Annotation of src/sys/kern/subr_pool.c, Revision 1.179
1.179 ! mlelstv 1: /* $NetBSD: subr_pool.c,v 1.178 2009/12/30 18:57:17 elad Exp $ */
1.1 pk 2:
3: /*-
1.161 ad 4: * Copyright (c) 1997, 1999, 2000, 2002, 2007, 2008 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
1.134 ad 9: * Simulation Facility, NASA Ames Research Center, and by Andrew Doran.
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: *
20: * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
21: * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
22: * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
23: * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
24: * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
25: * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
26: * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
27: * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28: * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
29: * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
30: * POSSIBILITY OF SUCH DAMAGE.
31: */
1.64 lukem 32:
33: #include <sys/cdefs.h>
1.179 ! mlelstv 34: __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.178 2009/12/30 18:57:17 elad Exp $");
1.24 scottr 35:
1.141 yamt 36: #include "opt_ddb.h"
1.25 thorpej 37: #include "opt_pool.h"
1.24 scottr 38: #include "opt_poollog.h"
1.28 thorpej 39: #include "opt_lockdebug.h"
1.1 pk 40:
41: #include <sys/param.h>
42: #include <sys/systm.h>
1.135 yamt 43: #include <sys/bitops.h>
1.1 pk 44: #include <sys/proc.h>
45: #include <sys/errno.h>
46: #include <sys/kernel.h>
47: #include <sys/malloc.h>
48: #include <sys/pool.h>
1.20 thorpej 49: #include <sys/syslog.h>
1.125 ad 50: #include <sys/debug.h>
1.134 ad 51: #include <sys/lockdebug.h>
52: #include <sys/xcall.h>
53: #include <sys/cpu.h>
1.145 ad 54: #include <sys/atomic.h>
1.3 pk 55:
56: #include <uvm/uvm.h>
57:
1.1 pk 58: /*
59: * Pool resource management utility.
1.3 pk 60: *
1.88 chs 61: * Memory is allocated in pages which are split into pieces according to
62: * the pool item size. Each page is kept on one of three lists in the
63: * pool structure: `pr_emptypages', `pr_fullpages' and `pr_partpages',
64: * for empty, full and partially-full pages respectively. The individual
65: * pool items are on a linked list headed by `ph_itemlist' in each page
66: * header. The memory for building the page list is either taken from
67: * the allocated pages themselves (for small pool items) or taken from
68: * an internal pool of page headers (`phpool').
1.1 pk 69: */
70:
1.3 pk 71: /* List of all pools */
1.173 rmind 72: static TAILQ_HEAD(, pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head);
1.134 ad 73:
1.3 pk 74: /* Private pool for page header structures */
1.97 yamt 75: #define PHPOOL_MAX 8
76: static struct pool phpool[PHPOOL_MAX];
1.135 yamt 77: #define PHPOOL_FREELIST_NELEM(idx) \
78: (((idx) == 0) ? 0 : BITMAP_SIZE * (1 << (idx)))
1.3 pk 79:
1.62 bjh21 80: #ifdef POOL_SUBPAGE
81: /* Pool of subpages for use by normal pools. */
82: static struct pool psppool;
83: #endif
84:
1.117 yamt 85: static SLIST_HEAD(, pool_allocator) pa_deferinitq =
86: SLIST_HEAD_INITIALIZER(pa_deferinitq);
87:
1.98 yamt 88: static void *pool_page_alloc_meta(struct pool *, int);
89: static void pool_page_free_meta(struct pool *, void *);
90:
91: /* allocator for pool metadata */
1.134 ad 92: struct pool_allocator pool_allocator_meta = {
1.117 yamt 93: pool_page_alloc_meta, pool_page_free_meta,
94: .pa_backingmapptr = &kmem_map,
1.98 yamt 95: };
96:
1.3 pk 97: /* # of seconds to retain page after last use */
98: int pool_inactive_time = 10;
99:
100: /* Next candidate for drainage (see pool_drain()) */
1.23 thorpej 101: static struct pool *drainpp;
102:
1.134 ad 103: /* This lock protects both pool_head and drainpp. */
104: static kmutex_t pool_head_lock;
105: static kcondvar_t pool_busy;
1.3 pk 106:
1.178 elad 107: /* This lock protects initialization of a potentially shared pool allocator */
108: static kmutex_t pool_allocator_lock;
109:
1.135 yamt 110: typedef uint32_t pool_item_bitmap_t;
111: #define BITMAP_SIZE (CHAR_BIT * sizeof(pool_item_bitmap_t))
112: #define BITMAP_MASK (BITMAP_SIZE - 1)
1.99 yamt 113:
1.3 pk 114: struct pool_item_header {
115: /* Page headers */
1.88 chs 116: LIST_ENTRY(pool_item_header)
1.3 pk 117: ph_pagelist; /* pool page list */
1.88 chs 118: SPLAY_ENTRY(pool_item_header)
119: ph_node; /* Off-page page headers */
1.128 christos 120: void * ph_page; /* this page's address */
1.151 yamt 121: uint32_t ph_time; /* last referenced */
1.135 yamt 122: uint16_t ph_nmissing; /* # of chunks in use */
1.141 yamt 123: uint16_t ph_off; /* start offset in page */
1.97 yamt 124: union {
125: /* !PR_NOTOUCH */
126: struct {
1.102 chs 127: LIST_HEAD(, pool_item)
1.97 yamt 128: phu_itemlist; /* chunk list for this page */
129: } phu_normal;
130: /* PR_NOTOUCH */
131: struct {
1.141 yamt 132: pool_item_bitmap_t phu_bitmap[1];
1.97 yamt 133: } phu_notouch;
134: } ph_u;
1.3 pk 135: };
1.97 yamt 136: #define ph_itemlist ph_u.phu_normal.phu_itemlist
1.135 yamt 137: #define ph_bitmap ph_u.phu_notouch.phu_bitmap
1.3 pk 138:
1.1 pk 139: struct pool_item {
1.3 pk 140: #ifdef DIAGNOSTIC
1.82 thorpej 141: u_int pi_magic;
1.33 chs 142: #endif
1.134 ad 143: #define PI_MAGIC 0xdeaddeadU
1.3 pk 144: /* Other entries use only this list entry */
1.102 chs 145: LIST_ENTRY(pool_item) pi_list;
1.3 pk 146: };
147:
1.53 thorpej 148: #define POOL_NEEDS_CATCHUP(pp) \
149: ((pp)->pr_nitems < (pp)->pr_minitems)
150:
1.43 thorpej 151: /*
152: * Pool cache management.
153: *
154: * Pool caches provide a way for constructed objects to be cached by the
155: * pool subsystem. This can lead to performance improvements by avoiding
156: * needless object construction/destruction; it is deferred until absolutely
157: * necessary.
158: *
1.134 ad 159: * Caches are grouped into cache groups. Each cache group references up
160: * to PCG_NUMOBJECTS constructed objects. When a cache allocates an
161: * object from the pool, it calls the object's constructor and places it
162: * into a cache group. When a cache group frees an object back to the
163: * pool, it first calls the object's destructor. This allows the object
164: * to persist in constructed form while freed to the cache.
165: *
166: * The pool references each cache, so that when a pool is drained by the
167: * pagedaemon, it can drain each individual cache as well. Each time a
168: * cache is drained, the most idle cache group is freed to the pool in
169: * its entirety.
1.43 thorpej 170: *
171: * Pool caches are layed on top of pools. By layering them, we can avoid
172: * the complexity of cache management for pools which would not benefit
173: * from it.
174: */
175:
1.142 ad 176: static struct pool pcg_normal_pool;
177: static struct pool pcg_large_pool;
1.134 ad 178: static struct pool cache_pool;
179: static struct pool cache_cpu_pool;
1.3 pk 180:
1.145 ad 181: /* List of all caches. */
182: TAILQ_HEAD(,pool_cache) pool_cache_head =
183: TAILQ_HEAD_INITIALIZER(pool_cache_head);
184:
1.162 ad 185: int pool_cache_disable; /* global disable for caching */
1.169 yamt 186: static const pcg_t pcg_dummy; /* zero sized: always empty, yet always full */
1.145 ad 187:
1.162 ad 188: static bool pool_cache_put_slow(pool_cache_cpu_t *, int,
189: void *);
190: static bool pool_cache_get_slow(pool_cache_cpu_t *, int,
191: void **, paddr_t *, int);
1.134 ad 192: static void pool_cache_cpu_init1(struct cpu_info *, pool_cache_t);
193: static void pool_cache_invalidate_groups(pool_cache_t, pcg_t *);
1.175 jym 194: static void pool_cache_invalidate_cpu(pool_cache_t, u_int);
1.134 ad 195: static void pool_cache_xcall(pool_cache_t);
1.3 pk 196:
1.42 thorpej 197: static int pool_catchup(struct pool *);
1.128 christos 198: static void pool_prime_page(struct pool *, void *,
1.55 thorpej 199: struct pool_item_header *);
1.88 chs 200: static void pool_update_curpage(struct pool *);
1.66 thorpej 201:
1.113 yamt 202: static int pool_grow(struct pool *, int);
1.117 yamt 203: static void *pool_allocator_alloc(struct pool *, int);
204: static void pool_allocator_free(struct pool *, void *);
1.3 pk 205:
1.97 yamt 206: static void pool_print_pagelist(struct pool *, struct pool_pagelist *,
1.88 chs 207: void (*)(const char *, ...));
1.42 thorpej 208: static void pool_print1(struct pool *, const char *,
209: void (*)(const char *, ...));
1.3 pk 210:
1.88 chs 211: static int pool_chk_page(struct pool *, const char *,
212: struct pool_item_header *);
213:
1.3 pk 214: /*
1.52 thorpej 215: * Pool log entry. An array of these is allocated in pool_init().
1.3 pk 216: */
217: struct pool_log {
218: const char *pl_file;
219: long pl_line;
220: int pl_action;
1.25 thorpej 221: #define PRLOG_GET 1
222: #define PRLOG_PUT 2
1.3 pk 223: void *pl_addr;
1.1 pk 224: };
225:
1.86 matt 226: #ifdef POOL_DIAGNOSTIC
1.3 pk 227: /* Number of entries in pool log buffers */
1.17 thorpej 228: #ifndef POOL_LOGSIZE
229: #define POOL_LOGSIZE 10
230: #endif
231:
232: int pool_logsize = POOL_LOGSIZE;
1.1 pk 233:
1.110 perry 234: static inline void
1.42 thorpej 235: pr_log(struct pool *pp, void *v, int action, const char *file, long line)
1.3 pk 236: {
1.179 ! mlelstv 237: int n;
1.3 pk 238: struct pool_log *pl;
239:
1.20 thorpej 240: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 241: return;
242:
1.179 ! mlelstv 243: if (pp->pr_log == NULL) {
! 244: if (kmem_map != NULL)
! 245: pp->pr_log = malloc(
! 246: pool_logsize * sizeof(struct pool_log),
! 247: M_TEMP, M_NOWAIT | M_ZERO);
! 248: if (pp->pr_log == NULL)
! 249: return;
! 250: pp->pr_curlogentry = 0;
! 251: pp->pr_logsize = pool_logsize;
! 252: }
! 253:
1.3 pk 254: /*
255: * Fill in the current entry. Wrap around and overwrite
256: * the oldest entry if necessary.
257: */
1.179 ! mlelstv 258: n = pp->pr_curlogentry;
1.3 pk 259: pl = &pp->pr_log[n];
260: pl->pl_file = file;
261: pl->pl_line = line;
262: pl->pl_action = action;
263: pl->pl_addr = v;
264: if (++n >= pp->pr_logsize)
265: n = 0;
266: pp->pr_curlogentry = n;
267: }
268:
269: static void
1.42 thorpej 270: pr_printlog(struct pool *pp, struct pool_item *pi,
271: void (*pr)(const char *, ...))
1.3 pk 272: {
273: int i = pp->pr_logsize;
274: int n = pp->pr_curlogentry;
275:
1.179 ! mlelstv 276: if (pp->pr_log == NULL)
1.3 pk 277: return;
278:
279: /*
280: * Print all entries in this pool's log.
281: */
282: while (i-- > 0) {
283: struct pool_log *pl = &pp->pr_log[n];
284: if (pl->pl_action != 0) {
1.25 thorpej 285: if (pi == NULL || pi == pl->pl_addr) {
286: (*pr)("\tlog entry %d:\n", i);
287: (*pr)("\t\taction = %s, addr = %p\n",
288: pl->pl_action == PRLOG_GET ? "get" : "put",
289: pl->pl_addr);
290: (*pr)("\t\tfile: %s at line %lu\n",
291: pl->pl_file, pl->pl_line);
292: }
1.3 pk 293: }
294: if (++n >= pp->pr_logsize)
295: n = 0;
296: }
297: }
1.25 thorpej 298:
1.110 perry 299: static inline void
1.42 thorpej 300: pr_enter(struct pool *pp, const char *file, long line)
1.25 thorpej 301: {
302:
1.34 thorpej 303: if (__predict_false(pp->pr_entered_file != NULL)) {
1.25 thorpej 304: printf("pool %s: reentrancy at file %s line %ld\n",
305: pp->pr_wchan, file, line);
306: printf(" previous entry at file %s line %ld\n",
307: pp->pr_entered_file, pp->pr_entered_line);
308: panic("pr_enter");
309: }
310:
311: pp->pr_entered_file = file;
312: pp->pr_entered_line = line;
313: }
314:
1.110 perry 315: static inline void
1.42 thorpej 316: pr_leave(struct pool *pp)
1.25 thorpej 317: {
318:
1.34 thorpej 319: if (__predict_false(pp->pr_entered_file == NULL)) {
1.25 thorpej 320: printf("pool %s not entered?\n", pp->pr_wchan);
321: panic("pr_leave");
322: }
323:
324: pp->pr_entered_file = NULL;
325: pp->pr_entered_line = 0;
326: }
327:
1.110 perry 328: static inline void
1.42 thorpej 329: pr_enter_check(struct pool *pp, void (*pr)(const char *, ...))
1.25 thorpej 330: {
331:
332: if (pp->pr_entered_file != NULL)
333: (*pr)("\n\tcurrently entered from file %s line %ld\n",
334: pp->pr_entered_file, pp->pr_entered_line);
335: }
1.3 pk 336: #else
1.25 thorpej 337: #define pr_log(pp, v, action, file, line)
338: #define pr_printlog(pp, pi, pr)
339: #define pr_enter(pp, file, line)
340: #define pr_leave(pp)
341: #define pr_enter_check(pp, pr)
1.59 thorpej 342: #endif /* POOL_DIAGNOSTIC */
1.3 pk 343:
1.135 yamt 344: static inline unsigned int
1.97 yamt 345: pr_item_notouch_index(const struct pool *pp, const struct pool_item_header *ph,
346: const void *v)
347: {
348: const char *cp = v;
1.135 yamt 349: unsigned int idx;
1.97 yamt 350:
351: KASSERT(pp->pr_roflags & PR_NOTOUCH);
1.128 christos 352: idx = (cp - (char *)ph->ph_page - ph->ph_off) / pp->pr_size;
1.97 yamt 353: KASSERT(idx < pp->pr_itemsperpage);
354: return idx;
355: }
356:
1.110 perry 357: static inline void
1.97 yamt 358: pr_item_notouch_put(const struct pool *pp, struct pool_item_header *ph,
359: void *obj)
360: {
1.135 yamt 361: unsigned int idx = pr_item_notouch_index(pp, ph, obj);
362: pool_item_bitmap_t *bitmap = ph->ph_bitmap + (idx / BITMAP_SIZE);
363: pool_item_bitmap_t mask = 1 << (idx & BITMAP_MASK);
1.97 yamt 364:
1.135 yamt 365: KASSERT((*bitmap & mask) == 0);
366: *bitmap |= mask;
1.97 yamt 367: }
368:
1.110 perry 369: static inline void *
1.97 yamt 370: pr_item_notouch_get(const struct pool *pp, struct pool_item_header *ph)
371: {
1.135 yamt 372: pool_item_bitmap_t *bitmap = ph->ph_bitmap;
373: unsigned int idx;
374: int i;
1.97 yamt 375:
1.135 yamt 376: for (i = 0; ; i++) {
377: int bit;
1.97 yamt 378:
1.135 yamt 379: KASSERT((i * BITMAP_SIZE) < pp->pr_itemsperpage);
380: bit = ffs32(bitmap[i]);
381: if (bit) {
382: pool_item_bitmap_t mask;
383:
384: bit--;
385: idx = (i * BITMAP_SIZE) + bit;
386: mask = 1 << bit;
387: KASSERT((bitmap[i] & mask) != 0);
388: bitmap[i] &= ~mask;
389: break;
390: }
391: }
392: KASSERT(idx < pp->pr_itemsperpage);
1.128 christos 393: return (char *)ph->ph_page + ph->ph_off + idx * pp->pr_size;
1.97 yamt 394: }
395:
1.135 yamt 396: static inline void
1.141 yamt 397: pr_item_notouch_init(const struct pool *pp, struct pool_item_header *ph)
1.135 yamt 398: {
399: pool_item_bitmap_t *bitmap = ph->ph_bitmap;
400: const int n = howmany(pp->pr_itemsperpage, BITMAP_SIZE);
401: int i;
402:
403: for (i = 0; i < n; i++) {
404: bitmap[i] = (pool_item_bitmap_t)-1;
405: }
406: }
407:
1.110 perry 408: static inline int
1.88 chs 409: phtree_compare(struct pool_item_header *a, struct pool_item_header *b)
410: {
1.121 yamt 411:
412: /*
413: * we consider pool_item_header with smaller ph_page bigger.
414: * (this unnatural ordering is for the benefit of pr_find_pagehead.)
415: */
416:
1.88 chs 417: if (a->ph_page < b->ph_page)
1.121 yamt 418: return (1);
419: else if (a->ph_page > b->ph_page)
1.88 chs 420: return (-1);
421: else
422: return (0);
423: }
424:
425: SPLAY_PROTOTYPE(phtree, pool_item_header, ph_node, phtree_compare);
426: SPLAY_GENERATE(phtree, pool_item_header, ph_node, phtree_compare);
427:
1.141 yamt 428: static inline struct pool_item_header *
429: pr_find_pagehead_noalign(struct pool *pp, void *v)
430: {
431: struct pool_item_header *ph, tmp;
432:
433: tmp.ph_page = (void *)(uintptr_t)v;
434: ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
435: if (ph == NULL) {
436: ph = SPLAY_ROOT(&pp->pr_phtree);
437: if (ph != NULL && phtree_compare(&tmp, ph) >= 0) {
438: ph = SPLAY_NEXT(phtree, &pp->pr_phtree, ph);
439: }
440: KASSERT(ph == NULL || phtree_compare(&tmp, ph) < 0);
441: }
442:
443: return ph;
444: }
445:
1.3 pk 446: /*
1.121 yamt 447: * Return the pool page header based on item address.
1.3 pk 448: */
1.110 perry 449: static inline struct pool_item_header *
1.121 yamt 450: pr_find_pagehead(struct pool *pp, void *v)
1.3 pk 451: {
1.88 chs 452: struct pool_item_header *ph, tmp;
1.3 pk 453:
1.121 yamt 454: if ((pp->pr_roflags & PR_NOALIGN) != 0) {
1.141 yamt 455: ph = pr_find_pagehead_noalign(pp, v);
1.121 yamt 456: } else {
1.128 christos 457: void *page =
458: (void *)((uintptr_t)v & pp->pr_alloc->pa_pagemask);
1.121 yamt 459:
460: if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
1.128 christos 461: ph = (struct pool_item_header *)((char *)page + pp->pr_phoffset);
1.121 yamt 462: } else {
463: tmp.ph_page = page;
464: ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
465: }
466: }
1.3 pk 467:
1.121 yamt 468: KASSERT(ph == NULL || ((pp->pr_roflags & PR_PHINPAGE) != 0) ||
1.128 christos 469: ((char *)ph->ph_page <= (char *)v &&
470: (char *)v < (char *)ph->ph_page + pp->pr_alloc->pa_pagesz));
1.88 chs 471: return ph;
1.3 pk 472: }
473:
1.101 thorpej 474: static void
475: pr_pagelist_free(struct pool *pp, struct pool_pagelist *pq)
476: {
477: struct pool_item_header *ph;
478:
479: while ((ph = LIST_FIRST(pq)) != NULL) {
480: LIST_REMOVE(ph, ph_pagelist);
481: pool_allocator_free(pp, ph->ph_page);
1.134 ad 482: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1.101 thorpej 483: pool_put(pp->pr_phpool, ph);
484: }
485: }
486:
1.3 pk 487: /*
488: * Remove a page from the pool.
489: */
1.110 perry 490: static inline void
1.61 chs 491: pr_rmpage(struct pool *pp, struct pool_item_header *ph,
492: struct pool_pagelist *pq)
1.3 pk 493: {
494:
1.134 ad 495: KASSERT(mutex_owned(&pp->pr_lock));
1.91 yamt 496:
1.3 pk 497: /*
1.7 thorpej 498: * If the page was idle, decrement the idle page count.
1.3 pk 499: */
1.6 thorpej 500: if (ph->ph_nmissing == 0) {
501: #ifdef DIAGNOSTIC
502: if (pp->pr_nidle == 0)
503: panic("pr_rmpage: nidle inconsistent");
1.20 thorpej 504: if (pp->pr_nitems < pp->pr_itemsperpage)
505: panic("pr_rmpage: nitems inconsistent");
1.6 thorpej 506: #endif
507: pp->pr_nidle--;
508: }
1.7 thorpej 509:
1.20 thorpej 510: pp->pr_nitems -= pp->pr_itemsperpage;
511:
1.7 thorpej 512: /*
1.101 thorpej 513: * Unlink the page from the pool and queue it for release.
1.7 thorpej 514: */
1.88 chs 515: LIST_REMOVE(ph, ph_pagelist);
1.91 yamt 516: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
517: SPLAY_REMOVE(phtree, &pp->pr_phtree, ph);
1.101 thorpej 518: LIST_INSERT_HEAD(pq, ph, ph_pagelist);
519:
1.7 thorpej 520: pp->pr_npages--;
521: pp->pr_npagefree++;
1.6 thorpej 522:
1.88 chs 523: pool_update_curpage(pp);
1.3 pk 524: }
525:
1.126 thorpej 526: static bool
1.117 yamt 527: pa_starved_p(struct pool_allocator *pa)
528: {
529:
530: if (pa->pa_backingmap != NULL) {
531: return vm_map_starved_p(pa->pa_backingmap);
532: }
1.127 thorpej 533: return false;
1.117 yamt 534: }
535:
536: static int
1.124 yamt 537: pool_reclaim_callback(struct callback_entry *ce, void *obj, void *arg)
1.117 yamt 538: {
539: struct pool *pp = obj;
540: struct pool_allocator *pa = pp->pr_alloc;
541:
542: KASSERT(&pp->pr_reclaimerentry == ce);
543: pool_reclaim(pp);
544: if (!pa_starved_p(pa)) {
545: return CALLBACK_CHAIN_ABORT;
546: }
547: return CALLBACK_CHAIN_CONTINUE;
548: }
549:
550: static void
551: pool_reclaim_register(struct pool *pp)
552: {
553: struct vm_map *map = pp->pr_alloc->pa_backingmap;
554: int s;
555:
556: if (map == NULL) {
557: return;
558: }
559:
560: s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */
561: callback_register(&vm_map_to_kernel(map)->vmk_reclaim_callback,
562: &pp->pr_reclaimerentry, pp, pool_reclaim_callback);
563: splx(s);
564: }
565:
566: static void
567: pool_reclaim_unregister(struct pool *pp)
568: {
569: struct vm_map *map = pp->pr_alloc->pa_backingmap;
570: int s;
571:
572: if (map == NULL) {
573: return;
574: }
575:
576: s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */
577: callback_unregister(&vm_map_to_kernel(map)->vmk_reclaim_callback,
578: &pp->pr_reclaimerentry);
579: splx(s);
580: }
581:
582: static void
583: pa_reclaim_register(struct pool_allocator *pa)
584: {
585: struct vm_map *map = *pa->pa_backingmapptr;
586: struct pool *pp;
587:
588: KASSERT(pa->pa_backingmap == NULL);
589: if (map == NULL) {
590: SLIST_INSERT_HEAD(&pa_deferinitq, pa, pa_q);
591: return;
592: }
593: pa->pa_backingmap = map;
594: TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
595: pool_reclaim_register(pp);
596: }
597: }
598:
1.3 pk 599: /*
1.94 simonb 600: * Initialize all the pools listed in the "pools" link set.
601: */
602: void
1.117 yamt 603: pool_subsystem_init(void)
1.94 simonb 604: {
1.117 yamt 605: struct pool_allocator *pa;
1.94 simonb 606:
1.134 ad 607: mutex_init(&pool_head_lock, MUTEX_DEFAULT, IPL_NONE);
1.179 ! mlelstv 608: mutex_init(&pool_allocator_lock, MUTEX_DEFAULT, IPL_NONE);
1.134 ad 609: cv_init(&pool_busy, "poolbusy");
610:
1.117 yamt 611: while ((pa = SLIST_FIRST(&pa_deferinitq)) != NULL) {
612: KASSERT(pa->pa_backingmapptr != NULL);
613: KASSERT(*pa->pa_backingmapptr != NULL);
614: SLIST_REMOVE_HEAD(&pa_deferinitq, pa_q);
615: pa_reclaim_register(pa);
616: }
1.134 ad 617:
1.156 ad 618: pool_init(&cache_pool, sizeof(struct pool_cache), coherency_unit,
1.134 ad 619: 0, 0, "pcache", &pool_allocator_nointr, IPL_NONE);
620:
1.156 ad 621: pool_init(&cache_cpu_pool, sizeof(pool_cache_cpu_t), coherency_unit,
1.134 ad 622: 0, 0, "pcachecpu", &pool_allocator_nointr, IPL_NONE);
1.94 simonb 623: }
624:
625: /*
1.3 pk 626: * Initialize the given pool resource structure.
627: *
628: * We export this routine to allow other kernel parts to declare
629: * static pools that must be initialized before malloc() is available.
630: */
631: void
1.42 thorpej 632: pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
1.129 ad 633: const char *wchan, struct pool_allocator *palloc, int ipl)
1.3 pk 634: {
1.116 simonb 635: struct pool *pp1;
1.92 enami 636: size_t trysize, phsize;
1.134 ad 637: int off, slack;
1.3 pk 638:
1.116 simonb 639: #ifdef DEBUG
640: /*
641: * Check that the pool hasn't already been initialised and
642: * added to the list of all pools.
643: */
1.145 ad 644: TAILQ_FOREACH(pp1, &pool_head, pr_poollist) {
1.116 simonb 645: if (pp == pp1)
646: panic("pool_init: pool %s already initialised",
647: wchan);
648: }
649: #endif
650:
1.25 thorpej 651: #ifdef POOL_DIAGNOSTIC
652: /*
653: * Always log if POOL_DIAGNOSTIC is defined.
654: */
655: if (pool_logsize != 0)
656: flags |= PR_LOGGING;
657: #endif
658:
1.66 thorpej 659: if (palloc == NULL)
660: palloc = &pool_allocator_kmem;
1.112 bjh21 661: #ifdef POOL_SUBPAGE
662: if (size > palloc->pa_pagesz) {
663: if (palloc == &pool_allocator_kmem)
664: palloc = &pool_allocator_kmem_fullpage;
665: else if (palloc == &pool_allocator_nointr)
666: palloc = &pool_allocator_nointr_fullpage;
667: }
1.66 thorpej 668: #endif /* POOL_SUBPAGE */
1.178 elad 669: mutex_enter(&pool_allocator_lock);
670: if (palloc->pa_refcnt++ == 0) {
1.112 bjh21 671: if (palloc->pa_pagesz == 0)
1.66 thorpej 672: palloc->pa_pagesz = PAGE_SIZE;
673:
674: TAILQ_INIT(&palloc->pa_list);
675:
1.134 ad 676: mutex_init(&palloc->pa_lock, MUTEX_DEFAULT, IPL_VM);
1.66 thorpej 677: palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
678: palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
1.117 yamt 679:
680: if (palloc->pa_backingmapptr != NULL) {
681: pa_reclaim_register(palloc);
682: }
1.4 thorpej 683: }
1.178 elad 684: mutex_exit(&pool_allocator_lock);
1.3 pk 685:
686: if (align == 0)
687: align = ALIGN(1);
1.14 thorpej 688:
1.120 yamt 689: if ((flags & PR_NOTOUCH) == 0 && size < sizeof(struct pool_item))
1.14 thorpej 690: size = sizeof(struct pool_item);
1.3 pk 691:
1.78 thorpej 692: size = roundup(size, align);
1.66 thorpej 693: #ifdef DIAGNOSTIC
694: if (size > palloc->pa_pagesz)
1.121 yamt 695: panic("pool_init: pool item size (%zu) too large", size);
1.66 thorpej 696: #endif
1.35 pk 697:
1.3 pk 698: /*
699: * Initialize the pool structure.
700: */
1.88 chs 701: LIST_INIT(&pp->pr_emptypages);
702: LIST_INIT(&pp->pr_fullpages);
703: LIST_INIT(&pp->pr_partpages);
1.134 ad 704: pp->pr_cache = NULL;
1.3 pk 705: pp->pr_curpage = NULL;
706: pp->pr_npages = 0;
707: pp->pr_minitems = 0;
708: pp->pr_minpages = 0;
709: pp->pr_maxpages = UINT_MAX;
1.20 thorpej 710: pp->pr_roflags = flags;
711: pp->pr_flags = 0;
1.35 pk 712: pp->pr_size = size;
1.3 pk 713: pp->pr_align = align;
714: pp->pr_wchan = wchan;
1.66 thorpej 715: pp->pr_alloc = palloc;
1.20 thorpej 716: pp->pr_nitems = 0;
717: pp->pr_nout = 0;
718: pp->pr_hardlimit = UINT_MAX;
719: pp->pr_hardlimit_warning = NULL;
1.31 thorpej 720: pp->pr_hardlimit_ratecap.tv_sec = 0;
721: pp->pr_hardlimit_ratecap.tv_usec = 0;
722: pp->pr_hardlimit_warning_last.tv_sec = 0;
723: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.68 thorpej 724: pp->pr_drain_hook = NULL;
725: pp->pr_drain_hook_arg = NULL;
1.125 ad 726: pp->pr_freecheck = NULL;
1.3 pk 727:
728: /*
729: * Decide whether to put the page header off page to avoid
1.92 enami 730: * wasting too large a part of the page or too big item.
731: * Off-page page headers go on a hash table, so we can match
732: * a returned item with its header based on the page address.
733: * We use 1/16 of the page size and about 8 times of the item
734: * size as the threshold (XXX: tune)
735: *
736: * However, we'll put the header into the page if we can put
737: * it without wasting any items.
738: *
739: * Silently enforce `0 <= ioff < align'.
1.3 pk 740: */
1.92 enami 741: pp->pr_itemoffset = ioff %= align;
742: /* See the comment below about reserved bytes. */
743: trysize = palloc->pa_pagesz - ((align - ioff) % align);
744: phsize = ALIGN(sizeof(struct pool_item_header));
1.121 yamt 745: if ((pp->pr_roflags & (PR_NOTOUCH | PR_NOALIGN)) == 0 &&
1.97 yamt 746: (pp->pr_size < MIN(palloc->pa_pagesz / 16, phsize << 3) ||
747: trysize / pp->pr_size == (trysize - phsize) / pp->pr_size)) {
1.3 pk 748: /* Use the end of the page for the page header */
1.20 thorpej 749: pp->pr_roflags |= PR_PHINPAGE;
1.92 enami 750: pp->pr_phoffset = off = palloc->pa_pagesz - phsize;
1.2 pk 751: } else {
1.3 pk 752: /* The page header will be taken from our page header pool */
753: pp->pr_phoffset = 0;
1.66 thorpej 754: off = palloc->pa_pagesz;
1.88 chs 755: SPLAY_INIT(&pp->pr_phtree);
1.2 pk 756: }
1.1 pk 757:
1.3 pk 758: /*
759: * Alignment is to take place at `ioff' within the item. This means
760: * we must reserve up to `align - 1' bytes on the page to allow
761: * appropriate positioning of each item.
762: */
763: pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
1.43 thorpej 764: KASSERT(pp->pr_itemsperpage != 0);
1.97 yamt 765: if ((pp->pr_roflags & PR_NOTOUCH)) {
766: int idx;
767:
768: for (idx = 0; pp->pr_itemsperpage > PHPOOL_FREELIST_NELEM(idx);
769: idx++) {
770: /* nothing */
771: }
772: if (idx >= PHPOOL_MAX) {
773: /*
774: * if you see this panic, consider to tweak
775: * PHPOOL_MAX and PHPOOL_FREELIST_NELEM.
776: */
777: panic("%s: too large itemsperpage(%d) for PR_NOTOUCH",
778: pp->pr_wchan, pp->pr_itemsperpage);
779: }
780: pp->pr_phpool = &phpool[idx];
781: } else if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
782: pp->pr_phpool = &phpool[0];
783: }
784: #if defined(DIAGNOSTIC)
785: else {
786: pp->pr_phpool = NULL;
787: }
788: #endif
1.3 pk 789:
790: /*
791: * Use the slack between the chunks and the page header
792: * for "cache coloring".
793: */
794: slack = off - pp->pr_itemsperpage * pp->pr_size;
795: pp->pr_maxcolor = (slack / align) * align;
796: pp->pr_curcolor = 0;
797:
798: pp->pr_nget = 0;
799: pp->pr_nfail = 0;
800: pp->pr_nput = 0;
801: pp->pr_npagealloc = 0;
802: pp->pr_npagefree = 0;
1.1 pk 803: pp->pr_hiwat = 0;
1.8 thorpej 804: pp->pr_nidle = 0;
1.134 ad 805: pp->pr_refcnt = 0;
1.3 pk 806:
1.179 ! mlelstv 807: pp->pr_log = NULL;
1.25 thorpej 808:
809: pp->pr_entered_file = NULL;
810: pp->pr_entered_line = 0;
1.3 pk 811:
1.157 ad 812: mutex_init(&pp->pr_lock, MUTEX_DEFAULT, ipl);
1.134 ad 813: cv_init(&pp->pr_cv, wchan);
814: pp->pr_ipl = ipl;
1.1 pk 815:
1.3 pk 816: /*
1.43 thorpej 817: * Initialize private page header pool and cache magazine pool if we
818: * haven't done so yet.
1.23 thorpej 819: * XXX LOCKING.
1.3 pk 820: */
1.97 yamt 821: if (phpool[0].pr_size == 0) {
822: int idx;
823: for (idx = 0; idx < PHPOOL_MAX; idx++) {
824: static char phpool_names[PHPOOL_MAX][6+1+6+1];
825: int nelem;
826: size_t sz;
827:
828: nelem = PHPOOL_FREELIST_NELEM(idx);
829: snprintf(phpool_names[idx], sizeof(phpool_names[idx]),
830: "phpool-%d", nelem);
831: sz = sizeof(struct pool_item_header);
832: if (nelem) {
1.135 yamt 833: sz = offsetof(struct pool_item_header,
834: ph_bitmap[howmany(nelem, BITMAP_SIZE)]);
1.97 yamt 835: }
836: pool_init(&phpool[idx], sz, 0, 0, 0,
1.129 ad 837: phpool_names[idx], &pool_allocator_meta, IPL_VM);
1.97 yamt 838: }
1.62 bjh21 839: #ifdef POOL_SUBPAGE
840: pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
1.129 ad 841: PR_RECURSIVE, "psppool", &pool_allocator_meta, IPL_VM);
1.62 bjh21 842: #endif
1.142 ad 843:
844: size = sizeof(pcg_t) +
845: (PCG_NOBJECTS_NORMAL - 1) * sizeof(pcgpair_t);
1.156 ad 846: pool_init(&pcg_normal_pool, size, coherency_unit, 0, 0,
1.142 ad 847: "pcgnormal", &pool_allocator_meta, IPL_VM);
848:
849: size = sizeof(pcg_t) +
850: (PCG_NOBJECTS_LARGE - 1) * sizeof(pcgpair_t);
1.156 ad 851: pool_init(&pcg_large_pool, size, coherency_unit, 0, 0,
1.142 ad 852: "pcglarge", &pool_allocator_meta, IPL_VM);
1.1 pk 853: }
854:
1.145 ad 855: /* Insert into the list of all pools. */
856: if (__predict_true(!cold))
1.134 ad 857: mutex_enter(&pool_head_lock);
1.145 ad 858: TAILQ_FOREACH(pp1, &pool_head, pr_poollist) {
859: if (strcmp(pp1->pr_wchan, pp->pr_wchan) > 0)
860: break;
861: }
862: if (pp1 == NULL)
863: TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist);
864: else
865: TAILQ_INSERT_BEFORE(pp1, pp, pr_poollist);
866: if (__predict_true(!cold))
1.134 ad 867: mutex_exit(&pool_head_lock);
868:
1.167 skrll 869: /* Insert this into the list of pools using this allocator. */
1.145 ad 870: if (__predict_true(!cold))
1.134 ad 871: mutex_enter(&palloc->pa_lock);
1.145 ad 872: TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
873: if (__predict_true(!cold))
1.134 ad 874: mutex_exit(&palloc->pa_lock);
1.66 thorpej 875:
1.117 yamt 876: pool_reclaim_register(pp);
1.1 pk 877: }
878:
879: /*
880: * De-commision a pool resource.
881: */
882: void
1.42 thorpej 883: pool_destroy(struct pool *pp)
1.1 pk 884: {
1.101 thorpej 885: struct pool_pagelist pq;
1.3 pk 886: struct pool_item_header *ph;
1.43 thorpej 887:
1.101 thorpej 888: /* Remove from global pool list */
1.134 ad 889: mutex_enter(&pool_head_lock);
890: while (pp->pr_refcnt != 0)
891: cv_wait(&pool_busy, &pool_head_lock);
1.145 ad 892: TAILQ_REMOVE(&pool_head, pp, pr_poollist);
1.101 thorpej 893: if (drainpp == pp)
894: drainpp = NULL;
1.134 ad 895: mutex_exit(&pool_head_lock);
1.101 thorpej 896:
897: /* Remove this pool from its allocator's list of pools. */
1.117 yamt 898: pool_reclaim_unregister(pp);
1.134 ad 899: mutex_enter(&pp->pr_alloc->pa_lock);
1.66 thorpej 900: TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
1.134 ad 901: mutex_exit(&pp->pr_alloc->pa_lock);
1.66 thorpej 902:
1.178 elad 903: mutex_enter(&pool_allocator_lock);
904: if (--pp->pr_alloc->pa_refcnt == 0)
905: mutex_destroy(&pp->pr_alloc->pa_lock);
906: mutex_exit(&pool_allocator_lock);
907:
1.134 ad 908: mutex_enter(&pp->pr_lock);
1.101 thorpej 909:
1.134 ad 910: KASSERT(pp->pr_cache == NULL);
1.3 pk 911:
912: #ifdef DIAGNOSTIC
1.20 thorpej 913: if (pp->pr_nout != 0) {
1.25 thorpej 914: pr_printlog(pp, NULL, printf);
1.80 provos 915: panic("pool_destroy: pool busy: still out: %u",
1.20 thorpej 916: pp->pr_nout);
1.3 pk 917: }
918: #endif
1.1 pk 919:
1.101 thorpej 920: KASSERT(LIST_EMPTY(&pp->pr_fullpages));
921: KASSERT(LIST_EMPTY(&pp->pr_partpages));
922:
1.3 pk 923: /* Remove all pages */
1.101 thorpej 924: LIST_INIT(&pq);
1.88 chs 925: while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1.101 thorpej 926: pr_rmpage(pp, ph, &pq);
927:
1.134 ad 928: mutex_exit(&pp->pr_lock);
1.3 pk 929:
1.101 thorpej 930: pr_pagelist_free(pp, &pq);
1.3 pk 931:
1.59 thorpej 932: #ifdef POOL_DIAGNOSTIC
1.179 ! mlelstv 933: if (pp->pr_log != NULL) {
1.3 pk 934: free(pp->pr_log, M_TEMP);
1.179 ! mlelstv 935: pp->pr_log = NULL;
! 936: }
1.59 thorpej 937: #endif
1.134 ad 938:
939: cv_destroy(&pp->pr_cv);
940: mutex_destroy(&pp->pr_lock);
1.1 pk 941: }
942:
1.68 thorpej 943: void
944: pool_set_drain_hook(struct pool *pp, void (*fn)(void *, int), void *arg)
945: {
946:
947: /* XXX no locking -- must be used just after pool_init() */
948: #ifdef DIAGNOSTIC
949: if (pp->pr_drain_hook != NULL)
950: panic("pool_set_drain_hook(%s): already set", pp->pr_wchan);
951: #endif
952: pp->pr_drain_hook = fn;
953: pp->pr_drain_hook_arg = arg;
954: }
955:
1.88 chs 956: static struct pool_item_header *
1.128 christos 957: pool_alloc_item_header(struct pool *pp, void *storage, int flags)
1.55 thorpej 958: {
959: struct pool_item_header *ph;
960:
961: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
1.128 christos 962: ph = (struct pool_item_header *) ((char *)storage + pp->pr_phoffset);
1.134 ad 963: else
1.97 yamt 964: ph = pool_get(pp->pr_phpool, flags);
1.55 thorpej 965:
966: return (ph);
967: }
1.1 pk 968:
969: /*
1.134 ad 970: * Grab an item from the pool.
1.1 pk 971: */
1.3 pk 972: void *
1.59 thorpej 973: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 974: _pool_get(struct pool *pp, int flags, const char *file, long line)
1.56 sommerfe 975: #else
976: pool_get(struct pool *pp, int flags)
977: #endif
1.1 pk 978: {
979: struct pool_item *pi;
1.3 pk 980: struct pool_item_header *ph;
1.55 thorpej 981: void *v;
1.1 pk 982:
1.2 pk 983: #ifdef DIAGNOSTIC
1.95 atatat 984: if (__predict_false(pp->pr_itemsperpage == 0))
985: panic("pool_get: pool %p: pr_itemsperpage is zero, "
986: "pool not initialized?", pp);
1.84 thorpej 987: if (__predict_false(curlwp == NULL && doing_shutdown == 0 &&
1.37 sommerfe 988: (flags & PR_WAITOK) != 0))
1.77 matt 989: panic("pool_get: %s: must have NOWAIT", pp->pr_wchan);
1.58 thorpej 990:
1.102 chs 991: #endif /* DIAGNOSTIC */
1.58 thorpej 992: #ifdef LOCKDEBUG
1.155 ad 993: if (flags & PR_WAITOK) {
1.154 yamt 994: ASSERT_SLEEPABLE();
1.155 ad 995: }
1.56 sommerfe 996: #endif
1.1 pk 997:
1.134 ad 998: mutex_enter(&pp->pr_lock);
1.25 thorpej 999: pr_enter(pp, file, line);
1.20 thorpej 1000:
1001: startover:
1002: /*
1003: * Check to see if we've reached the hard limit. If we have,
1004: * and we can wait, then wait until an item has been returned to
1005: * the pool.
1006: */
1007: #ifdef DIAGNOSTIC
1.34 thorpej 1008: if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
1.25 thorpej 1009: pr_leave(pp);
1.134 ad 1010: mutex_exit(&pp->pr_lock);
1.20 thorpej 1011: panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
1012: }
1013: #endif
1.34 thorpej 1014: if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
1.68 thorpej 1015: if (pp->pr_drain_hook != NULL) {
1016: /*
1017: * Since the drain hook is going to free things
1018: * back to the pool, unlock, call the hook, re-lock,
1019: * and check the hardlimit condition again.
1020: */
1021: pr_leave(pp);
1.134 ad 1022: mutex_exit(&pp->pr_lock);
1.68 thorpej 1023: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
1.134 ad 1024: mutex_enter(&pp->pr_lock);
1.68 thorpej 1025: pr_enter(pp, file, line);
1026: if (pp->pr_nout < pp->pr_hardlimit)
1027: goto startover;
1028: }
1029:
1.29 sommerfe 1030: if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
1.20 thorpej 1031: /*
1032: * XXX: A warning isn't logged in this case. Should
1033: * it be?
1034: */
1035: pp->pr_flags |= PR_WANTED;
1.25 thorpej 1036: pr_leave(pp);
1.134 ad 1037: cv_wait(&pp->pr_cv, &pp->pr_lock);
1.25 thorpej 1038: pr_enter(pp, file, line);
1.20 thorpej 1039: goto startover;
1040: }
1.31 thorpej 1041:
1042: /*
1043: * Log a message that the hard limit has been hit.
1044: */
1045: if (pp->pr_hardlimit_warning != NULL &&
1046: ratecheck(&pp->pr_hardlimit_warning_last,
1047: &pp->pr_hardlimit_ratecap))
1048: log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
1.21 thorpej 1049:
1050: pp->pr_nfail++;
1051:
1.25 thorpej 1052: pr_leave(pp);
1.134 ad 1053: mutex_exit(&pp->pr_lock);
1.20 thorpej 1054: return (NULL);
1055: }
1056:
1.3 pk 1057: /*
1058: * The convention we use is that if `curpage' is not NULL, then
1059: * it points at a non-empty bucket. In particular, `curpage'
1060: * never points at a page header which has PR_PHINPAGE set and
1061: * has no items in its bucket.
1062: */
1.20 thorpej 1063: if ((ph = pp->pr_curpage) == NULL) {
1.113 yamt 1064: int error;
1065:
1.20 thorpej 1066: #ifdef DIAGNOSTIC
1067: if (pp->pr_nitems != 0) {
1.134 ad 1068: mutex_exit(&pp->pr_lock);
1.20 thorpej 1069: printf("pool_get: %s: curpage NULL, nitems %u\n",
1070: pp->pr_wchan, pp->pr_nitems);
1.80 provos 1071: panic("pool_get: nitems inconsistent");
1.20 thorpej 1072: }
1073: #endif
1074:
1.21 thorpej 1075: /*
1076: * Call the back-end page allocator for more memory.
1077: * Release the pool lock, as the back-end page allocator
1078: * may block.
1079: */
1.25 thorpej 1080: pr_leave(pp);
1.113 yamt 1081: error = pool_grow(pp, flags);
1082: pr_enter(pp, file, line);
1083: if (error != 0) {
1.21 thorpej 1084: /*
1.55 thorpej 1085: * We were unable to allocate a page or item
1086: * header, but we released the lock during
1087: * allocation, so perhaps items were freed
1088: * back to the pool. Check for this case.
1.21 thorpej 1089: */
1090: if (pp->pr_curpage != NULL)
1091: goto startover;
1.15 pk 1092:
1.117 yamt 1093: pp->pr_nfail++;
1.25 thorpej 1094: pr_leave(pp);
1.134 ad 1095: mutex_exit(&pp->pr_lock);
1.117 yamt 1096: return (NULL);
1.1 pk 1097: }
1.3 pk 1098:
1.20 thorpej 1099: /* Start the allocation process over. */
1100: goto startover;
1.3 pk 1101: }
1.97 yamt 1102: if (pp->pr_roflags & PR_NOTOUCH) {
1103: #ifdef DIAGNOSTIC
1104: if (__predict_false(ph->ph_nmissing == pp->pr_itemsperpage)) {
1105: pr_leave(pp);
1.134 ad 1106: mutex_exit(&pp->pr_lock);
1.97 yamt 1107: panic("pool_get: %s: page empty", pp->pr_wchan);
1108: }
1109: #endif
1110: v = pr_item_notouch_get(pp, ph);
1111: #ifdef POOL_DIAGNOSTIC
1112: pr_log(pp, v, PRLOG_GET, file, line);
1113: #endif
1114: } else {
1.102 chs 1115: v = pi = LIST_FIRST(&ph->ph_itemlist);
1.97 yamt 1116: if (__predict_false(v == NULL)) {
1117: pr_leave(pp);
1.134 ad 1118: mutex_exit(&pp->pr_lock);
1.97 yamt 1119: panic("pool_get: %s: page empty", pp->pr_wchan);
1120: }
1.20 thorpej 1121: #ifdef DIAGNOSTIC
1.97 yamt 1122: if (__predict_false(pp->pr_nitems == 0)) {
1123: pr_leave(pp);
1.134 ad 1124: mutex_exit(&pp->pr_lock);
1.97 yamt 1125: printf("pool_get: %s: items on itemlist, nitems %u\n",
1126: pp->pr_wchan, pp->pr_nitems);
1127: panic("pool_get: nitems inconsistent");
1128: }
1.65 enami 1129: #endif
1.56 sommerfe 1130:
1.65 enami 1131: #ifdef POOL_DIAGNOSTIC
1.97 yamt 1132: pr_log(pp, v, PRLOG_GET, file, line);
1.65 enami 1133: #endif
1.3 pk 1134:
1.65 enami 1135: #ifdef DIAGNOSTIC
1.97 yamt 1136: if (__predict_false(pi->pi_magic != PI_MAGIC)) {
1137: pr_printlog(pp, pi, printf);
1138: panic("pool_get(%s): free list modified: "
1139: "magic=%x; page %p; item addr %p\n",
1140: pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
1141: }
1.3 pk 1142: #endif
1143:
1.97 yamt 1144: /*
1145: * Remove from item list.
1146: */
1.102 chs 1147: LIST_REMOVE(pi, pi_list);
1.97 yamt 1148: }
1.20 thorpej 1149: pp->pr_nitems--;
1150: pp->pr_nout++;
1.6 thorpej 1151: if (ph->ph_nmissing == 0) {
1152: #ifdef DIAGNOSTIC
1.34 thorpej 1153: if (__predict_false(pp->pr_nidle == 0))
1.6 thorpej 1154: panic("pool_get: nidle inconsistent");
1155: #endif
1156: pp->pr_nidle--;
1.88 chs 1157:
1158: /*
1159: * This page was previously empty. Move it to the list of
1160: * partially-full pages. This page is already curpage.
1161: */
1162: LIST_REMOVE(ph, ph_pagelist);
1163: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.6 thorpej 1164: }
1.3 pk 1165: ph->ph_nmissing++;
1.97 yamt 1166: if (ph->ph_nmissing == pp->pr_itemsperpage) {
1.21 thorpej 1167: #ifdef DIAGNOSTIC
1.97 yamt 1168: if (__predict_false((pp->pr_roflags & PR_NOTOUCH) == 0 &&
1.102 chs 1169: !LIST_EMPTY(&ph->ph_itemlist))) {
1.25 thorpej 1170: pr_leave(pp);
1.134 ad 1171: mutex_exit(&pp->pr_lock);
1.21 thorpej 1172: panic("pool_get: %s: nmissing inconsistent",
1173: pp->pr_wchan);
1174: }
1175: #endif
1.3 pk 1176: /*
1.88 chs 1177: * This page is now full. Move it to the full list
1178: * and select a new current page.
1.3 pk 1179: */
1.88 chs 1180: LIST_REMOVE(ph, ph_pagelist);
1181: LIST_INSERT_HEAD(&pp->pr_fullpages, ph, ph_pagelist);
1182: pool_update_curpage(pp);
1.1 pk 1183: }
1.3 pk 1184:
1185: pp->pr_nget++;
1.111 christos 1186: pr_leave(pp);
1.20 thorpej 1187:
1188: /*
1189: * If we have a low water mark and we are now below that low
1190: * water mark, add more items to the pool.
1191: */
1.53 thorpej 1192: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1193: /*
1194: * XXX: Should we log a warning? Should we set up a timeout
1195: * to try again in a second or so? The latter could break
1196: * a caller's assumptions about interrupt protection, etc.
1197: */
1198: }
1199:
1.134 ad 1200: mutex_exit(&pp->pr_lock);
1.125 ad 1201: KASSERT((((vaddr_t)v + pp->pr_itemoffset) & (pp->pr_align - 1)) == 0);
1202: FREECHECK_OUT(&pp->pr_freecheck, v);
1.1 pk 1203: return (v);
1204: }
1205:
1206: /*
1.43 thorpej 1207: * Internal version of pool_put(). Pool is already locked/entered.
1.1 pk 1208: */
1.43 thorpej 1209: static void
1.101 thorpej 1210: pool_do_put(struct pool *pp, void *v, struct pool_pagelist *pq)
1.1 pk 1211: {
1212: struct pool_item *pi = v;
1.3 pk 1213: struct pool_item_header *ph;
1214:
1.134 ad 1215: KASSERT(mutex_owned(&pp->pr_lock));
1.125 ad 1216: FREECHECK_IN(&pp->pr_freecheck, v);
1.134 ad 1217: LOCKDEBUG_MEM_CHECK(v, pp->pr_size);
1.61 chs 1218:
1.30 thorpej 1219: #ifdef DIAGNOSTIC
1.34 thorpej 1220: if (__predict_false(pp->pr_nout == 0)) {
1.30 thorpej 1221: printf("pool %s: putting with none out\n",
1222: pp->pr_wchan);
1223: panic("pool_put");
1224: }
1225: #endif
1.3 pk 1226:
1.121 yamt 1227: if (__predict_false((ph = pr_find_pagehead(pp, v)) == NULL)) {
1.25 thorpej 1228: pr_printlog(pp, NULL, printf);
1.3 pk 1229: panic("pool_put: %s: page header missing", pp->pr_wchan);
1230: }
1.28 thorpej 1231:
1.3 pk 1232: /*
1233: * Return to item list.
1234: */
1.97 yamt 1235: if (pp->pr_roflags & PR_NOTOUCH) {
1236: pr_item_notouch_put(pp, ph, v);
1237: } else {
1.2 pk 1238: #ifdef DIAGNOSTIC
1.97 yamt 1239: pi->pi_magic = PI_MAGIC;
1.3 pk 1240: #endif
1.32 chs 1241: #ifdef DEBUG
1.97 yamt 1242: {
1243: int i, *ip = v;
1.32 chs 1244:
1.97 yamt 1245: for (i = 0; i < pp->pr_size / sizeof(int); i++) {
1246: *ip++ = PI_MAGIC;
1247: }
1.32 chs 1248: }
1249: #endif
1250:
1.102 chs 1251: LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.97 yamt 1252: }
1.79 thorpej 1253: KDASSERT(ph->ph_nmissing != 0);
1.3 pk 1254: ph->ph_nmissing--;
1255: pp->pr_nput++;
1.20 thorpej 1256: pp->pr_nitems++;
1257: pp->pr_nout--;
1.3 pk 1258:
1259: /* Cancel "pool empty" condition if it exists */
1260: if (pp->pr_curpage == NULL)
1261: pp->pr_curpage = ph;
1262:
1263: if (pp->pr_flags & PR_WANTED) {
1264: pp->pr_flags &= ~PR_WANTED;
1.134 ad 1265: cv_broadcast(&pp->pr_cv);
1.3 pk 1266: }
1267:
1268: /*
1.88 chs 1269: * If this page is now empty, do one of two things:
1.21 thorpej 1270: *
1.88 chs 1271: * (1) If we have more pages than the page high water mark,
1.96 thorpej 1272: * free the page back to the system. ONLY CONSIDER
1.90 thorpej 1273: * FREEING BACK A PAGE IF WE HAVE MORE THAN OUR MINIMUM PAGE
1274: * CLAIM.
1.21 thorpej 1275: *
1.88 chs 1276: * (2) Otherwise, move the page to the empty page list.
1277: *
1278: * Either way, select a new current page (so we use a partially-full
1279: * page if one is available).
1.3 pk 1280: */
1281: if (ph->ph_nmissing == 0) {
1.6 thorpej 1282: pp->pr_nidle++;
1.90 thorpej 1283: if (pp->pr_npages > pp->pr_minpages &&
1.152 yamt 1284: pp->pr_npages > pp->pr_maxpages) {
1.101 thorpej 1285: pr_rmpage(pp, ph, pq);
1.3 pk 1286: } else {
1.88 chs 1287: LIST_REMOVE(ph, ph_pagelist);
1288: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.3 pk 1289:
1.21 thorpej 1290: /*
1291: * Update the timestamp on the page. A page must
1292: * be idle for some period of time before it can
1293: * be reclaimed by the pagedaemon. This minimizes
1294: * ping-pong'ing for memory.
1.151 yamt 1295: *
1296: * note for 64-bit time_t: truncating to 32-bit is not
1297: * a problem for our usage.
1.21 thorpej 1298: */
1.151 yamt 1299: ph->ph_time = time_uptime;
1.1 pk 1300: }
1.88 chs 1301: pool_update_curpage(pp);
1.1 pk 1302: }
1.88 chs 1303:
1.21 thorpej 1304: /*
1.88 chs 1305: * If the page was previously completely full, move it to the
1306: * partially-full list and make it the current page. The next
1307: * allocation will get the item from this page, instead of
1308: * further fragmenting the pool.
1.21 thorpej 1309: */
1310: else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
1.88 chs 1311: LIST_REMOVE(ph, ph_pagelist);
1312: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.21 thorpej 1313: pp->pr_curpage = ph;
1314: }
1.43 thorpej 1315: }
1316:
1317: /*
1.134 ad 1318: * Return resource to the pool.
1.43 thorpej 1319: */
1.59 thorpej 1320: #ifdef POOL_DIAGNOSTIC
1.43 thorpej 1321: void
1322: _pool_put(struct pool *pp, void *v, const char *file, long line)
1323: {
1.101 thorpej 1324: struct pool_pagelist pq;
1325:
1326: LIST_INIT(&pq);
1.43 thorpej 1327:
1.134 ad 1328: mutex_enter(&pp->pr_lock);
1.43 thorpej 1329: pr_enter(pp, file, line);
1330:
1.56 sommerfe 1331: pr_log(pp, v, PRLOG_PUT, file, line);
1332:
1.101 thorpej 1333: pool_do_put(pp, v, &pq);
1.21 thorpej 1334:
1.25 thorpej 1335: pr_leave(pp);
1.134 ad 1336: mutex_exit(&pp->pr_lock);
1.101 thorpej 1337:
1.102 chs 1338: pr_pagelist_free(pp, &pq);
1.1 pk 1339: }
1.57 sommerfe 1340: #undef pool_put
1.59 thorpej 1341: #endif /* POOL_DIAGNOSTIC */
1.1 pk 1342:
1.56 sommerfe 1343: void
1344: pool_put(struct pool *pp, void *v)
1345: {
1.101 thorpej 1346: struct pool_pagelist pq;
1347:
1348: LIST_INIT(&pq);
1.56 sommerfe 1349:
1.134 ad 1350: mutex_enter(&pp->pr_lock);
1.101 thorpej 1351: pool_do_put(pp, v, &pq);
1.134 ad 1352: mutex_exit(&pp->pr_lock);
1.56 sommerfe 1353:
1.102 chs 1354: pr_pagelist_free(pp, &pq);
1.56 sommerfe 1355: }
1.57 sommerfe 1356:
1.59 thorpej 1357: #ifdef POOL_DIAGNOSTIC
1.57 sommerfe 1358: #define pool_put(h, v) _pool_put((h), (v), __FILE__, __LINE__)
1.56 sommerfe 1359: #endif
1.74 thorpej 1360:
1361: /*
1.113 yamt 1362: * pool_grow: grow a pool by a page.
1363: *
1364: * => called with pool locked.
1365: * => unlock and relock the pool.
1366: * => return with pool locked.
1367: */
1368:
1369: static int
1370: pool_grow(struct pool *pp, int flags)
1371: {
1372: struct pool_item_header *ph = NULL;
1373: char *cp;
1374:
1.134 ad 1375: mutex_exit(&pp->pr_lock);
1.113 yamt 1376: cp = pool_allocator_alloc(pp, flags);
1377: if (__predict_true(cp != NULL)) {
1378: ph = pool_alloc_item_header(pp, cp, flags);
1379: }
1380: if (__predict_false(cp == NULL || ph == NULL)) {
1381: if (cp != NULL) {
1382: pool_allocator_free(pp, cp);
1383: }
1.134 ad 1384: mutex_enter(&pp->pr_lock);
1.113 yamt 1385: return ENOMEM;
1386: }
1387:
1.134 ad 1388: mutex_enter(&pp->pr_lock);
1.113 yamt 1389: pool_prime_page(pp, cp, ph);
1390: pp->pr_npagealloc++;
1391: return 0;
1392: }
1393:
1394: /*
1.74 thorpej 1395: * Add N items to the pool.
1396: */
1397: int
1398: pool_prime(struct pool *pp, int n)
1399: {
1.75 simonb 1400: int newpages;
1.113 yamt 1401: int error = 0;
1.74 thorpej 1402:
1.134 ad 1403: mutex_enter(&pp->pr_lock);
1.74 thorpej 1404:
1405: newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1406:
1407: while (newpages-- > 0) {
1.113 yamt 1408: error = pool_grow(pp, PR_NOWAIT);
1409: if (error) {
1.74 thorpej 1410: break;
1411: }
1412: pp->pr_minpages++;
1413: }
1414:
1415: if (pp->pr_minpages >= pp->pr_maxpages)
1416: pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */
1417:
1.134 ad 1418: mutex_exit(&pp->pr_lock);
1.113 yamt 1419: return error;
1.74 thorpej 1420: }
1.55 thorpej 1421:
1422: /*
1.3 pk 1423: * Add a page worth of items to the pool.
1.21 thorpej 1424: *
1425: * Note, we must be called with the pool descriptor LOCKED.
1.3 pk 1426: */
1.55 thorpej 1427: static void
1.128 christos 1428: pool_prime_page(struct pool *pp, void *storage, struct pool_item_header *ph)
1.3 pk 1429: {
1430: struct pool_item *pi;
1.128 christos 1431: void *cp = storage;
1.125 ad 1432: const unsigned int align = pp->pr_align;
1433: const unsigned int ioff = pp->pr_itemoffset;
1.55 thorpej 1434: int n;
1.36 pk 1435:
1.134 ad 1436: KASSERT(mutex_owned(&pp->pr_lock));
1.91 yamt 1437:
1.66 thorpej 1438: #ifdef DIAGNOSTIC
1.121 yamt 1439: if ((pp->pr_roflags & PR_NOALIGN) == 0 &&
1.150 skrll 1440: ((uintptr_t)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
1.36 pk 1441: panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1.66 thorpej 1442: #endif
1.3 pk 1443:
1444: /*
1445: * Insert page header.
1446: */
1.88 chs 1447: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.102 chs 1448: LIST_INIT(&ph->ph_itemlist);
1.3 pk 1449: ph->ph_page = storage;
1450: ph->ph_nmissing = 0;
1.151 yamt 1451: ph->ph_time = time_uptime;
1.88 chs 1452: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1453: SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
1.3 pk 1454:
1.6 thorpej 1455: pp->pr_nidle++;
1456:
1.3 pk 1457: /*
1458: * Color this page.
1459: */
1.141 yamt 1460: ph->ph_off = pp->pr_curcolor;
1461: cp = (char *)cp + ph->ph_off;
1.3 pk 1462: if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
1463: pp->pr_curcolor = 0;
1464:
1465: /*
1466: * Adjust storage to apply aligment to `pr_itemoffset' in each item.
1467: */
1468: if (ioff != 0)
1.128 christos 1469: cp = (char *)cp + align - ioff;
1.3 pk 1470:
1.125 ad 1471: KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
1472:
1.3 pk 1473: /*
1474: * Insert remaining chunks on the bucket list.
1475: */
1476: n = pp->pr_itemsperpage;
1.20 thorpej 1477: pp->pr_nitems += n;
1.3 pk 1478:
1.97 yamt 1479: if (pp->pr_roflags & PR_NOTOUCH) {
1.141 yamt 1480: pr_item_notouch_init(pp, ph);
1.97 yamt 1481: } else {
1482: while (n--) {
1483: pi = (struct pool_item *)cp;
1.78 thorpej 1484:
1.97 yamt 1485: KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
1.3 pk 1486:
1.97 yamt 1487: /* Insert on page list */
1.102 chs 1488: LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.3 pk 1489: #ifdef DIAGNOSTIC
1.97 yamt 1490: pi->pi_magic = PI_MAGIC;
1.3 pk 1491: #endif
1.128 christos 1492: cp = (char *)cp + pp->pr_size;
1.125 ad 1493:
1494: KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
1.97 yamt 1495: }
1.3 pk 1496: }
1497:
1498: /*
1499: * If the pool was depleted, point at the new page.
1500: */
1501: if (pp->pr_curpage == NULL)
1502: pp->pr_curpage = ph;
1503:
1504: if (++pp->pr_npages > pp->pr_hiwat)
1505: pp->pr_hiwat = pp->pr_npages;
1506: }
1507:
1.20 thorpej 1508: /*
1.52 thorpej 1509: * Used by pool_get() when nitems drops below the low water mark. This
1.88 chs 1510: * is used to catch up pr_nitems with the low water mark.
1.20 thorpej 1511: *
1.21 thorpej 1512: * Note 1, we never wait for memory here, we let the caller decide what to do.
1.20 thorpej 1513: *
1.73 thorpej 1514: * Note 2, we must be called with the pool already locked, and we return
1.20 thorpej 1515: * with it locked.
1516: */
1517: static int
1.42 thorpej 1518: pool_catchup(struct pool *pp)
1.20 thorpej 1519: {
1520: int error = 0;
1521:
1.54 thorpej 1522: while (POOL_NEEDS_CATCHUP(pp)) {
1.113 yamt 1523: error = pool_grow(pp, PR_NOWAIT);
1524: if (error) {
1.20 thorpej 1525: break;
1526: }
1527: }
1.113 yamt 1528: return error;
1.20 thorpej 1529: }
1530:
1.88 chs 1531: static void
1532: pool_update_curpage(struct pool *pp)
1533: {
1534:
1535: pp->pr_curpage = LIST_FIRST(&pp->pr_partpages);
1536: if (pp->pr_curpage == NULL) {
1537: pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages);
1538: }
1.168 yamt 1539: KASSERT((pp->pr_curpage == NULL && pp->pr_nitems == 0) ||
1540: (pp->pr_curpage != NULL && pp->pr_nitems > 0));
1.88 chs 1541: }
1542:
1.3 pk 1543: void
1.42 thorpej 1544: pool_setlowat(struct pool *pp, int n)
1.3 pk 1545: {
1.15 pk 1546:
1.134 ad 1547: mutex_enter(&pp->pr_lock);
1.21 thorpej 1548:
1.3 pk 1549: pp->pr_minitems = n;
1.15 pk 1550: pp->pr_minpages = (n == 0)
1551: ? 0
1.18 thorpej 1552: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.20 thorpej 1553:
1554: /* Make sure we're caught up with the newly-set low water mark. */
1.75 simonb 1555: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1556: /*
1557: * XXX: Should we log a warning? Should we set up a timeout
1558: * to try again in a second or so? The latter could break
1559: * a caller's assumptions about interrupt protection, etc.
1560: */
1561: }
1.21 thorpej 1562:
1.134 ad 1563: mutex_exit(&pp->pr_lock);
1.3 pk 1564: }
1565:
1566: void
1.42 thorpej 1567: pool_sethiwat(struct pool *pp, int n)
1.3 pk 1568: {
1.15 pk 1569:
1.134 ad 1570: mutex_enter(&pp->pr_lock);
1.21 thorpej 1571:
1.15 pk 1572: pp->pr_maxpages = (n == 0)
1573: ? 0
1.18 thorpej 1574: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1575:
1.134 ad 1576: mutex_exit(&pp->pr_lock);
1.3 pk 1577: }
1578:
1.20 thorpej 1579: void
1.42 thorpej 1580: pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1.20 thorpej 1581: {
1582:
1.134 ad 1583: mutex_enter(&pp->pr_lock);
1.20 thorpej 1584:
1585: pp->pr_hardlimit = n;
1586: pp->pr_hardlimit_warning = warnmess;
1.31 thorpej 1587: pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1588: pp->pr_hardlimit_warning_last.tv_sec = 0;
1589: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.20 thorpej 1590:
1591: /*
1.21 thorpej 1592: * In-line version of pool_sethiwat(), because we don't want to
1593: * release the lock.
1.20 thorpej 1594: */
1595: pp->pr_maxpages = (n == 0)
1596: ? 0
1597: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1598:
1.134 ad 1599: mutex_exit(&pp->pr_lock);
1.20 thorpej 1600: }
1.3 pk 1601:
1602: /*
1603: * Release all complete pages that have not been used recently.
1604: */
1.66 thorpej 1605: int
1.59 thorpej 1606: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 1607: _pool_reclaim(struct pool *pp, const char *file, long line)
1.56 sommerfe 1608: #else
1609: pool_reclaim(struct pool *pp)
1610: #endif
1.3 pk 1611: {
1612: struct pool_item_header *ph, *phnext;
1.61 chs 1613: struct pool_pagelist pq;
1.151 yamt 1614: uint32_t curtime;
1.134 ad 1615: bool klock;
1616: int rv;
1.3 pk 1617:
1.68 thorpej 1618: if (pp->pr_drain_hook != NULL) {
1619: /*
1620: * The drain hook must be called with the pool unlocked.
1621: */
1622: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT);
1623: }
1624:
1.134 ad 1625: /*
1.157 ad 1626: * XXXSMP Because we do not want to cause non-MPSAFE code
1627: * to block.
1.134 ad 1628: */
1629: if (pp->pr_ipl == IPL_SOFTNET || pp->pr_ipl == IPL_SOFTCLOCK ||
1630: pp->pr_ipl == IPL_SOFTSERIAL) {
1631: KERNEL_LOCK(1, NULL);
1632: klock = true;
1633: } else
1634: klock = false;
1635:
1636: /* Reclaim items from the pool's cache (if any). */
1637: if (pp->pr_cache != NULL)
1638: pool_cache_invalidate(pp->pr_cache);
1639:
1640: if (mutex_tryenter(&pp->pr_lock) == 0) {
1641: if (klock) {
1642: KERNEL_UNLOCK_ONE(NULL);
1643: }
1.66 thorpej 1644: return (0);
1.134 ad 1645: }
1.25 thorpej 1646: pr_enter(pp, file, line);
1.68 thorpej 1647:
1.88 chs 1648: LIST_INIT(&pq);
1.43 thorpej 1649:
1.151 yamt 1650: curtime = time_uptime;
1.21 thorpej 1651:
1.88 chs 1652: for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
1653: phnext = LIST_NEXT(ph, ph_pagelist);
1.3 pk 1654:
1655: /* Check our minimum page claim */
1656: if (pp->pr_npages <= pp->pr_minpages)
1657: break;
1658:
1.88 chs 1659: KASSERT(ph->ph_nmissing == 0);
1.151 yamt 1660: if (curtime - ph->ph_time < pool_inactive_time
1.117 yamt 1661: && !pa_starved_p(pp->pr_alloc))
1.88 chs 1662: continue;
1.21 thorpej 1663:
1.88 chs 1664: /*
1665: * If freeing this page would put us below
1666: * the low water mark, stop now.
1667: */
1668: if ((pp->pr_nitems - pp->pr_itemsperpage) <
1669: pp->pr_minitems)
1670: break;
1.21 thorpej 1671:
1.88 chs 1672: pr_rmpage(pp, ph, &pq);
1.3 pk 1673: }
1674:
1.25 thorpej 1675: pr_leave(pp);
1.134 ad 1676: mutex_exit(&pp->pr_lock);
1677:
1678: if (LIST_EMPTY(&pq))
1679: rv = 0;
1680: else {
1681: pr_pagelist_free(pp, &pq);
1682: rv = 1;
1683: }
1684:
1685: if (klock) {
1686: KERNEL_UNLOCK_ONE(NULL);
1687: }
1.66 thorpej 1688:
1.134 ad 1689: return (rv);
1.3 pk 1690: }
1691:
1692: /*
1.134 ad 1693: * Drain pools, one at a time. This is a two stage process;
1694: * drain_start kicks off a cross call to drain CPU-level caches
1695: * if the pool has an associated pool_cache. drain_end waits
1696: * for those cross calls to finish, and then drains the cache
1697: * (if any) and pool.
1.131 ad 1698: *
1.134 ad 1699: * Note, must never be called from interrupt context.
1.3 pk 1700: */
1701: void
1.134 ad 1702: pool_drain_start(struct pool **ppp, uint64_t *wp)
1.3 pk 1703: {
1704: struct pool *pp;
1.134 ad 1705:
1.145 ad 1706: KASSERT(!TAILQ_EMPTY(&pool_head));
1.3 pk 1707:
1.61 chs 1708: pp = NULL;
1.134 ad 1709:
1710: /* Find next pool to drain, and add a reference. */
1711: mutex_enter(&pool_head_lock);
1712: do {
1713: if (drainpp == NULL) {
1.145 ad 1714: drainpp = TAILQ_FIRST(&pool_head);
1.134 ad 1715: }
1716: if (drainpp != NULL) {
1717: pp = drainpp;
1.145 ad 1718: drainpp = TAILQ_NEXT(pp, pr_poollist);
1.134 ad 1719: }
1720: /*
1721: * Skip completely idle pools. We depend on at least
1722: * one pool in the system being active.
1723: */
1724: } while (pp == NULL || pp->pr_npages == 0);
1725: pp->pr_refcnt++;
1726: mutex_exit(&pool_head_lock);
1727:
1728: /* If there is a pool_cache, drain CPU level caches. */
1729: *ppp = pp;
1730: if (pp->pr_cache != NULL) {
1731: *wp = xc_broadcast(0, (xcfunc_t)pool_cache_xcall,
1732: pp->pr_cache, NULL);
1733: }
1734: }
1735:
1736: void
1737: pool_drain_end(struct pool *pp, uint64_t where)
1738: {
1739:
1740: if (pp == NULL)
1741: return;
1742:
1743: KASSERT(pp->pr_refcnt > 0);
1744:
1745: /* Wait for remote draining to complete. */
1746: if (pp->pr_cache != NULL)
1747: xc_wait(where);
1748:
1749: /* Drain the cache (if any) and pool.. */
1750: pool_reclaim(pp);
1751:
1752: /* Finally, unlock the pool. */
1753: mutex_enter(&pool_head_lock);
1754: pp->pr_refcnt--;
1755: cv_broadcast(&pool_busy);
1756: mutex_exit(&pool_head_lock);
1.3 pk 1757: }
1758:
1759: /*
1760: * Diagnostic helpers.
1761: */
1762: void
1.42 thorpej 1763: pool_print(struct pool *pp, const char *modif)
1.21 thorpej 1764: {
1765:
1.25 thorpej 1766: pool_print1(pp, modif, printf);
1.21 thorpej 1767: }
1768:
1.25 thorpej 1769: void
1.108 yamt 1770: pool_printall(const char *modif, void (*pr)(const char *, ...))
1771: {
1772: struct pool *pp;
1773:
1.145 ad 1774: TAILQ_FOREACH(pp, &pool_head, pr_poollist) {
1.108 yamt 1775: pool_printit(pp, modif, pr);
1776: }
1777: }
1778:
1779: void
1.42 thorpej 1780: pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.25 thorpej 1781: {
1782:
1783: if (pp == NULL) {
1784: (*pr)("Must specify a pool to print.\n");
1785: return;
1786: }
1787:
1788: pool_print1(pp, modif, pr);
1789: }
1790:
1.21 thorpej 1791: static void
1.124 yamt 1792: pool_print_pagelist(struct pool *pp, struct pool_pagelist *pl,
1.97 yamt 1793: void (*pr)(const char *, ...))
1.88 chs 1794: {
1795: struct pool_item_header *ph;
1796: #ifdef DIAGNOSTIC
1797: struct pool_item *pi;
1798: #endif
1799:
1800: LIST_FOREACH(ph, pl, ph_pagelist) {
1.151 yamt 1801: (*pr)("\t\tpage %p, nmissing %d, time %" PRIu32 "\n",
1802: ph->ph_page, ph->ph_nmissing, ph->ph_time);
1.88 chs 1803: #ifdef DIAGNOSTIC
1.97 yamt 1804: if (!(pp->pr_roflags & PR_NOTOUCH)) {
1.102 chs 1805: LIST_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1.97 yamt 1806: if (pi->pi_magic != PI_MAGIC) {
1807: (*pr)("\t\t\titem %p, magic 0x%x\n",
1808: pi, pi->pi_magic);
1809: }
1.88 chs 1810: }
1811: }
1812: #endif
1813: }
1814: }
1815:
1816: static void
1.42 thorpej 1817: pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.3 pk 1818: {
1.25 thorpej 1819: struct pool_item_header *ph;
1.134 ad 1820: pool_cache_t pc;
1821: pcg_t *pcg;
1822: pool_cache_cpu_t *cc;
1823: uint64_t cpuhit, cpumiss;
1.44 thorpej 1824: int i, print_log = 0, print_pagelist = 0, print_cache = 0;
1.25 thorpej 1825: char c;
1826:
1827: while ((c = *modif++) != '\0') {
1828: if (c == 'l')
1829: print_log = 1;
1830: if (c == 'p')
1831: print_pagelist = 1;
1.44 thorpej 1832: if (c == 'c')
1833: print_cache = 1;
1.25 thorpej 1834: }
1835:
1.134 ad 1836: if ((pc = pp->pr_cache) != NULL) {
1837: (*pr)("POOL CACHE");
1838: } else {
1839: (*pr)("POOL");
1840: }
1841:
1842: (*pr)(" %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1.25 thorpej 1843: pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1844: pp->pr_roflags);
1.66 thorpej 1845: (*pr)("\talloc %p\n", pp->pr_alloc);
1.25 thorpej 1846: (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1847: pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1848: (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1849: pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1850:
1.134 ad 1851: (*pr)("\tnget %lu, nfail %lu, nput %lu\n",
1.25 thorpej 1852: pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1853: (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1854: pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1855:
1856: if (print_pagelist == 0)
1857: goto skip_pagelist;
1858:
1.88 chs 1859: if ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1860: (*pr)("\n\tempty page list:\n");
1.97 yamt 1861: pool_print_pagelist(pp, &pp->pr_emptypages, pr);
1.88 chs 1862: if ((ph = LIST_FIRST(&pp->pr_fullpages)) != NULL)
1863: (*pr)("\n\tfull page list:\n");
1.97 yamt 1864: pool_print_pagelist(pp, &pp->pr_fullpages, pr);
1.88 chs 1865: if ((ph = LIST_FIRST(&pp->pr_partpages)) != NULL)
1866: (*pr)("\n\tpartial-page list:\n");
1.97 yamt 1867: pool_print_pagelist(pp, &pp->pr_partpages, pr);
1.88 chs 1868:
1.25 thorpej 1869: if (pp->pr_curpage == NULL)
1870: (*pr)("\tno current page\n");
1871: else
1872: (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1873:
1874: skip_pagelist:
1875: if (print_log == 0)
1876: goto skip_log;
1877:
1878: (*pr)("\n");
1879: if ((pp->pr_roflags & PR_LOGGING) == 0)
1880: (*pr)("\tno log\n");
1.122 christos 1881: else {
1.25 thorpej 1882: pr_printlog(pp, NULL, pr);
1.122 christos 1883: }
1.3 pk 1884:
1.25 thorpej 1885: skip_log:
1.44 thorpej 1886:
1.102 chs 1887: #define PR_GROUPLIST(pcg) \
1888: (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail); \
1.142 ad 1889: for (i = 0; i < pcg->pcg_size; i++) { \
1.102 chs 1890: if (pcg->pcg_objects[i].pcgo_pa != \
1891: POOL_PADDR_INVALID) { \
1892: (*pr)("\t\t\t%p, 0x%llx\n", \
1893: pcg->pcg_objects[i].pcgo_va, \
1894: (unsigned long long) \
1895: pcg->pcg_objects[i].pcgo_pa); \
1896: } else { \
1897: (*pr)("\t\t\t%p\n", \
1898: pcg->pcg_objects[i].pcgo_va); \
1899: } \
1900: }
1901:
1.134 ad 1902: if (pc != NULL) {
1903: cpuhit = 0;
1904: cpumiss = 0;
1905: for (i = 0; i < MAXCPUS; i++) {
1906: if ((cc = pc->pc_cpus[i]) == NULL)
1907: continue;
1908: cpuhit += cc->cc_hits;
1909: cpumiss += cc->cc_misses;
1910: }
1911: (*pr)("\tcpu layer hits %llu misses %llu\n", cpuhit, cpumiss);
1912: (*pr)("\tcache layer hits %llu misses %llu\n",
1913: pc->pc_hits, pc->pc_misses);
1914: (*pr)("\tcache layer entry uncontended %llu contended %llu\n",
1915: pc->pc_hits + pc->pc_misses - pc->pc_contended,
1916: pc->pc_contended);
1917: (*pr)("\tcache layer empty groups %u full groups %u\n",
1918: pc->pc_nempty, pc->pc_nfull);
1919: if (print_cache) {
1920: (*pr)("\tfull cache groups:\n");
1921: for (pcg = pc->pc_fullgroups; pcg != NULL;
1922: pcg = pcg->pcg_next) {
1923: PR_GROUPLIST(pcg);
1924: }
1925: (*pr)("\tempty cache groups:\n");
1926: for (pcg = pc->pc_emptygroups; pcg != NULL;
1927: pcg = pcg->pcg_next) {
1928: PR_GROUPLIST(pcg);
1929: }
1.103 chs 1930: }
1.44 thorpej 1931: }
1.102 chs 1932: #undef PR_GROUPLIST
1.44 thorpej 1933:
1.88 chs 1934: pr_enter_check(pp, pr);
1935: }
1936:
1937: static int
1938: pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
1939: {
1940: struct pool_item *pi;
1.128 christos 1941: void *page;
1.88 chs 1942: int n;
1943:
1.121 yamt 1944: if ((pp->pr_roflags & PR_NOALIGN) == 0) {
1.128 christos 1945: page = (void *)((uintptr_t)ph & pp->pr_alloc->pa_pagemask);
1.121 yamt 1946: if (page != ph->ph_page &&
1947: (pp->pr_roflags & PR_PHINPAGE) != 0) {
1948: if (label != NULL)
1949: printf("%s: ", label);
1950: printf("pool(%p:%s): page inconsistency: page %p;"
1951: " at page head addr %p (p %p)\n", pp,
1952: pp->pr_wchan, ph->ph_page,
1953: ph, page);
1954: return 1;
1955: }
1.88 chs 1956: }
1.3 pk 1957:
1.97 yamt 1958: if ((pp->pr_roflags & PR_NOTOUCH) != 0)
1959: return 0;
1960:
1.102 chs 1961: for (pi = LIST_FIRST(&ph->ph_itemlist), n = 0;
1.88 chs 1962: pi != NULL;
1.102 chs 1963: pi = LIST_NEXT(pi,pi_list), n++) {
1.88 chs 1964:
1965: #ifdef DIAGNOSTIC
1966: if (pi->pi_magic != PI_MAGIC) {
1967: if (label != NULL)
1968: printf("%s: ", label);
1969: printf("pool(%s): free list modified: magic=%x;"
1.121 yamt 1970: " page %p; item ordinal %d; addr %p\n",
1.88 chs 1971: pp->pr_wchan, pi->pi_magic, ph->ph_page,
1.121 yamt 1972: n, pi);
1.88 chs 1973: panic("pool");
1974: }
1975: #endif
1.121 yamt 1976: if ((pp->pr_roflags & PR_NOALIGN) != 0) {
1977: continue;
1978: }
1.128 christos 1979: page = (void *)((uintptr_t)pi & pp->pr_alloc->pa_pagemask);
1.88 chs 1980: if (page == ph->ph_page)
1981: continue;
1982:
1983: if (label != NULL)
1984: printf("%s: ", label);
1985: printf("pool(%p:%s): page inconsistency: page %p;"
1986: " item ordinal %d; addr %p (p %p)\n", pp,
1987: pp->pr_wchan, ph->ph_page,
1988: n, pi, page);
1989: return 1;
1990: }
1991: return 0;
1.3 pk 1992: }
1993:
1.88 chs 1994:
1.3 pk 1995: int
1.42 thorpej 1996: pool_chk(struct pool *pp, const char *label)
1.3 pk 1997: {
1998: struct pool_item_header *ph;
1999: int r = 0;
2000:
1.134 ad 2001: mutex_enter(&pp->pr_lock);
1.88 chs 2002: LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
2003: r = pool_chk_page(pp, label, ph);
2004: if (r) {
2005: goto out;
2006: }
2007: }
2008: LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
2009: r = pool_chk_page(pp, label, ph);
2010: if (r) {
1.3 pk 2011: goto out;
2012: }
1.88 chs 2013: }
2014: LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
2015: r = pool_chk_page(pp, label, ph);
2016: if (r) {
1.3 pk 2017: goto out;
2018: }
2019: }
1.88 chs 2020:
1.3 pk 2021: out:
1.134 ad 2022: mutex_exit(&pp->pr_lock);
1.3 pk 2023: return (r);
1.43 thorpej 2024: }
2025:
2026: /*
2027: * pool_cache_init:
2028: *
2029: * Initialize a pool cache.
1.134 ad 2030: */
2031: pool_cache_t
2032: pool_cache_init(size_t size, u_int align, u_int align_offset, u_int flags,
2033: const char *wchan, struct pool_allocator *palloc, int ipl,
2034: int (*ctor)(void *, void *, int), void (*dtor)(void *, void *), void *arg)
2035: {
2036: pool_cache_t pc;
2037:
2038: pc = pool_get(&cache_pool, PR_WAITOK);
2039: if (pc == NULL)
2040: return NULL;
2041:
2042: pool_cache_bootstrap(pc, size, align, align_offset, flags, wchan,
2043: palloc, ipl, ctor, dtor, arg);
2044:
2045: return pc;
2046: }
2047:
2048: /*
2049: * pool_cache_bootstrap:
1.43 thorpej 2050: *
1.134 ad 2051: * Kernel-private version of pool_cache_init(). The caller
2052: * provides initial storage.
1.43 thorpej 2053: */
2054: void
1.134 ad 2055: pool_cache_bootstrap(pool_cache_t pc, size_t size, u_int align,
2056: u_int align_offset, u_int flags, const char *wchan,
2057: struct pool_allocator *palloc, int ipl,
2058: int (*ctor)(void *, void *, int), void (*dtor)(void *, void *),
1.43 thorpej 2059: void *arg)
2060: {
1.134 ad 2061: CPU_INFO_ITERATOR cii;
1.145 ad 2062: pool_cache_t pc1;
1.134 ad 2063: struct cpu_info *ci;
2064: struct pool *pp;
2065:
2066: pp = &pc->pc_pool;
2067: if (palloc == NULL && ipl == IPL_NONE)
2068: palloc = &pool_allocator_nointr;
2069: pool_init(pp, size, align, align_offset, flags, wchan, palloc, ipl);
1.157 ad 2070: mutex_init(&pc->pc_lock, MUTEX_DEFAULT, ipl);
1.43 thorpej 2071:
1.134 ad 2072: if (ctor == NULL) {
2073: ctor = (int (*)(void *, void *, int))nullop;
2074: }
2075: if (dtor == NULL) {
2076: dtor = (void (*)(void *, void *))nullop;
2077: }
1.43 thorpej 2078:
1.134 ad 2079: pc->pc_emptygroups = NULL;
2080: pc->pc_fullgroups = NULL;
2081: pc->pc_partgroups = NULL;
1.43 thorpej 2082: pc->pc_ctor = ctor;
2083: pc->pc_dtor = dtor;
2084: pc->pc_arg = arg;
1.134 ad 2085: pc->pc_hits = 0;
1.48 thorpej 2086: pc->pc_misses = 0;
1.134 ad 2087: pc->pc_nempty = 0;
2088: pc->pc_npart = 0;
2089: pc->pc_nfull = 0;
2090: pc->pc_contended = 0;
2091: pc->pc_refcnt = 0;
1.136 yamt 2092: pc->pc_freecheck = NULL;
1.134 ad 2093:
1.142 ad 2094: if ((flags & PR_LARGECACHE) != 0) {
2095: pc->pc_pcgsize = PCG_NOBJECTS_LARGE;
1.163 ad 2096: pc->pc_pcgpool = &pcg_large_pool;
1.142 ad 2097: } else {
2098: pc->pc_pcgsize = PCG_NOBJECTS_NORMAL;
1.163 ad 2099: pc->pc_pcgpool = &pcg_normal_pool;
1.142 ad 2100: }
2101:
1.134 ad 2102: /* Allocate per-CPU caches. */
2103: memset(pc->pc_cpus, 0, sizeof(pc->pc_cpus));
2104: pc->pc_ncpu = 0;
1.139 ad 2105: if (ncpu < 2) {
1.137 ad 2106: /* XXX For sparc: boot CPU is not attached yet. */
2107: pool_cache_cpu_init1(curcpu(), pc);
2108: } else {
2109: for (CPU_INFO_FOREACH(cii, ci)) {
2110: pool_cache_cpu_init1(ci, pc);
2111: }
1.134 ad 2112: }
1.145 ad 2113:
2114: /* Add to list of all pools. */
2115: if (__predict_true(!cold))
1.134 ad 2116: mutex_enter(&pool_head_lock);
1.145 ad 2117: TAILQ_FOREACH(pc1, &pool_cache_head, pc_cachelist) {
2118: if (strcmp(pc1->pc_pool.pr_wchan, pc->pc_pool.pr_wchan) > 0)
2119: break;
2120: }
2121: if (pc1 == NULL)
2122: TAILQ_INSERT_TAIL(&pool_cache_head, pc, pc_cachelist);
2123: else
2124: TAILQ_INSERT_BEFORE(pc1, pc, pc_cachelist);
2125: if (__predict_true(!cold))
1.134 ad 2126: mutex_exit(&pool_head_lock);
1.145 ad 2127:
2128: membar_sync();
2129: pp->pr_cache = pc;
1.43 thorpej 2130: }
2131:
2132: /*
2133: * pool_cache_destroy:
2134: *
2135: * Destroy a pool cache.
2136: */
2137: void
1.134 ad 2138: pool_cache_destroy(pool_cache_t pc)
1.43 thorpej 2139: {
1.134 ad 2140: struct pool *pp = &pc->pc_pool;
1.175 jym 2141: u_int i;
1.134 ad 2142:
2143: /* Remove it from the global list. */
2144: mutex_enter(&pool_head_lock);
2145: while (pc->pc_refcnt != 0)
2146: cv_wait(&pool_busy, &pool_head_lock);
1.145 ad 2147: TAILQ_REMOVE(&pool_cache_head, pc, pc_cachelist);
1.134 ad 2148: mutex_exit(&pool_head_lock);
1.43 thorpej 2149:
2150: /* First, invalidate the entire cache. */
2151: pool_cache_invalidate(pc);
2152:
1.134 ad 2153: /* Disassociate it from the pool. */
2154: mutex_enter(&pp->pr_lock);
2155: pp->pr_cache = NULL;
2156: mutex_exit(&pp->pr_lock);
2157:
2158: /* Destroy per-CPU data */
1.175 jym 2159: for (i = 0; i < MAXCPUS; i++)
2160: pool_cache_invalidate_cpu(pc, i);
1.134 ad 2161:
2162: /* Finally, destroy it. */
2163: mutex_destroy(&pc->pc_lock);
2164: pool_destroy(pp);
2165: pool_put(&cache_pool, pc);
2166: }
2167:
2168: /*
2169: * pool_cache_cpu_init1:
2170: *
2171: * Called for each pool_cache whenever a new CPU is attached.
2172: */
2173: static void
2174: pool_cache_cpu_init1(struct cpu_info *ci, pool_cache_t pc)
2175: {
2176: pool_cache_cpu_t *cc;
1.137 ad 2177: int index;
1.134 ad 2178:
1.137 ad 2179: index = ci->ci_index;
2180:
2181: KASSERT(index < MAXCPUS);
1.134 ad 2182:
1.137 ad 2183: if ((cc = pc->pc_cpus[index]) != NULL) {
2184: KASSERT(cc->cc_cpuindex == index);
1.134 ad 2185: return;
2186: }
2187:
2188: /*
2189: * The first CPU is 'free'. This needs to be the case for
2190: * bootstrap - we may not be able to allocate yet.
2191: */
2192: if (pc->pc_ncpu == 0) {
2193: cc = &pc->pc_cpu0;
2194: pc->pc_ncpu = 1;
2195: } else {
2196: mutex_enter(&pc->pc_lock);
2197: pc->pc_ncpu++;
2198: mutex_exit(&pc->pc_lock);
2199: cc = pool_get(&cache_cpu_pool, PR_WAITOK);
2200: }
2201:
2202: cc->cc_ipl = pc->pc_pool.pr_ipl;
2203: cc->cc_iplcookie = makeiplcookie(cc->cc_ipl);
2204: cc->cc_cache = pc;
1.137 ad 2205: cc->cc_cpuindex = index;
1.134 ad 2206: cc->cc_hits = 0;
2207: cc->cc_misses = 0;
1.169 yamt 2208: cc->cc_current = __UNCONST(&pcg_dummy);
2209: cc->cc_previous = __UNCONST(&pcg_dummy);
1.134 ad 2210:
1.137 ad 2211: pc->pc_cpus[index] = cc;
1.43 thorpej 2212: }
2213:
1.134 ad 2214: /*
2215: * pool_cache_cpu_init:
2216: *
2217: * Called whenever a new CPU is attached.
2218: */
2219: void
2220: pool_cache_cpu_init(struct cpu_info *ci)
1.43 thorpej 2221: {
1.134 ad 2222: pool_cache_t pc;
2223:
2224: mutex_enter(&pool_head_lock);
1.145 ad 2225: TAILQ_FOREACH(pc, &pool_cache_head, pc_cachelist) {
1.134 ad 2226: pc->pc_refcnt++;
2227: mutex_exit(&pool_head_lock);
1.43 thorpej 2228:
1.134 ad 2229: pool_cache_cpu_init1(ci, pc);
1.43 thorpej 2230:
1.134 ad 2231: mutex_enter(&pool_head_lock);
2232: pc->pc_refcnt--;
2233: cv_broadcast(&pool_busy);
2234: }
2235: mutex_exit(&pool_head_lock);
1.43 thorpej 2236: }
2237:
1.134 ad 2238: /*
2239: * pool_cache_reclaim:
2240: *
2241: * Reclaim memory from a pool cache.
2242: */
2243: bool
2244: pool_cache_reclaim(pool_cache_t pc)
1.43 thorpej 2245: {
2246:
1.134 ad 2247: return pool_reclaim(&pc->pc_pool);
2248: }
1.43 thorpej 2249:
1.136 yamt 2250: static void
2251: pool_cache_destruct_object1(pool_cache_t pc, void *object)
2252: {
2253:
2254: (*pc->pc_dtor)(pc->pc_arg, object);
2255: pool_put(&pc->pc_pool, object);
2256: }
2257:
1.134 ad 2258: /*
2259: * pool_cache_destruct_object:
2260: *
2261: * Force destruction of an object and its release back into
2262: * the pool.
2263: */
2264: void
2265: pool_cache_destruct_object(pool_cache_t pc, void *object)
2266: {
2267:
1.136 yamt 2268: FREECHECK_IN(&pc->pc_freecheck, object);
2269:
2270: pool_cache_destruct_object1(pc, object);
1.43 thorpej 2271: }
2272:
1.134 ad 2273: /*
2274: * pool_cache_invalidate_groups:
2275: *
2276: * Invalidate a chain of groups and destruct all objects.
2277: */
1.102 chs 2278: static void
1.134 ad 2279: pool_cache_invalidate_groups(pool_cache_t pc, pcg_t *pcg)
1.102 chs 2280: {
1.134 ad 2281: void *object;
2282: pcg_t *next;
2283: int i;
2284:
2285: for (; pcg != NULL; pcg = next) {
2286: next = pcg->pcg_next;
2287:
2288: for (i = 0; i < pcg->pcg_avail; i++) {
2289: object = pcg->pcg_objects[i].pcgo_va;
1.136 yamt 2290: pool_cache_destruct_object1(pc, object);
1.134 ad 2291: }
1.102 chs 2292:
1.142 ad 2293: if (pcg->pcg_size == PCG_NOBJECTS_LARGE) {
2294: pool_put(&pcg_large_pool, pcg);
2295: } else {
2296: KASSERT(pcg->pcg_size == PCG_NOBJECTS_NORMAL);
2297: pool_put(&pcg_normal_pool, pcg);
2298: }
1.102 chs 2299: }
2300: }
2301:
1.43 thorpej 2302: /*
1.134 ad 2303: * pool_cache_invalidate:
1.43 thorpej 2304: *
1.134 ad 2305: * Invalidate a pool cache (destruct and release all of the
2306: * cached objects). Does not reclaim objects from the pool.
1.176 thorpej 2307: *
2308: * Note: For pool caches that provide constructed objects, there
2309: * is an assumption that another level of synchronization is occurring
2310: * between the input to the constructor and the cache invalidation.
1.43 thorpej 2311: */
1.134 ad 2312: void
2313: pool_cache_invalidate(pool_cache_t pc)
2314: {
2315: pcg_t *full, *empty, *part;
1.176 thorpej 2316: uint64_t where;
2317:
1.177 jym 2318: if (ncpu < 2 || !mp_online) {
1.176 thorpej 2319: /*
2320: * We might be called early enough in the boot process
2321: * for the CPU data structures to not be fully initialized.
2322: * In this case, simply gather the local CPU's cache now
2323: * since it will be the only one running.
2324: */
2325: pool_cache_xcall(pc);
2326: } else {
2327: /*
2328: * Gather all of the CPU-specific caches into the
2329: * global cache.
2330: */
2331: where = xc_broadcast(0, (xcfunc_t)pool_cache_xcall, pc, NULL);
2332: xc_wait(where);
2333: }
1.134 ad 2334:
2335: mutex_enter(&pc->pc_lock);
2336: full = pc->pc_fullgroups;
2337: empty = pc->pc_emptygroups;
2338: part = pc->pc_partgroups;
2339: pc->pc_fullgroups = NULL;
2340: pc->pc_emptygroups = NULL;
2341: pc->pc_partgroups = NULL;
2342: pc->pc_nfull = 0;
2343: pc->pc_nempty = 0;
2344: pc->pc_npart = 0;
2345: mutex_exit(&pc->pc_lock);
2346:
2347: pool_cache_invalidate_groups(pc, full);
2348: pool_cache_invalidate_groups(pc, empty);
2349: pool_cache_invalidate_groups(pc, part);
2350: }
2351:
1.175 jym 2352: /*
2353: * pool_cache_invalidate_cpu:
2354: *
2355: * Invalidate all CPU-bound cached objects in pool cache, the CPU being
2356: * identified by its associated index.
2357: * It is caller's responsibility to ensure that no operation is
2358: * taking place on this pool cache while doing this invalidation.
2359: * WARNING: as no inter-CPU locking is enforced, trying to invalidate
2360: * pool cached objects from a CPU different from the one currently running
2361: * may result in an undefined behaviour.
2362: */
2363: static void
2364: pool_cache_invalidate_cpu(pool_cache_t pc, u_int index)
2365: {
2366:
2367: pool_cache_cpu_t *cc;
2368: pcg_t *pcg;
2369:
2370: if ((cc = pc->pc_cpus[index]) == NULL)
2371: return;
2372:
2373: if ((pcg = cc->cc_current) != &pcg_dummy) {
2374: pcg->pcg_next = NULL;
2375: pool_cache_invalidate_groups(pc, pcg);
2376: }
2377: if ((pcg = cc->cc_previous) != &pcg_dummy) {
2378: pcg->pcg_next = NULL;
2379: pool_cache_invalidate_groups(pc, pcg);
2380: }
2381: if (cc != &pc->pc_cpu0)
2382: pool_put(&cache_cpu_pool, cc);
2383:
2384: }
2385:
1.134 ad 2386: void
2387: pool_cache_set_drain_hook(pool_cache_t pc, void (*fn)(void *, int), void *arg)
2388: {
2389:
2390: pool_set_drain_hook(&pc->pc_pool, fn, arg);
2391: }
2392:
2393: void
2394: pool_cache_setlowat(pool_cache_t pc, int n)
2395: {
2396:
2397: pool_setlowat(&pc->pc_pool, n);
2398: }
2399:
2400: void
2401: pool_cache_sethiwat(pool_cache_t pc, int n)
2402: {
2403:
2404: pool_sethiwat(&pc->pc_pool, n);
2405: }
2406:
2407: void
2408: pool_cache_sethardlimit(pool_cache_t pc, int n, const char *warnmess, int ratecap)
2409: {
2410:
2411: pool_sethardlimit(&pc->pc_pool, n, warnmess, ratecap);
2412: }
2413:
1.162 ad 2414: static bool __noinline
2415: pool_cache_get_slow(pool_cache_cpu_t *cc, int s, void **objectp,
1.134 ad 2416: paddr_t *pap, int flags)
1.43 thorpej 2417: {
1.134 ad 2418: pcg_t *pcg, *cur;
2419: uint64_t ncsw;
2420: pool_cache_t pc;
1.43 thorpej 2421: void *object;
1.58 thorpej 2422:
1.168 yamt 2423: KASSERT(cc->cc_current->pcg_avail == 0);
2424: KASSERT(cc->cc_previous->pcg_avail == 0);
2425:
1.134 ad 2426: pc = cc->cc_cache;
2427: cc->cc_misses++;
1.43 thorpej 2428:
1.134 ad 2429: /*
2430: * Nothing was available locally. Try and grab a group
2431: * from the cache.
2432: */
1.162 ad 2433: if (__predict_false(!mutex_tryenter(&pc->pc_lock))) {
1.134 ad 2434: ncsw = curlwp->l_ncsw;
2435: mutex_enter(&pc->pc_lock);
2436: pc->pc_contended++;
1.43 thorpej 2437:
1.134 ad 2438: /*
2439: * If we context switched while locking, then
2440: * our view of the per-CPU data is invalid:
2441: * retry.
2442: */
2443: if (curlwp->l_ncsw != ncsw) {
2444: mutex_exit(&pc->pc_lock);
1.162 ad 2445: return true;
1.43 thorpej 2446: }
1.102 chs 2447: }
1.43 thorpej 2448:
1.162 ad 2449: if (__predict_true((pcg = pc->pc_fullgroups) != NULL)) {
1.43 thorpej 2450: /*
1.134 ad 2451: * If there's a full group, release our empty
2452: * group back to the cache. Install the full
2453: * group as cc_current and return.
1.43 thorpej 2454: */
1.162 ad 2455: if (__predict_true((cur = cc->cc_current) != &pcg_dummy)) {
1.134 ad 2456: KASSERT(cur->pcg_avail == 0);
2457: cur->pcg_next = pc->pc_emptygroups;
2458: pc->pc_emptygroups = cur;
2459: pc->pc_nempty++;
1.87 thorpej 2460: }
1.142 ad 2461: KASSERT(pcg->pcg_avail == pcg->pcg_size);
1.134 ad 2462: cc->cc_current = pcg;
2463: pc->pc_fullgroups = pcg->pcg_next;
2464: pc->pc_hits++;
2465: pc->pc_nfull--;
2466: mutex_exit(&pc->pc_lock);
1.162 ad 2467: return true;
1.134 ad 2468: }
2469:
2470: /*
2471: * Nothing available locally or in cache. Take the slow
2472: * path: fetch a new object from the pool and construct
2473: * it.
2474: */
2475: pc->pc_misses++;
2476: mutex_exit(&pc->pc_lock);
1.162 ad 2477: splx(s);
1.134 ad 2478:
2479: object = pool_get(&pc->pc_pool, flags);
2480: *objectp = object;
1.162 ad 2481: if (__predict_false(object == NULL))
2482: return false;
1.125 ad 2483:
1.162 ad 2484: if (__predict_false((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0)) {
1.134 ad 2485: pool_put(&pc->pc_pool, object);
2486: *objectp = NULL;
1.162 ad 2487: return false;
1.43 thorpej 2488: }
2489:
1.134 ad 2490: KASSERT((((vaddr_t)object + pc->pc_pool.pr_itemoffset) &
2491: (pc->pc_pool.pr_align - 1)) == 0);
1.43 thorpej 2492:
1.134 ad 2493: if (pap != NULL) {
2494: #ifdef POOL_VTOPHYS
2495: *pap = POOL_VTOPHYS(object);
2496: #else
2497: *pap = POOL_PADDR_INVALID;
2498: #endif
1.102 chs 2499: }
1.43 thorpej 2500:
1.125 ad 2501: FREECHECK_OUT(&pc->pc_freecheck, object);
1.162 ad 2502: return false;
1.43 thorpej 2503: }
2504:
2505: /*
1.134 ad 2506: * pool_cache_get{,_paddr}:
1.43 thorpej 2507: *
1.134 ad 2508: * Get an object from a pool cache (optionally returning
2509: * the physical address of the object).
1.43 thorpej 2510: */
1.134 ad 2511: void *
2512: pool_cache_get_paddr(pool_cache_t pc, int flags, paddr_t *pap)
1.43 thorpej 2513: {
1.134 ad 2514: pool_cache_cpu_t *cc;
2515: pcg_t *pcg;
2516: void *object;
1.60 thorpej 2517: int s;
1.43 thorpej 2518:
1.134 ad 2519: #ifdef LOCKDEBUG
1.155 ad 2520: if (flags & PR_WAITOK) {
1.154 yamt 2521: ASSERT_SLEEPABLE();
1.155 ad 2522: }
1.134 ad 2523: #endif
1.125 ad 2524:
1.162 ad 2525: /* Lock out interrupts and disable preemption. */
2526: s = splvm();
1.165 yamt 2527: while (/* CONSTCOND */ true) {
1.134 ad 2528: /* Try and allocate an object from the current group. */
1.162 ad 2529: cc = pc->pc_cpus[curcpu()->ci_index];
2530: KASSERT(cc->cc_cache == pc);
1.134 ad 2531: pcg = cc->cc_current;
1.162 ad 2532: if (__predict_true(pcg->pcg_avail > 0)) {
1.134 ad 2533: object = pcg->pcg_objects[--pcg->pcg_avail].pcgo_va;
1.162 ad 2534: if (__predict_false(pap != NULL))
1.134 ad 2535: *pap = pcg->pcg_objects[pcg->pcg_avail].pcgo_pa;
1.148 yamt 2536: #if defined(DIAGNOSTIC)
1.134 ad 2537: pcg->pcg_objects[pcg->pcg_avail].pcgo_va = NULL;
1.163 ad 2538: KASSERT(pcg->pcg_avail < pcg->pcg_size);
1.134 ad 2539: KASSERT(object != NULL);
1.163 ad 2540: #endif
1.134 ad 2541: cc->cc_hits++;
1.162 ad 2542: splx(s);
1.134 ad 2543: FREECHECK_OUT(&pc->pc_freecheck, object);
2544: return object;
1.43 thorpej 2545: }
2546:
2547: /*
1.134 ad 2548: * That failed. If the previous group isn't empty, swap
2549: * it with the current group and allocate from there.
1.43 thorpej 2550: */
1.134 ad 2551: pcg = cc->cc_previous;
1.162 ad 2552: if (__predict_true(pcg->pcg_avail > 0)) {
1.134 ad 2553: cc->cc_previous = cc->cc_current;
2554: cc->cc_current = pcg;
2555: continue;
1.43 thorpej 2556: }
2557:
1.134 ad 2558: /*
2559: * Can't allocate from either group: try the slow path.
2560: * If get_slow() allocated an object for us, or if
1.162 ad 2561: * no more objects are available, it will return false.
1.134 ad 2562: * Otherwise, we need to retry.
2563: */
1.165 yamt 2564: if (!pool_cache_get_slow(cc, s, &object, pap, flags))
2565: break;
2566: }
1.43 thorpej 2567:
1.134 ad 2568: return object;
1.51 thorpej 2569: }
2570:
1.162 ad 2571: static bool __noinline
2572: pool_cache_put_slow(pool_cache_cpu_t *cc, int s, void *object)
1.51 thorpej 2573: {
1.163 ad 2574: pcg_t *pcg, *cur;
1.134 ad 2575: uint64_t ncsw;
2576: pool_cache_t pc;
1.51 thorpej 2577:
1.168 yamt 2578: KASSERT(cc->cc_current->pcg_avail == cc->cc_current->pcg_size);
2579: KASSERT(cc->cc_previous->pcg_avail == cc->cc_previous->pcg_size);
2580:
1.134 ad 2581: pc = cc->cc_cache;
1.171 ad 2582: pcg = NULL;
1.134 ad 2583: cc->cc_misses++;
1.43 thorpej 2584:
1.171 ad 2585: /*
2586: * If there are no empty groups in the cache then allocate one
2587: * while still unlocked.
2588: */
2589: if (__predict_false(pc->pc_emptygroups == NULL)) {
2590: if (__predict_true(!pool_cache_disable)) {
2591: pcg = pool_get(pc->pc_pcgpool, PR_NOWAIT);
2592: }
2593: if (__predict_true(pcg != NULL)) {
2594: pcg->pcg_avail = 0;
2595: pcg->pcg_size = pc->pc_pcgsize;
2596: }
2597: }
2598:
1.162 ad 2599: /* Lock the cache. */
2600: if (__predict_false(!mutex_tryenter(&pc->pc_lock))) {
1.164 ad 2601: ncsw = curlwp->l_ncsw;
1.134 ad 2602: mutex_enter(&pc->pc_lock);
2603: pc->pc_contended++;
1.162 ad 2604:
1.163 ad 2605: /*
2606: * If we context switched while locking, then our view of
2607: * the per-CPU data is invalid: retry.
2608: */
2609: if (__predict_false(curlwp->l_ncsw != ncsw)) {
2610: mutex_exit(&pc->pc_lock);
1.171 ad 2611: if (pcg != NULL) {
2612: pool_put(pc->pc_pcgpool, pcg);
2613: }
1.163 ad 2614: return true;
2615: }
1.162 ad 2616: }
1.102 chs 2617:
1.163 ad 2618: /* If there are no empty groups in the cache then allocate one. */
1.171 ad 2619: if (pcg == NULL && pc->pc_emptygroups != NULL) {
2620: pcg = pc->pc_emptygroups;
1.163 ad 2621: pc->pc_emptygroups = pcg->pcg_next;
2622: pc->pc_nempty--;
1.134 ad 2623: }
1.130 ad 2624:
1.162 ad 2625: /*
2626: * If there's a empty group, release our full group back
2627: * to the cache. Install the empty group to the local CPU
2628: * and return.
2629: */
1.163 ad 2630: if (pcg != NULL) {
1.134 ad 2631: KASSERT(pcg->pcg_avail == 0);
1.162 ad 2632: if (__predict_false(cc->cc_previous == &pcg_dummy)) {
1.146 ad 2633: cc->cc_previous = pcg;
2634: } else {
1.162 ad 2635: cur = cc->cc_current;
2636: if (__predict_true(cur != &pcg_dummy)) {
1.163 ad 2637: KASSERT(cur->pcg_avail == cur->pcg_size);
1.146 ad 2638: cur->pcg_next = pc->pc_fullgroups;
2639: pc->pc_fullgroups = cur;
2640: pc->pc_nfull++;
2641: }
2642: cc->cc_current = pcg;
2643: }
1.163 ad 2644: pc->pc_hits++;
1.134 ad 2645: mutex_exit(&pc->pc_lock);
1.162 ad 2646: return true;
1.102 chs 2647: }
1.105 christos 2648:
1.134 ad 2649: /*
1.162 ad 2650: * Nothing available locally or in cache, and we didn't
2651: * allocate an empty group. Take the slow path and destroy
2652: * the object here and now.
1.134 ad 2653: */
2654: pc->pc_misses++;
2655: mutex_exit(&pc->pc_lock);
1.162 ad 2656: splx(s);
2657: pool_cache_destruct_object(pc, object);
1.105 christos 2658:
1.162 ad 2659: return false;
1.134 ad 2660: }
1.102 chs 2661:
1.43 thorpej 2662: /*
1.134 ad 2663: * pool_cache_put{,_paddr}:
1.43 thorpej 2664: *
1.134 ad 2665: * Put an object back to the pool cache (optionally caching the
2666: * physical address of the object).
1.43 thorpej 2667: */
1.101 thorpej 2668: void
1.134 ad 2669: pool_cache_put_paddr(pool_cache_t pc, void *object, paddr_t pa)
1.43 thorpej 2670: {
1.134 ad 2671: pool_cache_cpu_t *cc;
2672: pcg_t *pcg;
2673: int s;
1.101 thorpej 2674:
1.172 yamt 2675: KASSERT(object != NULL);
1.134 ad 2676: FREECHECK_IN(&pc->pc_freecheck, object);
1.101 thorpej 2677:
1.162 ad 2678: /* Lock out interrupts and disable preemption. */
2679: s = splvm();
1.165 yamt 2680: while (/* CONSTCOND */ true) {
1.134 ad 2681: /* If the current group isn't full, release it there. */
1.162 ad 2682: cc = pc->pc_cpus[curcpu()->ci_index];
2683: KASSERT(cc->cc_cache == pc);
1.134 ad 2684: pcg = cc->cc_current;
1.162 ad 2685: if (__predict_true(pcg->pcg_avail < pcg->pcg_size)) {
1.134 ad 2686: pcg->pcg_objects[pcg->pcg_avail].pcgo_va = object;
2687: pcg->pcg_objects[pcg->pcg_avail].pcgo_pa = pa;
2688: pcg->pcg_avail++;
2689: cc->cc_hits++;
1.162 ad 2690: splx(s);
1.134 ad 2691: return;
2692: }
1.43 thorpej 2693:
1.134 ad 2694: /*
1.162 ad 2695: * That failed. If the previous group isn't full, swap
1.134 ad 2696: * it with the current group and try again.
2697: */
2698: pcg = cc->cc_previous;
1.162 ad 2699: if (__predict_true(pcg->pcg_avail < pcg->pcg_size)) {
1.134 ad 2700: cc->cc_previous = cc->cc_current;
2701: cc->cc_current = pcg;
2702: continue;
2703: }
1.43 thorpej 2704:
1.134 ad 2705: /*
2706: * Can't free to either group: try the slow path.
2707: * If put_slow() releases the object for us, it
1.162 ad 2708: * will return false. Otherwise we need to retry.
1.134 ad 2709: */
1.165 yamt 2710: if (!pool_cache_put_slow(cc, s, object))
2711: break;
2712: }
1.43 thorpej 2713: }
2714:
2715: /*
1.134 ad 2716: * pool_cache_xcall:
1.43 thorpej 2717: *
1.134 ad 2718: * Transfer objects from the per-CPU cache to the global cache.
2719: * Run within a cross-call thread.
1.43 thorpej 2720: */
2721: static void
1.134 ad 2722: pool_cache_xcall(pool_cache_t pc)
1.43 thorpej 2723: {
1.134 ad 2724: pool_cache_cpu_t *cc;
2725: pcg_t *prev, *cur, **list;
1.162 ad 2726: int s;
1.134 ad 2727:
1.162 ad 2728: s = splvm();
2729: mutex_enter(&pc->pc_lock);
2730: cc = pc->pc_cpus[curcpu()->ci_index];
1.134 ad 2731: cur = cc->cc_current;
1.169 yamt 2732: cc->cc_current = __UNCONST(&pcg_dummy);
1.134 ad 2733: prev = cc->cc_previous;
1.169 yamt 2734: cc->cc_previous = __UNCONST(&pcg_dummy);
1.162 ad 2735: if (cur != &pcg_dummy) {
1.142 ad 2736: if (cur->pcg_avail == cur->pcg_size) {
1.134 ad 2737: list = &pc->pc_fullgroups;
2738: pc->pc_nfull++;
2739: } else if (cur->pcg_avail == 0) {
2740: list = &pc->pc_emptygroups;
2741: pc->pc_nempty++;
2742: } else {
2743: list = &pc->pc_partgroups;
2744: pc->pc_npart++;
2745: }
2746: cur->pcg_next = *list;
2747: *list = cur;
2748: }
1.162 ad 2749: if (prev != &pcg_dummy) {
1.142 ad 2750: if (prev->pcg_avail == prev->pcg_size) {
1.134 ad 2751: list = &pc->pc_fullgroups;
2752: pc->pc_nfull++;
2753: } else if (prev->pcg_avail == 0) {
2754: list = &pc->pc_emptygroups;
2755: pc->pc_nempty++;
2756: } else {
2757: list = &pc->pc_partgroups;
2758: pc->pc_npart++;
2759: }
2760: prev->pcg_next = *list;
2761: *list = prev;
2762: }
2763: mutex_exit(&pc->pc_lock);
2764: splx(s);
1.3 pk 2765: }
1.66 thorpej 2766:
2767: /*
2768: * Pool backend allocators.
2769: *
2770: * Each pool has a backend allocator that handles allocation, deallocation,
2771: * and any additional draining that might be needed.
2772: *
2773: * We provide two standard allocators:
2774: *
2775: * pool_allocator_kmem - the default when no allocator is specified
2776: *
2777: * pool_allocator_nointr - used for pools that will not be accessed
2778: * in interrupt context.
2779: */
2780: void *pool_page_alloc(struct pool *, int);
2781: void pool_page_free(struct pool *, void *);
2782:
1.112 bjh21 2783: #ifdef POOL_SUBPAGE
2784: struct pool_allocator pool_allocator_kmem_fullpage = {
2785: pool_page_alloc, pool_page_free, 0,
1.117 yamt 2786: .pa_backingmapptr = &kmem_map,
1.112 bjh21 2787: };
2788: #else
1.66 thorpej 2789: struct pool_allocator pool_allocator_kmem = {
2790: pool_page_alloc, pool_page_free, 0,
1.117 yamt 2791: .pa_backingmapptr = &kmem_map,
1.66 thorpej 2792: };
1.112 bjh21 2793: #endif
1.66 thorpej 2794:
2795: void *pool_page_alloc_nointr(struct pool *, int);
2796: void pool_page_free_nointr(struct pool *, void *);
2797:
1.112 bjh21 2798: #ifdef POOL_SUBPAGE
2799: struct pool_allocator pool_allocator_nointr_fullpage = {
2800: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1.117 yamt 2801: .pa_backingmapptr = &kernel_map,
1.112 bjh21 2802: };
2803: #else
1.66 thorpej 2804: struct pool_allocator pool_allocator_nointr = {
2805: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1.117 yamt 2806: .pa_backingmapptr = &kernel_map,
1.66 thorpej 2807: };
1.112 bjh21 2808: #endif
1.66 thorpej 2809:
2810: #ifdef POOL_SUBPAGE
2811: void *pool_subpage_alloc(struct pool *, int);
2812: void pool_subpage_free(struct pool *, void *);
2813:
1.112 bjh21 2814: struct pool_allocator pool_allocator_kmem = {
2815: pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE,
1.117 yamt 2816: .pa_backingmapptr = &kmem_map,
1.112 bjh21 2817: };
2818:
2819: void *pool_subpage_alloc_nointr(struct pool *, int);
2820: void pool_subpage_free_nointr(struct pool *, void *);
2821:
2822: struct pool_allocator pool_allocator_nointr = {
2823: pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE,
1.117 yamt 2824: .pa_backingmapptr = &kmem_map,
1.66 thorpej 2825: };
2826: #endif /* POOL_SUBPAGE */
2827:
1.117 yamt 2828: static void *
2829: pool_allocator_alloc(struct pool *pp, int flags)
1.66 thorpej 2830: {
1.117 yamt 2831: struct pool_allocator *pa = pp->pr_alloc;
1.66 thorpej 2832: void *res;
2833:
1.117 yamt 2834: res = (*pa->pa_alloc)(pp, flags);
2835: if (res == NULL && (flags & PR_WAITOK) == 0) {
1.66 thorpej 2836: /*
1.117 yamt 2837: * We only run the drain hook here if PR_NOWAIT.
2838: * In other cases, the hook will be run in
2839: * pool_reclaim().
1.66 thorpej 2840: */
1.117 yamt 2841: if (pp->pr_drain_hook != NULL) {
2842: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
2843: res = (*pa->pa_alloc)(pp, flags);
1.66 thorpej 2844: }
1.117 yamt 2845: }
2846: return res;
1.66 thorpej 2847: }
2848:
1.117 yamt 2849: static void
1.66 thorpej 2850: pool_allocator_free(struct pool *pp, void *v)
2851: {
2852: struct pool_allocator *pa = pp->pr_alloc;
2853:
2854: (*pa->pa_free)(pp, v);
2855: }
2856:
2857: void *
1.124 yamt 2858: pool_page_alloc(struct pool *pp, int flags)
1.66 thorpej 2859: {
1.127 thorpej 2860: bool waitok = (flags & PR_WAITOK) ? true : false;
1.66 thorpej 2861:
1.100 yamt 2862: return ((void *) uvm_km_alloc_poolpage_cache(kmem_map, waitok));
1.66 thorpej 2863: }
2864:
2865: void
1.124 yamt 2866: pool_page_free(struct pool *pp, void *v)
1.66 thorpej 2867: {
2868:
1.98 yamt 2869: uvm_km_free_poolpage_cache(kmem_map, (vaddr_t) v);
2870: }
2871:
2872: static void *
1.124 yamt 2873: pool_page_alloc_meta(struct pool *pp, int flags)
1.98 yamt 2874: {
1.127 thorpej 2875: bool waitok = (flags & PR_WAITOK) ? true : false;
1.98 yamt 2876:
1.100 yamt 2877: return ((void *) uvm_km_alloc_poolpage(kmem_map, waitok));
1.98 yamt 2878: }
2879:
2880: static void
1.124 yamt 2881: pool_page_free_meta(struct pool *pp, void *v)
1.98 yamt 2882: {
2883:
1.100 yamt 2884: uvm_km_free_poolpage(kmem_map, (vaddr_t) v);
1.66 thorpej 2885: }
2886:
2887: #ifdef POOL_SUBPAGE
2888: /* Sub-page allocator, for machines with large hardware pages. */
2889: void *
2890: pool_subpage_alloc(struct pool *pp, int flags)
2891: {
1.134 ad 2892: return pool_get(&psppool, flags);
1.66 thorpej 2893: }
2894:
2895: void
2896: pool_subpage_free(struct pool *pp, void *v)
2897: {
2898: pool_put(&psppool, v);
2899: }
2900:
2901: /* We don't provide a real nointr allocator. Maybe later. */
2902: void *
1.112 bjh21 2903: pool_subpage_alloc_nointr(struct pool *pp, int flags)
1.66 thorpej 2904: {
2905:
2906: return (pool_subpage_alloc(pp, flags));
2907: }
2908:
2909: void
1.112 bjh21 2910: pool_subpage_free_nointr(struct pool *pp, void *v)
1.66 thorpej 2911: {
2912:
2913: pool_subpage_free(pp, v);
2914: }
1.112 bjh21 2915: #endif /* POOL_SUBPAGE */
1.66 thorpej 2916: void *
1.124 yamt 2917: pool_page_alloc_nointr(struct pool *pp, int flags)
1.66 thorpej 2918: {
1.127 thorpej 2919: bool waitok = (flags & PR_WAITOK) ? true : false;
1.66 thorpej 2920:
1.100 yamt 2921: return ((void *) uvm_km_alloc_poolpage_cache(kernel_map, waitok));
1.66 thorpej 2922: }
2923:
2924: void
1.124 yamt 2925: pool_page_free_nointr(struct pool *pp, void *v)
1.66 thorpej 2926: {
2927:
1.98 yamt 2928: uvm_km_free_poolpage_cache(kernel_map, (vaddr_t) v);
1.66 thorpej 2929: }
1.141 yamt 2930:
2931: #if defined(DDB)
2932: static bool
2933: pool_in_page(struct pool *pp, struct pool_item_header *ph, uintptr_t addr)
2934: {
2935:
2936: return (uintptr_t)ph->ph_page <= addr &&
2937: addr < (uintptr_t)ph->ph_page + pp->pr_alloc->pa_pagesz;
2938: }
2939:
1.143 yamt 2940: static bool
2941: pool_in_item(struct pool *pp, void *item, uintptr_t addr)
2942: {
2943:
2944: return (uintptr_t)item <= addr && addr < (uintptr_t)item + pp->pr_size;
2945: }
2946:
2947: static bool
2948: pool_in_cg(struct pool *pp, struct pool_cache_group *pcg, uintptr_t addr)
2949: {
2950: int i;
2951:
2952: if (pcg == NULL) {
2953: return false;
2954: }
1.144 yamt 2955: for (i = 0; i < pcg->pcg_avail; i++) {
1.143 yamt 2956: if (pool_in_item(pp, pcg->pcg_objects[i].pcgo_va, addr)) {
2957: return true;
2958: }
2959: }
2960: return false;
2961: }
2962:
2963: static bool
2964: pool_allocated(struct pool *pp, struct pool_item_header *ph, uintptr_t addr)
2965: {
2966:
2967: if ((pp->pr_roflags & PR_NOTOUCH) != 0) {
2968: unsigned int idx = pr_item_notouch_index(pp, ph, (void *)addr);
2969: pool_item_bitmap_t *bitmap =
2970: ph->ph_bitmap + (idx / BITMAP_SIZE);
2971: pool_item_bitmap_t mask = 1 << (idx & BITMAP_MASK);
2972:
2973: return (*bitmap & mask) == 0;
2974: } else {
2975: struct pool_item *pi;
2976:
2977: LIST_FOREACH(pi, &ph->ph_itemlist, pi_list) {
2978: if (pool_in_item(pp, pi, addr)) {
2979: return false;
2980: }
2981: }
2982: return true;
2983: }
2984: }
2985:
1.141 yamt 2986: void
2987: pool_whatis(uintptr_t addr, void (*pr)(const char *, ...))
2988: {
2989: struct pool *pp;
2990:
1.145 ad 2991: TAILQ_FOREACH(pp, &pool_head, pr_poollist) {
1.141 yamt 2992: struct pool_item_header *ph;
2993: uintptr_t item;
1.143 yamt 2994: bool allocated = true;
2995: bool incache = false;
2996: bool incpucache = false;
2997: char cpucachestr[32];
1.141 yamt 2998:
2999: if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
3000: LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
3001: if (pool_in_page(pp, ph, addr)) {
3002: goto found;
3003: }
3004: }
3005: LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
3006: if (pool_in_page(pp, ph, addr)) {
1.143 yamt 3007: allocated =
3008: pool_allocated(pp, ph, addr);
3009: goto found;
3010: }
3011: }
3012: LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
3013: if (pool_in_page(pp, ph, addr)) {
3014: allocated = false;
1.141 yamt 3015: goto found;
3016: }
3017: }
3018: continue;
3019: } else {
3020: ph = pr_find_pagehead_noalign(pp, (void *)addr);
3021: if (ph == NULL || !pool_in_page(pp, ph, addr)) {
3022: continue;
3023: }
1.143 yamt 3024: allocated = pool_allocated(pp, ph, addr);
1.141 yamt 3025: }
3026: found:
1.143 yamt 3027: if (allocated && pp->pr_cache) {
3028: pool_cache_t pc = pp->pr_cache;
3029: struct pool_cache_group *pcg;
3030: int i;
3031:
3032: for (pcg = pc->pc_fullgroups; pcg != NULL;
3033: pcg = pcg->pcg_next) {
3034: if (pool_in_cg(pp, pcg, addr)) {
3035: incache = true;
3036: goto print;
3037: }
3038: }
3039: for (i = 0; i < MAXCPUS; i++) {
3040: pool_cache_cpu_t *cc;
3041:
3042: if ((cc = pc->pc_cpus[i]) == NULL) {
3043: continue;
3044: }
3045: if (pool_in_cg(pp, cc->cc_current, addr) ||
3046: pool_in_cg(pp, cc->cc_previous, addr)) {
3047: struct cpu_info *ci =
1.170 ad 3048: cpu_lookup(i);
1.143 yamt 3049:
3050: incpucache = true;
3051: snprintf(cpucachestr,
3052: sizeof(cpucachestr),
3053: "cached by CPU %u",
1.153 martin 3054: ci->ci_index);
1.143 yamt 3055: goto print;
3056: }
3057: }
3058: }
3059: print:
1.141 yamt 3060: item = (uintptr_t)ph->ph_page + ph->ph_off;
3061: item = item + rounddown(addr - item, pp->pr_size);
1.143 yamt 3062: (*pr)("%p is %p+%zu in POOL '%s' (%s)\n",
1.141 yamt 3063: (void *)addr, item, (size_t)(addr - item),
1.143 yamt 3064: pp->pr_wchan,
3065: incpucache ? cpucachestr :
3066: incache ? "cached" : allocated ? "allocated" : "free");
1.141 yamt 3067: }
3068: }
3069: #endif /* defined(DDB) */
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