Annotation of src/sys/kern/subr_pool.c, Revision 1.128.2.4
1.128.2.4! ad 1: /* $NetBSD: subr_pool.c,v 1.128.2.3 2007/03/21 20:10:22 ad Exp $ */
1.1 pk 2:
3: /*-
1.128.2.2 ad 4: * Copyright (c) 1997, 1999, 2000, 2002, 2007 The NetBSD Foundation, Inc.
1.1 pk 5: * All rights reserved.
6: *
7: * This code is derived from software contributed to The NetBSD Foundation
1.20 thorpej 8: * by Paul Kranenburg; by Jason R. Thorpe of the Numerical Aerospace
9: * Simulation Facility, NASA Ames Research Center.
1.1 pk 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:
1.13 christos 21: * This product includes software developed by the NetBSD
22: * Foundation, Inc. and its contributors.
1.1 pk 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: */
1.64 lukem 39:
40: #include <sys/cdefs.h>
1.128.2.4! ad 41: __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.128.2.3 2007/03/21 20:10:22 ad Exp $");
1.24 scottr 42:
1.25 thorpej 43: #include "opt_pool.h"
1.24 scottr 44: #include "opt_poollog.h"
1.28 thorpej 45: #include "opt_lockdebug.h"
1.1 pk 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>
1.20 thorpej 55: #include <sys/syslog.h>
1.125 ad 56: #include <sys/debug.h>
1.3 pk 57:
58: #include <uvm/uvm.h>
59:
1.1 pk 60: /*
61: * Pool resource management utility.
1.3 pk 62: *
1.88 chs 63: * Memory is allocated in pages which are split into pieces according to
64: * the pool item size. Each page is kept on one of three lists in the
65: * pool structure: `pr_emptypages', `pr_fullpages' and `pr_partpages',
66: * for empty, full and partially-full pages respectively. The individual
67: * pool items are on a linked list headed by `ph_itemlist' in each page
68: * header. The memory for building the page list is either taken from
69: * the allocated pages themselves (for small pool items) or taken from
70: * an internal pool of page headers (`phpool').
1.1 pk 71: */
72:
1.3 pk 73: /* List of all pools */
1.102 chs 74: LIST_HEAD(,pool) pool_head = LIST_HEAD_INITIALIZER(pool_head);
1.3 pk 75:
76: /* Private pool for page header structures */
1.97 yamt 77: #define PHPOOL_MAX 8
78: static struct pool phpool[PHPOOL_MAX];
79: #define PHPOOL_FREELIST_NELEM(idx) (((idx) == 0) ? 0 : (1 << (idx)))
1.3 pk 80:
1.62 bjh21 81: #ifdef POOL_SUBPAGE
82: /* Pool of subpages for use by normal pools. */
83: static struct pool psppool;
84: #endif
85:
1.117 yamt 86: static SLIST_HEAD(, pool_allocator) pa_deferinitq =
87: SLIST_HEAD_INITIALIZER(pa_deferinitq);
88:
1.98 yamt 89: static void *pool_page_alloc_meta(struct pool *, int);
90: static void pool_page_free_meta(struct pool *, void *);
91:
92: /* allocator for pool metadata */
93: static struct pool_allocator pool_allocator_meta = {
1.117 yamt 94: pool_page_alloc_meta, pool_page_free_meta,
95: .pa_backingmapptr = &kmem_map,
1.98 yamt 96: };
97:
1.3 pk 98: /* # of seconds to retain page after last use */
99: int pool_inactive_time = 10;
100:
101: /* Next candidate for drainage (see pool_drain()) */
1.23 thorpej 102: static struct pool *drainpp;
103:
1.128.2.2 ad 104: /* This lock protects both pool_head and drainpp. */
105: static kmutex_t pool_head_lock;
1.3 pk 106:
1.99 yamt 107: typedef uint8_t pool_item_freelist_t;
108:
1.3 pk 109: struct pool_item_header {
110: /* Page headers */
1.88 chs 111: LIST_ENTRY(pool_item_header)
1.3 pk 112: ph_pagelist; /* pool page list */
1.88 chs 113: SPLAY_ENTRY(pool_item_header)
114: ph_node; /* Off-page page headers */
1.128 christos 115: void * ph_page; /* this page's address */
1.3 pk 116: struct timeval ph_time; /* last referenced */
1.97 yamt 117: union {
118: /* !PR_NOTOUCH */
119: struct {
1.102 chs 120: LIST_HEAD(, pool_item)
1.97 yamt 121: phu_itemlist; /* chunk list for this page */
122: } phu_normal;
123: /* PR_NOTOUCH */
124: struct {
125: uint16_t
126: phu_off; /* start offset in page */
1.99 yamt 127: pool_item_freelist_t
1.97 yamt 128: phu_firstfree; /* first free item */
1.99 yamt 129: /*
130: * XXX it might be better to use
131: * a simple bitmap and ffs(3)
132: */
1.97 yamt 133: } phu_notouch;
134: } ph_u;
135: uint16_t ph_nmissing; /* # of chunks in use */
1.3 pk 136: };
1.97 yamt 137: #define ph_itemlist ph_u.phu_normal.phu_itemlist
138: #define ph_off ph_u.phu_notouch.phu_off
139: #define ph_firstfree ph_u.phu_notouch.phu_firstfree
1.3 pk 140:
1.1 pk 141: struct pool_item {
1.3 pk 142: #ifdef DIAGNOSTIC
1.82 thorpej 143: u_int pi_magic;
1.33 chs 144: #endif
1.82 thorpej 145: #define PI_MAGIC 0xdeadbeefU
1.3 pk 146: /* Other entries use only this list entry */
1.102 chs 147: LIST_ENTRY(pool_item) pi_list;
1.3 pk 148: };
149:
1.53 thorpej 150: #define POOL_NEEDS_CATCHUP(pp) \
151: ((pp)->pr_nitems < (pp)->pr_minitems)
152:
1.43 thorpej 153: /*
154: * Pool cache management.
155: *
156: * Pool caches provide a way for constructed objects to be cached by the
157: * pool subsystem. This can lead to performance improvements by avoiding
158: * needless object construction/destruction; it is deferred until absolutely
159: * necessary.
160: *
161: * Caches are grouped into cache groups. Each cache group references
162: * up to 16 constructed objects. When a cache allocates an object
163: * from the pool, it calls the object's constructor and places it into
164: * a cache group. When a cache group frees an object back to the pool,
165: * it first calls the object's destructor. This allows the object to
166: * persist in constructed form while freed to the cache.
167: *
168: * Multiple caches may exist for each pool. This allows a single
169: * object type to have multiple constructed forms. The pool references
170: * each cache, so that when a pool is drained by the pagedaemon, it can
171: * drain each individual cache as well. Each time a cache is drained,
172: * the most idle cache group is freed to the pool in its entirety.
173: *
174: * Pool caches are layed on top of pools. By layering them, we can avoid
175: * the complexity of cache management for pools which would not benefit
176: * from it.
177: */
178:
179: /* The cache group pool. */
180: static struct pool pcgpool;
1.3 pk 181:
1.102 chs 182: static void pool_cache_reclaim(struct pool_cache *, struct pool_pagelist *,
183: struct pool_cache_grouplist *);
184: static void pcg_grouplist_free(struct pool_cache_grouplist *);
1.3 pk 185:
1.42 thorpej 186: static int pool_catchup(struct pool *);
1.128 christos 187: static void pool_prime_page(struct pool *, void *,
1.55 thorpej 188: struct pool_item_header *);
1.88 chs 189: static void pool_update_curpage(struct pool *);
1.66 thorpej 190:
1.113 yamt 191: static int pool_grow(struct pool *, int);
1.117 yamt 192: static void *pool_allocator_alloc(struct pool *, int);
193: static void pool_allocator_free(struct pool *, void *);
1.3 pk 194:
1.97 yamt 195: static void pool_print_pagelist(struct pool *, struct pool_pagelist *,
1.88 chs 196: void (*)(const char *, ...));
1.42 thorpej 197: static void pool_print1(struct pool *, const char *,
198: void (*)(const char *, ...));
1.3 pk 199:
1.88 chs 200: static int pool_chk_page(struct pool *, const char *,
201: struct pool_item_header *);
202:
1.3 pk 203: /*
1.52 thorpej 204: * Pool log entry. An array of these is allocated in pool_init().
1.3 pk 205: */
206: struct pool_log {
207: const char *pl_file;
208: long pl_line;
209: int pl_action;
1.25 thorpej 210: #define PRLOG_GET 1
211: #define PRLOG_PUT 2
1.3 pk 212: void *pl_addr;
1.1 pk 213: };
214:
1.86 matt 215: #ifdef POOL_DIAGNOSTIC
1.3 pk 216: /* Number of entries in pool log buffers */
1.17 thorpej 217: #ifndef POOL_LOGSIZE
218: #define POOL_LOGSIZE 10
219: #endif
220:
221: int pool_logsize = POOL_LOGSIZE;
1.1 pk 222:
1.110 perry 223: static inline void
1.42 thorpej 224: pr_log(struct pool *pp, void *v, int action, const char *file, long line)
1.3 pk 225: {
226: int n = pp->pr_curlogentry;
227: struct pool_log *pl;
228:
1.20 thorpej 229: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 230: return;
231:
232: /*
233: * Fill in the current entry. Wrap around and overwrite
234: * the oldest entry if necessary.
235: */
236: pl = &pp->pr_log[n];
237: pl->pl_file = file;
238: pl->pl_line = line;
239: pl->pl_action = action;
240: pl->pl_addr = v;
241: if (++n >= pp->pr_logsize)
242: n = 0;
243: pp->pr_curlogentry = n;
244: }
245:
246: static void
1.42 thorpej 247: pr_printlog(struct pool *pp, struct pool_item *pi,
248: void (*pr)(const char *, ...))
1.3 pk 249: {
250: int i = pp->pr_logsize;
251: int n = pp->pr_curlogentry;
252:
1.20 thorpej 253: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 254: return;
255:
256: /*
257: * Print all entries in this pool's log.
258: */
259: while (i-- > 0) {
260: struct pool_log *pl = &pp->pr_log[n];
261: if (pl->pl_action != 0) {
1.25 thorpej 262: if (pi == NULL || pi == pl->pl_addr) {
263: (*pr)("\tlog entry %d:\n", i);
264: (*pr)("\t\taction = %s, addr = %p\n",
265: pl->pl_action == PRLOG_GET ? "get" : "put",
266: pl->pl_addr);
267: (*pr)("\t\tfile: %s at line %lu\n",
268: pl->pl_file, pl->pl_line);
269: }
1.3 pk 270: }
271: if (++n >= pp->pr_logsize)
272: n = 0;
273: }
274: }
1.25 thorpej 275:
1.110 perry 276: static inline void
1.42 thorpej 277: pr_enter(struct pool *pp, const char *file, long line)
1.25 thorpej 278: {
279:
1.34 thorpej 280: if (__predict_false(pp->pr_entered_file != NULL)) {
1.25 thorpej 281: printf("pool %s: reentrancy at file %s line %ld\n",
282: pp->pr_wchan, file, line);
283: printf(" previous entry at file %s line %ld\n",
284: pp->pr_entered_file, pp->pr_entered_line);
285: panic("pr_enter");
286: }
287:
288: pp->pr_entered_file = file;
289: pp->pr_entered_line = line;
290: }
291:
1.110 perry 292: static inline void
1.42 thorpej 293: pr_leave(struct pool *pp)
1.25 thorpej 294: {
295:
1.34 thorpej 296: if (__predict_false(pp->pr_entered_file == NULL)) {
1.25 thorpej 297: printf("pool %s not entered?\n", pp->pr_wchan);
298: panic("pr_leave");
299: }
300:
301: pp->pr_entered_file = NULL;
302: pp->pr_entered_line = 0;
303: }
304:
1.110 perry 305: static inline void
1.42 thorpej 306: pr_enter_check(struct pool *pp, void (*pr)(const char *, ...))
1.25 thorpej 307: {
308:
309: if (pp->pr_entered_file != NULL)
310: (*pr)("\n\tcurrently entered from file %s line %ld\n",
311: pp->pr_entered_file, pp->pr_entered_line);
312: }
1.3 pk 313: #else
1.25 thorpej 314: #define pr_log(pp, v, action, file, line)
315: #define pr_printlog(pp, pi, pr)
316: #define pr_enter(pp, file, line)
317: #define pr_leave(pp)
318: #define pr_enter_check(pp, pr)
1.59 thorpej 319: #endif /* POOL_DIAGNOSTIC */
1.3 pk 320:
1.110 perry 321: static inline int
1.97 yamt 322: pr_item_notouch_index(const struct pool *pp, const struct pool_item_header *ph,
323: const void *v)
324: {
325: const char *cp = v;
326: int idx;
327:
328: KASSERT(pp->pr_roflags & PR_NOTOUCH);
1.128 christos 329: idx = (cp - (char *)ph->ph_page - ph->ph_off) / pp->pr_size;
1.97 yamt 330: KASSERT(idx < pp->pr_itemsperpage);
331: return idx;
332: }
333:
1.99 yamt 334: #define PR_FREELIST_ALIGN(p) \
335: roundup((uintptr_t)(p), sizeof(pool_item_freelist_t))
336: #define PR_FREELIST(ph) ((pool_item_freelist_t *)PR_FREELIST_ALIGN((ph) + 1))
337: #define PR_INDEX_USED ((pool_item_freelist_t)-1)
338: #define PR_INDEX_EOL ((pool_item_freelist_t)-2)
1.97 yamt 339:
1.110 perry 340: static inline void
1.97 yamt 341: pr_item_notouch_put(const struct pool *pp, struct pool_item_header *ph,
342: void *obj)
343: {
344: int idx = pr_item_notouch_index(pp, ph, obj);
1.99 yamt 345: pool_item_freelist_t *freelist = PR_FREELIST(ph);
1.97 yamt 346:
347: KASSERT(freelist[idx] == PR_INDEX_USED);
348: freelist[idx] = ph->ph_firstfree;
349: ph->ph_firstfree = idx;
350: }
351:
1.110 perry 352: static inline void *
1.97 yamt 353: pr_item_notouch_get(const struct pool *pp, struct pool_item_header *ph)
354: {
355: int idx = ph->ph_firstfree;
1.99 yamt 356: pool_item_freelist_t *freelist = PR_FREELIST(ph);
1.97 yamt 357:
358: KASSERT(freelist[idx] != PR_INDEX_USED);
359: ph->ph_firstfree = freelist[idx];
360: freelist[idx] = PR_INDEX_USED;
361:
1.128 christos 362: return (char *)ph->ph_page + ph->ph_off + idx * pp->pr_size;
1.97 yamt 363: }
364:
1.110 perry 365: static inline int
1.88 chs 366: phtree_compare(struct pool_item_header *a, struct pool_item_header *b)
367: {
1.121 yamt 368:
369: /*
370: * we consider pool_item_header with smaller ph_page bigger.
371: * (this unnatural ordering is for the benefit of pr_find_pagehead.)
372: */
373:
1.88 chs 374: if (a->ph_page < b->ph_page)
1.121 yamt 375: return (1);
376: else if (a->ph_page > b->ph_page)
1.88 chs 377: return (-1);
378: else
379: return (0);
380: }
381:
382: SPLAY_PROTOTYPE(phtree, pool_item_header, ph_node, phtree_compare);
383: SPLAY_GENERATE(phtree, pool_item_header, ph_node, phtree_compare);
384:
1.3 pk 385: /*
1.121 yamt 386: * Return the pool page header based on item address.
1.3 pk 387: */
1.110 perry 388: static inline struct pool_item_header *
1.121 yamt 389: pr_find_pagehead(struct pool *pp, void *v)
1.3 pk 390: {
1.88 chs 391: struct pool_item_header *ph, tmp;
1.3 pk 392:
1.121 yamt 393: if ((pp->pr_roflags & PR_NOALIGN) != 0) {
1.128 christos 394: tmp.ph_page = (void *)(uintptr_t)v;
1.121 yamt 395: ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
396: if (ph == NULL) {
397: ph = SPLAY_ROOT(&pp->pr_phtree);
398: if (ph != NULL && phtree_compare(&tmp, ph) >= 0) {
399: ph = SPLAY_NEXT(phtree, &pp->pr_phtree, ph);
400: }
401: KASSERT(ph == NULL || phtree_compare(&tmp, ph) < 0);
402: }
403: } else {
1.128 christos 404: void *page =
405: (void *)((uintptr_t)v & pp->pr_alloc->pa_pagemask);
1.121 yamt 406:
407: if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
1.128 christos 408: ph = (struct pool_item_header *)((char *)page + pp->pr_phoffset);
1.121 yamt 409: } else {
410: tmp.ph_page = page;
411: ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
412: }
413: }
1.3 pk 414:
1.121 yamt 415: KASSERT(ph == NULL || ((pp->pr_roflags & PR_PHINPAGE) != 0) ||
1.128 christos 416: ((char *)ph->ph_page <= (char *)v &&
417: (char *)v < (char *)ph->ph_page + pp->pr_alloc->pa_pagesz));
1.88 chs 418: return ph;
1.3 pk 419: }
420:
1.101 thorpej 421: static void
422: pr_pagelist_free(struct pool *pp, struct pool_pagelist *pq)
423: {
424: struct pool_item_header *ph;
425:
426: while ((ph = LIST_FIRST(pq)) != NULL) {
427: LIST_REMOVE(ph, ph_pagelist);
428: pool_allocator_free(pp, ph->ph_page);
1.128.2.2 ad 429: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1.101 thorpej 430: pool_put(pp->pr_phpool, ph);
431: }
432: }
433:
1.3 pk 434: /*
435: * Remove a page from the pool.
436: */
1.110 perry 437: static inline void
1.61 chs 438: pr_rmpage(struct pool *pp, struct pool_item_header *ph,
439: struct pool_pagelist *pq)
1.3 pk 440: {
441:
1.128.2.2 ad 442: KASSERT(mutex_owned(&pp->pr_lock));
1.91 yamt 443:
1.3 pk 444: /*
1.7 thorpej 445: * If the page was idle, decrement the idle page count.
1.3 pk 446: */
1.6 thorpej 447: if (ph->ph_nmissing == 0) {
448: #ifdef DIAGNOSTIC
449: if (pp->pr_nidle == 0)
450: panic("pr_rmpage: nidle inconsistent");
1.20 thorpej 451: if (pp->pr_nitems < pp->pr_itemsperpage)
452: panic("pr_rmpage: nitems inconsistent");
1.6 thorpej 453: #endif
454: pp->pr_nidle--;
455: }
1.7 thorpej 456:
1.20 thorpej 457: pp->pr_nitems -= pp->pr_itemsperpage;
458:
1.7 thorpej 459: /*
1.101 thorpej 460: * Unlink the page from the pool and queue it for release.
1.7 thorpej 461: */
1.88 chs 462: LIST_REMOVE(ph, ph_pagelist);
1.91 yamt 463: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
464: SPLAY_REMOVE(phtree, &pp->pr_phtree, ph);
1.101 thorpej 465: LIST_INSERT_HEAD(pq, ph, ph_pagelist);
466:
1.7 thorpej 467: pp->pr_npages--;
468: pp->pr_npagefree++;
1.6 thorpej 469:
1.88 chs 470: pool_update_curpage(pp);
1.3 pk 471: }
472:
1.126 thorpej 473: static bool
1.117 yamt 474: pa_starved_p(struct pool_allocator *pa)
475: {
476:
477: if (pa->pa_backingmap != NULL) {
478: return vm_map_starved_p(pa->pa_backingmap);
479: }
1.127 thorpej 480: return false;
1.117 yamt 481: }
482:
483: static int
1.124 yamt 484: pool_reclaim_callback(struct callback_entry *ce, void *obj, void *arg)
1.117 yamt 485: {
486: struct pool *pp = obj;
487: struct pool_allocator *pa = pp->pr_alloc;
488:
489: KASSERT(&pp->pr_reclaimerentry == ce);
490: pool_reclaim(pp);
491: if (!pa_starved_p(pa)) {
492: return CALLBACK_CHAIN_ABORT;
493: }
494: return CALLBACK_CHAIN_CONTINUE;
495: }
496:
497: static void
498: pool_reclaim_register(struct pool *pp)
499: {
500: struct vm_map *map = pp->pr_alloc->pa_backingmap;
501: int s;
502:
503: if (map == NULL) {
504: return;
505: }
506:
507: s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */
508: callback_register(&vm_map_to_kernel(map)->vmk_reclaim_callback,
509: &pp->pr_reclaimerentry, pp, pool_reclaim_callback);
510: splx(s);
511: }
512:
513: static void
514: pool_reclaim_unregister(struct pool *pp)
515: {
516: struct vm_map *map = pp->pr_alloc->pa_backingmap;
517: int s;
518:
519: if (map == NULL) {
520: return;
521: }
522:
523: s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */
524: callback_unregister(&vm_map_to_kernel(map)->vmk_reclaim_callback,
525: &pp->pr_reclaimerentry);
526: splx(s);
527: }
528:
529: static void
530: pa_reclaim_register(struct pool_allocator *pa)
531: {
532: struct vm_map *map = *pa->pa_backingmapptr;
533: struct pool *pp;
534:
535: KASSERT(pa->pa_backingmap == NULL);
536: if (map == NULL) {
537: SLIST_INSERT_HEAD(&pa_deferinitq, pa, pa_q);
538: return;
539: }
540: pa->pa_backingmap = map;
541: TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
542: pool_reclaim_register(pp);
543: }
544: }
545:
1.3 pk 546: /*
1.94 simonb 547: * Initialize all the pools listed in the "pools" link set.
548: */
549: void
1.117 yamt 550: pool_subsystem_init(void)
1.94 simonb 551: {
1.117 yamt 552: struct pool_allocator *pa;
1.94 simonb 553: __link_set_decl(pools, struct link_pool_init);
554: struct link_pool_init * const *pi;
555:
1.128.2.2 ad 556: mutex_init(&pool_head_lock, MUTEX_DEFAULT, IPL_NONE);
557:
1.94 simonb 558: __link_set_foreach(pi, pools)
559: pool_init((*pi)->pp, (*pi)->size, (*pi)->align,
560: (*pi)->align_offset, (*pi)->flags, (*pi)->wchan,
1.128.2.1 ad 561: (*pi)->palloc, (*pi)->ipl);
1.117 yamt 562:
563: while ((pa = SLIST_FIRST(&pa_deferinitq)) != NULL) {
564: KASSERT(pa->pa_backingmapptr != NULL);
565: KASSERT(*pa->pa_backingmapptr != NULL);
566: SLIST_REMOVE_HEAD(&pa_deferinitq, pa_q);
567: pa_reclaim_register(pa);
568: }
1.94 simonb 569: }
570:
571: /*
1.3 pk 572: * Initialize the given pool resource structure.
573: *
574: * We export this routine to allow other kernel parts to declare
575: * static pools that must be initialized before malloc() is available.
576: */
577: void
1.42 thorpej 578: pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
1.128.2.1 ad 579: const char *wchan, struct pool_allocator *palloc, int ipl)
1.3 pk 580: {
1.116 simonb 581: #ifdef DEBUG
582: struct pool *pp1;
583: #endif
1.92 enami 584: size_t trysize, phsize;
1.128.2.2 ad 585: int off, slack;
1.3 pk 586:
1.99 yamt 587: KASSERT((1UL << (CHAR_BIT * sizeof(pool_item_freelist_t))) - 2 >=
588: PHPOOL_FREELIST_NELEM(PHPOOL_MAX - 1));
589:
1.116 simonb 590: #ifdef DEBUG
591: /*
592: * Check that the pool hasn't already been initialised and
593: * added to the list of all pools.
594: */
595: LIST_FOREACH(pp1, &pool_head, pr_poollist) {
596: if (pp == pp1)
597: panic("pool_init: pool %s already initialised",
598: wchan);
599: }
600: #endif
601:
1.25 thorpej 602: #ifdef POOL_DIAGNOSTIC
603: /*
604: * Always log if POOL_DIAGNOSTIC is defined.
605: */
606: if (pool_logsize != 0)
607: flags |= PR_LOGGING;
608: #endif
609:
1.66 thorpej 610: if (palloc == NULL)
611: palloc = &pool_allocator_kmem;
1.112 bjh21 612: #ifdef POOL_SUBPAGE
613: if (size > palloc->pa_pagesz) {
614: if (palloc == &pool_allocator_kmem)
615: palloc = &pool_allocator_kmem_fullpage;
616: else if (palloc == &pool_allocator_nointr)
617: palloc = &pool_allocator_nointr_fullpage;
618: }
1.66 thorpej 619: #endif /* POOL_SUBPAGE */
620: if ((palloc->pa_flags & PA_INITIALIZED) == 0) {
1.112 bjh21 621: if (palloc->pa_pagesz == 0)
1.66 thorpej 622: palloc->pa_pagesz = PAGE_SIZE;
623:
624: TAILQ_INIT(&palloc->pa_list);
625:
1.128.2.2 ad 626: mutex_init(&palloc->pa_lock, MUTEX_DRIVER, IPL_VM);
1.66 thorpej 627: palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
628: palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
1.117 yamt 629:
630: if (palloc->pa_backingmapptr != NULL) {
631: pa_reclaim_register(palloc);
632: }
1.66 thorpej 633: palloc->pa_flags |= PA_INITIALIZED;
1.4 thorpej 634: }
1.3 pk 635:
636: if (align == 0)
637: align = ALIGN(1);
1.14 thorpej 638:
1.120 yamt 639: if ((flags & PR_NOTOUCH) == 0 && size < sizeof(struct pool_item))
1.14 thorpej 640: size = sizeof(struct pool_item);
1.3 pk 641:
1.78 thorpej 642: size = roundup(size, align);
1.66 thorpej 643: #ifdef DIAGNOSTIC
644: if (size > palloc->pa_pagesz)
1.121 yamt 645: panic("pool_init: pool item size (%zu) too large", size);
1.66 thorpej 646: #endif
1.35 pk 647:
1.3 pk 648: /*
649: * Initialize the pool structure.
650: */
1.88 chs 651: LIST_INIT(&pp->pr_emptypages);
652: LIST_INIT(&pp->pr_fullpages);
653: LIST_INIT(&pp->pr_partpages);
1.102 chs 654: LIST_INIT(&pp->pr_cachelist);
1.3 pk 655: pp->pr_curpage = NULL;
656: pp->pr_npages = 0;
657: pp->pr_minitems = 0;
658: pp->pr_minpages = 0;
659: pp->pr_maxpages = UINT_MAX;
1.20 thorpej 660: pp->pr_roflags = flags;
661: pp->pr_flags = 0;
1.35 pk 662: pp->pr_size = size;
1.3 pk 663: pp->pr_align = align;
664: pp->pr_wchan = wchan;
1.66 thorpej 665: pp->pr_alloc = palloc;
1.20 thorpej 666: pp->pr_nitems = 0;
667: pp->pr_nout = 0;
668: pp->pr_hardlimit = UINT_MAX;
669: pp->pr_hardlimit_warning = NULL;
1.31 thorpej 670: pp->pr_hardlimit_ratecap.tv_sec = 0;
671: pp->pr_hardlimit_ratecap.tv_usec = 0;
672: pp->pr_hardlimit_warning_last.tv_sec = 0;
673: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.68 thorpej 674: pp->pr_drain_hook = NULL;
675: pp->pr_drain_hook_arg = NULL;
1.125 ad 676: pp->pr_freecheck = NULL;
1.3 pk 677:
678: /*
679: * Decide whether to put the page header off page to avoid
1.92 enami 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'.
1.3 pk 690: */
1.92 enami 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));
1.121 yamt 695: if ((pp->pr_roflags & (PR_NOTOUCH | PR_NOALIGN)) == 0 &&
1.97 yamt 696: (pp->pr_size < MIN(palloc->pa_pagesz / 16, phsize << 3) ||
697: trysize / pp->pr_size == (trysize - phsize) / pp->pr_size)) {
1.3 pk 698: /* Use the end of the page for the page header */
1.20 thorpej 699: pp->pr_roflags |= PR_PHINPAGE;
1.92 enami 700: pp->pr_phoffset = off = palloc->pa_pagesz - phsize;
1.2 pk 701: } else {
1.3 pk 702: /* The page header will be taken from our page header pool */
703: pp->pr_phoffset = 0;
1.66 thorpej 704: off = palloc->pa_pagesz;
1.88 chs 705: SPLAY_INIT(&pp->pr_phtree);
1.2 pk 706: }
1.1 pk 707:
1.3 pk 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;
1.43 thorpej 714: KASSERT(pp->pr_itemsperpage != 0);
1.97 yamt 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
1.3 pk 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;
1.1 pk 753: pp->pr_hiwat = 0;
1.8 thorpej 754: pp->pr_nidle = 0;
1.3 pk 755:
1.59 thorpej 756: #ifdef POOL_DIAGNOSTIC
1.25 thorpej 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)
1.20 thorpej 761: pp->pr_roflags &= ~PR_LOGGING;
1.3 pk 762: pp->pr_curlogentry = 0;
763: pp->pr_logsize = pool_logsize;
764: }
1.59 thorpej 765: #endif
1.25 thorpej 766:
767: pp->pr_entered_file = NULL;
768: pp->pr_entered_line = 0;
1.3 pk 769:
1.128.2.2 ad 770: mutex_init(&pp->pr_lock, MUTEX_DRIVER, ipl);
771: cv_init(&pp->pr_cv, wchan);
772: pp->pr_ipl = ipl;
773:
1.3 pk 774: /*
1.43 thorpej 775: * Initialize private page header pool and cache magazine pool if we
776: * haven't done so yet.
1.23 thorpej 777: * XXX LOCKING.
1.3 pk 778: */
1.97 yamt 779: if (phpool[0].pr_size == 0) {
780: int idx;
781: for (idx = 0; idx < PHPOOL_MAX; idx++) {
782: static char phpool_names[PHPOOL_MAX][6+1+6+1];
783: int nelem;
784: size_t sz;
785:
786: nelem = PHPOOL_FREELIST_NELEM(idx);
787: snprintf(phpool_names[idx], sizeof(phpool_names[idx]),
788: "phpool-%d", nelem);
789: sz = sizeof(struct pool_item_header);
790: if (nelem) {
791: sz = PR_FREELIST_ALIGN(sz)
1.99 yamt 792: + nelem * sizeof(pool_item_freelist_t);
1.97 yamt 793: }
794: pool_init(&phpool[idx], sz, 0, 0, 0,
1.128.2.1 ad 795: phpool_names[idx], &pool_allocator_meta, IPL_VM);
1.97 yamt 796: }
1.62 bjh21 797: #ifdef POOL_SUBPAGE
798: pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
1.128.2.1 ad 799: PR_RECURSIVE, "psppool", &pool_allocator_meta, IPL_VM);
1.62 bjh21 800: #endif
1.43 thorpej 801: pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
1.128.2.1 ad 802: 0, "pcgpool", &pool_allocator_meta, IPL_VM);
1.1 pk 803: }
804:
1.128.2.2 ad 805: if (__predict_true(!cold)) {
806: /* Insert into the list of all pools. */
807: mutex_enter(&pool_head_lock);
808: LIST_INSERT_HEAD(&pool_head, pp, pr_poollist);
809: mutex_exit(&pool_head_lock);
810:
811: /* Insert this into the list of pools using this allocator. */
812: mutex_enter(&palloc->pa_lock);
813: TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
814: mutex_exit(&palloc->pa_lock);
815: } else {
816: LIST_INSERT_HEAD(&pool_head, pp, pr_poollist);
817: TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
818: }
819:
1.117 yamt 820: pool_reclaim_register(pp);
1.1 pk 821: }
822:
823: /*
824: * De-commision a pool resource.
825: */
826: void
1.42 thorpej 827: pool_destroy(struct pool *pp)
1.1 pk 828: {
1.101 thorpej 829: struct pool_pagelist pq;
1.3 pk 830: struct pool_item_header *ph;
1.43 thorpej 831:
1.101 thorpej 832: /* Remove from global pool list */
1.128.2.2 ad 833: mutex_enter(&pool_head_lock);
1.102 chs 834: LIST_REMOVE(pp, pr_poollist);
1.101 thorpej 835: if (drainpp == pp)
836: drainpp = NULL;
1.128.2.2 ad 837: mutex_exit(&pool_head_lock);
1.101 thorpej 838:
839: /* Remove this pool from its allocator's list of pools. */
1.117 yamt 840: pool_reclaim_unregister(pp);
1.128.2.2 ad 841: mutex_enter(&pp->pr_alloc->pa_lock);
1.66 thorpej 842: TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
1.128.2.2 ad 843: mutex_exit(&pp->pr_alloc->pa_lock);
1.66 thorpej 844:
1.128.2.2 ad 845: mutex_enter(&pp->pr_lock);
1.101 thorpej 846:
1.102 chs 847: KASSERT(LIST_EMPTY(&pp->pr_cachelist));
1.3 pk 848:
849: #ifdef DIAGNOSTIC
1.20 thorpej 850: if (pp->pr_nout != 0) {
1.25 thorpej 851: pr_printlog(pp, NULL, printf);
1.80 provos 852: panic("pool_destroy: pool busy: still out: %u",
1.20 thorpej 853: pp->pr_nout);
1.3 pk 854: }
855: #endif
1.1 pk 856:
1.101 thorpej 857: KASSERT(LIST_EMPTY(&pp->pr_fullpages));
858: KASSERT(LIST_EMPTY(&pp->pr_partpages));
859:
1.3 pk 860: /* Remove all pages */
1.101 thorpej 861: LIST_INIT(&pq);
1.88 chs 862: while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1.101 thorpej 863: pr_rmpage(pp, ph, &pq);
864:
1.128.2.2 ad 865: mutex_exit(&pp->pr_lock);
1.3 pk 866:
1.101 thorpej 867: pr_pagelist_free(pp, &pq);
1.3 pk 868:
1.59 thorpej 869: #ifdef POOL_DIAGNOSTIC
1.20 thorpej 870: if ((pp->pr_roflags & PR_LOGGING) != 0)
1.3 pk 871: free(pp->pr_log, M_TEMP);
1.59 thorpej 872: #endif
1.128.2.2 ad 873:
874: cv_destroy(&pp->pr_cv);
875: mutex_destroy(&pp->pr_lock);
1.1 pk 876: }
877:
1.68 thorpej 878: void
879: pool_set_drain_hook(struct pool *pp, void (*fn)(void *, int), void *arg)
880: {
881:
882: /* XXX no locking -- must be used just after pool_init() */
883: #ifdef DIAGNOSTIC
884: if (pp->pr_drain_hook != NULL)
885: panic("pool_set_drain_hook(%s): already set", pp->pr_wchan);
886: #endif
887: pp->pr_drain_hook = fn;
888: pp->pr_drain_hook_arg = arg;
889: }
890:
1.88 chs 891: static struct pool_item_header *
1.128 christos 892: pool_alloc_item_header(struct pool *pp, void *storage, int flags)
1.55 thorpej 893: {
894: struct pool_item_header *ph;
895:
896: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
1.128 christos 897: ph = (struct pool_item_header *) ((char *)storage + pp->pr_phoffset);
1.128.2.2 ad 898: else
1.97 yamt 899: ph = pool_get(pp->pr_phpool, flags);
1.55 thorpej 900:
901: return (ph);
902: }
1.1 pk 903:
904: /*
1.3 pk 905: * Grab an item from the pool; must be called at appropriate spl level
1.1 pk 906: */
1.3 pk 907: void *
1.59 thorpej 908: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 909: _pool_get(struct pool *pp, int flags, const char *file, long line)
1.56 sommerfe 910: #else
911: pool_get(struct pool *pp, int flags)
912: #endif
1.1 pk 913: {
914: struct pool_item *pi;
1.3 pk 915: struct pool_item_header *ph;
1.55 thorpej 916: void *v;
1.1 pk 917:
1.2 pk 918: #ifdef DIAGNOSTIC
1.95 atatat 919: if (__predict_false(pp->pr_itemsperpage == 0))
920: panic("pool_get: pool %p: pr_itemsperpage is zero, "
921: "pool not initialized?", pp);
1.84 thorpej 922: if (__predict_false(curlwp == NULL && doing_shutdown == 0 &&
1.37 sommerfe 923: (flags & PR_WAITOK) != 0))
1.77 matt 924: panic("pool_get: %s: must have NOWAIT", pp->pr_wchan);
1.58 thorpej 925:
1.102 chs 926: #endif /* DIAGNOSTIC */
1.58 thorpej 927: #ifdef LOCKDEBUG
928: if (flags & PR_WAITOK)
1.119 yamt 929: ASSERT_SLEEPABLE(NULL, "pool_get(PR_WAITOK)");
1.56 sommerfe 930: #endif
1.1 pk 931:
1.128.2.2 ad 932: mutex_enter(&pp->pr_lock);
1.25 thorpej 933: pr_enter(pp, file, line);
1.20 thorpej 934:
935: startover:
936: /*
937: * Check to see if we've reached the hard limit. If we have,
938: * and we can wait, then wait until an item has been returned to
939: * the pool.
940: */
941: #ifdef DIAGNOSTIC
1.34 thorpej 942: if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
1.25 thorpej 943: pr_leave(pp);
1.128.2.2 ad 944: mutex_exit(&pp->pr_lock);
1.20 thorpej 945: panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
946: }
947: #endif
1.34 thorpej 948: if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
1.68 thorpej 949: if (pp->pr_drain_hook != NULL) {
950: /*
951: * Since the drain hook is going to free things
952: * back to the pool, unlock, call the hook, re-lock,
953: * and check the hardlimit condition again.
954: */
955: pr_leave(pp);
1.128.2.2 ad 956: mutex_exit(&pp->pr_lock);
1.68 thorpej 957: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
1.128.2.2 ad 958: mutex_enter(&pp->pr_lock);
1.68 thorpej 959: pr_enter(pp, file, line);
960: if (pp->pr_nout < pp->pr_hardlimit)
961: goto startover;
962: }
963:
1.29 sommerfe 964: if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
1.20 thorpej 965: /*
966: * XXX: A warning isn't logged in this case. Should
967: * it be?
968: */
969: pp->pr_flags |= PR_WANTED;
1.25 thorpej 970: pr_leave(pp);
1.128.2.2 ad 971: cv_wait(&pp->pr_cv, &pp->pr_lock);
1.25 thorpej 972: pr_enter(pp, file, line);
1.20 thorpej 973: goto startover;
974: }
1.31 thorpej 975:
976: /*
977: * Log a message that the hard limit has been hit.
978: */
979: if (pp->pr_hardlimit_warning != NULL &&
980: ratecheck(&pp->pr_hardlimit_warning_last,
981: &pp->pr_hardlimit_ratecap))
982: log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
1.21 thorpej 983:
984: pp->pr_nfail++;
985:
1.25 thorpej 986: pr_leave(pp);
1.128.2.2 ad 987: mutex_exit(&pp->pr_lock);
1.20 thorpej 988: return (NULL);
989: }
990:
1.3 pk 991: /*
992: * The convention we use is that if `curpage' is not NULL, then
993: * it points at a non-empty bucket. In particular, `curpage'
994: * never points at a page header which has PR_PHINPAGE set and
995: * has no items in its bucket.
996: */
1.20 thorpej 997: if ((ph = pp->pr_curpage) == NULL) {
1.113 yamt 998: int error;
999:
1.20 thorpej 1000: #ifdef DIAGNOSTIC
1001: if (pp->pr_nitems != 0) {
1.128.2.2 ad 1002: mutex_exit(&pp->pr_lock);
1.20 thorpej 1003: printf("pool_get: %s: curpage NULL, nitems %u\n",
1004: pp->pr_wchan, pp->pr_nitems);
1.80 provos 1005: panic("pool_get: nitems inconsistent");
1.20 thorpej 1006: }
1007: #endif
1008:
1.21 thorpej 1009: /*
1010: * Call the back-end page allocator for more memory.
1011: * Release the pool lock, as the back-end page allocator
1012: * may block.
1013: */
1.25 thorpej 1014: pr_leave(pp);
1.113 yamt 1015: error = pool_grow(pp, flags);
1016: pr_enter(pp, file, line);
1017: if (error != 0) {
1.21 thorpej 1018: /*
1.55 thorpej 1019: * We were unable to allocate a page or item
1020: * header, but we released the lock during
1021: * allocation, so perhaps items were freed
1022: * back to the pool. Check for this case.
1.21 thorpej 1023: */
1024: if (pp->pr_curpage != NULL)
1025: goto startover;
1.15 pk 1026:
1.117 yamt 1027: pp->pr_nfail++;
1.25 thorpej 1028: pr_leave(pp);
1.128.2.2 ad 1029: mutex_exit(&pp->pr_lock);
1.117 yamt 1030: return (NULL);
1.1 pk 1031: }
1.3 pk 1032:
1.20 thorpej 1033: /* Start the allocation process over. */
1034: goto startover;
1.3 pk 1035: }
1.97 yamt 1036: if (pp->pr_roflags & PR_NOTOUCH) {
1037: #ifdef DIAGNOSTIC
1038: if (__predict_false(ph->ph_nmissing == pp->pr_itemsperpage)) {
1039: pr_leave(pp);
1.128.2.2 ad 1040: mutex_exit(&pp->pr_lock);
1.97 yamt 1041: panic("pool_get: %s: page empty", pp->pr_wchan);
1042: }
1043: #endif
1044: v = pr_item_notouch_get(pp, ph);
1045: #ifdef POOL_DIAGNOSTIC
1046: pr_log(pp, v, PRLOG_GET, file, line);
1047: #endif
1048: } else {
1.102 chs 1049: v = pi = LIST_FIRST(&ph->ph_itemlist);
1.97 yamt 1050: if (__predict_false(v == NULL)) {
1051: pr_leave(pp);
1.128.2.2 ad 1052: mutex_exit(&pp->pr_lock);
1.97 yamt 1053: panic("pool_get: %s: page empty", pp->pr_wchan);
1054: }
1.20 thorpej 1055: #ifdef DIAGNOSTIC
1.97 yamt 1056: if (__predict_false(pp->pr_nitems == 0)) {
1057: pr_leave(pp);
1.128.2.2 ad 1058: mutex_exit(&pp->pr_lock);
1.97 yamt 1059: printf("pool_get: %s: items on itemlist, nitems %u\n",
1060: pp->pr_wchan, pp->pr_nitems);
1061: panic("pool_get: nitems inconsistent");
1062: }
1.65 enami 1063: #endif
1.56 sommerfe 1064:
1.65 enami 1065: #ifdef POOL_DIAGNOSTIC
1.97 yamt 1066: pr_log(pp, v, PRLOG_GET, file, line);
1.65 enami 1067: #endif
1.3 pk 1068:
1.65 enami 1069: #ifdef DIAGNOSTIC
1.97 yamt 1070: if (__predict_false(pi->pi_magic != PI_MAGIC)) {
1071: pr_printlog(pp, pi, printf);
1072: panic("pool_get(%s): free list modified: "
1073: "magic=%x; page %p; item addr %p\n",
1074: pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
1075: }
1.3 pk 1076: #endif
1077:
1.97 yamt 1078: /*
1079: * Remove from item list.
1080: */
1.102 chs 1081: LIST_REMOVE(pi, pi_list);
1.97 yamt 1082: }
1.20 thorpej 1083: pp->pr_nitems--;
1084: pp->pr_nout++;
1.6 thorpej 1085: if (ph->ph_nmissing == 0) {
1086: #ifdef DIAGNOSTIC
1.34 thorpej 1087: if (__predict_false(pp->pr_nidle == 0))
1.6 thorpej 1088: panic("pool_get: nidle inconsistent");
1089: #endif
1090: pp->pr_nidle--;
1.88 chs 1091:
1092: /*
1093: * This page was previously empty. Move it to the list of
1094: * partially-full pages. This page is already curpage.
1095: */
1096: LIST_REMOVE(ph, ph_pagelist);
1097: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.6 thorpej 1098: }
1.3 pk 1099: ph->ph_nmissing++;
1.97 yamt 1100: if (ph->ph_nmissing == pp->pr_itemsperpage) {
1.21 thorpej 1101: #ifdef DIAGNOSTIC
1.97 yamt 1102: if (__predict_false((pp->pr_roflags & PR_NOTOUCH) == 0 &&
1.102 chs 1103: !LIST_EMPTY(&ph->ph_itemlist))) {
1.25 thorpej 1104: pr_leave(pp);
1.128.2.2 ad 1105: mutex_exit(&pp->pr_lock);
1.21 thorpej 1106: panic("pool_get: %s: nmissing inconsistent",
1107: pp->pr_wchan);
1108: }
1109: #endif
1.3 pk 1110: /*
1.88 chs 1111: * This page is now full. Move it to the full list
1112: * and select a new current page.
1.3 pk 1113: */
1.88 chs 1114: LIST_REMOVE(ph, ph_pagelist);
1115: LIST_INSERT_HEAD(&pp->pr_fullpages, ph, ph_pagelist);
1116: pool_update_curpage(pp);
1.1 pk 1117: }
1.3 pk 1118:
1119: pp->pr_nget++;
1.111 christos 1120: pr_leave(pp);
1.20 thorpej 1121:
1122: /*
1123: * If we have a low water mark and we are now below that low
1124: * water mark, add more items to the pool.
1125: */
1.53 thorpej 1126: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1127: /*
1128: * XXX: Should we log a warning? Should we set up a timeout
1129: * to try again in a second or so? The latter could break
1130: * a caller's assumptions about interrupt protection, etc.
1131: */
1132: }
1133:
1.128.2.2 ad 1134: mutex_exit(&pp->pr_lock);
1.125 ad 1135: KASSERT((((vaddr_t)v + pp->pr_itemoffset) & (pp->pr_align - 1)) == 0);
1136: FREECHECK_OUT(&pp->pr_freecheck, v);
1.1 pk 1137: return (v);
1138: }
1139:
1140: /*
1.43 thorpej 1141: * Internal version of pool_put(). Pool is already locked/entered.
1.1 pk 1142: */
1.43 thorpej 1143: static void
1.101 thorpej 1144: pool_do_put(struct pool *pp, void *v, struct pool_pagelist *pq)
1.1 pk 1145: {
1146: struct pool_item *pi = v;
1.3 pk 1147: struct pool_item_header *ph;
1148:
1.128.2.2 ad 1149: KASSERT(mutex_owned(&pp->pr_lock));
1.125 ad 1150: FREECHECK_IN(&pp->pr_freecheck, v);
1.61 chs 1151:
1.30 thorpej 1152: #ifdef DIAGNOSTIC
1.34 thorpej 1153: if (__predict_false(pp->pr_nout == 0)) {
1.30 thorpej 1154: printf("pool %s: putting with none out\n",
1155: pp->pr_wchan);
1156: panic("pool_put");
1157: }
1158: #endif
1.3 pk 1159:
1.121 yamt 1160: if (__predict_false((ph = pr_find_pagehead(pp, v)) == NULL)) {
1.25 thorpej 1161: pr_printlog(pp, NULL, printf);
1.3 pk 1162: panic("pool_put: %s: page header missing", pp->pr_wchan);
1163: }
1.28 thorpej 1164:
1.3 pk 1165: /*
1166: * Return to item list.
1167: */
1.97 yamt 1168: if (pp->pr_roflags & PR_NOTOUCH) {
1169: pr_item_notouch_put(pp, ph, v);
1170: } else {
1.2 pk 1171: #ifdef DIAGNOSTIC
1.97 yamt 1172: pi->pi_magic = PI_MAGIC;
1.3 pk 1173: #endif
1.32 chs 1174: #ifdef DEBUG
1.97 yamt 1175: {
1176: int i, *ip = v;
1.32 chs 1177:
1.97 yamt 1178: for (i = 0; i < pp->pr_size / sizeof(int); i++) {
1179: *ip++ = PI_MAGIC;
1180: }
1.32 chs 1181: }
1182: #endif
1183:
1.102 chs 1184: LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.97 yamt 1185: }
1.79 thorpej 1186: KDASSERT(ph->ph_nmissing != 0);
1.3 pk 1187: ph->ph_nmissing--;
1188: pp->pr_nput++;
1.20 thorpej 1189: pp->pr_nitems++;
1190: pp->pr_nout--;
1.3 pk 1191:
1192: /* Cancel "pool empty" condition if it exists */
1193: if (pp->pr_curpage == NULL)
1194: pp->pr_curpage = ph;
1195:
1196: if (pp->pr_flags & PR_WANTED) {
1197: pp->pr_flags &= ~PR_WANTED;
1.15 pk 1198: if (ph->ph_nmissing == 0)
1199: pp->pr_nidle++;
1.128.2.4! ad 1200: cv_broadcast(&pp->pr_cv);
1.3 pk 1201: return;
1202: }
1203:
1204: /*
1.88 chs 1205: * If this page is now empty, do one of two things:
1.21 thorpej 1206: *
1.88 chs 1207: * (1) If we have more pages than the page high water mark,
1.96 thorpej 1208: * free the page back to the system. ONLY CONSIDER
1.90 thorpej 1209: * FREEING BACK A PAGE IF WE HAVE MORE THAN OUR MINIMUM PAGE
1210: * CLAIM.
1.21 thorpej 1211: *
1.88 chs 1212: * (2) Otherwise, move the page to the empty page list.
1213: *
1214: * Either way, select a new current page (so we use a partially-full
1215: * page if one is available).
1.3 pk 1216: */
1217: if (ph->ph_nmissing == 0) {
1.6 thorpej 1218: pp->pr_nidle++;
1.90 thorpej 1219: if (pp->pr_npages > pp->pr_minpages &&
1220: (pp->pr_npages > pp->pr_maxpages ||
1.117 yamt 1221: pa_starved_p(pp->pr_alloc))) {
1.101 thorpej 1222: pr_rmpage(pp, ph, pq);
1.3 pk 1223: } else {
1.88 chs 1224: LIST_REMOVE(ph, ph_pagelist);
1225: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.3 pk 1226:
1.21 thorpej 1227: /*
1228: * Update the timestamp on the page. A page must
1229: * be idle for some period of time before it can
1230: * be reclaimed by the pagedaemon. This minimizes
1231: * ping-pong'ing for memory.
1232: */
1.118 kardel 1233: getmicrotime(&ph->ph_time);
1.1 pk 1234: }
1.88 chs 1235: pool_update_curpage(pp);
1.1 pk 1236: }
1.88 chs 1237:
1.21 thorpej 1238: /*
1.88 chs 1239: * If the page was previously completely full, move it to the
1240: * partially-full list and make it the current page. The next
1241: * allocation will get the item from this page, instead of
1242: * further fragmenting the pool.
1.21 thorpej 1243: */
1244: else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
1.88 chs 1245: LIST_REMOVE(ph, ph_pagelist);
1246: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.21 thorpej 1247: pp->pr_curpage = ph;
1248: }
1.43 thorpej 1249: }
1250:
1251: /*
1252: * Return resource to the pool; must be called at appropriate spl level
1253: */
1.59 thorpej 1254: #ifdef POOL_DIAGNOSTIC
1.43 thorpej 1255: void
1256: _pool_put(struct pool *pp, void *v, const char *file, long line)
1257: {
1.101 thorpej 1258: struct pool_pagelist pq;
1259:
1260: LIST_INIT(&pq);
1.43 thorpej 1261:
1.128.2.2 ad 1262: mutex_enter(&pp->pr_lock);
1.43 thorpej 1263: pr_enter(pp, file, line);
1264:
1.56 sommerfe 1265: pr_log(pp, v, PRLOG_PUT, file, line);
1266:
1.101 thorpej 1267: pool_do_put(pp, v, &pq);
1.21 thorpej 1268:
1.25 thorpej 1269: pr_leave(pp);
1.128.2.2 ad 1270: mutex_exit(&pp->pr_lock);
1.101 thorpej 1271:
1.102 chs 1272: pr_pagelist_free(pp, &pq);
1.1 pk 1273: }
1.57 sommerfe 1274: #undef pool_put
1.59 thorpej 1275: #endif /* POOL_DIAGNOSTIC */
1.1 pk 1276:
1.56 sommerfe 1277: void
1278: pool_put(struct pool *pp, void *v)
1279: {
1.101 thorpej 1280: struct pool_pagelist pq;
1281:
1282: LIST_INIT(&pq);
1.56 sommerfe 1283:
1.128.2.2 ad 1284: mutex_enter(&pp->pr_lock);
1.101 thorpej 1285: pool_do_put(pp, v, &pq);
1.128.2.2 ad 1286: mutex_exit(&pp->pr_lock);
1.56 sommerfe 1287:
1.102 chs 1288: pr_pagelist_free(pp, &pq);
1.56 sommerfe 1289: }
1.57 sommerfe 1290:
1.59 thorpej 1291: #ifdef POOL_DIAGNOSTIC
1.57 sommerfe 1292: #define pool_put(h, v) _pool_put((h), (v), __FILE__, __LINE__)
1.56 sommerfe 1293: #endif
1.74 thorpej 1294:
1295: /*
1.113 yamt 1296: * pool_grow: grow a pool by a page.
1297: *
1298: * => called with pool locked.
1299: * => unlock and relock the pool.
1300: * => return with pool locked.
1301: */
1302:
1303: static int
1304: pool_grow(struct pool *pp, int flags)
1305: {
1306: struct pool_item_header *ph = NULL;
1307: char *cp;
1308:
1.128.2.2 ad 1309: mutex_exit(&pp->pr_lock);
1.113 yamt 1310: cp = pool_allocator_alloc(pp, flags);
1311: if (__predict_true(cp != NULL)) {
1312: ph = pool_alloc_item_header(pp, cp, flags);
1313: }
1314: if (__predict_false(cp == NULL || ph == NULL)) {
1315: if (cp != NULL) {
1316: pool_allocator_free(pp, cp);
1317: }
1.128.2.2 ad 1318: mutex_enter(&pp->pr_lock);
1.113 yamt 1319: return ENOMEM;
1320: }
1321:
1.128.2.2 ad 1322: mutex_enter(&pp->pr_lock);
1.113 yamt 1323: pool_prime_page(pp, cp, ph);
1324: pp->pr_npagealloc++;
1325: return 0;
1326: }
1327:
1328: /*
1.74 thorpej 1329: * Add N items to the pool.
1330: */
1331: int
1332: pool_prime(struct pool *pp, int n)
1333: {
1.75 simonb 1334: int newpages;
1.113 yamt 1335: int error = 0;
1.74 thorpej 1336:
1.128.2.2 ad 1337: mutex_enter(&pp->pr_lock);
1.74 thorpej 1338:
1339: newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1340:
1341: while (newpages-- > 0) {
1.113 yamt 1342: error = pool_grow(pp, PR_NOWAIT);
1343: if (error) {
1.74 thorpej 1344: break;
1345: }
1346: pp->pr_minpages++;
1347: }
1348:
1349: if (pp->pr_minpages >= pp->pr_maxpages)
1350: pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */
1351:
1.128.2.2 ad 1352: mutex_exit(&pp->pr_lock);
1.113 yamt 1353: return error;
1.74 thorpej 1354: }
1.55 thorpej 1355:
1356: /*
1.3 pk 1357: * Add a page worth of items to the pool.
1.21 thorpej 1358: *
1359: * Note, we must be called with the pool descriptor LOCKED.
1.3 pk 1360: */
1.55 thorpej 1361: static void
1.128 christos 1362: pool_prime_page(struct pool *pp, void *storage, struct pool_item_header *ph)
1.3 pk 1363: {
1364: struct pool_item *pi;
1.128 christos 1365: void *cp = storage;
1.125 ad 1366: const unsigned int align = pp->pr_align;
1367: const unsigned int ioff = pp->pr_itemoffset;
1.55 thorpej 1368: int n;
1.36 pk 1369:
1.128.2.2 ad 1370: KASSERT(mutex_owned(&pp->pr_lock));
1.91 yamt 1371:
1.66 thorpej 1372: #ifdef DIAGNOSTIC
1.121 yamt 1373: if ((pp->pr_roflags & PR_NOALIGN) == 0 &&
1374: ((uintptr_t)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
1.36 pk 1375: panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1.66 thorpej 1376: #endif
1.3 pk 1377:
1378: /*
1379: * Insert page header.
1380: */
1.88 chs 1381: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.102 chs 1382: LIST_INIT(&ph->ph_itemlist);
1.3 pk 1383: ph->ph_page = storage;
1384: ph->ph_nmissing = 0;
1.118 kardel 1385: getmicrotime(&ph->ph_time);
1.88 chs 1386: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1387: SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
1.3 pk 1388:
1.6 thorpej 1389: pp->pr_nidle++;
1390:
1.3 pk 1391: /*
1392: * Color this page.
1393: */
1.128 christos 1394: cp = (char *)cp + pp->pr_curcolor;
1.3 pk 1395: if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
1396: pp->pr_curcolor = 0;
1397:
1398: /*
1399: * Adjust storage to apply aligment to `pr_itemoffset' in each item.
1400: */
1401: if (ioff != 0)
1.128 christos 1402: cp = (char *)cp + align - ioff;
1.3 pk 1403:
1.125 ad 1404: KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
1405:
1.3 pk 1406: /*
1407: * Insert remaining chunks on the bucket list.
1408: */
1409: n = pp->pr_itemsperpage;
1.20 thorpej 1410: pp->pr_nitems += n;
1.3 pk 1411:
1.97 yamt 1412: if (pp->pr_roflags & PR_NOTOUCH) {
1.99 yamt 1413: pool_item_freelist_t *freelist = PR_FREELIST(ph);
1.97 yamt 1414: int i;
1415:
1.128 christos 1416: ph->ph_off = (char *)cp - (char *)storage;
1.97 yamt 1417: ph->ph_firstfree = 0;
1418: for (i = 0; i < n - 1; i++)
1419: freelist[i] = i + 1;
1420: freelist[n - 1] = PR_INDEX_EOL;
1421: } else {
1422: while (n--) {
1423: pi = (struct pool_item *)cp;
1.78 thorpej 1424:
1.97 yamt 1425: KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
1.3 pk 1426:
1.97 yamt 1427: /* Insert on page list */
1.102 chs 1428: LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.3 pk 1429: #ifdef DIAGNOSTIC
1.97 yamt 1430: pi->pi_magic = PI_MAGIC;
1.3 pk 1431: #endif
1.128 christos 1432: cp = (char *)cp + pp->pr_size;
1.125 ad 1433:
1434: KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
1.97 yamt 1435: }
1.3 pk 1436: }
1437:
1438: /*
1439: * If the pool was depleted, point at the new page.
1440: */
1441: if (pp->pr_curpage == NULL)
1442: pp->pr_curpage = ph;
1443:
1444: if (++pp->pr_npages > pp->pr_hiwat)
1445: pp->pr_hiwat = pp->pr_npages;
1446: }
1447:
1.20 thorpej 1448: /*
1.52 thorpej 1449: * Used by pool_get() when nitems drops below the low water mark. This
1.88 chs 1450: * is used to catch up pr_nitems with the low water mark.
1.20 thorpej 1451: *
1.21 thorpej 1452: * Note 1, we never wait for memory here, we let the caller decide what to do.
1.20 thorpej 1453: *
1.73 thorpej 1454: * Note 2, we must be called with the pool already locked, and we return
1.20 thorpej 1455: * with it locked.
1456: */
1457: static int
1.42 thorpej 1458: pool_catchup(struct pool *pp)
1.20 thorpej 1459: {
1460: int error = 0;
1461:
1.54 thorpej 1462: while (POOL_NEEDS_CATCHUP(pp)) {
1.113 yamt 1463: error = pool_grow(pp, PR_NOWAIT);
1464: if (error) {
1.20 thorpej 1465: break;
1466: }
1467: }
1.113 yamt 1468: return error;
1.20 thorpej 1469: }
1470:
1.88 chs 1471: static void
1472: pool_update_curpage(struct pool *pp)
1473: {
1474:
1475: pp->pr_curpage = LIST_FIRST(&pp->pr_partpages);
1476: if (pp->pr_curpage == NULL) {
1477: pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages);
1478: }
1479: }
1480:
1.3 pk 1481: void
1.42 thorpej 1482: pool_setlowat(struct pool *pp, int n)
1.3 pk 1483: {
1.15 pk 1484:
1.128.2.2 ad 1485: mutex_enter(&pp->pr_lock);
1.21 thorpej 1486:
1.3 pk 1487: pp->pr_minitems = n;
1.15 pk 1488: pp->pr_minpages = (n == 0)
1489: ? 0
1.18 thorpej 1490: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.20 thorpej 1491:
1492: /* Make sure we're caught up with the newly-set low water mark. */
1.75 simonb 1493: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1494: /*
1495: * XXX: Should we log a warning? Should we set up a timeout
1496: * to try again in a second or so? The latter could break
1497: * a caller's assumptions about interrupt protection, etc.
1498: */
1499: }
1.21 thorpej 1500:
1.128.2.2 ad 1501: mutex_exit(&pp->pr_lock);
1.3 pk 1502: }
1503:
1504: void
1.42 thorpej 1505: pool_sethiwat(struct pool *pp, int n)
1.3 pk 1506: {
1.15 pk 1507:
1.128.2.2 ad 1508: mutex_enter(&pp->pr_lock);
1.21 thorpej 1509:
1.15 pk 1510: pp->pr_maxpages = (n == 0)
1511: ? 0
1.18 thorpej 1512: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1513:
1.128.2.2 ad 1514: mutex_exit(&pp->pr_lock);
1.3 pk 1515: }
1516:
1.20 thorpej 1517: void
1.42 thorpej 1518: pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1.20 thorpej 1519: {
1520:
1.128.2.2 ad 1521: mutex_enter(&pp->pr_lock);
1.20 thorpej 1522:
1523: pp->pr_hardlimit = n;
1524: pp->pr_hardlimit_warning = warnmess;
1.31 thorpej 1525: pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1526: pp->pr_hardlimit_warning_last.tv_sec = 0;
1527: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.20 thorpej 1528:
1529: /*
1.21 thorpej 1530: * In-line version of pool_sethiwat(), because we don't want to
1531: * release the lock.
1.20 thorpej 1532: */
1533: pp->pr_maxpages = (n == 0)
1534: ? 0
1535: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1536:
1.128.2.2 ad 1537: mutex_exit(&pp->pr_lock);
1.20 thorpej 1538: }
1.3 pk 1539:
1540: /*
1541: * Release all complete pages that have not been used recently.
1542: */
1.66 thorpej 1543: int
1.59 thorpej 1544: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 1545: _pool_reclaim(struct pool *pp, const char *file, long line)
1.56 sommerfe 1546: #else
1547: pool_reclaim(struct pool *pp)
1548: #endif
1.3 pk 1549: {
1550: struct pool_item_header *ph, *phnext;
1.43 thorpej 1551: struct pool_cache *pc;
1.61 chs 1552: struct pool_pagelist pq;
1.102 chs 1553: struct pool_cache_grouplist pcgl;
1554: struct timeval curtime, diff;
1.3 pk 1555:
1.68 thorpej 1556: if (pp->pr_drain_hook != NULL) {
1557: /*
1558: * The drain hook must be called with the pool unlocked.
1559: */
1560: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT);
1561: }
1562:
1.128.2.2 ad 1563: if (mutex_tryenter(&pp->pr_lock) == 0)
1.66 thorpej 1564: return (0);
1.25 thorpej 1565: pr_enter(pp, file, line);
1.68 thorpej 1566:
1.88 chs 1567: LIST_INIT(&pq);
1.102 chs 1568: LIST_INIT(&pcgl);
1.3 pk 1569:
1.43 thorpej 1570: /*
1571: * Reclaim items from the pool's caches.
1572: */
1.102 chs 1573: LIST_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
1574: pool_cache_reclaim(pc, &pq, &pcgl);
1.43 thorpej 1575:
1.118 kardel 1576: getmicrotime(&curtime);
1.21 thorpej 1577:
1.88 chs 1578: for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
1579: phnext = LIST_NEXT(ph, ph_pagelist);
1.3 pk 1580:
1581: /* Check our minimum page claim */
1582: if (pp->pr_npages <= pp->pr_minpages)
1583: break;
1584:
1.88 chs 1585: KASSERT(ph->ph_nmissing == 0);
1586: timersub(&curtime, &ph->ph_time, &diff);
1.117 yamt 1587: if (diff.tv_sec < pool_inactive_time
1588: && !pa_starved_p(pp->pr_alloc))
1.88 chs 1589: continue;
1.21 thorpej 1590:
1.88 chs 1591: /*
1592: * If freeing this page would put us below
1593: * the low water mark, stop now.
1594: */
1595: if ((pp->pr_nitems - pp->pr_itemsperpage) <
1596: pp->pr_minitems)
1597: break;
1.21 thorpej 1598:
1.88 chs 1599: pr_rmpage(pp, ph, &pq);
1.3 pk 1600: }
1601:
1.25 thorpej 1602: pr_leave(pp);
1.128.2.2 ad 1603: mutex_exit(&pp->pr_lock);
1.102 chs 1604: if (LIST_EMPTY(&pq) && LIST_EMPTY(&pcgl))
1605: return 0;
1.66 thorpej 1606:
1.101 thorpej 1607: pr_pagelist_free(pp, &pq);
1.102 chs 1608: pcg_grouplist_free(&pcgl);
1.66 thorpej 1609: return (1);
1.3 pk 1610: }
1611:
1612: /*
1613: * Drain pools, one at a time.
1.21 thorpej 1614: *
1615: * Note, we must never be called from an interrupt context.
1.3 pk 1616: */
1617: void
1.124 yamt 1618: pool_drain(void *arg)
1.3 pk 1619: {
1620: struct pool *pp;
1.23 thorpej 1621: int s;
1.3 pk 1622:
1.61 chs 1623: pp = NULL;
1.128.2.2 ad 1624: s = splvm(); /* XXX why? */
1625: mutex_enter(&pool_head_lock);
1.61 chs 1626: if (drainpp == NULL) {
1.102 chs 1627: drainpp = LIST_FIRST(&pool_head);
1.61 chs 1628: }
1629: if (drainpp) {
1630: pp = drainpp;
1.102 chs 1631: drainpp = LIST_NEXT(pp, pr_poollist);
1.61 chs 1632: }
1.128.2.2 ad 1633: mutex_exit(&pool_head_lock);
1.115 christos 1634: if (pp)
1635: pool_reclaim(pp);
1.61 chs 1636: splx(s);
1.3 pk 1637: }
1638:
1639: /*
1640: * Diagnostic helpers.
1641: */
1642: void
1.42 thorpej 1643: pool_print(struct pool *pp, const char *modif)
1.21 thorpej 1644: {
1645:
1.128.2.2 ad 1646: if (mutex_tryenter(&pp->pr_lock) == 0) {
1.25 thorpej 1647: printf("pool %s is locked; try again later\n",
1648: pp->pr_wchan);
1649: return;
1650: }
1651: pool_print1(pp, modif, printf);
1.128.2.2 ad 1652: mutex_exit(&pp->pr_lock);
1.21 thorpej 1653: }
1654:
1.25 thorpej 1655: void
1.108 yamt 1656: pool_printall(const char *modif, void (*pr)(const char *, ...))
1657: {
1658: struct pool *pp;
1659:
1.128.2.2 ad 1660: if (mutex_tryenter(&pool_head_lock) == 0) {
1.108 yamt 1661: (*pr)("WARNING: pool_head_slock is locked\n");
1662: } else {
1.128.2.2 ad 1663: mutex_exit(&pool_head_lock);
1.108 yamt 1664: }
1665:
1666: LIST_FOREACH(pp, &pool_head, pr_poollist) {
1667: pool_printit(pp, modif, pr);
1668: }
1669: }
1670:
1671: void
1.42 thorpej 1672: pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.25 thorpej 1673: {
1674:
1675: if (pp == NULL) {
1676: (*pr)("Must specify a pool to print.\n");
1677: return;
1678: }
1679:
1680: /*
1681: * Called from DDB; interrupts should be blocked, and all
1682: * other processors should be paused. We can skip locking
1683: * the pool in this case.
1684: *
1.128.2.2 ad 1685: * We do a mutex_tryenter() just to print the lock
1.25 thorpej 1686: * status, however.
1687: */
1688:
1.128.2.2 ad 1689: if (mutex_tryenter(&pp->pr_lock) == 0)
1.25 thorpej 1690: (*pr)("WARNING: pool %s is locked\n", pp->pr_wchan);
1691: else
1.128.2.2 ad 1692: mutex_exit(&pp->pr_lock);
1.25 thorpej 1693:
1694: pool_print1(pp, modif, pr);
1695: }
1696:
1.21 thorpej 1697: static void
1.124 yamt 1698: pool_print_pagelist(struct pool *pp, struct pool_pagelist *pl,
1.97 yamt 1699: void (*pr)(const char *, ...))
1.88 chs 1700: {
1701: struct pool_item_header *ph;
1702: #ifdef DIAGNOSTIC
1703: struct pool_item *pi;
1704: #endif
1705:
1706: LIST_FOREACH(ph, pl, ph_pagelist) {
1707: (*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n",
1708: ph->ph_page, ph->ph_nmissing,
1709: (u_long)ph->ph_time.tv_sec,
1710: (u_long)ph->ph_time.tv_usec);
1711: #ifdef DIAGNOSTIC
1.97 yamt 1712: if (!(pp->pr_roflags & PR_NOTOUCH)) {
1.102 chs 1713: LIST_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1.97 yamt 1714: if (pi->pi_magic != PI_MAGIC) {
1715: (*pr)("\t\t\titem %p, magic 0x%x\n",
1716: pi, pi->pi_magic);
1717: }
1.88 chs 1718: }
1719: }
1720: #endif
1721: }
1722: }
1723:
1724: static void
1.42 thorpej 1725: pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.3 pk 1726: {
1.25 thorpej 1727: struct pool_item_header *ph;
1.44 thorpej 1728: struct pool_cache *pc;
1729: struct pool_cache_group *pcg;
1730: int i, print_log = 0, print_pagelist = 0, print_cache = 0;
1.25 thorpej 1731: char c;
1732:
1733: while ((c = *modif++) != '\0') {
1734: if (c == 'l')
1735: print_log = 1;
1736: if (c == 'p')
1737: print_pagelist = 1;
1.44 thorpej 1738: if (c == 'c')
1739: print_cache = 1;
1.25 thorpej 1740: }
1741:
1742: (*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1743: pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1744: pp->pr_roflags);
1.66 thorpej 1745: (*pr)("\talloc %p\n", pp->pr_alloc);
1.25 thorpej 1746: (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1747: pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1748: (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1749: pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1750:
1751: (*pr)("\n\tnget %lu, nfail %lu, nput %lu\n",
1752: pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1753: (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1754: pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1755:
1756: if (print_pagelist == 0)
1757: goto skip_pagelist;
1758:
1.88 chs 1759: if ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1760: (*pr)("\n\tempty page list:\n");
1.97 yamt 1761: pool_print_pagelist(pp, &pp->pr_emptypages, pr);
1.88 chs 1762: if ((ph = LIST_FIRST(&pp->pr_fullpages)) != NULL)
1763: (*pr)("\n\tfull page list:\n");
1.97 yamt 1764: pool_print_pagelist(pp, &pp->pr_fullpages, pr);
1.88 chs 1765: if ((ph = LIST_FIRST(&pp->pr_partpages)) != NULL)
1766: (*pr)("\n\tpartial-page list:\n");
1.97 yamt 1767: pool_print_pagelist(pp, &pp->pr_partpages, pr);
1.88 chs 1768:
1.25 thorpej 1769: if (pp->pr_curpage == NULL)
1770: (*pr)("\tno current page\n");
1771: else
1772: (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1773:
1774: skip_pagelist:
1775: if (print_log == 0)
1776: goto skip_log;
1777:
1778: (*pr)("\n");
1779: if ((pp->pr_roflags & PR_LOGGING) == 0)
1780: (*pr)("\tno log\n");
1.122 christos 1781: else {
1.25 thorpej 1782: pr_printlog(pp, NULL, pr);
1.122 christos 1783: }
1.3 pk 1784:
1.25 thorpej 1785: skip_log:
1.44 thorpej 1786: if (print_cache == 0)
1787: goto skip_cache;
1788:
1.102 chs 1789: #define PR_GROUPLIST(pcg) \
1790: (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail); \
1791: for (i = 0; i < PCG_NOBJECTS; i++) { \
1792: if (pcg->pcg_objects[i].pcgo_pa != \
1793: POOL_PADDR_INVALID) { \
1794: (*pr)("\t\t\t%p, 0x%llx\n", \
1795: pcg->pcg_objects[i].pcgo_va, \
1796: (unsigned long long) \
1797: pcg->pcg_objects[i].pcgo_pa); \
1798: } else { \
1799: (*pr)("\t\t\t%p\n", \
1800: pcg->pcg_objects[i].pcgo_va); \
1801: } \
1802: }
1803:
1804: LIST_FOREACH(pc, &pp->pr_cachelist, pc_poollist) {
1.103 chs 1805: (*pr)("\tcache %p\n", pc);
1.48 thorpej 1806: (*pr)("\t hits %lu misses %lu ngroups %lu nitems %lu\n",
1807: pc->pc_hits, pc->pc_misses, pc->pc_ngroups, pc->pc_nitems);
1.102 chs 1808: (*pr)("\t full groups:\n");
1.103 chs 1809: LIST_FOREACH(pcg, &pc->pc_fullgroups, pcg_list) {
1.102 chs 1810: PR_GROUPLIST(pcg);
1.103 chs 1811: }
1.102 chs 1812: (*pr)("\t partial groups:\n");
1.103 chs 1813: LIST_FOREACH(pcg, &pc->pc_partgroups, pcg_list) {
1.102 chs 1814: PR_GROUPLIST(pcg);
1.103 chs 1815: }
1.102 chs 1816: (*pr)("\t empty groups:\n");
1.103 chs 1817: LIST_FOREACH(pcg, &pc->pc_emptygroups, pcg_list) {
1.102 chs 1818: PR_GROUPLIST(pcg);
1.103 chs 1819: }
1.44 thorpej 1820: }
1.102 chs 1821: #undef PR_GROUPLIST
1.44 thorpej 1822:
1823: skip_cache:
1.88 chs 1824: pr_enter_check(pp, pr);
1825: }
1826:
1827: static int
1828: pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
1829: {
1830: struct pool_item *pi;
1.128 christos 1831: void *page;
1.88 chs 1832: int n;
1833:
1.121 yamt 1834: if ((pp->pr_roflags & PR_NOALIGN) == 0) {
1.128 christos 1835: page = (void *)((uintptr_t)ph & pp->pr_alloc->pa_pagemask);
1.121 yamt 1836: if (page != ph->ph_page &&
1837: (pp->pr_roflags & PR_PHINPAGE) != 0) {
1838: if (label != NULL)
1839: printf("%s: ", label);
1840: printf("pool(%p:%s): page inconsistency: page %p;"
1841: " at page head addr %p (p %p)\n", pp,
1842: pp->pr_wchan, ph->ph_page,
1843: ph, page);
1844: return 1;
1845: }
1.88 chs 1846: }
1.3 pk 1847:
1.97 yamt 1848: if ((pp->pr_roflags & PR_NOTOUCH) != 0)
1849: return 0;
1850:
1.102 chs 1851: for (pi = LIST_FIRST(&ph->ph_itemlist), n = 0;
1.88 chs 1852: pi != NULL;
1.102 chs 1853: pi = LIST_NEXT(pi,pi_list), n++) {
1.88 chs 1854:
1855: #ifdef DIAGNOSTIC
1856: if (pi->pi_magic != PI_MAGIC) {
1857: if (label != NULL)
1858: printf("%s: ", label);
1859: printf("pool(%s): free list modified: magic=%x;"
1.121 yamt 1860: " page %p; item ordinal %d; addr %p\n",
1.88 chs 1861: pp->pr_wchan, pi->pi_magic, ph->ph_page,
1.121 yamt 1862: n, pi);
1.88 chs 1863: panic("pool");
1864: }
1865: #endif
1.121 yamt 1866: if ((pp->pr_roflags & PR_NOALIGN) != 0) {
1867: continue;
1868: }
1.128 christos 1869: page = (void *)((uintptr_t)pi & pp->pr_alloc->pa_pagemask);
1.88 chs 1870: if (page == ph->ph_page)
1871: continue;
1872:
1873: if (label != NULL)
1874: printf("%s: ", label);
1875: printf("pool(%p:%s): page inconsistency: page %p;"
1876: " item ordinal %d; addr %p (p %p)\n", pp,
1877: pp->pr_wchan, ph->ph_page,
1878: n, pi, page);
1879: return 1;
1880: }
1881: return 0;
1.3 pk 1882: }
1883:
1.88 chs 1884:
1.3 pk 1885: int
1.42 thorpej 1886: pool_chk(struct pool *pp, const char *label)
1.3 pk 1887: {
1888: struct pool_item_header *ph;
1889: int r = 0;
1890:
1.128.2.2 ad 1891: mutex_enter(&pp->pr_lock);
1.88 chs 1892: LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
1893: r = pool_chk_page(pp, label, ph);
1894: if (r) {
1895: goto out;
1896: }
1897: }
1898: LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
1899: r = pool_chk_page(pp, label, ph);
1900: if (r) {
1.3 pk 1901: goto out;
1902: }
1.88 chs 1903: }
1904: LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
1905: r = pool_chk_page(pp, label, ph);
1906: if (r) {
1.3 pk 1907: goto out;
1908: }
1909: }
1.88 chs 1910:
1.3 pk 1911: out:
1.128.2.2 ad 1912: mutex_exit(&pp->pr_lock);
1.3 pk 1913: return (r);
1.43 thorpej 1914: }
1915:
1916: /*
1917: * pool_cache_init:
1918: *
1919: * Initialize a pool cache.
1920: *
1921: * NOTE: If the pool must be protected from interrupts, we expect
1922: * to be called at the appropriate interrupt priority level.
1923: */
1924: void
1925: pool_cache_init(struct pool_cache *pc, struct pool *pp,
1926: int (*ctor)(void *, void *, int),
1927: void (*dtor)(void *, void *),
1928: void *arg)
1929: {
1930:
1.102 chs 1931: LIST_INIT(&pc->pc_emptygroups);
1932: LIST_INIT(&pc->pc_fullgroups);
1933: LIST_INIT(&pc->pc_partgroups);
1.128.2.2 ad 1934: mutex_init(&pc->pc_lock, MUTEX_DRIVER, pp->pr_ipl);
1.43 thorpej 1935:
1936: pc->pc_pool = pp;
1937:
1938: pc->pc_ctor = ctor;
1939: pc->pc_dtor = dtor;
1940: pc->pc_arg = arg;
1941:
1.48 thorpej 1942: pc->pc_hits = 0;
1943: pc->pc_misses = 0;
1944:
1945: pc->pc_ngroups = 0;
1946:
1947: pc->pc_nitems = 0;
1948:
1.128.2.2 ad 1949: if (__predict_true(!cold)) {
1950: mutex_enter(&pp->pr_lock);
1951: LIST_INSERT_HEAD(&pp->pr_cachelist, pc, pc_poollist);
1952: mutex_exit(&pp->pr_lock);
1953: } else
1954: LIST_INSERT_HEAD(&pp->pr_cachelist, pc, pc_poollist);
1.43 thorpej 1955: }
1956:
1957: /*
1958: * pool_cache_destroy:
1959: *
1960: * Destroy a pool cache.
1961: */
1962: void
1963: pool_cache_destroy(struct pool_cache *pc)
1964: {
1965: struct pool *pp = pc->pc_pool;
1966:
1967: /* First, invalidate the entire cache. */
1968: pool_cache_invalidate(pc);
1969:
1970: /* ...and remove it from the pool's cache list. */
1.128.2.2 ad 1971: mutex_enter(&pp->pr_lock);
1.102 chs 1972: LIST_REMOVE(pc, pc_poollist);
1.128.2.2 ad 1973: mutex_exit(&pp->pr_lock);
1974:
1975: mutex_destroy(&pc->pc_lock);
1.43 thorpej 1976: }
1977:
1.110 perry 1978: static inline void *
1.87 thorpej 1979: pcg_get(struct pool_cache_group *pcg, paddr_t *pap)
1.43 thorpej 1980: {
1981: void *object;
1982: u_int idx;
1983:
1984: KASSERT(pcg->pcg_avail <= PCG_NOBJECTS);
1.45 thorpej 1985: KASSERT(pcg->pcg_avail != 0);
1.43 thorpej 1986: idx = --pcg->pcg_avail;
1987:
1.87 thorpej 1988: KASSERT(pcg->pcg_objects[idx].pcgo_va != NULL);
1989: object = pcg->pcg_objects[idx].pcgo_va;
1990: if (pap != NULL)
1991: *pap = pcg->pcg_objects[idx].pcgo_pa;
1992: pcg->pcg_objects[idx].pcgo_va = NULL;
1.43 thorpej 1993:
1994: return (object);
1995: }
1996:
1.110 perry 1997: static inline void
1.87 thorpej 1998: pcg_put(struct pool_cache_group *pcg, void *object, paddr_t pa)
1.43 thorpej 1999: {
2000: u_int idx;
2001:
2002: KASSERT(pcg->pcg_avail < PCG_NOBJECTS);
2003: idx = pcg->pcg_avail++;
2004:
1.87 thorpej 2005: KASSERT(pcg->pcg_objects[idx].pcgo_va == NULL);
2006: pcg->pcg_objects[idx].pcgo_va = object;
2007: pcg->pcg_objects[idx].pcgo_pa = pa;
1.43 thorpej 2008: }
2009:
1.102 chs 2010: static void
2011: pcg_grouplist_free(struct pool_cache_grouplist *pcgl)
2012: {
2013: struct pool_cache_group *pcg;
2014:
2015: while ((pcg = LIST_FIRST(pcgl)) != NULL) {
2016: LIST_REMOVE(pcg, pcg_list);
2017: pool_put(&pcgpool, pcg);
2018: }
2019: }
2020:
1.43 thorpej 2021: /*
1.87 thorpej 2022: * pool_cache_get{,_paddr}:
1.43 thorpej 2023: *
1.87 thorpej 2024: * Get an object from a pool cache (optionally returning
2025: * the physical address of the object).
1.43 thorpej 2026: */
2027: void *
1.87 thorpej 2028: pool_cache_get_paddr(struct pool_cache *pc, int flags, paddr_t *pap)
1.43 thorpej 2029: {
2030: struct pool_cache_group *pcg;
2031: void *object;
1.58 thorpej 2032:
2033: #ifdef LOCKDEBUG
2034: if (flags & PR_WAITOK)
1.119 yamt 2035: ASSERT_SLEEPABLE(NULL, "pool_cache_get(PR_WAITOK)");
1.58 thorpej 2036: #endif
1.43 thorpej 2037:
1.128.2.2 ad 2038: mutex_enter(&pc->pc_lock);
1.43 thorpej 2039:
1.102 chs 2040: pcg = LIST_FIRST(&pc->pc_partgroups);
2041: if (pcg == NULL) {
2042: pcg = LIST_FIRST(&pc->pc_fullgroups);
2043: if (pcg != NULL) {
2044: LIST_REMOVE(pcg, pcg_list);
2045: LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list);
1.43 thorpej 2046: }
1.102 chs 2047: }
2048: if (pcg == NULL) {
1.43 thorpej 2049:
2050: /*
2051: * No groups with any available objects. Allocate
2052: * a new object, construct it, and return it to
2053: * the caller. We will allocate a group, if necessary,
2054: * when the object is freed back to the cache.
2055: */
1.48 thorpej 2056: pc->pc_misses++;
1.128.2.2 ad 2057: mutex_exit(&pc->pc_lock);
1.43 thorpej 2058: object = pool_get(pc->pc_pool, flags);
2059: if (object != NULL && pc->pc_ctor != NULL) {
2060: if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) {
2061: pool_put(pc->pc_pool, object);
2062: return (NULL);
2063: }
2064: }
1.125 ad 2065: KASSERT((((vaddr_t)object + pc->pc_pool->pr_itemoffset) &
2066: (pc->pc_pool->pr_align - 1)) == 0);
1.87 thorpej 2067: if (object != NULL && pap != NULL) {
2068: #ifdef POOL_VTOPHYS
2069: *pap = POOL_VTOPHYS(object);
2070: #else
2071: *pap = POOL_PADDR_INVALID;
2072: #endif
2073: }
1.125 ad 2074:
2075: FREECHECK_OUT(&pc->pc_freecheck, object);
1.43 thorpej 2076: return (object);
2077: }
2078:
1.48 thorpej 2079: pc->pc_hits++;
2080: pc->pc_nitems--;
1.87 thorpej 2081: object = pcg_get(pcg, pap);
1.43 thorpej 2082:
1.102 chs 2083: if (pcg->pcg_avail == 0) {
2084: LIST_REMOVE(pcg, pcg_list);
2085: LIST_INSERT_HEAD(&pc->pc_emptygroups, pcg, pcg_list);
2086: }
1.128.2.2 ad 2087: mutex_exit(&pc->pc_lock);
1.43 thorpej 2088:
1.125 ad 2089: KASSERT((((vaddr_t)object + pc->pc_pool->pr_itemoffset) &
2090: (pc->pc_pool->pr_align - 1)) == 0);
2091: FREECHECK_OUT(&pc->pc_freecheck, object);
1.43 thorpej 2092: return (object);
2093: }
2094:
2095: /*
1.87 thorpej 2096: * pool_cache_put{,_paddr}:
1.43 thorpej 2097: *
1.87 thorpej 2098: * Put an object back to the pool cache (optionally caching the
2099: * physical address of the object).
1.43 thorpej 2100: */
2101: void
1.87 thorpej 2102: pool_cache_put_paddr(struct pool_cache *pc, void *object, paddr_t pa)
1.43 thorpej 2103: {
2104: struct pool_cache_group *pcg;
2105:
1.125 ad 2106: FREECHECK_IN(&pc->pc_freecheck, object);
2107:
1.109 christos 2108: if (__predict_false((pc->pc_pool->pr_flags & PR_WANTED) != 0)) {
2109: goto destruct;
2110: }
2111:
1.128.2.2 ad 2112: mutex_enter(&pc->pc_lock);
1.43 thorpej 2113:
1.102 chs 2114: pcg = LIST_FIRST(&pc->pc_partgroups);
2115: if (pcg == NULL) {
2116: pcg = LIST_FIRST(&pc->pc_emptygroups);
2117: if (pcg != NULL) {
2118: LIST_REMOVE(pcg, pcg_list);
2119: LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list);
1.43 thorpej 2120: }
1.102 chs 2121: }
2122: if (pcg == NULL) {
1.43 thorpej 2123:
2124: /*
2125: * No empty groups to free the object to. Attempt to
1.47 thorpej 2126: * allocate one.
1.43 thorpej 2127: */
1.128.2.2 ad 2128: mutex_exit(&pc->pc_lock);
1.43 thorpej 2129: pcg = pool_get(&pcgpool, PR_NOWAIT);
1.102 chs 2130: if (pcg == NULL) {
1.109 christos 2131: destruct:
1.102 chs 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;
1.43 thorpej 2139: }
1.102 chs 2140: memset(pcg, 0, sizeof(*pcg));
1.128.2.2 ad 2141: mutex_enter(&pc->pc_lock);
1.102 chs 2142: pc->pc_ngroups++;
2143: LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list);
1.43 thorpej 2144: }
2145:
1.48 thorpej 2146: pc->pc_nitems++;
1.87 thorpej 2147: pcg_put(pcg, object, pa);
1.43 thorpej 2148:
1.102 chs 2149: if (pcg->pcg_avail == PCG_NOBJECTS) {
2150: LIST_REMOVE(pcg, pcg_list);
2151: LIST_INSERT_HEAD(&pc->pc_fullgroups, pcg, pcg_list);
2152: }
1.128.2.2 ad 2153: mutex_exit(&pc->pc_lock);
1.51 thorpej 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);
1.43 thorpej 2169: }
2170:
1.102 chs 2171: static void
1.106 christos 2172: pool_do_cache_invalidate_grouplist(struct pool_cache_grouplist *pcgsl,
1.105 christos 2173: struct pool_cache *pc, struct pool_pagelist *pq,
1.106 christos 2174: struct pool_cache_grouplist *pcgdl)
1.102 chs 2175: {
1.106 christos 2176: struct pool_cache_group *pcg, *npcg;
1.102 chs 2177: void *object;
2178:
1.106 christos 2179: for (pcg = LIST_FIRST(pcgsl); pcg != NULL; pcg = npcg) {
1.102 chs 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: }
1.103 chs 2188: pc->pc_ngroups--;
1.102 chs 2189: LIST_REMOVE(pcg, pcg_list);
1.106 christos 2190: LIST_INSERT_HEAD(pcgdl, pcg, pcg_list);
1.102 chs 2191: }
1.105 christos 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:
1.128.2.2 ad 2199: KASSERT(mutex_owned(&pc->pc_lock));
2200: KASSERT(mutex_owned(&pc->pc_pool->pr_lock));
1.105 christos 2201:
1.106 christos 2202: pool_do_cache_invalidate_grouplist(&pc->pc_fullgroups, pc, pq, pcgl);
2203: pool_do_cache_invalidate_grouplist(&pc->pc_partgroups, pc, pq, pcgl);
1.103 chs 2204:
2205: KASSERT(LIST_EMPTY(&pc->pc_partgroups));
2206: KASSERT(LIST_EMPTY(&pc->pc_fullgroups));
2207: KASSERT(pc->pc_nitems == 0);
1.102 chs 2208: }
2209:
1.43 thorpej 2210: /*
1.101 thorpej 2211: * pool_cache_invalidate:
1.43 thorpej 2212: *
1.101 thorpej 2213: * Invalidate a pool cache (destruct and release all of the
2214: * cached objects).
1.43 thorpej 2215: */
1.101 thorpej 2216: void
2217: pool_cache_invalidate(struct pool_cache *pc)
1.43 thorpej 2218: {
1.101 thorpej 2219: struct pool_pagelist pq;
1.102 chs 2220: struct pool_cache_grouplist pcgl;
1.101 thorpej 2221:
2222: LIST_INIT(&pq);
1.102 chs 2223: LIST_INIT(&pcgl);
1.101 thorpej 2224:
1.128.2.2 ad 2225: mutex_enter(&pc->pc_lock);
2226: mutex_enter(&pc->pc_pool->pr_lock);
1.43 thorpej 2227:
1.102 chs 2228: pool_do_cache_invalidate(pc, &pq, &pcgl);
1.43 thorpej 2229:
1.128.2.2 ad 2230: mutex_exit(&pc->pc_pool->pr_lock);
2231: mutex_exit(&pc->pc_lock);
1.43 thorpej 2232:
1.102 chs 2233: pr_pagelist_free(pc->pc_pool, &pq);
2234: pcg_grouplist_free(&pcgl);
1.43 thorpej 2235: }
2236:
2237: /*
2238: * pool_cache_reclaim:
2239: *
2240: * Reclaim a pool cache for pool_reclaim().
2241: */
2242: static void
1.102 chs 2243: pool_cache_reclaim(struct pool_cache *pc, struct pool_pagelist *pq,
2244: struct pool_cache_grouplist *pcgl)
1.43 thorpej 2245: {
1.101 thorpej 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: */
1.128.2.2 ad 2253: if (mutex_tryenter(&pc->pc_lock) == 0)
1.101 thorpej 2254: return;
2255:
1.102 chs 2256: pool_do_cache_invalidate(pc, pq, pcgl);
1.43 thorpej 2257:
1.128.2.2 ad 2258: mutex_exit(&pc->pc_lock);
1.3 pk 2259: }
1.66 thorpej 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:
1.112 bjh21 2277: #ifdef POOL_SUBPAGE
2278: struct pool_allocator pool_allocator_kmem_fullpage = {
2279: pool_page_alloc, pool_page_free, 0,
1.117 yamt 2280: .pa_backingmapptr = &kmem_map,
1.112 bjh21 2281: };
2282: #else
1.66 thorpej 2283: struct pool_allocator pool_allocator_kmem = {
2284: pool_page_alloc, pool_page_free, 0,
1.117 yamt 2285: .pa_backingmapptr = &kmem_map,
1.66 thorpej 2286: };
1.112 bjh21 2287: #endif
1.66 thorpej 2288:
2289: void *pool_page_alloc_nointr(struct pool *, int);
2290: void pool_page_free_nointr(struct pool *, void *);
2291:
1.112 bjh21 2292: #ifdef POOL_SUBPAGE
2293: struct pool_allocator pool_allocator_nointr_fullpage = {
2294: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1.117 yamt 2295: .pa_backingmapptr = &kernel_map,
1.112 bjh21 2296: };
2297: #else
1.66 thorpej 2298: struct pool_allocator pool_allocator_nointr = {
2299: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1.117 yamt 2300: .pa_backingmapptr = &kernel_map,
1.66 thorpej 2301: };
1.112 bjh21 2302: #endif
1.66 thorpej 2303:
2304: #ifdef POOL_SUBPAGE
2305: void *pool_subpage_alloc(struct pool *, int);
2306: void pool_subpage_free(struct pool *, void *);
2307:
1.112 bjh21 2308: struct pool_allocator pool_allocator_kmem = {
2309: pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE,
1.117 yamt 2310: .pa_backingmapptr = &kmem_map,
1.112 bjh21 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,
1.117 yamt 2318: .pa_backingmapptr = &kmem_map,
1.66 thorpej 2319: };
2320: #endif /* POOL_SUBPAGE */
2321:
1.117 yamt 2322: static void *
2323: pool_allocator_alloc(struct pool *pp, int flags)
1.66 thorpej 2324: {
1.117 yamt 2325: struct pool_allocator *pa = pp->pr_alloc;
1.66 thorpej 2326: void *res;
2327:
1.117 yamt 2328: res = (*pa->pa_alloc)(pp, flags);
2329: if (res == NULL && (flags & PR_WAITOK) == 0) {
1.66 thorpej 2330: /*
1.117 yamt 2331: * We only run the drain hook here if PR_NOWAIT.
2332: * In other cases, the hook will be run in
2333: * pool_reclaim().
1.66 thorpej 2334: */
1.117 yamt 2335: if (pp->pr_drain_hook != NULL) {
2336: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
2337: res = (*pa->pa_alloc)(pp, flags);
1.66 thorpej 2338: }
1.117 yamt 2339: }
2340: return res;
1.66 thorpej 2341: }
2342:
1.117 yamt 2343: static void
1.66 thorpej 2344: pool_allocator_free(struct pool *pp, void *v)
2345: {
2346: struct pool_allocator *pa = pp->pr_alloc;
2347:
2348: (*pa->pa_free)(pp, v);
2349: }
2350:
2351: void *
1.124 yamt 2352: pool_page_alloc(struct pool *pp, int flags)
1.66 thorpej 2353: {
1.127 thorpej 2354: bool waitok = (flags & PR_WAITOK) ? true : false;
1.66 thorpej 2355:
1.100 yamt 2356: return ((void *) uvm_km_alloc_poolpage_cache(kmem_map, waitok));
1.66 thorpej 2357: }
2358:
2359: void
1.124 yamt 2360: pool_page_free(struct pool *pp, void *v)
1.66 thorpej 2361: {
2362:
1.98 yamt 2363: uvm_km_free_poolpage_cache(kmem_map, (vaddr_t) v);
2364: }
2365:
2366: static void *
1.124 yamt 2367: pool_page_alloc_meta(struct pool *pp, int flags)
1.98 yamt 2368: {
1.127 thorpej 2369: bool waitok = (flags & PR_WAITOK) ? true : false;
1.98 yamt 2370:
1.100 yamt 2371: return ((void *) uvm_km_alloc_poolpage(kmem_map, waitok));
1.98 yamt 2372: }
2373:
2374: static void
1.124 yamt 2375: pool_page_free_meta(struct pool *pp, void *v)
1.98 yamt 2376: {
2377:
1.100 yamt 2378: uvm_km_free_poolpage(kmem_map, (vaddr_t) v);
1.66 thorpej 2379: }
2380:
2381: #ifdef POOL_SUBPAGE
2382: /* Sub-page allocator, for machines with large hardware pages. */
2383: void *
2384: pool_subpage_alloc(struct pool *pp, int flags)
2385: {
1.128.2.2 ad 2386: return pool_get(&psppool, flags);
1.66 thorpej 2387: }
2388:
2389: void
2390: pool_subpage_free(struct pool *pp, void *v)
2391: {
2392: pool_put(&psppool, v);
2393: }
2394:
2395: /* We don't provide a real nointr allocator. Maybe later. */
2396: void *
1.112 bjh21 2397: pool_subpage_alloc_nointr(struct pool *pp, int flags)
1.66 thorpej 2398: {
2399:
2400: return (pool_subpage_alloc(pp, flags));
2401: }
2402:
2403: void
1.112 bjh21 2404: pool_subpage_free_nointr(struct pool *pp, void *v)
1.66 thorpej 2405: {
2406:
2407: pool_subpage_free(pp, v);
2408: }
1.112 bjh21 2409: #endif /* POOL_SUBPAGE */
1.66 thorpej 2410: void *
1.124 yamt 2411: pool_page_alloc_nointr(struct pool *pp, int flags)
1.66 thorpej 2412: {
1.127 thorpej 2413: bool waitok = (flags & PR_WAITOK) ? true : false;
1.66 thorpej 2414:
1.100 yamt 2415: return ((void *) uvm_km_alloc_poolpage_cache(kernel_map, waitok));
1.66 thorpej 2416: }
2417:
2418: void
1.124 yamt 2419: pool_page_free_nointr(struct pool *pp, void *v)
1.66 thorpej 2420: {
2421:
1.98 yamt 2422: uvm_km_free_poolpage_cache(kernel_map, (vaddr_t) v);
1.66 thorpej 2423: }
CVSweb <webmaster@jp.NetBSD.org>
![[BACK]](/icons/back.gif){kind=icon}
Return to subr_pool.c CVS log ![[TXT]](/icons/text.gif){kind=icon}
![[DIR]](/icons/dir.gif){kind=icon}
Up to [cvs.NetBSD.org] / src / sys / kern