xref: /titanic_52/usr/src/lib/libumem/common/umem.c (revision ae115bc77f6fcde83175c75b4206dc2e50747966)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 /*
30  * based on usr/src/uts/common/os/kmem.c r1.64 from 2001/12/18
31  *
32  * The slab allocator, as described in the following two papers:
33  *
34  *	Jeff Bonwick,
35  *	The Slab Allocator: An Object-Caching Kernel Memory Allocator.
36  *	Proceedings of the Summer 1994 Usenix Conference.
37  *	Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
38  *
39  *	Jeff Bonwick and Jonathan Adams,
40  *	Magazines and vmem: Extending the Slab Allocator to Many CPUs and
41  *	Arbitrary Resources.
42  *	Proceedings of the 2001 Usenix Conference.
43  *	Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
44  *
45  * 1. Overview
46  * -----------
47  * umem is very close to kmem in implementation.  There are four major
48  * areas of divergence:
49  *
50  *	* Initialization
51  *
52  *	* CPU handling
53  *
54  *	* umem_update()
55  *
56  *	* KM_SLEEP v.s. UMEM_NOFAIL
57  *
58  *	* lock ordering
59  *
60  * 2. Initialization
61  * -----------------
62  * kmem is initialized early on in boot, and knows that no one will call
63  * into it before it is ready.  umem does not have these luxuries. Instead,
64  * initialization is divided into two phases:
65  *
66  *	* library initialization, and
67  *
68  *	* first use
69  *
70  * umem's full initialization happens at the time of the first allocation
71  * request (via malloc() and friends, umem_alloc(), or umem_zalloc()),
72  * or the first call to umem_cache_create().
73  *
74  * umem_free(), and umem_cache_alloc() do not require special handling,
75  * since the only way to get valid arguments for them is to successfully
76  * call a function from the first group.
77  *
78  * 2.1. Library Initialization: umem_startup()
79  * -------------------------------------------
80  * umem_startup() is libumem.so's .init section.  It calls pthread_atfork()
81  * to install the handlers necessary for umem's Fork1-Safety.  Because of
82  * race condition issues, all other pre-umem_init() initialization is done
83  * statically (i.e. by the dynamic linker).
84  *
85  * For standalone use, umem_startup() returns everything to its initial
86  * state.
87  *
88  * 2.2. First use: umem_init()
89  * ------------------------------
90  * The first time any memory allocation function is used, we have to
91  * create the backing caches and vmem arenas which are needed for it.
92  * umem_init() is the central point for that task.  When it completes,
93  * umem_ready is either UMEM_READY (all set) or UMEM_READY_INIT_FAILED (unable
94  * to initialize, probably due to lack of memory).
95  *
96  * There are four different paths from which umem_init() is called:
97  *
98  *	* from umem_alloc() or umem_zalloc(), with 0 < size < UMEM_MAXBUF,
99  *
100  *	* from umem_alloc() or umem_zalloc(), with size > UMEM_MAXBUF,
101  *
102  *	* from umem_cache_create(), and
103  *
104  *	* from memalign(), with align > UMEM_ALIGN.
105  *
106  * The last three just check if umem is initialized, and call umem_init()
107  * if it is not.  For performance reasons, the first case is more complicated.
108  *
109  * 2.2.1. umem_alloc()/umem_zalloc(), with 0 < size < UMEM_MAXBUF
110  * -----------------------------------------------------------------
111  * In this case, umem_cache_alloc(&umem_null_cache, ...) is called.
112  * There is special case code in which causes any allocation on
113  * &umem_null_cache to fail by returning (NULL), regardless of the
114  * flags argument.
115  *
116  * So umem_cache_alloc() returns NULL, and umem_alloc()/umem_zalloc() call
117  * umem_alloc_retry().  umem_alloc_retry() sees that the allocation
118  * was agains &umem_null_cache, and calls umem_init().
119  *
120  * If initialization is successful, umem_alloc_retry() returns 1, which
121  * causes umem_alloc()/umem_zalloc() to start over, which causes it to load
122  * the (now valid) cache pointer from umem_alloc_table.
123  *
124  * 2.2.2. Dealing with race conditions
125  * -----------------------------------
126  * There are a couple race conditions resulting from the initialization
127  * code that we have to guard against:
128  *
129  *	* In umem_cache_create(), there is a special UMC_INTERNAL cflag
130  *	that is passed for caches created during initialization.  It
131  *	is illegal for a user to try to create a UMC_INTERNAL cache.
132  *	This allows initialization to proceed, but any other
133  *	umem_cache_create()s will block by calling umem_init().
134  *
135  *	* Since umem_null_cache has a 1-element cache_cpu, it's cache_cpu_mask
136  *	is always zero.  umem_cache_alloc uses cp->cache_cpu_mask to
137  *	mask the cpu number.  This prevents a race between grabbing a
138  *	cache pointer out of umem_alloc_table and growing the cpu array.
139  *
140  *
141  * 3. CPU handling
142  * ---------------
143  * kmem uses the CPU's sequence number to determine which "cpu cache" to
144  * use for an allocation.  Currently, there is no way to get the sequence
145  * number in userspace.
146  *
147  * umem keeps track of cpu information in umem_cpus, an array of umem_max_ncpus
148  * umem_cpu_t structures.  CURCPU() is a a "hint" function, which we then mask
149  * with either umem_cpu_mask or cp->cache_cpu_mask to find the actual "cpu" id.
150  * The mechanics of this is all in the CPU(mask) macro.
151  *
152  * Currently, umem uses _lwp_self() as its hint.
153  *
154  *
155  * 4. The update thread
156  * --------------------
157  * kmem uses a task queue, kmem_taskq, to do periodic maintenance on
158  * every kmem cache.  vmem has a periodic timeout for hash table resizing.
159  * The kmem_taskq also provides a separate context for kmem_cache_reap()'s
160  * to be done in, avoiding issues of the context of kmem_reap() callers.
161  *
162  * Instead, umem has the concept of "updates", which are asynchronous requests
163  * for work attached to single caches.  All caches with pending work are
164  * on a doubly linked list rooted at the umem_null_cache.  All update state
165  * is protected by the umem_update_lock mutex, and the umem_update_cv is used
166  * for notification between threads.
167  *
168  * 4.1. Cache states with regards to updates
169  * -----------------------------------------
170  * A given cache is in one of three states:
171  *
172  * Inactive		cache_uflags is zero, cache_u{next,prev} are NULL
173  *
174  * Work Requested	cache_uflags is non-zero (but UMU_ACTIVE is not set),
175  *			cache_u{next,prev} link the cache onto the global
176  *			update list
177  *
178  * Active		cache_uflags has UMU_ACTIVE set, cache_u{next,prev}
179  *			are NULL, and either umem_update_thr or
180  *			umem_st_update_thr are actively doing work on the
181  *			cache.
182  *
183  * An update can be added to any cache in any state -- if the cache is
184  * Inactive, it transitions to being Work Requested.  If the cache is
185  * Active, the worker will notice the new update and act on it before
186  * transitioning the cache to the Inactive state.
187  *
188  * If a cache is in the Active state, UMU_NOTIFY can be set, which asks
189  * the worker to broadcast the umem_update_cv when it has finished.
190  *
191  * 4.2. Update interface
192  * ---------------------
193  * umem_add_update() adds an update to a particular cache.
194  * umem_updateall() adds an update to all caches.
195  * umem_remove_updates() returns a cache to the Inactive state.
196  *
197  * umem_process_updates() process all caches in the Work Requested state.
198  *
199  * 4.3. Reaping
200  * ------------
201  * When umem_reap() is called (at the time of heap growth), it schedule
202  * UMU_REAP updates on every cache.  It then checks to see if the update
203  * thread exists (umem_update_thr != 0).  If it is, it broadcasts
204  * the umem_update_cv to wake the update thread up, and returns.
205  *
206  * If the update thread does not exist (umem_update_thr == 0), and the
207  * program currently has multiple threads, umem_reap() attempts to create
208  * a new update thread.
209  *
210  * If the process is not multithreaded, or the creation fails, umem_reap()
211  * calls umem_st_update() to do an inline update.
212  *
213  * 4.4. The update thread
214  * ----------------------
215  * The update thread spends most of its time in cond_timedwait() on the
216  * umem_update_cv.  It wakes up under two conditions:
217  *
218  *	* The timedwait times out, in which case it needs to run a global
219  *	update, or
220  *
221  *	* someone cond_broadcast(3THR)s the umem_update_cv, in which case
222  *	it needs to check if there are any caches in the Work Requested
223  *	state.
224  *
225  * When it is time for another global update, umem calls umem_cache_update()
226  * on every cache, then calls vmem_update(), which tunes the vmem structures.
227  * umem_cache_update() can request further work using umem_add_update().
228  *
229  * After any work from the global update completes, the update timer is
230  * reset to umem_reap_interval seconds in the future.  This makes the
231  * updates self-throttling.
232  *
233  * Reaps are similarly self-throttling.  After a UMU_REAP update has
234  * been scheduled on all caches, umem_reap() sets a flag and wakes up the
235  * update thread.  The update thread notices the flag, and resets the
236  * reap state.
237  *
238  * 4.5. Inline updates
239  * -------------------
240  * If the update thread is not running, umem_st_update() is used instead.  It
241  * immediately does a global update (as above), then calls
242  * umem_process_updates() to process both the reaps that umem_reap() added and
243  * any work generated by the global update.  Afterwards, it resets the reap
244  * state.
245  *
246  * While the umem_st_update() is running, umem_st_update_thr holds the thread
247  * id of the thread performing the update.
248  *
249  * 4.6. Updates and fork1()
250  * ------------------------
251  * umem has fork1() pre- and post-handlers which lock up (and release) every
252  * mutex in every cache.  They also lock up the umem_update_lock.  Since
253  * fork1() only copies over a single lwp, other threads (including the update
254  * thread) could have been actively using a cache in the parent.  This
255  * can lead to inconsistencies in the child process.
256  *
257  * Because we locked all of the mutexes, the only possible inconsistancies are:
258  *
259  *	* a umem_cache_alloc() could leak its buffer.
260  *
261  *	* a caller of umem_depot_alloc() could leak a magazine, and all the
262  *	buffers contained in it.
263  *
264  *	* a cache could be in the Active update state.  In the child, there
265  *	would be no thread actually working on it.
266  *
267  *	* a umem_hash_rescale() could leak the new hash table.
268  *
269  *	* a umem_magazine_resize() could be in progress.
270  *
271  *	* a umem_reap() could be in progress.
272  *
273  * The memory leaks we can't do anything about.  umem_release_child() resets
274  * the update state, moves any caches in the Active state to the Work Requested
275  * state.  This might cause some updates to be re-run, but UMU_REAP and
276  * UMU_HASH_RESCALE are effectively idempotent, and the worst that can
277  * happen from umem_magazine_resize() is resizing the magazine twice in close
278  * succession.
279  *
280  * Much of the cleanup in umem_release_child() is skipped if
281  * umem_st_update_thr == thr_self().  This is so that applications which call
282  * fork1() from a cache callback does not break.  Needless to say, any such
283  * application is tremendously broken.
284  *
285  *
286  * 5. KM_SLEEP v.s. UMEM_NOFAIL
287  * ----------------------------
288  * Allocations against kmem and vmem have two basic modes:  SLEEP and
289  * NOSLEEP.  A sleeping allocation is will go to sleep (waiting for
290  * more memory) instead of failing (returning NULL).
291  *
292  * SLEEP allocations presume an extremely multithreaded model, with
293  * a lot of allocation and deallocation activity.  umem cannot presume
294  * that its clients have any particular type of behavior.  Instead,
295  * it provides two types of allocations:
296  *
297  *	* UMEM_DEFAULT, equivalent to KM_NOSLEEP (i.e. return NULL on
298  *	failure)
299  *
300  *	* UMEM_NOFAIL, which, on failure, calls an optional callback
301  *	(registered with umem_nofail_callback()).
302  *
303  * The callback is invoked with no locks held, and can do an arbitrary
304  * amount of work.  It then has a choice between:
305  *
306  *	* Returning UMEM_CALLBACK_RETRY, which will cause the allocation
307  *	to be restarted.
308  *
309  *	* Returning UMEM_CALLBACK_EXIT(status), which will cause exit(2)
310  *	to be invoked with status.  If multiple threads attempt to do
311  *	this simultaneously, only one will call exit(2).
312  *
313  *	* Doing some kind of non-local exit (thr_exit(3thr), longjmp(3C),
314  *	etc.)
315  *
316  * The default callback returns UMEM_CALLBACK_EXIT(255).
317  *
318  * To have these callbacks without risk of state corruption (in the case of
319  * a non-local exit), we have to ensure that the callbacks get invoked
320  * close to the original allocation, with no inconsistent state or held
321  * locks.  The following steps are taken:
322  *
323  *	* All invocations of vmem are VM_NOSLEEP.
324  *
325  *	* All constructor callbacks (which can themselves to allocations)
326  *	are passed UMEM_DEFAULT as their required allocation argument.  This
327  *	way, the constructor will fail, allowing the highest-level allocation
328  *	invoke the nofail callback.
329  *
330  *	If a constructor callback _does_ do a UMEM_NOFAIL allocation, and
331  *	the nofail callback does a non-local exit, we will leak the
332  *	partially-constructed buffer.
333  *
334  *
335  * 6. Lock Ordering
336  * ----------------
337  * umem has a few more locks than kmem does, mostly in the update path.  The
338  * overall lock ordering (earlier locks must be acquired first) is:
339  *
340  *	umem_init_lock
341  *
342  *	vmem_list_lock
343  *	vmem_nosleep_lock.vmpl_mutex
344  *	vmem_t's:
345  *		vm_lock
346  *	sbrk_lock
347  *
348  *	umem_cache_lock
349  *	umem_update_lock
350  *	umem_flags_lock
351  *	umem_cache_t's:
352  *		cache_cpu[*].cc_lock
353  *		cache_depot_lock
354  *		cache_lock
355  *	umem_log_header_t's:
356  *		lh_cpu[*].clh_lock
357  *		lh_lock
358  */
359 
360 #include "c_synonyms.h"
361 #include <umem_impl.h>
362 #include <sys/vmem_impl_user.h>
363 #include "umem_base.h"
364 #include "vmem_base.h"
365 
366 #include <sys/processor.h>
367 #include <sys/sysmacros.h>
368 
369 #include <alloca.h>
370 #include <errno.h>
371 #include <limits.h>
372 #include <stdio.h>
373 #include <stdlib.h>
374 #include <string.h>
375 #include <strings.h>
376 #include <signal.h>
377 #include <unistd.h>
378 #include <atomic.h>
379 
380 #include "misc.h"
381 
382 #define	UMEM_VMFLAGS(umflag)	(VM_NOSLEEP)
383 
384 size_t pagesize;
385 
386 /*
387  * The default set of caches to back umem_alloc().
388  * These sizes should be reevaluated periodically.
389  *
390  * We want allocations that are multiples of the coherency granularity
391  * (64 bytes) to be satisfied from a cache which is a multiple of 64
392  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
393  * the next kmem_cache_size greater than or equal to it must be a
394  * multiple of 64.
395  *
396  * This table must be in sorted order, from smallest to highest.  The
397  * highest slot must be UMEM_MAXBUF, and every slot afterwards must be
398  * zero.
399  */
400 static int umem_alloc_sizes[] = {
401 #ifdef _LP64
402 	1 * 8,
403 	1 * 16,
404 	2 * 16,
405 	3 * 16,
406 #else
407 	1 * 8,
408 	2 * 8,
409 	3 * 8,
410 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
411 #endif
412 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
413 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
414 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
415 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
416 	P2ALIGN(8192 / 7, 64),
417 	P2ALIGN(8192 / 6, 64),
418 	P2ALIGN(8192 / 5, 64),
419 	P2ALIGN(8192 / 4, 64), 2304,
420 	P2ALIGN(8192 / 3, 64),
421 	P2ALIGN(8192 / 2, 64), 4544,
422 	P2ALIGN(8192 / 1, 64), 9216,
423 	4096 * 3,
424 	UMEM_MAXBUF,				/* = 8192 * 2 */
425 	/* 24 slots for user expansion */
426 	0, 0, 0, 0, 0, 0, 0, 0,
427 	0, 0, 0, 0, 0, 0, 0, 0,
428 	0, 0, 0, 0, 0, 0, 0, 0,
429 };
430 #define	NUM_ALLOC_SIZES (sizeof (umem_alloc_sizes) / sizeof (*umem_alloc_sizes))
431 
432 static umem_magtype_t umem_magtype[] = {
433 	{ 1,	8,	3200,	65536	},
434 	{ 3,	16,	256,	32768	},
435 	{ 7,	32,	64,	16384	},
436 	{ 15,	64,	0,	8192	},
437 	{ 31,	64,	0,	4096	},
438 	{ 47,	64,	0,	2048	},
439 	{ 63,	64,	0,	1024	},
440 	{ 95,	64,	0,	512	},
441 	{ 143,	64,	0,	0	},
442 };
443 
444 /*
445  * umem tunables
446  */
447 uint32_t umem_max_ncpus;	/* # of CPU caches. */
448 
449 uint32_t umem_stack_depth = 15; /* # stack frames in a bufctl_audit */
450 uint32_t umem_reap_interval = 10; /* max reaping rate (seconds) */
451 uint_t umem_depot_contention = 2; /* max failed trylocks per real interval */
452 uint_t umem_abort = 1;		/* whether to abort on error */
453 uint_t umem_output = 0;		/* whether to write to standard error */
454 uint_t umem_logging = 0;	/* umem_log_enter() override */
455 uint32_t umem_mtbf = 0;		/* mean time between failures [default: off] */
456 size_t umem_transaction_log_size; /* size of transaction log */
457 size_t umem_content_log_size;	/* size of content log */
458 size_t umem_failure_log_size;	/* failure log [4 pages per CPU] */
459 size_t umem_slab_log_size;	/* slab create log [4 pages per CPU] */
460 size_t umem_content_maxsave = 256; /* UMF_CONTENTS max bytes to log */
461 size_t umem_lite_minsize = 0;	/* minimum buffer size for UMF_LITE */
462 size_t umem_lite_maxalign = 1024; /* maximum buffer alignment for UMF_LITE */
463 size_t umem_maxverify;		/* maximum bytes to inspect in debug routines */
464 size_t umem_minfirewall;	/* hardware-enforced redzone threshold */
465 
466 uint_t umem_flags = 0;
467 
468 mutex_t			umem_init_lock;		/* locks initialization */
469 cond_t			umem_init_cv;		/* initialization CV */
470 thread_t		umem_init_thr;		/* thread initializing */
471 int			umem_init_env_ready;	/* environ pre-initted */
472 int			umem_ready = UMEM_READY_STARTUP;
473 
474 static umem_nofail_callback_t *nofail_callback;
475 static mutex_t		umem_nofail_exit_lock;
476 static thread_t		umem_nofail_exit_thr;
477 
478 static umem_cache_t	*umem_slab_cache;
479 static umem_cache_t	*umem_bufctl_cache;
480 static umem_cache_t	*umem_bufctl_audit_cache;
481 
482 mutex_t			umem_flags_lock;
483 
484 static vmem_t		*heap_arena;
485 static vmem_alloc_t	*heap_alloc;
486 static vmem_free_t	*heap_free;
487 
488 static vmem_t		*umem_internal_arena;
489 static vmem_t		*umem_cache_arena;
490 static vmem_t		*umem_hash_arena;
491 static vmem_t		*umem_log_arena;
492 static vmem_t		*umem_oversize_arena;
493 static vmem_t		*umem_va_arena;
494 static vmem_t		*umem_default_arena;
495 static vmem_t		*umem_firewall_va_arena;
496 static vmem_t		*umem_firewall_arena;
497 
498 vmem_t			*umem_memalign_arena;
499 
500 umem_log_header_t *umem_transaction_log;
501 umem_log_header_t *umem_content_log;
502 umem_log_header_t *umem_failure_log;
503 umem_log_header_t *umem_slab_log;
504 
505 extern thread_t _thr_self(void);
506 #define	CPUHINT()		(_thr_self())
507 #define	CPUHINT_MAX()		INT_MAX
508 
509 #define	CPU(mask)		(umem_cpus + (CPUHINT() & (mask)))
510 static umem_cpu_t umem_startup_cpu = {	/* initial, single, cpu */
511 	UMEM_CACHE_SIZE(0),
512 	0
513 };
514 
515 static uint32_t umem_cpu_mask = 0;			/* global cpu mask */
516 static umem_cpu_t *umem_cpus = &umem_startup_cpu;	/* cpu list */
517 
518 volatile uint32_t umem_reaping;
519 
520 thread_t		umem_update_thr;
521 struct timeval		umem_update_next;	/* timeofday of next update */
522 volatile thread_t	umem_st_update_thr;	/* only used when single-thd */
523 
524 #define	IN_UPDATE()	(thr_self() == umem_update_thr || \
525 			    thr_self() == umem_st_update_thr)
526 #define	IN_REAP()	IN_UPDATE()
527 
528 mutex_t			umem_update_lock;	/* cache_u{next,prev,flags} */
529 cond_t			umem_update_cv;
530 
531 volatile hrtime_t umem_reap_next;	/* min hrtime of next reap */
532 
533 mutex_t			umem_cache_lock;	/* inter-cache linkage only */
534 
535 #ifdef UMEM_STANDALONE
536 umem_cache_t		umem_null_cache;
537 static const umem_cache_t umem_null_cache_template = {
538 #else
539 umem_cache_t		umem_null_cache = {
540 #endif
541 	0, 0, 0, 0, 0,
542 	0, 0,
543 	0, 0,
544 	0, 0,
545 	"invalid_cache",
546 	0, 0,
547 	NULL, NULL, NULL, NULL,
548 	NULL,
549 	0, 0, 0, 0,
550 	&umem_null_cache, &umem_null_cache,
551 	&umem_null_cache, &umem_null_cache,
552 	0,
553 	DEFAULTMUTEX,				/* start of slab layer */
554 	0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
555 	&umem_null_cache.cache_nullslab,
556 	{
557 		&umem_null_cache,
558 		NULL,
559 		&umem_null_cache.cache_nullslab,
560 		&umem_null_cache.cache_nullslab,
561 		NULL,
562 		-1,
563 		0
564 	},
565 	NULL,
566 	NULL,
567 	DEFAULTMUTEX,				/* start of depot layer */
568 	NULL, {
569 		NULL, 0, 0, 0, 0
570 	}, {
571 		NULL, 0, 0, 0, 0
572 	}, {
573 		{
574 			DEFAULTMUTEX,		/* start of CPU cache */
575 			0, 0, NULL, NULL, -1, -1, 0
576 		}
577 	}
578 };
579 
580 #define	ALLOC_TABLE_4 \
581 	&umem_null_cache, &umem_null_cache, &umem_null_cache, &umem_null_cache
582 
583 #define	ALLOC_TABLE_64 \
584 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
585 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
586 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
587 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4
588 
589 #define	ALLOC_TABLE_1024 \
590 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
591 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
592 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
593 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64
594 
595 static umem_cache_t *umem_alloc_table[UMEM_MAXBUF >> UMEM_ALIGN_SHIFT] = {
596 	ALLOC_TABLE_1024,
597 	ALLOC_TABLE_1024
598 };
599 
600 
601 /* Used to constrain audit-log stack traces */
602 caddr_t			umem_min_stack;
603 caddr_t			umem_max_stack;
604 
605 
606 /*
607  * we use the _ versions, since we don't want to be cancelled.
608  * Actually, this is automatically taken care of by including "mtlib.h".
609  */
610 extern int _cond_wait(cond_t *cv, mutex_t *mutex);
611 
612 #define	UMERR_MODIFIED	0	/* buffer modified while on freelist */
613 #define	UMERR_REDZONE	1	/* redzone violation (write past end of buf) */
614 #define	UMERR_DUPFREE	2	/* freed a buffer twice */
615 #define	UMERR_BADADDR	3	/* freed a bad (unallocated) address */
616 #define	UMERR_BADBUFTAG	4	/* buftag corrupted */
617 #define	UMERR_BADBUFCTL	5	/* bufctl corrupted */
618 #define	UMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
619 #define	UMERR_BADSIZE	7	/* alloc size != free size */
620 #define	UMERR_BADBASE	8	/* buffer base address wrong */
621 
622 struct {
623 	hrtime_t	ump_timestamp;	/* timestamp of error */
624 	int		ump_error;	/* type of umem error (UMERR_*) */
625 	void		*ump_buffer;	/* buffer that induced abort */
626 	void		*ump_realbuf;	/* real start address for buffer */
627 	umem_cache_t	*ump_cache;	/* buffer's cache according to client */
628 	umem_cache_t	*ump_realcache;	/* actual cache containing buffer */
629 	umem_slab_t	*ump_slab;	/* slab accoring to umem_findslab() */
630 	umem_bufctl_t	*ump_bufctl;	/* bufctl */
631 } umem_abort_info;
632 
633 static void
634 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
635 {
636 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
637 	uint64_t *buf = buf_arg;
638 
639 	while (buf < bufend)
640 		*buf++ = pattern;
641 }
642 
643 static void *
644 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
645 {
646 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
647 	uint64_t *buf;
648 
649 	for (buf = buf_arg; buf < bufend; buf++)
650 		if (*buf != pattern)
651 			return (buf);
652 	return (NULL);
653 }
654 
655 static void *
656 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
657 {
658 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
659 	uint64_t *buf;
660 
661 	for (buf = buf_arg; buf < bufend; buf++) {
662 		if (*buf != old) {
663 			copy_pattern(old, buf_arg,
664 			    (char *)buf - (char *)buf_arg);
665 			return (buf);
666 		}
667 		*buf = new;
668 	}
669 
670 	return (NULL);
671 }
672 
673 void
674 umem_cache_applyall(void (*func)(umem_cache_t *))
675 {
676 	umem_cache_t *cp;
677 
678 	(void) mutex_lock(&umem_cache_lock);
679 	for (cp = umem_null_cache.cache_next; cp != &umem_null_cache;
680 	    cp = cp->cache_next)
681 		func(cp);
682 	(void) mutex_unlock(&umem_cache_lock);
683 }
684 
685 static void
686 umem_add_update_unlocked(umem_cache_t *cp, int flags)
687 {
688 	umem_cache_t *cnext, *cprev;
689 
690 	flags &= ~UMU_ACTIVE;
691 
692 	if (!flags)
693 		return;
694 
695 	if (cp->cache_uflags & UMU_ACTIVE) {
696 		cp->cache_uflags |= flags;
697 	} else {
698 		if (cp->cache_unext != NULL) {
699 			ASSERT(cp->cache_uflags != 0);
700 			cp->cache_uflags |= flags;
701 		} else {
702 			ASSERT(cp->cache_uflags == 0);
703 			cp->cache_uflags = flags;
704 			cp->cache_unext = cnext = &umem_null_cache;
705 			cp->cache_uprev = cprev = umem_null_cache.cache_uprev;
706 			cnext->cache_uprev = cp;
707 			cprev->cache_unext = cp;
708 		}
709 	}
710 }
711 
712 static void
713 umem_add_update(umem_cache_t *cp, int flags)
714 {
715 	(void) mutex_lock(&umem_update_lock);
716 
717 	umem_add_update_unlocked(cp, flags);
718 
719 	if (!IN_UPDATE())
720 		(void) cond_broadcast(&umem_update_cv);
721 
722 	(void) mutex_unlock(&umem_update_lock);
723 }
724 
725 /*
726  * Remove a cache from the update list, waiting for any in-progress work to
727  * complete first.
728  */
729 static void
730 umem_remove_updates(umem_cache_t *cp)
731 {
732 	(void) mutex_lock(&umem_update_lock);
733 
734 	/*
735 	 * Get it out of the active state
736 	 */
737 	while (cp->cache_uflags & UMU_ACTIVE) {
738 		ASSERT(cp->cache_unext == NULL);
739 
740 		cp->cache_uflags |= UMU_NOTIFY;
741 
742 		/*
743 		 * Make sure the update state is sane, before we wait
744 		 */
745 		ASSERT(umem_update_thr != 0 || umem_st_update_thr != 0);
746 		ASSERT(umem_update_thr != thr_self() &&
747 		    umem_st_update_thr != thr_self());
748 
749 		(void) _cond_wait(&umem_update_cv, &umem_update_lock);
750 	}
751 	/*
752 	 * Get it out of the Work Requested state
753 	 */
754 	if (cp->cache_unext != NULL) {
755 		cp->cache_uprev->cache_unext = cp->cache_unext;
756 		cp->cache_unext->cache_uprev = cp->cache_uprev;
757 		cp->cache_uprev = cp->cache_unext = NULL;
758 		cp->cache_uflags = 0;
759 	}
760 	/*
761 	 * Make sure it is in the Inactive state
762 	 */
763 	ASSERT(cp->cache_unext == NULL && cp->cache_uflags == 0);
764 	(void) mutex_unlock(&umem_update_lock);
765 }
766 
767 static void
768 umem_updateall(int flags)
769 {
770 	umem_cache_t *cp;
771 
772 	/*
773 	 * NOTE:  To prevent deadlock, umem_cache_lock is always acquired first.
774 	 *
775 	 * (umem_add_update is called from things run via umem_cache_applyall)
776 	 */
777 	(void) mutex_lock(&umem_cache_lock);
778 	(void) mutex_lock(&umem_update_lock);
779 
780 	for (cp = umem_null_cache.cache_next; cp != &umem_null_cache;
781 	    cp = cp->cache_next)
782 		umem_add_update_unlocked(cp, flags);
783 
784 	if (!IN_UPDATE())
785 		(void) cond_broadcast(&umem_update_cv);
786 
787 	(void) mutex_unlock(&umem_update_lock);
788 	(void) mutex_unlock(&umem_cache_lock);
789 }
790 
791 /*
792  * Debugging support.  Given a buffer address, find its slab.
793  */
794 static umem_slab_t *
795 umem_findslab(umem_cache_t *cp, void *buf)
796 {
797 	umem_slab_t *sp;
798 
799 	(void) mutex_lock(&cp->cache_lock);
800 	for (sp = cp->cache_nullslab.slab_next;
801 	    sp != &cp->cache_nullslab; sp = sp->slab_next) {
802 		if (UMEM_SLAB_MEMBER(sp, buf)) {
803 			(void) mutex_unlock(&cp->cache_lock);
804 			return (sp);
805 		}
806 	}
807 	(void) mutex_unlock(&cp->cache_lock);
808 
809 	return (NULL);
810 }
811 
812 static void
813 umem_error(int error, umem_cache_t *cparg, void *bufarg)
814 {
815 	umem_buftag_t *btp = NULL;
816 	umem_bufctl_t *bcp = NULL;
817 	umem_cache_t *cp = cparg;
818 	umem_slab_t *sp;
819 	uint64_t *off;
820 	void *buf = bufarg;
821 
822 	int old_logging = umem_logging;
823 
824 	umem_logging = 0;	/* stop logging when a bad thing happens */
825 
826 	umem_abort_info.ump_timestamp = gethrtime();
827 
828 	sp = umem_findslab(cp, buf);
829 	if (sp == NULL) {
830 		for (cp = umem_null_cache.cache_prev; cp != &umem_null_cache;
831 		    cp = cp->cache_prev) {
832 			if ((sp = umem_findslab(cp, buf)) != NULL)
833 				break;
834 		}
835 	}
836 
837 	if (sp == NULL) {
838 		cp = NULL;
839 		error = UMERR_BADADDR;
840 	} else {
841 		if (cp != cparg)
842 			error = UMERR_BADCACHE;
843 		else
844 			buf = (char *)bufarg - ((uintptr_t)bufarg -
845 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
846 		if (buf != bufarg)
847 			error = UMERR_BADBASE;
848 		if (cp->cache_flags & UMF_BUFTAG)
849 			btp = UMEM_BUFTAG(cp, buf);
850 		if (cp->cache_flags & UMF_HASH) {
851 			(void) mutex_lock(&cp->cache_lock);
852 			for (bcp = *UMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
853 				if (bcp->bc_addr == buf)
854 					break;
855 			(void) mutex_unlock(&cp->cache_lock);
856 			if (bcp == NULL && btp != NULL)
857 				bcp = btp->bt_bufctl;
858 			if (umem_findslab(cp->cache_bufctl_cache, bcp) ==
859 			    NULL || P2PHASE((uintptr_t)bcp, UMEM_ALIGN) ||
860 			    bcp->bc_addr != buf) {
861 				error = UMERR_BADBUFCTL;
862 				bcp = NULL;
863 			}
864 		}
865 	}
866 
867 	umem_abort_info.ump_error = error;
868 	umem_abort_info.ump_buffer = bufarg;
869 	umem_abort_info.ump_realbuf = buf;
870 	umem_abort_info.ump_cache = cparg;
871 	umem_abort_info.ump_realcache = cp;
872 	umem_abort_info.ump_slab = sp;
873 	umem_abort_info.ump_bufctl = bcp;
874 
875 	umem_printf("umem allocator: ");
876 
877 	switch (error) {
878 
879 	case UMERR_MODIFIED:
880 		umem_printf("buffer modified after being freed\n");
881 		off = verify_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
882 		if (off == NULL)	/* shouldn't happen */
883 			off = buf;
884 		umem_printf("modification occurred at offset 0x%lx "
885 		    "(0x%llx replaced by 0x%llx)\n",
886 		    (uintptr_t)off - (uintptr_t)buf,
887 		    (longlong_t)UMEM_FREE_PATTERN, (longlong_t)*off);
888 		break;
889 
890 	case UMERR_REDZONE:
891 		umem_printf("redzone violation: write past end of buffer\n");
892 		break;
893 
894 	case UMERR_BADADDR:
895 		umem_printf("invalid free: buffer not in cache\n");
896 		break;
897 
898 	case UMERR_DUPFREE:
899 		umem_printf("duplicate free: buffer freed twice\n");
900 		break;
901 
902 	case UMERR_BADBUFTAG:
903 		umem_printf("boundary tag corrupted\n");
904 		umem_printf("bcp ^ bxstat = %lx, should be %lx\n",
905 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
906 		    UMEM_BUFTAG_FREE);
907 		break;
908 
909 	case UMERR_BADBUFCTL:
910 		umem_printf("bufctl corrupted\n");
911 		break;
912 
913 	case UMERR_BADCACHE:
914 		umem_printf("buffer freed to wrong cache\n");
915 		umem_printf("buffer was allocated from %s,\n", cp->cache_name);
916 		umem_printf("caller attempting free to %s.\n",
917 		    cparg->cache_name);
918 		break;
919 
920 	case UMERR_BADSIZE:
921 		umem_printf("bad free: free size (%u) != alloc size (%u)\n",
922 		    UMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
923 		    UMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
924 		break;
925 
926 	case UMERR_BADBASE:
927 		umem_printf("bad free: free address (%p) != alloc address "
928 		    "(%p)\n", bufarg, buf);
929 		break;
930 	}
931 
932 	umem_printf("buffer=%p  bufctl=%p  cache: %s\n",
933 	    bufarg, (void *)bcp, cparg->cache_name);
934 
935 	if (bcp != NULL && (cp->cache_flags & UMF_AUDIT) &&
936 	    error != UMERR_BADBUFCTL) {
937 		int d;
938 		timespec_t ts;
939 		hrtime_t diff;
940 		umem_bufctl_audit_t *bcap = (umem_bufctl_audit_t *)bcp;
941 
942 		diff = umem_abort_info.ump_timestamp - bcap->bc_timestamp;
943 		ts.tv_sec = diff / NANOSEC;
944 		ts.tv_nsec = diff % NANOSEC;
945 
946 		umem_printf("previous transaction on buffer %p:\n", buf);
947 		umem_printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
948 		    (void *)(intptr_t)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
949 		    (void *)sp, cp->cache_name);
950 		for (d = 0; d < MIN(bcap->bc_depth, umem_stack_depth); d++) {
951 			(void) print_sym((void *)bcap->bc_stack[d]);
952 			umem_printf("\n");
953 		}
954 	}
955 
956 	umem_err_recoverable("umem: heap corruption detected");
957 
958 	umem_logging = old_logging;	/* resume logging */
959 }
960 
961 void
962 umem_nofail_callback(umem_nofail_callback_t *cb)
963 {
964 	nofail_callback = cb;
965 }
966 
967 static int
968 umem_alloc_retry(umem_cache_t *cp, int umflag)
969 {
970 	if (cp == &umem_null_cache) {
971 		if (umem_init())
972 			return (1);				/* retry */
973 		/*
974 		 * Initialization failed.  Do normal failure processing.
975 		 */
976 	}
977 	if (umflag & UMEM_NOFAIL) {
978 		int def_result = UMEM_CALLBACK_EXIT(255);
979 		int result = def_result;
980 		umem_nofail_callback_t *callback = nofail_callback;
981 
982 		if (callback != NULL)
983 			result = callback();
984 
985 		if (result == UMEM_CALLBACK_RETRY)
986 			return (1);
987 
988 		if ((result & ~0xFF) != UMEM_CALLBACK_EXIT(0)) {
989 			log_message("nofail callback returned %x\n", result);
990 			result = def_result;
991 		}
992 
993 		/*
994 		 * only one thread will call exit
995 		 */
996 		if (umem_nofail_exit_thr == thr_self())
997 			umem_panic("recursive UMEM_CALLBACK_EXIT()\n");
998 
999 		(void) mutex_lock(&umem_nofail_exit_lock);
1000 		umem_nofail_exit_thr = thr_self();
1001 		exit(result & 0xFF);
1002 		/*NOTREACHED*/
1003 	}
1004 	return (0);
1005 }
1006 
1007 static umem_log_header_t *
1008 umem_log_init(size_t logsize)
1009 {
1010 	umem_log_header_t *lhp;
1011 	int nchunks = 4 * umem_max_ncpus;
1012 	size_t lhsize = offsetof(umem_log_header_t, lh_cpu[umem_max_ncpus]);
1013 	int i;
1014 
1015 	if (logsize == 0)
1016 		return (NULL);
1017 
1018 	/*
1019 	 * Make sure that lhp->lh_cpu[] is nicely aligned
1020 	 * to prevent false sharing of cache lines.
1021 	 */
1022 	lhsize = P2ROUNDUP(lhsize, UMEM_ALIGN);
1023 	lhp = vmem_xalloc(umem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1024 	    NULL, NULL, VM_NOSLEEP);
1025 	if (lhp == NULL)
1026 		goto fail;
1027 
1028 	bzero(lhp, lhsize);
1029 
1030 	(void) mutex_init(&lhp->lh_lock, USYNC_THREAD, NULL);
1031 	lhp->lh_nchunks = nchunks;
1032 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks, PAGESIZE);
1033 	if (lhp->lh_chunksize == 0)
1034 		lhp->lh_chunksize = PAGESIZE;
1035 
1036 	lhp->lh_base = vmem_alloc(umem_log_arena,
1037 	    lhp->lh_chunksize * nchunks, VM_NOSLEEP);
1038 	if (lhp->lh_base == NULL)
1039 		goto fail;
1040 
1041 	lhp->lh_free = vmem_alloc(umem_log_arena,
1042 	    nchunks * sizeof (int), VM_NOSLEEP);
1043 	if (lhp->lh_free == NULL)
1044 		goto fail;
1045 
1046 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1047 
1048 	for (i = 0; i < umem_max_ncpus; i++) {
1049 		umem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1050 		(void) mutex_init(&clhp->clh_lock, USYNC_THREAD, NULL);
1051 		clhp->clh_chunk = i;
1052 	}
1053 
1054 	for (i = umem_max_ncpus; i < nchunks; i++)
1055 		lhp->lh_free[i] = i;
1056 
1057 	lhp->lh_head = umem_max_ncpus;
1058 	lhp->lh_tail = 0;
1059 
1060 	return (lhp);
1061 
1062 fail:
1063 	if (lhp != NULL) {
1064 		if (lhp->lh_base != NULL)
1065 			vmem_free(umem_log_arena, lhp->lh_base,
1066 			    lhp->lh_chunksize * nchunks);
1067 
1068 		vmem_xfree(umem_log_arena, lhp, lhsize);
1069 	}
1070 	return (NULL);
1071 }
1072 
1073 static void *
1074 umem_log_enter(umem_log_header_t *lhp, void *data, size_t size)
1075 {
1076 	void *logspace;
1077 	umem_cpu_log_header_t *clhp =
1078 	    &lhp->lh_cpu[CPU(umem_cpu_mask)->cpu_number];
1079 
1080 	if (lhp == NULL || umem_logging == 0)
1081 		return (NULL);
1082 
1083 	(void) mutex_lock(&clhp->clh_lock);
1084 	clhp->clh_hits++;
1085 	if (size > clhp->clh_avail) {
1086 		(void) mutex_lock(&lhp->lh_lock);
1087 		lhp->lh_hits++;
1088 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1089 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1090 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1091 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1092 		clhp->clh_current = lhp->lh_base +
1093 		    clhp->clh_chunk * lhp->lh_chunksize;
1094 		clhp->clh_avail = lhp->lh_chunksize;
1095 		if (size > lhp->lh_chunksize)
1096 			size = lhp->lh_chunksize;
1097 		(void) mutex_unlock(&lhp->lh_lock);
1098 	}
1099 	logspace = clhp->clh_current;
1100 	clhp->clh_current += size;
1101 	clhp->clh_avail -= size;
1102 	bcopy(data, logspace, size);
1103 	(void) mutex_unlock(&clhp->clh_lock);
1104 	return (logspace);
1105 }
1106 
1107 #define	UMEM_AUDIT(lp, cp, bcp)						\
1108 {									\
1109 	umem_bufctl_audit_t *_bcp = (umem_bufctl_audit_t *)(bcp);	\
1110 	_bcp->bc_timestamp = gethrtime();				\
1111 	_bcp->bc_thread = thr_self();					\
1112 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, umem_stack_depth,	\
1113 	    (cp != NULL) && (cp->cache_flags & UMF_CHECKSIGNAL));	\
1114 	_bcp->bc_lastlog = umem_log_enter((lp), _bcp,			\
1115 	    UMEM_BUFCTL_AUDIT_SIZE);					\
1116 }
1117 
1118 static void
1119 umem_log_event(umem_log_header_t *lp, umem_cache_t *cp,
1120 	umem_slab_t *sp, void *addr)
1121 {
1122 	umem_bufctl_audit_t *bcp;
1123 	UMEM_LOCAL_BUFCTL_AUDIT(&bcp);
1124 
1125 	bzero(bcp, UMEM_BUFCTL_AUDIT_SIZE);
1126 	bcp->bc_addr = addr;
1127 	bcp->bc_slab = sp;
1128 	bcp->bc_cache = cp;
1129 	UMEM_AUDIT(lp, cp, bcp);
1130 }
1131 
1132 /*
1133  * Create a new slab for cache cp.
1134  */
1135 static umem_slab_t *
1136 umem_slab_create(umem_cache_t *cp, int umflag)
1137 {
1138 	size_t slabsize = cp->cache_slabsize;
1139 	size_t chunksize = cp->cache_chunksize;
1140 	int cache_flags = cp->cache_flags;
1141 	size_t color, chunks;
1142 	char *buf, *slab;
1143 	umem_slab_t *sp;
1144 	umem_bufctl_t *bcp;
1145 	vmem_t *vmp = cp->cache_arena;
1146 
1147 	color = cp->cache_color + cp->cache_align;
1148 	if (color > cp->cache_maxcolor)
1149 		color = cp->cache_mincolor;
1150 	cp->cache_color = color;
1151 
1152 	slab = vmem_alloc(vmp, slabsize, UMEM_VMFLAGS(umflag));
1153 
1154 	if (slab == NULL)
1155 		goto vmem_alloc_failure;
1156 
1157 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1158 
1159 	if (!(cp->cache_cflags & UMC_NOTOUCH) &&
1160 	    (cp->cache_flags & UMF_DEADBEEF))
1161 		copy_pattern(UMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1162 
1163 	if (cache_flags & UMF_HASH) {
1164 		if ((sp = _umem_cache_alloc(umem_slab_cache, umflag)) == NULL)
1165 			goto slab_alloc_failure;
1166 		chunks = (slabsize - color) / chunksize;
1167 	} else {
1168 		sp = UMEM_SLAB(cp, slab);
1169 		chunks = (slabsize - sizeof (umem_slab_t) - color) / chunksize;
1170 	}
1171 
1172 	sp->slab_cache	= cp;
1173 	sp->slab_head	= NULL;
1174 	sp->slab_refcnt	= 0;
1175 	sp->slab_base	= buf = slab + color;
1176 	sp->slab_chunks	= chunks;
1177 
1178 	ASSERT(chunks > 0);
1179 	while (chunks-- != 0) {
1180 		if (cache_flags & UMF_HASH) {
1181 			bcp = _umem_cache_alloc(cp->cache_bufctl_cache, umflag);
1182 			if (bcp == NULL)
1183 				goto bufctl_alloc_failure;
1184 			if (cache_flags & UMF_AUDIT) {
1185 				umem_bufctl_audit_t *bcap =
1186 				    (umem_bufctl_audit_t *)bcp;
1187 				bzero(bcap, UMEM_BUFCTL_AUDIT_SIZE);
1188 				bcap->bc_cache = cp;
1189 			}
1190 			bcp->bc_addr = buf;
1191 			bcp->bc_slab = sp;
1192 		} else {
1193 			bcp = UMEM_BUFCTL(cp, buf);
1194 		}
1195 		if (cache_flags & UMF_BUFTAG) {
1196 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1197 			btp->bt_redzone = UMEM_REDZONE_PATTERN;
1198 			btp->bt_bufctl = bcp;
1199 			btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
1200 			if (cache_flags & UMF_DEADBEEF) {
1201 				copy_pattern(UMEM_FREE_PATTERN, buf,
1202 				    cp->cache_verify);
1203 			}
1204 		}
1205 		bcp->bc_next = sp->slab_head;
1206 		sp->slab_head = bcp;
1207 		buf += chunksize;
1208 	}
1209 
1210 	umem_log_event(umem_slab_log, cp, sp, slab);
1211 
1212 	return (sp);
1213 
1214 bufctl_alloc_failure:
1215 
1216 	while ((bcp = sp->slab_head) != NULL) {
1217 		sp->slab_head = bcp->bc_next;
1218 		_umem_cache_free(cp->cache_bufctl_cache, bcp);
1219 	}
1220 	_umem_cache_free(umem_slab_cache, sp);
1221 
1222 slab_alloc_failure:
1223 
1224 	vmem_free(vmp, slab, slabsize);
1225 
1226 vmem_alloc_failure:
1227 
1228 	umem_log_event(umem_failure_log, cp, NULL, NULL);
1229 	atomic_add_64(&cp->cache_alloc_fail, 1);
1230 
1231 	return (NULL);
1232 }
1233 
1234 /*
1235  * Destroy a slab.
1236  */
1237 static void
1238 umem_slab_destroy(umem_cache_t *cp, umem_slab_t *sp)
1239 {
1240 	vmem_t *vmp = cp->cache_arena;
1241 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1242 
1243 	if (cp->cache_flags & UMF_HASH) {
1244 		umem_bufctl_t *bcp;
1245 		while ((bcp = sp->slab_head) != NULL) {
1246 			sp->slab_head = bcp->bc_next;
1247 			_umem_cache_free(cp->cache_bufctl_cache, bcp);
1248 		}
1249 		_umem_cache_free(umem_slab_cache, sp);
1250 	}
1251 	vmem_free(vmp, slab, cp->cache_slabsize);
1252 }
1253 
1254 /*
1255  * Allocate a raw (unconstructed) buffer from cp's slab layer.
1256  */
1257 static void *
1258 umem_slab_alloc(umem_cache_t *cp, int umflag)
1259 {
1260 	umem_bufctl_t *bcp, **hash_bucket;
1261 	umem_slab_t *sp;
1262 	void *buf;
1263 
1264 	(void) mutex_lock(&cp->cache_lock);
1265 	cp->cache_slab_alloc++;
1266 	sp = cp->cache_freelist;
1267 	ASSERT(sp->slab_cache == cp);
1268 	if (sp->slab_head == NULL) {
1269 		/*
1270 		 * The freelist is empty.  Create a new slab.
1271 		 */
1272 		(void) mutex_unlock(&cp->cache_lock);
1273 		if (cp == &umem_null_cache)
1274 			return (NULL);
1275 		if ((sp = umem_slab_create(cp, umflag)) == NULL)
1276 			return (NULL);
1277 		(void) mutex_lock(&cp->cache_lock);
1278 		cp->cache_slab_create++;
1279 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1280 			cp->cache_bufmax = cp->cache_buftotal;
1281 		sp->slab_next = cp->cache_freelist;
1282 		sp->slab_prev = cp->cache_freelist->slab_prev;
1283 		sp->slab_next->slab_prev = sp;
1284 		sp->slab_prev->slab_next = sp;
1285 		cp->cache_freelist = sp;
1286 	}
1287 
1288 	sp->slab_refcnt++;
1289 	ASSERT(sp->slab_refcnt <= sp->slab_chunks);
1290 
1291 	/*
1292 	 * If we're taking the last buffer in the slab,
1293 	 * remove the slab from the cache's freelist.
1294 	 */
1295 	bcp = sp->slab_head;
1296 	if ((sp->slab_head = bcp->bc_next) == NULL) {
1297 		cp->cache_freelist = sp->slab_next;
1298 		ASSERT(sp->slab_refcnt == sp->slab_chunks);
1299 	}
1300 
1301 	if (cp->cache_flags & UMF_HASH) {
1302 		/*
1303 		 * Add buffer to allocated-address hash table.
1304 		 */
1305 		buf = bcp->bc_addr;
1306 		hash_bucket = UMEM_HASH(cp, buf);
1307 		bcp->bc_next = *hash_bucket;
1308 		*hash_bucket = bcp;
1309 		if ((cp->cache_flags & (UMF_AUDIT | UMF_BUFTAG)) == UMF_AUDIT) {
1310 			UMEM_AUDIT(umem_transaction_log, cp, bcp);
1311 		}
1312 	} else {
1313 		buf = UMEM_BUF(cp, bcp);
1314 	}
1315 
1316 	ASSERT(UMEM_SLAB_MEMBER(sp, buf));
1317 
1318 	(void) mutex_unlock(&cp->cache_lock);
1319 
1320 	return (buf);
1321 }
1322 
1323 /*
1324  * Free a raw (unconstructed) buffer to cp's slab layer.
1325  */
1326 static void
1327 umem_slab_free(umem_cache_t *cp, void *buf)
1328 {
1329 	umem_slab_t *sp;
1330 	umem_bufctl_t *bcp, **prev_bcpp;
1331 
1332 	ASSERT(buf != NULL);
1333 
1334 	(void) mutex_lock(&cp->cache_lock);
1335 	cp->cache_slab_free++;
1336 
1337 	if (cp->cache_flags & UMF_HASH) {
1338 		/*
1339 		 * Look up buffer in allocated-address hash table.
1340 		 */
1341 		prev_bcpp = UMEM_HASH(cp, buf);
1342 		while ((bcp = *prev_bcpp) != NULL) {
1343 			if (bcp->bc_addr == buf) {
1344 				*prev_bcpp = bcp->bc_next;
1345 				sp = bcp->bc_slab;
1346 				break;
1347 			}
1348 			cp->cache_lookup_depth++;
1349 			prev_bcpp = &bcp->bc_next;
1350 		}
1351 	} else {
1352 		bcp = UMEM_BUFCTL(cp, buf);
1353 		sp = UMEM_SLAB(cp, buf);
1354 	}
1355 
1356 	if (bcp == NULL || sp->slab_cache != cp || !UMEM_SLAB_MEMBER(sp, buf)) {
1357 		(void) mutex_unlock(&cp->cache_lock);
1358 		umem_error(UMERR_BADADDR, cp, buf);
1359 		return;
1360 	}
1361 
1362 	if ((cp->cache_flags & (UMF_AUDIT | UMF_BUFTAG)) == UMF_AUDIT) {
1363 		if (cp->cache_flags & UMF_CONTENTS)
1364 			((umem_bufctl_audit_t *)bcp)->bc_contents =
1365 			    umem_log_enter(umem_content_log, buf,
1366 			    cp->cache_contents);
1367 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
1368 	}
1369 
1370 	/*
1371 	 * If this slab isn't currently on the freelist, put it there.
1372 	 */
1373 	if (sp->slab_head == NULL) {
1374 		ASSERT(sp->slab_refcnt == sp->slab_chunks);
1375 		ASSERT(cp->cache_freelist != sp);
1376 		sp->slab_next->slab_prev = sp->slab_prev;
1377 		sp->slab_prev->slab_next = sp->slab_next;
1378 		sp->slab_next = cp->cache_freelist;
1379 		sp->slab_prev = cp->cache_freelist->slab_prev;
1380 		sp->slab_next->slab_prev = sp;
1381 		sp->slab_prev->slab_next = sp;
1382 		cp->cache_freelist = sp;
1383 	}
1384 
1385 	bcp->bc_next = sp->slab_head;
1386 	sp->slab_head = bcp;
1387 
1388 	ASSERT(sp->slab_refcnt >= 1);
1389 	if (--sp->slab_refcnt == 0) {
1390 		/*
1391 		 * There are no outstanding allocations from this slab,
1392 		 * so we can reclaim the memory.
1393 		 */
1394 		sp->slab_next->slab_prev = sp->slab_prev;
1395 		sp->slab_prev->slab_next = sp->slab_next;
1396 		if (sp == cp->cache_freelist)
1397 			cp->cache_freelist = sp->slab_next;
1398 		cp->cache_slab_destroy++;
1399 		cp->cache_buftotal -= sp->slab_chunks;
1400 		(void) mutex_unlock(&cp->cache_lock);
1401 		umem_slab_destroy(cp, sp);
1402 		return;
1403 	}
1404 	(void) mutex_unlock(&cp->cache_lock);
1405 }
1406 
1407 static int
1408 umem_cache_alloc_debug(umem_cache_t *cp, void *buf, int umflag)
1409 {
1410 	umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1411 	umem_bufctl_audit_t *bcp = (umem_bufctl_audit_t *)btp->bt_bufctl;
1412 	uint32_t mtbf;
1413 	int flags_nfatal;
1414 
1415 	if (btp->bt_bxstat != ((intptr_t)bcp ^ UMEM_BUFTAG_FREE)) {
1416 		umem_error(UMERR_BADBUFTAG, cp, buf);
1417 		return (-1);
1418 	}
1419 
1420 	btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_ALLOC;
1421 
1422 	if ((cp->cache_flags & UMF_HASH) && bcp->bc_addr != buf) {
1423 		umem_error(UMERR_BADBUFCTL, cp, buf);
1424 		return (-1);
1425 	}
1426 
1427 	btp->bt_redzone = UMEM_REDZONE_PATTERN;
1428 
1429 	if (cp->cache_flags & UMF_DEADBEEF) {
1430 		if (verify_and_copy_pattern(UMEM_FREE_PATTERN,
1431 		    UMEM_UNINITIALIZED_PATTERN, buf, cp->cache_verify)) {
1432 			umem_error(UMERR_MODIFIED, cp, buf);
1433 			return (-1);
1434 		}
1435 	}
1436 
1437 	if ((mtbf = umem_mtbf | cp->cache_mtbf) != 0 &&
1438 	    gethrtime() % mtbf == 0 &&
1439 	    (umflag & (UMEM_FATAL_FLAGS)) == 0) {
1440 		umem_log_event(umem_failure_log, cp, NULL, NULL);
1441 	} else {
1442 		mtbf = 0;
1443 	}
1444 
1445 	/*
1446 	 * We do not pass fatal flags on to the constructor.  This prevents
1447 	 * leaking buffers in the event of a subordinate constructor failing.
1448 	 */
1449 	flags_nfatal = UMEM_DEFAULT;
1450 	if (mtbf || (cp->cache_constructor != NULL &&
1451 	    cp->cache_constructor(buf, cp->cache_private, flags_nfatal) != 0)) {
1452 		atomic_add_64(&cp->cache_alloc_fail, 1);
1453 		btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
1454 		copy_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
1455 		umem_slab_free(cp, buf);
1456 		return (-1);
1457 	}
1458 
1459 	if (cp->cache_flags & UMF_AUDIT) {
1460 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
1461 	}
1462 
1463 	return (0);
1464 }
1465 
1466 static int
1467 umem_cache_free_debug(umem_cache_t *cp, void *buf)
1468 {
1469 	umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1470 	umem_bufctl_audit_t *bcp = (umem_bufctl_audit_t *)btp->bt_bufctl;
1471 	umem_slab_t *sp;
1472 
1473 	if (btp->bt_bxstat != ((intptr_t)bcp ^ UMEM_BUFTAG_ALLOC)) {
1474 		if (btp->bt_bxstat == ((intptr_t)bcp ^ UMEM_BUFTAG_FREE)) {
1475 			umem_error(UMERR_DUPFREE, cp, buf);
1476 			return (-1);
1477 		}
1478 		sp = umem_findslab(cp, buf);
1479 		if (sp == NULL || sp->slab_cache != cp)
1480 			umem_error(UMERR_BADADDR, cp, buf);
1481 		else
1482 			umem_error(UMERR_REDZONE, cp, buf);
1483 		return (-1);
1484 	}
1485 
1486 	btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
1487 
1488 	if ((cp->cache_flags & UMF_HASH) && bcp->bc_addr != buf) {
1489 		umem_error(UMERR_BADBUFCTL, cp, buf);
1490 		return (-1);
1491 	}
1492 
1493 	if (btp->bt_redzone != UMEM_REDZONE_PATTERN) {
1494 		umem_error(UMERR_REDZONE, cp, buf);
1495 		return (-1);
1496 	}
1497 
1498 	if (cp->cache_flags & UMF_AUDIT) {
1499 		if (cp->cache_flags & UMF_CONTENTS)
1500 			bcp->bc_contents = umem_log_enter(umem_content_log,
1501 			    buf, cp->cache_contents);
1502 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
1503 	}
1504 
1505 	if (cp->cache_destructor != NULL)
1506 		cp->cache_destructor(buf, cp->cache_private);
1507 
1508 	if (cp->cache_flags & UMF_DEADBEEF)
1509 		copy_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
1510 
1511 	return (0);
1512 }
1513 
1514 /*
1515  * Free each object in magazine mp to cp's slab layer, and free mp itself.
1516  */
1517 static void
1518 umem_magazine_destroy(umem_cache_t *cp, umem_magazine_t *mp, int nrounds)
1519 {
1520 	int round;
1521 
1522 	ASSERT(cp->cache_next == NULL || IN_UPDATE());
1523 
1524 	for (round = 0; round < nrounds; round++) {
1525 		void *buf = mp->mag_round[round];
1526 
1527 		if ((cp->cache_flags & UMF_DEADBEEF) &&
1528 		    verify_pattern(UMEM_FREE_PATTERN, buf,
1529 		    cp->cache_verify) != NULL) {
1530 			umem_error(UMERR_MODIFIED, cp, buf);
1531 			continue;
1532 		}
1533 
1534 		if (!(cp->cache_flags & UMF_BUFTAG) &&
1535 		    cp->cache_destructor != NULL)
1536 			cp->cache_destructor(buf, cp->cache_private);
1537 
1538 		umem_slab_free(cp, buf);
1539 	}
1540 	ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
1541 	_umem_cache_free(cp->cache_magtype->mt_cache, mp);
1542 }
1543 
1544 /*
1545  * Allocate a magazine from the depot.
1546  */
1547 static umem_magazine_t *
1548 umem_depot_alloc(umem_cache_t *cp, umem_maglist_t *mlp)
1549 {
1550 	umem_magazine_t *mp;
1551 
1552 	/*
1553 	 * If we can't get the depot lock without contention,
1554 	 * update our contention count.  We use the depot
1555 	 * contention rate to determine whether we need to
1556 	 * increase the magazine size for better scalability.
1557 	 */
1558 	if (mutex_trylock(&cp->cache_depot_lock) != 0) {
1559 		(void) mutex_lock(&cp->cache_depot_lock);
1560 		cp->cache_depot_contention++;
1561 	}
1562 
1563 	if ((mp = mlp->ml_list) != NULL) {
1564 		ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
1565 		mlp->ml_list = mp->mag_next;
1566 		if (--mlp->ml_total < mlp->ml_min)
1567 			mlp->ml_min = mlp->ml_total;
1568 		mlp->ml_alloc++;
1569 	}
1570 
1571 	(void) mutex_unlock(&cp->cache_depot_lock);
1572 
1573 	return (mp);
1574 }
1575 
1576 /*
1577  * Free a magazine to the depot.
1578  */
1579 static void
1580 umem_depot_free(umem_cache_t *cp, umem_maglist_t *mlp, umem_magazine_t *mp)
1581 {
1582 	(void) mutex_lock(&cp->cache_depot_lock);
1583 	ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
1584 	mp->mag_next = mlp->ml_list;
1585 	mlp->ml_list = mp;
1586 	mlp->ml_total++;
1587 	(void) mutex_unlock(&cp->cache_depot_lock);
1588 }
1589 
1590 /*
1591  * Update the working set statistics for cp's depot.
1592  */
1593 static void
1594 umem_depot_ws_update(umem_cache_t *cp)
1595 {
1596 	(void) mutex_lock(&cp->cache_depot_lock);
1597 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
1598 	cp->cache_full.ml_min = cp->cache_full.ml_total;
1599 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
1600 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
1601 	(void) mutex_unlock(&cp->cache_depot_lock);
1602 }
1603 
1604 /*
1605  * Reap all magazines that have fallen out of the depot's working set.
1606  */
1607 static void
1608 umem_depot_ws_reap(umem_cache_t *cp)
1609 {
1610 	long reap;
1611 	umem_magazine_t *mp;
1612 
1613 	ASSERT(cp->cache_next == NULL || IN_REAP());
1614 
1615 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
1616 	while (reap-- && (mp = umem_depot_alloc(cp, &cp->cache_full)) != NULL)
1617 		umem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
1618 
1619 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
1620 	while (reap-- && (mp = umem_depot_alloc(cp, &cp->cache_empty)) != NULL)
1621 		umem_magazine_destroy(cp, mp, 0);
1622 }
1623 
1624 static void
1625 umem_cpu_reload(umem_cpu_cache_t *ccp, umem_magazine_t *mp, int rounds)
1626 {
1627 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
1628 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
1629 	ASSERT(ccp->cc_magsize > 0);
1630 
1631 	ccp->cc_ploaded = ccp->cc_loaded;
1632 	ccp->cc_prounds = ccp->cc_rounds;
1633 	ccp->cc_loaded = mp;
1634 	ccp->cc_rounds = rounds;
1635 }
1636 
1637 /*
1638  * Allocate a constructed object from cache cp.
1639  */
1640 #pragma weak umem_cache_alloc = _umem_cache_alloc
1641 void *
1642 _umem_cache_alloc(umem_cache_t *cp, int umflag)
1643 {
1644 	umem_cpu_cache_t *ccp;
1645 	umem_magazine_t *fmp;
1646 	void *buf;
1647 	int flags_nfatal;
1648 
1649 retry:
1650 	ccp = UMEM_CPU_CACHE(cp, CPU(cp->cache_cpu_mask));
1651 	(void) mutex_lock(&ccp->cc_lock);
1652 	for (;;) {
1653 		/*
1654 		 * If there's an object available in the current CPU's
1655 		 * loaded magazine, just take it and return.
1656 		 */
1657 		if (ccp->cc_rounds > 0) {
1658 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
1659 			ccp->cc_alloc++;
1660 			(void) mutex_unlock(&ccp->cc_lock);
1661 			if ((ccp->cc_flags & UMF_BUFTAG) &&
1662 			    umem_cache_alloc_debug(cp, buf, umflag) == -1) {
1663 				if (umem_alloc_retry(cp, umflag)) {
1664 					goto retry;
1665 				}
1666 
1667 				return (NULL);
1668 			}
1669 			return (buf);
1670 		}
1671 
1672 		/*
1673 		 * The loaded magazine is empty.  If the previously loaded
1674 		 * magazine was full, exchange them and try again.
1675 		 */
1676 		if (ccp->cc_prounds > 0) {
1677 			umem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
1678 			continue;
1679 		}
1680 
1681 		/*
1682 		 * If the magazine layer is disabled, break out now.
1683 		 */
1684 		if (ccp->cc_magsize == 0)
1685 			break;
1686 
1687 		/*
1688 		 * Try to get a full magazine from the depot.
1689 		 */
1690 		fmp = umem_depot_alloc(cp, &cp->cache_full);
1691 		if (fmp != NULL) {
1692 			if (ccp->cc_ploaded != NULL)
1693 				umem_depot_free(cp, &cp->cache_empty,
1694 				    ccp->cc_ploaded);
1695 			umem_cpu_reload(ccp, fmp, ccp->cc_magsize);
1696 			continue;
1697 		}
1698 
1699 		/*
1700 		 * There are no full magazines in the depot,
1701 		 * so fall through to the slab layer.
1702 		 */
1703 		break;
1704 	}
1705 	(void) mutex_unlock(&ccp->cc_lock);
1706 
1707 	/*
1708 	 * We couldn't allocate a constructed object from the magazine layer,
1709 	 * so get a raw buffer from the slab layer and apply its constructor.
1710 	 */
1711 	buf = umem_slab_alloc(cp, umflag);
1712 
1713 	if (buf == NULL) {
1714 		if (cp == &umem_null_cache)
1715 			return (NULL);
1716 		if (umem_alloc_retry(cp, umflag)) {
1717 			goto retry;
1718 		}
1719 
1720 		return (NULL);
1721 	}
1722 
1723 	if (cp->cache_flags & UMF_BUFTAG) {
1724 		/*
1725 		 * Let umem_cache_alloc_debug() apply the constructor for us.
1726 		 */
1727 		if (umem_cache_alloc_debug(cp, buf, umflag) == -1) {
1728 			if (umem_alloc_retry(cp, umflag)) {
1729 				goto retry;
1730 			}
1731 			return (NULL);
1732 		}
1733 		return (buf);
1734 	}
1735 
1736 	/*
1737 	 * We do not pass fatal flags on to the constructor.  This prevents
1738 	 * leaking buffers in the event of a subordinate constructor failing.
1739 	 */
1740 	flags_nfatal = UMEM_DEFAULT;
1741 	if (cp->cache_constructor != NULL &&
1742 	    cp->cache_constructor(buf, cp->cache_private, flags_nfatal) != 0) {
1743 		atomic_add_64(&cp->cache_alloc_fail, 1);
1744 		umem_slab_free(cp, buf);
1745 
1746 		if (umem_alloc_retry(cp, umflag)) {
1747 			goto retry;
1748 		}
1749 		return (NULL);
1750 	}
1751 
1752 	return (buf);
1753 }
1754 
1755 /*
1756  * Free a constructed object to cache cp.
1757  */
1758 #pragma weak umem_cache_free = _umem_cache_free
1759 void
1760 _umem_cache_free(umem_cache_t *cp, void *buf)
1761 {
1762 	umem_cpu_cache_t *ccp = UMEM_CPU_CACHE(cp, CPU(cp->cache_cpu_mask));
1763 	umem_magazine_t *emp;
1764 	umem_magtype_t *mtp;
1765 
1766 	if (ccp->cc_flags & UMF_BUFTAG)
1767 		if (umem_cache_free_debug(cp, buf) == -1)
1768 			return;
1769 
1770 	(void) mutex_lock(&ccp->cc_lock);
1771 	for (;;) {
1772 		/*
1773 		 * If there's a slot available in the current CPU's
1774 		 * loaded magazine, just put the object there and return.
1775 		 */
1776 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
1777 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
1778 			ccp->cc_free++;
1779 			(void) mutex_unlock(&ccp->cc_lock);
1780 			return;
1781 		}
1782 
1783 		/*
1784 		 * The loaded magazine is full.  If the previously loaded
1785 		 * magazine was empty, exchange them and try again.
1786 		 */
1787 		if (ccp->cc_prounds == 0) {
1788 			umem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
1789 			continue;
1790 		}
1791 
1792 		/*
1793 		 * If the magazine layer is disabled, break out now.
1794 		 */
1795 		if (ccp->cc_magsize == 0)
1796 			break;
1797 
1798 		/*
1799 		 * Try to get an empty magazine from the depot.
1800 		 */
1801 		emp = umem_depot_alloc(cp, &cp->cache_empty);
1802 		if (emp != NULL) {
1803 			if (ccp->cc_ploaded != NULL)
1804 				umem_depot_free(cp, &cp->cache_full,
1805 				    ccp->cc_ploaded);
1806 			umem_cpu_reload(ccp, emp, 0);
1807 			continue;
1808 		}
1809 
1810 		/*
1811 		 * There are no empty magazines in the depot,
1812 		 * so try to allocate a new one.  We must drop all locks
1813 		 * across umem_cache_alloc() because lower layers may
1814 		 * attempt to allocate from this cache.
1815 		 */
1816 		mtp = cp->cache_magtype;
1817 		(void) mutex_unlock(&ccp->cc_lock);
1818 		emp = _umem_cache_alloc(mtp->mt_cache, UMEM_DEFAULT);
1819 		(void) mutex_lock(&ccp->cc_lock);
1820 
1821 		if (emp != NULL) {
1822 			/*
1823 			 * We successfully allocated an empty magazine.
1824 			 * However, we had to drop ccp->cc_lock to do it,
1825 			 * so the cache's magazine size may have changed.
1826 			 * If so, free the magazine and try again.
1827 			 */
1828 			if (ccp->cc_magsize != mtp->mt_magsize) {
1829 				(void) mutex_unlock(&ccp->cc_lock);
1830 				_umem_cache_free(mtp->mt_cache, emp);
1831 				(void) mutex_lock(&ccp->cc_lock);
1832 				continue;
1833 			}
1834 
1835 			/*
1836 			 * We got a magazine of the right size.  Add it to
1837 			 * the depot and try the whole dance again.
1838 			 */
1839 			umem_depot_free(cp, &cp->cache_empty, emp);
1840 			continue;
1841 		}
1842 
1843 		/*
1844 		 * We couldn't allocate an empty magazine,
1845 		 * so fall through to the slab layer.
1846 		 */
1847 		break;
1848 	}
1849 	(void) mutex_unlock(&ccp->cc_lock);
1850 
1851 	/*
1852 	 * We couldn't free our constructed object to the magazine layer,
1853 	 * so apply its destructor and free it to the slab layer.
1854 	 * Note that if UMF_BUFTAG is in effect, umem_cache_free_debug()
1855 	 * will have already applied the destructor.
1856 	 */
1857 	if (!(cp->cache_flags & UMF_BUFTAG) && cp->cache_destructor != NULL)
1858 		cp->cache_destructor(buf, cp->cache_private);
1859 
1860 	umem_slab_free(cp, buf);
1861 }
1862 
1863 #pragma weak umem_zalloc = _umem_zalloc
1864 void *
1865 _umem_zalloc(size_t size, int umflag)
1866 {
1867 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
1868 	void *buf;
1869 
1870 retry:
1871 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
1872 		umem_cache_t *cp = umem_alloc_table[index];
1873 		buf = _umem_cache_alloc(cp, umflag);
1874 		if (buf != NULL) {
1875 			if (cp->cache_flags & UMF_BUFTAG) {
1876 				umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1877 				((uint8_t *)buf)[size] = UMEM_REDZONE_BYTE;
1878 				((uint32_t *)btp)[1] = UMEM_SIZE_ENCODE(size);
1879 			}
1880 			bzero(buf, size);
1881 		} else if (umem_alloc_retry(cp, umflag))
1882 			goto retry;
1883 	} else {
1884 		buf = _umem_alloc(size, umflag);	/* handles failure */
1885 		if (buf != NULL)
1886 			bzero(buf, size);
1887 	}
1888 	return (buf);
1889 }
1890 
1891 #pragma weak umem_alloc = _umem_alloc
1892 void *
1893 _umem_alloc(size_t size, int umflag)
1894 {
1895 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
1896 	void *buf;
1897 umem_alloc_retry:
1898 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
1899 		umem_cache_t *cp = umem_alloc_table[index];
1900 		buf = _umem_cache_alloc(cp, umflag);
1901 		if ((cp->cache_flags & UMF_BUFTAG) && buf != NULL) {
1902 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1903 			((uint8_t *)buf)[size] = UMEM_REDZONE_BYTE;
1904 			((uint32_t *)btp)[1] = UMEM_SIZE_ENCODE(size);
1905 		}
1906 		if (buf == NULL && umem_alloc_retry(cp, umflag))
1907 			goto umem_alloc_retry;
1908 		return (buf);
1909 	}
1910 	if (size == 0)
1911 		return (NULL);
1912 	if (umem_oversize_arena == NULL) {
1913 		if (umem_init())
1914 			ASSERT(umem_oversize_arena != NULL);
1915 		else
1916 			return (NULL);
1917 	}
1918 	buf = vmem_alloc(umem_oversize_arena, size, UMEM_VMFLAGS(umflag));
1919 	if (buf == NULL) {
1920 		umem_log_event(umem_failure_log, NULL, NULL, (void *)size);
1921 		if (umem_alloc_retry(NULL, umflag))
1922 			goto umem_alloc_retry;
1923 	}
1924 	return (buf);
1925 }
1926 
1927 #pragma weak umem_alloc_align = _umem_alloc_align
1928 void *
1929 _umem_alloc_align(size_t size, size_t align, int umflag)
1930 {
1931 	void *buf;
1932 
1933 	if (size == 0)
1934 		return (NULL);
1935 	if ((align & (align - 1)) != 0)
1936 		return (NULL);
1937 	if (align < UMEM_ALIGN)
1938 		align = UMEM_ALIGN;
1939 
1940 umem_alloc_align_retry:
1941 	if (umem_memalign_arena == NULL) {
1942 		if (umem_init())
1943 			ASSERT(umem_oversize_arena != NULL);
1944 		else
1945 			return (NULL);
1946 	}
1947 	buf = vmem_xalloc(umem_memalign_arena, size, align, 0, 0, NULL, NULL,
1948 	    UMEM_VMFLAGS(umflag));
1949 	if (buf == NULL) {
1950 		umem_log_event(umem_failure_log, NULL, NULL, (void *)size);
1951 		if (umem_alloc_retry(NULL, umflag))
1952 			goto umem_alloc_align_retry;
1953 	}
1954 	return (buf);
1955 }
1956 
1957 #pragma weak umem_free = _umem_free
1958 void
1959 _umem_free(void *buf, size_t size)
1960 {
1961 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
1962 
1963 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
1964 		umem_cache_t *cp = umem_alloc_table[index];
1965 		if (cp->cache_flags & UMF_BUFTAG) {
1966 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1967 			uint32_t *ip = (uint32_t *)btp;
1968 			if (ip[1] != UMEM_SIZE_ENCODE(size)) {
1969 				if (*(uint64_t *)buf == UMEM_FREE_PATTERN) {
1970 					umem_error(UMERR_DUPFREE, cp, buf);
1971 					return;
1972 				}
1973 				if (UMEM_SIZE_VALID(ip[1])) {
1974 					ip[0] = UMEM_SIZE_ENCODE(size);
1975 					umem_error(UMERR_BADSIZE, cp, buf);
1976 				} else {
1977 					umem_error(UMERR_REDZONE, cp, buf);
1978 				}
1979 				return;
1980 			}
1981 			if (((uint8_t *)buf)[size] != UMEM_REDZONE_BYTE) {
1982 				umem_error(UMERR_REDZONE, cp, buf);
1983 				return;
1984 			}
1985 			btp->bt_redzone = UMEM_REDZONE_PATTERN;
1986 		}
1987 		_umem_cache_free(cp, buf);
1988 	} else {
1989 		if (buf == NULL && size == 0)
1990 			return;
1991 		vmem_free(umem_oversize_arena, buf, size);
1992 	}
1993 }
1994 
1995 #pragma weak umem_free_align = _umem_free_align
1996 void
1997 _umem_free_align(void *buf, size_t size)
1998 {
1999 	if (buf == NULL && size == 0)
2000 		return;
2001 	vmem_xfree(umem_memalign_arena, buf, size);
2002 }
2003 
2004 static void *
2005 umem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
2006 {
2007 	size_t realsize = size + vmp->vm_quantum;
2008 
2009 	/*
2010 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
2011 	 * vm_quantum will cause integer wraparound.  Check for this, and
2012 	 * blow off the firewall page in this case.  Note that such a
2013 	 * giant allocation (the entire address space) can never be
2014 	 * satisfied, so it will either fail immediately (VM_NOSLEEP)
2015 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
2016 	 * corresponding check in umem_firewall_va_free().
2017 	 */
2018 	if (realsize < size)
2019 		realsize = size;
2020 
2021 	return (vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT));
2022 }
2023 
2024 static void
2025 umem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
2026 {
2027 	vmem_free(vmp, addr, size + vmp->vm_quantum);
2028 }
2029 
2030 /*
2031  * Reclaim all unused memory from a cache.
2032  */
2033 static void
2034 umem_cache_reap(umem_cache_t *cp)
2035 {
2036 	/*
2037 	 * Ask the cache's owner to free some memory if possible.
2038 	 * The idea is to handle things like the inode cache, which
2039 	 * typically sits on a bunch of memory that it doesn't truly
2040 	 * *need*.  Reclaim policy is entirely up to the owner; this
2041 	 * callback is just an advisory plea for help.
2042 	 */
2043 	if (cp->cache_reclaim != NULL)
2044 		cp->cache_reclaim(cp->cache_private);
2045 
2046 	umem_depot_ws_reap(cp);
2047 }
2048 
2049 /*
2050  * Purge all magazines from a cache and set its magazine limit to zero.
2051  * All calls are serialized by being done by the update thread, except for
2052  * the final call from umem_cache_destroy().
2053  */
2054 static void
2055 umem_cache_magazine_purge(umem_cache_t *cp)
2056 {
2057 	umem_cpu_cache_t *ccp;
2058 	umem_magazine_t *mp, *pmp;
2059 	int rounds, prounds, cpu_seqid;
2060 
2061 	ASSERT(cp->cache_next == NULL || IN_UPDATE());
2062 
2063 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
2064 		ccp = &cp->cache_cpu[cpu_seqid];
2065 
2066 		(void) mutex_lock(&ccp->cc_lock);
2067 		mp = ccp->cc_loaded;
2068 		pmp = ccp->cc_ploaded;
2069 		rounds = ccp->cc_rounds;
2070 		prounds = ccp->cc_prounds;
2071 		ccp->cc_loaded = NULL;
2072 		ccp->cc_ploaded = NULL;
2073 		ccp->cc_rounds = -1;
2074 		ccp->cc_prounds = -1;
2075 		ccp->cc_magsize = 0;
2076 		(void) mutex_unlock(&ccp->cc_lock);
2077 
2078 		if (mp)
2079 			umem_magazine_destroy(cp, mp, rounds);
2080 		if (pmp)
2081 			umem_magazine_destroy(cp, pmp, prounds);
2082 	}
2083 
2084 	/*
2085 	 * Updating the working set statistics twice in a row has the
2086 	 * effect of setting the working set size to zero, so everything
2087 	 * is eligible for reaping.
2088 	 */
2089 	umem_depot_ws_update(cp);
2090 	umem_depot_ws_update(cp);
2091 
2092 	umem_depot_ws_reap(cp);
2093 }
2094 
2095 /*
2096  * Enable per-cpu magazines on a cache.
2097  */
2098 static void
2099 umem_cache_magazine_enable(umem_cache_t *cp)
2100 {
2101 	int cpu_seqid;
2102 
2103 	if (cp->cache_flags & UMF_NOMAGAZINE)
2104 		return;
2105 
2106 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
2107 		umem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
2108 		(void) mutex_lock(&ccp->cc_lock);
2109 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
2110 		(void) mutex_unlock(&ccp->cc_lock);
2111 	}
2112 
2113 }
2114 
2115 /*
2116  * Recompute a cache's magazine size.  The trade-off is that larger magazines
2117  * provide a higher transfer rate with the depot, while smaller magazines
2118  * reduce memory consumption.  Magazine resizing is an expensive operation;
2119  * it should not be done frequently.
2120  *
2121  * Changes to the magazine size are serialized by only having one thread
2122  * doing updates. (the update thread)
2123  *
2124  * Note: at present this only grows the magazine size.  It might be useful
2125  * to allow shrinkage too.
2126  */
2127 static void
2128 umem_cache_magazine_resize(umem_cache_t *cp)
2129 {
2130 	umem_magtype_t *mtp = cp->cache_magtype;
2131 
2132 	ASSERT(IN_UPDATE());
2133 
2134 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
2135 		umem_cache_magazine_purge(cp);
2136 		(void) mutex_lock(&cp->cache_depot_lock);
2137 		cp->cache_magtype = ++mtp;
2138 		cp->cache_depot_contention_prev =
2139 		    cp->cache_depot_contention + INT_MAX;
2140 		(void) mutex_unlock(&cp->cache_depot_lock);
2141 		umem_cache_magazine_enable(cp);
2142 	}
2143 }
2144 
2145 /*
2146  * Rescale a cache's hash table, so that the table size is roughly the
2147  * cache size.  We want the average lookup time to be extremely small.
2148  */
2149 static void
2150 umem_hash_rescale(umem_cache_t *cp)
2151 {
2152 	umem_bufctl_t **old_table, **new_table, *bcp;
2153 	size_t old_size, new_size, h;
2154 
2155 	ASSERT(IN_UPDATE());
2156 
2157 	new_size = MAX(UMEM_HASH_INITIAL,
2158 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
2159 	old_size = cp->cache_hash_mask + 1;
2160 
2161 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
2162 		return;
2163 
2164 	new_table = vmem_alloc(umem_hash_arena, new_size * sizeof (void *),
2165 	    VM_NOSLEEP);
2166 	if (new_table == NULL)
2167 		return;
2168 	bzero(new_table, new_size * sizeof (void *));
2169 
2170 	(void) mutex_lock(&cp->cache_lock);
2171 
2172 	old_size = cp->cache_hash_mask + 1;
2173 	old_table = cp->cache_hash_table;
2174 
2175 	cp->cache_hash_mask = new_size - 1;
2176 	cp->cache_hash_table = new_table;
2177 	cp->cache_rescale++;
2178 
2179 	for (h = 0; h < old_size; h++) {
2180 		bcp = old_table[h];
2181 		while (bcp != NULL) {
2182 			void *addr = bcp->bc_addr;
2183 			umem_bufctl_t *next_bcp = bcp->bc_next;
2184 			umem_bufctl_t **hash_bucket = UMEM_HASH(cp, addr);
2185 			bcp->bc_next = *hash_bucket;
2186 			*hash_bucket = bcp;
2187 			bcp = next_bcp;
2188 		}
2189 	}
2190 
2191 	(void) mutex_unlock(&cp->cache_lock);
2192 
2193 	vmem_free(umem_hash_arena, old_table, old_size * sizeof (void *));
2194 }
2195 
2196 /*
2197  * Perform periodic maintenance on a cache: hash rescaling,
2198  * depot working-set update, and magazine resizing.
2199  */
2200 void
2201 umem_cache_update(umem_cache_t *cp)
2202 {
2203 	int update_flags = 0;
2204 
2205 	ASSERT(MUTEX_HELD(&umem_cache_lock));
2206 
2207 	/*
2208 	 * If the cache has become much larger or smaller than its hash table,
2209 	 * fire off a request to rescale the hash table.
2210 	 */
2211 	(void) mutex_lock(&cp->cache_lock);
2212 
2213 	if ((cp->cache_flags & UMF_HASH) &&
2214 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
2215 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
2216 	    cp->cache_hash_mask > UMEM_HASH_INITIAL)))
2217 		update_flags |= UMU_HASH_RESCALE;
2218 
2219 	(void) mutex_unlock(&cp->cache_lock);
2220 
2221 	/*
2222 	 * Update the depot working set statistics.
2223 	 */
2224 	umem_depot_ws_update(cp);
2225 
2226 	/*
2227 	 * If there's a lot of contention in the depot,
2228 	 * increase the magazine size.
2229 	 */
2230 	(void) mutex_lock(&cp->cache_depot_lock);
2231 
2232 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
2233 	    (int)(cp->cache_depot_contention -
2234 	    cp->cache_depot_contention_prev) > umem_depot_contention)
2235 		update_flags |= UMU_MAGAZINE_RESIZE;
2236 
2237 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
2238 
2239 	(void) mutex_unlock(&cp->cache_depot_lock);
2240 
2241 	if (update_flags)
2242 		umem_add_update(cp, update_flags);
2243 }
2244 
2245 /*
2246  * Runs all pending updates.
2247  *
2248  * The update lock must be held on entrance, and will be held on exit.
2249  */
2250 void
2251 umem_process_updates(void)
2252 {
2253 	ASSERT(MUTEX_HELD(&umem_update_lock));
2254 
2255 	while (umem_null_cache.cache_unext != &umem_null_cache) {
2256 		int notify = 0;
2257 		umem_cache_t *cp = umem_null_cache.cache_unext;
2258 
2259 		cp->cache_uprev->cache_unext = cp->cache_unext;
2260 		cp->cache_unext->cache_uprev = cp->cache_uprev;
2261 		cp->cache_uprev = cp->cache_unext = NULL;
2262 
2263 		ASSERT(!(cp->cache_uflags & UMU_ACTIVE));
2264 
2265 		while (cp->cache_uflags) {
2266 			int uflags = (cp->cache_uflags |= UMU_ACTIVE);
2267 			(void) mutex_unlock(&umem_update_lock);
2268 
2269 			/*
2270 			 * The order here is important.  Each step can speed up
2271 			 * later steps.
2272 			 */
2273 
2274 			if (uflags & UMU_HASH_RESCALE)
2275 				umem_hash_rescale(cp);
2276 
2277 			if (uflags & UMU_MAGAZINE_RESIZE)
2278 				umem_cache_magazine_resize(cp);
2279 
2280 			if (uflags & UMU_REAP)
2281 				umem_cache_reap(cp);
2282 
2283 			(void) mutex_lock(&umem_update_lock);
2284 
2285 			/*
2286 			 * check if anyone has requested notification
2287 			 */
2288 			if (cp->cache_uflags & UMU_NOTIFY) {
2289 				uflags |= UMU_NOTIFY;
2290 				notify = 1;
2291 			}
2292 			cp->cache_uflags &= ~uflags;
2293 		}
2294 		if (notify)
2295 			(void) cond_broadcast(&umem_update_cv);
2296 	}
2297 }
2298 
2299 #ifndef UMEM_STANDALONE
2300 static void
2301 umem_st_update(void)
2302 {
2303 	ASSERT(MUTEX_HELD(&umem_update_lock));
2304 	ASSERT(umem_update_thr == 0 && umem_st_update_thr == 0);
2305 
2306 	umem_st_update_thr = thr_self();
2307 
2308 	(void) mutex_unlock(&umem_update_lock);
2309 
2310 	vmem_update(NULL);
2311 	umem_cache_applyall(umem_cache_update);
2312 
2313 	(void) mutex_lock(&umem_update_lock);
2314 
2315 	umem_process_updates();	/* does all of the requested work */
2316 
2317 	umem_reap_next = gethrtime() +
2318 	    (hrtime_t)umem_reap_interval * NANOSEC;
2319 
2320 	umem_reaping = UMEM_REAP_DONE;
2321 
2322 	umem_st_update_thr = 0;
2323 }
2324 #endif
2325 
2326 /*
2327  * Reclaim all unused memory from all caches.  Called from vmem when memory
2328  * gets tight.  Must be called with no locks held.
2329  *
2330  * This just requests a reap on all caches, and notifies the update thread.
2331  */
2332 void
2333 umem_reap(void)
2334 {
2335 #ifndef UMEM_STANDALONE
2336 	extern int __nthreads(void);
2337 #endif
2338 
2339 	if (umem_ready != UMEM_READY || umem_reaping != UMEM_REAP_DONE ||
2340 	    gethrtime() < umem_reap_next)
2341 		return;
2342 
2343 	(void) mutex_lock(&umem_update_lock);
2344 
2345 	if (umem_reaping != UMEM_REAP_DONE || gethrtime() < umem_reap_next) {
2346 		(void) mutex_unlock(&umem_update_lock);
2347 		return;
2348 	}
2349 	umem_reaping = UMEM_REAP_ADDING;	/* lock out other reaps */
2350 
2351 	(void) mutex_unlock(&umem_update_lock);
2352 
2353 	umem_updateall(UMU_REAP);
2354 
2355 	(void) mutex_lock(&umem_update_lock);
2356 
2357 	umem_reaping = UMEM_REAP_ACTIVE;
2358 
2359 	/* Standalone is single-threaded */
2360 #ifndef UMEM_STANDALONE
2361 	if (umem_update_thr == 0) {
2362 		/*
2363 		 * The update thread does not exist.  If the process is
2364 		 * multi-threaded, create it.  If not, or the creation fails,
2365 		 * do the update processing inline.
2366 		 */
2367 		ASSERT(umem_st_update_thr == 0);
2368 
2369 		if (__nthreads() <= 1 || umem_create_update_thread() == 0)
2370 			umem_st_update();
2371 	}
2372 
2373 	(void) cond_broadcast(&umem_update_cv);	/* wake up the update thread */
2374 #endif
2375 
2376 	(void) mutex_unlock(&umem_update_lock);
2377 }
2378 
2379 umem_cache_t *
2380 umem_cache_create(
2381 	char *name,		/* descriptive name for this cache */
2382 	size_t bufsize,		/* size of the objects it manages */
2383 	size_t align,		/* required object alignment */
2384 	umem_constructor_t *constructor, /* object constructor */
2385 	umem_destructor_t *destructor, /* object destructor */
2386 	umem_reclaim_t *reclaim, /* memory reclaim callback */
2387 	void *private,		/* pass-thru arg for constr/destr/reclaim */
2388 	vmem_t *vmp,		/* vmem source for slab allocation */
2389 	int cflags)		/* cache creation flags */
2390 {
2391 	int cpu_seqid;
2392 	size_t chunksize;
2393 	umem_cache_t *cp, *cnext, *cprev;
2394 	umem_magtype_t *mtp;
2395 	size_t csize;
2396 	size_t phase;
2397 
2398 	/*
2399 	 * The init thread is allowed to create internal and quantum caches.
2400 	 *
2401 	 * Other threads must wait until until initialization is complete.
2402 	 */
2403 	if (umem_init_thr == thr_self())
2404 		ASSERT((cflags & (UMC_INTERNAL | UMC_QCACHE)) != 0);
2405 	else {
2406 		ASSERT(!(cflags & UMC_INTERNAL));
2407 		if (umem_ready != UMEM_READY && umem_init() == 0) {
2408 			errno = EAGAIN;
2409 			return (NULL);
2410 		}
2411 	}
2412 
2413 	csize = UMEM_CACHE_SIZE(umem_max_ncpus);
2414 	phase = P2NPHASE(csize, UMEM_CPU_CACHE_SIZE);
2415 
2416 	if (vmp == NULL)
2417 		vmp = umem_default_arena;
2418 
2419 	ASSERT(P2PHASE(phase, UMEM_ALIGN) == 0);
2420 
2421 	/*
2422 	 * Check that the arguments are reasonable
2423 	 */
2424 	if ((align & (align - 1)) != 0 || align > vmp->vm_quantum ||
2425 	    ((cflags & UMC_NOHASH) && (cflags & UMC_NOTOUCH)) ||
2426 	    name == NULL || bufsize == 0) {
2427 		errno = EINVAL;
2428 		return (NULL);
2429 	}
2430 
2431 	/*
2432 	 * If align == 0, we set it to the minimum required alignment.
2433 	 *
2434 	 * If align < UMEM_ALIGN, we round it up to UMEM_ALIGN, unless
2435 	 * UMC_NOTOUCH was passed.
2436 	 */
2437 	if (align == 0) {
2438 		if (P2ROUNDUP(bufsize, UMEM_ALIGN) >= UMEM_SECOND_ALIGN)
2439 			align = UMEM_SECOND_ALIGN;
2440 		else
2441 			align = UMEM_ALIGN;
2442 	} else if (align < UMEM_ALIGN && (cflags & UMC_NOTOUCH) == 0)
2443 		align = UMEM_ALIGN;
2444 
2445 
2446 	/*
2447 	 * Get a umem_cache structure.  We arrange that cp->cache_cpu[]
2448 	 * is aligned on a UMEM_CPU_CACHE_SIZE boundary to prevent
2449 	 * false sharing of per-CPU data.
2450 	 */
2451 	cp = vmem_xalloc(umem_cache_arena, csize, UMEM_CPU_CACHE_SIZE, phase,
2452 	    0, NULL, NULL, VM_NOSLEEP);
2453 
2454 	if (cp == NULL) {
2455 		errno = EAGAIN;
2456 		return (NULL);
2457 	}
2458 
2459 	bzero(cp, csize);
2460 
2461 	(void) mutex_lock(&umem_flags_lock);
2462 	if (umem_flags & UMF_RANDOMIZE)
2463 		umem_flags = (((umem_flags | ~UMF_RANDOM) + 1) & UMF_RANDOM) |
2464 		    UMF_RANDOMIZE;
2465 	cp->cache_flags = umem_flags | (cflags & UMF_DEBUG);
2466 	(void) mutex_unlock(&umem_flags_lock);
2467 
2468 	/*
2469 	 * Make sure all the various flags are reasonable.
2470 	 */
2471 	if (cp->cache_flags & UMF_LITE) {
2472 		if (bufsize >= umem_lite_minsize &&
2473 		    align <= umem_lite_maxalign &&
2474 		    P2PHASE(bufsize, umem_lite_maxalign) != 0) {
2475 			cp->cache_flags |= UMF_BUFTAG;
2476 			cp->cache_flags &= ~(UMF_AUDIT | UMF_FIREWALL);
2477 		} else {
2478 			cp->cache_flags &= ~UMF_DEBUG;
2479 		}
2480 	}
2481 
2482 	if ((cflags & UMC_QCACHE) && (cp->cache_flags & UMF_AUDIT))
2483 		cp->cache_flags |= UMF_NOMAGAZINE;
2484 
2485 	if (cflags & UMC_NODEBUG)
2486 		cp->cache_flags &= ~UMF_DEBUG;
2487 
2488 	if (cflags & UMC_NOTOUCH)
2489 		cp->cache_flags &= ~UMF_TOUCH;
2490 
2491 	if (cflags & UMC_NOHASH)
2492 		cp->cache_flags &= ~(UMF_AUDIT | UMF_FIREWALL);
2493 
2494 	if (cflags & UMC_NOMAGAZINE)
2495 		cp->cache_flags |= UMF_NOMAGAZINE;
2496 
2497 	if ((cp->cache_flags & UMF_AUDIT) && !(cflags & UMC_NOTOUCH))
2498 		cp->cache_flags |= UMF_REDZONE;
2499 
2500 	if ((cp->cache_flags & UMF_BUFTAG) && bufsize >= umem_minfirewall &&
2501 	    !(cp->cache_flags & UMF_LITE) && !(cflags & UMC_NOHASH))
2502 		cp->cache_flags |= UMF_FIREWALL;
2503 
2504 	if (vmp != umem_default_arena || umem_firewall_arena == NULL)
2505 		cp->cache_flags &= ~UMF_FIREWALL;
2506 
2507 	if (cp->cache_flags & UMF_FIREWALL) {
2508 		cp->cache_flags &= ~UMF_BUFTAG;
2509 		cp->cache_flags |= UMF_NOMAGAZINE;
2510 		ASSERT(vmp == umem_default_arena);
2511 		vmp = umem_firewall_arena;
2512 	}
2513 
2514 	/*
2515 	 * Set cache properties.
2516 	 */
2517 	(void) strncpy(cp->cache_name, name, sizeof (cp->cache_name) - 1);
2518 	cp->cache_bufsize = bufsize;
2519 	cp->cache_align = align;
2520 	cp->cache_constructor = constructor;
2521 	cp->cache_destructor = destructor;
2522 	cp->cache_reclaim = reclaim;
2523 	cp->cache_private = private;
2524 	cp->cache_arena = vmp;
2525 	cp->cache_cflags = cflags;
2526 	cp->cache_cpu_mask = umem_cpu_mask;
2527 
2528 	/*
2529 	 * Determine the chunk size.
2530 	 */
2531 	chunksize = bufsize;
2532 
2533 	if (align >= UMEM_ALIGN) {
2534 		chunksize = P2ROUNDUP(chunksize, UMEM_ALIGN);
2535 		cp->cache_bufctl = chunksize - UMEM_ALIGN;
2536 	}
2537 
2538 	if (cp->cache_flags & UMF_BUFTAG) {
2539 		cp->cache_bufctl = chunksize;
2540 		cp->cache_buftag = chunksize;
2541 		chunksize += sizeof (umem_buftag_t);
2542 	}
2543 
2544 	if (cp->cache_flags & UMF_DEADBEEF) {
2545 		cp->cache_verify = MIN(cp->cache_buftag, umem_maxverify);
2546 		if (cp->cache_flags & UMF_LITE)
2547 			cp->cache_verify = MIN(cp->cache_verify, UMEM_ALIGN);
2548 	}
2549 
2550 	cp->cache_contents = MIN(cp->cache_bufctl, umem_content_maxsave);
2551 
2552 	cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
2553 
2554 	if (chunksize < bufsize) {
2555 		errno = ENOMEM;
2556 		goto fail;
2557 	}
2558 
2559 	/*
2560 	 * Now that we know the chunk size, determine the optimal slab size.
2561 	 */
2562 	if (vmp == umem_firewall_arena) {
2563 		cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
2564 		cp->cache_mincolor = cp->cache_slabsize - chunksize;
2565 		cp->cache_maxcolor = cp->cache_mincolor;
2566 		cp->cache_flags |= UMF_HASH;
2567 		ASSERT(!(cp->cache_flags & UMF_BUFTAG));
2568 	} else if ((cflags & UMC_NOHASH) || (!(cflags & UMC_NOTOUCH) &&
2569 	    !(cp->cache_flags & UMF_AUDIT) &&
2570 	    chunksize < vmp->vm_quantum / UMEM_VOID_FRACTION)) {
2571 		cp->cache_slabsize = vmp->vm_quantum;
2572 		cp->cache_mincolor = 0;
2573 		cp->cache_maxcolor =
2574 		    (cp->cache_slabsize - sizeof (umem_slab_t)) % chunksize;
2575 
2576 		if (chunksize + sizeof (umem_slab_t) > cp->cache_slabsize) {
2577 			errno = EINVAL;
2578 			goto fail;
2579 		}
2580 		ASSERT(!(cp->cache_flags & UMF_AUDIT));
2581 	} else {
2582 		size_t chunks, bestfit, waste, slabsize;
2583 		size_t minwaste = LONG_MAX;
2584 
2585 		for (chunks = 1; chunks <= UMEM_VOID_FRACTION; chunks++) {
2586 			slabsize = P2ROUNDUP(chunksize * chunks,
2587 			    vmp->vm_quantum);
2588 			/*
2589 			 * check for overflow
2590 			 */
2591 			if ((slabsize / chunks) < chunksize) {
2592 				errno = ENOMEM;
2593 				goto fail;
2594 			}
2595 			chunks = slabsize / chunksize;
2596 			waste = (slabsize % chunksize) / chunks;
2597 			if (waste < minwaste) {
2598 				minwaste = waste;
2599 				bestfit = slabsize;
2600 			}
2601 		}
2602 		if (cflags & UMC_QCACHE)
2603 			bestfit = MAX(1 << highbit(3 * vmp->vm_qcache_max), 64);
2604 		cp->cache_slabsize = bestfit;
2605 		cp->cache_mincolor = 0;
2606 		cp->cache_maxcolor = bestfit % chunksize;
2607 		cp->cache_flags |= UMF_HASH;
2608 	}
2609 
2610 	if (cp->cache_flags & UMF_HASH) {
2611 		ASSERT(!(cflags & UMC_NOHASH));
2612 		cp->cache_bufctl_cache = (cp->cache_flags & UMF_AUDIT) ?
2613 		    umem_bufctl_audit_cache : umem_bufctl_cache;
2614 	}
2615 
2616 	if (cp->cache_maxcolor >= vmp->vm_quantum)
2617 		cp->cache_maxcolor = vmp->vm_quantum - 1;
2618 
2619 	cp->cache_color = cp->cache_mincolor;
2620 
2621 	/*
2622 	 * Initialize the rest of the slab layer.
2623 	 */
2624 	(void) mutex_init(&cp->cache_lock, USYNC_THREAD, NULL);
2625 
2626 	cp->cache_freelist = &cp->cache_nullslab;
2627 	cp->cache_nullslab.slab_cache = cp;
2628 	cp->cache_nullslab.slab_refcnt = -1;
2629 	cp->cache_nullslab.slab_next = &cp->cache_nullslab;
2630 	cp->cache_nullslab.slab_prev = &cp->cache_nullslab;
2631 
2632 	if (cp->cache_flags & UMF_HASH) {
2633 		cp->cache_hash_table = vmem_alloc(umem_hash_arena,
2634 		    UMEM_HASH_INITIAL * sizeof (void *), VM_NOSLEEP);
2635 		if (cp->cache_hash_table == NULL) {
2636 			errno = EAGAIN;
2637 			goto fail_lock;
2638 		}
2639 		bzero(cp->cache_hash_table,
2640 		    UMEM_HASH_INITIAL * sizeof (void *));
2641 		cp->cache_hash_mask = UMEM_HASH_INITIAL - 1;
2642 		cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
2643 	}
2644 
2645 	/*
2646 	 * Initialize the depot.
2647 	 */
2648 	(void) mutex_init(&cp->cache_depot_lock, USYNC_THREAD, NULL);
2649 
2650 	for (mtp = umem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
2651 		continue;
2652 
2653 	cp->cache_magtype = mtp;
2654 
2655 	/*
2656 	 * Initialize the CPU layer.
2657 	 */
2658 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
2659 		umem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
2660 		(void) mutex_init(&ccp->cc_lock, USYNC_THREAD, NULL);
2661 		ccp->cc_flags = cp->cache_flags;
2662 		ccp->cc_rounds = -1;
2663 		ccp->cc_prounds = -1;
2664 	}
2665 
2666 	/*
2667 	 * Add the cache to the global list.  This makes it visible
2668 	 * to umem_update(), so the cache must be ready for business.
2669 	 */
2670 	(void) mutex_lock(&umem_cache_lock);
2671 	cp->cache_next = cnext = &umem_null_cache;
2672 	cp->cache_prev = cprev = umem_null_cache.cache_prev;
2673 	cnext->cache_prev = cp;
2674 	cprev->cache_next = cp;
2675 	(void) mutex_unlock(&umem_cache_lock);
2676 
2677 	if (umem_ready == UMEM_READY)
2678 		umem_cache_magazine_enable(cp);
2679 
2680 	return (cp);
2681 
2682 fail_lock:
2683 	(void) mutex_destroy(&cp->cache_lock);
2684 fail:
2685 	vmem_xfree(umem_cache_arena, cp, csize);
2686 	return (NULL);
2687 }
2688 
2689 void
2690 umem_cache_destroy(umem_cache_t *cp)
2691 {
2692 	int cpu_seqid;
2693 
2694 	/*
2695 	 * Remove the cache from the global cache list so that no new updates
2696 	 * will be scheduled on its behalf, wait for any pending tasks to
2697 	 * complete, purge the cache, and then destroy it.
2698 	 */
2699 	(void) mutex_lock(&umem_cache_lock);
2700 	cp->cache_prev->cache_next = cp->cache_next;
2701 	cp->cache_next->cache_prev = cp->cache_prev;
2702 	cp->cache_prev = cp->cache_next = NULL;
2703 	(void) mutex_unlock(&umem_cache_lock);
2704 
2705 	umem_remove_updates(cp);
2706 
2707 	umem_cache_magazine_purge(cp);
2708 
2709 	(void) mutex_lock(&cp->cache_lock);
2710 	if (cp->cache_buftotal != 0)
2711 		log_message("umem_cache_destroy: '%s' (%p) not empty\n",
2712 		    cp->cache_name, (void *)cp);
2713 	cp->cache_reclaim = NULL;
2714 	/*
2715 	 * The cache is now dead.  There should be no further activity.
2716 	 * We enforce this by setting land mines in the constructor and
2717 	 * destructor routines that induce a segmentation fault if invoked.
2718 	 */
2719 	cp->cache_constructor = (umem_constructor_t *)1;
2720 	cp->cache_destructor = (umem_destructor_t *)2;
2721 	(void) mutex_unlock(&cp->cache_lock);
2722 
2723 	if (cp->cache_hash_table != NULL)
2724 		vmem_free(umem_hash_arena, cp->cache_hash_table,
2725 		    (cp->cache_hash_mask + 1) * sizeof (void *));
2726 
2727 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++)
2728 		(void) mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
2729 
2730 	(void) mutex_destroy(&cp->cache_depot_lock);
2731 	(void) mutex_destroy(&cp->cache_lock);
2732 
2733 	vmem_free(umem_cache_arena, cp, UMEM_CACHE_SIZE(umem_max_ncpus));
2734 }
2735 
2736 void
2737 umem_alloc_sizes_clear(void)
2738 {
2739 	int i;
2740 
2741 	umem_alloc_sizes[0] = UMEM_MAXBUF;
2742 	for (i = 1; i < NUM_ALLOC_SIZES; i++)
2743 		umem_alloc_sizes[i] = 0;
2744 }
2745 
2746 void
2747 umem_alloc_sizes_add(size_t size_arg)
2748 {
2749 	int i, j;
2750 	size_t size = size_arg;
2751 
2752 	if (size == 0) {
2753 		log_message("size_add: cannot add zero-sized cache\n",
2754 		    size, UMEM_MAXBUF);
2755 		return;
2756 	}
2757 
2758 	if (size > UMEM_MAXBUF) {
2759 		log_message("size_add: %ld > %d, cannot add\n", size,
2760 		    UMEM_MAXBUF);
2761 		return;
2762 	}
2763 
2764 	if (umem_alloc_sizes[NUM_ALLOC_SIZES - 1] != 0) {
2765 		log_message("size_add: no space in alloc_table for %d\n",
2766 		    size);
2767 		return;
2768 	}
2769 
2770 	if (P2PHASE(size, UMEM_ALIGN) != 0) {
2771 		size = P2ROUNDUP(size, UMEM_ALIGN);
2772 		log_message("size_add: rounding %d up to %d\n", size_arg,
2773 		    size);
2774 	}
2775 
2776 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
2777 		int cur = umem_alloc_sizes[i];
2778 		if (cur == size) {
2779 			log_message("size_add: %ld already in table\n",
2780 			    size);
2781 			return;
2782 		}
2783 		if (cur > size)
2784 			break;
2785 	}
2786 
2787 	for (j = NUM_ALLOC_SIZES - 1; j > i; j--)
2788 		umem_alloc_sizes[j] = umem_alloc_sizes[j-1];
2789 	umem_alloc_sizes[i] = size;
2790 }
2791 
2792 void
2793 umem_alloc_sizes_remove(size_t size)
2794 {
2795 	int i;
2796 
2797 	if (size == UMEM_MAXBUF) {
2798 		log_message("size_remove: cannot remove %ld\n", size);
2799 		return;
2800 	}
2801 
2802 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
2803 		int cur = umem_alloc_sizes[i];
2804 		if (cur == size)
2805 			break;
2806 		else if (cur > size || cur == 0) {
2807 			log_message("size_remove: %ld not found in table\n",
2808 			    size);
2809 			return;
2810 		}
2811 	}
2812 
2813 	for (; i + 1 < NUM_ALLOC_SIZES; i++)
2814 		umem_alloc_sizes[i] = umem_alloc_sizes[i+1];
2815 	umem_alloc_sizes[i] = 0;
2816 }
2817 
2818 static int
2819 umem_cache_init(void)
2820 {
2821 	int i;
2822 	size_t size, max_size;
2823 	umem_cache_t *cp;
2824 	umem_magtype_t *mtp;
2825 	char name[UMEM_CACHE_NAMELEN + 1];
2826 	umem_cache_t *umem_alloc_caches[NUM_ALLOC_SIZES];
2827 
2828 	for (i = 0; i < sizeof (umem_magtype) / sizeof (*mtp); i++) {
2829 		mtp = &umem_magtype[i];
2830 		(void) snprintf(name, sizeof (name), "umem_magazine_%d",
2831 		    mtp->mt_magsize);
2832 		mtp->mt_cache = umem_cache_create(name,
2833 		    (mtp->mt_magsize + 1) * sizeof (void *),
2834 		    mtp->mt_align, NULL, NULL, NULL, NULL,
2835 		    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
2836 		if (mtp->mt_cache == NULL)
2837 			return (0);
2838 	}
2839 
2840 	umem_slab_cache = umem_cache_create("umem_slab_cache",
2841 	    sizeof (umem_slab_t), 0, NULL, NULL, NULL, NULL,
2842 	    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
2843 
2844 	if (umem_slab_cache == NULL)
2845 		return (0);
2846 
2847 	umem_bufctl_cache = umem_cache_create("umem_bufctl_cache",
2848 	    sizeof (umem_bufctl_t), 0, NULL, NULL, NULL, NULL,
2849 	    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
2850 
2851 	if (umem_bufctl_cache == NULL)
2852 		return (0);
2853 
2854 	/*
2855 	 * The size of the umem_bufctl_audit structure depends upon
2856 	 * umem_stack_depth.   See umem_impl.h for details on the size
2857 	 * restrictions.
2858 	 */
2859 
2860 	size = UMEM_BUFCTL_AUDIT_SIZE_DEPTH(umem_stack_depth);
2861 	max_size = UMEM_BUFCTL_AUDIT_MAX_SIZE;
2862 
2863 	if (size > max_size) {			/* too large -- truncate */
2864 		int max_frames = UMEM_MAX_STACK_DEPTH;
2865 
2866 		ASSERT(UMEM_BUFCTL_AUDIT_SIZE_DEPTH(max_frames) <= max_size);
2867 
2868 		umem_stack_depth = max_frames;
2869 		size = UMEM_BUFCTL_AUDIT_SIZE_DEPTH(umem_stack_depth);
2870 	}
2871 
2872 	umem_bufctl_audit_cache = umem_cache_create("umem_bufctl_audit_cache",
2873 	    size, 0, NULL, NULL, NULL, NULL, umem_internal_arena,
2874 	    UMC_NOHASH | UMC_INTERNAL);
2875 
2876 	if (umem_bufctl_audit_cache == NULL)
2877 		return (0);
2878 
2879 	if (vmem_backend & VMEM_BACKEND_MMAP)
2880 		umem_va_arena = vmem_create("umem_va",
2881 		    NULL, 0, pagesize,
2882 		    vmem_alloc, vmem_free, heap_arena,
2883 		    8 * pagesize, VM_NOSLEEP);
2884 	else
2885 		umem_va_arena = heap_arena;
2886 
2887 	if (umem_va_arena == NULL)
2888 		return (0);
2889 
2890 	umem_default_arena = vmem_create("umem_default",
2891 	    NULL, 0, pagesize,
2892 	    heap_alloc, heap_free, umem_va_arena,
2893 	    0, VM_NOSLEEP);
2894 
2895 	if (umem_default_arena == NULL)
2896 		return (0);
2897 
2898 	/*
2899 	 * make sure the umem_alloc table initializer is correct
2900 	 */
2901 	i = sizeof (umem_alloc_table) / sizeof (*umem_alloc_table);
2902 	ASSERT(umem_alloc_table[i - 1] == &umem_null_cache);
2903 
2904 	/*
2905 	 * Create the default caches to back umem_alloc()
2906 	 */
2907 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
2908 		size_t cache_size = umem_alloc_sizes[i];
2909 		size_t align = 0;
2910 
2911 		if (cache_size == 0)
2912 			break;		/* 0 terminates the list */
2913 
2914 		/*
2915 		 * If they allocate a multiple of the coherency granularity,
2916 		 * they get a coherency-granularity-aligned address.
2917 		 */
2918 		if (IS_P2ALIGNED(cache_size, 64))
2919 			align = 64;
2920 		if (IS_P2ALIGNED(cache_size, pagesize))
2921 			align = pagesize;
2922 		(void) snprintf(name, sizeof (name), "umem_alloc_%lu",
2923 		    (long)cache_size);
2924 
2925 		cp = umem_cache_create(name, cache_size, align,
2926 		    NULL, NULL, NULL, NULL, NULL, UMC_INTERNAL);
2927 		if (cp == NULL)
2928 			return (0);
2929 
2930 		umem_alloc_caches[i] = cp;
2931 	}
2932 
2933 	/*
2934 	 * Initialization cannot fail at this point.  Make the caches
2935 	 * visible to umem_alloc() and friends.
2936 	 */
2937 	size = UMEM_ALIGN;
2938 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
2939 		size_t cache_size = umem_alloc_sizes[i];
2940 
2941 		if (cache_size == 0)
2942 			break;		/* 0 terminates the list */
2943 
2944 		cp = umem_alloc_caches[i];
2945 
2946 		while (size <= cache_size) {
2947 			umem_alloc_table[(size - 1) >> UMEM_ALIGN_SHIFT] = cp;
2948 			size += UMEM_ALIGN;
2949 		}
2950 	}
2951 	ASSERT(size - UMEM_ALIGN == UMEM_MAXBUF);
2952 	return (1);
2953 }
2954 
2955 /*
2956  * umem_startup() is called early on, and must be called explicitly if we're
2957  * the standalone version.
2958  */
2959 #ifdef UMEM_STANDALONE
2960 void
2961 #else
2962 #pragma init(umem_startup)
2963 static void
2964 #endif
2965 umem_startup(caddr_t start, size_t len, size_t pagesize, caddr_t minstack,
2966     caddr_t maxstack)
2967 {
2968 #ifdef UMEM_STANDALONE
2969 	int idx;
2970 	/* Standalone doesn't fork */
2971 #else
2972 	umem_forkhandler_init(); /* register the fork handler */
2973 #endif
2974 
2975 #ifdef __lint
2976 	/* make lint happy */
2977 	minstack = maxstack;
2978 #endif
2979 
2980 #ifdef UMEM_STANDALONE
2981 	umem_ready = UMEM_READY_STARTUP;
2982 	umem_init_env_ready = 0;
2983 
2984 	umem_min_stack = minstack;
2985 	umem_max_stack = maxstack;
2986 
2987 	nofail_callback = NULL;
2988 	umem_slab_cache = NULL;
2989 	umem_bufctl_cache = NULL;
2990 	umem_bufctl_audit_cache = NULL;
2991 	heap_arena = NULL;
2992 	heap_alloc = NULL;
2993 	heap_free = NULL;
2994 	umem_internal_arena = NULL;
2995 	umem_cache_arena = NULL;
2996 	umem_hash_arena = NULL;
2997 	umem_log_arena = NULL;
2998 	umem_oversize_arena = NULL;
2999 	umem_va_arena = NULL;
3000 	umem_default_arena = NULL;
3001 	umem_firewall_va_arena = NULL;
3002 	umem_firewall_arena = NULL;
3003 	umem_memalign_arena = NULL;
3004 	umem_transaction_log = NULL;
3005 	umem_content_log = NULL;
3006 	umem_failure_log = NULL;
3007 	umem_slab_log = NULL;
3008 	umem_cpu_mask = 0;
3009 
3010 	umem_cpus = &umem_startup_cpu;
3011 	umem_startup_cpu.cpu_cache_offset = UMEM_CACHE_SIZE(0);
3012 	umem_startup_cpu.cpu_number = 0;
3013 
3014 	bcopy(&umem_null_cache_template, &umem_null_cache,
3015 	    sizeof (umem_cache_t));
3016 
3017 	for (idx = 0; idx < (UMEM_MAXBUF >> UMEM_ALIGN_SHIFT); idx++)
3018 		umem_alloc_table[idx] = &umem_null_cache;
3019 #endif
3020 
3021 	/*
3022 	 * Perform initialization specific to the way we've been compiled
3023 	 * (library or standalone)
3024 	 */
3025 	umem_type_init(start, len, pagesize);
3026 
3027 	vmem_startup();
3028 }
3029 
3030 int
3031 umem_init(void)
3032 {
3033 	size_t maxverify, minfirewall;
3034 	size_t size;
3035 	int idx;
3036 	umem_cpu_t *new_cpus;
3037 
3038 	vmem_t *memalign_arena, *oversize_arena;
3039 
3040 	if (thr_self() != umem_init_thr) {
3041 		/*
3042 		 * The usual case -- non-recursive invocation of umem_init().
3043 		 */
3044 		(void) mutex_lock(&umem_init_lock);
3045 		if (umem_ready != UMEM_READY_STARTUP) {
3046 			/*
3047 			 * someone else beat us to initializing umem.  Wait
3048 			 * for them to complete, then return.
3049 			 */
3050 			while (umem_ready == UMEM_READY_INITING)
3051 				(void) _cond_wait(&umem_init_cv,
3052 				    &umem_init_lock);
3053 			ASSERT(umem_ready == UMEM_READY ||
3054 			    umem_ready == UMEM_READY_INIT_FAILED);
3055 			(void) mutex_unlock(&umem_init_lock);
3056 			return (umem_ready == UMEM_READY);
3057 		}
3058 
3059 		ASSERT(umem_ready == UMEM_READY_STARTUP);
3060 		ASSERT(umem_init_env_ready == 0);
3061 
3062 		umem_ready = UMEM_READY_INITING;
3063 		umem_init_thr = thr_self();
3064 
3065 		(void) mutex_unlock(&umem_init_lock);
3066 		umem_setup_envvars(0);		/* can recurse -- see below */
3067 		if (umem_init_env_ready) {
3068 			/*
3069 			 * initialization was completed already
3070 			 */
3071 			ASSERT(umem_ready == UMEM_READY ||
3072 			    umem_ready == UMEM_READY_INIT_FAILED);
3073 			ASSERT(umem_init_thr == 0);
3074 			return (umem_ready == UMEM_READY);
3075 		}
3076 	} else if (!umem_init_env_ready) {
3077 		/*
3078 		 * The umem_setup_envvars() call (above) makes calls into
3079 		 * the dynamic linker and directly into user-supplied code.
3080 		 * Since we cannot know what that code will do, we could be
3081 		 * recursively invoked (by, say, a malloc() call in the code
3082 		 * itself, or in a (C++) _init section it causes to be fired).
3083 		 *
3084 		 * This code is where we end up if such recursion occurs.  We
3085 		 * first clean up any partial results in the envvar code, then
3086 		 * proceed to finish initialization processing in the recursive
3087 		 * call.  The original call will notice this, and return
3088 		 * immediately.
3089 		 */
3090 		umem_setup_envvars(1);		/* clean up any partial state */
3091 	} else {
3092 		umem_panic(
3093 		    "recursive allocation while initializing umem\n");
3094 	}
3095 	umem_init_env_ready = 1;
3096 
3097 	/*
3098 	 * From this point until we finish, recursion into umem_init() will
3099 	 * cause a umem_panic().
3100 	 */
3101 	maxverify = minfirewall = ULONG_MAX;
3102 
3103 	/* LINTED constant condition */
3104 	if (sizeof (umem_cpu_cache_t) != UMEM_CPU_CACHE_SIZE) {
3105 		umem_panic("sizeof (umem_cpu_cache_t) = %d, should be %d\n",
3106 		    sizeof (umem_cpu_cache_t), UMEM_CPU_CACHE_SIZE);
3107 	}
3108 
3109 	umem_max_ncpus = umem_get_max_ncpus();
3110 
3111 	/*
3112 	 * load tunables from environment
3113 	 */
3114 	umem_process_envvars();
3115 
3116 	if (issetugid())
3117 		umem_mtbf = 0;
3118 
3119 	/*
3120 	 * set up vmem
3121 	 */
3122 	if (!(umem_flags & UMF_AUDIT))
3123 		vmem_no_debug();
3124 
3125 	heap_arena = vmem_heap_arena(&heap_alloc, &heap_free);
3126 
3127 	pagesize = heap_arena->vm_quantum;
3128 
3129 	umem_internal_arena = vmem_create("umem_internal", NULL, 0, pagesize,
3130 	    heap_alloc, heap_free, heap_arena, 0, VM_NOSLEEP);
3131 
3132 	umem_default_arena = umem_internal_arena;
3133 
3134 	if (umem_internal_arena == NULL)
3135 		goto fail;
3136 
3137 	umem_cache_arena = vmem_create("umem_cache", NULL, 0, UMEM_ALIGN,
3138 	    vmem_alloc, vmem_free, umem_internal_arena, 0, VM_NOSLEEP);
3139 
3140 	umem_hash_arena = vmem_create("umem_hash", NULL, 0, UMEM_ALIGN,
3141 	    vmem_alloc, vmem_free, umem_internal_arena, 0, VM_NOSLEEP);
3142 
3143 	umem_log_arena = vmem_create("umem_log", NULL, 0, UMEM_ALIGN,
3144 	    heap_alloc, heap_free, heap_arena, 0, VM_NOSLEEP);
3145 
3146 	umem_firewall_va_arena = vmem_create("umem_firewall_va",
3147 	    NULL, 0, pagesize,
3148 	    umem_firewall_va_alloc, umem_firewall_va_free, heap_arena,
3149 	    0, VM_NOSLEEP);
3150 
3151 	if (umem_cache_arena == NULL || umem_hash_arena == NULL ||
3152 	    umem_log_arena == NULL || umem_firewall_va_arena == NULL)
3153 		goto fail;
3154 
3155 	umem_firewall_arena = vmem_create("umem_firewall", NULL, 0, pagesize,
3156 	    heap_alloc, heap_free, umem_firewall_va_arena, 0,
3157 	    VM_NOSLEEP);
3158 
3159 	if (umem_firewall_arena == NULL)
3160 		goto fail;
3161 
3162 	oversize_arena = vmem_create("umem_oversize", NULL, 0, pagesize,
3163 	    heap_alloc, heap_free, minfirewall < ULONG_MAX ?
3164 	    umem_firewall_va_arena : heap_arena, 0, VM_NOSLEEP);
3165 
3166 	memalign_arena = vmem_create("umem_memalign", NULL, 0, UMEM_ALIGN,
3167 	    heap_alloc, heap_free, minfirewall < ULONG_MAX ?
3168 	    umem_firewall_va_arena : heap_arena, 0, VM_NOSLEEP);
3169 
3170 	if (oversize_arena == NULL || memalign_arena == NULL)
3171 		goto fail;
3172 
3173 	if (umem_max_ncpus > CPUHINT_MAX())
3174 		umem_max_ncpus = CPUHINT_MAX();
3175 
3176 	while ((umem_max_ncpus & (umem_max_ncpus - 1)) != 0)
3177 		umem_max_ncpus++;
3178 
3179 	if (umem_max_ncpus == 0)
3180 		umem_max_ncpus = 1;
3181 
3182 	size = umem_max_ncpus * sizeof (umem_cpu_t);
3183 	new_cpus = vmem_alloc(umem_internal_arena, size, VM_NOSLEEP);
3184 	if (new_cpus == NULL)
3185 		goto fail;
3186 
3187 	bzero(new_cpus, size);
3188 	for (idx = 0; idx < umem_max_ncpus; idx++) {
3189 		new_cpus[idx].cpu_number = idx;
3190 		new_cpus[idx].cpu_cache_offset = UMEM_CACHE_SIZE(idx);
3191 	}
3192 	umem_cpus = new_cpus;
3193 	umem_cpu_mask = (umem_max_ncpus - 1);
3194 
3195 	if (umem_maxverify == 0)
3196 		umem_maxverify = maxverify;
3197 
3198 	if (umem_minfirewall == 0)
3199 		umem_minfirewall = minfirewall;
3200 
3201 	/*
3202 	 * Set up updating and reaping
3203 	 */
3204 	umem_reap_next = gethrtime() + NANOSEC;
3205 
3206 #ifndef UMEM_STANDALONE
3207 	(void) gettimeofday(&umem_update_next, NULL);
3208 #endif
3209 
3210 	/*
3211 	 * Set up logging -- failure here is okay, since it will just disable
3212 	 * the logs
3213 	 */
3214 	if (umem_logging) {
3215 		umem_transaction_log = umem_log_init(umem_transaction_log_size);
3216 		umem_content_log = umem_log_init(umem_content_log_size);
3217 		umem_failure_log = umem_log_init(umem_failure_log_size);
3218 		umem_slab_log = umem_log_init(umem_slab_log_size);
3219 	}
3220 
3221 	/*
3222 	 * Set up caches -- if successful, initialization cannot fail, since
3223 	 * allocations from other threads can now succeed.
3224 	 */
3225 	if (umem_cache_init() == 0) {
3226 		log_message("unable to create initial caches\n");
3227 		goto fail;
3228 	}
3229 	umem_oversize_arena = oversize_arena;
3230 	umem_memalign_arena = memalign_arena;
3231 
3232 	umem_cache_applyall(umem_cache_magazine_enable);
3233 
3234 	/*
3235 	 * initialization done, ready to go
3236 	 */
3237 	(void) mutex_lock(&umem_init_lock);
3238 	umem_ready = UMEM_READY;
3239 	umem_init_thr = 0;
3240 	(void) cond_broadcast(&umem_init_cv);
3241 	(void) mutex_unlock(&umem_init_lock);
3242 	return (1);
3243 
3244 fail:
3245 	log_message("umem initialization failed\n");
3246 
3247 	(void) mutex_lock(&umem_init_lock);
3248 	umem_ready = UMEM_READY_INIT_FAILED;
3249 	umem_init_thr = 0;
3250 	(void) cond_broadcast(&umem_init_cv);
3251 	(void) mutex_unlock(&umem_init_lock);
3252 	return (0);
3253 }
3254