vm_reserv.c 41.2 KB
Newer Older
1
/*-
2
3
 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
 *
4
 * Copyright (c) 2002-2006 Rice University
Alan Cox's avatar
Alan Cox committed
5
 * Copyright (c) 2007-2011 Alan L. Cox <alc@cs.rice.edu>
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
 * All rights reserved.
 *
 * This software was developed for the FreeBSD Project by Alan L. Cox,
 * Olivier Crameri, Peter Druschel, Sitaram Iyer, and Juan Navarro.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 * 1. Redistributions of source code must retain the above copyright
 *    notice, this list of conditions and the following disclaimer.
 * 2. Redistributions in binary form must reproduce the above copyright
 *    notice, this list of conditions and the following disclaimer in the
 *    documentation and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
 * A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT
 * HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
 * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY
 * WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
 * POSSIBILITY OF SUCH DAMAGE.
 */

/*
 *	Superpage reservation management module
36
37
38
 *
 * Any external functions defined by this module are only to be used by the
 * virtual memory system.
39
40
41
42
43
44
45
46
47
48
49
50
51
 */

#include <sys/cdefs.h>
__FBSDID("$FreeBSD$");

#include "opt_vm.h"

#include <sys/param.h>
#include <sys/kernel.h>
#include <sys/lock.h>
#include <sys/malloc.h>
#include <sys/mutex.h>
#include <sys/queue.h>
52
#include <sys/rwlock.h>
53
54
55
#include <sys/sbuf.h>
#include <sys/sysctl.h>
#include <sys/systm.h>
56
57
#include <sys/counter.h>
#include <sys/ktr.h>
58
#include <sys/vmmeter.h>
59
#include <sys/smp.h>
60
61
62
63
64

#include <vm/vm.h>
#include <vm/vm_param.h>
#include <vm/vm_object.h>
#include <vm/vm_page.h>
65
#include <vm/vm_pageout.h>
66
#include <vm/vm_phys.h>
67
#include <vm/vm_pagequeue.h>
68
#include <vm/vm_radix.h>
69
70
71
72
#include <vm/vm_reserv.h>

/*
 * The reservation system supports the speculative allocation of large physical
73
 * pages ("superpages").  Speculative allocation enables the fully automatic
74
75
76
77
78
79
 * utilization of superpages by the virtual memory system.  In other words, no
 * programmatic directives are required to use superpages.
 */

#if VM_NRESERVLEVEL > 0

80
81
82
83
#ifndef VM_LEVEL_0_ORDER_MAX
#define	VM_LEVEL_0_ORDER_MAX	VM_LEVEL_0_ORDER
#endif

84
85
86
87
/*
 * The number of small pages that are contained in a level 0 reservation
 */
#define	VM_LEVEL_0_NPAGES	(1 << VM_LEVEL_0_ORDER)
88
#define	VM_LEVEL_0_NPAGES_MAX	(1 << VM_LEVEL_0_ORDER_MAX)
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107

/*
 * The number of bits by which a physical address is shifted to obtain the
 * reservation number
 */
#define	VM_LEVEL_0_SHIFT	(VM_LEVEL_0_ORDER + PAGE_SHIFT)

/*
 * The size of a level 0 reservation in bytes
 */
#define	VM_LEVEL_0_SIZE		(1 << VM_LEVEL_0_SHIFT)

/*
 * Computes the index of the small page underlying the given (object, pindex)
 * within the reservation's array of small pages.
 */
#define	VM_RESERV_INDEX(object, pindex)	\
    (((object)->pg_color + (pindex)) & (VM_LEVEL_0_NPAGES - 1))

Alan Cox's avatar
Alan Cox committed
108
109
110
111
112
113
114
115
116
117
118
119
120
121
/*
 * The size of a population map entry
 */
typedef	u_long		popmap_t;

/*
 * The number of bits in a population map entry
 */
#define	NBPOPMAP	(NBBY * sizeof(popmap_t))

/*
 * The number of population map entries in a reservation
 */
#define	NPOPMAP		howmany(VM_LEVEL_0_NPAGES, NBPOPMAP)
122
#define	NPOPMAP_MAX	howmany(VM_LEVEL_0_NPAGES_MAX, NBPOPMAP)
Alan Cox's avatar
Alan Cox committed
123

124
125
126
127
128
129
/*
 * Number of elapsed ticks before we update the LRU queue position.  Used
 * to reduce contention and churn on the list.
 */
#define	PARTPOPSLOP	1

130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
/*
 * Clear a bit in the population map.
 */
static __inline void
popmap_clear(popmap_t popmap[], int i)
{

	popmap[i / NBPOPMAP] &= ~(1UL << (i % NBPOPMAP));
}

/*
 * Set a bit in the population map.
 */
static __inline void
popmap_set(popmap_t popmap[], int i)
{

	popmap[i / NBPOPMAP] |= 1UL << (i % NBPOPMAP);
}

/*
 * Is a bit in the population map clear?
 */
static __inline boolean_t
popmap_is_clear(popmap_t popmap[], int i)
{

	return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) == 0);
}

/*
 * Is a bit in the population map set?
 */
static __inline boolean_t
popmap_is_set(popmap_t popmap[], int i)
{

	return ((popmap[i / NBPOPMAP] & (1UL << (i % NBPOPMAP))) != 0);
}

170
171
172
173
174
175
176
177
/*
 * The reservation structure
 *
 * A reservation structure is constructed whenever a large physical page is
 * speculatively allocated to an object.  The reservation provides the small
 * physical pages for the range [pindex, pindex + VM_LEVEL_0_NPAGES) of offsets
 * within that object.  The reservation's "popcnt" tracks the number of these
 * small physical pages that are in use at any given time.  When and if the
178
 * reservation is not fully utilized, it appears in the queue of partially
179
180
181
 * populated reservations.  The reservation always appears on the containing
 * object's list of reservations.
 *
182
 * A partially populated reservation can be broken and reclaimed at any time.
183
 *
184
185
 * r - vm_reserv_lock
 * d - vm_reserv_domain_lock
186
187
 * o - vm_reserv_object_lock
 * c - constant after boot
188
189
 */
struct vm_reserv {
190
191
192
193
194
	struct mtx	lock;			/* reservation lock. */
	TAILQ_ENTRY(vm_reserv) partpopq;	/* (d) per-domain queue. */
	LIST_ENTRY(vm_reserv) objq;		/* (o, r) object queue */
	vm_object_t	object;			/* (o, r) containing object */
	vm_pindex_t	pindex;			/* (o, r) offset in object */
195
	vm_page_t	pages;			/* (c) first page  */
196
197
	uint16_t	domain;			/* (c) NUMA domain. */
	uint16_t	popcnt;			/* (r) # of pages in use */
198
	int		lasttick;		/* (r) last pop update tick. */
199
	char		inpartpopq;		/* (d) */
200
	popmap_t	popmap[NPOPMAP_MAX];	/* (r) bit vector, used pages */
201
202
};

203
204
205
206
207
208
209
210
211
212
213
214
215
#define	vm_reserv_lockptr(rv)		(&(rv)->lock)
#define	vm_reserv_assert_locked(rv)					\
	    mtx_assert(vm_reserv_lockptr(rv), MA_OWNED)
#define	vm_reserv_lock(rv)		mtx_lock(vm_reserv_lockptr(rv))
#define	vm_reserv_trylock(rv)		mtx_trylock(vm_reserv_lockptr(rv))
#define	vm_reserv_unlock(rv)		mtx_unlock(vm_reserv_lockptr(rv))

static struct mtx_padalign vm_reserv_domain_locks[MAXMEMDOM];

#define	vm_reserv_domain_lockptr(d)	&vm_reserv_domain_locks[(d)]
#define	vm_reserv_domain_lock(d)	mtx_lock(vm_reserv_domain_lockptr(d))
#define	vm_reserv_domain_unlock(d)	mtx_unlock(vm_reserv_domain_lockptr(d))

216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
/*
 * The reservation array
 *
 * This array is analoguous in function to vm_page_array.  It differs in the
 * respect that it may contain a greater number of useful reservation
 * structures than there are (physical) superpages.  These "invalid"
 * reservation structures exist to trade-off space for time in the
 * implementation of vm_reserv_from_page().  Invalid reservation structures are
 * distinguishable from "valid" reservation structures by inspecting the
 * reservation's "pages" field.  Invalid reservation structures have a NULL
 * "pages" field.
 *
 * vm_reserv_from_page() maps a small (physical) page to an element of this
 * array by computing a physical reservation number from the page's physical
 * address.  The physical reservation number is used as the array index.
 *
 * An "active" reservation is a valid reservation structure that has a non-NULL
 * "object" field and a non-zero "popcnt" field.  In other words, every active
 * reservation belongs to a particular object.  Moreover, every active
 * reservation has an entry in the containing object's list of reservations.  
 */
static vm_reserv_t vm_reserv_array;

/*
240
 * The partially populated reservation queue
241
 *
242
243
244
 * This queue enables the fast recovery of an unused free small page from a
 * partially populated reservation.  The reservation at the head of this queue
 * is the least recently changed, partially populated reservation.
245
246
247
 *
 * Access to this queue is synchronized by the free page queue lock.
 */
248
static TAILQ_HEAD(, vm_reserv) vm_rvq_partpop[MAXMEMDOM];
249
250
251

static SYSCTL_NODE(_vm, OID_AUTO, reserv, CTLFLAG_RD, 0, "Reservation Info");

252
253
254
static counter_u64_t vm_reserv_broken = EARLY_COUNTER;
SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, broken, CTLFLAG_RD,
    &vm_reserv_broken, "Cumulative number of broken reservations");
255

256
257
258
static counter_u64_t vm_reserv_freed = EARLY_COUNTER;
SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, freed, CTLFLAG_RD,
    &vm_reserv_freed, "Cumulative number of freed reservations");
259

260
261
262
263
264
static int sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS);

SYSCTL_PROC(_vm_reserv, OID_AUTO, fullpop, CTLTYPE_INT | CTLFLAG_RD, NULL, 0,
    sysctl_vm_reserv_fullpop, "I", "Current number of full reservations");

265
266
267
static int sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS);

SYSCTL_OID(_vm_reserv, OID_AUTO, partpopq, CTLTYPE_STRING | CTLFLAG_RD, NULL, 0,
268
    sysctl_vm_reserv_partpopq, "A", "Partially populated reservation queues");
269

270
271
272
static counter_u64_t vm_reserv_reclaimed = EARLY_COUNTER;
SYSCTL_COUNTER_U64(_vm_reserv, OID_AUTO, reclaimed, CTLFLAG_RD,
    &vm_reserv_reclaimed, "Cumulative number of reclaimed reservations");
273

274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
/*
 * The object lock pool is used to synchronize the rvq.  We can not use a
 * pool mutex because it is required before malloc works.
 *
 * The "hash" function could be made faster without divide and modulo.
 */
#define	VM_RESERV_OBJ_LOCK_COUNT	MAXCPU

struct mtx_padalign vm_reserv_object_mtx[VM_RESERV_OBJ_LOCK_COUNT];

#define	vm_reserv_object_lock_idx(object)			\
	    (((uintptr_t)object / sizeof(*object)) % VM_RESERV_OBJ_LOCK_COUNT)
#define	vm_reserv_object_lock_ptr(object)			\
	    &vm_reserv_object_mtx[vm_reserv_object_lock_idx((object))]
#define	vm_reserv_object_lock(object)				\
	    mtx_lock(vm_reserv_object_lock_ptr((object)))
#define	vm_reserv_object_unlock(object)				\
	    mtx_unlock(vm_reserv_object_lock_ptr((object)))

293
static void		vm_reserv_break(vm_reserv_t rv);
Alan Cox's avatar
Alan Cox committed
294
static void		vm_reserv_depopulate(vm_reserv_t rv, int index);
295
296
297
static vm_reserv_t	vm_reserv_from_page(vm_page_t m);
static boolean_t	vm_reserv_has_pindex(vm_reserv_t rv,
			    vm_pindex_t pindex);
Alan Cox's avatar
Alan Cox committed
298
static void		vm_reserv_populate(vm_reserv_t rv, int index);
299
static void		vm_reserv_reclaim(vm_reserv_t rv);
300

301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
/*
 * Returns the current number of full reservations.
 *
 * Since the number of full reservations is computed without acquiring the
 * free page queue lock, the returned value may be inexact.
 */
static int
sysctl_vm_reserv_fullpop(SYSCTL_HANDLER_ARGS)
{
	vm_paddr_t paddr;
	struct vm_phys_seg *seg;
	vm_reserv_t rv;
	int fullpop, segind;

	fullpop = 0;
	for (segind = 0; segind < vm_phys_nsegs; segind++) {
		seg = &vm_phys_segs[segind];
		paddr = roundup2(seg->start, VM_LEVEL_0_SIZE);
Aleksandr Rybalko's avatar
Aleksandr Rybalko committed
319
320
		while (paddr + VM_LEVEL_0_SIZE > paddr && paddr +
		    VM_LEVEL_0_SIZE <= seg->end) {
321
322
323
324
325
326
327
328
			rv = &vm_reserv_array[paddr >> VM_LEVEL_0_SHIFT];
			fullpop += rv->popcnt == VM_LEVEL_0_NPAGES;
			paddr += VM_LEVEL_0_SIZE;
		}
	}
	return (sysctl_handle_int(oidp, &fullpop, 0, req));
}

329
/*
330
 * Describes the current state of the partially populated reservation queue.
331
332
333
334
335
336
 */
static int
sysctl_vm_reserv_partpopq(SYSCTL_HANDLER_ARGS)
{
	struct sbuf sbuf;
	vm_reserv_t rv;
337
	int counter, error, domain, level, unused_pages;
338

339
340
341
	error = sysctl_wire_old_buffer(req, 0);
	if (error != 0)
		return (error);
342
	sbuf_new_for_sysctl(&sbuf, NULL, 128, req);
343
344
345
346
347
	sbuf_printf(&sbuf, "\nDOMAIN    LEVEL     SIZE  NUMBER\n\n");
	for (domain = 0; domain < vm_ndomains; domain++) {
		for (level = -1; level <= VM_NRESERVLEVEL - 2; level++) {
			counter = 0;
			unused_pages = 0;
348
			vm_reserv_domain_lock(domain);
349
350
351
352
			TAILQ_FOREACH(rv, &vm_rvq_partpop[domain], partpopq) {
				counter++;
				unused_pages += VM_LEVEL_0_NPAGES - rv->popcnt;
			}
353
			vm_reserv_domain_unlock(domain);
354
355
356
			sbuf_printf(&sbuf, "%6d, %7d, %6dK, %6d\n",
			    domain, level,
			    unused_pages * ((int)PAGE_SIZE / 1024), counter);
357
358
		}
	}
359
	error = sbuf_finish(&sbuf);
360
361
362
363
	sbuf_delete(&sbuf);
	return (error);
}

364
365
366
367
368
369
370
371
/*
 * Remove a reservation from the object's objq.
 */
static void
vm_reserv_remove(vm_reserv_t rv)
{
	vm_object_t object;

372
373
374
	vm_reserv_assert_locked(rv);
	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
	KASSERT(rv->object != NULL,
	    ("vm_reserv_remove: reserv %p is free", rv));
	KASSERT(!rv->inpartpopq,
	    ("vm_reserv_remove: reserv %p's inpartpopq is TRUE", rv));
	object = rv->object;
	vm_reserv_object_lock(object);
	LIST_REMOVE(rv, objq);
	rv->object = NULL;
	vm_reserv_object_unlock(object);
}

/*
 * Insert a new reservation into the object's objq.
 */
static void
vm_reserv_insert(vm_reserv_t rv, vm_object_t object, vm_pindex_t pindex)
{
	int i;

394
395
396
397
398
	vm_reserv_assert_locked(rv);
	CTR6(KTR_VM,
	    "%s: rv %p(%p) object %p new %p popcnt %d",
	    __FUNCTION__, rv, rv->pages, rv->object, object,
	   rv->popcnt);
399
400
401
402
403
404
405
406
407
408
409
410
	KASSERT(rv->object == NULL,
	    ("vm_reserv_insert: reserv %p isn't free", rv));
	KASSERT(rv->popcnt == 0,
	    ("vm_reserv_insert: reserv %p's popcnt is corrupted", rv));
	KASSERT(!rv->inpartpopq,
	    ("vm_reserv_insert: reserv %p's inpartpopq is TRUE", rv));
	for (i = 0; i < NPOPMAP; i++)
		KASSERT(rv->popmap[i] == 0,
		    ("vm_reserv_insert: reserv %p's popmap is corrupted", rv));
	vm_reserv_object_lock(object);
	rv->pindex = pindex;
	rv->object = object;
411
	rv->lasttick = ticks;
412
413
414
415
	LIST_INSERT_HEAD(&object->rvq, rv, objq);
	vm_reserv_object_unlock(object);
}

416
417
418
/*
 * Reduces the given reservation's population count.  If the population count
 * becomes zero, the reservation is destroyed.  Additionally, moves the
419
 * reservation to the tail of the partially populated reservation queue if the
420
421
422
 * population count is non-zero.
 */
static void
Alan Cox's avatar
Alan Cox committed
423
vm_reserv_depopulate(vm_reserv_t rv, int index)
424
{
425
	struct vm_domain *vmd;
426

427
428
429
	vm_reserv_assert_locked(rv);
	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
430
431
	KASSERT(rv->object != NULL,
	    ("vm_reserv_depopulate: reserv %p is free", rv));
432
	KASSERT(popmap_is_set(rv->popmap, index),
433
434
	    ("vm_reserv_depopulate: reserv %p's popmap[%d] is clear", rv,
	    index));
435
436
	KASSERT(rv->popcnt > 0,
	    ("vm_reserv_depopulate: reserv %p's popcnt is corrupted", rv));
437
	KASSERT(rv->domain < vm_ndomains,
438
439
	    ("vm_reserv_depopulate: reserv %p's domain is corrupted %d",
	    rv, rv->domain));
440
	if (rv->popcnt == VM_LEVEL_0_NPAGES) {
441
442
443
444
		KASSERT(rv->pages->psind == 1,
		    ("vm_reserv_depopulate: reserv %p is already demoted",
		    rv));
		rv->pages->psind = 0;
445
	}
446
	popmap_clear(rv->popmap, index);
447
	rv->popcnt--;
448
449
450
451
452
453
454
455
456
457
458
459
460
	if ((unsigned)(ticks - rv->lasttick) >= PARTPOPSLOP ||
	    rv->popcnt == 0) {
		vm_reserv_domain_lock(rv->domain);
		if (rv->inpartpopq) {
			TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
			rv->inpartpopq = FALSE;
		}
		if (rv->popcnt != 0) {
			rv->inpartpopq = TRUE;
			TAILQ_INSERT_TAIL(&vm_rvq_partpop[rv->domain], rv, partpopq);
		}
		vm_reserv_domain_unlock(rv->domain);
		rv->lasttick = ticks;
461
462
	}
	vmd = VM_DOMAIN(rv->domain);
463
	if (rv->popcnt == 0) {
464
		vm_reserv_remove(rv);
465
		vm_domain_free_lock(vmd);
466
		vm_phys_free_pages(rv->pages, VM_LEVEL_0_ORDER);
467
468
		vm_domain_free_unlock(vmd);
		counter_u64_add(vm_reserv_freed, 1);
469
	}
470
	vm_domain_freecnt_inc(vmd, 1);
471
472
473
474
475
476
477
478
479
480
481
482
}

/*
 * Returns the reservation to which the given page might belong.
 */
static __inline vm_reserv_t
vm_reserv_from_page(vm_page_t m)
{

	return (&vm_reserv_array[VM_PAGE_TO_PHYS(m) >> VM_LEVEL_0_SHIFT]);
}

483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
/*
 * Returns an existing reservation or NULL and initialized successor pointer.
 */
static vm_reserv_t
vm_reserv_from_object(vm_object_t object, vm_pindex_t pindex,
    vm_page_t mpred, vm_page_t *msuccp)
{
	vm_reserv_t rv;
	vm_page_t msucc;

	msucc = NULL;
	if (mpred != NULL) {
		KASSERT(mpred->object == object,
		    ("vm_reserv_from_object: object doesn't contain mpred"));
		KASSERT(mpred->pindex < pindex,
		    ("vm_reserv_from_object: mpred doesn't precede pindex"));
		rv = vm_reserv_from_page(mpred);
		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
			goto found;
		msucc = TAILQ_NEXT(mpred, listq);
	} else
		msucc = TAILQ_FIRST(&object->memq);
	if (msucc != NULL) {
		KASSERT(msucc->pindex > pindex,
		    ("vm_reserv_from_object: msucc doesn't succeed pindex"));
		rv = vm_reserv_from_page(msucc);
		if (rv->object == object && vm_reserv_has_pindex(rv, pindex))
			goto found;
	}
	rv = NULL;

found:
	*msuccp = msucc;

	return (rv);
}

520
521
522
523
524
525
526
527
528
529
530
531
532
/*
 * Returns TRUE if the given reservation contains the given page index and
 * FALSE otherwise.
 */
static __inline boolean_t
vm_reserv_has_pindex(vm_reserv_t rv, vm_pindex_t pindex)
{

	return (((pindex - rv->pindex) & ~(VM_LEVEL_0_NPAGES - 1)) == 0);
}

/*
 * Increases the given reservation's population count.  Moves the reservation
533
 * to the tail of the partially populated reservation queue.
534
535
536
537
 *
 * The free page queue must be locked.
 */
static void
Alan Cox's avatar
Alan Cox committed
538
vm_reserv_populate(vm_reserv_t rv, int index)
539
540
{

541
542
543
	vm_reserv_assert_locked(rv);
	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
544
545
	KASSERT(rv->object != NULL,
	    ("vm_reserv_populate: reserv %p is free", rv));
546
	KASSERT(popmap_is_clear(rv->popmap, index),
547
548
	    ("vm_reserv_populate: reserv %p's popmap[%d] is set", rv,
	    index));
549
550
	KASSERT(rv->popcnt < VM_LEVEL_0_NPAGES,
	    ("vm_reserv_populate: reserv %p is already full", rv));
551
552
	KASSERT(rv->pages->psind == 0,
	    ("vm_reserv_populate: reserv %p is already promoted", rv));
553
	KASSERT(rv->domain < vm_ndomains,
554
555
	    ("vm_reserv_populate: reserv %p's domain is corrupted %d",
	    rv, rv->domain));
556
557
	popmap_set(rv->popmap, index);
	rv->popcnt++;
558
559
560
561
	if ((unsigned)(ticks - rv->lasttick) < PARTPOPSLOP &&
	    rv->inpartpopq && rv->popcnt != VM_LEVEL_0_NPAGES)
		return;
	rv->lasttick = ticks;
562
	vm_reserv_domain_lock(rv->domain);
563
	if (rv->inpartpopq) {
564
		TAILQ_REMOVE(&vm_rvq_partpop[rv->domain], rv, partpopq);
565
566
567
568
		rv->inpartpopq = FALSE;
	}
	if (rv->popcnt < VM_LEVEL_0_NPAGES) {
		rv->inpartpopq = TRUE;
569
		TAILQ_INSERT_TAIL(&vm_rvq_partpop[rv->domain], rv, partpopq);
570
571
572
573
	} else {
		KASSERT(rv->pages->psind == 0,
		    ("vm_reserv_populate: reserv %p is already promoted",
		    rv));
574
		rv->pages->psind = 1;
575
576
	}
	vm_reserv_domain_unlock(rv->domain);
577
578
579
}

/*
580
 * Allocates a contiguous set of physical pages of the given size "npages"
581
 * from existing or newly created reservations.  All of the physical pages
582
583
584
585
586
587
 * must be at or above the given physical address "low" and below the given
 * physical address "high".  The given value "alignment" determines the
 * alignment of the first physical page in the set.  If the given value
 * "boundary" is non-zero, then the set of physical pages cannot cross any
 * physical address boundary that is a multiple of that value.  Both
 * "alignment" and "boundary" must be a power of two.
588
 *
589
590
591
 * The page "mpred" must immediately precede the offset "pindex" within the
 * specified object.
 *
592
 * The object must be locked.
593
594
 */
vm_page_t
595
596
597
vm_reserv_alloc_contig(vm_object_t object, vm_pindex_t pindex, int domain,
    int req, vm_page_t mpred, u_long npages, vm_paddr_t low, vm_paddr_t high,
    u_long alignment, vm_paddr_t boundary)
598
{
599
	struct vm_domain *vmd;
600
	vm_paddr_t pa, size;
601
	vm_page_t m, m_ret, msucc;
602
603
	vm_pindex_t first, leftcap, rightcap;
	vm_reserv_t rv;
604
605
	u_long allocpages, maxpages, minpages;
	int i, index, n;
606

607
	VM_OBJECT_ASSERT_WLOCKED(object);
608
	KASSERT(npages != 0, ("vm_reserv_alloc_contig: npages is 0"));
609
610

	/*
611
	 * Is a reservation fundamentally impossible?
612
613
	 */
	if (pindex < VM_RESERV_INDEX(object, pindex) ||
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
	    pindex + npages > object->size)
		return (NULL);

	/*
	 * All reservations of a particular size have the same alignment.
	 * Assuming that the first page is allocated from a reservation, the
	 * least significant bits of its physical address can be determined
	 * from its offset from the beginning of the reservation and the size
	 * of the reservation.
	 *
	 * Could the specified index within a reservation of the smallest
	 * possible size satisfy the alignment and boundary requirements?
	 */
	pa = VM_RESERV_INDEX(object, pindex) << PAGE_SHIFT;
	if ((pa & (alignment - 1)) != 0)
		return (NULL);
	size = npages << PAGE_SHIFT;
	if (((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
632
633
634
		return (NULL);

	/*
635
	 * Look for an existing reservation.
636
	 */
637
	rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
	if (rv != NULL) {
		KASSERT(object != kernel_object || rv->domain == domain,
		    ("vm_reserv_alloc_contig: domain mismatch"));
		index = VM_RESERV_INDEX(object, pindex);
		/* Does the allocation fit within the reservation? */
		if (index + npages > VM_LEVEL_0_NPAGES)
			return (NULL);
		domain = rv->domain;
		vmd = VM_DOMAIN(domain);
		vm_reserv_lock(rv);
		/* Handle reclaim race. */
		if (rv->object != object)
			goto out;
		m = &rv->pages[index];
		pa = VM_PAGE_TO_PHYS(m);
		if (pa < low || pa + size > high ||
		    (pa & (alignment - 1)) != 0 ||
		    ((pa ^ (pa + size - 1)) & ~(boundary - 1)) != 0)
			goto out;
		/* Handle vm_page_rename(m, new_object, ...). */
		for (i = 0; i < npages; i++)
			if (popmap_is_set(rv->popmap, index + i))
				goto out;
		if (!vm_domain_allocate(vmd, req, npages))
			goto out;
		for (i = 0; i < npages; i++)
			vm_reserv_populate(rv, index + i);
		vm_reserv_unlock(rv);
		return (m);
out:
		vm_reserv_unlock(rv);
669
		return (NULL);
670
	}
671
672

	/*
673
	 * Could at least one reservation fit between the first index to the
674
675
	 * left that can be used ("leftcap") and the first index to the right
	 * that cannot be used ("rightcap")?
676
677
678
679
	 *
	 * We must synchronize with the reserv object lock to protect the
	 * pindex/object of the resulting reservations against rename while
	 * we are inspecting.
680
681
	 */
	first = pindex - VM_RESERV_INDEX(object, pindex);
682
683
684
685
	minpages = VM_RESERV_INDEX(object, pindex) + npages;
	maxpages = roundup2(minpages, VM_LEVEL_0_NPAGES);
	allocpages = maxpages;
	vm_reserv_object_lock(object);
686
687
688
689
690
	if (mpred != NULL) {
		if ((rv = vm_reserv_from_page(mpred))->object != object)
			leftcap = mpred->pindex + 1;
		else
			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
691
692
		if (leftcap > first) {
			vm_reserv_object_unlock(object);
693
			return (NULL);
694
		}
695
696
697
698
699
700
701
	}
	if (msucc != NULL) {
		if ((rv = vm_reserv_from_page(msucc))->object != object)
			rightcap = msucc->pindex;
		else
			rightcap = rv->pindex;
		if (first + maxpages > rightcap) {
702
703
			if (maxpages == VM_LEVEL_0_NPAGES) {
				vm_reserv_object_unlock(object);
704
				return (NULL);
705
			}
706
707
708
709
710
711
712

			/*
			 * At least one reservation will fit between "leftcap"
			 * and "rightcap".  However, a reservation for the
			 * last of the requested pages will not fit.  Reduce
			 * the size of the upcoming allocation accordingly.
			 */
713
714
715
			allocpages = minpages;
		}
	}
716
	vm_reserv_object_unlock(object);
717
718

	/*
719
	 * Would the last new reservation extend past the end of the object?
720
	 */
721
	if (first + maxpages > object->size) {
722
		/*
723
724
		 * Don't allocate the last new reservation if the object is a
		 * vnode or backed by another object that is a vnode. 
725
726
727
		 */
		if (object->type == OBJT_VNODE ||
		    (object->backing_object != NULL &&
728
729
730
731
732
		    object->backing_object->type == OBJT_VNODE)) {
			if (maxpages == VM_LEVEL_0_NPAGES)
				return (NULL);
			allocpages = minpages;
		}
733
734
735
736
		/* Speculate that the object may grow. */
	}

	/*
737
738
739
740
741
	 * Allocate the physical pages.  The alignment and boundary specified
	 * for this allocation may be different from the alignment and
	 * boundary specified for the requested pages.  For instance, the
	 * specified index may not be the first page within the first new
	 * reservation.
742
	 */
743
744
745
746
747
748
749
750
751
752
753
754
755
	m = NULL;
	vmd = VM_DOMAIN(domain);
	if (vm_domain_allocate(vmd, req, npages)) {
		vm_domain_free_lock(vmd);
		m = vm_phys_alloc_contig(domain, allocpages, low, high,
		    ulmax(alignment, VM_LEVEL_0_SIZE),
		    boundary > VM_LEVEL_0_SIZE ? boundary : 0);
		vm_domain_free_unlock(vmd);
		if (m == NULL) {
			vm_domain_freecnt_inc(vmd, npages);
			return (NULL);
		}
	} else
756
		return (NULL);
757
	KASSERT(vm_phys_domain(m) == domain,
758
	    ("vm_reserv_alloc_contig: Page domain does not match requested."));
759
760
761
762
763
764
765

	/*
	 * The allocated physical pages always begin at a reservation
	 * boundary, but they do not always end at a reservation boundary.
	 * Initialize every reservation that is completely covered by the
	 * allocated physical pages.
	 */
766
767
768
	m_ret = NULL;
	index = VM_RESERV_INDEX(object, pindex);
	do {
769
770
		rv = vm_reserv_from_page(m);
		KASSERT(rv->pages == m,
771
		    ("vm_reserv_alloc_contig: reserv %p's pages is corrupted",
772
		    rv));
773
		vm_reserv_lock(rv);
774
		vm_reserv_insert(rv, object, first);
775
776
		n = ulmin(VM_LEVEL_0_NPAGES - index, npages);
		for (i = 0; i < n; i++)
Alan Cox's avatar
Alan Cox committed
777
			vm_reserv_populate(rv, index + i);
778
779
780
781
782
		npages -= n;
		if (m_ret == NULL) {
			m_ret = &rv->pages[index];
			index = 0;
		}
783
		vm_reserv_unlock(rv);
784
785
786
		m += VM_LEVEL_0_NPAGES;
		first += VM_LEVEL_0_NPAGES;
		allocpages -= VM_LEVEL_0_NPAGES;
787
	} while (allocpages >= VM_LEVEL_0_NPAGES);
788
	return (m_ret);
789
790
791
}

/*
792
 * Allocate a physical page from an existing or newly created reservation.
793
794
795
796
797
798
799
 *
 * The page "mpred" must immediately precede the offset "pindex" within the
 * specified object.
 *
 * The object must be locked.
 */
vm_page_t
800
801
vm_reserv_alloc_page(vm_object_t object, vm_pindex_t pindex, int domain,
    int req, vm_page_t mpred)
802
{
803
	struct vm_domain *vmd;
804
	vm_page_t m, msucc;
805
806
	vm_pindex_t first, leftcap, rightcap;
	vm_reserv_t rv;
807
	int index;
808

809
	VM_OBJECT_ASSERT_WLOCKED(object);
810
811
812
813
814
815
816
817
818

	/*
	 * Is a reservation fundamentally impossible?
	 */
	if (pindex < VM_RESERV_INDEX(object, pindex) ||
	    pindex >= object->size)
		return (NULL);

	/*
819
	 * Look for an existing reservation.
820
	 */
821
	rv = vm_reserv_from_object(object, pindex, mpred, &msucc);
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
	if (rv != NULL) {
		KASSERT(object != kernel_object || rv->domain == domain,
		    ("vm_reserv_alloc_page: domain mismatch"));
		domain = rv->domain;
		vmd = VM_DOMAIN(domain);
		index = VM_RESERV_INDEX(object, pindex);
		m = &rv->pages[index];
		vm_reserv_lock(rv);
		/* Handle reclaim race. */
		if (rv->object != object ||
		    /* Handle vm_page_rename(m, new_object, ...). */
		    popmap_is_set(rv->popmap, index)) {
			m = NULL;
			goto out;
		}
		if (vm_domain_allocate(vmd, req, 1) == 0)
			m = NULL;
		else
			vm_reserv_populate(rv, index);
out:
		vm_reserv_unlock(rv);
		return (m);
	}
845
846
847
848

	/*
	 * Could a reservation fit between the first index to the left that
	 * can be used and the first index to the right that cannot be used?
849
850
851
852
	 *
	 * We must synchronize with the reserv object lock to protect the
	 * pindex/object of the resulting reservations against rename while
	 * we are inspecting.
853
854
	 */
	first = pindex - VM_RESERV_INDEX(object, pindex);
855
	vm_reserv_object_lock(object);
856
857
858
859
860
	if (mpred != NULL) {
		if ((rv = vm_reserv_from_page(mpred))->object != object)
			leftcap = mpred->pindex + 1;
		else
			leftcap = rv->pindex + VM_LEVEL_0_NPAGES;
861
862
		if (leftcap > first) {
			vm_reserv_object_unlock(object);
863
			return (NULL);
864
		}
865
866
867
868
869
870
	}
	if (msucc != NULL) {
		if ((rv = vm_reserv_from_page(msucc))->object != object)
			rightcap = msucc->pindex;
		else
			rightcap = rv->pindex;
871
872
		if (first + VM_LEVEL_0_NPAGES > rightcap) {
			vm_reserv_object_unlock(object);
873
			return (NULL);
874
		}
875
	}
876
	vm_reserv_object_unlock(object);
877
878
879
880
881
882
883
884
885
886
887
888
889
890

	/*
	 * Would a new reservation extend past the end of the object? 
	 */
	if (first + VM_LEVEL_0_NPAGES > object->size) {
		/*
		 * Don't allocate a new reservation if the object is a vnode or
		 * backed by another object that is a vnode. 
		 */
		if (object->type == OBJT_VNODE ||
		    (object->backing_object != NULL &&
		    object->backing_object->type == OBJT_VNODE))
			return (NULL);
		/* Speculate that the object may grow. */
891
	}
892
893
894
895

	/*
	 * Allocate and populate the new reservation.
	 */
896
897
898
899
900
901
902
903
904
905
906
907
	m = NULL;
	vmd = VM_DOMAIN(domain);
	if (vm_domain_allocate(vmd, req, 1)) {
		vm_domain_free_lock(vmd);
		m = vm_phys_alloc_pages(domain, VM_FREEPOOL_DEFAULT,
		    VM_LEVEL_0_ORDER);
		vm_domain_free_unlock(vmd);
		if (m == NULL) {
			vm_domain_freecnt_inc(vmd, 1);
			return (NULL);
		}
	} else
908
909
		return (NULL);
	rv = vm_reserv_from_page(m);
910
	vm_reserv_lock(rv);
911
912
	KASSERT(rv->pages == m,
	    ("vm_reserv_alloc_page: reserv %p's pages is corrupted", rv));
913
	vm_reserv_insert(rv, object, first);
Alan Cox's avatar
Alan Cox committed
914
915
	index = VM_RESERV_INDEX(object, pindex);
	vm_reserv_populate(rv, index);
916
917
	vm_reserv_unlock(rv);

Alan Cox's avatar
Alan Cox committed
918
	return (&rv->pages[index]);
919
920
}

Alan Cox's avatar
Alan Cox committed
921
/*
922
923
924
 * Breaks the given reservation.  All free pages in the reservation
 * are returned to the physical memory allocator.  The reservation's
 * population count and map are reset to their initial state.
Alan Cox's avatar
Alan Cox committed
925
 *
926
 * The given reservation must not be in the partially populated reservation
Alan Cox's avatar
Alan Cox committed
927
928
929
 * queue.  The free page queue lock must be held.
 */
static void
930
vm_reserv_break(vm_reserv_t rv)
Alan Cox's avatar
Alan Cox committed
931
{
932
933
	u_long changes;
	int bitpos, hi, i, lo;
Alan Cox's avatar
Alan Cox committed
934

935
936
937
	vm_reserv_assert_locked(rv);
	CTR5(KTR_VM, "%s: rv %p object %p popcnt %d inpartpop %d",
	    __FUNCTION__, rv, rv->object, rv->popcnt, rv->inpartpopq);
938
	vm_reserv_remove(rv);
939
	rv->pages->psind = 0;
940
941
942
943
944
945
946
947
948
949
950
951
952
	hi = lo = -1;
	for (i = 0; i <= NPOPMAP; i++) {
		/*
		 * "changes" is a bitmask that marks where a new sequence of
		 * 0s or 1s begins in popmap[i], with last bit in popmap[i-1]
		 * considered to be 1 if and only if lo == hi.  The bits of
		 * popmap[-1] and popmap[NPOPMAP] are considered all 1s.
		 */
		if (i == NPOPMAP)
			changes = lo != hi;
		else {
			changes = rv->popmap[i];
			changes ^= (changes << 1) | (lo == hi);
Alan Cox's avatar
Alan Cox committed
953
954
			rv->popmap[i] = 0;
		}
955
956
957
958
959
960
961
962
963
964
965
966
967
		while (changes != 0) {
			/*
			 * If the next change marked begins a run of 0s, set
			 * lo to mark that position.  Otherwise set hi and
			 * free pages from lo up to hi.
			 */
			bitpos = ffsl(changes) - 1;
			changes ^= 1UL << bitpos;
			if (lo == hi)
				lo = NBPOPMAP * i + bitpos;
			else {
				hi = NBPOPMAP * i + bitpos;
				vm_domain_free_lock(VM_DOMAIN(rv->domain));
968
				vm_phys_enqueue_contig(&rv->pages[lo], hi - lo);
969
970
971
				vm_domain_free_unlock(VM_DOMAIN(rv->domain));
				lo = hi;
			}
Alan Cox's avatar
Alan Cox committed
972
		}
973
974
	}
	rv->popcnt = 0;
975
	counter_u64_add(vm_reserv_broken, 1);
Alan Cox's avatar
Alan Cox committed
976
977
}

978
979
980
981
982
983
984
985
/*
 * Breaks all reservations belonging to the given object.
 */
void
vm_reserv_break_all(vm_object_t object)
{
	vm_reserv_t rv;

986
987
988
989
990
991
992
	/*
	 * This access of object->rvq is unsynchronized so that the
	 * object rvq lock can nest after the domain_free lock.  We
	 * must check for races in the results.  However, the object
	 * lock prevents new additions, so we are guaranteed that when
	 * it returns NULL the object is properly empty.
	 */
993
	while ((rv = LIST_FIRST(&object->rvq)) != NULL) {
994
		vm_reserv_lock(rv);
995
		/* Reclaim race. */
996
997
		if (rv->object != object) {
			vm_reserv_unlock(rv);
998
			continue;
999
1000
		}
		vm_reserv_domain_lock(rv->domain);
For faster browsing, not all history is shown. View entire blame