2 * Copyright (c) 1991 Regents of the University of California.
4 * Copyright (c) 2003-2011 The DragonFly Project. All rights reserved.
6 * This code is derived from software contributed to Berkeley by
7 * The Mach Operating System project at Carnegie-Mellon University.
9 * This code is derived from software contributed to The DragonFly Project
10 * by Matthew Dillon <dillon@backplane.com>
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
15 * 1. Redistributions of source code must retain the above copyright
16 * notice, this list of conditions and the following disclaimer.
17 * 2. Redistributions in binary form must reproduce the above copyright
18 * notice, this list of conditions and the following disclaimer in the
19 * documentation and/or other materials provided with the distribution.
20 * 3. Neither the name of the University nor the names of its contributors
21 * may be used to endorse or promote products derived from this software
22 * without specific prior written permission.
24 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
25 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
26 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
27 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
28 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
29 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
30 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
31 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
32 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
33 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 * from: @(#)vm_page.c 7.4 (Berkeley) 5/7/91
37 * $FreeBSD: src/sys/vm/vm_page.c,v 1.147.2.18 2002/03/10 05:03:19 alc Exp $
41 * Copyright (c) 1987, 1990 Carnegie-Mellon University.
42 * All rights reserved.
44 * Authors: Avadis Tevanian, Jr., Michael Wayne Young
46 * Permission to use, copy, modify and distribute this software and
47 * its documentation is hereby granted, provided that both the copyright
48 * notice and this permission notice appear in all copies of the
49 * software, derivative works or modified versions, and any portions
50 * thereof, and that both notices appear in supporting documentation.
52 * CARNEGIE MELLON ALLOWS FREE USE OF THIS SOFTWARE IN ITS "AS IS"
53 * CONDITION. CARNEGIE MELLON DISCLAIMS ANY LIABILITY OF ANY KIND
54 * FOR ANY DAMAGES WHATSOEVER RESULTING FROM THE USE OF THIS SOFTWARE.
56 * Carnegie Mellon requests users of this software to return to
58 * Software Distribution Coordinator or Software.Distribution@CS.CMU.EDU
59 * School of Computer Science
60 * Carnegie Mellon University
61 * Pittsburgh PA 15213-3890
63 * any improvements or extensions that they make and grant Carnegie the
64 * rights to redistribute these changes.
67 * Resident memory management module. The module manipulates 'VM pages'.
68 * A VM page is the core building block for memory management.
71 #include <sys/param.h>
72 #include <sys/systm.h>
73 #include <sys/malloc.h>
75 #include <sys/vmmeter.h>
76 #include <sys/vnode.h>
77 #include <sys/kernel.h>
78 #include <sys/alist.h>
79 #include <sys/sysctl.h>
80 #include <sys/cpu_topology.h>
83 #include <vm/vm_param.h>
85 #include <vm/vm_kern.h>
87 #include <vm/vm_map.h>
88 #include <vm/vm_object.h>
89 #include <vm/vm_page.h>
90 #include <vm/vm_pageout.h>
91 #include <vm/vm_pager.h>
92 #include <vm/vm_extern.h>
93 #include <vm/swap_pager.h>
95 #include <machine/inttypes.h>
96 #include <machine/md_var.h>
97 #include <machine/specialreg.h>
98 #include <machine/bus_dma.h>
100 #include <vm/vm_page2.h>
101 #include <sys/spinlock2.h>
104 * SET - Minimum required set associative size, must be a power of 2. We
105 * want this to match or exceed the set-associativeness of the cpu.
107 * GRP - A larger set that allows bleed-over into the domains of other
108 * nearby cpus. Also must be a power of 2. Used by the page zeroing
109 * code to smooth things out a bit.
111 #define PQ_SET_ASSOC 16
112 #define PQ_SET_ASSOC_MASK (PQ_SET_ASSOC - 1)
114 #define PQ_GRP_ASSOC (PQ_SET_ASSOC * 2)
115 #define PQ_GRP_ASSOC_MASK (PQ_GRP_ASSOC - 1)
117 static void vm_page_queue_init(void);
118 static void vm_page_free_wakeup(void);
119 static vm_page_t vm_page_select_cache(u_short pg_color);
120 static vm_page_t _vm_page_list_find2(int basequeue, int index);
121 static void _vm_page_deactivate_locked(vm_page_t m, int athead);
122 static void vm_numa_add_topology_mem(cpu_node_t *cpup, int physid, long bytes);
125 * Array of tailq lists
127 __cachealign struct vpgqueues vm_page_queues[PQ_COUNT];
129 static volatile int vm_pages_waiting;
130 static struct alist vm_contig_alist;
131 static struct almeta vm_contig_ameta[ALIST_RECORDS_65536];
132 static struct spinlock vm_contig_spin = SPINLOCK_INITIALIZER(&vm_contig_spin, "vm_contig_spin");
134 static u_long vm_dma_reserved = 0;
135 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved);
136 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0,
137 "Memory reserved for DMA");
138 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD,
139 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA");
141 static int vm_contig_verbose = 0;
142 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose);
144 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
145 vm_pindex_t, pindex);
148 vm_page_queue_init(void)
152 for (i = 0; i < PQ_L2_SIZE; i++)
153 vm_page_queues[PQ_FREE+i].cnt_offset =
154 offsetof(struct vmstats, v_free_count);
155 for (i = 0; i < PQ_L2_SIZE; i++)
156 vm_page_queues[PQ_CACHE+i].cnt_offset =
157 offsetof(struct vmstats, v_cache_count);
158 for (i = 0; i < PQ_L2_SIZE; i++)
159 vm_page_queues[PQ_INACTIVE+i].cnt_offset =
160 offsetof(struct vmstats, v_inactive_count);
161 for (i = 0; i < PQ_L2_SIZE; i++)
162 vm_page_queues[PQ_ACTIVE+i].cnt_offset =
163 offsetof(struct vmstats, v_active_count);
164 for (i = 0; i < PQ_L2_SIZE; i++)
165 vm_page_queues[PQ_HOLD+i].cnt_offset =
166 offsetof(struct 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");
176 * note: place in initialized data section? Is this necessary?
178 vm_pindex_t first_page = 0;
179 vm_pindex_t vm_page_array_size = 0;
180 vm_page_t vm_page_array = NULL;
181 vm_paddr_t vm_low_phys_reserved;
186 * Sets the page size, perhaps based upon the memory size.
187 * Must be called before any use of page-size dependent functions.
190 vm_set_page_size(void)
192 if (vmstats.v_page_size == 0)
193 vmstats.v_page_size = PAGE_SIZE;
194 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
195 panic("vm_set_page_size: page size not a power of two");
201 * Add a new page to the freelist for use by the system. New pages
202 * are added to both the head and tail of the associated free page
203 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
204 * requests pull 'recent' adds (higher physical addresses) first.
206 * Beware that the page zeroing daemon will also be running soon after
207 * boot, moving pages from the head to the tail of the PQ_FREE queues.
209 * Must be called in a critical section.
212 vm_add_new_page(vm_paddr_t pa)
214 struct vpgqueues *vpq;
217 m = PHYS_TO_VM_PAGE(pa);
220 m->pat_mode = PAT_WRITE_BACK;
221 m->pc = (pa >> PAGE_SHIFT);
224 * Twist for cpu localization in addition to page coloring, so
225 * different cpus selecting by m->queue get different page colors.
227 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE);
228 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE));
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_long(&vmstats.v_page_count, 1);
237 atomic_add_long(&vmstats.v_dma_pages, 1);
240 atomic_add_long(&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_long(&vmstats.v_page_count, 1);
252 atomic_add_long(&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, x86 still needs this workspace internally within the
340 * minidump code. In theory, they are not needed on x86, 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 if (bootverbose && ctob(physmem) >= 400LL*1024*1024*1024)
388 kprintf("initializing vm_page_array ");
389 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
390 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE);
391 vm_page_array = (vm_page_t)mapped;
393 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
395 * since pmap_map on amd64 returns stuff out of a direct-map region,
396 * we have to manually add these pages to the minidump tracking so
397 * that they can be dumped, including the vm_page_array.
400 pa < phys_avail[biggestone].phys_end;
407 * Clear all of the page structures, run basic initialization so
408 * PHYS_TO_VM_PAGE() operates properly even on pages not in the
411 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
412 vm_page_array_size = page_range;
413 if (bootverbose && ctob(physmem) >= 400LL*1024*1024*1024)
414 kprintf("size = 0x%zx\n", vm_page_array_size);
416 m = &vm_page_array[0];
417 pa = ptoa(first_page);
418 for (i = 0; i < page_range; ++i) {
419 spin_init(&m->spin, "vm_page");
426 * Construct the free queue(s) in ascending order (by physical
427 * address) so that the first 16MB of physical memory is allocated
428 * last rather than first. On large-memory machines, this avoids
429 * the exhaustion of low physical memory before isa_dma_init has run.
431 vmstats.v_page_count = 0;
432 vmstats.v_free_count = 0;
433 for (i = 0; phys_avail[i].phys_end && npages > 0; ++i) {
434 pa = phys_avail[i].phys_beg;
438 last_pa = phys_avail[i].phys_end;
439 while (pa < last_pa && npages-- > 0) {
445 virtual2_start = vaddr;
447 virtual_start = vaddr;
448 mycpu->gd_vmstats = vmstats;
452 * (called from early boot only)
454 * Reorganize VM pages based on numa data. May be called as many times as
455 * necessary. Will reorganize the vm_page_t page color and related queue(s)
456 * to allow vm_page_alloc() to choose pages based on socket affinity.
458 * NOTE: This function is only called while we are still in UP mode, so
459 * we only need a critical section to protect the queues (which
460 * saves a lot of time, there are likely a ton of pages).
463 vm_numa_organize(vm_paddr_t ran_beg, vm_paddr_t bytes, int physid)
468 struct vpgqueues *vpq;
476 * Check if no physical information, or there was only one socket
477 * (so don't waste time doing nothing!).
479 if (cpu_topology_phys_ids <= 1 ||
480 cpu_topology_core_ids == 0) {
485 * Setup for our iteration. Note that ACPI may iterate CPU
486 * sockets starting at 0 or 1 or some other number. The
487 * cpu_topology code mod's it against the socket count.
489 ran_end = ran_beg + bytes;
491 socket_mod = PQ_L2_SIZE / cpu_topology_phys_ids;
492 socket_value = (physid % cpu_topology_phys_ids) * socket_mod;
493 mend = &vm_page_array[vm_page_array_size];
498 * Adjust cpu_topology's phys_mem parameter
501 vm_numa_add_topology_mem(root_cpu_node, physid, (long)bytes);
504 * Adjust vm_page->pc and requeue all affected pages. The
505 * allocator will then be able to localize memory allocations
508 for (i = 0; phys_avail[i].phys_end; ++i) {
509 scan_beg = phys_avail[i].phys_beg;
510 scan_end = phys_avail[i].phys_end;
511 if (scan_end <= ran_beg)
513 if (scan_beg >= ran_end)
515 if (scan_beg < ran_beg)
517 if (scan_end > ran_end)
519 if (atop(scan_end) > first_page + vm_page_array_size)
520 scan_end = ptoa(first_page + vm_page_array_size);
522 m = PHYS_TO_VM_PAGE(scan_beg);
523 while (scan_beg < scan_end) {
525 if (m->queue != PQ_NONE) {
526 vpq = &vm_page_queues[m->queue];
527 TAILQ_REMOVE(&vpq->pl, m, pageq);
529 /* queue doesn't change, no need to adj cnt */
532 m->pc += socket_value;
535 vpq = &vm_page_queues[m->queue];
536 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
538 /* queue doesn't change, no need to adj cnt */
541 m->pc += socket_value;
544 scan_beg += PAGE_SIZE;
553 * (called from early boot only)
555 * Don't allow the NUMA organization to leave vm_page_queues[] nodes
556 * completely empty for a logical cpu. Doing so would force allocations
557 * on that cpu to always borrow from a nearby cpu, create unnecessary
558 * contention, and cause vm_page_alloc() to iterate more queues and run more
561 * This situation can occur when memory sticks are not entirely populated,
562 * populated at different densities, or in naturally assymetric systems
563 * such as the 2990WX. There could very well be many vm_page_queues[]
564 * entries with *NO* pages assigned to them.
566 * Fixing this up ensures that each logical CPU has roughly the same
567 * sized memory pool, and more importantly ensures that logical CPUs
568 * do not wind up with an empty memory pool.
570 * At them moment we just iterate the other queues and borrow pages,
571 * moving them into the queues for cpus with severe deficits even though
572 * the memory might not be local to those cpus. I am not doing this in
573 * a 'smart' way, its effectively UMA style (sorta, since its page-by-page
574 * whereas real UMA typically exchanges address bits 8-10 with high address
575 * bits). But it works extremely well and gives us fairly good deterministic
576 * results on the cpu cores associated with these secondary nodes.
579 vm_numa_organize_finalize(void)
581 struct vpgqueues *vpq;
592 * Machines might not use an exact power of 2 for phys_ids,
593 * core_ids, ht_ids, etc. This can slightly reduce the actual
594 * range of indices in vm_page_queues[] that are nominally used.
596 if (cpu_topology_ht_ids) {
597 scale_lim = PQ_L2_SIZE / cpu_topology_phys_ids;
598 scale_lim = scale_lim / cpu_topology_core_ids;
599 scale_lim = scale_lim / cpu_topology_ht_ids;
600 scale_lim = scale_lim * cpu_topology_ht_ids;
601 scale_lim = scale_lim * cpu_topology_core_ids;
602 scale_lim = scale_lim * cpu_topology_phys_ids;
604 scale_lim = PQ_L2_SIZE;
608 * Calculate an average, set hysteresis for balancing from
609 * 10% below the average to the average.
612 for (i = 0; i < scale_lim; ++i) {
613 lcnt_hi += vm_page_queues[i].lcnt;
615 lcnt_hi /= scale_lim;
616 lcnt_lo = lcnt_hi - lcnt_hi / 10;
618 kprintf("vm_page: avg %ld pages per queue, %d queues\n",
622 for (i = 0; i < scale_lim; ++i) {
623 vpq = &vm_page_queues[PQ_FREE + i];
624 while (vpq->lcnt < lcnt_lo) {
625 struct vpgqueues *vptmp;
627 iter = (iter + 1) & PQ_L2_MASK;
628 vptmp = &vm_page_queues[PQ_FREE + iter];
629 if (vptmp->lcnt < lcnt_hi)
631 m = TAILQ_FIRST(&vptmp->pl);
632 KKASSERT(m->queue == PQ_FREE + iter);
633 TAILQ_REMOVE(&vptmp->pl, m, pageq);
635 /* queue doesn't change, no need to adj cnt */
639 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
648 vm_numa_add_topology_mem(cpu_node_t *cpup, int physid, long bytes)
655 cpup->phys_mem += bytes;
659 * All members should have the same chipid, so we only need
660 * to pull out one member.
662 if (CPUMASK_TESTNZERO(cpup->members)) {
663 cpuid = BSFCPUMASK(cpup->members);
665 get_chip_ID_from_APICID(CPUID_TO_APICID(cpuid))) {
666 cpup->phys_mem += bytes;
673 * Just inherit from the parent node
675 cpup->phys_mem = cpup->parent_node->phys_mem;
678 for (i = 0; i < MAXCPU && cpup->child_node[i]; ++i)
679 vm_numa_add_topology_mem(cpup->child_node[i], physid, bytes);
683 * We tended to reserve a ton of memory for contigmalloc(). Now that most
684 * drivers have initialized we want to return most the remaining free
685 * reserve back to the VM page queues so they can be used for normal
688 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool.
691 vm_page_startup_finish(void *dummy __unused)
700 spin_lock(&vm_contig_spin);
702 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
703 if (bfree <= vm_dma_reserved / PAGE_SIZE)
709 * Figure out how much of the initial reserve we have to
710 * free in order to reach our target.
712 bfree -= vm_dma_reserved / PAGE_SIZE;
714 blk += count - bfree;
719 * Calculate the nearest power of 2 <= count.
721 for (xcount = 1; xcount <= count; xcount <<= 1)
724 blk += count - xcount;
728 * Allocate the pages from the alist, then free them to
729 * the normal VM page queues.
731 * Pages allocated from the alist are wired. We have to
732 * busy, unwire, and free them. We must also adjust
733 * vm_low_phys_reserved before freeing any pages to prevent
736 rblk = alist_alloc(&vm_contig_alist, blk, count);
738 kprintf("vm_page_startup_finish: Unable to return "
739 "dma space @0x%08x/%d -> 0x%08x\n",
743 atomic_add_long(&vmstats.v_dma_pages, -(long)count);
744 spin_unlock(&vm_contig_spin);
746 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
747 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m);
749 vm_page_busy_wait(m, FALSE, "cpgfr");
750 vm_page_unwire(m, 0);
755 spin_lock(&vm_contig_spin);
757 spin_unlock(&vm_contig_spin);
760 * Print out how much DMA space drivers have already allocated and
761 * how much is left over.
763 kprintf("DMA space used: %jdk, remaining available: %jdk\n",
764 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) *
766 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024));
768 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY,
769 vm_page_startup_finish, NULL);
773 * Scan comparison function for Red-Black tree scans. An inclusive
774 * (start,end) is expected. Other fields are not used.
777 rb_vm_page_scancmp(struct vm_page *p, void *data)
779 struct rb_vm_page_scan_info *info = data;
781 if (p->pindex < info->start_pindex)
783 if (p->pindex > info->end_pindex)
789 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
791 if (p1->pindex < p2->pindex)
793 if (p1->pindex > p2->pindex)
799 vm_page_init(vm_page_t m)
801 /* do nothing for now. Called from pmap_page_init() */
805 * Each page queue has its own spin lock, which is fairly optimal for
806 * allocating and freeing pages at least.
808 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
809 * queue spinlock via this function. Also note that m->queue cannot change
810 * unless both the page and queue are locked.
814 _vm_page_queue_spin_lock(vm_page_t m)
819 if (queue != PQ_NONE) {
820 spin_lock(&vm_page_queues[queue].spin);
821 KKASSERT(queue == m->queue);
827 _vm_page_queue_spin_unlock(vm_page_t m)
833 if (queue != PQ_NONE)
834 spin_unlock(&vm_page_queues[queue].spin);
839 _vm_page_queues_spin_lock(u_short queue)
842 if (queue != PQ_NONE)
843 spin_lock(&vm_page_queues[queue].spin);
849 _vm_page_queues_spin_unlock(u_short queue)
852 if (queue != PQ_NONE)
853 spin_unlock(&vm_page_queues[queue].spin);
857 vm_page_queue_spin_lock(vm_page_t m)
859 _vm_page_queue_spin_lock(m);
863 vm_page_queues_spin_lock(u_short queue)
865 _vm_page_queues_spin_lock(queue);
869 vm_page_queue_spin_unlock(vm_page_t m)
871 _vm_page_queue_spin_unlock(m);
875 vm_page_queues_spin_unlock(u_short queue)
877 _vm_page_queues_spin_unlock(queue);
881 * This locks the specified vm_page and its queue in the proper order
882 * (page first, then queue). The queue may change so the caller must
887 _vm_page_and_queue_spin_lock(vm_page_t m)
889 vm_page_spin_lock(m);
890 _vm_page_queue_spin_lock(m);
895 _vm_page_and_queue_spin_unlock(vm_page_t m)
897 _vm_page_queues_spin_unlock(m->queue);
898 vm_page_spin_unlock(m);
902 vm_page_and_queue_spin_unlock(vm_page_t m)
904 _vm_page_and_queue_spin_unlock(m);
908 vm_page_and_queue_spin_lock(vm_page_t m)
910 _vm_page_and_queue_spin_lock(m);
914 * Helper function removes vm_page from its current queue.
915 * Returns the base queue the page used to be on.
917 * The vm_page and the queue must be spinlocked.
918 * This function will unlock the queue but leave the page spinlocked.
920 static __inline u_short
921 _vm_page_rem_queue_spinlocked(vm_page_t m)
923 struct vpgqueues *pq;
929 if (queue != PQ_NONE) {
930 pq = &vm_page_queues[queue];
931 TAILQ_REMOVE(&pq->pl, m, pageq);
934 * Adjust our pcpu stats. In order for the nominal low-memory
935 * algorithms to work properly we don't let any pcpu stat get
936 * too negative before we force it to be rolled-up into the
937 * global stats. Otherwise our pageout and vm_wait tests
940 * The idea here is to reduce unnecessary SMP cache
941 * mastership changes in the global vmstats, which can be
942 * particularly bad in multi-socket systems.
944 cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
945 atomic_add_long(cnt, -1);
946 if (*cnt < -VMMETER_SLOP_COUNT) {
947 u_long copy = atomic_swap_long(cnt, 0);
948 cnt = (long *)((char *)&vmstats + pq->cnt_offset);
949 atomic_add_long(cnt, copy);
950 cnt = (long *)((char *)&mycpu->gd_vmstats +
952 atomic_add_long(cnt, copy);
958 vm_page_queues_spin_unlock(oqueue); /* intended */
964 * Helper function places the vm_page on the specified queue. Generally
965 * speaking only PQ_FREE pages are placed at the head, to allow them to
966 * be allocated sooner rather than later on the assumption that they
969 * The vm_page must be spinlocked.
970 * This function will return with both the page and the queue locked.
973 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
975 struct vpgqueues *pq;
978 KKASSERT(m->queue == PQ_NONE);
980 if (queue != PQ_NONE) {
981 vm_page_queues_spin_lock(queue);
982 pq = &vm_page_queues[queue];
986 * Adjust our pcpu stats. If a system entity really needs
987 * to incorporate the count it will call vmstats_rollup()
988 * to roll it all up into the global vmstats strufture.
990 cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
991 atomic_add_long(cnt, 1);
994 * PQ_FREE is always handled LIFO style to try to provide
995 * cache-hot pages to programs.
998 if (queue - m->pc == PQ_FREE) {
999 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1000 } else if (athead) {
1001 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1003 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1005 /* leave the queue spinlocked */
1010 * Wait until page is no longer BUSY. If also_m_busy is TRUE we wait
1011 * until the page is no longer BUSY or SBUSY (busy_count field is 0).
1013 * Returns TRUE if it had to sleep, FALSE if we did not. Only one sleep
1014 * call will be made before returning.
1016 * This function does NOT busy the page and on return the page is not
1017 * guaranteed to be available.
1020 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
1022 u_int32_t busy_count;
1025 busy_count = m->busy_count;
1028 if ((busy_count & PBUSY_LOCKED) == 0 &&
1029 (also_m_busy == 0 || (busy_count & PBUSY_MASK) == 0)) {
1032 tsleep_interlock(m, 0);
1033 if (atomic_cmpset_int(&m->busy_count, busy_count,
1034 busy_count | PBUSY_WANTED)) {
1035 atomic_set_int(&m->flags, PG_REFERENCED);
1036 tsleep(m, PINTERLOCKED, msg, 0);
1043 * This calculates and returns a page color given an optional VM object and
1044 * either a pindex or an iterator. We attempt to return a cpu-localized
1045 * pg_color that is still roughly 16-way set-associative. The CPU topology
1046 * is used if it was probed.
1048 * The caller may use the returned value to index into e.g. PQ_FREE when
1049 * allocating a page in order to nominally obtain pages that are hopefully
1050 * already localized to the requesting cpu. This function is not able to
1051 * provide any sort of guarantee of this, but does its best to improve
1052 * hardware cache management performance.
1054 * WARNING! The caller must mask the returned value with PQ_L2_MASK.
1057 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex)
1060 int object_pg_color;
1063 * WARNING! cpu_topology_core_ids might not be a power of two.
1064 * We also shouldn't make assumptions about
1065 * cpu_topology_phys_ids either.
1067 * WARNING! ncpus might not be known at this time (during early
1068 * boot), and might be set to 1.
1070 * General format: [phys_id][core_id][cpuid][set-associativity]
1071 * (but uses modulo, so not necessarily precise bit masks)
1073 object_pg_color = object ? object->pg_color : 0;
1075 if (cpu_topology_ht_ids) {
1084 * Translate cpuid to socket, core, and hyperthread id.
1086 phys_id = get_cpu_phys_id(cpuid);
1087 core_id = get_cpu_core_id(cpuid);
1088 ht_id = get_cpu_ht_id(cpuid);
1091 * Calculate pg_color for our array index.
1093 * physcale - socket multiplier.
1094 * grpscale - core multiplier (cores per socket)
1095 * cpu* - cpus per core
1097 * WARNING! In early boot, ncpus has not yet been
1098 * initialized and may be set to (1).
1100 * WARNING! physcale must match the organization that
1101 * vm_numa_organize() creates to ensure that
1102 * we properly localize allocations to the
1105 physcale = PQ_L2_SIZE / cpu_topology_phys_ids;
1106 grpscale = physcale / cpu_topology_core_ids;
1107 cpuscale = grpscale / cpu_topology_ht_ids;
1109 pg_color = phys_id * physcale;
1110 pg_color += core_id * grpscale;
1111 pg_color += ht_id * cpuscale;
1112 pg_color += (pindex + object_pg_color) % cpuscale;
1116 pg_color += (pindex + object_pg_color) % grpsize;
1121 /* 3->9, 4->8, 5->10, 6->12, 7->14 */
1126 pg_color += (pindex + object_pg_color) % grpsize;
1131 * Unknown topology, distribute things evenly.
1133 * WARNING! In early boot, ncpus has not yet been
1134 * initialized and may be set to (1).
1138 cpuscale = PQ_L2_SIZE / ncpus;
1140 pg_color = cpuid * cpuscale;
1141 pg_color += (pindex + object_pg_color) % cpuscale;
1143 return (pg_color & PQ_L2_MASK);
1147 * Wait until BUSY can be set, then set it. If also_m_busy is TRUE we
1148 * also wait for m->busy_count to become 0 before setting PBUSY_LOCKED.
1151 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
1152 int also_m_busy, const char *msg
1155 u_int32_t busy_count;
1158 busy_count = m->busy_count;
1160 if (busy_count & PBUSY_LOCKED) {
1161 tsleep_interlock(m, 0);
1162 if (atomic_cmpset_int(&m->busy_count, busy_count,
1163 busy_count | PBUSY_WANTED)) {
1164 atomic_set_int(&m->flags, PG_REFERENCED);
1165 tsleep(m, PINTERLOCKED, msg, 0);
1167 } else if (also_m_busy && busy_count) {
1168 tsleep_interlock(m, 0);
1169 if (atomic_cmpset_int(&m->busy_count, busy_count,
1170 busy_count | PBUSY_WANTED)) {
1171 atomic_set_int(&m->flags, PG_REFERENCED);
1172 tsleep(m, PINTERLOCKED, msg, 0);
1175 if (atomic_cmpset_int(&m->busy_count, busy_count,
1176 busy_count | PBUSY_LOCKED)) {
1177 #ifdef VM_PAGE_DEBUG
1178 m->busy_func = func;
1179 m->busy_line = lineno;
1188 * Attempt to set BUSY. If also_m_busy is TRUE we only succeed if
1189 * m->busy_count is also 0.
1191 * Returns non-zero on failure.
1194 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
1197 u_int32_t busy_count;
1200 busy_count = m->busy_count;
1202 if (busy_count & PBUSY_LOCKED)
1204 if (also_m_busy && (busy_count & PBUSY_MASK) != 0)
1206 if (atomic_cmpset_int(&m->busy_count, busy_count,
1207 busy_count | PBUSY_LOCKED)) {
1208 #ifdef VM_PAGE_DEBUG
1209 m->busy_func = func;
1210 m->busy_line = lineno;
1218 * Clear the BUSY flag and return non-zero to indicate to the caller
1219 * that a wakeup() should be performed.
1221 * The vm_page must be spinlocked and will remain spinlocked on return.
1222 * The related queue must NOT be spinlocked (which could deadlock us).
1228 _vm_page_wakeup(vm_page_t m)
1230 u_int32_t busy_count;
1233 busy_count = m->busy_count;
1235 if (atomic_cmpset_int(&m->busy_count, busy_count,
1237 ~(PBUSY_LOCKED | PBUSY_WANTED))) {
1241 return((int)(busy_count & PBUSY_WANTED));
1245 * Clear the BUSY flag and wakeup anyone waiting for the page. This
1246 * is typically the last call you make on a page before moving onto
1250 vm_page_wakeup(vm_page_t m)
1252 KASSERT(m->busy_count & PBUSY_LOCKED,
1253 ("vm_page_wakeup: page not busy!!!"));
1254 vm_page_spin_lock(m);
1255 if (_vm_page_wakeup(m)) {
1256 vm_page_spin_unlock(m);
1259 vm_page_spin_unlock(m);
1264 * Holding a page keeps it from being reused. Other parts of the system
1265 * can still disassociate the page from its current object and free it, or
1266 * perform read or write I/O on it and/or otherwise manipulate the page,
1267 * but if the page is held the VM system will leave the page and its data
1268 * intact and not reuse the page for other purposes until the last hold
1269 * reference is released. (see vm_page_wire() if you want to prevent the
1270 * page from being disassociated from its object too).
1272 * The caller must still validate the contents of the page and, if necessary,
1273 * wait for any pending I/O (e.g. vm_page_sleep_busy() loop) to complete
1274 * before manipulating the page.
1276 * XXX get vm_page_spin_lock() here and move FREE->HOLD if necessary
1279 vm_page_hold(vm_page_t m)
1281 vm_page_spin_lock(m);
1282 atomic_add_int(&m->hold_count, 1);
1283 if (m->queue - m->pc == PQ_FREE) {
1284 _vm_page_queue_spin_lock(m);
1285 _vm_page_rem_queue_spinlocked(m);
1286 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
1287 _vm_page_queue_spin_unlock(m);
1289 vm_page_spin_unlock(m);
1293 * The opposite of vm_page_hold(). If the page is on the HOLD queue
1294 * it was freed while held and must be moved back to the FREE queue.
1297 vm_page_unhold(vm_page_t m)
1299 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE,
1300 ("vm_page_unhold: pg %p illegal hold_count (%d) or on FREE queue (%d)",
1301 m, m->hold_count, m->queue - m->pc));
1302 vm_page_spin_lock(m);
1303 atomic_add_int(&m->hold_count, -1);
1304 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
1305 _vm_page_queue_spin_lock(m);
1306 _vm_page_rem_queue_spinlocked(m);
1307 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1);
1308 _vm_page_queue_spin_unlock(m);
1310 vm_page_spin_unlock(m);
1316 * Create a fictitious page with the specified physical address and
1317 * memory attribute. The memory attribute is the only the machine-
1318 * dependent aspect of a fictitious page that must be initialized.
1322 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1325 if ((m->flags & PG_FICTITIOUS) != 0) {
1327 * The page's memattr might have changed since the
1328 * previous initialization. Update the pmap to the
1333 m->phys_addr = paddr;
1335 /* Fictitious pages don't use "segind". */
1336 /* Fictitious pages don't use "order" or "pool". */
1337 m->flags = PG_FICTITIOUS | PG_UNMANAGED;
1338 m->busy_count = PBUSY_LOCKED;
1340 spin_init(&m->spin, "fake_page");
1343 pmap_page_set_memattr(m, memattr);
1347 * Inserts the given vm_page into the object and object list.
1349 * The pagetables are not updated but will presumably fault the page
1350 * in if necessary, or if a kernel page the caller will at some point
1351 * enter the page into the kernel's pmap. We are not allowed to block
1352 * here so we *can't* do this anyway.
1354 * This routine may not block.
1355 * This routine must be called with the vm_object held.
1356 * This routine must be called with a critical section held.
1358 * This routine returns TRUE if the page was inserted into the object
1359 * successfully, and FALSE if the page already exists in the object.
1362 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1364 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object));
1365 if (m->object != NULL)
1366 panic("vm_page_insert: already inserted");
1368 atomic_add_int(&object->generation, 1);
1371 * Record the object/offset pair in this page and add the
1372 * pv_list_count of the page to the object.
1374 * The vm_page spin lock is required for interactions with the pmap.
1376 vm_page_spin_lock(m);
1379 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
1382 vm_page_spin_unlock(m);
1385 ++object->resident_page_count;
1386 ++mycpu->gd_vmtotal.t_rm;
1387 vm_page_spin_unlock(m);
1390 * Since we are inserting a new and possibly dirty page,
1391 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
1393 if ((m->valid & m->dirty) ||
1394 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT)))
1395 vm_object_set_writeable_dirty(object);
1398 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
1400 swap_pager_page_inserted(m);
1405 * Removes the given vm_page_t from the (object,index) table
1407 * The underlying pmap entry (if any) is NOT removed here.
1408 * This routine may not block.
1410 * The page must be BUSY and will remain BUSY on return.
1411 * No other requirements.
1413 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
1417 vm_page_remove(vm_page_t m)
1421 if (m->object == NULL) {
1425 if ((m->busy_count & PBUSY_LOCKED) == 0)
1426 panic("vm_page_remove: page not busy");
1430 vm_object_hold(object);
1433 * Remove the page from the object and update the object.
1435 * The vm_page spin lock is required for interactions with the pmap.
1437 vm_page_spin_lock(m);
1438 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
1439 --object->resident_page_count;
1440 --mycpu->gd_vmtotal.t_rm;
1442 atomic_add_int(&object->generation, 1);
1443 vm_page_spin_unlock(m);
1445 vm_object_drop(object);
1449 * Locate and return the page at (object, pindex), or NULL if the
1450 * page could not be found.
1452 * The caller must hold the vm_object token.
1455 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1460 * Search the hash table for this object/offset pair
1462 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1463 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1464 KKASSERT(m == NULL || (m->object == object && m->pindex == pindex));
1469 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1471 int also_m_busy, const char *msg
1474 u_int32_t busy_count;
1477 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1478 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1480 KKASSERT(m->object == object && m->pindex == pindex);
1481 busy_count = m->busy_count;
1483 if (busy_count & PBUSY_LOCKED) {
1484 tsleep_interlock(m, 0);
1485 if (atomic_cmpset_int(&m->busy_count, busy_count,
1486 busy_count | PBUSY_WANTED)) {
1487 atomic_set_int(&m->flags, PG_REFERENCED);
1488 tsleep(m, PINTERLOCKED, msg, 0);
1489 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1492 } else if (also_m_busy && busy_count) {
1493 tsleep_interlock(m, 0);
1494 if (atomic_cmpset_int(&m->busy_count, busy_count,
1495 busy_count | PBUSY_WANTED)) {
1496 atomic_set_int(&m->flags, PG_REFERENCED);
1497 tsleep(m, PINTERLOCKED, msg, 0);
1498 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1501 } else if (atomic_cmpset_int(&m->busy_count, busy_count,
1502 busy_count | PBUSY_LOCKED)) {
1503 #ifdef VM_PAGE_DEBUG
1504 m->busy_func = func;
1505 m->busy_line = lineno;
1514 * Attempt to lookup and busy a page.
1516 * Returns NULL if the page could not be found
1518 * Returns a vm_page and error == TRUE if the page exists but could not
1521 * Returns a vm_page and error == FALSE on success.
1524 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1526 int also_m_busy, int *errorp
1529 u_int32_t busy_count;
1532 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1533 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1536 KKASSERT(m->object == object && m->pindex == pindex);
1537 busy_count = m->busy_count;
1539 if (busy_count & PBUSY_LOCKED) {
1543 if (also_m_busy && busy_count) {
1547 if (atomic_cmpset_int(&m->busy_count, busy_count,
1548 busy_count | PBUSY_LOCKED)) {
1549 #ifdef VM_PAGE_DEBUG
1550 m->busy_func = func;
1551 m->busy_line = lineno;
1560 * Returns a page that is only soft-busied for use by the caller in
1561 * a read-only fashion. Returns NULL if the page could not be found,
1562 * the soft busy could not be obtained, or the page data is invalid.
1565 vm_page_lookup_sbusy_try(struct vm_object *object, vm_pindex_t pindex,
1566 int pgoff, int pgbytes)
1570 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1571 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1573 if ((m->valid != VM_PAGE_BITS_ALL &&
1574 !vm_page_is_valid(m, pgoff, pgbytes)) ||
1575 (m->flags & PG_FICTITIOUS)) {
1577 } else if (vm_page_sbusy_try(m)) {
1579 } else if ((m->valid != VM_PAGE_BITS_ALL &&
1580 !vm_page_is_valid(m, pgoff, pgbytes)) ||
1581 (m->flags & PG_FICTITIOUS)) {
1582 vm_page_sbusy_drop(m);
1590 * Caller must hold the related vm_object
1593 vm_page_next(vm_page_t m)
1597 next = vm_page_rb_tree_RB_NEXT(m);
1598 if (next && next->pindex != m->pindex + 1)
1606 * Move the given vm_page from its current object to the specified
1607 * target object/offset. The page must be busy and will remain so
1610 * new_object must be held.
1611 * This routine might block. XXX ?
1613 * NOTE: Swap associated with the page must be invalidated by the move. We
1614 * have to do this for several reasons: (1) we aren't freeing the
1615 * page, (2) we are dirtying the page, (3) the VM system is probably
1616 * moving the page from object A to B, and will then later move
1617 * the backing store from A to B and we can't have a conflict.
1619 * NOTE: We *always* dirty the page. It is necessary both for the
1620 * fact that we moved it, and because we may be invalidating
1621 * swap. If the page is on the cache, we have to deactivate it
1622 * or vm_page_dirty() will panic. Dirty pages are not allowed
1626 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1628 KKASSERT(m->busy_count & PBUSY_LOCKED);
1629 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object));
1631 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object));
1634 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1635 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1636 new_object, new_pindex);
1638 if (m->queue - m->pc == PQ_CACHE)
1639 vm_page_deactivate(m);
1644 * vm_page_unqueue() without any wakeup. This routine is used when a page
1645 * is to remain BUSYied by the caller.
1647 * This routine may not block.
1650 vm_page_unqueue_nowakeup(vm_page_t m)
1652 vm_page_and_queue_spin_lock(m);
1653 (void)_vm_page_rem_queue_spinlocked(m);
1654 vm_page_spin_unlock(m);
1658 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1661 * This routine may not block.
1664 vm_page_unqueue(vm_page_t m)
1668 vm_page_and_queue_spin_lock(m);
1669 queue = _vm_page_rem_queue_spinlocked(m);
1670 if (queue == PQ_FREE || queue == PQ_CACHE) {
1671 vm_page_spin_unlock(m);
1672 pagedaemon_wakeup();
1674 vm_page_spin_unlock(m);
1679 * vm_page_list_find()
1681 * Find a page on the specified queue with color optimization.
1683 * The page coloring optimization attempts to locate a page that does
1684 * not overload other nearby pages in the object in the cpu's L1 or L2
1685 * caches. We need this optimization because cpu caches tend to be
1686 * physical caches, while object spaces tend to be virtual.
1688 * The page coloring optimization also, very importantly, tries to localize
1689 * memory to cpus and physical sockets.
1691 * On MP systems each PQ_FREE and PQ_CACHE color queue has its own spinlock
1692 * and the algorithm is adjusted to localize allocations on a per-core basis.
1693 * This is done by 'twisting' the colors.
1695 * The page is returned spinlocked and removed from its queue (it will
1696 * be on PQ_NONE), or NULL. The page is not BUSY'd. The caller
1697 * is responsible for dealing with the busy-page case (usually by
1698 * deactivating the page and looping).
1700 * NOTE: This routine is carefully inlined. A non-inlined version
1701 * is available for outside callers but the only critical path is
1702 * from within this source file.
1704 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1705 * represent stable storage, allowing us to order our locks vm_page
1706 * first, then queue.
1710 _vm_page_list_find(int basequeue, int index)
1715 m = TAILQ_FIRST(&vm_page_queues[basequeue+index].pl);
1717 m = _vm_page_list_find2(basequeue, index);
1720 vm_page_and_queue_spin_lock(m);
1721 if (m->queue == basequeue + index) {
1722 _vm_page_rem_queue_spinlocked(m);
1723 /* vm_page_t spin held, no queue spin */
1726 vm_page_and_queue_spin_unlock(m);
1732 * If we could not find the page in the desired queue try to find it in
1736 _vm_page_list_find2(int basequeue, int index)
1738 struct vpgqueues *pq;
1740 int pqmask = PQ_SET_ASSOC_MASK >> 1;
1744 index &= PQ_L2_MASK;
1745 pq = &vm_page_queues[basequeue];
1748 * Run local sets of 16, 32, 64, 128, and the whole queue if all
1749 * else fails (PQ_L2_MASK which is 255).
1752 pqmask = (pqmask << 1) | 1;
1753 for (i = 0; i <= pqmask; ++i) {
1754 pqi = (index & ~pqmask) | ((index + i) & pqmask);
1755 m = TAILQ_FIRST(&pq[pqi].pl);
1757 _vm_page_and_queue_spin_lock(m);
1758 if (m->queue == basequeue + pqi) {
1759 _vm_page_rem_queue_spinlocked(m);
1762 _vm_page_and_queue_spin_unlock(m);
1767 } while (pqmask != PQ_L2_MASK);
1773 * Returns a vm_page candidate for allocation. The page is not busied so
1774 * it can move around. The caller must busy the page (and typically
1775 * deactivate it if it cannot be busied!)
1777 * Returns a spinlocked vm_page that has been removed from its queue.
1780 vm_page_list_find(int basequeue, int index)
1782 return(_vm_page_list_find(basequeue, index));
1786 * Find a page on the cache queue with color optimization, remove it
1787 * from the queue, and busy it. The returned page will not be spinlocked.
1789 * A candidate failure will be deactivated. Candidates can fail due to
1790 * being busied by someone else, in which case they will be deactivated.
1792 * This routine may not block.
1796 vm_page_select_cache(u_short pg_color)
1801 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK);
1805 * (m) has been removed from its queue and spinlocked
1807 if (vm_page_busy_try(m, TRUE)) {
1808 _vm_page_deactivate_locked(m, 0);
1809 vm_page_spin_unlock(m);
1812 * We successfully busied the page
1814 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 &&
1815 m->hold_count == 0 &&
1816 m->wire_count == 0 &&
1817 (m->dirty & m->valid) == 0) {
1818 vm_page_spin_unlock(m);
1819 pagedaemon_wakeup();
1824 * The page cannot be recycled, deactivate it.
1826 _vm_page_deactivate_locked(m, 0);
1827 if (_vm_page_wakeup(m)) {
1828 vm_page_spin_unlock(m);
1831 vm_page_spin_unlock(m);
1839 * Find a free page. We attempt to inline the nominal case and fall back
1840 * to _vm_page_select_free() otherwise. A busied page is removed from
1841 * the queue and returned.
1843 * This routine may not block.
1845 static __inline vm_page_t
1846 vm_page_select_free(u_short pg_color)
1851 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK);
1854 if (vm_page_busy_try(m, TRUE)) {
1856 * Various mechanisms such as a pmap_collect can
1857 * result in a busy page on the free queue. We
1858 * have to move the page out of the way so we can
1859 * retry the allocation. If the other thread is not
1860 * allocating the page then m->valid will remain 0 and
1861 * the pageout daemon will free the page later on.
1863 * Since we could not busy the page, however, we
1864 * cannot make assumptions as to whether the page
1865 * will be allocated by the other thread or not,
1866 * so all we can do is deactivate it to move it out
1867 * of the way. In particular, if the other thread
1868 * wires the page it may wind up on the inactive
1869 * queue and the pageout daemon will have to deal
1870 * with that case too.
1872 _vm_page_deactivate_locked(m, 0);
1873 vm_page_spin_unlock(m);
1876 * Theoretically if we are able to busy the page
1877 * atomic with the queue removal (using the vm_page
1878 * lock) nobody else should be able to mess with the
1881 KKASSERT((m->flags & (PG_UNMANAGED |
1882 PG_NEED_COMMIT)) == 0);
1883 KASSERT(m->hold_count == 0, ("m->hold_count is not zero "
1884 "pg %p q=%d flags=%08x hold=%d wire=%d",
1885 m, m->queue, m->flags, m->hold_count, m->wire_count));
1886 KKASSERT(m->wire_count == 0);
1887 vm_page_spin_unlock(m);
1888 pagedaemon_wakeup();
1890 /* return busied and removed page */
1900 * Allocate and return a memory cell associated with this VM object/offset
1901 * pair. If object is NULL an unassociated page will be allocated.
1903 * The returned page will be busied and removed from its queues. This
1904 * routine can block and may return NULL if a race occurs and the page
1905 * is found to already exist at the specified (object, pindex).
1907 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
1908 * VM_ALLOC_QUICK like normal but cannot use cache
1909 * VM_ALLOC_SYSTEM greater free drain
1910 * VM_ALLOC_INTERRUPT allow free list to be completely drained
1911 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
1912 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
1913 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
1914 * (see vm_page_grab())
1915 * VM_ALLOC_USE_GD ok to use per-gd cache
1917 * VM_ALLOC_CPU(n) allocate using specified cpu localization
1919 * The object must be held if not NULL
1920 * This routine may not block
1922 * Additional special handling is required when called from an interrupt
1923 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
1927 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
1937 * Special per-cpu free VM page cache. The pages are pre-busied
1938 * and pre-zerod for us.
1940 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
1942 if (gd->gd_vmpg_count) {
1943 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
1955 * CPU localization algorithm. Break the page queues up by physical
1956 * id and core id (note that two cpu threads will have the same core
1957 * id, and core_id != gd_cpuid).
1959 * This is nowhere near perfect, for example the last pindex in a
1960 * subgroup will overflow into the next cpu or package. But this
1961 * should get us good page reuse locality in heavy mixed loads.
1963 * (may be executed before the APs are started, so other GDs might
1966 if (page_req & VM_ALLOC_CPU_SPEC)
1967 cpuid_local = VM_ALLOC_GETCPU(page_req);
1969 cpuid_local = mycpu->gd_cpuid;
1971 pg_color = vm_get_pg_color(cpuid_local, object, pindex);
1974 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
1975 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
1978 * Certain system threads (pageout daemon, buf_daemon's) are
1979 * allowed to eat deeper into the free page list.
1981 if (curthread->td_flags & TDF_SYSTHREAD)
1982 page_req |= VM_ALLOC_SYSTEM;
1985 * Impose various limitations. Note that the v_free_reserved test
1986 * must match the opposite of vm_page_count_target() to avoid
1987 * livelocks, be careful.
1991 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved ||
1992 ((page_req & VM_ALLOC_INTERRUPT) &&
1993 gd->gd_vmstats.v_free_count > 0) ||
1994 ((page_req & VM_ALLOC_SYSTEM) &&
1995 gd->gd_vmstats.v_cache_count == 0 &&
1996 gd->gd_vmstats.v_free_count >
1997 gd->gd_vmstats.v_interrupt_free_min)
2000 * The free queue has sufficient free pages to take one out.
2002 m = vm_page_select_free(pg_color);
2003 } else if (page_req & VM_ALLOC_NORMAL) {
2005 * Allocatable from the cache (non-interrupt only). On
2006 * success, we must free the page and try again, thus
2007 * ensuring that vmstats.v_*_free_min counters are replenished.
2010 if (curthread->td_preempted) {
2011 kprintf("vm_page_alloc(): warning, attempt to allocate"
2012 " cache page from preempting interrupt\n");
2015 m = vm_page_select_cache(pg_color);
2018 m = vm_page_select_cache(pg_color);
2021 * On success move the page into the free queue and loop.
2023 * Only do this if we can safely acquire the vm_object lock,
2024 * because this is effectively a random page and the caller
2025 * might be holding the lock shared, we don't want to
2029 KASSERT(m->dirty == 0,
2030 ("Found dirty cache page %p", m));
2031 if ((obj = m->object) != NULL) {
2032 if (vm_object_hold_try(obj)) {
2033 vm_page_protect(m, VM_PROT_NONE);
2035 /* m->object NULL here */
2036 vm_object_drop(obj);
2038 vm_page_deactivate(m);
2042 vm_page_protect(m, VM_PROT_NONE);
2049 * On failure return NULL
2051 atomic_add_int(&vm_pageout_deficit, 1);
2052 pagedaemon_wakeup();
2056 * No pages available, wakeup the pageout daemon and give up.
2058 atomic_add_int(&vm_pageout_deficit, 1);
2059 pagedaemon_wakeup();
2064 * v_free_count can race so loop if we don't find the expected
2073 * Good page found. The page has already been busied for us and
2074 * removed from its queues.
2076 KASSERT(m->dirty == 0,
2077 ("vm_page_alloc: free/cache page %p was dirty", m));
2078 KKASSERT(m->queue == PQ_NONE);
2084 * Initialize the structure, inheriting some flags but clearing
2085 * all the rest. The page has already been busied for us.
2087 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK);
2089 KKASSERT(m->wire_count == 0);
2090 KKASSERT((m->busy_count & PBUSY_MASK) == 0);
2095 * Caller must be holding the object lock (asserted by
2096 * vm_page_insert()).
2098 * NOTE: Inserting a page here does not insert it into any pmaps
2099 * (which could cause us to block allocating memory).
2101 * NOTE: If no object an unassociated page is allocated, m->pindex
2102 * can be used by the caller for any purpose.
2105 if (vm_page_insert(m, object, pindex) == FALSE) {
2107 if ((page_req & VM_ALLOC_NULL_OK) == 0)
2108 panic("PAGE RACE %p[%ld]/%p",
2109 object, (long)pindex, m);
2117 * Don't wakeup too often - wakeup the pageout daemon when
2118 * we would be nearly out of memory.
2120 pagedaemon_wakeup();
2123 * A BUSY page is returned.
2129 * Returns number of pages available in our DMA memory reserve
2130 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf)
2133 vm_contig_avail_pages(void)
2138 spin_lock(&vm_contig_spin);
2139 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
2140 spin_unlock(&vm_contig_spin);
2146 * Attempt to allocate contiguous physical memory with the specified
2150 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
2151 unsigned long alignment, unsigned long boundary,
2152 unsigned long size, vm_memattr_t memattr)
2158 static vm_pindex_t contig_rover;
2161 alignment >>= PAGE_SHIFT;
2164 boundary >>= PAGE_SHIFT;
2167 size = (size + PAGE_MASK) >> PAGE_SHIFT;
2171 * Disabled temporarily until we find a solution for DRM (a flag
2172 * to always use the free space reserve, for performance).
2174 if (high == BUS_SPACE_MAXADDR && alignment <= PAGE_SIZE &&
2175 boundary <= PAGE_SIZE && size == 1 &&
2176 memattr == VM_MEMATTR_DEFAULT) {
2178 * Any page will work, use vm_page_alloc()
2179 * (e.g. when used from kmem_alloc_attr())
2181 m = vm_page_alloc(NULL, (contig_rover++) & 0x7FFFFFFF,
2182 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2183 VM_ALLOC_INTERRUPT);
2184 m->valid = VM_PAGE_BITS_ALL;
2191 * Use the low-memory dma reserve
2193 spin_lock(&vm_contig_spin);
2194 blk = alist_alloc(&vm_contig_alist, 0, size);
2195 if (blk == ALIST_BLOCK_NONE) {
2196 spin_unlock(&vm_contig_spin);
2198 kprintf("vm_page_alloc_contig: %ldk nospace\n",
2199 (size << PAGE_SHIFT) / 1024);
2204 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
2205 alist_free(&vm_contig_alist, blk, size);
2206 spin_unlock(&vm_contig_spin);
2208 kprintf("vm_page_alloc_contig: %ldk high "
2210 (size << PAGE_SHIFT) / 1024,
2215 spin_unlock(&vm_contig_spin);
2216 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
2218 if (vm_contig_verbose) {
2219 kprintf("vm_page_alloc_contig: %016jx/%ldk "
2220 "(%016jx-%016jx al=%lu bo=%lu pgs=%lu attr=%d\n",
2221 (intmax_t)m->phys_addr,
2222 (size << PAGE_SHIFT) / 1024,
2223 low, high, alignment, boundary, size, memattr);
2225 if (memattr != VM_MEMATTR_DEFAULT) {
2226 for (i = 0;i < size; i++)
2227 pmap_page_set_memattr(&m[i], memattr);
2233 * Free contiguously allocated pages. The pages will be wired but not busy.
2234 * When freeing to the alist we leave them wired and not busy.
2237 vm_page_free_contig(vm_page_t m, unsigned long size)
2239 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
2240 vm_pindex_t start = pa >> PAGE_SHIFT;
2241 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
2243 if (vm_contig_verbose) {
2244 kprintf("vm_page_free_contig: %016jx/%ldk\n",
2245 (intmax_t)pa, size / 1024);
2247 if (pa < vm_low_phys_reserved) {
2248 KKASSERT(pa + size <= vm_low_phys_reserved);
2249 spin_lock(&vm_contig_spin);
2250 alist_free(&vm_contig_alist, start, pages);
2251 spin_unlock(&vm_contig_spin);
2254 vm_page_busy_wait(m, FALSE, "cpgfr");
2255 vm_page_unwire(m, 0);
2266 * Wait for sufficient free memory for nominal heavy memory use kernel
2269 * WARNING! Be sure never to call this in any vm_pageout code path, which
2270 * will trivially deadlock the system.
2273 vm_wait_nominal(void)
2275 while (vm_page_count_min(0))
2280 * Test if vm_wait_nominal() would block.
2283 vm_test_nominal(void)
2285 if (vm_page_count_min(0))
2291 * Block until free pages are available for allocation, called in various
2292 * places before memory allocations.
2294 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
2295 * more generous then that.
2301 * never wait forever
2305 lwkt_gettoken(&vm_token);
2307 if (curthread == pagethread ||
2308 curthread == emergpager) {
2310 * The pageout daemon itself needs pages, this is bad.
2312 if (vm_page_count_min(0)) {
2313 vm_pageout_pages_needed = 1;
2314 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
2318 * Wakeup the pageout daemon if necessary and wait.
2320 * Do not wait indefinitely for the target to be reached,
2321 * as load might prevent it from being reached any time soon.
2322 * But wait a little to try to slow down page allocations
2323 * and to give more important threads (the pagedaemon)
2324 * allocation priority.
2326 if (vm_page_count_target()) {
2327 if (vm_pages_needed == 0) {
2328 vm_pages_needed = 1;
2329 wakeup(&vm_pages_needed);
2331 ++vm_pages_waiting; /* SMP race ok */
2332 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
2335 lwkt_reltoken(&vm_token);
2339 * Block until free pages are available for allocation
2341 * Called only from vm_fault so that processes page faulting can be
2345 vm_wait_pfault(void)
2348 * Wakeup the pageout daemon if necessary and wait.
2350 * Do not wait indefinitely for the target to be reached,
2351 * as load might prevent it from being reached any time soon.
2352 * But wait a little to try to slow down page allocations
2353 * and to give more important threads (the pagedaemon)
2354 * allocation priority.
2356 if (vm_page_count_min(0)) {
2357 lwkt_gettoken(&vm_token);
2358 while (vm_page_count_severe()) {
2359 if (vm_page_count_target()) {
2362 if (vm_pages_needed == 0) {
2363 vm_pages_needed = 1;
2364 wakeup(&vm_pages_needed);
2366 ++vm_pages_waiting; /* SMP race ok */
2367 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
2370 * Do not stay stuck in the loop if the system is trying
2371 * to kill the process.
2374 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
2378 lwkt_reltoken(&vm_token);
2383 * Put the specified page on the active list (if appropriate). Ensure
2384 * that act_count is at least ACT_INIT but do not otherwise mess with it.
2386 * The caller should be holding the page busied ? XXX
2387 * This routine may not block.
2390 vm_page_activate(vm_page_t m)
2394 vm_page_spin_lock(m);
2395 if (m->queue - m->pc != PQ_ACTIVE) {
2396 _vm_page_queue_spin_lock(m);
2397 oqueue = _vm_page_rem_queue_spinlocked(m);
2398 /* page is left spinlocked, queue is unlocked */
2400 if (oqueue == PQ_CACHE)
2401 mycpu->gd_cnt.v_reactivated++;
2402 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2403 if (m->act_count < ACT_INIT)
2404 m->act_count = ACT_INIT;
2405 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
2407 _vm_page_and_queue_spin_unlock(m);
2408 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
2409 pagedaemon_wakeup();
2411 if (m->act_count < ACT_INIT)
2412 m->act_count = ACT_INIT;
2413 vm_page_spin_unlock(m);
2418 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2419 * routine is called when a page has been added to the cache or free
2422 * This routine may not block.
2424 static __inline void
2425 vm_page_free_wakeup(void)
2427 globaldata_t gd = mycpu;
2430 * If the pageout daemon itself needs pages, then tell it that
2431 * there are some free.
2433 if (vm_pageout_pages_needed &&
2434 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >=
2435 gd->gd_vmstats.v_pageout_free_min
2437 vm_pageout_pages_needed = 0;
2438 wakeup(&vm_pageout_pages_needed);
2442 * Wakeup processes that are waiting on memory.
2444 * Generally speaking we want to wakeup stuck processes as soon as
2445 * possible. !vm_page_count_min(0) is the absolute minimum point
2446 * where we can do this. Wait a bit longer to reduce degenerate
2447 * re-blocking (vm_page_free_hysteresis). The target check is just
2448 * to make sure the min-check w/hysteresis does not exceed the
2451 if (vm_pages_waiting) {
2452 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2453 !vm_page_count_target()) {
2454 vm_pages_waiting = 0;
2455 wakeup(&vmstats.v_free_count);
2456 ++mycpu->gd_cnt.v_ppwakeups;
2459 if (!vm_page_count_target()) {
2461 * Plenty of pages are free, wakeup everyone.
2463 vm_pages_waiting = 0;
2464 wakeup(&vmstats.v_free_count);
2465 ++mycpu->gd_cnt.v_ppwakeups;
2466 } else if (!vm_page_count_min(0)) {
2468 * Some pages are free, wakeup someone.
2470 int wcount = vm_pages_waiting;
2473 vm_pages_waiting = wcount;
2474 wakeup_one(&vmstats.v_free_count);
2475 ++mycpu->gd_cnt.v_ppwakeups;
2482 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
2483 * it from its VM object.
2485 * The vm_page must be BUSY on entry. BUSY will be released on
2486 * return (the page will have been freed).
2489 vm_page_free_toq(vm_page_t m)
2491 mycpu->gd_cnt.v_tfree++;
2492 KKASSERT((m->flags & PG_MAPPED) == 0);
2493 KKASSERT(m->busy_count & PBUSY_LOCKED);
2495 if ((m->busy_count & PBUSY_MASK) || ((m->queue - m->pc) == PQ_FREE)) {
2496 kprintf("vm_page_free: pindex(%lu), busy %08x, "
2498 (u_long)m->pindex, m->busy_count, m->hold_count);
2499 if ((m->queue - m->pc) == PQ_FREE)
2500 panic("vm_page_free: freeing free page");
2502 panic("vm_page_free: freeing busy page");
2506 * Remove from object, spinlock the page and its queues and
2507 * remove from any queue. No queue spinlock will be held
2508 * after this section (because the page was removed from any
2512 vm_page_and_queue_spin_lock(m);
2513 _vm_page_rem_queue_spinlocked(m);
2516 * No further management of fictitious pages occurs beyond object
2517 * and queue removal.
2519 if ((m->flags & PG_FICTITIOUS) != 0) {
2520 vm_page_spin_unlock(m);
2528 if (m->wire_count != 0) {
2529 if (m->wire_count > 1) {
2531 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2532 m->wire_count, (long)m->pindex);
2534 panic("vm_page_free: freeing wired page");
2538 * Clear the UNMANAGED flag when freeing an unmanaged page.
2539 * Clear the NEED_COMMIT flag
2541 if (m->flags & PG_UNMANAGED)
2542 vm_page_flag_clear(m, PG_UNMANAGED);
2543 if (m->flags & PG_NEED_COMMIT)
2544 vm_page_flag_clear(m, PG_NEED_COMMIT);
2546 if (m->hold_count != 0) {
2547 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2549 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1);
2553 * This sequence allows us to clear BUSY while still holding
2554 * its spin lock, which reduces contention vs allocators. We
2555 * must not leave the queue locked or _vm_page_wakeup() may
2558 _vm_page_queue_spin_unlock(m);
2559 if (_vm_page_wakeup(m)) {
2560 vm_page_spin_unlock(m);
2563 vm_page_spin_unlock(m);
2565 vm_page_free_wakeup();
2569 * vm_page_unmanage()
2571 * Prevent PV management from being done on the page. The page is
2572 * removed from the paging queues as if it were wired, and as a
2573 * consequence of no longer being managed the pageout daemon will not
2574 * touch it (since there is no way to locate the pte mappings for the
2575 * page). madvise() calls that mess with the pmap will also no longer
2576 * operate on the page.
2578 * Beyond that the page is still reasonably 'normal'. Freeing the page
2579 * will clear the flag.
2581 * This routine is used by OBJT_PHYS objects - objects using unswappable
2582 * physical memory as backing store rather then swap-backed memory and
2583 * will eventually be extended to support 4MB unmanaged physical
2586 * Caller must be holding the page busy.
2589 vm_page_unmanage(vm_page_t m)
2591 KKASSERT(m->busy_count & PBUSY_LOCKED);
2592 if ((m->flags & PG_UNMANAGED) == 0) {
2593 if (m->wire_count == 0)
2596 vm_page_flag_set(m, PG_UNMANAGED);
2600 * Mark this page as wired down by yet another map, removing it from
2601 * paging queues as necessary.
2603 * Caller must be holding the page busy.
2606 vm_page_wire(vm_page_t m)
2609 * Only bump the wire statistics if the page is not already wired,
2610 * and only unqueue the page if it is on some queue (if it is unmanaged
2611 * it is already off the queues). Don't do anything with fictitious
2612 * pages because they are always wired.
2614 KKASSERT(m->busy_count & PBUSY_LOCKED);
2615 if ((m->flags & PG_FICTITIOUS) == 0) {
2616 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2617 if ((m->flags & PG_UNMANAGED) == 0)
2619 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count, 1);
2621 KASSERT(m->wire_count != 0,
2622 ("vm_page_wire: wire_count overflow m=%p", m));
2627 * Release one wiring of this page, potentially enabling it to be paged again.
2629 * Many pages placed on the inactive queue should actually go
2630 * into the cache, but it is difficult to figure out which. What
2631 * we do instead, if the inactive target is well met, is to put
2632 * clean pages at the head of the inactive queue instead of the tail.
2633 * This will cause them to be moved to the cache more quickly and
2634 * if not actively re-referenced, freed more quickly. If we just
2635 * stick these pages at the end of the inactive queue, heavy filesystem
2636 * meta-data accesses can cause an unnecessary paging load on memory bound
2637 * processes. This optimization causes one-time-use metadata to be
2638 * reused more quickly.
2640 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2641 * the inactive queue. This helps the pageout daemon determine memory
2642 * pressure and act on out-of-memory situations more quickly.
2644 * BUT, if we are in a low-memory situation we have no choice but to
2645 * put clean pages on the cache queue.
2647 * A number of routines use vm_page_unwire() to guarantee that the page
2648 * will go into either the inactive or active queues, and will NEVER
2649 * be placed in the cache - for example, just after dirtying a page.
2650 * dirty pages in the cache are not allowed.
2652 * This routine may not block.
2655 vm_page_unwire(vm_page_t m, int activate)
2657 KKASSERT(m->busy_count & PBUSY_LOCKED);
2658 if (m->flags & PG_FICTITIOUS) {
2660 } else if (m->wire_count <= 0) {
2661 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2663 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2664 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count,-1);
2665 if (m->flags & PG_UNMANAGED) {
2667 } else if (activate || (m->flags & PG_NEED_COMMIT)) {
2668 vm_page_spin_lock(m);
2669 _vm_page_add_queue_spinlocked(m,
2670 PQ_ACTIVE + m->pc, 0);
2671 _vm_page_and_queue_spin_unlock(m);
2673 vm_page_spin_lock(m);
2674 vm_page_flag_clear(m, PG_WINATCFLS);
2675 _vm_page_add_queue_spinlocked(m,
2676 PQ_INACTIVE + m->pc, 0);
2677 ++vm_swapcache_inactive_heuristic;
2678 _vm_page_and_queue_spin_unlock(m);
2685 * Move the specified page to the inactive queue. If the page has
2686 * any associated swap, the swap is deallocated.
2688 * Normally athead is 0 resulting in LRU operation. athead is set
2689 * to 1 if we want this page to be 'as if it were placed in the cache',
2690 * except without unmapping it from the process address space.
2692 * vm_page's spinlock must be held on entry and will remain held on return.
2693 * This routine may not block.
2696 _vm_page_deactivate_locked(vm_page_t m, int athead)
2701 * Ignore if already inactive.
2703 if (m->queue - m->pc == PQ_INACTIVE)
2705 _vm_page_queue_spin_lock(m);
2706 oqueue = _vm_page_rem_queue_spinlocked(m);
2708 if (m->wire_count == 0 && (m->flags & PG_UNMANAGED) == 0) {
2709 if (oqueue == PQ_CACHE)
2710 mycpu->gd_cnt.v_reactivated++;
2711 vm_page_flag_clear(m, PG_WINATCFLS);
2712 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2714 ++vm_swapcache_inactive_heuristic;
2716 /* NOTE: PQ_NONE if condition not taken */
2717 _vm_page_queue_spin_unlock(m);
2718 /* leaves vm_page spinlocked */
2722 * Attempt to deactivate a page.
2727 vm_page_deactivate(vm_page_t m)
2729 vm_page_spin_lock(m);
2730 _vm_page_deactivate_locked(m, 0);
2731 vm_page_spin_unlock(m);
2735 vm_page_deactivate_locked(vm_page_t m)
2737 _vm_page_deactivate_locked(m, 0);
2741 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it.
2743 * This function returns non-zero if it successfully moved the page to
2746 * This function unconditionally unbusies the page on return.
2749 vm_page_try_to_cache(vm_page_t m)
2751 vm_page_spin_lock(m);
2752 if (m->dirty || m->hold_count || m->wire_count ||
2753 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2754 if (_vm_page_wakeup(m)) {
2755 vm_page_spin_unlock(m);
2758 vm_page_spin_unlock(m);
2762 vm_page_spin_unlock(m);
2765 * Page busied by us and no longer spinlocked. Dirty pages cannot
2766 * be moved to the cache.
2768 vm_page_test_dirty(m);
2769 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2778 * Attempt to free the page. If we cannot free it, we do nothing.
2779 * 1 is returned on success, 0 on failure.
2784 vm_page_try_to_free(vm_page_t m)
2786 vm_page_spin_lock(m);
2787 if (vm_page_busy_try(m, TRUE)) {
2788 vm_page_spin_unlock(m);
2793 * The page can be in any state, including already being on the free
2794 * queue. Check to see if it really can be freed.
2796 if (m->dirty || /* can't free if it is dirty */
2797 m->hold_count || /* or held (XXX may be wrong) */
2798 m->wire_count || /* or wired */
2799 (m->flags & (PG_UNMANAGED | /* or unmanaged */
2800 PG_NEED_COMMIT)) || /* or needs a commit */
2801 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */
2802 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */
2803 if (_vm_page_wakeup(m)) {
2804 vm_page_spin_unlock(m);
2807 vm_page_spin_unlock(m);
2811 vm_page_spin_unlock(m);
2814 * We can probably free the page.
2816 * Page busied by us and no longer spinlocked. Dirty pages will
2817 * not be freed by this function. We have to re-test the
2818 * dirty bit after cleaning out the pmaps.
2820 vm_page_test_dirty(m);
2821 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2825 vm_page_protect(m, VM_PROT_NONE);
2826 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2837 * Put the specified page onto the page cache queue (if appropriate).
2839 * The page must be busy, and this routine will release the busy and
2840 * possibly even free the page.
2843 vm_page_cache(vm_page_t m)
2846 * Not suitable for the cache
2848 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2849 (m->busy_count & PBUSY_MASK) ||
2850 m->wire_count || m->hold_count) {
2856 * Already in the cache (and thus not mapped)
2858 if ((m->queue - m->pc) == PQ_CACHE) {
2859 KKASSERT((m->flags & PG_MAPPED) == 0);
2865 * Caller is required to test m->dirty, but note that the act of
2866 * removing the page from its maps can cause it to become dirty
2867 * on an SMP system due to another cpu running in usermode.
2870 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2875 * Remove all pmaps and indicate that the page is not
2876 * writeable or mapped. Our vm_page_protect() call may
2877 * have blocked (especially w/ VM_PROT_NONE), so recheck
2880 vm_page_protect(m, VM_PROT_NONE);
2881 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2882 (m->busy_count & PBUSY_MASK) ||
2883 m->wire_count || m->hold_count) {
2885 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2886 vm_page_deactivate(m);
2889 _vm_page_and_queue_spin_lock(m);
2890 _vm_page_rem_queue_spinlocked(m);
2891 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
2892 _vm_page_queue_spin_unlock(m);
2893 if (_vm_page_wakeup(m)) {
2894 vm_page_spin_unlock(m);
2897 vm_page_spin_unlock(m);
2899 vm_page_free_wakeup();
2904 * vm_page_dontneed()
2906 * Cache, deactivate, or do nothing as appropriate. This routine
2907 * is typically used by madvise() MADV_DONTNEED.
2909 * Generally speaking we want to move the page into the cache so
2910 * it gets reused quickly. However, this can result in a silly syndrome
2911 * due to the page recycling too quickly. Small objects will not be
2912 * fully cached. On the otherhand, if we move the page to the inactive
2913 * queue we wind up with a problem whereby very large objects
2914 * unnecessarily blow away our inactive and cache queues.
2916 * The solution is to move the pages based on a fixed weighting. We
2917 * either leave them alone, deactivate them, or move them to the cache,
2918 * where moving them to the cache has the highest weighting.
2919 * By forcing some pages into other queues we eventually force the
2920 * system to balance the queues, potentially recovering other unrelated
2921 * space from active. The idea is to not force this to happen too
2924 * The page must be busied.
2927 vm_page_dontneed(vm_page_t m)
2929 static int dnweight;
2936 * occassionally leave the page alone
2938 if ((dnw & 0x01F0) == 0 ||
2939 m->queue - m->pc == PQ_INACTIVE ||
2940 m->queue - m->pc == PQ_CACHE
2942 if (m->act_count >= ACT_INIT)
2948 * If vm_page_dontneed() is inactivating a page, it must clear
2949 * the referenced flag; otherwise the pagedaemon will see references
2950 * on the page in the inactive queue and reactivate it. Until the
2951 * page can move to the cache queue, madvise's job is not done.
2953 vm_page_flag_clear(m, PG_REFERENCED);
2954 pmap_clear_reference(m);
2957 vm_page_test_dirty(m);
2959 if (m->dirty || (dnw & 0x0070) == 0) {
2961 * Deactivate the page 3 times out of 32.
2966 * Cache the page 28 times out of every 32. Note that
2967 * the page is deactivated instead of cached, but placed
2968 * at the head of the queue instead of the tail.
2972 vm_page_spin_lock(m);
2973 _vm_page_deactivate_locked(m, head);
2974 vm_page_spin_unlock(m);
2978 * These routines manipulate the 'soft busy' count for a page. A soft busy
2979 * is almost like a hard BUSY except that it allows certain compatible
2980 * operations to occur on the page while it is busy. For example, a page
2981 * undergoing a write can still be mapped read-only.
2983 * We also use soft-busy to quickly pmap_enter shared read-only pages
2984 * without having to hold the page locked.
2986 * The soft-busy count can be > 1 in situations where multiple threads
2987 * are pmap_enter()ing the same page simultaneously, or when two buffer
2988 * cache buffers overlap the same page.
2990 * The caller must hold the page BUSY when making these two calls.
2993 vm_page_io_start(vm_page_t m)
2997 ocount = atomic_fetchadd_int(&m->busy_count, 1);
2998 KKASSERT(ocount & PBUSY_LOCKED);
3002 vm_page_io_finish(vm_page_t m)
3006 ocount = atomic_fetchadd_int(&m->busy_count, -1);
3007 KKASSERT(ocount & PBUSY_MASK);
3009 if (((ocount - 1) & (PBUSY_LOCKED | PBUSY_MASK)) == 0)
3015 * Attempt to soft-busy a page. The page must not be PBUSY_LOCKED.
3017 * We can't use fetchadd here because we might race a hard-busy and the
3018 * page freeing code asserts on a non-zero soft-busy count (even if only
3021 * Returns 0 on success, non-zero on failure.
3024 vm_page_sbusy_try(vm_page_t m)
3029 ocount = m->busy_count;
3031 if (ocount & PBUSY_LOCKED)
3033 if (atomic_cmpset_int(&m->busy_count, ocount, ocount + 1))
3038 if (m->busy_count & PBUSY_LOCKED)
3040 ocount = atomic_fetchadd_int(&m->busy_count, 1);
3041 if (ocount & PBUSY_LOCKED) {
3042 vm_page_sbusy_drop(m);
3050 * Indicate that a clean VM page requires a filesystem commit and cannot
3051 * be reused. Used by tmpfs.
3054 vm_page_need_commit(vm_page_t m)
3056 vm_page_flag_set(m, PG_NEED_COMMIT);
3057 vm_object_set_writeable_dirty(m->object);
3061 vm_page_clear_commit(vm_page_t m)
3063 vm_page_flag_clear(m, PG_NEED_COMMIT);
3067 * Grab a page, blocking if it is busy and allocating a page if necessary.
3068 * A busy page is returned or NULL. The page may or may not be valid and
3069 * might not be on a queue (the caller is responsible for the disposition of
3072 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
3073 * page will be zero'd and marked valid.
3075 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
3076 * valid even if it already exists.
3078 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
3079 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
3080 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
3082 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
3083 * always returned if we had blocked.
3085 * This routine may not be called from an interrupt.
3087 * No other requirements.
3090 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3096 KKASSERT(allocflags &
3097 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
3098 vm_object_hold_shared(object);
3100 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
3102 vm_page_sleep_busy(m, TRUE, "pgrbwt");
3103 if ((allocflags & VM_ALLOC_RETRY) == 0) {
3108 } else if (m == NULL) {
3110 vm_object_upgrade(object);
3113 if (allocflags & VM_ALLOC_RETRY)
3114 allocflags |= VM_ALLOC_NULL_OK;
3115 m = vm_page_alloc(object, pindex,
3116 allocflags & ~VM_ALLOC_RETRY);
3120 if ((allocflags & VM_ALLOC_RETRY) == 0)
3129 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
3131 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
3132 * valid even if already valid.
3134 * NOTE! We have removed all of the PG_ZERO optimizations and also
3135 * removed the idle zeroing code. These optimizations actually
3136 * slow things down on modern cpus because the zerod area is
3137 * likely uncached, placing a memory-access burden on the
3138 * accesors taking the fault.
3140 * By always zeroing the page in-line with the fault, no
3141 * dynamic ram reads are needed and the caches are hot, ready
3142 * for userland to access the memory.
3144 if (m->valid == 0) {
3145 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
3146 pmap_zero_page(VM_PAGE_TO_PHYS(m));
3147 m->valid = VM_PAGE_BITS_ALL;
3149 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
3150 pmap_zero_page(VM_PAGE_TO_PHYS(m));
3151 m->valid = VM_PAGE_BITS_ALL;
3154 vm_object_drop(object);
3159 * Mapping function for valid bits or for dirty bits in
3160 * a page. May not block.
3162 * Inputs are required to range within a page.
3168 vm_page_bits(int base, int size)
3174 base + size <= PAGE_SIZE,
3175 ("vm_page_bits: illegal base/size %d/%d", base, size)
3178 if (size == 0) /* handle degenerate case */
3181 first_bit = base >> DEV_BSHIFT;
3182 last_bit = (base + size - 1) >> DEV_BSHIFT;
3184 return ((2 << last_bit) - (1 << first_bit));
3188 * Sets portions of a page valid and clean. The arguments are expected
3189 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3190 * of any partial chunks touched by the range. The invalid portion of
3191 * such chunks will be zero'd.
3193 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
3194 * align base to DEV_BSIZE so as not to mark clean a partially
3195 * truncated device block. Otherwise the dirty page status might be
3198 * This routine may not block.
3200 * (base + size) must be less then or equal to PAGE_SIZE.
3203 _vm_page_zero_valid(vm_page_t m, int base, int size)
3208 if (size == 0) /* handle degenerate case */
3212 * If the base is not DEV_BSIZE aligned and the valid
3213 * bit is clear, we have to zero out a portion of the
3217 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
3218 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
3220 pmap_zero_page_area(
3228 * If the ending offset is not DEV_BSIZE aligned and the
3229 * valid bit is clear, we have to zero out a portion of
3233 endoff = base + size;
3235 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
3236 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
3238 pmap_zero_page_area(
3241 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
3247 * Set valid, clear dirty bits. If validating the entire
3248 * page we can safely clear the pmap modify bit. We also
3249 * use this opportunity to clear the PG_NOSYNC flag. If a process
3250 * takes a write fault on a MAP_NOSYNC memory area the flag will
3253 * We set valid bits inclusive of any overlap, but we can only
3254 * clear dirty bits for DEV_BSIZE chunks that are fully within
3257 * Page must be busied?
3258 * No other requirements.
3261 vm_page_set_valid(vm_page_t m, int base, int size)
3263 _vm_page_zero_valid(m, base, size);
3264 m->valid |= vm_page_bits(base, size);
3269 * Set valid bits and clear dirty bits.
3271 * Page must be busied by caller.
3273 * NOTE: This function does not clear the pmap modified bit.
3274 * Also note that e.g. NFS may use a byte-granular base
3277 * No other requirements.
3280 vm_page_set_validclean(vm_page_t m, int base, int size)
3284 _vm_page_zero_valid(m, base, size);
3285 pagebits = vm_page_bits(base, size);
3286 m->valid |= pagebits;
3287 m->dirty &= ~pagebits;
3288 if (base == 0 && size == PAGE_SIZE) {
3289 /*pmap_clear_modify(m);*/
3290 vm_page_flag_clear(m, PG_NOSYNC);
3295 * Set valid & dirty. Used by buwrite()
3297 * Page must be busied by caller.
3300 vm_page_set_validdirty(vm_page_t m, int base, int size)
3304 pagebits = vm_page_bits(base, size);
3305 m->valid |= pagebits;
3306 m->dirty |= pagebits;
3308 vm_object_set_writeable_dirty(m->object);
3314 * NOTE: This function does not clear the pmap modified bit.
3315 * Also note that e.g. NFS may use a byte-granular base
3318 * Page must be busied?
3319 * No other requirements.
3322 vm_page_clear_dirty(vm_page_t m, int base, int size)
3324 m->dirty &= ~vm_page_bits(base, size);
3325 if (base == 0 && size == PAGE_SIZE) {
3326 /*pmap_clear_modify(m);*/
3327 vm_page_flag_clear(m, PG_NOSYNC);
3332 * Make the page all-dirty.
3334 * Also make sure the related object and vnode reflect the fact that the
3335 * object may now contain a dirty page.
3337 * Page must be busied?
3338 * No other requirements.
3341 vm_page_dirty(vm_page_t m)
3344 int pqtype = m->queue - m->pc;
3346 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
3347 ("vm_page_dirty: page in free/cache queue!"));
3348 if (m->dirty != VM_PAGE_BITS_ALL) {
3349 m->dirty = VM_PAGE_BITS_ALL;
3351 vm_object_set_writeable_dirty(m->object);
3356 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3357 * valid and dirty bits for the effected areas are cleared.
3359 * Page must be busied?
3361 * No other requirements.
3364 vm_page_set_invalid(vm_page_t m, int base, int size)
3368 bits = vm_page_bits(base, size);
3371 atomic_add_int(&m->object->generation, 1);
3375 * The kernel assumes that the invalid portions of a page contain
3376 * garbage, but such pages can be mapped into memory by user code.
3377 * When this occurs, we must zero out the non-valid portions of the
3378 * page so user code sees what it expects.
3380 * Pages are most often semi-valid when the end of a file is mapped
3381 * into memory and the file's size is not page aligned.
3383 * Page must be busied?
3384 * No other requirements.
3387 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3393 * Scan the valid bits looking for invalid sections that
3394 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
3395 * valid bit may be set ) have already been zerod by
3396 * vm_page_set_validclean().
3398 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3399 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3400 (m->valid & (1 << i))
3403 pmap_zero_page_area(
3406 (i - b) << DEV_BSHIFT
3414 * setvalid is TRUE when we can safely set the zero'd areas
3415 * as being valid. We can do this if there are no cache consistency
3416 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3419 m->valid = VM_PAGE_BITS_ALL;
3423 * Is a (partial) page valid? Note that the case where size == 0
3424 * will return FALSE in the degenerate case where the page is entirely
3425 * invalid, and TRUE otherwise.
3428 * No other requirements.
3431 vm_page_is_valid(vm_page_t m, int base, int size)
3433 int bits = vm_page_bits(base, size);
3435 if (m->valid && ((m->valid & bits) == bits))
3442 * update dirty bits from pmap/mmu. May not block.
3444 * Caller must hold the page busy
3447 vm_page_test_dirty(vm_page_t m)
3449 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
3454 #include "opt_ddb.h"
3456 #include <ddb/ddb.h>
3458 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3460 db_printf("vmstats.v_free_count: %ld\n", vmstats.v_free_count);
3461 db_printf("vmstats.v_cache_count: %ld\n", vmstats.v_cache_count);
3462 db_printf("vmstats.v_inactive_count: %ld\n", vmstats.v_inactive_count);
3463 db_printf("vmstats.v_active_count: %ld\n", vmstats.v_active_count);
3464 db_printf("vmstats.v_wire_count: %ld\n", vmstats.v_wire_count);
3465 db_printf("vmstats.v_free_reserved: %ld\n", vmstats.v_free_reserved);
3466 db_printf("vmstats.v_free_min: %ld\n", vmstats.v_free_min);
3467 db_printf("vmstats.v_free_target: %ld\n", vmstats.v_free_target);
3468 db_printf("vmstats.v_cache_min: %ld\n", vmstats.v_cache_min);
3469 db_printf("vmstats.v_inactive_target: %ld\n",
3470 vmstats.v_inactive_target);
3473 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3476 db_printf("PQ_FREE:");
3477 for (i = 0; i < PQ_L2_SIZE; i++) {
3478 db_printf(" %d", vm_page_queues[PQ_FREE + i].lcnt);
3482 db_printf("PQ_CACHE:");
3483 for(i = 0; i < PQ_L2_SIZE; i++) {
3484 db_printf(" %d", vm_page_queues[PQ_CACHE + i].lcnt);
3488 db_printf("PQ_ACTIVE:");
3489 for(i = 0; i < PQ_L2_SIZE; i++) {
3490 db_printf(" %d", vm_page_queues[PQ_ACTIVE + i].lcnt);
3494 db_printf("PQ_INACTIVE:");
3495 for(i = 0; i < PQ_L2_SIZE; i++) {
3496 db_printf(" %d", vm_page_queues[PQ_INACTIVE + i].lcnt);