xref: /illumos-gate/usr/src/uts/common/vm/vm_as.c (revision 2570281c)
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 2010 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  * Copyright 2018 Joyent, Inc.
25  * Copyright (c) 2016 by Delphix. All rights reserved.
26  */
27 
28 /*	Copyright (c) 1984, 1986, 1987, 1988, 1989 AT&T	*/
29 /*	  All Rights Reserved  	*/
30 
31 /*
32  * University Copyright- Copyright (c) 1982, 1986, 1988
33  * The Regents of the University of California
34  * All Rights Reserved
35  *
36  * University Acknowledgment- Portions of this document are derived from
37  * software developed by the University of California, Berkeley, and its
38  * contributors.
39  */
40 
41 /*
42  * VM - address spaces.
43  */
44 
45 #include <sys/types.h>
46 #include <sys/t_lock.h>
47 #include <sys/param.h>
48 #include <sys/errno.h>
49 #include <sys/systm.h>
50 #include <sys/mman.h>
51 #include <sys/sysmacros.h>
52 #include <sys/cpuvar.h>
53 #include <sys/sysinfo.h>
54 #include <sys/kmem.h>
55 #include <sys/vnode.h>
56 #include <sys/vmsystm.h>
57 #include <sys/cmn_err.h>
58 #include <sys/debug.h>
59 #include <sys/vtrace.h>
60 
61 #include <vm/hat.h>
62 #include <vm/as.h>
63 #include <vm/seg.h>
64 #include <vm/seg_vn.h>
65 #include <vm/seg_dev.h>
66 #include <vm/seg_kmem.h>
67 #include <vm/seg_map.h>
68 #include <vm/seg_spt.h>
69 #include <vm/seg_hole.h>
70 #include <vm/page.h>
71 
72 clock_t deadlk_wait = 1; /* number of ticks to wait before retrying */
73 
74 static struct kmem_cache *as_cache;
75 
76 static void as_setwatchprot(struct as *, caddr_t, size_t, uint_t);
77 static void as_clearwatchprot(struct as *, caddr_t, size_t);
78 
79 
80 /*
81  * Verifying the segment lists is very time-consuming; it may not be
82  * desirable always to define VERIFY_SEGLIST when DEBUG is set.
83  */
84 #ifdef DEBUG
85 #define	VERIFY_SEGLIST
86 int do_as_verify = 0;
87 #endif
88 
89 /*
90  * Allocate a new callback data structure entry and fill in the events of
91  * interest, the address range of interest, and the callback argument.
92  * Link the entry on the as->a_callbacks list. A callback entry for the
93  * entire address space may be specified with vaddr = 0 and size = -1.
94  *
95  * CALLERS RESPONSIBILITY: If not calling from within the process context for
96  * the specified as, the caller must guarantee persistence of the specified as
97  * for the duration of this function (eg. pages being locked within the as
98  * will guarantee persistence).
99  */
100 int
as_add_callback(struct as * as,void (* cb_func)(),void * arg,uint_t events,caddr_t vaddr,size_t size,int sleepflag)101 as_add_callback(struct as *as, void (*cb_func)(), void *arg, uint_t events,
102     caddr_t vaddr, size_t size, int sleepflag)
103 {
104 	struct as_callback 	*current_head, *cb;
105 	caddr_t 		saddr;
106 	size_t 			rsize;
107 
108 	/* callback function and an event are mandatory */
109 	if ((cb_func == NULL) || ((events & AS_ALL_EVENT) == 0))
110 		return (EINVAL);
111 
112 	/* Adding a callback after as_free has been called is not allowed */
113 	if (as == &kas)
114 		return (ENOMEM);
115 
116 	/*
117 	 * vaddr = 0 and size = -1 is used to indicate that the callback range
118 	 * is the entire address space so no rounding is done in that case.
119 	 */
120 	if (size != -1) {
121 		saddr = (caddr_t)((uintptr_t)vaddr & (uintptr_t)PAGEMASK);
122 		rsize = (((size_t)(vaddr + size) + PAGEOFFSET) & PAGEMASK) -
123 		    (size_t)saddr;
124 		/* check for wraparound */
125 		if (saddr + rsize < saddr)
126 			return (ENOMEM);
127 	} else {
128 		if (vaddr != 0)
129 			return (EINVAL);
130 		saddr = vaddr;
131 		rsize = size;
132 	}
133 
134 	/* Allocate and initialize a callback entry */
135 	cb = kmem_zalloc(sizeof (struct as_callback), sleepflag);
136 	if (cb == NULL)
137 		return (EAGAIN);
138 
139 	cb->ascb_func = cb_func;
140 	cb->ascb_arg = arg;
141 	cb->ascb_events = events;
142 	cb->ascb_saddr = saddr;
143 	cb->ascb_len = rsize;
144 
145 	/* Add the entry to the list */
146 	mutex_enter(&as->a_contents);
147 	current_head = as->a_callbacks;
148 	as->a_callbacks = cb;
149 	cb->ascb_next = current_head;
150 
151 	/*
152 	 * The call to this function may lose in a race with
153 	 * a pertinent event - eg. a thread does long term memory locking
154 	 * but before the callback is added another thread executes as_unmap.
155 	 * A broadcast here resolves that.
156 	 */
157 	if ((cb->ascb_events & AS_UNMAPWAIT_EVENT) && AS_ISUNMAPWAIT(as)) {
158 		AS_CLRUNMAPWAIT(as);
159 		cv_broadcast(&as->a_cv);
160 	}
161 
162 	mutex_exit(&as->a_contents);
163 	return (0);
164 }
165 
166 /*
167  * Search the callback list for an entry which pertains to arg.
168  *
169  * This is called from within the client upon completion of the callback.
170  * RETURN VALUES:
171  *	AS_CALLBACK_DELETED  (callback entry found and deleted)
172  *	AS_CALLBACK_NOTFOUND (no callback entry found - this is ok)
173  *	AS_CALLBACK_DELETE_DEFERRED (callback is in process, delete of this
174  *			entry will be made in as_do_callbacks)
175  *
176  * If as_delete_callback encounters a matching entry with AS_CALLBACK_CALLED
177  * set, it indicates that as_do_callbacks is processing this entry.  The
178  * AS_ALL_EVENT events are cleared in the entry, and a broadcast is made
179  * to unblock as_do_callbacks, in case it is blocked.
180  *
181  * CALLERS RESPONSIBILITY: If not calling from within the process context for
182  * the specified as, the caller must guarantee persistence of the specified as
183  * for the duration of this function (eg. pages being locked within the as
184  * will guarantee persistence).
185  */
186 uint_t
as_delete_callback(struct as * as,void * arg)187 as_delete_callback(struct as *as, void *arg)
188 {
189 	struct as_callback **prevcb = &as->a_callbacks;
190 	struct as_callback *cb;
191 	uint_t rc = AS_CALLBACK_NOTFOUND;
192 
193 	mutex_enter(&as->a_contents);
194 	for (cb = as->a_callbacks; cb; prevcb = &cb->ascb_next, cb = *prevcb) {
195 		if (cb->ascb_arg != arg)
196 			continue;
197 
198 		/*
199 		 * If the events indicate AS_CALLBACK_CALLED, just clear
200 		 * AS_ALL_EVENT in the events field and wakeup the thread
201 		 * that may be waiting in as_do_callbacks.  as_do_callbacks
202 		 * will take care of removing this entry from the list.  In
203 		 * that case, return AS_CALLBACK_DELETE_DEFERRED.  Otherwise
204 		 * (AS_CALLBACK_CALLED not set), just remove it from the
205 		 * list, return the memory and return AS_CALLBACK_DELETED.
206 		 */
207 		if ((cb->ascb_events & AS_CALLBACK_CALLED) != 0) {
208 			/* leave AS_CALLBACK_CALLED */
209 			cb->ascb_events &= ~AS_ALL_EVENT;
210 			rc = AS_CALLBACK_DELETE_DEFERRED;
211 			cv_broadcast(&as->a_cv);
212 		} else {
213 			*prevcb = cb->ascb_next;
214 			kmem_free(cb, sizeof (struct as_callback));
215 			rc = AS_CALLBACK_DELETED;
216 		}
217 		break;
218 	}
219 	mutex_exit(&as->a_contents);
220 	return (rc);
221 }
222 
223 /*
224  * Searches the as callback list for a matching entry.
225  * Returns a pointer to the first matching callback, or NULL if
226  * nothing is found.
227  * This function never sleeps so it is ok to call it with more
228  * locks held but the (required) a_contents mutex.
229  *
230  * See also comment on as_do_callbacks below.
231  */
232 static struct as_callback *
as_find_callback(struct as * as,uint_t events,caddr_t event_addr,size_t event_len)233 as_find_callback(struct as *as, uint_t events, caddr_t event_addr,
234     size_t event_len)
235 {
236 	struct as_callback	*cb;
237 
238 	ASSERT(MUTEX_HELD(&as->a_contents));
239 	for (cb = as->a_callbacks; cb != NULL; cb = cb->ascb_next) {
240 		/*
241 		 * If the callback has not already been called, then
242 		 * check if events or address range pertains.  An event_len
243 		 * of zero means do an unconditional callback.
244 		 */
245 		if (((cb->ascb_events & AS_CALLBACK_CALLED) != 0) ||
246 		    ((event_len != 0) && (((cb->ascb_events & events) == 0) ||
247 		    (event_addr + event_len < cb->ascb_saddr) ||
248 		    (event_addr > (cb->ascb_saddr + cb->ascb_len))))) {
249 			continue;
250 		}
251 		break;
252 	}
253 	return (cb);
254 }
255 
256 /*
257  * Executes a given callback and removes it from the callback list for
258  * this address space.
259  * This function may sleep so the caller must drop all locks except
260  * a_contents before calling this func.
261  *
262  * See also comments on as_do_callbacks below.
263  */
264 static void
as_execute_callback(struct as * as,struct as_callback * cb,uint_t events)265 as_execute_callback(struct as *as, struct as_callback *cb,
266     uint_t events)
267 {
268 	struct as_callback **prevcb;
269 	void	*cb_arg;
270 
271 	ASSERT(MUTEX_HELD(&as->a_contents) && (cb->ascb_events & events));
272 	cb->ascb_events |= AS_CALLBACK_CALLED;
273 	mutex_exit(&as->a_contents);
274 	(*cb->ascb_func)(as, cb->ascb_arg, events);
275 	mutex_enter(&as->a_contents);
276 	/*
277 	 * the callback function is required to delete the callback
278 	 * when the callback function determines it is OK for
279 	 * this thread to continue. as_delete_callback will clear
280 	 * the AS_ALL_EVENT in the events field when it is deleted.
281 	 * If the callback function called as_delete_callback,
282 	 * events will already be cleared and there will be no blocking.
283 	 */
284 	while ((cb->ascb_events & events) != 0) {
285 		cv_wait(&as->a_cv, &as->a_contents);
286 	}
287 	/*
288 	 * This entry needs to be taken off the list. Normally, the
289 	 * callback func itself does that, but unfortunately the list
290 	 * may have changed while the callback was running because the
291 	 * a_contents mutex was dropped and someone else other than the
292 	 * callback func itself could have called as_delete_callback,
293 	 * so we have to search to find this entry again.  The entry
294 	 * must have AS_CALLBACK_CALLED, and have the same 'arg'.
295 	 */
296 	cb_arg = cb->ascb_arg;
297 	prevcb = &as->a_callbacks;
298 	for (cb = as->a_callbacks; cb != NULL;
299 	    prevcb = &cb->ascb_next, cb = *prevcb) {
300 		if (((cb->ascb_events & AS_CALLBACK_CALLED) == 0) ||
301 		    (cb_arg != cb->ascb_arg)) {
302 			continue;
303 		}
304 		*prevcb = cb->ascb_next;
305 		kmem_free(cb, sizeof (struct as_callback));
306 		break;
307 	}
308 }
309 
310 /*
311  * Check the callback list for a matching event and intersection of
312  * address range. If there is a match invoke the callback.  Skip an entry if:
313  *    - a callback is already in progress for this entry (AS_CALLBACK_CALLED)
314  *    - not event of interest
315  *    - not address range of interest
316  *
317  * An event_len of zero indicates a request for an unconditional callback
318  * (regardless of event), only the AS_CALLBACK_CALLED is checked.  The
319  * a_contents lock must be dropped before a callback, so only one callback
320  * can be done before returning. Return -1 (true) if a callback was
321  * executed and removed from the list, else return 0 (false).
322  *
323  * The logically separate parts, i.e. finding a matching callback and
324  * executing a given callback have been separated into two functions
325  * so that they can be called with different sets of locks held beyond
326  * the always-required a_contents. as_find_callback does not sleep so
327  * it is ok to call it if more locks than a_contents (i.e. the a_lock
328  * rwlock) are held. as_execute_callback on the other hand may sleep
329  * so all locks beyond a_contents must be dropped by the caller if one
330  * does not want to end comatose.
331  */
332 static int
as_do_callbacks(struct as * as,uint_t events,caddr_t event_addr,size_t event_len)333 as_do_callbacks(struct as *as, uint_t events, caddr_t event_addr,
334     size_t event_len)
335 {
336 	struct as_callback *cb;
337 
338 	if ((cb = as_find_callback(as, events, event_addr, event_len))) {
339 		as_execute_callback(as, cb, events);
340 		return (-1);
341 	}
342 	return (0);
343 }
344 
345 /*
346  * Search for the segment containing addr. If a segment containing addr
347  * exists, that segment is returned.  If no such segment exists, and
348  * the list spans addresses greater than addr, then the first segment
349  * whose base is greater than addr is returned; otherwise, NULL is
350  * returned unless tail is true, in which case the last element of the
351  * list is returned.
352  *
353  * a_seglast is used to cache the last found segment for repeated
354  * searches to the same addr (which happens frequently).
355  */
356 struct seg *
as_findseg(struct as * as,caddr_t addr,int tail)357 as_findseg(struct as *as, caddr_t addr, int tail)
358 {
359 	struct seg *seg = as->a_seglast;
360 	avl_index_t where;
361 
362 	ASSERT(AS_LOCK_HELD(as));
363 
364 	if (seg != NULL &&
365 	    seg->s_base <= addr &&
366 	    addr < seg->s_base + seg->s_size)
367 		return (seg);
368 
369 	seg = avl_find(&as->a_segtree, &addr, &where);
370 	if (seg != NULL)
371 		return (as->a_seglast = seg);
372 
373 	seg = avl_nearest(&as->a_segtree, where, AVL_AFTER);
374 	if (seg == NULL && tail)
375 		seg = avl_last(&as->a_segtree);
376 	return (as->a_seglast = seg);
377 }
378 
379 #ifdef VERIFY_SEGLIST
380 /*
381  * verify that the linked list is coherent
382  */
383 static void
as_verify(struct as * as)384 as_verify(struct as *as)
385 {
386 	struct seg *seg, *seglast, *p, *n;
387 	uint_t nsegs = 0;
388 
389 	if (do_as_verify == 0)
390 		return;
391 
392 	seglast = as->a_seglast;
393 
394 	for (seg = AS_SEGFIRST(as); seg != NULL; seg = AS_SEGNEXT(as, seg)) {
395 		ASSERT(seg->s_as == as);
396 		p = AS_SEGPREV(as, seg);
397 		n = AS_SEGNEXT(as, seg);
398 		ASSERT(p == NULL || p->s_as == as);
399 		ASSERT(p == NULL || p->s_base < seg->s_base);
400 		ASSERT(n == NULL || n->s_base > seg->s_base);
401 		ASSERT(n != NULL || seg == avl_last(&as->a_segtree));
402 		if (seg == seglast)
403 			seglast = NULL;
404 		nsegs++;
405 	}
406 	ASSERT(seglast == NULL);
407 	ASSERT(avl_numnodes(&as->a_segtree) == nsegs);
408 }
409 #endif /* VERIFY_SEGLIST */
410 
411 /*
412  * Add a new segment to the address space. The avl_find()
413  * may be expensive so we attempt to use last segment accessed
414  * in as_gap() as an insertion point.
415  */
416 int
as_addseg(struct as * as,struct seg * newseg)417 as_addseg(struct as  *as, struct seg *newseg)
418 {
419 	struct seg *seg;
420 	caddr_t addr;
421 	caddr_t eaddr;
422 	avl_index_t where;
423 
424 	ASSERT(AS_WRITE_HELD(as));
425 
426 	as->a_updatedir = 1;	/* inform /proc */
427 	gethrestime(&as->a_updatetime);
428 
429 	if (as->a_lastgaphl != NULL) {
430 		struct seg *hseg = NULL;
431 		struct seg *lseg = NULL;
432 
433 		if (as->a_lastgaphl->s_base > newseg->s_base) {
434 			hseg = as->a_lastgaphl;
435 			lseg = AVL_PREV(&as->a_segtree, hseg);
436 		} else {
437 			lseg = as->a_lastgaphl;
438 			hseg = AVL_NEXT(&as->a_segtree, lseg);
439 		}
440 
441 		if (hseg && lseg && lseg->s_base < newseg->s_base &&
442 		    hseg->s_base > newseg->s_base) {
443 			avl_insert_here(&as->a_segtree, newseg, lseg,
444 			    AVL_AFTER);
445 			as->a_lastgaphl = NULL;
446 			as->a_seglast = newseg;
447 			return (0);
448 		}
449 		as->a_lastgaphl = NULL;
450 	}
451 
452 	addr = newseg->s_base;
453 	eaddr = addr + newseg->s_size;
454 
455 	seg = avl_find(&as->a_segtree, &addr, &where);
456 
457 	if (seg == NULL)
458 		seg = avl_nearest(&as->a_segtree, where, AVL_AFTER);
459 
460 	if (seg == NULL)
461 		seg = avl_last(&as->a_segtree);
462 
463 	if (seg != NULL) {
464 		caddr_t base = seg->s_base;
465 
466 		/*
467 		 * If top of seg is below the requested address, then
468 		 * the insertion point is at the end of the linked list,
469 		 * and seg points to the tail of the list.  Otherwise,
470 		 * the insertion point is immediately before seg.
471 		 */
472 		if (base + seg->s_size > addr) {
473 			if (addr >= base || eaddr > base) {
474 				return (-1);	/* overlapping segment */
475 			}
476 		}
477 	}
478 	as->a_seglast = newseg;
479 	avl_insert(&as->a_segtree, newseg, where);
480 
481 #ifdef VERIFY_SEGLIST
482 	as_verify(as);
483 #endif
484 	return (0);
485 }
486 
487 struct seg *
as_removeseg(struct as * as,struct seg * seg)488 as_removeseg(struct as *as, struct seg *seg)
489 {
490 	avl_tree_t *t;
491 
492 	ASSERT(AS_WRITE_HELD(as));
493 
494 	as->a_updatedir = 1;	/* inform /proc */
495 	gethrestime(&as->a_updatetime);
496 
497 	if (seg == NULL)
498 		return (NULL);
499 
500 	t = &as->a_segtree;
501 	if (as->a_seglast == seg)
502 		as->a_seglast = NULL;
503 	as->a_lastgaphl = NULL;
504 
505 	/*
506 	 * if this segment is at an address higher than
507 	 * a_lastgap, set a_lastgap to the next segment (NULL if last segment)
508 	 */
509 	if (as->a_lastgap &&
510 	    (seg == as->a_lastgap || seg->s_base > as->a_lastgap->s_base))
511 		as->a_lastgap = AVL_NEXT(t, seg);
512 
513 	/*
514 	 * remove the segment from the seg tree
515 	 */
516 	avl_remove(t, seg);
517 
518 #ifdef VERIFY_SEGLIST
519 	as_verify(as);
520 #endif
521 	return (seg);
522 }
523 
524 /*
525  * Find a segment containing addr.
526  */
527 struct seg *
as_segat(struct as * as,caddr_t addr)528 as_segat(struct as *as, caddr_t addr)
529 {
530 	struct seg *seg = as->a_seglast;
531 
532 	ASSERT(AS_LOCK_HELD(as));
533 
534 	if (seg != NULL && seg->s_base <= addr &&
535 	    addr < seg->s_base + seg->s_size)
536 		return (seg);
537 
538 	seg = avl_find(&as->a_segtree, &addr, NULL);
539 	return (seg);
540 }
541 
542 /*
543  * Serialize all searches for holes in an address space to
544  * prevent two or more threads from allocating the same virtual
545  * address range.  The address space must not be "read/write"
546  * locked by the caller since we may block.
547  */
548 void
as_rangelock(struct as * as)549 as_rangelock(struct as *as)
550 {
551 	mutex_enter(&as->a_contents);
552 	while (AS_ISCLAIMGAP(as))
553 		cv_wait(&as->a_cv, &as->a_contents);
554 	AS_SETCLAIMGAP(as);
555 	mutex_exit(&as->a_contents);
556 }
557 
558 /*
559  * Release hold on a_state & AS_CLAIMGAP and signal any other blocked threads.
560  */
561 void
as_rangeunlock(struct as * as)562 as_rangeunlock(struct as *as)
563 {
564 	mutex_enter(&as->a_contents);
565 	AS_CLRCLAIMGAP(as);
566 	cv_signal(&as->a_cv);
567 	mutex_exit(&as->a_contents);
568 }
569 
570 /*
571  * compar segments (or just an address) by segment address range
572  */
573 static int
as_segcompar(const void * x,const void * y)574 as_segcompar(const void *x, const void *y)
575 {
576 	struct seg *a = (struct seg *)x;
577 	struct seg *b = (struct seg *)y;
578 
579 	if (a->s_base < b->s_base)
580 		return (-1);
581 	if (a->s_base >= b->s_base + b->s_size)
582 		return (1);
583 	return (0);
584 }
585 
586 
587 void
as_avlinit(struct as * as)588 as_avlinit(struct as *as)
589 {
590 	avl_create(&as->a_segtree, as_segcompar, sizeof (struct seg),
591 	    offsetof(struct seg, s_tree));
592 	avl_create(&as->a_wpage, wp_compare, sizeof (struct watched_page),
593 	    offsetof(struct watched_page, wp_link));
594 }
595 
596 /*ARGSUSED*/
597 static int
as_constructor(void * buf,void * cdrarg,int kmflags)598 as_constructor(void *buf, void *cdrarg, int kmflags)
599 {
600 	struct as *as = buf;
601 
602 	mutex_init(&as->a_contents, NULL, MUTEX_DEFAULT, NULL);
603 	cv_init(&as->a_cv, NULL, CV_DEFAULT, NULL);
604 	rw_init(&as->a_lock, NULL, RW_DEFAULT, NULL);
605 	as_avlinit(as);
606 	return (0);
607 }
608 
609 /*ARGSUSED1*/
610 static void
as_destructor(void * buf,void * cdrarg)611 as_destructor(void *buf, void *cdrarg)
612 {
613 	struct as *as = buf;
614 
615 	avl_destroy(&as->a_segtree);
616 	mutex_destroy(&as->a_contents);
617 	cv_destroy(&as->a_cv);
618 	rw_destroy(&as->a_lock);
619 }
620 
621 void
as_init(void)622 as_init(void)
623 {
624 	as_cache = kmem_cache_create("as_cache", sizeof (struct as), 0,
625 	    as_constructor, as_destructor, NULL, NULL, NULL, 0);
626 }
627 
628 /*
629  * Allocate and initialize an address space data structure.
630  * We call hat_alloc to allow any machine dependent
631  * information in the hat structure to be initialized.
632  */
633 struct as *
as_alloc(void)634 as_alloc(void)
635 {
636 	struct as *as;
637 
638 	as = kmem_cache_alloc(as_cache, KM_SLEEP);
639 
640 	as->a_flags		= 0;
641 	as->a_vbits		= 0;
642 	as->a_hrm		= NULL;
643 	as->a_seglast		= NULL;
644 	as->a_size		= 0;
645 	as->a_resvsize		= 0;
646 	as->a_updatedir		= 0;
647 	gethrestime(&as->a_updatetime);
648 	as->a_objectdir		= NULL;
649 	as->a_sizedir		= 0;
650 	as->a_userlimit		= (caddr_t)USERLIMIT;
651 	as->a_lastgap		= NULL;
652 	as->a_lastgaphl		= NULL;
653 	as->a_callbacks		= NULL;
654 	as->a_proc		= NULL;
655 
656 	AS_LOCK_ENTER(as, RW_WRITER);
657 	as->a_hat = hat_alloc(as);	/* create hat for default system mmu */
658 	AS_LOCK_EXIT(as);
659 
660 	return (as);
661 }
662 
663 /*
664  * Free an address space data structure.
665  * Need to free the hat first and then
666  * all the segments on this as and finally
667  * the space for the as struct itself.
668  */
669 void
as_free(struct as * as)670 as_free(struct as *as)
671 {
672 	struct hat *hat = as->a_hat;
673 	struct seg *seg, *next;
674 	boolean_t free_started = B_FALSE;
675 
676 top:
677 	/*
678 	 * Invoke ALL callbacks. as_do_callbacks will do one callback
679 	 * per call, and not return (-1) until the callback has completed.
680 	 * When as_do_callbacks returns zero, all callbacks have completed.
681 	 */
682 	mutex_enter(&as->a_contents);
683 	while (as->a_callbacks && as_do_callbacks(as, AS_ALL_EVENT, 0, 0))
684 		;
685 
686 	mutex_exit(&as->a_contents);
687 	AS_LOCK_ENTER(as, RW_WRITER);
688 
689 	if (!free_started) {
690 		free_started = B_TRUE;
691 		hat_free_start(hat);
692 	}
693 	for (seg = AS_SEGFIRST(as); seg != NULL; seg = next) {
694 		int err;
695 
696 		next = AS_SEGNEXT(as, seg);
697 retry:
698 		err = SEGOP_UNMAP(seg, seg->s_base, seg->s_size);
699 		if (err == EAGAIN) {
700 			mutex_enter(&as->a_contents);
701 			if (as->a_callbacks) {
702 				AS_LOCK_EXIT(as);
703 			} else if (!AS_ISNOUNMAPWAIT(as)) {
704 				/*
705 				 * Memory is currently locked. Wait for a
706 				 * cv_signal that it has been unlocked, then
707 				 * try the operation again.
708 				 */
709 				if (AS_ISUNMAPWAIT(as) == 0)
710 					cv_broadcast(&as->a_cv);
711 				AS_SETUNMAPWAIT(as);
712 				AS_LOCK_EXIT(as);
713 				while (AS_ISUNMAPWAIT(as))
714 					cv_wait(&as->a_cv, &as->a_contents);
715 			} else {
716 				/*
717 				 * We may have raced with
718 				 * segvn_reclaim()/segspt_reclaim(). In this
719 				 * case clean nounmapwait flag and retry since
720 				 * softlockcnt in this segment may be already
721 				 * 0.  We don't drop as writer lock so our
722 				 * number of retries without sleeping should
723 				 * be very small. See segvn_reclaim() for
724 				 * more comments.
725 				 */
726 				AS_CLRNOUNMAPWAIT(as);
727 				mutex_exit(&as->a_contents);
728 				goto retry;
729 			}
730 			mutex_exit(&as->a_contents);
731 			goto top;
732 		} else {
733 			/*
734 			 * We do not expect any other error return at this
735 			 * time. This is similar to an ASSERT in seg_unmap()
736 			 */
737 			ASSERT(err == 0);
738 		}
739 	}
740 	hat_free_end(hat);
741 	AS_LOCK_EXIT(as);
742 
743 	/* /proc stuff */
744 	ASSERT(avl_numnodes(&as->a_wpage) == 0);
745 	if (as->a_objectdir) {
746 		kmem_free(as->a_objectdir, as->a_sizedir * sizeof (vnode_t *));
747 		as->a_objectdir = NULL;
748 		as->a_sizedir = 0;
749 	}
750 
751 	/*
752 	 * Free the struct as back to kmem.  Assert it has no segments.
753 	 */
754 	ASSERT(avl_numnodes(&as->a_segtree) == 0);
755 	kmem_cache_free(as_cache, as);
756 }
757 
758 int
as_dup(struct as * as,struct proc * forkedproc)759 as_dup(struct as *as, struct proc *forkedproc)
760 {
761 	struct as *newas;
762 	struct seg *seg, *newseg;
763 	size_t	purgesize = 0;
764 	int error;
765 
766 	AS_LOCK_ENTER(as, RW_WRITER);
767 	as_clearwatch(as);
768 	newas = as_alloc();
769 	newas->a_userlimit = as->a_userlimit;
770 	newas->a_proc = forkedproc;
771 
772 	AS_LOCK_ENTER(newas, RW_WRITER);
773 
774 	(void) hat_dup(as->a_hat, newas->a_hat, NULL, 0, HAT_DUP_SRD);
775 
776 	for (seg = AS_SEGFIRST(as); seg != NULL; seg = AS_SEGNEXT(as, seg)) {
777 
778 		if (seg->s_flags & S_PURGE) {
779 			purgesize += seg->s_size;
780 			continue;
781 		}
782 
783 		newseg = seg_alloc(newas, seg->s_base, seg->s_size);
784 		if (newseg == NULL) {
785 			AS_LOCK_EXIT(newas);
786 			as_setwatch(as);
787 			AS_LOCK_EXIT(as);
788 			as_free(newas);
789 			return (-1);
790 		}
791 		if ((error = SEGOP_DUP(seg, newseg)) != 0) {
792 			/*
793 			 * We call seg_free() on the new seg
794 			 * because the segment is not set up
795 			 * completely; i.e. it has no ops.
796 			 */
797 			as_setwatch(as);
798 			AS_LOCK_EXIT(as);
799 			seg_free(newseg);
800 			AS_LOCK_EXIT(newas);
801 			as_free(newas);
802 			return (error);
803 		}
804 		if ((newseg->s_flags & S_HOLE) == 0) {
805 			newas->a_size += seg->s_size;
806 		}
807 	}
808 	newas->a_resvsize = as->a_resvsize - purgesize;
809 
810 	error = hat_dup(as->a_hat, newas->a_hat, NULL, 0, HAT_DUP_ALL);
811 
812 	AS_LOCK_EXIT(newas);
813 
814 	as_setwatch(as);
815 	AS_LOCK_EXIT(as);
816 	if (error != 0) {
817 		as_free(newas);
818 		return (error);
819 	}
820 	forkedproc->p_as = newas;
821 	return (0);
822 }
823 
824 /*
825  * Handle a ``fault'' at addr for size bytes.
826  */
827 faultcode_t
as_fault(struct hat * hat,struct as * as,caddr_t addr,size_t size,enum fault_type type,enum seg_rw rw)828 as_fault(struct hat *hat, struct as *as, caddr_t addr, size_t size,
829     enum fault_type type, enum seg_rw rw)
830 {
831 	struct seg *seg;
832 	caddr_t raddr;			/* rounded down addr */
833 	size_t rsize;			/* rounded up size */
834 	size_t ssize;
835 	faultcode_t res = 0;
836 	caddr_t addrsav;
837 	struct seg *segsav;
838 	int as_lock_held;
839 	klwp_t *lwp = ttolwp(curthread);
840 
841 
842 
843 retry:
844 	/*
845 	 * Indicate that the lwp is not to be stopped while waiting for a
846 	 * pagefault.  This is to avoid deadlock while debugging a process
847 	 * via /proc over NFS (in particular).
848 	 */
849 	if (lwp != NULL)
850 		lwp->lwp_nostop++;
851 
852 	/*
853 	 * same length must be used when we softlock and softunlock.  We
854 	 * don't support softunlocking lengths less than the original length
855 	 * when there is largepage support.  See seg_dev.c for more
856 	 * comments.
857 	 */
858 	switch (type) {
859 
860 	case F_SOFTLOCK:
861 		CPU_STATS_ADD_K(vm, softlock, 1);
862 		break;
863 
864 	case F_SOFTUNLOCK:
865 		break;
866 
867 	case F_PROT:
868 		CPU_STATS_ADD_K(vm, prot_fault, 1);
869 		break;
870 
871 	case F_INVAL:
872 		CPU_STATS_ENTER_K();
873 		CPU_STATS_ADDQ(CPU, vm, as_fault, 1);
874 		if (as == &kas)
875 			CPU_STATS_ADDQ(CPU, vm, kernel_asflt, 1);
876 		CPU_STATS_EXIT_K();
877 		break;
878 	}
879 
880 	raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
881 	rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) -
882 	    (size_t)raddr;
883 
884 	/*
885 	 * XXX -- Don't grab the as lock for segkmap. We should grab it for
886 	 * correctness, but then we could be stuck holding this lock for
887 	 * a LONG time if the fault needs to be resolved on a slow
888 	 * filesystem, and then no-one will be able to exec new commands,
889 	 * as exec'ing requires the write lock on the as.
890 	 */
891 	if (as == &kas && segkmap && segkmap->s_base <= raddr &&
892 	    raddr + size < segkmap->s_base + segkmap->s_size) {
893 		seg = segkmap;
894 		as_lock_held = 0;
895 	} else {
896 		AS_LOCK_ENTER(as, RW_READER);
897 
898 		seg = as_segat(as, raddr);
899 		if (seg == NULL) {
900 			AS_LOCK_EXIT(as);
901 			if (lwp != NULL)
902 				lwp->lwp_nostop--;
903 			return (FC_NOMAP);
904 		}
905 
906 		as_lock_held = 1;
907 	}
908 
909 	addrsav = raddr;
910 	segsav = seg;
911 
912 	for (; rsize != 0; rsize -= ssize, raddr += ssize) {
913 		if (raddr >= seg->s_base + seg->s_size) {
914 			seg = AS_SEGNEXT(as, seg);
915 			if (seg == NULL || raddr != seg->s_base) {
916 				res = FC_NOMAP;
917 				break;
918 			}
919 		}
920 		if (raddr + rsize > seg->s_base + seg->s_size)
921 			ssize = seg->s_base + seg->s_size - raddr;
922 		else
923 			ssize = rsize;
924 
925 		res = SEGOP_FAULT(hat, seg, raddr, ssize, type, rw);
926 		if (res != 0)
927 			break;
928 	}
929 
930 	/*
931 	 * If we were SOFTLOCKing and encountered a failure,
932 	 * we must SOFTUNLOCK the range we already did. (Maybe we
933 	 * should just panic if we are SOFTLOCKing or even SOFTUNLOCKing
934 	 * right here...)
935 	 */
936 	if (res != 0 && type == F_SOFTLOCK) {
937 		for (seg = segsav; addrsav < raddr; addrsav += ssize) {
938 			if (addrsav >= seg->s_base + seg->s_size)
939 				seg = AS_SEGNEXT(as, seg);
940 			ASSERT(seg != NULL);
941 			/*
942 			 * Now call the fault routine again to perform the
943 			 * unlock using S_OTHER instead of the rw variable
944 			 * since we never got a chance to touch the pages.
945 			 */
946 			if (raddr > seg->s_base + seg->s_size)
947 				ssize = seg->s_base + seg->s_size - addrsav;
948 			else
949 				ssize = raddr - addrsav;
950 			(void) SEGOP_FAULT(hat, seg, addrsav, ssize,
951 			    F_SOFTUNLOCK, S_OTHER);
952 		}
953 	}
954 	if (as_lock_held)
955 		AS_LOCK_EXIT(as);
956 	if (lwp != NULL)
957 		lwp->lwp_nostop--;
958 
959 	/*
960 	 * If the lower levels returned EDEADLK for a fault,
961 	 * It means that we should retry the fault.  Let's wait
962 	 * a bit also to let the deadlock causing condition clear.
963 	 * This is part of a gross hack to work around a design flaw
964 	 * in the ufs/sds logging code and should go away when the
965 	 * logging code is re-designed to fix the problem. See bug
966 	 * 4125102 for details of the problem.
967 	 */
968 	if (FC_ERRNO(res) == EDEADLK) {
969 		delay(deadlk_wait);
970 		res = 0;
971 		goto retry;
972 	}
973 	return (res);
974 }
975 
976 
977 
978 /*
979  * Asynchronous ``fault'' at addr for size bytes.
980  */
981 faultcode_t
as_faulta(struct as * as,caddr_t addr,size_t size)982 as_faulta(struct as *as, caddr_t addr, size_t size)
983 {
984 	struct seg *seg;
985 	caddr_t raddr;			/* rounded down addr */
986 	size_t rsize;			/* rounded up size */
987 	faultcode_t res = 0;
988 	klwp_t *lwp = ttolwp(curthread);
989 
990 retry:
991 	/*
992 	 * Indicate that the lwp is not to be stopped while waiting
993 	 * for a pagefault.  This is to avoid deadlock while debugging
994 	 * a process via /proc over NFS (in particular).
995 	 */
996 	if (lwp != NULL)
997 		lwp->lwp_nostop++;
998 
999 	raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
1000 	rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) -
1001 	    (size_t)raddr;
1002 
1003 	AS_LOCK_ENTER(as, RW_READER);
1004 	seg = as_segat(as, raddr);
1005 	if (seg == NULL) {
1006 		AS_LOCK_EXIT(as);
1007 		if (lwp != NULL)
1008 			lwp->lwp_nostop--;
1009 		return (FC_NOMAP);
1010 	}
1011 
1012 	for (; rsize != 0; rsize -= PAGESIZE, raddr += PAGESIZE) {
1013 		if (raddr >= seg->s_base + seg->s_size) {
1014 			seg = AS_SEGNEXT(as, seg);
1015 			if (seg == NULL || raddr != seg->s_base) {
1016 				res = FC_NOMAP;
1017 				break;
1018 			}
1019 		}
1020 		res = SEGOP_FAULTA(seg, raddr);
1021 		if (res != 0)
1022 			break;
1023 	}
1024 	AS_LOCK_EXIT(as);
1025 	if (lwp != NULL)
1026 		lwp->lwp_nostop--;
1027 	/*
1028 	 * If the lower levels returned EDEADLK for a fault,
1029 	 * It means that we should retry the fault.  Let's wait
1030 	 * a bit also to let the deadlock causing condition clear.
1031 	 * This is part of a gross hack to work around a design flaw
1032 	 * in the ufs/sds logging code and should go away when the
1033 	 * logging code is re-designed to fix the problem. See bug
1034 	 * 4125102 for details of the problem.
1035 	 */
1036 	if (FC_ERRNO(res) == EDEADLK) {
1037 		delay(deadlk_wait);
1038 		res = 0;
1039 		goto retry;
1040 	}
1041 	return (res);
1042 }
1043 
1044 /*
1045  * Set the virtual mapping for the interval from [addr : addr + size)
1046  * in address space `as' to have the specified protection.
1047  * It is ok for the range to cross over several segments,
1048  * as long as they are contiguous.
1049  */
1050 int
as_setprot(struct as * as,caddr_t addr,size_t size,uint_t prot)1051 as_setprot(struct as *as, caddr_t addr, size_t size, uint_t prot)
1052 {
1053 	struct seg *seg;
1054 	struct as_callback *cb;
1055 	size_t ssize;
1056 	caddr_t raddr;			/* rounded down addr */
1057 	size_t rsize;			/* rounded up size */
1058 	int error = 0, writer = 0;
1059 	caddr_t saveraddr;
1060 	size_t saversize;
1061 
1062 setprot_top:
1063 	raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
1064 	rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) -
1065 	    (size_t)raddr;
1066 
1067 	if (raddr + rsize < raddr)		/* check for wraparound */
1068 		return (ENOMEM);
1069 
1070 	saveraddr = raddr;
1071 	saversize = rsize;
1072 
1073 	/*
1074 	 * Normally we only lock the as as a reader. But
1075 	 * if due to setprot the segment driver needs to split
1076 	 * a segment it will return IE_RETRY. Therefore we re-acquire
1077 	 * the as lock as a writer so the segment driver can change
1078 	 * the seg list. Also the segment driver will return IE_RETRY
1079 	 * after it has changed the segment list so we therefore keep
1080 	 * locking as a writer. Since these opeartions should be rare
1081 	 * want to only lock as a writer when necessary.
1082 	 */
1083 	if (writer || avl_numnodes(&as->a_wpage) != 0) {
1084 		AS_LOCK_ENTER(as, RW_WRITER);
1085 	} else {
1086 		AS_LOCK_ENTER(as, RW_READER);
1087 	}
1088 
1089 	as_clearwatchprot(as, raddr, rsize);
1090 	seg = as_segat(as, raddr);
1091 	if (seg == NULL) {
1092 		as_setwatch(as);
1093 		AS_LOCK_EXIT(as);
1094 		return (ENOMEM);
1095 	}
1096 
1097 	for (; rsize != 0; rsize -= ssize, raddr += ssize) {
1098 		if (raddr >= seg->s_base + seg->s_size) {
1099 			seg = AS_SEGNEXT(as, seg);
1100 			if (seg == NULL || raddr != seg->s_base) {
1101 				error = ENOMEM;
1102 				break;
1103 			}
1104 		}
1105 		if ((raddr + rsize) > (seg->s_base + seg->s_size))
1106 			ssize = seg->s_base + seg->s_size - raddr;
1107 		else
1108 			ssize = rsize;
1109 retry:
1110 		error = SEGOP_SETPROT(seg, raddr, ssize, prot);
1111 
1112 		if (error == IE_NOMEM) {
1113 			error = EAGAIN;
1114 			break;
1115 		}
1116 
1117 		if (error == IE_RETRY) {
1118 			AS_LOCK_EXIT(as);
1119 			writer = 1;
1120 			goto setprot_top;
1121 		}
1122 
1123 		if (error == EAGAIN) {
1124 			/*
1125 			 * Make sure we have a_lock as writer.
1126 			 */
1127 			if (writer == 0) {
1128 				AS_LOCK_EXIT(as);
1129 				writer = 1;
1130 				goto setprot_top;
1131 			}
1132 
1133 			/*
1134 			 * Memory is currently locked.  It must be unlocked
1135 			 * before this operation can succeed through a retry.
1136 			 * The possible reasons for locked memory and
1137 			 * corresponding strategies for unlocking are:
1138 			 * (1) Normal I/O
1139 			 *	wait for a signal that the I/O operation
1140 			 *	has completed and the memory is unlocked.
1141 			 * (2) Asynchronous I/O
1142 			 *	The aio subsystem does not unlock pages when
1143 			 *	the I/O is completed. Those pages are unlocked
1144 			 *	when the application calls aiowait/aioerror.
1145 			 *	So, to prevent blocking forever, cv_broadcast()
1146 			 *	is done to wake up aio_cleanup_thread.
1147 			 *	Subsequently, segvn_reclaim will be called, and
1148 			 *	that will do AS_CLRUNMAPWAIT() and wake us up.
1149 			 * (3) Long term page locking:
1150 			 *	Drivers intending to have pages locked for a
1151 			 *	period considerably longer than for normal I/O
1152 			 *	(essentially forever) may have registered for a
1153 			 *	callback so they may unlock these pages on
1154 			 *	request. This is needed to allow this operation
1155 			 *	to succeed. Each entry on the callback list is
1156 			 *	examined. If the event or address range pertains
1157 			 *	the callback is invoked (unless it already is in
1158 			 *	progress). The a_contents lock must be dropped
1159 			 *	before the callback, so only one callback can
1160 			 *	be done at a time. Go to the top and do more
1161 			 *	until zero is returned. If zero is returned,
1162 			 *	either there were no callbacks for this event
1163 			 *	or they were already in progress.
1164 			 */
1165 			mutex_enter(&as->a_contents);
1166 			if (as->a_callbacks &&
1167 			    (cb = as_find_callback(as, AS_SETPROT_EVENT,
1168 			    seg->s_base, seg->s_size))) {
1169 				AS_LOCK_EXIT(as);
1170 				as_execute_callback(as, cb, AS_SETPROT_EVENT);
1171 			} else if (!AS_ISNOUNMAPWAIT(as)) {
1172 				if (AS_ISUNMAPWAIT(as) == 0)
1173 					cv_broadcast(&as->a_cv);
1174 				AS_SETUNMAPWAIT(as);
1175 				AS_LOCK_EXIT(as);
1176 				while (AS_ISUNMAPWAIT(as))
1177 					cv_wait(&as->a_cv, &as->a_contents);
1178 			} else {
1179 				/*
1180 				 * We may have raced with
1181 				 * segvn_reclaim()/segspt_reclaim(). In this
1182 				 * case clean nounmapwait flag and retry since
1183 				 * softlockcnt in this segment may be already
1184 				 * 0.  We don't drop as writer lock so our
1185 				 * number of retries without sleeping should
1186 				 * be very small. See segvn_reclaim() for
1187 				 * more comments.
1188 				 */
1189 				AS_CLRNOUNMAPWAIT(as);
1190 				mutex_exit(&as->a_contents);
1191 				goto retry;
1192 			}
1193 			mutex_exit(&as->a_contents);
1194 			goto setprot_top;
1195 		} else if (error != 0)
1196 			break;
1197 	}
1198 	if (error != 0) {
1199 		as_setwatch(as);
1200 	} else {
1201 		as_setwatchprot(as, saveraddr, saversize, prot);
1202 	}
1203 	AS_LOCK_EXIT(as);
1204 	return (error);
1205 }
1206 
1207 /*
1208  * Check to make sure that the interval [addr, addr + size)
1209  * in address space `as' has at least the specified protection.
1210  * It is ok for the range to cross over several segments, as long
1211  * as they are contiguous.
1212  */
1213 int
as_checkprot(struct as * as,caddr_t addr,size_t size,uint_t prot)1214 as_checkprot(struct as *as, caddr_t addr, size_t size, uint_t prot)
1215 {
1216 	struct seg *seg;
1217 	size_t ssize;
1218 	caddr_t raddr;			/* rounded down addr */
1219 	size_t rsize;			/* rounded up size */
1220 	int error = 0;
1221 
1222 	raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
1223 	rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) -
1224 	    (size_t)raddr;
1225 
1226 	if (raddr + rsize < raddr)		/* check for wraparound */
1227 		return (ENOMEM);
1228 
1229 	/*
1230 	 * This is ugly as sin...
1231 	 * Normally, we only acquire the address space readers lock.
1232 	 * However, if the address space has watchpoints present,
1233 	 * we must acquire the writer lock on the address space for
1234 	 * the benefit of as_clearwatchprot() and as_setwatchprot().
1235 	 */
1236 	if (avl_numnodes(&as->a_wpage) != 0)
1237 		AS_LOCK_ENTER(as, RW_WRITER);
1238 	else
1239 		AS_LOCK_ENTER(as, RW_READER);
1240 	as_clearwatchprot(as, raddr, rsize);
1241 	seg = as_segat(as, raddr);
1242 	if (seg == NULL) {
1243 		as_setwatch(as);
1244 		AS_LOCK_EXIT(as);
1245 		return (ENOMEM);
1246 	}
1247 
1248 	for (; rsize != 0; rsize -= ssize, raddr += ssize) {
1249 		if (raddr >= seg->s_base + seg->s_size) {
1250 			seg = AS_SEGNEXT(as, seg);
1251 			if (seg == NULL || raddr != seg->s_base) {
1252 				error = ENOMEM;
1253 				break;
1254 			}
1255 		}
1256 		if ((raddr + rsize) > (seg->s_base + seg->s_size))
1257 			ssize = seg->s_base + seg->s_size - raddr;
1258 		else
1259 			ssize = rsize;
1260 
1261 		error = SEGOP_CHECKPROT(seg, raddr, ssize, prot);
1262 		if (error != 0)
1263 			break;
1264 	}
1265 	as_setwatch(as);
1266 	AS_LOCK_EXIT(as);
1267 	return (error);
1268 }
1269 
1270 int
as_unmap(struct as * as,caddr_t addr,size_t size)1271 as_unmap(struct as *as, caddr_t addr, size_t size)
1272 {
1273 	struct seg *seg, *seg_next;
1274 	struct as_callback *cb;
1275 	caddr_t raddr, eaddr;
1276 	size_t ssize, rsize = 0;
1277 	int err;
1278 
1279 top:
1280 	raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
1281 	eaddr = (caddr_t)(((uintptr_t)(addr + size) + PAGEOFFSET) &
1282 	    (uintptr_t)PAGEMASK);
1283 
1284 	AS_LOCK_ENTER(as, RW_WRITER);
1285 
1286 	as->a_updatedir = 1;	/* inform /proc */
1287 	gethrestime(&as->a_updatetime);
1288 
1289 	/*
1290 	 * Use as_findseg to find the first segment in the range, then
1291 	 * step through the segments in order, following s_next.
1292 	 */
1293 	as_clearwatchprot(as, raddr, eaddr - raddr);
1294 
1295 	for (seg = as_findseg(as, raddr, 0); seg != NULL; seg = seg_next) {
1296 		const boolean_t is_hole = ((seg->s_flags & S_HOLE) != 0);
1297 
1298 		if (eaddr <= seg->s_base)
1299 			break;		/* eaddr was in a gap; all done */
1300 
1301 		/* this is implied by the test above */
1302 		ASSERT(raddr < eaddr);
1303 
1304 		if (raddr < seg->s_base)
1305 			raddr = seg->s_base; 	/* raddr was in a gap */
1306 
1307 		if (eaddr > (seg->s_base + seg->s_size))
1308 			ssize = seg->s_base + seg->s_size - raddr;
1309 		else
1310 			ssize = eaddr - raddr;
1311 
1312 		/*
1313 		 * Save next segment pointer since seg can be
1314 		 * destroyed during the segment unmap operation.
1315 		 */
1316 		seg_next = AS_SEGNEXT(as, seg);
1317 
1318 		/*
1319 		 * We didn't count /dev/null mappings, so ignore them here.
1320 		 * We'll handle MAP_NORESERVE cases in segvn_unmap(). (Again,
1321 		 * we have to do this check here while we have seg.)
1322 		 */
1323 		rsize = 0;
1324 		if (!SEG_IS_DEVNULL_MAPPING(seg) &&
1325 		    !SEG_IS_PARTIAL_RESV(seg))
1326 			rsize = ssize;
1327 
1328 retry:
1329 		err = SEGOP_UNMAP(seg, raddr, ssize);
1330 		if (err == EAGAIN) {
1331 			/*
1332 			 * Memory is currently locked.  It must be unlocked
1333 			 * before this operation can succeed through a retry.
1334 			 * The possible reasons for locked memory and
1335 			 * corresponding strategies for unlocking are:
1336 			 * (1) Normal I/O
1337 			 *	wait for a signal that the I/O operation
1338 			 *	has completed and the memory is unlocked.
1339 			 * (2) Asynchronous I/O
1340 			 *	The aio subsystem does not unlock pages when
1341 			 *	the I/O is completed. Those pages are unlocked
1342 			 *	when the application calls aiowait/aioerror.
1343 			 *	So, to prevent blocking forever, cv_broadcast()
1344 			 *	is done to wake up aio_cleanup_thread.
1345 			 *	Subsequently, segvn_reclaim will be called, and
1346 			 *	that will do AS_CLRUNMAPWAIT() and wake us up.
1347 			 * (3) Long term page locking:
1348 			 *	Drivers intending to have pages locked for a
1349 			 *	period considerably longer than for normal I/O
1350 			 *	(essentially forever) may have registered for a
1351 			 *	callback so they may unlock these pages on
1352 			 *	request. This is needed to allow this operation
1353 			 *	to succeed. Each entry on the callback list is
1354 			 *	examined. If the event or address range pertains
1355 			 *	the callback is invoked (unless it already is in
1356 			 *	progress). The a_contents lock must be dropped
1357 			 *	before the callback, so only one callback can
1358 			 *	be done at a time. Go to the top and do more
1359 			 *	until zero is returned. If zero is returned,
1360 			 *	either there were no callbacks for this event
1361 			 *	or they were already in progress.
1362 			 */
1363 			mutex_enter(&as->a_contents);
1364 			if (as->a_callbacks &&
1365 			    (cb = as_find_callback(as, AS_UNMAP_EVENT,
1366 			    seg->s_base, seg->s_size))) {
1367 				AS_LOCK_EXIT(as);
1368 				as_execute_callback(as, cb, AS_UNMAP_EVENT);
1369 			} else if (!AS_ISNOUNMAPWAIT(as)) {
1370 				if (AS_ISUNMAPWAIT(as) == 0)
1371 					cv_broadcast(&as->a_cv);
1372 				AS_SETUNMAPWAIT(as);
1373 				AS_LOCK_EXIT(as);
1374 				while (AS_ISUNMAPWAIT(as))
1375 					cv_wait(&as->a_cv, &as->a_contents);
1376 			} else {
1377 				/*
1378 				 * We may have raced with
1379 				 * segvn_reclaim()/segspt_reclaim(). In this
1380 				 * case clean nounmapwait flag and retry since
1381 				 * softlockcnt in this segment may be already
1382 				 * 0.  We don't drop as writer lock so our
1383 				 * number of retries without sleeping should
1384 				 * be very small. See segvn_reclaim() for
1385 				 * more comments.
1386 				 */
1387 				AS_CLRNOUNMAPWAIT(as);
1388 				mutex_exit(&as->a_contents);
1389 				goto retry;
1390 			}
1391 			mutex_exit(&as->a_contents);
1392 			goto top;
1393 		} else if (err == IE_RETRY) {
1394 			AS_LOCK_EXIT(as);
1395 			goto top;
1396 		} else if (err) {
1397 			as_setwatch(as);
1398 			AS_LOCK_EXIT(as);
1399 			return (-1);
1400 		}
1401 
1402 		if (!is_hole) {
1403 			as->a_size -= ssize;
1404 			if (rsize)
1405 				as->a_resvsize -= rsize;
1406 		}
1407 		raddr += ssize;
1408 	}
1409 	AS_LOCK_EXIT(as);
1410 	return (0);
1411 }
1412 
1413 static int
as_map_segvn_segs(struct as * as,caddr_t addr,size_t size,uint_t szcvec,segcreate_func_t crfp,struct segvn_crargs * vn_a,boolean_t * segcreated)1414 as_map_segvn_segs(struct as *as, caddr_t addr, size_t size, uint_t szcvec,
1415     segcreate_func_t crfp, struct segvn_crargs *vn_a, boolean_t *segcreated)
1416 {
1417 	uint_t szc, nszc, save_szcvec;
1418 	int error;
1419 	caddr_t a, eaddr;
1420 	size_t pgsz = 0;
1421 	const boolean_t do_off = (vn_a->vp != NULL || vn_a->amp != NULL);
1422 
1423 	ASSERT(AS_WRITE_HELD(as));
1424 	ASSERT(IS_P2ALIGNED(addr, PAGESIZE));
1425 	ASSERT(IS_P2ALIGNED(size, PAGESIZE));
1426 	ASSERT(vn_a->vp == NULL || vn_a->amp == NULL);
1427 
1428 	if (!do_off) {
1429 		vn_a->offset = 0;
1430 	}
1431 
1432 	if (szcvec <= 1) {
1433 		struct seg *seg, *segref;
1434 
1435 		seg = segref = seg_alloc(as, addr, size);
1436 		if (seg == NULL) {
1437 			return (ENOMEM);
1438 		}
1439 		vn_a->szc = 0;
1440 		error = (*crfp)(&seg, vn_a);
1441 		if (error != 0) {
1442 			VERIFY3P(seg, ==, segref);
1443 			seg_free(seg);
1444 		} else {
1445 			as->a_size += size;
1446 			as->a_resvsize += size;
1447 		}
1448 		return (error);
1449 	}
1450 
1451 	eaddr = addr + size;
1452 	save_szcvec = szcvec;
1453 	szcvec >>= 1;
1454 	szc = 0;
1455 	nszc = 0;
1456 	while (szcvec) {
1457 		if ((szcvec & 0x1) == 0) {
1458 			nszc++;
1459 			szcvec >>= 1;
1460 			continue;
1461 		}
1462 		nszc++;
1463 		pgsz = page_get_pagesize(nszc);
1464 		a = (caddr_t)P2ROUNDUP((uintptr_t)addr, pgsz);
1465 		if (a != addr) {
1466 			struct seg *seg, *segref;
1467 			size_t segsize;
1468 
1469 			ASSERT(a < eaddr);
1470 
1471 			segsize = a - addr;
1472 			seg = segref = seg_alloc(as, addr, segsize);
1473 			if (seg == NULL) {
1474 				return (ENOMEM);
1475 			}
1476 			vn_a->szc = szc;
1477 			error = (*crfp)(&seg, vn_a);
1478 			if (error != 0) {
1479 				VERIFY3P(seg, ==, segref);
1480 				seg_free(seg);
1481 				return (error);
1482 			}
1483 			as->a_size += segsize;
1484 			as->a_resvsize += segsize;
1485 			*segcreated = B_TRUE;
1486 			if (do_off) {
1487 				vn_a->offset += segsize;
1488 			}
1489 			addr = a;
1490 		}
1491 		szc = nszc;
1492 		szcvec >>= 1;
1493 	}
1494 
1495 	ASSERT(addr < eaddr);
1496 	szcvec = save_szcvec | 1; /* add 8K pages */
1497 	while (szcvec) {
1498 		a = (caddr_t)P2ALIGN((uintptr_t)eaddr, pgsz);
1499 		ASSERT(a >= addr);
1500 		if (a != addr) {
1501 			struct seg *seg, *segref;
1502 			size_t segsize;
1503 
1504 			segsize = a - addr;
1505 			seg = segref = seg_alloc(as, addr, segsize);
1506 			if (seg == NULL) {
1507 				return (ENOMEM);
1508 			}
1509 			vn_a->szc = szc;
1510 			error = (*crfp)(&seg, vn_a);
1511 			if (error != 0) {
1512 				VERIFY3P(seg, ==, segref);
1513 				seg_free(seg);
1514 				return (error);
1515 			}
1516 			as->a_size += segsize;
1517 			as->a_resvsize += segsize;
1518 			*segcreated = B_TRUE;
1519 			if (do_off) {
1520 				vn_a->offset += segsize;
1521 			}
1522 			addr = a;
1523 		}
1524 		szcvec &= ~(1 << szc);
1525 		if (szcvec) {
1526 			szc = highbit(szcvec) - 1;
1527 			pgsz = page_get_pagesize(szc);
1528 		}
1529 	}
1530 	ASSERT(addr == eaddr);
1531 
1532 	return (0);
1533 }
1534 
1535 static int
as_map_vnsegs(struct as * as,caddr_t addr,size_t size,segcreate_func_t crfp,struct segvn_crargs * vn_a,boolean_t * segcreated)1536 as_map_vnsegs(struct as *as, caddr_t addr, size_t size,
1537     segcreate_func_t crfp, struct segvn_crargs *vn_a, boolean_t *segcreated)
1538 {
1539 	uint_t mapflags = vn_a->flags & (MAP_TEXT | MAP_INITDATA);
1540 	int type = (vn_a->type == MAP_SHARED) ? MAPPGSZC_SHM : MAPPGSZC_PRIVM;
1541 	uint_t szcvec = map_pgszcvec(addr, size, (uintptr_t)addr, mapflags,
1542 	    type, 0);
1543 	int error;
1544 	struct vattr va;
1545 	u_offset_t eoff;
1546 	size_t save_size = 0;
1547 	extern size_t textrepl_size_thresh;
1548 
1549 	ASSERT(AS_WRITE_HELD(as));
1550 	ASSERT(IS_P2ALIGNED(addr, PAGESIZE));
1551 	ASSERT(IS_P2ALIGNED(size, PAGESIZE));
1552 	ASSERT(vn_a->vp != NULL);
1553 	ASSERT(vn_a->amp == NULL);
1554 
1555 again:
1556 	if (szcvec <= 1) {
1557 		struct seg *seg, *segref;
1558 
1559 		seg = segref = seg_alloc(as, addr, size);
1560 		if (seg == NULL) {
1561 			return (ENOMEM);
1562 		}
1563 		vn_a->szc = 0;
1564 		error = (*crfp)(&seg, vn_a);
1565 		if (error != 0) {
1566 			VERIFY3P(seg, ==, segref);
1567 			seg_free(seg);
1568 		} else {
1569 			as->a_size += size;
1570 			as->a_resvsize += size;
1571 		}
1572 		return (error);
1573 	}
1574 
1575 	va.va_mask = AT_SIZE;
1576 	if (VOP_GETATTR(vn_a->vp, &va, ATTR_HINT, vn_a->cred, NULL) != 0) {
1577 		szcvec = 0;
1578 		goto again;
1579 	}
1580 	eoff = vn_a->offset & PAGEMASK;
1581 	if (eoff >= va.va_size) {
1582 		szcvec = 0;
1583 		goto again;
1584 	}
1585 	eoff += size;
1586 	if (btopr(va.va_size) < btopr(eoff)) {
1587 		save_size = size;
1588 		size = va.va_size - (vn_a->offset & PAGEMASK);
1589 		size = P2ROUNDUP_TYPED(size, PAGESIZE, size_t);
1590 		szcvec = map_pgszcvec(addr, size, (uintptr_t)addr, mapflags,
1591 		    type, 0);
1592 		if (szcvec <= 1) {
1593 			size = save_size;
1594 			goto again;
1595 		}
1596 	}
1597 
1598 	if (size > textrepl_size_thresh) {
1599 		vn_a->flags |= _MAP_TEXTREPL;
1600 	}
1601 	error = as_map_segvn_segs(as, addr, size, szcvec, crfp, vn_a,
1602 	    segcreated);
1603 	if (error != 0) {
1604 		return (error);
1605 	}
1606 	if (save_size) {
1607 		addr += size;
1608 		size = save_size - size;
1609 		szcvec = 0;
1610 		goto again;
1611 	}
1612 	return (0);
1613 }
1614 
1615 /*
1616  * as_map_ansegs: shared or private anonymous memory.  Note that the flags
1617  * passed to map_pgszvec cannot be MAP_INITDATA, for anon.
1618  */
1619 static int
as_map_ansegs(struct as * as,caddr_t addr,size_t size,segcreate_func_t crfp,struct segvn_crargs * vn_a,boolean_t * segcreated)1620 as_map_ansegs(struct as *as, caddr_t addr, size_t size,
1621     segcreate_func_t crfp, struct segvn_crargs *vn_a, boolean_t *segcreated)
1622 {
1623 	uint_t szcvec;
1624 	uchar_t type = 0;
1625 
1626 	ASSERT(vn_a->type == MAP_SHARED || vn_a->type == MAP_PRIVATE);
1627 	if (vn_a->type == MAP_SHARED) {
1628 		type = MAPPGSZC_SHM;
1629 	} else if (vn_a->type == MAP_PRIVATE) {
1630 		if (vn_a->szc == AS_MAP_HEAP) {
1631 			type = MAPPGSZC_HEAP;
1632 		} else if (vn_a->szc == AS_MAP_STACK) {
1633 			type = MAPPGSZC_STACK;
1634 		} else {
1635 			type = MAPPGSZC_PRIVM;
1636 		}
1637 	}
1638 	szcvec = map_pgszcvec(addr, size, vn_a->amp == NULL ?
1639 	    (uintptr_t)addr : (uintptr_t)P2ROUNDUP(vn_a->offset, PAGESIZE),
1640 	    (vn_a->flags & MAP_TEXT), type, 0);
1641 	ASSERT(AS_WRITE_HELD(as));
1642 	ASSERT(IS_P2ALIGNED(addr, PAGESIZE));
1643 	ASSERT(IS_P2ALIGNED(size, PAGESIZE));
1644 	ASSERT(vn_a->vp == NULL);
1645 
1646 	return (as_map_segvn_segs(as, addr, size, szcvec,
1647 	    crfp, vn_a, segcreated));
1648 }
1649 
1650 int
as_map(struct as * as,caddr_t addr,size_t size,segcreate_func_t crfp,void * argsp)1651 as_map(struct as *as, caddr_t addr, size_t size, segcreate_func_t crfp,
1652     void *argsp)
1653 {
1654 	AS_LOCK_ENTER(as, RW_WRITER);
1655 	return (as_map_locked(as, addr, size, crfp, argsp));
1656 }
1657 
1658 int
as_map_locked(struct as * as,caddr_t addr,size_t size,segcreate_func_t crfp,void * argsp)1659 as_map_locked(struct as *as, caddr_t addr, size_t size, segcreate_func_t crfp,
1660     void *argsp)
1661 {
1662 	caddr_t raddr;			/* rounded down addr */
1663 	size_t rsize;			/* rounded up size */
1664 	int error;
1665 	boolean_t is_hole = B_FALSE;
1666 	/*
1667 	 * The use of a_proc is preferred to handle the case where curproc is
1668 	 * a door_call server and is allocating memory in the client's (a_proc)
1669 	 * address space.
1670 	 * When creating a shared memory segment a_proc will be NULL so we
1671 	 * fallback to curproc in that case.
1672 	 */
1673 	struct proc *p = (as->a_proc == NULL) ? curproc : as->a_proc;
1674 	struct segvn_crargs crargs;
1675 
1676 	raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
1677 	rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) -
1678 	    (size_t)raddr;
1679 
1680 	/*
1681 	 * check for wrap around
1682 	 */
1683 	if ((raddr + rsize < raddr) || (as->a_size > (ULONG_MAX - size))) {
1684 		AS_LOCK_EXIT(as);
1685 		return (ENOMEM);
1686 	}
1687 
1688 	as->a_updatedir = 1;	/* inform /proc */
1689 	gethrestime(&as->a_updatetime);
1690 
1691 	if (as != &kas) {
1692 		/*
1693 		 * Ensure that the virtual size of the process will not exceed
1694 		 * the configured limit.  Since seg_hole segments will later
1695 		 * set the S_HOLE flag indicating their status as a hole in the
1696 		 * AS, they are excluded from this check.
1697 		 */
1698 		if (as->a_size + rsize > (size_t)p->p_vmem_ctl &&
1699 		    !AS_MAP_CHECK_SEGHOLE(crfp)) {
1700 			AS_LOCK_EXIT(as);
1701 
1702 			(void) rctl_action(rctlproc_legacy[RLIMIT_VMEM],
1703 			    p->p_rctls, p, RCA_UNSAFE_ALL);
1704 			return (ENOMEM);
1705 		}
1706 	}
1707 
1708 	if (AS_MAP_CHECK_VNODE_LPOOB(crfp, argsp)) {
1709 		boolean_t do_unmap = B_FALSE;
1710 
1711 		crargs = *(struct segvn_crargs *)argsp;
1712 		error = as_map_vnsegs(as, raddr, rsize, crfp, &crargs,
1713 		    &do_unmap);
1714 		if (error != 0) {
1715 			AS_LOCK_EXIT(as);
1716 			if (do_unmap) {
1717 				(void) as_unmap(as, addr, size);
1718 			}
1719 			return (error);
1720 		}
1721 	} else if (AS_MAP_CHECK_ANON_LPOOB(crfp, argsp)) {
1722 		boolean_t do_unmap = B_FALSE;
1723 
1724 		crargs = *(struct segvn_crargs *)argsp;
1725 		error = as_map_ansegs(as, raddr, rsize, crfp, &crargs,
1726 		    &do_unmap);
1727 		if (error != 0) {
1728 			AS_LOCK_EXIT(as);
1729 			if (do_unmap) {
1730 				(void) as_unmap(as, addr, size);
1731 			}
1732 			return (error);
1733 		}
1734 	} else {
1735 		struct seg *seg, *segref;
1736 
1737 		seg = segref = seg_alloc(as, addr, size);
1738 		if (seg == NULL) {
1739 			AS_LOCK_EXIT(as);
1740 			return (ENOMEM);
1741 		}
1742 
1743 		/*
1744 		 * It is possible that the segment creation routine will free
1745 		 * 'seg' as part of a more advanced operation, such as when
1746 		 * segvn concatenates adjacent segments together.  When this
1747 		 * occurs, the seg*_create routine must communicate the
1748 		 * resulting segment out via the 'struct seg **' parameter.
1749 		 *
1750 		 * If segment creation fails, it must not free the passed-in
1751 		 * segment, nor alter the argument pointer.
1752 		 */
1753 		error = (*crfp)(&seg, argsp);
1754 		if (error != 0) {
1755 			VERIFY3P(seg, ==, segref);
1756 			seg_free(seg);
1757 			AS_LOCK_EXIT(as);
1758 			return (error);
1759 		}
1760 
1761 		/*
1762 		 * Check if the resulting segment represents a hole in the
1763 		 * address space, rather than contributing to the AS size.
1764 		 */
1765 		is_hole = ((seg->s_flags & S_HOLE) != 0);
1766 
1767 		/* Add size now so as_unmap will work if as_ctl fails. */
1768 		if (!is_hole) {
1769 			as->a_size += rsize;
1770 			as->a_resvsize += rsize;
1771 		}
1772 	}
1773 
1774 	as_setwatch(as);
1775 
1776 	/*
1777 	 * Establish memory locks for the segment if the address space is
1778 	 * locked, provided it's not an explicit hole in the AS.
1779 	 */
1780 	mutex_enter(&as->a_contents);
1781 	if (AS_ISPGLCK(as) && !is_hole) {
1782 		mutex_exit(&as->a_contents);
1783 		AS_LOCK_EXIT(as);
1784 		error = as_ctl(as, addr, size, MC_LOCK, 0, 0, NULL, 0);
1785 		if (error != 0)
1786 			(void) as_unmap(as, addr, size);
1787 	} else {
1788 		mutex_exit(&as->a_contents);
1789 		AS_LOCK_EXIT(as);
1790 	}
1791 	return (error);
1792 }
1793 
1794 
1795 /*
1796  * Delete all segments in the address space marked with S_PURGE.
1797  * This is currently used for Sparc V9 nofault ASI segments (seg_nf.c).
1798  * These segments are deleted as a first step before calls to as_gap(), so
1799  * that they don't affect mmap() or shmat().
1800  */
1801 void
as_purge(struct as * as)1802 as_purge(struct as *as)
1803 {
1804 	struct seg *seg;
1805 	struct seg *next_seg;
1806 
1807 	/*
1808 	 * the setting of NEEDSPURGE is protect by as_rangelock(), so
1809 	 * no need to grab a_contents mutex for this check
1810 	 */
1811 	if ((as->a_flags & AS_NEEDSPURGE) == 0)
1812 		return;
1813 
1814 	AS_LOCK_ENTER(as, RW_WRITER);
1815 	next_seg = NULL;
1816 	seg = AS_SEGFIRST(as);
1817 	while (seg != NULL) {
1818 		next_seg = AS_SEGNEXT(as, seg);
1819 		if (seg->s_flags & S_PURGE)
1820 			SEGOP_UNMAP(seg, seg->s_base, seg->s_size);
1821 		seg = next_seg;
1822 	}
1823 	AS_LOCK_EXIT(as);
1824 
1825 	mutex_enter(&as->a_contents);
1826 	as->a_flags &= ~AS_NEEDSPURGE;
1827 	mutex_exit(&as->a_contents);
1828 }
1829 
1830 /*
1831  * Find a hole within [*basep, *basep + *lenp), which contains a mappable
1832  * range of addresses at least "minlen" long, where the base of the range is
1833  * at "off" phase from an "align" boundary and there is space for a
1834  * "redzone"-sized redzone on eithe rside of the range.  Thus,
1835  * if align was 4M and off was 16k, the user wants a hole which will start
1836  * 16k into a 4M page.
1837  *
1838  * If flags specifies AH_HI, the hole will have the highest possible address
1839  * in the range.  We use the as->a_lastgap field to figure out where to
1840  * start looking for a gap.
1841  *
1842  * Otherwise, the gap will have the lowest possible address.
1843  *
1844  * If flags specifies AH_CONTAIN, the hole will contain the address addr.
1845  *
1846  * If an adequate hole is found, *basep and *lenp are set to reflect the part of
1847  * the hole that is within range, and 0 is returned. On failure, -1 is returned.
1848  *
1849  * NOTE: This routine is not correct when base+len overflows caddr_t.
1850  */
1851 int
as_gap_aligned(struct as * as,size_t minlen,caddr_t * basep,size_t * lenp,uint_t flags,caddr_t addr,size_t align,size_t redzone,size_t off)1852 as_gap_aligned(struct as *as, size_t minlen, caddr_t *basep, size_t *lenp,
1853     uint_t flags, caddr_t addr, size_t align, size_t redzone, size_t off)
1854 {
1855 	caddr_t lobound = *basep;
1856 	caddr_t hibound = lobound + *lenp;
1857 	struct seg *lseg, *hseg;
1858 	caddr_t lo, hi;
1859 	int forward;
1860 	caddr_t save_base;
1861 	size_t save_len;
1862 	size_t save_minlen;
1863 	size_t save_redzone;
1864 	int fast_path = 1;
1865 
1866 	save_base = *basep;
1867 	save_len = *lenp;
1868 	save_minlen = minlen;
1869 	save_redzone = redzone;
1870 
1871 	/*
1872 	 * For the first pass/fast_path, just add align and redzone into
1873 	 * minlen since if we get an allocation, we can guarantee that it
1874 	 * will fit the alignment and redzone requested.
1875 	 * This increases the chance that hibound will be adjusted to
1876 	 * a_lastgap->s_base which will likely allow us to find an
1877 	 * acceptable hole in the address space quicker.
1878 	 * If we can't find a hole with this fast_path, then we look for
1879 	 * smaller holes in which the alignment and offset may allow
1880 	 * the allocation to fit.
1881 	 */
1882 	minlen += align;
1883 	minlen += 2 * redzone;
1884 	redzone = 0;
1885 
1886 	AS_LOCK_ENTER(as, RW_READER);
1887 	if (AS_SEGFIRST(as) == NULL) {
1888 		if (valid_va_range_aligned(basep, lenp, minlen, flags & AH_DIR,
1889 		    align, redzone, off)) {
1890 			AS_LOCK_EXIT(as);
1891 			return (0);
1892 		} else {
1893 			AS_LOCK_EXIT(as);
1894 			*basep = save_base;
1895 			*lenp = save_len;
1896 			return (-1);
1897 		}
1898 	}
1899 
1900 retry:
1901 	/*
1902 	 * Set up to iterate over all the inter-segment holes in the given
1903 	 * direction.  lseg is NULL for the lowest-addressed hole and hseg is
1904 	 * NULL for the highest-addressed hole.  If moving backwards, we reset
1905 	 * sseg to denote the highest-addressed segment.
1906 	 */
1907 	forward = (flags & AH_DIR) == AH_LO;
1908 	if (forward) {
1909 		hseg = as_findseg(as, lobound, 1);
1910 		lseg = AS_SEGPREV(as, hseg);
1911 	} else {
1912 
1913 		/*
1914 		 * If allocating at least as much as the last allocation,
1915 		 * use a_lastgap's base as a better estimate of hibound.
1916 		 */
1917 		if (as->a_lastgap &&
1918 		    minlen >= as->a_lastgap->s_size &&
1919 		    hibound >= as->a_lastgap->s_base)
1920 			hibound = as->a_lastgap->s_base;
1921 
1922 		hseg = as_findseg(as, hibound, 1);
1923 		if (hseg->s_base + hseg->s_size < hibound) {
1924 			lseg = hseg;
1925 			hseg = NULL;
1926 		} else {
1927 			lseg = AS_SEGPREV(as, hseg);
1928 		}
1929 	}
1930 
1931 	for (;;) {
1932 		/*
1933 		 * Set lo and hi to the hole's boundaries.  (We should really
1934 		 * use MAXADDR in place of hibound in the expression below,
1935 		 * but can't express it easily; using hibound in its place is
1936 		 * harmless.)
1937 		 */
1938 		lo = (lseg == NULL) ? 0 : lseg->s_base + lseg->s_size;
1939 		hi = (hseg == NULL) ? hibound : hseg->s_base;
1940 		/*
1941 		 * If the iteration has moved past the interval from lobound
1942 		 * to hibound it's pointless to continue.
1943 		 */
1944 		if ((forward && lo > hibound) || (!forward && hi < lobound))
1945 			break;
1946 		else if (lo > hibound || hi < lobound)
1947 			goto cont;
1948 		/*
1949 		 * Candidate hole lies at least partially within the allowable
1950 		 * range.  Restrict it to fall completely within that range,
1951 		 * i.e., to [max(lo, lobound), min(hi, hibound)].
1952 		 */
1953 		if (lo < lobound)
1954 			lo = lobound;
1955 		if (hi > hibound)
1956 			hi = hibound;
1957 		/*
1958 		 * Verify that the candidate hole is big enough and meets
1959 		 * hardware constraints.  If the hole is too small, no need
1960 		 * to do the further checks since they will fail.
1961 		 */
1962 		*basep = lo;
1963 		*lenp = hi - lo;
1964 		if (*lenp >= minlen && valid_va_range_aligned(basep, lenp,
1965 		    minlen, forward ? AH_LO : AH_HI, align, redzone, off) &&
1966 		    ((flags & AH_CONTAIN) == 0 ||
1967 		    (*basep <= addr && *basep + *lenp > addr))) {
1968 			if (!forward)
1969 				as->a_lastgap = hseg;
1970 			if (hseg != NULL)
1971 				as->a_lastgaphl = hseg;
1972 			else
1973 				as->a_lastgaphl = lseg;
1974 			AS_LOCK_EXIT(as);
1975 			return (0);
1976 		}
1977 	cont:
1978 		/*
1979 		 * Move to the next hole.
1980 		 */
1981 		if (forward) {
1982 			lseg = hseg;
1983 			if (lseg == NULL)
1984 				break;
1985 			hseg = AS_SEGNEXT(as, hseg);
1986 		} else {
1987 			hseg = lseg;
1988 			if (hseg == NULL)
1989 				break;
1990 			lseg = AS_SEGPREV(as, lseg);
1991 		}
1992 	}
1993 	if (fast_path && (align != 0 || save_redzone != 0)) {
1994 		fast_path = 0;
1995 		minlen = save_minlen;
1996 		redzone = save_redzone;
1997 		goto retry;
1998 	}
1999 	*basep = save_base;
2000 	*lenp = save_len;
2001 	AS_LOCK_EXIT(as);
2002 	return (-1);
2003 }
2004 
2005 /*
2006  * Find a hole of at least size minlen within [*basep, *basep + *lenp).
2007  *
2008  * If flags specifies AH_HI, the hole will have the highest possible address
2009  * in the range.  We use the as->a_lastgap field to figure out where to
2010  * start looking for a gap.
2011  *
2012  * Otherwise, the gap will have the lowest possible address.
2013  *
2014  * If flags specifies AH_CONTAIN, the hole will contain the address addr.
2015  *
2016  * If an adequate hole is found, base and len are set to reflect the part of
2017  * the hole that is within range, and 0 is returned, otherwise,
2018  * -1 is returned.
2019  *
2020  * NOTE: This routine is not correct when base+len overflows caddr_t.
2021  */
2022 int
as_gap(struct as * as,size_t minlen,caddr_t * basep,size_t * lenp,uint_t flags,caddr_t addr)2023 as_gap(struct as *as, size_t minlen, caddr_t *basep, size_t *lenp, uint_t flags,
2024     caddr_t addr)
2025 {
2026 
2027 	return (as_gap_aligned(as, minlen, basep, lenp, flags, addr, 0, 0, 0));
2028 }
2029 
2030 /*
2031  * Return the next range within [base, base + len) that is backed
2032  * with "real memory".  Skip holes and non-seg_vn segments.
2033  * We're lazy and only return one segment at a time.
2034  */
2035 int
as_memory(struct as * as,caddr_t * basep,size_t * lenp)2036 as_memory(struct as *as, caddr_t *basep, size_t *lenp)
2037 {
2038 	extern struct seg_ops segspt_shmops;	/* needs a header file */
2039 	struct seg *seg;
2040 	caddr_t addr, eaddr;
2041 	caddr_t segend;
2042 
2043 	AS_LOCK_ENTER(as, RW_READER);
2044 
2045 	addr = *basep;
2046 	eaddr = addr + *lenp;
2047 
2048 	seg = as_findseg(as, addr, 0);
2049 	if (seg != NULL)
2050 		addr = MAX(seg->s_base, addr);
2051 
2052 	for (;;) {
2053 		if (seg == NULL || addr >= eaddr || eaddr <= seg->s_base) {
2054 			AS_LOCK_EXIT(as);
2055 			return (EINVAL);
2056 		}
2057 
2058 		if (seg->s_ops == &segvn_ops) {
2059 			segend = seg->s_base + seg->s_size;
2060 			break;
2061 		}
2062 
2063 		/*
2064 		 * We do ISM by looking into the private data
2065 		 * to determine the real size of the segment.
2066 		 */
2067 		if (seg->s_ops == &segspt_shmops) {
2068 			segend = seg->s_base + spt_realsize(seg);
2069 			if (addr < segend)
2070 				break;
2071 		}
2072 
2073 		seg = AS_SEGNEXT(as, seg);
2074 
2075 		if (seg != NULL)
2076 			addr = seg->s_base;
2077 	}
2078 
2079 	*basep = addr;
2080 
2081 	if (segend > eaddr)
2082 		*lenp = eaddr - addr;
2083 	else
2084 		*lenp = segend - addr;
2085 
2086 	AS_LOCK_EXIT(as);
2087 	return (0);
2088 }
2089 
2090 /*
2091  * Swap the pages associated with the address space as out to
2092  * secondary storage, returning the number of bytes actually
2093  * swapped.
2094  *
2095  * The value returned is intended to correlate well with the process's
2096  * memory requirements.  Its usefulness for this purpose depends on
2097  * how well the segment-level routines do at returning accurate
2098  * information.
2099  */
2100 size_t
as_swapout(struct as * as)2101 as_swapout(struct as *as)
2102 {
2103 	struct seg *seg;
2104 	size_t swpcnt = 0;
2105 
2106 	/*
2107 	 * Kernel-only processes have given up their address
2108 	 * spaces.  Of course, we shouldn't be attempting to
2109 	 * swap out such processes in the first place...
2110 	 */
2111 	if (as == NULL)
2112 		return (0);
2113 
2114 	AS_LOCK_ENTER(as, RW_READER);
2115 
2116 	/*
2117 	 * Free all mapping resources associated with the address
2118 	 * space.  The segment-level swapout routines capitalize
2119 	 * on this unmapping by scavanging pages that have become
2120 	 * unmapped here.
2121 	 */
2122 	hat_swapout(as->a_hat);
2123 
2124 	/*
2125 	 * Call the swapout routines of all segments in the address
2126 	 * space to do the actual work, accumulating the amount of
2127 	 * space reclaimed.
2128 	 */
2129 	for (seg = AS_SEGFIRST(as); seg != NULL; seg = AS_SEGNEXT(as, seg)) {
2130 		struct seg_ops *ov = seg->s_ops;
2131 
2132 		/*
2133 		 * We have to check to see if the seg has
2134 		 * an ops vector because the seg may have
2135 		 * been in the middle of being set up when
2136 		 * the process was picked for swapout.
2137 		 */
2138 		if ((ov != NULL) && (ov->swapout != NULL))
2139 			swpcnt += SEGOP_SWAPOUT(seg);
2140 	}
2141 	AS_LOCK_EXIT(as);
2142 	return (swpcnt);
2143 }
2144 
2145 /*
2146  * Determine whether data from the mappings in interval [addr, addr + size)
2147  * are in the primary memory (core) cache.
2148  */
2149 int
as_incore(struct as * as,caddr_t addr,size_t size,char * vec,size_t * sizep)2150 as_incore(struct as *as, caddr_t addr,
2151     size_t size, char *vec, size_t *sizep)
2152 {
2153 	struct seg *seg;
2154 	size_t ssize;
2155 	caddr_t raddr;		/* rounded down addr */
2156 	size_t rsize;		/* rounded up size */
2157 	size_t isize;			/* iteration size */
2158 	int error = 0;		/* result, assume success */
2159 
2160 	*sizep = 0;
2161 	raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
2162 	rsize = ((((size_t)addr + size) + PAGEOFFSET) & PAGEMASK) -
2163 	    (size_t)raddr;
2164 
2165 	if (raddr + rsize < raddr)		/* check for wraparound */
2166 		return (ENOMEM);
2167 
2168 	AS_LOCK_ENTER(as, RW_READER);
2169 	seg = as_segat(as, raddr);
2170 	if (seg == NULL) {
2171 		AS_LOCK_EXIT(as);
2172 		return (-1);
2173 	}
2174 
2175 	for (; rsize != 0; rsize -= ssize, raddr += ssize) {
2176 		if (raddr >= seg->s_base + seg->s_size) {
2177 			seg = AS_SEGNEXT(as, seg);
2178 			if (seg == NULL || raddr != seg->s_base) {
2179 				error = -1;
2180 				break;
2181 			}
2182 		}
2183 		if ((raddr + rsize) > (seg->s_base + seg->s_size))
2184 			ssize = seg->s_base + seg->s_size - raddr;
2185 		else
2186 			ssize = rsize;
2187 		*sizep += isize = SEGOP_INCORE(seg, raddr, ssize, vec);
2188 		if (isize != ssize) {
2189 			error = -1;
2190 			break;
2191 		}
2192 		vec += btopr(ssize);
2193 	}
2194 	AS_LOCK_EXIT(as);
2195 	return (error);
2196 }
2197 
2198 static void
as_segunlock(struct seg * seg,caddr_t addr,int attr,ulong_t * bitmap,size_t position,size_t npages)2199 as_segunlock(struct seg *seg, caddr_t addr, int attr,
2200     ulong_t *bitmap, size_t position, size_t npages)
2201 {
2202 	caddr_t	range_start;
2203 	size_t	pos1 = position;
2204 	size_t	pos2;
2205 	size_t	size;
2206 	size_t  end_pos = npages + position;
2207 
2208 	while (bt_range(bitmap, &pos1, &pos2, end_pos)) {
2209 		size = ptob((pos2 - pos1));
2210 		range_start = (caddr_t)((uintptr_t)addr +
2211 		    ptob(pos1 - position));
2212 
2213 		(void) SEGOP_LOCKOP(seg, range_start, size, attr, MC_UNLOCK,
2214 		    (ulong_t *)NULL, (size_t)NULL);
2215 		pos1 = pos2;
2216 	}
2217 }
2218 
2219 static void
as_unlockerr(struct as * as,int attr,ulong_t * mlock_map,caddr_t raddr,size_t rsize)2220 as_unlockerr(struct as *as, int attr, ulong_t *mlock_map,
2221     caddr_t raddr, size_t rsize)
2222 {
2223 	struct seg *seg = as_segat(as, raddr);
2224 	size_t ssize;
2225 
2226 	while (rsize != 0) {
2227 		if (raddr >= seg->s_base + seg->s_size)
2228 			seg = AS_SEGNEXT(as, seg);
2229 
2230 		if ((raddr + rsize) > (seg->s_base + seg->s_size))
2231 			ssize = seg->s_base + seg->s_size - raddr;
2232 		else
2233 			ssize = rsize;
2234 
2235 		as_segunlock(seg, raddr, attr, mlock_map, 0, btopr(ssize));
2236 
2237 		rsize -= ssize;
2238 		raddr += ssize;
2239 	}
2240 }
2241 
2242 /*
2243  * Cache control operations over the interval [addr, addr + size) in
2244  * address space "as".
2245  */
2246 /*ARGSUSED*/
2247 int
as_ctl(struct as * as,caddr_t addr,size_t size,int func,int attr,uintptr_t arg,ulong_t * lock_map,size_t pos)2248 as_ctl(struct as *as, caddr_t addr, size_t size, int func, int attr,
2249     uintptr_t arg, ulong_t *lock_map, size_t pos)
2250 {
2251 	struct seg *seg;	/* working segment */
2252 	caddr_t raddr;		/* rounded down addr */
2253 	caddr_t initraddr;	/* saved initial rounded down addr */
2254 	size_t rsize;		/* rounded up size */
2255 	size_t initrsize;	/* saved initial rounded up size */
2256 	size_t ssize;		/* size of seg */
2257 	int error = 0;			/* result */
2258 	size_t mlock_size;	/* size of bitmap */
2259 	ulong_t *mlock_map;	/* pointer to bitmap used */
2260 				/* to represent the locked */
2261 				/* pages. */
2262 
2263 	mlock_size = 0;
2264 	mlock_map = NULL;
2265 retry:
2266 	if (error == IE_RETRY)
2267 		AS_LOCK_ENTER(as, RW_WRITER);
2268 	else
2269 		AS_LOCK_ENTER(as, RW_READER);
2270 
2271 	/*
2272 	 * If these are address space lock/unlock operations, loop over
2273 	 * all segments in the address space, as appropriate.
2274 	 */
2275 	if (func == MC_LOCKAS) {
2276 		size_t npages, idx;
2277 		size_t rlen = 0;	/* rounded as length */
2278 
2279 		idx = pos;
2280 
2281 		if (arg & MCL_FUTURE) {
2282 			mutex_enter(&as->a_contents);
2283 			AS_SETPGLCK(as);
2284 			mutex_exit(&as->a_contents);
2285 		}
2286 		if ((arg & MCL_CURRENT) == 0) {
2287 			AS_LOCK_EXIT(as);
2288 			return (0);
2289 		}
2290 
2291 		seg = AS_SEGFIRST(as);
2292 		if (seg == NULL) {
2293 			AS_LOCK_EXIT(as);
2294 			return (0);
2295 		}
2296 
2297 		do {
2298 			raddr = (caddr_t)((uintptr_t)seg->s_base &
2299 			    (uintptr_t)PAGEMASK);
2300 			rlen += (((uintptr_t)(seg->s_base + seg->s_size) +
2301 			    PAGEOFFSET) & PAGEMASK) - (uintptr_t)raddr;
2302 		} while ((seg = AS_SEGNEXT(as, seg)) != NULL);
2303 
2304 		mlock_size = BT_BITOUL(btopr(rlen));
2305 		if ((mlock_map = (ulong_t *)kmem_zalloc(mlock_size *
2306 		    sizeof (ulong_t), KM_NOSLEEP)) == NULL) {
2307 				AS_LOCK_EXIT(as);
2308 				return (EAGAIN);
2309 		}
2310 
2311 		for (seg = AS_SEGFIRST(as); seg; seg = AS_SEGNEXT(as, seg)) {
2312 			if ((seg->s_flags & S_HOLE) != 0) {
2313 				continue;
2314 			}
2315 			error = SEGOP_LOCKOP(seg, seg->s_base,
2316 			    seg->s_size, attr, MC_LOCK, mlock_map, pos);
2317 			if (error != 0)
2318 				break;
2319 			pos += seg_pages(seg);
2320 		}
2321 
2322 		if (error) {
2323 			for (seg = AS_SEGFIRST(as); seg != NULL;
2324 			    seg = AS_SEGNEXT(as, seg)) {
2325 
2326 				raddr = (caddr_t)((uintptr_t)seg->s_base &
2327 				    (uintptr_t)PAGEMASK);
2328 				npages = seg_pages(seg);
2329 				as_segunlock(seg, raddr, attr, mlock_map,
2330 				    idx, npages);
2331 				idx += npages;
2332 			}
2333 		}
2334 
2335 		kmem_free(mlock_map, mlock_size * sizeof (ulong_t));
2336 		AS_LOCK_EXIT(as);
2337 		goto lockerr;
2338 	} else if (func == MC_UNLOCKAS) {
2339 		mutex_enter(&as->a_contents);
2340 		AS_CLRPGLCK(as);
2341 		mutex_exit(&as->a_contents);
2342 
2343 		for (seg = AS_SEGFIRST(as); seg; seg = AS_SEGNEXT(as, seg)) {
2344 			if ((seg->s_flags & S_HOLE) != 0) {
2345 				continue;
2346 			}
2347 			error = SEGOP_LOCKOP(seg, seg->s_base,
2348 			    seg->s_size, attr, MC_UNLOCK, NULL, 0);
2349 			if (error != 0)
2350 				break;
2351 		}
2352 
2353 		AS_LOCK_EXIT(as);
2354 		goto lockerr;
2355 	}
2356 
2357 	/*
2358 	 * Normalize addresses and sizes.
2359 	 */
2360 	initraddr = raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
2361 	initrsize = rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) -
2362 	    (size_t)raddr;
2363 
2364 	if (raddr + rsize < raddr) {		/* check for wraparound */
2365 		AS_LOCK_EXIT(as);
2366 		return (ENOMEM);
2367 	}
2368 
2369 	/*
2370 	 * Get initial segment.
2371 	 */
2372 	if ((seg = as_segat(as, raddr)) == NULL) {
2373 		AS_LOCK_EXIT(as);
2374 		return (ENOMEM);
2375 	}
2376 
2377 	if (func == MC_LOCK) {
2378 		mlock_size = BT_BITOUL(btopr(rsize));
2379 		if ((mlock_map = (ulong_t *)kmem_zalloc(mlock_size *
2380 		    sizeof (ulong_t), KM_NOSLEEP)) == NULL) {
2381 				AS_LOCK_EXIT(as);
2382 				return (EAGAIN);
2383 		}
2384 	}
2385 
2386 	/*
2387 	 * Loop over all segments.  If a hole in the address range is
2388 	 * discovered, then fail.  For each segment, perform the appropriate
2389 	 * control operation.
2390 	 */
2391 	while (rsize != 0) {
2392 
2393 		/*
2394 		 * Make sure there's no hole, calculate the portion
2395 		 * of the next segment to be operated over.
2396 		 */
2397 		if (raddr >= seg->s_base + seg->s_size) {
2398 			seg = AS_SEGNEXT(as, seg);
2399 			if (seg == NULL || raddr != seg->s_base) {
2400 				if (func == MC_LOCK) {
2401 					as_unlockerr(as, attr, mlock_map,
2402 					    initraddr, initrsize - rsize);
2403 					kmem_free(mlock_map,
2404 					    mlock_size * sizeof (ulong_t));
2405 				}
2406 				AS_LOCK_EXIT(as);
2407 				return (ENOMEM);
2408 			}
2409 		}
2410 		if ((raddr + rsize) > (seg->s_base + seg->s_size))
2411 			ssize = seg->s_base + seg->s_size - raddr;
2412 		else
2413 			ssize = rsize;
2414 
2415 		/*
2416 		 * Dispatch on specific function.
2417 		 */
2418 		switch (func) {
2419 
2420 		/*
2421 		 * Synchronize cached data from mappings with backing
2422 		 * objects.
2423 		 */
2424 		case MC_SYNC:
2425 			if (error = SEGOP_SYNC(seg, raddr, ssize,
2426 			    attr, (uint_t)arg)) {
2427 				AS_LOCK_EXIT(as);
2428 				return (error);
2429 			}
2430 			break;
2431 
2432 		/*
2433 		 * Lock pages in memory.
2434 		 */
2435 		case MC_LOCK:
2436 			if (error = SEGOP_LOCKOP(seg, raddr, ssize,
2437 			    attr, func, mlock_map, pos)) {
2438 				as_unlockerr(as, attr, mlock_map, initraddr,
2439 				    initrsize - rsize + ssize);
2440 				kmem_free(mlock_map, mlock_size *
2441 				    sizeof (ulong_t));
2442 				AS_LOCK_EXIT(as);
2443 				goto lockerr;
2444 			}
2445 			break;
2446 
2447 		/*
2448 		 * Unlock mapped pages.
2449 		 */
2450 		case MC_UNLOCK:
2451 			(void) SEGOP_LOCKOP(seg, raddr, ssize, attr, func,
2452 			    (ulong_t *)NULL, (size_t)NULL);
2453 			break;
2454 
2455 		/*
2456 		 * Store VM advise for mapped pages in segment layer.
2457 		 */
2458 		case MC_ADVISE:
2459 			error = SEGOP_ADVISE(seg, raddr, ssize, (uint_t)arg);
2460 
2461 			/*
2462 			 * Check for regular errors and special retry error
2463 			 */
2464 			if (error) {
2465 				if (error == IE_RETRY) {
2466 					/*
2467 					 * Need to acquire writers lock, so
2468 					 * have to drop readers lock and start
2469 					 * all over again
2470 					 */
2471 					AS_LOCK_EXIT(as);
2472 					goto retry;
2473 				} else if (error == IE_REATTACH) {
2474 					/*
2475 					 * Find segment for current address
2476 					 * because current segment just got
2477 					 * split or concatenated
2478 					 */
2479 					seg = as_segat(as, raddr);
2480 					if (seg == NULL) {
2481 						AS_LOCK_EXIT(as);
2482 						return (ENOMEM);
2483 					}
2484 				} else {
2485 					/*
2486 					 * Regular error
2487 					 */
2488 					AS_LOCK_EXIT(as);
2489 					return (error);
2490 				}
2491 			}
2492 			break;
2493 
2494 		case MC_INHERIT_ZERO:
2495 			if (seg->s_ops->inherit == NULL) {
2496 				error = ENOTSUP;
2497 			} else {
2498 				error = SEGOP_INHERIT(seg, raddr, ssize,
2499 				    SEGP_INH_ZERO);
2500 			}
2501 			if (error != 0) {
2502 				AS_LOCK_EXIT(as);
2503 				return (error);
2504 			}
2505 			break;
2506 
2507 		/*
2508 		 * Can't happen.
2509 		 */
2510 		default:
2511 			panic("as_ctl: bad operation %d", func);
2512 			/*NOTREACHED*/
2513 		}
2514 
2515 		rsize -= ssize;
2516 		raddr += ssize;
2517 	}
2518 
2519 	if (func == MC_LOCK)
2520 		kmem_free(mlock_map, mlock_size * sizeof (ulong_t));
2521 	AS_LOCK_EXIT(as);
2522 	return (0);
2523 lockerr:
2524 
2525 	/*
2526 	 * If the lower levels returned EDEADLK for a segment lockop,
2527 	 * it means that we should retry the operation.  Let's wait
2528 	 * a bit also to let the deadlock causing condition clear.
2529 	 * This is part of a gross hack to work around a design flaw
2530 	 * in the ufs/sds logging code and should go away when the
2531 	 * logging code is re-designed to fix the problem. See bug
2532 	 * 4125102 for details of the problem.
2533 	 */
2534 	if (error == EDEADLK) {
2535 		delay(deadlk_wait);
2536 		error = 0;
2537 		goto retry;
2538 	}
2539 	return (error);
2540 }
2541 
2542 int
fc_decode(faultcode_t fault_err)2543 fc_decode(faultcode_t fault_err)
2544 {
2545 	int error = 0;
2546 
2547 	switch (FC_CODE(fault_err)) {
2548 	case FC_OBJERR:
2549 		error = FC_ERRNO(fault_err);
2550 		break;
2551 	case FC_PROT:
2552 		error = EACCES;
2553 		break;
2554 	default:
2555 		error = EFAULT;
2556 		break;
2557 	}
2558 	return (error);
2559 }
2560 
2561 /*
2562  * Pagelock pages from a range that spans more than 1 segment.  Obtain shadow
2563  * lists from each segment and copy them to one contiguous shadow list (plist)
2564  * as expected by the caller.  Save pointers to per segment shadow lists at
2565  * the tail of plist so that they can be used during as_pageunlock().
2566  */
2567 static int
as_pagelock_segs(struct as * as,struct seg * seg,struct page *** ppp,caddr_t addr,size_t size,enum seg_rw rw)2568 as_pagelock_segs(struct as *as, struct seg *seg, struct page ***ppp,
2569     caddr_t addr, size_t size, enum seg_rw rw)
2570 {
2571 	caddr_t sv_addr = addr;
2572 	size_t sv_size = size;
2573 	struct seg *sv_seg = seg;
2574 	ulong_t segcnt = 1;
2575 	ulong_t cnt;
2576 	size_t ssize;
2577 	pgcnt_t npages = btop(size);
2578 	page_t **plist;
2579 	page_t **pl;
2580 	int error;
2581 	caddr_t eaddr;
2582 	faultcode_t fault_err = 0;
2583 	pgcnt_t pl_off;
2584 	extern struct seg_ops segspt_shmops;
2585 
2586 	ASSERT(AS_LOCK_HELD(as));
2587 	ASSERT(seg != NULL);
2588 	ASSERT(addr >= seg->s_base && addr < seg->s_base + seg->s_size);
2589 	ASSERT(addr + size > seg->s_base + seg->s_size);
2590 	ASSERT(IS_P2ALIGNED(size, PAGESIZE));
2591 	ASSERT(IS_P2ALIGNED(addr, PAGESIZE));
2592 
2593 	/*
2594 	 * Count the number of segments covered by the range we are about to
2595 	 * lock. The segment count is used to size the shadow list we return
2596 	 * back to the caller.
2597 	 */
2598 	for (; size != 0; size -= ssize, addr += ssize) {
2599 		if (addr >= seg->s_base + seg->s_size) {
2600 
2601 			seg = AS_SEGNEXT(as, seg);
2602 			if (seg == NULL || addr != seg->s_base) {
2603 				AS_LOCK_EXIT(as);
2604 				return (EFAULT);
2605 			}
2606 			/*
2607 			 * Do a quick check if subsequent segments
2608 			 * will most likely support pagelock.
2609 			 */
2610 			if (seg->s_ops == &segvn_ops) {
2611 				vnode_t *vp;
2612 
2613 				if (SEGOP_GETVP(seg, addr, &vp) != 0 ||
2614 				    vp != NULL) {
2615 					AS_LOCK_EXIT(as);
2616 					goto slow;
2617 				}
2618 			} else if (seg->s_ops != &segspt_shmops) {
2619 				AS_LOCK_EXIT(as);
2620 				goto slow;
2621 			}
2622 			segcnt++;
2623 		}
2624 		if (addr + size > seg->s_base + seg->s_size) {
2625 			ssize = seg->s_base + seg->s_size - addr;
2626 		} else {
2627 			ssize = size;
2628 		}
2629 	}
2630 	ASSERT(segcnt > 1);
2631 
2632 	plist = kmem_zalloc((npages + segcnt) * sizeof (page_t *), KM_SLEEP);
2633 
2634 	addr = sv_addr;
2635 	size = sv_size;
2636 	seg = sv_seg;
2637 
2638 	for (cnt = 0, pl_off = 0; size != 0; size -= ssize, addr += ssize) {
2639 		if (addr >= seg->s_base + seg->s_size) {
2640 			seg = AS_SEGNEXT(as, seg);
2641 			ASSERT(seg != NULL && addr == seg->s_base);
2642 			cnt++;
2643 			ASSERT(cnt < segcnt);
2644 		}
2645 		if (addr + size > seg->s_base + seg->s_size) {
2646 			ssize = seg->s_base + seg->s_size - addr;
2647 		} else {
2648 			ssize = size;
2649 		}
2650 		pl = &plist[npages + cnt];
2651 		error = SEGOP_PAGELOCK(seg, addr, ssize, (page_t ***)pl,
2652 		    L_PAGELOCK, rw);
2653 		if (error) {
2654 			break;
2655 		}
2656 		ASSERT(plist[npages + cnt] != NULL);
2657 		ASSERT(pl_off + btop(ssize) <= npages);
2658 		bcopy(plist[npages + cnt], &plist[pl_off],
2659 		    btop(ssize) * sizeof (page_t *));
2660 		pl_off += btop(ssize);
2661 	}
2662 
2663 	if (size == 0) {
2664 		AS_LOCK_EXIT(as);
2665 		ASSERT(cnt == segcnt - 1);
2666 		*ppp = plist;
2667 		return (0);
2668 	}
2669 
2670 	/*
2671 	 * one of pagelock calls failed. The error type is in error variable.
2672 	 * Unlock what we've locked so far and retry with F_SOFTLOCK if error
2673 	 * type is either EFAULT or ENOTSUP. Otherwise just return the error
2674 	 * back to the caller.
2675 	 */
2676 
2677 	eaddr = addr;
2678 	seg = sv_seg;
2679 
2680 	for (cnt = 0, addr = sv_addr; addr < eaddr; addr += ssize) {
2681 		if (addr >= seg->s_base + seg->s_size) {
2682 			seg = AS_SEGNEXT(as, seg);
2683 			ASSERT(seg != NULL && addr == seg->s_base);
2684 			cnt++;
2685 			ASSERT(cnt < segcnt);
2686 		}
2687 		if (eaddr > seg->s_base + seg->s_size) {
2688 			ssize = seg->s_base + seg->s_size - addr;
2689 		} else {
2690 			ssize = eaddr - addr;
2691 		}
2692 		pl = &plist[npages + cnt];
2693 		ASSERT(*pl != NULL);
2694 		(void) SEGOP_PAGELOCK(seg, addr, ssize, (page_t ***)pl,
2695 		    L_PAGEUNLOCK, rw);
2696 	}
2697 
2698 	AS_LOCK_EXIT(as);
2699 
2700 	kmem_free(plist, (npages + segcnt) * sizeof (page_t *));
2701 
2702 	if (error != ENOTSUP && error != EFAULT) {
2703 		return (error);
2704 	}
2705 
2706 slow:
2707 	/*
2708 	 * If we are here because pagelock failed due to the need to cow fault
2709 	 * in the pages we want to lock F_SOFTLOCK will do this job and in
2710 	 * next as_pagelock() call for this address range pagelock will
2711 	 * hopefully succeed.
2712 	 */
2713 	fault_err = as_fault(as->a_hat, as, sv_addr, sv_size, F_SOFTLOCK, rw);
2714 	if (fault_err != 0) {
2715 		return (fc_decode(fault_err));
2716 	}
2717 	*ppp = NULL;
2718 
2719 	return (0);
2720 }
2721 
2722 /*
2723  * lock pages in a given address space. Return shadow list. If
2724  * the list is NULL, the MMU mapping is also locked.
2725  */
2726 int
as_pagelock(struct as * as,struct page *** ppp,caddr_t addr,size_t size,enum seg_rw rw)2727 as_pagelock(struct as *as, struct page ***ppp, caddr_t addr,
2728     size_t size, enum seg_rw rw)
2729 {
2730 	size_t rsize;
2731 	caddr_t raddr;
2732 	faultcode_t fault_err;
2733 	struct seg *seg;
2734 	int err;
2735 
2736 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_AS_LOCK_START,
2737 	    "as_pagelock_start: addr %p size %ld", addr, size);
2738 
2739 	raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
2740 	rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) -
2741 	    (size_t)raddr;
2742 
2743 	/*
2744 	 * if the request crosses two segments let
2745 	 * as_fault handle it.
2746 	 */
2747 	AS_LOCK_ENTER(as, RW_READER);
2748 
2749 	seg = as_segat(as, raddr);
2750 	if (seg == NULL) {
2751 		AS_LOCK_EXIT(as);
2752 		return (EFAULT);
2753 	}
2754 	ASSERT(raddr >= seg->s_base && raddr < seg->s_base + seg->s_size);
2755 	if (raddr + rsize > seg->s_base + seg->s_size) {
2756 		return (as_pagelock_segs(as, seg, ppp, raddr, rsize, rw));
2757 	}
2758 	if (raddr + rsize <= raddr) {
2759 		AS_LOCK_EXIT(as);
2760 		return (EFAULT);
2761 	}
2762 
2763 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEG_LOCK_START,
2764 	    "seg_lock_1_start: raddr %p rsize %ld", raddr, rsize);
2765 
2766 	/*
2767 	 * try to lock pages and pass back shadow list
2768 	 */
2769 	err = SEGOP_PAGELOCK(seg, raddr, rsize, ppp, L_PAGELOCK, rw);
2770 
2771 	TRACE_0(TR_FAC_PHYSIO, TR_PHYSIO_SEG_LOCK_END, "seg_lock_1_end");
2772 
2773 	AS_LOCK_EXIT(as);
2774 
2775 	if (err == 0 || (err != ENOTSUP && err != EFAULT)) {
2776 		return (err);
2777 	}
2778 
2779 	/*
2780 	 * Use F_SOFTLOCK to lock the pages because pagelock failed either due
2781 	 * to no pagelock support for this segment or pages need to be cow
2782 	 * faulted in. If fault is needed F_SOFTLOCK will do this job for
2783 	 * this as_pagelock() call and in the next as_pagelock() call for the
2784 	 * same address range pagelock call will hopefull succeed.
2785 	 */
2786 	fault_err = as_fault(as->a_hat, as, addr, size, F_SOFTLOCK, rw);
2787 	if (fault_err != 0) {
2788 		return (fc_decode(fault_err));
2789 	}
2790 	*ppp = NULL;
2791 
2792 	TRACE_0(TR_FAC_PHYSIO, TR_PHYSIO_AS_LOCK_END, "as_pagelock_end");
2793 	return (0);
2794 }
2795 
2796 /*
2797  * unlock pages locked by as_pagelock_segs().  Retrieve per segment shadow
2798  * lists from the end of plist and call pageunlock interface for each segment.
2799  * Drop as lock and free plist.
2800  */
2801 static void
as_pageunlock_segs(struct as * as,struct seg * seg,caddr_t addr,size_t size,struct page ** plist,enum seg_rw rw)2802 as_pageunlock_segs(struct as *as, struct seg *seg, caddr_t addr, size_t size,
2803     struct page **plist, enum seg_rw rw)
2804 {
2805 	ulong_t cnt;
2806 	caddr_t eaddr = addr + size;
2807 	pgcnt_t npages = btop(size);
2808 	size_t ssize;
2809 	page_t **pl;
2810 
2811 	ASSERT(AS_LOCK_HELD(as));
2812 	ASSERT(seg != NULL);
2813 	ASSERT(addr >= seg->s_base && addr < seg->s_base + seg->s_size);
2814 	ASSERT(addr + size > seg->s_base + seg->s_size);
2815 	ASSERT(IS_P2ALIGNED(size, PAGESIZE));
2816 	ASSERT(IS_P2ALIGNED(addr, PAGESIZE));
2817 	ASSERT(plist != NULL);
2818 
2819 	for (cnt = 0; addr < eaddr; addr += ssize) {
2820 		if (addr >= seg->s_base + seg->s_size) {
2821 			seg = AS_SEGNEXT(as, seg);
2822 			ASSERT(seg != NULL && addr == seg->s_base);
2823 			cnt++;
2824 		}
2825 		if (eaddr > seg->s_base + seg->s_size) {
2826 			ssize = seg->s_base + seg->s_size - addr;
2827 		} else {
2828 			ssize = eaddr - addr;
2829 		}
2830 		pl = &plist[npages + cnt];
2831 		ASSERT(*pl != NULL);
2832 		(void) SEGOP_PAGELOCK(seg, addr, ssize, (page_t ***)pl,
2833 		    L_PAGEUNLOCK, rw);
2834 	}
2835 	ASSERT(cnt > 0);
2836 	AS_LOCK_EXIT(as);
2837 
2838 	cnt++;
2839 	kmem_free(plist, (npages + cnt) * sizeof (page_t *));
2840 }
2841 
2842 /*
2843  * unlock pages in a given address range
2844  */
2845 void
as_pageunlock(struct as * as,struct page ** pp,caddr_t addr,size_t size,enum seg_rw rw)2846 as_pageunlock(struct as *as, struct page **pp, caddr_t addr, size_t size,
2847     enum seg_rw rw)
2848 {
2849 	struct seg *seg;
2850 	size_t rsize;
2851 	caddr_t raddr;
2852 
2853 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_AS_UNLOCK_START,
2854 	    "as_pageunlock_start: addr %p size %ld", addr, size);
2855 
2856 	/*
2857 	 * if the shadow list is NULL, as_pagelock was
2858 	 * falling back to as_fault
2859 	 */
2860 	if (pp == NULL) {
2861 		(void) as_fault(as->a_hat, as, addr, size, F_SOFTUNLOCK, rw);
2862 		return;
2863 	}
2864 
2865 	raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
2866 	rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) -
2867 	    (size_t)raddr;
2868 
2869 	AS_LOCK_ENTER(as, RW_READER);
2870 	seg = as_segat(as, raddr);
2871 	ASSERT(seg != NULL);
2872 
2873 	TRACE_2(TR_FAC_PHYSIO, TR_PHYSIO_SEG_UNLOCK_START,
2874 	    "seg_unlock_start: raddr %p rsize %ld", raddr, rsize);
2875 
2876 	ASSERT(raddr >= seg->s_base && raddr < seg->s_base + seg->s_size);
2877 	if (raddr + rsize <= seg->s_base + seg->s_size) {
2878 		SEGOP_PAGELOCK(seg, raddr, rsize, &pp, L_PAGEUNLOCK, rw);
2879 	} else {
2880 		as_pageunlock_segs(as, seg, raddr, rsize, pp, rw);
2881 		return;
2882 	}
2883 	AS_LOCK_EXIT(as);
2884 	TRACE_0(TR_FAC_PHYSIO, TR_PHYSIO_AS_UNLOCK_END, "as_pageunlock_end");
2885 }
2886 
2887 int
as_setpagesize(struct as * as,caddr_t addr,size_t size,uint_t szc,boolean_t wait)2888 as_setpagesize(struct as *as, caddr_t addr, size_t size, uint_t szc,
2889     boolean_t wait)
2890 {
2891 	struct seg *seg;
2892 	size_t ssize;
2893 	caddr_t raddr;			/* rounded down addr */
2894 	size_t rsize;			/* rounded up size */
2895 	int error = 0;
2896 	size_t pgsz = page_get_pagesize(szc);
2897 
2898 setpgsz_top:
2899 	if (!IS_P2ALIGNED(addr, pgsz) || !IS_P2ALIGNED(size, pgsz)) {
2900 		return (EINVAL);
2901 	}
2902 
2903 	raddr = addr;
2904 	rsize = size;
2905 
2906 	if (raddr + rsize < raddr)		/* check for wraparound */
2907 		return (ENOMEM);
2908 
2909 	AS_LOCK_ENTER(as, RW_WRITER);
2910 	as_clearwatchprot(as, raddr, rsize);
2911 	seg = as_segat(as, raddr);
2912 	if (seg == NULL) {
2913 		as_setwatch(as);
2914 		AS_LOCK_EXIT(as);
2915 		return (ENOMEM);
2916 	}
2917 
2918 	for (; rsize != 0; rsize -= ssize, raddr += ssize) {
2919 		if (raddr >= seg->s_base + seg->s_size) {
2920 			seg = AS_SEGNEXT(as, seg);
2921 			if (seg == NULL || raddr != seg->s_base) {
2922 				error = ENOMEM;
2923 				break;
2924 			}
2925 		}
2926 		if ((raddr + rsize) > (seg->s_base + seg->s_size)) {
2927 			ssize = seg->s_base + seg->s_size - raddr;
2928 		} else {
2929 			ssize = rsize;
2930 		}
2931 
2932 retry:
2933 		error = SEGOP_SETPAGESIZE(seg, raddr, ssize, szc);
2934 
2935 		if (error == IE_NOMEM) {
2936 			error = EAGAIN;
2937 			break;
2938 		}
2939 
2940 		if (error == IE_RETRY) {
2941 			AS_LOCK_EXIT(as);
2942 			goto setpgsz_top;
2943 		}
2944 
2945 		if (error == ENOTSUP) {
2946 			error = EINVAL;
2947 			break;
2948 		}
2949 
2950 		if (wait && (error == EAGAIN)) {
2951 			/*
2952 			 * Memory is currently locked.  It must be unlocked
2953 			 * before this operation can succeed through a retry.
2954 			 * The possible reasons for locked memory and
2955 			 * corresponding strategies for unlocking are:
2956 			 * (1) Normal I/O
2957 			 *	wait for a signal that the I/O operation
2958 			 *	has completed and the memory is unlocked.
2959 			 * (2) Asynchronous I/O
2960 			 *	The aio subsystem does not unlock pages when
2961 			 *	the I/O is completed. Those pages are unlocked
2962 			 *	when the application calls aiowait/aioerror.
2963 			 *	So, to prevent blocking forever, cv_broadcast()
2964 			 *	is done to wake up aio_cleanup_thread.
2965 			 *	Subsequently, segvn_reclaim will be called, and
2966 			 *	that will do AS_CLRUNMAPWAIT() and wake us up.
2967 			 * (3) Long term page locking:
2968 			 *	This is not relevant for as_setpagesize()
2969 			 *	because we cannot change the page size for
2970 			 *	driver memory. The attempt to do so will
2971 			 *	fail with a different error than EAGAIN so
2972 			 *	there's no need to trigger as callbacks like
2973 			 *	as_unmap, as_setprot or as_free would do.
2974 			 */
2975 			mutex_enter(&as->a_contents);
2976 			if (!AS_ISNOUNMAPWAIT(as)) {
2977 				if (AS_ISUNMAPWAIT(as) == 0) {
2978 					cv_broadcast(&as->a_cv);
2979 				}
2980 				AS_SETUNMAPWAIT(as);
2981 				AS_LOCK_EXIT(as);
2982 				while (AS_ISUNMAPWAIT(as)) {
2983 					cv_wait(&as->a_cv, &as->a_contents);
2984 				}
2985 			} else {
2986 				/*
2987 				 * We may have raced with
2988 				 * segvn_reclaim()/segspt_reclaim(). In this
2989 				 * case clean nounmapwait flag and retry since
2990 				 * softlockcnt in this segment may be already
2991 				 * 0.  We don't drop as writer lock so our
2992 				 * number of retries without sleeping should
2993 				 * be very small. See segvn_reclaim() for
2994 				 * more comments.
2995 				 */
2996 				AS_CLRNOUNMAPWAIT(as);
2997 				mutex_exit(&as->a_contents);
2998 				goto retry;
2999 			}
3000 			mutex_exit(&as->a_contents);
3001 			goto setpgsz_top;
3002 		} else if (error != 0) {
3003 			break;
3004 		}
3005 	}
3006 	as_setwatch(as);
3007 	AS_LOCK_EXIT(as);
3008 	return (error);
3009 }
3010 
3011 /*
3012  * as_iset3_default_lpsize() just calls SEGOP_SETPAGESIZE() on all segments
3013  * in its chunk where s_szc is less than the szc we want to set.
3014  */
3015 static int
as_iset3_default_lpsize(struct as * as,caddr_t raddr,size_t rsize,uint_t szc,int * retry)3016 as_iset3_default_lpsize(struct as *as, caddr_t raddr, size_t rsize, uint_t szc,
3017     int *retry)
3018 {
3019 	struct seg *seg;
3020 	size_t ssize;
3021 	int error;
3022 
3023 	ASSERT(AS_WRITE_HELD(as));
3024 
3025 	seg = as_segat(as, raddr);
3026 	if (seg == NULL) {
3027 		panic("as_iset3_default_lpsize: no seg");
3028 	}
3029 
3030 	for (; rsize != 0; rsize -= ssize, raddr += ssize) {
3031 		if (raddr >= seg->s_base + seg->s_size) {
3032 			seg = AS_SEGNEXT(as, seg);
3033 			if (seg == NULL || raddr != seg->s_base) {
3034 				panic("as_iset3_default_lpsize: as changed");
3035 			}
3036 		}
3037 		if ((raddr + rsize) > (seg->s_base + seg->s_size)) {
3038 			ssize = seg->s_base + seg->s_size - raddr;
3039 		} else {
3040 			ssize = rsize;
3041 		}
3042 
3043 		if (szc > seg->s_szc) {
3044 			error = SEGOP_SETPAGESIZE(seg, raddr, ssize, szc);
3045 			/* Only retry on EINVAL segments that have no vnode. */
3046 			if (error == EINVAL) {
3047 				vnode_t *vp = NULL;
3048 				if ((SEGOP_GETTYPE(seg, raddr) & MAP_SHARED) &&
3049 				    (SEGOP_GETVP(seg, raddr, &vp) != 0 ||
3050 				    vp == NULL)) {
3051 					*retry = 1;
3052 				} else {
3053 					*retry = 0;
3054 				}
3055 			}
3056 			if (error) {
3057 				return (error);
3058 			}
3059 		}
3060 	}
3061 	return (0);
3062 }
3063 
3064 /*
3065  * as_iset2_default_lpsize() calls as_iset3_default_lpsize() to set the
3066  * pagesize on each segment in its range, but if any fails with EINVAL,
3067  * then it reduces the pagesizes to the next size in the bitmap and
3068  * retries as_iset3_default_lpsize(). The reason why the code retries
3069  * smaller allowed sizes on EINVAL is because (a) the anon offset may not
3070  * match the bigger sizes, and (b) it's hard to get this offset (to begin
3071  * with) to pass to map_pgszcvec().
3072  */
3073 static int
as_iset2_default_lpsize(struct as * as,caddr_t addr,size_t size,uint_t szc,uint_t szcvec)3074 as_iset2_default_lpsize(struct as *as, caddr_t addr, size_t size, uint_t szc,
3075     uint_t szcvec)
3076 {
3077 	int error;
3078 	int retry;
3079 
3080 	ASSERT(AS_WRITE_HELD(as));
3081 
3082 	for (;;) {
3083 		error = as_iset3_default_lpsize(as, addr, size, szc, &retry);
3084 		if (error == EINVAL && retry) {
3085 			szcvec &= ~(1 << szc);
3086 			if (szcvec <= 1) {
3087 				return (EINVAL);
3088 			}
3089 			szc = highbit(szcvec) - 1;
3090 		} else {
3091 			return (error);
3092 		}
3093 	}
3094 }
3095 
3096 /*
3097  * as_iset1_default_lpsize() breaks its chunk into areas where existing
3098  * segments have a smaller szc than we want to set. For each such area,
3099  * it calls as_iset2_default_lpsize()
3100  */
3101 static int
as_iset1_default_lpsize(struct as * as,caddr_t raddr,size_t rsize,uint_t szc,uint_t szcvec)3102 as_iset1_default_lpsize(struct as *as, caddr_t raddr, size_t rsize, uint_t szc,
3103     uint_t szcvec)
3104 {
3105 	struct seg *seg;
3106 	size_t ssize;
3107 	caddr_t setaddr = raddr;
3108 	size_t setsize = 0;
3109 	int set;
3110 	int error;
3111 
3112 	ASSERT(AS_WRITE_HELD(as));
3113 
3114 	seg = as_segat(as, raddr);
3115 	if (seg == NULL) {
3116 		panic("as_iset1_default_lpsize: no seg");
3117 	}
3118 	if (seg->s_szc < szc) {
3119 		set = 1;
3120 	} else {
3121 		set = 0;
3122 	}
3123 
3124 	for (; rsize != 0; rsize -= ssize, raddr += ssize, setsize += ssize) {
3125 		if (raddr >= seg->s_base + seg->s_size) {
3126 			seg = AS_SEGNEXT(as, seg);
3127 			if (seg == NULL || raddr != seg->s_base) {
3128 				panic("as_iset1_default_lpsize: as changed");
3129 			}
3130 			if (seg->s_szc >= szc && set) {
3131 				ASSERT(setsize != 0);
3132 				error = as_iset2_default_lpsize(as,
3133 				    setaddr, setsize, szc, szcvec);
3134 				if (error) {
3135 					return (error);
3136 				}
3137 				set = 0;
3138 			} else if (seg->s_szc < szc && !set) {
3139 				setaddr = raddr;
3140 				setsize = 0;
3141 				set = 1;
3142 			}
3143 		}
3144 		if ((raddr + rsize) > (seg->s_base + seg->s_size)) {
3145 			ssize = seg->s_base + seg->s_size - raddr;
3146 		} else {
3147 			ssize = rsize;
3148 		}
3149 	}
3150 	error = 0;
3151 	if (set) {
3152 		ASSERT(setsize != 0);
3153 		error = as_iset2_default_lpsize(as, setaddr, setsize,
3154 		    szc, szcvec);
3155 	}
3156 	return (error);
3157 }
3158 
3159 /*
3160  * as_iset_default_lpsize() breaks its chunk according to the size code bitmap
3161  * returned by map_pgszcvec() (similar to as_map_segvn_segs()), and passes each
3162  * chunk to as_iset1_default_lpsize().
3163  */
3164 static int
as_iset_default_lpsize(struct as * as,caddr_t addr,size_t size,int flags,int type)3165 as_iset_default_lpsize(struct as *as, caddr_t addr, size_t size, int flags,
3166     int type)
3167 {
3168 	int rtype = (type & MAP_SHARED) ? MAPPGSZC_SHM : MAPPGSZC_PRIVM;
3169 	uint_t szcvec = map_pgszcvec(addr, size, (uintptr_t)addr,
3170 	    flags, rtype, 1);
3171 	uint_t szc;
3172 	uint_t nszc;
3173 	int error;
3174 	caddr_t a;
3175 	caddr_t eaddr;
3176 	size_t segsize;
3177 	size_t pgsz;
3178 	uint_t save_szcvec;
3179 
3180 	ASSERT(AS_WRITE_HELD(as));
3181 	ASSERT(IS_P2ALIGNED(addr, PAGESIZE));
3182 	ASSERT(IS_P2ALIGNED(size, PAGESIZE));
3183 
3184 	szcvec &= ~1;
3185 	if (szcvec <= 1) {	/* skip if base page size */
3186 		return (0);
3187 	}
3188 
3189 	/* Get the pagesize of the first larger page size. */
3190 	szc = lowbit(szcvec) - 1;
3191 	pgsz = page_get_pagesize(szc);
3192 	eaddr = addr + size;
3193 	addr = (caddr_t)P2ROUNDUP((uintptr_t)addr, pgsz);
3194 	eaddr = (caddr_t)P2ALIGN((uintptr_t)eaddr, pgsz);
3195 
3196 	save_szcvec = szcvec;
3197 	szcvec >>= (szc + 1);
3198 	nszc = szc;
3199 	while (szcvec) {
3200 		if ((szcvec & 0x1) == 0) {
3201 			nszc++;
3202 			szcvec >>= 1;
3203 			continue;
3204 		}
3205 		nszc++;
3206 		pgsz = page_get_pagesize(nszc);
3207 		a = (caddr_t)P2ROUNDUP((uintptr_t)addr, pgsz);
3208 		if (a != addr) {
3209 			ASSERT(szc > 0);
3210 			ASSERT(a < eaddr);
3211 			segsize = a - addr;
3212 			error = as_iset1_default_lpsize(as, addr, segsize, szc,
3213 			    save_szcvec);
3214 			if (error) {
3215 				return (error);
3216 			}
3217 			addr = a;
3218 		}
3219 		szc = nszc;
3220 		szcvec >>= 1;
3221 	}
3222 
3223 	ASSERT(addr < eaddr);
3224 	szcvec = save_szcvec;
3225 	while (szcvec) {
3226 		a = (caddr_t)P2ALIGN((uintptr_t)eaddr, pgsz);
3227 		ASSERT(a >= addr);
3228 		if (a != addr) {
3229 			ASSERT(szc > 0);
3230 			segsize = a - addr;
3231 			error = as_iset1_default_lpsize(as, addr, segsize, szc,
3232 			    save_szcvec);
3233 			if (error) {
3234 				return (error);
3235 			}
3236 			addr = a;
3237 		}
3238 		szcvec &= ~(1 << szc);
3239 		if (szcvec) {
3240 			szc = highbit(szcvec) - 1;
3241 			pgsz = page_get_pagesize(szc);
3242 		}
3243 	}
3244 	ASSERT(addr == eaddr);
3245 
3246 	return (0);
3247 }
3248 
3249 /*
3250  * Set the default large page size for the range. Called via memcntl with
3251  * page size set to 0. as_set_default_lpsize breaks the range down into
3252  * chunks with the same type/flags, ignores-non segvn segments, and passes
3253  * each chunk to as_iset_default_lpsize().
3254  */
3255 int
as_set_default_lpsize(struct as * as,caddr_t addr,size_t size)3256 as_set_default_lpsize(struct as *as, caddr_t addr, size_t size)
3257 {
3258 	struct seg *seg;
3259 	caddr_t raddr;
3260 	size_t rsize;
3261 	size_t ssize;
3262 	int rtype, rflags;
3263 	int stype, sflags;
3264 	int error;
3265 	caddr_t	setaddr;
3266 	size_t setsize;
3267 	int segvn;
3268 
3269 	if (size == 0)
3270 		return (0);
3271 
3272 	AS_LOCK_ENTER(as, RW_WRITER);
3273 again:
3274 	error = 0;
3275 
3276 	raddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
3277 	rsize = (((size_t)(addr + size) + PAGEOFFSET) & PAGEMASK) -
3278 	    (size_t)raddr;
3279 
3280 	if (raddr + rsize < raddr) {		/* check for wraparound */
3281 		AS_LOCK_EXIT(as);
3282 		return (ENOMEM);
3283 	}
3284 	as_clearwatchprot(as, raddr, rsize);
3285 	seg = as_segat(as, raddr);
3286 	if (seg == NULL) {
3287 		as_setwatch(as);
3288 		AS_LOCK_EXIT(as);
3289 		return (ENOMEM);
3290 	}
3291 	if (seg->s_ops == &segvn_ops) {
3292 		rtype = SEGOP_GETTYPE(seg, addr);
3293 		rflags = rtype & (MAP_TEXT | MAP_INITDATA);
3294 		rtype = rtype & (MAP_SHARED | MAP_PRIVATE);
3295 		segvn = 1;
3296 	} else {
3297 		segvn = 0;
3298 	}
3299 	setaddr = raddr;
3300 	setsize = 0;
3301 
3302 	for (; rsize != 0; rsize -= ssize, raddr += ssize, setsize += ssize) {
3303 		if (raddr >= (seg->s_base + seg->s_size)) {
3304 			seg = AS_SEGNEXT(as, seg);
3305 			if (seg == NULL || raddr != seg->s_base) {
3306 				error = ENOMEM;
3307 				break;
3308 			}
3309 			if (seg->s_ops == &segvn_ops) {
3310 				stype = SEGOP_GETTYPE(seg, raddr);
3311 				sflags = stype & (MAP_TEXT | MAP_INITDATA);
3312 				stype &= (MAP_SHARED | MAP_PRIVATE);
3313 				if (segvn && (rflags != sflags ||
3314 				    rtype != stype)) {
3315 					/*
3316 					 * The next segment is also segvn but
3317 					 * has different flags and/or type.
3318 					 */
3319 					ASSERT(setsize != 0);
3320 					error = as_iset_default_lpsize(as,
3321 					    setaddr, setsize, rflags, rtype);
3322 					if (error) {
3323 						break;
3324 					}
3325 					rflags = sflags;
3326 					rtype = stype;
3327 					setaddr = raddr;
3328 					setsize = 0;
3329 				} else if (!segvn) {
3330 					rflags = sflags;
3331 					rtype = stype;
3332 					setaddr = raddr;
3333 					setsize = 0;
3334 					segvn = 1;
3335 				}
3336 			} else if (segvn) {
3337 				/* The next segment is not segvn. */
3338 				ASSERT(setsize != 0);
3339 				error = as_iset_default_lpsize(as,
3340 				    setaddr, setsize, rflags, rtype);
3341 				if (error) {
3342 					break;
3343 				}
3344 				segvn = 0;
3345 			}
3346 		}
3347 		if ((raddr + rsize) > (seg->s_base + seg->s_size)) {
3348 			ssize = seg->s_base + seg->s_size - raddr;
3349 		} else {
3350 			ssize = rsize;
3351 		}
3352 	}
3353 	if (error == 0 && segvn) {
3354 		/* The last chunk when rsize == 0. */
3355 		ASSERT(setsize != 0);
3356 		error = as_iset_default_lpsize(as, setaddr, setsize,
3357 		    rflags, rtype);
3358 	}
3359 
3360 	if (error == IE_RETRY) {
3361 		goto again;
3362 	} else if (error == IE_NOMEM) {
3363 		error = EAGAIN;
3364 	} else if (error == ENOTSUP) {
3365 		error = EINVAL;
3366 	} else if (error == EAGAIN) {
3367 		mutex_enter(&as->a_contents);
3368 		if (!AS_ISNOUNMAPWAIT(as)) {
3369 			if (AS_ISUNMAPWAIT(as) == 0) {
3370 				cv_broadcast(&as->a_cv);
3371 			}
3372 			AS_SETUNMAPWAIT(as);
3373 			AS_LOCK_EXIT(as);
3374 			while (AS_ISUNMAPWAIT(as)) {
3375 				cv_wait(&as->a_cv, &as->a_contents);
3376 			}
3377 			mutex_exit(&as->a_contents);
3378 			AS_LOCK_ENTER(as, RW_WRITER);
3379 		} else {
3380 			/*
3381 			 * We may have raced with
3382 			 * segvn_reclaim()/segspt_reclaim(). In this case
3383 			 * clean nounmapwait flag and retry since softlockcnt
3384 			 * in this segment may be already 0.  We don't drop as
3385 			 * writer lock so our number of retries without
3386 			 * sleeping should be very small. See segvn_reclaim()
3387 			 * for more comments.
3388 			 */
3389 			AS_CLRNOUNMAPWAIT(as);
3390 			mutex_exit(&as->a_contents);
3391 		}
3392 		goto again;
3393 	}
3394 
3395 	as_setwatch(as);
3396 	AS_LOCK_EXIT(as);
3397 	return (error);
3398 }
3399 
3400 /*
3401  * Setup all of the uninitialized watched pages that we can.
3402  */
3403 void
as_setwatch(struct as * as)3404 as_setwatch(struct as *as)
3405 {
3406 	struct watched_page *pwp;
3407 	struct seg *seg;
3408 	caddr_t vaddr;
3409 	uint_t prot;
3410 	int  err, retrycnt;
3411 
3412 	if (avl_numnodes(&as->a_wpage) == 0)
3413 		return;
3414 
3415 	ASSERT(AS_WRITE_HELD(as));
3416 
3417 	for (pwp = avl_first(&as->a_wpage); pwp != NULL;
3418 	    pwp = AVL_NEXT(&as->a_wpage, pwp)) {
3419 		retrycnt = 0;
3420 	retry:
3421 		vaddr = pwp->wp_vaddr;
3422 		if (pwp->wp_oprot != 0 ||	/* already set up */
3423 		    (seg = as_segat(as, vaddr)) == NULL ||
3424 		    SEGOP_GETPROT(seg, vaddr, 0, &prot) != 0)
3425 			continue;
3426 
3427 		pwp->wp_oprot = prot;
3428 		if (pwp->wp_read)
3429 			prot &= ~(PROT_READ|PROT_WRITE|PROT_EXEC);
3430 		if (pwp->wp_write)
3431 			prot &= ~PROT_WRITE;
3432 		if (pwp->wp_exec)
3433 			prot &= ~(PROT_READ|PROT_WRITE|PROT_EXEC);
3434 		if (!(pwp->wp_flags & WP_NOWATCH) && prot != pwp->wp_oprot) {
3435 			err = SEGOP_SETPROT(seg, vaddr, PAGESIZE, prot);
3436 			if (err == IE_RETRY) {
3437 				pwp->wp_oprot = 0;
3438 				ASSERT(retrycnt == 0);
3439 				retrycnt++;
3440 				goto retry;
3441 			}
3442 		}
3443 		pwp->wp_prot = prot;
3444 	}
3445 }
3446 
3447 /*
3448  * Clear all of the watched pages in the address space.
3449  */
3450 void
as_clearwatch(struct as * as)3451 as_clearwatch(struct as *as)
3452 {
3453 	struct watched_page *pwp;
3454 	struct seg *seg;
3455 	caddr_t vaddr;
3456 	uint_t prot;
3457 	int err, retrycnt;
3458 
3459 	if (avl_numnodes(&as->a_wpage) == 0)
3460 		return;
3461 
3462 	ASSERT(AS_WRITE_HELD(as));
3463 
3464 	for (pwp = avl_first(&as->a_wpage); pwp != NULL;
3465 	    pwp = AVL_NEXT(&as->a_wpage, pwp)) {
3466 		retrycnt = 0;
3467 	retry:
3468 		vaddr = pwp->wp_vaddr;
3469 		if (pwp->wp_oprot == 0 ||	/* not set up */
3470 		    (seg = as_segat(as, vaddr)) == NULL)
3471 			continue;
3472 
3473 		if ((prot = pwp->wp_oprot) != pwp->wp_prot) {
3474 			err = SEGOP_SETPROT(seg, vaddr, PAGESIZE, prot);
3475 			if (err == IE_RETRY) {
3476 				ASSERT(retrycnt == 0);
3477 				retrycnt++;
3478 				goto retry;
3479 			}
3480 		}
3481 		pwp->wp_oprot = 0;
3482 		pwp->wp_prot = 0;
3483 	}
3484 }
3485 
3486 /*
3487  * Force a new setup for all the watched pages in the range.
3488  */
3489 static void
as_setwatchprot(struct as * as,caddr_t addr,size_t size,uint_t prot)3490 as_setwatchprot(struct as *as, caddr_t addr, size_t size, uint_t prot)
3491 {
3492 	struct watched_page *pwp;
3493 	struct watched_page tpw;
3494 	caddr_t eaddr = addr + size;
3495 	caddr_t vaddr;
3496 	struct seg *seg;
3497 	int err, retrycnt;
3498 	uint_t	wprot;
3499 	avl_index_t where;
3500 
3501 	if (avl_numnodes(&as->a_wpage) == 0)
3502 		return;
3503 
3504 	ASSERT(AS_WRITE_HELD(as));
3505 
3506 	tpw.wp_vaddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
3507 	if ((pwp = avl_find(&as->a_wpage, &tpw, &where)) == NULL)
3508 		pwp = avl_nearest(&as->a_wpage, where, AVL_AFTER);
3509 
3510 	while (pwp != NULL && pwp->wp_vaddr < eaddr) {
3511 		retrycnt = 0;
3512 		vaddr = pwp->wp_vaddr;
3513 
3514 		wprot = prot;
3515 		if (pwp->wp_read)
3516 			wprot &= ~(PROT_READ|PROT_WRITE|PROT_EXEC);
3517 		if (pwp->wp_write)
3518 			wprot &= ~PROT_WRITE;
3519 		if (pwp->wp_exec)
3520 			wprot &= ~(PROT_READ|PROT_WRITE|PROT_EXEC);
3521 		if (!(pwp->wp_flags & WP_NOWATCH) && wprot != pwp->wp_oprot) {
3522 		retry:
3523 			seg = as_segat(as, vaddr);
3524 			if (seg == NULL) {
3525 				panic("as_setwatchprot: no seg");
3526 				/*NOTREACHED*/
3527 			}
3528 			err = SEGOP_SETPROT(seg, vaddr, PAGESIZE, wprot);
3529 			if (err == IE_RETRY) {
3530 				ASSERT(retrycnt == 0);
3531 				retrycnt++;
3532 				goto retry;
3533 			}
3534 		}
3535 		pwp->wp_oprot = prot;
3536 		pwp->wp_prot = wprot;
3537 
3538 		pwp = AVL_NEXT(&as->a_wpage, pwp);
3539 	}
3540 }
3541 
3542 /*
3543  * Clear all of the watched pages in the range.
3544  */
3545 static void
as_clearwatchprot(struct as * as,caddr_t addr,size_t size)3546 as_clearwatchprot(struct as *as, caddr_t addr, size_t size)
3547 {
3548 	caddr_t eaddr = addr + size;
3549 	struct watched_page *pwp;
3550 	struct watched_page tpw;
3551 	uint_t prot;
3552 	struct seg *seg;
3553 	int err, retrycnt;
3554 	avl_index_t where;
3555 
3556 	if (avl_numnodes(&as->a_wpage) == 0)
3557 		return;
3558 
3559 	tpw.wp_vaddr = (caddr_t)((uintptr_t)addr & (uintptr_t)PAGEMASK);
3560 	if ((pwp = avl_find(&as->a_wpage, &tpw, &where)) == NULL)
3561 		pwp = avl_nearest(&as->a_wpage, where, AVL_AFTER);
3562 
3563 	ASSERT(AS_WRITE_HELD(as));
3564 
3565 	while (pwp != NULL && pwp->wp_vaddr < eaddr) {
3566 
3567 		if ((prot = pwp->wp_oprot) != 0) {
3568 			retrycnt = 0;
3569 
3570 			if (prot != pwp->wp_prot) {
3571 			retry:
3572 				seg = as_segat(as, pwp->wp_vaddr);
3573 				if (seg == NULL)
3574 					continue;
3575 				err = SEGOP_SETPROT(seg, pwp->wp_vaddr,
3576 				    PAGESIZE, prot);
3577 				if (err == IE_RETRY) {
3578 					ASSERT(retrycnt == 0);
3579 					retrycnt++;
3580 					goto retry;
3581 
3582 				}
3583 			}
3584 			pwp->wp_oprot = 0;
3585 			pwp->wp_prot = 0;
3586 		}
3587 
3588 		pwp = AVL_NEXT(&as->a_wpage, pwp);
3589 	}
3590 }
3591 
3592 void
as_signal_proc(struct as * as,k_siginfo_t * siginfo)3593 as_signal_proc(struct as *as, k_siginfo_t *siginfo)
3594 {
3595 	struct proc *p;
3596 
3597 	mutex_enter(&pidlock);
3598 	for (p = practive; p; p = p->p_next) {
3599 		if (p->p_as == as) {
3600 			mutex_enter(&p->p_lock);
3601 			if (p->p_as == as)
3602 				sigaddq(p, NULL, siginfo, KM_NOSLEEP);
3603 			mutex_exit(&p->p_lock);
3604 		}
3605 	}
3606 	mutex_exit(&pidlock);
3607 }
3608 
3609 /*
3610  * return memory object ID
3611  */
3612 int
as_getmemid(struct as * as,caddr_t addr,memid_t * memidp)3613 as_getmemid(struct as *as, caddr_t addr, memid_t *memidp)
3614 {
3615 	struct seg	*seg;
3616 	int		sts;
3617 
3618 	AS_LOCK_ENTER(as, RW_READER);
3619 	seg = as_segat(as, addr);
3620 	if (seg == NULL) {
3621 		AS_LOCK_EXIT(as);
3622 		return (EFAULT);
3623 	}
3624 	/*
3625 	 * catch old drivers which may not support getmemid
3626 	 */
3627 	if (seg->s_ops->getmemid == NULL) {
3628 		AS_LOCK_EXIT(as);
3629 		return (ENODEV);
3630 	}
3631 
3632 	sts = SEGOP_GETMEMID(seg, addr, memidp);
3633 
3634 	AS_LOCK_EXIT(as);
3635 	return (sts);
3636 }
3637