xref: /illumos-gate/usr/src/lib/libumem/common/umem.c (revision 38849194)
17c478bd9Sstevel@tonic-gate /*
27c478bd9Sstevel@tonic-gate  * CDDL HEADER START
37c478bd9Sstevel@tonic-gate  *
47c478bd9Sstevel@tonic-gate  * The contents of this file are subject to the terms of the
5789d94c2Sjwadams  * Common Development and Distribution License (the "License").
6789d94c2Sjwadams  * You may not use this file except in compliance with the License.
77c478bd9Sstevel@tonic-gate  *
87c478bd9Sstevel@tonic-gate  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
97c478bd9Sstevel@tonic-gate  * or http://www.opensolaris.org/os/licensing.
107c478bd9Sstevel@tonic-gate  * See the License for the specific language governing permissions
117c478bd9Sstevel@tonic-gate  * and limitations under the License.
127c478bd9Sstevel@tonic-gate  *
137c478bd9Sstevel@tonic-gate  * When distributing Covered Code, include this CDDL HEADER in each
147c478bd9Sstevel@tonic-gate  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
157c478bd9Sstevel@tonic-gate  * If applicable, add the following below this CDDL HEADER, with the
167c478bd9Sstevel@tonic-gate  * fields enclosed by brackets "[]" replaced with your own identifying
177c478bd9Sstevel@tonic-gate  * information: Portions Copyright [yyyy] [name of copyright owner]
187c478bd9Sstevel@tonic-gate  *
197c478bd9Sstevel@tonic-gate  * CDDL HEADER END
207c478bd9Sstevel@tonic-gate  */
21e8031f0aSraf 
227c478bd9Sstevel@tonic-gate /*
23a574db85Sraf  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24*38849194SRobert Mustacchi  * Copyright 2012 Joyent, Inc.  All rights reserved.
257c478bd9Sstevel@tonic-gate  * Use is subject to license terms.
267c478bd9Sstevel@tonic-gate  */
277c478bd9Sstevel@tonic-gate 
287c478bd9Sstevel@tonic-gate /*
297c478bd9Sstevel@tonic-gate  * based on usr/src/uts/common/os/kmem.c r1.64 from 2001/12/18
307c478bd9Sstevel@tonic-gate  *
317c478bd9Sstevel@tonic-gate  * The slab allocator, as described in the following two papers:
327c478bd9Sstevel@tonic-gate  *
337c478bd9Sstevel@tonic-gate  *	Jeff Bonwick,
347c478bd9Sstevel@tonic-gate  *	The Slab Allocator: An Object-Caching Kernel Memory Allocator.
357c478bd9Sstevel@tonic-gate  *	Proceedings of the Summer 1994 Usenix Conference.
367c478bd9Sstevel@tonic-gate  *	Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
377c478bd9Sstevel@tonic-gate  *
387c478bd9Sstevel@tonic-gate  *	Jeff Bonwick and Jonathan Adams,
397c478bd9Sstevel@tonic-gate  *	Magazines and vmem: Extending the Slab Allocator to Many CPUs and
407c478bd9Sstevel@tonic-gate  *	Arbitrary Resources.
417c478bd9Sstevel@tonic-gate  *	Proceedings of the 2001 Usenix Conference.
427c478bd9Sstevel@tonic-gate  *	Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
437c478bd9Sstevel@tonic-gate  *
447c478bd9Sstevel@tonic-gate  * 1. Overview
457c478bd9Sstevel@tonic-gate  * -----------
467c478bd9Sstevel@tonic-gate  * umem is very close to kmem in implementation.  There are four major
477c478bd9Sstevel@tonic-gate  * areas of divergence:
487c478bd9Sstevel@tonic-gate  *
497c478bd9Sstevel@tonic-gate  *	* Initialization
507c478bd9Sstevel@tonic-gate  *
517c478bd9Sstevel@tonic-gate  *	* CPU handling
527c478bd9Sstevel@tonic-gate  *
537c478bd9Sstevel@tonic-gate  *	* umem_update()
547c478bd9Sstevel@tonic-gate  *
557c478bd9Sstevel@tonic-gate  *	* KM_SLEEP v.s. UMEM_NOFAIL
567c478bd9Sstevel@tonic-gate  *
571c326e94Sjwadams  *	* lock ordering
587c478bd9Sstevel@tonic-gate  *
597c478bd9Sstevel@tonic-gate  * 2. Initialization
607c478bd9Sstevel@tonic-gate  * -----------------
617c478bd9Sstevel@tonic-gate  * kmem is initialized early on in boot, and knows that no one will call
627c478bd9Sstevel@tonic-gate  * into it before it is ready.  umem does not have these luxuries. Instead,
637c478bd9Sstevel@tonic-gate  * initialization is divided into two phases:
647c478bd9Sstevel@tonic-gate  *
657c478bd9Sstevel@tonic-gate  *	* library initialization, and
667c478bd9Sstevel@tonic-gate  *
677c478bd9Sstevel@tonic-gate  *	* first use
687c478bd9Sstevel@tonic-gate  *
697c478bd9Sstevel@tonic-gate  * umem's full initialization happens at the time of the first allocation
707c478bd9Sstevel@tonic-gate  * request (via malloc() and friends, umem_alloc(), or umem_zalloc()),
717c478bd9Sstevel@tonic-gate  * or the first call to umem_cache_create().
727c478bd9Sstevel@tonic-gate  *
737c478bd9Sstevel@tonic-gate  * umem_free(), and umem_cache_alloc() do not require special handling,
747c478bd9Sstevel@tonic-gate  * since the only way to get valid arguments for them is to successfully
757c478bd9Sstevel@tonic-gate  * call a function from the first group.
767c478bd9Sstevel@tonic-gate  *
777c478bd9Sstevel@tonic-gate  * 2.1. Library Initialization: umem_startup()
787c478bd9Sstevel@tonic-gate  * -------------------------------------------
797c478bd9Sstevel@tonic-gate  * umem_startup() is libumem.so's .init section.  It calls pthread_atfork()
807c478bd9Sstevel@tonic-gate  * to install the handlers necessary for umem's Fork1-Safety.  Because of
817c478bd9Sstevel@tonic-gate  * race condition issues, all other pre-umem_init() initialization is done
827c478bd9Sstevel@tonic-gate  * statically (i.e. by the dynamic linker).
837c478bd9Sstevel@tonic-gate  *
847c478bd9Sstevel@tonic-gate  * For standalone use, umem_startup() returns everything to its initial
857c478bd9Sstevel@tonic-gate  * state.
867c478bd9Sstevel@tonic-gate  *
877c478bd9Sstevel@tonic-gate  * 2.2. First use: umem_init()
887c478bd9Sstevel@tonic-gate  * ------------------------------
897c478bd9Sstevel@tonic-gate  * The first time any memory allocation function is used, we have to
907c478bd9Sstevel@tonic-gate  * create the backing caches and vmem arenas which are needed for it.
917c478bd9Sstevel@tonic-gate  * umem_init() is the central point for that task.  When it completes,
927c478bd9Sstevel@tonic-gate  * umem_ready is either UMEM_READY (all set) or UMEM_READY_INIT_FAILED (unable
937c478bd9Sstevel@tonic-gate  * to initialize, probably due to lack of memory).
947c478bd9Sstevel@tonic-gate  *
957c478bd9Sstevel@tonic-gate  * There are four different paths from which umem_init() is called:
967c478bd9Sstevel@tonic-gate  *
977c478bd9Sstevel@tonic-gate  *	* from umem_alloc() or umem_zalloc(), with 0 < size < UMEM_MAXBUF,
987c478bd9Sstevel@tonic-gate  *
997c478bd9Sstevel@tonic-gate  *	* from umem_alloc() or umem_zalloc(), with size > UMEM_MAXBUF,
1007c478bd9Sstevel@tonic-gate  *
1017c478bd9Sstevel@tonic-gate  *	* from umem_cache_create(), and
1027c478bd9Sstevel@tonic-gate  *
1037c478bd9Sstevel@tonic-gate  *	* from memalign(), with align > UMEM_ALIGN.
1047c478bd9Sstevel@tonic-gate  *
1057c478bd9Sstevel@tonic-gate  * The last three just check if umem is initialized, and call umem_init()
1067c478bd9Sstevel@tonic-gate  * if it is not.  For performance reasons, the first case is more complicated.
1077c478bd9Sstevel@tonic-gate  *
1087c478bd9Sstevel@tonic-gate  * 2.2.1. umem_alloc()/umem_zalloc(), with 0 < size < UMEM_MAXBUF
1097c478bd9Sstevel@tonic-gate  * -----------------------------------------------------------------
1107c478bd9Sstevel@tonic-gate  * In this case, umem_cache_alloc(&umem_null_cache, ...) is called.
1117c478bd9Sstevel@tonic-gate  * There is special case code in which causes any allocation on
1127c478bd9Sstevel@tonic-gate  * &umem_null_cache to fail by returning (NULL), regardless of the
1137c478bd9Sstevel@tonic-gate  * flags argument.
1147c478bd9Sstevel@tonic-gate  *
1157c478bd9Sstevel@tonic-gate  * So umem_cache_alloc() returns NULL, and umem_alloc()/umem_zalloc() call
1167c478bd9Sstevel@tonic-gate  * umem_alloc_retry().  umem_alloc_retry() sees that the allocation
1177c478bd9Sstevel@tonic-gate  * was agains &umem_null_cache, and calls umem_init().
1187c478bd9Sstevel@tonic-gate  *
1197c478bd9Sstevel@tonic-gate  * If initialization is successful, umem_alloc_retry() returns 1, which
1207c478bd9Sstevel@tonic-gate  * causes umem_alloc()/umem_zalloc() to start over, which causes it to load
1217c478bd9Sstevel@tonic-gate  * the (now valid) cache pointer from umem_alloc_table.
1227c478bd9Sstevel@tonic-gate  *
1237c478bd9Sstevel@tonic-gate  * 2.2.2. Dealing with race conditions
1247c478bd9Sstevel@tonic-gate  * -----------------------------------
1257c478bd9Sstevel@tonic-gate  * There are a couple race conditions resulting from the initialization
1267c478bd9Sstevel@tonic-gate  * code that we have to guard against:
1277c478bd9Sstevel@tonic-gate  *
1287c478bd9Sstevel@tonic-gate  *	* In umem_cache_create(), there is a special UMC_INTERNAL cflag
1297c478bd9Sstevel@tonic-gate  *	that is passed for caches created during initialization.  It
1307c478bd9Sstevel@tonic-gate  *	is illegal for a user to try to create a UMC_INTERNAL cache.
1317c478bd9Sstevel@tonic-gate  *	This allows initialization to proceed, but any other
1327c478bd9Sstevel@tonic-gate  *	umem_cache_create()s will block by calling umem_init().
1337c478bd9Sstevel@tonic-gate  *
1347c478bd9Sstevel@tonic-gate  *	* Since umem_null_cache has a 1-element cache_cpu, it's cache_cpu_mask
1357c478bd9Sstevel@tonic-gate  *	is always zero.  umem_cache_alloc uses cp->cache_cpu_mask to
1367c478bd9Sstevel@tonic-gate  *	mask the cpu number.  This prevents a race between grabbing a
1377c478bd9Sstevel@tonic-gate  *	cache pointer out of umem_alloc_table and growing the cpu array.
1387c478bd9Sstevel@tonic-gate  *
1397c478bd9Sstevel@tonic-gate  *
1407c478bd9Sstevel@tonic-gate  * 3. CPU handling
1417c478bd9Sstevel@tonic-gate  * ---------------
1427c478bd9Sstevel@tonic-gate  * kmem uses the CPU's sequence number to determine which "cpu cache" to
1437c478bd9Sstevel@tonic-gate  * use for an allocation.  Currently, there is no way to get the sequence
1447c478bd9Sstevel@tonic-gate  * number in userspace.
1457c478bd9Sstevel@tonic-gate  *
1467c478bd9Sstevel@tonic-gate  * umem keeps track of cpu information in umem_cpus, an array of umem_max_ncpus
1477c478bd9Sstevel@tonic-gate  * umem_cpu_t structures.  CURCPU() is a a "hint" function, which we then mask
1487c478bd9Sstevel@tonic-gate  * with either umem_cpu_mask or cp->cache_cpu_mask to find the actual "cpu" id.
1497c478bd9Sstevel@tonic-gate  * The mechanics of this is all in the CPU(mask) macro.
1507c478bd9Sstevel@tonic-gate  *
1517c478bd9Sstevel@tonic-gate  * Currently, umem uses _lwp_self() as its hint.
1527c478bd9Sstevel@tonic-gate  *
1537c478bd9Sstevel@tonic-gate  *
1547c478bd9Sstevel@tonic-gate  * 4. The update thread
1557c478bd9Sstevel@tonic-gate  * --------------------
1567c478bd9Sstevel@tonic-gate  * kmem uses a task queue, kmem_taskq, to do periodic maintenance on
1577c478bd9Sstevel@tonic-gate  * every kmem cache.  vmem has a periodic timeout for hash table resizing.
1587c478bd9Sstevel@tonic-gate  * The kmem_taskq also provides a separate context for kmem_cache_reap()'s
1597c478bd9Sstevel@tonic-gate  * to be done in, avoiding issues of the context of kmem_reap() callers.
1607c478bd9Sstevel@tonic-gate  *
1617c478bd9Sstevel@tonic-gate  * Instead, umem has the concept of "updates", which are asynchronous requests
1627c478bd9Sstevel@tonic-gate  * for work attached to single caches.  All caches with pending work are
1637c478bd9Sstevel@tonic-gate  * on a doubly linked list rooted at the umem_null_cache.  All update state
1647c478bd9Sstevel@tonic-gate  * is protected by the umem_update_lock mutex, and the umem_update_cv is used
1657c478bd9Sstevel@tonic-gate  * for notification between threads.
1667c478bd9Sstevel@tonic-gate  *
1677c478bd9Sstevel@tonic-gate  * 4.1. Cache states with regards to updates
1687c478bd9Sstevel@tonic-gate  * -----------------------------------------
1697c478bd9Sstevel@tonic-gate  * A given cache is in one of three states:
1707c478bd9Sstevel@tonic-gate  *
1717c478bd9Sstevel@tonic-gate  * Inactive		cache_uflags is zero, cache_u{next,prev} are NULL
1727c478bd9Sstevel@tonic-gate  *
1737c478bd9Sstevel@tonic-gate  * Work Requested	cache_uflags is non-zero (but UMU_ACTIVE is not set),
1747c478bd9Sstevel@tonic-gate  *			cache_u{next,prev} link the cache onto the global
1757c478bd9Sstevel@tonic-gate  *			update list
1767c478bd9Sstevel@tonic-gate  *
1777c478bd9Sstevel@tonic-gate  * Active		cache_uflags has UMU_ACTIVE set, cache_u{next,prev}
1787c478bd9Sstevel@tonic-gate  *			are NULL, and either umem_update_thr or
1797c478bd9Sstevel@tonic-gate  *			umem_st_update_thr are actively doing work on the
1807c478bd9Sstevel@tonic-gate  *			cache.
1817c478bd9Sstevel@tonic-gate  *
1827c478bd9Sstevel@tonic-gate  * An update can be added to any cache in any state -- if the cache is
1837c478bd9Sstevel@tonic-gate  * Inactive, it transitions to being Work Requested.  If the cache is
1847c478bd9Sstevel@tonic-gate  * Active, the worker will notice the new update and act on it before
1857c478bd9Sstevel@tonic-gate  * transitioning the cache to the Inactive state.
1867c478bd9Sstevel@tonic-gate  *
1877c478bd9Sstevel@tonic-gate  * If a cache is in the Active state, UMU_NOTIFY can be set, which asks
1887c478bd9Sstevel@tonic-gate  * the worker to broadcast the umem_update_cv when it has finished.
1897c478bd9Sstevel@tonic-gate  *
1907c478bd9Sstevel@tonic-gate  * 4.2. Update interface
1917c478bd9Sstevel@tonic-gate  * ---------------------
1927c478bd9Sstevel@tonic-gate  * umem_add_update() adds an update to a particular cache.
1937c478bd9Sstevel@tonic-gate  * umem_updateall() adds an update to all caches.
1947c478bd9Sstevel@tonic-gate  * umem_remove_updates() returns a cache to the Inactive state.
1957c478bd9Sstevel@tonic-gate  *
1967c478bd9Sstevel@tonic-gate  * umem_process_updates() process all caches in the Work Requested state.
1977c478bd9Sstevel@tonic-gate  *
1987c478bd9Sstevel@tonic-gate  * 4.3. Reaping
1997c478bd9Sstevel@tonic-gate  * ------------
2007c478bd9Sstevel@tonic-gate  * When umem_reap() is called (at the time of heap growth), it schedule
2017c478bd9Sstevel@tonic-gate  * UMU_REAP updates on every cache.  It then checks to see if the update
2027c478bd9Sstevel@tonic-gate  * thread exists (umem_update_thr != 0).  If it is, it broadcasts
2037c478bd9Sstevel@tonic-gate  * the umem_update_cv to wake the update thread up, and returns.
2047c478bd9Sstevel@tonic-gate  *
2057c478bd9Sstevel@tonic-gate  * If the update thread does not exist (umem_update_thr == 0), and the
2067c478bd9Sstevel@tonic-gate  * program currently has multiple threads, umem_reap() attempts to create
2077c478bd9Sstevel@tonic-gate  * a new update thread.
2087c478bd9Sstevel@tonic-gate  *
2097c478bd9Sstevel@tonic-gate  * If the process is not multithreaded, or the creation fails, umem_reap()
2107c478bd9Sstevel@tonic-gate  * calls umem_st_update() to do an inline update.
2117c478bd9Sstevel@tonic-gate  *
2127c478bd9Sstevel@tonic-gate  * 4.4. The update thread
2137c478bd9Sstevel@tonic-gate  * ----------------------
2147c478bd9Sstevel@tonic-gate  * The update thread spends most of its time in cond_timedwait() on the
2157c478bd9Sstevel@tonic-gate  * umem_update_cv.  It wakes up under two conditions:
2167c478bd9Sstevel@tonic-gate  *
2177c478bd9Sstevel@tonic-gate  *	* The timedwait times out, in which case it needs to run a global
2187c478bd9Sstevel@tonic-gate  *	update, or
2197c478bd9Sstevel@tonic-gate  *
2207c478bd9Sstevel@tonic-gate  *	* someone cond_broadcast(3THR)s the umem_update_cv, in which case
2217c478bd9Sstevel@tonic-gate  *	it needs to check if there are any caches in the Work Requested
2227c478bd9Sstevel@tonic-gate  *	state.
2237c478bd9Sstevel@tonic-gate  *
2247c478bd9Sstevel@tonic-gate  * When it is time for another global update, umem calls umem_cache_update()
2257c478bd9Sstevel@tonic-gate  * on every cache, then calls vmem_update(), which tunes the vmem structures.
2267c478bd9Sstevel@tonic-gate  * umem_cache_update() can request further work using umem_add_update().
2277c478bd9Sstevel@tonic-gate  *
2287c478bd9Sstevel@tonic-gate  * After any work from the global update completes, the update timer is
2297c478bd9Sstevel@tonic-gate  * reset to umem_reap_interval seconds in the future.  This makes the
2307c478bd9Sstevel@tonic-gate  * updates self-throttling.
2317c478bd9Sstevel@tonic-gate  *
2327c478bd9Sstevel@tonic-gate  * Reaps are similarly self-throttling.  After a UMU_REAP update has
2337c478bd9Sstevel@tonic-gate  * been scheduled on all caches, umem_reap() sets a flag and wakes up the
2347c478bd9Sstevel@tonic-gate  * update thread.  The update thread notices the flag, and resets the
2357c478bd9Sstevel@tonic-gate  * reap state.
2367c478bd9Sstevel@tonic-gate  *
2377c478bd9Sstevel@tonic-gate  * 4.5. Inline updates
2387c478bd9Sstevel@tonic-gate  * -------------------
2397c478bd9Sstevel@tonic-gate  * If the update thread is not running, umem_st_update() is used instead.  It
2407c478bd9Sstevel@tonic-gate  * immediately does a global update (as above), then calls
2417c478bd9Sstevel@tonic-gate  * umem_process_updates() to process both the reaps that umem_reap() added and
2427c478bd9Sstevel@tonic-gate  * any work generated by the global update.  Afterwards, it resets the reap
2437c478bd9Sstevel@tonic-gate  * state.
2447c478bd9Sstevel@tonic-gate  *
2457c478bd9Sstevel@tonic-gate  * While the umem_st_update() is running, umem_st_update_thr holds the thread
2467c478bd9Sstevel@tonic-gate  * id of the thread performing the update.
2477c478bd9Sstevel@tonic-gate  *
2487c478bd9Sstevel@tonic-gate  * 4.6. Updates and fork1()
2497c478bd9Sstevel@tonic-gate  * ------------------------
2507c478bd9Sstevel@tonic-gate  * umem has fork1() pre- and post-handlers which lock up (and release) every
2517c478bd9Sstevel@tonic-gate  * mutex in every cache.  They also lock up the umem_update_lock.  Since
2527c478bd9Sstevel@tonic-gate  * fork1() only copies over a single lwp, other threads (including the update
2537c478bd9Sstevel@tonic-gate  * thread) could have been actively using a cache in the parent.  This
2547c478bd9Sstevel@tonic-gate  * can lead to inconsistencies in the child process.
2557c478bd9Sstevel@tonic-gate  *
2567c478bd9Sstevel@tonic-gate  * Because we locked all of the mutexes, the only possible inconsistancies are:
2577c478bd9Sstevel@tonic-gate  *
2587c478bd9Sstevel@tonic-gate  *	* a umem_cache_alloc() could leak its buffer.
2597c478bd9Sstevel@tonic-gate  *
2607c478bd9Sstevel@tonic-gate  *	* a caller of umem_depot_alloc() could leak a magazine, and all the
2617c478bd9Sstevel@tonic-gate  *	buffers contained in it.
2627c478bd9Sstevel@tonic-gate  *
2637c478bd9Sstevel@tonic-gate  *	* a cache could be in the Active update state.  In the child, there
2647c478bd9Sstevel@tonic-gate  *	would be no thread actually working on it.
2657c478bd9Sstevel@tonic-gate  *
2667c478bd9Sstevel@tonic-gate  *	* a umem_hash_rescale() could leak the new hash table.
2677c478bd9Sstevel@tonic-gate  *
2687c478bd9Sstevel@tonic-gate  *	* a umem_magazine_resize() could be in progress.
2697c478bd9Sstevel@tonic-gate  *
2707c478bd9Sstevel@tonic-gate  *	* a umem_reap() could be in progress.
2717c478bd9Sstevel@tonic-gate  *
2727c478bd9Sstevel@tonic-gate  * The memory leaks we can't do anything about.  umem_release_child() resets
2737c478bd9Sstevel@tonic-gate  * the update state, moves any caches in the Active state to the Work Requested
2747c478bd9Sstevel@tonic-gate  * state.  This might cause some updates to be re-run, but UMU_REAP and
2757c478bd9Sstevel@tonic-gate  * UMU_HASH_RESCALE are effectively idempotent, and the worst that can
2767c478bd9Sstevel@tonic-gate  * happen from umem_magazine_resize() is resizing the magazine twice in close
2777c478bd9Sstevel@tonic-gate  * succession.
2787c478bd9Sstevel@tonic-gate  *
2797c478bd9Sstevel@tonic-gate  * Much of the cleanup in umem_release_child() is skipped if
2807c478bd9Sstevel@tonic-gate  * umem_st_update_thr == thr_self().  This is so that applications which call
2817c478bd9Sstevel@tonic-gate  * fork1() from a cache callback does not break.  Needless to say, any such
2827c478bd9Sstevel@tonic-gate  * application is tremendously broken.
2837c478bd9Sstevel@tonic-gate  *
2847c478bd9Sstevel@tonic-gate  *
2857c478bd9Sstevel@tonic-gate  * 5. KM_SLEEP v.s. UMEM_NOFAIL
2867c478bd9Sstevel@tonic-gate  * ----------------------------
2877c478bd9Sstevel@tonic-gate  * Allocations against kmem and vmem have two basic modes:  SLEEP and
2887c478bd9Sstevel@tonic-gate  * NOSLEEP.  A sleeping allocation is will go to sleep (waiting for
2897c478bd9Sstevel@tonic-gate  * more memory) instead of failing (returning NULL).
2907c478bd9Sstevel@tonic-gate  *
2917c478bd9Sstevel@tonic-gate  * SLEEP allocations presume an extremely multithreaded model, with
2927c478bd9Sstevel@tonic-gate  * a lot of allocation and deallocation activity.  umem cannot presume
2937c478bd9Sstevel@tonic-gate  * that its clients have any particular type of behavior.  Instead,
2947c478bd9Sstevel@tonic-gate  * it provides two types of allocations:
2957c478bd9Sstevel@tonic-gate  *
2967c478bd9Sstevel@tonic-gate  *	* UMEM_DEFAULT, equivalent to KM_NOSLEEP (i.e. return NULL on
2977c478bd9Sstevel@tonic-gate  *	failure)
2987c478bd9Sstevel@tonic-gate  *
2997c478bd9Sstevel@tonic-gate  *	* UMEM_NOFAIL, which, on failure, calls an optional callback
3007c478bd9Sstevel@tonic-gate  *	(registered with umem_nofail_callback()).
3017c478bd9Sstevel@tonic-gate  *
3027c478bd9Sstevel@tonic-gate  * The callback is invoked with no locks held, and can do an arbitrary
3037c478bd9Sstevel@tonic-gate  * amount of work.  It then has a choice between:
3047c478bd9Sstevel@tonic-gate  *
3057c478bd9Sstevel@tonic-gate  *	* Returning UMEM_CALLBACK_RETRY, which will cause the allocation
3067c478bd9Sstevel@tonic-gate  *	to be restarted.
3077c478bd9Sstevel@tonic-gate  *
3087c478bd9Sstevel@tonic-gate  *	* Returning UMEM_CALLBACK_EXIT(status), which will cause exit(2)
3097c478bd9Sstevel@tonic-gate  *	to be invoked with status.  If multiple threads attempt to do
3107c478bd9Sstevel@tonic-gate  *	this simultaneously, only one will call exit(2).
3117c478bd9Sstevel@tonic-gate  *
3127c478bd9Sstevel@tonic-gate  *	* Doing some kind of non-local exit (thr_exit(3thr), longjmp(3C),
3137c478bd9Sstevel@tonic-gate  *	etc.)
3147c478bd9Sstevel@tonic-gate  *
3157c478bd9Sstevel@tonic-gate  * The default callback returns UMEM_CALLBACK_EXIT(255).
3167c478bd9Sstevel@tonic-gate  *
3177c478bd9Sstevel@tonic-gate  * To have these callbacks without risk of state corruption (in the case of
3187c478bd9Sstevel@tonic-gate  * a non-local exit), we have to ensure that the callbacks get invoked
3197c478bd9Sstevel@tonic-gate  * close to the original allocation, with no inconsistent state or held
3207c478bd9Sstevel@tonic-gate  * locks.  The following steps are taken:
3217c478bd9Sstevel@tonic-gate  *
3227c478bd9Sstevel@tonic-gate  *	* All invocations of vmem are VM_NOSLEEP.
3237c478bd9Sstevel@tonic-gate  *
3247c478bd9Sstevel@tonic-gate  *	* All constructor callbacks (which can themselves to allocations)
3257c478bd9Sstevel@tonic-gate  *	are passed UMEM_DEFAULT as their required allocation argument.  This
3267c478bd9Sstevel@tonic-gate  *	way, the constructor will fail, allowing the highest-level allocation
3277c478bd9Sstevel@tonic-gate  *	invoke the nofail callback.
3287c478bd9Sstevel@tonic-gate  *
3297c478bd9Sstevel@tonic-gate  *	If a constructor callback _does_ do a UMEM_NOFAIL allocation, and
3307c478bd9Sstevel@tonic-gate  *	the nofail callback does a non-local exit, we will leak the
3317c478bd9Sstevel@tonic-gate  *	partially-constructed buffer.
3321c326e94Sjwadams  *
3331c326e94Sjwadams  *
3341c326e94Sjwadams  * 6. Lock Ordering
3351c326e94Sjwadams  * ----------------
3361c326e94Sjwadams  * umem has a few more locks than kmem does, mostly in the update path.  The
3371c326e94Sjwadams  * overall lock ordering (earlier locks must be acquired first) is:
3381c326e94Sjwadams  *
3391c326e94Sjwadams  *	umem_init_lock
3401c326e94Sjwadams  *
3411c326e94Sjwadams  *	vmem_list_lock
3421c326e94Sjwadams  *	vmem_nosleep_lock.vmpl_mutex
3431c326e94Sjwadams  *	vmem_t's:
3441c326e94Sjwadams  *		vm_lock
345789d94c2Sjwadams  *	sbrk_lock
3461c326e94Sjwadams  *
3471c326e94Sjwadams  *	umem_cache_lock
3481c326e94Sjwadams  *	umem_update_lock
3491c326e94Sjwadams  *	umem_flags_lock
3501c326e94Sjwadams  *	umem_cache_t's:
3511c326e94Sjwadams  *		cache_cpu[*].cc_lock
3521c326e94Sjwadams  *		cache_depot_lock
3531c326e94Sjwadams  *		cache_lock
3541c326e94Sjwadams  *	umem_log_header_t's:
3551c326e94Sjwadams  *		lh_cpu[*].clh_lock
3561c326e94Sjwadams  *		lh_lock
357*38849194SRobert Mustacchi  *
358*38849194SRobert Mustacchi  * 7. Changing UMEM_MAXBUF
359*38849194SRobert Mustacchi  * -----------------------
360*38849194SRobert Mustacchi  *
361*38849194SRobert Mustacchi  * When changing UMEM_MAXBUF extra care has to be taken. It is not sufficient to
362*38849194SRobert Mustacchi  * simply increase this number. First, one must update the umem_alloc_table to
363*38849194SRobert Mustacchi  * have the appropriate number of entires based upon the new size. If this is
364*38849194SRobert Mustacchi  * not done, this will lead to libumem blowing an assertion.
365*38849194SRobert Mustacchi  *
366*38849194SRobert Mustacchi  * The second place to update, which is not required, is the umem_alloc_sizes.
367*38849194SRobert Mustacchi  * These determine the default cache sizes that we're going to support.
3687c478bd9Sstevel@tonic-gate  */
3697c478bd9Sstevel@tonic-gate 
3707c478bd9Sstevel@tonic-gate #include <umem_impl.h>
3717c478bd9Sstevel@tonic-gate #include <sys/vmem_impl_user.h>
3727c478bd9Sstevel@tonic-gate #include "umem_base.h"
3737c478bd9Sstevel@tonic-gate #include "vmem_base.h"
3747c478bd9Sstevel@tonic-gate 
3757c478bd9Sstevel@tonic-gate #include <sys/processor.h>
3767c478bd9Sstevel@tonic-gate #include <sys/sysmacros.h>
3777c478bd9Sstevel@tonic-gate 
3787c478bd9Sstevel@tonic-gate #include <alloca.h>
3797c478bd9Sstevel@tonic-gate #include <errno.h>
3807c478bd9Sstevel@tonic-gate #include <limits.h>
3817c478bd9Sstevel@tonic-gate #include <stdio.h>
3827c478bd9Sstevel@tonic-gate #include <stdlib.h>
3837c478bd9Sstevel@tonic-gate #include <string.h>
3847c478bd9Sstevel@tonic-gate #include <strings.h>
3857c478bd9Sstevel@tonic-gate #include <signal.h>
3867c478bd9Sstevel@tonic-gate #include <unistd.h>
3877c478bd9Sstevel@tonic-gate #include <atomic.h>
3887c478bd9Sstevel@tonic-gate 
3897c478bd9Sstevel@tonic-gate #include "misc.h"
3907c478bd9Sstevel@tonic-gate 
3917c478bd9Sstevel@tonic-gate #define	UMEM_VMFLAGS(umflag)	(VM_NOSLEEP)
3927c478bd9Sstevel@tonic-gate 
3937c478bd9Sstevel@tonic-gate size_t pagesize;
3947c478bd9Sstevel@tonic-gate 
3957c478bd9Sstevel@tonic-gate /*
3967c478bd9Sstevel@tonic-gate  * The default set of caches to back umem_alloc().
3977c478bd9Sstevel@tonic-gate  * These sizes should be reevaluated periodically.
3987c478bd9Sstevel@tonic-gate  *
3997c478bd9Sstevel@tonic-gate  * We want allocations that are multiples of the coherency granularity
4007c478bd9Sstevel@tonic-gate  * (64 bytes) to be satisfied from a cache which is a multiple of 64
4017c478bd9Sstevel@tonic-gate  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
4027c478bd9Sstevel@tonic-gate  * the next kmem_cache_size greater than or equal to it must be a
4037c478bd9Sstevel@tonic-gate  * multiple of 64.
404789d94c2Sjwadams  *
405789d94c2Sjwadams  * This table must be in sorted order, from smallest to highest.  The
406789d94c2Sjwadams  * highest slot must be UMEM_MAXBUF, and every slot afterwards must be
407789d94c2Sjwadams  * zero.
4087c478bd9Sstevel@tonic-gate  */
409789d94c2Sjwadams static int umem_alloc_sizes[] = {
4107c478bd9Sstevel@tonic-gate #ifdef _LP64
4117c478bd9Sstevel@tonic-gate 	1 * 8,
4127c478bd9Sstevel@tonic-gate 	1 * 16,
4137c478bd9Sstevel@tonic-gate 	2 * 16,
4147c478bd9Sstevel@tonic-gate 	3 * 16,
4157c478bd9Sstevel@tonic-gate #else
4167c478bd9Sstevel@tonic-gate 	1 * 8,
4177c478bd9Sstevel@tonic-gate 	2 * 8,
4187c478bd9Sstevel@tonic-gate 	3 * 8,
4197c478bd9Sstevel@tonic-gate 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
4207c478bd9Sstevel@tonic-gate #endif
4217c478bd9Sstevel@tonic-gate 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
4227c478bd9Sstevel@tonic-gate 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
4237c478bd9Sstevel@tonic-gate 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
4247c478bd9Sstevel@tonic-gate 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
4257c478bd9Sstevel@tonic-gate 	P2ALIGN(8192 / 7, 64),
4267c478bd9Sstevel@tonic-gate 	P2ALIGN(8192 / 6, 64),
4277c478bd9Sstevel@tonic-gate 	P2ALIGN(8192 / 5, 64),
428789d94c2Sjwadams 	P2ALIGN(8192 / 4, 64), 2304,
4297c478bd9Sstevel@tonic-gate 	P2ALIGN(8192 / 3, 64),
430789d94c2Sjwadams 	P2ALIGN(8192 / 2, 64), 4544,
431789d94c2Sjwadams 	P2ALIGN(8192 / 1, 64), 9216,
4327c478bd9Sstevel@tonic-gate 	4096 * 3,
433*38849194SRobert Mustacchi 	8192 * 2,				/* = 8192 * 2 */
434*38849194SRobert Mustacchi 	24576, 32768, 40960, 49152, 57344, 65536, 73728, 81920,
435*38849194SRobert Mustacchi 	90112, 98304, 106496, 114688, 122880, UMEM_MAXBUF, /* 128k */
436789d94c2Sjwadams 	/* 24 slots for user expansion */
437789d94c2Sjwadams 	0, 0, 0, 0, 0, 0, 0, 0,
438789d94c2Sjwadams 	0, 0, 0, 0, 0, 0, 0, 0,
439789d94c2Sjwadams 	0, 0, 0, 0, 0, 0, 0, 0,
4407c478bd9Sstevel@tonic-gate };
4417c478bd9Sstevel@tonic-gate #define	NUM_ALLOC_SIZES (sizeof (umem_alloc_sizes) / sizeof (*umem_alloc_sizes))
4427c478bd9Sstevel@tonic-gate 
4437c478bd9Sstevel@tonic-gate static umem_magtype_t umem_magtype[] = {
4447c478bd9Sstevel@tonic-gate 	{ 1,	8,	3200,	65536	},
4457c478bd9Sstevel@tonic-gate 	{ 3,	16,	256,	32768	},
4467c478bd9Sstevel@tonic-gate 	{ 7,	32,	64,	16384	},
4477c478bd9Sstevel@tonic-gate 	{ 15,	64,	0,	8192	},
4487c478bd9Sstevel@tonic-gate 	{ 31,	64,	0,	4096	},
4497c478bd9Sstevel@tonic-gate 	{ 47,	64,	0,	2048	},
4507c478bd9Sstevel@tonic-gate 	{ 63,	64,	0,	1024	},
4517c478bd9Sstevel@tonic-gate 	{ 95,	64,	0,	512	},
4527c478bd9Sstevel@tonic-gate 	{ 143,	64,	0,	0	},
4537c478bd9Sstevel@tonic-gate };
4547c478bd9Sstevel@tonic-gate 
4557c478bd9Sstevel@tonic-gate /*
4567c478bd9Sstevel@tonic-gate  * umem tunables
4577c478bd9Sstevel@tonic-gate  */
4587c478bd9Sstevel@tonic-gate uint32_t umem_max_ncpus;	/* # of CPU caches. */
4597c478bd9Sstevel@tonic-gate 
4607c478bd9Sstevel@tonic-gate uint32_t umem_stack_depth = 15; /* # stack frames in a bufctl_audit */
4617c478bd9Sstevel@tonic-gate uint32_t umem_reap_interval = 10; /* max reaping rate (seconds) */
4627c478bd9Sstevel@tonic-gate uint_t umem_depot_contention = 2; /* max failed trylocks per real interval */
4637c478bd9Sstevel@tonic-gate uint_t umem_abort = 1;		/* whether to abort on error */
4647c478bd9Sstevel@tonic-gate uint_t umem_output = 0;		/* whether to write to standard error */
4657c478bd9Sstevel@tonic-gate uint_t umem_logging = 0;	/* umem_log_enter() override */
4667c478bd9Sstevel@tonic-gate uint32_t umem_mtbf = 0;		/* mean time between failures [default: off] */
4677c478bd9Sstevel@tonic-gate size_t umem_transaction_log_size; /* size of transaction log */
4687c478bd9Sstevel@tonic-gate size_t umem_content_log_size;	/* size of content log */
4697c478bd9Sstevel@tonic-gate size_t umem_failure_log_size;	/* failure log [4 pages per CPU] */
4707c478bd9Sstevel@tonic-gate size_t umem_slab_log_size;	/* slab create log [4 pages per CPU] */
4717c478bd9Sstevel@tonic-gate size_t umem_content_maxsave = 256; /* UMF_CONTENTS max bytes to log */
4727c478bd9Sstevel@tonic-gate size_t umem_lite_minsize = 0;	/* minimum buffer size for UMF_LITE */
4737c478bd9Sstevel@tonic-gate size_t umem_lite_maxalign = 1024; /* maximum buffer alignment for UMF_LITE */
4747c478bd9Sstevel@tonic-gate size_t umem_maxverify;		/* maximum bytes to inspect in debug routines */
4757c478bd9Sstevel@tonic-gate size_t umem_minfirewall;	/* hardware-enforced redzone threshold */
4767c478bd9Sstevel@tonic-gate 
4777c478bd9Sstevel@tonic-gate uint_t umem_flags = 0;
4787c478bd9Sstevel@tonic-gate 
4797c478bd9Sstevel@tonic-gate mutex_t			umem_init_lock;		/* locks initialization */
4807c478bd9Sstevel@tonic-gate cond_t			umem_init_cv;		/* initialization CV */
4817c478bd9Sstevel@tonic-gate thread_t		umem_init_thr;		/* thread initializing */
4827c478bd9Sstevel@tonic-gate int			umem_init_env_ready;	/* environ pre-initted */
4837c478bd9Sstevel@tonic-gate int			umem_ready = UMEM_READY_STARTUP;
4847c478bd9Sstevel@tonic-gate 
4857c478bd9Sstevel@tonic-gate static umem_nofail_callback_t *nofail_callback;
4867c478bd9Sstevel@tonic-gate static mutex_t		umem_nofail_exit_lock;
4877c478bd9Sstevel@tonic-gate static thread_t		umem_nofail_exit_thr;
4887c478bd9Sstevel@tonic-gate 
4897c478bd9Sstevel@tonic-gate static umem_cache_t	*umem_slab_cache;
4907c478bd9Sstevel@tonic-gate static umem_cache_t	*umem_bufctl_cache;
4917c478bd9Sstevel@tonic-gate static umem_cache_t	*umem_bufctl_audit_cache;
4927c478bd9Sstevel@tonic-gate 
4937c478bd9Sstevel@tonic-gate mutex_t			umem_flags_lock;
4947c478bd9Sstevel@tonic-gate 
4957c478bd9Sstevel@tonic-gate static vmem_t		*heap_arena;
4967c478bd9Sstevel@tonic-gate static vmem_alloc_t	*heap_alloc;
4977c478bd9Sstevel@tonic-gate static vmem_free_t	*heap_free;
4987c478bd9Sstevel@tonic-gate 
4997c478bd9Sstevel@tonic-gate static vmem_t		*umem_internal_arena;
5007c478bd9Sstevel@tonic-gate static vmem_t		*umem_cache_arena;
5017c478bd9Sstevel@tonic-gate static vmem_t		*umem_hash_arena;
5027c478bd9Sstevel@tonic-gate static vmem_t		*umem_log_arena;
5037c478bd9Sstevel@tonic-gate static vmem_t		*umem_oversize_arena;
5047c478bd9Sstevel@tonic-gate static vmem_t		*umem_va_arena;
5057c478bd9Sstevel@tonic-gate static vmem_t		*umem_default_arena;
5067c478bd9Sstevel@tonic-gate static vmem_t		*umem_firewall_va_arena;
5077c478bd9Sstevel@tonic-gate static vmem_t		*umem_firewall_arena;
5087c478bd9Sstevel@tonic-gate 
5097c478bd9Sstevel@tonic-gate vmem_t			*umem_memalign_arena;
5107c478bd9Sstevel@tonic-gate 
5117c478bd9Sstevel@tonic-gate umem_log_header_t *umem_transaction_log;
5127c478bd9Sstevel@tonic-gate umem_log_header_t *umem_content_log;
5137c478bd9Sstevel@tonic-gate umem_log_header_t *umem_failure_log;
5147c478bd9Sstevel@tonic-gate umem_log_header_t *umem_slab_log;
5157c478bd9Sstevel@tonic-gate 
5167257d1b4Sraf #define	CPUHINT()		(thr_self())
5177c478bd9Sstevel@tonic-gate #define	CPUHINT_MAX()		INT_MAX
5187c478bd9Sstevel@tonic-gate 
5197c478bd9Sstevel@tonic-gate #define	CPU(mask)		(umem_cpus + (CPUHINT() & (mask)))
5207c478bd9Sstevel@tonic-gate static umem_cpu_t umem_startup_cpu = {	/* initial, single, cpu */
5217c478bd9Sstevel@tonic-gate 	UMEM_CACHE_SIZE(0),
5227c478bd9Sstevel@tonic-gate 	0
5237c478bd9Sstevel@tonic-gate };
5247c478bd9Sstevel@tonic-gate 
5257c478bd9Sstevel@tonic-gate static uint32_t umem_cpu_mask = 0;			/* global cpu mask */
5267c478bd9Sstevel@tonic-gate static umem_cpu_t *umem_cpus = &umem_startup_cpu;	/* cpu list */
5277c478bd9Sstevel@tonic-gate 
5287c478bd9Sstevel@tonic-gate volatile uint32_t umem_reaping;
5297c478bd9Sstevel@tonic-gate 
5307c478bd9Sstevel@tonic-gate thread_t		umem_update_thr;
5317c478bd9Sstevel@tonic-gate struct timeval		umem_update_next;	/* timeofday of next update */
5327c478bd9Sstevel@tonic-gate volatile thread_t	umem_st_update_thr;	/* only used when single-thd */
5337c478bd9Sstevel@tonic-gate 
5347c478bd9Sstevel@tonic-gate #define	IN_UPDATE()	(thr_self() == umem_update_thr || \
5357c478bd9Sstevel@tonic-gate 			    thr_self() == umem_st_update_thr)
5367c478bd9Sstevel@tonic-gate #define	IN_REAP()	IN_UPDATE()
5377c478bd9Sstevel@tonic-gate 
5387c478bd9Sstevel@tonic-gate mutex_t			umem_update_lock;	/* cache_u{next,prev,flags} */
5397c478bd9Sstevel@tonic-gate cond_t			umem_update_cv;
5407c478bd9Sstevel@tonic-gate 
5417c478bd9Sstevel@tonic-gate volatile hrtime_t umem_reap_next;	/* min hrtime of next reap */
5427c478bd9Sstevel@tonic-gate 
5437c478bd9Sstevel@tonic-gate mutex_t			umem_cache_lock;	/* inter-cache linkage only */
5447c478bd9Sstevel@tonic-gate 
5457c478bd9Sstevel@tonic-gate #ifdef UMEM_STANDALONE
5467c478bd9Sstevel@tonic-gate umem_cache_t		umem_null_cache;
5477c478bd9Sstevel@tonic-gate static const umem_cache_t umem_null_cache_template = {
5487c478bd9Sstevel@tonic-gate #else
5497c478bd9Sstevel@tonic-gate umem_cache_t		umem_null_cache = {
5507c478bd9Sstevel@tonic-gate #endif
5517c478bd9Sstevel@tonic-gate 	0, 0, 0, 0, 0,
5527c478bd9Sstevel@tonic-gate 	0, 0,
5537c478bd9Sstevel@tonic-gate 	0, 0,
5547c478bd9Sstevel@tonic-gate 	0, 0,
5557c478bd9Sstevel@tonic-gate 	"invalid_cache",
5567c478bd9Sstevel@tonic-gate 	0, 0,
5577c478bd9Sstevel@tonic-gate 	NULL, NULL, NULL, NULL,
5587c478bd9Sstevel@tonic-gate 	NULL,
5597c478bd9Sstevel@tonic-gate 	0, 0, 0, 0,
5607c478bd9Sstevel@tonic-gate 	&umem_null_cache, &umem_null_cache,
5617c478bd9Sstevel@tonic-gate 	&umem_null_cache, &umem_null_cache,
5627c478bd9Sstevel@tonic-gate 	0,
5637c478bd9Sstevel@tonic-gate 	DEFAULTMUTEX,				/* start of slab layer */
5647c478bd9Sstevel@tonic-gate 	0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
5657c478bd9Sstevel@tonic-gate 	&umem_null_cache.cache_nullslab,
5667c478bd9Sstevel@tonic-gate 	{
5677c478bd9Sstevel@tonic-gate 		&umem_null_cache,
5687c478bd9Sstevel@tonic-gate 		NULL,
5697c478bd9Sstevel@tonic-gate 		&umem_null_cache.cache_nullslab,
5707c478bd9Sstevel@tonic-gate 		&umem_null_cache.cache_nullslab,
5717c478bd9Sstevel@tonic-gate 		NULL,
5727c478bd9Sstevel@tonic-gate 		-1,
5737c478bd9Sstevel@tonic-gate 		0
5747c478bd9Sstevel@tonic-gate 	},
5757c478bd9Sstevel@tonic-gate 	NULL,
5767c478bd9Sstevel@tonic-gate 	NULL,
5777c478bd9Sstevel@tonic-gate 	DEFAULTMUTEX,				/* start of depot layer */
5787c478bd9Sstevel@tonic-gate 	NULL, {
5797c478bd9Sstevel@tonic-gate 		NULL, 0, 0, 0, 0
5807c478bd9Sstevel@tonic-gate 	}, {
5817c478bd9Sstevel@tonic-gate 		NULL, 0, 0, 0, 0
5827c478bd9Sstevel@tonic-gate 	}, {
5837c478bd9Sstevel@tonic-gate 		{
5847c478bd9Sstevel@tonic-gate 			DEFAULTMUTEX,		/* start of CPU cache */
5857c478bd9Sstevel@tonic-gate 			0, 0, NULL, NULL, -1, -1, 0
5867c478bd9Sstevel@tonic-gate 		}
5877c478bd9Sstevel@tonic-gate 	}
5887c478bd9Sstevel@tonic-gate };
5897c478bd9Sstevel@tonic-gate 
5907c478bd9Sstevel@tonic-gate #define	ALLOC_TABLE_4 \
5917c478bd9Sstevel@tonic-gate 	&umem_null_cache, &umem_null_cache, &umem_null_cache, &umem_null_cache
5927c478bd9Sstevel@tonic-gate 
5937c478bd9Sstevel@tonic-gate #define	ALLOC_TABLE_64 \
5947c478bd9Sstevel@tonic-gate 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
5957c478bd9Sstevel@tonic-gate 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
5967c478bd9Sstevel@tonic-gate 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
5977c478bd9Sstevel@tonic-gate 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4
5987c478bd9Sstevel@tonic-gate 
5997c478bd9Sstevel@tonic-gate #define	ALLOC_TABLE_1024 \
6007c478bd9Sstevel@tonic-gate 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
6017c478bd9Sstevel@tonic-gate 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
6027c478bd9Sstevel@tonic-gate 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
6037c478bd9Sstevel@tonic-gate 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64
6047c478bd9Sstevel@tonic-gate 
6057c478bd9Sstevel@tonic-gate static umem_cache_t *umem_alloc_table[UMEM_MAXBUF >> UMEM_ALIGN_SHIFT] = {
606*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
607*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
608*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
609*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
610*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
611*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
612*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
613*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
614*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
615*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
616*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
617*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
618*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
619*38849194SRobert Mustacchi 	ALLOC_TABLE_1024,
6207c478bd9Sstevel@tonic-gate 	ALLOC_TABLE_1024,
6217c478bd9Sstevel@tonic-gate 	ALLOC_TABLE_1024
6227c478bd9Sstevel@tonic-gate };
6237c478bd9Sstevel@tonic-gate 
6247c478bd9Sstevel@tonic-gate 
6257c478bd9Sstevel@tonic-gate /* Used to constrain audit-log stack traces */
6267c478bd9Sstevel@tonic-gate caddr_t			umem_min_stack;
6277c478bd9Sstevel@tonic-gate caddr_t			umem_max_stack;
6287c478bd9Sstevel@tonic-gate 
6297c478bd9Sstevel@tonic-gate 
6307c478bd9Sstevel@tonic-gate #define	UMERR_MODIFIED	0	/* buffer modified while on freelist */
6317c478bd9Sstevel@tonic-gate #define	UMERR_REDZONE	1	/* redzone violation (write past end of buf) */
6327c478bd9Sstevel@tonic-gate #define	UMERR_DUPFREE	2	/* freed a buffer twice */
6337c478bd9Sstevel@tonic-gate #define	UMERR_BADADDR	3	/* freed a bad (unallocated) address */
6347c478bd9Sstevel@tonic-gate #define	UMERR_BADBUFTAG	4	/* buftag corrupted */
6357c478bd9Sstevel@tonic-gate #define	UMERR_BADBUFCTL	5	/* bufctl corrupted */
6367c478bd9Sstevel@tonic-gate #define	UMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
6377c478bd9Sstevel@tonic-gate #define	UMERR_BADSIZE	7	/* alloc size != free size */
6387c478bd9Sstevel@tonic-gate #define	UMERR_BADBASE	8	/* buffer base address wrong */
6397c478bd9Sstevel@tonic-gate 
6407c478bd9Sstevel@tonic-gate struct {
6417c478bd9Sstevel@tonic-gate 	hrtime_t	ump_timestamp;	/* timestamp of error */
6427c478bd9Sstevel@tonic-gate 	int		ump_error;	/* type of umem error (UMERR_*) */
6437c478bd9Sstevel@tonic-gate 	void		*ump_buffer;	/* buffer that induced abort */
6447c478bd9Sstevel@tonic-gate 	void		*ump_realbuf;	/* real start address for buffer */
6457c478bd9Sstevel@tonic-gate 	umem_cache_t	*ump_cache;	/* buffer's cache according to client */
6467c478bd9Sstevel@tonic-gate 	umem_cache_t	*ump_realcache;	/* actual cache containing buffer */
6477c478bd9Sstevel@tonic-gate 	umem_slab_t	*ump_slab;	/* slab accoring to umem_findslab() */
6487c478bd9Sstevel@tonic-gate 	umem_bufctl_t	*ump_bufctl;	/* bufctl */
6497c478bd9Sstevel@tonic-gate } umem_abort_info;
6507c478bd9Sstevel@tonic-gate 
6517c478bd9Sstevel@tonic-gate static void
6527c478bd9Sstevel@tonic-gate copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
6537c478bd9Sstevel@tonic-gate {
6547c478bd9Sstevel@tonic-gate 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
6557c478bd9Sstevel@tonic-gate 	uint64_t *buf = buf_arg;
6567c478bd9Sstevel@tonic-gate 
6577c478bd9Sstevel@tonic-gate 	while (buf < bufend)
6587c478bd9Sstevel@tonic-gate 		*buf++ = pattern;
6597c478bd9Sstevel@tonic-gate }
6607c478bd9Sstevel@tonic-gate 
6617c478bd9Sstevel@tonic-gate static void *
6627c478bd9Sstevel@tonic-gate verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
6637c478bd9Sstevel@tonic-gate {
6647c478bd9Sstevel@tonic-gate 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
6657c478bd9Sstevel@tonic-gate 	uint64_t *buf;
6667c478bd9Sstevel@tonic-gate 
6677c478bd9Sstevel@tonic-gate 	for (buf = buf_arg; buf < bufend; buf++)
6687c478bd9Sstevel@tonic-gate 		if (*buf != pattern)
6697c478bd9Sstevel@tonic-gate 			return (buf);
6707c478bd9Sstevel@tonic-gate 	return (NULL);
6717c478bd9Sstevel@tonic-gate }
6727c478bd9Sstevel@tonic-gate 
6737c478bd9Sstevel@tonic-gate static void *
6747c478bd9Sstevel@tonic-gate verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
6757c478bd9Sstevel@tonic-gate {
6767c478bd9Sstevel@tonic-gate 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
6777c478bd9Sstevel@tonic-gate 	uint64_t *buf;
6787c478bd9Sstevel@tonic-gate 
6797c478bd9Sstevel@tonic-gate 	for (buf = buf_arg; buf < bufend; buf++) {
6807c478bd9Sstevel@tonic-gate 		if (*buf != old) {
6817c478bd9Sstevel@tonic-gate 			copy_pattern(old, buf_arg,
6827c478bd9Sstevel@tonic-gate 			    (char *)buf - (char *)buf_arg);
6837c478bd9Sstevel@tonic-gate 			return (buf);
6847c478bd9Sstevel@tonic-gate 		}
6857c478bd9Sstevel@tonic-gate 		*buf = new;
6867c478bd9Sstevel@tonic-gate 	}
6877c478bd9Sstevel@tonic-gate 
6887c478bd9Sstevel@tonic-gate 	return (NULL);
6897c478bd9Sstevel@tonic-gate }
6907c478bd9Sstevel@tonic-gate 
6917c478bd9Sstevel@tonic-gate void
6927c478bd9Sstevel@tonic-gate umem_cache_applyall(void (*func)(umem_cache_t *))
6937c478bd9Sstevel@tonic-gate {
6947c478bd9Sstevel@tonic-gate 	umem_cache_t *cp;
6957c478bd9Sstevel@tonic-gate 
6967c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&umem_cache_lock);
6977c478bd9Sstevel@tonic-gate 	for (cp = umem_null_cache.cache_next; cp != &umem_null_cache;
6987c478bd9Sstevel@tonic-gate 	    cp = cp->cache_next)
6997c478bd9Sstevel@tonic-gate 		func(cp);
7007c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&umem_cache_lock);
7017c478bd9Sstevel@tonic-gate }
7027c478bd9Sstevel@tonic-gate 
7037c478bd9Sstevel@tonic-gate static void
7047c478bd9Sstevel@tonic-gate umem_add_update_unlocked(umem_cache_t *cp, int flags)
7057c478bd9Sstevel@tonic-gate {
7067c478bd9Sstevel@tonic-gate 	umem_cache_t *cnext, *cprev;
7077c478bd9Sstevel@tonic-gate 
7087c478bd9Sstevel@tonic-gate 	flags &= ~UMU_ACTIVE;
7097c478bd9Sstevel@tonic-gate 
7107c478bd9Sstevel@tonic-gate 	if (!flags)
7117c478bd9Sstevel@tonic-gate 		return;
7127c478bd9Sstevel@tonic-gate 
7137c478bd9Sstevel@tonic-gate 	if (cp->cache_uflags & UMU_ACTIVE) {
7147c478bd9Sstevel@tonic-gate 		cp->cache_uflags |= flags;
7157c478bd9Sstevel@tonic-gate 	} else {
7167c478bd9Sstevel@tonic-gate 		if (cp->cache_unext != NULL) {
7177c478bd9Sstevel@tonic-gate 			ASSERT(cp->cache_uflags != 0);
7187c478bd9Sstevel@tonic-gate 			cp->cache_uflags |= flags;
7197c478bd9Sstevel@tonic-gate 		} else {
7207c478bd9Sstevel@tonic-gate 			ASSERT(cp->cache_uflags == 0);
7217c478bd9Sstevel@tonic-gate 			cp->cache_uflags = flags;
7227c478bd9Sstevel@tonic-gate 			cp->cache_unext = cnext = &umem_null_cache;
7237c478bd9Sstevel@tonic-gate 			cp->cache_uprev = cprev = umem_null_cache.cache_uprev;
7247c478bd9Sstevel@tonic-gate 			cnext->cache_uprev = cp;
7257c478bd9Sstevel@tonic-gate 			cprev->cache_unext = cp;
7267c478bd9Sstevel@tonic-gate 		}
7277c478bd9Sstevel@tonic-gate 	}
7287c478bd9Sstevel@tonic-gate }
7297c478bd9Sstevel@tonic-gate 
7307c478bd9Sstevel@tonic-gate static void
7317c478bd9Sstevel@tonic-gate umem_add_update(umem_cache_t *cp, int flags)
7327c478bd9Sstevel@tonic-gate {
7337c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&umem_update_lock);
7347c478bd9Sstevel@tonic-gate 
7357c478bd9Sstevel@tonic-gate 	umem_add_update_unlocked(cp, flags);
7367c478bd9Sstevel@tonic-gate 
7377c478bd9Sstevel@tonic-gate 	if (!IN_UPDATE())
7387c478bd9Sstevel@tonic-gate 		(void) cond_broadcast(&umem_update_cv);
7397c478bd9Sstevel@tonic-gate 
7407c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&umem_update_lock);
7417c478bd9Sstevel@tonic-gate }
7427c478bd9Sstevel@tonic-gate 
7437c478bd9Sstevel@tonic-gate /*
7447c478bd9Sstevel@tonic-gate  * Remove a cache from the update list, waiting for any in-progress work to
7457c478bd9Sstevel@tonic-gate  * complete first.
7467c478bd9Sstevel@tonic-gate  */
7477c478bd9Sstevel@tonic-gate static void
7487c478bd9Sstevel@tonic-gate umem_remove_updates(umem_cache_t *cp)
7497c478bd9Sstevel@tonic-gate {
7507c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&umem_update_lock);
7517c478bd9Sstevel@tonic-gate 
7527c478bd9Sstevel@tonic-gate 	/*
7537c478bd9Sstevel@tonic-gate 	 * Get it out of the active state
7547c478bd9Sstevel@tonic-gate 	 */
7557c478bd9Sstevel@tonic-gate 	while (cp->cache_uflags & UMU_ACTIVE) {
756a574db85Sraf 		int cancel_state;
757a574db85Sraf 
7587c478bd9Sstevel@tonic-gate 		ASSERT(cp->cache_unext == NULL);
7597c478bd9Sstevel@tonic-gate 
7607c478bd9Sstevel@tonic-gate 		cp->cache_uflags |= UMU_NOTIFY;
7617c478bd9Sstevel@tonic-gate 
7627c478bd9Sstevel@tonic-gate 		/*
7637c478bd9Sstevel@tonic-gate 		 * Make sure the update state is sane, before we wait
7647c478bd9Sstevel@tonic-gate 		 */
7657c478bd9Sstevel@tonic-gate 		ASSERT(umem_update_thr != 0 || umem_st_update_thr != 0);
7667c478bd9Sstevel@tonic-gate 		ASSERT(umem_update_thr != thr_self() &&
7677c478bd9Sstevel@tonic-gate 		    umem_st_update_thr != thr_self());
7687c478bd9Sstevel@tonic-gate 
769a574db85Sraf 		(void) pthread_setcancelstate(PTHREAD_CANCEL_DISABLE,
770a574db85Sraf 		    &cancel_state);
771a574db85Sraf 		(void) cond_wait(&umem_update_cv, &umem_update_lock);
772a574db85Sraf 		(void) pthread_setcancelstate(cancel_state, NULL);
7737c478bd9Sstevel@tonic-gate 	}
7747c478bd9Sstevel@tonic-gate 	/*
7757c478bd9Sstevel@tonic-gate 	 * Get it out of the Work Requested state
7767c478bd9Sstevel@tonic-gate 	 */
7777c478bd9Sstevel@tonic-gate 	if (cp->cache_unext != NULL) {
7787c478bd9Sstevel@tonic-gate 		cp->cache_uprev->cache_unext = cp->cache_unext;
7797c478bd9Sstevel@tonic-gate 		cp->cache_unext->cache_uprev = cp->cache_uprev;
7807c478bd9Sstevel@tonic-gate 		cp->cache_uprev = cp->cache_unext = NULL;
7817c478bd9Sstevel@tonic-gate 		cp->cache_uflags = 0;
7827c478bd9Sstevel@tonic-gate 	}
7837c478bd9Sstevel@tonic-gate 	/*
7847c478bd9Sstevel@tonic-gate 	 * Make sure it is in the Inactive state
7857c478bd9Sstevel@tonic-gate 	 */
7867c478bd9Sstevel@tonic-gate 	ASSERT(cp->cache_unext == NULL && cp->cache_uflags == 0);
7877c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&umem_update_lock);
7887c478bd9Sstevel@tonic-gate }
7897c478bd9Sstevel@tonic-gate 
7907c478bd9Sstevel@tonic-gate static void
7917c478bd9Sstevel@tonic-gate umem_updateall(int flags)
7927c478bd9Sstevel@tonic-gate {
7937c478bd9Sstevel@tonic-gate 	umem_cache_t *cp;
7947c478bd9Sstevel@tonic-gate 
7957c478bd9Sstevel@tonic-gate 	/*
7967c478bd9Sstevel@tonic-gate 	 * NOTE:  To prevent deadlock, umem_cache_lock is always acquired first.
7977c478bd9Sstevel@tonic-gate 	 *
7987c478bd9Sstevel@tonic-gate 	 * (umem_add_update is called from things run via umem_cache_applyall)
7997c478bd9Sstevel@tonic-gate 	 */
8007c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&umem_cache_lock);
8017c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&umem_update_lock);
8027c478bd9Sstevel@tonic-gate 
8037c478bd9Sstevel@tonic-gate 	for (cp = umem_null_cache.cache_next; cp != &umem_null_cache;
8047c478bd9Sstevel@tonic-gate 	    cp = cp->cache_next)
8057c478bd9Sstevel@tonic-gate 		umem_add_update_unlocked(cp, flags);
8067c478bd9Sstevel@tonic-gate 
8077c478bd9Sstevel@tonic-gate 	if (!IN_UPDATE())
8087c478bd9Sstevel@tonic-gate 		(void) cond_broadcast(&umem_update_cv);
8097c478bd9Sstevel@tonic-gate 
8107c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&umem_update_lock);
8117c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&umem_cache_lock);
8127c478bd9Sstevel@tonic-gate }
8137c478bd9Sstevel@tonic-gate 
8147c478bd9Sstevel@tonic-gate /*
8157c478bd9Sstevel@tonic-gate  * Debugging support.  Given a buffer address, find its slab.
8167c478bd9Sstevel@tonic-gate  */
8177c478bd9Sstevel@tonic-gate static umem_slab_t *
8187c478bd9Sstevel@tonic-gate umem_findslab(umem_cache_t *cp, void *buf)
8197c478bd9Sstevel@tonic-gate {
8207c478bd9Sstevel@tonic-gate 	umem_slab_t *sp;
8217c478bd9Sstevel@tonic-gate 
8227c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&cp->cache_lock);
8237c478bd9Sstevel@tonic-gate 	for (sp = cp->cache_nullslab.slab_next;
8247c478bd9Sstevel@tonic-gate 	    sp != &cp->cache_nullslab; sp = sp->slab_next) {
8257c478bd9Sstevel@tonic-gate 		if (UMEM_SLAB_MEMBER(sp, buf)) {
8267c478bd9Sstevel@tonic-gate 			(void) mutex_unlock(&cp->cache_lock);
8277c478bd9Sstevel@tonic-gate 			return (sp);
8287c478bd9Sstevel@tonic-gate 		}
8297c478bd9Sstevel@tonic-gate 	}
8307c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&cp->cache_lock);
8317c478bd9Sstevel@tonic-gate 
8327c478bd9Sstevel@tonic-gate 	return (NULL);
8337c478bd9Sstevel@tonic-gate }
8347c478bd9Sstevel@tonic-gate 
8357c478bd9Sstevel@tonic-gate static void
8367c478bd9Sstevel@tonic-gate umem_error(int error, umem_cache_t *cparg, void *bufarg)
8377c478bd9Sstevel@tonic-gate {
8387c478bd9Sstevel@tonic-gate 	umem_buftag_t *btp = NULL;
8397c478bd9Sstevel@tonic-gate 	umem_bufctl_t *bcp = NULL;
8407c478bd9Sstevel@tonic-gate 	umem_cache_t *cp = cparg;
8417c478bd9Sstevel@tonic-gate 	umem_slab_t *sp;
8427c478bd9Sstevel@tonic-gate 	uint64_t *off;
8437c478bd9Sstevel@tonic-gate 	void *buf = bufarg;
8447c478bd9Sstevel@tonic-gate 
8457c478bd9Sstevel@tonic-gate 	int old_logging = umem_logging;
8467c478bd9Sstevel@tonic-gate 
8477c478bd9Sstevel@tonic-gate 	umem_logging = 0;	/* stop logging when a bad thing happens */
8487c478bd9Sstevel@tonic-gate 
8497c478bd9Sstevel@tonic-gate 	umem_abort_info.ump_timestamp = gethrtime();
8507c478bd9Sstevel@tonic-gate 
8517c478bd9Sstevel@tonic-gate 	sp = umem_findslab(cp, buf);
8527c478bd9Sstevel@tonic-gate 	if (sp == NULL) {
8537c478bd9Sstevel@tonic-gate 		for (cp = umem_null_cache.cache_prev; cp != &umem_null_cache;
8547c478bd9Sstevel@tonic-gate 		    cp = cp->cache_prev) {
8557c478bd9Sstevel@tonic-gate 			if ((sp = umem_findslab(cp, buf)) != NULL)
8567c478bd9Sstevel@tonic-gate 				break;
8577c478bd9Sstevel@tonic-gate 		}
8587c478bd9Sstevel@tonic-gate 	}
8597c478bd9Sstevel@tonic-gate 
8607c478bd9Sstevel@tonic-gate 	if (sp == NULL) {
8617c478bd9Sstevel@tonic-gate 		cp = NULL;
8627c478bd9Sstevel@tonic-gate 		error = UMERR_BADADDR;
8637c478bd9Sstevel@tonic-gate 	} else {
8647c478bd9Sstevel@tonic-gate 		if (cp != cparg)
8657c478bd9Sstevel@tonic-gate 			error = UMERR_BADCACHE;
8667c478bd9Sstevel@tonic-gate 		else
8677c478bd9Sstevel@tonic-gate 			buf = (char *)bufarg - ((uintptr_t)bufarg -
8687c478bd9Sstevel@tonic-gate 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
8697c478bd9Sstevel@tonic-gate 		if (buf != bufarg)
8707c478bd9Sstevel@tonic-gate 			error = UMERR_BADBASE;
8717c478bd9Sstevel@tonic-gate 		if (cp->cache_flags & UMF_BUFTAG)
8727c478bd9Sstevel@tonic-gate 			btp = UMEM_BUFTAG(cp, buf);
8737c478bd9Sstevel@tonic-gate 		if (cp->cache_flags & UMF_HASH) {
8747c478bd9Sstevel@tonic-gate 			(void) mutex_lock(&cp->cache_lock);
8757c478bd9Sstevel@tonic-gate 			for (bcp = *UMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
8767c478bd9Sstevel@tonic-gate 				if (bcp->bc_addr == buf)
8777c478bd9Sstevel@tonic-gate 					break;
8787c478bd9Sstevel@tonic-gate 			(void) mutex_unlock(&cp->cache_lock);
8797c478bd9Sstevel@tonic-gate 			if (bcp == NULL && btp != NULL)
8807c478bd9Sstevel@tonic-gate 				bcp = btp->bt_bufctl;
8817c478bd9Sstevel@tonic-gate 			if (umem_findslab(cp->cache_bufctl_cache, bcp) ==
8827c478bd9Sstevel@tonic-gate 			    NULL || P2PHASE((uintptr_t)bcp, UMEM_ALIGN) ||
8837c478bd9Sstevel@tonic-gate 			    bcp->bc_addr != buf) {
8847c478bd9Sstevel@tonic-gate 				error = UMERR_BADBUFCTL;
8857c478bd9Sstevel@tonic-gate 				bcp = NULL;
8867c478bd9Sstevel@tonic-gate 			}
8877c478bd9Sstevel@tonic-gate 		}
8887c478bd9Sstevel@tonic-gate 	}
8897c478bd9Sstevel@tonic-gate 
8907c478bd9Sstevel@tonic-gate 	umem_abort_info.ump_error = error;
8917c478bd9Sstevel@tonic-gate 	umem_abort_info.ump_buffer = bufarg;
8927c478bd9Sstevel@tonic-gate 	umem_abort_info.ump_realbuf = buf;
8937c478bd9Sstevel@tonic-gate 	umem_abort_info.ump_cache = cparg;
8947c478bd9Sstevel@tonic-gate 	umem_abort_info.ump_realcache = cp;
8957c478bd9Sstevel@tonic-gate 	umem_abort_info.ump_slab = sp;
8967c478bd9Sstevel@tonic-gate 	umem_abort_info.ump_bufctl = bcp;
8977c478bd9Sstevel@tonic-gate 
8987c478bd9Sstevel@tonic-gate 	umem_printf("umem allocator: ");
8997c478bd9Sstevel@tonic-gate 
9007c478bd9Sstevel@tonic-gate 	switch (error) {
9017c478bd9Sstevel@tonic-gate 
9027c478bd9Sstevel@tonic-gate 	case UMERR_MODIFIED:
9037c478bd9Sstevel@tonic-gate 		umem_printf("buffer modified after being freed\n");
9047c478bd9Sstevel@tonic-gate 		off = verify_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
9057c478bd9Sstevel@tonic-gate 		if (off == NULL)	/* shouldn't happen */
9067c478bd9Sstevel@tonic-gate 			off = buf;
9077c478bd9Sstevel@tonic-gate 		umem_printf("modification occurred at offset 0x%lx "
9087c478bd9Sstevel@tonic-gate 		    "(0x%llx replaced by 0x%llx)\n",
9097c478bd9Sstevel@tonic-gate 		    (uintptr_t)off - (uintptr_t)buf,
9107c478bd9Sstevel@tonic-gate 		    (longlong_t)UMEM_FREE_PATTERN, (longlong_t)*off);
9117c478bd9Sstevel@tonic-gate 		break;
9127c478bd9Sstevel@tonic-gate 
9137c478bd9Sstevel@tonic-gate 	case UMERR_REDZONE:
9147c478bd9Sstevel@tonic-gate 		umem_printf("redzone violation: write past end of buffer\n");
9157c478bd9Sstevel@tonic-gate 		break;
9167c478bd9Sstevel@tonic-gate 
9177c478bd9Sstevel@tonic-gate 	case UMERR_BADADDR:
9187c478bd9Sstevel@tonic-gate 		umem_printf("invalid free: buffer not in cache\n");
9197c478bd9Sstevel@tonic-gate 		break;
9207c478bd9Sstevel@tonic-gate 
9217c478bd9Sstevel@tonic-gate 	case UMERR_DUPFREE:
9227c478bd9Sstevel@tonic-gate 		umem_printf("duplicate free: buffer freed twice\n");
9237c478bd9Sstevel@tonic-gate 		break;
9247c478bd9Sstevel@tonic-gate 
9257c478bd9Sstevel@tonic-gate 	case UMERR_BADBUFTAG:
9267c478bd9Sstevel@tonic-gate 		umem_printf("boundary tag corrupted\n");
9277c478bd9Sstevel@tonic-gate 		umem_printf("bcp ^ bxstat = %lx, should be %lx\n",
9287c478bd9Sstevel@tonic-gate 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
9297c478bd9Sstevel@tonic-gate 		    UMEM_BUFTAG_FREE);
9307c478bd9Sstevel@tonic-gate 		break;
9317c478bd9Sstevel@tonic-gate 
9327c478bd9Sstevel@tonic-gate 	case UMERR_BADBUFCTL:
9337c478bd9Sstevel@tonic-gate 		umem_printf("bufctl corrupted\n");
9347c478bd9Sstevel@tonic-gate 		break;
9357c478bd9Sstevel@tonic-gate 
9367c478bd9Sstevel@tonic-gate 	case UMERR_BADCACHE:
9377c478bd9Sstevel@tonic-gate 		umem_printf("buffer freed to wrong cache\n");
9387c478bd9Sstevel@tonic-gate 		umem_printf("buffer was allocated from %s,\n", cp->cache_name);
9397c478bd9Sstevel@tonic-gate 		umem_printf("caller attempting free to %s.\n",
9407c478bd9Sstevel@tonic-gate 		    cparg->cache_name);
9417c478bd9Sstevel@tonic-gate 		break;
9427c478bd9Sstevel@tonic-gate 
9437c478bd9Sstevel@tonic-gate 	case UMERR_BADSIZE:
9447c478bd9Sstevel@tonic-gate 		umem_printf("bad free: free size (%u) != alloc size (%u)\n",
9457c478bd9Sstevel@tonic-gate 		    UMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
9467c478bd9Sstevel@tonic-gate 		    UMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
9477c478bd9Sstevel@tonic-gate 		break;
9487c478bd9Sstevel@tonic-gate 
9497c478bd9Sstevel@tonic-gate 	case UMERR_BADBASE:
9507c478bd9Sstevel@tonic-gate 		umem_printf("bad free: free address (%p) != alloc address "
9517c478bd9Sstevel@tonic-gate 		    "(%p)\n", bufarg, buf);
9527c478bd9Sstevel@tonic-gate 		break;
9537c478bd9Sstevel@tonic-gate 	}
9547c478bd9Sstevel@tonic-gate 
9557c478bd9Sstevel@tonic-gate 	umem_printf("buffer=%p  bufctl=%p  cache: %s\n",
9567c478bd9Sstevel@tonic-gate 	    bufarg, (void *)bcp, cparg->cache_name);
9577c478bd9Sstevel@tonic-gate 
9587c478bd9Sstevel@tonic-gate 	if (bcp != NULL && (cp->cache_flags & UMF_AUDIT) &&
9597c478bd9Sstevel@tonic-gate 	    error != UMERR_BADBUFCTL) {
9607c478bd9Sstevel@tonic-gate 		int d;
9617c478bd9Sstevel@tonic-gate 		timespec_t ts;
9627c478bd9Sstevel@tonic-gate 		hrtime_t diff;
9637c478bd9Sstevel@tonic-gate 		umem_bufctl_audit_t *bcap = (umem_bufctl_audit_t *)bcp;
9647c478bd9Sstevel@tonic-gate 
9657c478bd9Sstevel@tonic-gate 		diff = umem_abort_info.ump_timestamp - bcap->bc_timestamp;
9667c478bd9Sstevel@tonic-gate 		ts.tv_sec = diff / NANOSEC;
9677c478bd9Sstevel@tonic-gate 		ts.tv_nsec = diff % NANOSEC;
9687c478bd9Sstevel@tonic-gate 
9697c478bd9Sstevel@tonic-gate 		umem_printf("previous transaction on buffer %p:\n", buf);
9707c478bd9Sstevel@tonic-gate 		umem_printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
9717c478bd9Sstevel@tonic-gate 		    (void *)(intptr_t)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
9727c478bd9Sstevel@tonic-gate 		    (void *)sp, cp->cache_name);
9737c478bd9Sstevel@tonic-gate 		for (d = 0; d < MIN(bcap->bc_depth, umem_stack_depth); d++) {
9747c478bd9Sstevel@tonic-gate 			(void) print_sym((void *)bcap->bc_stack[d]);
9757c478bd9Sstevel@tonic-gate 			umem_printf("\n");
9767c478bd9Sstevel@tonic-gate 		}
9777c478bd9Sstevel@tonic-gate 	}
9787c478bd9Sstevel@tonic-gate 
9797c478bd9Sstevel@tonic-gate 	umem_err_recoverable("umem: heap corruption detected");
9807c478bd9Sstevel@tonic-gate 
9817c478bd9Sstevel@tonic-gate 	umem_logging = old_logging;	/* resume logging */
9827c478bd9Sstevel@tonic-gate }
9837c478bd9Sstevel@tonic-gate 
9847c478bd9Sstevel@tonic-gate void
9857c478bd9Sstevel@tonic-gate umem_nofail_callback(umem_nofail_callback_t *cb)
9867c478bd9Sstevel@tonic-gate {
9877c478bd9Sstevel@tonic-gate 	nofail_callback = cb;
9887c478bd9Sstevel@tonic-gate }
9897c478bd9Sstevel@tonic-gate 
9907c478bd9Sstevel@tonic-gate static int
9917c478bd9Sstevel@tonic-gate umem_alloc_retry(umem_cache_t *cp, int umflag)
9927c478bd9Sstevel@tonic-gate {
9937c478bd9Sstevel@tonic-gate 	if (cp == &umem_null_cache) {
9947c478bd9Sstevel@tonic-gate 		if (umem_init())
9957c478bd9Sstevel@tonic-gate 			return (1);				/* retry */
9967c478bd9Sstevel@tonic-gate 		/*
9977c478bd9Sstevel@tonic-gate 		 * Initialization failed.  Do normal failure processing.
9987c478bd9Sstevel@tonic-gate 		 */
9997c478bd9Sstevel@tonic-gate 	}
10007c478bd9Sstevel@tonic-gate 	if (umflag & UMEM_NOFAIL) {
10017c478bd9Sstevel@tonic-gate 		int def_result = UMEM_CALLBACK_EXIT(255);
10027c478bd9Sstevel@tonic-gate 		int result = def_result;
10037c478bd9Sstevel@tonic-gate 		umem_nofail_callback_t *callback = nofail_callback;
10047c478bd9Sstevel@tonic-gate 
10057c478bd9Sstevel@tonic-gate 		if (callback != NULL)
10067c478bd9Sstevel@tonic-gate 			result = callback();
10077c478bd9Sstevel@tonic-gate 
10087c478bd9Sstevel@tonic-gate 		if (result == UMEM_CALLBACK_RETRY)
10097c478bd9Sstevel@tonic-gate 			return (1);
10107c478bd9Sstevel@tonic-gate 
10117c478bd9Sstevel@tonic-gate 		if ((result & ~0xFF) != UMEM_CALLBACK_EXIT(0)) {
10127c478bd9Sstevel@tonic-gate 			log_message("nofail callback returned %x\n", result);
10137c478bd9Sstevel@tonic-gate 			result = def_result;
10147c478bd9Sstevel@tonic-gate 		}
10157c478bd9Sstevel@tonic-gate 
10167c478bd9Sstevel@tonic-gate 		/*
10177c478bd9Sstevel@tonic-gate 		 * only one thread will call exit
10187c478bd9Sstevel@tonic-gate 		 */
10197c478bd9Sstevel@tonic-gate 		if (umem_nofail_exit_thr == thr_self())
10207c478bd9Sstevel@tonic-gate 			umem_panic("recursive UMEM_CALLBACK_EXIT()\n");
10217c478bd9Sstevel@tonic-gate 
10227c478bd9Sstevel@tonic-gate 		(void) mutex_lock(&umem_nofail_exit_lock);
10237c478bd9Sstevel@tonic-gate 		umem_nofail_exit_thr = thr_self();
10247c478bd9Sstevel@tonic-gate 		exit(result & 0xFF);
10257c478bd9Sstevel@tonic-gate 		/*NOTREACHED*/
10267c478bd9Sstevel@tonic-gate 	}
10277c478bd9Sstevel@tonic-gate 	return (0);
10287c478bd9Sstevel@tonic-gate }
10297c478bd9Sstevel@tonic-gate 
10307c478bd9Sstevel@tonic-gate static umem_log_header_t *
10317c478bd9Sstevel@tonic-gate umem_log_init(size_t logsize)
10327c478bd9Sstevel@tonic-gate {
10337c478bd9Sstevel@tonic-gate 	umem_log_header_t *lhp;
10347c478bd9Sstevel@tonic-gate 	int nchunks = 4 * umem_max_ncpus;
10357c478bd9Sstevel@tonic-gate 	size_t lhsize = offsetof(umem_log_header_t, lh_cpu[umem_max_ncpus]);
10367c478bd9Sstevel@tonic-gate 	int i;
10377c478bd9Sstevel@tonic-gate 
10387c478bd9Sstevel@tonic-gate 	if (logsize == 0)
10397c478bd9Sstevel@tonic-gate 		return (NULL);
10407c478bd9Sstevel@tonic-gate 
10417c478bd9Sstevel@tonic-gate 	/*
10427c478bd9Sstevel@tonic-gate 	 * Make sure that lhp->lh_cpu[] is nicely aligned
10437c478bd9Sstevel@tonic-gate 	 * to prevent false sharing of cache lines.
10447c478bd9Sstevel@tonic-gate 	 */
10457c478bd9Sstevel@tonic-gate 	lhsize = P2ROUNDUP(lhsize, UMEM_ALIGN);
10467c478bd9Sstevel@tonic-gate 	lhp = vmem_xalloc(umem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
10477c478bd9Sstevel@tonic-gate 	    NULL, NULL, VM_NOSLEEP);
10487c478bd9Sstevel@tonic-gate 	if (lhp == NULL)
10497c478bd9Sstevel@tonic-gate 		goto fail;
10507c478bd9Sstevel@tonic-gate 
10517c478bd9Sstevel@tonic-gate 	bzero(lhp, lhsize);
10527c478bd9Sstevel@tonic-gate 
10537c478bd9Sstevel@tonic-gate 	(void) mutex_init(&lhp->lh_lock, USYNC_THREAD, NULL);
10547c478bd9Sstevel@tonic-gate 	lhp->lh_nchunks = nchunks;
10557c478bd9Sstevel@tonic-gate 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks, PAGESIZE);
10567c478bd9Sstevel@tonic-gate 	if (lhp->lh_chunksize == 0)
10577c478bd9Sstevel@tonic-gate 		lhp->lh_chunksize = PAGESIZE;
10587c478bd9Sstevel@tonic-gate 
10597c478bd9Sstevel@tonic-gate 	lhp->lh_base = vmem_alloc(umem_log_arena,
10607c478bd9Sstevel@tonic-gate 	    lhp->lh_chunksize * nchunks, VM_NOSLEEP);
10617c478bd9Sstevel@tonic-gate 	if (lhp->lh_base == NULL)
10627c478bd9Sstevel@tonic-gate 		goto fail;
10637c478bd9Sstevel@tonic-gate 
10647c478bd9Sstevel@tonic-gate 	lhp->lh_free = vmem_alloc(umem_log_arena,
10657c478bd9Sstevel@tonic-gate 	    nchunks * sizeof (int), VM_NOSLEEP);
10667c478bd9Sstevel@tonic-gate 	if (lhp->lh_free == NULL)
10677c478bd9Sstevel@tonic-gate 		goto fail;
10687c478bd9Sstevel@tonic-gate 
10697c478bd9Sstevel@tonic-gate 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
10707c478bd9Sstevel@tonic-gate 
10717c478bd9Sstevel@tonic-gate 	for (i = 0; i < umem_max_ncpus; i++) {
10727c478bd9Sstevel@tonic-gate 		umem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
10737c478bd9Sstevel@tonic-gate 		(void) mutex_init(&clhp->clh_lock, USYNC_THREAD, NULL);
10747c478bd9Sstevel@tonic-gate 		clhp->clh_chunk = i;
10757c478bd9Sstevel@tonic-gate 	}
10767c478bd9Sstevel@tonic-gate 
10777c478bd9Sstevel@tonic-gate 	for (i = umem_max_ncpus; i < nchunks; i++)
10787c478bd9Sstevel@tonic-gate 		lhp->lh_free[i] = i;
10797c478bd9Sstevel@tonic-gate 
10807c478bd9Sstevel@tonic-gate 	lhp->lh_head = umem_max_ncpus;
10817c478bd9Sstevel@tonic-gate 	lhp->lh_tail = 0;
10827c478bd9Sstevel@tonic-gate 
10837c478bd9Sstevel@tonic-gate 	return (lhp);
10847c478bd9Sstevel@tonic-gate 
10857c478bd9Sstevel@tonic-gate fail:
10867c478bd9Sstevel@tonic-gate 	if (lhp != NULL) {
10877c478bd9Sstevel@tonic-gate 		if (lhp->lh_base != NULL)
10887c478bd9Sstevel@tonic-gate 			vmem_free(umem_log_arena, lhp->lh_base,
10897c478bd9Sstevel@tonic-gate 			    lhp->lh_chunksize * nchunks);
10907c478bd9Sstevel@tonic-gate 
10917c478bd9Sstevel@tonic-gate 		vmem_xfree(umem_log_arena, lhp, lhsize);
10927c478bd9Sstevel@tonic-gate 	}
10937c478bd9Sstevel@tonic-gate 	return (NULL);
10947c478bd9Sstevel@tonic-gate }
10957c478bd9Sstevel@tonic-gate 
10967c478bd9Sstevel@tonic-gate static void *
10977c478bd9Sstevel@tonic-gate umem_log_enter(umem_log_header_t *lhp, void *data, size_t size)
10987c478bd9Sstevel@tonic-gate {
10997c478bd9Sstevel@tonic-gate 	void *logspace;
11007c478bd9Sstevel@tonic-gate 	umem_cpu_log_header_t *clhp =
11017c478bd9Sstevel@tonic-gate 	    &lhp->lh_cpu[CPU(umem_cpu_mask)->cpu_number];
11027c478bd9Sstevel@tonic-gate 
11037c478bd9Sstevel@tonic-gate 	if (lhp == NULL || umem_logging == 0)
11047c478bd9Sstevel@tonic-gate 		return (NULL);
11057c478bd9Sstevel@tonic-gate 
11067c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&clhp->clh_lock);
11077c478bd9Sstevel@tonic-gate 	clhp->clh_hits++;
11087c478bd9Sstevel@tonic-gate 	if (size > clhp->clh_avail) {
11097c478bd9Sstevel@tonic-gate 		(void) mutex_lock(&lhp->lh_lock);
11107c478bd9Sstevel@tonic-gate 		lhp->lh_hits++;
11117c478bd9Sstevel@tonic-gate 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
11127c478bd9Sstevel@tonic-gate 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
11137c478bd9Sstevel@tonic-gate 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
11147c478bd9Sstevel@tonic-gate 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
11157c478bd9Sstevel@tonic-gate 		clhp->clh_current = lhp->lh_base +
11167c478bd9Sstevel@tonic-gate 		    clhp->clh_chunk * lhp->lh_chunksize;
11177c478bd9Sstevel@tonic-gate 		clhp->clh_avail = lhp->lh_chunksize;
11187c478bd9Sstevel@tonic-gate 		if (size > lhp->lh_chunksize)
11197c478bd9Sstevel@tonic-gate 			size = lhp->lh_chunksize;
11207c478bd9Sstevel@tonic-gate 		(void) mutex_unlock(&lhp->lh_lock);
11217c478bd9Sstevel@tonic-gate 	}
11227c478bd9Sstevel@tonic-gate 	logspace = clhp->clh_current;
11237c478bd9Sstevel@tonic-gate 	clhp->clh_current += size;
11247c478bd9Sstevel@tonic-gate 	clhp->clh_avail -= size;
11257c478bd9Sstevel@tonic-gate 	bcopy(data, logspace, size);
11267c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&clhp->clh_lock);
11277c478bd9Sstevel@tonic-gate 	return (logspace);
11287c478bd9Sstevel@tonic-gate }
11297c478bd9Sstevel@tonic-gate 
11307c478bd9Sstevel@tonic-gate #define	UMEM_AUDIT(lp, cp, bcp)						\
11317c478bd9Sstevel@tonic-gate {									\
11327c478bd9Sstevel@tonic-gate 	umem_bufctl_audit_t *_bcp = (umem_bufctl_audit_t *)(bcp);	\
11337c478bd9Sstevel@tonic-gate 	_bcp->bc_timestamp = gethrtime();				\
11347c478bd9Sstevel@tonic-gate 	_bcp->bc_thread = thr_self();					\
11357c478bd9Sstevel@tonic-gate 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, umem_stack_depth,	\
11367c478bd9Sstevel@tonic-gate 	    (cp != NULL) && (cp->cache_flags & UMF_CHECKSIGNAL));	\
11377c478bd9Sstevel@tonic-gate 	_bcp->bc_lastlog = umem_log_enter((lp), _bcp,			\
11387c478bd9Sstevel@tonic-gate 	    UMEM_BUFCTL_AUDIT_SIZE);					\
11397c478bd9Sstevel@tonic-gate }
11407c478bd9Sstevel@tonic-gate 
11417c478bd9Sstevel@tonic-gate static void
11427c478bd9Sstevel@tonic-gate umem_log_event(umem_log_header_t *lp, umem_cache_t *cp,
11437c478bd9Sstevel@tonic-gate 	umem_slab_t *sp, void *addr)
11447c478bd9Sstevel@tonic-gate {
11457c478bd9Sstevel@tonic-gate 	umem_bufctl_audit_t *bcp;
11467c478bd9Sstevel@tonic-gate 	UMEM_LOCAL_BUFCTL_AUDIT(&bcp);
11477c478bd9Sstevel@tonic-gate 
11487c478bd9Sstevel@tonic-gate 	bzero(bcp, UMEM_BUFCTL_AUDIT_SIZE);
11497c478bd9Sstevel@tonic-gate 	bcp->bc_addr = addr;
11507c478bd9Sstevel@tonic-gate 	bcp->bc_slab = sp;
11517c478bd9Sstevel@tonic-gate 	bcp->bc_cache = cp;
11527c478bd9Sstevel@tonic-gate 	UMEM_AUDIT(lp, cp, bcp);
11537c478bd9Sstevel@tonic-gate }
11547c478bd9Sstevel@tonic-gate 
11557c478bd9Sstevel@tonic-gate /*
11567c478bd9Sstevel@tonic-gate  * Create a new slab for cache cp.
11577c478bd9Sstevel@tonic-gate  */
11587c478bd9Sstevel@tonic-gate static umem_slab_t *
11597c478bd9Sstevel@tonic-gate umem_slab_create(umem_cache_t *cp, int umflag)
11607c478bd9Sstevel@tonic-gate {
11617c478bd9Sstevel@tonic-gate 	size_t slabsize = cp->cache_slabsize;
11627c478bd9Sstevel@tonic-gate 	size_t chunksize = cp->cache_chunksize;
11637c478bd9Sstevel@tonic-gate 	int cache_flags = cp->cache_flags;
11647c478bd9Sstevel@tonic-gate 	size_t color, chunks;
11657c478bd9Sstevel@tonic-gate 	char *buf, *slab;
11667c478bd9Sstevel@tonic-gate 	umem_slab_t *sp;
11677c478bd9Sstevel@tonic-gate 	umem_bufctl_t *bcp;
11687c478bd9Sstevel@tonic-gate 	vmem_t *vmp = cp->cache_arena;
11697c478bd9Sstevel@tonic-gate 
11707c478bd9Sstevel@tonic-gate 	color = cp->cache_color + cp->cache_align;
11717c478bd9Sstevel@tonic-gate 	if (color > cp->cache_maxcolor)
11727c478bd9Sstevel@tonic-gate 		color = cp->cache_mincolor;
11737c478bd9Sstevel@tonic-gate 	cp->cache_color = color;
11747c478bd9Sstevel@tonic-gate 
11757c478bd9Sstevel@tonic-gate 	slab = vmem_alloc(vmp, slabsize, UMEM_VMFLAGS(umflag));
11767c478bd9Sstevel@tonic-gate 
11777c478bd9Sstevel@tonic-gate 	if (slab == NULL)
11787c478bd9Sstevel@tonic-gate 		goto vmem_alloc_failure;
11797c478bd9Sstevel@tonic-gate 
11807c478bd9Sstevel@tonic-gate 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
11817c478bd9Sstevel@tonic-gate 
11827c478bd9Sstevel@tonic-gate 	if (!(cp->cache_cflags & UMC_NOTOUCH) &&
11837c478bd9Sstevel@tonic-gate 	    (cp->cache_flags & UMF_DEADBEEF))
11847c478bd9Sstevel@tonic-gate 		copy_pattern(UMEM_UNINITIALIZED_PATTERN, slab, slabsize);
11857c478bd9Sstevel@tonic-gate 
11867c478bd9Sstevel@tonic-gate 	if (cache_flags & UMF_HASH) {
11877c478bd9Sstevel@tonic-gate 		if ((sp = _umem_cache_alloc(umem_slab_cache, umflag)) == NULL)
11887c478bd9Sstevel@tonic-gate 			goto slab_alloc_failure;
11897c478bd9Sstevel@tonic-gate 		chunks = (slabsize - color) / chunksize;
11907c478bd9Sstevel@tonic-gate 	} else {
11917c478bd9Sstevel@tonic-gate 		sp = UMEM_SLAB(cp, slab);
11927c478bd9Sstevel@tonic-gate 		chunks = (slabsize - sizeof (umem_slab_t) - color) / chunksize;
11937c478bd9Sstevel@tonic-gate 	}
11947c478bd9Sstevel@tonic-gate 
11957c478bd9Sstevel@tonic-gate 	sp->slab_cache	= cp;
11967c478bd9Sstevel@tonic-gate 	sp->slab_head	= NULL;
11977c478bd9Sstevel@tonic-gate 	sp->slab_refcnt	= 0;
11987c478bd9Sstevel@tonic-gate 	sp->slab_base	= buf = slab + color;
11997c478bd9Sstevel@tonic-gate 	sp->slab_chunks	= chunks;
12007c478bd9Sstevel@tonic-gate 
12017c478bd9Sstevel@tonic-gate 	ASSERT(chunks > 0);
12027c478bd9Sstevel@tonic-gate 	while (chunks-- != 0) {
12037c478bd9Sstevel@tonic-gate 		if (cache_flags & UMF_HASH) {
12047c478bd9Sstevel@tonic-gate 			bcp = _umem_cache_alloc(cp->cache_bufctl_cache, umflag);
12057c478bd9Sstevel@tonic-gate 			if (bcp == NULL)
12067c478bd9Sstevel@tonic-gate 				goto bufctl_alloc_failure;
12077c478bd9Sstevel@tonic-gate 			if (cache_flags & UMF_AUDIT) {
12087c478bd9Sstevel@tonic-gate 				umem_bufctl_audit_t *bcap =
12097c478bd9Sstevel@tonic-gate 				    (umem_bufctl_audit_t *)bcp;
12107c478bd9Sstevel@tonic-gate 				bzero(bcap, UMEM_BUFCTL_AUDIT_SIZE);
12117c478bd9Sstevel@tonic-gate 				bcap->bc_cache = cp;
12127c478bd9Sstevel@tonic-gate 			}
12137c478bd9Sstevel@tonic-gate 			bcp->bc_addr = buf;
12147c478bd9Sstevel@tonic-gate 			bcp->bc_slab = sp;
12157c478bd9Sstevel@tonic-gate 		} else {
12167c478bd9Sstevel@tonic-gate 			bcp = UMEM_BUFCTL(cp, buf);
12177c478bd9Sstevel@tonic-gate 		}
12187c478bd9Sstevel@tonic-gate 		if (cache_flags & UMF_BUFTAG) {
12197c478bd9Sstevel@tonic-gate 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
12207c478bd9Sstevel@tonic-gate 			btp->bt_redzone = UMEM_REDZONE_PATTERN;
12217c478bd9Sstevel@tonic-gate 			btp->bt_bufctl = bcp;
12227c478bd9Sstevel@tonic-gate 			btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
12237c478bd9Sstevel@tonic-gate 			if (cache_flags & UMF_DEADBEEF) {
12247c478bd9Sstevel@tonic-gate 				copy_pattern(UMEM_FREE_PATTERN, buf,
12257c478bd9Sstevel@tonic-gate 				    cp->cache_verify);
12267c478bd9Sstevel@tonic-gate 			}
12277c478bd9Sstevel@tonic-gate 		}
12287c478bd9Sstevel@tonic-gate 		bcp->bc_next = sp->slab_head;
12297c478bd9Sstevel@tonic-gate 		sp->slab_head = bcp;
12307c478bd9Sstevel@tonic-gate 		buf += chunksize;
12317c478bd9Sstevel@tonic-gate 	}
12327c478bd9Sstevel@tonic-gate 
12337c478bd9Sstevel@tonic-gate 	umem_log_event(umem_slab_log, cp, sp, slab);
12347c478bd9Sstevel@tonic-gate 
12357c478bd9Sstevel@tonic-gate 	return (sp);
12367c478bd9Sstevel@tonic-gate 
12377c478bd9Sstevel@tonic-gate bufctl_alloc_failure:
12387c478bd9Sstevel@tonic-gate 
12397c478bd9Sstevel@tonic-gate 	while ((bcp = sp->slab_head) != NULL) {
12407c478bd9Sstevel@tonic-gate 		sp->slab_head = bcp->bc_next;
12417c478bd9Sstevel@tonic-gate 		_umem_cache_free(cp->cache_bufctl_cache, bcp);
12427c478bd9Sstevel@tonic-gate 	}
12437c478bd9Sstevel@tonic-gate 	_umem_cache_free(umem_slab_cache, sp);
12447c478bd9Sstevel@tonic-gate 
12457c478bd9Sstevel@tonic-gate slab_alloc_failure:
12467c478bd9Sstevel@tonic-gate 
12477c478bd9Sstevel@tonic-gate 	vmem_free(vmp, slab, slabsize);
12487c478bd9Sstevel@tonic-gate 
12497c478bd9Sstevel@tonic-gate vmem_alloc_failure:
12507c478bd9Sstevel@tonic-gate 
12517c478bd9Sstevel@tonic-gate 	umem_log_event(umem_failure_log, cp, NULL, NULL);
12527c478bd9Sstevel@tonic-gate 	atomic_add_64(&cp->cache_alloc_fail, 1);
12537c478bd9Sstevel@tonic-gate 
12547c478bd9Sstevel@tonic-gate 	return (NULL);
12557c478bd9Sstevel@tonic-gate }
12567c478bd9Sstevel@tonic-gate 
12577c478bd9Sstevel@tonic-gate /*
12587c478bd9Sstevel@tonic-gate  * Destroy a slab.
12597c478bd9Sstevel@tonic-gate  */
12607c478bd9Sstevel@tonic-gate static void
12617c478bd9Sstevel@tonic-gate umem_slab_destroy(umem_cache_t *cp, umem_slab_t *sp)
12627c478bd9Sstevel@tonic-gate {
12637c478bd9Sstevel@tonic-gate 	vmem_t *vmp = cp->cache_arena;
12647c478bd9Sstevel@tonic-gate 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
12657c478bd9Sstevel@tonic-gate 
12667c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_HASH) {
12677c478bd9Sstevel@tonic-gate 		umem_bufctl_t *bcp;
12687c478bd9Sstevel@tonic-gate 		while ((bcp = sp->slab_head) != NULL) {
12697c478bd9Sstevel@tonic-gate 			sp->slab_head = bcp->bc_next;
12707c478bd9Sstevel@tonic-gate 			_umem_cache_free(cp->cache_bufctl_cache, bcp);
12717c478bd9Sstevel@tonic-gate 		}
12727c478bd9Sstevel@tonic-gate 		_umem_cache_free(umem_slab_cache, sp);
12737c478bd9Sstevel@tonic-gate 	}
12747c478bd9Sstevel@tonic-gate 	vmem_free(vmp, slab, cp->cache_slabsize);
12757c478bd9Sstevel@tonic-gate }
12767c478bd9Sstevel@tonic-gate 
12777c478bd9Sstevel@tonic-gate /*
12787c478bd9Sstevel@tonic-gate  * Allocate a raw (unconstructed) buffer from cp's slab layer.
12797c478bd9Sstevel@tonic-gate  */
12807c478bd9Sstevel@tonic-gate static void *
12817c478bd9Sstevel@tonic-gate umem_slab_alloc(umem_cache_t *cp, int umflag)
12827c478bd9Sstevel@tonic-gate {
12837c478bd9Sstevel@tonic-gate 	umem_bufctl_t *bcp, **hash_bucket;
12847c478bd9Sstevel@tonic-gate 	umem_slab_t *sp;
12857c478bd9Sstevel@tonic-gate 	void *buf;
12867c478bd9Sstevel@tonic-gate 
12877c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&cp->cache_lock);
12887c478bd9Sstevel@tonic-gate 	cp->cache_slab_alloc++;
12897c478bd9Sstevel@tonic-gate 	sp = cp->cache_freelist;
12907c478bd9Sstevel@tonic-gate 	ASSERT(sp->slab_cache == cp);
12917c478bd9Sstevel@tonic-gate 	if (sp->slab_head == NULL) {
12927c478bd9Sstevel@tonic-gate 		/*
12937c478bd9Sstevel@tonic-gate 		 * The freelist is empty.  Create a new slab.
12947c478bd9Sstevel@tonic-gate 		 */
12957c478bd9Sstevel@tonic-gate 		(void) mutex_unlock(&cp->cache_lock);
12967c478bd9Sstevel@tonic-gate 		if (cp == &umem_null_cache)
12977c478bd9Sstevel@tonic-gate 			return (NULL);
12987c478bd9Sstevel@tonic-gate 		if ((sp = umem_slab_create(cp, umflag)) == NULL)
12997c478bd9Sstevel@tonic-gate 			return (NULL);
13007c478bd9Sstevel@tonic-gate 		(void) mutex_lock(&cp->cache_lock);
13017c478bd9Sstevel@tonic-gate 		cp->cache_slab_create++;
13027c478bd9Sstevel@tonic-gate 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
13037c478bd9Sstevel@tonic-gate 			cp->cache_bufmax = cp->cache_buftotal;
13047c478bd9Sstevel@tonic-gate 		sp->slab_next = cp->cache_freelist;
13057c478bd9Sstevel@tonic-gate 		sp->slab_prev = cp->cache_freelist->slab_prev;
13067c478bd9Sstevel@tonic-gate 		sp->slab_next->slab_prev = sp;
13077c478bd9Sstevel@tonic-gate 		sp->slab_prev->slab_next = sp;
13087c478bd9Sstevel@tonic-gate 		cp->cache_freelist = sp;
13097c478bd9Sstevel@tonic-gate 	}
13107c478bd9Sstevel@tonic-gate 
13117c478bd9Sstevel@tonic-gate 	sp->slab_refcnt++;
13127c478bd9Sstevel@tonic-gate 	ASSERT(sp->slab_refcnt <= sp->slab_chunks);
13137c478bd9Sstevel@tonic-gate 
13147c478bd9Sstevel@tonic-gate 	/*
13157c478bd9Sstevel@tonic-gate 	 * If we're taking the last buffer in the slab,
13167c478bd9Sstevel@tonic-gate 	 * remove the slab from the cache's freelist.
13177c478bd9Sstevel@tonic-gate 	 */
13187c478bd9Sstevel@tonic-gate 	bcp = sp->slab_head;
13197c478bd9Sstevel@tonic-gate 	if ((sp->slab_head = bcp->bc_next) == NULL) {
13207c478bd9Sstevel@tonic-gate 		cp->cache_freelist = sp->slab_next;
13217c478bd9Sstevel@tonic-gate 		ASSERT(sp->slab_refcnt == sp->slab_chunks);
13227c478bd9Sstevel@tonic-gate 	}
13237c478bd9Sstevel@tonic-gate 
13247c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_HASH) {
13257c478bd9Sstevel@tonic-gate 		/*
13267c478bd9Sstevel@tonic-gate 		 * Add buffer to allocated-address hash table.
13277c478bd9Sstevel@tonic-gate 		 */
13287c478bd9Sstevel@tonic-gate 		buf = bcp->bc_addr;
13297c478bd9Sstevel@tonic-gate 		hash_bucket = UMEM_HASH(cp, buf);
13307c478bd9Sstevel@tonic-gate 		bcp->bc_next = *hash_bucket;
13317c478bd9Sstevel@tonic-gate 		*hash_bucket = bcp;
13327c478bd9Sstevel@tonic-gate 		if ((cp->cache_flags & (UMF_AUDIT | UMF_BUFTAG)) == UMF_AUDIT) {
13337c478bd9Sstevel@tonic-gate 			UMEM_AUDIT(umem_transaction_log, cp, bcp);
13347c478bd9Sstevel@tonic-gate 		}
13357c478bd9Sstevel@tonic-gate 	} else {
13367c478bd9Sstevel@tonic-gate 		buf = UMEM_BUF(cp, bcp);
13377c478bd9Sstevel@tonic-gate 	}
13387c478bd9Sstevel@tonic-gate 
13397c478bd9Sstevel@tonic-gate 	ASSERT(UMEM_SLAB_MEMBER(sp, buf));
13407c478bd9Sstevel@tonic-gate 
13417c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&cp->cache_lock);
13427c478bd9Sstevel@tonic-gate 
13437c478bd9Sstevel@tonic-gate 	return (buf);
13447c478bd9Sstevel@tonic-gate }
13457c478bd9Sstevel@tonic-gate 
13467c478bd9Sstevel@tonic-gate /*
13477c478bd9Sstevel@tonic-gate  * Free a raw (unconstructed) buffer to cp's slab layer.
13487c478bd9Sstevel@tonic-gate  */
13497c478bd9Sstevel@tonic-gate static void
13507c478bd9Sstevel@tonic-gate umem_slab_free(umem_cache_t *cp, void *buf)
13517c478bd9Sstevel@tonic-gate {
13527c478bd9Sstevel@tonic-gate 	umem_slab_t *sp;
13537c478bd9Sstevel@tonic-gate 	umem_bufctl_t *bcp, **prev_bcpp;
13547c478bd9Sstevel@tonic-gate 
13557c478bd9Sstevel@tonic-gate 	ASSERT(buf != NULL);
13567c478bd9Sstevel@tonic-gate 
13577c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&cp->cache_lock);
13587c478bd9Sstevel@tonic-gate 	cp->cache_slab_free++;
13597c478bd9Sstevel@tonic-gate 
13607c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_HASH) {
13617c478bd9Sstevel@tonic-gate 		/*
13627c478bd9Sstevel@tonic-gate 		 * Look up buffer in allocated-address hash table.
13637c478bd9Sstevel@tonic-gate 		 */
13647c478bd9Sstevel@tonic-gate 		prev_bcpp = UMEM_HASH(cp, buf);
13657c478bd9Sstevel@tonic-gate 		while ((bcp = *prev_bcpp) != NULL) {
13667c478bd9Sstevel@tonic-gate 			if (bcp->bc_addr == buf) {
13677c478bd9Sstevel@tonic-gate 				*prev_bcpp = bcp->bc_next;
13687c478bd9Sstevel@tonic-gate 				sp = bcp->bc_slab;
13697c478bd9Sstevel@tonic-gate 				break;
13707c478bd9Sstevel@tonic-gate 			}
13717c478bd9Sstevel@tonic-gate 			cp->cache_lookup_depth++;
13727c478bd9Sstevel@tonic-gate 			prev_bcpp = &bcp->bc_next;
13737c478bd9Sstevel@tonic-gate 		}
13747c478bd9Sstevel@tonic-gate 	} else {
13757c478bd9Sstevel@tonic-gate 		bcp = UMEM_BUFCTL(cp, buf);
13767c478bd9Sstevel@tonic-gate 		sp = UMEM_SLAB(cp, buf);
13777c478bd9Sstevel@tonic-gate 	}
13787c478bd9Sstevel@tonic-gate 
13797c478bd9Sstevel@tonic-gate 	if (bcp == NULL || sp->slab_cache != cp || !UMEM_SLAB_MEMBER(sp, buf)) {
13807c478bd9Sstevel@tonic-gate 		(void) mutex_unlock(&cp->cache_lock);
13817c478bd9Sstevel@tonic-gate 		umem_error(UMERR_BADADDR, cp, buf);
13827c478bd9Sstevel@tonic-gate 		return;
13837c478bd9Sstevel@tonic-gate 	}
13847c478bd9Sstevel@tonic-gate 
13857c478bd9Sstevel@tonic-gate 	if ((cp->cache_flags & (UMF_AUDIT | UMF_BUFTAG)) == UMF_AUDIT) {
13867c478bd9Sstevel@tonic-gate 		if (cp->cache_flags & UMF_CONTENTS)
13877c478bd9Sstevel@tonic-gate 			((umem_bufctl_audit_t *)bcp)->bc_contents =
13887c478bd9Sstevel@tonic-gate 			    umem_log_enter(umem_content_log, buf,
13897c478bd9Sstevel@tonic-gate 			    cp->cache_contents);
13907c478bd9Sstevel@tonic-gate 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
13917c478bd9Sstevel@tonic-gate 	}
13927c478bd9Sstevel@tonic-gate 
13937c478bd9Sstevel@tonic-gate 	/*
13947c478bd9Sstevel@tonic-gate 	 * If this slab isn't currently on the freelist, put it there.
13957c478bd9Sstevel@tonic-gate 	 */
13967c478bd9Sstevel@tonic-gate 	if (sp->slab_head == NULL) {
13977c478bd9Sstevel@tonic-gate 		ASSERT(sp->slab_refcnt == sp->slab_chunks);
13987c478bd9Sstevel@tonic-gate 		ASSERT(cp->cache_freelist != sp);
13997c478bd9Sstevel@tonic-gate 		sp->slab_next->slab_prev = sp->slab_prev;
14007c478bd9Sstevel@tonic-gate 		sp->slab_prev->slab_next = sp->slab_next;
14017c478bd9Sstevel@tonic-gate 		sp->slab_next = cp->cache_freelist;
14027c478bd9Sstevel@tonic-gate 		sp->slab_prev = cp->cache_freelist->slab_prev;
14037c478bd9Sstevel@tonic-gate 		sp->slab_next->slab_prev = sp;
14047c478bd9Sstevel@tonic-gate 		sp->slab_prev->slab_next = sp;
14057c478bd9Sstevel@tonic-gate 		cp->cache_freelist = sp;
14067c478bd9Sstevel@tonic-gate 	}
14077c478bd9Sstevel@tonic-gate 
14087c478bd9Sstevel@tonic-gate 	bcp->bc_next = sp->slab_head;
14097c478bd9Sstevel@tonic-gate 	sp->slab_head = bcp;
14107c478bd9Sstevel@tonic-gate 
14117c478bd9Sstevel@tonic-gate 	ASSERT(sp->slab_refcnt >= 1);
14127c478bd9Sstevel@tonic-gate 	if (--sp->slab_refcnt == 0) {
14137c478bd9Sstevel@tonic-gate 		/*
14147c478bd9Sstevel@tonic-gate 		 * There are no outstanding allocations from this slab,
14157c478bd9Sstevel@tonic-gate 		 * so we can reclaim the memory.
14167c478bd9Sstevel@tonic-gate 		 */
14177c478bd9Sstevel@tonic-gate 		sp->slab_next->slab_prev = sp->slab_prev;
14187c478bd9Sstevel@tonic-gate 		sp->slab_prev->slab_next = sp->slab_next;
14197c478bd9Sstevel@tonic-gate 		if (sp == cp->cache_freelist)
14207c478bd9Sstevel@tonic-gate 			cp->cache_freelist = sp->slab_next;
14217c478bd9Sstevel@tonic-gate 		cp->cache_slab_destroy++;
14227c478bd9Sstevel@tonic-gate 		cp->cache_buftotal -= sp->slab_chunks;
14237c478bd9Sstevel@tonic-gate 		(void) mutex_unlock(&cp->cache_lock);
14247c478bd9Sstevel@tonic-gate 		umem_slab_destroy(cp, sp);
14257c478bd9Sstevel@tonic-gate 		return;
14267c478bd9Sstevel@tonic-gate 	}
14277c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&cp->cache_lock);
14287c478bd9Sstevel@tonic-gate }
14297c478bd9Sstevel@tonic-gate 
14307c478bd9Sstevel@tonic-gate static int
14317c478bd9Sstevel@tonic-gate umem_cache_alloc_debug(umem_cache_t *cp, void *buf, int umflag)
14327c478bd9Sstevel@tonic-gate {
14337c478bd9Sstevel@tonic-gate 	umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
14347c478bd9Sstevel@tonic-gate 	umem_bufctl_audit_t *bcp = (umem_bufctl_audit_t *)btp->bt_bufctl;
14357c478bd9Sstevel@tonic-gate 	uint32_t mtbf;
14367c478bd9Sstevel@tonic-gate 	int flags_nfatal;
14377c478bd9Sstevel@tonic-gate 
14387c478bd9Sstevel@tonic-gate 	if (btp->bt_bxstat != ((intptr_t)bcp ^ UMEM_BUFTAG_FREE)) {
14397c478bd9Sstevel@tonic-gate 		umem_error(UMERR_BADBUFTAG, cp, buf);
14407c478bd9Sstevel@tonic-gate 		return (-1);
14417c478bd9Sstevel@tonic-gate 	}
14427c478bd9Sstevel@tonic-gate 
14437c478bd9Sstevel@tonic-gate 	btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_ALLOC;
14447c478bd9Sstevel@tonic-gate 
14457c478bd9Sstevel@tonic-gate 	if ((cp->cache_flags & UMF_HASH) && bcp->bc_addr != buf) {
14467c478bd9Sstevel@tonic-gate 		umem_error(UMERR_BADBUFCTL, cp, buf);
14477c478bd9Sstevel@tonic-gate 		return (-1);
14487c478bd9Sstevel@tonic-gate 	}
14497c478bd9Sstevel@tonic-gate 
14507c478bd9Sstevel@tonic-gate 	btp->bt_redzone = UMEM_REDZONE_PATTERN;
14517c478bd9Sstevel@tonic-gate 
14527c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_DEADBEEF) {
14537c478bd9Sstevel@tonic-gate 		if (verify_and_copy_pattern(UMEM_FREE_PATTERN,
14547c478bd9Sstevel@tonic-gate 		    UMEM_UNINITIALIZED_PATTERN, buf, cp->cache_verify)) {
14557c478bd9Sstevel@tonic-gate 			umem_error(UMERR_MODIFIED, cp, buf);
14567c478bd9Sstevel@tonic-gate 			return (-1);
14577c478bd9Sstevel@tonic-gate 		}
14587c478bd9Sstevel@tonic-gate 	}
14597c478bd9Sstevel@tonic-gate 
14607c478bd9Sstevel@tonic-gate 	if ((mtbf = umem_mtbf | cp->cache_mtbf) != 0 &&
14617c478bd9Sstevel@tonic-gate 	    gethrtime() % mtbf == 0 &&
14627c478bd9Sstevel@tonic-gate 	    (umflag & (UMEM_FATAL_FLAGS)) == 0) {
14637c478bd9Sstevel@tonic-gate 		umem_log_event(umem_failure_log, cp, NULL, NULL);
14647c478bd9Sstevel@tonic-gate 	} else {
14657c478bd9Sstevel@tonic-gate 		mtbf = 0;
14667c478bd9Sstevel@tonic-gate 	}
14677c478bd9Sstevel@tonic-gate 
14687c478bd9Sstevel@tonic-gate 	/*
14697c478bd9Sstevel@tonic-gate 	 * We do not pass fatal flags on to the constructor.  This prevents
14707c478bd9Sstevel@tonic-gate 	 * leaking buffers in the event of a subordinate constructor failing.
14717c478bd9Sstevel@tonic-gate 	 */
14727c478bd9Sstevel@tonic-gate 	flags_nfatal = UMEM_DEFAULT;
14737c478bd9Sstevel@tonic-gate 	if (mtbf || (cp->cache_constructor != NULL &&
14747c478bd9Sstevel@tonic-gate 	    cp->cache_constructor(buf, cp->cache_private, flags_nfatal) != 0)) {
14757c478bd9Sstevel@tonic-gate 		atomic_add_64(&cp->cache_alloc_fail, 1);
14767c478bd9Sstevel@tonic-gate 		btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
14777c478bd9Sstevel@tonic-gate 		copy_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
14787c478bd9Sstevel@tonic-gate 		umem_slab_free(cp, buf);
14797c478bd9Sstevel@tonic-gate 		return (-1);
14807c478bd9Sstevel@tonic-gate 	}
14817c478bd9Sstevel@tonic-gate 
14827c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_AUDIT) {
14837c478bd9Sstevel@tonic-gate 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
14847c478bd9Sstevel@tonic-gate 	}
14857c478bd9Sstevel@tonic-gate 
14867c478bd9Sstevel@tonic-gate 	return (0);
14877c478bd9Sstevel@tonic-gate }
14887c478bd9Sstevel@tonic-gate 
14897c478bd9Sstevel@tonic-gate static int
14907c478bd9Sstevel@tonic-gate umem_cache_free_debug(umem_cache_t *cp, void *buf)
14917c478bd9Sstevel@tonic-gate {
14927c478bd9Sstevel@tonic-gate 	umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
14937c478bd9Sstevel@tonic-gate 	umem_bufctl_audit_t *bcp = (umem_bufctl_audit_t *)btp->bt_bufctl;
14947c478bd9Sstevel@tonic-gate 	umem_slab_t *sp;
14957c478bd9Sstevel@tonic-gate 
14967c478bd9Sstevel@tonic-gate 	if (btp->bt_bxstat != ((intptr_t)bcp ^ UMEM_BUFTAG_ALLOC)) {
14977c478bd9Sstevel@tonic-gate 		if (btp->bt_bxstat == ((intptr_t)bcp ^ UMEM_BUFTAG_FREE)) {
14987c478bd9Sstevel@tonic-gate 			umem_error(UMERR_DUPFREE, cp, buf);
14997c478bd9Sstevel@tonic-gate 			return (-1);
15007c478bd9Sstevel@tonic-gate 		}
15017c478bd9Sstevel@tonic-gate 		sp = umem_findslab(cp, buf);
15027c478bd9Sstevel@tonic-gate 		if (sp == NULL || sp->slab_cache != cp)
15037c478bd9Sstevel@tonic-gate 			umem_error(UMERR_BADADDR, cp, buf);
15047c478bd9Sstevel@tonic-gate 		else
15057c478bd9Sstevel@tonic-gate 			umem_error(UMERR_REDZONE, cp, buf);
15067c478bd9Sstevel@tonic-gate 		return (-1);
15077c478bd9Sstevel@tonic-gate 	}
15087c478bd9Sstevel@tonic-gate 
15097c478bd9Sstevel@tonic-gate 	btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
15107c478bd9Sstevel@tonic-gate 
15117c478bd9Sstevel@tonic-gate 	if ((cp->cache_flags & UMF_HASH) && bcp->bc_addr != buf) {
15127c478bd9Sstevel@tonic-gate 		umem_error(UMERR_BADBUFCTL, cp, buf);
15137c478bd9Sstevel@tonic-gate 		return (-1);
15147c478bd9Sstevel@tonic-gate 	}
15157c478bd9Sstevel@tonic-gate 
15167c478bd9Sstevel@tonic-gate 	if (btp->bt_redzone != UMEM_REDZONE_PATTERN) {
15177c478bd9Sstevel@tonic-gate 		umem_error(UMERR_REDZONE, cp, buf);
15187c478bd9Sstevel@tonic-gate 		return (-1);
15197c478bd9Sstevel@tonic-gate 	}
15207c478bd9Sstevel@tonic-gate 
15217c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_AUDIT) {
15227c478bd9Sstevel@tonic-gate 		if (cp->cache_flags & UMF_CONTENTS)
15237c478bd9Sstevel@tonic-gate 			bcp->bc_contents = umem_log_enter(umem_content_log,
15247c478bd9Sstevel@tonic-gate 			    buf, cp->cache_contents);
15257c478bd9Sstevel@tonic-gate 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
15267c478bd9Sstevel@tonic-gate 	}
15277c478bd9Sstevel@tonic-gate 
15287c478bd9Sstevel@tonic-gate 	if (cp->cache_destructor != NULL)
15297c478bd9Sstevel@tonic-gate 		cp->cache_destructor(buf, cp->cache_private);
15307c478bd9Sstevel@tonic-gate 
15317c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_DEADBEEF)
15327c478bd9Sstevel@tonic-gate 		copy_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
15337c478bd9Sstevel@tonic-gate 
15347c478bd9Sstevel@tonic-gate 	return (0);
15357c478bd9Sstevel@tonic-gate }
15367c478bd9Sstevel@tonic-gate 
15377c478bd9Sstevel@tonic-gate /*
15387c478bd9Sstevel@tonic-gate  * Free each object in magazine mp to cp's slab layer, and free mp itself.
15397c478bd9Sstevel@tonic-gate  */
15407c478bd9Sstevel@tonic-gate static void
15417c478bd9Sstevel@tonic-gate umem_magazine_destroy(umem_cache_t *cp, umem_magazine_t *mp, int nrounds)
15427c478bd9Sstevel@tonic-gate {
15437c478bd9Sstevel@tonic-gate 	int round;
15447c478bd9Sstevel@tonic-gate 
15457c478bd9Sstevel@tonic-gate 	ASSERT(cp->cache_next == NULL || IN_UPDATE());
15467c478bd9Sstevel@tonic-gate 
15477c478bd9Sstevel@tonic-gate 	for (round = 0; round < nrounds; round++) {
15487c478bd9Sstevel@tonic-gate 		void *buf = mp->mag_round[round];
15497c478bd9Sstevel@tonic-gate 
15507c478bd9Sstevel@tonic-gate 		if ((cp->cache_flags & UMF_DEADBEEF) &&
15517c478bd9Sstevel@tonic-gate 		    verify_pattern(UMEM_FREE_PATTERN, buf,
15527c478bd9Sstevel@tonic-gate 		    cp->cache_verify) != NULL) {
15537c478bd9Sstevel@tonic-gate 			umem_error(UMERR_MODIFIED, cp, buf);
15547c478bd9Sstevel@tonic-gate 			continue;
15557c478bd9Sstevel@tonic-gate 		}
15567c478bd9Sstevel@tonic-gate 
15577c478bd9Sstevel@tonic-gate 		if (!(cp->cache_flags & UMF_BUFTAG) &&
15587c478bd9Sstevel@tonic-gate 		    cp->cache_destructor != NULL)
15597c478bd9Sstevel@tonic-gate 			cp->cache_destructor(buf, cp->cache_private);
15607c478bd9Sstevel@tonic-gate 
15617c478bd9Sstevel@tonic-gate 		umem_slab_free(cp, buf);
15627c478bd9Sstevel@tonic-gate 	}
15637c478bd9Sstevel@tonic-gate 	ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
15647c478bd9Sstevel@tonic-gate 	_umem_cache_free(cp->cache_magtype->mt_cache, mp);
15657c478bd9Sstevel@tonic-gate }
15667c478bd9Sstevel@tonic-gate 
15677c478bd9Sstevel@tonic-gate /*
15687c478bd9Sstevel@tonic-gate  * Allocate a magazine from the depot.
15697c478bd9Sstevel@tonic-gate  */
15707c478bd9Sstevel@tonic-gate static umem_magazine_t *
15717c478bd9Sstevel@tonic-gate umem_depot_alloc(umem_cache_t *cp, umem_maglist_t *mlp)
15727c478bd9Sstevel@tonic-gate {
15737c478bd9Sstevel@tonic-gate 	umem_magazine_t *mp;
15747c478bd9Sstevel@tonic-gate 
15757c478bd9Sstevel@tonic-gate 	/*
15767c478bd9Sstevel@tonic-gate 	 * If we can't get the depot lock without contention,
15777c478bd9Sstevel@tonic-gate 	 * update our contention count.  We use the depot
15787c478bd9Sstevel@tonic-gate 	 * contention rate to determine whether we need to
15797c478bd9Sstevel@tonic-gate 	 * increase the magazine size for better scalability.
15807c478bd9Sstevel@tonic-gate 	 */
15817c478bd9Sstevel@tonic-gate 	if (mutex_trylock(&cp->cache_depot_lock) != 0) {
15827c478bd9Sstevel@tonic-gate 		(void) mutex_lock(&cp->cache_depot_lock);
15837c478bd9Sstevel@tonic-gate 		cp->cache_depot_contention++;
15847c478bd9Sstevel@tonic-gate 	}
15857c478bd9Sstevel@tonic-gate 
15867c478bd9Sstevel@tonic-gate 	if ((mp = mlp->ml_list) != NULL) {
15877c478bd9Sstevel@tonic-gate 		ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
15887c478bd9Sstevel@tonic-gate 		mlp->ml_list = mp->mag_next;
15897c478bd9Sstevel@tonic-gate 		if (--mlp->ml_total < mlp->ml_min)
15907c478bd9Sstevel@tonic-gate 			mlp->ml_min = mlp->ml_total;
15917c478bd9Sstevel@tonic-gate 		mlp->ml_alloc++;
15927c478bd9Sstevel@tonic-gate 	}
15937c478bd9Sstevel@tonic-gate 
15947c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&cp->cache_depot_lock);
15957c478bd9Sstevel@tonic-gate 
15967c478bd9Sstevel@tonic-gate 	return (mp);
15977c478bd9Sstevel@tonic-gate }
15987c478bd9Sstevel@tonic-gate 
15997c478bd9Sstevel@tonic-gate /*
16007c478bd9Sstevel@tonic-gate  * Free a magazine to the depot.
16017c478bd9Sstevel@tonic-gate  */
16027c478bd9Sstevel@tonic-gate static void
16037c478bd9Sstevel@tonic-gate umem_depot_free(umem_cache_t *cp, umem_maglist_t *mlp, umem_magazine_t *mp)
16047c478bd9Sstevel@tonic-gate {
16057c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&cp->cache_depot_lock);
16067c478bd9Sstevel@tonic-gate 	ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
16077c478bd9Sstevel@tonic-gate 	mp->mag_next = mlp->ml_list;
16087c478bd9Sstevel@tonic-gate 	mlp->ml_list = mp;
16097c478bd9Sstevel@tonic-gate 	mlp->ml_total++;
16107c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&cp->cache_depot_lock);
16117c478bd9Sstevel@tonic-gate }
16127c478bd9Sstevel@tonic-gate 
16137c478bd9Sstevel@tonic-gate /*
16147c478bd9Sstevel@tonic-gate  * Update the working set statistics for cp's depot.
16157c478bd9Sstevel@tonic-gate  */
16167c478bd9Sstevel@tonic-gate static void
16177c478bd9Sstevel@tonic-gate umem_depot_ws_update(umem_cache_t *cp)
16187c478bd9Sstevel@tonic-gate {
16197c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&cp->cache_depot_lock);
16207c478bd9Sstevel@tonic-gate 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
16217c478bd9Sstevel@tonic-gate 	cp->cache_full.ml_min = cp->cache_full.ml_total;
16227c478bd9Sstevel@tonic-gate 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
16237c478bd9Sstevel@tonic-gate 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
16247c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&cp->cache_depot_lock);
16257c478bd9Sstevel@tonic-gate }
16267c478bd9Sstevel@tonic-gate 
16277c478bd9Sstevel@tonic-gate /*
16287c478bd9Sstevel@tonic-gate  * Reap all magazines that have fallen out of the depot's working set.
16297c478bd9Sstevel@tonic-gate  */
16307c478bd9Sstevel@tonic-gate static void
16317c478bd9Sstevel@tonic-gate umem_depot_ws_reap(umem_cache_t *cp)
16327c478bd9Sstevel@tonic-gate {
16337c478bd9Sstevel@tonic-gate 	long reap;
16347c478bd9Sstevel@tonic-gate 	umem_magazine_t *mp;
16357c478bd9Sstevel@tonic-gate 
16367c478bd9Sstevel@tonic-gate 	ASSERT(cp->cache_next == NULL || IN_REAP());
16377c478bd9Sstevel@tonic-gate 
16387c478bd9Sstevel@tonic-gate 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
16397c478bd9Sstevel@tonic-gate 	while (reap-- && (mp = umem_depot_alloc(cp, &cp->cache_full)) != NULL)
16407c478bd9Sstevel@tonic-gate 		umem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
16417c478bd9Sstevel@tonic-gate 
16427c478bd9Sstevel@tonic-gate 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
16437c478bd9Sstevel@tonic-gate 	while (reap-- && (mp = umem_depot_alloc(cp, &cp->cache_empty)) != NULL)
16447c478bd9Sstevel@tonic-gate 		umem_magazine_destroy(cp, mp, 0);
16457c478bd9Sstevel@tonic-gate }
16467c478bd9Sstevel@tonic-gate 
16477c478bd9Sstevel@tonic-gate static void
16487c478bd9Sstevel@tonic-gate umem_cpu_reload(umem_cpu_cache_t *ccp, umem_magazine_t *mp, int rounds)
16497c478bd9Sstevel@tonic-gate {
16507c478bd9Sstevel@tonic-gate 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
16517c478bd9Sstevel@tonic-gate 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
16527c478bd9Sstevel@tonic-gate 	ASSERT(ccp->cc_magsize > 0);
16537c478bd9Sstevel@tonic-gate 
16547c478bd9Sstevel@tonic-gate 	ccp->cc_ploaded = ccp->cc_loaded;
16557c478bd9Sstevel@tonic-gate 	ccp->cc_prounds = ccp->cc_rounds;
16567c478bd9Sstevel@tonic-gate 	ccp->cc_loaded = mp;
16577c478bd9Sstevel@tonic-gate 	ccp->cc_rounds = rounds;
16587c478bd9Sstevel@tonic-gate }
16597c478bd9Sstevel@tonic-gate 
16607c478bd9Sstevel@tonic-gate /*
16617c478bd9Sstevel@tonic-gate  * Allocate a constructed object from cache cp.
16627c478bd9Sstevel@tonic-gate  */
16637c478bd9Sstevel@tonic-gate #pragma weak umem_cache_alloc = _umem_cache_alloc
16647c478bd9Sstevel@tonic-gate void *
16657c478bd9Sstevel@tonic-gate _umem_cache_alloc(umem_cache_t *cp, int umflag)
16667c478bd9Sstevel@tonic-gate {
16677c478bd9Sstevel@tonic-gate 	umem_cpu_cache_t *ccp;
16687c478bd9Sstevel@tonic-gate 	umem_magazine_t *fmp;
16697c478bd9Sstevel@tonic-gate 	void *buf;
16707c478bd9Sstevel@tonic-gate 	int flags_nfatal;
16717c478bd9Sstevel@tonic-gate 
16727c478bd9Sstevel@tonic-gate retry:
16737c478bd9Sstevel@tonic-gate 	ccp = UMEM_CPU_CACHE(cp, CPU(cp->cache_cpu_mask));
16747c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&ccp->cc_lock);
16757c478bd9Sstevel@tonic-gate 	for (;;) {
16767c478bd9Sstevel@tonic-gate 		/*
16777c478bd9Sstevel@tonic-gate 		 * If there's an object available in the current CPU's
16787c478bd9Sstevel@tonic-gate 		 * loaded magazine, just take it and return.
16797c478bd9Sstevel@tonic-gate 		 */
16807c478bd9Sstevel@tonic-gate 		if (ccp->cc_rounds > 0) {
16817c478bd9Sstevel@tonic-gate 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
16827c478bd9Sstevel@tonic-gate 			ccp->cc_alloc++;
16837c478bd9Sstevel@tonic-gate 			(void) mutex_unlock(&ccp->cc_lock);
16847c478bd9Sstevel@tonic-gate 			if ((ccp->cc_flags & UMF_BUFTAG) &&
16857c478bd9Sstevel@tonic-gate 			    umem_cache_alloc_debug(cp, buf, umflag) == -1) {
16867c478bd9Sstevel@tonic-gate 				if (umem_alloc_retry(cp, umflag)) {
16877c478bd9Sstevel@tonic-gate 					goto retry;
16887c478bd9Sstevel@tonic-gate 				}
16897c478bd9Sstevel@tonic-gate 
16907c478bd9Sstevel@tonic-gate 				return (NULL);
16917c478bd9Sstevel@tonic-gate 			}
16927c478bd9Sstevel@tonic-gate 			return (buf);
16937c478bd9Sstevel@tonic-gate 		}
16947c478bd9Sstevel@tonic-gate 
16957c478bd9Sstevel@tonic-gate 		/*
16967c478bd9Sstevel@tonic-gate 		 * The loaded magazine is empty.  If the previously loaded
16977c478bd9Sstevel@tonic-gate 		 * magazine was full, exchange them and try again.
16987c478bd9Sstevel@tonic-gate 		 */
16997c478bd9Sstevel@tonic-gate 		if (ccp->cc_prounds > 0) {
17007c478bd9Sstevel@tonic-gate 			umem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
17017c478bd9Sstevel@tonic-gate 			continue;
17027c478bd9Sstevel@tonic-gate 		}
17037c478bd9Sstevel@tonic-gate 
17047c478bd9Sstevel@tonic-gate 		/*
17057c478bd9Sstevel@tonic-gate 		 * If the magazine layer is disabled, break out now.
17067c478bd9Sstevel@tonic-gate 		 */
17077c478bd9Sstevel@tonic-gate 		if (ccp->cc_magsize == 0)
17087c478bd9Sstevel@tonic-gate 			break;
17097c478bd9Sstevel@tonic-gate 
17107c478bd9Sstevel@tonic-gate 		/*
17117c478bd9Sstevel@tonic-gate 		 * Try to get a full magazine from the depot.
17127c478bd9Sstevel@tonic-gate 		 */
17137c478bd9Sstevel@tonic-gate 		fmp = umem_depot_alloc(cp, &cp->cache_full);
17147c478bd9Sstevel@tonic-gate 		if (fmp != NULL) {
17157c478bd9Sstevel@tonic-gate 			if (ccp->cc_ploaded != NULL)
17167c478bd9Sstevel@tonic-gate 				umem_depot_free(cp, &cp->cache_empty,
17177c478bd9Sstevel@tonic-gate 				    ccp->cc_ploaded);
17187c478bd9Sstevel@tonic-gate 			umem_cpu_reload(ccp, fmp, ccp->cc_magsize);
17197c478bd9Sstevel@tonic-gate 			continue;
17207c478bd9Sstevel@tonic-gate 		}
17217c478bd9Sstevel@tonic-gate 
17227c478bd9Sstevel@tonic-gate 		/*
17237c478bd9Sstevel@tonic-gate 		 * There are no full magazines in the depot,
17247c478bd9Sstevel@tonic-gate 		 * so fall through to the slab layer.
17257c478bd9Sstevel@tonic-gate 		 */
17267c478bd9Sstevel@tonic-gate 		break;
17277c478bd9Sstevel@tonic-gate 	}
17287c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&ccp->cc_lock);
17297c478bd9Sstevel@tonic-gate 
17307c478bd9Sstevel@tonic-gate 	/*
17317c478bd9Sstevel@tonic-gate 	 * We couldn't allocate a constructed object from the magazine layer,
17327c478bd9Sstevel@tonic-gate 	 * so get a raw buffer from the slab layer and apply its constructor.
17337c478bd9Sstevel@tonic-gate 	 */
17347c478bd9Sstevel@tonic-gate 	buf = umem_slab_alloc(cp, umflag);
17357c478bd9Sstevel@tonic-gate 
17367c478bd9Sstevel@tonic-gate 	if (buf == NULL) {
17377c478bd9Sstevel@tonic-gate 		if (cp == &umem_null_cache)
17387c478bd9Sstevel@tonic-gate 			return (NULL);
17397c478bd9Sstevel@tonic-gate 		if (umem_alloc_retry(cp, umflag)) {
17407c478bd9Sstevel@tonic-gate 			goto retry;
17417c478bd9Sstevel@tonic-gate 		}
17427c478bd9Sstevel@tonic-gate 
17437c478bd9Sstevel@tonic-gate 		return (NULL);
17447c478bd9Sstevel@tonic-gate 	}
17457c478bd9Sstevel@tonic-gate 
17467c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_BUFTAG) {
17477c478bd9Sstevel@tonic-gate 		/*
17487c478bd9Sstevel@tonic-gate 		 * Let umem_cache_alloc_debug() apply the constructor for us.
17497c478bd9Sstevel@tonic-gate 		 */
17507c478bd9Sstevel@tonic-gate 		if (umem_cache_alloc_debug(cp, buf, umflag) == -1) {
17517c478bd9Sstevel@tonic-gate 			if (umem_alloc_retry(cp, umflag)) {
17527c478bd9Sstevel@tonic-gate 				goto retry;
17537c478bd9Sstevel@tonic-gate 			}
17547c478bd9Sstevel@tonic-gate 			return (NULL);
17557c478bd9Sstevel@tonic-gate 		}
17567c478bd9Sstevel@tonic-gate 		return (buf);
17577c478bd9Sstevel@tonic-gate 	}
17587c478bd9Sstevel@tonic-gate 
17597c478bd9Sstevel@tonic-gate 	/*
17607c478bd9Sstevel@tonic-gate 	 * We do not pass fatal flags on to the constructor.  This prevents
17617c478bd9Sstevel@tonic-gate 	 * leaking buffers in the event of a subordinate constructor failing.
17627c478bd9Sstevel@tonic-gate 	 */
17637c478bd9Sstevel@tonic-gate 	flags_nfatal = UMEM_DEFAULT;
17647c478bd9Sstevel@tonic-gate 	if (cp->cache_constructor != NULL &&
17657c478bd9Sstevel@tonic-gate 	    cp->cache_constructor(buf, cp->cache_private, flags_nfatal) != 0) {
17667c478bd9Sstevel@tonic-gate 		atomic_add_64(&cp->cache_alloc_fail, 1);
17677c478bd9Sstevel@tonic-gate 		umem_slab_free(cp, buf);
17687c478bd9Sstevel@tonic-gate 
17697c478bd9Sstevel@tonic-gate 		if (umem_alloc_retry(cp, umflag)) {
17707c478bd9Sstevel@tonic-gate 			goto retry;
17717c478bd9Sstevel@tonic-gate 		}
17727c478bd9Sstevel@tonic-gate 		return (NULL);
17737c478bd9Sstevel@tonic-gate 	}
17747c478bd9Sstevel@tonic-gate 
17757c478bd9Sstevel@tonic-gate 	return (buf);
17767c478bd9Sstevel@tonic-gate }
17777c478bd9Sstevel@tonic-gate 
17787c478bd9Sstevel@tonic-gate /*
17797c478bd9Sstevel@tonic-gate  * Free a constructed object to cache cp.
17807c478bd9Sstevel@tonic-gate  */
17817c478bd9Sstevel@tonic-gate #pragma weak umem_cache_free = _umem_cache_free
17827c478bd9Sstevel@tonic-gate void
17837c478bd9Sstevel@tonic-gate _umem_cache_free(umem_cache_t *cp, void *buf)
17847c478bd9Sstevel@tonic-gate {
17857c478bd9Sstevel@tonic-gate 	umem_cpu_cache_t *ccp = UMEM_CPU_CACHE(cp, CPU(cp->cache_cpu_mask));
17867c478bd9Sstevel@tonic-gate 	umem_magazine_t *emp;
17877c478bd9Sstevel@tonic-gate 	umem_magtype_t *mtp;
17887c478bd9Sstevel@tonic-gate 
17897c478bd9Sstevel@tonic-gate 	if (ccp->cc_flags & UMF_BUFTAG)
17907c478bd9Sstevel@tonic-gate 		if (umem_cache_free_debug(cp, buf) == -1)
17917c478bd9Sstevel@tonic-gate 			return;
17927c478bd9Sstevel@tonic-gate 
17937c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&ccp->cc_lock);
17947c478bd9Sstevel@tonic-gate 	for (;;) {
17957c478bd9Sstevel@tonic-gate 		/*
17967c478bd9Sstevel@tonic-gate 		 * If there's a slot available in the current CPU's
17977c478bd9Sstevel@tonic-gate 		 * loaded magazine, just put the object there and return.
17987c478bd9Sstevel@tonic-gate 		 */
17997c478bd9Sstevel@tonic-gate 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
18007c478bd9Sstevel@tonic-gate 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
18017c478bd9Sstevel@tonic-gate 			ccp->cc_free++;
18027c478bd9Sstevel@tonic-gate 			(void) mutex_unlock(&ccp->cc_lock);
18037c478bd9Sstevel@tonic-gate 			return;
18047c478bd9Sstevel@tonic-gate 		}
18057c478bd9Sstevel@tonic-gate 
18067c478bd9Sstevel@tonic-gate 		/*
18077c478bd9Sstevel@tonic-gate 		 * The loaded magazine is full.  If the previously loaded
18087c478bd9Sstevel@tonic-gate 		 * magazine was empty, exchange them and try again.
18097c478bd9Sstevel@tonic-gate 		 */
18107c478bd9Sstevel@tonic-gate 		if (ccp->cc_prounds == 0) {
18117c478bd9Sstevel@tonic-gate 			umem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
18127c478bd9Sstevel@tonic-gate 			continue;
18137c478bd9Sstevel@tonic-gate 		}
18147c478bd9Sstevel@tonic-gate 
18157c478bd9Sstevel@tonic-gate 		/*
18167c478bd9Sstevel@tonic-gate 		 * If the magazine layer is disabled, break out now.
18177c478bd9Sstevel@tonic-gate 		 */
18187c478bd9Sstevel@tonic-gate 		if (ccp->cc_magsize == 0)
18197c478bd9Sstevel@tonic-gate 			break;
18207c478bd9Sstevel@tonic-gate 
18217c478bd9Sstevel@tonic-gate 		/*
18227c478bd9Sstevel@tonic-gate 		 * Try to get an empty magazine from the depot.
18237c478bd9Sstevel@tonic-gate 		 */
18247c478bd9Sstevel@tonic-gate 		emp = umem_depot_alloc(cp, &cp->cache_empty);
18257c478bd9Sstevel@tonic-gate 		if (emp != NULL) {
18267c478bd9Sstevel@tonic-gate 			if (ccp->cc_ploaded != NULL)
18277c478bd9Sstevel@tonic-gate 				umem_depot_free(cp, &cp->cache_full,
18287c478bd9Sstevel@tonic-gate 				    ccp->cc_ploaded);
18297c478bd9Sstevel@tonic-gate 			umem_cpu_reload(ccp, emp, 0);
18307c478bd9Sstevel@tonic-gate 			continue;
18317c478bd9Sstevel@tonic-gate 		}
18327c478bd9Sstevel@tonic-gate 
18337c478bd9Sstevel@tonic-gate 		/*
18347c478bd9Sstevel@tonic-gate 		 * There are no empty magazines in the depot,
18357c478bd9Sstevel@tonic-gate 		 * so try to allocate a new one.  We must drop all locks
18367c478bd9Sstevel@tonic-gate 		 * across umem_cache_alloc() because lower layers may
18377c478bd9Sstevel@tonic-gate 		 * attempt to allocate from this cache.
18387c478bd9Sstevel@tonic-gate 		 */
18397c478bd9Sstevel@tonic-gate 		mtp = cp->cache_magtype;
18407c478bd9Sstevel@tonic-gate 		(void) mutex_unlock(&ccp->cc_lock);
18417c478bd9Sstevel@tonic-gate 		emp = _umem_cache_alloc(mtp->mt_cache, UMEM_DEFAULT);
18427c478bd9Sstevel@tonic-gate 		(void) mutex_lock(&ccp->cc_lock);
18437c478bd9Sstevel@tonic-gate 
18447c478bd9Sstevel@tonic-gate 		if (emp != NULL) {
18457c478bd9Sstevel@tonic-gate 			/*
18467c478bd9Sstevel@tonic-gate 			 * We successfully allocated an empty magazine.
18477c478bd9Sstevel@tonic-gate 			 * However, we had to drop ccp->cc_lock to do it,
18487c478bd9Sstevel@tonic-gate 			 * so the cache's magazine size may have changed.
18497c478bd9Sstevel@tonic-gate 			 * If so, free the magazine and try again.
18507c478bd9Sstevel@tonic-gate 			 */
18517c478bd9Sstevel@tonic-gate 			if (ccp->cc_magsize != mtp->mt_magsize) {
18527c478bd9Sstevel@tonic-gate 				(void) mutex_unlock(&ccp->cc_lock);
18537c478bd9Sstevel@tonic-gate 				_umem_cache_free(mtp->mt_cache, emp);
18547c478bd9Sstevel@tonic-gate 				(void) mutex_lock(&ccp->cc_lock);
18557c478bd9Sstevel@tonic-gate 				continue;
18567c478bd9Sstevel@tonic-gate 			}
18577c478bd9Sstevel@tonic-gate 
18587c478bd9Sstevel@tonic-gate 			/*
18597c478bd9Sstevel@tonic-gate 			 * We got a magazine of the right size.  Add it to
18607c478bd9Sstevel@tonic-gate 			 * the depot and try the whole dance again.
18617c478bd9Sstevel@tonic-gate 			 */
18627c478bd9Sstevel@tonic-gate 			umem_depot_free(cp, &cp->cache_empty, emp);
18637c478bd9Sstevel@tonic-gate 			continue;
18647c478bd9Sstevel@tonic-gate 		}
18657c478bd9Sstevel@tonic-gate 
18667c478bd9Sstevel@tonic-gate 		/*
18677c478bd9Sstevel@tonic-gate 		 * We couldn't allocate an empty magazine,
18687c478bd9Sstevel@tonic-gate 		 * so fall through to the slab layer.
18697c478bd9Sstevel@tonic-gate 		 */
18707c478bd9Sstevel@tonic-gate 		break;
18717c478bd9Sstevel@tonic-gate 	}
18727c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&ccp->cc_lock);
18737c478bd9Sstevel@tonic-gate 
18747c478bd9Sstevel@tonic-gate 	/*
18757c478bd9Sstevel@tonic-gate 	 * We couldn't free our constructed object to the magazine layer,
18767c478bd9Sstevel@tonic-gate 	 * so apply its destructor and free it to the slab layer.
18777c478bd9Sstevel@tonic-gate 	 * Note that if UMF_BUFTAG is in effect, umem_cache_free_debug()
18787c478bd9Sstevel@tonic-gate 	 * will have already applied the destructor.
18797c478bd9Sstevel@tonic-gate 	 */
18807c478bd9Sstevel@tonic-gate 	if (!(cp->cache_flags & UMF_BUFTAG) && cp->cache_destructor != NULL)
18817c478bd9Sstevel@tonic-gate 		cp->cache_destructor(buf, cp->cache_private);
18827c478bd9Sstevel@tonic-gate 
18837c478bd9Sstevel@tonic-gate 	umem_slab_free(cp, buf);
18847c478bd9Sstevel@tonic-gate }
18857c478bd9Sstevel@tonic-gate 
18867c478bd9Sstevel@tonic-gate #pragma weak umem_zalloc = _umem_zalloc
18877c478bd9Sstevel@tonic-gate void *
18887c478bd9Sstevel@tonic-gate _umem_zalloc(size_t size, int umflag)
18897c478bd9Sstevel@tonic-gate {
18907c478bd9Sstevel@tonic-gate 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
18917c478bd9Sstevel@tonic-gate 	void *buf;
18927c478bd9Sstevel@tonic-gate 
18937c478bd9Sstevel@tonic-gate retry:
18947c478bd9Sstevel@tonic-gate 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
18957c478bd9Sstevel@tonic-gate 		umem_cache_t *cp = umem_alloc_table[index];
18967c478bd9Sstevel@tonic-gate 		buf = _umem_cache_alloc(cp, umflag);
18977c478bd9Sstevel@tonic-gate 		if (buf != NULL) {
18987c478bd9Sstevel@tonic-gate 			if (cp->cache_flags & UMF_BUFTAG) {
18997c478bd9Sstevel@tonic-gate 				umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
19007c478bd9Sstevel@tonic-gate 				((uint8_t *)buf)[size] = UMEM_REDZONE_BYTE;
19017c478bd9Sstevel@tonic-gate 				((uint32_t *)btp)[1] = UMEM_SIZE_ENCODE(size);
19027c478bd9Sstevel@tonic-gate 			}
19037c478bd9Sstevel@tonic-gate 			bzero(buf, size);
19047c478bd9Sstevel@tonic-gate 		} else if (umem_alloc_retry(cp, umflag))
19057c478bd9Sstevel@tonic-gate 			goto retry;
19067c478bd9Sstevel@tonic-gate 	} else {
19077c478bd9Sstevel@tonic-gate 		buf = _umem_alloc(size, umflag);	/* handles failure */
19087c478bd9Sstevel@tonic-gate 		if (buf != NULL)
19097c478bd9Sstevel@tonic-gate 			bzero(buf, size);
19107c478bd9Sstevel@tonic-gate 	}
19117c478bd9Sstevel@tonic-gate 	return (buf);
19127c478bd9Sstevel@tonic-gate }
19137c478bd9Sstevel@tonic-gate 
19147c478bd9Sstevel@tonic-gate #pragma weak umem_alloc = _umem_alloc
19157c478bd9Sstevel@tonic-gate void *
19167c478bd9Sstevel@tonic-gate _umem_alloc(size_t size, int umflag)
19177c478bd9Sstevel@tonic-gate {
19187c478bd9Sstevel@tonic-gate 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
19197c478bd9Sstevel@tonic-gate 	void *buf;
19207c478bd9Sstevel@tonic-gate umem_alloc_retry:
19217c478bd9Sstevel@tonic-gate 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
19227c478bd9Sstevel@tonic-gate 		umem_cache_t *cp = umem_alloc_table[index];
19237c478bd9Sstevel@tonic-gate 		buf = _umem_cache_alloc(cp, umflag);
19247c478bd9Sstevel@tonic-gate 		if ((cp->cache_flags & UMF_BUFTAG) && buf != NULL) {
19257c478bd9Sstevel@tonic-gate 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
19267c478bd9Sstevel@tonic-gate 			((uint8_t *)buf)[size] = UMEM_REDZONE_BYTE;
19277c478bd9Sstevel@tonic-gate 			((uint32_t *)btp)[1] = UMEM_SIZE_ENCODE(size);
19287c478bd9Sstevel@tonic-gate 		}
19297c478bd9Sstevel@tonic-gate 		if (buf == NULL && umem_alloc_retry(cp, umflag))
19307c478bd9Sstevel@tonic-gate 			goto umem_alloc_retry;
19317c478bd9Sstevel@tonic-gate 		return (buf);
19327c478bd9Sstevel@tonic-gate 	}
19337c478bd9Sstevel@tonic-gate 	if (size == 0)
19347c478bd9Sstevel@tonic-gate 		return (NULL);
19357c478bd9Sstevel@tonic-gate 	if (umem_oversize_arena == NULL) {
19367c478bd9Sstevel@tonic-gate 		if (umem_init())
19377c478bd9Sstevel@tonic-gate 			ASSERT(umem_oversize_arena != NULL);
19387c478bd9Sstevel@tonic-gate 		else
19397c478bd9Sstevel@tonic-gate 			return (NULL);
19407c478bd9Sstevel@tonic-gate 	}
19417c478bd9Sstevel@tonic-gate 	buf = vmem_alloc(umem_oversize_arena, size, UMEM_VMFLAGS(umflag));
19427c478bd9Sstevel@tonic-gate 	if (buf == NULL) {
19437c478bd9Sstevel@tonic-gate 		umem_log_event(umem_failure_log, NULL, NULL, (void *)size);
19447c478bd9Sstevel@tonic-gate 		if (umem_alloc_retry(NULL, umflag))
19457c478bd9Sstevel@tonic-gate 			goto umem_alloc_retry;
19467c478bd9Sstevel@tonic-gate 	}
19477c478bd9Sstevel@tonic-gate 	return (buf);
19487c478bd9Sstevel@tonic-gate }
19497c478bd9Sstevel@tonic-gate 
19507c478bd9Sstevel@tonic-gate #pragma weak umem_alloc_align = _umem_alloc_align
19517c478bd9Sstevel@tonic-gate void *
19527c478bd9Sstevel@tonic-gate _umem_alloc_align(size_t size, size_t align, int umflag)
19537c478bd9Sstevel@tonic-gate {
19547c478bd9Sstevel@tonic-gate 	void *buf;
19557c478bd9Sstevel@tonic-gate 
19567c478bd9Sstevel@tonic-gate 	if (size == 0)
19577c478bd9Sstevel@tonic-gate 		return (NULL);
19587c478bd9Sstevel@tonic-gate 	if ((align & (align - 1)) != 0)
19597c478bd9Sstevel@tonic-gate 		return (NULL);
19607c478bd9Sstevel@tonic-gate 	if (align < UMEM_ALIGN)
19617c478bd9Sstevel@tonic-gate 		align = UMEM_ALIGN;
19627c478bd9Sstevel@tonic-gate 
19637c478bd9Sstevel@tonic-gate umem_alloc_align_retry:
19647c478bd9Sstevel@tonic-gate 	if (umem_memalign_arena == NULL) {
19657c478bd9Sstevel@tonic-gate 		if (umem_init())
19667c478bd9Sstevel@tonic-gate 			ASSERT(umem_oversize_arena != NULL);
19677c478bd9Sstevel@tonic-gate 		else
19687c478bd9Sstevel@tonic-gate 			return (NULL);
19697c478bd9Sstevel@tonic-gate 	}
19707c478bd9Sstevel@tonic-gate 	buf = vmem_xalloc(umem_memalign_arena, size, align, 0, 0, NULL, NULL,
19717c478bd9Sstevel@tonic-gate 	    UMEM_VMFLAGS(umflag));
19727c478bd9Sstevel@tonic-gate 	if (buf == NULL) {
19737c478bd9Sstevel@tonic-gate 		umem_log_event(umem_failure_log, NULL, NULL, (void *)size);
19747c478bd9Sstevel@tonic-gate 		if (umem_alloc_retry(NULL, umflag))
19757c478bd9Sstevel@tonic-gate 			goto umem_alloc_align_retry;
19767c478bd9Sstevel@tonic-gate 	}
19777c478bd9Sstevel@tonic-gate 	return (buf);
19787c478bd9Sstevel@tonic-gate }
19797c478bd9Sstevel@tonic-gate 
19807c478bd9Sstevel@tonic-gate #pragma weak umem_free = _umem_free
19817c478bd9Sstevel@tonic-gate void
19827c478bd9Sstevel@tonic-gate _umem_free(void *buf, size_t size)
19837c478bd9Sstevel@tonic-gate {
19847c478bd9Sstevel@tonic-gate 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
19857c478bd9Sstevel@tonic-gate 
19867c478bd9Sstevel@tonic-gate 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
19877c478bd9Sstevel@tonic-gate 		umem_cache_t *cp = umem_alloc_table[index];
19887c478bd9Sstevel@tonic-gate 		if (cp->cache_flags & UMF_BUFTAG) {
19897c478bd9Sstevel@tonic-gate 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
19907c478bd9Sstevel@tonic-gate 			uint32_t *ip = (uint32_t *)btp;
19917c478bd9Sstevel@tonic-gate 			if (ip[1] != UMEM_SIZE_ENCODE(size)) {
19927c478bd9Sstevel@tonic-gate 				if (*(uint64_t *)buf == UMEM_FREE_PATTERN) {
19937c478bd9Sstevel@tonic-gate 					umem_error(UMERR_DUPFREE, cp, buf);
19947c478bd9Sstevel@tonic-gate 					return;
19957c478bd9Sstevel@tonic-gate 				}
19967c478bd9Sstevel@tonic-gate 				if (UMEM_SIZE_VALID(ip[1])) {
19977c478bd9Sstevel@tonic-gate 					ip[0] = UMEM_SIZE_ENCODE(size);
19987c478bd9Sstevel@tonic-gate 					umem_error(UMERR_BADSIZE, cp, buf);
19997c478bd9Sstevel@tonic-gate 				} else {
20007c478bd9Sstevel@tonic-gate 					umem_error(UMERR_REDZONE, cp, buf);
20017c478bd9Sstevel@tonic-gate 				}
20027c478bd9Sstevel@tonic-gate 				return;
20037c478bd9Sstevel@tonic-gate 			}
20047c478bd9Sstevel@tonic-gate 			if (((uint8_t *)buf)[size] != UMEM_REDZONE_BYTE) {
20057c478bd9Sstevel@tonic-gate 				umem_error(UMERR_REDZONE, cp, buf);
20067c478bd9Sstevel@tonic-gate 				return;
20077c478bd9Sstevel@tonic-gate 			}
20087c478bd9Sstevel@tonic-gate 			btp->bt_redzone = UMEM_REDZONE_PATTERN;
20097c478bd9Sstevel@tonic-gate 		}
20107c478bd9Sstevel@tonic-gate 		_umem_cache_free(cp, buf);
20117c478bd9Sstevel@tonic-gate 	} else {
20127c478bd9Sstevel@tonic-gate 		if (buf == NULL && size == 0)
20137c478bd9Sstevel@tonic-gate 			return;
20147c478bd9Sstevel@tonic-gate 		vmem_free(umem_oversize_arena, buf, size);
20157c478bd9Sstevel@tonic-gate 	}
20167c478bd9Sstevel@tonic-gate }
20177c478bd9Sstevel@tonic-gate 
20187c478bd9Sstevel@tonic-gate #pragma weak umem_free_align = _umem_free_align
20197c478bd9Sstevel@tonic-gate void
20207c478bd9Sstevel@tonic-gate _umem_free_align(void *buf, size_t size)
20217c478bd9Sstevel@tonic-gate {
20227c478bd9Sstevel@tonic-gate 	if (buf == NULL && size == 0)
20237c478bd9Sstevel@tonic-gate 		return;
20247c478bd9Sstevel@tonic-gate 	vmem_xfree(umem_memalign_arena, buf, size);
20257c478bd9Sstevel@tonic-gate }
20267c478bd9Sstevel@tonic-gate 
20277c478bd9Sstevel@tonic-gate static void *
20287c478bd9Sstevel@tonic-gate umem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
20297c478bd9Sstevel@tonic-gate {
20307c478bd9Sstevel@tonic-gate 	size_t realsize = size + vmp->vm_quantum;
20317c478bd9Sstevel@tonic-gate 
20327c478bd9Sstevel@tonic-gate 	/*
20337c478bd9Sstevel@tonic-gate 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
20347c478bd9Sstevel@tonic-gate 	 * vm_quantum will cause integer wraparound.  Check for this, and
20357c478bd9Sstevel@tonic-gate 	 * blow off the firewall page in this case.  Note that such a
20367c478bd9Sstevel@tonic-gate 	 * giant allocation (the entire address space) can never be
20377c478bd9Sstevel@tonic-gate 	 * satisfied, so it will either fail immediately (VM_NOSLEEP)
20387c478bd9Sstevel@tonic-gate 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
20397c478bd9Sstevel@tonic-gate 	 * corresponding check in umem_firewall_va_free().
20407c478bd9Sstevel@tonic-gate 	 */
20417c478bd9Sstevel@tonic-gate 	if (realsize < size)
20427c478bd9Sstevel@tonic-gate 		realsize = size;
20437c478bd9Sstevel@tonic-gate 
20447c478bd9Sstevel@tonic-gate 	return (vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT));
20457c478bd9Sstevel@tonic-gate }
20467c478bd9Sstevel@tonic-gate 
20477c478bd9Sstevel@tonic-gate static void
20487c478bd9Sstevel@tonic-gate umem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
20497c478bd9Sstevel@tonic-gate {
20507c478bd9Sstevel@tonic-gate 	vmem_free(vmp, addr, size + vmp->vm_quantum);
20517c478bd9Sstevel@tonic-gate }
20527c478bd9Sstevel@tonic-gate 
20537c478bd9Sstevel@tonic-gate /*
20547c478bd9Sstevel@tonic-gate  * Reclaim all unused memory from a cache.
20557c478bd9Sstevel@tonic-gate  */
20567c478bd9Sstevel@tonic-gate static void
20577c478bd9Sstevel@tonic-gate umem_cache_reap(umem_cache_t *cp)
20587c478bd9Sstevel@tonic-gate {
20597c478bd9Sstevel@tonic-gate 	/*
20607c478bd9Sstevel@tonic-gate 	 * Ask the cache's owner to free some memory if possible.
20617c478bd9Sstevel@tonic-gate 	 * The idea is to handle things like the inode cache, which
20627c478bd9Sstevel@tonic-gate 	 * typically sits on a bunch of memory that it doesn't truly
20637c478bd9Sstevel@tonic-gate 	 * *need*.  Reclaim policy is entirely up to the owner; this
20647c478bd9Sstevel@tonic-gate 	 * callback is just an advisory plea for help.
20657c478bd9Sstevel@tonic-gate 	 */
20667c478bd9Sstevel@tonic-gate 	if (cp->cache_reclaim != NULL)
20677c478bd9Sstevel@tonic-gate 		cp->cache_reclaim(cp->cache_private);
20687c478bd9Sstevel@tonic-gate 
20697c478bd9Sstevel@tonic-gate 	umem_depot_ws_reap(cp);
20707c478bd9Sstevel@tonic-gate }
20717c478bd9Sstevel@tonic-gate 
20727c478bd9Sstevel@tonic-gate /*
20737c478bd9Sstevel@tonic-gate  * Purge all magazines from a cache and set its magazine limit to zero.
20747c478bd9Sstevel@tonic-gate  * All calls are serialized by being done by the update thread, except for
20757c478bd9Sstevel@tonic-gate  * the final call from umem_cache_destroy().
20767c478bd9Sstevel@tonic-gate  */
20777c478bd9Sstevel@tonic-gate static void
20787c478bd9Sstevel@tonic-gate umem_cache_magazine_purge(umem_cache_t *cp)
20797c478bd9Sstevel@tonic-gate {
20807c478bd9Sstevel@tonic-gate 	umem_cpu_cache_t *ccp;
20817c478bd9Sstevel@tonic-gate 	umem_magazine_t *mp, *pmp;
20827c478bd9Sstevel@tonic-gate 	int rounds, prounds, cpu_seqid;
20837c478bd9Sstevel@tonic-gate 
20847c478bd9Sstevel@tonic-gate 	ASSERT(cp->cache_next == NULL || IN_UPDATE());
20857c478bd9Sstevel@tonic-gate 
20867c478bd9Sstevel@tonic-gate 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
20877c478bd9Sstevel@tonic-gate 		ccp = &cp->cache_cpu[cpu_seqid];
20887c478bd9Sstevel@tonic-gate 
20897c478bd9Sstevel@tonic-gate 		(void) mutex_lock(&ccp->cc_lock);
20907c478bd9Sstevel@tonic-gate 		mp = ccp->cc_loaded;
20917c478bd9Sstevel@tonic-gate 		pmp = ccp->cc_ploaded;
20927c478bd9Sstevel@tonic-gate 		rounds = ccp->cc_rounds;
20937c478bd9Sstevel@tonic-gate 		prounds = ccp->cc_prounds;
20947c478bd9Sstevel@tonic-gate 		ccp->cc_loaded = NULL;
20957c478bd9Sstevel@tonic-gate 		ccp->cc_ploaded = NULL;
20967c478bd9Sstevel@tonic-gate 		ccp->cc_rounds = -1;
20977c478bd9Sstevel@tonic-gate 		ccp->cc_prounds = -1;
20987c478bd9Sstevel@tonic-gate 		ccp->cc_magsize = 0;
20997c478bd9Sstevel@tonic-gate 		(void) mutex_unlock(&ccp->cc_lock);
21007c478bd9Sstevel@tonic-gate 
21017c478bd9Sstevel@tonic-gate 		if (mp)
21027c478bd9Sstevel@tonic-gate 			umem_magazine_destroy(cp, mp, rounds);
21037c478bd9Sstevel@tonic-gate 		if (pmp)
21047c478bd9Sstevel@tonic-gate 			umem_magazine_destroy(cp, pmp, prounds);
21057c478bd9Sstevel@tonic-gate 	}
21067c478bd9Sstevel@tonic-gate 
21077c478bd9Sstevel@tonic-gate 	/*
21087c478bd9Sstevel@tonic-gate 	 * Updating the working set statistics twice in a row has the
21097c478bd9Sstevel@tonic-gate 	 * effect of setting the working set size to zero, so everything
21107c478bd9Sstevel@tonic-gate 	 * is eligible for reaping.
21117c478bd9Sstevel@tonic-gate 	 */
21127c478bd9Sstevel@tonic-gate 	umem_depot_ws_update(cp);
21137c478bd9Sstevel@tonic-gate 	umem_depot_ws_update(cp);
21147c478bd9Sstevel@tonic-gate 
21157c478bd9Sstevel@tonic-gate 	umem_depot_ws_reap(cp);
21167c478bd9Sstevel@tonic-gate }
21177c478bd9Sstevel@tonic-gate 
21187c478bd9Sstevel@tonic-gate /*
21197c478bd9Sstevel@tonic-gate  * Enable per-cpu magazines on a cache.
21207c478bd9Sstevel@tonic-gate  */
21217c478bd9Sstevel@tonic-gate static void
21227c478bd9Sstevel@tonic-gate umem_cache_magazine_enable(umem_cache_t *cp)
21237c478bd9Sstevel@tonic-gate {
21247c478bd9Sstevel@tonic-gate 	int cpu_seqid;
21257c478bd9Sstevel@tonic-gate 
21267c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_NOMAGAZINE)
21277c478bd9Sstevel@tonic-gate 		return;
21287c478bd9Sstevel@tonic-gate 
21297c478bd9Sstevel@tonic-gate 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
21307c478bd9Sstevel@tonic-gate 		umem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
21317c478bd9Sstevel@tonic-gate 		(void) mutex_lock(&ccp->cc_lock);
21327c478bd9Sstevel@tonic-gate 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
21337c478bd9Sstevel@tonic-gate 		(void) mutex_unlock(&ccp->cc_lock);
21347c478bd9Sstevel@tonic-gate 	}
21357c478bd9Sstevel@tonic-gate 
21367c478bd9Sstevel@tonic-gate }
21377c478bd9Sstevel@tonic-gate 
21387c478bd9Sstevel@tonic-gate /*
21397c478bd9Sstevel@tonic-gate  * Recompute a cache's magazine size.  The trade-off is that larger magazines
21407c478bd9Sstevel@tonic-gate  * provide a higher transfer rate with the depot, while smaller magazines
21417c478bd9Sstevel@tonic-gate  * reduce memory consumption.  Magazine resizing is an expensive operation;
21427c478bd9Sstevel@tonic-gate  * it should not be done frequently.
21437c478bd9Sstevel@tonic-gate  *
21447c478bd9Sstevel@tonic-gate  * Changes to the magazine size are serialized by only having one thread
21457c478bd9Sstevel@tonic-gate  * doing updates. (the update thread)
21467c478bd9Sstevel@tonic-gate  *
21477c478bd9Sstevel@tonic-gate  * Note: at present this only grows the magazine size.  It might be useful
21487c478bd9Sstevel@tonic-gate  * to allow shrinkage too.
21497c478bd9Sstevel@tonic-gate  */
21507c478bd9Sstevel@tonic-gate static void
21517c478bd9Sstevel@tonic-gate umem_cache_magazine_resize(umem_cache_t *cp)
21527c478bd9Sstevel@tonic-gate {
21537c478bd9Sstevel@tonic-gate 	umem_magtype_t *mtp = cp->cache_magtype;
21547c478bd9Sstevel@tonic-gate 
21557c478bd9Sstevel@tonic-gate 	ASSERT(IN_UPDATE());
21567c478bd9Sstevel@tonic-gate 
21577c478bd9Sstevel@tonic-gate 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
21587c478bd9Sstevel@tonic-gate 		umem_cache_magazine_purge(cp);
21597c478bd9Sstevel@tonic-gate 		(void) mutex_lock(&cp->cache_depot_lock);
21607c478bd9Sstevel@tonic-gate 		cp->cache_magtype = ++mtp;
21617c478bd9Sstevel@tonic-gate 		cp->cache_depot_contention_prev =
21627c478bd9Sstevel@tonic-gate 		    cp->cache_depot_contention + INT_MAX;
21637c478bd9Sstevel@tonic-gate 		(void) mutex_unlock(&cp->cache_depot_lock);
21647c478bd9Sstevel@tonic-gate 		umem_cache_magazine_enable(cp);
21657c478bd9Sstevel@tonic-gate 	}
21667c478bd9Sstevel@tonic-gate }
21677c478bd9Sstevel@tonic-gate 
21687c478bd9Sstevel@tonic-gate /*
21697c478bd9Sstevel@tonic-gate  * Rescale a cache's hash table, so that the table size is roughly the
21707c478bd9Sstevel@tonic-gate  * cache size.  We want the average lookup time to be extremely small.
21717c478bd9Sstevel@tonic-gate  */
21727c478bd9Sstevel@tonic-gate static void
21737c478bd9Sstevel@tonic-gate umem_hash_rescale(umem_cache_t *cp)
21747c478bd9Sstevel@tonic-gate {
21757c478bd9Sstevel@tonic-gate 	umem_bufctl_t **old_table, **new_table, *bcp;
21767c478bd9Sstevel@tonic-gate 	size_t old_size, new_size, h;
21777c478bd9Sstevel@tonic-gate 
21787c478bd9Sstevel@tonic-gate 	ASSERT(IN_UPDATE());
21797c478bd9Sstevel@tonic-gate 
21807c478bd9Sstevel@tonic-gate 	new_size = MAX(UMEM_HASH_INITIAL,
21817c478bd9Sstevel@tonic-gate 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
21827c478bd9Sstevel@tonic-gate 	old_size = cp->cache_hash_mask + 1;
21837c478bd9Sstevel@tonic-gate 
21847c478bd9Sstevel@tonic-gate 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
21857c478bd9Sstevel@tonic-gate 		return;
21867c478bd9Sstevel@tonic-gate 
21877c478bd9Sstevel@tonic-gate 	new_table = vmem_alloc(umem_hash_arena, new_size * sizeof (void *),
21887c478bd9Sstevel@tonic-gate 	    VM_NOSLEEP);
21897c478bd9Sstevel@tonic-gate 	if (new_table == NULL)
21907c478bd9Sstevel@tonic-gate 		return;
21917c478bd9Sstevel@tonic-gate 	bzero(new_table, new_size * sizeof (void *));
21927c478bd9Sstevel@tonic-gate 
21937c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&cp->cache_lock);
21947c478bd9Sstevel@tonic-gate 
21957c478bd9Sstevel@tonic-gate 	old_size = cp->cache_hash_mask + 1;
21967c478bd9Sstevel@tonic-gate 	old_table = cp->cache_hash_table;
21977c478bd9Sstevel@tonic-gate 
21987c478bd9Sstevel@tonic-gate 	cp->cache_hash_mask = new_size - 1;
21997c478bd9Sstevel@tonic-gate 	cp->cache_hash_table = new_table;
22007c478bd9Sstevel@tonic-gate 	cp->cache_rescale++;
22017c478bd9Sstevel@tonic-gate 
22027c478bd9Sstevel@tonic-gate 	for (h = 0; h < old_size; h++) {
22037c478bd9Sstevel@tonic-gate 		bcp = old_table[h];
22047c478bd9Sstevel@tonic-gate 		while (bcp != NULL) {
22057c478bd9Sstevel@tonic-gate 			void *addr = bcp->bc_addr;
22067c478bd9Sstevel@tonic-gate 			umem_bufctl_t *next_bcp = bcp->bc_next;
22077c478bd9Sstevel@tonic-gate 			umem_bufctl_t **hash_bucket = UMEM_HASH(cp, addr);
22087c478bd9Sstevel@tonic-gate 			bcp->bc_next = *hash_bucket;
22097c478bd9Sstevel@tonic-gate 			*hash_bucket = bcp;
22107c478bd9Sstevel@tonic-gate 			bcp = next_bcp;
22117c478bd9Sstevel@tonic-gate 		}
22127c478bd9Sstevel@tonic-gate 	}
22137c478bd9Sstevel@tonic-gate 
22147c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&cp->cache_lock);
22157c478bd9Sstevel@tonic-gate 
22167c478bd9Sstevel@tonic-gate 	vmem_free(umem_hash_arena, old_table, old_size * sizeof (void *));
22177c478bd9Sstevel@tonic-gate }
22187c478bd9Sstevel@tonic-gate 
22197c478bd9Sstevel@tonic-gate /*
22207c478bd9Sstevel@tonic-gate  * Perform periodic maintenance on a cache: hash rescaling,
22217c478bd9Sstevel@tonic-gate  * depot working-set update, and magazine resizing.
22227c478bd9Sstevel@tonic-gate  */
22237c478bd9Sstevel@tonic-gate void
22247c478bd9Sstevel@tonic-gate umem_cache_update(umem_cache_t *cp)
22257c478bd9Sstevel@tonic-gate {
22267c478bd9Sstevel@tonic-gate 	int update_flags = 0;
22277c478bd9Sstevel@tonic-gate 
22287c478bd9Sstevel@tonic-gate 	ASSERT(MUTEX_HELD(&umem_cache_lock));
22297c478bd9Sstevel@tonic-gate 
22307c478bd9Sstevel@tonic-gate 	/*
22317c478bd9Sstevel@tonic-gate 	 * If the cache has become much larger or smaller than its hash table,
22327c478bd9Sstevel@tonic-gate 	 * fire off a request to rescale the hash table.
22337c478bd9Sstevel@tonic-gate 	 */
22347c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&cp->cache_lock);
22357c478bd9Sstevel@tonic-gate 
22367c478bd9Sstevel@tonic-gate 	if ((cp->cache_flags & UMF_HASH) &&
22377c478bd9Sstevel@tonic-gate 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
22387c478bd9Sstevel@tonic-gate 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
22397c478bd9Sstevel@tonic-gate 	    cp->cache_hash_mask > UMEM_HASH_INITIAL)))
22407c478bd9Sstevel@tonic-gate 		update_flags |= UMU_HASH_RESCALE;
22417c478bd9Sstevel@tonic-gate 
22427c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&cp->cache_lock);
22437c478bd9Sstevel@tonic-gate 
22447c478bd9Sstevel@tonic-gate 	/*
22457c478bd9Sstevel@tonic-gate 	 * Update the depot working set statistics.
22467c478bd9Sstevel@tonic-gate 	 */
22477c478bd9Sstevel@tonic-gate 	umem_depot_ws_update(cp);
22487c478bd9Sstevel@tonic-gate 
22497c478bd9Sstevel@tonic-gate 	/*
22507c478bd9Sstevel@tonic-gate 	 * If there's a lot of contention in the depot,
22517c478bd9Sstevel@tonic-gate 	 * increase the magazine size.
22527c478bd9Sstevel@tonic-gate 	 */
22537c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&cp->cache_depot_lock);
22547c478bd9Sstevel@tonic-gate 
22557c478bd9Sstevel@tonic-gate 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
22567c478bd9Sstevel@tonic-gate 	    (int)(cp->cache_depot_contention -
22577c478bd9Sstevel@tonic-gate 	    cp->cache_depot_contention_prev) > umem_depot_contention)
22587c478bd9Sstevel@tonic-gate 		update_flags |= UMU_MAGAZINE_RESIZE;
22597c478bd9Sstevel@tonic-gate 
22607c478bd9Sstevel@tonic-gate 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
22617c478bd9Sstevel@tonic-gate 
22627c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&cp->cache_depot_lock);
22637c478bd9Sstevel@tonic-gate 
22647c478bd9Sstevel@tonic-gate 	if (update_flags)
22657c478bd9Sstevel@tonic-gate 		umem_add_update(cp, update_flags);
22667c478bd9Sstevel@tonic-gate }
22677c478bd9Sstevel@tonic-gate 
22687c478bd9Sstevel@tonic-gate /*
22697c478bd9Sstevel@tonic-gate  * Runs all pending updates.
22707c478bd9Sstevel@tonic-gate  *
22717c478bd9Sstevel@tonic-gate  * The update lock must be held on entrance, and will be held on exit.
22727c478bd9Sstevel@tonic-gate  */
22737c478bd9Sstevel@tonic-gate void
22747c478bd9Sstevel@tonic-gate umem_process_updates(void)
22757c478bd9Sstevel@tonic-gate {
22767c478bd9Sstevel@tonic-gate 	ASSERT(MUTEX_HELD(&umem_update_lock));
22777c478bd9Sstevel@tonic-gate 
22787c478bd9Sstevel@tonic-gate 	while (umem_null_cache.cache_unext != &umem_null_cache) {
22797c478bd9Sstevel@tonic-gate 		int notify = 0;
22807c478bd9Sstevel@tonic-gate 		umem_cache_t *cp = umem_null_cache.cache_unext;
22817c478bd9Sstevel@tonic-gate 
22827c478bd9Sstevel@tonic-gate 		cp->cache_uprev->cache_unext = cp->cache_unext;
22837c478bd9Sstevel@tonic-gate 		cp->cache_unext->cache_uprev = cp->cache_uprev;
22847c478bd9Sstevel@tonic-gate 		cp->cache_uprev = cp->cache_unext = NULL;
22857c478bd9Sstevel@tonic-gate 
22867c478bd9Sstevel@tonic-gate 		ASSERT(!(cp->cache_uflags & UMU_ACTIVE));
22877c478bd9Sstevel@tonic-gate 
22887c478bd9Sstevel@tonic-gate 		while (cp->cache_uflags) {
22897c478bd9Sstevel@tonic-gate 			int uflags = (cp->cache_uflags |= UMU_ACTIVE);
22907c478bd9Sstevel@tonic-gate 			(void) mutex_unlock(&umem_update_lock);
22917c478bd9Sstevel@tonic-gate 
22927c478bd9Sstevel@tonic-gate 			/*
22937c478bd9Sstevel@tonic-gate 			 * The order here is important.  Each step can speed up
22947c478bd9Sstevel@tonic-gate 			 * later steps.
22957c478bd9Sstevel@tonic-gate 			 */
22967c478bd9Sstevel@tonic-gate 
22977c478bd9Sstevel@tonic-gate 			if (uflags & UMU_HASH_RESCALE)
22987c478bd9Sstevel@tonic-gate 				umem_hash_rescale(cp);
22997c478bd9Sstevel@tonic-gate 
23007c478bd9Sstevel@tonic-gate 			if (uflags & UMU_MAGAZINE_RESIZE)
23017c478bd9Sstevel@tonic-gate 				umem_cache_magazine_resize(cp);
23027c478bd9Sstevel@tonic-gate 
23037c478bd9Sstevel@tonic-gate 			if (uflags & UMU_REAP)
23047c478bd9Sstevel@tonic-gate 				umem_cache_reap(cp);
23057c478bd9Sstevel@tonic-gate 
23067c478bd9Sstevel@tonic-gate 			(void) mutex_lock(&umem_update_lock);
23077c478bd9Sstevel@tonic-gate 
23087c478bd9Sstevel@tonic-gate 			/*
23097c478bd9Sstevel@tonic-gate 			 * check if anyone has requested notification
23107c478bd9Sstevel@tonic-gate 			 */
23117c478bd9Sstevel@tonic-gate 			if (cp->cache_uflags & UMU_NOTIFY) {
23127c478bd9Sstevel@tonic-gate 				uflags |= UMU_NOTIFY;
23137c478bd9Sstevel@tonic-gate 				notify = 1;
23147c478bd9Sstevel@tonic-gate 			}
23157c478bd9Sstevel@tonic-gate 			cp->cache_uflags &= ~uflags;
23167c478bd9Sstevel@tonic-gate 		}
23177c478bd9Sstevel@tonic-gate 		if (notify)
23187c478bd9Sstevel@tonic-gate 			(void) cond_broadcast(&umem_update_cv);
23197c478bd9Sstevel@tonic-gate 	}
23207c478bd9Sstevel@tonic-gate }
23217c478bd9Sstevel@tonic-gate 
23227c478bd9Sstevel@tonic-gate #ifndef UMEM_STANDALONE
23237c478bd9Sstevel@tonic-gate static void
23247c478bd9Sstevel@tonic-gate umem_st_update(void)
23257c478bd9Sstevel@tonic-gate {
23267c478bd9Sstevel@tonic-gate 	ASSERT(MUTEX_HELD(&umem_update_lock));
23277c478bd9Sstevel@tonic-gate 	ASSERT(umem_update_thr == 0 && umem_st_update_thr == 0);
23287c478bd9Sstevel@tonic-gate 
23297c478bd9Sstevel@tonic-gate 	umem_st_update_thr = thr_self();
23307c478bd9Sstevel@tonic-gate 
23317c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&umem_update_lock);
23327c478bd9Sstevel@tonic-gate 
23337c478bd9Sstevel@tonic-gate 	vmem_update(NULL);
23347c478bd9Sstevel@tonic-gate 	umem_cache_applyall(umem_cache_update);
23357c478bd9Sstevel@tonic-gate 
23367c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&umem_update_lock);
23377c478bd9Sstevel@tonic-gate 
23387c478bd9Sstevel@tonic-gate 	umem_process_updates();	/* does all of the requested work */
23397c478bd9Sstevel@tonic-gate 
23407c478bd9Sstevel@tonic-gate 	umem_reap_next = gethrtime() +
23417c478bd9Sstevel@tonic-gate 	    (hrtime_t)umem_reap_interval * NANOSEC;
23427c478bd9Sstevel@tonic-gate 
23437c478bd9Sstevel@tonic-gate 	umem_reaping = UMEM_REAP_DONE;
23447c478bd9Sstevel@tonic-gate 
23457c478bd9Sstevel@tonic-gate 	umem_st_update_thr = 0;
23467c478bd9Sstevel@tonic-gate }
23477c478bd9Sstevel@tonic-gate #endif
23487c478bd9Sstevel@tonic-gate 
23497c478bd9Sstevel@tonic-gate /*
23507c478bd9Sstevel@tonic-gate  * Reclaim all unused memory from all caches.  Called from vmem when memory
23517c478bd9Sstevel@tonic-gate  * gets tight.  Must be called with no locks held.
23527c478bd9Sstevel@tonic-gate  *
23537c478bd9Sstevel@tonic-gate  * This just requests a reap on all caches, and notifies the update thread.
23547c478bd9Sstevel@tonic-gate  */
23557c478bd9Sstevel@tonic-gate void
23567c478bd9Sstevel@tonic-gate umem_reap(void)
23577c478bd9Sstevel@tonic-gate {
23587c478bd9Sstevel@tonic-gate #ifndef UMEM_STANDALONE
23597c478bd9Sstevel@tonic-gate 	extern int __nthreads(void);
23607c478bd9Sstevel@tonic-gate #endif
23617c478bd9Sstevel@tonic-gate 
23627c478bd9Sstevel@tonic-gate 	if (umem_ready != UMEM_READY || umem_reaping != UMEM_REAP_DONE ||
23637c478bd9Sstevel@tonic-gate 	    gethrtime() < umem_reap_next)
23647c478bd9Sstevel@tonic-gate 		return;
23657c478bd9Sstevel@tonic-gate 
23667c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&umem_update_lock);
23677c478bd9Sstevel@tonic-gate 
23687c478bd9Sstevel@tonic-gate 	if (umem_reaping != UMEM_REAP_DONE || gethrtime() < umem_reap_next) {
23697c478bd9Sstevel@tonic-gate 		(void) mutex_unlock(&umem_update_lock);
23707c478bd9Sstevel@tonic-gate 		return;
23717c478bd9Sstevel@tonic-gate 	}
23727c478bd9Sstevel@tonic-gate 	umem_reaping = UMEM_REAP_ADDING;	/* lock out other reaps */
23737c478bd9Sstevel@tonic-gate 
23747c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&umem_update_lock);
23757c478bd9Sstevel@tonic-gate 
23767c478bd9Sstevel@tonic-gate 	umem_updateall(UMU_REAP);
23777c478bd9Sstevel@tonic-gate 
23787c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&umem_update_lock);
23797c478bd9Sstevel@tonic-gate 
23807c478bd9Sstevel@tonic-gate 	umem_reaping = UMEM_REAP_ACTIVE;
23817c478bd9Sstevel@tonic-gate 
23827c478bd9Sstevel@tonic-gate 	/* Standalone is single-threaded */
23837c478bd9Sstevel@tonic-gate #ifndef UMEM_STANDALONE
23847c478bd9Sstevel@tonic-gate 	if (umem_update_thr == 0) {
23857c478bd9Sstevel@tonic-gate 		/*
23867c478bd9Sstevel@tonic-gate 		 * The update thread does not exist.  If the process is
23877c478bd9Sstevel@tonic-gate 		 * multi-threaded, create it.  If not, or the creation fails,
23887c478bd9Sstevel@tonic-gate 		 * do the update processing inline.
23897c478bd9Sstevel@tonic-gate 		 */
23907c478bd9Sstevel@tonic-gate 		ASSERT(umem_st_update_thr == 0);
23917c478bd9Sstevel@tonic-gate 
23927c478bd9Sstevel@tonic-gate 		if (__nthreads() <= 1 || umem_create_update_thread() == 0)
23937c478bd9Sstevel@tonic-gate 			umem_st_update();
23947c478bd9Sstevel@tonic-gate 	}
23957c478bd9Sstevel@tonic-gate 
23967c478bd9Sstevel@tonic-gate 	(void) cond_broadcast(&umem_update_cv);	/* wake up the update thread */
23977c478bd9Sstevel@tonic-gate #endif
23987c478bd9Sstevel@tonic-gate 
23997c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&umem_update_lock);
24007c478bd9Sstevel@tonic-gate }
24017c478bd9Sstevel@tonic-gate 
24027c478bd9Sstevel@tonic-gate umem_cache_t *
24037c478bd9Sstevel@tonic-gate umem_cache_create(
24047c478bd9Sstevel@tonic-gate 	char *name,		/* descriptive name for this cache */
24057c478bd9Sstevel@tonic-gate 	size_t bufsize,		/* size of the objects it manages */
24067c478bd9Sstevel@tonic-gate 	size_t align,		/* required object alignment */
24077c478bd9Sstevel@tonic-gate 	umem_constructor_t *constructor, /* object constructor */
24087c478bd9Sstevel@tonic-gate 	umem_destructor_t *destructor, /* object destructor */
24097c478bd9Sstevel@tonic-gate 	umem_reclaim_t *reclaim, /* memory reclaim callback */
24107c478bd9Sstevel@tonic-gate 	void *private,		/* pass-thru arg for constr/destr/reclaim */
24117c478bd9Sstevel@tonic-gate 	vmem_t *vmp,		/* vmem source for slab allocation */
24127c478bd9Sstevel@tonic-gate 	int cflags)		/* cache creation flags */
24137c478bd9Sstevel@tonic-gate {
24147c478bd9Sstevel@tonic-gate 	int cpu_seqid;
24157c478bd9Sstevel@tonic-gate 	size_t chunksize;
24167c478bd9Sstevel@tonic-gate 	umem_cache_t *cp, *cnext, *cprev;
24177c478bd9Sstevel@tonic-gate 	umem_magtype_t *mtp;
24187c478bd9Sstevel@tonic-gate 	size_t csize;
24197c478bd9Sstevel@tonic-gate 	size_t phase;
24207c478bd9Sstevel@tonic-gate 
24217c478bd9Sstevel@tonic-gate 	/*
24227c478bd9Sstevel@tonic-gate 	 * The init thread is allowed to create internal and quantum caches.
24237c478bd9Sstevel@tonic-gate 	 *
24247c478bd9Sstevel@tonic-gate 	 * Other threads must wait until until initialization is complete.
24257c478bd9Sstevel@tonic-gate 	 */
24267c478bd9Sstevel@tonic-gate 	if (umem_init_thr == thr_self())
24277c478bd9Sstevel@tonic-gate 		ASSERT((cflags & (UMC_INTERNAL | UMC_QCACHE)) != 0);
24287c478bd9Sstevel@tonic-gate 	else {
24297c478bd9Sstevel@tonic-gate 		ASSERT(!(cflags & UMC_INTERNAL));
24307c478bd9Sstevel@tonic-gate 		if (umem_ready != UMEM_READY && umem_init() == 0) {
24317c478bd9Sstevel@tonic-gate 			errno = EAGAIN;
24327c478bd9Sstevel@tonic-gate 			return (NULL);
24337c478bd9Sstevel@tonic-gate 		}
24347c478bd9Sstevel@tonic-gate 	}
24357c478bd9Sstevel@tonic-gate 
24367c478bd9Sstevel@tonic-gate 	csize = UMEM_CACHE_SIZE(umem_max_ncpus);
24377c478bd9Sstevel@tonic-gate 	phase = P2NPHASE(csize, UMEM_CPU_CACHE_SIZE);
24387c478bd9Sstevel@tonic-gate 
24397c478bd9Sstevel@tonic-gate 	if (vmp == NULL)
24407c478bd9Sstevel@tonic-gate 		vmp = umem_default_arena;
24417c478bd9Sstevel@tonic-gate 
24427c478bd9Sstevel@tonic-gate 	ASSERT(P2PHASE(phase, UMEM_ALIGN) == 0);
24437c478bd9Sstevel@tonic-gate 
24447c478bd9Sstevel@tonic-gate 	/*
24457c478bd9Sstevel@tonic-gate 	 * Check that the arguments are reasonable
24467c478bd9Sstevel@tonic-gate 	 */
24477c478bd9Sstevel@tonic-gate 	if ((align & (align - 1)) != 0 || align > vmp->vm_quantum ||
24487c478bd9Sstevel@tonic-gate 	    ((cflags & UMC_NOHASH) && (cflags & UMC_NOTOUCH)) ||
24497c478bd9Sstevel@tonic-gate 	    name == NULL || bufsize == 0) {
24507c478bd9Sstevel@tonic-gate 		errno = EINVAL;
24517c478bd9Sstevel@tonic-gate 		return (NULL);
24527c478bd9Sstevel@tonic-gate 	}
24537c478bd9Sstevel@tonic-gate 
24547c478bd9Sstevel@tonic-gate 	/*
24557c478bd9Sstevel@tonic-gate 	 * If align == 0, we set it to the minimum required alignment.
24567c478bd9Sstevel@tonic-gate 	 *
24577c478bd9Sstevel@tonic-gate 	 * If align < UMEM_ALIGN, we round it up to UMEM_ALIGN, unless
24587c478bd9Sstevel@tonic-gate 	 * UMC_NOTOUCH was passed.
24597c478bd9Sstevel@tonic-gate 	 */
24607c478bd9Sstevel@tonic-gate 	if (align == 0) {
24617c478bd9Sstevel@tonic-gate 		if (P2ROUNDUP(bufsize, UMEM_ALIGN) >= UMEM_SECOND_ALIGN)
24627c478bd9Sstevel@tonic-gate 			align = UMEM_SECOND_ALIGN;
24637c478bd9Sstevel@tonic-gate 		else
24647c478bd9Sstevel@tonic-gate 			align = UMEM_ALIGN;
24657c478bd9Sstevel@tonic-gate 	} else if (align < UMEM_ALIGN && (cflags & UMC_NOTOUCH) == 0)
24667c478bd9Sstevel@tonic-gate 		align = UMEM_ALIGN;
24677c478bd9Sstevel@tonic-gate 
24687c478bd9Sstevel@tonic-gate 
24697c478bd9Sstevel@tonic-gate 	/*
24707c478bd9Sstevel@tonic-gate 	 * Get a umem_cache structure.  We arrange that cp->cache_cpu[]
24717c478bd9Sstevel@tonic-gate 	 * is aligned on a UMEM_CPU_CACHE_SIZE boundary to prevent
24727c478bd9Sstevel@tonic-gate 	 * false sharing of per-CPU data.
24737c478bd9Sstevel@tonic-gate 	 */
24747c478bd9Sstevel@tonic-gate 	cp = vmem_xalloc(umem_cache_arena, csize, UMEM_CPU_CACHE_SIZE, phase,
24757c478bd9Sstevel@tonic-gate 	    0, NULL, NULL, VM_NOSLEEP);
24767c478bd9Sstevel@tonic-gate 
24777c478bd9Sstevel@tonic-gate 	if (cp == NULL) {
24787c478bd9Sstevel@tonic-gate 		errno = EAGAIN;
24797c478bd9Sstevel@tonic-gate 		return (NULL);
24807c478bd9Sstevel@tonic-gate 	}
24817c478bd9Sstevel@tonic-gate 
24827c478bd9Sstevel@tonic-gate 	bzero(cp, csize);
24837c478bd9Sstevel@tonic-gate 
24847c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&umem_flags_lock);
24857c478bd9Sstevel@tonic-gate 	if (umem_flags & UMF_RANDOMIZE)
24867c478bd9Sstevel@tonic-gate 		umem_flags = (((umem_flags | ~UMF_RANDOM) + 1) & UMF_RANDOM) |
24877c478bd9Sstevel@tonic-gate 		    UMF_RANDOMIZE;
24887c478bd9Sstevel@tonic-gate 	cp->cache_flags = umem_flags | (cflags & UMF_DEBUG);
24897c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&umem_flags_lock);
24907c478bd9Sstevel@tonic-gate 
24917c478bd9Sstevel@tonic-gate 	/*
24927c478bd9Sstevel@tonic-gate 	 * Make sure all the various flags are reasonable.
24937c478bd9Sstevel@tonic-gate 	 */
24947c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_LITE) {
24957c478bd9Sstevel@tonic-gate 		if (bufsize >= umem_lite_minsize &&
24967c478bd9Sstevel@tonic-gate 		    align <= umem_lite_maxalign &&
24977c478bd9Sstevel@tonic-gate 		    P2PHASE(bufsize, umem_lite_maxalign) != 0) {
24987c478bd9Sstevel@tonic-gate 			cp->cache_flags |= UMF_BUFTAG;
24997c478bd9Sstevel@tonic-gate 			cp->cache_flags &= ~(UMF_AUDIT | UMF_FIREWALL);
25007c478bd9Sstevel@tonic-gate 		} else {
25017c478bd9Sstevel@tonic-gate 			cp->cache_flags &= ~UMF_DEBUG;
25027c478bd9Sstevel@tonic-gate 		}
25037c478bd9Sstevel@tonic-gate 	}
25047c478bd9Sstevel@tonic-gate 
25057c478bd9Sstevel@tonic-gate 	if ((cflags & UMC_QCACHE) && (cp->cache_flags & UMF_AUDIT))
25067c478bd9Sstevel@tonic-gate 		cp->cache_flags |= UMF_NOMAGAZINE;
25077c478bd9Sstevel@tonic-gate 
25087c478bd9Sstevel@tonic-gate 	if (cflags & UMC_NODEBUG)
25097c478bd9Sstevel@tonic-gate 		cp->cache_flags &= ~UMF_DEBUG;
25107c478bd9Sstevel@tonic-gate 
25117c478bd9Sstevel@tonic-gate 	if (cflags & UMC_NOTOUCH)
25127c478bd9Sstevel@tonic-gate 		cp->cache_flags &= ~UMF_TOUCH;
25137c478bd9Sstevel@tonic-gate 
25147c478bd9Sstevel@tonic-gate 	if (cflags & UMC_NOHASH)
25157c478bd9Sstevel@tonic-gate 		cp->cache_flags &= ~(UMF_AUDIT | UMF_FIREWALL);
25167c478bd9Sstevel@tonic-gate 
25177c478bd9Sstevel@tonic-gate 	if (cflags & UMC_NOMAGAZINE)
25187c478bd9Sstevel@tonic-gate 		cp->cache_flags |= UMF_NOMAGAZINE;
25197c478bd9Sstevel@tonic-gate 
25207c478bd9Sstevel@tonic-gate 	if ((cp->cache_flags & UMF_AUDIT) && !(cflags & UMC_NOTOUCH))
25217c478bd9Sstevel@tonic-gate 		cp->cache_flags |= UMF_REDZONE;
25227c478bd9Sstevel@tonic-gate 
25237c478bd9Sstevel@tonic-gate 	if ((cp->cache_flags & UMF_BUFTAG) && bufsize >= umem_minfirewall &&
25247c478bd9Sstevel@tonic-gate 	    !(cp->cache_flags & UMF_LITE) && !(cflags & UMC_NOHASH))
25257c478bd9Sstevel@tonic-gate 		cp->cache_flags |= UMF_FIREWALL;
25267c478bd9Sstevel@tonic-gate 
25277c478bd9Sstevel@tonic-gate 	if (vmp != umem_default_arena || umem_firewall_arena == NULL)
25287c478bd9Sstevel@tonic-gate 		cp->cache_flags &= ~UMF_FIREWALL;
25297c478bd9Sstevel@tonic-gate 
25307c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_FIREWALL) {
25317c478bd9Sstevel@tonic-gate 		cp->cache_flags &= ~UMF_BUFTAG;
25327c478bd9Sstevel@tonic-gate 		cp->cache_flags |= UMF_NOMAGAZINE;
25337c478bd9Sstevel@tonic-gate 		ASSERT(vmp == umem_default_arena);
25347c478bd9Sstevel@tonic-gate 		vmp = umem_firewall_arena;
25357c478bd9Sstevel@tonic-gate 	}
25367c478bd9Sstevel@tonic-gate 
25377c478bd9Sstevel@tonic-gate 	/*
25387c478bd9Sstevel@tonic-gate 	 * Set cache properties.
25397c478bd9Sstevel@tonic-gate 	 */
25407c478bd9Sstevel@tonic-gate 	(void) strncpy(cp->cache_name, name, sizeof (cp->cache_name) - 1);
25417c478bd9Sstevel@tonic-gate 	cp->cache_bufsize = bufsize;
25427c478bd9Sstevel@tonic-gate 	cp->cache_align = align;
25437c478bd9Sstevel@tonic-gate 	cp->cache_constructor = constructor;
25447c478bd9Sstevel@tonic-gate 	cp->cache_destructor = destructor;
25457c478bd9Sstevel@tonic-gate 	cp->cache_reclaim = reclaim;
25467c478bd9Sstevel@tonic-gate 	cp->cache_private = private;
25477c478bd9Sstevel@tonic-gate 	cp->cache_arena = vmp;
25487c478bd9Sstevel@tonic-gate 	cp->cache_cflags = cflags;
25497c478bd9Sstevel@tonic-gate 	cp->cache_cpu_mask = umem_cpu_mask;
25507c478bd9Sstevel@tonic-gate 
25517c478bd9Sstevel@tonic-gate 	/*
25527c478bd9Sstevel@tonic-gate 	 * Determine the chunk size.
25537c478bd9Sstevel@tonic-gate 	 */
25547c478bd9Sstevel@tonic-gate 	chunksize = bufsize;
25557c478bd9Sstevel@tonic-gate 
25567c478bd9Sstevel@tonic-gate 	if (align >= UMEM_ALIGN) {
25577c478bd9Sstevel@tonic-gate 		chunksize = P2ROUNDUP(chunksize, UMEM_ALIGN);
25587c478bd9Sstevel@tonic-gate 		cp->cache_bufctl = chunksize - UMEM_ALIGN;
25597c478bd9Sstevel@tonic-gate 	}
25607c478bd9Sstevel@tonic-gate 
25617c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_BUFTAG) {
25627c478bd9Sstevel@tonic-gate 		cp->cache_bufctl = chunksize;
25637c478bd9Sstevel@tonic-gate 		cp->cache_buftag = chunksize;
25647c478bd9Sstevel@tonic-gate 		chunksize += sizeof (umem_buftag_t);
25657c478bd9Sstevel@tonic-gate 	}
25667c478bd9Sstevel@tonic-gate 
25677c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_DEADBEEF) {
25687c478bd9Sstevel@tonic-gate 		cp->cache_verify = MIN(cp->cache_buftag, umem_maxverify);
25697c478bd9Sstevel@tonic-gate 		if (cp->cache_flags & UMF_LITE)
25707c478bd9Sstevel@tonic-gate 			cp->cache_verify = MIN(cp->cache_verify, UMEM_ALIGN);
25717c478bd9Sstevel@tonic-gate 	}
25727c478bd9Sstevel@tonic-gate 
25737c478bd9Sstevel@tonic-gate 	cp->cache_contents = MIN(cp->cache_bufctl, umem_content_maxsave);
25747c478bd9Sstevel@tonic-gate 
25757c478bd9Sstevel@tonic-gate 	cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
25767c478bd9Sstevel@tonic-gate 
25777c478bd9Sstevel@tonic-gate 	if (chunksize < bufsize) {
25787c478bd9Sstevel@tonic-gate 		errno = ENOMEM;
25797c478bd9Sstevel@tonic-gate 		goto fail;
25807c478bd9Sstevel@tonic-gate 	}
25817c478bd9Sstevel@tonic-gate 
25827c478bd9Sstevel@tonic-gate 	/*
25837c478bd9Sstevel@tonic-gate 	 * Now that we know the chunk size, determine the optimal slab size.
25847c478bd9Sstevel@tonic-gate 	 */
25857c478bd9Sstevel@tonic-gate 	if (vmp == umem_firewall_arena) {
25867c478bd9Sstevel@tonic-gate 		cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
25877c478bd9Sstevel@tonic-gate 		cp->cache_mincolor = cp->cache_slabsize - chunksize;
25887c478bd9Sstevel@tonic-gate 		cp->cache_maxcolor = cp->cache_mincolor;
25897c478bd9Sstevel@tonic-gate 		cp->cache_flags |= UMF_HASH;
25907c478bd9Sstevel@tonic-gate 		ASSERT(!(cp->cache_flags & UMF_BUFTAG));
25917c478bd9Sstevel@tonic-gate 	} else if ((cflags & UMC_NOHASH) || (!(cflags & UMC_NOTOUCH) &&
25927c478bd9Sstevel@tonic-gate 	    !(cp->cache_flags & UMF_AUDIT) &&
25937c478bd9Sstevel@tonic-gate 	    chunksize < vmp->vm_quantum / UMEM_VOID_FRACTION)) {
25947c478bd9Sstevel@tonic-gate 		cp->cache_slabsize = vmp->vm_quantum;
25957c478bd9Sstevel@tonic-gate 		cp->cache_mincolor = 0;
25967c478bd9Sstevel@tonic-gate 		cp->cache_maxcolor =
25977c478bd9Sstevel@tonic-gate 		    (cp->cache_slabsize - sizeof (umem_slab_t)) % chunksize;
25987c478bd9Sstevel@tonic-gate 
25997c478bd9Sstevel@tonic-gate 		if (chunksize + sizeof (umem_slab_t) > cp->cache_slabsize) {
26007c478bd9Sstevel@tonic-gate 			errno = EINVAL;
26017c478bd9Sstevel@tonic-gate 			goto fail;
26027c478bd9Sstevel@tonic-gate 		}
26037c478bd9Sstevel@tonic-gate 		ASSERT(!(cp->cache_flags & UMF_AUDIT));
26047c478bd9Sstevel@tonic-gate 	} else {
26057c478bd9Sstevel@tonic-gate 		size_t chunks, bestfit, waste, slabsize;
26067c478bd9Sstevel@tonic-gate 		size_t minwaste = LONG_MAX;
26077c478bd9Sstevel@tonic-gate 
26087c478bd9Sstevel@tonic-gate 		for (chunks = 1; chunks <= UMEM_VOID_FRACTION; chunks++) {
26097c478bd9Sstevel@tonic-gate 			slabsize = P2ROUNDUP(chunksize * chunks,
26107c478bd9Sstevel@tonic-gate 			    vmp->vm_quantum);
26117c478bd9Sstevel@tonic-gate 			/*
26127c478bd9Sstevel@tonic-gate 			 * check for overflow
26137c478bd9Sstevel@tonic-gate 			 */
26147c478bd9Sstevel@tonic-gate 			if ((slabsize / chunks) < chunksize) {
26157c478bd9Sstevel@tonic-gate 				errno = ENOMEM;
26167c478bd9Sstevel@tonic-gate 				goto fail;
26177c478bd9Sstevel@tonic-gate 			}
26187c478bd9Sstevel@tonic-gate 			chunks = slabsize / chunksize;
26197c478bd9Sstevel@tonic-gate 			waste = (slabsize % chunksize) / chunks;
26207c478bd9Sstevel@tonic-gate 			if (waste < minwaste) {
26217c478bd9Sstevel@tonic-gate 				minwaste = waste;
26227c478bd9Sstevel@tonic-gate 				bestfit = slabsize;
26237c478bd9Sstevel@tonic-gate 			}
26247c478bd9Sstevel@tonic-gate 		}
26257c478bd9Sstevel@tonic-gate 		if (cflags & UMC_QCACHE)
26267c478bd9Sstevel@tonic-gate 			bestfit = MAX(1 << highbit(3 * vmp->vm_qcache_max), 64);
26277c478bd9Sstevel@tonic-gate 		cp->cache_slabsize = bestfit;
26287c478bd9Sstevel@tonic-gate 		cp->cache_mincolor = 0;
26297c478bd9Sstevel@tonic-gate 		cp->cache_maxcolor = bestfit % chunksize;
26307c478bd9Sstevel@tonic-gate 		cp->cache_flags |= UMF_HASH;
26317c478bd9Sstevel@tonic-gate 	}
26327c478bd9Sstevel@tonic-gate 
26337c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_HASH) {
26347c478bd9Sstevel@tonic-gate 		ASSERT(!(cflags & UMC_NOHASH));
26357c478bd9Sstevel@tonic-gate 		cp->cache_bufctl_cache = (cp->cache_flags & UMF_AUDIT) ?
26367c478bd9Sstevel@tonic-gate 		    umem_bufctl_audit_cache : umem_bufctl_cache;
26377c478bd9Sstevel@tonic-gate 	}
26387c478bd9Sstevel@tonic-gate 
26397c478bd9Sstevel@tonic-gate 	if (cp->cache_maxcolor >= vmp->vm_quantum)
26407c478bd9Sstevel@tonic-gate 		cp->cache_maxcolor = vmp->vm_quantum - 1;
26417c478bd9Sstevel@tonic-gate 
26427c478bd9Sstevel@tonic-gate 	cp->cache_color = cp->cache_mincolor;
26437c478bd9Sstevel@tonic-gate 
26447c478bd9Sstevel@tonic-gate 	/*
26457c478bd9Sstevel@tonic-gate 	 * Initialize the rest of the slab layer.
26467c478bd9Sstevel@tonic-gate 	 */
26477c478bd9Sstevel@tonic-gate 	(void) mutex_init(&cp->cache_lock, USYNC_THREAD, NULL);
26487c478bd9Sstevel@tonic-gate 
26497c478bd9Sstevel@tonic-gate 	cp->cache_freelist = &cp->cache_nullslab;
26507c478bd9Sstevel@tonic-gate 	cp->cache_nullslab.slab_cache = cp;
26517c478bd9Sstevel@tonic-gate 	cp->cache_nullslab.slab_refcnt = -1;
26527c478bd9Sstevel@tonic-gate 	cp->cache_nullslab.slab_next = &cp->cache_nullslab;
26537c478bd9Sstevel@tonic-gate 	cp->cache_nullslab.slab_prev = &cp->cache_nullslab;
26547c478bd9Sstevel@tonic-gate 
26557c478bd9Sstevel@tonic-gate 	if (cp->cache_flags & UMF_HASH) {
26567c478bd9Sstevel@tonic-gate 		cp->cache_hash_table = vmem_alloc(umem_hash_arena,
26577c478bd9Sstevel@tonic-gate 		    UMEM_HASH_INITIAL * sizeof (void *), VM_NOSLEEP);
26587c478bd9Sstevel@tonic-gate 		if (cp->cache_hash_table == NULL) {
26597c478bd9Sstevel@tonic-gate 			errno = EAGAIN;
26607c478bd9Sstevel@tonic-gate 			goto fail_lock;
26617c478bd9Sstevel@tonic-gate 		}
26627c478bd9Sstevel@tonic-gate 		bzero(cp->cache_hash_table,
26637c478bd9Sstevel@tonic-gate 		    UMEM_HASH_INITIAL * sizeof (void *));
26647c478bd9Sstevel@tonic-gate 		cp->cache_hash_mask = UMEM_HASH_INITIAL - 1;
26657c478bd9Sstevel@tonic-gate 		cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
26667c478bd9Sstevel@tonic-gate 	}
26677c478bd9Sstevel@tonic-gate 
26687c478bd9Sstevel@tonic-gate 	/*
26697c478bd9Sstevel@tonic-gate 	 * Initialize the depot.
26707c478bd9Sstevel@tonic-gate 	 */
26717c478bd9Sstevel@tonic-gate 	(void) mutex_init(&cp->cache_depot_lock, USYNC_THREAD, NULL);
26727c478bd9Sstevel@tonic-gate 
26737c478bd9Sstevel@tonic-gate 	for (mtp = umem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
26747c478bd9Sstevel@tonic-gate 		continue;
26757c478bd9Sstevel@tonic-gate 
26767c478bd9Sstevel@tonic-gate 	cp->cache_magtype = mtp;
26777c478bd9Sstevel@tonic-gate 
26787c478bd9Sstevel@tonic-gate 	/*
26797c478bd9Sstevel@tonic-gate 	 * Initialize the CPU layer.
26807c478bd9Sstevel@tonic-gate 	 */
26817c478bd9Sstevel@tonic-gate 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
26827c478bd9Sstevel@tonic-gate 		umem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
26837c478bd9Sstevel@tonic-gate 		(void) mutex_init(&ccp->cc_lock, USYNC_THREAD, NULL);
26847c478bd9Sstevel@tonic-gate 		ccp->cc_flags = cp->cache_flags;
26857c478bd9Sstevel@tonic-gate 		ccp->cc_rounds = -1;
26867c478bd9Sstevel@tonic-gate 		ccp->cc_prounds = -1;
26877c478bd9Sstevel@tonic-gate 	}
26887c478bd9Sstevel@tonic-gate 
26897c478bd9Sstevel@tonic-gate 	/*
26907c478bd9Sstevel@tonic-gate 	 * Add the cache to the global list.  This makes it visible
26917c478bd9Sstevel@tonic-gate 	 * to umem_update(), so the cache must be ready for business.
26927c478bd9Sstevel@tonic-gate 	 */
26937c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&umem_cache_lock);
26947c478bd9Sstevel@tonic-gate 	cp->cache_next = cnext = &umem_null_cache;
26957c478bd9Sstevel@tonic-gate 	cp->cache_prev = cprev = umem_null_cache.cache_prev;
26967c478bd9Sstevel@tonic-gate 	cnext->cache_prev = cp;
26977c478bd9Sstevel@tonic-gate 	cprev->cache_next = cp;
26987c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&umem_cache_lock);
26997c478bd9Sstevel@tonic-gate 
27007c478bd9Sstevel@tonic-gate 	if (umem_ready == UMEM_READY)
27017c478bd9Sstevel@tonic-gate 		umem_cache_magazine_enable(cp);
27027c478bd9Sstevel@tonic-gate 
27037c478bd9Sstevel@tonic-gate 	return (cp);
27047c478bd9Sstevel@tonic-gate 
27057c478bd9Sstevel@tonic-gate fail_lock:
27067c478bd9Sstevel@tonic-gate 	(void) mutex_destroy(&cp->cache_lock);
27077c478bd9Sstevel@tonic-gate fail:
27087c478bd9Sstevel@tonic-gate 	vmem_xfree(umem_cache_arena, cp, csize);
27097c478bd9Sstevel@tonic-gate 	return (NULL);
27107c478bd9Sstevel@tonic-gate }
27117c478bd9Sstevel@tonic-gate 
27127c478bd9Sstevel@tonic-gate void
27137c478bd9Sstevel@tonic-gate umem_cache_destroy(umem_cache_t *cp)
27147c478bd9Sstevel@tonic-gate {
27157c478bd9Sstevel@tonic-gate 	int cpu_seqid;
27167c478bd9Sstevel@tonic-gate 
27177c478bd9Sstevel@tonic-gate 	/*
27187c478bd9Sstevel@tonic-gate 	 * Remove the cache from the global cache list so that no new updates
27197c478bd9Sstevel@tonic-gate 	 * will be scheduled on its behalf, wait for any pending tasks to
27207c478bd9Sstevel@tonic-gate 	 * complete, purge the cache, and then destroy it.
27217c478bd9Sstevel@tonic-gate 	 */
27227c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&umem_cache_lock);
27237c478bd9Sstevel@tonic-gate 	cp->cache_prev->cache_next = cp->cache_next;
27247c478bd9Sstevel@tonic-gate 	cp->cache_next->cache_prev = cp->cache_prev;
27257c478bd9Sstevel@tonic-gate 	cp->cache_prev = cp->cache_next = NULL;
27267c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&umem_cache_lock);
27277c478bd9Sstevel@tonic-gate 
27287c478bd9Sstevel@tonic-gate 	umem_remove_updates(cp);
27297c478bd9Sstevel@tonic-gate 
27307c478bd9Sstevel@tonic-gate 	umem_cache_magazine_purge(cp);
27317c478bd9Sstevel@tonic-gate 
27327c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&cp->cache_lock);
27337c478bd9Sstevel@tonic-gate 	if (cp->cache_buftotal != 0)
27347c478bd9Sstevel@tonic-gate 		log_message("umem_cache_destroy: '%s' (%p) not empty\n",
27357c478bd9Sstevel@tonic-gate 		    cp->cache_name, (void *)cp);
27367c478bd9Sstevel@tonic-gate 	cp->cache_reclaim = NULL;
27377c478bd9Sstevel@tonic-gate 	/*
27387c478bd9Sstevel@tonic-gate 	 * The cache is now dead.  There should be no further activity.
27397c478bd9Sstevel@tonic-gate 	 * We enforce this by setting land mines in the constructor and
27407c478bd9Sstevel@tonic-gate 	 * destructor routines that induce a segmentation fault if invoked.
27417c478bd9Sstevel@tonic-gate 	 */
27427c478bd9Sstevel@tonic-gate 	cp->cache_constructor = (umem_constructor_t *)1;
27437c478bd9Sstevel@tonic-gate 	cp->cache_destructor = (umem_destructor_t *)2;
27447c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&cp->cache_lock);
27457c478bd9Sstevel@tonic-gate 
27467c478bd9Sstevel@tonic-gate 	if (cp->cache_hash_table != NULL)
27477c478bd9Sstevel@tonic-gate 		vmem_free(umem_hash_arena, cp->cache_hash_table,
27487c478bd9Sstevel@tonic-gate 		    (cp->cache_hash_mask + 1) * sizeof (void *));
27497c478bd9Sstevel@tonic-gate 
27507c478bd9Sstevel@tonic-gate 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++)
27517c478bd9Sstevel@tonic-gate 		(void) mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
27527c478bd9Sstevel@tonic-gate 
27537c478bd9Sstevel@tonic-gate 	(void) mutex_destroy(&cp->cache_depot_lock);
27547c478bd9Sstevel@tonic-gate 	(void) mutex_destroy(&cp->cache_lock);
27557c478bd9Sstevel@tonic-gate 
27567c478bd9Sstevel@tonic-gate 	vmem_free(umem_cache_arena, cp, UMEM_CACHE_SIZE(umem_max_ncpus));
27577c478bd9Sstevel@tonic-gate }
27587c478bd9Sstevel@tonic-gate 
2759789d94c2Sjwadams void
2760789d94c2Sjwadams umem_alloc_sizes_clear(void)
2761789d94c2Sjwadams {
2762789d94c2Sjwadams 	int i;
2763789d94c2Sjwadams 
2764789d94c2Sjwadams 	umem_alloc_sizes[0] = UMEM_MAXBUF;
2765789d94c2Sjwadams 	for (i = 1; i < NUM_ALLOC_SIZES; i++)
2766789d94c2Sjwadams 		umem_alloc_sizes[i] = 0;
2767789d94c2Sjwadams }
2768789d94c2Sjwadams 
2769789d94c2Sjwadams void
2770789d94c2Sjwadams umem_alloc_sizes_add(size_t size_arg)
2771789d94c2Sjwadams {
2772789d94c2Sjwadams 	int i, j;
2773789d94c2Sjwadams 	size_t size = size_arg;
2774789d94c2Sjwadams 
2775789d94c2Sjwadams 	if (size == 0) {
2776789d94c2Sjwadams 		log_message("size_add: cannot add zero-sized cache\n",
2777789d94c2Sjwadams 		    size, UMEM_MAXBUF);
2778789d94c2Sjwadams 		return;
2779789d94c2Sjwadams 	}
2780789d94c2Sjwadams 
2781789d94c2Sjwadams 	if (size > UMEM_MAXBUF) {
2782789d94c2Sjwadams 		log_message("size_add: %ld > %d, cannot add\n", size,
2783789d94c2Sjwadams 		    UMEM_MAXBUF);
2784789d94c2Sjwadams 		return;
2785789d94c2Sjwadams 	}
2786789d94c2Sjwadams 
2787789d94c2Sjwadams 	if (umem_alloc_sizes[NUM_ALLOC_SIZES - 1] != 0) {
2788789d94c2Sjwadams 		log_message("size_add: no space in alloc_table for %d\n",
2789789d94c2Sjwadams 		    size);
2790789d94c2Sjwadams 		return;
2791789d94c2Sjwadams 	}
2792789d94c2Sjwadams 
2793789d94c2Sjwadams 	if (P2PHASE(size, UMEM_ALIGN) != 0) {
2794789d94c2Sjwadams 		size = P2ROUNDUP(size, UMEM_ALIGN);
2795789d94c2Sjwadams 		log_message("size_add: rounding %d up to %d\n", size_arg,
2796789d94c2Sjwadams 		    size);
2797789d94c2Sjwadams 	}
2798789d94c2Sjwadams 
2799789d94c2Sjwadams 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
2800789d94c2Sjwadams 		int cur = umem_alloc_sizes[i];
2801789d94c2Sjwadams 		if (cur == size) {
2802789d94c2Sjwadams 			log_message("size_add: %ld already in table\n",
2803789d94c2Sjwadams 			    size);
2804789d94c2Sjwadams 			return;
2805789d94c2Sjwadams 		}
2806789d94c2Sjwadams 		if (cur > size)
2807789d94c2Sjwadams 			break;
2808789d94c2Sjwadams 	}
2809789d94c2Sjwadams 
2810789d94c2Sjwadams 	for (j = NUM_ALLOC_SIZES - 1; j > i; j--)
2811789d94c2Sjwadams 		umem_alloc_sizes[j] = umem_alloc_sizes[j-1];
2812789d94c2Sjwadams 	umem_alloc_sizes[i] = size;
2813789d94c2Sjwadams }
2814789d94c2Sjwadams 
2815789d94c2Sjwadams void
2816789d94c2Sjwadams umem_alloc_sizes_remove(size_t size)
2817789d94c2Sjwadams {
2818789d94c2Sjwadams 	int i;
2819789d94c2Sjwadams 
2820789d94c2Sjwadams 	if (size == UMEM_MAXBUF) {
2821789d94c2Sjwadams 		log_message("size_remove: cannot remove %ld\n", size);
2822789d94c2Sjwadams 		return;
2823789d94c2Sjwadams 	}
2824789d94c2Sjwadams 
2825789d94c2Sjwadams 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
2826789d94c2Sjwadams 		int cur = umem_alloc_sizes[i];
2827789d94c2Sjwadams 		if (cur == size)
2828789d94c2Sjwadams 			break;
2829789d94c2Sjwadams 		else if (cur > size || cur == 0) {
2830789d94c2Sjwadams 			log_message("size_remove: %ld not found in table\n",
2831789d94c2Sjwadams 			    size);
2832789d94c2Sjwadams 			return;
2833789d94c2Sjwadams 		}
2834789d94c2Sjwadams 	}
2835789d94c2Sjwadams 
2836789d94c2Sjwadams 	for (; i + 1 < NUM_ALLOC_SIZES; i++)
2837789d94c2Sjwadams 		umem_alloc_sizes[i] = umem_alloc_sizes[i+1];
2838789d94c2Sjwadams 	umem_alloc_sizes[i] = 0;
2839789d94c2Sjwadams }
2840789d94c2Sjwadams 
28417c478bd9Sstevel@tonic-gate static int
28427c478bd9Sstevel@tonic-gate umem_cache_init(void)
28437c478bd9Sstevel@tonic-gate {
28447c478bd9Sstevel@tonic-gate 	int i;
28457c478bd9Sstevel@tonic-gate 	size_t size, max_size;
28467c478bd9Sstevel@tonic-gate 	umem_cache_t *cp;
28477c478bd9Sstevel@tonic-gate 	umem_magtype_t *mtp;
28487c478bd9Sstevel@tonic-gate 	char name[UMEM_CACHE_NAMELEN + 1];
28497c478bd9Sstevel@tonic-gate 	umem_cache_t *umem_alloc_caches[NUM_ALLOC_SIZES];
28507c478bd9Sstevel@tonic-gate 
28517c478bd9Sstevel@tonic-gate 	for (i = 0; i < sizeof (umem_magtype) / sizeof (*mtp); i++) {
28527c478bd9Sstevel@tonic-gate 		mtp = &umem_magtype[i];
28537c478bd9Sstevel@tonic-gate 		(void) snprintf(name, sizeof (name), "umem_magazine_%d",
28547c478bd9Sstevel@tonic-gate 		    mtp->mt_magsize);
28557c478bd9Sstevel@tonic-gate 		mtp->mt_cache = umem_cache_create(name,
28567c478bd9Sstevel@tonic-gate 		    (mtp->mt_magsize + 1) * sizeof (void *),
28577c478bd9Sstevel@tonic-gate 		    mtp->mt_align, NULL, NULL, NULL, NULL,
28587c478bd9Sstevel@tonic-gate 		    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
28597c478bd9Sstevel@tonic-gate 		if (mtp->mt_cache == NULL)
28607c478bd9Sstevel@tonic-gate 			return (0);
28617c478bd9Sstevel@tonic-gate 	}
28627c478bd9Sstevel@tonic-gate 
28637c478bd9Sstevel@tonic-gate 	umem_slab_cache = umem_cache_create("umem_slab_cache",
28647c478bd9Sstevel@tonic-gate 	    sizeof (umem_slab_t), 0, NULL, NULL, NULL, NULL,
28657c478bd9Sstevel@tonic-gate 	    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
28667c478bd9Sstevel@tonic-gate 
28677c478bd9Sstevel@tonic-gate 	if (umem_slab_cache == NULL)
28687c478bd9Sstevel@tonic-gate 		return (0);
28697c478bd9Sstevel@tonic-gate 
28707c478bd9Sstevel@tonic-gate 	umem_bufctl_cache = umem_cache_create("umem_bufctl_cache",
28717c478bd9Sstevel@tonic-gate 	    sizeof (umem_bufctl_t), 0, NULL, NULL, NULL, NULL,
28727c478bd9Sstevel@tonic-gate 	    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
28737c478bd9Sstevel@tonic-gate 
28747c478bd9Sstevel@tonic-gate 	if (umem_bufctl_cache == NULL)
28757c478bd9Sstevel@tonic-gate 		return (0);
28767c478bd9Sstevel@tonic-gate 
28777c478bd9Sstevel@tonic-gate 	/*
28787c478bd9Sstevel@tonic-gate 	 * The size of the umem_bufctl_audit structure depends upon
28797c478bd9Sstevel@tonic-gate 	 * umem_stack_depth.   See umem_impl.h for details on the size
28807c478bd9Sstevel@tonic-gate 	 * restrictions.
28817c478bd9Sstevel@tonic-gate 	 */
28827c478bd9Sstevel@tonic-gate 
28837c478bd9Sstevel@tonic-gate 	size = UMEM_BUFCTL_AUDIT_SIZE_DEPTH(umem_stack_depth);
28847c478bd9Sstevel@tonic-gate 	max_size = UMEM_BUFCTL_AUDIT_MAX_SIZE;
28857c478bd9Sstevel@tonic-gate 
28867c478bd9Sstevel@tonic-gate 	if (size > max_size) {			/* too large -- truncate */
28877c478bd9Sstevel@tonic-gate 		int max_frames = UMEM_MAX_STACK_DEPTH;
28887c478bd9Sstevel@tonic-gate 
28897c478bd9Sstevel@tonic-gate 		ASSERT(UMEM_BUFCTL_AUDIT_SIZE_DEPTH(max_frames) <= max_size);
28907c478bd9Sstevel@tonic-gate 
28917c478bd9Sstevel@tonic-gate 		umem_stack_depth = max_frames;
28927c478bd9Sstevel@tonic-gate 		size = UMEM_BUFCTL_AUDIT_SIZE_DEPTH(umem_stack_depth);
28937c478bd9Sstevel@tonic-gate 	}
28947c478bd9Sstevel@tonic-gate 
28957c478bd9Sstevel@tonic-gate 	umem_bufctl_audit_cache = umem_cache_create("umem_bufctl_audit_cache",
28967c478bd9Sstevel@tonic-gate 	    size, 0, NULL, NULL, NULL, NULL, umem_internal_arena,
28977c478bd9Sstevel@tonic-gate 	    UMC_NOHASH | UMC_INTERNAL);
28987c478bd9Sstevel@tonic-gate 
28997c478bd9Sstevel@tonic-gate 	if (umem_bufctl_audit_cache == NULL)
29007c478bd9Sstevel@tonic-gate 		return (0);
29017c478bd9Sstevel@tonic-gate 
29027c478bd9Sstevel@tonic-gate 	if (vmem_backend & VMEM_BACKEND_MMAP)
29037c478bd9Sstevel@tonic-gate 		umem_va_arena = vmem_create("umem_va",
29047c478bd9Sstevel@tonic-gate 		    NULL, 0, pagesize,
29057c478bd9Sstevel@tonic-gate 		    vmem_alloc, vmem_free, heap_arena,
29067c478bd9Sstevel@tonic-gate 		    8 * pagesize, VM_NOSLEEP);
29077c478bd9Sstevel@tonic-gate 	else
29087c478bd9Sstevel@tonic-gate 		umem_va_arena = heap_arena;
29097c478bd9Sstevel@tonic-gate 
29107c478bd9Sstevel@tonic-gate 	if (umem_va_arena == NULL)
29117c478bd9Sstevel@tonic-gate 		return (0);
29127c478bd9Sstevel@tonic-gate 
29137c478bd9Sstevel@tonic-gate 	umem_default_arena = vmem_create("umem_default",
29147c478bd9Sstevel@tonic-gate 	    NULL, 0, pagesize,
29157c478bd9Sstevel@tonic-gate 	    heap_alloc, heap_free, umem_va_arena,
29167c478bd9Sstevel@tonic-gate 	    0, VM_NOSLEEP);
29177c478bd9Sstevel@tonic-gate 
29187c478bd9Sstevel@tonic-gate 	if (umem_default_arena == NULL)
29197c478bd9Sstevel@tonic-gate 		return (0);
29207c478bd9Sstevel@tonic-gate 
29217c478bd9Sstevel@tonic-gate 	/*
29227c478bd9Sstevel@tonic-gate 	 * make sure the umem_alloc table initializer is correct
29237c478bd9Sstevel@tonic-gate 	 */
29247c478bd9Sstevel@tonic-gate 	i = sizeof (umem_alloc_table) / sizeof (*umem_alloc_table);
29257c478bd9Sstevel@tonic-gate 	ASSERT(umem_alloc_table[i - 1] == &umem_null_cache);
29267c478bd9Sstevel@tonic-gate 
29277c478bd9Sstevel@tonic-gate 	/*
29287c478bd9Sstevel@tonic-gate 	 * Create the default caches to back umem_alloc()
29297c478bd9Sstevel@tonic-gate 	 */
29307c478bd9Sstevel@tonic-gate 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
29317c478bd9Sstevel@tonic-gate 		size_t cache_size = umem_alloc_sizes[i];
29327c478bd9Sstevel@tonic-gate 		size_t align = 0;
2933789d94c2Sjwadams 
2934789d94c2Sjwadams 		if (cache_size == 0)
2935789d94c2Sjwadams 			break;		/* 0 terminates the list */
2936789d94c2Sjwadams 
29377c478bd9Sstevel@tonic-gate 		/*
29387c478bd9Sstevel@tonic-gate 		 * If they allocate a multiple of the coherency granularity,
29397c478bd9Sstevel@tonic-gate 		 * they get a coherency-granularity-aligned address.
29407c478bd9Sstevel@tonic-gate 		 */
29417c478bd9Sstevel@tonic-gate 		if (IS_P2ALIGNED(cache_size, 64))
29427c478bd9Sstevel@tonic-gate 			align = 64;
29437c478bd9Sstevel@tonic-gate 		if (IS_P2ALIGNED(cache_size, pagesize))
29447c478bd9Sstevel@tonic-gate 			align = pagesize;
29457c478bd9Sstevel@tonic-gate 		(void) snprintf(name, sizeof (name), "umem_alloc_%lu",
29467c478bd9Sstevel@tonic-gate 		    (long)cache_size);
29477c478bd9Sstevel@tonic-gate 
29487c478bd9Sstevel@tonic-gate 		cp = umem_cache_create(name, cache_size, align,
29497c478bd9Sstevel@tonic-gate 		    NULL, NULL, NULL, NULL, NULL, UMC_INTERNAL);
29507c478bd9Sstevel@tonic-gate 		if (cp == NULL)
29517c478bd9Sstevel@tonic-gate 			return (0);
29527c478bd9Sstevel@tonic-gate 
29537c478bd9Sstevel@tonic-gate 		umem_alloc_caches[i] = cp;
29547c478bd9Sstevel@tonic-gate 	}
29557c478bd9Sstevel@tonic-gate 
29567c478bd9Sstevel@tonic-gate 	/*
29577c478bd9Sstevel@tonic-gate 	 * Initialization cannot fail at this point.  Make the caches
29587c478bd9Sstevel@tonic-gate 	 * visible to umem_alloc() and friends.
29597c478bd9Sstevel@tonic-gate 	 */
29607c478bd9Sstevel@tonic-gate 	size = UMEM_ALIGN;
29617c478bd9Sstevel@tonic-gate 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
29627c478bd9Sstevel@tonic-gate 		size_t cache_size = umem_alloc_sizes[i];
29637c478bd9Sstevel@tonic-gate 
2964789d94c2Sjwadams 		if (cache_size == 0)
2965789d94c2Sjwadams 			break;		/* 0 terminates the list */
2966789d94c2Sjwadams 
29677c478bd9Sstevel@tonic-gate 		cp = umem_alloc_caches[i];
29687c478bd9Sstevel@tonic-gate 
29697c478bd9Sstevel@tonic-gate 		while (size <= cache_size) {
29707c478bd9Sstevel@tonic-gate 			umem_alloc_table[(size - 1) >> UMEM_ALIGN_SHIFT] = cp;
29717c478bd9Sstevel@tonic-gate 			size += UMEM_ALIGN;
29727c478bd9Sstevel@tonic-gate 		}
29737c478bd9Sstevel@tonic-gate 	}
2974789d94c2Sjwadams 	ASSERT(size - UMEM_ALIGN == UMEM_MAXBUF);
29757c478bd9Sstevel@tonic-gate 	return (1);
29767c478bd9Sstevel@tonic-gate }
29777c478bd9Sstevel@tonic-gate 
29787c478bd9Sstevel@tonic-gate /*
29797c478bd9Sstevel@tonic-gate  * umem_startup() is called early on, and must be called explicitly if we're
29807c478bd9Sstevel@tonic-gate  * the standalone version.
29817c478bd9Sstevel@tonic-gate  */
29827c478bd9Sstevel@tonic-gate #ifdef UMEM_STANDALONE
29837c478bd9Sstevel@tonic-gate void
29847c478bd9Sstevel@tonic-gate #else
29857c478bd9Sstevel@tonic-gate #pragma init(umem_startup)
29867c478bd9Sstevel@tonic-gate static void
29877c478bd9Sstevel@tonic-gate #endif
29887c478bd9Sstevel@tonic-gate umem_startup(caddr_t start, size_t len, size_t pagesize, caddr_t minstack,
29897c478bd9Sstevel@tonic-gate     caddr_t maxstack)
29907c478bd9Sstevel@tonic-gate {
29917c478bd9Sstevel@tonic-gate #ifdef UMEM_STANDALONE
29927c478bd9Sstevel@tonic-gate 	int idx;
29937c478bd9Sstevel@tonic-gate 	/* Standalone doesn't fork */
29947c478bd9Sstevel@tonic-gate #else
29957c478bd9Sstevel@tonic-gate 	umem_forkhandler_init(); /* register the fork handler */
29967c478bd9Sstevel@tonic-gate #endif
29977c478bd9Sstevel@tonic-gate 
29987c478bd9Sstevel@tonic-gate #ifdef __lint
29997c478bd9Sstevel@tonic-gate 	/* make lint happy */
30007c478bd9Sstevel@tonic-gate 	minstack = maxstack;
30017c478bd9Sstevel@tonic-gate #endif
30027c478bd9Sstevel@tonic-gate 
30037c478bd9Sstevel@tonic-gate #ifdef UMEM_STANDALONE
30047c478bd9Sstevel@tonic-gate 	umem_ready = UMEM_READY_STARTUP;
30057c478bd9Sstevel@tonic-gate 	umem_init_env_ready = 0;
30067c478bd9Sstevel@tonic-gate 
30077c478bd9Sstevel@tonic-gate 	umem_min_stack = minstack;
30087c478bd9Sstevel@tonic-gate 	umem_max_stack = maxstack;
30097c478bd9Sstevel@tonic-gate 
30107c478bd9Sstevel@tonic-gate 	nofail_callback = NULL;
30117c478bd9Sstevel@tonic-gate 	umem_slab_cache = NULL;
30127c478bd9Sstevel@tonic-gate 	umem_bufctl_cache = NULL;
30137c478bd9Sstevel@tonic-gate 	umem_bufctl_audit_cache = NULL;
30147c478bd9Sstevel@tonic-gate 	heap_arena = NULL;
30157c478bd9Sstevel@tonic-gate 	heap_alloc = NULL;
30167c478bd9Sstevel@tonic-gate 	heap_free = NULL;
30177c478bd9Sstevel@tonic-gate 	umem_internal_arena = NULL;
30187c478bd9Sstevel@tonic-gate 	umem_cache_arena = NULL;
30197c478bd9Sstevel@tonic-gate 	umem_hash_arena = NULL;
30207c478bd9Sstevel@tonic-gate 	umem_log_arena = NULL;
30217c478bd9Sstevel@tonic-gate 	umem_oversize_arena = NULL;
30227c478bd9Sstevel@tonic-gate 	umem_va_arena = NULL;
30237c478bd9Sstevel@tonic-gate 	umem_default_arena = NULL;
30247c478bd9Sstevel@tonic-gate 	umem_firewall_va_arena = NULL;
30257c478bd9Sstevel@tonic-gate 	umem_firewall_arena = NULL;
30267c478bd9Sstevel@tonic-gate 	umem_memalign_arena = NULL;
30277c478bd9Sstevel@tonic-gate 	umem_transaction_log = NULL;
30287c478bd9Sstevel@tonic-gate 	umem_content_log = NULL;
30297c478bd9Sstevel@tonic-gate 	umem_failure_log = NULL;
30307c478bd9Sstevel@tonic-gate 	umem_slab_log = NULL;
30317c478bd9Sstevel@tonic-gate 	umem_cpu_mask = 0;
30327c478bd9Sstevel@tonic-gate 
30337c478bd9Sstevel@tonic-gate 	umem_cpus = &umem_startup_cpu;
30347c478bd9Sstevel@tonic-gate 	umem_startup_cpu.cpu_cache_offset = UMEM_CACHE_SIZE(0);
30357c478bd9Sstevel@tonic-gate 	umem_startup_cpu.cpu_number = 0;
30367c478bd9Sstevel@tonic-gate 
30377c478bd9Sstevel@tonic-gate 	bcopy(&umem_null_cache_template, &umem_null_cache,
30387c478bd9Sstevel@tonic-gate 	    sizeof (umem_cache_t));
30397c478bd9Sstevel@tonic-gate 
30407c478bd9Sstevel@tonic-gate 	for (idx = 0; idx < (UMEM_MAXBUF >> UMEM_ALIGN_SHIFT); idx++)
30417c478bd9Sstevel@tonic-gate 		umem_alloc_table[idx] = &umem_null_cache;
30427c478bd9Sstevel@tonic-gate #endif
30437c478bd9Sstevel@tonic-gate 
30447c478bd9Sstevel@tonic-gate 	/*
30457c478bd9Sstevel@tonic-gate 	 * Perform initialization specific to the way we've been compiled
30467c478bd9Sstevel@tonic-gate 	 * (library or standalone)
30477c478bd9Sstevel@tonic-gate 	 */
30487c478bd9Sstevel@tonic-gate 	umem_type_init(start, len, pagesize);
30497c478bd9Sstevel@tonic-gate 
30507c478bd9Sstevel@tonic-gate 	vmem_startup();
30517c478bd9Sstevel@tonic-gate }
30527c478bd9Sstevel@tonic-gate 
30537c478bd9Sstevel@tonic-gate int
30547c478bd9Sstevel@tonic-gate umem_init(void)
30557c478bd9Sstevel@tonic-gate {
30567c478bd9Sstevel@tonic-gate 	size_t maxverify, minfirewall;
30577c478bd9Sstevel@tonic-gate 	size_t size;
30587c478bd9Sstevel@tonic-gate 	int idx;
30597c478bd9Sstevel@tonic-gate 	umem_cpu_t *new_cpus;
30607c478bd9Sstevel@tonic-gate 
30617c478bd9Sstevel@tonic-gate 	vmem_t *memalign_arena, *oversize_arena;
30627c478bd9Sstevel@tonic-gate 
30637c478bd9Sstevel@tonic-gate 	if (thr_self() != umem_init_thr) {
30647c478bd9Sstevel@tonic-gate 		/*
30657c478bd9Sstevel@tonic-gate 		 * The usual case -- non-recursive invocation of umem_init().
30667c478bd9Sstevel@tonic-gate 		 */
30677c478bd9Sstevel@tonic-gate 		(void) mutex_lock(&umem_init_lock);
30687c478bd9Sstevel@tonic-gate 		if (umem_ready != UMEM_READY_STARTUP) {
30697c478bd9Sstevel@tonic-gate 			/*
30707c478bd9Sstevel@tonic-gate 			 * someone else beat us to initializing umem.  Wait
30717c478bd9Sstevel@tonic-gate 			 * for them to complete, then return.
30727c478bd9Sstevel@tonic-gate 			 */
3073a574db85Sraf 			while (umem_ready == UMEM_READY_INITING) {
3074a574db85Sraf 				int cancel_state;
3075a574db85Sraf 
3076a574db85Sraf 				(void) pthread_setcancelstate(
3077a574db85Sraf 				    PTHREAD_CANCEL_DISABLE, &cancel_state);
3078a574db85Sraf 				(void) cond_wait(&umem_init_cv,
30797c478bd9Sstevel@tonic-gate 				    &umem_init_lock);
3080a574db85Sraf 				(void) pthread_setcancelstate(
3081a574db85Sraf 				    cancel_state, NULL);
3082a574db85Sraf 			}
30837c478bd9Sstevel@tonic-gate 			ASSERT(umem_ready == UMEM_READY ||
30847c478bd9Sstevel@tonic-gate 			    umem_ready == UMEM_READY_INIT_FAILED);
30857c478bd9Sstevel@tonic-gate 			(void) mutex_unlock(&umem_init_lock);
30867c478bd9Sstevel@tonic-gate 			return (umem_ready == UMEM_READY);
30877c478bd9Sstevel@tonic-gate 		}
30887c478bd9Sstevel@tonic-gate 
30897c478bd9Sstevel@tonic-gate 		ASSERT(umem_ready == UMEM_READY_STARTUP);
30907c478bd9Sstevel@tonic-gate 		ASSERT(umem_init_env_ready == 0);
30917c478bd9Sstevel@tonic-gate 
30927c478bd9Sstevel@tonic-gate 		umem_ready = UMEM_READY_INITING;
30937c478bd9Sstevel@tonic-gate 		umem_init_thr = thr_self();
30947c478bd9Sstevel@tonic-gate 
30957c478bd9Sstevel@tonic-gate 		(void) mutex_unlock(&umem_init_lock);
30967c478bd9Sstevel@tonic-gate 		umem_setup_envvars(0);		/* can recurse -- see below */
30977c478bd9Sstevel@tonic-gate 		if (umem_init_env_ready) {
30987c478bd9Sstevel@tonic-gate 			/*
30997c478bd9Sstevel@tonic-gate 			 * initialization was completed already
31007c478bd9Sstevel@tonic-gate 			 */
31017c478bd9Sstevel@tonic-gate 			ASSERT(umem_ready == UMEM_READY ||
31027c478bd9Sstevel@tonic-gate 			    umem_ready == UMEM_READY_INIT_FAILED);
31037c478bd9Sstevel@tonic-gate 			ASSERT(umem_init_thr == 0);
31047c478bd9Sstevel@tonic-gate 			return (umem_ready == UMEM_READY);
31057c478bd9Sstevel@tonic-gate 		}
31067c478bd9Sstevel@tonic-gate 	} else if (!umem_init_env_ready) {
31077c478bd9Sstevel@tonic-gate 		/*
31087c478bd9Sstevel@tonic-gate 		 * The umem_setup_envvars() call (above) makes calls into
31097c478bd9Sstevel@tonic-gate 		 * the dynamic linker and directly into user-supplied code.
31107c478bd9Sstevel@tonic-gate 		 * Since we cannot know what that code will do, we could be
31117c478bd9Sstevel@tonic-gate 		 * recursively invoked (by, say, a malloc() call in the code
31127c478bd9Sstevel@tonic-gate 		 * itself, or in a (C++) _init section it causes to be fired).
31137c478bd9Sstevel@tonic-gate 		 *
31147c478bd9Sstevel@tonic-gate 		 * This code is where we end up if such recursion occurs.  We
31157c478bd9Sstevel@tonic-gate 		 * first clean up any partial results in the envvar code, then
31167c478bd9Sstevel@tonic-gate 		 * proceed to finish initialization processing in the recursive
31177c478bd9Sstevel@tonic-gate 		 * call.  The original call will notice this, and return
31187c478bd9Sstevel@tonic-gate 		 * immediately.
31197c478bd9Sstevel@tonic-gate 		 */
31207c478bd9Sstevel@tonic-gate 		umem_setup_envvars(1);		/* clean up any partial state */
31217c478bd9Sstevel@tonic-gate 	} else {
31227c478bd9Sstevel@tonic-gate 		umem_panic(
31237c478bd9Sstevel@tonic-gate 		    "recursive allocation while initializing umem\n");
31247c478bd9Sstevel@tonic-gate 	}
31257c478bd9Sstevel@tonic-gate 	umem_init_env_ready = 1;
31267c478bd9Sstevel@tonic-gate 
31277c478bd9Sstevel@tonic-gate 	/*
31287c478bd9Sstevel@tonic-gate 	 * From this point until we finish, recursion into umem_init() will
31297c478bd9Sstevel@tonic-gate 	 * cause a umem_panic().
31307c478bd9Sstevel@tonic-gate 	 */
31317c478bd9Sstevel@tonic-gate 	maxverify = minfirewall = ULONG_MAX;
31327c478bd9Sstevel@tonic-gate 
31337c478bd9Sstevel@tonic-gate 	/* LINTED constant condition */
31347c478bd9Sstevel@tonic-gate 	if (sizeof (umem_cpu_cache_t) != UMEM_CPU_CACHE_SIZE) {
31357c478bd9Sstevel@tonic-gate 		umem_panic("sizeof (umem_cpu_cache_t) = %d, should be %d\n",
31367c478bd9Sstevel@tonic-gate 		    sizeof (umem_cpu_cache_t), UMEM_CPU_CACHE_SIZE);
31377c478bd9Sstevel@tonic-gate 	}
31387c478bd9Sstevel@tonic-gate 
31397c478bd9Sstevel@tonic-gate 	umem_max_ncpus = umem_get_max_ncpus();
31407c478bd9Sstevel@tonic-gate 
31417c478bd9Sstevel@tonic-gate 	/*
31427c478bd9Sstevel@tonic-gate 	 * load tunables from environment
31437c478bd9Sstevel@tonic-gate 	 */
31447c478bd9Sstevel@tonic-gate 	umem_process_envvars();
31457c478bd9Sstevel@tonic-gate 
31467c478bd9Sstevel@tonic-gate 	if (issetugid())
31477c478bd9Sstevel@tonic-gate 		umem_mtbf = 0;
31487c478bd9Sstevel@tonic-gate 
31497c478bd9Sstevel@tonic-gate 	/*
31507c478bd9Sstevel@tonic-gate 	 * set up vmem
31517c478bd9Sstevel@tonic-gate 	 */
31527c478bd9Sstevel@tonic-gate 	if (!(umem_flags & UMF_AUDIT))
31537c478bd9Sstevel@tonic-gate 		vmem_no_debug();
31547c478bd9Sstevel@tonic-gate 
31557c478bd9Sstevel@tonic-gate 	heap_arena = vmem_heap_arena(&heap_alloc, &heap_free);
31567c478bd9Sstevel@tonic-gate 
31577c478bd9Sstevel@tonic-gate 	pagesize = heap_arena->vm_quantum;
31587c478bd9Sstevel@tonic-gate 
31597c478bd9Sstevel@tonic-gate 	umem_internal_arena = vmem_create("umem_internal", NULL, 0, pagesize,
31607c478bd9Sstevel@tonic-gate 	    heap_alloc, heap_free, heap_arena, 0, VM_NOSLEEP);
31617c478bd9Sstevel@tonic-gate 
31627c478bd9Sstevel@tonic-gate 	umem_default_arena = umem_internal_arena;
31637c478bd9Sstevel@tonic-gate 
31647c478bd9Sstevel@tonic-gate 	if (umem_internal_arena == NULL)
31657c478bd9Sstevel@tonic-gate 		goto fail;
31667c478bd9Sstevel@tonic-gate 
31677c478bd9Sstevel@tonic-gate 	umem_cache_arena = vmem_create("umem_cache", NULL, 0, UMEM_ALIGN,
31687c478bd9Sstevel@tonic-gate 	    vmem_alloc, vmem_free, umem_internal_arena, 0, VM_NOSLEEP);
31697c478bd9Sstevel@tonic-gate 
31707c478bd9Sstevel@tonic-gate 	umem_hash_arena = vmem_create("umem_hash", NULL, 0, UMEM_ALIGN,
31717c478bd9Sstevel@tonic-gate 	    vmem_alloc, vmem_free, umem_internal_arena, 0, VM_NOSLEEP);
31727c478bd9Sstevel@tonic-gate 
31737c478bd9Sstevel@tonic-gate 	umem_log_arena = vmem_create("umem_log", NULL, 0, UMEM_ALIGN,
31747c478bd9Sstevel@tonic-gate 	    heap_alloc, heap_free, heap_arena, 0, VM_NOSLEEP);
31757c478bd9Sstevel@tonic-gate 
31767c478bd9Sstevel@tonic-gate 	umem_firewall_va_arena = vmem_create("umem_firewall_va",
31777c478bd9Sstevel@tonic-gate 	    NULL, 0, pagesize,
31787c478bd9Sstevel@tonic-gate 	    umem_firewall_va_alloc, umem_firewall_va_free, heap_arena,
31797c478bd9Sstevel@tonic-gate 	    0, VM_NOSLEEP);
31807c478bd9Sstevel@tonic-gate 
31817c478bd9Sstevel@tonic-gate 	if (umem_cache_arena == NULL || umem_hash_arena == NULL ||
31827c478bd9Sstevel@tonic-gate 	    umem_log_arena == NULL || umem_firewall_va_arena == NULL)
31837c478bd9Sstevel@tonic-gate 		goto fail;
31847c478bd9Sstevel@tonic-gate 
31857c478bd9Sstevel@tonic-gate 	umem_firewall_arena = vmem_create("umem_firewall", NULL, 0, pagesize,
31867c478bd9Sstevel@tonic-gate 	    heap_alloc, heap_free, umem_firewall_va_arena, 0,
31877c478bd9Sstevel@tonic-gate 	    VM_NOSLEEP);
31887c478bd9Sstevel@tonic-gate 
31897c478bd9Sstevel@tonic-gate 	if (umem_firewall_arena == NULL)
31907c478bd9Sstevel@tonic-gate 		goto fail;
31917c478bd9Sstevel@tonic-gate 
31927c478bd9Sstevel@tonic-gate 	oversize_arena = vmem_create("umem_oversize", NULL, 0, pagesize,
31937c478bd9Sstevel@tonic-gate 	    heap_alloc, heap_free, minfirewall < ULONG_MAX ?
31947c478bd9Sstevel@tonic-gate 	    umem_firewall_va_arena : heap_arena, 0, VM_NOSLEEP);
31957c478bd9Sstevel@tonic-gate 
31967c478bd9Sstevel@tonic-gate 	memalign_arena = vmem_create("umem_memalign", NULL, 0, UMEM_ALIGN,
31977c478bd9Sstevel@tonic-gate 	    heap_alloc, heap_free, minfirewall < ULONG_MAX ?
31987c478bd9Sstevel@tonic-gate 	    umem_firewall_va_arena : heap_arena, 0, VM_NOSLEEP);
31997c478bd9Sstevel@tonic-gate 
32007c478bd9Sstevel@tonic-gate 	if (oversize_arena == NULL || memalign_arena == NULL)
32017c478bd9Sstevel@tonic-gate 		goto fail;
32027c478bd9Sstevel@tonic-gate 
32037c478bd9Sstevel@tonic-gate 	if (umem_max_ncpus > CPUHINT_MAX())
32047c478bd9Sstevel@tonic-gate 		umem_max_ncpus = CPUHINT_MAX();
32057c478bd9Sstevel@tonic-gate 
32067c478bd9Sstevel@tonic-gate 	while ((umem_max_ncpus & (umem_max_ncpus - 1)) != 0)
32077c478bd9Sstevel@tonic-gate 		umem_max_ncpus++;
32087c478bd9Sstevel@tonic-gate 
32097c478bd9Sstevel@tonic-gate 	if (umem_max_ncpus == 0)
32107c478bd9Sstevel@tonic-gate 		umem_max_ncpus = 1;
32117c478bd9Sstevel@tonic-gate 
32127c478bd9Sstevel@tonic-gate 	size = umem_max_ncpus * sizeof (umem_cpu_t);
32137c478bd9Sstevel@tonic-gate 	new_cpus = vmem_alloc(umem_internal_arena, size, VM_NOSLEEP);
32147c478bd9Sstevel@tonic-gate 	if (new_cpus == NULL)
32157c478bd9Sstevel@tonic-gate 		goto fail;
32167c478bd9Sstevel@tonic-gate 
32177c478bd9Sstevel@tonic-gate 	bzero(new_cpus, size);
32187c478bd9Sstevel@tonic-gate 	for (idx = 0; idx < umem_max_ncpus; idx++) {
32197c478bd9Sstevel@tonic-gate 		new_cpus[idx].cpu_number = idx;
32207c478bd9Sstevel@tonic-gate 		new_cpus[idx].cpu_cache_offset = UMEM_CACHE_SIZE(idx);
32217c478bd9Sstevel@tonic-gate 	}
32227c478bd9Sstevel@tonic-gate 	umem_cpus = new_cpus;
32237c478bd9Sstevel@tonic-gate 	umem_cpu_mask = (umem_max_ncpus - 1);
32247c478bd9Sstevel@tonic-gate 
32257c478bd9Sstevel@tonic-gate 	if (umem_maxverify == 0)
32267c478bd9Sstevel@tonic-gate 		umem_maxverify = maxverify;
32277c478bd9Sstevel@tonic-gate 
32287c478bd9Sstevel@tonic-gate 	if (umem_minfirewall == 0)
32297c478bd9Sstevel@tonic-gate 		umem_minfirewall = minfirewall;
32307c478bd9Sstevel@tonic-gate 
32317c478bd9Sstevel@tonic-gate 	/*
32327c478bd9Sstevel@tonic-gate 	 * Set up updating and reaping
32337c478bd9Sstevel@tonic-gate 	 */
32347c478bd9Sstevel@tonic-gate 	umem_reap_next = gethrtime() + NANOSEC;
32357c478bd9Sstevel@tonic-gate 
32367c478bd9Sstevel@tonic-gate #ifndef UMEM_STANDALONE
32377c478bd9Sstevel@tonic-gate 	(void) gettimeofday(&umem_update_next, NULL);
32387c478bd9Sstevel@tonic-gate #endif
32397c478bd9Sstevel@tonic-gate 
32407c478bd9Sstevel@tonic-gate 	/*
32417c478bd9Sstevel@tonic-gate 	 * Set up logging -- failure here is okay, since it will just disable
32427c478bd9Sstevel@tonic-gate 	 * the logs
32437c478bd9Sstevel@tonic-gate 	 */
32447c478bd9Sstevel@tonic-gate 	if (umem_logging) {
32457c478bd9Sstevel@tonic-gate 		umem_transaction_log = umem_log_init(umem_transaction_log_size);
32467c478bd9Sstevel@tonic-gate 		umem_content_log = umem_log_init(umem_content_log_size);
32477c478bd9Sstevel@tonic-gate 		umem_failure_log = umem_log_init(umem_failure_log_size);
32487c478bd9Sstevel@tonic-gate 		umem_slab_log = umem_log_init(umem_slab_log_size);
32497c478bd9Sstevel@tonic-gate 	}
32507c478bd9Sstevel@tonic-gate 
32517c478bd9Sstevel@tonic-gate 	/*
32527c478bd9Sstevel@tonic-gate 	 * Set up caches -- if successful, initialization cannot fail, since
32537c478bd9Sstevel@tonic-gate 	 * allocations from other threads can now succeed.
32547c478bd9Sstevel@tonic-gate 	 */
32557c478bd9Sstevel@tonic-gate 	if (umem_cache_init() == 0) {
32567c478bd9Sstevel@tonic-gate 		log_message("unable to create initial caches\n");
32577c478bd9Sstevel@tonic-gate 		goto fail;
32587c478bd9Sstevel@tonic-gate 	}
32597c478bd9Sstevel@tonic-gate 	umem_oversize_arena = oversize_arena;
32607c478bd9Sstevel@tonic-gate 	umem_memalign_arena = memalign_arena;
32617c478bd9Sstevel@tonic-gate 
32627c478bd9Sstevel@tonic-gate 	umem_cache_applyall(umem_cache_magazine_enable);
32637c478bd9Sstevel@tonic-gate 
32647c478bd9Sstevel@tonic-gate 	/*
32657c478bd9Sstevel@tonic-gate 	 * initialization done, ready to go
32667c478bd9Sstevel@tonic-gate 	 */
32677c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&umem_init_lock);
32687c478bd9Sstevel@tonic-gate 	umem_ready = UMEM_READY;
32697c478bd9Sstevel@tonic-gate 	umem_init_thr = 0;
32707c478bd9Sstevel@tonic-gate 	(void) cond_broadcast(&umem_init_cv);
32717c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&umem_init_lock);
32727c478bd9Sstevel@tonic-gate 	return (1);
32737c478bd9Sstevel@tonic-gate 
32747c478bd9Sstevel@tonic-gate fail:
32757c478bd9Sstevel@tonic-gate 	log_message("umem initialization failed\n");
32767c478bd9Sstevel@tonic-gate 
32777c478bd9Sstevel@tonic-gate 	(void) mutex_lock(&umem_init_lock);
32787c478bd9Sstevel@tonic-gate 	umem_ready = UMEM_READY_INIT_FAILED;
32797c478bd9Sstevel@tonic-gate 	umem_init_thr = 0;
32807c478bd9Sstevel@tonic-gate 	(void) cond_broadcast(&umem_init_cv);
32817c478bd9Sstevel@tonic-gate 	(void) mutex_unlock(&umem_init_lock);
32827c478bd9Sstevel@tonic-gate 	return (0);
32837c478bd9Sstevel@tonic-gate }
3284