Annotation of src/sys/kern/subr_pool.c, Revision 1.137.2.2
1.137.2.2! ad 1: /* $NetBSD: subr_pool.c,v 1.137.2.1 2007/12/08 17:57:46 ad Exp $ */
1.1 pk 2:
3: /*-
1.134 ad 4: * Copyright (c) 1997, 1999, 2000, 2002, 2007 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: * 3. All advertising materials mentioning features or use of this software
20: * must display the following acknowledgement:
1.13 christos 21: * This product includes software developed by the NetBSD
22: * Foundation, Inc. and its contributors.
1.1 pk 23: * 4. Neither the name of The NetBSD Foundation nor the names of its
24: * contributors may be used to endorse or promote products derived
25: * from this software without specific prior written permission.
26: *
27: * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
28: * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
29: * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
30: * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
31: * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
32: * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
33: * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
34: * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
35: * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
36: * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
37: * POSSIBILITY OF SUCH DAMAGE.
38: */
1.64 lukem 39:
40: #include <sys/cdefs.h>
1.137.2.2! ad 41: __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.137.2.1 2007/12/08 17:57:46 ad Exp $");
1.24 scottr 42:
1.25 thorpej 43: #include "opt_pool.h"
1.24 scottr 44: #include "opt_poollog.h"
1.28 thorpej 45: #include "opt_lockdebug.h"
1.1 pk 46:
47: #include <sys/param.h>
48: #include <sys/systm.h>
1.135 yamt 49: #include <sys/bitops.h>
1.1 pk 50: #include <sys/proc.h>
51: #include <sys/errno.h>
52: #include <sys/kernel.h>
53: #include <sys/malloc.h>
54: #include <sys/lock.h>
55: #include <sys/pool.h>
1.20 thorpej 56: #include <sys/syslog.h>
1.125 ad 57: #include <sys/debug.h>
1.134 ad 58: #include <sys/lockdebug.h>
59: #include <sys/xcall.h>
60: #include <sys/cpu.h>
1.3 pk 61:
62: #include <uvm/uvm.h>
63:
1.1 pk 64: /*
65: * Pool resource management utility.
1.3 pk 66: *
1.88 chs 67: * Memory is allocated in pages which are split into pieces according to
68: * the pool item size. Each page is kept on one of three lists in the
69: * pool structure: `pr_emptypages', `pr_fullpages' and `pr_partpages',
70: * for empty, full and partially-full pages respectively. The individual
71: * pool items are on a linked list headed by `ph_itemlist' in each page
72: * header. The memory for building the page list is either taken from
73: * the allocated pages themselves (for small pool items) or taken from
74: * an internal pool of page headers (`phpool').
1.1 pk 75: */
76:
1.3 pk 77: /* List of all pools */
1.102 chs 78: LIST_HEAD(,pool) pool_head = LIST_HEAD_INITIALIZER(pool_head);
1.3 pk 79:
80: /* Private pool for page header structures */
1.97 yamt 81: #define PHPOOL_MAX 8
82: static struct pool phpool[PHPOOL_MAX];
1.135 yamt 83: #define PHPOOL_FREELIST_NELEM(idx) \
84: (((idx) == 0) ? 0 : BITMAP_SIZE * (1 << (idx)))
1.3 pk 85:
1.62 bjh21 86: #ifdef POOL_SUBPAGE
87: /* Pool of subpages for use by normal pools. */
88: static struct pool psppool;
89: #endif
90:
1.117 yamt 91: static SLIST_HEAD(, pool_allocator) pa_deferinitq =
92: SLIST_HEAD_INITIALIZER(pa_deferinitq);
93:
1.98 yamt 94: static void *pool_page_alloc_meta(struct pool *, int);
95: static void pool_page_free_meta(struct pool *, void *);
96:
97: /* allocator for pool metadata */
1.134 ad 98: struct pool_allocator pool_allocator_meta = {
1.117 yamt 99: pool_page_alloc_meta, pool_page_free_meta,
100: .pa_backingmapptr = &kmem_map,
1.98 yamt 101: };
102:
1.3 pk 103: /* # of seconds to retain page after last use */
104: int pool_inactive_time = 10;
105:
106: /* Next candidate for drainage (see pool_drain()) */
1.23 thorpej 107: static struct pool *drainpp;
108:
1.134 ad 109: /* This lock protects both pool_head and drainpp. */
110: static kmutex_t pool_head_lock;
111: static kcondvar_t pool_busy;
1.3 pk 112:
1.135 yamt 113: typedef uint32_t pool_item_bitmap_t;
114: #define BITMAP_SIZE (CHAR_BIT * sizeof(pool_item_bitmap_t))
115: #define BITMAP_MASK (BITMAP_SIZE - 1)
1.99 yamt 116:
1.3 pk 117: struct pool_item_header {
118: /* Page headers */
1.88 chs 119: LIST_ENTRY(pool_item_header)
1.3 pk 120: ph_pagelist; /* pool page list */
1.88 chs 121: SPLAY_ENTRY(pool_item_header)
122: ph_node; /* Off-page page headers */
1.128 christos 123: void * ph_page; /* this page's address */
1.3 pk 124: struct timeval ph_time; /* last referenced */
1.135 yamt 125: uint16_t ph_nmissing; /* # of chunks in use */
1.97 yamt 126: union {
127: /* !PR_NOTOUCH */
128: struct {
1.102 chs 129: LIST_HEAD(, pool_item)
1.97 yamt 130: phu_itemlist; /* chunk list for this page */
131: } phu_normal;
132: /* PR_NOTOUCH */
133: struct {
1.135 yamt 134: uint16_t phu_off; /* start offset in page */
135: pool_item_bitmap_t phu_bitmap[];
1.97 yamt 136: } phu_notouch;
137: } ph_u;
1.3 pk 138: };
1.97 yamt 139: #define ph_itemlist ph_u.phu_normal.phu_itemlist
140: #define ph_off ph_u.phu_notouch.phu_off
1.135 yamt 141: #define ph_bitmap ph_u.phu_notouch.phu_bitmap
1.3 pk 142:
1.1 pk 143: struct pool_item {
1.3 pk 144: #ifdef DIAGNOSTIC
1.82 thorpej 145: u_int pi_magic;
1.33 chs 146: #endif
1.134 ad 147: #define PI_MAGIC 0xdeaddeadU
1.3 pk 148: /* Other entries use only this list entry */
1.102 chs 149: LIST_ENTRY(pool_item) pi_list;
1.3 pk 150: };
151:
1.53 thorpej 152: #define POOL_NEEDS_CATCHUP(pp) \
153: ((pp)->pr_nitems < (pp)->pr_minitems)
154:
1.43 thorpej 155: /*
156: * Pool cache management.
157: *
158: * Pool caches provide a way for constructed objects to be cached by the
159: * pool subsystem. This can lead to performance improvements by avoiding
160: * needless object construction/destruction; it is deferred until absolutely
161: * necessary.
162: *
1.134 ad 163: * Caches are grouped into cache groups. Each cache group references up
164: * to PCG_NUMOBJECTS constructed objects. When a cache allocates an
165: * object from the pool, it calls the object's constructor and places it
166: * into a cache group. When a cache group frees an object back to the
167: * pool, it first calls the object's destructor. This allows the object
168: * to persist in constructed form while freed to the cache.
169: *
170: * The pool references each cache, so that when a pool is drained by the
171: * pagedaemon, it can drain each individual cache as well. Each time a
172: * cache is drained, the most idle cache group is freed to the pool in
173: * its entirety.
1.43 thorpej 174: *
175: * Pool caches are layed on top of pools. By layering them, we can avoid
176: * the complexity of cache management for pools which would not benefit
177: * from it.
178: */
179:
180: static struct pool pcgpool;
1.134 ad 181: static struct pool cache_pool;
182: static struct pool cache_cpu_pool;
1.3 pk 183:
1.137.2.2! ad 184: /* List of all caches. */
! 185: LIST_HEAD(,pool_cache) pool_cache_head =
! 186: LIST_HEAD_INITIALIZER(pool_cache_head);
! 187:
! 188: int pool_cache_disable;
! 189:
! 190:
1.134 ad 191: static pool_cache_cpu_t *pool_cache_put_slow(pool_cache_cpu_t *, int *,
192: void *, paddr_t);
193: static pool_cache_cpu_t *pool_cache_get_slow(pool_cache_cpu_t *, int *,
194: void **, paddr_t *, int);
195: static void pool_cache_cpu_init1(struct cpu_info *, pool_cache_t);
196: static void pool_cache_invalidate_groups(pool_cache_t, pcg_t *);
197: static void pool_cache_xcall(pool_cache_t);
1.3 pk 198:
1.42 thorpej 199: static int pool_catchup(struct pool *);
1.128 christos 200: static void pool_prime_page(struct pool *, void *,
1.55 thorpej 201: struct pool_item_header *);
1.88 chs 202: static void pool_update_curpage(struct pool *);
1.66 thorpej 203:
1.113 yamt 204: static int pool_grow(struct pool *, int);
1.117 yamt 205: static void *pool_allocator_alloc(struct pool *, int);
206: static void pool_allocator_free(struct pool *, void *);
1.3 pk 207:
1.97 yamt 208: static void pool_print_pagelist(struct pool *, struct pool_pagelist *,
1.88 chs 209: void (*)(const char *, ...));
1.42 thorpej 210: static void pool_print1(struct pool *, const char *,
211: void (*)(const char *, ...));
1.3 pk 212:
1.88 chs 213: static int pool_chk_page(struct pool *, const char *,
214: struct pool_item_header *);
215:
1.3 pk 216: /*
1.52 thorpej 217: * Pool log entry. An array of these is allocated in pool_init().
1.3 pk 218: */
219: struct pool_log {
220: const char *pl_file;
221: long pl_line;
222: int pl_action;
1.25 thorpej 223: #define PRLOG_GET 1
224: #define PRLOG_PUT 2
1.3 pk 225: void *pl_addr;
1.1 pk 226: };
227:
1.86 matt 228: #ifdef POOL_DIAGNOSTIC
1.3 pk 229: /* Number of entries in pool log buffers */
1.17 thorpej 230: #ifndef POOL_LOGSIZE
231: #define POOL_LOGSIZE 10
232: #endif
233:
234: int pool_logsize = POOL_LOGSIZE;
1.1 pk 235:
1.110 perry 236: static inline void
1.42 thorpej 237: pr_log(struct pool *pp, void *v, int action, const char *file, long line)
1.3 pk 238: {
239: int n = pp->pr_curlogentry;
240: struct pool_log *pl;
241:
1.20 thorpej 242: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 243: return;
244:
245: /*
246: * Fill in the current entry. Wrap around and overwrite
247: * the oldest entry if necessary.
248: */
249: pl = &pp->pr_log[n];
250: pl->pl_file = file;
251: pl->pl_line = line;
252: pl->pl_action = action;
253: pl->pl_addr = v;
254: if (++n >= pp->pr_logsize)
255: n = 0;
256: pp->pr_curlogentry = n;
257: }
258:
259: static void
1.42 thorpej 260: pr_printlog(struct pool *pp, struct pool_item *pi,
261: void (*pr)(const char *, ...))
1.3 pk 262: {
263: int i = pp->pr_logsize;
264: int n = pp->pr_curlogentry;
265:
1.20 thorpej 266: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 267: return;
268:
269: /*
270: * Print all entries in this pool's log.
271: */
272: while (i-- > 0) {
273: struct pool_log *pl = &pp->pr_log[n];
274: if (pl->pl_action != 0) {
1.25 thorpej 275: if (pi == NULL || pi == pl->pl_addr) {
276: (*pr)("\tlog entry %d:\n", i);
277: (*pr)("\t\taction = %s, addr = %p\n",
278: pl->pl_action == PRLOG_GET ? "get" : "put",
279: pl->pl_addr);
280: (*pr)("\t\tfile: %s at line %lu\n",
281: pl->pl_file, pl->pl_line);
282: }
1.3 pk 283: }
284: if (++n >= pp->pr_logsize)
285: n = 0;
286: }
287: }
1.25 thorpej 288:
1.110 perry 289: static inline void
1.42 thorpej 290: pr_enter(struct pool *pp, const char *file, long line)
1.25 thorpej 291: {
292:
1.34 thorpej 293: if (__predict_false(pp->pr_entered_file != NULL)) {
1.25 thorpej 294: printf("pool %s: reentrancy at file %s line %ld\n",
295: pp->pr_wchan, file, line);
296: printf(" previous entry at file %s line %ld\n",
297: pp->pr_entered_file, pp->pr_entered_line);
298: panic("pr_enter");
299: }
300:
301: pp->pr_entered_file = file;
302: pp->pr_entered_line = line;
303: }
304:
1.110 perry 305: static inline void
1.42 thorpej 306: pr_leave(struct pool *pp)
1.25 thorpej 307: {
308:
1.34 thorpej 309: if (__predict_false(pp->pr_entered_file == NULL)) {
1.25 thorpej 310: printf("pool %s not entered?\n", pp->pr_wchan);
311: panic("pr_leave");
312: }
313:
314: pp->pr_entered_file = NULL;
315: pp->pr_entered_line = 0;
316: }
317:
1.110 perry 318: static inline void
1.42 thorpej 319: pr_enter_check(struct pool *pp, void (*pr)(const char *, ...))
1.25 thorpej 320: {
321:
322: if (pp->pr_entered_file != NULL)
323: (*pr)("\n\tcurrently entered from file %s line %ld\n",
324: pp->pr_entered_file, pp->pr_entered_line);
325: }
1.3 pk 326: #else
1.25 thorpej 327: #define pr_log(pp, v, action, file, line)
328: #define pr_printlog(pp, pi, pr)
329: #define pr_enter(pp, file, line)
330: #define pr_leave(pp)
331: #define pr_enter_check(pp, pr)
1.59 thorpej 332: #endif /* POOL_DIAGNOSTIC */
1.3 pk 333:
1.135 yamt 334: static inline unsigned int
1.97 yamt 335: pr_item_notouch_index(const struct pool *pp, const struct pool_item_header *ph,
336: const void *v)
337: {
338: const char *cp = v;
1.135 yamt 339: unsigned int idx;
1.97 yamt 340:
341: KASSERT(pp->pr_roflags & PR_NOTOUCH);
1.128 christos 342: idx = (cp - (char *)ph->ph_page - ph->ph_off) / pp->pr_size;
1.97 yamt 343: KASSERT(idx < pp->pr_itemsperpage);
344: return idx;
345: }
346:
1.110 perry 347: static inline void
1.97 yamt 348: pr_item_notouch_put(const struct pool *pp, struct pool_item_header *ph,
349: void *obj)
350: {
1.135 yamt 351: unsigned int idx = pr_item_notouch_index(pp, ph, obj);
352: pool_item_bitmap_t *bitmap = ph->ph_bitmap + (idx / BITMAP_SIZE);
353: pool_item_bitmap_t mask = 1 << (idx & BITMAP_MASK);
1.97 yamt 354:
1.135 yamt 355: KASSERT((*bitmap & mask) == 0);
356: *bitmap |= mask;
1.97 yamt 357: }
358:
1.110 perry 359: static inline void *
1.97 yamt 360: pr_item_notouch_get(const struct pool *pp, struct pool_item_header *ph)
361: {
1.135 yamt 362: pool_item_bitmap_t *bitmap = ph->ph_bitmap;
363: unsigned int idx;
364: int i;
1.97 yamt 365:
1.135 yamt 366: for (i = 0; ; i++) {
367: int bit;
1.97 yamt 368:
1.135 yamt 369: KASSERT((i * BITMAP_SIZE) < pp->pr_itemsperpage);
370: bit = ffs32(bitmap[i]);
371: if (bit) {
372: pool_item_bitmap_t mask;
373:
374: bit--;
375: idx = (i * BITMAP_SIZE) + bit;
376: mask = 1 << bit;
377: KASSERT((bitmap[i] & mask) != 0);
378: bitmap[i] &= ~mask;
379: break;
380: }
381: }
382: KASSERT(idx < pp->pr_itemsperpage);
1.128 christos 383: return (char *)ph->ph_page + ph->ph_off + idx * pp->pr_size;
1.97 yamt 384: }
385:
1.135 yamt 386: static inline void
387: pr_item_notouch_init(const struct pool *pp, struct pool_item_header *ph)
388: {
389: pool_item_bitmap_t *bitmap = ph->ph_bitmap;
390: const int n = howmany(pp->pr_itemsperpage, BITMAP_SIZE);
391: int i;
392:
393: for (i = 0; i < n; i++) {
394: bitmap[i] = (pool_item_bitmap_t)-1;
395: }
396: }
397:
1.110 perry 398: static inline int
1.88 chs 399: phtree_compare(struct pool_item_header *a, struct pool_item_header *b)
400: {
1.121 yamt 401:
402: /*
403: * we consider pool_item_header with smaller ph_page bigger.
404: * (this unnatural ordering is for the benefit of pr_find_pagehead.)
405: */
406:
1.88 chs 407: if (a->ph_page < b->ph_page)
1.121 yamt 408: return (1);
409: else if (a->ph_page > b->ph_page)
1.88 chs 410: return (-1);
411: else
412: return (0);
413: }
414:
415: SPLAY_PROTOTYPE(phtree, pool_item_header, ph_node, phtree_compare);
416: SPLAY_GENERATE(phtree, pool_item_header, ph_node, phtree_compare);
417:
1.3 pk 418: /*
1.121 yamt 419: * Return the pool page header based on item address.
1.3 pk 420: */
1.110 perry 421: static inline struct pool_item_header *
1.121 yamt 422: pr_find_pagehead(struct pool *pp, void *v)
1.3 pk 423: {
1.88 chs 424: struct pool_item_header *ph, tmp;
1.3 pk 425:
1.121 yamt 426: if ((pp->pr_roflags & PR_NOALIGN) != 0) {
1.128 christos 427: tmp.ph_page = (void *)(uintptr_t)v;
1.121 yamt 428: ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
429: if (ph == NULL) {
430: ph = SPLAY_ROOT(&pp->pr_phtree);
431: if (ph != NULL && phtree_compare(&tmp, ph) >= 0) {
432: ph = SPLAY_NEXT(phtree, &pp->pr_phtree, ph);
433: }
434: KASSERT(ph == NULL || phtree_compare(&tmp, ph) < 0);
435: }
436: } else {
1.128 christos 437: void *page =
438: (void *)((uintptr_t)v & pp->pr_alloc->pa_pagemask);
1.121 yamt 439:
440: if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
1.128 christos 441: ph = (struct pool_item_header *)((char *)page + pp->pr_phoffset);
1.121 yamt 442: } else {
443: tmp.ph_page = page;
444: ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
445: }
446: }
1.3 pk 447:
1.121 yamt 448: KASSERT(ph == NULL || ((pp->pr_roflags & PR_PHINPAGE) != 0) ||
1.128 christos 449: ((char *)ph->ph_page <= (char *)v &&
450: (char *)v < (char *)ph->ph_page + pp->pr_alloc->pa_pagesz));
1.88 chs 451: return ph;
1.3 pk 452: }
453:
1.101 thorpej 454: static void
455: pr_pagelist_free(struct pool *pp, struct pool_pagelist *pq)
456: {
457: struct pool_item_header *ph;
458:
459: while ((ph = LIST_FIRST(pq)) != NULL) {
460: LIST_REMOVE(ph, ph_pagelist);
461: pool_allocator_free(pp, ph->ph_page);
1.134 ad 462: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1.101 thorpej 463: pool_put(pp->pr_phpool, ph);
464: }
465: }
466:
1.3 pk 467: /*
468: * Remove a page from the pool.
469: */
1.110 perry 470: static inline void
1.61 chs 471: pr_rmpage(struct pool *pp, struct pool_item_header *ph,
472: struct pool_pagelist *pq)
1.3 pk 473: {
474:
1.134 ad 475: KASSERT(mutex_owned(&pp->pr_lock));
1.91 yamt 476:
1.3 pk 477: /*
1.7 thorpej 478: * If the page was idle, decrement the idle page count.
1.3 pk 479: */
1.6 thorpej 480: if (ph->ph_nmissing == 0) {
481: #ifdef DIAGNOSTIC
482: if (pp->pr_nidle == 0)
483: panic("pr_rmpage: nidle inconsistent");
1.20 thorpej 484: if (pp->pr_nitems < pp->pr_itemsperpage)
485: panic("pr_rmpage: nitems inconsistent");
1.6 thorpej 486: #endif
487: pp->pr_nidle--;
488: }
1.7 thorpej 489:
1.20 thorpej 490: pp->pr_nitems -= pp->pr_itemsperpage;
491:
1.7 thorpej 492: /*
1.101 thorpej 493: * Unlink the page from the pool and queue it for release.
1.7 thorpej 494: */
1.88 chs 495: LIST_REMOVE(ph, ph_pagelist);
1.91 yamt 496: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
497: SPLAY_REMOVE(phtree, &pp->pr_phtree, ph);
1.101 thorpej 498: LIST_INSERT_HEAD(pq, ph, ph_pagelist);
499:
1.7 thorpej 500: pp->pr_npages--;
501: pp->pr_npagefree++;
1.6 thorpej 502:
1.88 chs 503: pool_update_curpage(pp);
1.3 pk 504: }
505:
1.126 thorpej 506: static bool
1.117 yamt 507: pa_starved_p(struct pool_allocator *pa)
508: {
509:
510: if (pa->pa_backingmap != NULL) {
511: return vm_map_starved_p(pa->pa_backingmap);
512: }
1.127 thorpej 513: return false;
1.117 yamt 514: }
515:
516: static int
1.124 yamt 517: pool_reclaim_callback(struct callback_entry *ce, void *obj, void *arg)
1.117 yamt 518: {
519: struct pool *pp = obj;
520: struct pool_allocator *pa = pp->pr_alloc;
521:
522: KASSERT(&pp->pr_reclaimerentry == ce);
523: pool_reclaim(pp);
524: if (!pa_starved_p(pa)) {
525: return CALLBACK_CHAIN_ABORT;
526: }
527: return CALLBACK_CHAIN_CONTINUE;
528: }
529:
530: static void
531: pool_reclaim_register(struct pool *pp)
532: {
533: struct vm_map *map = pp->pr_alloc->pa_backingmap;
534: int s;
535:
536: if (map == NULL) {
537: return;
538: }
539:
540: s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */
541: callback_register(&vm_map_to_kernel(map)->vmk_reclaim_callback,
542: &pp->pr_reclaimerentry, pp, pool_reclaim_callback);
543: splx(s);
544: }
545:
546: static void
547: pool_reclaim_unregister(struct pool *pp)
548: {
549: struct vm_map *map = pp->pr_alloc->pa_backingmap;
550: int s;
551:
552: if (map == NULL) {
553: return;
554: }
555:
556: s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */
557: callback_unregister(&vm_map_to_kernel(map)->vmk_reclaim_callback,
558: &pp->pr_reclaimerentry);
559: splx(s);
560: }
561:
562: static void
563: pa_reclaim_register(struct pool_allocator *pa)
564: {
565: struct vm_map *map = *pa->pa_backingmapptr;
566: struct pool *pp;
567:
568: KASSERT(pa->pa_backingmap == NULL);
569: if (map == NULL) {
570: SLIST_INSERT_HEAD(&pa_deferinitq, pa, pa_q);
571: return;
572: }
573: pa->pa_backingmap = map;
574: TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
575: pool_reclaim_register(pp);
576: }
577: }
578:
1.3 pk 579: /*
1.94 simonb 580: * Initialize all the pools listed in the "pools" link set.
581: */
582: void
1.117 yamt 583: pool_subsystem_init(void)
1.94 simonb 584: {
1.117 yamt 585: struct pool_allocator *pa;
1.94 simonb 586: __link_set_decl(pools, struct link_pool_init);
587: struct link_pool_init * const *pi;
588:
1.134 ad 589: mutex_init(&pool_head_lock, MUTEX_DEFAULT, IPL_NONE);
590: cv_init(&pool_busy, "poolbusy");
591:
1.94 simonb 592: __link_set_foreach(pi, pools)
593: pool_init((*pi)->pp, (*pi)->size, (*pi)->align,
594: (*pi)->align_offset, (*pi)->flags, (*pi)->wchan,
1.129 ad 595: (*pi)->palloc, (*pi)->ipl);
1.117 yamt 596:
597: while ((pa = SLIST_FIRST(&pa_deferinitq)) != NULL) {
598: KASSERT(pa->pa_backingmapptr != NULL);
599: KASSERT(*pa->pa_backingmapptr != NULL);
600: SLIST_REMOVE_HEAD(&pa_deferinitq, pa_q);
601: pa_reclaim_register(pa);
602: }
1.134 ad 603:
604: pool_init(&cache_pool, sizeof(struct pool_cache), CACHE_LINE_SIZE,
605: 0, 0, "pcache", &pool_allocator_nointr, IPL_NONE);
606:
607: pool_init(&cache_cpu_pool, sizeof(pool_cache_cpu_t), CACHE_LINE_SIZE,
608: 0, 0, "pcachecpu", &pool_allocator_nointr, IPL_NONE);
1.94 simonb 609: }
610:
611: /*
1.3 pk 612: * Initialize the given pool resource structure.
613: *
614: * We export this routine to allow other kernel parts to declare
615: * static pools that must be initialized before malloc() is available.
616: */
617: void
1.42 thorpej 618: pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
1.129 ad 619: const char *wchan, struct pool_allocator *palloc, int ipl)
1.3 pk 620: {
1.116 simonb 621: #ifdef DEBUG
622: struct pool *pp1;
623: #endif
1.92 enami 624: size_t trysize, phsize;
1.134 ad 625: int off, slack;
1.3 pk 626:
1.116 simonb 627: #ifdef DEBUG
628: /*
629: * Check that the pool hasn't already been initialised and
630: * added to the list of all pools.
631: */
632: LIST_FOREACH(pp1, &pool_head, pr_poollist) {
633: if (pp == pp1)
634: panic("pool_init: pool %s already initialised",
635: wchan);
636: }
637: #endif
638:
1.25 thorpej 639: #ifdef POOL_DIAGNOSTIC
640: /*
641: * Always log if POOL_DIAGNOSTIC is defined.
642: */
643: if (pool_logsize != 0)
644: flags |= PR_LOGGING;
645: #endif
646:
1.66 thorpej 647: if (palloc == NULL)
648: palloc = &pool_allocator_kmem;
1.112 bjh21 649: #ifdef POOL_SUBPAGE
650: if (size > palloc->pa_pagesz) {
651: if (palloc == &pool_allocator_kmem)
652: palloc = &pool_allocator_kmem_fullpage;
653: else if (palloc == &pool_allocator_nointr)
654: palloc = &pool_allocator_nointr_fullpage;
655: }
1.66 thorpej 656: #endif /* POOL_SUBPAGE */
657: if ((palloc->pa_flags & PA_INITIALIZED) == 0) {
1.112 bjh21 658: if (palloc->pa_pagesz == 0)
1.66 thorpej 659: palloc->pa_pagesz = PAGE_SIZE;
660:
661: TAILQ_INIT(&palloc->pa_list);
662:
1.134 ad 663: mutex_init(&palloc->pa_lock, MUTEX_DEFAULT, IPL_VM);
1.66 thorpej 664: palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
665: palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
1.117 yamt 666:
667: if (palloc->pa_backingmapptr != NULL) {
668: pa_reclaim_register(palloc);
669: }
1.66 thorpej 670: palloc->pa_flags |= PA_INITIALIZED;
1.4 thorpej 671: }
1.3 pk 672:
673: if (align == 0)
674: align = ALIGN(1);
1.14 thorpej 675:
1.120 yamt 676: if ((flags & PR_NOTOUCH) == 0 && size < sizeof(struct pool_item))
1.14 thorpej 677: size = sizeof(struct pool_item);
1.3 pk 678:
1.78 thorpej 679: size = roundup(size, align);
1.66 thorpej 680: #ifdef DIAGNOSTIC
681: if (size > palloc->pa_pagesz)
1.121 yamt 682: panic("pool_init: pool item size (%zu) too large", size);
1.66 thorpej 683: #endif
1.35 pk 684:
1.3 pk 685: /*
686: * Initialize the pool structure.
687: */
1.88 chs 688: LIST_INIT(&pp->pr_emptypages);
689: LIST_INIT(&pp->pr_fullpages);
690: LIST_INIT(&pp->pr_partpages);
1.134 ad 691: pp->pr_cache = NULL;
1.3 pk 692: pp->pr_curpage = NULL;
693: pp->pr_npages = 0;
694: pp->pr_minitems = 0;
695: pp->pr_minpages = 0;
696: pp->pr_maxpages = UINT_MAX;
1.20 thorpej 697: pp->pr_roflags = flags;
698: pp->pr_flags = 0;
1.35 pk 699: pp->pr_size = size;
1.3 pk 700: pp->pr_align = align;
701: pp->pr_wchan = wchan;
1.66 thorpej 702: pp->pr_alloc = palloc;
1.20 thorpej 703: pp->pr_nitems = 0;
704: pp->pr_nout = 0;
705: pp->pr_hardlimit = UINT_MAX;
706: pp->pr_hardlimit_warning = NULL;
1.31 thorpej 707: pp->pr_hardlimit_ratecap.tv_sec = 0;
708: pp->pr_hardlimit_ratecap.tv_usec = 0;
709: pp->pr_hardlimit_warning_last.tv_sec = 0;
710: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.68 thorpej 711: pp->pr_drain_hook = NULL;
712: pp->pr_drain_hook_arg = NULL;
1.125 ad 713: pp->pr_freecheck = NULL;
1.3 pk 714:
715: /*
716: * Decide whether to put the page header off page to avoid
1.92 enami 717: * wasting too large a part of the page or too big item.
718: * Off-page page headers go on a hash table, so we can match
719: * a returned item with its header based on the page address.
720: * We use 1/16 of the page size and about 8 times of the item
721: * size as the threshold (XXX: tune)
722: *
723: * However, we'll put the header into the page if we can put
724: * it without wasting any items.
725: *
726: * Silently enforce `0 <= ioff < align'.
1.3 pk 727: */
1.92 enami 728: pp->pr_itemoffset = ioff %= align;
729: /* See the comment below about reserved bytes. */
730: trysize = palloc->pa_pagesz - ((align - ioff) % align);
731: phsize = ALIGN(sizeof(struct pool_item_header));
1.121 yamt 732: if ((pp->pr_roflags & (PR_NOTOUCH | PR_NOALIGN)) == 0 &&
1.97 yamt 733: (pp->pr_size < MIN(palloc->pa_pagesz / 16, phsize << 3) ||
734: trysize / pp->pr_size == (trysize - phsize) / pp->pr_size)) {
1.3 pk 735: /* Use the end of the page for the page header */
1.20 thorpej 736: pp->pr_roflags |= PR_PHINPAGE;
1.92 enami 737: pp->pr_phoffset = off = palloc->pa_pagesz - phsize;
1.2 pk 738: } else {
1.3 pk 739: /* The page header will be taken from our page header pool */
740: pp->pr_phoffset = 0;
1.66 thorpej 741: off = palloc->pa_pagesz;
1.88 chs 742: SPLAY_INIT(&pp->pr_phtree);
1.2 pk 743: }
1.1 pk 744:
1.3 pk 745: /*
746: * Alignment is to take place at `ioff' within the item. This means
747: * we must reserve up to `align - 1' bytes on the page to allow
748: * appropriate positioning of each item.
749: */
750: pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
1.43 thorpej 751: KASSERT(pp->pr_itemsperpage != 0);
1.97 yamt 752: if ((pp->pr_roflags & PR_NOTOUCH)) {
753: int idx;
754:
755: for (idx = 0; pp->pr_itemsperpage > PHPOOL_FREELIST_NELEM(idx);
756: idx++) {
757: /* nothing */
758: }
759: if (idx >= PHPOOL_MAX) {
760: /*
761: * if you see this panic, consider to tweak
762: * PHPOOL_MAX and PHPOOL_FREELIST_NELEM.
763: */
764: panic("%s: too large itemsperpage(%d) for PR_NOTOUCH",
765: pp->pr_wchan, pp->pr_itemsperpage);
766: }
767: pp->pr_phpool = &phpool[idx];
768: } else if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
769: pp->pr_phpool = &phpool[0];
770: }
771: #if defined(DIAGNOSTIC)
772: else {
773: pp->pr_phpool = NULL;
774: }
775: #endif
1.3 pk 776:
777: /*
778: * Use the slack between the chunks and the page header
779: * for "cache coloring".
780: */
781: slack = off - pp->pr_itemsperpage * pp->pr_size;
782: pp->pr_maxcolor = (slack / align) * align;
783: pp->pr_curcolor = 0;
784:
785: pp->pr_nget = 0;
786: pp->pr_nfail = 0;
787: pp->pr_nput = 0;
788: pp->pr_npagealloc = 0;
789: pp->pr_npagefree = 0;
1.1 pk 790: pp->pr_hiwat = 0;
1.8 thorpej 791: pp->pr_nidle = 0;
1.134 ad 792: pp->pr_refcnt = 0;
1.3 pk 793:
1.59 thorpej 794: #ifdef POOL_DIAGNOSTIC
1.25 thorpej 795: if (flags & PR_LOGGING) {
796: if (kmem_map == NULL ||
797: (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log),
798: M_TEMP, M_NOWAIT)) == NULL)
1.20 thorpej 799: pp->pr_roflags &= ~PR_LOGGING;
1.3 pk 800: pp->pr_curlogentry = 0;
801: pp->pr_logsize = pool_logsize;
802: }
1.59 thorpej 803: #endif
1.25 thorpej 804:
805: pp->pr_entered_file = NULL;
806: pp->pr_entered_line = 0;
1.3 pk 807:
1.137.2.1 ad 808: /*
809: * XXXAD hack to prevent IP input processing from blocking.
810: */
811: if (ipl == IPL_SOFTNET) {
812: mutex_init(&pp->pr_lock, MUTEX_DEFAULT, IPL_VM);
813: } else {
814: mutex_init(&pp->pr_lock, MUTEX_DEFAULT, ipl);
815: }
1.134 ad 816: cv_init(&pp->pr_cv, wchan);
817: pp->pr_ipl = ipl;
1.1 pk 818:
1.3 pk 819: /*
1.43 thorpej 820: * Initialize private page header pool and cache magazine pool if we
821: * haven't done so yet.
1.23 thorpej 822: * XXX LOCKING.
1.3 pk 823: */
1.97 yamt 824: if (phpool[0].pr_size == 0) {
825: int idx;
826: for (idx = 0; idx < PHPOOL_MAX; idx++) {
827: static char phpool_names[PHPOOL_MAX][6+1+6+1];
828: int nelem;
829: size_t sz;
830:
831: nelem = PHPOOL_FREELIST_NELEM(idx);
832: snprintf(phpool_names[idx], sizeof(phpool_names[idx]),
833: "phpool-%d", nelem);
834: sz = sizeof(struct pool_item_header);
835: if (nelem) {
1.135 yamt 836: sz = offsetof(struct pool_item_header,
837: ph_bitmap[howmany(nelem, BITMAP_SIZE)]);
1.97 yamt 838: }
839: pool_init(&phpool[idx], sz, 0, 0, 0,
1.129 ad 840: phpool_names[idx], &pool_allocator_meta, IPL_VM);
1.97 yamt 841: }
1.62 bjh21 842: #ifdef POOL_SUBPAGE
843: pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
1.129 ad 844: PR_RECURSIVE, "psppool", &pool_allocator_meta, IPL_VM);
1.62 bjh21 845: #endif
1.134 ad 846: pool_init(&pcgpool, sizeof(pcg_t), CACHE_LINE_SIZE, 0, 0,
847: "cachegrp", &pool_allocator_meta, IPL_VM);
1.1 pk 848: }
849:
1.134 ad 850: if (__predict_true(!cold)) {
851: /* Insert into the list of all pools. */
852: mutex_enter(&pool_head_lock);
853: LIST_INSERT_HEAD(&pool_head, pp, pr_poollist);
854: mutex_exit(&pool_head_lock);
855:
856: /* Insert this into the list of pools using this allocator. */
857: mutex_enter(&palloc->pa_lock);
858: TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
859: mutex_exit(&palloc->pa_lock);
860: } else {
861: LIST_INSERT_HEAD(&pool_head, pp, pr_poollist);
862: TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
863: }
1.66 thorpej 864:
1.117 yamt 865: pool_reclaim_register(pp);
1.1 pk 866: }
867:
868: /*
869: * De-commision a pool resource.
870: */
871: void
1.42 thorpej 872: pool_destroy(struct pool *pp)
1.1 pk 873: {
1.101 thorpej 874: struct pool_pagelist pq;
1.3 pk 875: struct pool_item_header *ph;
1.43 thorpej 876:
1.101 thorpej 877: /* Remove from global pool list */
1.134 ad 878: mutex_enter(&pool_head_lock);
879: while (pp->pr_refcnt != 0)
880: cv_wait(&pool_busy, &pool_head_lock);
1.102 chs 881: LIST_REMOVE(pp, pr_poollist);
1.101 thorpej 882: if (drainpp == pp)
883: drainpp = NULL;
1.134 ad 884: mutex_exit(&pool_head_lock);
1.101 thorpej 885:
886: /* Remove this pool from its allocator's list of pools. */
1.117 yamt 887: pool_reclaim_unregister(pp);
1.134 ad 888: mutex_enter(&pp->pr_alloc->pa_lock);
1.66 thorpej 889: TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
1.134 ad 890: mutex_exit(&pp->pr_alloc->pa_lock);
1.66 thorpej 891:
1.134 ad 892: mutex_enter(&pp->pr_lock);
1.101 thorpej 893:
1.134 ad 894: KASSERT(pp->pr_cache == NULL);
1.3 pk 895:
896: #ifdef DIAGNOSTIC
1.20 thorpej 897: if (pp->pr_nout != 0) {
1.25 thorpej 898: pr_printlog(pp, NULL, printf);
1.80 provos 899: panic("pool_destroy: pool busy: still out: %u",
1.20 thorpej 900: pp->pr_nout);
1.3 pk 901: }
902: #endif
1.1 pk 903:
1.101 thorpej 904: KASSERT(LIST_EMPTY(&pp->pr_fullpages));
905: KASSERT(LIST_EMPTY(&pp->pr_partpages));
906:
1.3 pk 907: /* Remove all pages */
1.101 thorpej 908: LIST_INIT(&pq);
1.88 chs 909: while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1.101 thorpej 910: pr_rmpage(pp, ph, &pq);
911:
1.134 ad 912: mutex_exit(&pp->pr_lock);
1.3 pk 913:
1.101 thorpej 914: pr_pagelist_free(pp, &pq);
1.3 pk 915:
1.59 thorpej 916: #ifdef POOL_DIAGNOSTIC
1.20 thorpej 917: if ((pp->pr_roflags & PR_LOGGING) != 0)
1.3 pk 918: free(pp->pr_log, M_TEMP);
1.59 thorpej 919: #endif
1.134 ad 920:
921: cv_destroy(&pp->pr_cv);
922: mutex_destroy(&pp->pr_lock);
1.1 pk 923: }
924:
1.68 thorpej 925: void
926: pool_set_drain_hook(struct pool *pp, void (*fn)(void *, int), void *arg)
927: {
928:
929: /* XXX no locking -- must be used just after pool_init() */
930: #ifdef DIAGNOSTIC
931: if (pp->pr_drain_hook != NULL)
932: panic("pool_set_drain_hook(%s): already set", pp->pr_wchan);
933: #endif
934: pp->pr_drain_hook = fn;
935: pp->pr_drain_hook_arg = arg;
936: }
937:
1.88 chs 938: static struct pool_item_header *
1.128 christos 939: pool_alloc_item_header(struct pool *pp, void *storage, int flags)
1.55 thorpej 940: {
941: struct pool_item_header *ph;
942:
943: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
1.128 christos 944: ph = (struct pool_item_header *) ((char *)storage + pp->pr_phoffset);
1.134 ad 945: else
1.97 yamt 946: ph = pool_get(pp->pr_phpool, flags);
1.55 thorpej 947:
948: return (ph);
949: }
1.1 pk 950:
951: /*
1.134 ad 952: * Grab an item from the pool.
1.1 pk 953: */
1.3 pk 954: void *
1.59 thorpej 955: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 956: _pool_get(struct pool *pp, int flags, const char *file, long line)
1.56 sommerfe 957: #else
958: pool_get(struct pool *pp, int flags)
959: #endif
1.1 pk 960: {
961: struct pool_item *pi;
1.3 pk 962: struct pool_item_header *ph;
1.55 thorpej 963: void *v;
1.1 pk 964:
1.2 pk 965: #ifdef DIAGNOSTIC
1.95 atatat 966: if (__predict_false(pp->pr_itemsperpage == 0))
967: panic("pool_get: pool %p: pr_itemsperpage is zero, "
968: "pool not initialized?", pp);
1.84 thorpej 969: if (__predict_false(curlwp == NULL && doing_shutdown == 0 &&
1.37 sommerfe 970: (flags & PR_WAITOK) != 0))
1.77 matt 971: panic("pool_get: %s: must have NOWAIT", pp->pr_wchan);
1.58 thorpej 972:
1.102 chs 973: #endif /* DIAGNOSTIC */
1.58 thorpej 974: #ifdef LOCKDEBUG
975: if (flags & PR_WAITOK)
1.119 yamt 976: ASSERT_SLEEPABLE(NULL, "pool_get(PR_WAITOK)");
1.56 sommerfe 977: #endif
1.1 pk 978:
1.134 ad 979: mutex_enter(&pp->pr_lock);
1.25 thorpej 980: pr_enter(pp, file, line);
1.20 thorpej 981:
982: startover:
983: /*
984: * Check to see if we've reached the hard limit. If we have,
985: * and we can wait, then wait until an item has been returned to
986: * the pool.
987: */
988: #ifdef DIAGNOSTIC
1.34 thorpej 989: if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
1.25 thorpej 990: pr_leave(pp);
1.134 ad 991: mutex_exit(&pp->pr_lock);
1.20 thorpej 992: panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
993: }
994: #endif
1.34 thorpej 995: if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
1.68 thorpej 996: if (pp->pr_drain_hook != NULL) {
997: /*
998: * Since the drain hook is going to free things
999: * back to the pool, unlock, call the hook, re-lock,
1000: * and check the hardlimit condition again.
1001: */
1002: pr_leave(pp);
1.134 ad 1003: mutex_exit(&pp->pr_lock);
1.68 thorpej 1004: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
1.134 ad 1005: mutex_enter(&pp->pr_lock);
1.68 thorpej 1006: pr_enter(pp, file, line);
1007: if (pp->pr_nout < pp->pr_hardlimit)
1008: goto startover;
1009: }
1010:
1.29 sommerfe 1011: if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
1.20 thorpej 1012: /*
1013: * XXX: A warning isn't logged in this case. Should
1014: * it be?
1015: */
1016: pp->pr_flags |= PR_WANTED;
1.25 thorpej 1017: pr_leave(pp);
1.134 ad 1018: cv_wait(&pp->pr_cv, &pp->pr_lock);
1.25 thorpej 1019: pr_enter(pp, file, line);
1.20 thorpej 1020: goto startover;
1021: }
1.31 thorpej 1022:
1023: /*
1024: * Log a message that the hard limit has been hit.
1025: */
1026: if (pp->pr_hardlimit_warning != NULL &&
1027: ratecheck(&pp->pr_hardlimit_warning_last,
1028: &pp->pr_hardlimit_ratecap))
1029: log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
1.21 thorpej 1030:
1031: pp->pr_nfail++;
1032:
1.25 thorpej 1033: pr_leave(pp);
1.134 ad 1034: mutex_exit(&pp->pr_lock);
1.20 thorpej 1035: return (NULL);
1036: }
1037:
1.3 pk 1038: /*
1039: * The convention we use is that if `curpage' is not NULL, then
1040: * it points at a non-empty bucket. In particular, `curpage'
1041: * never points at a page header which has PR_PHINPAGE set and
1042: * has no items in its bucket.
1043: */
1.20 thorpej 1044: if ((ph = pp->pr_curpage) == NULL) {
1.113 yamt 1045: int error;
1046:
1.20 thorpej 1047: #ifdef DIAGNOSTIC
1048: if (pp->pr_nitems != 0) {
1.134 ad 1049: mutex_exit(&pp->pr_lock);
1.20 thorpej 1050: printf("pool_get: %s: curpage NULL, nitems %u\n",
1051: pp->pr_wchan, pp->pr_nitems);
1.80 provos 1052: panic("pool_get: nitems inconsistent");
1.20 thorpej 1053: }
1054: #endif
1055:
1.21 thorpej 1056: /*
1057: * Call the back-end page allocator for more memory.
1058: * Release the pool lock, as the back-end page allocator
1059: * may block.
1060: */
1.25 thorpej 1061: pr_leave(pp);
1.113 yamt 1062: error = pool_grow(pp, flags);
1063: pr_enter(pp, file, line);
1064: if (error != 0) {
1.21 thorpej 1065: /*
1.55 thorpej 1066: * We were unable to allocate a page or item
1067: * header, but we released the lock during
1068: * allocation, so perhaps items were freed
1069: * back to the pool. Check for this case.
1.21 thorpej 1070: */
1071: if (pp->pr_curpage != NULL)
1072: goto startover;
1.15 pk 1073:
1.117 yamt 1074: pp->pr_nfail++;
1.25 thorpej 1075: pr_leave(pp);
1.134 ad 1076: mutex_exit(&pp->pr_lock);
1.117 yamt 1077: return (NULL);
1.1 pk 1078: }
1.3 pk 1079:
1.20 thorpej 1080: /* Start the allocation process over. */
1081: goto startover;
1.3 pk 1082: }
1.97 yamt 1083: if (pp->pr_roflags & PR_NOTOUCH) {
1084: #ifdef DIAGNOSTIC
1085: if (__predict_false(ph->ph_nmissing == pp->pr_itemsperpage)) {
1086: pr_leave(pp);
1.134 ad 1087: mutex_exit(&pp->pr_lock);
1.97 yamt 1088: panic("pool_get: %s: page empty", pp->pr_wchan);
1089: }
1090: #endif
1091: v = pr_item_notouch_get(pp, ph);
1092: #ifdef POOL_DIAGNOSTIC
1093: pr_log(pp, v, PRLOG_GET, file, line);
1094: #endif
1095: } else {
1.102 chs 1096: v = pi = LIST_FIRST(&ph->ph_itemlist);
1.97 yamt 1097: if (__predict_false(v == NULL)) {
1098: pr_leave(pp);
1.134 ad 1099: mutex_exit(&pp->pr_lock);
1.97 yamt 1100: panic("pool_get: %s: page empty", pp->pr_wchan);
1101: }
1.20 thorpej 1102: #ifdef DIAGNOSTIC
1.97 yamt 1103: if (__predict_false(pp->pr_nitems == 0)) {
1104: pr_leave(pp);
1.134 ad 1105: mutex_exit(&pp->pr_lock);
1.97 yamt 1106: printf("pool_get: %s: items on itemlist, nitems %u\n",
1107: pp->pr_wchan, pp->pr_nitems);
1108: panic("pool_get: nitems inconsistent");
1109: }
1.65 enami 1110: #endif
1.56 sommerfe 1111:
1.65 enami 1112: #ifdef POOL_DIAGNOSTIC
1.97 yamt 1113: pr_log(pp, v, PRLOG_GET, file, line);
1.65 enami 1114: #endif
1.3 pk 1115:
1.65 enami 1116: #ifdef DIAGNOSTIC
1.97 yamt 1117: if (__predict_false(pi->pi_magic != PI_MAGIC)) {
1118: pr_printlog(pp, pi, printf);
1119: panic("pool_get(%s): free list modified: "
1120: "magic=%x; page %p; item addr %p\n",
1121: pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
1122: }
1.3 pk 1123: #endif
1124:
1.97 yamt 1125: /*
1126: * Remove from item list.
1127: */
1.102 chs 1128: LIST_REMOVE(pi, pi_list);
1.97 yamt 1129: }
1.20 thorpej 1130: pp->pr_nitems--;
1131: pp->pr_nout++;
1.6 thorpej 1132: if (ph->ph_nmissing == 0) {
1133: #ifdef DIAGNOSTIC
1.34 thorpej 1134: if (__predict_false(pp->pr_nidle == 0))
1.6 thorpej 1135: panic("pool_get: nidle inconsistent");
1136: #endif
1137: pp->pr_nidle--;
1.88 chs 1138:
1139: /*
1140: * This page was previously empty. Move it to the list of
1141: * partially-full pages. This page is already curpage.
1142: */
1143: LIST_REMOVE(ph, ph_pagelist);
1144: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.6 thorpej 1145: }
1.3 pk 1146: ph->ph_nmissing++;
1.97 yamt 1147: if (ph->ph_nmissing == pp->pr_itemsperpage) {
1.21 thorpej 1148: #ifdef DIAGNOSTIC
1.97 yamt 1149: if (__predict_false((pp->pr_roflags & PR_NOTOUCH) == 0 &&
1.102 chs 1150: !LIST_EMPTY(&ph->ph_itemlist))) {
1.25 thorpej 1151: pr_leave(pp);
1.134 ad 1152: mutex_exit(&pp->pr_lock);
1.21 thorpej 1153: panic("pool_get: %s: nmissing inconsistent",
1154: pp->pr_wchan);
1155: }
1156: #endif
1.3 pk 1157: /*
1.88 chs 1158: * This page is now full. Move it to the full list
1159: * and select a new current page.
1.3 pk 1160: */
1.88 chs 1161: LIST_REMOVE(ph, ph_pagelist);
1162: LIST_INSERT_HEAD(&pp->pr_fullpages, ph, ph_pagelist);
1163: pool_update_curpage(pp);
1.1 pk 1164: }
1.3 pk 1165:
1166: pp->pr_nget++;
1.111 christos 1167: pr_leave(pp);
1.20 thorpej 1168:
1169: /*
1170: * If we have a low water mark and we are now below that low
1171: * water mark, add more items to the pool.
1172: */
1.53 thorpej 1173: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1174: /*
1175: * XXX: Should we log a warning? Should we set up a timeout
1176: * to try again in a second or so? The latter could break
1177: * a caller's assumptions about interrupt protection, etc.
1178: */
1179: }
1180:
1.134 ad 1181: mutex_exit(&pp->pr_lock);
1.125 ad 1182: KASSERT((((vaddr_t)v + pp->pr_itemoffset) & (pp->pr_align - 1)) == 0);
1183: FREECHECK_OUT(&pp->pr_freecheck, v);
1.1 pk 1184: return (v);
1185: }
1186:
1187: /*
1.43 thorpej 1188: * Internal version of pool_put(). Pool is already locked/entered.
1.1 pk 1189: */
1.43 thorpej 1190: static void
1.101 thorpej 1191: pool_do_put(struct pool *pp, void *v, struct pool_pagelist *pq)
1.1 pk 1192: {
1193: struct pool_item *pi = v;
1.3 pk 1194: struct pool_item_header *ph;
1195:
1.134 ad 1196: KASSERT(mutex_owned(&pp->pr_lock));
1.125 ad 1197: FREECHECK_IN(&pp->pr_freecheck, v);
1.134 ad 1198: LOCKDEBUG_MEM_CHECK(v, pp->pr_size);
1.61 chs 1199:
1.30 thorpej 1200: #ifdef DIAGNOSTIC
1.34 thorpej 1201: if (__predict_false(pp->pr_nout == 0)) {
1.30 thorpej 1202: printf("pool %s: putting with none out\n",
1203: pp->pr_wchan);
1204: panic("pool_put");
1205: }
1206: #endif
1.3 pk 1207:
1.121 yamt 1208: if (__predict_false((ph = pr_find_pagehead(pp, v)) == NULL)) {
1.25 thorpej 1209: pr_printlog(pp, NULL, printf);
1.3 pk 1210: panic("pool_put: %s: page header missing", pp->pr_wchan);
1211: }
1.28 thorpej 1212:
1.3 pk 1213: /*
1214: * Return to item list.
1215: */
1.97 yamt 1216: if (pp->pr_roflags & PR_NOTOUCH) {
1217: pr_item_notouch_put(pp, ph, v);
1218: } else {
1.2 pk 1219: #ifdef DIAGNOSTIC
1.97 yamt 1220: pi->pi_magic = PI_MAGIC;
1.3 pk 1221: #endif
1.32 chs 1222: #ifdef DEBUG
1.97 yamt 1223: {
1224: int i, *ip = v;
1.32 chs 1225:
1.97 yamt 1226: for (i = 0; i < pp->pr_size / sizeof(int); i++) {
1227: *ip++ = PI_MAGIC;
1228: }
1.32 chs 1229: }
1230: #endif
1231:
1.102 chs 1232: LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.97 yamt 1233: }
1.79 thorpej 1234: KDASSERT(ph->ph_nmissing != 0);
1.3 pk 1235: ph->ph_nmissing--;
1236: pp->pr_nput++;
1.20 thorpej 1237: pp->pr_nitems++;
1238: pp->pr_nout--;
1.3 pk 1239:
1240: /* Cancel "pool empty" condition if it exists */
1241: if (pp->pr_curpage == NULL)
1242: pp->pr_curpage = ph;
1243:
1244: if (pp->pr_flags & PR_WANTED) {
1245: pp->pr_flags &= ~PR_WANTED;
1.15 pk 1246: if (ph->ph_nmissing == 0)
1247: pp->pr_nidle++;
1.134 ad 1248: cv_broadcast(&pp->pr_cv);
1.3 pk 1249: return;
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 &&
1268: (pp->pr_npages > pp->pr_maxpages ||
1.117 yamt 1269: pa_starved_p(pp->pr_alloc))) {
1.101 thorpej 1270: pr_rmpage(pp, ph, pq);
1.3 pk 1271: } else {
1.88 chs 1272: LIST_REMOVE(ph, ph_pagelist);
1273: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.3 pk 1274:
1.21 thorpej 1275: /*
1276: * Update the timestamp on the page. A page must
1277: * be idle for some period of time before it can
1278: * be reclaimed by the pagedaemon. This minimizes
1279: * ping-pong'ing for memory.
1280: */
1.118 kardel 1281: getmicrotime(&ph->ph_time);
1.1 pk 1282: }
1.88 chs 1283: pool_update_curpage(pp);
1.1 pk 1284: }
1.88 chs 1285:
1.21 thorpej 1286: /*
1.88 chs 1287: * If the page was previously completely full, move it to the
1288: * partially-full list and make it the current page. The next
1289: * allocation will get the item from this page, instead of
1290: * further fragmenting the pool.
1.21 thorpej 1291: */
1292: else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
1.88 chs 1293: LIST_REMOVE(ph, ph_pagelist);
1294: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.21 thorpej 1295: pp->pr_curpage = ph;
1296: }
1.43 thorpej 1297: }
1298:
1299: /*
1.134 ad 1300: * Return resource to the pool.
1.43 thorpej 1301: */
1.59 thorpej 1302: #ifdef POOL_DIAGNOSTIC
1.43 thorpej 1303: void
1304: _pool_put(struct pool *pp, void *v, const char *file, long line)
1305: {
1.101 thorpej 1306: struct pool_pagelist pq;
1307:
1308: LIST_INIT(&pq);
1.43 thorpej 1309:
1.134 ad 1310: mutex_enter(&pp->pr_lock);
1.43 thorpej 1311: pr_enter(pp, file, line);
1312:
1.56 sommerfe 1313: pr_log(pp, v, PRLOG_PUT, file, line);
1314:
1.101 thorpej 1315: pool_do_put(pp, v, &pq);
1.21 thorpej 1316:
1.25 thorpej 1317: pr_leave(pp);
1.134 ad 1318: mutex_exit(&pp->pr_lock);
1.101 thorpej 1319:
1.102 chs 1320: pr_pagelist_free(pp, &pq);
1.1 pk 1321: }
1.57 sommerfe 1322: #undef pool_put
1.59 thorpej 1323: #endif /* POOL_DIAGNOSTIC */
1.1 pk 1324:
1.56 sommerfe 1325: void
1326: pool_put(struct pool *pp, void *v)
1327: {
1.101 thorpej 1328: struct pool_pagelist pq;
1329:
1330: LIST_INIT(&pq);
1.56 sommerfe 1331:
1.134 ad 1332: mutex_enter(&pp->pr_lock);
1.101 thorpej 1333: pool_do_put(pp, v, &pq);
1.134 ad 1334: mutex_exit(&pp->pr_lock);
1.56 sommerfe 1335:
1.102 chs 1336: pr_pagelist_free(pp, &pq);
1.56 sommerfe 1337: }
1.57 sommerfe 1338:
1.59 thorpej 1339: #ifdef POOL_DIAGNOSTIC
1.57 sommerfe 1340: #define pool_put(h, v) _pool_put((h), (v), __FILE__, __LINE__)
1.56 sommerfe 1341: #endif
1.74 thorpej 1342:
1343: /*
1.113 yamt 1344: * pool_grow: grow a pool by a page.
1345: *
1346: * => called with pool locked.
1347: * => unlock and relock the pool.
1348: * => return with pool locked.
1349: */
1350:
1351: static int
1352: pool_grow(struct pool *pp, int flags)
1353: {
1354: struct pool_item_header *ph = NULL;
1355: char *cp;
1356:
1.134 ad 1357: mutex_exit(&pp->pr_lock);
1.113 yamt 1358: cp = pool_allocator_alloc(pp, flags);
1359: if (__predict_true(cp != NULL)) {
1360: ph = pool_alloc_item_header(pp, cp, flags);
1361: }
1362: if (__predict_false(cp == NULL || ph == NULL)) {
1363: if (cp != NULL) {
1364: pool_allocator_free(pp, cp);
1365: }
1.134 ad 1366: mutex_enter(&pp->pr_lock);
1.113 yamt 1367: return ENOMEM;
1368: }
1369:
1.134 ad 1370: mutex_enter(&pp->pr_lock);
1.113 yamt 1371: pool_prime_page(pp, cp, ph);
1372: pp->pr_npagealloc++;
1373: return 0;
1374: }
1375:
1376: /*
1.74 thorpej 1377: * Add N items to the pool.
1378: */
1379: int
1380: pool_prime(struct pool *pp, int n)
1381: {
1.75 simonb 1382: int newpages;
1.113 yamt 1383: int error = 0;
1.74 thorpej 1384:
1.134 ad 1385: mutex_enter(&pp->pr_lock);
1.74 thorpej 1386:
1387: newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1388:
1389: while (newpages-- > 0) {
1.113 yamt 1390: error = pool_grow(pp, PR_NOWAIT);
1391: if (error) {
1.74 thorpej 1392: break;
1393: }
1394: pp->pr_minpages++;
1395: }
1396:
1397: if (pp->pr_minpages >= pp->pr_maxpages)
1398: pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */
1399:
1.134 ad 1400: mutex_exit(&pp->pr_lock);
1.113 yamt 1401: return error;
1.74 thorpej 1402: }
1.55 thorpej 1403:
1404: /*
1.3 pk 1405: * Add a page worth of items to the pool.
1.21 thorpej 1406: *
1407: * Note, we must be called with the pool descriptor LOCKED.
1.3 pk 1408: */
1.55 thorpej 1409: static void
1.128 christos 1410: pool_prime_page(struct pool *pp, void *storage, struct pool_item_header *ph)
1.3 pk 1411: {
1412: struct pool_item *pi;
1.128 christos 1413: void *cp = storage;
1.125 ad 1414: const unsigned int align = pp->pr_align;
1415: const unsigned int ioff = pp->pr_itemoffset;
1.55 thorpej 1416: int n;
1.36 pk 1417:
1.134 ad 1418: KASSERT(mutex_owned(&pp->pr_lock));
1.91 yamt 1419:
1.66 thorpej 1420: #ifdef DIAGNOSTIC
1.121 yamt 1421: if ((pp->pr_roflags & PR_NOALIGN) == 0 &&
1422: ((uintptr_t)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
1.36 pk 1423: panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1.66 thorpej 1424: #endif
1.3 pk 1425:
1426: /*
1427: * Insert page header.
1428: */
1.88 chs 1429: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.102 chs 1430: LIST_INIT(&ph->ph_itemlist);
1.3 pk 1431: ph->ph_page = storage;
1432: ph->ph_nmissing = 0;
1.118 kardel 1433: getmicrotime(&ph->ph_time);
1.88 chs 1434: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1435: SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
1.3 pk 1436:
1.6 thorpej 1437: pp->pr_nidle++;
1438:
1.3 pk 1439: /*
1440: * Color this page.
1441: */
1.128 christos 1442: cp = (char *)cp + pp->pr_curcolor;
1.3 pk 1443: if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
1444: pp->pr_curcolor = 0;
1445:
1446: /*
1447: * Adjust storage to apply aligment to `pr_itemoffset' in each item.
1448: */
1449: if (ioff != 0)
1.128 christos 1450: cp = (char *)cp + align - ioff;
1.3 pk 1451:
1.125 ad 1452: KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
1453:
1.3 pk 1454: /*
1455: * Insert remaining chunks on the bucket list.
1456: */
1457: n = pp->pr_itemsperpage;
1.20 thorpej 1458: pp->pr_nitems += n;
1.3 pk 1459:
1.97 yamt 1460: if (pp->pr_roflags & PR_NOTOUCH) {
1.135 yamt 1461: pr_item_notouch_init(pp, ph);
1.97 yamt 1462: } else {
1463: while (n--) {
1464: pi = (struct pool_item *)cp;
1.78 thorpej 1465:
1.97 yamt 1466: KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
1.3 pk 1467:
1.97 yamt 1468: /* Insert on page list */
1.102 chs 1469: LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.3 pk 1470: #ifdef DIAGNOSTIC
1.97 yamt 1471: pi->pi_magic = PI_MAGIC;
1.3 pk 1472: #endif
1.128 christos 1473: cp = (char *)cp + pp->pr_size;
1.125 ad 1474:
1475: KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
1.97 yamt 1476: }
1.3 pk 1477: }
1478:
1479: /*
1480: * If the pool was depleted, point at the new page.
1481: */
1482: if (pp->pr_curpage == NULL)
1483: pp->pr_curpage = ph;
1484:
1485: if (++pp->pr_npages > pp->pr_hiwat)
1486: pp->pr_hiwat = pp->pr_npages;
1487: }
1488:
1.20 thorpej 1489: /*
1.52 thorpej 1490: * Used by pool_get() when nitems drops below the low water mark. This
1.88 chs 1491: * is used to catch up pr_nitems with the low water mark.
1.20 thorpej 1492: *
1.21 thorpej 1493: * Note 1, we never wait for memory here, we let the caller decide what to do.
1.20 thorpej 1494: *
1.73 thorpej 1495: * Note 2, we must be called with the pool already locked, and we return
1.20 thorpej 1496: * with it locked.
1497: */
1498: static int
1.42 thorpej 1499: pool_catchup(struct pool *pp)
1.20 thorpej 1500: {
1501: int error = 0;
1502:
1.54 thorpej 1503: while (POOL_NEEDS_CATCHUP(pp)) {
1.113 yamt 1504: error = pool_grow(pp, PR_NOWAIT);
1505: if (error) {
1.20 thorpej 1506: break;
1507: }
1508: }
1.113 yamt 1509: return error;
1.20 thorpej 1510: }
1511:
1.88 chs 1512: static void
1513: pool_update_curpage(struct pool *pp)
1514: {
1515:
1516: pp->pr_curpage = LIST_FIRST(&pp->pr_partpages);
1517: if (pp->pr_curpage == NULL) {
1518: pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages);
1519: }
1520: }
1521:
1.3 pk 1522: void
1.42 thorpej 1523: pool_setlowat(struct pool *pp, int n)
1.3 pk 1524: {
1.15 pk 1525:
1.134 ad 1526: mutex_enter(&pp->pr_lock);
1.21 thorpej 1527:
1.3 pk 1528: pp->pr_minitems = n;
1.15 pk 1529: pp->pr_minpages = (n == 0)
1530: ? 0
1.18 thorpej 1531: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.20 thorpej 1532:
1533: /* Make sure we're caught up with the newly-set low water mark. */
1.75 simonb 1534: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1535: /*
1536: * XXX: Should we log a warning? Should we set up a timeout
1537: * to try again in a second or so? The latter could break
1538: * a caller's assumptions about interrupt protection, etc.
1539: */
1540: }
1.21 thorpej 1541:
1.134 ad 1542: mutex_exit(&pp->pr_lock);
1.3 pk 1543: }
1544:
1545: void
1.42 thorpej 1546: pool_sethiwat(struct pool *pp, int n)
1.3 pk 1547: {
1.15 pk 1548:
1.134 ad 1549: mutex_enter(&pp->pr_lock);
1.21 thorpej 1550:
1.15 pk 1551: pp->pr_maxpages = (n == 0)
1552: ? 0
1.18 thorpej 1553: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1554:
1.134 ad 1555: mutex_exit(&pp->pr_lock);
1.3 pk 1556: }
1557:
1.20 thorpej 1558: void
1.42 thorpej 1559: pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1.20 thorpej 1560: {
1561:
1.134 ad 1562: mutex_enter(&pp->pr_lock);
1.20 thorpej 1563:
1564: pp->pr_hardlimit = n;
1565: pp->pr_hardlimit_warning = warnmess;
1.31 thorpej 1566: pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1567: pp->pr_hardlimit_warning_last.tv_sec = 0;
1568: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.20 thorpej 1569:
1570: /*
1.21 thorpej 1571: * In-line version of pool_sethiwat(), because we don't want to
1572: * release the lock.
1.20 thorpej 1573: */
1574: pp->pr_maxpages = (n == 0)
1575: ? 0
1576: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1577:
1.134 ad 1578: mutex_exit(&pp->pr_lock);
1.20 thorpej 1579: }
1.3 pk 1580:
1581: /*
1582: * Release all complete pages that have not been used recently.
1583: */
1.66 thorpej 1584: int
1.59 thorpej 1585: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 1586: _pool_reclaim(struct pool *pp, const char *file, long line)
1.56 sommerfe 1587: #else
1588: pool_reclaim(struct pool *pp)
1589: #endif
1.3 pk 1590: {
1591: struct pool_item_header *ph, *phnext;
1.61 chs 1592: struct pool_pagelist pq;
1.102 chs 1593: struct timeval curtime, diff;
1.134 ad 1594: bool klock;
1595: int rv;
1.3 pk 1596:
1.68 thorpej 1597: if (pp->pr_drain_hook != NULL) {
1598: /*
1599: * The drain hook must be called with the pool unlocked.
1600: */
1601: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT);
1602: }
1603:
1.134 ad 1604: /*
1605: * XXXSMP Because mutexes at IPL_SOFTXXX are still spinlocks,
1606: * and we are called from the pagedaemon without kernel_lock.
1607: * Does not apply to IPL_SOFTBIO.
1608: */
1609: if (pp->pr_ipl == IPL_SOFTNET || pp->pr_ipl == IPL_SOFTCLOCK ||
1610: pp->pr_ipl == IPL_SOFTSERIAL) {
1611: KERNEL_LOCK(1, NULL);
1612: klock = true;
1613: } else
1614: klock = false;
1615:
1616: /* Reclaim items from the pool's cache (if any). */
1617: if (pp->pr_cache != NULL)
1618: pool_cache_invalidate(pp->pr_cache);
1619:
1620: if (mutex_tryenter(&pp->pr_lock) == 0) {
1621: if (klock) {
1622: KERNEL_UNLOCK_ONE(NULL);
1623: }
1.66 thorpej 1624: return (0);
1.134 ad 1625: }
1.25 thorpej 1626: pr_enter(pp, file, line);
1.68 thorpej 1627:
1.88 chs 1628: LIST_INIT(&pq);
1.43 thorpej 1629:
1.118 kardel 1630: getmicrotime(&curtime);
1.21 thorpej 1631:
1.88 chs 1632: for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
1633: phnext = LIST_NEXT(ph, ph_pagelist);
1.3 pk 1634:
1635: /* Check our minimum page claim */
1636: if (pp->pr_npages <= pp->pr_minpages)
1637: break;
1638:
1.88 chs 1639: KASSERT(ph->ph_nmissing == 0);
1640: timersub(&curtime, &ph->ph_time, &diff);
1.117 yamt 1641: if (diff.tv_sec < pool_inactive_time
1642: && !pa_starved_p(pp->pr_alloc))
1.88 chs 1643: continue;
1.21 thorpej 1644:
1.88 chs 1645: /*
1646: * If freeing this page would put us below
1647: * the low water mark, stop now.
1648: */
1649: if ((pp->pr_nitems - pp->pr_itemsperpage) <
1650: pp->pr_minitems)
1651: break;
1.21 thorpej 1652:
1.88 chs 1653: pr_rmpage(pp, ph, &pq);
1.3 pk 1654: }
1655:
1.25 thorpej 1656: pr_leave(pp);
1.134 ad 1657: mutex_exit(&pp->pr_lock);
1658:
1659: if (LIST_EMPTY(&pq))
1660: rv = 0;
1661: else {
1662: pr_pagelist_free(pp, &pq);
1663: rv = 1;
1664: }
1665:
1666: if (klock) {
1667: KERNEL_UNLOCK_ONE(NULL);
1668: }
1.66 thorpej 1669:
1.134 ad 1670: return (rv);
1.3 pk 1671: }
1672:
1673: /*
1.134 ad 1674: * Drain pools, one at a time. This is a two stage process;
1675: * drain_start kicks off a cross call to drain CPU-level caches
1676: * if the pool has an associated pool_cache. drain_end waits
1677: * for those cross calls to finish, and then drains the cache
1678: * (if any) and pool.
1.131 ad 1679: *
1.134 ad 1680: * Note, must never be called from interrupt context.
1.3 pk 1681: */
1682: void
1.134 ad 1683: pool_drain_start(struct pool **ppp, uint64_t *wp)
1.3 pk 1684: {
1685: struct pool *pp;
1.134 ad 1686:
1687: KASSERT(!LIST_EMPTY(&pool_head));
1.3 pk 1688:
1.61 chs 1689: pp = NULL;
1.134 ad 1690:
1691: /* Find next pool to drain, and add a reference. */
1692: mutex_enter(&pool_head_lock);
1693: do {
1694: if (drainpp == NULL) {
1695: drainpp = LIST_FIRST(&pool_head);
1696: }
1697: if (drainpp != NULL) {
1698: pp = drainpp;
1699: drainpp = LIST_NEXT(pp, pr_poollist);
1700: }
1701: /*
1702: * Skip completely idle pools. We depend on at least
1703: * one pool in the system being active.
1704: */
1705: } while (pp == NULL || pp->pr_npages == 0);
1706: pp->pr_refcnt++;
1707: mutex_exit(&pool_head_lock);
1708:
1709: /* If there is a pool_cache, drain CPU level caches. */
1710: *ppp = pp;
1711: if (pp->pr_cache != NULL) {
1712: *wp = xc_broadcast(0, (xcfunc_t)pool_cache_xcall,
1713: pp->pr_cache, NULL);
1714: }
1715: }
1716:
1717: void
1718: pool_drain_end(struct pool *pp, uint64_t where)
1719: {
1720:
1721: if (pp == NULL)
1722: return;
1723:
1724: KASSERT(pp->pr_refcnt > 0);
1725:
1726: /* Wait for remote draining to complete. */
1727: if (pp->pr_cache != NULL)
1728: xc_wait(where);
1729:
1730: /* Drain the cache (if any) and pool.. */
1731: pool_reclaim(pp);
1732:
1733: /* Finally, unlock the pool. */
1734: mutex_enter(&pool_head_lock);
1735: pp->pr_refcnt--;
1736: cv_broadcast(&pool_busy);
1737: mutex_exit(&pool_head_lock);
1.3 pk 1738: }
1739:
1740: /*
1741: * Diagnostic helpers.
1742: */
1743: void
1.42 thorpej 1744: pool_print(struct pool *pp, const char *modif)
1.21 thorpej 1745: {
1746:
1.25 thorpej 1747: pool_print1(pp, modif, printf);
1.21 thorpej 1748: }
1749:
1.25 thorpej 1750: void
1.108 yamt 1751: pool_printall(const char *modif, void (*pr)(const char *, ...))
1752: {
1753: struct pool *pp;
1754:
1755: LIST_FOREACH(pp, &pool_head, pr_poollist) {
1756: pool_printit(pp, modif, pr);
1757: }
1758: }
1759:
1760: void
1.42 thorpej 1761: pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.25 thorpej 1762: {
1763:
1764: if (pp == NULL) {
1765: (*pr)("Must specify a pool to print.\n");
1766: return;
1767: }
1768:
1769: pool_print1(pp, modif, pr);
1770: }
1771:
1.21 thorpej 1772: static void
1.124 yamt 1773: pool_print_pagelist(struct pool *pp, struct pool_pagelist *pl,
1.97 yamt 1774: void (*pr)(const char *, ...))
1.88 chs 1775: {
1776: struct pool_item_header *ph;
1777: #ifdef DIAGNOSTIC
1778: struct pool_item *pi;
1779: #endif
1780:
1781: LIST_FOREACH(ph, pl, ph_pagelist) {
1782: (*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n",
1783: ph->ph_page, ph->ph_nmissing,
1784: (u_long)ph->ph_time.tv_sec,
1785: (u_long)ph->ph_time.tv_usec);
1786: #ifdef DIAGNOSTIC
1.97 yamt 1787: if (!(pp->pr_roflags & PR_NOTOUCH)) {
1.102 chs 1788: LIST_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1.97 yamt 1789: if (pi->pi_magic != PI_MAGIC) {
1790: (*pr)("\t\t\titem %p, magic 0x%x\n",
1791: pi, pi->pi_magic);
1792: }
1.88 chs 1793: }
1794: }
1795: #endif
1796: }
1797: }
1798:
1799: static void
1.42 thorpej 1800: pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.3 pk 1801: {
1.25 thorpej 1802: struct pool_item_header *ph;
1.134 ad 1803: pool_cache_t pc;
1804: pcg_t *pcg;
1805: pool_cache_cpu_t *cc;
1806: uint64_t cpuhit, cpumiss;
1.44 thorpej 1807: int i, print_log = 0, print_pagelist = 0, print_cache = 0;
1.25 thorpej 1808: char c;
1809:
1810: while ((c = *modif++) != '\0') {
1811: if (c == 'l')
1812: print_log = 1;
1813: if (c == 'p')
1814: print_pagelist = 1;
1.44 thorpej 1815: if (c == 'c')
1816: print_cache = 1;
1.25 thorpej 1817: }
1818:
1.134 ad 1819: if ((pc = pp->pr_cache) != NULL) {
1820: (*pr)("POOL CACHE");
1821: } else {
1822: (*pr)("POOL");
1823: }
1824:
1825: (*pr)(" %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1.25 thorpej 1826: pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1827: pp->pr_roflags);
1.66 thorpej 1828: (*pr)("\talloc %p\n", pp->pr_alloc);
1.25 thorpej 1829: (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1830: pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1831: (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1832: pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1833:
1.134 ad 1834: (*pr)("\tnget %lu, nfail %lu, nput %lu\n",
1.25 thorpej 1835: pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1836: (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1837: pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1838:
1839: if (print_pagelist == 0)
1840: goto skip_pagelist;
1841:
1.88 chs 1842: if ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1843: (*pr)("\n\tempty page list:\n");
1.97 yamt 1844: pool_print_pagelist(pp, &pp->pr_emptypages, pr);
1.88 chs 1845: if ((ph = LIST_FIRST(&pp->pr_fullpages)) != NULL)
1846: (*pr)("\n\tfull page list:\n");
1.97 yamt 1847: pool_print_pagelist(pp, &pp->pr_fullpages, pr);
1.88 chs 1848: if ((ph = LIST_FIRST(&pp->pr_partpages)) != NULL)
1849: (*pr)("\n\tpartial-page list:\n");
1.97 yamt 1850: pool_print_pagelist(pp, &pp->pr_partpages, pr);
1.88 chs 1851:
1.25 thorpej 1852: if (pp->pr_curpage == NULL)
1853: (*pr)("\tno current page\n");
1854: else
1855: (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1856:
1857: skip_pagelist:
1858: if (print_log == 0)
1859: goto skip_log;
1860:
1861: (*pr)("\n");
1862: if ((pp->pr_roflags & PR_LOGGING) == 0)
1863: (*pr)("\tno log\n");
1.122 christos 1864: else {
1.25 thorpej 1865: pr_printlog(pp, NULL, pr);
1.122 christos 1866: }
1.3 pk 1867:
1.25 thorpej 1868: skip_log:
1.44 thorpej 1869:
1.102 chs 1870: #define PR_GROUPLIST(pcg) \
1871: (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail); \
1872: for (i = 0; i < PCG_NOBJECTS; i++) { \
1873: if (pcg->pcg_objects[i].pcgo_pa != \
1874: POOL_PADDR_INVALID) { \
1875: (*pr)("\t\t\t%p, 0x%llx\n", \
1876: pcg->pcg_objects[i].pcgo_va, \
1877: (unsigned long long) \
1878: pcg->pcg_objects[i].pcgo_pa); \
1879: } else { \
1880: (*pr)("\t\t\t%p\n", \
1881: pcg->pcg_objects[i].pcgo_va); \
1882: } \
1883: }
1884:
1.134 ad 1885: if (pc != NULL) {
1886: cpuhit = 0;
1887: cpumiss = 0;
1888: for (i = 0; i < MAXCPUS; i++) {
1889: if ((cc = pc->pc_cpus[i]) == NULL)
1890: continue;
1891: cpuhit += cc->cc_hits;
1892: cpumiss += cc->cc_misses;
1893: }
1894: (*pr)("\tcpu layer hits %llu misses %llu\n", cpuhit, cpumiss);
1895: (*pr)("\tcache layer hits %llu misses %llu\n",
1896: pc->pc_hits, pc->pc_misses);
1897: (*pr)("\tcache layer entry uncontended %llu contended %llu\n",
1898: pc->pc_hits + pc->pc_misses - pc->pc_contended,
1899: pc->pc_contended);
1900: (*pr)("\tcache layer empty groups %u full groups %u\n",
1901: pc->pc_nempty, pc->pc_nfull);
1902: if (print_cache) {
1903: (*pr)("\tfull cache groups:\n");
1904: for (pcg = pc->pc_fullgroups; pcg != NULL;
1905: pcg = pcg->pcg_next) {
1906: PR_GROUPLIST(pcg);
1907: }
1908: (*pr)("\tempty cache groups:\n");
1909: for (pcg = pc->pc_emptygroups; pcg != NULL;
1910: pcg = pcg->pcg_next) {
1911: PR_GROUPLIST(pcg);
1912: }
1.103 chs 1913: }
1.44 thorpej 1914: }
1.102 chs 1915: #undef PR_GROUPLIST
1.44 thorpej 1916:
1.88 chs 1917: pr_enter_check(pp, pr);
1918: }
1919:
1920: static int
1921: pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
1922: {
1923: struct pool_item *pi;
1.128 christos 1924: void *page;
1.88 chs 1925: int n;
1926:
1.121 yamt 1927: if ((pp->pr_roflags & PR_NOALIGN) == 0) {
1.128 christos 1928: page = (void *)((uintptr_t)ph & pp->pr_alloc->pa_pagemask);
1.121 yamt 1929: if (page != ph->ph_page &&
1930: (pp->pr_roflags & PR_PHINPAGE) != 0) {
1931: if (label != NULL)
1932: printf("%s: ", label);
1933: printf("pool(%p:%s): page inconsistency: page %p;"
1934: " at page head addr %p (p %p)\n", pp,
1935: pp->pr_wchan, ph->ph_page,
1936: ph, page);
1937: return 1;
1938: }
1.88 chs 1939: }
1.3 pk 1940:
1.97 yamt 1941: if ((pp->pr_roflags & PR_NOTOUCH) != 0)
1942: return 0;
1943:
1.102 chs 1944: for (pi = LIST_FIRST(&ph->ph_itemlist), n = 0;
1.88 chs 1945: pi != NULL;
1.102 chs 1946: pi = LIST_NEXT(pi,pi_list), n++) {
1.88 chs 1947:
1948: #ifdef DIAGNOSTIC
1949: if (pi->pi_magic != PI_MAGIC) {
1950: if (label != NULL)
1951: printf("%s: ", label);
1952: printf("pool(%s): free list modified: magic=%x;"
1.121 yamt 1953: " page %p; item ordinal %d; addr %p\n",
1.88 chs 1954: pp->pr_wchan, pi->pi_magic, ph->ph_page,
1.121 yamt 1955: n, pi);
1.88 chs 1956: panic("pool");
1957: }
1958: #endif
1.121 yamt 1959: if ((pp->pr_roflags & PR_NOALIGN) != 0) {
1960: continue;
1961: }
1.128 christos 1962: page = (void *)((uintptr_t)pi & pp->pr_alloc->pa_pagemask);
1.88 chs 1963: if (page == ph->ph_page)
1964: continue;
1965:
1966: if (label != NULL)
1967: printf("%s: ", label);
1968: printf("pool(%p:%s): page inconsistency: page %p;"
1969: " item ordinal %d; addr %p (p %p)\n", pp,
1970: pp->pr_wchan, ph->ph_page,
1971: n, pi, page);
1972: return 1;
1973: }
1974: return 0;
1.3 pk 1975: }
1976:
1.88 chs 1977:
1.3 pk 1978: int
1.42 thorpej 1979: pool_chk(struct pool *pp, const char *label)
1.3 pk 1980: {
1981: struct pool_item_header *ph;
1982: int r = 0;
1983:
1.134 ad 1984: mutex_enter(&pp->pr_lock);
1.88 chs 1985: LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
1986: r = pool_chk_page(pp, label, ph);
1987: if (r) {
1988: goto out;
1989: }
1990: }
1991: LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
1992: r = pool_chk_page(pp, label, ph);
1993: if (r) {
1.3 pk 1994: goto out;
1995: }
1.88 chs 1996: }
1997: LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
1998: r = pool_chk_page(pp, label, ph);
1999: if (r) {
1.3 pk 2000: goto out;
2001: }
2002: }
1.88 chs 2003:
1.3 pk 2004: out:
1.134 ad 2005: mutex_exit(&pp->pr_lock);
1.3 pk 2006: return (r);
1.43 thorpej 2007: }
2008:
2009: /*
2010: * pool_cache_init:
2011: *
2012: * Initialize a pool cache.
1.134 ad 2013: */
2014: pool_cache_t
2015: pool_cache_init(size_t size, u_int align, u_int align_offset, u_int flags,
2016: const char *wchan, struct pool_allocator *palloc, int ipl,
2017: int (*ctor)(void *, void *, int), void (*dtor)(void *, void *), void *arg)
2018: {
2019: pool_cache_t pc;
2020:
2021: pc = pool_get(&cache_pool, PR_WAITOK);
2022: if (pc == NULL)
2023: return NULL;
2024:
2025: pool_cache_bootstrap(pc, size, align, align_offset, flags, wchan,
2026: palloc, ipl, ctor, dtor, arg);
2027:
2028: return pc;
2029: }
2030:
2031: /*
2032: * pool_cache_bootstrap:
1.43 thorpej 2033: *
1.134 ad 2034: * Kernel-private version of pool_cache_init(). The caller
2035: * provides initial storage.
1.43 thorpej 2036: */
2037: void
1.134 ad 2038: pool_cache_bootstrap(pool_cache_t pc, size_t size, u_int align,
2039: u_int align_offset, u_int flags, const char *wchan,
2040: struct pool_allocator *palloc, int ipl,
2041: int (*ctor)(void *, void *, int), void (*dtor)(void *, void *),
1.43 thorpej 2042: void *arg)
2043: {
1.134 ad 2044: CPU_INFO_ITERATOR cii;
2045: struct cpu_info *ci;
2046: struct pool *pp;
2047:
2048: pp = &pc->pc_pool;
2049: if (palloc == NULL && ipl == IPL_NONE)
2050: palloc = &pool_allocator_nointr;
2051: pool_init(pp, size, align, align_offset, flags, wchan, palloc, ipl);
1.43 thorpej 2052:
1.137.2.1 ad 2053: /*
2054: * XXXAD hack to prevent IP input processing from blocking.
2055: */
2056: if (ipl == IPL_SOFTNET) {
2057: mutex_init(&pc->pc_lock, MUTEX_DEFAULT, IPL_VM);
2058: } else {
2059: mutex_init(&pc->pc_lock, MUTEX_DEFAULT, ipl);
2060: }
1.43 thorpej 2061:
1.134 ad 2062: if (ctor == NULL) {
2063: ctor = (int (*)(void *, void *, int))nullop;
2064: }
2065: if (dtor == NULL) {
2066: dtor = (void (*)(void *, void *))nullop;
2067: }
1.43 thorpej 2068:
1.134 ad 2069: pc->pc_emptygroups = NULL;
2070: pc->pc_fullgroups = NULL;
2071: pc->pc_partgroups = NULL;
1.43 thorpej 2072: pc->pc_ctor = ctor;
2073: pc->pc_dtor = dtor;
2074: pc->pc_arg = arg;
1.134 ad 2075: pc->pc_hits = 0;
1.48 thorpej 2076: pc->pc_misses = 0;
1.134 ad 2077: pc->pc_nempty = 0;
2078: pc->pc_npart = 0;
2079: pc->pc_nfull = 0;
2080: pc->pc_contended = 0;
2081: pc->pc_refcnt = 0;
1.136 yamt 2082: pc->pc_freecheck = NULL;
1.134 ad 2083:
2084: /* Allocate per-CPU caches. */
2085: memset(pc->pc_cpus, 0, sizeof(pc->pc_cpus));
2086: pc->pc_ncpu = 0;
1.137 ad 2087: if (ncpu == 0) {
2088: /* XXX For sparc: boot CPU is not attached yet. */
2089: pool_cache_cpu_init1(curcpu(), pc);
2090: } else {
2091: for (CPU_INFO_FOREACH(cii, ci)) {
2092: pool_cache_cpu_init1(ci, pc);
2093: }
1.134 ad 2094: }
2095:
2096: if (__predict_true(!cold)) {
2097: mutex_enter(&pp->pr_lock);
2098: pp->pr_cache = pc;
2099: mutex_exit(&pp->pr_lock);
2100: mutex_enter(&pool_head_lock);
2101: LIST_INSERT_HEAD(&pool_cache_head, pc, pc_cachelist);
2102: mutex_exit(&pool_head_lock);
2103: } else {
2104: pp->pr_cache = pc;
2105: LIST_INSERT_HEAD(&pool_cache_head, pc, pc_cachelist);
2106: }
1.43 thorpej 2107: }
2108:
2109: /*
2110: * pool_cache_destroy:
2111: *
2112: * Destroy a pool cache.
2113: */
2114: void
1.134 ad 2115: pool_cache_destroy(pool_cache_t pc)
1.43 thorpej 2116: {
1.134 ad 2117: struct pool *pp = &pc->pc_pool;
2118: pool_cache_cpu_t *cc;
2119: pcg_t *pcg;
2120: int i;
2121:
2122: /* Remove it from the global list. */
2123: mutex_enter(&pool_head_lock);
2124: while (pc->pc_refcnt != 0)
2125: cv_wait(&pool_busy, &pool_head_lock);
2126: LIST_REMOVE(pc, pc_cachelist);
2127: mutex_exit(&pool_head_lock);
1.43 thorpej 2128:
2129: /* First, invalidate the entire cache. */
2130: pool_cache_invalidate(pc);
2131:
1.134 ad 2132: /* Disassociate it from the pool. */
2133: mutex_enter(&pp->pr_lock);
2134: pp->pr_cache = NULL;
2135: mutex_exit(&pp->pr_lock);
2136:
2137: /* Destroy per-CPU data */
2138: for (i = 0; i < MAXCPUS; i++) {
2139: if ((cc = pc->pc_cpus[i]) == NULL)
2140: continue;
2141: if ((pcg = cc->cc_current) != NULL) {
2142: pcg->pcg_next = NULL;
2143: pool_cache_invalidate_groups(pc, pcg);
2144: }
2145: if ((pcg = cc->cc_previous) != NULL) {
2146: pcg->pcg_next = NULL;
2147: pool_cache_invalidate_groups(pc, pcg);
2148: }
2149: if (cc != &pc->pc_cpu0)
2150: pool_put(&cache_cpu_pool, cc);
2151: }
2152:
2153: /* Finally, destroy it. */
2154: mutex_destroy(&pc->pc_lock);
2155: pool_destroy(pp);
2156: pool_put(&cache_pool, pc);
2157: }
2158:
2159: /*
2160: * pool_cache_cpu_init1:
2161: *
2162: * Called for each pool_cache whenever a new CPU is attached.
2163: */
2164: static void
2165: pool_cache_cpu_init1(struct cpu_info *ci, pool_cache_t pc)
2166: {
2167: pool_cache_cpu_t *cc;
1.137 ad 2168: int index;
1.134 ad 2169:
1.137 ad 2170: index = ci->ci_index;
2171:
2172: KASSERT(index < MAXCPUS);
1.134 ad 2173: KASSERT(((uintptr_t)pc->pc_cpus & (CACHE_LINE_SIZE - 1)) == 0);
2174:
1.137 ad 2175: if ((cc = pc->pc_cpus[index]) != NULL) {
2176: KASSERT(cc->cc_cpuindex == index);
1.134 ad 2177: return;
2178: }
2179:
2180: /*
2181: * The first CPU is 'free'. This needs to be the case for
2182: * bootstrap - we may not be able to allocate yet.
2183: */
2184: if (pc->pc_ncpu == 0) {
2185: cc = &pc->pc_cpu0;
2186: pc->pc_ncpu = 1;
2187: } else {
2188: mutex_enter(&pc->pc_lock);
2189: pc->pc_ncpu++;
2190: mutex_exit(&pc->pc_lock);
2191: cc = pool_get(&cache_cpu_pool, PR_WAITOK);
2192: }
2193:
2194: cc->cc_ipl = pc->pc_pool.pr_ipl;
2195: cc->cc_iplcookie = makeiplcookie(cc->cc_ipl);
2196: cc->cc_cache = pc;
1.137 ad 2197: cc->cc_cpuindex = index;
1.134 ad 2198: cc->cc_hits = 0;
2199: cc->cc_misses = 0;
2200: cc->cc_current = NULL;
2201: cc->cc_previous = NULL;
2202:
1.137 ad 2203: pc->pc_cpus[index] = cc;
1.43 thorpej 2204: }
2205:
1.134 ad 2206: /*
2207: * pool_cache_cpu_init:
2208: *
2209: * Called whenever a new CPU is attached.
2210: */
2211: void
2212: pool_cache_cpu_init(struct cpu_info *ci)
1.43 thorpej 2213: {
1.134 ad 2214: pool_cache_t pc;
2215:
2216: mutex_enter(&pool_head_lock);
2217: LIST_FOREACH(pc, &pool_cache_head, pc_cachelist) {
2218: pc->pc_refcnt++;
2219: mutex_exit(&pool_head_lock);
1.43 thorpej 2220:
1.134 ad 2221: pool_cache_cpu_init1(ci, pc);
1.43 thorpej 2222:
1.134 ad 2223: mutex_enter(&pool_head_lock);
2224: pc->pc_refcnt--;
2225: cv_broadcast(&pool_busy);
2226: }
2227: mutex_exit(&pool_head_lock);
1.43 thorpej 2228: }
2229:
1.134 ad 2230: /*
2231: * pool_cache_reclaim:
2232: *
2233: * Reclaim memory from a pool cache.
2234: */
2235: bool
2236: pool_cache_reclaim(pool_cache_t pc)
1.43 thorpej 2237: {
2238:
1.134 ad 2239: return pool_reclaim(&pc->pc_pool);
2240: }
1.43 thorpej 2241:
1.136 yamt 2242: static void
2243: pool_cache_destruct_object1(pool_cache_t pc, void *object)
2244: {
2245:
2246: (*pc->pc_dtor)(pc->pc_arg, object);
2247: pool_put(&pc->pc_pool, object);
2248: }
2249:
1.134 ad 2250: /*
2251: * pool_cache_destruct_object:
2252: *
2253: * Force destruction of an object and its release back into
2254: * the pool.
2255: */
2256: void
2257: pool_cache_destruct_object(pool_cache_t pc, void *object)
2258: {
2259:
1.136 yamt 2260: FREECHECK_IN(&pc->pc_freecheck, object);
2261:
2262: pool_cache_destruct_object1(pc, object);
1.43 thorpej 2263: }
2264:
1.134 ad 2265: /*
2266: * pool_cache_invalidate_groups:
2267: *
2268: * Invalidate a chain of groups and destruct all objects.
2269: */
1.102 chs 2270: static void
1.134 ad 2271: pool_cache_invalidate_groups(pool_cache_t pc, pcg_t *pcg)
1.102 chs 2272: {
1.134 ad 2273: void *object;
2274: pcg_t *next;
2275: int i;
2276:
2277: for (; pcg != NULL; pcg = next) {
2278: next = pcg->pcg_next;
2279:
2280: for (i = 0; i < pcg->pcg_avail; i++) {
2281: object = pcg->pcg_objects[i].pcgo_va;
1.136 yamt 2282: pool_cache_destruct_object1(pc, object);
1.134 ad 2283: }
1.102 chs 2284:
2285: pool_put(&pcgpool, pcg);
2286: }
2287: }
2288:
1.43 thorpej 2289: /*
1.134 ad 2290: * pool_cache_invalidate:
1.43 thorpej 2291: *
1.134 ad 2292: * Invalidate a pool cache (destruct and release all of the
2293: * cached objects). Does not reclaim objects from the pool.
1.43 thorpej 2294: */
1.134 ad 2295: void
2296: pool_cache_invalidate(pool_cache_t pc)
2297: {
2298: pcg_t *full, *empty, *part;
2299:
2300: mutex_enter(&pc->pc_lock);
2301: full = pc->pc_fullgroups;
2302: empty = pc->pc_emptygroups;
2303: part = pc->pc_partgroups;
2304: pc->pc_fullgroups = NULL;
2305: pc->pc_emptygroups = NULL;
2306: pc->pc_partgroups = NULL;
2307: pc->pc_nfull = 0;
2308: pc->pc_nempty = 0;
2309: pc->pc_npart = 0;
2310: mutex_exit(&pc->pc_lock);
2311:
2312: pool_cache_invalidate_groups(pc, full);
2313: pool_cache_invalidate_groups(pc, empty);
2314: pool_cache_invalidate_groups(pc, part);
2315: }
2316:
2317: void
2318: pool_cache_set_drain_hook(pool_cache_t pc, void (*fn)(void *, int), void *arg)
2319: {
2320:
2321: pool_set_drain_hook(&pc->pc_pool, fn, arg);
2322: }
2323:
2324: void
2325: pool_cache_setlowat(pool_cache_t pc, int n)
2326: {
2327:
2328: pool_setlowat(&pc->pc_pool, n);
2329: }
2330:
2331: void
2332: pool_cache_sethiwat(pool_cache_t pc, int n)
2333: {
2334:
2335: pool_sethiwat(&pc->pc_pool, n);
2336: }
2337:
2338: void
2339: pool_cache_sethardlimit(pool_cache_t pc, int n, const char *warnmess, int ratecap)
2340: {
2341:
2342: pool_sethardlimit(&pc->pc_pool, n, warnmess, ratecap);
2343: }
2344:
2345: static inline pool_cache_cpu_t *
2346: pool_cache_cpu_enter(pool_cache_t pc, int *s)
2347: {
2348: pool_cache_cpu_t *cc;
2349:
2350: /*
2351: * Prevent other users of the cache from accessing our
2352: * CPU-local data. To avoid touching shared state, we
2353: * pull the neccessary information from CPU local data.
2354: */
1.137 ad 2355: crit_enter();
2356: cc = pc->pc_cpus[curcpu()->ci_index];
1.134 ad 2357: KASSERT(cc->cc_cache == pc);
1.137 ad 2358: if (cc->cc_ipl != IPL_NONE) {
1.134 ad 2359: *s = splraiseipl(cc->cc_iplcookie);
2360: }
2361: KASSERT(((uintptr_t)cc & (CACHE_LINE_SIZE - 1)) == 0);
2362:
2363: return cc;
2364: }
2365:
2366: static inline void
2367: pool_cache_cpu_exit(pool_cache_cpu_t *cc, int *s)
2368: {
2369:
2370: /* No longer need exclusive access to the per-CPU data. */
1.137 ad 2371: if (cc->cc_ipl != IPL_NONE) {
1.134 ad 2372: splx(*s);
2373: }
1.137 ad 2374: crit_exit();
1.134 ad 2375: }
2376:
2377: #if __GNUC_PREREQ__(3, 0)
2378: __attribute ((noinline))
2379: #endif
2380: pool_cache_cpu_t *
2381: pool_cache_get_slow(pool_cache_cpu_t *cc, int *s, void **objectp,
2382: paddr_t *pap, int flags)
1.43 thorpej 2383: {
1.134 ad 2384: pcg_t *pcg, *cur;
2385: uint64_t ncsw;
2386: pool_cache_t pc;
1.43 thorpej 2387: void *object;
1.58 thorpej 2388:
1.134 ad 2389: pc = cc->cc_cache;
2390: cc->cc_misses++;
1.43 thorpej 2391:
1.134 ad 2392: /*
2393: * Nothing was available locally. Try and grab a group
2394: * from the cache.
2395: */
2396: if (!mutex_tryenter(&pc->pc_lock)) {
2397: ncsw = curlwp->l_ncsw;
2398: mutex_enter(&pc->pc_lock);
2399: pc->pc_contended++;
1.43 thorpej 2400:
1.134 ad 2401: /*
2402: * If we context switched while locking, then
2403: * our view of the per-CPU data is invalid:
2404: * retry.
2405: */
2406: if (curlwp->l_ncsw != ncsw) {
2407: mutex_exit(&pc->pc_lock);
2408: pool_cache_cpu_exit(cc, s);
2409: return pool_cache_cpu_enter(pc, s);
1.43 thorpej 2410: }
1.102 chs 2411: }
1.43 thorpej 2412:
1.134 ad 2413: if ((pcg = pc->pc_fullgroups) != NULL) {
1.43 thorpej 2414: /*
1.134 ad 2415: * If there's a full group, release our empty
2416: * group back to the cache. Install the full
2417: * group as cc_current and return.
1.43 thorpej 2418: */
1.134 ad 2419: if ((cur = cc->cc_current) != NULL) {
2420: KASSERT(cur->pcg_avail == 0);
2421: cur->pcg_next = pc->pc_emptygroups;
2422: pc->pc_emptygroups = cur;
2423: pc->pc_nempty++;
1.87 thorpej 2424: }
1.134 ad 2425: KASSERT(pcg->pcg_avail == PCG_NOBJECTS);
2426: cc->cc_current = pcg;
2427: pc->pc_fullgroups = pcg->pcg_next;
2428: pc->pc_hits++;
2429: pc->pc_nfull--;
2430: mutex_exit(&pc->pc_lock);
2431: return cc;
2432: }
2433:
2434: /*
2435: * Nothing available locally or in cache. Take the slow
2436: * path: fetch a new object from the pool and construct
2437: * it.
2438: */
2439: pc->pc_misses++;
2440: mutex_exit(&pc->pc_lock);
2441: pool_cache_cpu_exit(cc, s);
2442:
2443: object = pool_get(&pc->pc_pool, flags);
2444: *objectp = object;
2445: if (object == NULL)
2446: return NULL;
1.125 ad 2447:
1.134 ad 2448: if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) {
2449: pool_put(&pc->pc_pool, object);
2450: *objectp = NULL;
2451: return NULL;
1.43 thorpej 2452: }
2453:
1.134 ad 2454: KASSERT((((vaddr_t)object + pc->pc_pool.pr_itemoffset) &
2455: (pc->pc_pool.pr_align - 1)) == 0);
1.43 thorpej 2456:
1.134 ad 2457: if (pap != NULL) {
2458: #ifdef POOL_VTOPHYS
2459: *pap = POOL_VTOPHYS(object);
2460: #else
2461: *pap = POOL_PADDR_INVALID;
2462: #endif
1.102 chs 2463: }
1.43 thorpej 2464:
1.125 ad 2465: FREECHECK_OUT(&pc->pc_freecheck, object);
1.134 ad 2466: return NULL;
1.43 thorpej 2467: }
2468:
2469: /*
1.134 ad 2470: * pool_cache_get{,_paddr}:
1.43 thorpej 2471: *
1.134 ad 2472: * Get an object from a pool cache (optionally returning
2473: * the physical address of the object).
1.43 thorpej 2474: */
1.134 ad 2475: void *
2476: pool_cache_get_paddr(pool_cache_t pc, int flags, paddr_t *pap)
1.43 thorpej 2477: {
1.134 ad 2478: pool_cache_cpu_t *cc;
2479: pcg_t *pcg;
2480: void *object;
1.60 thorpej 2481: int s;
1.43 thorpej 2482:
1.134 ad 2483: #ifdef LOCKDEBUG
2484: if (flags & PR_WAITOK)
2485: ASSERT_SLEEPABLE(NULL, "pool_cache_get(PR_WAITOK)");
2486: #endif
1.125 ad 2487:
1.134 ad 2488: cc = pool_cache_cpu_enter(pc, &s);
2489: do {
2490: /* Try and allocate an object from the current group. */
2491: pcg = cc->cc_current;
2492: if (pcg != NULL && pcg->pcg_avail > 0) {
2493: object = pcg->pcg_objects[--pcg->pcg_avail].pcgo_va;
2494: if (pap != NULL)
2495: *pap = pcg->pcg_objects[pcg->pcg_avail].pcgo_pa;
2496: pcg->pcg_objects[pcg->pcg_avail].pcgo_va = NULL;
2497: KASSERT(pcg->pcg_avail <= PCG_NOBJECTS);
2498: KASSERT(object != NULL);
2499: cc->cc_hits++;
2500: pool_cache_cpu_exit(cc, &s);
2501: FREECHECK_OUT(&pc->pc_freecheck, object);
2502: return object;
1.43 thorpej 2503: }
2504:
2505: /*
1.134 ad 2506: * That failed. If the previous group isn't empty, swap
2507: * it with the current group and allocate from there.
1.43 thorpej 2508: */
1.134 ad 2509: pcg = cc->cc_previous;
2510: if (pcg != NULL && pcg->pcg_avail > 0) {
2511: cc->cc_previous = cc->cc_current;
2512: cc->cc_current = pcg;
2513: continue;
1.43 thorpej 2514: }
2515:
1.134 ad 2516: /*
2517: * Can't allocate from either group: try the slow path.
2518: * If get_slow() allocated an object for us, or if
2519: * no more objects are available, it will return NULL.
2520: * Otherwise, we need to retry.
2521: */
2522: cc = pool_cache_get_slow(cc, &s, &object, pap, flags);
2523: } while (cc != NULL);
1.43 thorpej 2524:
1.134 ad 2525: return object;
1.51 thorpej 2526: }
2527:
1.134 ad 2528: #if __GNUC_PREREQ__(3, 0)
2529: __attribute ((noinline))
2530: #endif
2531: pool_cache_cpu_t *
2532: pool_cache_put_slow(pool_cache_cpu_t *cc, int *s, void *object, paddr_t pa)
1.51 thorpej 2533: {
1.134 ad 2534: pcg_t *pcg, *cur;
2535: uint64_t ncsw;
2536: pool_cache_t pc;
1.51 thorpej 2537:
1.134 ad 2538: pc = cc->cc_cache;
2539: cc->cc_misses++;
1.43 thorpej 2540:
1.134 ad 2541: /*
2542: * No free slots locally. Try to grab an empty, unused
2543: * group from the cache.
2544: */
2545: if (!mutex_tryenter(&pc->pc_lock)) {
2546: ncsw = curlwp->l_ncsw;
2547: mutex_enter(&pc->pc_lock);
2548: pc->pc_contended++;
1.102 chs 2549:
1.134 ad 2550: /*
2551: * If we context switched while locking, then
2552: * our view of the per-CPU data is invalid:
2553: * retry.
2554: */
2555: if (curlwp->l_ncsw != ncsw) {
2556: mutex_exit(&pc->pc_lock);
2557: pool_cache_cpu_exit(cc, s);
2558: return pool_cache_cpu_enter(pc, s);
2559: }
2560: }
1.130 ad 2561:
1.134 ad 2562: if ((pcg = pc->pc_emptygroups) != NULL) {
2563: /*
2564: * If there's a empty group, release our full
2565: * group back to the cache. Install the empty
2566: * group as cc_current and return.
2567: */
2568: if ((cur = cc->cc_current) != NULL) {
2569: KASSERT(cur->pcg_avail == PCG_NOBJECTS);
2570: cur->pcg_next = pc->pc_fullgroups;
2571: pc->pc_fullgroups = cur;
2572: pc->pc_nfull++;
1.102 chs 2573: }
1.134 ad 2574: KASSERT(pcg->pcg_avail == 0);
2575: cc->cc_current = pcg;
2576: pc->pc_emptygroups = pcg->pcg_next;
2577: pc->pc_hits++;
2578: pc->pc_nempty--;
2579: mutex_exit(&pc->pc_lock);
2580: return cc;
1.102 chs 2581: }
1.105 christos 2582:
1.134 ad 2583: /*
2584: * Nothing available locally or in cache. Take the
2585: * slow path and try to allocate a new group that we
2586: * can release to.
2587: */
2588: pc->pc_misses++;
2589: mutex_exit(&pc->pc_lock);
2590: pool_cache_cpu_exit(cc, s);
1.105 christos 2591:
1.134 ad 2592: /*
2593: * If we can't allocate a new group, just throw the
2594: * object away.
2595: */
1.137.2.2! ad 2596: if (pool_cache_disable) {
! 2597: pcg = NULL;
! 2598: } else {
! 2599: pcg = pool_get(&pcgpool, PR_NOWAIT);
! 2600: }
1.134 ad 2601: if (pcg == NULL) {
2602: pool_cache_destruct_object(pc, object);
2603: return NULL;
2604: }
2605: #ifdef DIAGNOSTIC
2606: memset(pcg, 0, sizeof(*pcg));
2607: #else
2608: pcg->pcg_avail = 0;
2609: #endif
1.105 christos 2610:
1.134 ad 2611: /*
2612: * Add the empty group to the cache and try again.
2613: */
2614: mutex_enter(&pc->pc_lock);
2615: pcg->pcg_next = pc->pc_emptygroups;
2616: pc->pc_emptygroups = pcg;
2617: pc->pc_nempty++;
2618: mutex_exit(&pc->pc_lock);
1.103 chs 2619:
1.134 ad 2620: return pool_cache_cpu_enter(pc, s);
2621: }
1.102 chs 2622:
1.43 thorpej 2623: /*
1.134 ad 2624: * pool_cache_put{,_paddr}:
1.43 thorpej 2625: *
1.134 ad 2626: * Put an object back to the pool cache (optionally caching the
2627: * physical address of the object).
1.43 thorpej 2628: */
1.101 thorpej 2629: void
1.134 ad 2630: pool_cache_put_paddr(pool_cache_t pc, void *object, paddr_t pa)
1.43 thorpej 2631: {
1.134 ad 2632: pool_cache_cpu_t *cc;
2633: pcg_t *pcg;
2634: int s;
1.101 thorpej 2635:
1.134 ad 2636: FREECHECK_IN(&pc->pc_freecheck, object);
1.101 thorpej 2637:
1.134 ad 2638: cc = pool_cache_cpu_enter(pc, &s);
2639: do {
2640: /* If the current group isn't full, release it there. */
2641: pcg = cc->cc_current;
2642: if (pcg != NULL && pcg->pcg_avail < PCG_NOBJECTS) {
2643: KASSERT(pcg->pcg_objects[pcg->pcg_avail].pcgo_va
2644: == NULL);
2645: pcg->pcg_objects[pcg->pcg_avail].pcgo_va = object;
2646: pcg->pcg_objects[pcg->pcg_avail].pcgo_pa = pa;
2647: pcg->pcg_avail++;
2648: cc->cc_hits++;
2649: pool_cache_cpu_exit(cc, &s);
2650: return;
2651: }
1.43 thorpej 2652:
1.134 ad 2653: /*
2654: * That failed. If the previous group is empty, swap
2655: * it with the current group and try again.
2656: */
2657: pcg = cc->cc_previous;
2658: if (pcg != NULL && pcg->pcg_avail == 0) {
2659: cc->cc_previous = cc->cc_current;
2660: cc->cc_current = pcg;
2661: continue;
2662: }
1.43 thorpej 2663:
1.134 ad 2664: /*
2665: * Can't free to either group: try the slow path.
2666: * If put_slow() releases the object for us, it
2667: * will return NULL. Otherwise we need to retry.
2668: */
2669: cc = pool_cache_put_slow(cc, &s, object, pa);
2670: } while (cc != NULL);
1.43 thorpej 2671: }
2672:
2673: /*
1.134 ad 2674: * pool_cache_xcall:
1.43 thorpej 2675: *
1.134 ad 2676: * Transfer objects from the per-CPU cache to the global cache.
2677: * Run within a cross-call thread.
1.43 thorpej 2678: */
2679: static void
1.134 ad 2680: pool_cache_xcall(pool_cache_t pc)
1.43 thorpej 2681: {
1.134 ad 2682: pool_cache_cpu_t *cc;
2683: pcg_t *prev, *cur, **list;
2684: int s = 0; /* XXXgcc */
2685:
2686: cc = pool_cache_cpu_enter(pc, &s);
2687: cur = cc->cc_current;
2688: cc->cc_current = NULL;
2689: prev = cc->cc_previous;
2690: cc->cc_previous = NULL;
2691: pool_cache_cpu_exit(cc, &s);
2692:
2693: /*
2694: * XXXSMP Go to splvm to prevent kernel_lock from being taken,
2695: * because locks at IPL_SOFTXXX are still spinlocks. Does not
2696: * apply to IPL_SOFTBIO. Cross-call threads do not take the
2697: * kernel_lock.
1.101 thorpej 2698: */
1.134 ad 2699: s = splvm();
2700: mutex_enter(&pc->pc_lock);
2701: if (cur != NULL) {
2702: if (cur->pcg_avail == PCG_NOBJECTS) {
2703: list = &pc->pc_fullgroups;
2704: pc->pc_nfull++;
2705: } else if (cur->pcg_avail == 0) {
2706: list = &pc->pc_emptygroups;
2707: pc->pc_nempty++;
2708: } else {
2709: list = &pc->pc_partgroups;
2710: pc->pc_npart++;
2711: }
2712: cur->pcg_next = *list;
2713: *list = cur;
2714: }
2715: if (prev != NULL) {
2716: if (prev->pcg_avail == PCG_NOBJECTS) {
2717: list = &pc->pc_fullgroups;
2718: pc->pc_nfull++;
2719: } else if (prev->pcg_avail == 0) {
2720: list = &pc->pc_emptygroups;
2721: pc->pc_nempty++;
2722: } else {
2723: list = &pc->pc_partgroups;
2724: pc->pc_npart++;
2725: }
2726: prev->pcg_next = *list;
2727: *list = prev;
2728: }
2729: mutex_exit(&pc->pc_lock);
2730: splx(s);
1.3 pk 2731: }
1.66 thorpej 2732:
2733: /*
2734: * Pool backend allocators.
2735: *
2736: * Each pool has a backend allocator that handles allocation, deallocation,
2737: * and any additional draining that might be needed.
2738: *
2739: * We provide two standard allocators:
2740: *
2741: * pool_allocator_kmem - the default when no allocator is specified
2742: *
2743: * pool_allocator_nointr - used for pools that will not be accessed
2744: * in interrupt context.
2745: */
2746: void *pool_page_alloc(struct pool *, int);
2747: void pool_page_free(struct pool *, void *);
2748:
1.112 bjh21 2749: #ifdef POOL_SUBPAGE
2750: struct pool_allocator pool_allocator_kmem_fullpage = {
2751: pool_page_alloc, pool_page_free, 0,
1.117 yamt 2752: .pa_backingmapptr = &kmem_map,
1.112 bjh21 2753: };
2754: #else
1.66 thorpej 2755: struct pool_allocator pool_allocator_kmem = {
2756: pool_page_alloc, pool_page_free, 0,
1.117 yamt 2757: .pa_backingmapptr = &kmem_map,
1.66 thorpej 2758: };
1.112 bjh21 2759: #endif
1.66 thorpej 2760:
2761: void *pool_page_alloc_nointr(struct pool *, int);
2762: void pool_page_free_nointr(struct pool *, void *);
2763:
1.112 bjh21 2764: #ifdef POOL_SUBPAGE
2765: struct pool_allocator pool_allocator_nointr_fullpage = {
2766: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1.117 yamt 2767: .pa_backingmapptr = &kernel_map,
1.112 bjh21 2768: };
2769: #else
1.66 thorpej 2770: struct pool_allocator pool_allocator_nointr = {
2771: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1.117 yamt 2772: .pa_backingmapptr = &kernel_map,
1.66 thorpej 2773: };
1.112 bjh21 2774: #endif
1.66 thorpej 2775:
2776: #ifdef POOL_SUBPAGE
2777: void *pool_subpage_alloc(struct pool *, int);
2778: void pool_subpage_free(struct pool *, void *);
2779:
1.112 bjh21 2780: struct pool_allocator pool_allocator_kmem = {
2781: pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE,
1.117 yamt 2782: .pa_backingmapptr = &kmem_map,
1.112 bjh21 2783: };
2784:
2785: void *pool_subpage_alloc_nointr(struct pool *, int);
2786: void pool_subpage_free_nointr(struct pool *, void *);
2787:
2788: struct pool_allocator pool_allocator_nointr = {
2789: pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE,
1.117 yamt 2790: .pa_backingmapptr = &kmem_map,
1.66 thorpej 2791: };
2792: #endif /* POOL_SUBPAGE */
2793:
1.117 yamt 2794: static void *
2795: pool_allocator_alloc(struct pool *pp, int flags)
1.66 thorpej 2796: {
1.117 yamt 2797: struct pool_allocator *pa = pp->pr_alloc;
1.66 thorpej 2798: void *res;
2799:
1.117 yamt 2800: res = (*pa->pa_alloc)(pp, flags);
2801: if (res == NULL && (flags & PR_WAITOK) == 0) {
1.66 thorpej 2802: /*
1.117 yamt 2803: * We only run the drain hook here if PR_NOWAIT.
2804: * In other cases, the hook will be run in
2805: * pool_reclaim().
1.66 thorpej 2806: */
1.117 yamt 2807: if (pp->pr_drain_hook != NULL) {
2808: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
2809: res = (*pa->pa_alloc)(pp, flags);
1.66 thorpej 2810: }
1.117 yamt 2811: }
2812: return res;
1.66 thorpej 2813: }
2814:
1.117 yamt 2815: static void
1.66 thorpej 2816: pool_allocator_free(struct pool *pp, void *v)
2817: {
2818: struct pool_allocator *pa = pp->pr_alloc;
2819:
2820: (*pa->pa_free)(pp, v);
2821: }
2822:
2823: void *
1.124 yamt 2824: pool_page_alloc(struct pool *pp, int flags)
1.66 thorpej 2825: {
1.127 thorpej 2826: bool waitok = (flags & PR_WAITOK) ? true : false;
1.66 thorpej 2827:
1.100 yamt 2828: return ((void *) uvm_km_alloc_poolpage_cache(kmem_map, waitok));
1.66 thorpej 2829: }
2830:
2831: void
1.124 yamt 2832: pool_page_free(struct pool *pp, void *v)
1.66 thorpej 2833: {
2834:
1.98 yamt 2835: uvm_km_free_poolpage_cache(kmem_map, (vaddr_t) v);
2836: }
2837:
2838: static void *
1.124 yamt 2839: pool_page_alloc_meta(struct pool *pp, int flags)
1.98 yamt 2840: {
1.127 thorpej 2841: bool waitok = (flags & PR_WAITOK) ? true : false;
1.98 yamt 2842:
1.100 yamt 2843: return ((void *) uvm_km_alloc_poolpage(kmem_map, waitok));
1.98 yamt 2844: }
2845:
2846: static void
1.124 yamt 2847: pool_page_free_meta(struct pool *pp, void *v)
1.98 yamt 2848: {
2849:
1.100 yamt 2850: uvm_km_free_poolpage(kmem_map, (vaddr_t) v);
1.66 thorpej 2851: }
2852:
2853: #ifdef POOL_SUBPAGE
2854: /* Sub-page allocator, for machines with large hardware pages. */
2855: void *
2856: pool_subpage_alloc(struct pool *pp, int flags)
2857: {
1.134 ad 2858: return pool_get(&psppool, flags);
1.66 thorpej 2859: }
2860:
2861: void
2862: pool_subpage_free(struct pool *pp, void *v)
2863: {
2864: pool_put(&psppool, v);
2865: }
2866:
2867: /* We don't provide a real nointr allocator. Maybe later. */
2868: void *
1.112 bjh21 2869: pool_subpage_alloc_nointr(struct pool *pp, int flags)
1.66 thorpej 2870: {
2871:
2872: return (pool_subpage_alloc(pp, flags));
2873: }
2874:
2875: void
1.112 bjh21 2876: pool_subpage_free_nointr(struct pool *pp, void *v)
1.66 thorpej 2877: {
2878:
2879: pool_subpage_free(pp, v);
2880: }
1.112 bjh21 2881: #endif /* POOL_SUBPAGE */
1.66 thorpej 2882: void *
1.124 yamt 2883: pool_page_alloc_nointr(struct pool *pp, int flags)
1.66 thorpej 2884: {
1.127 thorpej 2885: bool waitok = (flags & PR_WAITOK) ? true : false;
1.66 thorpej 2886:
1.100 yamt 2887: return ((void *) uvm_km_alloc_poolpage_cache(kernel_map, waitok));
1.66 thorpej 2888: }
2889:
2890: void
1.124 yamt 2891: pool_page_free_nointr(struct pool *pp, void *v)
1.66 thorpej 2892: {
2893:
1.98 yamt 2894: uvm_km_free_poolpage_cache(kernel_map, (vaddr_t) v);
1.66 thorpej 2895: }
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