kernel - Introduce lightweight buffers
[dragonfly.git] / sys / vm / vm_fault.c
CommitLineData
984263bc
MD
1/*
2 * Copyright (c) 1991, 1993
3 * The Regents of the University of California. All rights reserved.
4 * Copyright (c) 1994 John S. Dyson
5 * All rights reserved.
6 * Copyright (c) 1994 David Greenman
7 * All rights reserved.
8 *
9 *
10 * This code is derived from software contributed to Berkeley by
11 * The Mach Operating System project at Carnegie-Mellon University.
12 *
13 * Redistribution and use in source and binary forms, with or without
14 * modification, are permitted provided that the following conditions
15 * are met:
16 * 1. Redistributions of source code must retain the above copyright
17 * notice, this list of conditions and the following disclaimer.
18 * 2. Redistributions in binary form must reproduce the above copyright
19 * notice, this list of conditions and the following disclaimer in the
20 * documentation and/or other materials provided with the distribution.
21 * 3. All advertising materials mentioning features or use of this software
22 * must display the following acknowledgement:
23 * This product includes software developed by the University of
24 * California, Berkeley and its contributors.
25 * 4. Neither the name of the University nor the names of its contributors
26 * may be used to endorse or promote products derived from this software
27 * without specific prior written permission.
28 *
29 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
30 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
31 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
32 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
33 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
34 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
35 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
36 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
37 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
38 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
39 * SUCH DAMAGE.
40 *
41 * from: @(#)vm_fault.c 8.4 (Berkeley) 1/12/94
42 *
43 *
44 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
45 * All rights reserved.
46 *
47 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
48 *
49 * Permission to use, copy, modify and distribute this software and
50 * its documentation is hereby granted, provided that both the copyright
51 * notice and this permission notice appear in all copies of the
52 * software, derivative works or modified versions, and any portions
53 * thereof, and that both notices appear in supporting documentation.
54 *
55 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
56 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
57 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
58 *
59 * Carnegie Mellon requests users of this software to return to
60 *
61 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
62 * School of Computer Science
63 * Carnegie Mellon University
64 * Pittsburgh PA 15213-3890
65 *
66 * any improvements or extensions that they make and grant Carnegie the
67 * rights to redistribute these changes.
68 *
69 * $FreeBSD: src/sys/vm/vm_fault.c,v 1.108.2.8 2002/02/26 05:49:27 silby Exp $
4ecf7cc9 70 * $DragonFly: src/sys/vm/vm_fault.c,v 1.47 2008/07/01 02:02:56 dillon Exp $
984263bc
MD
71 */
72
73/*
74 * Page fault handling module.
75 */
76
77#include <sys/param.h>
78#include <sys/systm.h>
46311ac2 79#include <sys/kernel.h>
984263bc
MD
80#include <sys/proc.h>
81#include <sys/vnode.h>
82#include <sys/resourcevar.h>
83#include <sys/vmmeter.h>
75f59a66 84#include <sys/vkernel.h>
75f59a66 85#include <sys/lock.h>
bc823b32 86#include <sys/sysctl.h>
984263bc 87
5c5185ae
SG
88#include <cpu/lwbuf.h>
89
984263bc
MD
90#include <vm/vm.h>
91#include <vm/vm_param.h>
984263bc
MD
92#include <vm/pmap.h>
93#include <vm/vm_map.h>
94#include <vm/vm_object.h>
95#include <vm/vm_page.h>
96#include <vm/vm_pageout.h>
97#include <vm/vm_kern.h>
98#include <vm/vm_pager.h>
99#include <vm/vnode_pager.h>
100#include <vm/vm_extern.h>
654a39f0
MD
101
102#include <sys/thread2.h>
12e4aaff 103#include <vm/vm_page2.h>
984263bc 104
984263bc
MD
105struct faultstate {
106 vm_page_t m;
107 vm_object_t object;
108 vm_pindex_t pindex;
72579d2e 109 vm_prot_t prot;
984263bc 110 vm_page_t first_m;
568e6804 111 vm_object_t first_object;
72579d2e 112 vm_prot_t first_prot;
984263bc
MD
113 vm_map_t map;
114 vm_map_entry_t entry;
115 int lookup_still_valid;
568e6804
MD
116 int didlimit;
117 int hardfault;
568e6804
MD
118 int fault_flags;
119 int map_generation;
120 boolean_t wired;
984263bc
MD
121 struct vnode *vp;
122};
123
1b9d3514
MD
124static int vm_fast_fault = 1;
125SYSCTL_INT(_vm, OID_AUTO, fast_fault, CTLFLAG_RW, &vm_fast_fault, 0, "");
cf1bb2a8
MD
126static int debug_cluster = 0;
127SYSCTL_INT(_vm, OID_AUTO, debug_cluster, CTLFLAG_RW, &debug_cluster, 0, "");
bc823b32 128
72579d2e 129static int vm_fault_object(struct faultstate *, vm_pindex_t, vm_prot_t);
4e7c41c5 130static int vm_fault_vpagetable(struct faultstate *, vm_pindex_t *, vpte_t, int);
1b9d3514 131#if 0
568e6804 132static int vm_fault_additional_pages (vm_page_t, int, int, vm_page_t *, int *);
1b9d3514 133#endif
72579d2e 134static int vm_fault_ratelimit(struct vmspace *);
1b9d3514
MD
135static void vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry,
136 int prot);
568e6804 137
984263bc
MD
138static __inline void
139release_page(struct faultstate *fs)
140{
984263bc 141 vm_page_deactivate(fs->m);
17cde63e 142 vm_page_wakeup(fs->m);
984263bc
MD
143 fs->m = NULL;
144}
145
146static __inline void
147unlock_map(struct faultstate *fs)
148{
aa542ad5 149 if (fs->lookup_still_valid && fs->map) {
a108bf71 150 vm_map_lookup_done(fs->map, fs->entry, 0);
984263bc
MD
151 fs->lookup_still_valid = FALSE;
152 }
153}
154
75f59a66
MD
155/*
156 * Clean up after a successful call to vm_fault_object() so another call
157 * to vm_fault_object() can be made.
158 */
984263bc 159static void
75f59a66 160_cleanup_successful_fault(struct faultstate *fs, int relock)
984263bc 161{
984263bc
MD
162 if (fs->object != fs->first_object) {
163 vm_page_free(fs->first_m);
75f59a66 164 vm_object_pip_wakeup(fs->object);
984263bc
MD
165 fs->first_m = NULL;
166 }
75f59a66
MD
167 fs->object = fs->first_object;
168 if (relock && fs->lookup_still_valid == FALSE) {
aa542ad5
MD
169 if (fs->map)
170 vm_map_lock_read(fs->map);
75f59a66
MD
171 fs->lookup_still_valid = TRUE;
172 }
173}
174
175static void
176_unlock_things(struct faultstate *fs, int dealloc)
177{
178 vm_object_pip_wakeup(fs->first_object);
179 _cleanup_successful_fault(fs, 0);
984263bc
MD
180 if (dealloc) {
181 vm_object_deallocate(fs->first_object);
bc823b32 182 fs->first_object = NULL;
984263bc
MD
183 }
184 unlock_map(fs);
185 if (fs->vp != NULL) {
186 vput(fs->vp);
187 fs->vp = NULL;
188 }
189}
190
191#define unlock_things(fs) _unlock_things(fs, 0)
192#define unlock_and_deallocate(fs) _unlock_things(fs, 1)
75f59a66 193#define cleanup_successful_fault(fs) _cleanup_successful_fault(fs, 1)
984263bc
MD
194
195/*
568e6804
MD
196 * TRYPAGER
197 *
198 * Determine if the pager for the current object *might* contain the page.
984263bc 199 *
568e6804
MD
200 * We only need to try the pager if this is not a default object (default
201 * objects are zero-fill and have no real pager), and if we are not taking
202 * a wiring fault or if the FS entry is wired.
984263bc 203 */
568e6804
MD
204#define TRYPAGER(fs) \
205 (fs->object->type != OBJT_DEFAULT && \
206 (((fs->fault_flags & VM_FAULT_WIRE_MASK) == 0) || fs->wired))
984263bc
MD
207
208/*
568e6804 209 * vm_fault:
984263bc 210 *
568e6804
MD
211 * Handle a page fault occuring at the given address, requiring the given
212 * permissions, in the map specified. If successful, the page is inserted
213 * into the associated physical map.
984263bc 214 *
568e6804 215 * NOTE: The given address should be truncated to the proper page address.
984263bc 216 *
568e6804
MD
217 * KERN_SUCCESS is returned if the page fault is handled; otherwise,
218 * a standard error specifying why the fault is fatal is returned.
984263bc 219 *
568e6804
MD
220 * The map in question must be referenced, and remains so.
221 * The caller may hold no locks.
984263bc
MD
222 */
223int
224vm_fault(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type, int fault_flags)
225{
984263bc 226 int result;
72579d2e 227 vm_pindex_t first_pindex;
984263bc 228 struct faultstate fs;
8d496bf9 229 int growstack;
984263bc 230
12e4aaff 231 mycpu->gd_cnt.v_vm_faults++;
568e6804
MD
232
233 fs.didlimit = 0;
234 fs.hardfault = 0;
235 fs.fault_flags = fault_flags;
8d496bf9 236 growstack = 1;
984263bc 237
06ecca5a 238RetryFault:
984263bc 239 /*
568e6804
MD
240 * Find the vm_map_entry representing the backing store and resolve
241 * the top level object and page index. This may have the side
242 * effect of executing a copy-on-write on the map entry and/or
243 * creating a shadow object, but will not COW any actual VM pages.
244 *
245 * On success fs.map is left read-locked and various other fields
246 * are initialized but not otherwise referenced or locked.
247 *
4e7c41c5
MD
248 * NOTE! vm_map_lookup will try to upgrade the fault_type to
249 * VM_FAULT_WRITE if the map entry is a virtual page table and also
250 * writable, so we can set the 'A'accessed bit in the virtual page
251 * table entry.
984263bc
MD
252 */
253 fs.map = map;
568e6804
MD
254 result = vm_map_lookup(&fs.map, vaddr, fault_type,
255 &fs.entry, &fs.first_object,
72579d2e 256 &first_pindex, &fs.first_prot, &fs.wired);
568e6804
MD
257
258 /*
259 * If the lookup failed or the map protections are incompatible,
260 * the fault generally fails. However, if the caller is trying
261 * to do a user wiring we have more work to do.
262 */
263 if (result != KERN_SUCCESS) {
8d496bf9
MD
264 if (result != KERN_PROTECTION_FAILURE ||
265 (fs.fault_flags & VM_FAULT_WIRE_MASK) != VM_FAULT_USER_WIRE)
266 {
267 if (result == KERN_INVALID_ADDRESS && growstack &&
268 map != &kernel_map && curproc != NULL) {
269 result = vm_map_growstack(curproc, vaddr);
270 if (result != KERN_SUCCESS)
271 return (KERN_FAILURE);
272 growstack = 0;
273 goto RetryFault;
274 }
275 return (result);
276 }
984263bc
MD
277
278 /*
279 * If we are user-wiring a r/w segment, and it is COW, then
568e6804
MD
280 * we need to do the COW operation. Note that we don't
281 * currently COW RO sections now, because it is NOT desirable
984263bc
MD
282 * to COW .text. We simply keep .text from ever being COW'ed
283 * and take the heat that one cannot debug wired .text sections.
284 */
285 result = vm_map_lookup(&fs.map, vaddr,
568e6804
MD
286 VM_PROT_READ|VM_PROT_WRITE|
287 VM_PROT_OVERRIDE_WRITE,
288 &fs.entry, &fs.first_object,
72579d2e
MD
289 &first_pindex, &fs.first_prot,
290 &fs.wired);
568e6804 291 if (result != KERN_SUCCESS)
984263bc 292 return result;
984263bc
MD
293
294 /*
295 * If we don't COW now, on a user wire, the user will never
296 * be able to write to the mapping. If we don't make this
297 * restriction, the bookkeeping would be nearly impossible.
298 */
299 if ((fs.entry->protection & VM_PROT_WRITE) == 0)
300 fs.entry->max_protection &= ~VM_PROT_WRITE;
301 }
302
568e6804
MD
303 /*
304 * fs.map is read-locked
305 *
306 * Misc checks. Save the map generation number to detect races.
307 */
308 fs.map_generation = fs.map->timestamp;
984263bc
MD
309
310 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
311 panic("vm_fault: fault on nofault entry, addr: %lx",
312 (u_long)vaddr);
313 }
314
315 /*
c40f2b75
MD
316 * A system map entry may return a NULL object. No object means
317 * no pager means an unrecoverable kernel fault.
318 */
319 if (fs.first_object == NULL) {
320 panic("vm_fault: unrecoverable fault at %p in entry %p",
321 (void *)vaddr, fs.entry);
322 }
323
324 /*
984263bc
MD
325 * Make a reference to this object to prevent its disposal while we
326 * are messing with it. Once we have the reference, the map is free
327 * to be diddled. Since objects reference their shadows (and copies),
328 * they will stay around as well.
329 *
330 * Bump the paging-in-progress count to prevent size changes (e.g.
331 * truncation operations) during I/O. This must be done after
332 * obtaining the vnode lock in order to avoid possible deadlocks.
333 */
334 vm_object_reference(fs.first_object);
335 fs.vp = vnode_pager_lock(fs.first_object);
336 vm_object_pip_add(fs.first_object, 1);
337
984263bc 338 fs.lookup_still_valid = TRUE;
984263bc 339 fs.first_m = NULL;
afeabdca 340 fs.object = fs.first_object; /* so unlock_and_deallocate works */
984263bc
MD
341
342 /*
568e6804 343 * If the entry is wired we cannot change the page protection.
984263bc 344 */
568e6804 345 if (fs.wired)
72579d2e 346 fault_type = fs.first_prot;
984263bc 347
568e6804 348 /*
75f59a66
MD
349 * The page we want is at (first_object, first_pindex), but if the
350 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
351 * page table to figure out the actual pindex.
352 *
353 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
354 * ONLY
568e6804 355 */
568e6804 356 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
72579d2e 357 result = vm_fault_vpagetable(&fs, &first_pindex,
4e7c41c5
MD
358 fs.entry->aux.master_pde,
359 fault_type);
568e6804
MD
360 if (result == KERN_TRY_AGAIN)
361 goto RetryFault;
75f59a66 362 if (result != KERN_SUCCESS)
568e6804 363 return (result);
568e6804 364 }
75f59a66 365
568e6804
MD
366 /*
367 * Now we have the actual (object, pindex), fault in the page. If
368 * vm_fault_object() fails it will unlock and deallocate the FS
75f59a66
MD
369 * data. If it succeeds everything remains locked and fs->object
370 * will have an additinal PIP count if it is not equal to
371 * fs->first_object
4e7c41c5
MD
372 *
373 * vm_fault_object will set fs->prot for the pmap operation. It is
374 * allowed to set VM_PROT_WRITE if fault_type == VM_PROT_READ if the
375 * page can be safely written. However, it will force a read-only
376 * mapping for a read fault if the memory is managed by a virtual
377 * page table.
568e6804 378 */
72579d2e 379 result = vm_fault_object(&fs, first_pindex, fault_type);
afeabdca 380
568e6804
MD
381 if (result == KERN_TRY_AGAIN)
382 goto RetryFault;
383 if (result != KERN_SUCCESS)
384 return (result);
385
386 /*
75f59a66
MD
387 * On success vm_fault_object() does not unlock or deallocate, and fs.m
388 * will contain a busied page.
568e6804
MD
389 *
390 * Enter the page into the pmap and do pmap-related adjustments.
391 */
392 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
393
1b9d3514
MD
394 /*
395 * Burst in a few more pages if possible. The fs.map should still
396 * be locked.
397 */
398 if (fault_flags & VM_FAULT_BURST) {
399 if ((fs.fault_flags & VM_FAULT_WIRE_MASK) == 0 &&
400 fs.wired == 0) {
401 vm_prefault(fs.map->pmap, vaddr, fs.entry, fs.prot);
402 }
568e6804 403 }
1b9d3514 404 unlock_things(&fs);
568e6804
MD
405
406 vm_page_flag_clear(fs.m, PG_ZERO);
17cde63e 407 vm_page_flag_set(fs.m, PG_REFERENCED);
568e6804
MD
408
409 /*
410 * If the page is not wired down, then put it where the pageout daemon
411 * can find it.
412 */
413 if (fs.fault_flags & VM_FAULT_WIRE_MASK) {
414 if (fs.wired)
415 vm_page_wire(fs.m);
416 else
417 vm_page_unwire(fs.m, 1);
418 } else {
419 vm_page_activate(fs.m);
420 }
421
fde7ac71 422 if (curthread->td_lwp) {
568e6804 423 if (fs.hardfault) {
fde7ac71 424 curthread->td_lwp->lwp_ru.ru_majflt++;
568e6804 425 } else {
fde7ac71 426 curthread->td_lwp->lwp_ru.ru_minflt++;
568e6804
MD
427 }
428 }
429
430 /*
431 * Unlock everything, and return
432 */
433 vm_page_wakeup(fs.m);
434 vm_object_deallocate(fs.first_object);
435
436 return (KERN_SUCCESS);
437}
438
439/*
5a0e2a66
MD
440 * Fault in the specified virtual address in the current process map,
441 * returning a held VM page or NULL. See vm_fault_page() for more
442 * information.
443 */
444vm_page_t
445vm_fault_page_quick(vm_offset_t va, vm_prot_t fault_type, int *errorp)
446{
287ebb09 447 struct lwp *lp = curthread->td_lwp;
5a0e2a66
MD
448 vm_page_t m;
449
287ebb09 450 m = vm_fault_page(&lp->lwp_vmspace->vm_map, va,
5a0e2a66
MD
451 fault_type, VM_FAULT_NORMAL, errorp);
452 return(m);
453}
454
455/*
456 * Fault in the specified virtual address in the specified map, doing all
4e158347
MD
457 * necessary manipulation of the object store and all necessary I/O. Return
458 * a held VM page or NULL, and set *errorp. The related pmap is not
459 * updated.
460 *
5a0e2a66
MD
461 * The returned page will be properly dirtied if VM_PROT_WRITE was specified,
462 * and marked PG_REFERENCED as well.
17cde63e
MD
463 *
464 * If the page cannot be faulted writable and VM_PROT_WRITE was specified, an
465 * error will be returned.
4e158347
MD
466 */
467vm_page_t
468vm_fault_page(vm_map_t map, vm_offset_t vaddr, vm_prot_t fault_type,
469 int fault_flags, int *errorp)
470{
4e158347
MD
471 vm_pindex_t first_pindex;
472 struct faultstate fs;
17cde63e
MD
473 int result;
474 vm_prot_t orig_fault_type = fault_type;
4e158347
MD
475
476 mycpu->gd_cnt.v_vm_faults++;
477
478 fs.didlimit = 0;
479 fs.hardfault = 0;
480 fs.fault_flags = fault_flags;
481 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
482
483RetryFault:
484 /*
485 * Find the vm_map_entry representing the backing store and resolve
486 * the top level object and page index. This may have the side
487 * effect of executing a copy-on-write on the map entry and/or
488 * creating a shadow object, but will not COW any actual VM pages.
489 *
490 * On success fs.map is left read-locked and various other fields
491 * are initialized but not otherwise referenced or locked.
492 *
493 * NOTE! vm_map_lookup will upgrade the fault_type to VM_FAULT_WRITE
494 * if the map entry is a virtual page table and also writable,
495 * so we can set the 'A'accessed bit in the virtual page table entry.
496 */
497 fs.map = map;
498 result = vm_map_lookup(&fs.map, vaddr, fault_type,
499 &fs.entry, &fs.first_object,
500 &first_pindex, &fs.first_prot, &fs.wired);
501
502 if (result != KERN_SUCCESS) {
503 *errorp = result;
504 return (NULL);
505 }
506
507 /*
508 * fs.map is read-locked
509 *
510 * Misc checks. Save the map generation number to detect races.
511 */
512 fs.map_generation = fs.map->timestamp;
513
514 if (fs.entry->eflags & MAP_ENTRY_NOFAULT) {
515 panic("vm_fault: fault on nofault entry, addr: %lx",
516 (u_long)vaddr);
517 }
518
519 /*
520 * A system map entry may return a NULL object. No object means
521 * no pager means an unrecoverable kernel fault.
522 */
523 if (fs.first_object == NULL) {
524 panic("vm_fault: unrecoverable fault at %p in entry %p",
525 (void *)vaddr, fs.entry);
526 }
527
528 /*
529 * Make a reference to this object to prevent its disposal while we
530 * are messing with it. Once we have the reference, the map is free
531 * to be diddled. Since objects reference their shadows (and copies),
532 * they will stay around as well.
533 *
534 * Bump the paging-in-progress count to prevent size changes (e.g.
535 * truncation operations) during I/O. This must be done after
536 * obtaining the vnode lock in order to avoid possible deadlocks.
537 */
538 vm_object_reference(fs.first_object);
539 fs.vp = vnode_pager_lock(fs.first_object);
540 vm_object_pip_add(fs.first_object, 1);
541
542 fs.lookup_still_valid = TRUE;
543 fs.first_m = NULL;
544 fs.object = fs.first_object; /* so unlock_and_deallocate works */
545
546 /*
547 * If the entry is wired we cannot change the page protection.
548 */
549 if (fs.wired)
550 fault_type = fs.first_prot;
551
552 /*
553 * The page we want is at (first_object, first_pindex), but if the
554 * vm_map_entry is VM_MAPTYPE_VPAGETABLE we have to traverse the
555 * page table to figure out the actual pindex.
556 *
557 * NOTE! DEVELOPMENT IN PROGRESS, THIS IS AN INITIAL IMPLEMENTATION
558 * ONLY
559 */
560 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
561 result = vm_fault_vpagetable(&fs, &first_pindex,
4e7c41c5
MD
562 fs.entry->aux.master_pde,
563 fault_type);
4e158347
MD
564 if (result == KERN_TRY_AGAIN)
565 goto RetryFault;
566 if (result != KERN_SUCCESS) {
567 *errorp = result;
568 return (NULL);
569 }
570 }
571
572 /*
573 * Now we have the actual (object, pindex), fault in the page. If
574 * vm_fault_object() fails it will unlock and deallocate the FS
575 * data. If it succeeds everything remains locked and fs->object
576 * will have an additinal PIP count if it is not equal to
577 * fs->first_object
578 */
579 result = vm_fault_object(&fs, first_pindex, fault_type);
580
581 if (result == KERN_TRY_AGAIN)
582 goto RetryFault;
583 if (result != KERN_SUCCESS) {
584 *errorp = result;
585 return(NULL);
586 }
587
17cde63e
MD
588 if ((orig_fault_type & VM_PROT_WRITE) &&
589 (fs.prot & VM_PROT_WRITE) == 0) {
590 *errorp = KERN_PROTECTION_FAILURE;
591 unlock_and_deallocate(&fs);
592 return(NULL);
593 }
594
4e158347
MD
595 /*
596 * On success vm_fault_object() does not unlock or deallocate, and fs.m
597 * will contain a busied page.
598 */
599 unlock_things(&fs);
600
601 /*
602 * Return a held page. We are not doing any pmap manipulation so do
5a0e2a66
MD
603 * not set PG_MAPPED. However, adjust the page flags according to
604 * the fault type because the caller may not use a managed pmapping
605 * (so we don't want to lose the fact that the page will be dirtied
606 * if a write fault was specified).
4e158347 607 */
5a0e2a66 608 vm_page_hold(fs.m);
4e158347 609 vm_page_flag_clear(fs.m, PG_ZERO);
5a0e2a66
MD
610 if (fault_type & VM_PROT_WRITE)
611 vm_page_dirty(fs.m);
4e158347
MD
612
613 /*
40d6ef3a
MD
614 * Update the pmap. We really only have to do this if a COW
615 * occured to replace the read-only page with the new page. For
616 * now just do it unconditionally. XXX
617 */
618 pmap_enter(fs.map->pmap, vaddr, fs.m, fs.prot, fs.wired);
17cde63e 619 vm_page_flag_set(fs.m, PG_REFERENCED);
40d6ef3a
MD
620
621 /*
4e158347
MD
622 * Unbusy the page by activating it. It remains held and will not
623 * be reclaimed.
624 */
625 vm_page_activate(fs.m);
626
627 if (curthread->td_lwp) {
628 if (fs.hardfault) {
629 curthread->td_lwp->lwp_ru.ru_majflt++;
630 } else {
631 curthread->td_lwp->lwp_ru.ru_minflt++;
632 }
633 }
634
635 /*
636 * Unlock everything, and return the held page.
637 */
638 vm_page_wakeup(fs.m);
639 vm_object_deallocate(fs.first_object);
640
641 *errorp = 0;
642 return(fs.m);
643}
644
645/*
17cde63e
MD
646 * Fault in the specified (object,offset), dirty the returned page as
647 * needed. If the requested fault_type cannot be done NULL and an
648 * error is returned.
aa542ad5
MD
649 */
650vm_page_t
651vm_fault_object_page(vm_object_t object, vm_ooffset_t offset,
652 vm_prot_t fault_type, int fault_flags, int *errorp)
653{
654 int result;
655 vm_pindex_t first_pindex;
656 struct faultstate fs;
657 struct vm_map_entry entry;
658
659 bzero(&entry, sizeof(entry));
660 entry.object.vm_object = object;
661 entry.maptype = VM_MAPTYPE_NORMAL;
662 entry.protection = entry.max_protection = fault_type;
663
664 fs.didlimit = 0;
665 fs.hardfault = 0;
666 fs.fault_flags = fault_flags;
667 fs.map = NULL;
668 KKASSERT((fault_flags & VM_FAULT_WIRE_MASK) == 0);
669
670RetryFault:
671
672 fs.first_object = object;
673 first_pindex = OFF_TO_IDX(offset);
674 fs.entry = &entry;
675 fs.first_prot = fault_type;
676 fs.wired = 0;
677 /*fs.map_generation = 0; unused */
678
679 /*
680 * Make a reference to this object to prevent its disposal while we
681 * are messing with it. Once we have the reference, the map is free
682 * to be diddled. Since objects reference their shadows (and copies),
683 * they will stay around as well.
684 *
685 * Bump the paging-in-progress count to prevent size changes (e.g.
686 * truncation operations) during I/O. This must be done after
687 * obtaining the vnode lock in order to avoid possible deadlocks.
688 */
689 vm_object_reference(fs.first_object);
690 fs.vp = vnode_pager_lock(fs.first_object);
691 vm_object_pip_add(fs.first_object, 1);
692
693 fs.lookup_still_valid = TRUE;
694 fs.first_m = NULL;
695 fs.object = fs.first_object; /* so unlock_and_deallocate works */
696
697#if 0
698 /* XXX future - ability to operate on VM object using vpagetable */
699 if (fs.entry->maptype == VM_MAPTYPE_VPAGETABLE) {
700 result = vm_fault_vpagetable(&fs, &first_pindex,
701 fs.entry->aux.master_pde,
702 fault_type);
703 if (result == KERN_TRY_AGAIN)
704 goto RetryFault;
705 if (result != KERN_SUCCESS) {
706 *errorp = result;
707 return (NULL);
708 }
709 }
710#endif
711
712 /*
713 * Now we have the actual (object, pindex), fault in the page. If
714 * vm_fault_object() fails it will unlock and deallocate the FS
715 * data. If it succeeds everything remains locked and fs->object
716 * will have an additinal PIP count if it is not equal to
717 * fs->first_object
718 */
719 result = vm_fault_object(&fs, first_pindex, fault_type);
720
721 if (result == KERN_TRY_AGAIN)
722 goto RetryFault;
723 if (result != KERN_SUCCESS) {
724 *errorp = result;
725 return(NULL);
726 }
727
17cde63e
MD
728 if ((fault_type & VM_PROT_WRITE) && (fs.prot & VM_PROT_WRITE) == 0) {
729 *errorp = KERN_PROTECTION_FAILURE;
730 unlock_and_deallocate(&fs);
731 return(NULL);
732 }
733
aa542ad5
MD
734 /*
735 * On success vm_fault_object() does not unlock or deallocate, and fs.m
736 * will contain a busied page.
737 */
738 unlock_things(&fs);
739
740 /*
741 * Return a held page. We are not doing any pmap manipulation so do
742 * not set PG_MAPPED. However, adjust the page flags according to
743 * the fault type because the caller may not use a managed pmapping
744 * (so we don't want to lose the fact that the page will be dirtied
745 * if a write fault was specified).
746 */
747 vm_page_hold(fs.m);
748 vm_page_flag_clear(fs.m, PG_ZERO);
749 if (fault_type & VM_PROT_WRITE)
750 vm_page_dirty(fs.m);
751
752 /*
753 * Indicate that the page was accessed.
754 */
755 vm_page_flag_set(fs.m, PG_REFERENCED);
756
757 /*
758 * Unbusy the page by activating it. It remains held and will not
759 * be reclaimed.
760 */
761 vm_page_activate(fs.m);
762
763 if (curthread->td_lwp) {
764 if (fs.hardfault) {
765 mycpu->gd_cnt.v_vm_faults++;
766 curthread->td_lwp->lwp_ru.ru_majflt++;
767 } else {
768 curthread->td_lwp->lwp_ru.ru_minflt++;
769 }
770 }
771
772 /*
773 * Unlock everything, and return the held page.
774 */
775 vm_page_wakeup(fs.m);
776 vm_object_deallocate(fs.first_object);
777
778 *errorp = 0;
779 return(fs.m);
780}
781
782/*
72579d2e 783 * Translate the virtual page number (first_pindex) that is relative
afeabdca
MD
784 * to the address space into a logical page number that is relative to the
785 * backing object. Use the virtual page table pointed to by (vpte).
786 *
787 * This implements an N-level page table. Any level can terminate the
788 * scan by setting VPTE_PS. A linear mapping is accomplished by setting
789 * VPTE_PS in the master page directory entry set via mcontrol(MADV_SETMAP).
790 */
791static
792int
4e7c41c5
MD
793vm_fault_vpagetable(struct faultstate *fs, vm_pindex_t *pindex,
794 vpte_t vpte, int fault_type)
afeabdca 795{
5c5185ae 796 struct lwbuf *lwb;
afeabdca 797 int vshift = 32 - PAGE_SHIFT; /* page index bits remaining */
72579d2e 798 int result = KERN_SUCCESS;
4e7c41c5 799 vpte_t *ptep;
afeabdca
MD
800
801 for (;;) {
4e7c41c5
MD
802 /*
803 * We cannot proceed if the vpte is not valid, not readable
804 * for a read fault, or not writable for a write fault.
805 */
afeabdca
MD
806 if ((vpte & VPTE_V) == 0) {
807 unlock_and_deallocate(fs);
808 return (KERN_FAILURE);
809 }
4e7c41c5
MD
810 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_R) == 0) {
811 unlock_and_deallocate(fs);
812 return (KERN_FAILURE);
813 }
814 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_W) == 0) {
815 unlock_and_deallocate(fs);
816 return (KERN_FAILURE);
817 }
afeabdca
MD
818 if ((vpte & VPTE_PS) || vshift == 0)
819 break;
820 KKASSERT(vshift >= VPTE_PAGE_BITS);
821
822 /*
4e7c41c5
MD
823 * Get the page table page. Nominally we only read the page
824 * table, but since we are actively setting VPTE_M and VPTE_A,
825 * tell vm_fault_object() that we are writing it.
826 *
827 * There is currently no real need to optimize this.
afeabdca 828 */
0035dca9
MD
829 result = vm_fault_object(fs, vpte >> PAGE_SHIFT,
830 VM_PROT_READ|VM_PROT_WRITE);
afeabdca
MD
831 if (result != KERN_SUCCESS)
832 return (result);
833
834 /*
835 * Process the returned fs.m and look up the page table
836 * entry in the page table page.
837 */
838 vshift -= VPTE_PAGE_BITS;
5c5185ae
SG
839 lwb = lwbuf_alloc(fs->m);
840 ptep = ((vpte_t *)lwbuf_kva(lwb) +
4e7c41c5
MD
841 ((*pindex >> vshift) & VPTE_PAGE_MASK));
842 vpte = *ptep;
843
844 /*
845 * Page table write-back. If the vpte is valid for the
846 * requested operation, do a write-back to the page table.
847 *
848 * XXX VPTE_M is not set properly for page directory pages.
849 * It doesn't get set in the page directory if the page table
850 * is modified during a read access.
851 */
852 if ((fault_type & VM_PROT_WRITE) && (vpte & VPTE_V) &&
853 (vpte & VPTE_W)) {
0035dca9 854 if ((vpte & (VPTE_M|VPTE_A)) != (VPTE_M|VPTE_A)) {
4e7c41c5
MD
855 atomic_set_int(ptep, VPTE_M|VPTE_A);
856 vm_page_dirty(fs->m);
857 }
858 }
859 if ((fault_type & VM_PROT_READ) && (vpte & VPTE_V) &&
860 (vpte & VPTE_R)) {
861 if ((vpte & VPTE_A) == 0) {
862 atomic_set_int(ptep, VPTE_A);
863 vm_page_dirty(fs->m);
864 }
865 }
5c5185ae 866 lwbuf_free(lwb);
afeabdca
MD
867 vm_page_flag_set(fs->m, PG_REFERENCED);
868 vm_page_activate(fs->m);
869 vm_page_wakeup(fs->m);
870 cleanup_successful_fault(fs);
871 }
afeabdca
MD
872 /*
873 * Combine remaining address bits with the vpte.
874 */
72579d2e
MD
875 *pindex = (vpte >> PAGE_SHIFT) +
876 (*pindex & ((1 << vshift) - 1));
afeabdca
MD
877 return (KERN_SUCCESS);
878}
879
880
881/*
72579d2e 882 * Do all operations required to fault-in (fs.first_object, pindex). Run
568e6804
MD
883 * through the shadow chain as necessary and do required COW or virtual
884 * copy operations. The caller has already fully resolved the vm_map_entry
885 * and, if appropriate, has created a copy-on-write layer. All we need to
886 * do is iterate the object chain.
887 *
888 * On failure (fs) is unlocked and deallocated and the caller may return or
75f59a66
MD
889 * retry depending on the failure code. On success (fs) is NOT unlocked or
890 * deallocated, fs.m will contained a resolved, busied page, and fs.object
891 * will have an additional PIP count if it is not equal to fs.first_object.
568e6804
MD
892 */
893static
894int
72579d2e
MD
895vm_fault_object(struct faultstate *fs,
896 vm_pindex_t first_pindex, vm_prot_t fault_type)
568e6804
MD
897{
898 vm_object_t next_object;
72579d2e 899 vm_pindex_t pindex;
568e6804 900
72579d2e
MD
901 fs->prot = fs->first_prot;
902 fs->object = fs->first_object;
903 pindex = first_pindex;
904
4e7c41c5
MD
905 /*
906 * If a read fault occurs we try to make the page writable if
907 * possible. There are three cases where we cannot make the
908 * page mapping writable:
909 *
910 * (1) The mapping is read-only or the VM object is read-only,
0035dca9 911 * fs->prot above will simply not have VM_PROT_WRITE set.
4e7c41c5
MD
912 *
913 * (2) If the mapping is a virtual page table we need to be able
70fc5283
MD
914 * to detect writes so we can set VPTE_M in the virtual page
915 * table.
4e7c41c5
MD
916 *
917 * (3) If the VM page is read-only or copy-on-write, upgrading would
918 * just result in an unnecessary COW fault.
0035dca9
MD
919 *
920 * VM_PROT_VPAGED is set if faulting via a virtual page table and
921 * causes adjustments to the 'M'odify bit to also turn off write
922 * access to force a re-fault.
4e7c41c5 923 */
0035dca9
MD
924 if (fs->entry->maptype == VM_MAPTYPE_VPAGETABLE) {
925 if ((fault_type & VM_PROT_WRITE) == 0)
926 fs->prot &= ~VM_PROT_WRITE;
4e7c41c5
MD
927 }
928
568e6804
MD
929 for (;;) {
930 /*
931 * If the object is dead, we stop here
932 */
933 if (fs->object->flags & OBJ_DEAD) {
934 unlock_and_deallocate(fs);
984263bc
MD
935 return (KERN_PROTECTION_FAILURE);
936 }
937
938 /*
06ecca5a
MD
939 * See if page is resident. spl protection is required
940 * to avoid an interrupt unbusy/free race against our
941 * lookup. We must hold the protection through a page
942 * allocation or busy.
984263bc 943 */
654a39f0 944 crit_enter();
72579d2e 945 fs->m = vm_page_lookup(fs->object, pindex);
568e6804 946 if (fs->m != NULL) {
06ecca5a 947 int queue;
984263bc
MD
948 /*
949 * Wait/Retry if the page is busy. We have to do this
950 * if the page is busy via either PG_BUSY or
951 * vm_page_t->busy because the vm_pager may be using
952 * vm_page_t->busy for pageouts ( and even pageins if
953 * it is the vnode pager ), and we could end up trying
954 * to pagein and pageout the same page simultaneously.
955 *
956 * We can theoretically allow the busy case on a read
957 * fault if the page is marked valid, but since such
958 * pages are typically already pmap'd, putting that
959 * special case in might be more effort then it is
960 * worth. We cannot under any circumstances mess
961 * around with a vm_page_t->busy page except, perhaps,
962 * to pmap it.
963 */
568e6804
MD
964 if ((fs->m->flags & PG_BUSY) || fs->m->busy) {
965 unlock_things(fs);
966 vm_page_sleep_busy(fs->m, TRUE, "vmpfw");
12e4aaff 967 mycpu->gd_cnt.v_intrans++;
568e6804 968 vm_object_deallocate(fs->first_object);
bc823b32 969 fs->first_object = NULL;
654a39f0 970 crit_exit();
568e6804 971 return (KERN_TRY_AGAIN);
984263bc
MD
972 }
973
568e6804
MD
974 /*
975 * If reactivating a page from PQ_CACHE we may have
976 * to rate-limit.
977 */
978 queue = fs->m->queue;
979 vm_page_unqueue_nowakeup(fs->m);
984263bc 980
568e6804
MD
981 if ((queue - fs->m->pc) == PQ_CACHE &&
982 vm_page_count_severe()) {
983 vm_page_activate(fs->m);
984 unlock_and_deallocate(fs);
659c6a07 985 vm_waitpfault();
654a39f0 986 crit_exit();
568e6804 987 return (KERN_TRY_AGAIN);
984263bc
MD
988 }
989
990 /*
991 * Mark page busy for other processes, and the
992 * pagedaemon. If it still isn't completely valid
cf1bb2a8
MD
993 * (readable), or if a read-ahead-mark is set on
994 * the VM page, jump to readrest, else we found the
568e6804 995 * page and can return.
06ecca5a
MD
996 *
997 * We can release the spl once we have marked the
998 * page busy.
984263bc 999 */
568e6804 1000 vm_page_busy(fs->m);
654a39f0 1001 crit_exit();
06ecca5a 1002
cf1bb2a8
MD
1003 if (fs->m->object != &kernel_object) {
1004 if ((fs->m->valid & VM_PAGE_BITS_ALL) !=
1005 VM_PAGE_BITS_ALL) {
1006 goto readrest;
1007 }
1008 if (fs->m->flags & PG_RAM) {
1009 if (debug_cluster)
1010 kprintf("R");
1011 vm_page_flag_clear(fs->m, PG_RAM);
1012 goto readrest;
1013 }
984263bc 1014 }
568e6804 1015 break; /* break to PAGE HAS BEEN FOUND */
984263bc
MD
1016 }
1017
1018 /*
1019 * Page is not resident, If this is the search termination
1020 * or the pager might contain the page, allocate a new page.
06ecca5a 1021 *
568e6804 1022 * NOTE: We are still in a critical section.
984263bc 1023 */
568e6804
MD
1024 if (TRYPAGER(fs) || fs->object == fs->first_object) {
1025 /*
1026 * If the page is beyond the object size we fail
1027 */
72579d2e 1028 if (pindex >= fs->object->size) {
654a39f0 1029 crit_exit();
568e6804 1030 unlock_and_deallocate(fs);
984263bc
MD
1031 return (KERN_PROTECTION_FAILURE);
1032 }
1033
1034 /*
46311ac2
MD
1035 * Ratelimit.
1036 */
568e6804
MD
1037 if (fs->didlimit == 0 && curproc != NULL) {
1038 int limticks;
1039
1040 limticks = vm_fault_ratelimit(curproc->p_vmspace);
46311ac2
MD
1041 if (limticks) {
1042 crit_exit();
568e6804 1043 unlock_and_deallocate(fs);
46311ac2 1044 tsleep(curproc, 0, "vmrate", limticks);
568e6804
MD
1045 fs->didlimit = 1;
1046 return (KERN_TRY_AGAIN);
46311ac2
MD
1047 }
1048 }
1049
1050 /*
984263bc
MD
1051 * Allocate a new page for this object/offset pair.
1052 */
568e6804 1053 fs->m = NULL;
984263bc 1054 if (!vm_page_count_severe()) {
72579d2e 1055 fs->m = vm_page_alloc(fs->object, pindex,
568e6804 1056 (fs->vp || fs->object->backing_object) ? VM_ALLOC_NORMAL : VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
984263bc 1057 }
568e6804 1058 if (fs->m == NULL) {
654a39f0 1059 crit_exit();
568e6804 1060 unlock_and_deallocate(fs);
659c6a07 1061 vm_waitpfault();
568e6804 1062 return (KERN_TRY_AGAIN);
984263bc
MD
1063 }
1064 }
654a39f0 1065 crit_exit();
984263bc
MD
1066
1067readrest:
1068 /*
1b9d3514 1069 * We have found an invalid or partially valid page, a
1c9602b3
MD
1070 * page with a read-ahead mark which might be partially or
1071 * fully valid (and maybe dirty too), or we have allocated
1072 * a new page.
984263bc
MD
1073 *
1074 * Attempt to fault-in the page if there is a chance that the
1075 * pager has it, and potentially fault in additional pages
1076 * at the same time.
06ecca5a
MD
1077 *
1078 * We are NOT in splvm here and if TRYPAGER is true then
1079 * fs.m will be non-NULL and will be PG_BUSY for us.
984263bc 1080 */
568e6804 1081 if (TRYPAGER(fs)) {
984263bc 1082 int rv;
1b9d3514 1083 int seqaccess;
568e6804 1084 u_char behavior = vm_map_entry_behavior(fs->entry);
984263bc 1085
1b9d3514
MD
1086 if (behavior == MAP_ENTRY_BEHAV_RANDOM)
1087 seqaccess = 0;
1088 else
1089 seqaccess = -1;
984263bc 1090
1b9d3514
MD
1091 /*
1092 * If sequential access is detected then attempt
1093 * to deactivate/cache pages behind the scan to
1094 * prevent resource hogging.
1095 *
1096 * Use of PG_RAM to detect sequential access
1097 * also simulates multi-zone sequential access
1098 * detection for free.
1099 *
1100 * NOTE: Partially valid dirty pages cannot be
1101 * deactivated without causing NFS picemeal
1102 * writes to barf.
1103 */
568e6804 1104 if ((fs->first_object->type != OBJT_DEVICE) &&
984263bc
MD
1105 (behavior == MAP_ENTRY_BEHAV_SEQUENTIAL ||
1106 (behavior != MAP_ENTRY_BEHAV_RANDOM &&
1b9d3514 1107 (fs->m->flags & PG_RAM)))
984263bc 1108 ) {
1b9d3514
MD
1109 vm_pindex_t scan_pindex;
1110 int scan_count = 16;
1111
1112 if (first_pindex < 16) {
1113 scan_pindex = 0;
1114 scan_count = 0;
1115 } else {
1116 scan_pindex = first_pindex - 16;
1117 if (scan_pindex < 16)
1118 scan_count = scan_pindex;
1119 else
1120 scan_count = 16;
1121 }
984263bc 1122
654a39f0 1123 crit_enter();
1b9d3514 1124 while (scan_count) {
984263bc 1125 vm_page_t mt;
568e6804 1126
1b9d3514
MD
1127 mt = vm_page_lookup(fs->first_object,
1128 scan_pindex);
1129 if (mt == NULL ||
1130 (mt->valid != VM_PAGE_BITS_ALL)) {
984263bc 1131 break;
1b9d3514 1132 }
984263bc 1133 if (mt->busy ||
1b9d3514
MD
1134 (mt->flags & (PG_BUSY | PG_FICTITIOUS | PG_UNMANAGED)) ||
1135 mt->hold_count ||
1136 mt->wire_count) {
1137 goto skip;
1138 }
984263bc
MD
1139 if (mt->dirty == 0)
1140 vm_page_test_dirty(mt);
1141 if (mt->dirty) {
17cde63e 1142 vm_page_busy(mt);
1b9d3514
MD
1143 vm_page_protect(mt,
1144 VM_PROT_NONE);
984263bc 1145 vm_page_deactivate(mt);
17cde63e 1146 vm_page_wakeup(mt);
984263bc
MD
1147 } else {
1148 vm_page_cache(mt);
1149 }
1b9d3514
MD
1150skip:
1151 --scan_count;
1152 --scan_pindex;
984263bc 1153 }
654a39f0 1154 crit_exit();
984263bc 1155
1b9d3514 1156 seqaccess = 1;
984263bc
MD
1157 }
1158
1159 /*
1b9d3514
MD
1160 * Avoid deadlocking against the map when doing I/O.
1161 * fs.object and the page is PG_BUSY'd.
984263bc 1162 */
1b9d3514 1163 unlock_map(fs);
984263bc
MD
1164
1165 /*
1b9d3514
MD
1166 * Acquire the page data. We still hold a ref on
1167 * fs.object and the page has been PG_BUSY's.
1168 *
1169 * The pager may replace the page (for example, in
1170 * order to enter a fictitious page into the
1171 * object). If it does so it is responsible for
1172 * cleaning up the passed page and properly setting
1173 * the new page PG_BUSY.
1c9602b3
MD
1174 *
1175 * If we got here through a PG_RAM read-ahead
1176 * mark the page may be partially dirty and thus
1177 * not freeable. Don't bother checking to see
1178 * if the pager has the page because we can't free
1179 * it anyway. We have to depend on the get_page
1180 * operation filling in any gaps whether there is
1181 * backing store or not.
984263bc 1182 */
1c9602b3 1183 rv = vm_pager_get_page(fs->object, &fs->m, seqaccess);
984263bc
MD
1184
1185 if (rv == VM_PAGER_OK) {
1186 /*
984263bc
MD
1187 * Relookup in case pager changed page. Pager
1188 * is responsible for disposition of old page
1189 * if moved.
06ecca5a
MD
1190 *
1191 * XXX other code segments do relookups too.
1192 * It's a bad abstraction that needs to be
1193 * fixed/removed.
984263bc 1194 */
72579d2e 1195 fs->m = vm_page_lookup(fs->object, pindex);
568e6804
MD
1196 if (fs->m == NULL) {
1197 unlock_and_deallocate(fs);
1198 return (KERN_TRY_AGAIN);
984263bc
MD
1199 }
1200
568e6804 1201 ++fs->hardfault;
984263bc
MD
1202 break; /* break to PAGE HAS BEEN FOUND */
1203 }
568e6804 1204
984263bc
MD
1205 /*
1206 * Remove the bogus page (which does not exist at this
1207 * object/offset); before doing so, we must get back
1208 * our object lock to preserve our invariant.
1209 *
1210 * Also wake up any other process that may want to bring
1211 * in this page.
1212 *
1213 * If this is the top-level object, we must leave the
1214 * busy page to prevent another process from rushing
1215 * past us, and inserting the page in that object at
1216 * the same time that we are.
1217 */
a0bc8638
MD
1218 if (rv == VM_PAGER_ERROR) {
1219 if (curproc)
086c1d7e 1220 kprintf("vm_fault: pager read error, pid %d (%s)\n", curproc->p_pid, curproc->p_comm);
a0bc8638 1221 else
086c1d7e 1222 kprintf("vm_fault: pager read error, thread %p (%s)\n", curthread, curproc->p_comm);
a0bc8638 1223 }
1b9d3514 1224
984263bc
MD
1225 /*
1226 * Data outside the range of the pager or an I/O error
a55afca2
MD
1227 *
1228 * The page may have been wired during the pagein,
1229 * e.g. by the buffer cache, and cannot simply be
1b9d3514 1230 * freed. Call vnode_pager_freepage() to deal with it.
984263bc
MD
1231 */
1232 /*
1233 * XXX - the check for kernel_map is a kludge to work
1234 * around having the machine panic on a kernel space
1235 * fault w/ I/O error.
1236 */
1b9d3514
MD
1237 if (((fs->map != &kernel_map) &&
1238 (rv == VM_PAGER_ERROR)) || (rv == VM_PAGER_BAD)) {
a55afca2 1239 vnode_pager_freepage(fs->m);
568e6804
MD
1240 fs->m = NULL;
1241 unlock_and_deallocate(fs);
1242 if (rv == VM_PAGER_ERROR)
1243 return (KERN_FAILURE);
1244 else
1245 return (KERN_PROTECTION_FAILURE);
1246 /* NOT REACHED */
984263bc 1247 }
568e6804 1248 if (fs->object != fs->first_object) {
a55afca2 1249 vnode_pager_freepage(fs->m);
568e6804 1250 fs->m = NULL;
984263bc
MD
1251 /*
1252 * XXX - we cannot just fall out at this
1253 * point, m has been freed and is invalid!
1254 */
1255 }
1256 }
1257
1258 /*
568e6804 1259 * We get here if the object has a default pager (or unwiring)
984263bc
MD
1260 * or the pager doesn't have the page.
1261 */
568e6804
MD
1262 if (fs->object == fs->first_object)
1263 fs->first_m = fs->m;
984263bc
MD
1264
1265 /*
1266 * Move on to the next object. Lock the next object before
1267 * unlocking the current one.
1268 */
72579d2e 1269 pindex += OFF_TO_IDX(fs->object->backing_object_offset);
568e6804 1270 next_object = fs->object->backing_object;
984263bc
MD
1271 if (next_object == NULL) {
1272 /*
1273 * If there's no object left, fill the page in the top
1274 * object with zeros.
1275 */
568e6804
MD
1276 if (fs->object != fs->first_object) {
1277 vm_object_pip_wakeup(fs->object);
984263bc 1278
568e6804 1279 fs->object = fs->first_object;
72579d2e 1280 pindex = first_pindex;
568e6804 1281 fs->m = fs->first_m;
984263bc 1282 }
568e6804 1283 fs->first_m = NULL;
984263bc
MD
1284
1285 /*
1286 * Zero the page if necessary and mark it valid.
1287 */
568e6804
MD
1288 if ((fs->m->flags & PG_ZERO) == 0) {
1289 vm_page_zero_fill(fs->m);
984263bc 1290 } else {
12e4aaff 1291 mycpu->gd_cnt.v_ozfod++;
984263bc 1292 }
12e4aaff 1293 mycpu->gd_cnt.v_zfod++;
568e6804 1294 fs->m->valid = VM_PAGE_BITS_ALL;
984263bc
MD
1295 break; /* break to PAGE HAS BEEN FOUND */
1296 } else {
568e6804
MD
1297 if (fs->object != fs->first_object) {
1298 vm_object_pip_wakeup(fs->object);
984263bc 1299 }
568e6804
MD
1300 KASSERT(fs->object != next_object, ("object loop %p", next_object));
1301 fs->object = next_object;
1302 vm_object_pip_add(fs->object, 1);
984263bc
MD
1303 }
1304 }
1305
984263bc
MD
1306 /*
1307 * PAGE HAS BEEN FOUND. [Loop invariant still holds -- the object lock
1308 * is held.]
1b9d3514 1309 *
984263bc
MD
1310 * If the page is being written, but isn't already owned by the
1311 * top-level object, we have to copy it into a new page owned by the
1312 * top-level object.
1313 */
1b9d3514
MD
1314 KASSERT((fs->m->flags & PG_BUSY) != 0,
1315 ("vm_fault: not busy after main loop"));
1316
568e6804 1317 if (fs->object != fs->first_object) {
984263bc
MD
1318 /*
1319 * We only really need to copy if we want to write it.
1320 */
984263bc
MD
1321 if (fault_type & VM_PROT_WRITE) {
1322 /*
1323 * This allows pages to be virtually copied from a
1324 * backing_object into the first_object, where the
1325 * backing object has no other refs to it, and cannot
1326 * gain any more refs. Instead of a bcopy, we just
1327 * move the page from the backing object to the
1328 * first object. Note that we must mark the page
1329 * dirty in the first object so that it will go out
1330 * to swap when needed.
1331 */
aa542ad5
MD
1332 if (
1333 /*
1334 * Map, if present, has not changed
1335 */
1336 (fs->map == NULL ||
1337 fs->map_generation == fs->map->timestamp) &&
984263bc
MD
1338 /*
1339 * Only one shadow object
1340 */
568e6804 1341 (fs->object->shadow_count == 1) &&
984263bc
MD
1342 /*
1343 * No COW refs, except us
1344 */
568e6804 1345 (fs->object->ref_count == 1) &&
984263bc
MD
1346 /*
1347 * No one else can look this object up
1348 */
568e6804 1349 (fs->object->handle == NULL) &&
984263bc
MD
1350 /*
1351 * No other ways to look the object up
1352 */
568e6804
MD
1353 ((fs->object->type == OBJT_DEFAULT) ||
1354 (fs->object->type == OBJT_SWAP)) &&
984263bc
MD
1355 /*
1356 * We don't chase down the shadow chain
1357 */
568e6804 1358 (fs->object == fs->first_object->backing_object) &&
984263bc
MD
1359
1360 /*
1361 * grab the lock if we need to
1362 */
568e6804 1363 (fs->lookup_still_valid ||
aa542ad5 1364 fs->map == NULL ||
568e6804 1365 lockmgr(&fs->map->lock, LK_EXCLUSIVE|LK_NOWAIT) == 0)
984263bc
MD
1366 ) {
1367
568e6804 1368 fs->lookup_still_valid = 1;
984263bc
MD
1369 /*
1370 * get rid of the unnecessary page
1371 */
568e6804
MD
1372 vm_page_protect(fs->first_m, VM_PROT_NONE);
1373 vm_page_free(fs->first_m);
1374 fs->first_m = NULL;
984263bc
MD
1375
1376 /*
1377 * grab the page and put it into the
1378 * process'es object. The page is
1379 * automatically made dirty.
1380 */
72579d2e 1381 vm_page_rename(fs->m, fs->first_object, first_pindex);
568e6804
MD
1382 fs->first_m = fs->m;
1383 vm_page_busy(fs->first_m);
1384 fs->m = NULL;
12e4aaff 1385 mycpu->gd_cnt.v_cow_optim++;
984263bc
MD
1386 } else {
1387 /*
1388 * Oh, well, lets copy it.
1389 */
568e6804 1390 vm_page_copy(fs->m, fs->first_m);
10192bae 1391 vm_page_event(fs->m, VMEVENT_COW);
984263bc
MD
1392 }
1393
568e6804 1394 if (fs->m) {
984263bc
MD
1395 /*
1396 * We no longer need the old page or object.
1397 */
568e6804 1398 release_page(fs);
984263bc
MD
1399 }
1400
1401 /*
568e6804 1402 * fs->object != fs->first_object due to above
984263bc
MD
1403 * conditional
1404 */
568e6804 1405 vm_object_pip_wakeup(fs->object);
984263bc
MD
1406
1407 /*
1408 * Only use the new page below...
1409 */
1410
12e4aaff 1411 mycpu->gd_cnt.v_cow_faults++;
568e6804
MD
1412 fs->m = fs->first_m;
1413 fs->object = fs->first_object;
72579d2e 1414 pindex = first_pindex;
984263bc 1415 } else {
568e6804
MD
1416 /*
1417 * If it wasn't a write fault avoid having to copy
1418 * the page by mapping it read-only.
1419 */
1420 fs->prot &= ~VM_PROT_WRITE;
984263bc
MD
1421 }
1422 }
1423
1424 /*
568e6804
MD
1425 * We may have had to unlock a map to do I/O. If we did then
1426 * lookup_still_valid will be FALSE. If the map generation count
1427 * also changed then all sorts of things could have happened while
1428 * we were doing the I/O and we need to retry.
984263bc
MD
1429 */
1430
568e6804 1431 if (!fs->lookup_still_valid &&
aa542ad5 1432 fs->map != NULL &&
568e6804
MD
1433 (fs->map->timestamp != fs->map_generation)) {
1434 release_page(fs);
1435 unlock_and_deallocate(fs);
1436 return (KERN_TRY_AGAIN);
1437 }
1438
984263bc 1439 /*
17cde63e
MD
1440 * If the fault is a write, we know that this page is being
1441 * written NOW so dirty it explicitly to save on pmap_is_modified()
1442 * calls later.
1443 *
1444 * If this is a NOSYNC mmap we do not want to set PG_NOSYNC
1445 * if the page is already dirty to prevent data written with
1446 * the expectation of being synced from not being synced.
1447 * Likewise if this entry does not request NOSYNC then make
1448 * sure the page isn't marked NOSYNC. Applications sharing
1449 * data should use the same flags to avoid ping ponging.
1450 *
1451 * Also tell the backing pager, if any, that it should remove
1452 * any swap backing since the page is now dirty.
984263bc 1453 */
568e6804 1454 if (fs->prot & VM_PROT_WRITE) {
568e6804 1455 vm_object_set_writeable_dirty(fs->m->object);
568e6804
MD
1456 if (fs->entry->eflags & MAP_ENTRY_NOSYNC) {
1457 if (fs->m->dirty == 0)
1458 vm_page_flag_set(fs->m, PG_NOSYNC);
984263bc 1459 } else {
568e6804 1460 vm_page_flag_clear(fs->m, PG_NOSYNC);
984263bc 1461 }
568e6804 1462 if (fs->fault_flags & VM_FAULT_DIRTY) {
654a39f0 1463 crit_enter();
568e6804 1464 vm_page_dirty(fs->m);
107e9bcc 1465 swap_pager_unswapped(fs->m);
654a39f0 1466 crit_exit();
984263bc
MD
1467 }
1468 }
1469
1470 /*
75f59a66
MD
1471 * Page had better still be busy. We are still locked up and
1472 * fs->object will have another PIP reference if it is not equal
1473 * to fs->first_object.
984263bc 1474 */
568e6804
MD
1475 KASSERT(fs->m->flags & PG_BUSY,
1476 ("vm_fault: page %p not busy!", fs->m));
984263bc 1477
984263bc
MD
1478 /*
1479 * Sanity check: page must be completely valid or it is not fit to
1480 * map into user space. vm_pager_get_pages() ensures this.
1481 */
568e6804
MD
1482 if (fs->m->valid != VM_PAGE_BITS_ALL) {
1483 vm_page_zero_invalid(fs->m, TRUE);
086c1d7e 1484 kprintf("Warning: page %p partially invalid on fault\n", fs->m);
984263bc
MD
1485 }
1486
984263bc 1487 return (KERN_SUCCESS);
984263bc
MD
1488}
1489
1490/*
f2d22ebf
MD
1491 * Wire down a range of virtual addresses in a map. The entry in question
1492 * should be marked in-transition and the map must be locked. We must
1493 * release the map temporarily while faulting-in the page to avoid a
1494 * deadlock. Note that the entry may be clipped while we are blocked but
1495 * will never be freed.
984263bc
MD
1496 */
1497int
f2d22ebf 1498vm_fault_wire(vm_map_t map, vm_map_entry_t entry, boolean_t user_wire)
984263bc 1499{
f2d22ebf
MD
1500 boolean_t fictitious;
1501 vm_offset_t start;
1502 vm_offset_t end;
5f910b2f 1503 vm_offset_t va;
f2d22ebf 1504 vm_paddr_t pa;
5f910b2f 1505 pmap_t pmap;
984263bc
MD
1506 int rv;
1507
1508 pmap = vm_map_pmap(map);
f2d22ebf
MD
1509 start = entry->start;
1510 end = entry->end;
1511 fictitious = entry->object.vm_object &&
1512 (entry->object.vm_object->type == OBJT_DEVICE);
984263bc 1513
f2d22ebf
MD
1514 vm_map_unlock(map);
1515 map->timestamp++;
984263bc
MD
1516
1517 /*
984263bc
MD
1518 * We simulate a fault to get the page and enter it in the physical
1519 * map.
1520 */
1521 for (va = start; va < end; va += PAGE_SIZE) {
f2d22ebf
MD
1522 if (user_wire) {
1523 rv = vm_fault(map, va, VM_PROT_READ,
1524 VM_FAULT_USER_WIRE);
1525 } else {
1526 rv = vm_fault(map, va, VM_PROT_READ|VM_PROT_WRITE,
1527 VM_FAULT_CHANGE_WIRING);
1528 }
984263bc 1529 if (rv) {
f2d22ebf
MD
1530 while (va > start) {
1531 va -= PAGE_SIZE;
1532 if ((pa = pmap_extract(pmap, va)) == 0)
1533 continue;
1534 pmap_change_wiring(pmap, va, FALSE);
1535 if (!fictitious)
1536 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
1537 }
f2d22ebf 1538 vm_map_lock(map);
984263bc
MD
1539 return (rv);
1540 }
1541 }
f2d22ebf 1542 vm_map_lock(map);
984263bc
MD
1543 return (KERN_SUCCESS);
1544}
1545
984263bc 1546/*
f2d22ebf
MD
1547 * Unwire a range of virtual addresses in a map. The map should be
1548 * locked.
984263bc
MD
1549 */
1550void
f2d22ebf 1551vm_fault_unwire(vm_map_t map, vm_map_entry_t entry)
984263bc 1552{
f2d22ebf
MD
1553 boolean_t fictitious;
1554 vm_offset_t start;
1555 vm_offset_t end;
6ef943a3
MD
1556 vm_offset_t va;
1557 vm_paddr_t pa;
5f910b2f 1558 pmap_t pmap;
984263bc
MD
1559
1560 pmap = vm_map_pmap(map);
f2d22ebf
MD
1561 start = entry->start;
1562 end = entry->end;
1563 fictitious = entry->object.vm_object &&
1564 (entry->object.vm_object->type == OBJT_DEVICE);
984263bc
MD
1565
1566 /*
1567 * Since the pages are wired down, we must be able to get their
1568 * mappings from the physical map system.
1569 */
984263bc
MD
1570 for (va = start; va < end; va += PAGE_SIZE) {
1571 pa = pmap_extract(pmap, va);
6ef943a3 1572 if (pa != 0) {
984263bc 1573 pmap_change_wiring(pmap, va, FALSE);
f2d22ebf
MD
1574 if (!fictitious)
1575 vm_page_unwire(PHYS_TO_VM_PAGE(pa), 1);
984263bc
MD
1576 }
1577 }
984263bc
MD
1578}
1579
1580/*
46311ac2
MD
1581 * Reduce the rate at which memory is allocated to a process based
1582 * on the perceived load on the VM system. As the load increases
1583 * the allocation burst rate goes down and the delay increases.
1584 *
1585 * Rate limiting does not apply when faulting active or inactive
1586 * pages. When faulting 'cache' pages, rate limiting only applies
1587 * if the system currently has a severe page deficit.
1588 *
1589 * XXX vm_pagesupply should be increased when a page is freed.
1590 *
1591 * We sleep up to 1/10 of a second.
1592 */
1593static int
1594vm_fault_ratelimit(struct vmspace *vmspace)
1595{
1596 if (vm_load_enable == 0)
1597 return(0);
1598 if (vmspace->vm_pagesupply > 0) {
1599 --vmspace->vm_pagesupply;
1600 return(0);
1601 }
1602#ifdef INVARIANTS
1603 if (vm_load_debug) {
086c1d7e 1604 kprintf("load %-4d give %d pgs, wait %d, pid %-5d (%s)\n",
46311ac2
MD
1605 vm_load,
1606 (1000 - vm_load ) / 10, vm_load * hz / 10000,
1607 curproc->p_pid, curproc->p_comm);
1608 }
1609#endif
1610 vmspace->vm_pagesupply = (1000 - vm_load) / 10;
1611 return(vm_load * hz / 10000);
1612}
1613
1614/*
984263bc
MD
1615 * Routine:
1616 * vm_fault_copy_entry
1617 * Function:
1618 * Copy all of the pages from a wired-down map entry to another.
1619 *
1620 * In/out conditions:
1621 * The source and destination maps must be locked for write.
1622 * The source map entry must be wired down (or be a sharing map
1623 * entry corresponding to a main map entry that is wired down).
1624 */
1625
1626void
57e43348
MD
1627vm_fault_copy_entry(vm_map_t dst_map, vm_map_t src_map,
1628 vm_map_entry_t dst_entry, vm_map_entry_t src_entry)
984263bc
MD
1629{
1630 vm_object_t dst_object;
1631 vm_object_t src_object;
1632 vm_ooffset_t dst_offset;
1633 vm_ooffset_t src_offset;
1634 vm_prot_t prot;
1635 vm_offset_t vaddr;
1636 vm_page_t dst_m;
1637 vm_page_t src_m;
1638
1639#ifdef lint
1640 src_map++;
1641#endif /* lint */
1642
1643 src_object = src_entry->object.vm_object;
1644 src_offset = src_entry->offset;
1645
1646 /*
1647 * Create the top-level object for the destination entry. (Doesn't
1648 * actually shadow anything - we copy the pages directly.)
1649 */
53025830
MD
1650 vm_map_entry_allocate_object(dst_entry);
1651 dst_object = dst_entry->object.vm_object;
984263bc
MD
1652
1653 prot = dst_entry->max_protection;
1654
1655 /*
1656 * Loop through all of the pages in the entry's range, copying each
1657 * one from the source object (it should be there) to the destination
1658 * object.
1659 */
1660 for (vaddr = dst_entry->start, dst_offset = 0;
1661 vaddr < dst_entry->end;
1662 vaddr += PAGE_SIZE, dst_offset += PAGE_SIZE) {
1663
1664 /*
1665 * Allocate a page in the destination object
1666 */
1667 do {
1668 dst_m = vm_page_alloc(dst_object,
1669 OFF_TO_IDX(dst_offset), VM_ALLOC_NORMAL);
1670 if (dst_m == NULL) {
4ecf7cc9 1671 vm_wait(0);
984263bc
MD
1672 }
1673 } while (dst_m == NULL);
1674
1675 /*
1676 * Find the page in the source object, and copy it in.
1677 * (Because the source is wired down, the page will be in
1678 * memory.)
1679 */
1680 src_m = vm_page_lookup(src_object,
1681 OFF_TO_IDX(dst_offset + src_offset));
1682 if (src_m == NULL)
1683 panic("vm_fault_copy_wired: page missing");
1684
1685 vm_page_copy(src_m, dst_m);
10192bae 1686 vm_page_event(src_m, VMEVENT_COW);
984263bc
MD
1687
1688 /*
1689 * Enter it in the pmap...
1690 */
1691
1692 vm_page_flag_clear(dst_m, PG_ZERO);
1693 pmap_enter(dst_map->pmap, vaddr, dst_m, prot, FALSE);
984263bc
MD
1694
1695 /*
1696 * Mark it no longer busy, and put it on the active list.
1697 */
1698 vm_page_activate(dst_m);
1699 vm_page_wakeup(dst_m);
1700 }
1701}
1702
1b9d3514 1703#if 0
984263bc
MD
1704
1705/*
1706 * This routine checks around the requested page for other pages that
1707 * might be able to be faulted in. This routine brackets the viable
1708 * pages for the pages to be paged in.
1709 *
1710 * Inputs:
1711 * m, rbehind, rahead
1712 *
1713 * Outputs:
1714 * marray (array of vm_page_t), reqpage (index of requested page)
1715 *
1716 * Return value:
1717 * number of pages in marray
1718 */
1719static int
57e43348 1720vm_fault_additional_pages(vm_page_t m, int rbehind, int rahead,
bc823b32 1721 vm_page_t *marray, int *reqpage)
984263bc
MD
1722{
1723 int i,j;
1724 vm_object_t object;
1725 vm_pindex_t pindex, startpindex, endpindex, tpindex;
1726 vm_page_t rtm;
1727 int cbehind, cahead;
1728
1729 object = m->object;
1730 pindex = m->pindex;
1731
1732 /*
1733 * we don't fault-ahead for device pager
1734 */
1735 if (object->type == OBJT_DEVICE) {
1736 *reqpage = 0;
1737 marray[0] = m;
1738 return 1;
1739 }
1740
1741 /*
1742 * if the requested page is not available, then give up now
1743 */
984263bc 1744 if (!vm_pager_has_page(object, pindex, &cbehind, &cahead)) {
17cde63e 1745 *reqpage = 0; /* not used by caller, fix compiler warn */
984263bc
MD
1746 return 0;
1747 }
1748
1749 if ((cbehind == 0) && (cahead == 0)) {
1750 *reqpage = 0;
1751 marray[0] = m;
1752 return 1;
1753 }
1754
1755 if (rahead > cahead) {
1756 rahead = cahead;
1757 }
1758
1759 if (rbehind > cbehind) {
1760 rbehind = cbehind;
1761 }
1762
1763 /*
bc823b32
MD
1764 * Do not do any readahead if we have insufficient free memory.
1765 *
1766 * XXX code was broken disabled before and has instability
1767 * with this conditonal fixed, so shortcut for now.
984263bc 1768 */
bc823b32 1769 if (burst_fault == 0 || vm_page_count_severe()) {
984263bc
MD
1770 marray[0] = m;
1771 *reqpage = 0;
1772 return 1;
1773 }
1774
1775 /*
1776 * scan backward for the read behind pages -- in memory
06ecca5a
MD
1777 *
1778 * Assume that if the page is not found an interrupt will not
1779 * create it. Theoretically interrupts can only remove (busy)
1780 * pages, not create new associations.
984263bc
MD
1781 */
1782 if (pindex > 0) {
1783 if (rbehind > pindex) {
1784 rbehind = pindex;
1785 startpindex = 0;
1786 } else {
1787 startpindex = pindex - rbehind;
1788 }
1789
654a39f0 1790 crit_enter();
bc823b32
MD
1791 for (tpindex = pindex; tpindex > startpindex; --tpindex) {
1792 if (vm_page_lookup(object, tpindex - 1))
984263bc
MD
1793 break;
1794 }
1795
bc823b32
MD
1796 i = 0;
1797 while (tpindex < pindex) {
1798 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM);
984263bc 1799 if (rtm == NULL) {
654a39f0 1800 crit_exit();
984263bc
MD
1801 for (j = 0; j < i; j++) {
1802 vm_page_free(marray[j]);
1803 }
1804 marray[0] = m;
1805 *reqpage = 0;
1806 return 1;
1807 }
984263bc 1808 marray[i] = rtm;
bc823b32
MD
1809 ++i;
1810 ++tpindex;
984263bc 1811 }
654a39f0 1812 crit_exit();
984263bc 1813 } else {
984263bc
MD
1814 i = 0;
1815 }
1816
bc823b32
MD
1817 /*
1818 * Assign requested page
1819 */
984263bc 1820 marray[i] = m;
984263bc 1821 *reqpage = i;
bc823b32 1822 ++i;
984263bc
MD
1823
1824 /*
bc823b32 1825 * Scan forwards for read-ahead pages
984263bc 1826 */
bc823b32 1827 tpindex = pindex + 1;
984263bc
MD
1828 endpindex = tpindex + rahead;
1829 if (endpindex > object->size)
1830 endpindex = object->size;
1831
654a39f0 1832 crit_enter();
bc823b32
MD
1833 while (tpindex < endpindex) {
1834 if (vm_page_lookup(object, tpindex))
984263bc 1835 break;
bc823b32
MD
1836 rtm = vm_page_alloc(object, tpindex, VM_ALLOC_SYSTEM);
1837 if (rtm == NULL)
984263bc 1838 break;
984263bc 1839 marray[i] = rtm;
bc823b32
MD
1840 ++i;
1841 ++tpindex;
984263bc 1842 }
654a39f0 1843 crit_exit();
984263bc 1844
bc823b32 1845 return (i);
984263bc 1846}
1b9d3514
MD
1847
1848#endif
1849
1850/*
1851 * vm_prefault() provides a quick way of clustering pagefaults into a
1852 * processes address space. It is a "cousin" of pmap_object_init_pt,
1853 * except it runs at page fault time instead of mmap time.
1854 *
1855 * This code used to be per-platform pmap_prefault(). It is now
1856 * machine-independent and enhanced to also pre-fault zero-fill pages
1857 * (see vm.fast_fault) as well as make them writable, which greatly
1858 * reduces the number of page faults programs incur.
1859 *
1860 * Application performance when pre-faulting zero-fill pages is heavily
1861 * dependent on the application. Very tiny applications like /bin/echo
1862 * lose a little performance while applications of any appreciable size
1863 * gain performance. Prefaulting multiple pages also reduces SMP
1864 * congestion and can improve SMP performance significantly.
1865 *
1866 * NOTE! prot may allow writing but this only applies to the top level
1867 * object. If we wind up mapping a page extracted from a backing
1868 * object we have to make sure it is read-only.
1869 *
1870 * NOTE! The caller has already handled any COW operations on the
1871 * vm_map_entry via the normal fault code. Do NOT call this
1872 * shortcut unless the normal fault code has run on this entry.
1873 */
1874#define PFBAK 4
1875#define PFFOR 4
1876#define PAGEORDER_SIZE (PFBAK+PFFOR)
1877
1878static int vm_prefault_pageorder[] = {
1879 -PAGE_SIZE, PAGE_SIZE,
1880 -2 * PAGE_SIZE, 2 * PAGE_SIZE,
1881 -3 * PAGE_SIZE, 3 * PAGE_SIZE,
1882 -4 * PAGE_SIZE, 4 * PAGE_SIZE
1883};
1884
1885static void
1886vm_prefault(pmap_t pmap, vm_offset_t addra, vm_map_entry_t entry, int prot)
1887{
1888 struct lwp *lp;
1889 vm_page_t m;
1890 vm_offset_t starta;
1891 vm_offset_t addr;
1892 vm_pindex_t index;
1893 vm_pindex_t pindex;
1894 vm_object_t object;
1895 int pprot;
1896 int i;
1897
1898 /*
1899 * We do not currently prefault mappings that use virtual page
1900 * tables. We do not prefault foreign pmaps.
1901 */
1902 if (entry->maptype == VM_MAPTYPE_VPAGETABLE)
1903 return;
1904 lp = curthread->td_lwp;
1905 if (lp == NULL || (pmap != vmspace_pmap(lp->lwp_vmspace)))
1906 return;
1907
1908 object = entry->object.vm_object;
1909
1910 starta = addra - PFBAK * PAGE_SIZE;
1911 if (starta < entry->start)
1912 starta = entry->start;
1913 else if (starta > addra)
1914 starta = 0;
1915
1916 /*
1917 * critical section protection is required to maintain the
1918 * page/object association, interrupts can free pages and remove
1919 * them from their objects.
1920 */
1921 crit_enter();
1922 for (i = 0; i < PAGEORDER_SIZE; i++) {
1923 vm_object_t lobject;
3bb7eedb 1924 int allocated = 0;
1b9d3514
MD
1925
1926 addr = addra + vm_prefault_pageorder[i];
1927 if (addr > addra + (PFFOR * PAGE_SIZE))
1928 addr = 0;
1929
1930 if (addr < starta || addr >= entry->end)
1931 continue;
1932
1933 if (pmap_prefault_ok(pmap, addr) == 0)
1934 continue;
1935
1936 /*
1937 * Follow the VM object chain to obtain the page to be mapped
1938 * into the pmap.
1939 *
1940 * If we reach the terminal object without finding a page
1941 * and we determine it would be advantageous, then allocate
1942 * a zero-fill page for the base object. The base object
1943 * is guaranteed to be OBJT_DEFAULT for this case.
3bb7eedb
MD
1944 *
1945 * In order to not have to check the pager via *haspage*()
1946 * we stop if any non-default object is encountered. e.g.
1947 * a vnode or swap object would stop the loop.
1b9d3514
MD
1948 */
1949 index = ((addr - entry->start) + entry->offset) >> PAGE_SHIFT;
1950 lobject = object;
1951 pindex = index;
1952 pprot = prot;
1953
1954 while ((m = vm_page_lookup(lobject, pindex)) == NULL) {
1955 if (lobject->type != OBJT_DEFAULT)
1956 break;
1957 if (lobject->backing_object == NULL) {
1958 if (vm_fast_fault == 0)
1959 break;
1960 if (vm_prefault_pageorder[i] < 0 ||
1961 (prot & VM_PROT_WRITE) == 0 ||
1962 vm_page_count_min(0)) {
1963 break;
1964 }
3bb7eedb 1965 /* note: allocate from base object */
1b9d3514
MD
1966 m = vm_page_alloc(object, index,
1967 VM_ALLOC_NORMAL | VM_ALLOC_ZERO);
1968
1969 if ((m->flags & PG_ZERO) == 0) {
1970 vm_page_zero_fill(m);
1971 } else {
1972 vm_page_flag_clear(m, PG_ZERO);
1973 mycpu->gd_cnt.v_ozfod++;
1974 }
1975 mycpu->gd_cnt.v_zfod++;
1976 m->valid = VM_PAGE_BITS_ALL;
3bb7eedb 1977 allocated = 1;
1b9d3514
MD
1978 pprot = prot;
1979 /* lobject = object .. not needed */
1980 break;
1981 }
1982 if (lobject->backing_object_offset & PAGE_MASK)
1983 break;
1984 pindex += lobject->backing_object_offset >> PAGE_SHIFT;
1985 lobject = lobject->backing_object;
1986 pprot &= ~VM_PROT_WRITE;
1987 }
1988 /*
1989 * NOTE: lobject now invalid (if we did a zero-fill we didn't
1990 * bother assigning lobject = object).
1991 *
1992 * Give-up if the page is not available.
1993 */
1994 if (m == NULL)
1995 break;
1996
1997 /*
1998 * Do not conditionalize on PG_RAM. If pages are present in
1999 * the VM system we assume optimal caching. If caching is
2000 * not optimal the I/O gravy train will be restarted when we
2001 * hit an unavailable page. We do not want to try to restart
2002 * the gravy train now because we really don't know how much
2003 * of the object has been cached. The cost for restarting
2004 * the gravy train should be low (since accesses will likely
2005 * be I/O bound anyway).
2006 *
2007 * The object must be marked dirty if we are mapping a
2008 * writable page.
2009 */
2010 if (pprot & VM_PROT_WRITE)
2011 vm_object_set_writeable_dirty(m->object);
2012
2013 /*
3bb7eedb
MD
2014 * Enter the page into the pmap if appropriate. If we had
2015 * allocated the page we have to place it on a queue. If not
2016 * we just have to make sure it isn't on the cache queue
2017 * (pages on the cache queue are not allowed to be mapped).
1b9d3514 2018 */
3bb7eedb
MD
2019 if (allocated) {
2020 pmap_enter(pmap, addr, m, pprot, 0);
2021 vm_page_deactivate(m);
2022 vm_page_wakeup(m);
2023 } else if (((m->valid & VM_PAGE_BITS_ALL) == VM_PAGE_BITS_ALL) &&
1b9d3514
MD
2024 (m->busy == 0) &&
2025 (m->flags & (PG_BUSY | PG_FICTITIOUS)) == 0) {
2026
2027 if ((m->queue - m->pc) == PQ_CACHE) {
2028 vm_page_deactivate(m);
2029 }
2030 vm_page_busy(m);
2031 pmap_enter(pmap, addr, m, pprot, 0);
2032 vm_page_wakeup(m);
2033 }
2034 }
2035 crit_exit();
2036}