Annotation of src/sys/kern/subr_pool.c, Revision 1.111.4.1
1.111.4.1! simonb 1: /* $NetBSD: subr_pool.c,v 1.111 2006/01/26 15:07:25 christos Exp $ */
1.1 pk 2:
3: /*-
1.43 thorpej 4: * Copyright (c) 1997, 1999, 2000 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>
1.111.4.1! simonb 41: __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.111 2006/01/26 15:07:25 christos 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>
1.3 pk 56:
57: #include <uvm/uvm.h>
58:
1.1 pk 59: /*
60: * Pool resource management utility.
1.3 pk 61: *
1.88 chs 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').
1.1 pk 70: */
71:
1.3 pk 72: /* List of all pools */
1.102 chs 73: LIST_HEAD(,pool) pool_head = LIST_HEAD_INITIALIZER(pool_head);
1.3 pk 74:
75: /* Private pool for page header structures */
1.97 yamt 76: #define PHPOOL_MAX 8
77: static struct pool phpool[PHPOOL_MAX];
78: #define PHPOOL_FREELIST_NELEM(idx) (((idx) == 0) ? 0 : (1 << (idx)))
1.3 pk 79:
1.62 bjh21 80: #ifdef POOL_SUBPAGE
81: /* Pool of subpages for use by normal pools. */
82: static struct pool psppool;
83: #endif
84:
1.98 yamt 85: static void *pool_page_alloc_meta(struct pool *, int);
86: static void pool_page_free_meta(struct pool *, void *);
87:
88: /* allocator for pool metadata */
89: static struct pool_allocator pool_allocator_meta = {
90: pool_page_alloc_meta, pool_page_free_meta
91: };
92:
1.3 pk 93: /* # of seconds to retain page after last use */
94: int pool_inactive_time = 10;
95:
96: /* Next candidate for drainage (see pool_drain()) */
1.23 thorpej 97: static struct pool *drainpp;
98:
99: /* This spin lock protects both pool_head and drainpp. */
100: struct simplelock pool_head_slock = SIMPLELOCK_INITIALIZER;
1.3 pk 101:
1.99 yamt 102: typedef uint8_t pool_item_freelist_t;
103:
1.3 pk 104: struct pool_item_header {
105: /* Page headers */
1.88 chs 106: LIST_ENTRY(pool_item_header)
1.3 pk 107: ph_pagelist; /* pool page list */
1.88 chs 108: SPLAY_ENTRY(pool_item_header)
109: ph_node; /* Off-page page headers */
1.3 pk 110: caddr_t ph_page; /* this page's address */
111: struct timeval ph_time; /* last referenced */
1.97 yamt 112: union {
113: /* !PR_NOTOUCH */
114: struct {
1.102 chs 115: LIST_HEAD(, pool_item)
1.97 yamt 116: phu_itemlist; /* chunk list for this page */
117: } phu_normal;
118: /* PR_NOTOUCH */
119: struct {
120: uint16_t
121: phu_off; /* start offset in page */
1.99 yamt 122: pool_item_freelist_t
1.97 yamt 123: phu_firstfree; /* first free item */
1.99 yamt 124: /*
125: * XXX it might be better to use
126: * a simple bitmap and ffs(3)
127: */
1.97 yamt 128: } phu_notouch;
129: } ph_u;
130: uint16_t ph_nmissing; /* # of chunks in use */
1.3 pk 131: };
1.97 yamt 132: #define ph_itemlist ph_u.phu_normal.phu_itemlist
133: #define ph_off ph_u.phu_notouch.phu_off
134: #define ph_firstfree ph_u.phu_notouch.phu_firstfree
1.3 pk 135:
1.1 pk 136: struct pool_item {
1.3 pk 137: #ifdef DIAGNOSTIC
1.82 thorpej 138: u_int pi_magic;
1.33 chs 139: #endif
1.82 thorpej 140: #define PI_MAGIC 0xdeadbeefU
1.3 pk 141: /* Other entries use only this list entry */
1.102 chs 142: LIST_ENTRY(pool_item) pi_list;
1.3 pk 143: };
144:
1.53 thorpej 145: #define POOL_NEEDS_CATCHUP(pp) \
146: ((pp)->pr_nitems < (pp)->pr_minitems)
147:
1.43 thorpej 148: /*
149: * Pool cache management.
150: *
151: * Pool caches provide a way for constructed objects to be cached by the
152: * pool subsystem. This can lead to performance improvements by avoiding
153: * needless object construction/destruction; it is deferred until absolutely
154: * necessary.
155: *
156: * Caches are grouped into cache groups. Each cache group references
157: * up to 16 constructed objects. When a cache allocates an object
158: * from the pool, it calls the object's constructor and places it into
159: * a cache group. When a cache group frees an object back to the pool,
160: * it first calls the object's destructor. This allows the object to
161: * persist in constructed form while freed to the cache.
162: *
163: * Multiple caches may exist for each pool. This allows a single
164: * object type to have multiple constructed forms. The pool references
165: * each cache, so that when a pool is drained by the pagedaemon, it can
166: * drain each individual cache as well. Each time a cache is drained,
167: * the most idle cache group is freed to the pool in its entirety.
168: *
169: * Pool caches are layed on top of pools. By layering them, we can avoid
170: * the complexity of cache management for pools which would not benefit
171: * from it.
172: */
173:
174: /* The cache group pool. */
175: static struct pool pcgpool;
1.3 pk 176:
1.102 chs 177: static void pool_cache_reclaim(struct pool_cache *, struct pool_pagelist *,
178: struct pool_cache_grouplist *);
179: static void pcg_grouplist_free(struct pool_cache_grouplist *);
1.3 pk 180:
1.42 thorpej 181: static int pool_catchup(struct pool *);
1.55 thorpej 182: static void pool_prime_page(struct pool *, caddr_t,
183: struct pool_item_header *);
1.88 chs 184: static void pool_update_curpage(struct pool *);
1.66 thorpej 185:
186: void *pool_allocator_alloc(struct pool *, int);
187: void pool_allocator_free(struct pool *, void *);
1.3 pk 188:
1.97 yamt 189: static void pool_print_pagelist(struct pool *, struct pool_pagelist *,
1.88 chs 190: void (*)(const char *, ...));
1.42 thorpej 191: static void pool_print1(struct pool *, const char *,
192: void (*)(const char *, ...));
1.3 pk 193:
1.88 chs 194: static int pool_chk_page(struct pool *, const char *,
195: struct pool_item_header *);
196:
1.3 pk 197: /*
1.52 thorpej 198: * Pool log entry. An array of these is allocated in pool_init().
1.3 pk 199: */
200: struct pool_log {
201: const char *pl_file;
202: long pl_line;
203: int pl_action;
1.25 thorpej 204: #define PRLOG_GET 1
205: #define PRLOG_PUT 2
1.3 pk 206: void *pl_addr;
1.1 pk 207: };
208:
1.86 matt 209: #ifdef POOL_DIAGNOSTIC
1.3 pk 210: /* Number of entries in pool log buffers */
1.17 thorpej 211: #ifndef POOL_LOGSIZE
212: #define POOL_LOGSIZE 10
213: #endif
214:
215: int pool_logsize = POOL_LOGSIZE;
1.1 pk 216:
1.110 perry 217: static inline void
1.42 thorpej 218: pr_log(struct pool *pp, void *v, int action, const char *file, long line)
1.3 pk 219: {
220: int n = pp->pr_curlogentry;
221: struct pool_log *pl;
222:
1.20 thorpej 223: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 224: return;
225:
226: /*
227: * Fill in the current entry. Wrap around and overwrite
228: * the oldest entry if necessary.
229: */
230: pl = &pp->pr_log[n];
231: pl->pl_file = file;
232: pl->pl_line = line;
233: pl->pl_action = action;
234: pl->pl_addr = v;
235: if (++n >= pp->pr_logsize)
236: n = 0;
237: pp->pr_curlogentry = n;
238: }
239:
240: static void
1.42 thorpej 241: pr_printlog(struct pool *pp, struct pool_item *pi,
242: void (*pr)(const char *, ...))
1.3 pk 243: {
244: int i = pp->pr_logsize;
245: int n = pp->pr_curlogentry;
246:
1.20 thorpej 247: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 248: return;
249:
250: /*
251: * Print all entries in this pool's log.
252: */
253: while (i-- > 0) {
254: struct pool_log *pl = &pp->pr_log[n];
255: if (pl->pl_action != 0) {
1.25 thorpej 256: if (pi == NULL || pi == pl->pl_addr) {
257: (*pr)("\tlog entry %d:\n", i);
258: (*pr)("\t\taction = %s, addr = %p\n",
259: pl->pl_action == PRLOG_GET ? "get" : "put",
260: pl->pl_addr);
261: (*pr)("\t\tfile: %s at line %lu\n",
262: pl->pl_file, pl->pl_line);
263: }
1.3 pk 264: }
265: if (++n >= pp->pr_logsize)
266: n = 0;
267: }
268: }
1.25 thorpej 269:
1.110 perry 270: static inline void
1.42 thorpej 271: pr_enter(struct pool *pp, const char *file, long line)
1.25 thorpej 272: {
273:
1.34 thorpej 274: if (__predict_false(pp->pr_entered_file != NULL)) {
1.25 thorpej 275: printf("pool %s: reentrancy at file %s line %ld\n",
276: pp->pr_wchan, file, line);
277: printf(" previous entry at file %s line %ld\n",
278: pp->pr_entered_file, pp->pr_entered_line);
279: panic("pr_enter");
280: }
281:
282: pp->pr_entered_file = file;
283: pp->pr_entered_line = line;
284: }
285:
1.110 perry 286: static inline void
1.42 thorpej 287: pr_leave(struct pool *pp)
1.25 thorpej 288: {
289:
1.34 thorpej 290: if (__predict_false(pp->pr_entered_file == NULL)) {
1.25 thorpej 291: printf("pool %s not entered?\n", pp->pr_wchan);
292: panic("pr_leave");
293: }
294:
295: pp->pr_entered_file = NULL;
296: pp->pr_entered_line = 0;
297: }
298:
1.110 perry 299: static inline void
1.42 thorpej 300: pr_enter_check(struct pool *pp, void (*pr)(const char *, ...))
1.25 thorpej 301: {
302:
303: if (pp->pr_entered_file != NULL)
304: (*pr)("\n\tcurrently entered from file %s line %ld\n",
305: pp->pr_entered_file, pp->pr_entered_line);
306: }
1.3 pk 307: #else
1.25 thorpej 308: #define pr_log(pp, v, action, file, line)
309: #define pr_printlog(pp, pi, pr)
310: #define pr_enter(pp, file, line)
311: #define pr_leave(pp)
312: #define pr_enter_check(pp, pr)
1.59 thorpej 313: #endif /* POOL_DIAGNOSTIC */
1.3 pk 314:
1.110 perry 315: static inline int
1.97 yamt 316: pr_item_notouch_index(const struct pool *pp, const struct pool_item_header *ph,
317: const void *v)
318: {
319: const char *cp = v;
320: int idx;
321:
322: KASSERT(pp->pr_roflags & PR_NOTOUCH);
323: idx = (cp - ph->ph_page - ph->ph_off) / pp->pr_size;
324: KASSERT(idx < pp->pr_itemsperpage);
325: return idx;
326: }
327:
1.99 yamt 328: #define PR_FREELIST_ALIGN(p) \
329: roundup((uintptr_t)(p), sizeof(pool_item_freelist_t))
330: #define PR_FREELIST(ph) ((pool_item_freelist_t *)PR_FREELIST_ALIGN((ph) + 1))
331: #define PR_INDEX_USED ((pool_item_freelist_t)-1)
332: #define PR_INDEX_EOL ((pool_item_freelist_t)-2)
1.97 yamt 333:
1.110 perry 334: static inline void
1.97 yamt 335: pr_item_notouch_put(const struct pool *pp, struct pool_item_header *ph,
336: void *obj)
337: {
338: int idx = pr_item_notouch_index(pp, ph, obj);
1.99 yamt 339: pool_item_freelist_t *freelist = PR_FREELIST(ph);
1.97 yamt 340:
341: KASSERT(freelist[idx] == PR_INDEX_USED);
342: freelist[idx] = ph->ph_firstfree;
343: ph->ph_firstfree = idx;
344: }
345:
1.110 perry 346: static inline void *
1.97 yamt 347: pr_item_notouch_get(const struct pool *pp, struct pool_item_header *ph)
348: {
349: int idx = ph->ph_firstfree;
1.99 yamt 350: pool_item_freelist_t *freelist = PR_FREELIST(ph);
1.97 yamt 351:
352: KASSERT(freelist[idx] != PR_INDEX_USED);
353: ph->ph_firstfree = freelist[idx];
354: freelist[idx] = PR_INDEX_USED;
355:
356: return ph->ph_page + ph->ph_off + idx * pp->pr_size;
357: }
358:
1.110 perry 359: static inline int
1.88 chs 360: phtree_compare(struct pool_item_header *a, struct pool_item_header *b)
361: {
362: if (a->ph_page < b->ph_page)
363: return (-1);
364: else if (a->ph_page > b->ph_page)
365: return (1);
366: else
367: return (0);
368: }
369:
370: SPLAY_PROTOTYPE(phtree, pool_item_header, ph_node, phtree_compare);
371: SPLAY_GENERATE(phtree, pool_item_header, ph_node, phtree_compare);
372:
1.3 pk 373: /*
374: * Return the pool page header based on page address.
375: */
1.110 perry 376: static inline struct pool_item_header *
1.42 thorpej 377: pr_find_pagehead(struct pool *pp, caddr_t page)
1.3 pk 378: {
1.88 chs 379: struct pool_item_header *ph, tmp;
1.3 pk 380:
1.20 thorpej 381: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
1.3 pk 382: return ((struct pool_item_header *)(page + pp->pr_phoffset));
383:
1.88 chs 384: tmp.ph_page = page;
385: ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
386: return ph;
1.3 pk 387: }
388:
1.101 thorpej 389: static void
390: pr_pagelist_free(struct pool *pp, struct pool_pagelist *pq)
391: {
392: struct pool_item_header *ph;
393: int s;
394:
395: while ((ph = LIST_FIRST(pq)) != NULL) {
396: LIST_REMOVE(ph, ph_pagelist);
397: pool_allocator_free(pp, ph->ph_page);
398: if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
399: s = splvm();
400: pool_put(pp->pr_phpool, ph);
401: splx(s);
402: }
403: }
404: }
405:
1.3 pk 406: /*
407: * Remove a page from the pool.
408: */
1.110 perry 409: static inline void
1.61 chs 410: pr_rmpage(struct pool *pp, struct pool_item_header *ph,
411: struct pool_pagelist *pq)
1.3 pk 412: {
413:
1.101 thorpej 414: LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
1.91 yamt 415:
1.3 pk 416: /*
1.7 thorpej 417: * If the page was idle, decrement the idle page count.
1.3 pk 418: */
1.6 thorpej 419: if (ph->ph_nmissing == 0) {
420: #ifdef DIAGNOSTIC
421: if (pp->pr_nidle == 0)
422: panic("pr_rmpage: nidle inconsistent");
1.20 thorpej 423: if (pp->pr_nitems < pp->pr_itemsperpage)
424: panic("pr_rmpage: nitems inconsistent");
1.6 thorpej 425: #endif
426: pp->pr_nidle--;
427: }
1.7 thorpej 428:
1.20 thorpej 429: pp->pr_nitems -= pp->pr_itemsperpage;
430:
1.7 thorpej 431: /*
1.101 thorpej 432: * Unlink the page from the pool and queue it for release.
1.7 thorpej 433: */
1.88 chs 434: LIST_REMOVE(ph, ph_pagelist);
1.91 yamt 435: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
436: SPLAY_REMOVE(phtree, &pp->pr_phtree, ph);
1.101 thorpej 437: LIST_INSERT_HEAD(pq, ph, ph_pagelist);
438:
1.7 thorpej 439: pp->pr_npages--;
440: pp->pr_npagefree++;
1.6 thorpej 441:
1.88 chs 442: pool_update_curpage(pp);
1.3 pk 443: }
444:
445: /*
1.94 simonb 446: * Initialize all the pools listed in the "pools" link set.
447: */
448: void
449: link_pool_init(void)
450: {
451: __link_set_decl(pools, struct link_pool_init);
452: struct link_pool_init * const *pi;
453:
454: __link_set_foreach(pi, pools)
455: pool_init((*pi)->pp, (*pi)->size, (*pi)->align,
456: (*pi)->align_offset, (*pi)->flags, (*pi)->wchan,
457: (*pi)->palloc);
458: }
459:
460: /*
1.3 pk 461: * Initialize the given pool resource structure.
462: *
463: * We export this routine to allow other kernel parts to declare
464: * static pools that must be initialized before malloc() is available.
465: */
466: void
1.42 thorpej 467: pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
1.66 thorpej 468: const char *wchan, struct pool_allocator *palloc)
1.3 pk 469: {
1.88 chs 470: int off, slack;
1.92 enami 471: size_t trysize, phsize;
1.93 dbj 472: int s;
1.3 pk 473:
1.99 yamt 474: KASSERT((1UL << (CHAR_BIT * sizeof(pool_item_freelist_t))) - 2 >=
475: PHPOOL_FREELIST_NELEM(PHPOOL_MAX - 1));
476:
1.25 thorpej 477: #ifdef POOL_DIAGNOSTIC
478: /*
479: * Always log if POOL_DIAGNOSTIC is defined.
480: */
481: if (pool_logsize != 0)
482: flags |= PR_LOGGING;
483: #endif
484:
1.66 thorpej 485: #ifdef POOL_SUBPAGE
486: /*
487: * XXX We don't provide a real `nointr' back-end
488: * yet; all sub-pages come from a kmem back-end.
489: * maybe some day...
490: */
491: if (palloc == NULL) {
492: extern struct pool_allocator pool_allocator_kmem_subpage;
493: palloc = &pool_allocator_kmem_subpage;
494: }
1.3 pk 495: /*
1.66 thorpej 496: * We'll assume any user-specified back-end allocator
497: * will deal with sub-pages, or simply don't care.
1.3 pk 498: */
1.66 thorpej 499: #else
500: if (palloc == NULL)
501: palloc = &pool_allocator_kmem;
502: #endif /* POOL_SUBPAGE */
503: if ((palloc->pa_flags & PA_INITIALIZED) == 0) {
504: if (palloc->pa_pagesz == 0) {
1.62 bjh21 505: #ifdef POOL_SUBPAGE
1.66 thorpej 506: if (palloc == &pool_allocator_kmem)
507: palloc->pa_pagesz = PAGE_SIZE;
508: else
509: palloc->pa_pagesz = POOL_SUBPAGE;
1.62 bjh21 510: #else
1.66 thorpej 511: palloc->pa_pagesz = PAGE_SIZE;
512: #endif /* POOL_SUBPAGE */
513: }
514:
515: TAILQ_INIT(&palloc->pa_list);
516:
517: simple_lock_init(&palloc->pa_slock);
518: palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
519: palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
520: palloc->pa_flags |= PA_INITIALIZED;
1.4 thorpej 521: }
1.3 pk 522:
523: if (align == 0)
524: align = ALIGN(1);
1.14 thorpej 525:
526: if (size < sizeof(struct pool_item))
527: size = sizeof(struct pool_item);
1.3 pk 528:
1.78 thorpej 529: size = roundup(size, align);
1.66 thorpej 530: #ifdef DIAGNOSTIC
531: if (size > palloc->pa_pagesz)
1.35 pk 532: panic("pool_init: pool item size (%lu) too large",
533: (u_long)size);
1.66 thorpej 534: #endif
1.35 pk 535:
1.3 pk 536: /*
537: * Initialize the pool structure.
538: */
1.88 chs 539: LIST_INIT(&pp->pr_emptypages);
540: LIST_INIT(&pp->pr_fullpages);
541: LIST_INIT(&pp->pr_partpages);
1.102 chs 542: LIST_INIT(&pp->pr_cachelist);
1.3 pk 543: pp->pr_curpage = NULL;
544: pp->pr_npages = 0;
545: pp->pr_minitems = 0;
546: pp->pr_minpages = 0;
547: pp->pr_maxpages = UINT_MAX;
1.20 thorpej 548: pp->pr_roflags = flags;
549: pp->pr_flags = 0;
1.35 pk 550: pp->pr_size = size;
1.3 pk 551: pp->pr_align = align;
552: pp->pr_wchan = wchan;
1.66 thorpej 553: pp->pr_alloc = palloc;
1.20 thorpej 554: pp->pr_nitems = 0;
555: pp->pr_nout = 0;
556: pp->pr_hardlimit = UINT_MAX;
557: pp->pr_hardlimit_warning = NULL;
1.31 thorpej 558: pp->pr_hardlimit_ratecap.tv_sec = 0;
559: pp->pr_hardlimit_ratecap.tv_usec = 0;
560: pp->pr_hardlimit_warning_last.tv_sec = 0;
561: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.68 thorpej 562: pp->pr_drain_hook = NULL;
563: pp->pr_drain_hook_arg = NULL;
1.3 pk 564:
565: /*
566: * Decide whether to put the page header off page to avoid
1.92 enami 567: * wasting too large a part of the page or too big item.
568: * Off-page page headers go on a hash table, so we can match
569: * a returned item with its header based on the page address.
570: * We use 1/16 of the page size and about 8 times of the item
571: * size as the threshold (XXX: tune)
572: *
573: * However, we'll put the header into the page if we can put
574: * it without wasting any items.
575: *
576: * Silently enforce `0 <= ioff < align'.
1.3 pk 577: */
1.92 enami 578: pp->pr_itemoffset = ioff %= align;
579: /* See the comment below about reserved bytes. */
580: trysize = palloc->pa_pagesz - ((align - ioff) % align);
581: phsize = ALIGN(sizeof(struct pool_item_header));
1.97 yamt 582: if ((pp->pr_roflags & PR_NOTOUCH) == 0 &&
583: (pp->pr_size < MIN(palloc->pa_pagesz / 16, phsize << 3) ||
584: trysize / pp->pr_size == (trysize - phsize) / pp->pr_size)) {
1.3 pk 585: /* Use the end of the page for the page header */
1.20 thorpej 586: pp->pr_roflags |= PR_PHINPAGE;
1.92 enami 587: pp->pr_phoffset = off = palloc->pa_pagesz - phsize;
1.2 pk 588: } else {
1.3 pk 589: /* The page header will be taken from our page header pool */
590: pp->pr_phoffset = 0;
1.66 thorpej 591: off = palloc->pa_pagesz;
1.88 chs 592: SPLAY_INIT(&pp->pr_phtree);
1.2 pk 593: }
1.1 pk 594:
1.3 pk 595: /*
596: * Alignment is to take place at `ioff' within the item. This means
597: * we must reserve up to `align - 1' bytes on the page to allow
598: * appropriate positioning of each item.
599: */
600: pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
1.43 thorpej 601: KASSERT(pp->pr_itemsperpage != 0);
1.97 yamt 602: if ((pp->pr_roflags & PR_NOTOUCH)) {
603: int idx;
604:
605: for (idx = 0; pp->pr_itemsperpage > PHPOOL_FREELIST_NELEM(idx);
606: idx++) {
607: /* nothing */
608: }
609: if (idx >= PHPOOL_MAX) {
610: /*
611: * if you see this panic, consider to tweak
612: * PHPOOL_MAX and PHPOOL_FREELIST_NELEM.
613: */
614: panic("%s: too large itemsperpage(%d) for PR_NOTOUCH",
615: pp->pr_wchan, pp->pr_itemsperpage);
616: }
617: pp->pr_phpool = &phpool[idx];
618: } else if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
619: pp->pr_phpool = &phpool[0];
620: }
621: #if defined(DIAGNOSTIC)
622: else {
623: pp->pr_phpool = NULL;
624: }
625: #endif
1.3 pk 626:
627: /*
628: * Use the slack between the chunks and the page header
629: * for "cache coloring".
630: */
631: slack = off - pp->pr_itemsperpage * pp->pr_size;
632: pp->pr_maxcolor = (slack / align) * align;
633: pp->pr_curcolor = 0;
634:
635: pp->pr_nget = 0;
636: pp->pr_nfail = 0;
637: pp->pr_nput = 0;
638: pp->pr_npagealloc = 0;
639: pp->pr_npagefree = 0;
1.1 pk 640: pp->pr_hiwat = 0;
1.8 thorpej 641: pp->pr_nidle = 0;
1.3 pk 642:
1.59 thorpej 643: #ifdef POOL_DIAGNOSTIC
1.25 thorpej 644: if (flags & PR_LOGGING) {
645: if (kmem_map == NULL ||
646: (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log),
647: M_TEMP, M_NOWAIT)) == NULL)
1.20 thorpej 648: pp->pr_roflags &= ~PR_LOGGING;
1.3 pk 649: pp->pr_curlogentry = 0;
650: pp->pr_logsize = pool_logsize;
651: }
1.59 thorpej 652: #endif
1.25 thorpej 653:
654: pp->pr_entered_file = NULL;
655: pp->pr_entered_line = 0;
1.3 pk 656:
1.21 thorpej 657: simple_lock_init(&pp->pr_slock);
1.1 pk 658:
1.3 pk 659: /*
1.43 thorpej 660: * Initialize private page header pool and cache magazine pool if we
661: * haven't done so yet.
1.23 thorpej 662: * XXX LOCKING.
1.3 pk 663: */
1.97 yamt 664: if (phpool[0].pr_size == 0) {
665: int idx;
666: for (idx = 0; idx < PHPOOL_MAX; idx++) {
667: static char phpool_names[PHPOOL_MAX][6+1+6+1];
668: int nelem;
669: size_t sz;
670:
671: nelem = PHPOOL_FREELIST_NELEM(idx);
672: snprintf(phpool_names[idx], sizeof(phpool_names[idx]),
673: "phpool-%d", nelem);
674: sz = sizeof(struct pool_item_header);
675: if (nelem) {
676: sz = PR_FREELIST_ALIGN(sz)
1.99 yamt 677: + nelem * sizeof(pool_item_freelist_t);
1.97 yamt 678: }
679: pool_init(&phpool[idx], sz, 0, 0, 0,
1.98 yamt 680: phpool_names[idx], &pool_allocator_meta);
1.97 yamt 681: }
1.62 bjh21 682: #ifdef POOL_SUBPAGE
683: pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
1.98 yamt 684: PR_RECURSIVE, "psppool", &pool_allocator_meta);
1.62 bjh21 685: #endif
1.43 thorpej 686: pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
1.98 yamt 687: 0, "pcgpool", &pool_allocator_meta);
1.1 pk 688: }
689:
1.23 thorpej 690: /* Insert into the list of all pools. */
691: simple_lock(&pool_head_slock);
1.102 chs 692: LIST_INSERT_HEAD(&pool_head, pp, pr_poollist);
1.23 thorpej 693: simple_unlock(&pool_head_slock);
1.66 thorpej 694:
695: /* Insert this into the list of pools using this allocator. */
1.93 dbj 696: s = splvm();
1.66 thorpej 697: simple_lock(&palloc->pa_slock);
698: TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
699: simple_unlock(&palloc->pa_slock);
1.93 dbj 700: splx(s);
1.1 pk 701: }
702:
703: /*
704: * De-commision a pool resource.
705: */
706: void
1.42 thorpej 707: pool_destroy(struct pool *pp)
1.1 pk 708: {
1.101 thorpej 709: struct pool_pagelist pq;
1.3 pk 710: struct pool_item_header *ph;
1.93 dbj 711: int s;
1.43 thorpej 712:
1.101 thorpej 713: /* Remove from global pool list */
714: simple_lock(&pool_head_slock);
1.102 chs 715: LIST_REMOVE(pp, pr_poollist);
1.101 thorpej 716: if (drainpp == pp)
717: drainpp = NULL;
718: simple_unlock(&pool_head_slock);
719:
720: /* Remove this pool from its allocator's list of pools. */
1.93 dbj 721: s = splvm();
1.66 thorpej 722: simple_lock(&pp->pr_alloc->pa_slock);
723: TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
724: simple_unlock(&pp->pr_alloc->pa_slock);
1.93 dbj 725: splx(s);
1.66 thorpej 726:
1.101 thorpej 727: s = splvm();
728: simple_lock(&pp->pr_slock);
729:
1.102 chs 730: KASSERT(LIST_EMPTY(&pp->pr_cachelist));
1.3 pk 731:
732: #ifdef DIAGNOSTIC
1.20 thorpej 733: if (pp->pr_nout != 0) {
1.25 thorpej 734: pr_printlog(pp, NULL, printf);
1.80 provos 735: panic("pool_destroy: pool busy: still out: %u",
1.20 thorpej 736: pp->pr_nout);
1.3 pk 737: }
738: #endif
1.1 pk 739:
1.101 thorpej 740: KASSERT(LIST_EMPTY(&pp->pr_fullpages));
741: KASSERT(LIST_EMPTY(&pp->pr_partpages));
742:
1.3 pk 743: /* Remove all pages */
1.101 thorpej 744: LIST_INIT(&pq);
1.88 chs 745: while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1.101 thorpej 746: pr_rmpage(pp, ph, &pq);
747:
748: simple_unlock(&pp->pr_slock);
749: splx(s);
1.3 pk 750:
1.101 thorpej 751: pr_pagelist_free(pp, &pq);
1.3 pk 752:
1.59 thorpej 753: #ifdef POOL_DIAGNOSTIC
1.20 thorpej 754: if ((pp->pr_roflags & PR_LOGGING) != 0)
1.3 pk 755: free(pp->pr_log, M_TEMP);
1.59 thorpej 756: #endif
1.1 pk 757: }
758:
1.68 thorpej 759: void
760: pool_set_drain_hook(struct pool *pp, void (*fn)(void *, int), void *arg)
761: {
762:
763: /* XXX no locking -- must be used just after pool_init() */
764: #ifdef DIAGNOSTIC
765: if (pp->pr_drain_hook != NULL)
766: panic("pool_set_drain_hook(%s): already set", pp->pr_wchan);
767: #endif
768: pp->pr_drain_hook = fn;
769: pp->pr_drain_hook_arg = arg;
770: }
771:
1.88 chs 772: static struct pool_item_header *
1.55 thorpej 773: pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags)
774: {
775: struct pool_item_header *ph;
776: int s;
777:
778: LOCK_ASSERT(simple_lock_held(&pp->pr_slock) == 0);
779:
780: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
781: ph = (struct pool_item_header *) (storage + pp->pr_phoffset);
782: else {
1.85 pk 783: s = splvm();
1.97 yamt 784: ph = pool_get(pp->pr_phpool, flags);
1.55 thorpej 785: splx(s);
786: }
787:
788: return (ph);
789: }
1.1 pk 790:
791: /*
1.3 pk 792: * Grab an item from the pool; must be called at appropriate spl level
1.1 pk 793: */
1.3 pk 794: void *
1.59 thorpej 795: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 796: _pool_get(struct pool *pp, int flags, const char *file, long line)
1.56 sommerfe 797: #else
798: pool_get(struct pool *pp, int flags)
799: #endif
1.1 pk 800: {
801: struct pool_item *pi;
1.3 pk 802: struct pool_item_header *ph;
1.55 thorpej 803: void *v;
1.1 pk 804:
1.2 pk 805: #ifdef DIAGNOSTIC
1.95 atatat 806: if (__predict_false(pp->pr_itemsperpage == 0))
807: panic("pool_get: pool %p: pr_itemsperpage is zero, "
808: "pool not initialized?", pp);
1.84 thorpej 809: if (__predict_false(curlwp == NULL && doing_shutdown == 0 &&
1.37 sommerfe 810: (flags & PR_WAITOK) != 0))
1.77 matt 811: panic("pool_get: %s: must have NOWAIT", pp->pr_wchan);
1.58 thorpej 812:
1.102 chs 813: #endif /* DIAGNOSTIC */
1.58 thorpej 814: #ifdef LOCKDEBUG
815: if (flags & PR_WAITOK)
816: simple_lock_only_held(NULL, "pool_get(PR_WAITOK)");
1.102 chs 817: SCHED_ASSERT_UNLOCKED();
1.56 sommerfe 818: #endif
1.1 pk 819:
1.21 thorpej 820: simple_lock(&pp->pr_slock);
1.25 thorpej 821: pr_enter(pp, file, line);
1.20 thorpej 822:
823: startover:
824: /*
825: * Check to see if we've reached the hard limit. If we have,
826: * and we can wait, then wait until an item has been returned to
827: * the pool.
828: */
829: #ifdef DIAGNOSTIC
1.34 thorpej 830: if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
1.25 thorpej 831: pr_leave(pp);
1.21 thorpej 832: simple_unlock(&pp->pr_slock);
1.20 thorpej 833: panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
834: }
835: #endif
1.34 thorpej 836: if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
1.68 thorpej 837: if (pp->pr_drain_hook != NULL) {
838: /*
839: * Since the drain hook is going to free things
840: * back to the pool, unlock, call the hook, re-lock,
841: * and check the hardlimit condition again.
842: */
843: pr_leave(pp);
844: simple_unlock(&pp->pr_slock);
845: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
846: simple_lock(&pp->pr_slock);
847: pr_enter(pp, file, line);
848: if (pp->pr_nout < pp->pr_hardlimit)
849: goto startover;
850: }
851:
1.29 sommerfe 852: if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
1.20 thorpej 853: /*
854: * XXX: A warning isn't logged in this case. Should
855: * it be?
856: */
857: pp->pr_flags |= PR_WANTED;
1.25 thorpej 858: pr_leave(pp);
1.40 sommerfe 859: ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
1.25 thorpej 860: pr_enter(pp, file, line);
1.20 thorpej 861: goto startover;
862: }
1.31 thorpej 863:
864: /*
865: * Log a message that the hard limit has been hit.
866: */
867: if (pp->pr_hardlimit_warning != NULL &&
868: ratecheck(&pp->pr_hardlimit_warning_last,
869: &pp->pr_hardlimit_ratecap))
870: log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
1.21 thorpej 871:
872: pp->pr_nfail++;
873:
1.25 thorpej 874: pr_leave(pp);
1.21 thorpej 875: simple_unlock(&pp->pr_slock);
1.20 thorpej 876: return (NULL);
877: }
878:
1.3 pk 879: /*
880: * The convention we use is that if `curpage' is not NULL, then
881: * it points at a non-empty bucket. In particular, `curpage'
882: * never points at a page header which has PR_PHINPAGE set and
883: * has no items in its bucket.
884: */
1.20 thorpej 885: if ((ph = pp->pr_curpage) == NULL) {
886: #ifdef DIAGNOSTIC
887: if (pp->pr_nitems != 0) {
1.21 thorpej 888: simple_unlock(&pp->pr_slock);
1.20 thorpej 889: printf("pool_get: %s: curpage NULL, nitems %u\n",
890: pp->pr_wchan, pp->pr_nitems);
1.80 provos 891: panic("pool_get: nitems inconsistent");
1.20 thorpej 892: }
893: #endif
894:
1.21 thorpej 895: /*
896: * Call the back-end page allocator for more memory.
897: * Release the pool lock, as the back-end page allocator
898: * may block.
899: */
1.25 thorpej 900: pr_leave(pp);
1.21 thorpej 901: simple_unlock(&pp->pr_slock);
1.66 thorpej 902: v = pool_allocator_alloc(pp, flags);
1.55 thorpej 903: if (__predict_true(v != NULL))
904: ph = pool_alloc_item_header(pp, v, flags);
1.15 pk 905:
1.55 thorpej 906: if (__predict_false(v == NULL || ph == NULL)) {
907: if (v != NULL)
1.66 thorpej 908: pool_allocator_free(pp, v);
1.55 thorpej 909:
1.91 yamt 910: simple_lock(&pp->pr_slock);
911: pr_enter(pp, file, line);
912:
1.21 thorpej 913: /*
1.55 thorpej 914: * We were unable to allocate a page or item
915: * header, but we released the lock during
916: * allocation, so perhaps items were freed
917: * back to the pool. Check for this case.
1.21 thorpej 918: */
919: if (pp->pr_curpage != NULL)
920: goto startover;
1.15 pk 921:
1.3 pk 922: if ((flags & PR_WAITOK) == 0) {
923: pp->pr_nfail++;
1.25 thorpej 924: pr_leave(pp);
1.21 thorpej 925: simple_unlock(&pp->pr_slock);
1.1 pk 926: return (NULL);
1.3 pk 927: }
928:
1.15 pk 929: /*
930: * Wait for items to be returned to this pool.
1.21 thorpej 931: *
1.109 christos 932: * wake up once a second and try again,
933: * as the check in pool_cache_put_paddr() is racy.
1.15 pk 934: */
1.1 pk 935: pp->pr_flags |= PR_WANTED;
1.66 thorpej 936: /* PA_WANTED is already set on the allocator. */
1.25 thorpej 937: pr_leave(pp);
1.109 christos 938: ltsleep(pp, PSWP, pp->pr_wchan, hz, &pp->pr_slock);
1.25 thorpej 939: pr_enter(pp, file, line);
1.20 thorpej 940: goto startover;
1.1 pk 941: }
1.3 pk 942:
1.15 pk 943: /* We have more memory; add it to the pool */
1.91 yamt 944: simple_lock(&pp->pr_slock);
945: pr_enter(pp, file, line);
1.55 thorpej 946: pool_prime_page(pp, v, ph);
1.15 pk 947: pp->pr_npagealloc++;
948:
1.20 thorpej 949: /* Start the allocation process over. */
950: goto startover;
1.3 pk 951: }
1.97 yamt 952: if (pp->pr_roflags & PR_NOTOUCH) {
953: #ifdef DIAGNOSTIC
954: if (__predict_false(ph->ph_nmissing == pp->pr_itemsperpage)) {
955: pr_leave(pp);
956: simple_unlock(&pp->pr_slock);
957: panic("pool_get: %s: page empty", pp->pr_wchan);
958: }
959: #endif
960: v = pr_item_notouch_get(pp, ph);
961: #ifdef POOL_DIAGNOSTIC
962: pr_log(pp, v, PRLOG_GET, file, line);
963: #endif
964: } else {
1.102 chs 965: v = pi = LIST_FIRST(&ph->ph_itemlist);
1.97 yamt 966: if (__predict_false(v == NULL)) {
967: pr_leave(pp);
968: simple_unlock(&pp->pr_slock);
969: panic("pool_get: %s: page empty", pp->pr_wchan);
970: }
1.20 thorpej 971: #ifdef DIAGNOSTIC
1.97 yamt 972: if (__predict_false(pp->pr_nitems == 0)) {
973: pr_leave(pp);
974: simple_unlock(&pp->pr_slock);
975: printf("pool_get: %s: items on itemlist, nitems %u\n",
976: pp->pr_wchan, pp->pr_nitems);
977: panic("pool_get: nitems inconsistent");
978: }
1.65 enami 979: #endif
1.56 sommerfe 980:
1.65 enami 981: #ifdef POOL_DIAGNOSTIC
1.97 yamt 982: pr_log(pp, v, PRLOG_GET, file, line);
1.65 enami 983: #endif
1.3 pk 984:
1.65 enami 985: #ifdef DIAGNOSTIC
1.97 yamt 986: if (__predict_false(pi->pi_magic != PI_MAGIC)) {
987: pr_printlog(pp, pi, printf);
988: panic("pool_get(%s): free list modified: "
989: "magic=%x; page %p; item addr %p\n",
990: pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
991: }
1.3 pk 992: #endif
993:
1.97 yamt 994: /*
995: * Remove from item list.
996: */
1.102 chs 997: LIST_REMOVE(pi, pi_list);
1.97 yamt 998: }
1.20 thorpej 999: pp->pr_nitems--;
1000: pp->pr_nout++;
1.6 thorpej 1001: if (ph->ph_nmissing == 0) {
1002: #ifdef DIAGNOSTIC
1.34 thorpej 1003: if (__predict_false(pp->pr_nidle == 0))
1.6 thorpej 1004: panic("pool_get: nidle inconsistent");
1005: #endif
1006: pp->pr_nidle--;
1.88 chs 1007:
1008: /*
1009: * This page was previously empty. Move it to the list of
1010: * partially-full pages. This page is already curpage.
1011: */
1012: LIST_REMOVE(ph, ph_pagelist);
1013: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.6 thorpej 1014: }
1.3 pk 1015: ph->ph_nmissing++;
1.97 yamt 1016: if (ph->ph_nmissing == pp->pr_itemsperpage) {
1.21 thorpej 1017: #ifdef DIAGNOSTIC
1.97 yamt 1018: if (__predict_false((pp->pr_roflags & PR_NOTOUCH) == 0 &&
1.102 chs 1019: !LIST_EMPTY(&ph->ph_itemlist))) {
1.25 thorpej 1020: pr_leave(pp);
1.21 thorpej 1021: simple_unlock(&pp->pr_slock);
1022: panic("pool_get: %s: nmissing inconsistent",
1023: pp->pr_wchan);
1024: }
1025: #endif
1.3 pk 1026: /*
1.88 chs 1027: * This page is now full. Move it to the full list
1028: * and select a new current page.
1.3 pk 1029: */
1.88 chs 1030: LIST_REMOVE(ph, ph_pagelist);
1031: LIST_INSERT_HEAD(&pp->pr_fullpages, ph, ph_pagelist);
1032: pool_update_curpage(pp);
1.1 pk 1033: }
1.3 pk 1034:
1035: pp->pr_nget++;
1.111 christos 1036: pr_leave(pp);
1.20 thorpej 1037:
1038: /*
1039: * If we have a low water mark and we are now below that low
1040: * water mark, add more items to the pool.
1041: */
1.53 thorpej 1042: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1043: /*
1044: * XXX: Should we log a warning? Should we set up a timeout
1045: * to try again in a second or so? The latter could break
1046: * a caller's assumptions about interrupt protection, etc.
1047: */
1048: }
1049:
1.21 thorpej 1050: simple_unlock(&pp->pr_slock);
1.1 pk 1051: return (v);
1052: }
1053:
1054: /*
1.43 thorpej 1055: * Internal version of pool_put(). Pool is already locked/entered.
1.1 pk 1056: */
1.43 thorpej 1057: static void
1.101 thorpej 1058: pool_do_put(struct pool *pp, void *v, struct pool_pagelist *pq)
1.1 pk 1059: {
1060: struct pool_item *pi = v;
1.3 pk 1061: struct pool_item_header *ph;
1062: caddr_t page;
1063:
1.61 chs 1064: LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
1.102 chs 1065: SCHED_ASSERT_UNLOCKED();
1.61 chs 1066:
1.66 thorpej 1067: page = (caddr_t)((u_long)v & pp->pr_alloc->pa_pagemask);
1.1 pk 1068:
1.30 thorpej 1069: #ifdef DIAGNOSTIC
1.34 thorpej 1070: if (__predict_false(pp->pr_nout == 0)) {
1.30 thorpej 1071: printf("pool %s: putting with none out\n",
1072: pp->pr_wchan);
1073: panic("pool_put");
1074: }
1075: #endif
1.3 pk 1076:
1.34 thorpej 1077: if (__predict_false((ph = pr_find_pagehead(pp, page)) == NULL)) {
1.25 thorpej 1078: pr_printlog(pp, NULL, printf);
1.3 pk 1079: panic("pool_put: %s: page header missing", pp->pr_wchan);
1080: }
1.28 thorpej 1081:
1082: #ifdef LOCKDEBUG
1083: /*
1084: * Check if we're freeing a locked simple lock.
1085: */
1086: simple_lock_freecheck((caddr_t)pi, ((caddr_t)pi) + pp->pr_size);
1087: #endif
1.3 pk 1088:
1089: /*
1090: * Return to item list.
1091: */
1.97 yamt 1092: if (pp->pr_roflags & PR_NOTOUCH) {
1093: pr_item_notouch_put(pp, ph, v);
1094: } else {
1.2 pk 1095: #ifdef DIAGNOSTIC
1.97 yamt 1096: pi->pi_magic = PI_MAGIC;
1.3 pk 1097: #endif
1.32 chs 1098: #ifdef DEBUG
1.97 yamt 1099: {
1100: int i, *ip = v;
1.32 chs 1101:
1.97 yamt 1102: for (i = 0; i < pp->pr_size / sizeof(int); i++) {
1103: *ip++ = PI_MAGIC;
1104: }
1.32 chs 1105: }
1106: #endif
1107:
1.102 chs 1108: LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.97 yamt 1109: }
1.79 thorpej 1110: KDASSERT(ph->ph_nmissing != 0);
1.3 pk 1111: ph->ph_nmissing--;
1112: pp->pr_nput++;
1.20 thorpej 1113: pp->pr_nitems++;
1114: pp->pr_nout--;
1.3 pk 1115:
1116: /* Cancel "pool empty" condition if it exists */
1117: if (pp->pr_curpage == NULL)
1118: pp->pr_curpage = ph;
1119:
1120: if (pp->pr_flags & PR_WANTED) {
1121: pp->pr_flags &= ~PR_WANTED;
1.15 pk 1122: if (ph->ph_nmissing == 0)
1123: pp->pr_nidle++;
1.3 pk 1124: wakeup((caddr_t)pp);
1125: return;
1126: }
1127:
1128: /*
1.88 chs 1129: * If this page is now empty, do one of two things:
1.21 thorpej 1130: *
1.88 chs 1131: * (1) If we have more pages than the page high water mark,
1.96 thorpej 1132: * free the page back to the system. ONLY CONSIDER
1.90 thorpej 1133: * FREEING BACK A PAGE IF WE HAVE MORE THAN OUR MINIMUM PAGE
1134: * CLAIM.
1.21 thorpej 1135: *
1.88 chs 1136: * (2) Otherwise, move the page to the empty page list.
1137: *
1138: * Either way, select a new current page (so we use a partially-full
1139: * page if one is available).
1.3 pk 1140: */
1141: if (ph->ph_nmissing == 0) {
1.6 thorpej 1142: pp->pr_nidle++;
1.90 thorpej 1143: if (pp->pr_npages > pp->pr_minpages &&
1144: (pp->pr_npages > pp->pr_maxpages ||
1145: (pp->pr_alloc->pa_flags & PA_WANT) != 0)) {
1.101 thorpej 1146: pr_rmpage(pp, ph, pq);
1.3 pk 1147: } else {
1.88 chs 1148: LIST_REMOVE(ph, ph_pagelist);
1149: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.3 pk 1150:
1.21 thorpej 1151: /*
1152: * Update the timestamp on the page. A page must
1153: * be idle for some period of time before it can
1154: * be reclaimed by the pagedaemon. This minimizes
1155: * ping-pong'ing for memory.
1156: */
1.111.4.1! simonb 1157: getmicrotime(&ph->ph_time);
1.1 pk 1158: }
1.88 chs 1159: pool_update_curpage(pp);
1.1 pk 1160: }
1.88 chs 1161:
1.21 thorpej 1162: /*
1.88 chs 1163: * If the page was previously completely full, move it to the
1164: * partially-full list and make it the current page. The next
1165: * allocation will get the item from this page, instead of
1166: * further fragmenting the pool.
1.21 thorpej 1167: */
1168: else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
1.88 chs 1169: LIST_REMOVE(ph, ph_pagelist);
1170: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.21 thorpej 1171: pp->pr_curpage = ph;
1172: }
1.43 thorpej 1173: }
1174:
1175: /*
1176: * Return resource to the pool; must be called at appropriate spl level
1177: */
1.59 thorpej 1178: #ifdef POOL_DIAGNOSTIC
1.43 thorpej 1179: void
1180: _pool_put(struct pool *pp, void *v, const char *file, long line)
1181: {
1.101 thorpej 1182: struct pool_pagelist pq;
1183:
1184: LIST_INIT(&pq);
1.43 thorpej 1185:
1186: simple_lock(&pp->pr_slock);
1187: pr_enter(pp, file, line);
1188:
1.56 sommerfe 1189: pr_log(pp, v, PRLOG_PUT, file, line);
1190:
1.101 thorpej 1191: pool_do_put(pp, v, &pq);
1.21 thorpej 1192:
1.25 thorpej 1193: pr_leave(pp);
1.21 thorpej 1194: simple_unlock(&pp->pr_slock);
1.101 thorpej 1195:
1.102 chs 1196: pr_pagelist_free(pp, &pq);
1.1 pk 1197: }
1.57 sommerfe 1198: #undef pool_put
1.59 thorpej 1199: #endif /* POOL_DIAGNOSTIC */
1.1 pk 1200:
1.56 sommerfe 1201: void
1202: pool_put(struct pool *pp, void *v)
1203: {
1.101 thorpej 1204: struct pool_pagelist pq;
1205:
1206: LIST_INIT(&pq);
1.56 sommerfe 1207:
1208: simple_lock(&pp->pr_slock);
1.101 thorpej 1209: pool_do_put(pp, v, &pq);
1210: simple_unlock(&pp->pr_slock);
1.56 sommerfe 1211:
1.102 chs 1212: pr_pagelist_free(pp, &pq);
1.56 sommerfe 1213: }
1.57 sommerfe 1214:
1.59 thorpej 1215: #ifdef POOL_DIAGNOSTIC
1.57 sommerfe 1216: #define pool_put(h, v) _pool_put((h), (v), __FILE__, __LINE__)
1.56 sommerfe 1217: #endif
1.74 thorpej 1218:
1219: /*
1220: * Add N items to the pool.
1221: */
1222: int
1223: pool_prime(struct pool *pp, int n)
1224: {
1.83 scw 1225: struct pool_item_header *ph = NULL;
1.74 thorpej 1226: caddr_t cp;
1.75 simonb 1227: int newpages;
1.74 thorpej 1228:
1229: simple_lock(&pp->pr_slock);
1230:
1231: newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1232:
1233: while (newpages-- > 0) {
1234: simple_unlock(&pp->pr_slock);
1235: cp = pool_allocator_alloc(pp, PR_NOWAIT);
1236: if (__predict_true(cp != NULL))
1237: ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1238:
1239: if (__predict_false(cp == NULL || ph == NULL)) {
1240: if (cp != NULL)
1241: pool_allocator_free(pp, cp);
1.91 yamt 1242: simple_lock(&pp->pr_slock);
1.74 thorpej 1243: break;
1244: }
1245:
1.91 yamt 1246: simple_lock(&pp->pr_slock);
1.74 thorpej 1247: pool_prime_page(pp, cp, ph);
1248: pp->pr_npagealloc++;
1249: pp->pr_minpages++;
1250: }
1251:
1252: if (pp->pr_minpages >= pp->pr_maxpages)
1253: pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */
1254:
1255: simple_unlock(&pp->pr_slock);
1256: return (0);
1257: }
1.55 thorpej 1258:
1259: /*
1.3 pk 1260: * Add a page worth of items to the pool.
1.21 thorpej 1261: *
1262: * Note, we must be called with the pool descriptor LOCKED.
1.3 pk 1263: */
1.55 thorpej 1264: static void
1265: pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph)
1.3 pk 1266: {
1267: struct pool_item *pi;
1268: caddr_t cp = storage;
1269: unsigned int align = pp->pr_align;
1270: unsigned int ioff = pp->pr_itemoffset;
1.55 thorpej 1271: int n;
1.36 pk 1272:
1.91 yamt 1273: LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
1274:
1.66 thorpej 1275: #ifdef DIAGNOSTIC
1276: if (((u_long)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
1.36 pk 1277: panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1.66 thorpej 1278: #endif
1.3 pk 1279:
1280: /*
1281: * Insert page header.
1282: */
1.88 chs 1283: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.102 chs 1284: LIST_INIT(&ph->ph_itemlist);
1.3 pk 1285: ph->ph_page = storage;
1286: ph->ph_nmissing = 0;
1.111.4.1! simonb 1287: getmicrotime(&ph->ph_time);
1.88 chs 1288: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1289: SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
1.3 pk 1290:
1.6 thorpej 1291: pp->pr_nidle++;
1292:
1.3 pk 1293: /*
1294: * Color this page.
1295: */
1296: cp = (caddr_t)(cp + pp->pr_curcolor);
1297: if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
1298: pp->pr_curcolor = 0;
1299:
1300: /*
1301: * Adjust storage to apply aligment to `pr_itemoffset' in each item.
1302: */
1303: if (ioff != 0)
1304: cp = (caddr_t)(cp + (align - ioff));
1305:
1306: /*
1307: * Insert remaining chunks on the bucket list.
1308: */
1309: n = pp->pr_itemsperpage;
1.20 thorpej 1310: pp->pr_nitems += n;
1.3 pk 1311:
1.97 yamt 1312: if (pp->pr_roflags & PR_NOTOUCH) {
1.99 yamt 1313: pool_item_freelist_t *freelist = PR_FREELIST(ph);
1.97 yamt 1314: int i;
1315:
1.99 yamt 1316: ph->ph_off = cp - storage;
1.97 yamt 1317: ph->ph_firstfree = 0;
1318: for (i = 0; i < n - 1; i++)
1319: freelist[i] = i + 1;
1320: freelist[n - 1] = PR_INDEX_EOL;
1321: } else {
1322: while (n--) {
1323: pi = (struct pool_item *)cp;
1.78 thorpej 1324:
1.97 yamt 1325: KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
1.3 pk 1326:
1.97 yamt 1327: /* Insert on page list */
1.102 chs 1328: LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.3 pk 1329: #ifdef DIAGNOSTIC
1.97 yamt 1330: pi->pi_magic = PI_MAGIC;
1.3 pk 1331: #endif
1.97 yamt 1332: cp = (caddr_t)(cp + pp->pr_size);
1333: }
1.3 pk 1334: }
1335:
1336: /*
1337: * If the pool was depleted, point at the new page.
1338: */
1339: if (pp->pr_curpage == NULL)
1340: pp->pr_curpage = ph;
1341:
1342: if (++pp->pr_npages > pp->pr_hiwat)
1343: pp->pr_hiwat = pp->pr_npages;
1344: }
1345:
1.20 thorpej 1346: /*
1.52 thorpej 1347: * Used by pool_get() when nitems drops below the low water mark. This
1.88 chs 1348: * is used to catch up pr_nitems with the low water mark.
1.20 thorpej 1349: *
1.21 thorpej 1350: * Note 1, we never wait for memory here, we let the caller decide what to do.
1.20 thorpej 1351: *
1.73 thorpej 1352: * Note 2, we must be called with the pool already locked, and we return
1.20 thorpej 1353: * with it locked.
1354: */
1355: static int
1.42 thorpej 1356: pool_catchup(struct pool *pp)
1.20 thorpej 1357: {
1.83 scw 1358: struct pool_item_header *ph = NULL;
1.20 thorpej 1359: caddr_t cp;
1360: int error = 0;
1361:
1.54 thorpej 1362: while (POOL_NEEDS_CATCHUP(pp)) {
1.20 thorpej 1363: /*
1.21 thorpej 1364: * Call the page back-end allocator for more memory.
1365: *
1366: * XXX: We never wait, so should we bother unlocking
1367: * the pool descriptor?
1.20 thorpej 1368: */
1.21 thorpej 1369: simple_unlock(&pp->pr_slock);
1.66 thorpej 1370: cp = pool_allocator_alloc(pp, PR_NOWAIT);
1.55 thorpej 1371: if (__predict_true(cp != NULL))
1372: ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1373: if (__predict_false(cp == NULL || ph == NULL)) {
1374: if (cp != NULL)
1.66 thorpej 1375: pool_allocator_free(pp, cp);
1.20 thorpej 1376: error = ENOMEM;
1.91 yamt 1377: simple_lock(&pp->pr_slock);
1.20 thorpej 1378: break;
1379: }
1.91 yamt 1380: simple_lock(&pp->pr_slock);
1.55 thorpej 1381: pool_prime_page(pp, cp, ph);
1.26 thorpej 1382: pp->pr_npagealloc++;
1.20 thorpej 1383: }
1384:
1385: return (error);
1386: }
1387:
1.88 chs 1388: static void
1389: pool_update_curpage(struct pool *pp)
1390: {
1391:
1392: pp->pr_curpage = LIST_FIRST(&pp->pr_partpages);
1393: if (pp->pr_curpage == NULL) {
1394: pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages);
1395: }
1396: }
1397:
1.3 pk 1398: void
1.42 thorpej 1399: pool_setlowat(struct pool *pp, int n)
1.3 pk 1400: {
1.15 pk 1401:
1.21 thorpej 1402: simple_lock(&pp->pr_slock);
1403:
1.3 pk 1404: pp->pr_minitems = n;
1.15 pk 1405: pp->pr_minpages = (n == 0)
1406: ? 0
1.18 thorpej 1407: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.20 thorpej 1408:
1409: /* Make sure we're caught up with the newly-set low water mark. */
1.75 simonb 1410: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1411: /*
1412: * XXX: Should we log a warning? Should we set up a timeout
1413: * to try again in a second or so? The latter could break
1414: * a caller's assumptions about interrupt protection, etc.
1415: */
1416: }
1.21 thorpej 1417:
1418: simple_unlock(&pp->pr_slock);
1.3 pk 1419: }
1420:
1421: void
1.42 thorpej 1422: pool_sethiwat(struct pool *pp, int n)
1.3 pk 1423: {
1.15 pk 1424:
1.21 thorpej 1425: simple_lock(&pp->pr_slock);
1426:
1.15 pk 1427: pp->pr_maxpages = (n == 0)
1428: ? 0
1.18 thorpej 1429: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1430:
1431: simple_unlock(&pp->pr_slock);
1.3 pk 1432: }
1433:
1.20 thorpej 1434: void
1.42 thorpej 1435: pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1.20 thorpej 1436: {
1437:
1.21 thorpej 1438: simple_lock(&pp->pr_slock);
1.20 thorpej 1439:
1440: pp->pr_hardlimit = n;
1441: pp->pr_hardlimit_warning = warnmess;
1.31 thorpej 1442: pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1443: pp->pr_hardlimit_warning_last.tv_sec = 0;
1444: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.20 thorpej 1445:
1446: /*
1.21 thorpej 1447: * In-line version of pool_sethiwat(), because we don't want to
1448: * release the lock.
1.20 thorpej 1449: */
1450: pp->pr_maxpages = (n == 0)
1451: ? 0
1452: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1453:
1454: simple_unlock(&pp->pr_slock);
1.20 thorpej 1455: }
1.3 pk 1456:
1457: /*
1458: * Release all complete pages that have not been used recently.
1459: */
1.66 thorpej 1460: int
1.59 thorpej 1461: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 1462: _pool_reclaim(struct pool *pp, const char *file, long line)
1.56 sommerfe 1463: #else
1464: pool_reclaim(struct pool *pp)
1465: #endif
1.3 pk 1466: {
1467: struct pool_item_header *ph, *phnext;
1.43 thorpej 1468: struct pool_cache *pc;
1.61 chs 1469: struct pool_pagelist pq;
1.102 chs 1470: struct pool_cache_grouplist pcgl;
1471: struct timeval curtime, diff;
1.3 pk 1472:
1.68 thorpej 1473: if (pp->pr_drain_hook != NULL) {
1474: /*
1475: * The drain hook must be called with the pool unlocked.
1476: */
1477: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT);
1478: }
1479:
1.21 thorpej 1480: if (simple_lock_try(&pp->pr_slock) == 0)
1.66 thorpej 1481: return (0);
1.25 thorpej 1482: pr_enter(pp, file, line);
1.68 thorpej 1483:
1.88 chs 1484: LIST_INIT(&pq);
1.102 chs 1485: LIST_INIT(&pcgl);
1.3 pk 1486:
1.43 thorpej 1487: /*
1488: * Reclaim items from the pool's caches.
1489: */
1.102 chs 1490: LIST_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
1491: pool_cache_reclaim(pc, &pq, &pcgl);
1.43 thorpej 1492:
1.111.4.1! simonb 1493: getmicrotime(&curtime);
1.21 thorpej 1494:
1.88 chs 1495: for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
1496: phnext = LIST_NEXT(ph, ph_pagelist);
1.3 pk 1497:
1498: /* Check our minimum page claim */
1499: if (pp->pr_npages <= pp->pr_minpages)
1500: break;
1501:
1.88 chs 1502: KASSERT(ph->ph_nmissing == 0);
1503: timersub(&curtime, &ph->ph_time, &diff);
1504: if (diff.tv_sec < pool_inactive_time)
1505: continue;
1.21 thorpej 1506:
1.88 chs 1507: /*
1508: * If freeing this page would put us below
1509: * the low water mark, stop now.
1510: */
1511: if ((pp->pr_nitems - pp->pr_itemsperpage) <
1512: pp->pr_minitems)
1513: break;
1.21 thorpej 1514:
1.88 chs 1515: pr_rmpage(pp, ph, &pq);
1.3 pk 1516: }
1517:
1.25 thorpej 1518: pr_leave(pp);
1.21 thorpej 1519: simple_unlock(&pp->pr_slock);
1.102 chs 1520: if (LIST_EMPTY(&pq) && LIST_EMPTY(&pcgl))
1521: return 0;
1.66 thorpej 1522:
1.101 thorpej 1523: pr_pagelist_free(pp, &pq);
1.102 chs 1524: pcg_grouplist_free(&pcgl);
1.66 thorpej 1525: return (1);
1.3 pk 1526: }
1527:
1528: /*
1529: * Drain pools, one at a time.
1.21 thorpej 1530: *
1531: * Note, we must never be called from an interrupt context.
1.3 pk 1532: */
1533: void
1.42 thorpej 1534: pool_drain(void *arg)
1.3 pk 1535: {
1536: struct pool *pp;
1.23 thorpej 1537: int s;
1.3 pk 1538:
1.61 chs 1539: pp = NULL;
1.49 thorpej 1540: s = splvm();
1.23 thorpej 1541: simple_lock(&pool_head_slock);
1.61 chs 1542: if (drainpp == NULL) {
1.102 chs 1543: drainpp = LIST_FIRST(&pool_head);
1.61 chs 1544: }
1545: if (drainpp) {
1546: pp = drainpp;
1.102 chs 1547: drainpp = LIST_NEXT(pp, pr_poollist);
1.61 chs 1548: }
1549: simple_unlock(&pool_head_slock);
1.63 chs 1550: pool_reclaim(pp);
1.61 chs 1551: splx(s);
1.3 pk 1552: }
1553:
1554: /*
1555: * Diagnostic helpers.
1556: */
1557: void
1.42 thorpej 1558: pool_print(struct pool *pp, const char *modif)
1.21 thorpej 1559: {
1560: int s;
1561:
1.49 thorpej 1562: s = splvm();
1.25 thorpej 1563: if (simple_lock_try(&pp->pr_slock) == 0) {
1564: printf("pool %s is locked; try again later\n",
1565: pp->pr_wchan);
1566: splx(s);
1567: return;
1568: }
1569: pool_print1(pp, modif, printf);
1.21 thorpej 1570: simple_unlock(&pp->pr_slock);
1571: splx(s);
1572: }
1573:
1.25 thorpej 1574: void
1.108 yamt 1575: pool_printall(const char *modif, void (*pr)(const char *, ...))
1576: {
1577: struct pool *pp;
1578:
1579: if (simple_lock_try(&pool_head_slock) == 0) {
1580: (*pr)("WARNING: pool_head_slock is locked\n");
1581: } else {
1582: simple_unlock(&pool_head_slock);
1583: }
1584:
1585: LIST_FOREACH(pp, &pool_head, pr_poollist) {
1586: pool_printit(pp, modif, pr);
1587: }
1588: }
1589:
1590: void
1.42 thorpej 1591: pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.25 thorpej 1592: {
1593:
1594: if (pp == NULL) {
1595: (*pr)("Must specify a pool to print.\n");
1596: return;
1597: }
1598:
1599: /*
1600: * Called from DDB; interrupts should be blocked, and all
1601: * other processors should be paused. We can skip locking
1602: * the pool in this case.
1603: *
1604: * We do a simple_lock_try() just to print the lock
1605: * status, however.
1606: */
1607:
1608: if (simple_lock_try(&pp->pr_slock) == 0)
1609: (*pr)("WARNING: pool %s is locked\n", pp->pr_wchan);
1610: else
1.107 yamt 1611: simple_unlock(&pp->pr_slock);
1.25 thorpej 1612:
1613: pool_print1(pp, modif, pr);
1614: }
1615:
1.21 thorpej 1616: static void
1.97 yamt 1617: pool_print_pagelist(struct pool *pp, struct pool_pagelist *pl,
1618: void (*pr)(const char *, ...))
1.88 chs 1619: {
1620: struct pool_item_header *ph;
1621: #ifdef DIAGNOSTIC
1622: struct pool_item *pi;
1623: #endif
1624:
1625: LIST_FOREACH(ph, pl, ph_pagelist) {
1626: (*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n",
1627: ph->ph_page, ph->ph_nmissing,
1628: (u_long)ph->ph_time.tv_sec,
1629: (u_long)ph->ph_time.tv_usec);
1630: #ifdef DIAGNOSTIC
1.97 yamt 1631: if (!(pp->pr_roflags & PR_NOTOUCH)) {
1.102 chs 1632: LIST_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1.97 yamt 1633: if (pi->pi_magic != PI_MAGIC) {
1634: (*pr)("\t\t\titem %p, magic 0x%x\n",
1635: pi, pi->pi_magic);
1636: }
1.88 chs 1637: }
1638: }
1639: #endif
1640: }
1641: }
1642:
1643: static void
1.42 thorpej 1644: pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.3 pk 1645: {
1.25 thorpej 1646: struct pool_item_header *ph;
1.44 thorpej 1647: struct pool_cache *pc;
1648: struct pool_cache_group *pcg;
1649: int i, print_log = 0, print_pagelist = 0, print_cache = 0;
1.25 thorpej 1650: char c;
1651:
1652: while ((c = *modif++) != '\0') {
1653: if (c == 'l')
1654: print_log = 1;
1655: if (c == 'p')
1656: print_pagelist = 1;
1.44 thorpej 1657: if (c == 'c')
1658: print_cache = 1;
1.25 thorpej 1659: }
1660:
1661: (*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1662: pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1663: pp->pr_roflags);
1.66 thorpej 1664: (*pr)("\talloc %p\n", pp->pr_alloc);
1.25 thorpej 1665: (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1666: pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1667: (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1668: pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1669:
1670: (*pr)("\n\tnget %lu, nfail %lu, nput %lu\n",
1671: pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1672: (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1673: pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1674:
1675: if (print_pagelist == 0)
1676: goto skip_pagelist;
1677:
1.88 chs 1678: if ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1679: (*pr)("\n\tempty page list:\n");
1.97 yamt 1680: pool_print_pagelist(pp, &pp->pr_emptypages, pr);
1.88 chs 1681: if ((ph = LIST_FIRST(&pp->pr_fullpages)) != NULL)
1682: (*pr)("\n\tfull page list:\n");
1.97 yamt 1683: pool_print_pagelist(pp, &pp->pr_fullpages, pr);
1.88 chs 1684: if ((ph = LIST_FIRST(&pp->pr_partpages)) != NULL)
1685: (*pr)("\n\tpartial-page list:\n");
1.97 yamt 1686: pool_print_pagelist(pp, &pp->pr_partpages, pr);
1.88 chs 1687:
1.25 thorpej 1688: if (pp->pr_curpage == NULL)
1689: (*pr)("\tno current page\n");
1690: else
1691: (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1692:
1693: skip_pagelist:
1694: if (print_log == 0)
1695: goto skip_log;
1696:
1697: (*pr)("\n");
1698: if ((pp->pr_roflags & PR_LOGGING) == 0)
1699: (*pr)("\tno log\n");
1700: else
1701: pr_printlog(pp, NULL, pr);
1.3 pk 1702:
1.25 thorpej 1703: skip_log:
1.44 thorpej 1704: if (print_cache == 0)
1705: goto skip_cache;
1706:
1.102 chs 1707: #define PR_GROUPLIST(pcg) \
1708: (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail); \
1709: for (i = 0; i < PCG_NOBJECTS; i++) { \
1710: if (pcg->pcg_objects[i].pcgo_pa != \
1711: POOL_PADDR_INVALID) { \
1712: (*pr)("\t\t\t%p, 0x%llx\n", \
1713: pcg->pcg_objects[i].pcgo_va, \
1714: (unsigned long long) \
1715: pcg->pcg_objects[i].pcgo_pa); \
1716: } else { \
1717: (*pr)("\t\t\t%p\n", \
1718: pcg->pcg_objects[i].pcgo_va); \
1719: } \
1720: }
1721:
1722: LIST_FOREACH(pc, &pp->pr_cachelist, pc_poollist) {
1.103 chs 1723: (*pr)("\tcache %p\n", pc);
1.48 thorpej 1724: (*pr)("\t hits %lu misses %lu ngroups %lu nitems %lu\n",
1725: pc->pc_hits, pc->pc_misses, pc->pc_ngroups, pc->pc_nitems);
1.102 chs 1726: (*pr)("\t full groups:\n");
1.103 chs 1727: LIST_FOREACH(pcg, &pc->pc_fullgroups, pcg_list) {
1.102 chs 1728: PR_GROUPLIST(pcg);
1.103 chs 1729: }
1.102 chs 1730: (*pr)("\t partial groups:\n");
1.103 chs 1731: LIST_FOREACH(pcg, &pc->pc_partgroups, pcg_list) {
1.102 chs 1732: PR_GROUPLIST(pcg);
1.103 chs 1733: }
1.102 chs 1734: (*pr)("\t empty groups:\n");
1.103 chs 1735: LIST_FOREACH(pcg, &pc->pc_emptygroups, pcg_list) {
1.102 chs 1736: PR_GROUPLIST(pcg);
1.103 chs 1737: }
1.44 thorpej 1738: }
1.102 chs 1739: #undef PR_GROUPLIST
1.44 thorpej 1740:
1741: skip_cache:
1.88 chs 1742: pr_enter_check(pp, pr);
1743: }
1744:
1745: static int
1746: pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
1747: {
1748: struct pool_item *pi;
1749: caddr_t page;
1750: int n;
1751:
1752: page = (caddr_t)((u_long)ph & pp->pr_alloc->pa_pagemask);
1753: if (page != ph->ph_page &&
1754: (pp->pr_roflags & PR_PHINPAGE) != 0) {
1755: if (label != NULL)
1756: printf("%s: ", label);
1757: printf("pool(%p:%s): page inconsistency: page %p;"
1758: " at page head addr %p (p %p)\n", pp,
1759: pp->pr_wchan, ph->ph_page,
1760: ph, page);
1761: return 1;
1762: }
1.3 pk 1763:
1.97 yamt 1764: if ((pp->pr_roflags & PR_NOTOUCH) != 0)
1765: return 0;
1766:
1.102 chs 1767: for (pi = LIST_FIRST(&ph->ph_itemlist), n = 0;
1.88 chs 1768: pi != NULL;
1.102 chs 1769: pi = LIST_NEXT(pi,pi_list), n++) {
1.88 chs 1770:
1771: #ifdef DIAGNOSTIC
1772: if (pi->pi_magic != PI_MAGIC) {
1773: if (label != NULL)
1774: printf("%s: ", label);
1775: printf("pool(%s): free list modified: magic=%x;"
1776: " page %p; item ordinal %d;"
1777: " addr %p (p %p)\n",
1778: pp->pr_wchan, pi->pi_magic, ph->ph_page,
1779: n, pi, page);
1780: panic("pool");
1781: }
1782: #endif
1783: page =
1784: (caddr_t)((u_long)pi & pp->pr_alloc->pa_pagemask);
1785: if (page == ph->ph_page)
1786: continue;
1787:
1788: if (label != NULL)
1789: printf("%s: ", label);
1790: printf("pool(%p:%s): page inconsistency: page %p;"
1791: " item ordinal %d; addr %p (p %p)\n", pp,
1792: pp->pr_wchan, ph->ph_page,
1793: n, pi, page);
1794: return 1;
1795: }
1796: return 0;
1.3 pk 1797: }
1798:
1.88 chs 1799:
1.3 pk 1800: int
1.42 thorpej 1801: pool_chk(struct pool *pp, const char *label)
1.3 pk 1802: {
1803: struct pool_item_header *ph;
1804: int r = 0;
1805:
1.21 thorpej 1806: simple_lock(&pp->pr_slock);
1.88 chs 1807: LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
1808: r = pool_chk_page(pp, label, ph);
1809: if (r) {
1810: goto out;
1811: }
1812: }
1813: LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
1814: r = pool_chk_page(pp, label, ph);
1815: if (r) {
1.3 pk 1816: goto out;
1817: }
1.88 chs 1818: }
1819: LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
1820: r = pool_chk_page(pp, label, ph);
1821: if (r) {
1.3 pk 1822: goto out;
1823: }
1824: }
1.88 chs 1825:
1.3 pk 1826: out:
1.21 thorpej 1827: simple_unlock(&pp->pr_slock);
1.3 pk 1828: return (r);
1.43 thorpej 1829: }
1830:
1831: /*
1832: * pool_cache_init:
1833: *
1834: * Initialize a pool cache.
1835: *
1836: * NOTE: If the pool must be protected from interrupts, we expect
1837: * to be called at the appropriate interrupt priority level.
1838: */
1839: void
1840: pool_cache_init(struct pool_cache *pc, struct pool *pp,
1841: int (*ctor)(void *, void *, int),
1842: void (*dtor)(void *, void *),
1843: void *arg)
1844: {
1845:
1.102 chs 1846: LIST_INIT(&pc->pc_emptygroups);
1847: LIST_INIT(&pc->pc_fullgroups);
1848: LIST_INIT(&pc->pc_partgroups);
1.43 thorpej 1849: simple_lock_init(&pc->pc_slock);
1850:
1851: pc->pc_pool = pp;
1852:
1853: pc->pc_ctor = ctor;
1854: pc->pc_dtor = dtor;
1855: pc->pc_arg = arg;
1856:
1.48 thorpej 1857: pc->pc_hits = 0;
1858: pc->pc_misses = 0;
1859:
1860: pc->pc_ngroups = 0;
1861:
1862: pc->pc_nitems = 0;
1863:
1.43 thorpej 1864: simple_lock(&pp->pr_slock);
1.102 chs 1865: LIST_INSERT_HEAD(&pp->pr_cachelist, pc, pc_poollist);
1.43 thorpej 1866: simple_unlock(&pp->pr_slock);
1867: }
1868:
1869: /*
1870: * pool_cache_destroy:
1871: *
1872: * Destroy a pool cache.
1873: */
1874: void
1875: pool_cache_destroy(struct pool_cache *pc)
1876: {
1877: struct pool *pp = pc->pc_pool;
1878:
1879: /* First, invalidate the entire cache. */
1880: pool_cache_invalidate(pc);
1881:
1882: /* ...and remove it from the pool's cache list. */
1883: simple_lock(&pp->pr_slock);
1.102 chs 1884: LIST_REMOVE(pc, pc_poollist);
1.43 thorpej 1885: simple_unlock(&pp->pr_slock);
1886: }
1887:
1.110 perry 1888: static inline void *
1.87 thorpej 1889: pcg_get(struct pool_cache_group *pcg, paddr_t *pap)
1.43 thorpej 1890: {
1891: void *object;
1892: u_int idx;
1893:
1894: KASSERT(pcg->pcg_avail <= PCG_NOBJECTS);
1.45 thorpej 1895: KASSERT(pcg->pcg_avail != 0);
1.43 thorpej 1896: idx = --pcg->pcg_avail;
1897:
1.87 thorpej 1898: KASSERT(pcg->pcg_objects[idx].pcgo_va != NULL);
1899: object = pcg->pcg_objects[idx].pcgo_va;
1900: if (pap != NULL)
1901: *pap = pcg->pcg_objects[idx].pcgo_pa;
1902: pcg->pcg_objects[idx].pcgo_va = NULL;
1.43 thorpej 1903:
1904: return (object);
1905: }
1906:
1.110 perry 1907: static inline void
1.87 thorpej 1908: pcg_put(struct pool_cache_group *pcg, void *object, paddr_t pa)
1.43 thorpej 1909: {
1910: u_int idx;
1911:
1912: KASSERT(pcg->pcg_avail < PCG_NOBJECTS);
1913: idx = pcg->pcg_avail++;
1914:
1.87 thorpej 1915: KASSERT(pcg->pcg_objects[idx].pcgo_va == NULL);
1916: pcg->pcg_objects[idx].pcgo_va = object;
1917: pcg->pcg_objects[idx].pcgo_pa = pa;
1.43 thorpej 1918: }
1919:
1.102 chs 1920: static void
1921: pcg_grouplist_free(struct pool_cache_grouplist *pcgl)
1922: {
1923: struct pool_cache_group *pcg;
1924: int s;
1925:
1926: s = splvm();
1927: while ((pcg = LIST_FIRST(pcgl)) != NULL) {
1928: LIST_REMOVE(pcg, pcg_list);
1929: pool_put(&pcgpool, pcg);
1930: }
1931: splx(s);
1932: }
1933:
1.43 thorpej 1934: /*
1.87 thorpej 1935: * pool_cache_get{,_paddr}:
1.43 thorpej 1936: *
1.87 thorpej 1937: * Get an object from a pool cache (optionally returning
1938: * the physical address of the object).
1.43 thorpej 1939: */
1940: void *
1.87 thorpej 1941: pool_cache_get_paddr(struct pool_cache *pc, int flags, paddr_t *pap)
1.43 thorpej 1942: {
1943: struct pool_cache_group *pcg;
1944: void *object;
1.58 thorpej 1945:
1946: #ifdef LOCKDEBUG
1947: if (flags & PR_WAITOK)
1948: simple_lock_only_held(NULL, "pool_cache_get(PR_WAITOK)");
1949: #endif
1.43 thorpej 1950:
1951: simple_lock(&pc->pc_slock);
1952:
1.102 chs 1953: pcg = LIST_FIRST(&pc->pc_partgroups);
1954: if (pcg == NULL) {
1955: pcg = LIST_FIRST(&pc->pc_fullgroups);
1956: if (pcg != NULL) {
1957: LIST_REMOVE(pcg, pcg_list);
1958: LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list);
1.43 thorpej 1959: }
1.102 chs 1960: }
1961: if (pcg == NULL) {
1.43 thorpej 1962:
1963: /*
1964: * No groups with any available objects. Allocate
1965: * a new object, construct it, and return it to
1966: * the caller. We will allocate a group, if necessary,
1967: * when the object is freed back to the cache.
1968: */
1.48 thorpej 1969: pc->pc_misses++;
1.43 thorpej 1970: simple_unlock(&pc->pc_slock);
1971: object = pool_get(pc->pc_pool, flags);
1972: if (object != NULL && pc->pc_ctor != NULL) {
1973: if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) {
1974: pool_put(pc->pc_pool, object);
1975: return (NULL);
1976: }
1977: }
1.87 thorpej 1978: if (object != NULL && pap != NULL) {
1979: #ifdef POOL_VTOPHYS
1980: *pap = POOL_VTOPHYS(object);
1981: #else
1982: *pap = POOL_PADDR_INVALID;
1983: #endif
1984: }
1.43 thorpej 1985: return (object);
1986: }
1987:
1.48 thorpej 1988: pc->pc_hits++;
1989: pc->pc_nitems--;
1.87 thorpej 1990: object = pcg_get(pcg, pap);
1.43 thorpej 1991:
1.102 chs 1992: if (pcg->pcg_avail == 0) {
1993: LIST_REMOVE(pcg, pcg_list);
1994: LIST_INSERT_HEAD(&pc->pc_emptygroups, pcg, pcg_list);
1995: }
1.43 thorpej 1996: simple_unlock(&pc->pc_slock);
1997:
1998: return (object);
1999: }
2000:
2001: /*
1.87 thorpej 2002: * pool_cache_put{,_paddr}:
1.43 thorpej 2003: *
1.87 thorpej 2004: * Put an object back to the pool cache (optionally caching the
2005: * physical address of the object).
1.43 thorpej 2006: */
2007: void
1.87 thorpej 2008: pool_cache_put_paddr(struct pool_cache *pc, void *object, paddr_t pa)
1.43 thorpej 2009: {
2010: struct pool_cache_group *pcg;
1.60 thorpej 2011: int s;
1.43 thorpej 2012:
1.109 christos 2013: if (__predict_false((pc->pc_pool->pr_flags & PR_WANTED) != 0)) {
2014: goto destruct;
2015: }
2016:
1.43 thorpej 2017: simple_lock(&pc->pc_slock);
2018:
1.102 chs 2019: pcg = LIST_FIRST(&pc->pc_partgroups);
2020: if (pcg == NULL) {
2021: pcg = LIST_FIRST(&pc->pc_emptygroups);
2022: if (pcg != NULL) {
2023: LIST_REMOVE(pcg, pcg_list);
2024: LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list);
1.43 thorpej 2025: }
1.102 chs 2026: }
2027: if (pcg == NULL) {
1.43 thorpej 2028:
2029: /*
2030: * No empty groups to free the object to. Attempt to
1.47 thorpej 2031: * allocate one.
1.43 thorpej 2032: */
1.47 thorpej 2033: simple_unlock(&pc->pc_slock);
1.60 thorpej 2034: s = splvm();
1.43 thorpej 2035: pcg = pool_get(&pcgpool, PR_NOWAIT);
1.60 thorpej 2036: splx(s);
1.102 chs 2037: if (pcg == NULL) {
1.109 christos 2038: destruct:
1.102 chs 2039:
2040: /*
2041: * Unable to allocate a cache group; destruct the object
2042: * and free it back to the pool.
2043: */
2044: pool_cache_destruct_object(pc, object);
2045: return;
1.43 thorpej 2046: }
1.102 chs 2047: memset(pcg, 0, sizeof(*pcg));
2048: simple_lock(&pc->pc_slock);
2049: pc->pc_ngroups++;
2050: LIST_INSERT_HEAD(&pc->pc_partgroups, pcg, pcg_list);
1.43 thorpej 2051: }
2052:
1.48 thorpej 2053: pc->pc_nitems++;
1.87 thorpej 2054: pcg_put(pcg, object, pa);
1.43 thorpej 2055:
1.102 chs 2056: if (pcg->pcg_avail == PCG_NOBJECTS) {
2057: LIST_REMOVE(pcg, pcg_list);
2058: LIST_INSERT_HEAD(&pc->pc_fullgroups, pcg, pcg_list);
2059: }
1.43 thorpej 2060: simple_unlock(&pc->pc_slock);
1.51 thorpej 2061: }
2062:
2063: /*
2064: * pool_cache_destruct_object:
2065: *
2066: * Force destruction of an object and its release back into
2067: * the pool.
2068: */
2069: void
2070: pool_cache_destruct_object(struct pool_cache *pc, void *object)
2071: {
2072:
2073: if (pc->pc_dtor != NULL)
2074: (*pc->pc_dtor)(pc->pc_arg, object);
2075: pool_put(pc->pc_pool, object);
1.43 thorpej 2076: }
2077:
1.102 chs 2078: static void
1.106 christos 2079: pool_do_cache_invalidate_grouplist(struct pool_cache_grouplist *pcgsl,
1.105 christos 2080: struct pool_cache *pc, struct pool_pagelist *pq,
1.106 christos 2081: struct pool_cache_grouplist *pcgdl)
1.102 chs 2082: {
1.106 christos 2083: struct pool_cache_group *pcg, *npcg;
1.102 chs 2084: void *object;
2085:
1.106 christos 2086: for (pcg = LIST_FIRST(pcgsl); pcg != NULL; pcg = npcg) {
1.102 chs 2087: npcg = LIST_NEXT(pcg, pcg_list);
2088: while (pcg->pcg_avail != 0) {
2089: pc->pc_nitems--;
2090: object = pcg_get(pcg, NULL);
2091: if (pc->pc_dtor != NULL)
2092: (*pc->pc_dtor)(pc->pc_arg, object);
2093: pool_do_put(pc->pc_pool, object, pq);
2094: }
1.103 chs 2095: pc->pc_ngroups--;
1.102 chs 2096: LIST_REMOVE(pcg, pcg_list);
1.106 christos 2097: LIST_INSERT_HEAD(pcgdl, pcg, pcg_list);
1.102 chs 2098: }
1.105 christos 2099: }
2100:
2101: static void
2102: pool_do_cache_invalidate(struct pool_cache *pc, struct pool_pagelist *pq,
2103: struct pool_cache_grouplist *pcgl)
2104: {
2105:
2106: LOCK_ASSERT(simple_lock_held(&pc->pc_slock));
2107: LOCK_ASSERT(simple_lock_held(&pc->pc_pool->pr_slock));
2108:
1.106 christos 2109: pool_do_cache_invalidate_grouplist(&pc->pc_fullgroups, pc, pq, pcgl);
2110: pool_do_cache_invalidate_grouplist(&pc->pc_partgroups, pc, pq, pcgl);
1.103 chs 2111:
2112: KASSERT(LIST_EMPTY(&pc->pc_partgroups));
2113: KASSERT(LIST_EMPTY(&pc->pc_fullgroups));
2114: KASSERT(pc->pc_nitems == 0);
1.102 chs 2115: }
2116:
1.43 thorpej 2117: /*
1.101 thorpej 2118: * pool_cache_invalidate:
1.43 thorpej 2119: *
1.101 thorpej 2120: * Invalidate a pool cache (destruct and release all of the
2121: * cached objects).
1.43 thorpej 2122: */
1.101 thorpej 2123: void
2124: pool_cache_invalidate(struct pool_cache *pc)
1.43 thorpej 2125: {
1.101 thorpej 2126: struct pool_pagelist pq;
1.102 chs 2127: struct pool_cache_grouplist pcgl;
1.101 thorpej 2128:
2129: LIST_INIT(&pq);
1.102 chs 2130: LIST_INIT(&pcgl);
1.101 thorpej 2131:
2132: simple_lock(&pc->pc_slock);
2133: simple_lock(&pc->pc_pool->pr_slock);
1.43 thorpej 2134:
1.102 chs 2135: pool_do_cache_invalidate(pc, &pq, &pcgl);
1.43 thorpej 2136:
1.101 thorpej 2137: simple_unlock(&pc->pc_pool->pr_slock);
2138: simple_unlock(&pc->pc_slock);
1.43 thorpej 2139:
1.102 chs 2140: pr_pagelist_free(pc->pc_pool, &pq);
2141: pcg_grouplist_free(&pcgl);
1.43 thorpej 2142: }
2143:
2144: /*
2145: * pool_cache_reclaim:
2146: *
2147: * Reclaim a pool cache for pool_reclaim().
2148: */
2149: static void
1.102 chs 2150: pool_cache_reclaim(struct pool_cache *pc, struct pool_pagelist *pq,
2151: struct pool_cache_grouplist *pcgl)
1.43 thorpej 2152: {
1.101 thorpej 2153:
2154: /*
2155: * We're locking in the wrong order (normally pool_cache -> pool,
2156: * but the pool is already locked when we get here), so we have
2157: * to use trylock. If we can't lock the pool_cache, it's not really
2158: * a big deal here.
2159: */
2160: if (simple_lock_try(&pc->pc_slock) == 0)
2161: return;
2162:
1.102 chs 2163: pool_do_cache_invalidate(pc, pq, pcgl);
1.43 thorpej 2164:
2165: simple_unlock(&pc->pc_slock);
1.3 pk 2166: }
1.66 thorpej 2167:
2168: /*
2169: * Pool backend allocators.
2170: *
2171: * Each pool has a backend allocator that handles allocation, deallocation,
2172: * and any additional draining that might be needed.
2173: *
2174: * We provide two standard allocators:
2175: *
2176: * pool_allocator_kmem - the default when no allocator is specified
2177: *
2178: * pool_allocator_nointr - used for pools that will not be accessed
2179: * in interrupt context.
2180: */
2181: void *pool_page_alloc(struct pool *, int);
2182: void pool_page_free(struct pool *, void *);
2183:
2184: struct pool_allocator pool_allocator_kmem = {
2185: pool_page_alloc, pool_page_free, 0,
2186: };
2187:
2188: void *pool_page_alloc_nointr(struct pool *, int);
2189: void pool_page_free_nointr(struct pool *, void *);
2190:
2191: struct pool_allocator pool_allocator_nointr = {
2192: pool_page_alloc_nointr, pool_page_free_nointr, 0,
2193: };
2194:
2195: #ifdef POOL_SUBPAGE
2196: void *pool_subpage_alloc(struct pool *, int);
2197: void pool_subpage_free(struct pool *, void *);
2198:
2199: struct pool_allocator pool_allocator_kmem_subpage = {
2200: pool_subpage_alloc, pool_subpage_free, 0,
2201: };
2202: #endif /* POOL_SUBPAGE */
2203:
2204: /*
2205: * We have at least three different resources for the same allocation and
2206: * each resource can be depleted. First, we have the ready elements in the
2207: * pool. Then we have the resource (typically a vm_map) for this allocator.
2208: * Finally, we have physical memory. Waiting for any of these can be
2209: * unnecessary when any other is freed, but the kernel doesn't support
2210: * sleeping on multiple wait channels, so we have to employ another strategy.
2211: *
2212: * The caller sleeps on the pool (so that it can be awakened when an item
2213: * is returned to the pool), but we set PA_WANT on the allocator. When a
2214: * page is returned to the allocator and PA_WANT is set, pool_allocator_free
2215: * will wake up all sleeping pools belonging to this allocator.
2216: *
2217: * XXX Thundering herd.
2218: */
2219: void *
2220: pool_allocator_alloc(struct pool *org, int flags)
2221: {
2222: struct pool_allocator *pa = org->pr_alloc;
2223: struct pool *pp, *start;
2224: int s, freed;
2225: void *res;
2226:
1.91 yamt 2227: LOCK_ASSERT(!simple_lock_held(&org->pr_slock));
2228:
1.66 thorpej 2229: do {
2230: if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
2231: return (res);
1.68 thorpej 2232: if ((flags & PR_WAITOK) == 0) {
2233: /*
2234: * We only run the drain hookhere if PR_NOWAIT.
2235: * In other cases, the hook will be run in
2236: * pool_reclaim().
2237: */
2238: if (org->pr_drain_hook != NULL) {
2239: (*org->pr_drain_hook)(org->pr_drain_hook_arg,
2240: flags);
2241: if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
2242: return (res);
2243: }
1.66 thorpej 2244: break;
1.68 thorpej 2245: }
1.66 thorpej 2246:
2247: /*
1.109 christos 2248: * Drain all pools, that use this allocator.
2249: * We do this to reclaim VA space.
1.66 thorpej 2250: * pa_alloc is responsible for waiting for
2251: * physical memory.
2252: *
2253: * XXX We risk looping forever if start if someone
2254: * calls pool_destroy on "start". But there is no
2255: * other way to have potentially sleeping pool_reclaim,
2256: * non-sleeping locks on pool_allocator, and some
2257: * stirring of drained pools in the allocator.
1.68 thorpej 2258: *
2259: * XXX Maybe we should use pool_head_slock for locking
2260: * the allocators?
1.66 thorpej 2261: */
2262: freed = 0;
2263:
2264: s = splvm();
2265: simple_lock(&pa->pa_slock);
2266: pp = start = TAILQ_FIRST(&pa->pa_list);
2267: do {
2268: TAILQ_REMOVE(&pa->pa_list, pp, pr_alloc_list);
2269: TAILQ_INSERT_TAIL(&pa->pa_list, pp, pr_alloc_list);
1.73 thorpej 2270: simple_unlock(&pa->pa_slock);
1.66 thorpej 2271: freed = pool_reclaim(pp);
1.73 thorpej 2272: simple_lock(&pa->pa_slock);
1.66 thorpej 2273: } while ((pp = TAILQ_FIRST(&pa->pa_list)) != start &&
2274: freed == 0);
2275:
2276: if (freed == 0) {
2277: /*
2278: * We set PA_WANT here, the caller will most likely
2279: * sleep waiting for pages (if not, this won't hurt
2280: * that much), and there is no way to set this in
2281: * the caller without violating locking order.
2282: */
2283: pa->pa_flags |= PA_WANT;
2284: }
2285: simple_unlock(&pa->pa_slock);
2286: splx(s);
2287: } while (freed);
2288: return (NULL);
2289: }
2290:
2291: void
2292: pool_allocator_free(struct pool *pp, void *v)
2293: {
2294: struct pool_allocator *pa = pp->pr_alloc;
2295: int s;
2296:
1.91 yamt 2297: LOCK_ASSERT(!simple_lock_held(&pp->pr_slock));
2298:
1.66 thorpej 2299: (*pa->pa_free)(pp, v);
2300:
2301: s = splvm();
2302: simple_lock(&pa->pa_slock);
2303: if ((pa->pa_flags & PA_WANT) == 0) {
2304: simple_unlock(&pa->pa_slock);
2305: splx(s);
2306: return;
2307: }
2308:
2309: TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
2310: simple_lock(&pp->pr_slock);
2311: if ((pp->pr_flags & PR_WANTED) != 0) {
2312: pp->pr_flags &= ~PR_WANTED;
2313: wakeup(pp);
2314: }
1.69 thorpej 2315: simple_unlock(&pp->pr_slock);
1.66 thorpej 2316: }
2317: pa->pa_flags &= ~PA_WANT;
2318: simple_unlock(&pa->pa_slock);
2319: splx(s);
2320: }
2321:
2322: void *
2323: pool_page_alloc(struct pool *pp, int flags)
2324: {
2325: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2326:
1.100 yamt 2327: return ((void *) uvm_km_alloc_poolpage_cache(kmem_map, waitok));
1.66 thorpej 2328: }
2329:
2330: void
2331: pool_page_free(struct pool *pp, void *v)
2332: {
2333:
1.98 yamt 2334: uvm_km_free_poolpage_cache(kmem_map, (vaddr_t) v);
2335: }
2336:
2337: static void *
2338: pool_page_alloc_meta(struct pool *pp, int flags)
2339: {
2340: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2341:
1.100 yamt 2342: return ((void *) uvm_km_alloc_poolpage(kmem_map, waitok));
1.98 yamt 2343: }
2344:
2345: static void
2346: pool_page_free_meta(struct pool *pp, void *v)
2347: {
2348:
1.100 yamt 2349: uvm_km_free_poolpage(kmem_map, (vaddr_t) v);
1.66 thorpej 2350: }
2351:
2352: #ifdef POOL_SUBPAGE
2353: /* Sub-page allocator, for machines with large hardware pages. */
2354: void *
2355: pool_subpage_alloc(struct pool *pp, int flags)
2356: {
1.93 dbj 2357: void *v;
2358: int s;
2359: s = splvm();
2360: v = pool_get(&psppool, flags);
2361: splx(s);
2362: return v;
1.66 thorpej 2363: }
2364:
2365: void
2366: pool_subpage_free(struct pool *pp, void *v)
2367: {
1.93 dbj 2368: int s;
2369: s = splvm();
1.66 thorpej 2370: pool_put(&psppool, v);
1.93 dbj 2371: splx(s);
1.66 thorpej 2372: }
2373:
2374: /* We don't provide a real nointr allocator. Maybe later. */
2375: void *
2376: pool_page_alloc_nointr(struct pool *pp, int flags)
2377: {
2378:
2379: return (pool_subpage_alloc(pp, flags));
2380: }
2381:
2382: void
2383: pool_page_free_nointr(struct pool *pp, void *v)
2384: {
2385:
2386: pool_subpage_free(pp, v);
2387: }
2388: #else
2389: void *
2390: pool_page_alloc_nointr(struct pool *pp, int flags)
2391: {
2392: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2393:
1.100 yamt 2394: return ((void *) uvm_km_alloc_poolpage_cache(kernel_map, waitok));
1.66 thorpej 2395: }
2396:
2397: void
2398: pool_page_free_nointr(struct pool *pp, void *v)
2399: {
2400:
1.98 yamt 2401: uvm_km_free_poolpage_cache(kernel_map, (vaddr_t) v);
1.66 thorpej 2402: }
2403: #endif /* POOL_SUBPAGE */
CVSweb <webmaster@jp.NetBSD.org>