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 int
bt_fill(vmem_t * vm,int flags)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 /*
302 * Pop a tag off of the freetag stack.
303 */
304 static bt_t *
bt_alloc(vmem_t * vm)305 bt_alloc(vmem_t *vm)
306 {
307 bt_t *bt;
308
309 VMEM_ASSERT_LOCKED(vm);
310 bt = LIST_FIRST(&vm->vm_freetags);
311 MPASS(bt != NULL);
312 LIST_REMOVE(bt, bt_freelist);
313 vm->vm_nfreetags--;
314
315 return bt;
316 }
317
318 /*
319 * Trim the per-vmem free list. Returns with the lock released to
320 * avoid allocator recursions.
321 */
322 static void
bt_freetrim(vmem_t * vm,int freelimit)323 bt_freetrim(vmem_t *vm, int freelimit)
324 {
325 LIST_HEAD(, vmem_btag) freetags;
326 bt_t *bt;
327
328 LIST_INIT(&freetags);
329 VMEM_ASSERT_LOCKED(vm);
330 while (vm->vm_nfreetags > freelimit) {
331 bt = LIST_FIRST(&vm->vm_freetags);
332 LIST_REMOVE(bt, bt_freelist);
333 vm->vm_nfreetags--;
334 LIST_INSERT_HEAD(&freetags, bt, bt_freelist);
335 }
336 VMEM_UNLOCK(vm);
337 while ((bt = LIST_FIRST(&freetags)) != NULL) {
338 LIST_REMOVE(bt, bt_freelist);
339 uma_zfree(vmem_bt_zone, bt);
340 }
341 }
342
343 static inline void
bt_free(vmem_t * vm,bt_t * bt)344 bt_free(vmem_t *vm, bt_t *bt)
345 {
346
347 VMEM_ASSERT_LOCKED(vm);
348 MPASS(LIST_FIRST(&vm->vm_freetags) != bt);
349 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
350 vm->vm_nfreetags++;
351 }
352
353 /*
354 * freelist[0] ... [1, 1]
355 * freelist[1] ... [2, 2]
356 * :
357 * freelist[29] ... [30, 30]
358 * freelist[30] ... [31, 31]
359 * freelist[31] ... [32, 63]
360 * freelist[33] ... [64, 127]
361 * :
362 * freelist[n] ... [(1 << (n - 26)), (1 << (n - 25)) - 1]
363 * :
364 */
365
366 static struct vmem_freelist *
bt_freehead_tofree(vmem_t * vm,vmem_size_t size)367 bt_freehead_tofree(vmem_t *vm, vmem_size_t size)
368 {
369 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
370 const int idx = SIZE2ORDER(qsize);
371
372 MPASS(size != 0 && qsize != 0);
373 MPASS((size & vm->vm_quantum_mask) == 0);
374 MPASS(idx >= 0);
375 MPASS(idx < VMEM_MAXORDER);
376
377 return &vm->vm_freelist[idx];
378 }
379
380 /*
381 * bt_freehead_toalloc: return the freelist for the given size and allocation
382 * strategy.
383 *
384 * For M_FIRSTFIT, return the list in which any blocks are large enough
385 * for the requested size. otherwise, return the list which can have blocks
386 * large enough for the requested size.
387 */
388 static struct vmem_freelist *
bt_freehead_toalloc(vmem_t * vm,vmem_size_t size,int strat)389 bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, int strat)
390 {
391 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
392 int idx = SIZE2ORDER(qsize);
393
394 MPASS(size != 0 && qsize != 0);
395 MPASS((size & vm->vm_quantum_mask) == 0);
396
397 if (strat == M_FIRSTFIT && ORDER2SIZE(idx) != qsize) {
398 idx++;
399 /* check too large request? */
400 }
401 MPASS(idx >= 0);
402 MPASS(idx < VMEM_MAXORDER);
403
404 return &vm->vm_freelist[idx];
405 }
406
407 /* ---- boundary tag hash */
408
409 static struct vmem_hashlist *
bt_hashhead(vmem_t * vm,vmem_addr_t addr)410 bt_hashhead(vmem_t *vm, vmem_addr_t addr)
411 {
412 struct vmem_hashlist *list;
413 unsigned int hash;
414
415 hash = hash32_buf(&addr, sizeof(addr), 0);
416 list = &vm->vm_hashlist[hash % vm->vm_hashsize];
417
418 return list;
419 }
420
421 static bt_t *
bt_lookupbusy(vmem_t * vm,vmem_addr_t addr)422 bt_lookupbusy(vmem_t *vm, vmem_addr_t addr)
423 {
424 struct vmem_hashlist *list;
425 bt_t *bt;
426
427 VMEM_ASSERT_LOCKED(vm);
428 list = bt_hashhead(vm, addr);
429 LIST_FOREACH(bt, list, bt_hashlist) {
430 if (bt->bt_start == addr) {
431 break;
432 }
433 }
434
435 return bt;
436 }
437
438 static void
bt_rembusy(vmem_t * vm,bt_t * bt)439 bt_rembusy(vmem_t *vm, bt_t *bt)
440 {
441
442 VMEM_ASSERT_LOCKED(vm);
443 MPASS(vm->vm_nbusytag > 0);
444 vm->vm_inuse -= bt->bt_size;
445 vm->vm_nbusytag--;
446 LIST_REMOVE(bt, bt_hashlist);
447 }
448
449 static void
bt_insbusy(vmem_t * vm,bt_t * bt)450 bt_insbusy(vmem_t *vm, bt_t *bt)
451 {
452 struct vmem_hashlist *list;
453
454 VMEM_ASSERT_LOCKED(vm);
455 MPASS(bt->bt_type == BT_TYPE_BUSY);
456
457 list = bt_hashhead(vm, bt->bt_start);
458 LIST_INSERT_HEAD(list, bt, bt_hashlist);
459 vm->vm_nbusytag++;
460 vm->vm_inuse += bt->bt_size;
461 }
462
463 /* ---- boundary tag list */
464
465 static void
bt_remseg(vmem_t * vm,bt_t * bt)466 bt_remseg(vmem_t *vm, bt_t *bt)
467 {
468
469 TAILQ_REMOVE(&vm->vm_seglist, bt, bt_seglist);
470 bt_free(vm, bt);
471 }
472
473 static void
bt_insseg(vmem_t * vm,bt_t * bt,bt_t * prev)474 bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev)
475 {
476
477 TAILQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist);
478 }
479
480 static void
bt_insseg_tail(vmem_t * vm,bt_t * bt)481 bt_insseg_tail(vmem_t *vm, bt_t *bt)
482 {
483
484 TAILQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist);
485 }
486
487 static void
bt_remfree(vmem_t * vm,bt_t * bt)488 bt_remfree(vmem_t *vm, bt_t *bt)
489 {
490
491 MPASS(bt->bt_type == BT_TYPE_FREE);
492
493 LIST_REMOVE(bt, bt_freelist);
494 }
495
496 static void
bt_insfree(vmem_t * vm,bt_t * bt)497 bt_insfree(vmem_t *vm, bt_t *bt)
498 {
499 struct vmem_freelist *list;
500
501 list = bt_freehead_tofree(vm, bt->bt_size);
502 LIST_INSERT_HEAD(list, bt, bt_freelist);
503 }
504
505 /* ---- vmem internal functions */
506
507 /*
508 * Import from the arena into the quantum cache in UMA.
509 *
510 * We use VMEM_ADDR_QCACHE_MIN instead of 0: uma_zalloc() returns 0 to indicate
511 * failure, so UMA can't be used to cache a resource with value 0.
512 */
513 static int
qc_import(void * arg,void ** store,int cnt,int domain,int flags)514 qc_import(void *arg, void **store, int cnt, int domain, int flags)
515 {
516 qcache_t *qc;
517 vmem_addr_t addr;
518 int i;
519
520 KASSERT((flags & M_WAITOK) == 0, ("blocking allocation"));
521
522 qc = arg;
523 for (i = 0; i < cnt; i++) {
524 if (vmem_xalloc(qc->qc_vmem, qc->qc_size, 0, 0, 0,
525 VMEM_ADDR_QCACHE_MIN, VMEM_ADDR_MAX, flags, &addr) != 0)
526 break;
527 store[i] = (void *)addr;
528 }
529 return (i);
530 }
531
532 /*
533 * Release memory from the UMA cache to the arena.
534 */
535 static void
qc_release(void * arg,void ** store,int cnt)536 qc_release(void *arg, void **store, int cnt)
537 {
538 qcache_t *qc;
539 int i;
540
541 qc = arg;
542 for (i = 0; i < cnt; i++)
543 vmem_xfree(qc->qc_vmem, (vmem_addr_t)store[i], qc->qc_size);
544 }
545
546 static void
qc_init(vmem_t * vm,vmem_size_t qcache_max)547 qc_init(vmem_t *vm, vmem_size_t qcache_max)
548 {
549 qcache_t *qc;
550 vmem_size_t size;
551 int qcache_idx_max;
552 int i;
553
554 MPASS((qcache_max & vm->vm_quantum_mask) == 0);
555 qcache_idx_max = MIN(qcache_max >> vm->vm_quantum_shift,
556 VMEM_QCACHE_IDX_MAX);
557 vm->vm_qcache_max = qcache_idx_max << vm->vm_quantum_shift;
558 for (i = 0; i < qcache_idx_max; i++) {
559 qc = &vm->vm_qcache[i];
560 size = (i + 1) << vm->vm_quantum_shift;
561 snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu",
562 vm->vm_name, size);
563 qc->qc_vmem = vm;
564 qc->qc_size = size;
565 qc->qc_cache = uma_zcache_create(qc->qc_name, size,
566 NULL, NULL, NULL, NULL, qc_import, qc_release, qc,
567 UMA_ZONE_VM);
568 MPASS(qc->qc_cache);
569 }
570 }
571
572 static void
qc_destroy(vmem_t * vm)573 qc_destroy(vmem_t *vm)
574 {
575 int qcache_idx_max;
576 int i;
577
578 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
579 for (i = 0; i < qcache_idx_max; i++)
580 uma_zdestroy(vm->vm_qcache[i].qc_cache);
581 }
582
583 static void
qc_drain(vmem_t * vm)584 qc_drain(vmem_t *vm)
585 {
586 int qcache_idx_max;
587 int i;
588
589 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
590 for (i = 0; i < qcache_idx_max; i++)
591 zone_drain(vm->vm_qcache[i].qc_cache);
592 }
593
594 #ifndef UMA_MD_SMALL_ALLOC
595
596 static struct mtx_padalign __exclusive_cache_line vmem_bt_lock;
597
598 /*
599 * vmem_bt_alloc: Allocate a new page of boundary tags.
600 *
601 * On architectures with uma_small_alloc there is no recursion; no address
602 * space need be allocated to allocate boundary tags. For the others, we
603 * must handle recursion. Boundary tags are necessary to allocate new
604 * boundary tags.
605 *
606 * UMA guarantees that enough tags are held in reserve to allocate a new
607 * page of kva. We dip into this reserve by specifying M_USE_RESERVE only
608 * when allocating the page to hold new boundary tags. In this way the
609 * reserve is automatically filled by the allocation that uses the reserve.
610 *
611 * We still have to guarantee that the new tags are allocated atomically since
612 * many threads may try concurrently. The bt_lock provides this guarantee.
613 * We convert WAITOK allocations to NOWAIT and then handle the blocking here
614 * on failure. It's ok to return NULL for a WAITOK allocation as UMA will
615 * loop again after checking to see if we lost the race to allocate.
616 *
617 * There is a small race between vmem_bt_alloc() returning the page and the
618 * zone lock being acquired to add the page to the zone. For WAITOK
619 * allocations we just pause briefly. NOWAIT may experience a transient
620 * failure. To alleviate this we permit a small number of simultaneous
621 * fills to proceed concurrently so NOWAIT is less likely to fail unless
622 * we are really out of KVA.
623 */
624 static void *
vmem_bt_alloc(uma_zone_t zone,vm_size_t bytes,int domain,uint8_t * pflag,int wait)625 vmem_bt_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
626 int wait)
627 {
628 vmem_addr_t addr;
629
630 *pflag = UMA_SLAB_KERNEL;
631
632 /*
633 * Single thread boundary tag allocation so that the address space
634 * and memory are added in one atomic operation.
635 */
636 mtx_lock(&vmem_bt_lock);
637 if (vmem_xalloc(vm_dom[domain].vmd_kernel_arena, bytes, 0, 0, 0,
638 VMEM_ADDR_MIN, VMEM_ADDR_MAX,
639 M_NOWAIT | M_NOVM | M_USE_RESERVE | M_BESTFIT, &addr) == 0) {
640 if (kmem_back_domain(domain, kernel_object, addr, bytes,
641 M_NOWAIT | M_USE_RESERVE) == 0) {
642 mtx_unlock(&vmem_bt_lock);
643 return ((void *)addr);
644 }
645 vmem_xfree(vm_dom[domain].vmd_kernel_arena, addr, bytes);
646 mtx_unlock(&vmem_bt_lock);
647 /*
648 * Out of memory, not address space. This may not even be
649 * possible due to M_USE_RESERVE page allocation.
650 */
651 if (wait & M_WAITOK)
652 vm_wait_domain(domain);
653 return (NULL);
654 }
655 mtx_unlock(&vmem_bt_lock);
656 /*
657 * We're either out of address space or lost a fill race.
658 */
659 if (wait & M_WAITOK)
660 pause("btalloc", 1);
661
662 return (NULL);
663 }
664
665 /*
666 * How many pages do we need to startup_alloc.
667 */
668 int
vmem_startup_count(void)669 vmem_startup_count(void)
670 {
671
672 return (howmany(BT_MAXALLOC,
673 UMA_SLAB_SPACE / sizeof(struct vmem_btag)));
674 }
675 #endif
676
677 void
vmem_startup(void)678 vmem_startup(void)
679 {
680
681 mtx_init(&vmem_list_lock, "vmem list lock", NULL, MTX_DEF);
682 vmem_zone = uma_zcreate("vmem",
683 sizeof(struct vmem), NULL, NULL, NULL, NULL,
684 UMA_ALIGN_PTR, UMA_ZONE_VM);
685 vmem_bt_zone = uma_zcreate("vmem btag",
686 sizeof(struct vmem_btag), NULL, NULL, NULL, NULL,
687 UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE);
688 #ifndef UMA_MD_SMALL_ALLOC
689 mtx_init(&vmem_bt_lock, "btag lock", NULL, MTX_DEF);
690 uma_prealloc(vmem_bt_zone, BT_MAXALLOC);
691 /*
692 * Reserve enough tags to allocate new tags. We allow multiple
693 * CPUs to attempt to allocate new tags concurrently to limit
694 * false restarts in UMA. vmem_bt_alloc() allocates from a per-domain
695 * arena, which may involve importing a range from the kernel arena,
696 * so we need to keep at least 2 * BT_MAXALLOC tags reserved.
697 */
698 uma_zone_reserve(vmem_bt_zone, 2 * BT_MAXALLOC * mp_ncpus);
699 uma_zone_set_allocf(vmem_bt_zone, vmem_bt_alloc);
700 #endif
701 }
702
703 /* ---- rehash */
704
705 static int
vmem_rehash(vmem_t * vm,vmem_size_t newhashsize)706 vmem_rehash(vmem_t *vm, vmem_size_t newhashsize)
707 {
708 bt_t *bt;
709 int i;
710 struct vmem_hashlist *newhashlist;
711 struct vmem_hashlist *oldhashlist;
712 vmem_size_t oldhashsize;
713
714 MPASS(newhashsize > 0);
715
716 newhashlist = malloc(sizeof(struct vmem_hashlist) * newhashsize,
717 M_VMEM, M_NOWAIT);
718 if (newhashlist == NULL)
719 return ENOMEM;
720 for (i = 0; i < newhashsize; i++) {
721 LIST_INIT(&newhashlist[i]);
722 }
723
724 VMEM_LOCK(vm);
725 oldhashlist = vm->vm_hashlist;
726 oldhashsize = vm->vm_hashsize;
727 vm->vm_hashlist = newhashlist;
728 vm->vm_hashsize = newhashsize;
729 if (oldhashlist == NULL) {
730 VMEM_UNLOCK(vm);
731 return 0;
732 }
733 for (i = 0; i < oldhashsize; i++) {
734 while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) {
735 bt_rembusy(vm, bt);
736 bt_insbusy(vm, bt);
737 }
738 }
739 VMEM_UNLOCK(vm);
740
741 if (oldhashlist != vm->vm_hash0) {
742 free(oldhashlist, M_VMEM);
743 }
744
745 return 0;
746 }
747
748 static void
vmem_periodic_kick(void * dummy)749 vmem_periodic_kick(void *dummy)
750 {
751
752 taskqueue_enqueue(taskqueue_thread, &vmem_periodic_wk);
753 }
754
755 static void
vmem_periodic(void * unused,int pending)756 vmem_periodic(void *unused, int pending)
757 {
758 vmem_t *vm;
759 vmem_size_t desired;
760 vmem_size_t current;
761
762 mtx_lock(&vmem_list_lock);
763 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
764 #ifdef DIAGNOSTIC
765 /* Convenient time to verify vmem state. */
766 if (enable_vmem_check == 1) {
767 VMEM_LOCK(vm);
768 vmem_check(vm);
769 VMEM_UNLOCK(vm);
770 }
771 #endif
772 desired = 1 << flsl(vm->vm_nbusytag);
773 desired = MIN(MAX(desired, VMEM_HASHSIZE_MIN),
774 VMEM_HASHSIZE_MAX);
775 current = vm->vm_hashsize;
776
777 /* Grow in powers of two. Shrink less aggressively. */
778 if (desired >= current * 2 || desired * 4 <= current)
779 vmem_rehash(vm, desired);
780
781 /*
782 * Periodically wake up threads waiting for resources,
783 * so they could ask for reclamation again.
784 */
785 VMEM_CONDVAR_BROADCAST(vm);
786 }
787 mtx_unlock(&vmem_list_lock);
788
789 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
790 vmem_periodic_kick, NULL);
791 }
792
793 static void
vmem_start_callout(void * unused)794 vmem_start_callout(void *unused)
795 {
796
797 TASK_INIT(&vmem_periodic_wk, 0, vmem_periodic, NULL);
798 vmem_periodic_interval = hz * 10;
799 callout_init(&vmem_periodic_ch, 1);
800 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
801 vmem_periodic_kick, NULL);
802 }
803 SYSINIT(vfs, SI_SUB_CONFIGURE, SI_ORDER_ANY, vmem_start_callout, NULL);
804
805 static void
vmem_add1(vmem_t * vm,vmem_addr_t addr,vmem_size_t size,int type)806 vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int type)
807 {
808 bt_t *btspan;
809 bt_t *btfree;
810
811 MPASS(type == BT_TYPE_SPAN || type == BT_TYPE_SPAN_STATIC);
812 MPASS((size & vm->vm_quantum_mask) == 0);
813
814 btspan = bt_alloc(vm);
815 btspan->bt_type = type;
816 btspan->bt_start = addr;
817 btspan->bt_size = size;
818 bt_insseg_tail(vm, btspan);
819
820 btfree = bt_alloc(vm);
821 btfree->bt_type = BT_TYPE_FREE;
822 btfree->bt_start = addr;
823 btfree->bt_size = size;
824 bt_insseg(vm, btfree, btspan);
825 bt_insfree(vm, btfree);
826
827 vm->vm_size += size;
828 }
829
830 static void
vmem_destroy1(vmem_t * vm)831 vmem_destroy1(vmem_t *vm)
832 {
833 bt_t *bt;
834
835 /*
836 * Drain per-cpu quantum caches.
837 */
838 qc_destroy(vm);
839
840 /*
841 * The vmem should now only contain empty segments.
842 */
843 VMEM_LOCK(vm);
844 MPASS(vm->vm_nbusytag == 0);
845
846 while ((bt = TAILQ_FIRST(&vm->vm_seglist)) != NULL)
847 bt_remseg(vm, bt);
848
849 if (vm->vm_hashlist != NULL && vm->vm_hashlist != vm->vm_hash0)
850 free(vm->vm_hashlist, M_VMEM);
851
852 bt_freetrim(vm, 0);
853
854 VMEM_CONDVAR_DESTROY(vm);
855 VMEM_LOCK_DESTROY(vm);
856 uma_zfree(vmem_zone, vm);
857 }
858
859 static int
vmem_import(vmem_t * vm,vmem_size_t size,vmem_size_t align,int flags)860 vmem_import(vmem_t *vm, vmem_size_t size, vmem_size_t align, int flags)
861 {
862 vmem_addr_t addr;
863 int error;
864
865 if (vm->vm_importfn == NULL)
866 return (EINVAL);
867
868 /*
869 * To make sure we get a span that meets the alignment we double it
870 * and add the size to the tail. This slightly overestimates.
871 */
872 if (align != vm->vm_quantum_mask + 1)
873 size = (align * 2) + size;
874 size = roundup(size, vm->vm_import_quantum);
875
876 if (vm->vm_limit != 0 && vm->vm_limit < vm->vm_size + size)
877 return (ENOMEM);
878
879 /*
880 * Hide MAXALLOC tags so we're guaranteed to be able to add this
881 * span and the tag we want to allocate from it.
882 */
883 MPASS(vm->vm_nfreetags >= BT_MAXALLOC);
884 vm->vm_nfreetags -= BT_MAXALLOC;
885 VMEM_UNLOCK(vm);
886 error = (vm->vm_importfn)(vm->vm_arg, size, flags, &addr);
887 VMEM_LOCK(vm);
888 vm->vm_nfreetags += BT_MAXALLOC;
889 if (error)
890 return (ENOMEM);
891
892 vmem_add1(vm, addr, size, BT_TYPE_SPAN);
893
894 return 0;
895 }
896
897 /*
898 * vmem_fit: check if a bt can satisfy the given restrictions.
899 *
900 * it's a caller's responsibility to ensure the region is big enough
901 * before calling us.
902 */
903 static int
vmem_fit(const bt_t * bt,vmem_size_t size,vmem_size_t align,vmem_size_t phase,vmem_size_t nocross,vmem_addr_t minaddr,vmem_addr_t maxaddr,vmem_addr_t * addrp)904 vmem_fit(const bt_t *bt, vmem_size_t size, vmem_size_t align,
905 vmem_size_t phase, vmem_size_t nocross, vmem_addr_t minaddr,
906 vmem_addr_t maxaddr, vmem_addr_t *addrp)
907 {
908 vmem_addr_t start;
909 vmem_addr_t end;
910
911 MPASS(size > 0);
912 MPASS(bt->bt_size >= size); /* caller's responsibility */
913
914 /*
915 * XXX assumption: vmem_addr_t and vmem_size_t are
916 * unsigned integer of the same size.
917 */
918
919 start = bt->bt_start;
920 if (start < minaddr) {
921 start = minaddr;
922 }
923 end = BT_END(bt);
924 if (end > maxaddr)
925 end = maxaddr;
926 if (start > end)
927 return (ENOMEM);
928
929 start = VMEM_ALIGNUP(start - phase, align) + phase;
930 if (start < bt->bt_start)
931 start += align;
932 if (VMEM_CROSS_P(start, start + size - 1, nocross)) {
933 MPASS(align < nocross);
934 start = VMEM_ALIGNUP(start - phase, nocross) + phase;
935 }
936 if (start <= end && end - start >= size - 1) {
937 MPASS((start & (align - 1)) == phase);
938 MPASS(!VMEM_CROSS_P(start, start + size - 1, nocross));
939 MPASS(minaddr <= start);
940 MPASS(maxaddr == 0 || start + size - 1 <= maxaddr);
941 MPASS(bt->bt_start <= start);
942 MPASS(BT_END(bt) - start >= size - 1);
943 *addrp = start;
944
945 return (0);
946 }
947 return (ENOMEM);
948 }
949
950 /*
951 * vmem_clip: Trim the boundary tag edges to the requested start and size.
952 */
953 static void
vmem_clip(vmem_t * vm,bt_t * bt,vmem_addr_t start,vmem_size_t size)954 vmem_clip(vmem_t *vm, bt_t *bt, vmem_addr_t start, vmem_size_t size)
955 {
956 bt_t *btnew;
957 bt_t *btprev;
958
959 VMEM_ASSERT_LOCKED(vm);
960 MPASS(bt->bt_type == BT_TYPE_FREE);
961 MPASS(bt->bt_size >= size);
962 bt_remfree(vm, bt);
963 if (bt->bt_start != start) {
964 btprev = bt_alloc(vm);
965 btprev->bt_type = BT_TYPE_FREE;
966 btprev->bt_start = bt->bt_start;
967 btprev->bt_size = start - bt->bt_start;
968 bt->bt_start = start;
969 bt->bt_size -= btprev->bt_size;
970 bt_insfree(vm, btprev);
971 bt_insseg(vm, btprev,
972 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
973 }
974 MPASS(bt->bt_start == start);
975 if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) {
976 /* split */
977 btnew = bt_alloc(vm);
978 btnew->bt_type = BT_TYPE_BUSY;
979 btnew->bt_start = bt->bt_start;
980 btnew->bt_size = size;
981 bt->bt_start = bt->bt_start + size;
982 bt->bt_size -= size;
983 bt_insfree(vm, bt);
984 bt_insseg(vm, btnew,
985 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
986 bt_insbusy(vm, btnew);
987 bt = btnew;
988 } else {
989 bt->bt_type = BT_TYPE_BUSY;
990 bt_insbusy(vm, bt);
991 }
992 MPASS(bt->bt_size >= size);
993 }
994
995 static int
vmem_try_fetch(vmem_t * vm,const vmem_size_t size,vmem_size_t align,int flags)996 vmem_try_fetch(vmem_t *vm, const vmem_size_t size, vmem_size_t align, int flags)
997 {
998 vmem_size_t avail;
999
1000 VMEM_ASSERT_LOCKED(vm);
1001
1002 /*
1003 * XXX it is possible to fail to meet xalloc constraints with the
1004 * imported region. It is up to the user to specify the
1005 * import quantum such that it can satisfy any allocation.
1006 */
1007 if (vmem_import(vm, size, align, flags) == 0)
1008 return (1);
1009
1010 /*
1011 * Try to free some space from the quantum cache or reclaim
1012 * functions if available.
1013 */
1014 if (vm->vm_qcache_max != 0 || vm->vm_reclaimfn != NULL) {
1015 avail = vm->vm_size - vm->vm_inuse;
1016 VMEM_UNLOCK(vm);
1017 if (vm->vm_qcache_max != 0)
1018 qc_drain(vm);
1019 if (vm->vm_reclaimfn != NULL)
1020 vm->vm_reclaimfn(vm, flags);
1021 VMEM_LOCK(vm);
1022 /* If we were successful retry even NOWAIT. */
1023 if (vm->vm_size - vm->vm_inuse > avail)
1024 return (1);
1025 }
1026 if ((flags & M_NOWAIT) != 0)
1027 return (0);
1028 VMEM_CONDVAR_WAIT(vm);
1029 return (1);
1030 }
1031
1032 static int
vmem_try_release(vmem_t * vm,struct vmem_btag * bt,const bool remfree)1033 vmem_try_release(vmem_t *vm, struct vmem_btag *bt, const bool remfree)
1034 {
1035 struct vmem_btag *prev;
1036
1037 MPASS(bt->bt_type == BT_TYPE_FREE);
1038
1039 if (vm->vm_releasefn == NULL)
1040 return (0);
1041
1042 prev = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1043 MPASS(prev != NULL);
1044 MPASS(prev->bt_type != BT_TYPE_FREE);
1045
1046 if (prev->bt_type == BT_TYPE_SPAN && prev->bt_size == bt->bt_size) {
1047 vmem_addr_t spanaddr;
1048 vmem_size_t spansize;
1049
1050 MPASS(prev->bt_start == bt->bt_start);
1051 spanaddr = prev->bt_start;
1052 spansize = prev->bt_size;
1053 if (remfree)
1054 bt_remfree(vm, bt);
1055 bt_remseg(vm, bt);
1056 bt_remseg(vm, prev);
1057 vm->vm_size -= spansize;
1058 VMEM_CONDVAR_BROADCAST(vm);
1059 bt_freetrim(vm, BT_MAXFREE);
1060 vm->vm_releasefn(vm->vm_arg, spanaddr, spansize);
1061 return (1);
1062 }
1063 return (0);
1064 }
1065
1066 static int
vmem_xalloc_nextfit(vmem_t * vm,const vmem_size_t size,vmem_size_t align,const vmem_size_t phase,const vmem_size_t nocross,int flags,vmem_addr_t * addrp)1067 vmem_xalloc_nextfit(vmem_t *vm, const vmem_size_t size, vmem_size_t align,
1068 const vmem_size_t phase, const vmem_size_t nocross, int flags,
1069 vmem_addr_t *addrp)
1070 {
1071 struct vmem_btag *bt, *cursor, *next, *prev;
1072 int error;
1073
1074 error = ENOMEM;
1075 VMEM_LOCK(vm);
1076 retry:
1077 /*
1078 * Make sure we have enough tags to complete the operation.
1079 */
1080 if (vm->vm_nfreetags < BT_MAXALLOC && bt_fill(vm, flags) != 0)
1081 goto out;
1082
1083 /*
1084 * Find the next free tag meeting our constraints. If one is found,
1085 * perform the allocation.
1086 */
1087 for (cursor = &vm->vm_cursor, bt = TAILQ_NEXT(cursor, bt_seglist);
1088 bt != cursor; bt = TAILQ_NEXT(bt, bt_seglist)) {
1089 if (bt == NULL)
1090 bt = TAILQ_FIRST(&vm->vm_seglist);
1091 if (bt->bt_type == BT_TYPE_FREE && bt->bt_size >= size &&
1092 (error = vmem_fit(bt, size, align, phase, nocross,
1093 VMEM_ADDR_MIN, VMEM_ADDR_MAX, addrp)) == 0) {
1094 vmem_clip(vm, bt, *addrp, size);
1095 break;
1096 }
1097 }
1098
1099 /*
1100 * Try to coalesce free segments around the cursor. If we succeed, and
1101 * have not yet satisfied the allocation request, try again with the
1102 * newly coalesced segment.
1103 */
1104 if ((next = TAILQ_NEXT(cursor, bt_seglist)) != NULL &&
1105 (prev = TAILQ_PREV(cursor, vmem_seglist, bt_seglist)) != NULL &&
1106 next->bt_type == BT_TYPE_FREE && prev->bt_type == BT_TYPE_FREE &&
1107 prev->bt_start + prev->bt_size == next->bt_start) {
1108 prev->bt_size += next->bt_size;
1109 bt_remfree(vm, next);
1110 bt_remseg(vm, next);
1111
1112 /*
1113 * The coalesced segment might be able to satisfy our request.
1114 * If not, we might need to release it from the arena.
1115 */
1116 if (error == ENOMEM && prev->bt_size >= size &&
1117 (error = vmem_fit(prev, size, align, phase, nocross,
1118 VMEM_ADDR_MIN, VMEM_ADDR_MAX, addrp)) == 0) {
1119 vmem_clip(vm, prev, *addrp, size);
1120 bt = prev;
1121 } else
1122 (void)vmem_try_release(vm, prev, true);
1123 }
1124
1125 /*
1126 * If the allocation was successful, advance the cursor.
1127 */
1128 if (error == 0) {
1129 TAILQ_REMOVE(&vm->vm_seglist, cursor, bt_seglist);
1130 for (; bt != NULL && bt->bt_start < *addrp + size;
1131 bt = TAILQ_NEXT(bt, bt_seglist))
1132 ;
1133 if (bt != NULL)
1134 TAILQ_INSERT_BEFORE(bt, cursor, bt_seglist);
1135 else
1136 TAILQ_INSERT_HEAD(&vm->vm_seglist, cursor, bt_seglist);
1137 }
1138
1139 /*
1140 * Attempt to bring additional resources into the arena. If that fails
1141 * and M_WAITOK is specified, sleep waiting for resources to be freed.
1142 */
1143 if (error == ENOMEM && vmem_try_fetch(vm, size, align, flags))
1144 goto retry;
1145
1146 out:
1147 VMEM_UNLOCK(vm);
1148 return (error);
1149 }
1150
1151 /* ---- vmem API */
1152
1153 void
vmem_set_import(vmem_t * vm,vmem_import_t * importfn,vmem_release_t * releasefn,void * arg,vmem_size_t import_quantum)1154 vmem_set_import(vmem_t *vm, vmem_import_t *importfn,
1155 vmem_release_t *releasefn, void *arg, vmem_size_t import_quantum)
1156 {
1157
1158 VMEM_LOCK(vm);
1159 vm->vm_importfn = importfn;
1160 vm->vm_releasefn = releasefn;
1161 vm->vm_arg = arg;
1162 vm->vm_import_quantum = import_quantum;
1163 VMEM_UNLOCK(vm);
1164 }
1165
1166 void
vmem_set_limit(vmem_t * vm,vmem_size_t limit)1167 vmem_set_limit(vmem_t *vm, vmem_size_t limit)
1168 {
1169
1170 VMEM_LOCK(vm);
1171 vm->vm_limit = limit;
1172 VMEM_UNLOCK(vm);
1173 }
1174
1175 void
vmem_set_reclaim(vmem_t * vm,vmem_reclaim_t * reclaimfn)1176 vmem_set_reclaim(vmem_t *vm, vmem_reclaim_t *reclaimfn)
1177 {
1178
1179 VMEM_LOCK(vm);
1180 vm->vm_reclaimfn = reclaimfn;
1181 VMEM_UNLOCK(vm);
1182 }
1183
1184 /*
1185 * vmem_init: Initializes vmem arena.
1186 */
1187 vmem_t *
vmem_init(vmem_t * vm,const char * name,vmem_addr_t base,vmem_size_t size,vmem_size_t quantum,vmem_size_t qcache_max,int flags)1188 vmem_init(vmem_t *vm, const char *name, vmem_addr_t base, vmem_size_t size,
1189 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1190 {
1191 int i;
1192
1193 MPASS(quantum > 0);
1194 MPASS((quantum & (quantum - 1)) == 0);
1195
1196 bzero(vm, sizeof(*vm));
1197
1198 VMEM_CONDVAR_INIT(vm, name);
1199 VMEM_LOCK_INIT(vm, name);
1200 vm->vm_nfreetags = 0;
1201 LIST_INIT(&vm->vm_freetags);
1202 strlcpy(vm->vm_name, name, sizeof(vm->vm_name));
1203 vm->vm_quantum_mask = quantum - 1;
1204 vm->vm_quantum_shift = flsl(quantum) - 1;
1205 vm->vm_nbusytag = 0;
1206 vm->vm_size = 0;
1207 vm->vm_limit = 0;
1208 vm->vm_inuse = 0;
1209 qc_init(vm, qcache_max);
1210
1211 TAILQ_INIT(&vm->vm_seglist);
1212 vm->vm_cursor.bt_start = vm->vm_cursor.bt_size = 0;
1213 vm->vm_cursor.bt_type = BT_TYPE_CURSOR;
1214 TAILQ_INSERT_TAIL(&vm->vm_seglist, &vm->vm_cursor, bt_seglist);
1215
1216 for (i = 0; i < VMEM_MAXORDER; i++)
1217 LIST_INIT(&vm->vm_freelist[i]);
1218
1219 memset(&vm->vm_hash0, 0, sizeof(vm->vm_hash0));
1220 vm->vm_hashsize = VMEM_HASHSIZE_MIN;
1221 vm->vm_hashlist = vm->vm_hash0;
1222
1223 if (size != 0) {
1224 if (vmem_add(vm, base, size, flags) != 0) {
1225 vmem_destroy1(vm);
1226 return NULL;
1227 }
1228 }
1229
1230 mtx_lock(&vmem_list_lock);
1231 LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist);
1232 mtx_unlock(&vmem_list_lock);
1233
1234 return vm;
1235 }
1236
1237 /*
1238 * vmem_create: create an arena.
1239 */
1240 vmem_t *
vmem_create(const char * name,vmem_addr_t base,vmem_size_t size,vmem_size_t quantum,vmem_size_t qcache_max,int flags)1241 vmem_create(const char *name, vmem_addr_t base, vmem_size_t size,
1242 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1243 {
1244
1245 vmem_t *vm;
1246
1247 vm = uma_zalloc(vmem_zone, flags & (M_WAITOK|M_NOWAIT));
1248 if (vm == NULL)
1249 return (NULL);
1250 if (vmem_init(vm, name, base, size, quantum, qcache_max,
1251 flags) == NULL)
1252 return (NULL);
1253 return (vm);
1254 }
1255
1256 void
vmem_destroy(vmem_t * vm)1257 vmem_destroy(vmem_t *vm)
1258 {
1259
1260 mtx_lock(&vmem_list_lock);
1261 LIST_REMOVE(vm, vm_alllist);
1262 mtx_unlock(&vmem_list_lock);
1263
1264 vmem_destroy1(vm);
1265 }
1266
1267 vmem_size_t
vmem_roundup_size(vmem_t * vm,vmem_size_t size)1268 vmem_roundup_size(vmem_t *vm, vmem_size_t size)
1269 {
1270
1271 return (size + vm->vm_quantum_mask) & ~vm->vm_quantum_mask;
1272 }
1273
1274 /*
1275 * vmem_alloc: allocate resource from the arena.
1276 */
1277 int
vmem_alloc(vmem_t * vm,vmem_size_t size,int flags,vmem_addr_t * addrp)1278 vmem_alloc(vmem_t *vm, vmem_size_t size, int flags, vmem_addr_t *addrp)
1279 {
1280 const int strat __unused = flags & VMEM_FITMASK;
1281 qcache_t *qc;
1282
1283 flags &= VMEM_FLAGS;
1284 MPASS(size > 0);
1285 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT || strat == M_NEXTFIT);
1286 if ((flags & M_NOWAIT) == 0)
1287 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_alloc");
1288
1289 if (size <= vm->vm_qcache_max) {
1290 /*
1291 * Resource 0 cannot be cached, so avoid a blocking allocation
1292 * in qc_import() and give the vmem_xalloc() call below a chance
1293 * to return 0.
1294 */
1295 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1296 *addrp = (vmem_addr_t)uma_zalloc(qc->qc_cache,
1297 (flags & ~M_WAITOK) | M_NOWAIT);
1298 if (__predict_true(*addrp != 0))
1299 return (0);
1300 }
1301
1302 return (vmem_xalloc(vm, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
1303 flags, addrp));
1304 }
1305
1306 int
vmem_xalloc(vmem_t * vm,const vmem_size_t size0,vmem_size_t align,const vmem_size_t phase,const vmem_size_t nocross,const vmem_addr_t minaddr,const vmem_addr_t maxaddr,int flags,vmem_addr_t * addrp)1307 vmem_xalloc(vmem_t *vm, const vmem_size_t size0, vmem_size_t align,
1308 const vmem_size_t phase, const vmem_size_t nocross,
1309 const vmem_addr_t minaddr, const vmem_addr_t maxaddr, int flags,
1310 vmem_addr_t *addrp)
1311 {
1312 const vmem_size_t size = vmem_roundup_size(vm, size0);
1313 struct vmem_freelist *list;
1314 struct vmem_freelist *first;
1315 struct vmem_freelist *end;
1316 bt_t *bt;
1317 int error;
1318 int strat;
1319
1320 flags &= VMEM_FLAGS;
1321 strat = flags & VMEM_FITMASK;
1322 MPASS(size0 > 0);
1323 MPASS(size > 0);
1324 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT || strat == M_NEXTFIT);
1325 MPASS((flags & (M_NOWAIT|M_WAITOK)) != (M_NOWAIT|M_WAITOK));
1326 if ((flags & M_NOWAIT) == 0)
1327 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_xalloc");
1328 MPASS((align & vm->vm_quantum_mask) == 0);
1329 MPASS((align & (align - 1)) == 0);
1330 MPASS((phase & vm->vm_quantum_mask) == 0);
1331 MPASS((nocross & vm->vm_quantum_mask) == 0);
1332 MPASS((nocross & (nocross - 1)) == 0);
1333 MPASS((align == 0 && phase == 0) || phase < align);
1334 MPASS(nocross == 0 || nocross >= size);
1335 MPASS(minaddr <= maxaddr);
1336 MPASS(!VMEM_CROSS_P(phase, phase + size - 1, nocross));
1337 if (strat == M_NEXTFIT)
1338 MPASS(minaddr == VMEM_ADDR_MIN && maxaddr == VMEM_ADDR_MAX);
1339
1340 if (align == 0)
1341 align = vm->vm_quantum_mask + 1;
1342 *addrp = 0;
1343
1344 /*
1345 * Next-fit allocations don't use the freelists.
1346 */
1347 if (strat == M_NEXTFIT)
1348 return (vmem_xalloc_nextfit(vm, size0, align, phase, nocross,
1349 flags, addrp));
1350
1351 end = &vm->vm_freelist[VMEM_MAXORDER];
1352 /*
1353 * choose a free block from which we allocate.
1354 */
1355 first = bt_freehead_toalloc(vm, size, strat);
1356 VMEM_LOCK(vm);
1357 for (;;) {
1358 /*
1359 * Make sure we have enough tags to complete the
1360 * operation.
1361 */
1362 if (vm->vm_nfreetags < BT_MAXALLOC &&
1363 bt_fill(vm, flags) != 0) {
1364 error = ENOMEM;
1365 break;
1366 }
1367
1368 /*
1369 * Scan freelists looking for a tag that satisfies the
1370 * allocation. If we're doing BESTFIT we may encounter
1371 * sizes below the request. If we're doing FIRSTFIT we
1372 * inspect only the first element from each list.
1373 */
1374 for (list = first; list < end; list++) {
1375 LIST_FOREACH(bt, list, bt_freelist) {
1376 if (bt->bt_size >= size) {
1377 error = vmem_fit(bt, size, align, phase,
1378 nocross, minaddr, maxaddr, addrp);
1379 if (error == 0) {
1380 vmem_clip(vm, bt, *addrp, size);
1381 goto out;
1382 }
1383 }
1384 /* FIRST skips to the next list. */
1385 if (strat == M_FIRSTFIT)
1386 break;
1387 }
1388 }
1389
1390 /*
1391 * Retry if the fast algorithm failed.
1392 */
1393 if (strat == M_FIRSTFIT) {
1394 strat = M_BESTFIT;
1395 first = bt_freehead_toalloc(vm, size, strat);
1396 continue;
1397 }
1398
1399 /*
1400 * Try a few measures to bring additional resources into the
1401 * arena. If all else fails, we will sleep waiting for
1402 * resources to be freed.
1403 */
1404 if (!vmem_try_fetch(vm, size, align, flags)) {
1405 error = ENOMEM;
1406 break;
1407 }
1408 }
1409 out:
1410 VMEM_UNLOCK(vm);
1411 if (error != 0 && (flags & M_NOWAIT) == 0)
1412 panic("failed to allocate waiting allocation\n");
1413
1414 return (error);
1415 }
1416
1417 /*
1418 * vmem_free: free the resource to the arena.
1419 */
1420 void
vmem_free(vmem_t * vm,vmem_addr_t addr,vmem_size_t size)1421 vmem_free(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1422 {
1423 qcache_t *qc;
1424 MPASS(size > 0);
1425
1426 if (size <= vm->vm_qcache_max &&
1427 __predict_true(addr >= VMEM_ADDR_QCACHE_MIN)) {
1428 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1429 uma_zfree(qc->qc_cache, (void *)addr);
1430 } else
1431 vmem_xfree(vm, addr, size);
1432 }
1433
1434 void
vmem_xfree(vmem_t * vm,vmem_addr_t addr,vmem_size_t size)1435 vmem_xfree(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1436 {
1437 bt_t *bt;
1438 bt_t *t;
1439
1440 MPASS(size > 0);
1441
1442 VMEM_LOCK(vm);
1443 bt = bt_lookupbusy(vm, addr);
1444 MPASS(bt != NULL);
1445 MPASS(bt->bt_start == addr);
1446 MPASS(bt->bt_size == vmem_roundup_size(vm, size) ||
1447 bt->bt_size - vmem_roundup_size(vm, size) <= vm->vm_quantum_mask);
1448 MPASS(bt->bt_type == BT_TYPE_BUSY);
1449 bt_rembusy(vm, bt);
1450 bt->bt_type = BT_TYPE_FREE;
1451
1452 /* coalesce */
1453 t = TAILQ_NEXT(bt, bt_seglist);
1454 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1455 MPASS(BT_END(bt) < t->bt_start); /* YYY */
1456 bt->bt_size += t->bt_size;
1457 bt_remfree(vm, t);
1458 bt_remseg(vm, t);
1459 }
1460 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1461 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1462 MPASS(BT_END(t) < bt->bt_start); /* YYY */
1463 bt->bt_size += t->bt_size;
1464 bt->bt_start = t->bt_start;
1465 bt_remfree(vm, t);
1466 bt_remseg(vm, t);
1467 }
1468
1469 if (!vmem_try_release(vm, bt, false)) {
1470 bt_insfree(vm, bt);
1471 VMEM_CONDVAR_BROADCAST(vm);
1472 bt_freetrim(vm, BT_MAXFREE);
1473 }
1474 }
1475
1476 /*
1477 * vmem_add:
1478 *
1479 */
1480 int
vmem_add(vmem_t * vm,vmem_addr_t addr,vmem_size_t size,int flags)1481 vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int flags)
1482 {
1483 int error;
1484
1485 error = 0;
1486 flags &= VMEM_FLAGS;
1487 VMEM_LOCK(vm);
1488 if (vm->vm_nfreetags >= BT_MAXALLOC || bt_fill(vm, flags) == 0)
1489 vmem_add1(vm, addr, size, BT_TYPE_SPAN_STATIC);
1490 else
1491 error = ENOMEM;
1492 VMEM_UNLOCK(vm);
1493
1494 return (error);
1495 }
1496
1497 /*
1498 * vmem_size: information about arenas size
1499 */
1500 vmem_size_t
vmem_size(vmem_t * vm,int typemask)1501 vmem_size(vmem_t *vm, int typemask)
1502 {
1503 int i;
1504
1505 switch (typemask) {
1506 case VMEM_ALLOC:
1507 return vm->vm_inuse;
1508 case VMEM_FREE:
1509 return vm->vm_size - vm->vm_inuse;
1510 case VMEM_FREE|VMEM_ALLOC:
1511 return vm->vm_size;
1512 case VMEM_MAXFREE:
1513 VMEM_LOCK(vm);
1514 for (i = VMEM_MAXORDER - 1; i >= 0; i--) {
1515 if (LIST_EMPTY(&vm->vm_freelist[i]))
1516 continue;
1517 VMEM_UNLOCK(vm);
1518 return ((vmem_size_t)ORDER2SIZE(i) <<
1519 vm->vm_quantum_shift);
1520 }
1521 VMEM_UNLOCK(vm);
1522 return (0);
1523 default:
1524 panic("vmem_size");
1525 }
1526 }
1527
1528 /* ---- debug */
1529
1530 #if defined(DDB) || defined(DIAGNOSTIC)
1531
1532 static void bt_dump(const bt_t *, int (*)(const char *, ...)
1533 __printflike(1, 2));
1534
1535 static const char *
bt_type_string(int type)1536 bt_type_string(int type)
1537 {
1538
1539 switch (type) {
1540 case BT_TYPE_BUSY:
1541 return "busy";
1542 case BT_TYPE_FREE:
1543 return "free";
1544 case BT_TYPE_SPAN:
1545 return "span";
1546 case BT_TYPE_SPAN_STATIC:
1547 return "static span";
1548 case BT_TYPE_CURSOR:
1549 return "cursor";
1550 default:
1551 break;
1552 }
1553 return "BOGUS";
1554 }
1555
1556 static void
bt_dump(const bt_t * bt,int (* pr)(const char *,...))1557 bt_dump(const bt_t *bt, int (*pr)(const char *, ...))
1558 {
1559
1560 (*pr)("\t%p: %jx %jx, %d(%s)\n",
1561 bt, (intmax_t)bt->bt_start, (intmax_t)bt->bt_size,
1562 bt->bt_type, bt_type_string(bt->bt_type));
1563 }
1564
1565 static void
1566 vmem_dump(const vmem_t *vm , int (*pr)(const char *, ...) __printflike(1, 2))
1567 {
1568 const bt_t *bt;
1569 int i;
1570
1571 (*pr)("vmem %p '%s'\n", vm, vm->vm_name);
1572 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1573 bt_dump(bt, pr);
1574 }
1575
1576 for (i = 0; i < VMEM_MAXORDER; i++) {
1577 const struct vmem_freelist *fl = &vm->vm_freelist[i];
1578
1579 if (LIST_EMPTY(fl)) {
1580 continue;
1581 }
1582
1583 (*pr)("freelist[%d]\n", i);
LIST_FOREACH(bt,fl,bt_freelist)1584 LIST_FOREACH(bt, fl, bt_freelist) {
1585 bt_dump(bt, pr);
1586 }
1587 }
1588 }
1589
1590 #endif /* defined(DDB) || defined(DIAGNOSTIC) */
1591
1592 #if defined(DDB)
1593 #include <ddb/ddb.h>
1594
1595 static bt_t *
vmem_whatis_lookup(vmem_t * vm,vmem_addr_t addr)1596 vmem_whatis_lookup(vmem_t *vm, vmem_addr_t addr)
1597 {
1598 bt_t *bt;
1599
1600 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1601 if (BT_ISSPAN_P(bt)) {
1602 continue;
1603 }
1604 if (bt->bt_start <= addr && addr <= BT_END(bt)) {
1605 return bt;
1606 }
1607 }
1608
1609 return NULL;
1610 }
1611
1612 void
vmem_whatis(vmem_addr_t addr,int (* pr)(const char *,...))1613 vmem_whatis(vmem_addr_t addr, int (*pr)(const char *, ...))
1614 {
1615 vmem_t *vm;
1616
1617 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1618 bt_t *bt;
1619
1620 bt = vmem_whatis_lookup(vm, addr);
1621 if (bt == NULL) {
1622 continue;
1623 }
1624 (*pr)("%p is %p+%zu in VMEM '%s' (%s)\n",
1625 (void *)addr, (void *)bt->bt_start,
1626 (vmem_size_t)(addr - bt->bt_start), vm->vm_name,
1627 (bt->bt_type == BT_TYPE_BUSY) ? "allocated" : "free");
1628 }
1629 }
1630
1631 void
vmem_printall(const char * modif,int (* pr)(const char *,...))1632 vmem_printall(const char *modif, int (*pr)(const char *, ...))
1633 {
1634 const vmem_t *vm;
1635
1636 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1637 vmem_dump(vm, pr);
1638 }
1639 }
1640
1641 void
vmem_print(vmem_addr_t addr,const char * modif,int (* pr)(const char *,...))1642 vmem_print(vmem_addr_t addr, const char *modif, int (*pr)(const char *, ...))
1643 {
1644 const vmem_t *vm = (const void *)addr;
1645
1646 vmem_dump(vm, pr);
1647 }
1648
DB_SHOW_COMMAND(vmemdump,vmemdump)1649 DB_SHOW_COMMAND(vmemdump, vmemdump)
1650 {
1651
1652 if (!have_addr) {
1653 db_printf("usage: show vmemdump <addr>\n");
1654 return;
1655 }
1656
1657 vmem_dump((const vmem_t *)addr, db_printf);
1658 }
1659
DB_SHOW_ALL_COMMAND(vmemdump,vmemdumpall)1660 DB_SHOW_ALL_COMMAND(vmemdump, vmemdumpall)
1661 {
1662 const vmem_t *vm;
1663
1664 LIST_FOREACH(vm, &vmem_list, vm_alllist)
1665 vmem_dump(vm, db_printf);
1666 }
1667
DB_SHOW_COMMAND(vmem,vmem_summ)1668 DB_SHOW_COMMAND(vmem, vmem_summ)
1669 {
1670 const vmem_t *vm = (const void *)addr;
1671 const bt_t *bt;
1672 size_t ft[VMEM_MAXORDER], ut[VMEM_MAXORDER];
1673 size_t fs[VMEM_MAXORDER], us[VMEM_MAXORDER];
1674 int ord;
1675
1676 if (!have_addr) {
1677 db_printf("usage: show vmem <addr>\n");
1678 return;
1679 }
1680
1681 db_printf("vmem %p '%s'\n", vm, vm->vm_name);
1682 db_printf("\tquantum:\t%zu\n", vm->vm_quantum_mask + 1);
1683 db_printf("\tsize:\t%zu\n", vm->vm_size);
1684 db_printf("\tinuse:\t%zu\n", vm->vm_inuse);
1685 db_printf("\tfree:\t%zu\n", vm->vm_size - vm->vm_inuse);
1686 db_printf("\tbusy tags:\t%d\n", vm->vm_nbusytag);
1687 db_printf("\tfree tags:\t%d\n", vm->vm_nfreetags);
1688
1689 memset(&ft, 0, sizeof(ft));
1690 memset(&ut, 0, sizeof(ut));
1691 memset(&fs, 0, sizeof(fs));
1692 memset(&us, 0, sizeof(us));
1693 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1694 ord = SIZE2ORDER(bt->bt_size >> vm->vm_quantum_shift);
1695 if (bt->bt_type == BT_TYPE_BUSY) {
1696 ut[ord]++;
1697 us[ord] += bt->bt_size;
1698 } else if (bt->bt_type == BT_TYPE_FREE) {
1699 ft[ord]++;
1700 fs[ord] += bt->bt_size;
1701 }
1702 }
1703 db_printf("\t\t\tinuse\tsize\t\tfree\tsize\n");
1704 for (ord = 0; ord < VMEM_MAXORDER; ord++) {
1705 if (ut[ord] == 0 && ft[ord] == 0)
1706 continue;
1707 db_printf("\t%-15zu %zu\t%-15zu %zu\t%-16zu\n",
1708 ORDER2SIZE(ord) << vm->vm_quantum_shift,
1709 ut[ord], us[ord], ft[ord], fs[ord]);
1710 }
1711 }
1712
DB_SHOW_ALL_COMMAND(vmem,vmem_summall)1713 DB_SHOW_ALL_COMMAND(vmem, vmem_summall)
1714 {
1715 const vmem_t *vm;
1716
1717 LIST_FOREACH(vm, &vmem_list, vm_alllist)
1718 vmem_summ((db_expr_t)vm, TRUE, count, modif);
1719 }
1720 #endif /* defined(DDB) */
1721
1722 #define vmem_printf printf
1723
1724 #if defined(DIAGNOSTIC)
1725
1726 static bool
vmem_check_sanity(vmem_t * vm)1727 vmem_check_sanity(vmem_t *vm)
1728 {
1729 const bt_t *bt, *bt2;
1730
1731 MPASS(vm != NULL);
1732
1733 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1734 if (bt->bt_start > BT_END(bt)) {
1735 printf("corrupted tag\n");
1736 bt_dump(bt, vmem_printf);
1737 return false;
1738 }
1739 }
1740 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1741 if (bt->bt_type == BT_TYPE_CURSOR) {
1742 if (bt->bt_start != 0 || bt->bt_size != 0) {
1743 printf("corrupted cursor\n");
1744 return false;
1745 }
1746 continue;
1747 }
1748 TAILQ_FOREACH(bt2, &vm->vm_seglist, bt_seglist) {
1749 if (bt == bt2) {
1750 continue;
1751 }
1752 if (bt2->bt_type == BT_TYPE_CURSOR) {
1753 continue;
1754 }
1755 if (BT_ISSPAN_P(bt) != BT_ISSPAN_P(bt2)) {
1756 continue;
1757 }
1758 if (bt->bt_start <= BT_END(bt2) &&
1759 bt2->bt_start <= BT_END(bt)) {
1760 printf("overwrapped tags\n");
1761 bt_dump(bt, vmem_printf);
1762 bt_dump(bt2, vmem_printf);
1763 return false;
1764 }
1765 }
1766 }
1767
1768 return true;
1769 }
1770
1771 static void
vmem_check(vmem_t * vm)1772 vmem_check(vmem_t *vm)
1773 {
1774
1775 if (!vmem_check_sanity(vm)) {
1776 panic("insanity vmem %p", vm);
1777 }
1778 }
1779
1780 #endif /* defined(DIAGNOSTIC) */
1781