Annotation of src/sys/kern/subr_pool.c, Revision 1.95
1.95 ! atatat 1: /* $NetBSD: subr_pool.c,v 1.94 2004/04/25 16:42:41 simonb 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>
1.95 ! atatat 41: __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.94 2004/04/25 16:42:41 simonb 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 */
1.5 thorpej 73: TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head);
1.3 pk 74:
75: /* Private pool for page header structures */
76: static struct pool phpool;
77:
1.62 bjh21 78: #ifdef POOL_SUBPAGE
79: /* Pool of subpages for use by normal pools. */
80: static struct pool psppool;
81: #endif
82:
1.3 pk 83: /* # of seconds to retain page after last use */
84: int pool_inactive_time = 10;
85:
86: /* Next candidate for drainage (see pool_drain()) */
1.23 thorpej 87: static struct pool *drainpp;
88:
89: /* This spin lock protects both pool_head and drainpp. */
90: struct simplelock pool_head_slock = SIMPLELOCK_INITIALIZER;
1.3 pk 91:
92: struct pool_item_header {
93: /* Page headers */
1.88 chs 94: LIST_ENTRY(pool_item_header)
1.3 pk 95: ph_pagelist; /* pool page list */
96: TAILQ_HEAD(,pool_item) ph_itemlist; /* chunk list for this page */
1.88 chs 97: SPLAY_ENTRY(pool_item_header)
98: ph_node; /* Off-page page headers */
1.79 thorpej 99: unsigned int ph_nmissing; /* # of chunks in use */
1.3 pk 100: caddr_t ph_page; /* this page's address */
101: struct timeval ph_time; /* last referenced */
102: };
103:
1.1 pk 104: struct pool_item {
1.3 pk 105: #ifdef DIAGNOSTIC
1.82 thorpej 106: u_int pi_magic;
1.33 chs 107: #endif
1.82 thorpej 108: #define PI_MAGIC 0xdeadbeefU
1.3 pk 109: /* Other entries use only this list entry */
110: TAILQ_ENTRY(pool_item) pi_list;
111: };
112:
1.53 thorpej 113: #define POOL_NEEDS_CATCHUP(pp) \
114: ((pp)->pr_nitems < (pp)->pr_minitems)
115:
1.43 thorpej 116: /*
117: * Pool cache management.
118: *
119: * Pool caches provide a way for constructed objects to be cached by the
120: * pool subsystem. This can lead to performance improvements by avoiding
121: * needless object construction/destruction; it is deferred until absolutely
122: * necessary.
123: *
124: * Caches are grouped into cache groups. Each cache group references
125: * up to 16 constructed objects. When a cache allocates an object
126: * from the pool, it calls the object's constructor and places it into
127: * a cache group. When a cache group frees an object back to the pool,
128: * it first calls the object's destructor. This allows the object to
129: * persist in constructed form while freed to the cache.
130: *
131: * Multiple caches may exist for each pool. This allows a single
132: * object type to have multiple constructed forms. The pool references
133: * each cache, so that when a pool is drained by the pagedaemon, it can
134: * drain each individual cache as well. Each time a cache is drained,
135: * the most idle cache group is freed to the pool in its entirety.
136: *
137: * Pool caches are layed on top of pools. By layering them, we can avoid
138: * the complexity of cache management for pools which would not benefit
139: * from it.
140: */
141:
142: /* The cache group pool. */
143: static struct pool pcgpool;
1.3 pk 144:
1.43 thorpej 145: static void pool_cache_reclaim(struct pool_cache *);
1.3 pk 146:
1.42 thorpej 147: static int pool_catchup(struct pool *);
1.55 thorpej 148: static void pool_prime_page(struct pool *, caddr_t,
149: struct pool_item_header *);
1.88 chs 150: static void pool_update_curpage(struct pool *);
1.66 thorpej 151:
152: void *pool_allocator_alloc(struct pool *, int);
153: void pool_allocator_free(struct pool *, void *);
1.3 pk 154:
1.88 chs 155: static void pool_print_pagelist(struct pool_pagelist *,
156: void (*)(const char *, ...));
1.42 thorpej 157: static void pool_print1(struct pool *, const char *,
158: void (*)(const char *, ...));
1.3 pk 159:
1.88 chs 160: static int pool_chk_page(struct pool *, const char *,
161: struct pool_item_header *);
162:
1.3 pk 163: /*
1.52 thorpej 164: * Pool log entry. An array of these is allocated in pool_init().
1.3 pk 165: */
166: struct pool_log {
167: const char *pl_file;
168: long pl_line;
169: int pl_action;
1.25 thorpej 170: #define PRLOG_GET 1
171: #define PRLOG_PUT 2
1.3 pk 172: void *pl_addr;
1.1 pk 173: };
174:
1.86 matt 175: #ifdef POOL_DIAGNOSTIC
1.3 pk 176: /* Number of entries in pool log buffers */
1.17 thorpej 177: #ifndef POOL_LOGSIZE
178: #define POOL_LOGSIZE 10
179: #endif
180:
181: int pool_logsize = POOL_LOGSIZE;
1.1 pk 182:
1.42 thorpej 183: static __inline void
184: pr_log(struct pool *pp, void *v, int action, const char *file, long line)
1.3 pk 185: {
186: int n = pp->pr_curlogentry;
187: struct pool_log *pl;
188:
1.20 thorpej 189: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 190: return;
191:
192: /*
193: * Fill in the current entry. Wrap around and overwrite
194: * the oldest entry if necessary.
195: */
196: pl = &pp->pr_log[n];
197: pl->pl_file = file;
198: pl->pl_line = line;
199: pl->pl_action = action;
200: pl->pl_addr = v;
201: if (++n >= pp->pr_logsize)
202: n = 0;
203: pp->pr_curlogentry = n;
204: }
205:
206: static void
1.42 thorpej 207: pr_printlog(struct pool *pp, struct pool_item *pi,
208: void (*pr)(const char *, ...))
1.3 pk 209: {
210: int i = pp->pr_logsize;
211: int n = pp->pr_curlogentry;
212:
1.20 thorpej 213: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 214: return;
215:
216: /*
217: * Print all entries in this pool's log.
218: */
219: while (i-- > 0) {
220: struct pool_log *pl = &pp->pr_log[n];
221: if (pl->pl_action != 0) {
1.25 thorpej 222: if (pi == NULL || pi == pl->pl_addr) {
223: (*pr)("\tlog entry %d:\n", i);
224: (*pr)("\t\taction = %s, addr = %p\n",
225: pl->pl_action == PRLOG_GET ? "get" : "put",
226: pl->pl_addr);
227: (*pr)("\t\tfile: %s at line %lu\n",
228: pl->pl_file, pl->pl_line);
229: }
1.3 pk 230: }
231: if (++n >= pp->pr_logsize)
232: n = 0;
233: }
234: }
1.25 thorpej 235:
1.42 thorpej 236: static __inline void
237: pr_enter(struct pool *pp, const char *file, long line)
1.25 thorpej 238: {
239:
1.34 thorpej 240: if (__predict_false(pp->pr_entered_file != NULL)) {
1.25 thorpej 241: printf("pool %s: reentrancy at file %s line %ld\n",
242: pp->pr_wchan, file, line);
243: printf(" previous entry at file %s line %ld\n",
244: pp->pr_entered_file, pp->pr_entered_line);
245: panic("pr_enter");
246: }
247:
248: pp->pr_entered_file = file;
249: pp->pr_entered_line = line;
250: }
251:
1.42 thorpej 252: static __inline void
253: pr_leave(struct pool *pp)
1.25 thorpej 254: {
255:
1.34 thorpej 256: if (__predict_false(pp->pr_entered_file == NULL)) {
1.25 thorpej 257: printf("pool %s not entered?\n", pp->pr_wchan);
258: panic("pr_leave");
259: }
260:
261: pp->pr_entered_file = NULL;
262: pp->pr_entered_line = 0;
263: }
264:
1.42 thorpej 265: static __inline void
266: pr_enter_check(struct pool *pp, void (*pr)(const char *, ...))
1.25 thorpej 267: {
268:
269: if (pp->pr_entered_file != NULL)
270: (*pr)("\n\tcurrently entered from file %s line %ld\n",
271: pp->pr_entered_file, pp->pr_entered_line);
272: }
1.3 pk 273: #else
1.25 thorpej 274: #define pr_log(pp, v, action, file, line)
275: #define pr_printlog(pp, pi, pr)
276: #define pr_enter(pp, file, line)
277: #define pr_leave(pp)
278: #define pr_enter_check(pp, pr)
1.59 thorpej 279: #endif /* POOL_DIAGNOSTIC */
1.3 pk 280:
1.88 chs 281: static __inline int
282: phtree_compare(struct pool_item_header *a, struct pool_item_header *b)
283: {
284: if (a->ph_page < b->ph_page)
285: return (-1);
286: else if (a->ph_page > b->ph_page)
287: return (1);
288: else
289: return (0);
290: }
291:
292: SPLAY_PROTOTYPE(phtree, pool_item_header, ph_node, phtree_compare);
293: SPLAY_GENERATE(phtree, pool_item_header, ph_node, phtree_compare);
294:
1.3 pk 295: /*
296: * Return the pool page header based on page address.
297: */
1.42 thorpej 298: static __inline struct pool_item_header *
299: pr_find_pagehead(struct pool *pp, caddr_t page)
1.3 pk 300: {
1.88 chs 301: struct pool_item_header *ph, tmp;
1.3 pk 302:
1.20 thorpej 303: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
1.3 pk 304: return ((struct pool_item_header *)(page + pp->pr_phoffset));
305:
1.88 chs 306: tmp.ph_page = page;
307: ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
308: return ph;
1.3 pk 309: }
310:
311: /*
312: * Remove a page from the pool.
313: */
1.42 thorpej 314: static __inline void
1.61 chs 315: pr_rmpage(struct pool *pp, struct pool_item_header *ph,
316: struct pool_pagelist *pq)
1.3 pk 317: {
1.61 chs 318: int s;
1.3 pk 319:
1.91 yamt 320: LOCK_ASSERT(!simple_lock_held(&pp->pr_slock) || pq != NULL);
321:
1.3 pk 322: /*
1.7 thorpej 323: * If the page was idle, decrement the idle page count.
1.3 pk 324: */
1.6 thorpej 325: if (ph->ph_nmissing == 0) {
326: #ifdef DIAGNOSTIC
327: if (pp->pr_nidle == 0)
328: panic("pr_rmpage: nidle inconsistent");
1.20 thorpej 329: if (pp->pr_nitems < pp->pr_itemsperpage)
330: panic("pr_rmpage: nitems inconsistent");
1.6 thorpej 331: #endif
332: pp->pr_nidle--;
333: }
1.7 thorpej 334:
1.20 thorpej 335: pp->pr_nitems -= pp->pr_itemsperpage;
336:
1.7 thorpej 337: /*
1.61 chs 338: * Unlink a page from the pool and release it (or queue it for release).
1.7 thorpej 339: */
1.88 chs 340: LIST_REMOVE(ph, ph_pagelist);
1.91 yamt 341: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
342: SPLAY_REMOVE(phtree, &pp->pr_phtree, ph);
1.61 chs 343: if (pq) {
1.88 chs 344: LIST_INSERT_HEAD(pq, ph, ph_pagelist);
1.61 chs 345: } else {
1.66 thorpej 346: pool_allocator_free(pp, ph->ph_page);
1.61 chs 347: if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
1.85 pk 348: s = splvm();
1.61 chs 349: pool_put(&phpool, ph);
350: splx(s);
351: }
352: }
1.7 thorpej 353: pp->pr_npages--;
354: pp->pr_npagefree++;
1.6 thorpej 355:
1.88 chs 356: pool_update_curpage(pp);
1.3 pk 357: }
358:
359: /*
1.94 simonb 360: * Initialize all the pools listed in the "pools" link set.
361: */
362: void
363: link_pool_init(void)
364: {
365: __link_set_decl(pools, struct link_pool_init);
366: struct link_pool_init * const *pi;
367:
368: __link_set_foreach(pi, pools)
369: pool_init((*pi)->pp, (*pi)->size, (*pi)->align,
370: (*pi)->align_offset, (*pi)->flags, (*pi)->wchan,
371: (*pi)->palloc);
372: }
373:
374: /*
1.3 pk 375: * Initialize the given pool resource structure.
376: *
377: * We export this routine to allow other kernel parts to declare
378: * static pools that must be initialized before malloc() is available.
379: */
380: void
1.42 thorpej 381: pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
1.66 thorpej 382: const char *wchan, struct pool_allocator *palloc)
1.3 pk 383: {
1.88 chs 384: int off, slack;
1.92 enami 385: size_t trysize, phsize;
1.93 dbj 386: int s;
1.3 pk 387:
1.25 thorpej 388: #ifdef POOL_DIAGNOSTIC
389: /*
390: * Always log if POOL_DIAGNOSTIC is defined.
391: */
392: if (pool_logsize != 0)
393: flags |= PR_LOGGING;
394: #endif
395:
1.66 thorpej 396: #ifdef POOL_SUBPAGE
397: /*
398: * XXX We don't provide a real `nointr' back-end
399: * yet; all sub-pages come from a kmem back-end.
400: * maybe some day...
401: */
402: if (palloc == NULL) {
403: extern struct pool_allocator pool_allocator_kmem_subpage;
404: palloc = &pool_allocator_kmem_subpage;
405: }
1.3 pk 406: /*
1.66 thorpej 407: * We'll assume any user-specified back-end allocator
408: * will deal with sub-pages, or simply don't care.
1.3 pk 409: */
1.66 thorpej 410: #else
411: if (palloc == NULL)
412: palloc = &pool_allocator_kmem;
413: #endif /* POOL_SUBPAGE */
414: if ((palloc->pa_flags & PA_INITIALIZED) == 0) {
415: if (palloc->pa_pagesz == 0) {
1.62 bjh21 416: #ifdef POOL_SUBPAGE
1.66 thorpej 417: if (palloc == &pool_allocator_kmem)
418: palloc->pa_pagesz = PAGE_SIZE;
419: else
420: palloc->pa_pagesz = POOL_SUBPAGE;
1.62 bjh21 421: #else
1.66 thorpej 422: palloc->pa_pagesz = PAGE_SIZE;
423: #endif /* POOL_SUBPAGE */
424: }
425:
426: TAILQ_INIT(&palloc->pa_list);
427:
428: simple_lock_init(&palloc->pa_slock);
429: palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
430: palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
431: palloc->pa_flags |= PA_INITIALIZED;
1.4 thorpej 432: }
1.3 pk 433:
434: if (align == 0)
435: align = ALIGN(1);
1.14 thorpej 436:
437: if (size < sizeof(struct pool_item))
438: size = sizeof(struct pool_item);
1.3 pk 439:
1.78 thorpej 440: size = roundup(size, align);
1.66 thorpej 441: #ifdef DIAGNOSTIC
442: if (size > palloc->pa_pagesz)
1.35 pk 443: panic("pool_init: pool item size (%lu) too large",
444: (u_long)size);
1.66 thorpej 445: #endif
1.35 pk 446:
1.3 pk 447: /*
448: * Initialize the pool structure.
449: */
1.88 chs 450: LIST_INIT(&pp->pr_emptypages);
451: LIST_INIT(&pp->pr_fullpages);
452: LIST_INIT(&pp->pr_partpages);
1.43 thorpej 453: TAILQ_INIT(&pp->pr_cachelist);
1.3 pk 454: pp->pr_curpage = NULL;
455: pp->pr_npages = 0;
456: pp->pr_minitems = 0;
457: pp->pr_minpages = 0;
458: pp->pr_maxpages = UINT_MAX;
1.20 thorpej 459: pp->pr_roflags = flags;
460: pp->pr_flags = 0;
1.35 pk 461: pp->pr_size = size;
1.3 pk 462: pp->pr_align = align;
463: pp->pr_wchan = wchan;
1.66 thorpej 464: pp->pr_alloc = palloc;
1.20 thorpej 465: pp->pr_nitems = 0;
466: pp->pr_nout = 0;
467: pp->pr_hardlimit = UINT_MAX;
468: pp->pr_hardlimit_warning = NULL;
1.31 thorpej 469: pp->pr_hardlimit_ratecap.tv_sec = 0;
470: pp->pr_hardlimit_ratecap.tv_usec = 0;
471: pp->pr_hardlimit_warning_last.tv_sec = 0;
472: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.68 thorpej 473: pp->pr_drain_hook = NULL;
474: pp->pr_drain_hook_arg = NULL;
1.3 pk 475:
476: /*
477: * Decide whether to put the page header off page to avoid
1.92 enami 478: * wasting too large a part of the page or too big item.
479: * Off-page page headers go on a hash table, so we can match
480: * a returned item with its header based on the page address.
481: * We use 1/16 of the page size and about 8 times of the item
482: * size as the threshold (XXX: tune)
483: *
484: * However, we'll put the header into the page if we can put
485: * it without wasting any items.
486: *
487: * Silently enforce `0 <= ioff < align'.
1.3 pk 488: */
1.92 enami 489: pp->pr_itemoffset = ioff %= align;
490: /* See the comment below about reserved bytes. */
491: trysize = palloc->pa_pagesz - ((align - ioff) % align);
492: phsize = ALIGN(sizeof(struct pool_item_header));
493: if (pp->pr_size < MIN(palloc->pa_pagesz / 16, phsize << 3) ||
494: trysize / pp->pr_size == (trysize - phsize) / pp->pr_size) {
1.3 pk 495: /* Use the end of the page for the page header */
1.20 thorpej 496: pp->pr_roflags |= PR_PHINPAGE;
1.92 enami 497: pp->pr_phoffset = off = palloc->pa_pagesz - phsize;
1.2 pk 498: } else {
1.3 pk 499: /* The page header will be taken from our page header pool */
500: pp->pr_phoffset = 0;
1.66 thorpej 501: off = palloc->pa_pagesz;
1.88 chs 502: SPLAY_INIT(&pp->pr_phtree);
1.2 pk 503: }
1.1 pk 504:
1.3 pk 505: /*
506: * Alignment is to take place at `ioff' within the item. This means
507: * we must reserve up to `align - 1' bytes on the page to allow
508: * appropriate positioning of each item.
509: */
510: pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
1.43 thorpej 511: KASSERT(pp->pr_itemsperpage != 0);
1.3 pk 512:
513: /*
514: * Use the slack between the chunks and the page header
515: * for "cache coloring".
516: */
517: slack = off - pp->pr_itemsperpage * pp->pr_size;
518: pp->pr_maxcolor = (slack / align) * align;
519: pp->pr_curcolor = 0;
520:
521: pp->pr_nget = 0;
522: pp->pr_nfail = 0;
523: pp->pr_nput = 0;
524: pp->pr_npagealloc = 0;
525: pp->pr_npagefree = 0;
1.1 pk 526: pp->pr_hiwat = 0;
1.8 thorpej 527: pp->pr_nidle = 0;
1.3 pk 528:
1.59 thorpej 529: #ifdef POOL_DIAGNOSTIC
1.25 thorpej 530: if (flags & PR_LOGGING) {
531: if (kmem_map == NULL ||
532: (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log),
533: M_TEMP, M_NOWAIT)) == NULL)
1.20 thorpej 534: pp->pr_roflags &= ~PR_LOGGING;
1.3 pk 535: pp->pr_curlogentry = 0;
536: pp->pr_logsize = pool_logsize;
537: }
1.59 thorpej 538: #endif
1.25 thorpej 539:
540: pp->pr_entered_file = NULL;
541: pp->pr_entered_line = 0;
1.3 pk 542:
1.21 thorpej 543: simple_lock_init(&pp->pr_slock);
1.1 pk 544:
1.3 pk 545: /*
1.43 thorpej 546: * Initialize private page header pool and cache magazine pool if we
547: * haven't done so yet.
1.23 thorpej 548: * XXX LOCKING.
1.3 pk 549: */
550: if (phpool.pr_size == 0) {
1.62 bjh21 551: #ifdef POOL_SUBPAGE
552: pool_init(&phpool, sizeof(struct pool_item_header), 0, 0, 0,
1.66 thorpej 553: "phpool", &pool_allocator_kmem);
1.62 bjh21 554: pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
1.66 thorpej 555: PR_RECURSIVE, "psppool", &pool_allocator_kmem);
1.62 bjh21 556: #else
1.3 pk 557: pool_init(&phpool, sizeof(struct pool_item_header), 0, 0,
1.66 thorpej 558: 0, "phpool", NULL);
1.62 bjh21 559: #endif
1.43 thorpej 560: pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
1.66 thorpej 561: 0, "pcgpool", NULL);
1.1 pk 562: }
563:
1.23 thorpej 564: /* Insert into the list of all pools. */
565: simple_lock(&pool_head_slock);
566: TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist);
567: simple_unlock(&pool_head_slock);
1.66 thorpej 568:
569: /* Insert this into the list of pools using this allocator. */
1.93 dbj 570: s = splvm();
1.66 thorpej 571: simple_lock(&palloc->pa_slock);
572: TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
573: simple_unlock(&palloc->pa_slock);
1.93 dbj 574: splx(s);
1.1 pk 575: }
576:
577: /*
578: * De-commision a pool resource.
579: */
580: void
1.42 thorpej 581: pool_destroy(struct pool *pp)
1.1 pk 582: {
1.3 pk 583: struct pool_item_header *ph;
1.43 thorpej 584: struct pool_cache *pc;
1.93 dbj 585: int s;
1.43 thorpej 586:
1.66 thorpej 587: /* Locking order: pool_allocator -> pool */
1.93 dbj 588: s = splvm();
1.66 thorpej 589: simple_lock(&pp->pr_alloc->pa_slock);
590: TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
591: simple_unlock(&pp->pr_alloc->pa_slock);
1.93 dbj 592: splx(s);
1.66 thorpej 593:
1.43 thorpej 594: /* Destroy all caches for this pool. */
595: while ((pc = TAILQ_FIRST(&pp->pr_cachelist)) != NULL)
596: pool_cache_destroy(pc);
1.3 pk 597:
598: #ifdef DIAGNOSTIC
1.20 thorpej 599: if (pp->pr_nout != 0) {
1.25 thorpej 600: pr_printlog(pp, NULL, printf);
1.80 provos 601: panic("pool_destroy: pool busy: still out: %u",
1.20 thorpej 602: pp->pr_nout);
1.3 pk 603: }
604: #endif
1.1 pk 605:
1.3 pk 606: /* Remove all pages */
1.88 chs 607: while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1.70 thorpej 608: pr_rmpage(pp, ph, NULL);
1.88 chs 609: KASSERT(LIST_EMPTY(&pp->pr_fullpages));
610: KASSERT(LIST_EMPTY(&pp->pr_partpages));
1.3 pk 611:
612: /* Remove from global pool list */
1.23 thorpej 613: simple_lock(&pool_head_slock);
1.3 pk 614: TAILQ_REMOVE(&pool_head, pp, pr_poollist);
1.61 chs 615: if (drainpp == pp) {
616: drainpp = NULL;
617: }
1.23 thorpej 618: simple_unlock(&pool_head_slock);
1.3 pk 619:
1.59 thorpej 620: #ifdef POOL_DIAGNOSTIC
1.20 thorpej 621: if ((pp->pr_roflags & PR_LOGGING) != 0)
1.3 pk 622: free(pp->pr_log, M_TEMP);
1.59 thorpej 623: #endif
1.1 pk 624: }
625:
1.68 thorpej 626: void
627: pool_set_drain_hook(struct pool *pp, void (*fn)(void *, int), void *arg)
628: {
629:
630: /* XXX no locking -- must be used just after pool_init() */
631: #ifdef DIAGNOSTIC
632: if (pp->pr_drain_hook != NULL)
633: panic("pool_set_drain_hook(%s): already set", pp->pr_wchan);
634: #endif
635: pp->pr_drain_hook = fn;
636: pp->pr_drain_hook_arg = arg;
637: }
638:
1.88 chs 639: static struct pool_item_header *
1.55 thorpej 640: pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags)
641: {
642: struct pool_item_header *ph;
643: int s;
644:
645: LOCK_ASSERT(simple_lock_held(&pp->pr_slock) == 0);
646:
647: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
648: ph = (struct pool_item_header *) (storage + pp->pr_phoffset);
649: else {
1.85 pk 650: s = splvm();
1.55 thorpej 651: ph = pool_get(&phpool, flags);
652: splx(s);
653: }
654:
655: return (ph);
656: }
1.1 pk 657:
658: /*
1.3 pk 659: * Grab an item from the pool; must be called at appropriate spl level
1.1 pk 660: */
1.3 pk 661: void *
1.59 thorpej 662: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 663: _pool_get(struct pool *pp, int flags, const char *file, long line)
1.56 sommerfe 664: #else
665: pool_get(struct pool *pp, int flags)
666: #endif
1.1 pk 667: {
668: struct pool_item *pi;
1.3 pk 669: struct pool_item_header *ph;
1.55 thorpej 670: void *v;
1.1 pk 671:
1.2 pk 672: #ifdef DIAGNOSTIC
1.95 ! atatat 673: if (__predict_false(pp->pr_itemsperpage == 0))
! 674: panic("pool_get: pool %p: pr_itemsperpage is zero, "
! 675: "pool not initialized?", pp);
1.84 thorpej 676: if (__predict_false(curlwp == NULL && doing_shutdown == 0 &&
1.37 sommerfe 677: (flags & PR_WAITOK) != 0))
1.77 matt 678: panic("pool_get: %s: must have NOWAIT", pp->pr_wchan);
1.58 thorpej 679:
680: #ifdef LOCKDEBUG
681: if (flags & PR_WAITOK)
682: simple_lock_only_held(NULL, "pool_get(PR_WAITOK)");
1.56 sommerfe 683: #endif
1.58 thorpej 684: #endif /* DIAGNOSTIC */
1.1 pk 685:
1.21 thorpej 686: simple_lock(&pp->pr_slock);
1.25 thorpej 687: pr_enter(pp, file, line);
1.20 thorpej 688:
689: startover:
690: /*
691: * Check to see if we've reached the hard limit. If we have,
692: * and we can wait, then wait until an item has been returned to
693: * the pool.
694: */
695: #ifdef DIAGNOSTIC
1.34 thorpej 696: if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
1.25 thorpej 697: pr_leave(pp);
1.21 thorpej 698: simple_unlock(&pp->pr_slock);
1.20 thorpej 699: panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
700: }
701: #endif
1.34 thorpej 702: if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
1.68 thorpej 703: if (pp->pr_drain_hook != NULL) {
704: /*
705: * Since the drain hook is going to free things
706: * back to the pool, unlock, call the hook, re-lock,
707: * and check the hardlimit condition again.
708: */
709: pr_leave(pp);
710: simple_unlock(&pp->pr_slock);
711: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
712: simple_lock(&pp->pr_slock);
713: pr_enter(pp, file, line);
714: if (pp->pr_nout < pp->pr_hardlimit)
715: goto startover;
716: }
717:
1.29 sommerfe 718: if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
1.20 thorpej 719: /*
720: * XXX: A warning isn't logged in this case. Should
721: * it be?
722: */
723: pp->pr_flags |= PR_WANTED;
1.25 thorpej 724: pr_leave(pp);
1.40 sommerfe 725: ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
1.25 thorpej 726: pr_enter(pp, file, line);
1.20 thorpej 727: goto startover;
728: }
1.31 thorpej 729:
730: /*
731: * Log a message that the hard limit has been hit.
732: */
733: if (pp->pr_hardlimit_warning != NULL &&
734: ratecheck(&pp->pr_hardlimit_warning_last,
735: &pp->pr_hardlimit_ratecap))
736: log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
1.21 thorpej 737:
738: pp->pr_nfail++;
739:
1.25 thorpej 740: pr_leave(pp);
1.21 thorpej 741: simple_unlock(&pp->pr_slock);
1.20 thorpej 742: return (NULL);
743: }
744:
1.3 pk 745: /*
746: * The convention we use is that if `curpage' is not NULL, then
747: * it points at a non-empty bucket. In particular, `curpage'
748: * never points at a page header which has PR_PHINPAGE set and
749: * has no items in its bucket.
750: */
1.20 thorpej 751: if ((ph = pp->pr_curpage) == NULL) {
752: #ifdef DIAGNOSTIC
753: if (pp->pr_nitems != 0) {
1.21 thorpej 754: simple_unlock(&pp->pr_slock);
1.20 thorpej 755: printf("pool_get: %s: curpage NULL, nitems %u\n",
756: pp->pr_wchan, pp->pr_nitems);
1.80 provos 757: panic("pool_get: nitems inconsistent");
1.20 thorpej 758: }
759: #endif
760:
1.21 thorpej 761: /*
762: * Call the back-end page allocator for more memory.
763: * Release the pool lock, as the back-end page allocator
764: * may block.
765: */
1.25 thorpej 766: pr_leave(pp);
1.21 thorpej 767: simple_unlock(&pp->pr_slock);
1.66 thorpej 768: v = pool_allocator_alloc(pp, flags);
1.55 thorpej 769: if (__predict_true(v != NULL))
770: ph = pool_alloc_item_header(pp, v, flags);
1.15 pk 771:
1.55 thorpej 772: if (__predict_false(v == NULL || ph == NULL)) {
773: if (v != NULL)
1.66 thorpej 774: pool_allocator_free(pp, v);
1.55 thorpej 775:
1.91 yamt 776: simple_lock(&pp->pr_slock);
777: pr_enter(pp, file, line);
778:
1.21 thorpej 779: /*
1.55 thorpej 780: * We were unable to allocate a page or item
781: * header, but we released the lock during
782: * allocation, so perhaps items were freed
783: * back to the pool. Check for this case.
1.21 thorpej 784: */
785: if (pp->pr_curpage != NULL)
786: goto startover;
1.15 pk 787:
1.3 pk 788: if ((flags & PR_WAITOK) == 0) {
789: pp->pr_nfail++;
1.25 thorpej 790: pr_leave(pp);
1.21 thorpej 791: simple_unlock(&pp->pr_slock);
1.1 pk 792: return (NULL);
1.3 pk 793: }
794:
1.15 pk 795: /*
796: * Wait for items to be returned to this pool.
1.21 thorpej 797: *
1.20 thorpej 798: * XXX: maybe we should wake up once a second and
799: * try again?
1.15 pk 800: */
1.1 pk 801: pp->pr_flags |= PR_WANTED;
1.66 thorpej 802: /* PA_WANTED is already set on the allocator. */
1.25 thorpej 803: pr_leave(pp);
1.40 sommerfe 804: ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
1.25 thorpej 805: pr_enter(pp, file, line);
1.20 thorpej 806: goto startover;
1.1 pk 807: }
1.3 pk 808:
1.15 pk 809: /* We have more memory; add it to the pool */
1.91 yamt 810: simple_lock(&pp->pr_slock);
811: pr_enter(pp, file, line);
1.55 thorpej 812: pool_prime_page(pp, v, ph);
1.15 pk 813: pp->pr_npagealloc++;
814:
1.20 thorpej 815: /* Start the allocation process over. */
816: goto startover;
1.3 pk 817: }
1.34 thorpej 818: if (__predict_false((v = pi = TAILQ_FIRST(&ph->ph_itemlist)) == NULL)) {
1.25 thorpej 819: pr_leave(pp);
1.21 thorpej 820: simple_unlock(&pp->pr_slock);
1.3 pk 821: panic("pool_get: %s: page empty", pp->pr_wchan);
1.21 thorpej 822: }
1.20 thorpej 823: #ifdef DIAGNOSTIC
1.34 thorpej 824: if (__predict_false(pp->pr_nitems == 0)) {
1.25 thorpej 825: pr_leave(pp);
1.21 thorpej 826: simple_unlock(&pp->pr_slock);
1.20 thorpej 827: printf("pool_get: %s: items on itemlist, nitems %u\n",
828: pp->pr_wchan, pp->pr_nitems);
1.80 provos 829: panic("pool_get: nitems inconsistent");
1.20 thorpej 830: }
1.65 enami 831: #endif
1.56 sommerfe 832:
1.65 enami 833: #ifdef POOL_DIAGNOSTIC
1.3 pk 834: pr_log(pp, v, PRLOG_GET, file, line);
1.65 enami 835: #endif
1.3 pk 836:
1.65 enami 837: #ifdef DIAGNOSTIC
1.34 thorpej 838: if (__predict_false(pi->pi_magic != PI_MAGIC)) {
1.25 thorpej 839: pr_printlog(pp, pi, printf);
1.3 pk 840: panic("pool_get(%s): free list modified: magic=%x; page %p;"
841: " item addr %p\n",
842: pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
843: }
844: #endif
845:
846: /*
847: * Remove from item list.
848: */
849: TAILQ_REMOVE(&ph->ph_itemlist, pi, pi_list);
1.20 thorpej 850: pp->pr_nitems--;
851: pp->pr_nout++;
1.6 thorpej 852: if (ph->ph_nmissing == 0) {
853: #ifdef DIAGNOSTIC
1.34 thorpej 854: if (__predict_false(pp->pr_nidle == 0))
1.6 thorpej 855: panic("pool_get: nidle inconsistent");
856: #endif
857: pp->pr_nidle--;
1.88 chs 858:
859: /*
860: * This page was previously empty. Move it to the list of
861: * partially-full pages. This page is already curpage.
862: */
863: LIST_REMOVE(ph, ph_pagelist);
864: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.6 thorpej 865: }
1.3 pk 866: ph->ph_nmissing++;
1.88 chs 867: if (TAILQ_EMPTY(&ph->ph_itemlist)) {
1.21 thorpej 868: #ifdef DIAGNOSTIC
1.34 thorpej 869: if (__predict_false(ph->ph_nmissing != pp->pr_itemsperpage)) {
1.25 thorpej 870: pr_leave(pp);
1.21 thorpej 871: simple_unlock(&pp->pr_slock);
872: panic("pool_get: %s: nmissing inconsistent",
873: pp->pr_wchan);
874: }
875: #endif
1.3 pk 876: /*
1.88 chs 877: * This page is now full. Move it to the full list
878: * and select a new current page.
1.3 pk 879: */
1.88 chs 880: LIST_REMOVE(ph, ph_pagelist);
881: LIST_INSERT_HEAD(&pp->pr_fullpages, ph, ph_pagelist);
882: pool_update_curpage(pp);
1.1 pk 883: }
1.3 pk 884:
885: pp->pr_nget++;
1.20 thorpej 886:
887: /*
888: * If we have a low water mark and we are now below that low
889: * water mark, add more items to the pool.
890: */
1.53 thorpej 891: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 892: /*
893: * XXX: Should we log a warning? Should we set up a timeout
894: * to try again in a second or so? The latter could break
895: * a caller's assumptions about interrupt protection, etc.
896: */
897: }
898:
1.25 thorpej 899: pr_leave(pp);
1.21 thorpej 900: simple_unlock(&pp->pr_slock);
1.1 pk 901: return (v);
902: }
903:
904: /*
1.43 thorpej 905: * Internal version of pool_put(). Pool is already locked/entered.
1.1 pk 906: */
1.43 thorpej 907: static void
1.56 sommerfe 908: pool_do_put(struct pool *pp, void *v)
1.1 pk 909: {
910: struct pool_item *pi = v;
1.3 pk 911: struct pool_item_header *ph;
912: caddr_t page;
1.21 thorpej 913: int s;
1.3 pk 914:
1.61 chs 915: LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
916:
1.66 thorpej 917: page = (caddr_t)((u_long)v & pp->pr_alloc->pa_pagemask);
1.1 pk 918:
1.30 thorpej 919: #ifdef DIAGNOSTIC
1.34 thorpej 920: if (__predict_false(pp->pr_nout == 0)) {
1.30 thorpej 921: printf("pool %s: putting with none out\n",
922: pp->pr_wchan);
923: panic("pool_put");
924: }
925: #endif
1.3 pk 926:
1.34 thorpej 927: if (__predict_false((ph = pr_find_pagehead(pp, page)) == NULL)) {
1.25 thorpej 928: pr_printlog(pp, NULL, printf);
1.3 pk 929: panic("pool_put: %s: page header missing", pp->pr_wchan);
930: }
1.28 thorpej 931:
932: #ifdef LOCKDEBUG
933: /*
934: * Check if we're freeing a locked simple lock.
935: */
936: simple_lock_freecheck((caddr_t)pi, ((caddr_t)pi) + pp->pr_size);
937: #endif
1.3 pk 938:
939: /*
940: * Return to item list.
941: */
1.2 pk 942: #ifdef DIAGNOSTIC
1.3 pk 943: pi->pi_magic = PI_MAGIC;
944: #endif
1.32 chs 945: #ifdef DEBUG
946: {
947: int i, *ip = v;
948:
949: for (i = 0; i < pp->pr_size / sizeof(int); i++) {
950: *ip++ = PI_MAGIC;
951: }
952: }
953: #endif
954:
1.3 pk 955: TAILQ_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.79 thorpej 956: KDASSERT(ph->ph_nmissing != 0);
1.3 pk 957: ph->ph_nmissing--;
958: pp->pr_nput++;
1.20 thorpej 959: pp->pr_nitems++;
960: pp->pr_nout--;
1.3 pk 961:
962: /* Cancel "pool empty" condition if it exists */
963: if (pp->pr_curpage == NULL)
964: pp->pr_curpage = ph;
965:
966: if (pp->pr_flags & PR_WANTED) {
967: pp->pr_flags &= ~PR_WANTED;
1.15 pk 968: if (ph->ph_nmissing == 0)
969: pp->pr_nidle++;
1.3 pk 970: wakeup((caddr_t)pp);
971: return;
972: }
973:
974: /*
1.88 chs 975: * If this page is now empty, do one of two things:
1.21 thorpej 976: *
1.88 chs 977: * (1) If we have more pages than the page high water mark,
1.90 thorpej 978: * or if we are flagged as immediately freeing back idle
979: * pages, free the page back to the system. ONLY CONSIDER
980: * FREEING BACK A PAGE IF WE HAVE MORE THAN OUR MINIMUM PAGE
981: * CLAIM.
1.21 thorpej 982: *
1.88 chs 983: * (2) Otherwise, move the page to the empty page list.
984: *
985: * Either way, select a new current page (so we use a partially-full
986: * page if one is available).
1.3 pk 987: */
988: if (ph->ph_nmissing == 0) {
1.6 thorpej 989: pp->pr_nidle++;
1.90 thorpej 990: if (pp->pr_npages > pp->pr_minpages &&
991: (pp->pr_npages > pp->pr_maxpages ||
992: (pp->pr_roflags & PR_IMMEDRELEASE) != 0 ||
993: (pp->pr_alloc->pa_flags & PA_WANT) != 0)) {
1.91 yamt 994: simple_unlock(&pp->pr_slock);
1.61 chs 995: pr_rmpage(pp, ph, NULL);
1.91 yamt 996: simple_lock(&pp->pr_slock);
1.3 pk 997: } else {
1.88 chs 998: LIST_REMOVE(ph, ph_pagelist);
999: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.3 pk 1000:
1.21 thorpej 1001: /*
1002: * Update the timestamp on the page. A page must
1003: * be idle for some period of time before it can
1004: * be reclaimed by the pagedaemon. This minimizes
1005: * ping-pong'ing for memory.
1006: */
1007: s = splclock();
1008: ph->ph_time = mono_time;
1009: splx(s);
1.1 pk 1010: }
1.88 chs 1011: pool_update_curpage(pp);
1.1 pk 1012: }
1.88 chs 1013:
1.21 thorpej 1014: /*
1.88 chs 1015: * If the page was previously completely full, move it to the
1016: * partially-full list and make it the current page. The next
1017: * allocation will get the item from this page, instead of
1018: * further fragmenting the pool.
1.21 thorpej 1019: */
1020: else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
1.88 chs 1021: LIST_REMOVE(ph, ph_pagelist);
1022: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.21 thorpej 1023: pp->pr_curpage = ph;
1024: }
1.43 thorpej 1025: }
1026:
1027: /*
1028: * Return resource to the pool; must be called at appropriate spl level
1029: */
1.59 thorpej 1030: #ifdef POOL_DIAGNOSTIC
1.43 thorpej 1031: void
1032: _pool_put(struct pool *pp, void *v, const char *file, long line)
1033: {
1034:
1035: simple_lock(&pp->pr_slock);
1036: pr_enter(pp, file, line);
1037:
1.56 sommerfe 1038: pr_log(pp, v, PRLOG_PUT, file, line);
1039:
1040: pool_do_put(pp, v);
1.21 thorpej 1041:
1.25 thorpej 1042: pr_leave(pp);
1.21 thorpej 1043: simple_unlock(&pp->pr_slock);
1.1 pk 1044: }
1.57 sommerfe 1045: #undef pool_put
1.59 thorpej 1046: #endif /* POOL_DIAGNOSTIC */
1.1 pk 1047:
1.56 sommerfe 1048: void
1049: pool_put(struct pool *pp, void *v)
1050: {
1051:
1052: simple_lock(&pp->pr_slock);
1053:
1054: pool_do_put(pp, v);
1055:
1056: simple_unlock(&pp->pr_slock);
1057: }
1.57 sommerfe 1058:
1.59 thorpej 1059: #ifdef POOL_DIAGNOSTIC
1.57 sommerfe 1060: #define pool_put(h, v) _pool_put((h), (v), __FILE__, __LINE__)
1.56 sommerfe 1061: #endif
1.74 thorpej 1062:
1063: /*
1064: * Add N items to the pool.
1065: */
1066: int
1067: pool_prime(struct pool *pp, int n)
1068: {
1.83 scw 1069: struct pool_item_header *ph = NULL;
1.74 thorpej 1070: caddr_t cp;
1.75 simonb 1071: int newpages;
1.74 thorpej 1072:
1073: simple_lock(&pp->pr_slock);
1074:
1075: newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1076:
1077: while (newpages-- > 0) {
1078: simple_unlock(&pp->pr_slock);
1079: cp = pool_allocator_alloc(pp, PR_NOWAIT);
1080: if (__predict_true(cp != NULL))
1081: ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1082:
1083: if (__predict_false(cp == NULL || ph == NULL)) {
1084: if (cp != NULL)
1085: pool_allocator_free(pp, cp);
1.91 yamt 1086: simple_lock(&pp->pr_slock);
1.74 thorpej 1087: break;
1088: }
1089:
1.91 yamt 1090: simple_lock(&pp->pr_slock);
1.74 thorpej 1091: pool_prime_page(pp, cp, ph);
1092: pp->pr_npagealloc++;
1093: pp->pr_minpages++;
1094: }
1095:
1096: if (pp->pr_minpages >= pp->pr_maxpages)
1097: pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */
1098:
1099: simple_unlock(&pp->pr_slock);
1100: return (0);
1101: }
1.55 thorpej 1102:
1103: /*
1.3 pk 1104: * Add a page worth of items to the pool.
1.21 thorpej 1105: *
1106: * Note, we must be called with the pool descriptor LOCKED.
1.3 pk 1107: */
1.55 thorpej 1108: static void
1109: pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph)
1.3 pk 1110: {
1111: struct pool_item *pi;
1112: caddr_t cp = storage;
1113: unsigned int align = pp->pr_align;
1114: unsigned int ioff = pp->pr_itemoffset;
1.55 thorpej 1115: int n;
1.89 yamt 1116: int s;
1.36 pk 1117:
1.91 yamt 1118: LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
1119:
1.66 thorpej 1120: #ifdef DIAGNOSTIC
1121: if (((u_long)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
1.36 pk 1122: panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1.66 thorpej 1123: #endif
1.3 pk 1124:
1125: /*
1126: * Insert page header.
1127: */
1.88 chs 1128: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.3 pk 1129: TAILQ_INIT(&ph->ph_itemlist);
1130: ph->ph_page = storage;
1131: ph->ph_nmissing = 0;
1.89 yamt 1132: s = splclock();
1133: ph->ph_time = mono_time;
1134: splx(s);
1.88 chs 1135: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1136: SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
1.3 pk 1137:
1.6 thorpej 1138: pp->pr_nidle++;
1139:
1.3 pk 1140: /*
1141: * Color this page.
1142: */
1143: cp = (caddr_t)(cp + pp->pr_curcolor);
1144: if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
1145: pp->pr_curcolor = 0;
1146:
1147: /*
1148: * Adjust storage to apply aligment to `pr_itemoffset' in each item.
1149: */
1150: if (ioff != 0)
1151: cp = (caddr_t)(cp + (align - ioff));
1152:
1153: /*
1154: * Insert remaining chunks on the bucket list.
1155: */
1156: n = pp->pr_itemsperpage;
1.20 thorpej 1157: pp->pr_nitems += n;
1.3 pk 1158:
1159: while (n--) {
1160: pi = (struct pool_item *)cp;
1.78 thorpej 1161:
1162: KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
1.3 pk 1163:
1164: /* Insert on page list */
1165: TAILQ_INSERT_TAIL(&ph->ph_itemlist, pi, pi_list);
1166: #ifdef DIAGNOSTIC
1167: pi->pi_magic = PI_MAGIC;
1168: #endif
1169: cp = (caddr_t)(cp + pp->pr_size);
1170: }
1171:
1172: /*
1173: * If the pool was depleted, point at the new page.
1174: */
1175: if (pp->pr_curpage == NULL)
1176: pp->pr_curpage = ph;
1177:
1178: if (++pp->pr_npages > pp->pr_hiwat)
1179: pp->pr_hiwat = pp->pr_npages;
1180: }
1181:
1.20 thorpej 1182: /*
1.52 thorpej 1183: * Used by pool_get() when nitems drops below the low water mark. This
1.88 chs 1184: * is used to catch up pr_nitems with the low water mark.
1.20 thorpej 1185: *
1.21 thorpej 1186: * Note 1, we never wait for memory here, we let the caller decide what to do.
1.20 thorpej 1187: *
1.73 thorpej 1188: * Note 2, we must be called with the pool already locked, and we return
1.20 thorpej 1189: * with it locked.
1190: */
1191: static int
1.42 thorpej 1192: pool_catchup(struct pool *pp)
1.20 thorpej 1193: {
1.83 scw 1194: struct pool_item_header *ph = NULL;
1.20 thorpej 1195: caddr_t cp;
1196: int error = 0;
1197:
1.54 thorpej 1198: while (POOL_NEEDS_CATCHUP(pp)) {
1.20 thorpej 1199: /*
1.21 thorpej 1200: * Call the page back-end allocator for more memory.
1201: *
1202: * XXX: We never wait, so should we bother unlocking
1203: * the pool descriptor?
1.20 thorpej 1204: */
1.21 thorpej 1205: simple_unlock(&pp->pr_slock);
1.66 thorpej 1206: cp = pool_allocator_alloc(pp, PR_NOWAIT);
1.55 thorpej 1207: if (__predict_true(cp != NULL))
1208: ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1209: if (__predict_false(cp == NULL || ph == NULL)) {
1210: if (cp != NULL)
1.66 thorpej 1211: pool_allocator_free(pp, cp);
1.20 thorpej 1212: error = ENOMEM;
1.91 yamt 1213: simple_lock(&pp->pr_slock);
1.20 thorpej 1214: break;
1215: }
1.91 yamt 1216: simple_lock(&pp->pr_slock);
1.55 thorpej 1217: pool_prime_page(pp, cp, ph);
1.26 thorpej 1218: pp->pr_npagealloc++;
1.20 thorpej 1219: }
1220:
1221: return (error);
1222: }
1223:
1.88 chs 1224: static void
1225: pool_update_curpage(struct pool *pp)
1226: {
1227:
1228: pp->pr_curpage = LIST_FIRST(&pp->pr_partpages);
1229: if (pp->pr_curpage == NULL) {
1230: pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages);
1231: }
1232: }
1233:
1.3 pk 1234: void
1.42 thorpej 1235: pool_setlowat(struct pool *pp, int n)
1.3 pk 1236: {
1.15 pk 1237:
1.21 thorpej 1238: simple_lock(&pp->pr_slock);
1239:
1.3 pk 1240: pp->pr_minitems = n;
1.15 pk 1241: pp->pr_minpages = (n == 0)
1242: ? 0
1.18 thorpej 1243: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.20 thorpej 1244:
1245: /* Make sure we're caught up with the newly-set low water mark. */
1.75 simonb 1246: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1247: /*
1248: * XXX: Should we log a warning? Should we set up a timeout
1249: * to try again in a second or so? The latter could break
1250: * a caller's assumptions about interrupt protection, etc.
1251: */
1252: }
1.21 thorpej 1253:
1254: simple_unlock(&pp->pr_slock);
1.3 pk 1255: }
1256:
1257: void
1.42 thorpej 1258: pool_sethiwat(struct pool *pp, int n)
1.3 pk 1259: {
1.15 pk 1260:
1.21 thorpej 1261: simple_lock(&pp->pr_slock);
1262:
1.15 pk 1263: pp->pr_maxpages = (n == 0)
1264: ? 0
1.18 thorpej 1265: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1266:
1267: simple_unlock(&pp->pr_slock);
1.3 pk 1268: }
1269:
1.20 thorpej 1270: void
1.42 thorpej 1271: pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1.20 thorpej 1272: {
1273:
1.21 thorpej 1274: simple_lock(&pp->pr_slock);
1.20 thorpej 1275:
1276: pp->pr_hardlimit = n;
1277: pp->pr_hardlimit_warning = warnmess;
1.31 thorpej 1278: pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1279: pp->pr_hardlimit_warning_last.tv_sec = 0;
1280: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.20 thorpej 1281:
1282: /*
1.21 thorpej 1283: * In-line version of pool_sethiwat(), because we don't want to
1284: * release the lock.
1.20 thorpej 1285: */
1286: pp->pr_maxpages = (n == 0)
1287: ? 0
1288: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1289:
1290: simple_unlock(&pp->pr_slock);
1.20 thorpej 1291: }
1.3 pk 1292:
1293: /*
1294: * Release all complete pages that have not been used recently.
1295: */
1.66 thorpej 1296: int
1.59 thorpej 1297: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 1298: _pool_reclaim(struct pool *pp, const char *file, long line)
1.56 sommerfe 1299: #else
1300: pool_reclaim(struct pool *pp)
1301: #endif
1.3 pk 1302: {
1303: struct pool_item_header *ph, *phnext;
1.43 thorpej 1304: struct pool_cache *pc;
1.21 thorpej 1305: struct timeval curtime;
1.61 chs 1306: struct pool_pagelist pq;
1.88 chs 1307: struct timeval diff;
1.21 thorpej 1308: int s;
1.3 pk 1309:
1.68 thorpej 1310: if (pp->pr_drain_hook != NULL) {
1311: /*
1312: * The drain hook must be called with the pool unlocked.
1313: */
1314: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT);
1315: }
1316:
1.21 thorpej 1317: if (simple_lock_try(&pp->pr_slock) == 0)
1.66 thorpej 1318: return (0);
1.25 thorpej 1319: pr_enter(pp, file, line);
1.68 thorpej 1320:
1.88 chs 1321: LIST_INIT(&pq);
1.3 pk 1322:
1.43 thorpej 1323: /*
1324: * Reclaim items from the pool's caches.
1325: */
1.61 chs 1326: TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
1.43 thorpej 1327: pool_cache_reclaim(pc);
1328:
1.21 thorpej 1329: s = splclock();
1330: curtime = mono_time;
1331: splx(s);
1332:
1.88 chs 1333: for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
1334: phnext = LIST_NEXT(ph, ph_pagelist);
1.3 pk 1335:
1336: /* Check our minimum page claim */
1337: if (pp->pr_npages <= pp->pr_minpages)
1338: break;
1339:
1.88 chs 1340: KASSERT(ph->ph_nmissing == 0);
1341: timersub(&curtime, &ph->ph_time, &diff);
1342: if (diff.tv_sec < pool_inactive_time)
1343: continue;
1.21 thorpej 1344:
1.88 chs 1345: /*
1346: * If freeing this page would put us below
1347: * the low water mark, stop now.
1348: */
1349: if ((pp->pr_nitems - pp->pr_itemsperpage) <
1350: pp->pr_minitems)
1351: break;
1.21 thorpej 1352:
1.88 chs 1353: pr_rmpage(pp, ph, &pq);
1.3 pk 1354: }
1355:
1.25 thorpej 1356: pr_leave(pp);
1.21 thorpej 1357: simple_unlock(&pp->pr_slock);
1.88 chs 1358: if (LIST_EMPTY(&pq))
1.66 thorpej 1359: return (0);
1360:
1.88 chs 1361: while ((ph = LIST_FIRST(&pq)) != NULL) {
1362: LIST_REMOVE(ph, ph_pagelist);
1.66 thorpej 1363: pool_allocator_free(pp, ph->ph_page);
1.61 chs 1364: if (pp->pr_roflags & PR_PHINPAGE) {
1365: continue;
1366: }
1.85 pk 1367: s = splvm();
1.61 chs 1368: pool_put(&phpool, ph);
1369: splx(s);
1370: }
1.66 thorpej 1371:
1372: return (1);
1.3 pk 1373: }
1374:
1375: /*
1376: * Drain pools, one at a time.
1.21 thorpej 1377: *
1378: * Note, we must never be called from an interrupt context.
1.3 pk 1379: */
1380: void
1.42 thorpej 1381: pool_drain(void *arg)
1.3 pk 1382: {
1383: struct pool *pp;
1.23 thorpej 1384: int s;
1.3 pk 1385:
1.61 chs 1386: pp = NULL;
1.49 thorpej 1387: s = splvm();
1.23 thorpej 1388: simple_lock(&pool_head_slock);
1.61 chs 1389: if (drainpp == NULL) {
1390: drainpp = TAILQ_FIRST(&pool_head);
1391: }
1392: if (drainpp) {
1393: pp = drainpp;
1394: drainpp = TAILQ_NEXT(pp, pr_poollist);
1395: }
1396: simple_unlock(&pool_head_slock);
1.63 chs 1397: pool_reclaim(pp);
1.61 chs 1398: splx(s);
1.3 pk 1399: }
1400:
1401: /*
1402: * Diagnostic helpers.
1403: */
1404: void
1.42 thorpej 1405: pool_print(struct pool *pp, const char *modif)
1.21 thorpej 1406: {
1407: int s;
1408:
1.49 thorpej 1409: s = splvm();
1.25 thorpej 1410: if (simple_lock_try(&pp->pr_slock) == 0) {
1411: printf("pool %s is locked; try again later\n",
1412: pp->pr_wchan);
1413: splx(s);
1414: return;
1415: }
1416: pool_print1(pp, modif, printf);
1.21 thorpej 1417: simple_unlock(&pp->pr_slock);
1418: splx(s);
1419: }
1420:
1.25 thorpej 1421: void
1.42 thorpej 1422: pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.25 thorpej 1423: {
1424: int didlock = 0;
1425:
1426: if (pp == NULL) {
1427: (*pr)("Must specify a pool to print.\n");
1428: return;
1429: }
1430:
1431: /*
1432: * Called from DDB; interrupts should be blocked, and all
1433: * other processors should be paused. We can skip locking
1434: * the pool in this case.
1435: *
1436: * We do a simple_lock_try() just to print the lock
1437: * status, however.
1438: */
1439:
1440: if (simple_lock_try(&pp->pr_slock) == 0)
1441: (*pr)("WARNING: pool %s is locked\n", pp->pr_wchan);
1442: else
1443: didlock = 1;
1444:
1445: pool_print1(pp, modif, pr);
1446:
1447: if (didlock)
1448: simple_unlock(&pp->pr_slock);
1449: }
1450:
1.21 thorpej 1451: static void
1.88 chs 1452: pool_print_pagelist(struct pool_pagelist *pl, void (*pr)(const char *, ...))
1453: {
1454: struct pool_item_header *ph;
1455: #ifdef DIAGNOSTIC
1456: struct pool_item *pi;
1457: #endif
1458:
1459: LIST_FOREACH(ph, pl, ph_pagelist) {
1460: (*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n",
1461: ph->ph_page, ph->ph_nmissing,
1462: (u_long)ph->ph_time.tv_sec,
1463: (u_long)ph->ph_time.tv_usec);
1464: #ifdef DIAGNOSTIC
1465: TAILQ_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1466: if (pi->pi_magic != PI_MAGIC) {
1467: (*pr)("\t\t\titem %p, magic 0x%x\n",
1468: pi, pi->pi_magic);
1469: }
1470: }
1471: #endif
1472: }
1473: }
1474:
1475: static void
1.42 thorpej 1476: pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.3 pk 1477: {
1.25 thorpej 1478: struct pool_item_header *ph;
1.44 thorpej 1479: struct pool_cache *pc;
1480: struct pool_cache_group *pcg;
1481: int i, print_log = 0, print_pagelist = 0, print_cache = 0;
1.25 thorpej 1482: char c;
1483:
1484: while ((c = *modif++) != '\0') {
1485: if (c == 'l')
1486: print_log = 1;
1487: if (c == 'p')
1488: print_pagelist = 1;
1.44 thorpej 1489: if (c == 'c')
1490: print_cache = 1;
1.25 thorpej 1491: }
1492:
1493: (*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1494: pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1495: pp->pr_roflags);
1.66 thorpej 1496: (*pr)("\talloc %p\n", pp->pr_alloc);
1.25 thorpej 1497: (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1498: pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1499: (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1500: pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1501:
1502: (*pr)("\n\tnget %lu, nfail %lu, nput %lu\n",
1503: pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1504: (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1505: pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1506:
1507: if (print_pagelist == 0)
1508: goto skip_pagelist;
1509:
1.88 chs 1510: if ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1511: (*pr)("\n\tempty page list:\n");
1512: pool_print_pagelist(&pp->pr_emptypages, pr);
1513: if ((ph = LIST_FIRST(&pp->pr_fullpages)) != NULL)
1514: (*pr)("\n\tfull page list:\n");
1515: pool_print_pagelist(&pp->pr_fullpages, pr);
1516: if ((ph = LIST_FIRST(&pp->pr_partpages)) != NULL)
1517: (*pr)("\n\tpartial-page list:\n");
1518: pool_print_pagelist(&pp->pr_partpages, pr);
1519:
1.25 thorpej 1520: if (pp->pr_curpage == NULL)
1521: (*pr)("\tno current page\n");
1522: else
1523: (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1524:
1525: skip_pagelist:
1526: if (print_log == 0)
1527: goto skip_log;
1528:
1529: (*pr)("\n");
1530: if ((pp->pr_roflags & PR_LOGGING) == 0)
1531: (*pr)("\tno log\n");
1532: else
1533: pr_printlog(pp, NULL, pr);
1.3 pk 1534:
1.25 thorpej 1535: skip_log:
1.44 thorpej 1536: if (print_cache == 0)
1537: goto skip_cache;
1538:
1.61 chs 1539: TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist) {
1.44 thorpej 1540: (*pr)("\tcache %p: allocfrom %p freeto %p\n", pc,
1541: pc->pc_allocfrom, pc->pc_freeto);
1.48 thorpej 1542: (*pr)("\t hits %lu misses %lu ngroups %lu nitems %lu\n",
1543: pc->pc_hits, pc->pc_misses, pc->pc_ngroups, pc->pc_nitems);
1.61 chs 1544: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.44 thorpej 1545: (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail);
1.87 thorpej 1546: for (i = 0; i < PCG_NOBJECTS; i++) {
1547: if (pcg->pcg_objects[i].pcgo_pa !=
1548: POOL_PADDR_INVALID) {
1549: (*pr)("\t\t\t%p, 0x%llx\n",
1550: pcg->pcg_objects[i].pcgo_va,
1551: (unsigned long long)
1552: pcg->pcg_objects[i].pcgo_pa);
1553: } else {
1554: (*pr)("\t\t\t%p\n",
1555: pcg->pcg_objects[i].pcgo_va);
1556: }
1557: }
1.44 thorpej 1558: }
1559: }
1560:
1561: skip_cache:
1.88 chs 1562: pr_enter_check(pp, pr);
1563: }
1564:
1565: static int
1566: pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
1567: {
1568: struct pool_item *pi;
1569: caddr_t page;
1570: int n;
1571:
1572: page = (caddr_t)((u_long)ph & pp->pr_alloc->pa_pagemask);
1573: if (page != ph->ph_page &&
1574: (pp->pr_roflags & PR_PHINPAGE) != 0) {
1575: if (label != NULL)
1576: printf("%s: ", label);
1577: printf("pool(%p:%s): page inconsistency: page %p;"
1578: " at page head addr %p (p %p)\n", pp,
1579: pp->pr_wchan, ph->ph_page,
1580: ph, page);
1581: return 1;
1582: }
1.3 pk 1583:
1.88 chs 1584: for (pi = TAILQ_FIRST(&ph->ph_itemlist), n = 0;
1585: pi != NULL;
1586: pi = TAILQ_NEXT(pi,pi_list), n++) {
1587:
1588: #ifdef DIAGNOSTIC
1589: if (pi->pi_magic != PI_MAGIC) {
1590: if (label != NULL)
1591: printf("%s: ", label);
1592: printf("pool(%s): free list modified: magic=%x;"
1593: " page %p; item ordinal %d;"
1594: " addr %p (p %p)\n",
1595: pp->pr_wchan, pi->pi_magic, ph->ph_page,
1596: n, pi, page);
1597: panic("pool");
1598: }
1599: #endif
1600: page =
1601: (caddr_t)((u_long)pi & pp->pr_alloc->pa_pagemask);
1602: if (page == ph->ph_page)
1603: continue;
1604:
1605: if (label != NULL)
1606: printf("%s: ", label);
1607: printf("pool(%p:%s): page inconsistency: page %p;"
1608: " item ordinal %d; addr %p (p %p)\n", pp,
1609: pp->pr_wchan, ph->ph_page,
1610: n, pi, page);
1611: return 1;
1612: }
1613: return 0;
1.3 pk 1614: }
1615:
1.88 chs 1616:
1.3 pk 1617: int
1.42 thorpej 1618: pool_chk(struct pool *pp, const char *label)
1.3 pk 1619: {
1620: struct pool_item_header *ph;
1621: int r = 0;
1622:
1.21 thorpej 1623: simple_lock(&pp->pr_slock);
1.88 chs 1624: LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
1625: r = pool_chk_page(pp, label, ph);
1626: if (r) {
1627: goto out;
1628: }
1629: }
1630: LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
1631: r = pool_chk_page(pp, label, ph);
1632: if (r) {
1.3 pk 1633: goto out;
1634: }
1.88 chs 1635: }
1636: LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
1637: r = pool_chk_page(pp, label, ph);
1638: if (r) {
1.3 pk 1639: goto out;
1640: }
1641: }
1.88 chs 1642:
1.3 pk 1643: out:
1.21 thorpej 1644: simple_unlock(&pp->pr_slock);
1.3 pk 1645: return (r);
1.43 thorpej 1646: }
1647:
1648: /*
1649: * pool_cache_init:
1650: *
1651: * Initialize a pool cache.
1652: *
1653: * NOTE: If the pool must be protected from interrupts, we expect
1654: * to be called at the appropriate interrupt priority level.
1655: */
1656: void
1657: pool_cache_init(struct pool_cache *pc, struct pool *pp,
1658: int (*ctor)(void *, void *, int),
1659: void (*dtor)(void *, void *),
1660: void *arg)
1661: {
1662:
1663: TAILQ_INIT(&pc->pc_grouplist);
1664: simple_lock_init(&pc->pc_slock);
1665:
1666: pc->pc_allocfrom = NULL;
1667: pc->pc_freeto = NULL;
1668: pc->pc_pool = pp;
1669:
1670: pc->pc_ctor = ctor;
1671: pc->pc_dtor = dtor;
1672: pc->pc_arg = arg;
1673:
1.48 thorpej 1674: pc->pc_hits = 0;
1675: pc->pc_misses = 0;
1676:
1677: pc->pc_ngroups = 0;
1678:
1679: pc->pc_nitems = 0;
1680:
1.43 thorpej 1681: simple_lock(&pp->pr_slock);
1682: TAILQ_INSERT_TAIL(&pp->pr_cachelist, pc, pc_poollist);
1683: simple_unlock(&pp->pr_slock);
1684: }
1685:
1686: /*
1687: * pool_cache_destroy:
1688: *
1689: * Destroy a pool cache.
1690: */
1691: void
1692: pool_cache_destroy(struct pool_cache *pc)
1693: {
1694: struct pool *pp = pc->pc_pool;
1695:
1696: /* First, invalidate the entire cache. */
1697: pool_cache_invalidate(pc);
1698:
1699: /* ...and remove it from the pool's cache list. */
1700: simple_lock(&pp->pr_slock);
1701: TAILQ_REMOVE(&pp->pr_cachelist, pc, pc_poollist);
1702: simple_unlock(&pp->pr_slock);
1703: }
1704:
1705: static __inline void *
1.87 thorpej 1706: pcg_get(struct pool_cache_group *pcg, paddr_t *pap)
1.43 thorpej 1707: {
1708: void *object;
1709: u_int idx;
1710:
1711: KASSERT(pcg->pcg_avail <= PCG_NOBJECTS);
1.45 thorpej 1712: KASSERT(pcg->pcg_avail != 0);
1.43 thorpej 1713: idx = --pcg->pcg_avail;
1714:
1.87 thorpej 1715: KASSERT(pcg->pcg_objects[idx].pcgo_va != NULL);
1716: object = pcg->pcg_objects[idx].pcgo_va;
1717: if (pap != NULL)
1718: *pap = pcg->pcg_objects[idx].pcgo_pa;
1719: pcg->pcg_objects[idx].pcgo_va = NULL;
1.43 thorpej 1720:
1721: return (object);
1722: }
1723:
1724: static __inline void
1.87 thorpej 1725: pcg_put(struct pool_cache_group *pcg, void *object, paddr_t pa)
1.43 thorpej 1726: {
1727: u_int idx;
1728:
1729: KASSERT(pcg->pcg_avail < PCG_NOBJECTS);
1730: idx = pcg->pcg_avail++;
1731:
1.87 thorpej 1732: KASSERT(pcg->pcg_objects[idx].pcgo_va == NULL);
1733: pcg->pcg_objects[idx].pcgo_va = object;
1734: pcg->pcg_objects[idx].pcgo_pa = pa;
1.43 thorpej 1735: }
1736:
1737: /*
1.87 thorpej 1738: * pool_cache_get{,_paddr}:
1.43 thorpej 1739: *
1.87 thorpej 1740: * Get an object from a pool cache (optionally returning
1741: * the physical address of the object).
1.43 thorpej 1742: */
1743: void *
1.87 thorpej 1744: pool_cache_get_paddr(struct pool_cache *pc, int flags, paddr_t *pap)
1.43 thorpej 1745: {
1746: struct pool_cache_group *pcg;
1747: void *object;
1.58 thorpej 1748:
1749: #ifdef LOCKDEBUG
1750: if (flags & PR_WAITOK)
1751: simple_lock_only_held(NULL, "pool_cache_get(PR_WAITOK)");
1752: #endif
1.43 thorpej 1753:
1754: simple_lock(&pc->pc_slock);
1755:
1756: if ((pcg = pc->pc_allocfrom) == NULL) {
1.61 chs 1757: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.43 thorpej 1758: if (pcg->pcg_avail != 0) {
1759: pc->pc_allocfrom = pcg;
1760: goto have_group;
1761: }
1762: }
1763:
1764: /*
1765: * No groups with any available objects. Allocate
1766: * a new object, construct it, and return it to
1767: * the caller. We will allocate a group, if necessary,
1768: * when the object is freed back to the cache.
1769: */
1.48 thorpej 1770: pc->pc_misses++;
1.43 thorpej 1771: simple_unlock(&pc->pc_slock);
1772: object = pool_get(pc->pc_pool, flags);
1773: if (object != NULL && pc->pc_ctor != NULL) {
1774: if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) {
1775: pool_put(pc->pc_pool, object);
1776: return (NULL);
1777: }
1778: }
1.87 thorpej 1779: if (object != NULL && pap != NULL) {
1780: #ifdef POOL_VTOPHYS
1781: *pap = POOL_VTOPHYS(object);
1782: #else
1783: *pap = POOL_PADDR_INVALID;
1784: #endif
1785: }
1.43 thorpej 1786: return (object);
1787: }
1788:
1789: have_group:
1.48 thorpej 1790: pc->pc_hits++;
1791: pc->pc_nitems--;
1.87 thorpej 1792: object = pcg_get(pcg, pap);
1.43 thorpej 1793:
1794: if (pcg->pcg_avail == 0)
1795: pc->pc_allocfrom = NULL;
1.45 thorpej 1796:
1.43 thorpej 1797: simple_unlock(&pc->pc_slock);
1798:
1799: return (object);
1800: }
1801:
1802: /*
1.87 thorpej 1803: * pool_cache_put{,_paddr}:
1.43 thorpej 1804: *
1.87 thorpej 1805: * Put an object back to the pool cache (optionally caching the
1806: * physical address of the object).
1.43 thorpej 1807: */
1808: void
1.87 thorpej 1809: pool_cache_put_paddr(struct pool_cache *pc, void *object, paddr_t pa)
1.43 thorpej 1810: {
1811: struct pool_cache_group *pcg;
1.60 thorpej 1812: int s;
1.43 thorpej 1813:
1814: simple_lock(&pc->pc_slock);
1815:
1816: if ((pcg = pc->pc_freeto) == NULL) {
1.61 chs 1817: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.43 thorpej 1818: if (pcg->pcg_avail != PCG_NOBJECTS) {
1819: pc->pc_freeto = pcg;
1820: goto have_group;
1821: }
1822: }
1823:
1824: /*
1825: * No empty groups to free the object to. Attempt to
1.47 thorpej 1826: * allocate one.
1.43 thorpej 1827: */
1.47 thorpej 1828: simple_unlock(&pc->pc_slock);
1.60 thorpej 1829: s = splvm();
1.43 thorpej 1830: pcg = pool_get(&pcgpool, PR_NOWAIT);
1.60 thorpej 1831: splx(s);
1.43 thorpej 1832: if (pcg != NULL) {
1833: memset(pcg, 0, sizeof(*pcg));
1.47 thorpej 1834: simple_lock(&pc->pc_slock);
1.48 thorpej 1835: pc->pc_ngroups++;
1.43 thorpej 1836: TAILQ_INSERT_TAIL(&pc->pc_grouplist, pcg, pcg_list);
1.47 thorpej 1837: if (pc->pc_freeto == NULL)
1838: pc->pc_freeto = pcg;
1.43 thorpej 1839: goto have_group;
1840: }
1841:
1842: /*
1843: * Unable to allocate a cache group; destruct the object
1844: * and free it back to the pool.
1845: */
1.51 thorpej 1846: pool_cache_destruct_object(pc, object);
1.43 thorpej 1847: return;
1848: }
1849:
1850: have_group:
1.48 thorpej 1851: pc->pc_nitems++;
1.87 thorpej 1852: pcg_put(pcg, object, pa);
1.43 thorpej 1853:
1854: if (pcg->pcg_avail == PCG_NOBJECTS)
1855: pc->pc_freeto = NULL;
1856:
1857: simple_unlock(&pc->pc_slock);
1.51 thorpej 1858: }
1859:
1860: /*
1861: * pool_cache_destruct_object:
1862: *
1863: * Force destruction of an object and its release back into
1864: * the pool.
1865: */
1866: void
1867: pool_cache_destruct_object(struct pool_cache *pc, void *object)
1868: {
1869:
1870: if (pc->pc_dtor != NULL)
1871: (*pc->pc_dtor)(pc->pc_arg, object);
1872: pool_put(pc->pc_pool, object);
1.43 thorpej 1873: }
1874:
1875: /*
1876: * pool_cache_do_invalidate:
1877: *
1878: * This internal function implements pool_cache_invalidate() and
1879: * pool_cache_reclaim().
1880: */
1881: static void
1882: pool_cache_do_invalidate(struct pool_cache *pc, int free_groups,
1.56 sommerfe 1883: void (*putit)(struct pool *, void *))
1.43 thorpej 1884: {
1885: struct pool_cache_group *pcg, *npcg;
1886: void *object;
1.60 thorpej 1887: int s;
1.43 thorpej 1888:
1889: for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
1890: pcg = npcg) {
1891: npcg = TAILQ_NEXT(pcg, pcg_list);
1892: while (pcg->pcg_avail != 0) {
1.48 thorpej 1893: pc->pc_nitems--;
1.87 thorpej 1894: object = pcg_get(pcg, NULL);
1.45 thorpej 1895: if (pcg->pcg_avail == 0 && pc->pc_allocfrom == pcg)
1896: pc->pc_allocfrom = NULL;
1.43 thorpej 1897: if (pc->pc_dtor != NULL)
1898: (*pc->pc_dtor)(pc->pc_arg, object);
1.56 sommerfe 1899: (*putit)(pc->pc_pool, object);
1.43 thorpej 1900: }
1901: if (free_groups) {
1.48 thorpej 1902: pc->pc_ngroups--;
1.43 thorpej 1903: TAILQ_REMOVE(&pc->pc_grouplist, pcg, pcg_list);
1.46 thorpej 1904: if (pc->pc_freeto == pcg)
1905: pc->pc_freeto = NULL;
1.60 thorpej 1906: s = splvm();
1.43 thorpej 1907: pool_put(&pcgpool, pcg);
1.60 thorpej 1908: splx(s);
1.43 thorpej 1909: }
1910: }
1911: }
1912:
1913: /*
1914: * pool_cache_invalidate:
1915: *
1916: * Invalidate a pool cache (destruct and release all of the
1917: * cached objects).
1918: */
1919: void
1920: pool_cache_invalidate(struct pool_cache *pc)
1921: {
1922:
1923: simple_lock(&pc->pc_slock);
1.56 sommerfe 1924: pool_cache_do_invalidate(pc, 0, pool_put);
1.43 thorpej 1925: simple_unlock(&pc->pc_slock);
1926: }
1927:
1928: /*
1929: * pool_cache_reclaim:
1930: *
1931: * Reclaim a pool cache for pool_reclaim().
1932: */
1933: static void
1934: pool_cache_reclaim(struct pool_cache *pc)
1935: {
1936:
1.47 thorpej 1937: simple_lock(&pc->pc_slock);
1.43 thorpej 1938: pool_cache_do_invalidate(pc, 1, pool_do_put);
1939: simple_unlock(&pc->pc_slock);
1.3 pk 1940: }
1.66 thorpej 1941:
1942: /*
1943: * Pool backend allocators.
1944: *
1945: * Each pool has a backend allocator that handles allocation, deallocation,
1946: * and any additional draining that might be needed.
1947: *
1948: * We provide two standard allocators:
1949: *
1950: * pool_allocator_kmem - the default when no allocator is specified
1951: *
1952: * pool_allocator_nointr - used for pools that will not be accessed
1953: * in interrupt context.
1954: */
1955: void *pool_page_alloc(struct pool *, int);
1956: void pool_page_free(struct pool *, void *);
1957:
1958: struct pool_allocator pool_allocator_kmem = {
1959: pool_page_alloc, pool_page_free, 0,
1960: };
1961:
1962: void *pool_page_alloc_nointr(struct pool *, int);
1963: void pool_page_free_nointr(struct pool *, void *);
1964:
1965: struct pool_allocator pool_allocator_nointr = {
1966: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1967: };
1968:
1969: #ifdef POOL_SUBPAGE
1970: void *pool_subpage_alloc(struct pool *, int);
1971: void pool_subpage_free(struct pool *, void *);
1972:
1973: struct pool_allocator pool_allocator_kmem_subpage = {
1974: pool_subpage_alloc, pool_subpage_free, 0,
1975: };
1976: #endif /* POOL_SUBPAGE */
1977:
1978: /*
1979: * We have at least three different resources for the same allocation and
1980: * each resource can be depleted. First, we have the ready elements in the
1981: * pool. Then we have the resource (typically a vm_map) for this allocator.
1982: * Finally, we have physical memory. Waiting for any of these can be
1983: * unnecessary when any other is freed, but the kernel doesn't support
1984: * sleeping on multiple wait channels, so we have to employ another strategy.
1985: *
1986: * The caller sleeps on the pool (so that it can be awakened when an item
1987: * is returned to the pool), but we set PA_WANT on the allocator. When a
1988: * page is returned to the allocator and PA_WANT is set, pool_allocator_free
1989: * will wake up all sleeping pools belonging to this allocator.
1990: *
1991: * XXX Thundering herd.
1992: */
1993: void *
1994: pool_allocator_alloc(struct pool *org, int flags)
1995: {
1996: struct pool_allocator *pa = org->pr_alloc;
1997: struct pool *pp, *start;
1998: int s, freed;
1999: void *res;
2000:
1.91 yamt 2001: LOCK_ASSERT(!simple_lock_held(&org->pr_slock));
2002:
1.66 thorpej 2003: do {
2004: if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
2005: return (res);
1.68 thorpej 2006: if ((flags & PR_WAITOK) == 0) {
2007: /*
2008: * We only run the drain hookhere if PR_NOWAIT.
2009: * In other cases, the hook will be run in
2010: * pool_reclaim().
2011: */
2012: if (org->pr_drain_hook != NULL) {
2013: (*org->pr_drain_hook)(org->pr_drain_hook_arg,
2014: flags);
2015: if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
2016: return (res);
2017: }
1.66 thorpej 2018: break;
1.68 thorpej 2019: }
1.66 thorpej 2020:
2021: /*
2022: * Drain all pools, except "org", that use this
2023: * allocator. We do this to reclaim VA space.
2024: * pa_alloc is responsible for waiting for
2025: * physical memory.
2026: *
2027: * XXX We risk looping forever if start if someone
2028: * calls pool_destroy on "start". But there is no
2029: * other way to have potentially sleeping pool_reclaim,
2030: * non-sleeping locks on pool_allocator, and some
2031: * stirring of drained pools in the allocator.
1.68 thorpej 2032: *
2033: * XXX Maybe we should use pool_head_slock for locking
2034: * the allocators?
1.66 thorpej 2035: */
2036: freed = 0;
2037:
2038: s = splvm();
2039: simple_lock(&pa->pa_slock);
2040: pp = start = TAILQ_FIRST(&pa->pa_list);
2041: do {
2042: TAILQ_REMOVE(&pa->pa_list, pp, pr_alloc_list);
2043: TAILQ_INSERT_TAIL(&pa->pa_list, pp, pr_alloc_list);
2044: if (pp == org)
2045: continue;
1.73 thorpej 2046: simple_unlock(&pa->pa_slock);
1.66 thorpej 2047: freed = pool_reclaim(pp);
1.73 thorpej 2048: simple_lock(&pa->pa_slock);
1.66 thorpej 2049: } while ((pp = TAILQ_FIRST(&pa->pa_list)) != start &&
2050: freed == 0);
2051:
2052: if (freed == 0) {
2053: /*
2054: * We set PA_WANT here, the caller will most likely
2055: * sleep waiting for pages (if not, this won't hurt
2056: * that much), and there is no way to set this in
2057: * the caller without violating locking order.
2058: */
2059: pa->pa_flags |= PA_WANT;
2060: }
2061: simple_unlock(&pa->pa_slock);
2062: splx(s);
2063: } while (freed);
2064: return (NULL);
2065: }
2066:
2067: void
2068: pool_allocator_free(struct pool *pp, void *v)
2069: {
2070: struct pool_allocator *pa = pp->pr_alloc;
2071: int s;
2072:
1.91 yamt 2073: LOCK_ASSERT(!simple_lock_held(&pp->pr_slock));
2074:
1.66 thorpej 2075: (*pa->pa_free)(pp, v);
2076:
2077: s = splvm();
2078: simple_lock(&pa->pa_slock);
2079: if ((pa->pa_flags & PA_WANT) == 0) {
2080: simple_unlock(&pa->pa_slock);
2081: splx(s);
2082: return;
2083: }
2084:
2085: TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
2086: simple_lock(&pp->pr_slock);
2087: if ((pp->pr_flags & PR_WANTED) != 0) {
2088: pp->pr_flags &= ~PR_WANTED;
2089: wakeup(pp);
2090: }
1.69 thorpej 2091: simple_unlock(&pp->pr_slock);
1.66 thorpej 2092: }
2093: pa->pa_flags &= ~PA_WANT;
2094: simple_unlock(&pa->pa_slock);
2095: splx(s);
2096: }
2097:
2098: void *
2099: pool_page_alloc(struct pool *pp, int flags)
2100: {
2101: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2102:
2103: return ((void *) uvm_km_alloc_poolpage(waitok));
2104: }
2105:
2106: void
2107: pool_page_free(struct pool *pp, void *v)
2108: {
2109:
2110: uvm_km_free_poolpage((vaddr_t) v);
2111: }
2112:
2113: #ifdef POOL_SUBPAGE
2114: /* Sub-page allocator, for machines with large hardware pages. */
2115: void *
2116: pool_subpage_alloc(struct pool *pp, int flags)
2117: {
1.93 dbj 2118: void *v;
2119: int s;
2120: s = splvm();
2121: v = pool_get(&psppool, flags);
2122: splx(s);
2123: return v;
1.66 thorpej 2124: }
2125:
2126: void
2127: pool_subpage_free(struct pool *pp, void *v)
2128: {
1.93 dbj 2129: int s;
2130: s = splvm();
1.66 thorpej 2131: pool_put(&psppool, v);
1.93 dbj 2132: splx(s);
1.66 thorpej 2133: }
2134:
2135: /* We don't provide a real nointr allocator. Maybe later. */
2136: void *
2137: pool_page_alloc_nointr(struct pool *pp, int flags)
2138: {
2139:
2140: return (pool_subpage_alloc(pp, flags));
2141: }
2142:
2143: void
2144: pool_page_free_nointr(struct pool *pp, void *v)
2145: {
2146:
2147: pool_subpage_free(pp, v);
2148: }
2149: #else
2150: void *
2151: pool_page_alloc_nointr(struct pool *pp, int flags)
2152: {
2153: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2154:
2155: return ((void *) uvm_km_alloc_poolpage1(kernel_map,
2156: uvm.kernel_object, waitok));
2157: }
2158:
2159: void
2160: pool_page_free_nointr(struct pool *pp, void *v)
2161: {
2162:
2163: uvm_km_free_poolpage1(kernel_map, (vaddr_t) v);
2164: }
2165: #endif /* POOL_SUBPAGE */
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