FreeBSD/Linux Kernel Cross Reference
sys/kern/subr_pool.c
1 /* $NetBSD: subr_pool.c,v 1.124 2006/11/01 10:17:58 yamt Exp $ */
2
3 /*-
4 * Copyright (c) 1997, 1999, 2000 The NetBSD Foundation, Inc.
5 * All rights reserved.
6 *
7 * This code is derived from software contributed to The NetBSD Foundation
8 * by Paul Kranenburg; by Jason R. Thorpe of the Numerical Aerospace
9 * Simulation Facility, NASA Ames Research Center.
10 *
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
14 * 1. Redistributions of source code must retain the above copyright
15 * notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 * notice, this list of conditions and the following disclaimer in the
18 * documentation and/or other materials provided with the distribution.
19 * 3. All advertising materials mentioning features or use of this software
20 * must display the following acknowledgement:
21 * This product includes software developed by the NetBSD
22 * Foundation, Inc. and its contributors.
23 * 4. Neither the name of The NetBSD Foundation nor the names of its
24 * contributors may be used to endorse or promote products derived
25 * from this software without specific prior written permission.
26 *
27 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
28 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
29 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
30 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
31 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
32 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
33 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
34 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
35 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
36 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
37 * POSSIBILITY OF SUCH DAMAGE.
38 */
39
40 #include <sys/cdefs.h>
41 __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.124 2006/11/01 10:17:58 yamt Exp $");
42
43 #include "opt_pool.h"
44 #include "opt_poollog.h"
45 #include "opt_lockdebug.h"
46
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/proc.h>
50 #include <sys/errno.h>
51 #include <sys/kernel.h>
52 #include <sys/malloc.h>
53 #include <sys/lock.h>
54 #include <sys/pool.h>
55 #include <sys/syslog.h>
56
57 #include <uvm/uvm.h>
58
59 /*
60 * Pool resource management utility.
61 *
62 * Memory is allocated in pages which are split into pieces according to
63 * the pool item size. Each page is kept on one of three lists in the
64 * pool structure: `pr_emptypages', `pr_fullpages' and `pr_partpages',
65 * for empty, full and partially-full pages respectively. The individual
66 * pool items are on a linked list headed by `ph_itemlist' in each page
67 * header. The memory for building the page list is either taken from
68 * the allocated pages themselves (for small pool items) or taken from
69 * an internal pool of page headers (`phpool').
70 */
71
72 /* List of all pools */
73 LIST_HEAD(,pool) pool_head = LIST_HEAD_INITIALIZER(pool_head);
74
75 /* Private pool for page header structures */
76 #define PHPOOL_MAX 8
77 static struct pool phpool[PHPOOL_MAX];
78 #define PHPOOL_FREELIST_NELEM(idx) (((idx) == 0) ? 0 : (1 << (idx)))
79
80 #ifdef POOL_SUBPAGE
81 /* Pool of subpages for use by normal pools. */
82 static struct pool psppool;
83 #endif
84
85 static SLIST_HEAD(, pool_allocator) pa_deferinitq =
86 SLIST_HEAD_INITIALIZER(pa_deferinitq);
87
88 static void *pool_page_alloc_meta(struct pool *, int);
89 static void pool_page_free_meta(struct pool *, void *);
90
91 /* allocator for pool metadata */
92 static struct pool_allocator pool_allocator_meta = {
93 pool_page_alloc_meta, pool_page_free_meta,
94 .pa_backingmapptr = &kmem_map,
95 };
96
97 /* # of seconds to retain page after last use */
98 int pool_inactive_time = 10;
99
100 /* Next candidate for drainage (see pool_drain()) */
101 static struct pool *drainpp;
102
103 /* This spin lock protects both pool_head and drainpp. */
104 struct simplelock pool_head_slock = SIMPLELOCK_INITIALIZER;
105
106 typedef uint8_t pool_item_freelist_t;
107
108 struct pool_item_header {
109 /* Page headers */
110 LIST_ENTRY(pool_item_header)
111 ph_pagelist; /* pool page list */
112 SPLAY_ENTRY(pool_item_header)
113 ph_node; /* Off-page page headers */
114 caddr_t ph_page; /* this page's address */
115 struct timeval ph_time; /* last referenced */
116 union {
117 /* !PR_NOTOUCH */
118 struct {
119 LIST_HEAD(, pool_item)
120 phu_itemlist; /* chunk list for this page */
121 } phu_normal;
122 /* PR_NOTOUCH */
123 struct {
124 uint16_t
125 phu_off; /* start offset in page */
126 pool_item_freelist_t
127 phu_firstfree; /* first free item */
128 /*
129 * XXX it might be better to use
130 * a simple bitmap and ffs(3)
131 */
132 } phu_notouch;
133 } ph_u;
134 uint16_t ph_nmissing; /* # of chunks in use */
135 };
136 #define ph_itemlist ph_u.phu_normal.phu_itemlist
137 #define ph_off ph_u.phu_notouch.phu_off
138 #define ph_firstfree ph_u.phu_notouch.phu_firstfree
139
140 struct pool_item {
141 #ifdef DIAGNOSTIC
142 u_int pi_magic;
143 #endif
144 #define PI_MAGIC 0xdeadbeefU
145 /* Other entries use only this list entry */
146 LIST_ENTRY(pool_item) pi_list;
147 };
148
149 #define POOL_NEEDS_CATCHUP(pp) \
150 ((pp)->pr_nitems < (pp)->pr_minitems)
151
152 /*
153 * Pool cache management.
154 *
155 * Pool caches provide a way for constructed objects to be cached by the
156 * pool subsystem. This can lead to performance improvements by avoiding
157 * needless object construction/destruction; it is deferred until absolutely
158 * necessary.
159 *
160 * Caches are grouped into cache groups. Each cache group references
161 * up to 16 constructed objects. When a cache allocates an object
162 * from the pool, it calls the object's constructor and places it into
163 * a cache group. When a cache group frees an object back to the pool,
164 * it first calls the object's destructor. This allows the object to
165 * persist in constructed form while freed to the cache.
166 *
167 * Multiple caches may exist for each pool. This allows a single
168 * object type to have multiple constructed forms. The pool references
169 * each cache, so that when a pool is drained by the pagedaemon, it can
170 * drain each individual cache as well. Each time a cache is drained,
171 * the most idle cache group is freed to the pool in its entirety.
172 *
173 * Pool caches are layed on top of pools. By layering them, we can avoid
174 * the complexity of cache management for pools which would not benefit
175 * from it.
176 */
177
178 /* The cache group pool. */
179 static struct pool pcgpool;
180
181 static void pool_cache_reclaim(struct pool_cache *, struct pool_pagelist *,
182 struct pool_cache_grouplist *);
183 static void pcg_grouplist_free(struct pool_cache_grouplist *);
184
185 static int pool_catchup(struct pool *);
186 static void pool_prime_page(struct pool *, caddr_t,
187 struct pool_item_header *);
188 static void pool_update_curpage(struct pool *);
189
190 static int pool_grow(struct pool *, int);
191 static void *pool_allocator_alloc(struct pool *, int);
192 static void pool_allocator_free(struct pool *, void *);
193
194 static void pool_print_pagelist(struct pool *, struct pool_pagelist *,
195 void (*)(const char *, ...));
196 static void pool_print1(struct pool *, const char *,
197 void (*)(const char *, ...));
198
199 static int pool_chk_page(struct pool *, const char *,
200 struct pool_item_header *);
201
202 /*
203 * Pool log entry. An array of these is allocated in pool_init().
204 */
205 struct pool_log {
206 const char *pl_file;
207 long pl_line;
208 int pl_action;
209 #define PRLOG_GET 1
210 #define PRLOG_PUT 2
211 void *pl_addr;
212 };
213
214 #ifdef POOL_DIAGNOSTIC
215 /* Number of entries in pool log buffers */
216 #ifndef POOL_LOGSIZE
217 #define POOL_LOGSIZE 10
218 #endif
219
220 int pool_logsize = POOL_LOGSIZE;
221
222 static inline void
223 pr_log(struct pool *pp, void *v, int action, const char *file, long line)
224 {
225 int n = pp->pr_curlogentry;
226 struct pool_log *pl;
227
228 if ((pp->pr_roflags & PR_LOGGING) == 0)
229 return;
230
231 /*
232 * Fill in the current entry. Wrap around and overwrite
233 * the oldest entry if necessary.
234 */
235 pl = &pp->pr_log[n];
236 pl->pl_file = file;
237 pl->pl_line = line;
238 pl->pl_action = action;
239 pl->pl_addr = v;
240 if (++n >= pp->pr_logsize)
241 n = 0;
242 pp->pr_curlogentry = n;
243 }
244
245 static void
246 pr_printlog(struct pool *pp, struct pool_item *pi,
247 void (*pr)(const char *, ...))
248 {
249 int i = pp->pr_logsize;
250 int n = pp->pr_curlogentry;
251
252 if ((pp->pr_roflags & PR_LOGGING) == 0)
253 return;
254
255 /*
256 * Print all entries in this pool's log.
257 */
258 while (i-- > 0) {
259 struct pool_log *pl = &pp->pr_log[n];
260 if (pl->pl_action != 0) {
261 if (pi == NULL || pi == pl->pl_addr) {
262 (*pr)("\tlog entry %d:\n", i);
263 (*pr)("\t\taction = %s, addr = %p\n",
264 pl->pl_action == PRLOG_GET ? "get" : "put",
265 pl->pl_addr);
266 (*pr)("\t\tfile: %s at line %lu\n",
267 pl->pl_file, pl->pl_line);
268 }
269 }
270 if (++n >= pp->pr_logsize)
271 n = 0;
272 }
273 }
274
275 static inline void
276 pr_enter(struct pool *pp, const char *file, long line)
277 {
278
279 if (__predict_false(pp->pr_entered_file != NULL)) {
280 printf("pool %s: reentrancy at file %s line %ld\n",
281 pp->pr_wchan, file, line);
282 printf(" previous entry at file %s line %ld\n",
283 pp->pr_entered_file, pp->pr_entered_line);
284 panic("pr_enter");
285 }
286
287 pp->pr_entered_file = file;
288 pp->pr_entered_line = line;
289 }
290
291 static inline void
292 pr_leave(struct pool *pp)
293 {
294
295 if (__predict_false(pp->pr_entered_file == NULL)) {
296 printf("pool %s not entered?\n", pp->pr_wchan);
297 panic("pr_leave");
298 }
299
300 pp->pr_entered_file = NULL;
301 pp->pr_entered_line = 0;
302 }
303
304 static inline void
305 pr_enter_check(struct pool *pp, void (*pr)(const char *, ...))
306 {
307
308 if (pp->pr_entered_file != NULL)
309 (*pr)("\n\tcurrently entered from file %s line %ld\n",
310 pp->pr_entered_file, pp->pr_entered_line);
311 }
312 #else
313 #define pr_log(pp, v, action, file, line)
314 #define pr_printlog(pp, pi, pr)
315 #define pr_enter(pp, file, line)
316 #define pr_leave(pp)
317 #define pr_enter_check(pp, pr)
318 #endif /* POOL_DIAGNOSTIC */
319
320 static inline int
321 pr_item_notouch_index(const struct pool *pp, const struct pool_item_header *ph,
322 const void *v)
323 {
324 const char *cp = v;
325 int idx;
326
327 KASSERT(pp->pr_roflags & PR_NOTOUCH);
328 idx = (cp - ph->ph_page - ph->ph_off) / pp->pr_size;
329 KASSERT(idx < pp->pr_itemsperpage);
330 return idx;
331 }
332
333 #define PR_FREELIST_ALIGN(p) \
334 roundup((uintptr_t)(p), sizeof(pool_item_freelist_t))
335 #define PR_FREELIST(ph) ((pool_item_freelist_t *)PR_FREELIST_ALIGN((ph) + 1))
336 #define PR_INDEX_USED ((pool_item_freelist_t)-1)
337 #define PR_INDEX_EOL ((pool_item_freelist_t)-2)
338
339 static inline void
340 pr_item_notouch_put(const struct pool *pp, struct pool_item_header *ph,
341 void *obj)
342 {
343 int idx = pr_item_notouch_index(pp, ph, obj);
344 pool_item_freelist_t *freelist = PR_FREELIST(ph);
345
346 KASSERT(freelist[idx] == PR_INDEX_USED);
347 freelist[idx] = ph->ph_firstfree;
348 ph->ph_firstfree = idx;
349 }
350
351 static inline void *
352 pr_item_notouch_get(const struct pool *pp, struct pool_item_header *ph)
353 {
354 int idx = ph->ph_firstfree;
355 pool_item_freelist_t *freelist = PR_FREELIST(ph);
356
357 KASSERT(freelist[idx] != PR_INDEX_USED);
358 ph->ph_firstfree = freelist[idx];
359 freelist[idx] = PR_INDEX_USED;
360
361 return ph->ph_page + ph->ph_off + idx * pp->pr_size;
362 }
363
364 static inline int
365 phtree_compare(struct pool_item_header *a, struct pool_item_header *b)
366 {
367
368 /*
369 * we consider pool_item_header with smaller ph_page bigger.
370 * (this unnatural ordering is for the benefit of pr_find_pagehead.)
371 */
372
373 if (a->ph_page < b->ph_page)
374 return (1);
375 else if (a->ph_page > b->ph_page)
376 return (-1);
377 else
378 return (0);
379 }
380
381 SPLAY_PROTOTYPE(phtree, pool_item_header, ph_node, phtree_compare);
382 SPLAY_GENERATE(phtree, pool_item_header, ph_node, phtree_compare);
383
384 /*
385 * Return the pool page header based on item address.
386 */
387 static inline struct pool_item_header *
388 pr_find_pagehead(struct pool *pp, void *v)
389 {
390 struct pool_item_header *ph, tmp;
391
392 if ((pp->pr_roflags & PR_NOALIGN) != 0) {
393 tmp.ph_page = (caddr_t)(uintptr_t)v;
394 ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
395 if (ph == NULL) {
396 ph = SPLAY_ROOT(&pp->pr_phtree);
397 if (ph != NULL && phtree_compare(&tmp, ph) >= 0) {
398 ph = SPLAY_NEXT(phtree, &pp->pr_phtree, ph);
399 }
400 KASSERT(ph == NULL || phtree_compare(&tmp, ph) < 0);
401 }
402 } else {
403 caddr_t page =
404 (caddr_t)((uintptr_t)v & pp->pr_alloc->pa_pagemask);
405
406 if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
407 ph = (void *)(page + pp->pr_phoffset);
408 } else {
409 tmp.ph_page = page;
410 ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
411 }
412 }
413
414 KASSERT(ph == NULL || ((pp->pr_roflags & PR_PHINPAGE) != 0) ||
415 (ph->ph_page <= (char *)v &&
416 (char *)v < ph->ph_page + pp->pr_alloc->pa_pagesz));
417 return ph;
418 }
419
420 static void
421 pr_pagelist_free(struct pool *pp, struct pool_pagelist *pq)
422 {
423 struct pool_item_header *ph;
424 int s;
425
426 while ((ph = LIST_FIRST(pq)) != NULL) {
427 LIST_REMOVE(ph, ph_pagelist);
428 pool_allocator_free(pp, ph->ph_page);
429 if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
430 s = splvm();
431 pool_put(pp->pr_phpool, ph);
432 splx(s);
433 }
434 }
435 }
436
437 /*
438 * Remove a page from the pool.
439 */
440 static inline void
441 pr_rmpage(struct pool *pp, struct pool_item_header *ph,
442 struct pool_pagelist *pq)
443 {
444
445 LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
446
447 /*
448 * If the page was idle, decrement the idle page count.
449 */
450 if (ph->ph_nmissing == 0) {
451 #ifdef DIAGNOSTIC
452 if (pp->pr_nidle == 0)
453 panic("pr_rmpage: nidle inconsistent");
454 if (pp->pr_nitems < pp->pr_itemsperpage)
455 panic("pr_rmpage: nitems inconsistent");
456 #endif
457 pp->pr_nidle--;
458 }
459
460 pp->pr_nitems -= pp->pr_itemsperpage;
461
462 /*
463 * Unlink the page from the pool and queue it for release.
464 */
465 LIST_REMOVE(ph, ph_pagelist);
466 if ((pp->pr_roflags & PR_PHINPAGE) == 0)
467 SPLAY_REMOVE(phtree, &pp->pr_phtree, ph);
468 LIST_INSERT_HEAD(pq, ph, ph_pagelist);
469
470 pp->pr_npages--;
471 pp->pr_npagefree++;
472
473 pool_update_curpage(pp);
474 }
475
476 static boolean_t
477 pa_starved_p(struct pool_allocator *pa)
478 {
479
480 if (pa->pa_backingmap != NULL) {
481 return vm_map_starved_p(pa->pa_backingmap);
482 }
483 return FALSE;
484 }
485
486 static int
487 pool_reclaim_callback(struct callback_entry *ce, void *obj, void *arg)
488 {
489 struct pool *pp = obj;
490 struct pool_allocator *pa = pp->pr_alloc;
491
492 KASSERT(&pp->pr_reclaimerentry == ce);
493 pool_reclaim(pp);
494 if (!pa_starved_p(pa)) {
495 return CALLBACK_CHAIN_ABORT;
496 }
497 return CALLBACK_CHAIN_CONTINUE;
498 }
499
500 static void
501 pool_reclaim_register(struct pool *pp)
502 {
503 struct vm_map *map = pp->pr_alloc->pa_backingmap;
504 int s;
505
506 if (map == NULL) {
507 return;
508 }
509
510 s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */
511 callback_register(&vm_map_to_kernel(map)->vmk_reclaim_callback,
512 &pp->pr_reclaimerentry, pp, pool_reclaim_callback);
513 splx(s);
514 }
515
516 static void
517 pool_reclaim_unregister(struct pool *pp)
518 {
519 struct vm_map *map = pp->pr_alloc->pa_backingmap;
520 int s;
521
522 if (map == NULL) {
523 return;
524 }
525
526 s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */
527 callback_unregister(&vm_map_to_kernel(map)->vmk_reclaim_callback,
528 &pp->pr_reclaimerentry);
529 splx(s);
530 }
531
532 static void
533 pa_reclaim_register(struct pool_allocator *pa)
534 {
535 struct vm_map *map = *pa->pa_backingmapptr;
536 struct pool *pp;
537
538 KASSERT(pa->pa_backingmap == NULL);
539 if (map == NULL) {
540 SLIST_INSERT_HEAD(&pa_deferinitq, pa, pa_q);
541 return;
542 }
543 pa->pa_backingmap = map;
544 TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
545 pool_reclaim_register(pp);
546 }
547 }
548
549 /*
550 * Initialize all the pools listed in the "pools" link set.
551 */
552 void
553 pool_subsystem_init(void)
554 {
555 struct pool_allocator *pa;
556 __link_set_decl(pools, struct link_pool_init);
557 struct link_pool_init * const *pi;
558
559 __link_set_foreach(pi, pools)
560 pool_init((*pi)->pp, (*pi)->size, (*pi)->align,
561 (*pi)->align_offset, (*pi)->flags, (*pi)->wchan,
562 (*pi)->palloc);
563
564 while ((pa = SLIST_FIRST(&pa_deferinitq)) != NULL) {
565 KASSERT(pa->pa_backingmapptr != NULL);
566 KASSERT(*pa->pa_backingmapptr != NULL);
567 SLIST_REMOVE_HEAD(&pa_deferinitq, pa_q);
568 pa_reclaim_register(pa);
569 }
570 }
571
572 /*
573 * Initialize the given pool resource structure.
574 *
575 * We export this routine to allow other kernel parts to declare
576 * static pools that must be initialized before malloc() is available.
577 */
578 void
579 pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
580 const char *wchan, struct pool_allocator *palloc)
581 {
582 #ifdef DEBUG
583 struct pool *pp1;
584 #endif
585 size_t trysize, phsize;
586 int off, slack, s;
587
588 KASSERT((1UL << (CHAR_BIT * sizeof(pool_item_freelist_t))) - 2 >=
589 PHPOOL_FREELIST_NELEM(PHPOOL_MAX - 1));
590
591 #ifdef DEBUG
592 /*
593 * Check that the pool hasn't already been initialised and
594 * added to the list of all pools.
595 */
596 LIST_FOREACH(pp1, &pool_head, pr_poollist) {
597 if (pp == pp1)
598 panic("pool_init: pool %s already initialised",
599 wchan);
600 }
601 #endif
602
603 #ifdef POOL_DIAGNOSTIC
604 /*
605 * Always log if POOL_DIAGNOSTIC is defined.
606 */
607 if (pool_logsize != 0)
608 flags |= PR_LOGGING;
609 #endif
610
611 if (palloc == NULL)
612 palloc = &pool_allocator_kmem;
613 #ifdef POOL_SUBPAGE
614 if (size > palloc->pa_pagesz) {
615 if (palloc == &pool_allocator_kmem)
616 palloc = &pool_allocator_kmem_fullpage;
617 else if (palloc == &pool_allocator_nointr)
618 palloc = &pool_allocator_nointr_fullpage;
619 }
620 #endif /* POOL_SUBPAGE */
621 if ((palloc->pa_flags & PA_INITIALIZED) == 0) {
622 if (palloc->pa_pagesz == 0)
623 palloc->pa_pagesz = PAGE_SIZE;
624
625 TAILQ_INIT(&palloc->pa_list);
626
627 simple_lock_init(&palloc->pa_slock);
628 palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
629 palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
630
631 if (palloc->pa_backingmapptr != NULL) {
632 pa_reclaim_register(palloc);
633 }
634 palloc->pa_flags |= PA_INITIALIZED;
635 }
636
637 if (align == 0)
638 align = ALIGN(1);
639
640 if ((flags & PR_NOTOUCH) == 0 && size < sizeof(struct pool_item))
641 size = sizeof(struct pool_item);
642
643 size = roundup(size, align);
644 #ifdef DIAGNOSTIC
645 if (size > palloc->pa_pagesz)
646 panic("pool_init: pool item size (%zu) too large", size);
647 #endif
648
649 /*
650 * Initialize the pool structure.
651 */
652 LIST_INIT(&pp->pr_emptypages);
653 LIST_INIT(&pp->pr_fullpages);
654 LIST_INIT(&pp->pr_partpages);
655 LIST_INIT(&pp->pr_cachelist);
656 pp->pr_curpage = NULL;
657 pp->pr_npages = 0;
658 pp->pr_minitems = 0;
659 pp->pr_minpages = 0;
660 pp->pr_maxpages = UINT_MAX;
661 pp->pr_roflags = flags;
662 pp->pr_flags = 0;
663 pp->pr_size = size;
664 pp->pr_align = align;
665 pp->pr_wchan = wchan;
666 pp->pr_alloc = palloc;
667 pp->pr_nitems = 0;
668 pp->pr_nout = 0;
669 pp->pr_hardlimit = UINT_MAX;
670 pp->pr_hardlimit_warning = NULL;
671 pp->pr_hardlimit_ratecap.tv_sec = 0;
672 pp->pr_hardlimit_ratecap.tv_usec = 0;
673 pp->pr_hardlimit_warning_last.tv_sec = 0;
674 pp->pr_hardlimit_warning_last.tv_usec = 0;
675 pp->pr_drain_hook = NULL;
676 pp->pr_drain_hook_arg = NULL;
677
678 /*
679 * Decide whether to put the page header off page to avoid
680 * wasting too large a part of the page or too big item.
681 * Off-page page headers go on a hash table, so we can match
682 * a returned item with its header based on the page address.
683 * We use 1/16 of the page size and about 8 times of the item
684 * size as the threshold (XXX: tune)
685 *
686 * However, we'll put the header into the page if we can put
687 * it without wasting any items.
688 *
689 * Silently enforce `0 <= ioff < align'.
690 */
691 pp->pr_itemoffset = ioff %= align;
692 /* See the comment below about reserved bytes. */
693 trysize = palloc->pa_pagesz - ((align - ioff) % align);
694 phsize = ALIGN(sizeof(struct pool_item_header));
695 if ((pp->pr_roflags & (PR_NOTOUCH | PR_NOALIGN)) == 0 &&
696 (pp->pr_size < MIN(palloc->pa_pagesz / 16, phsize << 3) ||
697 trysize / pp->pr_size == (trysize - phsize) / pp->pr_size)) {
698 /* Use the end of the page for the page header */
699 pp->pr_roflags |= PR_PHINPAGE;
700 pp->pr_phoffset = off = palloc->pa_pagesz - phsize;
701 } else {
702 /* The page header will be taken from our page header pool */
703 pp->pr_phoffset = 0;
704 off = palloc->pa_pagesz;
705 SPLAY_INIT(&pp->pr_phtree);
706 }
707
708 /*
709 * Alignment is to take place at `ioff' within the item. This means
710 * we must reserve up to `align - 1' bytes on the page to allow
711 * appropriate positioning of each item.
712 */
713 pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
714 KASSERT(pp->pr_itemsperpage != 0);
715 if ((pp->pr_roflags & PR_NOTOUCH)) {
716 int idx;
717
718 for (idx = 0; pp->pr_itemsperpage > PHPOOL_FREELIST_NELEM(idx);
719 idx++) {
720 /* nothing */
721 }
722 if (idx >= PHPOOL_MAX) {
723 /*
724 * if you see this panic, consider to tweak
725 * PHPOOL_MAX and PHPOOL_FREELIST_NELEM.
726 */
727 panic("%s: too large itemsperpage(%d) for PR_NOTOUCH",
728 pp->pr_wchan, pp->pr_itemsperpage);
729 }
730 pp->pr_phpool = &phpool[idx];
731 } else if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
732 pp->pr_phpool = &phpool[0];
733 }
734 #if defined(DIAGNOSTIC)
735 else {
736 pp->pr_phpool = NULL;
737 }
738 #endif
739
740 /*
741 * Use the slack between the chunks and the page header
742 * for "cache coloring".
743 */
744 slack = off - pp->pr_itemsperpage * pp->pr_size;
745 pp->pr_maxcolor = (slack / align) * align;
746 pp->pr_curcolor = 0;
747
748 pp->pr_nget = 0;
749 pp->pr_nfail = 0;
750 pp->pr_nput = 0;
751 pp->pr_npagealloc = 0;
752 pp->pr_npagefree = 0;
753 pp->pr_hiwat = 0;
754 pp->pr_nidle = 0;
755
756 #ifdef POOL_DIAGNOSTIC
757 if (flags & PR_LOGGING) {
758 if (kmem_map == NULL ||
759 (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log),
760 M_TEMP, M_NOWAIT)) == NULL)
761 pp->pr_roflags &= ~PR_LOGGING;
762 pp->pr_curlogentry = 0;
763 pp->pr_logsize = pool_logsize;
764 }
765 #endif
766
767 pp->pr_entered_file = NULL;
768 pp->pr_entered_line = 0;
769
770 simple_lock_init(&pp->pr_slock);
771
772 /*
773 * Initialize private page header pool and cache magazine pool if we
774 * haven't done so yet.
775 * XXX LOCKING.
776 */
777 if (phpool[0].pr_size == 0) {
778 int idx;
779 for (idx = 0; idx < PHPOOL_MAX; idx++) {
780 static char phpool_names[PHPOOL_MAX][6+1+6+1];
781 int nelem;
782 size_t sz;
783
784 nelem = PHPOOL_FREELIST_NELEM(idx);
785 snprintf(phpool_names[idx], sizeof(phpool_names[idx]),
786 "phpool-%d", nelem);
787 sz = sizeof(struct pool_item_header);
788 if (nelem) {
789 sz = PR_FREELIST_ALIGN(sz)
790 + nelem * sizeof(pool_item_freelist_t);
791 }
792 pool_init(&phpool[idx], sz, 0, 0, 0,
793 phpool_names[idx], &pool_allocator_meta);
794 }
795 #ifdef POOL_SUBPAGE
796 pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
797 PR_RECURSIVE, "psppool", &pool_allocator_meta);
798 #endif
799 pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
800 0, "pcgpool", &pool_allocator_meta);
801 }
802
803 /* Insert into the list of all pools. */
804 simple_lock(&pool_head_slock);
805 LIST_INSERT_HEAD(&pool_head, pp, pr_poollist);
806 simple_unlock(&pool_head_slock);
807
808 /* Insert this into the list of pools using this allocator. */
809 s = splvm();
810 simple_lock(&palloc->pa_slock);
811 TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
812 simple_unlock(&palloc->pa_slock);
813 splx(s);
814 pool_reclaim_register(pp);
815 }
816
817 /*
818 * De-commision a pool resource.
819 */
820 void
821 pool_destroy(struct pool *pp)
822 {
823 struct pool_pagelist pq;
824 struct pool_item_header *ph;
825 int s;
826
827 /* Remove from global pool list */
828 simple_lock(&pool_head_slock);
829 LIST_REMOVE(pp, pr_poollist);
830 if (drainpp == pp)
831 drainpp = NULL;
832 simple_unlock(&pool_head_slock);
833
834 /* Remove this pool from its allocator's list of pools. */
835 pool_reclaim_unregister(pp);
836 s = splvm();
837 simple_lock(&pp->pr_alloc->pa_slock);
838 TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
839 simple_unlock(&pp->pr_alloc->pa_slock);
840 splx(s);
841
842 s = splvm();
843 simple_lock(&pp->pr_slock);
844
845 KASSERT(LIST_EMPTY(&pp->pr_cachelist));
846
847 #ifdef DIAGNOSTIC
848 if (pp->pr_nout != 0) {
849 pr_printlog(pp, NULL, printf);
850 panic("pool_destroy: pool busy: still out: %u",
851 pp->pr_nout);
852 }
853 #endif
854
855 KASSERT(LIST_EMPTY(&pp->pr_fullpages));
856 KASSERT(LIST_EMPTY(&pp->pr_partpages));
857
858 /* Remove all pages */
859 LIST_INIT(&pq);
860 while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
861 pr_rmpage(pp, ph, &pq);
862
863 simple_unlock(&pp->pr_slock);
864 splx(s);
865
866 pr_pagelist_free(pp, &pq);
867
868 #ifdef POOL_DIAGNOSTIC
869 if ((pp->pr_roflags & PR_LOGGING) != 0)
870 free(pp->pr_log, M_TEMP);
871 #endif
872 }
873
874 void
875 pool_set_drain_hook(struct pool *pp, void (*fn)(void *, int), void *arg)
876 {
877
878 /* XXX no locking -- must be used just after pool_init() */
879 #ifdef DIAGNOSTIC
880 if (pp->pr_drain_hook != NULL)
881 panic("pool_set_drain_hook(%s): already set", pp->pr_wchan);
882 #endif
883 pp->pr_drain_hook = fn;
884 pp->pr_drain_hook_arg = arg;
885 }
886
887 static struct pool_item_header *
888 pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags)
889 {
890 struct pool_item_header *ph;
891 int s;
892
893 LOCK_ASSERT(simple_lock_held(&pp->pr_slock) == 0);
894
895 if ((pp->pr_roflags & PR_PHINPAGE) != 0)
896 ph = (struct pool_item_header *) (storage + pp->pr_phoffset);
897 else {
898 s = splvm();
899 ph = pool_get(pp->pr_phpool, flags);
900 splx(s);
901 }
902
903 return (ph);
904 }
905
906 /*
907 * Grab an item from the pool; must be called at appropriate spl level
908 */
909 void *
910 #ifdef POOL_DIAGNOSTIC
911 _pool_get(struct pool *pp, int flags, const char *file, long line)
912 #else
913 pool_get(struct pool *pp, int flags)
914 #endif
915 {
916 struct pool_item *pi;
917 struct pool_item_header *ph;
918 void *v;
919
920 #ifdef DIAGNOSTIC
921 if (__predict_false(pp->pr_itemsperpage == 0))
922 panic("pool_get: pool %p: pr_itemsperpage is zero, "
923 "pool not initialized?", pp);
924 if (__predict_false(curlwp == NULL && doing_shutdown == 0 &&
925 (flags & PR_WAITOK) != 0))
926 panic("pool_get: %s: must have NOWAIT", pp->pr_wchan);
927
928 #endif /* DIAGNOSTIC */
929 #ifdef LOCKDEBUG
930 if (flags & PR_WAITOK)
931 ASSERT_SLEEPABLE(NULL, "pool_get(PR_WAITOK)");
932 SCHED_ASSERT_UNLOCKED();
933 #endif
934
935 simple_lock(&pp->pr_slock);
936 pr_enter(pp, file, line);
937
938 startover:
939 /*
940 * Check to see if we've reached the hard limit. If we have,
941 * and we can wait, then wait until an item has been returned to
942 * the pool.
943 */
944 #ifdef DIAGNOSTIC
945 if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
946 pr_leave(pp);
947 simple_unlock(&pp->pr_slock);
948 panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
949 }
950 #endif
951 if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
952 if (pp->pr_drain_hook != NULL) {
953 /*
954 * Since the drain hook is going to free things
955 * back to the pool, unlock, call the hook, re-lock,
956 * and check the hardlimit condition again.
957 */
958 pr_leave(pp);
959 simple_unlock(&pp->pr_slock);
960 (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
961 simple_lock(&pp->pr_slock);
962 pr_enter(pp, file, line);
963 if (pp->pr_nout < pp->pr_hardlimit)
964 goto startover;
965 }
966
967 if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
968 /*
969 * XXX: A warning isn't logged in this case. Should
970 * it be?
971 */
972 pp->pr_flags |= PR_WANTED;
973 pr_leave(pp);
974 ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
975 pr_enter(pp, file, line);
976 goto startover;
977 }
978
979 /*
980 * Log a message that the hard limit has been hit.
981 */
982 if (pp->pr_hardlimit_warning != NULL &&
983 ratecheck(&pp->pr_hardlimit_warning_last,
984 &pp->pr_hardlimit_ratecap))
985 log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
986
987 pp->pr_nfail++;
988
989 pr_leave(pp);
990 simple_unlock(&pp->pr_slock);
991 return (NULL);
992 }
993
994 /*
995 * The convention we use is that if `curpage' is not NULL, then
996 * it points at a non-empty bucket. In particular, `curpage'
997 * never points at a page header which has PR_PHINPAGE set and
998 * has no items in its bucket.
999 */
1000 if ((ph = pp->pr_curpage) == NULL) {
1001 int error;
1002
1003 #ifdef DIAGNOSTIC
1004 if (pp->pr_nitems != 0) {
1005 simple_unlock(&pp->pr_slock);
1006 printf("pool_get: %s: curpage NULL, nitems %u\n",
1007 pp->pr_wchan, pp->pr_nitems);
1008 panic("pool_get: nitems inconsistent");
1009 }
1010 #endif
1011
1012 /*
1013 * Call the back-end page allocator for more memory.
1014 * Release the pool lock, as the back-end page allocator
1015 * may block.
1016 */
1017 pr_leave(pp);
1018 error = pool_grow(pp, flags);
1019 pr_enter(pp, file, line);
1020 if (error != 0) {
1021 /*
1022 * We were unable to allocate a page or item
1023 * header, but we released the lock during
1024 * allocation, so perhaps items were freed
1025 * back to the pool. Check for this case.
1026 */
1027 if (pp->pr_curpage != NULL)
1028 goto startover;
1029
1030 pp->pr_nfail++;
1031 pr_leave(pp);
1032 simple_unlock(&pp->pr_slock);
1033 return (NULL);
1034 }
1035
1036 /* Start the allocation process over. */
1037 goto startover;
1038 }
1039 if (pp->pr_roflags & PR_NOTOUCH) {
1040 #ifdef DIAGNOSTIC
1041 if (__predict_false(ph->ph_nmissing == pp->pr_itemsperpage)) {
1042 pr_leave(pp);
1043 simple_unlock(&pp->pr_slock);
1044 panic("pool_get: %s: page empty", pp->pr_wchan);
1045 }
1046 #endif
1047 v = pr_item_notouch_get(pp, ph);
1048 #ifdef POOL_DIAGNOSTIC
1049 pr_log(pp, v, PRLOG_GET, file, line);
1050 #endif
1051 } else {
1052 v = pi = LIST_FIRST(&ph->ph_itemlist);
1053 if (__predict_false(v == NULL)) {
1054 pr_leave(pp);
1055 simple_unlock(&pp->pr_slock);
1056 panic("pool_get: %s: page empty", pp->pr_wchan);
1057 }
1058 #ifdef DIAGNOSTIC
1059 if (__predict_false(pp->pr_nitems == 0)) {
1060 pr_leave(pp);
1061 simple_unlock(&pp->pr_slock);
1062 printf("pool_get: %s: items on itemlist, nitems %u\n",
1063 pp->pr_wchan, pp->pr_nitems);
1064 panic("pool_get: nitems inconsistent");
1065 }
1066 #endif
1067
1068 #ifdef POOL_DIAGNOSTIC
1069 pr_log(pp, v, PRLOG_GET, file, line);
1070 #endif
1071
1072 #ifdef DIAGNOSTIC
1073 if (__predict_false(pi->pi_magic != PI_MAGIC)) {
1074 pr_printlog(pp, pi, printf);
1075 panic("pool_get(%s): free list modified: "
1076 "magic=%x; page %p; item addr %p\n",
1077 pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
1078 }
1079 #endif
1080
1081 /*
1082 * Remove from item list.
1083 */
1084 LIST_REMOVE(pi, pi_list);
1085 }
1086 pp->pr_nitems--;
1087 pp->pr_nout++;
1088 if (ph->ph_nmissing == 0) {
1089 #ifdef DIAGNOSTIC
1090 if (__predict_false(pp->pr_nidle == 0))
1091 panic("pool_get: nidle inconsistent");
1092 #endif
1093 pp->pr_nidle--;
1094
1095 /*
1096 * This page was previously empty. Move it to the list of
1097 * partially-full pages. This page is already curpage.
1098 */
1099 LIST_REMOVE(ph, ph_pagelist);
1100 LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1101 }
1102 ph->ph_nmissing++;
1103 if (ph->ph_nmissing == pp->pr_itemsperpage) {
1104 #ifdef DIAGNOSTIC
1105 if (__predict_false((pp->pr_roflags & PR_NOTOUCH) == 0 &&
1106 !LIST_EMPTY(&ph->ph_itemlist))) {
1107 pr_leave(pp);
1108 simple_unlock(&pp->pr_slock);
1109 panic("pool_get: %s: nmissing inconsistent",
1110 pp->pr_wchan);
1111 }
1112 #endif
1113 /*
1114 * This page is now full. Move it to the full list
1115 * and select a new current page.
1116 */
1117 LIST_REMOVE(ph, ph_pagelist);
1118 LIST_INSERT_HEAD(&pp->pr_fullpages, ph, ph_pagelist);
1119 pool_update_curpage(pp);
1120 }
1121
1122 pp->pr_nget++;
1123 pr_leave(pp);
1124
1125 /*
1126 * If we have a low water mark and we are now below that low
1127 * water mark, add more items to the pool.
1128 */
1129 if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1130 /*
1131 * XXX: Should we log a warning? Should we set up a timeout
1132 * to try again in a second or so? The latter could break
1133 * a caller's assumptions about interrupt protection, etc.
1134 */
1135 }
1136
1137 simple_unlock(&pp->pr_slock);
1138 return (v);
1139 }
1140
1141 /*
1142 * Internal version of pool_put(). Pool is already locked/entered.
1143 */
1144 static void
1145 pool_do_put(struct pool *pp, void *v, struct pool_pagelist *pq)
1146 {
1147 struct pool_item *pi = v;
1148 struct pool_item_header *ph;
1149
1150 LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
1151 SCHED_ASSERT_UNLOCKED();
1152
1153 #ifdef DIAGNOSTIC
1154 if (__predict_false(pp->pr_nout == 0)) {
1155 printf("pool %s: putting with none out\n",
1156 pp->pr_wchan);
1157 panic("pool_put");
1158 }
1159 #endif
1160
1161 if (__predict_false((ph = pr_find_pagehead(pp, v)) == NULL)) {
1162 pr_printlog(pp, NULL, printf);
1163 panic("pool_put: %s: page header missing", pp->pr_wchan);
1164 }
1165
1166 #ifdef LOCKDEBUG
1167 /*
1168 * Check if we're freeing a locked simple lock.
1169 */
1170 simple_lock_freecheck((caddr_t)pi, ((caddr_t)pi) + pp->pr_size);
1171 #endif
1172
1173 /*
1174 * Return to item list.
1175 */
1176 if (pp->pr_roflags & PR_NOTOUCH) {
1177 pr_item_notouch_put(pp, ph, v);
1178 } else {
1179 #ifdef DIAGNOSTIC
1180 pi->pi_magic = PI_MAGIC;
1181 #endif
1182 #ifdef DEBUG
1183 {
1184 int i, *ip = v;
1185
1186 for (i = 0; i < pp->pr_size / sizeof(int); i++) {
1187 *ip++ = PI_MAGIC;
1188 }
1189 }
1190 #endif
1191
1192 LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1193 }
1194 KDASSERT(ph->ph_nmissing != 0);
1195 ph->ph_nmissing--;
1196 pp->pr_nput++;
1197 pp->pr_nitems++;
1198 pp->pr_nout--;
1199
1200 /* Cancel "pool empty" condition if it exists */
1201 if (pp->pr_curpage == NULL)
1202 pp->pr_curpage = ph;
1203
1204 if (pp->pr_flags & PR_WANTED) {
1205 pp->pr_flags &= ~PR_WANTED;
1206 if (ph->ph_nmissing == 0)
1207 pp->pr_nidle++;
1208 wakeup((caddr_t)pp);
1209 return;
1210 }
1211
1212 /*
1213 * If this page is now empty, do one of two things:
1214 *
1215 * (1) If we have more pages than the page high water mark,
1216 * free the page back to the system. ONLY CONSIDER
1217 * FREEING BACK A PAGE IF WE HAVE MORE THAN OUR MINIMUM PAGE
1218 * CLAIM.
1219 *
1220 * (2) Otherwise, move the page to the empty page list.
1221 *
1222 * Either way, select a new current page (so we use a partially-full
1223 * page if one is available).
1224 */
1225 if (ph->ph_nmissing == 0) {
1226 pp->pr_nidle++;
1227 if (pp->pr_npages > pp->pr_minpages &&
1228 (pp->pr_npages > pp->pr_maxpages ||
1229 pa_starved_p(pp->pr_alloc))) {
1230 pr_rmpage(pp, ph, pq);
1231 } else {
1232 LIST_REMOVE(ph, ph_pagelist);
1233 LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1234
1235 /*
1236 * Update the timestamp on the page. A page must
1237 * be idle for some period of time before it can
1238 * be reclaimed by the pagedaemon. This minimizes
1239 * ping-pong'ing for memory.
1240 */
1241 getmicrotime(&ph->ph_time);
1242 }
1243 pool_update_curpage(pp);
1244 }
1245
1246 /*
1247 * If the page was previously completely full, move it to the
1248 * partially-full list and make it the current page. The next
1249 * allocation will get the item from this page, instead of
1250 * further fragmenting the pool.
1251 */
1252 else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
1253 LIST_REMOVE(ph, ph_pagelist);
1254 LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1255 pp->pr_curpage = ph;
1256 }
1257 }
1258
1259 /*
1260 * Return resource to the pool; must be called at appropriate spl level
1261 */
1262 #ifdef POOL_DIAGNOSTIC
1263 void
1264 _pool_put(struct pool *pp, void *v, const char *file, long line)
1265 {
1266 struct pool_pagelist pq;
1267
1268 LIST_INIT(&pq);
1269
1270 simple_lock(&pp->pr_slock);
1271 pr_enter(pp, file, line);
1272
1273 pr_log(pp, v, PRLOG_PUT, file, line);
1274
1275 pool_do_put(pp, v, &pq);
1276
1277 pr_leave(pp);
1278 simple_unlock(&pp->pr_slock);
1279
1280 pr_pagelist_free(pp, &pq);
1281 }
1282 #undef pool_put
1283 #endif /* POOL_DIAGNOSTIC */
1284
1285 void
1286 pool_put(struct pool *pp, void *v)
1287 {
1288 struct pool_pagelist pq;
1289
1290 LIST_INIT(&pq);
1291
1292 simple_lock(&pp->pr_slock);
1293 pool_do_put(pp, v, &pq);
1294 simple_unlock(&pp->pr_slock);
1295
1296 pr_pagelist_free(pp, &pq);
1297 }
1298
1299 #ifdef POOL_DIAGNOSTIC
1300 #define pool_put(h, v) _pool_put((h), (v), __FILE__, __LINE__)
1301 #endif
1302
1303 /*
1304 * pool_grow: grow a pool by a page.
1305 *
1306 * => called with pool locked.
1307 * => unlock and relock the pool.
1308 * => return with pool locked.
1309 */
1310
1311 static int
1312 pool_grow(struct pool *pp, int flags)
1313 {
1314 struct pool_item_header *ph = NULL;
1315 char *cp;
1316
1317 simple_unlock(&pp->pr_slock);
1318 cp = pool_allocator_alloc(pp, flags);
1319 if (__predict_true(cp != NULL)) {
1320 ph = pool_alloc_item_header(pp, cp, flags);
1321 }
1322 if (__predict_false(cp == NULL || ph == NULL)) {
1323 if (cp != NULL) {
1324 pool_allocator_free(pp, cp);
1325 }
1326 simple_lock(&pp->pr_slock);
1327 return ENOMEM;
1328 }
1329
1330 simple_lock(&pp->pr_slock);
1331 pool_prime_page(pp, cp, ph);
1332 pp->pr_npagealloc++;
1333 return 0;
1334 }
1335
1336 /*
1337 * Add N items to the pool.
1338 */
1339 int
1340 pool_prime(struct pool *pp, int n)
1341 {
1342 int newpages;
1343 int error = 0;
1344
1345 simple_lock(&pp->pr_slock);
1346
1347 newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1348
1349 while (newpages-- > 0) {
1350 error = pool_grow(pp, PR_NOWAIT);
1351 if (error) {
1352 break;
1353 }
1354 pp->pr_minpages++;
1355 }
1356
1357 if (pp->pr_minpages >= pp->pr_maxpages)
1358 pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */
1359
1360 simple_unlock(&pp->pr_slock);
1361 return error;
1362 }
1363
1364 /*
1365 * Add a page worth of items to the pool.
1366 *
1367 * Note, we must be called with the pool descriptor LOCKED.
1368 */
1369 static void
1370 pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph)
1371 {
1372 struct pool_item *pi;
1373 caddr_t cp = storage;
1374 unsigned int align = pp->pr_align;
1375 unsigned int ioff = pp->pr_itemoffset;
1376 int n;
1377
1378 LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
1379
1380 #ifdef DIAGNOSTIC
1381 if ((pp->pr_roflags & PR_NOALIGN) == 0 &&
1382 ((uintptr_t)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
1383 panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1384 #endif
1385
1386 /*
1387 * Insert page header.
1388 */
1389 LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1390 LIST_INIT(&ph->ph_itemlist);
1391 ph->ph_page = storage;
1392 ph->ph_nmissing = 0;
1393 getmicrotime(&ph->ph_time);
1394 if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1395 SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
1396
1397 pp->pr_nidle++;
1398
1399 /*
1400 * Color this page.
1401 */
1402 cp = (caddr_t)(cp + pp->pr_curcolor);
1403 if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
1404 pp->pr_curcolor = 0;
1405
1406 /*
1407 * Adjust storage to apply aligment to `pr_itemoffset' in each item.
1408 */
1409 if (ioff != 0)
1410 cp = (caddr_t)(cp + (align - ioff));
1411
1412 /*
1413 * Insert remaining chunks on the bucket list.
1414 */
1415 n = pp->pr_itemsperpage;
1416 pp->pr_nitems += n;
1417
1418 if (pp->pr_roflags & PR_NOTOUCH) {
1419 pool_item_freelist_t *freelist = PR_FREELIST(ph);
1420 int i;
1421
1422 ph->ph_off = cp - storage;
1423 ph->ph_firstfree = 0;
1424 for (i = 0; i < n - 1; i++)
1425 freelist[i] = i + 1;
1426 freelist[n - 1] = PR_INDEX_EOL;
1427 } else {
1428 while (n--) {
1429 pi = (struct pool_item *)cp;
1430
1431 KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
1432
1433 /* Insert on page list */
1434 LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1435 #ifdef DIAGNOSTIC
1436 pi->pi_magic = PI_MAGIC;
1437 #endif
1438 cp = (caddr_t)(cp + pp->pr_size);
1439 }
1440 }
1441
1442 /*
1443 * If the pool was depleted, point at the new page.
1444 */
1445 if (pp->pr_curpage == NULL)
1446 pp->pr_curpage = ph;
1447
1448 if (++pp->pr_npages > pp->pr_hiwat)
1449 pp->pr_hiwat = pp->pr_npages;
1450 }
1451
1452 /*
1453 * Used by pool_get() when nitems drops below the low water mark. This
1454 * is used to catch up pr_nitems with the low water mark.
1455 *
1456 * Note 1, we never wait for memory here, we let the caller decide what to do.
1457 *
1458 * Note 2, we must be called with the pool already locked, and we return
1459 * with it locked.
1460 */
1461 static int
1462 pool_catchup(struct pool *pp)
1463 {
1464 int error = 0;
1465
1466 while (POOL_NEEDS_CATCHUP(pp)) {
1467 error = pool_grow(pp, PR_NOWAIT);
1468 if (error) {
1469 break;
1470 }
1471 }
1472 return error;
1473 }
1474
1475 static void
1476 pool_update_curpage(struct pool *pp)
1477 {
1478
1479 pp->pr_curpage = LIST_FIRST(&pp->pr_partpages);
1480 if (pp->pr_curpage == NULL) {
1481 pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages);
1482 }
1483 }
1484
1485 void
1486 pool_setlowat(struct pool *pp, int n)
1487 {
1488
1489 simple_lock(&pp->pr_slock);
1490
1491 pp->pr_minitems = n;
1492 pp->pr_minpages = (n == 0)
1493 ? 0
1494 : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1495
1496 /* Make sure we're caught up with the newly-set low water mark. */
1497 if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1498 /*
1499 * XXX: Should we log a warning? Should we set up a timeout
1500 * to try again in a second or so? The latter could break
1501 * a caller's assumptions about interrupt protection, etc.
1502 */
1503 }
1504
1505 simple_unlock(&pp->pr_slock);
1506 }
1507
1508 void
1509 pool_sethiwat(struct pool *pp, int n)
1510 {
1511
1512 simple_lock(&pp->pr_slock);
1513
1514 pp->pr_maxpages = (n == 0)
1515 ? 0
1516 : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1517
1518 simple_unlock(&pp->pr_slock);
1519 }
1520
1521 void
1522 pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1523 {
1524
1525 simple_lock(&pp->pr_slock);
1526
1527 pp->pr_hardlimit = n;
1528 pp->pr_hardlimit_warning = warnmess;
1529 pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1530 pp->pr_hardlimit_warning_last.tv_sec = 0;
1531 pp->pr_hardlimit_warning_last.tv_usec = 0;
1532
1533 /*
1534 * In-line version of pool_sethiwat(), because we don't want to
1535 * release the lock.
1536 */
1537 pp->pr_maxpages = (n == 0)
1538 ? 0
1539 : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1540
1541 simple_unlock(&pp->pr_slock);
1542 }
1543
1544 /*
1545 * Release all complete pages that have not been used recently.
1546 */
1547 int
1548 #ifdef POOL_DIAGNOSTIC
1549 _pool_reclaim(struct pool *pp, const char *file, long line)
1550 #else
1551 pool_reclaim(struct pool *pp)
1552 #endif
1553 {
1554 struct pool_item_header *ph, *phnext;
1555 struct pool_cache *pc;
1556 struct pool_pagelist pq;
1557 struct pool_cache_grouplist pcgl;
1558 struct timeval curtime, diff;
1559
1560 if (pp->pr_drain_hook != NULL) {
1561 /*
1562 * The drain hook must be called with the pool unlocked.
1563 */
1564 (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT);
1565 }
1566
1567 if (simple_lock_try(&pp->pr_slock) == 0)
1568 return (0);
1569 pr_enter(pp, file, line);
1570
1571 LIST_INIT(&pq);
1572 LIST_INIT(&pcgl);
1573
1574 /*
1575 * Reclaim items from the pool's caches.
1576 */
1577 LIST_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
1578 pool_cache_reclaim(pc, &pq, &pcgl);
1579
1580 getmicrotime(&curtime);
1581
1582 for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
1583 phnext = LIST_NEXT(ph, ph_pagelist);
1584
1585 /* Check our minimum page claim */
1586 if (pp->pr_npages <= pp->pr_minpages)
1587 break;
1588
1589 KASSERT(ph->ph_nmissing == 0);
1590 timersub(&curtime, &ph->ph_time, &diff);
1591 if (diff.tv_sec < pool_inactive_time
1592 && !pa_starved_p(pp->pr_alloc))
1593 continue;
1594
1595 /*
1596 * If freeing this page would put us below
1597 * the low water mark, stop now.
1598 */
1599 if ((pp->pr_nitems - pp->pr_itemsperpage) <
1600 pp->pr_minitems)
1601 break;
1602
1603 pr_rmpage(pp, ph, &pq);
1604 }
1605
1606 pr_leave(pp);
1607 simple_unlock(&pp->pr_slock);
1608 if (LIST_EMPTY(&pq) && LIST_EMPTY(&pcgl))
1609 return 0;
1610
1611 pr_pagelist_free(pp, &pq);
1612 pcg_grouplist_free(&pcgl);
1613 return (1);
1614 }
1615
1616 /*
1617 * Drain pools, one at a time.
1618 *
1619 * Note, we must never be called from an interrupt context.
1620 */
1621 void
1622 pool_drain(void *arg)
1623 {
1624 struct pool *pp;
1625 int s;
1626
1627 pp = NULL;
1628 s = splvm();
1629 simple_lock(&pool_head_slock);
1630 if (drainpp == NULL) {
1631 drainpp = LIST_FIRST(&pool_head);
1632 }
1633 if (drainpp) {
1634 pp = drainpp;
1635 drainpp = LIST_NEXT(pp, pr_poollist);
1636 }
1637 simple_unlock(&pool_head_slock);
1638 if (pp)
1639 pool_reclaim(pp);
1640 splx(s);
1641 }
1642
1643 /*
1644 * Diagnostic helpers.
1645 */
1646 void
1647 pool_print(struct pool *pp, const char *modif)
1648 {
1649 int s;
1650
1651 s = splvm();
1652 if (simple_lock_try(&pp->pr_slock) == 0) {
1653 printf("pool %s is locked; try again later\n",
1654 pp->pr_wchan);
1655 splx(s);
1656 return;
1657 }
1658 pool_print1(pp, modif, printf);
1659 simple_unlock(&pp->pr_slock);
1660 splx(s);
1661 }
1662
1663 void
1664 pool_printall(const char *modif, void (*pr)(const char *, ...))
1665 {
1666 struct pool *pp;
1667
1668 if (simple_lock_try(&pool_head_slock) == 0) {
1669 (*pr)("WARNING: pool_head_slock is locked\n");
1670 } else {
1671 simple_unlock(&pool_head_slock);
1672 }
1673
1674 LIST_FOREACH(pp, &pool_head, pr_poollist) {
1675 pool_printit(pp, modif, pr);
1676 }
1677 }
1678
1679 void
1680 pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1681 {
1682
1683 if (pp == NULL) {
1684 (*pr)("Must specify a pool to print.\n");
1685 return;
1686 }
1687
1688 /*
1689 * Called from DDB; interrupts should be blocked, and all
1690 * other processors should be paused. We can skip locking
1691 * the pool in this case.
1692 *
1693 * We do a simple_lock_try() just to print the lock
1694 * status, however.
1695 */
1696
1697 if (simple_lock_try(&pp->pr_slock) == 0)
1698 (*pr)("WARNING: pool %s is locked\n", pp->pr_wchan);
1699 else
1700 simple_unlock(&pp->pr_slock);
1701
1702 pool_print1(pp, modif, pr);
1703 }
1704
1705 static void
1706 pool_print_pagelist(struct pool *pp, struct pool_pagelist *pl,
1707 void (*pr)(const char *, ...))
1708 {
1709 struct pool_item_header *ph;
1710 #ifdef DIAGNOSTIC
1711 struct pool_item *pi;
1712 #endif
1713
1714 LIST_FOREACH(ph, pl, ph_pagelist) {
1715 (*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n",
1716 ph->ph_page, ph->ph_nmissing,
1717 (u_long)ph->ph_time.tv_sec,
1718 (u_long)ph->ph_time.tv_usec);
1719 #ifdef DIAGNOSTIC
1720 if (!(pp->pr_roflags & PR_NOTOUCH)) {
1721 LIST_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1722 if (pi->pi_magic != PI_MAGIC) {
1723 (*pr)("\t\t\titem %p, magic 0x%x\n",
1724 pi, pi->pi_magic);
1725 }
1726 }
1727 }
1728 #endif
1729 }
1730 }
1731
1732 static void
1733 pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1734 {
1735 struct pool_item_header *ph;
1736 struct pool_cache *pc;
1737 struct pool_cache_group *pcg;
1738 int i, print_log = 0, print_pagelist = 0, print_cache = 0;
1739 char c;
1740
1741 while ((c = *modif++) != '\0') {
1742 if (c == 'l')
1743 print_log = 1;
1744 if (c == 'p')
1745 print_pagelist = 1;
1746 if (c == 'c')
1747 print_cache = 1;
1748 }
1749
1750 (*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1751 pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1752 pp->pr_roflags);
1753 (*pr)("\talloc %p\n", pp->pr_alloc);
1754 (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1755 pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1756 (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1757 pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1758
1759 (*pr)("\n\tnget %lu, nfail %lu, nput %lu\n",
1760 pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1761 (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1762 pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1763
1764 if (print_pagelist == 0)
1765 goto skip_pagelist;
1766
1767 if ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1768 (*pr)("\n\tempty page list:\n");
1769 pool_print_pagelist(pp, &pp->pr_emptypages, pr);
1770 if ((ph = LIST_FIRST(&pp->pr_fullpages)) != NULL)
1771 (*pr)("\n\tfull page list:\n");
1772 pool_print_pagelist(pp, &pp->pr_fullpages, pr);
1773 if ((ph = LIST_FIRST(&pp->pr_partpages)) != NULL)
1774 (*pr)("\n\tpartial-page list:\n");
1775 pool_print_pagelist(pp, &pp->pr_partpages, pr);
1776
1777 if (pp->pr_curpage == NULL)
1778 (*pr)("\tno current page\n");
1779 else
1780 (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1781
1782 skip_pagelist:
1783 if (print_log == 0)
1784 goto skip_log;
1785
1786 (*pr)("\n");
1787 if ((pp->pr_roflags & PR_LOGGING) == 0)
1788 (*pr)("\tno log\n");
1789 else {
1790 pr_printlog(pp, NULL, pr);
1791 }
1792
1793 skip_log:
1794 if (print_cache == 0)
1795 goto skip_cache;
1796
1797 #define PR_GROUPLIST(pcg) \
1798 (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail); \
1799 for (i = 0; i < PCG_NOBJECTS; i++) { \
1800 if (pcg->pcg_objects[i].pcgo_pa != \
1801 POOL_PADDR_INVALID) { \
1802 (*pr)("\t\t\t%p, 0x%llx\n", \
1803 pcg->pcg_objects[i].pcgo_va, \
1804 (unsigned long long) \
1805 pcg->pcg_objects[i].pcgo_pa); \
1806 } else { \
1807 (*pr)("\t\t\t%p\n", \
1808 pcg->pcg_objects[i].pcgo_va); \
1809 } \
1810 }
1811
1812 LIST_FOREACH(pc, &pp->pr_cachelist, pc_poollist) {
1813 (*pr)("\tcache %p\n", pc);
1814 (*pr)("\t hits %lu misses %lu ngroups %lu nitems %lu\n",
1815 pc->pc_hits, pc->pc_misses, pc->pc_ngroups, pc->pc_nitems);
1816 (*pr)("\t full groups:\n");
1817 LIST_FOREACH(pcg, &pc->pc_fullgroups, pcg_list) {
1818 PR_GROUPLIST(pcg);
1819 }
1820 (*pr)("\t partial groups:\n");
1821 LIST_FOREACH(pcg, &pc->pc_partgroups, pcg_list) {
1822 PR_GROUPLIST(pcg);
1823 }
1824 (*pr)("\t empty groups:\n");
1825 LIST_FOREACH(pcg, &pc->pc_emptygroups, pcg_list) {
1826 PR_GROUPLIST(pcg);
1827 }
1828 }
1829 #undef PR_GROUPLIST
1830
1831 skip_cache:
1832 pr_enter_check(pp, pr);
1833 }
1834
1835 static int
1836 pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
1837 {
1838 struct pool_item *pi;
1839 caddr_t page;
1840 int n;
1841
1842 if ((pp->pr_roflags & PR_NOALIGN) == 0) {
1843 page = (caddr_t)((uintptr_t)ph & pp->pr_alloc->pa_pagemask);
1844 if (page != ph->ph_page &&
1845 (pp->pr_roflags & PR_PHINPAGE) != 0) {
1846 if (label != NULL)
1847 printf("%s: ", label);
1848 printf("pool(%p:%s): page inconsistency: page %p;"
1849 " at page head addr %p (p %p)\n", pp,
1850 pp->pr_wchan, ph->ph_page,
1851 ph, page);
1852 return 1;
1853 }
1854 }
1855
1856 if ((pp->pr_roflags & PR_NOTOUCH) != 0)
1857 return 0;
1858
1859 for (pi = LIST_FIRST(&ph->ph_itemlist), n = 0;
1860 pi != NULL;
1861 pi = LIST_NEXT(pi,pi_list), n++) {
1862
1863 #ifdef DIAGNOSTIC
1864 if (pi->pi_magic != PI_MAGIC) {
1865 if (label != NULL)
1866 printf("%s: ", label);
1867 printf("pool(%s): free list modified: magic=%x;"
1868 " page %p; item ordinal %d; addr %p\n",
1869 pp->pr_wchan, pi->pi_magic, ph->ph_page,
1870 n, pi);
1871 panic("pool");
1872 }
1873 #endif
1874 if ((pp->pr_roflags & PR_NOALIGN) != 0) {
1875 continue;
1876 }
1877 page = (caddr_t)((uintptr_t)pi & pp->pr_alloc->pa_pagemask);
1878 if (page == ph->ph_page)
1879 continue;
1880
1881 if (label != NULL)
1882 printf("%s: ", label);
1883 printf("pool(%p:%s): page inconsistency: page %p;"
1884 " item ordinal %d; addr %p (p %p)\n", pp,
1885 pp->pr_wchan, ph->ph_page,
1886 n, pi, page);
1887 return 1;
1888 }
1889 return 0;
1890 }
1891
1892
1893 int
1894 pool_chk(struct pool *pp, const char *label)
1895 {
1896 struct pool_item_header *ph;
1897 int r = 0;
1898
1899 simple_lock(&pp->pr_slock);
1900 LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
1901 r = pool_chk_page(pp, label, ph);
1902 if (r) {
1903 goto out;
1904 }
1905 }
1906 LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
1907 r = pool_chk_page(pp, label, ph);
1908 if (r) {
1909 goto out;
1910 }
1911 }
1912 LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
1913 r = pool_chk_page(pp, label, ph);
1914 if (r) {
1915 goto out;
1916 }
1917 }
1918
1919 out:
1920 simple_unlock(&pp->pr_slock);
1921 return (r);
1922 }
1923
1924 /*
1925 * pool_cache_init:
1926 *
1927 * Initialize a pool cache.
1928 *
1929 * NOTE: If the pool must be protected from interrupts, we expect
1930 * to be called at the appropriate interrupt priority level.
1931 */
1932 void
1933 pool_cache_init(struct pool_cache *pc, struct pool *pp,
1934 int (*ctor)(void *, void *, int),
1935 void (*dtor)(void *, void *),
1936 void *arg)
1937 {
1938
1939 LIST_INIT(&pc->pc_emptygroups);
1940 LIST_INIT(&pc->pc_fullgroups);
1941 LIST_INIT(&pc->pc_partgroups);
1942 simple_lock_init(&pc->pc_slock);
1943
1944 pc->pc_pool = pp;
1945
1946 pc->pc_ctor = ctor;
1947 pc->pc_dtor = dtor;
1948 pc->pc_arg = arg;
1949
1950 pc->pc_hits = 0;
1951 pc->pc_misses = 0;
1952
1953 pc->pc_ngroups = 0;
1954
1955 pc->pc_nitems = 0;
1956
1957 simple_lock(&pp->pr_slock);
1958 LIST_INSERT_HEAD(&pp->pr_cachelist, pc, pc_poollist);
1959 simple_unlock(&pp->pr_slock);
1960 }
1961
1962 /*
1963 * pool_cache_destroy:
1964 *
1965 * Destroy a pool cache.
1966 */
1967 void
1968 pool_cache_destroy(struct pool_cache *pc)
1969 {
1970 struct pool *pp = pc->pc_pool;
1971
1972 /* First, invalidate the entire cache. */
1973 pool_cache_invalidate(pc);
1974
1975 /* ...and remove it from the pool's cache list. */
1976 simple_lock(&pp->pr_slock);
1977 LIST_REMOVE(pc, pc_poollist);
1978 simple_unlock(&pp->pr_slock);
1979 }
1980
1981 static inline void *
1982 pcg_get(struct pool_cache_group *pcg, paddr_t *pap)
1983 {
1984 void *object;
1985 u_int idx;
1986
1987 KASSERT(pcg->pcg_avail <= PCG_NOBJECTS);
1988 KASSERT(pcg->pcg_avail != 0);
1989 idx = --pcg->pcg_avail;
1990
1991 KASSERT(pcg->pcg_objects[idx].pcgo_va != NULL);
1992 object = pcg->pcg_objects[idx].pcgo_va;
1993 if (pap != NULL)
1994 *pap = pcg->pcg_objects[idx].pcgo_pa;
1995 pcg->pcg_objects[idx].pcgo_va = NULL;
1996
1997 return (object);
1998 }
1999
2000 static inline void
2001 pcg_put(struct pool_cache_group *pcg, void *object, paddr_t pa)
2002 {
2003 u_int idx;
2004
2005 KASSERT(pcg->pcg_avail < PCG_NOBJECTS);
2006 idx = pcg->pcg_avail++;
2007
2008 KASSERT(pcg->pcg_objects[idx].pcgo_va == NULL);
2009 pcg->pcg_objects[idx].pcgo_va = object;
2010 pcg->pcg_objects[idx].pcgo_pa = pa;
2011 }
2012
2013 static void
2014 pcg_grouplist_free(struct pool_cache_grouplist *pcgl)
2015 {
2016 struct pool_cache_group *pcg;
2017 int s;
2018
2019 s = splvm();
2020 while ((pcg = LIST_FIRST(pcgl)) != NULL) {
2021 LIST_REMOVE(pcg, pcg_list);
2022 pool_put(&pcgpool, pcg);
2023 }
2024 splx(s);
2025 }
2026
2027 /*
2028 * pool_cache_get{,_paddr}:
2029 *
2030 * Get an object from a pool cache (optionally returning
2031 * the physical address of the object).
2032 */
2033 void *
2034 pool_cache_get_paddr(struct pool_cache *pc, int flags, paddr_t *pap)
2035 {
2036 struct pool_cache_group *pcg;
2037 void *object;
2038
2039 #ifdef LOCKDEBUG
2040 if (flags & PR_WAITOK)
2041 ASSERT_SLEEPABLE(NULL, "pool_cache_get(PR_WAITOK)");
2042 #endif
2043
2044 simple_lock(&pc->pc_slock);
2045
2046 pcg = LIST_FIRST(&pc->pc_partgroups);
2047 if (pcg == NULL) {
2048 pcg = LIST_FIRST(&pc->pc_fullgroups);
2049 if (pcg != NULL) {
2050 LIST_REMOVE(pcg, pcg_list);
2051 LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list);
2052 }
2053 }
2054 if (pcg == NULL) {
2055
2056 /*
2057 * No groups with any available objects. Allocate
2058 * a new object, construct it, and return it to
2059 * the caller. We will allocate a group, if necessary,
2060 * when the object is freed back to the cache.
2061 */
2062 pc->pc_misses++;
2063 simple_unlock(&pc->pc_slock);
2064 object = pool_get(pc->pc_pool, flags);
2065 if (object != NULL && pc->pc_ctor != NULL) {
2066 if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) {
2067 pool_put(pc->pc_pool, object);
2068 return (NULL);
2069 }
2070 }
2071 if (object != NULL && pap != NULL) {
2072 #ifdef POOL_VTOPHYS
2073 *pap = POOL_VTOPHYS(object);
2074 #else
2075 *pap = POOL_PADDR_INVALID;
2076 #endif
2077 }
2078 return (object);
2079 }
2080
2081 pc->pc_hits++;
2082 pc->pc_nitems--;
2083 object = pcg_get(pcg, pap);
2084
2085 if (pcg->pcg_avail == 0) {
2086 LIST_REMOVE(pcg, pcg_list);
2087 LIST_INSERT_HEAD(&pc->pc_emptygroups, pcg, pcg_list);
2088 }
2089 simple_unlock(&pc->pc_slock);
2090
2091 return (object);
2092 }
2093
2094 /*
2095 * pool_cache_put{,_paddr}:
2096 *
2097 * Put an object back to the pool cache (optionally caching the
2098 * physical address of the object).
2099 */
2100 void
2101 pool_cache_put_paddr(struct pool_cache *pc, void *object, paddr_t pa)
2102 {
2103 struct pool_cache_group *pcg;
2104 int s;
2105
2106 if (__predict_false((pc->pc_pool->pr_flags & PR_WANTED) != 0)) {
2107 goto destruct;
2108 }
2109
2110 simple_lock(&pc->pc_slock);
2111
2112 pcg = LIST_FIRST(&pc->pc_partgroups);
2113 if (pcg == NULL) {
2114 pcg = LIST_FIRST(&pc->pc_emptygroups);
2115 if (pcg != NULL) {
2116 LIST_REMOVE(pcg, pcg_list);
2117 LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list);
2118 }
2119 }
2120 if (pcg == NULL) {
2121
2122 /*
2123 * No empty groups to free the object to. Attempt to
2124 * allocate one.
2125 */
2126 simple_unlock(&pc->pc_slock);
2127 s = splvm();
2128 pcg = pool_get(&pcgpool, PR_NOWAIT);
2129 splx(s);
2130 if (pcg == NULL) {
2131 destruct:
2132
2133 /*
2134 * Unable to allocate a cache group; destruct the object
2135 * and free it back to the pool.
2136 */
2137 pool_cache_destruct_object(pc, object);
2138 return;
2139 }
2140 memset(pcg, 0, sizeof(*pcg));
2141 simple_lock(&pc->pc_slock);
2142 pc->pc_ngroups++;
2143 LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list);
2144 }
2145
2146 pc->pc_nitems++;
2147 pcg_put(pcg, object, pa);
2148
2149 if (pcg->pcg_avail == PCG_NOBJECTS) {
2150 LIST_REMOVE(pcg, pcg_list);
2151 LIST_INSERT_HEAD(&pc->pc_fullgroups, pcg, pcg_list);
2152 }
2153 simple_unlock(&pc->pc_slock);
2154 }
2155
2156 /*
2157 * pool_cache_destruct_object:
2158 *
2159 * Force destruction of an object and its release back into
2160 * the pool.
2161 */
2162 void
2163 pool_cache_destruct_object(struct pool_cache *pc, void *object)
2164 {
2165
2166 if (pc->pc_dtor != NULL)
2167 (*pc->pc_dtor)(pc->pc_arg, object);
2168 pool_put(pc->pc_pool, object);
2169 }
2170
2171 static void
2172 pool_do_cache_invalidate_grouplist(struct pool_cache_grouplist *pcgsl,
2173 struct pool_cache *pc, struct pool_pagelist *pq,
2174 struct pool_cache_grouplist *pcgdl)
2175 {
2176 struct pool_cache_group *pcg, *npcg;
2177 void *object;
2178
2179 for (pcg = LIST_FIRST(pcgsl); pcg != NULL; pcg = npcg) {
2180 npcg = LIST_NEXT(pcg, pcg_list);
2181 while (pcg->pcg_avail != 0) {
2182 pc->pc_nitems--;
2183 object = pcg_get(pcg, NULL);
2184 if (pc->pc_dtor != NULL)
2185 (*pc->pc_dtor)(pc->pc_arg, object);
2186 pool_do_put(pc->pc_pool, object, pq);
2187 }
2188 pc->pc_ngroups--;
2189 LIST_REMOVE(pcg, pcg_list);
2190 LIST_INSERT_HEAD(pcgdl, pcg, pcg_list);
2191 }
2192 }
2193
2194 static void
2195 pool_do_cache_invalidate(struct pool_cache *pc, struct pool_pagelist *pq,
2196 struct pool_cache_grouplist *pcgl)
2197 {
2198
2199 LOCK_ASSERT(simple_lock_held(&pc->pc_slock));
2200 LOCK_ASSERT(simple_lock_held(&pc->pc_pool->pr_slock));
2201
2202 pool_do_cache_invalidate_grouplist(&pc->pc_fullgroups, pc, pq, pcgl);
2203 pool_do_cache_invalidate_grouplist(&pc->pc_partgroups, pc, pq, pcgl);
2204
2205 KASSERT(LIST_EMPTY(&pc->pc_partgroups));
2206 KASSERT(LIST_EMPTY(&pc->pc_fullgroups));
2207 KASSERT(pc->pc_nitems == 0);
2208 }
2209
2210 /*
2211 * pool_cache_invalidate:
2212 *
2213 * Invalidate a pool cache (destruct and release all of the
2214 * cached objects).
2215 */
2216 void
2217 pool_cache_invalidate(struct pool_cache *pc)
2218 {
2219 struct pool_pagelist pq;
2220 struct pool_cache_grouplist pcgl;
2221
2222 LIST_INIT(&pq);
2223 LIST_INIT(&pcgl);
2224
2225 simple_lock(&pc->pc_slock);
2226 simple_lock(&pc->pc_pool->pr_slock);
2227
2228 pool_do_cache_invalidate(pc, &pq, &pcgl);
2229
2230 simple_unlock(&pc->pc_pool->pr_slock);
2231 simple_unlock(&pc->pc_slock);
2232
2233 pr_pagelist_free(pc->pc_pool, &pq);
2234 pcg_grouplist_free(&pcgl);
2235 }
2236
2237 /*
2238 * pool_cache_reclaim:
2239 *
2240 * Reclaim a pool cache for pool_reclaim().
2241 */
2242 static void
2243 pool_cache_reclaim(struct pool_cache *pc, struct pool_pagelist *pq,
2244 struct pool_cache_grouplist *pcgl)
2245 {
2246
2247 /*
2248 * We're locking in the wrong order (normally pool_cache -> pool,
2249 * but the pool is already locked when we get here), so we have
2250 * to use trylock. If we can't lock the pool_cache, it's not really
2251 * a big deal here.
2252 */
2253 if (simple_lock_try(&pc->pc_slock) == 0)
2254 return;
2255
2256 pool_do_cache_invalidate(pc, pq, pcgl);
2257
2258 simple_unlock(&pc->pc_slock);
2259 }
2260
2261 /*
2262 * Pool backend allocators.
2263 *
2264 * Each pool has a backend allocator that handles allocation, deallocation,
2265 * and any additional draining that might be needed.
2266 *
2267 * We provide two standard allocators:
2268 *
2269 * pool_allocator_kmem - the default when no allocator is specified
2270 *
2271 * pool_allocator_nointr - used for pools that will not be accessed
2272 * in interrupt context.
2273 */
2274 void *pool_page_alloc(struct pool *, int);
2275 void pool_page_free(struct pool *, void *);
2276
2277 #ifdef POOL_SUBPAGE
2278 struct pool_allocator pool_allocator_kmem_fullpage = {
2279 pool_page_alloc, pool_page_free, 0,
2280 .pa_backingmapptr = &kmem_map,
2281 };
2282 #else
2283 struct pool_allocator pool_allocator_kmem = {
2284 pool_page_alloc, pool_page_free, 0,
2285 .pa_backingmapptr = &kmem_map,
2286 };
2287 #endif
2288
2289 void *pool_page_alloc_nointr(struct pool *, int);
2290 void pool_page_free_nointr(struct pool *, void *);
2291
2292 #ifdef POOL_SUBPAGE
2293 struct pool_allocator pool_allocator_nointr_fullpage = {
2294 pool_page_alloc_nointr, pool_page_free_nointr, 0,
2295 .pa_backingmapptr = &kernel_map,
2296 };
2297 #else
2298 struct pool_allocator pool_allocator_nointr = {
2299 pool_page_alloc_nointr, pool_page_free_nointr, 0,
2300 .pa_backingmapptr = &kernel_map,
2301 };
2302 #endif
2303
2304 #ifdef POOL_SUBPAGE
2305 void *pool_subpage_alloc(struct pool *, int);
2306 void pool_subpage_free(struct pool *, void *);
2307
2308 struct pool_allocator pool_allocator_kmem = {
2309 pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE,
2310 .pa_backingmapptr = &kmem_map,
2311 };
2312
2313 void *pool_subpage_alloc_nointr(struct pool *, int);
2314 void pool_subpage_free_nointr(struct pool *, void *);
2315
2316 struct pool_allocator pool_allocator_nointr = {
2317 pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE,
2318 .pa_backingmapptr = &kmem_map,
2319 };
2320 #endif /* POOL_SUBPAGE */
2321
2322 static void *
2323 pool_allocator_alloc(struct pool *pp, int flags)
2324 {
2325 struct pool_allocator *pa = pp->pr_alloc;
2326 void *res;
2327
2328 LOCK_ASSERT(!simple_lock_held(&pp->pr_slock));
2329
2330 res = (*pa->pa_alloc)(pp, flags);
2331 if (res == NULL && (flags & PR_WAITOK) == 0) {
2332 /*
2333 * We only run the drain hook here if PR_NOWAIT.
2334 * In other cases, the hook will be run in
2335 * pool_reclaim().
2336 */
2337 if (pp->pr_drain_hook != NULL) {
2338 (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
2339 res = (*pa->pa_alloc)(pp, flags);
2340 }
2341 }
2342 return res;
2343 }
2344
2345 static void
2346 pool_allocator_free(struct pool *pp, void *v)
2347 {
2348 struct pool_allocator *pa = pp->pr_alloc;
2349
2350 LOCK_ASSERT(!simple_lock_held(&pp->pr_slock));
2351
2352 (*pa->pa_free)(pp, v);
2353 }
2354
2355 void *
2356 pool_page_alloc(struct pool *pp, int flags)
2357 {
2358 boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2359
2360 return ((void *) uvm_km_alloc_poolpage_cache(kmem_map, waitok));
2361 }
2362
2363 void
2364 pool_page_free(struct pool *pp, void *v)
2365 {
2366
2367 uvm_km_free_poolpage_cache(kmem_map, (vaddr_t) v);
2368 }
2369
2370 static void *
2371 pool_page_alloc_meta(struct pool *pp, int flags)
2372 {
2373 boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2374
2375 return ((void *) uvm_km_alloc_poolpage(kmem_map, waitok));
2376 }
2377
2378 static void
2379 pool_page_free_meta(struct pool *pp, void *v)
2380 {
2381
2382 uvm_km_free_poolpage(kmem_map, (vaddr_t) v);
2383 }
2384
2385 #ifdef POOL_SUBPAGE
2386 /* Sub-page allocator, for machines with large hardware pages. */
2387 void *
2388 pool_subpage_alloc(struct pool *pp, int flags)
2389 {
2390 void *v;
2391 int s;
2392 s = splvm();
2393 v = pool_get(&psppool, flags);
2394 splx(s);
2395 return v;
2396 }
2397
2398 void
2399 pool_subpage_free(struct pool *pp, void *v)
2400 {
2401 int s;
2402 s = splvm();
2403 pool_put(&psppool, v);
2404 splx(s);
2405 }
2406
2407 /* We don't provide a real nointr allocator. Maybe later. */
2408 void *
2409 pool_subpage_alloc_nointr(struct pool *pp, int flags)
2410 {
2411
2412 return (pool_subpage_alloc(pp, flags));
2413 }
2414
2415 void
2416 pool_subpage_free_nointr(struct pool *pp, void *v)
2417 {
2418
2419 pool_subpage_free(pp, v);
2420 }
2421 #endif /* POOL_SUBPAGE */
2422 void *
2423 pool_page_alloc_nointr(struct pool *pp, int flags)
2424 {
2425 boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2426
2427 return ((void *) uvm_km_alloc_poolpage_cache(kernel_map, waitok));
2428 }
2429
2430 void
2431 pool_page_free_nointr(struct pool *pp, void *v)
2432 {
2433
2434 uvm_km_free_poolpage_cache(kernel_map, (vaddr_t) v);
2435 }
Cache object: d2c73153b0a19b03b11a23c743c03ed5
|