Annotation of src/sys/kern/subr_pool.c, Revision 1.93
1.93 ! dbj 1: /* $NetBSD: subr_pool.c,v 1.92 2004/02/22 00:19:48 enami 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>
1.93 ! dbj 41: __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.92 2004/02/22 00:19:48 enami 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: /*
360: * Initialize the given pool resource structure.
361: *
362: * We export this routine to allow other kernel parts to declare
363: * static pools that must be initialized before malloc() is available.
364: */
365: void
1.42 thorpej 366: pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
1.66 thorpej 367: const char *wchan, struct pool_allocator *palloc)
1.3 pk 368: {
1.88 chs 369: int off, slack;
1.92 enami 370: size_t trysize, phsize;
1.93 ! dbj 371: int s;
1.3 pk 372:
1.25 thorpej 373: #ifdef POOL_DIAGNOSTIC
374: /*
375: * Always log if POOL_DIAGNOSTIC is defined.
376: */
377: if (pool_logsize != 0)
378: flags |= PR_LOGGING;
379: #endif
380:
1.66 thorpej 381: #ifdef POOL_SUBPAGE
382: /*
383: * XXX We don't provide a real `nointr' back-end
384: * yet; all sub-pages come from a kmem back-end.
385: * maybe some day...
386: */
387: if (palloc == NULL) {
388: extern struct pool_allocator pool_allocator_kmem_subpage;
389: palloc = &pool_allocator_kmem_subpage;
390: }
1.3 pk 391: /*
1.66 thorpej 392: * We'll assume any user-specified back-end allocator
393: * will deal with sub-pages, or simply don't care.
1.3 pk 394: */
1.66 thorpej 395: #else
396: if (palloc == NULL)
397: palloc = &pool_allocator_kmem;
398: #endif /* POOL_SUBPAGE */
399: if ((palloc->pa_flags & PA_INITIALIZED) == 0) {
400: if (palloc->pa_pagesz == 0) {
1.62 bjh21 401: #ifdef POOL_SUBPAGE
1.66 thorpej 402: if (palloc == &pool_allocator_kmem)
403: palloc->pa_pagesz = PAGE_SIZE;
404: else
405: palloc->pa_pagesz = POOL_SUBPAGE;
1.62 bjh21 406: #else
1.66 thorpej 407: palloc->pa_pagesz = PAGE_SIZE;
408: #endif /* POOL_SUBPAGE */
409: }
410:
411: TAILQ_INIT(&palloc->pa_list);
412:
413: simple_lock_init(&palloc->pa_slock);
414: palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
415: palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
416: palloc->pa_flags |= PA_INITIALIZED;
1.4 thorpej 417: }
1.3 pk 418:
419: if (align == 0)
420: align = ALIGN(1);
1.14 thorpej 421:
422: if (size < sizeof(struct pool_item))
423: size = sizeof(struct pool_item);
1.3 pk 424:
1.78 thorpej 425: size = roundup(size, align);
1.66 thorpej 426: #ifdef DIAGNOSTIC
427: if (size > palloc->pa_pagesz)
1.35 pk 428: panic("pool_init: pool item size (%lu) too large",
429: (u_long)size);
1.66 thorpej 430: #endif
1.35 pk 431:
1.3 pk 432: /*
433: * Initialize the pool structure.
434: */
1.88 chs 435: LIST_INIT(&pp->pr_emptypages);
436: LIST_INIT(&pp->pr_fullpages);
437: LIST_INIT(&pp->pr_partpages);
1.43 thorpej 438: TAILQ_INIT(&pp->pr_cachelist);
1.3 pk 439: pp->pr_curpage = NULL;
440: pp->pr_npages = 0;
441: pp->pr_minitems = 0;
442: pp->pr_minpages = 0;
443: pp->pr_maxpages = UINT_MAX;
1.20 thorpej 444: pp->pr_roflags = flags;
445: pp->pr_flags = 0;
1.35 pk 446: pp->pr_size = size;
1.3 pk 447: pp->pr_align = align;
448: pp->pr_wchan = wchan;
1.66 thorpej 449: pp->pr_alloc = palloc;
1.20 thorpej 450: pp->pr_nitems = 0;
451: pp->pr_nout = 0;
452: pp->pr_hardlimit = UINT_MAX;
453: pp->pr_hardlimit_warning = NULL;
1.31 thorpej 454: pp->pr_hardlimit_ratecap.tv_sec = 0;
455: pp->pr_hardlimit_ratecap.tv_usec = 0;
456: pp->pr_hardlimit_warning_last.tv_sec = 0;
457: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.68 thorpej 458: pp->pr_drain_hook = NULL;
459: pp->pr_drain_hook_arg = NULL;
1.3 pk 460:
461: /*
462: * Decide whether to put the page header off page to avoid
1.92 enami 463: * wasting too large a part of the page or too big item.
464: * Off-page page headers go on a hash table, so we can match
465: * a returned item with its header based on the page address.
466: * We use 1/16 of the page size and about 8 times of the item
467: * size as the threshold (XXX: tune)
468: *
469: * However, we'll put the header into the page if we can put
470: * it without wasting any items.
471: *
472: * Silently enforce `0 <= ioff < align'.
1.3 pk 473: */
1.92 enami 474: pp->pr_itemoffset = ioff %= align;
475: /* See the comment below about reserved bytes. */
476: trysize = palloc->pa_pagesz - ((align - ioff) % align);
477: phsize = ALIGN(sizeof(struct pool_item_header));
478: if (pp->pr_size < MIN(palloc->pa_pagesz / 16, phsize << 3) ||
479: trysize / pp->pr_size == (trysize - phsize) / pp->pr_size) {
1.3 pk 480: /* Use the end of the page for the page header */
1.20 thorpej 481: pp->pr_roflags |= PR_PHINPAGE;
1.92 enami 482: pp->pr_phoffset = off = palloc->pa_pagesz - phsize;
1.2 pk 483: } else {
1.3 pk 484: /* The page header will be taken from our page header pool */
485: pp->pr_phoffset = 0;
1.66 thorpej 486: off = palloc->pa_pagesz;
1.88 chs 487: SPLAY_INIT(&pp->pr_phtree);
1.2 pk 488: }
1.1 pk 489:
1.3 pk 490: /*
491: * Alignment is to take place at `ioff' within the item. This means
492: * we must reserve up to `align - 1' bytes on the page to allow
493: * appropriate positioning of each item.
494: */
495: pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
1.43 thorpej 496: KASSERT(pp->pr_itemsperpage != 0);
1.3 pk 497:
498: /*
499: * Use the slack between the chunks and the page header
500: * for "cache coloring".
501: */
502: slack = off - pp->pr_itemsperpage * pp->pr_size;
503: pp->pr_maxcolor = (slack / align) * align;
504: pp->pr_curcolor = 0;
505:
506: pp->pr_nget = 0;
507: pp->pr_nfail = 0;
508: pp->pr_nput = 0;
509: pp->pr_npagealloc = 0;
510: pp->pr_npagefree = 0;
1.1 pk 511: pp->pr_hiwat = 0;
1.8 thorpej 512: pp->pr_nidle = 0;
1.3 pk 513:
1.59 thorpej 514: #ifdef POOL_DIAGNOSTIC
1.25 thorpej 515: if (flags & PR_LOGGING) {
516: if (kmem_map == NULL ||
517: (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log),
518: M_TEMP, M_NOWAIT)) == NULL)
1.20 thorpej 519: pp->pr_roflags &= ~PR_LOGGING;
1.3 pk 520: pp->pr_curlogentry = 0;
521: pp->pr_logsize = pool_logsize;
522: }
1.59 thorpej 523: #endif
1.25 thorpej 524:
525: pp->pr_entered_file = NULL;
526: pp->pr_entered_line = 0;
1.3 pk 527:
1.21 thorpej 528: simple_lock_init(&pp->pr_slock);
1.1 pk 529:
1.3 pk 530: /*
1.43 thorpej 531: * Initialize private page header pool and cache magazine pool if we
532: * haven't done so yet.
1.23 thorpej 533: * XXX LOCKING.
1.3 pk 534: */
535: if (phpool.pr_size == 0) {
1.62 bjh21 536: #ifdef POOL_SUBPAGE
537: pool_init(&phpool, sizeof(struct pool_item_header), 0, 0, 0,
1.66 thorpej 538: "phpool", &pool_allocator_kmem);
1.62 bjh21 539: pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
1.66 thorpej 540: PR_RECURSIVE, "psppool", &pool_allocator_kmem);
1.62 bjh21 541: #else
1.3 pk 542: pool_init(&phpool, sizeof(struct pool_item_header), 0, 0,
1.66 thorpej 543: 0, "phpool", NULL);
1.62 bjh21 544: #endif
1.43 thorpej 545: pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
1.66 thorpej 546: 0, "pcgpool", NULL);
1.1 pk 547: }
548:
1.23 thorpej 549: /* Insert into the list of all pools. */
550: simple_lock(&pool_head_slock);
551: TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist);
552: simple_unlock(&pool_head_slock);
1.66 thorpej 553:
554: /* Insert this into the list of pools using this allocator. */
1.93 ! dbj 555: s = splvm();
1.66 thorpej 556: simple_lock(&palloc->pa_slock);
557: TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
558: simple_unlock(&palloc->pa_slock);
1.93 ! dbj 559: splx(s);
1.1 pk 560: }
561:
562: /*
563: * De-commision a pool resource.
564: */
565: void
1.42 thorpej 566: pool_destroy(struct pool *pp)
1.1 pk 567: {
1.3 pk 568: struct pool_item_header *ph;
1.43 thorpej 569: struct pool_cache *pc;
1.93 ! dbj 570: int s;
1.43 thorpej 571:
1.66 thorpej 572: /* Locking order: pool_allocator -> pool */
1.93 ! dbj 573: s = splvm();
1.66 thorpej 574: simple_lock(&pp->pr_alloc->pa_slock);
575: TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
576: simple_unlock(&pp->pr_alloc->pa_slock);
1.93 ! dbj 577: splx(s);
1.66 thorpej 578:
1.43 thorpej 579: /* Destroy all caches for this pool. */
580: while ((pc = TAILQ_FIRST(&pp->pr_cachelist)) != NULL)
581: pool_cache_destroy(pc);
1.3 pk 582:
583: #ifdef DIAGNOSTIC
1.20 thorpej 584: if (pp->pr_nout != 0) {
1.25 thorpej 585: pr_printlog(pp, NULL, printf);
1.80 provos 586: panic("pool_destroy: pool busy: still out: %u",
1.20 thorpej 587: pp->pr_nout);
1.3 pk 588: }
589: #endif
1.1 pk 590:
1.3 pk 591: /* Remove all pages */
1.88 chs 592: while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1.70 thorpej 593: pr_rmpage(pp, ph, NULL);
1.88 chs 594: KASSERT(LIST_EMPTY(&pp->pr_fullpages));
595: KASSERT(LIST_EMPTY(&pp->pr_partpages));
1.3 pk 596:
597: /* Remove from global pool list */
1.23 thorpej 598: simple_lock(&pool_head_slock);
1.3 pk 599: TAILQ_REMOVE(&pool_head, pp, pr_poollist);
1.61 chs 600: if (drainpp == pp) {
601: drainpp = NULL;
602: }
1.23 thorpej 603: simple_unlock(&pool_head_slock);
1.3 pk 604:
1.59 thorpej 605: #ifdef POOL_DIAGNOSTIC
1.20 thorpej 606: if ((pp->pr_roflags & PR_LOGGING) != 0)
1.3 pk 607: free(pp->pr_log, M_TEMP);
1.59 thorpej 608: #endif
1.1 pk 609: }
610:
1.68 thorpej 611: void
612: pool_set_drain_hook(struct pool *pp, void (*fn)(void *, int), void *arg)
613: {
614:
615: /* XXX no locking -- must be used just after pool_init() */
616: #ifdef DIAGNOSTIC
617: if (pp->pr_drain_hook != NULL)
618: panic("pool_set_drain_hook(%s): already set", pp->pr_wchan);
619: #endif
620: pp->pr_drain_hook = fn;
621: pp->pr_drain_hook_arg = arg;
622: }
623:
1.88 chs 624: static struct pool_item_header *
1.55 thorpej 625: pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags)
626: {
627: struct pool_item_header *ph;
628: int s;
629:
630: LOCK_ASSERT(simple_lock_held(&pp->pr_slock) == 0);
631:
632: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
633: ph = (struct pool_item_header *) (storage + pp->pr_phoffset);
634: else {
1.85 pk 635: s = splvm();
1.55 thorpej 636: ph = pool_get(&phpool, flags);
637: splx(s);
638: }
639:
640: return (ph);
641: }
1.1 pk 642:
643: /*
1.3 pk 644: * Grab an item from the pool; must be called at appropriate spl level
1.1 pk 645: */
1.3 pk 646: void *
1.59 thorpej 647: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 648: _pool_get(struct pool *pp, int flags, const char *file, long line)
1.56 sommerfe 649: #else
650: pool_get(struct pool *pp, int flags)
651: #endif
1.1 pk 652: {
653: struct pool_item *pi;
1.3 pk 654: struct pool_item_header *ph;
1.55 thorpej 655: void *v;
1.1 pk 656:
1.2 pk 657: #ifdef DIAGNOSTIC
1.84 thorpej 658: if (__predict_false(curlwp == NULL && doing_shutdown == 0 &&
1.37 sommerfe 659: (flags & PR_WAITOK) != 0))
1.77 matt 660: panic("pool_get: %s: must have NOWAIT", pp->pr_wchan);
1.58 thorpej 661:
662: #ifdef LOCKDEBUG
663: if (flags & PR_WAITOK)
664: simple_lock_only_held(NULL, "pool_get(PR_WAITOK)");
1.56 sommerfe 665: #endif
1.58 thorpej 666: #endif /* DIAGNOSTIC */
1.1 pk 667:
1.21 thorpej 668: simple_lock(&pp->pr_slock);
1.25 thorpej 669: pr_enter(pp, file, line);
1.20 thorpej 670:
671: startover:
672: /*
673: * Check to see if we've reached the hard limit. If we have,
674: * and we can wait, then wait until an item has been returned to
675: * the pool.
676: */
677: #ifdef DIAGNOSTIC
1.34 thorpej 678: if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
1.25 thorpej 679: pr_leave(pp);
1.21 thorpej 680: simple_unlock(&pp->pr_slock);
1.20 thorpej 681: panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
682: }
683: #endif
1.34 thorpej 684: if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
1.68 thorpej 685: if (pp->pr_drain_hook != NULL) {
686: /*
687: * Since the drain hook is going to free things
688: * back to the pool, unlock, call the hook, re-lock,
689: * and check the hardlimit condition again.
690: */
691: pr_leave(pp);
692: simple_unlock(&pp->pr_slock);
693: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
694: simple_lock(&pp->pr_slock);
695: pr_enter(pp, file, line);
696: if (pp->pr_nout < pp->pr_hardlimit)
697: goto startover;
698: }
699:
1.29 sommerfe 700: if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
1.20 thorpej 701: /*
702: * XXX: A warning isn't logged in this case. Should
703: * it be?
704: */
705: pp->pr_flags |= PR_WANTED;
1.25 thorpej 706: pr_leave(pp);
1.40 sommerfe 707: ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
1.25 thorpej 708: pr_enter(pp, file, line);
1.20 thorpej 709: goto startover;
710: }
1.31 thorpej 711:
712: /*
713: * Log a message that the hard limit has been hit.
714: */
715: if (pp->pr_hardlimit_warning != NULL &&
716: ratecheck(&pp->pr_hardlimit_warning_last,
717: &pp->pr_hardlimit_ratecap))
718: log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
1.21 thorpej 719:
720: pp->pr_nfail++;
721:
1.25 thorpej 722: pr_leave(pp);
1.21 thorpej 723: simple_unlock(&pp->pr_slock);
1.20 thorpej 724: return (NULL);
725: }
726:
1.3 pk 727: /*
728: * The convention we use is that if `curpage' is not NULL, then
729: * it points at a non-empty bucket. In particular, `curpage'
730: * never points at a page header which has PR_PHINPAGE set and
731: * has no items in its bucket.
732: */
1.20 thorpej 733: if ((ph = pp->pr_curpage) == NULL) {
734: #ifdef DIAGNOSTIC
735: if (pp->pr_nitems != 0) {
1.21 thorpej 736: simple_unlock(&pp->pr_slock);
1.20 thorpej 737: printf("pool_get: %s: curpage NULL, nitems %u\n",
738: pp->pr_wchan, pp->pr_nitems);
1.80 provos 739: panic("pool_get: nitems inconsistent");
1.20 thorpej 740: }
741: #endif
742:
1.21 thorpej 743: /*
744: * Call the back-end page allocator for more memory.
745: * Release the pool lock, as the back-end page allocator
746: * may block.
747: */
1.25 thorpej 748: pr_leave(pp);
1.21 thorpej 749: simple_unlock(&pp->pr_slock);
1.66 thorpej 750: v = pool_allocator_alloc(pp, flags);
1.55 thorpej 751: if (__predict_true(v != NULL))
752: ph = pool_alloc_item_header(pp, v, flags);
1.15 pk 753:
1.55 thorpej 754: if (__predict_false(v == NULL || ph == NULL)) {
755: if (v != NULL)
1.66 thorpej 756: pool_allocator_free(pp, v);
1.55 thorpej 757:
1.91 yamt 758: simple_lock(&pp->pr_slock);
759: pr_enter(pp, file, line);
760:
1.21 thorpej 761: /*
1.55 thorpej 762: * We were unable to allocate a page or item
763: * header, but we released the lock during
764: * allocation, so perhaps items were freed
765: * back to the pool. Check for this case.
1.21 thorpej 766: */
767: if (pp->pr_curpage != NULL)
768: goto startover;
1.15 pk 769:
1.3 pk 770: if ((flags & PR_WAITOK) == 0) {
771: pp->pr_nfail++;
1.25 thorpej 772: pr_leave(pp);
1.21 thorpej 773: simple_unlock(&pp->pr_slock);
1.1 pk 774: return (NULL);
1.3 pk 775: }
776:
1.15 pk 777: /*
778: * Wait for items to be returned to this pool.
1.21 thorpej 779: *
1.20 thorpej 780: * XXX: maybe we should wake up once a second and
781: * try again?
1.15 pk 782: */
1.1 pk 783: pp->pr_flags |= PR_WANTED;
1.66 thorpej 784: /* PA_WANTED is already set on the allocator. */
1.25 thorpej 785: pr_leave(pp);
1.40 sommerfe 786: ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
1.25 thorpej 787: pr_enter(pp, file, line);
1.20 thorpej 788: goto startover;
1.1 pk 789: }
1.3 pk 790:
1.15 pk 791: /* We have more memory; add it to the pool */
1.91 yamt 792: simple_lock(&pp->pr_slock);
793: pr_enter(pp, file, line);
1.55 thorpej 794: pool_prime_page(pp, v, ph);
1.15 pk 795: pp->pr_npagealloc++;
796:
1.20 thorpej 797: /* Start the allocation process over. */
798: goto startover;
1.3 pk 799: }
1.34 thorpej 800: if (__predict_false((v = pi = TAILQ_FIRST(&ph->ph_itemlist)) == NULL)) {
1.25 thorpej 801: pr_leave(pp);
1.21 thorpej 802: simple_unlock(&pp->pr_slock);
1.3 pk 803: panic("pool_get: %s: page empty", pp->pr_wchan);
1.21 thorpej 804: }
1.20 thorpej 805: #ifdef DIAGNOSTIC
1.34 thorpej 806: if (__predict_false(pp->pr_nitems == 0)) {
1.25 thorpej 807: pr_leave(pp);
1.21 thorpej 808: simple_unlock(&pp->pr_slock);
1.20 thorpej 809: printf("pool_get: %s: items on itemlist, nitems %u\n",
810: pp->pr_wchan, pp->pr_nitems);
1.80 provos 811: panic("pool_get: nitems inconsistent");
1.20 thorpej 812: }
1.65 enami 813: #endif
1.56 sommerfe 814:
1.65 enami 815: #ifdef POOL_DIAGNOSTIC
1.3 pk 816: pr_log(pp, v, PRLOG_GET, file, line);
1.65 enami 817: #endif
1.3 pk 818:
1.65 enami 819: #ifdef DIAGNOSTIC
1.34 thorpej 820: if (__predict_false(pi->pi_magic != PI_MAGIC)) {
1.25 thorpej 821: pr_printlog(pp, pi, printf);
1.3 pk 822: panic("pool_get(%s): free list modified: magic=%x; page %p;"
823: " item addr %p\n",
824: pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
825: }
826: #endif
827:
828: /*
829: * Remove from item list.
830: */
831: TAILQ_REMOVE(&ph->ph_itemlist, pi, pi_list);
1.20 thorpej 832: pp->pr_nitems--;
833: pp->pr_nout++;
1.6 thorpej 834: if (ph->ph_nmissing == 0) {
835: #ifdef DIAGNOSTIC
1.34 thorpej 836: if (__predict_false(pp->pr_nidle == 0))
1.6 thorpej 837: panic("pool_get: nidle inconsistent");
838: #endif
839: pp->pr_nidle--;
1.88 chs 840:
841: /*
842: * This page was previously empty. Move it to the list of
843: * partially-full pages. This page is already curpage.
844: */
845: LIST_REMOVE(ph, ph_pagelist);
846: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.6 thorpej 847: }
1.3 pk 848: ph->ph_nmissing++;
1.88 chs 849: if (TAILQ_EMPTY(&ph->ph_itemlist)) {
1.21 thorpej 850: #ifdef DIAGNOSTIC
1.34 thorpej 851: if (__predict_false(ph->ph_nmissing != pp->pr_itemsperpage)) {
1.25 thorpej 852: pr_leave(pp);
1.21 thorpej 853: simple_unlock(&pp->pr_slock);
854: panic("pool_get: %s: nmissing inconsistent",
855: pp->pr_wchan);
856: }
857: #endif
1.3 pk 858: /*
1.88 chs 859: * This page is now full. Move it to the full list
860: * and select a new current page.
1.3 pk 861: */
1.88 chs 862: LIST_REMOVE(ph, ph_pagelist);
863: LIST_INSERT_HEAD(&pp->pr_fullpages, ph, ph_pagelist);
864: pool_update_curpage(pp);
1.1 pk 865: }
1.3 pk 866:
867: pp->pr_nget++;
1.20 thorpej 868:
869: /*
870: * If we have a low water mark and we are now below that low
871: * water mark, add more items to the pool.
872: */
1.53 thorpej 873: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 874: /*
875: * XXX: Should we log a warning? Should we set up a timeout
876: * to try again in a second or so? The latter could break
877: * a caller's assumptions about interrupt protection, etc.
878: */
879: }
880:
1.25 thorpej 881: pr_leave(pp);
1.21 thorpej 882: simple_unlock(&pp->pr_slock);
1.1 pk 883: return (v);
884: }
885:
886: /*
1.43 thorpej 887: * Internal version of pool_put(). Pool is already locked/entered.
1.1 pk 888: */
1.43 thorpej 889: static void
1.56 sommerfe 890: pool_do_put(struct pool *pp, void *v)
1.1 pk 891: {
892: struct pool_item *pi = v;
1.3 pk 893: struct pool_item_header *ph;
894: caddr_t page;
1.21 thorpej 895: int s;
1.3 pk 896:
1.61 chs 897: LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
898:
1.66 thorpej 899: page = (caddr_t)((u_long)v & pp->pr_alloc->pa_pagemask);
1.1 pk 900:
1.30 thorpej 901: #ifdef DIAGNOSTIC
1.34 thorpej 902: if (__predict_false(pp->pr_nout == 0)) {
1.30 thorpej 903: printf("pool %s: putting with none out\n",
904: pp->pr_wchan);
905: panic("pool_put");
906: }
907: #endif
1.3 pk 908:
1.34 thorpej 909: if (__predict_false((ph = pr_find_pagehead(pp, page)) == NULL)) {
1.25 thorpej 910: pr_printlog(pp, NULL, printf);
1.3 pk 911: panic("pool_put: %s: page header missing", pp->pr_wchan);
912: }
1.28 thorpej 913:
914: #ifdef LOCKDEBUG
915: /*
916: * Check if we're freeing a locked simple lock.
917: */
918: simple_lock_freecheck((caddr_t)pi, ((caddr_t)pi) + pp->pr_size);
919: #endif
1.3 pk 920:
921: /*
922: * Return to item list.
923: */
1.2 pk 924: #ifdef DIAGNOSTIC
1.3 pk 925: pi->pi_magic = PI_MAGIC;
926: #endif
1.32 chs 927: #ifdef DEBUG
928: {
929: int i, *ip = v;
930:
931: for (i = 0; i < pp->pr_size / sizeof(int); i++) {
932: *ip++ = PI_MAGIC;
933: }
934: }
935: #endif
936:
1.3 pk 937: TAILQ_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.79 thorpej 938: KDASSERT(ph->ph_nmissing != 0);
1.3 pk 939: ph->ph_nmissing--;
940: pp->pr_nput++;
1.20 thorpej 941: pp->pr_nitems++;
942: pp->pr_nout--;
1.3 pk 943:
944: /* Cancel "pool empty" condition if it exists */
945: if (pp->pr_curpage == NULL)
946: pp->pr_curpage = ph;
947:
948: if (pp->pr_flags & PR_WANTED) {
949: pp->pr_flags &= ~PR_WANTED;
1.15 pk 950: if (ph->ph_nmissing == 0)
951: pp->pr_nidle++;
1.3 pk 952: wakeup((caddr_t)pp);
953: return;
954: }
955:
956: /*
1.88 chs 957: * If this page is now empty, do one of two things:
1.21 thorpej 958: *
1.88 chs 959: * (1) If we have more pages than the page high water mark,
1.90 thorpej 960: * or if we are flagged as immediately freeing back idle
961: * pages, free the page back to the system. ONLY CONSIDER
962: * FREEING BACK A PAGE IF WE HAVE MORE THAN OUR MINIMUM PAGE
963: * CLAIM.
1.21 thorpej 964: *
1.88 chs 965: * (2) Otherwise, move the page to the empty page list.
966: *
967: * Either way, select a new current page (so we use a partially-full
968: * page if one is available).
1.3 pk 969: */
970: if (ph->ph_nmissing == 0) {
1.6 thorpej 971: pp->pr_nidle++;
1.90 thorpej 972: if (pp->pr_npages > pp->pr_minpages &&
973: (pp->pr_npages > pp->pr_maxpages ||
974: (pp->pr_roflags & PR_IMMEDRELEASE) != 0 ||
975: (pp->pr_alloc->pa_flags & PA_WANT) != 0)) {
1.91 yamt 976: simple_unlock(&pp->pr_slock);
1.61 chs 977: pr_rmpage(pp, ph, NULL);
1.91 yamt 978: simple_lock(&pp->pr_slock);
1.3 pk 979: } else {
1.88 chs 980: LIST_REMOVE(ph, ph_pagelist);
981: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.3 pk 982:
1.21 thorpej 983: /*
984: * Update the timestamp on the page. A page must
985: * be idle for some period of time before it can
986: * be reclaimed by the pagedaemon. This minimizes
987: * ping-pong'ing for memory.
988: */
989: s = splclock();
990: ph->ph_time = mono_time;
991: splx(s);
1.1 pk 992: }
1.88 chs 993: pool_update_curpage(pp);
1.1 pk 994: }
1.88 chs 995:
1.21 thorpej 996: /*
1.88 chs 997: * If the page was previously completely full, move it to the
998: * partially-full list and make it the current page. The next
999: * allocation will get the item from this page, instead of
1000: * further fragmenting the pool.
1.21 thorpej 1001: */
1002: else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
1.88 chs 1003: LIST_REMOVE(ph, ph_pagelist);
1004: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.21 thorpej 1005: pp->pr_curpage = ph;
1006: }
1.43 thorpej 1007: }
1008:
1009: /*
1010: * Return resource to the pool; must be called at appropriate spl level
1011: */
1.59 thorpej 1012: #ifdef POOL_DIAGNOSTIC
1.43 thorpej 1013: void
1014: _pool_put(struct pool *pp, void *v, const char *file, long line)
1015: {
1016:
1017: simple_lock(&pp->pr_slock);
1018: pr_enter(pp, file, line);
1019:
1.56 sommerfe 1020: pr_log(pp, v, PRLOG_PUT, file, line);
1021:
1022: pool_do_put(pp, v);
1.21 thorpej 1023:
1.25 thorpej 1024: pr_leave(pp);
1.21 thorpej 1025: simple_unlock(&pp->pr_slock);
1.1 pk 1026: }
1.57 sommerfe 1027: #undef pool_put
1.59 thorpej 1028: #endif /* POOL_DIAGNOSTIC */
1.1 pk 1029:
1.56 sommerfe 1030: void
1031: pool_put(struct pool *pp, void *v)
1032: {
1033:
1034: simple_lock(&pp->pr_slock);
1035:
1036: pool_do_put(pp, v);
1037:
1038: simple_unlock(&pp->pr_slock);
1039: }
1.57 sommerfe 1040:
1.59 thorpej 1041: #ifdef POOL_DIAGNOSTIC
1.57 sommerfe 1042: #define pool_put(h, v) _pool_put((h), (v), __FILE__, __LINE__)
1.56 sommerfe 1043: #endif
1.74 thorpej 1044:
1045: /*
1046: * Add N items to the pool.
1047: */
1048: int
1049: pool_prime(struct pool *pp, int n)
1050: {
1.83 scw 1051: struct pool_item_header *ph = NULL;
1.74 thorpej 1052: caddr_t cp;
1.75 simonb 1053: int newpages;
1.74 thorpej 1054:
1055: simple_lock(&pp->pr_slock);
1056:
1057: newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1058:
1059: while (newpages-- > 0) {
1060: simple_unlock(&pp->pr_slock);
1061: cp = pool_allocator_alloc(pp, PR_NOWAIT);
1062: if (__predict_true(cp != NULL))
1063: ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1064:
1065: if (__predict_false(cp == NULL || ph == NULL)) {
1066: if (cp != NULL)
1067: pool_allocator_free(pp, cp);
1.91 yamt 1068: simple_lock(&pp->pr_slock);
1.74 thorpej 1069: break;
1070: }
1071:
1.91 yamt 1072: simple_lock(&pp->pr_slock);
1.74 thorpej 1073: pool_prime_page(pp, cp, ph);
1074: pp->pr_npagealloc++;
1075: pp->pr_minpages++;
1076: }
1077:
1078: if (pp->pr_minpages >= pp->pr_maxpages)
1079: pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */
1080:
1081: simple_unlock(&pp->pr_slock);
1082: return (0);
1083: }
1.55 thorpej 1084:
1085: /*
1.3 pk 1086: * Add a page worth of items to the pool.
1.21 thorpej 1087: *
1088: * Note, we must be called with the pool descriptor LOCKED.
1.3 pk 1089: */
1.55 thorpej 1090: static void
1091: pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph)
1.3 pk 1092: {
1093: struct pool_item *pi;
1094: caddr_t cp = storage;
1095: unsigned int align = pp->pr_align;
1096: unsigned int ioff = pp->pr_itemoffset;
1.55 thorpej 1097: int n;
1.89 yamt 1098: int s;
1.36 pk 1099:
1.91 yamt 1100: LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
1101:
1.66 thorpej 1102: #ifdef DIAGNOSTIC
1103: if (((u_long)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
1.36 pk 1104: panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1.66 thorpej 1105: #endif
1.3 pk 1106:
1107: /*
1108: * Insert page header.
1109: */
1.88 chs 1110: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.3 pk 1111: TAILQ_INIT(&ph->ph_itemlist);
1112: ph->ph_page = storage;
1113: ph->ph_nmissing = 0;
1.89 yamt 1114: s = splclock();
1115: ph->ph_time = mono_time;
1116: splx(s);
1.88 chs 1117: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1118: SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
1.3 pk 1119:
1.6 thorpej 1120: pp->pr_nidle++;
1121:
1.3 pk 1122: /*
1123: * Color this page.
1124: */
1125: cp = (caddr_t)(cp + pp->pr_curcolor);
1126: if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
1127: pp->pr_curcolor = 0;
1128:
1129: /*
1130: * Adjust storage to apply aligment to `pr_itemoffset' in each item.
1131: */
1132: if (ioff != 0)
1133: cp = (caddr_t)(cp + (align - ioff));
1134:
1135: /*
1136: * Insert remaining chunks on the bucket list.
1137: */
1138: n = pp->pr_itemsperpage;
1.20 thorpej 1139: pp->pr_nitems += n;
1.3 pk 1140:
1141: while (n--) {
1142: pi = (struct pool_item *)cp;
1.78 thorpej 1143:
1144: KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
1.3 pk 1145:
1146: /* Insert on page list */
1147: TAILQ_INSERT_TAIL(&ph->ph_itemlist, pi, pi_list);
1148: #ifdef DIAGNOSTIC
1149: pi->pi_magic = PI_MAGIC;
1150: #endif
1151: cp = (caddr_t)(cp + pp->pr_size);
1152: }
1153:
1154: /*
1155: * If the pool was depleted, point at the new page.
1156: */
1157: if (pp->pr_curpage == NULL)
1158: pp->pr_curpage = ph;
1159:
1160: if (++pp->pr_npages > pp->pr_hiwat)
1161: pp->pr_hiwat = pp->pr_npages;
1162: }
1163:
1.20 thorpej 1164: /*
1.52 thorpej 1165: * Used by pool_get() when nitems drops below the low water mark. This
1.88 chs 1166: * is used to catch up pr_nitems with the low water mark.
1.20 thorpej 1167: *
1.21 thorpej 1168: * Note 1, we never wait for memory here, we let the caller decide what to do.
1.20 thorpej 1169: *
1.73 thorpej 1170: * Note 2, we must be called with the pool already locked, and we return
1.20 thorpej 1171: * with it locked.
1172: */
1173: static int
1.42 thorpej 1174: pool_catchup(struct pool *pp)
1.20 thorpej 1175: {
1.83 scw 1176: struct pool_item_header *ph = NULL;
1.20 thorpej 1177: caddr_t cp;
1178: int error = 0;
1179:
1.54 thorpej 1180: while (POOL_NEEDS_CATCHUP(pp)) {
1.20 thorpej 1181: /*
1.21 thorpej 1182: * Call the page back-end allocator for more memory.
1183: *
1184: * XXX: We never wait, so should we bother unlocking
1185: * the pool descriptor?
1.20 thorpej 1186: */
1.21 thorpej 1187: simple_unlock(&pp->pr_slock);
1.66 thorpej 1188: cp = pool_allocator_alloc(pp, PR_NOWAIT);
1.55 thorpej 1189: if (__predict_true(cp != NULL))
1190: ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1191: if (__predict_false(cp == NULL || ph == NULL)) {
1192: if (cp != NULL)
1.66 thorpej 1193: pool_allocator_free(pp, cp);
1.20 thorpej 1194: error = ENOMEM;
1.91 yamt 1195: simple_lock(&pp->pr_slock);
1.20 thorpej 1196: break;
1197: }
1.91 yamt 1198: simple_lock(&pp->pr_slock);
1.55 thorpej 1199: pool_prime_page(pp, cp, ph);
1.26 thorpej 1200: pp->pr_npagealloc++;
1.20 thorpej 1201: }
1202:
1203: return (error);
1204: }
1205:
1.88 chs 1206: static void
1207: pool_update_curpage(struct pool *pp)
1208: {
1209:
1210: pp->pr_curpage = LIST_FIRST(&pp->pr_partpages);
1211: if (pp->pr_curpage == NULL) {
1212: pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages);
1213: }
1214: }
1215:
1.3 pk 1216: void
1.42 thorpej 1217: pool_setlowat(struct pool *pp, int n)
1.3 pk 1218: {
1.15 pk 1219:
1.21 thorpej 1220: simple_lock(&pp->pr_slock);
1221:
1.3 pk 1222: pp->pr_minitems = n;
1.15 pk 1223: pp->pr_minpages = (n == 0)
1224: ? 0
1.18 thorpej 1225: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.20 thorpej 1226:
1227: /* Make sure we're caught up with the newly-set low water mark. */
1.75 simonb 1228: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1229: /*
1230: * XXX: Should we log a warning? Should we set up a timeout
1231: * to try again in a second or so? The latter could break
1232: * a caller's assumptions about interrupt protection, etc.
1233: */
1234: }
1.21 thorpej 1235:
1236: simple_unlock(&pp->pr_slock);
1.3 pk 1237: }
1238:
1239: void
1.42 thorpej 1240: pool_sethiwat(struct pool *pp, int n)
1.3 pk 1241: {
1.15 pk 1242:
1.21 thorpej 1243: simple_lock(&pp->pr_slock);
1244:
1.15 pk 1245: pp->pr_maxpages = (n == 0)
1246: ? 0
1.18 thorpej 1247: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1248:
1249: simple_unlock(&pp->pr_slock);
1.3 pk 1250: }
1251:
1.20 thorpej 1252: void
1.42 thorpej 1253: pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1.20 thorpej 1254: {
1255:
1.21 thorpej 1256: simple_lock(&pp->pr_slock);
1.20 thorpej 1257:
1258: pp->pr_hardlimit = n;
1259: pp->pr_hardlimit_warning = warnmess;
1.31 thorpej 1260: pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1261: pp->pr_hardlimit_warning_last.tv_sec = 0;
1262: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.20 thorpej 1263:
1264: /*
1.21 thorpej 1265: * In-line version of pool_sethiwat(), because we don't want to
1266: * release the lock.
1.20 thorpej 1267: */
1268: pp->pr_maxpages = (n == 0)
1269: ? 0
1270: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1271:
1272: simple_unlock(&pp->pr_slock);
1.20 thorpej 1273: }
1.3 pk 1274:
1275: /*
1276: * Release all complete pages that have not been used recently.
1277: */
1.66 thorpej 1278: int
1.59 thorpej 1279: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 1280: _pool_reclaim(struct pool *pp, const char *file, long line)
1.56 sommerfe 1281: #else
1282: pool_reclaim(struct pool *pp)
1283: #endif
1.3 pk 1284: {
1285: struct pool_item_header *ph, *phnext;
1.43 thorpej 1286: struct pool_cache *pc;
1.21 thorpej 1287: struct timeval curtime;
1.61 chs 1288: struct pool_pagelist pq;
1.88 chs 1289: struct timeval diff;
1.21 thorpej 1290: int s;
1.3 pk 1291:
1.68 thorpej 1292: if (pp->pr_drain_hook != NULL) {
1293: /*
1294: * The drain hook must be called with the pool unlocked.
1295: */
1296: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT);
1297: }
1298:
1.21 thorpej 1299: if (simple_lock_try(&pp->pr_slock) == 0)
1.66 thorpej 1300: return (0);
1.25 thorpej 1301: pr_enter(pp, file, line);
1.68 thorpej 1302:
1.88 chs 1303: LIST_INIT(&pq);
1.3 pk 1304:
1.43 thorpej 1305: /*
1306: * Reclaim items from the pool's caches.
1307: */
1.61 chs 1308: TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
1.43 thorpej 1309: pool_cache_reclaim(pc);
1310:
1.21 thorpej 1311: s = splclock();
1312: curtime = mono_time;
1313: splx(s);
1314:
1.88 chs 1315: for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
1316: phnext = LIST_NEXT(ph, ph_pagelist);
1.3 pk 1317:
1318: /* Check our minimum page claim */
1319: if (pp->pr_npages <= pp->pr_minpages)
1320: break;
1321:
1.88 chs 1322: KASSERT(ph->ph_nmissing == 0);
1323: timersub(&curtime, &ph->ph_time, &diff);
1324: if (diff.tv_sec < pool_inactive_time)
1325: continue;
1.21 thorpej 1326:
1.88 chs 1327: /*
1328: * If freeing this page would put us below
1329: * the low water mark, stop now.
1330: */
1331: if ((pp->pr_nitems - pp->pr_itemsperpage) <
1332: pp->pr_minitems)
1333: break;
1.21 thorpej 1334:
1.88 chs 1335: pr_rmpage(pp, ph, &pq);
1.3 pk 1336: }
1337:
1.25 thorpej 1338: pr_leave(pp);
1.21 thorpej 1339: simple_unlock(&pp->pr_slock);
1.88 chs 1340: if (LIST_EMPTY(&pq))
1.66 thorpej 1341: return (0);
1342:
1.88 chs 1343: while ((ph = LIST_FIRST(&pq)) != NULL) {
1344: LIST_REMOVE(ph, ph_pagelist);
1.66 thorpej 1345: pool_allocator_free(pp, ph->ph_page);
1.61 chs 1346: if (pp->pr_roflags & PR_PHINPAGE) {
1347: continue;
1348: }
1.85 pk 1349: s = splvm();
1.61 chs 1350: pool_put(&phpool, ph);
1351: splx(s);
1352: }
1.66 thorpej 1353:
1354: return (1);
1.3 pk 1355: }
1356:
1357: /*
1358: * Drain pools, one at a time.
1.21 thorpej 1359: *
1360: * Note, we must never be called from an interrupt context.
1.3 pk 1361: */
1362: void
1.42 thorpej 1363: pool_drain(void *arg)
1.3 pk 1364: {
1365: struct pool *pp;
1.23 thorpej 1366: int s;
1.3 pk 1367:
1.61 chs 1368: pp = NULL;
1.49 thorpej 1369: s = splvm();
1.23 thorpej 1370: simple_lock(&pool_head_slock);
1.61 chs 1371: if (drainpp == NULL) {
1372: drainpp = TAILQ_FIRST(&pool_head);
1373: }
1374: if (drainpp) {
1375: pp = drainpp;
1376: drainpp = TAILQ_NEXT(pp, pr_poollist);
1377: }
1378: simple_unlock(&pool_head_slock);
1.63 chs 1379: pool_reclaim(pp);
1.61 chs 1380: splx(s);
1.3 pk 1381: }
1382:
1383: /*
1384: * Diagnostic helpers.
1385: */
1386: void
1.42 thorpej 1387: pool_print(struct pool *pp, const char *modif)
1.21 thorpej 1388: {
1389: int s;
1390:
1.49 thorpej 1391: s = splvm();
1.25 thorpej 1392: if (simple_lock_try(&pp->pr_slock) == 0) {
1393: printf("pool %s is locked; try again later\n",
1394: pp->pr_wchan);
1395: splx(s);
1396: return;
1397: }
1398: pool_print1(pp, modif, printf);
1.21 thorpej 1399: simple_unlock(&pp->pr_slock);
1400: splx(s);
1401: }
1402:
1.25 thorpej 1403: void
1.42 thorpej 1404: pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.25 thorpej 1405: {
1406: int didlock = 0;
1407:
1408: if (pp == NULL) {
1409: (*pr)("Must specify a pool to print.\n");
1410: return;
1411: }
1412:
1413: /*
1414: * Called from DDB; interrupts should be blocked, and all
1415: * other processors should be paused. We can skip locking
1416: * the pool in this case.
1417: *
1418: * We do a simple_lock_try() just to print the lock
1419: * status, however.
1420: */
1421:
1422: if (simple_lock_try(&pp->pr_slock) == 0)
1423: (*pr)("WARNING: pool %s is locked\n", pp->pr_wchan);
1424: else
1425: didlock = 1;
1426:
1427: pool_print1(pp, modif, pr);
1428:
1429: if (didlock)
1430: simple_unlock(&pp->pr_slock);
1431: }
1432:
1.21 thorpej 1433: static void
1.88 chs 1434: pool_print_pagelist(struct pool_pagelist *pl, void (*pr)(const char *, ...))
1435: {
1436: struct pool_item_header *ph;
1437: #ifdef DIAGNOSTIC
1438: struct pool_item *pi;
1439: #endif
1440:
1441: LIST_FOREACH(ph, pl, ph_pagelist) {
1442: (*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n",
1443: ph->ph_page, ph->ph_nmissing,
1444: (u_long)ph->ph_time.tv_sec,
1445: (u_long)ph->ph_time.tv_usec);
1446: #ifdef DIAGNOSTIC
1447: TAILQ_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1448: if (pi->pi_magic != PI_MAGIC) {
1449: (*pr)("\t\t\titem %p, magic 0x%x\n",
1450: pi, pi->pi_magic);
1451: }
1452: }
1453: #endif
1454: }
1455: }
1456:
1457: static void
1.42 thorpej 1458: pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.3 pk 1459: {
1.25 thorpej 1460: struct pool_item_header *ph;
1.44 thorpej 1461: struct pool_cache *pc;
1462: struct pool_cache_group *pcg;
1463: int i, print_log = 0, print_pagelist = 0, print_cache = 0;
1.25 thorpej 1464: char c;
1465:
1466: while ((c = *modif++) != '\0') {
1467: if (c == 'l')
1468: print_log = 1;
1469: if (c == 'p')
1470: print_pagelist = 1;
1.44 thorpej 1471: if (c == 'c')
1472: print_cache = 1;
1.25 thorpej 1473: }
1474:
1475: (*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1476: pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1477: pp->pr_roflags);
1.66 thorpej 1478: (*pr)("\talloc %p\n", pp->pr_alloc);
1.25 thorpej 1479: (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1480: pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1481: (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1482: pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1483:
1484: (*pr)("\n\tnget %lu, nfail %lu, nput %lu\n",
1485: pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1486: (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1487: pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1488:
1489: if (print_pagelist == 0)
1490: goto skip_pagelist;
1491:
1.88 chs 1492: if ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1493: (*pr)("\n\tempty page list:\n");
1494: pool_print_pagelist(&pp->pr_emptypages, pr);
1495: if ((ph = LIST_FIRST(&pp->pr_fullpages)) != NULL)
1496: (*pr)("\n\tfull page list:\n");
1497: pool_print_pagelist(&pp->pr_fullpages, pr);
1498: if ((ph = LIST_FIRST(&pp->pr_partpages)) != NULL)
1499: (*pr)("\n\tpartial-page list:\n");
1500: pool_print_pagelist(&pp->pr_partpages, pr);
1501:
1.25 thorpej 1502: if (pp->pr_curpage == NULL)
1503: (*pr)("\tno current page\n");
1504: else
1505: (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1506:
1507: skip_pagelist:
1508: if (print_log == 0)
1509: goto skip_log;
1510:
1511: (*pr)("\n");
1512: if ((pp->pr_roflags & PR_LOGGING) == 0)
1513: (*pr)("\tno log\n");
1514: else
1515: pr_printlog(pp, NULL, pr);
1.3 pk 1516:
1.25 thorpej 1517: skip_log:
1.44 thorpej 1518: if (print_cache == 0)
1519: goto skip_cache;
1520:
1.61 chs 1521: TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist) {
1.44 thorpej 1522: (*pr)("\tcache %p: allocfrom %p freeto %p\n", pc,
1523: pc->pc_allocfrom, pc->pc_freeto);
1.48 thorpej 1524: (*pr)("\t hits %lu misses %lu ngroups %lu nitems %lu\n",
1525: pc->pc_hits, pc->pc_misses, pc->pc_ngroups, pc->pc_nitems);
1.61 chs 1526: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.44 thorpej 1527: (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail);
1.87 thorpej 1528: for (i = 0; i < PCG_NOBJECTS; i++) {
1529: if (pcg->pcg_objects[i].pcgo_pa !=
1530: POOL_PADDR_INVALID) {
1531: (*pr)("\t\t\t%p, 0x%llx\n",
1532: pcg->pcg_objects[i].pcgo_va,
1533: (unsigned long long)
1534: pcg->pcg_objects[i].pcgo_pa);
1535: } else {
1536: (*pr)("\t\t\t%p\n",
1537: pcg->pcg_objects[i].pcgo_va);
1538: }
1539: }
1.44 thorpej 1540: }
1541: }
1542:
1543: skip_cache:
1.88 chs 1544: pr_enter_check(pp, pr);
1545: }
1546:
1547: static int
1548: pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
1549: {
1550: struct pool_item *pi;
1551: caddr_t page;
1552: int n;
1553:
1554: page = (caddr_t)((u_long)ph & pp->pr_alloc->pa_pagemask);
1555: if (page != ph->ph_page &&
1556: (pp->pr_roflags & PR_PHINPAGE) != 0) {
1557: if (label != NULL)
1558: printf("%s: ", label);
1559: printf("pool(%p:%s): page inconsistency: page %p;"
1560: " at page head addr %p (p %p)\n", pp,
1561: pp->pr_wchan, ph->ph_page,
1562: ph, page);
1563: return 1;
1564: }
1.3 pk 1565:
1.88 chs 1566: for (pi = TAILQ_FIRST(&ph->ph_itemlist), n = 0;
1567: pi != NULL;
1568: pi = TAILQ_NEXT(pi,pi_list), n++) {
1569:
1570: #ifdef DIAGNOSTIC
1571: if (pi->pi_magic != PI_MAGIC) {
1572: if (label != NULL)
1573: printf("%s: ", label);
1574: printf("pool(%s): free list modified: magic=%x;"
1575: " page %p; item ordinal %d;"
1576: " addr %p (p %p)\n",
1577: pp->pr_wchan, pi->pi_magic, ph->ph_page,
1578: n, pi, page);
1579: panic("pool");
1580: }
1581: #endif
1582: page =
1583: (caddr_t)((u_long)pi & pp->pr_alloc->pa_pagemask);
1584: if (page == ph->ph_page)
1585: continue;
1586:
1587: if (label != NULL)
1588: printf("%s: ", label);
1589: printf("pool(%p:%s): page inconsistency: page %p;"
1590: " item ordinal %d; addr %p (p %p)\n", pp,
1591: pp->pr_wchan, ph->ph_page,
1592: n, pi, page);
1593: return 1;
1594: }
1595: return 0;
1.3 pk 1596: }
1597:
1.88 chs 1598:
1.3 pk 1599: int
1.42 thorpej 1600: pool_chk(struct pool *pp, const char *label)
1.3 pk 1601: {
1602: struct pool_item_header *ph;
1603: int r = 0;
1604:
1.21 thorpej 1605: simple_lock(&pp->pr_slock);
1.88 chs 1606: LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
1607: r = pool_chk_page(pp, label, ph);
1608: if (r) {
1609: goto out;
1610: }
1611: }
1612: LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
1613: r = pool_chk_page(pp, label, ph);
1614: if (r) {
1.3 pk 1615: goto out;
1616: }
1.88 chs 1617: }
1618: LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
1619: r = pool_chk_page(pp, label, ph);
1620: if (r) {
1.3 pk 1621: goto out;
1622: }
1623: }
1.88 chs 1624:
1.3 pk 1625: out:
1.21 thorpej 1626: simple_unlock(&pp->pr_slock);
1.3 pk 1627: return (r);
1.43 thorpej 1628: }
1629:
1630: /*
1631: * pool_cache_init:
1632: *
1633: * Initialize a pool cache.
1634: *
1635: * NOTE: If the pool must be protected from interrupts, we expect
1636: * to be called at the appropriate interrupt priority level.
1637: */
1638: void
1639: pool_cache_init(struct pool_cache *pc, struct pool *pp,
1640: int (*ctor)(void *, void *, int),
1641: void (*dtor)(void *, void *),
1642: void *arg)
1643: {
1644:
1645: TAILQ_INIT(&pc->pc_grouplist);
1646: simple_lock_init(&pc->pc_slock);
1647:
1648: pc->pc_allocfrom = NULL;
1649: pc->pc_freeto = NULL;
1650: pc->pc_pool = pp;
1651:
1652: pc->pc_ctor = ctor;
1653: pc->pc_dtor = dtor;
1654: pc->pc_arg = arg;
1655:
1.48 thorpej 1656: pc->pc_hits = 0;
1657: pc->pc_misses = 0;
1658:
1659: pc->pc_ngroups = 0;
1660:
1661: pc->pc_nitems = 0;
1662:
1.43 thorpej 1663: simple_lock(&pp->pr_slock);
1664: TAILQ_INSERT_TAIL(&pp->pr_cachelist, pc, pc_poollist);
1665: simple_unlock(&pp->pr_slock);
1666: }
1667:
1668: /*
1669: * pool_cache_destroy:
1670: *
1671: * Destroy a pool cache.
1672: */
1673: void
1674: pool_cache_destroy(struct pool_cache *pc)
1675: {
1676: struct pool *pp = pc->pc_pool;
1677:
1678: /* First, invalidate the entire cache. */
1679: pool_cache_invalidate(pc);
1680:
1681: /* ...and remove it from the pool's cache list. */
1682: simple_lock(&pp->pr_slock);
1683: TAILQ_REMOVE(&pp->pr_cachelist, pc, pc_poollist);
1684: simple_unlock(&pp->pr_slock);
1685: }
1686:
1687: static __inline void *
1.87 thorpej 1688: pcg_get(struct pool_cache_group *pcg, paddr_t *pap)
1.43 thorpej 1689: {
1690: void *object;
1691: u_int idx;
1692:
1693: KASSERT(pcg->pcg_avail <= PCG_NOBJECTS);
1.45 thorpej 1694: KASSERT(pcg->pcg_avail != 0);
1.43 thorpej 1695: idx = --pcg->pcg_avail;
1696:
1.87 thorpej 1697: KASSERT(pcg->pcg_objects[idx].pcgo_va != NULL);
1698: object = pcg->pcg_objects[idx].pcgo_va;
1699: if (pap != NULL)
1700: *pap = pcg->pcg_objects[idx].pcgo_pa;
1701: pcg->pcg_objects[idx].pcgo_va = NULL;
1.43 thorpej 1702:
1703: return (object);
1704: }
1705:
1706: static __inline void
1.87 thorpej 1707: pcg_put(struct pool_cache_group *pcg, void *object, paddr_t pa)
1.43 thorpej 1708: {
1709: u_int idx;
1710:
1711: KASSERT(pcg->pcg_avail < PCG_NOBJECTS);
1712: idx = pcg->pcg_avail++;
1713:
1.87 thorpej 1714: KASSERT(pcg->pcg_objects[idx].pcgo_va == NULL);
1715: pcg->pcg_objects[idx].pcgo_va = object;
1716: pcg->pcg_objects[idx].pcgo_pa = pa;
1.43 thorpej 1717: }
1718:
1719: /*
1.87 thorpej 1720: * pool_cache_get{,_paddr}:
1.43 thorpej 1721: *
1.87 thorpej 1722: * Get an object from a pool cache (optionally returning
1723: * the physical address of the object).
1.43 thorpej 1724: */
1725: void *
1.87 thorpej 1726: pool_cache_get_paddr(struct pool_cache *pc, int flags, paddr_t *pap)
1.43 thorpej 1727: {
1728: struct pool_cache_group *pcg;
1729: void *object;
1.58 thorpej 1730:
1731: #ifdef LOCKDEBUG
1732: if (flags & PR_WAITOK)
1733: simple_lock_only_held(NULL, "pool_cache_get(PR_WAITOK)");
1734: #endif
1.43 thorpej 1735:
1736: simple_lock(&pc->pc_slock);
1737:
1738: if ((pcg = pc->pc_allocfrom) == NULL) {
1.61 chs 1739: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.43 thorpej 1740: if (pcg->pcg_avail != 0) {
1741: pc->pc_allocfrom = pcg;
1742: goto have_group;
1743: }
1744: }
1745:
1746: /*
1747: * No groups with any available objects. Allocate
1748: * a new object, construct it, and return it to
1749: * the caller. We will allocate a group, if necessary,
1750: * when the object is freed back to the cache.
1751: */
1.48 thorpej 1752: pc->pc_misses++;
1.43 thorpej 1753: simple_unlock(&pc->pc_slock);
1754: object = pool_get(pc->pc_pool, flags);
1755: if (object != NULL && pc->pc_ctor != NULL) {
1756: if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) {
1757: pool_put(pc->pc_pool, object);
1758: return (NULL);
1759: }
1760: }
1.87 thorpej 1761: if (object != NULL && pap != NULL) {
1762: #ifdef POOL_VTOPHYS
1763: *pap = POOL_VTOPHYS(object);
1764: #else
1765: *pap = POOL_PADDR_INVALID;
1766: #endif
1767: }
1.43 thorpej 1768: return (object);
1769: }
1770:
1771: have_group:
1.48 thorpej 1772: pc->pc_hits++;
1773: pc->pc_nitems--;
1.87 thorpej 1774: object = pcg_get(pcg, pap);
1.43 thorpej 1775:
1776: if (pcg->pcg_avail == 0)
1777: pc->pc_allocfrom = NULL;
1.45 thorpej 1778:
1.43 thorpej 1779: simple_unlock(&pc->pc_slock);
1780:
1781: return (object);
1782: }
1783:
1784: /*
1.87 thorpej 1785: * pool_cache_put{,_paddr}:
1.43 thorpej 1786: *
1.87 thorpej 1787: * Put an object back to the pool cache (optionally caching the
1788: * physical address of the object).
1.43 thorpej 1789: */
1790: void
1.87 thorpej 1791: pool_cache_put_paddr(struct pool_cache *pc, void *object, paddr_t pa)
1.43 thorpej 1792: {
1793: struct pool_cache_group *pcg;
1.60 thorpej 1794: int s;
1.43 thorpej 1795:
1796: simple_lock(&pc->pc_slock);
1797:
1798: if ((pcg = pc->pc_freeto) == NULL) {
1.61 chs 1799: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.43 thorpej 1800: if (pcg->pcg_avail != PCG_NOBJECTS) {
1801: pc->pc_freeto = pcg;
1802: goto have_group;
1803: }
1804: }
1805:
1806: /*
1807: * No empty groups to free the object to. Attempt to
1.47 thorpej 1808: * allocate one.
1.43 thorpej 1809: */
1.47 thorpej 1810: simple_unlock(&pc->pc_slock);
1.60 thorpej 1811: s = splvm();
1.43 thorpej 1812: pcg = pool_get(&pcgpool, PR_NOWAIT);
1.60 thorpej 1813: splx(s);
1.43 thorpej 1814: if (pcg != NULL) {
1815: memset(pcg, 0, sizeof(*pcg));
1.47 thorpej 1816: simple_lock(&pc->pc_slock);
1.48 thorpej 1817: pc->pc_ngroups++;
1.43 thorpej 1818: TAILQ_INSERT_TAIL(&pc->pc_grouplist, pcg, pcg_list);
1.47 thorpej 1819: if (pc->pc_freeto == NULL)
1820: pc->pc_freeto = pcg;
1.43 thorpej 1821: goto have_group;
1822: }
1823:
1824: /*
1825: * Unable to allocate a cache group; destruct the object
1826: * and free it back to the pool.
1827: */
1.51 thorpej 1828: pool_cache_destruct_object(pc, object);
1.43 thorpej 1829: return;
1830: }
1831:
1832: have_group:
1.48 thorpej 1833: pc->pc_nitems++;
1.87 thorpej 1834: pcg_put(pcg, object, pa);
1.43 thorpej 1835:
1836: if (pcg->pcg_avail == PCG_NOBJECTS)
1837: pc->pc_freeto = NULL;
1838:
1839: simple_unlock(&pc->pc_slock);
1.51 thorpej 1840: }
1841:
1842: /*
1843: * pool_cache_destruct_object:
1844: *
1845: * Force destruction of an object and its release back into
1846: * the pool.
1847: */
1848: void
1849: pool_cache_destruct_object(struct pool_cache *pc, void *object)
1850: {
1851:
1852: if (pc->pc_dtor != NULL)
1853: (*pc->pc_dtor)(pc->pc_arg, object);
1854: pool_put(pc->pc_pool, object);
1.43 thorpej 1855: }
1856:
1857: /*
1858: * pool_cache_do_invalidate:
1859: *
1860: * This internal function implements pool_cache_invalidate() and
1861: * pool_cache_reclaim().
1862: */
1863: static void
1864: pool_cache_do_invalidate(struct pool_cache *pc, int free_groups,
1.56 sommerfe 1865: void (*putit)(struct pool *, void *))
1.43 thorpej 1866: {
1867: struct pool_cache_group *pcg, *npcg;
1868: void *object;
1.60 thorpej 1869: int s;
1.43 thorpej 1870:
1871: for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
1872: pcg = npcg) {
1873: npcg = TAILQ_NEXT(pcg, pcg_list);
1874: while (pcg->pcg_avail != 0) {
1.48 thorpej 1875: pc->pc_nitems--;
1.87 thorpej 1876: object = pcg_get(pcg, NULL);
1.45 thorpej 1877: if (pcg->pcg_avail == 0 && pc->pc_allocfrom == pcg)
1878: pc->pc_allocfrom = NULL;
1.43 thorpej 1879: if (pc->pc_dtor != NULL)
1880: (*pc->pc_dtor)(pc->pc_arg, object);
1.56 sommerfe 1881: (*putit)(pc->pc_pool, object);
1.43 thorpej 1882: }
1883: if (free_groups) {
1.48 thorpej 1884: pc->pc_ngroups--;
1.43 thorpej 1885: TAILQ_REMOVE(&pc->pc_grouplist, pcg, pcg_list);
1.46 thorpej 1886: if (pc->pc_freeto == pcg)
1887: pc->pc_freeto = NULL;
1.60 thorpej 1888: s = splvm();
1.43 thorpej 1889: pool_put(&pcgpool, pcg);
1.60 thorpej 1890: splx(s);
1.43 thorpej 1891: }
1892: }
1893: }
1894:
1895: /*
1896: * pool_cache_invalidate:
1897: *
1898: * Invalidate a pool cache (destruct and release all of the
1899: * cached objects).
1900: */
1901: void
1902: pool_cache_invalidate(struct pool_cache *pc)
1903: {
1904:
1905: simple_lock(&pc->pc_slock);
1.56 sommerfe 1906: pool_cache_do_invalidate(pc, 0, pool_put);
1.43 thorpej 1907: simple_unlock(&pc->pc_slock);
1908: }
1909:
1910: /*
1911: * pool_cache_reclaim:
1912: *
1913: * Reclaim a pool cache for pool_reclaim().
1914: */
1915: static void
1916: pool_cache_reclaim(struct pool_cache *pc)
1917: {
1918:
1.47 thorpej 1919: simple_lock(&pc->pc_slock);
1.43 thorpej 1920: pool_cache_do_invalidate(pc, 1, pool_do_put);
1921: simple_unlock(&pc->pc_slock);
1.3 pk 1922: }
1.66 thorpej 1923:
1924: /*
1925: * Pool backend allocators.
1926: *
1927: * Each pool has a backend allocator that handles allocation, deallocation,
1928: * and any additional draining that might be needed.
1929: *
1930: * We provide two standard allocators:
1931: *
1932: * pool_allocator_kmem - the default when no allocator is specified
1933: *
1934: * pool_allocator_nointr - used for pools that will not be accessed
1935: * in interrupt context.
1936: */
1937: void *pool_page_alloc(struct pool *, int);
1938: void pool_page_free(struct pool *, void *);
1939:
1940: struct pool_allocator pool_allocator_kmem = {
1941: pool_page_alloc, pool_page_free, 0,
1942: };
1943:
1944: void *pool_page_alloc_nointr(struct pool *, int);
1945: void pool_page_free_nointr(struct pool *, void *);
1946:
1947: struct pool_allocator pool_allocator_nointr = {
1948: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1949: };
1950:
1951: #ifdef POOL_SUBPAGE
1952: void *pool_subpage_alloc(struct pool *, int);
1953: void pool_subpage_free(struct pool *, void *);
1954:
1955: struct pool_allocator pool_allocator_kmem_subpage = {
1956: pool_subpage_alloc, pool_subpage_free, 0,
1957: };
1958: #endif /* POOL_SUBPAGE */
1959:
1960: /*
1961: * We have at least three different resources for the same allocation and
1962: * each resource can be depleted. First, we have the ready elements in the
1963: * pool. Then we have the resource (typically a vm_map) for this allocator.
1964: * Finally, we have physical memory. Waiting for any of these can be
1965: * unnecessary when any other is freed, but the kernel doesn't support
1966: * sleeping on multiple wait channels, so we have to employ another strategy.
1967: *
1968: * The caller sleeps on the pool (so that it can be awakened when an item
1969: * is returned to the pool), but we set PA_WANT on the allocator. When a
1970: * page is returned to the allocator and PA_WANT is set, pool_allocator_free
1971: * will wake up all sleeping pools belonging to this allocator.
1972: *
1973: * XXX Thundering herd.
1974: */
1975: void *
1976: pool_allocator_alloc(struct pool *org, int flags)
1977: {
1978: struct pool_allocator *pa = org->pr_alloc;
1979: struct pool *pp, *start;
1980: int s, freed;
1981: void *res;
1982:
1.91 yamt 1983: LOCK_ASSERT(!simple_lock_held(&org->pr_slock));
1984:
1.66 thorpej 1985: do {
1986: if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
1987: return (res);
1.68 thorpej 1988: if ((flags & PR_WAITOK) == 0) {
1989: /*
1990: * We only run the drain hookhere if PR_NOWAIT.
1991: * In other cases, the hook will be run in
1992: * pool_reclaim().
1993: */
1994: if (org->pr_drain_hook != NULL) {
1995: (*org->pr_drain_hook)(org->pr_drain_hook_arg,
1996: flags);
1997: if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
1998: return (res);
1999: }
1.66 thorpej 2000: break;
1.68 thorpej 2001: }
1.66 thorpej 2002:
2003: /*
2004: * Drain all pools, except "org", that use this
2005: * allocator. We do this to reclaim VA space.
2006: * pa_alloc is responsible for waiting for
2007: * physical memory.
2008: *
2009: * XXX We risk looping forever if start if someone
2010: * calls pool_destroy on "start". But there is no
2011: * other way to have potentially sleeping pool_reclaim,
2012: * non-sleeping locks on pool_allocator, and some
2013: * stirring of drained pools in the allocator.
1.68 thorpej 2014: *
2015: * XXX Maybe we should use pool_head_slock for locking
2016: * the allocators?
1.66 thorpej 2017: */
2018: freed = 0;
2019:
2020: s = splvm();
2021: simple_lock(&pa->pa_slock);
2022: pp = start = TAILQ_FIRST(&pa->pa_list);
2023: do {
2024: TAILQ_REMOVE(&pa->pa_list, pp, pr_alloc_list);
2025: TAILQ_INSERT_TAIL(&pa->pa_list, pp, pr_alloc_list);
2026: if (pp == org)
2027: continue;
1.73 thorpej 2028: simple_unlock(&pa->pa_slock);
1.66 thorpej 2029: freed = pool_reclaim(pp);
1.73 thorpej 2030: simple_lock(&pa->pa_slock);
1.66 thorpej 2031: } while ((pp = TAILQ_FIRST(&pa->pa_list)) != start &&
2032: freed == 0);
2033:
2034: if (freed == 0) {
2035: /*
2036: * We set PA_WANT here, the caller will most likely
2037: * sleep waiting for pages (if not, this won't hurt
2038: * that much), and there is no way to set this in
2039: * the caller without violating locking order.
2040: */
2041: pa->pa_flags |= PA_WANT;
2042: }
2043: simple_unlock(&pa->pa_slock);
2044: splx(s);
2045: } while (freed);
2046: return (NULL);
2047: }
2048:
2049: void
2050: pool_allocator_free(struct pool *pp, void *v)
2051: {
2052: struct pool_allocator *pa = pp->pr_alloc;
2053: int s;
2054:
1.91 yamt 2055: LOCK_ASSERT(!simple_lock_held(&pp->pr_slock));
2056:
1.66 thorpej 2057: (*pa->pa_free)(pp, v);
2058:
2059: s = splvm();
2060: simple_lock(&pa->pa_slock);
2061: if ((pa->pa_flags & PA_WANT) == 0) {
2062: simple_unlock(&pa->pa_slock);
2063: splx(s);
2064: return;
2065: }
2066:
2067: TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
2068: simple_lock(&pp->pr_slock);
2069: if ((pp->pr_flags & PR_WANTED) != 0) {
2070: pp->pr_flags &= ~PR_WANTED;
2071: wakeup(pp);
2072: }
1.69 thorpej 2073: simple_unlock(&pp->pr_slock);
1.66 thorpej 2074: }
2075: pa->pa_flags &= ~PA_WANT;
2076: simple_unlock(&pa->pa_slock);
2077: splx(s);
2078: }
2079:
2080: void *
2081: pool_page_alloc(struct pool *pp, int flags)
2082: {
2083: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2084:
2085: return ((void *) uvm_km_alloc_poolpage(waitok));
2086: }
2087:
2088: void
2089: pool_page_free(struct pool *pp, void *v)
2090: {
2091:
2092: uvm_km_free_poolpage((vaddr_t) v);
2093: }
2094:
2095: #ifdef POOL_SUBPAGE
2096: /* Sub-page allocator, for machines with large hardware pages. */
2097: void *
2098: pool_subpage_alloc(struct pool *pp, int flags)
2099: {
1.93 ! dbj 2100: void *v;
! 2101: int s;
! 2102: s = splvm();
! 2103: v = pool_get(&psppool, flags);
! 2104: splx(s);
! 2105: return v;
1.66 thorpej 2106: }
2107:
2108: void
2109: pool_subpage_free(struct pool *pp, void *v)
2110: {
1.93 ! dbj 2111: int s;
! 2112: s = splvm();
1.66 thorpej 2113: pool_put(&psppool, v);
1.93 ! dbj 2114: splx(s);
1.66 thorpej 2115: }
2116:
2117: /* We don't provide a real nointr allocator. Maybe later. */
2118: void *
2119: pool_page_alloc_nointr(struct pool *pp, int flags)
2120: {
2121:
2122: return (pool_subpage_alloc(pp, flags));
2123: }
2124:
2125: void
2126: pool_page_free_nointr(struct pool *pp, void *v)
2127: {
2128:
2129: pool_subpage_free(pp, v);
2130: }
2131: #else
2132: void *
2133: pool_page_alloc_nointr(struct pool *pp, int flags)
2134: {
2135: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2136:
2137: return ((void *) uvm_km_alloc_poolpage1(kernel_map,
2138: uvm.kernel_object, waitok));
2139: }
2140:
2141: void
2142: pool_page_free_nointr(struct pool *pp, void *v)
2143: {
2144:
2145: uvm_km_free_poolpage1(kernel_map, (vaddr_t) v);
2146: }
2147: #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