrelocp to page_lookup_create()
3313 		 * so that it doesn't relocate. We relocate here
3314 		 * later only after we make sure we can lock all
3315 		 * pages in the range we handle and they are all
3316 		 * aligned.
3317 		 */
3318 		pp = page_lookup_create(vp, off, SE_SHARED, newpp, NULL, 0);
3319 		ASSERT(pp != NULL);
3320 		ASSERT(!PP_ISFREE(pp));
3321 		ASSERT(pp->p_vnode == vp);
3322 		ASSERT(pp->p_offset == off);
3323 		if (pp == newpp) {
3324 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[1]);
3325 			page_sub(&pplist, pp);
3326 			ASSERT(PAGE_EXCL(pp));
3327 			ASSERT(page_iolock_assert(pp));
3328 			page_list_concat(&io_pplist, &pp);
3329 			off += PAGESIZE;
3330 			continue;
3331 		}
3332 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[2]);
3333 		pfn = page_pptonum(pp);
3334 		pszc = pp->p_szc;
3335 		if (pszc >= szc && targ_pplist == NULL && io_pplist == NULL &&
3336 		    IS_P2ALIGNED(pfn, pages)) {
3337 			ASSERT(repl_pplist == NULL);
3338 			ASSERT(done_pplist == NULL);
3339 			ASSERT(pplist == *ppplist);
3340 			page_unlock(pp);
3341 			page_free_replacement_page(pplist);
3342 			page_create_putback(pages);
3343 			*ppplist = NULL;
3344 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[3]);
3345 			return (1);
3346 		}
3347 		if (pszc >= szc) {
3348 			page_unlock(pp);
3349 			segvn_faultvnmpss_align_err1++;
3350 			goto out;
3351 		}
3352 		ppages = page_get_pagecnt(pszc);
3353 		if (!IS_P2ALIGNED(pfn, ppages)) {
3354 			ASSERT(pszc > 0);
3355 			/*
3356 			 * sizing down to pszc won't help.
3357 			 */
3358 			page_unlock(pp);
3359 			segvn_faultvnmpss_align_err2++;
3360 			goto out;
3361 		}
3362 		pfn = page_pptonum(newpp);
3363 		if (!IS_P2ALIGNED(pfn, ppages)) {
3364 			ASSERT(pszc > 0);
3365 			/*
3366 			 * sizing down to pszc won't help.
3367 			 */
3368 			page_unlock(pp);
3369 			segvn_faultvnmpss_align_err3++;
3370 			goto out;
3371 		}
3372 		if (!PAGE_EXCL(pp)) {
3373 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[4]);
3374 			page_unlock(pp);
3375 			*downsize = 1;
3376 			*ret_pszc = pp->p_szc;
3377 			goto out;
3378 		}
3379 		targpp = pp;
3380 		if (io_pplist != NULL) {
3381 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[5]);
3382 			io_len = off - io_off;
3383 			/*
3384 			 * Some file systems like NFS don't check EOF
3385 			 * conditions in VOP_PAGEIO(). Check it here
3386 			 * now that pages are locked SE_EXCL. Any file
3387 			 * truncation will wait until the pages are
3388 			 * unlocked so no need to worry that file will
3389 			 * be truncated after we check its size here.
3390 			 * XXX fix NFS to remove this check.
3391 			 */
3392 			va.va_mask = AT_SIZE;
3393 			if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL)) {
3394 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[6]);
3395 				page_unlock(targpp);
3396 				goto out;
3397 			}
3398 			if (btopr(va.va_size) < btopr(io_off + io_len)) {
3399 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[7]);
3400 				*downsize = 1;
3401 				*ret_pszc = 0;
3402 				page_unlock(targpp);
3403 				goto out;
3404 			}
3405 			io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3406 				B_READ, svd->cred, NULL);
3407 			if (io_err) {
3408 				VM_STAT_ADD(segvnvmstats.fill_vp_pages[8]);
3409 				page_unlock(targpp);
3410 				if (io_err == EDEADLK) {
3411 					segvn_vmpss_pageio_deadlk_err++;
3412 				}
3413 				goto out;
3414 			}
3415 			nios++;
3416 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[9]);
3417 			while (io_pplist != NULL) {
3418 				pp = io_pplist;
3419 				page_sub(&io_pplist, pp);
3420 				ASSERT(page_iolock_assert(pp));
3421 				page_io_unlock(pp);
3422 				pgidx = (pp->p_offset - start_off) >>
3423 				    PAGESHIFT;
3424 				ASSERT(pgidx < pages);
3425 				ppa[pgidx] = pp;
3426 				page_list_concat(&done_pplist, &pp);
3427 			}
3428 		}
3429 		pp = targpp;
3430 		ASSERT(PAGE_EXCL(pp));
3431 		ASSERT(pp->p_szc <= pszc);
3432 		if (pszc != 0 && !group_page_trylock(pp, SE_EXCL)) {
3433 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[10]);
3434 			page_unlock(pp);
3435 			*downsize = 1;
3436 			*ret_pszc = pp->p_szc;
3437 			goto out;
3438 		}
3439 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[11]);
3440 		/*
3441 		 * page szc chould have changed before the entire group was
3442 		 * locked. reread page szc.
3443 		 */
3444 		pszc = pp->p_szc;
3445 		ppages = page_get_pagecnt(pszc);
3446 
3447 		/* link just the roots */
3448 		page_list_concat(&targ_pplist, &pp);
3449 		page_sub(&pplist, newpp);
3450 		page_list_concat(&repl_pplist, &newpp);
3451 		off += PAGESIZE;
3452 		while (--ppages != 0) {
3453 			newpp = pplist;
3454 			page_sub(&pplist, newpp);
3455 			off += PAGESIZE;
3456 		}
3457 		io_off = off;
3458 	}
3459 	if (io_pplist != NULL) {
3460 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[12]);
3461 		io_len = eoff - io_off;
3462 		va.va_mask = AT_SIZE;
3463 		if (VOP_GETATTR(vp, &va, ATTR_HINT, svd->cred, NULL) != 0) {
3464 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[13]);
3465 			goto out;
3466 		}
3467 		if (btopr(va.va_size) < btopr(io_off + io_len)) {
3468 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[14]);
3469 			*downsize = 1;
3470 			*ret_pszc = 0;
3471 			goto out;
3472 		}
3473 		io_err = VOP_PAGEIO(vp, io_pplist, io_off, io_len,
3474 		    B_READ, svd->cred, NULL);
3475 		if (io_err) {
3476 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[15]);
3477 			if (io_err == EDEADLK) {
3478 				segvn_vmpss_pageio_deadlk_err++;
3479 			}
3480 			goto out;
3481 		}
3482 		nios++;
3483 		while (io_pplist != NULL) {
3484 			pp = io_pplist;
3485 			page_sub(&io_pplist, pp);
3486 			ASSERT(page_iolock_assert(pp));
3487 			page_io_unlock(pp);
3488 			pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3489 			ASSERT(pgidx < pages);
3490 			ppa[pgidx] = pp;
3491 		}
3492 	}
3493 	/*
3494 	 * we're now bound to succeed or panic.
3495 	 * remove pages from done_pplist. it's not needed anymore.
3496 	 */
3497 	while (done_pplist != NULL) {
3498 		pp = done_pplist;
3499 		page_sub(&done_pplist, pp);
3500 	}
3501 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[16]);
3502 	ASSERT(pplist == NULL);
3503 	*ppplist = NULL;
3504 	while (targ_pplist != NULL) {
3505 		int ret;
3506 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[17]);
3507 		ASSERT(repl_pplist);
3508 		pp = targ_pplist;
3509 		page_sub(&targ_pplist, pp);
3510 		pgidx = (pp->p_offset - start_off) >> PAGESHIFT;
3511 		newpp = repl_pplist;
3512 		page_sub(&repl_pplist, newpp);
3513 #ifdef DEBUG
3514 		pfn = page_pptonum(pp);
3515 		pszc = pp->p_szc;
3516 		ppages = page_get_pagecnt(pszc);
3517 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3518 		pfn = page_pptonum(newpp);
3519 		ASSERT(IS_P2ALIGNED(pfn, ppages));
3520 		ASSERT(P2PHASE(pfn, pages) == pgidx);
3521 #endif
3522 		nreloc = 0;
3523 		ret = page_relocate(&pp, &newpp, 0, 1, &nreloc, NULL);
3524 		if (ret != 0 || nreloc == 0) {
3525 			panic("segvn_fill_vp_pages: "
3526 			    "page_relocate failed");
3527 		}
3528 		pp = newpp;
3529 		while (nreloc-- != 0) {
3530 			ASSERT(PAGE_EXCL(pp));
3531 			ASSERT(pp->p_vnode == vp);
3532 			ASSERT(pgidx ==
3533 			    ((pp->p_offset - start_off) >> PAGESHIFT));
3534 			ppa[pgidx++] = pp;
3535 			pp++;
3536 		}
3537 	}
3538 
3539 	if (svd->type == MAP_PRIVATE) {
3540 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[18]);
3541 		for (i = 0; i < pages; i++) {
3542 			ASSERT(ppa[i] != NULL);
3543 			ASSERT(PAGE_EXCL(ppa[i]));
3544 			ASSERT(ppa[i]->p_vnode == vp);
3545 			ASSERT(ppa[i]->p_offset ==
3546 			    start_off + (i << PAGESHIFT));
3547 			page_downgrade(ppa[i]);
3548 		}
3549 		ppa[pages] = NULL;
3550 	} else {
3551 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[19]);
3552 		/*
3553 		 * the caller will still call VOP_GETPAGE() for shared segments
3554 		 * to check FS write permissions. For private segments we map
3555 		 * file read only anyway.  so no VOP_GETPAGE is needed.
3556 		 */
3557 		for (i = 0; i < pages; i++) {
3558 			ASSERT(ppa[i] != NULL);
3559 			ASSERT(PAGE_EXCL(ppa[i]));
3560 			ASSERT(ppa[i]->p_vnode == vp);
3561 			ASSERT(ppa[i]->p_offset ==
3562 			    start_off + (i << PAGESHIFT));
3563 			page_unlock(ppa[i]);
3564 		}
3565 		ppa[0] = NULL;
3566 	}
3567 
3568 	return (1);
3569 out:
3570 	/*
3571 	 * Do the cleanup. Unlock target pages we didn't relocate. They are
3572 	 * linked on targ_pplist by root pages. reassemble unused replacement
3573 	 * and io pages back to pplist.
3574 	 */
3575 	if (io_pplist != NULL) {
3576 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[20]);
3577 		pp = io_pplist;
3578 		do {
3579 			ASSERT(pp->p_vnode == vp);
3580 			ASSERT(pp->p_offset == io_off);
3581 			ASSERT(page_iolock_assert(pp));
3582 			page_io_unlock(pp);
3583 			page_hashout(pp, NULL);
3584 			io_off += PAGESIZE;
3585 		} while ((pp = pp->p_next) != io_pplist);
3586 		page_list_concat(&io_pplist, &pplist);
3587 		pplist = io_pplist;
3588 	}
3589 	tmp_pplist = NULL;
3590 	while (targ_pplist != NULL) {
3591 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[21]);
3592 		pp = targ_pplist;
3593 		ASSERT(PAGE_EXCL(pp));
3594 		page_sub(&targ_pplist, pp);
3595 
3596 		pszc = pp->p_szc;
3597 		ppages = page_get_pagecnt(pszc);
3598 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3599 
3600 		if (pszc != 0) {
3601 			group_page_unlock(pp);
3602 		}
3603 		page_unlock(pp);
3604 
3605 		pp = repl_pplist;
3606 		ASSERT(pp != NULL);
3607 		ASSERT(PAGE_EXCL(pp));
3608 		ASSERT(pp->p_szc == szc);
3609 		page_sub(&repl_pplist, pp);
3610 
3611 		ASSERT(IS_P2ALIGNED(page_pptonum(pp), ppages));
3612 
3613 		/* relink replacement page */
3614 		page_list_concat(&tmp_pplist, &pp);
3615 		while (--ppages != 0) {
3616 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[22]);
3617 			pp++;
3618 			ASSERT(PAGE_EXCL(pp));
3619 			ASSERT(pp->p_szc == szc);
3620 			page_list_concat(&tmp_pplist, &pp);
3621 		}
3622 	}
3623 	if (tmp_pplist != NULL) {
3624 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[23]);
3625 		page_list_concat(&tmp_pplist, &pplist);
3626 		pplist = tmp_pplist;
3627 	}
3628 	/*
3629 	 * at this point all pages are either on done_pplist or
3630 	 * pplist. They can't be all on done_pplist otherwise
3631 	 * we'd've been done.
3632 	 */
3633 	ASSERT(pplist != NULL);
3634 	if (nios != 0) {
3635 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[24]);
3636 		pp = pplist;
3637 		do {
3638 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[25]);
3639 			ASSERT(pp->p_szc == szc);
3640 			ASSERT(PAGE_EXCL(pp));
3641 			ASSERT(pp->p_vnode != vp);
3642 			pp->p_szc = 0;
3643 		} while ((pp = pp->p_next) != pplist);
3644 
3645 		pp = done_pplist;
3646 		do {
3647 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[26]);
3648 			ASSERT(pp->p_szc == szc);
3649 			ASSERT(PAGE_EXCL(pp));
3650 			ASSERT(pp->p_vnode == vp);
3651 			pp->p_szc = 0;
3652 		} while ((pp = pp->p_next) != done_pplist);
3653 
3654 		while (pplist != NULL) {
3655 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[27]);
3656 			pp = pplist;
3657 			page_sub(&pplist, pp);
3658 			page_free(pp, 0);
3659 		}
3660 
3661 		while (done_pplist != NULL) {
3662 			VM_STAT_ADD(segvnvmstats.fill_vp_pages[28]);
3663 			pp = done_pplist;
3664 			page_sub(&done_pplist, pp);
3665 			page_unlock(pp);
3666 		}
3667 		*ppplist = NULL;
3668 		return (0);
3669 	}
3670 	ASSERT(pplist == *ppplist);
3671 	if (io_err) {
3672 		VM_STAT_ADD(segvnvmstats.fill_vp_pages[29]);
3673 		/*
3674 		 * don't downsize on io error.
3675 		 * see if vop_getpage succeeds.
3676 		 * pplist may still be used in this case
3677 		 * for relocations.
3678 		 */
3679 		return (0);
3680 	}
3681 	VM_STAT_ADD(segvnvmstats.fill_vp_pages[30]);
3682 	page_free_replacement_page(pplist);
3683 	page_create_putback(pages);
3684 	*ppplist = NULL;
3685 	return (0);
3686 }
3687 
3688 int segvn_anypgsz = 0;
3689 
3690 #define	SEGVN_RESTORE_SOFTLOCK_VP(type, pages) 				\
3691 		if ((type) == F_SOFTLOCK) {				\
3692 			atomic_add_long((ulong_t *)&(svd)->softlockcnt, \
3693 			    -(pages));					\
3694 		}
3695 
3696 #define	SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot)		\
3697 		if (IS_VMODSORT((ppa)[0]->p_vnode)) {			\
3698 			if ((rw) == S_WRITE) {				\
3699 				for (i = 0; i < (pages); i++) {		\
3700 					ASSERT((ppa)[i]->p_vnode ==	\
3701 					    (ppa)[0]->p_vnode);		\
3702 					hat_setmod((ppa)[i]);		\
3703 				}					\
3704 			} else if ((rw) != S_OTHER &&			\
3705 			    ((prot) & (vpprot) & PROT_WRITE)) {		\
3706 				for (i = 0; i < (pages); i++) {		\
3707 					ASSERT((ppa)[i]->p_vnode ==	\
3708 					    (ppa)[0]->p_vnode);		\
3709 					if (!hat_ismod((ppa)[i])) {	\
3710 						prot &= ~PROT_WRITE;	\
3711 						break;			\
3712 					}				\
3713 				}					\
3714 			}						\
3715 		}
3716 
3717 #ifdef  VM_STATS
3718 
3719 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)					\
3720 		VM_STAT_ADD(segvnvmstats.fltvnpages[(idx)]);
3721 
3722 #else /* VM_STATS */
3723 
3724 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)
3725 
3726 #endif
3727 
3728 static faultcode_t
3729 segvn_fault_vnodepages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
3730     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
3731     caddr_t eaddr, int brkcow)
3732 {
3733 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
3734 	struct anon_map *amp = svd->amp;
3735 	uchar_t segtype = svd->type;
3736 	uint_t szc = seg->s_szc;
3737 	size_t pgsz = page_get_pagesize(szc);
3738 	size_t maxpgsz = pgsz;
3739 	pgcnt_t pages = btop(pgsz);
3740 	pgcnt_t maxpages = pages;
3741 	size_t ppasize = (pages + 1) * sizeof (page_t *);
3742 	caddr_t a = lpgaddr;
3743 	caddr_t	maxlpgeaddr = lpgeaddr;
3744 	u_offset_t off = svd->offset + (uintptr_t)(a - seg->s_base);
3745 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
3746 	struct vpage *vpage = (svd->vpage != NULL) ?
3747 	    &svd->vpage[seg_page(seg, a)] : NULL;
3748 	vnode_t *vp = svd->vp;
3749 	page_t **ppa;
3750 	uint_t	pszc;
3751 	size_t	ppgsz;
3752 	pgcnt_t	ppages;
3753 	faultcode_t err = 0;
3754 	int ierr;
3755 	int vop_size_err = 0;
3756 	uint_t protchk, prot, vpprot;
3757 	ulong_t i;
3758 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
3759 	anon_sync_obj_t an_cookie;
3760 	enum seg_rw arw;
3761 	int alloc_failed = 0;
3762 	int adjszc_chk;
3763 	struct vattr va;
3764 	int xhat = 0;
3765 	page_t *pplist;
3766 	pfn_t pfn;
3767 	int physcontig;
3768 	int upgrdfail;
3769 	int segvn_anypgsz_vnode = 0; /* for now map vnode with 2 page sizes */
3770 	int tron = (svd->tr_state == SEGVN_TR_ON);
3771 
3772 	ASSERT(szc != 0);
3773 	ASSERT(vp != NULL);
3774 	ASSERT(brkcow == 0 || amp != NULL);
3775 	ASSERT(tron == 0 || amp != NULL);
3776 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
3777 	ASSERT(!(svd->flags & MAP_NORESERVE));
3778 	ASSERT(type != F_SOFTUNLOCK);
3779 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
3780 	ASSERT(amp == NULL || IS_P2ALIGNED(aindx, maxpages));
3781 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
3782 	ASSERT(seg->s_szc < NBBY * sizeof (int));
3783 	ASSERT(type != F_SOFTLOCK || lpgeaddr - a == maxpgsz);
3784 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
3785 
3786 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltvnpages[0]);
3787 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltvnpages[1]);
3788 
3789 	if (svd->flags & MAP_TEXT) {
3790 		hat_flag |= HAT_LOAD_TEXT;
3791 	}
3792 
3793 	if (svd->pageprot) {
3794 		switch (rw) {
3795 		case S_READ:
3796 			protchk = PROT_READ;
3797 			break;
3798 		case S_WRITE:
3799 			protchk = PROT_WRITE;
3800 			break;
3801 		case S_EXEC:
3802 			protchk = PROT_EXEC;
3803 			break;
3804 		case S_OTHER:
3805 		default:
3806 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
3807 			break;
3808 		}
3809 	} else {
3810 		prot = svd->prot;
3811 		/* caller has already done segment level protection check. */
3812 	}
3813 
3814 	if (seg->s_as->a_hat != hat) {
3815 		xhat = 1;
3816 	}
3817 
3818 	if (rw == S_WRITE && segtype == MAP_PRIVATE) {
3819 		SEGVN_VMSTAT_FLTVNPAGES(2);
3820 		arw = S_READ;
3821 	} else {
3822 		arw = rw;
3823 	}
3824 
3825 	ppa = kmem_alloc(ppasize, KM_SLEEP);
3826 
3827 	VM_STAT_COND_ADD(amp != NULL, segvnvmstats.fltvnpages[3]);
3828 
3829 	for (;;) {
3830 		adjszc_chk = 0;
3831 		for (; a < lpgeaddr; a += pgsz, off += pgsz, aindx += pages) {
3832 			if (adjszc_chk) {
3833 				while (szc < seg->s_szc) {
3834 					uintptr_t e;
3835 					uint_t tszc;
3836 					tszc = segvn_anypgsz_vnode ? szc + 1 :
3837 					    seg->s_szc;
3838 					ppgsz = page_get_pagesize(tszc);
3839 					if (!IS_P2ALIGNED(a, ppgsz) ||
3840 					    ((alloc_failed >> tszc) & 0x1)) {
3841 						break;
3842 					}
3843 					SEGVN_VMSTAT_FLTVNPAGES(4);
3844 					szc = tszc;
3845 					pgsz = ppgsz;
3846 					pages = btop(pgsz);
3847 					e = P2ROUNDUP((uintptr_t)eaddr, pgsz);
3848 					lpgeaddr = (caddr_t)e;
3849 				}
3850 			}
3851 
3852 		again:
3853 			if (IS_P2ALIGNED(a, maxpgsz) && amp != NULL) {
3854 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
3855 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
3856 				anon_array_enter(amp, aindx, &an_cookie);
3857 				if (anon_get_ptr(amp->ahp, aindx) != NULL) {
3858 					SEGVN_VMSTAT_FLTVNPAGES(5);
3859 					ASSERT(anon_pages(amp->ahp, aindx,
3860 					    maxpages) == maxpages);
3861 					anon_array_exit(&an_cookie);
3862 					ANON_LOCK_EXIT(&amp->a_rwlock);
3863 					err = segvn_fault_anonpages(hat, seg,
3864 					    a, a + maxpgsz, type, rw,
3865 					    MAX(a, addr),
3866 					    MIN(a + maxpgsz, eaddr), brkcow);
3867 					if (err != 0) {
3868 						SEGVN_VMSTAT_FLTVNPAGES(6);
3869 						goto out;
3870 					}
3871 					if (szc < seg->s_szc) {
3872 						szc = seg->s_szc;
3873 						pgsz = maxpgsz;
3874 						pages = maxpages;
3875 						lpgeaddr = maxlpgeaddr;
3876 					}
3877 					goto next;
3878 				} else {
3879 					ASSERT(anon_pages(amp->ahp, aindx,
3880 					    maxpages) == 0);
3881 					SEGVN_VMSTAT_FLTVNPAGES(7);
3882 					anon_array_exit(&an_cookie);
3883 					ANON_LOCK_EXIT(&amp->a_rwlock);
3884 				}
3885 			}
3886 			ASSERT(!brkcow || IS_P2ALIGNED(a, maxpgsz));
3887 			ASSERT(!tron || IS_P2ALIGNED(a, maxpgsz));
3888 
3889 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
3890 				ASSERT(vpage != NULL);
3891 				prot = VPP_PROT(vpage);
3892 				ASSERT(sameprot(seg, a, maxpgsz));
3893 				if ((prot & protchk) == 0) {
3894 					SEGVN_VMSTAT_FLTVNPAGES(8);
3895 					err = FC_PROT;
3896 					goto out;
3897 				}
3898 			}
3899 			if (type == F_SOFTLOCK) {
3900 				atomic_add_long((ulong_t *)&svd->softlockcnt,
3901 				    pages);
3902 			}
3903 
3904 			pplist = NULL;
3905 			physcontig = 0;
3906 			ppa[0] = NULL;
3907 			if (!brkcow && !tron && szc &&
3908 			    !page_exists_physcontig(vp, off, szc,
3909 			    segtype == MAP_PRIVATE ? ppa : NULL)) {
3910 				SEGVN_VMSTAT_FLTVNPAGES(9);
3911 				if (page_alloc_pages(vp, seg, a, &pplist, NULL,
3912 				    szc, 0, 0) && type != F_SOFTLOCK) {
3913 					SEGVN_VMSTAT_FLTVNPAGES(10);
3914 					pszc = 0;
3915 					ierr = -1;
3916 					alloc_failed |= (1 << szc);
3917 					break;
3918 				}
3919 				if (pplist != NULL &&
3920 				    vp->v_mpssdata == SEGVN_PAGEIO) {
3921 					int downsize;
3922 					SEGVN_VMSTAT_FLTVNPAGES(11);
3923 					physcontig = segvn_fill_vp_pages(svd,
3924 					    vp, off, szc, ppa, &pplist,
3925 					    &pszc, &downsize);
3926 					ASSERT(!physcontig || pplist == NULL);
3927 					if (!physcontig && downsize &&
3928 					    type != F_SOFTLOCK) {
3929 						ASSERT(pplist == NULL);
3930 						SEGVN_VMSTAT_FLTVNPAGES(12);
3931 						ierr = -1;
3932 						break;
3933 					}
3934 					ASSERT(!physcontig ||
3935 					    segtype == MAP_PRIVATE ||
3936 					    ppa[0] == NULL);
3937 					if (physcontig && ppa[0] == NULL) {
3938 						physcontig = 0;
3939 					}
3940 				}
3941 			} else if (!brkcow && !tron && szc && ppa[0] != NULL) {
3942 				SEGVN_VMSTAT_FLTVNPAGES(13);
3943 				ASSERT(segtype == MAP_PRIVATE);
3944 				physcontig = 1;
3945 			}
3946 
3947 			if (!physcontig) {
3948 				SEGVN_VMSTAT_FLTVNPAGES(14);
3949 				ppa[0] = NULL;
3950 				ierr = VOP_GETPAGE(vp, (offset_t)off, pgsz,
3951 				    &vpprot, ppa, pgsz, seg, a, arw,
3952 				    svd->cred, NULL);
3953 #ifdef DEBUG
3954 				if (ierr == 0) {
3955 					for (i = 0; i < pages; i++) {
3956 						ASSERT(PAGE_LOCKED(ppa[i]));
3957 						ASSERT(!PP_ISFREE(ppa[i]));
3958 						ASSERT(ppa[i]->p_vnode == vp);
3959 						ASSERT(ppa[i]->p_offset ==
3960 						    off + (i << PAGESHIFT));
3961 					}
3962 				}
3963 #endif /* DEBUG */
3964 				if (segtype == MAP_PRIVATE) {
3965 					SEGVN_VMSTAT_FLTVNPAGES(15);
3966 					vpprot &= ~PROT_WRITE;
3967 				}
3968 			} else {
3969 				ASSERT(segtype == MAP_PRIVATE);
3970 				SEGVN_VMSTAT_FLTVNPAGES(16);
3971 				vpprot = PROT_ALL & ~PROT_WRITE;
3972 				ierr = 0;
3973 			}
3974 
3975 			if (ierr != 0) {
3976 				SEGVN_VMSTAT_FLTVNPAGES(17);
3977 				if (pplist != NULL) {
3978 					SEGVN_VMSTAT_FLTVNPAGES(18);
3979 					page_free_replacement_page(pplist);
3980 					page_create_putback(pages);
3981 				}
3982 				SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
3983 				if (a + pgsz <= eaddr) {
3984 					SEGVN_VMSTAT_FLTVNPAGES(19);
3985 					err = FC_MAKE_ERR(ierr);
3986 					goto out;
3987 				}
3988 				va.va_mask = AT_SIZE;
3989 				if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL)) {
3990 					SEGVN_VMSTAT_FLTVNPAGES(20);
3991 					err = FC_MAKE_ERR(EIO);
3992 					goto out;
3993 				}
3994 				if (btopr(va.va_size) >= btopr(off + pgsz)) {
3995 					SEGVN_VMSTAT_FLTVNPAGES(21);
3996 					err = FC_MAKE_ERR(ierr);
3997 					goto out;
3998 				}
3999 				if (btopr(va.va_size) <
4000 				    btopr(off + (eaddr - a))) {
4001 					SEGVN_VMSTAT_FLTVNPAGES(22);
4002 					err = FC_MAKE_ERR(ierr);
4003 					goto out;
4004 				}
4005 				if (brkcow || tron || type == F_SOFTLOCK) {
4006 					/* can't reduce map area */
4007 					SEGVN_VMSTAT_FLTVNPAGES(23);
4008 					vop_size_err = 1;
4009 					goto out;
4010 				}
4011 				SEGVN_VMSTAT_FLTVNPAGES(24);
4012 				ASSERT(szc != 0);
4013 				pszc = 0;
4014 				ierr = -1;
4015 				break;
4016 			}
4017 
4018 			if (amp != NULL) {
4019 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4020 				anon_array_enter(amp, aindx, &an_cookie);
4021 			}
4022 			if (amp != NULL &&
4023 			    anon_get_ptr(amp->ahp, aindx) != NULL) {
4024 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4025 
4026 				SEGVN_VMSTAT_FLTVNPAGES(25);
4027 				ASSERT(anon_pages(amp->ahp, taindx,
4028 				    maxpages) == maxpages);
4029 				for (i = 0; i < pages; i++) {
4030 					page_unlock(ppa[i]);
4031 				}
4032 				anon_array_exit(&an_cookie);
4033 				ANON_LOCK_EXIT(&amp->a_rwlock);
4034 				if (pplist != NULL) {
4035 					page_free_replacement_page(pplist);
4036 					page_create_putback(pages);
4037 				}
4038 				SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4039 				if (szc < seg->s_szc) {
4040 					SEGVN_VMSTAT_FLTVNPAGES(26);
4041 					/*
4042 					 * For private segments SOFTLOCK
4043 					 * either always breaks cow (any rw
4044 					 * type except S_READ_NOCOW) or
4045 					 * address space is locked as writer
4046 					 * (S_READ_NOCOW case) and anon slots
4047 					 * can't show up on second check.
4048 					 * Therefore if we are here for
4049 					 * SOFTLOCK case it must be a cow
4050 					 * break but cow break never reduces
4051 					 * szc. text replication (tron) in
4052 					 * this case works as cow break.
4053 					 * Thus the assert below.
4054 					 */
4055 					ASSERT(!brkcow && !tron &&
4056 					    type != F_SOFTLOCK);
4057 					pszc = seg->s_szc;
4058 					ierr = -2;
4059 					break;
4060 				}
4061 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4062 				goto again;
4063 			}
4064 #ifdef DEBUG
4065 			if (amp != NULL) {
4066 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4067 				ASSERT(!anon_pages(amp->ahp, taindx, maxpages));
4068 			}
4069 #endif /* DEBUG */
4070 
4071 			if (brkcow || tron) {
4072 				ASSERT(amp != NULL);
4073 				ASSERT(pplist == NULL);
4074 				ASSERT(szc == seg->s_szc);
4075 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4076 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
4077 				SEGVN_VMSTAT_FLTVNPAGES(27);
4078 				ierr = anon_map_privatepages(amp, aindx, szc,
4079 				    seg, a, prot, ppa, vpage, segvn_anypgsz,
4080 				    tron ? PG_LOCAL : 0, svd->cred);
4081 				if (ierr != 0) {
4082 					SEGVN_VMSTAT_FLTVNPAGES(28);
4083 					anon_array_exit(&an_cookie);
4084 					ANON_LOCK_EXIT(&amp->a_rwlock);
4085 					SEGVN_RESTORE_SOFTLOCK_VP(type, pages);
4086 					err = FC_MAKE_ERR(ierr);
4087 					goto out;
4088 				}
4089 
4090 				ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4091 				/*
4092 				 * p_szc can't be changed for locked
4093 				 * swapfs pages.
4094 				 */
4095 				ASSERT(svd->rcookie ==
4096 				    HAT_INVALID_REGION_COOKIE);
4097 				hat_memload_array(hat, a, pgsz, ppa, prot,
4098 				    hat_flag);
4099 
4100 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4101 					SEGVN_VMSTAT_FLTVNPAGES(29);
4102 					for (i = 0; i < pages; i++) {
4103 						page_unlock(ppa[i]);
4104 					}
4105 				}
4106 				anon_array_exit(&an_cookie);
4107 				ANON_LOCK_EXIT(&amp->a_rwlock);
4108 				goto next;
4109 			}
4110 
4111 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
4112 			    (!svd->pageprot && svd->prot == (prot & vpprot)));
4113 
4114 			pfn = page_pptonum(ppa[0]);
4115 			/*
4116 			 * hat_page_demote() needs an SE_EXCL lock on one of
4117 			 * constituent page_t's and it decreases root's p_szc
4118 			 * last. This means if root's p_szc is equal szc and
4119 			 * all its constituent pages are locked
4120 			 * hat_page_demote() that could have changed p_szc to
4121 			 * szc is already done and no new have page_demote()
4122 			 * can start for this large page.
4123 			 */
4124 
4125 			/*
4126 			 * we need to make sure same mapping size is used for
4127 			 * the same address range if there's a possibility the
4128 			 * adddress is already mapped because hat layer panics
4129 			 * when translation is loaded for the range already
4130 			 * mapped with a different page size.  We achieve it
4131 			 * by always using largest page size possible subject
4132 			 * to the constraints of page size, segment page size
4133 			 * and page alignment.  Since mappings are invalidated
4134 			 * when those constraints change and make it
4135 			 * impossible to use previously used mapping size no
4136 			 * mapping size conflicts should happen.
4137 			 */
4138 
4139 		chkszc:
4140 			if ((pszc = ppa[0]->p_szc) == szc &&
4141 			    IS_P2ALIGNED(pfn, pages)) {
4142 
4143 				SEGVN_VMSTAT_FLTVNPAGES(30);
4144 #ifdef DEBUG
4145 				for (i = 0; i < pages; i++) {
4146 					ASSERT(PAGE_LOCKED(ppa[i]));
4147 					ASSERT(!PP_ISFREE(ppa[i]));
4148 					ASSERT(page_pptonum(ppa[i]) ==
4149 					    pfn + i);
4150 					ASSERT(ppa[i]->p_szc == szc);
4151 					ASSERT(ppa[i]->p_vnode == vp);
4152 					ASSERT(ppa[i]->p_offset ==
4153 					    off + (i << PAGESHIFT));
4154 				}
4155 #endif /* DEBUG */
4156 				/*
4157 				 * All pages are of szc we need and they are
4158 				 * all locked so they can't change szc. load
4159 				 * translations.
4160 				 *
4161 				 * if page got promoted since last check
4162 				 * we don't need pplist.
4163 				 */
4164 				if (pplist != NULL) {
4165 					page_free_replacement_page(pplist);
4166 					page_create_putback(pages);
4167 				}
4168 				if (PP_ISMIGRATE(ppa[0])) {
4169 					page_migrate(seg, a, ppa, pages);
4170 				}
4171 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4172 				    prot, vpprot);
4173 				if (!xhat) {
4174 					hat_memload_array_region(hat, a, pgsz,
4175 					    ppa, prot & vpprot, hat_flag,
4176 					    svd->rcookie);
4177 				} else {
4178 					/*
4179 					 * avoid large xhat mappings to FS
4180 					 * pages so that hat_page_demote()
4181 					 * doesn't need to check for xhat
4182 					 * large mappings.
4183 					 * Don't use regions with xhats.
4184 					 */
4185 					for (i = 0; i < pages; i++) {
4186 						hat_memload(hat,
4187 						    a + (i << PAGESHIFT),
4188 						    ppa[i], prot & vpprot,
4189 						    hat_flag);
4190 					}
4191 				}
4192 
4193 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4194 					for (i = 0; i < pages; i++) {
4195 						page_unlock(ppa[i]);
4196 					}
4197 				}
4198 				if (amp != NULL) {
4199 					anon_array_exit(&an_cookie);
4200 					ANON_LOCK_EXIT(&amp->a_rwlock);
4201 				}
4202 				goto next;
4203 			}
4204 
4205 			/*
4206 			 * See if upsize is possible.
4207 			 */
4208 			if (pszc > szc && szc < seg->s_szc &&
4209 			    (segvn_anypgsz_vnode || pszc >= seg->s_szc)) {
4210 				pgcnt_t aphase;
4211 				uint_t pszc1 = MIN(pszc, seg->s_szc);
4212 				ppgsz = page_get_pagesize(pszc1);
4213 				ppages = btop(ppgsz);
4214 				aphase = btop(P2PHASE((uintptr_t)a, ppgsz));
4215 
4216 				ASSERT(type != F_SOFTLOCK);
4217 
4218 				SEGVN_VMSTAT_FLTVNPAGES(31);
4219 				if (aphase != P2PHASE(pfn, ppages)) {
4220 					segvn_faultvnmpss_align_err4++;
4221 				} else {
4222 					SEGVN_VMSTAT_FLTVNPAGES(32);
4223 					if (pplist != NULL) {
4224 						page_t *pl = pplist;
4225 						page_free_replacement_page(pl);
4226 						page_create_putback(pages);
4227 					}
4228 					for (i = 0; i < pages; i++) {
4229 						page_unlock(ppa[i]);
4230 					}
4231 					if (amp != NULL) {
4232 						anon_array_exit(&an_cookie);
4233 						ANON_LOCK_EXIT(&amp->a_rwlock);
4234 					}
4235 					pszc = pszc1;
4236 					ierr = -2;
4237 					break;
4238 				}
4239 			}
4240 
4241 			/*
4242 			 * check if we should use smallest mapping size.
4243 			 */
4244 			upgrdfail = 0;
4245 			if (szc == 0 || xhat ||
4246 			    (pszc >= szc &&
4247 			    !IS_P2ALIGNED(pfn, pages)) ||
4248 			    (pszc < szc &&
4249 			    !segvn_full_szcpages(ppa, szc, &upgrdfail,
4250 			    &pszc))) {
4251 
4252 				if (upgrdfail && type != F_SOFTLOCK) {
4253 					/*
4254 					 * segvn_full_szcpages failed to lock
4255 					 * all pages EXCL. Size down.
4256 					 */
4257 					ASSERT(pszc < szc);
4258 
4259 					SEGVN_VMSTAT_FLTVNPAGES(33);
4260 
4261 					if (pplist != NULL) {
4262 						page_t *pl = pplist;
4263 						page_free_replacement_page(pl);
4264 						page_create_putback(pages);
4265 					}
4266 
4267 					for (i = 0; i < pages; i++) {
4268 						page_unlock(ppa[i]);
4269 					}
4270 					if (amp != NULL) {
4271 						anon_array_exit(&an_cookie);
4272 						ANON_LOCK_EXIT(&amp->a_rwlock);
4273 					}
4274 					ierr = -1;
4275 					break;
4276 				}
4277 				if (szc != 0 && !xhat && !upgrdfail) {
4278 					segvn_faultvnmpss_align_err5++;
4279 				}
4280 				SEGVN_VMSTAT_FLTVNPAGES(34);
4281 				if (pplist != NULL) {
4282 					page_free_replacement_page(pplist);
4283 					page_create_putback(pages);
4284 				}
4285 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4286 				    prot, vpprot);
4287 				if (upgrdfail && segvn_anypgsz_vnode) {
4288 					/* SOFTLOCK case */
4289 					hat_memload_array_region(hat, a, pgsz,
4290 					    ppa, prot & vpprot, hat_flag,
4291 					    svd->rcookie);
4292 				} else {
4293 					for (i = 0; i < pages; i++) {
4294 						hat_memload_region(hat,
4295 						    a + (i << PAGESHIFT),
4296 						    ppa[i], prot & vpprot,
4297 						    hat_flag, svd->rcookie);
4298 					}
4299 				}
4300 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4301 					for (i = 0; i < pages; i++) {
4302 						page_unlock(ppa[i]);
4303 					}
4304 				}
4305 				if (amp != NULL) {
4306 					anon_array_exit(&an_cookie);
4307 					ANON_LOCK_EXIT(&amp->a_rwlock);
4308 				}
4309 				goto next;
4310 			}
4311 
4312 			if (pszc == szc) {
4313 				/*
4314 				 * segvn_full_szcpages() upgraded pages szc.
4315 				 */
4316 				ASSERT(pszc == ppa[0]->p_szc);
4317 				ASSERT(IS_P2ALIGNED(pfn, pages));
4318 				goto chkszc;
4319 			}
4320 
4321 			if (pszc > szc) {
4322 				kmutex_t *szcmtx;
4323 				SEGVN_VMSTAT_FLTVNPAGES(35);
4324 				/*
4325 				 * p_szc of ppa[0] can change since we haven't
4326 				 * locked all constituent pages. Call
4327 				 * page_lock_szc() to prevent szc changes.
4328 				 * This should be a rare case that happens when
4329 				 * multiple segments use a different page size
4330 				 * to map the same file offsets.
4331 				 */
4332 				szcmtx = page_szc_lock(ppa[0]);
4333 				pszc = ppa[0]->p_szc;
4334 				ASSERT(szcmtx != NULL || pszc == 0);
4335 				ASSERT(ppa[0]->p_szc <= pszc);
4336 				if (pszc <= szc) {
4337 					SEGVN_VMSTAT_FLTVNPAGES(36);
4338 					if (szcmtx != NULL) {
4339 						mutex_exit(szcmtx);
4340 					}
4341 					goto chkszc;
4342 				}
4343 				if (pplist != NULL) {
4344 					/*
4345 					 * page got promoted since last check.
4346 					 * we don't need preaalocated large
4347 					 * page.
4348 					 */
4349 					SEGVN_VMSTAT_FLTVNPAGES(37);
4350 					page_free_replacement_page(pplist);
4351 					page_create_putback(pages);
4352 				}
4353 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4354 				    prot, vpprot);
4355 				hat_memload_array_region(hat, a, pgsz, ppa,
4356 				    prot & vpprot, hat_flag, svd->rcookie);
4357 				mutex_exit(szcmtx);
4358 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4359 					for (i = 0; i < pages; i++) {
4360 						page_unlock(ppa[i]);
4361 					}
4362 				}
4363 				if (amp != NULL) {
4364 					anon_array_exit(&an_cookie);
4365 					ANON_LOCK_EXIT(&amp->a_rwlock);
4366 				}
4367 				goto next;
4368 			}
4369 
4370 			/*
4371 			 * if page got demoted since last check
4372 			 * we could have not allocated larger page.
4373 			 * allocate now.
4374 			 */
4375 			if (pplist == NULL &&
4376 			    page_alloc_pages(vp, seg, a, &pplist, NULL,
4377 			    szc, 0, 0) && type != F_SOFTLOCK) {
4378 				SEGVN_VMSTAT_FLTVNPAGES(38);
4379 				for (i = 0; i < pages; i++) {
4380 					page_unlock(ppa[i]);
4381 				}
4382 				if (amp != NULL) {
4383 					anon_array_exit(&an_cookie);
4384 					ANON_LOCK_EXIT(&amp->a_rwlock);
4385 				}
4386 				ierr = -1;
4387 				alloc_failed |= (1 << szc);
4388 				break;
4389 			}
4390 
4391 			SEGVN_VMSTAT_FLTVNPAGES(39);
4392 
4393 			if (pplist != NULL) {
4394 				segvn_relocate_pages(ppa, pplist);
4395 #ifdef DEBUG
4396 			} else {
4397 				ASSERT(type == F_SOFTLOCK);
4398 				SEGVN_VMSTAT_FLTVNPAGES(40);
4399 #endif /* DEBUG */
4400 			}
4401 
4402 			SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot);
4403 
4404 			if (pplist == NULL && segvn_anypgsz_vnode == 0) {
4405 				ASSERT(type == F_SOFTLOCK);
4406 				for (i = 0; i < pages; i++) {
4407 					ASSERT(ppa[i]->p_szc < szc);
4408 					hat_memload_region(hat,
4409 					    a + (i << PAGESHIFT),
4410 					    ppa[i], prot & vpprot, hat_flag,
4411 					    svd->rcookie);
4412 				}
4413 			} else {
4414 				ASSERT(pplist != NULL || type == F_SOFTLOCK);
4415 				hat_memload_array_region(hat, a, pgsz, ppa,
4416 				    prot & vpprot, hat_flag, svd->rcookie);
4417 			}
4418 			if (!(hat_flag & HAT_LOAD_LOCK)) {
4419 				for (i = 0; i < pages; i++) {
4420 					ASSERT(PAGE_SHARED(ppa[i]));
4421 					page_unlock(ppa[i]);
4422 				}
4423 			}
4424 			if (amp != NULL) {
4425 				anon_array_exit(&an_cookie);
4426 				ANON_LOCK_EXIT(&amp->a_rwlock);
4427 			}
4428 
4429 		next:
4430 			if (vpage != NULL) {
4431 				vpage += pages;
4432 			}
4433 			adjszc_chk = 1;
4434 		}
4435 		if (a == lpgeaddr)
4436 			break;
4437 		ASSERT(a < lpgeaddr);
4438 
4439 		ASSERT(!brkcow && !tron && type != F_SOFTLOCK);
4440 
4441 		/*
4442 		 * ierr == -1 means we failed to map with a large page.
4443 		 * (either due to allocation/relocation failures or
4444 		 * misalignment with other mappings to this file.
4445 		 *
4446 		 * ierr == -2 means some other thread allocated a large page
4447 		 * after we gave up tp map with a large page.  retry with
4448 		 * larger mapping.
4449 		 */
4450 		ASSERT(ierr == -1 || ierr == -2);
4451 		ASSERT(ierr == -2 || szc != 0);
4452 		ASSERT(ierr == -1 || szc < seg->s_szc);
4453 		if (ierr == -2) {
4454 			SEGVN_VMSTAT_FLTVNPAGES(41);
4455 			ASSERT(pszc > szc && pszc <= seg->s_szc);
4456 			szc = pszc;
4457 		} else if (segvn_anypgsz_vnode) {
4458 			SEGVN_VMSTAT_FLTVNPAGES(42);
4459 			szc--;
4460 		} else {
4461 			SEGVN_VMSTAT_FLTVNPAGES(43);
4462 			ASSERT(pszc < szc);
4463 			/*
4464 			 * other process created pszc large page.
4465 			 * but we still have to drop to 0 szc.
4466 			 */
4467 			szc = 0;
4468 		}
4469 
4470 		pgsz = page_get_pagesize(szc);
4471 		pages = btop(pgsz);
4472 		if (ierr == -2) {
4473 			/*
4474 			 * Size up case. Note lpgaddr may only be needed for
4475 			 * softlock case so we don't adjust it here.
4476 			 */
4477 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4478 			ASSERT(a >= lpgaddr);
4479 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4480 			off = svd->offset + (uintptr_t)(a - seg->s_base);
4481 			aindx = svd->anon_index + seg_page(seg, a);
4482 			vpage = (svd->vpage != NULL) ?
4483 			    &svd->vpage[seg_page(seg, a)] : NULL;
4484 		} else {
4485 			/*
4486 			 * Size down case. Note lpgaddr may only be needed for
4487 			 * softlock case so we don't adjust it here.
4488 			 */
4489 			ASSERT(IS_P2ALIGNED(a, pgsz));
4490 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4491 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4492 			ASSERT(a < lpgeaddr);
4493 			if (a < addr) {
4494 				SEGVN_VMSTAT_FLTVNPAGES(44);
4495 				/*
4496 				 * The beginning of the large page region can
4497 				 * be pulled to the right to make a smaller
4498 				 * region. We haven't yet faulted a single
4499 				 * page.
4500 				 */
4501 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4502 				ASSERT(a >= lpgaddr);
4503 				off = svd->offset +
4504 				    (uintptr_t)(a - seg->s_base);
4505 				aindx = svd->anon_index + seg_page(seg, a);
4506 				vpage = (svd->vpage != NULL) ?
4507 				    &svd->vpage[seg_page(seg, a)] : NULL;
4508 			}
4509 		}
4510 	}
4511 out:
4512 	kmem_free(ppa, ppasize);
4513 	if (!err && !vop_size_err) {
4514 		SEGVN_VMSTAT_FLTVNPAGES(45);
4515 		return (0);
4516 	}
4517 	if (type == F_SOFTLOCK && a > lpgaddr) {
4518 		SEGVN_VMSTAT_FLTVNPAGES(46);
4519 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4520 	}
4521 	if (!vop_size_err) {
4522 		SEGVN_VMSTAT_FLTVNPAGES(47);
4523 		return (err);
4524 	}
4525 	ASSERT(brkcow || tron || type == F_SOFTLOCK);
4526 	/*
4527 	 * Large page end is mapped beyond the end of file and it's a cow
4528 	 * fault (can be a text replication induced cow) or softlock so we can't
4529 	 * reduce the map area.  For now just demote the segment. This should
4530 	 * really only happen if the end of the file changed after the mapping
4531 	 * was established since when large page segments are created we make
4532 	 * sure they don't extend beyond the end of the file.
4533 	 */
4534 	SEGVN_VMSTAT_FLTVNPAGES(48);
4535 
4536 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4537 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4538 	err = 0;
4539 	if (seg->s_szc != 0) {
4540 		segvn_fltvnpages_clrszc_cnt++;
4541 		ASSERT(svd->softlockcnt == 0);
4542 		err = segvn_clrszc(seg);
4543 		if (err != 0) {
4544 			segvn_fltvnpages_clrszc_err++;
4545 		}
4546 	}
4547 	ASSERT(err || seg->s_szc == 0);
4548 	SEGVN_LOCK_DOWNGRADE(seg->s_as, &svd->lock);
4549 	/* segvn_fault will do its job as if szc had been zero to begin with */
4550 	return (err == 0 ? IE_RETRY : FC_MAKE_ERR(err));
4551 }
4552 
4553 /*
4554  * This routine will attempt to fault in one large page.
4555  * it will use smaller pages if that fails.
4556  * It should only be called for pure anonymous segments.
4557  */
4558 static faultcode_t
4559 segvn_fault_anonpages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
4560     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
4561     caddr_t eaddr, int brkcow)
4562 {
4563 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4564 	struct anon_map *amp = svd->amp;
4565 	uchar_t segtype = svd->type;
4566 	uint_t szc = seg->s_szc;
4567 	size_t pgsz = page_get_pagesize(szc);
4568 	size_t maxpgsz = pgsz;
4569 	pgcnt_t pages = btop(pgsz);
4570 	uint_t ppaszc = szc;
4571 	caddr_t a = lpgaddr;
4572 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
4573 	struct vpage *vpage = (svd->vpage != NULL) ?
4574 	    &svd->vpage[seg_page(seg, a)] : NULL;
4575 	page_t **ppa;
4576 	uint_t	ppa_szc;
4577 	faultcode_t err;
4578 	int ierr;
4579 	uint_t protchk, prot, vpprot;
4580 	ulong_t i;
4581 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
4582 	anon_sync_obj_t cookie;
4583 	int adjszc_chk;
4584 	int pgflags = (svd->tr_state == SEGVN_TR_ON) ? PG_LOCAL : 0;
4585 
4586 	ASSERT(szc != 0);
4587 	ASSERT(amp != NULL);
4588 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
4589 	ASSERT(!(svd->flags & MAP_NORESERVE));
4590 	ASSERT(type != F_SOFTUNLOCK);
4591 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
4592 	ASSERT(!brkcow || svd->tr_state == SEGVN_TR_OFF);
4593 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
4594 
4595 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
4596 
4597 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltanpages[0]);
4598 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltanpages[1]);
4599 
4600 	if (svd->flags & MAP_TEXT) {
4601 		hat_flag |= HAT_LOAD_TEXT;
4602 	}
4603 
4604 	if (svd->pageprot) {
4605 		switch (rw) {
4606 		case S_READ:
4607 			protchk = PROT_READ;
4608 			break;
4609 		case S_WRITE:
4610 			protchk = PROT_WRITE;
4611 			break;
4612 		case S_EXEC:
4613 			protchk = PROT_EXEC;
4614 			break;
4615 		case S_OTHER:
4616 		default:
4617 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4618 			break;
4619 		}
4620 		VM_STAT_ADD(segvnvmstats.fltanpages[2]);
4621 	} else {
4622 		prot = svd->prot;
4623 		/* caller has already done segment level protection check. */
4624 	}
4625 
4626 	ppa = kmem_cache_alloc(segvn_szc_cache[ppaszc], KM_SLEEP);
4627 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4628 	for (;;) {
4629 		adjszc_chk = 0;
4630 		for (; a < lpgeaddr; a += pgsz, aindx += pages) {
4631 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
4632 				VM_STAT_ADD(segvnvmstats.fltanpages[3]);
4633 				ASSERT(vpage != NULL);
4634 				prot = VPP_PROT(vpage);
4635 				ASSERT(sameprot(seg, a, maxpgsz));
4636 				if ((prot & protchk) == 0) {
4637 					err = FC_PROT;
4638 					goto error;
4639 				}
4640 			}
4641 			if (adjszc_chk && IS_P2ALIGNED(a, maxpgsz) &&
4642 			    pgsz < maxpgsz) {
4643 				ASSERT(a > lpgaddr);
4644 				szc = seg->s_szc;
4645 				pgsz = maxpgsz;
4646 				pages = btop(pgsz);
4647 				ASSERT(IS_P2ALIGNED(aindx, pages));
4648 				lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr,
4649 				    pgsz);
4650 			}
4651 			if (type == F_SOFTLOCK) {
4652 				atomic_add_long((ulong_t *)&svd->softlockcnt,
4653 				    pages);
4654 			}
4655 			anon_array_enter(amp, aindx, &cookie);
4656 			ppa_szc = (uint_t)-1;
4657 			ierr = anon_map_getpages(amp, aindx, szc, seg, a,
4658 			    prot, &vpprot, ppa, &ppa_szc, vpage, rw, brkcow,
4659 			    segvn_anypgsz, pgflags, svd->cred);
4660 			if (ierr != 0) {
4661 				anon_array_exit(&cookie);
4662 				VM_STAT_ADD(segvnvmstats.fltanpages[4]);
4663 				if (type == F_SOFTLOCK) {
4664 					atomic_add_long(
4665 					    (ulong_t *)&svd->softlockcnt,
4666 					    -pages);
4667 				}
4668 				if (ierr > 0) {
4669 					VM_STAT_ADD(segvnvmstats.fltanpages[6]);
4670 					err = FC_MAKE_ERR(ierr);
4671 					goto error;
4672 				}
4673 				break;
4674 			}
4675 
4676 			ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4677 
4678 			ASSERT(segtype == MAP_SHARED ||
4679 			    ppa[0]->p_szc <= szc);
4680 			ASSERT(segtype == MAP_PRIVATE ||
4681 			    ppa[0]->p_szc >= szc);
4682 
4683 			/*
4684 			 * Handle pages that have been marked for migration
4685 			 */
4686 			if (lgrp_optimizations())
4687 				page_migrate(seg, a, ppa, pages);
4688 
4689 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
4690 
4691 			if (segtype == MAP_SHARED) {
4692 				vpprot |= PROT_WRITE;
4693 			}
4694 
4695 			hat_memload_array(hat, a, pgsz, ppa,
4696 			    prot & vpprot, hat_flag);
4697 
4698 			if (hat_flag & HAT_LOAD_LOCK) {
4699 				VM_STAT_ADD(segvnvmstats.fltanpages[7]);
4700 			} else {
4701 				VM_STAT_ADD(segvnvmstats.fltanpages[8]);
4702 				for (i = 0; i < pages; i++)
4703 					page_unlock(ppa[i]);
4704 			}
4705 			if (vpage != NULL)
4706 				vpage += pages;
4707 
4708 			anon_array_exit(&cookie);
4709 			adjszc_chk = 1;
4710 		}
4711 		if (a == lpgeaddr)
4712 			break;
4713 		ASSERT(a < lpgeaddr);
4714 		/*
4715 		 * ierr == -1 means we failed to allocate a large page.
4716 		 * so do a size down operation.
4717 		 *
4718 		 * ierr == -2 means some other process that privately shares
4719 		 * pages with this process has allocated a larger page and we
4720 		 * need to retry with larger pages. So do a size up
4721 		 * operation. This relies on the fact that large pages are
4722 		 * never partially shared i.e. if we share any constituent
4723 		 * page of a large page with another process we must share the
4724 		 * entire large page. Note this cannot happen for SOFTLOCK
4725 		 * case, unless current address (a) is at the beginning of the
4726 		 * next page size boundary because the other process couldn't
4727 		 * have relocated locked pages.
4728 		 */
4729 		ASSERT(ierr == -1 || ierr == -2);
4730 
4731 		if (segvn_anypgsz) {
4732 			ASSERT(ierr == -2 || szc != 0);
4733 			ASSERT(ierr == -1 || szc < seg->s_szc);
4734 			szc = (ierr == -1) ? szc - 1 : szc + 1;
4735 		} else {
4736 			/*
4737 			 * For non COW faults and segvn_anypgsz == 0
4738 			 * we need to be careful not to loop forever
4739 			 * if existing page is found with szc other
4740 			 * than 0 or seg->s_szc. This could be due
4741 			 * to page relocations on behalf of DR or
4742 			 * more likely large page creation. For this
4743 			 * case simply re-size to existing page's szc
4744 			 * if returned by anon_map_getpages().
4745 			 */
4746 			if (ppa_szc == (uint_t)-1) {
4747 				szc = (ierr == -1) ? 0 : seg->s_szc;
4748 			} else {
4749 				ASSERT(ppa_szc <= seg->s_szc);
4750 				ASSERT(ierr == -2 || ppa_szc < szc);
4751 				ASSERT(ierr == -1 || ppa_szc > szc);
4752 				szc = ppa_szc;
4753 			}
4754 		}
4755 
4756 		pgsz = page_get_pagesize(szc);
4757 		pages = btop(pgsz);
4758 		ASSERT(type != F_SOFTLOCK || ierr == -1 ||
4759 		    (IS_P2ALIGNED(a, pgsz) && IS_P2ALIGNED(lpgeaddr, pgsz)));
4760 		if (type == F_SOFTLOCK) {
4761 			/*
4762 			 * For softlocks we cannot reduce the fault area
4763 			 * (calculated based on the largest page size for this
4764 			 * segment) for size down and a is already next
4765 			 * page size aligned as assertted above for size
4766 			 * ups. Therefore just continue in case of softlock.
4767 			 */
4768 			VM_STAT_ADD(segvnvmstats.fltanpages[9]);
4769 			continue; /* keep lint happy */
4770 		} else if (ierr == -2) {
4771 
4772 			/*
4773 			 * Size up case. Note lpgaddr may only be needed for
4774 			 * softlock case so we don't adjust it here.
4775 			 */
4776 			VM_STAT_ADD(segvnvmstats.fltanpages[10]);
4777 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4778 			ASSERT(a >= lpgaddr);
4779 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4780 			aindx = svd->anon_index + seg_page(seg, a);
4781 			vpage = (svd->vpage != NULL) ?
4782 			    &svd->vpage[seg_page(seg, a)] : NULL;
4783 		} else {
4784 			/*
4785 			 * Size down case. Note lpgaddr may only be needed for
4786 			 * softlock case so we don't adjust it here.
4787 			 */
4788 			VM_STAT_ADD(segvnvmstats.fltanpages[11]);
4789 			ASSERT(IS_P2ALIGNED(a, pgsz));
4790 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4791 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4792 			ASSERT(a < lpgeaddr);
4793 			if (a < addr) {
4794 				/*
4795 				 * The beginning of the large page region can
4796 				 * be pulled to the right to make a smaller
4797 				 * region. We haven't yet faulted a single
4798 				 * page.
4799 				 */
4800 				VM_STAT_ADD(segvnvmstats.fltanpages[12]);
4801 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4802 				ASSERT(a >= lpgaddr);
4803 				aindx = svd->anon_index + seg_page(seg, a);
4804 				vpage = (svd->vpage != NULL) ?
4805 				    &svd->vpage[seg_page(seg, a)] : NULL;
4806 			}
4807 		}
4808 	}
4809 	VM_STAT_ADD(segvnvmstats.fltanpages[13]);
4810 	ANON_LOCK_EXIT(&amp->a_rwlock);
4811 	kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4812 	return (0);
4813 error:
4814 	VM_STAT_ADD(segvnvmstats.fltanpages[14]);
4815 	ANON_LOCK_EXIT(&amp->a_rwlock);
4816 	kmem_cache_free(segvn_szc_cache[ppaszc], ppa);
4817 	if (type == F_SOFTLOCK && a > lpgaddr) {
4818 		VM_STAT_ADD(segvnvmstats.fltanpages[15]);
4819 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4820 	}
4821 	return (err);
4822 }
4823 
4824 int fltadvice = 1;	/* set to free behind pages for sequential access */
4825 
4826 /*
4827  * This routine is called via a machine specific fault handling routine.
4828  * It is also called by software routines wishing to lock or unlock
4829  * a range of addresses.
4830  *
4831  * Here is the basic algorithm:
4832  *	If unlocking
4833  *		Call segvn_softunlock
4834  *		Return
4835  *	endif
4836  *	Checking and set up work
4837  *	If we will need some non-anonymous pages
4838  *		Call VOP_GETPAGE over the range of non-anonymous pages
4839  *	endif
4840  *	Loop over all addresses requested
4841  *		Call segvn_faultpage passing in page list
4842  *		    to load up translations and handle anonymous pages
4843  *	endloop
4844  *	Load up translation to any additional pages in page list not
4845  *	    already handled that fit into this segment
4846  */
4847 static faultcode_t
4848 segvn_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
4849     enum fault_type type, enum seg_rw rw)
4850 {
4851 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4852 	page_t **plp, **ppp, *pp;
4853 	u_offset_t off;
4854 	caddr_t a;
4855 	struct vpage *vpage;
4856 	uint_t vpprot, prot;
4857 	int err;
4858 	page_t *pl[PVN_GETPAGE_NUM + 1];
4859 	size_t plsz, pl_alloc_sz;
4860 	size_t page;
4861 	ulong_t anon_index;
4862 	struct anon_map *amp;
4863 	int dogetpage = 0;
4864 	caddr_t	lpgaddr, lpgeaddr;
4865 	size_t pgsz;
4866 	anon_sync_obj_t cookie;
4867 	int brkcow = BREAK_COW_SHARE(rw, type, svd->type);
4868 
4869 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
4870 	ASSERT(svd->amp == NULL || svd->rcookie == HAT_INVALID_REGION_COOKIE);
4871 
4872 	/*
4873 	 * First handle the easy stuff
4874 	 */
4875 	if (type == F_SOFTUNLOCK) {
4876 		if (rw == S_READ_NOCOW) {
4877 			rw = S_READ;
4878 			ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
4879 		}
4880 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4881 		pgsz = (seg->s_szc == 0) ? PAGESIZE :
4882 		    page_get_pagesize(seg->s_szc);
4883 		VM_STAT_COND_ADD(pgsz > PAGESIZE, segvnvmstats.fltanpages[16]);
4884 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
4885 		segvn_softunlock(seg, lpgaddr, lpgeaddr - lpgaddr, rw);
4886 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4887 		return (0);
4888 	}
4889 
4890 	ASSERT(svd->tr_state == SEGVN_TR_OFF ||
4891 	    !HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
4892 	if (brkcow == 0) {
4893 		if (svd->tr_state == SEGVN_TR_INIT) {
4894 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4895 			if (svd->tr_state == SEGVN_TR_INIT) {
4896 				ASSERT(svd->vp != NULL && svd->amp == NULL);
4897 				ASSERT(svd->flags & MAP_TEXT);
4898 				ASSERT(svd->type == MAP_PRIVATE);
4899 				segvn_textrepl(seg);
4900 				ASSERT(svd->tr_state != SEGVN_TR_INIT);
4901 				ASSERT(svd->tr_state != SEGVN_TR_ON ||
4902 				    svd->amp != NULL);
4903 			}
4904 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4905 		}
4906 	} else if (svd->tr_state != SEGVN_TR_OFF) {
4907 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4908 
4909 		if (rw == S_WRITE && svd->tr_state != SEGVN_TR_OFF) {
4910 			ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
4911 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4912 			return (FC_PROT);
4913 		}
4914 
4915 		if (svd->tr_state == SEGVN_TR_ON) {
4916 			ASSERT(svd->vp != NULL && svd->amp != NULL);
4917 			segvn_textunrepl(seg, 0);
4918 			ASSERT(svd->amp == NULL &&
4919 			    svd->tr_state == SEGVN_TR_OFF);
4920 		} else if (svd->tr_state != SEGVN_TR_OFF) {
4921 			svd->tr_state = SEGVN_TR_OFF;
4922 		}
4923 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
4924 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4925 	}
4926 
4927 top:
4928 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4929 
4930 	/*
4931 	 * If we have the same protections for the entire segment,
4932 	 * insure that the access being attempted is legitimate.
4933 	 */
4934 
4935 	if (svd->pageprot == 0) {
4936 		uint_t protchk;
4937 
4938 		switch (rw) {
4939 		case S_READ:
4940 		case S_READ_NOCOW:
4941 			protchk = PROT_READ;
4942 			break;
4943 		case S_WRITE:
4944 			protchk = PROT_WRITE;
4945 			break;
4946 		case S_EXEC:
4947 			protchk = PROT_EXEC;
4948 			break;
4949 		case S_OTHER:
4950 		default:
4951 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4952 			break;
4953 		}
4954 
4955 		if ((svd->prot & protchk) == 0) {
4956 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4957 			return (FC_PROT);	/* illegal access type */
4958 		}
4959 	}
4960 
4961 	if (brkcow && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
4962 		/* this must be SOFTLOCK S_READ fault */
4963 		ASSERT(svd->amp == NULL);
4964 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
4965 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4966 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4967 		if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
4968 			/*
4969 			 * this must be the first ever non S_READ_NOCOW
4970 			 * softlock for this segment.
4971 			 */
4972 			ASSERT(svd->softlockcnt == 0);
4973 			hat_leave_region(seg->s_as->a_hat, svd->rcookie,
4974 			    HAT_REGION_TEXT);
4975 			svd->rcookie = HAT_INVALID_REGION_COOKIE;
4976 		}
4977 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4978 		goto top;
4979 	}
4980 
4981 	/*
4982 	 * We can't allow the long term use of softlocks for vmpss segments,
4983 	 * because in some file truncation cases we should be able to demote
4984 	 * the segment, which requires that there are no softlocks.  The
4985 	 * only case where it's ok to allow a SOFTLOCK fault against a vmpss
4986 	 * segment is S_READ_NOCOW, where the caller holds the address space
4987 	 * locked as writer and calls softunlock before dropping the as lock.
4988 	 * S_READ_NOCOW is used by /proc to read memory from another user.
4989 	 *
4990 	 * Another deadlock between SOFTLOCK and file truncation can happen
4991 	 * because segvn_fault_vnodepages() calls the FS one pagesize at
4992 	 * a time. A second VOP_GETPAGE() call by segvn_fault_vnodepages()
4993 	 * can cause a deadlock because the first set of page_t's remain
4994 	 * locked SE_SHARED.  To avoid this, we demote segments on a first
4995 	 * SOFTLOCK if they have a length greater than the segment's
4996 	 * page size.
4997 	 *
4998 	 * So for now, we only avoid demoting a segment on a SOFTLOCK when
4999 	 * the access type is S_READ_NOCOW and the fault length is less than
5000 	 * or equal to the segment's page size. While this is quite restrictive,
5001 	 * it should be the most common case of SOFTLOCK against a vmpss
5002 	 * segment.
5003 	 *
5004 	 * For S_READ_NOCOW, it's safe not to do a copy on write because the
5005 	 * caller makes sure no COW will be caused by another thread for a
5006 	 * softlocked page.
5007 	 */
5008 	if (type == F_SOFTLOCK && svd->vp != NULL && seg->s_szc != 0) {
5009 		int demote = 0;
5010 
5011 		if (rw != S_READ_NOCOW) {
5012 			demote = 1;
5013 		}
5014 		if (!demote && len > PAGESIZE) {
5015 			pgsz = page_get_pagesize(seg->s_szc);
5016 			CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr,
5017 			    lpgeaddr);
5018 			if (lpgeaddr - lpgaddr > pgsz) {
5019 				demote = 1;
5020 			}
5021 		}
5022 
5023 		ASSERT(demote || AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5024 
5025 		if (demote) {
5026 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5027 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5028 			if (seg->s_szc != 0) {
5029 				segvn_vmpss_clrszc_cnt++;
5030 				ASSERT(svd->softlockcnt == 0);
5031 				err = segvn_clrszc(seg);
5032 				if (err) {
5033 					segvn_vmpss_clrszc_err++;
5034 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5035 					return (FC_MAKE_ERR(err));
5036 				}
5037 			}
5038 			ASSERT(seg->s_szc == 0);
5039 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5040 			goto top;
5041 		}
5042 	}
5043 
5044 	/*
5045 	 * Check to see if we need to allocate an anon_map structure.
5046 	 */
5047 	if (svd->amp == NULL && (svd->vp == NULL || brkcow)) {
5048 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5049 		/*
5050 		 * Drop the "read" lock on the segment and acquire
5051 		 * the "write" version since we have to allocate the
5052 		 * anon_map.
5053 		 */
5054 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5055 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5056 
5057 		if (svd->amp == NULL) {
5058 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
5059 			svd->amp->a_szc = seg->s_szc;
5060 		}
5061 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5062 
5063 		/*
5064 		 * Start all over again since segment protections
5065 		 * may have changed after we dropped the "read" lock.
5066 		 */
5067 		goto top;
5068 	}
5069 
5070 	/*
5071 	 * S_READ_NOCOW vs S_READ distinction was
5072 	 * only needed for the code above. After
5073 	 * that we treat it as S_READ.
5074 	 */
5075 	if (rw == S_READ_NOCOW) {
5076 		ASSERT(type == F_SOFTLOCK);
5077 		ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5078 		rw = S_READ;
5079 	}
5080 
5081 	amp = svd->amp;
5082 
5083 	/*
5084 	 * MADV_SEQUENTIAL work is ignored for large page segments.
5085 	 */
5086 	if (seg->s_szc != 0) {
5087 		pgsz = page_get_pagesize(seg->s_szc);
5088 		ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
5089 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
5090 		if (svd->vp == NULL) {
5091 			err = segvn_fault_anonpages(hat, seg, lpgaddr,
5092 			    lpgeaddr, type, rw, addr, addr + len, brkcow);
5093 		} else {
5094 			err = segvn_fault_vnodepages(hat, seg, lpgaddr,
5095 			    lpgeaddr, type, rw, addr, addr + len, brkcow);
5096 			if (err == IE_RETRY) {
5097 				ASSERT(seg->s_szc == 0);
5098 				ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
5099 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5100 				goto top;
5101 			}
5102 		}
5103 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5104 		return (err);
5105 	}
5106 
5107 	page = seg_page(seg, addr);
5108 	if (amp != NULL) {
5109 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5110 		anon_index = svd->anon_index + page;
5111 
5112 		if (type == F_PROT && rw == S_READ &&
5113 		    svd->tr_state == SEGVN_TR_OFF &&
5114 		    svd->type == MAP_PRIVATE && svd->pageprot == 0) {
5115 			size_t index = anon_index;
5116 			struct anon *ap;
5117 
5118 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5119 			/*
5120 			 * The fast path could apply to S_WRITE also, except
5121 			 * that the protection fault could be caused by lazy
5122 			 * tlb flush when ro->rw. In this case, the pte is
5123 			 * RW already. But RO in the other cpu's tlb causes
5124 			 * the fault. Since hat_chgprot won't do anything if
5125 			 * pte doesn't change, we may end up faulting
5126 			 * indefinitely until the RO tlb entry gets replaced.
5127 			 */
5128 			for (a = addr; a < addr + len; a += PAGESIZE, index++) {
5129 				anon_array_enter(amp, index, &cookie);
5130 				ap = anon_get_ptr(amp->ahp, index);
5131 				anon_array_exit(&cookie);
5132 				if ((ap == NULL) || (ap->an_refcnt != 1)) {
5133 					ANON_LOCK_EXIT(&amp->a_rwlock);
5134 					goto slow;
5135 				}
5136 			}
5137 			hat_chgprot(seg->s_as->a_hat, addr, len, svd->prot);
5138 			ANON_LOCK_EXIT(&amp->a_rwlock);
5139 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5140 			return (0);
5141 		}
5142 	}
5143 slow:
5144 
5145 	if (svd->vpage == NULL)
5146 		vpage = NULL;
5147 	else
5148 		vpage = &svd->vpage[page];
5149 
5150 	off = svd->offset + (uintptr_t)(addr - seg->s_base);
5151 
5152 	/*
5153 	 * If MADV_SEQUENTIAL has been set for the particular page we
5154 	 * are faulting on, free behind all pages in the segment and put
5155 	 * them on the free list.
5156 	 */
5157 
5158 	if ((page != 0) && fltadvice && svd->tr_state != SEGVN_TR_ON) {
5159 		struct vpage *vpp;
5160 		ulong_t fanon_index;
5161 		size_t fpage;
5162 		u_offset_t pgoff, fpgoff;
5163 		struct vnode *fvp;
5164 		struct anon *fap = NULL;
5165 
5166 		if (svd->advice == MADV_SEQUENTIAL ||
5167 		    (svd->pageadvice &&
5168 		    VPP_ADVICE(vpage) == MADV_SEQUENTIAL)) {
5169 			pgoff = off - PAGESIZE;
5170 			fpage = page - 1;
5171 			if (vpage != NULL)
5172 				vpp = &svd->vpage[fpage];
5173 			if (amp != NULL)
5174 				fanon_index = svd->anon_index + fpage;
5175 
5176 			while (pgoff > svd->offset) {
5177 				if (svd->advice != MADV_SEQUENTIAL &&
5178 				    (!svd->pageadvice || (vpage &&
5179 				    VPP_ADVICE(vpp) != MADV_SEQUENTIAL)))
5180 					break;
5181 
5182 				/*
5183 				 * If this is an anon page, we must find the
5184 				 * correct <vp, offset> for it
5185 				 */
5186 				fap = NULL;
5187 				if (amp != NULL) {
5188 					ANON_LOCK_ENTER(&amp->a_rwlock,
5189 					    RW_READER);
5190 					anon_array_enter(amp, fanon_index,
5191 					    &cookie);
5192 					fap = anon_get_ptr(amp->ahp,
5193 					    fanon_index);
5194 					if (fap != NULL) {
5195 						swap_xlate(fap, &fvp, &fpgoff);
5196 					} else {
5197 						fpgoff = pgoff;
5198 						fvp = svd->vp;
5199 					}
5200 					anon_array_exit(&cookie);
5201 					ANON_LOCK_EXIT(&amp->a_rwlock);
5202 				} else {
5203 					fpgoff = pgoff;
5204 					fvp = svd->vp;
5205 				}
5206 				if (fvp == NULL)
5207 					break;	/* XXX */
5208 				/*
5209 				 * Skip pages that are free or have an
5210 				 * "exclusive" lock.
5211 				 */
5212 				pp = page_lookup_nowait(fvp, fpgoff, SE_SHARED);
5213 				if (pp == NULL)
5214 					break;
5215 				/*
5216 				 * We don't need the page_struct_lock to test
5217 				 * as this is only advisory; even if we
5218 				 * acquire it someone might race in and lock
5219 				 * the page after we unlock and before the
5220 				 * PUTPAGE, then VOP_PUTPAGE will do nothing.
5221 				 */
5222 				if (pp->p_lckcnt == 0 && pp->p_cowcnt == 0) {
5223 					/*
5224 					 * Hold the vnode before releasing
5225 					 * the page lock to prevent it from
5226 					 * being freed and re-used by some
5227 					 * other thread.
5228 					 */
5229 					VN_HOLD(fvp);
5230 					page_unlock(pp);
5231 					/*
5232 					 * We should build a page list
5233 					 * to kluster putpages XXX
5234 					 */
5235 					(void) VOP_PUTPAGE(fvp,
5236 					    (offset_t)fpgoff, PAGESIZE,
5237 					    (B_DONTNEED|B_FREE|B_ASYNC),
5238 					    svd->cred, NULL);
5239 					VN_RELE(fvp);
5240 				} else {
5241 					/*
5242 					 * XXX - Should the loop terminate if
5243 					 * the page is `locked'?
5244 					 */
5245 					page_unlock(pp);
5246 				}
5247 				--vpp;
5248 				--fanon_index;
5249 				pgoff -= PAGESIZE;
5250 			}
5251 		}
5252 	}
5253 
5254 	plp = pl;
5255 	*plp = NULL;
5256 	pl_alloc_sz = 0;
5257 
5258 	/*
5259 	 * See if we need to call VOP_GETPAGE for
5260 	 * *any* of the range being faulted on.
5261 	 * We can skip all of this work if there
5262 	 * was no original vnode.
5263 	 */
5264 	if (svd->vp != NULL) {
5265 		u_offset_t vp_off;
5266 		size_t vp_len;
5267 		struct anon *ap;
5268 		vnode_t *vp;
5269 
5270 		vp_off = off;
5271 		vp_len = len;
5272 
5273 		if (amp == NULL)
5274 			dogetpage = 1;
5275 		else {
5276 			/*
5277 			 * Only acquire reader lock to prevent amp->ahp
5278 			 * from being changed.  It's ok to miss pages,
5279 			 * hence we don't do anon_array_enter
5280 			 */
5281 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5282 			ap = anon_get_ptr(amp->ahp, anon_index);
5283 
5284 			if (len <= PAGESIZE)
5285 				/* inline non_anon() */
5286 				dogetpage = (ap == NULL);
5287 			else
5288 				dogetpage = non_anon(amp->ahp, anon_index,
5289 				    &vp_off, &vp_len);
5290 			ANON_LOCK_EXIT(&amp->a_rwlock);
5291 		}
5292 
5293 		if (dogetpage) {
5294 			enum seg_rw arw;
5295 			struct as *as = seg->s_as;
5296 
5297 			if (len > ptob((sizeof (pl) / sizeof (pl[0])) - 1)) {
5298 				/*
5299 				 * Page list won't fit in local array,
5300 				 * allocate one of the needed size.
5301 				 */
5302 				pl_alloc_sz =
5303 				    (btop(len) + 1) * sizeof (page_t *);
5304 				plp = kmem_alloc(pl_alloc_sz, KM_SLEEP);
5305 				plp[0] = NULL;
5306 				plsz = len;
5307 			} else if (rw == S_WRITE && svd->type == MAP_PRIVATE ||
5308 			    svd->tr_state == SEGVN_TR_ON || rw == S_OTHER ||
5309 			    (((size_t)(addr + PAGESIZE) <
5310 			    (size_t)(seg->s_base + seg->s_size)) &&
5311 			    hat_probe(as->a_hat, addr + PAGESIZE))) {
5312 				/*
5313 				 * Ask VOP_GETPAGE to return the exact number
5314 				 * of pages if
5315 				 * (a) this is a COW fault, or
5316 				 * (b) this is a software fault, or
5317 				 * (c) next page is already mapped.
5318 				 */
5319 				plsz = len;
5320 			} else {
5321 				/*
5322 				 * Ask VOP_GETPAGE to return adjacent pages
5323 				 * within the segment.
5324 				 */
5325 				plsz = MIN((size_t)PVN_GETPAGE_SZ, (size_t)
5326 				    ((seg->s_base + seg->s_size) - addr));
5327 				ASSERT((addr + plsz) <=
5328 				    (seg->s_base + seg->s_size));
5329 			}
5330 
5331 			/*
5332 			 * Need to get some non-anonymous pages.
5333 			 * We need to make only one call to GETPAGE to do
5334 			 * this to prevent certain deadlocking conditions
5335 			 * when we are doing locking.  In this case
5336 			 * non_anon() should have picked up the smallest
5337 			 * range which includes all the non-anonymous
5338 			 * pages in the requested range.  We have to
5339 			 * be careful regarding which rw flag to pass in
5340 			 * because on a private mapping, the underlying
5341 			 * object is never allowed to be written.
5342 			 */
5343 			if (rw == S_WRITE && svd->type == MAP_PRIVATE) {
5344 				arw = S_READ;
5345 			} else {
5346 				arw = rw;
5347 			}
5348 			vp = svd->vp;
5349 			TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5350 			    "segvn_getpage:seg %p addr %p vp %p",
5351 			    seg, addr, vp);
5352 			err = VOP_GETPAGE(vp, (offset_t)vp_off, vp_len,
5353 			    &vpprot, plp, plsz, seg, addr + (vp_off - off), arw,
5354 			    svd->cred, NULL);
5355 			if (err) {
5356 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5357 				segvn_pagelist_rele(plp);
5358 				if (pl_alloc_sz)
5359 					kmem_free(plp, pl_alloc_sz);
5360 				return (FC_MAKE_ERR(err));
5361 			}
5362 			if (svd->type == MAP_PRIVATE)
5363 				vpprot &= ~PROT_WRITE;
5364 		}
5365 	}
5366 
5367 	/*
5368 	 * N.B. at this time the plp array has all the needed non-anon
5369 	 * pages in addition to (possibly) having some adjacent pages.
5370 	 */
5371 
5372 	/*
5373 	 * Always acquire the anon_array_lock to prevent
5374 	 * 2 threads from allocating separate anon slots for
5375 	 * the same "addr".
5376 	 *
5377 	 * If this is a copy-on-write fault and we don't already
5378 	 * have the anon_array_lock, acquire it to prevent the
5379 	 * fault routine from handling multiple copy-on-write faults
5380 	 * on the same "addr" in the same address space.
5381 	 *
5382 	 * Only one thread should deal with the fault since after
5383 	 * it is handled, the other threads can acquire a translation
5384 	 * to the newly created private page.  This prevents two or
5385 	 * more threads from creating different private pages for the
5386 	 * same fault.
5387 	 *
5388 	 * We grab "serialization" lock here if this is a MAP_PRIVATE segment
5389 	 * to prevent deadlock between this thread and another thread
5390 	 * which has soft-locked this page and wants to acquire serial_lock.
5391 	 * ( bug 4026339 )
5392 	 *
5393 	 * The fix for bug 4026339 becomes unnecessary when using the
5394 	 * locking scheme with per amp rwlock and a global set of hash
5395 	 * lock, anon_array_lock.  If we steal a vnode page when low
5396 	 * on memory and upgrad the page lock through page_rename,
5397 	 * then the page is PAGE_HANDLED, nothing needs to be done
5398 	 * for this page after returning from segvn_faultpage.
5399 	 *
5400 	 * But really, the page lock should be downgraded after
5401 	 * the stolen page is page_rename'd.
5402 	 */
5403 
5404 	if (amp != NULL)
5405 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5406 
5407 	/*
5408 	 * Ok, now loop over the address range and handle faults
5409 	 */
5410 	for (a = addr; a < addr + len; a += PAGESIZE, off += PAGESIZE) {
5411 		err = segvn_faultpage(hat, seg, a, off, vpage, plp, vpprot,
5412 		    type, rw, brkcow);
5413 		if (err) {
5414 			if (amp != NULL)
5415 				ANON_LOCK_EXIT(&amp->a_rwlock);
5416 			if (type == F_SOFTLOCK && a > addr) {
5417 				segvn_softunlock(seg, addr, (a - addr),
5418 				    S_OTHER);
5419 			}
5420 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5421 			segvn_pagelist_rele(plp);
5422 			if (pl_alloc_sz)
5423 				kmem_free(plp, pl_alloc_sz);
5424 			return (err);
5425 		}
5426 		if (vpage) {
5427 			vpage++;
5428 		} else if (svd->vpage) {
5429 			page = seg_page(seg, addr);
5430 			vpage = &svd->vpage[++page];
5431 		}
5432 	}
5433 
5434 	/* Didn't get pages from the underlying fs so we're done */
5435 	if (!dogetpage)
5436 		goto done;
5437 
5438 	/*
5439 	 * Now handle any other pages in the list returned.
5440 	 * If the page can be used, load up the translations now.
5441 	 * Note that the for loop will only be entered if "plp"
5442 	 * is pointing to a non-NULL page pointer which means that
5443 	 * VOP_GETPAGE() was called and vpprot has been initialized.
5444 	 */
5445 	if (svd->pageprot == 0)
5446 		prot = svd->prot & vpprot;
5447 
5448 
5449 	/*
5450 	 * Large Files: diff should be unsigned value because we started
5451 	 * supporting > 2GB segment sizes from 2.5.1 and when a
5452 	 * large file of size > 2GB gets mapped to address space
5453 	 * the diff value can be > 2GB.
5454 	 */
5455 
5456 	for (ppp = plp; (pp = *ppp) != NULL; ppp++) {
5457 		size_t diff;
5458 		struct anon *ap;
5459 		int anon_index;
5460 		anon_sync_obj_t cookie;
5461 		int hat_flag = HAT_LOAD_ADV;
5462 
5463 		if (svd->flags & MAP_TEXT) {
5464 			hat_flag |= HAT_LOAD_TEXT;
5465 		}
5466 
5467 		if (pp == PAGE_HANDLED)
5468 			continue;
5469 
5470 		if (svd->tr_state != SEGVN_TR_ON &&
5471 		    pp->p_offset >=  svd->offset &&
5472 		    pp->p_offset < svd->offset + seg->s_size) {
5473 
5474 			diff = pp->p_offset - svd->offset;
5475 
5476 			/*
5477 			 * Large Files: Following is the assertion
5478 			 * validating the above cast.
5479 			 */
5480 			ASSERT(svd->vp == pp->p_vnode);
5481 
5482 			page = btop(diff);
5483 			if (svd->pageprot)
5484 				prot = VPP_PROT(&svd->vpage[page]) & vpprot;
5485 
5486 			/*
5487 			 * Prevent other threads in the address space from
5488 			 * creating private pages (i.e., allocating anon slots)
5489 			 * while we are in the process of loading translations
5490 			 * to additional pages returned by the underlying
5491 			 * object.
5492 			 */
5493 			if (amp != NULL) {
5494 				anon_index = svd->anon_index + page;
5495 				anon_array_enter(amp, anon_index, &cookie);
5496 				ap = anon_get_ptr(amp->ahp, anon_index);
5497 			}
5498 			if ((amp == NULL) || (ap == NULL)) {
5499 				if (IS_VMODSORT(pp->p_vnode) ||
5500 				    enable_mbit_wa) {
5501 					if (rw == S_WRITE)
5502 						hat_setmod(pp);
5503 					else if (rw != S_OTHER &&
5504 					    !hat_ismod(pp))
5505 						prot &= ~PROT_WRITE;
5506 				}
5507 				/*
5508 				 * Skip mapping read ahead pages marked
5509 				 * for migration, so they will get migrated
5510 				 * properly on fault
5511 				 */
5512 				ASSERT(amp == NULL ||
5513 				    svd->rcookie == HAT_INVALID_REGION_COOKIE);
5514 				if ((prot & PROT_READ) && !PP_ISMIGRATE(pp)) {
5515 					hat_memload_region(hat,
5516 					    seg->s_base + diff,
5517 					    pp, prot, hat_flag,
5518 					    svd->rcookie);
5519 				}
5520 			}
5521 			if (amp != NULL)
5522 				anon_array_exit(&cookie);
5523 		}
5524 		page_unlock(pp);
5525 	}
5526 done:
5527 	if (amp != NULL)
5528 		ANON_LOCK_EXIT(&amp->a_rwlock);
5529 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5530 	if (pl_alloc_sz)
5531 		kmem_free(plp, pl_alloc_sz);
5532 	return (0);
5533 }
5534 
5535 /*
5536  * This routine is used to start I/O on pages asynchronously.  XXX it will
5537  * only create PAGESIZE pages. At fault time they will be relocated into
5538  * larger pages.
5539  */
5540 static faultcode_t
5541 segvn_faulta(struct seg *seg, caddr_t addr)
5542 {
5543 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5544 	int err;
5545 	struct anon_map *amp;
5546 	vnode_t *vp;
5547 
5548 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5549 
5550 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5551 	if ((amp = svd->amp) != NULL) {
5552 		struct anon *ap;
5553 
5554 		/*
5555 		 * Reader lock to prevent amp->ahp from being changed.
5556 		 * This is advisory, it's ok to miss a page, so
5557 		 * we don't do anon_array_enter lock.
5558 		 */
5559 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5560 		if ((ap = anon_get_ptr(amp->ahp,
5561 		    svd->anon_index + seg_page(seg, addr))) != NULL) {
5562 
5563 			err = anon_getpage(&ap, NULL, NULL,
5564 			    0, seg, addr, S_READ, svd->cred);
5565 
5566 			ANON_LOCK_EXIT(&amp->a_rwlock);
5567 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5568 			if (err)
5569 				return (FC_MAKE_ERR(err));
5570 			return (0);
5571 		}
5572 		ANON_LOCK_EXIT(&amp->a_rwlock);
5573 	}
5574 
5575 	if (svd->vp == NULL) {
5576 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5577 		return (0);			/* zfod page - do nothing now */
5578 	}
5579 
5580 	vp = svd->vp;
5581 	TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5582 	    "segvn_getpage:seg %p addr %p vp %p", seg, addr, vp);
5583 	err = VOP_GETPAGE(vp,
5584 	    (offset_t)(svd->offset + (uintptr_t)(addr - seg->s_base)),
5585 	    PAGESIZE, NULL, NULL, 0, seg, addr,
5586 	    S_OTHER, svd->cred, NULL);
5587 
5588 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5589 	if (err)
5590 		return (FC_MAKE_ERR(err));
5591 	return (0);
5592 }
5593 
5594 static int
5595 segvn_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
5596 {
5597 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5598 	struct vpage *cvp, *svp, *evp;
5599 	struct vnode *vp;
5600 	size_t pgsz;
5601 	pgcnt_t pgcnt;
5602 	anon_sync_obj_t cookie;
5603 	int unload_done = 0;
5604 
5605 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5606 
5607 	if ((svd->maxprot & prot) != prot)
5608 		return (EACCES);			/* violated maxprot */
5609 
5610 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5611 
5612 	/* return if prot is the same */
5613 	if (!svd->pageprot && svd->prot == prot) {
5614 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5615 		return (0);
5616 	}
5617 
5618 	/*
5619 	 * Since we change protections we first have to flush the cache.
5620 	 * This makes sure all the pagelock calls have to recheck
5621 	 * protections.
5622 	 */
5623 	if (svd->softlockcnt > 0) {
5624 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5625 
5626 		/*
5627 		 * If this is shared segment non 0 softlockcnt
5628 		 * means locked pages are still in use.
5629 		 */
5630 		if (svd->type == MAP_SHARED) {
5631 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5632 			return (EAGAIN);
5633 		}
5634 
5635 		/*
5636 		 * Since we do have the segvn writers lock nobody can fill
5637 		 * the cache with entries belonging to this seg during
5638 		 * the purge. The flush either succeeds or we still have
5639 		 * pending I/Os.
5640 		 */
5641 		segvn_purge(seg);
5642 		if (svd->softlockcnt > 0) {
5643 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5644 			return (EAGAIN);
5645 		}
5646 	}
5647 
5648 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5649 		ASSERT(svd->amp == NULL);
5650 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5651 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5652 		    HAT_REGION_TEXT);
5653 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
5654 		unload_done = 1;
5655 	} else if (svd->tr_state == SEGVN_TR_INIT) {
5656 		svd->tr_state = SEGVN_TR_OFF;
5657 	} else if (svd->tr_state == SEGVN_TR_ON) {
5658 		ASSERT(svd->amp != NULL);
5659 		segvn_textunrepl(seg, 0);
5660 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5661 		unload_done = 1;
5662 	}
5663 
5664 	if ((prot & PROT_WRITE) && svd->type == MAP_SHARED &&
5665 	    svd->vp != NULL && (svd->vp->v_flag & VVMEXEC)) {
5666 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
5667 		segvn_inval_trcache(svd->vp);
5668 	}
5669 	if (seg->s_szc != 0) {
5670 		int err;
5671 		pgsz = page_get_pagesize(seg->s_szc);
5672 		pgcnt = pgsz >> PAGESHIFT;
5673 		ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
5674 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
5675 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5676 			ASSERT(seg->s_base != addr || seg->s_size != len);
5677 			/*
5678 			 * If we are holding the as lock as a reader then
5679 			 * we need to return IE_RETRY and let the as
5680 			 * layer drop and re-acquire the lock as a writer.
5681 			 */
5682 			if (AS_READ_HELD(seg->s_as, &seg->s_as->a_lock))
5683 				return (IE_RETRY);
5684 			VM_STAT_ADD(segvnvmstats.demoterange[1]);
5685 			if (svd->type == MAP_PRIVATE || svd->vp != NULL) {
5686 				err = segvn_demote_range(seg, addr, len,
5687 				    SDR_END, 0);
5688 			} else {
5689 				uint_t szcvec = map_pgszcvec(seg->s_base,
5690 				    pgsz, (uintptr_t)seg->s_base,
5691 				    (svd->flags & MAP_TEXT), MAPPGSZC_SHM, 0);
5692 				err = segvn_demote_range(seg, addr, len,
5693 				    SDR_END, szcvec);
5694 			}
5695 			if (err == 0)
5696 				return (IE_RETRY);
5697 			if (err == ENOMEM)
5698 				return (IE_NOMEM);
5699 			return (err);
5700 		}
5701 	}
5702 
5703 
5704 	/*
5705 	 * If it's a private mapping and we're making it writable then we
5706 	 * may have to reserve the additional swap space now. If we are
5707 	 * making writable only a part of the segment then we use its vpage
5708 	 * array to keep a record of the pages for which we have reserved
5709 	 * swap. In this case we set the pageswap field in the segment's
5710 	 * segvn structure to record this.
5711 	 *
5712 	 * If it's a private mapping to a file (i.e., vp != NULL) and we're
5713 	 * removing write permission on the entire segment and we haven't
5714 	 * modified any pages, we can release the swap space.
5715 	 */
5716 	if (svd->type == MAP_PRIVATE) {
5717 		if (prot & PROT_WRITE) {
5718 			if (!(svd->flags & MAP_NORESERVE) &&
5719 			    !(svd->swresv && svd->pageswap == 0)) {
5720 				size_t sz = 0;
5721 
5722 				/*
5723 				 * Start by determining how much swap
5724 				 * space is required.
5725 				 */
5726 				if (addr == seg->s_base &&
5727 				    len == seg->s_size &&
5728 				    svd->pageswap == 0) {
5729 					/* The whole segment */
5730 					sz = seg->s_size;
5731 				} else {
5732 					/*
5733 					 * Make sure that the vpage array
5734 					 * exists, and make a note of the
5735 					 * range of elements corresponding
5736 					 * to len.
5737 					 */
5738 					segvn_vpage(seg);
5739 					svp = &svd->vpage[seg_page(seg, addr)];
5740 					evp = &svd->vpage[seg_page(seg,
5741 					    addr + len)];
5742 
5743 					if (svd->pageswap == 0) {
5744 						/*
5745 						 * This is the first time we've
5746 						 * asked for a part of this
5747 						 * segment, so we need to
5748 						 * reserve everything we've
5749 						 * been asked for.
5750 						 */
5751 						sz = len;
5752 					} else {
5753 						/*
5754 						 * We have to count the number
5755 						 * of pages required.
5756 						 */
5757 						for (cvp = svp;  cvp < evp;
5758 						    cvp++) {
5759 							if (!VPP_ISSWAPRES(cvp))
5760 								sz++;
5761 						}
5762 						sz <<= PAGESHIFT;
5763 					}
5764 				}
5765 
5766 				/* Try to reserve the necessary swap. */
5767 				if (anon_resv_zone(sz,
5768 				    seg->s_as->a_proc->p_zone) == 0) {
5769 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5770 					return (IE_NOMEM);
5771 				}
5772 
5773 				/*
5774 				 * Make a note of how much swap space
5775 				 * we've reserved.
5776 				 */
5777 				if (svd->pageswap == 0 && sz == seg->s_size) {
5778 					svd->swresv = sz;
5779 				} else {
5780 					ASSERT(svd->vpage != NULL);
5781 					svd->swresv += sz;
5782 					svd->pageswap = 1;
5783 					for (cvp = svp; cvp < evp; cvp++) {
5784 						if (!VPP_ISSWAPRES(cvp))
5785 							VPP_SETSWAPRES(cvp);
5786 					}
5787 				}
5788 			}
5789 		} else {
5790 			/*
5791 			 * Swap space is released only if this segment
5792 			 * does not map anonymous memory, since read faults
5793 			 * on such segments still need an anon slot to read
5794 			 * in the data.
5795 			 */
5796 			if (svd->swresv != 0 && svd->vp != NULL &&
5797 			    svd->amp == NULL && addr == seg->s_base &&
5798 			    len == seg->s_size && svd->pageprot == 0) {
5799 				ASSERT(svd->pageswap == 0);
5800 				anon_unresv_zone(svd->swresv,
5801 				    seg->s_as->a_proc->p_zone);
5802 				svd->swresv = 0;
5803 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
5804 				    "anon proc:%p %lu %u", seg, 0, 0);
5805 			}
5806 		}
5807 	}
5808 
5809 	if (addr == seg->s_base && len == seg->s_size && svd->vpage == NULL) {
5810 		if (svd->prot == prot) {
5811 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5812 			return (0);			/* all done */
5813 		}
5814 		svd->prot = (uchar_t)prot;
5815 	} else if (svd->type == MAP_PRIVATE) {
5816 		struct anon *ap = NULL;
5817 		page_t *pp;
5818 		u_offset_t offset, off;
5819 		struct anon_map *amp;
5820 		ulong_t anon_idx = 0;
5821 
5822 		/*
5823 		 * A vpage structure exists or else the change does not
5824 		 * involve the entire segment.  Establish a vpage structure
5825 		 * if none is there.  Then, for each page in the range,
5826 		 * adjust its individual permissions.  Note that write-
5827 		 * enabling a MAP_PRIVATE page can affect the claims for
5828 		 * locked down memory.  Overcommitting memory terminates
5829 		 * the operation.
5830 		 */
5831 		segvn_vpage(seg);
5832 		svd->pageprot = 1;
5833 		if ((amp = svd->amp) != NULL) {
5834 			anon_idx = svd->anon_index + seg_page(seg, addr);
5835 			ASSERT(seg->s_szc == 0 ||
5836 			    IS_P2ALIGNED(anon_idx, pgcnt));
5837 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5838 		}
5839 
5840 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
5841 		evp = &svd->vpage[seg_page(seg, addr + len)];
5842 
5843 		/*
5844 		 * See Statement at the beginning of segvn_lockop regarding
5845 		 * the way cowcnts and lckcnts are handled.
5846 		 */
5847 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5848 
5849 			if (seg->s_szc != 0) {
5850 				if (amp != NULL) {
5851 					anon_array_enter(amp, anon_idx,
5852 					    &cookie);
5853 				}
5854 				if (IS_P2ALIGNED(anon_idx, pgcnt) &&
5855 				    !segvn_claim_pages(seg, svp, offset,
5856 				    anon_idx, prot)) {
5857 					if (amp != NULL) {
5858 						anon_array_exit(&cookie);
5859 					}
5860 					break;
5861 				}
5862 				if (amp != NULL) {
5863 					anon_array_exit(&cookie);
5864 				}
5865 				anon_idx++;
5866 			} else {
5867 				if (amp != NULL) {
5868 					anon_array_enter(amp, anon_idx,
5869 					    &cookie);
5870 					ap = anon_get_ptr(amp->ahp, anon_idx++);
5871 				}
5872 
5873 				if (VPP_ISPPLOCK(svp) &&
5874 				    VPP_PROT(svp) != prot) {
5875 
5876 					if (amp == NULL || ap == NULL) {
5877 						vp = svd->vp;
5878 						off = offset;
5879 					} else
5880 						swap_xlate(ap, &vp, &off);
5881 					if (amp != NULL)
5882 						anon_array_exit(&cookie);
5883 
5884 					if ((pp = page_lookup(vp, off,
5885 					    SE_SHARED)) == NULL) {
5886 						panic("segvn_setprot: no page");
5887 						/*NOTREACHED*/
5888 					}
5889 					ASSERT(seg->s_szc == 0);
5890 					if ((VPP_PROT(svp) ^ prot) &
5891 					    PROT_WRITE) {
5892 						if (prot & PROT_WRITE) {
5893 							if (!page_addclaim(
5894 							    pp)) {
5895 								page_unlock(pp);
5896 								break;
5897 							}
5898 						} else {
5899 							if (!page_subclaim(
5900 							    pp)) {
5901 								page_unlock(pp);
5902 								break;
5903 							}
5904 						}
5905 					}
5906 					page_unlock(pp);
5907 				} else if (amp != NULL)
5908 					anon_array_exit(&cookie);
5909 			}
5910 			VPP_SETPROT(svp, prot);
5911 			offset += PAGESIZE;
5912 		}
5913 		if (amp != NULL)
5914 			ANON_LOCK_EXIT(&amp->a_rwlock);
5915 
5916 		/*
5917 		 * Did we terminate prematurely?  If so, simply unload
5918 		 * the translations to the things we've updated so far.
5919 		 */
5920 		if (svp != evp) {
5921 			if (unload_done) {
5922 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5923 				return (IE_NOMEM);
5924 			}
5925 			len = (svp - &svd->vpage[seg_page(seg, addr)]) *
5926 			    PAGESIZE;
5927 			ASSERT(seg->s_szc == 0 || IS_P2ALIGNED(len, pgsz));
5928 			if (len != 0)
5929 				hat_unload(seg->s_as->a_hat, addr,
5930 				    len, HAT_UNLOAD);
5931 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5932 			return (IE_NOMEM);
5933 		}
5934 	} else {
5935 		segvn_vpage(seg);
5936 		svd->pageprot = 1;
5937 		evp = &svd->vpage[seg_page(seg, addr + len)];
5938 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5939 			VPP_SETPROT(svp, prot);
5940 		}
5941 	}
5942 
5943 	if (unload_done) {
5944 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5945 		return (0);
5946 	}
5947 
5948 	if (((prot & PROT_WRITE) != 0 &&
5949 	    (svd->vp != NULL || svd->type == MAP_PRIVATE)) ||
5950 	    (prot & ~PROT_USER) == PROT_NONE) {
5951 		/*
5952 		 * Either private or shared data with write access (in
5953 		 * which case we need to throw out all former translations
5954 		 * so that we get the right translations set up on fault
5955 		 * and we don't allow write access to any copy-on-write pages
5956 		 * that might be around or to prevent write access to pages
5957 		 * representing holes in a file), or we don't have permission
5958 		 * to access the memory at all (in which case we have to
5959 		 * unload any current translations that might exist).
5960 		 */
5961 		hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
5962 	} else {
5963 		/*
5964 		 * A shared mapping or a private mapping in which write
5965 		 * protection is going to be denied - just change all the
5966 		 * protections over the range of addresses in question.
5967 		 * segvn does not support any other attributes other
5968 		 * than prot so we can use hat_chgattr.
5969 		 */
5970 		hat_chgattr(seg->s_as->a_hat, addr, len, prot);
5971 	}
5972 
5973 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5974 
5975 	return (0);
5976 }
5977 
5978 /*
5979  * segvn_setpagesize is called via SEGOP_SETPAGESIZE from as_setpagesize,
5980  * to determine if the seg is capable of mapping the requested szc.
5981  */
5982 static int
5983 segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
5984 {
5985 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5986 	struct segvn_data *nsvd;
5987 	struct anon_map *amp = svd->amp;
5988 	struct seg *nseg;
5989 	caddr_t eaddr = addr + len, a;
5990 	size_t pgsz = page_get_pagesize(szc);
5991 	pgcnt_t pgcnt = page_get_pagecnt(szc);
5992 	int err;
5993 	u_offset_t off = svd->offset + (uintptr_t)(addr - seg->s_base);
5994 	extern struct vnode kvp;
5995 
5996 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5997 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
5998 
5999 	if (seg->s_szc == szc || segvn_lpg_disable != 0) {
6000 		return (0);
6001 	}
6002 
6003 	/*
6004 	 * addr should always be pgsz aligned but eaddr may be misaligned if
6005 	 * it's at the end of the segment.
6006 	 *
6007 	 * XXX we should assert this condition since as_setpagesize() logic
6008 	 * guarantees it.
6009 	 */
6010 	if (!IS_P2ALIGNED(addr, pgsz) ||
6011 	    (!IS_P2ALIGNED(eaddr, pgsz) &&
6012 	    eaddr != seg->s_base + seg->s_size)) {
6013 
6014 		segvn_setpgsz_align_err++;
6015 		return (EINVAL);
6016 	}
6017 
6018 	if (amp != NULL && svd->type == MAP_SHARED) {
6019 		ulong_t an_idx = svd->anon_index + seg_page(seg, addr);
6020 		if (!IS_P2ALIGNED(an_idx, pgcnt)) {
6021 
6022 			segvn_setpgsz_anon_align_err++;
6023 			return (EINVAL);
6024 		}
6025 	}
6026 
6027 	if ((svd->flags & MAP_NORESERVE) || seg->s_as == &kas ||
6028 	    szc > segvn_maxpgszc) {
6029 		return (EINVAL);
6030 	}
6031 
6032 	/* paranoid check */
6033 	if (svd->vp != NULL &&
6034 	    (IS_SWAPFSVP(svd->vp) || VN_ISKAS(svd->vp))) {
6035 		return (EINVAL);
6036 	}
6037 
6038 	if (seg->s_szc == 0 && svd->vp != NULL &&
6039 	    map_addr_vacalign_check(addr, off)) {
6040 		return (EINVAL);
6041 	}
6042 
6043 	/*
6044 	 * Check that protections are the same within new page
6045 	 * size boundaries.
6046 	 */
6047 	if (svd->pageprot) {
6048 		for (a = addr; a < eaddr; a += pgsz) {
6049 			if ((a + pgsz) > eaddr) {
6050 				if (!sameprot(seg, a, eaddr - a)) {
6051 					return (EINVAL);
6052 				}
6053 			} else {
6054 				if (!sameprot(seg, a, pgsz)) {
6055 					return (EINVAL);
6056 				}
6057 			}
6058 		}
6059 	}
6060 
6061 	/*
6062 	 * Since we are changing page size we first have to flush
6063 	 * the cache. This makes sure all the pagelock calls have
6064 	 * to recheck protections.
6065 	 */
6066 	if (svd->softlockcnt > 0) {
6067 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6068 
6069 		/*
6070 		 * If this is shared segment non 0 softlockcnt
6071 		 * means locked pages are still in use.
6072 		 */
6073 		if (svd->type == MAP_SHARED) {
6074 			return (EAGAIN);
6075 		}
6076 
6077 		/*
6078 		 * Since we do have the segvn writers lock nobody can fill
6079 		 * the cache with entries belonging to this seg during
6080 		 * the purge. The flush either succeeds or we still have
6081 		 * pending I/Os.
6082 		 */
6083 		segvn_purge(seg);
6084 		if (svd->softlockcnt > 0) {
6085 			return (EAGAIN);
6086 		}
6087 	}
6088 
6089 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6090 		ASSERT(svd->amp == NULL);
6091 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6092 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6093 		    HAT_REGION_TEXT);
6094 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6095 	} else if (svd->tr_state == SEGVN_TR_INIT) {
6096 		svd->tr_state = SEGVN_TR_OFF;
6097 	} else if (svd->tr_state == SEGVN_TR_ON) {
6098 		ASSERT(svd->amp != NULL);
6099 		segvn_textunrepl(seg, 1);
6100 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6101 		amp = NULL;
6102 	}
6103 
6104 	/*
6105 	 * Operation for sub range of existing segment.
6106 	 */
6107 	if (addr != seg->s_base || eaddr != (seg->s_base + seg->s_size)) {
6108 		if (szc < seg->s_szc) {
6109 			VM_STAT_ADD(segvnvmstats.demoterange[2]);
6110 			err = segvn_demote_range(seg, addr, len, SDR_RANGE, 0);
6111 			if (err == 0) {
6112 				return (IE_RETRY);
6113 			}
6114 			if (err == ENOMEM) {
6115 				return (IE_NOMEM);
6116 			}
6117 			return (err);
6118 		}
6119 		if (addr != seg->s_base) {
6120 			nseg = segvn_split_seg(seg, addr);
6121 			if (eaddr != (nseg->s_base + nseg->s_size)) {
6122 				/* eaddr is szc aligned */
6123 				(void) segvn_split_seg(nseg, eaddr);
6124 			}
6125 			return (IE_RETRY);
6126 		}
6127 		if (eaddr != (seg->s_base + seg->s_size)) {
6128 			/* eaddr is szc aligned */
6129 			(void) segvn_split_seg(seg, eaddr);
6130 		}
6131 		return (IE_RETRY);
6132 	}
6133 
6134 	/*
6135 	 * Break any low level sharing and reset seg->s_szc to 0.
6136 	 */
6137 	if ((err = segvn_clrszc(seg)) != 0) {
6138 		if (err == ENOMEM) {
6139 			err = IE_NOMEM;
6140 		}
6141 		return (err);
6142 	}
6143 	ASSERT(seg->s_szc == 0);
6144 
6145 	/*
6146 	 * If the end of the current segment is not pgsz aligned
6147 	 * then attempt to concatenate with the next segment.
6148 	 */
6149 	if (!IS_P2ALIGNED(eaddr, pgsz)) {
6150 		nseg = AS_SEGNEXT(seg->s_as, seg);
6151 		if (nseg == NULL || nseg == seg || eaddr != nseg->s_base) {
6152 			return (ENOMEM);
6153 		}
6154 		if (nseg->s_ops != &segvn_ops) {
6155 			return (EINVAL);
6156 		}
6157 		nsvd = (struct segvn_data *)nseg->s_data;
6158 		if (nsvd->softlockcnt > 0) {
6159 			/*
6160 			 * If this is shared segment non 0 softlockcnt
6161 			 * means locked pages are still in use.
6162 			 */
6163 			if (nsvd->type == MAP_SHARED) {
6164 				return (EAGAIN);
6165 			}
6166 			segvn_purge(nseg);
6167 			if (nsvd->softlockcnt > 0) {
6168 				return (EAGAIN);
6169 			}
6170 		}
6171 		err = segvn_clrszc(nseg);
6172 		if (err == ENOMEM) {
6173 			err = IE_NOMEM;
6174 		}
6175 		if (err != 0) {
6176 			return (err);
6177 		}
6178 		ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6179 		err = segvn_concat(seg, nseg, 1);
6180 		if (err == -1) {
6181 			return (EINVAL);
6182 		}
6183 		if (err == -2) {
6184 			return (IE_NOMEM);
6185 		}
6186 		return (IE_RETRY);
6187 	}
6188 
6189 	/*
6190 	 * May need to re-align anon array to
6191 	 * new szc.
6192 	 */
6193 	if (amp != NULL) {
6194 		if (!IS_P2ALIGNED(svd->anon_index, pgcnt)) {
6195 			struct anon_hdr *nahp;
6196 
6197 			ASSERT(svd->type == MAP_PRIVATE);
6198 
6199 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6200 			ASSERT(amp->refcnt == 1);
6201 			nahp = anon_create(btop(amp->size), ANON_NOSLEEP);
6202 			if (nahp == NULL) {
6203 				ANON_LOCK_EXIT(&amp->a_rwlock);
6204 				return (IE_NOMEM);
6205 			}
6206 			if (anon_copy_ptr(amp->ahp, svd->anon_index,
6207 			    nahp, 0, btop(seg->s_size), ANON_NOSLEEP)) {
6208 				anon_release(nahp, btop(amp->size));
6209 				ANON_LOCK_EXIT(&amp->a_rwlock);
6210 				return (IE_NOMEM);
6211 			}
6212 			anon_release(amp->ahp, btop(amp->size));
6213 			amp->ahp = nahp;
6214 			svd->anon_index = 0;
6215 			ANON_LOCK_EXIT(&amp->a_rwlock);
6216 		}
6217 	}
6218 	if (svd->vp != NULL && szc != 0) {
6219 		struct vattr va;
6220 		u_offset_t eoffpage = svd->offset;
6221 		va.va_mask = AT_SIZE;
6222 		eoffpage += seg->s_size;
6223 		eoffpage = btopr(eoffpage);
6224 		if (VOP_GETATTR(svd->vp, &va, 0, svd->cred, NULL) != 0) {
6225 			segvn_setpgsz_getattr_err++;
6226 			return (EINVAL);
6227 		}
6228 		if (btopr(va.va_size) < eoffpage) {
6229 			segvn_setpgsz_eof_err++;
6230 			return (EINVAL);
6231 		}
6232 		if (amp != NULL) {
6233 			/*
6234 			 * anon_fill_cow_holes() may call VOP_GETPAGE().
6235 			 * don't take anon map lock here to avoid holding it
6236 			 * across VOP_GETPAGE() calls that may call back into
6237 			 * segvn for klsutering checks. We don't really need
6238 			 * anon map lock here since it's a private segment and
6239 			 * we hold as level lock as writers.
6240 			 */
6241 			if ((err = anon_fill_cow_holes(seg, seg->s_base,
6242 			    amp->ahp, svd->anon_index, svd->vp, svd->offset,
6243 			    seg->s_size, szc, svd->prot, svd->vpage,
6244 			    svd->cred)) != 0) {
6245 				return (EINVAL);
6246 			}
6247 		}
6248 		segvn_setvnode_mpss(svd->vp);
6249 	}
6250 
6251 	if (amp != NULL) {
6252 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6253 		if (svd->type == MAP_PRIVATE) {
6254 			amp->a_szc = szc;
6255 		} else if (szc > amp->a_szc) {
6256 			amp->a_szc = szc;
6257 		}
6258 		ANON_LOCK_EXIT(&amp->a_rwlock);
6259 	}
6260 
6261 	seg->s_szc = szc;
6262 
6263 	return (0);
6264 }
6265 
6266 static int
6267 segvn_clrszc(struct seg *seg)
6268 {
6269 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6270 	struct anon_map *amp = svd->amp;
6271 	size_t pgsz;
6272 	pgcnt_t pages;
6273 	int err = 0;
6274 	caddr_t a = seg->s_base;
6275 	caddr_t ea = a + seg->s_size;
6276 	ulong_t an_idx = svd->anon_index;
6277 	vnode_t *vp = svd->vp;
6278 	struct vpage *vpage = svd->vpage;
6279 	page_t *anon_pl[1 + 1], *pp;
6280 	struct anon *ap, *oldap;
6281 	uint_t prot = svd->prot, vpprot;
6282 	int pageflag = 0;
6283 
6284 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6285 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
6286 	ASSERT(svd->softlockcnt == 0);
6287 
6288 	if (vp == NULL && amp == NULL) {
6289 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6290 		seg->s_szc = 0;
6291 		return (0);
6292 	}
6293 
6294 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6295 		ASSERT(svd->amp == NULL);
6296 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6297 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6298 		    HAT_REGION_TEXT);
6299 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6300 	} else if (svd->tr_state == SEGVN_TR_ON) {
6301 		ASSERT(svd->amp != NULL);
6302 		segvn_textunrepl(seg, 1);
6303 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6304 		amp = NULL;
6305 	} else {
6306 		if (svd->tr_state != SEGVN_TR_OFF) {
6307 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
6308 			svd->tr_state = SEGVN_TR_OFF;
6309 		}
6310 
6311 		/*
6312 		 * do HAT_UNLOAD_UNMAP since we are changing the pagesize.
6313 		 * unload argument is 0 when we are freeing the segment
6314 		 * and unload was already done.
6315 		 */
6316 		hat_unload(seg->s_as->a_hat, seg->s_base, seg->s_size,
6317 		    HAT_UNLOAD_UNMAP);
6318 	}
6319 
6320 	if (amp == NULL || svd->type == MAP_SHARED) {
6321 		seg->s_szc = 0;
6322 		return (0);
6323 	}
6324 
6325 	pgsz = page_get_pagesize(seg->s_szc);
6326 	pages = btop(pgsz);
6327 
6328 	/*
6329 	 * XXX anon rwlock is not really needed because this is a
6330 	 * private segment and we are writers.
6331 	 */
6332 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6333 
6334 	for (; a < ea; a += pgsz, an_idx += pages) {
6335 		if ((oldap = anon_get_ptr(amp->ahp, an_idx)) != NULL) {
6336 			ASSERT(vpage != NULL || svd->pageprot == 0);
6337 			if (vpage != NULL) {
6338 				ASSERT(sameprot(seg, a, pgsz));
6339 				prot = VPP_PROT(vpage);
6340 				pageflag = VPP_ISPPLOCK(vpage) ? LOCK_PAGE : 0;
6341 			}
6342 			if (seg->s_szc != 0) {
6343 				ASSERT(vp == NULL || anon_pages(amp->ahp,
6344 				    an_idx, pages) == pages);
6345 				if ((err = anon_map_demotepages(amp, an_idx,
6346 				    seg, a, prot, vpage, svd->cred)) != 0) {
6347 					goto out;
6348 				}
6349 			} else {
6350 				if (oldap->an_refcnt == 1) {
6351 					continue;
6352 				}
6353 				if ((err = anon_getpage(&oldap, &vpprot,
6354 				    anon_pl, PAGESIZE, seg, a, S_READ,
6355 				    svd->cred))) {
6356 					goto out;
6357 				}
6358 				if ((pp = anon_private(&ap, seg, a, prot,
6359 				    anon_pl[0], pageflag, svd->cred)) == NULL) {
6360 					err = ENOMEM;
6361 					goto out;
6362 				}
6363 				anon_decref(oldap);
6364 				(void) anon_set_ptr(amp->ahp, an_idx, ap,
6365 				    ANON_SLEEP);
6366 				page_unlock(pp);
6367 			}
6368 		}
6369 		vpage = (vpage == NULL) ? NULL : vpage + pages;
6370 	}
6371 
6372 	amp->a_szc = 0;
6373 	seg->s_szc = 0;
6374 out:
6375 	ANON_LOCK_EXIT(&amp->a_rwlock);
6376 	return (err);
6377 }
6378 
6379 static int
6380 segvn_claim_pages(
6381 	struct seg *seg,
6382 	struct vpage *svp,
6383 	u_offset_t off,
6384 	ulong_t anon_idx,
6385 	uint_t prot)
6386 {
6387 	pgcnt_t	pgcnt = page_get_pagecnt(seg->s_szc);
6388 	size_t ppasize = (pgcnt + 1) * sizeof (page_t *);
6389 	page_t	**ppa;
6390 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6391 	struct anon_map *amp = svd->amp;
6392 	struct vpage *evp = svp + pgcnt;
6393 	caddr_t addr = ((uintptr_t)(svp - svd->vpage) << PAGESHIFT)
6394 	    + seg->s_base;
6395 	struct anon *ap;
6396 	struct vnode *vp = svd->vp;
6397 	page_t *pp;
6398 	pgcnt_t pg_idx, i;
6399 	int err = 0;
6400 	anoff_t aoff;
6401 	int anon = (amp != NULL) ? 1 : 0;
6402 
6403 	ASSERT(svd->type == MAP_PRIVATE);
6404 	ASSERT(svd->vpage != NULL);
6405 	ASSERT(seg->s_szc != 0);
6406 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
6407 	ASSERT(amp == NULL || IS_P2ALIGNED(anon_idx, pgcnt));
6408 	ASSERT(sameprot(seg, addr, pgcnt << PAGESHIFT));
6409 
6410 	if (VPP_PROT(svp) == prot)
6411 		return (1);
6412 	if (!((VPP_PROT(svp) ^ prot) & PROT_WRITE))
6413 		return (1);
6414 
6415 	ppa = kmem_alloc(ppasize, KM_SLEEP);
6416 	if (anon && vp != NULL) {
6417 		if (anon_get_ptr(amp->ahp, anon_idx) == NULL) {
6418 			anon = 0;
6419 			ASSERT(!anon_pages(amp->ahp, anon_idx, pgcnt));
6420 		}
6421 		ASSERT(!anon ||
6422 		    anon_pages(amp->ahp, anon_idx, pgcnt) == pgcnt);
6423 	}
6424 
6425 	for (*ppa = NULL, pg_idx = 0; svp < evp; svp++, anon_idx++) {
6426 		if (!VPP_ISPPLOCK(svp))
6427 			continue;
6428 		if (anon) {
6429 			ap = anon_get_ptr(amp->ahp, anon_idx);
6430 			if (ap == NULL) {
6431 				panic("segvn_claim_pages: no anon slot");
6432 			}
6433 			swap_xlate(ap, &vp, &aoff);
6434 			off = (u_offset_t)aoff;
6435 		}
6436 		ASSERT(vp != NULL);
6437 		if ((pp = page_lookup(vp,
6438 		    (u_offset_t)off, SE_SHARED)) == NULL) {
6439 			panic("segvn_claim_pages: no page");
6440 		}
6441 		ppa[pg_idx++] = pp;
6442 		off += PAGESIZE;
6443 	}
6444 
6445 	if (ppa[0] == NULL) {
6446 		kmem_free(ppa, ppasize);
6447 		return (1);
6448 	}
6449 
6450 	ASSERT(pg_idx <= pgcnt);
6451 	ppa[pg_idx] = NULL;
6452 
6453 	if (prot & PROT_WRITE)
6454 		err = page_addclaim_pages(ppa);
6455 	else
6456 		err = page_subclaim_pages(ppa);
6457 
6458 	for (i = 0; i < pg_idx; i++) {
6459 		ASSERT(ppa[i] != NULL);
6460 		page_unlock(ppa[i]);
6461 	}
6462 
6463 	kmem_free(ppa, ppasize);
6464 	return (err);
6465 }
6466 
6467 /*
6468  * Returns right (upper address) segment if split occurred.
6469  * If the address is equal to the beginning or end of its segment it returns
6470  * the current segment.
6471  */
6472 static struct seg *
6473 segvn_split_seg(struct seg *seg, caddr_t addr)
6474 {
6475 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6476 	struct seg *nseg;
6477 	size_t nsize;
6478 	struct segvn_data *nsvd;
6479 
6480 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6481 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6482 
6483 	ASSERT(addr >= seg->s_base);
6484 	ASSERT(addr <= seg->s_base + seg->s_size);
6485 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6486 
6487 	if (addr == seg->s_base || addr == seg->s_base + seg->s_size)
6488 		return (seg);
6489 
6490 	nsize = seg->s_base + seg->s_size - addr;
6491 	seg->s_size = addr - seg->s_base;
6492 	nseg = seg_alloc(seg->s_as, addr, nsize);
6493 	ASSERT(nseg != NULL);
6494 	nseg->s_ops = seg->s_ops;
6495 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
6496 	nseg->s_data = (void *)nsvd;
6497 	nseg->s_szc = seg->s_szc;
6498 	*nsvd = *svd;
6499 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6500 	nsvd->seg = nseg;
6501 	rw_init(&nsvd->lock, NULL, RW_DEFAULT, NULL);
6502 
6503 	if (nsvd->vp != NULL) {
6504 		VN_HOLD(nsvd->vp);
6505 		nsvd->offset = svd->offset +
6506 		    (uintptr_t)(nseg->s_base - seg->s_base);
6507 		if (nsvd->type == MAP_SHARED)
6508 			lgrp_shm_policy_init(NULL, nsvd->vp);
6509 	} else {
6510 		/*
6511 		 * The offset for an anonymous segment has no signifigance in
6512 		 * terms of an offset into a file. If we were to use the above
6513 		 * calculation instead, the structures read out of
6514 		 * /proc/<pid>/xmap would be more difficult to decipher since
6515 		 * it would be unclear whether two seemingly contiguous
6516 		 * prxmap_t structures represented different segments or a
6517 		 * single segment that had been split up into multiple prxmap_t
6518 		 * structures (e.g. if some part of the segment had not yet
6519 		 * been faulted in).
6520 		 */
6521 		nsvd->offset = 0;
6522 	}
6523 
6524 	ASSERT(svd->softlockcnt == 0);
6525 	ASSERT(svd->softlockcnt_sbase == 0);
6526 	ASSERT(svd->softlockcnt_send == 0);
6527 	crhold(svd->cred);
6528 
6529 	if (svd->vpage != NULL) {
6530 		size_t bytes = vpgtob(seg_pages(seg));
6531 		size_t nbytes = vpgtob(seg_pages(nseg));
6532 		struct vpage *ovpage = svd->vpage;
6533 
6534 		svd->vpage = kmem_alloc(bytes, KM_SLEEP);
6535 		bcopy(ovpage, svd->vpage, bytes);
6536 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
6537 		bcopy(ovpage + seg_pages(seg), nsvd->vpage, nbytes);
6538 		kmem_free(ovpage, bytes + nbytes);
6539 	}
6540 	if (svd->amp != NULL && svd->type == MAP_PRIVATE) {
6541 		struct anon_map *oamp = svd->amp, *namp;
6542 		struct anon_hdr *nahp;
6543 
6544 		ANON_LOCK_ENTER(&oamp->a_rwlock, RW_WRITER);
6545 		ASSERT(oamp->refcnt == 1);
6546 		nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
6547 		(void) anon_copy_ptr(oamp->ahp, svd->anon_index,
6548 		    nahp, 0, btop(seg->s_size), ANON_SLEEP);
6549 
6550 		namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
6551 		namp->a_szc = nseg->s_szc;
6552 		(void) anon_copy_ptr(oamp->ahp,
6553 		    svd->anon_index + btop(seg->s_size),
6554 		    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
6555 		anon_release(oamp->ahp, btop(oamp->size));
6556 		oamp->ahp = nahp;
6557 		oamp->size = seg->s_size;
6558 		svd->anon_index = 0;
6559 		nsvd->amp = namp;
6560 		nsvd->anon_index = 0;
6561 		ANON_LOCK_EXIT(&oamp->a_rwlock);
6562 	} else if (svd->amp != NULL) {
6563 		pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6564 		ASSERT(svd->amp == nsvd->amp);
6565 		ASSERT(seg->s_szc <= svd->amp->a_szc);
6566 		nsvd->anon_index = svd->anon_index + seg_pages(seg);
6567 		ASSERT(IS_P2ALIGNED(nsvd->anon_index, pgcnt));
6568 		ANON_LOCK_ENTER(&svd->amp->a_rwlock, RW_WRITER);
6569 		svd->amp->refcnt++;
6570 		ANON_LOCK_EXIT(&svd->amp->a_rwlock);
6571 	}
6572 
6573 	/*
6574 	 * Split the amount of swap reserved.
6575 	 */
6576 	if (svd->swresv) {
6577 		/*
6578 		 * For MAP_NORESERVE, only allocate swap reserve for pages
6579 		 * being used.  Other segments get enough to cover whole
6580 		 * segment.
6581 		 */
6582 		if (svd->flags & MAP_NORESERVE) {
6583 			size_t	oswresv;
6584 
6585 			ASSERT(svd->amp);
6586 			oswresv = svd->swresv;
6587 			svd->swresv = ptob(anon_pages(svd->amp->ahp,
6588 			    svd->anon_index, btop(seg->s_size)));
6589 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
6590 			    nsvd->anon_index, btop(nseg->s_size)));
6591 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
6592 		} else {
6593 			if (svd->pageswap) {
6594 				svd->swresv = segvn_count_swap_by_vpages(seg);
6595 				ASSERT(nsvd->swresv >= svd->swresv);
6596 				nsvd->swresv -= svd->swresv;
6597 			} else {
6598 				ASSERT(svd->swresv == seg->s_size +
6599 				    nseg->s_size);
6600 				svd->swresv = seg->s_size;
6601 				nsvd->swresv = nseg->s_size;
6602 			}
6603 		}
6604 	}
6605 
6606 	return (nseg);
6607 }
6608 
6609 /*
6610  * called on memory operations (unmap, setprot, setpagesize) for a subset
6611  * of a large page segment to either demote the memory range (SDR_RANGE)
6612  * or the ends (SDR_END) by addr/len.
6613  *
6614  * returns 0 on success. returns errno, including ENOMEM, on failure.
6615  */
6616 static int
6617 segvn_demote_range(
6618 	struct seg *seg,
6619 	caddr_t addr,
6620 	size_t len,
6621 	int flag,
6622 	uint_t szcvec)
6623 {
6624 	caddr_t eaddr = addr + len;
6625 	caddr_t lpgaddr, lpgeaddr;
6626 	struct seg *nseg;
6627 	struct seg *badseg1 = NULL;
6628 	struct seg *badseg2 = NULL;
6629 	size_t pgsz;
6630 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6631 	int err;
6632 	uint_t szc = seg->s_szc;
6633 	uint_t tszcvec;
6634 
6635 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6636 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6637 	ASSERT(szc != 0);
6638 	pgsz = page_get_pagesize(szc);
6639 	ASSERT(seg->s_base != addr || seg->s_size != len);
6640 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6641 	ASSERT(svd->softlockcnt == 0);
6642 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6643 	ASSERT(szcvec == 0 || (flag == SDR_END && svd->type == MAP_SHARED));
6644 
6645 	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
6646 	ASSERT(flag == SDR_RANGE || eaddr < lpgeaddr || addr > lpgaddr);
6647 	if (flag == SDR_RANGE) {
6648 		/* demote entire range */
6649 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6650 		(void) segvn_split_seg(nseg, lpgeaddr);
6651 		ASSERT(badseg1->s_base == lpgaddr);
6652 		ASSERT(badseg1->s_size == lpgeaddr - lpgaddr);
6653 	} else if (addr != lpgaddr) {
6654 		ASSERT(flag == SDR_END);
6655 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6656 		if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz &&
6657 		    eaddr < lpgaddr + 2 * pgsz) {
6658 			(void) segvn_split_seg(nseg, lpgeaddr);
6659 			ASSERT(badseg1->s_base == lpgaddr);
6660 			ASSERT(badseg1->s_size == 2 * pgsz);
6661 		} else {
6662 			nseg = segvn_split_seg(nseg, lpgaddr + pgsz);
6663 			ASSERT(badseg1->s_base == lpgaddr);
6664 			ASSERT(badseg1->s_size == pgsz);
6665 			if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz) {
6666 				ASSERT(lpgeaddr - lpgaddr > 2 * pgsz);
6667 				nseg = segvn_split_seg(nseg, lpgeaddr - pgsz);
6668 				badseg2 = nseg;
6669 				(void) segvn_split_seg(nseg, lpgeaddr);
6670 				ASSERT(badseg2->s_base == lpgeaddr - pgsz);
6671 				ASSERT(badseg2->s_size == pgsz);
6672 			}
6673 		}
6674 	} else {
6675 		ASSERT(flag == SDR_END);
6676 		ASSERT(eaddr < lpgeaddr);
6677 		badseg1 = nseg = segvn_split_seg(seg, lpgeaddr - pgsz);
6678 		(void) segvn_split_seg(nseg, lpgeaddr);
6679 		ASSERT(badseg1->s_base == lpgeaddr - pgsz);
6680 		ASSERT(badseg1->s_size == pgsz);
6681 	}
6682 
6683 	ASSERT(badseg1 != NULL);
6684 	ASSERT(badseg1->s_szc == szc);
6685 	ASSERT(flag == SDR_RANGE || badseg1->s_size == pgsz ||
6686 	    badseg1->s_size == 2 * pgsz);
6687 	ASSERT(sameprot(badseg1, badseg1->s_base, pgsz));
6688 	ASSERT(badseg1->s_size == pgsz ||
6689 	    sameprot(badseg1, badseg1->s_base + pgsz, pgsz));
6690 	if (err = segvn_clrszc(badseg1)) {
6691 		return (err);
6692 	}
6693 	ASSERT(badseg1->s_szc == 0);
6694 
6695 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6696 		uint_t tszc = highbit(tszcvec) - 1;
6697 		caddr_t ta = MAX(addr, badseg1->s_base);
6698 		caddr_t te;
6699 		size_t tpgsz = page_get_pagesize(tszc);
6700 
6701 		ASSERT(svd->type == MAP_SHARED);
6702 		ASSERT(flag == SDR_END);
6703 		ASSERT(tszc < szc && tszc > 0);
6704 
6705 		if (eaddr > badseg1->s_base + badseg1->s_size) {
6706 			te = badseg1->s_base + badseg1->s_size;
6707 		} else {
6708 			te = eaddr;
6709 		}
6710 
6711 		ASSERT(ta <= te);
6712 		badseg1->s_szc = tszc;
6713 		if (!IS_P2ALIGNED(ta, tpgsz) || !IS_P2ALIGNED(te, tpgsz)) {
6714 			if (badseg2 != NULL) {
6715 				err = segvn_demote_range(badseg1, ta, te - ta,
6716 				    SDR_END, tszcvec);
6717 				if (err != 0) {
6718 					return (err);
6719 				}
6720 			} else {
6721 				return (segvn_demote_range(badseg1, ta,
6722 				    te - ta, SDR_END, tszcvec));
6723 			}
6724 		}
6725 	}
6726 
6727 	if (badseg2 == NULL)
6728 		return (0);
6729 	ASSERT(badseg2->s_szc == szc);
6730 	ASSERT(badseg2->s_size == pgsz);
6731 	ASSERT(sameprot(badseg2, badseg2->s_base, badseg2->s_size));
6732 	if (err = segvn_clrszc(badseg2)) {
6733 		return (err);
6734 	}
6735 	ASSERT(badseg2->s_szc == 0);
6736 
6737 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6738 		uint_t tszc = highbit(tszcvec) - 1;
6739 		size_t tpgsz = page_get_pagesize(tszc);
6740 
6741 		ASSERT(svd->type == MAP_SHARED);
6742 		ASSERT(flag == SDR_END);
6743 		ASSERT(tszc < szc && tszc > 0);
6744 		ASSERT(badseg2->s_base > addr);
6745 		ASSERT(eaddr > badseg2->s_base);
6746 		ASSERT(eaddr < badseg2->s_base + badseg2->s_size);
6747 
6748 		badseg2->s_szc = tszc;
6749 		if (!IS_P2ALIGNED(eaddr, tpgsz)) {
6750 			return (segvn_demote_range(badseg2, badseg2->s_base,
6751 			    eaddr - badseg2->s_base, SDR_END, tszcvec));
6752 		}
6753 	}
6754 
6755 	return (0);
6756 }
6757 
6758 static int
6759 segvn_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
6760 {
6761 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6762 	struct vpage *vp, *evp;
6763 
6764 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6765 
6766 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6767 	/*
6768 	 * If segment protection can be used, simply check against them.
6769 	 */
6770 	if (svd->pageprot == 0) {
6771 		int err;
6772 
6773 		err = ((svd->prot & prot) != prot) ? EACCES : 0;
6774 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6775 		return (err);
6776 	}
6777 
6778 	/*
6779 	 * Have to check down to the vpage level.
6780 	 */
6781 	evp = &svd->vpage[seg_page(seg, addr + len)];
6782 	for (vp = &svd->vpage[seg_page(seg, addr)]; vp < evp; vp++) {
6783 		if ((VPP_PROT(vp) & prot) != prot) {
6784 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6785 			return (EACCES);
6786 		}
6787 	}
6788 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6789 	return (0);
6790 }
6791 
6792 static int
6793 segvn_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
6794 {
6795 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6796 	size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
6797 
6798 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6799 
6800 	if (pgno != 0) {
6801 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6802 		if (svd->pageprot == 0) {
6803 			do {
6804 				protv[--pgno] = svd->prot;
6805 			} while (pgno != 0);
6806 		} else {
6807 			size_t pgoff = seg_page(seg, addr);
6808 
6809 			do {
6810 				pgno--;
6811 				protv[pgno] = VPP_PROT(&svd->vpage[pgno+pgoff]);
6812 			} while (pgno != 0);
6813 		}
6814 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6815 	}
6816 	return (0);
6817 }
6818 
6819 static u_offset_t
6820 segvn_getoffset(struct seg *seg, caddr_t addr)
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 	return (svd->offset + (uintptr_t)(addr - seg->s_base));
6827 }
6828 
6829 /*ARGSUSED*/
6830 static int
6831 segvn_gettype(struct seg *seg, caddr_t addr)
6832 {
6833 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6834 
6835 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6836 
6837 	return (svd->type | (svd->flags & (MAP_NORESERVE | MAP_TEXT |
6838 	    MAP_INITDATA)));
6839 }
6840 
6841 /*ARGSUSED*/
6842 static int
6843 segvn_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
6844 {
6845 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6846 
6847 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6848 
6849 	*vpp = svd->vp;
6850 	return (0);
6851 }
6852 
6853 /*
6854  * Check to see if it makes sense to do kluster/read ahead to
6855  * addr + delta relative to the mapping at addr.  We assume here
6856  * that delta is a signed PAGESIZE'd multiple (which can be negative).
6857  *
6858  * For segvn, we currently "approve" of the action if we are
6859  * still in the segment and it maps from the same vp/off,
6860  * or if the advice stored in segvn_data or vpages allows it.
6861  * Currently, klustering is not allowed only if MADV_RANDOM is set.
6862  */
6863 static int
6864 segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
6865 {
6866 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6867 	struct anon *oap, *ap;
6868 	ssize_t pd;
6869 	size_t page;
6870 	struct vnode *vp1, *vp2;
6871 	u_offset_t off1, off2;
6872 	struct anon_map *amp;
6873 
6874 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6875 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6876 	    SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
6877 
6878 	if (addr + delta < seg->s_base ||
6879 	    addr + delta >= (seg->s_base + seg->s_size))
6880 		return (-1);		/* exceeded segment bounds */
6881 
6882 	pd = delta / (ssize_t)PAGESIZE;	/* divide to preserve sign bit */
6883 	page = seg_page(seg, addr);
6884 
6885 	/*
6886 	 * Check to see if either of the pages addr or addr + delta
6887 	 * have advice set that prevents klustering (if MADV_RANDOM advice
6888 	 * is set for entire segment, or MADV_SEQUENTIAL is set and delta
6889 	 * is negative).
6890 	 */
6891 	if (svd->advice == MADV_RANDOM ||
6892 	    svd->advice == MADV_SEQUENTIAL && delta < 0)
6893 		return (-1);
6894 	else if (svd->pageadvice && svd->vpage) {
6895 		struct vpage *bvpp, *evpp;
6896 
6897 		bvpp = &svd->vpage[page];
6898 		evpp = &svd->vpage[page + pd];
6899 		if (VPP_ADVICE(bvpp) == MADV_RANDOM ||
6900 		    VPP_ADVICE(evpp) == MADV_SEQUENTIAL && delta < 0)
6901 			return (-1);
6902 		if (VPP_ADVICE(bvpp) != VPP_ADVICE(evpp) &&
6903 		    VPP_ADVICE(evpp) == MADV_RANDOM)
6904 			return (-1);
6905 	}
6906 
6907 	if (svd->type == MAP_SHARED)
6908 		return (0);		/* shared mapping - all ok */
6909 
6910 	if ((amp = svd->amp) == NULL)
6911 		return (0);		/* off original vnode */
6912 
6913 	page += svd->anon_index;
6914 
6915 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
6916 
6917 	oap = anon_get_ptr(amp->ahp, page);
6918 	ap = anon_get_ptr(amp->ahp, page + pd);
6919 
6920 	ANON_LOCK_EXIT(&amp->a_rwlock);
6921 
6922 	if ((oap == NULL && ap != NULL) || (oap != NULL && ap == NULL)) {
6923 		return (-1);		/* one with and one without an anon */
6924 	}
6925 
6926 	if (oap == NULL) {		/* implies that ap == NULL */
6927 		return (0);		/* off original vnode */
6928 	}
6929 
6930 	/*
6931 	 * Now we know we have two anon pointers - check to
6932 	 * see if they happen to be properly allocated.
6933 	 */
6934 
6935 	/*
6936 	 * XXX We cheat here and don't lock the anon slots. We can't because
6937 	 * we may have been called from the anon layer which might already
6938 	 * have locked them. We are holding a refcnt on the slots so they
6939 	 * can't disappear. The worst that will happen is we'll get the wrong
6940 	 * names (vp, off) for the slots and make a poor klustering decision.
6941 	 */
6942 	swap_xlate(ap, &vp1, &off1);
6943 	swap_xlate(oap, &vp2, &off2);
6944 
6945 
6946 	if (!VOP_CMP(vp1, vp2, NULL) || off1 - off2 != delta)
6947 		return (-1);
6948 	return (0);
6949 }
6950 
6951 /*
6952  * Swap the pages of seg out to secondary storage, returning the
6953  * number of bytes of storage freed.
6954  *
6955  * The basic idea is first to unload all translations and then to call
6956  * VOP_PUTPAGE() for all newly-unmapped pages, to push them out to the
6957  * swap device.  Pages to which other segments have mappings will remain
6958  * mapped and won't be swapped.  Our caller (as_swapout) has already
6959  * performed the unloading step.
6960  *
6961  * The value returned is intended to correlate well with the process's
6962  * memory requirements.  However, there are some caveats:
6963  * 1)	When given a shared segment as argument, this routine will
6964  *	only succeed in swapping out pages for the last sharer of the
6965  *	segment.  (Previous callers will only have decremented mapping
6966  *	reference counts.)
6967  * 2)	We assume that the hat layer maintains a large enough translation
6968  *	cache to capture process reference patterns.
6969  */
6970 static size_t
6971 segvn_swapout(struct seg *seg)
6972 {
6973 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6974 	struct anon_map *amp;
6975 	pgcnt_t pgcnt = 0;
6976 	pgcnt_t npages;
6977 	pgcnt_t page;
6978 	ulong_t anon_index;
6979 
6980 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6981 
6982 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6983 	/*
6984 	 * Find pages unmapped by our caller and force them
6985 	 * out to the virtual swap device.
6986 	 */
6987 	if ((amp = svd->amp) != NULL)
6988 		anon_index = svd->anon_index;
6989 	npages = seg->s_size >> PAGESHIFT;
6990 	for (page = 0; page < npages; page++) {
6991 		page_t *pp;
6992 		struct anon *ap;
6993 		struct vnode *vp;
6994 		u_offset_t off;
6995 		anon_sync_obj_t cookie;
6996 
6997 		/*
6998 		 * Obtain <vp, off> pair for the page, then look it up.
6999 		 *
7000 		 * Note that this code is willing to consider regular
7001 		 * pages as well as anon pages.  Is this appropriate here?
7002 		 */
7003 		ap = NULL;
7004 		if (amp != NULL) {
7005 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7006 			if (anon_array_try_enter(amp, anon_index + page,
7007 			    &cookie)) {
7008 				ANON_LOCK_EXIT(&amp->a_rwlock);
7009 				continue;
7010 			}
7011 			ap = anon_get_ptr(amp->ahp, anon_index + page);
7012 			if (ap != NULL) {
7013 				swap_xlate(ap, &vp, &off);
7014 			} else {
7015 				vp = svd->vp;
7016 				off = svd->offset + ptob(page);
7017 			}
7018 			anon_array_exit(&cookie);
7019 			ANON_LOCK_EXIT(&amp->a_rwlock);
7020 		} else {
7021 			vp = svd->vp;
7022 			off = svd->offset + ptob(page);
7023 		}
7024 		if (vp == NULL) {		/* untouched zfod page */
7025 			ASSERT(ap == NULL);
7026 			continue;
7027 		}
7028 
7029 		pp = page_lookup_nowait(vp, off, SE_SHARED);
7030 		if (pp == NULL)
7031 			continue;
7032 
7033 
7034 		/*
7035 		 * Examine the page to see whether it can be tossed out,
7036 		 * keeping track of how many we've found.
7037 		 */
7038 		if (!page_tryupgrade(pp)) {
7039 			/*
7040 			 * If the page has an i/o lock and no mappings,
7041 			 * it's very likely that the page is being
7042 			 * written out as a result of klustering.
7043 			 * Assume this is so and take credit for it here.
7044 			 */
7045 			if (!page_io_trylock(pp)) {
7046 				if (!hat_page_is_mapped(pp))
7047 					pgcnt++;
7048 			} else {
7049 				page_io_unlock(pp);
7050 			}
7051 			page_unlock(pp);
7052 			continue;
7053 		}
7054 		ASSERT(!page_iolock_assert(pp));
7055 
7056 
7057 		/*
7058 		 * Skip if page is locked or has mappings.
7059 		 * We don't need the page_struct_lock to look at lckcnt
7060 		 * and cowcnt because the page is exclusive locked.
7061 		 */
7062 		if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
7063 		    hat_page_is_mapped(pp)) {
7064 			page_unlock(pp);
7065 			continue;
7066 		}
7067 
7068 		/*
7069 		 * dispose skips large pages so try to demote first.
7070 		 */
7071 		if (pp->p_szc != 0 && !page_try_demote_pages(pp)) {
7072 			page_unlock(pp);
7073 			/*
7074 			 * XXX should skip the remaining page_t's of this
7075 			 * large page.
7076 			 */
7077 			continue;
7078 		}
7079 
7080 		ASSERT(pp->p_szc == 0);
7081 
7082 		/*
7083 		 * No longer mapped -- we can toss it out.  How
7084 		 * we do so depends on whether or not it's dirty.
7085 		 */
7086 		if (hat_ismod(pp) && pp->p_vnode) {
7087 			/*
7088 			 * We must clean the page before it can be
7089 			 * freed.  Setting B_FREE will cause pvn_done
7090 			 * to free the page when the i/o completes.
7091 			 * XXX:	This also causes it to be accounted
7092 			 *	as a pageout instead of a swap: need
7093 			 *	B_SWAPOUT bit to use instead of B_FREE.
7094 			 *
7095 			 * Hold the vnode before releasing the page lock
7096 			 * to prevent it from being freed and re-used by
7097 			 * some other thread.
7098 			 */
7099 			VN_HOLD(vp);
7100 			page_unlock(pp);
7101 
7102 			/*
7103 			 * Queue all i/o requests for the pageout thread
7104 			 * to avoid saturating the pageout devices.
7105 			 */
7106 			if (!queue_io_request(vp, off))
7107 				VN_RELE(vp);
7108 		} else {
7109 			/*
7110 			 * The page was clean, free it.
7111 			 *
7112 			 * XXX:	Can we ever encounter modified pages
7113 			 *	with no associated vnode here?
7114 			 */
7115 			ASSERT(pp->p_vnode != NULL);
7116 			/*LINTED: constant in conditional context*/
7117 			VN_DISPOSE(pp, B_FREE, 0, kcred);
7118 		}
7119 
7120 		/*
7121 		 * Credit now even if i/o is in progress.
7122 		 */
7123 		pgcnt++;
7124 	}
7125 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7126 
7127 	/*
7128 	 * Wakeup pageout to initiate i/o on all queued requests.
7129 	 */
7130 	cv_signal_pageout();
7131 	return (ptob(pgcnt));
7132 }
7133 
7134 /*
7135  * Synchronize primary storage cache with real object in virtual memory.
7136  *
7137  * XXX - Anonymous pages should not be sync'ed out at all.
7138  */
7139 static int
7140 segvn_sync(struct seg *seg, caddr_t addr, size_t len, int attr, uint_t flags)
7141 {
7142 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7143 	struct vpage *vpp;
7144 	page_t *pp;
7145 	u_offset_t offset;
7146 	struct vnode *vp;
7147 	u_offset_t off;
7148 	caddr_t eaddr;
7149 	int bflags;
7150 	int err = 0;
7151 	int segtype;
7152 	int pageprot;
7153 	int prot;
7154 	ulong_t anon_index;
7155 	struct anon_map *amp;
7156 	struct anon *ap;
7157 	anon_sync_obj_t cookie;
7158 
7159 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7160 
7161 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7162 
7163 	if (svd->softlockcnt > 0) {
7164 		/*
7165 		 * If this is shared segment non 0 softlockcnt
7166 		 * means locked pages are still in use.
7167 		 */
7168 		if (svd->type == MAP_SHARED) {
7169 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7170 			return (EAGAIN);
7171 		}
7172 
7173 		/*
7174 		 * flush all pages from seg cache
7175 		 * otherwise we may deadlock in swap_putpage
7176 		 * for B_INVAL page (4175402).
7177 		 *
7178 		 * Even if we grab segvn WRITER's lock
7179 		 * here, there might be another thread which could've
7180 		 * successfully performed lookup/insert just before
7181 		 * we acquired the lock here.  So, grabbing either
7182 		 * lock here is of not much use.  Until we devise
7183 		 * a strategy at upper layers to solve the
7184 		 * synchronization issues completely, we expect
7185 		 * applications to handle this appropriately.
7186 		 */
7187 		segvn_purge(seg);
7188 		if (svd->softlockcnt > 0) {
7189 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7190 			return (EAGAIN);
7191 		}
7192 	} else if (svd->type == MAP_SHARED && svd->amp != NULL &&
7193 	    svd->amp->a_softlockcnt > 0) {
7194 		/*
7195 		 * Try to purge this amp's entries from pcache. It will
7196 		 * succeed only if other segments that share the amp have no
7197 		 * outstanding softlock's.
7198 		 */
7199 		segvn_purge(seg);
7200 		if (svd->amp->a_softlockcnt > 0 || svd->softlockcnt > 0) {
7201 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7202 			return (EAGAIN);
7203 		}
7204 	}
7205 
7206 	vpp = svd->vpage;
7207 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7208 	bflags = ((flags & MS_ASYNC) ? B_ASYNC : 0) |
7209 	    ((flags & MS_INVALIDATE) ? B_INVAL : 0);
7210 
7211 	if (attr) {
7212 		pageprot = attr & ~(SHARED|PRIVATE);
7213 		segtype = (attr & SHARED) ? MAP_SHARED : MAP_PRIVATE;
7214 
7215 		/*
7216 		 * We are done if the segment types don't match
7217 		 * or if we have segment level protections and
7218 		 * they don't match.
7219 		 */
7220 		if (svd->type != segtype) {
7221 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7222 			return (0);
7223 		}
7224 		if (vpp == NULL) {
7225 			if (svd->prot != pageprot) {
7226 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7227 				return (0);
7228 			}
7229 			prot = svd->prot;
7230 		} else
7231 			vpp = &svd->vpage[seg_page(seg, addr)];
7232 
7233 	} else if (svd->vp && svd->amp == NULL &&
7234 	    (flags & MS_INVALIDATE) == 0) {
7235 
7236 		/*
7237 		 * No attributes, no anonymous pages and MS_INVALIDATE flag
7238 		 * is not on, just use one big request.
7239 		 */
7240 		err = VOP_PUTPAGE(svd->vp, (offset_t)offset, len,
7241 		    bflags, svd->cred, NULL);
7242 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7243 		return (err);
7244 	}
7245 
7246 	if ((amp = svd->amp) != NULL)
7247 		anon_index = svd->anon_index + seg_page(seg, addr);
7248 
7249 	for (eaddr = addr + len; addr < eaddr; addr += PAGESIZE) {
7250 		ap = NULL;
7251 		if (amp != NULL) {
7252 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7253 			anon_array_enter(amp, anon_index, &cookie);
7254 			ap = anon_get_ptr(amp->ahp, anon_index++);
7255 			if (ap != NULL) {
7256 				swap_xlate(ap, &vp, &off);
7257 			} else {
7258 				vp = svd->vp;
7259 				off = offset;
7260 			}
7261 			anon_array_exit(&cookie);
7262 			ANON_LOCK_EXIT(&amp->a_rwlock);
7263 		} else {
7264 			vp = svd->vp;
7265 			off = offset;
7266 		}
7267 		offset += PAGESIZE;
7268 
7269 		if (vp == NULL)		/* untouched zfod page */
7270 			continue;
7271 
7272 		if (attr) {
7273 			if (vpp) {
7274 				prot = VPP_PROT(vpp);
7275 				vpp++;
7276 			}
7277 			if (prot != pageprot) {
7278 				continue;
7279 			}
7280 		}
7281 
7282 		/*
7283 		 * See if any of these pages are locked --  if so, then we
7284 		 * will have to truncate an invalidate request at the first
7285 		 * locked one. We don't need the page_struct_lock to test
7286 		 * as this is only advisory; even if we acquire it someone
7287 		 * might race in and lock the page after we unlock and before
7288 		 * we do the PUTPAGE, then PUTPAGE simply does nothing.
7289 		 */
7290 		if (flags & MS_INVALIDATE) {
7291 			if ((pp = page_lookup(vp, off, SE_SHARED)) != NULL) {
7292 				if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
7293 					page_unlock(pp);
7294 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7295 					return (EBUSY);
7296 				}
7297 				if (ap != NULL && pp->p_szc != 0 &&
7298 				    page_tryupgrade(pp)) {
7299 					if (pp->p_lckcnt == 0 &&
7300 					    pp->p_cowcnt == 0) {
7301 						/*
7302 						 * swapfs VN_DISPOSE() won't
7303 						 * invalidate large pages.
7304 						 * Attempt to demote.
7305 						 * XXX can't help it if it
7306 						 * fails. But for swapfs
7307 						 * pages it is no big deal.
7308 						 */
7309 						(void) page_try_demote_pages(
7310 						    pp);
7311 					}
7312 				}
7313 				page_unlock(pp);
7314 			}
7315 		} else if (svd->type == MAP_SHARED && amp != NULL) {
7316 			/*
7317 			 * Avoid writing out to disk ISM's large pages
7318 			 * because segspt_free_pages() relies on NULL an_pvp
7319 			 * of anon slots of such pages.
7320 			 */
7321 
7322 			ASSERT(svd->vp == NULL);
7323 			/*
7324 			 * swapfs uses page_lookup_nowait if not freeing or
7325 			 * invalidating and skips a page if
7326 			 * page_lookup_nowait returns NULL.
7327 			 */
7328 			pp = page_lookup_nowait(vp, off, SE_SHARED);
7329 			if (pp == NULL) {
7330 				continue;
7331 			}
7332 			if (pp->p_szc != 0) {
7333 				page_unlock(pp);
7334 				continue;
7335 			}
7336 
7337 			/*
7338 			 * Note ISM pages are created large so (vp, off)'s
7339 			 * page cannot suddenly become large after we unlock
7340 			 * pp.
7341 			 */
7342 			page_unlock(pp);
7343 		}
7344 		/*
7345 		 * XXX - Should ultimately try to kluster
7346 		 * calls to VOP_PUTPAGE() for performance.
7347 		 */
7348 		VN_HOLD(vp);
7349 		err = VOP_PUTPAGE(vp, (offset_t)off, PAGESIZE,
7350 		    bflags, svd->cred, NULL);
7351 		VN_RELE(vp);
7352 		if (err)
7353 			break;
7354 	}
7355 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7356 	return (err);
7357 }
7358 
7359 /*
7360  * Determine if we have data corresponding to pages in the
7361  * primary storage virtual memory cache (i.e., "in core").
7362  */
7363 static size_t
7364 segvn_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
7365 {
7366 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7367 	struct vnode *vp, *avp;
7368 	u_offset_t offset, aoffset;
7369 	size_t p, ep;
7370 	int ret;
7371 	struct vpage *vpp;
7372 	page_t *pp;
7373 	uint_t start;
7374 	struct anon_map *amp;		/* XXX - for locknest */
7375 	struct anon *ap;
7376 	uint_t attr;
7377 	anon_sync_obj_t cookie;
7378 
7379 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7380 
7381 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7382 	if (svd->amp == NULL && svd->vp == NULL) {
7383 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7384 		bzero(vec, btopr(len));
7385 		return (len);	/* no anonymous pages created yet */
7386 	}
7387 
7388 	p = seg_page(seg, addr);
7389 	ep = seg_page(seg, addr + len);
7390 	start = svd->vp ? SEG_PAGE_VNODEBACKED : 0;
7391 
7392 	amp = svd->amp;
7393 	for (; p < ep; p++, addr += PAGESIZE) {
7394 		vpp = (svd->vpage) ? &svd->vpage[p]: NULL;
7395 		ret = start;
7396 		ap = NULL;
7397 		avp = NULL;
7398 		/* Grab the vnode/offset for the anon slot */
7399 		if (amp != NULL) {
7400 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7401 			anon_array_enter(amp, svd->anon_index + p, &cookie);
7402 			ap = anon_get_ptr(amp->ahp, svd->anon_index + p);
7403 			if (ap != NULL) {
7404 				swap_xlate(ap, &avp, &aoffset);
7405 			}
7406 			anon_array_exit(&cookie);
7407 			ANON_LOCK_EXIT(&amp->a_rwlock);
7408 		}
7409 		if ((avp != NULL) && page_exists(avp, aoffset)) {
7410 			/* A page exists for the anon slot */
7411 			ret |= SEG_PAGE_INCORE;
7412 
7413 			/*
7414 			 * If page is mapped and writable
7415 			 */
7416 			attr = (uint_t)0;
7417 			if ((hat_getattr(seg->s_as->a_hat, addr,
7418 			    &attr) != -1) && (attr & PROT_WRITE)) {
7419 				ret |= SEG_PAGE_ANON;
7420 			}
7421 			/*
7422 			 * Don't get page_struct lock for lckcnt and cowcnt,
7423 			 * since this is purely advisory.
7424 			 */
7425 			if ((pp = page_lookup_nowait(avp, aoffset,
7426 			    SE_SHARED)) != NULL) {
7427 				if (pp->p_lckcnt)
7428 					ret |= SEG_PAGE_SOFTLOCK;
7429 				if (pp->p_cowcnt)
7430 					ret |= SEG_PAGE_HASCOW;
7431 				page_unlock(pp);
7432 			}
7433 		}
7434 
7435 		/* Gather vnode statistics */
7436 		vp = svd->vp;
7437 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7438 
7439 		if (vp != NULL) {
7440 			/*
7441 			 * Try to obtain a "shared" lock on the page
7442 			 * without blocking.  If this fails, determine
7443 			 * if the page is in memory.
7444 			 */
7445 			pp = page_lookup_nowait(vp, offset, SE_SHARED);
7446 			if ((pp == NULL) && (page_exists(vp, offset))) {
7447 				/* Page is incore, and is named */
7448 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7449 			}
7450 			/*
7451 			 * Don't get page_struct lock for lckcnt and cowcnt,
7452 			 * since this is purely advisory.
7453 			 */
7454 			if (pp != NULL) {
7455 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7456 				if (pp->p_lckcnt)
7457 					ret |= SEG_PAGE_SOFTLOCK;
7458 				if (pp->p_cowcnt)
7459 					ret |= SEG_PAGE_HASCOW;
7460 				page_unlock(pp);
7461 			}
7462 		}
7463 
7464 		/* Gather virtual page information */
7465 		if (vpp) {
7466 			if (VPP_ISPPLOCK(vpp))
7467 				ret |= SEG_PAGE_LOCKED;
7468 			vpp++;
7469 		}
7470 
7471 		*vec++ = (char)ret;
7472 	}
7473 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7474 	return (len);
7475 }
7476 
7477 /*
7478  * Statement for p_cowcnts/p_lckcnts.
7479  *
7480  * p_cowcnt is updated while mlock/munlocking MAP_PRIVATE and PROT_WRITE region
7481  * irrespective of the following factors or anything else:
7482  *
7483  *	(1) anon slots are populated or not
7484  *	(2) cow is broken or not
7485  *	(3) refcnt on ap is 1 or greater than 1
7486  *
7487  * If it's not MAP_PRIVATE and PROT_WRITE, p_lckcnt is updated during mlock
7488  * and munlock.
7489  *
7490  *
7491  * Handling p_cowcnts/p_lckcnts during copy-on-write fault:
7492  *
7493  *	if vpage has PROT_WRITE
7494  *		transfer cowcnt on the oldpage -> cowcnt on the newpage
7495  *	else
7496  *		transfer lckcnt on the oldpage -> lckcnt on the newpage
7497  *
7498  *	During copy-on-write, decrement p_cowcnt on the oldpage and increment
7499  *	p_cowcnt on the newpage *if* the corresponding vpage has PROT_WRITE.
7500  *
7501  *	We may also break COW if softlocking on read access in the physio case.
7502  *	In this case, vpage may not have PROT_WRITE. So, we need to decrement
7503  *	p_lckcnt on the oldpage and increment p_lckcnt on the newpage *if* the
7504  *	vpage doesn't have PROT_WRITE.
7505  *
7506  *
7507  * Handling p_cowcnts/p_lckcnts during mprotect on mlocked region:
7508  *
7509  * 	If a MAP_PRIVATE region loses PROT_WRITE, we decrement p_cowcnt and
7510  *	increment p_lckcnt by calling page_subclaim() which takes care of
7511  * 	availrmem accounting and p_lckcnt overflow.
7512  *
7513  *	If a MAP_PRIVATE region gains PROT_WRITE, we decrement p_lckcnt and
7514  *	increment p_cowcnt by calling page_addclaim() which takes care of
7515  *	availrmem availability and p_cowcnt overflow.
7516  */
7517 
7518 /*
7519  * Lock down (or unlock) pages mapped by this segment.
7520  *
7521  * XXX only creates PAGESIZE pages if anon slots are not initialized.
7522  * At fault time they will be relocated into larger pages.
7523  */
7524 static int
7525 segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
7526     int attr, int op, ulong_t *lockmap, size_t pos)
7527 {
7528 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7529 	struct vpage *vpp;
7530 	struct vpage *evp;
7531 	page_t *pp;
7532 	u_offset_t offset;
7533 	u_offset_t off;
7534 	int segtype;
7535 	int pageprot;
7536 	int claim;
7537 	struct vnode *vp;
7538 	ulong_t anon_index;
7539 	struct anon_map *amp;
7540 	struct anon *ap;
7541 	struct vattr va;
7542 	anon_sync_obj_t cookie;
7543 	struct kshmid *sp = NULL;
7544 	struct proc	*p = curproc;
7545 	kproject_t	*proj = NULL;
7546 	int chargeproc = 1;
7547 	size_t locked_bytes = 0;
7548 	size_t unlocked_bytes = 0;
7549 	int err = 0;
7550 
7551 	/*
7552 	 * Hold write lock on address space because may split or concatenate
7553 	 * segments
7554 	 */
7555 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7556 
7557 	/*
7558 	 * If this is a shm, use shm's project and zone, else use
7559 	 * project and zone of calling process
7560 	 */
7561 
7562 	/* Determine if this segment backs a sysV shm */
7563 	if (svd->amp != NULL && svd->amp->a_sp != NULL) {
7564 		ASSERT(svd->type == MAP_SHARED);
7565 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
7566 		sp = svd->amp->a_sp;
7567 		proj = sp->shm_perm.ipc_proj;
7568 		chargeproc = 0;
7569 	}
7570 
7571 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7572 	if (attr) {
7573 		pageprot = attr & ~(SHARED|PRIVATE);
7574 		segtype = attr & SHARED ? MAP_SHARED : MAP_PRIVATE;
7575 
7576 		/*
7577 		 * We are done if the segment types don't match
7578 		 * or if we have segment level protections and
7579 		 * they don't match.
7580 		 */
7581 		if (svd->type != segtype) {
7582 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7583 			return (0);
7584 		}
7585 		if (svd->pageprot == 0 && svd->prot != pageprot) {
7586 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7587 			return (0);
7588 		}
7589 	}
7590 
7591 	if (op == MC_LOCK) {
7592 		if (svd->tr_state == SEGVN_TR_INIT) {
7593 			svd->tr_state = SEGVN_TR_OFF;
7594 		} else if (svd->tr_state == SEGVN_TR_ON) {
7595 			ASSERT(svd->amp != NULL);
7596 			segvn_textunrepl(seg, 0);
7597 			ASSERT(svd->amp == NULL &&
7598 			    svd->tr_state == SEGVN_TR_OFF);
7599 		}
7600 	}
7601 
7602 	/*
7603 	 * If we're locking, then we must create a vpage structure if
7604 	 * none exists.  If we're unlocking, then check to see if there
7605 	 * is a vpage --  if not, then we could not have locked anything.
7606 	 */
7607 
7608 	if ((vpp = svd->vpage) == NULL) {
7609 		if (op == MC_LOCK)
7610 			segvn_vpage(seg);
7611 		else {
7612 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7613 			return (0);
7614 		}
7615 	}
7616 
7617 	/*
7618 	 * The anonymous data vector (i.e., previously
7619 	 * unreferenced mapping to swap space) can be allocated
7620 	 * by lazily testing for its existence.
7621 	 */
7622 	if (op == MC_LOCK && svd->amp == NULL && svd->vp == NULL) {
7623 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
7624 		svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
7625 		svd->amp->a_szc = seg->s_szc;
7626 	}
7627 
7628 	if ((amp = svd->amp) != NULL) {
7629 		anon_index = svd->anon_index + seg_page(seg, addr);
7630 	}
7631 
7632 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7633 	evp = &svd->vpage[seg_page(seg, addr + len)];
7634 
7635 	if (sp != NULL)
7636 		mutex_enter(&sp->shm_mlock);
7637 
7638 	/* determine number of unlocked bytes in range for lock operation */
7639 	if (op == MC_LOCK) {
7640 
7641 		if (sp == NULL) {
7642 			for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7643 			    vpp++) {
7644 				if (!VPP_ISPPLOCK(vpp))
7645 					unlocked_bytes += PAGESIZE;
7646 			}
7647 		} else {
7648 			ulong_t		i_idx, i_edx;
7649 			anon_sync_obj_t	i_cookie;
7650 			struct anon	*i_ap;
7651 			struct vnode	*i_vp;
7652 			u_offset_t	i_off;
7653 
7654 			/* Only count sysV pages once for locked memory */
7655 			i_edx = svd->anon_index + seg_page(seg, addr + len);
7656 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7657 			for (i_idx = anon_index; i_idx < i_edx; i_idx++) {
7658 				anon_array_enter(amp, i_idx, &i_cookie);
7659 				i_ap = anon_get_ptr(amp->ahp, i_idx);
7660 				if (i_ap == NULL) {
7661 					unlocked_bytes += PAGESIZE;
7662 					anon_array_exit(&i_cookie);
7663 					continue;
7664 				}
7665 				swap_xlate(i_ap, &i_vp, &i_off);
7666 				anon_array_exit(&i_cookie);
7667 				pp = page_lookup(i_vp, i_off, SE_SHARED);
7668 				if (pp == NULL) {
7669 					unlocked_bytes += PAGESIZE;
7670 					continue;
7671 				} else if (pp->p_lckcnt == 0)
7672 					unlocked_bytes += PAGESIZE;
7673 				page_unlock(pp);
7674 			}
7675 			ANON_LOCK_EXIT(&amp->a_rwlock);
7676 		}
7677 
7678 		mutex_enter(&p->p_lock);
7679 		err = rctl_incr_locked_mem(p, proj, unlocked_bytes,
7680 		    chargeproc);
7681 		mutex_exit(&p->p_lock);
7682 
7683 		if (err) {
7684 			if (sp != NULL)
7685 				mutex_exit(&sp->shm_mlock);
7686 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7687 			return (err);
7688 		}
7689 	}
7690 	/*
7691 	 * Loop over all pages in the range.  Process if we're locking and
7692 	 * page has not already been locked in this mapping; or if we're
7693 	 * unlocking and the page has been locked.
7694 	 */
7695 	for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7696 	    vpp++, pos++, addr += PAGESIZE, offset += PAGESIZE, anon_index++) {
7697 		if ((attr == 0 || VPP_PROT(vpp) == pageprot) &&
7698 		    ((op == MC_LOCK && !VPP_ISPPLOCK(vpp)) ||
7699 		    (op == MC_UNLOCK && VPP_ISPPLOCK(vpp)))) {
7700 
7701 			if (amp != NULL)
7702 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7703 			/*
7704 			 * If this isn't a MAP_NORESERVE segment and
7705 			 * we're locking, allocate anon slots if they
7706 			 * don't exist.  The page is brought in later on.
7707 			 */
7708 			if (op == MC_LOCK && svd->vp == NULL &&
7709 			    ((svd->flags & MAP_NORESERVE) == 0) &&
7710 			    amp != NULL &&
7711 			    ((ap = anon_get_ptr(amp->ahp, anon_index))
7712 			    == NULL)) {
7713 				anon_array_enter(amp, anon_index, &cookie);
7714 
7715 				if ((ap = anon_get_ptr(amp->ahp,
7716 				    anon_index)) == NULL) {
7717 					pp = anon_zero(seg, addr, &ap,
7718 					    svd->cred);
7719 					if (pp == NULL) {
7720 						anon_array_exit(&cookie);
7721 						ANON_LOCK_EXIT(&amp->a_rwlock);
7722 						err = ENOMEM;
7723 						goto out;
7724 					}
7725 					ASSERT(anon_get_ptr(amp->ahp,
7726 					    anon_index) == NULL);
7727 					(void) anon_set_ptr(amp->ahp,
7728 					    anon_index, ap, ANON_SLEEP);
7729 					page_unlock(pp);
7730 				}
7731 				anon_array_exit(&cookie);
7732 			}
7733 
7734 			/*
7735 			 * Get name for page, accounting for
7736 			 * existence of private copy.
7737 			 */
7738 			ap = NULL;
7739 			if (amp != NULL) {
7740 				anon_array_enter(amp, anon_index, &cookie);
7741 				ap = anon_get_ptr(amp->ahp, anon_index);
7742 				if (ap != NULL) {
7743 					swap_xlate(ap, &vp, &off);
7744 				} else {
7745 					if (svd->vp == NULL &&
7746 					    (svd->flags & MAP_NORESERVE)) {
7747 						anon_array_exit(&cookie);
7748 						ANON_LOCK_EXIT(&amp->a_rwlock);
7749 						continue;
7750 					}
7751 					vp = svd->vp;
7752 					off = offset;
7753 				}
7754 				if (op != MC_LOCK || ap == NULL) {
7755 					anon_array_exit(&cookie);
7756 					ANON_LOCK_EXIT(&amp->a_rwlock);
7757 				}
7758 			} else {
7759 				vp = svd->vp;
7760 				off = offset;
7761 			}
7762 
7763 			/*
7764 			 * Get page frame.  It's ok if the page is
7765 			 * not available when we're unlocking, as this
7766 			 * may simply mean that a page we locked got
7767 			 * truncated out of existence after we locked it.
7768 			 *
7769 			 * Invoke VOP_GETPAGE() to obtain the page struct
7770 			 * since we may need to read it from disk if its
7771 			 * been paged out.
7772 			 */
7773 			if (op != MC_LOCK)
7774 				pp = page_lookup(vp, off, SE_SHARED);
7775 			else {
7776 				page_t *pl[1 + 1];
7777 				int error;
7778 
7779 				ASSERT(vp != NULL);
7780 
7781 				error = VOP_GETPAGE(vp, (offset_t)off, PAGESIZE,
7782 				    (uint_t *)NULL, pl, PAGESIZE, seg, addr,
7783 				    S_OTHER, svd->cred, NULL);
7784 
7785 				if (error && ap != NULL) {
7786 					anon_array_exit(&cookie);
7787 					ANON_LOCK_EXIT(&amp->a_rwlock);
7788 				}
7789 
7790 				/*
7791 				 * If the error is EDEADLK then we must bounce
7792 				 * up and drop all vm subsystem locks and then
7793 				 * retry the operation later
7794 				 * This behavior is a temporary measure because
7795 				 * ufs/sds logging is badly designed and will
7796 				 * deadlock if we don't allow this bounce to
7797 				 * happen.  The real solution is to re-design
7798 				 * the logging code to work properly.  See bug
7799 				 * 4125102 for details of the problem.
7800 				 */
7801 				if (error == EDEADLK) {
7802 					err = error;
7803 					goto out;
7804 				}
7805 				/*
7806 				 * Quit if we fail to fault in the page.  Treat
7807 				 * the failure as an error, unless the addr
7808 				 * is mapped beyond the end of a file.
7809 				 */
7810 				if (error && svd->vp) {
7811 					va.va_mask = AT_SIZE;
7812 					if (VOP_GETATTR(svd->vp, &va, 0,
7813 					    svd->cred, NULL) != 0) {
7814 						err = EIO;
7815 						goto out;
7816 					}
7817 					if (btopr(va.va_size) >=
7818 					    btopr(off + 1)) {
7819 						err = EIO;
7820 						goto out;
7821 					}
7822 					goto out;
7823 
7824 				} else if (error) {
7825 					err = EIO;
7826 					goto out;
7827 				}
7828 				pp = pl[0];
7829 				ASSERT(pp != NULL);
7830 			}
7831 
7832 			/*
7833 			 * See Statement at the beginning of this routine.
7834 			 *
7835 			 * claim is always set if MAP_PRIVATE and PROT_WRITE
7836 			 * irrespective of following factors:
7837 			 *
7838 			 * (1) anon slots are populated or not
7839 			 * (2) cow is broken or not
7840 			 * (3) refcnt on ap is 1 or greater than 1
7841 			 *
7842 			 * See 4140683 for details
7843 			 */
7844 			claim = ((VPP_PROT(vpp) & PROT_WRITE) &&
7845 			    (svd->type == MAP_PRIVATE));
7846 
7847 			/*
7848 			 * Perform page-level operation appropriate to
7849 			 * operation.  If locking, undo the SOFTLOCK
7850 			 * performed to bring the page into memory
7851 			 * after setting the lock.  If unlocking,
7852 			 * and no page was found, account for the claim
7853 			 * separately.
7854 			 */
7855 			if (op == MC_LOCK) {
7856 				int ret = 1;	/* Assume success */
7857 
7858 				ASSERT(!VPP_ISPPLOCK(vpp));
7859 
7860 				ret = page_pp_lock(pp, claim, 0);
7861 				if (ap != NULL) {
7862 					if (ap->an_pvp != NULL) {
7863 						anon_swap_free(ap, pp);
7864 					}
7865 					anon_array_exit(&cookie);
7866 					ANON_LOCK_EXIT(&amp->a_rwlock);
7867 				}
7868 				if (ret == 0) {
7869 					/* locking page failed */
7870 					page_unlock(pp);
7871 					err = EAGAIN;
7872 					goto out;
7873 				}
7874 				VPP_SETPPLOCK(vpp);
7875 				if (sp != NULL) {
7876 					if (pp->p_lckcnt == 1)
7877 						locked_bytes += PAGESIZE;
7878 				} else
7879 					locked_bytes += PAGESIZE;
7880 
7881 				if (lockmap != (ulong_t *)NULL)
7882 					BT_SET(lockmap, pos);
7883 
7884 				page_unlock(pp);
7885 			} else {
7886 				ASSERT(VPP_ISPPLOCK(vpp));
7887 				if (pp != NULL) {
7888 					/* sysV pages should be locked */
7889 					ASSERT(sp == NULL || pp->p_lckcnt > 0);
7890 					page_pp_unlock(pp, claim, 0);
7891 					if (sp != NULL) {
7892 						if (pp->p_lckcnt == 0)
7893 							unlocked_bytes
7894 							    += PAGESIZE;
7895 					} else
7896 						unlocked_bytes += PAGESIZE;
7897 					page_unlock(pp);
7898 				} else {
7899 					ASSERT(sp == NULL);
7900 					unlocked_bytes += PAGESIZE;
7901 				}
7902 				VPP_CLRPPLOCK(vpp);
7903 			}
7904 		}
7905 	}
7906 out:
7907 	if (op == MC_LOCK) {
7908 		/* Credit back bytes that did not get locked */
7909 		if ((unlocked_bytes - locked_bytes) > 0) {
7910 			if (proj == NULL)
7911 				mutex_enter(&p->p_lock);
7912 			rctl_decr_locked_mem(p, proj,
7913 			    (unlocked_bytes - locked_bytes), chargeproc);
7914 			if (proj == NULL)
7915 				mutex_exit(&p->p_lock);
7916 		}
7917 
7918 	} else {
7919 		/* Account bytes that were unlocked */
7920 		if (unlocked_bytes > 0) {
7921 			if (proj == NULL)
7922 				mutex_enter(&p->p_lock);
7923 			rctl_decr_locked_mem(p, proj, unlocked_bytes,
7924 			    chargeproc);
7925 			if (proj == NULL)
7926 				mutex_exit(&p->p_lock);
7927 		}
7928 	}
7929 	if (sp != NULL)
7930 		mutex_exit(&sp->shm_mlock);
7931 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7932 
7933 	return (err);
7934 }
7935 
7936 /*
7937  * Set advice from user for specified pages
7938  * There are 5 types of advice:
7939  *	MADV_NORMAL	- Normal (default) behavior (whatever that is)
7940  *	MADV_RANDOM	- Random page references
7941  *				do not allow readahead or 'klustering'
7942  *	MADV_SEQUENTIAL	- Sequential page references
7943  *				Pages previous to the one currently being
7944  *				accessed (determined by fault) are 'not needed'
7945  *				and are freed immediately
7946  *	MADV_WILLNEED	- Pages are likely to be used (fault ahead in mctl)
7947  *	MADV_DONTNEED	- Pages are not needed (synced out in mctl)
7948  *	MADV_FREE	- Contents can be discarded
7949  *	MADV_ACCESS_DEFAULT- Default access
7950  *	MADV_ACCESS_LWP	- Next LWP will access heavily
7951  *	MADV_ACCESS_MANY- Many LWPs or processes will access heavily
7952  */
7953 static int
7954 segvn_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
7955 {
7956 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7957 	size_t page;
7958 	int err = 0;
7959 	int already_set;
7960 	struct anon_map *amp;
7961 	ulong_t anon_index;
7962 	struct seg *next;
7963 	lgrp_mem_policy_t policy;
7964 	struct seg *prev;
7965 	struct vnode *vp;
7966 
7967 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7968 
7969 	/*
7970 	 * In case of MADV_FREE, we won't be modifying any segment private
7971 	 * data structures; so, we only need to grab READER's lock
7972 	 */
7973 	if (behav != MADV_FREE) {
7974 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7975 		if (svd->tr_state != SEGVN_TR_OFF) {
7976 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7977 			return (0);
7978 		}
7979 	} else {
7980 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7981 	}
7982 
7983 	/*
7984 	 * Large pages are assumed to be only turned on when accesses to the
7985 	 * segment's address range have spatial and temporal locality. That
7986 	 * justifies ignoring MADV_SEQUENTIAL for large page segments.
7987 	 * Also, ignore advice affecting lgroup memory allocation
7988 	 * if don't need to do lgroup optimizations on this system
7989 	 */
7990 
7991 	if ((behav == MADV_SEQUENTIAL &&
7992 	    (seg->s_szc != 0 || HAT_IS_REGION_COOKIE_VALID(svd->rcookie))) ||
7993 	    (!lgrp_optimizations() && (behav == MADV_ACCESS_DEFAULT ||
7994 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY))) {
7995 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7996 		return (0);
7997 	}
7998 
7999 	if (behav == MADV_SEQUENTIAL || behav == MADV_ACCESS_DEFAULT ||
8000 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY) {
8001 		/*
8002 		 * Since we are going to unload hat mappings
8003 		 * we first have to flush the cache. Otherwise
8004 		 * this might lead to system panic if another
8005 		 * thread is doing physio on the range whose
8006 		 * mappings are unloaded by madvise(3C).
8007 		 */
8008 		if (svd->softlockcnt > 0) {
8009 			/*
8010 			 * If this is shared segment non 0 softlockcnt
8011 			 * means locked pages are still in use.
8012 			 */
8013 			if (svd->type == MAP_SHARED) {
8014 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8015 				return (EAGAIN);
8016 			}
8017 			/*
8018 			 * Since we do have the segvn writers lock
8019 			 * nobody can fill the cache with entries
8020 			 * belonging to this seg during the purge.
8021 			 * The flush either succeeds or we still
8022 			 * have pending I/Os. In the later case,
8023 			 * madvise(3C) fails.
8024 			 */
8025 			segvn_purge(seg);
8026 			if (svd->softlockcnt > 0) {
8027 				/*
8028 				 * Since madvise(3C) is advisory and
8029 				 * it's not part of UNIX98, madvise(3C)
8030 				 * failure here doesn't cause any hardship.
8031 				 * Note that we don't block in "as" layer.
8032 				 */
8033 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8034 				return (EAGAIN);
8035 			}
8036 		} else if (svd->type == MAP_SHARED && svd->amp != NULL &&
8037 		    svd->amp->a_softlockcnt > 0) {
8038 			/*
8039 			 * Try to purge this amp's entries from pcache. It
8040 			 * will succeed only if other segments that share the
8041 			 * amp have no outstanding softlock's.
8042 			 */
8043 			segvn_purge(seg);
8044 		}
8045 	}
8046 
8047 	amp = svd->amp;
8048 	vp = svd->vp;
8049 	if (behav == MADV_FREE) {
8050 		/*
8051 		 * MADV_FREE is not supported for segments with
8052 		 * underlying object; if anonmap is NULL, anon slots
8053 		 * are not yet populated and there is nothing for
8054 		 * us to do. As MADV_FREE is advisory, we don't
8055 		 * return error in either case.
8056 		 */
8057 		if (vp != NULL || amp == NULL) {
8058 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8059 			return (0);
8060 		}
8061 
8062 		segvn_purge(seg);
8063 
8064 		page = seg_page(seg, addr);
8065 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8066 		anon_disclaim(amp, svd->anon_index + page, len);
8067 		ANON_LOCK_EXIT(&amp->a_rwlock);
8068 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8069 		return (0);
8070 	}
8071 
8072 	/*
8073 	 * If advice is to be applied to entire segment,
8074 	 * use advice field in seg_data structure
8075 	 * otherwise use appropriate vpage entry.
8076 	 */
8077 	if ((addr == seg->s_base) && (len == seg->s_size)) {
8078 		switch (behav) {
8079 		case MADV_ACCESS_LWP:
8080 		case MADV_ACCESS_MANY:
8081 		case MADV_ACCESS_DEFAULT:
8082 			/*
8083 			 * Set memory allocation policy for this segment
8084 			 */
8085 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8086 			if (svd->type == MAP_SHARED)
8087 				already_set = lgrp_shm_policy_set(policy, amp,
8088 				    svd->anon_index, vp, svd->offset, len);
8089 			else {
8090 				/*
8091 				 * For private memory, need writers lock on
8092 				 * address space because the segment may be
8093 				 * split or concatenated when changing policy
8094 				 */
8095 				if (AS_READ_HELD(seg->s_as,
8096 				    &seg->s_as->a_lock)) {
8097 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8098 					return (IE_RETRY);
8099 				}
8100 
8101 				already_set = lgrp_privm_policy_set(policy,
8102 				    &svd->policy_info, len);
8103 			}
8104 
8105 			/*
8106 			 * If policy set already and it shouldn't be reapplied,
8107 			 * don't do anything.
8108 			 */
8109 			if (already_set &&
8110 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8111 				break;
8112 
8113 			/*
8114 			 * Mark any existing pages in given range for
8115 			 * migration
8116 			 */
8117 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8118 			    vp, svd->offset, 1);
8119 
8120 			/*
8121 			 * If same policy set already or this is a shared
8122 			 * memory segment, don't need to try to concatenate
8123 			 * segment with adjacent ones.
8124 			 */
8125 			if (already_set || svd->type == MAP_SHARED)
8126 				break;
8127 
8128 			/*
8129 			 * Try to concatenate this segment with previous
8130 			 * one and next one, since we changed policy for
8131 			 * this one and it may be compatible with adjacent
8132 			 * ones now.
8133 			 */
8134 			prev = AS_SEGPREV(seg->s_as, seg);
8135 			next = AS_SEGNEXT(seg->s_as, seg);
8136 
8137 			if (next && next->s_ops == &segvn_ops &&
8138 			    addr + len == next->s_base)
8139 				(void) segvn_concat(seg, next, 1);
8140 
8141 			if (prev && prev->s_ops == &segvn_ops &&
8142 			    addr == prev->s_base + prev->s_size) {
8143 				/*
8144 				 * Drop lock for private data of current
8145 				 * segment before concatenating (deleting) it
8146 				 * and return IE_REATTACH to tell as_ctl() that
8147 				 * current segment has changed
8148 				 */
8149 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8150 				if (!segvn_concat(prev, seg, 1))
8151 					err = IE_REATTACH;
8152 
8153 				return (err);
8154 			}
8155 			break;
8156 
8157 		case MADV_SEQUENTIAL:
8158 			/*
8159 			 * unloading mapping guarantees
8160 			 * detection in segvn_fault
8161 			 */
8162 			ASSERT(seg->s_szc == 0);
8163 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8164 			hat_unload(seg->s_as->a_hat, addr, len,
8165 			    HAT_UNLOAD);
8166 			/* FALLTHROUGH */
8167 		case MADV_NORMAL:
8168 		case MADV_RANDOM:
8169 			svd->advice = (uchar_t)behav;
8170 			svd->pageadvice = 0;
8171 			break;
8172 		case MADV_WILLNEED:	/* handled in memcntl */
8173 		case MADV_DONTNEED:	/* handled in memcntl */
8174 		case MADV_FREE:		/* handled above */
8175 			break;
8176 		default:
8177 			err = EINVAL;
8178 		}
8179 	} else {
8180 		caddr_t			eaddr;
8181 		struct seg		*new_seg;
8182 		struct segvn_data	*new_svd;
8183 		u_offset_t		off;
8184 		caddr_t			oldeaddr;
8185 
8186 		page = seg_page(seg, addr);
8187 
8188 		segvn_vpage(seg);
8189 
8190 		switch (behav) {
8191 			struct vpage *bvpp, *evpp;
8192 
8193 		case MADV_ACCESS_LWP:
8194 		case MADV_ACCESS_MANY:
8195 		case MADV_ACCESS_DEFAULT:
8196 			/*
8197 			 * Set memory allocation policy for portion of this
8198 			 * segment
8199 			 */
8200 
8201 			/*
8202 			 * Align address and length of advice to page
8203 			 * boundaries for large pages
8204 			 */
8205 			if (seg->s_szc != 0) {
8206 				size_t	pgsz;
8207 
8208 				pgsz = page_get_pagesize(seg->s_szc);
8209 				addr = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
8210 				len = P2ROUNDUP(len, pgsz);
8211 			}
8212 
8213 			/*
8214 			 * Check to see whether policy is set already
8215 			 */
8216 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8217 
8218 			anon_index = svd->anon_index + page;
8219 			off = svd->offset + (uintptr_t)(addr - seg->s_base);
8220 
8221 			if (svd->type == MAP_SHARED)
8222 				already_set = lgrp_shm_policy_set(policy, amp,
8223 				    anon_index, vp, off, len);
8224 			else
8225 				already_set =
8226 				    (policy == svd->policy_info.mem_policy);
8227 
8228 			/*
8229 			 * If policy set already and it shouldn't be reapplied,
8230 			 * don't do anything.
8231 			 */
8232 			if (already_set &&
8233 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8234 				break;
8235 
8236 			/*
8237 			 * For private memory, need writers lock on
8238 			 * address space because the segment may be
8239 			 * split or concatenated when changing policy
8240 			 */
8241 			if (svd->type == MAP_PRIVATE &&
8242 			    AS_READ_HELD(seg->s_as, &seg->s_as->a_lock)) {
8243 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8244 				return (IE_RETRY);
8245 			}
8246 
8247 			/*
8248 			 * Mark any existing pages in given range for
8249 			 * migration
8250 			 */
8251 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8252 			    vp, svd->offset, 1);
8253 
8254 			/*
8255 			 * Don't need to try to split or concatenate
8256 			 * segments, since policy is same or this is a shared
8257 			 * memory segment
8258 			 */
8259 			if (already_set || svd->type == MAP_SHARED)
8260 				break;
8261 
8262 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
8263 				ASSERT(svd->amp == NULL);
8264 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
8265 				ASSERT(svd->softlockcnt == 0);
8266 				hat_leave_region(seg->s_as->a_hat, svd->rcookie,
8267 				    HAT_REGION_TEXT);
8268 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
8269 			}
8270 
8271 			/*
8272 			 * Split off new segment if advice only applies to a
8273 			 * portion of existing segment starting in middle
8274 			 */
8275 			new_seg = NULL;
8276 			eaddr = addr + len;
8277 			oldeaddr = seg->s_base + seg->s_size;
8278 			if (addr > seg->s_base) {
8279 				/*
8280 				 * Must flush I/O page cache
8281 				 * before splitting segment
8282 				 */
8283 				if (svd->softlockcnt > 0)
8284 					segvn_purge(seg);
8285 
8286 				/*
8287 				 * Split segment and return IE_REATTACH to tell
8288 				 * as_ctl() that current segment changed
8289 				 */
8290 				new_seg = segvn_split_seg(seg, addr);
8291 				new_svd = (struct segvn_data *)new_seg->s_data;
8292 				err = IE_REATTACH;
8293 
8294 				/*
8295 				 * If new segment ends where old one
8296 				 * did, try to concatenate the new
8297 				 * segment with next one.
8298 				 */
8299 				if (eaddr == oldeaddr) {
8300 					/*
8301 					 * Set policy for new segment
8302 					 */
8303 					(void) lgrp_privm_policy_set(policy,
8304 					    &new_svd->policy_info,
8305 					    new_seg->s_size);
8306 
8307 					next = AS_SEGNEXT(new_seg->s_as,
8308 					    new_seg);
8309 
8310 					if (next &&
8311 					    next->s_ops == &segvn_ops &&
8312 					    eaddr == next->s_base)
8313 						(void) segvn_concat(new_seg,
8314 						    next, 1);
8315 				}
8316 			}
8317 
8318 			/*
8319 			 * Split off end of existing segment if advice only
8320 			 * applies to a portion of segment ending before
8321 			 * end of the existing segment
8322 			 */
8323 			if (eaddr < oldeaddr) {
8324 				/*
8325 				 * Must flush I/O page cache
8326 				 * before splitting segment
8327 				 */
8328 				if (svd->softlockcnt > 0)
8329 					segvn_purge(seg);
8330 
8331 				/*
8332 				 * If beginning of old segment was already
8333 				 * split off, use new segment to split end off
8334 				 * from.
8335 				 */
8336 				if (new_seg != NULL && new_seg != seg) {
8337 					/*
8338 					 * Split segment
8339 					 */
8340 					(void) segvn_split_seg(new_seg, eaddr);
8341 
8342 					/*
8343 					 * Set policy for new segment
8344 					 */
8345 					(void) lgrp_privm_policy_set(policy,
8346 					    &new_svd->policy_info,
8347 					    new_seg->s_size);
8348 				} else {
8349 					/*
8350 					 * Split segment and return IE_REATTACH
8351 					 * to tell as_ctl() that current
8352 					 * segment changed
8353 					 */
8354 					(void) segvn_split_seg(seg, eaddr);
8355 					err = IE_REATTACH;
8356 
8357 					(void) lgrp_privm_policy_set(policy,
8358 					    &svd->policy_info, seg->s_size);
8359 
8360 					/*
8361 					 * If new segment starts where old one
8362 					 * did, try to concatenate it with
8363 					 * previous segment.
8364 					 */
8365 					if (addr == seg->s_base) {
8366 						prev = AS_SEGPREV(seg->s_as,
8367 						    seg);
8368 
8369 						/*
8370 						 * Drop lock for private data
8371 						 * of current segment before
8372 						 * concatenating (deleting) it
8373 						 */
8374 						if (prev &&
8375 						    prev->s_ops ==
8376 						    &segvn_ops &&
8377 						    addr == prev->s_base +
8378 						    prev->s_size) {
8379 							SEGVN_LOCK_EXIT(
8380 							    seg->s_as,
8381 							    &svd->lock);
8382 							(void) segvn_concat(
8383 							    prev, seg, 1);
8384 							return (err);
8385 						}
8386 					}
8387 				}
8388 			}
8389 			break;
8390 		case MADV_SEQUENTIAL:
8391 			ASSERT(seg->s_szc == 0);
8392 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8393 			hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
8394 			/* FALLTHROUGH */
8395 		case MADV_NORMAL:
8396 		case MADV_RANDOM:
8397 			bvpp = &svd->vpage[page];
8398 			evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8399 			for (; bvpp < evpp; bvpp++)
8400 				VPP_SETADVICE(bvpp, behav);
8401 			svd->advice = MADV_NORMAL;
8402 			break;
8403 		case MADV_WILLNEED:	/* handled in memcntl */
8404 		case MADV_DONTNEED:	/* handled in memcntl */
8405 		case MADV_FREE:		/* handled above */
8406 			break;
8407 		default:
8408 			err = EINVAL;
8409 		}
8410 	}
8411 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8412 	return (err);
8413 }
8414 
8415 /*
8416  * Create a vpage structure for this seg.
8417  */
8418 static void
8419 segvn_vpage(struct seg *seg)
8420 {
8421 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8422 	struct vpage *vp, *evp;
8423 
8424 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8425 
8426 	/*
8427 	 * If no vpage structure exists, allocate one.  Copy the protections
8428 	 * and the advice from the segment itself to the individual pages.
8429 	 */
8430 	if (svd->vpage == NULL) {
8431 		svd->pageadvice = 1;
8432 		svd->vpage = kmem_zalloc(seg_pages(seg) * sizeof (struct vpage),
8433 		    KM_SLEEP);
8434 		evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
8435 		for (vp = svd->vpage; vp < evp; vp++) {
8436 			VPP_SETPROT(vp, svd->prot);
8437 			VPP_SETADVICE(vp, svd->advice);
8438 		}
8439 	}
8440 }
8441 
8442 /*
8443  * Dump the pages belonging to this segvn segment.
8444  */
8445 static void
8446 segvn_dump(struct seg *seg)
8447 {
8448 	struct segvn_data *svd;
8449 	page_t *pp;
8450 	struct anon_map *amp;
8451 	ulong_t	anon_index;
8452 	struct vnode *vp;
8453 	u_offset_t off, offset;
8454 	pfn_t pfn;
8455 	pgcnt_t page, npages;
8456 	caddr_t addr;
8457 
8458 	npages = seg_pages(seg);
8459 	svd = (struct segvn_data *)seg->s_data;
8460 	vp = svd->vp;
8461 	off = offset = svd->offset;
8462 	addr = seg->s_base;
8463 
8464 	if ((amp = svd->amp) != NULL) {
8465 		anon_index = svd->anon_index;
8466 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8467 	}
8468 
8469 	for (page = 0; page < npages; page++, offset += PAGESIZE) {
8470 		struct anon *ap;
8471 		int we_own_it = 0;
8472 
8473 		if (amp && (ap = anon_get_ptr(svd->amp->ahp, anon_index++))) {
8474 			swap_xlate_nopanic(ap, &vp, &off);
8475 		} else {
8476 			vp = svd->vp;
8477 			off = offset;
8478 		}
8479 
8480 		/*
8481 		 * If pp == NULL, the page either does not exist
8482 		 * or is exclusively locked.  So determine if it
8483 		 * exists before searching for it.
8484 		 */
8485 
8486 		if ((pp = page_lookup_nowait(vp, off, SE_SHARED)))
8487 			we_own_it = 1;
8488 		else
8489 			pp = page_exists(vp, off);
8490 
8491 		if (pp) {
8492 			pfn = page_pptonum(pp);
8493 			dump_addpage(seg->s_as, addr, pfn);
8494 			if (we_own_it)
8495 				page_unlock(pp);
8496 		}
8497 		addr += PAGESIZE;
8498 		dump_timeleft = dump_timeout;
8499 	}
8500 
8501 	if (amp != NULL)
8502 		ANON_LOCK_EXIT(&amp->a_rwlock);
8503 }
8504 
8505 #ifdef DEBUG
8506 static uint32_t segvn_pglock_mtbf = 0;
8507 #endif
8508 
8509 #define	PCACHE_SHWLIST		((page_t *)-2)
8510 #define	NOPCACHE_SHWLIST	((page_t *)-1)
8511 
8512 /*
8513  * Lock/Unlock anon pages over a given range. Return shadow list. This routine
8514  * uses global segment pcache to cache shadow lists (i.e. pp arrays) of pages
8515  * to avoid the overhead of per page locking, unlocking for subsequent IOs to
8516  * the same parts of the segment. Currently shadow list creation is only
8517  * supported for pure anon segments. MAP_PRIVATE segment pcache entries are
8518  * tagged with segment pointer, starting virtual address and length. This
8519  * approach for MAP_SHARED segments may add many pcache entries for the same
8520  * set of pages and lead to long hash chains that decrease pcache lookup
8521  * performance. To avoid this issue for shared segments shared anon map and
8522  * starting anon index are used for pcache entry tagging. This allows all
8523  * segments to share pcache entries for the same anon range and reduces pcache
8524  * chain's length as well as memory overhead from duplicate shadow lists and
8525  * pcache entries.
8526  *
8527  * softlockcnt field in segvn_data structure counts the number of F_SOFTLOCK'd
8528  * pages via segvn_fault() and pagelock'd pages via this routine. But pagelock
8529  * part of softlockcnt accounting is done differently for private and shared
8530  * segments. In private segment case softlock is only incremented when a new
8531  * shadow list is created but not when an existing one is found via
8532  * seg_plookup(). pcache entries have reference count incremented/decremented
8533  * by each seg_plookup()/seg_pinactive() operation. Only entries that have 0
8534  * reference count can be purged (and purging is needed before segment can be
8535  * freed). When a private segment pcache entry is purged segvn_reclaim() will
8536  * decrement softlockcnt. Since in private segment case each of its pcache
8537  * entries only belongs to this segment we can expect that when
8538  * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8539  * segment purge will succeed and softlockcnt will drop to 0. In shared
8540  * segment case reference count in pcache entry counts active locks from many
8541  * different segments so we can't expect segment purging to succeed even when
8542  * segvn_pagelock(L_PAGEUNLOCK) was called for all outstanding IOs in this
8543  * segment. To be able to determine when there're no pending pagelocks in
8544  * shared segment case we don't rely on purging to make softlockcnt drop to 0
8545  * but instead softlockcnt is incremented and decremented for every
8546  * segvn_pagelock(L_PAGELOCK/L_PAGEUNLOCK) call regardless if a new shadow
8547  * list was created or an existing one was found. When softlockcnt drops to 0
8548  * this segment no longer has any claims for pcached shadow lists and the
8549  * segment can be freed even if there're still active pcache entries
8550  * shared by this segment anon map. Shared segment pcache entries belong to
8551  * anon map and are typically removed when anon map is freed after all
8552  * processes destroy the segments that use this anon map.
8553  */
8554 static int
8555 segvn_pagelock(struct seg *seg, caddr_t addr, size_t len, struct page ***ppp,
8556     enum lock_type type, enum seg_rw rw)
8557 {
8558 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8559 	size_t np;
8560 	pgcnt_t adjustpages;
8561 	pgcnt_t npages;
8562 	ulong_t anon_index;
8563 	uint_t protchk = (rw == S_READ) ? PROT_READ : PROT_WRITE;
8564 	uint_t error;
8565 	struct anon_map *amp;
8566 	pgcnt_t anpgcnt;
8567 	struct page **pplist, **pl, *pp;
8568 	caddr_t a;
8569 	size_t page;
8570 	caddr_t lpgaddr, lpgeaddr;
8571 	anon_sync_obj_t cookie;
8572 	int anlock;
8573 	struct anon_map *pamp;
8574 	caddr_t paddr;
8575 	seg_preclaim_cbfunc_t preclaim_callback;
8576 	size_t pgsz;
8577 	int use_pcache;
8578 	size_t wlen;
8579 	uint_t pflags = 0;
8580 	int sftlck_sbase = 0;
8581 	int sftlck_send = 0;
8582 
8583 #ifdef DEBUG
8584 	if (type == L_PAGELOCK && segvn_pglock_mtbf) {
8585 		hrtime_t ts = gethrtime();
8586 		if ((ts % segvn_pglock_mtbf) == 0) {
8587 			return (ENOTSUP);
8588 		}
8589 		if ((ts % segvn_pglock_mtbf) == 1) {
8590 			return (EFAULT);
8591 		}
8592 	}
8593 #endif
8594 
8595 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_START,
8596 	    "segvn_pagelock: start seg %p addr %p", seg, addr);
8597 
8598 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
8599 	ASSERT(type == L_PAGELOCK || type == L_PAGEUNLOCK);
8600 
8601 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8602 
8603 	/*
8604 	 * for now we only support pagelock to anon memory. We would have to
8605 	 * check protections for vnode objects and call into the vnode driver.
8606 	 * That's too much for a fast path. Let the fault entry point handle
8607 	 * it.
8608 	 */
8609 	if (svd->vp != NULL) {
8610 		if (type == L_PAGELOCK) {
8611 			error = ENOTSUP;
8612 			goto out;
8613 		}
8614 		panic("segvn_pagelock(L_PAGEUNLOCK): vp != NULL");
8615 	}
8616 	if ((amp = svd->amp) == NULL) {
8617 		if (type == L_PAGELOCK) {
8618 			error = EFAULT;
8619 			goto out;
8620 		}
8621 		panic("segvn_pagelock(L_PAGEUNLOCK): amp == NULL");
8622 	}
8623 	if (rw != S_READ && rw != S_WRITE) {
8624 		if (type == L_PAGELOCK) {
8625 			error = ENOTSUP;
8626 			goto out;
8627 		}
8628 		panic("segvn_pagelock(L_PAGEUNLOCK): bad rw");
8629 	}
8630 
8631 	if (seg->s_szc != 0) {
8632 		/*
8633 		 * We are adjusting the pagelock region to the large page size
8634 		 * boundary because the unlocked part of a large page cannot
8635 		 * be freed anyway unless all constituent pages of a large
8636 		 * page are locked. Bigger regions reduce pcache chain length
8637 		 * and improve lookup performance. The tradeoff is that the
8638 		 * very first segvn_pagelock() call for a given page is more
8639 		 * expensive if only 1 page_t is needed for IO. This is only
8640 		 * an issue if pcache entry doesn't get reused by several
8641 		 * subsequent calls. We optimize here for the case when pcache
8642 		 * is heavily used by repeated IOs to the same address range.
8643 		 *
8644 		 * Note segment's page size cannot change while we are holding
8645 		 * as lock.  And then it cannot change while softlockcnt is
8646 		 * not 0. This will allow us to correctly recalculate large
8647 		 * page size region for the matching pageunlock/reclaim call
8648 		 * since as_pageunlock() caller must always match
8649 		 * as_pagelock() call's addr and len.
8650 		 *
8651 		 * For pageunlock *ppp points to the pointer of page_t that
8652 		 * corresponds to the real unadjusted start address. Similar
8653 		 * for pagelock *ppp must point to the pointer of page_t that
8654 		 * corresponds to the real unadjusted start address.
8655 		 */
8656 		pgsz = page_get_pagesize(seg->s_szc);
8657 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
8658 		adjustpages = btop((uintptr_t)(addr - lpgaddr));
8659 	} else if (len < segvn_pglock_comb_thrshld) {
8660 		lpgaddr = addr;
8661 		lpgeaddr = addr + len;
8662 		adjustpages = 0;
8663 		pgsz = PAGESIZE;
8664 	} else {
8665 		/*
8666 		 * Align the address range of large enough requests to allow
8667 		 * combining of different shadow lists into 1 to reduce memory
8668 		 * overhead from potentially overlapping large shadow lists
8669 		 * (worst case is we have a 1MB IO into buffers with start
8670 		 * addresses separated by 4K).  Alignment is only possible if
8671 		 * padded chunks have sufficient access permissions. Note
8672 		 * permissions won't change between L_PAGELOCK and
8673 		 * L_PAGEUNLOCK calls since non 0 softlockcnt will force
8674 		 * segvn_setprot() to wait until softlockcnt drops to 0. This
8675 		 * allows us to determine in L_PAGEUNLOCK the same range we
8676 		 * computed in L_PAGELOCK.
8677 		 *
8678 		 * If alignment is limited by segment ends set
8679 		 * sftlck_sbase/sftlck_send flags. In L_PAGELOCK case when
8680 		 * these flags are set bump softlockcnt_sbase/softlockcnt_send
8681 		 * per segment counters. In L_PAGEUNLOCK case decrease
8682 		 * softlockcnt_sbase/softlockcnt_send counters if
8683 		 * sftlck_sbase/sftlck_send flags are set.  When
8684 		 * softlockcnt_sbase/softlockcnt_send are non 0
8685 		 * segvn_concat()/segvn_extend_prev()/segvn_extend_next()
8686 		 * won't merge the segments. This restriction combined with
8687 		 * restriction on segment unmapping and splitting for segments
8688 		 * that have non 0 softlockcnt allows L_PAGEUNLOCK to
8689 		 * correctly determine the same range that was previously
8690 		 * locked by matching L_PAGELOCK.
8691 		 */
8692 		pflags = SEGP_PSHIFT | (segvn_pglock_comb_bshift << 16);
8693 		pgsz = PAGESIZE;
8694 		if (svd->type == MAP_PRIVATE) {
8695 			lpgaddr = (caddr_t)P2ALIGN((uintptr_t)addr,
8696 			    segvn_pglock_comb_balign);
8697 			if (lpgaddr < seg->s_base) {
8698 				lpgaddr = seg->s_base;
8699 				sftlck_sbase = 1;
8700 			}
8701 		} else {
8702 			ulong_t aix = svd->anon_index + seg_page(seg, addr);
8703 			ulong_t aaix = P2ALIGN(aix, segvn_pglock_comb_palign);
8704 			if (aaix < svd->anon_index) {
8705 				lpgaddr = seg->s_base;
8706 				sftlck_sbase = 1;
8707 			} else {
8708 				lpgaddr = addr - ptob(aix - aaix);
8709 				ASSERT(lpgaddr >= seg->s_base);
8710 			}
8711 		}
8712 		if (svd->pageprot && lpgaddr != addr) {
8713 			struct vpage *vp = &svd->vpage[seg_page(seg, lpgaddr)];
8714 			struct vpage *evp = &svd->vpage[seg_page(seg, addr)];
8715 			while (vp < evp) {
8716 				if ((VPP_PROT(vp) & protchk) == 0) {
8717 					break;
8718 				}
8719 				vp++;
8720 			}
8721 			if (vp < evp) {
8722 				lpgaddr = addr;
8723 				pflags = 0;
8724 			}
8725 		}
8726 		lpgeaddr = addr + len;
8727 		if (pflags) {
8728 			if (svd->type == MAP_PRIVATE) {
8729 				lpgeaddr = (caddr_t)P2ROUNDUP(
8730 				    (uintptr_t)lpgeaddr,
8731 				    segvn_pglock_comb_balign);
8732 			} else {
8733 				ulong_t aix = svd->anon_index +
8734 				    seg_page(seg, lpgeaddr);
8735 				ulong_t aaix = P2ROUNDUP(aix,
8736 				    segvn_pglock_comb_palign);
8737 				if (aaix < aix) {
8738 					lpgeaddr = 0;
8739 				} else {
8740 					lpgeaddr += ptob(aaix - aix);
8741 				}
8742 			}
8743 			if (lpgeaddr == 0 ||
8744 			    lpgeaddr > seg->s_base + seg->s_size) {
8745 				lpgeaddr = seg->s_base + seg->s_size;
8746 				sftlck_send = 1;
8747 			}
8748 		}
8749 		if (svd->pageprot && lpgeaddr != addr + len) {
8750 			struct vpage *vp;
8751 			struct vpage *evp;
8752 
8753 			vp = &svd->vpage[seg_page(seg, addr + len)];
8754 			evp = &svd->vpage[seg_page(seg, lpgeaddr)];
8755 
8756 			while (vp < evp) {
8757 				if ((VPP_PROT(vp) & protchk) == 0) {
8758 					break;
8759 				}
8760 				vp++;
8761 			}
8762 			if (vp < evp) {
8763 				lpgeaddr = addr + len;
8764 			}
8765 		}
8766 		adjustpages = btop((uintptr_t)(addr - lpgaddr));
8767 	}
8768 
8769 	/*
8770 	 * For MAP_SHARED segments we create pcache entries tagged by amp and
8771 	 * anon index so that we can share pcache entries with other segments
8772 	 * that map this amp.  For private segments pcache entries are tagged
8773 	 * with segment and virtual address.
8774 	 */
8775 	if (svd->type == MAP_SHARED) {
8776 		pamp = amp;
8777 		paddr = (caddr_t)((lpgaddr - seg->s_base) +
8778 		    ptob(svd->anon_index));
8779 		preclaim_callback = shamp_reclaim;
8780 	} else {
8781 		pamp = NULL;
8782 		paddr = lpgaddr;
8783 		preclaim_callback = segvn_reclaim;
8784 	}
8785 
8786 	if (type == L_PAGEUNLOCK) {
8787 		VM_STAT_ADD(segvnvmstats.pagelock[0]);
8788 
8789 		/*
8790 		 * update hat ref bits for /proc. We need to make sure
8791 		 * that threads tracing the ref and mod bits of the
8792 		 * address space get the right data.
8793 		 * Note: page ref and mod bits are updated at reclaim time
8794 		 */
8795 		if (seg->s_as->a_vbits) {
8796 			for (a = addr; a < addr + len; a += PAGESIZE) {
8797 				if (rw == S_WRITE) {
8798 					hat_setstat(seg->s_as, a,
8799 					    PAGESIZE, P_REF | P_MOD);
8800 				} else {
8801 					hat_setstat(seg->s_as, a,
8802 					    PAGESIZE, P_REF);
8803 				}
8804 			}
8805 		}
8806 
8807 		/*
8808 		 * Check the shadow list entry after the last page used in
8809 		 * this IO request. If it's NOPCACHE_SHWLIST the shadow list
8810 		 * was not inserted into pcache and is not large page
8811 		 * adjusted.  In this case call reclaim callback directly and
8812 		 * don't adjust the shadow list start and size for large
8813 		 * pages.
8814 		 */
8815 		npages = btop(len);
8816 		if ((*ppp)[npages] == NOPCACHE_SHWLIST) {
8817 			void *ptag;
8818 			if (pamp != NULL) {
8819 				ASSERT(svd->type == MAP_SHARED);
8820 				ptag = (void *)pamp;
8821 				paddr = (caddr_t)((addr - seg->s_base) +
8822 				    ptob(svd->anon_index));
8823 			} else {
8824 				ptag = (void *)seg;
8825 				paddr = addr;
8826 			}
8827 			(*preclaim_callback)(ptag, paddr, len, *ppp, rw, 0);
8828 		} else {
8829 			ASSERT((*ppp)[npages] == PCACHE_SHWLIST ||
8830 			    IS_SWAPFSVP((*ppp)[npages]->p_vnode));
8831 			len = lpgeaddr - lpgaddr;
8832 			npages = btop(len);
8833 			seg_pinactive(seg, pamp, paddr, len,
8834 			    *ppp - adjustpages, rw, pflags, preclaim_callback);
8835 		}
8836 
8837 		if (pamp != NULL) {
8838 			ASSERT(svd->type == MAP_SHARED);
8839 			ASSERT(svd->softlockcnt >= npages);
8840 			atomic_add_long((ulong_t *)&svd->softlockcnt, -npages);
8841 		}
8842 
8843 		if (sftlck_sbase) {
8844 			ASSERT(svd->softlockcnt_sbase > 0);
8845 			atomic_add_long((ulong_t *)&svd->softlockcnt_sbase, -1);
8846 		}
8847 		if (sftlck_send) {
8848 			ASSERT(svd->softlockcnt_send > 0);
8849 			atomic_add_long((ulong_t *)&svd->softlockcnt_send, -1);
8850 		}
8851 
8852 		/*
8853 		 * If someone is blocked while unmapping, we purge
8854 		 * segment page cache and thus reclaim pplist synchronously
8855 		 * without waiting for seg_pasync_thread. This speeds up
8856 		 * unmapping in cases where munmap(2) is called, while
8857 		 * raw async i/o is still in progress or where a thread
8858 		 * exits on data fault in a multithreaded application.
8859 		 */
8860 		if (AS_ISUNMAPWAIT(seg->s_as)) {
8861 			if (svd->softlockcnt == 0) {
8862 				mutex_enter(&seg->s_as->a_contents);
8863 				if (AS_ISUNMAPWAIT(seg->s_as)) {
8864 					AS_CLRUNMAPWAIT(seg->s_as);
8865 					cv_broadcast(&seg->s_as->a_cv);
8866 				}
8867 				mutex_exit(&seg->s_as->a_contents);
8868 			} else if (pamp == NULL) {
8869 				/*
8870 				 * softlockcnt is not 0 and this is a
8871 				 * MAP_PRIVATE segment. Try to purge its
8872 				 * pcache entries to reduce softlockcnt.
8873 				 * If it drops to 0 segvn_reclaim()
8874 				 * will wake up a thread waiting on
8875 				 * unmapwait flag.
8876 				 *
8877 				 * We don't purge MAP_SHARED segments with non
8878 				 * 0 softlockcnt since IO is still in progress
8879 				 * for such segments.
8880 				 */
8881 				ASSERT(svd->type == MAP_PRIVATE);
8882 				segvn_purge(seg);
8883 			}
8884 		}
8885 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8886 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_UNLOCK_END,
8887 		    "segvn_pagelock: unlock seg %p addr %p", seg, addr);
8888 		return (0);
8889 	}
8890 
8891 	/* The L_PAGELOCK case ... */
8892 
8893 	VM_STAT_ADD(segvnvmstats.pagelock[1]);
8894 
8895 	/*
8896 	 * For MAP_SHARED segments we have to check protections before
8897 	 * seg_plookup() since pcache entries may be shared by many segments
8898 	 * with potentially different page protections.
8899 	 */
8900 	if (pamp != NULL) {
8901 		ASSERT(svd->type == MAP_SHARED);
8902 		if (svd->pageprot == 0) {
8903 			if ((svd->prot & protchk) == 0) {
8904 				error = EACCES;
8905 				goto out;
8906 			}
8907 		} else {
8908 			/*
8909 			 * check page protections
8910 			 */
8911 			caddr_t ea;
8912 
8913 			if (seg->s_szc) {
8914 				a = lpgaddr;
8915 				ea = lpgeaddr;
8916 			} else {
8917 				a = addr;
8918 				ea = addr + len;
8919 			}
8920 			for (; a < ea; a += pgsz) {
8921 				struct vpage *vp;
8922 
8923 				ASSERT(seg->s_szc == 0 ||
8924 				    sameprot(seg, a, pgsz));
8925 				vp = &svd->vpage[seg_page(seg, a)];
8926 				if ((VPP_PROT(vp) & protchk) == 0) {
8927 					error = EACCES;
8928 					goto out;
8929 				}
8930 			}
8931 		}
8932 	}
8933 
8934 	/*
8935 	 * try to find pages in segment page cache
8936 	 */
8937 	pplist = seg_plookup(seg, pamp, paddr, lpgeaddr - lpgaddr, rw, pflags);
8938 	if (pplist != NULL) {
8939 		if (pamp != NULL) {
8940 			npages = btop((uintptr_t)(lpgeaddr - lpgaddr));
8941 			ASSERT(svd->type == MAP_SHARED);
8942 			atomic_add_long((ulong_t *)&svd->softlockcnt,
8943 			    npages);
8944 		}
8945 		if (sftlck_sbase) {
8946 			atomic_add_long((ulong_t *)&svd->softlockcnt_sbase, 1);
8947 		}
8948 		if (sftlck_send) {
8949 			atomic_add_long((ulong_t *)&svd->softlockcnt_send, 1);
8950 		}
8951 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8952 		*ppp = pplist + adjustpages;
8953 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_HIT_END,
8954 		    "segvn_pagelock: cache hit seg %p addr %p", seg, addr);
8955 		return (0);
8956 	}
8957 
8958 	/*
8959 	 * For MAP_SHARED segments we already verified above that segment
8960 	 * protections allow this pagelock operation.
8961 	 */
8962 	if (pamp == NULL) {
8963 		ASSERT(svd->type == MAP_PRIVATE);
8964 		if (svd->pageprot == 0) {
8965 			if ((svd->prot & protchk) == 0) {
8966 				error = EACCES;
8967 				goto out;
8968 			}
8969 			if (svd->prot & PROT_WRITE) {
8970 				wlen = lpgeaddr - lpgaddr;
8971 			} else {
8972 				wlen = 0;
8973 				ASSERT(rw == S_READ);
8974 			}
8975 		} else {
8976 			int wcont = 1;
8977 			/*
8978 			 * check page protections
8979 			 */
8980 			for (a = lpgaddr, wlen = 0; a < lpgeaddr; a += pgsz) {
8981 				struct vpage *vp;
8982 
8983 				ASSERT(seg->s_szc == 0 ||
8984 				    sameprot(seg, a, pgsz));
8985 				vp = &svd->vpage[seg_page(seg, a)];
8986 				if ((VPP_PROT(vp) & protchk) == 0) {
8987 					error = EACCES;
8988 					goto out;
8989 				}
8990 				if (wcont && (VPP_PROT(vp) & PROT_WRITE)) {
8991 					wlen += pgsz;
8992 				} else {
8993 					wcont = 0;
8994 					ASSERT(rw == S_READ);
8995 				}
8996 			}
8997 		}
8998 		ASSERT(rw == S_READ || wlen == lpgeaddr - lpgaddr);
8999 		ASSERT(rw == S_WRITE || wlen <= lpgeaddr - lpgaddr);
9000 	}
9001 
9002 	/*
9003 	 * Only build large page adjusted shadow list if we expect to insert
9004 	 * it into pcache. For large enough pages it's a big overhead to
9005 	 * create a shadow list of the entire large page. But this overhead
9006 	 * should be amortized over repeated pcache hits on subsequent reuse
9007 	 * of this shadow list (IO into any range within this shadow list will
9008 	 * find it in pcache since we large page align the request for pcache
9009 	 * lookups). pcache performance is improved with bigger shadow lists
9010 	 * as it reduces the time to pcache the entire big segment and reduces
9011 	 * pcache chain length.
9012 	 */
9013 	if (seg_pinsert_check(seg, pamp, paddr,
9014 	    lpgeaddr - lpgaddr, pflags) == SEGP_SUCCESS) {
9015 		addr = lpgaddr;
9016 		len = lpgeaddr - lpgaddr;
9017 		use_pcache = 1;
9018 	} else {
9019 		use_pcache = 0;
9020 		/*
9021 		 * Since this entry will not be inserted into the pcache, we
9022 		 * will not do any adjustments to the starting address or
9023 		 * size of the memory to be locked.
9024 		 */
9025 		adjustpages = 0;
9026 	}
9027 	npages = btop(len);
9028 
9029 	pplist = kmem_alloc(sizeof (page_t *) * (npages + 1), KM_SLEEP);
9030 	pl = pplist;
9031 	*ppp = pplist + adjustpages;
9032 	/*
9033 	 * If use_pcache is 0 this shadow list is not large page adjusted.
9034 	 * Record this info in the last entry of shadow array so that
9035 	 * L_PAGEUNLOCK can determine if it should large page adjust the
9036 	 * address range to find the real range that was locked.
9037 	 */
9038 	pl[npages] = use_pcache ? PCACHE_SHWLIST : NOPCACHE_SHWLIST;
9039 
9040 	page = seg_page(seg, addr);
9041 	anon_index = svd->anon_index + page;
9042 
9043 	anlock = 0;
9044 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9045 	ASSERT(amp->a_szc >= seg->s_szc);
9046 	anpgcnt = page_get_pagecnt(amp->a_szc);
9047 	for (a = addr; a < addr + len; a += PAGESIZE, anon_index++) {
9048 		struct anon *ap;
9049 		struct vnode *vp;
9050 		u_offset_t off;
9051 
9052 		/*
9053 		 * Lock and unlock anon array only once per large page.
9054 		 * anon_array_enter() locks the root anon slot according to
9055 		 * a_szc which can't change while anon map is locked.  We lock
9056 		 * anon the first time through this loop and each time we
9057 		 * reach anon index that corresponds to a root of a large
9058 		 * page.
9059 		 */
9060 		if (a == addr || P2PHASE(anon_index, anpgcnt) == 0) {
9061 			ASSERT(anlock == 0);
9062 			anon_array_enter(amp, anon_index, &cookie);
9063 			anlock = 1;
9064 		}
9065 		ap = anon_get_ptr(amp->ahp, anon_index);
9066 
9067 		/*
9068 		 * We must never use seg_pcache for COW pages
9069 		 * because we might end up with original page still
9070 		 * lying in seg_pcache even after private page is
9071 		 * created. This leads to data corruption as
9072 		 * aio_write refers to the page still in cache
9073 		 * while all other accesses refer to the private
9074 		 * page.
9075 		 */
9076 		if (ap == NULL || ap->an_refcnt != 1) {
9077 			struct vpage *vpage;
9078 
9079 			if (seg->s_szc) {
9080 				error = EFAULT;
9081 				break;
9082 			}
9083 			if (svd->vpage != NULL) {
9084 				vpage = &svd->vpage[seg_page(seg, a)];
9085 			} else {
9086 				vpage = NULL;
9087 			}
9088 			ASSERT(anlock);
9089 			anon_array_exit(&cookie);
9090 			anlock = 0;
9091 			pp = NULL;
9092 			error = segvn_faultpage(seg->s_as->a_hat, seg, a, 0,
9093 			    vpage, &pp, 0, F_INVAL, rw, 1);
9094 			if (error) {
9095 				error = fc_decode(error);
9096 				break;
9097 			}
9098 			anon_array_enter(amp, anon_index, &cookie);
9099 			anlock = 1;
9100 			ap = anon_get_ptr(amp->ahp, anon_index);
9101 			if (ap == NULL || ap->an_refcnt != 1) {
9102 				error = EFAULT;
9103 				break;
9104 			}
9105 		}
9106 		swap_xlate(ap, &vp, &off);
9107 		pp = page_lookup_nowait(vp, off, SE_SHARED);
9108 		if (pp == NULL) {
9109 			error = EFAULT;
9110 			break;
9111 		}
9112 		if (ap->an_pvp != NULL) {
9113 			anon_swap_free(ap, pp);
9114 		}
9115 		/*
9116 		 * Unlock anon if this is the last slot in a large page.
9117 		 */
9118 		if (P2PHASE(anon_index, anpgcnt) == anpgcnt - 1) {
9119 			ASSERT(anlock);
9120 			anon_array_exit(&cookie);
9121 			anlock = 0;
9122 		}
9123 		*pplist++ = pp;
9124 	}
9125 	if (anlock) {		/* Ensure the lock is dropped */
9126 		anon_array_exit(&cookie);
9127 	}
9128 	ANON_LOCK_EXIT(&amp->a_rwlock);
9129 
9130 	if (a >= addr + len) {
9131 		atomic_add_long((ulong_t *)&svd->softlockcnt, npages);
9132 		if (pamp != NULL) {
9133 			ASSERT(svd->type == MAP_SHARED);
9134 			atomic_add_long((ulong_t *)&pamp->a_softlockcnt,
9135 			    npages);
9136 			wlen = len;
9137 		}
9138 		if (sftlck_sbase) {
9139 			atomic_add_long((ulong_t *)&svd->softlockcnt_sbase, 1);
9140 		}
9141 		if (sftlck_send) {
9142 			atomic_add_long((ulong_t *)&svd->softlockcnt_send, 1);
9143 		}
9144 		if (use_pcache) {
9145 			(void) seg_pinsert(seg, pamp, paddr, len, wlen, pl,
9146 			    rw, pflags, preclaim_callback);
9147 		}
9148 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9149 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_FILL_END,
9150 		    "segvn_pagelock: cache fill seg %p addr %p", seg, addr);
9151 		return (0);
9152 	}
9153 
9154 	pplist = pl;
9155 	np = ((uintptr_t)(a - addr)) >> PAGESHIFT;
9156 	while (np > (uint_t)0) {
9157 		ASSERT(PAGE_LOCKED(*pplist));
9158 		page_unlock(*pplist);
9159 		np--;
9160 		pplist++;
9161 	}
9162 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9163 out:
9164 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9165 	*ppp = NULL;
9166 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
9167 	    "segvn_pagelock: cache miss seg %p addr %p", seg, addr);
9168 	return (error);
9169 }
9170 
9171 /*
9172  * purge any cached pages in the I/O page cache
9173  */
9174 static void
9175 segvn_purge(struct seg *seg)
9176 {
9177 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9178 
9179 	/*
9180 	 * pcache is only used by pure anon segments.
9181 	 */
9182 	if (svd->amp == NULL || svd->vp != NULL) {
9183 		return;
9184 	}
9185 
9186 	/*
9187 	 * For MAP_SHARED segments non 0 segment's softlockcnt means
9188 	 * active IO is still in progress via this segment. So we only
9189 	 * purge MAP_SHARED segments when their softlockcnt is 0.
9190 	 */
9191 	if (svd->type == MAP_PRIVATE) {
9192 		if (svd->softlockcnt) {
9193 			seg_ppurge(seg, NULL, 0);
9194 		}
9195 	} else if (svd->softlockcnt == 0 && svd->amp->a_softlockcnt != 0) {
9196 		seg_ppurge(seg, svd->amp, 0);
9197 	}
9198 }
9199 
9200 /*
9201  * If async argument is not 0 we are called from pcache async thread and don't
9202  * hold AS lock.
9203  */
9204 
9205 /*ARGSUSED*/
9206 static int
9207 segvn_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9208 	enum seg_rw rw, int async)
9209 {
9210 	struct seg *seg = (struct seg *)ptag;
9211 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9212 	pgcnt_t np, npages;
9213 	struct page **pl;
9214 
9215 	npages = np = btop(len);
9216 	ASSERT(npages);
9217 
9218 	ASSERT(svd->vp == NULL && svd->amp != NULL);
9219 	ASSERT(svd->softlockcnt >= npages);
9220 	ASSERT(async || AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9221 
9222 	pl = pplist;
9223 
9224 	ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9225 	ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9226 
9227 	while (np > (uint_t)0) {
9228 		if (rw == S_WRITE) {
9229 			hat_setrefmod(*pplist);
9230 		} else {
9231 			hat_setref(*pplist);
9232 		}
9233 		page_unlock(*pplist);
9234 		np--;
9235 		pplist++;
9236 	}
9237 
9238 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9239 
9240 	/*
9241 	 * If we are pcache async thread we don't hold AS lock. This means if
9242 	 * softlockcnt drops to 0 after the decrement below address space may
9243 	 * get freed. We can't allow it since after softlock derement to 0 we
9244 	 * still need to access as structure for possible wakeup of unmap
9245 	 * waiters. To prevent the disappearance of as we take this segment
9246 	 * segfree_syncmtx. segvn_free() also takes this mutex as a barrier to
9247 	 * make sure this routine completes before segment is freed.
9248 	 *
9249 	 * The second complication we have to deal with in async case is a
9250 	 * possibility of missed wake up of unmap wait thread. When we don't
9251 	 * hold as lock here we may take a_contents lock before unmap wait
9252 	 * thread that was first to see softlockcnt was still not 0. As a
9253 	 * result we'll fail to wake up an unmap wait thread. To avoid this
9254 	 * race we set nounmapwait flag in as structure if we drop softlockcnt
9255 	 * to 0 when we were called by pcache async thread.  unmapwait thread
9256 	 * will not block if this flag is set.
9257 	 */
9258 	if (async) {
9259 		mutex_enter(&svd->segfree_syncmtx);
9260 	}
9261 
9262 	if (!atomic_add_long_nv((ulong_t *)&svd->softlockcnt, -npages)) {
9263 		if (async || AS_ISUNMAPWAIT(seg->s_as)) {
9264 			mutex_enter(&seg->s_as->a_contents);
9265 			if (async) {
9266 				AS_SETNOUNMAPWAIT(seg->s_as);
9267 			}
9268 			if (AS_ISUNMAPWAIT(seg->s_as)) {
9269 				AS_CLRUNMAPWAIT(seg->s_as);
9270 				cv_broadcast(&seg->s_as->a_cv);
9271 			}
9272 			mutex_exit(&seg->s_as->a_contents);
9273 		}
9274 	}
9275 
9276 	if (async) {
9277 		mutex_exit(&svd->segfree_syncmtx);
9278 	}
9279 	return (0);
9280 }
9281 
9282 /*ARGSUSED*/
9283 static int
9284 shamp_reclaim(void *ptag, caddr_t addr, size_t len, struct page **pplist,
9285 	enum seg_rw rw, int async)
9286 {
9287 	amp_t *amp = (amp_t *)ptag;
9288 	pgcnt_t np, npages;
9289 	struct page **pl;
9290 
9291 	npages = np = btop(len);
9292 	ASSERT(npages);
9293 	ASSERT(amp->a_softlockcnt >= npages);
9294 
9295 	pl = pplist;
9296 
9297 	ASSERT(pl[np] == NOPCACHE_SHWLIST || pl[np] == PCACHE_SHWLIST);
9298 	ASSERT(!async || pl[np] == PCACHE_SHWLIST);
9299 
9300 	while (np > (uint_t)0) {
9301 		if (rw == S_WRITE) {
9302 			hat_setrefmod(*pplist);
9303 		} else {
9304 			hat_setref(*pplist);
9305 		}
9306 		page_unlock(*pplist);
9307 		np--;
9308 		pplist++;
9309 	}
9310 
9311 	kmem_free(pl, sizeof (page_t *) * (npages + 1));
9312 
9313 	/*
9314 	 * If somebody sleeps in anonmap_purge() wake them up if a_softlockcnt
9315 	 * drops to 0. anon map can't be freed until a_softlockcnt drops to 0
9316 	 * and anonmap_purge() acquires a_purgemtx.
9317 	 */
9318 	mutex_enter(&amp->a_purgemtx);
9319 	if (!atomic_add_long_nv((ulong_t *)&amp->a_softlockcnt, -npages) &&
9320 	    amp->a_purgewait) {
9321 		amp->a_purgewait = 0;
9322 		cv_broadcast(&amp->a_purgecv);
9323 	}
9324 	mutex_exit(&amp->a_purgemtx);
9325 	return (0);
9326 }
9327 
9328 /*
9329  * get a memory ID for an addr in a given segment
9330  *
9331  * XXX only creates PAGESIZE pages if anon slots are not initialized.
9332  * At fault time they will be relocated into larger pages.
9333  */
9334 static int
9335 segvn_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
9336 {
9337 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9338 	struct anon 	*ap = NULL;
9339 	ulong_t		anon_index;
9340 	struct anon_map	*amp;
9341 	anon_sync_obj_t cookie;
9342 
9343 	if (svd->type == MAP_PRIVATE) {
9344 		memidp->val[0] = (uintptr_t)seg->s_as;
9345 		memidp->val[1] = (uintptr_t)addr;
9346 		return (0);
9347 	}
9348 
9349 	if (svd->type == MAP_SHARED) {
9350 		if (svd->vp) {
9351 			memidp->val[0] = (uintptr_t)svd->vp;
9352 			memidp->val[1] = (u_longlong_t)svd->offset +
9353 			    (uintptr_t)(addr - seg->s_base);
9354 			return (0);
9355 		} else {
9356 
9357 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
9358 			if ((amp = svd->amp) != NULL) {
9359 				anon_index = svd->anon_index +
9360 				    seg_page(seg, addr);
9361 			}
9362 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9363 
9364 			ASSERT(amp != NULL);
9365 
9366 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
9367 			anon_array_enter(amp, anon_index, &cookie);
9368 			ap = anon_get_ptr(amp->ahp, anon_index);
9369 			if (ap == NULL) {
9370 				page_t		*pp;
9371 
9372 				pp = anon_zero(seg, addr, &ap, svd->cred);
9373 				if (pp == NULL) {
9374 					anon_array_exit(&cookie);
9375 					ANON_LOCK_EXIT(&amp->a_rwlock);
9376 					return (ENOMEM);
9377 				}
9378 				ASSERT(anon_get_ptr(amp->ahp, anon_index)
9379 				    == NULL);
9380 				(void) anon_set_ptr(amp->ahp, anon_index,
9381 				    ap, ANON_SLEEP);
9382 				page_unlock(pp);
9383 			}
9384 
9385 			anon_array_exit(&cookie);
9386 			ANON_LOCK_EXIT(&amp->a_rwlock);
9387 
9388 			memidp->val[0] = (uintptr_t)ap;
9389 			memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
9390 			return (0);
9391 		}
9392 	}
9393 	return (EINVAL);
9394 }
9395 
9396 static int
9397 sameprot(struct seg *seg, caddr_t a, size_t len)
9398 {
9399 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
9400 	struct vpage *vpage;
9401 	spgcnt_t pages = btop(len);
9402 	uint_t prot;
9403 
9404 	if (svd->pageprot == 0)
9405 		return (1);
9406 
9407 	ASSERT(svd->vpage != NULL);
9408 
9409 	vpage = &svd->vpage[seg_page(seg, a)];
9410 	prot = VPP_PROT(vpage);
9411 	vpage++;
9412 	pages--;
9413 	while (pages-- > 0) {
9414 		if (prot != VPP_PROT(vpage))
9415 			return (0);
9416 		vpage++;
9417 	}
9418 	return (1);
9419 }
9420 
9421 /*
9422  * Get memory allocation policy info for specified address in given segment
9423  */
9424 static lgrp_mem_policy_info_t *
9425 segvn_getpolicy(struct seg *seg, caddr_t addr)
9426 {
9427 	struct anon_map		*amp;
9428 	ulong_t			anon_index;
9429 	lgrp_mem_policy_info_t	*policy_info;
9430 	struct segvn_data	*svn_data;
9431 	u_offset_t		vn_off;
9432 	vnode_t			*vp;
9433 
9434 	ASSERT(seg != NULL);
9435 
9436 	svn_data = (struct segvn_data *)seg->s_data;
9437 	if (svn_data == NULL)
9438 		return (NULL);
9439 
9440 	/*
9441 	 * Get policy info for private or shared memory
9442 	 */
9443 	if (svn_data->type != MAP_SHARED) {
9444 		if (svn_data->tr_state != SEGVN_TR_ON) {
9445 			policy_info = &svn_data->policy_info;
9446 		} else {
9447 			policy_info = &svn_data->tr_policy_info;
9448 			ASSERT(policy_info->mem_policy ==
9449 			    LGRP_MEM_POLICY_NEXT_SEG);
9450 		}
9451 	} else {
9452 		amp = svn_data->amp;
9453 		anon_index = svn_data->anon_index + seg_page(seg, addr);
9454 		vp = svn_data->vp;
9455 		vn_off = svn_data->offset + (uintptr_t)(addr - seg->s_base);
9456 		policy_info = lgrp_shm_policy_get(amp, anon_index, vp, vn_off);
9457 	}
9458 
9459 	return (policy_info);
9460 }
9461 
9462 /*ARGSUSED*/
9463 static int
9464 segvn_capable(struct seg *seg, segcapability_t capability)
9465 {
9466 	return (0);
9467 }
9468 
9469 /*
9470  * Bind text vnode segment to an amp. If we bind successfully mappings will be
9471  * established to per vnode mapping per lgroup amp pages instead of to vnode
9472  * pages. There's one amp per vnode text mapping per lgroup. Many processes
9473  * may share the same text replication amp. If a suitable amp doesn't already
9474  * exist in svntr hash table create a new one.  We may fail to bind to amp if
9475  * segment is not eligible for text replication.  Code below first checks for
9476  * these conditions. If binding is successful segment tr_state is set to on
9477  * and svd->amp points to the amp to use. Otherwise tr_state is set to off and
9478  * svd->amp remains as NULL.
9479  */
9480 static void
9481 segvn_textrepl(struct seg *seg)
9482 {
9483 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
9484 	vnode_t			*vp = svd->vp;
9485 	u_offset_t		off = svd->offset;
9486 	size_t			size = seg->s_size;
9487 	u_offset_t		eoff = off + size;
9488 	uint_t			szc = seg->s_szc;
9489 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9490 	svntr_t			*svntrp;
9491 	struct vattr		va;
9492 	proc_t			*p = seg->s_as->a_proc;
9493 	lgrp_id_t		lgrp_id;
9494 	lgrp_id_t		olid;
9495 	int			first;
9496 	struct anon_map		*amp;
9497 
9498 	ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9499 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9500 	ASSERT(p != NULL);
9501 	ASSERT(svd->tr_state == SEGVN_TR_INIT);
9502 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9503 	ASSERT(svd->flags & MAP_TEXT);
9504 	ASSERT(svd->type == MAP_PRIVATE);
9505 	ASSERT(vp != NULL && svd->amp == NULL);
9506 	ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
9507 	ASSERT(!(svd->flags & MAP_NORESERVE) && svd->swresv == 0);
9508 	ASSERT(seg->s_as != &kas);
9509 	ASSERT(off < eoff);
9510 	ASSERT(svntr_hashtab != NULL);
9511 
9512 	/*
9513 	 * If numa optimizations are no longer desired bail out.
9514 	 */
9515 	if (!lgrp_optimizations()) {
9516 		svd->tr_state = SEGVN_TR_OFF;
9517 		return;
9518 	}
9519 
9520 	/*
9521 	 * Avoid creating anon maps with size bigger than the file size.
9522 	 * If VOP_GETATTR() call fails bail out.
9523 	 */
9524 	va.va_mask = AT_SIZE | AT_MTIME | AT_CTIME;
9525 	if (VOP_GETATTR(vp, &va, 0, svd->cred, NULL) != 0) {
9526 		svd->tr_state = SEGVN_TR_OFF;
9527 		SEGVN_TR_ADDSTAT(gaerr);
9528 		return;
9529 	}
9530 	if (btopr(va.va_size) < btopr(eoff)) {
9531 		svd->tr_state = SEGVN_TR_OFF;
9532 		SEGVN_TR_ADDSTAT(overmap);
9533 		return;
9534 	}
9535 
9536 	/*
9537 	 * VVMEXEC may not be set yet if exec() prefaults text segment. Set
9538 	 * this flag now before vn_is_mapped(V_WRITE) so that MAP_SHARED
9539 	 * mapping that checks if trcache for this vnode needs to be
9540 	 * invalidated can't miss us.
9541 	 */
9542 	if (!(vp->v_flag & VVMEXEC)) {
9543 		mutex_enter(&vp->v_lock);
9544 		vp->v_flag |= VVMEXEC;
9545 		mutex_exit(&vp->v_lock);
9546 	}
9547 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9548 	/*
9549 	 * Bail out if potentially MAP_SHARED writable mappings exist to this
9550 	 * vnode.  We don't want to use old file contents from existing
9551 	 * replicas if this mapping was established after the original file
9552 	 * was changed.
9553 	 */
9554 	if (vn_is_mapped(vp, V_WRITE)) {
9555 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9556 		svd->tr_state = SEGVN_TR_OFF;
9557 		SEGVN_TR_ADDSTAT(wrcnt);
9558 		return;
9559 	}
9560 	svntrp = svntr_hashtab[hash].tr_head;
9561 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9562 		ASSERT(svntrp->tr_refcnt != 0);
9563 		if (svntrp->tr_vp != vp) {
9564 			continue;
9565 		}
9566 
9567 		/*
9568 		 * Bail out if the file or its attributes were changed after
9569 		 * this replication entry was created since we need to use the
9570 		 * latest file contents. Note that mtime test alone is not
9571 		 * sufficient because a user can explicitly change mtime via
9572 		 * utimes(2) interfaces back to the old value after modifiying
9573 		 * the file contents. To detect this case we also have to test
9574 		 * ctime which among other things records the time of the last
9575 		 * mtime change by utimes(2). ctime is not changed when the file
9576 		 * is only read or executed so we expect that typically existing
9577 		 * replication amp's can be used most of the time.
9578 		 */
9579 		if (!svntrp->tr_valid ||
9580 		    svntrp->tr_mtime.tv_sec != va.va_mtime.tv_sec ||
9581 		    svntrp->tr_mtime.tv_nsec != va.va_mtime.tv_nsec ||
9582 		    svntrp->tr_ctime.tv_sec != va.va_ctime.tv_sec ||
9583 		    svntrp->tr_ctime.tv_nsec != va.va_ctime.tv_nsec) {
9584 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9585 			svd->tr_state = SEGVN_TR_OFF;
9586 			SEGVN_TR_ADDSTAT(stale);
9587 			return;
9588 		}
9589 		/*
9590 		 * if off, eoff and szc match current segment we found the
9591 		 * existing entry we can use.
9592 		 */
9593 		if (svntrp->tr_off == off && svntrp->tr_eoff == eoff &&
9594 		    svntrp->tr_szc == szc) {
9595 			break;
9596 		}
9597 		/*
9598 		 * Don't create different but overlapping in file offsets
9599 		 * entries to avoid replication of the same file pages more
9600 		 * than once per lgroup.
9601 		 */
9602 		if ((off >= svntrp->tr_off && off < svntrp->tr_eoff) ||
9603 		    (eoff > svntrp->tr_off && eoff <= svntrp->tr_eoff)) {
9604 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9605 			svd->tr_state = SEGVN_TR_OFF;
9606 			SEGVN_TR_ADDSTAT(overlap);
9607 			return;
9608 		}
9609 	}
9610 	/*
9611 	 * If we didn't find existing entry create a new one.
9612 	 */
9613 	if (svntrp == NULL) {
9614 		svntrp = kmem_cache_alloc(svntr_cache, KM_NOSLEEP);
9615 		if (svntrp == NULL) {
9616 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9617 			svd->tr_state = SEGVN_TR_OFF;
9618 			SEGVN_TR_ADDSTAT(nokmem);
9619 			return;
9620 		}
9621 #ifdef DEBUG
9622 		{
9623 			lgrp_id_t i;
9624 			for (i = 0; i < NLGRPS_MAX; i++) {
9625 				ASSERT(svntrp->tr_amp[i] == NULL);
9626 			}
9627 		}
9628 #endif /* DEBUG */
9629 		svntrp->tr_vp = vp;
9630 		svntrp->tr_off = off;
9631 		svntrp->tr_eoff = eoff;
9632 		svntrp->tr_szc = szc;
9633 		svntrp->tr_valid = 1;
9634 		svntrp->tr_mtime = va.va_mtime;
9635 		svntrp->tr_ctime = va.va_ctime;
9636 		svntrp->tr_refcnt = 0;
9637 		svntrp->tr_next = svntr_hashtab[hash].tr_head;
9638 		svntr_hashtab[hash].tr_head = svntrp;
9639 	}
9640 	first = 1;
9641 again:
9642 	/*
9643 	 * We want to pick a replica with pages on main thread's (t_tid = 1,
9644 	 * aka T1) lgrp. Currently text replication is only optimized for
9645 	 * workloads that either have all threads of a process on the same
9646 	 * lgrp or execute their large text primarily on main thread.
9647 	 */
9648 	lgrp_id = p->p_t1_lgrpid;
9649 	if (lgrp_id == LGRP_NONE) {
9650 		/*
9651 		 * In case exec() prefaults text on non main thread use
9652 		 * current thread lgrpid.  It will become main thread anyway
9653 		 * soon.
9654 		 */
9655 		lgrp_id = lgrp_home_id(curthread);
9656 	}
9657 	/*
9658 	 * Set p_tr_lgrpid to lgrpid if it hasn't been set yet.  Otherwise
9659 	 * just set it to NLGRPS_MAX if it's different from current process T1
9660 	 * home lgrp.  p_tr_lgrpid is used to detect if process uses text
9661 	 * replication and T1 new home is different from lgrp used for text
9662 	 * replication. When this happens asyncronous segvn thread rechecks if
9663 	 * segments should change lgrps used for text replication.  If we fail
9664 	 * to set p_tr_lgrpid with cas32 then set it to NLGRPS_MAX without cas
9665 	 * if it's not already NLGRPS_MAX and not equal lgrp_id we want to
9666 	 * use.  We don't need to use cas in this case because another thread
9667 	 * that races in between our non atomic check and set may only change
9668 	 * p_tr_lgrpid to NLGRPS_MAX at this point.
9669 	 */
9670 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9671 	olid = p->p_tr_lgrpid;
9672 	if (lgrp_id != olid && olid != NLGRPS_MAX) {
9673 		lgrp_id_t nlid = (olid == LGRP_NONE) ? lgrp_id : NLGRPS_MAX;
9674 		if (cas32((uint32_t *)&p->p_tr_lgrpid, olid, nlid) != olid) {
9675 			olid = p->p_tr_lgrpid;
9676 			ASSERT(olid != LGRP_NONE);
9677 			if (olid != lgrp_id && olid != NLGRPS_MAX) {
9678 				p->p_tr_lgrpid = NLGRPS_MAX;
9679 			}
9680 		}
9681 		ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9682 		membar_producer();
9683 		/*
9684 		 * lgrp_move_thread() won't schedule async recheck after
9685 		 * p->p_t1_lgrpid update unless p->p_tr_lgrpid is not
9686 		 * LGRP_NONE. Recheck p_t1_lgrpid once now that p->p_tr_lgrpid
9687 		 * is not LGRP_NONE.
9688 		 */
9689 		if (first && p->p_t1_lgrpid != LGRP_NONE &&
9690 		    p->p_t1_lgrpid != lgrp_id) {
9691 			first = 0;
9692 			goto again;
9693 		}
9694 	}
9695 	/*
9696 	 * If no amp was created yet for lgrp_id create a new one as long as
9697 	 * we have enough memory to afford it.
9698 	 */
9699 	if ((amp = svntrp->tr_amp[lgrp_id]) == NULL) {
9700 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9701 		if (trmem > segvn_textrepl_max_bytes) {
9702 			SEGVN_TR_ADDSTAT(normem);
9703 			goto fail;
9704 		}
9705 		if (anon_try_resv_zone(size, NULL) == 0) {
9706 			SEGVN_TR_ADDSTAT(noanon);
9707 			goto fail;
9708 		}
9709 		amp = anonmap_alloc(size, size, ANON_NOSLEEP);
9710 		if (amp == NULL) {
9711 			anon_unresv_zone(size, NULL);
9712 			SEGVN_TR_ADDSTAT(nokmem);
9713 			goto fail;
9714 		}
9715 		ASSERT(amp->refcnt == 1);
9716 		amp->a_szc = szc;
9717 		svntrp->tr_amp[lgrp_id] = amp;
9718 		SEGVN_TR_ADDSTAT(newamp);
9719 	}
9720 	svntrp->tr_refcnt++;
9721 	ASSERT(svd->svn_trnext == NULL);
9722 	ASSERT(svd->svn_trprev == NULL);
9723 	svd->svn_trnext = svntrp->tr_svnhead;
9724 	svd->svn_trprev = NULL;
9725 	if (svntrp->tr_svnhead != NULL) {
9726 		svntrp->tr_svnhead->svn_trprev = svd;
9727 	}
9728 	svntrp->tr_svnhead = svd;
9729 	ASSERT(amp->a_szc == szc && amp->size == size && amp->swresv == size);
9730 	ASSERT(amp->refcnt >= 1);
9731 	svd->amp = amp;
9732 	svd->anon_index = 0;
9733 	svd->tr_policy_info.mem_policy = LGRP_MEM_POLICY_NEXT_SEG;
9734 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
9735 	svd->tr_state = SEGVN_TR_ON;
9736 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9737 	SEGVN_TR_ADDSTAT(repl);
9738 	return;
9739 fail:
9740 	ASSERT(segvn_textrepl_bytes >= size);
9741 	atomic_add_long(&segvn_textrepl_bytes, -size);
9742 	ASSERT(svntrp != NULL);
9743 	ASSERT(svntrp->tr_amp[lgrp_id] == NULL);
9744 	if (svntrp->tr_refcnt == 0) {
9745 		ASSERT(svntrp == svntr_hashtab[hash].tr_head);
9746 		svntr_hashtab[hash].tr_head = svntrp->tr_next;
9747 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9748 		kmem_cache_free(svntr_cache, svntrp);
9749 	} else {
9750 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9751 	}
9752 	svd->tr_state = SEGVN_TR_OFF;
9753 }
9754 
9755 /*
9756  * Convert seg back to regular vnode mapping seg by unbinding it from its text
9757  * replication amp.  This routine is most typically called when segment is
9758  * unmapped but can also be called when segment no longer qualifies for text
9759  * replication (e.g. due to protection changes). If unload_unmap is set use
9760  * HAT_UNLOAD_UNMAP flag in hat_unload_callback().  If we are the last user of
9761  * svntr free all its anon maps and remove it from the hash table.
9762  */
9763 static void
9764 segvn_textunrepl(struct seg *seg, int unload_unmap)
9765 {
9766 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
9767 	vnode_t			*vp = svd->vp;
9768 	u_offset_t		off = svd->offset;
9769 	size_t			size = seg->s_size;
9770 	u_offset_t		eoff = off + size;
9771 	uint_t			szc = seg->s_szc;
9772 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9773 	svntr_t			*svntrp;
9774 	svntr_t			**prv_svntrp;
9775 	lgrp_id_t		lgrp_id = svd->tr_policy_info.mem_lgrpid;
9776 	lgrp_id_t		i;
9777 
9778 	ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9779 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
9780 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9781 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9782 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9783 	ASSERT(svd->amp != NULL);
9784 	ASSERT(svd->amp->refcnt >= 1);
9785 	ASSERT(svd->anon_index == 0);
9786 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9787 	ASSERT(svntr_hashtab != NULL);
9788 
9789 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9790 	prv_svntrp = &svntr_hashtab[hash].tr_head;
9791 	for (; (svntrp = *prv_svntrp) != NULL; prv_svntrp = &svntrp->tr_next) {
9792 		ASSERT(svntrp->tr_refcnt != 0);
9793 		if (svntrp->tr_vp == vp && svntrp->tr_off == off &&
9794 		    svntrp->tr_eoff == eoff && svntrp->tr_szc == szc) {
9795 			break;
9796 		}
9797 	}
9798 	if (svntrp == NULL) {
9799 		panic("segvn_textunrepl: svntr record not found");
9800 	}
9801 	if (svntrp->tr_amp[lgrp_id] != svd->amp) {
9802 		panic("segvn_textunrepl: amp mismatch");
9803 	}
9804 	svd->tr_state = SEGVN_TR_OFF;
9805 	svd->amp = NULL;
9806 	if (svd->svn_trprev == NULL) {
9807 		ASSERT(svntrp->tr_svnhead == svd);
9808 		svntrp->tr_svnhead = svd->svn_trnext;
9809 		if (svntrp->tr_svnhead != NULL) {
9810 			svntrp->tr_svnhead->svn_trprev = NULL;
9811 		}
9812 		svd->svn_trnext = NULL;
9813 	} else {
9814 		svd->svn_trprev->svn_trnext = svd->svn_trnext;
9815 		if (svd->svn_trnext != NULL) {
9816 			svd->svn_trnext->svn_trprev = svd->svn_trprev;
9817 			svd->svn_trnext = NULL;
9818 		}
9819 		svd->svn_trprev = NULL;
9820 	}
9821 	if (--svntrp->tr_refcnt) {
9822 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9823 		goto done;
9824 	}
9825 	*prv_svntrp = svntrp->tr_next;
9826 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9827 	for (i = 0; i < NLGRPS_MAX; i++) {
9828 		struct anon_map *amp = svntrp->tr_amp[i];
9829 		if (amp == NULL) {
9830 			continue;
9831 		}
9832 		ASSERT(amp->refcnt == 1);
9833 		ASSERT(amp->swresv == size);
9834 		ASSERT(amp->size == size);
9835 		ASSERT(amp->a_szc == szc);
9836 		if (amp->a_szc != 0) {
9837 			anon_free_pages(amp->ahp, 0, size, szc);
9838 		} else {
9839 			anon_free(amp->ahp, 0, size);
9840 		}
9841 		svntrp->tr_amp[i] = NULL;
9842 		ASSERT(segvn_textrepl_bytes >= size);
9843 		atomic_add_long(&segvn_textrepl_bytes, -size);
9844 		anon_unresv_zone(amp->swresv, NULL);
9845 		amp->refcnt = 0;
9846 		anonmap_free(amp);
9847 	}
9848 	kmem_cache_free(svntr_cache, svntrp);
9849 done:
9850 	hat_unload_callback(seg->s_as->a_hat, seg->s_base, size,
9851 	    unload_unmap ? HAT_UNLOAD_UNMAP : 0, NULL);
9852 }
9853 
9854 /*
9855  * This is called when a MAP_SHARED writable mapping is created to a vnode
9856  * that is currently used for execution (VVMEXEC flag is set). In this case we
9857  * need to prevent further use of existing replicas.
9858  */
9859 static void
9860 segvn_inval_trcache(vnode_t *vp)
9861 {
9862 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9863 	svntr_t			*svntrp;
9864 
9865 	ASSERT(vp->v_flag & VVMEXEC);
9866 
9867 	if (svntr_hashtab == NULL) {
9868 		return;
9869 	}
9870 
9871 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9872 	svntrp = svntr_hashtab[hash].tr_head;
9873 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9874 		ASSERT(svntrp->tr_refcnt != 0);
9875 		if (svntrp->tr_vp == vp && svntrp->tr_valid) {
9876 			svntrp->tr_valid = 0;
9877 		}
9878 	}
9879 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9880 }
9881 
9882 static void
9883 segvn_trasync_thread(void)
9884 {
9885 	callb_cpr_t cpr_info;
9886 	kmutex_t cpr_lock;	/* just for CPR stuff */
9887 
9888 	mutex_init(&cpr_lock, NULL, MUTEX_DEFAULT, NULL);
9889 
9890 	CALLB_CPR_INIT(&cpr_info, &cpr_lock,
9891 	    callb_generic_cpr, "segvn_async");
9892 
9893 	if (segvn_update_textrepl_interval == 0) {
9894 		segvn_update_textrepl_interval = segvn_update_tr_time * hz;
9895 	} else {
9896 		segvn_update_textrepl_interval *= hz;
9897 	}
9898 	(void) timeout(segvn_trupdate_wakeup, NULL,
9899 	    segvn_update_textrepl_interval);
9900 
9901 	for (;;) {
9902 		mutex_enter(&cpr_lock);
9903 		CALLB_CPR_SAFE_BEGIN(&cpr_info);
9904 		mutex_exit(&cpr_lock);
9905 		sema_p(&segvn_trasync_sem);
9906 		mutex_enter(&cpr_lock);
9907 		CALLB_CPR_SAFE_END(&cpr_info, &cpr_lock);
9908 		mutex_exit(&cpr_lock);
9909 		segvn_trupdate();
9910 	}
9911 }
9912 
9913 static uint64_t segvn_lgrp_trthr_migrs_snpsht = 0;
9914 
9915 static void
9916 segvn_trupdate_wakeup(void *dummy)
9917 {
9918 	uint64_t cur_lgrp_trthr_migrs = lgrp_get_trthr_migrations();
9919 
9920 	if (cur_lgrp_trthr_migrs != segvn_lgrp_trthr_migrs_snpsht) {
9921 		segvn_lgrp_trthr_migrs_snpsht = cur_lgrp_trthr_migrs;
9922 		sema_v(&segvn_trasync_sem);
9923 	}
9924 
9925 	if (!segvn_disable_textrepl_update &&
9926 	    segvn_update_textrepl_interval != 0) {
9927 		(void) timeout(segvn_trupdate_wakeup, dummy,
9928 		    segvn_update_textrepl_interval);
9929 	}
9930 }
9931 
9932 static void
9933 segvn_trupdate(void)
9934 {
9935 	ulong_t		hash;
9936 	svntr_t		*svntrp;
9937 	segvn_data_t	*svd;
9938 
9939 	ASSERT(svntr_hashtab != NULL);
9940 
9941 	for (hash = 0; hash < svntr_hashtab_sz; hash++) {
9942 		mutex_enter(&svntr_hashtab[hash].tr_lock);
9943 		svntrp = svntr_hashtab[hash].tr_head;
9944 		for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9945 			ASSERT(svntrp->tr_refcnt != 0);
9946 			svd = svntrp->tr_svnhead;
9947 			for (; svd != NULL; svd = svd->svn_trnext) {
9948 				segvn_trupdate_seg(svd->seg, svd, svntrp,
9949 				    hash);
9950 			}
9951 		}
9952 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9953 	}
9954 }
9955 
9956 static void
9957 segvn_trupdate_seg(struct seg *seg,
9958 	segvn_data_t *svd,
9959 	svntr_t *svntrp,
9960 	ulong_t hash)
9961 {
9962 	proc_t			*p;
9963 	lgrp_id_t		lgrp_id;
9964 	struct as		*as;
9965 	size_t			size;
9966 	struct anon_map		*amp;
9967 
9968 	ASSERT(svd->vp != NULL);
9969 	ASSERT(svd->vp == svntrp->tr_vp);
9970 	ASSERT(svd->offset == svntrp->tr_off);
9971 	ASSERT(svd->offset + seg->s_size == svntrp->tr_eoff);
9972 	ASSERT(seg != NULL);
9973 	ASSERT(svd->seg == seg);
9974 	ASSERT(seg->s_data == (void *)svd);
9975 	ASSERT(seg->s_szc == svntrp->tr_szc);
9976 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9977 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9978 	ASSERT(svd->amp != NULL);
9979 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
9980 	ASSERT(svd->tr_policy_info.mem_lgrpid != LGRP_NONE);
9981 	ASSERT(svd->tr_policy_info.mem_lgrpid < NLGRPS_MAX);
9982 	ASSERT(svntrp->tr_amp[svd->tr_policy_info.mem_lgrpid] == svd->amp);
9983 	ASSERT(svntrp->tr_refcnt != 0);
9984 	ASSERT(mutex_owned(&svntr_hashtab[hash].tr_lock));
9985 
9986 	as = seg->s_as;
9987 	ASSERT(as != NULL && as != &kas);
9988 	p = as->a_proc;
9989 	ASSERT(p != NULL);
9990 	ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9991 	lgrp_id = p->p_t1_lgrpid;
9992 	if (lgrp_id == LGRP_NONE) {
9993 		return;
9994 	}
9995 	ASSERT(lgrp_id < NLGRPS_MAX);
9996 	if (svd->tr_policy_info.mem_lgrpid == lgrp_id) {
9997 		return;
9998 	}
9999 
10000 	/*
10001 	 * Use tryenter locking since we are locking as/seg and svntr hash
10002 	 * lock in reverse from syncrounous thread order.
10003 	 */
10004 	if (!AS_LOCK_TRYENTER(as, &as->a_lock, RW_READER)) {
10005 		SEGVN_TR_ADDSTAT(nolock);
10006 		if (segvn_lgrp_trthr_migrs_snpsht) {
10007 			segvn_lgrp_trthr_migrs_snpsht = 0;
10008 		}
10009 		return;
10010 	}
10011 	if (!SEGVN_LOCK_TRYENTER(seg->s_as, &svd->lock, RW_WRITER)) {
10012 		AS_LOCK_EXIT(as, &as->a_lock);
10013 		SEGVN_TR_ADDSTAT(nolock);
10014 		if (segvn_lgrp_trthr_migrs_snpsht) {
10015 			segvn_lgrp_trthr_migrs_snpsht = 0;
10016 		}
10017 		return;
10018 	}
10019 	size = seg->s_size;
10020 	if (svntrp->tr_amp[lgrp_id] == NULL) {
10021 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
10022 		if (trmem > segvn_textrepl_max_bytes) {
10023 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10024 			AS_LOCK_EXIT(as, &as->a_lock);
10025 			atomic_add_long(&segvn_textrepl_bytes, -size);
10026 			SEGVN_TR_ADDSTAT(normem);
10027 			return;
10028 		}
10029 		if (anon_try_resv_zone(size, NULL) == 0) {
10030 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10031 			AS_LOCK_EXIT(as, &as->a_lock);
10032 			atomic_add_long(&segvn_textrepl_bytes, -size);
10033 			SEGVN_TR_ADDSTAT(noanon);
10034 			return;
10035 		}
10036 		amp = anonmap_alloc(size, size, KM_NOSLEEP);
10037 		if (amp == NULL) {
10038 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10039 			AS_LOCK_EXIT(as, &as->a_lock);
10040 			atomic_add_long(&segvn_textrepl_bytes, -size);
10041 			anon_unresv_zone(size, NULL);
10042 			SEGVN_TR_ADDSTAT(nokmem);
10043 			return;
10044 		}
10045 		ASSERT(amp->refcnt == 1);
10046 		amp->a_szc = seg->s_szc;
10047 		svntrp->tr_amp[lgrp_id] = amp;
10048 	}
10049 	/*
10050 	 * We don't need to drop the bucket lock but here we give other
10051 	 * threads a chance.  svntr and svd can't be unlinked as long as
10052 	 * segment lock is held as a writer and AS held as well.  After we
10053 	 * retake bucket lock we'll continue from where we left. We'll be able
10054 	 * to reach the end of either list since new entries are always added
10055 	 * to the beginning of the lists.
10056 	 */
10057 	mutex_exit(&svntr_hashtab[hash].tr_lock);
10058 	hat_unload_callback(as->a_hat, seg->s_base, size, 0, NULL);
10059 	mutex_enter(&svntr_hashtab[hash].tr_lock);
10060 
10061 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10062 	ASSERT(svd->amp != NULL);
10063 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
10064 	ASSERT(svd->tr_policy_info.mem_lgrpid != lgrp_id);
10065 	ASSERT(svd->amp != svntrp->tr_amp[lgrp_id]);
10066 
10067 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
10068 	svd->amp = svntrp->tr_amp[lgrp_id];
10069 	p->p_tr_lgrpid = NLGRPS_MAX;
10070 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
10071 	AS_LOCK_EXIT(as, &as->a_lock);
10072 
10073 	ASSERT(svntrp->tr_refcnt != 0);
10074 	ASSERT(svd->vp == svntrp->tr_vp);
10075 	ASSERT(svd->tr_policy_info.mem_lgrpid == lgrp_id);
10076 	ASSERT(svd->amp != NULL && svd->amp == svntrp->tr_amp[lgrp_id]);
10077 	ASSERT(svd->seg == seg);
10078 	ASSERT(svd->tr_state == SEGVN_TR_ON);
10079 
10080 	SEGVN_TR_ADDSTAT(asyncrepl);
10081 }
10082