2 * Copyright (c) 2002-2006 Rice University
3 * Copyright (c) 2007 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
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
41 #include <sys/param.h>
42 #include <sys/kernel.h>
44 #include <sys/malloc.h>
45 #include <sys/mutex.h>
46 #include <sys/queue.h>
48 #include <sys/sysctl.h>
49 #include <sys/systm.h>
52 #include <vm/vm_param.h>
53 #include <vm/vm_object.h>
54 #include <vm/vm_page.h>
55 #include <vm/vm_phys.h>
56 #include <vm/vm_reserv.h>
59 * The reservation system supports the speculative allocation of large physical
60 * pages ("superpages"). Speculative allocation enables the fully-automatic
61 * utilization of superpages by the virtual memory system. In other words, no
62 * programmatic directives are required to use superpages.
65 #if VM_NRESERVLEVEL > 0
68 * The number of small pages that are contained in a level 0 reservation
70 #define VM_LEVEL_0_NPAGES (1 << VM_LEVEL_0_ORDER)
73 * The number of bits by which a physical address is shifted to obtain the
76 #define VM_LEVEL_0_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT)
79 * The size of a level 0 reservation in bytes
81 #define VM_LEVEL_0_SIZE (1 << VM_LEVEL_0_SHIFT)
84 * Computes the index of the small page underlying the given (object, pindex)
85 * within the reservation's array of small pages.
87 #define VM_RESERV_INDEX(object, pindex) \
88 (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
91 * The reservation structure
93 * A reservation structure is constructed whenever a large physical page is
94 * speculatively allocated to an object. The reservation provides the small
95 * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
96 * within that object. The reservation's "popcnt" tracks the number of these
97 * small physical pages that are in use at any given time. When and if the
98 * reservation is not fully utilized, it appears in the queue of partially-
99 * populated reservations. The reservation always appears on the containing
100 * object's list of reservations.
102 * A partially-populated reservation can be broken and reclaimed at any time.
105 TAILQ_ENTRY(vm_reserv) partpopq;
106 LIST_ENTRY(vm_reserv) objq;
107 vm_object_t object; /* containing object */
108 vm_pindex_t pindex; /* offset within object */
109 vm_page_t pages; /* first page of a superpage */
110 int popcnt; /* # of pages in use */
115 * The reservation array
117 * This array is analoguous in function to vm_page_array. It differs in the
118 * respect that it may contain a greater number of useful reservation
119 * structures than there are (physical) superpages. These "invalid"
120 * reservation structures exist to trade-off space for time in the
121 * implementation of vm_reserv_from_page(). Invalid reservation structures are
122 * distinguishable from "valid" reservation structures by inspecting the
123 * reservation's "pages" field. Invalid reservation structures have a NULL
126 * vm_reserv_from_page() maps a small (physical) page to an element of this
127 * array by computing a physical reservation number from the page's physical
128 * address. The physical reservation number is used as the array index.
130 * An "active" reservation is a valid reservation structure that has a non-NULL
131 * "object" field and a non-zero "popcnt" field. In other words, every active
132 * reservation belongs to a particular object. Moreover, every active
133 * reservation has an entry in the containing object's list of reservations.
135 static vm_reserv_t vm_reserv_array;
138 * The partially-populated reservation queue
140 * This queue enables the fast recovery of an unused cached or free small page
141 * from a partially-populated reservation. The head of this queue is either
142 * the least-recently-populated or most-recently-depopulated reservation.
144 * Access to this queue is synchronized by the free page queue lock.
146 static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop =
147 TAILQ_HEAD_INITIALIZER(vm_rvq_partpop);
149 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
151 static long vm_reserv_broken;
152 SYSCTL_LONG(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
153 &vm_reserv_broken, 0, "Cumulative number of broken reservations");
155 static long vm_reserv_freed;
156 SYSCTL_LONG(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
157 &vm_reserv_freed, 0, "Cumulative number of freed reservations");
159 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
161 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
162 sysctl_vm_reserv_partpopq, "A", "Partially-populated reservation queues");
164 static long vm_reserv_reclaimed;
165 SYSCTL_LONG(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
166 &vm_reserv_reclaimed, 0, "Cumulative number of reclaimed reservations");
168 static void vm_reserv_depopulate(vm_reserv_t rv);
169 static vm_reserv_t vm_reserv_from_page(vm_page_t m);
170 static boolean_t vm_reserv_has_pindex(vm_reserv_t rv,
172 static void vm_reserv_populate(vm_reserv_t rv);
175 * Describes the current state of the partially-populated reservation queue.
178 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
183 const int cbufsize = (VM_NRESERVLEVEL + 1) * 81;
184 int counter, error, level, unused_pages;
186 cbuf = malloc(cbufsize, M_TEMP, M_WAITOK | M_ZERO);
187 sbuf_new(&sbuf, cbuf, cbufsize, SBUF_FIXEDLEN);
188 sbuf_printf(&sbuf, "\nLEVEL SIZE NUMBER\n\n");
189 for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
192 mtx_lock(&vm_page_queue_free_mtx);
193 TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) {
195 unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
197 mtx_unlock(&vm_page_queue_free_mtx);
198 sbuf_printf(&sbuf, "%5.5d: %6.6dK, %6.6d\n", level,
199 unused_pages * (PAGE_SIZE / 1024), counter);
202 error = SYSCTL_OUT(req, sbuf_data(&sbuf), sbuf_len(&sbuf));
209 * Reduces the given reservation's population count. If the population count
210 * becomes zero, the reservation is destroyed. Additionally, moves the
211 * reservation to the head of the partially-populated reservations queue if the
212 * population count is non-zero.
214 * The free page queue lock must be held.
217 vm_reserv_depopulate(vm_reserv_t rv)
220 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
221 KASSERT(rv->object != NULL,
222 ("vm_reserv_depopulate: reserv %p is free", rv));
223 KASSERT(rv->popcnt > 0,
224 ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
225 if (rv->inpartpopq) {
226 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
227 rv->inpartpopq = FALSE;
230 if (rv->popcnt == 0) {
231 LIST_REMOVE(rv, objq);
233 vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
236 rv->inpartpopq = TRUE;
237 TAILQ_INSERT_HEAD(&vm_rvq_partpop, rv, partpopq);
242 * Returns the reservation to which the given page might belong.
244 static __inline vm_reserv_t
245 vm_reserv_from_page(vm_page_t m)
248 return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
252 * Returns TRUE if the given reservation contains the given page index and
255 static __inline boolean_t
256 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
259 return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
263 * Increases the given reservation's population count. Moves the reservation
264 * to the tail of the partially-populated reservation queue.
266 * The free page queue must be locked.
269 vm_reserv_populate(vm_reserv_t rv)
272 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
273 KASSERT(rv->object != NULL,
274 ("vm_reserv_populate: reserv %p is free", rv));
275 KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
276 ("vm_reserv_populate: reserv %p is already full", rv));
277 if (rv->inpartpopq) {
278 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
279 rv->inpartpopq = FALSE;
282 if (rv->popcnt < VM_LEVEL_0_NPAGES) {
283 rv->inpartpopq = TRUE;
284 TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
289 * Allocates a page from an existing or newly-created reservation.
291 * The object and free page queue must be locked.
294 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex)
296 vm_page_t m, mpred, msucc;
297 vm_pindex_t first, leftcap, rightcap;
300 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
303 * Is a reservation fundamentally not possible?
305 VM_OBJECT_LOCK_ASSERT(object, MA_OWNED);
306 if (pindex < VM_RESERV_INDEX(object, pindex) ||
307 pindex >= object->size)
311 * Look for an existing reservation.
314 mpred = object->root;
315 while (mpred != NULL) {
316 KASSERT(mpred->pindex != pindex,
317 ("vm_reserv_alloc_page: pindex already allocated"));
318 rv = vm_reserv_from_page(mpred);
319 if (rv->object == object && vm_reserv_has_pindex(rv, pindex)) {
320 m = &rv->pages[VM_RESERV_INDEX(object, pindex)];
321 /* Handle vm_page_rename(m, new_object, ...). */
322 if ((m->flags & (PG_CACHED | PG_FREE)) == 0)
324 vm_reserv_populate(rv);
326 } else if (mpred->pindex < pindex) {
328 (msucc = TAILQ_NEXT(mpred, listq)) == NULL)
330 KASSERT(msucc->pindex != pindex,
331 ("vm_reserv_alloc_page: pindex already allocated"));
332 rv = vm_reserv_from_page(msucc);
333 if (rv->object == object &&
334 vm_reserv_has_pindex(rv, pindex)) {
335 m = &rv->pages[VM_RESERV_INDEX(object, pindex)];
336 /* Handle vm_page_rename(m, new_object, ...). */
337 if ((m->flags & (PG_CACHED | PG_FREE)) == 0)
339 vm_reserv_populate(rv);
341 } else if (pindex < msucc->pindex)
343 } else if (msucc == NULL) {
345 mpred = TAILQ_PREV(msucc, pglist, listq);
349 mpred = object->root = vm_page_splay(pindex, object->root);
353 * Determine the first index to the left that can be used.
357 else if ((rv = vm_reserv_from_page(mpred))->object != object)
358 leftcap = mpred->pindex + 1;
360 leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
363 * Determine the first index to the right that cannot be used.
366 rightcap = pindex + VM_LEVEL_0_NPAGES;
367 else if ((rv = vm_reserv_from_page(msucc))->object != object)
368 rightcap = msucc->pindex;
370 rightcap = rv->pindex;
373 * Determine if a reservation fits between the first index to
374 * the left that can be used and the first index to the right
375 * that cannot be used.
377 first = pindex - VM_RESERV_INDEX(object, pindex);
378 if (first < leftcap || first + VM_LEVEL_0_NPAGES > rightcap)
382 * Would a new reservation extend past the end of the given object?
384 if (object->size < first + VM_LEVEL_0_NPAGES) {
386 * Don't allocate a new reservation if the object is a vnode or
387 * backed by another object that is a vnode.
389 if (object->type == OBJT_VNODE ||
390 (object->backing_object != NULL &&
391 object->backing_object->type == OBJT_VNODE))
393 /* Speculate that the object may grow. */
397 * Allocate a new reservation.
399 m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER);
401 rv = vm_reserv_from_page(m);
402 KASSERT(rv->pages == m,
403 ("vm_reserv_alloc_page: reserv %p's pages is corrupted",
405 KASSERT(rv->object == NULL,
406 ("vm_reserv_alloc_page: reserv %p isn't free", rv));
407 LIST_INSERT_HEAD(&object->rvq, rv, objq);
410 KASSERT(rv->popcnt == 0,
411 ("vm_reserv_alloc_page: reserv %p's popcnt is corrupted",
413 KASSERT(!rv->inpartpopq,
414 ("vm_reserv_alloc_page: reserv %p's inpartpopq is TRUE",
416 vm_reserv_populate(rv);
417 m = &rv->pages[VM_RESERV_INDEX(object, pindex)];
423 * Breaks all reservations belonging to the given object.
426 vm_reserv_break_all(vm_object_t object)
431 mtx_lock(&vm_page_queue_free_mtx);
432 while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
433 KASSERT(rv->object == object,
434 ("vm_reserv_break_all: reserv %p is corrupted", rv));
435 if (rv->inpartpopq) {
436 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
437 rv->inpartpopq = FALSE;
439 LIST_REMOVE(rv, objq);
441 for (i = 0; i < VM_LEVEL_0_NPAGES; i++) {
442 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
443 vm_phys_free_pages(&rv->pages[i], 0);
447 KASSERT(rv->popcnt == 0,
448 ("vm_reserv_break_all: reserv %p's popcnt is corrupted",
452 mtx_unlock(&vm_page_queue_free_mtx);
456 * Frees the given page if it belongs to a reservation. Returns TRUE if the
457 * page is freed and FALSE otherwise.
459 * The free page queue lock must be held.
462 vm_reserv_free_page(vm_page_t m)
466 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
467 rv = vm_reserv_from_page(m);
468 if (rv->object != NULL) {
469 vm_reserv_depopulate(rv);
476 * Initializes the reservation management system. Specifically, initializes
477 * the reservation array.
479 * Requires that vm_page_array and first_page are initialized!
488 * Initialize the reservation array. Specifically, initialize the
489 * "pages" field for every element that has an underlying superpage.
491 for (i = 0; phys_avail[i + 1] != 0; i += 2) {
492 paddr = roundup2(phys_avail[i], VM_LEVEL_0_SIZE);
493 while (paddr + VM_LEVEL_0_SIZE <= phys_avail[i + 1]) {
494 vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages =
495 PHYS_TO_VM_PAGE(paddr);
496 paddr += VM_LEVEL_0_SIZE;
502 * Returns a reservation level if the given page belongs to a fully-populated
503 * reservation and -1 otherwise.
506 vm_reserv_level_iffullpop(vm_page_t m)
510 rv = vm_reserv_from_page(m);
511 return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
515 * Prepare for the reactivation of a cached page.
517 * First, suppose that the given page "m" was allocated individually, i.e., not
518 * as part of a reservation, and cached. Then, suppose a reservation
519 * containing "m" is allocated by the same object. Although "m" and the
520 * reservation belong to the same object, "m"'s pindex may not match the
523 * The free page queue must be locked.
526 vm_reserv_reactivate_page(vm_page_t m)
531 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
532 rv = vm_reserv_from_page(m);
533 if (rv->object == NULL)
535 KASSERT((m->flags & PG_CACHED) != 0,
536 ("vm_reserv_uncache_page: page %p is not cached", m));
537 if (m->object == rv->object &&
538 m->pindex - rv->pindex == VM_RESERV_INDEX(m->object, m->pindex))
539 vm_reserv_populate(rv);
541 KASSERT(rv->inpartpopq,
542 ("vm_reserv_uncache_page: reserv %p's inpartpopq is FALSE",
544 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
545 rv->inpartpopq = FALSE;
546 LIST_REMOVE(rv, objq);
548 /* Don't vm_phys_free_pages(m, 0). */
549 m_index = m - rv->pages;
550 for (i = 0; i < m_index; i++) {
551 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
552 vm_phys_free_pages(&rv->pages[i], 0);
556 for (i++; i < VM_LEVEL_0_NPAGES; i++) {
557 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
558 vm_phys_free_pages(&rv->pages[i], 0);
562 KASSERT(rv->popcnt == 0,
563 ("vm_reserv_uncache_page: reserv %p's popcnt is corrupted",
571 * Breaks the reservation at the head of the partially-populated reservation
572 * queue, releasing its cached and free pages to the physical memory
573 * allocator. Returns TRUE if a reservation is broken and FALSE otherwise.
575 * The free page queue lock must be held.
578 vm_reserv_reclaim(void)
583 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
584 if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) {
585 KASSERT(rv->inpartpopq,
586 ("vm_reserv_reclaim: reserv %p's inpartpopq is corrupted",
588 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
589 rv->inpartpopq = FALSE;
590 KASSERT(rv->object != NULL,
591 ("vm_reserv_reclaim: reserv %p is free", rv));
592 LIST_REMOVE(rv, objq);
594 for (i = 0; i < VM_LEVEL_0_NPAGES; i++) {
595 if ((rv->pages[i].flags & (PG_CACHED | PG_FREE)) != 0)
596 vm_phys_free_pages(&rv->pages[i], 0);
600 KASSERT(rv->popcnt == 0,
601 ("vm_reserv_reclaim: reserv %p's popcnt is corrupted",
603 vm_reserv_reclaimed++;
610 * Transfers the reservation underlying the given page to a new object.
612 * The object must be locked.
615 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
616 vm_pindex_t old_object_offset)
620 VM_OBJECT_LOCK_ASSERT(new_object, MA_OWNED);
621 rv = vm_reserv_from_page(m);
622 if (rv->object == old_object) {
623 mtx_lock(&vm_page_queue_free_mtx);
624 if (rv->object == old_object) {
625 LIST_REMOVE(rv, objq);
626 LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
627 rv->object = new_object;
628 rv->pindex -= old_object_offset;
630 mtx_unlock(&vm_page_queue_free_mtx);
635 * Allocates the virtual and physical memory required by the reservation
636 * management system's data structures, in particular, the reservation array.
639 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
645 * Calculate the size (in bytes) of the reservation array. Round up
646 * from "high_water" because every small page is mapped to an element
647 * in the reservation array based on its physical address. Thus, the
648 * number of elements in the reservation array can be greater than the
649 * number of superpages.
651 size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
654 * Allocate and map the physical memory for the reservation array. The
655 * next available virtual address is returned by reference.
657 new_end = end - round_page(size);
658 vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
659 VM_PROT_READ | VM_PROT_WRITE);
660 bzero(vm_reserv_array, size);
663 * Return the next available physical address.
668 #endif /* VM_NRESERVLEVEL > 0 */