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 * SET - Minimum required set associative size, must be a power of 2. We
100 * want this to match or exceed the set-associativeness of the cpu.
102 * GRP - A larger set that allows bleed-over into the domains of other
103 * nearby cpus. Also must be a power of 2. Used by the page zeroing
104 * code to smooth things out a bit.
106 #define PQ_SET_ASSOC 16
107 #define PQ_SET_ASSOC_MASK (PQ_SET_ASSOC - 1)
109 #define PQ_GRP_ASSOC (PQ_SET_ASSOC * 2)
110 #define PQ_GRP_ASSOC_MASK (PQ_GRP_ASSOC - 1)
112 static void vm_page_queue_init(void);
113 static void vm_page_free_wakeup(void);
114 static vm_page_t vm_page_select_cache(u_short pg_color);
115 static vm_page_t _vm_page_list_find2(int basequeue, int index);
116 static void _vm_page_deactivate_locked(vm_page_t m, int athead);
118 MALLOC_DEFINE(M_ACTIONHASH, "acthash", "vmpage action hash");
121 * Array of tailq lists
123 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT];
125 LIST_HEAD(vm_page_action_list, vm_page_action);
128 * Action hash for user umtx support. Contention is governed by both
129 * tsleep/wakeup handling (kern/kern_synch.c) and action_hash[] below.
130 * Because action_hash[] represents active table locks, a modest fixed
131 * value well in excess of MAXCPU works here.
133 * There is also scan overhead depending on the number of threads in
134 * umtx*() calls, so we also size the hash table based on maxproc.
136 struct vm_page_action_hash {
137 struct vm_page_action_list list;
141 #define VMACTION_MINHSIZE 256
143 struct vm_page_action_hash *action_hash;
144 static int vmaction_hsize;
145 static int vmaction_hmask;
147 static volatile int vm_pages_waiting;
148 static struct alist vm_contig_alist;
149 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536];
150 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin");
152 static u_long vm_dma_reserved = 0;
153 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved);
154 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0,
155 "Memory reserved for DMA");
156 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD,
157 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA");
159 static int vm_contig_verbose = 0;
160 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose);
162 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
163 vm_pindex_t, pindex);
166 vm_page_queue_init(void)
170 for (i = 0; i < PQ_L2_SIZE; i++)
171 vm_page_queues[PQ_FREE+i].cnt_offset =
172 offsetof(struct vmstats, v_free_count);
173 for (i = 0; i < PQ_L2_SIZE; i++)
174 vm_page_queues[PQ_CACHE+i].cnt_offset =
175 offsetof(struct vmstats, v_cache_count);
176 for (i = 0; i < PQ_L2_SIZE; i++)
177 vm_page_queues[PQ_INACTIVE+i].cnt_offset =
178 offsetof(struct vmstats, v_inactive_count);
179 for (i = 0; i < PQ_L2_SIZE; i++)
180 vm_page_queues[PQ_ACTIVE+i].cnt_offset =
181 offsetof(struct vmstats, v_active_count);
182 for (i = 0; i < PQ_L2_SIZE; i++)
183 vm_page_queues[PQ_HOLD+i].cnt_offset =
184 offsetof(struct vmstats, v_active_count);
185 /* PQ_NONE has no queue */
187 for (i = 0; i < PQ_COUNT; i++) {
188 TAILQ_INIT(&vm_page_queues[i].pl);
189 spin_init(&vm_page_queues[i].spin, "vm_page_queue_init");
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.
566 * Also setup the action_hash[] table here (which is only used by userland)
569 vm_page_startup_finish(void *dummy __unused)
579 spin_lock(&vm_contig_spin);
581 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
582 if (bfree <= vm_dma_reserved / PAGE_SIZE)
588 * Figure out how much of the initial reserve we have to
589 * free in order to reach our target.
591 bfree -= vm_dma_reserved / PAGE_SIZE;
593 blk += count - bfree;
598 * Calculate the nearest power of 2 <= count.
600 for (xcount = 1; xcount <= count; xcount <<= 1)
603 blk += count - xcount;
607 * Allocate the pages from the alist, then free them to
608 * the normal VM page queues.
610 * Pages allocated from the alist are wired. We have to
611 * busy, unwire, and free them. We must also adjust
612 * vm_low_phys_reserved before freeing any pages to prevent
615 rblk = alist_alloc(&vm_contig_alist, blk, count);
617 kprintf("vm_page_startup_finish: Unable to return "
618 "dma space @0x%08x/%d -> 0x%08x\n",
622 atomic_add_int(&vmstats.v_dma_pages, -count);
623 spin_unlock(&vm_contig_spin);
625 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
626 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m);
628 vm_page_busy_wait(m, FALSE, "cpgfr");
629 vm_page_unwire(m, 0);
634 spin_lock(&vm_contig_spin);
636 spin_unlock(&vm_contig_spin);
639 * Print out how much DMA space drivers have already allocated and
640 * how much is left over.
642 kprintf("DMA space used: %jdk, remaining available: %jdk\n",
643 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) *
645 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024));
648 * Scale the action_hash[] array. Primary contention occurs due
649 * to cpu locks, scaled to ncpus, and scan overhead may be incurred
650 * depending on the number of threads, which we scale to maxproc.
652 * NOTE: Action lock might recurse due to callback, so allow
655 vmaction_hsize = VMACTION_MINHSIZE;
656 if (vmaction_hsize < ncpus * 2)
657 vmaction_hsize = ncpus * 2;
658 if (vmaction_hsize < maxproc / 16)
659 vmaction_hsize = maxproc / 16;
661 while (vmaction_hmask < vmaction_hsize)
662 vmaction_hmask = (vmaction_hmask << 1) | 1;
663 vmaction_hsize = vmaction_hmask + 1;
665 action_hash = kmalloc(sizeof(action_hash[0]) * vmaction_hsize,
669 for (i = 0; i < vmaction_hsize; i++) {
670 LIST_INIT(&action_hash[i].list);
671 lockinit(&action_hash[i].lk, "actlk", 0, LK_CANRECURSE);
674 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY,
675 vm_page_startup_finish, NULL);
679 * Scan comparison function for Red-Black tree scans. An inclusive
680 * (start,end) is expected. Other fields are not used.
683 rb_vm_page_scancmp(struct vm_page *p, void *data)
685 struct rb_vm_page_scan_info *info = data;
687 if (p->pindex < info->start_pindex)
689 if (p->pindex > info->end_pindex)
695 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
697 if (p1->pindex < p2->pindex)
699 if (p1->pindex > p2->pindex)
705 vm_page_init(vm_page_t m)
707 /* do nothing for now. Called from pmap_page_init() */
711 * Each page queue has its own spin lock, which is fairly optimal for
712 * allocating and freeing pages at least.
714 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
715 * queue spinlock via this function. Also note that m->queue cannot change
716 * unless both the page and queue are locked.
720 _vm_page_queue_spin_lock(vm_page_t m)
725 if (queue != PQ_NONE) {
726 spin_lock(&vm_page_queues[queue].spin);
727 KKASSERT(queue == m->queue);
733 _vm_page_queue_spin_unlock(vm_page_t m)
739 if (queue != PQ_NONE)
740 spin_unlock(&vm_page_queues[queue].spin);
745 _vm_page_queues_spin_lock(u_short queue)
748 if (queue != PQ_NONE)
749 spin_lock(&vm_page_queues[queue].spin);
755 _vm_page_queues_spin_unlock(u_short queue)
758 if (queue != PQ_NONE)
759 spin_unlock(&vm_page_queues[queue].spin);
763 vm_page_queue_spin_lock(vm_page_t m)
765 _vm_page_queue_spin_lock(m);
769 vm_page_queues_spin_lock(u_short queue)
771 _vm_page_queues_spin_lock(queue);
775 vm_page_queue_spin_unlock(vm_page_t m)
777 _vm_page_queue_spin_unlock(m);
781 vm_page_queues_spin_unlock(u_short queue)
783 _vm_page_queues_spin_unlock(queue);
787 * This locks the specified vm_page and its queue in the proper order
788 * (page first, then queue). The queue may change so the caller must
793 _vm_page_and_queue_spin_lock(vm_page_t m)
795 vm_page_spin_lock(m);
796 _vm_page_queue_spin_lock(m);
801 _vm_page_and_queue_spin_unlock(vm_page_t m)
803 _vm_page_queues_spin_unlock(m->queue);
804 vm_page_spin_unlock(m);
808 vm_page_and_queue_spin_unlock(vm_page_t m)
810 _vm_page_and_queue_spin_unlock(m);
814 vm_page_and_queue_spin_lock(vm_page_t m)
816 _vm_page_and_queue_spin_lock(m);
820 * Helper function removes vm_page from its current queue.
821 * Returns the base queue the page used to be on.
823 * The vm_page and the queue must be spinlocked.
824 * This function will unlock the queue but leave the page spinlocked.
826 static __inline u_short
827 _vm_page_rem_queue_spinlocked(vm_page_t m)
829 struct vpgqueues *pq;
835 if (queue != PQ_NONE) {
836 pq = &vm_page_queues[queue];
837 TAILQ_REMOVE(&pq->pl, m, pageq);
840 * Adjust our pcpu stats. In order for the nominal low-memory
841 * algorithms to work properly we don't let any pcpu stat get
842 * too negative before we force it to be rolled-up into the
843 * global stats. Otherwise our pageout and vm_wait tests
846 * The idea here is to reduce unnecessary SMP cache
847 * mastership changes in the global vmstats, which can be
848 * particularly bad in multi-socket systems.
850 cnt = (int *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
851 atomic_add_int(cnt, -1);
852 if (*cnt < -VMMETER_SLOP_COUNT) {
853 u_int copy = atomic_swap_int(cnt, 0);
854 cnt = (int *)((char *)&vmstats + pq->cnt_offset);
855 atomic_add_int(cnt, copy);
856 cnt = (int *)((char *)&mycpu->gd_vmstats +
858 atomic_add_int(cnt, copy);
864 vm_page_queues_spin_unlock(oqueue); /* intended */
870 * Helper function places the vm_page on the specified queue. Generally
871 * speaking only PQ_FREE pages are placed at the head, to allow them to
872 * be allocated sooner rather than later on the assumption that they
875 * The vm_page must be spinlocked.
876 * This function will return with both the page and the queue locked.
879 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
881 struct vpgqueues *pq;
884 KKASSERT(m->queue == PQ_NONE);
886 if (queue != PQ_NONE) {
887 vm_page_queues_spin_lock(queue);
888 pq = &vm_page_queues[queue];
892 * Adjust our pcpu stats. If a system entity really needs
893 * to incorporate the count it will call vmstats_rollup()
894 * to roll it all up into the global vmstats strufture.
896 cnt = (int *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
897 atomic_add_int(cnt, 1);
900 * PQ_FREE is always handled LIFO style to try to provide
901 * cache-hot pages to programs.
904 if (queue - m->pc == PQ_FREE) {
905 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
907 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
909 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
911 /* leave the queue spinlocked */
916 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
917 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
918 * did not. Only one sleep call will be made before returning.
920 * This function does NOT busy the page and on return the page is not
921 * guaranteed to be available.
924 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
932 if ((flags & PG_BUSY) == 0 &&
933 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
936 tsleep_interlock(m, 0);
937 if (atomic_cmpset_int(&m->flags, flags,
938 flags | PG_WANTED | PG_REFERENCED)) {
939 tsleep(m, PINTERLOCKED, msg, 0);
946 * This calculates and returns a page color given an optional VM object and
947 * either a pindex or an iterator. We attempt to return a cpu-localized
948 * pg_color that is still roughly 16-way set-associative. The CPU topology
949 * is used if it was probed.
951 * The caller may use the returned value to index into e.g. PQ_FREE when
952 * allocating a page in order to nominally obtain pages that are hopefully
953 * already localized to the requesting cpu. This function is not able to
954 * provide any sort of guarantee of this, but does its best to improve
955 * hardware cache management performance.
957 * WARNING! The caller must mask the returned value with PQ_L2_MASK.
960 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex)
967 phys_id = get_cpu_phys_id(cpuid);
968 core_id = get_cpu_core_id(cpuid);
969 object_pg_color = object ? object->pg_color : 0;
971 if (cpu_topology_phys_ids && cpu_topology_core_ids) {
975 * Break us down by socket and cpu
977 pg_color = phys_id * PQ_L2_SIZE / cpu_topology_phys_ids;
978 pg_color += core_id * PQ_L2_SIZE /
979 (cpu_topology_core_ids * cpu_topology_phys_ids);
982 * Calculate remaining component for object/queue color
984 grpsize = PQ_L2_SIZE / (cpu_topology_core_ids *
985 cpu_topology_phys_ids);
987 pg_color += (pindex + object_pg_color) % grpsize;
992 /* 3->9, 4->8, 5->10, 6->12, 7->14 */
997 pg_color += (pindex + object_pg_color) % grpsize;
1001 * Unknown topology, distribute things evenly.
1003 pg_color = cpuid * PQ_L2_SIZE / ncpus;
1004 pg_color += pindex + object_pg_color;
1006 return (pg_color & PQ_L2_MASK);
1010 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
1011 * also wait for m->busy to become 0 before setting PG_BUSY.
1014 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
1015 int also_m_busy, const char *msg
1023 if (flags & PG_BUSY) {
1024 tsleep_interlock(m, 0);
1025 if (atomic_cmpset_int(&m->flags, flags,
1026 flags | PG_WANTED | PG_REFERENCED)) {
1027 tsleep(m, PINTERLOCKED, msg, 0);
1029 } else if (also_m_busy && (flags & PG_SBUSY)) {
1030 tsleep_interlock(m, 0);
1031 if (atomic_cmpset_int(&m->flags, flags,
1032 flags | PG_WANTED | PG_REFERENCED)) {
1033 tsleep(m, PINTERLOCKED, msg, 0);
1036 if (atomic_cmpset_int(&m->flags, flags,
1038 #ifdef VM_PAGE_DEBUG
1039 m->busy_func = func;
1040 m->busy_line = lineno;
1049 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
1052 * Returns non-zero on failure.
1055 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
1063 if (flags & PG_BUSY)
1065 if (also_m_busy && (flags & PG_SBUSY))
1067 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1068 #ifdef VM_PAGE_DEBUG
1069 m->busy_func = func;
1070 m->busy_line = lineno;
1078 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
1079 * that a wakeup() should be performed.
1081 * The vm_page must be spinlocked and will remain spinlocked on return.
1082 * The related queue must NOT be spinlocked (which could deadlock us).
1088 _vm_page_wakeup(vm_page_t m)
1095 if (atomic_cmpset_int(&m->flags, flags,
1096 flags & ~(PG_BUSY | PG_WANTED))) {
1100 return(flags & PG_WANTED);
1104 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
1105 * is typically the last call you make on a page before moving onto
1109 vm_page_wakeup(vm_page_t m)
1111 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
1112 vm_page_spin_lock(m);
1113 if (_vm_page_wakeup(m)) {
1114 vm_page_spin_unlock(m);
1117 vm_page_spin_unlock(m);
1122 * Holding a page keeps it from being reused. Other parts of the system
1123 * can still disassociate the page from its current object and free it, or
1124 * perform read or write I/O on it and/or otherwise manipulate the page,
1125 * but if the page is held the VM system will leave the page and its data
1126 * intact and not reuse the page for other purposes until the last hold
1127 * reference is released. (see vm_page_wire() if you want to prevent the
1128 * page from being disassociated from its object too).
1130 * The caller must still validate the contents of the page and, if necessary,
1131 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
1132 * before manipulating the page.
1134 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
1137 vm_page_hold(vm_page_t m)
1139 vm_page_spin_lock(m);
1140 atomic_add_int(&m->hold_count, 1);
1141 if (m->queue - m->pc == PQ_FREE) {
1142 _vm_page_queue_spin_lock(m);
1143 _vm_page_rem_queue_spinlocked(m);
1144 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
1145 _vm_page_queue_spin_unlock(m);
1147 vm_page_spin_unlock(m);
1151 * The opposite of vm_page_hold(). If the page is on the HOLD queue
1152 * it was freed while held and must be moved back to the FREE queue.
1155 vm_page_unhold(vm_page_t m)
1157 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE,
1158 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)",
1159 m, m->hold_count, m->queue - m->pc));
1160 vm_page_spin_lock(m);
1161 atomic_add_int(&m->hold_count, -1);
1162 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
1163 _vm_page_queue_spin_lock(m);
1164 _vm_page_rem_queue_spinlocked(m);
1165 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1);
1166 _vm_page_queue_spin_unlock(m);
1168 vm_page_spin_unlock(m);
1174 * Create a fictitious page with the specified physical address and
1175 * memory attribute. The memory attribute is the only the machine-
1176 * dependent aspect of a fictitious page that must be initialized.
1180 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1183 if ((m->flags & PG_FICTITIOUS) != 0) {
1185 * The page's memattr might have changed since the
1186 * previous initialization. Update the pmap to the
1191 m->phys_addr = paddr;
1193 /* Fictitious pages don't use "segind". */
1194 /* Fictitious pages don't use "order" or "pool". */
1195 m->flags = PG_FICTITIOUS | PG_UNMANAGED | PG_BUSY;
1197 spin_init(&m->spin, "fake_page");
1200 pmap_page_set_memattr(m, memattr);
1204 * Inserts the given vm_page into the object and object list.
1206 * The pagetables are not updated but will presumably fault the page
1207 * in if necessary, or if a kernel page the caller will at some point
1208 * enter the page into the kernel's pmap. We are not allowed to block
1209 * here so we *can't* do this anyway.
1211 * This routine may not block.
1212 * This routine must be called with the vm_object held.
1213 * This routine must be called with a critical section held.
1215 * This routine returns TRUE if the page was inserted into the object
1216 * successfully, and FALSE if the page already exists in the object.
1219 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1221 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object));
1222 if (m->object != NULL)
1223 panic("vm_page_insert: already inserted");
1225 atomic_add_int(&object->generation, 1);
1228 * Record the object/offset pair in this page and add the
1229 * pv_list_count of the page to the object.
1231 * The vm_page spin lock is required for interactions with the pmap.
1233 vm_page_spin_lock(m);
1236 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
1239 vm_page_spin_unlock(m);
1242 ++object->resident_page_count;
1243 ++mycpu->gd_vmtotal.t_rm;
1244 vm_page_spin_unlock(m);
1247 * Since we are inserting a new and possibly dirty page,
1248 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
1250 if ((m->valid & m->dirty) ||
1251 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT)))
1252 vm_object_set_writeable_dirty(object);
1255 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
1257 swap_pager_page_inserted(m);
1262 * Removes the given vm_page_t from the (object,index) table
1264 * The underlying pmap entry (if any) is NOT removed here.
1265 * This routine may not block.
1267 * The page must be BUSY and will remain BUSY on return.
1268 * No other requirements.
1270 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
1274 vm_page_remove(vm_page_t m)
1278 if (m->object == NULL) {
1282 if ((m->flags & PG_BUSY) == 0)
1283 panic("vm_page_remove: page not busy");
1287 vm_object_hold(object);
1290 * Remove the page from the object and update the object.
1292 * The vm_page spin lock is required for interactions with the pmap.
1294 vm_page_spin_lock(m);
1295 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
1296 --object->resident_page_count;
1297 --mycpu->gd_vmtotal.t_rm;
1299 atomic_add_int(&object->generation, 1);
1300 vm_page_spin_unlock(m);
1302 vm_object_drop(object);
1306 * Locate and return the page at (object, pindex), or NULL if the
1307 * page could not be found.
1309 * The caller must hold the vm_object token.
1312 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1317 * Search the hash table for this object/offset pair
1319 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1320 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1321 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1326 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1328 int also_m_busy, const char *msg
1334 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1335 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1337 KKASSERT(m->object == object && m->pindex == pindex);
1340 if (flags & PG_BUSY) {
1341 tsleep_interlock(m, 0);
1342 if (atomic_cmpset_int(&m->flags, flags,
1343 flags | PG_WANTED | PG_REFERENCED)) {
1344 tsleep(m, PINTERLOCKED, msg, 0);
1345 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1348 } else if (also_m_busy && (flags & PG_SBUSY)) {
1349 tsleep_interlock(m, 0);
1350 if (atomic_cmpset_int(&m->flags, flags,
1351 flags | PG_WANTED | PG_REFERENCED)) {
1352 tsleep(m, PINTERLOCKED, msg, 0);
1353 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1356 } else if (atomic_cmpset_int(&m->flags, flags,
1358 #ifdef VM_PAGE_DEBUG
1359 m->busy_func = func;
1360 m->busy_line = lineno;
1369 * Attempt to lookup and busy a page.
1371 * Returns NULL if the page could not be found
1373 * Returns a vm_page and error == TRUE if the page exists but could not
1376 * Returns a vm_page and error == FALSE on success.
1379 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1381 int also_m_busy, int *errorp
1387 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1388 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1391 KKASSERT(m->object == object && m->pindex == pindex);
1394 if (flags & PG_BUSY) {
1398 if (also_m_busy && (flags & PG_SBUSY)) {
1402 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1403 #ifdef VM_PAGE_DEBUG
1404 m->busy_func = func;
1405 m->busy_line = lineno;
1414 * Attempt to repurpose the passed-in page. If the passed-in page cannot
1415 * be repurposed it will be released, *must_reenter will be set to 1, and
1416 * this function will fall-through to vm_page_lookup_busy_try().
1418 * The passed-in page must be wired and not busy. The returned page will
1419 * be busied and not wired.
1421 * A different page may be returned. The returned page will be busied and
1424 * NULL can be returned. If so, the required page could not be busied.
1425 * The passed-in page will be unwired.
1428 vm_page_repurpose(struct vm_object *object, vm_pindex_t pindex,
1429 int also_m_busy, int *errorp, vm_page_t m,
1430 int *must_reenter, int *iswired)
1434 * Do not mess with pages in a complex state, such as pages
1435 * which are mapped, as repurposing such pages can be more
1436 * expensive than simply allocatin a new one.
1438 * NOTE: Soft-busying can deadlock against putpages or I/O
1439 * so we only allow hard-busying here.
1441 KKASSERT(also_m_busy == FALSE);
1442 vm_page_busy_wait(m, also_m_busy, "biodep");
1444 if ((m->flags & (PG_UNMANAGED | PG_MAPPED |
1445 PG_FICTITIOUS | PG_SBUSY)) ||
1446 m->busy || m->wire_count != 1 || m->hold_count) {
1447 vm_page_unwire(m, 0);
1449 /* fall through to normal lookup */
1450 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
1451 vm_page_unwire(m, 0);
1452 vm_page_deactivate(m);
1454 /* fall through to normal lookup */
1457 * We can safely repurpose the page. It should
1458 * already be unqueued.
1460 KKASSERT(m->queue == PQ_NONE && m->dirty == 0);
1464 if (vm_page_insert(m, object, pindex)) {
1470 vm_page_unwire(m, 0);
1472 /* fall through to normal lookup */
1477 * Cannot repurpose page, attempt to locate the desired page. May
1482 m = vm_page_lookup_busy_try(object, pindex, also_m_busy, errorp);
1488 * Caller must hold the related vm_object
1491 vm_page_next(vm_page_t m)
1495 next = vm_page_rb_tree_RB_NEXT(m);
1496 if (next && next->pindex != m->pindex + 1)
1504 * Move the given vm_page from its current object to the specified
1505 * target object/offset. The page must be busy and will remain so
1508 * new_object must be held.
1509 * This routine might block. XXX ?
1511 * NOTE: Swap associated with the page must be invalidated by the move. We
1512 * have to do this for several reasons: (1) we aren't freeing the
1513 * page, (2) we are dirtying the page, (3) the VM system is probably
1514 * moving the page from object A to B, and will then later move
1515 * the backing store from A to B and we can't have a conflict.
1517 * NOTE: We *always* dirty the page. It is necessary both for the
1518 * fact that we moved it, and because we may be invalidating
1519 * swap. If the page is on the cache, we have to deactivate it
1520 * or vm_page_dirty() will panic. Dirty pages are not allowed
1524 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1526 KKASSERT(m->flags & PG_BUSY);
1527 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object));
1529 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object));
1532 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1533 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1534 new_object, new_pindex);
1536 if (m->queue - m->pc == PQ_CACHE)
1537 vm_page_deactivate(m);
1542 * vm_page_unqueue() without any wakeup. This routine is used when a page
1543 * is to remain BUSYied by the caller.
1545 * This routine may not block.
1548 vm_page_unqueue_nowakeup(vm_page_t m)
1550 vm_page_and_queue_spin_lock(m);
1551 (void)_vm_page_rem_queue_spinlocked(m);
1552 vm_page_spin_unlock(m);
1556 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1559 * This routine may not block.
1562 vm_page_unqueue(vm_page_t m)
1566 vm_page_and_queue_spin_lock(m);
1567 queue = _vm_page_rem_queue_spinlocked(m);
1568 if (queue == PQ_FREE || queue == PQ_CACHE) {
1569 vm_page_spin_unlock(m);
1570 pagedaemon_wakeup();
1572 vm_page_spin_unlock(m);
1577 * vm_page_list_find()
1579 * Find a page on the specified queue with color optimization.
1581 * The page coloring optimization attempts to locate a page that does
1582 * not overload other nearby pages in the object in the cpu's L1 or L2
1583 * caches. We need this optimization because cpu caches tend to be
1584 * physical caches, while object spaces tend to be virtual.
1586 * The page coloring optimization also, very importantly, tries to localize
1587 * memory to cpus and physical sockets.
1589 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1590 * and the algorithm is adjusted to localize allocations on a per-core basis.
1591 * This is done by 'twisting' the colors.
1593 * The page is returned spinlocked and removed from its queue (it will
1594 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1595 * is responsible for dealing with the busy-page case (usually by
1596 * deactivating the page and looping).
1598 * NOTE: This routine is carefully inlined. A non-inlined version
1599 * is available for outside callers but the only critical path is
1600 * from within this source file.
1602 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1603 * represent stable storage, allowing us to order our locks vm_page
1604 * first, then queue.
1608 _vm_page_list_find(int basequeue, int index)
1613 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1615 m = _vm_page_list_find2(basequeue, index);
1618 vm_page_and_queue_spin_lock(m);
1619 if (m->queue == basequeue + index) {
1620 _vm_page_rem_queue_spinlocked(m);
1621 /* vm_page_t spin held, no queue spin */
1624 vm_page_and_queue_spin_unlock(m);
1630 * If we could not find the page in the desired queue try to find it in
1634 _vm_page_list_find2(int basequeue, int index)
1636 struct vpgqueues *pq;
1638 int pqmask = PQ_SET_ASSOC_MASK >> 1;
1642 index &= PQ_L2_MASK;
1643 pq = &vm_page_queues[basequeue];
1646 * Run local sets of 16, 32, 64, 128, and the whole queue if all
1647 * else fails (PQ_L2_MASK which is 255).
1650 pqmask = (pqmask << 1) | 1;
1651 for (i = 0; i <= pqmask; ++i) {
1652 pqi = (index & ~pqmask) | ((index + i) & pqmask);
1653 m = TAILQ_FIRST(&pq[pqi].pl);
1655 _vm_page_and_queue_spin_lock(m);
1656 if (m->queue == basequeue + pqi) {
1657 _vm_page_rem_queue_spinlocked(m);
1660 _vm_page_and_queue_spin_unlock(m);
1665 } while (pqmask != PQ_L2_MASK);
1671 * Returns a vm_page candidate for allocation. The page is not busied so
1672 * it can move around. The caller must busy the page (and typically
1673 * deactivate it if it cannot be busied!)
1675 * Returns a spinlocked vm_page that has been removed from its queue.
1678 vm_page_list_find(int basequeue, int index)
1680 return(_vm_page_list_find(basequeue, index));
1684 * Find a page on the cache queue with color optimization, remove it
1685 * from the queue, and busy it. The returned page will not be spinlocked.
1687 * A candidate failure will be deactivated. Candidates can fail due to
1688 * being busied by someone else, in which case they will be deactivated.
1690 * This routine may not block.
1694 vm_page_select_cache(u_short pg_color)
1699 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK);
1703 * (m) has been removed from its queue and spinlocked
1705 if (vm_page_busy_try(m, TRUE)) {
1706 _vm_page_deactivate_locked(m, 0);
1707 vm_page_spin_unlock(m);
1710 * We successfully busied the page
1712 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 &&
1713 m->hold_count == 0 &&
1714 m->wire_count == 0 &&
1715 (m->dirty & m->valid) == 0) {
1716 vm_page_spin_unlock(m);
1717 pagedaemon_wakeup();
1722 * The page cannot be recycled, deactivate it.
1724 _vm_page_deactivate_locked(m, 0);
1725 if (_vm_page_wakeup(m)) {
1726 vm_page_spin_unlock(m);
1729 vm_page_spin_unlock(m);
1737 * Find a free page. We attempt to inline the nominal case and fall back
1738 * to _vm_page_select_free() otherwise. A busied page is removed from
1739 * the queue and returned.
1741 * This routine may not block.
1743 static __inline vm_page_t
1744 vm_page_select_free(u_short pg_color)
1749 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK);
1752 if (vm_page_busy_try(m, TRUE)) {
1754 * Various mechanisms such as a pmap_collect can
1755 * result in a busy page on the free queue. We
1756 * have to move the page out of the way so we can
1757 * retry the allocation. If the other thread is not
1758 * allocating the page then m->valid will remain 0 and
1759 * the pageout daemon will free the page later on.
1761 * Since we could not busy the page, however, we
1762 * cannot make assumptions as to whether the page
1763 * will be allocated by the other thread or not,
1764 * so all we can do is deactivate it to move it out
1765 * of the way. In particular, if the other thread
1766 * wires the page it may wind up on the inactive
1767 * queue and the pageout daemon will have to deal
1768 * with that case too.
1770 _vm_page_deactivate_locked(m, 0);
1771 vm_page_spin_unlock(m);
1774 * Theoretically if we are able to busy the page
1775 * atomic with the queue removal (using the vm_page
1776 * lock) nobody else should be able to mess with the
1779 KKASSERT((m->flags & (PG_UNMANAGED |
1780 PG_NEED_COMMIT)) == 0);
1781 KASSERT(m->hold_count == 0, ("m->hold_count is not zero "
1782 "pg %p q=%d flags=%08x hold=%d wire=%d",
1783 m, m->queue, m->flags, m->hold_count, m->wire_count));
1784 KKASSERT(m->wire_count == 0);
1785 vm_page_spin_unlock(m);
1786 pagedaemon_wakeup();
1788 /* return busied and removed page */
1798 * Allocate and return a memory cell associated with this VM object/offset
1799 * pair. If object is NULL an unassociated page will be allocated.
1801 * The returned page will be busied and removed from its queues. This
1802 * routine can block and may return NULL if a race occurs and the page
1803 * is found to already exist at the specified (object, pindex).
1805 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1806 * VM_ALLOC_QUICK like normal but cannot use cache
1807 * VM_ALLOC_SYSTEM greater free drain
1808 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1809 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1810 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1811 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1812 * (see vm_page_grab())
1813 * VM_ALLOC_USE_GD ok to use per-gd cache
1815 * VM_ALLOC_CPU(n) allocate using specified cpu localization
1817 * The object must be held if not NULL
1818 * This routine may not block
1820 * Additional special handling is required when called from an interrupt
1821 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1825 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1835 * Special per-cpu free VM page cache. The pages are pre-busied
1836 * and pre-zerod for us.
1838 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1840 if (gd->gd_vmpg_count) {
1841 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1853 * CPU localization algorithm. Break the page queues up by physical
1854 * id and core id (note that two cpu threads will have the same core
1855 * id, and core_id != gd_cpuid).
1857 * This is nowhere near perfect, for example the last pindex in a
1858 * subgroup will overflow into the next cpu or package. But this
1859 * should get us good page reuse locality in heavy mixed loads.
1861 * (may be executed before the APs are started, so other GDs might
1864 if (page_req & VM_ALLOC_CPU_SPEC)
1865 cpuid_local = VM_ALLOC_GETCPU(page_req);
1867 cpuid_local = mycpu->gd_cpuid;
1869 pg_color = vm_get_pg_color(cpuid_local, object, pindex);
1872 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1873 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1876 * Certain system threads (pageout daemon, buf_daemon's) are
1877 * allowed to eat deeper into the free page list.
1879 if (curthread->td_flags & TDF_SYSTHREAD)
1880 page_req |= VM_ALLOC_SYSTEM;
1883 * Impose various limitations. Note that the v_free_reserved test
1884 * must match the opposite of vm_page_count_target() to avoid
1885 * livelocks, be careful.
1889 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved ||
1890 ((page_req & VM_ALLOC_INTERRUPT) &&
1891 gd->gd_vmstats.v_free_count > 0) ||
1892 ((page_req & VM_ALLOC_SYSTEM) &&
1893 gd->gd_vmstats.v_cache_count == 0 &&
1894 gd->gd_vmstats.v_free_count >
1895 gd->gd_vmstats.v_interrupt_free_min)
1898 * The free queue has sufficient free pages to take one out.
1900 m = vm_page_select_free(pg_color);
1901 } else if (page_req & VM_ALLOC_NORMAL) {
1903 * Allocatable from the cache (non-interrupt only). On
1904 * success, we must free the page and try again, thus
1905 * ensuring that vmstats.v_*_free_min counters are replenished.
1908 if (curthread->td_preempted) {
1909 kprintf("vm_page_alloc(): warning, attempt to allocate"
1910 " cache page from preempting interrupt\n");
1913 m = vm_page_select_cache(pg_color);
1916 m = vm_page_select_cache(pg_color);
1919 * On success move the page into the free queue and loop.
1921 * Only do this if we can safely acquire the vm_object lock,
1922 * because this is effectively a random page and the caller
1923 * might be holding the lock shared, we don't want to
1927 KASSERT(m->dirty == 0,
1928 ("Found dirty cache page %p", m));
1929 if ((obj = m->object) != NULL) {
1930 if (vm_object_hold_try(obj)) {
1931 vm_page_protect(m, VM_PROT_NONE);
1933 /* m->object NULL here */
1934 vm_object_drop(obj);
1936 vm_page_deactivate(m);
1940 vm_page_protect(m, VM_PROT_NONE);
1947 * On failure return NULL
1949 atomic_add_int(&vm_pageout_deficit, 1);
1950 pagedaemon_wakeup();
1954 * No pages available, wakeup the pageout daemon and give up.
1956 atomic_add_int(&vm_pageout_deficit, 1);
1957 pagedaemon_wakeup();
1962 * v_free_count can race so loop if we don't find the expected
1971 * Good page found. The page has already been busied for us and
1972 * removed from its queues.
1974 KASSERT(m->dirty == 0,
1975 ("vm_page_alloc: free/cache page %p was dirty", m));
1976 KKASSERT(m->queue == PQ_NONE);
1982 * Initialize the structure, inheriting some flags but clearing
1983 * all the rest. The page has already been busied for us.
1985 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK);
1987 KKASSERT(m->wire_count == 0);
1988 KKASSERT(m->busy == 0);
1993 * Caller must be holding the object lock (asserted by
1994 * vm_page_insert()).
1996 * NOTE: Inserting a page here does not insert it into any pmaps
1997 * (which could cause us to block allocating memory).
1999 * NOTE: If no object an unassociated page is allocated, m->pindex
2000 * can be used by the caller for any purpose.
2003 if (vm_page_insert(m, object, pindex) == FALSE) {
2005 if ((page_req & VM_ALLOC_NULL_OK) == 0)
2006 panic("PAGE RACE %p[%ld]/%p",
2007 object, (long)pindex, m);
2015 * Don't wakeup too often - wakeup the pageout daemon when
2016 * we would be nearly out of memory.
2018 pagedaemon_wakeup();
2021 * A PG_BUSY page is returned.
2027 * Returns number of pages available in our DMA memory reserve
2028 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf)
2031 vm_contig_avail_pages(void)
2036 spin_lock(&vm_contig_spin);
2037 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
2038 spin_unlock(&vm_contig_spin);
2044 * Attempt to allocate contiguous physical memory with the specified
2048 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
2049 unsigned long alignment, unsigned long boundary,
2050 unsigned long size, vm_memattr_t memattr)
2056 alignment >>= PAGE_SHIFT;
2059 boundary >>= PAGE_SHIFT;
2062 size = (size + PAGE_MASK) >> PAGE_SHIFT;
2064 spin_lock(&vm_contig_spin);
2065 blk = alist_alloc(&vm_contig_alist, 0, size);
2066 if (blk == ALIST_BLOCK_NONE) {
2067 spin_unlock(&vm_contig_spin);
2069 kprintf("vm_page_alloc_contig: %ldk nospace\n",
2070 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
2074 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
2075 alist_free(&vm_contig_alist, blk, size);
2076 spin_unlock(&vm_contig_spin);
2078 kprintf("vm_page_alloc_contig: %ldk high "
2080 (size + PAGE_MASK) * (PAGE_SIZE / 1024),
2085 spin_unlock(&vm_contig_spin);
2086 if (vm_contig_verbose) {
2087 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
2088 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT,
2089 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
2092 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
2093 if (memattr != VM_MEMATTR_DEFAULT)
2094 for (i = 0;i < size;i++)
2095 pmap_page_set_memattr(&m[i], memattr);
2100 * Free contiguously allocated pages. The pages will be wired but not busy.
2101 * When freeing to the alist we leave them wired and not busy.
2104 vm_page_free_contig(vm_page_t m, unsigned long size)
2106 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
2107 vm_pindex_t start = pa >> PAGE_SHIFT;
2108 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
2110 if (vm_contig_verbose) {
2111 kprintf("vm_page_free_contig: %016jx/%ldk\n",
2112 (intmax_t)pa, size / 1024);
2114 if (pa < vm_low_phys_reserved) {
2115 KKASSERT(pa + size <= vm_low_phys_reserved);
2116 spin_lock(&vm_contig_spin);
2117 alist_free(&vm_contig_alist, start, pages);
2118 spin_unlock(&vm_contig_spin);
2121 vm_page_busy_wait(m, FALSE, "cpgfr");
2122 vm_page_unwire(m, 0);
2133 * Wait for sufficient free memory for nominal heavy memory use kernel
2136 * WARNING! Be sure never to call this in any vm_pageout code path, which
2137 * will trivially deadlock the system.
2140 vm_wait_nominal(void)
2142 while (vm_page_count_min(0))
2147 * Test if vm_wait_nominal() would block.
2150 vm_test_nominal(void)
2152 if (vm_page_count_min(0))
2158 * Block until free pages are available for allocation, called in various
2159 * places before memory allocations.
2161 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
2162 * more generous then that.
2168 * never wait forever
2172 lwkt_gettoken(&vm_token);
2174 if (curthread == pagethread) {
2176 * The pageout daemon itself needs pages, this is bad.
2178 if (vm_page_count_min(0)) {
2179 vm_pageout_pages_needed = 1;
2180 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
2184 * Wakeup the pageout daemon if necessary and wait.
2186 * Do not wait indefinitely for the target to be reached,
2187 * as load might prevent it from being reached any time soon.
2188 * But wait a little to try to slow down page allocations
2189 * and to give more important threads (the pagedaemon)
2190 * allocation priority.
2192 if (vm_page_count_target()) {
2193 if (vm_pages_needed == 0) {
2194 vm_pages_needed = 1;
2195 wakeup(&vm_pages_needed);
2197 ++vm_pages_waiting; /* SMP race ok */
2198 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
2201 lwkt_reltoken(&vm_token);
2205 * Block until free pages are available for allocation
2207 * Called only from vm_fault so that processes page faulting can be
2211 vm_wait_pfault(void)
2214 * Wakeup the pageout daemon if necessary and wait.
2216 * Do not wait indefinitely for the target to be reached,
2217 * as load might prevent it from being reached any time soon.
2218 * But wait a little to try to slow down page allocations
2219 * and to give more important threads (the pagedaemon)
2220 * allocation priority.
2222 if (vm_page_count_min(0)) {
2223 lwkt_gettoken(&vm_token);
2224 while (vm_page_count_severe()) {
2225 if (vm_page_count_target()) {
2228 if (vm_pages_needed == 0) {
2229 vm_pages_needed = 1;
2230 wakeup(&vm_pages_needed);
2232 ++vm_pages_waiting; /* SMP race ok */
2233 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
2236 * Do not stay stuck in the loop if the system is trying
2237 * to kill the process.
2240 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
2244 lwkt_reltoken(&vm_token);
2249 * Put the specified page on the active list (if appropriate). Ensure
2250 * that act_count is at least ACT_INIT but do not otherwise mess with it.
2252 * The caller should be holding the page busied ? XXX
2253 * This routine may not block.
2256 vm_page_activate(vm_page_t m)
2260 vm_page_spin_lock(m);
2261 if (m->queue - m->pc != PQ_ACTIVE) {
2262 _vm_page_queue_spin_lock(m);
2263 oqueue = _vm_page_rem_queue_spinlocked(m);
2264 /* page is left spinlocked, queue is unlocked */
2266 if (oqueue == PQ_CACHE)
2267 mycpu->gd_cnt.v_reactivated++;
2268 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2269 if (m->act_count < ACT_INIT)
2270 m->act_count = ACT_INIT;
2271 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
2273 _vm_page_and_queue_spin_unlock(m);
2274 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
2275 pagedaemon_wakeup();
2277 if (m->act_count < ACT_INIT)
2278 m->act_count = ACT_INIT;
2279 vm_page_spin_unlock(m);
2284 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2285 * routine is called when a page has been added to the cache or free
2288 * This routine may not block.
2290 static __inline void
2291 vm_page_free_wakeup(void)
2293 globaldata_t gd = mycpu;
2296 * If the pageout daemon itself needs pages, then tell it that
2297 * there are some free.
2299 if (vm_pageout_pages_needed &&
2300 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >=
2301 gd->gd_vmstats.v_pageout_free_min
2303 vm_pageout_pages_needed = 0;
2304 wakeup(&vm_pageout_pages_needed);
2308 * Wakeup processes that are waiting on memory.
2310 * Generally speaking we want to wakeup stuck processes as soon as
2311 * possible. !vm_page_count_min(0) is the absolute minimum point
2312 * where we can do this. Wait a bit longer to reduce degenerate
2313 * re-blocking (vm_page_free_hysteresis). The target check is just
2314 * to make sure the min-check w/hysteresis does not exceed the
2317 if (vm_pages_waiting) {
2318 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2319 !vm_page_count_target()) {
2320 vm_pages_waiting = 0;
2321 wakeup(&vmstats.v_free_count);
2322 ++mycpu->gd_cnt.v_ppwakeups;
2325 if (!vm_page_count_target()) {
2327 * Plenty of pages are free, wakeup everyone.
2329 vm_pages_waiting = 0;
2330 wakeup(&vmstats.v_free_count);
2331 ++mycpu->gd_cnt.v_ppwakeups;
2332 } else if (!vm_page_count_min(0)) {
2334 * Some pages are free, wakeup someone.
2336 int wcount = vm_pages_waiting;
2339 vm_pages_waiting = wcount;
2340 wakeup_one(&vmstats.v_free_count);
2341 ++mycpu->gd_cnt.v_ppwakeups;
2348 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
2349 * it from its VM object.
2351 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
2352 * return (the page will have been freed).
2355 vm_page_free_toq(vm_page_t m)
2357 mycpu->gd_cnt.v_tfree++;
2358 KKASSERT((m->flags & PG_MAPPED) == 0);
2359 KKASSERT(m->flags & PG_BUSY);
2361 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
2362 kprintf("vm_page_free: pindex(%lu), busy(%d), "
2363 "PG_BUSY(%d), hold(%d)\n",
2364 (u_long)m->pindex, m->busy,
2365 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count);
2366 if ((m->queue - m->pc) == PQ_FREE)
2367 panic("vm_page_free: freeing free page");
2369 panic("vm_page_free: freeing busy page");
2373 * Remove from object, spinlock the page and its queues and
2374 * remove from any queue. No queue spinlock will be held
2375 * after this section (because the page was removed from any
2379 vm_page_and_queue_spin_lock(m);
2380 _vm_page_rem_queue_spinlocked(m);
2383 * No further management of fictitious pages occurs beyond object
2384 * and queue removal.
2386 if ((m->flags & PG_FICTITIOUS) != 0) {
2387 vm_page_spin_unlock(m);
2395 if (m->wire_count != 0) {
2396 if (m->wire_count > 1) {
2398 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2399 m->wire_count, (long)m->pindex);
2401 panic("vm_page_free: freeing wired page");
2405 * Clear the UNMANAGED flag when freeing an unmanaged page.
2406 * Clear the NEED_COMMIT flag
2408 if (m->flags & PG_UNMANAGED)
2409 vm_page_flag_clear(m, PG_UNMANAGED);
2410 if (m->flags & PG_NEED_COMMIT)
2411 vm_page_flag_clear(m, PG_NEED_COMMIT);
2413 if (m->hold_count != 0) {
2414 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2416 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1);
2420 * This sequence allows us to clear PG_BUSY while still holding
2421 * its spin lock, which reduces contention vs allocators. We
2422 * must not leave the queue locked or _vm_page_wakeup() may
2425 _vm_page_queue_spin_unlock(m);
2426 if (_vm_page_wakeup(m)) {
2427 vm_page_spin_unlock(m);
2430 vm_page_spin_unlock(m);
2432 vm_page_free_wakeup();
2436 * vm_page_unmanage()
2438 * Prevent PV management from being done on the page. The page is
2439 * removed from the paging queues as if it were wired, and as a
2440 * consequence of no longer being managed the pageout daemon will not
2441 * touch it (since there is no way to locate the pte mappings for the
2442 * page). madvise() calls that mess with the pmap will also no longer
2443 * operate on the page.
2445 * Beyond that the page is still reasonably 'normal'. Freeing the page
2446 * will clear the flag.
2448 * This routine is used by OBJT_PHYS objects - objects using unswappable
2449 * physical memory as backing store rather then swap-backed memory and
2450 * will eventually be extended to support 4MB unmanaged physical
2453 * Caller must be holding the page busy.
2456 vm_page_unmanage(vm_page_t m)
2458 KKASSERT(m->flags & PG_BUSY);
2459 if ((m->flags & PG_UNMANAGED) == 0) {
2460 if (m->wire_count == 0)
2463 vm_page_flag_set(m, PG_UNMANAGED);
2467 * Mark this page as wired down by yet another map, removing it from
2468 * paging queues as necessary.
2470 * Caller must be holding the page busy.
2473 vm_page_wire(vm_page_t m)
2476 * Only bump the wire statistics if the page is not already wired,
2477 * and only unqueue the page if it is on some queue (if it is unmanaged
2478 * it is already off the queues). Don't do anything with fictitious
2479 * pages because they are always wired.
2481 KKASSERT(m->flags & PG_BUSY);
2482 if ((m->flags & PG_FICTITIOUS) == 0) {
2483 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2484 if ((m->flags & PG_UNMANAGED) == 0)
2486 atomic_add_int(&mycpu->gd_vmstats_adj.v_wire_count, 1);
2488 KASSERT(m->wire_count != 0,
2489 ("vm_page_wire: wire_count overflow m=%p", m));
2494 * Release one wiring of this page, potentially enabling it to be paged again.
2496 * Many pages placed on the inactive queue should actually go
2497 * into the cache, but it is difficult to figure out which. What
2498 * we do instead, if the inactive target is well met, is to put
2499 * clean pages at the head of the inactive queue instead of the tail.
2500 * This will cause them to be moved to the cache more quickly and
2501 * if not actively re-referenced, freed more quickly. If we just
2502 * stick these pages at the end of the inactive queue, heavy filesystem
2503 * meta-data accesses can cause an unnecessary paging load on memory bound
2504 * processes. This optimization causes one-time-use metadata to be
2505 * reused more quickly.
2507 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2508 * the inactive queue. This helps the pageout daemon determine memory
2509 * pressure and act on out-of-memory situations more quickly.
2511 * BUT, if we are in a low-memory situation we have no choice but to
2512 * put clean pages on the cache queue.
2514 * A number of routines use vm_page_unwire() to guarantee that the page
2515 * will go into either the inactive or active queues, and will NEVER
2516 * be placed in the cache - for example, just after dirtying a page.
2517 * dirty pages in the cache are not allowed.
2519 * This routine may not block.
2522 vm_page_unwire(vm_page_t m, int activate)
2524 KKASSERT(m->flags & PG_BUSY);
2525 if (m->flags & PG_FICTITIOUS) {
2527 } else if (m->wire_count <= 0) {
2528 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2530 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2531 atomic_add_int(&mycpu->gd_vmstats_adj.v_wire_count, -1);
2532 if (m->flags & PG_UNMANAGED) {
2534 } else if (activate || (m->flags & PG_NEED_COMMIT)) {
2535 vm_page_spin_lock(m);
2536 _vm_page_add_queue_spinlocked(m,
2537 PQ_ACTIVE + m->pc, 0);
2538 _vm_page_and_queue_spin_unlock(m);
2540 vm_page_spin_lock(m);
2541 vm_page_flag_clear(m, PG_WINATCFLS);
2542 _vm_page_add_queue_spinlocked(m,
2543 PQ_INACTIVE + m->pc, 0);
2544 ++vm_swapcache_inactive_heuristic;
2545 _vm_page_and_queue_spin_unlock(m);
2552 * Move the specified page to the inactive queue. If the page has
2553 * any associated swap, the swap is deallocated.
2555 * Normally athead is 0 resulting in LRU operation. athead is set
2556 * to 1 if we want this page to be 'as if it were placed in the cache',
2557 * except without unmapping it from the process address space.
2559 * vm_page's spinlock must be held on entry and will remain held on return.
2560 * This routine may not block.
2563 _vm_page_deactivate_locked(vm_page_t m, int athead)
2568 * Ignore if already inactive.
2570 if (m->queue - m->pc == PQ_INACTIVE)
2572 _vm_page_queue_spin_lock(m);
2573 oqueue = _vm_page_rem_queue_spinlocked(m);
2575 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2576 if (oqueue == PQ_CACHE)
2577 mycpu->gd_cnt.v_reactivated++;
2578 vm_page_flag_clear(m, PG_WINATCFLS);
2579 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2581 ++vm_swapcache_inactive_heuristic;
2583 /* NOTE: PQ_NONE if condition not taken */
2584 _vm_page_queue_spin_unlock(m);
2585 /* leaves vm_page spinlocked */
2589 * Attempt to deactivate a page.
2594 vm_page_deactivate(vm_page_t m)
2596 vm_page_spin_lock(m);
2597 _vm_page_deactivate_locked(m, 0);
2598 vm_page_spin_unlock(m);
2602 vm_page_deactivate_locked(vm_page_t m)
2604 _vm_page_deactivate_locked(m, 0);
2608 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it.
2610 * This function returns non-zero if it successfully moved the page to
2613 * This function unconditionally unbusies the page on return.
2616 vm_page_try_to_cache(vm_page_t m)
2618 vm_page_spin_lock(m);
2619 if (m->dirty || m->hold_count || m->wire_count ||
2620 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2621 if (_vm_page_wakeup(m)) {
2622 vm_page_spin_unlock(m);
2625 vm_page_spin_unlock(m);
2629 vm_page_spin_unlock(m);
2632 * Page busied by us and no longer spinlocked. Dirty pages cannot
2633 * be moved to the cache.
2635 vm_page_test_dirty(m);
2636 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2645 * Attempt to free the page. If we cannot free it, we do nothing.
2646 * 1 is returned on success, 0 on failure.
2651 vm_page_try_to_free(vm_page_t m)
2653 vm_page_spin_lock(m);
2654 if (vm_page_busy_try(m, TRUE)) {
2655 vm_page_spin_unlock(m);
2660 * The page can be in any state, including already being on the free
2661 * queue. Check to see if it really can be freed.
2663 if (m->dirty || /* can't free if it is dirty */
2664 m->hold_count || /* or held (XXX may be wrong) */
2665 m->wire_count || /* or wired */
2666 (m->flags & (PG_UNMANAGED | /* or unmanaged */
2667 PG_NEED_COMMIT)) || /* or needs a commit */
2668 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */
2669 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */
2670 if (_vm_page_wakeup(m)) {
2671 vm_page_spin_unlock(m);
2674 vm_page_spin_unlock(m);
2678 vm_page_spin_unlock(m);
2681 * We can probably free the page.
2683 * Page busied by us and no longer spinlocked. Dirty pages will
2684 * not be freed by this function. We have to re-test the
2685 * dirty bit after cleaning out the pmaps.
2687 vm_page_test_dirty(m);
2688 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2692 vm_page_protect(m, VM_PROT_NONE);
2693 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2704 * Put the specified page onto the page cache queue (if appropriate).
2706 * The page must be busy, and this routine will release the busy and
2707 * possibly even free the page.
2710 vm_page_cache(vm_page_t m)
2713 * Not suitable for the cache
2715 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2716 m->busy || m->wire_count || m->hold_count) {
2722 * Already in the cache (and thus not mapped)
2724 if ((m->queue - m->pc) == PQ_CACHE) {
2725 KKASSERT((m->flags & PG_MAPPED) == 0);
2731 * Caller is required to test m->dirty, but note that the act of
2732 * removing the page from its maps can cause it to become dirty
2733 * on an SMP system due to another cpu running in usermode.
2736 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2741 * Remove all pmaps and indicate that the page is not
2742 * writeable or mapped. Our vm_page_protect() call may
2743 * have blocked (especially w/ VM_PROT_NONE), so recheck
2746 vm_page_protect(m, VM_PROT_NONE);
2747 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2748 m->busy || m->wire_count || m->hold_count) {
2750 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2751 vm_page_deactivate(m);
2754 _vm_page_and_queue_spin_lock(m);
2755 _vm_page_rem_queue_spinlocked(m);
2756 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2757 _vm_page_queue_spin_unlock(m);
2758 if (_vm_page_wakeup(m)) {
2759 vm_page_spin_unlock(m);
2762 vm_page_spin_unlock(m);
2764 vm_page_free_wakeup();
2769 * vm_page_dontneed()
2771 * Cache, deactivate, or do nothing as appropriate. This routine
2772 * is typically used by madvise() MADV_DONTNEED.
2774 * Generally speaking we want to move the page into the cache so
2775 * it gets reused quickly. However, this can result in a silly syndrome
2776 * due to the page recycling too quickly. Small objects will not be
2777 * fully cached. On the otherhand, if we move the page to the inactive
2778 * queue we wind up with a problem whereby very large objects
2779 * unnecessarily blow away our inactive and cache queues.
2781 * The solution is to move the pages based on a fixed weighting. We
2782 * either leave them alone, deactivate them, or move them to the cache,
2783 * where moving them to the cache has the highest weighting.
2784 * By forcing some pages into other queues we eventually force the
2785 * system to balance the queues, potentially recovering other unrelated
2786 * space from active. The idea is to not force this to happen too
2789 * The page must be busied.
2792 vm_page_dontneed(vm_page_t m)
2794 static int dnweight;
2801 * occassionally leave the page alone
2803 if ((dnw & 0x01F0) == 0 ||
2804 m->queue - m->pc == PQ_INACTIVE ||
2805 m->queue - m->pc == PQ_CACHE
2807 if (m->act_count >= ACT_INIT)
2813 * If vm_page_dontneed() is inactivating a page, it must clear
2814 * the referenced flag; otherwise the pagedaemon will see references
2815 * on the page in the inactive queue and reactivate it. Until the
2816 * page can move to the cache queue, madvise's job is not done.
2818 vm_page_flag_clear(m, PG_REFERENCED);
2819 pmap_clear_reference(m);
2822 vm_page_test_dirty(m);
2824 if (m->dirty || (dnw & 0x0070) == 0) {
2826 * Deactivate the page 3 times out of 32.
2831 * Cache the page 28 times out of every 32. Note that
2832 * the page is deactivated instead of cached, but placed
2833 * at the head of the queue instead of the tail.
2837 vm_page_spin_lock(m);
2838 _vm_page_deactivate_locked(m, head);
2839 vm_page_spin_unlock(m);
2843 * These routines manipulate the 'soft busy' count for a page. A soft busy
2844 * is almost like PG_BUSY except that it allows certain compatible operations
2845 * to occur on the page while it is busy. For example, a page undergoing a
2846 * write can still be mapped read-only.
2848 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2849 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2850 * busy bit is cleared.
2852 * The caller must hold the page BUSY when making these two calls.
2855 vm_page_io_start(vm_page_t m)
2857 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2858 atomic_add_char(&m->busy, 1);
2859 vm_page_flag_set(m, PG_SBUSY);
2863 vm_page_io_finish(vm_page_t m)
2865 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2866 atomic_subtract_char(&m->busy, 1);
2868 vm_page_flag_clear(m, PG_SBUSY);
2872 * Indicate that a clean VM page requires a filesystem commit and cannot
2873 * be reused. Used by tmpfs.
2876 vm_page_need_commit(vm_page_t m)
2878 vm_page_flag_set(m, PG_NEED_COMMIT);
2879 vm_object_set_writeable_dirty(m->object);
2883 vm_page_clear_commit(vm_page_t m)
2885 vm_page_flag_clear(m, PG_NEED_COMMIT);
2889 * Grab a page, blocking if it is busy and allocating a page if necessary.
2890 * A busy page is returned or NULL. The page may or may not be valid and
2891 * might not be on a queue (the caller is responsible for the disposition of
2894 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2895 * page will be zero'd and marked valid.
2897 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2898 * valid even if it already exists.
2900 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2901 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2902 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2904 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2905 * always returned if we had blocked.
2907 * This routine may not be called from an interrupt.
2909 * No other requirements.
2912 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2918 KKASSERT(allocflags &
2919 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2920 vm_object_hold_shared(object);
2922 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2924 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2925 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2930 } else if (m == NULL) {
2932 vm_object_upgrade(object);
2935 if (allocflags & VM_ALLOC_RETRY)
2936 allocflags |= VM_ALLOC_NULL_OK;
2937 m = vm_page_alloc(object, pindex,
2938 allocflags & ~VM_ALLOC_RETRY);
2942 if ((allocflags & VM_ALLOC_RETRY) == 0)
2951 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2953 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2954 * valid even if already valid.
2956 * NOTE! We have removed all of the PG_ZERO optimizations and also
2957 * removed the idle zeroing code. These optimizations actually
2958 * slow things down on modern cpus because the zerod area is
2959 * likely uncached, placing a memory-access burden on the
2960 * accesors taking the fault.
2962 * By always zeroing the page in-line with the fault, no
2963 * dynamic ram reads are needed and the caches are hot, ready
2964 * for userland to access the memory.
2966 if (m->valid == 0) {
2967 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2968 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2969 m->valid = VM_PAGE_BITS_ALL;
2971 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2972 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2973 m->valid = VM_PAGE_BITS_ALL;
2976 vm_object_drop(object);
2981 * Mapping function for valid bits or for dirty bits in
2982 * a page. May not block.
2984 * Inputs are required to range within a page.
2990 vm_page_bits(int base, int size)
2996 base + size <= PAGE_SIZE,
2997 ("vm_page_bits: illegal base/size %d/%d", base, size)
3000 if (size == 0) /* handle degenerate case */
3003 first_bit = base >> DEV_BSHIFT;
3004 last_bit = (base + size - 1) >> DEV_BSHIFT;
3006 return ((2 << last_bit) - (1 << first_bit));
3010 * Sets portions of a page valid and clean. The arguments are expected
3011 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3012 * of any partial chunks touched by the range. The invalid portion of
3013 * such chunks will be zero'd.
3015 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
3016 * align base to DEV_BSIZE so as not to mark clean a partially
3017 * truncated device block. Otherwise the dirty page status might be
3020 * This routine may not block.
3022 * (base + size) must be less then or equal to PAGE_SIZE.
3025 _vm_page_zero_valid(vm_page_t m, int base, int size)
3030 if (size == 0) /* handle degenerate case */
3034 * If the base is not DEV_BSIZE aligned and the valid
3035 * bit is clear, we have to zero out a portion of the
3039 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
3040 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
3042 pmap_zero_page_area(
3050 * If the ending offset is not DEV_BSIZE aligned and the
3051 * valid bit is clear, we have to zero out a portion of
3055 endoff = base + size;
3057 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
3058 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
3060 pmap_zero_page_area(
3063 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
3069 * Set valid, clear dirty bits. If validating the entire
3070 * page we can safely clear the pmap modify bit. We also
3071 * use this opportunity to clear the PG_NOSYNC flag. If a process
3072 * takes a write fault on a MAP_NOSYNC memory area the flag will
3075 * We set valid bits inclusive of any overlap, but we can only
3076 * clear dirty bits for DEV_BSIZE chunks that are fully within
3079 * Page must be busied?
3080 * No other requirements.
3083 vm_page_set_valid(vm_page_t m, int base, int size)
3085 _vm_page_zero_valid(m, base, size);
3086 m->valid |= vm_page_bits(base, size);
3091 * Set valid bits and clear dirty bits.
3093 * Page must be busied by caller.
3095 * NOTE: This function does not clear the pmap modified bit.
3096 * Also note that e.g. NFS may use a byte-granular base
3099 * No other requirements.
3102 vm_page_set_validclean(vm_page_t m, int base, int size)
3106 _vm_page_zero_valid(m, base, size);
3107 pagebits = vm_page_bits(base, size);
3108 m->valid |= pagebits;
3109 m->dirty &= ~pagebits;
3110 if (base == 0 && size == PAGE_SIZE) {
3111 /*pmap_clear_modify(m);*/
3112 vm_page_flag_clear(m, PG_NOSYNC);
3117 * Set valid & dirty. Used by buwrite()
3119 * Page must be busied by caller.
3122 vm_page_set_validdirty(vm_page_t m, int base, int size)
3126 pagebits = vm_page_bits(base, size);
3127 m->valid |= pagebits;
3128 m->dirty |= pagebits;
3130 vm_object_set_writeable_dirty(m->object);
3136 * NOTE: This function does not clear the pmap modified bit.
3137 * Also note that e.g. NFS may use a byte-granular base
3140 * Page must be busied?
3141 * No other requirements.
3144 vm_page_clear_dirty(vm_page_t m, int base, int size)
3146 m->dirty &= ~vm_page_bits(base, size);
3147 if (base == 0 && size == PAGE_SIZE) {
3148 /*pmap_clear_modify(m);*/
3149 vm_page_flag_clear(m, PG_NOSYNC);
3154 * Make the page all-dirty.
3156 * Also make sure the related object and vnode reflect the fact that the
3157 * object may now contain a dirty page.
3159 * Page must be busied?
3160 * No other requirements.
3163 vm_page_dirty(vm_page_t m)
3166 int pqtype = m->queue - m->pc;
3168 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
3169 ("vm_page_dirty: page in free/cache queue!"));
3170 if (m->dirty != VM_PAGE_BITS_ALL) {
3171 m->dirty = VM_PAGE_BITS_ALL;
3173 vm_object_set_writeable_dirty(m->object);
3178 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3179 * valid and dirty bits for the effected areas are cleared.
3181 * Page must be busied?
3183 * No other requirements.
3186 vm_page_set_invalid(vm_page_t m, int base, int size)
3190 bits = vm_page_bits(base, size);
3193 atomic_add_int(&m->object->generation, 1);
3197 * The kernel assumes that the invalid portions of a page contain
3198 * garbage, but such pages can be mapped into memory by user code.
3199 * When this occurs, we must zero out the non-valid portions of the
3200 * page so user code sees what it expects.
3202 * Pages are most often semi-valid when the end of a file is mapped
3203 * into memory and the file's size is not page aligned.
3205 * Page must be busied?
3206 * No other requirements.
3209 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3215 * Scan the valid bits looking for invalid sections that
3216 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
3217 * valid bit may be set ) have already been zerod by
3218 * vm_page_set_validclean().
3220 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3221 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3222 (m->valid & (1 << i))
3225 pmap_zero_page_area(
3228 (i - b) << DEV_BSHIFT
3236 * setvalid is TRUE when we can safely set the zero'd areas
3237 * as being valid. We can do this if there are no cache consistency
3238 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3241 m->valid = VM_PAGE_BITS_ALL;
3245 * Is a (partial) page valid? Note that the case where size == 0
3246 * will return FALSE in the degenerate case where the page is entirely
3247 * invalid, and TRUE otherwise.
3250 * No other requirements.
3253 vm_page_is_valid(vm_page_t m, int base, int size)
3255 int bits = vm_page_bits(base, size);
3257 if (m->valid && ((m->valid & bits) == bits))
3264 * update dirty bits from pmap/mmu. May not block.
3266 * Caller must hold the page busy
3269 vm_page_test_dirty(vm_page_t m)
3271 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
3277 * Register an action, associating it with its vm_page
3280 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
3282 struct vm_page_action_hash *hash;
3285 hv = (int)((intptr_t)action->m >> 8) & vmaction_hmask;
3286 hash = &action_hash[hv];
3288 lockmgr(&hash->lk, LK_EXCLUSIVE);
3289 vm_page_flag_set(action->m, PG_ACTIONLIST);
3290 action->event = event;
3291 LIST_INSERT_HEAD(&hash->list, action, entry);
3292 lockmgr(&hash->lk, LK_RELEASE);
3296 * Unregister an action, disassociating it from its related vm_page
3299 vm_page_unregister_action(vm_page_action_t action)
3301 struct vm_page_action_hash *hash;
3304 hv = (int)((intptr_t)action->m >> 8) & vmaction_hmask;
3305 hash = &action_hash[hv];
3306 lockmgr(&hash->lk, LK_EXCLUSIVE);
3307 if (action->event != VMEVENT_NONE) {
3308 action->event = VMEVENT_NONE;
3309 LIST_REMOVE(action, entry);
3311 if (LIST_EMPTY(&hash->list))
3312 vm_page_flag_clear(action->m, PG_ACTIONLIST);
3314 lockmgr(&hash->lk, LK_RELEASE);
3318 * Issue an event on a VM page. Corresponding action structures are
3319 * removed from the page's list and called.
3321 * If the vm_page has no more pending action events we clear its
3322 * PG_ACTIONLIST flag.
3325 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
3327 struct vm_page_action_hash *hash;
3328 struct vm_page_action *scan;
3329 struct vm_page_action *next;
3333 hv = (int)((intptr_t)m >> 8) & vmaction_hmask;
3334 hash = &action_hash[hv];
3337 lockmgr(&hash->lk, LK_EXCLUSIVE);
3338 LIST_FOREACH_MUTABLE(scan, &hash->list, entry, next) {
3340 if (scan->event == event) {
3341 scan->event = VMEVENT_NONE;
3342 LIST_REMOVE(scan, entry);
3343 scan->func(m, scan);
3351 vm_page_flag_clear(m, PG_ACTIONLIST);
3352 lockmgr(&hash->lk, LK_RELEASE);
3355 #include "opt_ddb.h"
3357 #include <ddb/ddb.h>
3359 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3361 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
3362 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
3363 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
3364 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
3365 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
3366 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
3367 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
3368 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
3369 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
3370 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
3373 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3376 db_printf("PQ_FREE:");
3377 for (i = 0; i < PQ_L2_SIZE; i++) {
3378 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
3382 db_printf("PQ_CACHE:");
3383 for(i = 0; i < PQ_L2_SIZE; i++) {
3384 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
3388 db_printf("PQ_ACTIVE:");
3389 for(i = 0; i < PQ_L2_SIZE; i++) {
3390 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3394 db_printf("PQ_INACTIVE:");
3395 for(i = 0; i < PQ_L2_SIZE; i++) {
3396 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);