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/11.1/sys/vm/vm_reserv.c 318716 2017-05-23 07:27:30Z markj $");
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
55 #include <vm/vm.h>
56 #include <vm/vm_param.h>
57 #include <vm/vm_object.h>
58 #include <vm/vm_page.h>
59 #include <vm/vm_phys.h>
60 #include <vm/vm_radix.h>
61 #include <vm/vm_reserv.h>
62
63 /*
64 * The reservation system supports the speculative allocation of large physical
65 * pages ("superpages"). Speculative allocation enables the fully automatic
66 * utilization of superpages by the virtual memory system. In other words, no
67 * programmatic directives are required to use superpages.
68 */
69
70 #if VM_NRESERVLEVEL > 0
71
72 /*
73 * The number of small pages that are contained in a level 0 reservation
74 */
75 #define VM_LEVEL_0_NPAGES (1 << VM_LEVEL_0_ORDER)
76
77 /*
78 * The number of bits by which a physical address is shifted to obtain the
79 * reservation number
80 */
81 #define VM_LEVEL_0_SHIFT (VM_LEVEL_0_ORDER + PAGE_SHIFT)
82
83 /*
84 * The size of a level 0 reservation in bytes
85 */
86 #define VM_LEVEL_0_SIZE (1 << VM_LEVEL_0_SHIFT)
87
88 /*
89 * Computes the index of the small page underlying the given (object, pindex)
90 * within the reservation's array of small pages.
91 */
92 #define VM_RESERV_INDEX(object, pindex) \
93 (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))
94
95 /*
96 * The size of a population map entry
97 */
98 typedef u_long popmap_t;
99
100 /*
101 * The number of bits in a population map entry
102 */
103 #define NBPOPMAP (NBBY * sizeof(popmap_t))
104
105 /*
106 * The number of population map entries in a reservation
107 */
108 #define NPOPMAP howmany(VM_LEVEL_0_NPAGES, NBPOPMAP)
109
110 /*
111 * Clear a bit in the population map.
112 */
113 static __inline void
114 popmap_clear(popmap_t popmap[], int i)
115 {
116
117 popmap[i / NBPOPMAP] &= ~(1UL << (i % NBPOPMAP));
118 }
119
120 /*
121 * Set a bit in the population map.
122 */
123 static __inline void
124 popmap_set(popmap_t popmap[], int i)
125 {
126
127 popmap[i / NBPOPMAP] |= 1UL << (i % NBPOPMAP);
128 }
129
130 /*
131 * Is a bit in the population map clear?
132 */
133 static __inline boolean_t
134 popmap_is_clear(popmap_t popmap[], int i)
135 {
136
137 return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) == 0);
138 }
139
140 /*
141 * Is a bit in the population map set?
142 */
143 static __inline boolean_t
144 popmap_is_set(popmap_t popmap[], int i)
145 {
146
147 return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) != 0);
148 }
149
150 /*
151 * The reservation structure
152 *
153 * A reservation structure is constructed whenever a large physical page is
154 * speculatively allocated to an object. The reservation provides the small
155 * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
156 * within that object. The reservation's "popcnt" tracks the number of these
157 * small physical pages that are in use at any given time. When and if the
158 * reservation is not fully utilized, it appears in the queue of partially
159 * populated reservations. The reservation always appears on the containing
160 * object's list of reservations.
161 *
162 * A partially populated reservation can be broken and reclaimed at any time.
163 */
164 struct vm_reserv {
165 TAILQ_ENTRY(vm_reserv) partpopq;
166 LIST_ENTRY(vm_reserv) objq;
167 vm_object_t object; /* containing object */
168 vm_pindex_t pindex; /* offset within object */
169 vm_page_t pages; /* first page of a superpage */
170 int popcnt; /* # of pages in use */
171 char inpartpopq;
172 popmap_t popmap[NPOPMAP]; /* bit vector of used pages */
173 };
174
175 /*
176 * The reservation array
177 *
178 * This array is analoguous in function to vm_page_array. It differs in the
179 * respect that it may contain a greater number of useful reservation
180 * structures than there are (physical) superpages. These "invalid"
181 * reservation structures exist to trade-off space for time in the
182 * implementation of vm_reserv_from_page(). Invalid reservation structures are
183 * distinguishable from "valid" reservation structures by inspecting the
184 * reservation's "pages" field. Invalid reservation structures have a NULL
185 * "pages" field.
186 *
187 * vm_reserv_from_page() maps a small (physical) page to an element of this
188 * array by computing a physical reservation number from the page's physical
189 * address. The physical reservation number is used as the array index.
190 *
191 * An "active" reservation is a valid reservation structure that has a non-NULL
192 * "object" field and a non-zero "popcnt" field. In other words, every active
193 * reservation belongs to a particular object. Moreover, every active
194 * reservation has an entry in the containing object's list of reservations.
195 */
196 static vm_reserv_t vm_reserv_array;
197
198 /*
199 * The partially populated reservation queue
200 *
201 * This queue enables the fast recovery of an unused free small page from a
202 * partially populated reservation. The reservation at the head of this queue
203 * is the least recently changed, partially populated reservation.
204 *
205 * Access to this queue is synchronized by the free page queue lock.
206 */
207 static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop =
208 TAILQ_HEAD_INITIALIZER(vm_rvq_partpop);
209
210 static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");
211
212 static long vm_reserv_broken;
213 SYSCTL_LONG(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
214 &vm_reserv_broken, 0, "Cumulative number of broken reservations");
215
216 static long vm_reserv_freed;
217 SYSCTL_LONG(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
218 &vm_reserv_freed, 0, "Cumulative number of freed reservations");
219
220 static int sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS);
221
222 SYSCTL_PROC(_vm_reserv, OID_AUTO, fullpop, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
223 sysctl_vm_reserv_fullpop, "I", "Current number of full reservations");
224
225 static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);
226
227 SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
228 sysctl_vm_reserv_partpopq, "A", "Partially populated reservation queues");
229
230 static long vm_reserv_reclaimed;
231 SYSCTL_LONG(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
232 &vm_reserv_reclaimed, 0, "Cumulative number of reclaimed reservations");
233
234 static void vm_reserv_break(vm_reserv_t rv, vm_page_t m);
235 static void vm_reserv_depopulate(vm_reserv_t rv, int index);
236 static vm_reserv_t vm_reserv_from_page(vm_page_t m);
237 static boolean_t vm_reserv_has_pindex(vm_reserv_t rv,
238 vm_pindex_t pindex);
239 static void vm_reserv_populate(vm_reserv_t rv, int index);
240 static void vm_reserv_reclaim(vm_reserv_t rv);
241
242 /*
243 * Returns the current number of full reservations.
244 *
245 * Since the number of full reservations is computed without acquiring the
246 * free page queue lock, the returned value may be inexact.
247 */
248 static int
249 sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS)
250 {
251 vm_paddr_t paddr;
252 struct vm_phys_seg *seg;
253 vm_reserv_t rv;
254 int fullpop, segind;
255
256 fullpop = 0;
257 for (segind = 0; segind < vm_phys_nsegs; segind++) {
258 seg = &vm_phys_segs[segind];
259 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
260 while (paddr + VM_LEVEL_0_SIZE <= seg->end) {
261 rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
262 fullpop += rv->popcnt == VM_LEVEL_0_NPAGES;
263 paddr += VM_LEVEL_0_SIZE;
264 }
265 }
266 return (sysctl_handle_int(oidp, &fullpop, 0, req));
267 }
268
269 /*
270 * Describes the current state of the partially populated reservation queue.
271 */
272 static int
273 sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
274 {
275 struct sbuf sbuf;
276 vm_reserv_t rv;
277 int counter, error, level, unused_pages;
278
279 error = sysctl_wire_old_buffer(req, 0);
280 if (error != 0)
281 return (error);
282 sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
283 sbuf_printf(&sbuf, "\nLEVEL SIZE NUMBER\n\n");
284 for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
285 counter = 0;
286 unused_pages = 0;
287 mtx_lock(&vm_page_queue_free_mtx);
288 TAILQ_FOREACH(rv, &vm_rvq_partpop/*[level]*/, partpopq) {
289 counter++;
290 unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
291 }
292 mtx_unlock(&vm_page_queue_free_mtx);
293 sbuf_printf(&sbuf, "%5d: %6dK, %6d\n", level,
294 unused_pages * ((int)PAGE_SIZE / 1024), counter);
295 }
296 error = sbuf_finish(&sbuf);
297 sbuf_delete(&sbuf);
298 return (error);
299 }
300
301 /*
302 * Reduces the given reservation's population count. If the population count
303 * becomes zero, the reservation is destroyed. Additionally, moves the
304 * reservation to the tail of the partially populated reservation queue if the
305 * population count is non-zero.
306 *
307 * The free page queue lock must be held.
308 */
309 static void
310 vm_reserv_depopulate(vm_reserv_t rv, int index)
311 {
312
313 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
314 KASSERT(rv->object != NULL,
315 ("vm_reserv_depopulate: reserv %p is free", rv));
316 KASSERT(popmap_is_set(rv->popmap, index),
317 ("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
318 index));
319 KASSERT(rv->popcnt > 0,
320 ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
321 if (rv->inpartpopq) {
322 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
323 rv->inpartpopq = FALSE;
324 } else {
325 KASSERT(rv->pages->psind == 1,
326 ("vm_reserv_depopulate: reserv %p is already demoted",
327 rv));
328 rv->pages->psind = 0;
329 }
330 popmap_clear(rv->popmap, index);
331 rv->popcnt--;
332 if (rv->popcnt == 0) {
333 LIST_REMOVE(rv, objq);
334 rv->object = NULL;
335 vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
336 vm_reserv_freed++;
337 } else {
338 rv->inpartpopq = TRUE;
339 TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
340 }
341 }
342
343 /*
344 * Returns the reservation to which the given page might belong.
345 */
346 static __inline vm_reserv_t
347 vm_reserv_from_page(vm_page_t m)
348 {
349
350 return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
351 }
352
353 /*
354 * Returns TRUE if the given reservation contains the given page index and
355 * FALSE otherwise.
356 */
357 static __inline boolean_t
358 vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
359 {
360
361 return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
362 }
363
364 /*
365 * Increases the given reservation's population count. Moves the reservation
366 * to the tail of the partially populated reservation queue.
367 *
368 * The free page queue must be locked.
369 */
370 static void
371 vm_reserv_populate(vm_reserv_t rv, int index)
372 {
373
374 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
375 KASSERT(rv->object != NULL,
376 ("vm_reserv_populate: reserv %p is free", rv));
377 KASSERT(popmap_is_clear(rv->popmap, index),
378 ("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
379 index));
380 KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
381 ("vm_reserv_populate: reserv %p is already full", rv));
382 KASSERT(rv->pages->psind == 0,
383 ("vm_reserv_populate: reserv %p is already promoted", rv));
384 if (rv->inpartpopq) {
385 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
386 rv->inpartpopq = FALSE;
387 }
388 popmap_set(rv->popmap, index);
389 rv->popcnt++;
390 if (rv->popcnt < VM_LEVEL_0_NPAGES) {
391 rv->inpartpopq = TRUE;
392 TAILQ_INSERT_TAIL(&vm_rvq_partpop, rv, partpopq);
393 } else
394 rv->pages->psind = 1;
395 }
396
397 /*
398 * Allocates a contiguous set of physical pages of the given size "npages"
399 * from existing or newly created reservations. All of the physical pages
400 * must be at or above the given physical address "low" and below the given
401 * physical address "high". The given value "alignment" determines the
402 * alignment of the first physical page in the set. If the given value
403 * "boundary" is non-zero, then the set of physical pages cannot cross any
404 * physical address boundary that is a multiple of that value. Both
405 * "alignment" and "boundary" must be a power of two.
406 *
407 * The page "mpred" must immediately precede the offset "pindex" within the
408 * specified object.
409 *
410 * The object and free page queue must be locked.
411 */
412 vm_page_t
413 vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, u_long npages,
414 vm_paddr_t low, vm_paddr_t high, u_long alignment, vm_paddr_t boundary,
415 vm_page_t mpred)
416 {
417 vm_paddr_t pa, size;
418 vm_page_t m, m_ret, msucc;
419 vm_pindex_t first, leftcap, rightcap;
420 vm_reserv_t rv;
421 u_long allocpages, maxpages, minpages;
422 int i, index, n;
423
424 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
425 VM_OBJECT_ASSERT_WLOCKED(object);
426 KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
427
428 /*
429 * Is a reservation fundamentally impossible?
430 */
431 if (pindex < VM_RESERV_INDEX(object, pindex) ||
432 pindex + npages > object->size)
433 return (NULL);
434
435 /*
436 * All reservations of a particular size have the same alignment.
437 * Assuming that the first page is allocated from a reservation, the
438 * least significant bits of its physical address can be determined
439 * from its offset from the beginning of the reservation and the size
440 * of the reservation.
441 *
442 * Could the specified index within a reservation of the smallest
443 * possible size satisfy the alignment and boundary requirements?
444 */
445 pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
446 if ((pa & (alignment - 1)) != 0)
447 return (NULL);
448 size = npages << PAGE_SHIFT;
449 if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
450 return (NULL);
451
452 /*
453 * Look for an existing reservation.
454 */
455 if (mpred != NULL) {
456 KASSERT(mpred->object == object,
457 ("vm_reserv_alloc_contig: object doesn't contain mpred"));
458 KASSERT(mpred->pindex < pindex,
459 ("vm_reserv_alloc_contig: mpred doesn't precede pindex"));
460 rv = vm_reserv_from_page(mpred);
461 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
462 goto found;
463 msucc = TAILQ_NEXT(mpred, listq);
464 } else
465 msucc = TAILQ_FIRST(&object->memq);
466 if (msucc != NULL) {
467 KASSERT(msucc->pindex > pindex,
468 ("vm_reserv_alloc_contig: msucc doesn't succeed pindex"));
469 rv = vm_reserv_from_page(msucc);
470 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
471 goto found;
472 }
473
474 /*
475 * Could at least one reservation fit between the first index to the
476 * left that can be used ("leftcap") and the first index to the right
477 * that cannot be used ("rightcap")?
478 */
479 first = pindex - VM_RESERV_INDEX(object, pindex);
480 if (mpred != NULL) {
481 if ((rv = vm_reserv_from_page(mpred))->object != object)
482 leftcap = mpred->pindex + 1;
483 else
484 leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
485 if (leftcap > first)
486 return (NULL);
487 }
488 minpages = VM_RESERV_INDEX(object, pindex) + npages;
489 maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
490 allocpages = maxpages;
491 if (msucc != NULL) {
492 if ((rv = vm_reserv_from_page(msucc))->object != object)
493 rightcap = msucc->pindex;
494 else
495 rightcap = rv->pindex;
496 if (first + maxpages > rightcap) {
497 if (maxpages == VM_LEVEL_0_NPAGES)
498 return (NULL);
499
500 /*
501 * At least one reservation will fit between "leftcap"
502 * and "rightcap". However, a reservation for the
503 * last of the requested pages will not fit. Reduce
504 * the size of the upcoming allocation accordingly.
505 */
506 allocpages = minpages;
507 }
508 }
509
510 /*
511 * Would the last new reservation extend past the end of the object?
512 */
513 if (first + maxpages > object->size) {
514 /*
515 * Don't allocate the last new reservation if the object is a
516 * vnode or backed by another object that is a vnode.
517 */
518 if (object->type == OBJT_VNODE ||
519 (object->backing_object != NULL &&
520 object->backing_object->type == OBJT_VNODE)) {
521 if (maxpages == VM_LEVEL_0_NPAGES)
522 return (NULL);
523 allocpages = minpages;
524 }
525 /* Speculate that the object may grow. */
526 }
527
528 /*
529 * Allocate the physical pages. The alignment and boundary specified
530 * for this allocation may be different from the alignment and
531 * boundary specified for the requested pages. For instance, the
532 * specified index may not be the first page within the first new
533 * reservation.
534 */
535 m = vm_phys_alloc_contig(allocpages, low, high, ulmax(alignment,
536 VM_LEVEL_0_SIZE), boundary > VM_LEVEL_0_SIZE ? boundary : 0);
537 if (m == NULL)
538 return (NULL);
539
540 /*
541 * The allocated physical pages always begin at a reservation
542 * boundary, but they do not always end at a reservation boundary.
543 * Initialize every reservation that is completely covered by the
544 * allocated physical pages.
545 */
546 m_ret = NULL;
547 index = VM_RESERV_INDEX(object, pindex);
548 do {
549 rv = vm_reserv_from_page(m);
550 KASSERT(rv->pages == m,
551 ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
552 rv));
553 KASSERT(rv->object == NULL,
554 ("vm_reserv_alloc_contig: reserv %p isn't free", rv));
555 LIST_INSERT_HEAD(&object->rvq, rv, objq);
556 rv->object = object;
557 rv->pindex = first;
558 KASSERT(rv->popcnt == 0,
559 ("vm_reserv_alloc_contig: reserv %p's popcnt is corrupted",
560 rv));
561 KASSERT(!rv->inpartpopq,
562 ("vm_reserv_alloc_contig: reserv %p's inpartpopq is TRUE",
563 rv));
564 for (i = 0; i < NPOPMAP; i++)
565 KASSERT(rv->popmap[i] == 0,
566 ("vm_reserv_alloc_contig: reserv %p's popmap is corrupted",
567 rv));
568 n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
569 for (i = 0; i < n; i++)
570 vm_reserv_populate(rv, index + i);
571 npages -= n;
572 if (m_ret == NULL) {
573 m_ret = &rv->pages[index];
574 index = 0;
575 }
576 m += VM_LEVEL_0_NPAGES;
577 first += VM_LEVEL_0_NPAGES;
578 allocpages -= VM_LEVEL_0_NPAGES;
579 } while (allocpages >= VM_LEVEL_0_NPAGES);
580 return (m_ret);
581
582 /*
583 * Found a matching reservation.
584 */
585 found:
586 index = VM_RESERV_INDEX(object, pindex);
587 /* Does the allocation fit within the reservation? */
588 if (index + npages > VM_LEVEL_0_NPAGES)
589 return (NULL);
590 m = &rv->pages[index];
591 pa = VM_PAGE_TO_PHYS(m);
592 if (pa < low || pa + size > high || (pa & (alignment - 1)) != 0 ||
593 ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
594 return (NULL);
595 /* Handle vm_page_rename(m, new_object, ...). */
596 for (i = 0; i < npages; i++)
597 if (popmap_is_set(rv->popmap, index + i))
598 return (NULL);
599 for (i = 0; i < npages; i++)
600 vm_reserv_populate(rv, index + i);
601 return (m);
602 }
603
604 /*
605 * Allocates a page from an existing or newly created reservation.
606 *
607 * The page "mpred" must immediately precede the offset "pindex" within the
608 * specified object.
609 *
610 * The object and free page queue must be locked.
611 */
612 vm_page_t
613 vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, vm_page_t mpred)
614 {
615 vm_page_t m, msucc;
616 vm_pindex_t first, leftcap, rightcap;
617 vm_reserv_t rv;
618 int i, index;
619
620 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
621 VM_OBJECT_ASSERT_WLOCKED(object);
622
623 /*
624 * Is a reservation fundamentally impossible?
625 */
626 if (pindex < VM_RESERV_INDEX(object, pindex) ||
627 pindex >= object->size)
628 return (NULL);
629
630 /*
631 * Look for an existing reservation.
632 */
633 if (mpred != NULL) {
634 KASSERT(mpred->object == object,
635 ("vm_reserv_alloc_page: object doesn't contain mpred"));
636 KASSERT(mpred->pindex < pindex,
637 ("vm_reserv_alloc_page: mpred doesn't precede pindex"));
638 rv = vm_reserv_from_page(mpred);
639 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
640 goto found;
641 msucc = TAILQ_NEXT(mpred, listq);
642 } else
643 msucc = TAILQ_FIRST(&object->memq);
644 if (msucc != NULL) {
645 KASSERT(msucc->pindex > pindex,
646 ("vm_reserv_alloc_page: msucc doesn't succeed pindex"));
647 rv = vm_reserv_from_page(msucc);
648 if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
649 goto found;
650 }
651
652 /*
653 * Could a reservation fit between the first index to the left that
654 * can be used and the first index to the right that cannot be used?
655 */
656 first = pindex - VM_RESERV_INDEX(object, pindex);
657 if (mpred != NULL) {
658 if ((rv = vm_reserv_from_page(mpred))->object != object)
659 leftcap = mpred->pindex + 1;
660 else
661 leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
662 if (leftcap > first)
663 return (NULL);
664 }
665 if (msucc != NULL) {
666 if ((rv = vm_reserv_from_page(msucc))->object != object)
667 rightcap = msucc->pindex;
668 else
669 rightcap = rv->pindex;
670 if (first + VM_LEVEL_0_NPAGES > rightcap)
671 return (NULL);
672 }
673
674 /*
675 * Would a new reservation extend past the end of the object?
676 */
677 if (first + VM_LEVEL_0_NPAGES > object->size) {
678 /*
679 * Don't allocate a new reservation if the object is a vnode or
680 * backed by another object that is a vnode.
681 */
682 if (object->type == OBJT_VNODE ||
683 (object->backing_object != NULL &&
684 object->backing_object->type == OBJT_VNODE))
685 return (NULL);
686 /* Speculate that the object may grow. */
687 }
688
689 /*
690 * Allocate and populate the new reservation.
691 */
692 m = vm_phys_alloc_pages(VM_FREEPOOL_DEFAULT, VM_LEVEL_0_ORDER);
693 if (m == NULL)
694 return (NULL);
695 rv = vm_reserv_from_page(m);
696 KASSERT(rv->pages == m,
697 ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
698 KASSERT(rv->object == NULL,
699 ("vm_reserv_alloc_page: reserv %p isn't free", rv));
700 LIST_INSERT_HEAD(&object->rvq, rv, objq);
701 rv->object = object;
702 rv->pindex = first;
703 KASSERT(rv->popcnt == 0,
704 ("vm_reserv_alloc_page: reserv %p's popcnt is corrupted", rv));
705 KASSERT(!rv->inpartpopq,
706 ("vm_reserv_alloc_page: reserv %p's inpartpopq is TRUE", rv));
707 for (i = 0; i < NPOPMAP; i++)
708 KASSERT(rv->popmap[i] == 0,
709 ("vm_reserv_alloc_page: reserv %p's popmap is corrupted",
710 rv));
711 index = VM_RESERV_INDEX(object, pindex);
712 vm_reserv_populate(rv, index);
713 return (&rv->pages[index]);
714
715 /*
716 * Found a matching reservation.
717 */
718 found:
719 index = VM_RESERV_INDEX(object, pindex);
720 m = &rv->pages[index];
721 /* Handle vm_page_rename(m, new_object, ...). */
722 if (popmap_is_set(rv->popmap, index))
723 return (NULL);
724 vm_reserv_populate(rv, index);
725 return (m);
726 }
727
728 /*
729 * Breaks the given reservation. Except for the specified free page, all free
730 * pages in the reservation are returned to the physical memory allocator.
731 * The reservation's population count and map are reset to their initial
732 * 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, vm_page_t m)
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 if (m != NULL) {
750 /*
751 * Since the reservation is being broken, there is no harm in
752 * abusing the population map to stop "m" from being returned
753 * to the physical memory allocator.
754 */
755 i = m - rv->pages;
756 KASSERT(popmap_is_clear(rv->popmap, i),
757 ("vm_reserv_break: reserv %p's popmap is corrupted", rv));
758 popmap_set(rv->popmap, i);
759 rv->popcnt++;
760 }
761 i = hi = 0;
762 do {
763 /* Find the next 0 bit. Any previous 0 bits are < "hi". */
764 lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
765 if (lo == 0) {
766 /* Redundantly clears bits < "hi". */
767 rv->popmap[i] = 0;
768 rv->popcnt -= NBPOPMAP - hi;
769 while (++i < NPOPMAP) {
770 lo = ffsl(~rv->popmap[i]);
771 if (lo == 0) {
772 rv->popmap[i] = 0;
773 rv->popcnt -= NBPOPMAP;
774 } else
775 break;
776 }
777 if (i == NPOPMAP)
778 break;
779 hi = 0;
780 }
781 KASSERT(lo > 0, ("vm_reserv_break: lo is %d", lo));
782 /* Convert from ffsl() to ordinary bit numbering. */
783 lo--;
784 if (lo > 0) {
785 /* Redundantly clears bits < "hi". */
786 rv->popmap[i] &= ~((1UL << lo) - 1);
787 rv->popcnt -= lo - hi;
788 }
789 begin_zeroes = NBPOPMAP * i + lo;
790 /* Find the next 1 bit. */
791 do
792 hi = ffsl(rv->popmap[i]);
793 while (hi == 0 && ++i < NPOPMAP);
794 if (i != NPOPMAP)
795 /* Convert from ffsl() to ordinary bit numbering. */
796 hi--;
797 vm_phys_free_contig(&rv->pages[begin_zeroes], NBPOPMAP * i +
798 hi - begin_zeroes);
799 } while (i < NPOPMAP);
800 KASSERT(rv->popcnt == 0,
801 ("vm_reserv_break: reserv %p's popcnt is corrupted", rv));
802 vm_reserv_broken++;
803 }
804
805 /*
806 * Breaks all reservations belonging to the given object.
807 */
808 void
809 vm_reserv_break_all(vm_object_t object)
810 {
811 vm_reserv_t rv;
812
813 mtx_lock(&vm_page_queue_free_mtx);
814 while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
815 KASSERT(rv->object == object,
816 ("vm_reserv_break_all: reserv %p is corrupted", rv));
817 if (rv->inpartpopq) {
818 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
819 rv->inpartpopq = FALSE;
820 }
821 vm_reserv_break(rv, NULL);
822 }
823 mtx_unlock(&vm_page_queue_free_mtx);
824 }
825
826 /*
827 * Frees the given page if it belongs to a reservation. Returns TRUE if the
828 * page is freed and FALSE otherwise.
829 *
830 * The free page queue lock must be held.
831 */
832 boolean_t
833 vm_reserv_free_page(vm_page_t m)
834 {
835 vm_reserv_t rv;
836
837 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
838 rv = vm_reserv_from_page(m);
839 if (rv->object == NULL)
840 return (FALSE);
841 vm_reserv_depopulate(rv, m - rv->pages);
842 return (TRUE);
843 }
844
845 /*
846 * Initializes the reservation management system. Specifically, initializes
847 * the reservation array.
848 *
849 * Requires that vm_page_array and first_page are initialized!
850 */
851 void
852 vm_reserv_init(void)
853 {
854 vm_paddr_t paddr;
855 struct vm_phys_seg *seg;
856 int segind;
857
858 /*
859 * Initialize the reservation array. Specifically, initialize the
860 * "pages" field for every element that has an underlying superpage.
861 */
862 for (segind = 0; segind < vm_phys_nsegs; segind++) {
863 seg = &vm_phys_segs[segind];
864 paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
865 while (paddr + VM_LEVEL_0_SIZE <= seg->end) {
866 vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT].pages =
867 PHYS_TO_VM_PAGE(paddr);
868 paddr += VM_LEVEL_0_SIZE;
869 }
870 }
871 }
872
873 /*
874 * Returns true if the given page belongs to a reservation and that page is
875 * free. Otherwise, returns false.
876 */
877 bool
878 vm_reserv_is_page_free(vm_page_t m)
879 {
880 vm_reserv_t rv;
881
882 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
883 rv = vm_reserv_from_page(m);
884 if (rv->object == NULL)
885 return (false);
886 return (popmap_is_clear(rv->popmap, m - rv->pages));
887 }
888
889 /*
890 * If the given page belongs to a reservation, returns the level of that
891 * reservation. Otherwise, returns -1.
892 */
893 int
894 vm_reserv_level(vm_page_t m)
895 {
896 vm_reserv_t rv;
897
898 rv = vm_reserv_from_page(m);
899 return (rv->object != NULL ? 0 : -1);
900 }
901
902 /*
903 * Returns a reservation level if the given page belongs to a fully populated
904 * reservation and -1 otherwise.
905 */
906 int
907 vm_reserv_level_iffullpop(vm_page_t m)
908 {
909 vm_reserv_t rv;
910
911 rv = vm_reserv_from_page(m);
912 return (rv->popcnt == VM_LEVEL_0_NPAGES ? 0 : -1);
913 }
914
915 /*
916 * Breaks the given partially populated reservation, releasing its free pages
917 * to the physical memory allocator.
918 *
919 * The free page queue lock must be held.
920 */
921 static void
922 vm_reserv_reclaim(vm_reserv_t rv)
923 {
924
925 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
926 KASSERT(rv->inpartpopq,
927 ("vm_reserv_reclaim: reserv %p's inpartpopq is FALSE", rv));
928 TAILQ_REMOVE(&vm_rvq_partpop, rv, partpopq);
929 rv->inpartpopq = FALSE;
930 vm_reserv_break(rv, NULL);
931 vm_reserv_reclaimed++;
932 }
933
934 /*
935 * Breaks the reservation at the head of the partially populated reservation
936 * queue, releasing its free pages to the physical memory allocator. Returns
937 * TRUE if a reservation is broken and FALSE otherwise.
938 *
939 * The free page queue lock must be held.
940 */
941 boolean_t
942 vm_reserv_reclaim_inactive(void)
943 {
944 vm_reserv_t rv;
945
946 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
947 if ((rv = TAILQ_FIRST(&vm_rvq_partpop)) != NULL) {
948 vm_reserv_reclaim(rv);
949 return (TRUE);
950 }
951 return (FALSE);
952 }
953
954 /*
955 * Searches the partially populated reservation queue for the least recently
956 * changed reservation with free pages that satisfy the given request for
957 * contiguous physical memory. If a satisfactory reservation is found, it is
958 * broken. Returns TRUE if a reservation is broken and FALSE otherwise.
959 *
960 * The free page queue lock must be held.
961 */
962 boolean_t
963 vm_reserv_reclaim_contig(u_long npages, vm_paddr_t low, vm_paddr_t high,
964 u_long alignment, vm_paddr_t boundary)
965 {
966 vm_paddr_t pa, size;
967 vm_reserv_t rv;
968 int hi, i, lo, low_index, next_free;
969
970 mtx_assert(&vm_page_queue_free_mtx, MA_OWNED);
971 if (npages > VM_LEVEL_0_NPAGES - 1)
972 return (FALSE);
973 size = npages << PAGE_SHIFT;
974 TAILQ_FOREACH(rv, &vm_rvq_partpop, partpopq) {
975 pa = VM_PAGE_TO_PHYS(&rv->pages[VM_LEVEL_0_NPAGES - 1]);
976 if (pa + PAGE_SIZE - size < low) {
977 /* This entire reservation is too low; go to next. */
978 continue;
979 }
980 pa = VM_PAGE_TO_PHYS(&rv->pages[0]);
981 if (pa + size > high) {
982 /* This entire reservation is too high; go to next. */
983 continue;
984 }
985 if (pa < low) {
986 /* Start the search for free pages at "low". */
987 low_index = (low + PAGE_MASK - pa) >> PAGE_SHIFT;
988 i = low_index / NBPOPMAP;
989 hi = low_index % NBPOPMAP;
990 } else
991 i = hi = 0;
992 do {
993 /* Find the next free page. */
994 lo = ffsl(~(((1UL << hi) - 1) | rv->popmap[i]));
995 while (lo == 0 && ++i < NPOPMAP)
996 lo = ffsl(~rv->popmap[i]);
997 if (i == NPOPMAP)
998 break;
999 /* Convert from ffsl() to ordinary bit numbering. */
1000 lo--;
1001 next_free = NBPOPMAP * i + lo;
1002 pa = VM_PAGE_TO_PHYS(&rv->pages[next_free]);
1003 KASSERT(pa >= low,
1004 ("vm_reserv_reclaim_contig: pa is too low"));
1005 if (pa + size > high) {
1006 /* The rest of this reservation is too high. */
1007 break;
1008 } else if ((pa & (alignment - 1)) != 0 ||
1009 ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0) {
1010 /*
1011 * The current page doesn't meet the alignment
1012 * and/or boundary requirements. Continue
1013 * searching this reservation until the rest
1014 * of its free pages are either excluded or
1015 * exhausted.
1016 */
1017 hi = lo + 1;
1018 if (hi >= NBPOPMAP) {
1019 hi = 0;
1020 i++;
1021 }
1022 continue;
1023 }
1024 /* Find the next used page. */
1025 hi = ffsl(rv->popmap[i] & ~((1UL << lo) - 1));
1026 while (hi == 0 && ++i < NPOPMAP) {
1027 if ((NBPOPMAP * i - next_free) * PAGE_SIZE >=
1028 size) {
1029 vm_reserv_reclaim(rv);
1030 return (TRUE);
1031 }
1032 hi = ffsl(rv->popmap[i]);
1033 }
1034 /* Convert from ffsl() to ordinary bit numbering. */
1035 if (i != NPOPMAP)
1036 hi--;
1037 if ((NBPOPMAP * i + hi - next_free) * PAGE_SIZE >=
1038 size) {
1039 vm_reserv_reclaim(rv);
1040 return (TRUE);
1041 }
1042 } while (i < NPOPMAP);
1043 }
1044 return (FALSE);
1045 }
1046
1047 /*
1048 * Transfers the reservation underlying the given page to a new object.
1049 *
1050 * The object must be locked.
1051 */
1052 void
1053 vm_reserv_rename(vm_page_t m, vm_object_t new_object, vm_object_t old_object,
1054 vm_pindex_t old_object_offset)
1055 {
1056 vm_reserv_t rv;
1057
1058 VM_OBJECT_ASSERT_WLOCKED(new_object);
1059 rv = vm_reserv_from_page(m);
1060 if (rv->object == old_object) {
1061 mtx_lock(&vm_page_queue_free_mtx);
1062 if (rv->object == old_object) {
1063 LIST_REMOVE(rv, objq);
1064 LIST_INSERT_HEAD(&new_object->rvq, rv, objq);
1065 rv->object = new_object;
1066 rv->pindex -= old_object_offset;
1067 }
1068 mtx_unlock(&vm_page_queue_free_mtx);
1069 }
1070 }
1071
1072 /*
1073 * Returns the size (in bytes) of a reservation of the specified level.
1074 */
1075 int
1076 vm_reserv_size(int level)
1077 {
1078
1079 switch (level) {
1080 case 0:
1081 return (VM_LEVEL_0_SIZE);
1082 case -1:
1083 return (PAGE_SIZE);
1084 default:
1085 return (0);
1086 }
1087 }
1088
1089 /*
1090 * Allocates the virtual and physical memory required by the reservation
1091 * management system's data structures, in particular, the reservation array.
1092 */
1093 vm_paddr_t
1094 vm_reserv_startup(vm_offset_t *vaddr, vm_paddr_t end, vm_paddr_t high_water)
1095 {
1096 vm_paddr_t new_end;
1097 size_t size;
1098
1099 /*
1100 * Calculate the size (in bytes) of the reservation array. Round up
1101 * from "high_water" because every small page is mapped to an element
1102 * in the reservation array based on its physical address. Thus, the
1103 * number of elements in the reservation array can be greater than the
1104 * number of superpages.
1105 */
1106 size = howmany(high_water, VM_LEVEL_0_SIZE) * sizeof(struct vm_reserv);
1107
1108 /*
1109 * Allocate and map the physical memory for the reservation array. The
1110 * next available virtual address is returned by reference.
1111 */
1112 new_end = end - round_page(size);
1113 vm_reserv_array = (void *)(uintptr_t)pmap_map(vaddr, new_end, end,
1114 VM_PROT_READ | VM_PROT_WRITE);
1115 bzero(vm_reserv_array, size);
1116
1117 /*
1118 * Return the next available physical address.
1119 */
1120 return (new_end);
1121 }
1122
1123 #endif /* VM_NRESERVLEVEL > 0 */
Cache object: 0cd76dd534e2fc2e242bd77addb3f6c4
|