FreeBSD/Linux Kernel Cross Reference
sys/kern/subr_vmem.c
1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c)2006,2007,2008,2009 YAMAMOTO Takashi,
5 * Copyright (c) 2013 EMC Corp.
6 * All rights reserved.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions and the following disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 * SUCH DAMAGE.
28 */
29
30 /*
31 * From:
32 * $NetBSD: vmem_impl.h,v 1.2 2013/01/29 21:26:24 para Exp $
33 * $NetBSD: subr_vmem.c,v 1.83 2013/03/06 11:20:10 yamt Exp $
34 */
35
36 /*
37 * reference:
38 * - Magazines and Vmem: Extending the Slab Allocator
39 * to Many CPUs and Arbitrary Resources
40 * http://www.usenix.org/event/usenix01/bonwick.html
41 */
42
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
45
46 #include "opt_ddb.h"
47
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/kernel.h>
51 #include <sys/queue.h>
52 #include <sys/callout.h>
53 #include <sys/hash.h>
54 #include <sys/lock.h>
55 #include <sys/malloc.h>
56 #include <sys/mutex.h>
57 #include <sys/smp.h>
58 #include <sys/condvar.h>
59 #include <sys/sysctl.h>
60 #include <sys/taskqueue.h>
61 #include <sys/vmem.h>
62 #include <sys/vmmeter.h>
63
64 #include "opt_vm.h"
65
66 #include <vm/uma.h>
67 #include <vm/vm.h>
68 #include <vm/pmap.h>
69 #include <vm/vm_map.h>
70 #include <vm/vm_object.h>
71 #include <vm/vm_kern.h>
72 #include <vm/vm_extern.h>
73 #include <vm/vm_param.h>
74 #include <vm/vm_page.h>
75 #include <vm/vm_pageout.h>
76 #include <vm/vm_phys.h>
77 #include <vm/vm_pagequeue.h>
78 #include <vm/uma_int.h>
79
80 int vmem_startup_count(void);
81
82 #define VMEM_OPTORDER 5
83 #define VMEM_OPTVALUE (1 << VMEM_OPTORDER)
84 #define VMEM_MAXORDER \
85 (VMEM_OPTVALUE - 1 + sizeof(vmem_size_t) * NBBY - VMEM_OPTORDER)
86
87 #define VMEM_HASHSIZE_MIN 16
88 #define VMEM_HASHSIZE_MAX 131072
89
90 #define VMEM_QCACHE_IDX_MAX 16
91
92 #define VMEM_FITMASK (M_BESTFIT | M_FIRSTFIT | M_NEXTFIT)
93
94 #define VMEM_FLAGS (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM | \
95 M_BESTFIT | M_FIRSTFIT | M_NEXTFIT)
96
97 #define BT_FLAGS (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM)
98
99 #define QC_NAME_MAX 16
100
101 /*
102 * Data structures private to vmem.
103 */
104 MALLOC_DEFINE(M_VMEM, "vmem", "vmem internal structures");
105
106 typedef struct vmem_btag bt_t;
107
108 TAILQ_HEAD(vmem_seglist, vmem_btag);
109 LIST_HEAD(vmem_freelist, vmem_btag);
110 LIST_HEAD(vmem_hashlist, vmem_btag);
111
112 struct qcache {
113 uma_zone_t qc_cache;
114 vmem_t *qc_vmem;
115 vmem_size_t qc_size;
116 char qc_name[QC_NAME_MAX];
117 };
118 typedef struct qcache qcache_t;
119 #define QC_POOL_TO_QCACHE(pool) ((qcache_t *)(pool->pr_qcache))
120
121 #define VMEM_NAME_MAX 16
122
123 /* boundary tag */
124 struct vmem_btag {
125 TAILQ_ENTRY(vmem_btag) bt_seglist;
126 union {
127 LIST_ENTRY(vmem_btag) u_freelist; /* BT_TYPE_FREE */
128 LIST_ENTRY(vmem_btag) u_hashlist; /* BT_TYPE_BUSY */
129 } bt_u;
130 #define bt_hashlist bt_u.u_hashlist
131 #define bt_freelist bt_u.u_freelist
132 vmem_addr_t bt_start;
133 vmem_size_t bt_size;
134 int bt_type;
135 };
136
137 /* vmem arena */
138 struct vmem {
139 struct mtx_padalign vm_lock;
140 struct cv vm_cv;
141 char vm_name[VMEM_NAME_MAX+1];
142 LIST_ENTRY(vmem) vm_alllist;
143 struct vmem_hashlist vm_hash0[VMEM_HASHSIZE_MIN];
144 struct vmem_freelist vm_freelist[VMEM_MAXORDER];
145 struct vmem_seglist vm_seglist;
146 struct vmem_hashlist *vm_hashlist;
147 vmem_size_t vm_hashsize;
148
149 /* Constant after init */
150 vmem_size_t vm_qcache_max;
151 vmem_size_t vm_quantum_mask;
152 vmem_size_t vm_import_quantum;
153 int vm_quantum_shift;
154
155 /* Written on alloc/free */
156 LIST_HEAD(, vmem_btag) vm_freetags;
157 int vm_nfreetags;
158 int vm_nbusytag;
159 vmem_size_t vm_inuse;
160 vmem_size_t vm_size;
161 vmem_size_t vm_limit;
162 struct vmem_btag vm_cursor;
163
164 /* Used on import. */
165 vmem_import_t *vm_importfn;
166 vmem_release_t *vm_releasefn;
167 void *vm_arg;
168
169 /* Space exhaustion callback. */
170 vmem_reclaim_t *vm_reclaimfn;
171
172 /* quantum cache */
173 qcache_t vm_qcache[VMEM_QCACHE_IDX_MAX];
174 };
175
176 #define BT_TYPE_SPAN 1 /* Allocated from importfn */
177 #define BT_TYPE_SPAN_STATIC 2 /* vmem_add() or create. */
178 #define BT_TYPE_FREE 3 /* Available space. */
179 #define BT_TYPE_BUSY 4 /* Used space. */
180 #define BT_TYPE_CURSOR 5 /* Cursor for nextfit allocations. */
181 #define BT_ISSPAN_P(bt) ((bt)->bt_type <= BT_TYPE_SPAN_STATIC)
182
183 #define BT_END(bt) ((bt)->bt_start + (bt)->bt_size - 1)
184
185 #if defined(DIAGNOSTIC)
186 static int enable_vmem_check = 1;
187 SYSCTL_INT(_debug, OID_AUTO, vmem_check, CTLFLAG_RWTUN,
188 &enable_vmem_check, 0, "Enable vmem check");
189 static void vmem_check(vmem_t *);
190 #endif
191
192 static struct callout vmem_periodic_ch;
193 static int vmem_periodic_interval;
194 static struct task vmem_periodic_wk;
195
196 static struct mtx_padalign __exclusive_cache_line vmem_list_lock;
197 static LIST_HEAD(, vmem) vmem_list = LIST_HEAD_INITIALIZER(vmem_list);
198 static uma_zone_t vmem_zone;
199
200 /* ---- misc */
201 #define VMEM_CONDVAR_INIT(vm, wchan) cv_init(&vm->vm_cv, wchan)
202 #define VMEM_CONDVAR_DESTROY(vm) cv_destroy(&vm->vm_cv)
203 #define VMEM_CONDVAR_WAIT(vm) cv_wait(&vm->vm_cv, &vm->vm_lock)
204 #define VMEM_CONDVAR_BROADCAST(vm) cv_broadcast(&vm->vm_cv)
205
206
207 #define VMEM_LOCK(vm) mtx_lock(&vm->vm_lock)
208 #define VMEM_TRYLOCK(vm) mtx_trylock(&vm->vm_lock)
209 #define VMEM_UNLOCK(vm) mtx_unlock(&vm->vm_lock)
210 #define VMEM_LOCK_INIT(vm, name) mtx_init(&vm->vm_lock, (name), NULL, MTX_DEF)
211 #define VMEM_LOCK_DESTROY(vm) mtx_destroy(&vm->vm_lock)
212 #define VMEM_ASSERT_LOCKED(vm) mtx_assert(&vm->vm_lock, MA_OWNED);
213
214 #define VMEM_ALIGNUP(addr, align) (-(-(addr) & -(align)))
215
216 #define VMEM_CROSS_P(addr1, addr2, boundary) \
217 ((((addr1) ^ (addr2)) & -(boundary)) != 0)
218
219 #define ORDER2SIZE(order) ((order) < VMEM_OPTVALUE ? ((order) + 1) : \
220 (vmem_size_t)1 << ((order) - (VMEM_OPTVALUE - VMEM_OPTORDER - 1)))
221 #define SIZE2ORDER(size) ((size) <= VMEM_OPTVALUE ? ((size) - 1) : \
222 (flsl(size) + (VMEM_OPTVALUE - VMEM_OPTORDER - 2)))
223
224 /*
225 * Maximum number of boundary tags that may be required to satisfy an
226 * allocation. Two may be required to import. Another two may be
227 * required to clip edges.
228 */
229 #define BT_MAXALLOC 4
230
231 /*
232 * Max free limits the number of locally cached boundary tags. We
233 * just want to avoid hitting the zone allocator for every call.
234 */
235 #define BT_MAXFREE (BT_MAXALLOC * 8)
236
237 /* Allocator for boundary tags. */
238 static uma_zone_t vmem_bt_zone;
239
240 /* boot time arena storage. */
241 static struct vmem kernel_arena_storage;
242 static struct vmem buffer_arena_storage;
243 static struct vmem transient_arena_storage;
244 /* kernel and kmem arenas are aliased for backwards KPI compat. */
245 vmem_t *kernel_arena = &kernel_arena_storage;
246 vmem_t *kmem_arena = &kernel_arena_storage;
247 vmem_t *buffer_arena = &buffer_arena_storage;
248 vmem_t *transient_arena = &transient_arena_storage;
249
250 #ifdef DEBUG_MEMGUARD
251 static struct vmem memguard_arena_storage;
252 vmem_t *memguard_arena = &memguard_arena_storage;
253 #endif
254
255 /*
256 * Fill the vmem's boundary tag cache. We guarantee that boundary tag
257 * allocation will not fail once bt_fill() passes. To do so we cache
258 * at least the maximum possible tag allocations in the arena.
259 */
260 static __noinline int
261 _bt_fill(vmem_t *vm, int flags)
262 {
263 bt_t *bt;
264
265 VMEM_ASSERT_LOCKED(vm);
266
267 /*
268 * Only allow the kernel arena and arenas derived from kernel arena to
269 * dip into reserve tags. They are where new tags come from.
270 */
271 flags &= BT_FLAGS;
272 if (vm != kernel_arena && vm->vm_arg != kernel_arena)
273 flags &= ~M_USE_RESERVE;
274
275 /*
276 * Loop until we meet the reserve. To minimize the lock shuffle
277 * and prevent simultaneous fills we first try a NOWAIT regardless
278 * of the caller's flags. Specify M_NOVM so we don't recurse while
279 * holding a vmem lock.
280 */
281 while (vm->vm_nfreetags < BT_MAXALLOC) {
282 bt = uma_zalloc(vmem_bt_zone,
283 (flags & M_USE_RESERVE) | M_NOWAIT | M_NOVM);
284 if (bt == NULL) {
285 VMEM_UNLOCK(vm);
286 bt = uma_zalloc(vmem_bt_zone, flags);
287 VMEM_LOCK(vm);
288 if (bt == NULL)
289 break;
290 }
291 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
292 vm->vm_nfreetags++;
293 }
294
295 if (vm->vm_nfreetags < BT_MAXALLOC)
296 return ENOMEM;
297
298 return 0;
299 }
300
301 static inline int
302 bt_fill(vmem_t *vm, int flags)
303 {
304 if (vm->vm_nfreetags >= BT_MAXALLOC)
305 return (0);
306 return (_bt_fill(vm, flags));
307 }
308
309 /*
310 * Pop a tag off of the freetag stack.
311 */
312 static bt_t *
313 bt_alloc(vmem_t *vm)
314 {
315 bt_t *bt;
316
317 VMEM_ASSERT_LOCKED(vm);
318 bt = LIST_FIRST(&vm->vm_freetags);
319 MPASS(bt != NULL);
320 LIST_REMOVE(bt, bt_freelist);
321 vm->vm_nfreetags--;
322
323 return bt;
324 }
325
326 /*
327 * Trim the per-vmem free list. Returns with the lock released to
328 * avoid allocator recursions.
329 */
330 static void
331 bt_freetrim(vmem_t *vm, int freelimit)
332 {
333 LIST_HEAD(, vmem_btag) freetags;
334 bt_t *bt;
335
336 LIST_INIT(&freetags);
337 VMEM_ASSERT_LOCKED(vm);
338 while (vm->vm_nfreetags > freelimit) {
339 bt = LIST_FIRST(&vm->vm_freetags);
340 LIST_REMOVE(bt, bt_freelist);
341 vm->vm_nfreetags--;
342 LIST_INSERT_HEAD(&freetags, bt, bt_freelist);
343 }
344 VMEM_UNLOCK(vm);
345 while ((bt = LIST_FIRST(&freetags)) != NULL) {
346 LIST_REMOVE(bt, bt_freelist);
347 uma_zfree(vmem_bt_zone, bt);
348 }
349 }
350
351 static inline void
352 bt_free(vmem_t *vm, bt_t *bt)
353 {
354
355 VMEM_ASSERT_LOCKED(vm);
356 MPASS(LIST_FIRST(&vm->vm_freetags) != bt);
357 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
358 vm->vm_nfreetags++;
359 }
360
361 /*
362 * Hide MAXALLOC tags before dropping the arena lock to ensure that a
363 * concurrent allocation attempt does not grab them.
364 */
365 static void
366 bt_save(vmem_t *vm)
367 {
368 KASSERT(vm->vm_nfreetags >= BT_MAXALLOC,
369 ("%s: insufficient free tags %d", __func__, vm->vm_nfreetags));
370 vm->vm_nfreetags -= BT_MAXALLOC;
371 }
372
373 static void
374 bt_restore(vmem_t *vm)
375 {
376 vm->vm_nfreetags += BT_MAXALLOC;
377 }
378
379 /*
380 * freelist[0] ... [1, 1]
381 * freelist[1] ... [2, 2]
382 * :
383 * freelist[29] ... [30, 30]
384 * freelist[30] ... [31, 31]
385 * freelist[31] ... [32, 63]
386 * freelist[33] ... [64, 127]
387 * :
388 * freelist[n] ... [(1 << (n - 26)), (1 << (n - 25)) - 1]
389 * :
390 */
391
392 static struct vmem_freelist *
393 bt_freehead_tofree(vmem_t *vm, vmem_size_t size)
394 {
395 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
396 const int idx = SIZE2ORDER(qsize);
397
398 MPASS(size != 0 && qsize != 0);
399 MPASS((size & vm->vm_quantum_mask) == 0);
400 MPASS(idx >= 0);
401 MPASS(idx < VMEM_MAXORDER);
402
403 return &vm->vm_freelist[idx];
404 }
405
406 /*
407 * bt_freehead_toalloc: return the freelist for the given size and allocation
408 * strategy.
409 *
410 * For M_FIRSTFIT, return the list in which any blocks are large enough
411 * for the requested size. otherwise, return the list which can have blocks
412 * large enough for the requested size.
413 */
414 static struct vmem_freelist *
415 bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, int strat)
416 {
417 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
418 int idx = SIZE2ORDER(qsize);
419
420 MPASS(size != 0 && qsize != 0);
421 MPASS((size & vm->vm_quantum_mask) == 0);
422
423 if (strat == M_FIRSTFIT && ORDER2SIZE(idx) != qsize) {
424 idx++;
425 /* check too large request? */
426 }
427 MPASS(idx >= 0);
428 MPASS(idx < VMEM_MAXORDER);
429
430 return &vm->vm_freelist[idx];
431 }
432
433 /* ---- boundary tag hash */
434
435 static struct vmem_hashlist *
436 bt_hashhead(vmem_t *vm, vmem_addr_t addr)
437 {
438 struct vmem_hashlist *list;
439 unsigned int hash;
440
441 hash = hash32_buf(&addr, sizeof(addr), 0);
442 list = &vm->vm_hashlist[hash % vm->vm_hashsize];
443
444 return list;
445 }
446
447 static bt_t *
448 bt_lookupbusy(vmem_t *vm, vmem_addr_t addr)
449 {
450 struct vmem_hashlist *list;
451 bt_t *bt;
452
453 VMEM_ASSERT_LOCKED(vm);
454 list = bt_hashhead(vm, addr);
455 LIST_FOREACH(bt, list, bt_hashlist) {
456 if (bt->bt_start == addr) {
457 break;
458 }
459 }
460
461 return bt;
462 }
463
464 static void
465 bt_rembusy(vmem_t *vm, bt_t *bt)
466 {
467
468 VMEM_ASSERT_LOCKED(vm);
469 MPASS(vm->vm_nbusytag > 0);
470 vm->vm_inuse -= bt->bt_size;
471 vm->vm_nbusytag--;
472 LIST_REMOVE(bt, bt_hashlist);
473 }
474
475 static void
476 bt_insbusy(vmem_t *vm, bt_t *bt)
477 {
478 struct vmem_hashlist *list;
479
480 VMEM_ASSERT_LOCKED(vm);
481 MPASS(bt->bt_type == BT_TYPE_BUSY);
482
483 list = bt_hashhead(vm, bt->bt_start);
484 LIST_INSERT_HEAD(list, bt, bt_hashlist);
485 vm->vm_nbusytag++;
486 vm->vm_inuse += bt->bt_size;
487 }
488
489 /* ---- boundary tag list */
490
491 static void
492 bt_remseg(vmem_t *vm, bt_t *bt)
493 {
494
495 MPASS(bt->bt_type != BT_TYPE_CURSOR);
496 TAILQ_REMOVE(&vm->vm_seglist, bt, bt_seglist);
497 bt_free(vm, bt);
498 }
499
500 static void
501 bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev)
502 {
503
504 TAILQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist);
505 }
506
507 static void
508 bt_insseg_tail(vmem_t *vm, bt_t *bt)
509 {
510
511 TAILQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist);
512 }
513
514 static void
515 bt_remfree(vmem_t *vm __unused, bt_t *bt)
516 {
517
518 MPASS(bt->bt_type == BT_TYPE_FREE);
519
520 LIST_REMOVE(bt, bt_freelist);
521 }
522
523 static void
524 bt_insfree(vmem_t *vm, bt_t *bt)
525 {
526 struct vmem_freelist *list;
527
528 list = bt_freehead_tofree(vm, bt->bt_size);
529 LIST_INSERT_HEAD(list, bt, bt_freelist);
530 }
531
532 /* ---- vmem internal functions */
533
534 /*
535 * Import from the arena into the quantum cache in UMA.
536 *
537 * We use VMEM_ADDR_QCACHE_MIN instead of 0: uma_zalloc() returns 0 to indicate
538 * failure, so UMA can't be used to cache a resource with value 0.
539 */
540 static int
541 qc_import(void *arg, void **store, int cnt, int domain, int flags)
542 {
543 qcache_t *qc;
544 vmem_addr_t addr;
545 int i;
546
547 KASSERT((flags & M_WAITOK) == 0, ("blocking allocation"));
548
549 qc = arg;
550 for (i = 0; i < cnt; i++) {
551 if (vmem_xalloc(qc->qc_vmem, qc->qc_size, 0, 0, 0,
552 VMEM_ADDR_QCACHE_MIN, VMEM_ADDR_MAX, flags, &addr) != 0)
553 break;
554 store[i] = (void *)addr;
555 }
556 return (i);
557 }
558
559 /*
560 * Release memory from the UMA cache to the arena.
561 */
562 static void
563 qc_release(void *arg, void **store, int cnt)
564 {
565 qcache_t *qc;
566 int i;
567
568 qc = arg;
569 for (i = 0; i < cnt; i++)
570 vmem_xfree(qc->qc_vmem, (vmem_addr_t)store[i], qc->qc_size);
571 }
572
573 static void
574 qc_init(vmem_t *vm, vmem_size_t qcache_max)
575 {
576 qcache_t *qc;
577 vmem_size_t size;
578 int qcache_idx_max;
579 int i;
580
581 MPASS((qcache_max & vm->vm_quantum_mask) == 0);
582 qcache_idx_max = MIN(qcache_max >> vm->vm_quantum_shift,
583 VMEM_QCACHE_IDX_MAX);
584 vm->vm_qcache_max = qcache_idx_max << vm->vm_quantum_shift;
585 for (i = 0; i < qcache_idx_max; i++) {
586 qc = &vm->vm_qcache[i];
587 size = (i + 1) << vm->vm_quantum_shift;
588 snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu",
589 vm->vm_name, size);
590 qc->qc_vmem = vm;
591 qc->qc_size = size;
592 qc->qc_cache = uma_zcache_create(qc->qc_name, size,
593 NULL, NULL, NULL, NULL, qc_import, qc_release, qc,
594 UMA_ZONE_VM);
595 MPASS(qc->qc_cache);
596 }
597 }
598
599 static void
600 qc_destroy(vmem_t *vm)
601 {
602 int qcache_idx_max;
603 int i;
604
605 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
606 for (i = 0; i < qcache_idx_max; i++)
607 uma_zdestroy(vm->vm_qcache[i].qc_cache);
608 }
609
610 static void
611 qc_drain(vmem_t *vm)
612 {
613 int qcache_idx_max;
614 int i;
615
616 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
617 for (i = 0; i < qcache_idx_max; i++)
618 zone_drain(vm->vm_qcache[i].qc_cache);
619 }
620
621 #ifndef UMA_MD_SMALL_ALLOC
622
623 static struct mtx_padalign __exclusive_cache_line vmem_bt_lock;
624
625 /*
626 * vmem_bt_alloc: Allocate a new page of boundary tags.
627 *
628 * On architectures with uma_small_alloc there is no recursion; no address
629 * space need be allocated to allocate boundary tags. For the others, we
630 * must handle recursion. Boundary tags are necessary to allocate new
631 * boundary tags.
632 *
633 * UMA guarantees that enough tags are held in reserve to allocate a new
634 * page of kva. We dip into this reserve by specifying M_USE_RESERVE only
635 * when allocating the page to hold new boundary tags. In this way the
636 * reserve is automatically filled by the allocation that uses the reserve.
637 *
638 * We still have to guarantee that the new tags are allocated atomically since
639 * many threads may try concurrently. The bt_lock provides this guarantee.
640 * We convert WAITOK allocations to NOWAIT and then handle the blocking here
641 * on failure. It's ok to return NULL for a WAITOK allocation as UMA will
642 * loop again after checking to see if we lost the race to allocate.
643 *
644 * There is a small race between vmem_bt_alloc() returning the page and the
645 * zone lock being acquired to add the page to the zone. For WAITOK
646 * allocations we just pause briefly. NOWAIT may experience a transient
647 * failure. To alleviate this we permit a small number of simultaneous
648 * fills to proceed concurrently so NOWAIT is less likely to fail unless
649 * we are really out of KVA.
650 */
651 static void *
652 vmem_bt_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
653 int wait)
654 {
655 vmem_addr_t addr;
656
657 *pflag = UMA_SLAB_KERNEL;
658
659 /*
660 * Single thread boundary tag allocation so that the address space
661 * and memory are added in one atomic operation.
662 */
663 mtx_lock(&vmem_bt_lock);
664 if (vmem_xalloc(vm_dom[domain].vmd_kernel_arena, bytes, 0, 0, 0,
665 VMEM_ADDR_MIN, VMEM_ADDR_MAX,
666 M_NOWAIT | M_NOVM | M_USE_RESERVE | M_BESTFIT, &addr) == 0) {
667 if (kmem_back_domain(domain, kernel_object, addr, bytes,
668 M_NOWAIT | M_USE_RESERVE) == 0) {
669 mtx_unlock(&vmem_bt_lock);
670 return ((void *)addr);
671 }
672 vmem_xfree(vm_dom[domain].vmd_kernel_arena, addr, bytes);
673 mtx_unlock(&vmem_bt_lock);
674 /*
675 * Out of memory, not address space. This may not even be
676 * possible due to M_USE_RESERVE page allocation.
677 */
678 if (wait & M_WAITOK)
679 vm_wait_domain(domain);
680 return (NULL);
681 }
682 mtx_unlock(&vmem_bt_lock);
683 /*
684 * We're either out of address space or lost a fill race.
685 */
686 if (wait & M_WAITOK)
687 pause("btalloc", 1);
688
689 return (NULL);
690 }
691
692 /*
693 * How many pages do we need to startup_alloc.
694 */
695 int
696 vmem_startup_count(void)
697 {
698
699 return (howmany(BT_MAXALLOC,
700 UMA_SLAB_SPACE / sizeof(struct vmem_btag)));
701 }
702 #endif
703
704 void
705 vmem_startup(void)
706 {
707
708 mtx_init(&vmem_list_lock, "vmem list lock", NULL, MTX_DEF);
709 vmem_zone = uma_zcreate("vmem",
710 sizeof(struct vmem), NULL, NULL, NULL, NULL,
711 UMA_ALIGN_PTR, UMA_ZONE_VM);
712 vmem_bt_zone = uma_zcreate("vmem btag",
713 sizeof(struct vmem_btag), NULL, NULL, NULL, NULL,
714 UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE);
715 #ifndef UMA_MD_SMALL_ALLOC
716 mtx_init(&vmem_bt_lock, "btag lock", NULL, MTX_DEF);
717 uma_prealloc(vmem_bt_zone, BT_MAXALLOC);
718 /*
719 * Reserve enough tags to allocate new tags. We allow multiple
720 * CPUs to attempt to allocate new tags concurrently to limit
721 * false restarts in UMA. vmem_bt_alloc() allocates from a per-domain
722 * arena, which may involve importing a range from the kernel arena,
723 * so we need to keep at least 2 * BT_MAXALLOC tags reserved.
724 */
725 uma_zone_reserve(vmem_bt_zone, 2 * BT_MAXALLOC * mp_ncpus);
726 uma_zone_set_allocf(vmem_bt_zone, vmem_bt_alloc);
727 #endif
728 }
729
730 /* ---- rehash */
731
732 static int
733 vmem_rehash(vmem_t *vm, vmem_size_t newhashsize)
734 {
735 bt_t *bt;
736 struct vmem_hashlist *newhashlist;
737 struct vmem_hashlist *oldhashlist;
738 vmem_size_t i, oldhashsize;
739
740 MPASS(newhashsize > 0);
741
742 newhashlist = malloc(sizeof(struct vmem_hashlist) * newhashsize,
743 M_VMEM, M_NOWAIT);
744 if (newhashlist == NULL)
745 return ENOMEM;
746 for (i = 0; i < newhashsize; i++) {
747 LIST_INIT(&newhashlist[i]);
748 }
749
750 VMEM_LOCK(vm);
751 oldhashlist = vm->vm_hashlist;
752 oldhashsize = vm->vm_hashsize;
753 vm->vm_hashlist = newhashlist;
754 vm->vm_hashsize = newhashsize;
755 if (oldhashlist == NULL) {
756 VMEM_UNLOCK(vm);
757 return 0;
758 }
759 for (i = 0; i < oldhashsize; i++) {
760 while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) {
761 bt_rembusy(vm, bt);
762 bt_insbusy(vm, bt);
763 }
764 }
765 VMEM_UNLOCK(vm);
766
767 if (oldhashlist != vm->vm_hash0)
768 free(oldhashlist, M_VMEM);
769
770 return 0;
771 }
772
773 static void
774 vmem_periodic_kick(void *dummy)
775 {
776
777 taskqueue_enqueue(taskqueue_thread, &vmem_periodic_wk);
778 }
779
780 static void
781 vmem_periodic(void *unused, int pending)
782 {
783 vmem_t *vm;
784 vmem_size_t desired;
785 vmem_size_t current;
786
787 mtx_lock(&vmem_list_lock);
788 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
789 #ifdef DIAGNOSTIC
790 /* Convenient time to verify vmem state. */
791 if (enable_vmem_check == 1) {
792 VMEM_LOCK(vm);
793 vmem_check(vm);
794 VMEM_UNLOCK(vm);
795 }
796 #endif
797 desired = 1 << flsl(vm->vm_nbusytag);
798 desired = MIN(MAX(desired, VMEM_HASHSIZE_MIN),
799 VMEM_HASHSIZE_MAX);
800 current = vm->vm_hashsize;
801
802 /* Grow in powers of two. Shrink less aggressively. */
803 if (desired >= current * 2 || desired * 4 <= current)
804 vmem_rehash(vm, desired);
805
806 /*
807 * Periodically wake up threads waiting for resources,
808 * so they could ask for reclamation again.
809 */
810 VMEM_CONDVAR_BROADCAST(vm);
811 }
812 mtx_unlock(&vmem_list_lock);
813
814 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
815 vmem_periodic_kick, NULL);
816 }
817
818 static void
819 vmem_start_callout(void *unused)
820 {
821
822 TASK_INIT(&vmem_periodic_wk, 0, vmem_periodic, NULL);
823 vmem_periodic_interval = hz * 10;
824 callout_init(&vmem_periodic_ch, 1);
825 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
826 vmem_periodic_kick, NULL);
827 }
828 SYSINIT(vfs, SI_SUB_CONFIGURE, SI_ORDER_ANY, vmem_start_callout, NULL);
829
830 static void
831 vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int type)
832 {
833 bt_t *btspan;
834 bt_t *btfree;
835
836 MPASS(type == BT_TYPE_SPAN || type == BT_TYPE_SPAN_STATIC);
837 MPASS((size & vm->vm_quantum_mask) == 0);
838
839 btspan = bt_alloc(vm);
840 btspan->bt_type = type;
841 btspan->bt_start = addr;
842 btspan->bt_size = size;
843 bt_insseg_tail(vm, btspan);
844
845 btfree = bt_alloc(vm);
846 btfree->bt_type = BT_TYPE_FREE;
847 btfree->bt_start = addr;
848 btfree->bt_size = size;
849 bt_insseg(vm, btfree, btspan);
850 bt_insfree(vm, btfree);
851
852 vm->vm_size += size;
853 }
854
855 static void
856 vmem_destroy1(vmem_t *vm)
857 {
858 bt_t *bt;
859
860 /*
861 * Drain per-cpu quantum caches.
862 */
863 qc_destroy(vm);
864
865 /*
866 * The vmem should now only contain empty segments.
867 */
868 VMEM_LOCK(vm);
869 MPASS(vm->vm_nbusytag == 0);
870
871 TAILQ_REMOVE(&vm->vm_seglist, &vm->vm_cursor, bt_seglist);
872 while ((bt = TAILQ_FIRST(&vm->vm_seglist)) != NULL)
873 bt_remseg(vm, bt);
874
875 if (vm->vm_hashlist != NULL && vm->vm_hashlist != vm->vm_hash0)
876 free(vm->vm_hashlist, M_VMEM);
877
878 bt_freetrim(vm, 0);
879
880 VMEM_CONDVAR_DESTROY(vm);
881 VMEM_LOCK_DESTROY(vm);
882 uma_zfree(vmem_zone, vm);
883 }
884
885 static int
886 vmem_import(vmem_t *vm, vmem_size_t size, vmem_size_t align, int flags)
887 {
888 vmem_addr_t addr;
889 int error;
890
891 if (vm->vm_importfn == NULL)
892 return (EINVAL);
893
894 /*
895 * To make sure we get a span that meets the alignment we double it
896 * and add the size to the tail. This slightly overestimates.
897 */
898 if (align != vm->vm_quantum_mask + 1)
899 size = (align * 2) + size;
900 size = roundup(size, vm->vm_import_quantum);
901
902 if (vm->vm_limit != 0 && vm->vm_limit < vm->vm_size + size)
903 return (ENOMEM);
904
905 bt_save(vm);
906 VMEM_UNLOCK(vm);
907 error = (vm->vm_importfn)(vm->vm_arg, size, flags, &addr);
908 VMEM_LOCK(vm);
909 bt_restore(vm);
910 if (error)
911 return (ENOMEM);
912
913 vmem_add1(vm, addr, size, BT_TYPE_SPAN);
914
915 return 0;
916 }
917
918 /*
919 * vmem_fit: check if a bt can satisfy the given restrictions.
920 *
921 * it's a caller's responsibility to ensure the region is big enough
922 * before calling us.
923 */
924 static int
925 vmem_fit(const bt_t *bt, vmem_size_t size, vmem_size_t align,
926 vmem_size_t phase, vmem_size_t nocross, vmem_addr_t minaddr,
927 vmem_addr_t maxaddr, vmem_addr_t *addrp)
928 {
929 vmem_addr_t start;
930 vmem_addr_t end;
931
932 MPASS(size > 0);
933 MPASS(bt->bt_size >= size); /* caller's responsibility */
934
935 /*
936 * XXX assumption: vmem_addr_t and vmem_size_t are
937 * unsigned integer of the same size.
938 */
939
940 start = bt->bt_start;
941 if (start < minaddr) {
942 start = minaddr;
943 }
944 end = BT_END(bt);
945 if (end > maxaddr)
946 end = maxaddr;
947 if (start > end)
948 return (ENOMEM);
949
950 start = VMEM_ALIGNUP(start - phase, align) + phase;
951 if (start < bt->bt_start)
952 start += align;
953 if (VMEM_CROSS_P(start, start + size - 1, nocross)) {
954 MPASS(align < nocross);
955 start = VMEM_ALIGNUP(start - phase, nocross) + phase;
956 }
957 if (start <= end && end - start >= size - 1) {
958 MPASS((start & (align - 1)) == phase);
959 MPASS(!VMEM_CROSS_P(start, start + size - 1, nocross));
960 MPASS(minaddr <= start);
961 MPASS(maxaddr == 0 || start + size - 1 <= maxaddr);
962 MPASS(bt->bt_start <= start);
963 MPASS(BT_END(bt) - start >= size - 1);
964 *addrp = start;
965
966 return (0);
967 }
968 return (ENOMEM);
969 }
970
971 /*
972 * vmem_clip: Trim the boundary tag edges to the requested start and size.
973 */
974 static void
975 vmem_clip(vmem_t *vm, bt_t *bt, vmem_addr_t start, vmem_size_t size)
976 {
977 bt_t *btnew;
978 bt_t *btprev;
979
980 VMEM_ASSERT_LOCKED(vm);
981 MPASS(bt->bt_type == BT_TYPE_FREE);
982 MPASS(bt->bt_size >= size);
983 bt_remfree(vm, bt);
984 if (bt->bt_start != start) {
985 btprev = bt_alloc(vm);
986 btprev->bt_type = BT_TYPE_FREE;
987 btprev->bt_start = bt->bt_start;
988 btprev->bt_size = start - bt->bt_start;
989 bt->bt_start = start;
990 bt->bt_size -= btprev->bt_size;
991 bt_insfree(vm, btprev);
992 bt_insseg(vm, btprev,
993 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
994 }
995 MPASS(bt->bt_start == start);
996 if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) {
997 /* split */
998 btnew = bt_alloc(vm);
999 btnew->bt_type = BT_TYPE_BUSY;
1000 btnew->bt_start = bt->bt_start;
1001 btnew->bt_size = size;
1002 bt->bt_start = bt->bt_start + size;
1003 bt->bt_size -= size;
1004 bt_insfree(vm, bt);
1005 bt_insseg(vm, btnew,
1006 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
1007 bt_insbusy(vm, btnew);
1008 bt = btnew;
1009 } else {
1010 bt->bt_type = BT_TYPE_BUSY;
1011 bt_insbusy(vm, bt);
1012 }
1013 MPASS(bt->bt_size >= size);
1014 }
1015
1016 static int
1017 vmem_try_fetch(vmem_t *vm, const vmem_size_t size, vmem_size_t align, int flags)
1018 {
1019 vmem_size_t avail;
1020
1021 VMEM_ASSERT_LOCKED(vm);
1022
1023 /*
1024 * XXX it is possible to fail to meet xalloc constraints with the
1025 * imported region. It is up to the user to specify the
1026 * import quantum such that it can satisfy any allocation.
1027 */
1028 if (vmem_import(vm, size, align, flags) == 0)
1029 return (1);
1030
1031 /*
1032 * Try to free some space from the quantum cache or reclaim
1033 * functions if available.
1034 */
1035 if (vm->vm_qcache_max != 0 || vm->vm_reclaimfn != NULL) {
1036 avail = vm->vm_size - vm->vm_inuse;
1037 bt_save(vm);
1038 VMEM_UNLOCK(vm);
1039 if (vm->vm_qcache_max != 0)
1040 qc_drain(vm);
1041 if (vm->vm_reclaimfn != NULL)
1042 vm->vm_reclaimfn(vm, flags);
1043 VMEM_LOCK(vm);
1044 bt_restore(vm);
1045 /* If we were successful retry even NOWAIT. */
1046 if (vm->vm_size - vm->vm_inuse > avail)
1047 return (1);
1048 }
1049 if ((flags & M_NOWAIT) != 0)
1050 return (0);
1051 bt_save(vm);
1052 VMEM_CONDVAR_WAIT(vm);
1053 bt_restore(vm);
1054 return (1);
1055 }
1056
1057 static int
1058 vmem_try_release(vmem_t *vm, struct vmem_btag *bt, const bool remfree)
1059 {
1060 struct vmem_btag *prev;
1061
1062 MPASS(bt->bt_type == BT_TYPE_FREE);
1063
1064 if (vm->vm_releasefn == NULL)
1065 return (0);
1066
1067 prev = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1068 MPASS(prev != NULL);
1069 MPASS(prev->bt_type != BT_TYPE_FREE);
1070
1071 if (prev->bt_type == BT_TYPE_SPAN && prev->bt_size == bt->bt_size) {
1072 vmem_addr_t spanaddr;
1073 vmem_size_t spansize;
1074
1075 MPASS(prev->bt_start == bt->bt_start);
1076 spanaddr = prev->bt_start;
1077 spansize = prev->bt_size;
1078 if (remfree)
1079 bt_remfree(vm, bt);
1080 bt_remseg(vm, bt);
1081 bt_remseg(vm, prev);
1082 vm->vm_size -= spansize;
1083 VMEM_CONDVAR_BROADCAST(vm);
1084 bt_freetrim(vm, BT_MAXFREE);
1085 vm->vm_releasefn(vm->vm_arg, spanaddr, spansize);
1086 return (1);
1087 }
1088 return (0);
1089 }
1090
1091 static int
1092 vmem_xalloc_nextfit(vmem_t *vm, const vmem_size_t size, vmem_size_t align,
1093 const vmem_size_t phase, const vmem_size_t nocross, int flags,
1094 vmem_addr_t *addrp)
1095 {
1096 struct vmem_btag *bt, *cursor, *next, *prev;
1097 int error;
1098
1099 error = ENOMEM;
1100 VMEM_LOCK(vm);
1101
1102 /*
1103 * Make sure we have enough tags to complete the operation.
1104 */
1105 if (bt_fill(vm, flags) != 0)
1106 goto out;
1107
1108 retry:
1109 /*
1110 * Find the next free tag meeting our constraints. If one is found,
1111 * perform the allocation.
1112 */
1113 for (cursor = &vm->vm_cursor, bt = TAILQ_NEXT(cursor, bt_seglist);
1114 bt != cursor; bt = TAILQ_NEXT(bt, bt_seglist)) {
1115 if (bt == NULL)
1116 bt = TAILQ_FIRST(&vm->vm_seglist);
1117 if (bt->bt_type == BT_TYPE_FREE && bt->bt_size >= size &&
1118 (error = vmem_fit(bt, size, align, phase, nocross,
1119 VMEM_ADDR_MIN, VMEM_ADDR_MAX, addrp)) == 0) {
1120 vmem_clip(vm, bt, *addrp, size);
1121 break;
1122 }
1123 }
1124
1125 /*
1126 * Try to coalesce free segments around the cursor. If we succeed, and
1127 * have not yet satisfied the allocation request, try again with the
1128 * newly coalesced segment.
1129 */
1130 if ((next = TAILQ_NEXT(cursor, bt_seglist)) != NULL &&
1131 (prev = TAILQ_PREV(cursor, vmem_seglist, bt_seglist)) != NULL &&
1132 next->bt_type == BT_TYPE_FREE && prev->bt_type == BT_TYPE_FREE &&
1133 prev->bt_start + prev->bt_size == next->bt_start) {
1134 prev->bt_size += next->bt_size;
1135 bt_remfree(vm, next);
1136 bt_remseg(vm, next);
1137
1138 /*
1139 * The coalesced segment might be able to satisfy our request.
1140 * If not, we might need to release it from the arena.
1141 */
1142 if (error == ENOMEM && prev->bt_size >= size &&
1143 (error = vmem_fit(prev, size, align, phase, nocross,
1144 VMEM_ADDR_MIN, VMEM_ADDR_MAX, addrp)) == 0) {
1145 vmem_clip(vm, prev, *addrp, size);
1146 bt = prev;
1147 } else
1148 (void)vmem_try_release(vm, prev, true);
1149 }
1150
1151 /*
1152 * If the allocation was successful, advance the cursor.
1153 */
1154 if (error == 0) {
1155 TAILQ_REMOVE(&vm->vm_seglist, cursor, bt_seglist);
1156 for (; bt != NULL && bt->bt_start < *addrp + size;
1157 bt = TAILQ_NEXT(bt, bt_seglist))
1158 ;
1159 if (bt != NULL)
1160 TAILQ_INSERT_BEFORE(bt, cursor, bt_seglist);
1161 else
1162 TAILQ_INSERT_HEAD(&vm->vm_seglist, cursor, bt_seglist);
1163 }
1164
1165 /*
1166 * Attempt to bring additional resources into the arena. If that fails
1167 * and M_WAITOK is specified, sleep waiting for resources to be freed.
1168 */
1169 if (error == ENOMEM && vmem_try_fetch(vm, size, align, flags))
1170 goto retry;
1171
1172 out:
1173 VMEM_UNLOCK(vm);
1174 return (error);
1175 }
1176
1177 /* ---- vmem API */
1178
1179 void
1180 vmem_set_import(vmem_t *vm, vmem_import_t *importfn,
1181 vmem_release_t *releasefn, void *arg, vmem_size_t import_quantum)
1182 {
1183
1184 VMEM_LOCK(vm);
1185 vm->vm_importfn = importfn;
1186 vm->vm_releasefn = releasefn;
1187 vm->vm_arg = arg;
1188 vm->vm_import_quantum = import_quantum;
1189 VMEM_UNLOCK(vm);
1190 }
1191
1192 void
1193 vmem_set_limit(vmem_t *vm, vmem_size_t limit)
1194 {
1195
1196 VMEM_LOCK(vm);
1197 vm->vm_limit = limit;
1198 VMEM_UNLOCK(vm);
1199 }
1200
1201 void
1202 vmem_set_reclaim(vmem_t *vm, vmem_reclaim_t *reclaimfn)
1203 {
1204
1205 VMEM_LOCK(vm);
1206 vm->vm_reclaimfn = reclaimfn;
1207 VMEM_UNLOCK(vm);
1208 }
1209
1210 /*
1211 * vmem_init: Initializes vmem arena.
1212 */
1213 vmem_t *
1214 vmem_init(vmem_t *vm, const char *name, vmem_addr_t base, vmem_size_t size,
1215 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1216 {
1217 vmem_size_t i;
1218
1219 MPASS(quantum > 0);
1220 MPASS((quantum & (quantum - 1)) == 0);
1221
1222 bzero(vm, sizeof(*vm));
1223
1224 VMEM_CONDVAR_INIT(vm, name);
1225 VMEM_LOCK_INIT(vm, name);
1226 vm->vm_nfreetags = 0;
1227 LIST_INIT(&vm->vm_freetags);
1228 strlcpy(vm->vm_name, name, sizeof(vm->vm_name));
1229 vm->vm_quantum_mask = quantum - 1;
1230 vm->vm_quantum_shift = flsl(quantum) - 1;
1231 vm->vm_nbusytag = 0;
1232 vm->vm_size = 0;
1233 vm->vm_limit = 0;
1234 vm->vm_inuse = 0;
1235 qc_init(vm, qcache_max);
1236
1237 TAILQ_INIT(&vm->vm_seglist);
1238 vm->vm_cursor.bt_start = vm->vm_cursor.bt_size = 0;
1239 vm->vm_cursor.bt_type = BT_TYPE_CURSOR;
1240 TAILQ_INSERT_TAIL(&vm->vm_seglist, &vm->vm_cursor, bt_seglist);
1241
1242 for (i = 0; i < VMEM_MAXORDER; i++)
1243 LIST_INIT(&vm->vm_freelist[i]);
1244
1245 memset(&vm->vm_hash0, 0, sizeof(vm->vm_hash0));
1246 vm->vm_hashsize = VMEM_HASHSIZE_MIN;
1247 vm->vm_hashlist = vm->vm_hash0;
1248
1249 if (size != 0) {
1250 if (vmem_add(vm, base, size, flags) != 0) {
1251 vmem_destroy1(vm);
1252 return NULL;
1253 }
1254 }
1255
1256 mtx_lock(&vmem_list_lock);
1257 LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist);
1258 mtx_unlock(&vmem_list_lock);
1259
1260 return vm;
1261 }
1262
1263 /*
1264 * vmem_create: create an arena.
1265 */
1266 vmem_t *
1267 vmem_create(const char *name, vmem_addr_t base, vmem_size_t size,
1268 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1269 {
1270
1271 vmem_t *vm;
1272
1273 vm = uma_zalloc(vmem_zone, flags & (M_WAITOK|M_NOWAIT));
1274 if (vm == NULL)
1275 return (NULL);
1276 if (vmem_init(vm, name, base, size, quantum, qcache_max,
1277 flags) == NULL)
1278 return (NULL);
1279 return (vm);
1280 }
1281
1282 void
1283 vmem_destroy(vmem_t *vm)
1284 {
1285
1286 mtx_lock(&vmem_list_lock);
1287 LIST_REMOVE(vm, vm_alllist);
1288 mtx_unlock(&vmem_list_lock);
1289
1290 vmem_destroy1(vm);
1291 }
1292
1293 vmem_size_t
1294 vmem_roundup_size(vmem_t *vm, vmem_size_t size)
1295 {
1296
1297 return (size + vm->vm_quantum_mask) & ~vm->vm_quantum_mask;
1298 }
1299
1300 /*
1301 * vmem_alloc: allocate resource from the arena.
1302 */
1303 int
1304 vmem_alloc(vmem_t *vm, vmem_size_t size, int flags, vmem_addr_t *addrp)
1305 {
1306 const int strat __unused = flags & VMEM_FITMASK;
1307 qcache_t *qc;
1308
1309 flags &= VMEM_FLAGS;
1310 MPASS(size > 0);
1311 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT || strat == M_NEXTFIT);
1312 if ((flags & M_NOWAIT) == 0)
1313 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_alloc");
1314
1315 if (size <= vm->vm_qcache_max) {
1316 /*
1317 * Resource 0 cannot be cached, so avoid a blocking allocation
1318 * in qc_import() and give the vmem_xalloc() call below a chance
1319 * to return 0.
1320 */
1321 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1322 *addrp = (vmem_addr_t)uma_zalloc(qc->qc_cache,
1323 (flags & ~M_WAITOK) | M_NOWAIT);
1324 if (__predict_true(*addrp != 0))
1325 return (0);
1326 }
1327
1328 return (vmem_xalloc(vm, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
1329 flags, addrp));
1330 }
1331
1332 int
1333 vmem_xalloc(vmem_t *vm, const vmem_size_t size0, vmem_size_t align,
1334 const vmem_size_t phase, const vmem_size_t nocross,
1335 const vmem_addr_t minaddr, const vmem_addr_t maxaddr, int flags,
1336 vmem_addr_t *addrp)
1337 {
1338 const vmem_size_t size = vmem_roundup_size(vm, size0);
1339 struct vmem_freelist *list;
1340 struct vmem_freelist *first;
1341 struct vmem_freelist *end;
1342 bt_t *bt;
1343 int error;
1344 int strat;
1345
1346 flags &= VMEM_FLAGS;
1347 strat = flags & VMEM_FITMASK;
1348 MPASS(size0 > 0);
1349 MPASS(size > 0);
1350 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT || strat == M_NEXTFIT);
1351 MPASS((flags & (M_NOWAIT|M_WAITOK)) != (M_NOWAIT|M_WAITOK));
1352 if ((flags & M_NOWAIT) == 0)
1353 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_xalloc");
1354 MPASS((align & vm->vm_quantum_mask) == 0);
1355 MPASS((align & (align - 1)) == 0);
1356 MPASS((phase & vm->vm_quantum_mask) == 0);
1357 MPASS((nocross & vm->vm_quantum_mask) == 0);
1358 MPASS((nocross & (nocross - 1)) == 0);
1359 MPASS((align == 0 && phase == 0) || phase < align);
1360 MPASS(nocross == 0 || nocross >= size);
1361 MPASS(minaddr <= maxaddr);
1362 MPASS(!VMEM_CROSS_P(phase, phase + size - 1, nocross));
1363 if (strat == M_NEXTFIT)
1364 MPASS(minaddr == VMEM_ADDR_MIN && maxaddr == VMEM_ADDR_MAX);
1365
1366 if (align == 0)
1367 align = vm->vm_quantum_mask + 1;
1368 *addrp = 0;
1369
1370 /*
1371 * Next-fit allocations don't use the freelists.
1372 */
1373 if (strat == M_NEXTFIT)
1374 return (vmem_xalloc_nextfit(vm, size0, align, phase, nocross,
1375 flags, addrp));
1376
1377 end = &vm->vm_freelist[VMEM_MAXORDER];
1378 /*
1379 * choose a free block from which we allocate.
1380 */
1381 first = bt_freehead_toalloc(vm, size, strat);
1382 VMEM_LOCK(vm);
1383
1384 /*
1385 * Make sure we have enough tags to complete the operation.
1386 */
1387 error = bt_fill(vm, flags);
1388 if (error != 0)
1389 goto out;
1390 for (;;) {
1391 /*
1392 * Scan freelists looking for a tag that satisfies the
1393 * allocation. If we're doing BESTFIT we may encounter
1394 * sizes below the request. If we're doing FIRSTFIT we
1395 * inspect only the first element from each list.
1396 */
1397 for (list = first; list < end; list++) {
1398 LIST_FOREACH(bt, list, bt_freelist) {
1399 if (bt->bt_size >= size) {
1400 error = vmem_fit(bt, size, align, phase,
1401 nocross, minaddr, maxaddr, addrp);
1402 if (error == 0) {
1403 vmem_clip(vm, bt, *addrp, size);
1404 goto out;
1405 }
1406 }
1407 /* FIRST skips to the next list. */
1408 if (strat == M_FIRSTFIT)
1409 break;
1410 }
1411 }
1412
1413 /*
1414 * Retry if the fast algorithm failed.
1415 */
1416 if (strat == M_FIRSTFIT) {
1417 strat = M_BESTFIT;
1418 first = bt_freehead_toalloc(vm, size, strat);
1419 continue;
1420 }
1421
1422 /*
1423 * Try a few measures to bring additional resources into the
1424 * arena. If all else fails, we will sleep waiting for
1425 * resources to be freed.
1426 */
1427 if (!vmem_try_fetch(vm, size, align, flags)) {
1428 error = ENOMEM;
1429 break;
1430 }
1431 }
1432 out:
1433 VMEM_UNLOCK(vm);
1434 if (error != 0 && (flags & M_NOWAIT) == 0)
1435 panic("failed to allocate waiting allocation\n");
1436
1437 return (error);
1438 }
1439
1440 /*
1441 * vmem_free: free the resource to the arena.
1442 */
1443 void
1444 vmem_free(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1445 {
1446 qcache_t *qc;
1447 MPASS(size > 0);
1448
1449 if (size <= vm->vm_qcache_max &&
1450 __predict_true(addr >= VMEM_ADDR_QCACHE_MIN)) {
1451 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1452 uma_zfree(qc->qc_cache, (void *)addr);
1453 } else
1454 vmem_xfree(vm, addr, size);
1455 }
1456
1457 void
1458 vmem_xfree(vmem_t *vm, vmem_addr_t addr, vmem_size_t size __unused)
1459 {
1460 bt_t *bt;
1461 bt_t *t;
1462
1463 MPASS(size > 0);
1464
1465 VMEM_LOCK(vm);
1466 bt = bt_lookupbusy(vm, addr);
1467 MPASS(bt != NULL);
1468 MPASS(bt->bt_start == addr);
1469 MPASS(bt->bt_size == vmem_roundup_size(vm, size) ||
1470 bt->bt_size - vmem_roundup_size(vm, size) <= vm->vm_quantum_mask);
1471 MPASS(bt->bt_type == BT_TYPE_BUSY);
1472 bt_rembusy(vm, bt);
1473 bt->bt_type = BT_TYPE_FREE;
1474
1475 /* coalesce */
1476 t = TAILQ_NEXT(bt, bt_seglist);
1477 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1478 MPASS(BT_END(bt) < t->bt_start); /* YYY */
1479 bt->bt_size += t->bt_size;
1480 bt_remfree(vm, t);
1481 bt_remseg(vm, t);
1482 }
1483 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1484 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1485 MPASS(BT_END(t) < bt->bt_start); /* YYY */
1486 bt->bt_size += t->bt_size;
1487 bt->bt_start = t->bt_start;
1488 bt_remfree(vm, t);
1489 bt_remseg(vm, t);
1490 }
1491
1492 if (!vmem_try_release(vm, bt, false)) {
1493 bt_insfree(vm, bt);
1494 VMEM_CONDVAR_BROADCAST(vm);
1495 bt_freetrim(vm, BT_MAXFREE);
1496 }
1497 }
1498
1499 /*
1500 * vmem_add:
1501 *
1502 */
1503 int
1504 vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int flags)
1505 {
1506 int error;
1507
1508 flags &= VMEM_FLAGS;
1509
1510 VMEM_LOCK(vm);
1511 error = bt_fill(vm, flags);
1512 if (error == 0)
1513 vmem_add1(vm, addr, size, BT_TYPE_SPAN_STATIC);
1514 VMEM_UNLOCK(vm);
1515
1516 return (error);
1517 }
1518
1519 /*
1520 * vmem_size: information about arenas size
1521 */
1522 vmem_size_t
1523 vmem_size(vmem_t *vm, int typemask)
1524 {
1525 int i;
1526
1527 switch (typemask) {
1528 case VMEM_ALLOC:
1529 return vm->vm_inuse;
1530 case VMEM_FREE:
1531 return vm->vm_size - vm->vm_inuse;
1532 case VMEM_FREE|VMEM_ALLOC:
1533 return vm->vm_size;
1534 case VMEM_MAXFREE:
1535 VMEM_LOCK(vm);
1536 for (i = VMEM_MAXORDER - 1; i >= 0; i--) {
1537 if (LIST_EMPTY(&vm->vm_freelist[i]))
1538 continue;
1539 VMEM_UNLOCK(vm);
1540 return ((vmem_size_t)ORDER2SIZE(i) <<
1541 vm->vm_quantum_shift);
1542 }
1543 VMEM_UNLOCK(vm);
1544 return (0);
1545 default:
1546 panic("vmem_size");
1547 }
1548 }
1549
1550 /* ---- debug */
1551
1552 #if defined(DDB) || defined(DIAGNOSTIC)
1553
1554 static void bt_dump(const bt_t *, int (*)(const char *, ...)
1555 __printflike(1, 2));
1556
1557 static const char *
1558 bt_type_string(int type)
1559 {
1560
1561 switch (type) {
1562 case BT_TYPE_BUSY:
1563 return "busy";
1564 case BT_TYPE_FREE:
1565 return "free";
1566 case BT_TYPE_SPAN:
1567 return "span";
1568 case BT_TYPE_SPAN_STATIC:
1569 return "static span";
1570 case BT_TYPE_CURSOR:
1571 return "cursor";
1572 default:
1573 break;
1574 }
1575 return "BOGUS";
1576 }
1577
1578 static void
1579 bt_dump(const bt_t *bt, int (*pr)(const char *, ...))
1580 {
1581
1582 (*pr)("\t%p: %jx %jx, %d(%s)\n",
1583 bt, (intmax_t)bt->bt_start, (intmax_t)bt->bt_size,
1584 bt->bt_type, bt_type_string(bt->bt_type));
1585 }
1586
1587 static void
1588 vmem_dump(const vmem_t *vm , int (*pr)(const char *, ...) __printflike(1, 2))
1589 {
1590 const bt_t *bt;
1591 int i;
1592
1593 (*pr)("vmem %p '%s'\n", vm, vm->vm_name);
1594 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1595 bt_dump(bt, pr);
1596 }
1597
1598 for (i = 0; i < VMEM_MAXORDER; i++) {
1599 const struct vmem_freelist *fl = &vm->vm_freelist[i];
1600
1601 if (LIST_EMPTY(fl)) {
1602 continue;
1603 }
1604
1605 (*pr)("freelist[%d]\n", i);
1606 LIST_FOREACH(bt, fl, bt_freelist) {
1607 bt_dump(bt, pr);
1608 }
1609 }
1610 }
1611
1612 #endif /* defined(DDB) || defined(DIAGNOSTIC) */
1613
1614 #if defined(DDB)
1615 #include <ddb/ddb.h>
1616
1617 static bt_t *
1618 vmem_whatis_lookup(vmem_t *vm, vmem_addr_t addr)
1619 {
1620 bt_t *bt;
1621
1622 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1623 if (BT_ISSPAN_P(bt)) {
1624 continue;
1625 }
1626 if (bt->bt_start <= addr && addr <= BT_END(bt)) {
1627 return bt;
1628 }
1629 }
1630
1631 return NULL;
1632 }
1633
1634 void
1635 vmem_whatis(vmem_addr_t addr, int (*pr)(const char *, ...))
1636 {
1637 vmem_t *vm;
1638
1639 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1640 bt_t *bt;
1641
1642 bt = vmem_whatis_lookup(vm, addr);
1643 if (bt == NULL) {
1644 continue;
1645 }
1646 (*pr)("%p is %p+%zu in VMEM '%s' (%s)\n",
1647 (void *)addr, (void *)bt->bt_start,
1648 (vmem_size_t)(addr - bt->bt_start), vm->vm_name,
1649 (bt->bt_type == BT_TYPE_BUSY) ? "allocated" : "free");
1650 }
1651 }
1652
1653 void
1654 vmem_printall(const char *modif, int (*pr)(const char *, ...))
1655 {
1656 const vmem_t *vm;
1657
1658 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1659 vmem_dump(vm, pr);
1660 }
1661 }
1662
1663 void
1664 vmem_print(vmem_addr_t addr, const char *modif, int (*pr)(const char *, ...))
1665 {
1666 const vmem_t *vm = (const void *)addr;
1667
1668 vmem_dump(vm, pr);
1669 }
1670
1671 DB_SHOW_COMMAND(vmemdump, vmemdump)
1672 {
1673
1674 if (!have_addr) {
1675 db_printf("usage: show vmemdump <addr>\n");
1676 return;
1677 }
1678
1679 vmem_dump((const vmem_t *)addr, db_printf);
1680 }
1681
1682 DB_SHOW_ALL_COMMAND(vmemdump, vmemdumpall)
1683 {
1684 const vmem_t *vm;
1685
1686 LIST_FOREACH(vm, &vmem_list, vm_alllist)
1687 vmem_dump(vm, db_printf);
1688 }
1689
1690 DB_SHOW_COMMAND(vmem, vmem_summ)
1691 {
1692 const vmem_t *vm = (const void *)addr;
1693 const bt_t *bt;
1694 size_t ft[VMEM_MAXORDER], ut[VMEM_MAXORDER];
1695 size_t fs[VMEM_MAXORDER], us[VMEM_MAXORDER];
1696 int ord;
1697
1698 if (!have_addr) {
1699 db_printf("usage: show vmem <addr>\n");
1700 return;
1701 }
1702
1703 db_printf("vmem %p '%s'\n", vm, vm->vm_name);
1704 db_printf("\tquantum:\t%zu\n", vm->vm_quantum_mask + 1);
1705 db_printf("\tsize:\t%zu\n", vm->vm_size);
1706 db_printf("\tinuse:\t%zu\n", vm->vm_inuse);
1707 db_printf("\tfree:\t%zu\n", vm->vm_size - vm->vm_inuse);
1708 db_printf("\tbusy tags:\t%d\n", vm->vm_nbusytag);
1709 db_printf("\tfree tags:\t%d\n", vm->vm_nfreetags);
1710
1711 memset(&ft, 0, sizeof(ft));
1712 memset(&ut, 0, sizeof(ut));
1713 memset(&fs, 0, sizeof(fs));
1714 memset(&us, 0, sizeof(us));
1715 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1716 ord = SIZE2ORDER(bt->bt_size >> vm->vm_quantum_shift);
1717 if (bt->bt_type == BT_TYPE_BUSY) {
1718 ut[ord]++;
1719 us[ord] += bt->bt_size;
1720 } else if (bt->bt_type == BT_TYPE_FREE) {
1721 ft[ord]++;
1722 fs[ord] += bt->bt_size;
1723 }
1724 }
1725 db_printf("\t\t\tinuse\tsize\t\tfree\tsize\n");
1726 for (ord = 0; ord < VMEM_MAXORDER; ord++) {
1727 if (ut[ord] == 0 && ft[ord] == 0)
1728 continue;
1729 db_printf("\t%-15zu %zu\t%-15zu %zu\t%-16zu\n",
1730 ORDER2SIZE(ord) << vm->vm_quantum_shift,
1731 ut[ord], us[ord], ft[ord], fs[ord]);
1732 }
1733 }
1734
1735 DB_SHOW_ALL_COMMAND(vmem, vmem_summall)
1736 {
1737 const vmem_t *vm;
1738
1739 LIST_FOREACH(vm, &vmem_list, vm_alllist)
1740 vmem_summ((db_expr_t)vm, TRUE, count, modif);
1741 }
1742 #endif /* defined(DDB) */
1743
1744 #define vmem_printf printf
1745
1746 #if defined(DIAGNOSTIC)
1747
1748 static bool
1749 vmem_check_sanity(vmem_t *vm)
1750 {
1751 const bt_t *bt, *bt2;
1752
1753 MPASS(vm != NULL);
1754
1755 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1756 if (bt->bt_start > BT_END(bt)) {
1757 printf("corrupted tag\n");
1758 bt_dump(bt, vmem_printf);
1759 return false;
1760 }
1761 }
1762 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1763 if (bt->bt_type == BT_TYPE_CURSOR) {
1764 if (bt->bt_start != 0 || bt->bt_size != 0) {
1765 printf("corrupted cursor\n");
1766 return false;
1767 }
1768 continue;
1769 }
1770 TAILQ_FOREACH(bt2, &vm->vm_seglist, bt_seglist) {
1771 if (bt == bt2) {
1772 continue;
1773 }
1774 if (bt2->bt_type == BT_TYPE_CURSOR) {
1775 continue;
1776 }
1777 if (BT_ISSPAN_P(bt) != BT_ISSPAN_P(bt2)) {
1778 continue;
1779 }
1780 if (bt->bt_start <= BT_END(bt2) &&
1781 bt2->bt_start <= BT_END(bt)) {
1782 printf("overwrapped tags\n");
1783 bt_dump(bt, vmem_printf);
1784 bt_dump(bt2, vmem_printf);
1785 return false;
1786 }
1787 }
1788 }
1789
1790 return true;
1791 }
1792
1793 static void
1794 vmem_check(vmem_t *vm)
1795 {
1796
1797 if (!vmem_check_sanity(vm)) {
1798 panic("insanity vmem %p", vm);
1799 }
1800 }
1801
1802 #endif /* defined(DIAGNOSTIC) */
Cache object: 84e3f6887c303ca36cbce61b7629e00b
|