2 * Copyright (c) 1991 Regents of the University of California.
5 * This code is derived from software contributed to Berkeley by
6 * The Mach Operating System project at Carnegie-Mellon University.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 * 3. Neither the name of the University nor the names of its contributors
17 * may be used to endorse or promote products derived from this software
18 * without specific prior written permission.
20 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
21 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
22 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
23 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
24 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
25 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
26 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
27 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
28 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
29 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
33 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $
37 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
38 * All rights reserved.
40 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
42 * Permission to use, copy, modify and distribute this software and
43 * its documentation is hereby granted, provided that both the copyright
44 * notice and this permission notice appear in all copies of the
45 * software, derivative works or modified versions, and any portions
46 * thereof, and that both notices appear in supporting documentation.
48 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
49 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
50 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
52 * Carnegie Mellon requests users of this software to return to
54 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
55 * School of Computer Science
56 * Carnegie Mellon University
57 * Pittsburgh PA 15213-3890
59 * any improvements or extensions that they make and grant Carnegie the
60 * rights to redistribute these changes.
63 * Resident memory management module. The module manipulates 'VM pages'.
64 * A VM page is the core building block for memory management.
67 #include <sys/param.h>
68 #include <sys/systm.h>
69 #include <sys/malloc.h>
71 #include <sys/vmmeter.h>
72 #include <sys/vnode.h>
73 #include <sys/kernel.h>
74 #include <sys/alist.h>
75 #include <sys/sysctl.h>
76 #include <sys/cpu_topology.h>
79 #include <vm/vm_param.h>
81 #include <vm/vm_kern.h>
83 #include <vm/vm_map.h>
84 #include <vm/vm_object.h>
85 #include <vm/vm_page.h>
86 #include <vm/vm_pageout.h>
87 #include <vm/vm_pager.h>
88 #include <vm/vm_extern.h>
89 #include <vm/swap_pager.h>
91 #include <machine/inttypes.h>
92 #include <machine/md_var.h>
93 #include <machine/specialreg.h>
95 #include <vm/vm_page2.h>
96 #include <sys/spinlock2.h>
99 * Action hash for user umtx support.
101 #define VMACTION_HSIZE 256
102 #define VMACTION_HMASK (VMACTION_HSIZE - 1)
105 * SET - Minimum required set associative size, must be a power of 2. We
106 * want this to match or exceed the set-associativeness of the cpu.
108 * GRP - A larger set that allows bleed-over into the domains of other
109 * nearby cpus. Also must be a power of 2. Used by the page zeroing
110 * code to smooth things out a bit.
112 #define PQ_SET_ASSOC 16
113 #define PQ_SET_ASSOC_MASK (PQ_SET_ASSOC - 1)
115 #define PQ_GRP_ASSOC (PQ_SET_ASSOC * 2)
116 #define PQ_GRP_ASSOC_MASK (PQ_GRP_ASSOC - 1)
118 static void vm_page_queue_init(void);
119 static void vm_page_free_wakeup(void);
120 static vm_page_t vm_page_select_cache(u_short pg_color);
121 static vm_page_t _vm_page_list_find2(int basequeue, int index);
122 static void _vm_page_deactivate_locked(vm_page_t m, int athead);
125 * Array of tailq lists
127 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT];
129 LIST_HEAD(vm_page_action_list, vm_page_action);
130 struct vm_page_action_list action_list[VMACTION_HSIZE];
131 static volatile int vm_pages_waiting;
133 static struct alist vm_contig_alist;
134 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536];
135 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin");
137 static u_long vm_dma_reserved = 0;
138 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved);
139 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0,
140 "Memory reserved for DMA");
141 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD,
142 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA");
144 static int vm_contig_verbose = 0;
145 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose);
147 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
148 vm_pindex_t, pindex);
151 vm_page_queue_init(void)
155 for (i = 0; i < PQ_L2_SIZE; i++)
156 vm_page_queues[PQ_FREE+i].cnt = &vmstats.v_free_count;
157 for (i = 0; i < PQ_L2_SIZE; i++)
158 vm_page_queues[PQ_CACHE+i].cnt = &vmstats.v_cache_count;
159 for (i = 0; i < PQ_L2_SIZE; i++)
160 vm_page_queues[PQ_INACTIVE+i].cnt = &vmstats.v_inactive_count;
161 for (i = 0; i < PQ_L2_SIZE; i++)
162 vm_page_queues[PQ_ACTIVE+i].cnt = &vmstats.v_active_count;
163 for (i = 0; i < PQ_L2_SIZE; i++)
164 vm_page_queues[PQ_HOLD+i].cnt = &vmstats.v_active_count;
165 /* PQ_NONE has no queue */
167 for (i = 0; i < PQ_COUNT; i++) {
168 TAILQ_INIT(&vm_page_queues[i].pl);
169 spin_init(&vm_page_queues[i].spin, "vm_page_queue_init");
172 for (i = 0; i < VMACTION_HSIZE; i++)
173 LIST_INIT(&action_list[i]);
177 * note: place in initialized data section? Is this necessary?
180 int vm_page_array_size = 0;
181 vm_page_t vm_page_array = NULL;
182 vm_paddr_t vm_low_phys_reserved;
187 * Sets the page size, perhaps based upon the memory size.
188 * Must be called before any use of page-size dependent functions.
191 vm_set_page_size(void)
193 if (vmstats.v_page_size == 0)
194 vmstats.v_page_size = PAGE_SIZE;
195 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
196 panic("vm_set_page_size: page size not a power of two");
202 * Add a new page to the freelist for use by the system. New pages
203 * are added to both the head and tail of the associated free page
204 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
205 * requests pull 'recent' adds (higher physical addresses) first.
207 * Beware that the page zeroing daemon will also be running soon after
208 * boot, moving pages from the head to the tail of the PQ_FREE queues.
210 * Must be called in a critical section.
213 vm_add_new_page(vm_paddr_t pa)
215 struct vpgqueues *vpq;
218 m = PHYS_TO_VM_PAGE(pa);
221 m->pc = (pa >> PAGE_SHIFT) & PQ_L2_MASK;
222 m->pat_mode = PAT_WRITE_BACK;
225 * Twist for cpu localization in addition to page coloring, so
226 * different cpus selecting by m->queue get different page colors.
228 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE) & PQ_L2_MASK;
229 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;
292 vaddr = round_page(vaddr);
295 * Make sure ranges are page-aligned.
297 for (i = 0; phys_avail[i].phys_end; ++i) {
298 phys_avail[i].phys_beg = round_page64(phys_avail[i].phys_beg);
299 phys_avail[i].phys_end = trunc_page64(phys_avail[i].phys_end);
300 if (phys_avail[i].phys_end < phys_avail[i].phys_beg)
301 phys_avail[i].phys_end = phys_avail[i].phys_beg;
305 * Locate largest block
307 for (i = 0; phys_avail[i].phys_end; ++i) {
308 vm_paddr_t size = phys_avail[i].phys_end -
309 phys_avail[i].phys_beg;
311 if (size > biggestsize) {
317 --i; /* adjust to last entry for use down below */
319 end = phys_avail[biggestone].phys_end;
320 end = trunc_page(end);
323 * Initialize the queue headers for the free queue, the active queue
324 * and the inactive queue.
326 vm_page_queue_init();
328 #if !defined(_KERNEL_VIRTUAL)
330 * VKERNELs don't support minidumps and as such don't need
333 * Allocate a bitmap to indicate that a random physical page
334 * needs to be included in a minidump.
336 * The amd64 port needs this to indicate which direct map pages
337 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
339 * However, i386 still needs this workspace internally within the
340 * minidump code. In theory, they are not needed on i386, but are
341 * included should the sf_buf code decide to use them.
343 page_range = phys_avail[i].phys_end / PAGE_SIZE;
344 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
345 end -= vm_page_dump_size;
346 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
347 VM_PROT_READ | VM_PROT_WRITE);
348 bzero((void *)vm_page_dump, vm_page_dump_size);
351 * Compute the number of pages of memory that will be available for
352 * use (taking into account the overhead of a page structure per
355 first_page = phys_avail[0].phys_beg / PAGE_SIZE;
356 page_range = phys_avail[i].phys_end / PAGE_SIZE - first_page;
357 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
359 #ifndef _KERNEL_VIRTUAL
361 * (only applies to real kernels)
363 * Reserve a large amount of low memory for potential 32-bit DMA
364 * space allocations. Once device initialization is complete we
365 * release most of it, but keep (vm_dma_reserved) memory reserved
366 * for later use. Typically for X / graphics. Through trial and
367 * error we find that GPUs usually requires ~60-100MB or so.
369 * By default, 128M is left in reserve on machines with 2G+ of ram.
371 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT;
372 if (vm_low_phys_reserved > total / 4)
373 vm_low_phys_reserved = total / 4;
374 if (vm_dma_reserved == 0) {
375 vm_dma_reserved = 128 * 1024 * 1024; /* 128MB */
376 if (vm_dma_reserved > total / 16)
377 vm_dma_reserved = total / 16;
380 alist_init(&vm_contig_alist, 65536, vm_contig_ameta,
381 ALIST_RECORDS_65536);
384 * Initialize the mem entry structures now, and put them in the free
387 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
388 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE);
389 vm_page_array = (vm_page_t)mapped;
391 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
393 * since pmap_map on amd64 returns stuff out of a direct-map region,
394 * we have to manually add these pages to the minidump tracking so
395 * that they can be dumped, including the vm_page_array.
398 pa < phys_avail[biggestone].phys_end;
405 * Clear all of the page structures, run basic initialization so
406 * PHYS_TO_VM_PAGE() operates properly even on pages not in the
409 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
410 vm_page_array_size = page_range;
412 m = &vm_page_array[0];
413 pa = ptoa(first_page);
414 for (i = 0; i < page_range; ++i) {
415 spin_init(&m->spin, "vm_page");
422 * Construct the free queue(s) in ascending order (by physical
423 * address) so that the first 16MB of physical memory is allocated
424 * last rather than first. On large-memory machines, this avoids
425 * the exhaustion of low physical memory before isa_dmainit has run.
427 vmstats.v_page_count = 0;
428 vmstats.v_free_count = 0;
429 for (i = 0; phys_avail[i].phys_end && npages > 0; ++i) {
430 pa = phys_avail[i].phys_beg;
434 last_pa = phys_avail[i].phys_end;
435 while (pa < last_pa && npages-- > 0) {
441 virtual2_start = vaddr;
443 virtual_start = vaddr;
447 * Reorganize VM pages based on numa data. May be called as many times as
448 * necessary. Will reorganize the vm_page_t page color and related queue(s)
449 * to allow vm_page_alloc() to choose pages based on socket affinity.
451 * NOTE: This function is only called while we are still in UP mode, so
452 * we only need a critical section to protect the queues (which
453 * saves a lot of time, there are likely a ton of pages).
456 vm_numa_organize(vm_paddr_t ran_beg, vm_paddr_t bytes, int physid)
461 struct vpgqueues *vpq;
469 * Check if no physical information, or there was only one socket
470 * (so don't waste time doing nothing!).
472 if (cpu_topology_phys_ids <= 1 ||
473 cpu_topology_core_ids == 0) {
478 * Setup for our iteration. Note that ACPI may iterate CPU
479 * sockets starting at 0 or 1 or some other number. The
480 * cpu_topology code mod's it against the socket count.
482 ran_end = ran_beg + bytes;
483 physid %= cpu_topology_phys_ids;
485 socket_mod = PQ_L2_SIZE / cpu_topology_phys_ids;
486 socket_value = physid * socket_mod;
487 mend = &vm_page_array[vm_page_array_size];
492 * Adjust vm_page->pc and requeue all affected pages. The
493 * allocator will then be able to localize memory allocations
496 for (i = 0; phys_avail[i].phys_end; ++i) {
497 scan_beg = phys_avail[i].phys_beg;
498 scan_end = phys_avail[i].phys_end;
499 if (scan_end <= ran_beg)
501 if (scan_beg >= ran_end)
503 if (scan_beg < ran_beg)
505 if (scan_end > ran_end)
507 if (atop(scan_end) > first_page + vm_page_array_size)
508 scan_end = ptoa(first_page + vm_page_array_size);
510 m = PHYS_TO_VM_PAGE(scan_beg);
511 while (scan_beg < scan_end) {
513 if (m->queue != PQ_NONE) {
514 vpq = &vm_page_queues[m->queue];
515 TAILQ_REMOVE(&vpq->pl, m, pageq);
517 /* queue doesn't change, no need to adj cnt */
518 /* atomic_add_int(vpq->cnt, -1); */
521 m->pc += socket_value;
524 vpq = &vm_page_queues[m->queue];
525 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
527 /* queue doesn't change, no need to adj cnt */
528 /* atomic_add_int(vpq->cnt, 1); */
531 m->pc += socket_value;
534 scan_beg += PAGE_SIZE;
542 * We tended to reserve a ton of memory for contigmalloc(). Now that most
543 * drivers have initialized we want to return most the remaining free
544 * reserve back to the VM page queues so they can be used for normal
547 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool.
550 vm_page_startup_finish(void *dummy __unused)
559 spin_lock(&vm_contig_spin);
561 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
562 if (bfree <= vm_dma_reserved / PAGE_SIZE)
568 * Figure out how much of the initial reserve we have to
569 * free in order to reach our target.
571 bfree -= vm_dma_reserved / PAGE_SIZE;
573 blk += count - bfree;
578 * Calculate the nearest power of 2 <= count.
580 for (xcount = 1; xcount <= count; xcount <<= 1)
583 blk += count - xcount;
587 * Allocate the pages from the alist, then free them to
588 * the normal VM page queues.
590 * Pages allocated from the alist are wired. We have to
591 * busy, unwire, and free them. We must also adjust
592 * vm_low_phys_reserved before freeing any pages to prevent
595 rblk = alist_alloc(&vm_contig_alist, blk, count);
597 kprintf("vm_page_startup_finish: Unable to return "
598 "dma space @0x%08x/%d -> 0x%08x\n",
602 atomic_add_int(&vmstats.v_dma_pages, -count);
603 spin_unlock(&vm_contig_spin);
605 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
606 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m);
608 vm_page_busy_wait(m, FALSE, "cpgfr");
609 vm_page_unwire(m, 0);
614 spin_lock(&vm_contig_spin);
616 spin_unlock(&vm_contig_spin);
619 * Print out how much DMA space drivers have already allocated and
620 * how much is left over.
622 kprintf("DMA space used: %jdk, remaining available: %jdk\n",
623 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) *
625 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024));
627 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY,
628 vm_page_startup_finish, NULL);
632 * Scan comparison function for Red-Black tree scans. An inclusive
633 * (start,end) is expected. Other fields are not used.
636 rb_vm_page_scancmp(struct vm_page *p, void *data)
638 struct rb_vm_page_scan_info *info = data;
640 if (p->pindex < info->start_pindex)
642 if (p->pindex > info->end_pindex)
648 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
650 if (p1->pindex < p2->pindex)
652 if (p1->pindex > p2->pindex)
658 vm_page_init(vm_page_t m)
660 /* do nothing for now. Called from pmap_page_init() */
664 * Each page queue has its own spin lock, which is fairly optimal for
665 * allocating and freeing pages at least.
667 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
668 * queue spinlock via this function. Also note that m->queue cannot change
669 * unless both the page and queue are locked.
673 _vm_page_queue_spin_lock(vm_page_t m)
678 if (queue != PQ_NONE) {
679 spin_lock(&vm_page_queues[queue].spin);
680 KKASSERT(queue == m->queue);
686 _vm_page_queue_spin_unlock(vm_page_t m)
692 if (queue != PQ_NONE)
693 spin_unlock(&vm_page_queues[queue].spin);
698 _vm_page_queues_spin_lock(u_short queue)
701 if (queue != PQ_NONE)
702 spin_lock(&vm_page_queues[queue].spin);
708 _vm_page_queues_spin_unlock(u_short queue)
711 if (queue != PQ_NONE)
712 spin_unlock(&vm_page_queues[queue].spin);
716 vm_page_queue_spin_lock(vm_page_t m)
718 _vm_page_queue_spin_lock(m);
722 vm_page_queues_spin_lock(u_short queue)
724 _vm_page_queues_spin_lock(queue);
728 vm_page_queue_spin_unlock(vm_page_t m)
730 _vm_page_queue_spin_unlock(m);
734 vm_page_queues_spin_unlock(u_short queue)
736 _vm_page_queues_spin_unlock(queue);
740 * This locks the specified vm_page and its queue in the proper order
741 * (page first, then queue). The queue may change so the caller must
746 _vm_page_and_queue_spin_lock(vm_page_t m)
748 vm_page_spin_lock(m);
749 _vm_page_queue_spin_lock(m);
754 _vm_page_and_queue_spin_unlock(vm_page_t m)
756 _vm_page_queues_spin_unlock(m->queue);
757 vm_page_spin_unlock(m);
761 vm_page_and_queue_spin_unlock(vm_page_t m)
763 _vm_page_and_queue_spin_unlock(m);
767 vm_page_and_queue_spin_lock(vm_page_t m)
769 _vm_page_and_queue_spin_lock(m);
773 * Helper function removes vm_page from its current queue.
774 * Returns the base queue the page used to be on.
776 * The vm_page and the queue must be spinlocked.
777 * This function will unlock the queue but leave the page spinlocked.
779 static __inline u_short
780 _vm_page_rem_queue_spinlocked(vm_page_t m)
782 struct vpgqueues *pq;
787 if (queue != PQ_NONE) {
788 pq = &vm_page_queues[queue];
789 TAILQ_REMOVE(&pq->pl, m, pageq);
790 atomic_add_int(pq->cnt, -1);
795 vm_page_queues_spin_unlock(oqueue); /* intended */
801 * Helper function places the vm_page on the specified queue.
803 * The vm_page must be spinlocked.
804 * This function will return with both the page and the queue locked.
807 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
809 struct vpgqueues *pq;
811 KKASSERT(m->queue == PQ_NONE);
813 if (queue != PQ_NONE) {
814 vm_page_queues_spin_lock(queue);
815 pq = &vm_page_queues[queue];
817 atomic_add_int(pq->cnt, 1);
821 * PQ_FREE is always handled LIFO style to try to provide
822 * cache-hot pages to programs.
824 if (queue - m->pc == PQ_FREE) {
825 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
827 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
829 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
831 /* leave the queue spinlocked */
836 * Wait until page is no longer PG_BUSY or (if also_m_busy is TRUE)
837 * m->busy is zero. Returns TRUE if it had to sleep, FALSE if we
838 * did not. Only one sleep call will be made before returning.
840 * This function does NOT busy the page and on return the page is not
841 * guaranteed to be available.
844 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
852 if ((flags & PG_BUSY) == 0 &&
853 (also_m_busy == 0 || (flags & PG_SBUSY) == 0)) {
856 tsleep_interlock(m, 0);
857 if (atomic_cmpset_int(&m->flags, flags,
858 flags | PG_WANTED | PG_REFERENCED)) {
859 tsleep(m, PINTERLOCKED, msg, 0);
866 * This calculates and returns a page color given an optional VM object and
867 * either a pindex or an iterator. We attempt to return a cpu-localized
868 * pg_color that is still roughly 16-way set-associative. The CPU topology
869 * is used if it was probed.
871 * The caller may use the returned value to index into e.g. PQ_FREE when
872 * allocating a page in order to nominally obtain pages that are hopefully
873 * already localized to the requesting cpu. This function is not able to
874 * provide any sort of guarantee of this, but does its best to improve
875 * hardware cache management performance.
877 * WARNING! The caller must mask the returned value with PQ_L2_MASK.
880 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex)
887 phys_id = get_cpu_phys_id(cpuid);
888 core_id = get_cpu_core_id(cpuid);
889 object_pg_color = object ? object->pg_color : 0;
891 if (cpu_topology_phys_ids && cpu_topology_core_ids) {
892 int grpsize = PQ_L2_SIZE / cpu_topology_phys_ids;
894 if (grpsize / cpu_topology_core_ids >= PQ_SET_ASSOC) {
896 * Enough space for a full break-down.
898 pg_color = phys_id * grpsize;
899 pg_color += core_id * grpsize / cpu_topology_core_ids;
900 pg_color += (pindex + object_pg_color) %
901 (grpsize / cpu_topology_core_ids);
904 * Not enough space, split up by physical package,
905 * then split up by core id but only down to a
906 * 16-set. If all else fails, force a 16-set.
908 pg_color = phys_id * grpsize;
910 pg_color += 16 * (core_id % (grpsize / 16));
915 pg_color += (pindex + object_pg_color) %
920 * Unknown topology, distribute things evenly.
922 pg_color = cpuid * PQ_L2_SIZE / ncpus;
923 pg_color += pindex + object_pg_color;
929 * Wait until PG_BUSY can be set, then set it. If also_m_busy is TRUE we
930 * also wait for m->busy to become 0 before setting PG_BUSY.
933 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
934 int also_m_busy, const char *msg
942 if (flags & PG_BUSY) {
943 tsleep_interlock(m, 0);
944 if (atomic_cmpset_int(&m->flags, flags,
945 flags | PG_WANTED | PG_REFERENCED)) {
946 tsleep(m, PINTERLOCKED, msg, 0);
948 } else if (also_m_busy && (flags & PG_SBUSY)) {
949 tsleep_interlock(m, 0);
950 if (atomic_cmpset_int(&m->flags, flags,
951 flags | PG_WANTED | PG_REFERENCED)) {
952 tsleep(m, PINTERLOCKED, msg, 0);
955 if (atomic_cmpset_int(&m->flags, flags,
959 m->busy_line = lineno;
968 * Attempt to set PG_BUSY. If also_m_busy is TRUE we only succeed if m->busy
971 * Returns non-zero on failure.
974 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
984 if (also_m_busy && (flags & PG_SBUSY))
986 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
989 m->busy_line = lineno;
997 * Clear the PG_BUSY flag and return non-zero to indicate to the caller
998 * that a wakeup() should be performed.
1000 * The vm_page must be spinlocked and will remain spinlocked on return.
1001 * The related queue must NOT be spinlocked (which could deadlock us).
1007 _vm_page_wakeup(vm_page_t m)
1014 if (atomic_cmpset_int(&m->flags, flags,
1015 flags & ~(PG_BUSY | PG_WANTED))) {
1019 return(flags & PG_WANTED);
1023 * Clear the PG_BUSY flag and wakeup anyone waiting for the page. This
1024 * is typically the last call you make on a page before moving onto
1028 vm_page_wakeup(vm_page_t m)
1030 KASSERT(m->flags & PG_BUSY, ("vm_page_wakeup: page not busy!!!"));
1031 vm_page_spin_lock(m);
1032 if (_vm_page_wakeup(m)) {
1033 vm_page_spin_unlock(m);
1036 vm_page_spin_unlock(m);
1041 * Holding a page keeps it from being reused. Other parts of the system
1042 * can still disassociate the page from its current object and free it, or
1043 * perform read or write I/O on it and/or otherwise manipulate the page,
1044 * but if the page is held the VM system will leave the page and its data
1045 * intact and not reuse the page for other purposes until the last hold
1046 * reference is released. (see vm_page_wire() if you want to prevent the
1047 * page from being disassociated from its object too).
1049 * The caller must still validate the contents of the page and, if necessary,
1050 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
1051 * before manipulating the page.
1053 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
1056 vm_page_hold(vm_page_t m)
1058 vm_page_spin_lock(m);
1059 atomic_add_int(&m->hold_count, 1);
1060 if (m->queue - m->pc == PQ_FREE) {
1061 _vm_page_queue_spin_lock(m);
1062 _vm_page_rem_queue_spinlocked(m);
1063 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
1064 _vm_page_queue_spin_unlock(m);
1066 vm_page_spin_unlock(m);
1070 * The opposite of vm_page_hold(). If the page is on the HOLD queue
1071 * it was freed while held and must be moved back to the FREE queue.
1074 vm_page_unhold(vm_page_t m)
1076 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE,
1077 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)",
1078 m, m->hold_count, m->queue - m->pc));
1079 vm_page_spin_lock(m);
1080 atomic_add_int(&m->hold_count, -1);
1081 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
1082 _vm_page_queue_spin_lock(m);
1083 _vm_page_rem_queue_spinlocked(m);
1084 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
1085 _vm_page_queue_spin_unlock(m);
1087 vm_page_spin_unlock(m);
1093 * Create a fictitious page with the specified physical address and
1094 * memory attribute. The memory attribute is the only the machine-
1095 * dependent aspect of a fictitious page that must be initialized.
1099 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1102 if ((m->flags & PG_FICTITIOUS) != 0) {
1104 * The page's memattr might have changed since the
1105 * previous initialization. Update the pmap to the
1110 m->phys_addr = paddr;
1112 /* Fictitious pages don't use "segind". */
1113 /* Fictitious pages don't use "order" or "pool". */
1114 m->flags = PG_FICTITIOUS | PG_UNMANAGED | PG_BUSY;
1116 spin_init(&m->spin, "fake_page");
1119 pmap_page_set_memattr(m, memattr);
1123 * Inserts the given vm_page into the object and object list.
1125 * The pagetables are not updated but will presumably fault the page
1126 * in if necessary, or if a kernel page the caller will at some point
1127 * enter the page into the kernel's pmap. We are not allowed to block
1128 * here so we *can't* do this anyway.
1130 * This routine may not block.
1131 * This routine must be called with the vm_object held.
1132 * This routine must be called with a critical section held.
1134 * This routine returns TRUE if the page was inserted into the object
1135 * successfully, and FALSE if the page already exists in the object.
1138 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1140 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object));
1141 if (m->object != NULL)
1142 panic("vm_page_insert: already inserted");
1144 object->generation++;
1147 * Record the object/offset pair in this page and add the
1148 * pv_list_count of the page to the object.
1150 * The vm_page spin lock is required for interactions with the pmap.
1152 vm_page_spin_lock(m);
1155 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
1158 vm_page_spin_unlock(m);
1161 ++object->resident_page_count;
1162 ++mycpu->gd_vmtotal.t_rm;
1163 /* atomic_add_int(&object->agg_pv_list_count, m->md.pv_list_count); */
1164 vm_page_spin_unlock(m);
1167 * Since we are inserting a new and possibly dirty page,
1168 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
1170 if ((m->valid & m->dirty) ||
1171 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT)))
1172 vm_object_set_writeable_dirty(object);
1175 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
1177 swap_pager_page_inserted(m);
1182 * Removes the given vm_page_t from the (object,index) table
1184 * The underlying pmap entry (if any) is NOT removed here.
1185 * This routine may not block.
1187 * The page must be BUSY and will remain BUSY on return.
1188 * No other requirements.
1190 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
1194 vm_page_remove(vm_page_t m)
1198 if (m->object == NULL) {
1202 if ((m->flags & PG_BUSY) == 0)
1203 panic("vm_page_remove: page not busy");
1207 vm_object_hold(object);
1210 * Remove the page from the object and update the object.
1212 * The vm_page spin lock is required for interactions with the pmap.
1214 vm_page_spin_lock(m);
1215 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
1216 --object->resident_page_count;
1217 --mycpu->gd_vmtotal.t_rm;
1218 /* atomic_add_int(&object->agg_pv_list_count, -m->md.pv_list_count); */
1220 vm_page_spin_unlock(m);
1222 object->generation++;
1224 vm_object_drop(object);
1228 * Locate and return the page at (object, pindex), or NULL if the
1229 * page could not be found.
1231 * The caller must hold the vm_object token.
1234 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1239 * Search the hash table for this object/offset pair
1241 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1242 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1243 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1248 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1250 int also_m_busy, const char *msg
1256 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1257 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1259 KKASSERT(m->object == object && m->pindex == pindex);
1262 if (flags & PG_BUSY) {
1263 tsleep_interlock(m, 0);
1264 if (atomic_cmpset_int(&m->flags, flags,
1265 flags | PG_WANTED | PG_REFERENCED)) {
1266 tsleep(m, PINTERLOCKED, msg, 0);
1267 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1270 } else if (also_m_busy && (flags & PG_SBUSY)) {
1271 tsleep_interlock(m, 0);
1272 if (atomic_cmpset_int(&m->flags, flags,
1273 flags | PG_WANTED | PG_REFERENCED)) {
1274 tsleep(m, PINTERLOCKED, msg, 0);
1275 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1278 } else if (atomic_cmpset_int(&m->flags, flags,
1280 #ifdef VM_PAGE_DEBUG
1281 m->busy_func = func;
1282 m->busy_line = lineno;
1291 * Attempt to lookup and busy a page.
1293 * Returns NULL if the page could not be found
1295 * Returns a vm_page and error == TRUE if the page exists but could not
1298 * Returns a vm_page and error == FALSE on success.
1301 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1303 int also_m_busy, int *errorp
1309 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1310 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1313 KKASSERT(m->object == object && m->pindex == pindex);
1316 if (flags & PG_BUSY) {
1320 if (also_m_busy && (flags & PG_SBUSY)) {
1324 if (atomic_cmpset_int(&m->flags, flags, flags | PG_BUSY)) {
1325 #ifdef VM_PAGE_DEBUG
1326 m->busy_func = func;
1327 m->busy_line = lineno;
1336 * Attempt to repurpose the passed-in page. If the passed-in page cannot
1337 * be repurposed it will be released, *must_reenter will be set to 1, and
1338 * this function will fall-through to vm_page_lookup_busy_try().
1340 * The passed-in page must be wired and not busy. The returned page will
1341 * be busied and not wired.
1343 * A different page may be returned. The returned page will be busied and
1346 * NULL can be returned. If so, the required page could not be busied.
1347 * The passed-in page will be unwired.
1350 vm_page_repurpose(struct vm_object *object, vm_pindex_t pindex,
1351 int also_m_busy, int *errorp, vm_page_t m,
1352 int *must_reenter, int *iswired)
1356 * Do not mess with pages in a complex state, such as pages
1357 * which are mapped, as repurposing such pages can be more
1358 * expensive than simply allocatin a new one.
1360 * NOTE: Soft-busying can deadlock against putpages or I/O
1361 * so we only allow hard-busying here.
1363 KKASSERT(also_m_busy == FALSE);
1364 vm_page_busy_wait(m, also_m_busy, "biodep");
1366 if ((m->flags & (PG_UNMANAGED | PG_MAPPED |
1367 PG_FICTITIOUS | PG_SBUSY)) ||
1368 m->busy || m->wire_count != 1 || m->hold_count) {
1369 vm_page_unwire(m, 0);
1371 /* fall through to normal lookup */
1372 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
1373 vm_page_unwire(m, 0);
1374 vm_page_deactivate(m);
1376 /* fall through to normal lookup */
1379 * We can safely repurpose the page. It should
1380 * already be unqueued.
1382 KKASSERT(m->queue == PQ_NONE && m->dirty == 0);
1386 if (vm_page_insert(m, object, pindex)) {
1392 vm_page_unwire(m, 0);
1394 /* fall through to normal lookup */
1399 * Cannot repurpose page, attempt to locate the desired page. May
1404 m = vm_page_lookup_busy_try(object, pindex, also_m_busy, errorp);
1410 * Caller must hold the related vm_object
1413 vm_page_next(vm_page_t m)
1417 next = vm_page_rb_tree_RB_NEXT(m);
1418 if (next && next->pindex != m->pindex + 1)
1426 * Move the given vm_page from its current object to the specified
1427 * target object/offset. The page must be busy and will remain so
1430 * new_object must be held.
1431 * This routine might block. XXX ?
1433 * NOTE: Swap associated with the page must be invalidated by the move. We
1434 * have to do this for several reasons: (1) we aren't freeing the
1435 * page, (2) we are dirtying the page, (3) the VM system is probably
1436 * moving the page from object A to B, and will then later move
1437 * the backing store from A to B and we can't have a conflict.
1439 * NOTE: We *always* dirty the page. It is necessary both for the
1440 * fact that we moved it, and because we may be invalidating
1441 * swap. If the page is on the cache, we have to deactivate it
1442 * or vm_page_dirty() will panic. Dirty pages are not allowed
1446 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1448 KKASSERT(m->flags & PG_BUSY);
1449 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object));
1451 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object));
1454 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1455 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1456 new_object, new_pindex);
1458 if (m->queue - m->pc == PQ_CACHE)
1459 vm_page_deactivate(m);
1464 * vm_page_unqueue() without any wakeup. This routine is used when a page
1465 * is to remain BUSYied by the caller.
1467 * This routine may not block.
1470 vm_page_unqueue_nowakeup(vm_page_t m)
1472 vm_page_and_queue_spin_lock(m);
1473 (void)_vm_page_rem_queue_spinlocked(m);
1474 vm_page_spin_unlock(m);
1478 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1481 * This routine may not block.
1484 vm_page_unqueue(vm_page_t m)
1488 vm_page_and_queue_spin_lock(m);
1489 queue = _vm_page_rem_queue_spinlocked(m);
1490 if (queue == PQ_FREE || queue == PQ_CACHE) {
1491 vm_page_spin_unlock(m);
1492 pagedaemon_wakeup();
1494 vm_page_spin_unlock(m);
1499 * vm_page_list_find()
1501 * Find a page on the specified queue with color optimization.
1503 * The page coloring optimization attempts to locate a page that does
1504 * not overload other nearby pages in the object in the cpu's L1 or L2
1505 * caches. We need this optimization because cpu caches tend to be
1506 * physical caches, while object spaces tend to be virtual.
1508 * The page coloring optimization also, very importantly, tries to localize
1509 * memory to cpus and physical sockets.
1511 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1512 * and the algorithm is adjusted to localize allocations on a per-core basis.
1513 * This is done by 'twisting' the colors.
1515 * The page is returned spinlocked and removed from its queue (it will
1516 * be on PQ_NONE), or NULL. The page is not PG_BUSY'd. The caller
1517 * is responsible for dealing with the busy-page case (usually by
1518 * deactivating the page and looping).
1520 * NOTE: This routine is carefully inlined. A non-inlined version
1521 * is available for outside callers but the only critical path is
1522 * from within this source file.
1524 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1525 * represent stable storage, allowing us to order our locks vm_page
1526 * first, then queue.
1530 _vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1536 m = TAILQ_LAST(&vm_page_queues[basequeue+index].pl,
1539 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1542 m = _vm_page_list_find2(basequeue, index);
1545 vm_page_and_queue_spin_lock(m);
1546 if (m->queue == basequeue + index) {
1547 _vm_page_rem_queue_spinlocked(m);
1548 /* vm_page_t spin held, no queue spin */
1551 vm_page_and_queue_spin_unlock(m);
1557 * If we could not find the page in the desired queue try to find it in
1561 _vm_page_list_find2(int basequeue, int index)
1563 struct vpgqueues *pq;
1565 int pqmask = PQ_SET_ASSOC_MASK >> 1;
1569 index &= PQ_L2_MASK;
1570 pq = &vm_page_queues[basequeue];
1573 * Run local sets of 16, 32, 64, 128, and the whole queue if all
1574 * else fails (PQ_L2_MASK which is 255).
1577 pqmask = (pqmask << 1) | 1;
1578 for (i = 0; i <= pqmask; ++i) {
1579 pqi = (index & ~pqmask) | ((index + i) & pqmask);
1580 m = TAILQ_FIRST(&pq[pqi].pl);
1582 _vm_page_and_queue_spin_lock(m);
1583 if (m->queue == basequeue + pqi) {
1584 _vm_page_rem_queue_spinlocked(m);
1587 _vm_page_and_queue_spin_unlock(m);
1592 } while (pqmask != PQ_L2_MASK);
1598 * Returns a vm_page candidate for allocation. The page is not busied so
1599 * it can move around. The caller must busy the page (and typically
1600 * deactivate it if it cannot be busied!)
1602 * Returns a spinlocked vm_page that has been removed from its queue.
1605 vm_page_list_find(int basequeue, int index, boolean_t prefer_zero)
1607 return(_vm_page_list_find(basequeue, index, prefer_zero));
1611 * Find a page on the cache queue with color optimization, remove it
1612 * from the queue, and busy it. The returned page will not be spinlocked.
1614 * A candidate failure will be deactivated. Candidates can fail due to
1615 * being busied by someone else, in which case they will be deactivated.
1617 * This routine may not block.
1621 vm_page_select_cache(u_short pg_color)
1626 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK, FALSE);
1630 * (m) has been removed from its queue and spinlocked
1632 if (vm_page_busy_try(m, TRUE)) {
1633 _vm_page_deactivate_locked(m, 0);
1634 vm_page_spin_unlock(m);
1637 * We successfully busied the page
1639 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 &&
1640 m->hold_count == 0 &&
1641 m->wire_count == 0 &&
1642 (m->dirty & m->valid) == 0) {
1643 vm_page_spin_unlock(m);
1644 pagedaemon_wakeup();
1649 * The page cannot be recycled, deactivate it.
1651 _vm_page_deactivate_locked(m, 0);
1652 if (_vm_page_wakeup(m)) {
1653 vm_page_spin_unlock(m);
1656 vm_page_spin_unlock(m);
1664 * Find a free or zero page, with specified preference. We attempt to
1665 * inline the nominal case and fall back to _vm_page_select_free()
1666 * otherwise. A busied page is removed from the queue and returned.
1668 * This routine may not block.
1670 static __inline vm_page_t
1671 vm_page_select_free(u_short pg_color, boolean_t prefer_zero)
1676 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK,
1680 if (vm_page_busy_try(m, TRUE)) {
1682 * Various mechanisms such as a pmap_collect can
1683 * result in a busy page on the free queue. We
1684 * have to move the page out of the way so we can
1685 * retry the allocation. If the other thread is not
1686 * allocating the page then m->valid will remain 0 and
1687 * the pageout daemon will free the page later on.
1689 * Since we could not busy the page, however, we
1690 * cannot make assumptions as to whether the page
1691 * will be allocated by the other thread or not,
1692 * so all we can do is deactivate it to move it out
1693 * of the way. In particular, if the other thread
1694 * wires the page it may wind up on the inactive
1695 * queue and the pageout daemon will have to deal
1696 * with that case too.
1698 _vm_page_deactivate_locked(m, 0);
1699 vm_page_spin_unlock(m);
1702 * Theoretically if we are able to busy the page
1703 * atomic with the queue removal (using the vm_page
1704 * lock) nobody else should be able to mess with the
1707 KKASSERT((m->flags & (PG_UNMANAGED |
1708 PG_NEED_COMMIT)) == 0);
1709 KASSERT(m->hold_count == 0, ("m->hold_count is not zero "
1710 "pg %p q=%d flags=%08x hold=%d wire=%d",
1711 m, m->queue, m->flags, m->hold_count, m->wire_count));
1712 KKASSERT(m->wire_count == 0);
1713 vm_page_spin_unlock(m);
1714 pagedaemon_wakeup();
1716 /* return busied and removed page */
1726 * Allocate and return a memory cell associated with this VM object/offset
1727 * pair. If object is NULL an unassociated page will be allocated.
1729 * The returned page will be busied and removed from its queues. This
1730 * routine can block and may return NULL if a race occurs and the page
1731 * is found to already exist at the specified (object, pindex).
1733 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1734 * VM_ALLOC_QUICK like normal but cannot use cache
1735 * VM_ALLOC_SYSTEM greater free drain
1736 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1737 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1738 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1739 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1740 * (see vm_page_grab())
1741 * VM_ALLOC_USE_GD ok to use per-gd cache
1743 * VM_ALLOC_CPU(n) allocate using specified cpu localization
1745 * The object must be held if not NULL
1746 * This routine may not block
1748 * Additional special handling is required when called from an interrupt
1749 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1753 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1762 * Special per-cpu free VM page cache. The pages are pre-busied
1763 * and pre-zerod for us.
1765 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1767 if (gd->gd_vmpg_count) {
1768 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1780 * CPU localization algorithm. Break the page queues up by physical
1781 * id and core id (note that two cpu threads will have the same core
1782 * id, and core_id != gd_cpuid).
1784 * This is nowhere near perfect, for example the last pindex in a
1785 * subgroup will overflow into the next cpu or package. But this
1786 * should get us good page reuse locality in heavy mixed loads.
1788 * (may be executed before the APs are started, so other GDs might
1791 if (page_req & VM_ALLOC_CPU_SPEC)
1792 cpuid_local = VM_ALLOC_GETCPU(page_req);
1794 cpuid_local = mycpu->gd_cpuid;
1796 pg_color = vm_get_pg_color(cpuid_local, object, pindex);
1799 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1800 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1803 * Certain system threads (pageout daemon, buf_daemon's) are
1804 * allowed to eat deeper into the free page list.
1806 if (curthread->td_flags & TDF_SYSTHREAD)
1807 page_req |= VM_ALLOC_SYSTEM;
1810 * Impose various limitations. Note that the v_free_reserved test
1811 * must match the opposite of vm_page_count_target() to avoid
1812 * livelocks, be careful.
1815 if (vmstats.v_free_count >= vmstats.v_free_reserved ||
1816 ((page_req & VM_ALLOC_INTERRUPT) && vmstats.v_free_count > 0) ||
1817 ((page_req & VM_ALLOC_SYSTEM) && vmstats.v_cache_count == 0 &&
1818 vmstats.v_free_count > vmstats.v_interrupt_free_min)
1821 * The free queue has sufficient free pages to take one out.
1823 if (page_req & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO))
1824 m = vm_page_select_free(pg_color, TRUE);
1826 m = vm_page_select_free(pg_color, FALSE);
1827 } else if (page_req & VM_ALLOC_NORMAL) {
1829 * Allocatable from the cache (non-interrupt only). On
1830 * success, we must free the page and try again, thus
1831 * ensuring that vmstats.v_*_free_min counters are replenished.
1834 if (curthread->td_preempted) {
1835 kprintf("vm_page_alloc(): warning, attempt to allocate"
1836 " cache page from preempting interrupt\n");
1839 m = vm_page_select_cache(pg_color);
1842 m = vm_page_select_cache(pg_color);
1845 * On success move the page into the free queue and loop.
1847 * Only do this if we can safely acquire the vm_object lock,
1848 * because this is effectively a random page and the caller
1849 * might be holding the lock shared, we don't want to
1853 KASSERT(m->dirty == 0,
1854 ("Found dirty cache page %p", m));
1855 if ((obj = m->object) != NULL) {
1856 if (vm_object_hold_try(obj)) {
1857 vm_page_protect(m, VM_PROT_NONE);
1859 /* m->object NULL here */
1860 vm_object_drop(obj);
1862 vm_page_deactivate(m);
1866 vm_page_protect(m, VM_PROT_NONE);
1873 * On failure return NULL
1875 #if defined(DIAGNOSTIC)
1876 if (vmstats.v_cache_count > 0)
1877 kprintf("vm_page_alloc(NORMAL): missing pages on cache queue: %d\n", vmstats.v_cache_count);
1879 atomic_add_int(&vm_pageout_deficit, 1);
1880 pagedaemon_wakeup();
1884 * No pages available, wakeup the pageout daemon and give up.
1886 atomic_add_int(&vm_pageout_deficit, 1);
1887 pagedaemon_wakeup();
1892 * v_free_count can race so loop if we don't find the expected
1899 * Good page found. The page has already been busied for us and
1900 * removed from its queues.
1902 KASSERT(m->dirty == 0,
1903 ("vm_page_alloc: free/cache page %p was dirty", m));
1904 KKASSERT(m->queue == PQ_NONE);
1910 * Initialize the structure, inheriting some flags but clearing
1911 * all the rest. The page has already been busied for us.
1913 vm_page_flag_clear(m, ~(PG_BUSY | PG_SBUSY));
1914 KKASSERT(m->wire_count == 0);
1915 KKASSERT(m->busy == 0);
1920 * Caller must be holding the object lock (asserted by
1921 * vm_page_insert()).
1923 * NOTE: Inserting a page here does not insert it into any pmaps
1924 * (which could cause us to block allocating memory).
1926 * NOTE: If no object an unassociated page is allocated, m->pindex
1927 * can be used by the caller for any purpose.
1930 if (vm_page_insert(m, object, pindex) == FALSE) {
1932 if ((page_req & VM_ALLOC_NULL_OK) == 0)
1933 panic("PAGE RACE %p[%ld]/%p",
1934 object, (long)pindex, m);
1942 * Don't wakeup too often - wakeup the pageout daemon when
1943 * we would be nearly out of memory.
1945 pagedaemon_wakeup();
1948 * A PG_BUSY page is returned.
1954 * Returns number of pages available in our DMA memory reserve
1955 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf)
1958 vm_contig_avail_pages(void)
1963 spin_lock(&vm_contig_spin);
1964 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
1965 spin_unlock(&vm_contig_spin);
1971 * Attempt to allocate contiguous physical memory with the specified
1975 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
1976 unsigned long alignment, unsigned long boundary,
1977 unsigned long size, vm_memattr_t memattr)
1983 alignment >>= PAGE_SHIFT;
1986 boundary >>= PAGE_SHIFT;
1989 size = (size + PAGE_MASK) >> PAGE_SHIFT;
1991 spin_lock(&vm_contig_spin);
1992 blk = alist_alloc(&vm_contig_alist, 0, size);
1993 if (blk == ALIST_BLOCK_NONE) {
1994 spin_unlock(&vm_contig_spin);
1996 kprintf("vm_page_alloc_contig: %ldk nospace\n",
1997 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
2001 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
2002 alist_free(&vm_contig_alist, blk, size);
2003 spin_unlock(&vm_contig_spin);
2005 kprintf("vm_page_alloc_contig: %ldk high "
2007 (size + PAGE_MASK) * (PAGE_SIZE / 1024),
2012 spin_unlock(&vm_contig_spin);
2013 if (vm_contig_verbose) {
2014 kprintf("vm_page_alloc_contig: %016jx/%ldk\n",
2015 (intmax_t)(vm_paddr_t)blk << PAGE_SHIFT,
2016 (size + PAGE_MASK) * (PAGE_SIZE / 1024));
2019 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
2020 if (memattr != VM_MEMATTR_DEFAULT)
2021 for (i = 0;i < size;i++)
2022 pmap_page_set_memattr(&m[i], memattr);
2027 * Free contiguously allocated pages. The pages will be wired but not busy.
2028 * When freeing to the alist we leave them wired and not busy.
2031 vm_page_free_contig(vm_page_t m, unsigned long size)
2033 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
2034 vm_pindex_t start = pa >> PAGE_SHIFT;
2035 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
2037 if (vm_contig_verbose) {
2038 kprintf("vm_page_free_contig: %016jx/%ldk\n",
2039 (intmax_t)pa, size / 1024);
2041 if (pa < vm_low_phys_reserved) {
2042 KKASSERT(pa + size <= vm_low_phys_reserved);
2043 spin_lock(&vm_contig_spin);
2044 alist_free(&vm_contig_alist, start, pages);
2045 spin_unlock(&vm_contig_spin);
2048 vm_page_busy_wait(m, FALSE, "cpgfr");
2049 vm_page_unwire(m, 0);
2060 * Wait for sufficient free memory for nominal heavy memory use kernel
2063 * WARNING! Be sure never to call this in any vm_pageout code path, which
2064 * will trivially deadlock the system.
2067 vm_wait_nominal(void)
2069 while (vm_page_count_min(0))
2074 * Test if vm_wait_nominal() would block.
2077 vm_test_nominal(void)
2079 if (vm_page_count_min(0))
2085 * Block until free pages are available for allocation, called in various
2086 * places before memory allocations.
2088 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
2089 * more generous then that.
2095 * never wait forever
2099 lwkt_gettoken(&vm_token);
2101 if (curthread == pagethread) {
2103 * The pageout daemon itself needs pages, this is bad.
2105 if (vm_page_count_min(0)) {
2106 vm_pageout_pages_needed = 1;
2107 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
2111 * Wakeup the pageout daemon if necessary and wait.
2113 * Do not wait indefinitely for the target to be reached,
2114 * as load might prevent it from being reached any time soon.
2115 * But wait a little to try to slow down page allocations
2116 * and to give more important threads (the pagedaemon)
2117 * allocation priority.
2119 if (vm_page_count_target()) {
2120 if (vm_pages_needed == 0) {
2121 vm_pages_needed = 1;
2122 wakeup(&vm_pages_needed);
2124 ++vm_pages_waiting; /* SMP race ok */
2125 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
2128 lwkt_reltoken(&vm_token);
2132 * Block until free pages are available for allocation
2134 * Called only from vm_fault so that processes page faulting can be
2138 vm_wait_pfault(void)
2141 * Wakeup the pageout daemon if necessary and wait.
2143 * Do not wait indefinitely for the target to be reached,
2144 * as load might prevent it from being reached any time soon.
2145 * But wait a little to try to slow down page allocations
2146 * and to give more important threads (the pagedaemon)
2147 * allocation priority.
2149 if (vm_page_count_min(0)) {
2150 lwkt_gettoken(&vm_token);
2151 while (vm_page_count_severe()) {
2152 if (vm_page_count_target()) {
2155 if (vm_pages_needed == 0) {
2156 vm_pages_needed = 1;
2157 wakeup(&vm_pages_needed);
2159 ++vm_pages_waiting; /* SMP race ok */
2160 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
2163 * Do not stay stuck in the loop if the system is trying
2164 * to kill the process.
2167 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
2171 lwkt_reltoken(&vm_token);
2176 * Put the specified page on the active list (if appropriate). Ensure
2177 * that act_count is at least ACT_INIT but do not otherwise mess with it.
2179 * The caller should be holding the page busied ? XXX
2180 * This routine may not block.
2183 vm_page_activate(vm_page_t m)
2187 vm_page_spin_lock(m);
2188 if (m->queue - m->pc != PQ_ACTIVE) {
2189 _vm_page_queue_spin_lock(m);
2190 oqueue = _vm_page_rem_queue_spinlocked(m);
2191 /* page is left spinlocked, queue is unlocked */
2193 if (oqueue == PQ_CACHE)
2194 mycpu->gd_cnt.v_reactivated++;
2195 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2196 if (m->act_count < ACT_INIT)
2197 m->act_count = ACT_INIT;
2198 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
2200 _vm_page_and_queue_spin_unlock(m);
2201 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
2202 pagedaemon_wakeup();
2204 if (m->act_count < ACT_INIT)
2205 m->act_count = ACT_INIT;
2206 vm_page_spin_unlock(m);
2211 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2212 * routine is called when a page has been added to the cache or free
2215 * This routine may not block.
2217 static __inline void
2218 vm_page_free_wakeup(void)
2221 * If the pageout daemon itself needs pages, then tell it that
2222 * there are some free.
2224 if (vm_pageout_pages_needed &&
2225 vmstats.v_cache_count + vmstats.v_free_count >=
2226 vmstats.v_pageout_free_min
2228 vm_pageout_pages_needed = 0;
2229 wakeup(&vm_pageout_pages_needed);
2233 * Wakeup processes that are waiting on memory.
2235 * Generally speaking we want to wakeup stuck processes as soon as
2236 * possible. !vm_page_count_min(0) is the absolute minimum point
2237 * where we can do this. Wait a bit longer to reduce degenerate
2238 * re-blocking (vm_page_free_hysteresis). The target check is just
2239 * to make sure the min-check w/hysteresis does not exceed the
2242 if (vm_pages_waiting) {
2243 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2244 !vm_page_count_target()) {
2245 vm_pages_waiting = 0;
2246 wakeup(&vmstats.v_free_count);
2247 ++mycpu->gd_cnt.v_ppwakeups;
2250 if (!vm_page_count_target()) {
2252 * Plenty of pages are free, wakeup everyone.
2254 vm_pages_waiting = 0;
2255 wakeup(&vmstats.v_free_count);
2256 ++mycpu->gd_cnt.v_ppwakeups;
2257 } else if (!vm_page_count_min(0)) {
2259 * Some pages are free, wakeup someone.
2261 int wcount = vm_pages_waiting;
2264 vm_pages_waiting = wcount;
2265 wakeup_one(&vmstats.v_free_count);
2266 ++mycpu->gd_cnt.v_ppwakeups;
2273 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
2274 * it from its VM object.
2276 * The vm_page must be PG_BUSY on entry. PG_BUSY will be released on
2277 * return (the page will have been freed).
2280 vm_page_free_toq(vm_page_t m)
2282 mycpu->gd_cnt.v_tfree++;
2283 KKASSERT((m->flags & PG_MAPPED) == 0);
2284 KKASSERT(m->flags & PG_BUSY);
2286 if (m->busy || ((m->queue - m->pc) == PQ_FREE)) {
2287 kprintf("vm_page_free: pindex(%lu), busy(%d), "
2288 "PG_BUSY(%d), hold(%d)\n",
2289 (u_long)m->pindex, m->busy,
2290 ((m->flags & PG_BUSY) ? 1 : 0), m->hold_count);
2291 if ((m->queue - m->pc) == PQ_FREE)
2292 panic("vm_page_free: freeing free page");
2294 panic("vm_page_free: freeing busy page");
2298 * Remove from object, spinlock the page and its queues and
2299 * remove from any queue. No queue spinlock will be held
2300 * after this section (because the page was removed from any
2304 vm_page_and_queue_spin_lock(m);
2305 _vm_page_rem_queue_spinlocked(m);
2308 * No further management of fictitious pages occurs beyond object
2309 * and queue removal.
2311 if ((m->flags & PG_FICTITIOUS) != 0) {
2312 vm_page_spin_unlock(m);
2320 if (m->wire_count != 0) {
2321 if (m->wire_count > 1) {
2323 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2324 m->wire_count, (long)m->pindex);
2326 panic("vm_page_free: freeing wired page");
2330 * Clear the UNMANAGED flag when freeing an unmanaged page.
2331 * Clear the NEED_COMMIT flag
2333 if (m->flags & PG_UNMANAGED)
2334 vm_page_flag_clear(m, PG_UNMANAGED);
2335 if (m->flags & PG_NEED_COMMIT)
2336 vm_page_flag_clear(m, PG_NEED_COMMIT);
2338 if (m->hold_count != 0) {
2339 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2341 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 0);
2345 * This sequence allows us to clear PG_BUSY while still holding
2346 * its spin lock, which reduces contention vs allocators. We
2347 * must not leave the queue locked or _vm_page_wakeup() may
2350 _vm_page_queue_spin_unlock(m);
2351 if (_vm_page_wakeup(m)) {
2352 vm_page_spin_unlock(m);
2355 vm_page_spin_unlock(m);
2357 vm_page_free_wakeup();
2361 * vm_page_unmanage()
2363 * Prevent PV management from being done on the page. The page is
2364 * removed from the paging queues as if it were wired, and as a
2365 * consequence of no longer being managed the pageout daemon will not
2366 * touch it (since there is no way to locate the pte mappings for the
2367 * page). madvise() calls that mess with the pmap will also no longer
2368 * operate on the page.
2370 * Beyond that the page is still reasonably 'normal'. Freeing the page
2371 * will clear the flag.
2373 * This routine is used by OBJT_PHYS objects - objects using unswappable
2374 * physical memory as backing store rather then swap-backed memory and
2375 * will eventually be extended to support 4MB unmanaged physical
2378 * Caller must be holding the page busy.
2381 vm_page_unmanage(vm_page_t m)
2383 KKASSERT(m->flags & PG_BUSY);
2384 if ((m->flags & PG_UNMANAGED) == 0) {
2385 if (m->wire_count == 0)
2388 vm_page_flag_set(m, PG_UNMANAGED);
2392 * Mark this page as wired down by yet another map, removing it from
2393 * paging queues as necessary.
2395 * Caller must be holding the page busy.
2398 vm_page_wire(vm_page_t m)
2401 * Only bump the wire statistics if the page is not already wired,
2402 * and only unqueue the page if it is on some queue (if it is unmanaged
2403 * it is already off the queues). Don't do anything with fictitious
2404 * pages because they are always wired.
2406 KKASSERT(m->flags & PG_BUSY);
2407 if ((m->flags & PG_FICTITIOUS) == 0) {
2408 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2409 if ((m->flags & PG_UNMANAGED) == 0)
2411 atomic_add_int(&vmstats.v_wire_count, 1);
2413 KASSERT(m->wire_count != 0,
2414 ("vm_page_wire: wire_count overflow m=%p", m));
2419 * Release one wiring of this page, potentially enabling it to be paged again.
2421 * Many pages placed on the inactive queue should actually go
2422 * into the cache, but it is difficult to figure out which. What
2423 * we do instead, if the inactive target is well met, is to put
2424 * clean pages at the head of the inactive queue instead of the tail.
2425 * This will cause them to be moved to the cache more quickly and
2426 * if not actively re-referenced, freed more quickly. If we just
2427 * stick these pages at the end of the inactive queue, heavy filesystem
2428 * meta-data accesses can cause an unnecessary paging load on memory bound
2429 * processes. This optimization causes one-time-use metadata to be
2430 * reused more quickly.
2432 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2433 * the inactive queue. This helps the pageout daemon determine memory
2434 * pressure and act on out-of-memory situations more quickly.
2436 * BUT, if we are in a low-memory situation we have no choice but to
2437 * put clean pages on the cache queue.
2439 * A number of routines use vm_page_unwire() to guarantee that the page
2440 * will go into either the inactive or active queues, and will NEVER
2441 * be placed in the cache - for example, just after dirtying a page.
2442 * dirty pages in the cache are not allowed.
2444 * This routine may not block.
2447 vm_page_unwire(vm_page_t m, int activate)
2449 KKASSERT(m->flags & PG_BUSY);
2450 if (m->flags & PG_FICTITIOUS) {
2452 } else if (m->wire_count <= 0) {
2453 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2455 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2456 atomic_add_int(&vmstats.v_wire_count, -1);
2457 if (m->flags & PG_UNMANAGED) {
2459 } else if (activate || (m->flags & PG_NEED_COMMIT)) {
2460 vm_page_spin_lock(m);
2461 _vm_page_add_queue_spinlocked(m,
2462 PQ_ACTIVE + m->pc, 0);
2463 _vm_page_and_queue_spin_unlock(m);
2465 vm_page_spin_lock(m);
2466 vm_page_flag_clear(m, PG_WINATCFLS);
2467 _vm_page_add_queue_spinlocked(m,
2468 PQ_INACTIVE + m->pc, 0);
2469 ++vm_swapcache_inactive_heuristic;
2470 _vm_page_and_queue_spin_unlock(m);
2477 * Move the specified page to the inactive queue. If the page has
2478 * any associated swap, the swap is deallocated.
2480 * Normally athead is 0 resulting in LRU operation. athead is set
2481 * to 1 if we want this page to be 'as if it were placed in the cache',
2482 * except without unmapping it from the process address space.
2484 * vm_page's spinlock must be held on entry and will remain held on return.
2485 * This routine may not block.
2488 _vm_page_deactivate_locked(vm_page_t m, int athead)
2493 * Ignore if already inactive.
2495 if (m->queue - m->pc == PQ_INACTIVE)
2497 _vm_page_queue_spin_lock(m);
2498 oqueue = _vm_page_rem_queue_spinlocked(m);
2500 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2501 if (oqueue == PQ_CACHE)
2502 mycpu->gd_cnt.v_reactivated++;
2503 vm_page_flag_clear(m, PG_WINATCFLS);
2504 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2506 ++vm_swapcache_inactive_heuristic;
2508 /* NOTE: PQ_NONE if condition not taken */
2509 _vm_page_queue_spin_unlock(m);
2510 /* leaves vm_page spinlocked */
2514 * Attempt to deactivate a page.
2519 vm_page_deactivate(vm_page_t m)
2521 vm_page_spin_lock(m);
2522 _vm_page_deactivate_locked(m, 0);
2523 vm_page_spin_unlock(m);
2527 vm_page_deactivate_locked(vm_page_t m)
2529 _vm_page_deactivate_locked(m, 0);
2533 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it.
2535 * This function returns non-zero if it successfully moved the page to
2538 * This function unconditionally unbusies the page on return.
2541 vm_page_try_to_cache(vm_page_t m)
2543 vm_page_spin_lock(m);
2544 if (m->dirty || m->hold_count || m->wire_count ||
2545 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2546 if (_vm_page_wakeup(m)) {
2547 vm_page_spin_unlock(m);
2550 vm_page_spin_unlock(m);
2554 vm_page_spin_unlock(m);
2557 * Page busied by us and no longer spinlocked. Dirty pages cannot
2558 * be moved to the cache.
2560 vm_page_test_dirty(m);
2561 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2570 * Attempt to free the page. If we cannot free it, we do nothing.
2571 * 1 is returned on success, 0 on failure.
2576 vm_page_try_to_free(vm_page_t m)
2578 vm_page_spin_lock(m);
2579 if (vm_page_busy_try(m, TRUE)) {
2580 vm_page_spin_unlock(m);
2585 * The page can be in any state, including already being on the free
2586 * queue. Check to see if it really can be freed.
2588 if (m->dirty || /* can't free if it is dirty */
2589 m->hold_count || /* or held (XXX may be wrong) */
2590 m->wire_count || /* or wired */
2591 (m->flags & (PG_UNMANAGED | /* or unmanaged */
2592 PG_NEED_COMMIT)) || /* or needs a commit */
2593 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */
2594 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */
2595 if (_vm_page_wakeup(m)) {
2596 vm_page_spin_unlock(m);
2599 vm_page_spin_unlock(m);
2603 vm_page_spin_unlock(m);
2606 * We can probably free the page.
2608 * Page busied by us and no longer spinlocked. Dirty pages will
2609 * not be freed by this function. We have to re-test the
2610 * dirty bit after cleaning out the pmaps.
2612 vm_page_test_dirty(m);
2613 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2617 vm_page_protect(m, VM_PROT_NONE);
2618 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2629 * Put the specified page onto the page cache queue (if appropriate).
2631 * The page must be busy, and this routine will release the busy and
2632 * possibly even free the page.
2635 vm_page_cache(vm_page_t m)
2638 * Not suitable for the cache
2640 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2641 m->busy || m->wire_count || m->hold_count) {
2647 * Already in the cache (and thus not mapped)
2649 if ((m->queue - m->pc) == PQ_CACHE) {
2650 KKASSERT((m->flags & PG_MAPPED) == 0);
2656 * Caller is required to test m->dirty, but note that the act of
2657 * removing the page from its maps can cause it to become dirty
2658 * on an SMP system due to another cpu running in usermode.
2661 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2666 * Remove all pmaps and indicate that the page is not
2667 * writeable or mapped. Our vm_page_protect() call may
2668 * have blocked (especially w/ VM_PROT_NONE), so recheck
2671 vm_page_protect(m, VM_PROT_NONE);
2672 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2673 m->busy || m->wire_count || m->hold_count) {
2675 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2676 vm_page_deactivate(m);
2679 _vm_page_and_queue_spin_lock(m);
2680 _vm_page_rem_queue_spinlocked(m);
2681 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2682 _vm_page_queue_spin_unlock(m);
2683 if (_vm_page_wakeup(m)) {
2684 vm_page_spin_unlock(m);
2687 vm_page_spin_unlock(m);
2689 vm_page_free_wakeup();
2694 * vm_page_dontneed()
2696 * Cache, deactivate, or do nothing as appropriate. This routine
2697 * is typically used by madvise() MADV_DONTNEED.
2699 * Generally speaking we want to move the page into the cache so
2700 * it gets reused quickly. However, this can result in a silly syndrome
2701 * due to the page recycling too quickly. Small objects will not be
2702 * fully cached. On the otherhand, if we move the page to the inactive
2703 * queue we wind up with a problem whereby very large objects
2704 * unnecessarily blow away our inactive and cache queues.
2706 * The solution is to move the pages based on a fixed weighting. We
2707 * either leave them alone, deactivate them, or move them to the cache,
2708 * where moving them to the cache has the highest weighting.
2709 * By forcing some pages into other queues we eventually force the
2710 * system to balance the queues, potentially recovering other unrelated
2711 * space from active. The idea is to not force this to happen too
2714 * The page must be busied.
2717 vm_page_dontneed(vm_page_t m)
2719 static int dnweight;
2726 * occassionally leave the page alone
2728 if ((dnw & 0x01F0) == 0 ||
2729 m->queue - m->pc == PQ_INACTIVE ||
2730 m->queue - m->pc == PQ_CACHE
2732 if (m->act_count >= ACT_INIT)
2738 * If vm_page_dontneed() is inactivating a page, it must clear
2739 * the referenced flag; otherwise the pagedaemon will see references
2740 * on the page in the inactive queue and reactivate it. Until the
2741 * page can move to the cache queue, madvise's job is not done.
2743 vm_page_flag_clear(m, PG_REFERENCED);
2744 pmap_clear_reference(m);
2747 vm_page_test_dirty(m);
2749 if (m->dirty || (dnw & 0x0070) == 0) {
2751 * Deactivate the page 3 times out of 32.
2756 * Cache the page 28 times out of every 32. Note that
2757 * the page is deactivated instead of cached, but placed
2758 * at the head of the queue instead of the tail.
2762 vm_page_spin_lock(m);
2763 _vm_page_deactivate_locked(m, head);
2764 vm_page_spin_unlock(m);
2768 * These routines manipulate the 'soft busy' count for a page. A soft busy
2769 * is almost like PG_BUSY except that it allows certain compatible operations
2770 * to occur on the page while it is busy. For example, a page undergoing a
2771 * write can still be mapped read-only.
2773 * Because vm_pages can overlap buffers m->busy can be > 1. m->busy is only
2774 * adjusted while the vm_page is PG_BUSY so the flash will occur when the
2775 * busy bit is cleared.
2778 vm_page_io_start(vm_page_t m)
2780 KASSERT(m->flags & PG_BUSY, ("vm_page_io_start: page not busy!!!"));
2781 atomic_add_char(&m->busy, 1);
2782 vm_page_flag_set(m, PG_SBUSY);
2786 vm_page_io_finish(vm_page_t m)
2788 KASSERT(m->flags & PG_BUSY, ("vm_page_io_finish: page not busy!!!"));
2789 atomic_subtract_char(&m->busy, 1);
2791 vm_page_flag_clear(m, PG_SBUSY);
2795 * Indicate that a clean VM page requires a filesystem commit and cannot
2796 * be reused. Used by tmpfs.
2799 vm_page_need_commit(vm_page_t m)
2801 vm_page_flag_set(m, PG_NEED_COMMIT);
2802 vm_object_set_writeable_dirty(m->object);
2806 vm_page_clear_commit(vm_page_t m)
2808 vm_page_flag_clear(m, PG_NEED_COMMIT);
2812 * Grab a page, blocking if it is busy and allocating a page if necessary.
2813 * A busy page is returned or NULL. The page may or may not be valid and
2814 * might not be on a queue (the caller is responsible for the disposition of
2817 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
2818 * page will be zero'd and marked valid.
2820 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
2821 * valid even if it already exists.
2823 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
2824 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
2825 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
2827 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
2828 * always returned if we had blocked.
2830 * This routine may not be called from an interrupt.
2832 * No other requirements.
2835 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
2841 KKASSERT(allocflags &
2842 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2843 vm_object_hold_shared(object);
2845 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
2847 vm_page_sleep_busy(m, TRUE, "pgrbwt");
2848 if ((allocflags & VM_ALLOC_RETRY) == 0) {
2853 } else if (m == NULL) {
2855 vm_object_upgrade(object);
2858 if (allocflags & VM_ALLOC_RETRY)
2859 allocflags |= VM_ALLOC_NULL_OK;
2860 m = vm_page_alloc(object, pindex,
2861 allocflags & ~VM_ALLOC_RETRY);
2865 if ((allocflags & VM_ALLOC_RETRY) == 0)
2874 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
2876 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
2877 * valid even if already valid.
2879 * NOTE! We have removed all of the PG_ZERO optimizations and also
2880 * removed the idle zeroing code. These optimizations actually
2881 * slow things down on modern cpus because the zerod area is
2882 * likely uncached, placing a memory-access burden on the
2883 * accesors taking the fault.
2885 * By always zeroing the page in-line with the fault, no
2886 * dynamic ram reads are needed and the caches are hot, ready
2887 * for userland to access the memory.
2889 if (m->valid == 0) {
2890 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
2891 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2892 m->valid = VM_PAGE_BITS_ALL;
2894 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
2895 pmap_zero_page(VM_PAGE_TO_PHYS(m));
2896 m->valid = VM_PAGE_BITS_ALL;
2899 vm_object_drop(object);
2904 * Mapping function for valid bits or for dirty bits in
2905 * a page. May not block.
2907 * Inputs are required to range within a page.
2913 vm_page_bits(int base, int size)
2919 base + size <= PAGE_SIZE,
2920 ("vm_page_bits: illegal base/size %d/%d", base, size)
2923 if (size == 0) /* handle degenerate case */
2926 first_bit = base >> DEV_BSHIFT;
2927 last_bit = (base + size - 1) >> DEV_BSHIFT;
2929 return ((2 << last_bit) - (1 << first_bit));
2933 * Sets portions of a page valid and clean. The arguments are expected
2934 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
2935 * of any partial chunks touched by the range. The invalid portion of
2936 * such chunks will be zero'd.
2938 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
2939 * align base to DEV_BSIZE so as not to mark clean a partially
2940 * truncated device block. Otherwise the dirty page status might be
2943 * This routine may not block.
2945 * (base + size) must be less then or equal to PAGE_SIZE.
2948 _vm_page_zero_valid(vm_page_t m, int base, int size)
2953 if (size == 0) /* handle degenerate case */
2957 * If the base is not DEV_BSIZE aligned and the valid
2958 * bit is clear, we have to zero out a portion of the
2962 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
2963 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
2965 pmap_zero_page_area(
2973 * If the ending offset is not DEV_BSIZE aligned and the
2974 * valid bit is clear, we have to zero out a portion of
2978 endoff = base + size;
2980 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
2981 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
2983 pmap_zero_page_area(
2986 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
2992 * Set valid, clear dirty bits. If validating the entire
2993 * page we can safely clear the pmap modify bit. We also
2994 * use this opportunity to clear the PG_NOSYNC flag. If a process
2995 * takes a write fault on a MAP_NOSYNC memory area the flag will
2998 * We set valid bits inclusive of any overlap, but we can only
2999 * clear dirty bits for DEV_BSIZE chunks that are fully within
3002 * Page must be busied?
3003 * No other requirements.
3006 vm_page_set_valid(vm_page_t m, int base, int size)
3008 _vm_page_zero_valid(m, base, size);
3009 m->valid |= vm_page_bits(base, size);
3014 * Set valid bits and clear dirty bits.
3016 * NOTE: This function does not clear the pmap modified bit.
3017 * Also note that e.g. NFS may use a byte-granular base
3020 * WARNING: Page must be busied? But vfs_clean_one_page() will call
3021 * this without necessarily busying the page (via bdwrite()).
3022 * So for now vm_token must also be held.
3024 * No other requirements.
3027 vm_page_set_validclean(vm_page_t m, int base, int size)
3031 _vm_page_zero_valid(m, base, size);
3032 pagebits = vm_page_bits(base, size);
3033 m->valid |= pagebits;
3034 m->dirty &= ~pagebits;
3035 if (base == 0 && size == PAGE_SIZE) {
3036 /*pmap_clear_modify(m);*/
3037 vm_page_flag_clear(m, PG_NOSYNC);
3042 * Set valid & dirty. Used by buwrite()
3044 * WARNING: Page must be busied? But vfs_dirty_one_page() will
3045 * call this function in buwrite() so for now vm_token must
3048 * No other requirements.
3051 vm_page_set_validdirty(vm_page_t m, int base, int size)
3055 pagebits = vm_page_bits(base, size);
3056 m->valid |= pagebits;
3057 m->dirty |= pagebits;
3059 vm_object_set_writeable_dirty(m->object);
3065 * NOTE: This function does not clear the pmap modified bit.
3066 * Also note that e.g. NFS may use a byte-granular base
3069 * Page must be busied?
3070 * No other requirements.
3073 vm_page_clear_dirty(vm_page_t m, int base, int size)
3075 m->dirty &= ~vm_page_bits(base, size);
3076 if (base == 0 && size == PAGE_SIZE) {
3077 /*pmap_clear_modify(m);*/
3078 vm_page_flag_clear(m, PG_NOSYNC);
3083 * Make the page all-dirty.
3085 * Also make sure the related object and vnode reflect the fact that the
3086 * object may now contain a dirty page.
3088 * Page must be busied?
3089 * No other requirements.
3092 vm_page_dirty(vm_page_t m)
3095 int pqtype = m->queue - m->pc;
3097 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
3098 ("vm_page_dirty: page in free/cache queue!"));
3099 if (m->dirty != VM_PAGE_BITS_ALL) {
3100 m->dirty = VM_PAGE_BITS_ALL;
3102 vm_object_set_writeable_dirty(m->object);
3107 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3108 * valid and dirty bits for the effected areas are cleared.
3110 * Page must be busied?
3112 * No other requirements.
3115 vm_page_set_invalid(vm_page_t m, int base, int size)
3119 bits = vm_page_bits(base, size);
3122 m->object->generation++;
3126 * The kernel assumes that the invalid portions of a page contain
3127 * garbage, but such pages can be mapped into memory by user code.
3128 * When this occurs, we must zero out the non-valid portions of the
3129 * page so user code sees what it expects.
3131 * Pages are most often semi-valid when the end of a file is mapped
3132 * into memory and the file's size is not page aligned.
3134 * Page must be busied?
3135 * No other requirements.
3138 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3144 * Scan the valid bits looking for invalid sections that
3145 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
3146 * valid bit may be set ) have already been zerod by
3147 * vm_page_set_validclean().
3149 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3150 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3151 (m->valid & (1 << i))
3154 pmap_zero_page_area(
3157 (i - b) << DEV_BSHIFT
3165 * setvalid is TRUE when we can safely set the zero'd areas
3166 * as being valid. We can do this if there are no cache consistency
3167 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3170 m->valid = VM_PAGE_BITS_ALL;
3174 * Is a (partial) page valid? Note that the case where size == 0
3175 * will return FALSE in the degenerate case where the page is entirely
3176 * invalid, and TRUE otherwise.
3179 * No other requirements.
3182 vm_page_is_valid(vm_page_t m, int base, int size)
3184 int bits = vm_page_bits(base, size);
3186 if (m->valid && ((m->valid & bits) == bits))
3193 * update dirty bits from pmap/mmu. May not block.
3195 * Caller must hold the page busy
3198 vm_page_test_dirty(vm_page_t m)
3200 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
3206 * Register an action, associating it with its vm_page
3209 vm_page_register_action(vm_page_action_t action, vm_page_event_t event)
3211 struct vm_page_action_list *list;
3214 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
3215 list = &action_list[hv];
3217 lwkt_gettoken(&vm_token);
3218 vm_page_flag_set(action->m, PG_ACTIONLIST);
3219 action->event = event;
3220 LIST_INSERT_HEAD(list, action, entry);
3221 lwkt_reltoken(&vm_token);
3225 * Unregister an action, disassociating it from its related vm_page
3228 vm_page_unregister_action(vm_page_action_t action)
3230 struct vm_page_action_list *list;
3233 lwkt_gettoken(&vm_token);
3234 if (action->event != VMEVENT_NONE) {
3235 action->event = VMEVENT_NONE;
3236 LIST_REMOVE(action, entry);
3238 hv = (int)((intptr_t)action->m >> 8) & VMACTION_HMASK;
3239 list = &action_list[hv];
3240 if (LIST_EMPTY(list))
3241 vm_page_flag_clear(action->m, PG_ACTIONLIST);
3243 lwkt_reltoken(&vm_token);
3247 * Issue an event on a VM page. Corresponding action structures are
3248 * removed from the page's list and called.
3250 * If the vm_page has no more pending action events we clear its
3251 * PG_ACTIONLIST flag.
3254 vm_page_event_internal(vm_page_t m, vm_page_event_t event)
3256 struct vm_page_action_list *list;
3257 struct vm_page_action *scan;
3258 struct vm_page_action *next;
3262 hv = (int)((intptr_t)m >> 8) & VMACTION_HMASK;
3263 list = &action_list[hv];
3266 lwkt_gettoken(&vm_token);
3267 LIST_FOREACH_MUTABLE(scan, list, entry, next) {
3269 if (scan->event == event) {
3270 scan->event = VMEVENT_NONE;
3271 LIST_REMOVE(scan, entry);
3272 scan->func(m, scan);
3280 vm_page_flag_clear(m, PG_ACTIONLIST);
3281 lwkt_reltoken(&vm_token);
3284 #include "opt_ddb.h"
3286 #include <sys/kernel.h>
3288 #include <ddb/ddb.h>
3290 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3292 db_printf("vmstats.v_free_count: %d\n", vmstats.v_free_count);
3293 db_printf("vmstats.v_cache_count: %d\n", vmstats.v_cache_count);
3294 db_printf("vmstats.v_inactive_count: %d\n", vmstats.v_inactive_count);
3295 db_printf("vmstats.v_active_count: %d\n", vmstats.v_active_count);
3296 db_printf("vmstats.v_wire_count: %d\n", vmstats.v_wire_count);
3297 db_printf("vmstats.v_free_reserved: %d\n", vmstats.v_free_reserved);
3298 db_printf("vmstats.v_free_min: %d\n", vmstats.v_free_min);
3299 db_printf("vmstats.v_free_target: %d\n", vmstats.v_free_target);
3300 db_printf("vmstats.v_cache_min: %d\n", vmstats.v_cache_min);
3301 db_printf("vmstats.v_inactive_target: %d\n", vmstats.v_inactive_target);
3304 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3307 db_printf("PQ_FREE:");
3308 for(i=0;i<PQ_L2_SIZE;i++) {
3309 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
3313 db_printf("PQ_CACHE:");
3314 for(i=0;i<PQ_L2_SIZE;i++) {
3315 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
3319 db_printf("PQ_ACTIVE:");
3320 for(i=0;i<PQ_L2_SIZE;i++) {
3321 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3325 db_printf("PQ_INACTIVE:");
3326 for(i=0;i<PQ_L2_SIZE;i++) {
3327 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);