xref: /titanic_44/usr/src/uts/common/vm/seg_vn.c (revision 54925bf60766fbb4f1f2d7c843721406a7b7a3fb)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
27 /*	  All Rights Reserved  	*/
28 
29 /*
30  * University Copyright- Copyright (c) 1982, 1986, 1988
31  * The Regents of the University of California
32  * All Rights Reserved
33  *
34  * University Acknowledgment- Portions of this document are derived from
35  * software developed by the University of California, Berkeley, and its
36  * contributors.
37  */
38 
39 #pragma ident	"%Z%%M%	%I%	%E% SMI"
40 
41 /*
42  * VM - shared or copy-on-write from a vnode/anonymous memory.
43  */
44 
45 #include <sys/types.h>
46 #include <sys/param.h>
47 #include <sys/t_lock.h>
48 #include <sys/errno.h>
49 #include <sys/systm.h>
50 #include <sys/mman.h>
51 #include <sys/debug.h>
52 #include <sys/cred.h>
53 #include <sys/vmsystm.h>
54 #include <sys/tuneable.h>
55 #include <sys/bitmap.h>
56 #include <sys/swap.h>
57 #include <sys/kmem.h>
58 #include <sys/sysmacros.h>
59 #include <sys/vtrace.h>
60 #include <sys/cmn_err.h>
61 #include <sys/callb.h>
62 #include <sys/vm.h>
63 #include <sys/dumphdr.h>
64 #include <sys/lgrp.h>
65 
66 #include <vm/hat.h>
67 #include <vm/as.h>
68 #include <vm/seg.h>
69 #include <vm/seg_vn.h>
70 #include <vm/pvn.h>
71 #include <vm/anon.h>
72 #include <vm/page.h>
73 #include <vm/vpage.h>
74 #include <sys/proc.h>
75 #include <sys/task.h>
76 #include <sys/project.h>
77 #include <sys/zone.h>
78 #include <sys/shm_impl.h>
79 /*
80  * Private seg op routines.
81  */
82 static int	segvn_dup(struct seg *seg, struct seg *newseg);
83 static int	segvn_unmap(struct seg *seg, caddr_t addr, size_t len);
84 static void	segvn_free(struct seg *seg);
85 static faultcode_t segvn_fault(struct hat *hat, struct seg *seg,
86 		    caddr_t addr, size_t len, enum fault_type type,
87 		    enum seg_rw rw);
88 static faultcode_t segvn_faulta(struct seg *seg, caddr_t addr);
89 static int	segvn_setprot(struct seg *seg, caddr_t addr,
90 		    size_t len, uint_t prot);
91 static int	segvn_checkprot(struct seg *seg, caddr_t addr,
92 		    size_t len, uint_t prot);
93 static int	segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta);
94 static size_t	segvn_swapout(struct seg *seg);
95 static int	segvn_sync(struct seg *seg, caddr_t addr, size_t len,
96 		    int attr, uint_t flags);
97 static size_t	segvn_incore(struct seg *seg, caddr_t addr, size_t len,
98 		    char *vec);
99 static int	segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
100 		    int attr, int op, ulong_t *lockmap, size_t pos);
101 static int	segvn_getprot(struct seg *seg, caddr_t addr, size_t len,
102 		    uint_t *protv);
103 static u_offset_t	segvn_getoffset(struct seg *seg, caddr_t addr);
104 static int	segvn_gettype(struct seg *seg, caddr_t addr);
105 static int	segvn_getvp(struct seg *seg, caddr_t addr,
106 		    struct vnode **vpp);
107 static int	segvn_advise(struct seg *seg, caddr_t addr, size_t len,
108 		    uint_t behav);
109 static void	segvn_dump(struct seg *seg);
110 static int	segvn_pagelock(struct seg *seg, caddr_t addr, size_t len,
111 		    struct page ***ppp, enum lock_type type, enum seg_rw rw);
112 static int	segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len,
113 		    uint_t szc);
114 static int	segvn_getmemid(struct seg *seg, caddr_t addr,
115 		    memid_t *memidp);
116 static lgrp_mem_policy_info_t	*segvn_getpolicy(struct seg *, caddr_t);
117 static int	segvn_capable(struct seg *seg, segcapability_t capable);
118 
119 struct	seg_ops segvn_ops = {
120 	segvn_dup,
121 	segvn_unmap,
122 	segvn_free,
123 	segvn_fault,
124 	segvn_faulta,
125 	segvn_setprot,
126 	segvn_checkprot,
127 	segvn_kluster,
128 	segvn_swapout,
129 	segvn_sync,
130 	segvn_incore,
131 	segvn_lockop,
132 	segvn_getprot,
133 	segvn_getoffset,
134 	segvn_gettype,
135 	segvn_getvp,
136 	segvn_advise,
137 	segvn_dump,
138 	segvn_pagelock,
139 	segvn_setpagesize,
140 	segvn_getmemid,
141 	segvn_getpolicy,
142 	segvn_capable,
143 };
144 
145 /*
146  * Common zfod structures, provided as a shorthand for others to use.
147  */
148 static segvn_crargs_t zfod_segvn_crargs =
149 	SEGVN_ZFOD_ARGS(PROT_ZFOD, PROT_ALL);
150 static segvn_crargs_t kzfod_segvn_crargs =
151 	SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_USER,
152 	PROT_ALL & ~PROT_USER);
153 static segvn_crargs_t stack_noexec_crargs =
154 	SEGVN_ZFOD_ARGS(PROT_ZFOD & ~PROT_EXEC, PROT_ALL);
155 
156 caddr_t	zfod_argsp = (caddr_t)&zfod_segvn_crargs;	/* user zfod argsp */
157 caddr_t	kzfod_argsp = (caddr_t)&kzfod_segvn_crargs;	/* kernel zfod argsp */
158 caddr_t	stack_exec_argsp = (caddr_t)&zfod_segvn_crargs;	/* executable stack */
159 caddr_t	stack_noexec_argsp = (caddr_t)&stack_noexec_crargs; /* noexec stack */
160 
161 #define	vpgtob(n)	((n) * sizeof (struct vpage))	/* For brevity */
162 
163 size_t	segvn_comb_thrshld = UINT_MAX;	/* patchable -- see 1196681 */
164 
165 static int	segvn_concat(struct seg *, struct seg *, int);
166 static int	segvn_extend_prev(struct seg *, struct seg *,
167 		    struct segvn_crargs *, size_t);
168 static int	segvn_extend_next(struct seg *, struct seg *,
169 		    struct segvn_crargs *, size_t);
170 static void	segvn_softunlock(struct seg *, caddr_t, size_t, enum seg_rw);
171 static void	segvn_pagelist_rele(page_t **);
172 static void	segvn_setvnode_mpss(vnode_t *);
173 static void	segvn_relocate_pages(page_t **, page_t *);
174 static int	segvn_full_szcpages(page_t **, uint_t, int *, uint_t *);
175 static int	segvn_fill_vp_pages(struct segvn_data *, vnode_t *, u_offset_t,
176     uint_t, page_t **, page_t **, uint_t *, int *);
177 static faultcode_t segvn_fault_vnodepages(struct hat *, struct seg *, caddr_t,
178     caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
179 static faultcode_t segvn_fault_anonpages(struct hat *, struct seg *, caddr_t,
180     caddr_t, enum fault_type, enum seg_rw, caddr_t, caddr_t, int);
181 static faultcode_t segvn_faultpage(struct hat *, struct seg *, caddr_t,
182     u_offset_t, struct vpage *, page_t **, uint_t,
183     enum fault_type, enum seg_rw, int, int);
184 static void	segvn_vpage(struct seg *);
185 
186 static void segvn_purge(struct seg *seg);
187 static int segvn_reclaim(struct seg *, caddr_t, size_t, struct page **,
188     enum seg_rw);
189 
190 static int sameprot(struct seg *, caddr_t, size_t);
191 
192 static int segvn_demote_range(struct seg *, caddr_t, size_t, int, uint_t);
193 static int segvn_clrszc(struct seg *);
194 static struct seg *segvn_split_seg(struct seg *, caddr_t);
195 static int segvn_claim_pages(struct seg *, struct vpage *, u_offset_t,
196     ulong_t, uint_t);
197 
198 static void segvn_hat_rgn_unload_callback(caddr_t, caddr_t, caddr_t,
199     size_t, void *, u_offset_t);
200 
201 static int segvn_slock_anonpages(page_t *, int);
202 static void segvn_sunlock_anonpages(page_t *, int);
203 
204 static struct kmem_cache *segvn_cache;
205 
206 #ifdef VM_STATS
207 static struct segvnvmstats_str {
208 	ulong_t	fill_vp_pages[31];
209 	ulong_t fltvnpages[49];
210 	ulong_t	fullszcpages[10];
211 	ulong_t	relocatepages[3];
212 	ulong_t	fltanpages[17];
213 	ulong_t pagelock[3];
214 	ulong_t	demoterange[3];
215 } segvnvmstats;
216 #endif /* VM_STATS */
217 
218 #define	SDR_RANGE	1		/* demote entire range */
219 #define	SDR_END		2		/* demote non aligned ends only */
220 
221 #define	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr) {	    \
222 		if ((len) != 0) { 		      	      		      \
223 			lpgaddr = (caddr_t)P2ALIGN((uintptr_t)(addr), pgsz);  \
224 			ASSERT(lpgaddr >= (seg)->s_base);	      	      \
225 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)((addr) +    \
226 			    (len)), pgsz);				      \
227 			ASSERT(lpgeaddr > lpgaddr);		      	      \
228 			ASSERT(lpgeaddr <= (seg)->s_base + (seg)->s_size);    \
229 		} else {					      	      \
230 			lpgeaddr = lpgaddr = (addr);	      		      \
231 		}							      \
232 	}
233 
234 /*ARGSUSED*/
235 static int
236 segvn_cache_constructor(void *buf, void *cdrarg, int kmflags)
237 {
238 	struct segvn_data *svd = buf;
239 
240 	rw_init(&svd->lock, NULL, RW_DEFAULT, NULL);
241 	mutex_init(&svd->segp_slock, NULL, MUTEX_DEFAULT, NULL);
242 	svd->svn_trnext = svd->svn_trprev = NULL;
243 	return (0);
244 }
245 
246 /*ARGSUSED1*/
247 static void
248 segvn_cache_destructor(void *buf, void *cdrarg)
249 {
250 	struct segvn_data *svd = buf;
251 
252 	rw_destroy(&svd->lock);
253 	mutex_destroy(&svd->segp_slock);
254 }
255 
256 /*ARGSUSED*/
257 static int
258 svntr_cache_constructor(void *buf, void *cdrarg, int kmflags)
259 {
260 	bzero(buf, sizeof (svntr_t));
261 	return (0);
262 }
263 
264 /*
265  * Patching this variable to non-zero allows the system to run with
266  * stacks marked as "not executable".  It's a bit of a kludge, but is
267  * provided as a tweakable for platforms that export those ABIs
268  * (e.g. sparc V8) that have executable stacks enabled by default.
269  * There are also some restrictions for platforms that don't actually
270  * implement 'noexec' protections.
271  *
272  * Once enabled, the system is (therefore) unable to provide a fully
273  * ABI-compliant execution environment, though practically speaking,
274  * most everything works.  The exceptions are generally some interpreters
275  * and debuggers that create executable code on the stack and jump
276  * into it (without explicitly mprotecting the address range to include
277  * PROT_EXEC).
278  *
279  * One important class of applications that are disabled are those
280  * that have been transformed into malicious agents using one of the
281  * numerous "buffer overflow" attacks.  See 4007890.
282  */
283 int noexec_user_stack = 0;
284 int noexec_user_stack_log = 1;
285 
286 int segvn_lpg_disable = 0;
287 uint_t segvn_maxpgszc = 0;
288 
289 ulong_t segvn_vmpss_clrszc_cnt;
290 ulong_t segvn_vmpss_clrszc_err;
291 ulong_t segvn_fltvnpages_clrszc_cnt;
292 ulong_t segvn_fltvnpages_clrszc_err;
293 ulong_t segvn_setpgsz_align_err;
294 ulong_t segvn_setpgsz_anon_align_err;
295 ulong_t segvn_setpgsz_getattr_err;
296 ulong_t segvn_setpgsz_eof_err;
297 ulong_t segvn_faultvnmpss_align_err1;
298 ulong_t segvn_faultvnmpss_align_err2;
299 ulong_t segvn_faultvnmpss_align_err3;
300 ulong_t segvn_faultvnmpss_align_err4;
301 ulong_t segvn_faultvnmpss_align_err5;
302 ulong_t	segvn_vmpss_pageio_deadlk_err;
303 
304 int segvn_use_regions = 1;
305 
306 /*
307  * Segvn supports text replication optimization for NUMA platforms. Text
308  * replica's are represented by anon maps (amp). There's one amp per text file
309  * region per lgroup. A process chooses the amp for each of its text mappings
310  * based on the lgroup assignment of its main thread (t_tid = 1). All
311  * processes that want a replica on a particular lgroup for the same text file
312  * mapping share the same amp. amp's are looked up in svntr_hashtab hash table
313  * with vp,off,size,szc used as a key. Text replication segments are read only
314  * MAP_PRIVATE|MAP_TEXT segments that map vnode. Replication is achieved by
315  * forcing COW faults from vnode to amp and mapping amp pages instead of vnode
316  * pages. Replication amp is assigned to a segment when it gets its first
317  * pagefault. To handle main thread lgroup rehoming segvn_trasync_thread
318  * rechecks periodically if the process still maps an amp local to the main
319  * thread. If not async thread forces process to remap to an amp in the new
320  * home lgroup of the main thread. Current text replication implementation
321  * only provides the benefit to workloads that do most of their work in the
322  * main thread of a process or all the threads of a process run in the same
323  * lgroup. To extend text replication benefit to different types of
324  * multithreaded workloads further work would be needed in the hat layer to
325  * allow the same virtual address in the same hat to simultaneously map
326  * different physical addresses (i.e. page table replication would be needed
327  * for x86).
328  *
329  * amp pages are used instead of vnode pages as long as segment has a very
330  * simple life cycle.  It's created via segvn_create(), handles S_EXEC
331  * (S_READ) pagefaults and is fully unmapped.  If anything more complicated
332  * happens such as protection is changed, real COW fault happens, pagesize is
333  * changed, MC_LOCK is requested or segment is partially unmapped we turn off
334  * text replication by converting the segment back to vnode only segment
335  * (unmap segment's address range and set svd->amp to NULL).
336  *
337  * The original file can be changed after amp is inserted into
338  * svntr_hashtab. Processes that are launched after the file is already
339  * changed can't use the replica's created prior to the file change. To
340  * implement this functionality hash entries are timestamped. Replica's can
341  * only be used if current file modification time is the same as the timestamp
342  * saved when hash entry was created. However just timestamps alone are not
343  * sufficient to detect file modification via mmap(MAP_SHARED) mappings. We
344  * deal with file changes via MAP_SHARED mappings differently. When writable
345  * MAP_SHARED mappings are created to vnodes marked as executable we mark all
346  * existing replica's for this vnode as not usable for future text
347  * mappings. And we don't create new replica's for files that currently have
348  * potentially writable MAP_SHARED mappings (i.e. vn_is_mapped(V_WRITE) is
349  * true).
350  */
351 
352 #define	SEGVN_TEXTREPL_MAXBYTES_FACTOR	(20)
353 size_t	segvn_textrepl_max_bytes_factor = SEGVN_TEXTREPL_MAXBYTES_FACTOR;
354 
355 static ulong_t			svntr_hashtab_sz = 512;
356 static svntr_bucket_t		*svntr_hashtab = NULL;
357 static struct kmem_cache	*svntr_cache;
358 static svntr_stats_t		*segvn_textrepl_stats;
359 static ksema_t 			segvn_trasync_sem;
360 
361 int				segvn_disable_textrepl = 1;
362 size_t				textrepl_size_thresh = (size_t)-1;
363 size_t				segvn_textrepl_bytes = 0;
364 size_t				segvn_textrepl_max_bytes = 0;
365 clock_t				segvn_update_textrepl_interval = 0;
366 int				segvn_update_tr_time = 10;
367 int				segvn_disable_textrepl_update = 0;
368 
369 static void segvn_textrepl(struct seg *);
370 static void segvn_textunrepl(struct seg *, int);
371 static void segvn_inval_trcache(vnode_t *);
372 static void segvn_trasync_thread(void);
373 static void segvn_trupdate_wakeup(void *);
374 static void segvn_trupdate(void);
375 static void segvn_trupdate_seg(struct seg *, segvn_data_t *, svntr_t *,
376     ulong_t);
377 
378 /*
379  * Initialize segvn data structures
380  */
381 void
382 segvn_init(void)
383 {
384 	uint_t maxszc;
385 	uint_t szc;
386 	size_t pgsz;
387 
388 	segvn_cache = kmem_cache_create("segvn_cache",
389 		sizeof (struct segvn_data), 0,
390 		segvn_cache_constructor, segvn_cache_destructor, NULL,
391 		NULL, NULL, 0);
392 
393 	if (segvn_lpg_disable == 0) {
394 		szc = maxszc = page_num_pagesizes() - 1;
395 		if (szc == 0) {
396 			segvn_lpg_disable = 1;
397 		}
398 		if (page_get_pagesize(0) != PAGESIZE) {
399 			panic("segvn_init: bad szc 0");
400 			/*NOTREACHED*/
401 		}
402 		while (szc != 0) {
403 			pgsz = page_get_pagesize(szc);
404 			if (pgsz <= PAGESIZE || !IS_P2ALIGNED(pgsz, pgsz)) {
405 				panic("segvn_init: bad szc %d", szc);
406 				/*NOTREACHED*/
407 			}
408 			szc--;
409 		}
410 		if (segvn_maxpgszc == 0 || segvn_maxpgszc > maxszc)
411 			segvn_maxpgszc = maxszc;
412 	}
413 
414 	if (segvn_use_regions && !hat_supported(HAT_SHARED_REGIONS, NULL))
415 		segvn_use_regions = 0;
416 
417 	/*
418 	 * For now shared regions and text replication segvn support
419 	 * are mutually exclusive. This is acceptable because
420 	 * currently significant benefit from text replication was
421 	 * only observed on AMD64 NUMA platforms (due to relatively
422 	 * small L2$ size) and currently we don't support shared
423 	 * regions on x86.
424 	 */
425 	if (segvn_use_regions && !segvn_disable_textrepl) {
426 		segvn_disable_textrepl = 1;
427 	}
428 
429 #if defined(_LP64)
430 	if (lgrp_optimizations() && textrepl_size_thresh != (size_t)-1 &&
431 	    !segvn_disable_textrepl) {
432 		ulong_t i;
433 		size_t hsz = svntr_hashtab_sz * sizeof (svntr_bucket_t);
434 
435 		svntr_cache = kmem_cache_create("svntr_cache",
436 		    sizeof (svntr_t), 0, svntr_cache_constructor, NULL,
437 		    NULL, NULL, NULL, 0);
438 		svntr_hashtab = kmem_zalloc(hsz, KM_SLEEP);
439 		for (i = 0; i < svntr_hashtab_sz; i++) {
440 			mutex_init(&svntr_hashtab[i].tr_lock,  NULL,
441 			    MUTEX_DEFAULT, NULL);
442 		}
443 		segvn_textrepl_max_bytes = ptob(physmem) /
444 		    segvn_textrepl_max_bytes_factor;
445 		segvn_textrepl_stats = kmem_zalloc(NCPU *
446 		    sizeof (svntr_stats_t), KM_SLEEP);
447 		sema_init(&segvn_trasync_sem, 0, NULL, SEMA_DEFAULT, NULL);
448 		(void) thread_create(NULL, 0, segvn_trasync_thread,
449 		    NULL, 0, &p0, TS_RUN, minclsyspri);
450 	}
451 #endif
452 }
453 
454 #define	SEGVN_PAGEIO	((void *)0x1)
455 #define	SEGVN_NOPAGEIO	((void *)0x2)
456 
457 static void
458 segvn_setvnode_mpss(vnode_t *vp)
459 {
460 	int err;
461 
462 	ASSERT(vp->v_mpssdata == NULL ||
463 	    vp->v_mpssdata == SEGVN_PAGEIO ||
464 	    vp->v_mpssdata == SEGVN_NOPAGEIO);
465 
466 	if (vp->v_mpssdata == NULL) {
467 		if (vn_vmpss_usepageio(vp)) {
468 			err = VOP_PAGEIO(vp, (page_t *)NULL,
469 			    (u_offset_t)0, 0, 0, CRED());
470 		} else {
471 			err = ENOSYS;
472 		}
473 		/*
474 		 * set v_mpssdata just once per vnode life
475 		 * so that it never changes.
476 		 */
477 		mutex_enter(&vp->v_lock);
478 		if (vp->v_mpssdata == NULL) {
479 			if (err == EINVAL) {
480 				vp->v_mpssdata = SEGVN_PAGEIO;
481 			} else {
482 				vp->v_mpssdata = SEGVN_NOPAGEIO;
483 			}
484 		}
485 		mutex_exit(&vp->v_lock);
486 	}
487 }
488 
489 int
490 segvn_create(struct seg *seg, void *argsp)
491 {
492 	struct segvn_crargs *a = (struct segvn_crargs *)argsp;
493 	struct segvn_data *svd;
494 	size_t swresv = 0;
495 	struct cred *cred;
496 	struct anon_map *amp;
497 	int error = 0;
498 	size_t pgsz;
499 	lgrp_mem_policy_t mpolicy = LGRP_MEM_POLICY_DEFAULT;
500 	int use_rgn = 0;
501 	int trok = 0;
502 
503 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
504 
505 	if (a->type != MAP_PRIVATE && a->type != MAP_SHARED) {
506 		panic("segvn_create type");
507 		/*NOTREACHED*/
508 	}
509 
510 	/*
511 	 * Check arguments.  If a shared anon structure is given then
512 	 * it is illegal to also specify a vp.
513 	 */
514 	if (a->amp != NULL && a->vp != NULL) {
515 		panic("segvn_create anon_map");
516 		/*NOTREACHED*/
517 	}
518 
519 	if (a->type == MAP_PRIVATE && (a->flags & MAP_TEXT) &&
520 	    a->vp != NULL && a->prot == (PROT_USER | PROT_READ | PROT_EXEC) &&
521 	    segvn_use_regions) {
522 		use_rgn = 1;
523 	}
524 
525 	/* MAP_NORESERVE on a MAP_SHARED segment is meaningless. */
526 	if (a->type == MAP_SHARED)
527 		a->flags &= ~MAP_NORESERVE;
528 
529 	if (a->szc != 0) {
530 		if (segvn_lpg_disable != 0 || (a->szc == AS_MAP_NO_LPOOB) ||
531 		    (a->amp != NULL && a->type == MAP_PRIVATE) ||
532 		    (a->flags & MAP_NORESERVE) || seg->s_as == &kas) {
533 			a->szc = 0;
534 		} else {
535 			if (a->szc > segvn_maxpgszc)
536 				a->szc = segvn_maxpgszc;
537 			pgsz = page_get_pagesize(a->szc);
538 			if (!IS_P2ALIGNED(seg->s_base, pgsz) ||
539 			    !IS_P2ALIGNED(seg->s_size, pgsz)) {
540 				a->szc = 0;
541 			} else if (a->vp != NULL) {
542 				extern struct vnode kvp;
543 				if (IS_SWAPFSVP(a->vp) || VN_ISKAS(a->vp)) {
544 					/*
545 					 * paranoid check.
546 					 * hat_page_demote() is not supported
547 					 * on swapfs pages.
548 					 */
549 					a->szc = 0;
550 				} else if (map_addr_vacalign_check(seg->s_base,
551 				    a->offset & PAGEMASK)) {
552 					a->szc = 0;
553 				}
554 			} else if (a->amp != NULL) {
555 				pgcnt_t anum = btopr(a->offset);
556 				pgcnt_t pgcnt = page_get_pagecnt(a->szc);
557 				if (!IS_P2ALIGNED(anum, pgcnt)) {
558 					a->szc = 0;
559 				}
560 			}
561 		}
562 	}
563 
564 	/*
565 	 * If segment may need private pages, reserve them now.
566 	 */
567 	if (!(a->flags & MAP_NORESERVE) && ((a->vp == NULL && a->amp == NULL) ||
568 	    (a->type == MAP_PRIVATE && (a->prot & PROT_WRITE)))) {
569 		if (anon_resv(seg->s_size) == 0)
570 			return (EAGAIN);
571 		swresv = seg->s_size;
572 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
573 			seg, swresv, 1);
574 	}
575 
576 	/*
577 	 * Reserve any mapping structures that may be required.
578 	 *
579 	 * Don't do it for segments that may use regions. It's currently a
580 	 * noop in the hat implementations anyway.
581 	 */
582 	if (!use_rgn) {
583 		hat_map(seg->s_as->a_hat, seg->s_base, seg->s_size, HAT_MAP);
584 	}
585 
586 	if (a->cred) {
587 		cred = a->cred;
588 		crhold(cred);
589 	} else {
590 		crhold(cred = CRED());
591 	}
592 
593 	/* Inform the vnode of the new mapping */
594 	if (a->vp != NULL) {
595 		error = VOP_ADDMAP(a->vp, a->offset & PAGEMASK,
596 		    seg->s_as, seg->s_base, seg->s_size, a->prot,
597 		    a->maxprot, a->type, cred);
598 		if (error) {
599 			if (swresv != 0) {
600 				anon_unresv(swresv);
601 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
602 					"anon proc:%p %lu %u",
603 					seg, swresv, 0);
604 			}
605 			crfree(cred);
606 			if (!use_rgn) {
607 				hat_unload(seg->s_as->a_hat, seg->s_base,
608 				    seg->s_size, HAT_UNLOAD_UNMAP);
609 			}
610 			return (error);
611 		}
612 		/*
613 		 * svntr_hashtab will be NULL if we support shared regions.
614 		 */
615 		trok = ((a->flags & MAP_TEXT) &&
616 		    (seg->s_size > textrepl_size_thresh ||
617 			(a->flags & _MAP_TEXTREPL)) &&
618 		    lgrp_optimizations() && svntr_hashtab != NULL &&
619 		    a->type == MAP_PRIVATE && swresv == 0 &&
620 		    !(a->flags & MAP_NORESERVE) &&
621 		    seg->s_as != &kas && a->vp->v_type == VREG);
622 
623 		ASSERT(!trok || !use_rgn);
624 	}
625 
626 	/*
627 	 * If more than one segment in the address space, and they're adjacent
628 	 * virtually, try to concatenate them.  Don't concatenate if an
629 	 * explicit anon_map structure was supplied (e.g., SystemV shared
630 	 * memory) or if we'll use text replication for this segment.
631 	 */
632 	if (a->amp == NULL && !use_rgn && !trok) {
633 		struct seg *pseg, *nseg;
634 		struct segvn_data *psvd, *nsvd;
635 		lgrp_mem_policy_t ppolicy, npolicy;
636 		uint_t	lgrp_mem_policy_flags = 0;
637 		extern lgrp_mem_policy_t lgrp_mem_default_policy;
638 
639 		/*
640 		 * Memory policy flags (lgrp_mem_policy_flags) is valid when
641 		 * extending stack/heap segments.
642 		 */
643 		if ((a->vp == NULL) && (a->type == MAP_PRIVATE) &&
644 		    !(a->flags & MAP_NORESERVE) && (seg->s_as != &kas)) {
645 			lgrp_mem_policy_flags = a->lgrp_mem_policy_flags;
646 		} else {
647 			/*
648 			 * Get policy when not extending it from another segment
649 			 */
650 			mpolicy = lgrp_mem_policy_default(seg->s_size, a->type);
651 		}
652 
653 		/*
654 		 * First, try to concatenate the previous and new segments
655 		 */
656 		pseg = AS_SEGPREV(seg->s_as, seg);
657 		if (pseg != NULL &&
658 		    pseg->s_base + pseg->s_size == seg->s_base &&
659 		    pseg->s_ops == &segvn_ops) {
660 			/*
661 			 * Get memory allocation policy from previous segment.
662 			 * When extension is specified (e.g. for heap) apply
663 			 * this policy to the new segment regardless of the
664 			 * outcome of segment concatenation.  Extension occurs
665 			 * for non-default policy otherwise default policy is
666 			 * used and is based on extended segment size.
667 			 */
668 			psvd = (struct segvn_data *)pseg->s_data;
669 			ppolicy = psvd->policy_info.mem_policy;
670 			if (lgrp_mem_policy_flags ==
671 			    LGRP_MP_FLAG_EXTEND_UP) {
672 				if (ppolicy != lgrp_mem_default_policy) {
673 					mpolicy = ppolicy;
674 				} else {
675 					mpolicy = lgrp_mem_policy_default(
676 					    pseg->s_size + seg->s_size,
677 					    a->type);
678 				}
679 			}
680 
681 			if (mpolicy == ppolicy &&
682 			    (pseg->s_size + seg->s_size <=
683 			    segvn_comb_thrshld || psvd->amp == NULL) &&
684 			    segvn_extend_prev(pseg, seg, a, swresv) == 0) {
685 				/*
686 				 * success! now try to concatenate
687 				 * with following seg
688 				 */
689 				crfree(cred);
690 				nseg = AS_SEGNEXT(pseg->s_as, pseg);
691 				if (nseg != NULL &&
692 				    nseg != pseg &&
693 				    nseg->s_ops == &segvn_ops &&
694 				    pseg->s_base + pseg->s_size ==
695 				    nseg->s_base)
696 					(void) segvn_concat(pseg, nseg, 0);
697 				ASSERT(pseg->s_szc == 0 ||
698 				    (a->szc == pseg->s_szc &&
699 				    IS_P2ALIGNED(pseg->s_base, pgsz) &&
700 				    IS_P2ALIGNED(pseg->s_size, pgsz)));
701 				return (0);
702 			}
703 		}
704 
705 		/*
706 		 * Failed, so try to concatenate with following seg
707 		 */
708 		nseg = AS_SEGNEXT(seg->s_as, seg);
709 		if (nseg != NULL &&
710 		    seg->s_base + seg->s_size == nseg->s_base &&
711 		    nseg->s_ops == &segvn_ops) {
712 			/*
713 			 * Get memory allocation policy from next segment.
714 			 * When extension is specified (e.g. for stack) apply
715 			 * this policy to the new segment regardless of the
716 			 * outcome of segment concatenation.  Extension occurs
717 			 * for non-default policy otherwise default policy is
718 			 * used and is based on extended segment size.
719 			 */
720 			nsvd = (struct segvn_data *)nseg->s_data;
721 			npolicy = nsvd->policy_info.mem_policy;
722 			if (lgrp_mem_policy_flags ==
723 			    LGRP_MP_FLAG_EXTEND_DOWN) {
724 				if (npolicy != lgrp_mem_default_policy) {
725 					mpolicy = npolicy;
726 				} else {
727 					mpolicy = lgrp_mem_policy_default(
728 					    nseg->s_size + seg->s_size,
729 					    a->type);
730 				}
731 			}
732 
733 			if (mpolicy == npolicy &&
734 			    segvn_extend_next(seg, nseg, a, swresv) == 0) {
735 				crfree(cred);
736 				ASSERT(nseg->s_szc == 0 ||
737 				    (a->szc == nseg->s_szc &&
738 				    IS_P2ALIGNED(nseg->s_base, pgsz) &&
739 				    IS_P2ALIGNED(nseg->s_size, pgsz)));
740 				return (0);
741 			}
742 		}
743 	}
744 
745 	if (a->vp != NULL) {
746 		VN_HOLD(a->vp);
747 		if (a->type == MAP_SHARED)
748 			lgrp_shm_policy_init(NULL, a->vp);
749 	}
750 	svd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
751 
752 	seg->s_ops = &segvn_ops;
753 	seg->s_data = (void *)svd;
754 	seg->s_szc = a->szc;
755 
756 	svd->seg = seg;
757 	svd->vp = a->vp;
758 	/*
759 	 * Anonymous mappings have no backing file so the offset is meaningless.
760 	 */
761 	svd->offset = a->vp ? (a->offset & PAGEMASK) : 0;
762 	svd->prot = a->prot;
763 	svd->maxprot = a->maxprot;
764 	svd->pageprot = 0;
765 	svd->type = a->type;
766 	svd->vpage = NULL;
767 	svd->cred = cred;
768 	svd->advice = MADV_NORMAL;
769 	svd->pageadvice = 0;
770 	svd->flags = (ushort_t)a->flags;
771 	svd->softlockcnt = 0;
772 	svd->rcookie = HAT_INVALID_REGION_COOKIE;
773 
774 	if (a->szc != 0 && a->vp != NULL) {
775 		segvn_setvnode_mpss(a->vp);
776 	}
777 	if (svd->type == MAP_SHARED && svd->vp != NULL &&
778 	    (svd->vp->v_flag & VVMEXEC) && (svd->prot & PROT_WRITE)) {
779 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
780 		segvn_inval_trcache(svd->vp);
781 	}
782 
783 	amp = a->amp;
784 	if ((svd->amp = amp) == NULL) {
785 		svd->anon_index = 0;
786 		if (svd->type == MAP_SHARED) {
787 			svd->swresv = 0;
788 			/*
789 			 * Shared mappings to a vp need no other setup.
790 			 * If we have a shared mapping to an anon_map object
791 			 * which hasn't been allocated yet,  allocate the
792 			 * struct now so that it will be properly shared
793 			 * by remembering the swap reservation there.
794 			 */
795 			if (a->vp == NULL) {
796 				svd->amp = anonmap_alloc(seg->s_size, swresv,
797 				    ANON_SLEEP);
798 				svd->amp->a_szc = seg->s_szc;
799 			}
800 		} else {
801 			/*
802 			 * Private mapping (with or without a vp).
803 			 * Allocate anon_map when needed.
804 			 */
805 			svd->swresv = swresv;
806 		}
807 	} else {
808 		pgcnt_t anon_num;
809 
810 		/*
811 		 * Mapping to an existing anon_map structure without a vp.
812 		 * For now we will insure that the segment size isn't larger
813 		 * than the size - offset gives us.  Later on we may wish to
814 		 * have the anon array dynamically allocated itself so that
815 		 * we don't always have to allocate all the anon pointer slots.
816 		 * This of course involves adding extra code to check that we
817 		 * aren't trying to use an anon pointer slot beyond the end
818 		 * of the currently allocated anon array.
819 		 */
820 		if ((amp->size - a->offset) < seg->s_size) {
821 			panic("segvn_create anon_map size");
822 			/*NOTREACHED*/
823 		}
824 
825 		anon_num = btopr(a->offset);
826 
827 		if (a->type == MAP_SHARED) {
828 			/*
829 			 * SHARED mapping to a given anon_map.
830 			 */
831 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
832 			amp->refcnt++;
833 			if (a->szc > amp->a_szc) {
834 				amp->a_szc = a->szc;
835 			}
836 			ANON_LOCK_EXIT(&amp->a_rwlock);
837 			svd->anon_index = anon_num;
838 			svd->swresv = 0;
839 		} else {
840 			/*
841 			 * PRIVATE mapping to a given anon_map.
842 			 * Make sure that all the needed anon
843 			 * structures are created (so that we will
844 			 * share the underlying pages if nothing
845 			 * is written by this mapping) and then
846 			 * duplicate the anon array as is done
847 			 * when a privately mapped segment is dup'ed.
848 			 */
849 			struct anon *ap;
850 			caddr_t addr;
851 			caddr_t eaddr;
852 			ulong_t	anon_idx;
853 			int hat_flag = HAT_LOAD;
854 
855 			if (svd->flags & MAP_TEXT) {
856 				hat_flag |= HAT_LOAD_TEXT;
857 			}
858 
859 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
860 			svd->amp->a_szc = seg->s_szc;
861 			svd->anon_index = 0;
862 			svd->swresv = swresv;
863 
864 			/*
865 			 * Prevent 2 threads from allocating anon
866 			 * slots simultaneously.
867 			 */
868 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
869 			eaddr = seg->s_base + seg->s_size;
870 
871 			for (anon_idx = anon_num, addr = seg->s_base;
872 			    addr < eaddr; addr += PAGESIZE, anon_idx++) {
873 				page_t *pp;
874 
875 				if ((ap = anon_get_ptr(amp->ahp,
876 				    anon_idx)) != NULL)
877 					continue;
878 
879 				/*
880 				 * Allocate the anon struct now.
881 				 * Might as well load up translation
882 				 * to the page while we're at it...
883 				 */
884 				pp = anon_zero(seg, addr, &ap, cred);
885 				if (ap == NULL || pp == NULL) {
886 					panic("segvn_create anon_zero");
887 					/*NOTREACHED*/
888 				}
889 
890 				/*
891 				 * Re-acquire the anon_map lock and
892 				 * initialize the anon array entry.
893 				 */
894 				ASSERT(anon_get_ptr(amp->ahp,
895 				    anon_idx) == NULL);
896 				(void) anon_set_ptr(amp->ahp, anon_idx, ap,
897 				    ANON_SLEEP);
898 
899 				ASSERT(seg->s_szc == 0);
900 				ASSERT(!IS_VMODSORT(pp->p_vnode));
901 
902 				ASSERT(use_rgn == 0);
903 				hat_memload(seg->s_as->a_hat, addr, pp,
904 					svd->prot & ~PROT_WRITE, hat_flag);
905 
906 				page_unlock(pp);
907 			}
908 			ASSERT(seg->s_szc == 0);
909 			anon_dup(amp->ahp, anon_num, svd->amp->ahp,
910 			    0, seg->s_size);
911 			ANON_LOCK_EXIT(&amp->a_rwlock);
912 		}
913 	}
914 
915 	/*
916 	 * Set default memory allocation policy for segment
917 	 *
918 	 * Always set policy for private memory at least for initialization
919 	 * even if this is a shared memory segment
920 	 */
921 	(void) lgrp_privm_policy_set(mpolicy, &svd->policy_info, seg->s_size);
922 
923 	if (svd->type == MAP_SHARED)
924 		(void) lgrp_shm_policy_set(mpolicy, svd->amp, svd->anon_index,
925 		    svd->vp, svd->offset, seg->s_size);
926 
927 	if (use_rgn) {
928 		ASSERT(!trok);
929 		ASSERT(svd->amp == NULL);
930 		svd->rcookie = hat_join_region(seg->s_as->a_hat, seg->s_base,
931 		    seg->s_size, (void *)svd->vp, svd->offset, svd->prot,
932 		    (uchar_t)seg->s_szc, segvn_hat_rgn_unload_callback,
933 		    HAT_REGION_TEXT);
934 	}
935 
936 	ASSERT(!trok || !(svd->prot & PROT_WRITE));
937 	svd->tr_state = trok ? SEGVN_TR_INIT : SEGVN_TR_OFF;
938 
939 	return (0);
940 }
941 
942 /*
943  * Concatenate two existing segments, if possible.
944  * Return 0 on success, -1 if two segments are not compatible
945  * or -2 on memory allocation failure.
946  * If amp_cat == 1 then try and concat segments with anon maps
947  */
948 static int
949 segvn_concat(struct seg *seg1, struct seg *seg2, int amp_cat)
950 {
951 	struct segvn_data *svd1 = seg1->s_data;
952 	struct segvn_data *svd2 = seg2->s_data;
953 	struct anon_map *amp1 = svd1->amp;
954 	struct anon_map *amp2 = svd2->amp;
955 	struct vpage *vpage1 = svd1->vpage;
956 	struct vpage *vpage2 = svd2->vpage, *nvpage = NULL;
957 	size_t size, nvpsize;
958 	pgcnt_t npages1, npages2;
959 
960 	ASSERT(seg1->s_as && seg2->s_as && seg1->s_as == seg2->s_as);
961 	ASSERT(AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
962 	ASSERT(seg1->s_ops == seg2->s_ops);
963 
964 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie) ||
965 	    HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
966 		return (-1);
967 	}
968 
969 	/* both segments exist, try to merge them */
970 #define	incompat(x)	(svd1->x != svd2->x)
971 	if (incompat(vp) || incompat(maxprot) ||
972 	    (!svd1->pageadvice && !svd2->pageadvice && incompat(advice)) ||
973 	    (!svd1->pageprot && !svd2->pageprot && incompat(prot)) ||
974 	    incompat(type) || incompat(cred) || incompat(flags) ||
975 	    seg1->s_szc != seg2->s_szc || incompat(policy_info.mem_policy) ||
976 	    (svd2->softlockcnt > 0))
977 		return (-1);
978 #undef incompat
979 
980 	/*
981 	 * vp == NULL implies zfod, offset doesn't matter
982 	 */
983 	if (svd1->vp != NULL &&
984 	    svd1->offset + seg1->s_size != svd2->offset) {
985 		return (-1);
986 	}
987 
988 	/*
989 	 * Don't concatenate if either segment uses text replication.
990 	 */
991 	if (svd1->tr_state != SEGVN_TR_OFF || svd2->tr_state != SEGVN_TR_OFF) {
992 		return (-1);
993 	}
994 
995 	/*
996 	 * Fail early if we're not supposed to concatenate
997 	 * segments with non NULL amp.
998 	 */
999 	if (amp_cat == 0 && (amp1 != NULL || amp2 != NULL)) {
1000 		return (-1);
1001 	}
1002 
1003 	if (svd1->vp == NULL && svd1->type == MAP_SHARED) {
1004 		if (amp1 != amp2) {
1005 			return (-1);
1006 		}
1007 		if (amp1 != NULL && svd1->anon_index + btop(seg1->s_size) !=
1008 		    svd2->anon_index) {
1009 			return (-1);
1010 		}
1011 		ASSERT(amp1 == NULL || amp1->refcnt >= 2);
1012 	}
1013 
1014 	/*
1015 	 * If either seg has vpages, create a new merged vpage array.
1016 	 */
1017 	if (vpage1 != NULL || vpage2 != NULL) {
1018 		struct vpage *vp;
1019 
1020 		npages1 = seg_pages(seg1);
1021 		npages2 = seg_pages(seg2);
1022 		nvpsize = vpgtob(npages1 + npages2);
1023 
1024 		if ((nvpage = kmem_zalloc(nvpsize, KM_NOSLEEP)) == NULL) {
1025 			return (-2);
1026 		}
1027 
1028 		if (vpage1 != NULL) {
1029 			bcopy(vpage1, nvpage, vpgtob(npages1));
1030 		} else {
1031 			for (vp = nvpage; vp < nvpage + npages1; vp++) {
1032 				VPP_SETPROT(vp, svd1->prot);
1033 				VPP_SETADVICE(vp, svd1->advice);
1034 			}
1035 		}
1036 
1037 		if (vpage2 != NULL) {
1038 			bcopy(vpage2, nvpage + npages1, vpgtob(npages2));
1039 		} else {
1040 			for (vp = nvpage + npages1;
1041 			    vp < nvpage + npages1 + npages2; vp++) {
1042 				VPP_SETPROT(vp, svd2->prot);
1043 				VPP_SETADVICE(vp, svd2->advice);
1044 			}
1045 		}
1046 	}
1047 
1048 	/*
1049 	 * If either segment has private pages, create a new merged anon
1050 	 * array. If mergeing shared anon segments just decrement anon map's
1051 	 * refcnt.
1052 	 */
1053 	if (amp1 != NULL && svd1->type == MAP_SHARED) {
1054 		ASSERT(amp1 == amp2 && svd1->vp == NULL);
1055 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1056 		ASSERT(amp1->refcnt >= 2);
1057 		amp1->refcnt--;
1058 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1059 		svd2->amp = NULL;
1060 	} else if (amp1 != NULL || amp2 != NULL) {
1061 		struct anon_hdr *nahp;
1062 		struct anon_map *namp = NULL;
1063 		size_t asize;
1064 
1065 		ASSERT(svd1->type == MAP_PRIVATE);
1066 
1067 		asize = seg1->s_size + seg2->s_size;
1068 		if ((nahp = anon_create(btop(asize), ANON_NOSLEEP)) == NULL) {
1069 			if (nvpage != NULL) {
1070 				kmem_free(nvpage, nvpsize);
1071 			}
1072 			return (-2);
1073 		}
1074 		if (amp1 != NULL) {
1075 			/*
1076 			 * XXX anon rwlock is not really needed because
1077 			 * this is a private segment and we are writers.
1078 			 */
1079 			ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1080 			ASSERT(amp1->refcnt == 1);
1081 			if (anon_copy_ptr(amp1->ahp, svd1->anon_index,
1082 			    nahp, 0, btop(seg1->s_size), ANON_NOSLEEP)) {
1083 				anon_release(nahp, btop(asize));
1084 				ANON_LOCK_EXIT(&amp1->a_rwlock);
1085 				if (nvpage != NULL) {
1086 					kmem_free(nvpage, nvpsize);
1087 				}
1088 				return (-2);
1089 			}
1090 		}
1091 		if (amp2 != NULL) {
1092 			ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1093 			ASSERT(amp2->refcnt == 1);
1094 			if (anon_copy_ptr(amp2->ahp, svd2->anon_index,
1095 			    nahp, btop(seg1->s_size), btop(seg2->s_size),
1096 			    ANON_NOSLEEP)) {
1097 				anon_release(nahp, btop(asize));
1098 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1099 				if (amp1 != NULL) {
1100 					ANON_LOCK_EXIT(&amp1->a_rwlock);
1101 				}
1102 				if (nvpage != NULL) {
1103 					kmem_free(nvpage, nvpsize);
1104 				}
1105 				return (-2);
1106 			}
1107 		}
1108 		if (amp1 != NULL) {
1109 			namp = amp1;
1110 			anon_release(amp1->ahp, btop(amp1->size));
1111 		}
1112 		if (amp2 != NULL) {
1113 			if (namp == NULL) {
1114 				ASSERT(amp1 == NULL);
1115 				namp = amp2;
1116 				anon_release(amp2->ahp, btop(amp2->size));
1117 			} else {
1118 				amp2->refcnt--;
1119 				ANON_LOCK_EXIT(&amp2->a_rwlock);
1120 				anonmap_free(amp2);
1121 			}
1122 			svd2->amp = NULL; /* needed for seg_free */
1123 		}
1124 		namp->ahp = nahp;
1125 		namp->size = asize;
1126 		svd1->amp = namp;
1127 		svd1->anon_index = 0;
1128 		ANON_LOCK_EXIT(&namp->a_rwlock);
1129 	}
1130 	/*
1131 	 * Now free the old vpage structures.
1132 	 */
1133 	if (nvpage != NULL) {
1134 		if (vpage1 != NULL) {
1135 			kmem_free(vpage1, vpgtob(npages1));
1136 		}
1137 		if (vpage2 != NULL) {
1138 			svd2->vpage = NULL;
1139 			kmem_free(vpage2, vpgtob(npages2));
1140 		}
1141 		if (svd2->pageprot) {
1142 			svd1->pageprot = 1;
1143 		}
1144 		if (svd2->pageadvice) {
1145 			svd1->pageadvice = 1;
1146 		}
1147 		svd1->vpage = nvpage;
1148 	}
1149 
1150 	/* all looks ok, merge segments */
1151 	svd1->swresv += svd2->swresv;
1152 	svd2->swresv = 0;  /* so seg_free doesn't release swap space */
1153 	size = seg2->s_size;
1154 	seg_free(seg2);
1155 	seg1->s_size += size;
1156 	return (0);
1157 }
1158 
1159 /*
1160  * Extend the previous segment (seg1) to include the
1161  * new segment (seg2 + a), if possible.
1162  * Return 0 on success.
1163  */
1164 static int
1165 segvn_extend_prev(seg1, seg2, a, swresv)
1166 	struct seg *seg1, *seg2;
1167 	struct segvn_crargs *a;
1168 	size_t swresv;
1169 {
1170 	struct segvn_data *svd1 = (struct segvn_data *)seg1->s_data;
1171 	size_t size;
1172 	struct anon_map *amp1;
1173 	struct vpage *new_vpage;
1174 
1175 	/*
1176 	 * We don't need any segment level locks for "segvn" data
1177 	 * since the address space is "write" locked.
1178 	 */
1179 	ASSERT(seg1->s_as && AS_WRITE_HELD(seg1->s_as, &seg1->s_as->a_lock));
1180 
1181 	if (HAT_IS_REGION_COOKIE_VALID(svd1->rcookie)) {
1182 		return (-1);
1183 	}
1184 
1185 	/* second segment is new, try to extend first */
1186 	/* XXX - should also check cred */
1187 	if (svd1->vp != a->vp || svd1->maxprot != a->maxprot ||
1188 	    (!svd1->pageprot && (svd1->prot != a->prot)) ||
1189 	    svd1->type != a->type || svd1->flags != a->flags ||
1190 	    seg1->s_szc != a->szc)
1191 		return (-1);
1192 
1193 	/* vp == NULL implies zfod, offset doesn't matter */
1194 	if (svd1->vp != NULL &&
1195 	    svd1->offset + seg1->s_size != (a->offset & PAGEMASK))
1196 		return (-1);
1197 
1198 	if (svd1->tr_state != SEGVN_TR_OFF) {
1199 		return (-1);
1200 	}
1201 
1202 	amp1 = svd1->amp;
1203 	if (amp1) {
1204 		pgcnt_t newpgs;
1205 
1206 		/*
1207 		 * Segment has private pages, can data structures
1208 		 * be expanded?
1209 		 *
1210 		 * Acquire the anon_map lock to prevent it from changing,
1211 		 * if it is shared.  This ensures that the anon_map
1212 		 * will not change while a thread which has a read/write
1213 		 * lock on an address space references it.
1214 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1215 		 * is 1.
1216 		 *
1217 		 * Can't grow a MAP_SHARED segment with an anonmap because
1218 		 * there may be existing anon slots where we want to extend
1219 		 * the segment and we wouldn't know what to do with them
1220 		 * (e.g., for tmpfs right thing is to just leave them there,
1221 		 * for /dev/zero they should be cleared out).
1222 		 */
1223 		if (svd1->type == MAP_SHARED)
1224 			return (-1);
1225 
1226 		ANON_LOCK_ENTER(&amp1->a_rwlock, RW_WRITER);
1227 		if (amp1->refcnt > 1) {
1228 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1229 			return (-1);
1230 		}
1231 		newpgs = anon_grow(amp1->ahp, &svd1->anon_index,
1232 		    btop(seg1->s_size), btop(seg2->s_size), ANON_NOSLEEP);
1233 
1234 		if (newpgs == 0) {
1235 			ANON_LOCK_EXIT(&amp1->a_rwlock);
1236 			return (-1);
1237 		}
1238 		amp1->size = ptob(newpgs);
1239 		ANON_LOCK_EXIT(&amp1->a_rwlock);
1240 	}
1241 	if (svd1->vpage != NULL) {
1242 		struct vpage *vp, *evp;
1243 		new_vpage =
1244 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1245 			KM_NOSLEEP);
1246 		if (new_vpage == NULL)
1247 			return (-1);
1248 		bcopy(svd1->vpage, new_vpage, vpgtob(seg_pages(seg1)));
1249 		kmem_free(svd1->vpage, vpgtob(seg_pages(seg1)));
1250 		svd1->vpage = new_vpage;
1251 
1252 		vp = new_vpage + seg_pages(seg1);
1253 		evp = vp + seg_pages(seg2);
1254 		for (; vp < evp; vp++)
1255 			VPP_SETPROT(vp, a->prot);
1256 	}
1257 	size = seg2->s_size;
1258 	seg_free(seg2);
1259 	seg1->s_size += size;
1260 	svd1->swresv += swresv;
1261 	if (svd1->pageprot && (a->prot & PROT_WRITE) &&
1262 	    svd1->type == MAP_SHARED && svd1->vp != NULL &&
1263 	    (svd1->vp->v_flag & VVMEXEC)) {
1264 		ASSERT(vn_is_mapped(svd1->vp, V_WRITE));
1265 		segvn_inval_trcache(svd1->vp);
1266 	}
1267 	return (0);
1268 }
1269 
1270 /*
1271  * Extend the next segment (seg2) to include the
1272  * new segment (seg1 + a), if possible.
1273  * Return 0 on success.
1274  */
1275 static int
1276 segvn_extend_next(
1277 	struct seg *seg1,
1278 	struct seg *seg2,
1279 	struct segvn_crargs *a,
1280 	size_t swresv)
1281 {
1282 	struct segvn_data *svd2 = (struct segvn_data *)seg2->s_data;
1283 	size_t size;
1284 	struct anon_map *amp2;
1285 	struct vpage *new_vpage;
1286 
1287 	/*
1288 	 * We don't need any segment level locks for "segvn" data
1289 	 * since the address space is "write" locked.
1290 	 */
1291 	ASSERT(seg2->s_as && AS_WRITE_HELD(seg2->s_as, &seg2->s_as->a_lock));
1292 
1293 	if (HAT_IS_REGION_COOKIE_VALID(svd2->rcookie)) {
1294 		return (-1);
1295 	}
1296 
1297 	/* first segment is new, try to extend second */
1298 	/* XXX - should also check cred */
1299 	if (svd2->vp != a->vp || svd2->maxprot != a->maxprot ||
1300 	    (!svd2->pageprot && (svd2->prot != a->prot)) ||
1301 	    svd2->type != a->type || svd2->flags != a->flags ||
1302 	    seg2->s_szc != a->szc)
1303 		return (-1);
1304 	/* vp == NULL implies zfod, offset doesn't matter */
1305 	if (svd2->vp != NULL &&
1306 	    (a->offset & PAGEMASK) + seg1->s_size != svd2->offset)
1307 		return (-1);
1308 
1309 	if (svd2->tr_state != SEGVN_TR_OFF) {
1310 		return (-1);
1311 	}
1312 
1313 	amp2 = svd2->amp;
1314 	if (amp2) {
1315 		pgcnt_t newpgs;
1316 
1317 		/*
1318 		 * Segment has private pages, can data structures
1319 		 * be expanded?
1320 		 *
1321 		 * Acquire the anon_map lock to prevent it from changing,
1322 		 * if it is shared.  This ensures that the anon_map
1323 		 * will not change while a thread which has a read/write
1324 		 * lock on an address space references it.
1325 		 *
1326 		 * XXX - Don't need the anon_map lock at all if "refcnt"
1327 		 * is 1.
1328 		 */
1329 		if (svd2->type == MAP_SHARED)
1330 			return (-1);
1331 
1332 		ANON_LOCK_ENTER(&amp2->a_rwlock, RW_WRITER);
1333 		if (amp2->refcnt > 1) {
1334 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1335 			return (-1);
1336 		}
1337 		newpgs = anon_grow(amp2->ahp, &svd2->anon_index,
1338 		    btop(seg2->s_size), btop(seg1->s_size),
1339 		    ANON_NOSLEEP | ANON_GROWDOWN);
1340 
1341 		if (newpgs == 0) {
1342 			ANON_LOCK_EXIT(&amp2->a_rwlock);
1343 			return (-1);
1344 		}
1345 		amp2->size = ptob(newpgs);
1346 		ANON_LOCK_EXIT(&amp2->a_rwlock);
1347 	}
1348 	if (svd2->vpage != NULL) {
1349 		struct vpage *vp, *evp;
1350 		new_vpage =
1351 		    kmem_zalloc(vpgtob(seg_pages(seg1) + seg_pages(seg2)),
1352 			KM_NOSLEEP);
1353 		if (new_vpage == NULL) {
1354 			/* Not merging segments so adjust anon_index back */
1355 			if (amp2)
1356 				svd2->anon_index += seg_pages(seg1);
1357 			return (-1);
1358 		}
1359 		bcopy(svd2->vpage, new_vpage + seg_pages(seg1),
1360 		    vpgtob(seg_pages(seg2)));
1361 		kmem_free(svd2->vpage, vpgtob(seg_pages(seg2)));
1362 		svd2->vpage = new_vpage;
1363 
1364 		vp = new_vpage;
1365 		evp = vp + seg_pages(seg1);
1366 		for (; vp < evp; vp++)
1367 			VPP_SETPROT(vp, a->prot);
1368 	}
1369 	size = seg1->s_size;
1370 	seg_free(seg1);
1371 	seg2->s_size += size;
1372 	seg2->s_base -= size;
1373 	svd2->offset -= size;
1374 	svd2->swresv += swresv;
1375 	if (svd2->pageprot && (a->prot & PROT_WRITE) &&
1376 	    svd2->type == MAP_SHARED && svd2->vp != NULL &&
1377 	    (svd2->vp->v_flag & VVMEXEC)) {
1378 		ASSERT(vn_is_mapped(svd2->vp, V_WRITE));
1379 		segvn_inval_trcache(svd2->vp);
1380 	}
1381 	return (0);
1382 }
1383 
1384 static int
1385 segvn_dup(struct seg *seg, struct seg *newseg)
1386 {
1387 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1388 	struct segvn_data *newsvd;
1389 	pgcnt_t npages = seg_pages(seg);
1390 	int error = 0;
1391 	uint_t prot;
1392 	size_t len;
1393 	struct anon_map *amp;
1394 
1395 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1396 
1397 	/*
1398 	 * If segment has anon reserved, reserve more for the new seg.
1399 	 * For a MAP_NORESERVE segment swresv will be a count of all the
1400 	 * allocated anon slots; thus we reserve for the child as many slots
1401 	 * as the parent has allocated. This semantic prevents the child or
1402 	 * parent from dieing during a copy-on-write fault caused by trying
1403 	 * to write a shared pre-existing anon page.
1404 	 */
1405 	if ((len = svd->swresv) != 0) {
1406 		if (anon_resv(svd->swresv) == 0)
1407 			return (ENOMEM);
1408 
1409 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
1410 			seg, len, 0);
1411 	}
1412 
1413 	newsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
1414 
1415 	newseg->s_ops = &segvn_ops;
1416 	newseg->s_data = (void *)newsvd;
1417 	newseg->s_szc = seg->s_szc;
1418 
1419 	newsvd->seg = newseg;
1420 	if ((newsvd->vp = svd->vp) != NULL) {
1421 		VN_HOLD(svd->vp);
1422 		if (svd->type == MAP_SHARED)
1423 			lgrp_shm_policy_init(NULL, svd->vp);
1424 	}
1425 	newsvd->offset = svd->offset;
1426 	newsvd->prot = svd->prot;
1427 	newsvd->maxprot = svd->maxprot;
1428 	newsvd->pageprot = svd->pageprot;
1429 	newsvd->type = svd->type;
1430 	newsvd->cred = svd->cred;
1431 	crhold(newsvd->cred);
1432 	newsvd->advice = svd->advice;
1433 	newsvd->pageadvice = svd->pageadvice;
1434 	newsvd->swresv = svd->swresv;
1435 	newsvd->flags = svd->flags;
1436 	newsvd->softlockcnt = 0;
1437 	newsvd->policy_info = svd->policy_info;
1438 	newsvd->rcookie = HAT_INVALID_REGION_COOKIE;
1439 
1440 	if ((amp = svd->amp) == NULL || svd->tr_state == SEGVN_TR_ON) {
1441 		/*
1442 		 * Not attaching to a shared anon object.
1443 		 */
1444 		ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie) ||
1445 		    svd->tr_state == SEGVN_TR_OFF);
1446 		if (svd->tr_state == SEGVN_TR_ON) {
1447 			ASSERT(newsvd->vp != NULL && amp != NULL);
1448 			newsvd->tr_state = SEGVN_TR_INIT;
1449 		} else {
1450 			newsvd->tr_state = svd->tr_state;
1451 		}
1452 		newsvd->amp = NULL;
1453 		newsvd->anon_index = 0;
1454 	} else {
1455 		/* regions for now are only used on pure vnode segments */
1456 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
1457 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1458 		newsvd->tr_state = SEGVN_TR_OFF;
1459 		if (svd->type == MAP_SHARED) {
1460 			newsvd->amp = amp;
1461 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1462 			amp->refcnt++;
1463 			ANON_LOCK_EXIT(&amp->a_rwlock);
1464 			newsvd->anon_index = svd->anon_index;
1465 		} else {
1466 			int reclaim = 1;
1467 
1468 			/*
1469 			 * Allocate and initialize new anon_map structure.
1470 			 */
1471 			newsvd->amp = anonmap_alloc(newseg->s_size, 0,
1472 			    ANON_SLEEP);
1473 			newsvd->amp->a_szc = newseg->s_szc;
1474 			newsvd->anon_index = 0;
1475 
1476 			/*
1477 			 * We don't have to acquire the anon_map lock
1478 			 * for the new segment (since it belongs to an
1479 			 * address space that is still not associated
1480 			 * with any process), or the segment in the old
1481 			 * address space (since all threads in it
1482 			 * are stopped while duplicating the address space).
1483 			 */
1484 
1485 			/*
1486 			 * The goal of the following code is to make sure that
1487 			 * softlocked pages do not end up as copy on write
1488 			 * pages.  This would cause problems where one
1489 			 * thread writes to a page that is COW and a different
1490 			 * thread in the same process has softlocked it.  The
1491 			 * softlock lock would move away from this process
1492 			 * because the write would cause this process to get
1493 			 * a copy (without the softlock).
1494 			 *
1495 			 * The strategy here is to just break the
1496 			 * sharing on pages that could possibly be
1497 			 * softlocked.
1498 			 */
1499 retry:
1500 			if (svd->softlockcnt) {
1501 				struct anon *ap, *newap;
1502 				size_t i;
1503 				uint_t vpprot;
1504 				page_t *anon_pl[1+1], *pp;
1505 				caddr_t addr;
1506 				ulong_t old_idx = svd->anon_index;
1507 				ulong_t new_idx = 0;
1508 
1509 				/*
1510 				 * The softlock count might be non zero
1511 				 * because some pages are still stuck in the
1512 				 * cache for lazy reclaim. Flush the cache
1513 				 * now. This should drop the count to zero.
1514 				 * [or there is really I/O going on to these
1515 				 * pages]. Note, we have the writers lock so
1516 				 * nothing gets inserted during the flush.
1517 				 */
1518 				if (reclaim == 1) {
1519 					segvn_purge(seg);
1520 					reclaim = 0;
1521 					goto retry;
1522 				}
1523 				i = btopr(seg->s_size);
1524 				addr = seg->s_base;
1525 				/*
1526 				 * XXX break cow sharing using PAGESIZE
1527 				 * pages. They will be relocated into larger
1528 				 * pages at fault time.
1529 				 */
1530 				while (i-- > 0) {
1531 					if (ap = anon_get_ptr(amp->ahp,
1532 					    old_idx)) {
1533 						error = anon_getpage(&ap,
1534 						    &vpprot, anon_pl, PAGESIZE,
1535 						    seg, addr, S_READ,
1536 						    svd->cred);
1537 						if (error) {
1538 							newsvd->vpage = NULL;
1539 							goto out;
1540 						}
1541 						/*
1542 						 * prot need not be computed
1543 						 * below 'cause anon_private is
1544 						 * going to ignore it anyway
1545 						 * as child doesn't inherit
1546 						 * pagelock from parent.
1547 						 */
1548 						prot = svd->pageprot ?
1549 						    VPP_PROT(
1550 						    &svd->vpage[
1551 						    seg_page(seg, addr)])
1552 						    : svd->prot;
1553 						pp = anon_private(&newap,
1554 						    newseg, addr, prot,
1555 						    anon_pl[0],	0,
1556 						    newsvd->cred);
1557 						if (pp == NULL) {
1558 							/* no mem abort */
1559 							newsvd->vpage = NULL;
1560 							error = ENOMEM;
1561 							goto out;
1562 						}
1563 						(void) anon_set_ptr(
1564 						    newsvd->amp->ahp, new_idx,
1565 						    newap, ANON_SLEEP);
1566 						page_unlock(pp);
1567 					}
1568 					addr += PAGESIZE;
1569 					old_idx++;
1570 					new_idx++;
1571 				}
1572 			} else {	/* common case */
1573 				if (seg->s_szc != 0) {
1574 					/*
1575 					 * If at least one of anon slots of a
1576 					 * large page exists then make sure
1577 					 * all anon slots of a large page
1578 					 * exist to avoid partial cow sharing
1579 					 * of a large page in the future.
1580 					 */
1581 					anon_dup_fill_holes(amp->ahp,
1582 					    svd->anon_index, newsvd->amp->ahp,
1583 					    0, seg->s_size, seg->s_szc,
1584 					    svd->vp != NULL);
1585 				} else {
1586 					anon_dup(amp->ahp, svd->anon_index,
1587 					    newsvd->amp->ahp, 0, seg->s_size);
1588 				}
1589 
1590 				hat_clrattr(seg->s_as->a_hat, seg->s_base,
1591 				    seg->s_size, PROT_WRITE);
1592 			}
1593 		}
1594 	}
1595 	/*
1596 	 * If necessary, create a vpage structure for the new segment.
1597 	 * Do not copy any page lock indications.
1598 	 */
1599 	if (svd->vpage != NULL) {
1600 		uint_t i;
1601 		struct vpage *ovp = svd->vpage;
1602 		struct vpage *nvp;
1603 
1604 		nvp = newsvd->vpage =
1605 		    kmem_alloc(vpgtob(npages), KM_SLEEP);
1606 		for (i = 0; i < npages; i++) {
1607 			*nvp = *ovp++;
1608 			VPP_CLRPPLOCK(nvp++);
1609 		}
1610 	} else
1611 		newsvd->vpage = NULL;
1612 
1613 	/* Inform the vnode of the new mapping */
1614 	if (newsvd->vp != NULL) {
1615 		error = VOP_ADDMAP(newsvd->vp, (offset_t)newsvd->offset,
1616 		    newseg->s_as, newseg->s_base, newseg->s_size, newsvd->prot,
1617 		    newsvd->maxprot, newsvd->type, newsvd->cred);
1618 	}
1619 out:
1620 	if (error == 0 && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1621 		ASSERT(newsvd->amp == NULL);
1622 		ASSERT(newsvd->tr_state == SEGVN_TR_OFF);
1623 		newsvd->rcookie = svd->rcookie;
1624 		hat_dup_region(newseg->s_as->a_hat, newsvd->rcookie);
1625 	}
1626 	return (error);
1627 }
1628 
1629 
1630 /*
1631  * callback function used by segvn_unmap to invoke free_vp_pages() for only
1632  * those pages actually processed by the HAT
1633  */
1634 extern int free_pages;
1635 
1636 static void
1637 segvn_hat_rgn_unload_callback(caddr_t saddr, caddr_t eaddr, caddr_t r_saddr,
1638     size_t r_size, void *r_obj, u_offset_t r_objoff)
1639 {
1640 	u_offset_t off;
1641 	size_t len;
1642 	vnode_t *vp = (vnode_t *)r_obj;
1643 
1644 	ASSERT(eaddr > saddr);
1645 	ASSERT(saddr >= r_saddr);
1646 	ASSERT(saddr < r_saddr + r_size);
1647 	ASSERT(eaddr > r_saddr);
1648 	ASSERT(eaddr <= r_saddr + r_size);
1649 	ASSERT(vp != NULL);
1650 
1651 	if (!free_pages) {
1652 		return;
1653 	}
1654 
1655 	len = eaddr - saddr;
1656 	off = (saddr - r_saddr) + r_objoff;
1657 	free_vp_pages(vp, off, len);
1658 }
1659 
1660 static void
1661 segvn_hat_unload_callback(hat_callback_t *cb)
1662 {
1663 	struct seg		*seg = cb->hcb_data;
1664 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
1665 	size_t			len;
1666 	u_offset_t		off;
1667 
1668 	ASSERT(svd->vp != NULL);
1669 	ASSERT(cb->hcb_end_addr > cb->hcb_start_addr);
1670 	ASSERT(cb->hcb_start_addr >= seg->s_base);
1671 
1672 	len = cb->hcb_end_addr - cb->hcb_start_addr;
1673 	off = cb->hcb_start_addr - seg->s_base;
1674 	free_vp_pages(svd->vp, svd->offset + off, len);
1675 }
1676 
1677 static int
1678 segvn_unmap(struct seg *seg, caddr_t addr, size_t len)
1679 {
1680 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
1681 	struct segvn_data *nsvd;
1682 	struct seg *nseg;
1683 	struct anon_map *amp;
1684 	pgcnt_t	opages;		/* old segment size in pages */
1685 	pgcnt_t	npages;		/* new segment size in pages */
1686 	pgcnt_t	dpages;		/* pages being deleted (unmapped) */
1687 	hat_callback_t callback;	/* used for free_vp_pages() */
1688 	hat_callback_t *cbp = NULL;
1689 	caddr_t nbase;
1690 	size_t nsize;
1691 	size_t oswresv;
1692 	int reclaim = 1;
1693 
1694 	/*
1695 	 * We don't need any segment level locks for "segvn" data
1696 	 * since the address space is "write" locked.
1697 	 */
1698 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
1699 
1700 	/*
1701 	 * Fail the unmap if pages are SOFTLOCKed through this mapping.
1702 	 * softlockcnt is protected from change by the as write lock.
1703 	 */
1704 retry:
1705 	if (svd->softlockcnt > 0) {
1706 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1707 		/*
1708 		 * since we do have the writers lock nobody can fill
1709 		 * the cache during the purge. The flush either succeeds
1710 		 * or we still have pending I/Os.
1711 		 */
1712 		if (reclaim == 1) {
1713 			segvn_purge(seg);
1714 			reclaim = 0;
1715 			goto retry;
1716 		}
1717 		return (EAGAIN);
1718 	}
1719 
1720 	/*
1721 	 * Check for bad sizes
1722 	 */
1723 	if (addr < seg->s_base || addr + len > seg->s_base + seg->s_size ||
1724 	    (len & PAGEOFFSET) || ((uintptr_t)addr & PAGEOFFSET)) {
1725 		panic("segvn_unmap");
1726 		/*NOTREACHED*/
1727 	}
1728 
1729 	if (seg->s_szc != 0) {
1730 		size_t pgsz = page_get_pagesize(seg->s_szc);
1731 		int err;
1732 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
1733 			ASSERT(seg->s_base != addr || seg->s_size != len);
1734 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1735 				ASSERT(svd->amp == NULL);
1736 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1737 				hat_leave_region(seg->s_as->a_hat,
1738 				    svd->rcookie, HAT_REGION_TEXT);
1739 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
1740 				/*
1741 				 * could pass a flag to segvn_demote_range()
1742 				 * below to tell it not to do any unloads but
1743 				 * this case is rare enough to not bother for
1744 				 * now.
1745 				 */
1746 			} else if (svd->tr_state == SEGVN_TR_INIT) {
1747 				svd->tr_state = SEGVN_TR_OFF;
1748 			} else if (svd->tr_state == SEGVN_TR_ON) {
1749 				ASSERT(svd->amp != NULL);
1750 				segvn_textunrepl(seg, 1);
1751 				ASSERT(svd->amp == NULL);
1752 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
1753 			}
1754 			VM_STAT_ADD(segvnvmstats.demoterange[0]);
1755 			err = segvn_demote_range(seg, addr, len, SDR_END, 0);
1756 			if (err == 0) {
1757 				return (IE_RETRY);
1758 			}
1759 			return (err);
1760 		}
1761 	}
1762 
1763 	/* Inform the vnode of the unmapping. */
1764 	if (svd->vp) {
1765 		int error;
1766 
1767 		error = VOP_DELMAP(svd->vp,
1768 			(offset_t)svd->offset + (uintptr_t)(addr - seg->s_base),
1769 			seg->s_as, addr, len, svd->prot, svd->maxprot,
1770 			svd->type, svd->cred);
1771 
1772 		if (error == EAGAIN)
1773 			return (error);
1774 	}
1775 
1776 	/*
1777 	 * Remove any page locks set through this mapping.
1778 	 * If text replication is not off no page locks could have been
1779 	 * established via this mapping.
1780 	 */
1781 	if (svd->tr_state == SEGVN_TR_OFF) {
1782 		(void) segvn_lockop(seg, addr, len, 0, MC_UNLOCK, NULL, 0);
1783 	}
1784 
1785 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
1786 		ASSERT(svd->amp == NULL);
1787 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
1788 		ASSERT(svd->type == MAP_PRIVATE);
1789 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
1790 		    HAT_REGION_TEXT);
1791 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
1792 	} else if (svd->tr_state == SEGVN_TR_ON) {
1793 		ASSERT(svd->amp != NULL);
1794 		ASSERT(svd->pageprot == 0 && !(svd->prot & PROT_WRITE));
1795 		segvn_textunrepl(seg, 1);
1796 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
1797 	} else {
1798 		if (svd->tr_state != SEGVN_TR_OFF) {
1799 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
1800 			svd->tr_state = SEGVN_TR_OFF;
1801 		}
1802 		/*
1803 		 * Unload any hardware translations in the range to be taken
1804 		 * out. Use a callback to invoke free_vp_pages() effectively.
1805 		 */
1806 		if (svd->vp != NULL && free_pages != 0) {
1807 			callback.hcb_data = seg;
1808 			callback.hcb_function = segvn_hat_unload_callback;
1809 			cbp = &callback;
1810 		}
1811 		hat_unload_callback(seg->s_as->a_hat, addr, len,
1812 		    HAT_UNLOAD_UNMAP, cbp);
1813 
1814 		if (svd->type == MAP_SHARED && svd->vp != NULL &&
1815 		    (svd->vp->v_flag & VVMEXEC) &&
1816 		    ((svd->prot & PROT_WRITE) || svd->pageprot)) {
1817 			segvn_inval_trcache(svd->vp);
1818 		}
1819 	}
1820 
1821 	/*
1822 	 * Check for entire segment
1823 	 */
1824 	if (addr == seg->s_base && len == seg->s_size) {
1825 		seg_free(seg);
1826 		return (0);
1827 	}
1828 
1829 	opages = seg_pages(seg);
1830 	dpages = btop(len);
1831 	npages = opages - dpages;
1832 	amp = svd->amp;
1833 	ASSERT(amp == NULL || amp->a_szc >= seg->s_szc);
1834 
1835 	/*
1836 	 * Check for beginning of segment
1837 	 */
1838 	if (addr == seg->s_base) {
1839 		if (svd->vpage != NULL) {
1840 			size_t nbytes;
1841 			struct vpage *ovpage;
1842 
1843 			ovpage = svd->vpage;	/* keep pointer to vpage */
1844 
1845 			nbytes = vpgtob(npages);
1846 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
1847 			bcopy(&ovpage[dpages], svd->vpage, nbytes);
1848 
1849 			/* free up old vpage */
1850 			kmem_free(ovpage, vpgtob(opages));
1851 		}
1852 		if (amp != NULL) {
1853 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1854 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
1855 				/*
1856 				 * Free up now unused parts of anon_map array.
1857 				 */
1858 				if (amp->a_szc == seg->s_szc) {
1859 					if (seg->s_szc != 0) {
1860 						anon_free_pages(amp->ahp,
1861 						    svd->anon_index, len,
1862 						    seg->s_szc);
1863 					} else {
1864 						anon_free(amp->ahp,
1865 						    svd->anon_index,
1866 						    len);
1867 					}
1868 				} else {
1869 					ASSERT(svd->type == MAP_SHARED);
1870 					ASSERT(amp->a_szc > seg->s_szc);
1871 					anon_shmap_free_pages(amp,
1872 					    svd->anon_index, len);
1873 				}
1874 
1875 				/*
1876 				 * Unreserve swap space for the
1877 				 * unmapped chunk of this segment in
1878 				 * case it's MAP_SHARED
1879 				 */
1880 				if (svd->type == MAP_SHARED) {
1881 					anon_unresv(len);
1882 					amp->swresv -= len;
1883 				}
1884 			}
1885 			ANON_LOCK_EXIT(&amp->a_rwlock);
1886 			svd->anon_index += dpages;
1887 		}
1888 		if (svd->vp != NULL)
1889 			svd->offset += len;
1890 
1891 		if (svd->swresv) {
1892 			if (svd->flags & MAP_NORESERVE) {
1893 				ASSERT(amp);
1894 				oswresv = svd->swresv;
1895 
1896 				svd->swresv = ptob(anon_pages(amp->ahp,
1897 				    svd->anon_index, npages));
1898 				anon_unresv(oswresv - svd->swresv);
1899 			} else {
1900 				anon_unresv(len);
1901 				svd->swresv -= len;
1902 			}
1903 			TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
1904 				seg, len, 0);
1905 		}
1906 
1907 		seg->s_base += len;
1908 		seg->s_size -= len;
1909 		return (0);
1910 	}
1911 
1912 	/*
1913 	 * Check for end of segment
1914 	 */
1915 	if (addr + len == seg->s_base + seg->s_size) {
1916 		if (svd->vpage != NULL) {
1917 			size_t nbytes;
1918 			struct vpage *ovpage;
1919 
1920 			ovpage = svd->vpage;	/* keep pointer to vpage */
1921 
1922 			nbytes = vpgtob(npages);
1923 			svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
1924 			bcopy(ovpage, svd->vpage, nbytes);
1925 
1926 			/* free up old vpage */
1927 			kmem_free(ovpage, vpgtob(opages));
1928 
1929 		}
1930 		if (amp != NULL) {
1931 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
1932 			if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
1933 				/*
1934 				 * Free up now unused parts of anon_map array.
1935 				 */
1936 				ulong_t an_idx = svd->anon_index + npages;
1937 				if (amp->a_szc == seg->s_szc) {
1938 					if (seg->s_szc != 0) {
1939 						anon_free_pages(amp->ahp,
1940 						    an_idx, len,
1941 						    seg->s_szc);
1942 					} else {
1943 						anon_free(amp->ahp, an_idx,
1944 						    len);
1945 					}
1946 				} else {
1947 					ASSERT(svd->type == MAP_SHARED);
1948 					ASSERT(amp->a_szc > seg->s_szc);
1949 					anon_shmap_free_pages(amp,
1950 					    an_idx, len);
1951 				}
1952 
1953 				/*
1954 				 * Unreserve swap space for the
1955 				 * unmapped chunk of this segment in
1956 				 * case it's MAP_SHARED
1957 				 */
1958 				if (svd->type == MAP_SHARED) {
1959 					anon_unresv(len);
1960 					amp->swresv -= len;
1961 				}
1962 			}
1963 			ANON_LOCK_EXIT(&amp->a_rwlock);
1964 		}
1965 
1966 		if (svd->swresv) {
1967 			if (svd->flags & MAP_NORESERVE) {
1968 				ASSERT(amp);
1969 				oswresv = svd->swresv;
1970 				svd->swresv = ptob(anon_pages(amp->ahp,
1971 					svd->anon_index, npages));
1972 				anon_unresv(oswresv - svd->swresv);
1973 			} else {
1974 				anon_unresv(len);
1975 				svd->swresv -= len;
1976 			}
1977 			TRACE_3(TR_FAC_VM, TR_ANON_PROC,
1978 				"anon proc:%p %lu %u", seg, len, 0);
1979 		}
1980 
1981 		seg->s_size -= len;
1982 		return (0);
1983 	}
1984 
1985 	/*
1986 	 * The section to go is in the middle of the segment,
1987 	 * have to make it into two segments.  nseg is made for
1988 	 * the high end while seg is cut down at the low end.
1989 	 */
1990 	nbase = addr + len;				/* new seg base */
1991 	nsize = (seg->s_base + seg->s_size) - nbase;	/* new seg size */
1992 	seg->s_size = addr - seg->s_base;		/* shrink old seg */
1993 	nseg = seg_alloc(seg->s_as, nbase, nsize);
1994 	if (nseg == NULL) {
1995 		panic("segvn_unmap seg_alloc");
1996 		/*NOTREACHED*/
1997 	}
1998 	nseg->s_ops = seg->s_ops;
1999 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
2000 	nseg->s_data = (void *)nsvd;
2001 	nseg->s_szc = seg->s_szc;
2002 	*nsvd = *svd;
2003 	nsvd->seg = nseg;
2004 	nsvd->offset = svd->offset + (uintptr_t)(nseg->s_base - seg->s_base);
2005 	nsvd->swresv = 0;
2006 	nsvd->softlockcnt = 0;
2007 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
2008 
2009 	if (svd->vp != NULL) {
2010 		VN_HOLD(nsvd->vp);
2011 		if (nsvd->type == MAP_SHARED)
2012 			lgrp_shm_policy_init(NULL, nsvd->vp);
2013 	}
2014 	crhold(svd->cred);
2015 
2016 	if (svd->vpage == NULL) {
2017 		nsvd->vpage = NULL;
2018 	} else {
2019 		/* need to split vpage into two arrays */
2020 		size_t nbytes;
2021 		struct vpage *ovpage;
2022 
2023 		ovpage = svd->vpage;		/* keep pointer to vpage */
2024 
2025 		npages = seg_pages(seg);	/* seg has shrunk */
2026 		nbytes = vpgtob(npages);
2027 		svd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2028 
2029 		bcopy(ovpage, svd->vpage, nbytes);
2030 
2031 		npages = seg_pages(nseg);
2032 		nbytes = vpgtob(npages);
2033 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
2034 
2035 		bcopy(&ovpage[opages - npages], nsvd->vpage, nbytes);
2036 
2037 		/* free up old vpage */
2038 		kmem_free(ovpage, vpgtob(opages));
2039 	}
2040 
2041 	if (amp == NULL) {
2042 		nsvd->amp = NULL;
2043 		nsvd->anon_index = 0;
2044 	} else {
2045 		/*
2046 		 * Need to create a new anon map for the new segment.
2047 		 * We'll also allocate a new smaller array for the old
2048 		 * smaller segment to save space.
2049 		 */
2050 		opages = btop((uintptr_t)(addr - seg->s_base));
2051 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2052 		if (amp->refcnt == 1 || svd->type == MAP_PRIVATE) {
2053 			/*
2054 			 * Free up now unused parts of anon_map array.
2055 			 */
2056 			ulong_t an_idx = svd->anon_index + opages;
2057 			if (amp->a_szc == seg->s_szc) {
2058 				if (seg->s_szc != 0) {
2059 					anon_free_pages(amp->ahp, an_idx, len,
2060 					    seg->s_szc);
2061 				} else {
2062 					anon_free(amp->ahp, an_idx,
2063 					    len);
2064 				}
2065 			} else {
2066 				ASSERT(svd->type == MAP_SHARED);
2067 				ASSERT(amp->a_szc > seg->s_szc);
2068 				anon_shmap_free_pages(amp, an_idx, len);
2069 			}
2070 
2071 			/*
2072 			 * Unreserve swap space for the
2073 			 * unmapped chunk of this segment in
2074 			 * case it's MAP_SHARED
2075 			 */
2076 			if (svd->type == MAP_SHARED) {
2077 				anon_unresv(len);
2078 				amp->swresv -= len;
2079 			}
2080 		}
2081 		nsvd->anon_index = svd->anon_index +
2082 		    btop((uintptr_t)(nseg->s_base - seg->s_base));
2083 		if (svd->type == MAP_SHARED) {
2084 			amp->refcnt++;
2085 			nsvd->amp = amp;
2086 		} else {
2087 			struct anon_map *namp;
2088 			struct anon_hdr *nahp;
2089 
2090 			ASSERT(svd->type == MAP_PRIVATE);
2091 			nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
2092 			namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
2093 			namp->a_szc = seg->s_szc;
2094 			(void) anon_copy_ptr(amp->ahp, svd->anon_index, nahp,
2095 			    0, btop(seg->s_size), ANON_SLEEP);
2096 			(void) anon_copy_ptr(amp->ahp, nsvd->anon_index,
2097 			    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
2098 			anon_release(amp->ahp, btop(amp->size));
2099 			svd->anon_index = 0;
2100 			nsvd->anon_index = 0;
2101 			amp->ahp = nahp;
2102 			amp->size = seg->s_size;
2103 			nsvd->amp = namp;
2104 		}
2105 		ANON_LOCK_EXIT(&amp->a_rwlock);
2106 	}
2107 	if (svd->swresv) {
2108 		if (svd->flags & MAP_NORESERVE) {
2109 			ASSERT(amp);
2110 			oswresv = svd->swresv;
2111 			svd->swresv = ptob(anon_pages(amp->ahp,
2112 				svd->anon_index, btop(seg->s_size)));
2113 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
2114 				nsvd->anon_index, btop(nseg->s_size)));
2115 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
2116 			anon_unresv(oswresv - (svd->swresv + nsvd->swresv));
2117 		} else {
2118 			if (seg->s_size + nseg->s_size + len != svd->swresv) {
2119 				panic("segvn_unmap: "
2120 				    "cannot split swap reservation");
2121 				/*NOTREACHED*/
2122 			}
2123 			anon_unresv(len);
2124 			svd->swresv = seg->s_size;
2125 			nsvd->swresv = nseg->s_size;
2126 		}
2127 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2128 			seg, len, 0);
2129 	}
2130 
2131 	return (0);			/* I'm glad that's all over with! */
2132 }
2133 
2134 static void
2135 segvn_free(struct seg *seg)
2136 {
2137 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2138 	pgcnt_t npages = seg_pages(seg);
2139 	struct anon_map *amp;
2140 	size_t len;
2141 
2142 	/*
2143 	 * We don't need any segment level locks for "segvn" data
2144 	 * since the address space is "write" locked.
2145 	 */
2146 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
2147 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
2148 
2149 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2150 
2151 	/*
2152 	 * Be sure to unlock pages. XXX Why do things get free'ed instead
2153 	 * of unmapped? XXX
2154 	 */
2155 	(void) segvn_lockop(seg, seg->s_base, seg->s_size,
2156 	    0, MC_UNLOCK, NULL, 0);
2157 
2158 	/*
2159 	 * Deallocate the vpage and anon pointers if necessary and possible.
2160 	 */
2161 	if (svd->vpage != NULL) {
2162 		kmem_free(svd->vpage, vpgtob(npages));
2163 		svd->vpage = NULL;
2164 	}
2165 	if ((amp = svd->amp) != NULL) {
2166 		/*
2167 		 * If there are no more references to this anon_map
2168 		 * structure, then deallocate the structure after freeing
2169 		 * up all the anon slot pointers that we can.
2170 		 */
2171 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
2172 		ASSERT(amp->a_szc >= seg->s_szc);
2173 		if (--amp->refcnt == 0) {
2174 			if (svd->type == MAP_PRIVATE) {
2175 				/*
2176 				 * Private - we only need to anon_free
2177 				 * the part that this segment refers to.
2178 				 */
2179 				if (seg->s_szc != 0) {
2180 					anon_free_pages(amp->ahp,
2181 					    svd->anon_index, seg->s_size,
2182 					    seg->s_szc);
2183 				} else {
2184 					anon_free(amp->ahp, svd->anon_index,
2185 					    seg->s_size);
2186 				}
2187 			} else {
2188 				/*
2189 				 * Shared - anon_free the entire
2190 				 * anon_map's worth of stuff and
2191 				 * release any swap reservation.
2192 				 */
2193 				if (amp->a_szc != 0) {
2194 					anon_shmap_free_pages(amp, 0,
2195 					    amp->size);
2196 				} else {
2197 					anon_free(amp->ahp, 0, amp->size);
2198 				}
2199 				if ((len = amp->swresv) != 0) {
2200 					anon_unresv(len);
2201 					TRACE_3(TR_FAC_VM, TR_ANON_PROC,
2202 						"anon proc:%p %lu %u",
2203 						seg, len, 0);
2204 				}
2205 			}
2206 			svd->amp = NULL;
2207 			ANON_LOCK_EXIT(&amp->a_rwlock);
2208 			anonmap_free(amp);
2209 		} else if (svd->type == MAP_PRIVATE) {
2210 			/*
2211 			 * We had a private mapping which still has
2212 			 * a held anon_map so just free up all the
2213 			 * anon slot pointers that we were using.
2214 			 */
2215 			if (seg->s_szc != 0) {
2216 				anon_free_pages(amp->ahp, svd->anon_index,
2217 				    seg->s_size, seg->s_szc);
2218 			} else {
2219 				anon_free(amp->ahp, svd->anon_index,
2220 				    seg->s_size);
2221 			}
2222 			ANON_LOCK_EXIT(&amp->a_rwlock);
2223 		} else {
2224 			ANON_LOCK_EXIT(&amp->a_rwlock);
2225 		}
2226 	}
2227 
2228 	/*
2229 	 * Release swap reservation.
2230 	 */
2231 	if ((len = svd->swresv) != 0) {
2232 		anon_unresv(svd->swresv);
2233 		TRACE_3(TR_FAC_VM, TR_ANON_PROC, "anon proc:%p %lu %u",
2234 			seg, len, 0);
2235 		svd->swresv = 0;
2236 	}
2237 	/*
2238 	 * Release claim on vnode, credentials, and finally free the
2239 	 * private data.
2240 	 */
2241 	if (svd->vp != NULL) {
2242 		if (svd->type == MAP_SHARED)
2243 			lgrp_shm_policy_fini(NULL, svd->vp);
2244 		VN_RELE(svd->vp);
2245 		svd->vp = NULL;
2246 	}
2247 	crfree(svd->cred);
2248 	svd->cred = NULL;
2249 
2250 	seg->s_data = NULL;
2251 	kmem_cache_free(segvn_cache, svd);
2252 }
2253 
2254 #ifdef DEBUG
2255 uint32_t segvn_slock_mtbf = 0;
2256 #endif
2257 
2258 ulong_t segvn_lpglck_limit = 0;
2259 
2260 /*
2261  * Support routines used by segvn_pagelock() and softlock faults for anonymous
2262  * pages to implement availrmem accounting in a way that makes sure the
2263  * same memory is accounted just once for all softlock/pagelock purposes.
2264  * This prevents a bug when availrmem is quickly incorrectly exausted from
2265  * several pagelocks to different parts of the same large page since each
2266  * pagelock has to decrement availrmem by the size of the entire large
2267  * page. Note those pages are not COW shared until softunlock/pageunlock so
2268  * we don't need to use cow style accounting here.  We also need to make sure
2269  * the entire large page is accounted even if softlock range is less than the
2270  * entire large page because large anon pages can't be demoted when any of
2271  * constituent pages is locked. The caller calls this routine for every page_t
2272  * it locks. The very first page in the range may not be the root page of a
2273  * large page. For all other pages it's guranteed we are going to visit the
2274  * root of a particular large page before any other constituent page as we are
2275  * locking sequential pages belonging to the same anon map. So we do all the
2276  * locking when the root is encountered except for the very first page.  Since
2277  * softlocking is not supported (except S_READ_NOCOW special case) for vmpss
2278  * segments and since vnode pages can be demoted without locking all
2279  * constituent pages vnode pages don't come here.  Unlocking relies on the
2280  * fact that pagesize can't change whenever any of constituent large pages is
2281  * locked at least SE_SHARED. This allows unlocking code to find the right
2282  * root and decrement availrmem by the same amount it was incremented when the
2283  * page was locked.
2284  */
2285 static int
2286 segvn_slock_anonpages(page_t *pp, int first)
2287 {
2288 	pgcnt_t		pages;
2289 	pfn_t		pfn;
2290 	uchar_t		szc = pp->p_szc;
2291 
2292 	ASSERT(PAGE_LOCKED(pp));
2293 	ASSERT(pp->p_vnode != NULL);
2294 	ASSERT(IS_SWAPFSVP(pp->p_vnode));
2295 
2296 	/*
2297 	 * pagesize won't change as long as any constituent page is locked.
2298 	 */
2299 	pages = page_get_pagecnt(pp->p_szc);
2300 	pfn = page_pptonum(pp);
2301 
2302 	if (!first) {
2303 		if (!IS_P2ALIGNED(pfn, pages)) {
2304 #ifdef DEBUG
2305 			pp = &pp[-(spgcnt_t)(pfn & (pages - 1))];
2306 			pfn = page_pptonum(pp);
2307 			ASSERT(IS_P2ALIGNED(pfn, pages));
2308 			ASSERT(pp->p_szc == szc);
2309 			ASSERT(pp->p_vnode != NULL);
2310 			ASSERT(IS_SWAPFSVP(pp->p_vnode));
2311 			ASSERT(pp->p_slckcnt != 0);
2312 #endif /* DEBUG */
2313 			return (1);
2314 		}
2315 	} else if (!IS_P2ALIGNED(pfn, pages)) {
2316 		pp = &pp[-(spgcnt_t)(pfn & (pages - 1))];
2317 #ifdef DEBUG
2318 		pfn = page_pptonum(pp);
2319 		ASSERT(IS_P2ALIGNED(pfn, pages));
2320 		ASSERT(pp->p_szc == szc);
2321 		ASSERT(pp->p_vnode != NULL);
2322 		ASSERT(IS_SWAPFSVP(pp->p_vnode));
2323 #endif /* DEBUG */
2324 	}
2325 
2326 #ifdef DEBUG
2327 	if (segvn_slock_mtbf && !(gethrtime() % segvn_slock_mtbf)) {
2328 		return (0);
2329 	}
2330 #endif /* DEBUG */
2331 
2332 	/*
2333 	 * pp is a root page.
2334 	 * We haven't locked this large page yet.
2335 	 */
2336 	page_struct_lock(pp);
2337 	if (pp->p_slckcnt != 0) {
2338 		if (pp->p_slckcnt < PAGE_SLOCK_MAXIMUM) {
2339 			pp->p_slckcnt++;
2340 			page_struct_unlock(pp);
2341 			return (1);
2342 		}
2343 		page_struct_unlock(pp);
2344 		segvn_lpglck_limit++;
2345 		return (0);
2346 	}
2347 	mutex_enter(&freemem_lock);
2348 	if (availrmem < tune.t_minarmem + pages) {
2349 		mutex_exit(&freemem_lock);
2350 		page_struct_unlock(pp);
2351 		return (0);
2352 	}
2353 	pp->p_slckcnt++;
2354 	availrmem -= pages;
2355 	mutex_exit(&freemem_lock);
2356 	page_struct_unlock(pp);
2357 	return (1);
2358 }
2359 
2360 static void
2361 segvn_sunlock_anonpages(page_t *pp, int first)
2362 {
2363 	pgcnt_t		pages;
2364 	pfn_t		pfn;
2365 
2366 	ASSERT(PAGE_LOCKED(pp));
2367 	ASSERT(pp->p_vnode != NULL);
2368 	ASSERT(IS_SWAPFSVP(pp->p_vnode));
2369 
2370 	/*
2371 	 * pagesize won't change as long as any constituent page is locked.
2372 	 */
2373 	pages = page_get_pagecnt(pp->p_szc);
2374 	pfn = page_pptonum(pp);
2375 
2376 	if (!first) {
2377 		if (!IS_P2ALIGNED(pfn, pages)) {
2378 			return;
2379 		}
2380 	} else if (!IS_P2ALIGNED(pfn, pages)) {
2381 		pp = &pp[-(spgcnt_t)(pfn & (pages - 1))];
2382 #ifdef DEBUG
2383 		pfn = page_pptonum(pp);
2384 		ASSERT(IS_P2ALIGNED(pfn, pages));
2385 #endif /* DEBUG */
2386 	}
2387 	ASSERT(pp->p_vnode != NULL);
2388 	ASSERT(IS_SWAPFSVP(pp->p_vnode));
2389 	ASSERT(pp->p_slckcnt != 0);
2390 	page_struct_lock(pp);
2391 	if (--pp->p_slckcnt == 0) {
2392 		mutex_enter(&freemem_lock);
2393 		availrmem += pages;
2394 		mutex_exit(&freemem_lock);
2395 	}
2396 	page_struct_unlock(pp);
2397 }
2398 
2399 /*
2400  * Do a F_SOFTUNLOCK call over the range requested.  The range must have
2401  * already been F_SOFTLOCK'ed.
2402  * Caller must always match addr and len of a softunlock with a previous
2403  * softlock with exactly the same addr and len.
2404  */
2405 static void
2406 segvn_softunlock(struct seg *seg, caddr_t addr, size_t len, enum seg_rw rw)
2407 {
2408 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2409 	page_t *pp;
2410 	caddr_t adr;
2411 	struct vnode *vp;
2412 	u_offset_t offset;
2413 	ulong_t anon_index;
2414 	struct anon_map *amp;
2415 	struct anon *ap = NULL;
2416 
2417 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
2418 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
2419 
2420 	if ((amp = svd->amp) != NULL)
2421 		anon_index = svd->anon_index + seg_page(seg, addr);
2422 
2423 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
2424 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2425 		hat_unlock_region(seg->s_as->a_hat, addr, len, svd->rcookie);
2426 	} else {
2427 		hat_unlock(seg->s_as->a_hat, addr, len);
2428 	}
2429 	for (adr = addr; adr < addr + len; adr += PAGESIZE) {
2430 		if (amp != NULL) {
2431 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
2432 			if ((ap = anon_get_ptr(amp->ahp, anon_index++))
2433 								!= NULL) {
2434 				swap_xlate(ap, &vp, &offset);
2435 			} else {
2436 				vp = svd->vp;
2437 				offset = svd->offset +
2438 				    (uintptr_t)(adr - seg->s_base);
2439 			}
2440 			ANON_LOCK_EXIT(&amp->a_rwlock);
2441 		} else {
2442 			vp = svd->vp;
2443 			offset = svd->offset +
2444 			    (uintptr_t)(adr - seg->s_base);
2445 		}
2446 
2447 		/*
2448 		 * Use page_find() instead of page_lookup() to
2449 		 * find the page since we know that it is locked.
2450 		 */
2451 		pp = page_find(vp, offset);
2452 		if (pp == NULL) {
2453 			panic(
2454 			    "segvn_softunlock: addr %p, ap %p, vp %p, off %llx",
2455 			    (void *)adr, (void *)ap, (void *)vp, offset);
2456 			/*NOTREACHED*/
2457 		}
2458 
2459 		if (rw == S_WRITE) {
2460 			hat_setrefmod(pp);
2461 			if (seg->s_as->a_vbits)
2462 				hat_setstat(seg->s_as, adr, PAGESIZE,
2463 				    P_REF | P_MOD);
2464 		} else if (rw != S_OTHER) {
2465 			hat_setref(pp);
2466 			if (seg->s_as->a_vbits)
2467 				hat_setstat(seg->s_as, adr, PAGESIZE, P_REF);
2468 		}
2469 		TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2470 			"segvn_fault:pp %p vp %p offset %llx", pp, vp, offset);
2471 		if (svd->vp == NULL) {
2472 			segvn_sunlock_anonpages(pp, adr == addr);
2473 		}
2474 		page_unlock(pp);
2475 	}
2476 	mutex_enter(&freemem_lock); /* for availrmem */
2477 	if (svd->vp != NULL) {
2478 		availrmem += btop(len);
2479 	}
2480 	segvn_pages_locked -= btop(len);
2481 	svd->softlockcnt -= btop(len);
2482 	mutex_exit(&freemem_lock);
2483 	if (svd->softlockcnt == 0) {
2484 		/*
2485 		 * All SOFTLOCKS are gone. Wakeup any waiting
2486 		 * unmappers so they can try again to unmap.
2487 		 * Check for waiters first without the mutex
2488 		 * held so we don't always grab the mutex on
2489 		 * softunlocks.
2490 		 */
2491 		if (AS_ISUNMAPWAIT(seg->s_as)) {
2492 			mutex_enter(&seg->s_as->a_contents);
2493 			if (AS_ISUNMAPWAIT(seg->s_as)) {
2494 				AS_CLRUNMAPWAIT(seg->s_as);
2495 				cv_broadcast(&seg->s_as->a_cv);
2496 			}
2497 			mutex_exit(&seg->s_as->a_contents);
2498 		}
2499 	}
2500 }
2501 
2502 #define	PAGE_HANDLED	((page_t *)-1)
2503 
2504 /*
2505  * Release all the pages in the NULL terminated ppp list
2506  * which haven't already been converted to PAGE_HANDLED.
2507  */
2508 static void
2509 segvn_pagelist_rele(page_t **ppp)
2510 {
2511 	for (; *ppp != NULL; ppp++) {
2512 		if (*ppp != PAGE_HANDLED)
2513 			page_unlock(*ppp);
2514 	}
2515 }
2516 
2517 static int stealcow = 1;
2518 
2519 /*
2520  * Workaround for viking chip bug.  See bug id 1220902.
2521  * To fix this down in pagefault() would require importing so
2522  * much as and segvn code as to be unmaintainable.
2523  */
2524 int enable_mbit_wa = 0;
2525 
2526 /*
2527  * Handles all the dirty work of getting the right
2528  * anonymous pages and loading up the translations.
2529  * This routine is called only from segvn_fault()
2530  * when looping over the range of addresses requested.
2531  *
2532  * The basic algorithm here is:
2533  * 	If this is an anon_zero case
2534  *		Call anon_zero to allocate page
2535  *		Load up translation
2536  *		Return
2537  *	endif
2538  *	If this is an anon page
2539  *		Use anon_getpage to get the page
2540  *	else
2541  *		Find page in pl[] list passed in
2542  *	endif
2543  *	If not a cow
2544  *		Load up the translation to the page
2545  *		return
2546  *	endif
2547  *	Call anon_private to handle cow
2548  *	Load up (writable) translation to new page
2549  */
2550 static faultcode_t
2551 segvn_faultpage(
2552 	struct hat *hat,		/* the hat to use for mapping */
2553 	struct seg *seg,		/* seg_vn of interest */
2554 	caddr_t addr,			/* address in as */
2555 	u_offset_t off,			/* offset in vp */
2556 	struct vpage *vpage,		/* pointer to vpage for vp, off */
2557 	page_t *pl[],			/* object source page pointer */
2558 	uint_t vpprot,			/* access allowed to object pages */
2559 	enum fault_type type,		/* type of fault */
2560 	enum seg_rw rw,			/* type of access at fault */
2561 	int brkcow,			/* we may need to break cow */
2562 	int first)			/* first page for this fault if 1 */
2563 {
2564 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
2565 	page_t *pp, **ppp;
2566 	uint_t pageflags = 0;
2567 	page_t *anon_pl[1 + 1];
2568 	page_t *opp = NULL;		/* original page */
2569 	uint_t prot;
2570 	int err;
2571 	int cow;
2572 	int claim;
2573 	int steal = 0;
2574 	ulong_t anon_index;
2575 	struct anon *ap, *oldap;
2576 	struct anon_map *amp;
2577 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
2578 	int anon_lock = 0;
2579 	anon_sync_obj_t cookie;
2580 
2581 	if (svd->flags & MAP_TEXT) {
2582 		hat_flag |= HAT_LOAD_TEXT;
2583 	}
2584 
2585 	ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
2586 	ASSERT(seg->s_szc == 0);
2587 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
2588 
2589 	/*
2590 	 * Initialize protection value for this page.
2591 	 * If we have per page protection values check it now.
2592 	 */
2593 	if (svd->pageprot) {
2594 		uint_t protchk;
2595 
2596 		switch (rw) {
2597 		case S_READ:
2598 			protchk = PROT_READ;
2599 			break;
2600 		case S_WRITE:
2601 			protchk = PROT_WRITE;
2602 			break;
2603 		case S_EXEC:
2604 			protchk = PROT_EXEC;
2605 			break;
2606 		case S_OTHER:
2607 		default:
2608 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
2609 			break;
2610 		}
2611 
2612 		prot = VPP_PROT(vpage);
2613 		if ((prot & protchk) == 0)
2614 			return (FC_PROT);	/* illegal access type */
2615 	} else {
2616 		prot = svd->prot;
2617 	}
2618 
2619 	if (type == F_SOFTLOCK && svd->vp != NULL) {
2620 		mutex_enter(&freemem_lock);
2621 		if (availrmem <= tune.t_minarmem) {
2622 			mutex_exit(&freemem_lock);
2623 			return (FC_MAKE_ERR(ENOMEM));	/* out of real memory */
2624 		} else {
2625 			availrmem--;
2626 			svd->softlockcnt++;
2627 			segvn_pages_locked++;
2628 		}
2629 		mutex_exit(&freemem_lock);
2630 	}
2631 
2632 	/*
2633 	 * Always acquire the anon array lock to prevent 2 threads from
2634 	 * allocating separate anon slots for the same "addr".
2635 	 */
2636 
2637 	if ((amp = svd->amp) != NULL) {
2638 		ASSERT(RW_READ_HELD(&amp->a_rwlock));
2639 		anon_index = svd->anon_index + seg_page(seg, addr);
2640 		anon_array_enter(amp, anon_index, &cookie);
2641 		anon_lock = 1;
2642 	}
2643 
2644 	if (svd->vp == NULL && amp != NULL) {
2645 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL) {
2646 			/*
2647 			 * Allocate a (normally) writable anonymous page of
2648 			 * zeroes. If no advance reservations, reserve now.
2649 			 */
2650 			if (svd->flags & MAP_NORESERVE) {
2651 				if (anon_resv_zone(ptob(1),
2652 				    seg->s_as->a_proc->p_zone)) {
2653 					atomic_add_long(&svd->swresv, ptob(1));
2654 				} else {
2655 					err = ENOMEM;
2656 					goto out;
2657 				}
2658 			}
2659 			if ((pp = anon_zero(seg, addr, &ap,
2660 			    svd->cred)) == NULL) {
2661 				err = ENOMEM;
2662 				goto out;	/* out of swap space */
2663 			}
2664 			/*
2665 			 * Re-acquire the anon_map lock and
2666 			 * initialize the anon array entry.
2667 			 */
2668 			(void) anon_set_ptr(amp->ahp, anon_index, ap,
2669 				ANON_SLEEP);
2670 
2671 			ASSERT(pp->p_szc == 0);
2672 
2673 			/*
2674 			 * Handle pages that have been marked for migration
2675 			 */
2676 			if (lgrp_optimizations())
2677 				page_migrate(seg, addr, &pp, 1);
2678 
2679 			if (type == F_SOFTLOCK) {
2680 				if (!segvn_slock_anonpages(pp, first)) {
2681 					page_unlock(pp);
2682 					err = ENOMEM;
2683 					goto out;
2684 				} else {
2685 					mutex_enter(&freemem_lock);
2686 					svd->softlockcnt++;
2687 					segvn_pages_locked++;
2688 					mutex_exit(&freemem_lock);
2689 				}
2690 			}
2691 
2692 			if (enable_mbit_wa) {
2693 				if (rw == S_WRITE)
2694 					hat_setmod(pp);
2695 				else if (!hat_ismod(pp))
2696 					prot &= ~PROT_WRITE;
2697 			}
2698 			/*
2699 			 * If AS_PAGLCK is set in a_flags (via memcntl(2)
2700 			 * with MC_LOCKAS, MCL_FUTURE) and this is a
2701 			 * MAP_NORESERVE segment, we may need to
2702 			 * permanently lock the page as it is being faulted
2703 			 * for the first time. The following text applies
2704 			 * only to MAP_NORESERVE segments:
2705 			 *
2706 			 * As per memcntl(2), if this segment was created
2707 			 * after MCL_FUTURE was applied (a "future"
2708 			 * segment), its pages must be locked.  If this
2709 			 * segment existed at MCL_FUTURE application (a
2710 			 * "past" segment), the interface is unclear.
2711 			 *
2712 			 * We decide to lock only if vpage is present:
2713 			 *
2714 			 * - "future" segments will have a vpage array (see
2715 			 *    as_map), and so will be locked as required
2716 			 *
2717 			 * - "past" segments may not have a vpage array,
2718 			 *    depending on whether events (such as
2719 			 *    mprotect) have occurred. Locking if vpage
2720 			 *    exists will preserve legacy behavior.  Not
2721 			 *    locking if vpage is absent, will not break
2722 			 *    the interface or legacy behavior.  Note that
2723 			 *    allocating vpage here if it's absent requires
2724 			 *    upgrading the segvn reader lock, the cost of
2725 			 *    which does not seem worthwhile.
2726 			 *
2727 			 * Usually testing and setting VPP_ISPPLOCK and
2728 			 * VPP_SETPPLOCK requires holding the segvn lock as
2729 			 * writer, but in this case all readers are
2730 			 * serializing on the anon array lock.
2731 			 */
2732 			if (AS_ISPGLCK(seg->s_as) && vpage != NULL &&
2733 			    (svd->flags & MAP_NORESERVE) &&
2734 			    !VPP_ISPPLOCK(vpage)) {
2735 				proc_t *p = seg->s_as->a_proc;
2736 				ASSERT(svd->type == MAP_PRIVATE);
2737 				mutex_enter(&p->p_lock);
2738 				if (rctl_incr_locked_mem(p, NULL, PAGESIZE,
2739 				    1) == 0) {
2740 					claim = VPP_PROT(vpage) & PROT_WRITE;
2741 					if (page_pp_lock(pp, claim, 0)) {
2742 						VPP_SETPPLOCK(vpage);
2743 					} else {
2744 						rctl_decr_locked_mem(p, NULL,
2745 						    PAGESIZE, 1);
2746 					}
2747 				}
2748 				mutex_exit(&p->p_lock);
2749 			}
2750 
2751 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2752 			hat_memload(hat, addr, pp, prot, hat_flag);
2753 
2754 			if (!(hat_flag & HAT_LOAD_LOCK))
2755 				page_unlock(pp);
2756 
2757 			anon_array_exit(&cookie);
2758 			return (0);
2759 		}
2760 	}
2761 
2762 	/*
2763 	 * Obtain the page structure via anon_getpage() if it is
2764 	 * a private copy of an object (the result of a previous
2765 	 * copy-on-write).
2766 	 */
2767 	if (amp != NULL) {
2768 		if ((ap = anon_get_ptr(amp->ahp, anon_index)) != NULL) {
2769 			err = anon_getpage(&ap, &vpprot, anon_pl, PAGESIZE,
2770 			    seg, addr, rw, svd->cred);
2771 			if (err)
2772 				goto out;
2773 
2774 			if (svd->type == MAP_SHARED) {
2775 				/*
2776 				 * If this is a shared mapping to an
2777 				 * anon_map, then ignore the write
2778 				 * permissions returned by anon_getpage().
2779 				 * They apply to the private mappings
2780 				 * of this anon_map.
2781 				 */
2782 				vpprot |= PROT_WRITE;
2783 			}
2784 			opp = anon_pl[0];
2785 		}
2786 	}
2787 
2788 	/*
2789 	 * Search the pl[] list passed in if it is from the
2790 	 * original object (i.e., not a private copy).
2791 	 */
2792 	if (opp == NULL) {
2793 		/*
2794 		 * Find original page.  We must be bringing it in
2795 		 * from the list in pl[].
2796 		 */
2797 		for (ppp = pl; (opp = *ppp) != NULL; ppp++) {
2798 			if (opp == PAGE_HANDLED)
2799 				continue;
2800 			ASSERT(opp->p_vnode == svd->vp); /* XXX */
2801 			if (opp->p_offset == off)
2802 				break;
2803 		}
2804 		if (opp == NULL) {
2805 			panic("segvn_faultpage not found");
2806 			/*NOTREACHED*/
2807 		}
2808 		*ppp = PAGE_HANDLED;
2809 
2810 	}
2811 
2812 	ASSERT(PAGE_LOCKED(opp));
2813 
2814 	TRACE_3(TR_FAC_VM, TR_SEGVN_FAULT,
2815 		"segvn_fault:pp %p vp %p offset %llx",
2816 		opp, NULL, 0);
2817 
2818 	/*
2819 	 * The fault is treated as a copy-on-write fault if a
2820 	 * write occurs on a private segment and the object
2821 	 * page (i.e., mapping) is write protected.  We assume
2822 	 * that fatal protection checks have already been made.
2823 	 */
2824 
2825 	if (brkcow) {
2826 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
2827 		cow = !(vpprot & PROT_WRITE);
2828 	} else if (svd->tr_state == SEGVN_TR_ON) {
2829 		/*
2830 		 * If we are doing text replication COW on first touch.
2831 		 */
2832 		ASSERT(amp != NULL);
2833 		ASSERT(svd->vp != NULL);
2834 		ASSERT(rw != S_WRITE);
2835 		cow = (ap == NULL);
2836 	} else {
2837 		cow = 0;
2838 	}
2839 
2840 	/*
2841 	 * If not a copy-on-write case load the translation
2842 	 * and return.
2843 	 */
2844 	if (cow == 0) {
2845 
2846 		/*
2847 		 * Handle pages that have been marked for migration
2848 		 */
2849 		if (lgrp_optimizations())
2850 			page_migrate(seg, addr, &opp, 1);
2851 
2852 		if (type == F_SOFTLOCK && svd->vp == NULL) {
2853 
2854 			ASSERT(opp->p_szc == 0 ||
2855 			    (svd->type == MAP_SHARED &&
2856 				amp != NULL && amp->a_szc != 0));
2857 
2858 			if (!segvn_slock_anonpages(opp, first)) {
2859 				page_unlock(opp);
2860 				err = ENOMEM;
2861 				goto out;
2862 			} else {
2863 				mutex_enter(&freemem_lock);
2864 				svd->softlockcnt++;
2865 				segvn_pages_locked++;
2866 				mutex_exit(&freemem_lock);
2867 			}
2868 		}
2869 		if (IS_VMODSORT(opp->p_vnode) || enable_mbit_wa) {
2870 			if (rw == S_WRITE)
2871 				hat_setmod(opp);
2872 			else if (rw != S_OTHER && !hat_ismod(opp))
2873 				prot &= ~PROT_WRITE;
2874 		}
2875 
2876 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
2877 		    (!svd->pageprot && svd->prot == (prot & vpprot)));
2878 		ASSERT(amp == NULL ||
2879 		    svd->rcookie == HAT_INVALID_REGION_COOKIE);
2880 		hat_memload_region(hat, addr, opp, prot & vpprot, hat_flag,
2881 		    svd->rcookie);
2882 
2883 		if (!(hat_flag & HAT_LOAD_LOCK))
2884 			page_unlock(opp);
2885 
2886 		if (anon_lock) {
2887 			anon_array_exit(&cookie);
2888 		}
2889 		return (0);
2890 	}
2891 
2892 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2893 
2894 	hat_setref(opp);
2895 
2896 	ASSERT(amp != NULL && anon_lock);
2897 
2898 	/*
2899 	 * Steal the page only if it isn't a private page
2900 	 * since stealing a private page is not worth the effort.
2901 	 */
2902 	if ((ap = anon_get_ptr(amp->ahp, anon_index)) == NULL)
2903 		steal = 1;
2904 
2905 	/*
2906 	 * Steal the original page if the following conditions are true:
2907 	 *
2908 	 * We are low on memory, the page is not private, page is not large,
2909 	 * not shared, not modified, not `locked' or if we have it `locked'
2910 	 * (i.e., p_cowcnt == 1 and p_lckcnt == 0, which also implies
2911 	 * that the page is not shared) and if it doesn't have any
2912 	 * translations. page_struct_lock isn't needed to look at p_cowcnt
2913 	 * and p_lckcnt because we first get exclusive lock on page.
2914 	 */
2915 	(void) hat_pagesync(opp, HAT_SYNC_DONTZERO | HAT_SYNC_STOPON_MOD);
2916 
2917 	if (stealcow && freemem < minfree && steal && opp->p_szc == 0 &&
2918 	    page_tryupgrade(opp) && !hat_ismod(opp) &&
2919 	    ((opp->p_lckcnt == 0 && opp->p_cowcnt == 0) ||
2920 	    (opp->p_lckcnt == 0 && opp->p_cowcnt == 1 &&
2921 	    vpage != NULL && VPP_ISPPLOCK(vpage)))) {
2922 		/*
2923 		 * Check if this page has other translations
2924 		 * after unloading our translation.
2925 		 */
2926 		if (hat_page_is_mapped(opp)) {
2927 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
2928 			hat_unload(seg->s_as->a_hat, addr, PAGESIZE,
2929 				HAT_UNLOAD);
2930 		}
2931 
2932 		/*
2933 		 * hat_unload() might sync back someone else's recent
2934 		 * modification, so check again.
2935 		 */
2936 		if (!hat_ismod(opp) && !hat_page_is_mapped(opp))
2937 			pageflags |= STEAL_PAGE;
2938 	}
2939 
2940 	/*
2941 	 * If we have a vpage pointer, see if it indicates that we have
2942 	 * ``locked'' the page we map -- if so, tell anon_private to
2943 	 * transfer the locking resource to the new page.
2944 	 *
2945 	 * See Statement at the beginning of segvn_lockop regarding
2946 	 * the way lockcnts/cowcnts are handled during COW.
2947 	 *
2948 	 */
2949 	if (vpage != NULL && VPP_ISPPLOCK(vpage))
2950 		pageflags |= LOCK_PAGE;
2951 
2952 	/*
2953 	 * Allocate a private page and perform the copy.
2954 	 * For MAP_NORESERVE reserve swap space now, unless this
2955 	 * is a cow fault on an existing anon page in which case
2956 	 * MAP_NORESERVE will have made advance reservations.
2957 	 */
2958 	if ((svd->flags & MAP_NORESERVE) && (ap == NULL)) {
2959 		if (anon_resv_zone(ptob(1), seg->s_as->a_proc->p_zone)) {
2960 			atomic_add_long(&svd->swresv, ptob(1));
2961 		} else {
2962 			page_unlock(opp);
2963 			err = ENOMEM;
2964 			goto out;
2965 		}
2966 	}
2967 	oldap = ap;
2968 	pp = anon_private(&ap, seg, addr, prot, opp, pageflags, svd->cred);
2969 	if (pp == NULL) {
2970 		err = ENOMEM;	/* out of swap space */
2971 		goto out;
2972 	}
2973 
2974 	/*
2975 	 * If we copied away from an anonymous page, then
2976 	 * we are one step closer to freeing up an anon slot.
2977 	 *
2978 	 * NOTE:  The original anon slot must be released while
2979 	 * holding the "anon_map" lock.  This is necessary to prevent
2980 	 * other threads from obtaining a pointer to the anon slot
2981 	 * which may be freed if its "refcnt" is 1.
2982 	 */
2983 	if (oldap != NULL)
2984 		anon_decref(oldap);
2985 
2986 	(void) anon_set_ptr(amp->ahp, anon_index, ap, ANON_SLEEP);
2987 
2988 	/*
2989 	 * Handle pages that have been marked for migration
2990 	 */
2991 	if (lgrp_optimizations())
2992 		page_migrate(seg, addr, &pp, 1);
2993 
2994 	ASSERT(pp->p_szc == 0);
2995 	if (type == F_SOFTLOCK && svd->vp == NULL) {
2996 		if (!segvn_slock_anonpages(pp, first)) {
2997 			page_unlock(pp);
2998 			err = ENOMEM;
2999 			goto out;
3000 		} else {
3001 			mutex_enter(&freemem_lock);
3002 			svd->softlockcnt++;
3003 			segvn_pages_locked++;
3004 			mutex_exit(&freemem_lock);
3005 		}
3006 	}
3007 
3008 	ASSERT(!IS_VMODSORT(pp->p_vnode));
3009 	if (enable_mbit_wa) {
3010 		if (rw == S_WRITE)
3011 			hat_setmod(pp);
3012 		else if (!hat_ismod(pp))
3013 			prot &= ~PROT_WRITE;
3014 	}
3015 
3016 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
3017 	hat_memload(hat, addr, pp, prot, hat_flag);
3018 
3019 	if (!(hat_flag & HAT_LOAD_LOCK))
3020 		page_unlock(pp);
3021 
3022 	ASSERT(anon_lock);
3023 	anon_array_exit(&cookie);
3024 	return (0);
3025 out:
3026 	if (anon_lock)
3027 		anon_array_exit(&cookie);
3028 
3029 	if (type == F_SOFTLOCK && svd->vp != NULL) {
3030 		mutex_enter(&freemem_lock);
3031 		availrmem++;
3032 		segvn_pages_locked--;
3033 		svd->softlockcnt--;
3034 		mutex_exit(&freemem_lock);
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 sucess and 0 on failure.
3255  *
3256  * If physically contiguos 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) != 0) {
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);
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) != 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);
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(type, pages) 		\
3691 		if ((type) == F_SOFTLOCK) {		\
3692 			mutex_enter(&freemem_lock);	\
3693 			availrmem += (pages);		\
3694 			segvn_pages_locked -= (pages);	\
3695 			svd->softlockcnt -= (pages);	\
3696 			mutex_exit(&freemem_lock);	\
3697 		}
3698 
3699 #define	SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot)		\
3700 		if (IS_VMODSORT((ppa)[0]->p_vnode)) {			\
3701 			if ((rw) == S_WRITE) {				\
3702 				for (i = 0; i < (pages); i++) {		\
3703 					ASSERT((ppa)[i]->p_vnode ==	\
3704 					    (ppa)[0]->p_vnode);		\
3705 					hat_setmod((ppa)[i]);		\
3706 				}					\
3707 			} else if ((rw) != S_OTHER &&			\
3708 			    ((prot) & (vpprot) & PROT_WRITE)) {		\
3709 				for (i = 0; i < (pages); i++) {		\
3710 					ASSERT((ppa)[i]->p_vnode ==	\
3711 					    (ppa)[0]->p_vnode);		\
3712 					if (!hat_ismod((ppa)[i])) {	\
3713 						prot &= ~PROT_WRITE;	\
3714 						break;			\
3715 					}				\
3716 				}					\
3717 			}						\
3718 		}
3719 
3720 #ifdef  VM_STATS
3721 
3722 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)					\
3723 		VM_STAT_ADD(segvnvmstats.fltvnpages[(idx)]);
3724 
3725 #else /* VM_STATS */
3726 
3727 #define	SEGVN_VMSTAT_FLTVNPAGES(idx)
3728 
3729 #endif
3730 
3731 static faultcode_t
3732 segvn_fault_vnodepages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
3733     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
3734     caddr_t eaddr, int brkcow)
3735 {
3736 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
3737 	struct anon_map *amp = svd->amp;
3738 	uchar_t segtype = svd->type;
3739 	uint_t szc = seg->s_szc;
3740 	size_t pgsz = page_get_pagesize(szc);
3741 	size_t maxpgsz = pgsz;
3742 	pgcnt_t pages = btop(pgsz);
3743 	pgcnt_t maxpages = pages;
3744 	size_t ppasize = (pages + 1) * sizeof (page_t *);
3745 	caddr_t a = lpgaddr;
3746 	caddr_t	maxlpgeaddr = lpgeaddr;
3747 	u_offset_t off = svd->offset + (uintptr_t)(a - seg->s_base);
3748 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
3749 	struct vpage *vpage = (svd->vpage != NULL) ?
3750 	    &svd->vpage[seg_page(seg, a)] : NULL;
3751 	vnode_t *vp = svd->vp;
3752 	page_t **ppa;
3753 	uint_t	pszc;
3754 	size_t	ppgsz;
3755 	pgcnt_t	ppages;
3756 	faultcode_t err = 0;
3757 	int ierr;
3758 	int vop_size_err = 0;
3759 	uint_t protchk, prot, vpprot;
3760 	ulong_t i;
3761 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
3762 	anon_sync_obj_t an_cookie;
3763 	enum seg_rw arw;
3764 	int alloc_failed = 0;
3765 	int adjszc_chk;
3766 	struct vattr va;
3767 	int xhat = 0;
3768 	page_t *pplist;
3769 	pfn_t pfn;
3770 	int physcontig;
3771 	int upgrdfail;
3772 	int segvn_anypgsz_vnode = 0; /* for now map vnode with 2 page sizes */
3773 	int tron = (svd->tr_state == SEGVN_TR_ON);
3774 
3775 	ASSERT(szc != 0);
3776 	ASSERT(vp != NULL);
3777 	ASSERT(brkcow == 0 || amp != NULL);
3778 	ASSERT(tron == 0 || amp != NULL);
3779 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
3780 	ASSERT(!(svd->flags & MAP_NORESERVE));
3781 	ASSERT(type != F_SOFTUNLOCK);
3782 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
3783 	ASSERT(amp == NULL || IS_P2ALIGNED(aindx, maxpages));
3784 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
3785 	ASSERT(seg->s_szc < NBBY * sizeof (int));
3786 	ASSERT(type != F_SOFTLOCK || lpgeaddr - a == maxpgsz);
3787 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
3788 
3789 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltvnpages[0]);
3790 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltvnpages[1]);
3791 
3792 	if (svd->flags & MAP_TEXT) {
3793 		hat_flag |= HAT_LOAD_TEXT;
3794 	}
3795 
3796 	if (svd->pageprot) {
3797 		switch (rw) {
3798 		case S_READ:
3799 			protchk = PROT_READ;
3800 			break;
3801 		case S_WRITE:
3802 			protchk = PROT_WRITE;
3803 			break;
3804 		case S_EXEC:
3805 			protchk = PROT_EXEC;
3806 			break;
3807 		case S_OTHER:
3808 		default:
3809 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
3810 			break;
3811 		}
3812 	} else {
3813 		prot = svd->prot;
3814 		/* caller has already done segment level protection check. */
3815 	}
3816 
3817 	if (seg->s_as->a_hat != hat) {
3818 		xhat = 1;
3819 	}
3820 
3821 	if (rw == S_WRITE && segtype == MAP_PRIVATE) {
3822 		SEGVN_VMSTAT_FLTVNPAGES(2);
3823 		arw = S_READ;
3824 	} else {
3825 		arw = rw;
3826 	}
3827 
3828 	ppa = kmem_alloc(ppasize, KM_SLEEP);
3829 
3830 	VM_STAT_COND_ADD(amp != NULL, segvnvmstats.fltvnpages[3]);
3831 
3832 	for (;;) {
3833 		adjszc_chk = 0;
3834 		for (; a < lpgeaddr; a += pgsz, off += pgsz, aindx += pages) {
3835 			if (adjszc_chk) {
3836 				while (szc < seg->s_szc) {
3837 					uintptr_t e;
3838 					uint_t tszc;
3839 					tszc = segvn_anypgsz_vnode ? szc + 1 :
3840 					    seg->s_szc;
3841 					ppgsz = page_get_pagesize(tszc);
3842 					if (!IS_P2ALIGNED(a, ppgsz) ||
3843 					    ((alloc_failed >> tszc) &
3844 						0x1)) {
3845 						break;
3846 					}
3847 					SEGVN_VMSTAT_FLTVNPAGES(4);
3848 					szc = tszc;
3849 					pgsz = ppgsz;
3850 					pages = btop(pgsz);
3851 					e = P2ROUNDUP((uintptr_t)eaddr, pgsz);
3852 					lpgeaddr = (caddr_t)e;
3853 				}
3854 			}
3855 
3856 		again:
3857 			if (IS_P2ALIGNED(a, maxpgsz) && amp != NULL) {
3858 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
3859 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
3860 				anon_array_enter(amp, aindx, &an_cookie);
3861 				if (anon_get_ptr(amp->ahp, aindx) != NULL) {
3862 					SEGVN_VMSTAT_FLTVNPAGES(5);
3863 					ASSERT(anon_pages(amp->ahp, aindx,
3864 					    maxpages) == maxpages);
3865 					anon_array_exit(&an_cookie);
3866 					ANON_LOCK_EXIT(&amp->a_rwlock);
3867 					err = segvn_fault_anonpages(hat, seg,
3868 					    a, a + maxpgsz, type, rw,
3869 					    MAX(a, addr),
3870 					    MIN(a + maxpgsz, eaddr), brkcow);
3871 					if (err != 0) {
3872 						SEGVN_VMSTAT_FLTVNPAGES(6);
3873 						goto out;
3874 					}
3875 					if (szc < seg->s_szc) {
3876 						szc = seg->s_szc;
3877 						pgsz = maxpgsz;
3878 						pages = maxpages;
3879 						lpgeaddr = maxlpgeaddr;
3880 					}
3881 					goto next;
3882 				} else {
3883 					ASSERT(anon_pages(amp->ahp, aindx,
3884 					    maxpages) == 0);
3885 					SEGVN_VMSTAT_FLTVNPAGES(7);
3886 					anon_array_exit(&an_cookie);
3887 					ANON_LOCK_EXIT(&amp->a_rwlock);
3888 				}
3889 			}
3890 			ASSERT(!brkcow || IS_P2ALIGNED(a, maxpgsz));
3891 			ASSERT(!tron || IS_P2ALIGNED(a, maxpgsz));
3892 
3893 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
3894 				ASSERT(vpage != NULL);
3895 				prot = VPP_PROT(vpage);
3896 				ASSERT(sameprot(seg, a, maxpgsz));
3897 				if ((prot & protchk) == 0) {
3898 					SEGVN_VMSTAT_FLTVNPAGES(8);
3899 					err = FC_PROT;
3900 					goto out;
3901 				}
3902 			}
3903 			if (type == F_SOFTLOCK) {
3904 				mutex_enter(&freemem_lock);
3905 				if (availrmem < tune.t_minarmem + pages) {
3906 					mutex_exit(&freemem_lock);
3907 					err = FC_MAKE_ERR(ENOMEM);
3908 					goto out;
3909 				} else {
3910 					availrmem -= pages;
3911 					segvn_pages_locked += pages;
3912 					svd->softlockcnt += pages;
3913 				}
3914 				mutex_exit(&freemem_lock);
3915 			}
3916 
3917 			pplist = NULL;
3918 			physcontig = 0;
3919 			ppa[0] = NULL;
3920 			if (!brkcow && !tron && szc &&
3921 			    !page_exists_physcontig(vp, off, szc,
3922 				segtype == MAP_PRIVATE ? ppa : NULL)) {
3923 				SEGVN_VMSTAT_FLTVNPAGES(9);
3924 				if (page_alloc_pages(vp, seg, a, &pplist, NULL,
3925 				    szc, 0, 0) && type != F_SOFTLOCK) {
3926 					SEGVN_VMSTAT_FLTVNPAGES(10);
3927 					pszc = 0;
3928 					ierr = -1;
3929 					alloc_failed |= (1 << szc);
3930 					break;
3931 				}
3932 				if (pplist != NULL &&
3933 				    vp->v_mpssdata == SEGVN_PAGEIO) {
3934 					int downsize;
3935 					SEGVN_VMSTAT_FLTVNPAGES(11);
3936 					physcontig = segvn_fill_vp_pages(svd,
3937 					    vp, off, szc, ppa, &pplist,
3938 					    &pszc, &downsize);
3939 					ASSERT(!physcontig || pplist == NULL);
3940 					if (!physcontig && downsize &&
3941 					    type != F_SOFTLOCK) {
3942 						ASSERT(pplist == NULL);
3943 						SEGVN_VMSTAT_FLTVNPAGES(12);
3944 						ierr = -1;
3945 						break;
3946 					}
3947 					ASSERT(!physcontig ||
3948 					    segtype == MAP_PRIVATE ||
3949 					    ppa[0] == NULL);
3950 					if (physcontig && ppa[0] == NULL) {
3951 						physcontig = 0;
3952 					}
3953 				}
3954 			} else if (!brkcow && !tron && szc && ppa[0] != NULL) {
3955 				SEGVN_VMSTAT_FLTVNPAGES(13);
3956 				ASSERT(segtype == MAP_PRIVATE);
3957 				physcontig = 1;
3958 			}
3959 
3960 			if (!physcontig) {
3961 				SEGVN_VMSTAT_FLTVNPAGES(14);
3962 				ppa[0] = NULL;
3963 				ierr = VOP_GETPAGE(vp, (offset_t)off, pgsz,
3964 				    &vpprot, ppa, pgsz, seg, a, arw,
3965 				    svd->cred);
3966 #ifdef DEBUG
3967 				if (ierr == 0) {
3968 					for (i = 0; i < pages; i++) {
3969 						ASSERT(PAGE_LOCKED(ppa[i]));
3970 						ASSERT(!PP_ISFREE(ppa[i]));
3971 						ASSERT(ppa[i]->p_vnode == vp);
3972 						ASSERT(ppa[i]->p_offset ==
3973 						    off + (i << PAGESHIFT));
3974 					}
3975 				}
3976 #endif /* DEBUG */
3977 				if (segtype == MAP_PRIVATE) {
3978 					SEGVN_VMSTAT_FLTVNPAGES(15);
3979 					vpprot &= ~PROT_WRITE;
3980 				}
3981 			} else {
3982 				ASSERT(segtype == MAP_PRIVATE);
3983 				SEGVN_VMSTAT_FLTVNPAGES(16);
3984 				vpprot = PROT_ALL & ~PROT_WRITE;
3985 				ierr = 0;
3986 			}
3987 
3988 			if (ierr != 0) {
3989 				SEGVN_VMSTAT_FLTVNPAGES(17);
3990 				if (pplist != NULL) {
3991 					SEGVN_VMSTAT_FLTVNPAGES(18);
3992 					page_free_replacement_page(pplist);
3993 					page_create_putback(pages);
3994 				}
3995 				SEGVN_RESTORE_SOFTLOCK(type, pages);
3996 				if (a + pgsz <= eaddr) {
3997 					SEGVN_VMSTAT_FLTVNPAGES(19);
3998 					err = FC_MAKE_ERR(ierr);
3999 					goto out;
4000 				}
4001 				va.va_mask = AT_SIZE;
4002 				if (VOP_GETATTR(vp, &va, 0, svd->cred) != 0) {
4003 					SEGVN_VMSTAT_FLTVNPAGES(20);
4004 					err = FC_MAKE_ERR(EIO);
4005 					goto out;
4006 				}
4007 				if (btopr(va.va_size) >= btopr(off + pgsz)) {
4008 					SEGVN_VMSTAT_FLTVNPAGES(21);
4009 					err = FC_MAKE_ERR(ierr);
4010 					goto out;
4011 				}
4012 				if (btopr(va.va_size) <
4013 				    btopr(off + (eaddr - a))) {
4014 					SEGVN_VMSTAT_FLTVNPAGES(22);
4015 					err = FC_MAKE_ERR(ierr);
4016 					goto out;
4017 				}
4018 				if (brkcow || tron || type == F_SOFTLOCK) {
4019 					/* can't reduce map area */
4020 					SEGVN_VMSTAT_FLTVNPAGES(23);
4021 					vop_size_err = 1;
4022 					goto out;
4023 				}
4024 				SEGVN_VMSTAT_FLTVNPAGES(24);
4025 				ASSERT(szc != 0);
4026 				pszc = 0;
4027 				ierr = -1;
4028 				break;
4029 			}
4030 
4031 			if (amp != NULL) {
4032 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4033 				anon_array_enter(amp, aindx, &an_cookie);
4034 			}
4035 			if (amp != NULL &&
4036 			    anon_get_ptr(amp->ahp, aindx) != NULL) {
4037 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4038 
4039 				SEGVN_VMSTAT_FLTVNPAGES(25);
4040 				ASSERT(anon_pages(amp->ahp, taindx,
4041 				    maxpages) == maxpages);
4042 				for (i = 0; i < pages; i++) {
4043 					page_unlock(ppa[i]);
4044 				}
4045 				anon_array_exit(&an_cookie);
4046 				ANON_LOCK_EXIT(&amp->a_rwlock);
4047 				if (pplist != NULL) {
4048 					page_free_replacement_page(pplist);
4049 					page_create_putback(pages);
4050 				}
4051 				SEGVN_RESTORE_SOFTLOCK(type, pages);
4052 				if (szc < seg->s_szc) {
4053 					SEGVN_VMSTAT_FLTVNPAGES(26);
4054 					/*
4055 					 * For private segments SOFTLOCK
4056 					 * either always breaks cow (any rw
4057 					 * type except S_READ_NOCOW) or
4058 					 * address space is locked as writer
4059 					 * (S_READ_NOCOW case) and anon slots
4060 					 * can't show up on second check.
4061 					 * Therefore if we are here for
4062 					 * SOFTLOCK case it must be a cow
4063 					 * break but cow break never reduces
4064 					 * szc. text replication (tron) in
4065 					 * this case works as cow break.
4066 					 * Thus the assert below.
4067 					 */
4068 					ASSERT(!brkcow && !tron &&
4069 					    type != F_SOFTLOCK);
4070 					pszc = seg->s_szc;
4071 					ierr = -2;
4072 					break;
4073 				}
4074 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4075 				goto again;
4076 			}
4077 #ifdef DEBUG
4078 			if (amp != NULL) {
4079 				ulong_t taindx = P2ALIGN(aindx, maxpages);
4080 				ASSERT(!anon_pages(amp->ahp, taindx, maxpages));
4081 			}
4082 #endif /* DEBUG */
4083 
4084 			if (brkcow || tron) {
4085 				ASSERT(amp != NULL);
4086 				ASSERT(pplist == NULL);
4087 				ASSERT(szc == seg->s_szc);
4088 				ASSERT(IS_P2ALIGNED(a, maxpgsz));
4089 				ASSERT(IS_P2ALIGNED(aindx, maxpages));
4090 				SEGVN_VMSTAT_FLTVNPAGES(27);
4091 				ierr = anon_map_privatepages(amp, aindx, szc,
4092 				    seg, a, prot, ppa, vpage, segvn_anypgsz,
4093 				    tron ? PG_LOCAL : 0, svd->cred);
4094 				if (ierr != 0) {
4095 					SEGVN_VMSTAT_FLTVNPAGES(28);
4096 					anon_array_exit(&an_cookie);
4097 					ANON_LOCK_EXIT(&amp->a_rwlock);
4098 					SEGVN_RESTORE_SOFTLOCK(type, pages);
4099 					err = FC_MAKE_ERR(ierr);
4100 					goto out;
4101 				}
4102 
4103 				ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4104 				/*
4105 				 * p_szc can't be changed for locked
4106 				 * swapfs pages.
4107 				 */
4108 				ASSERT(svd->rcookie ==
4109 				    HAT_INVALID_REGION_COOKIE);
4110 				hat_memload_array(hat, a, pgsz, ppa, prot,
4111 				    hat_flag);
4112 
4113 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4114 					SEGVN_VMSTAT_FLTVNPAGES(29);
4115 					for (i = 0; i < pages; i++) {
4116 						page_unlock(ppa[i]);
4117 					}
4118 				}
4119 				anon_array_exit(&an_cookie);
4120 				ANON_LOCK_EXIT(&amp->a_rwlock);
4121 				goto next;
4122 			}
4123 
4124 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE ||
4125 			    (!svd->pageprot && svd->prot == (prot & vpprot)));
4126 
4127 			pfn = page_pptonum(ppa[0]);
4128 			/*
4129 			 * hat_page_demote() needs an SE_EXCL lock on one of
4130 			 * constituent page_t's and it decreases root's p_szc
4131 			 * last. This means if root's p_szc is equal szc and
4132 			 * all its constituent pages are locked
4133 			 * hat_page_demote() that could have changed p_szc to
4134 			 * szc is already done and no new have page_demote()
4135 			 * can start for this large page.
4136 			 */
4137 
4138 			/*
4139 			 * we need to make sure same mapping size is used for
4140 			 * the same address range if there's a possibility the
4141 			 * adddress is already mapped because hat layer panics
4142 			 * when translation is loaded for the range already
4143 			 * mapped with a different page size.  We achieve it
4144 			 * by always using largest page size possible subject
4145 			 * to the constraints of page size, segment page size
4146 			 * and page alignment.  Since mappings are invalidated
4147 			 * when those constraints change and make it
4148 			 * impossible to use previously used mapping size no
4149 			 * mapping size conflicts should happen.
4150 			 */
4151 
4152 		chkszc:
4153 			if ((pszc = ppa[0]->p_szc) == szc &&
4154 			    IS_P2ALIGNED(pfn, pages)) {
4155 
4156 				SEGVN_VMSTAT_FLTVNPAGES(30);
4157 #ifdef DEBUG
4158 				for (i = 0; i < pages; i++) {
4159 					ASSERT(PAGE_LOCKED(ppa[i]));
4160 					ASSERT(!PP_ISFREE(ppa[i]));
4161 					ASSERT(page_pptonum(ppa[i]) ==
4162 					    pfn + i);
4163 					ASSERT(ppa[i]->p_szc == szc);
4164 					ASSERT(ppa[i]->p_vnode == vp);
4165 					ASSERT(ppa[i]->p_offset ==
4166 					    off + (i << PAGESHIFT));
4167 				}
4168 #endif /* DEBUG */
4169 				/*
4170 				 * All pages are of szc we need and they are
4171 				 * all locked so they can't change szc. load
4172 				 * translations.
4173 				 *
4174 				 * if page got promoted since last check
4175 				 * we don't need pplist.
4176 				 */
4177 				if (pplist != NULL) {
4178 					page_free_replacement_page(pplist);
4179 					page_create_putback(pages);
4180 				}
4181 				if (PP_ISMIGRATE(ppa[0])) {
4182 					page_migrate(seg, a, ppa, pages);
4183 				}
4184 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4185 				    prot, vpprot);
4186 				if (!xhat) {
4187 					hat_memload_array_region(hat, a, pgsz,
4188 					    ppa, prot & vpprot, hat_flag,
4189 					    svd->rcookie);
4190 				} else {
4191 					/*
4192 					 * avoid large xhat mappings to FS
4193 					 * pages so that hat_page_demote()
4194 					 * doesn't need to check for xhat
4195 					 * large mappings.
4196 					 * Don't use regions with xhats.
4197 					 */
4198 					for (i = 0; i < pages; i++) {
4199 						hat_memload(hat,
4200 						    a + (i << PAGESHIFT),
4201 						    ppa[i], prot & vpprot,
4202 						    hat_flag);
4203 					}
4204 				}
4205 
4206 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4207 					for (i = 0; i < pages; i++) {
4208 						page_unlock(ppa[i]);
4209 					}
4210 				}
4211 				if (amp != NULL) {
4212 					anon_array_exit(&an_cookie);
4213 					ANON_LOCK_EXIT(&amp->a_rwlock);
4214 				}
4215 				goto next;
4216 			}
4217 
4218 			/*
4219 			 * See if upsize is possible.
4220 			 */
4221 			if (pszc > szc && szc < seg->s_szc &&
4222 			    (segvn_anypgsz_vnode || pszc >= seg->s_szc)) {
4223 				pgcnt_t aphase;
4224 				uint_t pszc1 = MIN(pszc, seg->s_szc);
4225 				ppgsz = page_get_pagesize(pszc1);
4226 				ppages = btop(ppgsz);
4227 				aphase = btop(P2PHASE((uintptr_t)a, ppgsz));
4228 
4229 				ASSERT(type != F_SOFTLOCK);
4230 
4231 				SEGVN_VMSTAT_FLTVNPAGES(31);
4232 				if (aphase != P2PHASE(pfn, ppages)) {
4233 					segvn_faultvnmpss_align_err4++;
4234 				} else {
4235 					SEGVN_VMSTAT_FLTVNPAGES(32);
4236 					if (pplist != NULL) {
4237 						page_t *pl = pplist;
4238 						page_free_replacement_page(pl);
4239 						page_create_putback(pages);
4240 					}
4241 					for (i = 0; i < pages; i++) {
4242 						page_unlock(ppa[i]);
4243 					}
4244 					if (amp != NULL) {
4245 						anon_array_exit(&an_cookie);
4246 						ANON_LOCK_EXIT(&amp->a_rwlock);
4247 					}
4248 					pszc = pszc1;
4249 					ierr = -2;
4250 					break;
4251 				}
4252 			}
4253 
4254 			/*
4255 			 * check if we should use smallest mapping size.
4256 			 */
4257 			upgrdfail = 0;
4258 			if (szc == 0 || xhat ||
4259 			    (pszc >= szc &&
4260 			    !IS_P2ALIGNED(pfn, pages)) ||
4261 			    (pszc < szc &&
4262 			    !segvn_full_szcpages(ppa, szc, &upgrdfail,
4263 				&pszc))) {
4264 
4265 				if (upgrdfail && type != F_SOFTLOCK) {
4266 					/*
4267 					 * segvn_full_szcpages failed to lock
4268 					 * all pages EXCL. Size down.
4269 					 */
4270 					ASSERT(pszc < szc);
4271 
4272 					SEGVN_VMSTAT_FLTVNPAGES(33);
4273 
4274 					if (pplist != NULL) {
4275 						page_t *pl = pplist;
4276 						page_free_replacement_page(pl);
4277 						page_create_putback(pages);
4278 					}
4279 
4280 					for (i = 0; i < pages; i++) {
4281 						page_unlock(ppa[i]);
4282 					}
4283 					if (amp != NULL) {
4284 						anon_array_exit(&an_cookie);
4285 						ANON_LOCK_EXIT(&amp->a_rwlock);
4286 					}
4287 					ierr = -1;
4288 					break;
4289 				}
4290 				if (szc != 0 && !xhat && !upgrdfail) {
4291 					segvn_faultvnmpss_align_err5++;
4292 				}
4293 				SEGVN_VMSTAT_FLTVNPAGES(34);
4294 				if (pplist != NULL) {
4295 					page_free_replacement_page(pplist);
4296 					page_create_putback(pages);
4297 				}
4298 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4299 				    prot, vpprot);
4300 				if (upgrdfail && segvn_anypgsz_vnode) {
4301 					/* SOFTLOCK case */
4302 					hat_memload_array_region(hat, a, pgsz,
4303 					    ppa, prot & vpprot, hat_flag,
4304 					    svd->rcookie);
4305 				} else {
4306 					for (i = 0; i < pages; i++) {
4307 						hat_memload_region(hat,
4308 						    a + (i << PAGESHIFT),
4309 						    ppa[i], prot & vpprot,
4310 						    hat_flag, svd->rcookie);
4311 					}
4312 				}
4313 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4314 					for (i = 0; i < pages; i++) {
4315 						page_unlock(ppa[i]);
4316 					}
4317 				}
4318 				if (amp != NULL) {
4319 					anon_array_exit(&an_cookie);
4320 					ANON_LOCK_EXIT(&amp->a_rwlock);
4321 				}
4322 				goto next;
4323 			}
4324 
4325 			if (pszc == szc) {
4326 				/*
4327 				 * segvn_full_szcpages() upgraded pages szc.
4328 				 */
4329 				ASSERT(pszc == ppa[0]->p_szc);
4330 				ASSERT(IS_P2ALIGNED(pfn, pages));
4331 				goto chkszc;
4332 			}
4333 
4334 			if (pszc > szc) {
4335 				kmutex_t *szcmtx;
4336 				SEGVN_VMSTAT_FLTVNPAGES(35);
4337 				/*
4338 				 * p_szc of ppa[0] can change since we haven't
4339 				 * locked all constituent pages. Call
4340 				 * page_lock_szc() to prevent szc changes.
4341 				 * This should be a rare case that happens when
4342 				 * multiple segments use a different page size
4343 				 * to map the same file offsets.
4344 				 */
4345 				szcmtx = page_szc_lock(ppa[0]);
4346 				pszc = ppa[0]->p_szc;
4347 				ASSERT(szcmtx != NULL || pszc == 0);
4348 				ASSERT(ppa[0]->p_szc <= pszc);
4349 				if (pszc <= szc) {
4350 					SEGVN_VMSTAT_FLTVNPAGES(36);
4351 					if (szcmtx != NULL) {
4352 						mutex_exit(szcmtx);
4353 					}
4354 					goto chkszc;
4355 				}
4356 				if (pplist != NULL) {
4357 					/*
4358 					 * page got promoted since last check.
4359 					 * we don't need preaalocated large
4360 					 * page.
4361 					 */
4362 					SEGVN_VMSTAT_FLTVNPAGES(37);
4363 					page_free_replacement_page(pplist);
4364 					page_create_putback(pages);
4365 				}
4366 				SEGVN_UPDATE_MODBITS(ppa, pages, rw,
4367 				    prot, vpprot);
4368 				hat_memload_array_region(hat, a, pgsz, ppa,
4369 				    prot & vpprot, hat_flag, svd->rcookie);
4370 				mutex_exit(szcmtx);
4371 				if (!(hat_flag & HAT_LOAD_LOCK)) {
4372 					for (i = 0; i < pages; i++) {
4373 						page_unlock(ppa[i]);
4374 					}
4375 				}
4376 				if (amp != NULL) {
4377 					anon_array_exit(&an_cookie);
4378 					ANON_LOCK_EXIT(&amp->a_rwlock);
4379 				}
4380 				goto next;
4381 			}
4382 
4383 			/*
4384 			 * if page got demoted since last check
4385 			 * we could have not allocated larger page.
4386 			 * allocate now.
4387 			 */
4388 			if (pplist == NULL &&
4389 			    page_alloc_pages(vp, seg, a, &pplist, NULL,
4390 				szc, 0, 0) && type != F_SOFTLOCK) {
4391 				SEGVN_VMSTAT_FLTVNPAGES(38);
4392 				for (i = 0; i < pages; i++) {
4393 					page_unlock(ppa[i]);
4394 				}
4395 				if (amp != NULL) {
4396 					anon_array_exit(&an_cookie);
4397 					ANON_LOCK_EXIT(&amp->a_rwlock);
4398 				}
4399 				ierr = -1;
4400 				alloc_failed |= (1 << szc);
4401 				break;
4402 			}
4403 
4404 			SEGVN_VMSTAT_FLTVNPAGES(39);
4405 
4406 			if (pplist != NULL) {
4407 				segvn_relocate_pages(ppa, pplist);
4408 #ifdef DEBUG
4409 			} else {
4410 				ASSERT(type == F_SOFTLOCK);
4411 				SEGVN_VMSTAT_FLTVNPAGES(40);
4412 #endif /* DEBUG */
4413 			}
4414 
4415 			SEGVN_UPDATE_MODBITS(ppa, pages, rw, prot, vpprot);
4416 
4417 			if (pplist == NULL && segvn_anypgsz_vnode == 0) {
4418 				ASSERT(type == F_SOFTLOCK);
4419 				for (i = 0; i < pages; i++) {
4420 					ASSERT(ppa[i]->p_szc < szc);
4421 					hat_memload_region(hat,
4422 					    a + (i << PAGESHIFT),
4423 					    ppa[i], prot & vpprot, hat_flag,
4424 					    svd->rcookie);
4425 				}
4426 			} else {
4427 				ASSERT(pplist != NULL || type == F_SOFTLOCK);
4428 				hat_memload_array_region(hat, a, pgsz, ppa,
4429 				    prot & vpprot, hat_flag, svd->rcookie);
4430 			}
4431 			if (!(hat_flag & HAT_LOAD_LOCK)) {
4432 				for (i = 0; i < pages; i++) {
4433 					ASSERT(PAGE_SHARED(ppa[i]));
4434 					page_unlock(ppa[i]);
4435 				}
4436 			}
4437 			if (amp != NULL) {
4438 				anon_array_exit(&an_cookie);
4439 				ANON_LOCK_EXIT(&amp->a_rwlock);
4440 			}
4441 
4442 		next:
4443 			if (vpage != NULL) {
4444 				vpage += pages;
4445 			}
4446 			adjszc_chk = 1;
4447 		}
4448 		if (a == lpgeaddr)
4449 			break;
4450 		ASSERT(a < lpgeaddr);
4451 
4452 		ASSERT(!brkcow && !tron && type != F_SOFTLOCK);
4453 
4454 		/*
4455 		 * ierr == -1 means we failed to map with a large page.
4456 		 * (either due to allocation/relocation failures or
4457 		 * misalignment with other mappings to this file.
4458 		 *
4459 		 * ierr == -2 means some other thread allocated a large page
4460 		 * after we gave up tp map with a large page.  retry with
4461 		 * larger mapping.
4462 		 */
4463 		ASSERT(ierr == -1 || ierr == -2);
4464 		ASSERT(ierr == -2 || szc != 0);
4465 		ASSERT(ierr == -1 || szc < seg->s_szc);
4466 		if (ierr == -2) {
4467 			SEGVN_VMSTAT_FLTVNPAGES(41);
4468 			ASSERT(pszc > szc && pszc <= seg->s_szc);
4469 			szc = pszc;
4470 		} else if (segvn_anypgsz_vnode) {
4471 			SEGVN_VMSTAT_FLTVNPAGES(42);
4472 			szc--;
4473 		} else {
4474 			SEGVN_VMSTAT_FLTVNPAGES(43);
4475 			ASSERT(pszc < szc);
4476 			/*
4477 			 * other process created pszc large page.
4478 			 * but we still have to drop to 0 szc.
4479 			 */
4480 			szc = 0;
4481 		}
4482 
4483 		pgsz = page_get_pagesize(szc);
4484 		pages = btop(pgsz);
4485 		if (ierr == -2) {
4486 			/*
4487 			 * Size up case. Note lpgaddr may only be needed for
4488 			 * softlock case so we don't adjust it here.
4489 			 */
4490 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4491 			ASSERT(a >= lpgaddr);
4492 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4493 			off = svd->offset + (uintptr_t)(a - seg->s_base);
4494 			aindx = svd->anon_index + seg_page(seg, a);
4495 			vpage = (svd->vpage != NULL) ?
4496 			    &svd->vpage[seg_page(seg, a)] : NULL;
4497 		} else {
4498 			/*
4499 			 * Size down case. Note lpgaddr may only be needed for
4500 			 * softlock case so we don't adjust it here.
4501 			 */
4502 			ASSERT(IS_P2ALIGNED(a, pgsz));
4503 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4504 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4505 			ASSERT(a < lpgeaddr);
4506 			if (a < addr) {
4507 				SEGVN_VMSTAT_FLTVNPAGES(44);
4508 				/*
4509 				 * The beginning of the large page region can
4510 				 * be pulled to the right to make a smaller
4511 				 * region. We haven't yet faulted a single
4512 				 * page.
4513 				 */
4514 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4515 				ASSERT(a >= lpgaddr);
4516 				off = svd->offset +
4517 				    (uintptr_t)(a - seg->s_base);
4518 				aindx = svd->anon_index + seg_page(seg, a);
4519 				vpage = (svd->vpage != NULL) ?
4520 				    &svd->vpage[seg_page(seg, a)] : NULL;
4521 			}
4522 		}
4523 	}
4524 out:
4525 	kmem_free(ppa, ppasize);
4526 	if (!err && !vop_size_err) {
4527 		SEGVN_VMSTAT_FLTVNPAGES(45);
4528 		return (0);
4529 	}
4530 	if (type == F_SOFTLOCK && a > lpgaddr) {
4531 		SEGVN_VMSTAT_FLTVNPAGES(46);
4532 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4533 	}
4534 	if (!vop_size_err) {
4535 		SEGVN_VMSTAT_FLTVNPAGES(47);
4536 		return (err);
4537 	}
4538 	ASSERT(brkcow || tron || type == F_SOFTLOCK);
4539 	/*
4540 	 * Large page end is mapped beyond the end of file and it's a cow
4541 	 * fault (can be a text replication induced cow) or softlock so we can't
4542 	 * reduce the map area.  For now just demote the segment. This should
4543 	 * really only happen if the end of the file changed after the mapping
4544 	 * was established since when large page segments are created we make
4545 	 * sure they don't extend beyond the end of the file.
4546 	 */
4547 	SEGVN_VMSTAT_FLTVNPAGES(48);
4548 
4549 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4550 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4551 	err = 0;
4552 	if (seg->s_szc != 0) {
4553 		segvn_fltvnpages_clrszc_cnt++;
4554 		ASSERT(svd->softlockcnt == 0);
4555 		err = segvn_clrszc(seg);
4556 		if (err != 0) {
4557 			segvn_fltvnpages_clrszc_err++;
4558 		}
4559 	}
4560 	ASSERT(err || seg->s_szc == 0);
4561 	SEGVN_LOCK_DOWNGRADE(seg->s_as, &svd->lock);
4562 	/* segvn_fault will do its job as if szc had been zero to begin with */
4563 	return (err == 0 ? IE_RETRY : FC_MAKE_ERR(err));
4564 }
4565 
4566 /*
4567  * This routine will attempt to fault in one large page.
4568  * it will use smaller pages if that fails.
4569  * It should only be called for pure anonymous segments.
4570  */
4571 static faultcode_t
4572 segvn_fault_anonpages(struct hat *hat, struct seg *seg, caddr_t lpgaddr,
4573     caddr_t lpgeaddr, enum fault_type type, enum seg_rw rw, caddr_t addr,
4574     caddr_t eaddr, int brkcow)
4575 {
4576 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4577 	struct anon_map *amp = svd->amp;
4578 	uchar_t segtype = svd->type;
4579 	uint_t szc = seg->s_szc;
4580 	size_t pgsz = page_get_pagesize(szc);
4581 	size_t maxpgsz = pgsz;
4582 	pgcnt_t pages = btop(pgsz);
4583 	size_t ppasize = pages * sizeof (page_t *);
4584 	caddr_t a = lpgaddr;
4585 	ulong_t aindx = svd->anon_index + seg_page(seg, a);
4586 	struct vpage *vpage = (svd->vpage != NULL) ?
4587 	    &svd->vpage[seg_page(seg, a)] : NULL;
4588 	page_t **ppa;
4589 	uint_t	ppa_szc;
4590 	faultcode_t err;
4591 	int ierr;
4592 	uint_t protchk, prot, vpprot;
4593 	ulong_t i;
4594 	int hat_flag = (type == F_SOFTLOCK) ? HAT_LOAD_LOCK : HAT_LOAD;
4595 	anon_sync_obj_t cookie;
4596 	int first = 1;
4597 	int adjszc_chk;
4598 	int purged = 0;
4599 	int pgflags = (svd->tr_state == SEGVN_TR_ON) ? PG_LOCAL : 0;
4600 
4601 	ASSERT(szc != 0);
4602 	ASSERT(amp != NULL);
4603 	ASSERT(enable_mbit_wa == 0); /* no mbit simulations with large pages */
4604 	ASSERT(!(svd->flags & MAP_NORESERVE));
4605 	ASSERT(type != F_SOFTUNLOCK);
4606 	ASSERT(IS_P2ALIGNED(a, maxpgsz));
4607 	ASSERT(!brkcow || svd->tr_state == SEGVN_TR_OFF);
4608 	ASSERT(svd->tr_state != SEGVN_TR_INIT);
4609 
4610 	ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
4611 
4612 	VM_STAT_COND_ADD(type == F_SOFTLOCK, segvnvmstats.fltanpages[0]);
4613 	VM_STAT_COND_ADD(type != F_SOFTLOCK, segvnvmstats.fltanpages[1]);
4614 
4615 	if (svd->flags & MAP_TEXT) {
4616 		hat_flag |= HAT_LOAD_TEXT;
4617 	}
4618 
4619 	if (svd->pageprot) {
4620 		switch (rw) {
4621 		case S_READ:
4622 			protchk = PROT_READ;
4623 			break;
4624 		case S_WRITE:
4625 			protchk = PROT_WRITE;
4626 			break;
4627 		case S_EXEC:
4628 			protchk = PROT_EXEC;
4629 			break;
4630 		case S_OTHER:
4631 		default:
4632 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
4633 			break;
4634 		}
4635 		VM_STAT_ADD(segvnvmstats.fltanpages[2]);
4636 	} else {
4637 		prot = svd->prot;
4638 		/* caller has already done segment level protection check. */
4639 	}
4640 
4641 	ppa = kmem_alloc(ppasize, KM_SLEEP);
4642 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
4643 	for (;;) {
4644 		adjszc_chk = 0;
4645 		for (; a < lpgeaddr; a += pgsz, aindx += pages) {
4646 			if (svd->pageprot != 0 && IS_P2ALIGNED(a, maxpgsz)) {
4647 				VM_STAT_ADD(segvnvmstats.fltanpages[3]);
4648 				ASSERT(vpage != NULL);
4649 				prot = VPP_PROT(vpage);
4650 				ASSERT(sameprot(seg, a, maxpgsz));
4651 				if ((prot & protchk) == 0) {
4652 					err = FC_PROT;
4653 					goto error;
4654 				}
4655 			}
4656 			if (adjszc_chk && IS_P2ALIGNED(a, maxpgsz) &&
4657 			    pgsz < maxpgsz) {
4658 				ASSERT(a > lpgaddr);
4659 				szc = seg->s_szc;
4660 				pgsz = maxpgsz;
4661 				pages = btop(pgsz);
4662 				ASSERT(IS_P2ALIGNED(aindx, pages));
4663 				lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr,
4664 				    pgsz);
4665 			}
4666 			if (type == F_SOFTLOCK && svd->vp != NULL) {
4667 				mutex_enter(&freemem_lock);
4668 				if (availrmem < tune.t_minarmem + pages) {
4669 					mutex_exit(&freemem_lock);
4670 					err = FC_MAKE_ERR(ENOMEM);
4671 					goto error;
4672 				} else {
4673 					availrmem -= pages;
4674 					segvn_pages_locked += pages;
4675 					svd->softlockcnt += pages;
4676 				}
4677 				mutex_exit(&freemem_lock);
4678 			}
4679 			anon_array_enter(amp, aindx, &cookie);
4680 			ppa_szc = (uint_t)-1;
4681 			ierr = anon_map_getpages(amp, aindx, szc, seg, a,
4682 				prot, &vpprot, ppa, &ppa_szc, vpage, rw, brkcow,
4683 				segvn_anypgsz, pgflags, svd->cred);
4684 			if (ierr != 0) {
4685 				anon_array_exit(&cookie);
4686 				VM_STAT_ADD(segvnvmstats.fltanpages[4]);
4687 				if (type == F_SOFTLOCK && svd->vp != NULL) {
4688 					VM_STAT_ADD(segvnvmstats.fltanpages[5]);
4689 					mutex_enter(&freemem_lock);
4690 					availrmem += pages;
4691 					segvn_pages_locked -= pages;
4692 					svd->softlockcnt -= pages;
4693 					mutex_exit(&freemem_lock);
4694 				}
4695 				if (ierr > 0) {
4696 					VM_STAT_ADD(segvnvmstats.fltanpages[6]);
4697 					err = FC_MAKE_ERR(ierr);
4698 					goto error;
4699 				}
4700 				break;
4701 			}
4702 
4703 			ASSERT(!IS_VMODSORT(ppa[0]->p_vnode));
4704 
4705 			ASSERT(segtype == MAP_SHARED ||
4706 			    ppa[0]->p_szc <= szc);
4707 			ASSERT(segtype == MAP_PRIVATE ||
4708 			    ppa[0]->p_szc >= szc);
4709 
4710 			/*
4711 			 * Handle pages that have been marked for migration
4712 			 */
4713 			if (lgrp_optimizations())
4714 				page_migrate(seg, a, ppa, pages);
4715 
4716 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
4717 			if (type == F_SOFTLOCK && svd->vp == NULL) {
4718 				/*
4719 				 * If all pages in ppa array belong to the same
4720 				 * large page call segvn_slock_anonpages()
4721 				 * just for ppa[0].
4722 				 */
4723 				for (i = 0; i < pages; i++) {
4724 					if (!segvn_slock_anonpages(ppa[i],
4725 					    i == 0 && first)) {
4726 						ulong_t j;
4727 						for (j = 0; j < i; j++) {
4728 							segvn_sunlock_anonpages(
4729 								ppa[j],
4730 								j == 0 &&
4731 								first);
4732 							page_unlock(ppa[j]);
4733 						}
4734 						for (j = i; j < pages; j++) {
4735 							page_unlock(ppa[j]);
4736 						}
4737 						anon_array_exit(&cookie);
4738 						err = FC_MAKE_ERR(ENOMEM);
4739 						goto error;
4740 					}
4741 					if (i == 0 && ppa[0]->p_szc >= szc) {
4742 						ASSERT(!(page_pptonum(ppa[0]) &
4743 						    (pages - 1)));
4744 						break;
4745 					}
4746 				}
4747 				first = 0;
4748 				mutex_enter(&freemem_lock);
4749 				svd->softlockcnt += pages;
4750 				segvn_pages_locked += pages;
4751 				mutex_exit(&freemem_lock);
4752 			}
4753 
4754 			if (segtype == MAP_SHARED) {
4755 				vpprot |= PROT_WRITE;
4756 			}
4757 
4758 			hat_memload_array(hat, a, pgsz, ppa,
4759 			    prot & vpprot, hat_flag);
4760 
4761 			if (hat_flag & HAT_LOAD_LOCK) {
4762 				VM_STAT_ADD(segvnvmstats.fltanpages[7]);
4763 			} else {
4764 				VM_STAT_ADD(segvnvmstats.fltanpages[8]);
4765 				for (i = 0; i < pages; i++)
4766 					page_unlock(ppa[i]);
4767 			}
4768 			if (vpage != NULL)
4769 				vpage += pages;
4770 
4771 			anon_array_exit(&cookie);
4772 			adjszc_chk = 1;
4773 		}
4774 		if (a == lpgeaddr)
4775 			break;
4776 		ASSERT(a < lpgeaddr);
4777 		/*
4778 		 * ierr == -1 means we failed to allocate a large page.
4779 		 * so do a size down operation.
4780 		 *
4781 		 * ierr == -2 means some other process that privately shares
4782 		 * pages with this process has allocated a larger page and we
4783 		 * need to retry with larger pages. So do a size up
4784 		 * operation. This relies on the fact that large pages are
4785 		 * never partially shared i.e. if we share any constituent
4786 		 * page of a large page with another process we must share the
4787 		 * entire large page. Note this cannot happen for SOFTLOCK
4788 		 * case, unless current address (a) is at the beginning of the
4789 		 * next page size boundary because the other process couldn't
4790 		 * have relocated locked pages.
4791 		 */
4792 		ASSERT(ierr == -1 || ierr == -2);
4793 		/*
4794 		 * For the very first relocation failure try to purge this
4795 		 * segment's cache so that the relocator can obtain an
4796 		 * exclusive lock on pages we want to relocate.
4797 		 */
4798 		if (!purged && ierr == -1 && ppa_szc != (uint_t)-1 &&
4799 		    svd->softlockcnt != 0) {
4800 			purged = 1;
4801 			segvn_purge(seg);
4802 			continue;
4803 		}
4804 
4805 		if (segvn_anypgsz) {
4806 			ASSERT(ierr == -2 || szc != 0);
4807 			ASSERT(ierr == -1 || szc < seg->s_szc);
4808 			szc = (ierr == -1) ? szc - 1 : szc + 1;
4809 		} else {
4810 			/*
4811 			 * For non COW faults and segvn_anypgsz == 0
4812 			 * we need to be careful not to loop forever
4813 			 * if existing page is found with szc other
4814 			 * than 0 or seg->s_szc. This could be due
4815 			 * to page relocations on behalf of DR or
4816 			 * more likely large page creation. For this
4817 			 * case simply re-size to existing page's szc
4818 			 * if returned by anon_map_getpages().
4819 			 */
4820 			if (ppa_szc == (uint_t)-1) {
4821 				szc = (ierr == -1) ? 0 : seg->s_szc;
4822 			} else {
4823 				ASSERT(ppa_szc <= seg->s_szc);
4824 				ASSERT(ierr == -2 || ppa_szc < szc);
4825 				ASSERT(ierr == -1 || ppa_szc > szc);
4826 				szc = ppa_szc;
4827 			}
4828 		}
4829 
4830 		pgsz = page_get_pagesize(szc);
4831 		pages = btop(pgsz);
4832 		ASSERT(type != F_SOFTLOCK || ierr == -1 ||
4833 		    (IS_P2ALIGNED(a, pgsz) && IS_P2ALIGNED(lpgeaddr, pgsz)));
4834 		if (type == F_SOFTLOCK) {
4835 			/*
4836 			 * For softlocks we cannot reduce the fault area
4837 			 * (calculated based on the largest page size for this
4838 			 * segment) for size down and a is already next
4839 			 * page size aligned as assertted above for size
4840 			 * ups. Therefore just continue in case of softlock.
4841 			 */
4842 			VM_STAT_ADD(segvnvmstats.fltanpages[9]);
4843 			continue; /* keep lint happy */
4844 		} else if (ierr == -2) {
4845 
4846 			/*
4847 			 * Size up case. Note lpgaddr may only be needed for
4848 			 * softlock case so we don't adjust it here.
4849 			 */
4850 			VM_STAT_ADD(segvnvmstats.fltanpages[10]);
4851 			a = (caddr_t)P2ALIGN((uintptr_t)a, pgsz);
4852 			ASSERT(a >= lpgaddr);
4853 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4854 			aindx = svd->anon_index + seg_page(seg, a);
4855 			vpage = (svd->vpage != NULL) ?
4856 			    &svd->vpage[seg_page(seg, a)] : NULL;
4857 		} else {
4858 			/*
4859 			 * Size down case. Note lpgaddr may only be needed for
4860 			 * softlock case so we don't adjust it here.
4861 			 */
4862 			VM_STAT_ADD(segvnvmstats.fltanpages[11]);
4863 			ASSERT(IS_P2ALIGNED(a, pgsz));
4864 			ASSERT(IS_P2ALIGNED(lpgeaddr, pgsz));
4865 			lpgeaddr = (caddr_t)P2ROUNDUP((uintptr_t)eaddr, pgsz);
4866 			ASSERT(a < lpgeaddr);
4867 			if (a < addr) {
4868 				/*
4869 				 * The beginning of the large page region can
4870 				 * be pulled to the right to make a smaller
4871 				 * region. We haven't yet faulted a single
4872 				 * page.
4873 				 */
4874 				VM_STAT_ADD(segvnvmstats.fltanpages[12]);
4875 				a = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
4876 				ASSERT(a >= lpgaddr);
4877 				aindx = svd->anon_index + seg_page(seg, a);
4878 				vpage = (svd->vpage != NULL) ?
4879 				    &svd->vpage[seg_page(seg, a)] : NULL;
4880 			}
4881 		}
4882 	}
4883 	VM_STAT_ADD(segvnvmstats.fltanpages[13]);
4884 	ANON_LOCK_EXIT(&amp->a_rwlock);
4885 	kmem_free(ppa, ppasize);
4886 	return (0);
4887 error:
4888 	VM_STAT_ADD(segvnvmstats.fltanpages[14]);
4889 	ANON_LOCK_EXIT(&amp->a_rwlock);
4890 	kmem_free(ppa, ppasize);
4891 	if (type == F_SOFTLOCK && a > lpgaddr) {
4892 		VM_STAT_ADD(segvnvmstats.fltanpages[15]);
4893 		segvn_softunlock(seg, lpgaddr, a - lpgaddr, S_OTHER);
4894 	}
4895 	return (err);
4896 }
4897 
4898 int fltadvice = 1;	/* set to free behind pages for sequential access */
4899 
4900 /*
4901  * This routine is called via a machine specific fault handling routine.
4902  * It is also called by software routines wishing to lock or unlock
4903  * a range of addresses.
4904  *
4905  * Here is the basic algorithm:
4906  *	If unlocking
4907  *		Call segvn_softunlock
4908  *		Return
4909  *	endif
4910  *	Checking and set up work
4911  *	If we will need some non-anonymous pages
4912  *		Call VOP_GETPAGE over the range of non-anonymous pages
4913  *	endif
4914  *	Loop over all addresses requested
4915  *		Call segvn_faultpage passing in page list
4916  *		    to load up translations and handle anonymous pages
4917  *	endloop
4918  *	Load up translation to any additional pages in page list not
4919  *	    already handled that fit into this segment
4920  */
4921 static faultcode_t
4922 segvn_fault(struct hat *hat, struct seg *seg, caddr_t addr, size_t len,
4923     enum fault_type type, enum seg_rw rw)
4924 {
4925 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
4926 	page_t **plp, **ppp, *pp;
4927 	u_offset_t off;
4928 	caddr_t a;
4929 	struct vpage *vpage;
4930 	uint_t vpprot, prot;
4931 	int err;
4932 	page_t *pl[PVN_GETPAGE_NUM + 1];
4933 	size_t plsz, pl_alloc_sz;
4934 	size_t page;
4935 	ulong_t anon_index;
4936 	struct anon_map *amp;
4937 	int dogetpage = 0;
4938 	caddr_t	lpgaddr, lpgeaddr;
4939 	size_t pgsz;
4940 	anon_sync_obj_t cookie;
4941 	int brkcow = BREAK_COW_SHARE(rw, type, svd->type);
4942 
4943 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
4944 	ASSERT(svd->amp == NULL || svd->rcookie == HAT_INVALID_REGION_COOKIE);
4945 
4946 	/*
4947 	 * First handle the easy stuff
4948 	 */
4949 	if (type == F_SOFTUNLOCK) {
4950 		if (rw == S_READ_NOCOW) {
4951 			rw = S_READ;
4952 			ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
4953 		}
4954 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
4955 		pgsz = (seg->s_szc == 0) ? PAGESIZE :
4956 		    page_get_pagesize(seg->s_szc);
4957 		VM_STAT_COND_ADD(pgsz > PAGESIZE, segvnvmstats.fltanpages[16]);
4958 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
4959 		segvn_softunlock(seg, lpgaddr, lpgeaddr - lpgaddr, rw);
4960 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4961 		return (0);
4962 	}
4963 
4964 	ASSERT(svd->tr_state == SEGVN_TR_OFF ||
4965 	    !HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
4966 	if (brkcow == 0) {
4967 		if (svd->tr_state == SEGVN_TR_INIT) {
4968 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4969 			if (svd->tr_state == SEGVN_TR_INIT) {
4970 				ASSERT(svd->vp != NULL && svd->amp == NULL);
4971 				ASSERT(svd->flags & MAP_TEXT);
4972 				ASSERT(svd->type == MAP_PRIVATE);
4973 				segvn_textrepl(seg);
4974 				ASSERT(svd->tr_state != SEGVN_TR_INIT);
4975 				ASSERT(svd->tr_state != SEGVN_TR_ON ||
4976 				    svd->amp != NULL);
4977 			}
4978 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4979 		}
4980 	} else if (svd->tr_state != SEGVN_TR_OFF) {
4981 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
4982 
4983 		if (rw == S_WRITE && svd->tr_state != SEGVN_TR_OFF) {
4984 			ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
4985 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4986 			return (FC_PROT);
4987 		}
4988 
4989 		if (svd->tr_state == SEGVN_TR_ON) {
4990 			ASSERT(svd->vp != NULL && svd->amp != NULL);
4991 			segvn_textunrepl(seg, 0);
4992 			ASSERT(svd->amp == NULL &&
4993 			    svd->tr_state == SEGVN_TR_OFF);
4994 		} else if (svd->tr_state != SEGVN_TR_OFF) {
4995 			svd->tr_state = SEGVN_TR_OFF;
4996 		}
4997 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
4998 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
4999 	}
5000 
5001 top:
5002 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5003 
5004 	/*
5005 	 * If we have the same protections for the entire segment,
5006 	 * insure that the access being attempted is legitimate.
5007 	 */
5008 
5009 	if (svd->pageprot == 0) {
5010 		uint_t protchk;
5011 
5012 		switch (rw) {
5013 		case S_READ:
5014 		case S_READ_NOCOW:
5015 			protchk = PROT_READ;
5016 			break;
5017 		case S_WRITE:
5018 			protchk = PROT_WRITE;
5019 			break;
5020 		case S_EXEC:
5021 			protchk = PROT_EXEC;
5022 			break;
5023 		case S_OTHER:
5024 		default:
5025 			protchk = PROT_READ | PROT_WRITE | PROT_EXEC;
5026 			break;
5027 		}
5028 
5029 		if ((svd->prot & protchk) == 0) {
5030 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5031 			return (FC_PROT);	/* illegal access type */
5032 		}
5033 	}
5034 
5035 	if (brkcow && HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5036 		/* this must be SOFTLOCK S_READ fault */
5037 		ASSERT(svd->amp == NULL);
5038 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5039 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5040 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5041 		if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5042 			/*
5043 			 * this must be the first ever non S_READ_NOCOW
5044 			 * softlock for this segment.
5045 			 */
5046 			ASSERT(svd->softlockcnt == 0);
5047 			hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5048 			    HAT_REGION_TEXT);
5049 			svd->rcookie = HAT_INVALID_REGION_COOKIE;
5050 		}
5051 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5052 		goto top;
5053 	}
5054 
5055 	/*
5056 	 * We can't allow the long term use of softlocks for vmpss segments,
5057 	 * because in some file truncation cases we should be able to demote
5058 	 * the segment, which requires that there are no softlocks.  The
5059 	 * only case where it's ok to allow a SOFTLOCK fault against a vmpss
5060 	 * segment is S_READ_NOCOW, where the caller holds the address space
5061 	 * locked as writer and calls softunlock before dropping the as lock.
5062 	 * S_READ_NOCOW is used by /proc to read memory from another user.
5063 	 *
5064 	 * Another deadlock between SOFTLOCK and file truncation can happen
5065 	 * because segvn_fault_vnodepages() calls the FS one pagesize at
5066 	 * a time. A second VOP_GETPAGE() call by segvn_fault_vnodepages()
5067 	 * can cause a deadlock because the first set of page_t's remain
5068 	 * locked SE_SHARED.  To avoid this, we demote segments on a first
5069 	 * SOFTLOCK if they have a length greater than the segment's
5070 	 * page size.
5071 	 *
5072 	 * So for now, we only avoid demoting a segment on a SOFTLOCK when
5073 	 * the access type is S_READ_NOCOW and the fault length is less than
5074 	 * or equal to the segment's page size. While this is quite restrictive,
5075 	 * it should be the most common case of SOFTLOCK against a vmpss
5076 	 * segment.
5077 	 *
5078 	 * For S_READ_NOCOW, it's safe not to do a copy on write because the
5079 	 * caller makes sure no COW will be caused by another thread for a
5080 	 * softlocked page.
5081 	 */
5082 	if (type == F_SOFTLOCK && svd->vp != NULL && seg->s_szc != 0) {
5083 		int demote = 0;
5084 
5085 		if (rw != S_READ_NOCOW) {
5086 			demote = 1;
5087 		}
5088 		if (!demote && len > PAGESIZE) {
5089 			pgsz = page_get_pagesize(seg->s_szc);
5090 			CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr,
5091 			    lpgeaddr);
5092 			if (lpgeaddr - lpgaddr > pgsz) {
5093 				demote = 1;
5094 			}
5095 		}
5096 
5097 		ASSERT(demote || AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5098 
5099 		if (demote) {
5100 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5101 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5102 			if (seg->s_szc != 0) {
5103 				segvn_vmpss_clrszc_cnt++;
5104 				ASSERT(svd->softlockcnt == 0);
5105 				err = segvn_clrszc(seg);
5106 				if (err) {
5107 					segvn_vmpss_clrszc_err++;
5108 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5109 					return (FC_MAKE_ERR(err));
5110 				}
5111 			}
5112 			ASSERT(seg->s_szc == 0);
5113 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5114 			goto top;
5115 		}
5116 	}
5117 
5118 	/*
5119 	 * Check to see if we need to allocate an anon_map structure.
5120 	 */
5121 	if (svd->amp == NULL && (svd->vp == NULL || brkcow)) {
5122 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5123 		/*
5124 		 * Drop the "read" lock on the segment and acquire
5125 		 * the "write" version since we have to allocate the
5126 		 * anon_map.
5127 		 */
5128 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5129 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5130 
5131 		if (svd->amp == NULL) {
5132 			svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
5133 			svd->amp->a_szc = seg->s_szc;
5134 		}
5135 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5136 
5137 		/*
5138 		 * Start all over again since segment protections
5139 		 * may have changed after we dropped the "read" lock.
5140 		 */
5141 		goto top;
5142 	}
5143 
5144 	/*
5145 	 * S_READ_NOCOW vs S_READ distinction was
5146 	 * only needed for the code above. After
5147 	 * that we treat it as S_READ.
5148 	 */
5149 	if (rw == S_READ_NOCOW) {
5150 		ASSERT(type == F_SOFTLOCK);
5151 		ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5152 		rw = S_READ;
5153 	}
5154 
5155 	amp = svd->amp;
5156 
5157 	/*
5158 	 * MADV_SEQUENTIAL work is ignored for large page segments.
5159 	 */
5160 	if (seg->s_szc != 0) {
5161 		pgsz = page_get_pagesize(seg->s_szc);
5162 		ASSERT(SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
5163 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
5164 		if (svd->vp == NULL) {
5165 			err = segvn_fault_anonpages(hat, seg, lpgaddr,
5166 			    lpgeaddr, type, rw, addr, addr + len, brkcow);
5167 		} else {
5168 			err = segvn_fault_vnodepages(hat, seg, lpgaddr,
5169 				lpgeaddr, type, rw, addr, addr + len, brkcow);
5170 			if (err == IE_RETRY) {
5171 				ASSERT(seg->s_szc == 0);
5172 				ASSERT(SEGVN_READ_HELD(seg->s_as, &svd->lock));
5173 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5174 				goto top;
5175 			}
5176 		}
5177 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5178 		return (err);
5179 	}
5180 
5181 	page = seg_page(seg, addr);
5182 	if (amp != NULL) {
5183 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
5184 		anon_index = svd->anon_index + page;
5185 
5186 		if (type == F_PROT && rw == S_READ &&
5187 		    svd->tr_state == SEGVN_TR_OFF &&
5188 		    svd->type == MAP_PRIVATE && svd->pageprot == 0) {
5189 			size_t index = anon_index;
5190 			struct anon *ap;
5191 
5192 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5193 			/*
5194 			 * The fast path could apply to S_WRITE also, except
5195 			 * that the protection fault could be caused by lazy
5196 			 * tlb flush when ro->rw. In this case, the pte is
5197 			 * RW already. But RO in the other cpu's tlb causes
5198 			 * the fault. Since hat_chgprot won't do anything if
5199 			 * pte doesn't change, we may end up faulting
5200 			 * indefinitely until the RO tlb entry gets replaced.
5201 			 */
5202 			for (a = addr; a < addr + len; a += PAGESIZE, index++) {
5203 				anon_array_enter(amp, index, &cookie);
5204 				ap = anon_get_ptr(amp->ahp, index);
5205 				anon_array_exit(&cookie);
5206 				if ((ap == NULL) || (ap->an_refcnt != 1)) {
5207 					ANON_LOCK_EXIT(&amp->a_rwlock);
5208 					goto slow;
5209 				}
5210 			}
5211 			hat_chgprot(seg->s_as->a_hat, addr, len, svd->prot);
5212 			ANON_LOCK_EXIT(&amp->a_rwlock);
5213 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5214 			return (0);
5215 		}
5216 	}
5217 slow:
5218 
5219 	if (svd->vpage == NULL)
5220 		vpage = NULL;
5221 	else
5222 		vpage = &svd->vpage[page];
5223 
5224 	off = svd->offset + (uintptr_t)(addr - seg->s_base);
5225 
5226 	/*
5227 	 * If MADV_SEQUENTIAL has been set for the particular page we
5228 	 * are faulting on, free behind all pages in the segment and put
5229 	 * them on the free list.
5230 	 */
5231 
5232 	if ((page != 0) && fltadvice && svd->tr_state != SEGVN_TR_ON) {
5233 		struct vpage *vpp;
5234 		ulong_t fanon_index;
5235 		size_t fpage;
5236 		u_offset_t pgoff, fpgoff;
5237 		struct vnode *fvp;
5238 		struct anon *fap = NULL;
5239 
5240 		if (svd->advice == MADV_SEQUENTIAL ||
5241 		    (svd->pageadvice &&
5242 		    VPP_ADVICE(vpage) == MADV_SEQUENTIAL)) {
5243 			pgoff = off - PAGESIZE;
5244 			fpage = page - 1;
5245 			if (vpage != NULL)
5246 				vpp = &svd->vpage[fpage];
5247 			if (amp != NULL)
5248 				fanon_index = svd->anon_index + fpage;
5249 
5250 			while (pgoff > svd->offset) {
5251 				if (svd->advice != MADV_SEQUENTIAL &&
5252 				    (!svd->pageadvice || (vpage &&
5253 				    VPP_ADVICE(vpp) != MADV_SEQUENTIAL)))
5254 					break;
5255 
5256 				/*
5257 				 * If this is an anon page, we must find the
5258 				 * correct <vp, offset> for it
5259 				 */
5260 				fap = NULL;
5261 				if (amp != NULL) {
5262 					ANON_LOCK_ENTER(&amp->a_rwlock,
5263 						RW_READER);
5264 					anon_array_enter(amp, fanon_index,
5265 						&cookie);
5266 					fap = anon_get_ptr(amp->ahp,
5267 					    fanon_index);
5268 					if (fap != NULL) {
5269 						swap_xlate(fap, &fvp, &fpgoff);
5270 					} else {
5271 						fpgoff = pgoff;
5272 						fvp = svd->vp;
5273 					}
5274 					anon_array_exit(&cookie);
5275 					ANON_LOCK_EXIT(&amp->a_rwlock);
5276 				} else {
5277 					fpgoff = pgoff;
5278 					fvp = svd->vp;
5279 				}
5280 				if (fvp == NULL)
5281 					break;	/* XXX */
5282 				/*
5283 				 * Skip pages that are free or have an
5284 				 * "exclusive" lock.
5285 				 */
5286 				pp = page_lookup_nowait(fvp, fpgoff, SE_SHARED);
5287 				if (pp == NULL)
5288 					break;
5289 				/*
5290 				 * We don't need the page_struct_lock to test
5291 				 * as this is only advisory; even if we
5292 				 * acquire it someone might race in and lock
5293 				 * the page after we unlock and before the
5294 				 * PUTPAGE, then VOP_PUTPAGE will do nothing.
5295 				 */
5296 				if (pp->p_lckcnt == 0 && pp->p_cowcnt == 0) {
5297 					/*
5298 					 * Hold the vnode before releasing
5299 					 * the page lock to prevent it from
5300 					 * being freed and re-used by some
5301 					 * other thread.
5302 					 */
5303 					VN_HOLD(fvp);
5304 					page_unlock(pp);
5305 					/*
5306 					 * We should build a page list
5307 					 * to kluster putpages XXX
5308 					 */
5309 					(void) VOP_PUTPAGE(fvp,
5310 					    (offset_t)fpgoff, PAGESIZE,
5311 					    (B_DONTNEED|B_FREE|B_ASYNC),
5312 					    svd->cred);
5313 					VN_RELE(fvp);
5314 				} else {
5315 					/*
5316 					 * XXX - Should the loop terminate if
5317 					 * the page is `locked'?
5318 					 */
5319 					page_unlock(pp);
5320 				}
5321 				--vpp;
5322 				--fanon_index;
5323 				pgoff -= PAGESIZE;
5324 			}
5325 		}
5326 	}
5327 
5328 	plp = pl;
5329 	*plp = NULL;
5330 	pl_alloc_sz = 0;
5331 
5332 	/*
5333 	 * See if we need to call VOP_GETPAGE for
5334 	 * *any* of the range being faulted on.
5335 	 * We can skip all of this work if there
5336 	 * was no original vnode.
5337 	 */
5338 	if (svd->vp != NULL) {
5339 		u_offset_t vp_off;
5340 		size_t vp_len;
5341 		struct anon *ap;
5342 		vnode_t *vp;
5343 
5344 		vp_off = off;
5345 		vp_len = len;
5346 
5347 		if (amp == NULL)
5348 			dogetpage = 1;
5349 		else {
5350 			/*
5351 			 * Only acquire reader lock to prevent amp->ahp
5352 			 * from being changed.  It's ok to miss pages,
5353 			 * hence we don't do anon_array_enter
5354 			 */
5355 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5356 			ap = anon_get_ptr(amp->ahp, anon_index);
5357 
5358 			if (len <= PAGESIZE)
5359 				/* inline non_anon() */
5360 				dogetpage = (ap == NULL);
5361 			else
5362 				dogetpage = non_anon(amp->ahp, anon_index,
5363 				    &vp_off, &vp_len);
5364 			ANON_LOCK_EXIT(&amp->a_rwlock);
5365 		}
5366 
5367 		if (dogetpage) {
5368 			enum seg_rw arw;
5369 			struct as *as = seg->s_as;
5370 
5371 			if (len > ptob((sizeof (pl) / sizeof (pl[0])) - 1)) {
5372 				/*
5373 				 * Page list won't fit in local array,
5374 				 * allocate one of the needed size.
5375 				 */
5376 				pl_alloc_sz =
5377 				    (btop(len) + 1) * sizeof (page_t *);
5378 				plp = kmem_alloc(pl_alloc_sz, KM_SLEEP);
5379 				plp[0] = NULL;
5380 				plsz = len;
5381 			} else if (rw == S_WRITE && svd->type == MAP_PRIVATE ||
5382 			    svd->tr_state == SEGVN_TR_ON || rw == S_OTHER ||
5383 			    (((size_t)(addr + PAGESIZE) <
5384 			    (size_t)(seg->s_base + seg->s_size)) &&
5385 			    hat_probe(as->a_hat, addr + PAGESIZE))) {
5386 				/*
5387 				 * Ask VOP_GETPAGE to return the exact number
5388 				 * of pages if
5389 				 * (a) this is a COW fault, or
5390 				 * (b) this is a software fault, or
5391 				 * (c) next page is already mapped.
5392 				 */
5393 				plsz = len;
5394 			} else {
5395 				/*
5396 				 * Ask VOP_GETPAGE to return adjacent pages
5397 				 * within the segment.
5398 				 */
5399 				plsz = MIN((size_t)PVN_GETPAGE_SZ, (size_t)
5400 					((seg->s_base + seg->s_size) - addr));
5401 				ASSERT((addr + plsz) <=
5402 				    (seg->s_base + seg->s_size));
5403 			}
5404 
5405 			/*
5406 			 * Need to get some non-anonymous pages.
5407 			 * We need to make only one call to GETPAGE to do
5408 			 * this to prevent certain deadlocking conditions
5409 			 * when we are doing locking.  In this case
5410 			 * non_anon() should have picked up the smallest
5411 			 * range which includes all the non-anonymous
5412 			 * pages in the requested range.  We have to
5413 			 * be careful regarding which rw flag to pass in
5414 			 * because on a private mapping, the underlying
5415 			 * object is never allowed to be written.
5416 			 */
5417 			if (rw == S_WRITE && svd->type == MAP_PRIVATE) {
5418 				arw = S_READ;
5419 			} else {
5420 				arw = rw;
5421 			}
5422 			vp = svd->vp;
5423 			TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5424 				"segvn_getpage:seg %p addr %p vp %p",
5425 				seg, addr, vp);
5426 			err = VOP_GETPAGE(vp, (offset_t)vp_off, vp_len,
5427 			    &vpprot, plp, plsz, seg, addr + (vp_off - off), arw,
5428 			    svd->cred);
5429 			if (err) {
5430 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5431 				segvn_pagelist_rele(plp);
5432 				if (pl_alloc_sz)
5433 					kmem_free(plp, pl_alloc_sz);
5434 				return (FC_MAKE_ERR(err));
5435 			}
5436 			if (svd->type == MAP_PRIVATE)
5437 				vpprot &= ~PROT_WRITE;
5438 		}
5439 	}
5440 
5441 	/*
5442 	 * N.B. at this time the plp array has all the needed non-anon
5443 	 * pages in addition to (possibly) having some adjacent pages.
5444 	 */
5445 
5446 	/*
5447 	 * Always acquire the anon_array_lock to prevent
5448 	 * 2 threads from allocating separate anon slots for
5449 	 * the same "addr".
5450 	 *
5451 	 * If this is a copy-on-write fault and we don't already
5452 	 * have the anon_array_lock, acquire it to prevent the
5453 	 * fault routine from handling multiple copy-on-write faults
5454 	 * on the same "addr" in the same address space.
5455 	 *
5456 	 * Only one thread should deal with the fault since after
5457 	 * it is handled, the other threads can acquire a translation
5458 	 * to the newly created private page.  This prevents two or
5459 	 * more threads from creating different private pages for the
5460 	 * same fault.
5461 	 *
5462 	 * We grab "serialization" lock here if this is a MAP_PRIVATE segment
5463 	 * to prevent deadlock between this thread and another thread
5464 	 * which has soft-locked this page and wants to acquire serial_lock.
5465 	 * ( bug 4026339 )
5466 	 *
5467 	 * The fix for bug 4026339 becomes unnecessary when using the
5468 	 * locking scheme with per amp rwlock and a global set of hash
5469 	 * lock, anon_array_lock.  If we steal a vnode page when low
5470 	 * on memory and upgrad the page lock through page_rename,
5471 	 * then the page is PAGE_HANDLED, nothing needs to be done
5472 	 * for this page after returning from segvn_faultpage.
5473 	 *
5474 	 * But really, the page lock should be downgraded after
5475 	 * the stolen page is page_rename'd.
5476 	 */
5477 
5478 	if (amp != NULL)
5479 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5480 
5481 	/*
5482 	 * Ok, now loop over the address range and handle faults
5483 	 */
5484 	for (a = addr; a < addr + len; a += PAGESIZE, off += PAGESIZE) {
5485 		err = segvn_faultpage(hat, seg, a, off, vpage, plp, vpprot,
5486 		    type, rw, brkcow, a == addr);
5487 		if (err) {
5488 			if (amp != NULL)
5489 				ANON_LOCK_EXIT(&amp->a_rwlock);
5490 			if (type == F_SOFTLOCK && a > addr) {
5491 				segvn_softunlock(seg, addr, (a - addr),
5492 				    S_OTHER);
5493 			}
5494 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5495 			segvn_pagelist_rele(plp);
5496 			if (pl_alloc_sz)
5497 				kmem_free(plp, pl_alloc_sz);
5498 			return (err);
5499 		}
5500 		if (vpage) {
5501 			vpage++;
5502 		} else if (svd->vpage) {
5503 			page = seg_page(seg, addr);
5504 			vpage = &svd->vpage[++page];
5505 		}
5506 	}
5507 
5508 	/* Didn't get pages from the underlying fs so we're done */
5509 	if (!dogetpage)
5510 		goto done;
5511 
5512 	/*
5513 	 * Now handle any other pages in the list returned.
5514 	 * If the page can be used, load up the translations now.
5515 	 * Note that the for loop will only be entered if "plp"
5516 	 * is pointing to a non-NULL page pointer which means that
5517 	 * VOP_GETPAGE() was called and vpprot has been initialized.
5518 	 */
5519 	if (svd->pageprot == 0)
5520 		prot = svd->prot & vpprot;
5521 
5522 
5523 	/*
5524 	 * Large Files: diff should be unsigned value because we started
5525 	 * supporting > 2GB segment sizes from 2.5.1 and when a
5526 	 * large file of size > 2GB gets mapped to address space
5527 	 * the diff value can be > 2GB.
5528 	 */
5529 
5530 	for (ppp = plp; (pp = *ppp) != NULL; ppp++) {
5531 		size_t diff;
5532 		struct anon *ap;
5533 		int anon_index;
5534 		anon_sync_obj_t cookie;
5535 		int hat_flag = HAT_LOAD_ADV;
5536 
5537 		if (svd->flags & MAP_TEXT) {
5538 			hat_flag |= HAT_LOAD_TEXT;
5539 		}
5540 
5541 		if (pp == PAGE_HANDLED)
5542 			continue;
5543 
5544 		if (svd->tr_state != SEGVN_TR_ON &&
5545 		    pp->p_offset >=  svd->offset &&
5546 		    pp->p_offset < svd->offset + seg->s_size) {
5547 
5548 			diff = pp->p_offset - svd->offset;
5549 
5550 			/*
5551 			 * Large Files: Following is the assertion
5552 			 * validating the above cast.
5553 			 */
5554 			ASSERT(svd->vp == pp->p_vnode);
5555 
5556 			page = btop(diff);
5557 			if (svd->pageprot)
5558 				prot = VPP_PROT(&svd->vpage[page]) & vpprot;
5559 
5560 			/*
5561 			 * Prevent other threads in the address space from
5562 			 * creating private pages (i.e., allocating anon slots)
5563 			 * while we are in the process of loading translations
5564 			 * to additional pages returned by the underlying
5565 			 * object.
5566 			 */
5567 			if (amp != NULL) {
5568 				anon_index = svd->anon_index + page;
5569 				anon_array_enter(amp, anon_index, &cookie);
5570 				ap = anon_get_ptr(amp->ahp, anon_index);
5571 			}
5572 			if ((amp == NULL) || (ap == NULL)) {
5573 				if (IS_VMODSORT(pp->p_vnode) ||
5574 				    enable_mbit_wa) {
5575 					if (rw == S_WRITE)
5576 						hat_setmod(pp);
5577 					else if (rw != S_OTHER &&
5578 					    !hat_ismod(pp))
5579 						prot &= ~PROT_WRITE;
5580 				}
5581 				/*
5582 				 * Skip mapping read ahead pages marked
5583 				 * for migration, so they will get migrated
5584 				 * properly on fault
5585 				 */
5586 				ASSERT(amp == NULL ||
5587 				    svd->rcookie == HAT_INVALID_REGION_COOKIE);
5588 				if ((prot & PROT_READ) && !PP_ISMIGRATE(pp)) {
5589 					hat_memload_region(hat,
5590 					    seg->s_base + diff,
5591 					    pp, prot, hat_flag,
5592 					    svd->rcookie);
5593 				}
5594 			}
5595 			if (amp != NULL)
5596 				anon_array_exit(&cookie);
5597 		}
5598 		page_unlock(pp);
5599 	}
5600 done:
5601 	if (amp != NULL)
5602 		ANON_LOCK_EXIT(&amp->a_rwlock);
5603 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5604 	if (pl_alloc_sz)
5605 		kmem_free(plp, pl_alloc_sz);
5606 	return (0);
5607 }
5608 
5609 /*
5610  * This routine is used to start I/O on pages asynchronously.  XXX it will
5611  * only create PAGESIZE pages. At fault time they will be relocated into
5612  * larger pages.
5613  */
5614 static faultcode_t
5615 segvn_faulta(struct seg *seg, caddr_t addr)
5616 {
5617 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5618 	int err;
5619 	struct anon_map *amp;
5620 	vnode_t *vp;
5621 
5622 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5623 
5624 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
5625 	if ((amp = svd->amp) != NULL) {
5626 		struct anon *ap;
5627 
5628 		/*
5629 		 * Reader lock to prevent amp->ahp from being changed.
5630 		 * This is advisory, it's ok to miss a page, so
5631 		 * we don't do anon_array_enter lock.
5632 		 */
5633 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
5634 		if ((ap = anon_get_ptr(amp->ahp,
5635 			svd->anon_index + seg_page(seg, addr))) != NULL) {
5636 
5637 			err = anon_getpage(&ap, NULL, NULL,
5638 			    0, seg, addr, S_READ, svd->cred);
5639 
5640 			ANON_LOCK_EXIT(&amp->a_rwlock);
5641 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5642 			if (err)
5643 				return (FC_MAKE_ERR(err));
5644 			return (0);
5645 		}
5646 		ANON_LOCK_EXIT(&amp->a_rwlock);
5647 	}
5648 
5649 	if (svd->vp == NULL) {
5650 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5651 		return (0);			/* zfod page - do nothing now */
5652 	}
5653 
5654 	vp = svd->vp;
5655 	TRACE_3(TR_FAC_VM, TR_SEGVN_GETPAGE,
5656 		"segvn_getpage:seg %p addr %p vp %p", seg, addr, vp);
5657 	err = VOP_GETPAGE(vp,
5658 	    (offset_t)(svd->offset + (uintptr_t)(addr - seg->s_base)),
5659 	    PAGESIZE, NULL, NULL, 0, seg, addr,
5660 	    S_OTHER, svd->cred);
5661 
5662 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5663 	if (err)
5664 		return (FC_MAKE_ERR(err));
5665 	return (0);
5666 }
5667 
5668 static int
5669 segvn_setprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
5670 {
5671 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5672 	struct vpage *svp, *evp;
5673 	struct vnode *vp;
5674 	size_t pgsz;
5675 	pgcnt_t pgcnt;
5676 	anon_sync_obj_t cookie;
5677 	int unload_done = 0;
5678 
5679 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
5680 
5681 	if ((svd->maxprot & prot) != prot)
5682 		return (EACCES);			/* violated maxprot */
5683 
5684 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
5685 
5686 	/* return if prot is the same */
5687 	if (!svd->pageprot && svd->prot == prot) {
5688 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5689 		return (0);
5690 	}
5691 
5692 	/*
5693 	 * Since we change protections we first have to flush the cache.
5694 	 * This makes sure all the pagelock calls have to recheck
5695 	 * protections.
5696 	 */
5697 	if (svd->softlockcnt > 0) {
5698 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5699 		/*
5700 		 * Since we do have the segvn writers lock nobody can fill
5701 		 * the cache with entries belonging to this seg during
5702 		 * the purge. The flush either succeeds or we still have
5703 		 * pending I/Os.
5704 		 */
5705 		segvn_purge(seg);
5706 		if (svd->softlockcnt > 0) {
5707 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5708 			return (EAGAIN);
5709 		}
5710 	}
5711 
5712 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
5713 		ASSERT(svd->amp == NULL);
5714 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
5715 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
5716 		    HAT_REGION_TEXT);
5717 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
5718 		unload_done = 1;
5719 	} else if (svd->tr_state == SEGVN_TR_INIT) {
5720 		svd->tr_state = SEGVN_TR_OFF;
5721 	} else if (svd->tr_state == SEGVN_TR_ON) {
5722 		ASSERT(svd->amp != NULL);
5723 		segvn_textunrepl(seg, 0);
5724 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
5725 		unload_done = 1;
5726 	}
5727 
5728 	if ((prot & PROT_WRITE) && svd->type == MAP_SHARED &&
5729 	    svd->vp != NULL && (svd->vp->v_flag & VVMEXEC)) {
5730 		ASSERT(vn_is_mapped(svd->vp, V_WRITE));
5731 		segvn_inval_trcache(svd->vp);
5732 	}
5733 	if (seg->s_szc != 0) {
5734 		int err;
5735 		pgsz = page_get_pagesize(seg->s_szc);
5736 		pgcnt = pgsz >> PAGESHIFT;
5737 		ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
5738 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
5739 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5740 			ASSERT(seg->s_base != addr || seg->s_size != len);
5741 			/*
5742 			 * If we are holding the as lock as a reader then
5743 			 * we need to return IE_RETRY and let the as
5744 			 * layer drop and re-aquire the lock as a writer.
5745 			 */
5746 			if (AS_READ_HELD(seg->s_as, &seg->s_as->a_lock))
5747 				return (IE_RETRY);
5748 			VM_STAT_ADD(segvnvmstats.demoterange[1]);
5749 			if (svd->type == MAP_PRIVATE || svd->vp != NULL) {
5750 				err = segvn_demote_range(seg, addr, len,
5751 				    SDR_END, 0);
5752 			} else {
5753 				uint_t szcvec = map_pgszcvec(seg->s_base,
5754 				    pgsz, (uintptr_t)seg->s_base,
5755 				    (svd->flags & MAP_TEXT), MAPPGSZC_SHM, 0);
5756 				err = segvn_demote_range(seg, addr, len,
5757 				    SDR_END, szcvec);
5758 			}
5759 			if (err == 0)
5760 				return (IE_RETRY);
5761 			if (err == ENOMEM)
5762 				return (IE_NOMEM);
5763 			return (err);
5764 		}
5765 	}
5766 
5767 
5768 	/*
5769 	 * If it's a private mapping and we're making it writable
5770 	 * and no swap space has been reserved, have to reserve
5771 	 * it all now.  If it's a private mapping to a file (i.e., vp != NULL)
5772 	 * and we're removing write permission on the entire segment and
5773 	 * we haven't modified any pages, we can release the swap space.
5774 	 */
5775 	if (svd->type == MAP_PRIVATE) {
5776 		if (prot & PROT_WRITE) {
5777 			size_t sz;
5778 			if (svd->swresv == 0 && !(svd->flags & MAP_NORESERVE)) {
5779 				if (anon_resv_zone(seg->s_size,
5780 				    seg->s_as->a_proc->p_zone) == 0) {
5781 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5782 					return (IE_NOMEM);
5783 				}
5784 				sz = svd->swresv = seg->s_size;
5785 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
5786 					"anon proc:%p %lu %u",
5787 					seg, sz, 1);
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 				anon_unresv_zone(svd->swresv,
5800 				    seg->s_as->a_proc->p_zone);
5801 				svd->swresv = 0;
5802 				TRACE_3(TR_FAC_VM, TR_ANON_PROC,
5803 					"anon proc:%p %lu %u",
5804 					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(pp)) {
5894 							page_unlock(pp);
5895 							break;
5896 						    }
5897 						} else {
5898 						    if (!page_subclaim(pp)) {
5899 							page_unlock(pp);
5900 							break;
5901 						    }
5902 						}
5903 					}
5904 					page_unlock(pp);
5905 				} else if (amp != NULL)
5906 					anon_array_exit(&cookie);
5907 			}
5908 			VPP_SETPROT(svp, prot);
5909 			offset += PAGESIZE;
5910 		}
5911 		if (amp != NULL)
5912 			ANON_LOCK_EXIT(&amp->a_rwlock);
5913 
5914 		/*
5915 		 * Did we terminate prematurely?  If so, simply unload
5916 		 * the translations to the things we've updated so far.
5917 		 */
5918 		if (svp != evp) {
5919 			if (unload_done) {
5920 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5921 				return (IE_NOMEM);
5922 			}
5923 			len = (svp - &svd->vpage[seg_page(seg, addr)]) *
5924 			    PAGESIZE;
5925 			ASSERT(seg->s_szc == 0 || IS_P2ALIGNED(len, pgsz));
5926 			if (len != 0)
5927 				hat_unload(seg->s_as->a_hat, addr,
5928 				    len, HAT_UNLOAD);
5929 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5930 			return (IE_NOMEM);
5931 		}
5932 	} else {
5933 		segvn_vpage(seg);
5934 		svd->pageprot = 1;
5935 		evp = &svd->vpage[seg_page(seg, addr + len)];
5936 		for (svp = &svd->vpage[seg_page(seg, addr)]; svp < evp; svp++) {
5937 			VPP_SETPROT(svp, prot);
5938 		}
5939 	}
5940 
5941 	if (unload_done) {
5942 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5943 		return (0);
5944 	}
5945 
5946 	if (((prot & PROT_WRITE) != 0 &&
5947 	    (svd->vp != NULL || svd->type == MAP_PRIVATE)) ||
5948 	    (prot & ~PROT_USER) == PROT_NONE) {
5949 		/*
5950 		 * Either private or shared data with write access (in
5951 		 * which case we need to throw out all former translations
5952 		 * so that we get the right translations set up on fault
5953 		 * and we don't allow write access to any copy-on-write pages
5954 		 * that might be around or to prevent write access to pages
5955 		 * representing holes in a file), or we don't have permission
5956 		 * to access the memory at all (in which case we have to
5957 		 * unload any current translations that might exist).
5958 		 */
5959 		hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
5960 	} else {
5961 		/*
5962 		 * A shared mapping or a private mapping in which write
5963 		 * protection is going to be denied - just change all the
5964 		 * protections over the range of addresses in question.
5965 		 * segvn does not support any other attributes other
5966 		 * than prot so we can use hat_chgattr.
5967 		 */
5968 		hat_chgattr(seg->s_as->a_hat, addr, len, prot);
5969 	}
5970 
5971 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
5972 
5973 	return (0);
5974 }
5975 
5976 /*
5977  * segvn_setpagesize is called via SEGOP_SETPAGESIZE from as_setpagesize,
5978  * to determine if the seg is capable of mapping the requested szc.
5979  */
5980 static int
5981 segvn_setpagesize(struct seg *seg, caddr_t addr, size_t len, uint_t szc)
5982 {
5983 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
5984 	struct segvn_data *nsvd;
5985 	struct anon_map *amp = svd->amp;
5986 	struct seg *nseg;
5987 	caddr_t eaddr = addr + len, a;
5988 	size_t pgsz = page_get_pagesize(szc);
5989 	pgcnt_t pgcnt = page_get_pagecnt(szc);
5990 	int err;
5991 	u_offset_t off = svd->offset + (uintptr_t)(addr - seg->s_base);
5992 	extern struct vnode kvp;
5993 
5994 	ASSERT(seg->s_as && AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
5995 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
5996 
5997 	if (seg->s_szc == szc || segvn_lpg_disable != 0) {
5998 		return (0);
5999 	}
6000 
6001 	/*
6002 	 * addr should always be pgsz aligned but eaddr may be misaligned if
6003 	 * it's at the end of the segment.
6004 	 *
6005 	 * XXX we should assert this condition since as_setpagesize() logic
6006 	 * guarantees it.
6007 	 */
6008 	if (!IS_P2ALIGNED(addr, pgsz) ||
6009 	    (!IS_P2ALIGNED(eaddr, pgsz) &&
6010 		eaddr != seg->s_base + seg->s_size)) {
6011 
6012 		segvn_setpgsz_align_err++;
6013 		return (EINVAL);
6014 	}
6015 
6016 	if (amp != NULL && svd->type == MAP_SHARED) {
6017 		ulong_t an_idx = svd->anon_index + seg_page(seg, addr);
6018 		if (!IS_P2ALIGNED(an_idx, pgcnt)) {
6019 
6020 			segvn_setpgsz_anon_align_err++;
6021 			return (EINVAL);
6022 		}
6023 	}
6024 
6025 	if ((svd->flags & MAP_NORESERVE) || seg->s_as == &kas ||
6026 	    szc > segvn_maxpgszc) {
6027 		return (EINVAL);
6028 	}
6029 
6030 	/* paranoid check */
6031 	if (svd->vp != NULL &&
6032 	    (IS_SWAPFSVP(svd->vp) || VN_ISKAS(svd->vp))) {
6033 		    return (EINVAL);
6034 	}
6035 
6036 	if (seg->s_szc == 0 && svd->vp != NULL &&
6037 	    map_addr_vacalign_check(addr, off)) {
6038 		return (EINVAL);
6039 	}
6040 
6041 	/*
6042 	 * Check that protections are the same within new page
6043 	 * size boundaries.
6044 	 */
6045 	if (svd->pageprot) {
6046 		for (a = addr; a < eaddr; a += pgsz) {
6047 			if ((a + pgsz) > eaddr) {
6048 				if (!sameprot(seg, a, eaddr - a)) {
6049 					return (EINVAL);
6050 				}
6051 			} else {
6052 				if (!sameprot(seg, a, pgsz)) {
6053 					return (EINVAL);
6054 				}
6055 			}
6056 		}
6057 	}
6058 
6059 	/*
6060 	 * Since we are changing page size we first have to flush
6061 	 * the cache. This makes sure all the pagelock calls have
6062 	 * to recheck protections.
6063 	 */
6064 	if (svd->softlockcnt > 0) {
6065 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6066 		/*
6067 		 * Since we do have the segvn writers lock nobody can fill
6068 		 * the cache with entries belonging to this seg during
6069 		 * the purge. The flush either succeeds or we still have
6070 		 * pending I/Os.
6071 		 */
6072 		segvn_purge(seg);
6073 		if (svd->softlockcnt > 0) {
6074 			return (EAGAIN);
6075 		}
6076 	}
6077 
6078 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6079 		ASSERT(svd->amp == NULL);
6080 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6081 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6082 		    HAT_REGION_TEXT);
6083 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6084 	} else if (svd->tr_state == SEGVN_TR_INIT) {
6085 		svd->tr_state = SEGVN_TR_OFF;
6086 	} else if (svd->tr_state == SEGVN_TR_ON) {
6087 		ASSERT(svd->amp != NULL);
6088 		segvn_textunrepl(seg, 1);
6089 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6090 		amp = NULL;
6091 	}
6092 
6093 	/*
6094 	 * Operation for sub range of existing segment.
6095 	 */
6096 	if (addr != seg->s_base || eaddr != (seg->s_base + seg->s_size)) {
6097 		if (szc < seg->s_szc) {
6098 			VM_STAT_ADD(segvnvmstats.demoterange[2]);
6099 			err = segvn_demote_range(seg, addr, len, SDR_RANGE, 0);
6100 			if (err == 0) {
6101 				return (IE_RETRY);
6102 			}
6103 			if (err == ENOMEM) {
6104 				return (IE_NOMEM);
6105 			}
6106 			return (err);
6107 		}
6108 		if (addr != seg->s_base) {
6109 			nseg = segvn_split_seg(seg, addr);
6110 			if (eaddr != (nseg->s_base + nseg->s_size)) {
6111 				/* eaddr is szc aligned */
6112 				(void) segvn_split_seg(nseg, eaddr);
6113 			}
6114 			return (IE_RETRY);
6115 		}
6116 		if (eaddr != (seg->s_base + seg->s_size)) {
6117 			/* eaddr is szc aligned */
6118 			(void) segvn_split_seg(seg, eaddr);
6119 		}
6120 		return (IE_RETRY);
6121 	}
6122 
6123 	/*
6124 	 * Break any low level sharing and reset seg->s_szc to 0.
6125 	 */
6126 	if ((err = segvn_clrszc(seg)) != 0) {
6127 		if (err == ENOMEM) {
6128 			err = IE_NOMEM;
6129 		}
6130 		return (err);
6131 	}
6132 	ASSERT(seg->s_szc == 0);
6133 
6134 	/*
6135 	 * If the end of the current segment is not pgsz aligned
6136 	 * then attempt to concatenate with the next segment.
6137 	 */
6138 	if (!IS_P2ALIGNED(eaddr, pgsz)) {
6139 		nseg = AS_SEGNEXT(seg->s_as, seg);
6140 		if (nseg == NULL || nseg == seg || eaddr != nseg->s_base) {
6141 			return (ENOMEM);
6142 		}
6143 		if (nseg->s_ops != &segvn_ops) {
6144 			return (EINVAL);
6145 		}
6146 		nsvd = (struct segvn_data *)nseg->s_data;
6147 		if (nsvd->softlockcnt > 0) {
6148 			segvn_purge(nseg);
6149 			if (nsvd->softlockcnt > 0) {
6150 				return (EAGAIN);
6151 			}
6152 		}
6153 		err = segvn_clrszc(nseg);
6154 		if (err == ENOMEM) {
6155 			err = IE_NOMEM;
6156 		}
6157 		if (err != 0) {
6158 			return (err);
6159 		}
6160 		ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6161 		err = segvn_concat(seg, nseg, 1);
6162 		if (err == -1) {
6163 			return (EINVAL);
6164 		}
6165 		if (err == -2) {
6166 			return (IE_NOMEM);
6167 		}
6168 		return (IE_RETRY);
6169 	}
6170 
6171 	/*
6172 	 * May need to re-align anon array to
6173 	 * new szc.
6174 	 */
6175 	if (amp != NULL) {
6176 		if (!IS_P2ALIGNED(svd->anon_index, pgcnt)) {
6177 			struct anon_hdr *nahp;
6178 
6179 			ASSERT(svd->type == MAP_PRIVATE);
6180 
6181 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6182 			ASSERT(amp->refcnt == 1);
6183 			nahp = anon_create(btop(amp->size), ANON_NOSLEEP);
6184 			if (nahp == NULL) {
6185 				ANON_LOCK_EXIT(&amp->a_rwlock);
6186 				return (IE_NOMEM);
6187 			}
6188 			if (anon_copy_ptr(amp->ahp, svd->anon_index,
6189 				nahp, 0, btop(seg->s_size), ANON_NOSLEEP)) {
6190 				anon_release(nahp, btop(amp->size));
6191 				ANON_LOCK_EXIT(&amp->a_rwlock);
6192 				return (IE_NOMEM);
6193 			}
6194 			anon_release(amp->ahp, btop(amp->size));
6195 			amp->ahp = nahp;
6196 			svd->anon_index = 0;
6197 			ANON_LOCK_EXIT(&amp->a_rwlock);
6198 		}
6199 	}
6200 	if (svd->vp != NULL && szc != 0) {
6201 		struct vattr va;
6202 		u_offset_t eoffpage = svd->offset;
6203 		va.va_mask = AT_SIZE;
6204 		eoffpage += seg->s_size;
6205 		eoffpage = btopr(eoffpage);
6206 		if (VOP_GETATTR(svd->vp, &va, 0, svd->cred) != 0) {
6207 			segvn_setpgsz_getattr_err++;
6208 			return (EINVAL);
6209 		}
6210 		if (btopr(va.va_size) < eoffpage) {
6211 			segvn_setpgsz_eof_err++;
6212 			return (EINVAL);
6213 		}
6214 		if (amp != NULL) {
6215 			/*
6216 			 * anon_fill_cow_holes() may call VOP_GETPAGE().
6217 			 * don't take anon map lock here to avoid holding it
6218 			 * across VOP_GETPAGE() calls that may call back into
6219 			 * segvn for klsutering checks. We don't really need
6220 			 * anon map lock here since it's a private segment and
6221 			 * we hold as level lock as writers.
6222 			 */
6223 			if ((err = anon_fill_cow_holes(seg, seg->s_base,
6224 			    amp->ahp, svd->anon_index, svd->vp, svd->offset,
6225 			    seg->s_size, szc, svd->prot, svd->vpage,
6226 			    svd->cred)) != 0) {
6227 				return (EINVAL);
6228 			}
6229 		}
6230 		segvn_setvnode_mpss(svd->vp);
6231 	}
6232 
6233 	if (amp != NULL) {
6234 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6235 		if (svd->type == MAP_PRIVATE) {
6236 			amp->a_szc = szc;
6237 		} else if (szc > amp->a_szc) {
6238 			amp->a_szc = szc;
6239 		}
6240 		ANON_LOCK_EXIT(&amp->a_rwlock);
6241 	}
6242 
6243 	seg->s_szc = szc;
6244 
6245 	return (0);
6246 }
6247 
6248 static int
6249 segvn_clrszc(struct seg *seg)
6250 {
6251 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6252 	struct anon_map *amp = svd->amp;
6253 	size_t pgsz;
6254 	pgcnt_t pages;
6255 	int err = 0;
6256 	caddr_t a = seg->s_base;
6257 	caddr_t ea = a + seg->s_size;
6258 	ulong_t an_idx = svd->anon_index;
6259 	vnode_t *vp = svd->vp;
6260 	struct vpage *vpage = svd->vpage;
6261 	page_t *anon_pl[1 + 1], *pp;
6262 	struct anon *ap, *oldap;
6263 	uint_t prot = svd->prot, vpprot;
6264 	int pageflag = 0;
6265 
6266 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6267 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
6268 	ASSERT(svd->softlockcnt == 0);
6269 
6270 	if (vp == NULL && amp == NULL) {
6271 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6272 		seg->s_szc = 0;
6273 		return (0);
6274 	}
6275 
6276 	if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
6277 		ASSERT(svd->amp == NULL);
6278 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
6279 		hat_leave_region(seg->s_as->a_hat, svd->rcookie,
6280 		    HAT_REGION_TEXT);
6281 		svd->rcookie = HAT_INVALID_REGION_COOKIE;
6282 	} else if (svd->tr_state == SEGVN_TR_ON) {
6283 		ASSERT(svd->amp != NULL);
6284 		segvn_textunrepl(seg, 1);
6285 		ASSERT(svd->amp == NULL && svd->tr_state == SEGVN_TR_OFF);
6286 		amp = NULL;
6287 	} else {
6288 		if (svd->tr_state != SEGVN_TR_OFF) {
6289 			ASSERT(svd->tr_state == SEGVN_TR_INIT);
6290 			svd->tr_state = SEGVN_TR_OFF;
6291 		}
6292 
6293 		/*
6294 		 * do HAT_UNLOAD_UNMAP since we are changing the pagesize.
6295 		 * unload argument is 0 when we are freeing the segment
6296 		 * and unload was already done.
6297 		 */
6298 		hat_unload(seg->s_as->a_hat, seg->s_base, seg->s_size,
6299 		    HAT_UNLOAD_UNMAP);
6300 	}
6301 
6302 	if (amp == NULL || svd->type == MAP_SHARED) {
6303 		seg->s_szc = 0;
6304 		return (0);
6305 	}
6306 
6307 	pgsz = page_get_pagesize(seg->s_szc);
6308 	pages = btop(pgsz);
6309 
6310 	/*
6311 	 * XXX anon rwlock is not really needed because this is a
6312 	 * private segment and we are writers.
6313 	 */
6314 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_WRITER);
6315 
6316 	for (; a < ea; a += pgsz, an_idx += pages) {
6317 		if ((oldap = anon_get_ptr(amp->ahp, an_idx)) != NULL) {
6318 			ASSERT(vpage != NULL || svd->pageprot == 0);
6319 			if (vpage != NULL) {
6320 				ASSERT(sameprot(seg, a, pgsz));
6321 				prot = VPP_PROT(vpage);
6322 				pageflag = VPP_ISPPLOCK(vpage) ? LOCK_PAGE : 0;
6323 			}
6324 			if (seg->s_szc != 0) {
6325 				ASSERT(vp == NULL || anon_pages(amp->ahp,
6326 				    an_idx, pages) == pages);
6327 				if ((err = anon_map_demotepages(amp, an_idx,
6328 				    seg, a, prot, vpage, svd->cred)) != 0) {
6329 					goto out;
6330 				}
6331 			} else {
6332 				if (oldap->an_refcnt == 1) {
6333 					continue;
6334 				}
6335 				if ((err = anon_getpage(&oldap, &vpprot,
6336 				    anon_pl, PAGESIZE, seg, a, S_READ,
6337 				    svd->cred))) {
6338 					goto out;
6339 				}
6340 				if ((pp = anon_private(&ap, seg, a, prot,
6341 				    anon_pl[0], pageflag, svd->cred)) == NULL) {
6342 					err = ENOMEM;
6343 					goto out;
6344 				}
6345 				anon_decref(oldap);
6346 				(void) anon_set_ptr(amp->ahp, an_idx, ap,
6347 				    ANON_SLEEP);
6348 				page_unlock(pp);
6349 			}
6350 		}
6351 		vpage = (vpage == NULL) ? NULL : vpage + pages;
6352 	}
6353 
6354 	amp->a_szc = 0;
6355 	seg->s_szc = 0;
6356 out:
6357 	ANON_LOCK_EXIT(&amp->a_rwlock);
6358 	return (err);
6359 }
6360 
6361 static int
6362 segvn_claim_pages(
6363 	struct seg *seg,
6364 	struct vpage *svp,
6365 	u_offset_t off,
6366 	ulong_t anon_idx,
6367 	uint_t prot)
6368 {
6369 	pgcnt_t	pgcnt = page_get_pagecnt(seg->s_szc);
6370 	size_t ppasize = (pgcnt + 1) * sizeof (page_t *);
6371 	page_t	**ppa;
6372 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6373 	struct anon_map *amp = svd->amp;
6374 	struct vpage *evp = svp + pgcnt;
6375 	caddr_t addr = ((uintptr_t)(svp - svd->vpage) << PAGESHIFT)
6376 	    + seg->s_base;
6377 	struct anon *ap;
6378 	struct vnode *vp = svd->vp;
6379 	page_t *pp;
6380 	pgcnt_t pg_idx, i;
6381 	int err = 0;
6382 	anoff_t aoff;
6383 	int anon = (amp != NULL) ? 1 : 0;
6384 
6385 	ASSERT(svd->type == MAP_PRIVATE);
6386 	ASSERT(svd->vpage != NULL);
6387 	ASSERT(seg->s_szc != 0);
6388 	ASSERT(IS_P2ALIGNED(pgcnt, pgcnt));
6389 	ASSERT(amp == NULL || IS_P2ALIGNED(anon_idx, pgcnt));
6390 	ASSERT(sameprot(seg, addr, pgcnt << PAGESHIFT));
6391 
6392 	if (VPP_PROT(svp) == prot)
6393 		return (1);
6394 	if (!((VPP_PROT(svp) ^ prot) & PROT_WRITE))
6395 		return (1);
6396 
6397 	ppa = kmem_alloc(ppasize, KM_SLEEP);
6398 	if (anon && vp != NULL) {
6399 		if (anon_get_ptr(amp->ahp, anon_idx) == NULL) {
6400 			anon = 0;
6401 			ASSERT(!anon_pages(amp->ahp, anon_idx, pgcnt));
6402 		}
6403 		ASSERT(!anon ||
6404 		    anon_pages(amp->ahp, anon_idx, pgcnt) == pgcnt);
6405 	}
6406 
6407 	for (*ppa = NULL, pg_idx = 0; svp < evp; svp++, anon_idx++) {
6408 		if (!VPP_ISPPLOCK(svp))
6409 			continue;
6410 		if (anon) {
6411 			ap = anon_get_ptr(amp->ahp, anon_idx);
6412 			if (ap == NULL) {
6413 				panic("segvn_claim_pages: no anon slot");
6414 			}
6415 			swap_xlate(ap, &vp, &aoff);
6416 			off = (u_offset_t)aoff;
6417 		}
6418 		ASSERT(vp != NULL);
6419 		if ((pp = page_lookup(vp,
6420 		    (u_offset_t)off, SE_SHARED)) == NULL) {
6421 			panic("segvn_claim_pages: no page");
6422 		}
6423 		ppa[pg_idx++] = pp;
6424 		off += PAGESIZE;
6425 	}
6426 
6427 	if (ppa[0] == NULL) {
6428 		kmem_free(ppa, ppasize);
6429 		return (1);
6430 	}
6431 
6432 	ASSERT(pg_idx <= pgcnt);
6433 	ppa[pg_idx] = NULL;
6434 
6435 	if (prot & PROT_WRITE)
6436 		err = page_addclaim_pages(ppa);
6437 	else
6438 		err = page_subclaim_pages(ppa);
6439 
6440 	for (i = 0; i < pg_idx; i++) {
6441 		ASSERT(ppa[i] != NULL);
6442 		page_unlock(ppa[i]);
6443 	}
6444 
6445 	kmem_free(ppa, ppasize);
6446 	return (err);
6447 }
6448 
6449 /*
6450  * Returns right (upper address) segment if split occured.
6451  * If the address is equal to the beginning or end of its segment it returns
6452  * the current segment.
6453  */
6454 static struct seg *
6455 segvn_split_seg(struct seg *seg, caddr_t addr)
6456 {
6457 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6458 	struct seg *nseg;
6459 	size_t nsize;
6460 	struct segvn_data *nsvd;
6461 
6462 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6463 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6464 
6465 	ASSERT(addr >= seg->s_base);
6466 	ASSERT(addr <= seg->s_base + seg->s_size);
6467 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6468 
6469 	if (addr == seg->s_base || addr == seg->s_base + seg->s_size)
6470 		return (seg);
6471 
6472 	nsize = seg->s_base + seg->s_size - addr;
6473 	seg->s_size = addr - seg->s_base;
6474 	nseg = seg_alloc(seg->s_as, addr, nsize);
6475 	ASSERT(nseg != NULL);
6476 	nseg->s_ops = seg->s_ops;
6477 	nsvd = kmem_cache_alloc(segvn_cache, KM_SLEEP);
6478 	nseg->s_data = (void *)nsvd;
6479 	nseg->s_szc = seg->s_szc;
6480 	*nsvd = *svd;
6481 	ASSERT(nsvd->rcookie == HAT_INVALID_REGION_COOKIE);
6482 	nsvd->seg = nseg;
6483 	rw_init(&nsvd->lock, NULL, RW_DEFAULT, NULL);
6484 
6485 	if (nsvd->vp != NULL) {
6486 		VN_HOLD(nsvd->vp);
6487 		nsvd->offset = svd->offset +
6488 		    (uintptr_t)(nseg->s_base - seg->s_base);
6489 		if (nsvd->type == MAP_SHARED)
6490 			lgrp_shm_policy_init(NULL, nsvd->vp);
6491 	} else {
6492 		/*
6493 		 * The offset for an anonymous segment has no signifigance in
6494 		 * terms of an offset into a file. If we were to use the above
6495 		 * calculation instead, the structures read out of
6496 		 * /proc/<pid>/xmap would be more difficult to decipher since
6497 		 * it would be unclear whether two seemingly contiguous
6498 		 * prxmap_t structures represented different segments or a
6499 		 * single segment that had been split up into multiple prxmap_t
6500 		 * structures (e.g. if some part of the segment had not yet
6501 		 * been faulted in).
6502 		 */
6503 		nsvd->offset = 0;
6504 	}
6505 
6506 	ASSERT(svd->softlockcnt == 0);
6507 	crhold(svd->cred);
6508 
6509 	if (svd->vpage != NULL) {
6510 		size_t bytes = vpgtob(seg_pages(seg));
6511 		size_t nbytes = vpgtob(seg_pages(nseg));
6512 		struct vpage *ovpage = svd->vpage;
6513 
6514 		svd->vpage = kmem_alloc(bytes, KM_SLEEP);
6515 		bcopy(ovpage, svd->vpage, bytes);
6516 		nsvd->vpage = kmem_alloc(nbytes, KM_SLEEP);
6517 		bcopy(ovpage + seg_pages(seg), nsvd->vpage, nbytes);
6518 		kmem_free(ovpage, bytes + nbytes);
6519 	}
6520 	if (svd->amp != NULL && svd->type == MAP_PRIVATE) {
6521 		struct anon_map *oamp = svd->amp, *namp;
6522 		struct anon_hdr *nahp;
6523 
6524 		ANON_LOCK_ENTER(&oamp->a_rwlock, RW_WRITER);
6525 		ASSERT(oamp->refcnt == 1);
6526 		nahp = anon_create(btop(seg->s_size), ANON_SLEEP);
6527 		(void) anon_copy_ptr(oamp->ahp, svd->anon_index,
6528 		    nahp, 0, btop(seg->s_size), ANON_SLEEP);
6529 
6530 		namp = anonmap_alloc(nseg->s_size, 0, ANON_SLEEP);
6531 		namp->a_szc = nseg->s_szc;
6532 		(void) anon_copy_ptr(oamp->ahp,
6533 		    svd->anon_index + btop(seg->s_size),
6534 		    namp->ahp, 0, btop(nseg->s_size), ANON_SLEEP);
6535 		anon_release(oamp->ahp, btop(oamp->size));
6536 		oamp->ahp = nahp;
6537 		oamp->size = seg->s_size;
6538 		svd->anon_index = 0;
6539 		nsvd->amp = namp;
6540 		nsvd->anon_index = 0;
6541 		ANON_LOCK_EXIT(&oamp->a_rwlock);
6542 	} else if (svd->amp != NULL) {
6543 		pgcnt_t pgcnt = page_get_pagecnt(seg->s_szc);
6544 		ASSERT(svd->amp == nsvd->amp);
6545 		ASSERT(seg->s_szc <= svd->amp->a_szc);
6546 		nsvd->anon_index = svd->anon_index + seg_pages(seg);
6547 		ASSERT(IS_P2ALIGNED(nsvd->anon_index, pgcnt));
6548 		ANON_LOCK_ENTER(&svd->amp->a_rwlock, RW_WRITER);
6549 		svd->amp->refcnt++;
6550 		ANON_LOCK_EXIT(&svd->amp->a_rwlock);
6551 	}
6552 
6553 	/*
6554 	 * Split amount of swap reserve
6555 	 */
6556 	if (svd->swresv) {
6557 		/*
6558 		 * For MAP_NORESERVE, only allocate swap reserve for pages
6559 		 * being used.  Other segments get enough to cover whole
6560 		 * segment.
6561 		 */
6562 		if (svd->flags & MAP_NORESERVE) {
6563 			size_t	oswresv;
6564 
6565 			ASSERT(svd->amp);
6566 			oswresv = svd->swresv;
6567 			svd->swresv = ptob(anon_pages(svd->amp->ahp,
6568 				svd->anon_index, btop(seg->s_size)));
6569 			nsvd->swresv = ptob(anon_pages(nsvd->amp->ahp,
6570 				nsvd->anon_index, btop(nseg->s_size)));
6571 			ASSERT(oswresv >= (svd->swresv + nsvd->swresv));
6572 		} else {
6573 			ASSERT(svd->swresv == seg->s_size + nseg->s_size);
6574 			svd->swresv = seg->s_size;
6575 			nsvd->swresv = nseg->s_size;
6576 		}
6577 	}
6578 
6579 	return (nseg);
6580 }
6581 
6582 /*
6583  * called on memory operations (unmap, setprot, setpagesize) for a subset
6584  * of a large page segment to either demote the memory range (SDR_RANGE)
6585  * or the ends (SDR_END) by addr/len.
6586  *
6587  * returns 0 on success. returns errno, including ENOMEM, on failure.
6588  */
6589 static int
6590 segvn_demote_range(
6591 	struct seg *seg,
6592 	caddr_t addr,
6593 	size_t len,
6594 	int flag,
6595 	uint_t szcvec)
6596 {
6597 	caddr_t eaddr = addr + len;
6598 	caddr_t lpgaddr, lpgeaddr;
6599 	struct seg *nseg;
6600 	struct seg *badseg1 = NULL;
6601 	struct seg *badseg2 = NULL;
6602 	size_t pgsz;
6603 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6604 	int err;
6605 	uint_t szc = seg->s_szc;
6606 	uint_t tszcvec;
6607 
6608 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock));
6609 	ASSERT(svd->tr_state == SEGVN_TR_OFF);
6610 	ASSERT(szc != 0);
6611 	pgsz = page_get_pagesize(szc);
6612 	ASSERT(seg->s_base != addr || seg->s_size != len);
6613 	ASSERT(addr >= seg->s_base && eaddr <= seg->s_base + seg->s_size);
6614 	ASSERT(svd->softlockcnt == 0);
6615 	ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
6616 	ASSERT(szcvec == 0 || (flag == SDR_END && svd->type == MAP_SHARED));
6617 
6618 	CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
6619 	ASSERT(flag == SDR_RANGE || eaddr < lpgeaddr || addr > lpgaddr);
6620 	if (flag == SDR_RANGE) {
6621 		/* demote entire range */
6622 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6623 		(void) segvn_split_seg(nseg, lpgeaddr);
6624 		ASSERT(badseg1->s_base == lpgaddr);
6625 		ASSERT(badseg1->s_size == lpgeaddr - lpgaddr);
6626 	} else if (addr != lpgaddr) {
6627 		ASSERT(flag == SDR_END);
6628 		badseg1 = nseg = segvn_split_seg(seg, lpgaddr);
6629 		if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz &&
6630 		    eaddr < lpgaddr + 2 * pgsz) {
6631 			(void) segvn_split_seg(nseg, lpgeaddr);
6632 			ASSERT(badseg1->s_base == lpgaddr);
6633 			ASSERT(badseg1->s_size == 2 * pgsz);
6634 		} else {
6635 			nseg = segvn_split_seg(nseg, lpgaddr + pgsz);
6636 			ASSERT(badseg1->s_base == lpgaddr);
6637 			ASSERT(badseg1->s_size == pgsz);
6638 			if (eaddr != lpgeaddr && eaddr > lpgaddr + pgsz) {
6639 				ASSERT(lpgeaddr - lpgaddr > 2 * pgsz);
6640 				nseg = segvn_split_seg(nseg, lpgeaddr - pgsz);
6641 				badseg2 = nseg;
6642 				(void) segvn_split_seg(nseg, lpgeaddr);
6643 				ASSERT(badseg2->s_base == lpgeaddr - pgsz);
6644 				ASSERT(badseg2->s_size == pgsz);
6645 			}
6646 		}
6647 	} else {
6648 		ASSERT(flag == SDR_END);
6649 		ASSERT(eaddr < lpgeaddr);
6650 		badseg1 = nseg = segvn_split_seg(seg, lpgeaddr - pgsz);
6651 		(void) segvn_split_seg(nseg, lpgeaddr);
6652 		ASSERT(badseg1->s_base == lpgeaddr - pgsz);
6653 		ASSERT(badseg1->s_size == pgsz);
6654 	}
6655 
6656 	ASSERT(badseg1 != NULL);
6657 	ASSERT(badseg1->s_szc == szc);
6658 	ASSERT(flag == SDR_RANGE || badseg1->s_size == pgsz ||
6659 	    badseg1->s_size == 2 * pgsz);
6660 	ASSERT(sameprot(badseg1, badseg1->s_base, pgsz));
6661 	ASSERT(badseg1->s_size == pgsz ||
6662 	    sameprot(badseg1, badseg1->s_base + pgsz, pgsz));
6663 	if (err = segvn_clrszc(badseg1)) {
6664 		return (err);
6665 	}
6666 	ASSERT(badseg1->s_szc == 0);
6667 
6668 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6669 		uint_t tszc = highbit(tszcvec) - 1;
6670 		caddr_t ta = MAX(addr, badseg1->s_base);
6671 		caddr_t te;
6672 		size_t tpgsz = page_get_pagesize(tszc);
6673 
6674 		ASSERT(svd->type == MAP_SHARED);
6675 		ASSERT(flag == SDR_END);
6676 		ASSERT(tszc < szc && tszc > 0);
6677 
6678 		if (eaddr > badseg1->s_base + badseg1->s_size) {
6679 			te = badseg1->s_base + badseg1->s_size;
6680 		} else {
6681 			te = eaddr;
6682 		}
6683 
6684 		ASSERT(ta <= te);
6685 		badseg1->s_szc = tszc;
6686 		if (!IS_P2ALIGNED(ta, tpgsz) || !IS_P2ALIGNED(te, tpgsz)) {
6687 			if (badseg2 != NULL) {
6688 				err = segvn_demote_range(badseg1, ta, te - ta,
6689 				    SDR_END, tszcvec);
6690 				if (err != 0) {
6691 					return (err);
6692 				}
6693 			} else {
6694 				return (segvn_demote_range(badseg1, ta,
6695 				    te - ta, SDR_END, tszcvec));
6696 			}
6697 		}
6698 	}
6699 
6700 	if (badseg2 == NULL)
6701 		return (0);
6702 	ASSERT(badseg2->s_szc == szc);
6703 	ASSERT(badseg2->s_size == pgsz);
6704 	ASSERT(sameprot(badseg2, badseg2->s_base, badseg2->s_size));
6705 	if (err = segvn_clrszc(badseg2)) {
6706 		return (err);
6707 	}
6708 	ASSERT(badseg2->s_szc == 0);
6709 
6710 	if (szc > 1 && (tszcvec = P2PHASE(szcvec, 1 << szc)) > 1) {
6711 		uint_t tszc = highbit(tszcvec) - 1;
6712 		size_t tpgsz = page_get_pagesize(tszc);
6713 
6714 		ASSERT(svd->type == MAP_SHARED);
6715 		ASSERT(flag == SDR_END);
6716 		ASSERT(tszc < szc && tszc > 0);
6717 		ASSERT(badseg2->s_base > addr);
6718 		ASSERT(eaddr > badseg2->s_base);
6719 		ASSERT(eaddr < badseg2->s_base + badseg2->s_size);
6720 
6721 		badseg2->s_szc = tszc;
6722 		if (!IS_P2ALIGNED(eaddr, tpgsz)) {
6723 			return (segvn_demote_range(badseg2, badseg2->s_base,
6724 			    eaddr - badseg2->s_base, SDR_END, tszcvec));
6725 		}
6726 	}
6727 
6728 	return (0);
6729 }
6730 
6731 static int
6732 segvn_checkprot(struct seg *seg, caddr_t addr, size_t len, uint_t prot)
6733 {
6734 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6735 	struct vpage *vp, *evp;
6736 
6737 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6738 
6739 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6740 	/*
6741 	 * If segment protection can be used, simply check against them.
6742 	 */
6743 	if (svd->pageprot == 0) {
6744 		int err;
6745 
6746 		err = ((svd->prot & prot) != prot) ? EACCES : 0;
6747 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6748 		return (err);
6749 	}
6750 
6751 	/*
6752 	 * Have to check down to the vpage level.
6753 	 */
6754 	evp = &svd->vpage[seg_page(seg, addr + len)];
6755 	for (vp = &svd->vpage[seg_page(seg, addr)]; vp < evp; vp++) {
6756 		if ((VPP_PROT(vp) & prot) != prot) {
6757 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6758 			return (EACCES);
6759 		}
6760 	}
6761 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6762 	return (0);
6763 }
6764 
6765 static int
6766 segvn_getprot(struct seg *seg, caddr_t addr, size_t len, uint_t *protv)
6767 {
6768 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6769 	size_t pgno = seg_page(seg, addr + len) - seg_page(seg, addr) + 1;
6770 
6771 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6772 
6773 	if (pgno != 0) {
6774 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6775 		if (svd->pageprot == 0) {
6776 			do
6777 				protv[--pgno] = svd->prot;
6778 			while (pgno != 0);
6779 		} else {
6780 			size_t pgoff = seg_page(seg, addr);
6781 
6782 			do {
6783 				pgno--;
6784 				protv[pgno] = VPP_PROT(&svd->vpage[pgno+pgoff]);
6785 			} while (pgno != 0);
6786 		}
6787 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
6788 	}
6789 	return (0);
6790 }
6791 
6792 static u_offset_t
6793 segvn_getoffset(struct seg *seg, caddr_t addr)
6794 {
6795 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6796 
6797 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6798 
6799 	return (svd->offset + (uintptr_t)(addr - seg->s_base));
6800 }
6801 
6802 /*ARGSUSED*/
6803 static int
6804 segvn_gettype(struct seg *seg, caddr_t addr)
6805 {
6806 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6807 
6808 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6809 
6810 	return (svd->type | (svd->flags & (MAP_NORESERVE | MAP_TEXT |
6811 	    MAP_INITDATA)));
6812 }
6813 
6814 /*ARGSUSED*/
6815 static int
6816 segvn_getvp(struct seg *seg, caddr_t addr, struct vnode **vpp)
6817 {
6818 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6819 
6820 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6821 
6822 	*vpp = svd->vp;
6823 	return (0);
6824 }
6825 
6826 /*
6827  * Check to see if it makes sense to do kluster/read ahead to
6828  * addr + delta relative to the mapping at addr.  We assume here
6829  * that delta is a signed PAGESIZE'd multiple (which can be negative).
6830  *
6831  * For segvn, we currently "approve" of the action if we are
6832  * still in the segment and it maps from the same vp/off,
6833  * or if the advice stored in segvn_data or vpages allows it.
6834  * Currently, klustering is not allowed only if MADV_RANDOM is set.
6835  */
6836 static int
6837 segvn_kluster(struct seg *seg, caddr_t addr, ssize_t delta)
6838 {
6839 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6840 	struct anon *oap, *ap;
6841 	ssize_t pd;
6842 	size_t page;
6843 	struct vnode *vp1, *vp2;
6844 	u_offset_t off1, off2;
6845 	struct anon_map *amp;
6846 
6847 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6848 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
6849 	    SEGVN_LOCK_HELD(seg->s_as, &svd->lock));
6850 
6851 	if (addr + delta < seg->s_base ||
6852 	    addr + delta >= (seg->s_base + seg->s_size))
6853 		return (-1);		/* exceeded segment bounds */
6854 
6855 	pd = delta / (ssize_t)PAGESIZE;	/* divide to preserve sign bit */
6856 	page = seg_page(seg, addr);
6857 
6858 	/*
6859 	 * Check to see if either of the pages addr or addr + delta
6860 	 * have advice set that prevents klustering (if MADV_RANDOM advice
6861 	 * is set for entire segment, or MADV_SEQUENTIAL is set and delta
6862 	 * is negative).
6863 	 */
6864 	if (svd->advice == MADV_RANDOM ||
6865 	    svd->advice == MADV_SEQUENTIAL && delta < 0)
6866 		return (-1);
6867 	else if (svd->pageadvice && svd->vpage) {
6868 		struct vpage *bvpp, *evpp;
6869 
6870 		bvpp = &svd->vpage[page];
6871 		evpp = &svd->vpage[page + pd];
6872 		if (VPP_ADVICE(bvpp) == MADV_RANDOM ||
6873 		    VPP_ADVICE(evpp) == MADV_SEQUENTIAL && delta < 0)
6874 			return (-1);
6875 		if (VPP_ADVICE(bvpp) != VPP_ADVICE(evpp) &&
6876 		    VPP_ADVICE(evpp) == MADV_RANDOM)
6877 			return (-1);
6878 	}
6879 
6880 	if (svd->type == MAP_SHARED)
6881 		return (0);		/* shared mapping - all ok */
6882 
6883 	if ((amp = svd->amp) == NULL)
6884 		return (0);		/* off original vnode */
6885 
6886 	page += svd->anon_index;
6887 
6888 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
6889 
6890 	oap = anon_get_ptr(amp->ahp, page);
6891 	ap = anon_get_ptr(amp->ahp, page + pd);
6892 
6893 	ANON_LOCK_EXIT(&amp->a_rwlock);
6894 
6895 	if ((oap == NULL && ap != NULL) || (oap != NULL && ap == NULL)) {
6896 		return (-1);		/* one with and one without an anon */
6897 	}
6898 
6899 	if (oap == NULL) {		/* implies that ap == NULL */
6900 		return (0);		/* off original vnode */
6901 	}
6902 
6903 	/*
6904 	 * Now we know we have two anon pointers - check to
6905 	 * see if they happen to be properly allocated.
6906 	 */
6907 
6908 	/*
6909 	 * XXX We cheat here and don't lock the anon slots. We can't because
6910 	 * we may have been called from the anon layer which might already
6911 	 * have locked them. We are holding a refcnt on the slots so they
6912 	 * can't disappear. The worst that will happen is we'll get the wrong
6913 	 * names (vp, off) for the slots and make a poor klustering decision.
6914 	 */
6915 	swap_xlate(ap, &vp1, &off1);
6916 	swap_xlate(oap, &vp2, &off2);
6917 
6918 
6919 	if (!VOP_CMP(vp1, vp2) || off1 - off2 != delta)
6920 		return (-1);
6921 	return (0);
6922 }
6923 
6924 /*
6925  * Swap the pages of seg out to secondary storage, returning the
6926  * number of bytes of storage freed.
6927  *
6928  * The basic idea is first to unload all translations and then to call
6929  * VOP_PUTPAGE() for all newly-unmapped pages, to push them out to the
6930  * swap device.  Pages to which other segments have mappings will remain
6931  * mapped and won't be swapped.  Our caller (as_swapout) has already
6932  * performed the unloading step.
6933  *
6934  * The value returned is intended to correlate well with the process's
6935  * memory requirements.  However, there are some caveats:
6936  * 1)	When given a shared segment as argument, this routine will
6937  *	only succeed in swapping out pages for the last sharer of the
6938  *	segment.  (Previous callers will only have decremented mapping
6939  *	reference counts.)
6940  * 2)	We assume that the hat layer maintains a large enough translation
6941  *	cache to capture process reference patterns.
6942  */
6943 static size_t
6944 segvn_swapout(struct seg *seg)
6945 {
6946 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
6947 	struct anon_map *amp;
6948 	pgcnt_t pgcnt = 0;
6949 	pgcnt_t npages;
6950 	pgcnt_t page;
6951 	ulong_t anon_index;
6952 
6953 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
6954 
6955 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
6956 	/*
6957 	 * Find pages unmapped by our caller and force them
6958 	 * out to the virtual swap device.
6959 	 */
6960 	if ((amp = svd->amp) != NULL)
6961 		anon_index = svd->anon_index;
6962 	npages = seg->s_size >> PAGESHIFT;
6963 	for (page = 0; page < npages; page++) {
6964 		page_t *pp;
6965 		struct anon *ap;
6966 		struct vnode *vp;
6967 		u_offset_t off;
6968 		anon_sync_obj_t cookie;
6969 
6970 		/*
6971 		 * Obtain <vp, off> pair for the page, then look it up.
6972 		 *
6973 		 * Note that this code is willing to consider regular
6974 		 * pages as well as anon pages.  Is this appropriate here?
6975 		 */
6976 		ap = NULL;
6977 		if (amp != NULL) {
6978 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
6979 			if (anon_array_try_enter(amp, anon_index + page,
6980 						&cookie)) {
6981 				ANON_LOCK_EXIT(&amp->a_rwlock);
6982 				continue;
6983 			}
6984 			ap = anon_get_ptr(amp->ahp, anon_index + page);
6985 			if (ap != NULL) {
6986 				swap_xlate(ap, &vp, &off);
6987 			} else {
6988 				vp = svd->vp;
6989 				off = svd->offset + ptob(page);
6990 			}
6991 			anon_array_exit(&cookie);
6992 			ANON_LOCK_EXIT(&amp->a_rwlock);
6993 		} else {
6994 			vp = svd->vp;
6995 			off = svd->offset + ptob(page);
6996 		}
6997 		if (vp == NULL) {		/* untouched zfod page */
6998 			ASSERT(ap == NULL);
6999 			continue;
7000 		}
7001 
7002 		pp = page_lookup_nowait(vp, off, SE_SHARED);
7003 		if (pp == NULL)
7004 			continue;
7005 
7006 
7007 		/*
7008 		 * Examine the page to see whether it can be tossed out,
7009 		 * keeping track of how many we've found.
7010 		 */
7011 		if (!page_tryupgrade(pp)) {
7012 			/*
7013 			 * If the page has an i/o lock and no mappings,
7014 			 * it's very likely that the page is being
7015 			 * written out as a result of klustering.
7016 			 * Assume this is so and take credit for it here.
7017 			 */
7018 			if (!page_io_trylock(pp)) {
7019 				if (!hat_page_is_mapped(pp))
7020 					pgcnt++;
7021 			} else {
7022 				page_io_unlock(pp);
7023 			}
7024 			page_unlock(pp);
7025 			continue;
7026 		}
7027 		ASSERT(!page_iolock_assert(pp));
7028 
7029 
7030 		/*
7031 		 * Skip if page is locked or has mappings.
7032 		 * We don't need the page_struct_lock to look at lckcnt
7033 		 * and cowcnt because the page is exclusive locked.
7034 		 */
7035 		if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0 ||
7036 		    hat_page_is_mapped(pp)) {
7037 			page_unlock(pp);
7038 			continue;
7039 		}
7040 
7041 		/*
7042 		 * dispose skips large pages so try to demote first.
7043 		 */
7044 		if (pp->p_szc != 0 && !page_try_demote_pages(pp)) {
7045 			page_unlock(pp);
7046 			/*
7047 			 * XXX should skip the remaining page_t's of this
7048 			 * large page.
7049 			 */
7050 			continue;
7051 		}
7052 
7053 		ASSERT(pp->p_szc == 0);
7054 
7055 		/*
7056 		 * No longer mapped -- we can toss it out.  How
7057 		 * we do so depends on whether or not it's dirty.
7058 		 */
7059 		if (hat_ismod(pp) && pp->p_vnode) {
7060 			/*
7061 			 * We must clean the page before it can be
7062 			 * freed.  Setting B_FREE will cause pvn_done
7063 			 * to free the page when the i/o completes.
7064 			 * XXX:	This also causes it to be accounted
7065 			 *	as a pageout instead of a swap: need
7066 			 *	B_SWAPOUT bit to use instead of B_FREE.
7067 			 *
7068 			 * Hold the vnode before releasing the page lock
7069 			 * to prevent it from being freed and re-used by
7070 			 * some other thread.
7071 			 */
7072 			VN_HOLD(vp);
7073 			page_unlock(pp);
7074 
7075 			/*
7076 			 * Queue all i/o requests for the pageout thread
7077 			 * to avoid saturating the pageout devices.
7078 			 */
7079 			if (!queue_io_request(vp, off))
7080 				VN_RELE(vp);
7081 		} else {
7082 			/*
7083 			 * The page was clean, free it.
7084 			 *
7085 			 * XXX:	Can we ever encounter modified pages
7086 			 *	with no associated vnode here?
7087 			 */
7088 			ASSERT(pp->p_vnode != NULL);
7089 			/*LINTED: constant in conditional context*/
7090 			VN_DISPOSE(pp, B_FREE, 0, kcred);
7091 		}
7092 
7093 		/*
7094 		 * Credit now even if i/o is in progress.
7095 		 */
7096 		pgcnt++;
7097 	}
7098 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7099 
7100 	/*
7101 	 * Wakeup pageout to initiate i/o on all queued requests.
7102 	 */
7103 	cv_signal_pageout();
7104 	return (ptob(pgcnt));
7105 }
7106 
7107 /*
7108  * Synchronize primary storage cache with real object in virtual memory.
7109  *
7110  * XXX - Anonymous pages should not be sync'ed out at all.
7111  */
7112 static int
7113 segvn_sync(struct seg *seg, caddr_t addr, size_t len, int attr, uint_t flags)
7114 {
7115 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7116 	struct vpage *vpp;
7117 	page_t *pp;
7118 	u_offset_t offset;
7119 	struct vnode *vp;
7120 	u_offset_t off;
7121 	caddr_t eaddr;
7122 	int bflags;
7123 	int err = 0;
7124 	int segtype;
7125 	int pageprot;
7126 	int prot;
7127 	ulong_t anon_index;
7128 	struct anon_map *amp;
7129 	struct anon *ap;
7130 	anon_sync_obj_t cookie;
7131 
7132 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7133 
7134 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7135 
7136 	if (svd->softlockcnt > 0) {
7137 		/*
7138 		 * flush all pages from seg cache
7139 		 * otherwise we may deadlock in swap_putpage
7140 		 * for B_INVAL page (4175402).
7141 		 *
7142 		 * Even if we grab segvn WRITER's lock or segp_slock
7143 		 * here, there might be another thread which could've
7144 		 * successfully performed lookup/insert just before
7145 		 * we acquired the lock here.  So, grabbing either
7146 		 * lock here is of not much use.  Until we devise
7147 		 * a strategy at upper layers to solve the
7148 		 * synchronization issues completely, we expect
7149 		 * applications to handle this appropriately.
7150 		 */
7151 		segvn_purge(seg);
7152 		if (svd->softlockcnt > 0) {
7153 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7154 			return (EAGAIN);
7155 		}
7156 	}
7157 
7158 	vpp = svd->vpage;
7159 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7160 	bflags = ((flags & MS_ASYNC) ? B_ASYNC : 0) |
7161 	    ((flags & MS_INVALIDATE) ? B_INVAL : 0);
7162 
7163 	if (attr) {
7164 		pageprot = attr & ~(SHARED|PRIVATE);
7165 		segtype = (attr & SHARED) ? MAP_SHARED : MAP_PRIVATE;
7166 
7167 		/*
7168 		 * We are done if the segment types don't match
7169 		 * or if we have segment level protections and
7170 		 * they don't match.
7171 		 */
7172 		if (svd->type != segtype) {
7173 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7174 			return (0);
7175 		}
7176 		if (vpp == NULL) {
7177 			if (svd->prot != pageprot) {
7178 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7179 				return (0);
7180 			}
7181 			prot = svd->prot;
7182 		} else
7183 			vpp = &svd->vpage[seg_page(seg, addr)];
7184 
7185 	} else if (svd->vp && svd->amp == NULL &&
7186 	    (flags & MS_INVALIDATE) == 0) {
7187 
7188 		/*
7189 		 * No attributes, no anonymous pages and MS_INVALIDATE flag
7190 		 * is not on, just use one big request.
7191 		 */
7192 		err = VOP_PUTPAGE(svd->vp, (offset_t)offset, len,
7193 		    bflags, svd->cred);
7194 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7195 		return (err);
7196 	}
7197 
7198 	if ((amp = svd->amp) != NULL)
7199 		anon_index = svd->anon_index + seg_page(seg, addr);
7200 
7201 	for (eaddr = addr + len; addr < eaddr; addr += PAGESIZE) {
7202 		ap = NULL;
7203 		if (amp != NULL) {
7204 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7205 			anon_array_enter(amp, anon_index, &cookie);
7206 			ap = anon_get_ptr(amp->ahp, anon_index++);
7207 			if (ap != NULL) {
7208 				swap_xlate(ap, &vp, &off);
7209 			} else {
7210 				vp = svd->vp;
7211 				off = offset;
7212 			}
7213 			anon_array_exit(&cookie);
7214 			ANON_LOCK_EXIT(&amp->a_rwlock);
7215 		} else {
7216 			vp = svd->vp;
7217 			off = offset;
7218 		}
7219 		offset += PAGESIZE;
7220 
7221 		if (vp == NULL)		/* untouched zfod page */
7222 			continue;
7223 
7224 		if (attr) {
7225 			if (vpp) {
7226 				prot = VPP_PROT(vpp);
7227 				vpp++;
7228 			}
7229 			if (prot != pageprot) {
7230 				continue;
7231 			}
7232 		}
7233 
7234 		/*
7235 		 * See if any of these pages are locked --  if so, then we
7236 		 * will have to truncate an invalidate request at the first
7237 		 * locked one. We don't need the page_struct_lock to test
7238 		 * as this is only advisory; even if we acquire it someone
7239 		 * might race in and lock the page after we unlock and before
7240 		 * we do the PUTPAGE, then PUTPAGE simply does nothing.
7241 		 */
7242 		if (flags & MS_INVALIDATE) {
7243 			if ((pp = page_lookup(vp, off, SE_SHARED)) != NULL) {
7244 				if (pp->p_lckcnt != 0 || pp->p_cowcnt != 0) {
7245 					page_unlock(pp);
7246 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7247 					return (EBUSY);
7248 				}
7249 				if (ap != NULL && pp->p_szc != 0 &&
7250 				    page_tryupgrade(pp)) {
7251 					if (pp->p_lckcnt == 0 &&
7252 					    pp->p_cowcnt == 0) {
7253 						/*
7254 						 * swapfs VN_DISPOSE() won't
7255 						 * invalidate large pages.
7256 						 * Attempt to demote.
7257 						 * XXX can't help it if it
7258 						 * fails. But for swapfs
7259 						 * pages it is no big deal.
7260 						 */
7261 						(void) page_try_demote_pages(
7262 						    pp);
7263 				    }
7264 				}
7265 				page_unlock(pp);
7266 			}
7267 		} else if (svd->type == MAP_SHARED && amp != NULL) {
7268 			/*
7269 			 * Avoid writting out to disk ISM's large pages
7270 			 * because segspt_free_pages() relies on NULL an_pvp
7271 			 * of anon slots of such pages.
7272 			 */
7273 
7274 			ASSERT(svd->vp == NULL);
7275 			/*
7276 			 * swapfs uses page_lookup_nowait if not freeing or
7277 			 * invalidating and skips a page if
7278 			 * page_lookup_nowait returns NULL.
7279 			 */
7280 			pp = page_lookup_nowait(vp, off, SE_SHARED);
7281 			if (pp == NULL) {
7282 				continue;
7283 			}
7284 			if (pp->p_szc != 0) {
7285 				page_unlock(pp);
7286 				continue;
7287 			}
7288 
7289 			/*
7290 			 * Note ISM pages are created large so (vp, off)'s
7291 			 * page cannot suddenly become large after we unlock
7292 			 * pp.
7293 			 */
7294 			page_unlock(pp);
7295 		}
7296 		/*
7297 		 * XXX - Should ultimately try to kluster
7298 		 * calls to VOP_PUTPAGE() for performance.
7299 		 */
7300 		VN_HOLD(vp);
7301 		err = VOP_PUTPAGE(vp, (offset_t)off, PAGESIZE,
7302 		    bflags, svd->cred);
7303 		VN_RELE(vp);
7304 		if (err)
7305 			break;
7306 	}
7307 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7308 	return (err);
7309 }
7310 
7311 /*
7312  * Determine if we have data corresponding to pages in the
7313  * primary storage virtual memory cache (i.e., "in core").
7314  */
7315 static size_t
7316 segvn_incore(struct seg *seg, caddr_t addr, size_t len, char *vec)
7317 {
7318 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7319 	struct vnode *vp, *avp;
7320 	u_offset_t offset, aoffset;
7321 	size_t p, ep;
7322 	int ret;
7323 	struct vpage *vpp;
7324 	page_t *pp;
7325 	uint_t start;
7326 	struct anon_map *amp;		/* XXX - for locknest */
7327 	struct anon *ap;
7328 	uint_t attr;
7329 	anon_sync_obj_t cookie;
7330 
7331 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7332 
7333 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7334 	if (svd->amp == NULL && svd->vp == NULL) {
7335 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7336 		bzero(vec, btopr(len));
7337 		return (len);	/* no anonymous pages created yet */
7338 	}
7339 
7340 	p = seg_page(seg, addr);
7341 	ep = seg_page(seg, addr + len);
7342 	start = svd->vp ? SEG_PAGE_VNODEBACKED : 0;
7343 
7344 	amp = svd->amp;
7345 	for (; p < ep; p++, addr += PAGESIZE) {
7346 		vpp = (svd->vpage) ? &svd->vpage[p]: NULL;
7347 		ret = start;
7348 		ap = NULL;
7349 		avp = NULL;
7350 		/* Grab the vnode/offset for the anon slot */
7351 		if (amp != NULL) {
7352 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7353 			anon_array_enter(amp, svd->anon_index + p, &cookie);
7354 			ap = anon_get_ptr(amp->ahp, svd->anon_index + p);
7355 			if (ap != NULL) {
7356 				swap_xlate(ap, &avp, &aoffset);
7357 			}
7358 			anon_array_exit(&cookie);
7359 			ANON_LOCK_EXIT(&amp->a_rwlock);
7360 		}
7361 		if ((avp != NULL) && page_exists(avp, aoffset)) {
7362 			/* A page exists for the anon slot */
7363 			ret |= SEG_PAGE_INCORE;
7364 
7365 			/*
7366 			 * If page is mapped and writable
7367 			 */
7368 			attr = (uint_t)0;
7369 			if ((hat_getattr(seg->s_as->a_hat, addr,
7370 			    &attr) != -1) && (attr & PROT_WRITE)) {
7371 				ret |= SEG_PAGE_ANON;
7372 			}
7373 			/*
7374 			 * Don't get page_struct lock for lckcnt and cowcnt,
7375 			 * since this is purely advisory.
7376 			 */
7377 			if ((pp = page_lookup_nowait(avp, aoffset,
7378 			    SE_SHARED)) != NULL) {
7379 				if (pp->p_lckcnt)
7380 					ret |= SEG_PAGE_SOFTLOCK;
7381 				if (pp->p_cowcnt)
7382 					ret |= SEG_PAGE_HASCOW;
7383 				page_unlock(pp);
7384 			}
7385 		}
7386 
7387 		/* Gather vnode statistics */
7388 		vp = svd->vp;
7389 		offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7390 
7391 		if (vp != NULL) {
7392 			/*
7393 			 * Try to obtain a "shared" lock on the page
7394 			 * without blocking.  If this fails, determine
7395 			 * if the page is in memory.
7396 			 */
7397 			pp = page_lookup_nowait(vp, offset, SE_SHARED);
7398 			if ((pp == NULL) && (page_exists(vp, offset))) {
7399 				/* Page is incore, and is named */
7400 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7401 			}
7402 			/*
7403 			 * Don't get page_struct lock for lckcnt and cowcnt,
7404 			 * since this is purely advisory.
7405 			 */
7406 			if (pp != NULL) {
7407 				ret |= (SEG_PAGE_INCORE | SEG_PAGE_VNODE);
7408 				if (pp->p_lckcnt)
7409 					ret |= SEG_PAGE_SOFTLOCK;
7410 				if (pp->p_cowcnt)
7411 					ret |= SEG_PAGE_HASCOW;
7412 				page_unlock(pp);
7413 			}
7414 		}
7415 
7416 		/* Gather virtual page information */
7417 		if (vpp) {
7418 			if (VPP_ISPPLOCK(vpp))
7419 				ret |= SEG_PAGE_LOCKED;
7420 			vpp++;
7421 		}
7422 
7423 		*vec++ = (char)ret;
7424 	}
7425 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7426 	return (len);
7427 }
7428 
7429 /*
7430  * Statement for p_cowcnts/p_lckcnts.
7431  *
7432  * p_cowcnt is updated while mlock/munlocking MAP_PRIVATE and PROT_WRITE region
7433  * irrespective of the following factors or anything else:
7434  *
7435  *	(1) anon slots are populated or not
7436  *	(2) cow is broken or not
7437  *	(3) refcnt on ap is 1 or greater than 1
7438  *
7439  * If it's not MAP_PRIVATE and PROT_WRITE, p_lckcnt is updated during mlock
7440  * and munlock.
7441  *
7442  *
7443  * Handling p_cowcnts/p_lckcnts during copy-on-write fault:
7444  *
7445  *	if vpage has PROT_WRITE
7446  *		transfer cowcnt on the oldpage -> cowcnt on the newpage
7447  *	else
7448  *		transfer lckcnt on the oldpage -> lckcnt on the newpage
7449  *
7450  *	During copy-on-write, decrement p_cowcnt on the oldpage and increment
7451  *	p_cowcnt on the newpage *if* the corresponding vpage has PROT_WRITE.
7452  *
7453  *	We may also break COW if softlocking on read access in the physio case.
7454  *	In this case, vpage may not have PROT_WRITE. So, we need to decrement
7455  *	p_lckcnt on the oldpage and increment p_lckcnt on the newpage *if* the
7456  *	vpage doesn't have PROT_WRITE.
7457  *
7458  *
7459  * Handling p_cowcnts/p_lckcnts during mprotect on mlocked region:
7460  *
7461  * 	If a MAP_PRIVATE region loses PROT_WRITE, we decrement p_cowcnt and
7462  *	increment p_lckcnt by calling page_subclaim() which takes care of
7463  * 	availrmem accounting and p_lckcnt overflow.
7464  *
7465  *	If a MAP_PRIVATE region gains PROT_WRITE, we decrement p_lckcnt and
7466  *	increment p_cowcnt by calling page_addclaim() which takes care of
7467  *	availrmem availability and p_cowcnt overflow.
7468  */
7469 
7470 /*
7471  * Lock down (or unlock) pages mapped by this segment.
7472  *
7473  * XXX only creates PAGESIZE pages if anon slots are not initialized.
7474  * At fault time they will be relocated into larger pages.
7475  */
7476 static int
7477 segvn_lockop(struct seg *seg, caddr_t addr, size_t len,
7478     int attr, int op, ulong_t *lockmap, size_t pos)
7479 {
7480 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7481 	struct vpage *vpp;
7482 	struct vpage *evp;
7483 	page_t *pp;
7484 	u_offset_t offset;
7485 	u_offset_t off;
7486 	int segtype;
7487 	int pageprot;
7488 	int claim;
7489 	struct vnode *vp;
7490 	ulong_t anon_index;
7491 	struct anon_map *amp;
7492 	struct anon *ap;
7493 	struct vattr va;
7494 	anon_sync_obj_t cookie;
7495 	struct kshmid *sp = NULL;
7496 	struct proc	*p = curproc;
7497 	kproject_t	*proj = NULL;
7498 	int chargeproc = 1;
7499 	size_t locked_bytes = 0;
7500 	size_t unlocked_bytes = 0;
7501 	int err = 0;
7502 
7503 	/*
7504 	 * Hold write lock on address space because may split or concatenate
7505 	 * segments
7506 	 */
7507 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7508 
7509 	/*
7510 	 * If this is a shm, use shm's project and zone, else use
7511 	 * project and zone of calling process
7512 	 */
7513 
7514 	/* Determine if this segment backs a sysV shm */
7515 	if (svd->amp != NULL && svd->amp->a_sp != NULL) {
7516 		ASSERT(svd->type == MAP_SHARED);
7517 		ASSERT(svd->tr_state == SEGVN_TR_OFF);
7518 		sp = svd->amp->a_sp;
7519 		proj = sp->shm_perm.ipc_proj;
7520 		chargeproc = 0;
7521 	}
7522 
7523 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7524 	if (attr) {
7525 		pageprot = attr & ~(SHARED|PRIVATE);
7526 		segtype = attr & SHARED ? MAP_SHARED : MAP_PRIVATE;
7527 
7528 		/*
7529 		 * We are done if the segment types don't match
7530 		 * or if we have segment level protections and
7531 		 * they don't match.
7532 		 */
7533 		if (svd->type != segtype) {
7534 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7535 			return (0);
7536 		}
7537 		if (svd->pageprot == 0 && svd->prot != pageprot) {
7538 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7539 			return (0);
7540 		}
7541 	}
7542 
7543 	if (op == MC_LOCK) {
7544 		if (svd->tr_state == SEGVN_TR_INIT) {
7545 			svd->tr_state = SEGVN_TR_OFF;
7546 		} else if (svd->tr_state == SEGVN_TR_ON) {
7547 			ASSERT(svd->amp != NULL);
7548 			segvn_textunrepl(seg, 0);
7549 			ASSERT(svd->amp == NULL &&
7550 			    svd->tr_state == SEGVN_TR_OFF);
7551 		}
7552 	}
7553 
7554 	/*
7555 	 * If we're locking, then we must create a vpage structure if
7556 	 * none exists.  If we're unlocking, then check to see if there
7557 	 * is a vpage --  if not, then we could not have locked anything.
7558 	 */
7559 
7560 	if ((vpp = svd->vpage) == NULL) {
7561 		if (op == MC_LOCK)
7562 			segvn_vpage(seg);
7563 		else {
7564 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7565 			return (0);
7566 		}
7567 	}
7568 
7569 	/*
7570 	 * The anonymous data vector (i.e., previously
7571 	 * unreferenced mapping to swap space) can be allocated
7572 	 * by lazily testing for its existence.
7573 	 */
7574 	if (op == MC_LOCK && svd->amp == NULL && svd->vp == NULL) {
7575 		ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
7576 		svd->amp = anonmap_alloc(seg->s_size, 0, ANON_SLEEP);
7577 		svd->amp->a_szc = seg->s_szc;
7578 	}
7579 
7580 	if ((amp = svd->amp) != NULL) {
7581 		anon_index = svd->anon_index + seg_page(seg, addr);
7582 	}
7583 
7584 	offset = svd->offset + (uintptr_t)(addr - seg->s_base);
7585 	evp = &svd->vpage[seg_page(seg, addr + len)];
7586 
7587 	if (sp != NULL)
7588 		mutex_enter(&sp->shm_mlock);
7589 
7590 	/* determine number of unlocked bytes in range for lock operation */
7591 	if (op == MC_LOCK) {
7592 
7593 		if (sp == NULL) {
7594 			for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7595 			    vpp++) {
7596 				if (!VPP_ISPPLOCK(vpp))
7597 					unlocked_bytes += PAGESIZE;
7598 			}
7599 		} else {
7600 			ulong_t		i_idx, i_edx;
7601 			anon_sync_obj_t	i_cookie;
7602 			struct anon	*i_ap;
7603 			struct vnode	*i_vp;
7604 			u_offset_t	i_off;
7605 
7606 			/* Only count sysV pages once for locked memory */
7607 			i_edx = svd->anon_index + seg_page(seg, addr + len);
7608 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7609 			for (i_idx = anon_index; i_idx < i_edx; i_idx++) {
7610 				anon_array_enter(amp, i_idx, &i_cookie);
7611 				i_ap = anon_get_ptr(amp->ahp, i_idx);
7612 				if (i_ap == NULL) {
7613 					unlocked_bytes += PAGESIZE;
7614 					anon_array_exit(&i_cookie);
7615 					continue;
7616 				}
7617 				swap_xlate(i_ap, &i_vp, &i_off);
7618 				anon_array_exit(&i_cookie);
7619 				pp = page_lookup(i_vp, i_off, SE_SHARED);
7620 				if (pp == NULL) {
7621 					unlocked_bytes += PAGESIZE;
7622 					continue;
7623 				} else if (pp->p_lckcnt == 0)
7624 					unlocked_bytes += PAGESIZE;
7625 				page_unlock(pp);
7626 			}
7627 			ANON_LOCK_EXIT(&amp->a_rwlock);
7628 		}
7629 
7630 		mutex_enter(&p->p_lock);
7631 		err = rctl_incr_locked_mem(p, proj, unlocked_bytes,
7632 		    chargeproc);
7633 		mutex_exit(&p->p_lock);
7634 
7635 		if (err) {
7636 			if (sp != NULL)
7637 				mutex_exit(&sp->shm_mlock);
7638 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7639 			return (err);
7640 		}
7641 	}
7642 	/*
7643 	 * Loop over all pages in the range.  Process if we're locking and
7644 	 * page has not already been locked in this mapping; or if we're
7645 	 * unlocking and the page has been locked.
7646 	 */
7647 	for (vpp = &svd->vpage[seg_page(seg, addr)]; vpp < evp;
7648 	    vpp++, pos++, addr += PAGESIZE, offset += PAGESIZE, anon_index++) {
7649 		if ((attr == 0 || VPP_PROT(vpp) == pageprot) &&
7650 		    ((op == MC_LOCK && !VPP_ISPPLOCK(vpp)) ||
7651 		    (op == MC_UNLOCK && VPP_ISPPLOCK(vpp)))) {
7652 
7653 			if (amp != NULL)
7654 				ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7655 			/*
7656 			 * If this isn't a MAP_NORESERVE segment and
7657 			 * we're locking, allocate anon slots if they
7658 			 * don't exist.  The page is brought in later on.
7659 			 */
7660 			if (op == MC_LOCK && svd->vp == NULL &&
7661 			    ((svd->flags & MAP_NORESERVE) == 0) &&
7662 			    amp != NULL &&
7663 			    ((ap = anon_get_ptr(amp->ahp, anon_index))
7664 								== NULL)) {
7665 				anon_array_enter(amp, anon_index, &cookie);
7666 
7667 				if ((ap = anon_get_ptr(amp->ahp,
7668 						anon_index)) == NULL) {
7669 					pp = anon_zero(seg, addr, &ap,
7670 					    svd->cred);
7671 					if (pp == NULL) {
7672 						anon_array_exit(&cookie);
7673 						ANON_LOCK_EXIT(&amp->a_rwlock);
7674 						err = ENOMEM;
7675 						goto out;
7676 					}
7677 					ASSERT(anon_get_ptr(amp->ahp,
7678 						anon_index) == NULL);
7679 					(void) anon_set_ptr(amp->ahp,
7680 						anon_index, ap, ANON_SLEEP);
7681 					page_unlock(pp);
7682 				}
7683 				anon_array_exit(&cookie);
7684 			}
7685 
7686 			/*
7687 			 * Get name for page, accounting for
7688 			 * existence of private copy.
7689 			 */
7690 			ap = NULL;
7691 			if (amp != NULL) {
7692 				anon_array_enter(amp, anon_index, &cookie);
7693 				ap = anon_get_ptr(amp->ahp, anon_index);
7694 				if (ap != NULL) {
7695 					swap_xlate(ap, &vp, &off);
7696 				} else {
7697 					if (svd->vp == NULL &&
7698 					    (svd->flags & MAP_NORESERVE)) {
7699 						anon_array_exit(&cookie);
7700 						ANON_LOCK_EXIT(&amp->a_rwlock);
7701 						continue;
7702 					}
7703 					vp = svd->vp;
7704 					off = offset;
7705 				}
7706 				anon_array_exit(&cookie);
7707 				ANON_LOCK_EXIT(&amp->a_rwlock);
7708 			} else {
7709 				vp = svd->vp;
7710 				off = offset;
7711 			}
7712 
7713 			/*
7714 			 * Get page frame.  It's ok if the page is
7715 			 * not available when we're unlocking, as this
7716 			 * may simply mean that a page we locked got
7717 			 * truncated out of existence after we locked it.
7718 			 *
7719 			 * Invoke VOP_GETPAGE() to obtain the page struct
7720 			 * since we may need to read it from disk if its
7721 			 * been paged out.
7722 			 */
7723 			if (op != MC_LOCK)
7724 				pp = page_lookup(vp, off, SE_SHARED);
7725 			else {
7726 				page_t *pl[1 + 1];
7727 				int error;
7728 
7729 				ASSERT(vp != NULL);
7730 
7731 				error = VOP_GETPAGE(vp, (offset_t)off, PAGESIZE,
7732 				    (uint_t *)NULL, pl, PAGESIZE, seg, addr,
7733 				    S_OTHER, svd->cred);
7734 
7735 				/*
7736 				 * If the error is EDEADLK then we must bounce
7737 				 * up and drop all vm subsystem locks and then
7738 				 * retry the operation later
7739 				 * This behavior is a temporary measure because
7740 				 * ufs/sds logging is badly designed and will
7741 				 * deadlock if we don't allow this bounce to
7742 				 * happen.  The real solution is to re-design
7743 				 * the logging code to work properly.  See bug
7744 				 * 4125102 for details of the problem.
7745 				 */
7746 				if (error == EDEADLK) {
7747 					err = error;
7748 					goto out;
7749 				}
7750 				/*
7751 				 * Quit if we fail to fault in the page.  Treat
7752 				 * the failure as an error, unless the addr
7753 				 * is mapped beyond the end of a file.
7754 				 */
7755 				if (error && svd->vp) {
7756 					va.va_mask = AT_SIZE;
7757 					if (VOP_GETATTR(svd->vp, &va, 0,
7758 					    svd->cred) != 0) {
7759 						err = EIO;
7760 						goto out;
7761 					}
7762 					if (btopr(va.va_size) >=
7763 					    btopr(off + 1)) {
7764 						err = EIO;
7765 						goto out;
7766 					}
7767 					goto out;
7768 
7769 				} else if (error) {
7770 					err = EIO;
7771 					goto out;
7772 				}
7773 				pp = pl[0];
7774 				ASSERT(pp != NULL);
7775 			}
7776 
7777 			/*
7778 			 * See Statement at the beginning of this routine.
7779 			 *
7780 			 * claim is always set if MAP_PRIVATE and PROT_WRITE
7781 			 * irrespective of following factors:
7782 			 *
7783 			 * (1) anon slots are populated or not
7784 			 * (2) cow is broken or not
7785 			 * (3) refcnt on ap is 1 or greater than 1
7786 			 *
7787 			 * See 4140683 for details
7788 			 */
7789 			claim = ((VPP_PROT(vpp) & PROT_WRITE) &&
7790 				(svd->type == MAP_PRIVATE));
7791 
7792 			/*
7793 			 * Perform page-level operation appropriate to
7794 			 * operation.  If locking, undo the SOFTLOCK
7795 			 * performed to bring the page into memory
7796 			 * after setting the lock.  If unlocking,
7797 			 * and no page was found, account for the claim
7798 			 * separately.
7799 			 */
7800 			if (op == MC_LOCK) {
7801 				int ret = 1;	/* Assume success */
7802 
7803 				ASSERT(!VPP_ISPPLOCK(vpp));
7804 
7805 				ret = page_pp_lock(pp, claim, 0);
7806 				if (ret == 0) {
7807 					/* locking page failed */
7808 					page_unlock(pp);
7809 					err = EAGAIN;
7810 					goto out;
7811 				}
7812 				VPP_SETPPLOCK(vpp);
7813 				if (sp != NULL) {
7814 					if (pp->p_lckcnt == 1)
7815 						locked_bytes += PAGESIZE;
7816 				} else
7817 					locked_bytes += PAGESIZE;
7818 
7819 				if (lockmap != (ulong_t *)NULL)
7820 					BT_SET(lockmap, pos);
7821 
7822 				page_unlock(pp);
7823 			} else {
7824 				ASSERT(VPP_ISPPLOCK(vpp));
7825 				if (pp != NULL) {
7826 					/* sysV pages should be locked */
7827 					ASSERT(sp == NULL || pp->p_lckcnt > 0);
7828 					page_pp_unlock(pp, claim, 0);
7829 					if (sp != NULL) {
7830 						if (pp->p_lckcnt == 0)
7831 							unlocked_bytes
7832 							    += PAGESIZE;
7833 					} else
7834 						unlocked_bytes += PAGESIZE;
7835 					page_unlock(pp);
7836 				} else {
7837 					ASSERT(sp == NULL);
7838 					unlocked_bytes += PAGESIZE;
7839 				}
7840 				VPP_CLRPPLOCK(vpp);
7841 			}
7842 		}
7843 	}
7844 out:
7845 	if (op == MC_LOCK) {
7846 		/* Credit back bytes that did not get locked */
7847 		if ((unlocked_bytes - locked_bytes) > 0) {
7848 			if (proj == NULL)
7849 				mutex_enter(&p->p_lock);
7850 			rctl_decr_locked_mem(p, proj,
7851 			    (unlocked_bytes - locked_bytes), chargeproc);
7852 			if (proj == NULL)
7853 				mutex_exit(&p->p_lock);
7854 		}
7855 
7856 	} else {
7857 		/* Account bytes that were unlocked */
7858 		if (unlocked_bytes > 0) {
7859 			if (proj == NULL)
7860 				mutex_enter(&p->p_lock);
7861 			rctl_decr_locked_mem(p, proj, unlocked_bytes,
7862 			    chargeproc);
7863 			if (proj == NULL)
7864 				mutex_exit(&p->p_lock);
7865 		}
7866 	}
7867 	if (sp != NULL)
7868 		mutex_exit(&sp->shm_mlock);
7869 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7870 
7871 	return (err);
7872 }
7873 
7874 /*
7875  * Set advice from user for specified pages
7876  * There are 5 types of advice:
7877  *	MADV_NORMAL	- Normal (default) behavior (whatever that is)
7878  *	MADV_RANDOM	- Random page references
7879  *				do not allow readahead or 'klustering'
7880  *	MADV_SEQUENTIAL	- Sequential page references
7881  *				Pages previous to the one currently being
7882  *				accessed (determined by fault) are 'not needed'
7883  *				and are freed immediately
7884  *	MADV_WILLNEED	- Pages are likely to be used (fault ahead in mctl)
7885  *	MADV_DONTNEED	- Pages are not needed (synced out in mctl)
7886  *	MADV_FREE	- Contents can be discarded
7887  *	MADV_ACCESS_DEFAULT- Default access
7888  *	MADV_ACCESS_LWP	- Next LWP will access heavily
7889  *	MADV_ACCESS_MANY- Many LWPs or processes will access heavily
7890  */
7891 static int
7892 segvn_advise(struct seg *seg, caddr_t addr, size_t len, uint_t behav)
7893 {
7894 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
7895 	size_t page;
7896 	int err = 0;
7897 	int already_set;
7898 	struct anon_map *amp;
7899 	ulong_t anon_index;
7900 	struct seg *next;
7901 	lgrp_mem_policy_t policy;
7902 	struct seg *prev;
7903 	struct vnode *vp;
7904 
7905 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
7906 
7907 	/*
7908 	 * In case of MADV_FREE, we won't be modifying any segment private
7909 	 * data structures; so, we only need to grab READER's lock
7910 	 */
7911 	if (behav != MADV_FREE) {
7912 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_WRITER);
7913 		if (svd->tr_state != SEGVN_TR_OFF) {
7914 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7915 			return (0);
7916 		}
7917 	} else {
7918 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
7919 	}
7920 
7921 	/*
7922 	 * Large pages are assumed to be only turned on when accesses to the
7923 	 * segment's address range have spatial and temporal locality. That
7924 	 * justifies ignoring MADV_SEQUENTIAL for large page segments.
7925 	 * Also, ignore advice affecting lgroup memory allocation
7926 	 * if don't need to do lgroup optimizations on this system
7927 	 */
7928 
7929 	if ((behav == MADV_SEQUENTIAL &&
7930 	    (seg->s_szc != 0 || HAT_IS_REGION_COOKIE_VALID(svd->rcookie))) ||
7931 	    (!lgrp_optimizations() && (behav == MADV_ACCESS_DEFAULT ||
7932 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY))) {
7933 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7934 		return (0);
7935 	}
7936 
7937 	if (behav == MADV_SEQUENTIAL || behav == MADV_ACCESS_DEFAULT ||
7938 	    behav == MADV_ACCESS_LWP || behav == MADV_ACCESS_MANY) {
7939 		/*
7940 		 * Since we are going to unload hat mappings
7941 		 * we first have to flush the cache. Otherwise
7942 		 * this might lead to system panic if another
7943 		 * thread is doing physio on the range whose
7944 		 * mappings are unloaded by madvise(3C).
7945 		 */
7946 		if (svd->softlockcnt > 0) {
7947 			/*
7948 			 * Since we do have the segvn writers lock
7949 			 * nobody can fill the cache with entries
7950 			 * belonging to this seg during the purge.
7951 			 * The flush either succeeds or we still
7952 			 * have pending I/Os. In the later case,
7953 			 * madvise(3C) fails.
7954 			 */
7955 			segvn_purge(seg);
7956 			if (svd->softlockcnt > 0) {
7957 				/*
7958 				 * Since madvise(3C) is advisory and
7959 				 * it's not part of UNIX98, madvise(3C)
7960 				 * failure here doesn't cause any hardship.
7961 				 * Note that we don't block in "as" layer.
7962 				 */
7963 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7964 				return (EAGAIN);
7965 			}
7966 		}
7967 	}
7968 
7969 	amp = svd->amp;
7970 	vp = svd->vp;
7971 	if (behav == MADV_FREE) {
7972 		/*
7973 		 * MADV_FREE is not supported for segments with
7974 		 * underlying object; if anonmap is NULL, anon slots
7975 		 * are not yet populated and there is nothing for
7976 		 * us to do. As MADV_FREE is advisory, we don't
7977 		 * return error in either case.
7978 		 */
7979 		if (vp != NULL || amp == NULL) {
7980 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7981 			return (0);
7982 		}
7983 
7984 		page = seg_page(seg, addr);
7985 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
7986 		anon_disclaim(amp, svd->anon_index + page, len, 0);
7987 		ANON_LOCK_EXIT(&amp->a_rwlock);
7988 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
7989 		return (0);
7990 	}
7991 
7992 	/*
7993 	 * If advice is to be applied to entire segment,
7994 	 * use advice field in seg_data structure
7995 	 * otherwise use appropriate vpage entry.
7996 	 */
7997 	if ((addr == seg->s_base) && (len == seg->s_size)) {
7998 		switch (behav) {
7999 		case MADV_ACCESS_LWP:
8000 		case MADV_ACCESS_MANY:
8001 		case MADV_ACCESS_DEFAULT:
8002 			/*
8003 			 * Set memory allocation policy for this segment
8004 			 */
8005 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8006 			if (svd->type == MAP_SHARED)
8007 				already_set = lgrp_shm_policy_set(policy, amp,
8008 				    svd->anon_index, vp, svd->offset, len);
8009 			else {
8010 				/*
8011 				 * For private memory, need writers lock on
8012 				 * address space because the segment may be
8013 				 * split or concatenated when changing policy
8014 				 */
8015 				if (AS_READ_HELD(seg->s_as,
8016 				    &seg->s_as->a_lock)) {
8017 					SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8018 					return (IE_RETRY);
8019 				}
8020 
8021 				already_set = lgrp_privm_policy_set(policy,
8022 				    &svd->policy_info, len);
8023 			}
8024 
8025 			/*
8026 			 * If policy set already and it shouldn't be reapplied,
8027 			 * don't do anything.
8028 			 */
8029 			if (already_set &&
8030 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8031 				break;
8032 
8033 			/*
8034 			 * Mark any existing pages in given range for
8035 			 * migration
8036 			 */
8037 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8038 			    vp, svd->offset, 1);
8039 
8040 			/*
8041 			 * If same policy set already or this is a shared
8042 			 * memory segment, don't need to try to concatenate
8043 			 * segment with adjacent ones.
8044 			 */
8045 			if (already_set || svd->type == MAP_SHARED)
8046 				break;
8047 
8048 			/*
8049 			 * Try to concatenate this segment with previous
8050 			 * one and next one, since we changed policy for
8051 			 * this one and it may be compatible with adjacent
8052 			 * ones now.
8053 			 */
8054 			prev = AS_SEGPREV(seg->s_as, seg);
8055 			next = AS_SEGNEXT(seg->s_as, seg);
8056 
8057 			if (next && next->s_ops == &segvn_ops &&
8058 			    addr + len == next->s_base)
8059 				(void) segvn_concat(seg, next, 1);
8060 
8061 			if (prev && prev->s_ops == &segvn_ops &&
8062 			    addr == prev->s_base + prev->s_size) {
8063 				/*
8064 				 * Drop lock for private data of current
8065 				 * segment before concatenating (deleting) it
8066 				 * and return IE_REATTACH to tell as_ctl() that
8067 				 * current segment has changed
8068 				 */
8069 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8070 				if (!segvn_concat(prev, seg, 1))
8071 					err = IE_REATTACH;
8072 
8073 				return (err);
8074 			}
8075 			break;
8076 
8077 		case MADV_SEQUENTIAL:
8078 			/*
8079 			 * unloading mapping guarantees
8080 			 * detection in segvn_fault
8081 			 */
8082 			ASSERT(seg->s_szc == 0);
8083 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8084 			hat_unload(seg->s_as->a_hat, addr, len,
8085 				HAT_UNLOAD);
8086 			/* FALLTHROUGH */
8087 		case MADV_NORMAL:
8088 		case MADV_RANDOM:
8089 			svd->advice = (uchar_t)behav;
8090 			svd->pageadvice = 0;
8091 			break;
8092 		case MADV_WILLNEED:	/* handled in memcntl */
8093 		case MADV_DONTNEED:	/* handled in memcntl */
8094 		case MADV_FREE:		/* handled above */
8095 			break;
8096 		default:
8097 			err = EINVAL;
8098 		}
8099 	} else {
8100 		caddr_t			eaddr;
8101 		struct seg		*new_seg;
8102 		struct segvn_data	*new_svd;
8103 		u_offset_t		off;
8104 		caddr_t			oldeaddr;
8105 
8106 		page = seg_page(seg, addr);
8107 
8108 		segvn_vpage(seg);
8109 
8110 		switch (behav) {
8111 			struct vpage *bvpp, *evpp;
8112 
8113 		case MADV_ACCESS_LWP:
8114 		case MADV_ACCESS_MANY:
8115 		case MADV_ACCESS_DEFAULT:
8116 			/*
8117 			 * Set memory allocation policy for portion of this
8118 			 * segment
8119 			 */
8120 
8121 			/*
8122 			 * Align address and length of advice to page
8123 			 * boundaries for large pages
8124 			 */
8125 			if (seg->s_szc != 0) {
8126 				size_t	pgsz;
8127 
8128 				pgsz = page_get_pagesize(seg->s_szc);
8129 				addr = (caddr_t)P2ALIGN((uintptr_t)addr, pgsz);
8130 				len = P2ROUNDUP(len, pgsz);
8131 			}
8132 
8133 			/*
8134 			 * Check to see whether policy is set already
8135 			 */
8136 			policy = lgrp_madv_to_policy(behav, len, svd->type);
8137 
8138 			anon_index = svd->anon_index + page;
8139 			off = svd->offset + (uintptr_t)(addr - seg->s_base);
8140 
8141 			if (svd->type == MAP_SHARED)
8142 				already_set = lgrp_shm_policy_set(policy, amp,
8143 				    anon_index, vp, off, len);
8144 			else
8145 				already_set =
8146 				    (policy == svd->policy_info.mem_policy);
8147 
8148 			/*
8149 			 * If policy set already and it shouldn't be reapplied,
8150 			 * don't do anything.
8151 			 */
8152 			if (already_set &&
8153 			    !LGRP_MEM_POLICY_REAPPLICABLE(policy))
8154 				break;
8155 
8156 			/*
8157 			 * For private memory, need writers lock on
8158 			 * address space because the segment may be
8159 			 * split or concatenated when changing policy
8160 			 */
8161 			if (svd->type == MAP_PRIVATE &&
8162 			    AS_READ_HELD(seg->s_as, &seg->s_as->a_lock)) {
8163 				SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8164 				return (IE_RETRY);
8165 			}
8166 
8167 			/*
8168 			 * Mark any existing pages in given range for
8169 			 * migration
8170 			 */
8171 			page_mark_migrate(seg, addr, len, amp, svd->anon_index,
8172 			    vp, svd->offset, 1);
8173 
8174 			/*
8175 			 * Don't need to try to split or concatenate
8176 			 * segments, since policy is same or this is a shared
8177 			 * memory segment
8178 			 */
8179 			if (already_set || svd->type == MAP_SHARED)
8180 				break;
8181 
8182 			if (HAT_IS_REGION_COOKIE_VALID(svd->rcookie)) {
8183 				ASSERT(svd->amp == NULL);
8184 				ASSERT(svd->tr_state == SEGVN_TR_OFF);
8185 				ASSERT(svd->softlockcnt == 0);
8186 				hat_leave_region(seg->s_as->a_hat, svd->rcookie,
8187 				    HAT_REGION_TEXT);
8188 				svd->rcookie = HAT_INVALID_REGION_COOKIE;
8189 			}
8190 
8191 			/*
8192 			 * Split off new segment if advice only applies to a
8193 			 * portion of existing segment starting in middle
8194 			 */
8195 			new_seg = NULL;
8196 			eaddr = addr + len;
8197 			oldeaddr = seg->s_base + seg->s_size;
8198 			if (addr > seg->s_base) {
8199 				/*
8200 				 * Must flush I/O page cache
8201 				 * before splitting segment
8202 				 */
8203 				if (svd->softlockcnt > 0)
8204 					segvn_purge(seg);
8205 
8206 				/*
8207 				 * Split segment and return IE_REATTACH to tell
8208 				 * as_ctl() that current segment changed
8209 				 */
8210 				new_seg = segvn_split_seg(seg, addr);
8211 				new_svd = (struct segvn_data *)new_seg->s_data;
8212 				err = IE_REATTACH;
8213 
8214 				/*
8215 				 * If new segment ends where old one
8216 				 * did, try to concatenate the new
8217 				 * segment with next one.
8218 				 */
8219 				if (eaddr == oldeaddr) {
8220 					/*
8221 					 * Set policy for new segment
8222 					 */
8223 					(void) lgrp_privm_policy_set(policy,
8224 					    &new_svd->policy_info,
8225 					    new_seg->s_size);
8226 
8227 					next = AS_SEGNEXT(new_seg->s_as,
8228 					    new_seg);
8229 
8230 					if (next &&
8231 					    next->s_ops == &segvn_ops &&
8232 					    eaddr == next->s_base)
8233 						(void) segvn_concat(new_seg,
8234 						    next, 1);
8235 				}
8236 			}
8237 
8238 			/*
8239 			 * Split off end of existing segment if advice only
8240 			 * applies to a portion of segment ending before
8241 			 * end of the existing segment
8242 			 */
8243 			if (eaddr < oldeaddr) {
8244 				/*
8245 				 * Must flush I/O page cache
8246 				 * before splitting segment
8247 				 */
8248 				if (svd->softlockcnt > 0)
8249 					segvn_purge(seg);
8250 
8251 				/*
8252 				 * If beginning of old segment was already
8253 				 * split off, use new segment to split end off
8254 				 * from.
8255 				 */
8256 				if (new_seg != NULL && new_seg != seg) {
8257 					/*
8258 					 * Split segment
8259 					 */
8260 					(void) segvn_split_seg(new_seg, eaddr);
8261 
8262 					/*
8263 					 * Set policy for new segment
8264 					 */
8265 					(void) lgrp_privm_policy_set(policy,
8266 					    &new_svd->policy_info,
8267 					    new_seg->s_size);
8268 				} else {
8269 					/*
8270 					 * Split segment and return IE_REATTACH
8271 					 * to tell as_ctl() that current
8272 					 * segment changed
8273 					 */
8274 					(void) segvn_split_seg(seg, eaddr);
8275 					err = IE_REATTACH;
8276 
8277 					(void) lgrp_privm_policy_set(policy,
8278 					    &svd->policy_info, seg->s_size);
8279 
8280 					/*
8281 					 * If new segment starts where old one
8282 					 * did, try to concatenate it with
8283 					 * previous segment.
8284 					 */
8285 					if (addr == seg->s_base) {
8286 						prev = AS_SEGPREV(seg->s_as,
8287 						    seg);
8288 
8289 						/*
8290 						 * Drop lock for private data
8291 						 * of current segment before
8292 						 * concatenating (deleting) it
8293 						 */
8294 						if (prev &&
8295 						    prev->s_ops ==
8296 						    &segvn_ops &&
8297 						    addr == prev->s_base +
8298 						    prev->s_size) {
8299 							SEGVN_LOCK_EXIT(
8300 							    seg->s_as,
8301 							    &svd->lock);
8302 							(void) segvn_concat(
8303 							    prev, seg, 1);
8304 							return (err);
8305 						}
8306 					}
8307 				}
8308 			}
8309 			break;
8310 		case MADV_SEQUENTIAL:
8311 			ASSERT(seg->s_szc == 0);
8312 			ASSERT(svd->rcookie == HAT_INVALID_REGION_COOKIE);
8313 			hat_unload(seg->s_as->a_hat, addr, len, HAT_UNLOAD);
8314 			/* FALLTHROUGH */
8315 		case MADV_NORMAL:
8316 		case MADV_RANDOM:
8317 			bvpp = &svd->vpage[page];
8318 			evpp = &svd->vpage[page + (len >> PAGESHIFT)];
8319 			for (; bvpp < evpp; bvpp++)
8320 				VPP_SETADVICE(bvpp, behav);
8321 			svd->advice = MADV_NORMAL;
8322 			break;
8323 		case MADV_WILLNEED:	/* handled in memcntl */
8324 		case MADV_DONTNEED:	/* handled in memcntl */
8325 		case MADV_FREE:		/* handled above */
8326 			break;
8327 		default:
8328 			err = EINVAL;
8329 		}
8330 	}
8331 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8332 	return (err);
8333 }
8334 
8335 /*
8336  * Create a vpage structure for this seg.
8337  */
8338 static void
8339 segvn_vpage(struct seg *seg)
8340 {
8341 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8342 	struct vpage *vp, *evp;
8343 
8344 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8345 
8346 	/*
8347 	 * If no vpage structure exists, allocate one.  Copy the protections
8348 	 * and the advice from the segment itself to the individual pages.
8349 	 */
8350 	if (svd->vpage == NULL) {
8351 		svd->pageadvice = 1;
8352 		svd->vpage = kmem_zalloc(seg_pages(seg) * sizeof (struct vpage),
8353 		    KM_SLEEP);
8354 		evp = &svd->vpage[seg_page(seg, seg->s_base + seg->s_size)];
8355 		for (vp = svd->vpage; vp < evp; vp++) {
8356 			VPP_SETPROT(vp, svd->prot);
8357 			VPP_SETADVICE(vp, svd->advice);
8358 		}
8359 	}
8360 }
8361 
8362 /*
8363  * Dump the pages belonging to this segvn segment.
8364  */
8365 static void
8366 segvn_dump(struct seg *seg)
8367 {
8368 	struct segvn_data *svd;
8369 	page_t *pp;
8370 	struct anon_map *amp;
8371 	ulong_t	anon_index;
8372 	struct vnode *vp;
8373 	u_offset_t off, offset;
8374 	pfn_t pfn;
8375 	pgcnt_t page, npages;
8376 	caddr_t addr;
8377 
8378 	npages = seg_pages(seg);
8379 	svd = (struct segvn_data *)seg->s_data;
8380 	vp = svd->vp;
8381 	off = offset = svd->offset;
8382 	addr = seg->s_base;
8383 
8384 	if ((amp = svd->amp) != NULL) {
8385 		anon_index = svd->anon_index;
8386 		ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8387 	}
8388 
8389 	for (page = 0; page < npages; page++, offset += PAGESIZE) {
8390 		struct anon *ap;
8391 		int we_own_it = 0;
8392 
8393 		if (amp && (ap = anon_get_ptr(svd->amp->ahp, anon_index++))) {
8394 			swap_xlate_nopanic(ap, &vp, &off);
8395 		} else {
8396 			vp = svd->vp;
8397 			off = offset;
8398 		}
8399 
8400 		/*
8401 		 * If pp == NULL, the page either does not exist
8402 		 * or is exclusively locked.  So determine if it
8403 		 * exists before searching for it.
8404 		 */
8405 
8406 		if ((pp = page_lookup_nowait(vp, off, SE_SHARED)))
8407 			we_own_it = 1;
8408 		else
8409 			pp = page_exists(vp, off);
8410 
8411 		if (pp) {
8412 			pfn = page_pptonum(pp);
8413 			dump_addpage(seg->s_as, addr, pfn);
8414 			if (we_own_it)
8415 				page_unlock(pp);
8416 		}
8417 		addr += PAGESIZE;
8418 		dump_timeleft = dump_timeout;
8419 	}
8420 
8421 	if (amp != NULL)
8422 		ANON_LOCK_EXIT(&amp->a_rwlock);
8423 }
8424 
8425 /*
8426  * lock/unlock anon pages over a given range. Return shadow list
8427  */
8428 static int
8429 segvn_pagelock(struct seg *seg, caddr_t addr, size_t len, struct page ***ppp,
8430     enum lock_type type, enum seg_rw rw)
8431 {
8432 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8433 	size_t np, adjustpages = 0, npages = (len >> PAGESHIFT);
8434 	ulong_t anon_index;
8435 	uint_t protchk;
8436 	uint_t error;
8437 	struct anon_map *amp;
8438 	struct page **pplist, **pl, *pp;
8439 	caddr_t a;
8440 	size_t page;
8441 	caddr_t lpgaddr, lpgeaddr;
8442 	pgcnt_t szc0_npages = 0;
8443 
8444 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_START,
8445 		"segvn_pagelock: start seg %p addr %p", seg, addr);
8446 
8447 	ASSERT(seg->s_as && AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
8448 	if (seg->s_szc != 0 && (type == L_PAGELOCK || type == L_PAGEUNLOCK)) {
8449 		/*
8450 		 * We are adjusting the pagelock region to the large page size
8451 		 * boundary because the unlocked part of a large page cannot
8452 		 * be freed anyway unless all constituent pages of a large
8453 		 * page are locked. Therefore this adjustment allows us to
8454 		 * decrement availrmem by the right value (note we don't want
8455 		 * to just decrement availrem by the large page size without
8456 		 * adjusting addr and len because then we may end up
8457 		 * decrementing availrmem by large page size for every
8458 		 * constituent page locked by a new as_pagelock call).
8459 		 * as_pageunlock caller must always match as_pagelock call's
8460 		 * addr and len.
8461 		 *
8462 		 * Note segment's page size cannot change while we are holding
8463 		 * as lock.  And then it cannot change while softlockcnt is
8464 		 * not 0. This will allow us to correctly recalculate large
8465 		 * page size region for the matching pageunlock/reclaim call.
8466 		 *
8467 		 * for pageunlock *ppp points to the pointer of page_t that
8468 		 * corresponds to the real unadjusted start address. Similar
8469 		 * for pagelock *ppp must point to the pointer of page_t that
8470 		 * corresponds to the real unadjusted start address.
8471 		 */
8472 		size_t pgsz = page_get_pagesize(seg->s_szc);
8473 		CALC_LPG_REGION(pgsz, seg, addr, len, lpgaddr, lpgeaddr);
8474 		adjustpages = ((uintptr_t)(addr - lpgaddr)) >> PAGESHIFT;
8475 	}
8476 
8477 	if (type == L_PAGEUNLOCK) {
8478 
8479 		/*
8480 		 * update hat ref bits for /proc. We need to make sure
8481 		 * that threads tracing the ref and mod bits of the
8482 		 * address space get the right data.
8483 		 * Note: page ref and mod bits are updated at reclaim time
8484 		 */
8485 		if (seg->s_as->a_vbits) {
8486 			for (a = addr; a < addr + len; a += PAGESIZE) {
8487 				if (rw == S_WRITE) {
8488 					hat_setstat(seg->s_as, a,
8489 					    PAGESIZE, P_REF | P_MOD);
8490 				} else {
8491 					hat_setstat(seg->s_as, a,
8492 					    PAGESIZE, P_REF);
8493 				}
8494 			}
8495 		}
8496 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8497 		if (seg->s_szc != 0) {
8498 			VM_STAT_ADD(segvnvmstats.pagelock[0]);
8499 			seg_pinactive(seg, lpgaddr, lpgeaddr - lpgaddr,
8500 			    *ppp - adjustpages, rw, segvn_reclaim);
8501 		} else {
8502 			seg_pinactive(seg, addr, len, *ppp, rw, segvn_reclaim);
8503 		}
8504 
8505 		/*
8506 		 * If someone is blocked while unmapping, we purge
8507 		 * segment page cache and thus reclaim pplist synchronously
8508 		 * without waiting for seg_pasync_thread. This speeds up
8509 		 * unmapping in cases where munmap(2) is called, while
8510 		 * raw async i/o is still in progress or where a thread
8511 		 * exits on data fault in a multithreaded application.
8512 		 */
8513 		if (AS_ISUNMAPWAIT(seg->s_as) && (svd->softlockcnt > 0)) {
8514 			/*
8515 			 * Even if we grab segvn WRITER's lock or segp_slock
8516 			 * here, there might be another thread which could've
8517 			 * successfully performed lookup/insert just before
8518 			 * we acquired the lock here.  So, grabbing either
8519 			 * lock here is of not much use.  Until we devise
8520 			 * a strategy at upper layers to solve the
8521 			 * synchronization issues completely, we expect
8522 			 * applications to handle this appropriately.
8523 			 */
8524 			segvn_purge(seg);
8525 		}
8526 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8527 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_UNLOCK_END,
8528 			"segvn_pagelock: unlock seg %p addr %p", seg, addr);
8529 		return (0);
8530 	} else if (type == L_PAGERECLAIM) {
8531 		VM_STAT_COND_ADD(seg->s_szc != 0, segvnvmstats.pagelock[1]);
8532 		SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8533 		(void) segvn_reclaim(seg, addr, len, *ppp, rw);
8534 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8535 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_UNLOCK_END,
8536 			"segvn_pagelock: reclaim seg %p addr %p", seg, addr);
8537 		return (0);
8538 	}
8539 
8540 	if (seg->s_szc != 0) {
8541 		VM_STAT_ADD(segvnvmstats.pagelock[2]);
8542 		addr = lpgaddr;
8543 		len = lpgeaddr - lpgaddr;
8544 		npages = (len >> PAGESHIFT);
8545 	}
8546 
8547 	/*
8548 	 * for now we only support pagelock to anon memory. We've to check
8549 	 * protections for vnode objects and call into the vnode driver.
8550 	 * That's too much for a fast path. Let the fault entry point handle it.
8551 	 */
8552 	if (svd->vp != NULL) {
8553 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
8554 		    "segvn_pagelock: mapped vnode seg %p addr %p", seg, addr);
8555 		*ppp = NULL;
8556 		return (ENOTSUP);
8557 	}
8558 
8559 	/*
8560 	 * if anonmap is not yet created, let the fault entry point populate it
8561 	 * with anon ptrs.
8562 	 */
8563 	if ((amp = svd->amp) == NULL) {
8564 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
8565 		    "segvn_pagelock: anonmap null seg %p addr %p", seg, addr);
8566 		*ppp = NULL;
8567 		return (EFAULT);
8568 	}
8569 
8570 	SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8571 
8572 	/*
8573 	 * we acquire segp_slock to prevent duplicate entries
8574 	 * in seg_pcache
8575 	 */
8576 	mutex_enter(&svd->segp_slock);
8577 
8578 	/*
8579 	 * try to find pages in segment page cache
8580 	 */
8581 	pplist = seg_plookup(seg, addr, len, rw);
8582 	if (pplist != NULL) {
8583 		mutex_exit(&svd->segp_slock);
8584 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8585 		*ppp = pplist + adjustpages;
8586 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_HIT_END,
8587 			"segvn_pagelock: cache hit seg %p addr %p", seg, addr);
8588 		return (0);
8589 	}
8590 
8591 	if (rw == S_READ) {
8592 		protchk = PROT_READ;
8593 	} else {
8594 		protchk = PROT_WRITE;
8595 	}
8596 
8597 	if (svd->pageprot == 0) {
8598 		if ((svd->prot & protchk) == 0) {
8599 			mutex_exit(&svd->segp_slock);
8600 			error = EFAULT;
8601 			goto out;
8602 		}
8603 	} else {
8604 		/*
8605 		 * check page protections
8606 		 */
8607 		for (a = addr; a < addr + len; a += PAGESIZE) {
8608 			struct vpage *vp;
8609 
8610 			vp = &svd->vpage[seg_page(seg, a)];
8611 			if ((VPP_PROT(vp) & protchk) == 0) {
8612 				mutex_exit(&svd->segp_slock);
8613 				error = EFAULT;
8614 				goto out;
8615 			}
8616 		}
8617 	}
8618 
8619 	/*
8620 	 * Avoid per page overhead of segvn_slock_anonpages() for small
8621 	 * pages. For large pages segvn_slock_anonpages() only does real
8622 	 * work once per large page.  The tradeoff is that we may decrement
8623 	 * availrmem more than once for the same page but this is ok
8624 	 * for small pages.
8625 	 */
8626 	if (seg->s_szc == 0) {
8627 		mutex_enter(&freemem_lock);
8628 		if (availrmem < tune.t_minarmem + npages) {
8629 			mutex_exit(&freemem_lock);
8630 			mutex_exit(&svd->segp_slock);
8631 			error = ENOMEM;
8632 			goto out;
8633 		}
8634 		availrmem -= npages;
8635 		mutex_exit(&freemem_lock);
8636 	}
8637 
8638 	pplist = kmem_alloc(sizeof (page_t *) * npages, KM_SLEEP);
8639 	pl = pplist;
8640 	*ppp = pplist + adjustpages;
8641 
8642 	page = seg_page(seg, addr);
8643 	anon_index = svd->anon_index + page;
8644 
8645 	ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8646 	for (a = addr; a < addr + len; a += PAGESIZE, anon_index++) {
8647 		struct anon *ap;
8648 		struct vnode *vp;
8649 		u_offset_t off;
8650 		anon_sync_obj_t cookie;
8651 
8652 		anon_array_enter(amp, anon_index, &cookie);
8653 		ap = anon_get_ptr(amp->ahp, anon_index);
8654 		if (ap == NULL) {
8655 			anon_array_exit(&cookie);
8656 			break;
8657 		} else {
8658 			/*
8659 			 * We must never use seg_pcache for COW pages
8660 			 * because we might end up with original page still
8661 			 * lying in seg_pcache even after private page is
8662 			 * created. This leads to data corruption as
8663 			 * aio_write refers to the page still in cache
8664 			 * while all other accesses refer to the private
8665 			 * page.
8666 			 */
8667 			if (ap->an_refcnt != 1) {
8668 				anon_array_exit(&cookie);
8669 				break;
8670 			}
8671 		}
8672 		swap_xlate(ap, &vp, &off);
8673 		anon_array_exit(&cookie);
8674 
8675 		pp = page_lookup_nowait(vp, off, SE_SHARED);
8676 		if (pp == NULL) {
8677 			break;
8678 		}
8679 		if (seg->s_szc != 0 || pp->p_szc != 0) {
8680 			if (!segvn_slock_anonpages(pp, a == addr)) {
8681 				page_unlock(pp);
8682 				break;
8683 			}
8684 		} else {
8685 			szc0_npages++;
8686 		}
8687 		*pplist++ = pp;
8688 	}
8689 	ANON_LOCK_EXIT(&amp->a_rwlock);
8690 
8691 	ASSERT(npages >= szc0_npages);
8692 
8693 	if (a >= addr + len) {
8694 		mutex_enter(&freemem_lock);
8695 		if (seg->s_szc == 0 && npages != szc0_npages) {
8696 			ASSERT(svd->type == MAP_SHARED && amp->a_szc > 0);
8697 			availrmem += (npages - szc0_npages);
8698 		}
8699 		svd->softlockcnt += npages;
8700 		segvn_pages_locked += npages;
8701 		mutex_exit(&freemem_lock);
8702 		(void) seg_pinsert(seg, addr, len, pl, rw, SEGP_ASYNC_FLUSH,
8703 			segvn_reclaim);
8704 		mutex_exit(&svd->segp_slock);
8705 		SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8706 		TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_FILL_END,
8707 		    "segvn_pagelock: cache fill seg %p addr %p", seg, addr);
8708 		return (0);
8709 	}
8710 
8711 	mutex_exit(&svd->segp_slock);
8712 	if (seg->s_szc == 0) {
8713 		mutex_enter(&freemem_lock);
8714 		availrmem += npages;
8715 		mutex_exit(&freemem_lock);
8716 	}
8717 	error = EFAULT;
8718 	pplist = pl;
8719 	np = ((uintptr_t)(a - addr)) >> PAGESHIFT;
8720 	while (np > (uint_t)0) {
8721 		ASSERT(PAGE_LOCKED(*pplist));
8722 		if (seg->s_szc != 0 || (*pplist)->p_szc != 0) {
8723 			segvn_sunlock_anonpages(*pplist, pplist == pl);
8724 		}
8725 		page_unlock(*pplist);
8726 		np--;
8727 		pplist++;
8728 	}
8729 	kmem_free(pl, sizeof (page_t *) * npages);
8730 out:
8731 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8732 	*ppp = NULL;
8733 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEGVN_MISS_END,
8734 		"segvn_pagelock: cache miss seg %p addr %p", seg, addr);
8735 	return (error);
8736 }
8737 
8738 /*
8739  * purge any cached pages in the I/O page cache
8740  */
8741 static void
8742 segvn_purge(struct seg *seg)
8743 {
8744 	seg_ppurge(seg);
8745 }
8746 
8747 static int
8748 segvn_reclaim(struct seg *seg, caddr_t addr, size_t len, struct page **pplist,
8749 	enum seg_rw rw)
8750 {
8751 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8752 	pgcnt_t np, npages;
8753 	struct page **pl;
8754 	pgcnt_t szc0_npages = 0;
8755 
8756 #ifdef lint
8757 	addr = addr;
8758 #endif
8759 
8760 	npages = np = (len >> PAGESHIFT);
8761 	ASSERT(npages);
8762 	pl = pplist;
8763 	if (seg->s_szc != 0) {
8764 		size_t pgsz = page_get_pagesize(seg->s_szc);
8765 		if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(len, pgsz)) {
8766 			panic("segvn_reclaim: unaligned addr or len");
8767 			/*NOTREACHED*/
8768 		}
8769 	}
8770 
8771 	ASSERT(svd->vp == NULL && svd->amp != NULL);
8772 
8773 	while (np > (uint_t)0) {
8774 		if (rw == S_WRITE) {
8775 			hat_setrefmod(*pplist);
8776 		} else {
8777 			hat_setref(*pplist);
8778 		}
8779 		if (seg->s_szc != 0 || (*pplist)->p_szc != 0) {
8780 			segvn_sunlock_anonpages(*pplist, pplist == pl);
8781 		} else {
8782 			szc0_npages++;
8783 		}
8784 		page_unlock(*pplist);
8785 		np--;
8786 		pplist++;
8787 	}
8788 	kmem_free(pl, sizeof (page_t *) * npages);
8789 
8790 	mutex_enter(&freemem_lock);
8791 	segvn_pages_locked -= npages;
8792 	svd->softlockcnt -= npages;
8793 	if (szc0_npages != 0) {
8794 		availrmem += szc0_npages;
8795 	}
8796 	mutex_exit(&freemem_lock);
8797 	if (svd->softlockcnt <= 0) {
8798 		if (AS_ISUNMAPWAIT(seg->s_as)) {
8799 			mutex_enter(&seg->s_as->a_contents);
8800 			if (AS_ISUNMAPWAIT(seg->s_as)) {
8801 				AS_CLRUNMAPWAIT(seg->s_as);
8802 				cv_broadcast(&seg->s_as->a_cv);
8803 			}
8804 			mutex_exit(&seg->s_as->a_contents);
8805 		}
8806 	}
8807 	return (0);
8808 }
8809 /*
8810  * get a memory ID for an addr in a given segment
8811  *
8812  * XXX only creates PAGESIZE pages if anon slots are not initialized.
8813  * At fault time they will be relocated into larger pages.
8814  */
8815 static int
8816 segvn_getmemid(struct seg *seg, caddr_t addr, memid_t *memidp)
8817 {
8818 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8819 	struct anon 	*ap = NULL;
8820 	ulong_t		anon_index;
8821 	struct anon_map	*amp;
8822 	anon_sync_obj_t cookie;
8823 
8824 	if (svd->type == MAP_PRIVATE) {
8825 		memidp->val[0] = (uintptr_t)seg->s_as;
8826 		memidp->val[1] = (uintptr_t)addr;
8827 		return (0);
8828 	}
8829 
8830 	if (svd->type == MAP_SHARED) {
8831 		if (svd->vp) {
8832 			memidp->val[0] = (uintptr_t)svd->vp;
8833 			memidp->val[1] = (u_longlong_t)svd->offset +
8834 			    (uintptr_t)(addr - seg->s_base);
8835 			return (0);
8836 		} else {
8837 
8838 			SEGVN_LOCK_ENTER(seg->s_as, &svd->lock, RW_READER);
8839 			if ((amp = svd->amp) != NULL) {
8840 				anon_index = svd->anon_index +
8841 				    seg_page(seg, addr);
8842 			}
8843 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
8844 
8845 			ASSERT(amp != NULL);
8846 
8847 			ANON_LOCK_ENTER(&amp->a_rwlock, RW_READER);
8848 			anon_array_enter(amp, anon_index, &cookie);
8849 			ap = anon_get_ptr(amp->ahp, anon_index);
8850 			if (ap == NULL) {
8851 				page_t		*pp;
8852 
8853 				pp = anon_zero(seg, addr, &ap, svd->cred);
8854 				if (pp == NULL) {
8855 					anon_array_exit(&cookie);
8856 					ANON_LOCK_EXIT(&amp->a_rwlock);
8857 					return (ENOMEM);
8858 				}
8859 				ASSERT(anon_get_ptr(amp->ahp, anon_index)
8860 								== NULL);
8861 				(void) anon_set_ptr(amp->ahp, anon_index,
8862 				    ap, ANON_SLEEP);
8863 				page_unlock(pp);
8864 			}
8865 
8866 			anon_array_exit(&cookie);
8867 			ANON_LOCK_EXIT(&amp->a_rwlock);
8868 
8869 			memidp->val[0] = (uintptr_t)ap;
8870 			memidp->val[1] = (uintptr_t)addr & PAGEOFFSET;
8871 			return (0);
8872 		}
8873 	}
8874 	return (EINVAL);
8875 }
8876 
8877 static int
8878 sameprot(struct seg *seg, caddr_t a, size_t len)
8879 {
8880 	struct segvn_data *svd = (struct segvn_data *)seg->s_data;
8881 	struct vpage *vpage;
8882 	spgcnt_t pages = btop(len);
8883 	uint_t prot;
8884 
8885 	if (svd->pageprot == 0)
8886 		return (1);
8887 
8888 	ASSERT(svd->vpage != NULL);
8889 
8890 	vpage = &svd->vpage[seg_page(seg, a)];
8891 	prot = VPP_PROT(vpage);
8892 	vpage++;
8893 	pages--;
8894 	while (pages-- > 0) {
8895 		if (prot != VPP_PROT(vpage))
8896 			return (0);
8897 		vpage++;
8898 	}
8899 	return (1);
8900 }
8901 
8902 /*
8903  * Get memory allocation policy info for specified address in given segment
8904  */
8905 static lgrp_mem_policy_info_t *
8906 segvn_getpolicy(struct seg *seg, caddr_t addr)
8907 {
8908 	struct anon_map		*amp;
8909 	ulong_t			anon_index;
8910 	lgrp_mem_policy_info_t	*policy_info;
8911 	struct segvn_data	*svn_data;
8912 	u_offset_t		vn_off;
8913 	vnode_t			*vp;
8914 
8915 	ASSERT(seg != NULL);
8916 
8917 	svn_data = (struct segvn_data *)seg->s_data;
8918 	if (svn_data == NULL)
8919 		return (NULL);
8920 
8921 	/*
8922 	 * Get policy info for private or shared memory
8923 	 */
8924 	if (svn_data->type != MAP_SHARED) {
8925 		if (svn_data->tr_state != SEGVN_TR_ON) {
8926 			policy_info = &svn_data->policy_info;
8927 		} else {
8928 			policy_info = &svn_data->tr_policy_info;
8929 			ASSERT(policy_info->mem_policy ==
8930 			    LGRP_MEM_POLICY_NEXT_SEG);
8931 		}
8932 	} else {
8933 		amp = svn_data->amp;
8934 		anon_index = svn_data->anon_index + seg_page(seg, addr);
8935 		vp = svn_data->vp;
8936 		vn_off = svn_data->offset + (uintptr_t)(addr - seg->s_base);
8937 		policy_info = lgrp_shm_policy_get(amp, anon_index, vp, vn_off);
8938 	}
8939 
8940 	return (policy_info);
8941 }
8942 
8943 /*ARGSUSED*/
8944 static int
8945 segvn_capable(struct seg *seg, segcapability_t capability)
8946 {
8947 	return (0);
8948 }
8949 
8950 /*
8951  * Bind text vnode segment to an amp. If we bind successfully mappings will be
8952  * established to per vnode mapping per lgroup amp pages instead of to vnode
8953  * pages. There's one amp per vnode text mapping per lgroup. Many processes
8954  * may share the same text replication amp. If a suitable amp doesn't already
8955  * exist in svntr hash table create a new one.  We may fail to bind to amp if
8956  * segment is not eligible for text replication.  Code below first checks for
8957  * these conditions. If binding is successful segment tr_state is set to on
8958  * and svd->amp points to the amp to use. Otherwise tr_state is set to off and
8959  * svd->amp remains as NULL.
8960  */
8961 static void
8962 segvn_textrepl(struct seg *seg)
8963 {
8964 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
8965 	vnode_t			*vp = svd->vp;
8966 	u_offset_t		off = svd->offset;
8967 	size_t			size = seg->s_size;
8968 	u_offset_t		eoff = off + size;
8969 	uint_t			szc = seg->s_szc;
8970 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
8971 	svntr_t			*svntrp;
8972 	struct vattr		va;
8973 	proc_t			*p = seg->s_as->a_proc;
8974 	lgrp_id_t		lgrp_id;
8975 	lgrp_id_t		olid;
8976 	int			first;
8977 	struct anon_map		*amp;
8978 
8979 	ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
8980 	ASSERT(SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
8981 	ASSERT(p != NULL);
8982 	ASSERT(svd->tr_state == SEGVN_TR_INIT);
8983 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
8984 	ASSERT(svd->flags & MAP_TEXT);
8985 	ASSERT(svd->type == MAP_PRIVATE);
8986 	ASSERT(vp != NULL && svd->amp == NULL);
8987 	ASSERT(!svd->pageprot && !(svd->prot & PROT_WRITE));
8988 	ASSERT(!(svd->flags & MAP_NORESERVE) && svd->swresv == 0);
8989 	ASSERT(seg->s_as != &kas);
8990 	ASSERT(off < eoff);
8991 	ASSERT(svntr_hashtab != NULL);
8992 
8993 	/*
8994 	 * If numa optimizations are no longer desired bail out.
8995 	 */
8996 	if (!lgrp_optimizations()) {
8997 		svd->tr_state = SEGVN_TR_OFF;
8998 		return;
8999 	}
9000 
9001 	/*
9002 	 * Avoid creating anon maps with size bigger than the file size.
9003 	 * If VOP_GETATTR() call fails bail out.
9004 	 */
9005 	va.va_mask = AT_SIZE | AT_MTIME | AT_CTIME;
9006 	if (VOP_GETATTR(vp, &va, 0, svd->cred) != 0) {
9007 		svd->tr_state = SEGVN_TR_OFF;
9008 		SEGVN_TR_ADDSTAT(gaerr);
9009 		return;
9010 	}
9011 	if (btopr(va.va_size) < btopr(eoff)) {
9012 		svd->tr_state = SEGVN_TR_OFF;
9013 		SEGVN_TR_ADDSTAT(overmap);
9014 		return;
9015 	}
9016 
9017 	/*
9018 	 * VVMEXEC may not be set yet if exec() prefaults text segment. Set
9019 	 * this flag now before vn_is_mapped(V_WRITE) so that MAP_SHARED
9020 	 * mapping that checks if trcache for this vnode needs to be
9021 	 * invalidated can't miss us.
9022 	 */
9023 	if (!(vp->v_flag & VVMEXEC)) {
9024 		mutex_enter(&vp->v_lock);
9025 		vp->v_flag |= VVMEXEC;
9026 		mutex_exit(&vp->v_lock);
9027 	}
9028 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9029 	/*
9030 	 * Bail out if potentially MAP_SHARED writable mappings exist to this
9031 	 * vnode.  We don't want to use old file contents from existing
9032 	 * replicas if this mapping was established after the original file
9033 	 * was changed.
9034 	 */
9035 	if (vn_is_mapped(vp, V_WRITE)) {
9036 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9037 		svd->tr_state = SEGVN_TR_OFF;
9038 		SEGVN_TR_ADDSTAT(wrcnt);
9039 		return;
9040 	}
9041 	svntrp = svntr_hashtab[hash].tr_head;
9042 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9043 		ASSERT(svntrp->tr_refcnt != 0);
9044 		if (svntrp->tr_vp != vp) {
9045 			continue;
9046 		}
9047 
9048 		/*
9049 		 * Bail out if the file or its attributes were changed after
9050 		 * this replication entry was created since we need to use the
9051 		 * latest file contents. Note that mtime test alone is not
9052 		 * sufficient because a user can explicitly change mtime via
9053 		 * utimes(2) interfaces back to the old value after modifiying
9054 		 * the file contents. To detect this case we also have to test
9055 		 * ctime which among other things records the time of the last
9056 		 * mtime change by utimes(2). ctime is not changed when the file
9057 		 * is only read or executed so we expect that typically existing
9058 		 * replication amp's can be used most of the time.
9059 		 */
9060 		if (!svntrp->tr_valid ||
9061 		    svntrp->tr_mtime.tv_sec != va.va_mtime.tv_sec ||
9062 		    svntrp->tr_mtime.tv_nsec != va.va_mtime.tv_nsec ||
9063 		    svntrp->tr_ctime.tv_sec != va.va_ctime.tv_sec ||
9064 		    svntrp->tr_ctime.tv_nsec != va.va_ctime.tv_nsec) {
9065 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9066 			svd->tr_state = SEGVN_TR_OFF;
9067 			SEGVN_TR_ADDSTAT(stale);
9068 			return;
9069 		}
9070 		/*
9071 		 * if off, eoff and szc match current segment we found the
9072 		 * existing entry we can use.
9073 		 */
9074 		if (svntrp->tr_off == off && svntrp->tr_eoff == eoff &&
9075 		    svntrp->tr_szc == szc) {
9076 			break;
9077 		}
9078 		/*
9079 		 * Don't create different but overlapping in file offsets
9080 		 * entries to avoid replication of the same file pages more
9081 		 * than once per lgroup.
9082 		 */
9083 		if ((off >= svntrp->tr_off && off < svntrp->tr_eoff) ||
9084 		    (eoff > svntrp->tr_off && eoff <= svntrp->tr_eoff)) {
9085 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9086 			svd->tr_state = SEGVN_TR_OFF;
9087 			SEGVN_TR_ADDSTAT(overlap);
9088 			return;
9089 		}
9090 	}
9091 	/*
9092 	 * If we didn't find existing entry create a new one.
9093 	 */
9094 	if (svntrp == NULL) {
9095 		svntrp = kmem_cache_alloc(svntr_cache, KM_NOSLEEP);
9096 		if (svntrp == NULL) {
9097 			mutex_exit(&svntr_hashtab[hash].tr_lock);
9098 			svd->tr_state = SEGVN_TR_OFF;
9099 			SEGVN_TR_ADDSTAT(nokmem);
9100 			return;
9101 		}
9102 #ifdef DEBUG
9103 		{
9104 			lgrp_id_t i;
9105 			for (i = 0; i < NLGRPS_MAX; i++) {
9106 				ASSERT(svntrp->tr_amp[i] == NULL);
9107 			}
9108 		}
9109 #endif /* DEBUG */
9110 		svntrp->tr_vp = vp;
9111 		svntrp->tr_off = off;
9112 		svntrp->tr_eoff = eoff;
9113 		svntrp->tr_szc = szc;
9114 		svntrp->tr_valid = 1;
9115 		svntrp->tr_mtime = va.va_mtime;
9116 		svntrp->tr_ctime = va.va_ctime;
9117 		svntrp->tr_refcnt = 0;
9118 		svntrp->tr_next = svntr_hashtab[hash].tr_head;
9119 		svntr_hashtab[hash].tr_head = svntrp;
9120 	}
9121 	first = 1;
9122 again:
9123 	/*
9124 	 * We want to pick a replica with pages on main thread's (t_tid = 1,
9125 	 * aka T1) lgrp. Currently text replication is only optimized for
9126 	 * workloads that either have all threads of a process on the same
9127 	 * lgrp or execute their large text primarily on main thread.
9128 	 */
9129 	lgrp_id = p->p_t1_lgrpid;
9130 	if (lgrp_id == LGRP_NONE) {
9131 		/*
9132 		 * In case exec() prefaults text on non main thread use
9133 		 * current thread lgrpid.  It will become main thread anyway
9134 		 * soon.
9135 		 */
9136 		lgrp_id = lgrp_home_id(curthread);
9137 	}
9138 	/*
9139 	 * Set p_tr_lgrpid to lgrpid if it hasn't been set yet.  Otherwise
9140 	 * just set it to NLGRPS_MAX if it's different from current process T1
9141 	 * home lgrp.  p_tr_lgrpid is used to detect if process uses text
9142 	 * replication and T1 new home is different from lgrp used for text
9143 	 * replication. When this happens asyncronous segvn thread rechecks if
9144 	 * segments should change lgrps used for text replication.  If we fail
9145 	 * to set p_tr_lgrpid with cas32 then set it to NLGRPS_MAX without cas
9146 	 * if it's not already NLGRPS_MAX and not equal lgrp_id we want to
9147 	 * use.  We don't need to use cas in this case because another thread
9148 	 * that races in between our non atomic check and set may only change
9149 	 * p_tr_lgrpid to NLGRPS_MAX at this point.
9150 	 */
9151 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9152 	olid = p->p_tr_lgrpid;
9153 	if (lgrp_id != olid && olid != NLGRPS_MAX) {
9154 		lgrp_id_t nlid = (olid == LGRP_NONE) ? lgrp_id : NLGRPS_MAX;
9155 		if (cas32((uint32_t *)&p->p_tr_lgrpid, olid, nlid) != olid) {
9156 			olid = p->p_tr_lgrpid;
9157 			ASSERT(olid != LGRP_NONE);
9158 			if (olid != lgrp_id && olid != NLGRPS_MAX) {
9159 				p->p_tr_lgrpid = NLGRPS_MAX;
9160 			}
9161 		}
9162 		ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9163 		membar_producer();
9164 		/*
9165 		 * lgrp_move_thread() won't schedule async recheck after
9166 		 * p->p_t1_lgrpid update unless p->p_tr_lgrpid is not
9167 		 * LGRP_NONE. Recheck p_t1_lgrpid once now that p->p_tr_lgrpid
9168 		 * is not LGRP_NONE.
9169 		 */
9170 		if (first && p->p_t1_lgrpid != LGRP_NONE &&
9171 		    p->p_t1_lgrpid != lgrp_id) {
9172 			first = 0;
9173 			goto again;
9174 		}
9175 	}
9176 	/*
9177 	 * If no amp was created yet for lgrp_id create a new one as long as
9178 	 * we have enough memory to afford it.
9179 	 */
9180 	if ((amp = svntrp->tr_amp[lgrp_id]) == NULL) {
9181 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9182 		if (trmem > segvn_textrepl_max_bytes) {
9183 			SEGVN_TR_ADDSTAT(normem);
9184 			goto fail;
9185 		}
9186 		if (anon_try_resv_zone(size, NULL) == 0) {
9187 			SEGVN_TR_ADDSTAT(noanon);
9188 			goto fail;
9189 		}
9190 		amp = anonmap_alloc(size, size, ANON_NOSLEEP);
9191 		if (amp == NULL) {
9192 			anon_unresv_zone(size, NULL);
9193 			SEGVN_TR_ADDSTAT(nokmem);
9194 			goto fail;
9195 		}
9196 		ASSERT(amp->refcnt == 1);
9197 		amp->a_szc = szc;
9198 		svntrp->tr_amp[lgrp_id] = amp;
9199 		SEGVN_TR_ADDSTAT(newamp);
9200 	}
9201 	svntrp->tr_refcnt++;
9202 	ASSERT(svd->svn_trnext == NULL);
9203 	ASSERT(svd->svn_trprev == NULL);
9204 	svd->svn_trnext = svntrp->tr_svnhead;
9205 	svd->svn_trprev = NULL;
9206 	if (svntrp->tr_svnhead != NULL) {
9207 		svntrp->tr_svnhead->svn_trprev = svd;
9208 	}
9209 	svntrp->tr_svnhead = svd;
9210 	ASSERT(amp->a_szc == szc && amp->size == size && amp->swresv == size);
9211 	ASSERT(amp->refcnt >= 1);
9212 	svd->amp = amp;
9213 	svd->anon_index = 0;
9214 	svd->tr_policy_info.mem_policy = LGRP_MEM_POLICY_NEXT_SEG;
9215 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
9216 	svd->tr_state = SEGVN_TR_ON;
9217 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9218 	SEGVN_TR_ADDSTAT(repl);
9219 	return;
9220 fail:
9221 	ASSERT(segvn_textrepl_bytes >= size);
9222 	atomic_add_long(&segvn_textrepl_bytes, -size);
9223 	ASSERT(svntrp != NULL);
9224 	ASSERT(svntrp->tr_amp[lgrp_id] == NULL);
9225 	if (svntrp->tr_refcnt == 0) {
9226 		ASSERT(svntrp == svntr_hashtab[hash].tr_head);
9227 		svntr_hashtab[hash].tr_head = svntrp->tr_next;
9228 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9229 		kmem_cache_free(svntr_cache, svntrp);
9230 	} else {
9231 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9232 	}
9233 	svd->tr_state = SEGVN_TR_OFF;
9234 }
9235 
9236 /*
9237  * Convert seg back to regular vnode mapping seg by unbinding it from its text
9238  * replication amp.  This routine is most typically called when segment is
9239  * unmapped but can also be called when segment no longer qualifies for text
9240  * replication (e.g. due to protection changes). If unload_unmap is set use
9241  * HAT_UNLOAD_UNMAP flag in hat_unload_callback().  If we are the last user of
9242  * svntr free all its anon maps and remove it from the hash table.
9243  */
9244 static void
9245 segvn_textunrepl(struct seg *seg, int unload_unmap)
9246 {
9247 	struct segvn_data	*svd = (struct segvn_data *)seg->s_data;
9248 	vnode_t			*vp = svd->vp;
9249 	u_offset_t		off = svd->offset;
9250 	size_t			size = seg->s_size;
9251 	u_offset_t		eoff = off + size;
9252 	uint_t			szc = seg->s_szc;
9253 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9254 	svntr_t			*svntrp;
9255 	svntr_t			**prv_svntrp;
9256 	lgrp_id_t		lgrp_id = svd->tr_policy_info.mem_lgrpid;
9257 	lgrp_id_t		i;
9258 
9259 	ASSERT(AS_LOCK_HELD(seg->s_as, &seg->s_as->a_lock));
9260 	ASSERT(AS_WRITE_HELD(seg->s_as, &seg->s_as->a_lock) ||
9261 	    SEGVN_WRITE_HELD(seg->s_as, &svd->lock));
9262 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9263 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9264 	ASSERT(svd->amp != NULL);
9265 	ASSERT(svd->amp->refcnt >= 1);
9266 	ASSERT(svd->anon_index == 0);
9267 	ASSERT(lgrp_id != LGRP_NONE && lgrp_id < NLGRPS_MAX);
9268 	ASSERT(svntr_hashtab != NULL);
9269 
9270 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9271 	prv_svntrp = &svntr_hashtab[hash].tr_head;
9272 	for (; (svntrp = *prv_svntrp) != NULL; prv_svntrp = &svntrp->tr_next) {
9273 		ASSERT(svntrp->tr_refcnt != 0);
9274 		if (svntrp->tr_vp == vp && svntrp->tr_off == off &&
9275 		    svntrp->tr_eoff == eoff && svntrp->tr_szc == szc) {
9276 			break;
9277 		}
9278 	}
9279 	if (svntrp == NULL) {
9280 		panic("segvn_textunrepl: svntr record not found");
9281 	}
9282 	if (svntrp->tr_amp[lgrp_id] != svd->amp) {
9283 		panic("segvn_textunrepl: amp mismatch");
9284 	}
9285 	svd->tr_state = SEGVN_TR_OFF;
9286 	svd->amp = NULL;
9287 	if (svd->svn_trprev == NULL) {
9288 		ASSERT(svntrp->tr_svnhead == svd);
9289 		svntrp->tr_svnhead = svd->svn_trnext;
9290 		if (svntrp->tr_svnhead != NULL) {
9291 			svntrp->tr_svnhead->svn_trprev = NULL;
9292 		}
9293 		svd->svn_trnext = NULL;
9294 	} else {
9295 		svd->svn_trprev->svn_trnext = svd->svn_trnext;
9296 		if (svd->svn_trnext != NULL) {
9297 			svd->svn_trnext->svn_trprev = svd->svn_trprev;
9298 			svd->svn_trnext = NULL;
9299 		}
9300 		svd->svn_trprev = NULL;
9301 	}
9302 	if (--svntrp->tr_refcnt) {
9303 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9304 		goto done;
9305 	}
9306 	*prv_svntrp = svntrp->tr_next;
9307 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9308 	for (i = 0; i < NLGRPS_MAX; i++) {
9309 		struct anon_map *amp = svntrp->tr_amp[i];
9310 		if (amp == NULL) {
9311 			continue;
9312 		}
9313 		ASSERT(amp->refcnt == 1);
9314 		ASSERT(amp->swresv == size);
9315 		ASSERT(amp->size == size);
9316 		ASSERT(amp->a_szc == szc);
9317 		if (amp->a_szc != 0) {
9318 			anon_free_pages(amp->ahp, 0, size, szc);
9319 		} else {
9320 			anon_free(amp->ahp, 0, size);
9321 		}
9322 		svntrp->tr_amp[i] = NULL;
9323 		ASSERT(segvn_textrepl_bytes >= size);
9324 		atomic_add_long(&segvn_textrepl_bytes, -size);
9325 		anon_unresv_zone(amp->swresv, NULL);
9326 		amp->refcnt = 0;
9327 		anonmap_free(amp);
9328 	}
9329 	kmem_cache_free(svntr_cache, svntrp);
9330 done:
9331 	hat_unload_callback(seg->s_as->a_hat, seg->s_base, size,
9332 	    unload_unmap ? HAT_UNLOAD_UNMAP : 0, NULL);
9333 }
9334 
9335 /*
9336  * This is called when a MAP_SHARED writabble mapping is created to a vnode
9337  * that is currently used for execution (VVMEXEC flag is set). In this case we
9338  * need to prevent further use of existing replicas.
9339  */
9340 static void
9341 segvn_inval_trcache(vnode_t *vp)
9342 {
9343 	ulong_t			hash = SVNTR_HASH_FUNC(vp);
9344 	svntr_t			*svntrp;
9345 
9346 	ASSERT(vp->v_flag & VVMEXEC);
9347 
9348 	if (svntr_hashtab == NULL) {
9349 		return;
9350 	}
9351 
9352 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9353 	svntrp = svntr_hashtab[hash].tr_head;
9354 	for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9355 		ASSERT(svntrp->tr_refcnt != 0);
9356 		if (svntrp->tr_vp == vp && svntrp->tr_valid) {
9357 			svntrp->tr_valid = 0;
9358 		}
9359 	}
9360 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9361 }
9362 
9363 static void
9364 segvn_trasync_thread(void)
9365 {
9366 	callb_cpr_t cpr_info;
9367 	kmutex_t cpr_lock;	/* just for CPR stuff */
9368 
9369 	mutex_init(&cpr_lock, NULL, MUTEX_DEFAULT, NULL);
9370 
9371 	CALLB_CPR_INIT(&cpr_info, &cpr_lock,
9372 	    callb_generic_cpr, "segvn_async");
9373 
9374 	if (segvn_update_textrepl_interval == 0) {
9375 		segvn_update_textrepl_interval = segvn_update_tr_time * hz;
9376 	} else {
9377 		segvn_update_textrepl_interval *= hz;
9378 	}
9379 	(void) timeout(segvn_trupdate_wakeup, NULL,
9380 	    segvn_update_textrepl_interval);
9381 
9382 	for (;;) {
9383 		mutex_enter(&cpr_lock);
9384 		CALLB_CPR_SAFE_BEGIN(&cpr_info);
9385 		mutex_exit(&cpr_lock);
9386 		sema_p(&segvn_trasync_sem);
9387 		mutex_enter(&cpr_lock);
9388 		CALLB_CPR_SAFE_END(&cpr_info, &cpr_lock);
9389 		mutex_exit(&cpr_lock);
9390 		segvn_trupdate();
9391 	}
9392 }
9393 
9394 static uint64_t segvn_lgrp_trthr_migrs_snpsht = 0;
9395 
9396 static void
9397 segvn_trupdate_wakeup(void *dummy)
9398 {
9399 	uint64_t cur_lgrp_trthr_migrs = lgrp_get_trthr_migrations();
9400 
9401 	if (cur_lgrp_trthr_migrs != segvn_lgrp_trthr_migrs_snpsht) {
9402 		segvn_lgrp_trthr_migrs_snpsht = cur_lgrp_trthr_migrs;
9403 		sema_v(&segvn_trasync_sem);
9404 	}
9405 
9406 	if (!segvn_disable_textrepl_update &&
9407 	    segvn_update_textrepl_interval != 0) {
9408 		(void) timeout(segvn_trupdate_wakeup, dummy,
9409 		    segvn_update_textrepl_interval);
9410 	}
9411 }
9412 
9413 static void
9414 segvn_trupdate(void)
9415 {
9416 	ulong_t		hash;
9417 	svntr_t		*svntrp;
9418 	segvn_data_t	*svd;
9419 
9420 	ASSERT(svntr_hashtab != NULL);
9421 
9422 	for (hash = 0; hash < svntr_hashtab_sz; hash++) {
9423 		mutex_enter(&svntr_hashtab[hash].tr_lock);
9424 		svntrp = svntr_hashtab[hash].tr_head;
9425 		for (; svntrp != NULL; svntrp = svntrp->tr_next) {
9426 			ASSERT(svntrp->tr_refcnt != 0);
9427 			svd = svntrp->tr_svnhead;
9428 			for (; svd != NULL; svd = svd->svn_trnext) {
9429 				segvn_trupdate_seg(svd->seg, svd, svntrp,
9430 				    hash);
9431 			}
9432 		}
9433 		mutex_exit(&svntr_hashtab[hash].tr_lock);
9434 	}
9435 }
9436 
9437 static void
9438 segvn_trupdate_seg(struct seg *seg,
9439 	segvn_data_t *svd,
9440 	svntr_t *svntrp,
9441 	ulong_t hash)
9442 {
9443 	proc_t			*p;
9444 	lgrp_id_t		lgrp_id;
9445 	struct as		*as;
9446 	size_t			size;
9447 	struct anon_map		*amp;
9448 
9449 	ASSERT(svd->vp != NULL);
9450 	ASSERT(svd->vp == svntrp->tr_vp);
9451 	ASSERT(svd->offset == svntrp->tr_off);
9452 	ASSERT(svd->offset + seg->s_size == svntrp->tr_eoff);
9453 	ASSERT(seg != NULL);
9454 	ASSERT(svd->seg == seg);
9455 	ASSERT(seg->s_data == (void *)svd);
9456 	ASSERT(seg->s_szc == svntrp->tr_szc);
9457 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9458 	ASSERT(!HAT_IS_REGION_COOKIE_VALID(svd->rcookie));
9459 	ASSERT(svd->amp != NULL);
9460 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
9461 	ASSERT(svd->tr_policy_info.mem_lgrpid != LGRP_NONE);
9462 	ASSERT(svd->tr_policy_info.mem_lgrpid < NLGRPS_MAX);
9463 	ASSERT(svntrp->tr_amp[svd->tr_policy_info.mem_lgrpid] == svd->amp);
9464 	ASSERT(svntrp->tr_refcnt != 0);
9465 	ASSERT(mutex_owned(&svntr_hashtab[hash].tr_lock));
9466 
9467 	as = seg->s_as;
9468 	ASSERT(as != NULL && as != &kas);
9469 	p = as->a_proc;
9470 	ASSERT(p != NULL);
9471 	ASSERT(p->p_tr_lgrpid != LGRP_NONE);
9472 	lgrp_id = p->p_t1_lgrpid;
9473 	if (lgrp_id == LGRP_NONE) {
9474 		return;
9475 	}
9476 	ASSERT(lgrp_id < NLGRPS_MAX);
9477 	if (svd->tr_policy_info.mem_lgrpid == lgrp_id) {
9478 		return;
9479 	}
9480 
9481 	/*
9482 	 * Use tryenter locking since we are locking as/seg and svntr hash
9483 	 * lock in reverse from syncrounous thread order.
9484 	 */
9485 	if (!AS_LOCK_TRYENTER(as, &as->a_lock, RW_READER)) {
9486 		SEGVN_TR_ADDSTAT(nolock);
9487 		if (segvn_lgrp_trthr_migrs_snpsht) {
9488 			segvn_lgrp_trthr_migrs_snpsht = 0;
9489 		}
9490 		return;
9491 	}
9492 	if (!SEGVN_LOCK_TRYENTER(seg->s_as, &svd->lock, RW_WRITER)) {
9493 		AS_LOCK_EXIT(as, &as->a_lock);
9494 		SEGVN_TR_ADDSTAT(nolock);
9495 		if (segvn_lgrp_trthr_migrs_snpsht) {
9496 			segvn_lgrp_trthr_migrs_snpsht = 0;
9497 		}
9498 		return;
9499 	}
9500 	size = seg->s_size;
9501 	if (svntrp->tr_amp[lgrp_id] == NULL) {
9502 		size_t trmem = atomic_add_long_nv(&segvn_textrepl_bytes, size);
9503 		if (trmem > segvn_textrepl_max_bytes) {
9504 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9505 			AS_LOCK_EXIT(as, &as->a_lock);
9506 			atomic_add_long(&segvn_textrepl_bytes, -size);
9507 			SEGVN_TR_ADDSTAT(normem);
9508 			return;
9509 		}
9510 		if (anon_try_resv_zone(size, NULL) == 0) {
9511 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9512 			AS_LOCK_EXIT(as, &as->a_lock);
9513 			atomic_add_long(&segvn_textrepl_bytes, -size);
9514 			SEGVN_TR_ADDSTAT(noanon);
9515 			return;
9516 		}
9517 		amp = anonmap_alloc(size, size, KM_NOSLEEP);
9518 		if (amp == NULL) {
9519 			SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9520 			AS_LOCK_EXIT(as, &as->a_lock);
9521 			atomic_add_long(&segvn_textrepl_bytes, -size);
9522 			anon_unresv_zone(size, NULL);
9523 			SEGVN_TR_ADDSTAT(nokmem);
9524 			return;
9525 		}
9526 		ASSERT(amp->refcnt == 1);
9527 		amp->a_szc = seg->s_szc;
9528 		svntrp->tr_amp[lgrp_id] = amp;
9529 	}
9530 	/*
9531 	 * We don't need to drop the bucket lock but here we give other
9532 	 * threads a chance.  svntr and svd can't be unlinked as long as
9533 	 * segment lock is held as a writer and AS held as well.  After we
9534 	 * retake bucket lock we'll continue from where we left. We'll be able
9535 	 * to reach the end of either list since new entries are always added
9536 	 * to the beginning of the lists.
9537 	 */
9538 	mutex_exit(&svntr_hashtab[hash].tr_lock);
9539 	hat_unload_callback(as->a_hat, seg->s_base, size, 0, NULL);
9540 	mutex_enter(&svntr_hashtab[hash].tr_lock);
9541 
9542 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9543 	ASSERT(svd->amp != NULL);
9544 	ASSERT(svd->tr_policy_info.mem_policy == LGRP_MEM_POLICY_NEXT_SEG);
9545 	ASSERT(svd->tr_policy_info.mem_lgrpid != lgrp_id);
9546 	ASSERT(svd->amp != svntrp->tr_amp[lgrp_id]);
9547 
9548 	svd->tr_policy_info.mem_lgrpid = lgrp_id;
9549 	svd->amp = svntrp->tr_amp[lgrp_id];
9550 	p->p_tr_lgrpid = NLGRPS_MAX;
9551 	SEGVN_LOCK_EXIT(seg->s_as, &svd->lock);
9552 	AS_LOCK_EXIT(as, &as->a_lock);
9553 
9554 	ASSERT(svntrp->tr_refcnt != 0);
9555 	ASSERT(svd->vp == svntrp->tr_vp);
9556 	ASSERT(svd->tr_policy_info.mem_lgrpid == lgrp_id);
9557 	ASSERT(svd->amp != NULL && svd->amp == svntrp->tr_amp[lgrp_id]);
9558 	ASSERT(svd->seg == seg);
9559 	ASSERT(svd->tr_state == SEGVN_TR_ON);
9560 
9561 	SEGVN_TR_ADDSTAT(asyncrepl);
9562 }
9563