Annotation of src/sys/kern/subr_pool.c, Revision 1.167
1.167 ! skrll 1: /* $NetBSD: subr_pool.c,v 1.166 2008/07/09 02:43:53 yamt Exp $ */
1.1 pk 2:
3: /*-
1.161 ad 4: * Copyright (c) 1997, 1999, 2000, 2002, 2007, 2008 The NetBSD Foundation, Inc.
1.1 pk 5: * All rights reserved.
6: *
7: * This code is derived from software contributed to The NetBSD Foundation
1.20 thorpej 8: * by Paul Kranenburg; by Jason R. Thorpe of the Numerical Aerospace
1.134 ad 9: * Simulation Facility, NASA Ames Research Center, and by Andrew Doran.
1.1 pk 10: *
11: * Redistribution and use in source and binary forms, with or without
12: * modification, are permitted provided that the following conditions
13: * are met:
14: * 1. Redistributions of source code must retain the above copyright
15: * notice, this list of conditions and the following disclaimer.
16: * 2. Redistributions in binary form must reproduce the above copyright
17: * notice, this list of conditions and the following disclaimer in the
18: * documentation and/or other materials provided with the distribution.
19: *
20: * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
21: * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
22: * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
23: * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
24: * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
25: * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
26: * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
27: * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
28: * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
29: * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
30: * POSSIBILITY OF SUCH DAMAGE.
31: */
1.64 lukem 32:
33: #include <sys/cdefs.h>
1.167 ! skrll 34: __KERNEL_RCSID(0, "$NetBSD: subr_pool.c,v 1.166 2008/07/09 02:43:53 yamt Exp $");
1.24 scottr 35:
1.141 yamt 36: #include "opt_ddb.h"
1.25 thorpej 37: #include "opt_pool.h"
1.24 scottr 38: #include "opt_poollog.h"
1.28 thorpej 39: #include "opt_lockdebug.h"
1.1 pk 40:
41: #include <sys/param.h>
42: #include <sys/systm.h>
1.135 yamt 43: #include <sys/bitops.h>
1.1 pk 44: #include <sys/proc.h>
45: #include <sys/errno.h>
46: #include <sys/kernel.h>
47: #include <sys/malloc.h>
48: #include <sys/pool.h>
1.20 thorpej 49: #include <sys/syslog.h>
1.125 ad 50: #include <sys/debug.h>
1.134 ad 51: #include <sys/lockdebug.h>
52: #include <sys/xcall.h>
53: #include <sys/cpu.h>
1.145 ad 54: #include <sys/atomic.h>
1.3 pk 55:
56: #include <uvm/uvm.h>
57:
1.1 pk 58: /*
59: * Pool resource management utility.
1.3 pk 60: *
1.88 chs 61: * Memory is allocated in pages which are split into pieces according to
62: * the pool item size. Each page is kept on one of three lists in the
63: * pool structure: `pr_emptypages', `pr_fullpages' and `pr_partpages',
64: * for empty, full and partially-full pages respectively. The individual
65: * pool items are on a linked list headed by `ph_itemlist' in each page
66: * header. The memory for building the page list is either taken from
67: * the allocated pages themselves (for small pool items) or taken from
68: * an internal pool of page headers (`phpool').
1.1 pk 69: */
70:
1.3 pk 71: /* List of all pools */
1.145 ad 72: TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head);
1.134 ad 73:
1.3 pk 74: /* Private pool for page header structures */
1.97 yamt 75: #define PHPOOL_MAX 8
76: static struct pool phpool[PHPOOL_MAX];
1.135 yamt 77: #define PHPOOL_FREELIST_NELEM(idx) \
78: (((idx) == 0) ? 0 : BITMAP_SIZE * (1 << (idx)))
1.3 pk 79:
1.62 bjh21 80: #ifdef POOL_SUBPAGE
81: /* Pool of subpages for use by normal pools. */
82: static struct pool psppool;
83: #endif
84:
1.117 yamt 85: static SLIST_HEAD(, pool_allocator) pa_deferinitq =
86: SLIST_HEAD_INITIALIZER(pa_deferinitq);
87:
1.98 yamt 88: static void *pool_page_alloc_meta(struct pool *, int);
89: static void pool_page_free_meta(struct pool *, void *);
90:
91: /* allocator for pool metadata */
1.134 ad 92: struct pool_allocator pool_allocator_meta = {
1.117 yamt 93: pool_page_alloc_meta, pool_page_free_meta,
94: .pa_backingmapptr = &kmem_map,
1.98 yamt 95: };
96:
1.3 pk 97: /* # of seconds to retain page after last use */
98: int pool_inactive_time = 10;
99:
100: /* Next candidate for drainage (see pool_drain()) */
1.23 thorpej 101: static struct pool *drainpp;
102:
1.134 ad 103: /* This lock protects both pool_head and drainpp. */
104: static kmutex_t pool_head_lock;
105: static kcondvar_t pool_busy;
1.3 pk 106:
1.135 yamt 107: typedef uint32_t pool_item_bitmap_t;
108: #define BITMAP_SIZE (CHAR_BIT * sizeof(pool_item_bitmap_t))
109: #define BITMAP_MASK (BITMAP_SIZE - 1)
1.99 yamt 110:
1.3 pk 111: struct pool_item_header {
112: /* Page headers */
1.88 chs 113: LIST_ENTRY(pool_item_header)
1.3 pk 114: ph_pagelist; /* pool page list */
1.88 chs 115: SPLAY_ENTRY(pool_item_header)
116: ph_node; /* Off-page page headers */
1.128 christos 117: void * ph_page; /* this page's address */
1.151 yamt 118: uint32_t ph_time; /* last referenced */
1.135 yamt 119: uint16_t ph_nmissing; /* # of chunks in use */
1.141 yamt 120: uint16_t ph_off; /* start offset in page */
1.97 yamt 121: union {
122: /* !PR_NOTOUCH */
123: struct {
1.102 chs 124: LIST_HEAD(, pool_item)
1.97 yamt 125: phu_itemlist; /* chunk list for this page */
126: } phu_normal;
127: /* PR_NOTOUCH */
128: struct {
1.141 yamt 129: pool_item_bitmap_t phu_bitmap[1];
1.97 yamt 130: } phu_notouch;
131: } ph_u;
1.3 pk 132: };
1.97 yamt 133: #define ph_itemlist ph_u.phu_normal.phu_itemlist
1.135 yamt 134: #define ph_bitmap ph_u.phu_notouch.phu_bitmap
1.3 pk 135:
1.1 pk 136: struct pool_item {
1.3 pk 137: #ifdef DIAGNOSTIC
1.82 thorpej 138: u_int pi_magic;
1.33 chs 139: #endif
1.134 ad 140: #define PI_MAGIC 0xdeaddeadU
1.3 pk 141: /* Other entries use only this list entry */
1.102 chs 142: LIST_ENTRY(pool_item) pi_list;
1.3 pk 143: };
144:
1.53 thorpej 145: #define POOL_NEEDS_CATCHUP(pp) \
146: ((pp)->pr_nitems < (pp)->pr_minitems)
147:
1.43 thorpej 148: /*
149: * Pool cache management.
150: *
151: * Pool caches provide a way for constructed objects to be cached by the
152: * pool subsystem. This can lead to performance improvements by avoiding
153: * needless object construction/destruction; it is deferred until absolutely
154: * necessary.
155: *
1.134 ad 156: * Caches are grouped into cache groups. Each cache group references up
157: * to PCG_NUMOBJECTS constructed objects. When a cache allocates an
158: * object from the pool, it calls the object's constructor and places it
159: * into a cache group. When a cache group frees an object back to the
160: * pool, it first calls the object's destructor. This allows the object
161: * to persist in constructed form while freed to the cache.
162: *
163: * The pool references each cache, so that when a pool is drained by the
164: * pagedaemon, it can drain each individual cache as well. Each time a
165: * cache is drained, the most idle cache group is freed to the pool in
166: * its entirety.
1.43 thorpej 167: *
168: * Pool caches are layed on top of pools. By layering them, we can avoid
169: * the complexity of cache management for pools which would not benefit
170: * from it.
171: */
172:
1.142 ad 173: static struct pool pcg_normal_pool;
174: static struct pool pcg_large_pool;
1.134 ad 175: static struct pool cache_pool;
176: static struct pool cache_cpu_pool;
1.3 pk 177:
1.145 ad 178: /* List of all caches. */
179: TAILQ_HEAD(,pool_cache) pool_cache_head =
180: TAILQ_HEAD_INITIALIZER(pool_cache_head);
181:
1.162 ad 182: int pool_cache_disable; /* global disable for caching */
183: static pcg_t pcg_dummy; /* zero sized: always empty, yet always full */
1.145 ad 184:
1.162 ad 185: static bool pool_cache_put_slow(pool_cache_cpu_t *, int,
186: void *);
187: static bool pool_cache_get_slow(pool_cache_cpu_t *, int,
188: void **, paddr_t *, int);
1.134 ad 189: static void pool_cache_cpu_init1(struct cpu_info *, pool_cache_t);
190: static void pool_cache_invalidate_groups(pool_cache_t, pcg_t *);
191: static void pool_cache_xcall(pool_cache_t);
1.3 pk 192:
1.42 thorpej 193: static int pool_catchup(struct pool *);
1.128 christos 194: static void pool_prime_page(struct pool *, void *,
1.55 thorpej 195: struct pool_item_header *);
1.88 chs 196: static void pool_update_curpage(struct pool *);
1.66 thorpej 197:
1.113 yamt 198: static int pool_grow(struct pool *, int);
1.117 yamt 199: static void *pool_allocator_alloc(struct pool *, int);
200: static void pool_allocator_free(struct pool *, void *);
1.3 pk 201:
1.97 yamt 202: static void pool_print_pagelist(struct pool *, struct pool_pagelist *,
1.88 chs 203: void (*)(const char *, ...));
1.42 thorpej 204: static void pool_print1(struct pool *, const char *,
205: void (*)(const char *, ...));
1.3 pk 206:
1.88 chs 207: static int pool_chk_page(struct pool *, const char *,
208: struct pool_item_header *);
209:
1.3 pk 210: /*
1.52 thorpej 211: * Pool log entry. An array of these is allocated in pool_init().
1.3 pk 212: */
213: struct pool_log {
214: const char *pl_file;
215: long pl_line;
216: int pl_action;
1.25 thorpej 217: #define PRLOG_GET 1
218: #define PRLOG_PUT 2
1.3 pk 219: void *pl_addr;
1.1 pk 220: };
221:
1.86 matt 222: #ifdef POOL_DIAGNOSTIC
1.3 pk 223: /* Number of entries in pool log buffers */
1.17 thorpej 224: #ifndef POOL_LOGSIZE
225: #define POOL_LOGSIZE 10
226: #endif
227:
228: int pool_logsize = POOL_LOGSIZE;
1.1 pk 229:
1.110 perry 230: static inline void
1.42 thorpej 231: pr_log(struct pool *pp, void *v, int action, const char *file, long line)
1.3 pk 232: {
233: int n = pp->pr_curlogentry;
234: struct pool_log *pl;
235:
1.20 thorpej 236: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 237: return;
238:
239: /*
240: * Fill in the current entry. Wrap around and overwrite
241: * the oldest entry if necessary.
242: */
243: pl = &pp->pr_log[n];
244: pl->pl_file = file;
245: pl->pl_line = line;
246: pl->pl_action = action;
247: pl->pl_addr = v;
248: if (++n >= pp->pr_logsize)
249: n = 0;
250: pp->pr_curlogentry = n;
251: }
252:
253: static void
1.42 thorpej 254: pr_printlog(struct pool *pp, struct pool_item *pi,
255: void (*pr)(const char *, ...))
1.3 pk 256: {
257: int i = pp->pr_logsize;
258: int n = pp->pr_curlogentry;
259:
1.20 thorpej 260: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 261: return;
262:
263: /*
264: * Print all entries in this pool's log.
265: */
266: while (i-- > 0) {
267: struct pool_log *pl = &pp->pr_log[n];
268: if (pl->pl_action != 0) {
1.25 thorpej 269: if (pi == NULL || pi == pl->pl_addr) {
270: (*pr)("\tlog entry %d:\n", i);
271: (*pr)("\t\taction = %s, addr = %p\n",
272: pl->pl_action == PRLOG_GET ? "get" : "put",
273: pl->pl_addr);
274: (*pr)("\t\tfile: %s at line %lu\n",
275: pl->pl_file, pl->pl_line);
276: }
1.3 pk 277: }
278: if (++n >= pp->pr_logsize)
279: n = 0;
280: }
281: }
1.25 thorpej 282:
1.110 perry 283: static inline void
1.42 thorpej 284: pr_enter(struct pool *pp, const char *file, long line)
1.25 thorpej 285: {
286:
1.34 thorpej 287: if (__predict_false(pp->pr_entered_file != NULL)) {
1.25 thorpej 288: printf("pool %s: reentrancy at file %s line %ld\n",
289: pp->pr_wchan, file, line);
290: printf(" previous entry at file %s line %ld\n",
291: pp->pr_entered_file, pp->pr_entered_line);
292: panic("pr_enter");
293: }
294:
295: pp->pr_entered_file = file;
296: pp->pr_entered_line = line;
297: }
298:
1.110 perry 299: static inline void
1.42 thorpej 300: pr_leave(struct pool *pp)
1.25 thorpej 301: {
302:
1.34 thorpej 303: if (__predict_false(pp->pr_entered_file == NULL)) {
1.25 thorpej 304: printf("pool %s not entered?\n", pp->pr_wchan);
305: panic("pr_leave");
306: }
307:
308: pp->pr_entered_file = NULL;
309: pp->pr_entered_line = 0;
310: }
311:
1.110 perry 312: static inline void
1.42 thorpej 313: pr_enter_check(struct pool *pp, void (*pr)(const char *, ...))
1.25 thorpej 314: {
315:
316: if (pp->pr_entered_file != NULL)
317: (*pr)("\n\tcurrently entered from file %s line %ld\n",
318: pp->pr_entered_file, pp->pr_entered_line);
319: }
1.3 pk 320: #else
1.25 thorpej 321: #define pr_log(pp, v, action, file, line)
322: #define pr_printlog(pp, pi, pr)
323: #define pr_enter(pp, file, line)
324: #define pr_leave(pp)
325: #define pr_enter_check(pp, pr)
1.59 thorpej 326: #endif /* POOL_DIAGNOSTIC */
1.3 pk 327:
1.135 yamt 328: static inline unsigned int
1.97 yamt 329: pr_item_notouch_index(const struct pool *pp, const struct pool_item_header *ph,
330: const void *v)
331: {
332: const char *cp = v;
1.135 yamt 333: unsigned int idx;
1.97 yamt 334:
335: KASSERT(pp->pr_roflags & PR_NOTOUCH);
1.128 christos 336: idx = (cp - (char *)ph->ph_page - ph->ph_off) / pp->pr_size;
1.97 yamt 337: KASSERT(idx < pp->pr_itemsperpage);
338: return idx;
339: }
340:
1.110 perry 341: static inline void
1.97 yamt 342: pr_item_notouch_put(const struct pool *pp, struct pool_item_header *ph,
343: void *obj)
344: {
1.135 yamt 345: unsigned int idx = pr_item_notouch_index(pp, ph, obj);
346: pool_item_bitmap_t *bitmap = ph->ph_bitmap + (idx / BITMAP_SIZE);
347: pool_item_bitmap_t mask = 1 << (idx & BITMAP_MASK);
1.97 yamt 348:
1.135 yamt 349: KASSERT((*bitmap & mask) == 0);
350: *bitmap |= mask;
1.97 yamt 351: }
352:
1.110 perry 353: static inline void *
1.97 yamt 354: pr_item_notouch_get(const struct pool *pp, struct pool_item_header *ph)
355: {
1.135 yamt 356: pool_item_bitmap_t *bitmap = ph->ph_bitmap;
357: unsigned int idx;
358: int i;
1.97 yamt 359:
1.135 yamt 360: for (i = 0; ; i++) {
361: int bit;
1.97 yamt 362:
1.135 yamt 363: KASSERT((i * BITMAP_SIZE) < pp->pr_itemsperpage);
364: bit = ffs32(bitmap[i]);
365: if (bit) {
366: pool_item_bitmap_t mask;
367:
368: bit--;
369: idx = (i * BITMAP_SIZE) + bit;
370: mask = 1 << bit;
371: KASSERT((bitmap[i] & mask) != 0);
372: bitmap[i] &= ~mask;
373: break;
374: }
375: }
376: KASSERT(idx < pp->pr_itemsperpage);
1.128 christos 377: return (char *)ph->ph_page + ph->ph_off + idx * pp->pr_size;
1.97 yamt 378: }
379:
1.135 yamt 380: static inline void
1.141 yamt 381: pr_item_notouch_init(const struct pool *pp, struct pool_item_header *ph)
1.135 yamt 382: {
383: pool_item_bitmap_t *bitmap = ph->ph_bitmap;
384: const int n = howmany(pp->pr_itemsperpage, BITMAP_SIZE);
385: int i;
386:
387: for (i = 0; i < n; i++) {
388: bitmap[i] = (pool_item_bitmap_t)-1;
389: }
390: }
391:
1.110 perry 392: static inline int
1.88 chs 393: phtree_compare(struct pool_item_header *a, struct pool_item_header *b)
394: {
1.121 yamt 395:
396: /*
397: * we consider pool_item_header with smaller ph_page bigger.
398: * (this unnatural ordering is for the benefit of pr_find_pagehead.)
399: */
400:
1.88 chs 401: if (a->ph_page < b->ph_page)
1.121 yamt 402: return (1);
403: else if (a->ph_page > b->ph_page)
1.88 chs 404: return (-1);
405: else
406: return (0);
407: }
408:
409: SPLAY_PROTOTYPE(phtree, pool_item_header, ph_node, phtree_compare);
410: SPLAY_GENERATE(phtree, pool_item_header, ph_node, phtree_compare);
411:
1.141 yamt 412: static inline struct pool_item_header *
413: pr_find_pagehead_noalign(struct pool *pp, void *v)
414: {
415: struct pool_item_header *ph, tmp;
416:
417: tmp.ph_page = (void *)(uintptr_t)v;
418: ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
419: if (ph == NULL) {
420: ph = SPLAY_ROOT(&pp->pr_phtree);
421: if (ph != NULL && phtree_compare(&tmp, ph) >= 0) {
422: ph = SPLAY_NEXT(phtree, &pp->pr_phtree, ph);
423: }
424: KASSERT(ph == NULL || phtree_compare(&tmp, ph) < 0);
425: }
426:
427: return ph;
428: }
429:
1.3 pk 430: /*
1.121 yamt 431: * Return the pool page header based on item address.
1.3 pk 432: */
1.110 perry 433: static inline struct pool_item_header *
1.121 yamt 434: pr_find_pagehead(struct pool *pp, void *v)
1.3 pk 435: {
1.88 chs 436: struct pool_item_header *ph, tmp;
1.3 pk 437:
1.121 yamt 438: if ((pp->pr_roflags & PR_NOALIGN) != 0) {
1.141 yamt 439: ph = pr_find_pagehead_noalign(pp, v);
1.121 yamt 440: } else {
1.128 christos 441: void *page =
442: (void *)((uintptr_t)v & pp->pr_alloc->pa_pagemask);
1.121 yamt 443:
444: if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
1.128 christos 445: ph = (struct pool_item_header *)((char *)page + pp->pr_phoffset);
1.121 yamt 446: } else {
447: tmp.ph_page = page;
448: ph = SPLAY_FIND(phtree, &pp->pr_phtree, &tmp);
449: }
450: }
1.3 pk 451:
1.121 yamt 452: KASSERT(ph == NULL || ((pp->pr_roflags & PR_PHINPAGE) != 0) ||
1.128 christos 453: ((char *)ph->ph_page <= (char *)v &&
454: (char *)v < (char *)ph->ph_page + pp->pr_alloc->pa_pagesz));
1.88 chs 455: return ph;
1.3 pk 456: }
457:
1.101 thorpej 458: static void
459: pr_pagelist_free(struct pool *pp, struct pool_pagelist *pq)
460: {
461: struct pool_item_header *ph;
462:
463: while ((ph = LIST_FIRST(pq)) != NULL) {
464: LIST_REMOVE(ph, ph_pagelist);
465: pool_allocator_free(pp, ph->ph_page);
1.134 ad 466: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1.101 thorpej 467: pool_put(pp->pr_phpool, ph);
468: }
469: }
470:
1.3 pk 471: /*
472: * Remove a page from the pool.
473: */
1.110 perry 474: static inline void
1.61 chs 475: pr_rmpage(struct pool *pp, struct pool_item_header *ph,
476: struct pool_pagelist *pq)
1.3 pk 477: {
478:
1.134 ad 479: KASSERT(mutex_owned(&pp->pr_lock));
1.91 yamt 480:
1.3 pk 481: /*
1.7 thorpej 482: * If the page was idle, decrement the idle page count.
1.3 pk 483: */
1.6 thorpej 484: if (ph->ph_nmissing == 0) {
485: #ifdef DIAGNOSTIC
486: if (pp->pr_nidle == 0)
487: panic("pr_rmpage: nidle inconsistent");
1.20 thorpej 488: if (pp->pr_nitems < pp->pr_itemsperpage)
489: panic("pr_rmpage: nitems inconsistent");
1.6 thorpej 490: #endif
491: pp->pr_nidle--;
492: }
1.7 thorpej 493:
1.20 thorpej 494: pp->pr_nitems -= pp->pr_itemsperpage;
495:
1.7 thorpej 496: /*
1.101 thorpej 497: * Unlink the page from the pool and queue it for release.
1.7 thorpej 498: */
1.88 chs 499: LIST_REMOVE(ph, ph_pagelist);
1.91 yamt 500: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
501: SPLAY_REMOVE(phtree, &pp->pr_phtree, ph);
1.101 thorpej 502: LIST_INSERT_HEAD(pq, ph, ph_pagelist);
503:
1.7 thorpej 504: pp->pr_npages--;
505: pp->pr_npagefree++;
1.6 thorpej 506:
1.88 chs 507: pool_update_curpage(pp);
1.3 pk 508: }
509:
1.126 thorpej 510: static bool
1.117 yamt 511: pa_starved_p(struct pool_allocator *pa)
512: {
513:
514: if (pa->pa_backingmap != NULL) {
515: return vm_map_starved_p(pa->pa_backingmap);
516: }
1.127 thorpej 517: return false;
1.117 yamt 518: }
519:
520: static int
1.124 yamt 521: pool_reclaim_callback(struct callback_entry *ce, void *obj, void *arg)
1.117 yamt 522: {
523: struct pool *pp = obj;
524: struct pool_allocator *pa = pp->pr_alloc;
525:
526: KASSERT(&pp->pr_reclaimerentry == ce);
527: pool_reclaim(pp);
528: if (!pa_starved_p(pa)) {
529: return CALLBACK_CHAIN_ABORT;
530: }
531: return CALLBACK_CHAIN_CONTINUE;
532: }
533:
534: static void
535: pool_reclaim_register(struct pool *pp)
536: {
537: struct vm_map *map = pp->pr_alloc->pa_backingmap;
538: int s;
539:
540: if (map == NULL) {
541: return;
542: }
543:
544: s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */
545: callback_register(&vm_map_to_kernel(map)->vmk_reclaim_callback,
546: &pp->pr_reclaimerentry, pp, pool_reclaim_callback);
547: splx(s);
548: }
549:
550: static void
551: pool_reclaim_unregister(struct pool *pp)
552: {
553: struct vm_map *map = pp->pr_alloc->pa_backingmap;
554: int s;
555:
556: if (map == NULL) {
557: return;
558: }
559:
560: s = splvm(); /* not necessary for INTRSAFE maps, but don't care. */
561: callback_unregister(&vm_map_to_kernel(map)->vmk_reclaim_callback,
562: &pp->pr_reclaimerentry);
563: splx(s);
564: }
565:
566: static void
567: pa_reclaim_register(struct pool_allocator *pa)
568: {
569: struct vm_map *map = *pa->pa_backingmapptr;
570: struct pool *pp;
571:
572: KASSERT(pa->pa_backingmap == NULL);
573: if (map == NULL) {
574: SLIST_INSERT_HEAD(&pa_deferinitq, pa, pa_q);
575: return;
576: }
577: pa->pa_backingmap = map;
578: TAILQ_FOREACH(pp, &pa->pa_list, pr_alloc_list) {
579: pool_reclaim_register(pp);
580: }
581: }
582:
1.3 pk 583: /*
1.94 simonb 584: * Initialize all the pools listed in the "pools" link set.
585: */
586: void
1.117 yamt 587: pool_subsystem_init(void)
1.94 simonb 588: {
1.117 yamt 589: struct pool_allocator *pa;
1.94 simonb 590: __link_set_decl(pools, struct link_pool_init);
591: struct link_pool_init * const *pi;
592:
1.134 ad 593: mutex_init(&pool_head_lock, MUTEX_DEFAULT, IPL_NONE);
594: cv_init(&pool_busy, "poolbusy");
595:
1.94 simonb 596: __link_set_foreach(pi, pools)
597: pool_init((*pi)->pp, (*pi)->size, (*pi)->align,
598: (*pi)->align_offset, (*pi)->flags, (*pi)->wchan,
1.129 ad 599: (*pi)->palloc, (*pi)->ipl);
1.117 yamt 600:
601: while ((pa = SLIST_FIRST(&pa_deferinitq)) != NULL) {
602: KASSERT(pa->pa_backingmapptr != NULL);
603: KASSERT(*pa->pa_backingmapptr != NULL);
604: SLIST_REMOVE_HEAD(&pa_deferinitq, pa_q);
605: pa_reclaim_register(pa);
606: }
1.134 ad 607:
1.156 ad 608: pool_init(&cache_pool, sizeof(struct pool_cache), coherency_unit,
1.134 ad 609: 0, 0, "pcache", &pool_allocator_nointr, IPL_NONE);
610:
1.156 ad 611: pool_init(&cache_cpu_pool, sizeof(pool_cache_cpu_t), coherency_unit,
1.134 ad 612: 0, 0, "pcachecpu", &pool_allocator_nointr, IPL_NONE);
1.94 simonb 613: }
614:
615: /*
1.3 pk 616: * Initialize the given pool resource structure.
617: *
618: * We export this routine to allow other kernel parts to declare
619: * static pools that must be initialized before malloc() is available.
620: */
621: void
1.42 thorpej 622: pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
1.129 ad 623: const char *wchan, struct pool_allocator *palloc, int ipl)
1.3 pk 624: {
1.116 simonb 625: struct pool *pp1;
1.92 enami 626: size_t trysize, phsize;
1.134 ad 627: int off, slack;
1.3 pk 628:
1.116 simonb 629: #ifdef DEBUG
630: /*
631: * Check that the pool hasn't already been initialised and
632: * added to the list of all pools.
633: */
1.145 ad 634: TAILQ_FOREACH(pp1, &pool_head, pr_poollist) {
1.116 simonb 635: if (pp == pp1)
636: panic("pool_init: pool %s already initialised",
637: wchan);
638: }
639: #endif
640:
1.25 thorpej 641: #ifdef POOL_DIAGNOSTIC
642: /*
643: * Always log if POOL_DIAGNOSTIC is defined.
644: */
645: if (pool_logsize != 0)
646: flags |= PR_LOGGING;
647: #endif
648:
1.66 thorpej 649: if (palloc == NULL)
650: palloc = &pool_allocator_kmem;
1.112 bjh21 651: #ifdef POOL_SUBPAGE
652: if (size > palloc->pa_pagesz) {
653: if (palloc == &pool_allocator_kmem)
654: palloc = &pool_allocator_kmem_fullpage;
655: else if (palloc == &pool_allocator_nointr)
656: palloc = &pool_allocator_nointr_fullpage;
657: }
1.66 thorpej 658: #endif /* POOL_SUBPAGE */
659: if ((palloc->pa_flags & PA_INITIALIZED) == 0) {
1.112 bjh21 660: if (palloc->pa_pagesz == 0)
1.66 thorpej 661: palloc->pa_pagesz = PAGE_SIZE;
662:
663: TAILQ_INIT(&palloc->pa_list);
664:
1.134 ad 665: mutex_init(&palloc->pa_lock, MUTEX_DEFAULT, IPL_VM);
1.66 thorpej 666: palloc->pa_pagemask = ~(palloc->pa_pagesz - 1);
667: palloc->pa_pageshift = ffs(palloc->pa_pagesz) - 1;
1.117 yamt 668:
669: if (palloc->pa_backingmapptr != NULL) {
670: pa_reclaim_register(palloc);
671: }
1.66 thorpej 672: palloc->pa_flags |= PA_INITIALIZED;
1.4 thorpej 673: }
1.3 pk 674:
675: if (align == 0)
676: align = ALIGN(1);
1.14 thorpej 677:
1.120 yamt 678: if ((flags & PR_NOTOUCH) == 0 && size < sizeof(struct pool_item))
1.14 thorpej 679: size = sizeof(struct pool_item);
1.3 pk 680:
1.78 thorpej 681: size = roundup(size, align);
1.66 thorpej 682: #ifdef DIAGNOSTIC
683: if (size > palloc->pa_pagesz)
1.121 yamt 684: panic("pool_init: pool item size (%zu) too large", size);
1.66 thorpej 685: #endif
1.35 pk 686:
1.3 pk 687: /*
688: * Initialize the pool structure.
689: */
1.88 chs 690: LIST_INIT(&pp->pr_emptypages);
691: LIST_INIT(&pp->pr_fullpages);
692: LIST_INIT(&pp->pr_partpages);
1.134 ad 693: pp->pr_cache = NULL;
1.3 pk 694: pp->pr_curpage = NULL;
695: pp->pr_npages = 0;
696: pp->pr_minitems = 0;
697: pp->pr_minpages = 0;
698: pp->pr_maxpages = UINT_MAX;
1.20 thorpej 699: pp->pr_roflags = flags;
700: pp->pr_flags = 0;
1.35 pk 701: pp->pr_size = size;
1.3 pk 702: pp->pr_align = align;
703: pp->pr_wchan = wchan;
1.66 thorpej 704: pp->pr_alloc = palloc;
1.20 thorpej 705: pp->pr_nitems = 0;
706: pp->pr_nout = 0;
707: pp->pr_hardlimit = UINT_MAX;
708: pp->pr_hardlimit_warning = NULL;
1.31 thorpej 709: pp->pr_hardlimit_ratecap.tv_sec = 0;
710: pp->pr_hardlimit_ratecap.tv_usec = 0;
711: pp->pr_hardlimit_warning_last.tv_sec = 0;
712: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.68 thorpej 713: pp->pr_drain_hook = NULL;
714: pp->pr_drain_hook_arg = NULL;
1.125 ad 715: pp->pr_freecheck = NULL;
1.3 pk 716:
717: /*
718: * Decide whether to put the page header off page to avoid
1.92 enami 719: * wasting too large a part of the page or too big item.
720: * Off-page page headers go on a hash table, so we can match
721: * a returned item with its header based on the page address.
722: * We use 1/16 of the page size and about 8 times of the item
723: * size as the threshold (XXX: tune)
724: *
725: * However, we'll put the header into the page if we can put
726: * it without wasting any items.
727: *
728: * Silently enforce `0 <= ioff < align'.
1.3 pk 729: */
1.92 enami 730: pp->pr_itemoffset = ioff %= align;
731: /* See the comment below about reserved bytes. */
732: trysize = palloc->pa_pagesz - ((align - ioff) % align);
733: phsize = ALIGN(sizeof(struct pool_item_header));
1.121 yamt 734: if ((pp->pr_roflags & (PR_NOTOUCH | PR_NOALIGN)) == 0 &&
1.97 yamt 735: (pp->pr_size < MIN(palloc->pa_pagesz / 16, phsize << 3) ||
736: trysize / pp->pr_size == (trysize - phsize) / pp->pr_size)) {
1.3 pk 737: /* Use the end of the page for the page header */
1.20 thorpej 738: pp->pr_roflags |= PR_PHINPAGE;
1.92 enami 739: pp->pr_phoffset = off = palloc->pa_pagesz - phsize;
1.2 pk 740: } else {
1.3 pk 741: /* The page header will be taken from our page header pool */
742: pp->pr_phoffset = 0;
1.66 thorpej 743: off = palloc->pa_pagesz;
1.88 chs 744: SPLAY_INIT(&pp->pr_phtree);
1.2 pk 745: }
1.1 pk 746:
1.3 pk 747: /*
748: * Alignment is to take place at `ioff' within the item. This means
749: * we must reserve up to `align - 1' bytes on the page to allow
750: * appropriate positioning of each item.
751: */
752: pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
1.43 thorpej 753: KASSERT(pp->pr_itemsperpage != 0);
1.97 yamt 754: if ((pp->pr_roflags & PR_NOTOUCH)) {
755: int idx;
756:
757: for (idx = 0; pp->pr_itemsperpage > PHPOOL_FREELIST_NELEM(idx);
758: idx++) {
759: /* nothing */
760: }
761: if (idx >= PHPOOL_MAX) {
762: /*
763: * if you see this panic, consider to tweak
764: * PHPOOL_MAX and PHPOOL_FREELIST_NELEM.
765: */
766: panic("%s: too large itemsperpage(%d) for PR_NOTOUCH",
767: pp->pr_wchan, pp->pr_itemsperpage);
768: }
769: pp->pr_phpool = &phpool[idx];
770: } else if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
771: pp->pr_phpool = &phpool[0];
772: }
773: #if defined(DIAGNOSTIC)
774: else {
775: pp->pr_phpool = NULL;
776: }
777: #endif
1.3 pk 778:
779: /*
780: * Use the slack between the chunks and the page header
781: * for "cache coloring".
782: */
783: slack = off - pp->pr_itemsperpage * pp->pr_size;
784: pp->pr_maxcolor = (slack / align) * align;
785: pp->pr_curcolor = 0;
786:
787: pp->pr_nget = 0;
788: pp->pr_nfail = 0;
789: pp->pr_nput = 0;
790: pp->pr_npagealloc = 0;
791: pp->pr_npagefree = 0;
1.1 pk 792: pp->pr_hiwat = 0;
1.8 thorpej 793: pp->pr_nidle = 0;
1.134 ad 794: pp->pr_refcnt = 0;
1.3 pk 795:
1.59 thorpej 796: #ifdef POOL_DIAGNOSTIC
1.25 thorpej 797: if (flags & PR_LOGGING) {
798: if (kmem_map == NULL ||
799: (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log),
800: M_TEMP, M_NOWAIT)) == NULL)
1.20 thorpej 801: pp->pr_roflags &= ~PR_LOGGING;
1.3 pk 802: pp->pr_curlogentry = 0;
803: pp->pr_logsize = pool_logsize;
804: }
1.59 thorpej 805: #endif
1.25 thorpej 806:
807: pp->pr_entered_file = NULL;
808: pp->pr_entered_line = 0;
1.3 pk 809:
1.157 ad 810: mutex_init(&pp->pr_lock, MUTEX_DEFAULT, ipl);
1.134 ad 811: cv_init(&pp->pr_cv, wchan);
812: pp->pr_ipl = ipl;
1.1 pk 813:
1.3 pk 814: /*
1.43 thorpej 815: * Initialize private page header pool and cache magazine pool if we
816: * haven't done so yet.
1.23 thorpej 817: * XXX LOCKING.
1.3 pk 818: */
1.97 yamt 819: if (phpool[0].pr_size == 0) {
820: int idx;
821: for (idx = 0; idx < PHPOOL_MAX; idx++) {
822: static char phpool_names[PHPOOL_MAX][6+1+6+1];
823: int nelem;
824: size_t sz;
825:
826: nelem = PHPOOL_FREELIST_NELEM(idx);
827: snprintf(phpool_names[idx], sizeof(phpool_names[idx]),
828: "phpool-%d", nelem);
829: sz = sizeof(struct pool_item_header);
830: if (nelem) {
1.135 yamt 831: sz = offsetof(struct pool_item_header,
832: ph_bitmap[howmany(nelem, BITMAP_SIZE)]);
1.97 yamt 833: }
834: pool_init(&phpool[idx], sz, 0, 0, 0,
1.129 ad 835: phpool_names[idx], &pool_allocator_meta, IPL_VM);
1.97 yamt 836: }
1.62 bjh21 837: #ifdef POOL_SUBPAGE
838: pool_init(&psppool, POOL_SUBPAGE, POOL_SUBPAGE, 0,
1.129 ad 839: PR_RECURSIVE, "psppool", &pool_allocator_meta, IPL_VM);
1.62 bjh21 840: #endif
1.142 ad 841:
842: size = sizeof(pcg_t) +
843: (PCG_NOBJECTS_NORMAL - 1) * sizeof(pcgpair_t);
1.156 ad 844: pool_init(&pcg_normal_pool, size, coherency_unit, 0, 0,
1.142 ad 845: "pcgnormal", &pool_allocator_meta, IPL_VM);
846:
847: size = sizeof(pcg_t) +
848: (PCG_NOBJECTS_LARGE - 1) * sizeof(pcgpair_t);
1.156 ad 849: pool_init(&pcg_large_pool, size, coherency_unit, 0, 0,
1.142 ad 850: "pcglarge", &pool_allocator_meta, IPL_VM);
1.1 pk 851: }
852:
1.145 ad 853: /* Insert into the list of all pools. */
854: if (__predict_true(!cold))
1.134 ad 855: mutex_enter(&pool_head_lock);
1.145 ad 856: TAILQ_FOREACH(pp1, &pool_head, pr_poollist) {
857: if (strcmp(pp1->pr_wchan, pp->pr_wchan) > 0)
858: break;
859: }
860: if (pp1 == NULL)
861: TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist);
862: else
863: TAILQ_INSERT_BEFORE(pp1, pp, pr_poollist);
864: if (__predict_true(!cold))
1.134 ad 865: mutex_exit(&pool_head_lock);
866:
1.167 ! skrll 867: /* Insert this into the list of pools using this allocator. */
1.145 ad 868: if (__predict_true(!cold))
1.134 ad 869: mutex_enter(&palloc->pa_lock);
1.145 ad 870: TAILQ_INSERT_TAIL(&palloc->pa_list, pp, pr_alloc_list);
871: if (__predict_true(!cold))
1.134 ad 872: mutex_exit(&palloc->pa_lock);
1.66 thorpej 873:
1.117 yamt 874: pool_reclaim_register(pp);
1.1 pk 875: }
876:
877: /*
878: * De-commision a pool resource.
879: */
880: void
1.42 thorpej 881: pool_destroy(struct pool *pp)
1.1 pk 882: {
1.101 thorpej 883: struct pool_pagelist pq;
1.3 pk 884: struct pool_item_header *ph;
1.43 thorpej 885:
1.101 thorpej 886: /* Remove from global pool list */
1.134 ad 887: mutex_enter(&pool_head_lock);
888: while (pp->pr_refcnt != 0)
889: cv_wait(&pool_busy, &pool_head_lock);
1.145 ad 890: TAILQ_REMOVE(&pool_head, pp, pr_poollist);
1.101 thorpej 891: if (drainpp == pp)
892: drainpp = NULL;
1.134 ad 893: mutex_exit(&pool_head_lock);
1.101 thorpej 894:
895: /* Remove this pool from its allocator's list of pools. */
1.117 yamt 896: pool_reclaim_unregister(pp);
1.134 ad 897: mutex_enter(&pp->pr_alloc->pa_lock);
1.66 thorpej 898: TAILQ_REMOVE(&pp->pr_alloc->pa_list, pp, pr_alloc_list);
1.134 ad 899: mutex_exit(&pp->pr_alloc->pa_lock);
1.66 thorpej 900:
1.134 ad 901: mutex_enter(&pp->pr_lock);
1.101 thorpej 902:
1.134 ad 903: KASSERT(pp->pr_cache == NULL);
1.3 pk 904:
905: #ifdef DIAGNOSTIC
1.20 thorpej 906: if (pp->pr_nout != 0) {
1.25 thorpej 907: pr_printlog(pp, NULL, printf);
1.80 provos 908: panic("pool_destroy: pool busy: still out: %u",
1.20 thorpej 909: pp->pr_nout);
1.3 pk 910: }
911: #endif
1.1 pk 912:
1.101 thorpej 913: KASSERT(LIST_EMPTY(&pp->pr_fullpages));
914: KASSERT(LIST_EMPTY(&pp->pr_partpages));
915:
1.3 pk 916: /* Remove all pages */
1.101 thorpej 917: LIST_INIT(&pq);
1.88 chs 918: while ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1.101 thorpej 919: pr_rmpage(pp, ph, &pq);
920:
1.134 ad 921: mutex_exit(&pp->pr_lock);
1.3 pk 922:
1.101 thorpej 923: pr_pagelist_free(pp, &pq);
1.3 pk 924:
1.59 thorpej 925: #ifdef POOL_DIAGNOSTIC
1.20 thorpej 926: if ((pp->pr_roflags & PR_LOGGING) != 0)
1.3 pk 927: free(pp->pr_log, M_TEMP);
1.59 thorpej 928: #endif
1.134 ad 929:
930: cv_destroy(&pp->pr_cv);
931: mutex_destroy(&pp->pr_lock);
1.1 pk 932: }
933:
1.68 thorpej 934: void
935: pool_set_drain_hook(struct pool *pp, void (*fn)(void *, int), void *arg)
936: {
937:
938: /* XXX no locking -- must be used just after pool_init() */
939: #ifdef DIAGNOSTIC
940: if (pp->pr_drain_hook != NULL)
941: panic("pool_set_drain_hook(%s): already set", pp->pr_wchan);
942: #endif
943: pp->pr_drain_hook = fn;
944: pp->pr_drain_hook_arg = arg;
945: }
946:
1.88 chs 947: static struct pool_item_header *
1.128 christos 948: pool_alloc_item_header(struct pool *pp, void *storage, int flags)
1.55 thorpej 949: {
950: struct pool_item_header *ph;
951:
952: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
1.128 christos 953: ph = (struct pool_item_header *) ((char *)storage + pp->pr_phoffset);
1.134 ad 954: else
1.97 yamt 955: ph = pool_get(pp->pr_phpool, flags);
1.55 thorpej 956:
957: return (ph);
958: }
1.1 pk 959:
960: /*
1.134 ad 961: * Grab an item from the pool.
1.1 pk 962: */
1.3 pk 963: void *
1.59 thorpej 964: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 965: _pool_get(struct pool *pp, int flags, const char *file, long line)
1.56 sommerfe 966: #else
967: pool_get(struct pool *pp, int flags)
968: #endif
1.1 pk 969: {
970: struct pool_item *pi;
1.3 pk 971: struct pool_item_header *ph;
1.55 thorpej 972: void *v;
1.1 pk 973:
1.2 pk 974: #ifdef DIAGNOSTIC
1.95 atatat 975: if (__predict_false(pp->pr_itemsperpage == 0))
976: panic("pool_get: pool %p: pr_itemsperpage is zero, "
977: "pool not initialized?", pp);
1.84 thorpej 978: if (__predict_false(curlwp == NULL && doing_shutdown == 0 &&
1.37 sommerfe 979: (flags & PR_WAITOK) != 0))
1.77 matt 980: panic("pool_get: %s: must have NOWAIT", pp->pr_wchan);
1.58 thorpej 981:
1.102 chs 982: #endif /* DIAGNOSTIC */
1.58 thorpej 983: #ifdef LOCKDEBUG
1.155 ad 984: if (flags & PR_WAITOK) {
1.154 yamt 985: ASSERT_SLEEPABLE();
1.155 ad 986: }
1.56 sommerfe 987: #endif
1.1 pk 988:
1.134 ad 989: mutex_enter(&pp->pr_lock);
1.25 thorpej 990: pr_enter(pp, file, line);
1.20 thorpej 991:
992: startover:
993: /*
994: * Check to see if we've reached the hard limit. If we have,
995: * and we can wait, then wait until an item has been returned to
996: * the pool.
997: */
998: #ifdef DIAGNOSTIC
1.34 thorpej 999: if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
1.25 thorpej 1000: pr_leave(pp);
1.134 ad 1001: mutex_exit(&pp->pr_lock);
1.20 thorpej 1002: panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
1003: }
1004: #endif
1.34 thorpej 1005: if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
1.68 thorpej 1006: if (pp->pr_drain_hook != NULL) {
1007: /*
1008: * Since the drain hook is going to free things
1009: * back to the pool, unlock, call the hook, re-lock,
1010: * and check the hardlimit condition again.
1011: */
1012: pr_leave(pp);
1.134 ad 1013: mutex_exit(&pp->pr_lock);
1.68 thorpej 1014: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
1.134 ad 1015: mutex_enter(&pp->pr_lock);
1.68 thorpej 1016: pr_enter(pp, file, line);
1017: if (pp->pr_nout < pp->pr_hardlimit)
1018: goto startover;
1019: }
1020:
1.29 sommerfe 1021: if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
1.20 thorpej 1022: /*
1023: * XXX: A warning isn't logged in this case. Should
1024: * it be?
1025: */
1026: pp->pr_flags |= PR_WANTED;
1.25 thorpej 1027: pr_leave(pp);
1.134 ad 1028: cv_wait(&pp->pr_cv, &pp->pr_lock);
1.25 thorpej 1029: pr_enter(pp, file, line);
1.20 thorpej 1030: goto startover;
1031: }
1.31 thorpej 1032:
1033: /*
1034: * Log a message that the hard limit has been hit.
1035: */
1036: if (pp->pr_hardlimit_warning != NULL &&
1037: ratecheck(&pp->pr_hardlimit_warning_last,
1038: &pp->pr_hardlimit_ratecap))
1039: log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
1.21 thorpej 1040:
1041: pp->pr_nfail++;
1042:
1.25 thorpej 1043: pr_leave(pp);
1.134 ad 1044: mutex_exit(&pp->pr_lock);
1.20 thorpej 1045: return (NULL);
1046: }
1047:
1.3 pk 1048: /*
1049: * The convention we use is that if `curpage' is not NULL, then
1050: * it points at a non-empty bucket. In particular, `curpage'
1051: * never points at a page header which has PR_PHINPAGE set and
1052: * has no items in its bucket.
1053: */
1.20 thorpej 1054: if ((ph = pp->pr_curpage) == NULL) {
1.113 yamt 1055: int error;
1056:
1.20 thorpej 1057: #ifdef DIAGNOSTIC
1058: if (pp->pr_nitems != 0) {
1.134 ad 1059: mutex_exit(&pp->pr_lock);
1.20 thorpej 1060: printf("pool_get: %s: curpage NULL, nitems %u\n",
1061: pp->pr_wchan, pp->pr_nitems);
1.80 provos 1062: panic("pool_get: nitems inconsistent");
1.20 thorpej 1063: }
1064: #endif
1065:
1.21 thorpej 1066: /*
1067: * Call the back-end page allocator for more memory.
1068: * Release the pool lock, as the back-end page allocator
1069: * may block.
1070: */
1.25 thorpej 1071: pr_leave(pp);
1.113 yamt 1072: error = pool_grow(pp, flags);
1073: pr_enter(pp, file, line);
1074: if (error != 0) {
1.21 thorpej 1075: /*
1.55 thorpej 1076: * We were unable to allocate a page or item
1077: * header, but we released the lock during
1078: * allocation, so perhaps items were freed
1079: * back to the pool. Check for this case.
1.21 thorpej 1080: */
1081: if (pp->pr_curpage != NULL)
1082: goto startover;
1.15 pk 1083:
1.117 yamt 1084: pp->pr_nfail++;
1.25 thorpej 1085: pr_leave(pp);
1.134 ad 1086: mutex_exit(&pp->pr_lock);
1.117 yamt 1087: return (NULL);
1.1 pk 1088: }
1.3 pk 1089:
1.20 thorpej 1090: /* Start the allocation process over. */
1091: goto startover;
1.3 pk 1092: }
1.97 yamt 1093: if (pp->pr_roflags & PR_NOTOUCH) {
1094: #ifdef DIAGNOSTIC
1095: if (__predict_false(ph->ph_nmissing == pp->pr_itemsperpage)) {
1096: pr_leave(pp);
1.134 ad 1097: mutex_exit(&pp->pr_lock);
1.97 yamt 1098: panic("pool_get: %s: page empty", pp->pr_wchan);
1099: }
1100: #endif
1101: v = pr_item_notouch_get(pp, ph);
1102: #ifdef POOL_DIAGNOSTIC
1103: pr_log(pp, v, PRLOG_GET, file, line);
1104: #endif
1105: } else {
1.102 chs 1106: v = pi = LIST_FIRST(&ph->ph_itemlist);
1.97 yamt 1107: if (__predict_false(v == NULL)) {
1108: pr_leave(pp);
1.134 ad 1109: mutex_exit(&pp->pr_lock);
1.97 yamt 1110: panic("pool_get: %s: page empty", pp->pr_wchan);
1111: }
1.20 thorpej 1112: #ifdef DIAGNOSTIC
1.97 yamt 1113: if (__predict_false(pp->pr_nitems == 0)) {
1114: pr_leave(pp);
1.134 ad 1115: mutex_exit(&pp->pr_lock);
1.97 yamt 1116: printf("pool_get: %s: items on itemlist, nitems %u\n",
1117: pp->pr_wchan, pp->pr_nitems);
1118: panic("pool_get: nitems inconsistent");
1119: }
1.65 enami 1120: #endif
1.56 sommerfe 1121:
1.65 enami 1122: #ifdef POOL_DIAGNOSTIC
1.97 yamt 1123: pr_log(pp, v, PRLOG_GET, file, line);
1.65 enami 1124: #endif
1.3 pk 1125:
1.65 enami 1126: #ifdef DIAGNOSTIC
1.97 yamt 1127: if (__predict_false(pi->pi_magic != PI_MAGIC)) {
1128: pr_printlog(pp, pi, printf);
1129: panic("pool_get(%s): free list modified: "
1130: "magic=%x; page %p; item addr %p\n",
1131: pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
1132: }
1.3 pk 1133: #endif
1134:
1.97 yamt 1135: /*
1136: * Remove from item list.
1137: */
1.102 chs 1138: LIST_REMOVE(pi, pi_list);
1.97 yamt 1139: }
1.20 thorpej 1140: pp->pr_nitems--;
1141: pp->pr_nout++;
1.6 thorpej 1142: if (ph->ph_nmissing == 0) {
1143: #ifdef DIAGNOSTIC
1.34 thorpej 1144: if (__predict_false(pp->pr_nidle == 0))
1.6 thorpej 1145: panic("pool_get: nidle inconsistent");
1146: #endif
1147: pp->pr_nidle--;
1.88 chs 1148:
1149: /*
1150: * This page was previously empty. Move it to the list of
1151: * partially-full pages. This page is already curpage.
1152: */
1153: LIST_REMOVE(ph, ph_pagelist);
1154: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.6 thorpej 1155: }
1.3 pk 1156: ph->ph_nmissing++;
1.97 yamt 1157: if (ph->ph_nmissing == pp->pr_itemsperpage) {
1.21 thorpej 1158: #ifdef DIAGNOSTIC
1.97 yamt 1159: if (__predict_false((pp->pr_roflags & PR_NOTOUCH) == 0 &&
1.102 chs 1160: !LIST_EMPTY(&ph->ph_itemlist))) {
1.25 thorpej 1161: pr_leave(pp);
1.134 ad 1162: mutex_exit(&pp->pr_lock);
1.21 thorpej 1163: panic("pool_get: %s: nmissing inconsistent",
1164: pp->pr_wchan);
1165: }
1166: #endif
1.3 pk 1167: /*
1.88 chs 1168: * This page is now full. Move it to the full list
1169: * and select a new current page.
1.3 pk 1170: */
1.88 chs 1171: LIST_REMOVE(ph, ph_pagelist);
1172: LIST_INSERT_HEAD(&pp->pr_fullpages, ph, ph_pagelist);
1173: pool_update_curpage(pp);
1.1 pk 1174: }
1.3 pk 1175:
1176: pp->pr_nget++;
1.111 christos 1177: pr_leave(pp);
1.20 thorpej 1178:
1179: /*
1180: * If we have a low water mark and we are now below that low
1181: * water mark, add more items to the pool.
1182: */
1.53 thorpej 1183: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1184: /*
1185: * XXX: Should we log a warning? Should we set up a timeout
1186: * to try again in a second or so? The latter could break
1187: * a caller's assumptions about interrupt protection, etc.
1188: */
1189: }
1190:
1.134 ad 1191: mutex_exit(&pp->pr_lock);
1.125 ad 1192: KASSERT((((vaddr_t)v + pp->pr_itemoffset) & (pp->pr_align - 1)) == 0);
1193: FREECHECK_OUT(&pp->pr_freecheck, v);
1.1 pk 1194: return (v);
1195: }
1196:
1197: /*
1.43 thorpej 1198: * Internal version of pool_put(). Pool is already locked/entered.
1.1 pk 1199: */
1.43 thorpej 1200: static void
1.101 thorpej 1201: pool_do_put(struct pool *pp, void *v, struct pool_pagelist *pq)
1.1 pk 1202: {
1203: struct pool_item *pi = v;
1.3 pk 1204: struct pool_item_header *ph;
1205:
1.134 ad 1206: KASSERT(mutex_owned(&pp->pr_lock));
1.125 ad 1207: FREECHECK_IN(&pp->pr_freecheck, v);
1.134 ad 1208: LOCKDEBUG_MEM_CHECK(v, pp->pr_size);
1.61 chs 1209:
1.30 thorpej 1210: #ifdef DIAGNOSTIC
1.34 thorpej 1211: if (__predict_false(pp->pr_nout == 0)) {
1.30 thorpej 1212: printf("pool %s: putting with none out\n",
1213: pp->pr_wchan);
1214: panic("pool_put");
1215: }
1216: #endif
1.3 pk 1217:
1.121 yamt 1218: if (__predict_false((ph = pr_find_pagehead(pp, v)) == NULL)) {
1.25 thorpej 1219: pr_printlog(pp, NULL, printf);
1.3 pk 1220: panic("pool_put: %s: page header missing", pp->pr_wchan);
1221: }
1.28 thorpej 1222:
1.3 pk 1223: /*
1224: * Return to item list.
1225: */
1.97 yamt 1226: if (pp->pr_roflags & PR_NOTOUCH) {
1227: pr_item_notouch_put(pp, ph, v);
1228: } else {
1.2 pk 1229: #ifdef DIAGNOSTIC
1.97 yamt 1230: pi->pi_magic = PI_MAGIC;
1.3 pk 1231: #endif
1.32 chs 1232: #ifdef DEBUG
1.97 yamt 1233: {
1234: int i, *ip = v;
1.32 chs 1235:
1.97 yamt 1236: for (i = 0; i < pp->pr_size / sizeof(int); i++) {
1237: *ip++ = PI_MAGIC;
1238: }
1.32 chs 1239: }
1240: #endif
1241:
1.102 chs 1242: LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.97 yamt 1243: }
1.79 thorpej 1244: KDASSERT(ph->ph_nmissing != 0);
1.3 pk 1245: ph->ph_nmissing--;
1246: pp->pr_nput++;
1.20 thorpej 1247: pp->pr_nitems++;
1248: pp->pr_nout--;
1.3 pk 1249:
1250: /* Cancel "pool empty" condition if it exists */
1251: if (pp->pr_curpage == NULL)
1252: pp->pr_curpage = ph;
1253:
1254: if (pp->pr_flags & PR_WANTED) {
1255: pp->pr_flags &= ~PR_WANTED;
1.134 ad 1256: cv_broadcast(&pp->pr_cv);
1.3 pk 1257: }
1258:
1259: /*
1.88 chs 1260: * If this page is now empty, do one of two things:
1.21 thorpej 1261: *
1.88 chs 1262: * (1) If we have more pages than the page high water mark,
1.96 thorpej 1263: * free the page back to the system. ONLY CONSIDER
1.90 thorpej 1264: * FREEING BACK A PAGE IF WE HAVE MORE THAN OUR MINIMUM PAGE
1265: * CLAIM.
1.21 thorpej 1266: *
1.88 chs 1267: * (2) Otherwise, move the page to the empty page list.
1268: *
1269: * Either way, select a new current page (so we use a partially-full
1270: * page if one is available).
1.3 pk 1271: */
1272: if (ph->ph_nmissing == 0) {
1.6 thorpej 1273: pp->pr_nidle++;
1.90 thorpej 1274: if (pp->pr_npages > pp->pr_minpages &&
1.152 yamt 1275: pp->pr_npages > pp->pr_maxpages) {
1.101 thorpej 1276: pr_rmpage(pp, ph, pq);
1.3 pk 1277: } else {
1.88 chs 1278: LIST_REMOVE(ph, ph_pagelist);
1279: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.3 pk 1280:
1.21 thorpej 1281: /*
1282: * Update the timestamp on the page. A page must
1283: * be idle for some period of time before it can
1284: * be reclaimed by the pagedaemon. This minimizes
1285: * ping-pong'ing for memory.
1.151 yamt 1286: *
1287: * note for 64-bit time_t: truncating to 32-bit is not
1288: * a problem for our usage.
1.21 thorpej 1289: */
1.151 yamt 1290: ph->ph_time = time_uptime;
1.1 pk 1291: }
1.88 chs 1292: pool_update_curpage(pp);
1.1 pk 1293: }
1.88 chs 1294:
1.21 thorpej 1295: /*
1.88 chs 1296: * If the page was previously completely full, move it to the
1297: * partially-full list and make it the current page. The next
1298: * allocation will get the item from this page, instead of
1299: * further fragmenting the pool.
1.21 thorpej 1300: */
1301: else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
1.88 chs 1302: LIST_REMOVE(ph, ph_pagelist);
1303: LIST_INSERT_HEAD(&pp->pr_partpages, ph, ph_pagelist);
1.21 thorpej 1304: pp->pr_curpage = ph;
1305: }
1.43 thorpej 1306: }
1307:
1308: /*
1.134 ad 1309: * Return resource to the pool.
1.43 thorpej 1310: */
1.59 thorpej 1311: #ifdef POOL_DIAGNOSTIC
1.43 thorpej 1312: void
1313: _pool_put(struct pool *pp, void *v, const char *file, long line)
1314: {
1.101 thorpej 1315: struct pool_pagelist pq;
1316:
1317: LIST_INIT(&pq);
1.43 thorpej 1318:
1.134 ad 1319: mutex_enter(&pp->pr_lock);
1.43 thorpej 1320: pr_enter(pp, file, line);
1321:
1.56 sommerfe 1322: pr_log(pp, v, PRLOG_PUT, file, line);
1323:
1.101 thorpej 1324: pool_do_put(pp, v, &pq);
1.21 thorpej 1325:
1.25 thorpej 1326: pr_leave(pp);
1.134 ad 1327: mutex_exit(&pp->pr_lock);
1.101 thorpej 1328:
1.102 chs 1329: pr_pagelist_free(pp, &pq);
1.1 pk 1330: }
1.57 sommerfe 1331: #undef pool_put
1.59 thorpej 1332: #endif /* POOL_DIAGNOSTIC */
1.1 pk 1333:
1.56 sommerfe 1334: void
1335: pool_put(struct pool *pp, void *v)
1336: {
1.101 thorpej 1337: struct pool_pagelist pq;
1338:
1339: LIST_INIT(&pq);
1.56 sommerfe 1340:
1.134 ad 1341: mutex_enter(&pp->pr_lock);
1.101 thorpej 1342: pool_do_put(pp, v, &pq);
1.134 ad 1343: mutex_exit(&pp->pr_lock);
1.56 sommerfe 1344:
1.102 chs 1345: pr_pagelist_free(pp, &pq);
1.56 sommerfe 1346: }
1.57 sommerfe 1347:
1.59 thorpej 1348: #ifdef POOL_DIAGNOSTIC
1.57 sommerfe 1349: #define pool_put(h, v) _pool_put((h), (v), __FILE__, __LINE__)
1.56 sommerfe 1350: #endif
1.74 thorpej 1351:
1352: /*
1.113 yamt 1353: * pool_grow: grow a pool by a page.
1354: *
1355: * => called with pool locked.
1356: * => unlock and relock the pool.
1357: * => return with pool locked.
1358: */
1359:
1360: static int
1361: pool_grow(struct pool *pp, int flags)
1362: {
1363: struct pool_item_header *ph = NULL;
1364: char *cp;
1365:
1.134 ad 1366: mutex_exit(&pp->pr_lock);
1.113 yamt 1367: cp = pool_allocator_alloc(pp, flags);
1368: if (__predict_true(cp != NULL)) {
1369: ph = pool_alloc_item_header(pp, cp, flags);
1370: }
1371: if (__predict_false(cp == NULL || ph == NULL)) {
1372: if (cp != NULL) {
1373: pool_allocator_free(pp, cp);
1374: }
1.134 ad 1375: mutex_enter(&pp->pr_lock);
1.113 yamt 1376: return ENOMEM;
1377: }
1378:
1.134 ad 1379: mutex_enter(&pp->pr_lock);
1.113 yamt 1380: pool_prime_page(pp, cp, ph);
1381: pp->pr_npagealloc++;
1382: return 0;
1383: }
1384:
1385: /*
1.74 thorpej 1386: * Add N items to the pool.
1387: */
1388: int
1389: pool_prime(struct pool *pp, int n)
1390: {
1.75 simonb 1391: int newpages;
1.113 yamt 1392: int error = 0;
1.74 thorpej 1393:
1.134 ad 1394: mutex_enter(&pp->pr_lock);
1.74 thorpej 1395:
1396: newpages = roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1397:
1398: while (newpages-- > 0) {
1.113 yamt 1399: error = pool_grow(pp, PR_NOWAIT);
1400: if (error) {
1.74 thorpej 1401: break;
1402: }
1403: pp->pr_minpages++;
1404: }
1405:
1406: if (pp->pr_minpages >= pp->pr_maxpages)
1407: pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */
1408:
1.134 ad 1409: mutex_exit(&pp->pr_lock);
1.113 yamt 1410: return error;
1.74 thorpej 1411: }
1.55 thorpej 1412:
1413: /*
1.3 pk 1414: * Add a page worth of items to the pool.
1.21 thorpej 1415: *
1416: * Note, we must be called with the pool descriptor LOCKED.
1.3 pk 1417: */
1.55 thorpej 1418: static void
1.128 christos 1419: pool_prime_page(struct pool *pp, void *storage, struct pool_item_header *ph)
1.3 pk 1420: {
1421: struct pool_item *pi;
1.128 christos 1422: void *cp = storage;
1.125 ad 1423: const unsigned int align = pp->pr_align;
1424: const unsigned int ioff = pp->pr_itemoffset;
1.55 thorpej 1425: int n;
1.36 pk 1426:
1.134 ad 1427: KASSERT(mutex_owned(&pp->pr_lock));
1.91 yamt 1428:
1.66 thorpej 1429: #ifdef DIAGNOSTIC
1.121 yamt 1430: if ((pp->pr_roflags & PR_NOALIGN) == 0 &&
1.150 skrll 1431: ((uintptr_t)cp & (pp->pr_alloc->pa_pagesz - 1)) != 0)
1.36 pk 1432: panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1.66 thorpej 1433: #endif
1.3 pk 1434:
1435: /*
1436: * Insert page header.
1437: */
1.88 chs 1438: LIST_INSERT_HEAD(&pp->pr_emptypages, ph, ph_pagelist);
1.102 chs 1439: LIST_INIT(&ph->ph_itemlist);
1.3 pk 1440: ph->ph_page = storage;
1441: ph->ph_nmissing = 0;
1.151 yamt 1442: ph->ph_time = time_uptime;
1.88 chs 1443: if ((pp->pr_roflags & PR_PHINPAGE) == 0)
1444: SPLAY_INSERT(phtree, &pp->pr_phtree, ph);
1.3 pk 1445:
1.6 thorpej 1446: pp->pr_nidle++;
1447:
1.3 pk 1448: /*
1449: * Color this page.
1450: */
1.141 yamt 1451: ph->ph_off = pp->pr_curcolor;
1452: cp = (char *)cp + ph->ph_off;
1.3 pk 1453: if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
1454: pp->pr_curcolor = 0;
1455:
1456: /*
1457: * Adjust storage to apply aligment to `pr_itemoffset' in each item.
1458: */
1459: if (ioff != 0)
1.128 christos 1460: cp = (char *)cp + align - ioff;
1.3 pk 1461:
1.125 ad 1462: KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
1463:
1.3 pk 1464: /*
1465: * Insert remaining chunks on the bucket list.
1466: */
1467: n = pp->pr_itemsperpage;
1.20 thorpej 1468: pp->pr_nitems += n;
1.3 pk 1469:
1.97 yamt 1470: if (pp->pr_roflags & PR_NOTOUCH) {
1.141 yamt 1471: pr_item_notouch_init(pp, ph);
1.97 yamt 1472: } else {
1473: while (n--) {
1474: pi = (struct pool_item *)cp;
1.78 thorpej 1475:
1.97 yamt 1476: KASSERT(((((vaddr_t)pi) + ioff) & (align - 1)) == 0);
1.3 pk 1477:
1.97 yamt 1478: /* Insert on page list */
1.102 chs 1479: LIST_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
1.3 pk 1480: #ifdef DIAGNOSTIC
1.97 yamt 1481: pi->pi_magic = PI_MAGIC;
1.3 pk 1482: #endif
1.128 christos 1483: cp = (char *)cp + pp->pr_size;
1.125 ad 1484:
1485: KASSERT((((vaddr_t)cp + ioff) & (align - 1)) == 0);
1.97 yamt 1486: }
1.3 pk 1487: }
1488:
1489: /*
1490: * If the pool was depleted, point at the new page.
1491: */
1492: if (pp->pr_curpage == NULL)
1493: pp->pr_curpage = ph;
1494:
1495: if (++pp->pr_npages > pp->pr_hiwat)
1496: pp->pr_hiwat = pp->pr_npages;
1497: }
1498:
1.20 thorpej 1499: /*
1.52 thorpej 1500: * Used by pool_get() when nitems drops below the low water mark. This
1.88 chs 1501: * is used to catch up pr_nitems with the low water mark.
1.20 thorpej 1502: *
1.21 thorpej 1503: * Note 1, we never wait for memory here, we let the caller decide what to do.
1.20 thorpej 1504: *
1.73 thorpej 1505: * Note 2, we must be called with the pool already locked, and we return
1.20 thorpej 1506: * with it locked.
1507: */
1508: static int
1.42 thorpej 1509: pool_catchup(struct pool *pp)
1.20 thorpej 1510: {
1511: int error = 0;
1512:
1.54 thorpej 1513: while (POOL_NEEDS_CATCHUP(pp)) {
1.113 yamt 1514: error = pool_grow(pp, PR_NOWAIT);
1515: if (error) {
1.20 thorpej 1516: break;
1517: }
1518: }
1.113 yamt 1519: return error;
1.20 thorpej 1520: }
1521:
1.88 chs 1522: static void
1523: pool_update_curpage(struct pool *pp)
1524: {
1525:
1526: pp->pr_curpage = LIST_FIRST(&pp->pr_partpages);
1527: if (pp->pr_curpage == NULL) {
1528: pp->pr_curpage = LIST_FIRST(&pp->pr_emptypages);
1529: }
1530: }
1531:
1.3 pk 1532: void
1.42 thorpej 1533: pool_setlowat(struct pool *pp, int n)
1.3 pk 1534: {
1.15 pk 1535:
1.134 ad 1536: mutex_enter(&pp->pr_lock);
1.21 thorpej 1537:
1.3 pk 1538: pp->pr_minitems = n;
1.15 pk 1539: pp->pr_minpages = (n == 0)
1540: ? 0
1.18 thorpej 1541: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.20 thorpej 1542:
1543: /* Make sure we're caught up with the newly-set low water mark. */
1.75 simonb 1544: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 1545: /*
1546: * XXX: Should we log a warning? Should we set up a timeout
1547: * to try again in a second or so? The latter could break
1548: * a caller's assumptions about interrupt protection, etc.
1549: */
1550: }
1.21 thorpej 1551:
1.134 ad 1552: mutex_exit(&pp->pr_lock);
1.3 pk 1553: }
1554:
1555: void
1.42 thorpej 1556: pool_sethiwat(struct pool *pp, int n)
1.3 pk 1557: {
1.15 pk 1558:
1.134 ad 1559: mutex_enter(&pp->pr_lock);
1.21 thorpej 1560:
1.15 pk 1561: pp->pr_maxpages = (n == 0)
1562: ? 0
1.18 thorpej 1563: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1564:
1.134 ad 1565: mutex_exit(&pp->pr_lock);
1.3 pk 1566: }
1567:
1.20 thorpej 1568: void
1.42 thorpej 1569: pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1.20 thorpej 1570: {
1571:
1.134 ad 1572: mutex_enter(&pp->pr_lock);
1.20 thorpej 1573:
1574: pp->pr_hardlimit = n;
1575: pp->pr_hardlimit_warning = warnmess;
1.31 thorpej 1576: pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1577: pp->pr_hardlimit_warning_last.tv_sec = 0;
1578: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.20 thorpej 1579:
1580: /*
1.21 thorpej 1581: * In-line version of pool_sethiwat(), because we don't want to
1582: * release the lock.
1.20 thorpej 1583: */
1584: pp->pr_maxpages = (n == 0)
1585: ? 0
1586: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1587:
1.134 ad 1588: mutex_exit(&pp->pr_lock);
1.20 thorpej 1589: }
1.3 pk 1590:
1591: /*
1592: * Release all complete pages that have not been used recently.
1593: */
1.66 thorpej 1594: int
1.59 thorpej 1595: #ifdef POOL_DIAGNOSTIC
1.42 thorpej 1596: _pool_reclaim(struct pool *pp, const char *file, long line)
1.56 sommerfe 1597: #else
1598: pool_reclaim(struct pool *pp)
1599: #endif
1.3 pk 1600: {
1601: struct pool_item_header *ph, *phnext;
1.61 chs 1602: struct pool_pagelist pq;
1.151 yamt 1603: uint32_t curtime;
1.134 ad 1604: bool klock;
1605: int rv;
1.3 pk 1606:
1.68 thorpej 1607: if (pp->pr_drain_hook != NULL) {
1608: /*
1609: * The drain hook must be called with the pool unlocked.
1610: */
1611: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, PR_NOWAIT);
1612: }
1613:
1.134 ad 1614: /*
1.157 ad 1615: * XXXSMP Because we do not want to cause non-MPSAFE code
1616: * to block.
1.134 ad 1617: */
1618: if (pp->pr_ipl == IPL_SOFTNET || pp->pr_ipl == IPL_SOFTCLOCK ||
1619: pp->pr_ipl == IPL_SOFTSERIAL) {
1620: KERNEL_LOCK(1, NULL);
1621: klock = true;
1622: } else
1623: klock = false;
1624:
1625: /* Reclaim items from the pool's cache (if any). */
1626: if (pp->pr_cache != NULL)
1627: pool_cache_invalidate(pp->pr_cache);
1628:
1629: if (mutex_tryenter(&pp->pr_lock) == 0) {
1630: if (klock) {
1631: KERNEL_UNLOCK_ONE(NULL);
1632: }
1.66 thorpej 1633: return (0);
1.134 ad 1634: }
1.25 thorpej 1635: pr_enter(pp, file, line);
1.68 thorpej 1636:
1.88 chs 1637: LIST_INIT(&pq);
1.43 thorpej 1638:
1.151 yamt 1639: curtime = time_uptime;
1.21 thorpej 1640:
1.88 chs 1641: for (ph = LIST_FIRST(&pp->pr_emptypages); ph != NULL; ph = phnext) {
1642: phnext = LIST_NEXT(ph, ph_pagelist);
1.3 pk 1643:
1644: /* Check our minimum page claim */
1645: if (pp->pr_npages <= pp->pr_minpages)
1646: break;
1647:
1.88 chs 1648: KASSERT(ph->ph_nmissing == 0);
1.151 yamt 1649: if (curtime - ph->ph_time < pool_inactive_time
1.117 yamt 1650: && !pa_starved_p(pp->pr_alloc))
1.88 chs 1651: continue;
1.21 thorpej 1652:
1.88 chs 1653: /*
1654: * If freeing this page would put us below
1655: * the low water mark, stop now.
1656: */
1657: if ((pp->pr_nitems - pp->pr_itemsperpage) <
1658: pp->pr_minitems)
1659: break;
1.21 thorpej 1660:
1.88 chs 1661: pr_rmpage(pp, ph, &pq);
1.3 pk 1662: }
1663:
1.25 thorpej 1664: pr_leave(pp);
1.134 ad 1665: mutex_exit(&pp->pr_lock);
1666:
1667: if (LIST_EMPTY(&pq))
1668: rv = 0;
1669: else {
1670: pr_pagelist_free(pp, &pq);
1671: rv = 1;
1672: }
1673:
1674: if (klock) {
1675: KERNEL_UNLOCK_ONE(NULL);
1676: }
1.66 thorpej 1677:
1.134 ad 1678: return (rv);
1.3 pk 1679: }
1680:
1681: /*
1.134 ad 1682: * Drain pools, one at a time. This is a two stage process;
1683: * drain_start kicks off a cross call to drain CPU-level caches
1684: * if the pool has an associated pool_cache. drain_end waits
1685: * for those cross calls to finish, and then drains the cache
1686: * (if any) and pool.
1.131 ad 1687: *
1.134 ad 1688: * Note, must never be called from interrupt context.
1.3 pk 1689: */
1690: void
1.134 ad 1691: pool_drain_start(struct pool **ppp, uint64_t *wp)
1.3 pk 1692: {
1693: struct pool *pp;
1.134 ad 1694:
1.145 ad 1695: KASSERT(!TAILQ_EMPTY(&pool_head));
1.3 pk 1696:
1.61 chs 1697: pp = NULL;
1.134 ad 1698:
1699: /* Find next pool to drain, and add a reference. */
1700: mutex_enter(&pool_head_lock);
1701: do {
1702: if (drainpp == NULL) {
1.145 ad 1703: drainpp = TAILQ_FIRST(&pool_head);
1.134 ad 1704: }
1705: if (drainpp != NULL) {
1706: pp = drainpp;
1.145 ad 1707: drainpp = TAILQ_NEXT(pp, pr_poollist);
1.134 ad 1708: }
1709: /*
1710: * Skip completely idle pools. We depend on at least
1711: * one pool in the system being active.
1712: */
1713: } while (pp == NULL || pp->pr_npages == 0);
1714: pp->pr_refcnt++;
1715: mutex_exit(&pool_head_lock);
1716:
1717: /* If there is a pool_cache, drain CPU level caches. */
1718: *ppp = pp;
1719: if (pp->pr_cache != NULL) {
1720: *wp = xc_broadcast(0, (xcfunc_t)pool_cache_xcall,
1721: pp->pr_cache, NULL);
1722: }
1723: }
1724:
1725: void
1726: pool_drain_end(struct pool *pp, uint64_t where)
1727: {
1728:
1729: if (pp == NULL)
1730: return;
1731:
1732: KASSERT(pp->pr_refcnt > 0);
1733:
1734: /* Wait for remote draining to complete. */
1735: if (pp->pr_cache != NULL)
1736: xc_wait(where);
1737:
1738: /* Drain the cache (if any) and pool.. */
1739: pool_reclaim(pp);
1740:
1741: /* Finally, unlock the pool. */
1742: mutex_enter(&pool_head_lock);
1743: pp->pr_refcnt--;
1744: cv_broadcast(&pool_busy);
1745: mutex_exit(&pool_head_lock);
1.3 pk 1746: }
1747:
1748: /*
1749: * Diagnostic helpers.
1750: */
1751: void
1.42 thorpej 1752: pool_print(struct pool *pp, const char *modif)
1.21 thorpej 1753: {
1754:
1.25 thorpej 1755: pool_print1(pp, modif, printf);
1.21 thorpej 1756: }
1757:
1.25 thorpej 1758: void
1.108 yamt 1759: pool_printall(const char *modif, void (*pr)(const char *, ...))
1760: {
1761: struct pool *pp;
1762:
1.145 ad 1763: TAILQ_FOREACH(pp, &pool_head, pr_poollist) {
1.108 yamt 1764: pool_printit(pp, modif, pr);
1765: }
1766: }
1767:
1768: void
1.42 thorpej 1769: pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.25 thorpej 1770: {
1771:
1772: if (pp == NULL) {
1773: (*pr)("Must specify a pool to print.\n");
1774: return;
1775: }
1776:
1777: pool_print1(pp, modif, pr);
1778: }
1779:
1.21 thorpej 1780: static void
1.124 yamt 1781: pool_print_pagelist(struct pool *pp, struct pool_pagelist *pl,
1.97 yamt 1782: void (*pr)(const char *, ...))
1.88 chs 1783: {
1784: struct pool_item_header *ph;
1785: #ifdef DIAGNOSTIC
1786: struct pool_item *pi;
1787: #endif
1788:
1789: LIST_FOREACH(ph, pl, ph_pagelist) {
1.151 yamt 1790: (*pr)("\t\tpage %p, nmissing %d, time %" PRIu32 "\n",
1791: ph->ph_page, ph->ph_nmissing, ph->ph_time);
1.88 chs 1792: #ifdef DIAGNOSTIC
1.97 yamt 1793: if (!(pp->pr_roflags & PR_NOTOUCH)) {
1.102 chs 1794: LIST_FOREACH(pi, &ph->ph_itemlist, pi_list) {
1.97 yamt 1795: if (pi->pi_magic != PI_MAGIC) {
1796: (*pr)("\t\t\titem %p, magic 0x%x\n",
1797: pi, pi->pi_magic);
1798: }
1.88 chs 1799: }
1800: }
1801: #endif
1802: }
1803: }
1804:
1805: static void
1.42 thorpej 1806: pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.3 pk 1807: {
1.25 thorpej 1808: struct pool_item_header *ph;
1.134 ad 1809: pool_cache_t pc;
1810: pcg_t *pcg;
1811: pool_cache_cpu_t *cc;
1812: uint64_t cpuhit, cpumiss;
1.44 thorpej 1813: int i, print_log = 0, print_pagelist = 0, print_cache = 0;
1.25 thorpej 1814: char c;
1815:
1816: while ((c = *modif++) != '\0') {
1817: if (c == 'l')
1818: print_log = 1;
1819: if (c == 'p')
1820: print_pagelist = 1;
1.44 thorpej 1821: if (c == 'c')
1822: print_cache = 1;
1.25 thorpej 1823: }
1824:
1.134 ad 1825: if ((pc = pp->pr_cache) != NULL) {
1826: (*pr)("POOL CACHE");
1827: } else {
1828: (*pr)("POOL");
1829: }
1830:
1831: (*pr)(" %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1.25 thorpej 1832: pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1833: pp->pr_roflags);
1.66 thorpej 1834: (*pr)("\talloc %p\n", pp->pr_alloc);
1.25 thorpej 1835: (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1836: pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1837: (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1838: pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1839:
1.134 ad 1840: (*pr)("\tnget %lu, nfail %lu, nput %lu\n",
1.25 thorpej 1841: pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1842: (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1843: pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1844:
1845: if (print_pagelist == 0)
1846: goto skip_pagelist;
1847:
1.88 chs 1848: if ((ph = LIST_FIRST(&pp->pr_emptypages)) != NULL)
1849: (*pr)("\n\tempty page list:\n");
1.97 yamt 1850: pool_print_pagelist(pp, &pp->pr_emptypages, pr);
1.88 chs 1851: if ((ph = LIST_FIRST(&pp->pr_fullpages)) != NULL)
1852: (*pr)("\n\tfull page list:\n");
1.97 yamt 1853: pool_print_pagelist(pp, &pp->pr_fullpages, pr);
1.88 chs 1854: if ((ph = LIST_FIRST(&pp->pr_partpages)) != NULL)
1855: (*pr)("\n\tpartial-page list:\n");
1.97 yamt 1856: pool_print_pagelist(pp, &pp->pr_partpages, pr);
1.88 chs 1857:
1.25 thorpej 1858: if (pp->pr_curpage == NULL)
1859: (*pr)("\tno current page\n");
1860: else
1861: (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1862:
1863: skip_pagelist:
1864: if (print_log == 0)
1865: goto skip_log;
1866:
1867: (*pr)("\n");
1868: if ((pp->pr_roflags & PR_LOGGING) == 0)
1869: (*pr)("\tno log\n");
1.122 christos 1870: else {
1.25 thorpej 1871: pr_printlog(pp, NULL, pr);
1.122 christos 1872: }
1.3 pk 1873:
1.25 thorpej 1874: skip_log:
1.44 thorpej 1875:
1.102 chs 1876: #define PR_GROUPLIST(pcg) \
1877: (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail); \
1.142 ad 1878: for (i = 0; i < pcg->pcg_size; i++) { \
1.102 chs 1879: if (pcg->pcg_objects[i].pcgo_pa != \
1880: POOL_PADDR_INVALID) { \
1881: (*pr)("\t\t\t%p, 0x%llx\n", \
1882: pcg->pcg_objects[i].pcgo_va, \
1883: (unsigned long long) \
1884: pcg->pcg_objects[i].pcgo_pa); \
1885: } else { \
1886: (*pr)("\t\t\t%p\n", \
1887: pcg->pcg_objects[i].pcgo_va); \
1888: } \
1889: }
1890:
1.134 ad 1891: if (pc != NULL) {
1892: cpuhit = 0;
1893: cpumiss = 0;
1894: for (i = 0; i < MAXCPUS; i++) {
1895: if ((cc = pc->pc_cpus[i]) == NULL)
1896: continue;
1897: cpuhit += cc->cc_hits;
1898: cpumiss += cc->cc_misses;
1899: }
1900: (*pr)("\tcpu layer hits %llu misses %llu\n", cpuhit, cpumiss);
1901: (*pr)("\tcache layer hits %llu misses %llu\n",
1902: pc->pc_hits, pc->pc_misses);
1903: (*pr)("\tcache layer entry uncontended %llu contended %llu\n",
1904: pc->pc_hits + pc->pc_misses - pc->pc_contended,
1905: pc->pc_contended);
1906: (*pr)("\tcache layer empty groups %u full groups %u\n",
1907: pc->pc_nempty, pc->pc_nfull);
1908: if (print_cache) {
1909: (*pr)("\tfull cache groups:\n");
1910: for (pcg = pc->pc_fullgroups; pcg != NULL;
1911: pcg = pcg->pcg_next) {
1912: PR_GROUPLIST(pcg);
1913: }
1914: (*pr)("\tempty cache groups:\n");
1915: for (pcg = pc->pc_emptygroups; pcg != NULL;
1916: pcg = pcg->pcg_next) {
1917: PR_GROUPLIST(pcg);
1918: }
1.103 chs 1919: }
1.44 thorpej 1920: }
1.102 chs 1921: #undef PR_GROUPLIST
1.44 thorpej 1922:
1.88 chs 1923: pr_enter_check(pp, pr);
1924: }
1925:
1926: static int
1927: pool_chk_page(struct pool *pp, const char *label, struct pool_item_header *ph)
1928: {
1929: struct pool_item *pi;
1.128 christos 1930: void *page;
1.88 chs 1931: int n;
1932:
1.121 yamt 1933: if ((pp->pr_roflags & PR_NOALIGN) == 0) {
1.128 christos 1934: page = (void *)((uintptr_t)ph & pp->pr_alloc->pa_pagemask);
1.121 yamt 1935: if (page != ph->ph_page &&
1936: (pp->pr_roflags & PR_PHINPAGE) != 0) {
1937: if (label != NULL)
1938: printf("%s: ", label);
1939: printf("pool(%p:%s): page inconsistency: page %p;"
1940: " at page head addr %p (p %p)\n", pp,
1941: pp->pr_wchan, ph->ph_page,
1942: ph, page);
1943: return 1;
1944: }
1.88 chs 1945: }
1.3 pk 1946:
1.97 yamt 1947: if ((pp->pr_roflags & PR_NOTOUCH) != 0)
1948: return 0;
1949:
1.102 chs 1950: for (pi = LIST_FIRST(&ph->ph_itemlist), n = 0;
1.88 chs 1951: pi != NULL;
1.102 chs 1952: pi = LIST_NEXT(pi,pi_list), n++) {
1.88 chs 1953:
1954: #ifdef DIAGNOSTIC
1955: if (pi->pi_magic != PI_MAGIC) {
1956: if (label != NULL)
1957: printf("%s: ", label);
1958: printf("pool(%s): free list modified: magic=%x;"
1.121 yamt 1959: " page %p; item ordinal %d; addr %p\n",
1.88 chs 1960: pp->pr_wchan, pi->pi_magic, ph->ph_page,
1.121 yamt 1961: n, pi);
1.88 chs 1962: panic("pool");
1963: }
1964: #endif
1.121 yamt 1965: if ((pp->pr_roflags & PR_NOALIGN) != 0) {
1966: continue;
1967: }
1.128 christos 1968: page = (void *)((uintptr_t)pi & pp->pr_alloc->pa_pagemask);
1.88 chs 1969: if (page == ph->ph_page)
1970: continue;
1971:
1972: if (label != NULL)
1973: printf("%s: ", label);
1974: printf("pool(%p:%s): page inconsistency: page %p;"
1975: " item ordinal %d; addr %p (p %p)\n", pp,
1976: pp->pr_wchan, ph->ph_page,
1977: n, pi, page);
1978: return 1;
1979: }
1980: return 0;
1.3 pk 1981: }
1982:
1.88 chs 1983:
1.3 pk 1984: int
1.42 thorpej 1985: pool_chk(struct pool *pp, const char *label)
1.3 pk 1986: {
1987: struct pool_item_header *ph;
1988: int r = 0;
1989:
1.134 ad 1990: mutex_enter(&pp->pr_lock);
1.88 chs 1991: LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
1992: r = pool_chk_page(pp, label, ph);
1993: if (r) {
1994: goto out;
1995: }
1996: }
1997: LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
1998: r = pool_chk_page(pp, label, ph);
1999: if (r) {
1.3 pk 2000: goto out;
2001: }
1.88 chs 2002: }
2003: LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
2004: r = pool_chk_page(pp, label, ph);
2005: if (r) {
1.3 pk 2006: goto out;
2007: }
2008: }
1.88 chs 2009:
1.3 pk 2010: out:
1.134 ad 2011: mutex_exit(&pp->pr_lock);
1.3 pk 2012: return (r);
1.43 thorpej 2013: }
2014:
2015: /*
2016: * pool_cache_init:
2017: *
2018: * Initialize a pool cache.
1.134 ad 2019: */
2020: pool_cache_t
2021: pool_cache_init(size_t size, u_int align, u_int align_offset, u_int flags,
2022: const char *wchan, struct pool_allocator *palloc, int ipl,
2023: int (*ctor)(void *, void *, int), void (*dtor)(void *, void *), void *arg)
2024: {
2025: pool_cache_t pc;
2026:
2027: pc = pool_get(&cache_pool, PR_WAITOK);
2028: if (pc == NULL)
2029: return NULL;
2030:
2031: pool_cache_bootstrap(pc, size, align, align_offset, flags, wchan,
2032: palloc, ipl, ctor, dtor, arg);
2033:
2034: return pc;
2035: }
2036:
2037: /*
2038: * pool_cache_bootstrap:
1.43 thorpej 2039: *
1.134 ad 2040: * Kernel-private version of pool_cache_init(). The caller
2041: * provides initial storage.
1.43 thorpej 2042: */
2043: void
1.134 ad 2044: pool_cache_bootstrap(pool_cache_t pc, size_t size, u_int align,
2045: u_int align_offset, u_int flags, const char *wchan,
2046: struct pool_allocator *palloc, int ipl,
2047: int (*ctor)(void *, void *, int), void (*dtor)(void *, void *),
1.43 thorpej 2048: void *arg)
2049: {
1.134 ad 2050: CPU_INFO_ITERATOR cii;
1.145 ad 2051: pool_cache_t pc1;
1.134 ad 2052: struct cpu_info *ci;
2053: struct pool *pp;
2054:
2055: pp = &pc->pc_pool;
2056: if (palloc == NULL && ipl == IPL_NONE)
2057: palloc = &pool_allocator_nointr;
2058: pool_init(pp, size, align, align_offset, flags, wchan, palloc, ipl);
1.157 ad 2059: mutex_init(&pc->pc_lock, MUTEX_DEFAULT, ipl);
1.43 thorpej 2060:
1.134 ad 2061: if (ctor == NULL) {
2062: ctor = (int (*)(void *, void *, int))nullop;
2063: }
2064: if (dtor == NULL) {
2065: dtor = (void (*)(void *, void *))nullop;
2066: }
1.43 thorpej 2067:
1.134 ad 2068: pc->pc_emptygroups = NULL;
2069: pc->pc_fullgroups = NULL;
2070: pc->pc_partgroups = NULL;
1.43 thorpej 2071: pc->pc_ctor = ctor;
2072: pc->pc_dtor = dtor;
2073: pc->pc_arg = arg;
1.134 ad 2074: pc->pc_hits = 0;
1.48 thorpej 2075: pc->pc_misses = 0;
1.134 ad 2076: pc->pc_nempty = 0;
2077: pc->pc_npart = 0;
2078: pc->pc_nfull = 0;
2079: pc->pc_contended = 0;
2080: pc->pc_refcnt = 0;
1.136 yamt 2081: pc->pc_freecheck = NULL;
1.134 ad 2082:
1.142 ad 2083: if ((flags & PR_LARGECACHE) != 0) {
2084: pc->pc_pcgsize = PCG_NOBJECTS_LARGE;
1.163 ad 2085: pc->pc_pcgpool = &pcg_large_pool;
1.142 ad 2086: } else {
2087: pc->pc_pcgsize = PCG_NOBJECTS_NORMAL;
1.163 ad 2088: pc->pc_pcgpool = &pcg_normal_pool;
1.142 ad 2089: }
2090:
1.134 ad 2091: /* Allocate per-CPU caches. */
2092: memset(pc->pc_cpus, 0, sizeof(pc->pc_cpus));
2093: pc->pc_ncpu = 0;
1.139 ad 2094: if (ncpu < 2) {
1.137 ad 2095: /* XXX For sparc: boot CPU is not attached yet. */
2096: pool_cache_cpu_init1(curcpu(), pc);
2097: } else {
2098: for (CPU_INFO_FOREACH(cii, ci)) {
2099: pool_cache_cpu_init1(ci, pc);
2100: }
1.134 ad 2101: }
1.145 ad 2102:
2103: /* Add to list of all pools. */
2104: if (__predict_true(!cold))
1.134 ad 2105: mutex_enter(&pool_head_lock);
1.145 ad 2106: TAILQ_FOREACH(pc1, &pool_cache_head, pc_cachelist) {
2107: if (strcmp(pc1->pc_pool.pr_wchan, pc->pc_pool.pr_wchan) > 0)
2108: break;
2109: }
2110: if (pc1 == NULL)
2111: TAILQ_INSERT_TAIL(&pool_cache_head, pc, pc_cachelist);
2112: else
2113: TAILQ_INSERT_BEFORE(pc1, pc, pc_cachelist);
2114: if (__predict_true(!cold))
1.134 ad 2115: mutex_exit(&pool_head_lock);
1.145 ad 2116:
2117: membar_sync();
2118: pp->pr_cache = pc;
1.43 thorpej 2119: }
2120:
2121: /*
2122: * pool_cache_destroy:
2123: *
2124: * Destroy a pool cache.
2125: */
2126: void
1.134 ad 2127: pool_cache_destroy(pool_cache_t pc)
1.43 thorpej 2128: {
1.134 ad 2129: struct pool *pp = &pc->pc_pool;
2130: pool_cache_cpu_t *cc;
2131: pcg_t *pcg;
2132: int i;
2133:
2134: /* Remove it from the global list. */
2135: mutex_enter(&pool_head_lock);
2136: while (pc->pc_refcnt != 0)
2137: cv_wait(&pool_busy, &pool_head_lock);
1.145 ad 2138: TAILQ_REMOVE(&pool_cache_head, pc, pc_cachelist);
1.134 ad 2139: mutex_exit(&pool_head_lock);
1.43 thorpej 2140:
2141: /* First, invalidate the entire cache. */
2142: pool_cache_invalidate(pc);
2143:
1.134 ad 2144: /* Disassociate it from the pool. */
2145: mutex_enter(&pp->pr_lock);
2146: pp->pr_cache = NULL;
2147: mutex_exit(&pp->pr_lock);
2148:
2149: /* Destroy per-CPU data */
2150: for (i = 0; i < MAXCPUS; i++) {
2151: if ((cc = pc->pc_cpus[i]) == NULL)
2152: continue;
1.162 ad 2153: if ((pcg = cc->cc_current) != &pcg_dummy) {
1.134 ad 2154: pcg->pcg_next = NULL;
2155: pool_cache_invalidate_groups(pc, pcg);
2156: }
1.162 ad 2157: if ((pcg = cc->cc_previous) != &pcg_dummy) {
1.134 ad 2158: pcg->pcg_next = NULL;
2159: pool_cache_invalidate_groups(pc, pcg);
2160: }
2161: if (cc != &pc->pc_cpu0)
2162: pool_put(&cache_cpu_pool, cc);
2163: }
2164:
2165: /* Finally, destroy it. */
2166: mutex_destroy(&pc->pc_lock);
2167: pool_destroy(pp);
2168: pool_put(&cache_pool, pc);
2169: }
2170:
2171: /*
2172: * pool_cache_cpu_init1:
2173: *
2174: * Called for each pool_cache whenever a new CPU is attached.
2175: */
2176: static void
2177: pool_cache_cpu_init1(struct cpu_info *ci, pool_cache_t pc)
2178: {
2179: pool_cache_cpu_t *cc;
1.137 ad 2180: int index;
1.134 ad 2181:
1.137 ad 2182: index = ci->ci_index;
2183:
2184: KASSERT(index < MAXCPUS);
1.134 ad 2185:
1.137 ad 2186: if ((cc = pc->pc_cpus[index]) != NULL) {
2187: KASSERT(cc->cc_cpuindex == index);
1.134 ad 2188: return;
2189: }
2190:
2191: /*
2192: * The first CPU is 'free'. This needs to be the case for
2193: * bootstrap - we may not be able to allocate yet.
2194: */
2195: if (pc->pc_ncpu == 0) {
2196: cc = &pc->pc_cpu0;
2197: pc->pc_ncpu = 1;
2198: } else {
2199: mutex_enter(&pc->pc_lock);
2200: pc->pc_ncpu++;
2201: mutex_exit(&pc->pc_lock);
2202: cc = pool_get(&cache_cpu_pool, PR_WAITOK);
2203: }
2204:
2205: cc->cc_ipl = pc->pc_pool.pr_ipl;
2206: cc->cc_iplcookie = makeiplcookie(cc->cc_ipl);
2207: cc->cc_cache = pc;
1.137 ad 2208: cc->cc_cpuindex = index;
1.134 ad 2209: cc->cc_hits = 0;
2210: cc->cc_misses = 0;
1.162 ad 2211: cc->cc_current = &pcg_dummy;
2212: cc->cc_previous = &pcg_dummy;
1.134 ad 2213:
1.137 ad 2214: pc->pc_cpus[index] = cc;
1.43 thorpej 2215: }
2216:
1.134 ad 2217: /*
2218: * pool_cache_cpu_init:
2219: *
2220: * Called whenever a new CPU is attached.
2221: */
2222: void
2223: pool_cache_cpu_init(struct cpu_info *ci)
1.43 thorpej 2224: {
1.134 ad 2225: pool_cache_t pc;
2226:
2227: mutex_enter(&pool_head_lock);
1.145 ad 2228: TAILQ_FOREACH(pc, &pool_cache_head, pc_cachelist) {
1.134 ad 2229: pc->pc_refcnt++;
2230: mutex_exit(&pool_head_lock);
1.43 thorpej 2231:
1.134 ad 2232: pool_cache_cpu_init1(ci, pc);
1.43 thorpej 2233:
1.134 ad 2234: mutex_enter(&pool_head_lock);
2235: pc->pc_refcnt--;
2236: cv_broadcast(&pool_busy);
2237: }
2238: mutex_exit(&pool_head_lock);
1.43 thorpej 2239: }
2240:
1.134 ad 2241: /*
2242: * pool_cache_reclaim:
2243: *
2244: * Reclaim memory from a pool cache.
2245: */
2246: bool
2247: pool_cache_reclaim(pool_cache_t pc)
1.43 thorpej 2248: {
2249:
1.134 ad 2250: return pool_reclaim(&pc->pc_pool);
2251: }
1.43 thorpej 2252:
1.136 yamt 2253: static void
2254: pool_cache_destruct_object1(pool_cache_t pc, void *object)
2255: {
2256:
2257: (*pc->pc_dtor)(pc->pc_arg, object);
2258: pool_put(&pc->pc_pool, object);
2259: }
2260:
1.134 ad 2261: /*
2262: * pool_cache_destruct_object:
2263: *
2264: * Force destruction of an object and its release back into
2265: * the pool.
2266: */
2267: void
2268: pool_cache_destruct_object(pool_cache_t pc, void *object)
2269: {
2270:
1.136 yamt 2271: FREECHECK_IN(&pc->pc_freecheck, object);
2272:
2273: pool_cache_destruct_object1(pc, object);
1.43 thorpej 2274: }
2275:
1.134 ad 2276: /*
2277: * pool_cache_invalidate_groups:
2278: *
2279: * Invalidate a chain of groups and destruct all objects.
2280: */
1.102 chs 2281: static void
1.134 ad 2282: pool_cache_invalidate_groups(pool_cache_t pc, pcg_t *pcg)
1.102 chs 2283: {
1.134 ad 2284: void *object;
2285: pcg_t *next;
2286: int i;
2287:
2288: for (; pcg != NULL; pcg = next) {
2289: next = pcg->pcg_next;
2290:
2291: for (i = 0; i < pcg->pcg_avail; i++) {
2292: object = pcg->pcg_objects[i].pcgo_va;
1.136 yamt 2293: pool_cache_destruct_object1(pc, object);
1.134 ad 2294: }
1.102 chs 2295:
1.142 ad 2296: if (pcg->pcg_size == PCG_NOBJECTS_LARGE) {
2297: pool_put(&pcg_large_pool, pcg);
2298: } else {
2299: KASSERT(pcg->pcg_size == PCG_NOBJECTS_NORMAL);
2300: pool_put(&pcg_normal_pool, pcg);
2301: }
1.102 chs 2302: }
2303: }
2304:
1.43 thorpej 2305: /*
1.134 ad 2306: * pool_cache_invalidate:
1.43 thorpej 2307: *
1.134 ad 2308: * Invalidate a pool cache (destruct and release all of the
2309: * cached objects). Does not reclaim objects from the pool.
1.43 thorpej 2310: */
1.134 ad 2311: void
2312: pool_cache_invalidate(pool_cache_t pc)
2313: {
2314: pcg_t *full, *empty, *part;
2315:
2316: mutex_enter(&pc->pc_lock);
2317: full = pc->pc_fullgroups;
2318: empty = pc->pc_emptygroups;
2319: part = pc->pc_partgroups;
2320: pc->pc_fullgroups = NULL;
2321: pc->pc_emptygroups = NULL;
2322: pc->pc_partgroups = NULL;
2323: pc->pc_nfull = 0;
2324: pc->pc_nempty = 0;
2325: pc->pc_npart = 0;
2326: mutex_exit(&pc->pc_lock);
2327:
2328: pool_cache_invalidate_groups(pc, full);
2329: pool_cache_invalidate_groups(pc, empty);
2330: pool_cache_invalidate_groups(pc, part);
2331: }
2332:
2333: void
2334: pool_cache_set_drain_hook(pool_cache_t pc, void (*fn)(void *, int), void *arg)
2335: {
2336:
2337: pool_set_drain_hook(&pc->pc_pool, fn, arg);
2338: }
2339:
2340: void
2341: pool_cache_setlowat(pool_cache_t pc, int n)
2342: {
2343:
2344: pool_setlowat(&pc->pc_pool, n);
2345: }
2346:
2347: void
2348: pool_cache_sethiwat(pool_cache_t pc, int n)
2349: {
2350:
2351: pool_sethiwat(&pc->pc_pool, n);
2352: }
2353:
2354: void
2355: pool_cache_sethardlimit(pool_cache_t pc, int n, const char *warnmess, int ratecap)
2356: {
2357:
2358: pool_sethardlimit(&pc->pc_pool, n, warnmess, ratecap);
2359: }
2360:
1.162 ad 2361: static bool __noinline
2362: pool_cache_get_slow(pool_cache_cpu_t *cc, int s, void **objectp,
1.134 ad 2363: paddr_t *pap, int flags)
1.43 thorpej 2364: {
1.134 ad 2365: pcg_t *pcg, *cur;
2366: uint64_t ncsw;
2367: pool_cache_t pc;
1.43 thorpej 2368: void *object;
1.58 thorpej 2369:
1.134 ad 2370: pc = cc->cc_cache;
2371: cc->cc_misses++;
1.43 thorpej 2372:
1.134 ad 2373: /*
2374: * Nothing was available locally. Try and grab a group
2375: * from the cache.
2376: */
1.162 ad 2377: if (__predict_false(!mutex_tryenter(&pc->pc_lock))) {
1.134 ad 2378: ncsw = curlwp->l_ncsw;
2379: mutex_enter(&pc->pc_lock);
2380: pc->pc_contended++;
1.43 thorpej 2381:
1.134 ad 2382: /*
2383: * If we context switched while locking, then
2384: * our view of the per-CPU data is invalid:
2385: * retry.
2386: */
2387: if (curlwp->l_ncsw != ncsw) {
2388: mutex_exit(&pc->pc_lock);
1.162 ad 2389: return true;
1.43 thorpej 2390: }
1.102 chs 2391: }
1.43 thorpej 2392:
1.162 ad 2393: if (__predict_true((pcg = pc->pc_fullgroups) != NULL)) {
1.43 thorpej 2394: /*
1.134 ad 2395: * If there's a full group, release our empty
2396: * group back to the cache. Install the full
2397: * group as cc_current and return.
1.43 thorpej 2398: */
1.162 ad 2399: if (__predict_true((cur = cc->cc_current) != &pcg_dummy)) {
1.134 ad 2400: KASSERT(cur->pcg_avail == 0);
2401: cur->pcg_next = pc->pc_emptygroups;
2402: pc->pc_emptygroups = cur;
2403: pc->pc_nempty++;
1.87 thorpej 2404: }
1.142 ad 2405: KASSERT(pcg->pcg_avail == pcg->pcg_size);
1.134 ad 2406: cc->cc_current = pcg;
2407: pc->pc_fullgroups = pcg->pcg_next;
2408: pc->pc_hits++;
2409: pc->pc_nfull--;
2410: mutex_exit(&pc->pc_lock);
1.162 ad 2411: return true;
1.134 ad 2412: }
2413:
2414: /*
2415: * Nothing available locally or in cache. Take the slow
2416: * path: fetch a new object from the pool and construct
2417: * it.
2418: */
2419: pc->pc_misses++;
2420: mutex_exit(&pc->pc_lock);
1.162 ad 2421: splx(s);
1.134 ad 2422:
2423: object = pool_get(&pc->pc_pool, flags);
2424: *objectp = object;
1.162 ad 2425: if (__predict_false(object == NULL))
2426: return false;
1.125 ad 2427:
1.162 ad 2428: if (__predict_false((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0)) {
1.134 ad 2429: pool_put(&pc->pc_pool, object);
2430: *objectp = NULL;
1.162 ad 2431: return false;
1.43 thorpej 2432: }
2433:
1.134 ad 2434: KASSERT((((vaddr_t)object + pc->pc_pool.pr_itemoffset) &
2435: (pc->pc_pool.pr_align - 1)) == 0);
1.43 thorpej 2436:
1.134 ad 2437: if (pap != NULL) {
2438: #ifdef POOL_VTOPHYS
2439: *pap = POOL_VTOPHYS(object);
2440: #else
2441: *pap = POOL_PADDR_INVALID;
2442: #endif
1.102 chs 2443: }
1.43 thorpej 2444:
1.125 ad 2445: FREECHECK_OUT(&pc->pc_freecheck, object);
1.162 ad 2446: return false;
1.43 thorpej 2447: }
2448:
2449: /*
1.134 ad 2450: * pool_cache_get{,_paddr}:
1.43 thorpej 2451: *
1.134 ad 2452: * Get an object from a pool cache (optionally returning
2453: * the physical address of the object).
1.43 thorpej 2454: */
1.134 ad 2455: void *
2456: pool_cache_get_paddr(pool_cache_t pc, int flags, paddr_t *pap)
1.43 thorpej 2457: {
1.134 ad 2458: pool_cache_cpu_t *cc;
2459: pcg_t *pcg;
2460: void *object;
1.60 thorpej 2461: int s;
1.43 thorpej 2462:
1.134 ad 2463: #ifdef LOCKDEBUG
1.155 ad 2464: if (flags & PR_WAITOK) {
1.154 yamt 2465: ASSERT_SLEEPABLE();
1.155 ad 2466: }
1.134 ad 2467: #endif
1.125 ad 2468:
1.162 ad 2469: /* Lock out interrupts and disable preemption. */
2470: s = splvm();
1.165 yamt 2471: while (/* CONSTCOND */ true) {
1.134 ad 2472: /* Try and allocate an object from the current group. */
1.162 ad 2473: cc = pc->pc_cpus[curcpu()->ci_index];
2474: KASSERT(cc->cc_cache == pc);
1.134 ad 2475: pcg = cc->cc_current;
1.162 ad 2476: if (__predict_true(pcg->pcg_avail > 0)) {
1.134 ad 2477: object = pcg->pcg_objects[--pcg->pcg_avail].pcgo_va;
1.162 ad 2478: if (__predict_false(pap != NULL))
1.134 ad 2479: *pap = pcg->pcg_objects[pcg->pcg_avail].pcgo_pa;
1.148 yamt 2480: #if defined(DIAGNOSTIC)
1.134 ad 2481: pcg->pcg_objects[pcg->pcg_avail].pcgo_va = NULL;
1.163 ad 2482: KASSERT(pcg->pcg_avail < pcg->pcg_size);
1.134 ad 2483: KASSERT(object != NULL);
1.163 ad 2484: #endif
1.134 ad 2485: cc->cc_hits++;
1.162 ad 2486: splx(s);
1.134 ad 2487: FREECHECK_OUT(&pc->pc_freecheck, object);
2488: return object;
1.43 thorpej 2489: }
2490:
2491: /*
1.134 ad 2492: * That failed. If the previous group isn't empty, swap
2493: * it with the current group and allocate from there.
1.43 thorpej 2494: */
1.134 ad 2495: pcg = cc->cc_previous;
1.162 ad 2496: if (__predict_true(pcg->pcg_avail > 0)) {
1.134 ad 2497: cc->cc_previous = cc->cc_current;
2498: cc->cc_current = pcg;
2499: continue;
1.43 thorpej 2500: }
2501:
1.134 ad 2502: /*
2503: * Can't allocate from either group: try the slow path.
2504: * If get_slow() allocated an object for us, or if
1.162 ad 2505: * no more objects are available, it will return false.
1.134 ad 2506: * Otherwise, we need to retry.
2507: */
1.165 yamt 2508: if (!pool_cache_get_slow(cc, s, &object, pap, flags))
2509: break;
2510: }
1.43 thorpej 2511:
1.134 ad 2512: return object;
1.51 thorpej 2513: }
2514:
1.162 ad 2515: static bool __noinline
2516: pool_cache_put_slow(pool_cache_cpu_t *cc, int s, void *object)
1.51 thorpej 2517: {
1.163 ad 2518: pcg_t *pcg, *cur;
1.134 ad 2519: uint64_t ncsw;
2520: pool_cache_t pc;
1.51 thorpej 2521:
1.134 ad 2522: pc = cc->cc_cache;
2523: cc->cc_misses++;
1.43 thorpej 2524:
1.162 ad 2525: /* Lock the cache. */
2526: if (__predict_false(!mutex_tryenter(&pc->pc_lock))) {
1.164 ad 2527: ncsw = curlwp->l_ncsw;
1.134 ad 2528: mutex_enter(&pc->pc_lock);
2529: pc->pc_contended++;
1.162 ad 2530:
1.163 ad 2531: /*
2532: * If we context switched while locking, then our view of
2533: * the per-CPU data is invalid: retry.
2534: */
2535: if (__predict_false(curlwp->l_ncsw != ncsw)) {
2536: mutex_exit(&pc->pc_lock);
2537: return true;
2538: }
1.162 ad 2539: }
1.102 chs 2540:
1.163 ad 2541: /* If there are no empty groups in the cache then allocate one. */
2542: if (__predict_false((pcg = pc->pc_emptygroups) == NULL)) {
2543: if (__predict_true(!pool_cache_disable)) {
2544: pcg = pool_get(pc->pc_pcgpool, PR_NOWAIT);
2545: }
2546: if (__predict_true(pcg != NULL)) {
2547: pcg->pcg_avail = 0;
2548: pcg->pcg_size = pc->pc_pcgsize;
2549: }
2550: } else {
2551: pc->pc_emptygroups = pcg->pcg_next;
2552: pc->pc_nempty--;
1.134 ad 2553: }
1.130 ad 2554:
1.162 ad 2555: /*
2556: * If there's a empty group, release our full group back
2557: * to the cache. Install the empty group to the local CPU
2558: * and return.
2559: */
1.163 ad 2560: if (pcg != NULL) {
1.134 ad 2561: KASSERT(pcg->pcg_avail == 0);
1.162 ad 2562: if (__predict_false(cc->cc_previous == &pcg_dummy)) {
1.146 ad 2563: cc->cc_previous = pcg;
2564: } else {
1.162 ad 2565: cur = cc->cc_current;
2566: if (__predict_true(cur != &pcg_dummy)) {
1.163 ad 2567: KASSERT(cur->pcg_avail == cur->pcg_size);
1.146 ad 2568: cur->pcg_next = pc->pc_fullgroups;
2569: pc->pc_fullgroups = cur;
2570: pc->pc_nfull++;
2571: }
2572: cc->cc_current = pcg;
2573: }
1.163 ad 2574: pc->pc_hits++;
1.134 ad 2575: mutex_exit(&pc->pc_lock);
1.162 ad 2576: return true;
1.102 chs 2577: }
1.105 christos 2578:
1.134 ad 2579: /*
1.162 ad 2580: * Nothing available locally or in cache, and we didn't
2581: * allocate an empty group. Take the slow path and destroy
2582: * the object here and now.
1.134 ad 2583: */
2584: pc->pc_misses++;
2585: mutex_exit(&pc->pc_lock);
1.162 ad 2586: splx(s);
2587: pool_cache_destruct_object(pc, object);
1.105 christos 2588:
1.162 ad 2589: return false;
1.134 ad 2590: }
1.102 chs 2591:
1.43 thorpej 2592: /*
1.134 ad 2593: * pool_cache_put{,_paddr}:
1.43 thorpej 2594: *
1.134 ad 2595: * Put an object back to the pool cache (optionally caching the
2596: * physical address of the object).
1.43 thorpej 2597: */
1.101 thorpej 2598: void
1.134 ad 2599: pool_cache_put_paddr(pool_cache_t pc, void *object, paddr_t pa)
1.43 thorpej 2600: {
1.134 ad 2601: pool_cache_cpu_t *cc;
2602: pcg_t *pcg;
2603: int s;
1.101 thorpej 2604:
1.134 ad 2605: FREECHECK_IN(&pc->pc_freecheck, object);
1.101 thorpej 2606:
1.162 ad 2607: /* Lock out interrupts and disable preemption. */
2608: s = splvm();
1.165 yamt 2609: while (/* CONSTCOND */ true) {
1.134 ad 2610: /* If the current group isn't full, release it there. */
1.162 ad 2611: cc = pc->pc_cpus[curcpu()->ci_index];
2612: KASSERT(cc->cc_cache == pc);
1.134 ad 2613: pcg = cc->cc_current;
1.162 ad 2614: if (__predict_true(pcg->pcg_avail < pcg->pcg_size)) {
1.134 ad 2615: pcg->pcg_objects[pcg->pcg_avail].pcgo_va = object;
2616: pcg->pcg_objects[pcg->pcg_avail].pcgo_pa = pa;
2617: pcg->pcg_avail++;
2618: cc->cc_hits++;
1.162 ad 2619: splx(s);
1.134 ad 2620: return;
2621: }
1.43 thorpej 2622:
1.134 ad 2623: /*
1.162 ad 2624: * That failed. If the previous group isn't full, swap
1.134 ad 2625: * it with the current group and try again.
2626: */
2627: pcg = cc->cc_previous;
1.162 ad 2628: if (__predict_true(pcg->pcg_avail < pcg->pcg_size)) {
1.134 ad 2629: cc->cc_previous = cc->cc_current;
2630: cc->cc_current = pcg;
2631: continue;
2632: }
1.43 thorpej 2633:
1.134 ad 2634: /*
2635: * Can't free to either group: try the slow path.
2636: * If put_slow() releases the object for us, it
1.162 ad 2637: * will return false. Otherwise we need to retry.
1.134 ad 2638: */
1.165 yamt 2639: if (!pool_cache_put_slow(cc, s, object))
2640: break;
2641: }
1.43 thorpej 2642: }
2643:
2644: /*
1.134 ad 2645: * pool_cache_xcall:
1.43 thorpej 2646: *
1.134 ad 2647: * Transfer objects from the per-CPU cache to the global cache.
2648: * Run within a cross-call thread.
1.43 thorpej 2649: */
2650: static void
1.134 ad 2651: pool_cache_xcall(pool_cache_t pc)
1.43 thorpej 2652: {
1.134 ad 2653: pool_cache_cpu_t *cc;
2654: pcg_t *prev, *cur, **list;
1.162 ad 2655: int s;
1.134 ad 2656:
1.162 ad 2657: s = splvm();
2658: mutex_enter(&pc->pc_lock);
2659: cc = pc->pc_cpus[curcpu()->ci_index];
1.134 ad 2660: cur = cc->cc_current;
1.162 ad 2661: cc->cc_current = &pcg_dummy;
1.134 ad 2662: prev = cc->cc_previous;
1.162 ad 2663: cc->cc_previous = &pcg_dummy;
2664: if (cur != &pcg_dummy) {
1.142 ad 2665: if (cur->pcg_avail == cur->pcg_size) {
1.134 ad 2666: list = &pc->pc_fullgroups;
2667: pc->pc_nfull++;
2668: } else if (cur->pcg_avail == 0) {
2669: list = &pc->pc_emptygroups;
2670: pc->pc_nempty++;
2671: } else {
2672: list = &pc->pc_partgroups;
2673: pc->pc_npart++;
2674: }
2675: cur->pcg_next = *list;
2676: *list = cur;
2677: }
1.162 ad 2678: if (prev != &pcg_dummy) {
1.142 ad 2679: if (prev->pcg_avail == prev->pcg_size) {
1.134 ad 2680: list = &pc->pc_fullgroups;
2681: pc->pc_nfull++;
2682: } else if (prev->pcg_avail == 0) {
2683: list = &pc->pc_emptygroups;
2684: pc->pc_nempty++;
2685: } else {
2686: list = &pc->pc_partgroups;
2687: pc->pc_npart++;
2688: }
2689: prev->pcg_next = *list;
2690: *list = prev;
2691: }
2692: mutex_exit(&pc->pc_lock);
2693: splx(s);
1.3 pk 2694: }
1.66 thorpej 2695:
2696: /*
2697: * Pool backend allocators.
2698: *
2699: * Each pool has a backend allocator that handles allocation, deallocation,
2700: * and any additional draining that might be needed.
2701: *
2702: * We provide two standard allocators:
2703: *
2704: * pool_allocator_kmem - the default when no allocator is specified
2705: *
2706: * pool_allocator_nointr - used for pools that will not be accessed
2707: * in interrupt context.
2708: */
2709: void *pool_page_alloc(struct pool *, int);
2710: void pool_page_free(struct pool *, void *);
2711:
1.112 bjh21 2712: #ifdef POOL_SUBPAGE
2713: struct pool_allocator pool_allocator_kmem_fullpage = {
2714: pool_page_alloc, pool_page_free, 0,
1.117 yamt 2715: .pa_backingmapptr = &kmem_map,
1.112 bjh21 2716: };
2717: #else
1.66 thorpej 2718: struct pool_allocator pool_allocator_kmem = {
2719: pool_page_alloc, pool_page_free, 0,
1.117 yamt 2720: .pa_backingmapptr = &kmem_map,
1.66 thorpej 2721: };
1.112 bjh21 2722: #endif
1.66 thorpej 2723:
2724: void *pool_page_alloc_nointr(struct pool *, int);
2725: void pool_page_free_nointr(struct pool *, void *);
2726:
1.112 bjh21 2727: #ifdef POOL_SUBPAGE
2728: struct pool_allocator pool_allocator_nointr_fullpage = {
2729: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1.117 yamt 2730: .pa_backingmapptr = &kernel_map,
1.112 bjh21 2731: };
2732: #else
1.66 thorpej 2733: struct pool_allocator pool_allocator_nointr = {
2734: pool_page_alloc_nointr, pool_page_free_nointr, 0,
1.117 yamt 2735: .pa_backingmapptr = &kernel_map,
1.66 thorpej 2736: };
1.112 bjh21 2737: #endif
1.66 thorpej 2738:
2739: #ifdef POOL_SUBPAGE
2740: void *pool_subpage_alloc(struct pool *, int);
2741: void pool_subpage_free(struct pool *, void *);
2742:
1.112 bjh21 2743: struct pool_allocator pool_allocator_kmem = {
2744: pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE,
1.117 yamt 2745: .pa_backingmapptr = &kmem_map,
1.112 bjh21 2746: };
2747:
2748: void *pool_subpage_alloc_nointr(struct pool *, int);
2749: void pool_subpage_free_nointr(struct pool *, void *);
2750:
2751: struct pool_allocator pool_allocator_nointr = {
2752: pool_subpage_alloc, pool_subpage_free, POOL_SUBPAGE,
1.117 yamt 2753: .pa_backingmapptr = &kmem_map,
1.66 thorpej 2754: };
2755: #endif /* POOL_SUBPAGE */
2756:
1.117 yamt 2757: static void *
2758: pool_allocator_alloc(struct pool *pp, int flags)
1.66 thorpej 2759: {
1.117 yamt 2760: struct pool_allocator *pa = pp->pr_alloc;
1.66 thorpej 2761: void *res;
2762:
1.117 yamt 2763: res = (*pa->pa_alloc)(pp, flags);
2764: if (res == NULL && (flags & PR_WAITOK) == 0) {
1.66 thorpej 2765: /*
1.117 yamt 2766: * We only run the drain hook here if PR_NOWAIT.
2767: * In other cases, the hook will be run in
2768: * pool_reclaim().
1.66 thorpej 2769: */
1.117 yamt 2770: if (pp->pr_drain_hook != NULL) {
2771: (*pp->pr_drain_hook)(pp->pr_drain_hook_arg, flags);
2772: res = (*pa->pa_alloc)(pp, flags);
1.66 thorpej 2773: }
1.117 yamt 2774: }
2775: return res;
1.66 thorpej 2776: }
2777:
1.117 yamt 2778: static void
1.66 thorpej 2779: pool_allocator_free(struct pool *pp, void *v)
2780: {
2781: struct pool_allocator *pa = pp->pr_alloc;
2782:
2783: (*pa->pa_free)(pp, v);
2784: }
2785:
2786: void *
1.124 yamt 2787: pool_page_alloc(struct pool *pp, int flags)
1.66 thorpej 2788: {
1.127 thorpej 2789: bool waitok = (flags & PR_WAITOK) ? true : false;
1.66 thorpej 2790:
1.100 yamt 2791: return ((void *) uvm_km_alloc_poolpage_cache(kmem_map, waitok));
1.66 thorpej 2792: }
2793:
2794: void
1.124 yamt 2795: pool_page_free(struct pool *pp, void *v)
1.66 thorpej 2796: {
2797:
1.98 yamt 2798: uvm_km_free_poolpage_cache(kmem_map, (vaddr_t) v);
2799: }
2800:
2801: static void *
1.124 yamt 2802: pool_page_alloc_meta(struct pool *pp, int flags)
1.98 yamt 2803: {
1.127 thorpej 2804: bool waitok = (flags & PR_WAITOK) ? true : false;
1.98 yamt 2805:
1.100 yamt 2806: return ((void *) uvm_km_alloc_poolpage(kmem_map, waitok));
1.98 yamt 2807: }
2808:
2809: static void
1.124 yamt 2810: pool_page_free_meta(struct pool *pp, void *v)
1.98 yamt 2811: {
2812:
1.100 yamt 2813: uvm_km_free_poolpage(kmem_map, (vaddr_t) v);
1.66 thorpej 2814: }
2815:
2816: #ifdef POOL_SUBPAGE
2817: /* Sub-page allocator, for machines with large hardware pages. */
2818: void *
2819: pool_subpage_alloc(struct pool *pp, int flags)
2820: {
1.134 ad 2821: return pool_get(&psppool, flags);
1.66 thorpej 2822: }
2823:
2824: void
2825: pool_subpage_free(struct pool *pp, void *v)
2826: {
2827: pool_put(&psppool, v);
2828: }
2829:
2830: /* We don't provide a real nointr allocator. Maybe later. */
2831: void *
1.112 bjh21 2832: pool_subpage_alloc_nointr(struct pool *pp, int flags)
1.66 thorpej 2833: {
2834:
2835: return (pool_subpage_alloc(pp, flags));
2836: }
2837:
2838: void
1.112 bjh21 2839: pool_subpage_free_nointr(struct pool *pp, void *v)
1.66 thorpej 2840: {
2841:
2842: pool_subpage_free(pp, v);
2843: }
1.112 bjh21 2844: #endif /* POOL_SUBPAGE */
1.66 thorpej 2845: void *
1.124 yamt 2846: pool_page_alloc_nointr(struct pool *pp, int flags)
1.66 thorpej 2847: {
1.127 thorpej 2848: bool waitok = (flags & PR_WAITOK) ? true : false;
1.66 thorpej 2849:
1.100 yamt 2850: return ((void *) uvm_km_alloc_poolpage_cache(kernel_map, waitok));
1.66 thorpej 2851: }
2852:
2853: void
1.124 yamt 2854: pool_page_free_nointr(struct pool *pp, void *v)
1.66 thorpej 2855: {
2856:
1.98 yamt 2857: uvm_km_free_poolpage_cache(kernel_map, (vaddr_t) v);
1.66 thorpej 2858: }
1.141 yamt 2859:
2860: #if defined(DDB)
2861: static bool
2862: pool_in_page(struct pool *pp, struct pool_item_header *ph, uintptr_t addr)
2863: {
2864:
2865: return (uintptr_t)ph->ph_page <= addr &&
2866: addr < (uintptr_t)ph->ph_page + pp->pr_alloc->pa_pagesz;
2867: }
2868:
1.143 yamt 2869: static bool
2870: pool_in_item(struct pool *pp, void *item, uintptr_t addr)
2871: {
2872:
2873: return (uintptr_t)item <= addr && addr < (uintptr_t)item + pp->pr_size;
2874: }
2875:
2876: static bool
2877: pool_in_cg(struct pool *pp, struct pool_cache_group *pcg, uintptr_t addr)
2878: {
2879: int i;
2880:
2881: if (pcg == NULL) {
2882: return false;
2883: }
1.144 yamt 2884: for (i = 0; i < pcg->pcg_avail; i++) {
1.143 yamt 2885: if (pool_in_item(pp, pcg->pcg_objects[i].pcgo_va, addr)) {
2886: return true;
2887: }
2888: }
2889: return false;
2890: }
2891:
2892: static bool
2893: pool_allocated(struct pool *pp, struct pool_item_header *ph, uintptr_t addr)
2894: {
2895:
2896: if ((pp->pr_roflags & PR_NOTOUCH) != 0) {
2897: unsigned int idx = pr_item_notouch_index(pp, ph, (void *)addr);
2898: pool_item_bitmap_t *bitmap =
2899: ph->ph_bitmap + (idx / BITMAP_SIZE);
2900: pool_item_bitmap_t mask = 1 << (idx & BITMAP_MASK);
2901:
2902: return (*bitmap & mask) == 0;
2903: } else {
2904: struct pool_item *pi;
2905:
2906: LIST_FOREACH(pi, &ph->ph_itemlist, pi_list) {
2907: if (pool_in_item(pp, pi, addr)) {
2908: return false;
2909: }
2910: }
2911: return true;
2912: }
2913: }
2914:
1.141 yamt 2915: void
2916: pool_whatis(uintptr_t addr, void (*pr)(const char *, ...))
2917: {
2918: struct pool *pp;
2919:
1.145 ad 2920: TAILQ_FOREACH(pp, &pool_head, pr_poollist) {
1.141 yamt 2921: struct pool_item_header *ph;
2922: uintptr_t item;
1.143 yamt 2923: bool allocated = true;
2924: bool incache = false;
2925: bool incpucache = false;
2926: char cpucachestr[32];
1.141 yamt 2927:
2928: if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
2929: LIST_FOREACH(ph, &pp->pr_fullpages, ph_pagelist) {
2930: if (pool_in_page(pp, ph, addr)) {
2931: goto found;
2932: }
2933: }
2934: LIST_FOREACH(ph, &pp->pr_partpages, ph_pagelist) {
2935: if (pool_in_page(pp, ph, addr)) {
1.143 yamt 2936: allocated =
2937: pool_allocated(pp, ph, addr);
2938: goto found;
2939: }
2940: }
2941: LIST_FOREACH(ph, &pp->pr_emptypages, ph_pagelist) {
2942: if (pool_in_page(pp, ph, addr)) {
2943: allocated = false;
1.141 yamt 2944: goto found;
2945: }
2946: }
2947: continue;
2948: } else {
2949: ph = pr_find_pagehead_noalign(pp, (void *)addr);
2950: if (ph == NULL || !pool_in_page(pp, ph, addr)) {
2951: continue;
2952: }
1.143 yamt 2953: allocated = pool_allocated(pp, ph, addr);
1.141 yamt 2954: }
2955: found:
1.143 yamt 2956: if (allocated && pp->pr_cache) {
2957: pool_cache_t pc = pp->pr_cache;
2958: struct pool_cache_group *pcg;
2959: int i;
2960:
2961: for (pcg = pc->pc_fullgroups; pcg != NULL;
2962: pcg = pcg->pcg_next) {
2963: if (pool_in_cg(pp, pcg, addr)) {
2964: incache = true;
2965: goto print;
2966: }
2967: }
2968: for (i = 0; i < MAXCPUS; i++) {
2969: pool_cache_cpu_t *cc;
2970:
2971: if ((cc = pc->pc_cpus[i]) == NULL) {
2972: continue;
2973: }
2974: if (pool_in_cg(pp, cc->cc_current, addr) ||
2975: pool_in_cg(pp, cc->cc_previous, addr)) {
2976: struct cpu_info *ci =
2977: cpu_lookup_byindex(i);
2978:
2979: incpucache = true;
2980: snprintf(cpucachestr,
2981: sizeof(cpucachestr),
2982: "cached by CPU %u",
1.153 martin 2983: ci->ci_index);
1.143 yamt 2984: goto print;
2985: }
2986: }
2987: }
2988: print:
1.141 yamt 2989: item = (uintptr_t)ph->ph_page + ph->ph_off;
2990: item = item + rounddown(addr - item, pp->pr_size);
1.143 yamt 2991: (*pr)("%p is %p+%zu in POOL '%s' (%s)\n",
1.141 yamt 2992: (void *)addr, item, (size_t)(addr - item),
1.143 yamt 2993: pp->pr_wchan,
2994: incpucache ? cpucachestr :
2995: incache ? "cached" : allocated ? "allocated" : "free");
1.141 yamt 2996: }
2997: }
2998: #endif /* defined(DDB) */
CVSweb <webmaster@jp.NetBSD.org>