Annotation of src/sys/kern/subr_pool.c, Revision 1.74.2.1
1.74.2.1! thorpej 1: /* $NetBSD: subr_pool.c,v 1.74 2002/03/09 18:06:55 thorpej Exp $ */
1.1 pk 2:
3: /*-
1.74.2.1! thorpej 4: * Copyright (c) 1997, 1999, 2000, 2002 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>
1.74.2.1! thorpej 41: __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.74 2002/03/09 18:06:55 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;
1.68 thorpej 459: pp->pr_drain_hook = NULL;
460: pp->pr_drain_hook_arg = NULL;
1.3 pk 461:
462: /*
463: * Decide whether to put the page header off page to avoid
464: * wasting too large a part of the page. Off-page page headers
465: * go on a hash table, so we can match a returned item
466: * with its header based on the page address.
467: * We use 1/16 of the page size as the threshold (XXX: tune)
468: */
1.66 thorpej 469: if (pp->pr_size < palloc->pa_pagesz/16) {
1.3 pk 470: /* Use the end of the page for the page header */
1.20 thorpej 471: pp->pr_roflags |= PR_PHINPAGE;
1.66 thorpej 472: pp->pr_phoffset = off = palloc->pa_pagesz -
473: ALIGN(sizeof(struct pool_item_header));
1.2 pk 474: } else {
1.3 pk 475: /* The page header will be taken from our page header pool */
476: pp->pr_phoffset = 0;
1.66 thorpej 477: off = palloc->pa_pagesz;
1.16 briggs 478: for (i = 0; i < PR_HASHTABSIZE; i++) {
479: LIST_INIT(&pp->pr_hashtab[i]);
480: }
1.2 pk 481: }
1.1 pk 482:
1.3 pk 483: /*
484: * Alignment is to take place at `ioff' within the item. This means
485: * we must reserve up to `align - 1' bytes on the page to allow
486: * appropriate positioning of each item.
487: *
488: * Silently enforce `0 <= ioff < align'.
489: */
490: pp->pr_itemoffset = ioff = ioff % align;
491: pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
1.43 thorpej 492: KASSERT(pp->pr_itemsperpage != 0);
1.3 pk 493:
494: /*
495: * Use the slack between the chunks and the page header
496: * for "cache coloring".
497: */
498: slack = off - pp->pr_itemsperpage * pp->pr_size;
499: pp->pr_maxcolor = (slack / align) * align;
500: pp->pr_curcolor = 0;
501:
502: pp->pr_nget = 0;
503: pp->pr_nfail = 0;
504: pp->pr_nput = 0;
505: pp->pr_npagealloc = 0;
506: pp->pr_npagefree = 0;
1.1 pk 507: pp->pr_hiwat = 0;
1.8 thorpej 508: pp->pr_nidle = 0;
1.3 pk 509:
1.59 thorpej 510: #ifdef POOL_DIAGNOSTIC
1.25 thorpej 511: if (flags & PR_LOGGING) {
512: if (kmem_map == NULL ||
513: (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log),
514: M_TEMP, M_NOWAIT)) == NULL)
1.20 thorpej 515: pp->pr_roflags &= ~PR_LOGGING;
1.3 pk 516: pp->pr_curlogentry = 0;
517: pp->pr_logsize = pool_logsize;
518: }
1.59 thorpej 519: #endif
1.25 thorpej 520:
521: pp->pr_entered_file = NULL;
522: pp->pr_entered_line = 0;
1.3 pk 523:
1.21 thorpej 524: simple_lock_init(&pp->pr_slock);
1.1 pk 525:
1.3 pk 526: /*
1.43 thorpej 527: * Initialize private page header pool and cache magazine pool if we
528: * haven't done so yet.
1.23 thorpej 529: * XXX LOCKING.
1.3 pk 530: */
531: if (phpool.pr_size == 0) {
1.62 bjh21 532: #ifdef POOL_SUBPAGE
533: pool_init(&phpool, sizeof(struct pool_item_header), 0, 0, 0,
1.66 thorpej 534: "phpool", &pool_allocator_kmem);
1.62 bjh21 535: pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
1.66 thorpej 536: PR_RECURSIVE, "psppool", &pool_allocator_kmem);
1.62 bjh21 537: #else
1.3 pk 538: pool_init(&phpool, sizeof(struct pool_item_header), 0, 0,
1.66 thorpej 539: 0, "phpool", NULL);
1.62 bjh21 540: #endif
1.43 thorpej 541: pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
1.66 thorpej 542: 0, "pcgpool", NULL);
1.1 pk 543: }
544:
1.23 thorpej 545: /* Insert into the list of all pools. */
546: simple_lock(&pool_head_slock);
547: TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist);
548: simple_unlock(&pool_head_slock);
1.66 thorpej 549:
550: /* Insert this into the list of pools using this allocator. */
551: simple_lock(&palloc->pa_slock);
552: TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
553: simple_unlock(&palloc->pa_slock);
1.1 pk 554: }
555:
556: /*
557: * De-commision a pool resource.
558: */
559: void
1.42 thorpej 560: pool_destroy(struct pool *pp)
1.1 pk 561: {
1.3 pk 562: struct pool_item_header *ph;
1.43 thorpej 563: struct pool_cache *pc;
564:
1.66 thorpej 565: /* Locking order: pool_allocator -> pool */
566: simple_lock(&pp->pr_alloc->pa_slock);
567: TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
568: simple_unlock(&pp->pr_alloc->pa_slock);
569:
1.43 thorpej 570: /* Destroy all caches for this pool. */
571: while ((pc = TAILQ_FIRST(&pp->pr_cachelist)) != NULL)
572: pool_cache_destroy(pc);
1.3 pk 573:
574: #ifdef DIAGNOSTIC
1.20 thorpej 575: if (pp->pr_nout != 0) {
1.25 thorpej 576: pr_printlog(pp, NULL, printf);
1.20 thorpej 577: panic("pool_destroy: pool busy: still out: %u\n",
578: pp->pr_nout);
1.3 pk 579: }
580: #endif
1.1 pk 581:
1.3 pk 582: /* Remove all pages */
1.70 thorpej 583: while ((ph = TAILQ_FIRST(&pp->pr_pagelist)) != NULL)
584: pr_rmpage(pp, ph, NULL);
1.3 pk 585:
586: /* Remove from global pool list */
1.23 thorpej 587: simple_lock(&pool_head_slock);
1.3 pk 588: TAILQ_REMOVE(&pool_head, pp, pr_poollist);
1.61 chs 589: if (drainpp == pp) {
590: drainpp = NULL;
591: }
1.23 thorpej 592: simple_unlock(&pool_head_slock);
1.3 pk 593:
1.59 thorpej 594: #ifdef POOL_DIAGNOSTIC
1.20 thorpej 595: if ((pp->pr_roflags & PR_LOGGING) != 0)
1.3 pk 596: free(pp->pr_log, M_TEMP);
1.59 thorpej 597: #endif
1.1 pk 598: }
599:
1.68 thorpej 600: void
601: pool_set_drain_hook(struct pool *pp, void (*fn)(void *, int), void *arg)
602: {
603:
604: /* XXX no locking -- must be used just after pool_init() */
605: #ifdef DIAGNOSTIC
606: if (pp->pr_drain_hook != NULL)
607: panic("pool_set_drain_hook(%s): already set", pp->pr_wchan);
608: #endif
609: pp->pr_drain_hook = fn;
610: pp->pr_drain_hook_arg = arg;
611: }
612:
1.55 thorpej 613: static __inline struct pool_item_header *
614: pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags)
615: {
616: struct pool_item_header *ph;
617: int s;
618:
619: LOCK_ASSERT(simple_lock_held(&pp->pr_slock) == 0);
620:
621: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
622: ph = (struct pool_item_header *) (storage + pp->pr_phoffset);
623: else {
624: s = splhigh();
625: ph = pool_get(&phpool, flags);
626: splx(s);
627: }
628:
629: return (ph);
630: }
1.1 pk 631:
632: /*
1.3 pk 633: * Grab an item from the pool; must be called at appropriate spl level
1.1 pk 634: */
1.3 pk 635: void *
1.59 thorpej 636: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 637: _pool_get(struct pool *pp, int flags, const char *file, long line)
1.56 sommerfe 638: #else
639: pool_get(struct pool *pp, int flags)
640: #endif
1.1 pk 641: {
642: struct pool_item *pi;
1.3 pk 643: struct pool_item_header *ph;
1.55 thorpej 644: void *v;
1.1 pk 645:
1.2 pk 646: #ifdef DIAGNOSTIC
1.37 sommerfe 647: if (__predict_false(curproc == NULL && doing_shutdown == 0 &&
648: (flags & PR_WAITOK) != 0))
1.3 pk 649: panic("pool_get: must have NOWAIT");
1.58 thorpej 650:
651: #ifdef LOCKDEBUG
652: if (flags & PR_WAITOK)
653: simple_lock_only_held(NULL, "pool_get(PR_WAITOK)");
1.56 sommerfe 654: #endif
1.58 thorpej 655: #endif /* DIAGNOSTIC */
1.1 pk 656:
1.21 thorpej 657: simple_lock(&pp->pr_slock);
1.25 thorpej 658: pr_enter(pp, file, line);
1.20 thorpej 659:
660: startover:
661: /*
662: * Check to see if we've reached the hard limit. If we have,
663: * and we can wait, then wait until an item has been returned to
664: * the pool.
665: */
666: #ifdef DIAGNOSTIC
1.34 thorpej 667: if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
1.25 thorpej 668: pr_leave(pp);
1.21 thorpej 669: simple_unlock(&pp->pr_slock);
1.20 thorpej 670: panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
671: }
672: #endif
1.34 thorpej 673: if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
1.68 thorpej 674: if (pp->pr_drain_hook != NULL) {
675: /*
676: * Since the drain hook is going to free things
677: * back to the pool, unlock, call the hook, re-lock,
678: * and check the hardlimit condition again.
679: */
680: pr_leave(pp);
681: simple_unlock(&pp->pr_slock);
682: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
683: simple_lock(&pp->pr_slock);
684: pr_enter(pp, file, line);
685: if (pp->pr_nout < pp->pr_hardlimit)
686: goto startover;
687: }
688:
1.29 sommerfe 689: if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
1.20 thorpej 690: /*
691: * XXX: A warning isn't logged in this case. Should
692: * it be?
693: */
694: pp->pr_flags |= PR_WANTED;
1.25 thorpej 695: pr_leave(pp);
1.40 sommerfe 696: ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
1.25 thorpej 697: pr_enter(pp, file, line);
1.20 thorpej 698: goto startover;
699: }
1.31 thorpej 700:
701: /*
702: * Log a message that the hard limit has been hit.
703: */
704: if (pp->pr_hardlimit_warning != NULL &&
705: ratecheck(&pp->pr_hardlimit_warning_last,
706: &pp->pr_hardlimit_ratecap))
707: log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
1.21 thorpej 708:
709: pp->pr_nfail++;
710:
1.25 thorpej 711: pr_leave(pp);
1.21 thorpej 712: simple_unlock(&pp->pr_slock);
1.20 thorpej 713: return (NULL);
714: }
715:
1.3 pk 716: /*
717: * The convention we use is that if `curpage' is not NULL, then
718: * it points at a non-empty bucket. In particular, `curpage'
719: * never points at a page header which has PR_PHINPAGE set and
720: * has no items in its bucket.
721: */
1.20 thorpej 722: if ((ph = pp->pr_curpage) == NULL) {
723: #ifdef DIAGNOSTIC
724: if (pp->pr_nitems != 0) {
1.21 thorpej 725: simple_unlock(&pp->pr_slock);
1.20 thorpej 726: printf("pool_get: %s: curpage NULL, nitems %u\n",
727: pp->pr_wchan, pp->pr_nitems);
728: panic("pool_get: nitems inconsistent\n");
729: }
730: #endif
731:
1.21 thorpej 732: /*
733: * Call the back-end page allocator for more memory.
734: * Release the pool lock, as the back-end page allocator
735: * may block.
736: */
1.25 thorpej 737: pr_leave(pp);
1.21 thorpej 738: simple_unlock(&pp->pr_slock);
1.66 thorpej 739: v = pool_allocator_alloc(pp, flags);
1.55 thorpej 740: if (__predict_true(v != NULL))
741: ph = pool_alloc_item_header(pp, v, flags);
1.21 thorpej 742: simple_lock(&pp->pr_slock);
1.25 thorpej 743: pr_enter(pp, file, line);
1.15 pk 744:
1.55 thorpej 745: if (__predict_false(v == NULL || ph == NULL)) {
746: if (v != NULL)
1.66 thorpej 747: pool_allocator_free(pp, v);
1.55 thorpej 748:
1.21 thorpej 749: /*
1.55 thorpej 750: * We were unable to allocate a page or item
751: * header, but we released the lock during
752: * allocation, so perhaps items were freed
753: * back to the pool. Check for this case.
1.21 thorpej 754: */
755: if (pp->pr_curpage != NULL)
756: goto startover;
1.15 pk 757:
1.3 pk 758: if ((flags & PR_WAITOK) == 0) {
759: pp->pr_nfail++;
1.25 thorpej 760: pr_leave(pp);
1.21 thorpej 761: simple_unlock(&pp->pr_slock);
1.1 pk 762: return (NULL);
1.3 pk 763: }
764:
1.15 pk 765: /*
766: * Wait for items to be returned to this pool.
1.21 thorpej 767: *
1.20 thorpej 768: * XXX: maybe we should wake up once a second and
769: * try again?
1.15 pk 770: */
1.1 pk 771: pp->pr_flags |= PR_WANTED;
1.66 thorpej 772: /* PA_WANTED is already set on the allocator. */
1.25 thorpej 773: pr_leave(pp);
1.40 sommerfe 774: ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
1.25 thorpej 775: pr_enter(pp, file, line);
1.20 thorpej 776: goto startover;
1.1 pk 777: }
1.3 pk 778:
1.15 pk 779: /* We have more memory; add it to the pool */
1.55 thorpej 780: pool_prime_page(pp, v, ph);
1.15 pk 781: pp->pr_npagealloc++;
782:
1.20 thorpej 783: /* Start the allocation process over. */
784: goto startover;
1.3 pk 785: }
786:
1.34 thorpej 787: if (__predict_false((v = pi = TAILQ_FIRST(&ph->ph_itemlist)) == NULL)) {
1.25 thorpej 788: pr_leave(pp);
1.21 thorpej 789: simple_unlock(&pp->pr_slock);
1.3 pk 790: panic("pool_get: %s: page empty", pp->pr_wchan);
1.21 thorpej 791: }
1.20 thorpej 792: #ifdef DIAGNOSTIC
1.34 thorpej 793: if (__predict_false(pp->pr_nitems == 0)) {
1.25 thorpej 794: pr_leave(pp);
1.21 thorpej 795: simple_unlock(&pp->pr_slock);
1.20 thorpej 796: printf("pool_get: %s: items on itemlist, nitems %u\n",
797: pp->pr_wchan, pp->pr_nitems);
798: panic("pool_get: nitems inconsistent\n");
799: }
1.65 enami 800: #endif
1.56 sommerfe 801:
1.65 enami 802: #ifdef POOL_DIAGNOSTIC
1.3 pk 803: pr_log(pp, v, PRLOG_GET, file, line);
1.65 enami 804: #endif
1.3 pk 805:
1.65 enami 806: #ifdef DIAGNOSTIC
1.34 thorpej 807: if (__predict_false(pi->pi_magic != PI_MAGIC)) {
1.25 thorpej 808: pr_printlog(pp, pi, printf);
1.3 pk 809: panic("pool_get(%s): free list modified: magic=%x; page %p;"
810: " item addr %p\n",
811: pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
812: }
813: #endif
814:
815: /*
816: * Remove from item list.
817: */
818: TAILQ_REMOVE(&ph->ph_itemlist, pi, pi_list);
1.20 thorpej 819: pp->pr_nitems--;
820: pp->pr_nout++;
1.6 thorpej 821: if (ph->ph_nmissing == 0) {
822: #ifdef DIAGNOSTIC
1.34 thorpej 823: if (__predict_false(pp->pr_nidle == 0))
1.6 thorpej 824: panic("pool_get: nidle inconsistent");
825: #endif
826: pp->pr_nidle--;
827: }
1.3 pk 828: ph->ph_nmissing++;
829: if (TAILQ_FIRST(&ph->ph_itemlist) == NULL) {
1.21 thorpej 830: #ifdef DIAGNOSTIC
1.34 thorpej 831: if (__predict_false(ph->ph_nmissing != pp->pr_itemsperpage)) {
1.25 thorpej 832: pr_leave(pp);
1.21 thorpej 833: simple_unlock(&pp->pr_slock);
834: panic("pool_get: %s: nmissing inconsistent",
835: pp->pr_wchan);
836: }
837: #endif
1.3 pk 838: /*
839: * Find a new non-empty page header, if any.
840: * Start search from the page head, to increase
841: * the chance for "high water" pages to be freed.
842: *
1.21 thorpej 843: * Migrate empty pages to the end of the list. This
844: * will speed the update of curpage as pages become
845: * idle. Empty pages intermingled with idle pages
846: * is no big deal. As soon as a page becomes un-empty,
847: * it will move back to the head of the list.
1.3 pk 848: */
849: TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
1.21 thorpej 850: TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist);
1.61 chs 851: TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist)
1.3 pk 852: if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
853: break;
854:
855: pp->pr_curpage = ph;
1.1 pk 856: }
1.3 pk 857:
858: pp->pr_nget++;
1.20 thorpej 859:
860: /*
861: * If we have a low water mark and we are now below that low
862: * water mark, add more items to the pool.
863: */
1.53 thorpej 864: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 865: /*
866: * XXX: Should we log a warning? Should we set up a timeout
867: * to try again in a second or so? The latter could break
868: * a caller's assumptions about interrupt protection, etc.
869: */
870: }
871:
1.25 thorpej 872: pr_leave(pp);
1.21 thorpej 873: simple_unlock(&pp->pr_slock);
1.74.2.1! thorpej 874: KASSERT((((vaddr_t)v + pp->pr_itemoffset) & (pp->pr_align - 1)) == 0);
1.1 pk 875: return (v);
876: }
877:
878: /*
1.43 thorpej 879: * Internal version of pool_put(). Pool is already locked/entered.
1.1 pk 880: */
1.43 thorpej 881: static void
1.56 sommerfe 882: pool_do_put(struct pool *pp, void *v)
1.1 pk 883: {
884: struct pool_item *pi = v;
1.3 pk 885: struct pool_item_header *ph;
886: caddr_t page;
1.21 thorpej 887: int s;
1.3 pk 888:
1.61 chs 889: LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
890:
1.66 thorpej 891: page = (caddr_t)((u_long)v & pp->pr_alloc->pa_pagemask);
1.1 pk 892:
1.30 thorpej 893: #ifdef DIAGNOSTIC
1.34 thorpej 894: if (__predict_false(pp->pr_nout == 0)) {
1.30 thorpej 895: printf("pool %s: putting with none out\n",
896: pp->pr_wchan);
897: panic("pool_put");
898: }
899: #endif
1.3 pk 900:
1.34 thorpej 901: if (__predict_false((ph = pr_find_pagehead(pp, page)) == NULL)) {
1.25 thorpej 902: pr_printlog(pp, NULL, printf);
1.3 pk 903: panic("pool_put: %s: page header missing", pp->pr_wchan);
904: }
1.28 thorpej 905:
906: #ifdef LOCKDEBUG
907: /*
908: * Check if we're freeing a locked simple lock.
909: */
910: simple_lock_freecheck((caddr_t)pi, ((caddr_t)pi) + pp->pr_size);
911: #endif
1.3 pk 912:
913: /*
914: * Return to item list.
915: */
1.2 pk 916: #ifdef DIAGNOSTIC
1.3 pk 917: pi->pi_magic = PI_MAGIC;
918: #endif
1.32 chs 919: #ifdef DEBUG
920: {
921: int i, *ip = v;
922:
923: for (i = 0; i < pp->pr_size / sizeof(int); i++) {
924: *ip++ = PI_MAGIC;
925: }
926: }
927: #endif
928:
1.3 pk 929: TAILQ_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
930: ph->ph_nmissing--;
931: pp->pr_nput++;
1.20 thorpej 932: pp->pr_nitems++;
933: pp->pr_nout--;
1.3 pk 934:
935: /* Cancel "pool empty" condition if it exists */
936: if (pp->pr_curpage == NULL)
937: pp->pr_curpage = ph;
938:
939: if (pp->pr_flags & PR_WANTED) {
940: pp->pr_flags &= ~PR_WANTED;
1.15 pk 941: if (ph->ph_nmissing == 0)
942: pp->pr_nidle++;
1.3 pk 943: wakeup((caddr_t)pp);
944: return;
945: }
946:
947: /*
1.21 thorpej 948: * If this page is now complete, do one of two things:
949: *
950: * (1) If we have more pages than the page high water
951: * mark, free the page back to the system.
952: *
953: * (2) Move it to the end of the page list, so that
954: * we minimize our chances of fragmenting the
955: * pool. Idle pages migrate to the end (along with
956: * completely empty pages, so that we find un-empty
957: * pages more quickly when we update curpage) of the
958: * list so they can be more easily swept up by
959: * the pagedaemon when pages are scarce.
1.3 pk 960: */
961: if (ph->ph_nmissing == 0) {
1.6 thorpej 962: pp->pr_nidle++;
1.71 thorpej 963: if (pp->pr_npages > pp->pr_maxpages ||
964: (pp->pr_alloc->pa_flags & PA_WANT) != 0) {
1.61 chs 965: pr_rmpage(pp, ph, NULL);
1.3 pk 966: } else {
967: TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
968: TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist);
969:
1.21 thorpej 970: /*
971: * Update the timestamp on the page. A page must
972: * be idle for some period of time before it can
973: * be reclaimed by the pagedaemon. This minimizes
974: * ping-pong'ing for memory.
975: */
976: s = splclock();
977: ph->ph_time = mono_time;
978: splx(s);
979:
980: /*
981: * Update the current page pointer. Just look for
982: * the first page with any free items.
983: *
984: * XXX: Maybe we want an option to look for the
985: * page with the fewest available items, to minimize
986: * fragmentation?
987: */
1.61 chs 988: TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist)
1.3 pk 989: if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
990: break;
1.1 pk 991:
1.3 pk 992: pp->pr_curpage = ph;
1.1 pk 993: }
994: }
1.21 thorpej 995: /*
996: * If the page has just become un-empty, move it to the head of
997: * the list, and make it the current page. The next allocation
998: * will get the item from this page, instead of further fragmenting
999: * the pool.
1000: */
1001: else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
1002: TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
1003: TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist);
1004: pp->pr_curpage = ph;
1005: }
1.43 thorpej 1006: }
1007:
1008: /*
1009: * Return resource to the pool; must be called at appropriate spl level
1010: */
1.59 thorpej 1011: #ifdef POOL_DIAGNOSTIC
1.43 thorpej 1012: void
1013: _pool_put(struct pool *pp, void *v, const char *file, long line)
1014: {
1015:
1016: simple_lock(&pp->pr_slock);
1017: pr_enter(pp, file, line);
1018:
1.56 sommerfe 1019: pr_log(pp, v, PRLOG_PUT, file, line);
1020:
1021: pool_do_put(pp, v);
1.21 thorpej 1022:
1.25 thorpej 1023: pr_leave(pp);
1.21 thorpej 1024: simple_unlock(&pp->pr_slock);
1.1 pk 1025: }
1.57 sommerfe 1026: #undef pool_put
1.59 thorpej 1027: #endif /* POOL_DIAGNOSTIC */
1.1 pk 1028:
1.56 sommerfe 1029: void
1030: pool_put(struct pool *pp, void *v)
1031: {
1032:
1033: simple_lock(&pp->pr_slock);
1034:
1035: pool_do_put(pp, v);
1036:
1037: simple_unlock(&pp->pr_slock);
1038: }
1.57 sommerfe 1039:
1.59 thorpej 1040: #ifdef POOL_DIAGNOSTIC
1.57 sommerfe 1041: #define pool_put(h, v) _pool_put((h), (v), __FILE__, __LINE__)
1.56 sommerfe 1042: #endif
1.74 thorpej 1043:
1044: /*
1045: * Add N items to the pool.
1046: */
1047: int
1048: pool_prime(struct pool *pp, int n)
1049: {
1050: struct pool_item_header *ph;
1051: caddr_t cp;
1052: int newpages, error = 0;
1053:
1054: simple_lock(&pp->pr_slock);
1055:
1056: newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1057:
1058: while (newpages-- > 0) {
1059: simple_unlock(&pp->pr_slock);
1060: cp = pool_allocator_alloc(pp, PR_NOWAIT);
1061: if (__predict_true(cp != NULL))
1062: ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1063: simple_lock(&pp->pr_slock);
1064:
1065: if (__predict_false(cp == NULL || ph == NULL)) {
1066: error = ENOMEM;
1067: if (cp != NULL)
1068: pool_allocator_free(pp, cp);
1069: break;
1070: }
1071:
1072: pool_prime_page(pp, cp, ph);
1073: pp->pr_npagealloc++;
1074: pp->pr_minpages++;
1075: }
1076:
1077: if (pp->pr_minpages >= pp->pr_maxpages)
1078: pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */
1079:
1080: simple_unlock(&pp->pr_slock);
1081: return (0);
1082: }
1.55 thorpej 1083:
1084: /*
1.3 pk 1085: * Add a page worth of items to the pool.
1.21 thorpej 1086: *
1087: * Note, we must be called with the pool descriptor LOCKED.
1.3 pk 1088: */
1.55 thorpej 1089: static void
1090: pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph)
1.3 pk 1091: {
1092: struct pool_item *pi;
1093: caddr_t cp = storage;
1.74.2.1! thorpej 1094: const unsigned int align = pp->pr_align;
! 1095: const unsigned int ioff = pp->pr_itemoffset;
! 1096: const unsigned int alignsize = roundup(pp->pr_size, align);
1.55 thorpej 1097: int n;
1.36 pk 1098:
1.66 thorpej 1099: #ifdef DIAGNOSTIC
1100: if (((u_long)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
1.36 pk 1101: panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1.66 thorpej 1102: #endif
1.3 pk 1103:
1.55 thorpej 1104: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1.3 pk 1105: LIST_INSERT_HEAD(&pp->pr_hashtab[PR_HASH_INDEX(pp, cp)],
1.55 thorpej 1106: ph, ph_hashlist);
1.3 pk 1107:
1108: /*
1109: * Insert page header.
1110: */
1111: TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist);
1112: TAILQ_INIT(&ph->ph_itemlist);
1113: ph->ph_page = storage;
1114: ph->ph_nmissing = 0;
1.21 thorpej 1115: memset(&ph->ph_time, 0, sizeof(ph->ph_time));
1.3 pk 1116:
1.6 thorpej 1117: pp->pr_nidle++;
1118:
1.3 pk 1119: /*
1120: * Color this page.
1121: */
1122: cp = (caddr_t)(cp + pp->pr_curcolor);
1123: if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
1124: pp->pr_curcolor = 0;
1125:
1126: /*
1127: * Adjust storage to apply aligment to `pr_itemoffset' in each item.
1128: */
1129: if (ioff != 0)
1130: cp = (caddr_t)(cp + (align - ioff));
1131:
1.74.2.1! thorpej 1132: KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
! 1133:
1.3 pk 1134: /*
1135: * Insert remaining chunks on the bucket list.
1136: */
1137: n = pp->pr_itemsperpage;
1.20 thorpej 1138: pp->pr_nitems += n;
1.3 pk 1139:
1140: while (n--) {
1141: pi = (struct pool_item *)cp;
1142:
1143: /* Insert on page list */
1144: TAILQ_INSERT_TAIL(&ph->ph_itemlist, pi, pi_list);
1145: #ifdef DIAGNOSTIC
1146: pi->pi_magic = PI_MAGIC;
1147: #endif
1.74.2.1! thorpej 1148: cp = (caddr_t)(cp + alignsize);
! 1149:
! 1150: KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
1.3 pk 1151: }
1152:
1153: /*
1154: * If the pool was depleted, point at the new page.
1155: */
1156: if (pp->pr_curpage == NULL)
1157: pp->pr_curpage = ph;
1158:
1159: if (++pp->pr_npages > pp->pr_hiwat)
1160: pp->pr_hiwat = pp->pr_npages;
1161: }
1162:
1.20 thorpej 1163: /*
1.52 thorpej 1164: * Used by pool_get() when nitems drops below the low water mark. This
1165: * is used to catch up nitmes with the low water mark.
1.20 thorpej 1166: *
1.21 thorpej 1167: * Note 1, we never wait for memory here, we let the caller decide what to do.
1.20 thorpej 1168: *
1.73 thorpej 1169: * Note 2, we must be called with the pool already locked, and we return
1.20 thorpej 1170: * with it locked.
1171: */
1172: static int
1.42 thorpej 1173: pool_catchup(struct pool *pp)
1.20 thorpej 1174: {
1.55 thorpej 1175: struct pool_item_header *ph;
1.20 thorpej 1176: caddr_t cp;
1177: int error = 0;
1178:
1.54 thorpej 1179: while (POOL_NEEDS_CATCHUP(pp)) {
1.20 thorpej 1180: /*
1.21 thorpej 1181: * Call the page back-end allocator for more memory.
1182: *
1183: * XXX: We never wait, so should we bother unlocking
1184: * the pool descriptor?
1.20 thorpej 1185: */
1.21 thorpej 1186: simple_unlock(&pp->pr_slock);
1.66 thorpej 1187: cp = pool_allocator_alloc(pp, PR_NOWAIT);
1.55 thorpej 1188: if (__predict_true(cp != NULL))
1189: ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1.21 thorpej 1190: simple_lock(&pp->pr_slock);
1.55 thorpej 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;
1195: break;
1196: }
1.55 thorpej 1197: pool_prime_page(pp, cp, ph);
1.26 thorpej 1198: pp->pr_npagealloc++;
1.20 thorpej 1199: }
1200:
1201: return (error);
1202: }
1203:
1.3 pk 1204: void
1.42 thorpej 1205: pool_setlowat(struct pool *pp, int n)
1.3 pk 1206: {
1.20 thorpej 1207: int error;
1.15 pk 1208:
1.21 thorpej 1209: simple_lock(&pp->pr_slock);
1210:
1.3 pk 1211: pp->pr_minitems = n;
1.15 pk 1212: pp->pr_minpages = (n == 0)
1213: ? 0
1.18 thorpej 1214: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.20 thorpej 1215:
1216: /* Make sure we're caught up with the newly-set low water mark. */
1.53 thorpej 1217: if (POOL_NEEDS_CATCHUP(pp) && (error = pool_catchup(pp) != 0)) {
1.20 thorpej 1218: /*
1219: * XXX: Should we log a warning? Should we set up a timeout
1220: * to try again in a second or so? The latter could break
1221: * a caller's assumptions about interrupt protection, etc.
1222: */
1223: }
1.21 thorpej 1224:
1225: simple_unlock(&pp->pr_slock);
1.3 pk 1226: }
1227:
1228: void
1.42 thorpej 1229: pool_sethiwat(struct pool *pp, int n)
1.3 pk 1230: {
1.15 pk 1231:
1.21 thorpej 1232: simple_lock(&pp->pr_slock);
1233:
1.15 pk 1234: pp->pr_maxpages = (n == 0)
1235: ? 0
1.18 thorpej 1236: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1237:
1238: simple_unlock(&pp->pr_slock);
1.3 pk 1239: }
1240:
1.20 thorpej 1241: void
1.42 thorpej 1242: pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1.20 thorpej 1243: {
1244:
1.21 thorpej 1245: simple_lock(&pp->pr_slock);
1.20 thorpej 1246:
1247: pp->pr_hardlimit = n;
1248: pp->pr_hardlimit_warning = warnmess;
1.31 thorpej 1249: pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1250: pp->pr_hardlimit_warning_last.tv_sec = 0;
1251: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.20 thorpej 1252:
1253: /*
1.21 thorpej 1254: * In-line version of pool_sethiwat(), because we don't want to
1255: * release the lock.
1.20 thorpej 1256: */
1257: pp->pr_maxpages = (n == 0)
1258: ? 0
1259: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1260:
1261: simple_unlock(&pp->pr_slock);
1.20 thorpej 1262: }
1.3 pk 1263:
1264: /*
1265: * Release all complete pages that have not been used recently.
1266: */
1.66 thorpej 1267: int
1.59 thorpej 1268: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 1269: _pool_reclaim(struct pool *pp, const char *file, long line)
1.56 sommerfe 1270: #else
1271: pool_reclaim(struct pool *pp)
1272: #endif
1.3 pk 1273: {
1274: struct pool_item_header *ph, *phnext;
1.43 thorpej 1275: struct pool_cache *pc;
1.21 thorpej 1276: struct timeval curtime;
1.61 chs 1277: struct pool_pagelist pq;
1.21 thorpej 1278: int s;
1.3 pk 1279:
1.68 thorpej 1280: if (pp->pr_drain_hook != NULL) {
1281: /*
1282: * The drain hook must be called with the pool unlocked.
1283: */
1284: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT);
1285: }
1286:
1.21 thorpej 1287: if (simple_lock_try(&pp->pr_slock) == 0)
1.66 thorpej 1288: return (0);
1.25 thorpej 1289: pr_enter(pp, file, line);
1.68 thorpej 1290:
1.61 chs 1291: TAILQ_INIT(&pq);
1.3 pk 1292:
1.43 thorpej 1293: /*
1294: * Reclaim items from the pool's caches.
1295: */
1.61 chs 1296: TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
1.43 thorpej 1297: pool_cache_reclaim(pc);
1298:
1.21 thorpej 1299: s = splclock();
1300: curtime = mono_time;
1301: splx(s);
1302:
1.3 pk 1303: for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; ph = phnext) {
1304: phnext = TAILQ_NEXT(ph, ph_pagelist);
1305:
1306: /* Check our minimum page claim */
1307: if (pp->pr_npages <= pp->pr_minpages)
1308: break;
1309:
1310: if (ph->ph_nmissing == 0) {
1311: struct timeval diff;
1312: timersub(&curtime, &ph->ph_time, &diff);
1313: if (diff.tv_sec < pool_inactive_time)
1314: continue;
1.21 thorpej 1315:
1316: /*
1317: * If freeing this page would put us below
1318: * the low water mark, stop now.
1319: */
1320: if ((pp->pr_nitems - pp->pr_itemsperpage) <
1321: pp->pr_minitems)
1322: break;
1323:
1.61 chs 1324: pr_rmpage(pp, ph, &pq);
1.3 pk 1325: }
1326: }
1327:
1.25 thorpej 1328: pr_leave(pp);
1.21 thorpej 1329: simple_unlock(&pp->pr_slock);
1.66 thorpej 1330: if (TAILQ_EMPTY(&pq))
1331: return (0);
1332:
1.61 chs 1333: while ((ph = TAILQ_FIRST(&pq)) != NULL) {
1334: TAILQ_REMOVE(&pq, ph, ph_pagelist);
1.66 thorpej 1335: pool_allocator_free(pp, ph->ph_page);
1.61 chs 1336: if (pp->pr_roflags & PR_PHINPAGE) {
1337: continue;
1338: }
1339: LIST_REMOVE(ph, ph_hashlist);
1340: s = splhigh();
1341: pool_put(&phpool, ph);
1342: splx(s);
1343: }
1.66 thorpej 1344:
1345: return (1);
1.3 pk 1346: }
1347:
1348: /*
1349: * Drain pools, one at a time.
1.21 thorpej 1350: *
1351: * Note, we must never be called from an interrupt context.
1.3 pk 1352: */
1353: void
1.42 thorpej 1354: pool_drain(void *arg)
1.3 pk 1355: {
1356: struct pool *pp;
1.23 thorpej 1357: int s;
1.3 pk 1358:
1.61 chs 1359: pp = NULL;
1.49 thorpej 1360: s = splvm();
1.23 thorpej 1361: simple_lock(&pool_head_slock);
1.61 chs 1362: if (drainpp == NULL) {
1363: drainpp = TAILQ_FIRST(&pool_head);
1364: }
1365: if (drainpp) {
1366: pp = drainpp;
1367: drainpp = TAILQ_NEXT(pp, pr_poollist);
1368: }
1369: simple_unlock(&pool_head_slock);
1.63 chs 1370: pool_reclaim(pp);
1.61 chs 1371: splx(s);
1.3 pk 1372: }
1373:
1374: /*
1375: * Diagnostic helpers.
1376: */
1377: void
1.42 thorpej 1378: pool_print(struct pool *pp, const char *modif)
1.21 thorpej 1379: {
1380: int s;
1381:
1.49 thorpej 1382: s = splvm();
1.25 thorpej 1383: if (simple_lock_try(&pp->pr_slock) == 0) {
1384: printf("pool %s is locked; try again later\n",
1385: pp->pr_wchan);
1386: splx(s);
1387: return;
1388: }
1389: pool_print1(pp, modif, printf);
1.21 thorpej 1390: simple_unlock(&pp->pr_slock);
1391: splx(s);
1392: }
1393:
1.25 thorpej 1394: void
1.42 thorpej 1395: pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.25 thorpej 1396: {
1397: int didlock = 0;
1398:
1399: if (pp == NULL) {
1400: (*pr)("Must specify a pool to print.\n");
1401: return;
1402: }
1403:
1404: /*
1405: * Called from DDB; interrupts should be blocked, and all
1406: * other processors should be paused. We can skip locking
1407: * the pool in this case.
1408: *
1409: * We do a simple_lock_try() just to print the lock
1410: * status, however.
1411: */
1412:
1413: if (simple_lock_try(&pp->pr_slock) == 0)
1414: (*pr)("WARNING: pool %s is locked\n", pp->pr_wchan);
1415: else
1416: didlock = 1;
1417:
1418: pool_print1(pp, modif, pr);
1419:
1420: if (didlock)
1421: simple_unlock(&pp->pr_slock);
1422: }
1423:
1.21 thorpej 1424: static void
1.42 thorpej 1425: pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.3 pk 1426: {
1.25 thorpej 1427: struct pool_item_header *ph;
1.44 thorpej 1428: struct pool_cache *pc;
1429: struct pool_cache_group *pcg;
1.25 thorpej 1430: #ifdef DIAGNOSTIC
1431: struct pool_item *pi;
1432: #endif
1.44 thorpej 1433: int i, print_log = 0, print_pagelist = 0, print_cache = 0;
1.25 thorpej 1434: char c;
1435:
1436: while ((c = *modif++) != '\0') {
1437: if (c == 'l')
1438: print_log = 1;
1439: if (c == 'p')
1440: print_pagelist = 1;
1.44 thorpej 1441: if (c == 'c')
1442: print_cache = 1;
1.25 thorpej 1443: modif++;
1444: }
1445:
1446: (*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1447: pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1448: pp->pr_roflags);
1.66 thorpej 1449: (*pr)("\talloc %p\n", pp->pr_alloc);
1.25 thorpej 1450: (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1451: pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1452: (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1453: pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1454:
1455: (*pr)("\n\tnget %lu, nfail %lu, nput %lu\n",
1456: pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1457: (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1458: pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1459:
1460: if (print_pagelist == 0)
1461: goto skip_pagelist;
1462:
1463: if ((ph = TAILQ_FIRST(&pp->pr_pagelist)) != NULL)
1464: (*pr)("\n\tpage list:\n");
1465: for (; ph != NULL; ph = TAILQ_NEXT(ph, ph_pagelist)) {
1466: (*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n",
1467: ph->ph_page, ph->ph_nmissing,
1468: (u_long)ph->ph_time.tv_sec,
1469: (u_long)ph->ph_time.tv_usec);
1470: #ifdef DIAGNOSTIC
1.61 chs 1471: TAILQ_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1.25 thorpej 1472: if (pi->pi_magic != PI_MAGIC) {
1473: (*pr)("\t\t\titem %p, magic 0x%x\n",
1474: pi, pi->pi_magic);
1475: }
1476: }
1477: #endif
1478: }
1479: if (pp->pr_curpage == NULL)
1480: (*pr)("\tno current page\n");
1481: else
1482: (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1483:
1484: skip_pagelist:
1485:
1486: if (print_log == 0)
1487: goto skip_log;
1488:
1489: (*pr)("\n");
1490: if ((pp->pr_roflags & PR_LOGGING) == 0)
1491: (*pr)("\tno log\n");
1492: else
1493: pr_printlog(pp, NULL, pr);
1.3 pk 1494:
1.25 thorpej 1495: skip_log:
1.44 thorpej 1496:
1497: if (print_cache == 0)
1498: goto skip_cache;
1499:
1.61 chs 1500: TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist) {
1.44 thorpej 1501: (*pr)("\tcache %p: allocfrom %p freeto %p\n", pc,
1502: pc->pc_allocfrom, pc->pc_freeto);
1.48 thorpej 1503: (*pr)("\t hits %lu misses %lu ngroups %lu nitems %lu\n",
1504: pc->pc_hits, pc->pc_misses, pc->pc_ngroups, pc->pc_nitems);
1.61 chs 1505: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.44 thorpej 1506: (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail);
1507: for (i = 0; i < PCG_NOBJECTS; i++)
1508: (*pr)("\t\t\t%p\n", pcg->pcg_objects[i]);
1509: }
1510: }
1511:
1512: skip_cache:
1.3 pk 1513:
1.25 thorpej 1514: pr_enter_check(pp, pr);
1.3 pk 1515: }
1516:
1517: int
1.42 thorpej 1518: pool_chk(struct pool *pp, const char *label)
1.3 pk 1519: {
1520: struct pool_item_header *ph;
1521: int r = 0;
1522:
1.21 thorpej 1523: simple_lock(&pp->pr_slock);
1.3 pk 1524:
1.61 chs 1525: TAILQ_FOREACH(ph, &pp->pr_pagelist, ph_pagelist) {
1.3 pk 1526: struct pool_item *pi;
1527: int n;
1528: caddr_t page;
1529:
1.66 thorpej 1530: page = (caddr_t)((u_long)ph & pp->pr_alloc->pa_pagemask);
1.20 thorpej 1531: if (page != ph->ph_page &&
1532: (pp->pr_roflags & PR_PHINPAGE) != 0) {
1.3 pk 1533: if (label != NULL)
1534: printf("%s: ", label);
1.16 briggs 1535: printf("pool(%p:%s): page inconsistency: page %p;"
1536: " at page head addr %p (p %p)\n", pp,
1.3 pk 1537: pp->pr_wchan, ph->ph_page,
1538: ph, page);
1539: r++;
1540: goto out;
1541: }
1542:
1543: for (pi = TAILQ_FIRST(&ph->ph_itemlist), n = 0;
1544: pi != NULL;
1545: pi = TAILQ_NEXT(pi,pi_list), n++) {
1546:
1547: #ifdef DIAGNOSTIC
1548: if (pi->pi_magic != PI_MAGIC) {
1549: if (label != NULL)
1550: printf("%s: ", label);
1551: printf("pool(%s): free list modified: magic=%x;"
1552: " page %p; item ordinal %d;"
1553: " addr %p (p %p)\n",
1554: pp->pr_wchan, pi->pi_magic, ph->ph_page,
1555: n, pi, page);
1556: panic("pool");
1557: }
1558: #endif
1.66 thorpej 1559: page =
1560: (caddr_t)((u_long)pi & pp->pr_alloc->pa_pagemask);
1.3 pk 1561: if (page == ph->ph_page)
1562: continue;
1563:
1564: if (label != NULL)
1565: printf("%s: ", label);
1.16 briggs 1566: printf("pool(%p:%s): page inconsistency: page %p;"
1567: " item ordinal %d; addr %p (p %p)\n", pp,
1.3 pk 1568: pp->pr_wchan, ph->ph_page,
1569: n, pi, page);
1570: r++;
1571: goto out;
1572: }
1573: }
1574: out:
1.21 thorpej 1575: simple_unlock(&pp->pr_slock);
1.3 pk 1576: return (r);
1.43 thorpej 1577: }
1578:
1579: /*
1580: * pool_cache_init:
1581: *
1582: * Initialize a pool cache.
1583: *
1584: * NOTE: If the pool must be protected from interrupts, we expect
1585: * to be called at the appropriate interrupt priority level.
1586: */
1587: void
1588: pool_cache_init(struct pool_cache *pc, struct pool *pp,
1589: int (*ctor)(void *, void *, int),
1590: void (*dtor)(void *, void *),
1591: void *arg)
1592: {
1593:
1594: TAILQ_INIT(&pc->pc_grouplist);
1595: simple_lock_init(&pc->pc_slock);
1596:
1597: pc->pc_allocfrom = NULL;
1598: pc->pc_freeto = NULL;
1599: pc->pc_pool = pp;
1600:
1601: pc->pc_ctor = ctor;
1602: pc->pc_dtor = dtor;
1603: pc->pc_arg = arg;
1604:
1.48 thorpej 1605: pc->pc_hits = 0;
1606: pc->pc_misses = 0;
1607:
1608: pc->pc_ngroups = 0;
1609:
1610: pc->pc_nitems = 0;
1611:
1.43 thorpej 1612: simple_lock(&pp->pr_slock);
1613: TAILQ_INSERT_TAIL(&pp->pr_cachelist, pc, pc_poollist);
1614: simple_unlock(&pp->pr_slock);
1615: }
1616:
1617: /*
1618: * pool_cache_destroy:
1619: *
1620: * Destroy a pool cache.
1621: */
1622: void
1623: pool_cache_destroy(struct pool_cache *pc)
1624: {
1625: struct pool *pp = pc->pc_pool;
1626:
1627: /* First, invalidate the entire cache. */
1628: pool_cache_invalidate(pc);
1629:
1630: /* ...and remove it from the pool's cache list. */
1631: simple_lock(&pp->pr_slock);
1632: TAILQ_REMOVE(&pp->pr_cachelist, pc, pc_poollist);
1633: simple_unlock(&pp->pr_slock);
1634: }
1635:
1636: static __inline void *
1637: pcg_get(struct pool_cache_group *pcg)
1638: {
1639: void *object;
1640: u_int idx;
1641:
1642: KASSERT(pcg->pcg_avail <= PCG_NOBJECTS);
1.45 thorpej 1643: KASSERT(pcg->pcg_avail != 0);
1.43 thorpej 1644: idx = --pcg->pcg_avail;
1645:
1646: KASSERT(pcg->pcg_objects[idx] != NULL);
1647: object = pcg->pcg_objects[idx];
1648: pcg->pcg_objects[idx] = NULL;
1649:
1650: return (object);
1651: }
1652:
1653: static __inline void
1654: pcg_put(struct pool_cache_group *pcg, void *object)
1655: {
1656: u_int idx;
1657:
1658: KASSERT(pcg->pcg_avail < PCG_NOBJECTS);
1659: idx = pcg->pcg_avail++;
1660:
1661: KASSERT(pcg->pcg_objects[idx] == NULL);
1662: pcg->pcg_objects[idx] = object;
1663: }
1664:
1665: /*
1666: * pool_cache_get:
1667: *
1668: * Get an object from a pool cache.
1669: */
1670: void *
1671: pool_cache_get(struct pool_cache *pc, int flags)
1672: {
1673: struct pool_cache_group *pcg;
1674: void *object;
1.58 thorpej 1675:
1676: #ifdef LOCKDEBUG
1677: if (flags & PR_WAITOK)
1678: simple_lock_only_held(NULL, "pool_cache_get(PR_WAITOK)");
1679: #endif
1.43 thorpej 1680:
1681: simple_lock(&pc->pc_slock);
1682:
1683: if ((pcg = pc->pc_allocfrom) == NULL) {
1.61 chs 1684: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.43 thorpej 1685: if (pcg->pcg_avail != 0) {
1686: pc->pc_allocfrom = pcg;
1687: goto have_group;
1688: }
1689: }
1690:
1691: /*
1692: * No groups with any available objects. Allocate
1693: * a new object, construct it, and return it to
1694: * the caller. We will allocate a group, if necessary,
1695: * when the object is freed back to the cache.
1696: */
1.48 thorpej 1697: pc->pc_misses++;
1.43 thorpej 1698: simple_unlock(&pc->pc_slock);
1699: object = pool_get(pc->pc_pool, flags);
1700: if (object != NULL && pc->pc_ctor != NULL) {
1701: if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) {
1702: pool_put(pc->pc_pool, object);
1703: return (NULL);
1704: }
1705: }
1.74.2.1! thorpej 1706: KASSERT((((vaddr_t)object + pc->pc_pool->pr_itemoffset) &
! 1707: (pc->pc_pool->pr_align - 1)) == 0);
1.43 thorpej 1708: return (object);
1709: }
1710:
1711: have_group:
1.48 thorpej 1712: pc->pc_hits++;
1713: pc->pc_nitems--;
1.43 thorpej 1714: object = pcg_get(pcg);
1715:
1716: if (pcg->pcg_avail == 0)
1717: pc->pc_allocfrom = NULL;
1.45 thorpej 1718:
1.43 thorpej 1719: simple_unlock(&pc->pc_slock);
1720:
1.74.2.1! thorpej 1721: KASSERT((((vaddr_t)object + pc->pc_pool->pr_itemoffset) &
! 1722: (pc->pc_pool->pr_align - 1)) == 0);
1.43 thorpej 1723: return (object);
1724: }
1725:
1726: /*
1727: * pool_cache_put:
1728: *
1729: * Put an object back to the pool cache.
1730: */
1731: void
1732: pool_cache_put(struct pool_cache *pc, void *object)
1733: {
1734: struct pool_cache_group *pcg;
1.60 thorpej 1735: int s;
1.43 thorpej 1736:
1737: simple_lock(&pc->pc_slock);
1738:
1739: if ((pcg = pc->pc_freeto) == NULL) {
1.61 chs 1740: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.43 thorpej 1741: if (pcg->pcg_avail != PCG_NOBJECTS) {
1742: pc->pc_freeto = pcg;
1743: goto have_group;
1744: }
1745: }
1746:
1747: /*
1748: * No empty groups to free the object to. Attempt to
1.47 thorpej 1749: * allocate one.
1.43 thorpej 1750: */
1.47 thorpej 1751: simple_unlock(&pc->pc_slock);
1.60 thorpej 1752: s = splvm();
1.43 thorpej 1753: pcg = pool_get(&pcgpool, PR_NOWAIT);
1.60 thorpej 1754: splx(s);
1.43 thorpej 1755: if (pcg != NULL) {
1756: memset(pcg, 0, sizeof(*pcg));
1.47 thorpej 1757: simple_lock(&pc->pc_slock);
1.48 thorpej 1758: pc->pc_ngroups++;
1.43 thorpej 1759: TAILQ_INSERT_TAIL(&pc->pc_grouplist, pcg, pcg_list);
1.47 thorpej 1760: if (pc->pc_freeto == NULL)
1761: pc->pc_freeto = pcg;
1.43 thorpej 1762: goto have_group;
1763: }
1764:
1765: /*
1766: * Unable to allocate a cache group; destruct the object
1767: * and free it back to the pool.
1768: */
1.51 thorpej 1769: pool_cache_destruct_object(pc, object);
1.43 thorpej 1770: return;
1771: }
1772:
1773: have_group:
1.48 thorpej 1774: pc->pc_nitems++;
1.43 thorpej 1775: pcg_put(pcg, object);
1776:
1777: if (pcg->pcg_avail == PCG_NOBJECTS)
1778: pc->pc_freeto = NULL;
1779:
1780: simple_unlock(&pc->pc_slock);
1.51 thorpej 1781: }
1782:
1783: /*
1784: * pool_cache_destruct_object:
1785: *
1786: * Force destruction of an object and its release back into
1787: * the pool.
1788: */
1789: void
1790: pool_cache_destruct_object(struct pool_cache *pc, void *object)
1791: {
1792:
1793: if (pc->pc_dtor != NULL)
1794: (*pc->pc_dtor)(pc->pc_arg, object);
1795: pool_put(pc->pc_pool, object);
1.43 thorpej 1796: }
1797:
1798: /*
1799: * pool_cache_do_invalidate:
1800: *
1801: * This internal function implements pool_cache_invalidate() and
1802: * pool_cache_reclaim().
1803: */
1804: static void
1805: pool_cache_do_invalidate(struct pool_cache *pc, int free_groups,
1.56 sommerfe 1806: void (*putit)(struct pool *, void *))
1.43 thorpej 1807: {
1808: struct pool_cache_group *pcg, *npcg;
1809: void *object;
1.60 thorpej 1810: int s;
1.43 thorpej 1811:
1812: for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
1813: pcg = npcg) {
1814: npcg = TAILQ_NEXT(pcg, pcg_list);
1815: while (pcg->pcg_avail != 0) {
1.48 thorpej 1816: pc->pc_nitems--;
1.43 thorpej 1817: object = pcg_get(pcg);
1.45 thorpej 1818: if (pcg->pcg_avail == 0 && pc->pc_allocfrom == pcg)
1819: pc->pc_allocfrom = NULL;
1.43 thorpej 1820: if (pc->pc_dtor != NULL)
1821: (*pc->pc_dtor)(pc->pc_arg, object);
1.56 sommerfe 1822: (*putit)(pc->pc_pool, object);
1.43 thorpej 1823: }
1824: if (free_groups) {
1.48 thorpej 1825: pc->pc_ngroups--;
1.43 thorpej 1826: TAILQ_REMOVE(&pc->pc_grouplist, pcg, pcg_list);
1.46 thorpej 1827: if (pc->pc_freeto == pcg)
1828: pc->pc_freeto = NULL;
1.60 thorpej 1829: s = splvm();
1.43 thorpej 1830: pool_put(&pcgpool, pcg);
1.60 thorpej 1831: splx(s);
1.43 thorpej 1832: }
1833: }
1834: }
1835:
1836: /*
1837: * pool_cache_invalidate:
1838: *
1839: * Invalidate a pool cache (destruct and release all of the
1840: * cached objects).
1841: */
1842: void
1843: pool_cache_invalidate(struct pool_cache *pc)
1844: {
1845:
1846: simple_lock(&pc->pc_slock);
1.56 sommerfe 1847: pool_cache_do_invalidate(pc, 0, pool_put);
1.43 thorpej 1848: simple_unlock(&pc->pc_slock);
1849: }
1850:
1851: /*
1852: * pool_cache_reclaim:
1853: *
1854: * Reclaim a pool cache for pool_reclaim().
1855: */
1856: static void
1857: pool_cache_reclaim(struct pool_cache *pc)
1858: {
1859:
1.47 thorpej 1860: simple_lock(&pc->pc_slock);
1.43 thorpej 1861: pool_cache_do_invalidate(pc, 1, pool_do_put);
1862: simple_unlock(&pc->pc_slock);
1.3 pk 1863: }
1.66 thorpej 1864:
1865: /*
1866: * Pool backend allocators.
1867: *
1868: * Each pool has a backend allocator that handles allocation, deallocation,
1869: * and any additional draining that might be needed.
1870: *
1871: * We provide two standard allocators:
1872: *
1873: * pool_allocator_kmem - the default when no allocator is specified
1874: *
1875: * pool_allocator_nointr - used for pools that will not be accessed
1876: * in interrupt context.
1877: */
1878: void *pool_page_alloc(struct pool *, int);
1879: void pool_page_free(struct pool *, void *);
1880:
1881: struct pool_allocator pool_allocator_kmem = {
1882: pool_page_alloc, pool_page_free, 0,
1883: };
1884:
1885: void *pool_page_alloc_nointr(struct pool *, int);
1886: void pool_page_free_nointr(struct pool *, void *);
1887:
1888: struct pool_allocator pool_allocator_nointr = {
1889: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1890: };
1891:
1892: #ifdef POOL_SUBPAGE
1893: void *pool_subpage_alloc(struct pool *, int);
1894: void pool_subpage_free(struct pool *, void *);
1895:
1896: struct pool_allocator pool_allocator_kmem_subpage = {
1897: pool_subpage_alloc, pool_subpage_free, 0,
1898: };
1899: #endif /* POOL_SUBPAGE */
1900:
1901: /*
1902: * We have at least three different resources for the same allocation and
1903: * each resource can be depleted. First, we have the ready elements in the
1904: * pool. Then we have the resource (typically a vm_map) for this allocator.
1905: * Finally, we have physical memory. Waiting for any of these can be
1906: * unnecessary when any other is freed, but the kernel doesn't support
1907: * sleeping on multiple wait channels, so we have to employ another strategy.
1908: *
1909: * The caller sleeps on the pool (so that it can be awakened when an item
1910: * is returned to the pool), but we set PA_WANT on the allocator. When a
1911: * page is returned to the allocator and PA_WANT is set, pool_allocator_free
1912: * will wake up all sleeping pools belonging to this allocator.
1913: *
1914: * XXX Thundering herd.
1915: */
1916: void *
1917: pool_allocator_alloc(struct pool *org, int flags)
1918: {
1919: struct pool_allocator *pa = org->pr_alloc;
1920: struct pool *pp, *start;
1921: int s, freed;
1922: void *res;
1923:
1924: do {
1925: if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
1926: return (res);
1.68 thorpej 1927: if ((flags & PR_WAITOK) == 0) {
1928: /*
1929: * We only run the drain hookhere if PR_NOWAIT.
1930: * In other cases, the hook will be run in
1931: * pool_reclaim().
1932: */
1933: if (org->pr_drain_hook != NULL) {
1934: (*org->pr_drain_hook)(org->pr_drain_hook_arg,
1935: flags);
1936: if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
1937: return (res);
1938: }
1.66 thorpej 1939: break;
1.68 thorpej 1940: }
1.66 thorpej 1941:
1942: /*
1943: * Drain all pools, except "org", that use this
1944: * allocator. We do this to reclaim VA space.
1945: * pa_alloc is responsible for waiting for
1946: * physical memory.
1947: *
1948: * XXX We risk looping forever if start if someone
1949: * calls pool_destroy on "start". But there is no
1950: * other way to have potentially sleeping pool_reclaim,
1951: * non-sleeping locks on pool_allocator, and some
1952: * stirring of drained pools in the allocator.
1.68 thorpej 1953: *
1954: * XXX Maybe we should use pool_head_slock for locking
1955: * the allocators?
1.66 thorpej 1956: */
1957: freed = 0;
1958:
1959: s = splvm();
1960: simple_lock(&pa->pa_slock);
1961: pp = start = TAILQ_FIRST(&pa->pa_list);
1962: do {
1963: TAILQ_REMOVE(&pa->pa_list, pp, pr_alloc_list);
1964: TAILQ_INSERT_TAIL(&pa->pa_list, pp, pr_alloc_list);
1965: if (pp == org)
1966: continue;
1.73 thorpej 1967: simple_unlock(&pa->pa_slock);
1.66 thorpej 1968: freed = pool_reclaim(pp);
1.73 thorpej 1969: simple_lock(&pa->pa_slock);
1.66 thorpej 1970: } while ((pp = TAILQ_FIRST(&pa->pa_list)) != start &&
1971: freed == 0);
1972:
1973: if (freed == 0) {
1974: /*
1975: * We set PA_WANT here, the caller will most likely
1976: * sleep waiting for pages (if not, this won't hurt
1977: * that much), and there is no way to set this in
1978: * the caller without violating locking order.
1979: */
1980: pa->pa_flags |= PA_WANT;
1981: }
1982: simple_unlock(&pa->pa_slock);
1983: splx(s);
1984: } while (freed);
1985: return (NULL);
1986: }
1987:
1988: void
1989: pool_allocator_free(struct pool *pp, void *v)
1990: {
1991: struct pool_allocator *pa = pp->pr_alloc;
1992: int s;
1993:
1994: (*pa->pa_free)(pp, v);
1995:
1996: s = splvm();
1997: simple_lock(&pa->pa_slock);
1998: if ((pa->pa_flags & PA_WANT) == 0) {
1999: simple_unlock(&pa->pa_slock);
2000: splx(s);
2001: return;
2002: }
2003:
2004: TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
2005: simple_lock(&pp->pr_slock);
2006: if ((pp->pr_flags & PR_WANTED) != 0) {
2007: pp->pr_flags &= ~PR_WANTED;
2008: wakeup(pp);
2009: }
1.69 thorpej 2010: simple_unlock(&pp->pr_slock);
1.66 thorpej 2011: }
2012: pa->pa_flags &= ~PA_WANT;
2013: simple_unlock(&pa->pa_slock);
2014: splx(s);
2015: }
2016:
2017: void *
2018: pool_page_alloc(struct pool *pp, int flags)
2019: {
2020: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2021:
2022: return ((void *) uvm_km_alloc_poolpage(waitok));
2023: }
2024:
2025: void
2026: pool_page_free(struct pool *pp, void *v)
2027: {
2028:
2029: uvm_km_free_poolpage((vaddr_t) v);
2030: }
2031:
2032: #ifdef POOL_SUBPAGE
2033: /* Sub-page allocator, for machines with large hardware pages. */
2034: void *
2035: pool_subpage_alloc(struct pool *pp, int flags)
2036: {
2037:
2038: return (pool_get(&psppool, flags));
2039: }
2040:
2041: void
2042: pool_subpage_free(struct pool *pp, void *v)
2043: {
2044:
2045: pool_put(&psppool, v);
2046: }
2047:
2048: /* We don't provide a real nointr allocator. Maybe later. */
2049: void *
2050: pool_page_alloc_nointr(struct pool *pp, int flags)
2051: {
2052:
2053: return (pool_subpage_alloc(pp, flags));
2054: }
2055:
2056: void
2057: pool_page_free_nointr(struct pool *pp, void *v)
2058: {
2059:
2060: pool_subpage_free(pp, v);
2061: }
2062: #else
2063: void *
2064: pool_page_alloc_nointr(struct pool *pp, int flags)
2065: {
2066: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2067:
2068: return ((void *) uvm_km_alloc_poolpage1(kernel_map,
2069: uvm.kernel_object, waitok));
2070: }
2071:
2072: void
2073: pool_page_free_nointr(struct pool *pp, void *v)
2074: {
2075:
2076: uvm_km_free_poolpage1(kernel_map, (vaddr_t) v);
2077: }
2078: #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