Annotation of src/sys/kern/subr_pool.c, Revision 1.54
1.54 ! thorpej 1: /* $NetBSD: subr_pool.c,v 1.53 2001/05/10 01:37:40 thorpej Exp $ */
1.1 pk 2:
3: /*-
1.43 thorpej 4: * Copyright (c) 1997, 1999, 2000 The NetBSD Foundation, Inc.
1.1 pk 5: * All rights reserved.
6: *
7: * This code is derived from software contributed to The NetBSD Foundation
1.20 thorpej 8: * by Paul Kranenburg; by Jason R. Thorpe of the Numerical Aerospace
9: * Simulation Facility, NASA Ames Research Center.
1.1 pk 10: *
11: * Redistribution and use in source and binary forms, with or without
12: * modification, are permitted provided that the following conditions
13: * are met:
14: * 1. Redistributions of source code must retain the above copyright
15: * notice, this list of conditions and the following disclaimer.
16: * 2. Redistributions in binary form must reproduce the above copyright
17: * notice, this list of conditions and the following disclaimer in the
18: * documentation and/or other materials provided with the distribution.
19: * 3. All advertising materials mentioning features or use of this software
20: * must display the following acknowledgement:
1.13 christos 21: * This product includes software developed by the NetBSD
22: * Foundation, Inc. and its contributors.
1.1 pk 23: * 4. Neither the name of The NetBSD Foundation nor the names of its
24: * contributors may be used to endorse or promote products derived
25: * from this software without specific prior written permission.
26: *
27: * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
28: * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
29: * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
30: * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
31: * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
32: * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
33: * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
34: * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
35: * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
36: * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
37: * POSSIBILITY OF SUCH DAMAGE.
38: */
1.24 scottr 39:
1.25 thorpej 40: #include "opt_pool.h"
1.24 scottr 41: #include "opt_poollog.h"
1.28 thorpej 42: #include "opt_lockdebug.h"
1.1 pk 43:
44: #include <sys/param.h>
45: #include <sys/systm.h>
46: #include <sys/proc.h>
47: #include <sys/errno.h>
48: #include <sys/kernel.h>
49: #include <sys/malloc.h>
50: #include <sys/lock.h>
51: #include <sys/pool.h>
1.20 thorpej 52: #include <sys/syslog.h>
1.3 pk 53:
54: #include <uvm/uvm.h>
55:
1.1 pk 56: /*
57: * Pool resource management utility.
1.3 pk 58: *
59: * Memory is allocated in pages which are split into pieces according
60: * to the pool item size. Each page is kept on a list headed by `pr_pagelist'
61: * in the pool structure and the individual pool items are on a linked list
62: * headed by `ph_itemlist' in each page header. The memory for building
63: * the page list is either taken from the allocated pages themselves (for
64: * small pool items) or taken from an internal pool of page headers (`phpool').
1.1 pk 65: */
66:
1.3 pk 67: /* List of all pools */
1.5 thorpej 68: TAILQ_HEAD(,pool) pool_head = TAILQ_HEAD_INITIALIZER(pool_head);
1.3 pk 69:
70: /* Private pool for page header structures */
71: static struct pool phpool;
72:
73: /* # of seconds to retain page after last use */
74: int pool_inactive_time = 10;
75:
76: /* Next candidate for drainage (see pool_drain()) */
1.23 thorpej 77: static struct pool *drainpp;
78:
79: /* This spin lock protects both pool_head and drainpp. */
80: struct simplelock pool_head_slock = SIMPLELOCK_INITIALIZER;
1.3 pk 81:
82: struct pool_item_header {
83: /* Page headers */
84: TAILQ_ENTRY(pool_item_header)
85: ph_pagelist; /* pool page list */
86: TAILQ_HEAD(,pool_item) ph_itemlist; /* chunk list for this page */
87: LIST_ENTRY(pool_item_header)
88: ph_hashlist; /* Off-page page headers */
89: int ph_nmissing; /* # of chunks in use */
90: caddr_t ph_page; /* this page's address */
91: struct timeval ph_time; /* last referenced */
92: };
93:
1.1 pk 94: struct pool_item {
1.3 pk 95: #ifdef DIAGNOSTIC
96: int pi_magic;
1.33 chs 97: #endif
1.25 thorpej 98: #define PI_MAGIC 0xdeadbeef
1.3 pk 99: /* Other entries use only this list entry */
100: TAILQ_ENTRY(pool_item) pi_list;
101: };
102:
1.25 thorpej 103: #define PR_HASH_INDEX(pp,addr) \
1.3 pk 104: (((u_long)(addr) >> (pp)->pr_pageshift) & (PR_HASHTABSIZE - 1))
105:
1.53 thorpej 106: #define POOL_NEEDS_CATCHUP(pp) \
107: ((pp)->pr_nitems < (pp)->pr_minitems)
108:
1.43 thorpej 109: /*
110: * Pool cache management.
111: *
112: * Pool caches provide a way for constructed objects to be cached by the
113: * pool subsystem. This can lead to performance improvements by avoiding
114: * needless object construction/destruction; it is deferred until absolutely
115: * necessary.
116: *
117: * Caches are grouped into cache groups. Each cache group references
118: * up to 16 constructed objects. When a cache allocates an object
119: * from the pool, it calls the object's constructor and places it into
120: * a cache group. When a cache group frees an object back to the pool,
121: * it first calls the object's destructor. This allows the object to
122: * persist in constructed form while freed to the cache.
123: *
124: * Multiple caches may exist for each pool. This allows a single
125: * object type to have multiple constructed forms. The pool references
126: * each cache, so that when a pool is drained by the pagedaemon, it can
127: * drain each individual cache as well. Each time a cache is drained,
128: * the most idle cache group is freed to the pool in its entirety.
129: *
130: * Pool caches are layed on top of pools. By layering them, we can avoid
131: * the complexity of cache management for pools which would not benefit
132: * from it.
133: */
134:
135: /* The cache group pool. */
136: static struct pool pcgpool;
137:
138: /* The pool cache group. */
139: #define PCG_NOBJECTS 16
140: struct pool_cache_group {
141: TAILQ_ENTRY(pool_cache_group)
142: pcg_list; /* link in the pool cache's group list */
143: u_int pcg_avail; /* # available objects */
144: /* pointers to the objects */
145: void *pcg_objects[PCG_NOBJECTS];
146: };
1.3 pk 147:
1.43 thorpej 148: static void pool_cache_reclaim(struct pool_cache *);
1.3 pk 149:
1.42 thorpej 150: static int pool_catchup(struct pool *);
1.50 enami 151: static int pool_prime_page(struct pool *, caddr_t, int);
1.42 thorpej 152: static void *pool_page_alloc(unsigned long, int, int);
153: static void pool_page_free(void *, unsigned long, int);
1.3 pk 154:
1.42 thorpej 155: static void pool_print1(struct pool *, const char *,
156: void (*)(const char *, ...));
1.3 pk 157:
158: /*
1.52 thorpej 159: * Pool log entry. An array of these is allocated in pool_init().
1.3 pk 160: */
161: struct pool_log {
162: const char *pl_file;
163: long pl_line;
164: int pl_action;
1.25 thorpej 165: #define PRLOG_GET 1
166: #define PRLOG_PUT 2
1.3 pk 167: void *pl_addr;
1.1 pk 168: };
169:
1.3 pk 170: /* Number of entries in pool log buffers */
1.17 thorpej 171: #ifndef POOL_LOGSIZE
172: #define POOL_LOGSIZE 10
173: #endif
174:
175: int pool_logsize = POOL_LOGSIZE;
1.1 pk 176:
1.25 thorpej 177: #ifdef DIAGNOSTIC
1.42 thorpej 178: static __inline void
179: pr_log(struct pool *pp, void *v, int action, const char *file, long line)
1.3 pk 180: {
181: int n = pp->pr_curlogentry;
182: struct pool_log *pl;
183:
1.20 thorpej 184: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 185: return;
186:
187: /*
188: * Fill in the current entry. Wrap around and overwrite
189: * the oldest entry if necessary.
190: */
191: pl = &pp->pr_log[n];
192: pl->pl_file = file;
193: pl->pl_line = line;
194: pl->pl_action = action;
195: pl->pl_addr = v;
196: if (++n >= pp->pr_logsize)
197: n = 0;
198: pp->pr_curlogentry = n;
199: }
200:
201: static void
1.42 thorpej 202: pr_printlog(struct pool *pp, struct pool_item *pi,
203: void (*pr)(const char *, ...))
1.3 pk 204: {
205: int i = pp->pr_logsize;
206: int n = pp->pr_curlogentry;
207:
1.20 thorpej 208: if ((pp->pr_roflags & PR_LOGGING) == 0)
1.3 pk 209: return;
210:
211: /*
212: * Print all entries in this pool's log.
213: */
214: while (i-- > 0) {
215: struct pool_log *pl = &pp->pr_log[n];
216: if (pl->pl_action != 0) {
1.25 thorpej 217: if (pi == NULL || pi == pl->pl_addr) {
218: (*pr)("\tlog entry %d:\n", i);
219: (*pr)("\t\taction = %s, addr = %p\n",
220: pl->pl_action == PRLOG_GET ? "get" : "put",
221: pl->pl_addr);
222: (*pr)("\t\tfile: %s at line %lu\n",
223: pl->pl_file, pl->pl_line);
224: }
1.3 pk 225: }
226: if (++n >= pp->pr_logsize)
227: n = 0;
228: }
229: }
1.25 thorpej 230:
1.42 thorpej 231: static __inline void
232: pr_enter(struct pool *pp, const char *file, long line)
1.25 thorpej 233: {
234:
1.34 thorpej 235: if (__predict_false(pp->pr_entered_file != NULL)) {
1.25 thorpej 236: printf("pool %s: reentrancy at file %s line %ld\n",
237: pp->pr_wchan, file, line);
238: printf(" previous entry at file %s line %ld\n",
239: pp->pr_entered_file, pp->pr_entered_line);
240: panic("pr_enter");
241: }
242:
243: pp->pr_entered_file = file;
244: pp->pr_entered_line = line;
245: }
246:
1.42 thorpej 247: static __inline void
248: pr_leave(struct pool *pp)
1.25 thorpej 249: {
250:
1.34 thorpej 251: if (__predict_false(pp->pr_entered_file == NULL)) {
1.25 thorpej 252: printf("pool %s not entered?\n", pp->pr_wchan);
253: panic("pr_leave");
254: }
255:
256: pp->pr_entered_file = NULL;
257: pp->pr_entered_line = 0;
258: }
259:
1.42 thorpej 260: static __inline void
261: pr_enter_check(struct pool *pp, void (*pr)(const char *, ...))
1.25 thorpej 262: {
263:
264: if (pp->pr_entered_file != NULL)
265: (*pr)("\n\tcurrently entered from file %s line %ld\n",
266: pp->pr_entered_file, pp->pr_entered_line);
267: }
1.3 pk 268: #else
1.25 thorpej 269: #define pr_log(pp, v, action, file, line)
270: #define pr_printlog(pp, pi, pr)
271: #define pr_enter(pp, file, line)
272: #define pr_leave(pp)
273: #define pr_enter_check(pp, pr)
274: #endif /* DIAGNOSTIC */
1.3 pk 275:
276: /*
277: * Return the pool page header based on page address.
278: */
1.42 thorpej 279: static __inline struct pool_item_header *
280: pr_find_pagehead(struct pool *pp, caddr_t page)
1.3 pk 281: {
282: struct pool_item_header *ph;
283:
1.20 thorpej 284: if ((pp->pr_roflags & PR_PHINPAGE) != 0)
1.3 pk 285: return ((struct pool_item_header *)(page + pp->pr_phoffset));
286:
287: for (ph = LIST_FIRST(&pp->pr_hashtab[PR_HASH_INDEX(pp, page)]);
288: ph != NULL;
289: ph = LIST_NEXT(ph, ph_hashlist)) {
290: if (ph->ph_page == page)
291: return (ph);
292: }
293: return (NULL);
294: }
295:
296: /*
297: * Remove a page from the pool.
298: */
1.42 thorpej 299: static __inline void
300: pr_rmpage(struct pool *pp, struct pool_item_header *ph)
1.3 pk 301: {
302:
303: /*
1.7 thorpej 304: * If the page was idle, decrement the idle page count.
1.3 pk 305: */
1.6 thorpej 306: if (ph->ph_nmissing == 0) {
307: #ifdef DIAGNOSTIC
308: if (pp->pr_nidle == 0)
309: panic("pr_rmpage: nidle inconsistent");
1.20 thorpej 310: if (pp->pr_nitems < pp->pr_itemsperpage)
311: panic("pr_rmpage: nitems inconsistent");
1.6 thorpej 312: #endif
313: pp->pr_nidle--;
314: }
1.7 thorpej 315:
1.20 thorpej 316: pp->pr_nitems -= pp->pr_itemsperpage;
317:
1.7 thorpej 318: /*
319: * Unlink a page from the pool and release it.
320: */
321: TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
322: (*pp->pr_free)(ph->ph_page, pp->pr_pagesz, pp->pr_mtype);
323: pp->pr_npages--;
324: pp->pr_npagefree++;
1.6 thorpej 325:
1.22 chs 326: if ((pp->pr_roflags & PR_PHINPAGE) == 0) {
1.27 pk 327: int s;
1.22 chs 328: LIST_REMOVE(ph, ph_hashlist);
1.27 pk 329: s = splhigh();
1.22 chs 330: pool_put(&phpool, ph);
1.27 pk 331: splx(s);
1.22 chs 332: }
333:
1.3 pk 334: if (pp->pr_curpage == ph) {
335: /*
336: * Find a new non-empty page header, if any.
337: * Start search from the page head, to increase the
338: * chance for "high water" pages to be freed.
339: */
340: for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
341: ph = TAILQ_NEXT(ph, ph_pagelist))
342: if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
343: break;
344:
345: pp->pr_curpage = ph;
1.21 thorpej 346: }
1.3 pk 347: }
348:
349: /*
350: * Initialize the given pool resource structure.
351: *
352: * We export this routine to allow other kernel parts to declare
353: * static pools that must be initialized before malloc() is available.
354: */
355: void
1.42 thorpej 356: pool_init(struct pool *pp, size_t size, u_int align, u_int ioff, int flags,
357: const char *wchan, size_t pagesz,
358: void *(*alloc)(unsigned long, int, int),
359: void (*release)(void *, unsigned long, int),
360: int mtype)
1.3 pk 361: {
1.16 briggs 362: int off, slack, i;
1.3 pk 363:
1.25 thorpej 364: #ifdef POOL_DIAGNOSTIC
365: /*
366: * Always log if POOL_DIAGNOSTIC is defined.
367: */
368: if (pool_logsize != 0)
369: flags |= PR_LOGGING;
370: #endif
371:
1.3 pk 372: /*
373: * Check arguments and construct default values.
374: */
1.36 pk 375: if (!powerof2(pagesz))
1.3 pk 376: panic("pool_init: page size invalid (%lx)\n", (u_long)pagesz);
377:
1.4 thorpej 378: if (alloc == NULL && release == NULL) {
1.3 pk 379: alloc = pool_page_alloc;
380: release = pool_page_free;
1.4 thorpej 381: pagesz = PAGE_SIZE; /* Rounds to PAGE_SIZE anyhow. */
382: } else if ((alloc != NULL && release != NULL) == 0) {
383: /* If you specifiy one, must specify both. */
384: panic("pool_init: must specify alloc and release together");
385: }
386:
1.3 pk 387: if (pagesz == 0)
388: pagesz = PAGE_SIZE;
389:
390: if (align == 0)
391: align = ALIGN(1);
1.14 thorpej 392:
393: if (size < sizeof(struct pool_item))
394: size = sizeof(struct pool_item);
1.3 pk 395:
1.35 pk 396: size = ALIGN(size);
1.43 thorpej 397: if (size > pagesz)
1.35 pk 398: panic("pool_init: pool item size (%lu) too large",
399: (u_long)size);
400:
1.3 pk 401: /*
402: * Initialize the pool structure.
403: */
404: TAILQ_INIT(&pp->pr_pagelist);
1.43 thorpej 405: TAILQ_INIT(&pp->pr_cachelist);
1.3 pk 406: pp->pr_curpage = NULL;
407: pp->pr_npages = 0;
408: pp->pr_minitems = 0;
409: pp->pr_minpages = 0;
410: pp->pr_maxpages = UINT_MAX;
1.20 thorpej 411: pp->pr_roflags = flags;
412: pp->pr_flags = 0;
1.35 pk 413: pp->pr_size = size;
1.3 pk 414: pp->pr_align = align;
415: pp->pr_wchan = wchan;
416: pp->pr_mtype = mtype;
417: pp->pr_alloc = alloc;
418: pp->pr_free = release;
419: pp->pr_pagesz = pagesz;
420: pp->pr_pagemask = ~(pagesz - 1);
421: pp->pr_pageshift = ffs(pagesz) - 1;
1.20 thorpej 422: pp->pr_nitems = 0;
423: pp->pr_nout = 0;
424: pp->pr_hardlimit = UINT_MAX;
425: pp->pr_hardlimit_warning = NULL;
1.31 thorpej 426: pp->pr_hardlimit_ratecap.tv_sec = 0;
427: pp->pr_hardlimit_ratecap.tv_usec = 0;
428: pp->pr_hardlimit_warning_last.tv_sec = 0;
429: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.3 pk 430:
431: /*
432: * Decide whether to put the page header off page to avoid
433: * wasting too large a part of the page. Off-page page headers
434: * go on a hash table, so we can match a returned item
435: * with its header based on the page address.
436: * We use 1/16 of the page size as the threshold (XXX: tune)
437: */
438: if (pp->pr_size < pagesz/16) {
439: /* Use the end of the page for the page header */
1.20 thorpej 440: pp->pr_roflags |= PR_PHINPAGE;
1.3 pk 441: pp->pr_phoffset = off =
442: pagesz - ALIGN(sizeof(struct pool_item_header));
1.2 pk 443: } else {
1.3 pk 444: /* The page header will be taken from our page header pool */
445: pp->pr_phoffset = 0;
446: off = pagesz;
1.16 briggs 447: for (i = 0; i < PR_HASHTABSIZE; i++) {
448: LIST_INIT(&pp->pr_hashtab[i]);
449: }
1.2 pk 450: }
1.1 pk 451:
1.3 pk 452: /*
453: * Alignment is to take place at `ioff' within the item. This means
454: * we must reserve up to `align - 1' bytes on the page to allow
455: * appropriate positioning of each item.
456: *
457: * Silently enforce `0 <= ioff < align'.
458: */
459: pp->pr_itemoffset = ioff = ioff % align;
460: pp->pr_itemsperpage = (off - ((align - ioff) % align)) / pp->pr_size;
1.43 thorpej 461: KASSERT(pp->pr_itemsperpage != 0);
1.3 pk 462:
463: /*
464: * Use the slack between the chunks and the page header
465: * for "cache coloring".
466: */
467: slack = off - pp->pr_itemsperpage * pp->pr_size;
468: pp->pr_maxcolor = (slack / align) * align;
469: pp->pr_curcolor = 0;
470:
471: pp->pr_nget = 0;
472: pp->pr_nfail = 0;
473: pp->pr_nput = 0;
474: pp->pr_npagealloc = 0;
475: pp->pr_npagefree = 0;
1.1 pk 476: pp->pr_hiwat = 0;
1.8 thorpej 477: pp->pr_nidle = 0;
1.3 pk 478:
1.25 thorpej 479: if (flags & PR_LOGGING) {
480: if (kmem_map == NULL ||
481: (pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log),
482: M_TEMP, M_NOWAIT)) == NULL)
1.20 thorpej 483: pp->pr_roflags &= ~PR_LOGGING;
1.3 pk 484: pp->pr_curlogentry = 0;
485: pp->pr_logsize = pool_logsize;
486: }
1.25 thorpej 487:
488: pp->pr_entered_file = NULL;
489: pp->pr_entered_line = 0;
1.3 pk 490:
1.21 thorpej 491: simple_lock_init(&pp->pr_slock);
1.1 pk 492:
1.3 pk 493: /*
1.43 thorpej 494: * Initialize private page header pool and cache magazine pool if we
495: * haven't done so yet.
1.23 thorpej 496: * XXX LOCKING.
1.3 pk 497: */
498: if (phpool.pr_size == 0) {
499: pool_init(&phpool, sizeof(struct pool_item_header), 0, 0,
1.43 thorpej 500: 0, "phpool", 0, 0, 0, 0);
501: pool_init(&pcgpool, sizeof(struct pool_cache_group), 0, 0,
502: 0, "pcgpool", 0, 0, 0, 0);
1.1 pk 503: }
504:
1.23 thorpej 505: /* Insert into the list of all pools. */
506: simple_lock(&pool_head_slock);
507: TAILQ_INSERT_TAIL(&pool_head, pp, pr_poollist);
508: simple_unlock(&pool_head_slock);
1.1 pk 509: }
510:
511: /*
512: * De-commision a pool resource.
513: */
514: void
1.42 thorpej 515: pool_destroy(struct pool *pp)
1.1 pk 516: {
1.3 pk 517: struct pool_item_header *ph;
1.43 thorpej 518: struct pool_cache *pc;
519:
520: /* Destroy all caches for this pool. */
521: while ((pc = TAILQ_FIRST(&pp->pr_cachelist)) != NULL)
522: pool_cache_destroy(pc);
1.3 pk 523:
524: #ifdef DIAGNOSTIC
1.20 thorpej 525: if (pp->pr_nout != 0) {
1.25 thorpej 526: pr_printlog(pp, NULL, printf);
1.20 thorpej 527: panic("pool_destroy: pool busy: still out: %u\n",
528: pp->pr_nout);
1.3 pk 529: }
530: #endif
1.1 pk 531:
1.3 pk 532: /* Remove all pages */
1.20 thorpej 533: if ((pp->pr_roflags & PR_STATIC) == 0)
1.3 pk 534: while ((ph = pp->pr_pagelist.tqh_first) != NULL)
535: pr_rmpage(pp, ph);
536:
537: /* Remove from global pool list */
1.23 thorpej 538: simple_lock(&pool_head_slock);
1.3 pk 539: TAILQ_REMOVE(&pool_head, pp, pr_poollist);
1.23 thorpej 540: /* XXX Only clear this if we were drainpp? */
1.3 pk 541: drainpp = NULL;
1.23 thorpej 542: simple_unlock(&pool_head_slock);
1.3 pk 543:
1.20 thorpej 544: if ((pp->pr_roflags & PR_LOGGING) != 0)
1.3 pk 545: free(pp->pr_log, M_TEMP);
1.2 pk 546:
1.20 thorpej 547: if (pp->pr_roflags & PR_FREEHEADER)
1.3 pk 548: free(pp, M_POOL);
1.1 pk 549: }
550:
551:
552: /*
1.3 pk 553: * Grab an item from the pool; must be called at appropriate spl level
1.1 pk 554: */
1.3 pk 555: void *
1.42 thorpej 556: _pool_get(struct pool *pp, int flags, const char *file, long line)
1.1 pk 557: {
558: void *v;
559: struct pool_item *pi;
1.3 pk 560: struct pool_item_header *ph;
1.1 pk 561:
1.2 pk 562: #ifdef DIAGNOSTIC
1.34 thorpej 563: if (__predict_false((pp->pr_roflags & PR_STATIC) &&
564: (flags & PR_MALLOCOK))) {
1.25 thorpej 565: pr_printlog(pp, NULL, printf);
1.2 pk 566: panic("pool_get: static");
1.3 pk 567: }
1.2 pk 568: #endif
569:
1.37 sommerfe 570: if (__predict_false(curproc == NULL && doing_shutdown == 0 &&
571: (flags & PR_WAITOK) != 0))
1.3 pk 572: panic("pool_get: must have NOWAIT");
1.1 pk 573:
1.21 thorpej 574: simple_lock(&pp->pr_slock);
1.25 thorpej 575: pr_enter(pp, file, line);
1.20 thorpej 576:
577: startover:
578: /*
579: * Check to see if we've reached the hard limit. If we have,
580: * and we can wait, then wait until an item has been returned to
581: * the pool.
582: */
583: #ifdef DIAGNOSTIC
1.34 thorpej 584: if (__predict_false(pp->pr_nout > pp->pr_hardlimit)) {
1.25 thorpej 585: pr_leave(pp);
1.21 thorpej 586: simple_unlock(&pp->pr_slock);
1.20 thorpej 587: panic("pool_get: %s: crossed hard limit", pp->pr_wchan);
588: }
589: #endif
1.34 thorpej 590: if (__predict_false(pp->pr_nout == pp->pr_hardlimit)) {
1.29 sommerfe 591: if ((flags & PR_WAITOK) && !(flags & PR_LIMITFAIL)) {
1.20 thorpej 592: /*
593: * XXX: A warning isn't logged in this case. Should
594: * it be?
595: */
596: pp->pr_flags |= PR_WANTED;
1.25 thorpej 597: pr_leave(pp);
1.40 sommerfe 598: ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
1.25 thorpej 599: pr_enter(pp, file, line);
1.20 thorpej 600: goto startover;
601: }
1.31 thorpej 602:
603: /*
604: * Log a message that the hard limit has been hit.
605: */
606: if (pp->pr_hardlimit_warning != NULL &&
607: ratecheck(&pp->pr_hardlimit_warning_last,
608: &pp->pr_hardlimit_ratecap))
609: log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning);
1.21 thorpej 610:
611: if (flags & PR_URGENT)
612: panic("pool_get: urgent");
613:
614: pp->pr_nfail++;
615:
1.25 thorpej 616: pr_leave(pp);
1.21 thorpej 617: simple_unlock(&pp->pr_slock);
1.20 thorpej 618: return (NULL);
619: }
620:
1.3 pk 621: /*
622: * The convention we use is that if `curpage' is not NULL, then
623: * it points at a non-empty bucket. In particular, `curpage'
624: * never points at a page header which has PR_PHINPAGE set and
625: * has no items in its bucket.
626: */
1.20 thorpej 627: if ((ph = pp->pr_curpage) == NULL) {
1.15 pk 628: void *v;
629:
1.20 thorpej 630: #ifdef DIAGNOSTIC
631: if (pp->pr_nitems != 0) {
1.21 thorpej 632: simple_unlock(&pp->pr_slock);
1.20 thorpej 633: printf("pool_get: %s: curpage NULL, nitems %u\n",
634: pp->pr_wchan, pp->pr_nitems);
635: panic("pool_get: nitems inconsistent\n");
636: }
637: #endif
638:
1.21 thorpej 639: /*
640: * Call the back-end page allocator for more memory.
641: * Release the pool lock, as the back-end page allocator
642: * may block.
643: */
1.25 thorpej 644: pr_leave(pp);
1.21 thorpej 645: simple_unlock(&pp->pr_slock);
646: v = (*pp->pr_alloc)(pp->pr_pagesz, flags, pp->pr_mtype);
647: simple_lock(&pp->pr_slock);
1.25 thorpej 648: pr_enter(pp, file, line);
1.15 pk 649:
1.21 thorpej 650: if (v == NULL) {
651: /*
652: * We were unable to allocate a page, but
653: * we released the lock during allocation,
654: * so perhaps items were freed back to the
655: * pool. Check for this case.
656: */
657: if (pp->pr_curpage != NULL)
658: goto startover;
1.15 pk 659:
1.3 pk 660: if (flags & PR_URGENT)
661: panic("pool_get: urgent");
1.21 thorpej 662:
1.3 pk 663: if ((flags & PR_WAITOK) == 0) {
664: pp->pr_nfail++;
1.25 thorpej 665: pr_leave(pp);
1.21 thorpej 666: simple_unlock(&pp->pr_slock);
1.1 pk 667: return (NULL);
1.3 pk 668: }
669:
1.15 pk 670: /*
671: * Wait for items to be returned to this pool.
1.21 thorpej 672: *
1.15 pk 673: * XXX: we actually want to wait just until
674: * the page allocator has memory again. Depending
675: * on this pool's usage, we might get stuck here
676: * for a long time.
1.20 thorpej 677: *
678: * XXX: maybe we should wake up once a second and
679: * try again?
1.15 pk 680: */
1.1 pk 681: pp->pr_flags |= PR_WANTED;
1.25 thorpej 682: pr_leave(pp);
1.40 sommerfe 683: ltsleep(pp, PSWP, pp->pr_wchan, 0, &pp->pr_slock);
1.25 thorpej 684: pr_enter(pp, file, line);
1.20 thorpej 685: goto startover;
1.1 pk 686: }
1.3 pk 687:
1.15 pk 688: /* We have more memory; add it to the pool */
1.50 enami 689: if (pool_prime_page(pp, v, flags & PR_WAITOK) != 0) {
690: /*
691: * Probably, we don't allowed to wait and
692: * couldn't allocate a page header.
693: */
694: (*pp->pr_free)(v, pp->pr_pagesz, pp->pr_mtype);
695: pp->pr_nfail++;
696: pr_leave(pp);
697: simple_unlock(&pp->pr_slock);
698: return (NULL);
699: }
1.15 pk 700: pp->pr_npagealloc++;
701:
1.20 thorpej 702: /* Start the allocation process over. */
703: goto startover;
1.3 pk 704: }
705:
1.34 thorpej 706: if (__predict_false((v = pi = TAILQ_FIRST(&ph->ph_itemlist)) == NULL)) {
1.25 thorpej 707: pr_leave(pp);
1.21 thorpej 708: simple_unlock(&pp->pr_slock);
1.3 pk 709: panic("pool_get: %s: page empty", pp->pr_wchan);
1.21 thorpej 710: }
1.20 thorpej 711: #ifdef DIAGNOSTIC
1.34 thorpej 712: if (__predict_false(pp->pr_nitems == 0)) {
1.25 thorpej 713: pr_leave(pp);
1.21 thorpej 714: simple_unlock(&pp->pr_slock);
1.20 thorpej 715: printf("pool_get: %s: items on itemlist, nitems %u\n",
716: pp->pr_wchan, pp->pr_nitems);
717: panic("pool_get: nitems inconsistent\n");
718: }
719: #endif
1.3 pk 720: pr_log(pp, v, PRLOG_GET, file, line);
721:
722: #ifdef DIAGNOSTIC
1.34 thorpej 723: if (__predict_false(pi->pi_magic != PI_MAGIC)) {
1.25 thorpej 724: pr_printlog(pp, pi, printf);
1.3 pk 725: panic("pool_get(%s): free list modified: magic=%x; page %p;"
726: " item addr %p\n",
727: pp->pr_wchan, pi->pi_magic, ph->ph_page, pi);
728: }
729: #endif
730:
731: /*
732: * Remove from item list.
733: */
734: TAILQ_REMOVE(&ph->ph_itemlist, pi, pi_list);
1.20 thorpej 735: pp->pr_nitems--;
736: pp->pr_nout++;
1.6 thorpej 737: if (ph->ph_nmissing == 0) {
738: #ifdef DIAGNOSTIC
1.34 thorpej 739: if (__predict_false(pp->pr_nidle == 0))
1.6 thorpej 740: panic("pool_get: nidle inconsistent");
741: #endif
742: pp->pr_nidle--;
743: }
1.3 pk 744: ph->ph_nmissing++;
745: if (TAILQ_FIRST(&ph->ph_itemlist) == NULL) {
1.21 thorpej 746: #ifdef DIAGNOSTIC
1.34 thorpej 747: if (__predict_false(ph->ph_nmissing != pp->pr_itemsperpage)) {
1.25 thorpej 748: pr_leave(pp);
1.21 thorpej 749: simple_unlock(&pp->pr_slock);
750: panic("pool_get: %s: nmissing inconsistent",
751: pp->pr_wchan);
752: }
753: #endif
1.3 pk 754: /*
755: * Find a new non-empty page header, if any.
756: * Start search from the page head, to increase
757: * the chance for "high water" pages to be freed.
758: *
1.21 thorpej 759: * Migrate empty pages to the end of the list. This
760: * will speed the update of curpage as pages become
761: * idle. Empty pages intermingled with idle pages
762: * is no big deal. As soon as a page becomes un-empty,
763: * it will move back to the head of the list.
1.3 pk 764: */
765: TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
1.21 thorpej 766: TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist);
767: for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
768: ph = TAILQ_NEXT(ph, ph_pagelist))
1.3 pk 769: if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
770: break;
771:
772: pp->pr_curpage = ph;
1.1 pk 773: }
1.3 pk 774:
775: pp->pr_nget++;
1.20 thorpej 776:
777: /*
778: * If we have a low water mark and we are now below that low
779: * water mark, add more items to the pool.
780: */
1.53 thorpej 781: if (POOL_NEEDS_CATCHUP(pp) && pool_catchup(pp) != 0) {
1.20 thorpej 782: /*
783: * XXX: Should we log a warning? Should we set up a timeout
784: * to try again in a second or so? The latter could break
785: * a caller's assumptions about interrupt protection, etc.
786: */
787: }
788:
1.25 thorpej 789: pr_leave(pp);
1.21 thorpej 790: simple_unlock(&pp->pr_slock);
1.1 pk 791: return (v);
792: }
793:
794: /*
1.43 thorpej 795: * Internal version of pool_put(). Pool is already locked/entered.
1.1 pk 796: */
1.43 thorpej 797: static void
798: pool_do_put(struct pool *pp, void *v, const char *file, long line)
1.1 pk 799: {
800: struct pool_item *pi = v;
1.3 pk 801: struct pool_item_header *ph;
802: caddr_t page;
1.21 thorpej 803: int s;
1.3 pk 804:
805: page = (caddr_t)((u_long)v & pp->pr_pagemask);
1.1 pk 806:
1.30 thorpej 807: #ifdef DIAGNOSTIC
1.34 thorpej 808: if (__predict_false(pp->pr_nout == 0)) {
1.30 thorpej 809: printf("pool %s: putting with none out\n",
810: pp->pr_wchan);
811: panic("pool_put");
812: }
813: #endif
1.3 pk 814:
815: pr_log(pp, v, PRLOG_PUT, file, line);
816:
1.34 thorpej 817: if (__predict_false((ph = pr_find_pagehead(pp, page)) == NULL)) {
1.25 thorpej 818: pr_printlog(pp, NULL, printf);
1.3 pk 819: panic("pool_put: %s: page header missing", pp->pr_wchan);
820: }
1.28 thorpej 821:
822: #ifdef LOCKDEBUG
823: /*
824: * Check if we're freeing a locked simple lock.
825: */
826: simple_lock_freecheck((caddr_t)pi, ((caddr_t)pi) + pp->pr_size);
827: #endif
1.3 pk 828:
829: /*
830: * Return to item list.
831: */
1.2 pk 832: #ifdef DIAGNOSTIC
1.3 pk 833: pi->pi_magic = PI_MAGIC;
834: #endif
1.32 chs 835: #ifdef DEBUG
836: {
837: int i, *ip = v;
838:
839: for (i = 0; i < pp->pr_size / sizeof(int); i++) {
840: *ip++ = PI_MAGIC;
841: }
842: }
843: #endif
844:
1.3 pk 845: TAILQ_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list);
846: ph->ph_nmissing--;
847: pp->pr_nput++;
1.20 thorpej 848: pp->pr_nitems++;
849: pp->pr_nout--;
1.3 pk 850:
851: /* Cancel "pool empty" condition if it exists */
852: if (pp->pr_curpage == NULL)
853: pp->pr_curpage = ph;
854:
855: if (pp->pr_flags & PR_WANTED) {
856: pp->pr_flags &= ~PR_WANTED;
1.15 pk 857: if (ph->ph_nmissing == 0)
858: pp->pr_nidle++;
1.3 pk 859: wakeup((caddr_t)pp);
860: return;
861: }
862:
863: /*
1.21 thorpej 864: * If this page is now complete, do one of two things:
865: *
866: * (1) If we have more pages than the page high water
867: * mark, free the page back to the system.
868: *
869: * (2) Move it to the end of the page list, so that
870: * we minimize our chances of fragmenting the
871: * pool. Idle pages migrate to the end (along with
872: * completely empty pages, so that we find un-empty
873: * pages more quickly when we update curpage) of the
874: * list so they can be more easily swept up by
875: * the pagedaemon when pages are scarce.
1.3 pk 876: */
877: if (ph->ph_nmissing == 0) {
1.6 thorpej 878: pp->pr_nidle++;
1.3 pk 879: if (pp->pr_npages > pp->pr_maxpages) {
880: pr_rmpage(pp, ph);
881: } else {
882: TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
883: TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist);
884:
1.21 thorpej 885: /*
886: * Update the timestamp on the page. A page must
887: * be idle for some period of time before it can
888: * be reclaimed by the pagedaemon. This minimizes
889: * ping-pong'ing for memory.
890: */
891: s = splclock();
892: ph->ph_time = mono_time;
893: splx(s);
894:
895: /*
896: * Update the current page pointer. Just look for
897: * the first page with any free items.
898: *
899: * XXX: Maybe we want an option to look for the
900: * page with the fewest available items, to minimize
901: * fragmentation?
902: */
1.3 pk 903: for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
904: ph = TAILQ_NEXT(ph, ph_pagelist))
905: if (TAILQ_FIRST(&ph->ph_itemlist) != NULL)
906: break;
1.1 pk 907:
1.3 pk 908: pp->pr_curpage = ph;
1.1 pk 909: }
910: }
1.21 thorpej 911: /*
912: * If the page has just become un-empty, move it to the head of
913: * the list, and make it the current page. The next allocation
914: * will get the item from this page, instead of further fragmenting
915: * the pool.
916: */
917: else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) {
918: TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist);
919: TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist);
920: pp->pr_curpage = ph;
921: }
1.43 thorpej 922: }
923:
924: /*
925: * Return resource to the pool; must be called at appropriate spl level
926: */
927: void
928: _pool_put(struct pool *pp, void *v, const char *file, long line)
929: {
930:
931: simple_lock(&pp->pr_slock);
932: pr_enter(pp, file, line);
933:
934: pool_do_put(pp, v, file, line);
1.21 thorpej 935:
1.25 thorpej 936: pr_leave(pp);
1.21 thorpej 937: simple_unlock(&pp->pr_slock);
1.1 pk 938: }
939:
940: /*
1.3 pk 941: * Add a page worth of items to the pool.
1.21 thorpej 942: *
943: * Note, we must be called with the pool descriptor LOCKED.
1.3 pk 944: */
1.50 enami 945: static int
946: pool_prime_page(struct pool *pp, caddr_t storage, int flags)
1.3 pk 947: {
948: struct pool_item *pi;
949: struct pool_item_header *ph;
950: caddr_t cp = storage;
951: unsigned int align = pp->pr_align;
952: unsigned int ioff = pp->pr_itemoffset;
1.27 pk 953: int s, n;
1.36 pk 954:
955: if (((u_long)cp & (pp->pr_pagesz - 1)) != 0)
956: panic("pool_prime_page: %s: unaligned page", pp->pr_wchan);
1.3 pk 957:
1.20 thorpej 958: if ((pp->pr_roflags & PR_PHINPAGE) != 0) {
1.3 pk 959: ph = (struct pool_item_header *)(cp + pp->pr_phoffset);
960: } else {
1.27 pk 961: s = splhigh();
1.50 enami 962: ph = pool_get(&phpool, flags);
1.27 pk 963: splx(s);
1.50 enami 964: if (ph == NULL)
965: return (ENOMEM);
1.3 pk 966: LIST_INSERT_HEAD(&pp->pr_hashtab[PR_HASH_INDEX(pp, cp)],
967: ph, ph_hashlist);
968: }
969:
970: /*
971: * Insert page header.
972: */
973: TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist);
974: TAILQ_INIT(&ph->ph_itemlist);
975: ph->ph_page = storage;
976: ph->ph_nmissing = 0;
1.21 thorpej 977: memset(&ph->ph_time, 0, sizeof(ph->ph_time));
1.3 pk 978:
1.6 thorpej 979: pp->pr_nidle++;
980:
1.3 pk 981: /*
982: * Color this page.
983: */
984: cp = (caddr_t)(cp + pp->pr_curcolor);
985: if ((pp->pr_curcolor += align) > pp->pr_maxcolor)
986: pp->pr_curcolor = 0;
987:
988: /*
989: * Adjust storage to apply aligment to `pr_itemoffset' in each item.
990: */
991: if (ioff != 0)
992: cp = (caddr_t)(cp + (align - ioff));
993:
994: /*
995: * Insert remaining chunks on the bucket list.
996: */
997: n = pp->pr_itemsperpage;
1.20 thorpej 998: pp->pr_nitems += n;
1.3 pk 999:
1000: while (n--) {
1001: pi = (struct pool_item *)cp;
1002:
1003: /* Insert on page list */
1004: TAILQ_INSERT_TAIL(&ph->ph_itemlist, pi, pi_list);
1005: #ifdef DIAGNOSTIC
1006: pi->pi_magic = PI_MAGIC;
1007: #endif
1008: cp = (caddr_t)(cp + pp->pr_size);
1009: }
1010:
1011: /*
1012: * If the pool was depleted, point at the new page.
1013: */
1014: if (pp->pr_curpage == NULL)
1015: pp->pr_curpage = ph;
1016:
1017: if (++pp->pr_npages > pp->pr_hiwat)
1018: pp->pr_hiwat = pp->pr_npages;
1.50 enami 1019:
1020: return (0);
1.3 pk 1021: }
1022:
1.20 thorpej 1023: /*
1.52 thorpej 1024: * Used by pool_get() when nitems drops below the low water mark. This
1025: * is used to catch up nitmes with the low water mark.
1.20 thorpej 1026: *
1.21 thorpej 1027: * Note 1, we never wait for memory here, we let the caller decide what to do.
1.20 thorpej 1028: *
1029: * Note 2, this doesn't work with static pools.
1030: *
1031: * Note 3, we must be called with the pool already locked, and we return
1032: * with it locked.
1033: */
1034: static int
1.42 thorpej 1035: pool_catchup(struct pool *pp)
1.20 thorpej 1036: {
1037: caddr_t cp;
1038: int error = 0;
1039:
1040: if (pp->pr_roflags & PR_STATIC) {
1041: /*
1042: * We dropped below the low water mark, and this is not a
1043: * good thing. Log a warning.
1.21 thorpej 1044: *
1045: * XXX: rate-limit this?
1.20 thorpej 1046: */
1047: printf("WARNING: static pool `%s' dropped below low water "
1048: "mark\n", pp->pr_wchan);
1049: return (0);
1050: }
1051:
1.54 ! thorpej 1052: while (POOL_NEEDS_CATCHUP(pp)) {
1.20 thorpej 1053: /*
1.21 thorpej 1054: * Call the page back-end allocator for more memory.
1055: *
1056: * XXX: We never wait, so should we bother unlocking
1057: * the pool descriptor?
1.20 thorpej 1058: */
1.21 thorpej 1059: simple_unlock(&pp->pr_slock);
1.20 thorpej 1060: cp = (*pp->pr_alloc)(pp->pr_pagesz, 0, pp->pr_mtype);
1.21 thorpej 1061: simple_lock(&pp->pr_slock);
1.34 thorpej 1062: if (__predict_false(cp == NULL)) {
1.20 thorpej 1063: error = ENOMEM;
1064: break;
1065: }
1.50 enami 1066: if ((error = pool_prime_page(pp, cp, PR_NOWAIT)) != 0) {
1067: (*pp->pr_free)(cp, pp->pr_pagesz, pp->pr_mtype);
1068: break;
1069: }
1.26 thorpej 1070: pp->pr_npagealloc++;
1.20 thorpej 1071: }
1072:
1073: return (error);
1074: }
1075:
1.3 pk 1076: void
1.42 thorpej 1077: pool_setlowat(struct pool *pp, int n)
1.3 pk 1078: {
1.20 thorpej 1079: int error;
1.15 pk 1080:
1.21 thorpej 1081: simple_lock(&pp->pr_slock);
1082:
1.3 pk 1083: pp->pr_minitems = n;
1.15 pk 1084: pp->pr_minpages = (n == 0)
1085: ? 0
1.18 thorpej 1086: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.20 thorpej 1087:
1088: /* Make sure we're caught up with the newly-set low water mark. */
1.53 thorpej 1089: if (POOL_NEEDS_CATCHUP(pp) && (error = pool_catchup(pp) != 0)) {
1.20 thorpej 1090: /*
1091: * XXX: Should we log a warning? Should we set up a timeout
1092: * to try again in a second or so? The latter could break
1093: * a caller's assumptions about interrupt protection, etc.
1094: */
1095: }
1.21 thorpej 1096:
1097: simple_unlock(&pp->pr_slock);
1.3 pk 1098: }
1099:
1100: void
1.42 thorpej 1101: pool_sethiwat(struct pool *pp, int n)
1.3 pk 1102: {
1.15 pk 1103:
1.21 thorpej 1104: simple_lock(&pp->pr_slock);
1105:
1.15 pk 1106: pp->pr_maxpages = (n == 0)
1107: ? 0
1.18 thorpej 1108: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1109:
1110: simple_unlock(&pp->pr_slock);
1.3 pk 1111: }
1112:
1.20 thorpej 1113: void
1.42 thorpej 1114: pool_sethardlimit(struct pool *pp, int n, const char *warnmess, int ratecap)
1.20 thorpej 1115: {
1116:
1.21 thorpej 1117: simple_lock(&pp->pr_slock);
1.20 thorpej 1118:
1119: pp->pr_hardlimit = n;
1120: pp->pr_hardlimit_warning = warnmess;
1.31 thorpej 1121: pp->pr_hardlimit_ratecap.tv_sec = ratecap;
1122: pp->pr_hardlimit_warning_last.tv_sec = 0;
1123: pp->pr_hardlimit_warning_last.tv_usec = 0;
1.20 thorpej 1124:
1125: /*
1.21 thorpej 1126: * In-line version of pool_sethiwat(), because we don't want to
1127: * release the lock.
1.20 thorpej 1128: */
1129: pp->pr_maxpages = (n == 0)
1130: ? 0
1131: : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage;
1.21 thorpej 1132:
1133: simple_unlock(&pp->pr_slock);
1.20 thorpej 1134: }
1.3 pk 1135:
1136: /*
1137: * Default page allocator.
1138: */
1139: static void *
1.42 thorpej 1140: pool_page_alloc(unsigned long sz, int flags, int mtype)
1.3 pk 1141: {
1.11 thorpej 1142: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
1.3 pk 1143:
1.11 thorpej 1144: return ((void *)uvm_km_alloc_poolpage(waitok));
1.3 pk 1145: }
1146:
1147: static void
1.42 thorpej 1148: pool_page_free(void *v, unsigned long sz, int mtype)
1.3 pk 1149: {
1150:
1.10 eeh 1151: uvm_km_free_poolpage((vaddr_t)v);
1.3 pk 1152: }
1.12 thorpej 1153:
1154: /*
1155: * Alternate pool page allocator for pools that know they will
1156: * never be accessed in interrupt context.
1157: */
1158: void *
1.42 thorpej 1159: pool_page_alloc_nointr(unsigned long sz, int flags, int mtype)
1.12 thorpej 1160: {
1161: boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE;
1162:
1163: return ((void *)uvm_km_alloc_poolpage1(kernel_map, uvm.kernel_object,
1164: waitok));
1165: }
1166:
1167: void
1.42 thorpej 1168: pool_page_free_nointr(void *v, unsigned long sz, int mtype)
1.12 thorpej 1169: {
1170:
1171: uvm_km_free_poolpage1(kernel_map, (vaddr_t)v);
1172: }
1173:
1.3 pk 1174:
1175: /*
1176: * Release all complete pages that have not been used recently.
1177: */
1178: void
1.42 thorpej 1179: _pool_reclaim(struct pool *pp, const char *file, long line)
1.3 pk 1180: {
1181: struct pool_item_header *ph, *phnext;
1.43 thorpej 1182: struct pool_cache *pc;
1.21 thorpej 1183: struct timeval curtime;
1184: int s;
1.3 pk 1185:
1.20 thorpej 1186: if (pp->pr_roflags & PR_STATIC)
1.3 pk 1187: return;
1188:
1.21 thorpej 1189: if (simple_lock_try(&pp->pr_slock) == 0)
1.3 pk 1190: return;
1.25 thorpej 1191: pr_enter(pp, file, line);
1.3 pk 1192:
1.43 thorpej 1193: /*
1194: * Reclaim items from the pool's caches.
1195: */
1196: for (pc = TAILQ_FIRST(&pp->pr_cachelist); pc != NULL;
1197: pc = TAILQ_NEXT(pc, pc_poollist))
1198: pool_cache_reclaim(pc);
1199:
1.21 thorpej 1200: s = splclock();
1201: curtime = mono_time;
1202: splx(s);
1203:
1.3 pk 1204: for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; ph = phnext) {
1205: phnext = TAILQ_NEXT(ph, ph_pagelist);
1206:
1207: /* Check our minimum page claim */
1208: if (pp->pr_npages <= pp->pr_minpages)
1209: break;
1210:
1211: if (ph->ph_nmissing == 0) {
1212: struct timeval diff;
1213: timersub(&curtime, &ph->ph_time, &diff);
1214: if (diff.tv_sec < pool_inactive_time)
1215: continue;
1.21 thorpej 1216:
1217: /*
1218: * If freeing this page would put us below
1219: * the low water mark, stop now.
1220: */
1221: if ((pp->pr_nitems - pp->pr_itemsperpage) <
1222: pp->pr_minitems)
1223: break;
1224:
1.3 pk 1225: pr_rmpage(pp, ph);
1226: }
1227: }
1228:
1.25 thorpej 1229: pr_leave(pp);
1.21 thorpej 1230: simple_unlock(&pp->pr_slock);
1.3 pk 1231: }
1232:
1233:
1234: /*
1235: * Drain pools, one at a time.
1.21 thorpej 1236: *
1237: * Note, we must never be called from an interrupt context.
1.3 pk 1238: */
1239: void
1.42 thorpej 1240: pool_drain(void *arg)
1.3 pk 1241: {
1242: struct pool *pp;
1.23 thorpej 1243: int s;
1.3 pk 1244:
1.49 thorpej 1245: s = splvm();
1.23 thorpej 1246: simple_lock(&pool_head_slock);
1247:
1248: if (drainpp == NULL && (drainpp = TAILQ_FIRST(&pool_head)) == NULL)
1249: goto out;
1.3 pk 1250:
1251: pp = drainpp;
1252: drainpp = TAILQ_NEXT(pp, pr_poollist);
1253:
1254: pool_reclaim(pp);
1.23 thorpej 1255:
1256: out:
1257: simple_unlock(&pool_head_slock);
1.3 pk 1258: splx(s);
1259: }
1260:
1261:
1262: /*
1263: * Diagnostic helpers.
1264: */
1265: void
1.42 thorpej 1266: pool_print(struct pool *pp, const char *modif)
1.21 thorpej 1267: {
1268: int s;
1269:
1.49 thorpej 1270: s = splvm();
1.25 thorpej 1271: if (simple_lock_try(&pp->pr_slock) == 0) {
1272: printf("pool %s is locked; try again later\n",
1273: pp->pr_wchan);
1274: splx(s);
1275: return;
1276: }
1277: pool_print1(pp, modif, printf);
1.21 thorpej 1278: simple_unlock(&pp->pr_slock);
1279: splx(s);
1280: }
1281:
1.25 thorpej 1282: void
1.42 thorpej 1283: pool_printit(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.25 thorpej 1284: {
1285: int didlock = 0;
1286:
1287: if (pp == NULL) {
1288: (*pr)("Must specify a pool to print.\n");
1289: return;
1290: }
1291:
1292: /*
1293: * Called from DDB; interrupts should be blocked, and all
1294: * other processors should be paused. We can skip locking
1295: * the pool in this case.
1296: *
1297: * We do a simple_lock_try() just to print the lock
1298: * status, however.
1299: */
1300:
1301: if (simple_lock_try(&pp->pr_slock) == 0)
1302: (*pr)("WARNING: pool %s is locked\n", pp->pr_wchan);
1303: else
1304: didlock = 1;
1305:
1306: pool_print1(pp, modif, pr);
1307:
1308: if (didlock)
1309: simple_unlock(&pp->pr_slock);
1310: }
1311:
1.21 thorpej 1312: static void
1.42 thorpej 1313: pool_print1(struct pool *pp, const char *modif, void (*pr)(const char *, ...))
1.3 pk 1314: {
1.25 thorpej 1315: struct pool_item_header *ph;
1.44 thorpej 1316: struct pool_cache *pc;
1317: struct pool_cache_group *pcg;
1.25 thorpej 1318: #ifdef DIAGNOSTIC
1319: struct pool_item *pi;
1320: #endif
1.44 thorpej 1321: int i, print_log = 0, print_pagelist = 0, print_cache = 0;
1.25 thorpej 1322: char c;
1323:
1324: while ((c = *modif++) != '\0') {
1325: if (c == 'l')
1326: print_log = 1;
1327: if (c == 'p')
1328: print_pagelist = 1;
1.44 thorpej 1329: if (c == 'c')
1330: print_cache = 1;
1.25 thorpej 1331: modif++;
1332: }
1333:
1334: (*pr)("POOL %s: size %u, align %u, ioff %u, roflags 0x%08x\n",
1335: pp->pr_wchan, pp->pr_size, pp->pr_align, pp->pr_itemoffset,
1336: pp->pr_roflags);
1337: (*pr)("\tpagesz %u, mtype %d\n", pp->pr_pagesz, pp->pr_mtype);
1338: (*pr)("\talloc %p, release %p\n", pp->pr_alloc, pp->pr_free);
1339: (*pr)("\tminitems %u, minpages %u, maxpages %u, npages %u\n",
1340: pp->pr_minitems, pp->pr_minpages, pp->pr_maxpages, pp->pr_npages);
1341: (*pr)("\titemsperpage %u, nitems %u, nout %u, hardlimit %u\n",
1342: pp->pr_itemsperpage, pp->pr_nitems, pp->pr_nout, pp->pr_hardlimit);
1343:
1344: (*pr)("\n\tnget %lu, nfail %lu, nput %lu\n",
1345: pp->pr_nget, pp->pr_nfail, pp->pr_nput);
1346: (*pr)("\tnpagealloc %lu, npagefree %lu, hiwat %u, nidle %lu\n",
1347: pp->pr_npagealloc, pp->pr_npagefree, pp->pr_hiwat, pp->pr_nidle);
1348:
1349: if (print_pagelist == 0)
1350: goto skip_pagelist;
1351:
1352: if ((ph = TAILQ_FIRST(&pp->pr_pagelist)) != NULL)
1353: (*pr)("\n\tpage list:\n");
1354: for (; ph != NULL; ph = TAILQ_NEXT(ph, ph_pagelist)) {
1355: (*pr)("\t\tpage %p, nmissing %d, time %lu,%lu\n",
1356: ph->ph_page, ph->ph_nmissing,
1357: (u_long)ph->ph_time.tv_sec,
1358: (u_long)ph->ph_time.tv_usec);
1359: #ifdef DIAGNOSTIC
1360: for (pi = TAILQ_FIRST(&ph->ph_itemlist); pi != NULL;
1361: pi = TAILQ_NEXT(pi, pi_list)) {
1362: if (pi->pi_magic != PI_MAGIC) {
1363: (*pr)("\t\t\titem %p, magic 0x%x\n",
1364: pi, pi->pi_magic);
1365: }
1366: }
1367: #endif
1368: }
1369: if (pp->pr_curpage == NULL)
1370: (*pr)("\tno current page\n");
1371: else
1372: (*pr)("\tcurpage %p\n", pp->pr_curpage->ph_page);
1373:
1374: skip_pagelist:
1375:
1376: if (print_log == 0)
1377: goto skip_log;
1378:
1379: (*pr)("\n");
1380: if ((pp->pr_roflags & PR_LOGGING) == 0)
1381: (*pr)("\tno log\n");
1382: else
1383: pr_printlog(pp, NULL, pr);
1.3 pk 1384:
1.25 thorpej 1385: skip_log:
1.44 thorpej 1386:
1387: if (print_cache == 0)
1388: goto skip_cache;
1389:
1390: for (pc = TAILQ_FIRST(&pp->pr_cachelist); pc != NULL;
1391: pc = TAILQ_NEXT(pc, pc_poollist)) {
1392: (*pr)("\tcache %p: allocfrom %p freeto %p\n", pc,
1393: pc->pc_allocfrom, pc->pc_freeto);
1.48 thorpej 1394: (*pr)("\t hits %lu misses %lu ngroups %lu nitems %lu\n",
1395: pc->pc_hits, pc->pc_misses, pc->pc_ngroups, pc->pc_nitems);
1.44 thorpej 1396: for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
1397: pcg = TAILQ_NEXT(pcg, pcg_list)) {
1398: (*pr)("\t\tgroup %p: avail %d\n", pcg, pcg->pcg_avail);
1399: for (i = 0; i < PCG_NOBJECTS; i++)
1400: (*pr)("\t\t\t%p\n", pcg->pcg_objects[i]);
1401: }
1402: }
1403:
1404: skip_cache:
1.3 pk 1405:
1.25 thorpej 1406: pr_enter_check(pp, pr);
1.3 pk 1407: }
1408:
1409: int
1.42 thorpej 1410: pool_chk(struct pool *pp, const char *label)
1.3 pk 1411: {
1412: struct pool_item_header *ph;
1413: int r = 0;
1414:
1.21 thorpej 1415: simple_lock(&pp->pr_slock);
1.3 pk 1416:
1417: for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL;
1418: ph = TAILQ_NEXT(ph, ph_pagelist)) {
1419:
1420: struct pool_item *pi;
1421: int n;
1422: caddr_t page;
1423:
1424: page = (caddr_t)((u_long)ph & pp->pr_pagemask);
1.20 thorpej 1425: if (page != ph->ph_page &&
1426: (pp->pr_roflags & PR_PHINPAGE) != 0) {
1.3 pk 1427: if (label != NULL)
1428: printf("%s: ", label);
1.16 briggs 1429: printf("pool(%p:%s): page inconsistency: page %p;"
1430: " at page head addr %p (p %p)\n", pp,
1.3 pk 1431: pp->pr_wchan, ph->ph_page,
1432: ph, page);
1433: r++;
1434: goto out;
1435: }
1436:
1437: for (pi = TAILQ_FIRST(&ph->ph_itemlist), n = 0;
1438: pi != NULL;
1439: pi = TAILQ_NEXT(pi,pi_list), n++) {
1440:
1441: #ifdef DIAGNOSTIC
1442: if (pi->pi_magic != PI_MAGIC) {
1443: if (label != NULL)
1444: printf("%s: ", label);
1445: printf("pool(%s): free list modified: magic=%x;"
1446: " page %p; item ordinal %d;"
1447: " addr %p (p %p)\n",
1448: pp->pr_wchan, pi->pi_magic, ph->ph_page,
1449: n, pi, page);
1450: panic("pool");
1451: }
1452: #endif
1453: page = (caddr_t)((u_long)pi & pp->pr_pagemask);
1454: if (page == ph->ph_page)
1455: continue;
1456:
1457: if (label != NULL)
1458: printf("%s: ", label);
1.16 briggs 1459: printf("pool(%p:%s): page inconsistency: page %p;"
1460: " item ordinal %d; addr %p (p %p)\n", pp,
1.3 pk 1461: pp->pr_wchan, ph->ph_page,
1462: n, pi, page);
1463: r++;
1464: goto out;
1465: }
1466: }
1467: out:
1.21 thorpej 1468: simple_unlock(&pp->pr_slock);
1.3 pk 1469: return (r);
1.43 thorpej 1470: }
1471:
1472: /*
1473: * pool_cache_init:
1474: *
1475: * Initialize a pool cache.
1476: *
1477: * NOTE: If the pool must be protected from interrupts, we expect
1478: * to be called at the appropriate interrupt priority level.
1479: */
1480: void
1481: pool_cache_init(struct pool_cache *pc, struct pool *pp,
1482: int (*ctor)(void *, void *, int),
1483: void (*dtor)(void *, void *),
1484: void *arg)
1485: {
1486:
1487: TAILQ_INIT(&pc->pc_grouplist);
1488: simple_lock_init(&pc->pc_slock);
1489:
1490: pc->pc_allocfrom = NULL;
1491: pc->pc_freeto = NULL;
1492: pc->pc_pool = pp;
1493:
1494: pc->pc_ctor = ctor;
1495: pc->pc_dtor = dtor;
1496: pc->pc_arg = arg;
1497:
1.48 thorpej 1498: pc->pc_hits = 0;
1499: pc->pc_misses = 0;
1500:
1501: pc->pc_ngroups = 0;
1502:
1503: pc->pc_nitems = 0;
1504:
1.43 thorpej 1505: simple_lock(&pp->pr_slock);
1506: TAILQ_INSERT_TAIL(&pp->pr_cachelist, pc, pc_poollist);
1507: simple_unlock(&pp->pr_slock);
1508: }
1509:
1510: /*
1511: * pool_cache_destroy:
1512: *
1513: * Destroy a pool cache.
1514: */
1515: void
1516: pool_cache_destroy(struct pool_cache *pc)
1517: {
1518: struct pool *pp = pc->pc_pool;
1519:
1520: /* First, invalidate the entire cache. */
1521: pool_cache_invalidate(pc);
1522:
1523: /* ...and remove it from the pool's cache list. */
1524: simple_lock(&pp->pr_slock);
1525: TAILQ_REMOVE(&pp->pr_cachelist, pc, pc_poollist);
1526: simple_unlock(&pp->pr_slock);
1527: }
1528:
1529: static __inline void *
1530: pcg_get(struct pool_cache_group *pcg)
1531: {
1532: void *object;
1533: u_int idx;
1534:
1535: KASSERT(pcg->pcg_avail <= PCG_NOBJECTS);
1.45 thorpej 1536: KASSERT(pcg->pcg_avail != 0);
1.43 thorpej 1537: idx = --pcg->pcg_avail;
1538:
1539: KASSERT(pcg->pcg_objects[idx] != NULL);
1540: object = pcg->pcg_objects[idx];
1541: pcg->pcg_objects[idx] = NULL;
1542:
1543: return (object);
1544: }
1545:
1546: static __inline void
1547: pcg_put(struct pool_cache_group *pcg, void *object)
1548: {
1549: u_int idx;
1550:
1551: KASSERT(pcg->pcg_avail < PCG_NOBJECTS);
1552: idx = pcg->pcg_avail++;
1553:
1554: KASSERT(pcg->pcg_objects[idx] == NULL);
1555: pcg->pcg_objects[idx] = object;
1556: }
1557:
1558: /*
1559: * pool_cache_get:
1560: *
1561: * Get an object from a pool cache.
1562: */
1563: void *
1564: pool_cache_get(struct pool_cache *pc, int flags)
1565: {
1566: struct pool_cache_group *pcg;
1567: void *object;
1568:
1569: simple_lock(&pc->pc_slock);
1570:
1571: if ((pcg = pc->pc_allocfrom) == NULL) {
1572: for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
1573: pcg = TAILQ_NEXT(pcg, pcg_list)) {
1574: if (pcg->pcg_avail != 0) {
1575: pc->pc_allocfrom = pcg;
1576: goto have_group;
1577: }
1578: }
1579:
1580: /*
1581: * No groups with any available objects. Allocate
1582: * a new object, construct it, and return it to
1583: * the caller. We will allocate a group, if necessary,
1584: * when the object is freed back to the cache.
1585: */
1.48 thorpej 1586: pc->pc_misses++;
1.43 thorpej 1587: simple_unlock(&pc->pc_slock);
1588: object = pool_get(pc->pc_pool, flags);
1589: if (object != NULL && pc->pc_ctor != NULL) {
1590: if ((*pc->pc_ctor)(pc->pc_arg, object, flags) != 0) {
1591: pool_put(pc->pc_pool, object);
1592: return (NULL);
1593: }
1594: }
1595: return (object);
1596: }
1597:
1598: have_group:
1.48 thorpej 1599: pc->pc_hits++;
1600: pc->pc_nitems--;
1.43 thorpej 1601: object = pcg_get(pcg);
1602:
1603: if (pcg->pcg_avail == 0)
1604: pc->pc_allocfrom = NULL;
1.45 thorpej 1605:
1.43 thorpej 1606: simple_unlock(&pc->pc_slock);
1607:
1608: return (object);
1609: }
1610:
1611: /*
1612: * pool_cache_put:
1613: *
1614: * Put an object back to the pool cache.
1615: */
1616: void
1617: pool_cache_put(struct pool_cache *pc, void *object)
1618: {
1619: struct pool_cache_group *pcg;
1620:
1621: simple_lock(&pc->pc_slock);
1622:
1623: if ((pcg = pc->pc_freeto) == NULL) {
1624: for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
1625: pcg = TAILQ_NEXT(pcg, pcg_list)) {
1626: if (pcg->pcg_avail != PCG_NOBJECTS) {
1627: pc->pc_freeto = pcg;
1628: goto have_group;
1629: }
1630: }
1631:
1632: /*
1633: * No empty groups to free the object to. Attempt to
1.47 thorpej 1634: * allocate one.
1.43 thorpej 1635: */
1.47 thorpej 1636: simple_unlock(&pc->pc_slock);
1.43 thorpej 1637: pcg = pool_get(&pcgpool, PR_NOWAIT);
1638: if (pcg != NULL) {
1639: memset(pcg, 0, sizeof(*pcg));
1.47 thorpej 1640: simple_lock(&pc->pc_slock);
1.48 thorpej 1641: pc->pc_ngroups++;
1.43 thorpej 1642: TAILQ_INSERT_TAIL(&pc->pc_grouplist, pcg, pcg_list);
1.47 thorpej 1643: if (pc->pc_freeto == NULL)
1644: pc->pc_freeto = pcg;
1.43 thorpej 1645: goto have_group;
1646: }
1647:
1648: /*
1649: * Unable to allocate a cache group; destruct the object
1650: * and free it back to the pool.
1651: */
1.51 thorpej 1652: pool_cache_destruct_object(pc, object);
1.43 thorpej 1653: return;
1654: }
1655:
1656: have_group:
1.48 thorpej 1657: pc->pc_nitems++;
1.43 thorpej 1658: pcg_put(pcg, object);
1659:
1660: if (pcg->pcg_avail == PCG_NOBJECTS)
1661: pc->pc_freeto = NULL;
1662:
1663: simple_unlock(&pc->pc_slock);
1.51 thorpej 1664: }
1665:
1666: /*
1667: * pool_cache_destruct_object:
1668: *
1669: * Force destruction of an object and its release back into
1670: * the pool.
1671: */
1672: void
1673: pool_cache_destruct_object(struct pool_cache *pc, void *object)
1674: {
1675:
1676: if (pc->pc_dtor != NULL)
1677: (*pc->pc_dtor)(pc->pc_arg, object);
1678: pool_put(pc->pc_pool, object);
1.43 thorpej 1679: }
1680:
1681: /*
1682: * pool_cache_do_invalidate:
1683: *
1684: * This internal function implements pool_cache_invalidate() and
1685: * pool_cache_reclaim().
1686: */
1687: static void
1688: pool_cache_do_invalidate(struct pool_cache *pc, int free_groups,
1689: void (*putit)(struct pool *, void *, const char *, long))
1690: {
1691: struct pool_cache_group *pcg, *npcg;
1692: void *object;
1693:
1694: for (pcg = TAILQ_FIRST(&pc->pc_grouplist); pcg != NULL;
1695: pcg = npcg) {
1696: npcg = TAILQ_NEXT(pcg, pcg_list);
1697: while (pcg->pcg_avail != 0) {
1.48 thorpej 1698: pc->pc_nitems--;
1.43 thorpej 1699: object = pcg_get(pcg);
1.45 thorpej 1700: if (pcg->pcg_avail == 0 && pc->pc_allocfrom == pcg)
1701: pc->pc_allocfrom = NULL;
1.43 thorpej 1702: if (pc->pc_dtor != NULL)
1703: (*pc->pc_dtor)(pc->pc_arg, object);
1704: (*putit)(pc->pc_pool, object, __FILE__, __LINE__);
1705: }
1706: if (free_groups) {
1.48 thorpej 1707: pc->pc_ngroups--;
1.43 thorpej 1708: TAILQ_REMOVE(&pc->pc_grouplist, pcg, pcg_list);
1.46 thorpej 1709: if (pc->pc_freeto == pcg)
1710: pc->pc_freeto = NULL;
1.43 thorpej 1711: pool_put(&pcgpool, pcg);
1712: }
1713: }
1714: }
1715:
1716: /*
1717: * pool_cache_invalidate:
1718: *
1719: * Invalidate a pool cache (destruct and release all of the
1720: * cached objects).
1721: */
1722: void
1723: pool_cache_invalidate(struct pool_cache *pc)
1724: {
1725:
1726: simple_lock(&pc->pc_slock);
1727: pool_cache_do_invalidate(pc, 0, _pool_put);
1728: simple_unlock(&pc->pc_slock);
1729: }
1730:
1731: /*
1732: * pool_cache_reclaim:
1733: *
1734: * Reclaim a pool cache for pool_reclaim().
1735: */
1736: static void
1737: pool_cache_reclaim(struct pool_cache *pc)
1738: {
1739:
1.47 thorpej 1740: simple_lock(&pc->pc_slock);
1.43 thorpej 1741: pool_cache_do_invalidate(pc, 1, pool_do_put);
1742: simple_unlock(&pc->pc_slock);
1.3 pk 1743: }
CVSweb <webmaster@jp.NetBSD.org>
![[BACK]](/icons/back.gif){kind=icon}
Return to subr_pool.c CVS log ![[TXT]](/icons/text.gif){kind=icon}
![[DIR]](/icons/dir.gif){kind=icon}
Up to [cvs.NetBSD.org] / src / sys / kern