2 * Copyright (c) 2003-2019 The DragonFly Project. All rights reserved.
3 * Copyright (c) 1991 Regents of the University of California.
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 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 struct vm_page **vm_page_hash;
136 static u_long vm_dma_reserved = 0;
137 TUNABLE_ULONG("vm.dma_reserved", &vm_dma_reserved);
138 SYSCTL_ULONG(_vm, OID_AUTO, dma_reserved, CTLFLAG_RD, &vm_dma_reserved, 0,
139 "Memory reserved for DMA");
140 SYSCTL_UINT(_vm, OID_AUTO, dma_free_pages, CTLFLAG_RD,
141 &vm_contig_alist.bl_free, 0, "Memory reserved for DMA");
143 static int vm_contig_verbose = 0;
144 TUNABLE_INT("vm.contig_verbose", &vm_contig_verbose);
146 RB_GENERATE2(vm_page_rb_tree, vm_page, rb_entry, rb_vm_page_compare,
147 vm_pindex_t, pindex);
150 vm_page_queue_init(void)
154 for (i = 0; i < PQ_L2_SIZE; i++)
155 vm_page_queues[PQ_FREE+i].cnt_offset =
156 offsetof(struct vmstats, v_free_count);
157 for (i = 0; i < PQ_L2_SIZE; i++)
158 vm_page_queues[PQ_CACHE+i].cnt_offset =
159 offsetof(struct vmstats, v_cache_count);
160 for (i = 0; i < PQ_L2_SIZE; i++)
161 vm_page_queues[PQ_INACTIVE+i].cnt_offset =
162 offsetof(struct vmstats, v_inactive_count);
163 for (i = 0; i < PQ_L2_SIZE; i++)
164 vm_page_queues[PQ_ACTIVE+i].cnt_offset =
165 offsetof(struct vmstats, v_active_count);
166 for (i = 0; i < PQ_L2_SIZE; i++)
167 vm_page_queues[PQ_HOLD+i].cnt_offset =
168 offsetof(struct vmstats, v_active_count);
169 /* PQ_NONE has no queue */
171 for (i = 0; i < PQ_COUNT; i++) {
172 TAILQ_INIT(&vm_page_queues[i].pl);
173 spin_init(&vm_page_queues[i].spin, "vm_page_queue_init");
178 * note: place in initialized data section? Is this necessary?
180 vm_pindex_t first_page = 0;
181 vm_pindex_t vm_page_array_size = 0;
182 vm_page_t vm_page_array = NULL;
183 vm_paddr_t vm_low_phys_reserved;
188 * Sets the page size, perhaps based upon the memory size.
189 * Must be called before any use of page-size dependent functions.
192 vm_set_page_size(void)
194 if (vmstats.v_page_size == 0)
195 vmstats.v_page_size = PAGE_SIZE;
196 if (((vmstats.v_page_size - 1) & vmstats.v_page_size) != 0)
197 panic("vm_set_page_size: page size not a power of two");
203 * Add a new page to the freelist for use by the system. New pages
204 * are added to both the head and tail of the associated free page
205 * queue in a bottom-up fashion, so both zero'd and non-zero'd page
206 * requests pull 'recent' adds (higher physical addresses) first.
208 * Beware that the page zeroing daemon will also be running soon after
209 * boot, moving pages from the head to the tail of the PQ_FREE queues.
211 * Must be called in a critical section.
214 vm_add_new_page(vm_paddr_t pa)
216 struct vpgqueues *vpq;
219 m = PHYS_TO_VM_PAGE(pa);
222 m->pat_mode = PAT_WRITE_BACK;
223 m->pc = (pa >> PAGE_SHIFT);
226 * Twist for cpu localization in addition to page coloring, so
227 * different cpus selecting by m->queue get different page colors.
229 m->pc ^= ((pa >> PAGE_SHIFT) / PQ_L2_SIZE);
230 m->pc ^= ((pa >> PAGE_SHIFT) / (PQ_L2_SIZE * PQ_L2_SIZE));
234 * Reserve a certain number of contiguous low memory pages for
235 * contigmalloc() to use.
237 if (pa < vm_low_phys_reserved) {
238 atomic_add_long(&vmstats.v_page_count, 1);
239 atomic_add_long(&vmstats.v_dma_pages, 1);
242 atomic_add_long(&vmstats.v_wire_count, 1);
243 alist_free(&vm_contig_alist, pa >> PAGE_SHIFT, 1);
250 m->queue = m->pc + PQ_FREE;
251 KKASSERT(m->dirty == 0);
253 atomic_add_long(&vmstats.v_page_count, 1);
254 atomic_add_long(&vmstats.v_free_count, 1);
255 vpq = &vm_page_queues[m->queue];
256 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
263 * Initializes the resident memory module.
265 * Preallocates memory for critical VM structures and arrays prior to
266 * kernel_map becoming available.
268 * Memory is allocated from (virtual2_start, virtual2_end) if available,
269 * otherwise memory is allocated from (virtual_start, virtual_end).
271 * On x86-64 (virtual_start, virtual_end) is only 2GB and may not be
272 * large enough to hold vm_page_array & other structures for machines with
273 * large amounts of ram, so we want to use virtual2* when available.
276 vm_page_startup(void)
278 vm_offset_t vaddr = virtual2_start ? virtual2_start : virtual_start;
281 vm_paddr_t page_range;
287 vm_paddr_t biggestone, biggestsize;
294 vaddr = round_page(vaddr);
297 * Make sure ranges are page-aligned.
299 for (i = 0; phys_avail[i].phys_end; ++i) {
300 phys_avail[i].phys_beg = round_page64(phys_avail[i].phys_beg);
301 phys_avail[i].phys_end = trunc_page64(phys_avail[i].phys_end);
302 if (phys_avail[i].phys_end < phys_avail[i].phys_beg)
303 phys_avail[i].phys_end = phys_avail[i].phys_beg;
307 * Locate largest block
309 for (i = 0; phys_avail[i].phys_end; ++i) {
310 vm_paddr_t size = phys_avail[i].phys_end -
311 phys_avail[i].phys_beg;
313 if (size > biggestsize) {
319 --i; /* adjust to last entry for use down below */
321 end = phys_avail[biggestone].phys_end;
322 end = trunc_page(end);
325 * Initialize the queue headers for the free queue, the active queue
326 * and the inactive queue.
328 vm_page_queue_init();
330 #if !defined(_KERNEL_VIRTUAL)
332 * VKERNELs don't support minidumps and as such don't need
335 * Allocate a bitmap to indicate that a random physical page
336 * needs to be included in a minidump.
338 * The amd64 port needs this to indicate which direct map pages
339 * need to be dumped, via calls to dump_add_page()/dump_drop_page().
341 * However, x86 still needs this workspace internally within the
342 * minidump code. In theory, they are not needed on x86, but are
343 * included should the sf_buf code decide to use them.
345 page_range = phys_avail[i].phys_end / PAGE_SIZE;
346 vm_page_dump_size = round_page(roundup2(page_range, NBBY) / NBBY);
347 end -= vm_page_dump_size;
348 vm_page_dump = (void *)pmap_map(&vaddr, end, end + vm_page_dump_size,
349 VM_PROT_READ | VM_PROT_WRITE);
350 bzero((void *)vm_page_dump, vm_page_dump_size);
353 * Compute the number of pages of memory that will be available for
354 * use (taking into account the overhead of a page structure per
357 first_page = phys_avail[0].phys_beg / PAGE_SIZE;
358 page_range = phys_avail[i].phys_end / PAGE_SIZE - first_page;
359 npages = (total - (page_range * sizeof(struct vm_page))) / PAGE_SIZE;
361 #ifndef _KERNEL_VIRTUAL
363 * (only applies to real kernels)
365 * Reserve a large amount of low memory for potential 32-bit DMA
366 * space allocations. Once device initialization is complete we
367 * release most of it, but keep (vm_dma_reserved) memory reserved
368 * for later use. Typically for X / graphics. Through trial and
369 * error we find that GPUs usually requires ~60-100MB or so.
371 * By default, 128M is left in reserve on machines with 2G+ of ram.
373 vm_low_phys_reserved = (vm_paddr_t)65536 << PAGE_SHIFT;
374 if (vm_low_phys_reserved > total / 4)
375 vm_low_phys_reserved = total / 4;
376 if (vm_dma_reserved == 0) {
377 vm_dma_reserved = 128 * 1024 * 1024; /* 128MB */
378 if (vm_dma_reserved > total / 16)
379 vm_dma_reserved = total / 16;
382 alist_init(&vm_contig_alist, 65536, vm_contig_ameta,
383 ALIST_RECORDS_65536);
386 * Initialize the mem entry structures now, and put them in the free
389 if (bootverbose && ctob(physmem) >= 400LL*1024*1024*1024)
390 kprintf("initializing vm_page_array ");
391 new_end = trunc_page(end - page_range * sizeof(struct vm_page));
392 mapped = pmap_map(&vaddr, new_end, end, VM_PROT_READ | VM_PROT_WRITE);
393 vm_page_array = (vm_page_t)mapped;
395 #if defined(__x86_64__) && !defined(_KERNEL_VIRTUAL)
397 * since pmap_map on amd64 returns stuff out of a direct-map region,
398 * we have to manually add these pages to the minidump tracking so
399 * that they can be dumped, including the vm_page_array.
402 pa < phys_avail[biggestone].phys_end;
409 * Clear all of the page structures, run basic initialization so
410 * PHYS_TO_VM_PAGE() operates properly even on pages not in the
413 bzero((caddr_t) vm_page_array, page_range * sizeof(struct vm_page));
414 vm_page_array_size = page_range;
415 if (bootverbose && ctob(physmem) >= 400LL*1024*1024*1024)
416 kprintf("size = 0x%zx\n", vm_page_array_size);
418 m = &vm_page_array[0];
419 pa = ptoa(first_page);
420 for (i = 0; i < page_range; ++i) {
421 spin_init(&m->spin, "vm_page");
428 * Construct the free queue(s) in ascending order (by physical
429 * address) so that the first 16MB of physical memory is allocated
430 * last rather than first. On large-memory machines, this avoids
431 * the exhaustion of low physical memory before isa_dma_init has run.
433 vmstats.v_page_count = 0;
434 vmstats.v_free_count = 0;
435 for (i = 0; phys_avail[i].phys_end && npages > 0; ++i) {
436 pa = phys_avail[i].phys_beg;
440 last_pa = phys_avail[i].phys_end;
441 while (pa < last_pa && npages-- > 0) {
447 virtual2_start = vaddr;
449 virtual_start = vaddr;
450 mycpu->gd_vmstats = vmstats;
454 * (called from early boot only)
456 * Reorganize VM pages based on numa data. May be called as many times as
457 * necessary. Will reorganize the vm_page_t page color and related queue(s)
458 * to allow vm_page_alloc() to choose pages based on socket affinity.
460 * NOTE: This function is only called while we are still in UP mode, so
461 * we only need a critical section to protect the queues (which
462 * saves a lot of time, there are likely a ton of pages).
465 vm_numa_organize(vm_paddr_t ran_beg, vm_paddr_t bytes, int physid)
470 struct vpgqueues *vpq;
478 * Check if no physical information, or there was only one socket
479 * (so don't waste time doing nothing!).
481 if (cpu_topology_phys_ids <= 1 ||
482 cpu_topology_core_ids == 0) {
487 * Setup for our iteration. Note that ACPI may iterate CPU
488 * sockets starting at 0 or 1 or some other number. The
489 * cpu_topology code mod's it against the socket count.
491 ran_end = ran_beg + bytes;
493 socket_mod = PQ_L2_SIZE / cpu_topology_phys_ids;
494 socket_value = (physid % cpu_topology_phys_ids) * socket_mod;
495 mend = &vm_page_array[vm_page_array_size];
500 * Adjust cpu_topology's phys_mem parameter
503 vm_numa_add_topology_mem(root_cpu_node, physid, (long)bytes);
506 * Adjust vm_page->pc and requeue all affected pages. The
507 * allocator will then be able to localize memory allocations
510 for (i = 0; phys_avail[i].phys_end; ++i) {
511 scan_beg = phys_avail[i].phys_beg;
512 scan_end = phys_avail[i].phys_end;
513 if (scan_end <= ran_beg)
515 if (scan_beg >= ran_end)
517 if (scan_beg < ran_beg)
519 if (scan_end > ran_end)
521 if (atop(scan_end) > first_page + vm_page_array_size)
522 scan_end = ptoa(first_page + vm_page_array_size);
524 m = PHYS_TO_VM_PAGE(scan_beg);
525 while (scan_beg < scan_end) {
527 if (m->queue != PQ_NONE) {
528 vpq = &vm_page_queues[m->queue];
529 TAILQ_REMOVE(&vpq->pl, m, pageq);
531 /* queue doesn't change, no need to adj cnt */
534 m->pc += socket_value;
537 vpq = &vm_page_queues[m->queue];
538 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
540 /* queue doesn't change, no need to adj cnt */
543 m->pc += socket_value;
546 scan_beg += PAGE_SIZE;
555 * (called from early boot only)
557 * Don't allow the NUMA organization to leave vm_page_queues[] nodes
558 * completely empty for a logical cpu. Doing so would force allocations
559 * on that cpu to always borrow from a nearby cpu, create unnecessary
560 * contention, and cause vm_page_alloc() to iterate more queues and run more
563 * This situation can occur when memory sticks are not entirely populated,
564 * populated at different densities, or in naturally assymetric systems
565 * such as the 2990WX. There could very well be many vm_page_queues[]
566 * entries with *NO* pages assigned to them.
568 * Fixing this up ensures that each logical CPU has roughly the same
569 * sized memory pool, and more importantly ensures that logical CPUs
570 * do not wind up with an empty memory pool.
572 * At them moment we just iterate the other queues and borrow pages,
573 * moving them into the queues for cpus with severe deficits even though
574 * the memory might not be local to those cpus. I am not doing this in
575 * a 'smart' way, its effectively UMA style (sorta, since its page-by-page
576 * whereas real UMA typically exchanges address bits 8-10 with high address
577 * bits). But it works extremely well and gives us fairly good deterministic
578 * results on the cpu cores associated with these secondary nodes.
581 vm_numa_organize_finalize(void)
583 struct vpgqueues *vpq;
594 * Machines might not use an exact power of 2 for phys_ids,
595 * core_ids, ht_ids, etc. This can slightly reduce the actual
596 * range of indices in vm_page_queues[] that are nominally used.
598 if (cpu_topology_ht_ids) {
599 scale_lim = PQ_L2_SIZE / cpu_topology_phys_ids;
600 scale_lim = scale_lim / cpu_topology_core_ids;
601 scale_lim = scale_lim / cpu_topology_ht_ids;
602 scale_lim = scale_lim * cpu_topology_ht_ids;
603 scale_lim = scale_lim * cpu_topology_core_ids;
604 scale_lim = scale_lim * cpu_topology_phys_ids;
606 scale_lim = PQ_L2_SIZE;
610 * Calculate an average, set hysteresis for balancing from
611 * 10% below the average to the average.
614 for (i = 0; i < scale_lim; ++i) {
615 lcnt_hi += vm_page_queues[i].lcnt;
617 lcnt_hi /= scale_lim;
618 lcnt_lo = lcnt_hi - lcnt_hi / 10;
620 kprintf("vm_page: avg %ld pages per queue, %d queues\n",
624 for (i = 0; i < scale_lim; ++i) {
625 vpq = &vm_page_queues[PQ_FREE + i];
626 while (vpq->lcnt < lcnt_lo) {
627 struct vpgqueues *vptmp;
629 iter = (iter + 1) & PQ_L2_MASK;
630 vptmp = &vm_page_queues[PQ_FREE + iter];
631 if (vptmp->lcnt < lcnt_hi)
633 m = TAILQ_FIRST(&vptmp->pl);
634 KKASSERT(m->queue == PQ_FREE + iter);
635 TAILQ_REMOVE(&vptmp->pl, m, pageq);
637 /* queue doesn't change, no need to adj cnt */
641 TAILQ_INSERT_HEAD(&vpq->pl, m, pageq);
650 vm_numa_add_topology_mem(cpu_node_t *cpup, int physid, long bytes)
657 cpup->phys_mem += bytes;
661 * All members should have the same chipid, so we only need
662 * to pull out one member.
664 if (CPUMASK_TESTNZERO(cpup->members)) {
665 cpuid = BSFCPUMASK(cpup->members);
667 get_chip_ID_from_APICID(CPUID_TO_APICID(cpuid))) {
668 cpup->phys_mem += bytes;
675 * Just inherit from the parent node
677 cpup->phys_mem = cpup->parent_node->phys_mem;
680 for (i = 0; i < MAXCPU && cpup->child_node[i]; ++i)
681 vm_numa_add_topology_mem(cpup->child_node[i], physid, bytes);
685 * We tended to reserve a ton of memory for contigmalloc(). Now that most
686 * drivers have initialized we want to return most the remaining free
687 * reserve back to the VM page queues so they can be used for normal
690 * We leave vm_dma_reserved bytes worth of free pages in the reserve pool.
693 vm_page_startup_finish(void *dummy __unused)
703 spin_lock(&vm_contig_spin);
705 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
706 if (bfree <= vm_dma_reserved / PAGE_SIZE)
712 * Figure out how much of the initial reserve we have to
713 * free in order to reach our target.
715 bfree -= vm_dma_reserved / PAGE_SIZE;
717 blk += count - bfree;
722 * Calculate the nearest power of 2 <= count.
724 for (xcount = 1; xcount <= count; xcount <<= 1)
727 blk += count - xcount;
731 * Allocate the pages from the alist, then free them to
732 * the normal VM page queues.
734 * Pages allocated from the alist are wired. We have to
735 * busy, unwire, and free them. We must also adjust
736 * vm_low_phys_reserved before freeing any pages to prevent
739 rblk = alist_alloc(&vm_contig_alist, blk, count);
741 kprintf("vm_page_startup_finish: Unable to return "
742 "dma space @0x%08x/%d -> 0x%08x\n",
746 atomic_add_long(&vmstats.v_dma_pages, -(long)count);
747 spin_unlock(&vm_contig_spin);
749 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
750 vm_low_phys_reserved = VM_PAGE_TO_PHYS(m);
752 vm_page_busy_wait(m, FALSE, "cpgfr");
753 vm_page_unwire(m, 0);
758 spin_lock(&vm_contig_spin);
760 spin_unlock(&vm_contig_spin);
763 * Print out how much DMA space drivers have already allocated and
764 * how much is left over.
766 kprintf("DMA space used: %jdk, remaining available: %jdk\n",
767 (intmax_t)(vmstats.v_dma_pages - vm_contig_alist.bl_free) *
769 (intmax_t)vm_contig_alist.bl_free * (PAGE_SIZE / 1024));
772 * hash table for vm_page_lookup_quick()
774 mp = (void *)kmem_alloc3(&kernel_map,
775 vm_page_array_size * sizeof(vm_page_t),
776 VM_SUBSYS_VMPGHASH, KM_CPU(0));
777 bzero(mp, vm_page_array_size * sizeof(vm_page_t));
781 SYSINIT(vm_pgend, SI_SUB_PROC0_POST, SI_ORDER_ANY,
782 vm_page_startup_finish, NULL);
786 * Scan comparison function for Red-Black tree scans. An inclusive
787 * (start,end) is expected. Other fields are not used.
790 rb_vm_page_scancmp(struct vm_page *p, void *data)
792 struct rb_vm_page_scan_info *info = data;
794 if (p->pindex < info->start_pindex)
796 if (p->pindex > info->end_pindex)
802 rb_vm_page_compare(struct vm_page *p1, struct vm_page *p2)
804 if (p1->pindex < p2->pindex)
806 if (p1->pindex > p2->pindex)
812 vm_page_init(vm_page_t m)
814 /* do nothing for now. Called from pmap_page_init() */
818 * Each page queue has its own spin lock, which is fairly optimal for
819 * allocating and freeing pages at least.
821 * The caller must hold the vm_page_spin_lock() before locking a vm_page's
822 * queue spinlock via this function. Also note that m->queue cannot change
823 * unless both the page and queue are locked.
827 _vm_page_queue_spin_lock(vm_page_t m)
832 if (queue != PQ_NONE) {
833 spin_lock(&vm_page_queues[queue].spin);
834 KKASSERT(queue == m->queue);
840 _vm_page_queue_spin_unlock(vm_page_t m)
846 if (queue != PQ_NONE)
847 spin_unlock(&vm_page_queues[queue].spin);
852 _vm_page_queues_spin_lock(u_short queue)
855 if (queue != PQ_NONE)
856 spin_lock(&vm_page_queues[queue].spin);
862 _vm_page_queues_spin_unlock(u_short queue)
865 if (queue != PQ_NONE)
866 spin_unlock(&vm_page_queues[queue].spin);
870 vm_page_queue_spin_lock(vm_page_t m)
872 _vm_page_queue_spin_lock(m);
876 vm_page_queues_spin_lock(u_short queue)
878 _vm_page_queues_spin_lock(queue);
882 vm_page_queue_spin_unlock(vm_page_t m)
884 _vm_page_queue_spin_unlock(m);
888 vm_page_queues_spin_unlock(u_short queue)
890 _vm_page_queues_spin_unlock(queue);
894 * This locks the specified vm_page and its queue in the proper order
895 * (page first, then queue). The queue may change so the caller must
900 _vm_page_and_queue_spin_lock(vm_page_t m)
902 vm_page_spin_lock(m);
903 _vm_page_queue_spin_lock(m);
908 _vm_page_and_queue_spin_unlock(vm_page_t m)
910 _vm_page_queues_spin_unlock(m->queue);
911 vm_page_spin_unlock(m);
915 vm_page_and_queue_spin_unlock(vm_page_t m)
917 _vm_page_and_queue_spin_unlock(m);
921 vm_page_and_queue_spin_lock(vm_page_t m)
923 _vm_page_and_queue_spin_lock(m);
927 * Helper function removes vm_page from its current queue.
928 * Returns the base queue the page used to be on.
930 * The vm_page and the queue must be spinlocked.
931 * This function will unlock the queue but leave the page spinlocked.
933 static __inline u_short
934 _vm_page_rem_queue_spinlocked(vm_page_t m)
936 struct vpgqueues *pq;
942 if (queue != PQ_NONE) {
943 pq = &vm_page_queues[queue];
944 TAILQ_REMOVE(&pq->pl, m, pageq);
947 * Adjust our pcpu stats. In order for the nominal low-memory
948 * algorithms to work properly we don't let any pcpu stat get
949 * too negative before we force it to be rolled-up into the
950 * global stats. Otherwise our pageout and vm_wait tests
953 * The idea here is to reduce unnecessary SMP cache
954 * mastership changes in the global vmstats, which can be
955 * particularly bad in multi-socket systems.
957 cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
958 atomic_add_long(cnt, -1);
959 if (*cnt < -VMMETER_SLOP_COUNT) {
960 u_long copy = atomic_swap_long(cnt, 0);
961 cnt = (long *)((char *)&vmstats + pq->cnt_offset);
962 atomic_add_long(cnt, copy);
963 cnt = (long *)((char *)&mycpu->gd_vmstats +
965 atomic_add_long(cnt, copy);
971 vm_page_queues_spin_unlock(oqueue); /* intended */
977 * Helper function places the vm_page on the specified queue. Generally
978 * speaking only PQ_FREE pages are placed at the head, to allow them to
979 * be allocated sooner rather than later on the assumption that they
982 * The vm_page must be spinlocked.
983 * This function will return with both the page and the queue locked.
986 _vm_page_add_queue_spinlocked(vm_page_t m, u_short queue, int athead)
988 struct vpgqueues *pq;
991 KKASSERT(m->queue == PQ_NONE);
993 if (queue != PQ_NONE) {
994 vm_page_queues_spin_lock(queue);
995 pq = &vm_page_queues[queue];
999 * Adjust our pcpu stats. If a system entity really needs
1000 * to incorporate the count it will call vmstats_rollup()
1001 * to roll it all up into the global vmstats strufture.
1003 cnt = (long *)((char *)&mycpu->gd_vmstats_adj + pq->cnt_offset);
1004 atomic_add_long(cnt, 1);
1007 * PQ_FREE is always handled LIFO style to try to provide
1008 * cache-hot pages to programs.
1011 if (queue - m->pc == PQ_FREE) {
1012 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1013 } else if (athead) {
1014 TAILQ_INSERT_HEAD(&pq->pl, m, pageq);
1016 TAILQ_INSERT_TAIL(&pq->pl, m, pageq);
1018 /* leave the queue spinlocked */
1023 * Wait until page is no longer BUSY. If also_m_busy is TRUE we wait
1024 * until the page is no longer BUSY or SBUSY (busy_count field is 0).
1026 * Returns TRUE if it had to sleep, FALSE if we did not. Only one sleep
1027 * call will be made before returning.
1029 * This function does NOT busy the page and on return the page is not
1030 * guaranteed to be available.
1033 vm_page_sleep_busy(vm_page_t m, int also_m_busy, const char *msg)
1035 u_int32_t busy_count;
1038 busy_count = m->busy_count;
1041 if ((busy_count & PBUSY_LOCKED) == 0 &&
1042 (also_m_busy == 0 || (busy_count & PBUSY_MASK) == 0)) {
1045 tsleep_interlock(m, 0);
1046 if (atomic_cmpset_int(&m->busy_count, busy_count,
1047 busy_count | PBUSY_WANTED)) {
1048 atomic_set_int(&m->flags, PG_REFERENCED);
1049 tsleep(m, PINTERLOCKED, msg, 0);
1056 * This calculates and returns a page color given an optional VM object and
1057 * either a pindex or an iterator. We attempt to return a cpu-localized
1058 * pg_color that is still roughly 16-way set-associative. The CPU topology
1059 * is used if it was probed.
1061 * The caller may use the returned value to index into e.g. PQ_FREE when
1062 * allocating a page in order to nominally obtain pages that are hopefully
1063 * already localized to the requesting cpu. This function is not able to
1064 * provide any sort of guarantee of this, but does its best to improve
1065 * hardware cache management performance.
1067 * WARNING! The caller must mask the returned value with PQ_L2_MASK.
1070 vm_get_pg_color(int cpuid, vm_object_t object, vm_pindex_t pindex)
1073 int object_pg_color;
1076 * WARNING! cpu_topology_core_ids might not be a power of two.
1077 * We also shouldn't make assumptions about
1078 * cpu_topology_phys_ids either.
1080 * WARNING! ncpus might not be known at this time (during early
1081 * boot), and might be set to 1.
1083 * General format: [phys_id][core_id][cpuid][set-associativity]
1084 * (but uses modulo, so not necessarily precise bit masks)
1086 object_pg_color = object ? object->pg_color : 0;
1088 if (cpu_topology_ht_ids) {
1097 * Translate cpuid to socket, core, and hyperthread id.
1099 phys_id = get_cpu_phys_id(cpuid);
1100 core_id = get_cpu_core_id(cpuid);
1101 ht_id = get_cpu_ht_id(cpuid);
1104 * Calculate pg_color for our array index.
1106 * physcale - socket multiplier.
1107 * grpscale - core multiplier (cores per socket)
1108 * cpu* - cpus per core
1110 * WARNING! In early boot, ncpus has not yet been
1111 * initialized and may be set to (1).
1113 * WARNING! physcale must match the organization that
1114 * vm_numa_organize() creates to ensure that
1115 * we properly localize allocations to the
1118 physcale = PQ_L2_SIZE / cpu_topology_phys_ids;
1119 grpscale = physcale / cpu_topology_core_ids;
1120 cpuscale = grpscale / cpu_topology_ht_ids;
1122 pg_color = phys_id * physcale;
1123 pg_color += core_id * grpscale;
1124 pg_color += ht_id * cpuscale;
1125 pg_color += (pindex + object_pg_color) % cpuscale;
1129 pg_color += (pindex + object_pg_color) % grpsize;
1134 /* 3->9, 4->8, 5->10, 6->12, 7->14 */
1139 pg_color += (pindex + object_pg_color) % grpsize;
1144 * Unknown topology, distribute things evenly.
1146 * WARNING! In early boot, ncpus has not yet been
1147 * initialized and may be set to (1).
1151 cpuscale = PQ_L2_SIZE / ncpus;
1153 pg_color = cpuid * cpuscale;
1154 pg_color += (pindex + object_pg_color) % cpuscale;
1156 return (pg_color & PQ_L2_MASK);
1160 * Wait until BUSY can be set, then set it. If also_m_busy is TRUE we
1161 * also wait for m->busy_count to become 0 before setting PBUSY_LOCKED.
1164 VM_PAGE_DEBUG_EXT(vm_page_busy_wait)(vm_page_t m,
1165 int also_m_busy, const char *msg
1168 u_int32_t busy_count;
1171 busy_count = m->busy_count;
1173 if (busy_count & PBUSY_LOCKED) {
1174 tsleep_interlock(m, 0);
1175 if (atomic_cmpset_int(&m->busy_count, busy_count,
1176 busy_count | PBUSY_WANTED)) {
1177 atomic_set_int(&m->flags, PG_REFERENCED);
1178 tsleep(m, PINTERLOCKED, msg, 0);
1180 } else if (also_m_busy && busy_count) {
1181 tsleep_interlock(m, 0);
1182 if (atomic_cmpset_int(&m->busy_count, busy_count,
1183 busy_count | PBUSY_WANTED)) {
1184 atomic_set_int(&m->flags, PG_REFERENCED);
1185 tsleep(m, PINTERLOCKED, msg, 0);
1188 if (atomic_cmpset_int(&m->busy_count, busy_count,
1189 busy_count | PBUSY_LOCKED)) {
1190 #ifdef VM_PAGE_DEBUG
1191 m->busy_func = func;
1192 m->busy_line = lineno;
1201 * Attempt to set BUSY. If also_m_busy is TRUE we only succeed if
1202 * m->busy_count is also 0.
1204 * Returns non-zero on failure.
1207 VM_PAGE_DEBUG_EXT(vm_page_busy_try)(vm_page_t m, int also_m_busy
1210 u_int32_t busy_count;
1213 busy_count = m->busy_count;
1215 if (busy_count & PBUSY_LOCKED)
1217 if (also_m_busy && (busy_count & PBUSY_MASK) != 0)
1219 if (atomic_cmpset_int(&m->busy_count, busy_count,
1220 busy_count | PBUSY_LOCKED)) {
1221 #ifdef VM_PAGE_DEBUG
1222 m->busy_func = func;
1223 m->busy_line = lineno;
1231 * Clear the BUSY flag and return non-zero to indicate to the caller
1232 * that a wakeup() should be performed.
1238 _vm_page_wakeup(vm_page_t m)
1240 u_int32_t busy_count;
1242 busy_count = m->busy_count;
1245 if (atomic_fcmpset_int(&m->busy_count, &busy_count,
1247 ~(PBUSY_LOCKED | PBUSY_WANTED))) {
1248 return((int)(busy_count & PBUSY_WANTED));
1255 * Clear the BUSY flag and wakeup anyone waiting for the page. This
1256 * is typically the last call you make on a page before moving onto
1260 vm_page_wakeup(vm_page_t m)
1262 KASSERT(m->busy_count & PBUSY_LOCKED,
1263 ("vm_page_wakeup: page not busy!!!"));
1264 if (_vm_page_wakeup(m))
1269 * Hold a page, preventing reuse. This is typically only called on pages
1270 * in a known state (either held busy, special, or interlocked in some
1271 * manner). Holding a page does not ensure that it remains valid, it only
1272 * prevents reuse. The page must not already be on the FREE queue or in
1273 * any danger of being moved to the FREE queue concurrent with this call.
1275 * Other parts of the system can still disassociate the page from its object
1276 * and attempt to free it, or perform read or write I/O on it and/or otherwise
1277 * manipulate the page, but if the page is held the VM system will leave the
1278 * page and its data intact and not cycle it through the FREE queue until
1279 * the last hold has been released.
1281 * (see vm_page_wire() if you want to prevent the page from being
1282 * disassociated from its object too).
1285 vm_page_hold(vm_page_t m)
1287 atomic_add_int(&m->hold_count, 1);
1288 KKASSERT(m->queue - m->pc != PQ_FREE);
1290 vm_page_spin_lock(m);
1291 atomic_add_int(&m->hold_count, 1);
1292 if (m->queue - m->pc == PQ_FREE) {
1293 _vm_page_queue_spin_lock(m);
1294 _vm_page_rem_queue_spinlocked(m);
1295 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
1296 _vm_page_queue_spin_unlock(m);
1298 vm_page_spin_unlock(m);
1303 * The opposite of vm_page_hold(). If the page is on the HOLD queue
1304 * it was freed while held and must be moved back to the FREE queue.
1306 * To avoid racing against vm_page_free*() we must re-test conditions
1307 * after obtaining the spin-lock. The initial test can also race a
1308 * vm_page_free*() that is in the middle of moving a page to PQ_HOLD,
1309 * leaving the page on PQ_HOLD with hold_count == 0. Rather than
1310 * throw a spin-lock in the critical path, we rely on the pageout
1311 * daemon to clean-up these loose ends.
1313 * More critically, the 'easy movement' between queues without busying
1314 * a vm_page is only allowed for PQ_FREE<->PQ_HOLD.
1317 vm_page_unhold(vm_page_t m)
1319 KASSERT(m->hold_count > 0 && m->queue - m->pc != PQ_FREE,
1320 ("vm_page_unhold: pg %p illegal hold_count (%d) or "
1321 "on FREE queue (%d)",
1322 m, m->hold_count, m->queue - m->pc));
1324 if (atomic_fetchadd_int(&m->hold_count, -1) == 1 &&
1325 m->queue - m->pc == PQ_HOLD) {
1326 vm_page_spin_lock(m);
1327 if (m->hold_count == 0 && m->queue - m->pc == PQ_HOLD) {
1328 _vm_page_queue_spin_lock(m);
1329 _vm_page_rem_queue_spinlocked(m);
1330 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1);
1331 _vm_page_queue_spin_unlock(m);
1333 vm_page_spin_unlock(m);
1338 * Create a fictitious page with the specified physical address and
1339 * memory attribute. The memory attribute is the only the machine-
1340 * dependent aspect of a fictitious page that must be initialized.
1343 vm_page_initfake(vm_page_t m, vm_paddr_t paddr, vm_memattr_t memattr)
1346 if ((m->flags & PG_FICTITIOUS) != 0) {
1348 * The page's memattr might have changed since the
1349 * previous initialization. Update the pmap to the
1354 m->phys_addr = paddr;
1356 /* Fictitious pages don't use "segind". */
1357 /* Fictitious pages don't use "order" or "pool". */
1358 m->flags = PG_FICTITIOUS | PG_UNMANAGED;
1359 m->busy_count = PBUSY_LOCKED;
1361 spin_init(&m->spin, "fake_page");
1364 pmap_page_set_memattr(m, memattr);
1368 * Inserts the given vm_page into the object and object list.
1370 * The pagetables are not updated but will presumably fault the page
1371 * in if necessary, or if a kernel page the caller will at some point
1372 * enter the page into the kernel's pmap. We are not allowed to block
1373 * here so we *can't* do this anyway.
1375 * This routine may not block.
1376 * This routine must be called with the vm_object held.
1377 * This routine must be called with a critical section held.
1379 * This routine returns TRUE if the page was inserted into the object
1380 * successfully, and FALSE if the page already exists in the object.
1383 vm_page_insert(vm_page_t m, vm_object_t object, vm_pindex_t pindex)
1385 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(object));
1386 if (m->object != NULL)
1387 panic("vm_page_insert: already inserted");
1389 atomic_add_int(&object->generation, 1);
1392 * Record the object/offset pair in this page and add the
1393 * pv_list_count of the page to the object.
1395 * The vm_page spin lock is required for interactions with the pmap.
1397 vm_page_spin_lock(m);
1400 if (vm_page_rb_tree_RB_INSERT(&object->rb_memq, m)) {
1403 vm_page_spin_unlock(m);
1406 ++object->resident_page_count;
1407 ++mycpu->gd_vmtotal.t_rm;
1408 vm_page_spin_unlock(m);
1411 * Since we are inserting a new and possibly dirty page,
1412 * update the object's OBJ_WRITEABLE and OBJ_MIGHTBEDIRTY flags.
1414 if ((m->valid & m->dirty) ||
1415 (m->flags & (PG_WRITEABLE | PG_NEED_COMMIT)))
1416 vm_object_set_writeable_dirty(object);
1419 * Checks for a swap assignment and sets PG_SWAPPED if appropriate.
1421 swap_pager_page_inserted(m);
1426 * Removes the given vm_page_t from the (object,index) table
1428 * The underlying pmap entry (if any) is NOT removed here.
1429 * This routine may not block.
1431 * The page must be BUSY and will remain BUSY on return.
1432 * No other requirements.
1434 * NOTE: FreeBSD side effect was to unbusy the page on return. We leave
1438 vm_page_remove(vm_page_t m)
1442 if (m->object == NULL) {
1446 if ((m->busy_count & PBUSY_LOCKED) == 0)
1447 panic("vm_page_remove: page not busy");
1451 vm_object_hold(object);
1454 * Remove the page from the object and update the object.
1456 * The vm_page spin lock is required for interactions with the pmap.
1458 vm_page_spin_lock(m);
1459 vm_page_rb_tree_RB_REMOVE(&object->rb_memq, m);
1460 --object->resident_page_count;
1461 --mycpu->gd_vmtotal.t_rm;
1463 atomic_add_int(&object->generation, 1);
1464 vm_page_spin_unlock(m);
1466 vm_object_drop(object);
1470 * Calculate the hash position for the vm_page hash heuristic.
1474 vm_page_hash_hash(vm_object_t object, vm_pindex_t pindex)
1478 hi = (uintptr_t)object % (uintptr_t)vm_page_array_size + pindex;
1479 hi %= vm_page_array_size;
1480 return (&vm_page_hash[hi]);
1484 * Heuristical page lookup that does not require any locks. Returns
1485 * a soft-busied page on success, NULL on failure.
1487 * Caller must lookup the page the slow way if NULL is returned.
1490 vm_page_hash_get(vm_object_t object, vm_pindex_t pindex)
1492 struct vm_page **mp;
1495 if (vm_page_hash == NULL)
1497 mp = vm_page_hash_hash(object, pindex);
1502 if (m->object != object || m->pindex != pindex)
1504 if (vm_page_sbusy_try(m))
1506 if (m->object != object || m->pindex != pindex) {
1514 * Enter page onto vm_page_hash[]. This is a heuristic, SMP collisions
1519 vm_page_hash_enter(vm_page_t m)
1521 struct vm_page **mp;
1524 m > &vm_page_array[0] &&
1525 m < &vm_page_array[vm_page_array_size]) {
1526 mp = vm_page_hash_hash(m->object, m->pindex);
1533 * Locate and return the page at (object, pindex), or NULL if the
1534 * page could not be found.
1536 * The caller must hold the vm_object token.
1539 vm_page_lookup(vm_object_t object, vm_pindex_t pindex)
1544 * Search the hash table for this object/offset pair
1546 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1547 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1549 KKASSERT(m->object == object && m->pindex == pindex);
1550 vm_page_hash_enter(m);
1556 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_wait)(struct vm_object *object,
1558 int also_m_busy, const char *msg
1561 u_int32_t busy_count;
1564 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1565 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1567 KKASSERT(m->object == object && m->pindex == pindex);
1568 busy_count = m->busy_count;
1570 if (busy_count & PBUSY_LOCKED) {
1571 tsleep_interlock(m, 0);
1572 if (atomic_cmpset_int(&m->busy_count, busy_count,
1573 busy_count | PBUSY_WANTED)) {
1574 atomic_set_int(&m->flags, PG_REFERENCED);
1575 tsleep(m, PINTERLOCKED, msg, 0);
1576 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1579 } else if (also_m_busy && busy_count) {
1580 tsleep_interlock(m, 0);
1581 if (atomic_cmpset_int(&m->busy_count, busy_count,
1582 busy_count | PBUSY_WANTED)) {
1583 atomic_set_int(&m->flags, PG_REFERENCED);
1584 tsleep(m, PINTERLOCKED, msg, 0);
1585 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq,
1588 } else if (atomic_cmpset_int(&m->busy_count, busy_count,
1589 busy_count | PBUSY_LOCKED)) {
1590 #ifdef VM_PAGE_DEBUG
1591 m->busy_func = func;
1592 m->busy_line = lineno;
1594 vm_page_hash_enter(m);
1602 * Attempt to lookup and busy a page.
1604 * Returns NULL if the page could not be found
1606 * Returns a vm_page and error == TRUE if the page exists but could not
1609 * Returns a vm_page and error == FALSE on success.
1612 VM_PAGE_DEBUG_EXT(vm_page_lookup_busy_try)(struct vm_object *object,
1614 int also_m_busy, int *errorp
1617 u_int32_t busy_count;
1620 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1621 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1624 KKASSERT(m->object == object && m->pindex == pindex);
1625 busy_count = m->busy_count;
1627 if (busy_count & PBUSY_LOCKED) {
1631 if (also_m_busy && busy_count) {
1635 if (atomic_cmpset_int(&m->busy_count, busy_count,
1636 busy_count | PBUSY_LOCKED)) {
1637 #ifdef VM_PAGE_DEBUG
1638 m->busy_func = func;
1639 m->busy_line = lineno;
1641 vm_page_hash_enter(m);
1649 * Returns a page that is only soft-busied for use by the caller in
1650 * a read-only fashion. Returns NULL if the page could not be found,
1651 * the soft busy could not be obtained, or the page data is invalid.
1654 vm_page_lookup_sbusy_try(struct vm_object *object, vm_pindex_t pindex,
1655 int pgoff, int pgbytes)
1659 ASSERT_LWKT_TOKEN_HELD(vm_object_token(object));
1660 m = vm_page_rb_tree_RB_LOOKUP(&object->rb_memq, pindex);
1662 if ((m->valid != VM_PAGE_BITS_ALL &&
1663 !vm_page_is_valid(m, pgoff, pgbytes)) ||
1664 (m->flags & PG_FICTITIOUS)) {
1666 } else if (vm_page_sbusy_try(m)) {
1668 } else if ((m->valid != VM_PAGE_BITS_ALL &&
1669 !vm_page_is_valid(m, pgoff, pgbytes)) ||
1670 (m->flags & PG_FICTITIOUS)) {
1671 vm_page_sbusy_drop(m);
1674 vm_page_hash_enter(m);
1681 * Caller must hold the related vm_object
1684 vm_page_next(vm_page_t m)
1688 next = vm_page_rb_tree_RB_NEXT(m);
1689 if (next && next->pindex != m->pindex + 1)
1697 * Move the given vm_page from its current object to the specified
1698 * target object/offset. The page must be busy and will remain so
1701 * new_object must be held.
1702 * This routine might block. XXX ?
1704 * NOTE: Swap associated with the page must be invalidated by the move. We
1705 * have to do this for several reasons: (1) we aren't freeing the
1706 * page, (2) we are dirtying the page, (3) the VM system is probably
1707 * moving the page from object A to B, and will then later move
1708 * the backing store from A to B and we can't have a conflict.
1710 * NOTE: We *always* dirty the page. It is necessary both for the
1711 * fact that we moved it, and because we may be invalidating
1712 * swap. If the page is on the cache, we have to deactivate it
1713 * or vm_page_dirty() will panic. Dirty pages are not allowed
1717 vm_page_rename(vm_page_t m, vm_object_t new_object, vm_pindex_t new_pindex)
1719 KKASSERT(m->busy_count & PBUSY_LOCKED);
1720 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(new_object));
1722 ASSERT_LWKT_TOKEN_HELD_EXCL(vm_object_token(m->object));
1725 if (vm_page_insert(m, new_object, new_pindex) == FALSE) {
1726 panic("vm_page_rename: target exists (%p,%"PRIu64")",
1727 new_object, new_pindex);
1729 if (m->queue - m->pc == PQ_CACHE)
1730 vm_page_deactivate(m);
1735 * vm_page_unqueue() without any wakeup. This routine is used when a page
1736 * is to remain BUSYied by the caller.
1738 * This routine may not block.
1741 vm_page_unqueue_nowakeup(vm_page_t m)
1743 vm_page_and_queue_spin_lock(m);
1744 (void)_vm_page_rem_queue_spinlocked(m);
1745 vm_page_spin_unlock(m);
1749 * vm_page_unqueue() - Remove a page from its queue, wakeup the pagedemon
1752 * This routine may not block.
1755 vm_page_unqueue(vm_page_t m)
1759 vm_page_and_queue_spin_lock(m);
1760 queue = _vm_page_rem_queue_spinlocked(m);
1761 if (queue == PQ_FREE || queue == PQ_CACHE) {
1762 vm_page_spin_unlock(m);
1763 pagedaemon_wakeup();
1765 vm_page_spin_unlock(m);
1770 * vm_page_list_find()
1772 * Find a page on the specified queue with color optimization.
1774 * The page coloring optimization attempts to locate a page that does
1775 * not overload other nearby pages in the object in the cpu's L1 or L2
1776 * caches. We need this optimization because cpu caches tend to be
1777 * physical caches, while object spaces tend to be virtual.
1779 * The page coloring optimization also, very importantly, tries to localize
1780 * memory to cpus and physical sockets.
1782 * Each PQ_FREE and PQ_CACHE color queue has its own spinlock and the
1783 * algorithm is adjusted to localize allocations on a per-core basis.
1784 * This is done by 'twisting' the colors.
1786 * The page is returned spinlocked and removed from its queue (it will
1787 * be on PQ_NONE), or NULL. The page is not BUSY'd. The caller
1788 * is responsible for dealing with the busy-page case (usually by
1789 * deactivating the page and looping).
1791 * NOTE: This routine is carefully inlined. A non-inlined version
1792 * is available for outside callers but the only critical path is
1793 * from within this source file.
1795 * NOTE: This routine assumes that the vm_pages found in PQ_CACHE and PQ_FREE
1796 * represent stable storage, allowing us to order our locks vm_page
1797 * first, then queue.
1801 _vm_page_list_find(int basequeue, int index)
1803 struct vpgqueues *pq;
1806 index &= PQ_L2_MASK;
1807 pq = &vm_page_queues[basequeue + index];
1810 * Try this cpu's colored queue first. Test for a page unlocked,
1811 * then lock the queue and locate a page. Note that the lock order
1812 * is reversed, but we do not want to dwadle on the page spinlock
1813 * anyway as it is held significantly longer than the queue spinlock.
1815 if (TAILQ_FIRST(&pq->pl)) {
1816 spin_lock(&pq->spin);
1817 TAILQ_FOREACH(m, &pq->pl, pageq) {
1818 if (spin_trylock(&m->spin) == 0)
1820 KKASSERT(m->queue == basequeue + index);
1821 _vm_page_rem_queue_spinlocked(m);
1824 spin_unlock(&pq->spin);
1828 * If we are unable to get a page, do a more involved NUMA-aware
1831 m = _vm_page_list_find2(basequeue, index);
1836 * If we could not find the page in the desired queue try to find it in
1837 * a nearby (NUMA-aware) queue.
1840 _vm_page_list_find2(int basequeue, int index)
1842 struct vpgqueues *pq;
1844 int pqmask = PQ_SET_ASSOC_MASK >> 1;
1848 index &= PQ_L2_MASK;
1849 pq = &vm_page_queues[basequeue];
1852 * Run local sets of 16, 32, 64, 128, and the whole queue if all
1853 * else fails (PQ_L2_MASK which is 255).
1855 * Test each queue unlocked first, then lock the queue and locate
1856 * a page. Note that the lock order is reversed, but we do not want
1857 * to dwadle on the page spinlock anyway as it is held significantly
1858 * longer than the queue spinlock.
1861 pqmask = (pqmask << 1) | 1;
1862 for (i = 0; i <= pqmask; ++i) {
1863 pqi = (index & ~pqmask) | ((index + i) & pqmask);
1864 if (TAILQ_FIRST(&pq[pqi].pl)) {
1865 spin_lock(&pq[pqi].spin);
1866 TAILQ_FOREACH(m, &pq[pqi].pl, pageq) {
1867 if (spin_trylock(&m->spin) == 0)
1869 KKASSERT(m->queue == basequeue + pqi);
1870 _vm_page_rem_queue_spinlocked(m);
1873 spin_unlock(&pq[pqi].spin);
1876 } while (pqmask != PQ_L2_MASK);
1882 * Returns a vm_page candidate for allocation. The page is not busied so
1883 * it can move around. The caller must busy the page (and typically
1884 * deactivate it if it cannot be busied!)
1886 * Returns a spinlocked vm_page that has been removed from its queue.
1889 vm_page_list_find(int basequeue, int index)
1891 return(_vm_page_list_find(basequeue, index));
1895 * Find a page on the cache queue with color optimization, remove it
1896 * from the queue, and busy it. The returned page will not be spinlocked.
1898 * A candidate failure will be deactivated. Candidates can fail due to
1899 * being busied by someone else, in which case they will be deactivated.
1901 * This routine may not block.
1905 vm_page_select_cache(u_short pg_color)
1910 m = _vm_page_list_find(PQ_CACHE, pg_color & PQ_L2_MASK);
1914 * (m) has been removed from its queue and spinlocked
1916 if (vm_page_busy_try(m, TRUE)) {
1917 _vm_page_deactivate_locked(m, 0);
1918 vm_page_spin_unlock(m);
1921 * We successfully busied the page
1923 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) == 0 &&
1924 m->hold_count == 0 &&
1925 m->wire_count == 0 &&
1926 (m->dirty & m->valid) == 0) {
1927 vm_page_spin_unlock(m);
1928 pagedaemon_wakeup();
1933 * The page cannot be recycled, deactivate it.
1935 _vm_page_deactivate_locked(m, 0);
1936 if (_vm_page_wakeup(m)) {
1937 vm_page_spin_unlock(m);
1940 vm_page_spin_unlock(m);
1948 * Find a free page. We attempt to inline the nominal case and fall back
1949 * to _vm_page_select_free() otherwise. A busied page is removed from
1950 * the queue and returned.
1952 * This routine may not block.
1954 static __inline vm_page_t
1955 vm_page_select_free(u_short pg_color)
1960 m = _vm_page_list_find(PQ_FREE, pg_color & PQ_L2_MASK);
1963 if (vm_page_busy_try(m, TRUE)) {
1965 * Various mechanisms such as a pmap_collect can
1966 * result in a busy page on the free queue. We
1967 * have to move the page out of the way so we can
1968 * retry the allocation. If the other thread is not
1969 * allocating the page then m->valid will remain 0 and
1970 * the pageout daemon will free the page later on.
1972 * Since we could not busy the page, however, we
1973 * cannot make assumptions as to whether the page
1974 * will be allocated by the other thread or not,
1975 * so all we can do is deactivate it to move it out
1976 * of the way. In particular, if the other thread
1977 * wires the page it may wind up on the inactive
1978 * queue and the pageout daemon will have to deal
1979 * with that case too.
1981 _vm_page_deactivate_locked(m, 0);
1982 vm_page_spin_unlock(m);
1985 * Theoretically if we are able to busy the page
1986 * atomic with the queue removal (using the vm_page
1987 * lock) nobody else should have been able to mess
1988 * with the page before us.
1990 * Assert the page state. Note that even though
1991 * wiring doesn't adjust queues, a page on the free
1992 * queue should never be wired at this point.
1994 KKASSERT((m->flags & (PG_UNMANAGED |
1995 PG_NEED_COMMIT)) == 0);
1996 KASSERT(m->hold_count == 0,
1997 ("m->hold_count is not zero "
1998 "pg %p q=%d flags=%08x hold=%d wire=%d",
1999 m, m->queue, m->flags,
2000 m->hold_count, m->wire_count));
2001 KKASSERT(m->wire_count == 0);
2002 vm_page_spin_unlock(m);
2003 pagedaemon_wakeup();
2005 /* return busied and removed page */
2015 * Allocate and return a memory cell associated with this VM object/offset
2016 * pair. If object is NULL an unassociated page will be allocated.
2018 * The returned page will be busied and removed from its queues. This
2019 * routine can block and may return NULL if a race occurs and the page
2020 * is found to already exist at the specified (object, pindex).
2022 * VM_ALLOC_NORMAL allow use of cache pages, nominal free drain
2023 * VM_ALLOC_QUICK like normal but cannot use cache
2024 * VM_ALLOC_SYSTEM greater free drain
2025 * VM_ALLOC_INTERRUPT allow free list to be completely drained
2026 * VM_ALLOC_ZERO advisory request for pre-zero'd page only
2027 * VM_ALLOC_FORCE_ZERO advisory request for pre-zero'd page only
2028 * VM_ALLOC_NULL_OK ok to return NULL on insertion collision
2029 * (see vm_page_grab())
2030 * VM_ALLOC_USE_GD ok to use per-gd cache
2032 * VM_ALLOC_CPU(n) allocate using specified cpu localization
2034 * The object must be held if not NULL
2035 * This routine may not block
2037 * Additional special handling is required when called from an interrupt
2038 * (VM_ALLOC_INTERRUPT). We are not allowed to mess with the page cache
2042 vm_page_alloc(vm_object_t object, vm_pindex_t pindex, int page_req)
2052 * Special per-cpu free VM page cache. The pages are pre-busied
2053 * and pre-zerod for us.
2055 if (gd->gd_vmpg_count && (page_req & VM_ALLOC_USE_GD)) {
2057 if (gd->gd_vmpg_count) {
2058 m = gd->gd_vmpg_array[--gd->gd_vmpg_count];
2070 * CPU localization algorithm. Break the page queues up by physical
2071 * id and core id (note that two cpu threads will have the same core
2072 * id, and core_id != gd_cpuid).
2074 * This is nowhere near perfect, for example the last pindex in a
2075 * subgroup will overflow into the next cpu or package. But this
2076 * should get us good page reuse locality in heavy mixed loads.
2078 * (may be executed before the APs are started, so other GDs might
2081 if (page_req & VM_ALLOC_CPU_SPEC)
2082 cpuid_local = VM_ALLOC_GETCPU(page_req);
2084 cpuid_local = mycpu->gd_cpuid;
2086 pg_color = vm_get_pg_color(cpuid_local, object, pindex);
2089 (VM_ALLOC_NORMAL|VM_ALLOC_QUICK|
2090 VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
2093 * Certain system threads (pageout daemon, buf_daemon's) are
2094 * allowed to eat deeper into the free page list.
2096 if (curthread->td_flags & TDF_SYSTHREAD)
2097 page_req |= VM_ALLOC_SYSTEM;
2100 * Impose various limitations. Note that the v_free_reserved test
2101 * must match the opposite of vm_page_count_target() to avoid
2102 * livelocks, be careful.
2106 if (gd->gd_vmstats.v_free_count >= gd->gd_vmstats.v_free_reserved ||
2107 ((page_req & VM_ALLOC_INTERRUPT) &&
2108 gd->gd_vmstats.v_free_count > 0) ||
2109 ((page_req & VM_ALLOC_SYSTEM) &&
2110 gd->gd_vmstats.v_cache_count == 0 &&
2111 gd->gd_vmstats.v_free_count >
2112 gd->gd_vmstats.v_interrupt_free_min)
2115 * The free queue has sufficient free pages to take one out.
2117 m = vm_page_select_free(pg_color);
2118 } else if (page_req & VM_ALLOC_NORMAL) {
2120 * Allocatable from the cache (non-interrupt only). On
2121 * success, we must free the page and try again, thus
2122 * ensuring that vmstats.v_*_free_min counters are replenished.
2125 if (curthread->td_preempted) {
2126 kprintf("vm_page_alloc(): warning, attempt to allocate"
2127 " cache page from preempting interrupt\n");
2130 m = vm_page_select_cache(pg_color);
2133 m = vm_page_select_cache(pg_color);
2136 * On success move the page into the free queue and loop.
2138 * Only do this if we can safely acquire the vm_object lock,
2139 * because this is effectively a random page and the caller
2140 * might be holding the lock shared, we don't want to
2144 KASSERT(m->dirty == 0,
2145 ("Found dirty cache page %p", m));
2146 if ((obj = m->object) != NULL) {
2147 if (vm_object_hold_try(obj)) {
2148 vm_page_protect(m, VM_PROT_NONE);
2150 /* m->object NULL here */
2151 vm_object_drop(obj);
2153 vm_page_deactivate(m);
2157 vm_page_protect(m, VM_PROT_NONE);
2164 * On failure return NULL
2166 atomic_add_int(&vm_pageout_deficit, 1);
2167 pagedaemon_wakeup();
2171 * No pages available, wakeup the pageout daemon and give up.
2173 atomic_add_int(&vm_pageout_deficit, 1);
2174 pagedaemon_wakeup();
2179 * v_free_count can race so loop if we don't find the expected
2188 * Good page found. The page has already been busied for us and
2189 * removed from its queues.
2191 KASSERT(m->dirty == 0,
2192 ("vm_page_alloc: free/cache page %p was dirty", m));
2193 KKASSERT(m->queue == PQ_NONE);
2199 * Initialize the structure, inheriting some flags but clearing
2200 * all the rest. The page has already been busied for us.
2202 vm_page_flag_clear(m, ~PG_KEEP_NEWPAGE_MASK);
2204 KKASSERT(m->wire_count == 0);
2205 KKASSERT((m->busy_count & PBUSY_MASK) == 0);
2210 * Caller must be holding the object lock (asserted by
2211 * vm_page_insert()).
2213 * NOTE: Inserting a page here does not insert it into any pmaps
2214 * (which could cause us to block allocating memory).
2216 * NOTE: If no object an unassociated page is allocated, m->pindex
2217 * can be used by the caller for any purpose.
2220 if (vm_page_insert(m, object, pindex) == FALSE) {
2222 if ((page_req & VM_ALLOC_NULL_OK) == 0)
2223 panic("PAGE RACE %p[%ld]/%p",
2224 object, (long)pindex, m);
2232 * Don't wakeup too often - wakeup the pageout daemon when
2233 * we would be nearly out of memory.
2235 pagedaemon_wakeup();
2238 * A BUSY page is returned.
2244 * Returns number of pages available in our DMA memory reserve
2245 * (adjusted with vm.dma_reserved=<value>m in /boot/loader.conf)
2248 vm_contig_avail_pages(void)
2253 spin_lock(&vm_contig_spin);
2254 bfree = alist_free_info(&vm_contig_alist, &blk, &count);
2255 spin_unlock(&vm_contig_spin);
2261 * Attempt to allocate contiguous physical memory with the specified
2265 vm_page_alloc_contig(vm_paddr_t low, vm_paddr_t high,
2266 unsigned long alignment, unsigned long boundary,
2267 unsigned long size, vm_memattr_t memattr)
2273 static vm_pindex_t contig_rover;
2276 alignment >>= PAGE_SHIFT;
2279 boundary >>= PAGE_SHIFT;
2282 size = (size + PAGE_MASK) >> PAGE_SHIFT;
2286 * Disabled temporarily until we find a solution for DRM (a flag
2287 * to always use the free space reserve, for performance).
2289 if (high == BUS_SPACE_MAXADDR && alignment <= PAGE_SIZE &&
2290 boundary <= PAGE_SIZE && size == 1 &&
2291 memattr == VM_MEMATTR_DEFAULT) {
2293 * Any page will work, use vm_page_alloc()
2294 * (e.g. when used from kmem_alloc_attr())
2296 m = vm_page_alloc(NULL, (contig_rover++) & 0x7FFFFFFF,
2297 VM_ALLOC_NORMAL | VM_ALLOC_SYSTEM |
2298 VM_ALLOC_INTERRUPT);
2299 m->valid = VM_PAGE_BITS_ALL;
2306 * Use the low-memory dma reserve
2308 spin_lock(&vm_contig_spin);
2309 blk = alist_alloc(&vm_contig_alist, 0, size);
2310 if (blk == ALIST_BLOCK_NONE) {
2311 spin_unlock(&vm_contig_spin);
2313 kprintf("vm_page_alloc_contig: %ldk nospace\n",
2314 (size << PAGE_SHIFT) / 1024);
2319 if (high && ((vm_paddr_t)(blk + size) << PAGE_SHIFT) > high) {
2320 alist_free(&vm_contig_alist, blk, size);
2321 spin_unlock(&vm_contig_spin);
2323 kprintf("vm_page_alloc_contig: %ldk high "
2325 (size << PAGE_SHIFT) / 1024,
2330 spin_unlock(&vm_contig_spin);
2331 m = PHYS_TO_VM_PAGE((vm_paddr_t)blk << PAGE_SHIFT);
2333 if (vm_contig_verbose) {
2334 kprintf("vm_page_alloc_contig: %016jx/%ldk "
2335 "(%016jx-%016jx al=%lu bo=%lu pgs=%lu attr=%d\n",
2336 (intmax_t)m->phys_addr,
2337 (size << PAGE_SHIFT) / 1024,
2338 low, high, alignment, boundary, size, memattr);
2340 if (memattr != VM_MEMATTR_DEFAULT) {
2341 for (i = 0;i < size; i++)
2342 pmap_page_set_memattr(&m[i], memattr);
2348 * Free contiguously allocated pages. The pages will be wired but not busy.
2349 * When freeing to the alist we leave them wired and not busy.
2352 vm_page_free_contig(vm_page_t m, unsigned long size)
2354 vm_paddr_t pa = VM_PAGE_TO_PHYS(m);
2355 vm_pindex_t start = pa >> PAGE_SHIFT;
2356 vm_pindex_t pages = (size + PAGE_MASK) >> PAGE_SHIFT;
2358 if (vm_contig_verbose) {
2359 kprintf("vm_page_free_contig: %016jx/%ldk\n",
2360 (intmax_t)pa, size / 1024);
2362 if (pa < vm_low_phys_reserved) {
2363 KKASSERT(pa + size <= vm_low_phys_reserved);
2364 spin_lock(&vm_contig_spin);
2365 alist_free(&vm_contig_alist, start, pages);
2366 spin_unlock(&vm_contig_spin);
2369 vm_page_busy_wait(m, FALSE, "cpgfr");
2370 vm_page_unwire(m, 0);
2381 * Wait for sufficient free memory for nominal heavy memory use kernel
2384 * WARNING! Be sure never to call this in any vm_pageout code path, which
2385 * will trivially deadlock the system.
2388 vm_wait_nominal(void)
2390 while (vm_page_count_min(0))
2395 * Test if vm_wait_nominal() would block.
2398 vm_test_nominal(void)
2400 if (vm_page_count_min(0))
2406 * Block until free pages are available for allocation, called in various
2407 * places before memory allocations.
2409 * The caller may loop if vm_page_count_min() == FALSE so we cannot be
2410 * more generous then that.
2416 * never wait forever
2420 lwkt_gettoken(&vm_token);
2422 if (curthread == pagethread ||
2423 curthread == emergpager) {
2425 * The pageout daemon itself needs pages, this is bad.
2427 if (vm_page_count_min(0)) {
2428 vm_pageout_pages_needed = 1;
2429 tsleep(&vm_pageout_pages_needed, 0, "VMWait", timo);
2433 * Wakeup the pageout daemon if necessary and wait.
2435 * Do not wait indefinitely for the target to be reached,
2436 * as load might prevent it from being reached any time soon.
2437 * But wait a little to try to slow down page allocations
2438 * and to give more important threads (the pagedaemon)
2439 * allocation priority.
2441 if (vm_page_count_target()) {
2442 if (vm_pages_needed == 0) {
2443 vm_pages_needed = 1;
2444 wakeup(&vm_pages_needed);
2446 ++vm_pages_waiting; /* SMP race ok */
2447 tsleep(&vmstats.v_free_count, 0, "vmwait", timo);
2450 lwkt_reltoken(&vm_token);
2454 * Block until free pages are available for allocation
2456 * Called only from vm_fault so that processes page faulting can be
2460 vm_wait_pfault(void)
2463 * Wakeup the pageout daemon if necessary and wait.
2465 * Do not wait indefinitely for the target to be reached,
2466 * as load might prevent it from being reached any time soon.
2467 * But wait a little to try to slow down page allocations
2468 * and to give more important threads (the pagedaemon)
2469 * allocation priority.
2471 if (vm_page_count_min(0)) {
2472 lwkt_gettoken(&vm_token);
2473 while (vm_page_count_severe()) {
2474 if (vm_page_count_target()) {
2477 if (vm_pages_needed == 0) {
2478 vm_pages_needed = 1;
2479 wakeup(&vm_pages_needed);
2481 ++vm_pages_waiting; /* SMP race ok */
2482 tsleep(&vmstats.v_free_count, 0, "pfault", hz);
2485 * Do not stay stuck in the loop if the system is trying
2486 * to kill the process.
2489 if (td->td_proc && (td->td_proc->p_flags & P_LOWMEMKILL))
2493 lwkt_reltoken(&vm_token);
2498 * Put the specified page on the active list (if appropriate). Ensure
2499 * that act_count is at least ACT_INIT but do not otherwise mess with it.
2501 * The caller should be holding the page busied ? XXX
2502 * This routine may not block.
2505 vm_page_activate(vm_page_t m)
2509 vm_page_spin_lock(m);
2510 if (m->queue - m->pc != PQ_ACTIVE) {
2511 _vm_page_queue_spin_lock(m);
2512 oqueue = _vm_page_rem_queue_spinlocked(m);
2513 /* page is left spinlocked, queue is unlocked */
2515 if (oqueue == PQ_CACHE)
2516 mycpu->gd_cnt.v_reactivated++;
2517 if ((m->flags & PG_UNMANAGED) == 0) {
2518 if (m->act_count < ACT_INIT)
2519 m->act_count = ACT_INIT;
2520 _vm_page_add_queue_spinlocked(m, PQ_ACTIVE + m->pc, 0);
2522 _vm_page_and_queue_spin_unlock(m);
2523 if (oqueue == PQ_CACHE || oqueue == PQ_FREE)
2524 pagedaemon_wakeup();
2526 if (m->act_count < ACT_INIT)
2527 m->act_count = ACT_INIT;
2528 vm_page_spin_unlock(m);
2533 * Helper routine for vm_page_free_toq() and vm_page_cache(). This
2534 * routine is called when a page has been added to the cache or free
2537 * This routine may not block.
2539 static __inline void
2540 vm_page_free_wakeup(void)
2542 globaldata_t gd = mycpu;
2545 * If the pageout daemon itself needs pages, then tell it that
2546 * there are some free.
2548 if (vm_pageout_pages_needed &&
2549 gd->gd_vmstats.v_cache_count + gd->gd_vmstats.v_free_count >=
2550 gd->gd_vmstats.v_pageout_free_min
2552 vm_pageout_pages_needed = 0;
2553 wakeup(&vm_pageout_pages_needed);
2557 * Wakeup processes that are waiting on memory.
2559 * Generally speaking we want to wakeup stuck processes as soon as
2560 * possible. !vm_page_count_min(0) is the absolute minimum point
2561 * where we can do this. Wait a bit longer to reduce degenerate
2562 * re-blocking (vm_page_free_hysteresis). The target check is just
2563 * to make sure the min-check w/hysteresis does not exceed the
2566 if (vm_pages_waiting) {
2567 if (!vm_page_count_min(vm_page_free_hysteresis) ||
2568 !vm_page_count_target()) {
2569 vm_pages_waiting = 0;
2570 wakeup(&vmstats.v_free_count);
2571 ++mycpu->gd_cnt.v_ppwakeups;
2574 if (!vm_page_count_target()) {
2576 * Plenty of pages are free, wakeup everyone.
2578 vm_pages_waiting = 0;
2579 wakeup(&vmstats.v_free_count);
2580 ++mycpu->gd_cnt.v_ppwakeups;
2581 } else if (!vm_page_count_min(0)) {
2583 * Some pages are free, wakeup someone.
2585 int wcount = vm_pages_waiting;
2588 vm_pages_waiting = wcount;
2589 wakeup_one(&vmstats.v_free_count);
2590 ++mycpu->gd_cnt.v_ppwakeups;
2597 * Returns the given page to the PQ_FREE or PQ_HOLD list and disassociates
2598 * it from its VM object.
2600 * The vm_page must be BUSY on entry. BUSY will be released on
2601 * return (the page will have been freed).
2604 vm_page_free_toq(vm_page_t m)
2606 mycpu->gd_cnt.v_tfree++;
2607 KKASSERT((m->flags & PG_MAPPED) == 0);
2608 KKASSERT(m->busy_count & PBUSY_LOCKED);
2610 if ((m->busy_count & PBUSY_MASK) || ((m->queue - m->pc) == PQ_FREE)) {
2611 kprintf("vm_page_free: pindex(%lu), busy %08x, "
2613 (u_long)m->pindex, m->busy_count, m->hold_count);
2614 if ((m->queue - m->pc) == PQ_FREE)
2615 panic("vm_page_free: freeing free page");
2617 panic("vm_page_free: freeing busy page");
2621 * Remove from object, spinlock the page and its queues and
2622 * remove from any queue. No queue spinlock will be held
2623 * after this section (because the page was removed from any
2627 vm_page_and_queue_spin_lock(m);
2628 _vm_page_rem_queue_spinlocked(m);
2631 * No further management of fictitious pages occurs beyond object
2632 * and queue removal.
2634 if ((m->flags & PG_FICTITIOUS) != 0) {
2635 vm_page_spin_unlock(m);
2643 if (m->wire_count != 0) {
2644 if (m->wire_count > 1) {
2646 "vm_page_free: invalid wire count (%d), pindex: 0x%lx",
2647 m->wire_count, (long)m->pindex);
2649 panic("vm_page_free: freeing wired page");
2653 * Clear the UNMANAGED flag when freeing an unmanaged page.
2654 * Clear the NEED_COMMIT flag
2656 if (m->flags & PG_UNMANAGED)
2657 vm_page_flag_clear(m, PG_UNMANAGED);
2658 if (m->flags & PG_NEED_COMMIT)
2659 vm_page_flag_clear(m, PG_NEED_COMMIT);
2661 if (m->hold_count != 0) {
2662 _vm_page_add_queue_spinlocked(m, PQ_HOLD + m->pc, 0);
2664 _vm_page_add_queue_spinlocked(m, PQ_FREE + m->pc, 1);
2668 * This sequence allows us to clear BUSY while still holding
2669 * its spin lock, which reduces contention vs allocators. We
2670 * must not leave the queue locked or _vm_page_wakeup() may
2673 _vm_page_queue_spin_unlock(m);
2674 if (_vm_page_wakeup(m)) {
2675 vm_page_spin_unlock(m);
2678 vm_page_spin_unlock(m);
2680 vm_page_free_wakeup();
2684 * vm_page_unmanage()
2686 * Prevent PV management from being done on the page. The page is
2687 * also removed from the paging queues, and as a consequence of no longer
2688 * being managed the pageout daemon will not touch it (since there is no
2689 * way to locate the pte mappings for the page). madvise() calls that
2690 * mess with the pmap will also no longer operate on the page.
2692 * Beyond that the page is still reasonably 'normal'. Freeing the page
2693 * will clear the flag.
2695 * This routine is used by OBJT_PHYS objects - objects using unswappable
2696 * physical memory as backing store rather then swap-backed memory and
2697 * will eventually be extended to support 4MB unmanaged physical
2700 * Caller must be holding the page busy.
2703 vm_page_unmanage(vm_page_t m)
2705 KKASSERT(m->busy_count & PBUSY_LOCKED);
2706 if ((m->flags & PG_UNMANAGED) == 0) {
2709 vm_page_flag_set(m, PG_UNMANAGED);
2713 * Mark this page as wired down by yet another map. We do not adjust the
2714 * queue the page is on, it will be checked for wiring as-needed.
2716 * Caller must be holding the page busy.
2719 vm_page_wire(vm_page_t m)
2722 * Only bump the wire statistics if the page is not already wired,
2723 * and only unqueue the page if it is on some queue (if it is unmanaged
2724 * it is already off the queues). Don't do anything with fictitious
2725 * pages because they are always wired.
2727 KKASSERT(m->busy_count & PBUSY_LOCKED);
2728 if ((m->flags & PG_FICTITIOUS) == 0) {
2729 if (atomic_fetchadd_int(&m->wire_count, 1) == 0) {
2730 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count, 1);
2732 KASSERT(m->wire_count != 0,
2733 ("vm_page_wire: wire_count overflow m=%p", m));
2738 * Release one wiring of this page, potentially enabling it to be paged again.
2740 * Note that wired pages are no longer unconditionally removed from the
2741 * paging queues, so the page may already be on a queue. Move the page
2742 * to the desired queue if necessary.
2744 * Many pages placed on the inactive queue should actually go
2745 * into the cache, but it is difficult to figure out which. What
2746 * we do instead, if the inactive target is well met, is to put
2747 * clean pages at the head of the inactive queue instead of the tail.
2748 * This will cause them to be moved to the cache more quickly and
2749 * if not actively re-referenced, freed more quickly. If we just
2750 * stick these pages at the end of the inactive queue, heavy filesystem
2751 * meta-data accesses can cause an unnecessary paging load on memory bound
2752 * processes. This optimization causes one-time-use metadata to be
2753 * reused more quickly.
2755 * Pages marked PG_NEED_COMMIT are always activated and never placed on
2756 * the inactive queue. This helps the pageout daemon determine memory
2757 * pressure and act on out-of-memory situations more quickly.
2759 * BUT, if we are in a low-memory situation we have no choice but to
2760 * put clean pages on the cache queue.
2762 * A number of routines use vm_page_unwire() to guarantee that the page
2763 * will go into either the inactive or active queues, and will NEVER
2764 * be placed in the cache - for example, just after dirtying a page.
2765 * dirty pages in the cache are not allowed.
2767 * This routine may not block.
2770 vm_page_unwire(vm_page_t m, int activate)
2772 KKASSERT(m->busy_count & PBUSY_LOCKED);
2773 if (m->flags & PG_FICTITIOUS) {
2775 } else if ((int)m->wire_count <= 0) {
2776 panic("vm_page_unwire: invalid wire count: %d", m->wire_count);
2778 if (atomic_fetchadd_int(&m->wire_count, -1) == 1) {
2779 atomic_add_long(&mycpu->gd_vmstats_adj.v_wire_count,-1);
2780 if (m->flags & PG_UNMANAGED) {
2782 } else if (activate || (m->flags & PG_NEED_COMMIT)) {
2783 vm_page_activate(m);
2785 vm_page_spin_lock(m);
2786 _vm_page_add_queue_spinlocked(m,
2787 PQ_ACTIVE + m->pc, 0);
2788 _vm_page_and_queue_spin_unlock(m);
2791 vm_page_deactivate(m);
2793 vm_page_spin_lock(m);
2794 vm_page_flag_clear(m, PG_WINATCFLS);
2795 _vm_page_add_queue_spinlocked(m,
2796 PQ_INACTIVE + m->pc, 0);
2797 _vm_page_and_queue_spin_unlock(m);
2805 * Move the specified page to the inactive queue.
2807 * Normally athead is 0 resulting in LRU operation. athead is set
2808 * to 1 if we want this page to be 'as if it were placed in the cache',
2809 * except without unmapping it from the process address space.
2811 * vm_page's spinlock must be held on entry and will remain held on return.
2812 * This routine may not block. The caller does not have to hold the page
2813 * busied but should have some sort of interlock on its validity.
2816 _vm_page_deactivate_locked(vm_page_t m, int athead)
2821 * Ignore if already inactive.
2823 if (m->queue - m->pc == PQ_INACTIVE)
2825 _vm_page_queue_spin_lock(m);
2826 oqueue = _vm_page_rem_queue_spinlocked(m);
2828 if ((m->flags & PG_UNMANAGED) == 0) {
2829 if (oqueue == PQ_CACHE)
2830 mycpu->gd_cnt.v_reactivated++;
2831 vm_page_flag_clear(m, PG_WINATCFLS);
2832 _vm_page_add_queue_spinlocked(m, PQ_INACTIVE + m->pc, athead);
2835 &vm_page_queues[PQ_INACTIVE + m->pc].adds, 1);
2838 /* NOTE: PQ_NONE if condition not taken */
2839 _vm_page_queue_spin_unlock(m);
2840 /* leaves vm_page spinlocked */
2844 * Attempt to deactivate a page.
2849 vm_page_deactivate(vm_page_t m)
2851 vm_page_spin_lock(m);
2852 _vm_page_deactivate_locked(m, 0);
2853 vm_page_spin_unlock(m);
2857 vm_page_deactivate_locked(vm_page_t m)
2859 _vm_page_deactivate_locked(m, 0);
2863 * Attempt to move a busied page to PQ_CACHE, then unconditionally unbusy it.
2865 * This function returns non-zero if it successfully moved the page to
2868 * This function unconditionally unbusies the page on return.
2871 vm_page_try_to_cache(vm_page_t m)
2874 * Shortcut if we obviously cannot move the page, or if the
2875 * page is already on the cache queue.
2877 if (m->dirty || m->hold_count || m->wire_count ||
2878 m->queue - m->pc == PQ_CACHE ||
2879 (m->flags & (PG_UNMANAGED | PG_NEED_COMMIT))) {
2885 * Page busied by us and no longer spinlocked. Dirty pages cannot
2886 * be moved to the cache.
2888 vm_page_test_dirty(m);
2889 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2898 * Attempt to free the page. If we cannot free it, we do nothing.
2899 * 1 is returned on success, 0 on failure.
2901 * Caller provides an unlocked/non-busied page.
2905 vm_page_try_to_free(vm_page_t m)
2907 if (vm_page_busy_try(m, TRUE))
2911 * The page can be in any state, including already being on the free
2912 * queue. Check to see if it really can be freed.
2914 if (m->dirty || /* can't free if it is dirty */
2915 m->hold_count || /* or held (XXX may be wrong) */
2916 m->wire_count || /* or wired */
2917 (m->flags & (PG_UNMANAGED | /* or unmanaged */
2918 PG_NEED_COMMIT)) || /* or needs a commit */
2919 m->queue - m->pc == PQ_FREE || /* already on PQ_FREE */
2920 m->queue - m->pc == PQ_HOLD) { /* already on PQ_HOLD */
2926 * We can probably free the page.
2928 * Page busied by us and no longer spinlocked. Dirty pages will
2929 * not be freed by this function. We have to re-test the
2930 * dirty bit after cleaning out the pmaps.
2932 vm_page_test_dirty(m);
2933 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2937 vm_page_protect(m, VM_PROT_NONE);
2938 if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2949 * Put the specified page onto the page cache queue (if appropriate).
2951 * The page must be busy, and this routine will release the busy and
2952 * possibly even free the page.
2955 vm_page_cache(vm_page_t m)
2958 * Not suitable for the cache
2960 if ((m->flags & (PG_UNMANAGED | PG_NEED_COMMIT)) ||
2961 (m->busy_count & PBUSY_MASK) ||
2962 m->wire_count || m->hold_count) {
2968 * Already in the cache (and thus not mapped)
2970 if ((m->queue - m->pc) == PQ_CACHE) {
2971 KKASSERT((m->flags & PG_MAPPED) == 0);
2977 * Caller is required to test m->dirty, but note that the act of
2978 * removing the page from its maps can cause it to become dirty
2979 * on an SMP system due to another cpu running in usermode.
2982 panic("vm_page_cache: caching a dirty page, pindex: %ld",
2987 * Remove all pmaps and indicate that the page is not
2988 * writeable or mapped. Our vm_page_protect() call may
2989 * have blocked (especially w/ VM_PROT_NONE), so recheck
2992 vm_page_protect(m, VM_PROT_NONE);
2993 if ((m->flags & (PG_UNMANAGED | PG_MAPPED)) ||
2994 (m->busy_count & PBUSY_MASK) ||
2995 m->wire_count || m->hold_count) {
2997 } else if (m->dirty || (m->flags & PG_NEED_COMMIT)) {
2998 vm_page_deactivate(m);
3001 _vm_page_and_queue_spin_lock(m);
3002 _vm_page_rem_queue_spinlocked(m);
3003 _vm_page_add_queue_spinlocked(m, PQ_CACHE + m->pc, 0);
3004 _vm_page_and_queue_spin_unlock(m);
3006 vm_page_free_wakeup();
3011 * vm_page_dontneed()
3013 * Cache, deactivate, or do nothing as appropriate. This routine
3014 * is typically used by madvise() MADV_DONTNEED.
3016 * Generally speaking we want to move the page into the cache so
3017 * it gets reused quickly. However, this can result in a silly syndrome
3018 * due to the page recycling too quickly. Small objects will not be
3019 * fully cached. On the otherhand, if we move the page to the inactive
3020 * queue we wind up with a problem whereby very large objects
3021 * unnecessarily blow away our inactive and cache queues.
3023 * The solution is to move the pages based on a fixed weighting. We
3024 * either leave them alone, deactivate them, or move them to the cache,
3025 * where moving them to the cache has the highest weighting.
3026 * By forcing some pages into other queues we eventually force the
3027 * system to balance the queues, potentially recovering other unrelated
3028 * space from active. The idea is to not force this to happen too
3031 * The page must be busied.
3034 vm_page_dontneed(vm_page_t m)
3036 static int dnweight;
3043 * occassionally leave the page alone
3045 if ((dnw & 0x01F0) == 0 ||
3046 m->queue - m->pc == PQ_INACTIVE ||
3047 m->queue - m->pc == PQ_CACHE
3049 if (m->act_count >= ACT_INIT)
3055 * If vm_page_dontneed() is inactivating a page, it must clear
3056 * the referenced flag; otherwise the pagedaemon will see references
3057 * on the page in the inactive queue and reactivate it. Until the
3058 * page can move to the cache queue, madvise's job is not done.
3060 vm_page_flag_clear(m, PG_REFERENCED);
3061 pmap_clear_reference(m);
3064 vm_page_test_dirty(m);
3066 if (m->dirty || (dnw & 0x0070) == 0) {
3068 * Deactivate the page 3 times out of 32.
3073 * Cache the page 28 times out of every 32. Note that
3074 * the page is deactivated instead of cached, but placed
3075 * at the head of the queue instead of the tail.
3079 vm_page_spin_lock(m);
3080 _vm_page_deactivate_locked(m, head);
3081 vm_page_spin_unlock(m);
3085 * These routines manipulate the 'soft busy' count for a page. A soft busy
3086 * is almost like a hard BUSY except that it allows certain compatible
3087 * operations to occur on the page while it is busy. For example, a page
3088 * undergoing a write can still be mapped read-only.
3090 * We also use soft-busy to quickly pmap_enter shared read-only pages
3091 * without having to hold the page locked.
3093 * The soft-busy count can be > 1 in situations where multiple threads
3094 * are pmap_enter()ing the same page simultaneously, or when two buffer
3095 * cache buffers overlap the same page.
3097 * The caller must hold the page BUSY when making these two calls.
3100 vm_page_io_start(vm_page_t m)
3104 ocount = atomic_fetchadd_int(&m->busy_count, 1);
3105 KKASSERT(ocount & PBUSY_LOCKED);
3109 vm_page_io_finish(vm_page_t m)
3113 ocount = atomic_fetchadd_int(&m->busy_count, -1);
3114 KKASSERT(ocount & PBUSY_MASK);
3116 if (((ocount - 1) & (PBUSY_LOCKED | PBUSY_MASK)) == 0)
3122 * Attempt to soft-busy a page. The page must not be PBUSY_LOCKED.
3124 * We can't use fetchadd here because we might race a hard-busy and the
3125 * page freeing code asserts on a non-zero soft-busy count (even if only
3128 * Returns 0 on success, non-zero on failure.
3131 vm_page_sbusy_try(vm_page_t m)
3136 ocount = m->busy_count;
3138 if (ocount & PBUSY_LOCKED)
3140 if (atomic_cmpset_int(&m->busy_count, ocount, ocount + 1))
3145 if (m->busy_count & PBUSY_LOCKED)
3147 ocount = atomic_fetchadd_int(&m->busy_count, 1);
3148 if (ocount & PBUSY_LOCKED) {
3149 vm_page_sbusy_drop(m);
3157 * Indicate that a clean VM page requires a filesystem commit and cannot
3158 * be reused. Used by tmpfs.
3161 vm_page_need_commit(vm_page_t m)
3163 vm_page_flag_set(m, PG_NEED_COMMIT);
3164 vm_object_set_writeable_dirty(m->object);
3168 vm_page_clear_commit(vm_page_t m)
3170 vm_page_flag_clear(m, PG_NEED_COMMIT);
3174 * Grab a page, blocking if it is busy and allocating a page if necessary.
3175 * A busy page is returned or NULL. The page may or may not be valid and
3176 * might not be on a queue (the caller is responsible for the disposition of
3179 * If VM_ALLOC_ZERO is specified and the grab must allocate a new page, the
3180 * page will be zero'd and marked valid.
3182 * If VM_ALLOC_FORCE_ZERO is specified the page will be zero'd and marked
3183 * valid even if it already exists.
3185 * If VM_ALLOC_RETRY is specified this routine will never return NULL. Also
3186 * note that VM_ALLOC_NORMAL must be specified if VM_ALLOC_RETRY is specified.
3187 * VM_ALLOC_NULL_OK is implied when VM_ALLOC_RETRY is specified.
3189 * This routine may block, but if VM_ALLOC_RETRY is not set then NULL is
3190 * always returned if we had blocked.
3192 * This routine may not be called from an interrupt.
3194 * No other requirements.
3197 vm_page_grab(vm_object_t object, vm_pindex_t pindex, int allocflags)
3203 KKASSERT(allocflags &
3204 (VM_ALLOC_NORMAL|VM_ALLOC_INTERRUPT|VM_ALLOC_SYSTEM));
3205 vm_object_hold_shared(object);
3207 m = vm_page_lookup_busy_try(object, pindex, TRUE, &error);
3209 vm_page_sleep_busy(m, TRUE, "pgrbwt");
3210 if ((allocflags & VM_ALLOC_RETRY) == 0) {
3215 } else if (m == NULL) {
3217 vm_object_upgrade(object);
3220 if (allocflags & VM_ALLOC_RETRY)
3221 allocflags |= VM_ALLOC_NULL_OK;
3222 m = vm_page_alloc(object, pindex,
3223 allocflags & ~VM_ALLOC_RETRY);
3227 if ((allocflags & VM_ALLOC_RETRY) == 0)
3236 * If VM_ALLOC_ZERO an invalid page will be zero'd and set valid.
3238 * If VM_ALLOC_FORCE_ZERO the page is unconditionally zero'd and set
3239 * valid even if already valid.
3241 * NOTE! We have removed all of the PG_ZERO optimizations and also
3242 * removed the idle zeroing code. These optimizations actually
3243 * slow things down on modern cpus because the zerod area is
3244 * likely uncached, placing a memory-access burden on the
3245 * accesors taking the fault.
3247 * By always zeroing the page in-line with the fault, no
3248 * dynamic ram reads are needed and the caches are hot, ready
3249 * for userland to access the memory.
3251 if (m->valid == 0) {
3252 if (allocflags & (VM_ALLOC_ZERO | VM_ALLOC_FORCE_ZERO)) {
3253 pmap_zero_page(VM_PAGE_TO_PHYS(m));
3254 m->valid = VM_PAGE_BITS_ALL;
3256 } else if (allocflags & VM_ALLOC_FORCE_ZERO) {
3257 pmap_zero_page(VM_PAGE_TO_PHYS(m));
3258 m->valid = VM_PAGE_BITS_ALL;
3261 vm_object_drop(object);
3266 * Mapping function for valid bits or for dirty bits in
3267 * a page. May not block.
3269 * Inputs are required to range within a page.
3275 vm_page_bits(int base, int size)
3281 base + size <= PAGE_SIZE,
3282 ("vm_page_bits: illegal base/size %d/%d", base, size)
3285 if (size == 0) /* handle degenerate case */
3288 first_bit = base >> DEV_BSHIFT;
3289 last_bit = (base + size - 1) >> DEV_BSHIFT;
3291 return ((2 << last_bit) - (1 << first_bit));
3295 * Sets portions of a page valid and clean. The arguments are expected
3296 * to be DEV_BSIZE aligned but if they aren't the bitmap is inclusive
3297 * of any partial chunks touched by the range. The invalid portion of
3298 * such chunks will be zero'd.
3300 * NOTE: When truncating a buffer vnode_pager_setsize() will automatically
3301 * align base to DEV_BSIZE so as not to mark clean a partially
3302 * truncated device block. Otherwise the dirty page status might be
3305 * This routine may not block.
3307 * (base + size) must be less then or equal to PAGE_SIZE.
3310 _vm_page_zero_valid(vm_page_t m, int base, int size)
3315 if (size == 0) /* handle degenerate case */
3319 * If the base is not DEV_BSIZE aligned and the valid
3320 * bit is clear, we have to zero out a portion of the
3324 if ((frag = base & ~(DEV_BSIZE - 1)) != base &&
3325 (m->valid & (1 << (base >> DEV_BSHIFT))) == 0
3327 pmap_zero_page_area(
3335 * If the ending offset is not DEV_BSIZE aligned and the
3336 * valid bit is clear, we have to zero out a portion of
3340 endoff = base + size;
3342 if ((frag = endoff & ~(DEV_BSIZE - 1)) != endoff &&
3343 (m->valid & (1 << (endoff >> DEV_BSHIFT))) == 0
3345 pmap_zero_page_area(
3348 DEV_BSIZE - (endoff & (DEV_BSIZE - 1))
3354 * Set valid, clear dirty bits. If validating the entire
3355 * page we can safely clear the pmap modify bit. We also
3356 * use this opportunity to clear the PG_NOSYNC flag. If a process
3357 * takes a write fault on a MAP_NOSYNC memory area the flag will
3360 * We set valid bits inclusive of any overlap, but we can only
3361 * clear dirty bits for DEV_BSIZE chunks that are fully within
3364 * Page must be busied?
3365 * No other requirements.
3368 vm_page_set_valid(vm_page_t m, int base, int size)
3370 _vm_page_zero_valid(m, base, size);
3371 m->valid |= vm_page_bits(base, size);
3376 * Set valid bits and clear dirty bits.
3378 * Page must be busied by caller.
3380 * NOTE: This function does not clear the pmap modified bit.
3381 * Also note that e.g. NFS may use a byte-granular base
3384 * No other requirements.
3387 vm_page_set_validclean(vm_page_t m, int base, int size)
3391 _vm_page_zero_valid(m, base, size);
3392 pagebits = vm_page_bits(base, size);
3393 m->valid |= pagebits;
3394 m->dirty &= ~pagebits;
3395 if (base == 0 && size == PAGE_SIZE) {
3396 /*pmap_clear_modify(m);*/
3397 vm_page_flag_clear(m, PG_NOSYNC);
3402 * Set valid & dirty. Used by buwrite()
3404 * Page must be busied by caller.
3407 vm_page_set_validdirty(vm_page_t m, int base, int size)
3411 pagebits = vm_page_bits(base, size);
3412 m->valid |= pagebits;
3413 m->dirty |= pagebits;
3415 vm_object_set_writeable_dirty(m->object);
3421 * NOTE: This function does not clear the pmap modified bit.
3422 * Also note that e.g. NFS may use a byte-granular base
3425 * Page must be busied?
3426 * No other requirements.
3429 vm_page_clear_dirty(vm_page_t m, int base, int size)
3431 m->dirty &= ~vm_page_bits(base, size);
3432 if (base == 0 && size == PAGE_SIZE) {
3433 /*pmap_clear_modify(m);*/
3434 vm_page_flag_clear(m, PG_NOSYNC);
3439 * Make the page all-dirty.
3441 * Also make sure the related object and vnode reflect the fact that the
3442 * object may now contain a dirty page.
3444 * Page must be busied?
3445 * No other requirements.
3448 vm_page_dirty(vm_page_t m)
3451 int pqtype = m->queue - m->pc;
3453 KASSERT(pqtype != PQ_CACHE && pqtype != PQ_FREE,
3454 ("vm_page_dirty: page in free/cache queue!"));
3455 if (m->dirty != VM_PAGE_BITS_ALL) {
3456 m->dirty = VM_PAGE_BITS_ALL;
3458 vm_object_set_writeable_dirty(m->object);
3463 * Invalidates DEV_BSIZE'd chunks within a page. Both the
3464 * valid and dirty bits for the effected areas are cleared.
3466 * Page must be busied?
3468 * No other requirements.
3471 vm_page_set_invalid(vm_page_t m, int base, int size)
3475 bits = vm_page_bits(base, size);
3478 atomic_add_int(&m->object->generation, 1);
3482 * The kernel assumes that the invalid portions of a page contain
3483 * garbage, but such pages can be mapped into memory by user code.
3484 * When this occurs, we must zero out the non-valid portions of the
3485 * page so user code sees what it expects.
3487 * Pages are most often semi-valid when the end of a file is mapped
3488 * into memory and the file's size is not page aligned.
3490 * Page must be busied?
3491 * No other requirements.
3494 vm_page_zero_invalid(vm_page_t m, boolean_t setvalid)
3500 * Scan the valid bits looking for invalid sections that
3501 * must be zerod. Invalid sub-DEV_BSIZE'd areas ( where the
3502 * valid bit may be set ) have already been zerod by
3503 * vm_page_set_validclean().
3505 for (b = i = 0; i <= PAGE_SIZE / DEV_BSIZE; ++i) {
3506 if (i == (PAGE_SIZE / DEV_BSIZE) ||
3507 (m->valid & (1 << i))
3510 pmap_zero_page_area(
3513 (i - b) << DEV_BSHIFT
3521 * setvalid is TRUE when we can safely set the zero'd areas
3522 * as being valid. We can do this if there are no cache consistency
3523 * issues. e.g. it is ok to do with UFS, but not ok to do with NFS.
3526 m->valid = VM_PAGE_BITS_ALL;
3530 * Is a (partial) page valid? Note that the case where size == 0
3531 * will return FALSE in the degenerate case where the page is entirely
3532 * invalid, and TRUE otherwise.
3535 * No other requirements.
3538 vm_page_is_valid(vm_page_t m, int base, int size)
3540 int bits = vm_page_bits(base, size);
3542 if (m->valid && ((m->valid & bits) == bits))
3549 * update dirty bits from pmap/mmu. May not block.
3551 * Caller must hold the page busy
3554 vm_page_test_dirty(vm_page_t m)
3556 if ((m->dirty != VM_PAGE_BITS_ALL) && pmap_is_modified(m)) {
3561 #include "opt_ddb.h"
3563 #include <ddb/ddb.h>
3565 DB_SHOW_COMMAND(page, vm_page_print_page_info)
3567 db_printf("vmstats.v_free_count: %ld\n", vmstats.v_free_count);
3568 db_printf("vmstats.v_cache_count: %ld\n", vmstats.v_cache_count);
3569 db_printf("vmstats.v_inactive_count: %ld\n", vmstats.v_inactive_count);
3570 db_printf("vmstats.v_active_count: %ld\n", vmstats.v_active_count);
3571 db_printf("vmstats.v_wire_count: %ld\n", vmstats.v_wire_count);
3572 db_printf("vmstats.v_free_reserved: %ld\n", vmstats.v_free_reserved);
3573 db_printf("vmstats.v_free_min: %ld\n", vmstats.v_free_min);
3574 db_printf("vmstats.v_free_target: %ld\n", vmstats.v_free_target);
3575 db_printf("vmstats.v_cache_min: %ld\n", vmstats.v_cache_min);
3576 db_printf("vmstats.v_inactive_target: %ld\n",
3577 vmstats.v_inactive_target);
3580 DB_SHOW_COMMAND(pageq, vm_page_print_pageq_info)
3583 db_printf("PQ_FREE:");
3584 for (i = 0; i < PQ_L2_SIZE; i++) {
3585 db_printf(" %ld", vm_page_queues[PQ_FREE + i].lcnt);
3589 db_printf("PQ_CACHE:");
3590 for(i = 0; i < PQ_L2_SIZE; i++) {
3591 db_printf(" %ld", vm_page_queues[PQ_CACHE + i].lcnt);
3595 db_printf("PQ_ACTIVE:");
3596 for(i = 0; i < PQ_L2_SIZE; i++) {
3597 db_printf(" %ld", vm_page_queues[PQ_ACTIVE + i].lcnt);
3601 db_printf("PQ_INACTIVE:");
3602 for(i = 0; i < PQ_L2_SIZE; i++) {
3603 db_printf(" %ld", vm_page_queues[PQ_INACTIVE + i].lcnt);