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);
119 * Array of tailq lists
121 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT];
123 static volatile int vm_pages_waiting;
124 static struct alist vm_contig_alist;
125 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536];
126 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin");
128 static u_long vm_dma_reserved = 0;
129 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved);
130 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0,
131 "Memory reserved for DMA");
132 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD,
133 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA");
135 static int vm_contig_verbose = 0;
136 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose);
138 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
139 vm_pindex_t, pindex);
142 vm_page_queue_init(void)
146 for (i = 0; i < PQ_L2_SIZE; i++)
147 vm_page_queues[PQ_FREE+i].cnt_offset =
148 offsetof(struct vmstats, v_free_count);
149 for (i = 0; i < PQ_L2_SIZE; i++)
150 vm_page_queues[PQ_CACHE+i].cnt_offset =
151 offsetof(struct vmstats, v_cache_count);
152 for (i = 0; i < PQ_L2_SIZE; i++)
153 vm_page_queues[PQ_INACTIVE+i].cnt_offset =
154 offsetof(struct vmstats, v_inactive_count);
155 for (i = 0; i < PQ_L2_SIZE; i++)
156 vm_page_queues[PQ_ACTIVE+i].cnt_offset =
157 offsetof(struct vmstats, v_active_count);
158 for (i = 0; i < PQ_L2_SIZE; i++)
159 vm_page_queues[PQ_HOLD+i].cnt_offset =
160 offsetof(struct vmstats, v_active_count);
161 /* PQ_NONE has no queue */
163 for (i = 0; i < PQ_COUNT; i++) {
164 TAILQ_INIT(&vm_page_queues[i].pl);
165 spin_init(&vm_page_queues[i].spin, "vm_page_queue_init");
170 * note: place in initialized data section? Is this necessary?
173 int vm_page_array_size = 0;
174 vm_page_t vm_page_array = NULL;
175 vm_paddr_t vm_low_phys_reserved;
180 * Sets the page size, perhaps based upon the memory size.
181 * Must be called before any use of page-size dependent functions.
184 vm_set_page_size(void)
186 if (vmstats.v_page_size == 0)
187 vmstats.v_page_size = PAGE_SIZE;
188 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
189 panic("vm_set_page_size: page size not a power of two");
195 * Add a new page to the freelist for use by the system. New pages
196 * are added to both the head and tail of the associated free page
197 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
198 * requests pull 'recent' adds (higher physical addresses) first.
200 * Beware that the page zeroing daemon will also be running soon after
201 * boot, moving pages from the head to the tail of the PQ_FREE queues.
203 * Must be called in a critical section.
206 vm_add_new_page(vm_paddr_t pa)
208 struct vpgqueues *vpq;
211 m = PHYS_TO_VM_PAGE(pa);
214 m->pat_mode = PAT_WRITE_BACK;
215 m->pc = (pa >> PAGE_SHIFT);
218 * Twist for cpu localization in addition to page coloring, so
219 * different cpus selecting by m->queue get different page colors.
221 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE);
222 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE));
226 * Reserve a certain number of contiguous low memory pages for
227 * contigmalloc() to use.
229 if (pa < vm_low_phys_reserved) {
230 atomic_add_int(&vmstats.v_page_count, 1);
231 atomic_add_int(&vmstats.v_dma_pages, 1);
234 atomic_add_int(&vmstats.v_wire_count, 1);
235 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1);
242 m->queue = m->pc + PQ_FREE;
243 KKASSERT(m->dirty == 0);
245 atomic_add_int(&vmstats.v_page_count, 1);
246 atomic_add_int(&vmstats.v_free_count, 1);
247 vpq = &vm_page_queues[m->queue];
248 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
255 * Initializes the resident memory module.
257 * Preallocates memory for critical VM structures and arrays prior to
258 * kernel_map becoming available.
260 * Memory is allocated from (virtual2_start, virtual2_end) if available,
261 * otherwise memory is allocated from (virtual_start, virtual_end).
263 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
264 * large enough to hold vm_page_array & other structures for machines with
265 * large amounts of ram, so we want to use virtual2* when available.
268 vm_page_startup(void)
270 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start;
273 vm_paddr_t page_range;
279 vm_paddr_t biggestone, biggestsize;
286 vaddr = round_page(vaddr);
289 * Make sure ranges are page-aligned.
291 for (i = 0; phys_avail[i].phys_end; ++i) {
292 phys_avail[i].phys_beg = round_page64(phys_avail[i].phys_beg);
293 phys_avail[i].phys_end = trunc_page64(phys_avail[i].phys_end);
294 if (phys_avail[i].phys_end < phys_avail[i].phys_beg)
295 phys_avail[i].phys_end = phys_avail[i].phys_beg;
299 * Locate largest block
301 for (i = 0; phys_avail[i].phys_end; ++i) {
302 vm_paddr_t size = phys_avail[i].phys_end -
303 phys_avail[i].phys_beg;
305 if (size > biggestsize) {
311 --i; /* adjust to last entry for use down below */
313 end = phys_avail[biggestone].phys_end;
314 end = trunc_page(end);
317 * Initialize the queue headers for the free queue, the active queue
318 * and the inactive queue.
320 vm_page_queue_init();
322 #if !defined(_KERNEL_VIRTUAL)
324 * VKERNELs don't support minidumps and as such don't need
327 * Allocate a bitmap to indicate that a random physical page
328 * needs to be included in a minidump.
330 * The amd64 port needs this to indicate which direct map pages
331 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
333 * However, i386 still needs this workspace internally within the
334 * minidump code. In theory, they are not needed on i386, but are
335 * included should the sf_buf code decide to use them.
337 page_range = phys_avail[i].phys_end / PAGE_SIZE;
338 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
339 end -= vm_page_dump_size;
340 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
341 VM_PROT_READ | VM_PROT_WRITE);
342 bzero((void *)vm_page_dump, vm_page_dump_size);
345 * Compute the number of pages of memory that will be available for
346 * use (taking into account the overhead of a page structure per
349 first_page = phys_avail[0].phys_beg / PAGE_SIZE;
350 page_range = phys_avail[i].phys_end / PAGE_SIZE - first_page;
351 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
353 #ifndef _KERNEL_VIRTUAL
355 * (only applies to real kernels)
357 * Reserve a large amount of low memory for potential 32-bit DMA
358 * space allocations. Once device initialization is complete we
359 * release most of it, but keep (vm_dma_reserved) memory reserved
360 * for later use. Typically for X / graphics. Through trial and
361 * error we find that GPUs usually requires ~60-100MB or so.
363 * By default, 128M is left in reserve on machines with 2G+ of ram.
365 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT;
366 if (vm_low_phys_reserved > total / 4)
367 vm_low_phys_reserved = total / 4;
368 if (vm_dma_reserved == 0) {
369 vm_dma_reserved = 128 * 1024 * 1024; /* 128MB */
370 if (vm_dma_reserved > total / 16)
371 vm_dma_reserved = total / 16;
374 alist_init(&vm_contig_alist, 65536, vm_contig_ameta,
375 ALIST_RECORDS_65536);
378 * Initialize the mem entry structures now, and put them in the free
381 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
382 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE);
383 vm_page_array = (vm_page_t)mapped;
385 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
387 * since pmap_map on amd64 returns stuff out of a direct-map region,
388 * we have to manually add these pages to the minidump tracking so
389 * that they can be dumped, including the vm_page_array.
392 pa < phys_avail[biggestone].phys_end;
399 * Clear all of the page structures, run basic initialization so
400 * PHYS_TO_VM_PAGE() operates properly even on pages not in the
403 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
404 vm_page_array_size = page_range;
406 m = &vm_page_array[0];
407 pa = ptoa(first_page);
408 for (i = 0; i < page_range; ++i) {
409 spin_init(&m->spin, "vm_page");
416 * Construct the free queue(s) in ascending order (by physical
417 * address) so that the first 16MB of physical memory is allocated
418 * last rather than first. On large-memory machines, this avoids
419 * the exhaustion of low physical memory before isa_dmainit has run.
421 vmstats.v_page_count = 0;
422 vmstats.v_free_count = 0;
423 for (i = 0; phys_avail[i].phys_end && npages > 0; ++i) {
424 pa = phys_avail[i].phys_beg;
428 last_pa = phys_avail[i].phys_end;
429 while (pa < last_pa && npages-- > 0) {
435 virtual2_start = vaddr;
437 virtual_start = vaddr;
438 mycpu->gd_vmstats = vmstats;
442 * Reorganize VM pages based on numa data. May be called as many times as
443 * necessary. Will reorganize the vm_page_t page color and related queue(s)
444 * to allow vm_page_alloc() to choose pages based on socket affinity.
446 * NOTE: This function is only called while we are still in UP mode, so
447 * we only need a critical section to protect the queues (which
448 * saves a lot of time, there are likely a ton of pages).
451 vm_numa_organize(vm_paddr_t ran_beg, vm_paddr_t bytes, int physid)
456 struct vpgqueues *vpq;
464 * Check if no physical information, or there was only one socket
465 * (so don't waste time doing nothing!).
467 if (cpu_topology_phys_ids <= 1 ||
468 cpu_topology_core_ids == 0) {
473 * Setup for our iteration. Note that ACPI may iterate CPU
474 * sockets starting at 0 or 1 or some other number. The
475 * cpu_topology code mod's it against the socket count.
477 ran_end = ran_beg + bytes;
478 physid %= cpu_topology_phys_ids;
480 socket_mod = PQ_L2_SIZE / cpu_topology_phys_ids;
481 socket_value = physid * socket_mod;
482 mend = &vm_page_array[vm_page_array_size];
487 * Adjust vm_page->pc and requeue all affected pages. The
488 * allocator will then be able to localize memory allocations
491 for (i = 0; phys_avail[i].phys_end; ++i) {
492 scan_beg = phys_avail[i].phys_beg;
493 scan_end = phys_avail[i].phys_end;
494 if (scan_end <= ran_beg)
496 if (scan_beg >= ran_end)
498 if (scan_beg < ran_beg)
500 if (scan_end > ran_end)
502 if (atop(scan_end) > first_page + vm_page_array_size)
503 scan_end = ptoa(first_page + vm_page_array_size);
505 m = PHYS_TO_VM_PAGE(scan_beg);
506 while (scan_beg < scan_end) {
508 if (m->queue != PQ_NONE) {
509 vpq = &vm_page_queues[m->queue];
510 TAILQ_REMOVE(&vpq->pl, m, pageq);
512 /* queue doesn't change, no need to adj cnt */
515 m->pc += socket_value;
518 vpq = &vm_page_queues[m->queue];
519 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
521 /* queue doesn't change, no need to adj cnt */
524 m->pc += socket_value;
527 scan_beg += PAGE_SIZE;
535 * We tended to reserve a ton of memory for contigmalloc(). Now that most
536 * drivers have initialized we want to return most the remaining free
537 * reserve back to the VM page queues so they can be used for normal
540 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool.
543 vm_page_startup_finish(void *dummy __unused)
552 spin_lock(&vm_contig_spin);
554 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
555 if (bfree <= vm_dma_reserved / PAGE_SIZE)
561 * Figure out how much of the initial reserve we have to
562 * free in order to reach our target.
564 bfree -= vm_dma_reserved / PAGE_SIZE;
566 blk += count - bfree;
571 * Calculate the nearest power of 2 <= count.
573 for (xcount = 1; xcount <= count; xcount <<= 1)
576 blk += count - xcount;
580 * Allocate the pages from the alist, then free them to
581 * the normal VM page queues.
583 * Pages allocated from the alist are wired. We have to
584 * busy, unwire, and free them. We must also adjust
585 * vm_low_phys_reserved before freeing any pages to prevent
588 rblk = alist_alloc(&vm_contig_alist, blk, count);
590 kprintf("vm_page_startup_finish: Unable to return "
591 "dma space @0x%08x/%d -> 0x%08x\n",
595 atomic_add_int(&vmstats.v_dma_pages, -count);
596 spin_unlock(&vm_contig_spin);
598 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
599 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m);
601 vm_page_busy_wait(m, FALSE, "cpgfr");
602 vm_page_unwire(m, 0);
607 spin_lock(&vm_contig_spin);
609 spin_unlock(&vm_contig_spin);
612 * Print out how much DMA space drivers have already allocated and
613 * how much is left over.
615 kprintf("DMA space used: %jdk, remaining available: %jdk\n",
616 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) *
618 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024));
620 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY,
621 vm_page_startup_finish, NULL);
625 * Scan comparison function for Red-Black tree scans. An inclusive
626 * (start,end) is expected. Other fields are not used.
629 rb_vm_page_scancmp(struct vm_page *p, void *data)
631 struct rb_vm_page_scan_info *info = data;
633 if (p->pindex < info->start_pindex)
635 if (p->pindex > info->end_pindex)
641 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
643 if (p1->pindex < p2->pindex)
645 if (p1->pindex > p2->pindex)
651 vm_page_init(vm_page_t m)
653 /* do nothing for now. Called from pmap_page_init() */
657 * Each page queue has its own spin lock, which is fairly optimal for
658 * allocating and freeing pages at least.
660 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
661 * queue spinlock via this function. Also note that m->queue cannot change
662 * unless both the page and queue are locked.
666 _vm_page_queue_spin_lock(vm_page_t m)
671 if (queue != PQ_NONE) {
672 spin_lock(&vm_page_queues[queue].spin);
673 KKASSERT(queue == m->queue);
679 _vm_page_queue_spin_unlock(vm_page_t m)
685 if (queue != PQ_NONE)
686 spin_unlock(&vm_page_queues[queue].spin);
691 _vm_page_queues_spin_lock(u_short queue)
694 if (queue != PQ_NONE)
695 spin_lock(&vm_page_queues[queue].spin);
701 _vm_page_queues_spin_unlock(u_short queue)
704 if (queue != PQ_NONE)
705 spin_unlock(&vm_page_queues[queue].spin);
709 vm_page_queue_spin_lock(vm_page_t m)
711 _vm_page_queue_spin_lock(m);
715 vm_page_queues_spin_lock(u_short queue)
717 _vm_page_queues_spin_lock(queue);
721 vm_page_queue_spin_unlock(vm_page_t m)
723 _vm_page_queue_spin_unlock(m);
727 vm_page_queues_spin_unlock(u_short queue)
729 _vm_page_queues_spin_unlock(queue);
733 * This locks the specified vm_page and its queue in the proper order
734 * (page first, then queue). The queue may change so the caller must
739 _vm_page_and_queue_spin_lock(vm_page_t m)
741 vm_page_spin_lock(m);
742 _vm_page_queue_spin_lock(m);
747 _vm_page_and_queue_spin_unlock(vm_page_t m)
749 _vm_page_queues_spin_unlock(m->queue);
750 vm_page_spin_unlock(m);
754 vm_page_and_queue_spin_unlock(vm_page_t m)
756 _vm_page_and_queue_spin_unlock(m);
760 vm_page_and_queue_spin_lock(vm_page_t m)
762 _vm_page_and_queue_spin_lock(m);
766 * Helper function removes vm_page from its current queue.
767 * Returns the base queue the page used to be on.
769 * The vm_page and the queue must be spinlocked.
770 * This function will unlock the queue but leave the page spinlocked.
772 static __inline u_short
773 _vm_page_rem_queue_spinlocked(vm_page_t m)
775 struct vpgqueues *pq;
781 if (queue != PQ_NONE) {
782 pq = &vm_page_queues[queue];
783 TAILQ_REMOVE(&pq->pl, m, pageq);
786 * Adjust our pcpu stats. In order for the nominal low-memory
787 * algorithms to work properly we don't let any pcpu stat get
788 * too negative before we force it to be rolled-up into the
789 * global stats. Otherwise our pageout and vm_wait tests
792 * The idea here is to reduce unnecessary SMP cache
793 * mastership changes in the global vmstats, which can be
794 * particularly bad in multi-socket systems.
796 cnt = (int *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
797 atomic_add_int(cnt, -1);
798 if (*cnt < -VMMETER_SLOP_COUNT) {
799 u_int copy = atomic_swap_int(cnt, 0);
800 cnt = (int *)((char *)&vmstats + pq->cnt_offset);
801 atomic_add_int(cnt, copy);
802 cnt = (int *)((char *)&mycpu->gd_vmstats +
804 atomic_add_int(cnt, copy);
810 vm_page_queues_spin_unlock(oqueue); /* intended */
816 * Helper function places the vm_page on the specified queue. Generally
817 * speaking only PQ_FREE pages are placed at the head, to allow them to
818 * be allocated sooner rather than later on the assumption that they
821 * The vm_page must be spinlocked.
822 * This function will return with both the page and the queue locked.
825 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
827 struct vpgqueues *pq;
830 KKASSERT(m->queue == PQ_NONE);
832 if (queue != PQ_NONE) {
833 vm_page_queues_spin_lock(queue);
834 pq = &vm_page_queues[queue];
838 * Adjust our pcpu stats. If a system entity really needs
839 * to incorporate the count it will call vmstats_rollup()
840 * to roll it all up into the global vmstats strufture.
842 cnt = (int *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
843 atomic_add_int(cnt, 1);
846 * PQ_FREE is always handled LIFO style to try to provide
847 * cache-hot pages to programs.
850 if (queue - m->pc == PQ_FREE) {
851 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
853 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
855 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
857 /* leave the queue spinlocked */
862 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
863 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
864 * did not. Only one sleep call will be made before returning.
866 * This function does NOT busy the page and on return the page is not
867 * guaranteed to be available.
870 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
878 if ((flags & PG_BUSY) == 0 &&
879 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
882 tsleep_interlock(m, 0);
883 if (atomic_cmpset_int(&m->flags, flags,
884 flags | PG_WANTED | PG_REFERENCED)) {
885 tsleep(m, PINTERLOCKED, msg, 0);
892 * This calculates and returns a page color given an optional VM object and
893 * either a pindex or an iterator. We attempt to return a cpu-localized
894 * pg_color that is still roughly 16-way set-associative. The CPU topology
895 * is used if it was probed.
897 * The caller may use the returned value to index into e.g. PQ_FREE when
898 * allocating a page in order to nominally obtain pages that are hopefully
899 * already localized to the requesting cpu. This function is not able to
900 * provide any sort of guarantee of this, but does its best to improve
901 * hardware cache management performance.
903 * WARNING! The caller must mask the returned value with PQ_L2_MASK.
906 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex)
913 phys_id = get_cpu_phys_id(cpuid);
914 core_id = get_cpu_core_id(cpuid);
915 object_pg_color = object ? object->pg_color : 0;
917 if (cpu_topology_phys_ids && cpu_topology_core_ids) {
921 * Break us down by socket and cpu
923 pg_color = phys_id * PQ_L2_SIZE / cpu_topology_phys_ids;
924 pg_color += core_id * PQ_L2_SIZE /
925 (cpu_topology_core_ids * cpu_topology_phys_ids);
928 * Calculate remaining component for object/queue color
930 grpsize = PQ_L2_SIZE / (cpu_topology_core_ids *
931 cpu_topology_phys_ids);
933 pg_color += (pindex + object_pg_color) % grpsize;
938 /* 3->9, 4->8, 5->10, 6->12, 7->14 */
943 pg_color += (pindex + object_pg_color) % grpsize;
947 * Unknown topology, distribute things evenly.
949 pg_color = cpuid * PQ_L2_SIZE / ncpus;
950 pg_color += pindex + object_pg_color;
952 return (pg_color & PQ_L2_MASK);
956 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
957 * also wait for m->busy to become 0 before setting PG_BUSY.
960 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
961 int also_m_busy, const char *msg
969 if (flags & PG_BUSY) {
970 tsleep_interlock(m, 0);
971 if (atomic_cmpset_int(&m->flags, flags,
972 flags | PG_WANTED | PG_REFERENCED)) {
973 tsleep(m, PINTERLOCKED, msg, 0);
975 } else if (also_m_busy && (flags & PG_SBUSY)) {
976 tsleep_interlock(m, 0);
977 if (atomic_cmpset_int(&m->flags, flags,
978 flags | PG_WANTED | PG_REFERENCED)) {
979 tsleep(m, PINTERLOCKED, msg, 0);
982 if (atomic_cmpset_int(&m->flags, flags,
986 m->busy_line = lineno;
995 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
998 * Returns non-zero on failure.
1001 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
1009 if (flags & PG_BUSY)
1011 if (also_m_busy && (flags & PG_SBUSY))
1013 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1014 #ifdef VM_PAGE_DEBUG
1015 m->busy_func = func;
1016 m->busy_line = lineno;
1024 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
1025 * that a wakeup() should be performed.
1027 * The vm_page must be spinlocked and will remain spinlocked on return.
1028 * The related queue must NOT be spinlocked (which could deadlock us).
1034 _vm_page_wakeup(vm_page_t m)
1041 if (atomic_cmpset_int(&m->flags, flags,
1042 flags & ~(PG_BUSY | PG_WANTED))) {
1046 return(flags & PG_WANTED);
1050 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
1051 * is typically the last call you make on a page before moving onto
1055 vm_page_wakeup(vm_page_t m)
1057 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
1058 vm_page_spin_lock(m);
1059 if (_vm_page_wakeup(m)) {
1060 vm_page_spin_unlock(m);
1063 vm_page_spin_unlock(m);
1068 * Holding a page keeps it from being reused. Other parts of the system
1069 * can still disassociate the page from its current object and free it, or
1070 * perform read or write I/O on it and/or otherwise manipulate the page,
1071 * but if the page is held the VM system will leave the page and its data
1072 * intact and not reuse the page for other purposes until the last hold
1073 * reference is released. (see vm_page_wire() if you want to prevent the
1074 * page from being disassociated from its object too).
1076 * The caller must still validate the contents of the page and, if necessary,
1077 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
1078 * before manipulating the page.
1080 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
1083 vm_page_hold(vm_page_t m)
1085 vm_page_spin_lock(m);
1086 atomic_add_int(&m->hold_count, 1);
1087 if (m->queue - m->pc == PQ_FREE) {
1088 _vm_page_queue_spin_lock(m);
1089 _vm_page_rem_queue_spinlocked(m);
1090 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
1091 _vm_page_queue_spin_unlock(m);
1093 vm_page_spin_unlock(m);
1097 * The opposite of vm_page_hold(). If the page is on the HOLD queue
1098 * it was freed while held and must be moved back to the FREE queue.
1101 vm_page_unhold(vm_page_t m)
1103 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE,
1104 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)",
1105 m, m->hold_count, m->queue - m->pc));
1106 vm_page_spin_lock(m);
1107 atomic_add_int(&m->hold_count, -1);
1108 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
1109 _vm_page_queue_spin_lock(m);
1110 _vm_page_rem_queue_spinlocked(m);
1111 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1);
1112 _vm_page_queue_spin_unlock(m);
1114 vm_page_spin_unlock(m);
1120 * Create a fictitious page with the specified physical address and
1121 * memory attribute. The memory attribute is the only the machine-
1122 * dependent aspect of a fictitious page that must be initialized.
1126 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1129 if ((m->flags & PG_FICTITIOUS) != 0) {
1131 * The page's memattr might have changed since the
1132 * previous initialization. Update the pmap to the
1137 m->phys_addr = paddr;
1139 /* Fictitious pages don't use "segind". */
1140 /* Fictitious pages don't use "order" or "pool". */
1141 m->flags = PG_FICTITIOUS | PG_UNMANAGED | PG_BUSY;
1143 spin_init(&m->spin, "fake_page");
1146 pmap_page_set_memattr(m, memattr);
1150 * Inserts the given vm_page into the object and object list.
1152 * The pagetables are not updated but will presumably fault the page
1153 * in if necessary, or if a kernel page the caller will at some point
1154 * enter the page into the kernel's pmap. We are not allowed to block
1155 * here so we *can't* do this anyway.
1157 * This routine may not block.
1158 * This routine must be called with the vm_object held.
1159 * This routine must be called with a critical section held.
1161 * This routine returns TRUE if the page was inserted into the object
1162 * successfully, and FALSE if the page already exists in the object.
1165 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1167 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object));
1168 if (m->object != NULL)
1169 panic("vm_page_insert: already inserted");
1171 atomic_add_int(&object->generation, 1);
1174 * Record the object/offset pair in this page and add the
1175 * pv_list_count of the page to the object.
1177 * The vm_page spin lock is required for interactions with the pmap.
1179 vm_page_spin_lock(m);
1182 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
1185 vm_page_spin_unlock(m);
1188 ++object->resident_page_count;
1189 ++mycpu->gd_vmtotal.t_rm;
1190 vm_page_spin_unlock(m);
1193 * Since we are inserting a new and possibly dirty page,
1194 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
1196 if ((m->valid & m->dirty) ||
1197 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT)))
1198 vm_object_set_writeable_dirty(object);
1201 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
1203 swap_pager_page_inserted(m);
1208 * Removes the given vm_page_t from the (object,index) table
1210 * The underlying pmap entry (if any) is NOT removed here.
1211 * This routine may not block.
1213 * The page must be BUSY and will remain BUSY on return.
1214 * No other requirements.
1216 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
1220 vm_page_remove(vm_page_t m)
1224 if (m->object == NULL) {
1228 if ((m->flags & PG_BUSY) == 0)
1229 panic("vm_page_remove: page not busy");
1233 vm_object_hold(object);
1236 * Remove the page from the object and update the object.
1238 * The vm_page spin lock is required for interactions with the pmap.
1240 vm_page_spin_lock(m);
1241 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
1242 --object->resident_page_count;
1243 --mycpu->gd_vmtotal.t_rm;
1245 atomic_add_int(&object->generation, 1);
1246 vm_page_spin_unlock(m);
1248 vm_object_drop(object);
1252 * Locate and return the page at (object, pindex), or NULL if the
1253 * page could not be found.
1255 * The caller must hold the vm_object token.
1258 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1263 * Search the hash table for this object/offset pair
1265 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1266 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1267 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1272 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1274 int also_m_busy, const char *msg
1280 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1281 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1283 KKASSERT(m->object == object && m->pindex == pindex);
1286 if (flags & PG_BUSY) {
1287 tsleep_interlock(m, 0);
1288 if (atomic_cmpset_int(&m->flags, flags,
1289 flags | PG_WANTED | PG_REFERENCED)) {
1290 tsleep(m, PINTERLOCKED, msg, 0);
1291 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1294 } else if (also_m_busy && (flags & PG_SBUSY)) {
1295 tsleep_interlock(m, 0);
1296 if (atomic_cmpset_int(&m->flags, flags,
1297 flags | PG_WANTED | PG_REFERENCED)) {
1298 tsleep(m, PINTERLOCKED, msg, 0);
1299 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1302 } else if (atomic_cmpset_int(&m->flags, flags,
1304 #ifdef VM_PAGE_DEBUG
1305 m->busy_func = func;
1306 m->busy_line = lineno;
1315 * Attempt to lookup and busy a page.
1317 * Returns NULL if the page could not be found
1319 * Returns a vm_page and error == TRUE if the page exists but could not
1322 * Returns a vm_page and error == FALSE on success.
1325 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1327 int also_m_busy, int *errorp
1333 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1334 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1337 KKASSERT(m->object == object && m->pindex == pindex);
1340 if (flags & PG_BUSY) {
1344 if (also_m_busy && (flags & PG_SBUSY)) {
1348 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1349 #ifdef VM_PAGE_DEBUG
1350 m->busy_func = func;
1351 m->busy_line = lineno;
1360 * Caller must hold the related vm_object
1363 vm_page_next(vm_page_t m)
1367 next = vm_page_rb_tree_RB_NEXT(m);
1368 if (next && next->pindex != m->pindex + 1)
1376 * Move the given vm_page from its current object to the specified
1377 * target object/offset. The page must be busy and will remain so
1380 * new_object must be held.
1381 * This routine might block. XXX ?
1383 * NOTE: Swap associated with the page must be invalidated by the move. We
1384 * have to do this for several reasons: (1) we aren't freeing the
1385 * page, (2) we are dirtying the page, (3) the VM system is probably
1386 * moving the page from object A to B, and will then later move
1387 * the backing store from A to B and we can't have a conflict.
1389 * NOTE: We *always* dirty the page. It is necessary both for the
1390 * fact that we moved it, and because we may be invalidating
1391 * swap. If the page is on the cache, we have to deactivate it
1392 * or vm_page_dirty() will panic. Dirty pages are not allowed
1396 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1398 KKASSERT(m->flags & PG_BUSY);
1399 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object));
1401 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object));
1404 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1405 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1406 new_object, new_pindex);
1408 if (m->queue - m->pc == PQ_CACHE)
1409 vm_page_deactivate(m);
1414 * vm_page_unqueue() without any wakeup. This routine is used when a page
1415 * is to remain BUSYied by the caller.
1417 * This routine may not block.
1420 vm_page_unqueue_nowakeup(vm_page_t m)
1422 vm_page_and_queue_spin_lock(m);
1423 (void)_vm_page_rem_queue_spinlocked(m);
1424 vm_page_spin_unlock(m);
1428 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1431 * This routine may not block.
1434 vm_page_unqueue(vm_page_t m)
1438 vm_page_and_queue_spin_lock(m);
1439 queue = _vm_page_rem_queue_spinlocked(m);
1440 if (queue == PQ_FREE || queue == PQ_CACHE) {
1441 vm_page_spin_unlock(m);
1442 pagedaemon_wakeup();
1444 vm_page_spin_unlock(m);
1449 * vm_page_list_find()
1451 * Find a page on the specified queue with color optimization.
1453 * The page coloring optimization attempts to locate a page that does
1454 * not overload other nearby pages in the object in the cpu's L1 or L2
1455 * caches. We need this optimization because cpu caches tend to be
1456 * physical caches, while object spaces tend to be virtual.
1458 * The page coloring optimization also, very importantly, tries to localize
1459 * memory to cpus and physical sockets.
1461 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1462 * and the algorithm is adjusted to localize allocations on a per-core basis.
1463 * This is done by 'twisting' the colors.
1465 * The page is returned spinlocked and removed from its queue (it will
1466 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1467 * is responsible for dealing with the busy-page case (usually by
1468 * deactivating the page and looping).
1470 * NOTE: This routine is carefully inlined. A non-inlined version
1471 * is available for outside callers but the only critical path is
1472 * from within this source file.
1474 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1475 * represent stable storage, allowing us to order our locks vm_page
1476 * first, then queue.
1480 _vm_page_list_find(int basequeue, int index)
1485 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1487 m = _vm_page_list_find2(basequeue, index);
1490 vm_page_and_queue_spin_lock(m);
1491 if (m->queue == basequeue + index) {
1492 _vm_page_rem_queue_spinlocked(m);
1493 /* vm_page_t spin held, no queue spin */
1496 vm_page_and_queue_spin_unlock(m);
1502 * If we could not find the page in the desired queue try to find it in
1506 _vm_page_list_find2(int basequeue, int index)
1508 struct vpgqueues *pq;
1510 int pqmask = PQ_SET_ASSOC_MASK >> 1;
1514 index &= PQ_L2_MASK;
1515 pq = &vm_page_queues[basequeue];
1518 * Run local sets of 16, 32, 64, 128, and the whole queue if all
1519 * else fails (PQ_L2_MASK which is 255).
1522 pqmask = (pqmask << 1) | 1;
1523 for (i = 0; i <= pqmask; ++i) {
1524 pqi = (index & ~pqmask) | ((index + i) & pqmask);
1525 m = TAILQ_FIRST(&pq[pqi].pl);
1527 _vm_page_and_queue_spin_lock(m);
1528 if (m->queue == basequeue + pqi) {
1529 _vm_page_rem_queue_spinlocked(m);
1532 _vm_page_and_queue_spin_unlock(m);
1537 } while (pqmask != PQ_L2_MASK);
1543 * Returns a vm_page candidate for allocation. The page is not busied so
1544 * it can move around. The caller must busy the page (and typically
1545 * deactivate it if it cannot be busied!)
1547 * Returns a spinlocked vm_page that has been removed from its queue.
1550 vm_page_list_find(int basequeue, int index)
1552 return(_vm_page_list_find(basequeue, index));
1556 * Find a page on the cache queue with color optimization, remove it
1557 * from the queue, and busy it. The returned page will not be spinlocked.
1559 * A candidate failure will be deactivated. Candidates can fail due to
1560 * being busied by someone else, in which case they will be deactivated.
1562 * This routine may not block.
1566 vm_page_select_cache(u_short pg_color)
1571 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK);
1575 * (m) has been removed from its queue and spinlocked
1577 if (vm_page_busy_try(m, TRUE)) {
1578 _vm_page_deactivate_locked(m, 0);
1579 vm_page_spin_unlock(m);
1582 * We successfully busied the page
1584 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 &&
1585 m->hold_count == 0 &&
1586 m->wire_count == 0 &&
1587 (m->dirty & m->valid) == 0) {
1588 vm_page_spin_unlock(m);
1589 pagedaemon_wakeup();
1594 * The page cannot be recycled, deactivate it.
1596 _vm_page_deactivate_locked(m, 0);
1597 if (_vm_page_wakeup(m)) {
1598 vm_page_spin_unlock(m);
1601 vm_page_spin_unlock(m);
1609 * Find a free page. We attempt to inline the nominal case and fall back
1610 * to _vm_page_select_free() otherwise. A busied page is removed from
1611 * the queue and returned.
1613 * This routine may not block.
1615 static __inline vm_page_t
1616 vm_page_select_free(u_short pg_color)
1621 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK);
1624 if (vm_page_busy_try(m, TRUE)) {
1626 * Various mechanisms such as a pmap_collect can
1627 * result in a busy page on the free queue. We
1628 * have to move the page out of the way so we can
1629 * retry the allocation. If the other thread is not
1630 * allocating the page then m->valid will remain 0 and
1631 * the pageout daemon will free the page later on.
1633 * Since we could not busy the page, however, we
1634 * cannot make assumptions as to whether the page
1635 * will be allocated by the other thread or not,
1636 * so all we can do is deactivate it to move it out
1637 * of the way. In particular, if the other thread
1638 * wires the page it may wind up on the inactive
1639 * queue and the pageout daemon will have to deal
1640 * with that case too.
1642 _vm_page_deactivate_locked(m, 0);
1643 vm_page_spin_unlock(m);
1646 * Theoretically if we are able to busy the page
1647 * atomic with the queue removal (using the vm_page
1648 * lock) nobody else should be able to mess with the
1651 KKASSERT((m->flags & (PG_UNMANAGED |
1652 PG_NEED_COMMIT)) == 0);
1653 KASSERT(m->hold_count == 0, ("m->hold_count is not zero "
1654 "pg %p q=%d flags=%08x hold=%d wire=%d",
1655 m, m->queue, m->flags, m->hold_count, m->wire_count));
1656 KKASSERT(m->wire_count == 0);
1657 vm_page_spin_unlock(m);
1658 pagedaemon_wakeup();
1660 /* return busied and removed page */
1670 * Allocate and return a memory cell associated with this VM object/offset
1671 * pair. If object is NULL an unassociated page will be allocated.
1673 * The returned page will be busied and removed from its queues. This
1674 * routine can block and may return NULL if a race occurs and the page
1675 * is found to already exist at the specified (object, pindex).
1677 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1678 * VM_ALLOC_QUICK like normal but cannot use cache
1679 * VM_ALLOC_SYSTEM greater free drain
1680 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1681 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1682 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1683 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1684 * (see vm_page_grab())
1685 * VM_ALLOC_USE_GD ok to use per-gd cache
1687 * VM_ALLOC_CPU(n) allocate using specified cpu localization
1689 * The object must be held if not NULL
1690 * This routine may not block
1692 * Additional special handling is required when called from an interrupt
1693 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1697 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1707 * Special per-cpu free VM page cache. The pages are pre-busied
1708 * and pre-zerod for us.
1710 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1712 if (gd->gd_vmpg_count) {
1713 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1725 * CPU localization algorithm. Break the page queues up by physical
1726 * id and core id (note that two cpu threads will have the same core
1727 * id, and core_id != gd_cpuid).
1729 * This is nowhere near perfect, for example the last pindex in a
1730 * subgroup will overflow into the next cpu or package. But this
1731 * should get us good page reuse locality in heavy mixed loads.
1733 * (may be executed before the APs are started, so other GDs might
1736 if (page_req & VM_ALLOC_CPU_SPEC)
1737 cpuid_local = VM_ALLOC_GETCPU(page_req);
1739 cpuid_local = mycpu->gd_cpuid;
1741 pg_color = vm_get_pg_color(cpuid_local, object, pindex);
1744 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1745 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1748 * Certain system threads (pageout daemon, buf_daemon's) are
1749 * allowed to eat deeper into the free page list.
1751 if (curthread->td_flags & TDF_SYSTHREAD)
1752 page_req |= VM_ALLOC_SYSTEM;
1755 * Impose various limitations. Note that the v_free_reserved test
1756 * must match the opposite of vm_page_count_target() to avoid
1757 * livelocks, be careful.
1761 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved ||
1762 ((page_req & VM_ALLOC_INTERRUPT) &&
1763 gd->gd_vmstats.v_free_count > 0) ||
1764 ((page_req & VM_ALLOC_SYSTEM) &&
1765 gd->gd_vmstats.v_cache_count == 0 &&
1766 gd->gd_vmstats.v_free_count >
1767 gd->gd_vmstats.v_interrupt_free_min)
1770 * The free queue has sufficient free pages to take one out.
1772 m = vm_page_select_free(pg_color);
1773 } else if (page_req & VM_ALLOC_NORMAL) {
1775 * Allocatable from the cache (non-interrupt only). On
1776 * success, we must free the page and try again, thus
1777 * ensuring that vmstats.v_*_free_min counters are replenished.
1780 if (curthread->td_preempted) {
1781 kprintf("vm_page_alloc(): warning, attempt to allocate"
1782 " cache page from preempting interrupt\n");
1785 m = vm_page_select_cache(pg_color);
1788 m = vm_page_select_cache(pg_color);
1791 * On success move the page into the free queue and loop.
1793 * Only do this if we can safely acquire the vm_object lock,
1794 * because this is effectively a random page and the caller
1795 * might be holding the lock shared, we don't want to
1799 KASSERT(m->dirty == 0,
1800 ("Found dirty cache page %p", m));
1801 if ((obj = m->object) != NULL) {
1802 if (vm_object_hold_try(obj)) {
1803 vm_page_protect(m, VM_PROT_NONE);
1805 /* m->object NULL here */
1806 vm_object_drop(obj);
1808 vm_page_deactivate(m);
1812 vm_page_protect(m, VM_PROT_NONE);
1819 * On failure return NULL
1821 atomic_add_int(&vm_pageout_deficit, 1);
1822 pagedaemon_wakeup();
1826 * No pages available, wakeup the pageout daemon and give up.
1828 atomic_add_int(&vm_pageout_deficit, 1);
1829 pagedaemon_wakeup();
1834 * v_free_count can race so loop if we don't find the expected
1843 * Good page found. The page has already been busied for us and
1844 * removed from its queues.
1846 KASSERT(m->dirty == 0,
1847 ("vm_page_alloc: free/cache page %p was dirty", m));
1848 KKASSERT(m->queue == PQ_NONE);
1854 * Initialize the structure, inheriting some flags but clearing
1855 * all the rest. The page has already been busied for us.
1857 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK);
1859 KKASSERT(m->wire_count == 0);
1860 KKASSERT(m->busy == 0);
1865 * Caller must be holding the object lock (asserted by
1866 * vm_page_insert()).
1868 * NOTE: Inserting a page here does not insert it into any pmaps
1869 * (which could cause us to block allocating memory).
1871 * NOTE: If no object an unassociated page is allocated, m->pindex
1872 * can be used by the caller for any purpose.
1875 if (vm_page_insert(m, object, pindex) == FALSE) {
1877 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1878 panic("PAGE RACE %p[%ld]/%p",
1879 object, (long)pindex, m);
1887 * Don't wakeup too often - wakeup the pageout daemon when
1888 * we would be nearly out of memory.
1890 pagedaemon_wakeup();
1893 * A PG_BUSY page is returned.
1899 * Returns number of pages available in our DMA memory reserve
1900 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf)
1903 vm_contig_avail_pages(void)
1908 spin_lock(&vm_contig_spin);
1909 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
1910 spin_unlock(&vm_contig_spin);
1916 * Attempt to allocate contiguous physical memory with the specified
1920 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
1921 unsigned long alignment, unsigned long boundary,
1922 unsigned long size, vm_memattr_t memattr)
1928 alignment >>= PAGE_SHIFT;
1931 boundary >>= PAGE_SHIFT;
1934 size = (size + PAGE_MASK) >> PAGE_SHIFT;
1936 spin_lock(&vm_contig_spin);
1937 blk = alist_alloc(&vm_contig_alist, 0, size);
1938 if (blk == ALIST_BLOCK_NONE) {
1939 spin_unlock(&vm_contig_spin);
1941 kprintf("vm_page_alloc_contig: %ldk nospace\n",
1942 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1946 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
1947 alist_free(&vm_contig_alist, blk, size);
1948 spin_unlock(&vm_contig_spin);
1950 kprintf("vm_page_alloc_contig: %ldk high "
1952 (size + PAGE_MASK) * (PAGE_SIZE / 1024),
1957 spin_unlock(&vm_contig_spin);
1958 if (vm_contig_verbose) {
1959 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
1960 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT,
1961 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1964 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
1965 if (memattr != VM_MEMATTR_DEFAULT)
1966 for (i = 0;i < size;i++)
1967 pmap_page_set_memattr(&m[i], memattr);
1972 * Free contiguously allocated pages. The pages will be wired but not busy.
1973 * When freeing to the alist we leave them wired and not busy.
1976 vm_page_free_contig(vm_page_t m, unsigned long size)
1978 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1979 vm_pindex_t start = pa >> PAGE_SHIFT;
1980 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
1982 if (vm_contig_verbose) {
1983 kprintf("vm_page_free_contig: %016jx/%ldk\n",
1984 (intmax_t)pa, size / 1024);
1986 if (pa < vm_low_phys_reserved) {
1987 KKASSERT(pa + size <= vm_low_phys_reserved);
1988 spin_lock(&vm_contig_spin);
1989 alist_free(&vm_contig_alist, start, pages);
1990 spin_unlock(&vm_contig_spin);
1993 vm_page_busy_wait(m, FALSE, "cpgfr");
1994 vm_page_unwire(m, 0);
2005 * Wait for sufficient free memory for nominal heavy memory use kernel
2008 * WARNING! Be sure never to call this in any vm_pageout code path, which
2009 * will trivially deadlock the system.
2012 vm_wait_nominal(void)
2014 while (vm_page_count_min(0))
2019 * Test if vm_wait_nominal() would block.
2022 vm_test_nominal(void)
2024 if (vm_page_count_min(0))
2030 * Block until free pages are available for allocation, called in various
2031 * places before memory allocations.
2033 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
2034 * more generous then that.
2040 * never wait forever
2044 lwkt_gettoken(&vm_token);
2046 if (curthread == pagethread ||
2047 curthread == emergpager) {
2049 * The pageout daemon itself needs pages, this is bad.
2051 if (vm_page_count_min(0)) {
2052 vm_pageout_pages_needed = 1;
2053 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
2057 * Wakeup the pageout daemon if necessary and wait.
2059 * Do not wait indefinitely for the target to be reached,
2060 * as load might prevent it from being reached any time soon.
2061 * But wait a little to try to slow down page allocations
2062 * and to give more important threads (the pagedaemon)
2063 * allocation priority.
2065 if (vm_page_count_target()) {
2066 if (vm_pages_needed == 0) {
2067 vm_pages_needed = 1;
2068 wakeup(&vm_pages_needed);
2070 ++vm_pages_waiting; /* SMP race ok */
2071 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
2074 lwkt_reltoken(&vm_token);
2078 * Block until free pages are available for allocation
2080 * Called only from vm_fault so that processes page faulting can be
2084 vm_wait_pfault(void)
2087 * Wakeup the pageout daemon if necessary and wait.
2089 * Do not wait indefinitely for the target to be reached,
2090 * as load might prevent it from being reached any time soon.
2091 * But wait a little to try to slow down page allocations
2092 * and to give more important threads (the pagedaemon)
2093 * allocation priority.
2095 if (vm_page_count_min(0)) {
2096 lwkt_gettoken(&vm_token);
2097 while (vm_page_count_severe()) {
2098 if (vm_page_count_target()) {
2101 if (vm_pages_needed == 0) {
2102 vm_pages_needed = 1;
2103 wakeup(&vm_pages_needed);
2105 ++vm_pages_waiting; /* SMP race ok */
2106 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
2109 * Do not stay stuck in the loop if the system is trying
2110 * to kill the process.
2113 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
2117 lwkt_reltoken(&vm_token);
2122 * Put the specified page on the active list (if appropriate). Ensure
2123 * that act_count is at least ACT_INIT but do not otherwise mess with it.
2125 * The caller should be holding the page busied ? XXX
2126 * This routine may not block.
2129 vm_page_activate(vm_page_t m)
2133 vm_page_spin_lock(m);
2134 if (m->queue - m->pc != PQ_ACTIVE) {
2135 _vm_page_queue_spin_lock(m);
2136 oqueue = _vm_page_rem_queue_spinlocked(m);
2137 /* page is left spinlocked, queue is unlocked */
2139 if (oqueue == PQ_CACHE)
2140 mycpu->gd_cnt.v_reactivated++;
2141 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2142 if (m->act_count < ACT_INIT)
2143 m->act_count = ACT_INIT;
2144 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
2146 _vm_page_and_queue_spin_unlock(m);
2147 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
2148 pagedaemon_wakeup();
2150 if (m->act_count < ACT_INIT)
2151 m->act_count = ACT_INIT;
2152 vm_page_spin_unlock(m);
2157 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2158 * routine is called when a page has been added to the cache or free
2161 * This routine may not block.
2163 static __inline void
2164 vm_page_free_wakeup(void)
2166 globaldata_t gd = mycpu;
2169 * If the pageout daemon itself needs pages, then tell it that
2170 * there are some free.
2172 if (vm_pageout_pages_needed &&
2173 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >=
2174 gd->gd_vmstats.v_pageout_free_min
2176 vm_pageout_pages_needed = 0;
2177 wakeup(&vm_pageout_pages_needed);
2181 * Wakeup processes that are waiting on memory.
2183 * Generally speaking we want to wakeup stuck processes as soon as
2184 * possible. !vm_page_count_min(0) is the absolute minimum point
2185 * where we can do this. Wait a bit longer to reduce degenerate
2186 * re-blocking (vm_page_free_hysteresis). The target check is just
2187 * to make sure the min-check w/hysteresis does not exceed the
2190 if (vm_pages_waiting) {
2191 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2192 !vm_page_count_target()) {
2193 vm_pages_waiting = 0;
2194 wakeup(&vmstats.v_free_count);
2195 ++mycpu->gd_cnt.v_ppwakeups;
2198 if (!vm_page_count_target()) {
2200 * Plenty of pages are free, wakeup everyone.
2202 vm_pages_waiting = 0;
2203 wakeup(&vmstats.v_free_count);
2204 ++mycpu->gd_cnt.v_ppwakeups;
2205 } else if (!vm_page_count_min(0)) {
2207 * Some pages are free, wakeup someone.
2209 int wcount = vm_pages_waiting;
2212 vm_pages_waiting = wcount;
2213 wakeup_one(&vmstats.v_free_count);
2214 ++mycpu->gd_cnt.v_ppwakeups;
2221 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
2222 * it from its VM object.
2224 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
2225 * return (the page will have been freed).
2228 vm_page_free_toq(vm_page_t m)
2230 mycpu->gd_cnt.v_tfree++;
2231 KKASSERT((m->flags & PG_MAPPED) == 0);
2232 KKASSERT(m->flags & PG_BUSY);
2234 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
2235 kprintf("vm_page_free: pindex(%lu), busy(%d), "
2236 "PG_BUSY(%d), hold(%d)\n",
2237 (u_long)m->pindex, m->busy,
2238 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count);
2239 if ((m->queue - m->pc) == PQ_FREE)
2240 panic("vm_page_free: freeing free page");
2242 panic("vm_page_free: freeing busy page");
2246 * Remove from object, spinlock the page and its queues and
2247 * remove from any queue. No queue spinlock will be held
2248 * after this section (because the page was removed from any
2252 vm_page_and_queue_spin_lock(m);
2253 _vm_page_rem_queue_spinlocked(m);
2256 * No further management of fictitious pages occurs beyond object
2257 * and queue removal.
2259 if ((m->flags & PG_FICTITIOUS) != 0) {
2260 vm_page_spin_unlock(m);
2268 if (m->wire_count != 0) {
2269 if (m->wire_count > 1) {
2271 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2272 m->wire_count, (long)m->pindex);
2274 panic("vm_page_free: freeing wired page");
2278 * Clear the UNMANAGED flag when freeing an unmanaged page.
2279 * Clear the NEED_COMMIT flag
2281 if (m->flags & PG_UNMANAGED)
2282 vm_page_flag_clear(m, PG_UNMANAGED);
2283 if (m->flags & PG_NEED_COMMIT)
2284 vm_page_flag_clear(m, PG_NEED_COMMIT);
2286 if (m->hold_count != 0) {
2287 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2289 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1);
2293 * This sequence allows us to clear PG_BUSY while still holding
2294 * its spin lock, which reduces contention vs allocators. We
2295 * must not leave the queue locked or _vm_page_wakeup() may
2298 _vm_page_queue_spin_unlock(m);
2299 if (_vm_page_wakeup(m)) {
2300 vm_page_spin_unlock(m);
2303 vm_page_spin_unlock(m);
2305 vm_page_free_wakeup();
2309 * vm_page_unmanage()
2311 * Prevent PV management from being done on the page. The page is
2312 * removed from the paging queues as if it were wired, and as a
2313 * consequence of no longer being managed the pageout daemon will not
2314 * touch it (since there is no way to locate the pte mappings for the
2315 * page). madvise() calls that mess with the pmap will also no longer
2316 * operate on the page.
2318 * Beyond that the page is still reasonably 'normal'. Freeing the page
2319 * will clear the flag.
2321 * This routine is used by OBJT_PHYS objects - objects using unswappable
2322 * physical memory as backing store rather then swap-backed memory and
2323 * will eventually be extended to support 4MB unmanaged physical
2326 * Caller must be holding the page busy.
2329 vm_page_unmanage(vm_page_t m)
2331 KKASSERT(m->flags & PG_BUSY);
2332 if ((m->flags & PG_UNMANAGED) == 0) {
2333 if (m->wire_count == 0)
2336 vm_page_flag_set(m, PG_UNMANAGED);
2340 * Mark this page as wired down by yet another map, removing it from
2341 * paging queues as necessary.
2343 * Caller must be holding the page busy.
2346 vm_page_wire(vm_page_t m)
2349 * Only bump the wire statistics if the page is not already wired,
2350 * and only unqueue the page if it is on some queue (if it is unmanaged
2351 * it is already off the queues). Don't do anything with fictitious
2352 * pages because they are always wired.
2354 KKASSERT(m->flags & PG_BUSY);
2355 if ((m->flags & PG_FICTITIOUS) == 0) {
2356 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2357 if ((m->flags & PG_UNMANAGED) == 0)
2359 atomic_add_int(&mycpu->gd_vmstats_adj.v_wire_count, 1);
2361 KASSERT(m->wire_count != 0,
2362 ("vm_page_wire: wire_count overflow m=%p", m));
2367 * Release one wiring of this page, potentially enabling it to be paged again.
2369 * Many pages placed on the inactive queue should actually go
2370 * into the cache, but it is difficult to figure out which. What
2371 * we do instead, if the inactive target is well met, is to put
2372 * clean pages at the head of the inactive queue instead of the tail.
2373 * This will cause them to be moved to the cache more quickly and
2374 * if not actively re-referenced, freed more quickly. If we just
2375 * stick these pages at the end of the inactive queue, heavy filesystem
2376 * meta-data accesses can cause an unnecessary paging load on memory bound
2377 * processes. This optimization causes one-time-use metadata to be
2378 * reused more quickly.
2380 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2381 * the inactive queue. This helps the pageout daemon determine memory
2382 * pressure and act on out-of-memory situations more quickly.
2384 * BUT, if we are in a low-memory situation we have no choice but to
2385 * put clean pages on the cache queue.
2387 * A number of routines use vm_page_unwire() to guarantee that the page
2388 * will go into either the inactive or active queues, and will NEVER
2389 * be placed in the cache - for example, just after dirtying a page.
2390 * dirty pages in the cache are not allowed.
2392 * This routine may not block.
2395 vm_page_unwire(vm_page_t m, int activate)
2397 KKASSERT(m->flags & PG_BUSY);
2398 if (m->flags & PG_FICTITIOUS) {
2400 } else if (m->wire_count <= 0) {
2401 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2403 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2404 atomic_add_int(&mycpu->gd_vmstats_adj.v_wire_count, -1);
2405 if (m->flags & PG_UNMANAGED) {
2407 } else if (activate || (m->flags & PG_NEED_COMMIT)) {
2408 vm_page_spin_lock(m);
2409 _vm_page_add_queue_spinlocked(m,
2410 PQ_ACTIVE + m->pc, 0);
2411 _vm_page_and_queue_spin_unlock(m);
2413 vm_page_spin_lock(m);
2414 vm_page_flag_clear(m, PG_WINATCFLS);
2415 _vm_page_add_queue_spinlocked(m,
2416 PQ_INACTIVE + m->pc, 0);
2417 ++vm_swapcache_inactive_heuristic;
2418 _vm_page_and_queue_spin_unlock(m);
2425 * Move the specified page to the inactive queue. If the page has
2426 * any associated swap, the swap is deallocated.
2428 * Normally athead is 0 resulting in LRU operation. athead is set
2429 * to 1 if we want this page to be 'as if it were placed in the cache',
2430 * except without unmapping it from the process address space.
2432 * vm_page's spinlock must be held on entry and will remain held on return.
2433 * This routine may not block.
2436 _vm_page_deactivate_locked(vm_page_t m, int athead)
2441 * Ignore if already inactive.
2443 if (m->queue - m->pc == PQ_INACTIVE)
2445 _vm_page_queue_spin_lock(m);
2446 oqueue = _vm_page_rem_queue_spinlocked(m);
2448 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2449 if (oqueue == PQ_CACHE)
2450 mycpu->gd_cnt.v_reactivated++;
2451 vm_page_flag_clear(m, PG_WINATCFLS);
2452 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2454 ++vm_swapcache_inactive_heuristic;
2456 /* NOTE: PQ_NONE if condition not taken */
2457 _vm_page_queue_spin_unlock(m);
2458 /* leaves vm_page spinlocked */
2462 * Attempt to deactivate a page.
2467 vm_page_deactivate(vm_page_t m)
2469 vm_page_spin_lock(m);
2470 _vm_page_deactivate_locked(m, 0);
2471 vm_page_spin_unlock(m);
2475 vm_page_deactivate_locked(vm_page_t m)
2477 _vm_page_deactivate_locked(m, 0);
2481 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it.
2483 * This function returns non-zero if it successfully moved the page to
2486 * This function unconditionally unbusies the page on return.
2489 vm_page_try_to_cache(vm_page_t m)
2491 vm_page_spin_lock(m);
2492 if (m->dirty || m->hold_count || m->wire_count ||
2493 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2494 if (_vm_page_wakeup(m)) {
2495 vm_page_spin_unlock(m);
2498 vm_page_spin_unlock(m);
2502 vm_page_spin_unlock(m);
2505 * Page busied by us and no longer spinlocked. Dirty pages cannot
2506 * be moved to the cache.
2508 vm_page_test_dirty(m);
2509 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2518 * Attempt to free the page. If we cannot free it, we do nothing.
2519 * 1 is returned on success, 0 on failure.
2524 vm_page_try_to_free(vm_page_t m)
2526 vm_page_spin_lock(m);
2527 if (vm_page_busy_try(m, TRUE)) {
2528 vm_page_spin_unlock(m);
2533 * The page can be in any state, including already being on the free
2534 * queue. Check to see if it really can be freed.
2536 if (m->dirty || /* can't free if it is dirty */
2537 m->hold_count || /* or held (XXX may be wrong) */
2538 m->wire_count || /* or wired */
2539 (m->flags & (PG_UNMANAGED | /* or unmanaged */
2540 PG_NEED_COMMIT)) || /* or needs a commit */
2541 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */
2542 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */
2543 if (_vm_page_wakeup(m)) {
2544 vm_page_spin_unlock(m);
2547 vm_page_spin_unlock(m);
2551 vm_page_spin_unlock(m);
2554 * We can probably free the page.
2556 * Page busied by us and no longer spinlocked. Dirty pages will
2557 * not be freed by this function. We have to re-test the
2558 * dirty bit after cleaning out the pmaps.
2560 vm_page_test_dirty(m);
2561 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2565 vm_page_protect(m, VM_PROT_NONE);
2566 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2577 * Put the specified page onto the page cache queue (if appropriate).
2579 * The page must be busy, and this routine will release the busy and
2580 * possibly even free the page.
2583 vm_page_cache(vm_page_t m)
2586 * Not suitable for the cache
2588 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2589 m->busy || m->wire_count || m->hold_count) {
2595 * Already in the cache (and thus not mapped)
2597 if ((m->queue - m->pc) == PQ_CACHE) {
2598 KKASSERT((m->flags & PG_MAPPED) == 0);
2604 * Caller is required to test m->dirty, but note that the act of
2605 * removing the page from its maps can cause it to become dirty
2606 * on an SMP system due to another cpu running in usermode.
2609 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2614 * Remove all pmaps and indicate that the page is not
2615 * writeable or mapped. Our vm_page_protect() call may
2616 * have blocked (especially w/ VM_PROT_NONE), so recheck
2619 vm_page_protect(m, VM_PROT_NONE);
2620 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2621 m->busy || m->wire_count || m->hold_count) {
2623 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2624 vm_page_deactivate(m);
2627 _vm_page_and_queue_spin_lock(m);
2628 _vm_page_rem_queue_spinlocked(m);
2629 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2630 _vm_page_queue_spin_unlock(m);
2631 if (_vm_page_wakeup(m)) {
2632 vm_page_spin_unlock(m);
2635 vm_page_spin_unlock(m);
2637 vm_page_free_wakeup();
2642 * vm_page_dontneed()
2644 * Cache, deactivate, or do nothing as appropriate. This routine
2645 * is typically used by madvise() MADV_DONTNEED.
2647 * Generally speaking we want to move the page into the cache so
2648 * it gets reused quickly. However, this can result in a silly syndrome
2649 * due to the page recycling too quickly. Small objects will not be
2650 * fully cached. On the otherhand, if we move the page to the inactive
2651 * queue we wind up with a problem whereby very large objects
2652 * unnecessarily blow away our inactive and cache queues.
2654 * The solution is to move the pages based on a fixed weighting. We
2655 * either leave them alone, deactivate them, or move them to the cache,
2656 * where moving them to the cache has the highest weighting.
2657 * By forcing some pages into other queues we eventually force the
2658 * system to balance the queues, potentially recovering other unrelated
2659 * space from active. The idea is to not force this to happen too
2662 * The page must be busied.
2665 vm_page_dontneed(vm_page_t m)
2667 static int dnweight;
2674 * occassionally leave the page alone
2676 if ((dnw & 0x01F0) == 0 ||
2677 m->queue - m->pc == PQ_INACTIVE ||
2678 m->queue - m->pc == PQ_CACHE
2680 if (m->act_count >= ACT_INIT)
2686 * If vm_page_dontneed() is inactivating a page, it must clear
2687 * the referenced flag; otherwise the pagedaemon will see references
2688 * on the page in the inactive queue and reactivate it. Until the
2689 * page can move to the cache queue, madvise's job is not done.
2691 vm_page_flag_clear(m, PG_REFERENCED);
2692 pmap_clear_reference(m);
2695 vm_page_test_dirty(m);
2697 if (m->dirty || (dnw & 0x0070) == 0) {
2699 * Deactivate the page 3 times out of 32.
2704 * Cache the page 28 times out of every 32. Note that
2705 * the page is deactivated instead of cached, but placed
2706 * at the head of the queue instead of the tail.
2710 vm_page_spin_lock(m);
2711 _vm_page_deactivate_locked(m, head);
2712 vm_page_spin_unlock(m);
2716 * These routines manipulate the 'soft busy' count for a page. A soft busy
2717 * is almost like PG_BUSY except that it allows certain compatible operations
2718 * to occur on the page while it is busy. For example, a page undergoing a
2719 * write can still be mapped read-only.
2721 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2722 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2723 * busy bit is cleared.
2725 * The caller must hold the page BUSY when making these two calls.
2728 vm_page_io_start(vm_page_t m)
2730 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2731 atomic_add_char(&m->busy, 1);
2732 vm_page_flag_set(m, PG_SBUSY);
2736 vm_page_io_finish(vm_page_t m)
2738 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2739 atomic_subtract_char(&m->busy, 1);
2741 vm_page_flag_clear(m, PG_SBUSY);
2745 * Indicate that a clean VM page requires a filesystem commit and cannot
2746 * be reused. Used by tmpfs.
2749 vm_page_need_commit(vm_page_t m)
2751 vm_page_flag_set(m, PG_NEED_COMMIT);
2752 vm_object_set_writeable_dirty(m->object);
2756 vm_page_clear_commit(vm_page_t m)
2758 vm_page_flag_clear(m, PG_NEED_COMMIT);
2762 * Grab a page, blocking if it is busy and allocating a page if necessary.
2763 * A busy page is returned or NULL. The page may or may not be valid and
2764 * might not be on a queue (the caller is responsible for the disposition of
2767 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2768 * page will be zero'd and marked valid.
2770 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2771 * valid even if it already exists.
2773 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2774 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2775 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2777 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2778 * always returned if we had blocked.
2780 * This routine may not be called from an interrupt.
2782 * No other requirements.
2785 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2791 KKASSERT(allocflags &
2792 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2793 vm_object_hold_shared(object);
2795 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2797 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2798 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2803 } else if (m == NULL) {
2805 vm_object_upgrade(object);
2808 if (allocflags & VM_ALLOC_RETRY)
2809 allocflags |= VM_ALLOC_NULL_OK;
2810 m = vm_page_alloc(object, pindex,
2811 allocflags & ~VM_ALLOC_RETRY);
2815 if ((allocflags & VM_ALLOC_RETRY) == 0)
2824 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2826 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2827 * valid even if already valid.
2829 * NOTE! We have removed all of the PG_ZERO optimizations and also
2830 * removed the idle zeroing code. These optimizations actually
2831 * slow things down on modern cpus because the zerod area is
2832 * likely uncached, placing a memory-access burden on the
2833 * accesors taking the fault.
2835 * By always zeroing the page in-line with the fault, no
2836 * dynamic ram reads are needed and the caches are hot, ready
2837 * for userland to access the memory.
2839 if (m->valid == 0) {
2840 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2841 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2842 m->valid = VM_PAGE_BITS_ALL;
2844 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2845 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2846 m->valid = VM_PAGE_BITS_ALL;
2849 vm_object_drop(object);
2854 * Mapping function for valid bits or for dirty bits in
2855 * a page. May not block.
2857 * Inputs are required to range within a page.
2863 vm_page_bits(int base, int size)
2869 base + size <= PAGE_SIZE,
2870 ("vm_page_bits: illegal base/size %d/%d", base, size)
2873 if (size == 0) /* handle degenerate case */
2876 first_bit = base >> DEV_BSHIFT;
2877 last_bit = (base + size - 1) >> DEV_BSHIFT;
2879 return ((2 << last_bit) - (1 << first_bit));
2883 * Sets portions of a page valid and clean. The arguments are expected
2884 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2885 * of any partial chunks touched by the range. The invalid portion of
2886 * such chunks will be zero'd.
2888 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2889 * align base to DEV_BSIZE so as not to mark clean a partially
2890 * truncated device block. Otherwise the dirty page status might be
2893 * This routine may not block.
2895 * (base + size) must be less then or equal to PAGE_SIZE.
2898 _vm_page_zero_valid(vm_page_t m, int base, int size)
2903 if (size == 0) /* handle degenerate case */
2907 * If the base is not DEV_BSIZE aligned and the valid
2908 * bit is clear, we have to zero out a portion of the
2912 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2913 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2915 pmap_zero_page_area(
2923 * If the ending offset is not DEV_BSIZE aligned and the
2924 * valid bit is clear, we have to zero out a portion of
2928 endoff = base + size;
2930 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2931 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
2933 pmap_zero_page_area(
2936 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
2942 * Set valid, clear dirty bits. If validating the entire
2943 * page we can safely clear the pmap modify bit. We also
2944 * use this opportunity to clear the PG_NOSYNC flag. If a process
2945 * takes a write fault on a MAP_NOSYNC memory area the flag will
2948 * We set valid bits inclusive of any overlap, but we can only
2949 * clear dirty bits for DEV_BSIZE chunks that are fully within
2952 * Page must be busied?
2953 * No other requirements.
2956 vm_page_set_valid(vm_page_t m, int base, int size)
2958 _vm_page_zero_valid(m, base, size);
2959 m->valid |= vm_page_bits(base, size);
2964 * Set valid bits and clear dirty bits.
2966 * Page must be busied by caller.
2968 * NOTE: This function does not clear the pmap modified bit.
2969 * Also note that e.g. NFS may use a byte-granular base
2972 * No other requirements.
2975 vm_page_set_validclean(vm_page_t m, int base, int size)
2979 _vm_page_zero_valid(m, base, size);
2980 pagebits = vm_page_bits(base, size);
2981 m->valid |= pagebits;
2982 m->dirty &= ~pagebits;
2983 if (base == 0 && size == PAGE_SIZE) {
2984 /*pmap_clear_modify(m);*/
2985 vm_page_flag_clear(m, PG_NOSYNC);
2990 * Set valid & dirty. Used by buwrite()
2992 * Page must be busied by caller.
2995 vm_page_set_validdirty(vm_page_t m, int base, int size)
2999 pagebits = vm_page_bits(base, size);
3000 m->valid |= pagebits;
3001 m->dirty |= pagebits;
3003 vm_object_set_writeable_dirty(m->object);
3009 * NOTE: This function does not clear the pmap modified bit.
3010 * Also note that e.g. NFS may use a byte-granular base
3013 * Page must be busied?
3014 * No other requirements.
3017 vm_page_clear_dirty(vm_page_t m, int base, int size)
3019 m->dirty &= ~vm_page_bits(base, size);
3020 if (base == 0 && size == PAGE_SIZE) {
3021 /*pmap_clear_modify(m);*/
3022 vm_page_flag_clear(m, PG_NOSYNC);
3027 * Make the page all-dirty.
3029 * Also make sure the related object and vnode reflect the fact that the
3030 * object may now contain a dirty page.
3032 * Page must be busied?
3033 * No other requirements.
3036 vm_page_dirty(vm_page_t m)
3039 int pqtype = m->queue - m->pc;
3041 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
3042 ("vm_page_dirty: page in free/cache queue!"));
3043 if (m->dirty != VM_PAGE_BITS_ALL) {
3044 m->dirty = VM_PAGE_BITS_ALL;
3046 vm_object_set_writeable_dirty(m->object);
3051 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3052 * valid and dirty bits for the effected areas are cleared.
3054 * Page must be busied?
3056 * No other requirements.
3059 vm_page_set_invalid(vm_page_t m, int base, int size)
3063 bits = vm_page_bits(base, size);
3066 atomic_add_int(&m->object->generation, 1);
3070 * The kernel assumes that the invalid portions of a page contain
3071 * garbage, but such pages can be mapped into memory by user code.
3072 * When this occurs, we must zero out the non-valid portions of the
3073 * page so user code sees what it expects.
3075 * Pages are most often semi-valid when the end of a file is mapped
3076 * into memory and the file's size is not page aligned.
3078 * Page must be busied?
3079 * No other requirements.
3082 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3088 * Scan the valid bits looking for invalid sections that
3089 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
3090 * valid bit may be set ) have already been zerod by
3091 * vm_page_set_validclean().
3093 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3094 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3095 (m->valid & (1 << i))
3098 pmap_zero_page_area(
3101 (i - b) << DEV_BSHIFT
3109 * setvalid is TRUE when we can safely set the zero'd areas
3110 * as being valid. We can do this if there are no cache consistency
3111 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3114 m->valid = VM_PAGE_BITS_ALL;
3118 * Is a (partial) page valid? Note that the case where size == 0
3119 * will return FALSE in the degenerate case where the page is entirely
3120 * invalid, and TRUE otherwise.
3123 * No other requirements.
3126 vm_page_is_valid(vm_page_t m, int base, int size)
3128 int bits = vm_page_bits(base, size);
3130 if (m->valid && ((m->valid & bits) == bits))
3137 * update dirty bits from pmap/mmu. May not block.
3139 * Caller must hold the page busy
3142 vm_page_test_dirty(vm_page_t m)
3144 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
3149 #include "opt_ddb.h"
3151 #include <ddb/ddb.h>
3153 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3155 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
3156 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
3157 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
3158 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
3159 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
3160 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
3161 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
3162 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
3163 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
3164 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
3167 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3170 db_printf("PQ_FREE:");
3171 for (i = 0; i < PQ_L2_SIZE; i++) {
3172 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
3176 db_printf("PQ_CACHE:");
3177 for(i = 0; i < PQ_L2_SIZE; i++) {
3178 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
3182 db_printf("PQ_ACTIVE:");
3183 for(i = 0; i < PQ_L2_SIZE; i++) {
3184 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3188 db_printf("PQ_INACTIVE:");
3189 for(i = 0; i < PQ_L2_SIZE; i++) {
3190 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);