1 /*-
2 * Copyright (c) 1991 Regents of the University of California.
3 * All rights reserved.
4 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
5 *
6 * This code is derived from software contributed to Berkeley by
7 * The Mach Operating System project at Carnegie-Mellon University.
8 *
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
17 * 4. Neither the name of the University nor the names of its contributors
18 * may be used to endorse or promote products derived from this software
19 * without specific prior written permission.
20 *
21 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * SUCH DAMAGE.
32 *
33 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
34 */
35
36 /*-
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
39 *
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
41 *
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
47 *
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
51 *
52 * Carnegie Mellon requests users of this software to return to
53 *
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
58 *
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
61 */
62
63 /*
64 * GENERAL RULES ON VM_PAGE MANIPULATION
65 *
66 * - A page queue lock is required when adding or removing a page from a
67 * page queue regardless of other locks or the busy state of a page.
68 *
69 * * In general, no thread besides the page daemon can acquire or
70 * hold more than one page queue lock at a time.
71 *
72 * * The page daemon can acquire and hold any pair of page queue
73 * locks in any order.
74 *
75 * - The object lock is required when inserting or removing
76 * pages from an object (vm_page_insert() or vm_page_remove()).
77 *
78 */
79
80 /*
81 * Resident memory management module.
82 */
83
84 #include <sys/cdefs.h>
85 __FBSDID("$FreeBSD$");
86
87 #include "opt_vm.h"
88
89 #include <sys/param.h>
90 #include <sys/systm.h>
91 #include <sys/lock.h>
92 #include <sys/kernel.h>
93 #include <sys/limits.h>
94 #include <sys/linker.h>
95 #include <sys/malloc.h>
96 #include <sys/mman.h>
97 #include <sys/msgbuf.h>
98 #include <sys/mutex.h>
99 #include <sys/proc.h>
100 #include <sys/rwlock.h>
101 #include <sys/sbuf.h>
102 #include <sys/sysctl.h>
103 #include <sys/vmmeter.h>
104 #include <sys/vnode.h>
105
106 #include <vm/vm.h>
107 #include <vm/pmap.h>
108 #include <vm/vm_param.h>
109 #include <vm/vm_kern.h>
110 #include <vm/vm_object.h>
111 #include <vm/vm_page.h>
112 #include <vm/vm_pageout.h>
113 #include <vm/vm_pager.h>
114 #include <vm/vm_phys.h>
115 #include <vm/vm_radix.h>
116 #include <vm/vm_reserv.h>
117 #include <vm/vm_extern.h>
118 #include <vm/uma.h>
119 #include <vm/uma_int.h>
120
121 #include <machine/md_var.h>
122
123 /*
124 * Associated with page of user-allocatable memory is a
125 * page structure.
126 */
127
128 struct vm_domain vm_dom[MAXMEMDOM];
129 struct mtx_padalign vm_page_queue_free_mtx;
130
131 struct mtx_padalign pa_lock[PA_LOCK_COUNT];
132
133 vm_page_t vm_page_array;
134 long vm_page_array_size;
135 long first_page;
136 int vm_page_zero_count;
137
138 static int boot_pages = UMA_BOOT_PAGES;
139 SYSCTL_INT(_vm, OID_AUTO, boot_pages, CTLFLAG_RDTUN | CTLFLAG_NOFETCH,
140 &boot_pages, 0,
141 "number of pages allocated for bootstrapping the VM system");
142
143 static int pa_tryrelock_restart;
144 SYSCTL_INT(_vm, OID_AUTO, tryrelock_restart, CTLFLAG_RD,
145 &pa_tryrelock_restart, 0, "Number of tryrelock restarts");
146
147 static TAILQ_HEAD(, vm_page) blacklist_head;
148 static int sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS);
149 SYSCTL_PROC(_vm, OID_AUTO, page_blacklist, CTLTYPE_STRING | CTLFLAG_RD |
150 CTLFLAG_MPSAFE, NULL, 0, sysctl_vm_page_blacklist, "A", "Blacklist pages");
151
152 /* Is the page daemon waiting for free pages? */
153 static int vm_pageout_pages_needed;
154
155 static uma_zone_t fakepg_zone;
156
157 static struct vnode *vm_page_alloc_init(vm_page_t m);
158 static void vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits);
159 static void vm_page_enqueue(uint8_t queue, vm_page_t m);
160 static void vm_page_free_wakeup(void);
161 static void vm_page_init_fakepg(void *dummy);
162 static int vm_page_insert_after(vm_page_t m, vm_object_t object,
163 vm_pindex_t pindex, vm_page_t mpred);
164 static void vm_page_insert_radixdone(vm_page_t m, vm_object_t object,
165 vm_page_t mpred);
166 static int vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
167 vm_paddr_t high);
168
169 SYSINIT(vm_page, SI_SUB_VM, SI_ORDER_SECOND, vm_page_init_fakepg, NULL);
170
171 static void
vm_page_init_fakepg(void * dummy)172 vm_page_init_fakepg(void *dummy)
173 {
174
175 fakepg_zone = uma_zcreate("fakepg", sizeof(struct vm_page), NULL, NULL,
176 NULL, NULL, UMA_ALIGN_PTR, UMA_ZONE_NOFREE | UMA_ZONE_VM);
177 }
178
179 /* Make sure that u_long is at least 64 bits when PAGE_SIZE is 32K. */
180 #if PAGE_SIZE == 32768
181 #ifdef CTASSERT
182 CTASSERT(sizeof(u_long) >= 8);
183 #endif
184 #endif
185
186
187 struct vm_page_percpu {
188 struct mtx vpp_lock;
189 struct pglist vpp_pages;
190 int vpp_cnt;
191 } __aligned(CACHE_LINE_SIZE);
192
193 struct vm_page_percpu page_percpu[MAXCPU] __aligned(CACHE_LINE_SIZE);
194
195 #define VM_PERCPU_MIN 128
196 #define VM_PERCPU_TARGET (VM_PERCPU_MIN * 2)
197 #define VM_PERCPU_MAX (VM_PERCPU_MIN * 3)
198
199 static void
vm_page_percpu_init(void)200 vm_page_percpu_init(void)
201 {
202 int i;
203
204 for (i = 0; i < MAXCPU; i++) {
205 mtx_init(&page_percpu[i].vpp_lock, "per-cpu free mtx", NULL,
206 MTX_DEF);
207 TAILQ_INIT(&page_percpu[i].vpp_pages);
208 page_percpu[i].vpp_cnt = 0;
209 }
210 }
211
212 static vm_page_t
vm_page_percpu_alloc(vm_object_t object)213 vm_page_percpu_alloc(vm_object_t object)
214 {
215 struct vm_page_percpu *ppcpu = &page_percpu[PCPU_GET(cpuid)];
216 vm_page_t m;
217
218 #if VM_NRESERVLEVEL > 0
219 /*
220 * Skip the cache of free pages for objects that have reservations
221 * so that they can still get superpages. This will never be set
222 * for objects populated via the filesystem buffercache.
223 */
224 if (object != NULL && (object->flags & OBJ_COLORED) != 0)
225 return (NULL);
226 #endif
227
228 mtx_lock(&ppcpu->vpp_lock);
229 if (ppcpu->vpp_cnt < VM_PERCPU_MIN) {
230 mtx_lock(&vm_page_queue_free_mtx);
231 while (!vm_page_count_min() &&
232 ppcpu->vpp_cnt < VM_PERCPU_TARGET) {
233 m = vm_phys_alloc_pages(object != NULL ?
234 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
235 if (m == NULL)
236 break;
237 vm_phys_freecnt_adj(m, -1);
238 ppcpu->vpp_cnt++;
239 TAILQ_INSERT_TAIL(&ppcpu->vpp_pages, m, plinks.q);
240 }
241 mtx_unlock(&vm_page_queue_free_mtx);
242 }
243 m = NULL;
244 if (ppcpu->vpp_cnt > 0) {
245 m = TAILQ_FIRST(&ppcpu->vpp_pages);
246 TAILQ_REMOVE(&ppcpu->vpp_pages, m, plinks.q);
247 ppcpu->vpp_cnt--;
248 }
249 mtx_unlock(&ppcpu->vpp_lock);
250
251 return (m);
252 }
253
254 static inline void vm_page_free_wakeup(void);
255
256 static void
vm_page_percpu_free(vm_page_t m)257 vm_page_percpu_free(vm_page_t m)
258 {
259 struct vm_page_percpu *ppcpu = &page_percpu[PCPU_GET(cpuid)];
260
261 mtx_lock(&ppcpu->vpp_lock);
262 TAILQ_INSERT_HEAD(&ppcpu->vpp_pages, m, plinks.q);
263 ppcpu->vpp_cnt++;
264 if (ppcpu->vpp_cnt > VM_PERCPU_MAX) {
265 mtx_lock(&vm_page_queue_free_mtx);
266 while (ppcpu->vpp_cnt > VM_PERCPU_TARGET) {
267 m = TAILQ_FIRST(&ppcpu->vpp_pages);
268 TAILQ_REMOVE(&ppcpu->vpp_pages, m, plinks.q);
269 ppcpu->vpp_cnt--;
270 vm_phys_freecnt_adj(m, 1);
271 #if VM_NRESERVLEVEL > 0
272 if (!vm_reserv_free_page(m))
273 #else
274 if (TRUE)
275 #endif
276 vm_phys_free_pages(m, 0);
277 }
278 vm_page_free_wakeup();
279 mtx_unlock(&vm_page_queue_free_mtx);
280 }
281 mtx_unlock(&ppcpu->vpp_lock);
282 }
283
284
285 /*
286 * Try to acquire a physical address lock while a pmap is locked. If we
287 * fail to trylock we unlock and lock the pmap directly and cache the
288 * locked pa in *locked. The caller should then restart their loop in case
289 * the virtual to physical mapping has changed.
290 */
291 int
vm_page_pa_tryrelock(pmap_t pmap,vm_paddr_t pa,vm_paddr_t * locked)292 vm_page_pa_tryrelock(pmap_t pmap, vm_paddr_t pa, vm_paddr_t *locked)
293 {
294 vm_paddr_t lockpa;
295
296 lockpa = *locked;
297 *locked = pa;
298 if (lockpa) {
299 PA_LOCK_ASSERT(lockpa, MA_OWNED);
300 if (PA_LOCKPTR(pa) == PA_LOCKPTR(lockpa))
301 return (0);
302 PA_UNLOCK(lockpa);
303 }
304 if (PA_TRYLOCK(pa))
305 return (0);
306 PMAP_UNLOCK(pmap);
307 atomic_add_int(&pa_tryrelock_restart, 1);
308 PA_LOCK(pa);
309 PMAP_LOCK(pmap);
310 return (EAGAIN);
311 }
312
313 /*
314 * vm_set_page_size:
315 *
316 * Sets the page size, perhaps based upon the memory
317 * size. Must be called before any use of page-size
318 * dependent functions.
319 */
320 void
vm_set_page_size(void)321 vm_set_page_size(void)
322 {
323 if (vm_cnt.v_page_size == 0)
324 vm_cnt.v_page_size = PAGE_SIZE;
325 if (((vm_cnt.v_page_size - 1) & vm_cnt.v_page_size) != 0)
326 panic("vm_set_page_size: page size not a power of two");
327 }
328
329 /*
330 * vm_page_blacklist_next:
331 *
332 * Find the next entry in the provided string of blacklist
333 * addresses. Entries are separated by space, comma, or newline.
334 * If an invalid integer is encountered then the rest of the
335 * string is skipped. Updates the list pointer to the next
336 * character, or NULL if the string is exhausted or invalid.
337 */
338 static vm_paddr_t
vm_page_blacklist_next(char ** list,char * end)339 vm_page_blacklist_next(char **list, char *end)
340 {
341 vm_paddr_t bad;
342 char *cp, *pos;
343
344 if (list == NULL || *list == NULL)
345 return (0);
346 if (**list =='\0') {
347 *list = NULL;
348 return (0);
349 }
350
351 /*
352 * If there's no end pointer then the buffer is coming from
353 * the kenv and we know it's null-terminated.
354 */
355 if (end == NULL)
356 end = *list + strlen(*list);
357
358 /* Ensure that strtoq() won't walk off the end */
359 if (*end != '\0') {
360 if (*end == '\n' || *end == ' ' || *end == ',')
361 *end = '\0';
362 else {
363 printf("Blacklist not terminated, skipping\n");
364 *list = NULL;
365 return (0);
366 }
367 }
368
369 for (pos = *list; *pos != '\0'; pos = cp) {
370 bad = strtoq(pos, &cp, 0);
371 if (*cp == '\0' || *cp == ' ' || *cp == ',' || *cp == '\n') {
372 if (bad == 0) {
373 if (++cp < end)
374 continue;
375 else
376 break;
377 }
378 } else
379 break;
380 if (*cp == '\0' || ++cp >= end)
381 *list = NULL;
382 else
383 *list = cp;
384 return (trunc_page(bad));
385 }
386 printf("Garbage in RAM blacklist, skipping\n");
387 *list = NULL;
388 return (0);
389 }
390
391 /*
392 * vm_page_blacklist_check:
393 *
394 * Iterate through the provided string of blacklist addresses, pulling
395 * each entry out of the physical allocator free list and putting it
396 * onto a list for reporting via the vm.page_blacklist sysctl.
397 */
398 static void
vm_page_blacklist_check(char * list,char * end)399 vm_page_blacklist_check(char *list, char *end)
400 {
401 vm_paddr_t pa;
402 vm_page_t m;
403 char *next;
404 int ret;
405
406 next = list;
407 while (next != NULL) {
408 if ((pa = vm_page_blacklist_next(&next, end)) == 0)
409 continue;
410 m = vm_phys_paddr_to_vm_page(pa);
411 if (m == NULL)
412 continue;
413 mtx_lock(&vm_page_queue_free_mtx);
414 ret = vm_phys_unfree_page(m);
415 mtx_unlock(&vm_page_queue_free_mtx);
416 if (ret == TRUE) {
417 TAILQ_INSERT_TAIL(&blacklist_head, m, listq);
418 if (bootverbose)
419 printf("Skipping page with pa 0x%jx\n",
420 (uintmax_t)pa);
421 }
422 }
423 }
424
425 /*
426 * vm_page_blacklist_load:
427 *
428 * Search for a special module named "ram_blacklist". It'll be a
429 * plain text file provided by the user via the loader directive
430 * of the same name.
431 */
432 static void
vm_page_blacklist_load(char ** list,char ** end)433 vm_page_blacklist_load(char **list, char **end)
434 {
435 void *mod;
436 u_char *ptr;
437 u_int len;
438
439 mod = NULL;
440 ptr = NULL;
441
442 mod = preload_search_by_type("ram_blacklist");
443 if (mod != NULL) {
444 ptr = preload_fetch_addr(mod);
445 len = preload_fetch_size(mod);
446 }
447 *list = ptr;
448 if (ptr != NULL)
449 *end = ptr + len;
450 else
451 *end = NULL;
452 return;
453 }
454
455 static int
sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)456 sysctl_vm_page_blacklist(SYSCTL_HANDLER_ARGS)
457 {
458 vm_page_t m;
459 struct sbuf sbuf;
460 int error, first;
461
462 first = 1;
463 error = sysctl_wire_old_buffer(req, 0);
464 if (error != 0)
465 return (error);
466 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
467 TAILQ_FOREACH(m, &blacklist_head, listq) {
468 sbuf_printf(&sbuf, "%s%#jx", first ? "" : ",",
469 (uintmax_t)m->phys_addr);
470 first = 0;
471 }
472 error = sbuf_finish(&sbuf);
473 sbuf_delete(&sbuf);
474 return (error);
475 }
476
477 static void
vm_page_domain_init(struct vm_domain * vmd)478 vm_page_domain_init(struct vm_domain *vmd)
479 {
480 struct vm_pagequeue *pq;
481 int i;
482
483 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_name) =
484 "vm inactive pagequeue";
485 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_INACTIVE].pq_vcnt) =
486 &vm_cnt.v_inactive_count;
487 *__DECONST(char **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_name) =
488 "vm active pagequeue";
489 *__DECONST(int **, &vmd->vmd_pagequeues[PQ_ACTIVE].pq_vcnt) =
490 &vm_cnt.v_active_count;
491 vmd->vmd_page_count = 0;
492 vmd->vmd_free_count = 0;
493 vmd->vmd_segs = 0;
494 vmd->vmd_oom = FALSE;
495 vmd->vmd_pass = 0;
496 for (i = 0; i < PQ_COUNT + PA_LOCK_COUNT; i++) {
497 pq = &vmd->vmd_pagequeues[i];
498 TAILQ_INIT(&pq->pq_pl);
499 mtx_init(&pq->pq_mutex, pq->pq_name, "vm pagequeue",
500 MTX_DEF | MTX_DUPOK);
501 if (i >= PQ_COUNT) {
502 pq->pq_cnt = 0;
503 *__DECONST(char **, &vmd->vmd_pagequeues[i].pq_name) =
504 "vm inactive pagequeue";
505 *__DECONST(int **, &vmd->vmd_pagequeues[i].pq_vcnt) =
506 &vm_cnt.v_inactive_count;
507 }
508 }
509 }
510
511 #define PAQLENTHRESH_SMALL_LWM 4
512 #define PAQLENTHRESH_MEDIUM_LWM 16
513 #define PAQLENTHRESH_LARGE_LWM 64
514
515 #define PAQLENTHRESH_SMALL_HWM 6
516 #define PAQLENTHRESH_MEDIUM_HWM 24
517 #define PAQLENTHRESH_LARGE_HWM 80
518
519 #define VM_PAGES_SMALL (1<<18)
520 #define VM_PAGES_MEDIUM (1<<21)
521
522 static int vm_paqlenthresh_lwm;
523 static int vm_paqlenthresh_hwm;
524
525 /*
526 * vm_page_startup:
527 *
528 * Initializes the resident memory module.
529 *
530 * Allocates memory for the page cells, and
531 * for the object/offset-to-page hash table headers.
532 * Each page cell is initialized and placed on the free list.
533 */
534 vm_offset_t
vm_page_startup(vm_offset_t vaddr)535 vm_page_startup(vm_offset_t vaddr)
536 {
537 vm_offset_t mapped;
538 vm_paddr_t page_range;
539 vm_paddr_t new_end;
540 int i;
541 vm_paddr_t pa;
542 vm_paddr_t last_pa;
543 char *list, *listend;
544 vm_paddr_t end;
545 vm_paddr_t biggestsize;
546 vm_paddr_t low_water, high_water;
547 int biggestone;
548
549 biggestsize = 0;
550 biggestone = 0;
551 vaddr = round_page(vaddr);
552
553 for (i = 0; phys_avail[i + 1]; i += 2) {
554 phys_avail[i] = round_page(phys_avail[i]);
555 phys_avail[i + 1] = trunc_page(phys_avail[i + 1]);
556 }
557
558 low_water = phys_avail[0];
559 high_water = phys_avail[1];
560
561 for (i = 0; i < vm_phys_nsegs; i++) {
562 if (vm_phys_segs[i].start < low_water)
563 low_water = vm_phys_segs[i].start;
564 if (vm_phys_segs[i].end > high_water)
565 high_water = vm_phys_segs[i].end;
566 }
567 for (i = 0; phys_avail[i + 1]; i += 2) {
568 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
569
570 if (size > biggestsize) {
571 biggestone = i;
572 biggestsize = size;
573 }
574 if (phys_avail[i] < low_water)
575 low_water = phys_avail[i];
576 if (phys_avail[i + 1] > high_water)
577 high_water = phys_avail[i + 1];
578 }
579
580 end = phys_avail[biggestone+1];
581
582 /*
583 * Initialize the page and queue locks.
584 */
585 mtx_init(&vm_page_queue_free_mtx, "vm page free queue", NULL, MTX_DEF);
586 for (i = 0; i < PA_LOCK_COUNT; i++)
587 mtx_init(&pa_lock[i], "vm page", NULL, MTX_DEF);
588 for (i = 0; i < vm_ndomains; i++)
589 vm_page_domain_init(&vm_dom[i]);
590
591 /*
592 * Allocate memory for use when boot strapping the kernel memory
593 * allocator.
594 *
595 * CTFLAG_RDTUN doesn't work during the early boot process, so we must
596 * manually fetch the value.
597 */
598 TUNABLE_INT_FETCH("vm.boot_pages", &boot_pages);
599 new_end = end - (boot_pages * UMA_SLAB_SIZE);
600 new_end = trunc_page(new_end);
601 mapped = pmap_map(&vaddr, new_end, end,
602 VM_PROT_READ | VM_PROT_WRITE);
603 bzero((void *)mapped, end - new_end);
604 uma_startup((void *)mapped, boot_pages);
605
606 #if defined(__aarch64__) || defined(__amd64__) || defined(__arm__) || \
607 defined(__i386__) || defined(__mips__)
608 /*
609 * Allocate a bitmap to indicate that a random physical page
610 * needs to be included in a minidump.
611 *
612 * The amd64 port needs this to indicate which direct map pages
613 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
614 *
615 * However, i386 still needs this workspace internally within the
616 * minidump code. In theory, they are not needed on i386, but are
617 * included should the sf_buf code decide to use them.
618 */
619 last_pa = 0;
620 for (i = 0; dump_avail[i + 1] != 0; i += 2)
621 if (dump_avail[i + 1] > last_pa)
622 last_pa = dump_avail[i + 1];
623 page_range = last_pa / PAGE_SIZE;
624 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
625 new_end -= vm_page_dump_size;
626 vm_page_dump = (void *)(uintptr_t)pmap_map(&vaddr, new_end,
627 new_end + vm_page_dump_size, VM_PROT_READ | VM_PROT_WRITE);
628 bzero((void *)vm_page_dump, vm_page_dump_size);
629 #endif
630 #ifdef __amd64__
631 /*
632 * Request that the physical pages underlying the message buffer be
633 * included in a crash dump. Since the message buffer is accessed
634 * through the direct map, they are not automatically included.
635 */
636 pa = DMAP_TO_PHYS((vm_offset_t)msgbufp->msg_ptr);
637 last_pa = pa + round_page(msgbufsize);
638 while (pa < last_pa) {
639 dump_add_page(pa);
640 pa += PAGE_SIZE;
641 }
642 #endif
643 /*
644 * Compute the number of pages of memory that will be available for
645 * use (taking into account the overhead of a page structure per
646 * page).
647 */
648 first_page = low_water / PAGE_SIZE;
649 #ifdef VM_PHYSSEG_SPARSE
650 page_range = 0;
651 for (i = 0; i < vm_phys_nsegs; i++) {
652 page_range += atop(vm_phys_segs[i].end -
653 vm_phys_segs[i].start);
654 }
655 for (i = 0; phys_avail[i + 1] != 0; i += 2)
656 page_range += atop(phys_avail[i + 1] - phys_avail[i]);
657 #elif defined(VM_PHYSSEG_DENSE)
658 page_range = high_water / PAGE_SIZE - first_page;
659 #else
660 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
661 #endif
662 end = new_end;
663
664 /*
665 * Reserve an unmapped guard page to trap access to vm_page_array[-1].
666 */
667 vaddr += PAGE_SIZE;
668
669 /*
670 * Initialize the mem entry structures now, and put them in the free
671 * queue.
672 */
673 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
674 mapped = pmap_map(&vaddr, new_end, end,
675 VM_PROT_READ | VM_PROT_WRITE);
676 vm_page_array = (vm_page_t) mapped;
677 #if VM_NRESERVLEVEL > 0
678 /*
679 * Allocate memory for the reservation management system's data
680 * structures.
681 */
682 new_end = vm_reserv_startup(&vaddr, new_end, high_water);
683 #endif
684 #if defined(__aarch64__) || defined(__amd64__) || defined(__mips__)
685 /*
686 * pmap_map on arm64, amd64, and mips can come out of the direct-map,
687 * not kvm like i386, so the pages must be tracked for a crashdump to
688 * include this data. This includes the vm_page_array and the early
689 * UMA bootstrap pages.
690 */
691 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
692 dump_add_page(pa);
693 #endif
694 phys_avail[biggestone + 1] = new_end;
695
696 /*
697 * Add physical memory segments corresponding to the available
698 * physical pages.
699 */
700 for (i = 0; phys_avail[i + 1] != 0; i += 2)
701 vm_phys_add_seg(phys_avail[i], phys_avail[i + 1]);
702
703 /*
704 * Clear all of the page structures
705 */
706 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
707 for (i = 0; i < page_range; i++)
708 vm_page_array[i].order = VM_NFREEORDER;
709 vm_page_array_size = page_range;
710
711 /*
712 * Initialize the physical memory allocator.
713 */
714 vm_phys_init();
715
716 /*
717 * Add every available physical page that is not blacklisted to
718 * the free lists.
719 */
720 vm_cnt.v_page_count = 0;
721 vm_cnt.v_free_count = 0;
722 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
723 pa = phys_avail[i];
724 last_pa = phys_avail[i + 1];
725 while (pa < last_pa) {
726 vm_phys_add_page(pa);
727 pa += PAGE_SIZE;
728 }
729 }
730
731 TAILQ_INIT(&blacklist_head);
732 vm_page_blacklist_load(&list, &listend);
733 vm_page_blacklist_check(list, listend);
734
735 list = kern_getenv("vm.blacklist");
736 vm_page_blacklist_check(list, NULL);
737
738 freeenv(list);
739 #if VM_NRESERVLEVEL > 0
740 /*
741 * Initialize the reservation management system.
742 */
743 vm_reserv_init();
744 if (vm_page_array_size < VM_PAGES_SMALL) {
745 vm_paqlenthresh_lwm = PAQLENTHRESH_SMALL_LWM;
746 vm_paqlenthresh_hwm = PAQLENTHRESH_SMALL_HWM;
747 } else if (vm_page_array_size < VM_PAGES_MEDIUM) {
748 vm_paqlenthresh_lwm = PAQLENTHRESH_MEDIUM_LWM;
749 vm_paqlenthresh_hwm = PAQLENTHRESH_MEDIUM_HWM;
750 } else {
751 vm_paqlenthresh_lwm = PAQLENTHRESH_LARGE_LWM;
752 vm_paqlenthresh_hwm = PAQLENTHRESH_LARGE_HWM;
753 }
754 #endif
755 vm_page_percpu_init();
756 return (vaddr);
757 }
758
759 void
vm_page_reference(vm_page_t m)760 vm_page_reference(vm_page_t m)
761 {
762
763 vm_page_aflag_set(m, PGA_REFERENCED);
764 }
765
766 /*
767 * vm_page_busy_downgrade:
768 *
769 * Downgrade an exclusive busy page into a single shared busy page.
770 */
771 void
vm_page_busy_downgrade(vm_page_t m)772 vm_page_busy_downgrade(vm_page_t m)
773 {
774 u_int x;
775
776 vm_page_assert_xbusied(m);
777
778 for (;;) {
779 x = m->busy_lock;
780 x &= VPB_BIT_WAITERS;
781 if (atomic_cmpset_rel_int(&m->busy_lock,
782 VPB_SINGLE_EXCLUSIVER | x, VPB_SHARERS_WORD(1) | x))
783 break;
784 }
785 }
786
787 /*
788 * vm_page_sbusied:
789 *
790 * Return a positive value if the page is shared busied, 0 otherwise.
791 */
792 int
vm_page_sbusied(vm_page_t m)793 vm_page_sbusied(vm_page_t m)
794 {
795 u_int x;
796
797 x = m->busy_lock;
798 return ((x & VPB_BIT_SHARED) != 0 && x != VPB_UNBUSIED);
799 }
800
801 /*
802 * vm_page_sunbusy:
803 *
804 * Shared unbusy a page.
805 */
806 void
vm_page_sunbusy(vm_page_t m)807 vm_page_sunbusy(vm_page_t m)
808 {
809 u_int x;
810
811 vm_page_assert_sbusied(m);
812
813 for (;;) {
814 x = m->busy_lock;
815 if (VPB_SHARERS(x) > 1) {
816 if (atomic_cmpset_int(&m->busy_lock, x,
817 x - VPB_ONE_SHARER))
818 break;
819 continue;
820 }
821 if ((x & VPB_BIT_WAITERS) == 0) {
822 KASSERT(x == VPB_SHARERS_WORD(1),
823 ("vm_page_sunbusy: invalid lock state"));
824 if (atomic_cmpset_int(&m->busy_lock,
825 VPB_SHARERS_WORD(1), VPB_UNBUSIED))
826 break;
827 continue;
828 }
829 KASSERT(x == (VPB_SHARERS_WORD(1) | VPB_BIT_WAITERS),
830 ("vm_page_sunbusy: invalid lock state for waiters"));
831
832 vm_page_lock(m);
833 if (!atomic_cmpset_int(&m->busy_lock, x, VPB_UNBUSIED)) {
834 vm_page_unlock(m);
835 continue;
836 }
837 wakeup(m);
838 vm_page_unlock(m);
839 break;
840 }
841 }
842
843 /*
844 * vm_page_busy_sleep:
845 *
846 * Sleep and release the page lock, using the page pointer as wchan.
847 * This is used to implement the hard-path of busying mechanism.
848 *
849 * The given page must be locked.
850 */
851 void
vm_page_busy_sleep(vm_page_t m,const char * wmesg)852 vm_page_busy_sleep(vm_page_t m, const char *wmesg)
853 {
854 u_int x;
855
856 vm_page_lock_assert(m, MA_OWNED);
857
858 x = m->busy_lock;
859 if (x == VPB_UNBUSIED) {
860 vm_page_unlock(m);
861 return;
862 }
863 if ((x & VPB_BIT_WAITERS) == 0 &&
864 !atomic_cmpset_int(&m->busy_lock, x, x | VPB_BIT_WAITERS)) {
865 vm_page_unlock(m);
866 return;
867 }
868 msleep(m, vm_page_lockptr(m), PVM | PDROP, wmesg, 0);
869 }
870
871 /*
872 * vm_page_trysbusy:
873 *
874 * Try to shared busy a page.
875 * If the operation succeeds 1 is returned otherwise 0.
876 * The operation never sleeps.
877 */
878 int
vm_page_trysbusy(vm_page_t m)879 vm_page_trysbusy(vm_page_t m)
880 {
881 u_int x;
882
883 for (;;) {
884 x = m->busy_lock;
885 if ((x & VPB_BIT_SHARED) == 0)
886 return (0);
887 if (atomic_cmpset_acq_int(&m->busy_lock, x, x + VPB_ONE_SHARER))
888 return (1);
889 }
890 }
891
892 /*
893 * vm_page_xunbusy_hard:
894 *
895 * Called after the first try the exclusive unbusy of a page failed.
896 * It is assumed that the waiters bit is on.
897 */
898 void
vm_page_xunbusy_hard(vm_page_t m)899 vm_page_xunbusy_hard(vm_page_t m)
900 {
901
902 vm_page_assert_xbusied(m);
903
904 vm_page_lock(m);
905 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
906 wakeup(m);
907 vm_page_unlock(m);
908 }
909
910 /*
911 * vm_page_flash:
912 *
913 * Wakeup anyone waiting for the page.
914 * The ownership bits do not change.
915 *
916 * The given page must be locked.
917 */
918 void
vm_page_flash(vm_page_t m)919 vm_page_flash(vm_page_t m)
920 {
921 u_int x;
922
923 vm_page_lock_assert(m, MA_OWNED);
924
925 for (;;) {
926 x = m->busy_lock;
927 if ((x & VPB_BIT_WAITERS) == 0)
928 return;
929 if (atomic_cmpset_int(&m->busy_lock, x,
930 x & (~VPB_BIT_WAITERS)))
931 break;
932 }
933 wakeup(m);
934 }
935
936 /*
937 * Keep page from being freed by the page daemon
938 * much of the same effect as wiring, except much lower
939 * overhead and should be used only for *very* temporary
940 * holding ("wiring").
941 */
942 void
vm_page_hold(vm_page_t mem)943 vm_page_hold(vm_page_t mem)
944 {
945
946 vm_page_lock_assert(mem, MA_OWNED);
947 mem->hold_count++;
948 }
949
950 void
vm_page_unhold(vm_page_t mem)951 vm_page_unhold(vm_page_t mem)
952 {
953
954 vm_page_lock_assert(mem, MA_OWNED);
955 KASSERT(mem->hold_count >= 1, ("vm_page_unhold: hold count < 0!!!"));
956 --mem->hold_count;
957 if (mem->hold_count == 0 && (mem->flags & PG_UNHOLDFREE) != 0)
958 vm_page_free_toq(mem);
959 }
960
961 /*
962 * vm_page_unhold_pages:
963 *
964 * Unhold each of the pages that is referenced by the given array.
965 */
966 void
vm_page_unhold_pages(vm_page_t * ma,int count)967 vm_page_unhold_pages(vm_page_t *ma, int count)
968 {
969 struct mtx *mtx, *new_mtx;
970
971 mtx = NULL;
972 for (; count != 0; count--) {
973 /*
974 * Avoid releasing and reacquiring the same page lock.
975 */
976 new_mtx = vm_page_lockptr(*ma);
977 if (mtx != new_mtx) {
978 if (mtx != NULL)
979 mtx_unlock(mtx);
980 mtx = new_mtx;
981 mtx_lock(mtx);
982 }
983 vm_page_unhold(*ma);
984 ma++;
985 }
986 if (mtx != NULL)
987 mtx_unlock(mtx);
988 }
989
990 vm_page_t
PHYS_TO_VM_PAGE(vm_paddr_t pa)991 PHYS_TO_VM_PAGE(vm_paddr_t pa)
992 {
993 vm_page_t m;
994
995 #ifdef VM_PHYSSEG_SPARSE
996 m = vm_phys_paddr_to_vm_page(pa);
997 if (m == NULL)
998 m = vm_phys_fictitious_to_vm_page(pa);
999 return (m);
1000 #elif defined(VM_PHYSSEG_DENSE)
1001 long pi;
1002
1003 pi = atop(pa);
1004 if (pi >= first_page && (pi - first_page) < vm_page_array_size) {
1005 m = &vm_page_array[pi - first_page];
1006 return (m);
1007 }
1008 return (vm_phys_fictitious_to_vm_page(pa));
1009 #else
1010 #error "Either VM_PHYSSEG_DENSE or VM_PHYSSEG_SPARSE must be defined."
1011 #endif
1012 }
1013
1014 /*
1015 * vm_page_getfake:
1016 *
1017 * Create a fictitious page with the specified physical address and
1018 * memory attribute. The memory attribute is the only the machine-
1019 * dependent aspect of a fictitious page that must be initialized.
1020 */
1021 vm_page_t
vm_page_getfake(vm_paddr_t paddr,vm_memattr_t memattr)1022 vm_page_getfake(vm_paddr_t paddr, vm_memattr_t memattr)
1023 {
1024 vm_page_t m;
1025
1026 m = uma_zalloc(fakepg_zone, M_WAITOK | M_ZERO);
1027 vm_page_initfake(m, paddr, memattr);
1028 return (m);
1029 }
1030
1031 void
vm_page_initfake(vm_page_t m,vm_paddr_t paddr,vm_memattr_t memattr)1032 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1033 {
1034
1035 if ((m->flags & PG_FICTITIOUS) != 0) {
1036 /*
1037 * The page's memattr might have changed since the
1038 * previous initialization. Update the pmap to the
1039 * new memattr.
1040 */
1041 goto memattr;
1042 }
1043 m->phys_addr = paddr;
1044 m->queue = PQ_NONE;
1045 /* Fictitious pages don't use "segind". */
1046 m->flags = PG_FICTITIOUS;
1047 /* Fictitious pages don't use "order" or "pool". */
1048 m->oflags = VPO_UNMANAGED;
1049 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1050 m->wire_count = 1;
1051 pmap_page_init(m);
1052 memattr:
1053 pmap_page_set_memattr(m, memattr);
1054 }
1055
1056 /*
1057 * vm_page_putfake:
1058 *
1059 * Release a fictitious page.
1060 */
1061 void
vm_page_putfake(vm_page_t m)1062 vm_page_putfake(vm_page_t m)
1063 {
1064
1065 KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed %p", m));
1066 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1067 ("vm_page_putfake: bad page %p", m));
1068 uma_zfree(fakepg_zone, m);
1069 }
1070
1071 /*
1072 * vm_page_updatefake:
1073 *
1074 * Update the given fictitious page to the specified physical address and
1075 * memory attribute.
1076 */
1077 void
vm_page_updatefake(vm_page_t m,vm_paddr_t paddr,vm_memattr_t memattr)1078 vm_page_updatefake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1079 {
1080
1081 KASSERT((m->flags & PG_FICTITIOUS) != 0,
1082 ("vm_page_updatefake: bad page %p", m));
1083 m->phys_addr = paddr;
1084 pmap_page_set_memattr(m, memattr);
1085 }
1086
1087 /*
1088 * vm_page_free:
1089 *
1090 * Free a page.
1091 */
1092 void
vm_page_free(vm_page_t m)1093 vm_page_free(vm_page_t m)
1094 {
1095
1096 m->flags &= ~PG_ZERO;
1097 vm_page_free_toq(m);
1098 }
1099
1100 /*
1101 * vm_page_free_zero:
1102 *
1103 * Free a page to the zerod-pages queue
1104 */
1105 void
vm_page_free_zero(vm_page_t m)1106 vm_page_free_zero(vm_page_t m)
1107 {
1108
1109 m->flags |= PG_ZERO;
1110 vm_page_free_toq(m);
1111 }
1112
1113 /*
1114 * Unbusy and handle the page queueing for a page from the VOP_GETPAGES()
1115 * array which was optionally read ahead or behind.
1116 */
1117 void
vm_page_readahead_finish(vm_page_t m)1118 vm_page_readahead_finish(vm_page_t m)
1119 {
1120
1121 /* We shouldn't put invalid pages on queues. */
1122 KASSERT(m->valid != 0, ("%s: %p is invalid", __func__, m));
1123
1124 /*
1125 * Since the page is not the actually needed one, whether it should
1126 * be activated or deactivated is not obvious. Empirical results
1127 * have shown that deactivating the page is usually the best choice,
1128 * unless the page is wanted by another thread.
1129 */
1130 vm_page_lock(m);
1131 if ((m->busy_lock & VPB_BIT_WAITERS) != 0)
1132 vm_page_activate(m);
1133 else
1134 vm_page_deactivate(m);
1135 vm_page_unlock(m);
1136 vm_page_xunbusy(m);
1137 }
1138
1139 /*
1140 * vm_page_sleep_if_busy:
1141 *
1142 * Sleep and release the page queues lock if the page is busied.
1143 * Returns TRUE if the thread slept.
1144 *
1145 * The given page must be unlocked and object containing it must
1146 * be locked.
1147 */
1148 int
vm_page_sleep_if_busy(vm_page_t m,const char * msg)1149 vm_page_sleep_if_busy(vm_page_t m, const char *msg)
1150 {
1151 vm_object_t obj;
1152
1153 vm_page_lock_assert(m, MA_NOTOWNED);
1154 VM_OBJECT_ASSERT_WLOCKED(m->object);
1155
1156 if (vm_page_busied(m)) {
1157 /*
1158 * The page-specific object must be cached because page
1159 * identity can change during the sleep, causing the
1160 * re-lock of a different object.
1161 * It is assumed that a reference to the object is already
1162 * held by the callers.
1163 */
1164 obj = m->object;
1165 vm_page_lock(m);
1166 VM_OBJECT_WUNLOCK(obj);
1167 vm_page_busy_sleep(m, msg);
1168 VM_OBJECT_WLOCK(obj);
1169 return (TRUE);
1170 }
1171 return (FALSE);
1172 }
1173
1174 /*
1175 * vm_page_dirty_KBI: [ internal use only ]
1176 *
1177 * Set all bits in the page's dirty field.
1178 *
1179 * The object containing the specified page must be locked if the
1180 * call is made from the machine-independent layer.
1181 *
1182 * See vm_page_clear_dirty_mask().
1183 *
1184 * This function should only be called by vm_page_dirty().
1185 */
1186 void
vm_page_dirty_KBI(vm_page_t m)1187 vm_page_dirty_KBI(vm_page_t m)
1188 {
1189
1190 /* These assertions refer to this operation by its public name. */
1191 KASSERT(m->valid == VM_PAGE_BITS_ALL,
1192 ("vm_page_dirty: page is invalid!"));
1193 m->dirty = VM_PAGE_BITS_ALL;
1194 }
1195
1196 /*
1197 * vm_page_insert: [ internal use only ]
1198 *
1199 * Inserts the given mem entry into the object and object list.
1200 *
1201 * The object must be locked.
1202 */
1203 int
vm_page_insert(vm_page_t m,vm_object_t object,vm_pindex_t pindex)1204 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1205 {
1206 vm_page_t mpred;
1207
1208 VM_OBJECT_ASSERT_WLOCKED(object);
1209 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1210 return (vm_page_insert_after(m, object, pindex, mpred));
1211 }
1212
1213 /*
1214 * vm_page_insert_after:
1215 *
1216 * Inserts the page "m" into the specified object at offset "pindex".
1217 *
1218 * The page "mpred" must immediately precede the offset "pindex" within
1219 * the specified object.
1220 *
1221 * The object must be locked.
1222 */
1223 static int
vm_page_insert_after(vm_page_t m,vm_object_t object,vm_pindex_t pindex,vm_page_t mpred)1224 vm_page_insert_after(vm_page_t m, vm_object_t object, vm_pindex_t pindex,
1225 vm_page_t mpred)
1226 {
1227 vm_pindex_t sidx;
1228 vm_object_t sobj;
1229 vm_page_t msucc;
1230
1231 VM_OBJECT_ASSERT_WLOCKED(object);
1232 KASSERT(m->object == NULL,
1233 ("vm_page_insert_after: page already inserted"));
1234 if (mpred != NULL) {
1235 KASSERT(mpred->object == object,
1236 ("vm_page_insert_after: object doesn't contain mpred"));
1237 KASSERT(mpred->pindex < pindex,
1238 ("vm_page_insert_after: mpred doesn't precede pindex"));
1239 msucc = TAILQ_NEXT(mpred, listq);
1240 } else
1241 msucc = TAILQ_FIRST(&object->memq);
1242 if (msucc != NULL)
1243 KASSERT(msucc->pindex > pindex,
1244 ("vm_page_insert_after: msucc doesn't succeed pindex"));
1245
1246 /*
1247 * Record the object/offset pair in this page
1248 */
1249 sobj = m->object;
1250 sidx = m->pindex;
1251 m->object = object;
1252 m->pindex = pindex;
1253
1254 /*
1255 * Now link into the object's ordered list of backed pages.
1256 */
1257 if (vm_radix_insert(&object->rtree, m)) {
1258 m->object = sobj;
1259 m->pindex = sidx;
1260 return (1);
1261 }
1262 vm_page_insert_radixdone(m, object, mpred);
1263 return (0);
1264 }
1265
1266 /*
1267 * vm_page_insert_radixdone:
1268 *
1269 * Complete page "m" insertion into the specified object after the
1270 * radix trie hooking.
1271 *
1272 * The page "mpred" must precede the offset "m->pindex" within the
1273 * specified object.
1274 *
1275 * The object must be locked.
1276 */
1277 static void
vm_page_insert_radixdone(vm_page_t m,vm_object_t object,vm_page_t mpred)1278 vm_page_insert_radixdone(vm_page_t m, vm_object_t object, vm_page_t mpred)
1279 {
1280
1281 VM_OBJECT_ASSERT_WLOCKED(object);
1282 KASSERT(object != NULL && m->object == object,
1283 ("vm_page_insert_radixdone: page %p has inconsistent object", m));
1284 if (mpred != NULL) {
1285 KASSERT(mpred->object == object,
1286 ("vm_page_insert_after: object doesn't contain mpred"));
1287 KASSERT(mpred->pindex < m->pindex,
1288 ("vm_page_insert_after: mpred doesn't precede pindex"));
1289 }
1290
1291 if (mpred != NULL)
1292 TAILQ_INSERT_AFTER(&object->memq, mpred, m, listq);
1293 else
1294 TAILQ_INSERT_HEAD(&object->memq, m, listq);
1295
1296 /*
1297 * Show that the object has one more resident page.
1298 */
1299 object->resident_page_count++;
1300
1301 /*
1302 * Hold the vnode until the last page is released.
1303 */
1304 if (object->resident_page_count == 1 && object->type == OBJT_VNODE)
1305 vhold(object->handle);
1306
1307 /*
1308 * Since we are inserting a new and possibly dirty page,
1309 * update the object's OBJ_MIGHTBEDIRTY flag.
1310 */
1311 if (pmap_page_is_write_mapped(m))
1312 vm_object_set_writeable_dirty(object);
1313 }
1314
1315 /*
1316 * vm_page_remove:
1317 *
1318 * Removes the given mem entry from the object/offset-page
1319 * table and the object page list, but do not invalidate/terminate
1320 * the backing store.
1321 *
1322 * The object must be locked. The page must be locked if it is managed.
1323 */
1324 void
vm_page_remove(vm_page_t m)1325 vm_page_remove(vm_page_t m)
1326 {
1327 vm_object_t object;
1328 boolean_t lockacq;
1329
1330 if ((m->oflags & VPO_UNMANAGED) == 0)
1331 vm_page_lock_assert(m, MA_OWNED);
1332 if ((object = m->object) == NULL)
1333 return;
1334 VM_OBJECT_ASSERT_WLOCKED(object);
1335 if (vm_page_xbusied(m)) {
1336 lockacq = FALSE;
1337 if ((m->oflags & VPO_UNMANAGED) != 0 &&
1338 !mtx_owned(vm_page_lockptr(m))) {
1339 lockacq = TRUE;
1340 vm_page_lock(m);
1341 }
1342 vm_page_flash(m);
1343 atomic_store_rel_int(&m->busy_lock, VPB_UNBUSIED);
1344 if (lockacq)
1345 vm_page_unlock(m);
1346 }
1347
1348 /*
1349 * Now remove from the object's list of backed pages.
1350 */
1351 vm_radix_remove(&object->rtree, m->pindex);
1352 TAILQ_REMOVE(&object->memq, m, listq);
1353
1354 /*
1355 * And show that the object has one fewer resident page.
1356 */
1357 object->resident_page_count--;
1358
1359 /*
1360 * The vnode may now be recycled.
1361 */
1362 if (object->resident_page_count == 0 && object->type == OBJT_VNODE)
1363 vdrop(object->handle);
1364
1365 m->object = NULL;
1366 }
1367
1368 /*
1369 * vm_page_lookup:
1370 *
1371 * Returns the page associated with the object/offset
1372 * pair specified; if none is found, NULL is returned.
1373 *
1374 * The object must be locked.
1375 */
1376 vm_page_t
vm_page_lookup(vm_object_t object,vm_pindex_t pindex)1377 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1378 {
1379
1380 VM_OBJECT_ASSERT_LOCKED(object);
1381 return (vm_radix_lookup(&object->rtree, pindex));
1382 }
1383
1384 /*
1385 * vm_page_find_least:
1386 *
1387 * Returns the page associated with the object with least pindex
1388 * greater than or equal to the parameter pindex, or NULL.
1389 *
1390 * The object must be locked.
1391 */
1392 vm_page_t
vm_page_find_least(vm_object_t object,vm_pindex_t pindex)1393 vm_page_find_least(vm_object_t object, vm_pindex_t pindex)
1394 {
1395 vm_page_t m;
1396
1397 VM_OBJECT_ASSERT_LOCKED(object);
1398 if ((m = TAILQ_FIRST(&object->memq)) != NULL && m->pindex < pindex)
1399 m = vm_radix_lookup_ge(&object->rtree, pindex);
1400 return (m);
1401 }
1402
1403 /*
1404 * Returns the given page's successor (by pindex) within the object if it is
1405 * resident; if none is found, NULL is returned.
1406 *
1407 * The object must be locked.
1408 */
1409 vm_page_t
vm_page_next(vm_page_t m)1410 vm_page_next(vm_page_t m)
1411 {
1412 vm_page_t next;
1413
1414 VM_OBJECT_ASSERT_WLOCKED(m->object);
1415 if ((next = TAILQ_NEXT(m, listq)) != NULL &&
1416 next->pindex != m->pindex + 1)
1417 next = NULL;
1418 return (next);
1419 }
1420
1421 /*
1422 * Returns the given page's predecessor (by pindex) within the object if it is
1423 * resident; if none is found, NULL is returned.
1424 *
1425 * The object must be locked.
1426 */
1427 vm_page_t
vm_page_prev(vm_page_t m)1428 vm_page_prev(vm_page_t m)
1429 {
1430 vm_page_t prev;
1431
1432 VM_OBJECT_ASSERT_WLOCKED(m->object);
1433 if ((prev = TAILQ_PREV(m, pglist, listq)) != NULL &&
1434 prev->pindex != m->pindex - 1)
1435 prev = NULL;
1436 return (prev);
1437 }
1438
1439 /*
1440 * Uses the page mnew as a replacement for an existing page at index
1441 * pindex which must be already present in the object.
1442 *
1443 * The existing page must not be on a paging queue.
1444 */
1445 vm_page_t
vm_page_replace(vm_page_t mnew,vm_object_t object,vm_pindex_t pindex)1446 vm_page_replace(vm_page_t mnew, vm_object_t object, vm_pindex_t pindex)
1447 {
1448 vm_page_t mold;
1449
1450 VM_OBJECT_ASSERT_WLOCKED(object);
1451 KASSERT(mnew->object == NULL,
1452 ("vm_page_replace: page already in object"));
1453
1454 /*
1455 * This function mostly follows vm_page_insert() and
1456 * vm_page_remove() without the radix, object count and vnode
1457 * dance. Double check such functions for more comments.
1458 */
1459
1460 mnew->object = object;
1461 mnew->pindex = pindex;
1462 mold = vm_radix_replace(&object->rtree, mnew);
1463 KASSERT(mold->queue == PQ_NONE,
1464 ("vm_page_replace: mold is on a paging queue"));
1465
1466 /* Keep the resident page list in sorted order. */
1467 TAILQ_INSERT_AFTER(&object->memq, mold, mnew, listq);
1468 TAILQ_REMOVE(&object->memq, mold, listq);
1469
1470 mold->object = NULL;
1471 vm_page_xunbusy(mold);
1472
1473 /*
1474 * The object's resident_page_count does not change because we have
1475 * swapped one page for another, but OBJ_MIGHTBEDIRTY.
1476 */
1477 if (pmap_page_is_write_mapped(mnew))
1478 vm_object_set_writeable_dirty(object);
1479 return (mold);
1480 }
1481
1482 /*
1483 * vm_page_rename:
1484 *
1485 * Move the given memory entry from its
1486 * current object to the specified target object/offset.
1487 *
1488 * Note: swap associated with the page must be invalidated by the move. We
1489 * have to do this for several reasons: (1) we aren't freeing the
1490 * page, (2) we are dirtying the page, (3) the VM system is probably
1491 * moving the page from object A to B, and will then later move
1492 * the backing store from A to B and we can't have a conflict.
1493 *
1494 * Note: we *always* dirty the page. It is necessary both for the
1495 * fact that we moved it, and because we may be invalidating
1496 * swap. If the page is on the cache, we have to deactivate it
1497 * or vm_page_dirty() will panic. Dirty pages are not allowed
1498 * on the cache.
1499 *
1500 * The objects must be locked.
1501 */
1502 int
vm_page_rename(vm_page_t m,vm_object_t new_object,vm_pindex_t new_pindex)1503 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1504 {
1505 vm_page_t mpred;
1506 vm_pindex_t opidx;
1507
1508 VM_OBJECT_ASSERT_WLOCKED(new_object);
1509
1510 mpred = vm_radix_lookup_le(&new_object->rtree, new_pindex);
1511 KASSERT(mpred == NULL || mpred->pindex != new_pindex,
1512 ("vm_page_rename: pindex already renamed"));
1513
1514 /*
1515 * Create a custom version of vm_page_insert() which does not depend
1516 * by m_prev and can cheat on the implementation aspects of the
1517 * function.
1518 */
1519 opidx = m->pindex;
1520 m->pindex = new_pindex;
1521 if (vm_radix_insert(&new_object->rtree, m)) {
1522 m->pindex = opidx;
1523 return (1);
1524 }
1525
1526 /*
1527 * The operation cannot fail anymore. The removal must happen before
1528 * the listq iterator is tainted.
1529 */
1530 m->pindex = opidx;
1531 vm_page_lock(m);
1532 vm_page_remove(m);
1533
1534 /* Return back to the new pindex to complete vm_page_insert(). */
1535 m->pindex = new_pindex;
1536 m->object = new_object;
1537 vm_page_unlock(m);
1538 vm_page_insert_radixdone(m, new_object, mpred);
1539 vm_page_dirty(m);
1540 return (0);
1541 }
1542
1543 /*
1544 * vm_page_alloc:
1545 *
1546 * Allocate and return a page that is associated with the specified
1547 * object and offset pair. By default, this page is exclusive busied.
1548 *
1549 * The caller must always specify an allocation class.
1550 *
1551 * allocation classes:
1552 * VM_ALLOC_NORMAL normal process request
1553 * VM_ALLOC_SYSTEM system *really* needs a page
1554 * VM_ALLOC_INTERRUPT interrupt time request
1555 *
1556 * optional allocation flags:
1557 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1558 * intends to allocate
1559 * VM_ALLOC_NOBUSY do not exclusive busy the page
1560 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1561 * VM_ALLOC_NOOBJ page is not associated with an object and
1562 * should not be exclusive busy
1563 * VM_ALLOC_SBUSY shared busy the allocated page
1564 * VM_ALLOC_WIRED wire the allocated page
1565 * VM_ALLOC_ZERO prefer a zeroed page
1566 *
1567 * This routine may not sleep.
1568 */
1569 vm_page_t
vm_page_alloc(vm_object_t object,vm_pindex_t pindex,int req)1570 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int req)
1571 {
1572 struct vnode *vp = NULL;
1573 vm_page_t m, mpred;
1574 int flags, req_class;
1575
1576 mpred = 0; /* XXX: pacify gcc */
1577 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1578 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1579 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1580 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1581 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1582 req));
1583 if (object != NULL)
1584 VM_OBJECT_ASSERT_WLOCKED(object);
1585
1586 req_class = req & VM_ALLOC_CLASS_MASK;
1587
1588 /*
1589 * The page daemon is allowed to dig deeper into the free page list.
1590 */
1591 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1592 req_class = VM_ALLOC_SYSTEM;
1593
1594 if (object != NULL) {
1595 mpred = vm_radix_lookup_le(&object->rtree, pindex);
1596 KASSERT(mpred == NULL || mpred->pindex != pindex,
1597 ("vm_page_alloc: pindex already allocated"));
1598 }
1599
1600 if ((m = vm_page_percpu_alloc(object)) != NULL) {
1601 flags = 0;
1602 goto gotit;
1603 }
1604 /*
1605 * The page allocation request can came from consumers which already
1606 * hold the free page queue mutex.
1607 */
1608 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
1609 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1610 (req_class == VM_ALLOC_SYSTEM &&
1611 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
1612 (req_class == VM_ALLOC_INTERRUPT &&
1613 vm_cnt.v_free_count > 0)) {
1614 /*
1615 * Allocate from the free queue if the number of free pages
1616 * exceeds the minimum for the request class.
1617 */
1618 #if VM_NRESERVLEVEL > 0
1619 if (object == NULL || (object->flags & (OBJ_COLORED |
1620 OBJ_FICTITIOUS)) != OBJ_COLORED || (m =
1621 vm_reserv_alloc_page(object, pindex, mpred)) == NULL) {
1622 #endif
1623 m = vm_phys_alloc_pages(object != NULL ?
1624 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT, 0);
1625 #if VM_NRESERVLEVEL > 0
1626 if (m == NULL && vm_reserv_reclaim_inactive()) {
1627 m = vm_phys_alloc_pages(object != NULL ?
1628 VM_FREEPOOL_DEFAULT : VM_FREEPOOL_DIRECT,
1629 0);
1630 }
1631 }
1632 #endif
1633 } else {
1634 /*
1635 * Not allocatable, give up.
1636 */
1637 mtx_unlock(&vm_page_queue_free_mtx);
1638 atomic_add_int(&vm_pageout_deficit,
1639 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
1640 pagedaemon_wakeup();
1641 return (NULL);
1642 }
1643
1644 /*
1645 * At this point we had better have found a good page.
1646 */
1647 KASSERT(m != NULL, ("vm_page_alloc: missing page"));
1648 KASSERT(m->queue == PQ_NONE,
1649 ("vm_page_alloc: page %p has unexpected queue %d", m, m->queue));
1650 KASSERT(m->wire_count == 0, ("vm_page_alloc: page %p is wired", m));
1651 KASSERT(m->hold_count == 0, ("vm_page_alloc: page %p is held", m));
1652 KASSERT(!vm_page_sbusied(m),
1653 ("vm_page_alloc: page %p is busy", m));
1654 KASSERT(m->dirty == 0, ("vm_page_alloc: page %p is dirty", m));
1655 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1656 ("vm_page_alloc: page %p has unexpected memattr %d", m,
1657 pmap_page_get_memattr(m)));
1658 KASSERT(m->valid == 0,
1659 ("vm_page_alloc: free page %p is valid", m));
1660 vm_phys_freecnt_adj(m, -1);
1661 if ((m->flags & PG_ZERO) != 0)
1662 vm_page_zero_count--;
1663 mtx_unlock(&vm_page_queue_free_mtx);
1664
1665 /*
1666 * Initialize the page. Only the PG_ZERO flag is inherited.
1667 */
1668 flags = 0;
1669 if ((req & VM_ALLOC_ZERO) != 0)
1670 flags = PG_ZERO;
1671 flags &= m->flags;
1672 gotit:
1673 if ((req & VM_ALLOC_NODUMP) != 0)
1674 flags |= PG_NODUMP;
1675 m->flags = flags;
1676 m->aflags = 0;
1677 m->oflags = object == NULL || (object->flags & OBJ_UNMANAGED) != 0 ?
1678 VPO_UNMANAGED : 0;
1679 m->busy_lock = VPB_UNBUSIED;
1680 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_NOOBJ | VM_ALLOC_SBUSY)) == 0)
1681 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1682 if ((req & VM_ALLOC_SBUSY) != 0)
1683 m->busy_lock = VPB_SHARERS_WORD(1);
1684 if (req & VM_ALLOC_WIRED) {
1685 /*
1686 * The page lock is not required for wiring a page until that
1687 * page is inserted into the object.
1688 */
1689 atomic_add_int(&vm_cnt.v_wire_count, 1);
1690 m->wire_count = 1;
1691 }
1692 m->act_count = 0;
1693
1694 if (object != NULL) {
1695 if (vm_page_insert_after(m, object, pindex, mpred)) {
1696 /* See the comment below about hold count. */
1697 if (vp != NULL)
1698 vdrop(vp);
1699 pagedaemon_wakeup();
1700 if (req & VM_ALLOC_WIRED) {
1701 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
1702 m->wire_count = 0;
1703 }
1704 m->object = NULL;
1705 m->oflags = VPO_UNMANAGED;
1706 vm_page_free(m);
1707 return (NULL);
1708 }
1709
1710 /* Ignore device objects; the pager sets "memattr" for them. */
1711 if (object->memattr != VM_MEMATTR_DEFAULT &&
1712 (object->flags & OBJ_FICTITIOUS) == 0)
1713 pmap_page_set_memattr(m, object->memattr);
1714 } else
1715 m->pindex = pindex;
1716
1717 /*
1718 * The following call to vdrop() must come after the above call
1719 * to vm_page_insert() in case both affect the same object and
1720 * vnode. Otherwise, the affected vnode's hold count could
1721 * temporarily become zero.
1722 */
1723 if (vp != NULL)
1724 vdrop(vp);
1725
1726 /*
1727 * Don't wakeup too often - wakeup the pageout daemon when
1728 * we would be nearly out of memory.
1729 */
1730 if (vm_paging_needed())
1731 pagedaemon_wakeup();
1732
1733 return (m);
1734 }
1735
1736 static void
vm_page_alloc_contig_vdrop(struct spglist * lst)1737 vm_page_alloc_contig_vdrop(struct spglist *lst)
1738 {
1739
1740 while (!SLIST_EMPTY(lst)) {
1741 vdrop((struct vnode *)SLIST_FIRST(lst)-> plinks.s.pv);
1742 SLIST_REMOVE_HEAD(lst, plinks.s.ss);
1743 }
1744 }
1745
1746 /*
1747 * vm_page_alloc_contig:
1748 *
1749 * Allocate a contiguous set of physical pages of the given size "npages"
1750 * from the free lists. All of the physical pages must be at or above
1751 * the given physical address "low" and below the given physical address
1752 * "high". The given value "alignment" determines the alignment of the
1753 * first physical page in the set. If the given value "boundary" is
1754 * non-zero, then the set of physical pages cannot cross any physical
1755 * address boundary that is a multiple of that value. Both "alignment"
1756 * and "boundary" must be a power of two.
1757 *
1758 * If the specified memory attribute, "memattr", is VM_MEMATTR_DEFAULT,
1759 * then the memory attribute setting for the physical pages is configured
1760 * to the object's memory attribute setting. Otherwise, the memory
1761 * attribute setting for the physical pages is configured to "memattr",
1762 * overriding the object's memory attribute setting. However, if the
1763 * object's memory attribute setting is not VM_MEMATTR_DEFAULT, then the
1764 * memory attribute setting for the physical pages cannot be configured
1765 * to VM_MEMATTR_DEFAULT.
1766 *
1767 * The caller must always specify an allocation class.
1768 *
1769 * allocation classes:
1770 * VM_ALLOC_NORMAL normal process request
1771 * VM_ALLOC_SYSTEM system *really* needs a page
1772 * VM_ALLOC_INTERRUPT interrupt time request
1773 *
1774 * optional allocation flags:
1775 * VM_ALLOC_NOBUSY do not exclusive busy the page
1776 * VM_ALLOC_NODUMP do not include the page in a kernel core dump
1777 * VM_ALLOC_NOOBJ page is not associated with an object and
1778 * should not be exclusive busy
1779 * VM_ALLOC_SBUSY shared busy the allocated page
1780 * VM_ALLOC_WIRED wire the allocated page
1781 * VM_ALLOC_ZERO prefer a zeroed page
1782 *
1783 * This routine may not sleep.
1784 */
1785 vm_page_t
vm_page_alloc_contig(vm_object_t object,vm_pindex_t pindex,int req,u_long npages,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary,vm_memattr_t memattr)1786 vm_page_alloc_contig(vm_object_t object, vm_pindex_t pindex, int req,
1787 u_long npages, vm_paddr_t low, vm_paddr_t high, u_long alignment,
1788 vm_paddr_t boundary, vm_memattr_t memattr)
1789 {
1790 struct vnode *drop;
1791 struct spglist deferred_vdrop_list;
1792 vm_page_t m, m_tmp, m_ret;
1793 u_int flags;
1794 int req_class;
1795
1796 KASSERT((object != NULL) == ((req & VM_ALLOC_NOOBJ) == 0) &&
1797 (object != NULL || (req & VM_ALLOC_SBUSY) == 0) &&
1798 ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) !=
1799 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)),
1800 ("vm_page_alloc: inconsistent object(%p)/req(%x)", (void *)object,
1801 req));
1802 if (object != NULL) {
1803 VM_OBJECT_ASSERT_WLOCKED(object);
1804 KASSERT(object->type == OBJT_PHYS,
1805 ("vm_page_alloc_contig: object %p isn't OBJT_PHYS",
1806 object));
1807 }
1808 KASSERT(npages > 0, ("vm_page_alloc_contig: npages is zero"));
1809 req_class = req & VM_ALLOC_CLASS_MASK;
1810
1811 /*
1812 * The page daemon is allowed to dig deeper into the free page list.
1813 */
1814 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1815 req_class = VM_ALLOC_SYSTEM;
1816
1817 SLIST_INIT(&deferred_vdrop_list);
1818 mtx_lock(&vm_page_queue_free_mtx);
1819 if (vm_cnt.v_free_count >= npages + vm_cnt.v_free_reserved ||
1820 (req_class == VM_ALLOC_SYSTEM &&
1821 (vm_cnt.v_free_count >= npages + vm_cnt.v_interrupt_free_min)) ||
1822 (req_class == VM_ALLOC_INTERRUPT && vm_cnt.v_free_count >= npages)) {
1823 #if VM_NRESERVLEVEL > 0
1824 retry:
1825 if (object == NULL || (object->flags & OBJ_COLORED) == 0 ||
1826 (m_ret = vm_reserv_alloc_contig(object, pindex, npages,
1827 low, high, alignment, boundary)) == NULL)
1828 #endif
1829 m_ret = vm_phys_alloc_contig(npages, low, high,
1830 alignment, boundary);
1831 } else {
1832 mtx_unlock(&vm_page_queue_free_mtx);
1833 atomic_add_int(&vm_pageout_deficit, npages);
1834 pagedaemon_wakeup();
1835 return (NULL);
1836 }
1837 if (m_ret != NULL)
1838 for (m = m_ret; m < &m_ret[npages]; m++) {
1839 drop = vm_page_alloc_init(m);
1840 if (drop != NULL) {
1841 /*
1842 * Enqueue the vnode for deferred vdrop().
1843 */
1844 m->plinks.s.pv = drop;
1845 SLIST_INSERT_HEAD(&deferred_vdrop_list, m,
1846 plinks.s.ss);
1847 }
1848 }
1849 else {
1850 #if VM_NRESERVLEVEL > 0
1851 if (vm_reserv_reclaim_contig(npages, low, high, alignment,
1852 boundary))
1853 goto retry;
1854 #endif
1855 }
1856 mtx_unlock(&vm_page_queue_free_mtx);
1857 if (m_ret == NULL)
1858 return (NULL);
1859
1860 /*
1861 * Initialize the pages. Only the PG_ZERO flag is inherited.
1862 */
1863 flags = 0;
1864 if ((req & VM_ALLOC_ZERO) != 0)
1865 flags = PG_ZERO;
1866 if ((req & VM_ALLOC_NODUMP) != 0)
1867 flags |= PG_NODUMP;
1868 if ((req & VM_ALLOC_WIRED) != 0)
1869 atomic_add_int(&vm_cnt.v_wire_count, npages);
1870 if (object != NULL) {
1871 if (object->memattr != VM_MEMATTR_DEFAULT &&
1872 memattr == VM_MEMATTR_DEFAULT)
1873 memattr = object->memattr;
1874 }
1875 for (m = m_ret; m < &m_ret[npages]; m++) {
1876 m->aflags = 0;
1877 m->flags = (m->flags | PG_NODUMP) & flags;
1878 m->busy_lock = VPB_UNBUSIED;
1879 if (object != NULL) {
1880 if ((req & (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
1881 m->busy_lock = VPB_SINGLE_EXCLUSIVER;
1882 if ((req & VM_ALLOC_SBUSY) != 0)
1883 m->busy_lock = VPB_SHARERS_WORD(1);
1884 }
1885 if ((req & VM_ALLOC_WIRED) != 0)
1886 m->wire_count = 1;
1887 /* Unmanaged pages don't use "act_count". */
1888 m->oflags = VPO_UNMANAGED;
1889 if (object != NULL) {
1890 if (vm_page_insert(m, object, pindex)) {
1891 vm_page_alloc_contig_vdrop(
1892 &deferred_vdrop_list);
1893 if (vm_paging_needed())
1894 pagedaemon_wakeup();
1895 if ((req & VM_ALLOC_WIRED) != 0)
1896 atomic_subtract_int(&vm_cnt.v_wire_count,
1897 npages);
1898 for (m_tmp = m, m = m_ret;
1899 m < &m_ret[npages]; m++) {
1900 if ((req & VM_ALLOC_WIRED) != 0)
1901 m->wire_count = 0;
1902 if (m >= m_tmp)
1903 m->object = NULL;
1904 vm_page_free(m);
1905 }
1906 return (NULL);
1907 }
1908 } else
1909 m->pindex = pindex;
1910 if (memattr != VM_MEMATTR_DEFAULT)
1911 pmap_page_set_memattr(m, memattr);
1912 pindex++;
1913 }
1914 vm_page_alloc_contig_vdrop(&deferred_vdrop_list);
1915 if (vm_paging_needed())
1916 pagedaemon_wakeup();
1917 return (m_ret);
1918 }
1919
1920 /*
1921 * Initialize a page that has been freshly dequeued from a freelist.
1922 * The caller has to drop the vnode returned, if it is not NULL.
1923 *
1924 * This function may only be used to initialize unmanaged pages.
1925 *
1926 * To be called with vm_page_queue_free_mtx held.
1927 */
1928 static struct vnode *
vm_page_alloc_init(vm_page_t m)1929 vm_page_alloc_init(vm_page_t m)
1930 {
1931 struct vnode *drop;
1932
1933 KASSERT(m->queue == PQ_NONE,
1934 ("vm_page_alloc_init: page %p has unexpected queue %d",
1935 m, m->queue));
1936 KASSERT(m->wire_count == 0,
1937 ("vm_page_alloc_init: page %p is wired", m));
1938 KASSERT(m->hold_count == 0,
1939 ("vm_page_alloc_init: page %p is held", m));
1940 KASSERT(!vm_page_sbusied(m),
1941 ("vm_page_alloc_init: page %p is busy", m));
1942 KASSERT(m->dirty == 0,
1943 ("vm_page_alloc_init: page %p is dirty", m));
1944 KASSERT(pmap_page_get_memattr(m) == VM_MEMATTR_DEFAULT,
1945 ("vm_page_alloc_init: page %p has unexpected memattr %d",
1946 m, pmap_page_get_memattr(m)));
1947 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
1948 drop = NULL;
1949 KASSERT(m->valid == 0,
1950 ("vm_page_alloc_init: free page %p is valid", m));
1951 vm_phys_freecnt_adj(m, -1);
1952 if ((m->flags & PG_ZERO) != 0)
1953 vm_page_zero_count--;
1954 return (drop);
1955 }
1956
1957 /*
1958 * vm_page_alloc_freelist:
1959 *
1960 * Allocate a physical page from the specified free page list.
1961 *
1962 * The caller must always specify an allocation class.
1963 *
1964 * allocation classes:
1965 * VM_ALLOC_NORMAL normal process request
1966 * VM_ALLOC_SYSTEM system *really* needs a page
1967 * VM_ALLOC_INTERRUPT interrupt time request
1968 *
1969 * optional allocation flags:
1970 * VM_ALLOC_COUNT(number) the number of additional pages that the caller
1971 * intends to allocate
1972 * VM_ALLOC_WIRED wire the allocated page
1973 * VM_ALLOC_ZERO prefer a zeroed page
1974 *
1975 * This routine may not sleep.
1976 */
1977 vm_page_t
vm_page_alloc_freelist(int flind,int req)1978 vm_page_alloc_freelist(int flind, int req)
1979 {
1980 struct vnode *drop;
1981 vm_page_t m;
1982 u_int flags;
1983 int req_class;
1984
1985 req_class = req & VM_ALLOC_CLASS_MASK;
1986
1987 /*
1988 * The page daemon is allowed to dig deeper into the free page list.
1989 */
1990 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
1991 req_class = VM_ALLOC_SYSTEM;
1992
1993 /*
1994 * Do not allocate reserved pages unless the req has asked for it.
1995 */
1996 mtx_lock_flags(&vm_page_queue_free_mtx, MTX_RECURSE);
1997 if (vm_cnt.v_free_count > vm_cnt.v_free_reserved ||
1998 (req_class == VM_ALLOC_SYSTEM &&
1999 vm_cnt.v_free_count > vm_cnt.v_interrupt_free_min) ||
2000 (req_class == VM_ALLOC_INTERRUPT && vm_cnt.v_free_count > 0))
2001 m = vm_phys_alloc_freelist_pages(flind, VM_FREEPOOL_DIRECT, 0);
2002 else {
2003 mtx_unlock(&vm_page_queue_free_mtx);
2004 atomic_add_int(&vm_pageout_deficit,
2005 max((u_int)req >> VM_ALLOC_COUNT_SHIFT, 1));
2006 pagedaemon_wakeup();
2007 return (NULL);
2008 }
2009 if (m == NULL) {
2010 mtx_unlock(&vm_page_queue_free_mtx);
2011 return (NULL);
2012 }
2013 drop = vm_page_alloc_init(m);
2014 mtx_unlock(&vm_page_queue_free_mtx);
2015
2016 /*
2017 * Initialize the page. Only the PG_ZERO flag is inherited.
2018 */
2019 m->aflags = 0;
2020 flags = 0;
2021 if ((req & VM_ALLOC_ZERO) != 0)
2022 flags = PG_ZERO;
2023 m->flags &= flags;
2024 if ((req & VM_ALLOC_WIRED) != 0) {
2025 /*
2026 * The page lock is not required for wiring a page that does
2027 * not belong to an object.
2028 */
2029 atomic_add_int(&vm_cnt.v_wire_count, 1);
2030 m->wire_count = 1;
2031 }
2032 /* Unmanaged pages don't use "act_count". */
2033 m->oflags = VPO_UNMANAGED;
2034 if (drop != NULL)
2035 vdrop(drop);
2036 if (vm_paging_needed())
2037 pagedaemon_wakeup();
2038 return (m);
2039 }
2040
2041 #define VPSC_ANY 0 /* No restrictions. */
2042 #define VPSC_NORESERV 1 /* Skip reservations; implies VPSC_NOSUPER. */
2043 #define VPSC_NOSUPER 2 /* Skip superpages. */
2044
2045 /*
2046 * vm_page_scan_contig:
2047 *
2048 * Scan vm_page_array[] between the specified entries "m_start" and
2049 * "m_end" for a run of contiguous physical pages that satisfy the
2050 * specified conditions, and return the lowest page in the run. The
2051 * specified "alignment" determines the alignment of the lowest physical
2052 * page in the run. If the specified "boundary" is non-zero, then the
2053 * run of physical pages cannot span a physical address that is a
2054 * multiple of "boundary".
2055 *
2056 * "m_end" is never dereferenced, so it need not point to a vm_page
2057 * structure within vm_page_array[].
2058 *
2059 * "npages" must be greater than zero. "m_start" and "m_end" must not
2060 * span a hole (or discontiguity) in the physical address space. Both
2061 * "alignment" and "boundary" must be a power of two.
2062 */
2063 vm_page_t
vm_page_scan_contig(u_long npages,vm_page_t m_start,vm_page_t m_end,u_long alignment,vm_paddr_t boundary,int options)2064 vm_page_scan_contig(u_long npages, vm_page_t m_start, vm_page_t m_end,
2065 u_long alignment, vm_paddr_t boundary, int options)
2066 {
2067 struct mtx *m_mtx, *new_mtx;
2068 vm_object_t object;
2069 vm_paddr_t pa;
2070 vm_page_t m, m_run;
2071 #if VM_NRESERVLEVEL > 0
2072 int level;
2073 #endif
2074 int m_inc, order, run_ext, run_len;
2075
2076 KASSERT(npages > 0, ("npages is 0"));
2077 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2078 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2079 m_run = NULL;
2080 run_len = 0;
2081 m_mtx = NULL;
2082 for (m = m_start; m < m_end && run_len < npages; m += m_inc) {
2083 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2084 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2085
2086 /*
2087 * If the current page would be the start of a run, check its
2088 * physical address against the end, alignment, and boundary
2089 * conditions. If it doesn't satisfy these conditions, either
2090 * terminate the scan or advance to the next page that
2091 * satisfies the failed condition.
2092 */
2093 if (run_len == 0) {
2094 KASSERT(m_run == NULL, ("m_run != NULL"));
2095 if (m + npages > m_end)
2096 break;
2097 pa = VM_PAGE_TO_PHYS(m);
2098 if ((pa & (alignment - 1)) != 0) {
2099 m_inc = atop(roundup2(pa, alignment) - pa);
2100 continue;
2101 }
2102 if (((pa ^ (pa + ptoa(npages) - 1)) & ~(boundary -
2103 1)) != 0) {
2104 m_inc = atop(roundup2(pa, boundary) - pa);
2105 continue;
2106 }
2107 } else
2108 KASSERT(m_run != NULL, ("m_run == NULL"));
2109
2110 /*
2111 * Avoid releasing and reacquiring the same page lock.
2112 */
2113 new_mtx = vm_page_lockptr(m);
2114 if (m_mtx != new_mtx) {
2115 if (m_mtx != NULL)
2116 mtx_unlock(m_mtx);
2117 m_mtx = new_mtx;
2118 mtx_lock(m_mtx);
2119 }
2120 m_inc = 1;
2121 retry:
2122 if (m->wire_count != 0 || m->hold_count != 0)
2123 run_ext = 0;
2124 #if VM_NRESERVLEVEL > 0
2125 else if ((level = vm_reserv_level(m)) >= 0 &&
2126 (options & VPSC_NORESERV) != 0) {
2127 run_ext = 0;
2128 /* Advance to the end of the reservation. */
2129 pa = VM_PAGE_TO_PHYS(m);
2130 m_inc = atop(roundup2(pa + 1, vm_reserv_size(level)) -
2131 pa);
2132 }
2133 #endif
2134 else if ((object = m->object) != NULL) {
2135 /*
2136 * The page is considered eligible for relocation if
2137 * and only if it could be laundered or reclaimed by
2138 * the page daemon.
2139 */
2140 if (!VM_OBJECT_TRYRLOCK(object)) {
2141 mtx_unlock(m_mtx);
2142 VM_OBJECT_RLOCK(object);
2143 mtx_lock(m_mtx);
2144 if (m->object != object) {
2145 /*
2146 * The page may have been freed.
2147 */
2148 VM_OBJECT_RUNLOCK(object);
2149 goto retry;
2150 } else if (m->wire_count != 0 ||
2151 m->hold_count != 0) {
2152 run_ext = 0;
2153 goto unlock;
2154 }
2155 }
2156 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2157 ("page %p is PG_UNHOLDFREE", m));
2158 /* Don't care: PG_NODUMP, PG_WINATCFLS, PG_ZERO. */
2159 if (object->type != OBJT_DEFAULT &&
2160 object->type != OBJT_SWAP &&
2161 object->type != OBJT_VNODE)
2162 run_ext = 0;
2163 else if (m != vm_page_lookup(object, m->pindex)) {
2164 /*
2165 * The page is cached or recently converted
2166 * from cached to free.
2167 */
2168 #if VM_NRESERVLEVEL > 0
2169 if (level >= 0) {
2170 /*
2171 * The page is reserved. Extend the
2172 * current run by one page.
2173 */
2174 run_ext = 1;
2175 } else
2176 #endif
2177 if ((order = m->order) < VM_NFREEORDER) {
2178 /*
2179 * The page is enqueued in the
2180 * physical memory allocator's cache/
2181 * free page queues. Moreover, it is
2182 * the first page in a power-of-two-
2183 * sized run of contiguous cache/free
2184 * pages. Add these pages to the end
2185 * of the current run, and jump
2186 * ahead.
2187 */
2188 run_ext = 1 << order;
2189 m_inc = 1 << order;
2190 } else
2191 run_ext = 0;
2192 #if VM_NRESERVLEVEL > 0
2193 } else if ((options & VPSC_NOSUPER) != 0 &&
2194 (level = vm_reserv_level_iffullpop(m)) >= 0) {
2195 run_ext = 0;
2196 /* Advance to the end of the superpage. */
2197 pa = VM_PAGE_TO_PHYS(m);
2198 m_inc = atop(roundup2(pa + 1,
2199 vm_reserv_size(level)) - pa);
2200 #endif
2201 } else if (object->memattr == VM_MEMATTR_DEFAULT &&
2202 m->queue != PQ_NONE && !vm_page_busied(m)) {
2203 /*
2204 * The page is allocated but eligible for
2205 * relocation. Extend the current run by one
2206 * page.
2207 */
2208 KASSERT(pmap_page_get_memattr(m) ==
2209 VM_MEMATTR_DEFAULT,
2210 ("page %p has an unexpected memattr", m));
2211 KASSERT((m->oflags & (VPO_SWAPINPROG |
2212 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2213 ("page %p has unexpected oflags", m));
2214 /* Don't care: VPO_NOSYNC. */
2215 run_ext = 1;
2216 } else
2217 run_ext = 0;
2218 unlock:
2219 VM_OBJECT_RUNLOCK(object);
2220 #if VM_NRESERVLEVEL > 0
2221 } else if (level >= 0) {
2222 /*
2223 * The page is reserved but not yet allocated. In
2224 * other words, it is still cached or free. Extend
2225 * the current run by one page.
2226 */
2227 run_ext = 1;
2228 #endif
2229 } else if ((order = m->order) < VM_NFREEORDER) {
2230 /*
2231 * The page is enqueued in the physical memory
2232 * allocator's cache/free page queues. Moreover, it
2233 * is the first page in a power-of-two-sized run of
2234 * contiguous cache/free pages. Add these pages to
2235 * the end of the current run, and jump ahead.
2236 */
2237 run_ext = 1 << order;
2238 m_inc = 1 << order;
2239 } else {
2240 /*
2241 * Skip the page for one of the following reasons: (1)
2242 * It is enqueued in the physical memory allocator's
2243 * cache/free page queues. However, it is not the
2244 * first page in a run of contiguous cache/free pages.
2245 * (This case rarely occurs because the scan is
2246 * performed in ascending order.) (2) It is not
2247 * reserved, and it is transitioning from free to
2248 * allocated. (Conversely, the transition from
2249 * allocated to free for managed pages is blocked by
2250 * the page lock.) (3) It is allocated but not
2251 * contained by an object and not wired, e.g.,
2252 * allocated by Xen's balloon driver.
2253 */
2254 run_ext = 0;
2255 }
2256
2257 /*
2258 * Extend or reset the current run of pages.
2259 */
2260 if (run_ext > 0) {
2261 if (run_len == 0)
2262 m_run = m;
2263 run_len += run_ext;
2264 } else {
2265 if (run_len > 0) {
2266 m_run = NULL;
2267 run_len = 0;
2268 }
2269 }
2270 }
2271 if (m_mtx != NULL)
2272 mtx_unlock(m_mtx);
2273 if (run_len >= npages)
2274 return (m_run);
2275 return (NULL);
2276 }
2277
2278 /*
2279 * vm_page_reclaim_run:
2280 *
2281 * Try to relocate each of the allocated virtual pages within the
2282 * specified run of physical pages to a new physical address. Free the
2283 * physical pages underlying the relocated virtual pages. A virtual page
2284 * is relocatable if and only if it could be laundered or reclaimed by
2285 * the page daemon. Whenever possible, a virtual page is relocated to a
2286 * physical address above "high".
2287 *
2288 * Returns 0 if every physical page within the run was already free or
2289 * just freed by a successful relocation. Otherwise, returns a non-zero
2290 * value indicating why the last attempt to relocate a virtual page was
2291 * unsuccessful.
2292 *
2293 * "req_class" must be an allocation class.
2294 */
2295 static int
vm_page_reclaim_run(int req_class,u_long npages,vm_page_t m_run,vm_paddr_t high)2296 vm_page_reclaim_run(int req_class, u_long npages, vm_page_t m_run,
2297 vm_paddr_t high)
2298 {
2299 struct mtx *m_mtx, *new_mtx;
2300 struct spglist free;
2301 vm_object_t object;
2302 vm_paddr_t pa;
2303 vm_page_t m, m_end, m_new;
2304 int error, order, req;
2305
2306 KASSERT((req_class & VM_ALLOC_CLASS_MASK) == req_class,
2307 ("req_class is not an allocation class"));
2308 SLIST_INIT(&free);
2309 error = 0;
2310 m = m_run;
2311 m_end = m_run + npages;
2312 m_mtx = NULL;
2313 for (; error == 0 && m < m_end; m++) {
2314 KASSERT((m->flags & (PG_FICTITIOUS | PG_MARKER)) == 0,
2315 ("page %p is PG_FICTITIOUS or PG_MARKER", m));
2316
2317 /*
2318 * Avoid releasing and reacquiring the same page lock.
2319 */
2320 new_mtx = vm_page_lockptr(m);
2321 if (m_mtx != new_mtx) {
2322 if (m_mtx != NULL)
2323 mtx_unlock(m_mtx);
2324 m_mtx = new_mtx;
2325 mtx_lock(m_mtx);
2326 }
2327 retry:
2328 if (m->wire_count != 0 || m->hold_count != 0)
2329 error = EBUSY;
2330 else if ((object = m->object) != NULL) {
2331 /*
2332 * The page is relocated if and only if it could be
2333 * laundered or reclaimed by the page daemon.
2334 */
2335 if (!VM_OBJECT_TRYWLOCK(object)) {
2336 mtx_unlock(m_mtx);
2337 VM_OBJECT_WLOCK(object);
2338 mtx_lock(m_mtx);
2339 if (m->object != object) {
2340 /*
2341 * The page may have been freed.
2342 */
2343 VM_OBJECT_WUNLOCK(object);
2344 goto retry;
2345 } else if (m->wire_count != 0 ||
2346 m->hold_count != 0) {
2347 error = EBUSY;
2348 goto unlock;
2349 }
2350 }
2351 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
2352 ("page %p is PG_UNHOLDFREE", m));
2353 /* Don't care: PG_NODUMP, PG_WINATCFLS, PG_ZERO. */
2354 if (object->type != OBJT_DEFAULT &&
2355 object->type != OBJT_SWAP &&
2356 object->type != OBJT_VNODE)
2357 error = EINVAL;
2358 else if (m != vm_page_lookup(object, m->pindex)) {
2359 /*
2360 * The page is cached or recently converted
2361 * from cached to free.
2362 */
2363 VM_OBJECT_WUNLOCK(object);
2364 goto cached;
2365 } else if (object->memattr != VM_MEMATTR_DEFAULT)
2366 error = EINVAL;
2367 else if (m->queue != PQ_NONE && !vm_page_busied(m)) {
2368 KASSERT(pmap_page_get_memattr(m) ==
2369 VM_MEMATTR_DEFAULT,
2370 ("page %p has an unexpected memattr", m));
2371 KASSERT((m->oflags & (VPO_SWAPINPROG |
2372 VPO_SWAPSLEEP | VPO_UNMANAGED)) == 0,
2373 ("page %p has unexpected oflags", m));
2374 /* Don't care: VPO_NOSYNC. */
2375 if (m->valid != 0) {
2376 /*
2377 * First, try to allocate a new page
2378 * that is above "high". Failing
2379 * that, try to allocate a new page
2380 * that is below "m_run". Allocate
2381 * the new page between the end of
2382 * "m_run" and "high" only as a last
2383 * resort.
2384 */
2385 req = req_class | VM_ALLOC_NOOBJ;
2386 if ((m->flags & PG_NODUMP) != 0)
2387 req |= VM_ALLOC_NODUMP;
2388 if (trunc_page(high) !=
2389 ~(vm_paddr_t)PAGE_MASK) {
2390 m_new = vm_page_alloc_contig(
2391 NULL, 0, req, 1,
2392 round_page(high),
2393 ~(vm_paddr_t)0,
2394 PAGE_SIZE, 0,
2395 VM_MEMATTR_DEFAULT);
2396 } else
2397 m_new = NULL;
2398 if (m_new == NULL) {
2399 pa = VM_PAGE_TO_PHYS(m_run);
2400 m_new = vm_page_alloc_contig(
2401 NULL, 0, req, 1,
2402 0, pa - 1, PAGE_SIZE, 0,
2403 VM_MEMATTR_DEFAULT);
2404 }
2405 if (m_new == NULL) {
2406 pa += ptoa(npages);
2407 m_new = vm_page_alloc_contig(
2408 NULL, 0, req, 1,
2409 pa, high, PAGE_SIZE, 0,
2410 VM_MEMATTR_DEFAULT);
2411 }
2412 if (m_new == NULL) {
2413 error = ENOMEM;
2414 goto unlock;
2415 }
2416 KASSERT(m_new->wire_count == 0,
2417 ("page %p is wired", m));
2418
2419 /*
2420 * Replace "m" with the new page. For
2421 * vm_page_replace(), "m" must be busy
2422 * and dequeued. Finally, change "m"
2423 * as if vm_page_free() was called.
2424 */
2425 if (object->ref_count != 0)
2426 pmap_remove_all(m);
2427 m_new->aflags = m->aflags;
2428 KASSERT(m_new->oflags == VPO_UNMANAGED,
2429 ("page %p is managed", m));
2430 m_new->oflags = m->oflags & VPO_NOSYNC;
2431 pmap_copy_page(m, m_new);
2432 m_new->valid = m->valid;
2433 m_new->dirty = m->dirty;
2434 m->flags &= ~PG_ZERO;
2435 vm_page_xbusy(m);
2436 vm_page_remque(m);
2437 vm_page_replace_checked(m_new, object,
2438 m->pindex, m);
2439 m->valid = 0;
2440 vm_page_undirty(m);
2441
2442 /*
2443 * The new page must be deactivated
2444 * before the object is unlocked.
2445 */
2446 new_mtx = vm_page_lockptr(m_new);
2447 if (m_mtx != new_mtx) {
2448 mtx_unlock(m_mtx);
2449 m_mtx = new_mtx;
2450 mtx_lock(m_mtx);
2451 }
2452 vm_page_deactivate(m_new);
2453 } else {
2454 m->flags &= ~PG_ZERO;
2455 vm_page_remque(m);
2456 vm_page_remove(m);
2457 KASSERT(m->dirty == 0,
2458 ("page %p is dirty", m));
2459 }
2460 SLIST_INSERT_HEAD(&free, m, plinks.s.ss);
2461 } else
2462 error = EBUSY;
2463 unlock:
2464 VM_OBJECT_WUNLOCK(object);
2465 } else {
2466 cached:
2467 mtx_lock(&vm_page_queue_free_mtx);
2468 order = m->order;
2469 if (order < VM_NFREEORDER) {
2470 /*
2471 * The page is enqueued in the physical memory
2472 * allocator's cache/free page queues.
2473 * Moreover, it is the first page in a power-
2474 * of-two-sized run of contiguous cache/free
2475 * pages. Jump ahead to the last page within
2476 * that run, and continue from there.
2477 */
2478 m += (1 << order) - 1;
2479 }
2480 #if VM_NRESERVLEVEL > 0
2481 else if (vm_reserv_is_page_free(m))
2482 order = 0;
2483 #endif
2484 mtx_unlock(&vm_page_queue_free_mtx);
2485 if (order == VM_NFREEORDER)
2486 error = EINVAL;
2487 }
2488 }
2489 if (m_mtx != NULL)
2490 mtx_unlock(m_mtx);
2491 if ((m = SLIST_FIRST(&free)) != NULL) {
2492 mtx_lock(&vm_page_queue_free_mtx);
2493 do {
2494 SLIST_REMOVE_HEAD(&free, plinks.s.ss);
2495 vm_phys_freecnt_adj(m, 1);
2496 #if VM_NRESERVLEVEL > 0
2497 if (!vm_reserv_free_page(m))
2498 #else
2499 if (true)
2500 #endif
2501 vm_phys_free_pages(m, 0);
2502 } while ((m = SLIST_FIRST(&free)) != NULL);
2503 vm_page_zero_idle_wakeup();
2504 vm_page_free_wakeup();
2505 mtx_unlock(&vm_page_queue_free_mtx);
2506 }
2507 return (error);
2508 }
2509
2510 #define NRUNS 16
2511
2512 CTASSERT(powerof2(NRUNS));
2513
2514 #define RUN_INDEX(count) ((count) & (NRUNS - 1))
2515
2516 #define MIN_RECLAIM 8
2517
2518 /*
2519 * vm_page_reclaim_contig:
2520 *
2521 * Reclaim allocated, contiguous physical memory satisfying the specified
2522 * conditions by relocating the virtual pages using that physical memory.
2523 * Returns true if reclamation is successful and false otherwise. Since
2524 * relocation requires the allocation of physical pages, reclamation may
2525 * fail due to a shortage of cache/free pages. When reclamation fails,
2526 * callers are expected to perform VM_WAIT before retrying a failed
2527 * allocation operation, e.g., vm_page_alloc_contig().
2528 *
2529 * The caller must always specify an allocation class through "req".
2530 *
2531 * allocation classes:
2532 * VM_ALLOC_NORMAL normal process request
2533 * VM_ALLOC_SYSTEM system *really* needs a page
2534 * VM_ALLOC_INTERRUPT interrupt time request
2535 *
2536 * The optional allocation flags are ignored.
2537 *
2538 * "npages" must be greater than zero. Both "alignment" and "boundary"
2539 * must be a power of two.
2540 */
2541 bool
vm_page_reclaim_contig(int req,u_long npages,vm_paddr_t low,vm_paddr_t high,u_long alignment,vm_paddr_t boundary)2542 vm_page_reclaim_contig(int req, u_long npages, vm_paddr_t low, vm_paddr_t high,
2543 u_long alignment, vm_paddr_t boundary)
2544 {
2545 vm_paddr_t curr_low;
2546 vm_page_t m_run, m_runs[NRUNS];
2547 u_long count, reclaimed;
2548 int error, i, options, req_class;
2549
2550 KASSERT(npages > 0, ("npages is 0"));
2551 KASSERT(powerof2(alignment), ("alignment is not a power of 2"));
2552 KASSERT(powerof2(boundary), ("boundary is not a power of 2"));
2553 req_class = req & VM_ALLOC_CLASS_MASK;
2554
2555 /*
2556 * The page daemon is allowed to dig deeper into the free page list.
2557 */
2558 if (curproc == pageproc && req_class != VM_ALLOC_INTERRUPT)
2559 req_class = VM_ALLOC_SYSTEM;
2560
2561 /*
2562 * Return if the number of cached and free pages cannot satisfy the
2563 * requested allocation.
2564 */
2565 count = vm_cnt.v_free_count;
2566 if (count < npages + vm_cnt.v_free_reserved || (count < npages +
2567 vm_cnt.v_interrupt_free_min && req_class == VM_ALLOC_SYSTEM) ||
2568 (count < npages && req_class == VM_ALLOC_INTERRUPT))
2569 return (false);
2570
2571 /*
2572 * Scan up to three times, relaxing the restrictions ("options") on
2573 * the reclamation of reservations and superpages each time.
2574 */
2575 for (options = VPSC_NORESERV;;) {
2576 /*
2577 * Find the highest runs that satisfy the given constraints
2578 * and restrictions, and record them in "m_runs".
2579 */
2580 curr_low = low;
2581 count = 0;
2582 for (;;) {
2583 m_run = vm_phys_scan_contig(npages, curr_low, high,
2584 alignment, boundary, options);
2585 if (m_run == NULL)
2586 break;
2587 curr_low = VM_PAGE_TO_PHYS(m_run) + ptoa(npages);
2588 m_runs[RUN_INDEX(count)] = m_run;
2589 count++;
2590 }
2591
2592 /*
2593 * Reclaim the highest runs in LIFO (descending) order until
2594 * the number of reclaimed pages, "reclaimed", is at least
2595 * MIN_RECLAIM. Reset "reclaimed" each time because each
2596 * reclamation is idempotent, and runs will (likely) recur
2597 * from one scan to the next as restrictions are relaxed.
2598 */
2599 reclaimed = 0;
2600 for (i = 0; count > 0 && i < NRUNS; i++) {
2601 count--;
2602 m_run = m_runs[RUN_INDEX(count)];
2603 error = vm_page_reclaim_run(req_class, npages, m_run,
2604 high);
2605 if (error == 0) {
2606 reclaimed += npages;
2607 if (reclaimed >= MIN_RECLAIM)
2608 return (true);
2609 }
2610 }
2611
2612 /*
2613 * Either relax the restrictions on the next scan or return if
2614 * the last scan had no restrictions.
2615 */
2616 if (options == VPSC_NORESERV)
2617 options = VPSC_NOSUPER;
2618 else if (options == VPSC_NOSUPER)
2619 options = VPSC_ANY;
2620 else if (options == VPSC_ANY)
2621 return (reclaimed != 0);
2622 }
2623 }
2624
2625 /*
2626 * vm_wait: (also see VM_WAIT macro)
2627 *
2628 * Sleep until free pages are available for allocation.
2629 * - Called in various places before memory allocations.
2630 */
2631 void
vm_wait(void)2632 vm_wait(void)
2633 {
2634
2635 mtx_lock(&vm_page_queue_free_mtx);
2636 if (curproc == pageproc) {
2637 vm_pageout_pages_needed = 1;
2638 msleep(&vm_pageout_pages_needed, &vm_page_queue_free_mtx,
2639 PDROP | PSWP, "VMWait", 0);
2640 } else {
2641 if (!vm_pages_needed) {
2642 vm_pages_needed = 1;
2643 wakeup(&vm_pages_needed);
2644 }
2645 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PVM,
2646 "vmwait", 0);
2647 }
2648 }
2649
2650 /*
2651 * vm_waitpfault: (also see VM_WAITPFAULT macro)
2652 *
2653 * Sleep until free pages are available for allocation.
2654 * - Called only in vm_fault so that processes page faulting
2655 * can be easily tracked.
2656 * - Sleeps at a lower priority than vm_wait() so that vm_wait()ing
2657 * processes will be able to grab memory first. Do not change
2658 * this balance without careful testing first.
2659 */
2660 void
vm_waitpfault(void)2661 vm_waitpfault(void)
2662 {
2663
2664 mtx_lock(&vm_page_queue_free_mtx);
2665 if (!vm_pages_needed) {
2666 vm_pages_needed = 1;
2667 wakeup(&vm_pages_needed);
2668 }
2669 msleep(&vm_cnt.v_free_count, &vm_page_queue_free_mtx, PDROP | PUSER,
2670 "pfault", 0);
2671 }
2672
2673 struct vm_pagequeue *
vm_page_pagequeue(vm_page_t m)2674 vm_page_pagequeue(vm_page_t m)
2675 {
2676 int queue = m->queue;
2677 struct vm_domain *dom = vm_phys_domain(m);
2678
2679 return (&dom->vmd_pagequeues[queue]);
2680 }
2681
2682 struct vm_pagequeue *
vm_page_pagequeue_deferred(vm_page_t m)2683 vm_page_pagequeue_deferred(vm_page_t m)
2684 {
2685 int queue = m->queue;
2686 struct vm_domain *dom = vm_phys_domain(m);
2687
2688 vm_page_lock_assert(m, MA_OWNED);
2689 if ((queue == PQ_INACTIVE) && (m->flags & PG_PAQUEUE))
2690 return (&dom->vmd_pagequeues[vm_page_queue_idx(m)]);
2691 else
2692 return (&dom->vmd_pagequeues[queue]);
2693 }
2694
2695 int
vm_page_queue_fixup_locked(vm_page_t m)2696 vm_page_queue_fixup_locked(vm_page_t m)
2697 {
2698 int merged, _cnt;
2699 struct vm_pagequeue *vpq, *lvpq;
2700 struct vm_domain *vmd;
2701 vm_page_t m1, mtmp;
2702
2703 vmd = vm_phys_domain(m);
2704 vm_page_lock_assert(m, MA_OWNED);
2705
2706 vpq = &vmd->vmd_pagequeues[PQ_INACTIVE];
2707
2708 merged = 0;
2709 lvpq = &vmd->vmd_pagequeues[vm_page_queue_idx(m)];
2710
2711 _cnt = 0;
2712 TAILQ_FOREACH_SAFE(m1, &lvpq->pq_pl, plinks.q, mtmp) {
2713 #ifdef INVARIANTS
2714 _cnt++;
2715 VM_ASSERT(m1->queue == PQ_INACTIVE);
2716 VM_ASSERT((m1->flags & PG_PAQUEUE) != 0);
2717 #endif
2718 if (m1->flags & PG_ATHEAD) {
2719 TAILQ_REMOVE(&lvpq->pq_pl, m1, plinks.q);
2720 TAILQ_INSERT_HEAD(&vpq->pq_pl, m1, plinks.q);
2721 }
2722 m1->flags &= ~(PG_PAQUEUE|PG_ATHEAD);
2723
2724 }
2725 #ifdef INVARIANTS
2726 VM_ASSERT(_cnt == lvpq->pq_cnt);
2727 #endif
2728
2729 TAILQ_CONCAT(&vpq->pq_pl, &lvpq->pq_pl, plinks.q);
2730 vpq->pq_cnt += lvpq->pq_cnt;
2731 merged += lvpq->pq_cnt;
2732 atomic_add_int(&vm_cnt.v_inactive_deferred_count, -lvpq->pq_cnt);
2733 lvpq->pq_cnt = 0;
2734
2735 return (merged);
2736 }
2737
2738 int
vm_page_queue_fixup(vm_page_t m)2739 vm_page_queue_fixup(vm_page_t m)
2740 {
2741 struct vm_domain *vmd;
2742 struct vm_pagequeue *vpq, *lvpq;
2743 int merged;
2744
2745 vmd = vm_phys_domain(m);
2746 vpq = &vmd->vmd_pagequeues[PQ_INACTIVE];
2747
2748 lvpq = &vmd->vmd_pagequeues[vm_page_queue_idx(m)];
2749
2750 if (lvpq->pq_cnt < vm_paqlenthresh_lwm) {
2751 return (0);
2752 } else if (lvpq->pq_cnt < vm_paqlenthresh_hwm) {
2753 if (!vm_pagequeue_trylock(vpq))
2754 return (0);
2755 } else
2756 vm_pagequeue_lock(vpq);
2757
2758 merged = vm_page_queue_fixup_locked(m);
2759 vm_pagequeue_unlock(vpq);
2760 return (merged);
2761 }
2762
2763 /*
2764 * vm_page_dequeue:
2765 *
2766 * Remove the given page from its current page queue.
2767 *
2768 * The page must be locked.
2769 */
2770 void
vm_page_dequeue(vm_page_t m)2771 vm_page_dequeue(vm_page_t m)
2772 {
2773 struct vm_pagequeue *pq;
2774
2775 vm_page_assert_locked(m);
2776 KASSERT(m->queue < PQ_COUNT, ("vm_page_dequeue: page %p is not queued",
2777 m));
2778 pq = vm_page_pagequeue_deferred(m);
2779 if (m->flags & PG_PAQUEUE) {
2780 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2781 vm_pagequeue_cnt_dec(pq);
2782 m->flags &= ~PG_PAQUEUE;
2783 atomic_add_int(&vm_cnt.v_inactive_deferred_count, -1);
2784 } else {
2785 vm_pagequeue_lock(pq);
2786 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2787 vm_pagequeue_cnt_dec(pq);
2788 vm_pagequeue_unlock(pq);
2789 }
2790 m->queue = PQ_NONE;
2791 }
2792
2793 /*
2794 * vm_page_dequeue_locked:
2795 *
2796 * Remove the given page from its current page queue.
2797 *
2798 * The page and page queue must be locked.
2799 */
2800 void
vm_page_dequeue_locked(vm_page_t m)2801 vm_page_dequeue_locked(vm_page_t m)
2802 {
2803 struct vm_pagequeue *pq;
2804
2805 vm_page_lock_assert(m, MA_OWNED);
2806 pq = vm_page_pagequeue_deferred(m);
2807 if (m->flags & PG_PAQUEUE) {
2808 m->flags &= ~PG_PAQUEUE;
2809 atomic_add_int(&vm_cnt.v_inactive_deferred_count, -1);
2810 } else
2811 vm_pagequeue_assert_locked(pq);
2812 m->queue = PQ_NONE;
2813 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2814 vm_pagequeue_cnt_dec(pq);
2815 }
2816
2817 /*
2818 * vm_page_enqueue:
2819 *
2820 * Add the given page to the specified page queue.
2821 *
2822 * The page must be locked.
2823 */
2824 static void
vm_page_enqueue(uint8_t queue,vm_page_t m)2825 vm_page_enqueue(uint8_t queue, vm_page_t m)
2826 {
2827 struct vm_pagequeue *pq;
2828
2829 vm_page_lock_assert(m, MA_OWNED);
2830 KASSERT(queue < PQ_COUNT,
2831 ("vm_page_enqueue: invalid queue %u request for page %p",
2832 queue, m));
2833 if (queue == PQ_INACTIVE) {
2834 /* look up deferred queue */
2835 pq = &vm_phys_domain(m)->vmd_pagequeues[vm_page_queue_idx(m)];
2836 m->queue = queue;
2837 /* mark page as on physically addressed deferred queue */
2838 m->flags |= PG_PAQUEUE;
2839 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2840 vm_pagequeue_cnt_inc(pq);
2841 atomic_fetchadd_int(&vm_cnt.v_inactive_deferred_count, 1);
2842 if (pq->pq_cnt >= vm_paqlenthresh_lwm)
2843 vm_page_queue_fixup(m);
2844 } else {
2845 pq = &vm_phys_domain(m)->vmd_pagequeues[queue];
2846 vm_pagequeue_lock(pq);
2847 m->queue = queue;
2848 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2849 vm_pagequeue_cnt_inc(pq);
2850 vm_pagequeue_unlock(pq);
2851 }
2852 }
2853
2854 /*
2855 * vm_page_requeue:
2856 *
2857 * Move the given page to the tail of its current page queue.
2858 *
2859 * The page must be locked.
2860 */
2861 void
vm_page_requeue(vm_page_t m)2862 vm_page_requeue(vm_page_t m)
2863 {
2864 struct vm_pagequeue *pq;
2865
2866 vm_page_lock_assert(m, MA_OWNED);
2867 KASSERT(m->queue != PQ_NONE,
2868 ("vm_page_requeue: page %p is not queued", m));
2869 pq = vm_page_pagequeue_deferred(m);
2870 if (m->flags & PG_PAQUEUE) {
2871 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2872 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2873 } else {
2874 vm_pagequeue_lock(pq);
2875 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2876 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2877 vm_pagequeue_unlock(pq);
2878 }
2879 }
2880
2881 /*
2882 * vm_page_requeue_locked:
2883 *
2884 * Move the given page to the tail of its current page queue.
2885 *
2886 * The page queue must be locked.
2887 */
2888 void
vm_page_requeue_locked(vm_page_t m)2889 vm_page_requeue_locked(vm_page_t m)
2890 {
2891 struct vm_pagequeue *pq;
2892
2893 KASSERT(m->queue != PQ_NONE,
2894 ("vm_page_requeue_locked: page %p is not queued", m));
2895 /* the page lock isn't held and the page isn't on
2896 * the inactive queue it should be moved by fixup
2897 */
2898 if (m->flags & PG_PAQUEUE)
2899 return;
2900 pq = vm_page_pagequeue(m);
2901 vm_pagequeue_assert_locked(pq);
2902 TAILQ_REMOVE(&pq->pq_pl, m, plinks.q);
2903 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
2904 }
2905
2906 /*
2907 * vm_page_activate:
2908 *
2909 * Put the specified page on the active list (if appropriate).
2910 * Ensure that act_count is at least ACT_INIT but do not otherwise
2911 * mess with it.
2912 *
2913 * The page must be locked.
2914 */
2915 void
vm_page_activate(vm_page_t m)2916 vm_page_activate(vm_page_t m)
2917 {
2918 int queue;
2919
2920 vm_page_lock_assert(m, MA_OWNED);
2921 if ((queue = m->queue) != PQ_ACTIVE) {
2922 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
2923 if (m->act_count < ACT_INIT)
2924 m->act_count = ACT_INIT;
2925 if (queue != PQ_NONE)
2926 vm_page_dequeue(m);
2927 vm_page_enqueue(PQ_ACTIVE, m);
2928 } else
2929 KASSERT(queue == PQ_NONE,
2930 ("vm_page_activate: wired page %p is queued", m));
2931 } else {
2932 if (m->act_count < ACT_INIT)
2933 m->act_count = ACT_INIT;
2934 }
2935 }
2936
2937 /*
2938 * vm_page_free_wakeup:
2939 *
2940 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2941 * routine is called when a page has been added to the cache or free
2942 * queues.
2943 *
2944 * The page queues must be locked.
2945 */
2946 static inline void
vm_page_free_wakeup(void)2947 vm_page_free_wakeup(void)
2948 {
2949
2950 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
2951 /*
2952 * if pageout daemon needs pages, then tell it that there are
2953 * some free.
2954 */
2955 if (vm_pageout_pages_needed &&
2956 vm_cnt.v_free_count >= vm_cnt.v_pageout_free_min) {
2957 wakeup(&vm_pageout_pages_needed);
2958 vm_pageout_pages_needed = 0;
2959 }
2960 /*
2961 * wakeup processes that are waiting on memory if we hit a
2962 * high water mark. And wakeup scheduler process if we have
2963 * lots of memory. this process will swapin processes.
2964 */
2965 if (vm_pages_needed && !vm_page_count_min()) {
2966 vm_pages_needed = 0;
2967 wakeup(&vm_cnt.v_free_count);
2968 }
2969 }
2970
2971 /*
2972 * vm_page_free_toq:
2973 *
2974 * Returns the given page to the free list,
2975 * disassociating it with any VM object.
2976 *
2977 * The object must be locked. The page must be locked if it is managed.
2978 */
2979 void
vm_page_free_toq(vm_page_t m)2980 vm_page_free_toq(vm_page_t m)
2981 {
2982 int can_cache;
2983
2984 if ((m->oflags & VPO_UNMANAGED) == 0) {
2985 vm_page_lock_assert(m, MA_OWNED);
2986 KASSERT(!pmap_page_is_mapped(m),
2987 ("vm_page_free_toq: freeing mapped page %p", m));
2988 } else
2989 KASSERT(m->queue == PQ_NONE,
2990 ("vm_page_free_toq: unmanaged page %p is queued", m));
2991 PCPU_INC(cnt.v_tfree);
2992
2993 if (vm_page_sbusied(m))
2994 panic("vm_page_free: freeing busy page %p", m);
2995
2996 if (m->object != NULL) {
2997 VM_OBJECT_ASSERT_LOCKED(m->object);
2998 can_cache = ((m->object->flags & OBJ_COLORED) == 0);
2999 } else
3000 can_cache = 0;
3001
3002 /*
3003 * Unqueue, then remove page. Note that we cannot destroy
3004 * the page here because we do not want to call the pager's
3005 * callback routine until after we've put the page on the
3006 * appropriate free queue.
3007 */
3008 vm_page_remque(m);
3009 vm_page_remove(m);
3010
3011 /*
3012 * If fictitious remove object association and
3013 * return, otherwise delay object association removal.
3014 */
3015 if ((m->flags & PG_FICTITIOUS) != 0) {
3016 return;
3017 }
3018
3019 m->valid = 0;
3020 vm_page_undirty(m);
3021
3022 if (m->wire_count != 0)
3023 panic("vm_page_free: freeing wired page %p", m);
3024 if (m->hold_count != 0) {
3025 m->flags &= ~PG_ZERO;
3026 KASSERT((m->flags & PG_UNHOLDFREE) == 0,
3027 ("vm_page_free: freeing PG_UNHOLDFREE page %p", m));
3028 m->flags |= PG_UNHOLDFREE;
3029 } else {
3030 /*
3031 * Restore the default memory attribute to the page.
3032 */
3033 if (pmap_page_get_memattr(m) != VM_MEMATTR_DEFAULT)
3034 pmap_page_set_memattr(m, VM_MEMATTR_DEFAULT);
3035
3036 if (can_cache) {
3037 vm_page_percpu_free(m);
3038 return;
3039 }
3040 /*
3041 * Insert the page into the physical memory allocator's
3042 * cache/free page queues.
3043 */
3044 mtx_lock(&vm_page_queue_free_mtx);
3045 vm_phys_freecnt_adj(m, 1);
3046 #if VM_NRESERVLEVEL > 0
3047 if (!vm_reserv_free_page(m))
3048 #else
3049 if (TRUE)
3050 #endif
3051 vm_phys_free_pages(m, 0);
3052 if ((m->flags & PG_ZERO) != 0)
3053 ++vm_page_zero_count;
3054 else
3055 vm_page_zero_idle_wakeup();
3056 vm_page_free_wakeup();
3057 mtx_unlock(&vm_page_queue_free_mtx);
3058 }
3059 }
3060
3061 /*
3062 * vm_page_wire:
3063 *
3064 * Mark this page as wired down by yet
3065 * another map, removing it from paging queues
3066 * as necessary.
3067 *
3068 * If the page is fictitious, then its wire count must remain one.
3069 *
3070 * The page must be locked.
3071 */
3072 void
vm_page_wire(vm_page_t m)3073 vm_page_wire(vm_page_t m)
3074 {
3075
3076 /*
3077 * Only bump the wire statistics if the page is not already wired,
3078 * and only unqueue the page if it is on some queue (if it is unmanaged
3079 * it is already off the queues).
3080 */
3081 vm_page_lock_assert(m, MA_OWNED);
3082 if ((m->flags & PG_FICTITIOUS) != 0) {
3083 KASSERT(m->wire_count == 1,
3084 ("vm_page_wire: fictitious page %p's wire count isn't one",
3085 m));
3086 return;
3087 }
3088 if (m->wire_count == 0) {
3089 KASSERT((m->oflags & VPO_UNMANAGED) == 0 ||
3090 m->queue == PQ_NONE,
3091 ("vm_page_wire: unmanaged page %p is queued", m));
3092 vm_page_remque(m);
3093 atomic_add_int(&vm_cnt.v_wire_count, 1);
3094 }
3095 m->wire_count++;
3096 KASSERT(m->wire_count != 0, ("vm_page_wire: wire_count overflow m=%p", m));
3097 }
3098
3099 /*
3100 * vm_page_unwire:
3101 *
3102 * Release one wiring of the specified page, potentially allowing it to be
3103 * paged out. Returns TRUE if the number of wirings transitions to zero and
3104 * FALSE otherwise.
3105 *
3106 * Only managed pages belonging to an object can be paged out. If the number
3107 * of wirings transitions to zero and the page is eligible for page out, then
3108 * the page is added to the specified paging queue (unless PQ_NONE is
3109 * specified).
3110 *
3111 * If a page is fictitious, then its wire count must always be one.
3112 *
3113 * A managed page must be locked.
3114 */
3115 boolean_t
vm_page_unwire(vm_page_t m,uint8_t queue)3116 vm_page_unwire(vm_page_t m, uint8_t queue)
3117 {
3118
3119 KASSERT(queue < PQ_COUNT || queue == PQ_NONE,
3120 ("vm_page_unwire: invalid queue %u request for page %p",
3121 queue, m));
3122 if ((m->oflags & VPO_UNMANAGED) == 0)
3123 vm_page_assert_locked(m);
3124 if ((m->flags & PG_FICTITIOUS) != 0) {
3125 KASSERT(m->wire_count == 1,
3126 ("vm_page_unwire: fictitious page %p's wire count isn't one", m));
3127 return (FALSE);
3128 }
3129 if (m->wire_count > 0) {
3130 m->wire_count--;
3131 if (m->wire_count == 0) {
3132 atomic_subtract_int(&vm_cnt.v_wire_count, 1);
3133 if ((m->oflags & VPO_UNMANAGED) == 0 &&
3134 m->object != NULL && queue != PQ_NONE) {
3135 if (queue == PQ_INACTIVE)
3136 m->flags &= ~PG_WINATCFLS;
3137 vm_page_enqueue(queue, m);
3138 }
3139 return (TRUE);
3140 } else
3141 return (FALSE);
3142 } else
3143 panic("vm_page_unwire: page %p's wire count is zero", m);
3144 }
3145
3146 /*
3147 * Move the specified page to the inactive queue.
3148 *
3149 * Many pages placed on the inactive queue should actually go
3150 * into the cache, but it is difficult to figure out which. What
3151 * we do instead, if the inactive target is well met, is to put
3152 * clean pages at the head of the inactive queue instead of the tail.
3153 * This will cause them to be moved to the cache more quickly and
3154 * if not actively re-referenced, reclaimed more quickly. If we just
3155 * stick these pages at the end of the inactive queue, heavy filesystem
3156 * meta-data accesses can cause an unnecessary paging load on memory bound
3157 * processes. This optimization causes one-time-use metadata to be
3158 * reused more quickly.
3159 *
3160 * Normally noreuse is FALSE, resulting in LRU operation. noreuse is set
3161 * to TRUE if we want this page to be 'as if it were placed in the cache',
3162 * except without unmapping it from the process address space. In
3163 * practice this is implemented by inserting the page at the head of the
3164 * queue, using a marker page to guide FIFO insertion ordering.
3165 *
3166 * The page must be locked.
3167 */
3168 static inline void
_vm_page_deactivate(vm_page_t m,boolean_t noreuse)3169 _vm_page_deactivate(vm_page_t m, boolean_t noreuse)
3170 {
3171 struct vm_pagequeue *pq;
3172 int queue;
3173
3174 vm_page_assert_locked(m);
3175
3176 /*
3177 * Ignore if the page is already inactive, unless it is unlikely to be
3178 * reactivated.
3179 */
3180 if ((queue = m->queue) == PQ_INACTIVE && !noreuse)
3181 return;
3182 if (m->wire_count == 0 && (m->oflags & VPO_UNMANAGED) == 0) {
3183 pq = &vm_phys_domain(m)->vmd_pagequeues[PQ_INACTIVE];
3184 /* Avoid multiple acquisitions of the inactive queue lock. */
3185 if (queue == PQ_INACTIVE) {
3186 vm_pagequeue_lock(pq);
3187 vm_page_dequeue_locked(m);
3188 } else {
3189 if (queue != PQ_NONE)
3190 vm_page_dequeue(m);
3191 m->flags &= ~PG_WINATCFLS;
3192 vm_pagequeue_lock(pq);
3193 }
3194 m->queue = PQ_INACTIVE;
3195 if (noreuse)
3196 TAILQ_INSERT_BEFORE(&vm_phys_domain(m)->vmd_inacthead,
3197 m, plinks.q);
3198 else
3199 TAILQ_INSERT_TAIL(&pq->pq_pl, m, plinks.q);
3200 vm_pagequeue_cnt_inc(pq);
3201 vm_pagequeue_unlock(pq);
3202 }
3203 }
3204
3205 /*
3206 * Move the specified page to the inactive queue.
3207 *
3208 * The page must be locked.
3209 */
3210 void
vm_page_deactivate(vm_page_t m)3211 vm_page_deactivate(vm_page_t m)
3212 {
3213
3214 _vm_page_deactivate(m, FALSE);
3215 }
3216
3217 /*
3218 * Move the specified page to the inactive queue with the expectation
3219 * that it is unlikely to be reused.
3220 *
3221 * The page must be locked.
3222 */
3223 void
vm_page_deactivate_noreuse(vm_page_t m)3224 vm_page_deactivate_noreuse(vm_page_t m)
3225 {
3226
3227 _vm_page_deactivate(m, TRUE);
3228 }
3229
3230 /*
3231 * vm_page_try_to_free()
3232 *
3233 * Attempt to free the page. If we cannot free it, we do nothing.
3234 * 1 is returned on success, 0 on failure.
3235 */
3236 int
vm_page_try_to_free(vm_page_t m)3237 vm_page_try_to_free(vm_page_t m)
3238 {
3239
3240 vm_page_lock_assert(m, MA_OWNED);
3241 if (m->object != NULL)
3242 VM_OBJECT_ASSERT_WLOCKED(m->object);
3243 if (m->dirty || m->hold_count || m->wire_count ||
3244 (m->oflags & VPO_UNMANAGED) != 0 || vm_page_busied(m))
3245 return (0);
3246 pmap_remove_all(m);
3247 if (m->dirty)
3248 return (0);
3249 vm_page_free(m);
3250 return (1);
3251 }
3252
3253 /*
3254 * vm_page_advise
3255 *
3256 * Deactivate or do nothing, as appropriate.
3257 *
3258 * The object and page must be locked.
3259 */
3260 void
vm_page_advise(vm_page_t m,int advice)3261 vm_page_advise(vm_page_t m, int advice)
3262 {
3263
3264 vm_page_assert_locked(m);
3265 VM_OBJECT_ASSERT_WLOCKED(m->object);
3266 if (advice == MADV_FREE)
3267 /*
3268 * Mark the page clean. This will allow the page to be freed
3269 * up by the system. However, such pages are often reused
3270 * quickly by malloc() so we do not do anything that would
3271 * cause a page fault if we can help it.
3272 *
3273 * Specifically, we do not try to actually free the page now
3274 * nor do we try to put it in the cache (which would cause a
3275 * page fault on reuse).
3276 *
3277 * But we do make the page as freeable as we can without
3278 * actually taking the step of unmapping it.
3279 */
3280 m->dirty = 0;
3281 else if (advice != MADV_DONTNEED)
3282 return;
3283
3284 /*
3285 * Clear any references to the page. Otherwise, the page daemon will
3286 * immediately reactivate the page.
3287 */
3288 vm_page_aflag_clear(m, PGA_REFERENCED);
3289
3290 if (advice != MADV_FREE && m->dirty == 0 && pmap_is_modified(m))
3291 vm_page_dirty(m);
3292
3293 /*
3294 * Place clean pages at the head of the inactive queue rather than the
3295 * tail, thus defeating the queue's LRU operation and ensuring that the
3296 * page will be reused quickly.
3297 */
3298 _vm_page_deactivate(m, m->dirty == 0);
3299 }
3300
3301 /*
3302 * Grab a page, waiting until we are waken up due to the page
3303 * changing state. We keep on waiting, if the page continues
3304 * to be in the object. If the page doesn't exist, first allocate it
3305 * and then conditionally zero it.
3306 *
3307 * This routine may sleep.
3308 *
3309 * The object must be locked on entry. The lock will, however, be released
3310 * and reacquired if the routine sleeps.
3311 */
3312 vm_page_t
vm_page_grab(vm_object_t object,vm_pindex_t pindex,int allocflags)3313 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3314 {
3315 vm_page_t m;
3316 int sleep;
3317
3318 VM_OBJECT_ASSERT_WLOCKED(object);
3319 KASSERT((allocflags & VM_ALLOC_SBUSY) == 0 ||
3320 (allocflags & VM_ALLOC_IGN_SBUSY) != 0,
3321 ("vm_page_grab: VM_ALLOC_SBUSY/VM_ALLOC_IGN_SBUSY mismatch"));
3322 retrylookup:
3323 if ((m = vm_page_lookup(object, pindex)) != NULL) {
3324 sleep = (allocflags & VM_ALLOC_IGN_SBUSY) != 0 ?
3325 vm_page_xbusied(m) : vm_page_busied(m);
3326 if (sleep) {
3327 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3328 return (NULL);
3329 /*
3330 * Reference the page before unlocking and
3331 * sleeping so that the page daemon is less
3332 * likely to reclaim it.
3333 */
3334 vm_page_aflag_set(m, PGA_REFERENCED);
3335 vm_page_lock(m);
3336 VM_OBJECT_WUNLOCK(object);
3337 vm_page_busy_sleep(m, "pgrbwt");
3338 VM_OBJECT_WLOCK(object);
3339 goto retrylookup;
3340 } else {
3341 if ((allocflags & VM_ALLOC_WIRED) != 0) {
3342 vm_page_lock(m);
3343 vm_page_wire(m);
3344 vm_page_unlock(m);
3345 }
3346 if ((allocflags &
3347 (VM_ALLOC_NOBUSY | VM_ALLOC_SBUSY)) == 0)
3348 vm_page_xbusy(m);
3349 if ((allocflags & VM_ALLOC_SBUSY) != 0)
3350 vm_page_sbusy(m);
3351 return (m);
3352 }
3353 }
3354 m = vm_page_alloc(object, pindex, allocflags);
3355 if (m == NULL) {
3356 if ((allocflags & VM_ALLOC_NOWAIT) != 0)
3357 return (NULL);
3358 VM_OBJECT_WUNLOCK(object);
3359 VM_WAIT;
3360 VM_OBJECT_WLOCK(object);
3361 goto retrylookup;
3362 } else if (m->valid != 0)
3363 return (m);
3364 if (allocflags & VM_ALLOC_ZERO && (m->flags & PG_ZERO) == 0)
3365 pmap_zero_page(m);
3366 return (m);
3367 }
3368
3369 /*
3370 * Mapping function for valid or dirty bits in a page.
3371 *
3372 * Inputs are required to range within a page.
3373 */
3374 vm_page_bits_t
vm_page_bits(int base,int size)3375 vm_page_bits(int base, int size)
3376 {
3377 int first_bit;
3378 int last_bit;
3379
3380 KASSERT(
3381 base + size <= PAGE_SIZE,
3382 ("vm_page_bits: illegal base/size %d/%d", base, size)
3383 );
3384
3385 if (size == 0) /* handle degenerate case */
3386 return (0);
3387
3388 first_bit = base >> DEV_BSHIFT;
3389 last_bit = (base + size - 1) >> DEV_BSHIFT;
3390
3391 return (((vm_page_bits_t)2 << last_bit) -
3392 ((vm_page_bits_t)1 << first_bit));
3393 }
3394
3395 /*
3396 * vm_page_set_valid_range:
3397 *
3398 * Sets portions of a page valid. The arguments are expected
3399 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3400 * of any partial chunks touched by the range. The invalid portion of
3401 * such chunks will be zeroed.
3402 *
3403 * (base + size) must be less then or equal to PAGE_SIZE.
3404 */
3405 void
vm_page_set_valid_range(vm_page_t m,int base,int size)3406 vm_page_set_valid_range(vm_page_t m, int base, int size)
3407 {
3408 int endoff, frag;
3409
3410 VM_OBJECT_ASSERT_WLOCKED(m->object);
3411 if (size == 0) /* handle degenerate case */
3412 return;
3413
3414 /*
3415 * If the base is not DEV_BSIZE aligned and the valid
3416 * bit is clear, we have to zero out a portion of the
3417 * first block.
3418 */
3419 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
3420 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0)
3421 pmap_zero_page_area(m, frag, base - frag);
3422
3423 /*
3424 * If the ending offset is not DEV_BSIZE aligned and the
3425 * valid bit is clear, we have to zero out a portion of
3426 * the last block.
3427 */
3428 endoff = base + size;
3429 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
3430 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0)
3431 pmap_zero_page_area(m, endoff,
3432 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3433
3434 /*
3435 * Assert that no previously invalid block that is now being validated
3436 * is already dirty.
3437 */
3438 KASSERT((~m->valid & vm_page_bits(base, size) & m->dirty) == 0,
3439 ("vm_page_set_valid_range: page %p is dirty", m));
3440
3441 /*
3442 * Set valid bits inclusive of any overlap.
3443 */
3444 m->valid |= vm_page_bits(base, size);
3445 }
3446
3447 /*
3448 * Clear the given bits from the specified page's dirty field.
3449 */
3450 static __inline void
vm_page_clear_dirty_mask(vm_page_t m,vm_page_bits_t pagebits)3451 vm_page_clear_dirty_mask(vm_page_t m, vm_page_bits_t pagebits)
3452 {
3453 uintptr_t addr;
3454 #if PAGE_SIZE < 16384
3455 int shift;
3456 #endif
3457
3458 /*
3459 * If the object is locked and the page is neither exclusive busy nor
3460 * write mapped, then the page's dirty field cannot possibly be
3461 * set by a concurrent pmap operation.
3462 */
3463 VM_OBJECT_ASSERT_WLOCKED(m->object);
3464 if (!vm_page_xbusied(m) && !pmap_page_is_write_mapped(m))
3465 m->dirty &= ~pagebits;
3466 else {
3467 /*
3468 * The pmap layer can call vm_page_dirty() without
3469 * holding a distinguished lock. The combination of
3470 * the object's lock and an atomic operation suffice
3471 * to guarantee consistency of the page dirty field.
3472 *
3473 * For PAGE_SIZE == 32768 case, compiler already
3474 * properly aligns the dirty field, so no forcible
3475 * alignment is needed. Only require existence of
3476 * atomic_clear_64 when page size is 32768.
3477 */
3478 addr = (uintptr_t)&m->dirty;
3479 #if PAGE_SIZE == 32768
3480 atomic_clear_64((uint64_t *)addr, pagebits);
3481 #elif PAGE_SIZE == 16384
3482 atomic_clear_32((uint32_t *)addr, pagebits);
3483 #else /* PAGE_SIZE <= 8192 */
3484 /*
3485 * Use a trick to perform a 32-bit atomic on the
3486 * containing aligned word, to not depend on the existence
3487 * of atomic_clear_{8, 16}.
3488 */
3489 shift = addr & (sizeof(uint32_t) - 1);
3490 #if BYTE_ORDER == BIG_ENDIAN
3491 shift = (sizeof(uint32_t) - sizeof(m->dirty) - shift) * NBBY;
3492 #else
3493 shift *= NBBY;
3494 #endif
3495 addr &= ~(sizeof(uint32_t) - 1);
3496 atomic_clear_32((uint32_t *)addr, pagebits << shift);
3497 #endif /* PAGE_SIZE */
3498 }
3499 }
3500
3501 /*
3502 * vm_page_set_validclean:
3503 *
3504 * Sets portions of a page valid and clean. The arguments are expected
3505 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3506 * of any partial chunks touched by the range. The invalid portion of
3507 * such chunks will be zero'd.
3508 *
3509 * (base + size) must be less then or equal to PAGE_SIZE.
3510 */
3511 void
vm_page_set_validclean(vm_page_t m,int base,int size)3512 vm_page_set_validclean(vm_page_t m, int base, int size)
3513 {
3514 vm_page_bits_t oldvalid, pagebits;
3515 int endoff, frag;
3516
3517 VM_OBJECT_ASSERT_WLOCKED(m->object);
3518 if (size == 0) /* handle degenerate case */
3519 return;
3520
3521 /*
3522 * If the base is not DEV_BSIZE aligned and the valid
3523 * bit is clear, we have to zero out a portion of the
3524 * first block.
3525 */
3526 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
3527 (m->valid & ((vm_page_bits_t)1 << (base >> DEV_BSHIFT))) == 0)
3528 pmap_zero_page_area(m, frag, base - frag);
3529
3530 /*
3531 * If the ending offset is not DEV_BSIZE aligned and the
3532 * valid bit is clear, we have to zero out a portion of
3533 * the last block.
3534 */
3535 endoff = base + size;
3536 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
3537 (m->valid & ((vm_page_bits_t)1 << (endoff >> DEV_BSHIFT))) == 0)
3538 pmap_zero_page_area(m, endoff,
3539 DEV_BSIZE - (endoff & (DEV_BSIZE - 1)));
3540
3541 /*
3542 * Set valid, clear dirty bits. If validating the entire
3543 * page we can safely clear the pmap modify bit. We also
3544 * use this opportunity to clear the VPO_NOSYNC flag. If a process
3545 * takes a write fault on a MAP_NOSYNC memory area the flag will
3546 * be set again.
3547 *
3548 * We set valid bits inclusive of any overlap, but we can only
3549 * clear dirty bits for DEV_BSIZE chunks that are fully within
3550 * the range.
3551 */
3552 oldvalid = m->valid;
3553 pagebits = vm_page_bits(base, size);
3554 m->valid |= pagebits;
3555 #if 0 /* NOT YET */
3556 if ((frag = base & (DEV_BSIZE - 1)) != 0) {
3557 frag = DEV_BSIZE - frag;
3558 base += frag;
3559 size -= frag;
3560 if (size < 0)
3561 size = 0;
3562 }
3563 pagebits = vm_page_bits(base, size & (DEV_BSIZE - 1));
3564 #endif
3565 if (base == 0 && size == PAGE_SIZE) {
3566 /*
3567 * The page can only be modified within the pmap if it is
3568 * mapped, and it can only be mapped if it was previously
3569 * fully valid.
3570 */
3571 if (oldvalid == VM_PAGE_BITS_ALL)
3572 /*
3573 * Perform the pmap_clear_modify() first. Otherwise,
3574 * a concurrent pmap operation, such as
3575 * pmap_protect(), could clear a modification in the
3576 * pmap and set the dirty field on the page before
3577 * pmap_clear_modify() had begun and after the dirty
3578 * field was cleared here.
3579 */
3580 pmap_clear_modify(m);
3581 m->dirty = 0;
3582 m->oflags &= ~VPO_NOSYNC;
3583 } else if (oldvalid != VM_PAGE_BITS_ALL)
3584 m->dirty &= ~pagebits;
3585 else
3586 vm_page_clear_dirty_mask(m, pagebits);
3587 }
3588
3589 void
vm_page_clear_dirty(vm_page_t m,int base,int size)3590 vm_page_clear_dirty(vm_page_t m, int base, int size)
3591 {
3592
3593 vm_page_clear_dirty_mask(m, vm_page_bits(base, size));
3594 }
3595
3596 /*
3597 * vm_page_set_invalid:
3598 *
3599 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3600 * valid and dirty bits for the effected areas are cleared.
3601 */
3602 void
vm_page_set_invalid(vm_page_t m,int base,int size)3603 vm_page_set_invalid(vm_page_t m, int base, int size)
3604 {
3605 vm_page_bits_t bits;
3606 vm_object_t object;
3607
3608 object = m->object;
3609 VM_OBJECT_ASSERT_WLOCKED(object);
3610 if (object->type == OBJT_VNODE && base == 0 && IDX_TO_OFF(m->pindex) +
3611 size >= object->un_pager.vnp.vnp_size)
3612 bits = VM_PAGE_BITS_ALL;
3613 else
3614 bits = vm_page_bits(base, size);
3615 if (object->ref_count != 0 && m->valid == VM_PAGE_BITS_ALL &&
3616 bits != 0)
3617 pmap_remove_all(m);
3618 KASSERT((bits == 0 && m->valid == VM_PAGE_BITS_ALL) ||
3619 !pmap_page_is_mapped(m),
3620 ("vm_page_set_invalid: page %p is mapped", m));
3621 m->valid &= ~bits;
3622 m->dirty &= ~bits;
3623 }
3624
3625 /*
3626 * vm_page_zero_invalid()
3627 *
3628 * The kernel assumes that the invalid portions of a page contain
3629 * garbage, but such pages can be mapped into memory by user code.
3630 * When this occurs, we must zero out the non-valid portions of the
3631 * page so user code sees what it expects.
3632 *
3633 * Pages are most often semi-valid when the end of a file is mapped
3634 * into memory and the file's size is not page aligned.
3635 */
3636 void
vm_page_zero_invalid(vm_page_t m,boolean_t setvalid)3637 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3638 {
3639 int b;
3640 int i;
3641
3642 VM_OBJECT_ASSERT_WLOCKED(m->object);
3643 /*
3644 * Scan the valid bits looking for invalid sections that
3645 * must be zeroed. Invalid sub-DEV_BSIZE'd areas ( where the
3646 * valid bit may be set ) have already been zeroed by
3647 * vm_page_set_validclean().
3648 */
3649 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3650 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3651 (m->valid & ((vm_page_bits_t)1 << i))) {
3652 if (i > b) {
3653 pmap_zero_page_area(m,
3654 b << DEV_BSHIFT, (i - b) << DEV_BSHIFT);
3655 }
3656 b = i + 1;
3657 }
3658 }
3659
3660 /*
3661 * setvalid is TRUE when we can safely set the zero'd areas
3662 * as being valid. We can do this if there are no cache consistancy
3663 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3664 */
3665 if (setvalid)
3666 m->valid = VM_PAGE_BITS_ALL;
3667 }
3668
3669 /*
3670 * vm_page_is_valid:
3671 *
3672 * Is (partial) page valid? Note that the case where size == 0
3673 * will return FALSE in the degenerate case where the page is
3674 * entirely invalid, and TRUE otherwise.
3675 */
3676 int
vm_page_is_valid(vm_page_t m,int base,int size)3677 vm_page_is_valid(vm_page_t m, int base, int size)
3678 {
3679 vm_page_bits_t bits;
3680
3681 VM_OBJECT_ASSERT_LOCKED(m->object);
3682 bits = vm_page_bits(base, size);
3683 return (m->valid != 0 && (m->valid & bits) == bits);
3684 }
3685
3686 /*
3687 * vm_page_ps_is_valid:
3688 *
3689 * Returns TRUE if the entire (super)page is valid and FALSE otherwise.
3690 */
3691 boolean_t
vm_page_ps_is_valid(vm_page_t m)3692 vm_page_ps_is_valid(vm_page_t m)
3693 {
3694 int i, npages;
3695
3696 VM_OBJECT_ASSERT_LOCKED(m->object);
3697 npages = atop(pagesizes[m->psind]);
3698
3699 /*
3700 * The physically contiguous pages that make up a superpage, i.e., a
3701 * page with a page size index ("psind") greater than zero, will
3702 * occupy adjacent entries in vm_page_array[].
3703 */
3704 for (i = 0; i < npages; i++) {
3705 if (m[i].valid != VM_PAGE_BITS_ALL)
3706 return (FALSE);
3707 }
3708 return (TRUE);
3709 }
3710
3711 /*
3712 * Set the page's dirty bits if the page is modified.
3713 */
3714 void
vm_page_test_dirty(vm_page_t m)3715 vm_page_test_dirty(vm_page_t m)
3716 {
3717
3718 VM_OBJECT_ASSERT_WLOCKED(m->object);
3719 if (m->dirty != VM_PAGE_BITS_ALL && pmap_is_modified(m))
3720 vm_page_dirty(m);
3721 }
3722
3723 void
vm_page_lock_KBI(vm_page_t m,const char * file,int line)3724 vm_page_lock_KBI(vm_page_t m, const char *file, int line)
3725 {
3726
3727 mtx_lock_flags_(vm_page_lockptr(m), 0, file, line);
3728 }
3729
3730 void
vm_page_unlock_KBI(vm_page_t m,const char * file,int line)3731 vm_page_unlock_KBI(vm_page_t m, const char *file, int line)
3732 {
3733
3734 mtx_unlock_flags_(vm_page_lockptr(m), 0, file, line);
3735 }
3736
3737 int
vm_page_trylock_KBI(vm_page_t m,const char * file,int line)3738 vm_page_trylock_KBI(vm_page_t m, const char *file, int line)
3739 {
3740
3741 return (mtx_trylock_flags_(vm_page_lockptr(m), 0, file, line));
3742 }
3743
3744 #if defined(INVARIANTS) || defined(INVARIANT_SUPPORT)
3745 void
vm_page_assert_locked_KBI(vm_page_t m,const char * file,int line)3746 vm_page_assert_locked_KBI(vm_page_t m, const char *file, int line)
3747 {
3748
3749 vm_page_lock_assert_KBI(m, MA_OWNED, file, line);
3750 }
3751
3752 void
vm_page_lock_assert_KBI(vm_page_t m,int a,const char * file,int line)3753 vm_page_lock_assert_KBI(vm_page_t m, int a, const char *file, int line)
3754 {
3755
3756 mtx_assert_(vm_page_lockptr(m), a, file, line);
3757 }
3758 #endif
3759
3760 #ifdef INVARIANTS
3761 void
vm_page_object_lock_assert(vm_page_t m)3762 vm_page_object_lock_assert(vm_page_t m)
3763 {
3764
3765 /*
3766 * Certain of the page's fields may only be modified by the
3767 * holder of the containing object's lock or the exclusive busy.
3768 * holder. Unfortunately, the holder of the write busy is
3769 * not recorded, and thus cannot be checked here.
3770 */
3771 if (m->object != NULL && !vm_page_xbusied(m))
3772 VM_OBJECT_ASSERT_WLOCKED(m->object);
3773 }
3774
3775 void
vm_page_assert_pga_writeable(vm_page_t m,uint8_t bits)3776 vm_page_assert_pga_writeable(vm_page_t m, uint8_t bits)
3777 {
3778
3779 if ((bits & PGA_WRITEABLE) == 0)
3780 return;
3781
3782 /*
3783 * The PGA_WRITEABLE flag can only be set if the page is
3784 * managed, is exclusively busied or the object is locked.
3785 * Currently, this flag is only set by pmap_enter().
3786 */
3787 KASSERT((m->oflags & VPO_UNMANAGED) == 0,
3788 ("PGA_WRITEABLE on unmanaged page"));
3789 if (!vm_page_xbusied(m))
3790 VM_OBJECT_ASSERT_LOCKED(m->object);
3791 }
3792 #endif
3793
3794 #include "opt_ddb.h"
3795 #ifdef DDB
3796 #include <sys/kernel.h>
3797
3798 #include <ddb/ddb.h>
3799
DB_SHOW_COMMAND(page,vm_page_print_page_info)3800 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3801 {
3802 db_printf("vm_cnt.v_free_count: %d\n", vm_cnt.v_free_count);
3803 db_printf("vm_cnt.v_inactive_count: %d\n", vm_cnt.v_inactive_count);
3804 db_printf("vm_cnt.v_active_count: %d\n", vm_cnt.v_active_count);
3805 db_printf("vm_cnt.v_wire_count: %d\n", vm_cnt.v_wire_count);
3806 db_printf("vm_cnt.v_free_reserved: %d\n", vm_cnt.v_free_reserved);
3807 db_printf("vm_cnt.v_free_min: %d\n", vm_cnt.v_free_min);
3808 db_printf("vm_cnt.v_free_target: %d\n", vm_cnt.v_free_target);
3809 db_printf("vm_cnt.v_inactive_target: %d\n", vm_cnt.v_inactive_target);
3810 }
3811
DB_SHOW_COMMAND(pageq,vm_page_print_pageq_info)3812 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3813 {
3814 int dom;
3815
3816 db_printf("pq_free %d\n", vm_cnt.v_free_count);
3817 for (dom = 0; dom < vm_ndomains; dom++) {
3818 db_printf(
3819 "dom %d page_cnt %d free %d pq_act %d pq_inact %d pass %d\n",
3820 dom,
3821 vm_dom[dom].vmd_page_count,
3822 vm_dom[dom].vmd_free_count,
3823 vm_dom[dom].vmd_pagequeues[PQ_ACTIVE].pq_cnt,
3824 vm_dom[dom].vmd_pagequeues[PQ_INACTIVE].pq_cnt,
3825 vm_dom[dom].vmd_pass);
3826 }
3827 }
3828
DB_SHOW_COMMAND(pginfo,vm_page_print_pginfo)3829 DB_SHOW_COMMAND(pginfo, vm_page_print_pginfo)
3830 {
3831 vm_page_t m;
3832 boolean_t phys;
3833
3834 if (!have_addr) {
3835 db_printf("show pginfo addr\n");
3836 return;
3837 }
3838
3839 phys = strchr(modif, 'p') != NULL;
3840 if (phys)
3841 m = PHYS_TO_VM_PAGE(addr);
3842 else
3843 m = (vm_page_t)addr;
3844 db_printf(
3845 "page %p obj %p pidx 0x%jx phys 0x%jx q %d hold %d wire %d\n"
3846 " af 0x%x of 0x%x f 0x%x act %d busy %x valid 0x%x dirty 0x%x\n",
3847 m, m->object, (uintmax_t)m->pindex, (uintmax_t)m->phys_addr,
3848 m->queue, m->hold_count, m->wire_count, m->aflags, m->oflags,
3849 m->flags, m->act_count, m->busy_lock, m->valid, m->dirty);
3850 }
3851 #endif /* DDB */
3852