2 * Copyright (c) 2002-2006 Rice University
3 * Copyright (c) 2007-2011 Alan L. Cox <alc@cs.rice.edu>
6 * This software was developed for the FreeBSD Project by Alan L. Cox,
7 * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
18 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
19 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
20 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
21 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
22 * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
23 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
24 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
25 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
26 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
28 * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29 * POSSIBILITY OF SUCH DAMAGE.
33 * Superpage reservation management module
35 * Any external functions defined by this module are only to be used by the
36 * virtual memory system.
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
44 #include <sys/param.h>
45 #include <sys/kernel.h>
47 #include <sys/malloc.h>
48 #include <sys/mutex.h>
49 #include <sys/queue.h>
50 #include <sys/rwlock.h>
52 #include <sys/sysctl.h>
53 #include <sys/systm.h>
54 #include <sys/vmmeter.h>
57 #include <vm/vm_param.h>
58 #include <vm/vm_object.h>
59 #include <vm/vm_page.h>
60 #include <vm/vm_phys.h>
61 #include <vm/vm_radix.h>
62 #include <vm/vm_reserv.h>
65 * The reservation system supports the speculative allocation of large physical
66 * pages ("superpages"). Speculative allocation enables the fully automatic
67 * utilization of superpages by the virtual memory system. In other words, no
68 * programmatic directives are required to use superpages.
71 #if VM_NRESERVLEVEL > 0
74 * The number of small pages that are contained in a level 0 reservation
76 #define VM_LEVEL_0_NPAGES (1 << VM_LEVEL_0_ORDER)
79 * The number of bits by which a physical address is shifted to obtain the
82 #define VM_LEVEL_0_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT)
85 * The size of a level 0 reservation in bytes
87 #define VM_LEVEL_0_SIZE (1 << VM_LEVEL_0_SHIFT)
90 * Computes the index of the small page underlying the given (object, pindex)
91 * within the reservation's array of small pages.
93 #define VM_RESERV_INDEX(object, pindex) \
94 (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
97 * The size of a population map entry
99 typedef u_long popmap_t;
102 * The number of bits in a population map entry
104 #define NBPOPMAP (NBBY * sizeof(popmap_t))
107 * The number of population map entries in a reservation
109 #define NPOPMAP howmany(VM_LEVEL_0_NPAGES, NBPOPMAP)
112 * Clear a bit in the population map.
115 popmap_clear(popmap_t popmap[], int i)
118 popmap[i / NBPOPMAP] &= ~(1UL << (i % NBPOPMAP));
122 * Set a bit in the population map.
125 popmap_set(popmap_t popmap[], int i)
128 popmap[i / NBPOPMAP] |= 1UL << (i % NBPOPMAP);
132 * Is a bit in the population map clear?
134 static __inline boolean_t
135 popmap_is_clear(popmap_t popmap[], int i)
138 return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) == 0);
142 * Is a bit in the population map set?
144 static __inline boolean_t
145 popmap_is_set(popmap_t popmap[], int i)
148 return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) != 0);
152 * The reservation structure
154 * A reservation structure is constructed whenever a large physical page is
155 * speculatively allocated to an object. The reservation provides the small
156 * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
157 * within that object. The reservation's "popcnt" tracks the number of these
158 * small physical pages that are in use at any given time. When and if the
159 * reservation is not fully utilized, it appears in the queue of partially
160 * populated reservations. The reservation always appears on the containing
161 * object's list of reservations.
163 * A partially populated reservation can be broken and reclaimed at any time.
166 TAILQ_ENTRY(vm_reserv) partpopq;
167 LIST_ENTRY(vm_reserv) objq;
168 vm_object_t object; /* containing object */
169 vm_pindex_t pindex; /* offset within object */
170 vm_page_t pages; /* first page of a superpage */
171 int popcnt; /* # of pages in use */
173 popmap_t popmap[NPOPMAP]; /* bit vector of used pages */
177 * The reservation array
179 * This array is analoguous in function to vm_page_array. It differs in the
180 * respect that it may contain a greater number of useful reservation
181 * structures than there are (physical) superpages. These "invalid"
182 * reservation structures exist to trade-off space for time in the
183 * implementation of vm_reserv_from_page(). Invalid reservation structures are
184 * distinguishable from "valid" reservation structures by inspecting the
185 * reservation's "pages" field. Invalid reservation structures have a NULL
188 * vm_reserv_from_page() maps a small (physical) page to an element of this
189 * array by computing a physical reservation number from the page's physical
190 * address. The physical reservation number is used as the array index.
192 * An "active" reservation is a valid reservation structure that has a non-NULL
193 * "object" field and a non-zero "popcnt" field. In other words, every active
194 * reservation belongs to a particular object. Moreover, every active
195 * reservation has an entry in the containing object's list of reservations.
197 static vm_reserv_t vm_reserv_array;
200 * The partially populated reservation queue
202 * This queue enables the fast recovery of an unused free small page from a
203 * partially populated reservation. The reservation at the head of this queue
204 * is the least recently changed, partially populated reservation.
206 * Access to this queue is synchronized by the free page queue lock.
208 static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop =
209 TAILQ_HEAD_INITIALIZER(vm_rvq_partpop);
211 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
213 static long vm_reserv_broken;
214 SYSCTL_LONG(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
215 &vm_reserv_broken, 0, "Cumulative number of broken reservations");
217 static long vm_reserv_freed;
218 SYSCTL_LONG(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
219 &vm_reserv_freed, 0, "Cumulative number of freed reservations");
221 static int sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS);
223 SYSCTL_PROC(_vm_reserv, OID_AUTO, fullpop, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
224 sysctl_vm_reserv_fullpop, "I", "Current number of full reservations");
226 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
228 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
229 sysctl_vm_reserv_partpopq, "A", "Partially populated reservation queues");
231 static long vm_reserv_reclaimed;
232 SYSCTL_LONG(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
233 &vm_reserv_reclaimed, 0, "Cumulative number of reclaimed reservations");
235 static void vm_reserv_break(vm_reserv_t rv, vm_page_t m);
236 static void vm_reserv_depopulate(vm_reserv_t rv, int index);
237 static vm_reserv_t vm_reserv_from_page(vm_page_t m);
238 static boolean_t vm_reserv_has_pindex(vm_reserv_t rv,
240 static void vm_reserv_populate(vm_reserv_t rv, int index);
241 static void vm_reserv_reclaim(vm_reserv_t rv);
244 * Returns the current number of full reservations.
246 * Since the number of full reservations is computed without acquiring the
247 * free page queue lock, the returned value may be inexact.
250 sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS)
253 struct vm_phys_seg *seg;
258 for (segind = 0; segind < vm_phys_nsegs; segind++) {
259 seg = &vm_phys_segs[segind];
260 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
261 while (paddr + VM_LEVEL_0_SIZE <= seg->end) {
262 rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
263 fullpop += rv->popcnt == VM_LEVEL_0_NPAGES;
264 paddr += VM_LEVEL_0_SIZE;
267 return (sysctl_handle_int(oidp, &fullpop, 0, req));
271 * Describes the current state of the partially populated reservation queue.
274 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
278 int counter, error, level, unused_pages;
280 error = sysctl_wire_old_buffer(req, 0);
283 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
284 sbuf_printf(&sbuf, "\nLEVEL SIZE NUMBER\n\n");
285 for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
288 mtx_lock(&vm_page_queue_free_mtx);
289 TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) {
291 unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
293 mtx_unlock(&vm_page_queue_free_mtx);
294 sbuf_printf(&sbuf, "%5d: %6dK, %6d\n", level,
295 unused_pages * ((int)PAGE_SIZE / 1024), counter);
297 error = sbuf_finish(&sbuf);
303 * Reduces the given reservation's population count. If the population count
304 * becomes zero, the reservation is destroyed. Additionally, moves the
305 * reservation to the tail of the partially populated reservation queue if the
306 * population count is non-zero.
308 * The free page queue lock must be held.
311 vm_reserv_depopulate(vm_reserv_t rv, int index)
314 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
315 KASSERT(rv->object != NULL,
316 ("vm_reserv_depopulate: reserv %p is free", rv));
317 KASSERT(popmap_is_set(rv->popmap, index),
318 ("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
320 KASSERT(rv->popcnt > 0,
321 ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
322 if (rv->inpartpopq) {
323 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
324 rv->inpartpopq = FALSE;
326 KASSERT(rv->pages->psind == 1,
327 ("vm_reserv_depopulate: reserv %p is already demoted",
329 rv->pages->psind = 0;
331 popmap_clear(rv->popmap, index);
333 if (rv->popcnt == 0) {
334 LIST_REMOVE(rv, objq);
336 vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
339 rv->inpartpopq = TRUE;
340 TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
345 * Returns the reservation to which the given page might belong.
347 static __inline vm_reserv_t
348 vm_reserv_from_page(vm_page_t m)
351 return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
355 * Returns TRUE if the given reservation contains the given page index and
358 static __inline boolean_t
359 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
362 return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
366 * Increases the given reservation's population count. Moves the reservation
367 * to the tail of the partially populated reservation queue.
369 * The free page queue must be locked.
372 vm_reserv_populate(vm_reserv_t rv, int index)
375 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
376 KASSERT(rv->object != NULL,
377 ("vm_reserv_populate: reserv %p is free", rv));
378 KASSERT(popmap_is_clear(rv->popmap, index),
379 ("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
381 KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
382 ("vm_reserv_populate: reserv %p is already full", rv));
383 KASSERT(rv->pages->psind == 0,
384 ("vm_reserv_populate: reserv %p is already promoted", rv));
385 if (rv->inpartpopq) {
386 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
387 rv->inpartpopq = FALSE;
389 popmap_set(rv->popmap, index);
391 if (rv->popcnt < VM_LEVEL_0_NPAGES) {
392 rv->inpartpopq = TRUE;
393 TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
395 rv->pages->psind = 1;
399 * Allocates a contiguous set of physical pages of the given size "npages"
400 * from existing or newly created reservations. All of the physical pages
401 * must be at or above the given physical address "low" and below the given
402 * physical address "high". The given value "alignment" determines the
403 * alignment of the first physical page in the set. If the given value
404 * "boundary" is non-zero, then the set of physical pages cannot cross any
405 * physical address boundary that is a multiple of that value. Both
406 * "alignment" and "boundary" must be a power of two.
408 * The page "mpred" must immediately precede the offset "pindex" within the
411 * The object and free page queue must be locked.
414 vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, u_long npages,
415 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
419 vm_page_t m, m_ret, msucc;
420 vm_pindex_t first, leftcap, rightcap;
422 u_long allocpages, maxpages, minpages;
425 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
426 VM_OBJECT_ASSERT_WLOCKED(object);
427 KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
430 * Is a reservation fundamentally impossible?
432 if (pindex < VM_RESERV_INDEX(object, pindex) ||
433 pindex + npages > object->size)
437 * All reservations of a particular size have the same alignment.
438 * Assuming that the first page is allocated from a reservation, the
439 * least significant bits of its physical address can be determined
440 * from its offset from the beginning of the reservation and the size
441 * of the reservation.
443 * Could the specified index within a reservation of the smallest
444 * possible size satisfy the alignment and boundary requirements?
446 pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
447 if ((pa & (alignment - 1)) != 0)
449 size = npages << PAGE_SHIFT;
450 if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
454 * Look for an existing reservation.
457 KASSERT(mpred->object == object,
458 ("vm_reserv_alloc_contig: object doesn't contain mpred"));
459 KASSERT(mpred->pindex < pindex,
460 ("vm_reserv_alloc_contig: mpred doesn't precede pindex"));
461 rv = vm_reserv_from_page(mpred);
462 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
464 msucc = TAILQ_NEXT(mpred, listq);
466 msucc = TAILQ_FIRST(&object->memq);
468 KASSERT(msucc->pindex > pindex,
469 ("vm_reserv_alloc_contig: msucc doesn't succeed pindex"));
470 rv = vm_reserv_from_page(msucc);
471 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
476 * Could at least one reservation fit between the first index to the
477 * left that can be used ("leftcap") and the first index to the right
478 * that cannot be used ("rightcap")?
480 first = pindex - VM_RESERV_INDEX(object, pindex);
482 if ((rv = vm_reserv_from_page(mpred))->object != object)
483 leftcap = mpred->pindex + 1;
485 leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
489 minpages = VM_RESERV_INDEX(object, pindex) + npages;
490 maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
491 allocpages = maxpages;
493 if ((rv = vm_reserv_from_page(msucc))->object != object)
494 rightcap = msucc->pindex;
496 rightcap = rv->pindex;
497 if (first + maxpages > rightcap) {
498 if (maxpages == VM_LEVEL_0_NPAGES)
502 * At least one reservation will fit between "leftcap"
503 * and "rightcap". However, a reservation for the
504 * last of the requested pages will not fit. Reduce
505 * the size of the upcoming allocation accordingly.
507 allocpages = minpages;
512 * Would the last new reservation extend past the end of the object?
514 if (first + maxpages > object->size) {
516 * Don't allocate the last new reservation if the object is a
517 * vnode or backed by another object that is a vnode.
519 if (object->type == OBJT_VNODE ||
520 (object->backing_object != NULL &&
521 object->backing_object->type == OBJT_VNODE)) {
522 if (maxpages == VM_LEVEL_0_NPAGES)
524 allocpages = minpages;
526 /* Speculate that the object may grow. */
530 * Allocate the physical pages. The alignment and boundary specified
531 * for this allocation may be different from the alignment and
532 * boundary specified for the requested pages. For instance, the
533 * specified index may not be the first page within the first new
536 m = vm_phys_alloc_contig(allocpages, low, high, ulmax(alignment,
537 VM_LEVEL_0_SIZE), boundary > VM_LEVEL_0_SIZE ? boundary : 0);
542 * The allocated physical pages always begin at a reservation
543 * boundary, but they do not always end at a reservation boundary.
544 * Initialize every reservation that is completely covered by the
545 * allocated physical pages.
548 index = VM_RESERV_INDEX(object, pindex);
550 rv = vm_reserv_from_page(m);
551 KASSERT(rv->pages == m,
552 ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
554 KASSERT(rv->object == NULL,
555 ("vm_reserv_alloc_contig: reserv %p isn't free", rv));
556 LIST_INSERT_HEAD(&object->rvq, rv, objq);
559 KASSERT(rv->popcnt == 0,
560 ("vm_reserv_alloc_contig: reserv %p's popcnt is corrupted",
562 KASSERT(!rv->inpartpopq,
563 ("vm_reserv_alloc_contig: reserv %p's inpartpopq is TRUE",
565 for (i = 0; i < NPOPMAP; i++)
566 KASSERT(rv->popmap[i] == 0,
567 ("vm_reserv_alloc_contig: reserv %p's popmap is corrupted",
569 n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
570 for (i = 0; i < n; i++)
571 vm_reserv_populate(rv, index + i);
574 m_ret = &rv->pages[index];
577 m += VM_LEVEL_0_NPAGES;
578 first += VM_LEVEL_0_NPAGES;
579 allocpages -= VM_LEVEL_0_NPAGES;
580 } while (allocpages >= VM_LEVEL_0_NPAGES);
584 * Found a matching reservation.
587 index = VM_RESERV_INDEX(object, pindex);
588 /* Does the allocation fit within the reservation? */
589 if (index + npages > VM_LEVEL_0_NPAGES)
591 m = &rv->pages[index];
592 pa = VM_PAGE_TO_PHYS(m);
593 if (pa < low || pa + size > high || (pa & (alignment - 1)) != 0 ||
594 ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
596 /* Handle vm_page_rename(m, new_object, ...). */
597 for (i = 0; i < npages; i++)
598 if (popmap_is_set(rv->popmap, index + i))
600 for (i = 0; i < npages; i++)
601 vm_reserv_populate(rv, index + i);
606 * Allocates a page from an existing or newly created reservation.
608 * The page "mpred" must immediately precede the offset "pindex" within the
611 * The object and free page queue must be locked.
614 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, vm_page_t mpred)
617 vm_pindex_t first, leftcap, rightcap;
621 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
622 VM_OBJECT_ASSERT_WLOCKED(object);
625 * Is a reservation fundamentally impossible?
627 if (pindex < VM_RESERV_INDEX(object, pindex) ||
628 pindex >= object->size)
632 * Look for an existing reservation.
635 KASSERT(mpred->object == object,
636 ("vm_reserv_alloc_page: object doesn't contain mpred"));
637 KASSERT(mpred->pindex < pindex,
638 ("vm_reserv_alloc_page: mpred doesn't precede pindex"));
639 rv = vm_reserv_from_page(mpred);
640 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
642 msucc = TAILQ_NEXT(mpred, listq);
644 msucc = TAILQ_FIRST(&object->memq);
646 KASSERT(msucc->pindex > pindex,
647 ("vm_reserv_alloc_page: msucc doesn't succeed pindex"));
648 rv = vm_reserv_from_page(msucc);
649 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
654 * Could a reservation fit between the first index to the left that
655 * can be used and the first index to the right that cannot be used?
657 first = pindex - VM_RESERV_INDEX(object, pindex);
659 if ((rv = vm_reserv_from_page(mpred))->object != object)
660 leftcap = mpred->pindex + 1;
662 leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
667 if ((rv = vm_reserv_from_page(msucc))->object != object)
668 rightcap = msucc->pindex;
670 rightcap = rv->pindex;
671 if (first + VM_LEVEL_0_NPAGES > rightcap)
676 * Would a new reservation extend past the end of the object?
678 if (first + VM_LEVEL_0_NPAGES > object->size) {
680 * Don't allocate a new reservation if the object is a vnode or
681 * backed by another object that is a vnode.
683 if (object->type == OBJT_VNODE ||
684 (object->backing_object != NULL &&
685 object->backing_object->type == OBJT_VNODE))
687 /* Speculate that the object may grow. */
691 * Allocate and populate the new reservation.
693 m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER);
696 rv = vm_reserv_from_page(m);
697 KASSERT(rv->pages == m,
698 ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
699 KASSERT(rv->object == NULL,
700 ("vm_reserv_alloc_page: reserv %p isn't free", rv));
701 LIST_INSERT_HEAD(&object->rvq, rv, objq);
704 KASSERT(rv->popcnt == 0,
705 ("vm_reserv_alloc_page: reserv %p's popcnt is corrupted", rv));
706 KASSERT(!rv->inpartpopq,
707 ("vm_reserv_alloc_page: reserv %p's inpartpopq is TRUE", rv));
708 for (i = 0; i < NPOPMAP; i++)
709 KASSERT(rv->popmap[i] == 0,
710 ("vm_reserv_alloc_page: reserv %p's popmap is corrupted",
712 index = VM_RESERV_INDEX(object, pindex);
713 vm_reserv_populate(rv, index);
714 return (&rv->pages[index]);
717 * Found a matching reservation.
720 index = VM_RESERV_INDEX(object, pindex);
721 m = &rv->pages[index];
722 /* Handle vm_page_rename(m, new_object, ...). */
723 if (popmap_is_set(rv->popmap, index))
725 vm_reserv_populate(rv, index);
730 * Breaks the given reservation. Except for the specified free page, all free
731 * pages in the reservation are returned to the physical memory allocator.
732 * The reservation's population count and map are reset to their initial
735 * The given reservation must not be in the partially populated reservation
736 * queue. The free page queue lock must be held.
739 vm_reserv_break(vm_reserv_t rv, vm_page_t m)
741 int begin_zeroes, hi, i, lo;
743 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
744 KASSERT(rv->object != NULL,
745 ("vm_reserv_break: reserv %p is free", rv));
746 KASSERT(!rv->inpartpopq,
747 ("vm_reserv_break: reserv %p's inpartpopq is TRUE", rv));
748 LIST_REMOVE(rv, objq);
752 * Since the reservation is being broken, there is no harm in
753 * abusing the population map to stop "m" from being returned
754 * to the physical memory allocator.
757 KASSERT(popmap_is_clear(rv->popmap, i),
758 ("vm_reserv_break: reserv %p's popmap is corrupted", rv));
759 popmap_set(rv->popmap, i);
764 /* Find the next 0 bit. Any previous 0 bits are < "hi". */
765 lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
767 /* Redundantly clears bits < "hi". */
769 rv->popcnt -= NBPOPMAP - hi;
770 while (++i < NPOPMAP) {
771 lo = ffsl(~rv->popmap[i]);
774 rv->popcnt -= NBPOPMAP;
782 KASSERT(lo > 0, ("vm_reserv_break: lo is %d", lo));
783 /* Convert from ffsl() to ordinary bit numbering. */
786 /* Redundantly clears bits < "hi". */
787 rv->popmap[i] &= ~((1UL << lo) - 1);
788 rv->popcnt -= lo - hi;
790 begin_zeroes = NBPOPMAP * i + lo;
791 /* Find the next 1 bit. */
793 hi = ffsl(rv->popmap[i]);
794 while (hi == 0 && ++i < NPOPMAP);
796 /* Convert from ffsl() to ordinary bit numbering. */
798 vm_phys_free_contig(&rv->pages[begin_zeroes], NBPOPMAP * i +
800 } while (i < NPOPMAP);
801 KASSERT(rv->popcnt == 0,
802 ("vm_reserv_break: reserv %p's popcnt is corrupted", rv));
807 * Breaks all reservations belonging to the given object.
810 vm_reserv_break_all(vm_object_t object)
814 mtx_lock(&vm_page_queue_free_mtx);
815 while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
816 KASSERT(rv->object == object,
817 ("vm_reserv_break_all: reserv %p is corrupted", rv));
818 if (rv->inpartpopq) {
819 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
820 rv->inpartpopq = FALSE;
822 vm_reserv_break(rv, NULL);
824 mtx_unlock(&vm_page_queue_free_mtx);
828 * Frees the given page if it belongs to a reservation. Returns TRUE if the
829 * page is freed and FALSE otherwise.
831 * The free page queue lock must be held.
834 vm_reserv_free_page(vm_page_t m)
838 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
839 rv = vm_reserv_from_page(m);
840 if (rv->object == NULL)
842 vm_reserv_depopulate(rv, m - rv->pages);
847 * Initializes the reservation management system. Specifically, initializes
848 * the reservation array.
850 * Requires that vm_page_array and first_page are initialized!
856 struct vm_phys_seg *seg;
860 * Initialize the reservation array. Specifically, initialize the
861 * "pages" field for every element that has an underlying superpage.
863 for (segind = 0; segind < vm_phys_nsegs; segind++) {
864 seg = &vm_phys_segs[segind];
865 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
866 while (paddr + VM_LEVEL_0_SIZE <= seg->end) {
867 vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages =
868 PHYS_TO_VM_PAGE(paddr);
869 paddr += VM_LEVEL_0_SIZE;
875 * Returns true if the given page belongs to a reservation and that page is
876 * free. Otherwise, returns false.
879 vm_reserv_is_page_free(vm_page_t m)
883 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
884 rv = vm_reserv_from_page(m);
885 if (rv->object == NULL)
887 return (popmap_is_clear(rv->popmap, m - rv->pages));
891 * If the given page belongs to a reservation, returns the level of that
892 * reservation. Otherwise, returns -1.
895 vm_reserv_level(vm_page_t m)
899 rv = vm_reserv_from_page(m);
900 return (rv->object != NULL ? 0 : -1);
904 * Returns a reservation level if the given page belongs to a fully populated
905 * reservation and -1 otherwise.
908 vm_reserv_level_iffullpop(vm_page_t m)
912 rv = vm_reserv_from_page(m);
913 return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
917 * Breaks the given partially populated reservation, releasing its free pages
918 * to the physical memory allocator.
920 * The free page queue lock must be held.
923 vm_reserv_reclaim(vm_reserv_t rv)
926 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
927 KASSERT(rv->inpartpopq,
928 ("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
929 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
930 rv->inpartpopq = FALSE;
931 vm_reserv_break(rv, NULL);
932 vm_reserv_reclaimed++;
936 * Breaks the reservation at the head of the partially populated reservation
937 * queue, releasing its free pages to the physical memory allocator. Returns
938 * TRUE if a reservation is broken and FALSE otherwise.
940 * The free page queue lock must be held.
943 vm_reserv_reclaim_inactive(void)
947 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
948 if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) {
949 vm_reserv_reclaim(rv);
956 * Searches the partially populated reservation queue for the least recently
957 * changed reservation with free pages that satisfy the given request for
958 * contiguous physical memory. If a satisfactory reservation is found, it is
959 * broken. Returns TRUE if a reservation is broken and FALSE otherwise.
961 * The free page queue lock must be held.
964 vm_reserv_reclaim_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
965 u_long alignment, vm_paddr_t boundary)
969 int hi, i, lo, low_index, next_free;
971 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
972 if (npages > VM_LEVEL_0_NPAGES - 1)
974 size = npages << PAGE_SHIFT;
975 TAILQ_FOREACH(rv, &vm_rvq_partpop, partpopq) {
976 pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]);
977 if (pa + PAGE_SIZE - size < low) {
978 /* This entire reservation is too low; go to next. */
981 pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
982 if (pa + size > high) {
983 /* This entire reservation is too high; go to next. */
987 /* Start the search for free pages at "low". */
988 low_index = (low + PAGE_MASK - pa) >> PAGE_SHIFT;
989 i = low_index / NBPOPMAP;
990 hi = low_index % NBPOPMAP;
994 /* Find the next free page. */
995 lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
996 while (lo == 0 && ++i < NPOPMAP)
997 lo = ffsl(~rv->popmap[i]);
1000 /* Convert from ffsl() to ordinary bit numbering. */
1002 next_free = NBPOPMAP * i + lo;
1003 pa = VM_PAGE_TO_PHYS(&rv->pages[next_free]);
1005 ("vm_reserv_reclaim_contig: pa is too low"));
1006 if (pa + size > high) {
1007 /* The rest of this reservation is too high. */
1009 } else if ((pa & (alignment - 1)) != 0 ||
1010 ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) {
1012 * The current page doesn't meet the alignment
1013 * and/or boundary requirements. Continue
1014 * searching this reservation until the rest
1015 * of its free pages are either excluded or
1019 if (hi >= NBPOPMAP) {
1025 /* Find the next used page. */
1026 hi = ffsl(rv->popmap[i] & ~((1UL << lo) - 1));
1027 while (hi == 0 && ++i < NPOPMAP) {
1028 if ((NBPOPMAP * i - next_free) * PAGE_SIZE >=
1030 vm_reserv_reclaim(rv);
1033 hi = ffsl(rv->popmap[i]);
1035 /* Convert from ffsl() to ordinary bit numbering. */
1038 if ((NBPOPMAP * i + hi - next_free) * PAGE_SIZE >=
1040 vm_reserv_reclaim(rv);
1043 } while (i < NPOPMAP);
1049 * Transfers the reservation underlying the given page to a new object.
1051 * The object must be locked.
1054 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
1055 vm_pindex_t old_object_offset)
1059 VM_OBJECT_ASSERT_WLOCKED(new_object);
1060 rv = vm_reserv_from_page(m);
1061 if (rv->object == old_object) {
1062 mtx_lock(&vm_page_queue_free_mtx);
1063 if (rv->object == old_object) {
1064 LIST_REMOVE(rv, objq);
1065 LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
1066 rv->object = new_object;
1067 rv->pindex -= old_object_offset;
1069 mtx_unlock(&vm_page_queue_free_mtx);
1074 * Returns the size (in bytes) of a reservation of the specified level.
1077 vm_reserv_size(int level)
1082 return (VM_LEVEL_0_SIZE);
1091 * Allocates the virtual and physical memory required by the reservation
1092 * management system's data structures, in particular, the reservation array.
1095 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
1101 * Calculate the size (in bytes) of the reservation array. Round up
1102 * from "high_water" because every small page is mapped to an element
1103 * in the reservation array based on its physical address. Thus, the
1104 * number of elements in the reservation array can be greater than the
1105 * number of superpages.
1107 size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
1110 * Allocate and map the physical memory for the reservation array. The
1111 * next available virtual address is returned by reference.
1113 new_end = end - round_page(size);
1114 vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
1115 VM_PROT_READ | VM_PROT_WRITE);
1116 bzero(vm_reserv_array, size);
1119 * Return the next available physical address.
1125 * Returns the superpage containing the given page.
1128 vm_reserv_to_superpage(vm_page_t m)
1132 VM_OBJECT_ASSERT_LOCKED(m->object);
1133 rv = vm_reserv_from_page(m);
1134 return (rv->object == m->object && rv->popcnt == VM_LEVEL_0_NPAGES ?
1138 #endif /* VM_NRESERVLEVEL > 0 */