Annotation of src/sys/kern/subr_pool.c, Revision 1.67
1.67 ! thorpej 1: /* $NetBSD: subr_pool.c,v 1.66 2002/03/08 20:48:41 thorpej 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>
1.67 ! thorpej 41: __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.66 2002/03/08 20:48:41 thorpej 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: *
62: * Memory is allocated in pages which are split into pieces according
63: * to the pool item size. Each page is kept on a list headed by `pr_pagelist'
64: * in the pool structure and the individual pool items are on a linked list
65: * headed by `ph_itemlist' in each page header. The memory for building
66: * the page list is either taken from the allocated pages themselves (for
67: * small pool items) or taken from an internal pool of page headers (`phpool').
1.1 pk 68: */
69:
1.3 pk 70: /* List of all pools */
1.5 thorpej 71: TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head);
1.3 pk 72:
73: /* Private pool for page header structures */
74: static struct pool phpool;
75:
1.62 bjh21 76: #ifdef POOL_SUBPAGE
77: /* Pool of subpages for use by normal pools. */
78: static struct pool psppool;
79: #endif
80:
1.3 pk 81: /* # of seconds to retain page after last use */
82: int pool_inactive_time = 10;
83:
84: /* Next candidate for drainage (see pool_drain()) */
1.23 thorpej 85: static struct pool *drainpp;
86:
87: /* This spin lock protects both pool_head and drainpp. */
88: struct simplelock pool_head_slock = SIMPLELOCK_INITIALIZER;
1.3 pk 89:
90: struct pool_item_header {
91: /* Page headers */
92: TAILQ_ENTRY(pool_item_header)
93: ph_pagelist; /* pool page list */
94: TAILQ_HEAD(,pool_item) ph_itemlist; /* chunk list for this page */
95: LIST_ENTRY(pool_item_header)
96: ph_hashlist; /* Off-page page headers */
97: int ph_nmissing; /* # of chunks in use */
98: caddr_t ph_page; /* this page's address */
99: struct timeval ph_time; /* last referenced */
100: };
1.61 chs 101: TAILQ_HEAD(pool_pagelist,pool_item_header);
1.3 pk 102:
1.1 pk 103: struct pool_item {
1.3 pk 104: #ifdef DIAGNOSTIC
105: int pi_magic;
1.33 chs 106: #endif
1.25 thorpej 107: #define PI_MAGIC 0xdeadbeef
1.3 pk 108: /* Other entries use only this list entry */
109: TAILQ_ENTRY(pool_item) pi_list;
110: };
111:
1.25 thorpej 112: #define PR_HASH_INDEX(pp,addr) \
1.66 thorpej 113: (((u_long)(addr) >> (pp)->pr_alloc->pa_pageshift) & \
114: (PR_HASHTABSIZE - 1))
1.3 pk 115:
1.53 thorpej 116: #define POOL_NEEDS_CATCHUP(pp) \
117: ((pp)->pr_nitems < (pp)->pr_minitems)
118:
1.43 thorpej 119: /*
120: * Pool cache management.
121: *
122: * Pool caches provide a way for constructed objects to be cached by the
123: * pool subsystem. This can lead to performance improvements by avoiding
124: * needless object construction/destruction; it is deferred until absolutely
125: * necessary.
126: *
127: * Caches are grouped into cache groups. Each cache group references
128: * up to 16 constructed objects. When a cache allocates an object
129: * from the pool, it calls the object's constructor and places it into
130: * a cache group. When a cache group frees an object back to the pool,
131: * it first calls the object's destructor. This allows the object to
132: * persist in constructed form while freed to the cache.
133: *
134: * Multiple caches may exist for each pool. This allows a single
135: * object type to have multiple constructed forms. The pool references
136: * each cache, so that when a pool is drained by the pagedaemon, it can
137: * drain each individual cache as well. Each time a cache is drained,
138: * the most idle cache group is freed to the pool in its entirety.
139: *
140: * Pool caches are layed on top of pools. By layering them, we can avoid
141: * the complexity of cache management for pools which would not benefit
142: * from it.
143: */
144:
145: /* The cache group pool. */
146: static struct pool pcgpool;
147:
148: /* The pool cache group. */
149: #define PCG_NOBJECTS 16
150: struct pool_cache_group {
151: TAILQ_ENTRY(pool_cache_group)
152: pcg_list; /* link in the pool cache's group list */
153: u_int pcg_avail; /* # available objects */
154: /* pointers to the objects */
155: void *pcg_objects[PCG_NOBJECTS];
156: };
1.3 pk 157:
1.43 thorpej 158: static void pool_cache_reclaim(struct pool_cache *);
1.3 pk 159:
1.42 thorpej 160: static int pool_catchup(struct pool *);
1.55 thorpej 161: static void pool_prime_page(struct pool *, caddr_t,
162: struct pool_item_header *);
1.66 thorpej 163:
164: void *pool_allocator_alloc(struct pool *, int);
165: void pool_allocator_free(struct pool *, void *);
1.3 pk 166:
1.42 thorpej 167: static void pool_print1(struct pool *, const char *,
168: void (*)(const char *, ...));
1.3 pk 169:
170: /*
1.52 thorpej 171: * Pool log entry. An array of these is allocated in pool_init().
1.3 pk 172: */
173: struct pool_log {
174: const char *pl_file;
175: long pl_line;
176: int pl_action;
1.25 thorpej 177: #define PRLOG_GET 1
178: #define PRLOG_PUT 2
1.3 pk 179: void *pl_addr;
1.1 pk 180: };
181:
1.3 pk 182: /* Number of entries in pool log buffers */
1.17 thorpej 183: #ifndef POOL_LOGSIZE
184: #define POOL_LOGSIZE 10
185: #endif
186:
187: int pool_logsize = POOL_LOGSIZE;
1.1 pk 188:
1.59 thorpej 189: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 190: static __inline void
191: pr_log(struct pool *pp, void *v, int action, const char *file, long line)
1.3 pk 192: {
193: int n = pp->pr_curlogentry;
194: struct pool_log *pl;
195:
1.20 thorpej 196: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 197: return;
198:
199: /*
200: * Fill in the current entry. Wrap around and overwrite
201: * the oldest entry if necessary.
202: */
203: pl = &pp->pr_log[n];
204: pl->pl_file = file;
205: pl->pl_line = line;
206: pl->pl_action = action;
207: pl->pl_addr = v;
208: if (++n >= pp->pr_logsize)
209: n = 0;
210: pp->pr_curlogentry = n;
211: }
212:
213: static void
1.42 thorpej 214: pr_printlog(struct pool *pp, struct pool_item *pi,
215: void (*pr)(const char *, ...))
1.3 pk 216: {
217: int i = pp->pr_logsize;
218: int n = pp->pr_curlogentry;
219:
1.20 thorpej 220: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 221: return;
222:
223: /*
224: * Print all entries in this pool's log.
225: */
226: while (i-- > 0) {
227: struct pool_log *pl = &pp->pr_log[n];
228: if (pl->pl_action != 0) {
1.25 thorpej 229: if (pi == NULL || pi == pl->pl_addr) {
230: (*pr)("\tlog entry %d:\n", i);
231: (*pr)("\t\taction = %s, addr = %p\n",
232: pl->pl_action == PRLOG_GET ? "get" : "put",
233: pl->pl_addr);
234: (*pr)("\t\tfile: %s at line %lu\n",
235: pl->pl_file, pl->pl_line);
236: }
1.3 pk 237: }
238: if (++n >= pp->pr_logsize)
239: n = 0;
240: }
241: }
1.25 thorpej 242:
1.42 thorpej 243: static __inline void
244: pr_enter(struct pool *pp, const char *file, long line)
1.25 thorpej 245: {
246:
1.34 thorpej 247: if (__predict_false(pp->pr_entered_file != NULL)) {
1.25 thorpej 248: printf("pool %s: reentrancy at file %s line %ld\n",
249: pp->pr_wchan, file, line);
250: printf(" previous entry at file %s line %ld\n",
251: pp->pr_entered_file, pp->pr_entered_line);
252: panic("pr_enter");
253: }
254:
255: pp->pr_entered_file = file;
256: pp->pr_entered_line = line;
257: }
258:
1.42 thorpej 259: static __inline void
260: pr_leave(struct pool *pp)
1.25 thorpej 261: {
262:
1.34 thorpej 263: if (__predict_false(pp->pr_entered_file == NULL)) {
1.25 thorpej 264: printf("pool %s not entered?\n", pp->pr_wchan);
265: panic("pr_leave");
266: }
267:
268: pp->pr_entered_file = NULL;
269: pp->pr_entered_line = 0;
270: }
271:
1.42 thorpej 272: static __inline void
273: pr_enter_check(struct pool *pp, void (*pr)(const char *, ...))
1.25 thorpej 274: {
275:
276: if (pp->pr_entered_file != NULL)
277: (*pr)("\n\tcurrently entered from file %s line %ld\n",
278: pp->pr_entered_file, pp->pr_entered_line);
279: }
1.3 pk 280: #else
1.25 thorpej 281: #define pr_log(pp, v, action, file, line)
282: #define pr_printlog(pp, pi, pr)
283: #define pr_enter(pp, file, line)
284: #define pr_leave(pp)
285: #define pr_enter_check(pp, pr)
1.59 thorpej 286: #endif /* POOL_DIAGNOSTIC */
1.3 pk 287:
288: /*
289: * Return the pool page header based on page address.
290: */
1.42 thorpej 291: static __inline struct pool_item_header *
292: pr_find_pagehead(struct pool *pp, caddr_t page)
1.3 pk 293: {
294: struct pool_item_header *ph;
295:
1.20 thorpej 296: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
1.3 pk 297: return ((struct pool_item_header *)(page + pp->pr_phoffset));
298:
299: for (ph = LIST_FIRST(&pp->pr_hashtab[PR_HASH_INDEX(pp, page)]);
300: ph != NULL;
301: ph = LIST_NEXT(ph, ph_hashlist)) {
302: if (ph->ph_page == page)
303: return (ph);
304: }
305: return (NULL);
306: }
307:
308: /*
309: * Remove a page from the pool.
310: */
1.42 thorpej 311: static __inline void
1.61 chs 312: pr_rmpage(struct pool *pp, struct pool_item_header *ph,
313: struct pool_pagelist *pq)
1.3 pk 314: {
1.61 chs 315: int s;
1.3 pk 316:
317: /*
1.7 thorpej 318: * If the page was idle, decrement the idle page count.
1.3 pk 319: */
1.6 thorpej 320: if (ph->ph_nmissing == 0) {
321: #ifdef DIAGNOSTIC
322: if (pp->pr_nidle == 0)
323: panic("pr_rmpage: nidle inconsistent");
1.20 thorpej 324: if (pp->pr_nitems < pp->pr_itemsperpage)
325: panic("pr_rmpage: nitems inconsistent");
1.6 thorpej 326: #endif
327: pp->pr_nidle--;
328: }
1.7 thorpej 329:
1.20 thorpej 330: pp->pr_nitems -= pp->pr_itemsperpage;
331:
1.7 thorpej 332: /*
1.61 chs 333: * Unlink a page from the pool and release it (or queue it for release).
1.7 thorpej 334: */
335: TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
1.61 chs 336: if (pq) {
337: TAILQ_INSERT_HEAD(pq, ph, ph_pagelist);
338: } else {
1.66 thorpej 339: pool_allocator_free(pp, ph->ph_page);
1.61 chs 340: if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
341: LIST_REMOVE(ph, ph_hashlist);
342: s = splhigh();
343: pool_put(&phpool, ph);
344: splx(s);
345: }
346: }
1.7 thorpej 347: pp->pr_npages--;
348: pp->pr_npagefree++;
1.6 thorpej 349:
1.3 pk 350: if (pp->pr_curpage == ph) {
351: /*
352: * Find a new non-empty page header, if any.
353: * Start search from the page head, to increase the
354: * chance for "high water" pages to be freed.
355: */
1.61 chs 356: TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist)
1.3 pk 357: if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
358: break;
359:
360: pp->pr_curpage = ph;
1.21 thorpej 361: }
1.3 pk 362: }
363:
364: /*
365: * Initialize the given pool resource structure.
366: *
367: * We export this routine to allow other kernel parts to declare
368: * static pools that must be initialized before malloc() is available.
369: */
370: void
1.42 thorpej 371: pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
1.66 thorpej 372: const char *wchan, struct pool_allocator *palloc)
1.3 pk 373: {
1.16 briggs 374: int off, slack, i;
1.3 pk 375:
1.25 thorpej 376: #ifdef POOL_DIAGNOSTIC
377: /*
378: * Always log if POOL_DIAGNOSTIC is defined.
379: */
380: if (pool_logsize != 0)
381: flags |= PR_LOGGING;
382: #endif
383:
1.66 thorpej 384: #ifdef POOL_SUBPAGE
385: /*
386: * XXX We don't provide a real `nointr' back-end
387: * yet; all sub-pages come from a kmem back-end.
388: * maybe some day...
389: */
390: if (palloc == NULL) {
391: extern struct pool_allocator pool_allocator_kmem_subpage;
392: palloc = &pool_allocator_kmem_subpage;
393: }
1.3 pk 394: /*
1.66 thorpej 395: * We'll assume any user-specified back-end allocator
396: * will deal with sub-pages, or simply don't care.
1.3 pk 397: */
1.66 thorpej 398: #else
399: if (palloc == NULL)
400: palloc = &pool_allocator_kmem;
401: #endif /* POOL_SUBPAGE */
402: if ((palloc->pa_flags & PA_INITIALIZED) == 0) {
403: if (palloc->pa_pagesz == 0) {
1.62 bjh21 404: #ifdef POOL_SUBPAGE
1.66 thorpej 405: if (palloc == &pool_allocator_kmem)
406: palloc->pa_pagesz = PAGE_SIZE;
407: else
408: palloc->pa_pagesz = POOL_SUBPAGE;
1.62 bjh21 409: #else
1.66 thorpej 410: palloc->pa_pagesz = PAGE_SIZE;
411: #endif /* POOL_SUBPAGE */
412: }
413:
414: TAILQ_INIT(&palloc->pa_list);
415:
416: simple_lock_init(&palloc->pa_slock);
417: palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
418: palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
419: palloc->pa_flags |= PA_INITIALIZED;
1.4 thorpej 420: }
1.3 pk 421:
422: if (align == 0)
423: align = ALIGN(1);
1.14 thorpej 424:
425: if (size < sizeof(struct pool_item))
426: size = sizeof(struct pool_item);
1.3 pk 427:
1.35 pk 428: size = ALIGN(size);
1.66 thorpej 429: #ifdef DIAGNOSTIC
430: if (size > palloc->pa_pagesz)
1.35 pk 431: panic("pool_init: pool item size (%lu) too large",
432: (u_long)size);
1.66 thorpej 433: #endif
1.35 pk 434:
1.3 pk 435: /*
436: * Initialize the pool structure.
437: */
438: TAILQ_INIT(&pp->pr_pagelist);
1.43 thorpej 439: TAILQ_INIT(&pp->pr_cachelist);
1.3 pk 440: pp->pr_curpage = NULL;
441: pp->pr_npages = 0;
442: pp->pr_minitems = 0;
443: pp->pr_minpages = 0;
444: pp->pr_maxpages = UINT_MAX;
1.20 thorpej 445: pp->pr_roflags = flags;
446: pp->pr_flags = 0;
1.35 pk 447: pp->pr_size = size;
1.3 pk 448: pp->pr_align = align;
449: pp->pr_wchan = wchan;
1.66 thorpej 450: pp->pr_alloc = palloc;
1.20 thorpej 451: pp->pr_nitems = 0;
452: pp->pr_nout = 0;
453: pp->pr_hardlimit = UINT_MAX;
454: pp->pr_hardlimit_warning = NULL;
1.31 thorpej 455: pp->pr_hardlimit_ratecap.tv_sec = 0;
456: pp->pr_hardlimit_ratecap.tv_usec = 0;
457: pp->pr_hardlimit_warning_last.tv_sec = 0;
458: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.3 pk 459:
460: /*
461: * Decide whether to put the page header off page to avoid
462: * wasting too large a part of the page. Off-page page headers
463: * go on a hash table, so we can match a returned item
464: * with its header based on the page address.
465: * We use 1/16 of the page size as the threshold (XXX: tune)
466: */
1.66 thorpej 467: if (pp->pr_size < palloc->pa_pagesz/16) {
1.3 pk 468: /* Use the end of the page for the page header */
1.20 thorpej 469: pp->pr_roflags |= PR_PHINPAGE;
1.66 thorpej 470: pp->pr_phoffset = off = palloc->pa_pagesz -
471: ALIGN(sizeof(struct pool_item_header));
1.2 pk 472: } else {
1.3 pk 473: /* The page header will be taken from our page header pool */
474: pp->pr_phoffset = 0;
1.66 thorpej 475: off = palloc->pa_pagesz;
1.16 briggs 476: for (i = 0; i < PR_HASHTABSIZE; i++) {
477: LIST_INIT(&pp->pr_hashtab[i]);
478: }
1.2 pk 479: }
1.1 pk 480:
1.3 pk 481: /*
482: * Alignment is to take place at `ioff' within the item. This means
483: * we must reserve up to `align - 1' bytes on the page to allow
484: * appropriate positioning of each item.
485: *
486: * Silently enforce `0 <= ioff < align'.
487: */
488: pp->pr_itemoffset = ioff = ioff % align;
489: pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
1.43 thorpej 490: KASSERT(pp->pr_itemsperpage != 0);
1.3 pk 491:
492: /*
493: * Use the slack between the chunks and the page header
494: * for "cache coloring".
495: */
496: slack = off - pp->pr_itemsperpage * pp->pr_size;
497: pp->pr_maxcolor = (slack / align) * align;
498: pp->pr_curcolor = 0;
499:
500: pp->pr_nget = 0;
501: pp->pr_nfail = 0;
502: pp->pr_nput = 0;
503: pp->pr_npagealloc = 0;
504: pp->pr_npagefree = 0;
1.1 pk 505: pp->pr_hiwat = 0;
1.8 thorpej 506: pp->pr_nidle = 0;
1.3 pk 507:
1.59 thorpej 508: #ifdef POOL_DIAGNOSTIC
1.25 thorpej 509: if (flags & PR_LOGGING) {
510: if (kmem_map == NULL ||
511: (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log),
512: M_TEMP, M_NOWAIT)) == NULL)
1.20 thorpej 513: pp->pr_roflags &= ~PR_LOGGING;
1.3 pk 514: pp->pr_curlogentry = 0;
515: pp->pr_logsize = pool_logsize;
516: }
1.59 thorpej 517: #endif
1.25 thorpej 518:
519: pp->pr_entered_file = NULL;
520: pp->pr_entered_line = 0;
1.3 pk 521:
1.21 thorpej 522: simple_lock_init(&pp->pr_slock);
1.1 pk 523:
1.3 pk 524: /*
1.43 thorpej 525: * Initialize private page header pool and cache magazine pool if we
526: * haven't done so yet.
1.23 thorpej 527: * XXX LOCKING.
1.3 pk 528: */
529: if (phpool.pr_size == 0) {
1.62 bjh21 530: #ifdef POOL_SUBPAGE
531: pool_init(&phpool, sizeof(struct pool_item_header), 0, 0, 0,
1.66 thorpej 532: "phpool", &pool_allocator_kmem);
1.62 bjh21 533: pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
1.66 thorpej 534: PR_RECURSIVE, "psppool", &pool_allocator_kmem);
1.62 bjh21 535: #else
1.3 pk 536: pool_init(&phpool, sizeof(struct pool_item_header), 0, 0,
1.66 thorpej 537: 0, "phpool", NULL);
1.62 bjh21 538: #endif
1.43 thorpej 539: pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
1.66 thorpej 540: 0, "pcgpool", NULL);
1.1 pk 541: }
542:
1.23 thorpej 543: /* Insert into the list of all pools. */
544: simple_lock(&pool_head_slock);
545: TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist);
546: simple_unlock(&pool_head_slock);
1.66 thorpej 547:
548: /* Insert this into the list of pools using this allocator. */
549: simple_lock(&palloc->pa_slock);
550: TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
551: simple_unlock(&palloc->pa_slock);
1.1 pk 552: }
553:
554: /*
555: * De-commision a pool resource.
556: */
557: void
1.42 thorpej 558: pool_destroy(struct pool *pp)
1.1 pk 559: {
1.3 pk 560: struct pool_item_header *ph;
1.43 thorpej 561: struct pool_cache *pc;
562:
1.66 thorpej 563: /* Locking order: pool_allocator -> pool */
564: simple_lock(&pp->pr_alloc->pa_slock);
565: TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
566: simple_unlock(&pp->pr_alloc->pa_slock);
567:
1.43 thorpej 568: /* Destroy all caches for this pool. */
569: while ((pc = TAILQ_FIRST(&pp->pr_cachelist)) != NULL)
570: pool_cache_destroy(pc);
1.3 pk 571:
572: #ifdef DIAGNOSTIC
1.20 thorpej 573: if (pp->pr_nout != 0) {
1.25 thorpej 574: pr_printlog(pp, NULL, printf);
1.20 thorpej 575: panic("pool_destroy: pool busy: still out: %u\n",
576: pp->pr_nout);
1.3 pk 577: }
578: #endif
1.1 pk 579:
1.3 pk 580: /* Remove all pages */
1.20 thorpej 581: if ((pp->pr_roflags & PR_STATIC) == 0)
1.61 chs 582: while ((ph = TAILQ_FIRST(&pp->pr_pagelist)) != NULL)
583: pr_rmpage(pp, ph, NULL);
1.3 pk 584:
585: /* Remove from global pool list */
1.23 thorpej 586: simple_lock(&pool_head_slock);
1.3 pk 587: TAILQ_REMOVE(&pool_head, pp, pr_poollist);
1.61 chs 588: if (drainpp == pp) {
589: drainpp = NULL;
590: }
1.23 thorpej 591: simple_unlock(&pool_head_slock);
1.3 pk 592:
1.59 thorpej 593: #ifdef POOL_DIAGNOSTIC
1.20 thorpej 594: if ((pp->pr_roflags & PR_LOGGING) != 0)
1.3 pk 595: free(pp->pr_log, M_TEMP);
1.59 thorpej 596: #endif
1.1 pk 597: }
598:
1.55 thorpej 599: static __inline struct pool_item_header *
600: pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags)
601: {
602: struct pool_item_header *ph;
603: int s;
604:
605: LOCK_ASSERT(simple_lock_held(&pp->pr_slock) == 0);
606:
607: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
608: ph = (struct pool_item_header *) (storage + pp->pr_phoffset);
609: else {
610: s = splhigh();
611: ph = pool_get(&phpool, flags);
612: splx(s);
613: }
614:
615: return (ph);
616: }
1.1 pk 617:
618: /*
1.3 pk 619: * Grab an item from the pool; must be called at appropriate spl level
1.1 pk 620: */
1.3 pk 621: void *
1.59 thorpej 622: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 623: _pool_get(struct pool *pp, int flags, const char *file, long line)
1.56 sommerfe 624: #else
625: pool_get(struct pool *pp, int flags)
626: #endif
1.1 pk 627: {
628: struct pool_item *pi;
1.3 pk 629: struct pool_item_header *ph;
1.55 thorpej 630: void *v;
1.1 pk 631:
1.2 pk 632: #ifdef DIAGNOSTIC
1.34 thorpej 633: if (__predict_false((pp->pr_roflags & PR_STATIC) &&
634: (flags & PR_MALLOCOK))) {
1.25 thorpej 635: pr_printlog(pp, NULL, printf);
1.2 pk 636: panic("pool_get: static");
1.3 pk 637: }
1.2 pk 638:
1.37 sommerfe 639: if (__predict_false(curproc == NULL && doing_shutdown == 0 &&
640: (flags & PR_WAITOK) != 0))
1.3 pk 641: panic("pool_get: must have NOWAIT");
1.58 thorpej 642:
643: #ifdef LOCKDEBUG
644: if (flags & PR_WAITOK)
645: simple_lock_only_held(NULL, "pool_get(PR_WAITOK)");
1.56 sommerfe 646: #endif
1.58 thorpej 647: #endif /* DIAGNOSTIC */
1.1 pk 648:
1.21 thorpej 649: simple_lock(&pp->pr_slock);
1.25 thorpej 650: pr_enter(pp, file, line);
1.20 thorpej 651:
652: startover:
653: /*
654: * Check to see if we've reached the hard limit. If we have,
655: * and we can wait, then wait until an item has been returned to
656: * the pool.
657: */
658: #ifdef DIAGNOSTIC
1.34 thorpej 659: if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
1.25 thorpej 660: pr_leave(pp);
1.21 thorpej 661: simple_unlock(&pp->pr_slock);
1.20 thorpej 662: panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
663: }
664: #endif
1.34 thorpej 665: if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
1.29 sommerfe 666: if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
1.20 thorpej 667: /*
668: * XXX: A warning isn't logged in this case. Should
669: * it be?
670: */
671: pp->pr_flags |= PR_WANTED;
1.25 thorpej 672: pr_leave(pp);
1.40 sommerfe 673: ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
1.25 thorpej 674: pr_enter(pp, file, line);
1.20 thorpej 675: goto startover;
676: }
1.31 thorpej 677:
678: /*
679: * Log a message that the hard limit has been hit.
680: */
681: if (pp->pr_hardlimit_warning != NULL &&
682: ratecheck(&pp->pr_hardlimit_warning_last,
683: &pp->pr_hardlimit_ratecap))
684: log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
1.21 thorpej 685:
686: pp->pr_nfail++;
687:
1.25 thorpej 688: pr_leave(pp);
1.21 thorpej 689: simple_unlock(&pp->pr_slock);
1.20 thorpej 690: return (NULL);
691: }
692:
1.3 pk 693: /*
694: * The convention we use is that if `curpage' is not NULL, then
695: * it points at a non-empty bucket. In particular, `curpage'
696: * never points at a page header which has PR_PHINPAGE set and
697: * has no items in its bucket.
698: */
1.20 thorpej 699: if ((ph = pp->pr_curpage) == NULL) {
700: #ifdef DIAGNOSTIC
701: if (pp->pr_nitems != 0) {
1.21 thorpej 702: simple_unlock(&pp->pr_slock);
1.20 thorpej 703: printf("pool_get: %s: curpage NULL, nitems %u\n",
704: pp->pr_wchan, pp->pr_nitems);
705: panic("pool_get: nitems inconsistent\n");
706: }
707: #endif
708:
1.21 thorpej 709: /*
710: * Call the back-end page allocator for more memory.
711: * Release the pool lock, as the back-end page allocator
712: * may block.
713: */
1.25 thorpej 714: pr_leave(pp);
1.21 thorpej 715: simple_unlock(&pp->pr_slock);
1.66 thorpej 716: v = pool_allocator_alloc(pp, flags);
1.55 thorpej 717: if (__predict_true(v != NULL))
718: ph = pool_alloc_item_header(pp, v, flags);
1.21 thorpej 719: simple_lock(&pp->pr_slock);
1.25 thorpej 720: pr_enter(pp, file, line);
1.15 pk 721:
1.55 thorpej 722: if (__predict_false(v == NULL || ph == NULL)) {
723: if (v != NULL)
1.66 thorpej 724: pool_allocator_free(pp, v);
1.55 thorpej 725:
1.21 thorpej 726: /*
1.55 thorpej 727: * We were unable to allocate a page or item
728: * header, but we released the lock during
729: * allocation, so perhaps items were freed
730: * back to the pool. Check for this case.
1.21 thorpej 731: */
732: if (pp->pr_curpage != NULL)
733: goto startover;
1.15 pk 734:
1.3 pk 735: if ((flags & PR_WAITOK) == 0) {
736: pp->pr_nfail++;
1.25 thorpej 737: pr_leave(pp);
1.21 thorpej 738: simple_unlock(&pp->pr_slock);
1.1 pk 739: return (NULL);
1.3 pk 740: }
741:
1.15 pk 742: /*
743: * Wait for items to be returned to this pool.
1.21 thorpej 744: *
1.20 thorpej 745: * XXX: maybe we should wake up once a second and
746: * try again?
1.15 pk 747: */
1.1 pk 748: pp->pr_flags |= PR_WANTED;
1.66 thorpej 749: /* PA_WANTED is already set on the allocator. */
1.25 thorpej 750: pr_leave(pp);
1.40 sommerfe 751: ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
1.25 thorpej 752: pr_enter(pp, file, line);
1.20 thorpej 753: goto startover;
1.1 pk 754: }
1.3 pk 755:
1.15 pk 756: /* We have more memory; add it to the pool */
1.55 thorpej 757: pool_prime_page(pp, v, ph);
1.15 pk 758: pp->pr_npagealloc++;
759:
1.20 thorpej 760: /* Start the allocation process over. */
761: goto startover;
1.3 pk 762: }
763:
1.34 thorpej 764: if (__predict_false((v = pi = TAILQ_FIRST(&ph->ph_itemlist)) == NULL)) {
1.25 thorpej 765: pr_leave(pp);
1.21 thorpej 766: simple_unlock(&pp->pr_slock);
1.3 pk 767: panic("pool_get: %s: page empty", pp->pr_wchan);
1.21 thorpej 768: }
1.20 thorpej 769: #ifdef DIAGNOSTIC
1.34 thorpej 770: if (__predict_false(pp->pr_nitems == 0)) {
1.25 thorpej 771: pr_leave(pp);
1.21 thorpej 772: simple_unlock(&pp->pr_slock);
1.20 thorpej 773: printf("pool_get: %s: items on itemlist, nitems %u\n",
774: pp->pr_wchan, pp->pr_nitems);
775: panic("pool_get: nitems inconsistent\n");
776: }
1.65 enami 777: #endif
1.56 sommerfe 778:
1.65 enami 779: #ifdef POOL_DIAGNOSTIC
1.3 pk 780: pr_log(pp, v, PRLOG_GET, file, line);
1.65 enami 781: #endif
1.3 pk 782:
1.65 enami 783: #ifdef DIAGNOSTIC
1.34 thorpej 784: if (__predict_false(pi->pi_magic != PI_MAGIC)) {
1.25 thorpej 785: pr_printlog(pp, pi, printf);
1.3 pk 786: panic("pool_get(%s): free list modified: magic=%x; page %p;"
787: " item addr %p\n",
788: pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
789: }
790: #endif
791:
792: /*
793: * Remove from item list.
794: */
795: TAILQ_REMOVE(&ph->ph_itemlist, pi, pi_list);
1.20 thorpej 796: pp->pr_nitems--;
797: pp->pr_nout++;
1.6 thorpej 798: if (ph->ph_nmissing == 0) {
799: #ifdef DIAGNOSTIC
1.34 thorpej 800: if (__predict_false(pp->pr_nidle == 0))
1.6 thorpej 801: panic("pool_get: nidle inconsistent");
802: #endif
803: pp->pr_nidle--;
804: }
1.3 pk 805: ph->ph_nmissing++;
806: if (TAILQ_FIRST(&ph->ph_itemlist) == NULL) {
1.21 thorpej 807: #ifdef DIAGNOSTIC
1.34 thorpej 808: if (__predict_false(ph->ph_nmissing != pp->pr_itemsperpage)) {
1.25 thorpej 809: pr_leave(pp);
1.21 thorpej 810: simple_unlock(&pp->pr_slock);
811: panic("pool_get: %s: nmissing inconsistent",
812: pp->pr_wchan);
813: }
814: #endif
1.3 pk 815: /*
816: * Find a new non-empty page header, if any.
817: * Start search from the page head, to increase
818: * the chance for "high water" pages to be freed.
819: *
1.21 thorpej 820: * Migrate empty pages to the end of the list. This
821: * will speed the update of curpage as pages become
822: * idle. Empty pages intermingled with idle pages
823: * is no big deal. As soon as a page becomes un-empty,
824: * it will move back to the head of the list.
1.3 pk 825: */
826: TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
1.21 thorpej 827: TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist);
1.61 chs 828: TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist)
1.3 pk 829: if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
830: break;
831:
832: pp->pr_curpage = ph;
1.1 pk 833: }
1.3 pk 834:
835: pp->pr_nget++;
1.20 thorpej 836:
837: /*
838: * If we have a low water mark and we are now below that low
839: * water mark, add more items to the pool.
840: */
1.53 thorpej 841: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 842: /*
843: * XXX: Should we log a warning? Should we set up a timeout
844: * to try again in a second or so? The latter could break
845: * a caller's assumptions about interrupt protection, etc.
846: */
847: }
848:
1.25 thorpej 849: pr_leave(pp);
1.21 thorpej 850: simple_unlock(&pp->pr_slock);
1.1 pk 851: return (v);
852: }
853:
854: /*
1.43 thorpej 855: * Internal version of pool_put(). Pool is already locked/entered.
1.1 pk 856: */
1.43 thorpej 857: static void
1.56 sommerfe 858: pool_do_put(struct pool *pp, void *v)
1.1 pk 859: {
860: struct pool_item *pi = v;
1.3 pk 861: struct pool_item_header *ph;
862: caddr_t page;
1.21 thorpej 863: int s;
1.3 pk 864:
1.61 chs 865: LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
866:
1.66 thorpej 867: page = (caddr_t)((u_long)v & pp->pr_alloc->pa_pagemask);
1.1 pk 868:
1.30 thorpej 869: #ifdef DIAGNOSTIC
1.34 thorpej 870: if (__predict_false(pp->pr_nout == 0)) {
1.30 thorpej 871: printf("pool %s: putting with none out\n",
872: pp->pr_wchan);
873: panic("pool_put");
874: }
875: #endif
1.3 pk 876:
1.34 thorpej 877: if (__predict_false((ph = pr_find_pagehead(pp, page)) == NULL)) {
1.25 thorpej 878: pr_printlog(pp, NULL, printf);
1.3 pk 879: panic("pool_put: %s: page header missing", pp->pr_wchan);
880: }
1.28 thorpej 881:
882: #ifdef LOCKDEBUG
883: /*
884: * Check if we're freeing a locked simple lock.
885: */
886: simple_lock_freecheck((caddr_t)pi, ((caddr_t)pi) + pp->pr_size);
887: #endif
1.3 pk 888:
889: /*
890: * Return to item list.
891: */
1.2 pk 892: #ifdef DIAGNOSTIC
1.3 pk 893: pi->pi_magic = PI_MAGIC;
894: #endif
1.32 chs 895: #ifdef DEBUG
896: {
897: int i, *ip = v;
898:
899: for (i = 0; i < pp->pr_size / sizeof(int); i++) {
900: *ip++ = PI_MAGIC;
901: }
902: }
903: #endif
904:
1.3 pk 905: TAILQ_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
906: ph->ph_nmissing--;
907: pp->pr_nput++;
1.20 thorpej 908: pp->pr_nitems++;
909: pp->pr_nout--;
1.3 pk 910:
911: /* Cancel "pool empty" condition if it exists */
912: if (pp->pr_curpage == NULL)
913: pp->pr_curpage = ph;
914:
915: if (pp->pr_flags & PR_WANTED) {
916: pp->pr_flags &= ~PR_WANTED;
1.15 pk 917: if (ph->ph_nmissing == 0)
918: pp->pr_nidle++;
1.3 pk 919: wakeup((caddr_t)pp);
920: return;
921: }
922:
923: /*
1.21 thorpej 924: * If this page is now complete, do one of two things:
925: *
926: * (1) If we have more pages than the page high water
927: * mark, free the page back to the system.
928: *
929: * (2) Move it to the end of the page list, so that
930: * we minimize our chances of fragmenting the
931: * pool. Idle pages migrate to the end (along with
932: * completely empty pages, so that we find un-empty
933: * pages more quickly when we update curpage) of the
934: * list so they can be more easily swept up by
935: * the pagedaemon when pages are scarce.
1.3 pk 936: */
937: if (ph->ph_nmissing == 0) {
1.6 thorpej 938: pp->pr_nidle++;
1.3 pk 939: if (pp->pr_npages > pp->pr_maxpages) {
1.61 chs 940: pr_rmpage(pp, ph, NULL);
1.3 pk 941: } else {
942: TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
943: TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist);
944:
1.21 thorpej 945: /*
946: * Update the timestamp on the page. A page must
947: * be idle for some period of time before it can
948: * be reclaimed by the pagedaemon. This minimizes
949: * ping-pong'ing for memory.
950: */
951: s = splclock();
952: ph->ph_time = mono_time;
953: splx(s);
954:
955: /*
956: * Update the current page pointer. Just look for
957: * the first page with any free items.
958: *
959: * XXX: Maybe we want an option to look for the
960: * page with the fewest available items, to minimize
961: * fragmentation?
962: */
1.61 chs 963: TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist)
1.3 pk 964: if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
965: break;
1.1 pk 966:
1.3 pk 967: pp->pr_curpage = ph;
1.1 pk 968: }
969: }
1.21 thorpej 970: /*
971: * If the page has just become un-empty, move it to the head of
972: * the list, and make it the current page. The next allocation
973: * will get the item from this page, instead of further fragmenting
974: * the pool.
975: */
976: else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
977: TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
978: TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist);
979: pp->pr_curpage = ph;
980: }
1.43 thorpej 981: }
982:
983: /*
984: * Return resource to the pool; must be called at appropriate spl level
985: */
1.59 thorpej 986: #ifdef POOL_DIAGNOSTIC
1.43 thorpej 987: void
988: _pool_put(struct pool *pp, void *v, const char *file, long line)
989: {
990:
991: simple_lock(&pp->pr_slock);
992: pr_enter(pp, file, line);
993:
1.56 sommerfe 994: pr_log(pp, v, PRLOG_PUT, file, line);
995:
996: pool_do_put(pp, v);
1.21 thorpej 997:
1.25 thorpej 998: pr_leave(pp);
1.21 thorpej 999: simple_unlock(&pp->pr_slock);
1.1 pk 1000: }
1.57 sommerfe 1001: #undef pool_put
1.59 thorpej 1002: #endif /* POOL_DIAGNOSTIC */
1.1 pk 1003:
1.56 sommerfe 1004: void
1005: pool_put(struct pool *pp, void *v)
1006: {
1007:
1008: simple_lock(&pp->pr_slock);
1009:
1010: pool_do_put(pp, v);
1011:
1012: simple_unlock(&pp->pr_slock);
1013: }
1.57 sommerfe 1014:
1.59 thorpej 1015: #ifdef POOL_DIAGNOSTIC
1.57 sommerfe 1016: #define pool_put(h, v) _pool_put((h), (v), __FILE__, __LINE__)
1.56 sommerfe 1017: #endif
1018:
1.1 pk 1019: /*
1.55 thorpej 1020: * Add N items to the pool.
1021: */
1022: int
1023: pool_prime(struct pool *pp, int n)
1024: {
1025: struct pool_item_header *ph;
1026: caddr_t cp;
1027: int newpages, error = 0;
1028:
1029: simple_lock(&pp->pr_slock);
1030:
1031: newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1032:
1033: while (newpages-- > 0) {
1034: simple_unlock(&pp->pr_slock);
1.66 thorpej 1035: cp = pool_allocator_alloc(pp, PR_NOWAIT);
1.55 thorpej 1036: if (__predict_true(cp != NULL))
1037: ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1038: simple_lock(&pp->pr_slock);
1039:
1040: if (__predict_false(cp == NULL || ph == NULL)) {
1041: error = ENOMEM;
1042: if (cp != NULL)
1.66 thorpej 1043: pool_allocator_free(pp, cp);
1.55 thorpej 1044: break;
1045: }
1046:
1047: pool_prime_page(pp, cp, ph);
1048: pp->pr_npagealloc++;
1049: pp->pr_minpages++;
1050: }
1051:
1052: if (pp->pr_minpages >= pp->pr_maxpages)
1053: pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */
1054:
1055: simple_unlock(&pp->pr_slock);
1056: return (0);
1057: }
1058:
1059: /*
1.3 pk 1060: * Add a page worth of items to the pool.
1.21 thorpej 1061: *
1062: * Note, we must be called with the pool descriptor LOCKED.
1.3 pk 1063: */
1.55 thorpej 1064: static void
1065: pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph)
1.3 pk 1066: {
1067: struct pool_item *pi;
1068: caddr_t cp = storage;
1069: unsigned int align = pp->pr_align;
1070: unsigned int ioff = pp->pr_itemoffset;
1.55 thorpej 1071: int n;
1.36 pk 1072:
1.66 thorpej 1073: #ifdef DIAGNOSTIC
1074: if (((u_long)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
1.36 pk 1075: panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1.66 thorpej 1076: #endif
1.3 pk 1077:
1.55 thorpej 1078: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1.3 pk 1079: LIST_INSERT_HEAD(&pp->pr_hashtab[PR_HASH_INDEX(pp, cp)],
1.55 thorpej 1080: ph, ph_hashlist);
1.3 pk 1081:
1082: /*
1083: * Insert page header.
1084: */
1085: TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist);
1086: TAILQ_INIT(&ph->ph_itemlist);
1087: ph->ph_page = storage;
1088: ph->ph_nmissing = 0;
1.21 thorpej 1089: memset(&ph->ph_time, 0, sizeof(ph->ph_time));
1.3 pk 1090:
1.6 thorpej 1091: pp->pr_nidle++;
1092:
1.3 pk 1093: /*
1094: * Color this page.
1095: */
1096: cp = (caddr_t)(cp + pp->pr_curcolor);
1097: if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
1098: pp->pr_curcolor = 0;
1099:
1100: /*
1101: * Adjust storage to apply aligment to `pr_itemoffset' in each item.
1102: */
1103: if (ioff != 0)
1104: cp = (caddr_t)(cp + (align - ioff));
1105:
1106: /*
1107: * Insert remaining chunks on the bucket list.
1108: */
1109: n = pp->pr_itemsperpage;
1.20 thorpej 1110: pp->pr_nitems += n;
1.3 pk 1111:
1112: while (n--) {
1113: pi = (struct pool_item *)cp;
1114:
1115: /* Insert on page list */
1116: TAILQ_INSERT_TAIL(&ph->ph_itemlist, pi, pi_list);
1117: #ifdef DIAGNOSTIC
1118: pi->pi_magic = PI_MAGIC;
1119: #endif
1120: cp = (caddr_t)(cp + pp->pr_size);
1121: }
1122:
1123: /*
1124: * If the pool was depleted, point at the new page.
1125: */
1126: if (pp->pr_curpage == NULL)
1127: pp->pr_curpage = ph;
1128:
1129: if (++pp->pr_npages > pp->pr_hiwat)
1130: pp->pr_hiwat = pp->pr_npages;
1131: }
1132:
1.20 thorpej 1133: /*
1.52 thorpej 1134: * Used by pool_get() when nitems drops below the low water mark. This
1135: * is used to catch up nitmes with the low water mark.
1.20 thorpej 1136: *
1.21 thorpej 1137: * Note 1, we never wait for memory here, we let the caller decide what to do.
1.20 thorpej 1138: *
1139: * Note 2, this doesn't work with static pools.
1140: *
1141: * Note 3, we must be called with the pool already locked, and we return
1142: * with it locked.
1143: */
1144: static int
1.42 thorpej 1145: pool_catchup(struct pool *pp)
1.20 thorpej 1146: {
1.55 thorpej 1147: struct pool_item_header *ph;
1.20 thorpej 1148: caddr_t cp;
1149: int error = 0;
1150:
1151: if (pp->pr_roflags & PR_STATIC) {
1152: /*
1153: * We dropped below the low water mark, and this is not a
1154: * good thing. Log a warning.
1.21 thorpej 1155: *
1156: * XXX: rate-limit this?
1.20 thorpej 1157: */
1158: printf("WARNING: static pool `%s' dropped below low water "
1159: "mark\n", pp->pr_wchan);
1160: return (0);
1161: }
1162:
1.54 thorpej 1163: while (POOL_NEEDS_CATCHUP(pp)) {
1.20 thorpej 1164: /*
1.21 thorpej 1165: * Call the page back-end allocator for more memory.
1166: *
1167: * XXX: We never wait, so should we bother unlocking
1168: * the pool descriptor?
1.20 thorpej 1169: */
1.21 thorpej 1170: simple_unlock(&pp->pr_slock);
1.66 thorpej 1171: cp = pool_allocator_alloc(pp, PR_NOWAIT);
1.55 thorpej 1172: if (__predict_true(cp != NULL))
1173: ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1.21 thorpej 1174: simple_lock(&pp->pr_slock);
1.55 thorpej 1175: if (__predict_false(cp == NULL || ph == NULL)) {
1176: if (cp != NULL)
1.66 thorpej 1177: pool_allocator_free(pp, cp);
1.20 thorpej 1178: error = ENOMEM;
1179: break;
1180: }
1.55 thorpej 1181: pool_prime_page(pp, cp, ph);
1.26 thorpej 1182: pp->pr_npagealloc++;
1.20 thorpej 1183: }
1184:
1185: return (error);
1186: }
1187:
1.3 pk 1188: void
1.42 thorpej 1189: pool_setlowat(struct pool *pp, int n)
1.3 pk 1190: {
1.20 thorpej 1191: int error;
1.15 pk 1192:
1.21 thorpej 1193: simple_lock(&pp->pr_slock);
1194:
1.3 pk 1195: pp->pr_minitems = n;
1.15 pk 1196: pp->pr_minpages = (n == 0)
1197: ? 0
1.18 thorpej 1198: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.20 thorpej 1199:
1200: /* Make sure we're caught up with the newly-set low water mark. */
1.53 thorpej 1201: if (POOL_NEEDS_CATCHUP(pp) && (error = pool_catchup(pp) != 0)) {
1.20 thorpej 1202: /*
1203: * XXX: Should we log a warning? Should we set up a timeout
1204: * to try again in a second or so? The latter could break
1205: * a caller's assumptions about interrupt protection, etc.
1206: */
1207: }
1.21 thorpej 1208:
1209: simple_unlock(&pp->pr_slock);
1.3 pk 1210: }
1211:
1212: void
1.42 thorpej 1213: pool_sethiwat(struct pool *pp, int n)
1.3 pk 1214: {
1.15 pk 1215:
1.21 thorpej 1216: simple_lock(&pp->pr_slock);
1217:
1.15 pk 1218: pp->pr_maxpages = (n == 0)
1219: ? 0
1.18 thorpej 1220: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1221:
1222: simple_unlock(&pp->pr_slock);
1.3 pk 1223: }
1224:
1.20 thorpej 1225: void
1.42 thorpej 1226: pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1.20 thorpej 1227: {
1228:
1.21 thorpej 1229: simple_lock(&pp->pr_slock);
1.20 thorpej 1230:
1231: pp->pr_hardlimit = n;
1232: pp->pr_hardlimit_warning = warnmess;
1.31 thorpej 1233: pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1234: pp->pr_hardlimit_warning_last.tv_sec = 0;
1235: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.20 thorpej 1236:
1237: /*
1.21 thorpej 1238: * In-line version of pool_sethiwat(), because we don't want to
1239: * release the lock.
1.20 thorpej 1240: */
1241: pp->pr_maxpages = (n == 0)
1242: ? 0
1243: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1244:
1245: simple_unlock(&pp->pr_slock);
1.20 thorpej 1246: }
1.3 pk 1247:
1248: /*
1249: * Release all complete pages that have not been used recently.
1250: */
1.66 thorpej 1251: int
1.59 thorpej 1252: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 1253: _pool_reclaim(struct pool *pp, const char *file, long line)
1.56 sommerfe 1254: #else
1255: pool_reclaim(struct pool *pp)
1256: #endif
1.3 pk 1257: {
1258: struct pool_item_header *ph, *phnext;
1.43 thorpej 1259: struct pool_cache *pc;
1.21 thorpej 1260: struct timeval curtime;
1.61 chs 1261: struct pool_pagelist pq;
1.21 thorpej 1262: int s;
1.3 pk 1263:
1.20 thorpej 1264: if (pp->pr_roflags & PR_STATIC)
1.66 thorpej 1265: return (0);
1.3 pk 1266:
1.21 thorpej 1267: if (simple_lock_try(&pp->pr_slock) == 0)
1.66 thorpej 1268: return (0);
1.25 thorpej 1269: pr_enter(pp, file, line);
1.61 chs 1270: TAILQ_INIT(&pq);
1.3 pk 1271:
1.43 thorpej 1272: /*
1273: * Reclaim items from the pool's caches.
1274: */
1.61 chs 1275: TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
1.43 thorpej 1276: pool_cache_reclaim(pc);
1277:
1.21 thorpej 1278: s = splclock();
1279: curtime = mono_time;
1280: splx(s);
1281:
1.3 pk 1282: for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; ph = phnext) {
1283: phnext = TAILQ_NEXT(ph, ph_pagelist);
1284:
1285: /* Check our minimum page claim */
1286: if (pp->pr_npages <= pp->pr_minpages)
1287: break;
1288:
1289: if (ph->ph_nmissing == 0) {
1290: struct timeval diff;
1291: timersub(&curtime, &ph->ph_time, &diff);
1292: if (diff.tv_sec < pool_inactive_time)
1293: continue;
1.21 thorpej 1294:
1295: /*
1296: * If freeing this page would put us below
1297: * the low water mark, stop now.
1298: */
1299: if ((pp->pr_nitems - pp->pr_itemsperpage) <
1300: pp->pr_minitems)
1301: break;
1302:
1.61 chs 1303: pr_rmpage(pp, ph, &pq);
1.3 pk 1304: }
1305: }
1306:
1.25 thorpej 1307: pr_leave(pp);
1.21 thorpej 1308: simple_unlock(&pp->pr_slock);
1.66 thorpej 1309: if (TAILQ_EMPTY(&pq))
1310: return (0);
1311:
1.61 chs 1312: while ((ph = TAILQ_FIRST(&pq)) != NULL) {
1313: TAILQ_REMOVE(&pq, ph, ph_pagelist);
1.66 thorpej 1314: pool_allocator_free(pp, ph->ph_page);
1.61 chs 1315: if (pp->pr_roflags & PR_PHINPAGE) {
1316: continue;
1317: }
1318: LIST_REMOVE(ph, ph_hashlist);
1319: s = splhigh();
1320: pool_put(&phpool, ph);
1321: splx(s);
1322: }
1.66 thorpej 1323:
1324: return (1);
1.3 pk 1325: }
1326:
1327: /*
1328: * Drain pools, one at a time.
1.21 thorpej 1329: *
1330: * Note, we must never be called from an interrupt context.
1.3 pk 1331: */
1332: void
1.42 thorpej 1333: pool_drain(void *arg)
1.3 pk 1334: {
1335: struct pool *pp;
1.23 thorpej 1336: int s;
1.3 pk 1337:
1.61 chs 1338: pp = NULL;
1.49 thorpej 1339: s = splvm();
1.23 thorpej 1340: simple_lock(&pool_head_slock);
1.61 chs 1341: if (drainpp == NULL) {
1342: drainpp = TAILQ_FIRST(&pool_head);
1343: }
1344: if (drainpp) {
1345: pp = drainpp;
1346: drainpp = TAILQ_NEXT(pp, pr_poollist);
1347: }
1348: simple_unlock(&pool_head_slock);
1.63 chs 1349: pool_reclaim(pp);
1.61 chs 1350: splx(s);
1.3 pk 1351: }
1352:
1353: /*
1354: * Diagnostic helpers.
1355: */
1356: void
1.42 thorpej 1357: pool_print(struct pool *pp, const char *modif)
1.21 thorpej 1358: {
1359: int s;
1360:
1.49 thorpej 1361: s = splvm();
1.25 thorpej 1362: if (simple_lock_try(&pp->pr_slock) == 0) {
1363: printf("pool %s is locked; try again later\n",
1364: pp->pr_wchan);
1365: splx(s);
1366: return;
1367: }
1368: pool_print1(pp, modif, printf);
1.21 thorpej 1369: simple_unlock(&pp->pr_slock);
1370: splx(s);
1371: }
1372:
1.25 thorpej 1373: void
1.42 thorpej 1374: pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.25 thorpej 1375: {
1376: int didlock = 0;
1377:
1378: if (pp == NULL) {
1379: (*pr)("Must specify a pool to print.\n");
1380: return;
1381: }
1382:
1383: /*
1384: * Called from DDB; interrupts should be blocked, and all
1385: * other processors should be paused. We can skip locking
1386: * the pool in this case.
1387: *
1388: * We do a simple_lock_try() just to print the lock
1389: * status, however.
1390: */
1391:
1392: if (simple_lock_try(&pp->pr_slock) == 0)
1393: (*pr)("WARNING: pool %s is locked\n", pp->pr_wchan);
1394: else
1395: didlock = 1;
1396:
1397: pool_print1(pp, modif, pr);
1398:
1399: if (didlock)
1400: simple_unlock(&pp->pr_slock);
1401: }
1402:
1.21 thorpej 1403: static void
1.42 thorpej 1404: pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.3 pk 1405: {
1.25 thorpej 1406: struct pool_item_header *ph;
1.44 thorpej 1407: struct pool_cache *pc;
1408: struct pool_cache_group *pcg;
1.25 thorpej 1409: #ifdef DIAGNOSTIC
1410: struct pool_item *pi;
1411: #endif
1.44 thorpej 1412: int i, print_log = 0, print_pagelist = 0, print_cache = 0;
1.25 thorpej 1413: char c;
1414:
1415: while ((c = *modif++) != '\0') {
1416: if (c == 'l')
1417: print_log = 1;
1418: if (c == 'p')
1419: print_pagelist = 1;
1.44 thorpej 1420: if (c == 'c')
1421: print_cache = 1;
1.25 thorpej 1422: modif++;
1423: }
1424:
1425: (*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1426: pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1427: pp->pr_roflags);
1.66 thorpej 1428: (*pr)("\talloc %p\n", pp->pr_alloc);
1.25 thorpej 1429: (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1430: pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1431: (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1432: pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1433:
1434: (*pr)("\n\tnget %lu, nfail %lu, nput %lu\n",
1435: pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1436: (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1437: pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1438:
1439: if (print_pagelist == 0)
1440: goto skip_pagelist;
1441:
1442: if ((ph = TAILQ_FIRST(&pp->pr_pagelist)) != NULL)
1443: (*pr)("\n\tpage list:\n");
1444: for (; ph != NULL; ph = TAILQ_NEXT(ph, ph_pagelist)) {
1445: (*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n",
1446: ph->ph_page, ph->ph_nmissing,
1447: (u_long)ph->ph_time.tv_sec,
1448: (u_long)ph->ph_time.tv_usec);
1449: #ifdef DIAGNOSTIC
1.61 chs 1450: TAILQ_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1.25 thorpej 1451: if (pi->pi_magic != PI_MAGIC) {
1452: (*pr)("\t\t\titem %p, magic 0x%x\n",
1453: pi, pi->pi_magic);
1454: }
1455: }
1456: #endif
1457: }
1458: if (pp->pr_curpage == NULL)
1459: (*pr)("\tno current page\n");
1460: else
1461: (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1462:
1463: skip_pagelist:
1464:
1465: if (print_log == 0)
1466: goto skip_log;
1467:
1468: (*pr)("\n");
1469: if ((pp->pr_roflags & PR_LOGGING) == 0)
1470: (*pr)("\tno log\n");
1471: else
1472: pr_printlog(pp, NULL, pr);
1.3 pk 1473:
1.25 thorpej 1474: skip_log:
1.44 thorpej 1475:
1476: if (print_cache == 0)
1477: goto skip_cache;
1478:
1.61 chs 1479: TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist) {
1.44 thorpej 1480: (*pr)("\tcache %p: allocfrom %p freeto %p\n", pc,
1481: pc->pc_allocfrom, pc->pc_freeto);
1.48 thorpej 1482: (*pr)("\t hits %lu misses %lu ngroups %lu nitems %lu\n",
1483: pc->pc_hits, pc->pc_misses, pc->pc_ngroups, pc->pc_nitems);
1.61 chs 1484: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.44 thorpej 1485: (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail);
1486: for (i = 0; i < PCG_NOBJECTS; i++)
1487: (*pr)("\t\t\t%p\n", pcg->pcg_objects[i]);
1488: }
1489: }
1490:
1491: skip_cache:
1.3 pk 1492:
1.25 thorpej 1493: pr_enter_check(pp, pr);
1.3 pk 1494: }
1495:
1496: int
1.42 thorpej 1497: pool_chk(struct pool *pp, const char *label)
1.3 pk 1498: {
1499: struct pool_item_header *ph;
1500: int r = 0;
1501:
1.21 thorpej 1502: simple_lock(&pp->pr_slock);
1.3 pk 1503:
1.61 chs 1504: TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist) {
1.3 pk 1505: struct pool_item *pi;
1506: int n;
1507: caddr_t page;
1508:
1.66 thorpej 1509: page = (caddr_t)((u_long)ph & pp->pr_alloc->pa_pagemask);
1.20 thorpej 1510: if (page != ph->ph_page &&
1511: (pp->pr_roflags & PR_PHINPAGE) != 0) {
1.3 pk 1512: if (label != NULL)
1513: printf("%s: ", label);
1.16 briggs 1514: printf("pool(%p:%s): page inconsistency: page %p;"
1515: " at page head addr %p (p %p)\n", pp,
1.3 pk 1516: pp->pr_wchan, ph->ph_page,
1517: ph, page);
1518: r++;
1519: goto out;
1520: }
1521:
1522: for (pi = TAILQ_FIRST(&ph->ph_itemlist), n = 0;
1523: pi != NULL;
1524: pi = TAILQ_NEXT(pi,pi_list), n++) {
1525:
1526: #ifdef DIAGNOSTIC
1527: if (pi->pi_magic != PI_MAGIC) {
1528: if (label != NULL)
1529: printf("%s: ", label);
1530: printf("pool(%s): free list modified: magic=%x;"
1531: " page %p; item ordinal %d;"
1532: " addr %p (p %p)\n",
1533: pp->pr_wchan, pi->pi_magic, ph->ph_page,
1534: n, pi, page);
1535: panic("pool");
1536: }
1537: #endif
1.66 thorpej 1538: page =
1539: (caddr_t)((u_long)pi & pp->pr_alloc->pa_pagemask);
1.3 pk 1540: if (page == ph->ph_page)
1541: continue;
1542:
1543: if (label != NULL)
1544: printf("%s: ", label);
1.16 briggs 1545: printf("pool(%p:%s): page inconsistency: page %p;"
1546: " item ordinal %d; addr %p (p %p)\n", pp,
1.3 pk 1547: pp->pr_wchan, ph->ph_page,
1548: n, pi, page);
1549: r++;
1550: goto out;
1551: }
1552: }
1553: out:
1.21 thorpej 1554: simple_unlock(&pp->pr_slock);
1.3 pk 1555: return (r);
1.43 thorpej 1556: }
1557:
1558: /*
1559: * pool_cache_init:
1560: *
1561: * Initialize a pool cache.
1562: *
1563: * NOTE: If the pool must be protected from interrupts, we expect
1564: * to be called at the appropriate interrupt priority level.
1565: */
1566: void
1567: pool_cache_init(struct pool_cache *pc, struct pool *pp,
1568: int (*ctor)(void *, void *, int),
1569: void (*dtor)(void *, void *),
1570: void *arg)
1571: {
1572:
1573: TAILQ_INIT(&pc->pc_grouplist);
1574: simple_lock_init(&pc->pc_slock);
1575:
1576: pc->pc_allocfrom = NULL;
1577: pc->pc_freeto = NULL;
1578: pc->pc_pool = pp;
1579:
1580: pc->pc_ctor = ctor;
1581: pc->pc_dtor = dtor;
1582: pc->pc_arg = arg;
1583:
1.48 thorpej 1584: pc->pc_hits = 0;
1585: pc->pc_misses = 0;
1586:
1587: pc->pc_ngroups = 0;
1588:
1589: pc->pc_nitems = 0;
1590:
1.43 thorpej 1591: simple_lock(&pp->pr_slock);
1592: TAILQ_INSERT_TAIL(&pp->pr_cachelist, pc, pc_poollist);
1593: simple_unlock(&pp->pr_slock);
1594: }
1595:
1596: /*
1597: * pool_cache_destroy:
1598: *
1599: * Destroy a pool cache.
1600: */
1601: void
1602: pool_cache_destroy(struct pool_cache *pc)
1603: {
1604: struct pool *pp = pc->pc_pool;
1605:
1606: /* First, invalidate the entire cache. */
1607: pool_cache_invalidate(pc);
1608:
1609: /* ...and remove it from the pool's cache list. */
1610: simple_lock(&pp->pr_slock);
1611: TAILQ_REMOVE(&pp->pr_cachelist, pc, pc_poollist);
1612: simple_unlock(&pp->pr_slock);
1613: }
1614:
1615: static __inline void *
1616: pcg_get(struct pool_cache_group *pcg)
1617: {
1618: void *object;
1619: u_int idx;
1620:
1621: KASSERT(pcg->pcg_avail <= PCG_NOBJECTS);
1.45 thorpej 1622: KASSERT(pcg->pcg_avail != 0);
1.43 thorpej 1623: idx = --pcg->pcg_avail;
1624:
1625: KASSERT(pcg->pcg_objects[idx] != NULL);
1626: object = pcg->pcg_objects[idx];
1627: pcg->pcg_objects[idx] = NULL;
1628:
1629: return (object);
1630: }
1631:
1632: static __inline void
1633: pcg_put(struct pool_cache_group *pcg, void *object)
1634: {
1635: u_int idx;
1636:
1637: KASSERT(pcg->pcg_avail < PCG_NOBJECTS);
1638: idx = pcg->pcg_avail++;
1639:
1640: KASSERT(pcg->pcg_objects[idx] == NULL);
1641: pcg->pcg_objects[idx] = object;
1642: }
1643:
1644: /*
1645: * pool_cache_get:
1646: *
1647: * Get an object from a pool cache.
1648: */
1649: void *
1650: pool_cache_get(struct pool_cache *pc, int flags)
1651: {
1652: struct pool_cache_group *pcg;
1653: void *object;
1.58 thorpej 1654:
1655: #ifdef LOCKDEBUG
1656: if (flags & PR_WAITOK)
1657: simple_lock_only_held(NULL, "pool_cache_get(PR_WAITOK)");
1658: #endif
1.43 thorpej 1659:
1660: simple_lock(&pc->pc_slock);
1661:
1662: if ((pcg = pc->pc_allocfrom) == NULL) {
1.61 chs 1663: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.43 thorpej 1664: if (pcg->pcg_avail != 0) {
1665: pc->pc_allocfrom = pcg;
1666: goto have_group;
1667: }
1668: }
1669:
1670: /*
1671: * No groups with any available objects. Allocate
1672: * a new object, construct it, and return it to
1673: * the caller. We will allocate a group, if necessary,
1674: * when the object is freed back to the cache.
1675: */
1.48 thorpej 1676: pc->pc_misses++;
1.43 thorpej 1677: simple_unlock(&pc->pc_slock);
1678: object = pool_get(pc->pc_pool, flags);
1679: if (object != NULL && pc->pc_ctor != NULL) {
1680: if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) {
1681: pool_put(pc->pc_pool, object);
1682: return (NULL);
1683: }
1684: }
1685: return (object);
1686: }
1687:
1688: have_group:
1.48 thorpej 1689: pc->pc_hits++;
1690: pc->pc_nitems--;
1.43 thorpej 1691: object = pcg_get(pcg);
1692:
1693: if (pcg->pcg_avail == 0)
1694: pc->pc_allocfrom = NULL;
1.45 thorpej 1695:
1.43 thorpej 1696: simple_unlock(&pc->pc_slock);
1697:
1698: return (object);
1699: }
1700:
1701: /*
1702: * pool_cache_put:
1703: *
1704: * Put an object back to the pool cache.
1705: */
1706: void
1707: pool_cache_put(struct pool_cache *pc, void *object)
1708: {
1709: struct pool_cache_group *pcg;
1.60 thorpej 1710: int s;
1.43 thorpej 1711:
1712: simple_lock(&pc->pc_slock);
1713:
1714: if ((pcg = pc->pc_freeto) == NULL) {
1.61 chs 1715: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.43 thorpej 1716: if (pcg->pcg_avail != PCG_NOBJECTS) {
1717: pc->pc_freeto = pcg;
1718: goto have_group;
1719: }
1720: }
1721:
1722: /*
1723: * No empty groups to free the object to. Attempt to
1.47 thorpej 1724: * allocate one.
1.43 thorpej 1725: */
1.47 thorpej 1726: simple_unlock(&pc->pc_slock);
1.60 thorpej 1727: s = splvm();
1.43 thorpej 1728: pcg = pool_get(&pcgpool, PR_NOWAIT);
1.60 thorpej 1729: splx(s);
1.43 thorpej 1730: if (pcg != NULL) {
1731: memset(pcg, 0, sizeof(*pcg));
1.47 thorpej 1732: simple_lock(&pc->pc_slock);
1.48 thorpej 1733: pc->pc_ngroups++;
1.43 thorpej 1734: TAILQ_INSERT_TAIL(&pc->pc_grouplist, pcg, pcg_list);
1.47 thorpej 1735: if (pc->pc_freeto == NULL)
1736: pc->pc_freeto = pcg;
1.43 thorpej 1737: goto have_group;
1738: }
1739:
1740: /*
1741: * Unable to allocate a cache group; destruct the object
1742: * and free it back to the pool.
1743: */
1.51 thorpej 1744: pool_cache_destruct_object(pc, object);
1.43 thorpej 1745: return;
1746: }
1747:
1748: have_group:
1.48 thorpej 1749: pc->pc_nitems++;
1.43 thorpej 1750: pcg_put(pcg, object);
1751:
1752: if (pcg->pcg_avail == PCG_NOBJECTS)
1753: pc->pc_freeto = NULL;
1754:
1755: simple_unlock(&pc->pc_slock);
1.51 thorpej 1756: }
1757:
1758: /*
1759: * pool_cache_destruct_object:
1760: *
1761: * Force destruction of an object and its release back into
1762: * the pool.
1763: */
1764: void
1765: pool_cache_destruct_object(struct pool_cache *pc, void *object)
1766: {
1767:
1768: if (pc->pc_dtor != NULL)
1769: (*pc->pc_dtor)(pc->pc_arg, object);
1770: pool_put(pc->pc_pool, object);
1.43 thorpej 1771: }
1772:
1773: /*
1774: * pool_cache_do_invalidate:
1775: *
1776: * This internal function implements pool_cache_invalidate() and
1777: * pool_cache_reclaim().
1778: */
1779: static void
1780: pool_cache_do_invalidate(struct pool_cache *pc, int free_groups,
1.56 sommerfe 1781: void (*putit)(struct pool *, void *))
1.43 thorpej 1782: {
1783: struct pool_cache_group *pcg, *npcg;
1784: void *object;
1.60 thorpej 1785: int s;
1.43 thorpej 1786:
1787: for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
1788: pcg = npcg) {
1789: npcg = TAILQ_NEXT(pcg, pcg_list);
1790: while (pcg->pcg_avail != 0) {
1.48 thorpej 1791: pc->pc_nitems--;
1.43 thorpej 1792: object = pcg_get(pcg);
1.45 thorpej 1793: if (pcg->pcg_avail == 0 && pc->pc_allocfrom == pcg)
1794: pc->pc_allocfrom = NULL;
1.43 thorpej 1795: if (pc->pc_dtor != NULL)
1796: (*pc->pc_dtor)(pc->pc_arg, object);
1.56 sommerfe 1797: (*putit)(pc->pc_pool, object);
1.43 thorpej 1798: }
1799: if (free_groups) {
1.48 thorpej 1800: pc->pc_ngroups--;
1.43 thorpej 1801: TAILQ_REMOVE(&pc->pc_grouplist, pcg, pcg_list);
1.46 thorpej 1802: if (pc->pc_freeto == pcg)
1803: pc->pc_freeto = NULL;
1.60 thorpej 1804: s = splvm();
1.43 thorpej 1805: pool_put(&pcgpool, pcg);
1.60 thorpej 1806: splx(s);
1.43 thorpej 1807: }
1808: }
1809: }
1810:
1811: /*
1812: * pool_cache_invalidate:
1813: *
1814: * Invalidate a pool cache (destruct and release all of the
1815: * cached objects).
1816: */
1817: void
1818: pool_cache_invalidate(struct pool_cache *pc)
1819: {
1820:
1821: simple_lock(&pc->pc_slock);
1.56 sommerfe 1822: pool_cache_do_invalidate(pc, 0, pool_put);
1.43 thorpej 1823: simple_unlock(&pc->pc_slock);
1824: }
1825:
1826: /*
1827: * pool_cache_reclaim:
1828: *
1829: * Reclaim a pool cache for pool_reclaim().
1830: */
1831: static void
1832: pool_cache_reclaim(struct pool_cache *pc)
1833: {
1834:
1.47 thorpej 1835: simple_lock(&pc->pc_slock);
1.43 thorpej 1836: pool_cache_do_invalidate(pc, 1, pool_do_put);
1837: simple_unlock(&pc->pc_slock);
1.3 pk 1838: }
1.66 thorpej 1839:
1840: /*
1841: * Pool backend allocators.
1842: *
1843: * Each pool has a backend allocator that handles allocation, deallocation,
1844: * and any additional draining that might be needed.
1845: *
1846: * We provide two standard allocators:
1847: *
1848: * pool_allocator_kmem - the default when no allocator is specified
1849: *
1850: * pool_allocator_nointr - used for pools that will not be accessed
1851: * in interrupt context.
1852: */
1853: void *pool_page_alloc(struct pool *, int);
1854: void pool_page_free(struct pool *, void *);
1855:
1856: struct pool_allocator pool_allocator_kmem = {
1857: pool_page_alloc, pool_page_free, 0,
1858: };
1859:
1860: void *pool_page_alloc_nointr(struct pool *, int);
1861: void pool_page_free_nointr(struct pool *, void *);
1862:
1863: struct pool_allocator pool_allocator_nointr = {
1864: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1865: };
1866:
1867: #ifdef POOL_SUBPAGE
1868: void *pool_subpage_alloc(struct pool *, int);
1869: void pool_subpage_free(struct pool *, void *);
1870:
1871: struct pool_allocator pool_allocator_kmem_subpage = {
1872: pool_subpage_alloc, pool_subpage_free, 0,
1873: };
1874: #endif /* POOL_SUBPAGE */
1875:
1876: /*
1877: * We have at least three different resources for the same allocation and
1878: * each resource can be depleted. First, we have the ready elements in the
1879: * pool. Then we have the resource (typically a vm_map) for this allocator.
1880: * Finally, we have physical memory. Waiting for any of these can be
1881: * unnecessary when any other is freed, but the kernel doesn't support
1882: * sleeping on multiple wait channels, so we have to employ another strategy.
1883: *
1884: * The caller sleeps on the pool (so that it can be awakened when an item
1885: * is returned to the pool), but we set PA_WANT on the allocator. When a
1886: * page is returned to the allocator and PA_WANT is set, pool_allocator_free
1887: * will wake up all sleeping pools belonging to this allocator.
1888: *
1889: * XXX Thundering herd.
1890: */
1891: void *
1892: pool_allocator_alloc(struct pool *org, int flags)
1893: {
1894: struct pool_allocator *pa = org->pr_alloc;
1895: struct pool *pp, *start;
1896: int s, freed;
1897: void *res;
1898:
1899: do {
1900: if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
1901: return (res);
1902: if ((flags & PR_WAITOK) == 0)
1903: break;
1904:
1905: /*
1906: * Drain all pools, except "org", that use this
1907: * allocator. We do this to reclaim VA space.
1908: * pa_alloc is responsible for waiting for
1909: * physical memory.
1910: *
1911: * XXX We risk looping forever if start if someone
1912: * calls pool_destroy on "start". But there is no
1913: * other way to have potentially sleeping pool_reclaim,
1914: * non-sleeping locks on pool_allocator, and some
1915: * stirring of drained pools in the allocator.
1916: */
1917: freed = 0;
1918:
1919: s = splvm();
1920: simple_lock(&pa->pa_slock);
1921: pp = start = TAILQ_FIRST(&pa->pa_list);
1922: do {
1923: TAILQ_REMOVE(&pa->pa_list, pp, pr_alloc_list);
1924: TAILQ_INSERT_TAIL(&pa->pa_list, pp, pr_alloc_list);
1925: if (pp == org)
1926: continue;
1927: simple_unlock(&pa->pa_list);
1928: freed = pool_reclaim(pp);
1929: simple_lock(&pa->pa_list);
1930: } while ((pp = TAILQ_FIRST(&pa->pa_list)) != start &&
1931: freed == 0);
1932:
1933: if (freed == 0) {
1934: /*
1935: * We set PA_WANT here, the caller will most likely
1936: * sleep waiting for pages (if not, this won't hurt
1937: * that much), and there is no way to set this in
1938: * the caller without violating locking order.
1939: */
1940: pa->pa_flags |= PA_WANT;
1941: }
1942: simple_unlock(&pa->pa_slock);
1943: splx(s);
1944: } while (freed);
1945: return (NULL);
1946: }
1947:
1948: void
1949: pool_allocator_free(struct pool *pp, void *v)
1950: {
1951: struct pool_allocator *pa = pp->pr_alloc;
1952: int s;
1953:
1954: (*pa->pa_free)(pp, v);
1955:
1956: s = splvm();
1957: simple_lock(&pa->pa_slock);
1958: if ((pa->pa_flags & PA_WANT) == 0) {
1959: simple_unlock(&pa->pa_slock);
1960: splx(s);
1961: return;
1962: }
1963:
1964: TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
1965: simple_lock(&pp->pr_slock);
1966: if ((pp->pr_flags & PR_WANTED) != 0) {
1967: pp->pr_flags &= ~PR_WANTED;
1968: wakeup(pp);
1969: }
1970: }
1971: pa->pa_flags &= ~PA_WANT;
1972: simple_unlock(&pa->pa_slock);
1973: splx(s);
1974: }
1975:
1976: void *
1977: pool_page_alloc(struct pool *pp, int flags)
1978: {
1979: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
1980:
1981: return ((void *) uvm_km_alloc_poolpage(waitok));
1982: }
1983:
1984: void
1985: pool_page_free(struct pool *pp, void *v)
1986: {
1987:
1988: uvm_km_free_poolpage((vaddr_t) v);
1989: }
1990:
1991: #ifdef POOL_SUBPAGE
1992: /* Sub-page allocator, for machines with large hardware pages. */
1993: void *
1994: pool_subpage_alloc(struct pool *pp, int flags)
1995: {
1996:
1997: return (pool_get(&psppool, flags));
1998: }
1999:
2000: void
2001: pool_subpage_free(struct pool *pp, void *v)
2002: {
2003:
2004: pool_put(&psppool, v);
2005: }
2006:
2007: /* We don't provide a real nointr allocator. Maybe later. */
2008: void *
2009: pool_page_alloc_nointr(struct pool *pp, int flags)
2010: {
2011:
2012: return (pool_subpage_alloc(pp, flags));
2013: }
2014:
2015: void
2016: pool_page_free_nointr(struct pool *pp, void *v)
2017: {
2018:
2019: pool_subpage_free(pp, v);
2020: }
2021: #else
2022: void *
2023: pool_page_alloc_nointr(struct pool *pp, int flags)
2024: {
2025: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2026:
2027: return ((void *) uvm_km_alloc_poolpage1(kernel_map,
2028: uvm.kernel_object, waitok));
2029: }
2030:
2031: void
2032: pool_page_free_nointr(struct pool *pp, void *v)
2033: {
2034:
2035: uvm_km_free_poolpage1(kernel_map, (vaddr_t) v);
2036: }
2037: #endif /* POOL_SUBPAGE */
CVSweb <webmaster@jp.NetBSD.org>