4 * Copyright (c) 1991 Regents of the University of California.
7 * This code is derived from software contributed to Berkeley by
8 * The Mach Operating System project at Carnegie-Mellon University.
10 * Redistribution and use in source and binary forms, with or without
11 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in the
17 * documentation and/or other materials provided with the distribution.
18 * 3. Neither the name of the University nor the names of its contributors
19 * may be used to endorse or promote products derived from this software
20 * without specific prior written permission.
22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
34 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
35 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $
39 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
40 * All rights reserved.
42 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
44 * Permission to use, copy, modify and distribute this software and
45 * its documentation is hereby granted, provided that both the copyright
46 * notice and this permission notice appear in all copies of the
47 * software, derivative works or modified versions, and any portions
48 * thereof, and that both notices appear in supporting documentation.
50 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
51 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
52 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
54 * Carnegie Mellon requests users of this software to return to
56 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
57 * School of Computer Science
58 * Carnegie Mellon University
59 * Pittsburgh PA 15213-3890
61 * any improvements or extensions that they make and grant Carnegie the
62 * rights to redistribute these changes.
65 * Resident memory management module. The module manipulates 'VM pages'.
66 * A VM page is the core building block for memory management.
69 #include <sys/param.h>
70 #include <sys/systm.h>
71 #include <sys/malloc.h>
73 #include <sys/vmmeter.h>
74 #include <sys/vnode.h>
75 #include <sys/kernel.h>
76 #include <sys/alist.h>
77 #include <sys/sysctl.h>
78 #include <sys/cpu_topology.h>
81 #include <vm/vm_param.h>
83 #include <vm/vm_kern.h>
85 #include <vm/vm_map.h>
86 #include <vm/vm_object.h>
87 #include <vm/vm_page.h>
88 #include <vm/vm_pageout.h>
89 #include <vm/vm_pager.h>
90 #include <vm/vm_extern.h>
91 #include <vm/swap_pager.h>
93 #include <machine/inttypes.h>
94 #include <machine/md_var.h>
95 #include <machine/specialreg.h>
97 #include <vm/vm_page2.h>
98 #include <sys/spinlock2.h>
101 * Action hash for user umtx support.
103 #define VMACTION_HSIZE 256
104 #define VMACTION_HMASK (VMACTION_HSIZE - 1)
107 * SET - Minimum required set associative size, must be a power of 2. We
108 * want this to match or exceed the set-associativeness of the cpu.
110 * GRP - A larger set that allows bleed-over into the domains of other
111 * nearby cpus. Also must be a power of 2. Used by the page zeroing
112 * code to smooth things out a bit.
114 #define PQ_SET_ASSOC 16
115 #define PQ_SET_ASSOC_MASK (PQ_SET_ASSOC - 1)
117 #define PQ_GRP_ASSOC (PQ_SET_ASSOC * 2)
118 #define PQ_GRP_ASSOC_MASK (PQ_GRP_ASSOC - 1)
120 static void vm_page_queue_init(void);
121 static void vm_page_free_wakeup(void);
122 static vm_page_t vm_page_select_cache(u_short pg_color);
123 static vm_page_t _vm_page_list_find2(int basequeue, int index);
124 static void _vm_page_deactivate_locked(vm_page_t m, int athead);
127 * Array of tailq lists
129 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT];
131 LIST_HEAD(vm_page_action_list, vm_page_action);
132 struct vm_page_action_list action_list[VMACTION_HSIZE];
133 static volatile int vm_pages_waiting;
135 static struct alist vm_contig_alist;
136 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536];
137 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin");
139 static u_long vm_dma_reserved = 0;
140 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved);
141 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0,
142 "Memory reserved for DMA");
143 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD,
144 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA");
146 static int vm_contig_verbose = 0;
147 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose);
149 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
150 vm_pindex_t, pindex);
153 vm_page_queue_init(void)
157 for (i = 0; i < PQ_L2_SIZE; i++)
158 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
159 for (i = 0; i < PQ_L2_SIZE; i++)
160 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
161 for (i = 0; i < PQ_L2_SIZE; i++)
162 vm_page_queues[PQ_INACTIVE+i].cnt = &vmstats.v_inactive_count;
163 for (i = 0; i < PQ_L2_SIZE; i++)
164 vm_page_queues[PQ_ACTIVE+i].cnt = &vmstats.v_active_count;
165 for (i = 0; i < PQ_L2_SIZE; i++)
166 vm_page_queues[PQ_HOLD+i].cnt = &vmstats.v_active_count;
167 /* PQ_NONE has no queue */
169 for (i = 0; i < PQ_COUNT; i++) {
170 TAILQ_INIT(&vm_page_queues[i].pl);
171 spin_init(&vm_page_queues[i].spin, "vm_page_queue_init");
174 for (i = 0; i < VMACTION_HSIZE; i++)
175 LIST_INIT(&action_list[i]);
179 * note: place in initialized data section? Is this necessary?
182 int vm_page_array_size = 0;
183 vm_page_t vm_page_array = NULL;
184 vm_paddr_t vm_low_phys_reserved;
189 * Sets the page size, perhaps based upon the memory size.
190 * Must be called before any use of page-size dependent functions.
193 vm_set_page_size(void)
195 if (vmstats.v_page_size == 0)
196 vmstats.v_page_size = PAGE_SIZE;
197 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
198 panic("vm_set_page_size: page size not a power of two");
204 * Add a new page to the freelist for use by the system. New pages
205 * are added to both the head and tail of the associated free page
206 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
207 * requests pull 'recent' adds (higher physical addresses) first.
209 * Beware that the page zeroing daemon will also be running soon after
210 * boot, moving pages from the head to the tail of the PQ_FREE queues.
212 * Must be called in a critical section.
215 vm_add_new_page(vm_paddr_t pa)
217 struct vpgqueues *vpq;
220 m = PHYS_TO_VM_PAGE(pa);
223 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
224 m->pat_mode = PAT_WRITE_BACK;
226 * Twist for cpu localization in addition to page coloring, so
227 * different cpus selecting by m->queue get different page colors.
229 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE) & PQ_L2_MASK;
230 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE)) & PQ_L2_MASK;
232 * Reserve a certain number of contiguous low memory pages for
233 * contigmalloc() to use.
235 if (pa < vm_low_phys_reserved) {
236 atomic_add_int(&vmstats.v_page_count, 1);
237 atomic_add_int(&vmstats.v_dma_pages, 1);
240 atomic_add_int(&vmstats.v_wire_count, 1);
241 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1);
248 m->queue = m->pc + PQ_FREE;
249 KKASSERT(m->dirty == 0);
251 atomic_add_int(&vmstats.v_page_count, 1);
252 atomic_add_int(&vmstats.v_free_count, 1);
253 vpq = &vm_page_queues[m->queue];
254 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
261 * Initializes the resident memory module.
263 * Preallocates memory for critical VM structures and arrays prior to
264 * kernel_map becoming available.
266 * Memory is allocated from (virtual2_start, virtual2_end) if available,
267 * otherwise memory is allocated from (virtual_start, virtual_end).
269 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
270 * large enough to hold vm_page_array & other structures for machines with
271 * large amounts of ram, so we want to use virtual2* when available.
274 vm_page_startup(void)
276 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start;
279 vm_paddr_t page_range;
285 vm_paddr_t biggestone, biggestsize;
291 vaddr = round_page(vaddr);
294 * Make sure ranges are page-aligned.
296 for (i = 0; phys_avail[i].phys_end; ++i) {
297 phys_avail[i].phys_beg = round_page64(phys_avail[i].phys_beg);
298 phys_avail[i].phys_end = trunc_page64(phys_avail[i].phys_end);
299 if (phys_avail[i].phys_end < phys_avail[i].phys_beg)
300 phys_avail[i].phys_end = phys_avail[i].phys_beg;
304 * Locate largest block
306 for (i = 0; phys_avail[i].phys_end; ++i) {
307 vm_paddr_t size = phys_avail[i].phys_end -
308 phys_avail[i].phys_beg;
310 if (size > biggestsize) {
317 end = phys_avail[biggestone].phys_end;
318 end = trunc_page(end);
321 * Initialize the queue headers for the free queue, the active queue
322 * and the inactive queue.
324 vm_page_queue_init();
326 #if !defined(_KERNEL_VIRTUAL)
328 * VKERNELs don't support minidumps and as such don't need
331 * Allocate a bitmap to indicate that a random physical page
332 * needs to be included in a minidump.
334 * The amd64 port needs this to indicate which direct map pages
335 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
337 * However, i386 still needs this workspace internally within the
338 * minidump code. In theory, they are not needed on i386, but are
339 * included should the sf_buf code decide to use them.
341 page_range = phys_avail[i-1].phys_end / PAGE_SIZE;
342 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
343 end -= vm_page_dump_size;
344 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
345 VM_PROT_READ | VM_PROT_WRITE);
346 bzero((void *)vm_page_dump, vm_page_dump_size);
349 * Compute the number of pages of memory that will be available for
350 * use (taking into account the overhead of a page structure per
353 first_page = phys_avail[0].phys_beg / PAGE_SIZE;
354 page_range = phys_avail[i-1].phys_end / PAGE_SIZE - first_page;
355 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
357 #ifndef _KERNEL_VIRTUAL
359 * (only applies to real kernels)
361 * Reserve a large amount of low memory for potential 32-bit DMA
362 * space allocations. Once device initialization is complete we
363 * release most of it, but keep (vm_dma_reserved) memory reserved
364 * for later use. Typically for X / graphics. Through trial and
365 * error we find that GPUs usually requires ~60-100MB or so.
367 * By default, 128M is left in reserve on machines with 2G+ of ram.
369 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT;
370 if (vm_low_phys_reserved > total / 4)
371 vm_low_phys_reserved = total / 4;
372 if (vm_dma_reserved == 0) {
373 vm_dma_reserved = 128 * 1024 * 1024; /* 128MB */
374 if (vm_dma_reserved > total / 16)
375 vm_dma_reserved = total / 16;
378 alist_init(&vm_contig_alist, 65536, vm_contig_ameta,
379 ALIST_RECORDS_65536);
382 * Initialize the mem entry structures now, and put them in the free
385 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
386 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE);
387 vm_page_array = (vm_page_t)mapped;
389 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
391 * since pmap_map on amd64 returns stuff out of a direct-map region,
392 * we have to manually add these pages to the minidump tracking so
393 * that they can be dumped, including the vm_page_array.
396 pa < phys_avail[biggestone].phys_end;
403 * Clear all of the page structures
405 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
406 vm_page_array_size = page_range;
409 * Construct the free queue(s) in ascending order (by physical
410 * address) so that the first 16MB of physical memory is allocated
411 * last rather than first. On large-memory machines, this avoids
412 * the exhaustion of low physical memory before isa_dmainit has run.
414 vmstats.v_page_count = 0;
415 vmstats.v_free_count = 0;
416 for (i = 0; phys_avail[i].phys_end && npages > 0; ++i) {
417 pa = phys_avail[i].phys_beg;
421 last_pa = phys_avail[i].phys_end;
422 while (pa < last_pa && npages-- > 0) {
428 virtual2_start = vaddr;
430 virtual_start = vaddr;
434 * We tended to reserve a ton of memory for contigmalloc(). Now that most
435 * drivers have initialized we want to return most the remaining free
436 * reserve back to the VM page queues so they can be used for normal
439 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool.
442 vm_page_startup_finish(void *dummy __unused)
451 spin_lock(&vm_contig_spin);
453 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
454 if (bfree <= vm_dma_reserved / PAGE_SIZE)
460 * Figure out how much of the initial reserve we have to
461 * free in order to reach our target.
463 bfree -= vm_dma_reserved / PAGE_SIZE;
465 blk += count - bfree;
470 * Calculate the nearest power of 2 <= count.
472 for (xcount = 1; xcount <= count; xcount <<= 1)
475 blk += count - xcount;
479 * Allocate the pages from the alist, then free them to
480 * the normal VM page queues.
482 * Pages allocated from the alist are wired. We have to
483 * busy, unwire, and free them. We must also adjust
484 * vm_low_phys_reserved before freeing any pages to prevent
487 rblk = alist_alloc(&vm_contig_alist, blk, count);
489 kprintf("vm_page_startup_finish: Unable to return "
490 "dma space @0x%08x/%d -> 0x%08x\n",
494 atomic_add_int(&vmstats.v_dma_pages, -count);
495 spin_unlock(&vm_contig_spin);
497 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
498 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m);
500 vm_page_busy_wait(m, FALSE, "cpgfr");
501 vm_page_unwire(m, 0);
506 spin_lock(&vm_contig_spin);
508 spin_unlock(&vm_contig_spin);
511 * Print out how much DMA space drivers have already allocated and
512 * how much is left over.
514 kprintf("DMA space used: %jdk, remaining available: %jdk\n",
515 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) *
517 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024));
519 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY,
520 vm_page_startup_finish, NULL);
524 * Scan comparison function for Red-Black tree scans. An inclusive
525 * (start,end) is expected. Other fields are not used.
528 rb_vm_page_scancmp(struct vm_page *p, void *data)
530 struct rb_vm_page_scan_info *info = data;
532 if (p->pindex < info->start_pindex)
534 if (p->pindex > info->end_pindex)
540 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
542 if (p1->pindex < p2->pindex)
544 if (p1->pindex > p2->pindex)
550 vm_page_init(vm_page_t m)
552 /* do nothing for now. Called from pmap_page_init() */
556 * Each page queue has its own spin lock, which is fairly optimal for
557 * allocating and freeing pages at least.
559 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
560 * queue spinlock via this function. Also note that m->queue cannot change
561 * unless both the page and queue are locked.
565 _vm_page_queue_spin_lock(vm_page_t m)
570 if (queue != PQ_NONE) {
571 spin_lock(&vm_page_queues[queue].spin);
572 KKASSERT(queue == m->queue);
578 _vm_page_queue_spin_unlock(vm_page_t m)
584 if (queue != PQ_NONE)
585 spin_unlock(&vm_page_queues[queue].spin);
590 _vm_page_queues_spin_lock(u_short queue)
593 if (queue != PQ_NONE)
594 spin_lock(&vm_page_queues[queue].spin);
600 _vm_page_queues_spin_unlock(u_short queue)
603 if (queue != PQ_NONE)
604 spin_unlock(&vm_page_queues[queue].spin);
608 vm_page_queue_spin_lock(vm_page_t m)
610 _vm_page_queue_spin_lock(m);
614 vm_page_queues_spin_lock(u_short queue)
616 _vm_page_queues_spin_lock(queue);
620 vm_page_queue_spin_unlock(vm_page_t m)
622 _vm_page_queue_spin_unlock(m);
626 vm_page_queues_spin_unlock(u_short queue)
628 _vm_page_queues_spin_unlock(queue);
632 * This locks the specified vm_page and its queue in the proper order
633 * (page first, then queue). The queue may change so the caller must
638 _vm_page_and_queue_spin_lock(vm_page_t m)
640 vm_page_spin_lock(m);
641 _vm_page_queue_spin_lock(m);
646 _vm_page_and_queue_spin_unlock(vm_page_t m)
648 _vm_page_queues_spin_unlock(m->queue);
649 vm_page_spin_unlock(m);
653 vm_page_and_queue_spin_unlock(vm_page_t m)
655 _vm_page_and_queue_spin_unlock(m);
659 vm_page_and_queue_spin_lock(vm_page_t m)
661 _vm_page_and_queue_spin_lock(m);
665 * Helper function removes vm_page from its current queue.
666 * Returns the base queue the page used to be on.
668 * The vm_page and the queue must be spinlocked.
669 * This function will unlock the queue but leave the page spinlocked.
671 static __inline u_short
672 _vm_page_rem_queue_spinlocked(vm_page_t m)
674 struct vpgqueues *pq;
679 if (queue != PQ_NONE) {
680 pq = &vm_page_queues[queue];
681 TAILQ_REMOVE(&pq->pl, m, pageq);
682 atomic_add_int(pq->cnt, -1);
686 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
688 vm_page_queues_spin_unlock(oqueue); /* intended */
694 * Helper function places the vm_page on the specified queue.
696 * The vm_page must be spinlocked.
697 * This function will return with both the page and the queue locked.
700 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
702 struct vpgqueues *pq;
704 KKASSERT(m->queue == PQ_NONE);
706 if (queue != PQ_NONE) {
707 vm_page_queues_spin_lock(queue);
708 pq = &vm_page_queues[queue];
710 atomic_add_int(pq->cnt, 1);
714 * PQ_FREE is always handled LIFO style to try to provide
715 * cache-hot pages to programs.
717 if (queue - m->pc == PQ_FREE) {
718 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
720 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
722 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
724 /* leave the queue spinlocked */
729 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
730 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
731 * did not. Only one sleep call will be made before returning.
733 * This function does NOT busy the page and on return the page is not
734 * guaranteed to be available.
737 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
745 if ((flags & PG_BUSY) == 0 &&
746 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
749 tsleep_interlock(m, 0);
750 if (atomic_cmpset_int(&m->flags, flags,
751 flags | PG_WANTED | PG_REFERENCED)) {
752 tsleep(m, PINTERLOCKED, msg, 0);
759 * This calculates and returns a page color given an optional VM object and
760 * either a pindex or an iterator. We attempt to return a cpu-localized
761 * pg_color that is still roughly 16-way set-associative. The CPU topology
762 * is used if it was probed.
764 * The caller may use the returned value to index into e.g. PQ_FREE when
765 * allocating a page in order to nominally obtain pages that are hopefully
766 * already localized to the requesting cpu. This function is not able to
767 * provide any sort of guarantee of this, but does its best to improve
768 * hardware cache management performance.
770 * WARNING! The caller must mask the returned value with PQ_L2_MASK.
773 vm_get_pg_color(globaldata_t gd, vm_object_t object, vm_pindex_t pindex)
780 phys_id = get_cpu_phys_id(gd->gd_cpuid);
781 core_id = get_cpu_core_id(gd->gd_cpuid);
782 object_pg_color = object ? object->pg_color : 0;
784 if (cpu_topology_phys_ids && cpu_topology_core_ids) {
785 int grpsize = PQ_L2_SIZE / cpu_topology_phys_ids;
787 if (grpsize / cpu_topology_core_ids >= PQ_SET_ASSOC) {
789 * Enough space for a full break-down.
791 pg_color = phys_id * grpsize;
792 pg_color += core_id * grpsize / cpu_topology_core_ids;
793 pg_color += (pindex + object_pg_color) %
794 (grpsize / cpu_topology_core_ids);
797 * Not enough space, split up by physical package,
798 * then split up by core id but only down to a
799 * 16-set. If all else fails, force a 16-set.
801 pg_color = phys_id * grpsize;
803 pg_color += 16 * (core_id % (grpsize / 16));
808 pg_color += (pindex + object_pg_color) %
813 * Unknown topology, distribute things evenly.
815 pg_color = gd->gd_cpuid * PQ_L2_SIZE / ncpus;
816 pg_color += pindex + object_pg_color;
822 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
823 * also wait for m->busy to become 0 before setting PG_BUSY.
826 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
827 int also_m_busy, const char *msg
835 if (flags & PG_BUSY) {
836 tsleep_interlock(m, 0);
837 if (atomic_cmpset_int(&m->flags, flags,
838 flags | PG_WANTED | PG_REFERENCED)) {
839 tsleep(m, PINTERLOCKED, msg, 0);
841 } else if (also_m_busy && (flags & PG_SBUSY)) {
842 tsleep_interlock(m, 0);
843 if (atomic_cmpset_int(&m->flags, flags,
844 flags | PG_WANTED | PG_REFERENCED)) {
845 tsleep(m, PINTERLOCKED, msg, 0);
848 if (atomic_cmpset_int(&m->flags, flags,
852 m->busy_line = lineno;
861 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
864 * Returns non-zero on failure.
867 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
877 if (also_m_busy && (flags & PG_SBUSY))
879 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
882 m->busy_line = lineno;
890 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
891 * that a wakeup() should be performed.
893 * The vm_page must be spinlocked and will remain spinlocked on return.
894 * The related queue must NOT be spinlocked (which could deadlock us).
900 _vm_page_wakeup(vm_page_t m)
907 if (atomic_cmpset_int(&m->flags, flags,
908 flags & ~(PG_BUSY | PG_WANTED))) {
912 return(flags & PG_WANTED);
916 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
917 * is typically the last call you make on a page before moving onto
921 vm_page_wakeup(vm_page_t m)
923 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
924 vm_page_spin_lock(m);
925 if (_vm_page_wakeup(m)) {
926 vm_page_spin_unlock(m);
929 vm_page_spin_unlock(m);
934 * Holding a page keeps it from being reused. Other parts of the system
935 * can still disassociate the page from its current object and free it, or
936 * perform read or write I/O on it and/or otherwise manipulate the page,
937 * but if the page is held the VM system will leave the page and its data
938 * intact and not reuse the page for other purposes until the last hold
939 * reference is released. (see vm_page_wire() if you want to prevent the
940 * page from being disassociated from its object too).
942 * The caller must still validate the contents of the page and, if necessary,
943 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
944 * before manipulating the page.
946 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
949 vm_page_hold(vm_page_t m)
951 vm_page_spin_lock(m);
952 atomic_add_int(&m->hold_count, 1);
953 if (m->queue - m->pc == PQ_FREE) {
954 _vm_page_queue_spin_lock(m);
955 _vm_page_rem_queue_spinlocked(m);
956 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
957 _vm_page_queue_spin_unlock(m);
959 vm_page_spin_unlock(m);
963 * The opposite of vm_page_hold(). If the page is on the HOLD queue
964 * it was freed while held and must be moved back to the FREE queue.
967 vm_page_unhold(vm_page_t m)
969 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE,
970 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)",
971 m, m->hold_count, m->queue - m->pc));
972 vm_page_spin_lock(m);
973 atomic_add_int(&m->hold_count, -1);
974 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
975 _vm_page_queue_spin_lock(m);
976 _vm_page_rem_queue_spinlocked(m);
977 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
978 _vm_page_queue_spin_unlock(m);
980 vm_page_spin_unlock(m);
986 * Create a fictitious page with the specified physical address and
987 * memory attribute. The memory attribute is the only the machine-
988 * dependent aspect of a fictitious page that must be initialized.
992 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
995 if ((m->flags & PG_FICTITIOUS) != 0) {
997 * The page's memattr might have changed since the
998 * previous initialization. Update the pmap to the
1003 m->phys_addr = paddr;
1005 /* Fictitious pages don't use "segind". */
1006 /* Fictitious pages don't use "order" or "pool". */
1007 m->flags = PG_FICTITIOUS | PG_UNMANAGED | PG_BUSY;
1011 pmap_page_set_memattr(m, memattr);
1015 * Inserts the given vm_page into the object and object list.
1017 * The pagetables are not updated but will presumably fault the page
1018 * in if necessary, or if a kernel page the caller will at some point
1019 * enter the page into the kernel's pmap. We are not allowed to block
1020 * here so we *can't* do this anyway.
1022 * This routine may not block.
1023 * This routine must be called with the vm_object held.
1024 * This routine must be called with a critical section held.
1026 * This routine returns TRUE if the page was inserted into the object
1027 * successfully, and FALSE if the page already exists in the object.
1030 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1032 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object));
1033 if (m->object != NULL)
1034 panic("vm_page_insert: already inserted");
1036 object->generation++;
1039 * Record the object/offset pair in this page and add the
1040 * pv_list_count of the page to the object.
1042 * The vm_page spin lock is required for interactions with the pmap.
1044 vm_page_spin_lock(m);
1047 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
1050 vm_page_spin_unlock(m);
1053 ++object->resident_page_count;
1054 ++mycpu->gd_vmtotal.t_rm;
1055 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
1056 vm_page_spin_unlock(m);
1059 * Since we are inserting a new and possibly dirty page,
1060 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
1062 if ((m->valid & m->dirty) ||
1063 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT)))
1064 vm_object_set_writeable_dirty(object);
1067 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
1069 swap_pager_page_inserted(m);
1074 * Removes the given vm_page_t from the (object,index) table
1076 * The underlying pmap entry (if any) is NOT removed here.
1077 * This routine may not block.
1079 * The page must be BUSY and will remain BUSY on return.
1080 * No other requirements.
1082 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
1086 vm_page_remove(vm_page_t m)
1090 if (m->object == NULL) {
1094 if ((m->flags & PG_BUSY) == 0)
1095 panic("vm_page_remove: page not busy");
1099 vm_object_hold(object);
1102 * Remove the page from the object and update the object.
1104 * The vm_page spin lock is required for interactions with the pmap.
1106 vm_page_spin_lock(m);
1107 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
1108 --object->resident_page_count;
1109 --mycpu->gd_vmtotal.t_rm;
1110 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
1112 vm_page_spin_unlock(m);
1114 object->generation++;
1116 vm_object_drop(object);
1120 * Locate and return the page at (object, pindex), or NULL if the
1121 * page could not be found.
1123 * The caller must hold the vm_object token.
1126 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1131 * Search the hash table for this object/offset pair
1133 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1134 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1135 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1140 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1142 int also_m_busy, const char *msg
1148 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1149 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1151 KKASSERT(m->object == object && m->pindex == pindex);
1154 if (flags & PG_BUSY) {
1155 tsleep_interlock(m, 0);
1156 if (atomic_cmpset_int(&m->flags, flags,
1157 flags | PG_WANTED | PG_REFERENCED)) {
1158 tsleep(m, PINTERLOCKED, msg, 0);
1159 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1162 } else if (also_m_busy && (flags & PG_SBUSY)) {
1163 tsleep_interlock(m, 0);
1164 if (atomic_cmpset_int(&m->flags, flags,
1165 flags | PG_WANTED | PG_REFERENCED)) {
1166 tsleep(m, PINTERLOCKED, msg, 0);
1167 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1170 } else if (atomic_cmpset_int(&m->flags, flags,
1172 #ifdef VM_PAGE_DEBUG
1173 m->busy_func = func;
1174 m->busy_line = lineno;
1183 * Attempt to lookup and busy a page.
1185 * Returns NULL if the page could not be found
1187 * Returns a vm_page and error == TRUE if the page exists but could not
1190 * Returns a vm_page and error == FALSE on success.
1193 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1195 int also_m_busy, int *errorp
1201 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1202 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1205 KKASSERT(m->object == object && m->pindex == pindex);
1208 if (flags & PG_BUSY) {
1212 if (also_m_busy && (flags & PG_SBUSY)) {
1216 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1217 #ifdef VM_PAGE_DEBUG
1218 m->busy_func = func;
1219 m->busy_line = lineno;
1228 * Attempt to repurpose the passed-in page. If the passed-in page cannot
1229 * be repurposed it will be released, *must_reenter will be set to 1, and
1230 * this function will fall-through to vm_page_lookup_busy_try().
1232 * The passed-in page must be wired and not busy. The returned page will
1233 * be busied and not wired.
1235 * A different page may be returned. The returned page will be busied and
1238 * NULL can be returned. If so, the required page could not be busied.
1239 * The passed-in page will be unwired.
1242 vm_page_repurpose(struct vm_object *object, vm_pindex_t pindex,
1243 int also_m_busy, int *errorp, vm_page_t m,
1244 int *must_reenter, int *iswired)
1248 * Do not mess with pages in a complex state, such as pages
1249 * which are mapped, as repurposing such pages can be more
1250 * expensive than simply allocatin a new one.
1252 * NOTE: Soft-busying can deadlock against putpages or I/O
1253 * so we only allow hard-busying here.
1255 KKASSERT(also_m_busy == FALSE);
1256 vm_page_busy_wait(m, also_m_busy, "biodep");
1258 if ((m->flags & (PG_UNMANAGED | PG_MAPPED |
1259 PG_FICTITIOUS | PG_SBUSY)) ||
1260 m->busy || m->wire_count != 1 || m->hold_count) {
1261 vm_page_unwire(m, 0);
1263 /* fall through to normal lookup */
1264 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
1265 vm_page_unwire(m, 0);
1266 vm_page_deactivate(m);
1268 /* fall through to normal lookup */
1271 * We can safely repurpose the page. It should
1272 * already be unqueued.
1274 KKASSERT(m->queue == PQ_NONE && m->dirty == 0);
1278 if (vm_page_insert(m, object, pindex)) {
1284 vm_page_unwire(m, 0);
1286 /* fall through to normal lookup */
1291 * Cannot repurpose page, attempt to locate the desired page. May
1296 m = vm_page_lookup_busy_try(object, pindex, also_m_busy, errorp);
1302 * Caller must hold the related vm_object
1305 vm_page_next(vm_page_t m)
1309 next = vm_page_rb_tree_RB_NEXT(m);
1310 if (next && next->pindex != m->pindex + 1)
1318 * Move the given vm_page from its current object to the specified
1319 * target object/offset. The page must be busy and will remain so
1322 * new_object must be held.
1323 * This routine might block. XXX ?
1325 * NOTE: Swap associated with the page must be invalidated by the move. We
1326 * have to do this for several reasons: (1) we aren't freeing the
1327 * page, (2) we are dirtying the page, (3) the VM system is probably
1328 * moving the page from object A to B, and will then later move
1329 * the backing store from A to B and we can't have a conflict.
1331 * NOTE: We *always* dirty the page. It is necessary both for the
1332 * fact that we moved it, and because we may be invalidating
1333 * swap. If the page is on the cache, we have to deactivate it
1334 * or vm_page_dirty() will panic. Dirty pages are not allowed
1338 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1340 KKASSERT(m->flags & PG_BUSY);
1341 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object));
1343 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object));
1346 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1347 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1348 new_object, new_pindex);
1350 if (m->queue - m->pc == PQ_CACHE)
1351 vm_page_deactivate(m);
1356 * vm_page_unqueue() without any wakeup. This routine is used when a page
1357 * is to remain BUSYied by the caller.
1359 * This routine may not block.
1362 vm_page_unqueue_nowakeup(vm_page_t m)
1364 vm_page_and_queue_spin_lock(m);
1365 (void)_vm_page_rem_queue_spinlocked(m);
1366 vm_page_spin_unlock(m);
1370 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1373 * This routine may not block.
1376 vm_page_unqueue(vm_page_t m)
1380 vm_page_and_queue_spin_lock(m);
1381 queue = _vm_page_rem_queue_spinlocked(m);
1382 if (queue == PQ_FREE || queue == PQ_CACHE) {
1383 vm_page_spin_unlock(m);
1384 pagedaemon_wakeup();
1386 vm_page_spin_unlock(m);
1391 * vm_page_list_find()
1393 * Find a page on the specified queue with color optimization.
1395 * The page coloring optimization attempts to locate a page that does
1396 * not overload other nearby pages in the object in the cpu's L1 or L2
1397 * caches. We need this optimization because cpu caches tend to be
1398 * physical caches, while object spaces tend to be virtual.
1400 * The page coloring optimization also, very importantly, tries to localize
1401 * memory to cpus and physical sockets.
1403 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1404 * and the algorithm is adjusted to localize allocations on a per-core basis.
1405 * This is done by 'twisting' the colors.
1407 * The page is returned spinlocked and removed from its queue (it will
1408 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1409 * is responsible for dealing with the busy-page case (usually by
1410 * deactivating the page and looping).
1412 * NOTE: This routine is carefully inlined. A non-inlined version
1413 * is available for outside callers but the only critical path is
1414 * from within this source file.
1416 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1417 * represent stable storage, allowing us to order our locks vm_page
1418 * first, then queue.
1422 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1428 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl,
1431 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1434 m = _vm_page_list_find2(basequeue, index);
1437 vm_page_and_queue_spin_lock(m);
1438 if (m->queue == basequeue + index) {
1439 _vm_page_rem_queue_spinlocked(m);
1440 /* vm_page_t spin held, no queue spin */
1443 vm_page_and_queue_spin_unlock(m);
1449 * If we could not find the page in the desired queue try to find it in
1453 _vm_page_list_find2(int basequeue, int index)
1455 struct vpgqueues *pq;
1457 int pqmask = PQ_SET_ASSOC_MASK >> 1;
1461 index &= PQ_L2_MASK;
1462 pq = &vm_page_queues[basequeue];
1465 * Run local sets of 16, 32, 64, 128, and the whole queue if all
1466 * else fails (PQ_L2_MASK which is 255).
1469 pqmask = (pqmask << 1) | 1;
1470 for (i = 0; i <= pqmask; ++i) {
1471 pqi = (index & ~pqmask) | ((index + i) & pqmask);
1472 m = TAILQ_FIRST(&pq[pqi].pl);
1474 _vm_page_and_queue_spin_lock(m);
1475 if (m->queue == basequeue + pqi) {
1476 _vm_page_rem_queue_spinlocked(m);
1479 _vm_page_and_queue_spin_unlock(m);
1484 } while (pqmask != PQ_L2_MASK);
1490 * Returns a vm_page candidate for allocation. The page is not busied so
1491 * it can move around. The caller must busy the page (and typically
1492 * deactivate it if it cannot be busied!)
1494 * Returns a spinlocked vm_page that has been removed from its queue.
1497 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1499 return(_vm_page_list_find(basequeue, index, prefer_zero));
1503 * Find a page on the cache queue with color optimization, remove it
1504 * from the queue, and busy it. The returned page will not be spinlocked.
1506 * A candidate failure will be deactivated. Candidates can fail due to
1507 * being busied by someone else, in which case they will be deactivated.
1509 * This routine may not block.
1513 vm_page_select_cache(u_short pg_color)
1518 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE);
1522 * (m) has been removed from its queue and spinlocked
1524 if (vm_page_busy_try(m, TRUE)) {
1525 _vm_page_deactivate_locked(m, 0);
1526 vm_page_spin_unlock(m);
1529 * We successfully busied the page
1531 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 &&
1532 m->hold_count == 0 &&
1533 m->wire_count == 0 &&
1534 (m->dirty & m->valid) == 0) {
1535 vm_page_spin_unlock(m);
1536 pagedaemon_wakeup();
1541 * The page cannot be recycled, deactivate it.
1543 _vm_page_deactivate_locked(m, 0);
1544 if (_vm_page_wakeup(m)) {
1545 vm_page_spin_unlock(m);
1548 vm_page_spin_unlock(m);
1556 * Find a free or zero page, with specified preference. We attempt to
1557 * inline the nominal case and fall back to _vm_page_select_free()
1558 * otherwise. A busied page is removed from the queue and returned.
1560 * This routine may not block.
1562 static __inline vm_page_t
1563 vm_page_select_free(u_short pg_color, boolean_t prefer_zero)
1568 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK,
1572 if (vm_page_busy_try(m, TRUE)) {
1574 * Various mechanisms such as a pmap_collect can
1575 * result in a busy page on the free queue. We
1576 * have to move the page out of the way so we can
1577 * retry the allocation. If the other thread is not
1578 * allocating the page then m->valid will remain 0 and
1579 * the pageout daemon will free the page later on.
1581 * Since we could not busy the page, however, we
1582 * cannot make assumptions as to whether the page
1583 * will be allocated by the other thread or not,
1584 * so all we can do is deactivate it to move it out
1585 * of the way. In particular, if the other thread
1586 * wires the page it may wind up on the inactive
1587 * queue and the pageout daemon will have to deal
1588 * with that case too.
1590 _vm_page_deactivate_locked(m, 0);
1591 vm_page_spin_unlock(m);
1594 * Theoretically if we are able to busy the page
1595 * atomic with the queue removal (using the vm_page
1596 * lock) nobody else should be able to mess with the
1599 KKASSERT((m->flags & (PG_UNMANAGED |
1600 PG_NEED_COMMIT)) == 0);
1601 KASSERT(m->hold_count == 0, ("m->hold_count is not zero "
1602 "pg %p q=%d flags=%08x hold=%d wire=%d",
1603 m, m->queue, m->flags, m->hold_count, m->wire_count));
1604 KKASSERT(m->wire_count == 0);
1605 vm_page_spin_unlock(m);
1606 pagedaemon_wakeup();
1608 /* return busied and removed page */
1618 * Allocate and return a memory cell associated with this VM object/offset
1619 * pair. If object is NULL an unassociated page will be allocated.
1621 * The returned page will be busied and removed from its queues. This
1622 * routine can block and may return NULL if a race occurs and the page
1623 * is found to already exist at the specified (object, pindex).
1625 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1626 * VM_ALLOC_QUICK like normal but cannot use cache
1627 * VM_ALLOC_SYSTEM greater free drain
1628 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1629 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1630 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1631 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1632 * (see vm_page_grab())
1633 * VM_ALLOC_USE_GD ok to use per-gd cache
1635 * The object must be held if not NULL
1636 * This routine may not block
1638 * Additional special handling is required when called from an interrupt
1639 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1643 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1645 globaldata_t gd = mycpu;
1652 * Special per-cpu free VM page cache. The pages are pre-busied
1653 * and pre-zerod for us.
1655 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1657 if (gd->gd_vmpg_count) {
1658 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1670 * CPU localization algorithm. Break the page queues up by physical
1671 * id and core id (note that two cpu threads will have the same core
1672 * id, and core_id != gd_cpuid).
1674 * This is nowhere near perfect, for example the last pindex in a
1675 * subgroup will overflow into the next cpu or package. But this
1676 * should get us good page reuse locality in heavy mixed loads.
1678 pg_color = vm_get_pg_color(gd, object, pindex);
1681 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1682 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1685 * Certain system threads (pageout daemon, buf_daemon's) are
1686 * allowed to eat deeper into the free page list.
1688 if (curthread->td_flags & TDF_SYSTHREAD)
1689 page_req |= VM_ALLOC_SYSTEM;
1692 * Impose various limitations. Note that the v_free_reserved test
1693 * must match the opposite of vm_page_count_target() to avoid
1694 * livelocks, be careful.
1697 if (vmstats.v_free_count >= vmstats.v_free_reserved ||
1698 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
1699 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
1700 vmstats.v_free_count > vmstats.v_interrupt_free_min)
1703 * The free queue has sufficient free pages to take one out.
1705 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO))
1706 m = vm_page_select_free(pg_color, TRUE);
1708 m = vm_page_select_free(pg_color, FALSE);
1709 } else if (page_req & VM_ALLOC_NORMAL) {
1711 * Allocatable from the cache (non-interrupt only). On
1712 * success, we must free the page and try again, thus
1713 * ensuring that vmstats.v_*_free_min counters are replenished.
1716 if (curthread->td_preempted) {
1717 kprintf("vm_page_alloc(): warning, attempt to allocate"
1718 " cache page from preempting interrupt\n");
1721 m = vm_page_select_cache(pg_color);
1724 m = vm_page_select_cache(pg_color);
1727 * On success move the page into the free queue and loop.
1729 * Only do this if we can safely acquire the vm_object lock,
1730 * because this is effectively a random page and the caller
1731 * might be holding the lock shared, we don't want to
1735 KASSERT(m->dirty == 0,
1736 ("Found dirty cache page %p", m));
1737 if ((obj = m->object) != NULL) {
1738 if (vm_object_hold_try(obj)) {
1739 vm_page_protect(m, VM_PROT_NONE);
1741 /* m->object NULL here */
1742 vm_object_drop(obj);
1744 vm_page_deactivate(m);
1748 vm_page_protect(m, VM_PROT_NONE);
1755 * On failure return NULL
1757 #if defined(DIAGNOSTIC)
1758 if (vmstats.v_cache_count > 0)
1759 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
1761 atomic_add_int(&vm_pageout_deficit, 1);
1762 pagedaemon_wakeup();
1766 * No pages available, wakeup the pageout daemon and give up.
1768 atomic_add_int(&vm_pageout_deficit, 1);
1769 pagedaemon_wakeup();
1774 * v_free_count can race so loop if we don't find the expected
1781 * Good page found. The page has already been busied for us and
1782 * removed from its queues.
1784 KASSERT(m->dirty == 0,
1785 ("vm_page_alloc: free/cache page %p was dirty", m));
1786 KKASSERT(m->queue == PQ_NONE);
1792 * Initialize the structure, inheriting some flags but clearing
1793 * all the rest. The page has already been busied for us.
1795 vm_page_flag_clear(m, ~(PG_BUSY | PG_SBUSY));
1796 KKASSERT(m->wire_count == 0);
1797 KKASSERT(m->busy == 0);
1802 * Caller must be holding the object lock (asserted by
1803 * vm_page_insert()).
1805 * NOTE: Inserting a page here does not insert it into any pmaps
1806 * (which could cause us to block allocating memory).
1808 * NOTE: If no object an unassociated page is allocated, m->pindex
1809 * can be used by the caller for any purpose.
1812 if (vm_page_insert(m, object, pindex) == FALSE) {
1814 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1815 panic("PAGE RACE %p[%ld]/%p",
1816 object, (long)pindex, m);
1824 * Don't wakeup too often - wakeup the pageout daemon when
1825 * we would be nearly out of memory.
1827 pagedaemon_wakeup();
1830 * A PG_BUSY page is returned.
1836 * Returns number of pages available in our DMA memory reserve
1837 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf)
1840 vm_contig_avail_pages(void)
1845 spin_lock(&vm_contig_spin);
1846 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
1847 spin_unlock(&vm_contig_spin);
1853 * Attempt to allocate contiguous physical memory with the specified
1857 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
1858 unsigned long alignment, unsigned long boundary,
1859 unsigned long size, vm_memattr_t memattr)
1865 alignment >>= PAGE_SHIFT;
1868 boundary >>= PAGE_SHIFT;
1871 size = (size + PAGE_MASK) >> PAGE_SHIFT;
1873 spin_lock(&vm_contig_spin);
1874 blk = alist_alloc(&vm_contig_alist, 0, size);
1875 if (blk == ALIST_BLOCK_NONE) {
1876 spin_unlock(&vm_contig_spin);
1878 kprintf("vm_page_alloc_contig: %ldk nospace\n",
1879 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1883 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
1884 alist_free(&vm_contig_alist, blk, size);
1885 spin_unlock(&vm_contig_spin);
1887 kprintf("vm_page_alloc_contig: %ldk high "
1889 (size + PAGE_MASK) * (PAGE_SIZE / 1024),
1894 spin_unlock(&vm_contig_spin);
1895 if (vm_contig_verbose) {
1896 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
1897 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT,
1898 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1901 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
1902 if (memattr != VM_MEMATTR_DEFAULT)
1903 for (i = 0;i < size;i++)
1904 pmap_page_set_memattr(&m[i], memattr);
1909 * Free contiguously allocated pages. The pages will be wired but not busy.
1910 * When freeing to the alist we leave them wired and not busy.
1913 vm_page_free_contig(vm_page_t m, unsigned long size)
1915 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1916 vm_pindex_t start = pa >> PAGE_SHIFT;
1917 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
1919 if (vm_contig_verbose) {
1920 kprintf("vm_page_free_contig: %016jx/%ldk\n",
1921 (intmax_t)pa, size / 1024);
1923 if (pa < vm_low_phys_reserved) {
1924 KKASSERT(pa + size <= vm_low_phys_reserved);
1925 spin_lock(&vm_contig_spin);
1926 alist_free(&vm_contig_alist, start, pages);
1927 spin_unlock(&vm_contig_spin);
1930 vm_page_busy_wait(m, FALSE, "cpgfr");
1931 vm_page_unwire(m, 0);
1942 * Wait for sufficient free memory for nominal heavy memory use kernel
1945 * WARNING! Be sure never to call this in any vm_pageout code path, which
1946 * will trivially deadlock the system.
1949 vm_wait_nominal(void)
1951 while (vm_page_count_min(0))
1956 * Test if vm_wait_nominal() would block.
1959 vm_test_nominal(void)
1961 if (vm_page_count_min(0))
1967 * Block until free pages are available for allocation, called in various
1968 * places before memory allocations.
1970 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
1971 * more generous then that.
1977 * never wait forever
1981 lwkt_gettoken(&vm_token);
1983 if (curthread == pagethread) {
1985 * The pageout daemon itself needs pages, this is bad.
1987 if (vm_page_count_min(0)) {
1988 vm_pageout_pages_needed = 1;
1989 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
1993 * Wakeup the pageout daemon if necessary and wait.
1995 * Do not wait indefinitely for the target to be reached,
1996 * as load might prevent it from being reached any time soon.
1997 * But wait a little to try to slow down page allocations
1998 * and to give more important threads (the pagedaemon)
1999 * allocation priority.
2001 if (vm_page_count_target()) {
2002 if (vm_pages_needed == 0) {
2003 vm_pages_needed = 1;
2004 wakeup(&vm_pages_needed);
2006 ++vm_pages_waiting; /* SMP race ok */
2007 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
2010 lwkt_reltoken(&vm_token);
2014 * Block until free pages are available for allocation
2016 * Called only from vm_fault so that processes page faulting can be
2020 vm_wait_pfault(void)
2023 * Wakeup the pageout daemon if necessary and wait.
2025 * Do not wait indefinitely for the target to be reached,
2026 * as load might prevent it from being reached any time soon.
2027 * But wait a little to try to slow down page allocations
2028 * and to give more important threads (the pagedaemon)
2029 * allocation priority.
2031 if (vm_page_count_min(0)) {
2032 lwkt_gettoken(&vm_token);
2033 while (vm_page_count_severe()) {
2034 if (vm_page_count_target()) {
2037 if (vm_pages_needed == 0) {
2038 vm_pages_needed = 1;
2039 wakeup(&vm_pages_needed);
2041 ++vm_pages_waiting; /* SMP race ok */
2042 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
2045 * Do not stay stuck in the loop if the system is trying
2046 * to kill the process.
2049 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
2053 lwkt_reltoken(&vm_token);
2058 * Put the specified page on the active list (if appropriate). Ensure
2059 * that act_count is at least ACT_INIT but do not otherwise mess with it.
2061 * The caller should be holding the page busied ? XXX
2062 * This routine may not block.
2065 vm_page_activate(vm_page_t m)
2069 vm_page_spin_lock(m);
2070 if (m->queue - m->pc != PQ_ACTIVE) {
2071 _vm_page_queue_spin_lock(m);
2072 oqueue = _vm_page_rem_queue_spinlocked(m);
2073 /* page is left spinlocked, queue is unlocked */
2075 if (oqueue == PQ_CACHE)
2076 mycpu->gd_cnt.v_reactivated++;
2077 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2078 if (m->act_count < ACT_INIT)
2079 m->act_count = ACT_INIT;
2080 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
2082 _vm_page_and_queue_spin_unlock(m);
2083 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
2084 pagedaemon_wakeup();
2086 if (m->act_count < ACT_INIT)
2087 m->act_count = ACT_INIT;
2088 vm_page_spin_unlock(m);
2093 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2094 * routine is called when a page has been added to the cache or free
2097 * This routine may not block.
2099 static __inline void
2100 vm_page_free_wakeup(void)
2103 * If the pageout daemon itself needs pages, then tell it that
2104 * there are some free.
2106 if (vm_pageout_pages_needed &&
2107 vmstats.v_cache_count + vmstats.v_free_count >=
2108 vmstats.v_pageout_free_min
2110 vm_pageout_pages_needed = 0;
2111 wakeup(&vm_pageout_pages_needed);
2115 * Wakeup processes that are waiting on memory.
2117 * Generally speaking we want to wakeup stuck processes as soon as
2118 * possible. !vm_page_count_min(0) is the absolute minimum point
2119 * where we can do this. Wait a bit longer to reduce degenerate
2120 * re-blocking (vm_page_free_hysteresis). The target check is just
2121 * to make sure the min-check w/hysteresis does not exceed the
2124 if (vm_pages_waiting) {
2125 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2126 !vm_page_count_target()) {
2127 vm_pages_waiting = 0;
2128 wakeup(&vmstats.v_free_count);
2129 ++mycpu->gd_cnt.v_ppwakeups;
2132 if (!vm_page_count_target()) {
2134 * Plenty of pages are free, wakeup everyone.
2136 vm_pages_waiting = 0;
2137 wakeup(&vmstats.v_free_count);
2138 ++mycpu->gd_cnt.v_ppwakeups;
2139 } else if (!vm_page_count_min(0)) {
2141 * Some pages are free, wakeup someone.
2143 int wcount = vm_pages_waiting;
2146 vm_pages_waiting = wcount;
2147 wakeup_one(&vmstats.v_free_count);
2148 ++mycpu->gd_cnt.v_ppwakeups;
2155 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
2156 * it from its VM object.
2158 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
2159 * return (the page will have been freed).
2162 vm_page_free_toq(vm_page_t m)
2164 mycpu->gd_cnt.v_tfree++;
2165 KKASSERT((m->flags & PG_MAPPED) == 0);
2166 KKASSERT(m->flags & PG_BUSY);
2168 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
2169 kprintf("vm_page_free: pindex(%lu), busy(%d), "
2170 "PG_BUSY(%d), hold(%d)\n",
2171 (u_long)m->pindex, m->busy,
2172 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count);
2173 if ((m->queue - m->pc) == PQ_FREE)
2174 panic("vm_page_free: freeing free page");
2176 panic("vm_page_free: freeing busy page");
2180 * Remove from object, spinlock the page and its queues and
2181 * remove from any queue. No queue spinlock will be held
2182 * after this section (because the page was removed from any
2186 vm_page_and_queue_spin_lock(m);
2187 _vm_page_rem_queue_spinlocked(m);
2190 * No further management of fictitious pages occurs beyond object
2191 * and queue removal.
2193 if ((m->flags & PG_FICTITIOUS) != 0) {
2194 vm_page_spin_unlock(m);
2202 if (m->wire_count != 0) {
2203 if (m->wire_count > 1) {
2205 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2206 m->wire_count, (long)m->pindex);
2208 panic("vm_page_free: freeing wired page");
2212 * Clear the UNMANAGED flag when freeing an unmanaged page.
2213 * Clear the NEED_COMMIT flag
2215 if (m->flags & PG_UNMANAGED)
2216 vm_page_flag_clear(m, PG_UNMANAGED);
2217 if (m->flags & PG_NEED_COMMIT)
2218 vm_page_flag_clear(m, PG_NEED_COMMIT);
2220 if (m->hold_count != 0) {
2221 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2223 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
2227 * This sequence allows us to clear PG_BUSY while still holding
2228 * its spin lock, which reduces contention vs allocators. We
2229 * must not leave the queue locked or _vm_page_wakeup() may
2232 _vm_page_queue_spin_unlock(m);
2233 if (_vm_page_wakeup(m)) {
2234 vm_page_spin_unlock(m);
2237 vm_page_spin_unlock(m);
2239 vm_page_free_wakeup();
2243 * vm_page_unmanage()
2245 * Prevent PV management from being done on the page. The page is
2246 * removed from the paging queues as if it were wired, and as a
2247 * consequence of no longer being managed the pageout daemon will not
2248 * touch it (since there is no way to locate the pte mappings for the
2249 * page). madvise() calls that mess with the pmap will also no longer
2250 * operate on the page.
2252 * Beyond that the page is still reasonably 'normal'. Freeing the page
2253 * will clear the flag.
2255 * This routine is used by OBJT_PHYS objects - objects using unswappable
2256 * physical memory as backing store rather then swap-backed memory and
2257 * will eventually be extended to support 4MB unmanaged physical
2260 * Caller must be holding the page busy.
2263 vm_page_unmanage(vm_page_t m)
2265 KKASSERT(m->flags & PG_BUSY);
2266 if ((m->flags & PG_UNMANAGED) == 0) {
2267 if (m->wire_count == 0)
2270 vm_page_flag_set(m, PG_UNMANAGED);
2274 * Mark this page as wired down by yet another map, removing it from
2275 * paging queues as necessary.
2277 * Caller must be holding the page busy.
2280 vm_page_wire(vm_page_t m)
2283 * Only bump the wire statistics if the page is not already wired,
2284 * and only unqueue the page if it is on some queue (if it is unmanaged
2285 * it is already off the queues). Don't do anything with fictitious
2286 * pages because they are always wired.
2288 KKASSERT(m->flags & PG_BUSY);
2289 if ((m->flags & PG_FICTITIOUS) == 0) {
2290 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2291 if ((m->flags & PG_UNMANAGED) == 0)
2293 atomic_add_int(&vmstats.v_wire_count, 1);
2295 KASSERT(m->wire_count != 0,
2296 ("vm_page_wire: wire_count overflow m=%p", m));
2301 * Release one wiring of this page, potentially enabling it to be paged again.
2303 * Many pages placed on the inactive queue should actually go
2304 * into the cache, but it is difficult to figure out which. What
2305 * we do instead, if the inactive target is well met, is to put
2306 * clean pages at the head of the inactive queue instead of the tail.
2307 * This will cause them to be moved to the cache more quickly and
2308 * if not actively re-referenced, freed more quickly. If we just
2309 * stick these pages at the end of the inactive queue, heavy filesystem
2310 * meta-data accesses can cause an unnecessary paging load on memory bound
2311 * processes. This optimization causes one-time-use metadata to be
2312 * reused more quickly.
2314 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2315 * the inactive queue. This helps the pageout daemon determine memory
2316 * pressure and act on out-of-memory situations more quickly.
2318 * BUT, if we are in a low-memory situation we have no choice but to
2319 * put clean pages on the cache queue.
2321 * A number of routines use vm_page_unwire() to guarantee that the page
2322 * will go into either the inactive or active queues, and will NEVER
2323 * be placed in the cache - for example, just after dirtying a page.
2324 * dirty pages in the cache are not allowed.
2326 * This routine may not block.
2329 vm_page_unwire(vm_page_t m, int activate)
2331 KKASSERT(m->flags & PG_BUSY);
2332 if (m->flags & PG_FICTITIOUS) {
2334 } else if (m->wire_count <= 0) {
2335 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2337 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2338 atomic_add_int(&vmstats.v_wire_count, -1);
2339 if (m->flags & PG_UNMANAGED) {
2341 } else if (activate || (m->flags & PG_NEED_COMMIT)) {
2342 vm_page_spin_lock(m);
2343 _vm_page_add_queue_spinlocked(m,
2344 PQ_ACTIVE + m->pc, 0);
2345 _vm_page_and_queue_spin_unlock(m);
2347 vm_page_spin_lock(m);
2348 vm_page_flag_clear(m, PG_WINATCFLS);
2349 _vm_page_add_queue_spinlocked(m,
2350 PQ_INACTIVE + m->pc, 0);
2351 ++vm_swapcache_inactive_heuristic;
2352 _vm_page_and_queue_spin_unlock(m);
2359 * Move the specified page to the inactive queue. If the page has
2360 * any associated swap, the swap is deallocated.
2362 * Normally athead is 0 resulting in LRU operation. athead is set
2363 * to 1 if we want this page to be 'as if it were placed in the cache',
2364 * except without unmapping it from the process address space.
2366 * vm_page's spinlock must be held on entry and will remain held on return.
2367 * This routine may not block.
2370 _vm_page_deactivate_locked(vm_page_t m, int athead)
2375 * Ignore if already inactive.
2377 if (m->queue - m->pc == PQ_INACTIVE)
2379 _vm_page_queue_spin_lock(m);
2380 oqueue = _vm_page_rem_queue_spinlocked(m);
2382 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2383 if (oqueue == PQ_CACHE)
2384 mycpu->gd_cnt.v_reactivated++;
2385 vm_page_flag_clear(m, PG_WINATCFLS);
2386 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2388 ++vm_swapcache_inactive_heuristic;
2390 /* NOTE: PQ_NONE if condition not taken */
2391 _vm_page_queue_spin_unlock(m);
2392 /* leaves vm_page spinlocked */
2396 * Attempt to deactivate a page.
2401 vm_page_deactivate(vm_page_t m)
2403 vm_page_spin_lock(m);
2404 _vm_page_deactivate_locked(m, 0);
2405 vm_page_spin_unlock(m);
2409 vm_page_deactivate_locked(vm_page_t m)
2411 _vm_page_deactivate_locked(m, 0);
2415 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it.
2417 * This function returns non-zero if it successfully moved the page to
2420 * This function unconditionally unbusies the page on return.
2423 vm_page_try_to_cache(vm_page_t m)
2425 vm_page_spin_lock(m);
2426 if (m->dirty || m->hold_count || m->wire_count ||
2427 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2428 if (_vm_page_wakeup(m)) {
2429 vm_page_spin_unlock(m);
2432 vm_page_spin_unlock(m);
2436 vm_page_spin_unlock(m);
2439 * Page busied by us and no longer spinlocked. Dirty pages cannot
2440 * be moved to the cache.
2442 vm_page_test_dirty(m);
2443 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2452 * Attempt to free the page. If we cannot free it, we do nothing.
2453 * 1 is returned on success, 0 on failure.
2458 vm_page_try_to_free(vm_page_t m)
2460 vm_page_spin_lock(m);
2461 if (vm_page_busy_try(m, TRUE)) {
2462 vm_page_spin_unlock(m);
2467 * The page can be in any state, including already being on the free
2468 * queue. Check to see if it really can be freed.
2470 if (m->dirty || /* can't free if it is dirty */
2471 m->hold_count || /* or held (XXX may be wrong) */
2472 m->wire_count || /* or wired */
2473 (m->flags & (PG_UNMANAGED | /* or unmanaged */
2474 PG_NEED_COMMIT)) || /* or needs a commit */
2475 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */
2476 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */
2477 if (_vm_page_wakeup(m)) {
2478 vm_page_spin_unlock(m);
2481 vm_page_spin_unlock(m);
2485 vm_page_spin_unlock(m);
2488 * We can probably free the page.
2490 * Page busied by us and no longer spinlocked. Dirty pages will
2491 * not be freed by this function. We have to re-test the
2492 * dirty bit after cleaning out the pmaps.
2494 vm_page_test_dirty(m);
2495 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2499 vm_page_protect(m, VM_PROT_NONE);
2500 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2511 * Put the specified page onto the page cache queue (if appropriate).
2513 * The page must be busy, and this routine will release the busy and
2514 * possibly even free the page.
2517 vm_page_cache(vm_page_t m)
2520 * Not suitable for the cache
2522 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2523 m->busy || m->wire_count || m->hold_count) {
2529 * Already in the cache (and thus not mapped)
2531 if ((m->queue - m->pc) == PQ_CACHE) {
2532 KKASSERT((m->flags & PG_MAPPED) == 0);
2538 * Caller is required to test m->dirty, but note that the act of
2539 * removing the page from its maps can cause it to become dirty
2540 * on an SMP system due to another cpu running in usermode.
2543 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2548 * Remove all pmaps and indicate that the page is not
2549 * writeable or mapped. Our vm_page_protect() call may
2550 * have blocked (especially w/ VM_PROT_NONE), so recheck
2553 vm_page_protect(m, VM_PROT_NONE);
2554 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2555 m->busy || m->wire_count || m->hold_count) {
2557 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2558 vm_page_deactivate(m);
2561 _vm_page_and_queue_spin_lock(m);
2562 _vm_page_rem_queue_spinlocked(m);
2563 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2564 _vm_page_queue_spin_unlock(m);
2565 if (_vm_page_wakeup(m)) {
2566 vm_page_spin_unlock(m);
2569 vm_page_spin_unlock(m);
2571 vm_page_free_wakeup();
2576 * vm_page_dontneed()
2578 * Cache, deactivate, or do nothing as appropriate. This routine
2579 * is typically used by madvise() MADV_DONTNEED.
2581 * Generally speaking we want to move the page into the cache so
2582 * it gets reused quickly. However, this can result in a silly syndrome
2583 * due to the page recycling too quickly. Small objects will not be
2584 * fully cached. On the otherhand, if we move the page to the inactive
2585 * queue we wind up with a problem whereby very large objects
2586 * unnecessarily blow away our inactive and cache queues.
2588 * The solution is to move the pages based on a fixed weighting. We
2589 * either leave them alone, deactivate them, or move them to the cache,
2590 * where moving them to the cache has the highest weighting.
2591 * By forcing some pages into other queues we eventually force the
2592 * system to balance the queues, potentially recovering other unrelated
2593 * space from active. The idea is to not force this to happen too
2596 * The page must be busied.
2599 vm_page_dontneed(vm_page_t m)
2601 static int dnweight;
2608 * occassionally leave the page alone
2610 if ((dnw & 0x01F0) == 0 ||
2611 m->queue - m->pc == PQ_INACTIVE ||
2612 m->queue - m->pc == PQ_CACHE
2614 if (m->act_count >= ACT_INIT)
2620 * If vm_page_dontneed() is inactivating a page, it must clear
2621 * the referenced flag; otherwise the pagedaemon will see references
2622 * on the page in the inactive queue and reactivate it. Until the
2623 * page can move to the cache queue, madvise's job is not done.
2625 vm_page_flag_clear(m, PG_REFERENCED);
2626 pmap_clear_reference(m);
2629 vm_page_test_dirty(m);
2631 if (m->dirty || (dnw & 0x0070) == 0) {
2633 * Deactivate the page 3 times out of 32.
2638 * Cache the page 28 times out of every 32. Note that
2639 * the page is deactivated instead of cached, but placed
2640 * at the head of the queue instead of the tail.
2644 vm_page_spin_lock(m);
2645 _vm_page_deactivate_locked(m, head);
2646 vm_page_spin_unlock(m);
2650 * These routines manipulate the 'soft busy' count for a page. A soft busy
2651 * is almost like PG_BUSY except that it allows certain compatible operations
2652 * to occur on the page while it is busy. For example, a page undergoing a
2653 * write can still be mapped read-only.
2655 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2656 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2657 * busy bit is cleared.
2660 vm_page_io_start(vm_page_t m)
2662 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2663 atomic_add_char(&m->busy, 1);
2664 vm_page_flag_set(m, PG_SBUSY);
2668 vm_page_io_finish(vm_page_t m)
2670 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2671 atomic_subtract_char(&m->busy, 1);
2673 vm_page_flag_clear(m, PG_SBUSY);
2677 * Indicate that a clean VM page requires a filesystem commit and cannot
2678 * be reused. Used by tmpfs.
2681 vm_page_need_commit(vm_page_t m)
2683 vm_page_flag_set(m, PG_NEED_COMMIT);
2684 vm_object_set_writeable_dirty(m->object);
2688 vm_page_clear_commit(vm_page_t m)
2690 vm_page_flag_clear(m, PG_NEED_COMMIT);
2694 * Grab a page, blocking if it is busy and allocating a page if necessary.
2695 * A busy page is returned or NULL. The page may or may not be valid and
2696 * might not be on a queue (the caller is responsible for the disposition of
2699 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2700 * page will be zero'd and marked valid.
2702 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2703 * valid even if it already exists.
2705 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2706 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2707 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2709 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2710 * always returned if we had blocked.
2712 * This routine may not be called from an interrupt.
2714 * No other requirements.
2717 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2723 KKASSERT(allocflags &
2724 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2725 vm_object_hold_shared(object);
2727 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2729 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2730 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2735 } else if (m == NULL) {
2737 vm_object_upgrade(object);
2740 if (allocflags & VM_ALLOC_RETRY)
2741 allocflags |= VM_ALLOC_NULL_OK;
2742 m = vm_page_alloc(object, pindex,
2743 allocflags & ~VM_ALLOC_RETRY);
2747 if ((allocflags & VM_ALLOC_RETRY) == 0)
2756 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2758 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2759 * valid even if already valid.
2761 * NOTE! We have removed all of the PG_ZERO optimizations and also
2762 * removed the idle zeroing code. These optimizations actually
2763 * slow things down on modern cpus because the zerod area is
2764 * likely uncached, placing a memory-access burden on the
2765 * accesors taking the fault.
2767 * By always zeroing the page in-line with the fault, no
2768 * dynamic ram reads are needed and the caches are hot, ready
2769 * for userland to access the memory.
2771 if (m->valid == 0) {
2772 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2773 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2774 m->valid = VM_PAGE_BITS_ALL;
2776 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2777 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2778 m->valid = VM_PAGE_BITS_ALL;
2781 vm_object_drop(object);
2786 * Mapping function for valid bits or for dirty bits in
2787 * a page. May not block.
2789 * Inputs are required to range within a page.
2795 vm_page_bits(int base, int size)
2801 base + size <= PAGE_SIZE,
2802 ("vm_page_bits: illegal base/size %d/%d", base, size)
2805 if (size == 0) /* handle degenerate case */
2808 first_bit = base >> DEV_BSHIFT;
2809 last_bit = (base + size - 1) >> DEV_BSHIFT;
2811 return ((2 << last_bit) - (1 << first_bit));
2815 * Sets portions of a page valid and clean. The arguments are expected
2816 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2817 * of any partial chunks touched by the range. The invalid portion of
2818 * such chunks will be zero'd.
2820 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2821 * align base to DEV_BSIZE so as not to mark clean a partially
2822 * truncated device block. Otherwise the dirty page status might be
2825 * This routine may not block.
2827 * (base + size) must be less then or equal to PAGE_SIZE.
2830 _vm_page_zero_valid(vm_page_t m, int base, int size)
2835 if (size == 0) /* handle degenerate case */
2839 * If the base is not DEV_BSIZE aligned and the valid
2840 * bit is clear, we have to zero out a portion of the
2844 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2845 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2847 pmap_zero_page_area(
2855 * If the ending offset is not DEV_BSIZE aligned and the
2856 * valid bit is clear, we have to zero out a portion of
2860 endoff = base + size;
2862 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2863 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
2865 pmap_zero_page_area(
2868 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
2874 * Set valid, clear dirty bits. If validating the entire
2875 * page we can safely clear the pmap modify bit. We also
2876 * use this opportunity to clear the PG_NOSYNC flag. If a process
2877 * takes a write fault on a MAP_NOSYNC memory area the flag will
2880 * We set valid bits inclusive of any overlap, but we can only
2881 * clear dirty bits for DEV_BSIZE chunks that are fully within
2884 * Page must be busied?
2885 * No other requirements.
2888 vm_page_set_valid(vm_page_t m, int base, int size)
2890 _vm_page_zero_valid(m, base, size);
2891 m->valid |= vm_page_bits(base, size);
2896 * Set valid bits and clear dirty bits.
2898 * NOTE: This function does not clear the pmap modified bit.
2899 * Also note that e.g. NFS may use a byte-granular base
2902 * WARNING: Page must be busied? But vfs_clean_one_page() will call
2903 * this without necessarily busying the page (via bdwrite()).
2904 * So for now vm_token must also be held.
2906 * No other requirements.
2909 vm_page_set_validclean(vm_page_t m, int base, int size)
2913 _vm_page_zero_valid(m, base, size);
2914 pagebits = vm_page_bits(base, size);
2915 m->valid |= pagebits;
2916 m->dirty &= ~pagebits;
2917 if (base == 0 && size == PAGE_SIZE) {
2918 /*pmap_clear_modify(m);*/
2919 vm_page_flag_clear(m, PG_NOSYNC);
2924 * Set valid & dirty. Used by buwrite()
2926 * WARNING: Page must be busied? But vfs_dirty_one_page() will
2927 * call this function in buwrite() so for now vm_token must
2930 * No other requirements.
2933 vm_page_set_validdirty(vm_page_t m, int base, int size)
2937 pagebits = vm_page_bits(base, size);
2938 m->valid |= pagebits;
2939 m->dirty |= pagebits;
2941 vm_object_set_writeable_dirty(m->object);
2947 * NOTE: This function does not clear the pmap modified bit.
2948 * Also note that e.g. NFS may use a byte-granular base
2951 * Page must be busied?
2952 * No other requirements.
2955 vm_page_clear_dirty(vm_page_t m, int base, int size)
2957 m->dirty &= ~vm_page_bits(base, size);
2958 if (base == 0 && size == PAGE_SIZE) {
2959 /*pmap_clear_modify(m);*/
2960 vm_page_flag_clear(m, PG_NOSYNC);
2965 * Make the page all-dirty.
2967 * Also make sure the related object and vnode reflect the fact that the
2968 * object may now contain a dirty page.
2970 * Page must be busied?
2971 * No other requirements.
2974 vm_page_dirty(vm_page_t m)
2977 int pqtype = m->queue - m->pc;
2979 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
2980 ("vm_page_dirty: page in free/cache queue!"));
2981 if (m->dirty != VM_PAGE_BITS_ALL) {
2982 m->dirty = VM_PAGE_BITS_ALL;
2984 vm_object_set_writeable_dirty(m->object);
2989 * Invalidates DEV_BSIZE'd chunks within a page. Both the
2990 * valid and dirty bits for the effected areas are cleared.
2992 * Page must be busied?
2994 * No other requirements.
2997 vm_page_set_invalid(vm_page_t m, int base, int size)
3001 bits = vm_page_bits(base, size);
3004 m->object->generation++;
3008 * The kernel assumes that the invalid portions of a page contain
3009 * garbage, but such pages can be mapped into memory by user code.
3010 * When this occurs, we must zero out the non-valid portions of the
3011 * page so user code sees what it expects.
3013 * Pages are most often semi-valid when the end of a file is mapped
3014 * into memory and the file's size is not page aligned.
3016 * Page must be busied?
3017 * No other requirements.
3020 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3026 * Scan the valid bits looking for invalid sections that
3027 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
3028 * valid bit may be set ) have already been zerod by
3029 * vm_page_set_validclean().
3031 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3032 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3033 (m->valid & (1 << i))
3036 pmap_zero_page_area(
3039 (i - b) << DEV_BSHIFT
3047 * setvalid is TRUE when we can safely set the zero'd areas
3048 * as being valid. We can do this if there are no cache consistency
3049 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3052 m->valid = VM_PAGE_BITS_ALL;
3056 * Is a (partial) page valid? Note that the case where size == 0
3057 * will return FALSE in the degenerate case where the page is entirely
3058 * invalid, and TRUE otherwise.
3061 * No other requirements.
3064 vm_page_is_valid(vm_page_t m, int base, int size)
3066 int bits = vm_page_bits(base, size);
3068 if (m->valid && ((m->valid & bits) == bits))
3075 * update dirty bits from pmap/mmu. May not block.
3077 * Caller must hold the page busy
3080 vm_page_test_dirty(vm_page_t m)
3082 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
3088 * Register an action, associating it with its vm_page
3091 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
3093 struct vm_page_action_list *list;
3096 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
3097 list = &action_list[hv];
3099 lwkt_gettoken(&vm_token);
3100 vm_page_flag_set(action->m, PG_ACTIONLIST);
3101 action->event = event;
3102 LIST_INSERT_HEAD(list, action, entry);
3103 lwkt_reltoken(&vm_token);
3107 * Unregister an action, disassociating it from its related vm_page
3110 vm_page_unregister_action(vm_page_action_t action)
3112 struct vm_page_action_list *list;
3115 lwkt_gettoken(&vm_token);
3116 if (action->event != VMEVENT_NONE) {
3117 action->event = VMEVENT_NONE;
3118 LIST_REMOVE(action, entry);
3120 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
3121 list = &action_list[hv];
3122 if (LIST_EMPTY(list))
3123 vm_page_flag_clear(action->m, PG_ACTIONLIST);
3125 lwkt_reltoken(&vm_token);
3129 * Issue an event on a VM page. Corresponding action structures are
3130 * removed from the page's list and called.
3132 * If the vm_page has no more pending action events we clear its
3133 * PG_ACTIONLIST flag.
3136 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
3138 struct vm_page_action_list *list;
3139 struct vm_page_action *scan;
3140 struct vm_page_action *next;
3144 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
3145 list = &action_list[hv];
3148 lwkt_gettoken(&vm_token);
3149 LIST_FOREACH_MUTABLE(scan, list, entry, next) {
3151 if (scan->event == event) {
3152 scan->event = VMEVENT_NONE;
3153 LIST_REMOVE(scan, entry);
3154 scan->func(m, scan);
3162 vm_page_flag_clear(m, PG_ACTIONLIST);
3163 lwkt_reltoken(&vm_token);
3166 #include "opt_ddb.h"
3168 #include <sys/kernel.h>
3170 #include <ddb/ddb.h>
3172 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3174 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
3175 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
3176 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
3177 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
3178 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
3179 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
3180 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
3181 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
3182 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
3183 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
3186 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3189 db_printf("PQ_FREE:");
3190 for(i=0;i<PQ_L2_SIZE;i++) {
3191 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
3195 db_printf("PQ_CACHE:");
3196 for(i=0;i<PQ_L2_SIZE;i++) {
3197 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
3201 db_printf("PQ_ACTIVE:");
3202 for(i=0;i<PQ_L2_SIZE;i++) {
3203 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3207 db_printf("PQ_INACTIVE:");
3208 for(i=0;i<PQ_L2_SIZE;i++) {
3209 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);