Please note that diffs are not public domain; they are subject to the copyright notices on the relevant files. =================================================================== RCS file: /ftp/cvs/cvsroot/src/sys/kern/subr_pool.c,v rcsdiff: /ftp/cvs/cvsroot/src/sys/kern/subr_pool.c,v: warning: Unknown phrases like `commitid ...;' are present. retrieving revision 1.15 retrieving revision 1.21 diff -u -p -r1.15 -r1.21 --- src/sys/kern/subr_pool.c 1998/09/29 18:09:29 1.15 +++ src/sys/kern/subr_pool.c 1999/03/31 23:23:48 1.21 @@ -1,11 +1,12 @@ -/* $NetBSD: subr_pool.c,v 1.15 1998/09/29 18:09:29 pk Exp $ */ +/* $NetBSD: subr_pool.c,v 1.21 1999/03/31 23:23:48 thorpej Exp $ */ /*- - * Copyright (c) 1997 The NetBSD Foundation, Inc. + * Copyright (c) 1997, 1999 The NetBSD Foundation, Inc. * All rights reserved. * * This code is derived from software contributed to The NetBSD Foundation - * by Paul Kranenburg. + * by Paul Kranenburg; by Jason R. Thorpe of the Numerical Aerospace + * Simulation Facility, NASA Ames Research Center. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions @@ -44,13 +45,12 @@ #include #include #include +#include #include #include -#if defined(UVM) #include -#endif /* * Pool resource management utility. @@ -61,7 +61,6 @@ * headed by `ph_itemlist' in each page header. The memory for building * the page list is either taken from the allocated pages themselves (for * small pool items) or taken from an internal pool of page headers (`phpool'). - * */ /* List of all pools */ @@ -106,11 +105,14 @@ struct pool_item { static struct pool_item_header *pr_find_pagehead __P((struct pool *, caddr_t)); static void pr_rmpage __P((struct pool *, struct pool_item_header *)); -static int pool_prime_page __P((struct pool *, caddr_t)); +static int pool_catchup __P((struct pool *)); +static void pool_prime_page __P((struct pool *, caddr_t)); static void *pool_page_alloc __P((unsigned long, int, int)); static void pool_page_free __P((void *, unsigned long, int)); -int pool_chk __P((struct pool *, char *)); +#if defined(POOL_DIAGNOSTIC) || defined(DEBUG) +static void pool_print1 __P((struct pool *, const char *)); +#endif #ifdef POOL_DIAGNOSTIC /* @@ -126,7 +128,11 @@ struct pool_log { }; /* Number of entries in pool log buffers */ -int pool_logsize = 10; +#ifndef POOL_LOGSIZE +#define POOL_LOGSIZE 10 +#endif + +int pool_logsize = POOL_LOGSIZE; static void pr_log __P((struct pool *, void *, int, const char *, long)); static void pr_printlog __P((struct pool *)); @@ -142,7 +148,7 @@ pr_log(pp, v, action, file, line) int n = pp->pr_curlogentry; struct pool_log *pl; - if ((pp->pr_flags & PR_LOGGING) == 0) + if ((pp->pr_roflags & PR_LOGGING) == 0) return; /* @@ -166,10 +172,10 @@ pr_printlog(pp) int i = pp->pr_logsize; int n = pp->pr_curlogentry; - if ((pp->pr_flags & PR_LOGGING) == 0) + if ((pp->pr_roflags & PR_LOGGING) == 0) return; - pool_print(pp, "printlog"); + pool_print1(pp, "printlog"); /* * Print all entries in this pool's log. @@ -204,7 +210,7 @@ pr_find_pagehead(pp, page) { struct pool_item_header *ph; - if ((pp->pr_flags & PR_PHINPAGE) != 0) + if ((pp->pr_roflags & PR_PHINPAGE) != 0) return ((struct pool_item_header *)(page + pp->pr_phoffset)); for (ph = LIST_FIRST(&pp->pr_hashtab[PR_HASH_INDEX(pp, page)]); @@ -232,10 +238,14 @@ pr_rmpage(pp, ph) #ifdef DIAGNOSTIC if (pp->pr_nidle == 0) panic("pr_rmpage: nidle inconsistent"); + if (pp->pr_nitems < pp->pr_itemsperpage) + panic("pr_rmpage: nitems inconsistent"); #endif pp->pr_nidle--; } + pp->pr_nitems -= pp->pr_itemsperpage; + /* * Unlink a page from the pool and release it. */ @@ -244,11 +254,6 @@ pr_rmpage(pp, ph) pp->pr_npages--; pp->pr_npagefree++; - if ((pp->pr_flags & PR_PHINPAGE) == 0) { - LIST_REMOVE(ph, ph_hashlist); - pool_put(&phpool, ph); - } - if (pp->pr_curpage == ph) { /* * Find a new non-empty page header, if any. @@ -262,6 +267,11 @@ pr_rmpage(pp, ph) pp->pr_curpage = ph; } + + if ((pp->pr_roflags & PR_PHINPAGE) == 0) { + LIST_REMOVE(ph, ph_hashlist); + pool_put(&phpool, ph); + } } /* @@ -273,7 +283,7 @@ pool_create(size, align, ioff, nitems, w u_int align; u_int ioff; int nitems; - char *wchan; + const char *wchan; size_t pagesz; void *(*alloc) __P((unsigned long, int, int)); void (*release) __P((void *, unsigned long, int)); @@ -318,13 +328,13 @@ pool_init(pp, size, align, ioff, flags, u_int align; u_int ioff; int flags; - char *wchan; + const char *wchan; size_t pagesz; void *(*alloc) __P((unsigned long, int, int)); void (*release) __P((void *, unsigned long, int)); int mtype; { - int off, slack; + int off, slack, i; /* * Check arguments and construct default values. @@ -360,7 +370,8 @@ pool_init(pp, size, align, ioff, flags, pp->pr_minitems = 0; pp->pr_minpages = 0; pp->pr_maxpages = UINT_MAX; - pp->pr_flags = flags; + pp->pr_roflags = flags; + pp->pr_flags = 0; pp->pr_size = ALIGN(size); pp->pr_align = align; pp->pr_wchan = wchan; @@ -370,6 +381,13 @@ pool_init(pp, size, align, ioff, flags, pp->pr_pagesz = pagesz; pp->pr_pagemask = ~(pagesz - 1); pp->pr_pageshift = ffs(pagesz) - 1; + pp->pr_nitems = 0; + pp->pr_nout = 0; + pp->pr_hardlimit = UINT_MAX; + pp->pr_hardlimit_warning = NULL; + pp->pr_hardlimit_ratecap = 0; + memset(&pp->pr_hardlimit_warning_last, 0, + sizeof(pp->pr_hardlimit_warning_last)); /* * Decide whether to put the page header off page to avoid @@ -380,14 +398,16 @@ pool_init(pp, size, align, ioff, flags, */ if (pp->pr_size < pagesz/16) { /* Use the end of the page for the page header */ - pp->pr_flags |= PR_PHINPAGE; + pp->pr_roflags |= PR_PHINPAGE; pp->pr_phoffset = off = pagesz - ALIGN(sizeof(struct pool_item_header)); } else { /* The page header will be taken from our page header pool */ pp->pr_phoffset = 0; off = pagesz; - memset(pp->pr_hashtab, 0, sizeof(pp->pr_hashtab)); + for (i = 0; i < PR_HASHTABSIZE; i++) { + LIST_INIT(&pp->pr_hashtab[i]); + } } /* @@ -421,14 +441,13 @@ pool_init(pp, size, align, ioff, flags, pp->pr_log = malloc(pool_logsize * sizeof(struct pool_log), M_TEMP, M_NOWAIT); if (pp->pr_log == NULL) - pp->pr_flags &= ~PR_LOGGING; + pp->pr_roflags &= ~PR_LOGGING; pp->pr_curlogentry = 0; pp->pr_logsize = pool_logsize; } #endif - simple_lock_init(&pp->pr_lock); - lockinit(&pp->pr_resourcelock, PSWP, wchan, 0, 0); + simple_lock_init(&pp->pr_slock); /* * Initialize private page header pool if we haven't done so yet. @@ -451,15 +470,15 @@ pool_destroy(pp) struct pool_item_header *ph; #ifdef DIAGNOSTIC - if (pp->pr_nget - pp->pr_nput != 0) { + if (pp->pr_nout != 0) { pr_printlog(pp); - panic("pool_destroy: pool busy: still out: %lu\n", - pp->pr_nget - pp->pr_nput); + panic("pool_destroy: pool busy: still out: %u\n", + pp->pr_nout); } #endif /* Remove all pages */ - if ((pp->pr_flags & PR_STATIC) == 0) + if ((pp->pr_roflags & PR_STATIC) == 0) while ((ph = pp->pr_pagelist.tqh_first) != NULL) pr_rmpage(pp, ph); @@ -468,11 +487,11 @@ pool_destroy(pp) drainpp = NULL; #ifdef POOL_DIAGNOSTIC - if ((pp->pr_flags & PR_LOGGING) != 0) + if ((pp->pr_roflags & PR_LOGGING) != 0) free(pp->pr_log, M_TEMP); #endif - if (pp->pr_flags & PR_FREEHEADER) + if (pp->pr_roflags & PR_FREEHEADER) free(pp, M_POOL); } @@ -499,73 +518,150 @@ pool_get(pp, flags) struct pool_item_header *ph; #ifdef DIAGNOSTIC - if ((pp->pr_flags & PR_STATIC) && (flags & PR_MALLOCOK)) { + if ((pp->pr_roflags & PR_STATIC) && (flags & PR_MALLOCOK)) { pr_printlog(pp); panic("pool_get: static"); } #endif - simple_lock(&pp->pr_lock); if (curproc == NULL && (flags & PR_WAITOK) != 0) panic("pool_get: must have NOWAIT"); + simple_lock(&pp->pr_slock); + + startover: + /* + * Check to see if we've reached the hard limit. If we have, + * and we can wait, then wait until an item has been returned to + * the pool. + */ +#ifdef DIAGNOSTIC + if (pp->pr_nout > pp->pr_hardlimit) { + simple_unlock(&pp->pr_slock); + panic("pool_get: %s: crossed hard limit", pp->pr_wchan); + } +#endif + if (pp->pr_nout == pp->pr_hardlimit) { + if (flags & PR_WAITOK) { + /* + * XXX: A warning isn't logged in this case. Should + * it be? + */ + pp->pr_flags |= PR_WANTED; + simple_unlock(&pp->pr_slock); + tsleep((caddr_t)pp, PSWP, pp->pr_wchan, 0); + simple_lock(&pp->pr_slock); + goto startover; + } + if (pp->pr_hardlimit_warning != NULL) { + /* + * Log a message that the hard limit has been hit. + */ + struct timeval curtime, logdiff; + int s = splclock(); + curtime = mono_time; + splx(s); + timersub(&curtime, &pp->pr_hardlimit_warning_last, + &logdiff); + if (logdiff.tv_sec >= pp->pr_hardlimit_ratecap) { + pp->pr_hardlimit_warning_last = curtime; + log(LOG_ERR, "%s\n", pp->pr_hardlimit_warning); + } + } + + if (flags & PR_URGENT) + panic("pool_get: urgent"); + + pp->pr_nfail++; + + simple_unlock(&pp->pr_slock); + return (NULL); + } + /* * The convention we use is that if `curpage' is not NULL, then * it points at a non-empty bucket. In particular, `curpage' * never points at a page header which has PR_PHINPAGE set and * has no items in its bucket. */ - while ((ph = pp->pr_curpage) == NULL) { + if ((ph = pp->pr_curpage) == NULL) { void *v; - int lkflags = LK_EXCLUSIVE | LK_INTERLOCK | - ((flags & PR_WAITOK) == 0 ? LK_NOWAIT : 0); - - /* Get long-term lock on pool */ - if (lockmgr(&pp->pr_resourcelock, lkflags, &pp->pr_lock) != 0) - return (NULL); - /* Check if pool became non-empty while we slept */ - if ((ph = pp->pr_curpage) != NULL) - goto again; +#ifdef DIAGNOSTIC + if (pp->pr_nitems != 0) { + simple_unlock(&pp->pr_slock); + printf("pool_get: %s: curpage NULL, nitems %u\n", + pp->pr_wchan, pp->pr_nitems); + panic("pool_get: nitems inconsistent\n"); + } +#endif - /* Call the page back-end allocator for more memory */ + /* + * Call the back-end page allocator for more memory. + * Release the pool lock, as the back-end page allocator + * may block. + */ + simple_unlock(&pp->pr_slock); v = (*pp->pr_alloc)(pp->pr_pagesz, flags, pp->pr_mtype); + simple_lock(&pp->pr_slock); + if (v == NULL) { + /* + * We were unable to allocate a page, but + * we released the lock during allocation, + * so perhaps items were freed back to the + * pool. Check for this case. + */ + if (pp->pr_curpage != NULL) + goto startover; + if (flags & PR_URGENT) panic("pool_get: urgent"); + if ((flags & PR_WAITOK) == 0) { pp->pr_nfail++; - lockmgr(&pp->pr_resourcelock, LK_RELEASE, NULL); + simple_unlock(&pp->pr_slock); return (NULL); } /* * Wait for items to be returned to this pool. + * * XXX: we actually want to wait just until * the page allocator has memory again. Depending * on this pool's usage, we might get stuck here * for a long time. + * + * XXX: maybe we should wake up once a second and + * try again? */ pp->pr_flags |= PR_WANTED; - lockmgr(&pp->pr_resourcelock, LK_RELEASE, NULL); + simple_unlock(&pp->pr_slock); tsleep((caddr_t)pp, PSWP, pp->pr_wchan, 0); - simple_lock(&pp->pr_lock); - continue; + simple_lock(&pp->pr_slock); + goto startover; } /* We have more memory; add it to the pool */ pp->pr_npagealloc++; pool_prime_page(pp, v); -again: - /* Re-acquire pool interlock */ - simple_lock(&pp->pr_lock); - lockmgr(&pp->pr_resourcelock, LK_RELEASE, NULL); + /* Start the allocation process over. */ + goto startover; } - if ((v = pi = TAILQ_FIRST(&ph->ph_itemlist)) == NULL) + if ((v = pi = TAILQ_FIRST(&ph->ph_itemlist)) == NULL) { + simple_unlock(&pp->pr_slock); panic("pool_get: %s: page empty", pp->pr_wchan); - + } +#ifdef DIAGNOSTIC + if (pp->pr_nitems == 0) { + simple_unlock(&pp->pr_slock); + printf("pool_get: %s: items on itemlist, nitems %u\n", + pp->pr_wchan, pp->pr_nitems); + panic("pool_get: nitems inconsistent\n"); + } +#endif pr_log(pp, v, PRLOG_GET, file, line); #ifdef DIAGNOSTIC @@ -581,6 +677,8 @@ again: * Remove from item list. */ TAILQ_REMOVE(&ph->ph_itemlist, pi, pi_list); + pp->pr_nitems--; + pp->pr_nout++; if (ph->ph_nmissing == 0) { #ifdef DIAGNOSTIC if (pp->pr_nidle == 0) @@ -590,17 +688,28 @@ again: } ph->ph_nmissing++; if (TAILQ_FIRST(&ph->ph_itemlist) == NULL) { +#ifdef DIAGNOSTIC + if (ph->ph_nmissing != pp->pr_itemsperpage) { + simple_unlock(&pp->pr_slock); + panic("pool_get: %s: nmissing inconsistent", + pp->pr_wchan); + } +#endif /* * Find a new non-empty page header, if any. * Start search from the page head, to increase * the chance for "high water" pages to be freed. * - * First, move the now empty page to the head of - * the page list. + * Migrate empty pages to the end of the list. This + * will speed the update of curpage as pages become + * idle. Empty pages intermingled with idle pages + * is no big deal. As soon as a page becomes un-empty, + * it will move back to the head of the list. */ TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist); - TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist); - while ((ph = TAILQ_NEXT(ph, ph_pagelist)) != NULL) + TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist); + for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; + ph = TAILQ_NEXT(ph, ph_pagelist)) if (TAILQ_FIRST(&ph->ph_itemlist) != NULL) break; @@ -608,7 +717,20 @@ again: } pp->pr_nget++; - simple_unlock(&pp->pr_lock); + + /* + * If we have a low water mark and we are now below that low + * water mark, add more items to the pool. + */ + if (pp->pr_nitems < pp->pr_minitems && pool_catchup(pp) != 0) { + /* + * XXX: Should we log a warning? Should we set up a timeout + * to try again in a second or so? The latter could break + * a caller's assumptions about interrupt protection, etc. + */ + } + + simple_unlock(&pp->pr_slock); return (v); } @@ -632,10 +754,11 @@ pool_put(pp, v) struct pool_item *pi = v; struct pool_item_header *ph; caddr_t page; + int s; page = (caddr_t)((u_long)v & pp->pr_pagemask); - simple_lock(&pp->pr_lock); + simple_lock(&pp->pr_slock); pr_log(pp, v, PRLOG_PUT, file, line); @@ -653,6 +776,8 @@ pool_put(pp, v) TAILQ_INSERT_HEAD(&ph->ph_itemlist, pi, pi_list); ph->ph_nmissing--; pp->pr_nput++; + pp->pr_nitems++; + pp->pr_nout--; /* Cancel "pool empty" condition if it exists */ if (pp->pr_curpage == NULL) @@ -662,29 +787,51 @@ pool_put(pp, v) pp->pr_flags &= ~PR_WANTED; if (ph->ph_nmissing == 0) pp->pr_nidle++; + simple_unlock(&pp->pr_slock); wakeup((caddr_t)pp); - simple_unlock(&pp->pr_lock); return; } /* - * If this page is now complete, move it to the end of the pagelist. - * If this page has just become un-empty, move it the head. + * If this page is now complete, do one of two things: + * + * (1) If we have more pages than the page high water + * mark, free the page back to the system. + * + * (2) Move it to the end of the page list, so that + * we minimize our chances of fragmenting the + * pool. Idle pages migrate to the end (along with + * completely empty pages, so that we find un-empty + * pages more quickly when we update curpage) of the + * list so they can be more easily swept up by + * the pagedaemon when pages are scarce. */ if (ph->ph_nmissing == 0) { pp->pr_nidle++; if (pp->pr_npages > pp->pr_maxpages) { -#if 0 - timeout(pool_drain, 0, pool_inactive_time*hz); -#else pr_rmpage(pp, ph); -#endif } else { TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist); TAILQ_INSERT_TAIL(&pp->pr_pagelist, ph, ph_pagelist); - ph->ph_time = time; - /* XXX - update curpage */ + /* + * Update the timestamp on the page. A page must + * be idle for some period of time before it can + * be reclaimed by the pagedaemon. This minimizes + * ping-pong'ing for memory. + */ + s = splclock(); + ph->ph_time = mono_time; + splx(s); + + /* + * Update the current page pointer. Just look for + * the first page with any free items. + * + * XXX: Maybe we want an option to look for the + * page with the fewest available items, to minimize + * fragmentation? + */ for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; ph = TAILQ_NEXT(ph, ph_pagelist)) if (TAILQ_FIRST(&ph->ph_itemlist) != NULL) @@ -693,8 +840,20 @@ pool_put(pp, v) pp->pr_curpage = ph; } } + /* + * If the page has just become un-empty, move it to the head of + * the list, and make it the current page. The next allocation + * will get the item from this page, instead of further fragmenting + * the pool. + */ + else if (ph->ph_nmissing == (pp->pr_itemsperpage - 1)) { + TAILQ_REMOVE(&pp->pr_pagelist, ph, ph_pagelist); + TAILQ_INSERT_HEAD(&pp->pr_pagelist, ph, ph_pagelist); + pp->pr_curpage = ph; + } + + simple_unlock(&pp->pr_slock); - simple_unlock(&pp->pr_lock); } /* @@ -710,28 +869,30 @@ pool_prime(pp, n, storage) int newnitems, newpages; #ifdef DIAGNOSTIC - if (storage && !(pp->pr_flags & PR_STATIC)) + if (storage && !(pp->pr_roflags & PR_STATIC)) panic("pool_prime: static"); /* !storage && static caught below */ #endif - (void)lockmgr(&pp->pr_resourcelock, LK_EXCLUSIVE, NULL); + simple_lock(&pp->pr_slock); + newnitems = pp->pr_minitems + n; newpages = - roundup(pp->pr_itemsperpage,newnitems) / pp->pr_itemsperpage + roundup(newnitems, pp->pr_itemsperpage) / pp->pr_itemsperpage - pp->pr_minpages; while (newpages-- > 0) { - - if (pp->pr_flags & PR_STATIC) { + if (pp->pr_roflags & PR_STATIC) { cp = storage; storage += pp->pr_pagesz; } else { + simple_unlock(&pp->pr_slock); cp = (*pp->pr_alloc)(pp->pr_pagesz, 0, pp->pr_mtype); + simple_lock(&pp->pr_slock); } if (cp == NULL) { - (void)lockmgr(&pp->pr_resourcelock, LK_RELEASE, NULL); + simple_unlock(&pp->pr_slock); return (ENOMEM); } @@ -744,14 +905,16 @@ pool_prime(pp, n, storage) if (pp->pr_minpages >= pp->pr_maxpages) pp->pr_maxpages = pp->pr_minpages + 1; /* XXX */ - (void)lockmgr(&pp->pr_resourcelock, LK_RELEASE, NULL); + simple_unlock(&pp->pr_slock); return (0); } /* * Add a page worth of items to the pool. + * + * Note, we must be called with the pool descriptor LOCKED. */ -int +static void pool_prime_page(pp, storage) struct pool *pp; caddr_t storage; @@ -763,9 +926,7 @@ pool_prime_page(pp, storage) unsigned int ioff = pp->pr_itemoffset; int n; - simple_lock(&pp->pr_lock); - - if ((pp->pr_flags & PR_PHINPAGE) != 0) { + if ((pp->pr_roflags & PR_PHINPAGE) != 0) { ph = (struct pool_item_header *)(cp + pp->pr_phoffset); } else { ph = pool_get(&phpool, PR_URGENT); @@ -780,7 +941,7 @@ pool_prime_page(pp, storage) TAILQ_INIT(&ph->ph_itemlist); ph->ph_page = storage; ph->ph_nmissing = 0; - ph->ph_time.tv_sec = ph->ph_time.tv_usec = 0; + memset(&ph->ph_time, 0, sizeof(ph->ph_time)); pp->pr_nidle++; @@ -801,6 +962,7 @@ pool_prime_page(pp, storage) * Insert remaining chunks on the bucket list. */ n = pp->pr_itemsperpage; + pp->pr_nitems += n; while (n--) { pi = (struct pool_item *)cp; @@ -821,9 +983,57 @@ pool_prime_page(pp, storage) if (++pp->pr_npages > pp->pr_hiwat) pp->pr_hiwat = pp->pr_npages; +} - simple_unlock(&pp->pr_lock); - return (0); +/* + * Like pool_prime(), except this is used by pool_get() when nitems + * drops below the low water mark. This is used to catch up nitmes + * with the low water mark. + * + * Note 1, we never wait for memory here, we let the caller decide what to do. + * + * Note 2, this doesn't work with static pools. + * + * Note 3, we must be called with the pool already locked, and we return + * with it locked. + */ +static int +pool_catchup(pp) + struct pool *pp; +{ + caddr_t cp; + int error = 0; + + if (pp->pr_roflags & PR_STATIC) { + /* + * We dropped below the low water mark, and this is not a + * good thing. Log a warning. + * + * XXX: rate-limit this? + */ + printf("WARNING: static pool `%s' dropped below low water " + "mark\n", pp->pr_wchan); + return (0); + } + + while (pp->pr_nitems < pp->pr_minitems) { + /* + * Call the page back-end allocator for more memory. + * + * XXX: We never wait, so should we bother unlocking + * the pool descriptor? + */ + simple_unlock(&pp->pr_slock); + cp = (*pp->pr_alloc)(pp->pr_pagesz, 0, pp->pr_mtype); + simple_lock(&pp->pr_slock); + if (cp == NULL) { + error = ENOMEM; + break; + } + pool_prime_page(pp, cp); + } + + return (error); } void @@ -831,13 +1041,25 @@ pool_setlowat(pp, n) pool_handle_t pp; int n; { + int error; + + simple_lock(&pp->pr_slock); - (void)lockmgr(&pp->pr_resourcelock, LK_EXCLUSIVE, NULL); pp->pr_minitems = n; pp->pr_minpages = (n == 0) ? 0 - : roundup(pp->pr_itemsperpage,n) / pp->pr_itemsperpage; - (void)lockmgr(&pp->pr_resourcelock, LK_RELEASE, NULL); + : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage; + + /* Make sure we're caught up with the newly-set low water mark. */ + if ((error = pool_catchup(pp)) != 0) { + /* + * XXX: Should we log a warning? Should we set up a timeout + * to try again in a second or so? The latter could break + * a caller's assumptions about interrupt protection, etc. + */ + } + + simple_unlock(&pp->pr_slock); } void @@ -846,13 +1068,41 @@ pool_sethiwat(pp, n) int n; { - (void)lockmgr(&pp->pr_resourcelock, LK_EXCLUSIVE, NULL); + simple_lock(&pp->pr_slock); + pp->pr_maxpages = (n == 0) ? 0 - : roundup(pp->pr_itemsperpage,n) / pp->pr_itemsperpage; - (void)lockmgr(&pp->pr_resourcelock, LK_RELEASE, NULL); + : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage; + + simple_unlock(&pp->pr_slock); } +void +pool_sethardlimit(pp, n, warnmess, ratecap) + pool_handle_t pp; + int n; + const char *warnmess; + int ratecap; +{ + + simple_lock(&pp->pr_slock); + + pp->pr_hardlimit = n; + pp->pr_hardlimit_warning = warnmess; + pp->pr_hardlimit_ratecap = ratecap; + memset(&pp->pr_hardlimit_warning_last, 0, + sizeof(pp->pr_hardlimit_warning_last)); + + /* + * In-line version of pool_sethiwat(), because we don't want to + * release the lock. + */ + pp->pr_maxpages = (n == 0) + ? 0 + : roundup(n, pp->pr_itemsperpage) / pp->pr_itemsperpage; + + simple_unlock(&pp->pr_slock); +} /* * Default page allocator. @@ -865,11 +1115,7 @@ pool_page_alloc(sz, flags, mtype) { boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE; -#if defined(UVM) return ((void *)uvm_km_alloc_poolpage(waitok)); -#else - return ((void *)kmem_alloc_poolpage(waitok)); -#endif } static void @@ -879,11 +1125,7 @@ pool_page_free(v, sz, mtype) int mtype; { -#if defined(UVM) uvm_km_free_poolpage((vaddr_t)v); -#else - kmem_free_poolpage((vaddr_t)v); -#endif } /* @@ -896,20 +1138,10 @@ pool_page_alloc_nointr(sz, flags, mtype) int flags; int mtype; { -#if defined(UVM) boolean_t waitok = (flags & PR_WAITOK) ? TRUE : FALSE; - /* - * With UVM, we can use the kernel_map. - */ return ((void *)uvm_km_alloc_poolpage1(kernel_map, uvm.kernel_object, waitok)); -#else - /* - * Can't do anything so cool with Mach VM. - */ - return (pool_page_alloc(sz, flags, mtype)); -#endif } void @@ -919,11 +1151,7 @@ pool_page_free_nointr(v, sz, mtype) int mtype; { -#if defined(UVM) uvm_km_free_poolpage1(kernel_map, (vaddr_t)v); -#else - pool_page_free(v, sz, mtype); -#endif } @@ -931,18 +1159,23 @@ pool_page_free_nointr(v, sz, mtype) * Release all complete pages that have not been used recently. */ void -pool_reclaim (pp) +pool_reclaim(pp) pool_handle_t pp; { struct pool_item_header *ph, *phnext; - struct timeval curtime = time; + struct timeval curtime; + int s; - if (pp->pr_flags & PR_STATIC) + if (pp->pr_roflags & PR_STATIC) return; - if (simple_lock_try(&pp->pr_lock) == 0) + if (simple_lock_try(&pp->pr_slock) == 0) return; + s = splclock(); + curtime = mono_time; + splx(s); + for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; ph = phnext) { phnext = TAILQ_NEXT(ph, ph_pagelist); @@ -955,16 +1188,27 @@ pool_reclaim (pp) timersub(&curtime, &ph->ph_time, &diff); if (diff.tv_sec < pool_inactive_time) continue; + + /* + * If freeing this page would put us below + * the low water mark, stop now. + */ + if ((pp->pr_nitems - pp->pr_itemsperpage) < + pp->pr_minitems) + break; + pr_rmpage(pp, ph); } } - simple_unlock(&pp->pr_lock); + simple_unlock(&pp->pr_slock); } /* * Drain pools, one at a time. + * + * Note, we must never be called from an interrupt context. */ void pool_drain(arg) @@ -988,14 +1232,28 @@ pool_drain(arg) } -#ifdef DEBUG +#if defined(POOL_DIAGNOSTIC) || defined(DEBUG) /* * Diagnostic helpers. */ void pool_print(pp, label) struct pool *pp; - char *label; + const char *label; +{ + int s; + + s = splimp(); + simple_lock(&pp->pr_slock); + pool_print1(pp, label); + simple_unlock(&pp->pr_slock); + splx(s); +} + +static void +pool_print1(pp, label) + struct pool *pp; + const char *label; { if (label != NULL) @@ -1024,7 +1282,7 @@ pool_chk(pp, label) struct pool_item_header *ph; int r = 0; - simple_lock(&pp->pr_lock); + simple_lock(&pp->pr_slock); for (ph = TAILQ_FIRST(&pp->pr_pagelist); ph != NULL; ph = TAILQ_NEXT(ph, ph_pagelist)) { @@ -1034,11 +1292,12 @@ pool_chk(pp, label) caddr_t page; page = (caddr_t)((u_long)ph & pp->pr_pagemask); - if (page != ph->ph_page) { + if (page != ph->ph_page && + (pp->pr_roflags & PR_PHINPAGE) != 0) { if (label != NULL) printf("%s: ", label); - printf("pool(%s): page inconsistency: page %p;" - " at page head addr %p (p %p)\n", + printf("pool(%p:%s): page inconsistency: page %p;" + " at page head addr %p (p %p)\n", pp, pp->pr_wchan, ph->ph_page, ph, page); r++; @@ -1067,8 +1326,8 @@ pool_chk(pp, label) if (label != NULL) printf("%s: ", label); - printf("pool(%s): page inconsistency: page %p;" - " item ordinal %d; addr %p (p %p)\n", + printf("pool(%p:%s): page inconsistency: page %p;" + " item ordinal %d; addr %p (p %p)\n", pp, pp->pr_wchan, ph->ph_page, n, pi, page); r++; @@ -1076,7 +1335,7 @@ pool_chk(pp, label) } } out: - simple_unlock(&pp->pr_lock); + simple_unlock(&pp->pr_slock); return (r); } -#endif +#endif /* POOL_DIAGNOSTIC || DEBUG */