Annotation of src/sys/kern/subr_pool.c, Revision 1.158.2.3
1.158.2.3! yamt 1: /* $NetBSD: subr_pool.c,v 1.158.2.2 2009/05/04 08:13:48 yamt Exp $ */
1.1 pk 2:
3: /*-
1.158.2.2 yamt 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>
1.158.2.3! yamt 34: __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.158.2.2 2009/05/04 08:13:48 yamt 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 */
1.158.2.3! yamt 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:
1.135 yamt 107: typedef uint32_t pool_item_bitmap_t;
108: #define BITMAP_SIZE (CHAR_BIT * sizeof(pool_item_bitmap_t))
109: #define BITMAP_MASK (BITMAP_SIZE - 1)
1.99 yamt 110:
1.3 pk 111: struct pool_item_header {
112: /* Page headers */
1.88 chs 113: LIST_ENTRY(pool_item_header)
1.3 pk 114: ph_pagelist; /* pool page list */
1.88 chs 115: SPLAY_ENTRY(pool_item_header)
116: ph_node; /* Off-page page headers */
1.128 christos 117: void * ph_page; /* this page's address */
1.151 yamt 118: uint32_t ph_time; /* last referenced */
1.135 yamt 119: uint16_t ph_nmissing; /* # of chunks in use */
1.141 yamt 120: uint16_t ph_off; /* start offset in page */
1.97 yamt 121: union {
122: /* !PR_NOTOUCH */
123: struct {
1.102 chs 124: LIST_HEAD(, pool_item)
1.97 yamt 125: phu_itemlist; /* chunk list for this page */
126: } phu_normal;
127: /* PR_NOTOUCH */
128: struct {
1.141 yamt 129: pool_item_bitmap_t phu_bitmap[1];
1.97 yamt 130: } phu_notouch;
131: } ph_u;
1.3 pk 132: };
1.97 yamt 133: #define ph_itemlist ph_u.phu_normal.phu_itemlist
1.135 yamt 134: #define ph_bitmap ph_u.phu_notouch.phu_bitmap
1.3 pk 135:
1.1 pk 136: struct pool_item {
1.3 pk 137: #ifdef DIAGNOSTIC
1.82 thorpej 138: u_int pi_magic;
1.33 chs 139: #endif
1.134 ad 140: #define PI_MAGIC 0xdeaddeadU
1.3 pk 141: /* Other entries use only this list entry */
1.102 chs 142: LIST_ENTRY(pool_item) pi_list;
1.3 pk 143: };
144:
1.53 thorpej 145: #define POOL_NEEDS_CATCHUP(pp) \
146: ((pp)->pr_nitems < (pp)->pr_minitems)
147:
1.43 thorpej 148: /*
149: * Pool cache management.
150: *
151: * Pool caches provide a way for constructed objects to be cached by the
152: * pool subsystem. This can lead to performance improvements by avoiding
153: * needless object construction/destruction; it is deferred until absolutely
154: * necessary.
155: *
1.134 ad 156: * Caches are grouped into cache groups. Each cache group references up
157: * to PCG_NUMOBJECTS constructed objects. When a cache allocates an
158: * object from the pool, it calls the object's constructor and places it
159: * into a cache group. When a cache group frees an object back to the
160: * pool, it first calls the object's destructor. This allows the object
161: * to persist in constructed form while freed to the cache.
162: *
163: * The pool references each cache, so that when a pool is drained by the
164: * pagedaemon, it can drain each individual cache as well. Each time a
165: * cache is drained, the most idle cache group is freed to the pool in
166: * its entirety.
1.43 thorpej 167: *
168: * Pool caches are layed on top of pools. By layering them, we can avoid
169: * the complexity of cache management for pools which would not benefit
170: * from it.
171: */
172:
1.142 ad 173: static struct pool pcg_normal_pool;
174: static struct pool pcg_large_pool;
1.134 ad 175: static struct pool cache_pool;
176: static struct pool cache_cpu_pool;
1.3 pk 177:
1.145 ad 178: /* List of all caches. */
179: TAILQ_HEAD(,pool_cache) pool_cache_head =
180: TAILQ_HEAD_INITIALIZER(pool_cache_head);
181:
1.158.2.2 yamt 182: int pool_cache_disable; /* global disable for caching */
183: static const pcg_t pcg_dummy; /* zero sized: always empty, yet always full */
1.145 ad 184:
1.158.2.2 yamt 185: static bool pool_cache_put_slow(pool_cache_cpu_t *, int,
186: void *);
187: static bool pool_cache_get_slow(pool_cache_cpu_t *, int,
188: void **, paddr_t *, int);
1.134 ad 189: static void pool_cache_cpu_init1(struct cpu_info *, pool_cache_t);
190: static void pool_cache_invalidate_groups(pool_cache_t, pcg_t *);
191: static void pool_cache_xcall(pool_cache_t);
1.3 pk 192:
1.42 thorpej 193: static int pool_catchup(struct pool *);
1.128 christos 194: static void pool_prime_page(struct pool *, void *,
1.55 thorpej 195: struct pool_item_header *);
1.88 chs 196: static void pool_update_curpage(struct pool *);
1.66 thorpej 197:
1.113 yamt 198: static int pool_grow(struct pool *, int);
1.117 yamt 199: static void *pool_allocator_alloc(struct pool *, int);
200: static void pool_allocator_free(struct pool *, void *);
1.3 pk 201:
1.97 yamt 202: static void pool_print_pagelist(struct pool *, struct pool_pagelist *,
1.88 chs 203: void (*)(const char *, ...));
1.42 thorpej 204: static void pool_print1(struct pool *, const char *,
205: void (*)(const char *, ...));
1.3 pk 206:
1.88 chs 207: static int pool_chk_page(struct pool *, const char *,
208: struct pool_item_header *);
209:
1.3 pk 210: /*
1.52 thorpej 211: * Pool log entry. An array of these is allocated in pool_init().
1.3 pk 212: */
213: struct pool_log {
214: const char *pl_file;
215: long pl_line;
216: int pl_action;
1.25 thorpej 217: #define PRLOG_GET 1
218: #define PRLOG_PUT 2
1.3 pk 219: void *pl_addr;
1.1 pk 220: };
221:
1.86 matt 222: #ifdef POOL_DIAGNOSTIC
1.3 pk 223: /* Number of entries in pool log buffers */
1.17 thorpej 224: #ifndef POOL_LOGSIZE
225: #define POOL_LOGSIZE 10
226: #endif
227:
228: int pool_logsize = POOL_LOGSIZE;
1.1 pk 229:
1.110 perry 230: static inline void
1.42 thorpej 231: pr_log(struct pool *pp, void *v, int action, const char *file, long line)
1.3 pk 232: {
233: int n = pp->pr_curlogentry;
234: struct pool_log *pl;
235:
1.20 thorpej 236: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 237: return;
238:
239: /*
240: * Fill in the current entry. Wrap around and overwrite
241: * the oldest entry if necessary.
242: */
243: pl = &pp->pr_log[n];
244: pl->pl_file = file;
245: pl->pl_line = line;
246: pl->pl_action = action;
247: pl->pl_addr = v;
248: if (++n >= pp->pr_logsize)
249: n = 0;
250: pp->pr_curlogentry = n;
251: }
252:
253: static void
1.42 thorpej 254: pr_printlog(struct pool *pp, struct pool_item *pi,
255: void (*pr)(const char *, ...))
1.3 pk 256: {
257: int i = pp->pr_logsize;
258: int n = pp->pr_curlogentry;
259:
1.20 thorpej 260: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 261: return;
262:
263: /*
264: * Print all entries in this pool's log.
265: */
266: while (i-- > 0) {
267: struct pool_log *pl = &pp->pr_log[n];
268: if (pl->pl_action != 0) {
1.25 thorpej 269: if (pi == NULL || pi == pl->pl_addr) {
270: (*pr)("\tlog entry %d:\n", i);
271: (*pr)("\t\taction = %s, addr = %p\n",
272: pl->pl_action == PRLOG_GET ? "get" : "put",
273: pl->pl_addr);
274: (*pr)("\t\tfile: %s at line %lu\n",
275: pl->pl_file, pl->pl_line);
276: }
1.3 pk 277: }
278: if (++n >= pp->pr_logsize)
279: n = 0;
280: }
281: }
1.25 thorpej 282:
1.110 perry 283: static inline void
1.42 thorpej 284: pr_enter(struct pool *pp, const char *file, long line)
1.25 thorpej 285: {
286:
1.34 thorpej 287: if (__predict_false(pp->pr_entered_file != NULL)) {
1.25 thorpej 288: printf("pool %s: reentrancy at file %s line %ld\n",
289: pp->pr_wchan, file, line);
290: printf(" previous entry at file %s line %ld\n",
291: pp->pr_entered_file, pp->pr_entered_line);
292: panic("pr_enter");
293: }
294:
295: pp->pr_entered_file = file;
296: pp->pr_entered_line = line;
297: }
298:
1.110 perry 299: static inline void
1.42 thorpej 300: pr_leave(struct pool *pp)
1.25 thorpej 301: {
302:
1.34 thorpej 303: if (__predict_false(pp->pr_entered_file == NULL)) {
1.25 thorpej 304: printf("pool %s not entered?\n", pp->pr_wchan);
305: panic("pr_leave");
306: }
307:
308: pp->pr_entered_file = NULL;
309: pp->pr_entered_line = 0;
310: }
311:
1.110 perry 312: static inline void
1.42 thorpej 313: pr_enter_check(struct pool *pp, void (*pr)(const char *, ...))
1.25 thorpej 314: {
315:
316: if (pp->pr_entered_file != NULL)
317: (*pr)("\n\tcurrently entered from file %s line %ld\n",
318: pp->pr_entered_file, pp->pr_entered_line);
319: }
1.3 pk 320: #else
1.25 thorpej 321: #define pr_log(pp, v, action, file, line)
322: #define pr_printlog(pp, pi, pr)
323: #define pr_enter(pp, file, line)
324: #define pr_leave(pp)
325: #define pr_enter_check(pp, pr)
1.59 thorpej 326: #endif /* POOL_DIAGNOSTIC */
1.3 pk 327:
1.135 yamt 328: static inline unsigned int
1.97 yamt 329: pr_item_notouch_index(const struct pool *pp, const struct pool_item_header *ph,
330: const void *v)
331: {
332: const char *cp = v;
1.135 yamt 333: unsigned int idx;
1.97 yamt 334:
335: KASSERT(pp->pr_roflags & PR_NOTOUCH);
1.128 christos 336: idx = (cp - (char *)ph->ph_page - ph->ph_off) / pp->pr_size;
1.97 yamt 337: KASSERT(idx < pp->pr_itemsperpage);
338: return idx;
339: }
340:
1.110 perry 341: static inline void
1.97 yamt 342: pr_item_notouch_put(const struct pool *pp, struct pool_item_header *ph,
343: void *obj)
344: {
1.135 yamt 345: unsigned int idx = pr_item_notouch_index(pp, ph, obj);
346: pool_item_bitmap_t *bitmap = ph->ph_bitmap + (idx / BITMAP_SIZE);
347: pool_item_bitmap_t mask = 1 << (idx & BITMAP_MASK);
1.97 yamt 348:
1.135 yamt 349: KASSERT((*bitmap & mask) == 0);
350: *bitmap |= mask;
1.97 yamt 351: }
352:
1.110 perry 353: static inline void *
1.97 yamt 354: pr_item_notouch_get(const struct pool *pp, struct pool_item_header *ph)
355: {
1.135 yamt 356: pool_item_bitmap_t *bitmap = ph->ph_bitmap;
357: unsigned int idx;
358: int i;
1.97 yamt 359:
1.135 yamt 360: for (i = 0; ; i++) {
361: int bit;
1.97 yamt 362:
1.135 yamt 363: KASSERT((i * BITMAP_SIZE) < pp->pr_itemsperpage);
364: bit = ffs32(bitmap[i]);
365: if (bit) {
366: pool_item_bitmap_t mask;
367:
368: bit--;
369: idx = (i * BITMAP_SIZE) + bit;
370: mask = 1 << bit;
371: KASSERT((bitmap[i] & mask) != 0);
372: bitmap[i] &= ~mask;
373: break;
374: }
375: }
376: KASSERT(idx < pp->pr_itemsperpage);
1.128 christos 377: return (char *)ph->ph_page + ph->ph_off + idx * pp->pr_size;
1.97 yamt 378: }
379:
1.135 yamt 380: static inline void
1.141 yamt 381: pr_item_notouch_init(const struct pool *pp, struct pool_item_header *ph)
1.135 yamt 382: {
383: pool_item_bitmap_t *bitmap = ph->ph_bitmap;
384: const int n = howmany(pp->pr_itemsperpage, BITMAP_SIZE);
385: int i;
386:
387: for (i = 0; i < n; i++) {
388: bitmap[i] = (pool_item_bitmap_t)-1;
389: }
390: }
391:
1.110 perry 392: static inline int
1.88 chs 393: phtree_compare(struct pool_item_header *a, struct pool_item_header *b)
394: {
1.121 yamt 395:
396: /*
397: * we consider pool_item_header with smaller ph_page bigger.
398: * (this unnatural ordering is for the benefit of pr_find_pagehead.)
399: */
400:
1.88 chs 401: if (a->ph_page < b->ph_page)
1.121 yamt 402: return (1);
403: else if (a->ph_page > b->ph_page)
1.88 chs 404: return (-1);
405: else
406: return (0);
407: }
408:
409: SPLAY_PROTOTYPE(phtree, pool_item_header, ph_node, phtree_compare);
410: SPLAY_GENERATE(phtree, pool_item_header, ph_node, phtree_compare);
411:
1.141 yamt 412: static inline struct pool_item_header *
413: pr_find_pagehead_noalign(struct pool *pp, void *v)
414: {
415: struct pool_item_header *ph, tmp;
416:
417: tmp.ph_page = (void *)(uintptr_t)v;
418: ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
419: if (ph == NULL) {
420: ph = SPLAY_ROOT(&pp->pr_phtree);
421: if (ph != NULL && phtree_compare(&tmp, ph) >= 0) {
422: ph = SPLAY_NEXT(phtree, &pp->pr_phtree, ph);
423: }
424: KASSERT(ph == NULL || phtree_compare(&tmp, ph) < 0);
425: }
426:
427: return ph;
428: }
429:
1.3 pk 430: /*
1.121 yamt 431: * Return the pool page header based on item address.
1.3 pk 432: */
1.110 perry 433: static inline struct pool_item_header *
1.121 yamt 434: pr_find_pagehead(struct pool *pp, void *v)
1.3 pk 435: {
1.88 chs 436: struct pool_item_header *ph, tmp;
1.3 pk 437:
1.121 yamt 438: if ((pp->pr_roflags & PR_NOALIGN) != 0) {
1.141 yamt 439: ph = pr_find_pagehead_noalign(pp, v);
1.121 yamt 440: } else {
1.128 christos 441: void *page =
442: (void *)((uintptr_t)v & pp->pr_alloc->pa_pagemask);
1.121 yamt 443:
444: if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
1.128 christos 445: ph = (struct pool_item_header *)((char *)page + pp->pr_phoffset);
1.121 yamt 446: } else {
447: tmp.ph_page = page;
448: ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
449: }
450: }
1.3 pk 451:
1.121 yamt 452: KASSERT(ph == NULL || ((pp->pr_roflags & PR_PHINPAGE) != 0) ||
1.128 christos 453: ((char *)ph->ph_page <= (char *)v &&
454: (char *)v < (char *)ph->ph_page + pp->pr_alloc->pa_pagesz));
1.88 chs 455: return ph;
1.3 pk 456: }
457:
1.101 thorpej 458: static void
459: pr_pagelist_free(struct pool *pp, struct pool_pagelist *pq)
460: {
461: struct pool_item_header *ph;
462:
463: while ((ph = LIST_FIRST(pq)) != NULL) {
464: LIST_REMOVE(ph, ph_pagelist);
465: pool_allocator_free(pp, ph->ph_page);
1.134 ad 466: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1.101 thorpej 467: pool_put(pp->pr_phpool, ph);
468: }
469: }
470:
1.3 pk 471: /*
472: * Remove a page from the pool.
473: */
1.110 perry 474: static inline void
1.61 chs 475: pr_rmpage(struct pool *pp, struct pool_item_header *ph,
476: struct pool_pagelist *pq)
1.3 pk 477: {
478:
1.134 ad 479: KASSERT(mutex_owned(&pp->pr_lock));
1.91 yamt 480:
1.3 pk 481: /*
1.7 thorpej 482: * If the page was idle, decrement the idle page count.
1.3 pk 483: */
1.6 thorpej 484: if (ph->ph_nmissing == 0) {
485: #ifdef DIAGNOSTIC
486: if (pp->pr_nidle == 0)
487: panic("pr_rmpage: nidle inconsistent");
1.20 thorpej 488: if (pp->pr_nitems < pp->pr_itemsperpage)
489: panic("pr_rmpage: nitems inconsistent");
1.6 thorpej 490: #endif
491: pp->pr_nidle--;
492: }
1.7 thorpej 493:
1.20 thorpej 494: pp->pr_nitems -= pp->pr_itemsperpage;
495:
1.7 thorpej 496: /*
1.101 thorpej 497: * Unlink the page from the pool and queue it for release.
1.7 thorpej 498: */
1.88 chs 499: LIST_REMOVE(ph, ph_pagelist);
1.91 yamt 500: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
501: SPLAY_REMOVE(phtree, &pp->pr_phtree, ph);
1.101 thorpej 502: LIST_INSERT_HEAD(pq, ph, ph_pagelist);
503:
1.7 thorpej 504: pp->pr_npages--;
505: pp->pr_npagefree++;
1.6 thorpej 506:
1.88 chs 507: pool_update_curpage(pp);
1.3 pk 508: }
509:
1.126 thorpej 510: static bool
1.117 yamt 511: pa_starved_p(struct pool_allocator *pa)
512: {
513:
514: if (pa->pa_backingmap != NULL) {
515: return vm_map_starved_p(pa->pa_backingmap);
516: }
1.127 thorpej 517: return false;
1.117 yamt 518: }
519:
520: static int
1.124 yamt 521: pool_reclaim_callback(struct callback_entry *ce, void *obj, void *arg)
1.117 yamt 522: {
523: struct pool *pp = obj;
524: struct pool_allocator *pa = pp->pr_alloc;
525:
526: KASSERT(&pp->pr_reclaimerentry == ce);
527: pool_reclaim(pp);
528: if (!pa_starved_p(pa)) {
529: return CALLBACK_CHAIN_ABORT;
530: }
531: return CALLBACK_CHAIN_CONTINUE;
532: }
533:
534: static void
535: pool_reclaim_register(struct pool *pp)
536: {
537: struct vm_map *map = pp->pr_alloc->pa_backingmap;
538: int s;
539:
540: if (map == NULL) {
541: return;
542: }
543:
544: s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */
545: callback_register(&vm_map_to_kernel(map)->vmk_reclaim_callback,
546: &pp->pr_reclaimerentry, pp, pool_reclaim_callback);
547: splx(s);
548: }
549:
550: static void
551: pool_reclaim_unregister(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_unregister(&vm_map_to_kernel(map)->vmk_reclaim_callback,
562: &pp->pr_reclaimerentry);
563: splx(s);
564: }
565:
566: static void
567: pa_reclaim_register(struct pool_allocator *pa)
568: {
569: struct vm_map *map = *pa->pa_backingmapptr;
570: struct pool *pp;
571:
572: KASSERT(pa->pa_backingmap == NULL);
573: if (map == NULL) {
574: SLIST_INSERT_HEAD(&pa_deferinitq, pa, pa_q);
575: return;
576: }
577: pa->pa_backingmap = map;
578: TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
579: pool_reclaim_register(pp);
580: }
581: }
582:
1.3 pk 583: /*
1.94 simonb 584: * Initialize all the pools listed in the "pools" link set.
585: */
586: void
1.117 yamt 587: pool_subsystem_init(void)
1.94 simonb 588: {
1.117 yamt 589: struct pool_allocator *pa;
1.94 simonb 590:
1.134 ad 591: mutex_init(&pool_head_lock, MUTEX_DEFAULT, IPL_NONE);
592: cv_init(&pool_busy, "poolbusy");
593:
1.117 yamt 594: while ((pa = SLIST_FIRST(&pa_deferinitq)) != NULL) {
595: KASSERT(pa->pa_backingmapptr != NULL);
596: KASSERT(*pa->pa_backingmapptr != NULL);
597: SLIST_REMOVE_HEAD(&pa_deferinitq, pa_q);
598: pa_reclaim_register(pa);
599: }
1.134 ad 600:
1.156 ad 601: pool_init(&cache_pool, sizeof(struct pool_cache), coherency_unit,
1.134 ad 602: 0, 0, "pcache", &pool_allocator_nointr, IPL_NONE);
603:
1.156 ad 604: pool_init(&cache_cpu_pool, sizeof(pool_cache_cpu_t), coherency_unit,
1.134 ad 605: 0, 0, "pcachecpu", &pool_allocator_nointr, IPL_NONE);
1.94 simonb 606: }
607:
608: /*
1.3 pk 609: * Initialize the given pool resource structure.
610: *
611: * We export this routine to allow other kernel parts to declare
612: * static pools that must be initialized before malloc() is available.
613: */
614: void
1.42 thorpej 615: pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
1.129 ad 616: const char *wchan, struct pool_allocator *palloc, int ipl)
1.3 pk 617: {
1.116 simonb 618: struct pool *pp1;
1.92 enami 619: size_t trysize, phsize;
1.134 ad 620: int off, slack;
1.3 pk 621:
1.116 simonb 622: #ifdef DEBUG
623: /*
624: * Check that the pool hasn't already been initialised and
625: * added to the list of all pools.
626: */
1.145 ad 627: TAILQ_FOREACH(pp1, &pool_head, pr_poollist) {
1.116 simonb 628: if (pp == pp1)
629: panic("pool_init: pool %s already initialised",
630: wchan);
631: }
632: #endif
633:
1.25 thorpej 634: #ifdef POOL_DIAGNOSTIC
635: /*
636: * Always log if POOL_DIAGNOSTIC is defined.
637: */
638: if (pool_logsize != 0)
639: flags |= PR_LOGGING;
640: #endif
641:
1.66 thorpej 642: if (palloc == NULL)
643: palloc = &pool_allocator_kmem;
1.112 bjh21 644: #ifdef POOL_SUBPAGE
645: if (size > palloc->pa_pagesz) {
646: if (palloc == &pool_allocator_kmem)
647: palloc = &pool_allocator_kmem_fullpage;
648: else if (palloc == &pool_allocator_nointr)
649: palloc = &pool_allocator_nointr_fullpage;
650: }
1.66 thorpej 651: #endif /* POOL_SUBPAGE */
652: if ((palloc->pa_flags & PA_INITIALIZED) == 0) {
1.112 bjh21 653: if (palloc->pa_pagesz == 0)
1.66 thorpej 654: palloc->pa_pagesz = PAGE_SIZE;
655:
656: TAILQ_INIT(&palloc->pa_list);
657:
1.134 ad 658: mutex_init(&palloc->pa_lock, MUTEX_DEFAULT, IPL_VM);
1.66 thorpej 659: palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
660: palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
1.117 yamt 661:
662: if (palloc->pa_backingmapptr != NULL) {
663: pa_reclaim_register(palloc);
664: }
1.66 thorpej 665: palloc->pa_flags |= PA_INITIALIZED;
1.4 thorpej 666: }
1.3 pk 667:
668: if (align == 0)
669: align = ALIGN(1);
1.14 thorpej 670:
1.120 yamt 671: if ((flags & PR_NOTOUCH) == 0 && size < sizeof(struct pool_item))
1.14 thorpej 672: size = sizeof(struct pool_item);
1.3 pk 673:
1.78 thorpej 674: size = roundup(size, align);
1.66 thorpej 675: #ifdef DIAGNOSTIC
676: if (size > palloc->pa_pagesz)
1.121 yamt 677: panic("pool_init: pool item size (%zu) too large", size);
1.66 thorpej 678: #endif
1.35 pk 679:
1.3 pk 680: /*
681: * Initialize the pool structure.
682: */
1.88 chs 683: LIST_INIT(&pp->pr_emptypages);
684: LIST_INIT(&pp->pr_fullpages);
685: LIST_INIT(&pp->pr_partpages);
1.134 ad 686: pp->pr_cache = NULL;
1.3 pk 687: pp->pr_curpage = NULL;
688: pp->pr_npages = 0;
689: pp->pr_minitems = 0;
690: pp->pr_minpages = 0;
691: pp->pr_maxpages = UINT_MAX;
1.20 thorpej 692: pp->pr_roflags = flags;
693: pp->pr_flags = 0;
1.35 pk 694: pp->pr_size = size;
1.3 pk 695: pp->pr_align = align;
696: pp->pr_wchan = wchan;
1.66 thorpej 697: pp->pr_alloc = palloc;
1.20 thorpej 698: pp->pr_nitems = 0;
699: pp->pr_nout = 0;
700: pp->pr_hardlimit = UINT_MAX;
701: pp->pr_hardlimit_warning = NULL;
1.31 thorpej 702: pp->pr_hardlimit_ratecap.tv_sec = 0;
703: pp->pr_hardlimit_ratecap.tv_usec = 0;
704: pp->pr_hardlimit_warning_last.tv_sec = 0;
705: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.68 thorpej 706: pp->pr_drain_hook = NULL;
707: pp->pr_drain_hook_arg = NULL;
1.125 ad 708: pp->pr_freecheck = NULL;
1.3 pk 709:
710: /*
711: * Decide whether to put the page header off page to avoid
1.92 enami 712: * wasting too large a part of the page or too big item.
713: * Off-page page headers go on a hash table, so we can match
714: * a returned item with its header based on the page address.
715: * We use 1/16 of the page size and about 8 times of the item
716: * size as the threshold (XXX: tune)
717: *
718: * However, we'll put the header into the page if we can put
719: * it without wasting any items.
720: *
721: * Silently enforce `0 <= ioff < align'.
1.3 pk 722: */
1.92 enami 723: pp->pr_itemoffset = ioff %= align;
724: /* See the comment below about reserved bytes. */
725: trysize = palloc->pa_pagesz - ((align - ioff) % align);
726: phsize = ALIGN(sizeof(struct pool_item_header));
1.121 yamt 727: if ((pp->pr_roflags & (PR_NOTOUCH | PR_NOALIGN)) == 0 &&
1.97 yamt 728: (pp->pr_size < MIN(palloc->pa_pagesz / 16, phsize << 3) ||
729: trysize / pp->pr_size == (trysize - phsize) / pp->pr_size)) {
1.3 pk 730: /* Use the end of the page for the page header */
1.20 thorpej 731: pp->pr_roflags |= PR_PHINPAGE;
1.92 enami 732: pp->pr_phoffset = off = palloc->pa_pagesz - phsize;
1.2 pk 733: } else {
1.3 pk 734: /* The page header will be taken from our page header pool */
735: pp->pr_phoffset = 0;
1.66 thorpej 736: off = palloc->pa_pagesz;
1.88 chs 737: SPLAY_INIT(&pp->pr_phtree);
1.2 pk 738: }
1.1 pk 739:
1.3 pk 740: /*
741: * Alignment is to take place at `ioff' within the item. This means
742: * we must reserve up to `align - 1' bytes on the page to allow
743: * appropriate positioning of each item.
744: */
745: pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
1.43 thorpej 746: KASSERT(pp->pr_itemsperpage != 0);
1.97 yamt 747: if ((pp->pr_roflags & PR_NOTOUCH)) {
748: int idx;
749:
750: for (idx = 0; pp->pr_itemsperpage > PHPOOL_FREELIST_NELEM(idx);
751: idx++) {
752: /* nothing */
753: }
754: if (idx >= PHPOOL_MAX) {
755: /*
756: * if you see this panic, consider to tweak
757: * PHPOOL_MAX and PHPOOL_FREELIST_NELEM.
758: */
759: panic("%s: too large itemsperpage(%d) for PR_NOTOUCH",
760: pp->pr_wchan, pp->pr_itemsperpage);
761: }
762: pp->pr_phpool = &phpool[idx];
763: } else if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
764: pp->pr_phpool = &phpool[0];
765: }
766: #if defined(DIAGNOSTIC)
767: else {
768: pp->pr_phpool = NULL;
769: }
770: #endif
1.3 pk 771:
772: /*
773: * Use the slack between the chunks and the page header
774: * for "cache coloring".
775: */
776: slack = off - pp->pr_itemsperpage * pp->pr_size;
777: pp->pr_maxcolor = (slack / align) * align;
778: pp->pr_curcolor = 0;
779:
780: pp->pr_nget = 0;
781: pp->pr_nfail = 0;
782: pp->pr_nput = 0;
783: pp->pr_npagealloc = 0;
784: pp->pr_npagefree = 0;
1.1 pk 785: pp->pr_hiwat = 0;
1.8 thorpej 786: pp->pr_nidle = 0;
1.134 ad 787: pp->pr_refcnt = 0;
1.3 pk 788:
1.59 thorpej 789: #ifdef POOL_DIAGNOSTIC
1.25 thorpej 790: if (flags & PR_LOGGING) {
791: if (kmem_map == NULL ||
792: (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log),
793: M_TEMP, M_NOWAIT)) == NULL)
1.20 thorpej 794: pp->pr_roflags &= ~PR_LOGGING;
1.3 pk 795: pp->pr_curlogentry = 0;
796: pp->pr_logsize = pool_logsize;
797: }
1.59 thorpej 798: #endif
1.25 thorpej 799:
800: pp->pr_entered_file = NULL;
801: pp->pr_entered_line = 0;
1.3 pk 802:
1.157 ad 803: mutex_init(&pp->pr_lock, MUTEX_DEFAULT, ipl);
1.134 ad 804: cv_init(&pp->pr_cv, wchan);
805: pp->pr_ipl = ipl;
1.1 pk 806:
1.3 pk 807: /*
1.43 thorpej 808: * Initialize private page header pool and cache magazine pool if we
809: * haven't done so yet.
1.23 thorpej 810: * XXX LOCKING.
1.3 pk 811: */
1.97 yamt 812: if (phpool[0].pr_size == 0) {
813: int idx;
814: for (idx = 0; idx < PHPOOL_MAX; idx++) {
815: static char phpool_names[PHPOOL_MAX][6+1+6+1];
816: int nelem;
817: size_t sz;
818:
819: nelem = PHPOOL_FREELIST_NELEM(idx);
820: snprintf(phpool_names[idx], sizeof(phpool_names[idx]),
821: "phpool-%d", nelem);
822: sz = sizeof(struct pool_item_header);
823: if (nelem) {
1.135 yamt 824: sz = offsetof(struct pool_item_header,
825: ph_bitmap[howmany(nelem, BITMAP_SIZE)]);
1.97 yamt 826: }
827: pool_init(&phpool[idx], sz, 0, 0, 0,
1.129 ad 828: phpool_names[idx], &pool_allocator_meta, IPL_VM);
1.97 yamt 829: }
1.62 bjh21 830: #ifdef POOL_SUBPAGE
831: pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
1.129 ad 832: PR_RECURSIVE, "psppool", &pool_allocator_meta, IPL_VM);
1.62 bjh21 833: #endif
1.142 ad 834:
835: size = sizeof(pcg_t) +
836: (PCG_NOBJECTS_NORMAL - 1) * sizeof(pcgpair_t);
1.156 ad 837: pool_init(&pcg_normal_pool, size, coherency_unit, 0, 0,
1.142 ad 838: "pcgnormal", &pool_allocator_meta, IPL_VM);
839:
840: size = sizeof(pcg_t) +
841: (PCG_NOBJECTS_LARGE - 1) * sizeof(pcgpair_t);
1.156 ad 842: pool_init(&pcg_large_pool, size, coherency_unit, 0, 0,
1.142 ad 843: "pcglarge", &pool_allocator_meta, IPL_VM);
1.1 pk 844: }
845:
1.145 ad 846: /* Insert into the list of all pools. */
847: if (__predict_true(!cold))
1.134 ad 848: mutex_enter(&pool_head_lock);
1.145 ad 849: TAILQ_FOREACH(pp1, &pool_head, pr_poollist) {
850: if (strcmp(pp1->pr_wchan, pp->pr_wchan) > 0)
851: break;
852: }
853: if (pp1 == NULL)
854: TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist);
855: else
856: TAILQ_INSERT_BEFORE(pp1, pp, pr_poollist);
857: if (__predict_true(!cold))
1.134 ad 858: mutex_exit(&pool_head_lock);
859:
1.158.2.2 yamt 860: /* Insert this into the list of pools using this allocator. */
1.145 ad 861: if (__predict_true(!cold))
1.134 ad 862: mutex_enter(&palloc->pa_lock);
1.145 ad 863: TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
864: if (__predict_true(!cold))
1.134 ad 865: mutex_exit(&palloc->pa_lock);
1.66 thorpej 866:
1.117 yamt 867: pool_reclaim_register(pp);
1.1 pk 868: }
869:
870: /*
871: * De-commision a pool resource.
872: */
873: void
1.42 thorpej 874: pool_destroy(struct pool *pp)
1.1 pk 875: {
1.101 thorpej 876: struct pool_pagelist pq;
1.3 pk 877: struct pool_item_header *ph;
1.43 thorpej 878:
1.101 thorpej 879: /* Remove from global pool list */
1.134 ad 880: mutex_enter(&pool_head_lock);
881: while (pp->pr_refcnt != 0)
882: cv_wait(&pool_busy, &pool_head_lock);
1.145 ad 883: TAILQ_REMOVE(&pool_head, pp, pr_poollist);
1.101 thorpej 884: if (drainpp == pp)
885: drainpp = NULL;
1.134 ad 886: mutex_exit(&pool_head_lock);
1.101 thorpej 887:
888: /* Remove this pool from its allocator's list of pools. */
1.117 yamt 889: pool_reclaim_unregister(pp);
1.134 ad 890: mutex_enter(&pp->pr_alloc->pa_lock);
1.66 thorpej 891: TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
1.134 ad 892: mutex_exit(&pp->pr_alloc->pa_lock);
1.66 thorpej 893:
1.134 ad 894: mutex_enter(&pp->pr_lock);
1.101 thorpej 895:
1.134 ad 896: KASSERT(pp->pr_cache == NULL);
1.3 pk 897:
898: #ifdef DIAGNOSTIC
1.20 thorpej 899: if (pp->pr_nout != 0) {
1.25 thorpej 900: pr_printlog(pp, NULL, printf);
1.80 provos 901: panic("pool_destroy: pool busy: still out: %u",
1.20 thorpej 902: pp->pr_nout);
1.3 pk 903: }
904: #endif
1.1 pk 905:
1.101 thorpej 906: KASSERT(LIST_EMPTY(&pp->pr_fullpages));
907: KASSERT(LIST_EMPTY(&pp->pr_partpages));
908:
1.3 pk 909: /* Remove all pages */
1.101 thorpej 910: LIST_INIT(&pq);
1.88 chs 911: while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1.101 thorpej 912: pr_rmpage(pp, ph, &pq);
913:
1.134 ad 914: mutex_exit(&pp->pr_lock);
1.3 pk 915:
1.101 thorpej 916: pr_pagelist_free(pp, &pq);
1.3 pk 917:
1.59 thorpej 918: #ifdef POOL_DIAGNOSTIC
1.20 thorpej 919: if ((pp->pr_roflags & PR_LOGGING) != 0)
1.3 pk 920: free(pp->pr_log, M_TEMP);
1.59 thorpej 921: #endif
1.134 ad 922:
923: cv_destroy(&pp->pr_cv);
924: mutex_destroy(&pp->pr_lock);
1.1 pk 925: }
926:
1.68 thorpej 927: void
928: pool_set_drain_hook(struct pool *pp, void (*fn)(void *, int), void *arg)
929: {
930:
931: /* XXX no locking -- must be used just after pool_init() */
932: #ifdef DIAGNOSTIC
933: if (pp->pr_drain_hook != NULL)
934: panic("pool_set_drain_hook(%s): already set", pp->pr_wchan);
935: #endif
936: pp->pr_drain_hook = fn;
937: pp->pr_drain_hook_arg = arg;
938: }
939:
1.88 chs 940: static struct pool_item_header *
1.128 christos 941: pool_alloc_item_header(struct pool *pp, void *storage, int flags)
1.55 thorpej 942: {
943: struct pool_item_header *ph;
944:
945: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
1.128 christos 946: ph = (struct pool_item_header *) ((char *)storage + pp->pr_phoffset);
1.134 ad 947: else
1.97 yamt 948: ph = pool_get(pp->pr_phpool, flags);
1.55 thorpej 949:
950: return (ph);
951: }
1.1 pk 952:
953: /*
1.134 ad 954: * Grab an item from the pool.
1.1 pk 955: */
1.3 pk 956: void *
1.59 thorpej 957: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 958: _pool_get(struct pool *pp, int flags, const char *file, long line)
1.56 sommerfe 959: #else
960: pool_get(struct pool *pp, int flags)
961: #endif
1.1 pk 962: {
963: struct pool_item *pi;
1.3 pk 964: struct pool_item_header *ph;
1.55 thorpej 965: void *v;
1.1 pk 966:
1.2 pk 967: #ifdef DIAGNOSTIC
1.95 atatat 968: if (__predict_false(pp->pr_itemsperpage == 0))
969: panic("pool_get: pool %p: pr_itemsperpage is zero, "
970: "pool not initialized?", pp);
1.84 thorpej 971: if (__predict_false(curlwp == NULL && doing_shutdown == 0 &&
1.37 sommerfe 972: (flags & PR_WAITOK) != 0))
1.77 matt 973: panic("pool_get: %s: must have NOWAIT", pp->pr_wchan);
1.58 thorpej 974:
1.102 chs 975: #endif /* DIAGNOSTIC */
1.58 thorpej 976: #ifdef LOCKDEBUG
1.155 ad 977: if (flags & PR_WAITOK) {
1.154 yamt 978: ASSERT_SLEEPABLE();
1.155 ad 979: }
1.56 sommerfe 980: #endif
1.1 pk 981:
1.134 ad 982: mutex_enter(&pp->pr_lock);
1.25 thorpej 983: pr_enter(pp, file, line);
1.20 thorpej 984:
985: startover:
986: /*
987: * Check to see if we've reached the hard limit. If we have,
988: * and we can wait, then wait until an item has been returned to
989: * the pool.
990: */
991: #ifdef DIAGNOSTIC
1.34 thorpej 992: if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
1.25 thorpej 993: pr_leave(pp);
1.134 ad 994: mutex_exit(&pp->pr_lock);
1.20 thorpej 995: panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
996: }
997: #endif
1.34 thorpej 998: if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
1.68 thorpej 999: if (pp->pr_drain_hook != NULL) {
1000: /*
1001: * Since the drain hook is going to free things
1002: * back to the pool, unlock, call the hook, re-lock,
1003: * and check the hardlimit condition again.
1004: */
1005: pr_leave(pp);
1.134 ad 1006: mutex_exit(&pp->pr_lock);
1.68 thorpej 1007: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
1.134 ad 1008: mutex_enter(&pp->pr_lock);
1.68 thorpej 1009: pr_enter(pp, file, line);
1010: if (pp->pr_nout < pp->pr_hardlimit)
1011: goto startover;
1012: }
1013:
1.29 sommerfe 1014: if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
1.20 thorpej 1015: /*
1016: * XXX: A warning isn't logged in this case. Should
1017: * it be?
1018: */
1019: pp->pr_flags |= PR_WANTED;
1.25 thorpej 1020: pr_leave(pp);
1.134 ad 1021: cv_wait(&pp->pr_cv, &pp->pr_lock);
1.25 thorpej 1022: pr_enter(pp, file, line);
1.20 thorpej 1023: goto startover;
1024: }
1.31 thorpej 1025:
1026: /*
1027: * Log a message that the hard limit has been hit.
1028: */
1029: if (pp->pr_hardlimit_warning != NULL &&
1030: ratecheck(&pp->pr_hardlimit_warning_last,
1031: &pp->pr_hardlimit_ratecap))
1032: log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
1.21 thorpej 1033:
1034: pp->pr_nfail++;
1035:
1.25 thorpej 1036: pr_leave(pp);
1.134 ad 1037: mutex_exit(&pp->pr_lock);
1.20 thorpej 1038: return (NULL);
1039: }
1040:
1.3 pk 1041: /*
1042: * The convention we use is that if `curpage' is not NULL, then
1043: * it points at a non-empty bucket. In particular, `curpage'
1044: * never points at a page header which has PR_PHINPAGE set and
1045: * has no items in its bucket.
1046: */
1.20 thorpej 1047: if ((ph = pp->pr_curpage) == NULL) {
1.113 yamt 1048: int error;
1049:
1.20 thorpej 1050: #ifdef DIAGNOSTIC
1051: if (pp->pr_nitems != 0) {
1.134 ad 1052: mutex_exit(&pp->pr_lock);
1.20 thorpej 1053: printf("pool_get: %s: curpage NULL, nitems %u\n",
1054: pp->pr_wchan, pp->pr_nitems);
1.80 provos 1055: panic("pool_get: nitems inconsistent");
1.20 thorpej 1056: }
1057: #endif
1058:
1.21 thorpej 1059: /*
1060: * Call the back-end page allocator for more memory.
1061: * Release the pool lock, as the back-end page allocator
1062: * may block.
1063: */
1.25 thorpej 1064: pr_leave(pp);
1.113 yamt 1065: error = pool_grow(pp, flags);
1066: pr_enter(pp, file, line);
1067: if (error != 0) {
1.21 thorpej 1068: /*
1.55 thorpej 1069: * We were unable to allocate a page or item
1070: * header, but we released the lock during
1071: * allocation, so perhaps items were freed
1072: * back to the pool. Check for this case.
1.21 thorpej 1073: */
1074: if (pp->pr_curpage != NULL)
1075: goto startover;
1.15 pk 1076:
1.117 yamt 1077: pp->pr_nfail++;
1.25 thorpej 1078: pr_leave(pp);
1.134 ad 1079: mutex_exit(&pp->pr_lock);
1.117 yamt 1080: return (NULL);
1.1 pk 1081: }
1.3 pk 1082:
1.20 thorpej 1083: /* Start the allocation process over. */
1084: goto startover;
1.3 pk 1085: }
1.97 yamt 1086: if (pp->pr_roflags & PR_NOTOUCH) {
1087: #ifdef DIAGNOSTIC
1088: if (__predict_false(ph->ph_nmissing == pp->pr_itemsperpage)) {
1089: pr_leave(pp);
1.134 ad 1090: mutex_exit(&pp->pr_lock);
1.97 yamt 1091: panic("pool_get: %s: page empty", pp->pr_wchan);
1092: }
1093: #endif
1094: v = pr_item_notouch_get(pp, ph);
1095: #ifdef POOL_DIAGNOSTIC
1096: pr_log(pp, v, PRLOG_GET, file, line);
1097: #endif
1098: } else {
1.102 chs 1099: v = pi = LIST_FIRST(&ph->ph_itemlist);
1.97 yamt 1100: if (__predict_false(v == NULL)) {
1101: pr_leave(pp);
1.134 ad 1102: mutex_exit(&pp->pr_lock);
1.97 yamt 1103: panic("pool_get: %s: page empty", pp->pr_wchan);
1104: }
1.20 thorpej 1105: #ifdef DIAGNOSTIC
1.97 yamt 1106: if (__predict_false(pp->pr_nitems == 0)) {
1107: pr_leave(pp);
1.134 ad 1108: mutex_exit(&pp->pr_lock);
1.97 yamt 1109: printf("pool_get: %s: items on itemlist, nitems %u\n",
1110: pp->pr_wchan, pp->pr_nitems);
1111: panic("pool_get: nitems inconsistent");
1112: }
1.65 enami 1113: #endif
1.56 sommerfe 1114:
1.65 enami 1115: #ifdef POOL_DIAGNOSTIC
1.97 yamt 1116: pr_log(pp, v, PRLOG_GET, file, line);
1.65 enami 1117: #endif
1.3 pk 1118:
1.65 enami 1119: #ifdef DIAGNOSTIC
1.97 yamt 1120: if (__predict_false(pi->pi_magic != PI_MAGIC)) {
1121: pr_printlog(pp, pi, printf);
1122: panic("pool_get(%s): free list modified: "
1123: "magic=%x; page %p; item addr %p\n",
1124: pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
1125: }
1.3 pk 1126: #endif
1127:
1.97 yamt 1128: /*
1129: * Remove from item list.
1130: */
1.102 chs 1131: LIST_REMOVE(pi, pi_list);
1.97 yamt 1132: }
1.20 thorpej 1133: pp->pr_nitems--;
1134: pp->pr_nout++;
1.6 thorpej 1135: if (ph->ph_nmissing == 0) {
1136: #ifdef DIAGNOSTIC
1.34 thorpej 1137: if (__predict_false(pp->pr_nidle == 0))
1.6 thorpej 1138: panic("pool_get: nidle inconsistent");
1139: #endif
1140: pp->pr_nidle--;
1.88 chs 1141:
1142: /*
1143: * This page was previously empty. Move it to the list of
1144: * partially-full pages. This page is already curpage.
1145: */
1146: LIST_REMOVE(ph, ph_pagelist);
1147: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.6 thorpej 1148: }
1.3 pk 1149: ph->ph_nmissing++;
1.97 yamt 1150: if (ph->ph_nmissing == pp->pr_itemsperpage) {
1.21 thorpej 1151: #ifdef DIAGNOSTIC
1.97 yamt 1152: if (__predict_false((pp->pr_roflags & PR_NOTOUCH) == 0 &&
1.102 chs 1153: !LIST_EMPTY(&ph->ph_itemlist))) {
1.25 thorpej 1154: pr_leave(pp);
1.134 ad 1155: mutex_exit(&pp->pr_lock);
1.21 thorpej 1156: panic("pool_get: %s: nmissing inconsistent",
1157: pp->pr_wchan);
1158: }
1159: #endif
1.3 pk 1160: /*
1.88 chs 1161: * This page is now full. Move it to the full list
1162: * and select a new current page.
1.3 pk 1163: */
1.88 chs 1164: LIST_REMOVE(ph, ph_pagelist);
1165: LIST_INSERT_HEAD(&pp->pr_fullpages, ph, ph_pagelist);
1166: pool_update_curpage(pp);
1.1 pk 1167: }
1.3 pk 1168:
1169: pp->pr_nget++;
1.111 christos 1170: pr_leave(pp);
1.20 thorpej 1171:
1172: /*
1173: * If we have a low water mark and we are now below that low
1174: * water mark, add more items to the pool.
1175: */
1.53 thorpej 1176: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1177: /*
1178: * XXX: Should we log a warning? Should we set up a timeout
1179: * to try again in a second or so? The latter could break
1180: * a caller's assumptions about interrupt protection, etc.
1181: */
1182: }
1183:
1.134 ad 1184: mutex_exit(&pp->pr_lock);
1.125 ad 1185: KASSERT((((vaddr_t)v + pp->pr_itemoffset) & (pp->pr_align - 1)) == 0);
1186: FREECHECK_OUT(&pp->pr_freecheck, v);
1.1 pk 1187: return (v);
1188: }
1189:
1190: /*
1.43 thorpej 1191: * Internal version of pool_put(). Pool is already locked/entered.
1.1 pk 1192: */
1.43 thorpej 1193: static void
1.101 thorpej 1194: pool_do_put(struct pool *pp, void *v, struct pool_pagelist *pq)
1.1 pk 1195: {
1196: struct pool_item *pi = v;
1.3 pk 1197: struct pool_item_header *ph;
1198:
1.134 ad 1199: KASSERT(mutex_owned(&pp->pr_lock));
1.125 ad 1200: FREECHECK_IN(&pp->pr_freecheck, v);
1.134 ad 1201: LOCKDEBUG_MEM_CHECK(v, pp->pr_size);
1.61 chs 1202:
1.30 thorpej 1203: #ifdef DIAGNOSTIC
1.34 thorpej 1204: if (__predict_false(pp->pr_nout == 0)) {
1.30 thorpej 1205: printf("pool %s: putting with none out\n",
1206: pp->pr_wchan);
1207: panic("pool_put");
1208: }
1209: #endif
1.3 pk 1210:
1.121 yamt 1211: if (__predict_false((ph = pr_find_pagehead(pp, v)) == NULL)) {
1.25 thorpej 1212: pr_printlog(pp, NULL, printf);
1.3 pk 1213: panic("pool_put: %s: page header missing", pp->pr_wchan);
1214: }
1.28 thorpej 1215:
1.3 pk 1216: /*
1217: * Return to item list.
1218: */
1.97 yamt 1219: if (pp->pr_roflags & PR_NOTOUCH) {
1220: pr_item_notouch_put(pp, ph, v);
1221: } else {
1.2 pk 1222: #ifdef DIAGNOSTIC
1.97 yamt 1223: pi->pi_magic = PI_MAGIC;
1.3 pk 1224: #endif
1.32 chs 1225: #ifdef DEBUG
1.97 yamt 1226: {
1227: int i, *ip = v;
1.32 chs 1228:
1.97 yamt 1229: for (i = 0; i < pp->pr_size / sizeof(int); i++) {
1230: *ip++ = PI_MAGIC;
1231: }
1.32 chs 1232: }
1233: #endif
1234:
1.102 chs 1235: LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.97 yamt 1236: }
1.79 thorpej 1237: KDASSERT(ph->ph_nmissing != 0);
1.3 pk 1238: ph->ph_nmissing--;
1239: pp->pr_nput++;
1.20 thorpej 1240: pp->pr_nitems++;
1241: pp->pr_nout--;
1.3 pk 1242:
1243: /* Cancel "pool empty" condition if it exists */
1244: if (pp->pr_curpage == NULL)
1245: pp->pr_curpage = ph;
1246:
1247: if (pp->pr_flags & PR_WANTED) {
1248: pp->pr_flags &= ~PR_WANTED;
1.134 ad 1249: cv_broadcast(&pp->pr_cv);
1.3 pk 1250: }
1251:
1252: /*
1.88 chs 1253: * If this page is now empty, do one of two things:
1.21 thorpej 1254: *
1.88 chs 1255: * (1) If we have more pages than the page high water mark,
1.96 thorpej 1256: * free the page back to the system. ONLY CONSIDER
1.90 thorpej 1257: * FREEING BACK A PAGE IF WE HAVE MORE THAN OUR MINIMUM PAGE
1258: * CLAIM.
1.21 thorpej 1259: *
1.88 chs 1260: * (2) Otherwise, move the page to the empty page list.
1261: *
1262: * Either way, select a new current page (so we use a partially-full
1263: * page if one is available).
1.3 pk 1264: */
1265: if (ph->ph_nmissing == 0) {
1.6 thorpej 1266: pp->pr_nidle++;
1.90 thorpej 1267: if (pp->pr_npages > pp->pr_minpages &&
1.152 yamt 1268: pp->pr_npages > pp->pr_maxpages) {
1.101 thorpej 1269: pr_rmpage(pp, ph, pq);
1.3 pk 1270: } else {
1.88 chs 1271: LIST_REMOVE(ph, ph_pagelist);
1272: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.3 pk 1273:
1.21 thorpej 1274: /*
1275: * Update the timestamp on the page. A page must
1276: * be idle for some period of time before it can
1277: * be reclaimed by the pagedaemon. This minimizes
1278: * ping-pong'ing for memory.
1.151 yamt 1279: *
1280: * note for 64-bit time_t: truncating to 32-bit is not
1281: * a problem for our usage.
1.21 thorpej 1282: */
1.151 yamt 1283: ph->ph_time = time_uptime;
1.1 pk 1284: }
1.88 chs 1285: pool_update_curpage(pp);
1.1 pk 1286: }
1.88 chs 1287:
1.21 thorpej 1288: /*
1.88 chs 1289: * If the page was previously completely full, move it to the
1290: * partially-full list and make it the current page. The next
1291: * allocation will get the item from this page, instead of
1292: * further fragmenting the pool.
1.21 thorpej 1293: */
1294: else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
1.88 chs 1295: LIST_REMOVE(ph, ph_pagelist);
1296: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.21 thorpej 1297: pp->pr_curpage = ph;
1298: }
1.43 thorpej 1299: }
1300:
1301: /*
1.134 ad 1302: * Return resource to the pool.
1.43 thorpej 1303: */
1.59 thorpej 1304: #ifdef POOL_DIAGNOSTIC
1.43 thorpej 1305: void
1306: _pool_put(struct pool *pp, void *v, const char *file, long line)
1307: {
1.101 thorpej 1308: struct pool_pagelist pq;
1309:
1310: LIST_INIT(&pq);
1.43 thorpej 1311:
1.134 ad 1312: mutex_enter(&pp->pr_lock);
1.43 thorpej 1313: pr_enter(pp, file, line);
1314:
1.56 sommerfe 1315: pr_log(pp, v, PRLOG_PUT, file, line);
1316:
1.101 thorpej 1317: pool_do_put(pp, v, &pq);
1.21 thorpej 1318:
1.25 thorpej 1319: pr_leave(pp);
1.134 ad 1320: mutex_exit(&pp->pr_lock);
1.101 thorpej 1321:
1.102 chs 1322: pr_pagelist_free(pp, &pq);
1.1 pk 1323: }
1.57 sommerfe 1324: #undef pool_put
1.59 thorpej 1325: #endif /* POOL_DIAGNOSTIC */
1.1 pk 1326:
1.56 sommerfe 1327: void
1328: pool_put(struct pool *pp, void *v)
1329: {
1.101 thorpej 1330: struct pool_pagelist pq;
1331:
1332: LIST_INIT(&pq);
1.56 sommerfe 1333:
1.134 ad 1334: mutex_enter(&pp->pr_lock);
1.101 thorpej 1335: pool_do_put(pp, v, &pq);
1.134 ad 1336: mutex_exit(&pp->pr_lock);
1.56 sommerfe 1337:
1.102 chs 1338: pr_pagelist_free(pp, &pq);
1.56 sommerfe 1339: }
1.57 sommerfe 1340:
1.59 thorpej 1341: #ifdef POOL_DIAGNOSTIC
1.57 sommerfe 1342: #define pool_put(h, v) _pool_put((h), (v), __FILE__, __LINE__)
1.56 sommerfe 1343: #endif
1.74 thorpej 1344:
1345: /*
1.113 yamt 1346: * pool_grow: grow a pool by a page.
1347: *
1348: * => called with pool locked.
1349: * => unlock and relock the pool.
1350: * => return with pool locked.
1351: */
1352:
1353: static int
1354: pool_grow(struct pool *pp, int flags)
1355: {
1356: struct pool_item_header *ph = NULL;
1357: char *cp;
1358:
1.134 ad 1359: mutex_exit(&pp->pr_lock);
1.113 yamt 1360: cp = pool_allocator_alloc(pp, flags);
1361: if (__predict_true(cp != NULL)) {
1362: ph = pool_alloc_item_header(pp, cp, flags);
1363: }
1364: if (__predict_false(cp == NULL || ph == NULL)) {
1365: if (cp != NULL) {
1366: pool_allocator_free(pp, cp);
1367: }
1.134 ad 1368: mutex_enter(&pp->pr_lock);
1.113 yamt 1369: return ENOMEM;
1370: }
1371:
1.134 ad 1372: mutex_enter(&pp->pr_lock);
1.113 yamt 1373: pool_prime_page(pp, cp, ph);
1374: pp->pr_npagealloc++;
1375: return 0;
1376: }
1377:
1378: /*
1.74 thorpej 1379: * Add N items to the pool.
1380: */
1381: int
1382: pool_prime(struct pool *pp, int n)
1383: {
1.75 simonb 1384: int newpages;
1.113 yamt 1385: int error = 0;
1.74 thorpej 1386:
1.134 ad 1387: mutex_enter(&pp->pr_lock);
1.74 thorpej 1388:
1389: newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1390:
1391: while (newpages-- > 0) {
1.113 yamt 1392: error = pool_grow(pp, PR_NOWAIT);
1393: if (error) {
1.74 thorpej 1394: break;
1395: }
1396: pp->pr_minpages++;
1397: }
1398:
1399: if (pp->pr_minpages >= pp->pr_maxpages)
1400: pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */
1401:
1.134 ad 1402: mutex_exit(&pp->pr_lock);
1.113 yamt 1403: return error;
1.74 thorpej 1404: }
1.55 thorpej 1405:
1406: /*
1.3 pk 1407: * Add a page worth of items to the pool.
1.21 thorpej 1408: *
1409: * Note, we must be called with the pool descriptor LOCKED.
1.3 pk 1410: */
1.55 thorpej 1411: static void
1.128 christos 1412: pool_prime_page(struct pool *pp, void *storage, struct pool_item_header *ph)
1.3 pk 1413: {
1414: struct pool_item *pi;
1.128 christos 1415: void *cp = storage;
1.125 ad 1416: const unsigned int align = pp->pr_align;
1417: const unsigned int ioff = pp->pr_itemoffset;
1.55 thorpej 1418: int n;
1.36 pk 1419:
1.134 ad 1420: KASSERT(mutex_owned(&pp->pr_lock));
1.91 yamt 1421:
1.66 thorpej 1422: #ifdef DIAGNOSTIC
1.121 yamt 1423: if ((pp->pr_roflags & PR_NOALIGN) == 0 &&
1.150 skrll 1424: ((uintptr_t)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
1.36 pk 1425: panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1.66 thorpej 1426: #endif
1.3 pk 1427:
1428: /*
1429: * Insert page header.
1430: */
1.88 chs 1431: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.102 chs 1432: LIST_INIT(&ph->ph_itemlist);
1.3 pk 1433: ph->ph_page = storage;
1434: ph->ph_nmissing = 0;
1.151 yamt 1435: ph->ph_time = time_uptime;
1.88 chs 1436: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1437: SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
1.3 pk 1438:
1.6 thorpej 1439: pp->pr_nidle++;
1440:
1.3 pk 1441: /*
1442: * Color this page.
1443: */
1.141 yamt 1444: ph->ph_off = pp->pr_curcolor;
1445: cp = (char *)cp + ph->ph_off;
1.3 pk 1446: if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
1447: pp->pr_curcolor = 0;
1448:
1449: /*
1450: * Adjust storage to apply aligment to `pr_itemoffset' in each item.
1451: */
1452: if (ioff != 0)
1.128 christos 1453: cp = (char *)cp + align - ioff;
1.3 pk 1454:
1.125 ad 1455: KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
1456:
1.3 pk 1457: /*
1458: * Insert remaining chunks on the bucket list.
1459: */
1460: n = pp->pr_itemsperpage;
1.20 thorpej 1461: pp->pr_nitems += n;
1.3 pk 1462:
1.97 yamt 1463: if (pp->pr_roflags & PR_NOTOUCH) {
1.141 yamt 1464: pr_item_notouch_init(pp, ph);
1.97 yamt 1465: } else {
1466: while (n--) {
1467: pi = (struct pool_item *)cp;
1.78 thorpej 1468:
1.97 yamt 1469: KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
1.3 pk 1470:
1.97 yamt 1471: /* Insert on page list */
1.102 chs 1472: LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.3 pk 1473: #ifdef DIAGNOSTIC
1.97 yamt 1474: pi->pi_magic = PI_MAGIC;
1.3 pk 1475: #endif
1.128 christos 1476: cp = (char *)cp + pp->pr_size;
1.125 ad 1477:
1478: KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
1.97 yamt 1479: }
1.3 pk 1480: }
1481:
1482: /*
1483: * If the pool was depleted, point at the new page.
1484: */
1485: if (pp->pr_curpage == NULL)
1486: pp->pr_curpage = ph;
1487:
1488: if (++pp->pr_npages > pp->pr_hiwat)
1489: pp->pr_hiwat = pp->pr_npages;
1490: }
1491:
1.20 thorpej 1492: /*
1.52 thorpej 1493: * Used by pool_get() when nitems drops below the low water mark. This
1.88 chs 1494: * is used to catch up pr_nitems with the low water mark.
1.20 thorpej 1495: *
1.21 thorpej 1496: * Note 1, we never wait for memory here, we let the caller decide what to do.
1.20 thorpej 1497: *
1.73 thorpej 1498: * Note 2, we must be called with the pool already locked, and we return
1.20 thorpej 1499: * with it locked.
1500: */
1501: static int
1.42 thorpej 1502: pool_catchup(struct pool *pp)
1.20 thorpej 1503: {
1504: int error = 0;
1505:
1.54 thorpej 1506: while (POOL_NEEDS_CATCHUP(pp)) {
1.113 yamt 1507: error = pool_grow(pp, PR_NOWAIT);
1508: if (error) {
1.20 thorpej 1509: break;
1510: }
1511: }
1.113 yamt 1512: return error;
1.20 thorpej 1513: }
1514:
1.88 chs 1515: static void
1516: pool_update_curpage(struct pool *pp)
1517: {
1518:
1519: pp->pr_curpage = LIST_FIRST(&pp->pr_partpages);
1520: if (pp->pr_curpage == NULL) {
1521: pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages);
1522: }
1.158.2.2 yamt 1523: KASSERT((pp->pr_curpage == NULL && pp->pr_nitems == 0) ||
1524: (pp->pr_curpage != NULL && pp->pr_nitems > 0));
1.88 chs 1525: }
1526:
1.3 pk 1527: void
1.42 thorpej 1528: pool_setlowat(struct pool *pp, int n)
1.3 pk 1529: {
1.15 pk 1530:
1.134 ad 1531: mutex_enter(&pp->pr_lock);
1.21 thorpej 1532:
1.3 pk 1533: pp->pr_minitems = n;
1.15 pk 1534: pp->pr_minpages = (n == 0)
1535: ? 0
1.18 thorpej 1536: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.20 thorpej 1537:
1538: /* Make sure we're caught up with the newly-set low water mark. */
1.75 simonb 1539: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1540: /*
1541: * XXX: Should we log a warning? Should we set up a timeout
1542: * to try again in a second or so? The latter could break
1543: * a caller's assumptions about interrupt protection, etc.
1544: */
1545: }
1.21 thorpej 1546:
1.134 ad 1547: mutex_exit(&pp->pr_lock);
1.3 pk 1548: }
1549:
1550: void
1.42 thorpej 1551: pool_sethiwat(struct pool *pp, int n)
1.3 pk 1552: {
1.15 pk 1553:
1.134 ad 1554: mutex_enter(&pp->pr_lock);
1.21 thorpej 1555:
1.15 pk 1556: pp->pr_maxpages = (n == 0)
1557: ? 0
1.18 thorpej 1558: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1559:
1.134 ad 1560: mutex_exit(&pp->pr_lock);
1.3 pk 1561: }
1562:
1.20 thorpej 1563: void
1.42 thorpej 1564: pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1.20 thorpej 1565: {
1566:
1.134 ad 1567: mutex_enter(&pp->pr_lock);
1.20 thorpej 1568:
1569: pp->pr_hardlimit = n;
1570: pp->pr_hardlimit_warning = warnmess;
1.31 thorpej 1571: pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1572: pp->pr_hardlimit_warning_last.tv_sec = 0;
1573: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.20 thorpej 1574:
1575: /*
1.21 thorpej 1576: * In-line version of pool_sethiwat(), because we don't want to
1577: * release the lock.
1.20 thorpej 1578: */
1579: pp->pr_maxpages = (n == 0)
1580: ? 0
1581: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1582:
1.134 ad 1583: mutex_exit(&pp->pr_lock);
1.20 thorpej 1584: }
1.3 pk 1585:
1586: /*
1587: * Release all complete pages that have not been used recently.
1588: */
1.66 thorpej 1589: int
1.59 thorpej 1590: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 1591: _pool_reclaim(struct pool *pp, const char *file, long line)
1.56 sommerfe 1592: #else
1593: pool_reclaim(struct pool *pp)
1594: #endif
1.3 pk 1595: {
1596: struct pool_item_header *ph, *phnext;
1.61 chs 1597: struct pool_pagelist pq;
1.151 yamt 1598: uint32_t curtime;
1.134 ad 1599: bool klock;
1600: int rv;
1.3 pk 1601:
1.68 thorpej 1602: if (pp->pr_drain_hook != NULL) {
1603: /*
1604: * The drain hook must be called with the pool unlocked.
1605: */
1606: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT);
1607: }
1608:
1.134 ad 1609: /*
1.157 ad 1610: * XXXSMP Because we do not want to cause non-MPSAFE code
1611: * to block.
1.134 ad 1612: */
1613: if (pp->pr_ipl == IPL_SOFTNET || pp->pr_ipl == IPL_SOFTCLOCK ||
1614: pp->pr_ipl == IPL_SOFTSERIAL) {
1615: KERNEL_LOCK(1, NULL);
1616: klock = true;
1617: } else
1618: klock = false;
1619:
1620: /* Reclaim items from the pool's cache (if any). */
1621: if (pp->pr_cache != NULL)
1622: pool_cache_invalidate(pp->pr_cache);
1623:
1624: if (mutex_tryenter(&pp->pr_lock) == 0) {
1625: if (klock) {
1626: KERNEL_UNLOCK_ONE(NULL);
1627: }
1.66 thorpej 1628: return (0);
1.134 ad 1629: }
1.25 thorpej 1630: pr_enter(pp, file, line);
1.68 thorpej 1631:
1.88 chs 1632: LIST_INIT(&pq);
1.43 thorpej 1633:
1.151 yamt 1634: curtime = time_uptime;
1.21 thorpej 1635:
1.88 chs 1636: for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
1637: phnext = LIST_NEXT(ph, ph_pagelist);
1.3 pk 1638:
1639: /* Check our minimum page claim */
1640: if (pp->pr_npages <= pp->pr_minpages)
1641: break;
1642:
1.88 chs 1643: KASSERT(ph->ph_nmissing == 0);
1.151 yamt 1644: if (curtime - ph->ph_time < pool_inactive_time
1.117 yamt 1645: && !pa_starved_p(pp->pr_alloc))
1.88 chs 1646: continue;
1.21 thorpej 1647:
1.88 chs 1648: /*
1649: * If freeing this page would put us below
1650: * the low water mark, stop now.
1651: */
1652: if ((pp->pr_nitems - pp->pr_itemsperpage) <
1653: pp->pr_minitems)
1654: break;
1.21 thorpej 1655:
1.88 chs 1656: pr_rmpage(pp, ph, &pq);
1.3 pk 1657: }
1658:
1.25 thorpej 1659: pr_leave(pp);
1.134 ad 1660: mutex_exit(&pp->pr_lock);
1661:
1662: if (LIST_EMPTY(&pq))
1663: rv = 0;
1664: else {
1665: pr_pagelist_free(pp, &pq);
1666: rv = 1;
1667: }
1668:
1669: if (klock) {
1670: KERNEL_UNLOCK_ONE(NULL);
1671: }
1.66 thorpej 1672:
1.134 ad 1673: return (rv);
1.3 pk 1674: }
1675:
1676: /*
1.134 ad 1677: * Drain pools, one at a time. This is a two stage process;
1678: * drain_start kicks off a cross call to drain CPU-level caches
1679: * if the pool has an associated pool_cache. drain_end waits
1680: * for those cross calls to finish, and then drains the cache
1681: * (if any) and pool.
1.131 ad 1682: *
1.134 ad 1683: * Note, must never be called from interrupt context.
1.3 pk 1684: */
1685: void
1.134 ad 1686: pool_drain_start(struct pool **ppp, uint64_t *wp)
1.3 pk 1687: {
1688: struct pool *pp;
1.134 ad 1689:
1.145 ad 1690: KASSERT(!TAILQ_EMPTY(&pool_head));
1.3 pk 1691:
1.61 chs 1692: pp = NULL;
1.134 ad 1693:
1694: /* Find next pool to drain, and add a reference. */
1695: mutex_enter(&pool_head_lock);
1696: do {
1697: if (drainpp == NULL) {
1.145 ad 1698: drainpp = TAILQ_FIRST(&pool_head);
1.134 ad 1699: }
1700: if (drainpp != NULL) {
1701: pp = drainpp;
1.145 ad 1702: drainpp = TAILQ_NEXT(pp, pr_poollist);
1.134 ad 1703: }
1704: /*
1705: * Skip completely idle pools. We depend on at least
1706: * one pool in the system being active.
1707: */
1708: } while (pp == NULL || pp->pr_npages == 0);
1709: pp->pr_refcnt++;
1710: mutex_exit(&pool_head_lock);
1711:
1712: /* If there is a pool_cache, drain CPU level caches. */
1713: *ppp = pp;
1714: if (pp->pr_cache != NULL) {
1715: *wp = xc_broadcast(0, (xcfunc_t)pool_cache_xcall,
1716: pp->pr_cache, NULL);
1717: }
1718: }
1719:
1720: void
1721: pool_drain_end(struct pool *pp, uint64_t where)
1722: {
1723:
1724: if (pp == NULL)
1725: return;
1726:
1727: KASSERT(pp->pr_refcnt > 0);
1728:
1729: /* Wait for remote draining to complete. */
1730: if (pp->pr_cache != NULL)
1731: xc_wait(where);
1732:
1733: /* Drain the cache (if any) and pool.. */
1734: pool_reclaim(pp);
1735:
1736: /* Finally, unlock the pool. */
1737: mutex_enter(&pool_head_lock);
1738: pp->pr_refcnt--;
1739: cv_broadcast(&pool_busy);
1740: mutex_exit(&pool_head_lock);
1.3 pk 1741: }
1742:
1743: /*
1744: * Diagnostic helpers.
1745: */
1746: void
1.42 thorpej 1747: pool_print(struct pool *pp, const char *modif)
1.21 thorpej 1748: {
1749:
1.25 thorpej 1750: pool_print1(pp, modif, printf);
1.21 thorpej 1751: }
1752:
1.25 thorpej 1753: void
1.108 yamt 1754: pool_printall(const char *modif, void (*pr)(const char *, ...))
1755: {
1756: struct pool *pp;
1757:
1.145 ad 1758: TAILQ_FOREACH(pp, &pool_head, pr_poollist) {
1.108 yamt 1759: pool_printit(pp, modif, pr);
1760: }
1761: }
1762:
1763: void
1.42 thorpej 1764: pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.25 thorpej 1765: {
1766:
1767: if (pp == NULL) {
1768: (*pr)("Must specify a pool to print.\n");
1769: return;
1770: }
1771:
1772: pool_print1(pp, modif, pr);
1773: }
1774:
1.21 thorpej 1775: static void
1.124 yamt 1776: pool_print_pagelist(struct pool *pp, struct pool_pagelist *pl,
1.97 yamt 1777: void (*pr)(const char *, ...))
1.88 chs 1778: {
1779: struct pool_item_header *ph;
1780: #ifdef DIAGNOSTIC
1781: struct pool_item *pi;
1782: #endif
1783:
1784: LIST_FOREACH(ph, pl, ph_pagelist) {
1.151 yamt 1785: (*pr)("\t\tpage %p, nmissing %d, time %" PRIu32 "\n",
1786: ph->ph_page, ph->ph_nmissing, ph->ph_time);
1.88 chs 1787: #ifdef DIAGNOSTIC
1.97 yamt 1788: if (!(pp->pr_roflags & PR_NOTOUCH)) {
1.102 chs 1789: LIST_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1.97 yamt 1790: if (pi->pi_magic != PI_MAGIC) {
1791: (*pr)("\t\t\titem %p, magic 0x%x\n",
1792: pi, pi->pi_magic);
1793: }
1.88 chs 1794: }
1795: }
1796: #endif
1797: }
1798: }
1799:
1800: static void
1.42 thorpej 1801: pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.3 pk 1802: {
1.25 thorpej 1803: struct pool_item_header *ph;
1.134 ad 1804: pool_cache_t pc;
1805: pcg_t *pcg;
1806: pool_cache_cpu_t *cc;
1807: uint64_t cpuhit, cpumiss;
1.44 thorpej 1808: int i, print_log = 0, print_pagelist = 0, print_cache = 0;
1.25 thorpej 1809: char c;
1810:
1811: while ((c = *modif++) != '\0') {
1812: if (c == 'l')
1813: print_log = 1;
1814: if (c == 'p')
1815: print_pagelist = 1;
1.44 thorpej 1816: if (c == 'c')
1817: print_cache = 1;
1.25 thorpej 1818: }
1819:
1.134 ad 1820: if ((pc = pp->pr_cache) != NULL) {
1821: (*pr)("POOL CACHE");
1822: } else {
1823: (*pr)("POOL");
1824: }
1825:
1826: (*pr)(" %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1.25 thorpej 1827: pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1828: pp->pr_roflags);
1.66 thorpej 1829: (*pr)("\talloc %p\n", pp->pr_alloc);
1.25 thorpej 1830: (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1831: pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1832: (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1833: pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1834:
1.134 ad 1835: (*pr)("\tnget %lu, nfail %lu, nput %lu\n",
1.25 thorpej 1836: pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1837: (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1838: pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1839:
1840: if (print_pagelist == 0)
1841: goto skip_pagelist;
1842:
1.88 chs 1843: if ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1844: (*pr)("\n\tempty page list:\n");
1.97 yamt 1845: pool_print_pagelist(pp, &pp->pr_emptypages, pr);
1.88 chs 1846: if ((ph = LIST_FIRST(&pp->pr_fullpages)) != NULL)
1847: (*pr)("\n\tfull page list:\n");
1.97 yamt 1848: pool_print_pagelist(pp, &pp->pr_fullpages, pr);
1.88 chs 1849: if ((ph = LIST_FIRST(&pp->pr_partpages)) != NULL)
1850: (*pr)("\n\tpartial-page list:\n");
1.97 yamt 1851: pool_print_pagelist(pp, &pp->pr_partpages, pr);
1.88 chs 1852:
1.25 thorpej 1853: if (pp->pr_curpage == NULL)
1854: (*pr)("\tno current page\n");
1855: else
1856: (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1857:
1858: skip_pagelist:
1859: if (print_log == 0)
1860: goto skip_log;
1861:
1862: (*pr)("\n");
1863: if ((pp->pr_roflags & PR_LOGGING) == 0)
1864: (*pr)("\tno log\n");
1.122 christos 1865: else {
1.25 thorpej 1866: pr_printlog(pp, NULL, pr);
1.122 christos 1867: }
1.3 pk 1868:
1.25 thorpej 1869: skip_log:
1.44 thorpej 1870:
1.102 chs 1871: #define PR_GROUPLIST(pcg) \
1872: (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail); \
1.142 ad 1873: for (i = 0; i < pcg->pcg_size; i++) { \
1.102 chs 1874: if (pcg->pcg_objects[i].pcgo_pa != \
1875: POOL_PADDR_INVALID) { \
1876: (*pr)("\t\t\t%p, 0x%llx\n", \
1877: pcg->pcg_objects[i].pcgo_va, \
1878: (unsigned long long) \
1879: pcg->pcg_objects[i].pcgo_pa); \
1880: } else { \
1881: (*pr)("\t\t\t%p\n", \
1882: pcg->pcg_objects[i].pcgo_va); \
1883: } \
1884: }
1885:
1.134 ad 1886: if (pc != NULL) {
1887: cpuhit = 0;
1888: cpumiss = 0;
1889: for (i = 0; i < MAXCPUS; i++) {
1890: if ((cc = pc->pc_cpus[i]) == NULL)
1891: continue;
1892: cpuhit += cc->cc_hits;
1893: cpumiss += cc->cc_misses;
1894: }
1895: (*pr)("\tcpu layer hits %llu misses %llu\n", cpuhit, cpumiss);
1896: (*pr)("\tcache layer hits %llu misses %llu\n",
1897: pc->pc_hits, pc->pc_misses);
1898: (*pr)("\tcache layer entry uncontended %llu contended %llu\n",
1899: pc->pc_hits + pc->pc_misses - pc->pc_contended,
1900: pc->pc_contended);
1901: (*pr)("\tcache layer empty groups %u full groups %u\n",
1902: pc->pc_nempty, pc->pc_nfull);
1903: if (print_cache) {
1904: (*pr)("\tfull cache groups:\n");
1905: for (pcg = pc->pc_fullgroups; pcg != NULL;
1906: pcg = pcg->pcg_next) {
1907: PR_GROUPLIST(pcg);
1908: }
1909: (*pr)("\tempty cache groups:\n");
1910: for (pcg = pc->pc_emptygroups; pcg != NULL;
1911: pcg = pcg->pcg_next) {
1912: PR_GROUPLIST(pcg);
1913: }
1.103 chs 1914: }
1.44 thorpej 1915: }
1.102 chs 1916: #undef PR_GROUPLIST
1.44 thorpej 1917:
1.88 chs 1918: pr_enter_check(pp, pr);
1919: }
1920:
1921: static int
1922: pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
1923: {
1924: struct pool_item *pi;
1.128 christos 1925: void *page;
1.88 chs 1926: int n;
1927:
1.121 yamt 1928: if ((pp->pr_roflags & PR_NOALIGN) == 0) {
1.128 christos 1929: page = (void *)((uintptr_t)ph & pp->pr_alloc->pa_pagemask);
1.121 yamt 1930: if (page != ph->ph_page &&
1931: (pp->pr_roflags & PR_PHINPAGE) != 0) {
1932: if (label != NULL)
1933: printf("%s: ", label);
1934: printf("pool(%p:%s): page inconsistency: page %p;"
1935: " at page head addr %p (p %p)\n", pp,
1936: pp->pr_wchan, ph->ph_page,
1937: ph, page);
1938: return 1;
1939: }
1.88 chs 1940: }
1.3 pk 1941:
1.97 yamt 1942: if ((pp->pr_roflags & PR_NOTOUCH) != 0)
1943: return 0;
1944:
1.102 chs 1945: for (pi = LIST_FIRST(&ph->ph_itemlist), n = 0;
1.88 chs 1946: pi != NULL;
1.102 chs 1947: pi = LIST_NEXT(pi,pi_list), n++) {
1.88 chs 1948:
1949: #ifdef DIAGNOSTIC
1950: if (pi->pi_magic != PI_MAGIC) {
1951: if (label != NULL)
1952: printf("%s: ", label);
1953: printf("pool(%s): free list modified: magic=%x;"
1.121 yamt 1954: " page %p; item ordinal %d; addr %p\n",
1.88 chs 1955: pp->pr_wchan, pi->pi_magic, ph->ph_page,
1.121 yamt 1956: n, pi);
1.88 chs 1957: panic("pool");
1958: }
1959: #endif
1.121 yamt 1960: if ((pp->pr_roflags & PR_NOALIGN) != 0) {
1961: continue;
1962: }
1.128 christos 1963: page = (void *)((uintptr_t)pi & pp->pr_alloc->pa_pagemask);
1.88 chs 1964: if (page == ph->ph_page)
1965: continue;
1966:
1967: if (label != NULL)
1968: printf("%s: ", label);
1969: printf("pool(%p:%s): page inconsistency: page %p;"
1970: " item ordinal %d; addr %p (p %p)\n", pp,
1971: pp->pr_wchan, ph->ph_page,
1972: n, pi, page);
1973: return 1;
1974: }
1975: return 0;
1.3 pk 1976: }
1977:
1.88 chs 1978:
1.3 pk 1979: int
1.42 thorpej 1980: pool_chk(struct pool *pp, const char *label)
1.3 pk 1981: {
1982: struct pool_item_header *ph;
1983: int r = 0;
1984:
1.134 ad 1985: mutex_enter(&pp->pr_lock);
1.88 chs 1986: LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
1987: r = pool_chk_page(pp, label, ph);
1988: if (r) {
1989: goto out;
1990: }
1991: }
1992: LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
1993: r = pool_chk_page(pp, label, ph);
1994: if (r) {
1.3 pk 1995: goto out;
1996: }
1.88 chs 1997: }
1998: LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
1999: r = pool_chk_page(pp, label, ph);
2000: if (r) {
1.3 pk 2001: goto out;
2002: }
2003: }
1.88 chs 2004:
1.3 pk 2005: out:
1.134 ad 2006: mutex_exit(&pp->pr_lock);
1.3 pk 2007: return (r);
1.43 thorpej 2008: }
2009:
2010: /*
2011: * pool_cache_init:
2012: *
2013: * Initialize a pool cache.
1.134 ad 2014: */
2015: pool_cache_t
2016: pool_cache_init(size_t size, u_int align, u_int align_offset, u_int flags,
2017: const char *wchan, struct pool_allocator *palloc, int ipl,
2018: int (*ctor)(void *, void *, int), void (*dtor)(void *, void *), void *arg)
2019: {
2020: pool_cache_t pc;
2021:
2022: pc = pool_get(&cache_pool, PR_WAITOK);
2023: if (pc == NULL)
2024: return NULL;
2025:
2026: pool_cache_bootstrap(pc, size, align, align_offset, flags, wchan,
2027: palloc, ipl, ctor, dtor, arg);
2028:
2029: return pc;
2030: }
2031:
2032: /*
2033: * pool_cache_bootstrap:
1.43 thorpej 2034: *
1.134 ad 2035: * Kernel-private version of pool_cache_init(). The caller
2036: * provides initial storage.
1.43 thorpej 2037: */
2038: void
1.134 ad 2039: pool_cache_bootstrap(pool_cache_t pc, size_t size, u_int align,
2040: u_int align_offset, u_int flags, const char *wchan,
2041: struct pool_allocator *palloc, int ipl,
2042: int (*ctor)(void *, void *, int), void (*dtor)(void *, void *),
1.43 thorpej 2043: void *arg)
2044: {
1.134 ad 2045: CPU_INFO_ITERATOR cii;
1.145 ad 2046: pool_cache_t pc1;
1.134 ad 2047: struct cpu_info *ci;
2048: struct pool *pp;
2049:
2050: pp = &pc->pc_pool;
2051: if (palloc == NULL && ipl == IPL_NONE)
2052: palloc = &pool_allocator_nointr;
2053: pool_init(pp, size, align, align_offset, flags, wchan, palloc, ipl);
1.157 ad 2054: mutex_init(&pc->pc_lock, MUTEX_DEFAULT, ipl);
1.43 thorpej 2055:
1.134 ad 2056: if (ctor == NULL) {
2057: ctor = (int (*)(void *, void *, int))nullop;
2058: }
2059: if (dtor == NULL) {
2060: dtor = (void (*)(void *, void *))nullop;
2061: }
1.43 thorpej 2062:
1.134 ad 2063: pc->pc_emptygroups = NULL;
2064: pc->pc_fullgroups = NULL;
2065: pc->pc_partgroups = NULL;
1.43 thorpej 2066: pc->pc_ctor = ctor;
2067: pc->pc_dtor = dtor;
2068: pc->pc_arg = arg;
1.134 ad 2069: pc->pc_hits = 0;
1.48 thorpej 2070: pc->pc_misses = 0;
1.134 ad 2071: pc->pc_nempty = 0;
2072: pc->pc_npart = 0;
2073: pc->pc_nfull = 0;
2074: pc->pc_contended = 0;
2075: pc->pc_refcnt = 0;
1.136 yamt 2076: pc->pc_freecheck = NULL;
1.134 ad 2077:
1.142 ad 2078: if ((flags & PR_LARGECACHE) != 0) {
2079: pc->pc_pcgsize = PCG_NOBJECTS_LARGE;
1.158.2.2 yamt 2080: pc->pc_pcgpool = &pcg_large_pool;
1.142 ad 2081: } else {
2082: pc->pc_pcgsize = PCG_NOBJECTS_NORMAL;
1.158.2.2 yamt 2083: pc->pc_pcgpool = &pcg_normal_pool;
1.142 ad 2084: }
2085:
1.134 ad 2086: /* Allocate per-CPU caches. */
2087: memset(pc->pc_cpus, 0, sizeof(pc->pc_cpus));
2088: pc->pc_ncpu = 0;
1.139 ad 2089: if (ncpu < 2) {
1.137 ad 2090: /* XXX For sparc: boot CPU is not attached yet. */
2091: pool_cache_cpu_init1(curcpu(), pc);
2092: } else {
2093: for (CPU_INFO_FOREACH(cii, ci)) {
2094: pool_cache_cpu_init1(ci, pc);
2095: }
1.134 ad 2096: }
1.145 ad 2097:
2098: /* Add to list of all pools. */
2099: if (__predict_true(!cold))
1.134 ad 2100: mutex_enter(&pool_head_lock);
1.145 ad 2101: TAILQ_FOREACH(pc1, &pool_cache_head, pc_cachelist) {
2102: if (strcmp(pc1->pc_pool.pr_wchan, pc->pc_pool.pr_wchan) > 0)
2103: break;
2104: }
2105: if (pc1 == NULL)
2106: TAILQ_INSERT_TAIL(&pool_cache_head, pc, pc_cachelist);
2107: else
2108: TAILQ_INSERT_BEFORE(pc1, pc, pc_cachelist);
2109: if (__predict_true(!cold))
1.134 ad 2110: mutex_exit(&pool_head_lock);
1.145 ad 2111:
2112: membar_sync();
2113: pp->pr_cache = pc;
1.43 thorpej 2114: }
2115:
2116: /*
2117: * pool_cache_destroy:
2118: *
2119: * Destroy a pool cache.
2120: */
2121: void
1.134 ad 2122: pool_cache_destroy(pool_cache_t pc)
1.43 thorpej 2123: {
1.134 ad 2124: struct pool *pp = &pc->pc_pool;
2125: pool_cache_cpu_t *cc;
2126: pcg_t *pcg;
2127: int i;
2128:
2129: /* Remove it from the global list. */
2130: mutex_enter(&pool_head_lock);
2131: while (pc->pc_refcnt != 0)
2132: cv_wait(&pool_busy, &pool_head_lock);
1.145 ad 2133: TAILQ_REMOVE(&pool_cache_head, pc, pc_cachelist);
1.134 ad 2134: mutex_exit(&pool_head_lock);
1.43 thorpej 2135:
2136: /* First, invalidate the entire cache. */
2137: pool_cache_invalidate(pc);
2138:
1.134 ad 2139: /* Disassociate it from the pool. */
2140: mutex_enter(&pp->pr_lock);
2141: pp->pr_cache = NULL;
2142: mutex_exit(&pp->pr_lock);
2143:
2144: /* Destroy per-CPU data */
2145: for (i = 0; i < MAXCPUS; i++) {
2146: if ((cc = pc->pc_cpus[i]) == NULL)
2147: continue;
1.158.2.2 yamt 2148: if ((pcg = cc->cc_current) != &pcg_dummy) {
1.134 ad 2149: pcg->pcg_next = NULL;
2150: pool_cache_invalidate_groups(pc, pcg);
2151: }
1.158.2.2 yamt 2152: if ((pcg = cc->cc_previous) != &pcg_dummy) {
1.134 ad 2153: pcg->pcg_next = NULL;
2154: pool_cache_invalidate_groups(pc, pcg);
2155: }
2156: if (cc != &pc->pc_cpu0)
2157: pool_put(&cache_cpu_pool, cc);
2158: }
2159:
2160: /* Finally, destroy it. */
2161: mutex_destroy(&pc->pc_lock);
2162: pool_destroy(pp);
2163: pool_put(&cache_pool, pc);
2164: }
2165:
2166: /*
2167: * pool_cache_cpu_init1:
2168: *
2169: * Called for each pool_cache whenever a new CPU is attached.
2170: */
2171: static void
2172: pool_cache_cpu_init1(struct cpu_info *ci, pool_cache_t pc)
2173: {
2174: pool_cache_cpu_t *cc;
1.137 ad 2175: int index;
1.134 ad 2176:
1.137 ad 2177: index = ci->ci_index;
2178:
2179: KASSERT(index < MAXCPUS);
1.134 ad 2180:
1.137 ad 2181: if ((cc = pc->pc_cpus[index]) != NULL) {
2182: KASSERT(cc->cc_cpuindex == index);
1.134 ad 2183: return;
2184: }
2185:
2186: /*
2187: * The first CPU is 'free'. This needs to be the case for
2188: * bootstrap - we may not be able to allocate yet.
2189: */
2190: if (pc->pc_ncpu == 0) {
2191: cc = &pc->pc_cpu0;
2192: pc->pc_ncpu = 1;
2193: } else {
2194: mutex_enter(&pc->pc_lock);
2195: pc->pc_ncpu++;
2196: mutex_exit(&pc->pc_lock);
2197: cc = pool_get(&cache_cpu_pool, PR_WAITOK);
2198: }
2199:
2200: cc->cc_ipl = pc->pc_pool.pr_ipl;
2201: cc->cc_iplcookie = makeiplcookie(cc->cc_ipl);
2202: cc->cc_cache = pc;
1.137 ad 2203: cc->cc_cpuindex = index;
1.134 ad 2204: cc->cc_hits = 0;
2205: cc->cc_misses = 0;
1.158.2.2 yamt 2206: cc->cc_current = __UNCONST(&pcg_dummy);
2207: cc->cc_previous = __UNCONST(&pcg_dummy);
1.134 ad 2208:
1.137 ad 2209: pc->pc_cpus[index] = cc;
1.43 thorpej 2210: }
2211:
1.134 ad 2212: /*
2213: * pool_cache_cpu_init:
2214: *
2215: * Called whenever a new CPU is attached.
2216: */
2217: void
2218: pool_cache_cpu_init(struct cpu_info *ci)
1.43 thorpej 2219: {
1.134 ad 2220: pool_cache_t pc;
2221:
2222: mutex_enter(&pool_head_lock);
1.145 ad 2223: TAILQ_FOREACH(pc, &pool_cache_head, pc_cachelist) {
1.134 ad 2224: pc->pc_refcnt++;
2225: mutex_exit(&pool_head_lock);
1.43 thorpej 2226:
1.134 ad 2227: pool_cache_cpu_init1(ci, pc);
1.43 thorpej 2228:
1.134 ad 2229: mutex_enter(&pool_head_lock);
2230: pc->pc_refcnt--;
2231: cv_broadcast(&pool_busy);
2232: }
2233: mutex_exit(&pool_head_lock);
1.43 thorpej 2234: }
2235:
1.134 ad 2236: /*
2237: * pool_cache_reclaim:
2238: *
2239: * Reclaim memory from a pool cache.
2240: */
2241: bool
2242: pool_cache_reclaim(pool_cache_t pc)
1.43 thorpej 2243: {
2244:
1.134 ad 2245: return pool_reclaim(&pc->pc_pool);
2246: }
1.43 thorpej 2247:
1.136 yamt 2248: static void
2249: pool_cache_destruct_object1(pool_cache_t pc, void *object)
2250: {
2251:
2252: (*pc->pc_dtor)(pc->pc_arg, object);
2253: pool_put(&pc->pc_pool, object);
2254: }
2255:
1.134 ad 2256: /*
2257: * pool_cache_destruct_object:
2258: *
2259: * Force destruction of an object and its release back into
2260: * the pool.
2261: */
2262: void
2263: pool_cache_destruct_object(pool_cache_t pc, void *object)
2264: {
2265:
1.136 yamt 2266: FREECHECK_IN(&pc->pc_freecheck, object);
2267:
2268: pool_cache_destruct_object1(pc, object);
1.43 thorpej 2269: }
2270:
1.134 ad 2271: /*
2272: * pool_cache_invalidate_groups:
2273: *
2274: * Invalidate a chain of groups and destruct all objects.
2275: */
1.102 chs 2276: static void
1.134 ad 2277: pool_cache_invalidate_groups(pool_cache_t pc, pcg_t *pcg)
1.102 chs 2278: {
1.134 ad 2279: void *object;
2280: pcg_t *next;
2281: int i;
2282:
2283: for (; pcg != NULL; pcg = next) {
2284: next = pcg->pcg_next;
2285:
2286: for (i = 0; i < pcg->pcg_avail; i++) {
2287: object = pcg->pcg_objects[i].pcgo_va;
1.136 yamt 2288: pool_cache_destruct_object1(pc, object);
1.134 ad 2289: }
1.102 chs 2290:
1.142 ad 2291: if (pcg->pcg_size == PCG_NOBJECTS_LARGE) {
2292: pool_put(&pcg_large_pool, pcg);
2293: } else {
2294: KASSERT(pcg->pcg_size == PCG_NOBJECTS_NORMAL);
2295: pool_put(&pcg_normal_pool, pcg);
2296: }
1.102 chs 2297: }
2298: }
2299:
1.43 thorpej 2300: /*
1.134 ad 2301: * pool_cache_invalidate:
1.43 thorpej 2302: *
1.134 ad 2303: * Invalidate a pool cache (destruct and release all of the
2304: * cached objects). Does not reclaim objects from the pool.
1.43 thorpej 2305: */
1.134 ad 2306: void
2307: pool_cache_invalidate(pool_cache_t pc)
2308: {
2309: pcg_t *full, *empty, *part;
2310:
2311: mutex_enter(&pc->pc_lock);
2312: full = pc->pc_fullgroups;
2313: empty = pc->pc_emptygroups;
2314: part = pc->pc_partgroups;
2315: pc->pc_fullgroups = NULL;
2316: pc->pc_emptygroups = NULL;
2317: pc->pc_partgroups = NULL;
2318: pc->pc_nfull = 0;
2319: pc->pc_nempty = 0;
2320: pc->pc_npart = 0;
2321: mutex_exit(&pc->pc_lock);
2322:
2323: pool_cache_invalidate_groups(pc, full);
2324: pool_cache_invalidate_groups(pc, empty);
2325: pool_cache_invalidate_groups(pc, part);
2326: }
2327:
2328: void
2329: pool_cache_set_drain_hook(pool_cache_t pc, void (*fn)(void *, int), void *arg)
2330: {
2331:
2332: pool_set_drain_hook(&pc->pc_pool, fn, arg);
2333: }
2334:
2335: void
2336: pool_cache_setlowat(pool_cache_t pc, int n)
2337: {
2338:
2339: pool_setlowat(&pc->pc_pool, n);
2340: }
2341:
2342: void
2343: pool_cache_sethiwat(pool_cache_t pc, int n)
2344: {
2345:
2346: pool_sethiwat(&pc->pc_pool, n);
2347: }
2348:
2349: void
2350: pool_cache_sethardlimit(pool_cache_t pc, int n, const char *warnmess, int ratecap)
2351: {
2352:
2353: pool_sethardlimit(&pc->pc_pool, n, warnmess, ratecap);
2354: }
2355:
1.158.2.2 yamt 2356: static bool __noinline
2357: pool_cache_get_slow(pool_cache_cpu_t *cc, int s, void **objectp,
1.134 ad 2358: paddr_t *pap, int flags)
1.43 thorpej 2359: {
1.134 ad 2360: pcg_t *pcg, *cur;
2361: uint64_t ncsw;
2362: pool_cache_t pc;
1.43 thorpej 2363: void *object;
1.58 thorpej 2364:
1.158.2.2 yamt 2365: KASSERT(cc->cc_current->pcg_avail == 0);
2366: KASSERT(cc->cc_previous->pcg_avail == 0);
2367:
1.134 ad 2368: pc = cc->cc_cache;
2369: cc->cc_misses++;
1.43 thorpej 2370:
1.134 ad 2371: /*
2372: * Nothing was available locally. Try and grab a group
2373: * from the cache.
2374: */
1.158.2.2 yamt 2375: if (__predict_false(!mutex_tryenter(&pc->pc_lock))) {
1.134 ad 2376: ncsw = curlwp->l_ncsw;
2377: mutex_enter(&pc->pc_lock);
2378: pc->pc_contended++;
1.43 thorpej 2379:
1.134 ad 2380: /*
2381: * If we context switched while locking, then
2382: * our view of the per-CPU data is invalid:
2383: * retry.
2384: */
2385: if (curlwp->l_ncsw != ncsw) {
2386: mutex_exit(&pc->pc_lock);
1.158.2.2 yamt 2387: return true;
1.43 thorpej 2388: }
1.102 chs 2389: }
1.43 thorpej 2390:
1.158.2.2 yamt 2391: if (__predict_true((pcg = pc->pc_fullgroups) != NULL)) {
1.43 thorpej 2392: /*
1.134 ad 2393: * If there's a full group, release our empty
2394: * group back to the cache. Install the full
2395: * group as cc_current and return.
1.43 thorpej 2396: */
1.158.2.2 yamt 2397: if (__predict_true((cur = cc->cc_current) != &pcg_dummy)) {
1.134 ad 2398: KASSERT(cur->pcg_avail == 0);
2399: cur->pcg_next = pc->pc_emptygroups;
2400: pc->pc_emptygroups = cur;
2401: pc->pc_nempty++;
1.87 thorpej 2402: }
1.142 ad 2403: KASSERT(pcg->pcg_avail == pcg->pcg_size);
1.134 ad 2404: cc->cc_current = pcg;
2405: pc->pc_fullgroups = pcg->pcg_next;
2406: pc->pc_hits++;
2407: pc->pc_nfull--;
2408: mutex_exit(&pc->pc_lock);
1.158.2.2 yamt 2409: return true;
1.134 ad 2410: }
2411:
2412: /*
2413: * Nothing available locally or in cache. Take the slow
2414: * path: fetch a new object from the pool and construct
2415: * it.
2416: */
2417: pc->pc_misses++;
2418: mutex_exit(&pc->pc_lock);
1.158.2.2 yamt 2419: splx(s);
1.134 ad 2420:
2421: object = pool_get(&pc->pc_pool, flags);
2422: *objectp = object;
1.158.2.2 yamt 2423: if (__predict_false(object == NULL))
2424: return false;
1.125 ad 2425:
1.158.2.2 yamt 2426: if (__predict_false((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0)) {
1.134 ad 2427: pool_put(&pc->pc_pool, object);
2428: *objectp = NULL;
1.158.2.2 yamt 2429: return false;
1.43 thorpej 2430: }
2431:
1.134 ad 2432: KASSERT((((vaddr_t)object + pc->pc_pool.pr_itemoffset) &
2433: (pc->pc_pool.pr_align - 1)) == 0);
1.43 thorpej 2434:
1.134 ad 2435: if (pap != NULL) {
2436: #ifdef POOL_VTOPHYS
2437: *pap = POOL_VTOPHYS(object);
2438: #else
2439: *pap = POOL_PADDR_INVALID;
2440: #endif
1.102 chs 2441: }
1.43 thorpej 2442:
1.125 ad 2443: FREECHECK_OUT(&pc->pc_freecheck, object);
1.158.2.2 yamt 2444: return false;
1.43 thorpej 2445: }
2446:
2447: /*
1.134 ad 2448: * pool_cache_get{,_paddr}:
1.43 thorpej 2449: *
1.134 ad 2450: * Get an object from a pool cache (optionally returning
2451: * the physical address of the object).
1.43 thorpej 2452: */
1.134 ad 2453: void *
2454: pool_cache_get_paddr(pool_cache_t pc, int flags, paddr_t *pap)
1.43 thorpej 2455: {
1.134 ad 2456: pool_cache_cpu_t *cc;
2457: pcg_t *pcg;
2458: void *object;
1.60 thorpej 2459: int s;
1.43 thorpej 2460:
1.134 ad 2461: #ifdef LOCKDEBUG
1.155 ad 2462: if (flags & PR_WAITOK) {
1.154 yamt 2463: ASSERT_SLEEPABLE();
1.155 ad 2464: }
1.134 ad 2465: #endif
1.125 ad 2466:
1.158.2.2 yamt 2467: /* Lock out interrupts and disable preemption. */
2468: s = splvm();
2469: while (/* CONSTCOND */ true) {
1.134 ad 2470: /* Try and allocate an object from the current group. */
1.158.2.2 yamt 2471: cc = pc->pc_cpus[curcpu()->ci_index];
2472: KASSERT(cc->cc_cache == pc);
1.134 ad 2473: pcg = cc->cc_current;
1.158.2.2 yamt 2474: if (__predict_true(pcg->pcg_avail > 0)) {
1.134 ad 2475: object = pcg->pcg_objects[--pcg->pcg_avail].pcgo_va;
1.158.2.2 yamt 2476: if (__predict_false(pap != NULL))
1.134 ad 2477: *pap = pcg->pcg_objects[pcg->pcg_avail].pcgo_pa;
1.148 yamt 2478: #if defined(DIAGNOSTIC)
1.134 ad 2479: pcg->pcg_objects[pcg->pcg_avail].pcgo_va = NULL;
1.158.2.2 yamt 2480: KASSERT(pcg->pcg_avail < pcg->pcg_size);
1.134 ad 2481: KASSERT(object != NULL);
1.158.2.2 yamt 2482: #endif
1.134 ad 2483: cc->cc_hits++;
1.158.2.2 yamt 2484: splx(s);
1.134 ad 2485: FREECHECK_OUT(&pc->pc_freecheck, object);
2486: return object;
1.43 thorpej 2487: }
2488:
2489: /*
1.134 ad 2490: * That failed. If the previous group isn't empty, swap
2491: * it with the current group and allocate from there.
1.43 thorpej 2492: */
1.134 ad 2493: pcg = cc->cc_previous;
1.158.2.2 yamt 2494: if (__predict_true(pcg->pcg_avail > 0)) {
1.134 ad 2495: cc->cc_previous = cc->cc_current;
2496: cc->cc_current = pcg;
2497: continue;
1.43 thorpej 2498: }
2499:
1.134 ad 2500: /*
2501: * Can't allocate from either group: try the slow path.
2502: * If get_slow() allocated an object for us, or if
1.158.2.2 yamt 2503: * no more objects are available, it will return false.
1.134 ad 2504: * Otherwise, we need to retry.
2505: */
1.158.2.2 yamt 2506: if (!pool_cache_get_slow(cc, s, &object, pap, flags))
2507: break;
2508: }
1.43 thorpej 2509:
1.134 ad 2510: return object;
1.51 thorpej 2511: }
2512:
1.158.2.2 yamt 2513: static bool __noinline
2514: pool_cache_put_slow(pool_cache_cpu_t *cc, int s, void *object)
1.51 thorpej 2515: {
1.134 ad 2516: pcg_t *pcg, *cur;
2517: uint64_t ncsw;
2518: pool_cache_t pc;
1.158.2.2 yamt 2519:
2520: KASSERT(cc->cc_current->pcg_avail == cc->cc_current->pcg_size);
2521: KASSERT(cc->cc_previous->pcg_avail == cc->cc_previous->pcg_size);
1.51 thorpej 2522:
1.134 ad 2523: pc = cc->cc_cache;
1.158.2.2 yamt 2524: pcg = NULL;
1.134 ad 2525: cc->cc_misses++;
1.43 thorpej 2526:
1.134 ad 2527: /*
1.158.2.2 yamt 2528: * If there are no empty groups in the cache then allocate one
2529: * while still unlocked.
1.134 ad 2530: */
1.158.2.2 yamt 2531: if (__predict_false(pc->pc_emptygroups == NULL)) {
2532: if (__predict_true(!pool_cache_disable)) {
2533: pcg = pool_get(pc->pc_pcgpool, PR_NOWAIT);
2534: }
2535: if (__predict_true(pcg != NULL)) {
2536: pcg->pcg_avail = 0;
2537: pcg->pcg_size = pc->pc_pcgsize;
2538: }
2539: }
2540:
2541: /* Lock the cache. */
2542: if (__predict_false(!mutex_tryenter(&pc->pc_lock))) {
1.134 ad 2543: ncsw = curlwp->l_ncsw;
2544: mutex_enter(&pc->pc_lock);
2545: pc->pc_contended++;
1.102 chs 2546:
1.134 ad 2547: /*
1.158.2.2 yamt 2548: * If we context switched while locking, then our view of
2549: * the per-CPU data is invalid: retry.
1.134 ad 2550: */
1.158.2.2 yamt 2551: if (__predict_false(curlwp->l_ncsw != ncsw)) {
1.134 ad 2552: mutex_exit(&pc->pc_lock);
1.158.2.2 yamt 2553: if (pcg != NULL) {
2554: pool_put(pc->pc_pcgpool, pcg);
2555: }
2556: return true;
1.134 ad 2557: }
2558: }
1.130 ad 2559:
1.158.2.2 yamt 2560: /* If there are no empty groups in the cache then allocate one. */
2561: if (pcg == NULL && pc->pc_emptygroups != NULL) {
2562: pcg = pc->pc_emptygroups;
1.134 ad 2563: pc->pc_emptygroups = pcg->pcg_next;
1.158.2.2 yamt 2564: pc->pc_nempty--;
2565: }
2566:
2567: /*
2568: * If there's a empty group, release our full group back
2569: * to the cache. Install the empty group to the local CPU
2570: * and return.
2571: */
2572: if (pcg != NULL) {
2573: KASSERT(pcg->pcg_avail == 0);
2574: if (__predict_false(cc->cc_previous == &pcg_dummy)) {
1.146 ad 2575: cc->cc_previous = pcg;
2576: } else {
1.158.2.2 yamt 2577: cur = cc->cc_current;
2578: if (__predict_true(cur != &pcg_dummy)) {
2579: KASSERT(cur->pcg_avail == cur->pcg_size);
1.146 ad 2580: cur->pcg_next = pc->pc_fullgroups;
2581: pc->pc_fullgroups = cur;
2582: pc->pc_nfull++;
2583: }
2584: cc->cc_current = pcg;
2585: }
1.134 ad 2586: pc->pc_hits++;
2587: mutex_exit(&pc->pc_lock);
1.158.2.2 yamt 2588: return true;
1.102 chs 2589: }
1.105 christos 2590:
1.134 ad 2591: /*
1.158.2.2 yamt 2592: * Nothing available locally or in cache, and we didn't
2593: * allocate an empty group. Take the slow path and destroy
2594: * the object here and now.
1.134 ad 2595: */
2596: pc->pc_misses++;
2597: mutex_exit(&pc->pc_lock);
1.158.2.2 yamt 2598: splx(s);
2599: pool_cache_destruct_object(pc, object);
1.103 chs 2600:
1.158.2.2 yamt 2601: return false;
1.134 ad 2602: }
1.102 chs 2603:
1.43 thorpej 2604: /*
1.134 ad 2605: * pool_cache_put{,_paddr}:
1.43 thorpej 2606: *
1.134 ad 2607: * Put an object back to the pool cache (optionally caching the
2608: * physical address of the object).
1.43 thorpej 2609: */
1.101 thorpej 2610: void
1.134 ad 2611: pool_cache_put_paddr(pool_cache_t pc, void *object, paddr_t pa)
1.43 thorpej 2612: {
1.134 ad 2613: pool_cache_cpu_t *cc;
2614: pcg_t *pcg;
2615: int s;
1.101 thorpej 2616:
1.158.2.2 yamt 2617: KASSERT(object != NULL);
1.134 ad 2618: FREECHECK_IN(&pc->pc_freecheck, object);
1.101 thorpej 2619:
1.158.2.2 yamt 2620: /* Lock out interrupts and disable preemption. */
2621: s = splvm();
2622: while (/* CONSTCOND */ true) {
1.134 ad 2623: /* If the current group isn't full, release it there. */
1.158.2.2 yamt 2624: cc = pc->pc_cpus[curcpu()->ci_index];
2625: KASSERT(cc->cc_cache == pc);
1.134 ad 2626: pcg = cc->cc_current;
1.158.2.2 yamt 2627: if (__predict_true(pcg->pcg_avail < pcg->pcg_size)) {
1.134 ad 2628: pcg->pcg_objects[pcg->pcg_avail].pcgo_va = object;
2629: pcg->pcg_objects[pcg->pcg_avail].pcgo_pa = pa;
2630: pcg->pcg_avail++;
2631: cc->cc_hits++;
1.158.2.2 yamt 2632: splx(s);
1.134 ad 2633: return;
2634: }
1.43 thorpej 2635:
1.134 ad 2636: /*
1.158.2.2 yamt 2637: * That failed. If the previous group isn't full, swap
1.134 ad 2638: * it with the current group and try again.
2639: */
2640: pcg = cc->cc_previous;
1.158.2.2 yamt 2641: if (__predict_true(pcg->pcg_avail < pcg->pcg_size)) {
1.134 ad 2642: cc->cc_previous = cc->cc_current;
2643: cc->cc_current = pcg;
2644: continue;
2645: }
1.43 thorpej 2646:
1.134 ad 2647: /*
2648: * Can't free to either group: try the slow path.
2649: * If put_slow() releases the object for us, it
1.158.2.2 yamt 2650: * will return false. Otherwise we need to retry.
1.134 ad 2651: */
1.158.2.2 yamt 2652: if (!pool_cache_put_slow(cc, s, object))
2653: break;
2654: }
1.43 thorpej 2655: }
2656:
2657: /*
1.134 ad 2658: * pool_cache_xcall:
1.43 thorpej 2659: *
1.134 ad 2660: * Transfer objects from the per-CPU cache to the global cache.
2661: * Run within a cross-call thread.
1.43 thorpej 2662: */
2663: static void
1.134 ad 2664: pool_cache_xcall(pool_cache_t pc)
1.43 thorpej 2665: {
1.134 ad 2666: pool_cache_cpu_t *cc;
2667: pcg_t *prev, *cur, **list;
1.158.2.2 yamt 2668: int s;
1.134 ad 2669:
2670: s = splvm();
2671: mutex_enter(&pc->pc_lock);
1.158.2.2 yamt 2672: cc = pc->pc_cpus[curcpu()->ci_index];
2673: cur = cc->cc_current;
2674: cc->cc_current = __UNCONST(&pcg_dummy);
2675: prev = cc->cc_previous;
2676: cc->cc_previous = __UNCONST(&pcg_dummy);
2677: if (cur != &pcg_dummy) {
1.142 ad 2678: if (cur->pcg_avail == cur->pcg_size) {
1.134 ad 2679: list = &pc->pc_fullgroups;
2680: pc->pc_nfull++;
2681: } else if (cur->pcg_avail == 0) {
2682: list = &pc->pc_emptygroups;
2683: pc->pc_nempty++;
2684: } else {
2685: list = &pc->pc_partgroups;
2686: pc->pc_npart++;
2687: }
2688: cur->pcg_next = *list;
2689: *list = cur;
2690: }
1.158.2.2 yamt 2691: if (prev != &pcg_dummy) {
1.142 ad 2692: if (prev->pcg_avail == prev->pcg_size) {
1.134 ad 2693: list = &pc->pc_fullgroups;
2694: pc->pc_nfull++;
2695: } else if (prev->pcg_avail == 0) {
2696: list = &pc->pc_emptygroups;
2697: pc->pc_nempty++;
2698: } else {
2699: list = &pc->pc_partgroups;
2700: pc->pc_npart++;
2701: }
2702: prev->pcg_next = *list;
2703: *list = prev;
2704: }
2705: mutex_exit(&pc->pc_lock);
2706: splx(s);
1.3 pk 2707: }
1.66 thorpej 2708:
2709: /*
2710: * Pool backend allocators.
2711: *
2712: * Each pool has a backend allocator that handles allocation, deallocation,
2713: * and any additional draining that might be needed.
2714: *
2715: * We provide two standard allocators:
2716: *
2717: * pool_allocator_kmem - the default when no allocator is specified
2718: *
2719: * pool_allocator_nointr - used for pools that will not be accessed
2720: * in interrupt context.
2721: */
2722: void *pool_page_alloc(struct pool *, int);
2723: void pool_page_free(struct pool *, void *);
2724:
1.112 bjh21 2725: #ifdef POOL_SUBPAGE
2726: struct pool_allocator pool_allocator_kmem_fullpage = {
2727: pool_page_alloc, pool_page_free, 0,
1.117 yamt 2728: .pa_backingmapptr = &kmem_map,
1.112 bjh21 2729: };
2730: #else
1.66 thorpej 2731: struct pool_allocator pool_allocator_kmem = {
2732: pool_page_alloc, pool_page_free, 0,
1.117 yamt 2733: .pa_backingmapptr = &kmem_map,
1.66 thorpej 2734: };
1.112 bjh21 2735: #endif
1.66 thorpej 2736:
2737: void *pool_page_alloc_nointr(struct pool *, int);
2738: void pool_page_free_nointr(struct pool *, void *);
2739:
1.112 bjh21 2740: #ifdef POOL_SUBPAGE
2741: struct pool_allocator pool_allocator_nointr_fullpage = {
2742: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1.117 yamt 2743: .pa_backingmapptr = &kernel_map,
1.112 bjh21 2744: };
2745: #else
1.66 thorpej 2746: struct pool_allocator pool_allocator_nointr = {
2747: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1.117 yamt 2748: .pa_backingmapptr = &kernel_map,
1.66 thorpej 2749: };
1.112 bjh21 2750: #endif
1.66 thorpej 2751:
2752: #ifdef POOL_SUBPAGE
2753: void *pool_subpage_alloc(struct pool *, int);
2754: void pool_subpage_free(struct pool *, void *);
2755:
1.112 bjh21 2756: struct pool_allocator pool_allocator_kmem = {
2757: pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE,
1.117 yamt 2758: .pa_backingmapptr = &kmem_map,
1.112 bjh21 2759: };
2760:
2761: void *pool_subpage_alloc_nointr(struct pool *, int);
2762: void pool_subpage_free_nointr(struct pool *, void *);
2763:
2764: struct pool_allocator pool_allocator_nointr = {
2765: pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE,
1.117 yamt 2766: .pa_backingmapptr = &kmem_map,
1.66 thorpej 2767: };
2768: #endif /* POOL_SUBPAGE */
2769:
1.117 yamt 2770: static void *
2771: pool_allocator_alloc(struct pool *pp, int flags)
1.66 thorpej 2772: {
1.117 yamt 2773: struct pool_allocator *pa = pp->pr_alloc;
1.66 thorpej 2774: void *res;
2775:
1.117 yamt 2776: res = (*pa->pa_alloc)(pp, flags);
2777: if (res == NULL && (flags & PR_WAITOK) == 0) {
1.66 thorpej 2778: /*
1.117 yamt 2779: * We only run the drain hook here if PR_NOWAIT.
2780: * In other cases, the hook will be run in
2781: * pool_reclaim().
1.66 thorpej 2782: */
1.117 yamt 2783: if (pp->pr_drain_hook != NULL) {
2784: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
2785: res = (*pa->pa_alloc)(pp, flags);
1.66 thorpej 2786: }
1.117 yamt 2787: }
2788: return res;
1.66 thorpej 2789: }
2790:
1.117 yamt 2791: static void
1.66 thorpej 2792: pool_allocator_free(struct pool *pp, void *v)
2793: {
2794: struct pool_allocator *pa = pp->pr_alloc;
2795:
2796: (*pa->pa_free)(pp, v);
2797: }
2798:
2799: void *
1.124 yamt 2800: pool_page_alloc(struct pool *pp, int flags)
1.66 thorpej 2801: {
1.127 thorpej 2802: bool waitok = (flags & PR_WAITOK) ? true : false;
1.66 thorpej 2803:
1.100 yamt 2804: return ((void *) uvm_km_alloc_poolpage_cache(kmem_map, waitok));
1.66 thorpej 2805: }
2806:
2807: void
1.124 yamt 2808: pool_page_free(struct pool *pp, void *v)
1.66 thorpej 2809: {
2810:
1.98 yamt 2811: uvm_km_free_poolpage_cache(kmem_map, (vaddr_t) v);
2812: }
2813:
2814: static void *
1.124 yamt 2815: pool_page_alloc_meta(struct pool *pp, int flags)
1.98 yamt 2816: {
1.127 thorpej 2817: bool waitok = (flags & PR_WAITOK) ? true : false;
1.98 yamt 2818:
1.100 yamt 2819: return ((void *) uvm_km_alloc_poolpage(kmem_map, waitok));
1.98 yamt 2820: }
2821:
2822: static void
1.124 yamt 2823: pool_page_free_meta(struct pool *pp, void *v)
1.98 yamt 2824: {
2825:
1.100 yamt 2826: uvm_km_free_poolpage(kmem_map, (vaddr_t) v);
1.66 thorpej 2827: }
2828:
2829: #ifdef POOL_SUBPAGE
2830: /* Sub-page allocator, for machines with large hardware pages. */
2831: void *
2832: pool_subpage_alloc(struct pool *pp, int flags)
2833: {
1.134 ad 2834: return pool_get(&psppool, flags);
1.66 thorpej 2835: }
2836:
2837: void
2838: pool_subpage_free(struct pool *pp, void *v)
2839: {
2840: pool_put(&psppool, v);
2841: }
2842:
2843: /* We don't provide a real nointr allocator. Maybe later. */
2844: void *
1.112 bjh21 2845: pool_subpage_alloc_nointr(struct pool *pp, int flags)
1.66 thorpej 2846: {
2847:
2848: return (pool_subpage_alloc(pp, flags));
2849: }
2850:
2851: void
1.112 bjh21 2852: pool_subpage_free_nointr(struct pool *pp, void *v)
1.66 thorpej 2853: {
2854:
2855: pool_subpage_free(pp, v);
2856: }
1.112 bjh21 2857: #endif /* POOL_SUBPAGE */
1.66 thorpej 2858: void *
1.124 yamt 2859: pool_page_alloc_nointr(struct pool *pp, int flags)
1.66 thorpej 2860: {
1.127 thorpej 2861: bool waitok = (flags & PR_WAITOK) ? true : false;
1.66 thorpej 2862:
1.100 yamt 2863: return ((void *) uvm_km_alloc_poolpage_cache(kernel_map, waitok));
1.66 thorpej 2864: }
2865:
2866: void
1.124 yamt 2867: pool_page_free_nointr(struct pool *pp, void *v)
1.66 thorpej 2868: {
2869:
1.98 yamt 2870: uvm_km_free_poolpage_cache(kernel_map, (vaddr_t) v);
1.66 thorpej 2871: }
1.141 yamt 2872:
2873: #if defined(DDB)
2874: static bool
2875: pool_in_page(struct pool *pp, struct pool_item_header *ph, uintptr_t addr)
2876: {
2877:
2878: return (uintptr_t)ph->ph_page <= addr &&
2879: addr < (uintptr_t)ph->ph_page + pp->pr_alloc->pa_pagesz;
2880: }
2881:
1.143 yamt 2882: static bool
2883: pool_in_item(struct pool *pp, void *item, uintptr_t addr)
2884: {
2885:
2886: return (uintptr_t)item <= addr && addr < (uintptr_t)item + pp->pr_size;
2887: }
2888:
2889: static bool
2890: pool_in_cg(struct pool *pp, struct pool_cache_group *pcg, uintptr_t addr)
2891: {
2892: int i;
2893:
2894: if (pcg == NULL) {
2895: return false;
2896: }
1.144 yamt 2897: for (i = 0; i < pcg->pcg_avail; i++) {
1.143 yamt 2898: if (pool_in_item(pp, pcg->pcg_objects[i].pcgo_va, addr)) {
2899: return true;
2900: }
2901: }
2902: return false;
2903: }
2904:
2905: static bool
2906: pool_allocated(struct pool *pp, struct pool_item_header *ph, uintptr_t addr)
2907: {
2908:
2909: if ((pp->pr_roflags & PR_NOTOUCH) != 0) {
2910: unsigned int idx = pr_item_notouch_index(pp, ph, (void *)addr);
2911: pool_item_bitmap_t *bitmap =
2912: ph->ph_bitmap + (idx / BITMAP_SIZE);
2913: pool_item_bitmap_t mask = 1 << (idx & BITMAP_MASK);
2914:
2915: return (*bitmap & mask) == 0;
2916: } else {
2917: struct pool_item *pi;
2918:
2919: LIST_FOREACH(pi, &ph->ph_itemlist, pi_list) {
2920: if (pool_in_item(pp, pi, addr)) {
2921: return false;
2922: }
2923: }
2924: return true;
2925: }
2926: }
2927:
1.141 yamt 2928: void
2929: pool_whatis(uintptr_t addr, void (*pr)(const char *, ...))
2930: {
2931: struct pool *pp;
2932:
1.145 ad 2933: TAILQ_FOREACH(pp, &pool_head, pr_poollist) {
1.141 yamt 2934: struct pool_item_header *ph;
2935: uintptr_t item;
1.143 yamt 2936: bool allocated = true;
2937: bool incache = false;
2938: bool incpucache = false;
2939: char cpucachestr[32];
1.141 yamt 2940:
2941: if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
2942: LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
2943: if (pool_in_page(pp, ph, addr)) {
2944: goto found;
2945: }
2946: }
2947: LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
2948: if (pool_in_page(pp, ph, addr)) {
1.143 yamt 2949: allocated =
2950: pool_allocated(pp, ph, addr);
2951: goto found;
2952: }
2953: }
2954: LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
2955: if (pool_in_page(pp, ph, addr)) {
2956: allocated = false;
1.141 yamt 2957: goto found;
2958: }
2959: }
2960: continue;
2961: } else {
2962: ph = pr_find_pagehead_noalign(pp, (void *)addr);
2963: if (ph == NULL || !pool_in_page(pp, ph, addr)) {
2964: continue;
2965: }
1.143 yamt 2966: allocated = pool_allocated(pp, ph, addr);
1.141 yamt 2967: }
2968: found:
1.143 yamt 2969: if (allocated && pp->pr_cache) {
2970: pool_cache_t pc = pp->pr_cache;
2971: struct pool_cache_group *pcg;
2972: int i;
2973:
2974: for (pcg = pc->pc_fullgroups; pcg != NULL;
2975: pcg = pcg->pcg_next) {
2976: if (pool_in_cg(pp, pcg, addr)) {
2977: incache = true;
2978: goto print;
2979: }
2980: }
2981: for (i = 0; i < MAXCPUS; i++) {
2982: pool_cache_cpu_t *cc;
2983:
2984: if ((cc = pc->pc_cpus[i]) == NULL) {
2985: continue;
2986: }
2987: if (pool_in_cg(pp, cc->cc_current, addr) ||
2988: pool_in_cg(pp, cc->cc_previous, addr)) {
2989: struct cpu_info *ci =
1.158.2.2 yamt 2990: cpu_lookup(i);
1.143 yamt 2991:
2992: incpucache = true;
2993: snprintf(cpucachestr,
2994: sizeof(cpucachestr),
2995: "cached by CPU %u",
1.153 martin 2996: ci->ci_index);
1.143 yamt 2997: goto print;
2998: }
2999: }
3000: }
3001: print:
1.141 yamt 3002: item = (uintptr_t)ph->ph_page + ph->ph_off;
3003: item = item + rounddown(addr - item, pp->pr_size);
1.143 yamt 3004: (*pr)("%p is %p+%zu in POOL '%s' (%s)\n",
1.141 yamt 3005: (void *)addr, item, (size_t)(addr - item),
1.143 yamt 3006: pp->pr_wchan,
3007: incpucache ? cpucachestr :
3008: incache ? "cached" : allocated ? "allocated" : "free");
1.141 yamt 3009: }
3010: }
3011: #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