Annotation of src/sys/kern/subr_pool.c, Revision 1.89
1.89 ! yamt 1: /* $NetBSD: subr_pool.c,v 1.88 2003/11/13 02:44:01 chs 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>
1.89 ! yamt 41: __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.88 2003/11/13 02:44:01 chs Exp $");
1.24 scottr 42:
1.25 thorpej 43: #include "opt_pool.h"
1.24 scottr 44: #include "opt_poollog.h"
1.28 thorpej 45: #include "opt_lockdebug.h"
1.1 pk 46:
47: #include <sys/param.h>
48: #include <sys/systm.h>
49: #include <sys/proc.h>
50: #include <sys/errno.h>
51: #include <sys/kernel.h>
52: #include <sys/malloc.h>
53: #include <sys/lock.h>
54: #include <sys/pool.h>
1.20 thorpej 55: #include <sys/syslog.h>
1.3 pk 56:
57: #include <uvm/uvm.h>
58:
1.1 pk 59: /*
60: * Pool resource management utility.
1.3 pk 61: *
1.88 chs 62: * Memory is allocated in pages which are split into pieces according to
63: * the pool item size. Each page is kept on one of three lists in the
64: * pool structure: `pr_emptypages', `pr_fullpages' and `pr_partpages',
65: * for empty, full and partially-full pages respectively. The individual
66: * pool items are on a linked list headed by `ph_itemlist' in each page
67: * header. The memory for building the page list is either taken from
68: * the allocated pages themselves (for small pool items) or taken from
69: * an internal pool of page headers (`phpool').
1.1 pk 70: */
71:
1.3 pk 72: /* List of all pools */
1.5 thorpej 73: TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head);
1.3 pk 74:
75: /* Private pool for page header structures */
76: static struct pool phpool;
77:
1.62 bjh21 78: #ifdef POOL_SUBPAGE
79: /* Pool of subpages for use by normal pools. */
80: static struct pool psppool;
81: #endif
82:
1.3 pk 83: /* # of seconds to retain page after last use */
84: int pool_inactive_time = 10;
85:
86: /* Next candidate for drainage (see pool_drain()) */
1.23 thorpej 87: static struct pool *drainpp;
88:
89: /* This spin lock protects both pool_head and drainpp. */
90: struct simplelock pool_head_slock = SIMPLELOCK_INITIALIZER;
1.3 pk 91:
92: struct pool_item_header {
93: /* Page headers */
1.88 chs 94: LIST_ENTRY(pool_item_header)
1.3 pk 95: ph_pagelist; /* pool page list */
96: TAILQ_HEAD(,pool_item) ph_itemlist; /* chunk list for this page */
1.88 chs 97: SPLAY_ENTRY(pool_item_header)
98: ph_node; /* Off-page page headers */
1.79 thorpej 99: unsigned int ph_nmissing; /* # of chunks in use */
1.3 pk 100: caddr_t ph_page; /* this page's address */
101: struct timeval ph_time; /* last referenced */
102: };
103:
1.1 pk 104: struct pool_item {
1.3 pk 105: #ifdef DIAGNOSTIC
1.82 thorpej 106: u_int pi_magic;
1.33 chs 107: #endif
1.82 thorpej 108: #define PI_MAGIC 0xdeadbeefU
1.3 pk 109: /* Other entries use only this list entry */
110: TAILQ_ENTRY(pool_item) pi_list;
111: };
112:
1.53 thorpej 113: #define POOL_NEEDS_CATCHUP(pp) \
114: ((pp)->pr_nitems < (pp)->pr_minitems)
115:
1.43 thorpej 116: /*
117: * Pool cache management.
118: *
119: * Pool caches provide a way for constructed objects to be cached by the
120: * pool subsystem. This can lead to performance improvements by avoiding
121: * needless object construction/destruction; it is deferred until absolutely
122: * necessary.
123: *
124: * Caches are grouped into cache groups. Each cache group references
125: * up to 16 constructed objects. When a cache allocates an object
126: * from the pool, it calls the object's constructor and places it into
127: * a cache group. When a cache group frees an object back to the pool,
128: * it first calls the object's destructor. This allows the object to
129: * persist in constructed form while freed to the cache.
130: *
131: * Multiple caches may exist for each pool. This allows a single
132: * object type to have multiple constructed forms. The pool references
133: * each cache, so that when a pool is drained by the pagedaemon, it can
134: * drain each individual cache as well. Each time a cache is drained,
135: * the most idle cache group is freed to the pool in its entirety.
136: *
137: * Pool caches are layed on top of pools. By layering them, we can avoid
138: * the complexity of cache management for pools which would not benefit
139: * from it.
140: */
141:
142: /* The cache group pool. */
143: static struct pool pcgpool;
1.3 pk 144:
1.43 thorpej 145: static void pool_cache_reclaim(struct pool_cache *);
1.3 pk 146:
1.42 thorpej 147: static int pool_catchup(struct pool *);
1.55 thorpej 148: static void pool_prime_page(struct pool *, caddr_t,
149: struct pool_item_header *);
1.88 chs 150: static void pool_update_curpage(struct pool *);
1.66 thorpej 151:
152: void *pool_allocator_alloc(struct pool *, int);
153: void pool_allocator_free(struct pool *, void *);
1.3 pk 154:
1.88 chs 155: static void pool_print_pagelist(struct pool_pagelist *,
156: void (*)(const char *, ...));
1.42 thorpej 157: static void pool_print1(struct pool *, const char *,
158: void (*)(const char *, ...));
1.3 pk 159:
1.88 chs 160: static int pool_chk_page(struct pool *, const char *,
161: struct pool_item_header *);
162:
1.3 pk 163: /*
1.52 thorpej 164: * Pool log entry. An array of these is allocated in pool_init().
1.3 pk 165: */
166: struct pool_log {
167: const char *pl_file;
168: long pl_line;
169: int pl_action;
1.25 thorpej 170: #define PRLOG_GET 1
171: #define PRLOG_PUT 2
1.3 pk 172: void *pl_addr;
1.1 pk 173: };
174:
1.86 matt 175: #ifdef POOL_DIAGNOSTIC
1.3 pk 176: /* Number of entries in pool log buffers */
1.17 thorpej 177: #ifndef POOL_LOGSIZE
178: #define POOL_LOGSIZE 10
179: #endif
180:
181: int pool_logsize = POOL_LOGSIZE;
1.1 pk 182:
1.42 thorpej 183: static __inline void
184: pr_log(struct pool *pp, void *v, int action, const char *file, long line)
1.3 pk 185: {
186: int n = pp->pr_curlogentry;
187: struct pool_log *pl;
188:
1.20 thorpej 189: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 190: return;
191:
192: /*
193: * Fill in the current entry. Wrap around and overwrite
194: * the oldest entry if necessary.
195: */
196: pl = &pp->pr_log[n];
197: pl->pl_file = file;
198: pl->pl_line = line;
199: pl->pl_action = action;
200: pl->pl_addr = v;
201: if (++n >= pp->pr_logsize)
202: n = 0;
203: pp->pr_curlogentry = n;
204: }
205:
206: static void
1.42 thorpej 207: pr_printlog(struct pool *pp, struct pool_item *pi,
208: void (*pr)(const char *, ...))
1.3 pk 209: {
210: int i = pp->pr_logsize;
211: int n = pp->pr_curlogentry;
212:
1.20 thorpej 213: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 214: return;
215:
216: /*
217: * Print all entries in this pool's log.
218: */
219: while (i-- > 0) {
220: struct pool_log *pl = &pp->pr_log[n];
221: if (pl->pl_action != 0) {
1.25 thorpej 222: if (pi == NULL || pi == pl->pl_addr) {
223: (*pr)("\tlog entry %d:\n", i);
224: (*pr)("\t\taction = %s, addr = %p\n",
225: pl->pl_action == PRLOG_GET ? "get" : "put",
226: pl->pl_addr);
227: (*pr)("\t\tfile: %s at line %lu\n",
228: pl->pl_file, pl->pl_line);
229: }
1.3 pk 230: }
231: if (++n >= pp->pr_logsize)
232: n = 0;
233: }
234: }
1.25 thorpej 235:
1.42 thorpej 236: static __inline void
237: pr_enter(struct pool *pp, const char *file, long line)
1.25 thorpej 238: {
239:
1.34 thorpej 240: if (__predict_false(pp->pr_entered_file != NULL)) {
1.25 thorpej 241: printf("pool %s: reentrancy at file %s line %ld\n",
242: pp->pr_wchan, file, line);
243: printf(" previous entry at file %s line %ld\n",
244: pp->pr_entered_file, pp->pr_entered_line);
245: panic("pr_enter");
246: }
247:
248: pp->pr_entered_file = file;
249: pp->pr_entered_line = line;
250: }
251:
1.42 thorpej 252: static __inline void
253: pr_leave(struct pool *pp)
1.25 thorpej 254: {
255:
1.34 thorpej 256: if (__predict_false(pp->pr_entered_file == NULL)) {
1.25 thorpej 257: printf("pool %s not entered?\n", pp->pr_wchan);
258: panic("pr_leave");
259: }
260:
261: pp->pr_entered_file = NULL;
262: pp->pr_entered_line = 0;
263: }
264:
1.42 thorpej 265: static __inline void
266: pr_enter_check(struct pool *pp, void (*pr)(const char *, ...))
1.25 thorpej 267: {
268:
269: if (pp->pr_entered_file != NULL)
270: (*pr)("\n\tcurrently entered from file %s line %ld\n",
271: pp->pr_entered_file, pp->pr_entered_line);
272: }
1.3 pk 273: #else
1.25 thorpej 274: #define pr_log(pp, v, action, file, line)
275: #define pr_printlog(pp, pi, pr)
276: #define pr_enter(pp, file, line)
277: #define pr_leave(pp)
278: #define pr_enter_check(pp, pr)
1.59 thorpej 279: #endif /* POOL_DIAGNOSTIC */
1.3 pk 280:
1.88 chs 281: static __inline int
282: phtree_compare(struct pool_item_header *a, struct pool_item_header *b)
283: {
284: if (a->ph_page < b->ph_page)
285: return (-1);
286: else if (a->ph_page > b->ph_page)
287: return (1);
288: else
289: return (0);
290: }
291:
292: SPLAY_PROTOTYPE(phtree, pool_item_header, ph_node, phtree_compare);
293: SPLAY_GENERATE(phtree, pool_item_header, ph_node, phtree_compare);
294:
1.3 pk 295: /*
296: * Return the pool page header based on page address.
297: */
1.42 thorpej 298: static __inline struct pool_item_header *
299: pr_find_pagehead(struct pool *pp, caddr_t page)
1.3 pk 300: {
1.88 chs 301: struct pool_item_header *ph, tmp;
1.3 pk 302:
1.20 thorpej 303: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
1.3 pk 304: return ((struct pool_item_header *)(page + pp->pr_phoffset));
305:
1.88 chs 306: tmp.ph_page = page;
307: ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
308: return ph;
1.3 pk 309: }
310:
311: /*
312: * Remove a page from the pool.
313: */
1.42 thorpej 314: static __inline void
1.61 chs 315: pr_rmpage(struct pool *pp, struct pool_item_header *ph,
316: struct pool_pagelist *pq)
1.3 pk 317: {
1.61 chs 318: int s;
1.3 pk 319:
320: /*
1.7 thorpej 321: * If the page was idle, decrement the idle page count.
1.3 pk 322: */
1.6 thorpej 323: if (ph->ph_nmissing == 0) {
324: #ifdef DIAGNOSTIC
325: if (pp->pr_nidle == 0)
326: panic("pr_rmpage: nidle inconsistent");
1.20 thorpej 327: if (pp->pr_nitems < pp->pr_itemsperpage)
328: panic("pr_rmpage: nitems inconsistent");
1.6 thorpej 329: #endif
330: pp->pr_nidle--;
331: }
1.7 thorpej 332:
1.20 thorpej 333: pp->pr_nitems -= pp->pr_itemsperpage;
334:
1.7 thorpej 335: /*
1.61 chs 336: * Unlink a page from the pool and release it (or queue it for release).
1.7 thorpej 337: */
1.88 chs 338: LIST_REMOVE(ph, ph_pagelist);
1.61 chs 339: if (pq) {
1.88 chs 340: LIST_INSERT_HEAD(pq, ph, ph_pagelist);
1.61 chs 341: } else {
1.66 thorpej 342: pool_allocator_free(pp, ph->ph_page);
1.61 chs 343: if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
1.88 chs 344: SPLAY_REMOVE(phtree, &pp->pr_phtree, ph);
1.85 pk 345: s = splvm();
1.61 chs 346: pool_put(&phpool, ph);
347: splx(s);
348: }
349: }
1.7 thorpej 350: pp->pr_npages--;
351: pp->pr_npagefree++;
1.6 thorpej 352:
1.88 chs 353: pool_update_curpage(pp);
1.3 pk 354: }
355:
356: /*
357: * Initialize the given pool resource structure.
358: *
359: * We export this routine to allow other kernel parts to declare
360: * static pools that must be initialized before malloc() is available.
361: */
362: void
1.42 thorpej 363: pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
1.66 thorpej 364: const char *wchan, struct pool_allocator *palloc)
1.3 pk 365: {
1.88 chs 366: int off, slack;
1.3 pk 367:
1.25 thorpej 368: #ifdef POOL_DIAGNOSTIC
369: /*
370: * Always log if POOL_DIAGNOSTIC is defined.
371: */
372: if (pool_logsize != 0)
373: flags |= PR_LOGGING;
374: #endif
375:
1.66 thorpej 376: #ifdef POOL_SUBPAGE
377: /*
378: * XXX We don't provide a real `nointr' back-end
379: * yet; all sub-pages come from a kmem back-end.
380: * maybe some day...
381: */
382: if (palloc == NULL) {
383: extern struct pool_allocator pool_allocator_kmem_subpage;
384: palloc = &pool_allocator_kmem_subpage;
385: }
1.3 pk 386: /*
1.66 thorpej 387: * We'll assume any user-specified back-end allocator
388: * will deal with sub-pages, or simply don't care.
1.3 pk 389: */
1.66 thorpej 390: #else
391: if (palloc == NULL)
392: palloc = &pool_allocator_kmem;
393: #endif /* POOL_SUBPAGE */
394: if ((palloc->pa_flags & PA_INITIALIZED) == 0) {
395: if (palloc->pa_pagesz == 0) {
1.62 bjh21 396: #ifdef POOL_SUBPAGE
1.66 thorpej 397: if (palloc == &pool_allocator_kmem)
398: palloc->pa_pagesz = PAGE_SIZE;
399: else
400: palloc->pa_pagesz = POOL_SUBPAGE;
1.62 bjh21 401: #else
1.66 thorpej 402: palloc->pa_pagesz = PAGE_SIZE;
403: #endif /* POOL_SUBPAGE */
404: }
405:
406: TAILQ_INIT(&palloc->pa_list);
407:
408: simple_lock_init(&palloc->pa_slock);
409: palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
410: palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
411: palloc->pa_flags |= PA_INITIALIZED;
1.4 thorpej 412: }
1.3 pk 413:
414: if (align == 0)
415: align = ALIGN(1);
1.14 thorpej 416:
417: if (size < sizeof(struct pool_item))
418: size = sizeof(struct pool_item);
1.3 pk 419:
1.78 thorpej 420: size = roundup(size, align);
1.66 thorpej 421: #ifdef DIAGNOSTIC
422: if (size > palloc->pa_pagesz)
1.35 pk 423: panic("pool_init: pool item size (%lu) too large",
424: (u_long)size);
1.66 thorpej 425: #endif
1.35 pk 426:
1.3 pk 427: /*
428: * Initialize the pool structure.
429: */
1.88 chs 430: LIST_INIT(&pp->pr_emptypages);
431: LIST_INIT(&pp->pr_fullpages);
432: LIST_INIT(&pp->pr_partpages);
1.43 thorpej 433: TAILQ_INIT(&pp->pr_cachelist);
1.3 pk 434: pp->pr_curpage = NULL;
435: pp->pr_npages = 0;
436: pp->pr_minitems = 0;
437: pp->pr_minpages = 0;
438: pp->pr_maxpages = UINT_MAX;
1.20 thorpej 439: pp->pr_roflags = flags;
440: pp->pr_flags = 0;
1.35 pk 441: pp->pr_size = size;
1.3 pk 442: pp->pr_align = align;
443: pp->pr_wchan = wchan;
1.66 thorpej 444: pp->pr_alloc = palloc;
1.20 thorpej 445: pp->pr_nitems = 0;
446: pp->pr_nout = 0;
447: pp->pr_hardlimit = UINT_MAX;
448: pp->pr_hardlimit_warning = NULL;
1.31 thorpej 449: pp->pr_hardlimit_ratecap.tv_sec = 0;
450: pp->pr_hardlimit_ratecap.tv_usec = 0;
451: pp->pr_hardlimit_warning_last.tv_sec = 0;
452: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.68 thorpej 453: pp->pr_drain_hook = NULL;
454: pp->pr_drain_hook_arg = NULL;
1.3 pk 455:
456: /*
457: * Decide whether to put the page header off page to avoid
458: * wasting too large a part of the page. Off-page page headers
459: * go on a hash table, so we can match a returned item
460: * with its header based on the page address.
461: * We use 1/16 of the page size as the threshold (XXX: tune)
462: */
1.66 thorpej 463: if (pp->pr_size < palloc->pa_pagesz/16) {
1.3 pk 464: /* Use the end of the page for the page header */
1.20 thorpej 465: pp->pr_roflags |= PR_PHINPAGE;
1.66 thorpej 466: pp->pr_phoffset = off = palloc->pa_pagesz -
467: ALIGN(sizeof(struct pool_item_header));
1.2 pk 468: } else {
1.3 pk 469: /* The page header will be taken from our page header pool */
470: pp->pr_phoffset = 0;
1.66 thorpej 471: off = palloc->pa_pagesz;
1.88 chs 472: SPLAY_INIT(&pp->pr_phtree);
1.2 pk 473: }
1.1 pk 474:
1.3 pk 475: /*
476: * Alignment is to take place at `ioff' within the item. This means
477: * we must reserve up to `align - 1' bytes on the page to allow
478: * appropriate positioning of each item.
479: *
480: * Silently enforce `0 <= ioff < align'.
481: */
482: pp->pr_itemoffset = ioff = ioff % align;
483: pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
1.43 thorpej 484: KASSERT(pp->pr_itemsperpage != 0);
1.3 pk 485:
486: /*
487: * Use the slack between the chunks and the page header
488: * for "cache coloring".
489: */
490: slack = off - pp->pr_itemsperpage * pp->pr_size;
491: pp->pr_maxcolor = (slack / align) * align;
492: pp->pr_curcolor = 0;
493:
494: pp->pr_nget = 0;
495: pp->pr_nfail = 0;
496: pp->pr_nput = 0;
497: pp->pr_npagealloc = 0;
498: pp->pr_npagefree = 0;
1.1 pk 499: pp->pr_hiwat = 0;
1.8 thorpej 500: pp->pr_nidle = 0;
1.3 pk 501:
1.59 thorpej 502: #ifdef POOL_DIAGNOSTIC
1.25 thorpej 503: if (flags & PR_LOGGING) {
504: if (kmem_map == NULL ||
505: (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log),
506: M_TEMP, M_NOWAIT)) == NULL)
1.20 thorpej 507: pp->pr_roflags &= ~PR_LOGGING;
1.3 pk 508: pp->pr_curlogentry = 0;
509: pp->pr_logsize = pool_logsize;
510: }
1.59 thorpej 511: #endif
1.25 thorpej 512:
513: pp->pr_entered_file = NULL;
514: pp->pr_entered_line = 0;
1.3 pk 515:
1.21 thorpej 516: simple_lock_init(&pp->pr_slock);
1.1 pk 517:
1.3 pk 518: /*
1.43 thorpej 519: * Initialize private page header pool and cache magazine pool if we
520: * haven't done so yet.
1.23 thorpej 521: * XXX LOCKING.
1.3 pk 522: */
523: if (phpool.pr_size == 0) {
1.62 bjh21 524: #ifdef POOL_SUBPAGE
525: pool_init(&phpool, sizeof(struct pool_item_header), 0, 0, 0,
1.66 thorpej 526: "phpool", &pool_allocator_kmem);
1.62 bjh21 527: pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
1.66 thorpej 528: PR_RECURSIVE, "psppool", &pool_allocator_kmem);
1.62 bjh21 529: #else
1.3 pk 530: pool_init(&phpool, sizeof(struct pool_item_header), 0, 0,
1.66 thorpej 531: 0, "phpool", NULL);
1.62 bjh21 532: #endif
1.43 thorpej 533: pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
1.66 thorpej 534: 0, "pcgpool", NULL);
1.1 pk 535: }
536:
1.23 thorpej 537: /* Insert into the list of all pools. */
538: simple_lock(&pool_head_slock);
539: TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist);
540: simple_unlock(&pool_head_slock);
1.66 thorpej 541:
542: /* Insert this into the list of pools using this allocator. */
543: simple_lock(&palloc->pa_slock);
544: TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
545: simple_unlock(&palloc->pa_slock);
1.1 pk 546: }
547:
548: /*
549: * De-commision a pool resource.
550: */
551: void
1.42 thorpej 552: pool_destroy(struct pool *pp)
1.1 pk 553: {
1.3 pk 554: struct pool_item_header *ph;
1.43 thorpej 555: struct pool_cache *pc;
556:
1.66 thorpej 557: /* Locking order: pool_allocator -> pool */
558: simple_lock(&pp->pr_alloc->pa_slock);
559: TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
560: simple_unlock(&pp->pr_alloc->pa_slock);
561:
1.43 thorpej 562: /* Destroy all caches for this pool. */
563: while ((pc = TAILQ_FIRST(&pp->pr_cachelist)) != NULL)
564: pool_cache_destroy(pc);
1.3 pk 565:
566: #ifdef DIAGNOSTIC
1.20 thorpej 567: if (pp->pr_nout != 0) {
1.25 thorpej 568: pr_printlog(pp, NULL, printf);
1.80 provos 569: panic("pool_destroy: pool busy: still out: %u",
1.20 thorpej 570: pp->pr_nout);
1.3 pk 571: }
572: #endif
1.1 pk 573:
1.3 pk 574: /* Remove all pages */
1.88 chs 575: while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1.70 thorpej 576: pr_rmpage(pp, ph, NULL);
1.88 chs 577: KASSERT(LIST_EMPTY(&pp->pr_fullpages));
578: KASSERT(LIST_EMPTY(&pp->pr_partpages));
1.3 pk 579:
580: /* Remove from global pool list */
1.23 thorpej 581: simple_lock(&pool_head_slock);
1.3 pk 582: TAILQ_REMOVE(&pool_head, pp, pr_poollist);
1.61 chs 583: if (drainpp == pp) {
584: drainpp = NULL;
585: }
1.23 thorpej 586: simple_unlock(&pool_head_slock);
1.3 pk 587:
1.59 thorpej 588: #ifdef POOL_DIAGNOSTIC
1.20 thorpej 589: if ((pp->pr_roflags & PR_LOGGING) != 0)
1.3 pk 590: free(pp->pr_log, M_TEMP);
1.59 thorpej 591: #endif
1.1 pk 592: }
593:
1.68 thorpej 594: void
595: pool_set_drain_hook(struct pool *pp, void (*fn)(void *, int), void *arg)
596: {
597:
598: /* XXX no locking -- must be used just after pool_init() */
599: #ifdef DIAGNOSTIC
600: if (pp->pr_drain_hook != NULL)
601: panic("pool_set_drain_hook(%s): already set", pp->pr_wchan);
602: #endif
603: pp->pr_drain_hook = fn;
604: pp->pr_drain_hook_arg = arg;
605: }
606:
1.88 chs 607: static struct pool_item_header *
1.55 thorpej 608: pool_alloc_item_header(struct pool *pp, caddr_t storage, int flags)
609: {
610: struct pool_item_header *ph;
611: int s;
612:
613: LOCK_ASSERT(simple_lock_held(&pp->pr_slock) == 0);
614:
615: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
616: ph = (struct pool_item_header *) (storage + pp->pr_phoffset);
617: else {
1.85 pk 618: s = splvm();
1.55 thorpej 619: ph = pool_get(&phpool, flags);
620: splx(s);
621: }
622:
623: return (ph);
624: }
1.1 pk 625:
626: /*
1.3 pk 627: * Grab an item from the pool; must be called at appropriate spl level
1.1 pk 628: */
1.3 pk 629: void *
1.59 thorpej 630: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 631: _pool_get(struct pool *pp, int flags, const char *file, long line)
1.56 sommerfe 632: #else
633: pool_get(struct pool *pp, int flags)
634: #endif
1.1 pk 635: {
636: struct pool_item *pi;
1.3 pk 637: struct pool_item_header *ph;
1.55 thorpej 638: void *v;
1.1 pk 639:
1.2 pk 640: #ifdef DIAGNOSTIC
1.84 thorpej 641: if (__predict_false(curlwp == NULL && doing_shutdown == 0 &&
1.37 sommerfe 642: (flags & PR_WAITOK) != 0))
1.77 matt 643: panic("pool_get: %s: must have NOWAIT", pp->pr_wchan);
1.58 thorpej 644:
645: #ifdef LOCKDEBUG
646: if (flags & PR_WAITOK)
647: simple_lock_only_held(NULL, "pool_get(PR_WAITOK)");
1.56 sommerfe 648: #endif
1.58 thorpej 649: #endif /* DIAGNOSTIC */
1.1 pk 650:
1.21 thorpej 651: simple_lock(&pp->pr_slock);
1.25 thorpej 652: pr_enter(pp, file, line);
1.20 thorpej 653:
654: startover:
655: /*
656: * Check to see if we've reached the hard limit. If we have,
657: * and we can wait, then wait until an item has been returned to
658: * the pool.
659: */
660: #ifdef DIAGNOSTIC
1.34 thorpej 661: if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
1.25 thorpej 662: pr_leave(pp);
1.21 thorpej 663: simple_unlock(&pp->pr_slock);
1.20 thorpej 664: panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
665: }
666: #endif
1.34 thorpej 667: if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
1.68 thorpej 668: if (pp->pr_drain_hook != NULL) {
669: /*
670: * Since the drain hook is going to free things
671: * back to the pool, unlock, call the hook, re-lock,
672: * and check the hardlimit condition again.
673: */
674: pr_leave(pp);
675: simple_unlock(&pp->pr_slock);
676: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
677: simple_lock(&pp->pr_slock);
678: pr_enter(pp, file, line);
679: if (pp->pr_nout < pp->pr_hardlimit)
680: goto startover;
681: }
682:
1.29 sommerfe 683: if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
1.20 thorpej 684: /*
685: * XXX: A warning isn't logged in this case. Should
686: * it be?
687: */
688: pp->pr_flags |= PR_WANTED;
1.25 thorpej 689: pr_leave(pp);
1.40 sommerfe 690: ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
1.25 thorpej 691: pr_enter(pp, file, line);
1.20 thorpej 692: goto startover;
693: }
1.31 thorpej 694:
695: /*
696: * Log a message that the hard limit has been hit.
697: */
698: if (pp->pr_hardlimit_warning != NULL &&
699: ratecheck(&pp->pr_hardlimit_warning_last,
700: &pp->pr_hardlimit_ratecap))
701: log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
1.21 thorpej 702:
703: pp->pr_nfail++;
704:
1.25 thorpej 705: pr_leave(pp);
1.21 thorpej 706: simple_unlock(&pp->pr_slock);
1.20 thorpej 707: return (NULL);
708: }
709:
1.3 pk 710: /*
711: * The convention we use is that if `curpage' is not NULL, then
712: * it points at a non-empty bucket. In particular, `curpage'
713: * never points at a page header which has PR_PHINPAGE set and
714: * has no items in its bucket.
715: */
1.20 thorpej 716: if ((ph = pp->pr_curpage) == NULL) {
717: #ifdef DIAGNOSTIC
718: if (pp->pr_nitems != 0) {
1.21 thorpej 719: simple_unlock(&pp->pr_slock);
1.20 thorpej 720: printf("pool_get: %s: curpage NULL, nitems %u\n",
721: pp->pr_wchan, pp->pr_nitems);
1.80 provos 722: panic("pool_get: nitems inconsistent");
1.20 thorpej 723: }
724: #endif
725:
1.21 thorpej 726: /*
727: * Call the back-end page allocator for more memory.
728: * Release the pool lock, as the back-end page allocator
729: * may block.
730: */
1.25 thorpej 731: pr_leave(pp);
1.21 thorpej 732: simple_unlock(&pp->pr_slock);
1.66 thorpej 733: v = pool_allocator_alloc(pp, flags);
1.55 thorpej 734: if (__predict_true(v != NULL))
735: ph = pool_alloc_item_header(pp, v, flags);
1.21 thorpej 736: simple_lock(&pp->pr_slock);
1.25 thorpej 737: pr_enter(pp, file, line);
1.15 pk 738:
1.55 thorpej 739: if (__predict_false(v == NULL || ph == NULL)) {
740: if (v != NULL)
1.66 thorpej 741: pool_allocator_free(pp, v);
1.55 thorpej 742:
1.21 thorpej 743: /*
1.55 thorpej 744: * We were unable to allocate a page or item
745: * header, but we released the lock during
746: * allocation, so perhaps items were freed
747: * back to the pool. Check for this case.
1.21 thorpej 748: */
749: if (pp->pr_curpage != NULL)
750: goto startover;
1.15 pk 751:
1.3 pk 752: if ((flags & PR_WAITOK) == 0) {
753: pp->pr_nfail++;
1.25 thorpej 754: pr_leave(pp);
1.21 thorpej 755: simple_unlock(&pp->pr_slock);
1.1 pk 756: return (NULL);
1.3 pk 757: }
758:
1.15 pk 759: /*
760: * Wait for items to be returned to this pool.
1.21 thorpej 761: *
1.20 thorpej 762: * XXX: maybe we should wake up once a second and
763: * try again?
1.15 pk 764: */
1.1 pk 765: pp->pr_flags |= PR_WANTED;
1.66 thorpej 766: /* PA_WANTED is already set on the allocator. */
1.25 thorpej 767: pr_leave(pp);
1.40 sommerfe 768: ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
1.25 thorpej 769: pr_enter(pp, file, line);
1.20 thorpej 770: goto startover;
1.1 pk 771: }
1.3 pk 772:
1.15 pk 773: /* We have more memory; add it to the pool */
1.55 thorpej 774: pool_prime_page(pp, v, ph);
1.15 pk 775: pp->pr_npagealloc++;
776:
1.20 thorpej 777: /* Start the allocation process over. */
778: goto startover;
1.3 pk 779: }
1.34 thorpej 780: if (__predict_false((v = pi = TAILQ_FIRST(&ph->ph_itemlist)) == NULL)) {
1.25 thorpej 781: pr_leave(pp);
1.21 thorpej 782: simple_unlock(&pp->pr_slock);
1.3 pk 783: panic("pool_get: %s: page empty", pp->pr_wchan);
1.21 thorpej 784: }
1.20 thorpej 785: #ifdef DIAGNOSTIC
1.34 thorpej 786: if (__predict_false(pp->pr_nitems == 0)) {
1.25 thorpej 787: pr_leave(pp);
1.21 thorpej 788: simple_unlock(&pp->pr_slock);
1.20 thorpej 789: printf("pool_get: %s: items on itemlist, nitems %u\n",
790: pp->pr_wchan, pp->pr_nitems);
1.80 provos 791: panic("pool_get: nitems inconsistent");
1.20 thorpej 792: }
1.65 enami 793: #endif
1.56 sommerfe 794:
1.65 enami 795: #ifdef POOL_DIAGNOSTIC
1.3 pk 796: pr_log(pp, v, PRLOG_GET, file, line);
1.65 enami 797: #endif
1.3 pk 798:
1.65 enami 799: #ifdef DIAGNOSTIC
1.34 thorpej 800: if (__predict_false(pi->pi_magic != PI_MAGIC)) {
1.25 thorpej 801: pr_printlog(pp, pi, printf);
1.3 pk 802: panic("pool_get(%s): free list modified: magic=%x; page %p;"
803: " item addr %p\n",
804: pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
805: }
806: #endif
807:
808: /*
809: * Remove from item list.
810: */
811: TAILQ_REMOVE(&ph->ph_itemlist, pi, pi_list);
1.20 thorpej 812: pp->pr_nitems--;
813: pp->pr_nout++;
1.6 thorpej 814: if (ph->ph_nmissing == 0) {
815: #ifdef DIAGNOSTIC
1.34 thorpej 816: if (__predict_false(pp->pr_nidle == 0))
1.6 thorpej 817: panic("pool_get: nidle inconsistent");
818: #endif
819: pp->pr_nidle--;
1.88 chs 820:
821: /*
822: * This page was previously empty. Move it to the list of
823: * partially-full pages. This page is already curpage.
824: */
825: LIST_REMOVE(ph, ph_pagelist);
826: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.6 thorpej 827: }
1.3 pk 828: ph->ph_nmissing++;
1.88 chs 829: if (TAILQ_EMPTY(&ph->ph_itemlist)) {
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: /*
1.88 chs 839: * This page is now full. Move it to the full list
840: * and select a new current page.
1.3 pk 841: */
1.88 chs 842: LIST_REMOVE(ph, ph_pagelist);
843: LIST_INSERT_HEAD(&pp->pr_fullpages, ph, ph_pagelist);
844: pool_update_curpage(pp);
1.1 pk 845: }
1.3 pk 846:
847: pp->pr_nget++;
1.20 thorpej 848:
849: /*
850: * If we have a low water mark and we are now below that low
851: * water mark, add more items to the pool.
852: */
1.53 thorpej 853: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 854: /*
855: * XXX: Should we log a warning? Should we set up a timeout
856: * to try again in a second or so? The latter could break
857: * a caller's assumptions about interrupt protection, etc.
858: */
859: }
860:
1.25 thorpej 861: pr_leave(pp);
1.21 thorpej 862: simple_unlock(&pp->pr_slock);
1.1 pk 863: return (v);
864: }
865:
866: /*
1.43 thorpej 867: * Internal version of pool_put(). Pool is already locked/entered.
1.1 pk 868: */
1.43 thorpej 869: static void
1.56 sommerfe 870: pool_do_put(struct pool *pp, void *v)
1.1 pk 871: {
872: struct pool_item *pi = v;
1.3 pk 873: struct pool_item_header *ph;
874: caddr_t page;
1.21 thorpej 875: int s;
1.3 pk 876:
1.61 chs 877: LOCK_ASSERT(simple_lock_held(&pp->pr_slock));
878:
1.66 thorpej 879: page = (caddr_t)((u_long)v & pp->pr_alloc->pa_pagemask);
1.1 pk 880:
1.30 thorpej 881: #ifdef DIAGNOSTIC
1.34 thorpej 882: if (__predict_false(pp->pr_nout == 0)) {
1.30 thorpej 883: printf("pool %s: putting with none out\n",
884: pp->pr_wchan);
885: panic("pool_put");
886: }
887: #endif
1.3 pk 888:
1.34 thorpej 889: if (__predict_false((ph = pr_find_pagehead(pp, page)) == NULL)) {
1.25 thorpej 890: pr_printlog(pp, NULL, printf);
1.3 pk 891: panic("pool_put: %s: page header missing", pp->pr_wchan);
892: }
1.28 thorpej 893:
894: #ifdef LOCKDEBUG
895: /*
896: * Check if we're freeing a locked simple lock.
897: */
898: simple_lock_freecheck((caddr_t)pi, ((caddr_t)pi) + pp->pr_size);
899: #endif
1.3 pk 900:
901: /*
902: * Return to item list.
903: */
1.2 pk 904: #ifdef DIAGNOSTIC
1.3 pk 905: pi->pi_magic = PI_MAGIC;
906: #endif
1.32 chs 907: #ifdef DEBUG
908: {
909: int i, *ip = v;
910:
911: for (i = 0; i < pp->pr_size / sizeof(int); i++) {
912: *ip++ = PI_MAGIC;
913: }
914: }
915: #endif
916:
1.3 pk 917: TAILQ_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.79 thorpej 918: KDASSERT(ph->ph_nmissing != 0);
1.3 pk 919: ph->ph_nmissing--;
920: pp->pr_nput++;
1.20 thorpej 921: pp->pr_nitems++;
922: pp->pr_nout--;
1.3 pk 923:
924: /* Cancel "pool empty" condition if it exists */
925: if (pp->pr_curpage == NULL)
926: pp->pr_curpage = ph;
927:
928: if (pp->pr_flags & PR_WANTED) {
929: pp->pr_flags &= ~PR_WANTED;
1.15 pk 930: if (ph->ph_nmissing == 0)
931: pp->pr_nidle++;
1.3 pk 932: wakeup((caddr_t)pp);
933: return;
934: }
935:
936: /*
1.88 chs 937: * If this page is now empty, do one of two things:
1.21 thorpej 938: *
1.88 chs 939: * (1) If we have more pages than the page high water mark,
940: * free the page back to the system.
1.21 thorpej 941: *
1.88 chs 942: * (2) Otherwise, move the page to the empty page list.
943: *
944: * Either way, select a new current page (so we use a partially-full
945: * page if one is available).
1.3 pk 946: */
947: if (ph->ph_nmissing == 0) {
1.6 thorpej 948: pp->pr_nidle++;
1.71 thorpej 949: if (pp->pr_npages > pp->pr_maxpages ||
950: (pp->pr_alloc->pa_flags & PA_WANT) != 0) {
1.61 chs 951: pr_rmpage(pp, ph, NULL);
1.3 pk 952: } else {
1.88 chs 953: LIST_REMOVE(ph, ph_pagelist);
954: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.3 pk 955:
1.21 thorpej 956: /*
957: * Update the timestamp on the page. A page must
958: * be idle for some period of time before it can
959: * be reclaimed by the pagedaemon. This minimizes
960: * ping-pong'ing for memory.
961: */
962: s = splclock();
963: ph->ph_time = mono_time;
964: splx(s);
1.1 pk 965: }
1.88 chs 966: pool_update_curpage(pp);
1.1 pk 967: }
1.88 chs 968:
1.21 thorpej 969: /*
1.88 chs 970: * If the page was previously completely full, move it to the
971: * partially-full list and make it the current page. The next
972: * allocation will get the item from this page, instead of
973: * further fragmenting the pool.
1.21 thorpej 974: */
975: else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
1.88 chs 976: LIST_REMOVE(ph, ph_pagelist);
977: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.21 thorpej 978: pp->pr_curpage = ph;
979: }
1.43 thorpej 980: }
981:
982: /*
983: * Return resource to the pool; must be called at appropriate spl level
984: */
1.59 thorpej 985: #ifdef POOL_DIAGNOSTIC
1.43 thorpej 986: void
987: _pool_put(struct pool *pp, void *v, const char *file, long line)
988: {
989:
990: simple_lock(&pp->pr_slock);
991: pr_enter(pp, file, line);
992:
1.56 sommerfe 993: pr_log(pp, v, PRLOG_PUT, file, line);
994:
995: pool_do_put(pp, v);
1.21 thorpej 996:
1.25 thorpej 997: pr_leave(pp);
1.21 thorpej 998: simple_unlock(&pp->pr_slock);
1.1 pk 999: }
1.57 sommerfe 1000: #undef pool_put
1.59 thorpej 1001: #endif /* POOL_DIAGNOSTIC */
1.1 pk 1002:
1.56 sommerfe 1003: void
1004: pool_put(struct pool *pp, void *v)
1005: {
1006:
1007: simple_lock(&pp->pr_slock);
1008:
1009: pool_do_put(pp, v);
1010:
1011: simple_unlock(&pp->pr_slock);
1012: }
1.57 sommerfe 1013:
1.59 thorpej 1014: #ifdef POOL_DIAGNOSTIC
1.57 sommerfe 1015: #define pool_put(h, v) _pool_put((h), (v), __FILE__, __LINE__)
1.56 sommerfe 1016: #endif
1.74 thorpej 1017:
1018: /*
1019: * Add N items to the pool.
1020: */
1021: int
1022: pool_prime(struct pool *pp, int n)
1023: {
1.83 scw 1024: struct pool_item_header *ph = NULL;
1.74 thorpej 1025: caddr_t cp;
1.75 simonb 1026: int newpages;
1.74 thorpej 1027:
1028: simple_lock(&pp->pr_slock);
1029:
1030: newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1031:
1032: while (newpages-- > 0) {
1033: simple_unlock(&pp->pr_slock);
1034: cp = pool_allocator_alloc(pp, PR_NOWAIT);
1035: if (__predict_true(cp != NULL))
1036: ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1037: simple_lock(&pp->pr_slock);
1038:
1039: if (__predict_false(cp == NULL || ph == NULL)) {
1040: if (cp != NULL)
1041: pool_allocator_free(pp, cp);
1042: break;
1043: }
1044:
1045: pool_prime_page(pp, cp, ph);
1046: pp->pr_npagealloc++;
1047: pp->pr_minpages++;
1048: }
1049:
1050: if (pp->pr_minpages >= pp->pr_maxpages)
1051: pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */
1052:
1053: simple_unlock(&pp->pr_slock);
1054: return (0);
1055: }
1.55 thorpej 1056:
1057: /*
1.3 pk 1058: * Add a page worth of items to the pool.
1.21 thorpej 1059: *
1060: * Note, we must be called with the pool descriptor LOCKED.
1.3 pk 1061: */
1.55 thorpej 1062: static void
1063: pool_prime_page(struct pool *pp, caddr_t storage, struct pool_item_header *ph)
1.3 pk 1064: {
1065: struct pool_item *pi;
1066: caddr_t cp = storage;
1067: unsigned int align = pp->pr_align;
1068: unsigned int ioff = pp->pr_itemoffset;
1.55 thorpej 1069: int n;
1.89 ! yamt 1070: int s;
1.36 pk 1071:
1.66 thorpej 1072: #ifdef DIAGNOSTIC
1073: if (((u_long)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
1.36 pk 1074: panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1.66 thorpej 1075: #endif
1.3 pk 1076:
1077: /*
1078: * Insert page header.
1079: */
1.88 chs 1080: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.3 pk 1081: TAILQ_INIT(&ph->ph_itemlist);
1082: ph->ph_page = storage;
1083: ph->ph_nmissing = 0;
1.89 ! yamt 1084: s = splclock();
! 1085: ph->ph_time = mono_time;
! 1086: splx(s);
1.88 chs 1087: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1088: SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
1.3 pk 1089:
1.6 thorpej 1090: pp->pr_nidle++;
1091:
1.3 pk 1092: /*
1093: * Color this page.
1094: */
1095: cp = (caddr_t)(cp + pp->pr_curcolor);
1096: if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
1097: pp->pr_curcolor = 0;
1098:
1099: /*
1100: * Adjust storage to apply aligment to `pr_itemoffset' in each item.
1101: */
1102: if (ioff != 0)
1103: cp = (caddr_t)(cp + (align - ioff));
1104:
1105: /*
1106: * Insert remaining chunks on the bucket list.
1107: */
1108: n = pp->pr_itemsperpage;
1.20 thorpej 1109: pp->pr_nitems += n;
1.3 pk 1110:
1111: while (n--) {
1112: pi = (struct pool_item *)cp;
1.78 thorpej 1113:
1114: KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
1.3 pk 1115:
1116: /* Insert on page list */
1117: TAILQ_INSERT_TAIL(&ph->ph_itemlist, pi, pi_list);
1118: #ifdef DIAGNOSTIC
1119: pi->pi_magic = PI_MAGIC;
1120: #endif
1121: cp = (caddr_t)(cp + pp->pr_size);
1122: }
1123:
1124: /*
1125: * If the pool was depleted, point at the new page.
1126: */
1127: if (pp->pr_curpage == NULL)
1128: pp->pr_curpage = ph;
1129:
1130: if (++pp->pr_npages > pp->pr_hiwat)
1131: pp->pr_hiwat = pp->pr_npages;
1132: }
1133:
1.20 thorpej 1134: /*
1.52 thorpej 1135: * Used by pool_get() when nitems drops below the low water mark. This
1.88 chs 1136: * is used to catch up pr_nitems with the low water mark.
1.20 thorpej 1137: *
1.21 thorpej 1138: * Note 1, we never wait for memory here, we let the caller decide what to do.
1.20 thorpej 1139: *
1.73 thorpej 1140: * Note 2, we must be called with the pool already locked, and we return
1.20 thorpej 1141: * with it locked.
1142: */
1143: static int
1.42 thorpej 1144: pool_catchup(struct pool *pp)
1.20 thorpej 1145: {
1.83 scw 1146: struct pool_item_header *ph = NULL;
1.20 thorpej 1147: caddr_t cp;
1148: int error = 0;
1149:
1.54 thorpej 1150: while (POOL_NEEDS_CATCHUP(pp)) {
1.20 thorpej 1151: /*
1.21 thorpej 1152: * Call the page back-end allocator for more memory.
1153: *
1154: * XXX: We never wait, so should we bother unlocking
1155: * the pool descriptor?
1.20 thorpej 1156: */
1.21 thorpej 1157: simple_unlock(&pp->pr_slock);
1.66 thorpej 1158: cp = pool_allocator_alloc(pp, PR_NOWAIT);
1.55 thorpej 1159: if (__predict_true(cp != NULL))
1160: ph = pool_alloc_item_header(pp, cp, PR_NOWAIT);
1.21 thorpej 1161: simple_lock(&pp->pr_slock);
1.55 thorpej 1162: if (__predict_false(cp == NULL || ph == NULL)) {
1163: if (cp != NULL)
1.66 thorpej 1164: pool_allocator_free(pp, cp);
1.20 thorpej 1165: error = ENOMEM;
1166: break;
1167: }
1.55 thorpej 1168: pool_prime_page(pp, cp, ph);
1.26 thorpej 1169: pp->pr_npagealloc++;
1.20 thorpej 1170: }
1171:
1172: return (error);
1173: }
1174:
1.88 chs 1175: static void
1176: pool_update_curpage(struct pool *pp)
1177: {
1178:
1179: pp->pr_curpage = LIST_FIRST(&pp->pr_partpages);
1180: if (pp->pr_curpage == NULL) {
1181: pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages);
1182: }
1183: }
1184:
1.3 pk 1185: void
1.42 thorpej 1186: pool_setlowat(struct pool *pp, int n)
1.3 pk 1187: {
1.15 pk 1188:
1.21 thorpej 1189: simple_lock(&pp->pr_slock);
1190:
1.3 pk 1191: pp->pr_minitems = n;
1.15 pk 1192: pp->pr_minpages = (n == 0)
1193: ? 0
1.18 thorpej 1194: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.20 thorpej 1195:
1196: /* Make sure we're caught up with the newly-set low water mark. */
1.75 simonb 1197: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1198: /*
1199: * XXX: Should we log a warning? Should we set up a timeout
1200: * to try again in a second or so? The latter could break
1201: * a caller's assumptions about interrupt protection, etc.
1202: */
1203: }
1.21 thorpej 1204:
1205: simple_unlock(&pp->pr_slock);
1.3 pk 1206: }
1207:
1208: void
1.42 thorpej 1209: pool_sethiwat(struct pool *pp, int n)
1.3 pk 1210: {
1.15 pk 1211:
1.21 thorpej 1212: simple_lock(&pp->pr_slock);
1213:
1.15 pk 1214: pp->pr_maxpages = (n == 0)
1215: ? 0
1.18 thorpej 1216: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1217:
1218: simple_unlock(&pp->pr_slock);
1.3 pk 1219: }
1220:
1.20 thorpej 1221: void
1.42 thorpej 1222: pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1.20 thorpej 1223: {
1224:
1.21 thorpej 1225: simple_lock(&pp->pr_slock);
1.20 thorpej 1226:
1227: pp->pr_hardlimit = n;
1228: pp->pr_hardlimit_warning = warnmess;
1.31 thorpej 1229: pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1230: pp->pr_hardlimit_warning_last.tv_sec = 0;
1231: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.20 thorpej 1232:
1233: /*
1.21 thorpej 1234: * In-line version of pool_sethiwat(), because we don't want to
1235: * release the lock.
1.20 thorpej 1236: */
1237: pp->pr_maxpages = (n == 0)
1238: ? 0
1239: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1240:
1241: simple_unlock(&pp->pr_slock);
1.20 thorpej 1242: }
1.3 pk 1243:
1244: /*
1245: * Release all complete pages that have not been used recently.
1246: */
1.66 thorpej 1247: int
1.59 thorpej 1248: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 1249: _pool_reclaim(struct pool *pp, const char *file, long line)
1.56 sommerfe 1250: #else
1251: pool_reclaim(struct pool *pp)
1252: #endif
1.3 pk 1253: {
1254: struct pool_item_header *ph, *phnext;
1.43 thorpej 1255: struct pool_cache *pc;
1.21 thorpej 1256: struct timeval curtime;
1.61 chs 1257: struct pool_pagelist pq;
1.88 chs 1258: struct timeval diff;
1.21 thorpej 1259: int s;
1.3 pk 1260:
1.68 thorpej 1261: if (pp->pr_drain_hook != NULL) {
1262: /*
1263: * The drain hook must be called with the pool unlocked.
1264: */
1265: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT);
1266: }
1267:
1.21 thorpej 1268: if (simple_lock_try(&pp->pr_slock) == 0)
1.66 thorpej 1269: return (0);
1.25 thorpej 1270: pr_enter(pp, file, line);
1.68 thorpej 1271:
1.88 chs 1272: LIST_INIT(&pq);
1.3 pk 1273:
1.43 thorpej 1274: /*
1275: * Reclaim items from the pool's caches.
1276: */
1.61 chs 1277: TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist)
1.43 thorpej 1278: pool_cache_reclaim(pc);
1279:
1.21 thorpej 1280: s = splclock();
1281: curtime = mono_time;
1282: splx(s);
1283:
1.88 chs 1284: for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
1285: phnext = LIST_NEXT(ph, ph_pagelist);
1.3 pk 1286:
1287: /* Check our minimum page claim */
1288: if (pp->pr_npages <= pp->pr_minpages)
1289: break;
1290:
1.88 chs 1291: KASSERT(ph->ph_nmissing == 0);
1292: timersub(&curtime, &ph->ph_time, &diff);
1293: if (diff.tv_sec < pool_inactive_time)
1294: continue;
1.21 thorpej 1295:
1.88 chs 1296: /*
1297: * If freeing this page would put us below
1298: * the low water mark, stop now.
1299: */
1300: if ((pp->pr_nitems - pp->pr_itemsperpage) <
1301: pp->pr_minitems)
1302: break;
1.21 thorpej 1303:
1.88 chs 1304: pr_rmpage(pp, ph, &pq);
1.3 pk 1305: }
1306:
1.25 thorpej 1307: pr_leave(pp);
1.21 thorpej 1308: simple_unlock(&pp->pr_slock);
1.88 chs 1309: if (LIST_EMPTY(&pq))
1.66 thorpej 1310: return (0);
1311:
1.88 chs 1312: while ((ph = LIST_FIRST(&pq)) != NULL) {
1313: LIST_REMOVE(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: }
1.88 chs 1318: SPLAY_REMOVE(phtree, &pp->pr_phtree, ph);
1.85 pk 1319: s = splvm();
1.61 chs 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
1.88 chs 1404: pool_print_pagelist(struct pool_pagelist *pl, void (*pr)(const char *, ...))
1405: {
1406: struct pool_item_header *ph;
1407: #ifdef DIAGNOSTIC
1408: struct pool_item *pi;
1409: #endif
1410:
1411: LIST_FOREACH(ph, pl, ph_pagelist) {
1412: (*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n",
1413: ph->ph_page, ph->ph_nmissing,
1414: (u_long)ph->ph_time.tv_sec,
1415: (u_long)ph->ph_time.tv_usec);
1416: #ifdef DIAGNOSTIC
1417: TAILQ_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1418: if (pi->pi_magic != PI_MAGIC) {
1419: (*pr)("\t\t\titem %p, magic 0x%x\n",
1420: pi, pi->pi_magic);
1421: }
1422: }
1423: #endif
1424: }
1425: }
1426:
1427: static void
1.42 thorpej 1428: pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.3 pk 1429: {
1.25 thorpej 1430: struct pool_item_header *ph;
1.44 thorpej 1431: struct pool_cache *pc;
1432: struct pool_cache_group *pcg;
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: }
1444:
1445: (*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1446: pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1447: pp->pr_roflags);
1.66 thorpej 1448: (*pr)("\talloc %p\n", pp->pr_alloc);
1.25 thorpej 1449: (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1450: pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1451: (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1452: pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1453:
1454: (*pr)("\n\tnget %lu, nfail %lu, nput %lu\n",
1455: pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1456: (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1457: pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1458:
1459: if (print_pagelist == 0)
1460: goto skip_pagelist;
1461:
1.88 chs 1462: if ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1463: (*pr)("\n\tempty page list:\n");
1464: pool_print_pagelist(&pp->pr_emptypages, pr);
1465: if ((ph = LIST_FIRST(&pp->pr_fullpages)) != NULL)
1466: (*pr)("\n\tfull page list:\n");
1467: pool_print_pagelist(&pp->pr_fullpages, pr);
1468: if ((ph = LIST_FIRST(&pp->pr_partpages)) != NULL)
1469: (*pr)("\n\tpartial-page list:\n");
1470: pool_print_pagelist(&pp->pr_partpages, pr);
1471:
1.25 thorpej 1472: if (pp->pr_curpage == NULL)
1473: (*pr)("\tno current page\n");
1474: else
1475: (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1476:
1477: skip_pagelist:
1478: if (print_log == 0)
1479: goto skip_log;
1480:
1481: (*pr)("\n");
1482: if ((pp->pr_roflags & PR_LOGGING) == 0)
1483: (*pr)("\tno log\n");
1484: else
1485: pr_printlog(pp, NULL, pr);
1.3 pk 1486:
1.25 thorpej 1487: skip_log:
1.44 thorpej 1488: if (print_cache == 0)
1489: goto skip_cache;
1490:
1.61 chs 1491: TAILQ_FOREACH(pc, &pp->pr_cachelist, pc_poollist) {
1.44 thorpej 1492: (*pr)("\tcache %p: allocfrom %p freeto %p\n", pc,
1493: pc->pc_allocfrom, pc->pc_freeto);
1.48 thorpej 1494: (*pr)("\t hits %lu misses %lu ngroups %lu nitems %lu\n",
1495: pc->pc_hits, pc->pc_misses, pc->pc_ngroups, pc->pc_nitems);
1.61 chs 1496: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.44 thorpej 1497: (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail);
1.87 thorpej 1498: for (i = 0; i < PCG_NOBJECTS; i++) {
1499: if (pcg->pcg_objects[i].pcgo_pa !=
1500: POOL_PADDR_INVALID) {
1501: (*pr)("\t\t\t%p, 0x%llx\n",
1502: pcg->pcg_objects[i].pcgo_va,
1503: (unsigned long long)
1504: pcg->pcg_objects[i].pcgo_pa);
1505: } else {
1506: (*pr)("\t\t\t%p\n",
1507: pcg->pcg_objects[i].pcgo_va);
1508: }
1509: }
1.44 thorpej 1510: }
1511: }
1512:
1513: skip_cache:
1.88 chs 1514: pr_enter_check(pp, pr);
1515: }
1516:
1517: static int
1518: pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
1519: {
1520: struct pool_item *pi;
1521: caddr_t page;
1522: int n;
1523:
1524: page = (caddr_t)((u_long)ph & pp->pr_alloc->pa_pagemask);
1525: if (page != ph->ph_page &&
1526: (pp->pr_roflags & PR_PHINPAGE) != 0) {
1527: if (label != NULL)
1528: printf("%s: ", label);
1529: printf("pool(%p:%s): page inconsistency: page %p;"
1530: " at page head addr %p (p %p)\n", pp,
1531: pp->pr_wchan, ph->ph_page,
1532: ph, page);
1533: return 1;
1534: }
1.3 pk 1535:
1.88 chs 1536: for (pi = TAILQ_FIRST(&ph->ph_itemlist), n = 0;
1537: pi != NULL;
1538: pi = TAILQ_NEXT(pi,pi_list), n++) {
1539:
1540: #ifdef DIAGNOSTIC
1541: if (pi->pi_magic != PI_MAGIC) {
1542: if (label != NULL)
1543: printf("%s: ", label);
1544: printf("pool(%s): free list modified: magic=%x;"
1545: " page %p; item ordinal %d;"
1546: " addr %p (p %p)\n",
1547: pp->pr_wchan, pi->pi_magic, ph->ph_page,
1548: n, pi, page);
1549: panic("pool");
1550: }
1551: #endif
1552: page =
1553: (caddr_t)((u_long)pi & pp->pr_alloc->pa_pagemask);
1554: if (page == ph->ph_page)
1555: continue;
1556:
1557: if (label != NULL)
1558: printf("%s: ", label);
1559: printf("pool(%p:%s): page inconsistency: page %p;"
1560: " item ordinal %d; addr %p (p %p)\n", pp,
1561: pp->pr_wchan, ph->ph_page,
1562: n, pi, page);
1563: return 1;
1564: }
1565: return 0;
1.3 pk 1566: }
1567:
1.88 chs 1568:
1.3 pk 1569: int
1.42 thorpej 1570: pool_chk(struct pool *pp, const char *label)
1.3 pk 1571: {
1572: struct pool_item_header *ph;
1573: int r = 0;
1574:
1.21 thorpej 1575: simple_lock(&pp->pr_slock);
1.88 chs 1576: LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
1577: r = pool_chk_page(pp, label, ph);
1578: if (r) {
1579: goto out;
1580: }
1581: }
1582: LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
1583: r = pool_chk_page(pp, label, ph);
1584: if (r) {
1.3 pk 1585: goto out;
1586: }
1.88 chs 1587: }
1588: LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
1589: r = pool_chk_page(pp, label, ph);
1590: if (r) {
1.3 pk 1591: goto out;
1592: }
1593: }
1.88 chs 1594:
1.3 pk 1595: out:
1.21 thorpej 1596: simple_unlock(&pp->pr_slock);
1.3 pk 1597: return (r);
1.43 thorpej 1598: }
1599:
1600: /*
1601: * pool_cache_init:
1602: *
1603: * Initialize a pool cache.
1604: *
1605: * NOTE: If the pool must be protected from interrupts, we expect
1606: * to be called at the appropriate interrupt priority level.
1607: */
1608: void
1609: pool_cache_init(struct pool_cache *pc, struct pool *pp,
1610: int (*ctor)(void *, void *, int),
1611: void (*dtor)(void *, void *),
1612: void *arg)
1613: {
1614:
1615: TAILQ_INIT(&pc->pc_grouplist);
1616: simple_lock_init(&pc->pc_slock);
1617:
1618: pc->pc_allocfrom = NULL;
1619: pc->pc_freeto = NULL;
1620: pc->pc_pool = pp;
1621:
1622: pc->pc_ctor = ctor;
1623: pc->pc_dtor = dtor;
1624: pc->pc_arg = arg;
1625:
1.48 thorpej 1626: pc->pc_hits = 0;
1627: pc->pc_misses = 0;
1628:
1629: pc->pc_ngroups = 0;
1630:
1631: pc->pc_nitems = 0;
1632:
1.43 thorpej 1633: simple_lock(&pp->pr_slock);
1634: TAILQ_INSERT_TAIL(&pp->pr_cachelist, pc, pc_poollist);
1635: simple_unlock(&pp->pr_slock);
1636: }
1637:
1638: /*
1639: * pool_cache_destroy:
1640: *
1641: * Destroy a pool cache.
1642: */
1643: void
1644: pool_cache_destroy(struct pool_cache *pc)
1645: {
1646: struct pool *pp = pc->pc_pool;
1647:
1648: /* First, invalidate the entire cache. */
1649: pool_cache_invalidate(pc);
1650:
1651: /* ...and remove it from the pool's cache list. */
1652: simple_lock(&pp->pr_slock);
1653: TAILQ_REMOVE(&pp->pr_cachelist, pc, pc_poollist);
1654: simple_unlock(&pp->pr_slock);
1655: }
1656:
1657: static __inline void *
1.87 thorpej 1658: pcg_get(struct pool_cache_group *pcg, paddr_t *pap)
1.43 thorpej 1659: {
1660: void *object;
1661: u_int idx;
1662:
1663: KASSERT(pcg->pcg_avail <= PCG_NOBJECTS);
1.45 thorpej 1664: KASSERT(pcg->pcg_avail != 0);
1.43 thorpej 1665: idx = --pcg->pcg_avail;
1666:
1.87 thorpej 1667: KASSERT(pcg->pcg_objects[idx].pcgo_va != NULL);
1668: object = pcg->pcg_objects[idx].pcgo_va;
1669: if (pap != NULL)
1670: *pap = pcg->pcg_objects[idx].pcgo_pa;
1671: pcg->pcg_objects[idx].pcgo_va = NULL;
1.43 thorpej 1672:
1673: return (object);
1674: }
1675:
1676: static __inline void
1.87 thorpej 1677: pcg_put(struct pool_cache_group *pcg, void *object, paddr_t pa)
1.43 thorpej 1678: {
1679: u_int idx;
1680:
1681: KASSERT(pcg->pcg_avail < PCG_NOBJECTS);
1682: idx = pcg->pcg_avail++;
1683:
1.87 thorpej 1684: KASSERT(pcg->pcg_objects[idx].pcgo_va == NULL);
1685: pcg->pcg_objects[idx].pcgo_va = object;
1686: pcg->pcg_objects[idx].pcgo_pa = pa;
1.43 thorpej 1687: }
1688:
1689: /*
1.87 thorpej 1690: * pool_cache_get{,_paddr}:
1.43 thorpej 1691: *
1.87 thorpej 1692: * Get an object from a pool cache (optionally returning
1693: * the physical address of the object).
1.43 thorpej 1694: */
1695: void *
1.87 thorpej 1696: pool_cache_get_paddr(struct pool_cache *pc, int flags, paddr_t *pap)
1.43 thorpej 1697: {
1698: struct pool_cache_group *pcg;
1699: void *object;
1.58 thorpej 1700:
1701: #ifdef LOCKDEBUG
1702: if (flags & PR_WAITOK)
1703: simple_lock_only_held(NULL, "pool_cache_get(PR_WAITOK)");
1704: #endif
1.43 thorpej 1705:
1706: simple_lock(&pc->pc_slock);
1707:
1708: if ((pcg = pc->pc_allocfrom) == NULL) {
1.61 chs 1709: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.43 thorpej 1710: if (pcg->pcg_avail != 0) {
1711: pc->pc_allocfrom = pcg;
1712: goto have_group;
1713: }
1714: }
1715:
1716: /*
1717: * No groups with any available objects. Allocate
1718: * a new object, construct it, and return it to
1719: * the caller. We will allocate a group, if necessary,
1720: * when the object is freed back to the cache.
1721: */
1.48 thorpej 1722: pc->pc_misses++;
1.43 thorpej 1723: simple_unlock(&pc->pc_slock);
1724: object = pool_get(pc->pc_pool, flags);
1725: if (object != NULL && pc->pc_ctor != NULL) {
1726: if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) {
1727: pool_put(pc->pc_pool, object);
1728: return (NULL);
1729: }
1730: }
1.87 thorpej 1731: if (object != NULL && pap != NULL) {
1732: #ifdef POOL_VTOPHYS
1733: *pap = POOL_VTOPHYS(object);
1734: #else
1735: *pap = POOL_PADDR_INVALID;
1736: #endif
1737: }
1.43 thorpej 1738: return (object);
1739: }
1740:
1741: have_group:
1.48 thorpej 1742: pc->pc_hits++;
1743: pc->pc_nitems--;
1.87 thorpej 1744: object = pcg_get(pcg, pap);
1.43 thorpej 1745:
1746: if (pcg->pcg_avail == 0)
1747: pc->pc_allocfrom = NULL;
1.45 thorpej 1748:
1.43 thorpej 1749: simple_unlock(&pc->pc_slock);
1750:
1751: return (object);
1752: }
1753:
1754: /*
1.87 thorpej 1755: * pool_cache_put{,_paddr}:
1.43 thorpej 1756: *
1.87 thorpej 1757: * Put an object back to the pool cache (optionally caching the
1758: * physical address of the object).
1.43 thorpej 1759: */
1760: void
1.87 thorpej 1761: pool_cache_put_paddr(struct pool_cache *pc, void *object, paddr_t pa)
1.43 thorpej 1762: {
1763: struct pool_cache_group *pcg;
1.60 thorpej 1764: int s;
1.43 thorpej 1765:
1766: simple_lock(&pc->pc_slock);
1767:
1768: if ((pcg = pc->pc_freeto) == NULL) {
1.61 chs 1769: TAILQ_FOREACH(pcg, &pc->pc_grouplist, pcg_list) {
1.43 thorpej 1770: if (pcg->pcg_avail != PCG_NOBJECTS) {
1771: pc->pc_freeto = pcg;
1772: goto have_group;
1773: }
1774: }
1775:
1776: /*
1777: * No empty groups to free the object to. Attempt to
1.47 thorpej 1778: * allocate one.
1.43 thorpej 1779: */
1.47 thorpej 1780: simple_unlock(&pc->pc_slock);
1.60 thorpej 1781: s = splvm();
1.43 thorpej 1782: pcg = pool_get(&pcgpool, PR_NOWAIT);
1.60 thorpej 1783: splx(s);
1.43 thorpej 1784: if (pcg != NULL) {
1785: memset(pcg, 0, sizeof(*pcg));
1.47 thorpej 1786: simple_lock(&pc->pc_slock);
1.48 thorpej 1787: pc->pc_ngroups++;
1.43 thorpej 1788: TAILQ_INSERT_TAIL(&pc->pc_grouplist, pcg, pcg_list);
1.47 thorpej 1789: if (pc->pc_freeto == NULL)
1790: pc->pc_freeto = pcg;
1.43 thorpej 1791: goto have_group;
1792: }
1793:
1794: /*
1795: * Unable to allocate a cache group; destruct the object
1796: * and free it back to the pool.
1797: */
1.51 thorpej 1798: pool_cache_destruct_object(pc, object);
1.43 thorpej 1799: return;
1800: }
1801:
1802: have_group:
1.48 thorpej 1803: pc->pc_nitems++;
1.87 thorpej 1804: pcg_put(pcg, object, pa);
1.43 thorpej 1805:
1806: if (pcg->pcg_avail == PCG_NOBJECTS)
1807: pc->pc_freeto = NULL;
1808:
1809: simple_unlock(&pc->pc_slock);
1.51 thorpej 1810: }
1811:
1812: /*
1813: * pool_cache_destruct_object:
1814: *
1815: * Force destruction of an object and its release back into
1816: * the pool.
1817: */
1818: void
1819: pool_cache_destruct_object(struct pool_cache *pc, void *object)
1820: {
1821:
1822: if (pc->pc_dtor != NULL)
1823: (*pc->pc_dtor)(pc->pc_arg, object);
1824: pool_put(pc->pc_pool, object);
1.43 thorpej 1825: }
1826:
1827: /*
1828: * pool_cache_do_invalidate:
1829: *
1830: * This internal function implements pool_cache_invalidate() and
1831: * pool_cache_reclaim().
1832: */
1833: static void
1834: pool_cache_do_invalidate(struct pool_cache *pc, int free_groups,
1.56 sommerfe 1835: void (*putit)(struct pool *, void *))
1.43 thorpej 1836: {
1837: struct pool_cache_group *pcg, *npcg;
1838: void *object;
1.60 thorpej 1839: int s;
1.43 thorpej 1840:
1841: for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
1842: pcg = npcg) {
1843: npcg = TAILQ_NEXT(pcg, pcg_list);
1844: while (pcg->pcg_avail != 0) {
1.48 thorpej 1845: pc->pc_nitems--;
1.87 thorpej 1846: object = pcg_get(pcg, NULL);
1.45 thorpej 1847: if (pcg->pcg_avail == 0 && pc->pc_allocfrom == pcg)
1848: pc->pc_allocfrom = NULL;
1.43 thorpej 1849: if (pc->pc_dtor != NULL)
1850: (*pc->pc_dtor)(pc->pc_arg, object);
1.56 sommerfe 1851: (*putit)(pc->pc_pool, object);
1.43 thorpej 1852: }
1853: if (free_groups) {
1.48 thorpej 1854: pc->pc_ngroups--;
1.43 thorpej 1855: TAILQ_REMOVE(&pc->pc_grouplist, pcg, pcg_list);
1.46 thorpej 1856: if (pc->pc_freeto == pcg)
1857: pc->pc_freeto = NULL;
1.60 thorpej 1858: s = splvm();
1.43 thorpej 1859: pool_put(&pcgpool, pcg);
1.60 thorpej 1860: splx(s);
1.43 thorpej 1861: }
1862: }
1863: }
1864:
1865: /*
1866: * pool_cache_invalidate:
1867: *
1868: * Invalidate a pool cache (destruct and release all of the
1869: * cached objects).
1870: */
1871: void
1872: pool_cache_invalidate(struct pool_cache *pc)
1873: {
1874:
1875: simple_lock(&pc->pc_slock);
1.56 sommerfe 1876: pool_cache_do_invalidate(pc, 0, pool_put);
1.43 thorpej 1877: simple_unlock(&pc->pc_slock);
1878: }
1879:
1880: /*
1881: * pool_cache_reclaim:
1882: *
1883: * Reclaim a pool cache for pool_reclaim().
1884: */
1885: static void
1886: pool_cache_reclaim(struct pool_cache *pc)
1887: {
1888:
1.47 thorpej 1889: simple_lock(&pc->pc_slock);
1.43 thorpej 1890: pool_cache_do_invalidate(pc, 1, pool_do_put);
1891: simple_unlock(&pc->pc_slock);
1.3 pk 1892: }
1.66 thorpej 1893:
1894: /*
1895: * Pool backend allocators.
1896: *
1897: * Each pool has a backend allocator that handles allocation, deallocation,
1898: * and any additional draining that might be needed.
1899: *
1900: * We provide two standard allocators:
1901: *
1902: * pool_allocator_kmem - the default when no allocator is specified
1903: *
1904: * pool_allocator_nointr - used for pools that will not be accessed
1905: * in interrupt context.
1906: */
1907: void *pool_page_alloc(struct pool *, int);
1908: void pool_page_free(struct pool *, void *);
1909:
1910: struct pool_allocator pool_allocator_kmem = {
1911: pool_page_alloc, pool_page_free, 0,
1912: };
1913:
1914: void *pool_page_alloc_nointr(struct pool *, int);
1915: void pool_page_free_nointr(struct pool *, void *);
1916:
1917: struct pool_allocator pool_allocator_nointr = {
1918: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1919: };
1920:
1921: #ifdef POOL_SUBPAGE
1922: void *pool_subpage_alloc(struct pool *, int);
1923: void pool_subpage_free(struct pool *, void *);
1924:
1925: struct pool_allocator pool_allocator_kmem_subpage = {
1926: pool_subpage_alloc, pool_subpage_free, 0,
1927: };
1928: #endif /* POOL_SUBPAGE */
1929:
1930: /*
1931: * We have at least three different resources for the same allocation and
1932: * each resource can be depleted. First, we have the ready elements in the
1933: * pool. Then we have the resource (typically a vm_map) for this allocator.
1934: * Finally, we have physical memory. Waiting for any of these can be
1935: * unnecessary when any other is freed, but the kernel doesn't support
1936: * sleeping on multiple wait channels, so we have to employ another strategy.
1937: *
1938: * The caller sleeps on the pool (so that it can be awakened when an item
1939: * is returned to the pool), but we set PA_WANT on the allocator. When a
1940: * page is returned to the allocator and PA_WANT is set, pool_allocator_free
1941: * will wake up all sleeping pools belonging to this allocator.
1942: *
1943: * XXX Thundering herd.
1944: */
1945: void *
1946: pool_allocator_alloc(struct pool *org, int flags)
1947: {
1948: struct pool_allocator *pa = org->pr_alloc;
1949: struct pool *pp, *start;
1950: int s, freed;
1951: void *res;
1952:
1953: do {
1954: if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
1955: return (res);
1.68 thorpej 1956: if ((flags & PR_WAITOK) == 0) {
1957: /*
1958: * We only run the drain hookhere if PR_NOWAIT.
1959: * In other cases, the hook will be run in
1960: * pool_reclaim().
1961: */
1962: if (org->pr_drain_hook != NULL) {
1963: (*org->pr_drain_hook)(org->pr_drain_hook_arg,
1964: flags);
1965: if ((res = (*pa->pa_alloc)(org, flags)) != NULL)
1966: return (res);
1967: }
1.66 thorpej 1968: break;
1.68 thorpej 1969: }
1.66 thorpej 1970:
1971: /*
1972: * Drain all pools, except "org", that use this
1973: * allocator. We do this to reclaim VA space.
1974: * pa_alloc is responsible for waiting for
1975: * physical memory.
1976: *
1977: * XXX We risk looping forever if start if someone
1978: * calls pool_destroy on "start". But there is no
1979: * other way to have potentially sleeping pool_reclaim,
1980: * non-sleeping locks on pool_allocator, and some
1981: * stirring of drained pools in the allocator.
1.68 thorpej 1982: *
1983: * XXX Maybe we should use pool_head_slock for locking
1984: * the allocators?
1.66 thorpej 1985: */
1986: freed = 0;
1987:
1988: s = splvm();
1989: simple_lock(&pa->pa_slock);
1990: pp = start = TAILQ_FIRST(&pa->pa_list);
1991: do {
1992: TAILQ_REMOVE(&pa->pa_list, pp, pr_alloc_list);
1993: TAILQ_INSERT_TAIL(&pa->pa_list, pp, pr_alloc_list);
1994: if (pp == org)
1995: continue;
1.73 thorpej 1996: simple_unlock(&pa->pa_slock);
1.66 thorpej 1997: freed = pool_reclaim(pp);
1.73 thorpej 1998: simple_lock(&pa->pa_slock);
1.66 thorpej 1999: } while ((pp = TAILQ_FIRST(&pa->pa_list)) != start &&
2000: freed == 0);
2001:
2002: if (freed == 0) {
2003: /*
2004: * We set PA_WANT here, the caller will most likely
2005: * sleep waiting for pages (if not, this won't hurt
2006: * that much), and there is no way to set this in
2007: * the caller without violating locking order.
2008: */
2009: pa->pa_flags |= PA_WANT;
2010: }
2011: simple_unlock(&pa->pa_slock);
2012: splx(s);
2013: } while (freed);
2014: return (NULL);
2015: }
2016:
2017: void
2018: pool_allocator_free(struct pool *pp, void *v)
2019: {
2020: struct pool_allocator *pa = pp->pr_alloc;
2021: int s;
2022:
2023: (*pa->pa_free)(pp, v);
2024:
2025: s = splvm();
2026: simple_lock(&pa->pa_slock);
2027: if ((pa->pa_flags & PA_WANT) == 0) {
2028: simple_unlock(&pa->pa_slock);
2029: splx(s);
2030: return;
2031: }
2032:
2033: TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
2034: simple_lock(&pp->pr_slock);
2035: if ((pp->pr_flags & PR_WANTED) != 0) {
2036: pp->pr_flags &= ~PR_WANTED;
2037: wakeup(pp);
2038: }
1.69 thorpej 2039: simple_unlock(&pp->pr_slock);
1.66 thorpej 2040: }
2041: pa->pa_flags &= ~PA_WANT;
2042: simple_unlock(&pa->pa_slock);
2043: splx(s);
2044: }
2045:
2046: void *
2047: pool_page_alloc(struct pool *pp, int flags)
2048: {
2049: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2050:
2051: return ((void *) uvm_km_alloc_poolpage(waitok));
2052: }
2053:
2054: void
2055: pool_page_free(struct pool *pp, void *v)
2056: {
2057:
2058: uvm_km_free_poolpage((vaddr_t) v);
2059: }
2060:
2061: #ifdef POOL_SUBPAGE
2062: /* Sub-page allocator, for machines with large hardware pages. */
2063: void *
2064: pool_subpage_alloc(struct pool *pp, int flags)
2065: {
2066:
2067: return (pool_get(&psppool, flags));
2068: }
2069:
2070: void
2071: pool_subpage_free(struct pool *pp, void *v)
2072: {
2073:
2074: pool_put(&psppool, v);
2075: }
2076:
2077: /* We don't provide a real nointr allocator. Maybe later. */
2078: void *
2079: pool_page_alloc_nointr(struct pool *pp, int flags)
2080: {
2081:
2082: return (pool_subpage_alloc(pp, flags));
2083: }
2084:
2085: void
2086: pool_page_free_nointr(struct pool *pp, void *v)
2087: {
2088:
2089: pool_subpage_free(pp, v);
2090: }
2091: #else
2092: void *
2093: pool_page_alloc_nointr(struct pool *pp, int flags)
2094: {
2095: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
2096:
2097: return ((void *) uvm_km_alloc_poolpage1(kernel_map,
2098: uvm.kernel_object, waitok));
2099: }
2100:
2101: void
2102: pool_page_free_nointr(struct pool *pp, void *v)
2103: {
2104:
2105: uvm_km_free_poolpage1(kernel_map, (vaddr_t) v);
2106: }
2107: #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