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?
172 vm_pindex_t first_page = 0;
173 vm_pindex_t 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_long(&vmstats.v_page_count, 1);
231 atomic_add_long(&vmstats.v_dma_pages, 1);
234 atomic_add_long(&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_long(&vmstats.v_page_count, 1);
246 atomic_add_long(&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_dma_init 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_long(&vmstats.v_dma_pages, -(long)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 = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
797 atomic_add_long(cnt, -1);
798 if (*cnt < -VMMETER_SLOP_COUNT) {
799 u_long copy = atomic_swap_long(cnt, 0);
800 cnt = (long *)((char *)&vmstats + pq->cnt_offset);
801 atomic_add_long(cnt, copy);
802 cnt = (long *)((char *)&mycpu->gd_vmstats +
804 atomic_add_long(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 = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
843 atomic_add_long(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 BUSY. If also_m_busy is TRUE we wait
863 * until the page is no longer BUSY or SBUSY (busy_count field is 0).
865 * Returns TRUE if it had to sleep, FALSE if we did not. Only one sleep
866 * call will be made before returning.
868 * This function does NOT busy the page and on return the page is not
869 * guaranteed to be available.
872 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
874 u_int32_t busy_count;
877 busy_count = m->busy_count;
880 if ((busy_count & PBUSY_LOCKED) == 0 &&
881 (also_m_busy == 0 || (busy_count & PBUSY_MASK) == 0)) {
884 tsleep_interlock(m, 0);
885 if (atomic_cmpset_int(&m->busy_count, busy_count,
886 busy_count | PBUSY_WANTED)) {
887 atomic_set_int(&m->flags, PG_REFERENCED);
888 tsleep(m, PINTERLOCKED, msg, 0);
895 * This calculates and returns a page color given an optional VM object and
896 * either a pindex or an iterator. We attempt to return a cpu-localized
897 * pg_color that is still roughly 16-way set-associative. The CPU topology
898 * is used if it was probed.
900 * The caller may use the returned value to index into e.g. PQ_FREE when
901 * allocating a page in order to nominally obtain pages that are hopefully
902 * already localized to the requesting cpu. This function is not able to
903 * provide any sort of guarantee of this, but does its best to improve
904 * hardware cache management performance.
906 * WARNING! The caller must mask the returned value with PQ_L2_MASK.
909 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex)
916 phys_id = get_cpu_phys_id(cpuid);
917 core_id = get_cpu_core_id(cpuid);
918 object_pg_color = object ? object->pg_color : 0;
920 if (cpu_topology_phys_ids && cpu_topology_core_ids) {
924 * Break us down by socket and cpu
926 pg_color = phys_id * PQ_L2_SIZE / cpu_topology_phys_ids;
927 pg_color += core_id * PQ_L2_SIZE /
928 (cpu_topology_core_ids * cpu_topology_phys_ids);
931 * Calculate remaining component for object/queue color
933 grpsize = PQ_L2_SIZE / (cpu_topology_core_ids *
934 cpu_topology_phys_ids);
936 pg_color += (pindex + object_pg_color) % grpsize;
941 /* 3->9, 4->8, 5->10, 6->12, 7->14 */
946 pg_color += (pindex + object_pg_color) % grpsize;
950 * Unknown topology, distribute things evenly.
952 pg_color = cpuid * PQ_L2_SIZE / ncpus;
953 pg_color += pindex + object_pg_color;
955 return (pg_color & PQ_L2_MASK);
959 * Wait until BUSY can be set, then set it. If also_m_busy is TRUE we
960 * also wait for m->busy_count to become 0 before setting PBUSY_LOCKED.
963 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
964 int also_m_busy, const char *msg
967 u_int32_t busy_count;
970 busy_count = m->busy_count;
972 if (busy_count & PBUSY_LOCKED) {
973 tsleep_interlock(m, 0);
974 if (atomic_cmpset_int(&m->busy_count, busy_count,
975 busy_count | PBUSY_WANTED)) {
976 atomic_set_int(&m->flags, PG_REFERENCED);
977 tsleep(m, PINTERLOCKED, msg, 0);
979 } else if (also_m_busy && busy_count) {
980 tsleep_interlock(m, 0);
981 if (atomic_cmpset_int(&m->busy_count, busy_count,
982 busy_count | PBUSY_WANTED)) {
983 atomic_set_int(&m->flags, PG_REFERENCED);
984 tsleep(m, PINTERLOCKED, msg, 0);
987 if (atomic_cmpset_int(&m->busy_count, busy_count,
988 busy_count | PBUSY_LOCKED)) {
991 m->busy_line = lineno;
1000 * Attempt to set BUSY. If also_m_busy is TRUE we only succeed if
1001 * m->busy_count is also 0.
1003 * Returns non-zero on failure.
1006 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
1009 u_int32_t busy_count;
1012 busy_count = m->busy_count;
1014 if (busy_count & PBUSY_LOCKED)
1016 if (also_m_busy && (busy_count & PBUSY_MASK) != 0)
1018 if (atomic_cmpset_int(&m->busy_count, busy_count,
1019 busy_count | PBUSY_LOCKED)) {
1020 #ifdef VM_PAGE_DEBUG
1021 m->busy_func = func;
1022 m->busy_line = lineno;
1030 * Clear the BUSY flag and return non-zero to indicate to the caller
1031 * that a wakeup() should be performed.
1033 * The vm_page must be spinlocked and will remain spinlocked on return.
1034 * The related queue must NOT be spinlocked (which could deadlock us).
1040 _vm_page_wakeup(vm_page_t m)
1042 u_int32_t busy_count;
1045 busy_count = m->busy_count;
1047 if (atomic_cmpset_int(&m->busy_count, busy_count,
1049 ~(PBUSY_LOCKED | PBUSY_WANTED))) {
1053 return((int)(busy_count & PBUSY_WANTED));
1057 * Clear the BUSY flag and wakeup anyone waiting for the page. This
1058 * is typically the last call you make on a page before moving onto
1062 vm_page_wakeup(vm_page_t m)
1064 KASSERT(m->busy_count & PBUSY_LOCKED,
1065 ("vm_page_wakeup: page not busy!!!"));
1066 vm_page_spin_lock(m);
1067 if (_vm_page_wakeup(m)) {
1068 vm_page_spin_unlock(m);
1071 vm_page_spin_unlock(m);
1076 * Holding a page keeps it from being reused. Other parts of the system
1077 * can still disassociate the page from its current object and free it, or
1078 * perform read or write I/O on it and/or otherwise manipulate the page,
1079 * but if the page is held the VM system will leave the page and its data
1080 * intact and not reuse the page for other purposes until the last hold
1081 * reference is released. (see vm_page_wire() if you want to prevent the
1082 * page from being disassociated from its object too).
1084 * The caller must still validate the contents of the page and, if necessary,
1085 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
1086 * before manipulating the page.
1088 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
1091 vm_page_hold(vm_page_t m)
1093 vm_page_spin_lock(m);
1094 atomic_add_int(&m->hold_count, 1);
1095 if (m->queue - m->pc == PQ_FREE) {
1096 _vm_page_queue_spin_lock(m);
1097 _vm_page_rem_queue_spinlocked(m);
1098 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
1099 _vm_page_queue_spin_unlock(m);
1101 vm_page_spin_unlock(m);
1105 * The opposite of vm_page_hold(). If the page is on the HOLD queue
1106 * it was freed while held and must be moved back to the FREE queue.
1109 vm_page_unhold(vm_page_t m)
1111 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE,
1112 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)",
1113 m, m->hold_count, m->queue - m->pc));
1114 vm_page_spin_lock(m);
1115 atomic_add_int(&m->hold_count, -1);
1116 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
1117 _vm_page_queue_spin_lock(m);
1118 _vm_page_rem_queue_spinlocked(m);
1119 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1);
1120 _vm_page_queue_spin_unlock(m);
1122 vm_page_spin_unlock(m);
1128 * Create a fictitious page with the specified physical address and
1129 * memory attribute. The memory attribute is the only the machine-
1130 * dependent aspect of a fictitious page that must be initialized.
1134 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1137 if ((m->flags & PG_FICTITIOUS) != 0) {
1139 * The page's memattr might have changed since the
1140 * previous initialization. Update the pmap to the
1145 m->phys_addr = paddr;
1147 /* Fictitious pages don't use "segind". */
1148 /* Fictitious pages don't use "order" or "pool". */
1149 m->flags = PG_FICTITIOUS | PG_UNMANAGED;
1150 m->busy_count = PBUSY_LOCKED;
1152 spin_init(&m->spin, "fake_page");
1155 pmap_page_set_memattr(m, memattr);
1159 * Inserts the given vm_page into the object and object list.
1161 * The pagetables are not updated but will presumably fault the page
1162 * in if necessary, or if a kernel page the caller will at some point
1163 * enter the page into the kernel's pmap. We are not allowed to block
1164 * here so we *can't* do this anyway.
1166 * This routine may not block.
1167 * This routine must be called with the vm_object held.
1168 * This routine must be called with a critical section held.
1170 * This routine returns TRUE if the page was inserted into the object
1171 * successfully, and FALSE if the page already exists in the object.
1174 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1176 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object));
1177 if (m->object != NULL)
1178 panic("vm_page_insert: already inserted");
1180 atomic_add_int(&object->generation, 1);
1183 * Record the object/offset pair in this page and add the
1184 * pv_list_count of the page to the object.
1186 * The vm_page spin lock is required for interactions with the pmap.
1188 vm_page_spin_lock(m);
1191 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
1194 vm_page_spin_unlock(m);
1197 ++object->resident_page_count;
1198 ++mycpu->gd_vmtotal.t_rm;
1199 vm_page_spin_unlock(m);
1202 * Since we are inserting a new and possibly dirty page,
1203 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
1205 if ((m->valid & m->dirty) ||
1206 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT)))
1207 vm_object_set_writeable_dirty(object);
1210 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
1212 swap_pager_page_inserted(m);
1217 * Removes the given vm_page_t from the (object,index) table
1219 * The underlying pmap entry (if any) is NOT removed here.
1220 * This routine may not block.
1222 * The page must be BUSY and will remain BUSY on return.
1223 * No other requirements.
1225 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
1229 vm_page_remove(vm_page_t m)
1233 if (m->object == NULL) {
1237 if ((m->busy_count & PBUSY_LOCKED) == 0)
1238 panic("vm_page_remove: page not busy");
1242 vm_object_hold(object);
1245 * Remove the page from the object and update the object.
1247 * The vm_page spin lock is required for interactions with the pmap.
1249 vm_page_spin_lock(m);
1250 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
1251 --object->resident_page_count;
1252 --mycpu->gd_vmtotal.t_rm;
1254 atomic_add_int(&object->generation, 1);
1255 vm_page_spin_unlock(m);
1257 vm_object_drop(object);
1261 * Locate and return the page at (object, pindex), or NULL if the
1262 * page could not be found.
1264 * The caller must hold the vm_object token.
1267 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1272 * Search the hash table for this object/offset pair
1274 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1275 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1276 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1281 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1283 int also_m_busy, const char *msg
1286 u_int32_t busy_count;
1289 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1290 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1292 KKASSERT(m->object == object && m->pindex == pindex);
1293 busy_count = m->busy_count;
1295 if (busy_count & PBUSY_LOCKED) {
1296 tsleep_interlock(m, 0);
1297 if (atomic_cmpset_int(&m->busy_count, busy_count,
1298 busy_count | PBUSY_WANTED)) {
1299 atomic_set_int(&m->flags, PG_REFERENCED);
1300 tsleep(m, PINTERLOCKED, msg, 0);
1301 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1304 } else if (also_m_busy && busy_count) {
1305 tsleep_interlock(m, 0);
1306 if (atomic_cmpset_int(&m->busy_count, busy_count,
1307 busy_count | PBUSY_WANTED)) {
1308 atomic_set_int(&m->flags, PG_REFERENCED);
1309 tsleep(m, PINTERLOCKED, msg, 0);
1310 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1313 } else if (atomic_cmpset_int(&m->busy_count, busy_count,
1314 busy_count | PBUSY_LOCKED)) {
1315 #ifdef VM_PAGE_DEBUG
1316 m->busy_func = func;
1317 m->busy_line = lineno;
1326 * Attempt to lookup and busy a page.
1328 * Returns NULL if the page could not be found
1330 * Returns a vm_page and error == TRUE if the page exists but could not
1333 * Returns a vm_page and error == FALSE on success.
1336 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1338 int also_m_busy, int *errorp
1341 u_int32_t busy_count;
1344 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1345 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1348 KKASSERT(m->object == object && m->pindex == pindex);
1349 busy_count = m->busy_count;
1351 if (busy_count & PBUSY_LOCKED) {
1355 if (also_m_busy && busy_count) {
1359 if (atomic_cmpset_int(&m->busy_count, busy_count,
1360 busy_count | PBUSY_LOCKED)) {
1361 #ifdef VM_PAGE_DEBUG
1362 m->busy_func = func;
1363 m->busy_line = lineno;
1372 * Returns a page that is only soft-busied for use by the caller in
1373 * a read-only fashion. Returns NULL if the page could not be found,
1374 * the soft busy could not be obtained, or the page data is invalid.
1377 vm_page_lookup_sbusy_try(struct vm_object *object, vm_pindex_t pindex,
1378 int pgoff, int pgbytes)
1382 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1383 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1385 if ((m->valid != VM_PAGE_BITS_ALL &&
1386 !vm_page_is_valid(m, pgoff, pgbytes)) ||
1387 (m->flags & PG_FICTITIOUS)) {
1389 } else if (vm_page_sbusy_try(m)) {
1391 } else if ((m->valid != VM_PAGE_BITS_ALL &&
1392 !vm_page_is_valid(m, pgoff, pgbytes)) ||
1393 (m->flags & PG_FICTITIOUS)) {
1394 vm_page_sbusy_drop(m);
1402 * Caller must hold the related vm_object
1405 vm_page_next(vm_page_t m)
1409 next = vm_page_rb_tree_RB_NEXT(m);
1410 if (next && next->pindex != m->pindex + 1)
1418 * Move the given vm_page from its current object to the specified
1419 * target object/offset. The page must be busy and will remain so
1422 * new_object must be held.
1423 * This routine might block. XXX ?
1425 * NOTE: Swap associated with the page must be invalidated by the move. We
1426 * have to do this for several reasons: (1) we aren't freeing the
1427 * page, (2) we are dirtying the page, (3) the VM system is probably
1428 * moving the page from object A to B, and will then later move
1429 * the backing store from A to B and we can't have a conflict.
1431 * NOTE: We *always* dirty the page. It is necessary both for the
1432 * fact that we moved it, and because we may be invalidating
1433 * swap. If the page is on the cache, we have to deactivate it
1434 * or vm_page_dirty() will panic. Dirty pages are not allowed
1438 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1440 KKASSERT(m->busy_count & PBUSY_LOCKED);
1441 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object));
1443 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object));
1446 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1447 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1448 new_object, new_pindex);
1450 if (m->queue - m->pc == PQ_CACHE)
1451 vm_page_deactivate(m);
1456 * vm_page_unqueue() without any wakeup. This routine is used when a page
1457 * is to remain BUSYied by the caller.
1459 * This routine may not block.
1462 vm_page_unqueue_nowakeup(vm_page_t m)
1464 vm_page_and_queue_spin_lock(m);
1465 (void)_vm_page_rem_queue_spinlocked(m);
1466 vm_page_spin_unlock(m);
1470 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1473 * This routine may not block.
1476 vm_page_unqueue(vm_page_t m)
1480 vm_page_and_queue_spin_lock(m);
1481 queue = _vm_page_rem_queue_spinlocked(m);
1482 if (queue == PQ_FREE || queue == PQ_CACHE) {
1483 vm_page_spin_unlock(m);
1484 pagedaemon_wakeup();
1486 vm_page_spin_unlock(m);
1491 * vm_page_list_find()
1493 * Find a page on the specified queue with color optimization.
1495 * The page coloring optimization attempts to locate a page that does
1496 * not overload other nearby pages in the object in the cpu's L1 or L2
1497 * caches. We need this optimization because cpu caches tend to be
1498 * physical caches, while object spaces tend to be virtual.
1500 * The page coloring optimization also, very importantly, tries to localize
1501 * memory to cpus and physical sockets.
1503 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1504 * and the algorithm is adjusted to localize allocations on a per-core basis.
1505 * This is done by 'twisting' the colors.
1507 * The page is returned spinlocked and removed from its queue (it will
1508 * be on PQ_NONE), or NULL. The page is not BUSY'd. The caller
1509 * is responsible for dealing with the busy-page case (usually by
1510 * deactivating the page and looping).
1512 * NOTE: This routine is carefully inlined. A non-inlined version
1513 * is available for outside callers but the only critical path is
1514 * from within this source file.
1516 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1517 * represent stable storage, allowing us to order our locks vm_page
1518 * first, then queue.
1522 _vm_page_list_find(int basequeue, int index)
1527 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1529 m = _vm_page_list_find2(basequeue, index);
1532 vm_page_and_queue_spin_lock(m);
1533 if (m->queue == basequeue + index) {
1534 _vm_page_rem_queue_spinlocked(m);
1535 /* vm_page_t spin held, no queue spin */
1538 vm_page_and_queue_spin_unlock(m);
1544 * If we could not find the page in the desired queue try to find it in
1548 _vm_page_list_find2(int basequeue, int index)
1550 struct vpgqueues *pq;
1552 int pqmask = PQ_SET_ASSOC_MASK >> 1;
1556 index &= PQ_L2_MASK;
1557 pq = &vm_page_queues[basequeue];
1560 * Run local sets of 16, 32, 64, 128, and the whole queue if all
1561 * else fails (PQ_L2_MASK which is 255).
1564 pqmask = (pqmask << 1) | 1;
1565 for (i = 0; i <= pqmask; ++i) {
1566 pqi = (index & ~pqmask) | ((index + i) & pqmask);
1567 m = TAILQ_FIRST(&pq[pqi].pl);
1569 _vm_page_and_queue_spin_lock(m);
1570 if (m->queue == basequeue + pqi) {
1571 _vm_page_rem_queue_spinlocked(m);
1574 _vm_page_and_queue_spin_unlock(m);
1579 } while (pqmask != PQ_L2_MASK);
1585 * Returns a vm_page candidate for allocation. The page is not busied so
1586 * it can move around. The caller must busy the page (and typically
1587 * deactivate it if it cannot be busied!)
1589 * Returns a spinlocked vm_page that has been removed from its queue.
1592 vm_page_list_find(int basequeue, int index)
1594 return(_vm_page_list_find(basequeue, index));
1598 * Find a page on the cache queue with color optimization, remove it
1599 * from the queue, and busy it. The returned page will not be spinlocked.
1601 * A candidate failure will be deactivated. Candidates can fail due to
1602 * being busied by someone else, in which case they will be deactivated.
1604 * This routine may not block.
1608 vm_page_select_cache(u_short pg_color)
1613 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK);
1617 * (m) has been removed from its queue and spinlocked
1619 if (vm_page_busy_try(m, TRUE)) {
1620 _vm_page_deactivate_locked(m, 0);
1621 vm_page_spin_unlock(m);
1624 * We successfully busied the page
1626 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 &&
1627 m->hold_count == 0 &&
1628 m->wire_count == 0 &&
1629 (m->dirty & m->valid) == 0) {
1630 vm_page_spin_unlock(m);
1631 pagedaemon_wakeup();
1636 * The page cannot be recycled, deactivate it.
1638 _vm_page_deactivate_locked(m, 0);
1639 if (_vm_page_wakeup(m)) {
1640 vm_page_spin_unlock(m);
1643 vm_page_spin_unlock(m);
1651 * Find a free page. We attempt to inline the nominal case and fall back
1652 * to _vm_page_select_free() otherwise. A busied page is removed from
1653 * the queue and returned.
1655 * This routine may not block.
1657 static __inline vm_page_t
1658 vm_page_select_free(u_short pg_color)
1663 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK);
1666 if (vm_page_busy_try(m, TRUE)) {
1668 * Various mechanisms such as a pmap_collect can
1669 * result in a busy page on the free queue. We
1670 * have to move the page out of the way so we can
1671 * retry the allocation. If the other thread is not
1672 * allocating the page then m->valid will remain 0 and
1673 * the pageout daemon will free the page later on.
1675 * Since we could not busy the page, however, we
1676 * cannot make assumptions as to whether the page
1677 * will be allocated by the other thread or not,
1678 * so all we can do is deactivate it to move it out
1679 * of the way. In particular, if the other thread
1680 * wires the page it may wind up on the inactive
1681 * queue and the pageout daemon will have to deal
1682 * with that case too.
1684 _vm_page_deactivate_locked(m, 0);
1685 vm_page_spin_unlock(m);
1688 * Theoretically if we are able to busy the page
1689 * atomic with the queue removal (using the vm_page
1690 * lock) nobody else should be able to mess with the
1693 KKASSERT((m->flags & (PG_UNMANAGED |
1694 PG_NEED_COMMIT)) == 0);
1695 KASSERT(m->hold_count == 0, ("m->hold_count is not zero "
1696 "pg %p q=%d flags=%08x hold=%d wire=%d",
1697 m, m->queue, m->flags, m->hold_count, m->wire_count));
1698 KKASSERT(m->wire_count == 0);
1699 vm_page_spin_unlock(m);
1700 pagedaemon_wakeup();
1702 /* return busied and removed page */
1712 * Allocate and return a memory cell associated with this VM object/offset
1713 * pair. If object is NULL an unassociated page will be allocated.
1715 * The returned page will be busied and removed from its queues. This
1716 * routine can block and may return NULL if a race occurs and the page
1717 * is found to already exist at the specified (object, pindex).
1719 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1720 * VM_ALLOC_QUICK like normal but cannot use cache
1721 * VM_ALLOC_SYSTEM greater free drain
1722 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1723 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1724 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1725 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1726 * (see vm_page_grab())
1727 * VM_ALLOC_USE_GD ok to use per-gd cache
1729 * VM_ALLOC_CPU(n) allocate using specified cpu localization
1731 * The object must be held if not NULL
1732 * This routine may not block
1734 * Additional special handling is required when called from an interrupt
1735 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1739 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1749 * Special per-cpu free VM page cache. The pages are pre-busied
1750 * and pre-zerod for us.
1752 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1754 if (gd->gd_vmpg_count) {
1755 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1767 * CPU localization algorithm. Break the page queues up by physical
1768 * id and core id (note that two cpu threads will have the same core
1769 * id, and core_id != gd_cpuid).
1771 * This is nowhere near perfect, for example the last pindex in a
1772 * subgroup will overflow into the next cpu or package. But this
1773 * should get us good page reuse locality in heavy mixed loads.
1775 * (may be executed before the APs are started, so other GDs might
1778 if (page_req & VM_ALLOC_CPU_SPEC)
1779 cpuid_local = VM_ALLOC_GETCPU(page_req);
1781 cpuid_local = mycpu->gd_cpuid;
1783 pg_color = vm_get_pg_color(cpuid_local, object, pindex);
1786 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1787 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1790 * Certain system threads (pageout daemon, buf_daemon's) are
1791 * allowed to eat deeper into the free page list.
1793 if (curthread->td_flags & TDF_SYSTHREAD)
1794 page_req |= VM_ALLOC_SYSTEM;
1797 * Impose various limitations. Note that the v_free_reserved test
1798 * must match the opposite of vm_page_count_target() to avoid
1799 * livelocks, be careful.
1803 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved ||
1804 ((page_req & VM_ALLOC_INTERRUPT) &&
1805 gd->gd_vmstats.v_free_count > 0) ||
1806 ((page_req & VM_ALLOC_SYSTEM) &&
1807 gd->gd_vmstats.v_cache_count == 0 &&
1808 gd->gd_vmstats.v_free_count >
1809 gd->gd_vmstats.v_interrupt_free_min)
1812 * The free queue has sufficient free pages to take one out.
1814 m = vm_page_select_free(pg_color);
1815 } else if (page_req & VM_ALLOC_NORMAL) {
1817 * Allocatable from the cache (non-interrupt only). On
1818 * success, we must free the page and try again, thus
1819 * ensuring that vmstats.v_*_free_min counters are replenished.
1822 if (curthread->td_preempted) {
1823 kprintf("vm_page_alloc(): warning, attempt to allocate"
1824 " cache page from preempting interrupt\n");
1827 m = vm_page_select_cache(pg_color);
1830 m = vm_page_select_cache(pg_color);
1833 * On success move the page into the free queue and loop.
1835 * Only do this if we can safely acquire the vm_object lock,
1836 * because this is effectively a random page and the caller
1837 * might be holding the lock shared, we don't want to
1841 KASSERT(m->dirty == 0,
1842 ("Found dirty cache page %p", m));
1843 if ((obj = m->object) != NULL) {
1844 if (vm_object_hold_try(obj)) {
1845 vm_page_protect(m, VM_PROT_NONE);
1847 /* m->object NULL here */
1848 vm_object_drop(obj);
1850 vm_page_deactivate(m);
1854 vm_page_protect(m, VM_PROT_NONE);
1861 * On failure return NULL
1863 atomic_add_int(&vm_pageout_deficit, 1);
1864 pagedaemon_wakeup();
1868 * No pages available, wakeup the pageout daemon and give up.
1870 atomic_add_int(&vm_pageout_deficit, 1);
1871 pagedaemon_wakeup();
1876 * v_free_count can race so loop if we don't find the expected
1885 * Good page found. The page has already been busied for us and
1886 * removed from its queues.
1888 KASSERT(m->dirty == 0,
1889 ("vm_page_alloc: free/cache page %p was dirty", m));
1890 KKASSERT(m->queue == PQ_NONE);
1896 * Initialize the structure, inheriting some flags but clearing
1897 * all the rest. The page has already been busied for us.
1899 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK);
1901 KKASSERT(m->wire_count == 0);
1902 KKASSERT((m->busy_count & PBUSY_MASK) == 0);
1907 * Caller must be holding the object lock (asserted by
1908 * vm_page_insert()).
1910 * NOTE: Inserting a page here does not insert it into any pmaps
1911 * (which could cause us to block allocating memory).
1913 * NOTE: If no object an unassociated page is allocated, m->pindex
1914 * can be used by the caller for any purpose.
1917 if (vm_page_insert(m, object, pindex) == FALSE) {
1919 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1920 panic("PAGE RACE %p[%ld]/%p",
1921 object, (long)pindex, m);
1929 * Don't wakeup too often - wakeup the pageout daemon when
1930 * we would be nearly out of memory.
1932 pagedaemon_wakeup();
1935 * A BUSY page is returned.
1941 * Returns number of pages available in our DMA memory reserve
1942 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf)
1945 vm_contig_avail_pages(void)
1950 spin_lock(&vm_contig_spin);
1951 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
1952 spin_unlock(&vm_contig_spin);
1958 * Attempt to allocate contiguous physical memory with the specified
1962 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
1963 unsigned long alignment, unsigned long boundary,
1964 unsigned long size, vm_memattr_t memattr)
1970 alignment >>= PAGE_SHIFT;
1973 boundary >>= PAGE_SHIFT;
1976 size = (size + PAGE_MASK) >> PAGE_SHIFT;
1978 spin_lock(&vm_contig_spin);
1979 blk = alist_alloc(&vm_contig_alist, 0, size);
1980 if (blk == ALIST_BLOCK_NONE) {
1981 spin_unlock(&vm_contig_spin);
1983 kprintf("vm_page_alloc_contig: %ldk nospace\n",
1984 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1988 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
1989 alist_free(&vm_contig_alist, blk, size);
1990 spin_unlock(&vm_contig_spin);
1992 kprintf("vm_page_alloc_contig: %ldk high "
1994 (size + PAGE_MASK) * (PAGE_SIZE / 1024),
1999 spin_unlock(&vm_contig_spin);
2000 if (vm_contig_verbose) {
2001 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
2002 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT,
2003 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
2006 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
2007 if (memattr != VM_MEMATTR_DEFAULT) {
2008 for (i = 0;i < size; i++)
2009 pmap_page_set_memattr(&m[i], memattr);
2015 * Free contiguously allocated pages. The pages will be wired but not busy.
2016 * When freeing to the alist we leave them wired and not busy.
2019 vm_page_free_contig(vm_page_t m, unsigned long size)
2021 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
2022 vm_pindex_t start = pa >> PAGE_SHIFT;
2023 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
2025 if (vm_contig_verbose) {
2026 kprintf("vm_page_free_contig: %016jx/%ldk\n",
2027 (intmax_t)pa, size / 1024);
2029 if (pa < vm_low_phys_reserved) {
2030 KKASSERT(pa + size <= vm_low_phys_reserved);
2031 spin_lock(&vm_contig_spin);
2032 alist_free(&vm_contig_alist, start, pages);
2033 spin_unlock(&vm_contig_spin);
2036 vm_page_busy_wait(m, FALSE, "cpgfr");
2037 vm_page_unwire(m, 0);
2048 * Wait for sufficient free memory for nominal heavy memory use kernel
2051 * WARNING! Be sure never to call this in any vm_pageout code path, which
2052 * will trivially deadlock the system.
2055 vm_wait_nominal(void)
2057 while (vm_page_count_min(0))
2062 * Test if vm_wait_nominal() would block.
2065 vm_test_nominal(void)
2067 if (vm_page_count_min(0))
2073 * Block until free pages are available for allocation, called in various
2074 * places before memory allocations.
2076 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
2077 * more generous then that.
2083 * never wait forever
2087 lwkt_gettoken(&vm_token);
2089 if (curthread == pagethread ||
2090 curthread == emergpager) {
2092 * The pageout daemon itself needs pages, this is bad.
2094 if (vm_page_count_min(0)) {
2095 vm_pageout_pages_needed = 1;
2096 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
2100 * Wakeup the pageout daemon if necessary and wait.
2102 * Do not wait indefinitely for the target to be reached,
2103 * as load might prevent it from being reached any time soon.
2104 * But wait a little to try to slow down page allocations
2105 * and to give more important threads (the pagedaemon)
2106 * allocation priority.
2108 if (vm_page_count_target()) {
2109 if (vm_pages_needed == 0) {
2110 vm_pages_needed = 1;
2111 wakeup(&vm_pages_needed);
2113 ++vm_pages_waiting; /* SMP race ok */
2114 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
2117 lwkt_reltoken(&vm_token);
2121 * Block until free pages are available for allocation
2123 * Called only from vm_fault so that processes page faulting can be
2127 vm_wait_pfault(void)
2130 * Wakeup the pageout daemon if necessary and wait.
2132 * Do not wait indefinitely for the target to be reached,
2133 * as load might prevent it from being reached any time soon.
2134 * But wait a little to try to slow down page allocations
2135 * and to give more important threads (the pagedaemon)
2136 * allocation priority.
2138 if (vm_page_count_min(0)) {
2139 lwkt_gettoken(&vm_token);
2140 while (vm_page_count_severe()) {
2141 if (vm_page_count_target()) {
2144 if (vm_pages_needed == 0) {
2145 vm_pages_needed = 1;
2146 wakeup(&vm_pages_needed);
2148 ++vm_pages_waiting; /* SMP race ok */
2149 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
2152 * Do not stay stuck in the loop if the system is trying
2153 * to kill the process.
2156 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
2160 lwkt_reltoken(&vm_token);
2165 * Put the specified page on the active list (if appropriate). Ensure
2166 * that act_count is at least ACT_INIT but do not otherwise mess with it.
2168 * The caller should be holding the page busied ? XXX
2169 * This routine may not block.
2172 vm_page_activate(vm_page_t m)
2176 vm_page_spin_lock(m);
2177 if (m->queue - m->pc != PQ_ACTIVE) {
2178 _vm_page_queue_spin_lock(m);
2179 oqueue = _vm_page_rem_queue_spinlocked(m);
2180 /* page is left spinlocked, queue is unlocked */
2182 if (oqueue == PQ_CACHE)
2183 mycpu->gd_cnt.v_reactivated++;
2184 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2185 if (m->act_count < ACT_INIT)
2186 m->act_count = ACT_INIT;
2187 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
2189 _vm_page_and_queue_spin_unlock(m);
2190 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
2191 pagedaemon_wakeup();
2193 if (m->act_count < ACT_INIT)
2194 m->act_count = ACT_INIT;
2195 vm_page_spin_unlock(m);
2200 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2201 * routine is called when a page has been added to the cache or free
2204 * This routine may not block.
2206 static __inline void
2207 vm_page_free_wakeup(void)
2209 globaldata_t gd = mycpu;
2212 * If the pageout daemon itself needs pages, then tell it that
2213 * there are some free.
2215 if (vm_pageout_pages_needed &&
2216 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >=
2217 gd->gd_vmstats.v_pageout_free_min
2219 vm_pageout_pages_needed = 0;
2220 wakeup(&vm_pageout_pages_needed);
2224 * Wakeup processes that are waiting on memory.
2226 * Generally speaking we want to wakeup stuck processes as soon as
2227 * possible. !vm_page_count_min(0) is the absolute minimum point
2228 * where we can do this. Wait a bit longer to reduce degenerate
2229 * re-blocking (vm_page_free_hysteresis). The target check is just
2230 * to make sure the min-check w/hysteresis does not exceed the
2233 if (vm_pages_waiting) {
2234 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2235 !vm_page_count_target()) {
2236 vm_pages_waiting = 0;
2237 wakeup(&vmstats.v_free_count);
2238 ++mycpu->gd_cnt.v_ppwakeups;
2241 if (!vm_page_count_target()) {
2243 * Plenty of pages are free, wakeup everyone.
2245 vm_pages_waiting = 0;
2246 wakeup(&vmstats.v_free_count);
2247 ++mycpu->gd_cnt.v_ppwakeups;
2248 } else if (!vm_page_count_min(0)) {
2250 * Some pages are free, wakeup someone.
2252 int wcount = vm_pages_waiting;
2255 vm_pages_waiting = wcount;
2256 wakeup_one(&vmstats.v_free_count);
2257 ++mycpu->gd_cnt.v_ppwakeups;
2264 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
2265 * it from its VM object.
2267 * The vm_page must be BUSY on entry. BUSY will be released on
2268 * return (the page will have been freed).
2271 vm_page_free_toq(vm_page_t m)
2273 mycpu->gd_cnt.v_tfree++;
2274 KKASSERT((m->flags & PG_MAPPED) == 0);
2275 KKASSERT(m->busy_count & PBUSY_LOCKED);
2277 if ((m->busy_count & PBUSY_MASK) || ((m->queue - m->pc) == PQ_FREE)) {
2278 kprintf("vm_page_free: pindex(%lu), busy %08x, "
2280 (u_long)m->pindex, m->busy_count, m->hold_count);
2281 if ((m->queue - m->pc) == PQ_FREE)
2282 panic("vm_page_free: freeing free page");
2284 panic("vm_page_free: freeing busy page");
2288 * Remove from object, spinlock the page and its queues and
2289 * remove from any queue. No queue spinlock will be held
2290 * after this section (because the page was removed from any
2294 vm_page_and_queue_spin_lock(m);
2295 _vm_page_rem_queue_spinlocked(m);
2298 * No further management of fictitious pages occurs beyond object
2299 * and queue removal.
2301 if ((m->flags & PG_FICTITIOUS) != 0) {
2302 vm_page_spin_unlock(m);
2310 if (m->wire_count != 0) {
2311 if (m->wire_count > 1) {
2313 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2314 m->wire_count, (long)m->pindex);
2316 panic("vm_page_free: freeing wired page");
2320 * Clear the UNMANAGED flag when freeing an unmanaged page.
2321 * Clear the NEED_COMMIT flag
2323 if (m->flags & PG_UNMANAGED)
2324 vm_page_flag_clear(m, PG_UNMANAGED);
2325 if (m->flags & PG_NEED_COMMIT)
2326 vm_page_flag_clear(m, PG_NEED_COMMIT);
2328 if (m->hold_count != 0) {
2329 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2331 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1);
2335 * This sequence allows us to clear BUSY while still holding
2336 * its spin lock, which reduces contention vs allocators. We
2337 * must not leave the queue locked or _vm_page_wakeup() may
2340 _vm_page_queue_spin_unlock(m);
2341 if (_vm_page_wakeup(m)) {
2342 vm_page_spin_unlock(m);
2345 vm_page_spin_unlock(m);
2347 vm_page_free_wakeup();
2351 * vm_page_unmanage()
2353 * Prevent PV management from being done on the page. The page is
2354 * removed from the paging queues as if it were wired, and as a
2355 * consequence of no longer being managed the pageout daemon will not
2356 * touch it (since there is no way to locate the pte mappings for the
2357 * page). madvise() calls that mess with the pmap will also no longer
2358 * operate on the page.
2360 * Beyond that the page is still reasonably 'normal'. Freeing the page
2361 * will clear the flag.
2363 * This routine is used by OBJT_PHYS objects - objects using unswappable
2364 * physical memory as backing store rather then swap-backed memory and
2365 * will eventually be extended to support 4MB unmanaged physical
2368 * Caller must be holding the page busy.
2371 vm_page_unmanage(vm_page_t m)
2373 KKASSERT(m->busy_count & PBUSY_LOCKED);
2374 if ((m->flags & PG_UNMANAGED) == 0) {
2375 if (m->wire_count == 0)
2378 vm_page_flag_set(m, PG_UNMANAGED);
2382 * Mark this page as wired down by yet another map, removing it from
2383 * paging queues as necessary.
2385 * Caller must be holding the page busy.
2388 vm_page_wire(vm_page_t m)
2391 * Only bump the wire statistics if the page is not already wired,
2392 * and only unqueue the page if it is on some queue (if it is unmanaged
2393 * it is already off the queues). Don't do anything with fictitious
2394 * pages because they are always wired.
2396 KKASSERT(m->busy_count & PBUSY_LOCKED);
2397 if ((m->flags & PG_FICTITIOUS) == 0) {
2398 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2399 if ((m->flags & PG_UNMANAGED) == 0)
2401 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count, 1);
2403 KASSERT(m->wire_count != 0,
2404 ("vm_page_wire: wire_count overflow m=%p", m));
2409 * Release one wiring of this page, potentially enabling it to be paged again.
2411 * Many pages placed on the inactive queue should actually go
2412 * into the cache, but it is difficult to figure out which. What
2413 * we do instead, if the inactive target is well met, is to put
2414 * clean pages at the head of the inactive queue instead of the tail.
2415 * This will cause them to be moved to the cache more quickly and
2416 * if not actively re-referenced, freed more quickly. If we just
2417 * stick these pages at the end of the inactive queue, heavy filesystem
2418 * meta-data accesses can cause an unnecessary paging load on memory bound
2419 * processes. This optimization causes one-time-use metadata to be
2420 * reused more quickly.
2422 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2423 * the inactive queue. This helps the pageout daemon determine memory
2424 * pressure and act on out-of-memory situations more quickly.
2426 * BUT, if we are in a low-memory situation we have no choice but to
2427 * put clean pages on the cache queue.
2429 * A number of routines use vm_page_unwire() to guarantee that the page
2430 * will go into either the inactive or active queues, and will NEVER
2431 * be placed in the cache - for example, just after dirtying a page.
2432 * dirty pages in the cache are not allowed.
2434 * This routine may not block.
2437 vm_page_unwire(vm_page_t m, int activate)
2439 KKASSERT(m->busy_count & PBUSY_LOCKED);
2440 if (m->flags & PG_FICTITIOUS) {
2442 } else if (m->wire_count <= 0) {
2443 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2445 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2446 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count,-1);
2447 if (m->flags & PG_UNMANAGED) {
2449 } else if (activate || (m->flags & PG_NEED_COMMIT)) {
2450 vm_page_spin_lock(m);
2451 _vm_page_add_queue_spinlocked(m,
2452 PQ_ACTIVE + m->pc, 0);
2453 _vm_page_and_queue_spin_unlock(m);
2455 vm_page_spin_lock(m);
2456 vm_page_flag_clear(m, PG_WINATCFLS);
2457 _vm_page_add_queue_spinlocked(m,
2458 PQ_INACTIVE + m->pc, 0);
2459 ++vm_swapcache_inactive_heuristic;
2460 _vm_page_and_queue_spin_unlock(m);
2467 * Move the specified page to the inactive queue. If the page has
2468 * any associated swap, the swap is deallocated.
2470 * Normally athead is 0 resulting in LRU operation. athead is set
2471 * to 1 if we want this page to be 'as if it were placed in the cache',
2472 * except without unmapping it from the process address space.
2474 * vm_page's spinlock must be held on entry and will remain held on return.
2475 * This routine may not block.
2478 _vm_page_deactivate_locked(vm_page_t m, int athead)
2483 * Ignore if already inactive.
2485 if (m->queue - m->pc == PQ_INACTIVE)
2487 _vm_page_queue_spin_lock(m);
2488 oqueue = _vm_page_rem_queue_spinlocked(m);
2490 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2491 if (oqueue == PQ_CACHE)
2492 mycpu->gd_cnt.v_reactivated++;
2493 vm_page_flag_clear(m, PG_WINATCFLS);
2494 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2496 ++vm_swapcache_inactive_heuristic;
2498 /* NOTE: PQ_NONE if condition not taken */
2499 _vm_page_queue_spin_unlock(m);
2500 /* leaves vm_page spinlocked */
2504 * Attempt to deactivate a page.
2509 vm_page_deactivate(vm_page_t m)
2511 vm_page_spin_lock(m);
2512 _vm_page_deactivate_locked(m, 0);
2513 vm_page_spin_unlock(m);
2517 vm_page_deactivate_locked(vm_page_t m)
2519 _vm_page_deactivate_locked(m, 0);
2523 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it.
2525 * This function returns non-zero if it successfully moved the page to
2528 * This function unconditionally unbusies the page on return.
2531 vm_page_try_to_cache(vm_page_t m)
2533 vm_page_spin_lock(m);
2534 if (m->dirty || m->hold_count || m->wire_count ||
2535 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2536 if (_vm_page_wakeup(m)) {
2537 vm_page_spin_unlock(m);
2540 vm_page_spin_unlock(m);
2544 vm_page_spin_unlock(m);
2547 * Page busied by us and no longer spinlocked. Dirty pages cannot
2548 * be moved to the cache.
2550 vm_page_test_dirty(m);
2551 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2560 * Attempt to free the page. If we cannot free it, we do nothing.
2561 * 1 is returned on success, 0 on failure.
2566 vm_page_try_to_free(vm_page_t m)
2568 vm_page_spin_lock(m);
2569 if (vm_page_busy_try(m, TRUE)) {
2570 vm_page_spin_unlock(m);
2575 * The page can be in any state, including already being on the free
2576 * queue. Check to see if it really can be freed.
2578 if (m->dirty || /* can't free if it is dirty */
2579 m->hold_count || /* or held (XXX may be wrong) */
2580 m->wire_count || /* or wired */
2581 (m->flags & (PG_UNMANAGED | /* or unmanaged */
2582 PG_NEED_COMMIT)) || /* or needs a commit */
2583 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */
2584 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */
2585 if (_vm_page_wakeup(m)) {
2586 vm_page_spin_unlock(m);
2589 vm_page_spin_unlock(m);
2593 vm_page_spin_unlock(m);
2596 * We can probably free the page.
2598 * Page busied by us and no longer spinlocked. Dirty pages will
2599 * not be freed by this function. We have to re-test the
2600 * dirty bit after cleaning out the pmaps.
2602 vm_page_test_dirty(m);
2603 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2607 vm_page_protect(m, VM_PROT_NONE);
2608 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2619 * Put the specified page onto the page cache queue (if appropriate).
2621 * The page must be busy, and this routine will release the busy and
2622 * possibly even free the page.
2625 vm_page_cache(vm_page_t m)
2628 * Not suitable for the cache
2630 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2631 (m->busy_count & PBUSY_MASK) ||
2632 m->wire_count || m->hold_count) {
2638 * Already in the cache (and thus not mapped)
2640 if ((m->queue - m->pc) == PQ_CACHE) {
2641 KKASSERT((m->flags & PG_MAPPED) == 0);
2647 * Caller is required to test m->dirty, but note that the act of
2648 * removing the page from its maps can cause it to become dirty
2649 * on an SMP system due to another cpu running in usermode.
2652 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2657 * Remove all pmaps and indicate that the page is not
2658 * writeable or mapped. Our vm_page_protect() call may
2659 * have blocked (especially w/ VM_PROT_NONE), so recheck
2662 vm_page_protect(m, VM_PROT_NONE);
2663 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2664 (m->busy_count & PBUSY_MASK) ||
2665 m->wire_count || m->hold_count) {
2667 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2668 vm_page_deactivate(m);
2671 _vm_page_and_queue_spin_lock(m);
2672 _vm_page_rem_queue_spinlocked(m);
2673 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2674 _vm_page_queue_spin_unlock(m);
2675 if (_vm_page_wakeup(m)) {
2676 vm_page_spin_unlock(m);
2679 vm_page_spin_unlock(m);
2681 vm_page_free_wakeup();
2686 * vm_page_dontneed()
2688 * Cache, deactivate, or do nothing as appropriate. This routine
2689 * is typically used by madvise() MADV_DONTNEED.
2691 * Generally speaking we want to move the page into the cache so
2692 * it gets reused quickly. However, this can result in a silly syndrome
2693 * due to the page recycling too quickly. Small objects will not be
2694 * fully cached. On the otherhand, if we move the page to the inactive
2695 * queue we wind up with a problem whereby very large objects
2696 * unnecessarily blow away our inactive and cache queues.
2698 * The solution is to move the pages based on a fixed weighting. We
2699 * either leave them alone, deactivate them, or move them to the cache,
2700 * where moving them to the cache has the highest weighting.
2701 * By forcing some pages into other queues we eventually force the
2702 * system to balance the queues, potentially recovering other unrelated
2703 * space from active. The idea is to not force this to happen too
2706 * The page must be busied.
2709 vm_page_dontneed(vm_page_t m)
2711 static int dnweight;
2718 * occassionally leave the page alone
2720 if ((dnw & 0x01F0) == 0 ||
2721 m->queue - m->pc == PQ_INACTIVE ||
2722 m->queue - m->pc == PQ_CACHE
2724 if (m->act_count >= ACT_INIT)
2730 * If vm_page_dontneed() is inactivating a page, it must clear
2731 * the referenced flag; otherwise the pagedaemon will see references
2732 * on the page in the inactive queue and reactivate it. Until the
2733 * page can move to the cache queue, madvise's job is not done.
2735 vm_page_flag_clear(m, PG_REFERENCED);
2736 pmap_clear_reference(m);
2739 vm_page_test_dirty(m);
2741 if (m->dirty || (dnw & 0x0070) == 0) {
2743 * Deactivate the page 3 times out of 32.
2748 * Cache the page 28 times out of every 32. Note that
2749 * the page is deactivated instead of cached, but placed
2750 * at the head of the queue instead of the tail.
2754 vm_page_spin_lock(m);
2755 _vm_page_deactivate_locked(m, head);
2756 vm_page_spin_unlock(m);
2760 * These routines manipulate the 'soft busy' count for a page. A soft busy
2761 * is almost like a hard BUSY except that it allows certain compatible
2762 * operations to occur on the page while it is busy. For example, a page
2763 * undergoing a write can still be mapped read-only.
2765 * We also use soft-busy to quickly pmap_enter shared read-only pages
2766 * without having to hold the page locked.
2768 * The soft-busy count can be > 1 in situations where multiple threads
2769 * are pmap_enter()ing the same page simultaneously, or when two buffer
2770 * cache buffers overlap the same page.
2772 * The caller must hold the page BUSY when making these two calls.
2775 vm_page_io_start(vm_page_t m)
2779 ocount = atomic_fetchadd_int(&m->busy_count, 1);
2780 KKASSERT(ocount & PBUSY_LOCKED);
2784 vm_page_io_finish(vm_page_t m)
2788 ocount = atomic_fetchadd_int(&m->busy_count, -1);
2789 KKASSERT(ocount & PBUSY_MASK);
2791 if (((ocount - 1) & (PBUSY_LOCKED | PBUSY_MASK)) == 0)
2797 * Attempt to soft-busy a page. The page must not be PBUSY_LOCKED.
2799 * Returns 0 on success, non-zero on failure.
2802 vm_page_sbusy_try(vm_page_t m)
2806 if (m->busy_count & PBUSY_LOCKED)
2808 ocount = atomic_fetchadd_int(&m->busy_count, 1);
2809 if (ocount & PBUSY_LOCKED) {
2810 vm_page_sbusy_drop(m);
2817 * Indicate that a clean VM page requires a filesystem commit and cannot
2818 * be reused. Used by tmpfs.
2821 vm_page_need_commit(vm_page_t m)
2823 vm_page_flag_set(m, PG_NEED_COMMIT);
2824 vm_object_set_writeable_dirty(m->object);
2828 vm_page_clear_commit(vm_page_t m)
2830 vm_page_flag_clear(m, PG_NEED_COMMIT);
2834 * Grab a page, blocking if it is busy and allocating a page if necessary.
2835 * A busy page is returned or NULL. The page may or may not be valid and
2836 * might not be on a queue (the caller is responsible for the disposition of
2839 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2840 * page will be zero'd and marked valid.
2842 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2843 * valid even if it already exists.
2845 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2846 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2847 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2849 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2850 * always returned if we had blocked.
2852 * This routine may not be called from an interrupt.
2854 * No other requirements.
2857 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2863 KKASSERT(allocflags &
2864 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2865 vm_object_hold_shared(object);
2867 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2869 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2870 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2875 } else if (m == NULL) {
2877 vm_object_upgrade(object);
2880 if (allocflags & VM_ALLOC_RETRY)
2881 allocflags |= VM_ALLOC_NULL_OK;
2882 m = vm_page_alloc(object, pindex,
2883 allocflags & ~VM_ALLOC_RETRY);
2887 if ((allocflags & VM_ALLOC_RETRY) == 0)
2896 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2898 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2899 * valid even if already valid.
2901 * NOTE! We have removed all of the PG_ZERO optimizations and also
2902 * removed the idle zeroing code. These optimizations actually
2903 * slow things down on modern cpus because the zerod area is
2904 * likely uncached, placing a memory-access burden on the
2905 * accesors taking the fault.
2907 * By always zeroing the page in-line with the fault, no
2908 * dynamic ram reads are needed and the caches are hot, ready
2909 * for userland to access the memory.
2911 if (m->valid == 0) {
2912 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2913 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2914 m->valid = VM_PAGE_BITS_ALL;
2916 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2917 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2918 m->valid = VM_PAGE_BITS_ALL;
2921 vm_object_drop(object);
2926 * Mapping function for valid bits or for dirty bits in
2927 * a page. May not block.
2929 * Inputs are required to range within a page.
2935 vm_page_bits(int base, int size)
2941 base + size <= PAGE_SIZE,
2942 ("vm_page_bits: illegal base/size %d/%d", base, size)
2945 if (size == 0) /* handle degenerate case */
2948 first_bit = base >> DEV_BSHIFT;
2949 last_bit = (base + size - 1) >> DEV_BSHIFT;
2951 return ((2 << last_bit) - (1 << first_bit));
2955 * Sets portions of a page valid and clean. The arguments are expected
2956 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2957 * of any partial chunks touched by the range. The invalid portion of
2958 * such chunks will be zero'd.
2960 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2961 * align base to DEV_BSIZE so as not to mark clean a partially
2962 * truncated device block. Otherwise the dirty page status might be
2965 * This routine may not block.
2967 * (base + size) must be less then or equal to PAGE_SIZE.
2970 _vm_page_zero_valid(vm_page_t m, int base, int size)
2975 if (size == 0) /* handle degenerate case */
2979 * If the base is not DEV_BSIZE aligned and the valid
2980 * bit is clear, we have to zero out a portion of the
2984 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2985 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2987 pmap_zero_page_area(
2995 * If the ending offset is not DEV_BSIZE aligned and the
2996 * valid bit is clear, we have to zero out a portion of
3000 endoff = base + size;
3002 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
3003 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
3005 pmap_zero_page_area(
3008 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
3014 * Set valid, clear dirty bits. If validating the entire
3015 * page we can safely clear the pmap modify bit. We also
3016 * use this opportunity to clear the PG_NOSYNC flag. If a process
3017 * takes a write fault on a MAP_NOSYNC memory area the flag will
3020 * We set valid bits inclusive of any overlap, but we can only
3021 * clear dirty bits for DEV_BSIZE chunks that are fully within
3024 * Page must be busied?
3025 * No other requirements.
3028 vm_page_set_valid(vm_page_t m, int base, int size)
3030 _vm_page_zero_valid(m, base, size);
3031 m->valid |= vm_page_bits(base, size);
3036 * Set valid bits and clear dirty bits.
3038 * Page must be busied by caller.
3040 * NOTE: This function does not clear the pmap modified bit.
3041 * Also note that e.g. NFS may use a byte-granular base
3044 * No other requirements.
3047 vm_page_set_validclean(vm_page_t m, int base, int size)
3051 _vm_page_zero_valid(m, base, size);
3052 pagebits = vm_page_bits(base, size);
3053 m->valid |= pagebits;
3054 m->dirty &= ~pagebits;
3055 if (base == 0 && size == PAGE_SIZE) {
3056 /*pmap_clear_modify(m);*/
3057 vm_page_flag_clear(m, PG_NOSYNC);
3062 * Set valid & dirty. Used by buwrite()
3064 * Page must be busied by caller.
3067 vm_page_set_validdirty(vm_page_t m, int base, int size)
3071 pagebits = vm_page_bits(base, size);
3072 m->valid |= pagebits;
3073 m->dirty |= pagebits;
3075 vm_object_set_writeable_dirty(m->object);
3081 * NOTE: This function does not clear the pmap modified bit.
3082 * Also note that e.g. NFS may use a byte-granular base
3085 * Page must be busied?
3086 * No other requirements.
3089 vm_page_clear_dirty(vm_page_t m, int base, int size)
3091 m->dirty &= ~vm_page_bits(base, size);
3092 if (base == 0 && size == PAGE_SIZE) {
3093 /*pmap_clear_modify(m);*/
3094 vm_page_flag_clear(m, PG_NOSYNC);
3099 * Make the page all-dirty.
3101 * Also make sure the related object and vnode reflect the fact that the
3102 * object may now contain a dirty page.
3104 * Page must be busied?
3105 * No other requirements.
3108 vm_page_dirty(vm_page_t m)
3111 int pqtype = m->queue - m->pc;
3113 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
3114 ("vm_page_dirty: page in free/cache queue!"));
3115 if (m->dirty != VM_PAGE_BITS_ALL) {
3116 m->dirty = VM_PAGE_BITS_ALL;
3118 vm_object_set_writeable_dirty(m->object);
3123 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3124 * valid and dirty bits for the effected areas are cleared.
3126 * Page must be busied?
3128 * No other requirements.
3131 vm_page_set_invalid(vm_page_t m, int base, int size)
3135 bits = vm_page_bits(base, size);
3138 atomic_add_int(&m->object->generation, 1);
3142 * The kernel assumes that the invalid portions of a page contain
3143 * garbage, but such pages can be mapped into memory by user code.
3144 * When this occurs, we must zero out the non-valid portions of the
3145 * page so user code sees what it expects.
3147 * Pages are most often semi-valid when the end of a file is mapped
3148 * into memory and the file's size is not page aligned.
3150 * Page must be busied?
3151 * No other requirements.
3154 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3160 * Scan the valid bits looking for invalid sections that
3161 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
3162 * valid bit may be set ) have already been zerod by
3163 * vm_page_set_validclean().
3165 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3166 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3167 (m->valid & (1 << i))
3170 pmap_zero_page_area(
3173 (i - b) << DEV_BSHIFT
3181 * setvalid is TRUE when we can safely set the zero'd areas
3182 * as being valid. We can do this if there are no cache consistency
3183 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3186 m->valid = VM_PAGE_BITS_ALL;
3190 * Is a (partial) page valid? Note that the case where size == 0
3191 * will return FALSE in the degenerate case where the page is entirely
3192 * invalid, and TRUE otherwise.
3195 * No other requirements.
3198 vm_page_is_valid(vm_page_t m, int base, int size)
3200 int bits = vm_page_bits(base, size);
3202 if (m->valid && ((m->valid & bits) == bits))
3209 * update dirty bits from pmap/mmu. May not block.
3211 * Caller must hold the page busy
3214 vm_page_test_dirty(vm_page_t m)
3216 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
3221 #include "opt_ddb.h"
3223 #include <ddb/ddb.h>
3225 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3227 db_printf("vmstats.v_free_count: %ld\n", vmstats.v_free_count);
3228 db_printf("vmstats.v_cache_count: %ld\n", vmstats.v_cache_count);
3229 db_printf("vmstats.v_inactive_count: %ld\n", vmstats.v_inactive_count);
3230 db_printf("vmstats.v_active_count: %ld\n", vmstats.v_active_count);
3231 db_printf("vmstats.v_wire_count: %ld\n", vmstats.v_wire_count);
3232 db_printf("vmstats.v_free_reserved: %ld\n", vmstats.v_free_reserved);
3233 db_printf("vmstats.v_free_min: %ld\n", vmstats.v_free_min);
3234 db_printf("vmstats.v_free_target: %ld\n", vmstats.v_free_target);
3235 db_printf("vmstats.v_cache_min: %ld\n", vmstats.v_cache_min);
3236 db_printf("vmstats.v_inactive_target: %ld\n",
3237 vmstats.v_inactive_target);
3240 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3243 db_printf("PQ_FREE:");
3244 for (i = 0; i < PQ_L2_SIZE; i++) {
3245 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
3249 db_printf("PQ_CACHE:");
3250 for(i = 0; i < PQ_L2_SIZE; i++) {
3251 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
3255 db_printf("PQ_ACTIVE:");
3256 for(i = 0; i < PQ_L2_SIZE; i++) {
3257 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3261 db_printf("PQ_INACTIVE:");
3262 for(i = 0; i < PQ_L2_SIZE; i++) {
3263 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);