FreeBSD/Linux Kernel Cross Reference
sys/vm/vm_reserv.c
1 /*-
2 * Copyright (c) 2002-2006 Rice University
3 * Copyright (c) 2007-2011 Alan L. Cox <alc@cs.rice.edu>
4 * All rights reserved.
5 *
6 * This software was developed for the FreeBSD Project by Alan L. Cox,
7 * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
8 *
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
11 * are met:
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.
17 *
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.
30 */
31
32 /*
33 * Superpage reservation management module
34 *
35 * Any external functions defined by this module are only to be used by the
36 * virtual memory system.
37 */
38
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD: releng/11.2/sys/vm/vm_reserv.c 331017 2018-03-15 19:08:33Z kevans $");
41
42 #include "opt_vm.h"
43
44 #include <sys/param.h>
45 #include <sys/kernel.h>
46 #include <sys/lock.h>
47 #include <sys/malloc.h>
48 #include <sys/mutex.h>
49 #include <sys/queue.h>
50 #include <sys/rwlock.h>
51 #include <sys/sbuf.h>
52 #include <sys/sysctl.h>
53 #include <sys/systm.h>
54 #include <sys/vmmeter.h>
55
56 #include <vm/vm.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>
63
64 /*
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.
69 */
70
71 #if VM_NRESERVLEVEL > 0
72
73 /*
74 * The number of small pages that are contained in a level 0 reservation
75 */
76 #define VM_LEVEL_0_NPAGES (1 << VM_LEVEL_0_ORDER)
77
78 /*
79 * The number of bits by which a physical address is shifted to obtain the
80 * reservation number
81 */
82 #define VM_LEVEL_0_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT)
83
84 /*
85 * The size of a level 0 reservation in bytes
86 */
87 #define VM_LEVEL_0_SIZE (1 << VM_LEVEL_0_SHIFT)
88
89 /*
90 * Computes the index of the small page underlying the given (object, pindex)
91 * within the reservation's array of small pages.
92 */
93 #define VM_RESERV_INDEX(object, pindex) \
94 (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
95
96 /*
97 * The size of a population map entry
98 */
99 typedef u_long popmap_t;
100
101 /*
102 * The number of bits in a population map entry
103 */
104 #define NBPOPMAP (NBBY * sizeof(popmap_t))
105
106 /*
107 * The number of population map entries in a reservation
108 */
109 #define NPOPMAP howmany(VM_LEVEL_0_NPAGES, NBPOPMAP)
110
111 /*
112 * Clear a bit in the population map.
113 */
114 static __inline void
115 popmap_clear(popmap_t popmap[], int i)
116 {
117
118 popmap[i / NBPOPMAP] &= ~(1UL << (i % NBPOPMAP));
119 }
120
121 /*
122 * Set a bit in the population map.
123 */
124 static __inline void
125 popmap_set(popmap_t popmap[], int i)
126 {
127
128 popmap[i / NBPOPMAP] |= 1UL << (i % NBPOPMAP);
129 }
130
131 /*
132 * Is a bit in the population map clear?
133 */
134 static __inline boolean_t
135 popmap_is_clear(popmap_t popmap[], int i)
136 {
137
138 return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) == 0);
139 }
140
141 /*
142 * Is a bit in the population map set?
143 */
144 static __inline boolean_t
145 popmap_is_set(popmap_t popmap[], int i)
146 {
147
148 return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) != 0);
149 }
150
151 /*
152 * The reservation structure
153 *
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.
162 *
163 * A partially populated reservation can be broken and reclaimed at any time.
164 */
165 struct vm_reserv {
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 */
172 char inpartpopq;
173 popmap_t popmap[NPOPMAP]; /* bit vector of used pages */
174 };
175
176 /*
177 * The reservation array
178 *
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
186 * "pages" field.
187 *
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.
191 *
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.
196 */
197 static vm_reserv_t vm_reserv_array;
198
199 /*
200 * The partially populated reservation queue
201 *
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.
205 *
206 * Access to this queue is synchronized by the free page queue lock.
207 */
208 static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop =
209 TAILQ_HEAD_INITIALIZER(vm_rvq_partpop);
210
211 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
212
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");
216
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");
220
221 static int sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS);
222
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");
225
226 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
227
228 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
229 sysctl_vm_reserv_partpopq, "A", "Partially populated reservation queues");
230
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");
234
235 static void vm_reserv_break(vm_reserv_t rv);
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,
239 vm_pindex_t pindex);
240 static void vm_reserv_populate(vm_reserv_t rv, int index);
241 static void vm_reserv_reclaim(vm_reserv_t rv);
242
243 /*
244 * Returns the current number of full reservations.
245 *
246 * Since the number of full reservations is computed without acquiring the
247 * free page queue lock, the returned value may be inexact.
248 */
249 static int
250 sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS)
251 {
252 vm_paddr_t paddr;
253 struct vm_phys_seg *seg;
254 vm_reserv_t rv;
255 int fullpop, segind;
256
257 fullpop = 0;
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;
265 }
266 }
267 return (sysctl_handle_int(oidp, &fullpop, 0, req));
268 }
269
270 /*
271 * Describes the current state of the partially populated reservation queue.
272 */
273 static int
274 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
275 {
276 struct sbuf sbuf;
277 vm_reserv_t rv;
278 int counter, error, level, unused_pages;
279
280 error = sysctl_wire_old_buffer(req, 0);
281 if (error != 0)
282 return (error);
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++) {
286 counter = 0;
287 unused_pages = 0;
288 mtx_lock(&vm_page_queue_free_mtx);
289 TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) {
290 counter++;
291 unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
292 }
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);
296 }
297 error = sbuf_finish(&sbuf);
298 sbuf_delete(&sbuf);
299 return (error);
300 }
301
302 /*
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.
307 *
308 * The free page queue lock must be held.
309 */
310 static void
311 vm_reserv_depopulate(vm_reserv_t rv, int index)
312 {
313
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,
319 index));
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;
325 } else {
326 KASSERT(rv->pages->psind == 1,
327 ("vm_reserv_depopulate: reserv %p is already demoted",
328 rv));
329 rv->pages->psind = 0;
330 }
331 popmap_clear(rv->popmap, index);
332 rv->popcnt--;
333 if (rv->popcnt == 0) {
334 LIST_REMOVE(rv, objq);
335 rv->object = NULL;
336 vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
337 vm_reserv_freed++;
338 } else {
339 rv->inpartpopq = TRUE;
340 TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
341 }
342 }
343
344 /*
345 * Returns the reservation to which the given page might belong.
346 */
347 static __inline vm_reserv_t
348 vm_reserv_from_page(vm_page_t m)
349 {
350
351 return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
352 }
353
354 /*
355 * Returns TRUE if the given reservation contains the given page index and
356 * FALSE otherwise.
357 */
358 static __inline boolean_t
359 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
360 {
361
362 return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
363 }
364
365 /*
366 * Increases the given reservation's population count. Moves the reservation
367 * to the tail of the partially populated reservation queue.
368 *
369 * The free page queue must be locked.
370 */
371 static void
372 vm_reserv_populate(vm_reserv_t rv, int index)
373 {
374
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,
380 index));
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;
388 }
389 popmap_set(rv->popmap, index);
390 rv->popcnt++;
391 if (rv->popcnt < VM_LEVEL_0_NPAGES) {
392 rv->inpartpopq = TRUE;
393 TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
394 } else
395 rv->pages->psind = 1;
396 }
397
398 /*
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.
407 *
408 * The page "mpred" must immediately precede the offset "pindex" within the
409 * specified object.
410 *
411 * The object and free page queue must be locked.
412 */
413 vm_page_t
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,
416 vm_page_t mpred)
417 {
418 vm_paddr_t pa, size;
419 vm_page_t m, m_ret, msucc;
420 vm_pindex_t first, leftcap, rightcap;
421 vm_reserv_t rv;
422 u_long allocpages, maxpages, minpages;
423 int i, index, n;
424
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"));
428
429 /*
430 * Is a reservation fundamentally impossible?
431 */
432 if (pindex < VM_RESERV_INDEX(object, pindex) ||
433 pindex + npages > object->size)
434 return (NULL);
435
436 /*
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.
442 *
443 * Could the specified index within a reservation of the smallest
444 * possible size satisfy the alignment and boundary requirements?
445 */
446 pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
447 if ((pa & (alignment - 1)) != 0)
448 return (NULL);
449 size = npages << PAGE_SHIFT;
450 if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
451 return (NULL);
452
453 /*
454 * Look for an existing reservation.
455 */
456 if (mpred != NULL) {
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))
463 goto found;
464 msucc = TAILQ_NEXT(mpred, listq);
465 } else
466 msucc = TAILQ_FIRST(&object->memq);
467 if (msucc != NULL) {
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))
472 goto found;
473 }
474
475 /*
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")?
479 */
480 first = pindex - VM_RESERV_INDEX(object, pindex);
481 if (mpred != NULL) {
482 if ((rv = vm_reserv_from_page(mpred))->object != object)
483 leftcap = mpred->pindex + 1;
484 else
485 leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
486 if (leftcap > first)
487 return (NULL);
488 }
489 minpages = VM_RESERV_INDEX(object, pindex) + npages;
490 maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
491 allocpages = maxpages;
492 if (msucc != NULL) {
493 if ((rv = vm_reserv_from_page(msucc))->object != object)
494 rightcap = msucc->pindex;
495 else
496 rightcap = rv->pindex;
497 if (first + maxpages > rightcap) {
498 if (maxpages == VM_LEVEL_0_NPAGES)
499 return (NULL);
500
501 /*
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.
506 */
507 allocpages = minpages;
508 }
509 }
510
511 /*
512 * Would the last new reservation extend past the end of the object?
513 */
514 if (first + maxpages > object->size) {
515 /*
516 * Don't allocate the last new reservation if the object is a
517 * vnode or backed by another object that is a vnode.
518 */
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)
523 return (NULL);
524 allocpages = minpages;
525 }
526 /* Speculate that the object may grow. */
527 }
528
529 /*
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
534 * reservation.
535 */
536 m = vm_phys_alloc_contig(allocpages, low, high, ulmax(alignment,
537 VM_LEVEL_0_SIZE), boundary > VM_LEVEL_0_SIZE ? boundary : 0);
538 if (m == NULL)
539 return (NULL);
540
541 /*
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.
546 */
547 m_ret = NULL;
548 index = VM_RESERV_INDEX(object, pindex);
549 do {
550 rv = vm_reserv_from_page(m);
551 KASSERT(rv->pages == m,
552 ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
553 rv));
554 KASSERT(rv->object == NULL,
555 ("vm_reserv_alloc_contig: reserv %p isn't free", rv));
556 LIST_INSERT_HEAD(&object->rvq, rv, objq);
557 rv->object = object;
558 rv->pindex = first;
559 KASSERT(rv->popcnt == 0,
560 ("vm_reserv_alloc_contig: reserv %p's popcnt is corrupted",
561 rv));
562 KASSERT(!rv->inpartpopq,
563 ("vm_reserv_alloc_contig: reserv %p's inpartpopq is TRUE",
564 rv));
565 for (i = 0; i < NPOPMAP; i++)
566 KASSERT(rv->popmap[i] == 0,
567 ("vm_reserv_alloc_contig: reserv %p's popmap is corrupted",
568 rv));
569 n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
570 for (i = 0; i < n; i++)
571 vm_reserv_populate(rv, index + i);
572 npages -= n;
573 if (m_ret == NULL) {
574 m_ret = &rv->pages[index];
575 index = 0;
576 }
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);
581 return (m_ret);
582
583 /*
584 * Found a matching reservation.
585 */
586 found:
587 index = VM_RESERV_INDEX(object, pindex);
588 /* Does the allocation fit within the reservation? */
589 if (index + npages > VM_LEVEL_0_NPAGES)
590 return (NULL);
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)
595 return (NULL);
596 /* Handle vm_page_rename(m, new_object, ...). */
597 for (i = 0; i < npages; i++)
598 if (popmap_is_set(rv->popmap, index + i))
599 return (NULL);
600 for (i = 0; i < npages; i++)
601 vm_reserv_populate(rv, index + i);
602 return (m);
603 }
604
605 /*
606 * Allocates a page from an existing or newly created reservation.
607 *
608 * The page "mpred" must immediately precede the offset "pindex" within the
609 * specified object.
610 *
611 * The object and free page queue must be locked.
612 */
613 vm_page_t
614 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, vm_page_t mpred)
615 {
616 vm_page_t m, msucc;
617 vm_pindex_t first, leftcap, rightcap;
618 vm_reserv_t rv;
619 int i, index;
620
621 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
622 VM_OBJECT_ASSERT_WLOCKED(object);
623
624 /*
625 * Is a reservation fundamentally impossible?
626 */
627 if (pindex < VM_RESERV_INDEX(object, pindex) ||
628 pindex >= object->size)
629 return (NULL);
630
631 /*
632 * Look for an existing reservation.
633 */
634 if (mpred != NULL) {
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))
641 goto found;
642 msucc = TAILQ_NEXT(mpred, listq);
643 } else
644 msucc = TAILQ_FIRST(&object->memq);
645 if (msucc != NULL) {
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))
650 goto found;
651 }
652
653 /*
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?
656 */
657 first = pindex - VM_RESERV_INDEX(object, pindex);
658 if (mpred != NULL) {
659 if ((rv = vm_reserv_from_page(mpred))->object != object)
660 leftcap = mpred->pindex + 1;
661 else
662 leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
663 if (leftcap > first)
664 return (NULL);
665 }
666 if (msucc != NULL) {
667 if ((rv = vm_reserv_from_page(msucc))->object != object)
668 rightcap = msucc->pindex;
669 else
670 rightcap = rv->pindex;
671 if (first + VM_LEVEL_0_NPAGES > rightcap)
672 return (NULL);
673 }
674
675 /*
676 * Would a new reservation extend past the end of the object?
677 */
678 if (first + VM_LEVEL_0_NPAGES > object->size) {
679 /*
680 * Don't allocate a new reservation if the object is a vnode or
681 * backed by another object that is a vnode.
682 */
683 if (object->type == OBJT_VNODE ||
684 (object->backing_object != NULL &&
685 object->backing_object->type == OBJT_VNODE))
686 return (NULL);
687 /* Speculate that the object may grow. */
688 }
689
690 /*
691 * Allocate and populate the new reservation.
692 */
693 m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER);
694 if (m == NULL)
695 return (NULL);
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);
702 rv->object = object;
703 rv->pindex = first;
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",
711 rv));
712 index = VM_RESERV_INDEX(object, pindex);
713 vm_reserv_populate(rv, index);
714 return (&rv->pages[index]);
715
716 /*
717 * Found a matching reservation.
718 */
719 found:
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))
724 return (NULL);
725 vm_reserv_populate(rv, index);
726 return (m);
727 }
728
729 /*
730 * Breaks the given reservation. All free pages in the reservation
731 * are returned to the physical memory allocator. The reservation's
732 * population count and map are reset to their initial state.
733 *
734 * The given reservation must not be in the partially populated reservation
735 * queue. The free page queue lock must be held.
736 */
737 static void
738 vm_reserv_break(vm_reserv_t rv)
739 {
740 int begin_zeroes, hi, i, lo;
741
742 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
743 KASSERT(rv->object != NULL,
744 ("vm_reserv_break: reserv %p is free", rv));
745 KASSERT(!rv->inpartpopq,
746 ("vm_reserv_break: reserv %p's inpartpopq is TRUE", rv));
747 LIST_REMOVE(rv, objq);
748 rv->object = NULL;
749 rv->pages->psind = 0;
750 i = hi = 0;
751 do {
752 /* Find the next 0 bit. Any previous 0 bits are < "hi". */
753 lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
754 if (lo == 0) {
755 /* Redundantly clears bits < "hi". */
756 rv->popmap[i] = 0;
757 rv->popcnt -= NBPOPMAP - hi;
758 while (++i < NPOPMAP) {
759 lo = ffsl(~rv->popmap[i]);
760 if (lo == 0) {
761 rv->popmap[i] = 0;
762 rv->popcnt -= NBPOPMAP;
763 } else
764 break;
765 }
766 if (i == NPOPMAP)
767 break;
768 hi = 0;
769 }
770 KASSERT(lo > 0, ("vm_reserv_break: lo is %d", lo));
771 /* Convert from ffsl() to ordinary bit numbering. */
772 lo--;
773 if (lo > 0) {
774 /* Redundantly clears bits < "hi". */
775 rv->popmap[i] &= ~((1UL << lo) - 1);
776 rv->popcnt -= lo - hi;
777 }
778 begin_zeroes = NBPOPMAP * i + lo;
779 /* Find the next 1 bit. */
780 do
781 hi = ffsl(rv->popmap[i]);
782 while (hi == 0 && ++i < NPOPMAP);
783 if (i != NPOPMAP)
784 /* Convert from ffsl() to ordinary bit numbering. */
785 hi--;
786 vm_phys_free_contig(&rv->pages[begin_zeroes], NBPOPMAP * i +
787 hi - begin_zeroes);
788 } while (i < NPOPMAP);
789 KASSERT(rv->popcnt == 0,
790 ("vm_reserv_break: reserv %p's popcnt is corrupted", rv));
791 vm_reserv_broken++;
792 }
793
794 /*
795 * Breaks all reservations belonging to the given object.
796 */
797 void
798 vm_reserv_break_all(vm_object_t object)
799 {
800 vm_reserv_t rv;
801
802 mtx_lock(&vm_page_queue_free_mtx);
803 while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
804 KASSERT(rv->object == object,
805 ("vm_reserv_break_all: reserv %p is corrupted", rv));
806 if (rv->inpartpopq) {
807 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
808 rv->inpartpopq = FALSE;
809 }
810 vm_reserv_break(rv);
811 }
812 mtx_unlock(&vm_page_queue_free_mtx);
813 }
814
815 /*
816 * Frees the given page if it belongs to a reservation. Returns TRUE if the
817 * page is freed and FALSE otherwise.
818 *
819 * The free page queue lock must be held.
820 */
821 boolean_t
822 vm_reserv_free_page(vm_page_t m)
823 {
824 vm_reserv_t rv;
825
826 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
827 rv = vm_reserv_from_page(m);
828 if (rv->object == NULL)
829 return (FALSE);
830 vm_reserv_depopulate(rv, m - rv->pages);
831 return (TRUE);
832 }
833
834 /*
835 * Initializes the reservation management system. Specifically, initializes
836 * the reservation array.
837 *
838 * Requires that vm_page_array and first_page are initialized!
839 */
840 void
841 vm_reserv_init(void)
842 {
843 vm_paddr_t paddr;
844 struct vm_phys_seg *seg;
845 int segind;
846
847 /*
848 * Initialize the reservation array. Specifically, initialize the
849 * "pages" field for every element that has an underlying superpage.
850 */
851 for (segind = 0; segind < vm_phys_nsegs; segind++) {
852 seg = &vm_phys_segs[segind];
853 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
854 while (paddr + VM_LEVEL_0_SIZE <= seg->end) {
855 vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages =
856 PHYS_TO_VM_PAGE(paddr);
857 paddr += VM_LEVEL_0_SIZE;
858 }
859 }
860 }
861
862 /*
863 * Returns true if the given page belongs to a reservation and that page is
864 * free. Otherwise, returns false.
865 */
866 bool
867 vm_reserv_is_page_free(vm_page_t m)
868 {
869 vm_reserv_t rv;
870
871 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
872 rv = vm_reserv_from_page(m);
873 if (rv->object == NULL)
874 return (false);
875 return (popmap_is_clear(rv->popmap, m - rv->pages));
876 }
877
878 /*
879 * If the given page belongs to a reservation, returns the level of that
880 * reservation. Otherwise, returns -1.
881 */
882 int
883 vm_reserv_level(vm_page_t m)
884 {
885 vm_reserv_t rv;
886
887 rv = vm_reserv_from_page(m);
888 return (rv->object != NULL ? 0 : -1);
889 }
890
891 /*
892 * Returns a reservation level if the given page belongs to a fully populated
893 * reservation and -1 otherwise.
894 */
895 int
896 vm_reserv_level_iffullpop(vm_page_t m)
897 {
898 vm_reserv_t rv;
899
900 rv = vm_reserv_from_page(m);
901 return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
902 }
903
904 /*
905 * Breaks the given partially populated reservation, releasing its free pages
906 * to the physical memory allocator.
907 *
908 * The free page queue lock must be held.
909 */
910 static void
911 vm_reserv_reclaim(vm_reserv_t rv)
912 {
913
914 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
915 KASSERT(rv->inpartpopq,
916 ("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
917 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
918 rv->inpartpopq = FALSE;
919 vm_reserv_break(rv);
920 vm_reserv_reclaimed++;
921 }
922
923 /*
924 * Breaks the reservation at the head of the partially populated reservation
925 * queue, releasing its free pages to the physical memory allocator. Returns
926 * TRUE if a reservation is broken and FALSE otherwise.
927 *
928 * The free page queue lock must be held.
929 */
930 boolean_t
931 vm_reserv_reclaim_inactive(void)
932 {
933 vm_reserv_t rv;
934
935 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
936 if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) {
937 vm_reserv_reclaim(rv);
938 return (TRUE);
939 }
940 return (FALSE);
941 }
942
943 /*
944 * Searches the partially populated reservation queue for the least recently
945 * changed reservation with free pages that satisfy the given request for
946 * contiguous physical memory. If a satisfactory reservation is found, it is
947 * broken. Returns TRUE if a reservation is broken and FALSE otherwise.
948 *
949 * The free page queue lock must be held.
950 */
951 boolean_t
952 vm_reserv_reclaim_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
953 u_long alignment, vm_paddr_t boundary)
954 {
955 vm_paddr_t pa, size;
956 vm_reserv_t rv;
957 int hi, i, lo, low_index, next_free;
958
959 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
960 if (npages > VM_LEVEL_0_NPAGES - 1)
961 return (FALSE);
962 size = npages << PAGE_SHIFT;
963 TAILQ_FOREACH(rv, &vm_rvq_partpop, partpopq) {
964 pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]);
965 if (pa + PAGE_SIZE - size < low) {
966 /* This entire reservation is too low; go to next. */
967 continue;
968 }
969 pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
970 if (pa + size > high) {
971 /* This entire reservation is too high; go to next. */
972 continue;
973 }
974 if (pa < low) {
975 /* Start the search for free pages at "low". */
976 low_index = (low + PAGE_MASK - pa) >> PAGE_SHIFT;
977 i = low_index / NBPOPMAP;
978 hi = low_index % NBPOPMAP;
979 } else
980 i = hi = 0;
981 do {
982 /* Find the next free page. */
983 lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
984 while (lo == 0 && ++i < NPOPMAP)
985 lo = ffsl(~rv->popmap[i]);
986 if (i == NPOPMAP)
987 break;
988 /* Convert from ffsl() to ordinary bit numbering. */
989 lo--;
990 next_free = NBPOPMAP * i + lo;
991 pa = VM_PAGE_TO_PHYS(&rv->pages[next_free]);
992 KASSERT(pa >= low,
993 ("vm_reserv_reclaim_contig: pa is too low"));
994 if (pa + size > high) {
995 /* The rest of this reservation is too high. */
996 break;
997 } else if ((pa & (alignment - 1)) != 0 ||
998 ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) {
999 /*
1000 * The current page doesn't meet the alignment
1001 * and/or boundary requirements. Continue
1002 * searching this reservation until the rest
1003 * of its free pages are either excluded or
1004 * exhausted.
1005 */
1006 hi = lo + 1;
1007 if (hi >= NBPOPMAP) {
1008 hi = 0;
1009 i++;
1010 }
1011 continue;
1012 }
1013 /* Find the next used page. */
1014 hi = ffsl(rv->popmap[i] & ~((1UL << lo) - 1));
1015 while (hi == 0 && ++i < NPOPMAP) {
1016 if ((NBPOPMAP * i - next_free) * PAGE_SIZE >=
1017 size) {
1018 vm_reserv_reclaim(rv);
1019 return (TRUE);
1020 }
1021 hi = ffsl(rv->popmap[i]);
1022 }
1023 /* Convert from ffsl() to ordinary bit numbering. */
1024 if (i != NPOPMAP)
1025 hi--;
1026 if ((NBPOPMAP * i + hi - next_free) * PAGE_SIZE >=
1027 size) {
1028 vm_reserv_reclaim(rv);
1029 return (TRUE);
1030 }
1031 } while (i < NPOPMAP);
1032 }
1033 return (FALSE);
1034 }
1035
1036 /*
1037 * Transfers the reservation underlying the given page to a new object.
1038 *
1039 * The object must be locked.
1040 */
1041 void
1042 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
1043 vm_pindex_t old_object_offset)
1044 {
1045 vm_reserv_t rv;
1046
1047 VM_OBJECT_ASSERT_WLOCKED(new_object);
1048 rv = vm_reserv_from_page(m);
1049 if (rv->object == old_object) {
1050 mtx_lock(&vm_page_queue_free_mtx);
1051 if (rv->object == old_object) {
1052 LIST_REMOVE(rv, objq);
1053 LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
1054 rv->object = new_object;
1055 rv->pindex -= old_object_offset;
1056 }
1057 mtx_unlock(&vm_page_queue_free_mtx);
1058 }
1059 }
1060
1061 /*
1062 * Returns the size (in bytes) of a reservation of the specified level.
1063 */
1064 int
1065 vm_reserv_size(int level)
1066 {
1067
1068 switch (level) {
1069 case 0:
1070 return (VM_LEVEL_0_SIZE);
1071 case -1:
1072 return (PAGE_SIZE);
1073 default:
1074 return (0);
1075 }
1076 }
1077
1078 /*
1079 * Allocates the virtual and physical memory required by the reservation
1080 * management system's data structures, in particular, the reservation array.
1081 */
1082 vm_paddr_t
1083 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
1084 {
1085 vm_paddr_t new_end;
1086 size_t size;
1087
1088 /*
1089 * Calculate the size (in bytes) of the reservation array. Round up
1090 * from "high_water" because every small page is mapped to an element
1091 * in the reservation array based on its physical address. Thus, the
1092 * number of elements in the reservation array can be greater than the
1093 * number of superpages.
1094 */
1095 size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
1096
1097 /*
1098 * Allocate and map the physical memory for the reservation array. The
1099 * next available virtual address is returned by reference.
1100 */
1101 new_end = end - round_page(size);
1102 vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
1103 VM_PROT_READ | VM_PROT_WRITE);
1104 bzero(vm_reserv_array, size);
1105
1106 /*
1107 * Return the next available physical address.
1108 */
1109 return (new_end);
1110 }
1111
1112 /*
1113 * Returns the superpage containing the given page.
1114 */
1115 vm_page_t
1116 vm_reserv_to_superpage(vm_page_t m)
1117 {
1118 vm_reserv_t rv;
1119
1120 VM_OBJECT_ASSERT_LOCKED(m->object);
1121 rv = vm_reserv_from_page(m);
1122 return (rv->object == m->object && rv->popcnt == VM_LEVEL_0_NPAGES ?
1123 rv->pages : NULL);
1124 }
1125
1126 #endif /* VM_NRESERVLEVEL > 0 */
Cache object: a480564175fae3eedc331ae78cab602c
|