2 * Copyright (c) 1991 Regents of the University of California.
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
33 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
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.
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.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
63 * Resident memory management module. The module manipulates 'VM pages'.
64 * A VM page is the core building block for memory management.
67 #include <sys/param.h>
68 #include <sys/systm.h>
69 #include <sys/malloc.h>
71 #include <sys/vmmeter.h>
72 #include <sys/vnode.h>
73 #include <sys/kernel.h>
74 #include <sys/alist.h>
75 #include <sys/sysctl.h>
76 #include <sys/cpu_topology.h>
79 #include <vm/vm_param.h>
81 #include <vm/vm_kern.h>
83 #include <vm/vm_map.h>
84 #include <vm/vm_object.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_pager.h>
88 #include <vm/vm_extern.h>
89 #include <vm/swap_pager.h>
91 #include <machine/inttypes.h>
92 #include <machine/md_var.h>
93 #include <machine/specialreg.h>
95 #include <vm/vm_page2.h>
96 #include <sys/spinlock2.h>
99 * Action hash for user umtx support.
101 #define VMACTION_HSIZE 256
102 #define VMACTION_HMASK (VMACTION_HSIZE - 1)
105 * SET - Minimum required set associative size, must be a power of 2. We
106 * want this to match or exceed the set-associativeness of the cpu.
108 * GRP - A larger set that allows bleed-over into the domains of other
109 * nearby cpus. Also must be a power of 2. Used by the page zeroing
110 * code to smooth things out a bit.
112 #define PQ_SET_ASSOC 16
113 #define PQ_SET_ASSOC_MASK (PQ_SET_ASSOC - 1)
115 #define PQ_GRP_ASSOC (PQ_SET_ASSOC * 2)
116 #define PQ_GRP_ASSOC_MASK (PQ_GRP_ASSOC - 1)
118 static void vm_page_queue_init(void);
119 static void vm_page_free_wakeup(void);
120 static vm_page_t vm_page_select_cache(u_short pg_color);
121 static vm_page_t _vm_page_list_find2(int basequeue, int index);
122 static void _vm_page_deactivate_locked(vm_page_t m, int athead);
125 * Array of tailq lists
127 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT];
129 LIST_HEAD(vm_page_action_list, vm_page_action);
131 struct vm_page_action_hash {
132 struct vm_page_action_list list;
136 struct vm_page_action_hash action_hash[VMACTION_HSIZE];
137 static volatile int vm_pages_waiting;
139 static struct alist vm_contig_alist;
140 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536];
141 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin");
143 static u_long vm_dma_reserved = 0;
144 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved);
145 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0,
146 "Memory reserved for DMA");
147 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD,
148 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA");
150 static int vm_contig_verbose = 0;
151 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose);
153 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
154 vm_pindex_t, pindex);
157 vm_page_queue_init(void)
161 for (i = 0; i < PQ_L2_SIZE; i++)
162 vm_page_queues[PQ_FREE+i].cnt_offset =
163 offsetof(struct vmstats, v_free_count);
164 for (i = 0; i < PQ_L2_SIZE; i++)
165 vm_page_queues[PQ_CACHE+i].cnt_offset =
166 offsetof(struct vmstats, v_cache_count);
167 for (i = 0; i < PQ_L2_SIZE; i++)
168 vm_page_queues[PQ_INACTIVE+i].cnt_offset =
169 offsetof(struct vmstats, v_inactive_count);
170 for (i = 0; i < PQ_L2_SIZE; i++)
171 vm_page_queues[PQ_ACTIVE+i].cnt_offset =
172 offsetof(struct vmstats, v_active_count);
173 for (i = 0; i < PQ_L2_SIZE; i++)
174 vm_page_queues[PQ_HOLD+i].cnt_offset =
175 offsetof(struct vmstats, v_active_count);
176 /* PQ_NONE has no queue */
178 for (i = 0; i < PQ_COUNT; i++) {
179 TAILQ_INIT(&vm_page_queues[i].pl);
180 spin_init(&vm_page_queues[i].spin, "vm_page_queue_init");
184 * NOTE: Action lock might recurse due to callback, so allow
187 for (i = 0; i < VMACTION_HSIZE; i++) {
188 LIST_INIT(&action_hash[i].list);
189 lockinit(&action_hash[i].lk, "actlk", 0, LK_CANRECURSE);
194 * note: place in initialized data section? Is this necessary?
197 int vm_page_array_size = 0;
198 vm_page_t vm_page_array = NULL;
199 vm_paddr_t vm_low_phys_reserved;
204 * Sets the page size, perhaps based upon the memory size.
205 * Must be called before any use of page-size dependent functions.
208 vm_set_page_size(void)
210 if (vmstats.v_page_size == 0)
211 vmstats.v_page_size = PAGE_SIZE;
212 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
213 panic("vm_set_page_size: page size not a power of two");
219 * Add a new page to the freelist for use by the system. New pages
220 * are added to both the head and tail of the associated free page
221 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
222 * requests pull 'recent' adds (higher physical addresses) first.
224 * Beware that the page zeroing daemon will also be running soon after
225 * boot, moving pages from the head to the tail of the PQ_FREE queues.
227 * Must be called in a critical section.
230 vm_add_new_page(vm_paddr_t pa)
232 struct vpgqueues *vpq;
235 m = PHYS_TO_VM_PAGE(pa);
238 m->pat_mode = PAT_WRITE_BACK;
239 m->pc = (pa >> PAGE_SHIFT);
242 * Twist for cpu localization in addition to page coloring, so
243 * different cpus selecting by m->queue get different page colors.
245 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE);
246 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE));
250 * Reserve a certain number of contiguous low memory pages for
251 * contigmalloc() to use.
253 if (pa < vm_low_phys_reserved) {
254 atomic_add_int(&vmstats.v_page_count, 1);
255 atomic_add_int(&vmstats.v_dma_pages, 1);
258 atomic_add_int(&vmstats.v_wire_count, 1);
259 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1);
266 m->queue = m->pc + PQ_FREE;
267 KKASSERT(m->dirty == 0);
269 atomic_add_int(&vmstats.v_page_count, 1);
270 atomic_add_int(&vmstats.v_free_count, 1);
271 vpq = &vm_page_queues[m->queue];
272 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
279 * Initializes the resident memory module.
281 * Preallocates memory for critical VM structures and arrays prior to
282 * kernel_map becoming available.
284 * Memory is allocated from (virtual2_start, virtual2_end) if available,
285 * otherwise memory is allocated from (virtual_start, virtual_end).
287 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
288 * large enough to hold vm_page_array & other structures for machines with
289 * large amounts of ram, so we want to use virtual2* when available.
292 vm_page_startup(void)
294 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start;
297 vm_paddr_t page_range;
303 vm_paddr_t biggestone, biggestsize;
310 vaddr = round_page(vaddr);
313 * Make sure ranges are page-aligned.
315 for (i = 0; phys_avail[i].phys_end; ++i) {
316 phys_avail[i].phys_beg = round_page64(phys_avail[i].phys_beg);
317 phys_avail[i].phys_end = trunc_page64(phys_avail[i].phys_end);
318 if (phys_avail[i].phys_end < phys_avail[i].phys_beg)
319 phys_avail[i].phys_end = phys_avail[i].phys_beg;
323 * Locate largest block
325 for (i = 0; phys_avail[i].phys_end; ++i) {
326 vm_paddr_t size = phys_avail[i].phys_end -
327 phys_avail[i].phys_beg;
329 if (size > biggestsize) {
335 --i; /* adjust to last entry for use down below */
337 end = phys_avail[biggestone].phys_end;
338 end = trunc_page(end);
341 * Initialize the queue headers for the free queue, the active queue
342 * and the inactive queue.
344 vm_page_queue_init();
346 #if !defined(_KERNEL_VIRTUAL)
348 * VKERNELs don't support minidumps and as such don't need
351 * Allocate a bitmap to indicate that a random physical page
352 * needs to be included in a minidump.
354 * The amd64 port needs this to indicate which direct map pages
355 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
357 * However, i386 still needs this workspace internally within the
358 * minidump code. In theory, they are not needed on i386, but are
359 * included should the sf_buf code decide to use them.
361 page_range = phys_avail[i].phys_end / PAGE_SIZE;
362 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
363 end -= vm_page_dump_size;
364 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
365 VM_PROT_READ | VM_PROT_WRITE);
366 bzero((void *)vm_page_dump, vm_page_dump_size);
369 * Compute the number of pages of memory that will be available for
370 * use (taking into account the overhead of a page structure per
373 first_page = phys_avail[0].phys_beg / PAGE_SIZE;
374 page_range = phys_avail[i].phys_end / PAGE_SIZE - first_page;
375 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
377 #ifndef _KERNEL_VIRTUAL
379 * (only applies to real kernels)
381 * Reserve a large amount of low memory for potential 32-bit DMA
382 * space allocations. Once device initialization is complete we
383 * release most of it, but keep (vm_dma_reserved) memory reserved
384 * for later use. Typically for X / graphics. Through trial and
385 * error we find that GPUs usually requires ~60-100MB or so.
387 * By default, 128M is left in reserve on machines with 2G+ of ram.
389 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT;
390 if (vm_low_phys_reserved > total / 4)
391 vm_low_phys_reserved = total / 4;
392 if (vm_dma_reserved == 0) {
393 vm_dma_reserved = 128 * 1024 * 1024; /* 128MB */
394 if (vm_dma_reserved > total / 16)
395 vm_dma_reserved = total / 16;
398 alist_init(&vm_contig_alist, 65536, vm_contig_ameta,
399 ALIST_RECORDS_65536);
402 * Initialize the mem entry structures now, and put them in the free
405 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
406 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE);
407 vm_page_array = (vm_page_t)mapped;
409 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
411 * since pmap_map on amd64 returns stuff out of a direct-map region,
412 * we have to manually add these pages to the minidump tracking so
413 * that they can be dumped, including the vm_page_array.
416 pa < phys_avail[biggestone].phys_end;
423 * Clear all of the page structures, run basic initialization so
424 * PHYS_TO_VM_PAGE() operates properly even on pages not in the
427 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
428 vm_page_array_size = page_range;
430 m = &vm_page_array[0];
431 pa = ptoa(first_page);
432 for (i = 0; i < page_range; ++i) {
433 spin_init(&m->spin, "vm_page");
440 * Construct the free queue(s) in ascending order (by physical
441 * address) so that the first 16MB of physical memory is allocated
442 * last rather than first. On large-memory machines, this avoids
443 * the exhaustion of low physical memory before isa_dmainit has run.
445 vmstats.v_page_count = 0;
446 vmstats.v_free_count = 0;
447 for (i = 0; phys_avail[i].phys_end && npages > 0; ++i) {
448 pa = phys_avail[i].phys_beg;
452 last_pa = phys_avail[i].phys_end;
453 while (pa < last_pa && npages-- > 0) {
459 virtual2_start = vaddr;
461 virtual_start = vaddr;
462 mycpu->gd_vmstats = vmstats;
466 * Reorganize VM pages based on numa data. May be called as many times as
467 * necessary. Will reorganize the vm_page_t page color and related queue(s)
468 * to allow vm_page_alloc() to choose pages based on socket affinity.
470 * NOTE: This function is only called while we are still in UP mode, so
471 * we only need a critical section to protect the queues (which
472 * saves a lot of time, there are likely a ton of pages).
475 vm_numa_organize(vm_paddr_t ran_beg, vm_paddr_t bytes, int physid)
480 struct vpgqueues *vpq;
488 * Check if no physical information, or there was only one socket
489 * (so don't waste time doing nothing!).
491 if (cpu_topology_phys_ids <= 1 ||
492 cpu_topology_core_ids == 0) {
497 * Setup for our iteration. Note that ACPI may iterate CPU
498 * sockets starting at 0 or 1 or some other number. The
499 * cpu_topology code mod's it against the socket count.
501 ran_end = ran_beg + bytes;
502 physid %= cpu_topology_phys_ids;
504 socket_mod = PQ_L2_SIZE / cpu_topology_phys_ids;
505 socket_value = physid * socket_mod;
506 mend = &vm_page_array[vm_page_array_size];
511 * Adjust vm_page->pc and requeue all affected pages. The
512 * allocator will then be able to localize memory allocations
515 for (i = 0; phys_avail[i].phys_end; ++i) {
516 scan_beg = phys_avail[i].phys_beg;
517 scan_end = phys_avail[i].phys_end;
518 if (scan_end <= ran_beg)
520 if (scan_beg >= ran_end)
522 if (scan_beg < ran_beg)
524 if (scan_end > ran_end)
526 if (atop(scan_end) > first_page + vm_page_array_size)
527 scan_end = ptoa(first_page + vm_page_array_size);
529 m = PHYS_TO_VM_PAGE(scan_beg);
530 while (scan_beg < scan_end) {
532 if (m->queue != PQ_NONE) {
533 vpq = &vm_page_queues[m->queue];
534 TAILQ_REMOVE(&vpq->pl, m, pageq);
536 /* queue doesn't change, no need to adj cnt */
539 m->pc += socket_value;
542 vpq = &vm_page_queues[m->queue];
543 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
545 /* queue doesn't change, no need to adj cnt */
548 m->pc += socket_value;
551 scan_beg += PAGE_SIZE;
559 * We tended to reserve a ton of memory for contigmalloc(). Now that most
560 * drivers have initialized we want to return most the remaining free
561 * reserve back to the VM page queues so they can be used for normal
564 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool.
567 vm_page_startup_finish(void *dummy __unused)
576 spin_lock(&vm_contig_spin);
578 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
579 if (bfree <= vm_dma_reserved / PAGE_SIZE)
585 * Figure out how much of the initial reserve we have to
586 * free in order to reach our target.
588 bfree -= vm_dma_reserved / PAGE_SIZE;
590 blk += count - bfree;
595 * Calculate the nearest power of 2 <= count.
597 for (xcount = 1; xcount <= count; xcount <<= 1)
600 blk += count - xcount;
604 * Allocate the pages from the alist, then free them to
605 * the normal VM page queues.
607 * Pages allocated from the alist are wired. We have to
608 * busy, unwire, and free them. We must also adjust
609 * vm_low_phys_reserved before freeing any pages to prevent
612 rblk = alist_alloc(&vm_contig_alist, blk, count);
614 kprintf("vm_page_startup_finish: Unable to return "
615 "dma space @0x%08x/%d -> 0x%08x\n",
619 atomic_add_int(&vmstats.v_dma_pages, -count);
620 spin_unlock(&vm_contig_spin);
622 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
623 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m);
625 vm_page_busy_wait(m, FALSE, "cpgfr");
626 vm_page_unwire(m, 0);
631 spin_lock(&vm_contig_spin);
633 spin_unlock(&vm_contig_spin);
636 * Print out how much DMA space drivers have already allocated and
637 * how much is left over.
639 kprintf("DMA space used: %jdk, remaining available: %jdk\n",
640 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) *
642 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024));
644 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY,
645 vm_page_startup_finish, NULL);
649 * Scan comparison function for Red-Black tree scans. An inclusive
650 * (start,end) is expected. Other fields are not used.
653 rb_vm_page_scancmp(struct vm_page *p, void *data)
655 struct rb_vm_page_scan_info *info = data;
657 if (p->pindex < info->start_pindex)
659 if (p->pindex > info->end_pindex)
665 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
667 if (p1->pindex < p2->pindex)
669 if (p1->pindex > p2->pindex)
675 vm_page_init(vm_page_t m)
677 /* do nothing for now. Called from pmap_page_init() */
681 * Each page queue has its own spin lock, which is fairly optimal for
682 * allocating and freeing pages at least.
684 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
685 * queue spinlock via this function. Also note that m->queue cannot change
686 * unless both the page and queue are locked.
690 _vm_page_queue_spin_lock(vm_page_t m)
695 if (queue != PQ_NONE) {
696 spin_lock(&vm_page_queues[queue].spin);
697 KKASSERT(queue == m->queue);
703 _vm_page_queue_spin_unlock(vm_page_t m)
709 if (queue != PQ_NONE)
710 spin_unlock(&vm_page_queues[queue].spin);
715 _vm_page_queues_spin_lock(u_short queue)
718 if (queue != PQ_NONE)
719 spin_lock(&vm_page_queues[queue].spin);
725 _vm_page_queues_spin_unlock(u_short queue)
728 if (queue != PQ_NONE)
729 spin_unlock(&vm_page_queues[queue].spin);
733 vm_page_queue_spin_lock(vm_page_t m)
735 _vm_page_queue_spin_lock(m);
739 vm_page_queues_spin_lock(u_short queue)
741 _vm_page_queues_spin_lock(queue);
745 vm_page_queue_spin_unlock(vm_page_t m)
747 _vm_page_queue_spin_unlock(m);
751 vm_page_queues_spin_unlock(u_short queue)
753 _vm_page_queues_spin_unlock(queue);
757 * This locks the specified vm_page and its queue in the proper order
758 * (page first, then queue). The queue may change so the caller must
763 _vm_page_and_queue_spin_lock(vm_page_t m)
765 vm_page_spin_lock(m);
766 _vm_page_queue_spin_lock(m);
771 _vm_page_and_queue_spin_unlock(vm_page_t m)
773 _vm_page_queues_spin_unlock(m->queue);
774 vm_page_spin_unlock(m);
778 vm_page_and_queue_spin_unlock(vm_page_t m)
780 _vm_page_and_queue_spin_unlock(m);
784 vm_page_and_queue_spin_lock(vm_page_t m)
786 _vm_page_and_queue_spin_lock(m);
790 * Helper function removes vm_page from its current queue.
791 * Returns the base queue the page used to be on.
793 * The vm_page and the queue must be spinlocked.
794 * This function will unlock the queue but leave the page spinlocked.
796 static __inline u_short
797 _vm_page_rem_queue_spinlocked(vm_page_t m)
799 struct vpgqueues *pq;
805 if (queue != PQ_NONE) {
806 pq = &vm_page_queues[queue];
807 TAILQ_REMOVE(&pq->pl, m, pageq);
810 * Adjust our pcpu stats. In order for the nominal low-memory
811 * algorithms to work properly we don't let any pcpu stat get
812 * too negative before we force it to be rolled-up into the
813 * global stats. Otherwise our pageout and vm_wait tests
816 * The idea here is to reduce unnecessary SMP cache
817 * mastership changes in the global vmstats, which can be
818 * particularly bad in multi-socket systems.
820 cnt = (int *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
821 atomic_add_int(cnt, -1);
822 if (*cnt < -VMMETER_SLOP_COUNT) {
823 u_int copy = atomic_swap_int(cnt, 0);
824 cnt = (int *)((char *)&vmstats + pq->cnt_offset);
825 atomic_add_int(cnt, copy);
826 cnt = (int *)((char *)&mycpu->gd_vmstats +
828 atomic_add_int(cnt, copy);
834 vm_page_queues_spin_unlock(oqueue); /* intended */
840 * Helper function places the vm_page on the specified queue.
842 * The vm_page must be spinlocked.
843 * This function will return with both the page and the queue locked.
846 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
848 struct vpgqueues *pq;
851 KKASSERT(m->queue == PQ_NONE);
853 if (queue != PQ_NONE) {
854 vm_page_queues_spin_lock(queue);
855 pq = &vm_page_queues[queue];
859 * Adjust our pcpu stats. If a system entity really needs
860 * to incorporate the count it will call vmstats_rollup()
861 * to roll it all up into the global vmstats strufture.
863 cnt = (int *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
864 atomic_add_int(cnt, 1);
867 * PQ_FREE is always handled LIFO style to try to provide
868 * cache-hot pages to programs.
871 if (queue - m->pc == PQ_FREE) {
872 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
874 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
876 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
878 /* leave the queue spinlocked */
883 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
884 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
885 * did not. Only one sleep call will be made before returning.
887 * This function does NOT busy the page and on return the page is not
888 * guaranteed to be available.
891 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
899 if ((flags & PG_BUSY) == 0 &&
900 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
903 tsleep_interlock(m, 0);
904 if (atomic_cmpset_int(&m->flags, flags,
905 flags | PG_WANTED | PG_REFERENCED)) {
906 tsleep(m, PINTERLOCKED, msg, 0);
913 * This calculates and returns a page color given an optional VM object and
914 * either a pindex or an iterator. We attempt to return a cpu-localized
915 * pg_color that is still roughly 16-way set-associative. The CPU topology
916 * is used if it was probed.
918 * The caller may use the returned value to index into e.g. PQ_FREE when
919 * allocating a page in order to nominally obtain pages that are hopefully
920 * already localized to the requesting cpu. This function is not able to
921 * provide any sort of guarantee of this, but does its best to improve
922 * hardware cache management performance.
924 * WARNING! The caller must mask the returned value with PQ_L2_MASK.
927 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex)
934 phys_id = get_cpu_phys_id(cpuid);
935 core_id = get_cpu_core_id(cpuid);
936 object_pg_color = object ? object->pg_color : 0;
938 if (cpu_topology_phys_ids && cpu_topology_core_ids) {
942 * Break us down by socket and cpu
944 pg_color = phys_id * PQ_L2_SIZE / cpu_topology_phys_ids;
945 pg_color += core_id * PQ_L2_SIZE /
946 (cpu_topology_core_ids * cpu_topology_phys_ids);
949 * Calculate remaining component for object/queue color
951 grpsize = PQ_L2_SIZE / (cpu_topology_core_ids *
952 cpu_topology_phys_ids);
954 pg_color += (pindex + object_pg_color) % grpsize;
959 /* 3->9, 4->8, 5->10, 6->12, 7->14 */
964 pg_color += (pindex + object_pg_color) % grpsize;
968 * Unknown topology, distribute things evenly.
970 pg_color = cpuid * PQ_L2_SIZE / ncpus;
971 pg_color += pindex + object_pg_color;
973 return (pg_color & PQ_L2_MASK);
977 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
978 * also wait for m->busy to become 0 before setting PG_BUSY.
981 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
982 int also_m_busy, const char *msg
990 if (flags & PG_BUSY) {
991 tsleep_interlock(m, 0);
992 if (atomic_cmpset_int(&m->flags, flags,
993 flags | PG_WANTED | PG_REFERENCED)) {
994 tsleep(m, PINTERLOCKED, msg, 0);
996 } else if (also_m_busy && (flags & PG_SBUSY)) {
997 tsleep_interlock(m, 0);
998 if (atomic_cmpset_int(&m->flags, flags,
999 flags | PG_WANTED | PG_REFERENCED)) {
1000 tsleep(m, PINTERLOCKED, msg, 0);
1003 if (atomic_cmpset_int(&m->flags, flags,
1005 #ifdef VM_PAGE_DEBUG
1006 m->busy_func = func;
1007 m->busy_line = lineno;
1016 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
1019 * Returns non-zero on failure.
1022 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
1030 if (flags & PG_BUSY)
1032 if (also_m_busy && (flags & PG_SBUSY))
1034 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1035 #ifdef VM_PAGE_DEBUG
1036 m->busy_func = func;
1037 m->busy_line = lineno;
1045 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
1046 * that a wakeup() should be performed.
1048 * The vm_page must be spinlocked and will remain spinlocked on return.
1049 * The related queue must NOT be spinlocked (which could deadlock us).
1055 _vm_page_wakeup(vm_page_t m)
1062 if (atomic_cmpset_int(&m->flags, flags,
1063 flags & ~(PG_BUSY | PG_WANTED))) {
1067 return(flags & PG_WANTED);
1071 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
1072 * is typically the last call you make on a page before moving onto
1076 vm_page_wakeup(vm_page_t m)
1078 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
1079 vm_page_spin_lock(m);
1080 if (_vm_page_wakeup(m)) {
1081 vm_page_spin_unlock(m);
1084 vm_page_spin_unlock(m);
1089 * Holding a page keeps it from being reused. Other parts of the system
1090 * can still disassociate the page from its current object and free it, or
1091 * perform read or write I/O on it and/or otherwise manipulate the page,
1092 * but if the page is held the VM system will leave the page and its data
1093 * intact and not reuse the page for other purposes until the last hold
1094 * reference is released. (see vm_page_wire() if you want to prevent the
1095 * page from being disassociated from its object too).
1097 * The caller must still validate the contents of the page and, if necessary,
1098 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
1099 * before manipulating the page.
1101 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
1104 vm_page_hold(vm_page_t m)
1106 vm_page_spin_lock(m);
1107 atomic_add_int(&m->hold_count, 1);
1108 if (m->queue - m->pc == PQ_FREE) {
1109 _vm_page_queue_spin_lock(m);
1110 _vm_page_rem_queue_spinlocked(m);
1111 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
1112 _vm_page_queue_spin_unlock(m);
1114 vm_page_spin_unlock(m);
1118 * The opposite of vm_page_hold(). If the page is on the HOLD queue
1119 * it was freed while held and must be moved back to the FREE queue.
1122 vm_page_unhold(vm_page_t m)
1124 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE,
1125 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)",
1126 m, m->hold_count, m->queue - m->pc));
1127 vm_page_spin_lock(m);
1128 atomic_add_int(&m->hold_count, -1);
1129 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
1130 _vm_page_queue_spin_lock(m);
1131 _vm_page_rem_queue_spinlocked(m);
1132 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
1133 _vm_page_queue_spin_unlock(m);
1135 vm_page_spin_unlock(m);
1141 * Create a fictitious page with the specified physical address and
1142 * memory attribute. The memory attribute is the only the machine-
1143 * dependent aspect of a fictitious page that must be initialized.
1147 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1150 if ((m->flags & PG_FICTITIOUS) != 0) {
1152 * The page's memattr might have changed since the
1153 * previous initialization. Update the pmap to the
1158 m->phys_addr = paddr;
1160 /* Fictitious pages don't use "segind". */
1161 /* Fictitious pages don't use "order" or "pool". */
1162 m->flags = PG_FICTITIOUS | PG_UNMANAGED | PG_BUSY;
1164 spin_init(&m->spin, "fake_page");
1167 pmap_page_set_memattr(m, memattr);
1171 * Inserts the given vm_page into the object and object list.
1173 * The pagetables are not updated but will presumably fault the page
1174 * in if necessary, or if a kernel page the caller will at some point
1175 * enter the page into the kernel's pmap. We are not allowed to block
1176 * here so we *can't* do this anyway.
1178 * This routine may not block.
1179 * This routine must be called with the vm_object held.
1180 * This routine must be called with a critical section held.
1182 * This routine returns TRUE if the page was inserted into the object
1183 * successfully, and FALSE if the page already exists in the object.
1186 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1188 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object));
1189 if (m->object != NULL)
1190 panic("vm_page_insert: already inserted");
1192 object->generation++;
1195 * Record the object/offset pair in this page and add the
1196 * pv_list_count of the page to the object.
1198 * The vm_page spin lock is required for interactions with the pmap.
1200 vm_page_spin_lock(m);
1203 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
1206 vm_page_spin_unlock(m);
1209 ++object->resident_page_count;
1210 ++mycpu->gd_vmtotal.t_rm;
1211 vm_page_spin_unlock(m);
1214 * Since we are inserting a new and possibly dirty page,
1215 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
1217 if ((m->valid & m->dirty) ||
1218 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT)))
1219 vm_object_set_writeable_dirty(object);
1222 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
1224 swap_pager_page_inserted(m);
1229 * Removes the given vm_page_t from the (object,index) table
1231 * The underlying pmap entry (if any) is NOT removed here.
1232 * This routine may not block.
1234 * The page must be BUSY and will remain BUSY on return.
1235 * No other requirements.
1237 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
1241 vm_page_remove(vm_page_t m)
1245 if (m->object == NULL) {
1249 if ((m->flags & PG_BUSY) == 0)
1250 panic("vm_page_remove: page not busy");
1254 vm_object_hold(object);
1257 * Remove the page from the object and update the object.
1259 * The vm_page spin lock is required for interactions with the pmap.
1261 vm_page_spin_lock(m);
1262 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
1263 --object->resident_page_count;
1264 --mycpu->gd_vmtotal.t_rm;
1266 vm_page_spin_unlock(m);
1268 object->generation++;
1270 vm_object_drop(object);
1274 * Locate and return the page at (object, pindex), or NULL if the
1275 * page could not be found.
1277 * The caller must hold the vm_object token.
1280 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1285 * Search the hash table for this object/offset pair
1287 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1288 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1289 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1294 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1296 int also_m_busy, const char *msg
1302 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1303 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1305 KKASSERT(m->object == object && m->pindex == pindex);
1308 if (flags & PG_BUSY) {
1309 tsleep_interlock(m, 0);
1310 if (atomic_cmpset_int(&m->flags, flags,
1311 flags | PG_WANTED | PG_REFERENCED)) {
1312 tsleep(m, PINTERLOCKED, msg, 0);
1313 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1316 } else if (also_m_busy && (flags & PG_SBUSY)) {
1317 tsleep_interlock(m, 0);
1318 if (atomic_cmpset_int(&m->flags, flags,
1319 flags | PG_WANTED | PG_REFERENCED)) {
1320 tsleep(m, PINTERLOCKED, msg, 0);
1321 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1324 } else if (atomic_cmpset_int(&m->flags, flags,
1326 #ifdef VM_PAGE_DEBUG
1327 m->busy_func = func;
1328 m->busy_line = lineno;
1337 * Attempt to lookup and busy a page.
1339 * Returns NULL if the page could not be found
1341 * Returns a vm_page and error == TRUE if the page exists but could not
1344 * Returns a vm_page and error == FALSE on success.
1347 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1349 int also_m_busy, int *errorp
1355 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1356 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1359 KKASSERT(m->object == object && m->pindex == pindex);
1362 if (flags & PG_BUSY) {
1366 if (also_m_busy && (flags & PG_SBUSY)) {
1370 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1371 #ifdef VM_PAGE_DEBUG
1372 m->busy_func = func;
1373 m->busy_line = lineno;
1382 * Attempt to repurpose the passed-in page. If the passed-in page cannot
1383 * be repurposed it will be released, *must_reenter will be set to 1, and
1384 * this function will fall-through to vm_page_lookup_busy_try().
1386 * The passed-in page must be wired and not busy. The returned page will
1387 * be busied and not wired.
1389 * A different page may be returned. The returned page will be busied and
1392 * NULL can be returned. If so, the required page could not be busied.
1393 * The passed-in page will be unwired.
1396 vm_page_repurpose(struct vm_object *object, vm_pindex_t pindex,
1397 int also_m_busy, int *errorp, vm_page_t m,
1398 int *must_reenter, int *iswired)
1402 * Do not mess with pages in a complex state, such as pages
1403 * which are mapped, as repurposing such pages can be more
1404 * expensive than simply allocatin a new one.
1406 * NOTE: Soft-busying can deadlock against putpages or I/O
1407 * so we only allow hard-busying here.
1409 KKASSERT(also_m_busy == FALSE);
1410 vm_page_busy_wait(m, also_m_busy, "biodep");
1412 if ((m->flags & (PG_UNMANAGED | PG_MAPPED |
1413 PG_FICTITIOUS | PG_SBUSY)) ||
1414 m->busy || m->wire_count != 1 || m->hold_count) {
1415 vm_page_unwire(m, 0);
1417 /* fall through to normal lookup */
1418 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
1419 vm_page_unwire(m, 0);
1420 vm_page_deactivate(m);
1422 /* fall through to normal lookup */
1425 * We can safely repurpose the page. It should
1426 * already be unqueued.
1428 KKASSERT(m->queue == PQ_NONE && m->dirty == 0);
1432 if (vm_page_insert(m, object, pindex)) {
1438 vm_page_unwire(m, 0);
1440 /* fall through to normal lookup */
1445 * Cannot repurpose page, attempt to locate the desired page. May
1450 m = vm_page_lookup_busy_try(object, pindex, also_m_busy, errorp);
1456 * Caller must hold the related vm_object
1459 vm_page_next(vm_page_t m)
1463 next = vm_page_rb_tree_RB_NEXT(m);
1464 if (next && next->pindex != m->pindex + 1)
1472 * Move the given vm_page from its current object to the specified
1473 * target object/offset. The page must be busy and will remain so
1476 * new_object must be held.
1477 * This routine might block. XXX ?
1479 * NOTE: Swap associated with the page must be invalidated by the move. We
1480 * have to do this for several reasons: (1) we aren't freeing the
1481 * page, (2) we are dirtying the page, (3) the VM system is probably
1482 * moving the page from object A to B, and will then later move
1483 * the backing store from A to B and we can't have a conflict.
1485 * NOTE: We *always* dirty the page. It is necessary both for the
1486 * fact that we moved it, and because we may be invalidating
1487 * swap. If the page is on the cache, we have to deactivate it
1488 * or vm_page_dirty() will panic. Dirty pages are not allowed
1492 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1494 KKASSERT(m->flags & PG_BUSY);
1495 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object));
1497 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object));
1500 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1501 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1502 new_object, new_pindex);
1504 if (m->queue - m->pc == PQ_CACHE)
1505 vm_page_deactivate(m);
1510 * vm_page_unqueue() without any wakeup. This routine is used when a page
1511 * is to remain BUSYied by the caller.
1513 * This routine may not block.
1516 vm_page_unqueue_nowakeup(vm_page_t m)
1518 vm_page_and_queue_spin_lock(m);
1519 (void)_vm_page_rem_queue_spinlocked(m);
1520 vm_page_spin_unlock(m);
1524 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1527 * This routine may not block.
1530 vm_page_unqueue(vm_page_t m)
1534 vm_page_and_queue_spin_lock(m);
1535 queue = _vm_page_rem_queue_spinlocked(m);
1536 if (queue == PQ_FREE || queue == PQ_CACHE) {
1537 vm_page_spin_unlock(m);
1538 pagedaemon_wakeup();
1540 vm_page_spin_unlock(m);
1545 * vm_page_list_find()
1547 * Find a page on the specified queue with color optimization.
1549 * The page coloring optimization attempts to locate a page that does
1550 * not overload other nearby pages in the object in the cpu's L1 or L2
1551 * caches. We need this optimization because cpu caches tend to be
1552 * physical caches, while object spaces tend to be virtual.
1554 * The page coloring optimization also, very importantly, tries to localize
1555 * memory to cpus and physical sockets.
1557 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1558 * and the algorithm is adjusted to localize allocations on a per-core basis.
1559 * This is done by 'twisting' the colors.
1561 * The page is returned spinlocked and removed from its queue (it will
1562 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1563 * is responsible for dealing with the busy-page case (usually by
1564 * deactivating the page and looping).
1566 * NOTE: This routine is carefully inlined. A non-inlined version
1567 * is available for outside callers but the only critical path is
1568 * from within this source file.
1570 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1571 * represent stable storage, allowing us to order our locks vm_page
1572 * first, then queue.
1576 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1582 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl,
1585 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1588 m = _vm_page_list_find2(basequeue, index);
1591 vm_page_and_queue_spin_lock(m);
1592 if (m->queue == basequeue + index) {
1593 _vm_page_rem_queue_spinlocked(m);
1594 /* vm_page_t spin held, no queue spin */
1597 vm_page_and_queue_spin_unlock(m);
1603 * If we could not find the page in the desired queue try to find it in
1607 _vm_page_list_find2(int basequeue, int index)
1609 struct vpgqueues *pq;
1611 int pqmask = PQ_SET_ASSOC_MASK >> 1;
1615 index &= PQ_L2_MASK;
1616 pq = &vm_page_queues[basequeue];
1619 * Run local sets of 16, 32, 64, 128, and the whole queue if all
1620 * else fails (PQ_L2_MASK which is 255).
1623 pqmask = (pqmask << 1) | 1;
1624 for (i = 0; i <= pqmask; ++i) {
1625 pqi = (index & ~pqmask) | ((index + i) & pqmask);
1626 m = TAILQ_FIRST(&pq[pqi].pl);
1628 _vm_page_and_queue_spin_lock(m);
1629 if (m->queue == basequeue + pqi) {
1630 _vm_page_rem_queue_spinlocked(m);
1633 _vm_page_and_queue_spin_unlock(m);
1638 } while (pqmask != PQ_L2_MASK);
1644 * Returns a vm_page candidate for allocation. The page is not busied so
1645 * it can move around. The caller must busy the page (and typically
1646 * deactivate it if it cannot be busied!)
1648 * Returns a spinlocked vm_page that has been removed from its queue.
1651 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1653 return(_vm_page_list_find(basequeue, index, prefer_zero));
1657 * Find a page on the cache queue with color optimization, remove it
1658 * from the queue, and busy it. The returned page will not be spinlocked.
1660 * A candidate failure will be deactivated. Candidates can fail due to
1661 * being busied by someone else, in which case they will be deactivated.
1663 * This routine may not block.
1667 vm_page_select_cache(u_short pg_color)
1672 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE);
1676 * (m) has been removed from its queue and spinlocked
1678 if (vm_page_busy_try(m, TRUE)) {
1679 _vm_page_deactivate_locked(m, 0);
1680 vm_page_spin_unlock(m);
1683 * We successfully busied the page
1685 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 &&
1686 m->hold_count == 0 &&
1687 m->wire_count == 0 &&
1688 (m->dirty & m->valid) == 0) {
1689 vm_page_spin_unlock(m);
1690 pagedaemon_wakeup();
1695 * The page cannot be recycled, deactivate it.
1697 _vm_page_deactivate_locked(m, 0);
1698 if (_vm_page_wakeup(m)) {
1699 vm_page_spin_unlock(m);
1702 vm_page_spin_unlock(m);
1710 * Find a free or zero page, with specified preference. We attempt to
1711 * inline the nominal case and fall back to _vm_page_select_free()
1712 * otherwise. A busied page is removed from the queue and returned.
1714 * This routine may not block.
1716 static __inline vm_page_t
1717 vm_page_select_free(u_short pg_color, boolean_t prefer_zero)
1722 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK,
1726 if (vm_page_busy_try(m, TRUE)) {
1728 * Various mechanisms such as a pmap_collect can
1729 * result in a busy page on the free queue. We
1730 * have to move the page out of the way so we can
1731 * retry the allocation. If the other thread is not
1732 * allocating the page then m->valid will remain 0 and
1733 * the pageout daemon will free the page later on.
1735 * Since we could not busy the page, however, we
1736 * cannot make assumptions as to whether the page
1737 * will be allocated by the other thread or not,
1738 * so all we can do is deactivate it to move it out
1739 * of the way. In particular, if the other thread
1740 * wires the page it may wind up on the inactive
1741 * queue and the pageout daemon will have to deal
1742 * with that case too.
1744 _vm_page_deactivate_locked(m, 0);
1745 vm_page_spin_unlock(m);
1748 * Theoretically if we are able to busy the page
1749 * atomic with the queue removal (using the vm_page
1750 * lock) nobody else should be able to mess with the
1753 KKASSERT((m->flags & (PG_UNMANAGED |
1754 PG_NEED_COMMIT)) == 0);
1755 KASSERT(m->hold_count == 0, ("m->hold_count is not zero "
1756 "pg %p q=%d flags=%08x hold=%d wire=%d",
1757 m, m->queue, m->flags, m->hold_count, m->wire_count));
1758 KKASSERT(m->wire_count == 0);
1759 vm_page_spin_unlock(m);
1760 pagedaemon_wakeup();
1762 /* return busied and removed page */
1772 * Allocate and return a memory cell associated with this VM object/offset
1773 * pair. If object is NULL an unassociated page will be allocated.
1775 * The returned page will be busied and removed from its queues. This
1776 * routine can block and may return NULL if a race occurs and the page
1777 * is found to already exist at the specified (object, pindex).
1779 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1780 * VM_ALLOC_QUICK like normal but cannot use cache
1781 * VM_ALLOC_SYSTEM greater free drain
1782 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1783 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1784 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1785 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1786 * (see vm_page_grab())
1787 * VM_ALLOC_USE_GD ok to use per-gd cache
1789 * VM_ALLOC_CPU(n) allocate using specified cpu localization
1791 * The object must be held if not NULL
1792 * This routine may not block
1794 * Additional special handling is required when called from an interrupt
1795 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1799 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1809 * Special per-cpu free VM page cache. The pages are pre-busied
1810 * and pre-zerod for us.
1812 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1814 if (gd->gd_vmpg_count) {
1815 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1827 * CPU localization algorithm. Break the page queues up by physical
1828 * id and core id (note that two cpu threads will have the same core
1829 * id, and core_id != gd_cpuid).
1831 * This is nowhere near perfect, for example the last pindex in a
1832 * subgroup will overflow into the next cpu or package. But this
1833 * should get us good page reuse locality in heavy mixed loads.
1835 * (may be executed before the APs are started, so other GDs might
1838 if (page_req & VM_ALLOC_CPU_SPEC)
1839 cpuid_local = VM_ALLOC_GETCPU(page_req);
1841 cpuid_local = mycpu->gd_cpuid;
1843 pg_color = vm_get_pg_color(cpuid_local, object, pindex);
1846 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1847 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1850 * Certain system threads (pageout daemon, buf_daemon's) are
1851 * allowed to eat deeper into the free page list.
1853 if (curthread->td_flags & TDF_SYSTHREAD)
1854 page_req |= VM_ALLOC_SYSTEM;
1857 * Impose various limitations. Note that the v_free_reserved test
1858 * must match the opposite of vm_page_count_target() to avoid
1859 * livelocks, be careful.
1863 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved ||
1864 ((page_req & VM_ALLOC_INTERRUPT) &&
1865 gd->gd_vmstats.v_free_count > 0) ||
1866 ((page_req & VM_ALLOC_SYSTEM) &&
1867 gd->gd_vmstats.v_cache_count == 0 &&
1868 gd->gd_vmstats.v_free_count >
1869 gd->gd_vmstats.v_interrupt_free_min)
1872 * The free queue has sufficient free pages to take one out.
1874 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO))
1875 m = vm_page_select_free(pg_color, TRUE);
1877 m = vm_page_select_free(pg_color, FALSE);
1878 } else if (page_req & VM_ALLOC_NORMAL) {
1880 * Allocatable from the cache (non-interrupt only). On
1881 * success, we must free the page and try again, thus
1882 * ensuring that vmstats.v_*_free_min counters are replenished.
1885 if (curthread->td_preempted) {
1886 kprintf("vm_page_alloc(): warning, attempt to allocate"
1887 " cache page from preempting interrupt\n");
1890 m = vm_page_select_cache(pg_color);
1893 m = vm_page_select_cache(pg_color);
1896 * On success move the page into the free queue and loop.
1898 * Only do this if we can safely acquire the vm_object lock,
1899 * because this is effectively a random page and the caller
1900 * might be holding the lock shared, we don't want to
1904 KASSERT(m->dirty == 0,
1905 ("Found dirty cache page %p", m));
1906 if ((obj = m->object) != NULL) {
1907 if (vm_object_hold_try(obj)) {
1908 vm_page_protect(m, VM_PROT_NONE);
1910 /* m->object NULL here */
1911 vm_object_drop(obj);
1913 vm_page_deactivate(m);
1917 vm_page_protect(m, VM_PROT_NONE);
1924 * On failure return NULL
1926 atomic_add_int(&vm_pageout_deficit, 1);
1927 pagedaemon_wakeup();
1931 * No pages available, wakeup the pageout daemon and give up.
1933 atomic_add_int(&vm_pageout_deficit, 1);
1934 pagedaemon_wakeup();
1939 * v_free_count can race so loop if we don't find the expected
1948 * Good page found. The page has already been busied for us and
1949 * removed from its queues.
1951 KASSERT(m->dirty == 0,
1952 ("vm_page_alloc: free/cache page %p was dirty", m));
1953 KKASSERT(m->queue == PQ_NONE);
1959 * Initialize the structure, inheriting some flags but clearing
1960 * all the rest. The page has already been busied for us.
1962 vm_page_flag_clear(m, ~(PG_BUSY | PG_SBUSY));
1963 KKASSERT(m->wire_count == 0);
1964 KKASSERT(m->busy == 0);
1969 * Caller must be holding the object lock (asserted by
1970 * vm_page_insert()).
1972 * NOTE: Inserting a page here does not insert it into any pmaps
1973 * (which could cause us to block allocating memory).
1975 * NOTE: If no object an unassociated page is allocated, m->pindex
1976 * can be used by the caller for any purpose.
1979 if (vm_page_insert(m, object, pindex) == FALSE) {
1981 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1982 panic("PAGE RACE %p[%ld]/%p",
1983 object, (long)pindex, m);
1991 * Don't wakeup too often - wakeup the pageout daemon when
1992 * we would be nearly out of memory.
1994 pagedaemon_wakeup();
1997 * A PG_BUSY page is returned.
2003 * Returns number of pages available in our DMA memory reserve
2004 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf)
2007 vm_contig_avail_pages(void)
2012 spin_lock(&vm_contig_spin);
2013 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
2014 spin_unlock(&vm_contig_spin);
2020 * Attempt to allocate contiguous physical memory with the specified
2024 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
2025 unsigned long alignment, unsigned long boundary,
2026 unsigned long size, vm_memattr_t memattr)
2032 alignment >>= PAGE_SHIFT;
2035 boundary >>= PAGE_SHIFT;
2038 size = (size + PAGE_MASK) >> PAGE_SHIFT;
2040 spin_lock(&vm_contig_spin);
2041 blk = alist_alloc(&vm_contig_alist, 0, size);
2042 if (blk == ALIST_BLOCK_NONE) {
2043 spin_unlock(&vm_contig_spin);
2045 kprintf("vm_page_alloc_contig: %ldk nospace\n",
2046 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
2050 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
2051 alist_free(&vm_contig_alist, blk, size);
2052 spin_unlock(&vm_contig_spin);
2054 kprintf("vm_page_alloc_contig: %ldk high "
2056 (size + PAGE_MASK) * (PAGE_SIZE / 1024),
2061 spin_unlock(&vm_contig_spin);
2062 if (vm_contig_verbose) {
2063 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
2064 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT,
2065 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
2068 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
2069 if (memattr != VM_MEMATTR_DEFAULT)
2070 for (i = 0;i < size;i++)
2071 pmap_page_set_memattr(&m[i], memattr);
2076 * Free contiguously allocated pages. The pages will be wired but not busy.
2077 * When freeing to the alist we leave them wired and not busy.
2080 vm_page_free_contig(vm_page_t m, unsigned long size)
2082 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
2083 vm_pindex_t start = pa >> PAGE_SHIFT;
2084 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
2086 if (vm_contig_verbose) {
2087 kprintf("vm_page_free_contig: %016jx/%ldk\n",
2088 (intmax_t)pa, size / 1024);
2090 if (pa < vm_low_phys_reserved) {
2091 KKASSERT(pa + size <= vm_low_phys_reserved);
2092 spin_lock(&vm_contig_spin);
2093 alist_free(&vm_contig_alist, start, pages);
2094 spin_unlock(&vm_contig_spin);
2097 vm_page_busy_wait(m, FALSE, "cpgfr");
2098 vm_page_unwire(m, 0);
2109 * Wait for sufficient free memory for nominal heavy memory use kernel
2112 * WARNING! Be sure never to call this in any vm_pageout code path, which
2113 * will trivially deadlock the system.
2116 vm_wait_nominal(void)
2118 while (vm_page_count_min(0))
2123 * Test if vm_wait_nominal() would block.
2126 vm_test_nominal(void)
2128 if (vm_page_count_min(0))
2134 * Block until free pages are available for allocation, called in various
2135 * places before memory allocations.
2137 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
2138 * more generous then that.
2144 * never wait forever
2148 lwkt_gettoken(&vm_token);
2150 if (curthread == pagethread) {
2152 * The pageout daemon itself needs pages, this is bad.
2154 if (vm_page_count_min(0)) {
2155 vm_pageout_pages_needed = 1;
2156 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
2160 * Wakeup the pageout daemon if necessary and wait.
2162 * Do not wait indefinitely for the target to be reached,
2163 * as load might prevent it from being reached any time soon.
2164 * But wait a little to try to slow down page allocations
2165 * and to give more important threads (the pagedaemon)
2166 * allocation priority.
2168 if (vm_page_count_target()) {
2169 if (vm_pages_needed == 0) {
2170 vm_pages_needed = 1;
2171 wakeup(&vm_pages_needed);
2173 ++vm_pages_waiting; /* SMP race ok */
2174 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
2177 lwkt_reltoken(&vm_token);
2181 * Block until free pages are available for allocation
2183 * Called only from vm_fault so that processes page faulting can be
2187 vm_wait_pfault(void)
2190 * Wakeup the pageout daemon if necessary and wait.
2192 * Do not wait indefinitely for the target to be reached,
2193 * as load might prevent it from being reached any time soon.
2194 * But wait a little to try to slow down page allocations
2195 * and to give more important threads (the pagedaemon)
2196 * allocation priority.
2198 if (vm_page_count_min(0)) {
2199 lwkt_gettoken(&vm_token);
2200 while (vm_page_count_severe()) {
2201 if (vm_page_count_target()) {
2204 if (vm_pages_needed == 0) {
2205 vm_pages_needed = 1;
2206 wakeup(&vm_pages_needed);
2208 ++vm_pages_waiting; /* SMP race ok */
2209 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
2212 * Do not stay stuck in the loop if the system is trying
2213 * to kill the process.
2216 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
2220 lwkt_reltoken(&vm_token);
2225 * Put the specified page on the active list (if appropriate). Ensure
2226 * that act_count is at least ACT_INIT but do not otherwise mess with it.
2228 * The caller should be holding the page busied ? XXX
2229 * This routine may not block.
2232 vm_page_activate(vm_page_t m)
2236 vm_page_spin_lock(m);
2237 if (m->queue - m->pc != PQ_ACTIVE) {
2238 _vm_page_queue_spin_lock(m);
2239 oqueue = _vm_page_rem_queue_spinlocked(m);
2240 /* page is left spinlocked, queue is unlocked */
2242 if (oqueue == PQ_CACHE)
2243 mycpu->gd_cnt.v_reactivated++;
2244 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2245 if (m->act_count < ACT_INIT)
2246 m->act_count = ACT_INIT;
2247 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
2249 _vm_page_and_queue_spin_unlock(m);
2250 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
2251 pagedaemon_wakeup();
2253 if (m->act_count < ACT_INIT)
2254 m->act_count = ACT_INIT;
2255 vm_page_spin_unlock(m);
2260 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2261 * routine is called when a page has been added to the cache or free
2264 * This routine may not block.
2266 static __inline void
2267 vm_page_free_wakeup(void)
2269 globaldata_t gd = mycpu;
2272 * If the pageout daemon itself needs pages, then tell it that
2273 * there are some free.
2275 if (vm_pageout_pages_needed &&
2276 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >=
2277 gd->gd_vmstats.v_pageout_free_min
2279 vm_pageout_pages_needed = 0;
2280 wakeup(&vm_pageout_pages_needed);
2284 * Wakeup processes that are waiting on memory.
2286 * Generally speaking we want to wakeup stuck processes as soon as
2287 * possible. !vm_page_count_min(0) is the absolute minimum point
2288 * where we can do this. Wait a bit longer to reduce degenerate
2289 * re-blocking (vm_page_free_hysteresis). The target check is just
2290 * to make sure the min-check w/hysteresis does not exceed the
2293 if (vm_pages_waiting) {
2294 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2295 !vm_page_count_target()) {
2296 vm_pages_waiting = 0;
2297 wakeup(&vmstats.v_free_count);
2298 ++mycpu->gd_cnt.v_ppwakeups;
2301 if (!vm_page_count_target()) {
2303 * Plenty of pages are free, wakeup everyone.
2305 vm_pages_waiting = 0;
2306 wakeup(&vmstats.v_free_count);
2307 ++mycpu->gd_cnt.v_ppwakeups;
2308 } else if (!vm_page_count_min(0)) {
2310 * Some pages are free, wakeup someone.
2312 int wcount = vm_pages_waiting;
2315 vm_pages_waiting = wcount;
2316 wakeup_one(&vmstats.v_free_count);
2317 ++mycpu->gd_cnt.v_ppwakeups;
2324 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
2325 * it from its VM object.
2327 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
2328 * return (the page will have been freed).
2331 vm_page_free_toq(vm_page_t m)
2333 mycpu->gd_cnt.v_tfree++;
2334 KKASSERT((m->flags & PG_MAPPED) == 0);
2335 KKASSERT(m->flags & PG_BUSY);
2337 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
2338 kprintf("vm_page_free: pindex(%lu), busy(%d), "
2339 "PG_BUSY(%d), hold(%d)\n",
2340 (u_long)m->pindex, m->busy,
2341 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count);
2342 if ((m->queue - m->pc) == PQ_FREE)
2343 panic("vm_page_free: freeing free page");
2345 panic("vm_page_free: freeing busy page");
2349 * Remove from object, spinlock the page and its queues and
2350 * remove from any queue. No queue spinlock will be held
2351 * after this section (because the page was removed from any
2355 vm_page_and_queue_spin_lock(m);
2356 _vm_page_rem_queue_spinlocked(m);
2359 * No further management of fictitious pages occurs beyond object
2360 * and queue removal.
2362 if ((m->flags & PG_FICTITIOUS) != 0) {
2363 vm_page_spin_unlock(m);
2371 if (m->wire_count != 0) {
2372 if (m->wire_count > 1) {
2374 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2375 m->wire_count, (long)m->pindex);
2377 panic("vm_page_free: freeing wired page");
2381 * Clear the UNMANAGED flag when freeing an unmanaged page.
2382 * Clear the NEED_COMMIT flag
2384 if (m->flags & PG_UNMANAGED)
2385 vm_page_flag_clear(m, PG_UNMANAGED);
2386 if (m->flags & PG_NEED_COMMIT)
2387 vm_page_flag_clear(m, PG_NEED_COMMIT);
2389 if (m->hold_count != 0) {
2390 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2392 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
2396 * This sequence allows us to clear PG_BUSY while still holding
2397 * its spin lock, which reduces contention vs allocators. We
2398 * must not leave the queue locked or _vm_page_wakeup() may
2401 _vm_page_queue_spin_unlock(m);
2402 if (_vm_page_wakeup(m)) {
2403 vm_page_spin_unlock(m);
2406 vm_page_spin_unlock(m);
2408 vm_page_free_wakeup();
2412 * vm_page_unmanage()
2414 * Prevent PV management from being done on the page. The page is
2415 * removed from the paging queues as if it were wired, and as a
2416 * consequence of no longer being managed the pageout daemon will not
2417 * touch it (since there is no way to locate the pte mappings for the
2418 * page). madvise() calls that mess with the pmap will also no longer
2419 * operate on the page.
2421 * Beyond that the page is still reasonably 'normal'. Freeing the page
2422 * will clear the flag.
2424 * This routine is used by OBJT_PHYS objects - objects using unswappable
2425 * physical memory as backing store rather then swap-backed memory and
2426 * will eventually be extended to support 4MB unmanaged physical
2429 * Caller must be holding the page busy.
2432 vm_page_unmanage(vm_page_t m)
2434 KKASSERT(m->flags & PG_BUSY);
2435 if ((m->flags & PG_UNMANAGED) == 0) {
2436 if (m->wire_count == 0)
2439 vm_page_flag_set(m, PG_UNMANAGED);
2443 * Mark this page as wired down by yet another map, removing it from
2444 * paging queues as necessary.
2446 * Caller must be holding the page busy.
2449 vm_page_wire(vm_page_t m)
2452 * Only bump the wire statistics if the page is not already wired,
2453 * and only unqueue the page if it is on some queue (if it is unmanaged
2454 * it is already off the queues). Don't do anything with fictitious
2455 * pages because they are always wired.
2457 KKASSERT(m->flags & PG_BUSY);
2458 if ((m->flags & PG_FICTITIOUS) == 0) {
2459 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2460 if ((m->flags & PG_UNMANAGED) == 0)
2462 atomic_add_int(&mycpu->gd_vmstats_adj.v_wire_count, 1);
2464 KASSERT(m->wire_count != 0,
2465 ("vm_page_wire: wire_count overflow m=%p", m));
2470 * Release one wiring of this page, potentially enabling it to be paged again.
2472 * Many pages placed on the inactive queue should actually go
2473 * into the cache, but it is difficult to figure out which. What
2474 * we do instead, if the inactive target is well met, is to put
2475 * clean pages at the head of the inactive queue instead of the tail.
2476 * This will cause them to be moved to the cache more quickly and
2477 * if not actively re-referenced, freed more quickly. If we just
2478 * stick these pages at the end of the inactive queue, heavy filesystem
2479 * meta-data accesses can cause an unnecessary paging load on memory bound
2480 * processes. This optimization causes one-time-use metadata to be
2481 * reused more quickly.
2483 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2484 * the inactive queue. This helps the pageout daemon determine memory
2485 * pressure and act on out-of-memory situations more quickly.
2487 * BUT, if we are in a low-memory situation we have no choice but to
2488 * put clean pages on the cache queue.
2490 * A number of routines use vm_page_unwire() to guarantee that the page
2491 * will go into either the inactive or active queues, and will NEVER
2492 * be placed in the cache - for example, just after dirtying a page.
2493 * dirty pages in the cache are not allowed.
2495 * This routine may not block.
2498 vm_page_unwire(vm_page_t m, int activate)
2500 KKASSERT(m->flags & PG_BUSY);
2501 if (m->flags & PG_FICTITIOUS) {
2503 } else if (m->wire_count <= 0) {
2504 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2506 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2507 atomic_add_int(&mycpu->gd_vmstats_adj.v_wire_count, -1);
2508 if (m->flags & PG_UNMANAGED) {
2510 } else if (activate || (m->flags & PG_NEED_COMMIT)) {
2511 vm_page_spin_lock(m);
2512 _vm_page_add_queue_spinlocked(m,
2513 PQ_ACTIVE + m->pc, 0);
2514 _vm_page_and_queue_spin_unlock(m);
2516 vm_page_spin_lock(m);
2517 vm_page_flag_clear(m, PG_WINATCFLS);
2518 _vm_page_add_queue_spinlocked(m,
2519 PQ_INACTIVE + m->pc, 0);
2520 ++vm_swapcache_inactive_heuristic;
2521 _vm_page_and_queue_spin_unlock(m);
2528 * Move the specified page to the inactive queue. If the page has
2529 * any associated swap, the swap is deallocated.
2531 * Normally athead is 0 resulting in LRU operation. athead is set
2532 * to 1 if we want this page to be 'as if it were placed in the cache',
2533 * except without unmapping it from the process address space.
2535 * vm_page's spinlock must be held on entry and will remain held on return.
2536 * This routine may not block.
2539 _vm_page_deactivate_locked(vm_page_t m, int athead)
2544 * Ignore if already inactive.
2546 if (m->queue - m->pc == PQ_INACTIVE)
2548 _vm_page_queue_spin_lock(m);
2549 oqueue = _vm_page_rem_queue_spinlocked(m);
2551 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2552 if (oqueue == PQ_CACHE)
2553 mycpu->gd_cnt.v_reactivated++;
2554 vm_page_flag_clear(m, PG_WINATCFLS);
2555 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2557 ++vm_swapcache_inactive_heuristic;
2559 /* NOTE: PQ_NONE if condition not taken */
2560 _vm_page_queue_spin_unlock(m);
2561 /* leaves vm_page spinlocked */
2565 * Attempt to deactivate a page.
2570 vm_page_deactivate(vm_page_t m)
2572 vm_page_spin_lock(m);
2573 _vm_page_deactivate_locked(m, 0);
2574 vm_page_spin_unlock(m);
2578 vm_page_deactivate_locked(vm_page_t m)
2580 _vm_page_deactivate_locked(m, 0);
2584 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it.
2586 * This function returns non-zero if it successfully moved the page to
2589 * This function unconditionally unbusies the page on return.
2592 vm_page_try_to_cache(vm_page_t m)
2594 vm_page_spin_lock(m);
2595 if (m->dirty || m->hold_count || m->wire_count ||
2596 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2597 if (_vm_page_wakeup(m)) {
2598 vm_page_spin_unlock(m);
2601 vm_page_spin_unlock(m);
2605 vm_page_spin_unlock(m);
2608 * Page busied by us and no longer spinlocked. Dirty pages cannot
2609 * be moved to the cache.
2611 vm_page_test_dirty(m);
2612 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2621 * Attempt to free the page. If we cannot free it, we do nothing.
2622 * 1 is returned on success, 0 on failure.
2627 vm_page_try_to_free(vm_page_t m)
2629 vm_page_spin_lock(m);
2630 if (vm_page_busy_try(m, TRUE)) {
2631 vm_page_spin_unlock(m);
2636 * The page can be in any state, including already being on the free
2637 * queue. Check to see if it really can be freed.
2639 if (m->dirty || /* can't free if it is dirty */
2640 m->hold_count || /* or held (XXX may be wrong) */
2641 m->wire_count || /* or wired */
2642 (m->flags & (PG_UNMANAGED | /* or unmanaged */
2643 PG_NEED_COMMIT)) || /* or needs a commit */
2644 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */
2645 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */
2646 if (_vm_page_wakeup(m)) {
2647 vm_page_spin_unlock(m);
2650 vm_page_spin_unlock(m);
2654 vm_page_spin_unlock(m);
2657 * We can probably free the page.
2659 * Page busied by us and no longer spinlocked. Dirty pages will
2660 * not be freed by this function. We have to re-test the
2661 * dirty bit after cleaning out the pmaps.
2663 vm_page_test_dirty(m);
2664 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2668 vm_page_protect(m, VM_PROT_NONE);
2669 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2680 * Put the specified page onto the page cache queue (if appropriate).
2682 * The page must be busy, and this routine will release the busy and
2683 * possibly even free the page.
2686 vm_page_cache(vm_page_t m)
2689 * Not suitable for the cache
2691 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2692 m->busy || m->wire_count || m->hold_count) {
2698 * Already in the cache (and thus not mapped)
2700 if ((m->queue - m->pc) == PQ_CACHE) {
2701 KKASSERT((m->flags & PG_MAPPED) == 0);
2707 * Caller is required to test m->dirty, but note that the act of
2708 * removing the page from its maps can cause it to become dirty
2709 * on an SMP system due to another cpu running in usermode.
2712 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2717 * Remove all pmaps and indicate that the page is not
2718 * writeable or mapped. Our vm_page_protect() call may
2719 * have blocked (especially w/ VM_PROT_NONE), so recheck
2722 vm_page_protect(m, VM_PROT_NONE);
2723 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2724 m->busy || m->wire_count || m->hold_count) {
2726 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2727 vm_page_deactivate(m);
2730 _vm_page_and_queue_spin_lock(m);
2731 _vm_page_rem_queue_spinlocked(m);
2732 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2733 _vm_page_queue_spin_unlock(m);
2734 if (_vm_page_wakeup(m)) {
2735 vm_page_spin_unlock(m);
2738 vm_page_spin_unlock(m);
2740 vm_page_free_wakeup();
2745 * vm_page_dontneed()
2747 * Cache, deactivate, or do nothing as appropriate. This routine
2748 * is typically used by madvise() MADV_DONTNEED.
2750 * Generally speaking we want to move the page into the cache so
2751 * it gets reused quickly. However, this can result in a silly syndrome
2752 * due to the page recycling too quickly. Small objects will not be
2753 * fully cached. On the otherhand, if we move the page to the inactive
2754 * queue we wind up with a problem whereby very large objects
2755 * unnecessarily blow away our inactive and cache queues.
2757 * The solution is to move the pages based on a fixed weighting. We
2758 * either leave them alone, deactivate them, or move them to the cache,
2759 * where moving them to the cache has the highest weighting.
2760 * By forcing some pages into other queues we eventually force the
2761 * system to balance the queues, potentially recovering other unrelated
2762 * space from active. The idea is to not force this to happen too
2765 * The page must be busied.
2768 vm_page_dontneed(vm_page_t m)
2770 static int dnweight;
2777 * occassionally leave the page alone
2779 if ((dnw & 0x01F0) == 0 ||
2780 m->queue - m->pc == PQ_INACTIVE ||
2781 m->queue - m->pc == PQ_CACHE
2783 if (m->act_count >= ACT_INIT)
2789 * If vm_page_dontneed() is inactivating a page, it must clear
2790 * the referenced flag; otherwise the pagedaemon will see references
2791 * on the page in the inactive queue and reactivate it. Until the
2792 * page can move to the cache queue, madvise's job is not done.
2794 vm_page_flag_clear(m, PG_REFERENCED);
2795 pmap_clear_reference(m);
2798 vm_page_test_dirty(m);
2800 if (m->dirty || (dnw & 0x0070) == 0) {
2802 * Deactivate the page 3 times out of 32.
2807 * Cache the page 28 times out of every 32. Note that
2808 * the page is deactivated instead of cached, but placed
2809 * at the head of the queue instead of the tail.
2813 vm_page_spin_lock(m);
2814 _vm_page_deactivate_locked(m, head);
2815 vm_page_spin_unlock(m);
2819 * These routines manipulate the 'soft busy' count for a page. A soft busy
2820 * is almost like PG_BUSY except that it allows certain compatible operations
2821 * to occur on the page while it is busy. For example, a page undergoing a
2822 * write can still be mapped read-only.
2824 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2825 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2826 * busy bit is cleared.
2829 vm_page_io_start(vm_page_t m)
2831 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2832 atomic_add_char(&m->busy, 1);
2833 vm_page_flag_set(m, PG_SBUSY);
2837 vm_page_io_finish(vm_page_t m)
2839 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2840 atomic_subtract_char(&m->busy, 1);
2842 vm_page_flag_clear(m, PG_SBUSY);
2846 * Indicate that a clean VM page requires a filesystem commit and cannot
2847 * be reused. Used by tmpfs.
2850 vm_page_need_commit(vm_page_t m)
2852 vm_page_flag_set(m, PG_NEED_COMMIT);
2853 vm_object_set_writeable_dirty(m->object);
2857 vm_page_clear_commit(vm_page_t m)
2859 vm_page_flag_clear(m, PG_NEED_COMMIT);
2863 * Grab a page, blocking if it is busy and allocating a page if necessary.
2864 * A busy page is returned or NULL. The page may or may not be valid and
2865 * might not be on a queue (the caller is responsible for the disposition of
2868 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2869 * page will be zero'd and marked valid.
2871 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2872 * valid even if it already exists.
2874 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2875 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2876 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2878 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2879 * always returned if we had blocked.
2881 * This routine may not be called from an interrupt.
2883 * No other requirements.
2886 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2892 KKASSERT(allocflags &
2893 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2894 vm_object_hold_shared(object);
2896 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2898 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2899 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2904 } else if (m == NULL) {
2906 vm_object_upgrade(object);
2909 if (allocflags & VM_ALLOC_RETRY)
2910 allocflags |= VM_ALLOC_NULL_OK;
2911 m = vm_page_alloc(object, pindex,
2912 allocflags & ~VM_ALLOC_RETRY);
2916 if ((allocflags & VM_ALLOC_RETRY) == 0)
2925 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2927 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2928 * valid even if already valid.
2930 * NOTE! We have removed all of the PG_ZERO optimizations and also
2931 * removed the idle zeroing code. These optimizations actually
2932 * slow things down on modern cpus because the zerod area is
2933 * likely uncached, placing a memory-access burden on the
2934 * accesors taking the fault.
2936 * By always zeroing the page in-line with the fault, no
2937 * dynamic ram reads are needed and the caches are hot, ready
2938 * for userland to access the memory.
2940 if (m->valid == 0) {
2941 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2942 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2943 m->valid = VM_PAGE_BITS_ALL;
2945 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2946 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2947 m->valid = VM_PAGE_BITS_ALL;
2950 vm_object_drop(object);
2955 * Mapping function for valid bits or for dirty bits in
2956 * a page. May not block.
2958 * Inputs are required to range within a page.
2964 vm_page_bits(int base, int size)
2970 base + size <= PAGE_SIZE,
2971 ("vm_page_bits: illegal base/size %d/%d", base, size)
2974 if (size == 0) /* handle degenerate case */
2977 first_bit = base >> DEV_BSHIFT;
2978 last_bit = (base + size - 1) >> DEV_BSHIFT;
2980 return ((2 << last_bit) - (1 << first_bit));
2984 * Sets portions of a page valid and clean. The arguments are expected
2985 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2986 * of any partial chunks touched by the range. The invalid portion of
2987 * such chunks will be zero'd.
2989 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2990 * align base to DEV_BSIZE so as not to mark clean a partially
2991 * truncated device block. Otherwise the dirty page status might be
2994 * This routine may not block.
2996 * (base + size) must be less then or equal to PAGE_SIZE.
2999 _vm_page_zero_valid(vm_page_t m, int base, int size)
3004 if (size == 0) /* handle degenerate case */
3008 * If the base is not DEV_BSIZE aligned and the valid
3009 * bit is clear, we have to zero out a portion of the
3013 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
3014 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
3016 pmap_zero_page_area(
3024 * If the ending offset is not DEV_BSIZE aligned and the
3025 * valid bit is clear, we have to zero out a portion of
3029 endoff = base + size;
3031 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
3032 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
3034 pmap_zero_page_area(
3037 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
3043 * Set valid, clear dirty bits. If validating the entire
3044 * page we can safely clear the pmap modify bit. We also
3045 * use this opportunity to clear the PG_NOSYNC flag. If a process
3046 * takes a write fault on a MAP_NOSYNC memory area the flag will
3049 * We set valid bits inclusive of any overlap, but we can only
3050 * clear dirty bits for DEV_BSIZE chunks that are fully within
3053 * Page must be busied?
3054 * No other requirements.
3057 vm_page_set_valid(vm_page_t m, int base, int size)
3059 _vm_page_zero_valid(m, base, size);
3060 m->valid |= vm_page_bits(base, size);
3065 * Set valid bits and clear dirty bits.
3067 * Page must be busied by caller.
3069 * NOTE: This function does not clear the pmap modified bit.
3070 * Also note that e.g. NFS may use a byte-granular base
3073 * No other requirements.
3076 vm_page_set_validclean(vm_page_t m, int base, int size)
3080 _vm_page_zero_valid(m, base, size);
3081 pagebits = vm_page_bits(base, size);
3082 m->valid |= pagebits;
3083 m->dirty &= ~pagebits;
3084 if (base == 0 && size == PAGE_SIZE) {
3085 /*pmap_clear_modify(m);*/
3086 vm_page_flag_clear(m, PG_NOSYNC);
3091 * Set valid & dirty. Used by buwrite()
3093 * Page must be busied by caller.
3096 vm_page_set_validdirty(vm_page_t m, int base, int size)
3100 pagebits = vm_page_bits(base, size);
3101 m->valid |= pagebits;
3102 m->dirty |= pagebits;
3104 vm_object_set_writeable_dirty(m->object);
3110 * NOTE: This function does not clear the pmap modified bit.
3111 * Also note that e.g. NFS may use a byte-granular base
3114 * Page must be busied?
3115 * No other requirements.
3118 vm_page_clear_dirty(vm_page_t m, int base, int size)
3120 m->dirty &= ~vm_page_bits(base, size);
3121 if (base == 0 && size == PAGE_SIZE) {
3122 /*pmap_clear_modify(m);*/
3123 vm_page_flag_clear(m, PG_NOSYNC);
3128 * Make the page all-dirty.
3130 * Also make sure the related object and vnode reflect the fact that the
3131 * object may now contain a dirty page.
3133 * Page must be busied?
3134 * No other requirements.
3137 vm_page_dirty(vm_page_t m)
3140 int pqtype = m->queue - m->pc;
3142 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
3143 ("vm_page_dirty: page in free/cache queue!"));
3144 if (m->dirty != VM_PAGE_BITS_ALL) {
3145 m->dirty = VM_PAGE_BITS_ALL;
3147 vm_object_set_writeable_dirty(m->object);
3152 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3153 * valid and dirty bits for the effected areas are cleared.
3155 * Page must be busied?
3157 * No other requirements.
3160 vm_page_set_invalid(vm_page_t m, int base, int size)
3164 bits = vm_page_bits(base, size);
3167 m->object->generation++;
3171 * The kernel assumes that the invalid portions of a page contain
3172 * garbage, but such pages can be mapped into memory by user code.
3173 * When this occurs, we must zero out the non-valid portions of the
3174 * page so user code sees what it expects.
3176 * Pages are most often semi-valid when the end of a file is mapped
3177 * into memory and the file's size is not page aligned.
3179 * Page must be busied?
3180 * No other requirements.
3183 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3189 * Scan the valid bits looking for invalid sections that
3190 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
3191 * valid bit may be set ) have already been zerod by
3192 * vm_page_set_validclean().
3194 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3195 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3196 (m->valid & (1 << i))
3199 pmap_zero_page_area(
3202 (i - b) << DEV_BSHIFT
3210 * setvalid is TRUE when we can safely set the zero'd areas
3211 * as being valid. We can do this if there are no cache consistency
3212 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3215 m->valid = VM_PAGE_BITS_ALL;
3219 * Is a (partial) page valid? Note that the case where size == 0
3220 * will return FALSE in the degenerate case where the page is entirely
3221 * invalid, and TRUE otherwise.
3224 * No other requirements.
3227 vm_page_is_valid(vm_page_t m, int base, int size)
3229 int bits = vm_page_bits(base, size);
3231 if (m->valid && ((m->valid & bits) == bits))
3238 * update dirty bits from pmap/mmu. May not block.
3240 * Caller must hold the page busy
3243 vm_page_test_dirty(vm_page_t m)
3245 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
3251 * Register an action, associating it with its vm_page
3254 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
3256 struct vm_page_action_hash *hash;
3259 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
3260 hash = &action_hash[hv];
3262 lockmgr(&hash->lk, LK_EXCLUSIVE);
3263 vm_page_flag_set(action->m, PG_ACTIONLIST);
3264 action->event = event;
3265 LIST_INSERT_HEAD(&hash->list, action, entry);
3266 lockmgr(&hash->lk, LK_RELEASE);
3270 * Unregister an action, disassociating it from its related vm_page
3273 vm_page_unregister_action(vm_page_action_t action)
3275 struct vm_page_action_hash *hash;
3278 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
3279 hash = &action_hash[hv];
3280 lockmgr(&hash->lk, LK_EXCLUSIVE);
3281 if (action->event != VMEVENT_NONE) {
3282 action->event = VMEVENT_NONE;
3283 LIST_REMOVE(action, entry);
3285 if (LIST_EMPTY(&hash->list))
3286 vm_page_flag_clear(action->m, PG_ACTIONLIST);
3288 lockmgr(&hash->lk, LK_RELEASE);
3292 * Issue an event on a VM page. Corresponding action structures are
3293 * removed from the page's list and called.
3295 * If the vm_page has no more pending action events we clear its
3296 * PG_ACTIONLIST flag.
3299 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
3301 struct vm_page_action_hash *hash;
3302 struct vm_page_action *scan;
3303 struct vm_page_action *next;
3307 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
3308 hash = &action_hash[hv];
3311 lockmgr(&hash->lk, LK_EXCLUSIVE);
3312 LIST_FOREACH_MUTABLE(scan, &hash->list, entry, next) {
3314 if (scan->event == event) {
3315 scan->event = VMEVENT_NONE;
3316 LIST_REMOVE(scan, entry);
3317 scan->func(m, scan);
3325 vm_page_flag_clear(m, PG_ACTIONLIST);
3326 lockmgr(&hash->lk, LK_RELEASE);
3329 #include "opt_ddb.h"
3331 #include <sys/kernel.h>
3333 #include <ddb/ddb.h>
3335 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3337 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
3338 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
3339 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
3340 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
3341 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
3342 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
3343 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
3344 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
3345 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
3346 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
3349 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3352 db_printf("PQ_FREE:");
3353 for (i = 0; i < PQ_L2_SIZE; i++) {
3354 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
3358 db_printf("PQ_CACHE:");
3359 for(i = 0; i < PQ_L2_SIZE; i++) {
3360 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
3364 db_printf("PQ_ACTIVE:");
3365 for(i = 0; i < PQ_L2_SIZE; i++) {
3366 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3370 db_printf("PQ_INACTIVE:");
3371 for(i = 0; i < PQ_L2_SIZE; i++) {
3372 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);