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 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
1212 vm_page_spin_unlock(m);
1215 * Since we are inserting a new and possibly dirty page,
1216 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
1218 if ((m->valid & m->dirty) ||
1219 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT)))
1220 vm_object_set_writeable_dirty(object);
1223 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
1225 swap_pager_page_inserted(m);
1230 * Removes the given vm_page_t from the (object,index) table
1232 * The underlying pmap entry (if any) is NOT removed here.
1233 * This routine may not block.
1235 * The page must be BUSY and will remain BUSY on return.
1236 * No other requirements.
1238 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
1242 vm_page_remove(vm_page_t m)
1246 if (m->object == NULL) {
1250 if ((m->flags & PG_BUSY) == 0)
1251 panic("vm_page_remove: page not busy");
1255 vm_object_hold(object);
1258 * Remove the page from the object and update the object.
1260 * The vm_page spin lock is required for interactions with the pmap.
1262 vm_page_spin_lock(m);
1263 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
1264 --object->resident_page_count;
1265 --mycpu->gd_vmtotal.t_rm;
1266 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
1268 vm_page_spin_unlock(m);
1270 object->generation++;
1272 vm_object_drop(object);
1276 * Locate and return the page at (object, pindex), or NULL if the
1277 * page could not be found.
1279 * The caller must hold the vm_object token.
1282 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1287 * Search the hash table for this object/offset pair
1289 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1290 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1291 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1296 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1298 int also_m_busy, const char *msg
1304 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1305 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1307 KKASSERT(m->object == object && m->pindex == pindex);
1310 if (flags & PG_BUSY) {
1311 tsleep_interlock(m, 0);
1312 if (atomic_cmpset_int(&m->flags, flags,
1313 flags | PG_WANTED | PG_REFERENCED)) {
1314 tsleep(m, PINTERLOCKED, msg, 0);
1315 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1318 } else if (also_m_busy && (flags & PG_SBUSY)) {
1319 tsleep_interlock(m, 0);
1320 if (atomic_cmpset_int(&m->flags, flags,
1321 flags | PG_WANTED | PG_REFERENCED)) {
1322 tsleep(m, PINTERLOCKED, msg, 0);
1323 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1326 } else if (atomic_cmpset_int(&m->flags, flags,
1328 #ifdef VM_PAGE_DEBUG
1329 m->busy_func = func;
1330 m->busy_line = lineno;
1339 * Attempt to lookup and busy a page.
1341 * Returns NULL if the page could not be found
1343 * Returns a vm_page and error == TRUE if the page exists but could not
1346 * Returns a vm_page and error == FALSE on success.
1349 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1351 int also_m_busy, int *errorp
1357 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1358 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1361 KKASSERT(m->object == object && m->pindex == pindex);
1364 if (flags & PG_BUSY) {
1368 if (also_m_busy && (flags & PG_SBUSY)) {
1372 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1373 #ifdef VM_PAGE_DEBUG
1374 m->busy_func = func;
1375 m->busy_line = lineno;
1384 * Attempt to repurpose the passed-in page. If the passed-in page cannot
1385 * be repurposed it will be released, *must_reenter will be set to 1, and
1386 * this function will fall-through to vm_page_lookup_busy_try().
1388 * The passed-in page must be wired and not busy. The returned page will
1389 * be busied and not wired.
1391 * A different page may be returned. The returned page will be busied and
1394 * NULL can be returned. If so, the required page could not be busied.
1395 * The passed-in page will be unwired.
1398 vm_page_repurpose(struct vm_object *object, vm_pindex_t pindex,
1399 int also_m_busy, int *errorp, vm_page_t m,
1400 int *must_reenter, int *iswired)
1404 * Do not mess with pages in a complex state, such as pages
1405 * which are mapped, as repurposing such pages can be more
1406 * expensive than simply allocatin a new one.
1408 * NOTE: Soft-busying can deadlock against putpages or I/O
1409 * so we only allow hard-busying here.
1411 KKASSERT(also_m_busy == FALSE);
1412 vm_page_busy_wait(m, also_m_busy, "biodep");
1414 if ((m->flags & (PG_UNMANAGED | PG_MAPPED |
1415 PG_FICTITIOUS | PG_SBUSY)) ||
1416 m->busy || m->wire_count != 1 || m->hold_count) {
1417 vm_page_unwire(m, 0);
1419 /* fall through to normal lookup */
1420 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
1421 vm_page_unwire(m, 0);
1422 vm_page_deactivate(m);
1424 /* fall through to normal lookup */
1427 * We can safely repurpose the page. It should
1428 * already be unqueued.
1430 KKASSERT(m->queue == PQ_NONE && m->dirty == 0);
1434 if (vm_page_insert(m, object, pindex)) {
1440 vm_page_unwire(m, 0);
1442 /* fall through to normal lookup */
1447 * Cannot repurpose page, attempt to locate the desired page. May
1452 m = vm_page_lookup_busy_try(object, pindex, also_m_busy, errorp);
1458 * Caller must hold the related vm_object
1461 vm_page_next(vm_page_t m)
1465 next = vm_page_rb_tree_RB_NEXT(m);
1466 if (next && next->pindex != m->pindex + 1)
1474 * Move the given vm_page from its current object to the specified
1475 * target object/offset. The page must be busy and will remain so
1478 * new_object must be held.
1479 * This routine might block. XXX ?
1481 * NOTE: Swap associated with the page must be invalidated by the move. We
1482 * have to do this for several reasons: (1) we aren't freeing the
1483 * page, (2) we are dirtying the page, (3) the VM system is probably
1484 * moving the page from object A to B, and will then later move
1485 * the backing store from A to B and we can't have a conflict.
1487 * NOTE: We *always* dirty the page. It is necessary both for the
1488 * fact that we moved it, and because we may be invalidating
1489 * swap. If the page is on the cache, we have to deactivate it
1490 * or vm_page_dirty() will panic. Dirty pages are not allowed
1494 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1496 KKASSERT(m->flags & PG_BUSY);
1497 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object));
1499 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object));
1502 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1503 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1504 new_object, new_pindex);
1506 if (m->queue - m->pc == PQ_CACHE)
1507 vm_page_deactivate(m);
1512 * vm_page_unqueue() without any wakeup. This routine is used when a page
1513 * is to remain BUSYied by the caller.
1515 * This routine may not block.
1518 vm_page_unqueue_nowakeup(vm_page_t m)
1520 vm_page_and_queue_spin_lock(m);
1521 (void)_vm_page_rem_queue_spinlocked(m);
1522 vm_page_spin_unlock(m);
1526 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1529 * This routine may not block.
1532 vm_page_unqueue(vm_page_t m)
1536 vm_page_and_queue_spin_lock(m);
1537 queue = _vm_page_rem_queue_spinlocked(m);
1538 if (queue == PQ_FREE || queue == PQ_CACHE) {
1539 vm_page_spin_unlock(m);
1540 pagedaemon_wakeup();
1542 vm_page_spin_unlock(m);
1547 * vm_page_list_find()
1549 * Find a page on the specified queue with color optimization.
1551 * The page coloring optimization attempts to locate a page that does
1552 * not overload other nearby pages in the object in the cpu's L1 or L2
1553 * caches. We need this optimization because cpu caches tend to be
1554 * physical caches, while object spaces tend to be virtual.
1556 * The page coloring optimization also, very importantly, tries to localize
1557 * memory to cpus and physical sockets.
1559 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1560 * and the algorithm is adjusted to localize allocations on a per-core basis.
1561 * This is done by 'twisting' the colors.
1563 * The page is returned spinlocked and removed from its queue (it will
1564 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1565 * is responsible for dealing with the busy-page case (usually by
1566 * deactivating the page and looping).
1568 * NOTE: This routine is carefully inlined. A non-inlined version
1569 * is available for outside callers but the only critical path is
1570 * from within this source file.
1572 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1573 * represent stable storage, allowing us to order our locks vm_page
1574 * first, then queue.
1578 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1584 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl,
1587 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1590 m = _vm_page_list_find2(basequeue, index);
1593 vm_page_and_queue_spin_lock(m);
1594 if (m->queue == basequeue + index) {
1595 _vm_page_rem_queue_spinlocked(m);
1596 /* vm_page_t spin held, no queue spin */
1599 vm_page_and_queue_spin_unlock(m);
1605 * If we could not find the page in the desired queue try to find it in
1609 _vm_page_list_find2(int basequeue, int index)
1611 struct vpgqueues *pq;
1613 int pqmask = PQ_SET_ASSOC_MASK >> 1;
1617 index &= PQ_L2_MASK;
1618 pq = &vm_page_queues[basequeue];
1621 * Run local sets of 16, 32, 64, 128, and the whole queue if all
1622 * else fails (PQ_L2_MASK which is 255).
1625 pqmask = (pqmask << 1) | 1;
1626 for (i = 0; i <= pqmask; ++i) {
1627 pqi = (index & ~pqmask) | ((index + i) & pqmask);
1628 m = TAILQ_FIRST(&pq[pqi].pl);
1630 _vm_page_and_queue_spin_lock(m);
1631 if (m->queue == basequeue + pqi) {
1632 _vm_page_rem_queue_spinlocked(m);
1635 _vm_page_and_queue_spin_unlock(m);
1640 } while (pqmask != PQ_L2_MASK);
1646 * Returns a vm_page candidate for allocation. The page is not busied so
1647 * it can move around. The caller must busy the page (and typically
1648 * deactivate it if it cannot be busied!)
1650 * Returns a spinlocked vm_page that has been removed from its queue.
1653 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1655 return(_vm_page_list_find(basequeue, index, prefer_zero));
1659 * Find a page on the cache queue with color optimization, remove it
1660 * from the queue, and busy it. The returned page will not be spinlocked.
1662 * A candidate failure will be deactivated. Candidates can fail due to
1663 * being busied by someone else, in which case they will be deactivated.
1665 * This routine may not block.
1669 vm_page_select_cache(u_short pg_color)
1674 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE);
1678 * (m) has been removed from its queue and spinlocked
1680 if (vm_page_busy_try(m, TRUE)) {
1681 _vm_page_deactivate_locked(m, 0);
1682 vm_page_spin_unlock(m);
1685 * We successfully busied the page
1687 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 &&
1688 m->hold_count == 0 &&
1689 m->wire_count == 0 &&
1690 (m->dirty & m->valid) == 0) {
1691 vm_page_spin_unlock(m);
1692 pagedaemon_wakeup();
1697 * The page cannot be recycled, deactivate it.
1699 _vm_page_deactivate_locked(m, 0);
1700 if (_vm_page_wakeup(m)) {
1701 vm_page_spin_unlock(m);
1704 vm_page_spin_unlock(m);
1712 * Find a free or zero page, with specified preference. We attempt to
1713 * inline the nominal case and fall back to _vm_page_select_free()
1714 * otherwise. A busied page is removed from the queue and returned.
1716 * This routine may not block.
1718 static __inline vm_page_t
1719 vm_page_select_free(u_short pg_color, boolean_t prefer_zero)
1724 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK,
1728 if (vm_page_busy_try(m, TRUE)) {
1730 * Various mechanisms such as a pmap_collect can
1731 * result in a busy page on the free queue. We
1732 * have to move the page out of the way so we can
1733 * retry the allocation. If the other thread is not
1734 * allocating the page then m->valid will remain 0 and
1735 * the pageout daemon will free the page later on.
1737 * Since we could not busy the page, however, we
1738 * cannot make assumptions as to whether the page
1739 * will be allocated by the other thread or not,
1740 * so all we can do is deactivate it to move it out
1741 * of the way. In particular, if the other thread
1742 * wires the page it may wind up on the inactive
1743 * queue and the pageout daemon will have to deal
1744 * with that case too.
1746 _vm_page_deactivate_locked(m, 0);
1747 vm_page_spin_unlock(m);
1750 * Theoretically if we are able to busy the page
1751 * atomic with the queue removal (using the vm_page
1752 * lock) nobody else should be able to mess with the
1755 KKASSERT((m->flags & (PG_UNMANAGED |
1756 PG_NEED_COMMIT)) == 0);
1757 KASSERT(m->hold_count == 0, ("m->hold_count is not zero "
1758 "pg %p q=%d flags=%08x hold=%d wire=%d",
1759 m, m->queue, m->flags, m->hold_count, m->wire_count));
1760 KKASSERT(m->wire_count == 0);
1761 vm_page_spin_unlock(m);
1762 pagedaemon_wakeup();
1764 /* return busied and removed page */
1774 * Allocate and return a memory cell associated with this VM object/offset
1775 * pair. If object is NULL an unassociated page will be allocated.
1777 * The returned page will be busied and removed from its queues. This
1778 * routine can block and may return NULL if a race occurs and the page
1779 * is found to already exist at the specified (object, pindex).
1781 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1782 * VM_ALLOC_QUICK like normal but cannot use cache
1783 * VM_ALLOC_SYSTEM greater free drain
1784 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1785 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1786 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1787 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1788 * (see vm_page_grab())
1789 * VM_ALLOC_USE_GD ok to use per-gd cache
1791 * VM_ALLOC_CPU(n) allocate using specified cpu localization
1793 * The object must be held if not NULL
1794 * This routine may not block
1796 * Additional special handling is required when called from an interrupt
1797 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1801 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1811 * Special per-cpu free VM page cache. The pages are pre-busied
1812 * and pre-zerod for us.
1814 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1816 if (gd->gd_vmpg_count) {
1817 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1829 * CPU localization algorithm. Break the page queues up by physical
1830 * id and core id (note that two cpu threads will have the same core
1831 * id, and core_id != gd_cpuid).
1833 * This is nowhere near perfect, for example the last pindex in a
1834 * subgroup will overflow into the next cpu or package. But this
1835 * should get us good page reuse locality in heavy mixed loads.
1837 * (may be executed before the APs are started, so other GDs might
1840 if (page_req & VM_ALLOC_CPU_SPEC)
1841 cpuid_local = VM_ALLOC_GETCPU(page_req);
1843 cpuid_local = mycpu->gd_cpuid;
1845 pg_color = vm_get_pg_color(cpuid_local, object, pindex);
1848 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1849 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1852 * Certain system threads (pageout daemon, buf_daemon's) are
1853 * allowed to eat deeper into the free page list.
1855 if (curthread->td_flags & TDF_SYSTHREAD)
1856 page_req |= VM_ALLOC_SYSTEM;
1859 * Impose various limitations. Note that the v_free_reserved test
1860 * must match the opposite of vm_page_count_target() to avoid
1861 * livelocks, be careful.
1865 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved ||
1866 ((page_req & VM_ALLOC_INTERRUPT) &&
1867 gd->gd_vmstats.v_free_count > 0) ||
1868 ((page_req & VM_ALLOC_SYSTEM) &&
1869 gd->gd_vmstats.v_cache_count == 0 &&
1870 gd->gd_vmstats.v_free_count >
1871 gd->gd_vmstats.v_interrupt_free_min)
1874 * The free queue has sufficient free pages to take one out.
1876 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO))
1877 m = vm_page_select_free(pg_color, TRUE);
1879 m = vm_page_select_free(pg_color, FALSE);
1880 } else if (page_req & VM_ALLOC_NORMAL) {
1882 * Allocatable from the cache (non-interrupt only). On
1883 * success, we must free the page and try again, thus
1884 * ensuring that vmstats.v_*_free_min counters are replenished.
1887 if (curthread->td_preempted) {
1888 kprintf("vm_page_alloc(): warning, attempt to allocate"
1889 " cache page from preempting interrupt\n");
1892 m = vm_page_select_cache(pg_color);
1895 m = vm_page_select_cache(pg_color);
1898 * On success move the page into the free queue and loop.
1900 * Only do this if we can safely acquire the vm_object lock,
1901 * because this is effectively a random page and the caller
1902 * might be holding the lock shared, we don't want to
1906 KASSERT(m->dirty == 0,
1907 ("Found dirty cache page %p", m));
1908 if ((obj = m->object) != NULL) {
1909 if (vm_object_hold_try(obj)) {
1910 vm_page_protect(m, VM_PROT_NONE);
1912 /* m->object NULL here */
1913 vm_object_drop(obj);
1915 vm_page_deactivate(m);
1919 vm_page_protect(m, VM_PROT_NONE);
1926 * On failure return NULL
1928 atomic_add_int(&vm_pageout_deficit, 1);
1929 pagedaemon_wakeup();
1933 * No pages available, wakeup the pageout daemon and give up.
1935 atomic_add_int(&vm_pageout_deficit, 1);
1936 pagedaemon_wakeup();
1941 * v_free_count can race so loop if we don't find the expected
1950 * Good page found. The page has already been busied for us and
1951 * removed from its queues.
1953 KASSERT(m->dirty == 0,
1954 ("vm_page_alloc: free/cache page %p was dirty", m));
1955 KKASSERT(m->queue == PQ_NONE);
1961 * Initialize the structure, inheriting some flags but clearing
1962 * all the rest. The page has already been busied for us.
1964 vm_page_flag_clear(m, ~(PG_BUSY | PG_SBUSY));
1965 KKASSERT(m->wire_count == 0);
1966 KKASSERT(m->busy == 0);
1971 * Caller must be holding the object lock (asserted by
1972 * vm_page_insert()).
1974 * NOTE: Inserting a page here does not insert it into any pmaps
1975 * (which could cause us to block allocating memory).
1977 * NOTE: If no object an unassociated page is allocated, m->pindex
1978 * can be used by the caller for any purpose.
1981 if (vm_page_insert(m, object, pindex) == FALSE) {
1983 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1984 panic("PAGE RACE %p[%ld]/%p",
1985 object, (long)pindex, m);
1993 * Don't wakeup too often - wakeup the pageout daemon when
1994 * we would be nearly out of memory.
1996 pagedaemon_wakeup();
1999 * A PG_BUSY page is returned.
2005 * Returns number of pages available in our DMA memory reserve
2006 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf)
2009 vm_contig_avail_pages(void)
2014 spin_lock(&vm_contig_spin);
2015 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
2016 spin_unlock(&vm_contig_spin);
2022 * Attempt to allocate contiguous physical memory with the specified
2026 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
2027 unsigned long alignment, unsigned long boundary,
2028 unsigned long size, vm_memattr_t memattr)
2034 alignment >>= PAGE_SHIFT;
2037 boundary >>= PAGE_SHIFT;
2040 size = (size + PAGE_MASK) >> PAGE_SHIFT;
2042 spin_lock(&vm_contig_spin);
2043 blk = alist_alloc(&vm_contig_alist, 0, size);
2044 if (blk == ALIST_BLOCK_NONE) {
2045 spin_unlock(&vm_contig_spin);
2047 kprintf("vm_page_alloc_contig: %ldk nospace\n",
2048 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
2052 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
2053 alist_free(&vm_contig_alist, blk, size);
2054 spin_unlock(&vm_contig_spin);
2056 kprintf("vm_page_alloc_contig: %ldk high "
2058 (size + PAGE_MASK) * (PAGE_SIZE / 1024),
2063 spin_unlock(&vm_contig_spin);
2064 if (vm_contig_verbose) {
2065 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
2066 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT,
2067 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
2070 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
2071 if (memattr != VM_MEMATTR_DEFAULT)
2072 for (i = 0;i < size;i++)
2073 pmap_page_set_memattr(&m[i], memattr);
2078 * Free contiguously allocated pages. The pages will be wired but not busy.
2079 * When freeing to the alist we leave them wired and not busy.
2082 vm_page_free_contig(vm_page_t m, unsigned long size)
2084 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
2085 vm_pindex_t start = pa >> PAGE_SHIFT;
2086 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
2088 if (vm_contig_verbose) {
2089 kprintf("vm_page_free_contig: %016jx/%ldk\n",
2090 (intmax_t)pa, size / 1024);
2092 if (pa < vm_low_phys_reserved) {
2093 KKASSERT(pa + size <= vm_low_phys_reserved);
2094 spin_lock(&vm_contig_spin);
2095 alist_free(&vm_contig_alist, start, pages);
2096 spin_unlock(&vm_contig_spin);
2099 vm_page_busy_wait(m, FALSE, "cpgfr");
2100 vm_page_unwire(m, 0);
2111 * Wait for sufficient free memory for nominal heavy memory use kernel
2114 * WARNING! Be sure never to call this in any vm_pageout code path, which
2115 * will trivially deadlock the system.
2118 vm_wait_nominal(void)
2120 while (vm_page_count_min(0))
2125 * Test if vm_wait_nominal() would block.
2128 vm_test_nominal(void)
2130 if (vm_page_count_min(0))
2136 * Block until free pages are available for allocation, called in various
2137 * places before memory allocations.
2139 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
2140 * more generous then that.
2146 * never wait forever
2150 lwkt_gettoken(&vm_token);
2152 if (curthread == pagethread) {
2154 * The pageout daemon itself needs pages, this is bad.
2156 if (vm_page_count_min(0)) {
2157 vm_pageout_pages_needed = 1;
2158 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
2162 * Wakeup the pageout daemon if necessary and wait.
2164 * Do not wait indefinitely for the target to be reached,
2165 * as load might prevent it from being reached any time soon.
2166 * But wait a little to try to slow down page allocations
2167 * and to give more important threads (the pagedaemon)
2168 * allocation priority.
2170 if (vm_page_count_target()) {
2171 if (vm_pages_needed == 0) {
2172 vm_pages_needed = 1;
2173 wakeup(&vm_pages_needed);
2175 ++vm_pages_waiting; /* SMP race ok */
2176 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
2179 lwkt_reltoken(&vm_token);
2183 * Block until free pages are available for allocation
2185 * Called only from vm_fault so that processes page faulting can be
2189 vm_wait_pfault(void)
2192 * Wakeup the pageout daemon if necessary and wait.
2194 * Do not wait indefinitely for the target to be reached,
2195 * as load might prevent it from being reached any time soon.
2196 * But wait a little to try to slow down page allocations
2197 * and to give more important threads (the pagedaemon)
2198 * allocation priority.
2200 if (vm_page_count_min(0)) {
2201 lwkt_gettoken(&vm_token);
2202 while (vm_page_count_severe()) {
2203 if (vm_page_count_target()) {
2206 if (vm_pages_needed == 0) {
2207 vm_pages_needed = 1;
2208 wakeup(&vm_pages_needed);
2210 ++vm_pages_waiting; /* SMP race ok */
2211 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
2214 * Do not stay stuck in the loop if the system is trying
2215 * to kill the process.
2218 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
2222 lwkt_reltoken(&vm_token);
2227 * Put the specified page on the active list (if appropriate). Ensure
2228 * that act_count is at least ACT_INIT but do not otherwise mess with it.
2230 * The caller should be holding the page busied ? XXX
2231 * This routine may not block.
2234 vm_page_activate(vm_page_t m)
2238 vm_page_spin_lock(m);
2239 if (m->queue - m->pc != PQ_ACTIVE) {
2240 _vm_page_queue_spin_lock(m);
2241 oqueue = _vm_page_rem_queue_spinlocked(m);
2242 /* page is left spinlocked, queue is unlocked */
2244 if (oqueue == PQ_CACHE)
2245 mycpu->gd_cnt.v_reactivated++;
2246 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2247 if (m->act_count < ACT_INIT)
2248 m->act_count = ACT_INIT;
2249 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
2251 _vm_page_and_queue_spin_unlock(m);
2252 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
2253 pagedaemon_wakeup();
2255 if (m->act_count < ACT_INIT)
2256 m->act_count = ACT_INIT;
2257 vm_page_spin_unlock(m);
2262 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2263 * routine is called when a page has been added to the cache or free
2266 * This routine may not block.
2268 static __inline void
2269 vm_page_free_wakeup(void)
2271 globaldata_t gd = mycpu;
2274 * If the pageout daemon itself needs pages, then tell it that
2275 * there are some free.
2277 if (vm_pageout_pages_needed &&
2278 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >=
2279 gd->gd_vmstats.v_pageout_free_min
2281 vm_pageout_pages_needed = 0;
2282 wakeup(&vm_pageout_pages_needed);
2286 * Wakeup processes that are waiting on memory.
2288 * Generally speaking we want to wakeup stuck processes as soon as
2289 * possible. !vm_page_count_min(0) is the absolute minimum point
2290 * where we can do this. Wait a bit longer to reduce degenerate
2291 * re-blocking (vm_page_free_hysteresis). The target check is just
2292 * to make sure the min-check w/hysteresis does not exceed the
2295 if (vm_pages_waiting) {
2296 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2297 !vm_page_count_target()) {
2298 vm_pages_waiting = 0;
2299 wakeup(&vmstats.v_free_count);
2300 ++mycpu->gd_cnt.v_ppwakeups;
2303 if (!vm_page_count_target()) {
2305 * Plenty of pages are free, wakeup everyone.
2307 vm_pages_waiting = 0;
2308 wakeup(&vmstats.v_free_count);
2309 ++mycpu->gd_cnt.v_ppwakeups;
2310 } else if (!vm_page_count_min(0)) {
2312 * Some pages are free, wakeup someone.
2314 int wcount = vm_pages_waiting;
2317 vm_pages_waiting = wcount;
2318 wakeup_one(&vmstats.v_free_count);
2319 ++mycpu->gd_cnt.v_ppwakeups;
2326 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
2327 * it from its VM object.
2329 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
2330 * return (the page will have been freed).
2333 vm_page_free_toq(vm_page_t m)
2335 mycpu->gd_cnt.v_tfree++;
2336 KKASSERT((m->flags & PG_MAPPED) == 0);
2337 KKASSERT(m->flags & PG_BUSY);
2339 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
2340 kprintf("vm_page_free: pindex(%lu), busy(%d), "
2341 "PG_BUSY(%d), hold(%d)\n",
2342 (u_long)m->pindex, m->busy,
2343 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count);
2344 if ((m->queue - m->pc) == PQ_FREE)
2345 panic("vm_page_free: freeing free page");
2347 panic("vm_page_free: freeing busy page");
2351 * Remove from object, spinlock the page and its queues and
2352 * remove from any queue. No queue spinlock will be held
2353 * after this section (because the page was removed from any
2357 vm_page_and_queue_spin_lock(m);
2358 _vm_page_rem_queue_spinlocked(m);
2361 * No further management of fictitious pages occurs beyond object
2362 * and queue removal.
2364 if ((m->flags & PG_FICTITIOUS) != 0) {
2365 vm_page_spin_unlock(m);
2373 if (m->wire_count != 0) {
2374 if (m->wire_count > 1) {
2376 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2377 m->wire_count, (long)m->pindex);
2379 panic("vm_page_free: freeing wired page");
2383 * Clear the UNMANAGED flag when freeing an unmanaged page.
2384 * Clear the NEED_COMMIT flag
2386 if (m->flags & PG_UNMANAGED)
2387 vm_page_flag_clear(m, PG_UNMANAGED);
2388 if (m->flags & PG_NEED_COMMIT)
2389 vm_page_flag_clear(m, PG_NEED_COMMIT);
2391 if (m->hold_count != 0) {
2392 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2394 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
2398 * This sequence allows us to clear PG_BUSY while still holding
2399 * its spin lock, which reduces contention vs allocators. We
2400 * must not leave the queue locked or _vm_page_wakeup() may
2403 _vm_page_queue_spin_unlock(m);
2404 if (_vm_page_wakeup(m)) {
2405 vm_page_spin_unlock(m);
2408 vm_page_spin_unlock(m);
2410 vm_page_free_wakeup();
2414 * vm_page_unmanage()
2416 * Prevent PV management from being done on the page. The page is
2417 * removed from the paging queues as if it were wired, and as a
2418 * consequence of no longer being managed the pageout daemon will not
2419 * touch it (since there is no way to locate the pte mappings for the
2420 * page). madvise() calls that mess with the pmap will also no longer
2421 * operate on the page.
2423 * Beyond that the page is still reasonably 'normal'. Freeing the page
2424 * will clear the flag.
2426 * This routine is used by OBJT_PHYS objects - objects using unswappable
2427 * physical memory as backing store rather then swap-backed memory and
2428 * will eventually be extended to support 4MB unmanaged physical
2431 * Caller must be holding the page busy.
2434 vm_page_unmanage(vm_page_t m)
2436 KKASSERT(m->flags & PG_BUSY);
2437 if ((m->flags & PG_UNMANAGED) == 0) {
2438 if (m->wire_count == 0)
2441 vm_page_flag_set(m, PG_UNMANAGED);
2445 * Mark this page as wired down by yet another map, removing it from
2446 * paging queues as necessary.
2448 * Caller must be holding the page busy.
2451 vm_page_wire(vm_page_t m)
2454 * Only bump the wire statistics if the page is not already wired,
2455 * and only unqueue the page if it is on some queue (if it is unmanaged
2456 * it is already off the queues). Don't do anything with fictitious
2457 * pages because they are always wired.
2459 KKASSERT(m->flags & PG_BUSY);
2460 if ((m->flags & PG_FICTITIOUS) == 0) {
2461 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2462 if ((m->flags & PG_UNMANAGED) == 0)
2464 atomic_add_int(&mycpu->gd_vmstats_adj.v_wire_count, 1);
2466 KASSERT(m->wire_count != 0,
2467 ("vm_page_wire: wire_count overflow m=%p", m));
2472 * Release one wiring of this page, potentially enabling it to be paged again.
2474 * Many pages placed on the inactive queue should actually go
2475 * into the cache, but it is difficult to figure out which. What
2476 * we do instead, if the inactive target is well met, is to put
2477 * clean pages at the head of the inactive queue instead of the tail.
2478 * This will cause them to be moved to the cache more quickly and
2479 * if not actively re-referenced, freed more quickly. If we just
2480 * stick these pages at the end of the inactive queue, heavy filesystem
2481 * meta-data accesses can cause an unnecessary paging load on memory bound
2482 * processes. This optimization causes one-time-use metadata to be
2483 * reused more quickly.
2485 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2486 * the inactive queue. This helps the pageout daemon determine memory
2487 * pressure and act on out-of-memory situations more quickly.
2489 * BUT, if we are in a low-memory situation we have no choice but to
2490 * put clean pages on the cache queue.
2492 * A number of routines use vm_page_unwire() to guarantee that the page
2493 * will go into either the inactive or active queues, and will NEVER
2494 * be placed in the cache - for example, just after dirtying a page.
2495 * dirty pages in the cache are not allowed.
2497 * This routine may not block.
2500 vm_page_unwire(vm_page_t m, int activate)
2502 KKASSERT(m->flags & PG_BUSY);
2503 if (m->flags & PG_FICTITIOUS) {
2505 } else if (m->wire_count <= 0) {
2506 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2508 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2509 atomic_add_int(&mycpu->gd_vmstats_adj.v_wire_count, -1);
2510 if (m->flags & PG_UNMANAGED) {
2512 } else if (activate || (m->flags & PG_NEED_COMMIT)) {
2513 vm_page_spin_lock(m);
2514 _vm_page_add_queue_spinlocked(m,
2515 PQ_ACTIVE + m->pc, 0);
2516 _vm_page_and_queue_spin_unlock(m);
2518 vm_page_spin_lock(m);
2519 vm_page_flag_clear(m, PG_WINATCFLS);
2520 _vm_page_add_queue_spinlocked(m,
2521 PQ_INACTIVE + m->pc, 0);
2522 ++vm_swapcache_inactive_heuristic;
2523 _vm_page_and_queue_spin_unlock(m);
2530 * Move the specified page to the inactive queue. If the page has
2531 * any associated swap, the swap is deallocated.
2533 * Normally athead is 0 resulting in LRU operation. athead is set
2534 * to 1 if we want this page to be 'as if it were placed in the cache',
2535 * except without unmapping it from the process address space.
2537 * vm_page's spinlock must be held on entry and will remain held on return.
2538 * This routine may not block.
2541 _vm_page_deactivate_locked(vm_page_t m, int athead)
2546 * Ignore if already inactive.
2548 if (m->queue - m->pc == PQ_INACTIVE)
2550 _vm_page_queue_spin_lock(m);
2551 oqueue = _vm_page_rem_queue_spinlocked(m);
2553 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2554 if (oqueue == PQ_CACHE)
2555 mycpu->gd_cnt.v_reactivated++;
2556 vm_page_flag_clear(m, PG_WINATCFLS);
2557 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2559 ++vm_swapcache_inactive_heuristic;
2561 /* NOTE: PQ_NONE if condition not taken */
2562 _vm_page_queue_spin_unlock(m);
2563 /* leaves vm_page spinlocked */
2567 * Attempt to deactivate a page.
2572 vm_page_deactivate(vm_page_t m)
2574 vm_page_spin_lock(m);
2575 _vm_page_deactivate_locked(m, 0);
2576 vm_page_spin_unlock(m);
2580 vm_page_deactivate_locked(vm_page_t m)
2582 _vm_page_deactivate_locked(m, 0);
2586 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it.
2588 * This function returns non-zero if it successfully moved the page to
2591 * This function unconditionally unbusies the page on return.
2594 vm_page_try_to_cache(vm_page_t m)
2596 vm_page_spin_lock(m);
2597 if (m->dirty || m->hold_count || m->wire_count ||
2598 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2599 if (_vm_page_wakeup(m)) {
2600 vm_page_spin_unlock(m);
2603 vm_page_spin_unlock(m);
2607 vm_page_spin_unlock(m);
2610 * Page busied by us and no longer spinlocked. Dirty pages cannot
2611 * be moved to the cache.
2613 vm_page_test_dirty(m);
2614 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2623 * Attempt to free the page. If we cannot free it, we do nothing.
2624 * 1 is returned on success, 0 on failure.
2629 vm_page_try_to_free(vm_page_t m)
2631 vm_page_spin_lock(m);
2632 if (vm_page_busy_try(m, TRUE)) {
2633 vm_page_spin_unlock(m);
2638 * The page can be in any state, including already being on the free
2639 * queue. Check to see if it really can be freed.
2641 if (m->dirty || /* can't free if it is dirty */
2642 m->hold_count || /* or held (XXX may be wrong) */
2643 m->wire_count || /* or wired */
2644 (m->flags & (PG_UNMANAGED | /* or unmanaged */
2645 PG_NEED_COMMIT)) || /* or needs a commit */
2646 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */
2647 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */
2648 if (_vm_page_wakeup(m)) {
2649 vm_page_spin_unlock(m);
2652 vm_page_spin_unlock(m);
2656 vm_page_spin_unlock(m);
2659 * We can probably free the page.
2661 * Page busied by us and no longer spinlocked. Dirty pages will
2662 * not be freed by this function. We have to re-test the
2663 * dirty bit after cleaning out the pmaps.
2665 vm_page_test_dirty(m);
2666 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2670 vm_page_protect(m, VM_PROT_NONE);
2671 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2682 * Put the specified page onto the page cache queue (if appropriate).
2684 * The page must be busy, and this routine will release the busy and
2685 * possibly even free the page.
2688 vm_page_cache(vm_page_t m)
2691 * Not suitable for the cache
2693 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2694 m->busy || m->wire_count || m->hold_count) {
2700 * Already in the cache (and thus not mapped)
2702 if ((m->queue - m->pc) == PQ_CACHE) {
2703 KKASSERT((m->flags & PG_MAPPED) == 0);
2709 * Caller is required to test m->dirty, but note that the act of
2710 * removing the page from its maps can cause it to become dirty
2711 * on an SMP system due to another cpu running in usermode.
2714 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2719 * Remove all pmaps and indicate that the page is not
2720 * writeable or mapped. Our vm_page_protect() call may
2721 * have blocked (especially w/ VM_PROT_NONE), so recheck
2724 vm_page_protect(m, VM_PROT_NONE);
2725 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2726 m->busy || m->wire_count || m->hold_count) {
2728 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2729 vm_page_deactivate(m);
2732 _vm_page_and_queue_spin_lock(m);
2733 _vm_page_rem_queue_spinlocked(m);
2734 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2735 _vm_page_queue_spin_unlock(m);
2736 if (_vm_page_wakeup(m)) {
2737 vm_page_spin_unlock(m);
2740 vm_page_spin_unlock(m);
2742 vm_page_free_wakeup();
2747 * vm_page_dontneed()
2749 * Cache, deactivate, or do nothing as appropriate. This routine
2750 * is typically used by madvise() MADV_DONTNEED.
2752 * Generally speaking we want to move the page into the cache so
2753 * it gets reused quickly. However, this can result in a silly syndrome
2754 * due to the page recycling too quickly. Small objects will not be
2755 * fully cached. On the otherhand, if we move the page to the inactive
2756 * queue we wind up with a problem whereby very large objects
2757 * unnecessarily blow away our inactive and cache queues.
2759 * The solution is to move the pages based on a fixed weighting. We
2760 * either leave them alone, deactivate them, or move them to the cache,
2761 * where moving them to the cache has the highest weighting.
2762 * By forcing some pages into other queues we eventually force the
2763 * system to balance the queues, potentially recovering other unrelated
2764 * space from active. The idea is to not force this to happen too
2767 * The page must be busied.
2770 vm_page_dontneed(vm_page_t m)
2772 static int dnweight;
2779 * occassionally leave the page alone
2781 if ((dnw & 0x01F0) == 0 ||
2782 m->queue - m->pc == PQ_INACTIVE ||
2783 m->queue - m->pc == PQ_CACHE
2785 if (m->act_count >= ACT_INIT)
2791 * If vm_page_dontneed() is inactivating a page, it must clear
2792 * the referenced flag; otherwise the pagedaemon will see references
2793 * on the page in the inactive queue and reactivate it. Until the
2794 * page can move to the cache queue, madvise's job is not done.
2796 vm_page_flag_clear(m, PG_REFERENCED);
2797 pmap_clear_reference(m);
2800 vm_page_test_dirty(m);
2802 if (m->dirty || (dnw & 0x0070) == 0) {
2804 * Deactivate the page 3 times out of 32.
2809 * Cache the page 28 times out of every 32. Note that
2810 * the page is deactivated instead of cached, but placed
2811 * at the head of the queue instead of the tail.
2815 vm_page_spin_lock(m);
2816 _vm_page_deactivate_locked(m, head);
2817 vm_page_spin_unlock(m);
2821 * These routines manipulate the 'soft busy' count for a page. A soft busy
2822 * is almost like PG_BUSY except that it allows certain compatible operations
2823 * to occur on the page while it is busy. For example, a page undergoing a
2824 * write can still be mapped read-only.
2826 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2827 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2828 * busy bit is cleared.
2831 vm_page_io_start(vm_page_t m)
2833 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2834 atomic_add_char(&m->busy, 1);
2835 vm_page_flag_set(m, PG_SBUSY);
2839 vm_page_io_finish(vm_page_t m)
2841 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2842 atomic_subtract_char(&m->busy, 1);
2844 vm_page_flag_clear(m, PG_SBUSY);
2848 * Indicate that a clean VM page requires a filesystem commit and cannot
2849 * be reused. Used by tmpfs.
2852 vm_page_need_commit(vm_page_t m)
2854 vm_page_flag_set(m, PG_NEED_COMMIT);
2855 vm_object_set_writeable_dirty(m->object);
2859 vm_page_clear_commit(vm_page_t m)
2861 vm_page_flag_clear(m, PG_NEED_COMMIT);
2865 * Grab a page, blocking if it is busy and allocating a page if necessary.
2866 * A busy page is returned or NULL. The page may or may not be valid and
2867 * might not be on a queue (the caller is responsible for the disposition of
2870 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2871 * page will be zero'd and marked valid.
2873 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2874 * valid even if it already exists.
2876 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2877 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2878 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2880 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2881 * always returned if we had blocked.
2883 * This routine may not be called from an interrupt.
2885 * No other requirements.
2888 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2894 KKASSERT(allocflags &
2895 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2896 vm_object_hold_shared(object);
2898 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2900 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2901 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2906 } else if (m == NULL) {
2908 vm_object_upgrade(object);
2911 if (allocflags & VM_ALLOC_RETRY)
2912 allocflags |= VM_ALLOC_NULL_OK;
2913 m = vm_page_alloc(object, pindex,
2914 allocflags & ~VM_ALLOC_RETRY);
2918 if ((allocflags & VM_ALLOC_RETRY) == 0)
2927 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2929 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2930 * valid even if already valid.
2932 * NOTE! We have removed all of the PG_ZERO optimizations and also
2933 * removed the idle zeroing code. These optimizations actually
2934 * slow things down on modern cpus because the zerod area is
2935 * likely uncached, placing a memory-access burden on the
2936 * accesors taking the fault.
2938 * By always zeroing the page in-line with the fault, no
2939 * dynamic ram reads are needed and the caches are hot, ready
2940 * for userland to access the memory.
2942 if (m->valid == 0) {
2943 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2944 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2945 m->valid = VM_PAGE_BITS_ALL;
2947 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2948 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2949 m->valid = VM_PAGE_BITS_ALL;
2952 vm_object_drop(object);
2957 * Mapping function for valid bits or for dirty bits in
2958 * a page. May not block.
2960 * Inputs are required to range within a page.
2966 vm_page_bits(int base, int size)
2972 base + size <= PAGE_SIZE,
2973 ("vm_page_bits: illegal base/size %d/%d", base, size)
2976 if (size == 0) /* handle degenerate case */
2979 first_bit = base >> DEV_BSHIFT;
2980 last_bit = (base + size - 1) >> DEV_BSHIFT;
2982 return ((2 << last_bit) - (1 << first_bit));
2986 * Sets portions of a page valid and clean. The arguments are expected
2987 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2988 * of any partial chunks touched by the range. The invalid portion of
2989 * such chunks will be zero'd.
2991 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2992 * align base to DEV_BSIZE so as not to mark clean a partially
2993 * truncated device block. Otherwise the dirty page status might be
2996 * This routine may not block.
2998 * (base + size) must be less then or equal to PAGE_SIZE.
3001 _vm_page_zero_valid(vm_page_t m, int base, int size)
3006 if (size == 0) /* handle degenerate case */
3010 * If the base is not DEV_BSIZE aligned and the valid
3011 * bit is clear, we have to zero out a portion of the
3015 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
3016 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
3018 pmap_zero_page_area(
3026 * If the ending offset is not DEV_BSIZE aligned and the
3027 * valid bit is clear, we have to zero out a portion of
3031 endoff = base + size;
3033 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
3034 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
3036 pmap_zero_page_area(
3039 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
3045 * Set valid, clear dirty bits. If validating the entire
3046 * page we can safely clear the pmap modify bit. We also
3047 * use this opportunity to clear the PG_NOSYNC flag. If a process
3048 * takes a write fault on a MAP_NOSYNC memory area the flag will
3051 * We set valid bits inclusive of any overlap, but we can only
3052 * clear dirty bits for DEV_BSIZE chunks that are fully within
3055 * Page must be busied?
3056 * No other requirements.
3059 vm_page_set_valid(vm_page_t m, int base, int size)
3061 _vm_page_zero_valid(m, base, size);
3062 m->valid |= vm_page_bits(base, size);
3067 * Set valid bits and clear dirty bits.
3069 * Page must be busied by caller.
3071 * NOTE: This function does not clear the pmap modified bit.
3072 * Also note that e.g. NFS may use a byte-granular base
3075 * No other requirements.
3078 vm_page_set_validclean(vm_page_t m, int base, int size)
3082 _vm_page_zero_valid(m, base, size);
3083 pagebits = vm_page_bits(base, size);
3084 m->valid |= pagebits;
3085 m->dirty &= ~pagebits;
3086 if (base == 0 && size == PAGE_SIZE) {
3087 /*pmap_clear_modify(m);*/
3088 vm_page_flag_clear(m, PG_NOSYNC);
3093 * Set valid & dirty. Used by buwrite()
3095 * Page must be busied by caller.
3098 vm_page_set_validdirty(vm_page_t m, int base, int size)
3102 pagebits = vm_page_bits(base, size);
3103 m->valid |= pagebits;
3104 m->dirty |= pagebits;
3106 vm_object_set_writeable_dirty(m->object);
3112 * NOTE: This function does not clear the pmap modified bit.
3113 * Also note that e.g. NFS may use a byte-granular base
3116 * Page must be busied?
3117 * No other requirements.
3120 vm_page_clear_dirty(vm_page_t m, int base, int size)
3122 m->dirty &= ~vm_page_bits(base, size);
3123 if (base == 0 && size == PAGE_SIZE) {
3124 /*pmap_clear_modify(m);*/
3125 vm_page_flag_clear(m, PG_NOSYNC);
3130 * Make the page all-dirty.
3132 * Also make sure the related object and vnode reflect the fact that the
3133 * object may now contain a dirty page.
3135 * Page must be busied?
3136 * No other requirements.
3139 vm_page_dirty(vm_page_t m)
3142 int pqtype = m->queue - m->pc;
3144 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
3145 ("vm_page_dirty: page in free/cache queue!"));
3146 if (m->dirty != VM_PAGE_BITS_ALL) {
3147 m->dirty = VM_PAGE_BITS_ALL;
3149 vm_object_set_writeable_dirty(m->object);
3154 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3155 * valid and dirty bits for the effected areas are cleared.
3157 * Page must be busied?
3159 * No other requirements.
3162 vm_page_set_invalid(vm_page_t m, int base, int size)
3166 bits = vm_page_bits(base, size);
3169 m->object->generation++;
3173 * The kernel assumes that the invalid portions of a page contain
3174 * garbage, but such pages can be mapped into memory by user code.
3175 * When this occurs, we must zero out the non-valid portions of the
3176 * page so user code sees what it expects.
3178 * Pages are most often semi-valid when the end of a file is mapped
3179 * into memory and the file's size is not page aligned.
3181 * Page must be busied?
3182 * No other requirements.
3185 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3191 * Scan the valid bits looking for invalid sections that
3192 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
3193 * valid bit may be set ) have already been zerod by
3194 * vm_page_set_validclean().
3196 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3197 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3198 (m->valid & (1 << i))
3201 pmap_zero_page_area(
3204 (i - b) << DEV_BSHIFT
3212 * setvalid is TRUE when we can safely set the zero'd areas
3213 * as being valid. We can do this if there are no cache consistency
3214 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3217 m->valid = VM_PAGE_BITS_ALL;
3221 * Is a (partial) page valid? Note that the case where size == 0
3222 * will return FALSE in the degenerate case where the page is entirely
3223 * invalid, and TRUE otherwise.
3226 * No other requirements.
3229 vm_page_is_valid(vm_page_t m, int base, int size)
3231 int bits = vm_page_bits(base, size);
3233 if (m->valid && ((m->valid & bits) == bits))
3240 * update dirty bits from pmap/mmu. May not block.
3242 * Caller must hold the page busy
3245 vm_page_test_dirty(vm_page_t m)
3247 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
3253 * Register an action, associating it with its vm_page
3256 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
3258 struct vm_page_action_hash *hash;
3261 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
3262 hash = &action_hash[hv];
3264 lockmgr(&hash->lk, LK_EXCLUSIVE);
3265 vm_page_flag_set(action->m, PG_ACTIONLIST);
3266 action->event = event;
3267 LIST_INSERT_HEAD(&hash->list, action, entry);
3268 lockmgr(&hash->lk, LK_RELEASE);
3272 * Unregister an action, disassociating it from its related vm_page
3275 vm_page_unregister_action(vm_page_action_t action)
3277 struct vm_page_action_hash *hash;
3280 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
3281 hash = &action_hash[hv];
3282 lockmgr(&hash->lk, LK_EXCLUSIVE);
3283 if (action->event != VMEVENT_NONE) {
3284 action->event = VMEVENT_NONE;
3285 LIST_REMOVE(action, entry);
3287 if (LIST_EMPTY(&hash->list))
3288 vm_page_flag_clear(action->m, PG_ACTIONLIST);
3290 lockmgr(&hash->lk, LK_RELEASE);
3294 * Issue an event on a VM page. Corresponding action structures are
3295 * removed from the page's list and called.
3297 * If the vm_page has no more pending action events we clear its
3298 * PG_ACTIONLIST flag.
3301 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
3303 struct vm_page_action_hash *hash;
3304 struct vm_page_action *scan;
3305 struct vm_page_action *next;
3309 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
3310 hash = &action_hash[hv];
3313 lockmgr(&hash->lk, LK_EXCLUSIVE);
3314 LIST_FOREACH_MUTABLE(scan, &hash->list, entry, next) {
3316 if (scan->event == event) {
3317 scan->event = VMEVENT_NONE;
3318 LIST_REMOVE(scan, entry);
3319 scan->func(m, scan);
3327 vm_page_flag_clear(m, PG_ACTIONLIST);
3328 lockmgr(&hash->lk, LK_RELEASE);
3331 #include "opt_ddb.h"
3333 #include <sys/kernel.h>
3335 #include <ddb/ddb.h>
3337 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3339 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
3340 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
3341 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
3342 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
3343 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
3344 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
3345 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
3346 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
3347 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
3348 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
3351 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3354 db_printf("PQ_FREE:");
3355 for (i = 0; i < PQ_L2_SIZE; i++) {
3356 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
3360 db_printf("PQ_CACHE:");
3361 for(i = 0; i < PQ_L2_SIZE; i++) {
3362 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
3366 db_printf("PQ_ACTIVE:");
3367 for(i = 0; i < PQ_L2_SIZE; i++) {
3368 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3372 db_printf("PQ_INACTIVE:");
3373 for(i = 0; i < PQ_L2_SIZE; i++) {
3374 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);