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];
255 /* too expensive time-wise in large-mem configurations */
256 if ((vpq->flipflop & 15) == 0) {
257 pmap_zero_page(VM_PAGE_TO_PHYS(m));
259 TAILQ_INSERT_TAIL(&vpq->pl, m, pageq);
263 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
274 * Initializes the resident memory module.
276 * Preallocates memory for critical VM structures and arrays prior to
277 * kernel_map becoming available.
279 * Memory is allocated from (virtual2_start, virtual2_end) if available,
280 * otherwise memory is allocated from (virtual_start, virtual_end).
282 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
283 * large enough to hold vm_page_array & other structures for machines with
284 * large amounts of ram, so we want to use virtual2* when available.
287 vm_page_startup(void)
289 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start;
292 vm_paddr_t page_range;
299 vm_paddr_t biggestone, biggestsize;
306 vaddr = round_page(vaddr);
308 for (i = 0; phys_avail[i + 1]; i += 2) {
309 phys_avail[i] = round_page64(phys_avail[i]);
310 phys_avail[i + 1] = trunc_page64(phys_avail[i + 1]);
313 for (i = 0; phys_avail[i + 1]; i += 2) {
314 vm_paddr_t size = phys_avail[i + 1] - phys_avail[i];
316 if (size > biggestsize) {
324 end = phys_avail[biggestone+1];
325 end = trunc_page(end);
328 * Initialize the queue headers for the free queue, the active queue
329 * and the inactive queue.
331 vm_page_queue_init();
333 #if !defined(_KERNEL_VIRTUAL)
335 * VKERNELs don't support minidumps and as such don't need
338 * Allocate a bitmap to indicate that a random physical page
339 * needs to be included in a minidump.
341 * The amd64 port needs this to indicate which direct map pages
342 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
344 * However, i386 still needs this workspace internally within the
345 * minidump code. In theory, they are not needed on i386, but are
346 * included should the sf_buf code decide to use them.
348 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE;
349 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
350 end -= vm_page_dump_size;
351 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
352 VM_PROT_READ | VM_PROT_WRITE);
353 bzero((void *)vm_page_dump, vm_page_dump_size);
356 * Compute the number of pages of memory that will be available for
357 * use (taking into account the overhead of a page structure per
360 first_page = phys_avail[0] / PAGE_SIZE;
361 page_range = phys_avail[(nblocks - 1) * 2 + 1] / PAGE_SIZE - first_page;
362 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
364 #ifndef _KERNEL_VIRTUAL
366 * (only applies to real kernels)
368 * Reserve a large amount of low memory for potential 32-bit DMA
369 * space allocations. Once device initialization is complete we
370 * release most of it, but keep (vm_dma_reserved) memory reserved
371 * for later use. Typically for X / graphics. Through trial and
372 * error we find that GPUs usually requires ~60-100MB or so.
374 * By default, 128M is left in reserve on machines with 2G+ of ram.
376 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT;
377 if (vm_low_phys_reserved > total / 4)
378 vm_low_phys_reserved = total / 4;
379 if (vm_dma_reserved == 0) {
380 vm_dma_reserved = 128 * 1024 * 1024; /* 128MB */
381 if (vm_dma_reserved > total / 16)
382 vm_dma_reserved = total / 16;
385 alist_init(&vm_contig_alist, 65536, vm_contig_ameta,
386 ALIST_RECORDS_65536);
389 * Initialize the mem entry structures now, and put them in the free
392 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
393 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE);
394 vm_page_array = (vm_page_t)mapped;
396 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
398 * since pmap_map on amd64 returns stuff out of a direct-map region,
399 * we have to manually add these pages to the minidump tracking so
400 * that they can be dumped, including the vm_page_array.
402 for (pa = new_end; pa < phys_avail[biggestone + 1]; pa += PAGE_SIZE)
407 * Clear all of the page structures
409 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
410 vm_page_array_size = page_range;
413 * Construct the free queue(s) in ascending order (by physical
414 * address) so that the first 16MB of physical memory is allocated
415 * last rather than first. On large-memory machines, this avoids
416 * the exhaustion of low physical memory before isa_dmainit has run.
418 vmstats.v_page_count = 0;
419 vmstats.v_free_count = 0;
420 for (i = 0; phys_avail[i + 1] && npages > 0; i += 2) {
425 last_pa = phys_avail[i + 1];
426 while (pa < last_pa && npages-- > 0) {
432 virtual2_start = vaddr;
434 virtual_start = vaddr;
438 * We tended to reserve a ton of memory for contigmalloc(). Now that most
439 * drivers have initialized we want to return most the remaining free
440 * reserve back to the VM page queues so they can be used for normal
443 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool.
446 vm_page_startup_finish(void *dummy __unused)
455 spin_lock(&vm_contig_spin);
457 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
458 if (bfree <= vm_dma_reserved / PAGE_SIZE)
464 * Figure out how much of the initial reserve we have to
465 * free in order to reach our target.
467 bfree -= vm_dma_reserved / PAGE_SIZE;
469 blk += count - bfree;
474 * Calculate the nearest power of 2 <= count.
476 for (xcount = 1; xcount <= count; xcount <<= 1)
479 blk += count - xcount;
483 * Allocate the pages from the alist, then free them to
484 * the normal VM page queues.
486 * Pages allocated from the alist are wired. We have to
487 * busy, unwire, and free them. We must also adjust
488 * vm_low_phys_reserved before freeing any pages to prevent
491 rblk = alist_alloc(&vm_contig_alist, blk, count);
493 kprintf("vm_page_startup_finish: Unable to return "
494 "dma space @0x%08x/%d -> 0x%08x\n",
498 atomic_add_int(&vmstats.v_dma_pages, -count);
499 spin_unlock(&vm_contig_spin);
501 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
502 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m);
504 vm_page_busy_wait(m, FALSE, "cpgfr");
505 vm_page_unwire(m, 0);
510 spin_lock(&vm_contig_spin);
512 spin_unlock(&vm_contig_spin);
515 * Print out how much DMA space drivers have already allocated and
516 * how much is left over.
518 kprintf("DMA space used: %jdk, remaining available: %jdk\n",
519 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) *
521 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024));
523 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY,
524 vm_page_startup_finish, NULL);
528 * Scan comparison function for Red-Black tree scans. An inclusive
529 * (start,end) is expected. Other fields are not used.
532 rb_vm_page_scancmp(struct vm_page *p, void *data)
534 struct rb_vm_page_scan_info *info = data;
536 if (p->pindex < info->start_pindex)
538 if (p->pindex > info->end_pindex)
544 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
546 if (p1->pindex < p2->pindex)
548 if (p1->pindex > p2->pindex)
554 vm_page_init(vm_page_t m)
556 /* do nothing for now. Called from pmap_page_init() */
560 * Each page queue has its own spin lock, which is fairly optimal for
561 * allocating and freeing pages at least.
563 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
564 * queue spinlock via this function. Also note that m->queue cannot change
565 * unless both the page and queue are locked.
569 _vm_page_queue_spin_lock(vm_page_t m)
574 if (queue != PQ_NONE) {
575 spin_lock(&vm_page_queues[queue].spin);
576 KKASSERT(queue == m->queue);
582 _vm_page_queue_spin_unlock(vm_page_t m)
588 if (queue != PQ_NONE)
589 spin_unlock(&vm_page_queues[queue].spin);
594 _vm_page_queues_spin_lock(u_short queue)
597 if (queue != PQ_NONE)
598 spin_lock(&vm_page_queues[queue].spin);
604 _vm_page_queues_spin_unlock(u_short queue)
607 if (queue != PQ_NONE)
608 spin_unlock(&vm_page_queues[queue].spin);
612 vm_page_queue_spin_lock(vm_page_t m)
614 _vm_page_queue_spin_lock(m);
618 vm_page_queues_spin_lock(u_short queue)
620 _vm_page_queues_spin_lock(queue);
624 vm_page_queue_spin_unlock(vm_page_t m)
626 _vm_page_queue_spin_unlock(m);
630 vm_page_queues_spin_unlock(u_short queue)
632 _vm_page_queues_spin_unlock(queue);
636 * This locks the specified vm_page and its queue in the proper order
637 * (page first, then queue). The queue may change so the caller must
642 _vm_page_and_queue_spin_lock(vm_page_t m)
644 vm_page_spin_lock(m);
645 _vm_page_queue_spin_lock(m);
650 _vm_page_and_queue_spin_unlock(vm_page_t m)
652 _vm_page_queues_spin_unlock(m->queue);
653 vm_page_spin_unlock(m);
657 vm_page_and_queue_spin_unlock(vm_page_t m)
659 _vm_page_and_queue_spin_unlock(m);
663 vm_page_and_queue_spin_lock(vm_page_t m)
665 _vm_page_and_queue_spin_lock(m);
669 * Helper function removes vm_page from its current queue.
670 * Returns the base queue the page used to be on.
672 * The vm_page and the queue must be spinlocked.
673 * This function will unlock the queue but leave the page spinlocked.
675 static __inline u_short
676 _vm_page_rem_queue_spinlocked(vm_page_t m)
678 struct vpgqueues *pq;
683 if (queue != PQ_NONE) {
684 pq = &vm_page_queues[queue];
685 TAILQ_REMOVE(&pq->pl, m, pageq);
686 atomic_add_int(pq->cnt, -1);
690 if ((queue - m->pc) == PQ_FREE && (m->flags & PG_ZERO))
692 if ((queue - m->pc) == PQ_CACHE || (queue - m->pc) == PQ_FREE)
694 vm_page_queues_spin_unlock(oqueue); /* intended */
700 * Helper function places the vm_page on the specified queue.
702 * The vm_page must be spinlocked.
703 * This function will return with both the page and the queue locked.
706 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
708 struct vpgqueues *pq;
710 KKASSERT(m->queue == PQ_NONE);
712 if (queue != PQ_NONE) {
713 vm_page_queues_spin_lock(queue);
714 pq = &vm_page_queues[queue];
716 atomic_add_int(pq->cnt, 1);
720 * Put zero'd pages on the end ( where we look for zero'd pages
721 * first ) and non-zerod pages at the head.
723 if (queue - m->pc == PQ_FREE) {
724 if (m->flags & PG_ZERO) {
725 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
728 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
731 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
733 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
735 /* leave the queue spinlocked */
740 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
741 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
742 * did not. Only one sleep call will be made before returning.
744 * This function does NOT busy the page and on return the page is not
745 * guaranteed to be available.
748 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
756 if ((flags & PG_BUSY) == 0 &&
757 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
760 tsleep_interlock(m, 0);
761 if (atomic_cmpset_int(&m->flags, flags,
762 flags | PG_WANTED | PG_REFERENCED)) {
763 tsleep(m, PINTERLOCKED, msg, 0);
770 * This calculates and returns a page color given an optional VM object and
771 * either a pindex or an iterator. We attempt to return a cpu-localized
772 * pg_color that is still roughly 16-way set-associative. The CPU topology
773 * is used if it was probed.
775 * The caller may use the returned value to index into e.g. PQ_FREE when
776 * allocating a page in order to nominally obtain pages that are hopefully
777 * already localized to the requesting cpu. This function is not able to
778 * provide any sort of guarantee of this, but does its best to improve
779 * hardware cache management performance.
781 * WARNING! The caller must mask the returned value with PQ_L2_MASK.
784 vm_get_pg_color(globaldata_t gd, vm_object_t object, vm_pindex_t pindex)
791 phys_id = get_cpu_phys_id(gd->gd_cpuid);
792 core_id = get_cpu_core_id(gd->gd_cpuid);
793 object_pg_color = object ? object->pg_color : 0;
795 if (cpu_topology_phys_ids && cpu_topology_core_ids) {
796 int grpsize = PQ_L2_SIZE / cpu_topology_phys_ids;
798 if (grpsize / cpu_topology_core_ids >= PQ_SET_ASSOC) {
800 * Enough space for a full break-down.
802 pg_color = phys_id * grpsize;
803 pg_color += core_id * grpsize / cpu_topology_core_ids;
804 pg_color += (pindex + object_pg_color) %
805 (grpsize / cpu_topology_core_ids);
808 * Not enough space, split up by physical package,
809 * then split up by core id but only down to a
810 * 16-set. If all else fails, force a 16-set.
812 pg_color = phys_id * grpsize;
814 pg_color += 16 * (core_id % (grpsize / 16));
819 pg_color += (pindex + object_pg_color) %
824 * Unknown topology, distribute things evenly.
826 pg_color = gd->gd_cpuid * PQ_L2_SIZE / ncpus;
827 pg_color += pindex + object_pg_color;
833 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
834 * also wait for m->busy to become 0 before setting PG_BUSY.
837 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
838 int also_m_busy, const char *msg
846 if (flags & PG_BUSY) {
847 tsleep_interlock(m, 0);
848 if (atomic_cmpset_int(&m->flags, flags,
849 flags | PG_WANTED | PG_REFERENCED)) {
850 tsleep(m, PINTERLOCKED, msg, 0);
852 } else if (also_m_busy && (flags & PG_SBUSY)) {
853 tsleep_interlock(m, 0);
854 if (atomic_cmpset_int(&m->flags, flags,
855 flags | PG_WANTED | PG_REFERENCED)) {
856 tsleep(m, PINTERLOCKED, msg, 0);
859 if (atomic_cmpset_int(&m->flags, flags,
863 m->busy_line = lineno;
872 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
875 * Returns non-zero on failure.
878 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
888 if (also_m_busy && (flags & PG_SBUSY))
890 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
893 m->busy_line = lineno;
901 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
902 * that a wakeup() should be performed.
904 * The vm_page must be spinlocked and will remain spinlocked on return.
905 * The related queue must NOT be spinlocked (which could deadlock us).
911 _vm_page_wakeup(vm_page_t m)
918 if (atomic_cmpset_int(&m->flags, flags,
919 flags & ~(PG_BUSY | PG_WANTED))) {
923 return(flags & PG_WANTED);
927 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
928 * is typically the last call you make on a page before moving onto
932 vm_page_wakeup(vm_page_t m)
934 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
935 vm_page_spin_lock(m);
936 if (_vm_page_wakeup(m)) {
937 vm_page_spin_unlock(m);
940 vm_page_spin_unlock(m);
945 * Holding a page keeps it from being reused. Other parts of the system
946 * can still disassociate the page from its current object and free it, or
947 * perform read or write I/O on it and/or otherwise manipulate the page,
948 * but if the page is held the VM system will leave the page and its data
949 * intact and not reuse the page for other purposes until the last hold
950 * reference is released. (see vm_page_wire() if you want to prevent the
951 * page from being disassociated from its object too).
953 * The caller must still validate the contents of the page and, if necessary,
954 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
955 * before manipulating the page.
957 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
960 vm_page_hold(vm_page_t m)
962 vm_page_spin_lock(m);
963 atomic_add_int(&m->hold_count, 1);
964 if (m->queue - m->pc == PQ_FREE) {
965 _vm_page_queue_spin_lock(m);
966 _vm_page_rem_queue_spinlocked(m);
967 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
968 _vm_page_queue_spin_unlock(m);
970 vm_page_spin_unlock(m);
974 * The opposite of vm_page_hold(). If the page is on the HOLD queue
975 * it was freed while held and must be moved back to the FREE queue.
978 vm_page_unhold(vm_page_t m)
980 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE,
981 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)",
982 m, m->hold_count, m->queue - m->pc));
983 vm_page_spin_lock(m);
984 atomic_add_int(&m->hold_count, -1);
985 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
986 _vm_page_queue_spin_lock(m);
987 _vm_page_rem_queue_spinlocked(m);
988 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
989 _vm_page_queue_spin_unlock(m);
991 vm_page_spin_unlock(m);
997 * Create a fictitious page with the specified physical address and
998 * memory attribute. The memory attribute is the only the machine-
999 * dependent aspect of a fictitious page that must be initialized.
1003 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1006 if ((m->flags & PG_FICTITIOUS) != 0) {
1008 * The page's memattr might have changed since the
1009 * previous initialization. Update the pmap to the
1014 m->phys_addr = paddr;
1016 /* Fictitious pages don't use "segind". */
1017 /* Fictitious pages don't use "order" or "pool". */
1018 m->flags = PG_FICTITIOUS | PG_UNMANAGED | PG_BUSY;
1022 pmap_page_set_memattr(m, memattr);
1026 * Inserts the given vm_page into the object and object list.
1028 * The pagetables are not updated but will presumably fault the page
1029 * in if necessary, or if a kernel page the caller will at some point
1030 * enter the page into the kernel's pmap. We are not allowed to block
1031 * here so we *can't* do this anyway.
1033 * This routine may not block.
1034 * This routine must be called with the vm_object held.
1035 * This routine must be called with a critical section held.
1037 * This routine returns TRUE if the page was inserted into the object
1038 * successfully, and FALSE if the page already exists in the object.
1041 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1043 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object));
1044 if (m->object != NULL)
1045 panic("vm_page_insert: already inserted");
1047 object->generation++;
1050 * Record the object/offset pair in this page and add the
1051 * pv_list_count of the page to the object.
1053 * The vm_page spin lock is required for interactions with the pmap.
1055 vm_page_spin_lock(m);
1058 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
1061 vm_page_spin_unlock(m);
1064 ++object->resident_page_count;
1065 ++mycpu->gd_vmtotal.t_rm;
1066 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
1067 vm_page_spin_unlock(m);
1070 * Since we are inserting a new and possibly dirty page,
1071 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
1073 if ((m->valid & m->dirty) ||
1074 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT)))
1075 vm_object_set_writeable_dirty(object);
1078 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
1080 swap_pager_page_inserted(m);
1085 * Removes the given vm_page_t from the (object,index) table
1087 * The underlying pmap entry (if any) is NOT removed here.
1088 * This routine may not block.
1090 * The page must be BUSY and will remain BUSY on return.
1091 * No other requirements.
1093 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
1097 vm_page_remove(vm_page_t m)
1101 if (m->object == NULL) {
1105 if ((m->flags & PG_BUSY) == 0)
1106 panic("vm_page_remove: page not busy");
1110 vm_object_hold(object);
1113 * Remove the page from the object and update the object.
1115 * The vm_page spin lock is required for interactions with the pmap.
1117 vm_page_spin_lock(m);
1118 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
1119 --object->resident_page_count;
1120 --mycpu->gd_vmtotal.t_rm;
1121 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
1123 vm_page_spin_unlock(m);
1125 object->generation++;
1127 vm_object_drop(object);
1131 * Locate and return the page at (object, pindex), or NULL if the
1132 * page could not be found.
1134 * The caller must hold the vm_object token.
1137 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1142 * Search the hash table for this object/offset pair
1144 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1145 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1146 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1151 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1153 int also_m_busy, const char *msg
1159 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1160 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1162 KKASSERT(m->object == object && m->pindex == pindex);
1165 if (flags & PG_BUSY) {
1166 tsleep_interlock(m, 0);
1167 if (atomic_cmpset_int(&m->flags, flags,
1168 flags | PG_WANTED | PG_REFERENCED)) {
1169 tsleep(m, PINTERLOCKED, msg, 0);
1170 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1173 } else if (also_m_busy && (flags & PG_SBUSY)) {
1174 tsleep_interlock(m, 0);
1175 if (atomic_cmpset_int(&m->flags, flags,
1176 flags | PG_WANTED | PG_REFERENCED)) {
1177 tsleep(m, PINTERLOCKED, msg, 0);
1178 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1181 } else if (atomic_cmpset_int(&m->flags, flags,
1183 #ifdef VM_PAGE_DEBUG
1184 m->busy_func = func;
1185 m->busy_line = lineno;
1194 * Attempt to lookup and busy a page.
1196 * Returns NULL if the page could not be found
1198 * Returns a vm_page and error == TRUE if the page exists but could not
1201 * Returns a vm_page and error == FALSE on success.
1204 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1206 int also_m_busy, int *errorp
1212 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1213 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1216 KKASSERT(m->object == object && m->pindex == pindex);
1219 if (flags & PG_BUSY) {
1223 if (also_m_busy && (flags & PG_SBUSY)) {
1227 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1228 #ifdef VM_PAGE_DEBUG
1229 m->busy_func = func;
1230 m->busy_line = lineno;
1239 * Attempt to repurpose the passed-in page. If the passed-in page cannot
1240 * be repurposed it will be released, *must_reenter will be set to 1, and
1241 * this function will fall-through to vm_page_lookup_busy_try().
1243 * The passed-in page must be wired and not busy. The returned page will
1244 * be busied and not wired.
1246 * A different page may be returned. The returned page will be busied and
1249 * NULL can be returned. If so, the required page could not be busied.
1250 * The passed-in page will be unwired.
1253 vm_page_repurpose(struct vm_object *object, vm_pindex_t pindex,
1254 int also_m_busy, int *errorp, vm_page_t m,
1255 int *must_reenter, int *iswired)
1258 vm_page_busy_wait(m, TRUE, "biodep");
1259 if ((m->flags & (PG_UNMANAGED | PG_MAPPED | PG_FICTITIOUS)) ||
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 m = vm_page_lookup_busy_try(object, pindex, also_m_busy, errorp);
1297 * Caller must hold the related vm_object
1300 vm_page_next(vm_page_t m)
1304 next = vm_page_rb_tree_RB_NEXT(m);
1305 if (next && next->pindex != m->pindex + 1)
1313 * Move the given vm_page from its current object to the specified
1314 * target object/offset. The page must be busy and will remain so
1317 * new_object must be held.
1318 * This routine might block. XXX ?
1320 * NOTE: Swap associated with the page must be invalidated by the move. We
1321 * have to do this for several reasons: (1) we aren't freeing the
1322 * page, (2) we are dirtying the page, (3) the VM system is probably
1323 * moving the page from object A to B, and will then later move
1324 * the backing store from A to B and we can't have a conflict.
1326 * NOTE: We *always* dirty the page. It is necessary both for the
1327 * fact that we moved it, and because we may be invalidating
1328 * swap. If the page is on the cache, we have to deactivate it
1329 * or vm_page_dirty() will panic. Dirty pages are not allowed
1333 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1335 KKASSERT(m->flags & PG_BUSY);
1336 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object));
1338 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object));
1341 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1342 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1343 new_object, new_pindex);
1345 if (m->queue - m->pc == PQ_CACHE)
1346 vm_page_deactivate(m);
1351 * vm_page_unqueue() without any wakeup. This routine is used when a page
1352 * is to remain BUSYied by the caller.
1354 * This routine may not block.
1357 vm_page_unqueue_nowakeup(vm_page_t m)
1359 vm_page_and_queue_spin_lock(m);
1360 (void)_vm_page_rem_queue_spinlocked(m);
1361 vm_page_spin_unlock(m);
1365 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1368 * This routine may not block.
1371 vm_page_unqueue(vm_page_t m)
1375 vm_page_and_queue_spin_lock(m);
1376 queue = _vm_page_rem_queue_spinlocked(m);
1377 if (queue == PQ_FREE || queue == PQ_CACHE) {
1378 vm_page_spin_unlock(m);
1379 pagedaemon_wakeup();
1381 vm_page_spin_unlock(m);
1386 * vm_page_list_find()
1388 * Find a page on the specified queue with color optimization.
1390 * The page coloring optimization attempts to locate a page that does
1391 * not overload other nearby pages in the object in the cpu's L1 or L2
1392 * caches. We need this optimization because cpu caches tend to be
1393 * physical caches, while object spaces tend to be virtual.
1395 * The page coloring optimization also, very importantly, tries to localize
1396 * memory to cpus and physical sockets.
1398 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1399 * and the algorithm is adjusted to localize allocations on a per-core basis.
1400 * This is done by 'twisting' the colors.
1402 * The page is returned spinlocked and removed from its queue (it will
1403 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1404 * is responsible for dealing with the busy-page case (usually by
1405 * deactivating the page and looping).
1407 * NOTE: This routine is carefully inlined. A non-inlined version
1408 * is available for outside callers but the only critical path is
1409 * from within this source file.
1411 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1412 * represent stable storage, allowing us to order our locks vm_page
1413 * first, then queue.
1417 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1423 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl,
1426 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1429 m = _vm_page_list_find2(basequeue, index);
1432 vm_page_and_queue_spin_lock(m);
1433 if (m->queue == basequeue + index) {
1434 _vm_page_rem_queue_spinlocked(m);
1435 /* vm_page_t spin held, no queue spin */
1438 vm_page_and_queue_spin_unlock(m);
1444 * If we could not find the page in the desired queue try to find it in
1448 _vm_page_list_find2(int basequeue, int index)
1450 struct vpgqueues *pq;
1452 int pqmask = PQ_SET_ASSOC_MASK >> 1;
1456 index &= PQ_L2_MASK;
1457 pq = &vm_page_queues[basequeue];
1460 * Run local sets of 16, 32, 64, 128, and the whole queue if all
1461 * else fails (PQ_L2_MASK which is 255).
1464 pqmask = (pqmask << 1) | 1;
1465 for (i = 0; i <= pqmask; ++i) {
1466 pqi = (index & ~pqmask) | ((index + i) & pqmask);
1467 m = TAILQ_FIRST(&pq[pqi].pl);
1469 _vm_page_and_queue_spin_lock(m);
1470 if (m->queue == basequeue + pqi) {
1471 _vm_page_rem_queue_spinlocked(m);
1474 _vm_page_and_queue_spin_unlock(m);
1479 } while (pqmask != PQ_L2_MASK);
1485 * Returns a vm_page candidate for allocation. The page is not busied so
1486 * it can move around. The caller must busy the page (and typically
1487 * deactivate it if it cannot be busied!)
1489 * Returns a spinlocked vm_page that has been removed from its queue.
1492 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1494 return(_vm_page_list_find(basequeue, index, prefer_zero));
1498 * Find a page on the cache queue with color optimization, remove it
1499 * from the queue, and busy it. The returned page will not be spinlocked.
1501 * A candidate failure will be deactivated. Candidates can fail due to
1502 * being busied by someone else, in which case they will be deactivated.
1504 * This routine may not block.
1508 vm_page_select_cache(u_short pg_color)
1513 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE);
1517 * (m) has been removed from its queue and spinlocked
1519 if (vm_page_busy_try(m, TRUE)) {
1520 _vm_page_deactivate_locked(m, 0);
1521 vm_page_spin_unlock(m);
1524 * We successfully busied the page
1526 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 &&
1527 m->hold_count == 0 &&
1528 m->wire_count == 0 &&
1529 (m->dirty & m->valid) == 0) {
1530 vm_page_spin_unlock(m);
1531 pagedaemon_wakeup();
1536 * The page cannot be recycled, deactivate it.
1538 _vm_page_deactivate_locked(m, 0);
1539 if (_vm_page_wakeup(m)) {
1540 vm_page_spin_unlock(m);
1543 vm_page_spin_unlock(m);
1551 * Find a free or zero page, with specified preference. We attempt to
1552 * inline the nominal case and fall back to _vm_page_select_free()
1553 * otherwise. A busied page is removed from the queue and returned.
1555 * This routine may not block.
1557 static __inline vm_page_t
1558 vm_page_select_free(u_short pg_color, boolean_t prefer_zero)
1563 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK,
1567 if (vm_page_busy_try(m, TRUE)) {
1569 * Various mechanisms such as a pmap_collect can
1570 * result in a busy page on the free queue. We
1571 * have to move the page out of the way so we can
1572 * retry the allocation. If the other thread is not
1573 * allocating the page then m->valid will remain 0 and
1574 * the pageout daemon will free the page later on.
1576 * Since we could not busy the page, however, we
1577 * cannot make assumptions as to whether the page
1578 * will be allocated by the other thread or not,
1579 * so all we can do is deactivate it to move it out
1580 * of the way. In particular, if the other thread
1581 * wires the page it may wind up on the inactive
1582 * queue and the pageout daemon will have to deal
1583 * with that case too.
1585 _vm_page_deactivate_locked(m, 0);
1586 vm_page_spin_unlock(m);
1589 * Theoretically if we are able to busy the page
1590 * atomic with the queue removal (using the vm_page
1591 * lock) nobody else should be able to mess with the
1594 KKASSERT((m->flags & (PG_UNMANAGED |
1595 PG_NEED_COMMIT)) == 0);
1596 KASSERT(m->hold_count == 0, ("m->hold_count is not zero "
1597 "pg %p q=%d flags=%08x hold=%d wire=%d",
1598 m, m->queue, m->flags, m->hold_count, m->wire_count));
1599 KKASSERT(m->wire_count == 0);
1600 vm_page_spin_unlock(m);
1601 pagedaemon_wakeup();
1603 /* return busied and removed page */
1611 * This implements a per-cpu cache of free, zero'd, ready-to-go pages.
1612 * The idea is to populate this cache prior to acquiring any locks so
1613 * we don't wind up potentially zeroing VM pages (under heavy loads) while
1614 * holding potentialy contending locks.
1616 * Note that we allocate the page uninserted into anything and use a pindex
1617 * of 0, the vm_page_alloc() will effectively add gd_cpuid so these
1618 * allocations should wind up being uncontended. However, we still want
1619 * to rove across PQ_L2_SIZE.
1622 vm_page_pcpu_cache(void)
1625 globaldata_t gd = mycpu;
1628 if (gd->gd_vmpg_count < GD_MINVMPG) {
1630 while (gd->gd_vmpg_count < GD_MAXVMPG) {
1631 m = vm_page_alloc(NULL, ticks & ~ncpus2_mask,
1632 VM_ALLOC_NULL_OK | VM_ALLOC_NORMAL |
1633 VM_ALLOC_NULL_OK | VM_ALLOC_ZERO);
1634 if (gd->gd_vmpg_count < GD_MAXVMPG) {
1635 if ((m->flags & PG_ZERO) == 0) {
1636 pmap_zero_page(VM_PAGE_TO_PHYS(m));
1637 vm_page_flag_set(m, PG_ZERO);
1639 gd->gd_vmpg_array[gd->gd_vmpg_count++] = m;
1652 * Allocate and return a memory cell associated with this VM object/offset
1653 * pair. If object is NULL an unassociated page will be allocated.
1655 * The returned page will be busied and removed from its queues. This
1656 * routine can block and may return NULL if a race occurs and the page
1657 * is found to already exist at the specified (object, pindex).
1659 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1660 * VM_ALLOC_QUICK like normal but cannot use cache
1661 * VM_ALLOC_SYSTEM greater free drain
1662 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1663 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1664 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1665 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1666 * (see vm_page_grab())
1667 * VM_ALLOC_USE_GD ok to use per-gd cache
1669 * The object must be held if not NULL
1670 * This routine may not block
1672 * Additional special handling is required when called from an interrupt
1673 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1677 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1679 globaldata_t gd = mycpu;
1686 * Special per-cpu free VM page cache. The pages are pre-busied
1687 * and pre-zerod for us.
1689 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1691 if (gd->gd_vmpg_count) {
1692 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1704 * CPU localization algorithm. Break the page queues up by physical
1705 * id and core id (note that two cpu threads will have the same core
1706 * id, and core_id != gd_cpuid).
1708 * This is nowhere near perfect, for example the last pindex in a
1709 * subgroup will overflow into the next cpu or package. But this
1710 * should get us good page reuse locality in heavy mixed loads.
1712 pg_color = vm_get_pg_color(gd, object, pindex);
1715 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1716 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1719 * Certain system threads (pageout daemon, buf_daemon's) are
1720 * allowed to eat deeper into the free page list.
1722 if (curthread->td_flags & TDF_SYSTHREAD)
1723 page_req |= VM_ALLOC_SYSTEM;
1726 * Impose various limitations. Note that the v_free_reserved test
1727 * must match the opposite of vm_page_count_target() to avoid
1728 * livelocks, be careful.
1731 if (vmstats.v_free_count >= vmstats.v_free_reserved ||
1732 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
1733 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
1734 vmstats.v_free_count > vmstats.v_interrupt_free_min)
1737 * The free queue has sufficient free pages to take one out.
1739 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO))
1740 m = vm_page_select_free(pg_color, TRUE);
1742 m = vm_page_select_free(pg_color, FALSE);
1743 } else if (page_req & VM_ALLOC_NORMAL) {
1745 * Allocatable from the cache (non-interrupt only). On
1746 * success, we must free the page and try again, thus
1747 * ensuring that vmstats.v_*_free_min counters are replenished.
1750 if (curthread->td_preempted) {
1751 kprintf("vm_page_alloc(): warning, attempt to allocate"
1752 " cache page from preempting interrupt\n");
1755 m = vm_page_select_cache(pg_color);
1758 m = vm_page_select_cache(pg_color);
1761 * On success move the page into the free queue and loop.
1763 * Only do this if we can safely acquire the vm_object lock,
1764 * because this is effectively a random page and the caller
1765 * might be holding the lock shared, we don't want to
1769 KASSERT(m->dirty == 0,
1770 ("Found dirty cache page %p", m));
1771 if ((obj = m->object) != NULL) {
1772 if (vm_object_hold_try(obj)) {
1773 vm_page_protect(m, VM_PROT_NONE);
1775 /* m->object NULL here */
1776 vm_object_drop(obj);
1778 vm_page_deactivate(m);
1782 vm_page_protect(m, VM_PROT_NONE);
1789 * On failure return NULL
1791 #if defined(DIAGNOSTIC)
1792 if (vmstats.v_cache_count > 0)
1793 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
1795 vm_pageout_deficit++;
1796 pagedaemon_wakeup();
1800 * No pages available, wakeup the pageout daemon and give up.
1802 vm_pageout_deficit++;
1803 pagedaemon_wakeup();
1808 * v_free_count can race so loop if we don't find the expected
1815 * Good page found. The page has already been busied for us and
1816 * removed from its queues.
1818 KASSERT(m->dirty == 0,
1819 ("vm_page_alloc: free/cache page %p was dirty", m));
1820 KKASSERT(m->queue == PQ_NONE);
1826 * Initialize the structure, inheriting some flags but clearing
1827 * all the rest. The page has already been busied for us.
1829 vm_page_flag_clear(m, ~(PG_ZERO | PG_BUSY | PG_SBUSY));
1830 KKASSERT(m->wire_count == 0);
1831 KKASSERT(m->busy == 0);
1836 * Caller must be holding the object lock (asserted by
1837 * vm_page_insert()).
1839 * NOTE: Inserting a page here does not insert it into any pmaps
1840 * (which could cause us to block allocating memory).
1842 * NOTE: If no object an unassociated page is allocated, m->pindex
1843 * can be used by the caller for any purpose.
1846 if (vm_page_insert(m, object, pindex) == FALSE) {
1848 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1849 panic("PAGE RACE %p[%ld]/%p",
1850 object, (long)pindex, m);
1858 * Don't wakeup too often - wakeup the pageout daemon when
1859 * we would be nearly out of memory.
1861 pagedaemon_wakeup();
1864 * A PG_BUSY page is returned.
1870 * Returns number of pages available in our DMA memory reserve
1871 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf)
1874 vm_contig_avail_pages(void)
1879 spin_lock(&vm_contig_spin);
1880 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
1881 spin_unlock(&vm_contig_spin);
1887 * Attempt to allocate contiguous physical memory with the specified
1891 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
1892 unsigned long alignment, unsigned long boundary,
1893 unsigned long size, vm_memattr_t memattr)
1899 alignment >>= PAGE_SHIFT;
1902 boundary >>= PAGE_SHIFT;
1905 size = (size + PAGE_MASK) >> PAGE_SHIFT;
1907 spin_lock(&vm_contig_spin);
1908 blk = alist_alloc(&vm_contig_alist, 0, size);
1909 if (blk == ALIST_BLOCK_NONE) {
1910 spin_unlock(&vm_contig_spin);
1912 kprintf("vm_page_alloc_contig: %ldk nospace\n",
1913 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1917 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
1918 alist_free(&vm_contig_alist, blk, size);
1919 spin_unlock(&vm_contig_spin);
1921 kprintf("vm_page_alloc_contig: %ldk high "
1923 (size + PAGE_MASK) * (PAGE_SIZE / 1024),
1928 spin_unlock(&vm_contig_spin);
1929 if (vm_contig_verbose) {
1930 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
1931 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT,
1932 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
1935 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
1936 if (memattr != VM_MEMATTR_DEFAULT)
1937 for (i = 0;i < size;i++)
1938 pmap_page_set_memattr(&m[i], memattr);
1943 * Free contiguously allocated pages. The pages will be wired but not busy.
1944 * When freeing to the alist we leave them wired and not busy.
1947 vm_page_free_contig(vm_page_t m, unsigned long size)
1949 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
1950 vm_pindex_t start = pa >> PAGE_SHIFT;
1951 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
1953 if (vm_contig_verbose) {
1954 kprintf("vm_page_free_contig: %016jx/%ldk\n",
1955 (intmax_t)pa, size / 1024);
1957 if (pa < vm_low_phys_reserved) {
1958 KKASSERT(pa + size <= vm_low_phys_reserved);
1959 spin_lock(&vm_contig_spin);
1960 alist_free(&vm_contig_alist, start, pages);
1961 spin_unlock(&vm_contig_spin);
1964 vm_page_busy_wait(m, FALSE, "cpgfr");
1965 vm_page_unwire(m, 0);
1976 * Wait for sufficient free memory for nominal heavy memory use kernel
1979 * WARNING! Be sure never to call this in any vm_pageout code path, which
1980 * will trivially deadlock the system.
1983 vm_wait_nominal(void)
1985 while (vm_page_count_min(0))
1990 * Test if vm_wait_nominal() would block.
1993 vm_test_nominal(void)
1995 if (vm_page_count_min(0))
2001 * Block until free pages are available for allocation, called in various
2002 * places before memory allocations.
2004 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
2005 * more generous then that.
2011 * never wait forever
2015 lwkt_gettoken(&vm_token);
2017 if (curthread == pagethread) {
2019 * The pageout daemon itself needs pages, this is bad.
2021 if (vm_page_count_min(0)) {
2022 vm_pageout_pages_needed = 1;
2023 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
2027 * Wakeup the pageout daemon if necessary and wait.
2029 * Do not wait indefinitely for the target to be reached,
2030 * as load might prevent it from being reached any time soon.
2031 * But wait a little to try to slow down page allocations
2032 * and to give more important threads (the pagedaemon)
2033 * allocation priority.
2035 if (vm_page_count_target()) {
2036 if (vm_pages_needed == 0) {
2037 vm_pages_needed = 1;
2038 wakeup(&vm_pages_needed);
2040 ++vm_pages_waiting; /* SMP race ok */
2041 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
2044 lwkt_reltoken(&vm_token);
2048 * Block until free pages are available for allocation
2050 * Called only from vm_fault so that processes page faulting can be
2054 vm_wait_pfault(void)
2057 * Wakeup the pageout daemon if necessary and wait.
2059 * Do not wait indefinitely for the target to be reached,
2060 * as load might prevent it from being reached any time soon.
2061 * But wait a little to try to slow down page allocations
2062 * and to give more important threads (the pagedaemon)
2063 * allocation priority.
2065 if (vm_page_count_min(0)) {
2066 lwkt_gettoken(&vm_token);
2067 while (vm_page_count_severe()) {
2068 if (vm_page_count_target()) {
2069 if (vm_pages_needed == 0) {
2070 vm_pages_needed = 1;
2071 wakeup(&vm_pages_needed);
2073 ++vm_pages_waiting; /* SMP race ok */
2074 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
2077 lwkt_reltoken(&vm_token);
2082 * Put the specified page on the active list (if appropriate). Ensure
2083 * that act_count is at least ACT_INIT but do not otherwise mess with it.
2085 * The caller should be holding the page busied ? XXX
2086 * This routine may not block.
2089 vm_page_activate(vm_page_t m)
2093 vm_page_spin_lock(m);
2094 if (m->queue - m->pc != PQ_ACTIVE) {
2095 _vm_page_queue_spin_lock(m);
2096 oqueue = _vm_page_rem_queue_spinlocked(m);
2097 /* page is left spinlocked, queue is unlocked */
2099 if (oqueue == PQ_CACHE)
2100 mycpu->gd_cnt.v_reactivated++;
2101 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2102 if (m->act_count < ACT_INIT)
2103 m->act_count = ACT_INIT;
2104 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
2106 _vm_page_and_queue_spin_unlock(m);
2107 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
2108 pagedaemon_wakeup();
2110 if (m->act_count < ACT_INIT)
2111 m->act_count = ACT_INIT;
2112 vm_page_spin_unlock(m);
2117 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2118 * routine is called when a page has been added to the cache or free
2121 * This routine may not block.
2123 static __inline void
2124 vm_page_free_wakeup(void)
2127 * If the pageout daemon itself needs pages, then tell it that
2128 * there are some free.
2130 if (vm_pageout_pages_needed &&
2131 vmstats.v_cache_count + vmstats.v_free_count >=
2132 vmstats.v_pageout_free_min
2134 vm_pageout_pages_needed = 0;
2135 wakeup(&vm_pageout_pages_needed);
2139 * Wakeup processes that are waiting on memory.
2141 * Generally speaking we want to wakeup stuck processes as soon as
2142 * possible. !vm_page_count_min(0) is the absolute minimum point
2143 * where we can do this. Wait a bit longer to reduce degenerate
2144 * re-blocking (vm_page_free_hysteresis). The target check is just
2145 * to make sure the min-check w/hysteresis does not exceed the
2148 if (vm_pages_waiting) {
2149 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2150 !vm_page_count_target()) {
2151 vm_pages_waiting = 0;
2152 wakeup(&vmstats.v_free_count);
2153 ++mycpu->gd_cnt.v_ppwakeups;
2156 if (!vm_page_count_target()) {
2158 * Plenty of pages are free, wakeup everyone.
2160 vm_pages_waiting = 0;
2161 wakeup(&vmstats.v_free_count);
2162 ++mycpu->gd_cnt.v_ppwakeups;
2163 } else if (!vm_page_count_min(0)) {
2165 * Some pages are free, wakeup someone.
2167 int wcount = vm_pages_waiting;
2170 vm_pages_waiting = wcount;
2171 wakeup_one(&vmstats.v_free_count);
2172 ++mycpu->gd_cnt.v_ppwakeups;
2179 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
2180 * it from its VM object.
2182 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
2183 * return (the page will have been freed).
2186 vm_page_free_toq(vm_page_t m)
2188 mycpu->gd_cnt.v_tfree++;
2189 KKASSERT((m->flags & PG_MAPPED) == 0);
2190 KKASSERT(m->flags & PG_BUSY);
2192 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
2193 kprintf("vm_page_free: pindex(%lu), busy(%d), "
2194 "PG_BUSY(%d), hold(%d)\n",
2195 (u_long)m->pindex, m->busy,
2196 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count);
2197 if ((m->queue - m->pc) == PQ_FREE)
2198 panic("vm_page_free: freeing free page");
2200 panic("vm_page_free: freeing busy page");
2204 * Remove from object, spinlock the page and its queues and
2205 * remove from any queue. No queue spinlock will be held
2206 * after this section (because the page was removed from any
2210 vm_page_and_queue_spin_lock(m);
2211 _vm_page_rem_queue_spinlocked(m);
2214 * No further management of fictitious pages occurs beyond object
2215 * and queue removal.
2217 if ((m->flags & PG_FICTITIOUS) != 0) {
2218 vm_page_spin_unlock(m);
2226 if (m->wire_count != 0) {
2227 if (m->wire_count > 1) {
2229 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2230 m->wire_count, (long)m->pindex);
2232 panic("vm_page_free: freeing wired page");
2236 * Clear the UNMANAGED flag when freeing an unmanaged page.
2237 * Clear the NEED_COMMIT flag
2239 if (m->flags & PG_UNMANAGED)
2240 vm_page_flag_clear(m, PG_UNMANAGED);
2241 if (m->flags & PG_NEED_COMMIT)
2242 vm_page_flag_clear(m, PG_NEED_COMMIT);
2244 if (m->hold_count != 0) {
2245 vm_page_flag_clear(m, PG_ZERO);
2246 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2248 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
2252 * This sequence allows us to clear PG_BUSY while still holding
2253 * its spin lock, which reduces contention vs allocators. We
2254 * must not leave the queue locked or _vm_page_wakeup() may
2257 _vm_page_queue_spin_unlock(m);
2258 if (_vm_page_wakeup(m)) {
2259 vm_page_spin_unlock(m);
2262 vm_page_spin_unlock(m);
2264 vm_page_free_wakeup();
2268 * vm_page_free_fromq_fast()
2270 * Remove a non-zero page from one of the free queues; the page is removed for
2271 * zeroing, so do not issue a wakeup.
2273 * Our zeroidle code is now per-cpu so only do a limited scan. We try to
2274 * stay within a single cpu's domain but we do a little statistical
2275 * improvement by encompassing two cpu's domains worst-case.
2278 vm_page_free_fromq_fast(void)
2280 globaldata_t gd = mycpu;
2286 qi = vm_get_pg_color(gd, NULL, ++gd->gd_quick_color);
2287 qi = qi & PQ_L2_MASK;
2290 * 16 = one cpu's domain
2291 * 32 = two cpu's domains
2292 * (note masking at bottom of loop!)
2294 for (i = 0; i < 10; ++i) {
2295 m = vm_page_list_find(PQ_FREE, qi, FALSE);
2296 /* page is returned spinlocked and removed from its queue */
2298 if (vm_page_busy_try(m, TRUE)) {
2300 * We were unable to busy the page, deactivate
2303 _vm_page_deactivate_locked(m, 0);
2304 vm_page_spin_unlock(m);
2305 } else if (m->flags & PG_ZERO) {
2307 * The page is already PG_ZERO, requeue it
2310 _vm_page_add_queue_spinlocked(m,
2313 vm_page_queue_spin_unlock(m);
2314 if (_vm_page_wakeup(m)) {
2315 vm_page_spin_unlock(m);
2318 vm_page_spin_unlock(m);
2322 * The page is not PG_ZERO'd so return it.
2324 KKASSERT((m->flags & (PG_UNMANAGED |
2325 PG_NEED_COMMIT)) == 0);
2326 KKASSERT(m->hold_count == 0);
2327 KKASSERT(m->wire_count == 0);
2328 vm_page_spin_unlock(m);
2338 * vm_page_unmanage()
2340 * Prevent PV management from being done on the page. The page is
2341 * removed from the paging queues as if it were wired, and as a
2342 * consequence of no longer being managed the pageout daemon will not
2343 * touch it (since there is no way to locate the pte mappings for the
2344 * page). madvise() calls that mess with the pmap will also no longer
2345 * operate on the page.
2347 * Beyond that the page is still reasonably 'normal'. Freeing the page
2348 * will clear the flag.
2350 * This routine is used by OBJT_PHYS objects - objects using unswappable
2351 * physical memory as backing store rather then swap-backed memory and
2352 * will eventually be extended to support 4MB unmanaged physical
2355 * Caller must be holding the page busy.
2358 vm_page_unmanage(vm_page_t m)
2360 KKASSERT(m->flags & PG_BUSY);
2361 if ((m->flags & PG_UNMANAGED) == 0) {
2362 if (m->wire_count == 0)
2365 vm_page_flag_set(m, PG_UNMANAGED);
2369 * Mark this page as wired down by yet another map, removing it from
2370 * paging queues as necessary.
2372 * Caller must be holding the page busy.
2375 vm_page_wire(vm_page_t m)
2378 * Only bump the wire statistics if the page is not already wired,
2379 * and only unqueue the page if it is on some queue (if it is unmanaged
2380 * it is already off the queues). Don't do anything with fictitious
2381 * pages because they are always wired.
2383 KKASSERT(m->flags & PG_BUSY);
2384 if ((m->flags & PG_FICTITIOUS) == 0) {
2385 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2386 if ((m->flags & PG_UNMANAGED) == 0)
2388 atomic_add_int(&vmstats.v_wire_count, 1);
2390 KASSERT(m->wire_count != 0,
2391 ("vm_page_wire: wire_count overflow m=%p", m));
2396 * Release one wiring of this page, potentially enabling it to be paged again.
2398 * Many pages placed on the inactive queue should actually go
2399 * into the cache, but it is difficult to figure out which. What
2400 * we do instead, if the inactive target is well met, is to put
2401 * clean pages at the head of the inactive queue instead of the tail.
2402 * This will cause them to be moved to the cache more quickly and
2403 * if not actively re-referenced, freed more quickly. If we just
2404 * stick these pages at the end of the inactive queue, heavy filesystem
2405 * meta-data accesses can cause an unnecessary paging load on memory bound
2406 * processes. This optimization causes one-time-use metadata to be
2407 * reused more quickly.
2409 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2410 * the inactive queue. This helps the pageout daemon determine memory
2411 * pressure and act on out-of-memory situations more quickly.
2413 * BUT, if we are in a low-memory situation we have no choice but to
2414 * put clean pages on the cache queue.
2416 * A number of routines use vm_page_unwire() to guarantee that the page
2417 * will go into either the inactive or active queues, and will NEVER
2418 * be placed in the cache - for example, just after dirtying a page.
2419 * dirty pages in the cache are not allowed.
2421 * This routine may not block.
2424 vm_page_unwire(vm_page_t m, int activate)
2426 KKASSERT(m->flags & PG_BUSY);
2427 if (m->flags & PG_FICTITIOUS) {
2429 } else if (m->wire_count <= 0) {
2430 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2432 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2433 atomic_add_int(&vmstats.v_wire_count, -1);
2434 if (m->flags & PG_UNMANAGED) {
2436 } else if (activate || (m->flags & PG_NEED_COMMIT)) {
2437 vm_page_spin_lock(m);
2438 _vm_page_add_queue_spinlocked(m,
2439 PQ_ACTIVE + m->pc, 0);
2440 _vm_page_and_queue_spin_unlock(m);
2442 vm_page_spin_lock(m);
2443 vm_page_flag_clear(m, PG_WINATCFLS);
2444 _vm_page_add_queue_spinlocked(m,
2445 PQ_INACTIVE + m->pc, 0);
2446 ++vm_swapcache_inactive_heuristic;
2447 _vm_page_and_queue_spin_unlock(m);
2454 * Move the specified page to the inactive queue. If the page has
2455 * any associated swap, the swap is deallocated.
2457 * Normally athead is 0 resulting in LRU operation. athead is set
2458 * to 1 if we want this page to be 'as if it were placed in the cache',
2459 * except without unmapping it from the process address space.
2461 * vm_page's spinlock must be held on entry and will remain held on return.
2462 * This routine may not block.
2465 _vm_page_deactivate_locked(vm_page_t m, int athead)
2470 * Ignore if already inactive.
2472 if (m->queue - m->pc == PQ_INACTIVE)
2474 _vm_page_queue_spin_lock(m);
2475 oqueue = _vm_page_rem_queue_spinlocked(m);
2477 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2478 if (oqueue == PQ_CACHE)
2479 mycpu->gd_cnt.v_reactivated++;
2480 vm_page_flag_clear(m, PG_WINATCFLS);
2481 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2483 ++vm_swapcache_inactive_heuristic;
2485 /* NOTE: PQ_NONE if condition not taken */
2486 _vm_page_queue_spin_unlock(m);
2487 /* leaves vm_page spinlocked */
2491 * Attempt to deactivate a page.
2496 vm_page_deactivate(vm_page_t m)
2498 vm_page_spin_lock(m);
2499 _vm_page_deactivate_locked(m, 0);
2500 vm_page_spin_unlock(m);
2504 vm_page_deactivate_locked(vm_page_t m)
2506 _vm_page_deactivate_locked(m, 0);
2510 * Attempt to move a page to PQ_CACHE.
2512 * Returns 0 on failure, 1 on success
2514 * The page should NOT be busied by the caller. This function will validate
2515 * whether the page can be safely moved to the cache.
2518 vm_page_try_to_cache(vm_page_t m)
2520 vm_page_spin_lock(m);
2521 if (vm_page_busy_try(m, TRUE)) {
2522 vm_page_spin_unlock(m);
2525 if (m->dirty || m->hold_count || m->wire_count ||
2526 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2527 if (_vm_page_wakeup(m)) {
2528 vm_page_spin_unlock(m);
2531 vm_page_spin_unlock(m);
2535 vm_page_spin_unlock(m);
2538 * Page busied by us and no longer spinlocked. Dirty pages cannot
2539 * be moved to the cache.
2541 vm_page_test_dirty(m);
2542 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2551 * Attempt to free the page. If we cannot free it, we do nothing.
2552 * 1 is returned on success, 0 on failure.
2557 vm_page_try_to_free(vm_page_t m)
2559 vm_page_spin_lock(m);
2560 if (vm_page_busy_try(m, TRUE)) {
2561 vm_page_spin_unlock(m);
2566 * The page can be in any state, including already being on the free
2567 * queue. Check to see if it really can be freed.
2569 if (m->dirty || /* can't free if it is dirty */
2570 m->hold_count || /* or held (XXX may be wrong) */
2571 m->wire_count || /* or wired */
2572 (m->flags & (PG_UNMANAGED | /* or unmanaged */
2573 PG_NEED_COMMIT)) || /* or needs a commit */
2574 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */
2575 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */
2576 if (_vm_page_wakeup(m)) {
2577 vm_page_spin_unlock(m);
2580 vm_page_spin_unlock(m);
2584 vm_page_spin_unlock(m);
2587 * We can probably free the page.
2589 * Page busied by us and no longer spinlocked. Dirty pages will
2590 * not be freed by this function. We have to re-test the
2591 * dirty bit after cleaning out the pmaps.
2593 vm_page_test_dirty(m);
2594 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2598 vm_page_protect(m, VM_PROT_NONE);
2599 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2610 * Put the specified page onto the page cache queue (if appropriate).
2612 * The page must be busy, and this routine will release the busy and
2613 * possibly even free the page.
2616 vm_page_cache(vm_page_t m)
2618 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2619 m->busy || m->wire_count || m->hold_count) {
2620 kprintf("vm_page_cache: attempting to cache busy/held page\n");
2626 * Already in the cache (and thus not mapped)
2628 if ((m->queue - m->pc) == PQ_CACHE) {
2629 KKASSERT((m->flags & PG_MAPPED) == 0);
2635 * Caller is required to test m->dirty, but note that the act of
2636 * removing the page from its maps can cause it to become dirty
2637 * on an SMP system due to another cpu running in usermode.
2640 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2645 * Remove all pmaps and indicate that the page is not
2646 * writeable or mapped. Our vm_page_protect() call may
2647 * have blocked (especially w/ VM_PROT_NONE), so recheck
2650 vm_page_protect(m, VM_PROT_NONE);
2651 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2652 m->busy || m->wire_count || m->hold_count) {
2654 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2655 vm_page_deactivate(m);
2658 _vm_page_and_queue_spin_lock(m);
2659 _vm_page_rem_queue_spinlocked(m);
2660 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2661 _vm_page_queue_spin_unlock(m);
2662 if (_vm_page_wakeup(m)) {
2663 vm_page_spin_unlock(m);
2666 vm_page_spin_unlock(m);
2668 vm_page_free_wakeup();
2673 * vm_page_dontneed()
2675 * Cache, deactivate, or do nothing as appropriate. This routine
2676 * is typically used by madvise() MADV_DONTNEED.
2678 * Generally speaking we want to move the page into the cache so
2679 * it gets reused quickly. However, this can result in a silly syndrome
2680 * due to the page recycling too quickly. Small objects will not be
2681 * fully cached. On the otherhand, if we move the page to the inactive
2682 * queue we wind up with a problem whereby very large objects
2683 * unnecessarily blow away our inactive and cache queues.
2685 * The solution is to move the pages based on a fixed weighting. We
2686 * either leave them alone, deactivate them, or move them to the cache,
2687 * where moving them to the cache has the highest weighting.
2688 * By forcing some pages into other queues we eventually force the
2689 * system to balance the queues, potentially recovering other unrelated
2690 * space from active. The idea is to not force this to happen too
2693 * The page must be busied.
2696 vm_page_dontneed(vm_page_t m)
2698 static int dnweight;
2705 * occassionally leave the page alone
2707 if ((dnw & 0x01F0) == 0 ||
2708 m->queue - m->pc == PQ_INACTIVE ||
2709 m->queue - m->pc == PQ_CACHE
2711 if (m->act_count >= ACT_INIT)
2717 * If vm_page_dontneed() is inactivating a page, it must clear
2718 * the referenced flag; otherwise the pagedaemon will see references
2719 * on the page in the inactive queue and reactivate it. Until the
2720 * page can move to the cache queue, madvise's job is not done.
2722 vm_page_flag_clear(m, PG_REFERENCED);
2723 pmap_clear_reference(m);
2726 vm_page_test_dirty(m);
2728 if (m->dirty || (dnw & 0x0070) == 0) {
2730 * Deactivate the page 3 times out of 32.
2735 * Cache the page 28 times out of every 32. Note that
2736 * the page is deactivated instead of cached, but placed
2737 * at the head of the queue instead of the tail.
2741 vm_page_spin_lock(m);
2742 _vm_page_deactivate_locked(m, head);
2743 vm_page_spin_unlock(m);
2747 * These routines manipulate the 'soft busy' count for a page. A soft busy
2748 * is almost like PG_BUSY except that it allows certain compatible operations
2749 * to occur on the page while it is busy. For example, a page undergoing a
2750 * write can still be mapped read-only.
2752 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2753 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2754 * busy bit is cleared.
2757 vm_page_io_start(vm_page_t m)
2759 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2760 atomic_add_char(&m->busy, 1);
2761 vm_page_flag_set(m, PG_SBUSY);
2765 vm_page_io_finish(vm_page_t m)
2767 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2768 atomic_subtract_char(&m->busy, 1);
2770 vm_page_flag_clear(m, PG_SBUSY);
2774 * Indicate that a clean VM page requires a filesystem commit and cannot
2775 * be reused. Used by tmpfs.
2778 vm_page_need_commit(vm_page_t m)
2780 vm_page_flag_set(m, PG_NEED_COMMIT);
2781 vm_object_set_writeable_dirty(m->object);
2785 vm_page_clear_commit(vm_page_t m)
2787 vm_page_flag_clear(m, PG_NEED_COMMIT);
2791 * Grab a page, blocking if it is busy and allocating a page if necessary.
2792 * A busy page is returned or NULL. The page may or may not be valid and
2793 * might not be on a queue (the caller is responsible for the disposition of
2796 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2797 * page will be zero'd and marked valid.
2799 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2800 * valid even if it already exists.
2802 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2803 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2804 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2806 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2807 * always returned if we had blocked.
2809 * This routine may not be called from an interrupt.
2811 * PG_ZERO is *ALWAYS* cleared by this routine.
2813 * No other requirements.
2816 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2822 KKASSERT(allocflags &
2823 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2824 vm_object_hold_shared(object);
2826 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2828 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2829 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2834 } else if (m == NULL) {
2836 vm_object_upgrade(object);
2839 if (allocflags & VM_ALLOC_RETRY)
2840 allocflags |= VM_ALLOC_NULL_OK;
2841 m = vm_page_alloc(object, pindex,
2842 allocflags & ~VM_ALLOC_RETRY);
2846 if ((allocflags & VM_ALLOC_RETRY) == 0)
2855 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2857 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2858 * valid even if already valid.
2860 if (m->valid == 0) {
2861 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2862 if ((m->flags & PG_ZERO) == 0)
2863 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2864 m->valid = VM_PAGE_BITS_ALL;
2866 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2867 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2868 m->valid = VM_PAGE_BITS_ALL;
2870 vm_page_flag_clear(m, PG_ZERO);
2872 vm_object_drop(object);
2877 * Mapping function for valid bits or for dirty bits in
2878 * a page. May not block.
2880 * Inputs are required to range within a page.
2886 vm_page_bits(int base, int size)
2892 base + size <= PAGE_SIZE,
2893 ("vm_page_bits: illegal base/size %d/%d", base, size)
2896 if (size == 0) /* handle degenerate case */
2899 first_bit = base >> DEV_BSHIFT;
2900 last_bit = (base + size - 1) >> DEV_BSHIFT;
2902 return ((2 << last_bit) - (1 << first_bit));
2906 * Sets portions of a page valid and clean. The arguments are expected
2907 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2908 * of any partial chunks touched by the range. The invalid portion of
2909 * such chunks will be zero'd.
2911 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2912 * align base to DEV_BSIZE so as not to mark clean a partially
2913 * truncated device block. Otherwise the dirty page status might be
2916 * This routine may not block.
2918 * (base + size) must be less then or equal to PAGE_SIZE.
2921 _vm_page_zero_valid(vm_page_t m, int base, int size)
2926 if (size == 0) /* handle degenerate case */
2930 * If the base is not DEV_BSIZE aligned and the valid
2931 * bit is clear, we have to zero out a portion of the
2935 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2936 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2938 pmap_zero_page_area(
2946 * If the ending offset is not DEV_BSIZE aligned and the
2947 * valid bit is clear, we have to zero out a portion of
2951 endoff = base + size;
2953 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2954 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
2956 pmap_zero_page_area(
2959 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
2965 * Set valid, clear dirty bits. If validating the entire
2966 * page we can safely clear the pmap modify bit. We also
2967 * use this opportunity to clear the PG_NOSYNC flag. If a process
2968 * takes a write fault on a MAP_NOSYNC memory area the flag will
2971 * We set valid bits inclusive of any overlap, but we can only
2972 * clear dirty bits for DEV_BSIZE chunks that are fully within
2975 * Page must be busied?
2976 * No other requirements.
2979 vm_page_set_valid(vm_page_t m, int base, int size)
2981 _vm_page_zero_valid(m, base, size);
2982 m->valid |= vm_page_bits(base, size);
2987 * Set valid bits and clear dirty bits.
2989 * NOTE: This function does not clear the pmap modified bit.
2990 * Also note that e.g. NFS may use a byte-granular base
2993 * WARNING: Page must be busied? But vfs_clean_one_page() will call
2994 * this without necessarily busying the page (via bdwrite()).
2995 * So for now vm_token must also be held.
2997 * No other requirements.
3000 vm_page_set_validclean(vm_page_t m, int base, int size)
3004 _vm_page_zero_valid(m, base, size);
3005 pagebits = vm_page_bits(base, size);
3006 m->valid |= pagebits;
3007 m->dirty &= ~pagebits;
3008 if (base == 0 && size == PAGE_SIZE) {
3009 /*pmap_clear_modify(m);*/
3010 vm_page_flag_clear(m, PG_NOSYNC);
3015 * Set valid & dirty. Used by buwrite()
3017 * WARNING: Page must be busied? But vfs_dirty_one_page() will
3018 * call this function in buwrite() so for now vm_token must
3021 * No other requirements.
3024 vm_page_set_validdirty(vm_page_t m, int base, int size)
3028 pagebits = vm_page_bits(base, size);
3029 m->valid |= pagebits;
3030 m->dirty |= pagebits;
3032 vm_object_set_writeable_dirty(m->object);
3038 * NOTE: This function does not clear the pmap modified bit.
3039 * Also note that e.g. NFS may use a byte-granular base
3042 * Page must be busied?
3043 * No other requirements.
3046 vm_page_clear_dirty(vm_page_t m, int base, int size)
3048 m->dirty &= ~vm_page_bits(base, size);
3049 if (base == 0 && size == PAGE_SIZE) {
3050 /*pmap_clear_modify(m);*/
3051 vm_page_flag_clear(m, PG_NOSYNC);
3056 * Make the page all-dirty.
3058 * Also make sure the related object and vnode reflect the fact that the
3059 * object may now contain a dirty page.
3061 * Page must be busied?
3062 * No other requirements.
3065 vm_page_dirty(vm_page_t m)
3068 int pqtype = m->queue - m->pc;
3070 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
3071 ("vm_page_dirty: page in free/cache queue!"));
3072 if (m->dirty != VM_PAGE_BITS_ALL) {
3073 m->dirty = VM_PAGE_BITS_ALL;
3075 vm_object_set_writeable_dirty(m->object);
3080 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3081 * valid and dirty bits for the effected areas are cleared.
3083 * Page must be busied?
3085 * No other requirements.
3088 vm_page_set_invalid(vm_page_t m, int base, int size)
3092 bits = vm_page_bits(base, size);
3095 m->object->generation++;
3099 * The kernel assumes that the invalid portions of a page contain
3100 * garbage, but such pages can be mapped into memory by user code.
3101 * When this occurs, we must zero out the non-valid portions of the
3102 * page so user code sees what it expects.
3104 * Pages are most often semi-valid when the end of a file is mapped
3105 * into memory and the file's size is not page aligned.
3107 * Page must be busied?
3108 * No other requirements.
3111 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3117 * Scan the valid bits looking for invalid sections that
3118 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
3119 * valid bit may be set ) have already been zerod by
3120 * vm_page_set_validclean().
3122 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3123 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3124 (m->valid & (1 << i))
3127 pmap_zero_page_area(
3130 (i - b) << DEV_BSHIFT
3138 * setvalid is TRUE when we can safely set the zero'd areas
3139 * as being valid. We can do this if there are no cache consistency
3140 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3143 m->valid = VM_PAGE_BITS_ALL;
3147 * Is a (partial) page valid? Note that the case where size == 0
3148 * will return FALSE in the degenerate case where the page is entirely
3149 * invalid, and TRUE otherwise.
3152 * No other requirements.
3155 vm_page_is_valid(vm_page_t m, int base, int size)
3157 int bits = vm_page_bits(base, size);
3159 if (m->valid && ((m->valid & bits) == bits))
3166 * update dirty bits from pmap/mmu. May not block.
3168 * Caller must hold the page busy
3171 vm_page_test_dirty(vm_page_t m)
3173 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
3179 * Register an action, associating it with its vm_page
3182 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
3184 struct vm_page_action_list *list;
3187 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
3188 list = &action_list[hv];
3190 lwkt_gettoken(&vm_token);
3191 vm_page_flag_set(action->m, PG_ACTIONLIST);
3192 action->event = event;
3193 LIST_INSERT_HEAD(list, action, entry);
3194 lwkt_reltoken(&vm_token);
3198 * Unregister an action, disassociating it from its related vm_page
3201 vm_page_unregister_action(vm_page_action_t action)
3203 struct vm_page_action_list *list;
3206 lwkt_gettoken(&vm_token);
3207 if (action->event != VMEVENT_NONE) {
3208 action->event = VMEVENT_NONE;
3209 LIST_REMOVE(action, entry);
3211 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
3212 list = &action_list[hv];
3213 if (LIST_EMPTY(list))
3214 vm_page_flag_clear(action->m, PG_ACTIONLIST);
3216 lwkt_reltoken(&vm_token);
3220 * Issue an event on a VM page. Corresponding action structures are
3221 * removed from the page's list and called.
3223 * If the vm_page has no more pending action events we clear its
3224 * PG_ACTIONLIST flag.
3227 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
3229 struct vm_page_action_list *list;
3230 struct vm_page_action *scan;
3231 struct vm_page_action *next;
3235 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
3236 list = &action_list[hv];
3239 lwkt_gettoken(&vm_token);
3240 LIST_FOREACH_MUTABLE(scan, list, entry, next) {
3242 if (scan->event == event) {
3243 scan->event = VMEVENT_NONE;
3244 LIST_REMOVE(scan, entry);
3245 scan->func(m, scan);
3253 vm_page_flag_clear(m, PG_ACTIONLIST);
3254 lwkt_reltoken(&vm_token);
3257 #include "opt_ddb.h"
3259 #include <sys/kernel.h>
3261 #include <ddb/ddb.h>
3263 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3265 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
3266 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
3267 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
3268 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
3269 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
3270 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
3271 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
3272 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
3273 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
3274 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
3277 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3280 db_printf("PQ_FREE:");
3281 for(i=0;i<PQ_L2_SIZE;i++) {
3282 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
3286 db_printf("PQ_CACHE:");
3287 for(i=0;i<PQ_L2_SIZE;i++) {
3288 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
3292 db_printf("PQ_ACTIVE:");
3293 for(i=0;i<PQ_L2_SIZE;i++) {
3294 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3298 db_printf("PQ_INACTIVE:");
3299 for(i=0;i<PQ_L2_SIZE;i++) {
3300 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);