xref: /titanic_52/usr/src/lib/libumem/common/umem.c (revision bdfc6d18da790deeec2e0eb09c625902defe2498)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2004 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  *
59  * 2. Initialization
60  * -----------------
61  * kmem is initialized early on in boot, and knows that no one will call
62  * into it before it is ready.  umem does not have these luxuries. Instead,
63  * initialization is divided into two phases:
64  *
65  *	* library initialization, and
66  *
67  *	* first use
68  *
69  * umem's full initialization happens at the time of the first allocation
70  * request (via malloc() and friends, umem_alloc(), or umem_zalloc()),
71  * or the first call to umem_cache_create().
72  *
73  * umem_free(), and umem_cache_alloc() do not require special handling,
74  * since the only way to get valid arguments for them is to successfully
75  * call a function from the first group.
76  *
77  * 2.1. Library Initialization: umem_startup()
78  * -------------------------------------------
79  * umem_startup() is libumem.so's .init section.  It calls pthread_atfork()
80  * to install the handlers necessary for umem's Fork1-Safety.  Because of
81  * race condition issues, all other pre-umem_init() initialization is done
82  * statically (i.e. by the dynamic linker).
83  *
84  * For standalone use, umem_startup() returns everything to its initial
85  * state.
86  *
87  * 2.2. First use: umem_init()
88  * ------------------------------
89  * The first time any memory allocation function is used, we have to
90  * create the backing caches and vmem arenas which are needed for it.
91  * umem_init() is the central point for that task.  When it completes,
92  * umem_ready is either UMEM_READY (all set) or UMEM_READY_INIT_FAILED (unable
93  * to initialize, probably due to lack of memory).
94  *
95  * There are four different paths from which umem_init() is called:
96  *
97  *	* from umem_alloc() or umem_zalloc(), with 0 < size < UMEM_MAXBUF,
98  *
99  *	* from umem_alloc() or umem_zalloc(), with size > UMEM_MAXBUF,
100  *
101  *	* from umem_cache_create(), and
102  *
103  *	* from memalign(), with align > UMEM_ALIGN.
104  *
105  * The last three just check if umem is initialized, and call umem_init()
106  * if it is not.  For performance reasons, the first case is more complicated.
107  *
108  * 2.2.1. umem_alloc()/umem_zalloc(), with 0 < size < UMEM_MAXBUF
109  * -----------------------------------------------------------------
110  * In this case, umem_cache_alloc(&umem_null_cache, ...) is called.
111  * There is special case code in which causes any allocation on
112  * &umem_null_cache to fail by returning (NULL), regardless of the
113  * flags argument.
114  *
115  * So umem_cache_alloc() returns NULL, and umem_alloc()/umem_zalloc() call
116  * umem_alloc_retry().  umem_alloc_retry() sees that the allocation
117  * was agains &umem_null_cache, and calls umem_init().
118  *
119  * If initialization is successful, umem_alloc_retry() returns 1, which
120  * causes umem_alloc()/umem_zalloc() to start over, which causes it to load
121  * the (now valid) cache pointer from umem_alloc_table.
122  *
123  * 2.2.2. Dealing with race conditions
124  * -----------------------------------
125  * There are a couple race conditions resulting from the initialization
126  * code that we have to guard against:
127  *
128  *	* In umem_cache_create(), there is a special UMC_INTERNAL cflag
129  *	that is passed for caches created during initialization.  It
130  *	is illegal for a user to try to create a UMC_INTERNAL cache.
131  *	This allows initialization to proceed, but any other
132  *	umem_cache_create()s will block by calling umem_init().
133  *
134  *	* Since umem_null_cache has a 1-element cache_cpu, it's cache_cpu_mask
135  *	is always zero.  umem_cache_alloc uses cp->cache_cpu_mask to
136  *	mask the cpu number.  This prevents a race between grabbing a
137  *	cache pointer out of umem_alloc_table and growing the cpu array.
138  *
139  *
140  * 3. CPU handling
141  * ---------------
142  * kmem uses the CPU's sequence number to determine which "cpu cache" to
143  * use for an allocation.  Currently, there is no way to get the sequence
144  * number in userspace.
145  *
146  * umem keeps track of cpu information in umem_cpus, an array of umem_max_ncpus
147  * umem_cpu_t structures.  CURCPU() is a a "hint" function, which we then mask
148  * with either umem_cpu_mask or cp->cache_cpu_mask to find the actual "cpu" id.
149  * The mechanics of this is all in the CPU(mask) macro.
150  *
151  * Currently, umem uses _lwp_self() as its hint.
152  *
153  *
154  * 4. The update thread
155  * --------------------
156  * kmem uses a task queue, kmem_taskq, to do periodic maintenance on
157  * every kmem cache.  vmem has a periodic timeout for hash table resizing.
158  * The kmem_taskq also provides a separate context for kmem_cache_reap()'s
159  * to be done in, avoiding issues of the context of kmem_reap() callers.
160  *
161  * Instead, umem has the concept of "updates", which are asynchronous requests
162  * for work attached to single caches.  All caches with pending work are
163  * on a doubly linked list rooted at the umem_null_cache.  All update state
164  * is protected by the umem_update_lock mutex, and the umem_update_cv is used
165  * for notification between threads.
166  *
167  * 4.1. Cache states with regards to updates
168  * -----------------------------------------
169  * A given cache is in one of three states:
170  *
171  * Inactive		cache_uflags is zero, cache_u{next,prev} are NULL
172  *
173  * Work Requested	cache_uflags is non-zero (but UMU_ACTIVE is not set),
174  *			cache_u{next,prev} link the cache onto the global
175  *			update list
176  *
177  * Active		cache_uflags has UMU_ACTIVE set, cache_u{next,prev}
178  *			are NULL, and either umem_update_thr or
179  *			umem_st_update_thr are actively doing work on the
180  *			cache.
181  *
182  * An update can be added to any cache in any state -- if the cache is
183  * Inactive, it transitions to being Work Requested.  If the cache is
184  * Active, the worker will notice the new update and act on it before
185  * transitioning the cache to the Inactive state.
186  *
187  * If a cache is in the Active state, UMU_NOTIFY can be set, which asks
188  * the worker to broadcast the umem_update_cv when it has finished.
189  *
190  * 4.2. Update interface
191  * ---------------------
192  * umem_add_update() adds an update to a particular cache.
193  * umem_updateall() adds an update to all caches.
194  * umem_remove_updates() returns a cache to the Inactive state.
195  *
196  * umem_process_updates() process all caches in the Work Requested state.
197  *
198  * 4.3. Reaping
199  * ------------
200  * When umem_reap() is called (at the time of heap growth), it schedule
201  * UMU_REAP updates on every cache.  It then checks to see if the update
202  * thread exists (umem_update_thr != 0).  If it is, it broadcasts
203  * the umem_update_cv to wake the update thread up, and returns.
204  *
205  * If the update thread does not exist (umem_update_thr == 0), and the
206  * program currently has multiple threads, umem_reap() attempts to create
207  * a new update thread.
208  *
209  * If the process is not multithreaded, or the creation fails, umem_reap()
210  * calls umem_st_update() to do an inline update.
211  *
212  * 4.4. The update thread
213  * ----------------------
214  * The update thread spends most of its time in cond_timedwait() on the
215  * umem_update_cv.  It wakes up under two conditions:
216  *
217  *	* The timedwait times out, in which case it needs to run a global
218  *	update, or
219  *
220  *	* someone cond_broadcast(3THR)s the umem_update_cv, in which case
221  *	it needs to check if there are any caches in the Work Requested
222  *	state.
223  *
224  * When it is time for another global update, umem calls umem_cache_update()
225  * on every cache, then calls vmem_update(), which tunes the vmem structures.
226  * umem_cache_update() can request further work using umem_add_update().
227  *
228  * After any work from the global update completes, the update timer is
229  * reset to umem_reap_interval seconds in the future.  This makes the
230  * updates self-throttling.
231  *
232  * Reaps are similarly self-throttling.  After a UMU_REAP update has
233  * been scheduled on all caches, umem_reap() sets a flag and wakes up the
234  * update thread.  The update thread notices the flag, and resets the
235  * reap state.
236  *
237  * 4.5. Inline updates
238  * -------------------
239  * If the update thread is not running, umem_st_update() is used instead.  It
240  * immediately does a global update (as above), then calls
241  * umem_process_updates() to process both the reaps that umem_reap() added and
242  * any work generated by the global update.  Afterwards, it resets the reap
243  * state.
244  *
245  * While the umem_st_update() is running, umem_st_update_thr holds the thread
246  * id of the thread performing the update.
247  *
248  * 4.6. Updates and fork1()
249  * ------------------------
250  * umem has fork1() pre- and post-handlers which lock up (and release) every
251  * mutex in every cache.  They also lock up the umem_update_lock.  Since
252  * fork1() only copies over a single lwp, other threads (including the update
253  * thread) could have been actively using a cache in the parent.  This
254  * can lead to inconsistencies in the child process.
255  *
256  * Because we locked all of the mutexes, the only possible inconsistancies are:
257  *
258  *	* a umem_cache_alloc() could leak its buffer.
259  *
260  *	* a caller of umem_depot_alloc() could leak a magazine, and all the
261  *	buffers contained in it.
262  *
263  *	* a cache could be in the Active update state.  In the child, there
264  *	would be no thread actually working on it.
265  *
266  *	* a umem_hash_rescale() could leak the new hash table.
267  *
268  *	* a umem_magazine_resize() could be in progress.
269  *
270  *	* a umem_reap() could be in progress.
271  *
272  * The memory leaks we can't do anything about.  umem_release_child() resets
273  * the update state, moves any caches in the Active state to the Work Requested
274  * state.  This might cause some updates to be re-run, but UMU_REAP and
275  * UMU_HASH_RESCALE are effectively idempotent, and the worst that can
276  * happen from umem_magazine_resize() is resizing the magazine twice in close
277  * succession.
278  *
279  * Much of the cleanup in umem_release_child() is skipped if
280  * umem_st_update_thr == thr_self().  This is so that applications which call
281  * fork1() from a cache callback does not break.  Needless to say, any such
282  * application is tremendously broken.
283  *
284  *
285  * 5. KM_SLEEP v.s. UMEM_NOFAIL
286  * ----------------------------
287  * Allocations against kmem and vmem have two basic modes:  SLEEP and
288  * NOSLEEP.  A sleeping allocation is will go to sleep (waiting for
289  * more memory) instead of failing (returning NULL).
290  *
291  * SLEEP allocations presume an extremely multithreaded model, with
292  * a lot of allocation and deallocation activity.  umem cannot presume
293  * that its clients have any particular type of behavior.  Instead,
294  * it provides two types of allocations:
295  *
296  *	* UMEM_DEFAULT, equivalent to KM_NOSLEEP (i.e. return NULL on
297  *	failure)
298  *
299  *	* UMEM_NOFAIL, which, on failure, calls an optional callback
300  *	(registered with umem_nofail_callback()).
301  *
302  * The callback is invoked with no locks held, and can do an arbitrary
303  * amount of work.  It then has a choice between:
304  *
305  *	* Returning UMEM_CALLBACK_RETRY, which will cause the allocation
306  *	to be restarted.
307  *
308  *	* Returning UMEM_CALLBACK_EXIT(status), which will cause exit(2)
309  *	to be invoked with status.  If multiple threads attempt to do
310  *	this simultaneously, only one will call exit(2).
311  *
312  *	* Doing some kind of non-local exit (thr_exit(3thr), longjmp(3C),
313  *	etc.)
314  *
315  * The default callback returns UMEM_CALLBACK_EXIT(255).
316  *
317  * To have these callbacks without risk of state corruption (in the case of
318  * a non-local exit), we have to ensure that the callbacks get invoked
319  * close to the original allocation, with no inconsistent state or held
320  * locks.  The following steps are taken:
321  *
322  *	* All invocations of vmem are VM_NOSLEEP.
323  *
324  *	* All constructor callbacks (which can themselves to allocations)
325  *	are passed UMEM_DEFAULT as their required allocation argument.  This
326  *	way, the constructor will fail, allowing the highest-level allocation
327  *	invoke the nofail callback.
328  *
329  *	If a constructor callback _does_ do a UMEM_NOFAIL allocation, and
330  *	the nofail callback does a non-local exit, we will leak the
331  *	partially-constructed buffer.
332  */
333 
334 #include "mtlib.h"
335 #include <umem_impl.h>
336 #include <sys/vmem_impl_user.h>
337 #include "umem_base.h"
338 #include "vmem_base.h"
339 
340 #include <sys/processor.h>
341 #include <sys/sysmacros.h>
342 
343 #include <alloca.h>
344 #include <errno.h>
345 #include <limits.h>
346 #include <stdio.h>
347 #include <stdlib.h>
348 #include <string.h>
349 #include <strings.h>
350 #include <signal.h>
351 #include <unistd.h>
352 #include <atomic.h>
353 
354 #include "misc.h"
355 
356 #define	UMEM_VMFLAGS(umflag)	(VM_NOSLEEP)
357 
358 size_t pagesize;
359 
360 /*
361  * The default set of caches to back umem_alloc().
362  * These sizes should be reevaluated periodically.
363  *
364  * We want allocations that are multiples of the coherency granularity
365  * (64 bytes) to be satisfied from a cache which is a multiple of 64
366  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
367  * the next kmem_cache_size greater than or equal to it must be a
368  * multiple of 64.
369  */
370 static const int umem_alloc_sizes[] = {
371 #ifdef _LP64
372 	1 * 8,
373 	1 * 16,
374 	2 * 16,
375 	3 * 16,
376 #else
377 	1 * 8,
378 	2 * 8,
379 	3 * 8,
380 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
381 #endif
382 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
383 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
384 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
385 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
386 	P2ALIGN(8192 / 7, 64),
387 	P2ALIGN(8192 / 6, 64),
388 	P2ALIGN(8192 / 5, 64),
389 	P2ALIGN(8192 / 4, 64),
390 	P2ALIGN(8192 / 3, 64),
391 	P2ALIGN(8192 / 2, 64),
392 	P2ALIGN(8192 / 1, 64),
393 	4096 * 3,
394 	8192 * 2,
395 };
396 #define	NUM_ALLOC_SIZES (sizeof (umem_alloc_sizes) / sizeof (*umem_alloc_sizes))
397 
398 #define	UMEM_MAXBUF	16384
399 
400 static umem_magtype_t umem_magtype[] = {
401 	{ 1,	8,	3200,	65536	},
402 	{ 3,	16,	256,	32768	},
403 	{ 7,	32,	64,	16384	},
404 	{ 15,	64,	0,	8192	},
405 	{ 31,	64,	0,	4096	},
406 	{ 47,	64,	0,	2048	},
407 	{ 63,	64,	0,	1024	},
408 	{ 95,	64,	0,	512	},
409 	{ 143,	64,	0,	0	},
410 };
411 
412 /*
413  * umem tunables
414  */
415 uint32_t umem_max_ncpus;	/* # of CPU caches. */
416 
417 uint32_t umem_stack_depth = 15; /* # stack frames in a bufctl_audit */
418 uint32_t umem_reap_interval = 10; /* max reaping rate (seconds) */
419 uint_t umem_depot_contention = 2; /* max failed trylocks per real interval */
420 uint_t umem_abort = 1;		/* whether to abort on error */
421 uint_t umem_output = 0;		/* whether to write to standard error */
422 uint_t umem_logging = 0;	/* umem_log_enter() override */
423 uint32_t umem_mtbf = 0;		/* mean time between failures [default: off] */
424 size_t umem_transaction_log_size; /* size of transaction log */
425 size_t umem_content_log_size;	/* size of content log */
426 size_t umem_failure_log_size;	/* failure log [4 pages per CPU] */
427 size_t umem_slab_log_size;	/* slab create log [4 pages per CPU] */
428 size_t umem_content_maxsave = 256; /* UMF_CONTENTS max bytes to log */
429 size_t umem_lite_minsize = 0;	/* minimum buffer size for UMF_LITE */
430 size_t umem_lite_maxalign = 1024; /* maximum buffer alignment for UMF_LITE */
431 size_t umem_maxverify;		/* maximum bytes to inspect in debug routines */
432 size_t umem_minfirewall;	/* hardware-enforced redzone threshold */
433 
434 uint_t umem_flags = 0;
435 
436 mutex_t			umem_init_lock;		/* locks initialization */
437 cond_t			umem_init_cv;		/* initialization CV */
438 thread_t		umem_init_thr;		/* thread initializing */
439 int			umem_init_env_ready;	/* environ pre-initted */
440 int			umem_ready = UMEM_READY_STARTUP;
441 
442 static umem_nofail_callback_t *nofail_callback;
443 static mutex_t		umem_nofail_exit_lock;
444 static thread_t		umem_nofail_exit_thr;
445 
446 static umem_cache_t	*umem_slab_cache;
447 static umem_cache_t	*umem_bufctl_cache;
448 static umem_cache_t	*umem_bufctl_audit_cache;
449 
450 mutex_t			umem_flags_lock;
451 
452 static vmem_t		*heap_arena;
453 static vmem_alloc_t	*heap_alloc;
454 static vmem_free_t	*heap_free;
455 
456 static vmem_t		*umem_internal_arena;
457 static vmem_t		*umem_cache_arena;
458 static vmem_t		*umem_hash_arena;
459 static vmem_t		*umem_log_arena;
460 static vmem_t		*umem_oversize_arena;
461 static vmem_t		*umem_va_arena;
462 static vmem_t		*umem_default_arena;
463 static vmem_t		*umem_firewall_va_arena;
464 static vmem_t		*umem_firewall_arena;
465 
466 vmem_t			*umem_memalign_arena;
467 
468 umem_log_header_t *umem_transaction_log;
469 umem_log_header_t *umem_content_log;
470 umem_log_header_t *umem_failure_log;
471 umem_log_header_t *umem_slab_log;
472 
473 extern thread_t _thr_self(void);
474 #define	CPUHINT()		(_thr_self())
475 #define	CPUHINT_MAX()		INT_MAX
476 
477 #define	CPU(mask)		(umem_cpus + (CPUHINT() & (mask)))
478 static umem_cpu_t umem_startup_cpu = {	/* initial, single, cpu */
479 	UMEM_CACHE_SIZE(0),
480 	0
481 };
482 
483 static uint32_t umem_cpu_mask = 0;			/* global cpu mask */
484 static umem_cpu_t *umem_cpus = &umem_startup_cpu;	/* cpu list */
485 
486 volatile uint32_t umem_reaping;
487 
488 thread_t		umem_update_thr;
489 struct timeval		umem_update_next;	/* timeofday of next update */
490 volatile thread_t	umem_st_update_thr;	/* only used when single-thd */
491 
492 #define	IN_UPDATE()	(thr_self() == umem_update_thr || \
493 			    thr_self() == umem_st_update_thr)
494 #define	IN_REAP()	IN_UPDATE()
495 
496 mutex_t			umem_update_lock;	/* cache_u{next,prev,flags} */
497 cond_t			umem_update_cv;
498 
499 volatile hrtime_t umem_reap_next;	/* min hrtime of next reap */
500 
501 mutex_t			umem_cache_lock;	/* inter-cache linkage only */
502 
503 #ifdef UMEM_STANDALONE
504 umem_cache_t		umem_null_cache;
505 static const umem_cache_t umem_null_cache_template = {
506 #else
507 umem_cache_t		umem_null_cache = {
508 #endif
509 	0, 0, 0, 0, 0,
510 	0, 0,
511 	0, 0,
512 	0, 0,
513 	"invalid_cache",
514 	0, 0,
515 	NULL, NULL, NULL, NULL,
516 	NULL,
517 	0, 0, 0, 0,
518 	&umem_null_cache, &umem_null_cache,
519 	&umem_null_cache, &umem_null_cache,
520 	0,
521 	DEFAULTMUTEX,				/* start of slab layer */
522 	0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
523 	&umem_null_cache.cache_nullslab,
524 	{
525 		&umem_null_cache,
526 		NULL,
527 		&umem_null_cache.cache_nullslab,
528 		&umem_null_cache.cache_nullslab,
529 		NULL,
530 		-1,
531 		0
532 	},
533 	NULL,
534 	NULL,
535 	DEFAULTMUTEX,				/* start of depot layer */
536 	NULL, {
537 		NULL, 0, 0, 0, 0
538 	}, {
539 		NULL, 0, 0, 0, 0
540 	}, {
541 		{
542 			DEFAULTMUTEX,		/* start of CPU cache */
543 			0, 0, NULL, NULL, -1, -1, 0
544 		}
545 	}
546 };
547 
548 #define	ALLOC_TABLE_4 \
549 	&umem_null_cache, &umem_null_cache, &umem_null_cache, &umem_null_cache
550 
551 #define	ALLOC_TABLE_64 \
552 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
553 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
554 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
555 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4
556 
557 #define	ALLOC_TABLE_1024 \
558 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
559 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
560 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
561 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64
562 
563 static umem_cache_t *umem_alloc_table[UMEM_MAXBUF >> UMEM_ALIGN_SHIFT] = {
564 	ALLOC_TABLE_1024,
565 	ALLOC_TABLE_1024
566 };
567 
568 
569 /* Used to constrain audit-log stack traces */
570 caddr_t			umem_min_stack;
571 caddr_t			umem_max_stack;
572 
573 
574 /*
575  * we use the _ versions, since we don't want to be cancelled.
576  * Actually, this is automatically taken care of by including "mtlib.h".
577  */
578 extern int _cond_wait(cond_t *cv, mutex_t *mutex);
579 
580 #define	UMERR_MODIFIED	0	/* buffer modified while on freelist */
581 #define	UMERR_REDZONE	1	/* redzone violation (write past end of buf) */
582 #define	UMERR_DUPFREE	2	/* freed a buffer twice */
583 #define	UMERR_BADADDR	3	/* freed a bad (unallocated) address */
584 #define	UMERR_BADBUFTAG	4	/* buftag corrupted */
585 #define	UMERR_BADBUFCTL	5	/* bufctl corrupted */
586 #define	UMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
587 #define	UMERR_BADSIZE	7	/* alloc size != free size */
588 #define	UMERR_BADBASE	8	/* buffer base address wrong */
589 
590 struct {
591 	hrtime_t	ump_timestamp;	/* timestamp of error */
592 	int		ump_error;	/* type of umem error (UMERR_*) */
593 	void		*ump_buffer;	/* buffer that induced abort */
594 	void		*ump_realbuf;	/* real start address for buffer */
595 	umem_cache_t	*ump_cache;	/* buffer's cache according to client */
596 	umem_cache_t	*ump_realcache;	/* actual cache containing buffer */
597 	umem_slab_t	*ump_slab;	/* slab accoring to umem_findslab() */
598 	umem_bufctl_t	*ump_bufctl;	/* bufctl */
599 } umem_abort_info;
600 
601 static void
602 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
603 {
604 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
605 	uint64_t *buf = buf_arg;
606 
607 	while (buf < bufend)
608 		*buf++ = pattern;
609 }
610 
611 static void *
612 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
613 {
614 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
615 	uint64_t *buf;
616 
617 	for (buf = buf_arg; buf < bufend; buf++)
618 		if (*buf != pattern)
619 			return (buf);
620 	return (NULL);
621 }
622 
623 static void *
624 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
625 {
626 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
627 	uint64_t *buf;
628 
629 	for (buf = buf_arg; buf < bufend; buf++) {
630 		if (*buf != old) {
631 			copy_pattern(old, buf_arg,
632 			    (char *)buf - (char *)buf_arg);
633 			return (buf);
634 		}
635 		*buf = new;
636 	}
637 
638 	return (NULL);
639 }
640 
641 void
642 umem_cache_applyall(void (*func)(umem_cache_t *))
643 {
644 	umem_cache_t *cp;
645 
646 	(void) mutex_lock(&umem_cache_lock);
647 	for (cp = umem_null_cache.cache_next; cp != &umem_null_cache;
648 	    cp = cp->cache_next)
649 		func(cp);
650 	(void) mutex_unlock(&umem_cache_lock);
651 }
652 
653 static void
654 umem_add_update_unlocked(umem_cache_t *cp, int flags)
655 {
656 	umem_cache_t *cnext, *cprev;
657 
658 	flags &= ~UMU_ACTIVE;
659 
660 	if (!flags)
661 		return;
662 
663 	if (cp->cache_uflags & UMU_ACTIVE) {
664 		cp->cache_uflags |= flags;
665 	} else {
666 		if (cp->cache_unext != NULL) {
667 			ASSERT(cp->cache_uflags != 0);
668 			cp->cache_uflags |= flags;
669 		} else {
670 			ASSERT(cp->cache_uflags == 0);
671 			cp->cache_uflags = flags;
672 			cp->cache_unext = cnext = &umem_null_cache;
673 			cp->cache_uprev = cprev = umem_null_cache.cache_uprev;
674 			cnext->cache_uprev = cp;
675 			cprev->cache_unext = cp;
676 		}
677 	}
678 }
679 
680 static void
681 umem_add_update(umem_cache_t *cp, int flags)
682 {
683 	(void) mutex_lock(&umem_update_lock);
684 
685 	umem_add_update_unlocked(cp, flags);
686 
687 	if (!IN_UPDATE())
688 		(void) cond_broadcast(&umem_update_cv);
689 
690 	(void) mutex_unlock(&umem_update_lock);
691 }
692 
693 /*
694  * Remove a cache from the update list, waiting for any in-progress work to
695  * complete first.
696  */
697 static void
698 umem_remove_updates(umem_cache_t *cp)
699 {
700 	(void) mutex_lock(&umem_update_lock);
701 
702 	/*
703 	 * Get it out of the active state
704 	 */
705 	while (cp->cache_uflags & UMU_ACTIVE) {
706 		ASSERT(cp->cache_unext == NULL);
707 
708 		cp->cache_uflags |= UMU_NOTIFY;
709 
710 		/*
711 		 * Make sure the update state is sane, before we wait
712 		 */
713 		ASSERT(umem_update_thr != 0 || umem_st_update_thr != 0);
714 		ASSERT(umem_update_thr != thr_self() &&
715 		    umem_st_update_thr != thr_self());
716 
717 		(void) _cond_wait(&umem_update_cv, &umem_update_lock);
718 	}
719 	/*
720 	 * Get it out of the Work Requested state
721 	 */
722 	if (cp->cache_unext != NULL) {
723 		cp->cache_uprev->cache_unext = cp->cache_unext;
724 		cp->cache_unext->cache_uprev = cp->cache_uprev;
725 		cp->cache_uprev = cp->cache_unext = NULL;
726 		cp->cache_uflags = 0;
727 	}
728 	/*
729 	 * Make sure it is in the Inactive state
730 	 */
731 	ASSERT(cp->cache_unext == NULL && cp->cache_uflags == 0);
732 	(void) mutex_unlock(&umem_update_lock);
733 }
734 
735 static void
736 umem_updateall(int flags)
737 {
738 	umem_cache_t *cp;
739 
740 	/*
741 	 * NOTE:  To prevent deadlock, umem_cache_lock is always acquired first.
742 	 *
743 	 * (umem_add_update is called from things run via umem_cache_applyall)
744 	 */
745 	(void) mutex_lock(&umem_cache_lock);
746 	(void) mutex_lock(&umem_update_lock);
747 
748 	for (cp = umem_null_cache.cache_next; cp != &umem_null_cache;
749 	    cp = cp->cache_next)
750 		umem_add_update_unlocked(cp, flags);
751 
752 	if (!IN_UPDATE())
753 		(void) cond_broadcast(&umem_update_cv);
754 
755 	(void) mutex_unlock(&umem_update_lock);
756 	(void) mutex_unlock(&umem_cache_lock);
757 }
758 
759 /*
760  * Debugging support.  Given a buffer address, find its slab.
761  */
762 static umem_slab_t *
763 umem_findslab(umem_cache_t *cp, void *buf)
764 {
765 	umem_slab_t *sp;
766 
767 	(void) mutex_lock(&cp->cache_lock);
768 	for (sp = cp->cache_nullslab.slab_next;
769 	    sp != &cp->cache_nullslab; sp = sp->slab_next) {
770 		if (UMEM_SLAB_MEMBER(sp, buf)) {
771 			(void) mutex_unlock(&cp->cache_lock);
772 			return (sp);
773 		}
774 	}
775 	(void) mutex_unlock(&cp->cache_lock);
776 
777 	return (NULL);
778 }
779 
780 static void
781 umem_error(int error, umem_cache_t *cparg, void *bufarg)
782 {
783 	umem_buftag_t *btp = NULL;
784 	umem_bufctl_t *bcp = NULL;
785 	umem_cache_t *cp = cparg;
786 	umem_slab_t *sp;
787 	uint64_t *off;
788 	void *buf = bufarg;
789 
790 	int old_logging = umem_logging;
791 
792 	umem_logging = 0;	/* stop logging when a bad thing happens */
793 
794 	umem_abort_info.ump_timestamp = gethrtime();
795 
796 	sp = umem_findslab(cp, buf);
797 	if (sp == NULL) {
798 		for (cp = umem_null_cache.cache_prev; cp != &umem_null_cache;
799 		    cp = cp->cache_prev) {
800 			if ((sp = umem_findslab(cp, buf)) != NULL)
801 				break;
802 		}
803 	}
804 
805 	if (sp == NULL) {
806 		cp = NULL;
807 		error = UMERR_BADADDR;
808 	} else {
809 		if (cp != cparg)
810 			error = UMERR_BADCACHE;
811 		else
812 			buf = (char *)bufarg - ((uintptr_t)bufarg -
813 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
814 		if (buf != bufarg)
815 			error = UMERR_BADBASE;
816 		if (cp->cache_flags & UMF_BUFTAG)
817 			btp = UMEM_BUFTAG(cp, buf);
818 		if (cp->cache_flags & UMF_HASH) {
819 			(void) mutex_lock(&cp->cache_lock);
820 			for (bcp = *UMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
821 				if (bcp->bc_addr == buf)
822 					break;
823 			(void) mutex_unlock(&cp->cache_lock);
824 			if (bcp == NULL && btp != NULL)
825 				bcp = btp->bt_bufctl;
826 			if (umem_findslab(cp->cache_bufctl_cache, bcp) ==
827 			    NULL || P2PHASE((uintptr_t)bcp, UMEM_ALIGN) ||
828 			    bcp->bc_addr != buf) {
829 				error = UMERR_BADBUFCTL;
830 				bcp = NULL;
831 			}
832 		}
833 	}
834 
835 	umem_abort_info.ump_error = error;
836 	umem_abort_info.ump_buffer = bufarg;
837 	umem_abort_info.ump_realbuf = buf;
838 	umem_abort_info.ump_cache = cparg;
839 	umem_abort_info.ump_realcache = cp;
840 	umem_abort_info.ump_slab = sp;
841 	umem_abort_info.ump_bufctl = bcp;
842 
843 	umem_printf("umem allocator: ");
844 
845 	switch (error) {
846 
847 	case UMERR_MODIFIED:
848 		umem_printf("buffer modified after being freed\n");
849 		off = verify_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
850 		if (off == NULL)	/* shouldn't happen */
851 			off = buf;
852 		umem_printf("modification occurred at offset 0x%lx "
853 		    "(0x%llx replaced by 0x%llx)\n",
854 		    (uintptr_t)off - (uintptr_t)buf,
855 		    (longlong_t)UMEM_FREE_PATTERN, (longlong_t)*off);
856 		break;
857 
858 	case UMERR_REDZONE:
859 		umem_printf("redzone violation: write past end of buffer\n");
860 		break;
861 
862 	case UMERR_BADADDR:
863 		umem_printf("invalid free: buffer not in cache\n");
864 		break;
865 
866 	case UMERR_DUPFREE:
867 		umem_printf("duplicate free: buffer freed twice\n");
868 		break;
869 
870 	case UMERR_BADBUFTAG:
871 		umem_printf("boundary tag corrupted\n");
872 		umem_printf("bcp ^ bxstat = %lx, should be %lx\n",
873 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
874 		    UMEM_BUFTAG_FREE);
875 		break;
876 
877 	case UMERR_BADBUFCTL:
878 		umem_printf("bufctl corrupted\n");
879 		break;
880 
881 	case UMERR_BADCACHE:
882 		umem_printf("buffer freed to wrong cache\n");
883 		umem_printf("buffer was allocated from %s,\n", cp->cache_name);
884 		umem_printf("caller attempting free to %s.\n",
885 		    cparg->cache_name);
886 		break;
887 
888 	case UMERR_BADSIZE:
889 		umem_printf("bad free: free size (%u) != alloc size (%u)\n",
890 		    UMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
891 		    UMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
892 		break;
893 
894 	case UMERR_BADBASE:
895 		umem_printf("bad free: free address (%p) != alloc address "
896 		    "(%p)\n", bufarg, buf);
897 		break;
898 	}
899 
900 	umem_printf("buffer=%p  bufctl=%p  cache: %s\n",
901 	    bufarg, (void *)bcp, cparg->cache_name);
902 
903 	if (bcp != NULL && (cp->cache_flags & UMF_AUDIT) &&
904 	    error != UMERR_BADBUFCTL) {
905 		int d;
906 		timespec_t ts;
907 		hrtime_t diff;
908 		umem_bufctl_audit_t *bcap = (umem_bufctl_audit_t *)bcp;
909 
910 		diff = umem_abort_info.ump_timestamp - bcap->bc_timestamp;
911 		ts.tv_sec = diff / NANOSEC;
912 		ts.tv_nsec = diff % NANOSEC;
913 
914 		umem_printf("previous transaction on buffer %p:\n", buf);
915 		umem_printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
916 		    (void *)(intptr_t)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
917 		    (void *)sp, cp->cache_name);
918 		for (d = 0; d < MIN(bcap->bc_depth, umem_stack_depth); d++) {
919 			(void) print_sym((void *)bcap->bc_stack[d]);
920 			umem_printf("\n");
921 		}
922 	}
923 
924 	umem_err_recoverable("umem: heap corruption detected");
925 
926 	umem_logging = old_logging;	/* resume logging */
927 }
928 
929 void
930 umem_nofail_callback(umem_nofail_callback_t *cb)
931 {
932 	nofail_callback = cb;
933 }
934 
935 static int
936 umem_alloc_retry(umem_cache_t *cp, int umflag)
937 {
938 	if (cp == &umem_null_cache) {
939 		if (umem_init())
940 			return (1);				/* retry */
941 		/*
942 		 * Initialization failed.  Do normal failure processing.
943 		 */
944 	}
945 	if (umflag & UMEM_NOFAIL) {
946 		int def_result = UMEM_CALLBACK_EXIT(255);
947 		int result = def_result;
948 		umem_nofail_callback_t *callback = nofail_callback;
949 
950 		if (callback != NULL)
951 			result = callback();
952 
953 		if (result == UMEM_CALLBACK_RETRY)
954 			return (1);
955 
956 		if ((result & ~0xFF) != UMEM_CALLBACK_EXIT(0)) {
957 			log_message("nofail callback returned %x\n", result);
958 			result = def_result;
959 		}
960 
961 		/*
962 		 * only one thread will call exit
963 		 */
964 		if (umem_nofail_exit_thr == thr_self())
965 			umem_panic("recursive UMEM_CALLBACK_EXIT()\n");
966 
967 		(void) mutex_lock(&umem_nofail_exit_lock);
968 		umem_nofail_exit_thr = thr_self();
969 		exit(result & 0xFF);
970 		/*NOTREACHED*/
971 	}
972 	return (0);
973 }
974 
975 static umem_log_header_t *
976 umem_log_init(size_t logsize)
977 {
978 	umem_log_header_t *lhp;
979 	int nchunks = 4 * umem_max_ncpus;
980 	size_t lhsize = offsetof(umem_log_header_t, lh_cpu[umem_max_ncpus]);
981 	int i;
982 
983 	if (logsize == 0)
984 		return (NULL);
985 
986 	/*
987 	 * Make sure that lhp->lh_cpu[] is nicely aligned
988 	 * to prevent false sharing of cache lines.
989 	 */
990 	lhsize = P2ROUNDUP(lhsize, UMEM_ALIGN);
991 	lhp = vmem_xalloc(umem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
992 	    NULL, NULL, VM_NOSLEEP);
993 	if (lhp == NULL)
994 		goto fail;
995 
996 	bzero(lhp, lhsize);
997 
998 	(void) mutex_init(&lhp->lh_lock, USYNC_THREAD, NULL);
999 	lhp->lh_nchunks = nchunks;
1000 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks, PAGESIZE);
1001 	if (lhp->lh_chunksize == 0)
1002 		lhp->lh_chunksize = PAGESIZE;
1003 
1004 	lhp->lh_base = vmem_alloc(umem_log_arena,
1005 	    lhp->lh_chunksize * nchunks, VM_NOSLEEP);
1006 	if (lhp->lh_base == NULL)
1007 		goto fail;
1008 
1009 	lhp->lh_free = vmem_alloc(umem_log_arena,
1010 	    nchunks * sizeof (int), VM_NOSLEEP);
1011 	if (lhp->lh_free == NULL)
1012 		goto fail;
1013 
1014 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1015 
1016 	for (i = 0; i < umem_max_ncpus; i++) {
1017 		umem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1018 		(void) mutex_init(&clhp->clh_lock, USYNC_THREAD, NULL);
1019 		clhp->clh_chunk = i;
1020 	}
1021 
1022 	for (i = umem_max_ncpus; i < nchunks; i++)
1023 		lhp->lh_free[i] = i;
1024 
1025 	lhp->lh_head = umem_max_ncpus;
1026 	lhp->lh_tail = 0;
1027 
1028 	return (lhp);
1029 
1030 fail:
1031 	if (lhp != NULL) {
1032 		if (lhp->lh_base != NULL)
1033 			vmem_free(umem_log_arena, lhp->lh_base,
1034 			    lhp->lh_chunksize * nchunks);
1035 
1036 		vmem_xfree(umem_log_arena, lhp, lhsize);
1037 	}
1038 	return (NULL);
1039 }
1040 
1041 static void *
1042 umem_log_enter(umem_log_header_t *lhp, void *data, size_t size)
1043 {
1044 	void *logspace;
1045 	umem_cpu_log_header_t *clhp =
1046 	    &lhp->lh_cpu[CPU(umem_cpu_mask)->cpu_number];
1047 
1048 	if (lhp == NULL || umem_logging == 0)
1049 		return (NULL);
1050 
1051 	(void) mutex_lock(&clhp->clh_lock);
1052 	clhp->clh_hits++;
1053 	if (size > clhp->clh_avail) {
1054 		(void) mutex_lock(&lhp->lh_lock);
1055 		lhp->lh_hits++;
1056 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1057 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1058 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1059 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1060 		clhp->clh_current = lhp->lh_base +
1061 		    clhp->clh_chunk * lhp->lh_chunksize;
1062 		clhp->clh_avail = lhp->lh_chunksize;
1063 		if (size > lhp->lh_chunksize)
1064 			size = lhp->lh_chunksize;
1065 		(void) mutex_unlock(&lhp->lh_lock);
1066 	}
1067 	logspace = clhp->clh_current;
1068 	clhp->clh_current += size;
1069 	clhp->clh_avail -= size;
1070 	bcopy(data, logspace, size);
1071 	(void) mutex_unlock(&clhp->clh_lock);
1072 	return (logspace);
1073 }
1074 
1075 #define	UMEM_AUDIT(lp, cp, bcp)						\
1076 {									\
1077 	umem_bufctl_audit_t *_bcp = (umem_bufctl_audit_t *)(bcp);	\
1078 	_bcp->bc_timestamp = gethrtime();				\
1079 	_bcp->bc_thread = thr_self();					\
1080 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, umem_stack_depth,	\
1081 	    (cp != NULL) && (cp->cache_flags & UMF_CHECKSIGNAL));	\
1082 	_bcp->bc_lastlog = umem_log_enter((lp), _bcp,			\
1083 	    UMEM_BUFCTL_AUDIT_SIZE);					\
1084 }
1085 
1086 static void
1087 umem_log_event(umem_log_header_t *lp, umem_cache_t *cp,
1088 	umem_slab_t *sp, void *addr)
1089 {
1090 	umem_bufctl_audit_t *bcp;
1091 	UMEM_LOCAL_BUFCTL_AUDIT(&bcp);
1092 
1093 	bzero(bcp, UMEM_BUFCTL_AUDIT_SIZE);
1094 	bcp->bc_addr = addr;
1095 	bcp->bc_slab = sp;
1096 	bcp->bc_cache = cp;
1097 	UMEM_AUDIT(lp, cp, bcp);
1098 }
1099 
1100 /*
1101  * Create a new slab for cache cp.
1102  */
1103 static umem_slab_t *
1104 umem_slab_create(umem_cache_t *cp, int umflag)
1105 {
1106 	size_t slabsize = cp->cache_slabsize;
1107 	size_t chunksize = cp->cache_chunksize;
1108 	int cache_flags = cp->cache_flags;
1109 	size_t color, chunks;
1110 	char *buf, *slab;
1111 	umem_slab_t *sp;
1112 	umem_bufctl_t *bcp;
1113 	vmem_t *vmp = cp->cache_arena;
1114 
1115 	color = cp->cache_color + cp->cache_align;
1116 	if (color > cp->cache_maxcolor)
1117 		color = cp->cache_mincolor;
1118 	cp->cache_color = color;
1119 
1120 	slab = vmem_alloc(vmp, slabsize, UMEM_VMFLAGS(umflag));
1121 
1122 	if (slab == NULL)
1123 		goto vmem_alloc_failure;
1124 
1125 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1126 
1127 	if (!(cp->cache_cflags & UMC_NOTOUCH) &&
1128 	    (cp->cache_flags & UMF_DEADBEEF))
1129 		copy_pattern(UMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1130 
1131 	if (cache_flags & UMF_HASH) {
1132 		if ((sp = _umem_cache_alloc(umem_slab_cache, umflag)) == NULL)
1133 			goto slab_alloc_failure;
1134 		chunks = (slabsize - color) / chunksize;
1135 	} else {
1136 		sp = UMEM_SLAB(cp, slab);
1137 		chunks = (slabsize - sizeof (umem_slab_t) - color) / chunksize;
1138 	}
1139 
1140 	sp->slab_cache	= cp;
1141 	sp->slab_head	= NULL;
1142 	sp->slab_refcnt	= 0;
1143 	sp->slab_base	= buf = slab + color;
1144 	sp->slab_chunks	= chunks;
1145 
1146 	ASSERT(chunks > 0);
1147 	while (chunks-- != 0) {
1148 		if (cache_flags & UMF_HASH) {
1149 			bcp = _umem_cache_alloc(cp->cache_bufctl_cache, umflag);
1150 			if (bcp == NULL)
1151 				goto bufctl_alloc_failure;
1152 			if (cache_flags & UMF_AUDIT) {
1153 				umem_bufctl_audit_t *bcap =
1154 				    (umem_bufctl_audit_t *)bcp;
1155 				bzero(bcap, UMEM_BUFCTL_AUDIT_SIZE);
1156 				bcap->bc_cache = cp;
1157 			}
1158 			bcp->bc_addr = buf;
1159 			bcp->bc_slab = sp;
1160 		} else {
1161 			bcp = UMEM_BUFCTL(cp, buf);
1162 		}
1163 		if (cache_flags & UMF_BUFTAG) {
1164 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1165 			btp->bt_redzone = UMEM_REDZONE_PATTERN;
1166 			btp->bt_bufctl = bcp;
1167 			btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
1168 			if (cache_flags & UMF_DEADBEEF) {
1169 				copy_pattern(UMEM_FREE_PATTERN, buf,
1170 				    cp->cache_verify);
1171 			}
1172 		}
1173 		bcp->bc_next = sp->slab_head;
1174 		sp->slab_head = bcp;
1175 		buf += chunksize;
1176 	}
1177 
1178 	umem_log_event(umem_slab_log, cp, sp, slab);
1179 
1180 	return (sp);
1181 
1182 bufctl_alloc_failure:
1183 
1184 	while ((bcp = sp->slab_head) != NULL) {
1185 		sp->slab_head = bcp->bc_next;
1186 		_umem_cache_free(cp->cache_bufctl_cache, bcp);
1187 	}
1188 	_umem_cache_free(umem_slab_cache, sp);
1189 
1190 slab_alloc_failure:
1191 
1192 	vmem_free(vmp, slab, slabsize);
1193 
1194 vmem_alloc_failure:
1195 
1196 	umem_log_event(umem_failure_log, cp, NULL, NULL);
1197 	atomic_add_64(&cp->cache_alloc_fail, 1);
1198 
1199 	return (NULL);
1200 }
1201 
1202 /*
1203  * Destroy a slab.
1204  */
1205 static void
1206 umem_slab_destroy(umem_cache_t *cp, umem_slab_t *sp)
1207 {
1208 	vmem_t *vmp = cp->cache_arena;
1209 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1210 
1211 	if (cp->cache_flags & UMF_HASH) {
1212 		umem_bufctl_t *bcp;
1213 		while ((bcp = sp->slab_head) != NULL) {
1214 			sp->slab_head = bcp->bc_next;
1215 			_umem_cache_free(cp->cache_bufctl_cache, bcp);
1216 		}
1217 		_umem_cache_free(umem_slab_cache, sp);
1218 	}
1219 	vmem_free(vmp, slab, cp->cache_slabsize);
1220 }
1221 
1222 /*
1223  * Allocate a raw (unconstructed) buffer from cp's slab layer.
1224  */
1225 static void *
1226 umem_slab_alloc(umem_cache_t *cp, int umflag)
1227 {
1228 	umem_bufctl_t *bcp, **hash_bucket;
1229 	umem_slab_t *sp;
1230 	void *buf;
1231 
1232 	(void) mutex_lock(&cp->cache_lock);
1233 	cp->cache_slab_alloc++;
1234 	sp = cp->cache_freelist;
1235 	ASSERT(sp->slab_cache == cp);
1236 	if (sp->slab_head == NULL) {
1237 		/*
1238 		 * The freelist is empty.  Create a new slab.
1239 		 */
1240 		(void) mutex_unlock(&cp->cache_lock);
1241 		if (cp == &umem_null_cache)
1242 			return (NULL);
1243 		if ((sp = umem_slab_create(cp, umflag)) == NULL)
1244 			return (NULL);
1245 		(void) mutex_lock(&cp->cache_lock);
1246 		cp->cache_slab_create++;
1247 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1248 			cp->cache_bufmax = cp->cache_buftotal;
1249 		sp->slab_next = cp->cache_freelist;
1250 		sp->slab_prev = cp->cache_freelist->slab_prev;
1251 		sp->slab_next->slab_prev = sp;
1252 		sp->slab_prev->slab_next = sp;
1253 		cp->cache_freelist = sp;
1254 	}
1255 
1256 	sp->slab_refcnt++;
1257 	ASSERT(sp->slab_refcnt <= sp->slab_chunks);
1258 
1259 	/*
1260 	 * If we're taking the last buffer in the slab,
1261 	 * remove the slab from the cache's freelist.
1262 	 */
1263 	bcp = sp->slab_head;
1264 	if ((sp->slab_head = bcp->bc_next) == NULL) {
1265 		cp->cache_freelist = sp->slab_next;
1266 		ASSERT(sp->slab_refcnt == sp->slab_chunks);
1267 	}
1268 
1269 	if (cp->cache_flags & UMF_HASH) {
1270 		/*
1271 		 * Add buffer to allocated-address hash table.
1272 		 */
1273 		buf = bcp->bc_addr;
1274 		hash_bucket = UMEM_HASH(cp, buf);
1275 		bcp->bc_next = *hash_bucket;
1276 		*hash_bucket = bcp;
1277 		if ((cp->cache_flags & (UMF_AUDIT | UMF_BUFTAG)) == UMF_AUDIT) {
1278 			UMEM_AUDIT(umem_transaction_log, cp, bcp);
1279 		}
1280 	} else {
1281 		buf = UMEM_BUF(cp, bcp);
1282 	}
1283 
1284 	ASSERT(UMEM_SLAB_MEMBER(sp, buf));
1285 
1286 	(void) mutex_unlock(&cp->cache_lock);
1287 
1288 	return (buf);
1289 }
1290 
1291 /*
1292  * Free a raw (unconstructed) buffer to cp's slab layer.
1293  */
1294 static void
1295 umem_slab_free(umem_cache_t *cp, void *buf)
1296 {
1297 	umem_slab_t *sp;
1298 	umem_bufctl_t *bcp, **prev_bcpp;
1299 
1300 	ASSERT(buf != NULL);
1301 
1302 	(void) mutex_lock(&cp->cache_lock);
1303 	cp->cache_slab_free++;
1304 
1305 	if (cp->cache_flags & UMF_HASH) {
1306 		/*
1307 		 * Look up buffer in allocated-address hash table.
1308 		 */
1309 		prev_bcpp = UMEM_HASH(cp, buf);
1310 		while ((bcp = *prev_bcpp) != NULL) {
1311 			if (bcp->bc_addr == buf) {
1312 				*prev_bcpp = bcp->bc_next;
1313 				sp = bcp->bc_slab;
1314 				break;
1315 			}
1316 			cp->cache_lookup_depth++;
1317 			prev_bcpp = &bcp->bc_next;
1318 		}
1319 	} else {
1320 		bcp = UMEM_BUFCTL(cp, buf);
1321 		sp = UMEM_SLAB(cp, buf);
1322 	}
1323 
1324 	if (bcp == NULL || sp->slab_cache != cp || !UMEM_SLAB_MEMBER(sp, buf)) {
1325 		(void) mutex_unlock(&cp->cache_lock);
1326 		umem_error(UMERR_BADADDR, cp, buf);
1327 		return;
1328 	}
1329 
1330 	if ((cp->cache_flags & (UMF_AUDIT | UMF_BUFTAG)) == UMF_AUDIT) {
1331 		if (cp->cache_flags & UMF_CONTENTS)
1332 			((umem_bufctl_audit_t *)bcp)->bc_contents =
1333 			    umem_log_enter(umem_content_log, buf,
1334 			    cp->cache_contents);
1335 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
1336 	}
1337 
1338 	/*
1339 	 * If this slab isn't currently on the freelist, put it there.
1340 	 */
1341 	if (sp->slab_head == NULL) {
1342 		ASSERT(sp->slab_refcnt == sp->slab_chunks);
1343 		ASSERT(cp->cache_freelist != sp);
1344 		sp->slab_next->slab_prev = sp->slab_prev;
1345 		sp->slab_prev->slab_next = sp->slab_next;
1346 		sp->slab_next = cp->cache_freelist;
1347 		sp->slab_prev = cp->cache_freelist->slab_prev;
1348 		sp->slab_next->slab_prev = sp;
1349 		sp->slab_prev->slab_next = sp;
1350 		cp->cache_freelist = sp;
1351 	}
1352 
1353 	bcp->bc_next = sp->slab_head;
1354 	sp->slab_head = bcp;
1355 
1356 	ASSERT(sp->slab_refcnt >= 1);
1357 	if (--sp->slab_refcnt == 0) {
1358 		/*
1359 		 * There are no outstanding allocations from this slab,
1360 		 * so we can reclaim the memory.
1361 		 */
1362 		sp->slab_next->slab_prev = sp->slab_prev;
1363 		sp->slab_prev->slab_next = sp->slab_next;
1364 		if (sp == cp->cache_freelist)
1365 			cp->cache_freelist = sp->slab_next;
1366 		cp->cache_slab_destroy++;
1367 		cp->cache_buftotal -= sp->slab_chunks;
1368 		(void) mutex_unlock(&cp->cache_lock);
1369 		umem_slab_destroy(cp, sp);
1370 		return;
1371 	}
1372 	(void) mutex_unlock(&cp->cache_lock);
1373 }
1374 
1375 static int
1376 umem_cache_alloc_debug(umem_cache_t *cp, void *buf, int umflag)
1377 {
1378 	umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1379 	umem_bufctl_audit_t *bcp = (umem_bufctl_audit_t *)btp->bt_bufctl;
1380 	uint32_t mtbf;
1381 	int flags_nfatal;
1382 
1383 	if (btp->bt_bxstat != ((intptr_t)bcp ^ UMEM_BUFTAG_FREE)) {
1384 		umem_error(UMERR_BADBUFTAG, cp, buf);
1385 		return (-1);
1386 	}
1387 
1388 	btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_ALLOC;
1389 
1390 	if ((cp->cache_flags & UMF_HASH) && bcp->bc_addr != buf) {
1391 		umem_error(UMERR_BADBUFCTL, cp, buf);
1392 		return (-1);
1393 	}
1394 
1395 	btp->bt_redzone = UMEM_REDZONE_PATTERN;
1396 
1397 	if (cp->cache_flags & UMF_DEADBEEF) {
1398 		if (verify_and_copy_pattern(UMEM_FREE_PATTERN,
1399 		    UMEM_UNINITIALIZED_PATTERN, buf, cp->cache_verify)) {
1400 			umem_error(UMERR_MODIFIED, cp, buf);
1401 			return (-1);
1402 		}
1403 	}
1404 
1405 	if ((mtbf = umem_mtbf | cp->cache_mtbf) != 0 &&
1406 	    gethrtime() % mtbf == 0 &&
1407 	    (umflag & (UMEM_FATAL_FLAGS)) == 0) {
1408 		umem_log_event(umem_failure_log, cp, NULL, NULL);
1409 	} else {
1410 		mtbf = 0;
1411 	}
1412 
1413 	/*
1414 	 * We do not pass fatal flags on to the constructor.  This prevents
1415 	 * leaking buffers in the event of a subordinate constructor failing.
1416 	 */
1417 	flags_nfatal = UMEM_DEFAULT;
1418 	if (mtbf || (cp->cache_constructor != NULL &&
1419 	    cp->cache_constructor(buf, cp->cache_private, flags_nfatal) != 0)) {
1420 		atomic_add_64(&cp->cache_alloc_fail, 1);
1421 		btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
1422 		copy_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
1423 		umem_slab_free(cp, buf);
1424 		return (-1);
1425 	}
1426 
1427 	if (cp->cache_flags & UMF_AUDIT) {
1428 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
1429 	}
1430 
1431 	return (0);
1432 }
1433 
1434 static int
1435 umem_cache_free_debug(umem_cache_t *cp, void *buf)
1436 {
1437 	umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1438 	umem_bufctl_audit_t *bcp = (umem_bufctl_audit_t *)btp->bt_bufctl;
1439 	umem_slab_t *sp;
1440 
1441 	if (btp->bt_bxstat != ((intptr_t)bcp ^ UMEM_BUFTAG_ALLOC)) {
1442 		if (btp->bt_bxstat == ((intptr_t)bcp ^ UMEM_BUFTAG_FREE)) {
1443 			umem_error(UMERR_DUPFREE, cp, buf);
1444 			return (-1);
1445 		}
1446 		sp = umem_findslab(cp, buf);
1447 		if (sp == NULL || sp->slab_cache != cp)
1448 			umem_error(UMERR_BADADDR, cp, buf);
1449 		else
1450 			umem_error(UMERR_REDZONE, cp, buf);
1451 		return (-1);
1452 	}
1453 
1454 	btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
1455 
1456 	if ((cp->cache_flags & UMF_HASH) && bcp->bc_addr != buf) {
1457 		umem_error(UMERR_BADBUFCTL, cp, buf);
1458 		return (-1);
1459 	}
1460 
1461 	if (btp->bt_redzone != UMEM_REDZONE_PATTERN) {
1462 		umem_error(UMERR_REDZONE, cp, buf);
1463 		return (-1);
1464 	}
1465 
1466 	if (cp->cache_flags & UMF_AUDIT) {
1467 		if (cp->cache_flags & UMF_CONTENTS)
1468 			bcp->bc_contents = umem_log_enter(umem_content_log,
1469 			    buf, cp->cache_contents);
1470 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
1471 	}
1472 
1473 	if (cp->cache_destructor != NULL)
1474 		cp->cache_destructor(buf, cp->cache_private);
1475 
1476 	if (cp->cache_flags & UMF_DEADBEEF)
1477 		copy_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
1478 
1479 	return (0);
1480 }
1481 
1482 /*
1483  * Free each object in magazine mp to cp's slab layer, and free mp itself.
1484  */
1485 static void
1486 umem_magazine_destroy(umem_cache_t *cp, umem_magazine_t *mp, int nrounds)
1487 {
1488 	int round;
1489 
1490 	ASSERT(cp->cache_next == NULL || IN_UPDATE());
1491 
1492 	for (round = 0; round < nrounds; round++) {
1493 		void *buf = mp->mag_round[round];
1494 
1495 		if ((cp->cache_flags & UMF_DEADBEEF) &&
1496 		    verify_pattern(UMEM_FREE_PATTERN, buf,
1497 		    cp->cache_verify) != NULL) {
1498 			umem_error(UMERR_MODIFIED, cp, buf);
1499 			continue;
1500 		}
1501 
1502 		if (!(cp->cache_flags & UMF_BUFTAG) &&
1503 		    cp->cache_destructor != NULL)
1504 			cp->cache_destructor(buf, cp->cache_private);
1505 
1506 		umem_slab_free(cp, buf);
1507 	}
1508 	ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
1509 	_umem_cache_free(cp->cache_magtype->mt_cache, mp);
1510 }
1511 
1512 /*
1513  * Allocate a magazine from the depot.
1514  */
1515 static umem_magazine_t *
1516 umem_depot_alloc(umem_cache_t *cp, umem_maglist_t *mlp)
1517 {
1518 	umem_magazine_t *mp;
1519 
1520 	/*
1521 	 * If we can't get the depot lock without contention,
1522 	 * update our contention count.  We use the depot
1523 	 * contention rate to determine whether we need to
1524 	 * increase the magazine size for better scalability.
1525 	 */
1526 	if (mutex_trylock(&cp->cache_depot_lock) != 0) {
1527 		(void) mutex_lock(&cp->cache_depot_lock);
1528 		cp->cache_depot_contention++;
1529 	}
1530 
1531 	if ((mp = mlp->ml_list) != NULL) {
1532 		ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
1533 		mlp->ml_list = mp->mag_next;
1534 		if (--mlp->ml_total < mlp->ml_min)
1535 			mlp->ml_min = mlp->ml_total;
1536 		mlp->ml_alloc++;
1537 	}
1538 
1539 	(void) mutex_unlock(&cp->cache_depot_lock);
1540 
1541 	return (mp);
1542 }
1543 
1544 /*
1545  * Free a magazine to the depot.
1546  */
1547 static void
1548 umem_depot_free(umem_cache_t *cp, umem_maglist_t *mlp, umem_magazine_t *mp)
1549 {
1550 	(void) mutex_lock(&cp->cache_depot_lock);
1551 	ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
1552 	mp->mag_next = mlp->ml_list;
1553 	mlp->ml_list = mp;
1554 	mlp->ml_total++;
1555 	(void) mutex_unlock(&cp->cache_depot_lock);
1556 }
1557 
1558 /*
1559  * Update the working set statistics for cp's depot.
1560  */
1561 static void
1562 umem_depot_ws_update(umem_cache_t *cp)
1563 {
1564 	(void) mutex_lock(&cp->cache_depot_lock);
1565 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
1566 	cp->cache_full.ml_min = cp->cache_full.ml_total;
1567 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
1568 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
1569 	(void) mutex_unlock(&cp->cache_depot_lock);
1570 }
1571 
1572 /*
1573  * Reap all magazines that have fallen out of the depot's working set.
1574  */
1575 static void
1576 umem_depot_ws_reap(umem_cache_t *cp)
1577 {
1578 	long reap;
1579 	umem_magazine_t *mp;
1580 
1581 	ASSERT(cp->cache_next == NULL || IN_REAP());
1582 
1583 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
1584 	while (reap-- && (mp = umem_depot_alloc(cp, &cp->cache_full)) != NULL)
1585 		umem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
1586 
1587 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
1588 	while (reap-- && (mp = umem_depot_alloc(cp, &cp->cache_empty)) != NULL)
1589 		umem_magazine_destroy(cp, mp, 0);
1590 }
1591 
1592 static void
1593 umem_cpu_reload(umem_cpu_cache_t *ccp, umem_magazine_t *mp, int rounds)
1594 {
1595 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
1596 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
1597 	ASSERT(ccp->cc_magsize > 0);
1598 
1599 	ccp->cc_ploaded = ccp->cc_loaded;
1600 	ccp->cc_prounds = ccp->cc_rounds;
1601 	ccp->cc_loaded = mp;
1602 	ccp->cc_rounds = rounds;
1603 }
1604 
1605 /*
1606  * Allocate a constructed object from cache cp.
1607  */
1608 #pragma weak umem_cache_alloc = _umem_cache_alloc
1609 void *
1610 _umem_cache_alloc(umem_cache_t *cp, int umflag)
1611 {
1612 	umem_cpu_cache_t *ccp;
1613 	umem_magazine_t *fmp;
1614 	void *buf;
1615 	int flags_nfatal;
1616 
1617 retry:
1618 	ccp = UMEM_CPU_CACHE(cp, CPU(cp->cache_cpu_mask));
1619 	(void) mutex_lock(&ccp->cc_lock);
1620 	for (;;) {
1621 		/*
1622 		 * If there's an object available in the current CPU's
1623 		 * loaded magazine, just take it and return.
1624 		 */
1625 		if (ccp->cc_rounds > 0) {
1626 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
1627 			ccp->cc_alloc++;
1628 			(void) mutex_unlock(&ccp->cc_lock);
1629 			if ((ccp->cc_flags & UMF_BUFTAG) &&
1630 			    umem_cache_alloc_debug(cp, buf, umflag) == -1) {
1631 				if (umem_alloc_retry(cp, umflag)) {
1632 					goto retry;
1633 				}
1634 
1635 				return (NULL);
1636 			}
1637 			return (buf);
1638 		}
1639 
1640 		/*
1641 		 * The loaded magazine is empty.  If the previously loaded
1642 		 * magazine was full, exchange them and try again.
1643 		 */
1644 		if (ccp->cc_prounds > 0) {
1645 			umem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
1646 			continue;
1647 		}
1648 
1649 		/*
1650 		 * If the magazine layer is disabled, break out now.
1651 		 */
1652 		if (ccp->cc_magsize == 0)
1653 			break;
1654 
1655 		/*
1656 		 * Try to get a full magazine from the depot.
1657 		 */
1658 		fmp = umem_depot_alloc(cp, &cp->cache_full);
1659 		if (fmp != NULL) {
1660 			if (ccp->cc_ploaded != NULL)
1661 				umem_depot_free(cp, &cp->cache_empty,
1662 				    ccp->cc_ploaded);
1663 			umem_cpu_reload(ccp, fmp, ccp->cc_magsize);
1664 			continue;
1665 		}
1666 
1667 		/*
1668 		 * There are no full magazines in the depot,
1669 		 * so fall through to the slab layer.
1670 		 */
1671 		break;
1672 	}
1673 	(void) mutex_unlock(&ccp->cc_lock);
1674 
1675 	/*
1676 	 * We couldn't allocate a constructed object from the magazine layer,
1677 	 * so get a raw buffer from the slab layer and apply its constructor.
1678 	 */
1679 	buf = umem_slab_alloc(cp, umflag);
1680 
1681 	if (buf == NULL) {
1682 		if (cp == &umem_null_cache)
1683 			return (NULL);
1684 		if (umem_alloc_retry(cp, umflag)) {
1685 			goto retry;
1686 		}
1687 
1688 		return (NULL);
1689 	}
1690 
1691 	if (cp->cache_flags & UMF_BUFTAG) {
1692 		/*
1693 		 * Let umem_cache_alloc_debug() apply the constructor for us.
1694 		 */
1695 		if (umem_cache_alloc_debug(cp, buf, umflag) == -1) {
1696 			if (umem_alloc_retry(cp, umflag)) {
1697 				goto retry;
1698 			}
1699 			return (NULL);
1700 		}
1701 		return (buf);
1702 	}
1703 
1704 	/*
1705 	 * We do not pass fatal flags on to the constructor.  This prevents
1706 	 * leaking buffers in the event of a subordinate constructor failing.
1707 	 */
1708 	flags_nfatal = UMEM_DEFAULT;
1709 	if (cp->cache_constructor != NULL &&
1710 	    cp->cache_constructor(buf, cp->cache_private, flags_nfatal) != 0) {
1711 		atomic_add_64(&cp->cache_alloc_fail, 1);
1712 		umem_slab_free(cp, buf);
1713 
1714 		if (umem_alloc_retry(cp, umflag)) {
1715 			goto retry;
1716 		}
1717 		return (NULL);
1718 	}
1719 
1720 	return (buf);
1721 }
1722 
1723 /*
1724  * Free a constructed object to cache cp.
1725  */
1726 #pragma weak umem_cache_free = _umem_cache_free
1727 void
1728 _umem_cache_free(umem_cache_t *cp, void *buf)
1729 {
1730 	umem_cpu_cache_t *ccp = UMEM_CPU_CACHE(cp, CPU(cp->cache_cpu_mask));
1731 	umem_magazine_t *emp;
1732 	umem_magtype_t *mtp;
1733 
1734 	if (ccp->cc_flags & UMF_BUFTAG)
1735 		if (umem_cache_free_debug(cp, buf) == -1)
1736 			return;
1737 
1738 	(void) mutex_lock(&ccp->cc_lock);
1739 	for (;;) {
1740 		/*
1741 		 * If there's a slot available in the current CPU's
1742 		 * loaded magazine, just put the object there and return.
1743 		 */
1744 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
1745 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
1746 			ccp->cc_free++;
1747 			(void) mutex_unlock(&ccp->cc_lock);
1748 			return;
1749 		}
1750 
1751 		/*
1752 		 * The loaded magazine is full.  If the previously loaded
1753 		 * magazine was empty, exchange them and try again.
1754 		 */
1755 		if (ccp->cc_prounds == 0) {
1756 			umem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
1757 			continue;
1758 		}
1759 
1760 		/*
1761 		 * If the magazine layer is disabled, break out now.
1762 		 */
1763 		if (ccp->cc_magsize == 0)
1764 			break;
1765 
1766 		/*
1767 		 * Try to get an empty magazine from the depot.
1768 		 */
1769 		emp = umem_depot_alloc(cp, &cp->cache_empty);
1770 		if (emp != NULL) {
1771 			if (ccp->cc_ploaded != NULL)
1772 				umem_depot_free(cp, &cp->cache_full,
1773 				    ccp->cc_ploaded);
1774 			umem_cpu_reload(ccp, emp, 0);
1775 			continue;
1776 		}
1777 
1778 		/*
1779 		 * There are no empty magazines in the depot,
1780 		 * so try to allocate a new one.  We must drop all locks
1781 		 * across umem_cache_alloc() because lower layers may
1782 		 * attempt to allocate from this cache.
1783 		 */
1784 		mtp = cp->cache_magtype;
1785 		(void) mutex_unlock(&ccp->cc_lock);
1786 		emp = _umem_cache_alloc(mtp->mt_cache, UMEM_DEFAULT);
1787 		(void) mutex_lock(&ccp->cc_lock);
1788 
1789 		if (emp != NULL) {
1790 			/*
1791 			 * We successfully allocated an empty magazine.
1792 			 * However, we had to drop ccp->cc_lock to do it,
1793 			 * so the cache's magazine size may have changed.
1794 			 * If so, free the magazine and try again.
1795 			 */
1796 			if (ccp->cc_magsize != mtp->mt_magsize) {
1797 				(void) mutex_unlock(&ccp->cc_lock);
1798 				_umem_cache_free(mtp->mt_cache, emp);
1799 				(void) mutex_lock(&ccp->cc_lock);
1800 				continue;
1801 			}
1802 
1803 			/*
1804 			 * We got a magazine of the right size.  Add it to
1805 			 * the depot and try the whole dance again.
1806 			 */
1807 			umem_depot_free(cp, &cp->cache_empty, emp);
1808 			continue;
1809 		}
1810 
1811 		/*
1812 		 * We couldn't allocate an empty magazine,
1813 		 * so fall through to the slab layer.
1814 		 */
1815 		break;
1816 	}
1817 	(void) mutex_unlock(&ccp->cc_lock);
1818 
1819 	/*
1820 	 * We couldn't free our constructed object to the magazine layer,
1821 	 * so apply its destructor and free it to the slab layer.
1822 	 * Note that if UMF_BUFTAG is in effect, umem_cache_free_debug()
1823 	 * will have already applied the destructor.
1824 	 */
1825 	if (!(cp->cache_flags & UMF_BUFTAG) && cp->cache_destructor != NULL)
1826 		cp->cache_destructor(buf, cp->cache_private);
1827 
1828 	umem_slab_free(cp, buf);
1829 }
1830 
1831 #pragma weak umem_zalloc = _umem_zalloc
1832 void *
1833 _umem_zalloc(size_t size, int umflag)
1834 {
1835 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
1836 	void *buf;
1837 
1838 retry:
1839 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
1840 		umem_cache_t *cp = umem_alloc_table[index];
1841 		buf = _umem_cache_alloc(cp, umflag);
1842 		if (buf != NULL) {
1843 			if (cp->cache_flags & UMF_BUFTAG) {
1844 				umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1845 				((uint8_t *)buf)[size] = UMEM_REDZONE_BYTE;
1846 				((uint32_t *)btp)[1] = UMEM_SIZE_ENCODE(size);
1847 			}
1848 			bzero(buf, size);
1849 		} else if (umem_alloc_retry(cp, umflag))
1850 			goto retry;
1851 	} else {
1852 		buf = _umem_alloc(size, umflag);	/* handles failure */
1853 		if (buf != NULL)
1854 			bzero(buf, size);
1855 	}
1856 	return (buf);
1857 }
1858 
1859 #pragma weak umem_alloc = _umem_alloc
1860 void *
1861 _umem_alloc(size_t size, int umflag)
1862 {
1863 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
1864 	void *buf;
1865 umem_alloc_retry:
1866 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
1867 		umem_cache_t *cp = umem_alloc_table[index];
1868 		buf = _umem_cache_alloc(cp, umflag);
1869 		if ((cp->cache_flags & UMF_BUFTAG) && buf != NULL) {
1870 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1871 			((uint8_t *)buf)[size] = UMEM_REDZONE_BYTE;
1872 			((uint32_t *)btp)[1] = UMEM_SIZE_ENCODE(size);
1873 		}
1874 		if (buf == NULL && umem_alloc_retry(cp, umflag))
1875 			goto umem_alloc_retry;
1876 		return (buf);
1877 	}
1878 	if (size == 0)
1879 		return (NULL);
1880 	if (umem_oversize_arena == NULL) {
1881 		if (umem_init())
1882 			ASSERT(umem_oversize_arena != NULL);
1883 		else
1884 			return (NULL);
1885 	}
1886 	buf = vmem_alloc(umem_oversize_arena, size, UMEM_VMFLAGS(umflag));
1887 	if (buf == NULL) {
1888 		umem_log_event(umem_failure_log, NULL, NULL, (void *)size);
1889 		if (umem_alloc_retry(NULL, umflag))
1890 			goto umem_alloc_retry;
1891 	}
1892 	return (buf);
1893 }
1894 
1895 #pragma weak umem_alloc_align = _umem_alloc_align
1896 void *
1897 _umem_alloc_align(size_t size, size_t align, int umflag)
1898 {
1899 	void *buf;
1900 
1901 	if (size == 0)
1902 		return (NULL);
1903 	if ((align & (align - 1)) != 0)
1904 		return (NULL);
1905 	if (align < UMEM_ALIGN)
1906 		align = UMEM_ALIGN;
1907 
1908 umem_alloc_align_retry:
1909 	if (umem_memalign_arena == NULL) {
1910 		if (umem_init())
1911 			ASSERT(umem_oversize_arena != NULL);
1912 		else
1913 			return (NULL);
1914 	}
1915 	buf = vmem_xalloc(umem_memalign_arena, size, align, 0, 0, NULL, NULL,
1916 	    UMEM_VMFLAGS(umflag));
1917 	if (buf == NULL) {
1918 		umem_log_event(umem_failure_log, NULL, NULL, (void *)size);
1919 		if (umem_alloc_retry(NULL, umflag))
1920 			goto umem_alloc_align_retry;
1921 	}
1922 	return (buf);
1923 }
1924 
1925 #pragma weak umem_free = _umem_free
1926 void
1927 _umem_free(void *buf, size_t size)
1928 {
1929 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
1930 
1931 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
1932 		umem_cache_t *cp = umem_alloc_table[index];
1933 		if (cp->cache_flags & UMF_BUFTAG) {
1934 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1935 			uint32_t *ip = (uint32_t *)btp;
1936 			if (ip[1] != UMEM_SIZE_ENCODE(size)) {
1937 				if (*(uint64_t *)buf == UMEM_FREE_PATTERN) {
1938 					umem_error(UMERR_DUPFREE, cp, buf);
1939 					return;
1940 				}
1941 				if (UMEM_SIZE_VALID(ip[1])) {
1942 					ip[0] = UMEM_SIZE_ENCODE(size);
1943 					umem_error(UMERR_BADSIZE, cp, buf);
1944 				} else {
1945 					umem_error(UMERR_REDZONE, cp, buf);
1946 				}
1947 				return;
1948 			}
1949 			if (((uint8_t *)buf)[size] != UMEM_REDZONE_BYTE) {
1950 				umem_error(UMERR_REDZONE, cp, buf);
1951 				return;
1952 			}
1953 			btp->bt_redzone = UMEM_REDZONE_PATTERN;
1954 		}
1955 		_umem_cache_free(cp, buf);
1956 	} else {
1957 		if (buf == NULL && size == 0)
1958 			return;
1959 		vmem_free(umem_oversize_arena, buf, size);
1960 	}
1961 }
1962 
1963 #pragma weak umem_free_align = _umem_free_align
1964 void
1965 _umem_free_align(void *buf, size_t size)
1966 {
1967 	if (buf == NULL && size == 0)
1968 		return;
1969 	vmem_xfree(umem_memalign_arena, buf, size);
1970 }
1971 
1972 static void *
1973 umem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
1974 {
1975 	size_t realsize = size + vmp->vm_quantum;
1976 
1977 	/*
1978 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
1979 	 * vm_quantum will cause integer wraparound.  Check for this, and
1980 	 * blow off the firewall page in this case.  Note that such a
1981 	 * giant allocation (the entire address space) can never be
1982 	 * satisfied, so it will either fail immediately (VM_NOSLEEP)
1983 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
1984 	 * corresponding check in umem_firewall_va_free().
1985 	 */
1986 	if (realsize < size)
1987 		realsize = size;
1988 
1989 	return (vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT));
1990 }
1991 
1992 static void
1993 umem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
1994 {
1995 	vmem_free(vmp, addr, size + vmp->vm_quantum);
1996 }
1997 
1998 /*
1999  * Reclaim all unused memory from a cache.
2000  */
2001 static void
2002 umem_cache_reap(umem_cache_t *cp)
2003 {
2004 	/*
2005 	 * Ask the cache's owner to free some memory if possible.
2006 	 * The idea is to handle things like the inode cache, which
2007 	 * typically sits on a bunch of memory that it doesn't truly
2008 	 * *need*.  Reclaim policy is entirely up to the owner; this
2009 	 * callback is just an advisory plea for help.
2010 	 */
2011 	if (cp->cache_reclaim != NULL)
2012 		cp->cache_reclaim(cp->cache_private);
2013 
2014 	umem_depot_ws_reap(cp);
2015 }
2016 
2017 /*
2018  * Purge all magazines from a cache and set its magazine limit to zero.
2019  * All calls are serialized by being done by the update thread, except for
2020  * the final call from umem_cache_destroy().
2021  */
2022 static void
2023 umem_cache_magazine_purge(umem_cache_t *cp)
2024 {
2025 	umem_cpu_cache_t *ccp;
2026 	umem_magazine_t *mp, *pmp;
2027 	int rounds, prounds, cpu_seqid;
2028 
2029 	ASSERT(cp->cache_next == NULL || IN_UPDATE());
2030 
2031 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
2032 		ccp = &cp->cache_cpu[cpu_seqid];
2033 
2034 		(void) mutex_lock(&ccp->cc_lock);
2035 		mp = ccp->cc_loaded;
2036 		pmp = ccp->cc_ploaded;
2037 		rounds = ccp->cc_rounds;
2038 		prounds = ccp->cc_prounds;
2039 		ccp->cc_loaded = NULL;
2040 		ccp->cc_ploaded = NULL;
2041 		ccp->cc_rounds = -1;
2042 		ccp->cc_prounds = -1;
2043 		ccp->cc_magsize = 0;
2044 		(void) mutex_unlock(&ccp->cc_lock);
2045 
2046 		if (mp)
2047 			umem_magazine_destroy(cp, mp, rounds);
2048 		if (pmp)
2049 			umem_magazine_destroy(cp, pmp, prounds);
2050 	}
2051 
2052 	/*
2053 	 * Updating the working set statistics twice in a row has the
2054 	 * effect of setting the working set size to zero, so everything
2055 	 * is eligible for reaping.
2056 	 */
2057 	umem_depot_ws_update(cp);
2058 	umem_depot_ws_update(cp);
2059 
2060 	umem_depot_ws_reap(cp);
2061 }
2062 
2063 /*
2064  * Enable per-cpu magazines on a cache.
2065  */
2066 static void
2067 umem_cache_magazine_enable(umem_cache_t *cp)
2068 {
2069 	int cpu_seqid;
2070 
2071 	if (cp->cache_flags & UMF_NOMAGAZINE)
2072 		return;
2073 
2074 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
2075 		umem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
2076 		(void) mutex_lock(&ccp->cc_lock);
2077 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
2078 		(void) mutex_unlock(&ccp->cc_lock);
2079 	}
2080 
2081 }
2082 
2083 /*
2084  * Recompute a cache's magazine size.  The trade-off is that larger magazines
2085  * provide a higher transfer rate with the depot, while smaller magazines
2086  * reduce memory consumption.  Magazine resizing is an expensive operation;
2087  * it should not be done frequently.
2088  *
2089  * Changes to the magazine size are serialized by only having one thread
2090  * doing updates. (the update thread)
2091  *
2092  * Note: at present this only grows the magazine size.  It might be useful
2093  * to allow shrinkage too.
2094  */
2095 static void
2096 umem_cache_magazine_resize(umem_cache_t *cp)
2097 {
2098 	umem_magtype_t *mtp = cp->cache_magtype;
2099 
2100 	ASSERT(IN_UPDATE());
2101 
2102 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
2103 		umem_cache_magazine_purge(cp);
2104 		(void) mutex_lock(&cp->cache_depot_lock);
2105 		cp->cache_magtype = ++mtp;
2106 		cp->cache_depot_contention_prev =
2107 		    cp->cache_depot_contention + INT_MAX;
2108 		(void) mutex_unlock(&cp->cache_depot_lock);
2109 		umem_cache_magazine_enable(cp);
2110 	}
2111 }
2112 
2113 /*
2114  * Rescale a cache's hash table, so that the table size is roughly the
2115  * cache size.  We want the average lookup time to be extremely small.
2116  */
2117 static void
2118 umem_hash_rescale(umem_cache_t *cp)
2119 {
2120 	umem_bufctl_t **old_table, **new_table, *bcp;
2121 	size_t old_size, new_size, h;
2122 
2123 	ASSERT(IN_UPDATE());
2124 
2125 	new_size = MAX(UMEM_HASH_INITIAL,
2126 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
2127 	old_size = cp->cache_hash_mask + 1;
2128 
2129 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
2130 		return;
2131 
2132 	new_table = vmem_alloc(umem_hash_arena, new_size * sizeof (void *),
2133 	    VM_NOSLEEP);
2134 	if (new_table == NULL)
2135 		return;
2136 	bzero(new_table, new_size * sizeof (void *));
2137 
2138 	(void) mutex_lock(&cp->cache_lock);
2139 
2140 	old_size = cp->cache_hash_mask + 1;
2141 	old_table = cp->cache_hash_table;
2142 
2143 	cp->cache_hash_mask = new_size - 1;
2144 	cp->cache_hash_table = new_table;
2145 	cp->cache_rescale++;
2146 
2147 	for (h = 0; h < old_size; h++) {
2148 		bcp = old_table[h];
2149 		while (bcp != NULL) {
2150 			void *addr = bcp->bc_addr;
2151 			umem_bufctl_t *next_bcp = bcp->bc_next;
2152 			umem_bufctl_t **hash_bucket = UMEM_HASH(cp, addr);
2153 			bcp->bc_next = *hash_bucket;
2154 			*hash_bucket = bcp;
2155 			bcp = next_bcp;
2156 		}
2157 	}
2158 
2159 	(void) mutex_unlock(&cp->cache_lock);
2160 
2161 	vmem_free(umem_hash_arena, old_table, old_size * sizeof (void *));
2162 }
2163 
2164 /*
2165  * Perform periodic maintenance on a cache: hash rescaling,
2166  * depot working-set update, and magazine resizing.
2167  */
2168 void
2169 umem_cache_update(umem_cache_t *cp)
2170 {
2171 	int update_flags = 0;
2172 
2173 	ASSERT(MUTEX_HELD(&umem_cache_lock));
2174 
2175 	/*
2176 	 * If the cache has become much larger or smaller than its hash table,
2177 	 * fire off a request to rescale the hash table.
2178 	 */
2179 	(void) mutex_lock(&cp->cache_lock);
2180 
2181 	if ((cp->cache_flags & UMF_HASH) &&
2182 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
2183 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
2184 	    cp->cache_hash_mask > UMEM_HASH_INITIAL)))
2185 		update_flags |= UMU_HASH_RESCALE;
2186 
2187 	(void) mutex_unlock(&cp->cache_lock);
2188 
2189 	/*
2190 	 * Update the depot working set statistics.
2191 	 */
2192 	umem_depot_ws_update(cp);
2193 
2194 	/*
2195 	 * If there's a lot of contention in the depot,
2196 	 * increase the magazine size.
2197 	 */
2198 	(void) mutex_lock(&cp->cache_depot_lock);
2199 
2200 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
2201 	    (int)(cp->cache_depot_contention -
2202 	    cp->cache_depot_contention_prev) > umem_depot_contention)
2203 		update_flags |= UMU_MAGAZINE_RESIZE;
2204 
2205 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
2206 
2207 	(void) mutex_unlock(&cp->cache_depot_lock);
2208 
2209 	if (update_flags)
2210 		umem_add_update(cp, update_flags);
2211 }
2212 
2213 /*
2214  * Runs all pending updates.
2215  *
2216  * The update lock must be held on entrance, and will be held on exit.
2217  */
2218 void
2219 umem_process_updates(void)
2220 {
2221 	ASSERT(MUTEX_HELD(&umem_update_lock));
2222 
2223 	while (umem_null_cache.cache_unext != &umem_null_cache) {
2224 		int notify = 0;
2225 		umem_cache_t *cp = umem_null_cache.cache_unext;
2226 
2227 		cp->cache_uprev->cache_unext = cp->cache_unext;
2228 		cp->cache_unext->cache_uprev = cp->cache_uprev;
2229 		cp->cache_uprev = cp->cache_unext = NULL;
2230 
2231 		ASSERT(!(cp->cache_uflags & UMU_ACTIVE));
2232 
2233 		while (cp->cache_uflags) {
2234 			int uflags = (cp->cache_uflags |= UMU_ACTIVE);
2235 			(void) mutex_unlock(&umem_update_lock);
2236 
2237 			/*
2238 			 * The order here is important.  Each step can speed up
2239 			 * later steps.
2240 			 */
2241 
2242 			if (uflags & UMU_HASH_RESCALE)
2243 				umem_hash_rescale(cp);
2244 
2245 			if (uflags & UMU_MAGAZINE_RESIZE)
2246 				umem_cache_magazine_resize(cp);
2247 
2248 			if (uflags & UMU_REAP)
2249 				umem_cache_reap(cp);
2250 
2251 			(void) mutex_lock(&umem_update_lock);
2252 
2253 			/*
2254 			 * check if anyone has requested notification
2255 			 */
2256 			if (cp->cache_uflags & UMU_NOTIFY) {
2257 				uflags |= UMU_NOTIFY;
2258 				notify = 1;
2259 			}
2260 			cp->cache_uflags &= ~uflags;
2261 		}
2262 		if (notify)
2263 			(void) cond_broadcast(&umem_update_cv);
2264 	}
2265 }
2266 
2267 #ifndef UMEM_STANDALONE
2268 static void
2269 umem_st_update(void)
2270 {
2271 	ASSERT(MUTEX_HELD(&umem_update_lock));
2272 	ASSERT(umem_update_thr == 0 && umem_st_update_thr == 0);
2273 
2274 	umem_st_update_thr = thr_self();
2275 
2276 	(void) mutex_unlock(&umem_update_lock);
2277 
2278 	vmem_update(NULL);
2279 	umem_cache_applyall(umem_cache_update);
2280 
2281 	(void) mutex_lock(&umem_update_lock);
2282 
2283 	umem_process_updates();	/* does all of the requested work */
2284 
2285 	umem_reap_next = gethrtime() +
2286 	    (hrtime_t)umem_reap_interval * NANOSEC;
2287 
2288 	umem_reaping = UMEM_REAP_DONE;
2289 
2290 	umem_st_update_thr = 0;
2291 }
2292 #endif
2293 
2294 /*
2295  * Reclaim all unused memory from all caches.  Called from vmem when memory
2296  * gets tight.  Must be called with no locks held.
2297  *
2298  * This just requests a reap on all caches, and notifies the update thread.
2299  */
2300 void
2301 umem_reap(void)
2302 {
2303 #ifndef UMEM_STANDALONE
2304 	extern int __nthreads(void);
2305 #endif
2306 
2307 	if (umem_ready != UMEM_READY || umem_reaping != UMEM_REAP_DONE ||
2308 	    gethrtime() < umem_reap_next)
2309 		return;
2310 
2311 	(void) mutex_lock(&umem_update_lock);
2312 
2313 	if (umem_reaping != UMEM_REAP_DONE || gethrtime() < umem_reap_next) {
2314 		(void) mutex_unlock(&umem_update_lock);
2315 		return;
2316 	}
2317 	umem_reaping = UMEM_REAP_ADDING;	/* lock out other reaps */
2318 
2319 	(void) mutex_unlock(&umem_update_lock);
2320 
2321 	umem_updateall(UMU_REAP);
2322 
2323 	(void) mutex_lock(&umem_update_lock);
2324 
2325 	umem_reaping = UMEM_REAP_ACTIVE;
2326 
2327 	/* Standalone is single-threaded */
2328 #ifndef UMEM_STANDALONE
2329 	if (umem_update_thr == 0) {
2330 		/*
2331 		 * The update thread does not exist.  If the process is
2332 		 * multi-threaded, create it.  If not, or the creation fails,
2333 		 * do the update processing inline.
2334 		 */
2335 		ASSERT(umem_st_update_thr == 0);
2336 
2337 		if (__nthreads() <= 1 || umem_create_update_thread() == 0)
2338 			umem_st_update();
2339 	}
2340 
2341 	(void) cond_broadcast(&umem_update_cv);	/* wake up the update thread */
2342 #endif
2343 
2344 	(void) mutex_unlock(&umem_update_lock);
2345 }
2346 
2347 umem_cache_t *
2348 umem_cache_create(
2349 	char *name,		/* descriptive name for this cache */
2350 	size_t bufsize,		/* size of the objects it manages */
2351 	size_t align,		/* required object alignment */
2352 	umem_constructor_t *constructor, /* object constructor */
2353 	umem_destructor_t *destructor, /* object destructor */
2354 	umem_reclaim_t *reclaim, /* memory reclaim callback */
2355 	void *private,		/* pass-thru arg for constr/destr/reclaim */
2356 	vmem_t *vmp,		/* vmem source for slab allocation */
2357 	int cflags)		/* cache creation flags */
2358 {
2359 	int cpu_seqid;
2360 	size_t chunksize;
2361 	umem_cache_t *cp, *cnext, *cprev;
2362 	umem_magtype_t *mtp;
2363 	size_t csize;
2364 	size_t phase;
2365 
2366 	/*
2367 	 * The init thread is allowed to create internal and quantum caches.
2368 	 *
2369 	 * Other threads must wait until until initialization is complete.
2370 	 */
2371 	if (umem_init_thr == thr_self())
2372 		ASSERT((cflags & (UMC_INTERNAL | UMC_QCACHE)) != 0);
2373 	else {
2374 		ASSERT(!(cflags & UMC_INTERNAL));
2375 		if (umem_ready != UMEM_READY && umem_init() == 0) {
2376 			errno = EAGAIN;
2377 			return (NULL);
2378 		}
2379 	}
2380 
2381 	csize = UMEM_CACHE_SIZE(umem_max_ncpus);
2382 	phase = P2NPHASE(csize, UMEM_CPU_CACHE_SIZE);
2383 
2384 	if (vmp == NULL)
2385 		vmp = umem_default_arena;
2386 
2387 	ASSERT(P2PHASE(phase, UMEM_ALIGN) == 0);
2388 
2389 	/*
2390 	 * Check that the arguments are reasonable
2391 	 */
2392 	if ((align & (align - 1)) != 0 || align > vmp->vm_quantum ||
2393 	    ((cflags & UMC_NOHASH) && (cflags & UMC_NOTOUCH)) ||
2394 	    name == NULL || bufsize == 0) {
2395 		errno = EINVAL;
2396 		return (NULL);
2397 	}
2398 
2399 	/*
2400 	 * If align == 0, we set it to the minimum required alignment.
2401 	 *
2402 	 * If align < UMEM_ALIGN, we round it up to UMEM_ALIGN, unless
2403 	 * UMC_NOTOUCH was passed.
2404 	 */
2405 	if (align == 0) {
2406 		if (P2ROUNDUP(bufsize, UMEM_ALIGN) >= UMEM_SECOND_ALIGN)
2407 			align = UMEM_SECOND_ALIGN;
2408 		else
2409 			align = UMEM_ALIGN;
2410 	} else if (align < UMEM_ALIGN && (cflags & UMC_NOTOUCH) == 0)
2411 		align = UMEM_ALIGN;
2412 
2413 
2414 	/*
2415 	 * Get a umem_cache structure.  We arrange that cp->cache_cpu[]
2416 	 * is aligned on a UMEM_CPU_CACHE_SIZE boundary to prevent
2417 	 * false sharing of per-CPU data.
2418 	 */
2419 	cp = vmem_xalloc(umem_cache_arena, csize, UMEM_CPU_CACHE_SIZE, phase,
2420 	    0, NULL, NULL, VM_NOSLEEP);
2421 
2422 	if (cp == NULL) {
2423 		errno = EAGAIN;
2424 		return (NULL);
2425 	}
2426 
2427 	bzero(cp, csize);
2428 
2429 	(void) mutex_lock(&umem_flags_lock);
2430 	if (umem_flags & UMF_RANDOMIZE)
2431 		umem_flags = (((umem_flags | ~UMF_RANDOM) + 1) & UMF_RANDOM) |
2432 		    UMF_RANDOMIZE;
2433 	cp->cache_flags = umem_flags | (cflags & UMF_DEBUG);
2434 	(void) mutex_unlock(&umem_flags_lock);
2435 
2436 	/*
2437 	 * Make sure all the various flags are reasonable.
2438 	 */
2439 	if (cp->cache_flags & UMF_LITE) {
2440 		if (bufsize >= umem_lite_minsize &&
2441 		    align <= umem_lite_maxalign &&
2442 		    P2PHASE(bufsize, umem_lite_maxalign) != 0) {
2443 			cp->cache_flags |= UMF_BUFTAG;
2444 			cp->cache_flags &= ~(UMF_AUDIT | UMF_FIREWALL);
2445 		} else {
2446 			cp->cache_flags &= ~UMF_DEBUG;
2447 		}
2448 	}
2449 
2450 	if ((cflags & UMC_QCACHE) && (cp->cache_flags & UMF_AUDIT))
2451 		cp->cache_flags |= UMF_NOMAGAZINE;
2452 
2453 	if (cflags & UMC_NODEBUG)
2454 		cp->cache_flags &= ~UMF_DEBUG;
2455 
2456 	if (cflags & UMC_NOTOUCH)
2457 		cp->cache_flags &= ~UMF_TOUCH;
2458 
2459 	if (cflags & UMC_NOHASH)
2460 		cp->cache_flags &= ~(UMF_AUDIT | UMF_FIREWALL);
2461 
2462 	if (cflags & UMC_NOMAGAZINE)
2463 		cp->cache_flags |= UMF_NOMAGAZINE;
2464 
2465 	if ((cp->cache_flags & UMF_AUDIT) && !(cflags & UMC_NOTOUCH))
2466 		cp->cache_flags |= UMF_REDZONE;
2467 
2468 	if ((cp->cache_flags & UMF_BUFTAG) && bufsize >= umem_minfirewall &&
2469 	    !(cp->cache_flags & UMF_LITE) && !(cflags & UMC_NOHASH))
2470 		cp->cache_flags |= UMF_FIREWALL;
2471 
2472 	if (vmp != umem_default_arena || umem_firewall_arena == NULL)
2473 		cp->cache_flags &= ~UMF_FIREWALL;
2474 
2475 	if (cp->cache_flags & UMF_FIREWALL) {
2476 		cp->cache_flags &= ~UMF_BUFTAG;
2477 		cp->cache_flags |= UMF_NOMAGAZINE;
2478 		ASSERT(vmp == umem_default_arena);
2479 		vmp = umem_firewall_arena;
2480 	}
2481 
2482 	/*
2483 	 * Set cache properties.
2484 	 */
2485 	(void) strncpy(cp->cache_name, name, sizeof (cp->cache_name) - 1);
2486 	cp->cache_bufsize = bufsize;
2487 	cp->cache_align = align;
2488 	cp->cache_constructor = constructor;
2489 	cp->cache_destructor = destructor;
2490 	cp->cache_reclaim = reclaim;
2491 	cp->cache_private = private;
2492 	cp->cache_arena = vmp;
2493 	cp->cache_cflags = cflags;
2494 	cp->cache_cpu_mask = umem_cpu_mask;
2495 
2496 	/*
2497 	 * Determine the chunk size.
2498 	 */
2499 	chunksize = bufsize;
2500 
2501 	if (align >= UMEM_ALIGN) {
2502 		chunksize = P2ROUNDUP(chunksize, UMEM_ALIGN);
2503 		cp->cache_bufctl = chunksize - UMEM_ALIGN;
2504 	}
2505 
2506 	if (cp->cache_flags & UMF_BUFTAG) {
2507 		cp->cache_bufctl = chunksize;
2508 		cp->cache_buftag = chunksize;
2509 		chunksize += sizeof (umem_buftag_t);
2510 	}
2511 
2512 	if (cp->cache_flags & UMF_DEADBEEF) {
2513 		cp->cache_verify = MIN(cp->cache_buftag, umem_maxverify);
2514 		if (cp->cache_flags & UMF_LITE)
2515 			cp->cache_verify = MIN(cp->cache_verify, UMEM_ALIGN);
2516 	}
2517 
2518 	cp->cache_contents = MIN(cp->cache_bufctl, umem_content_maxsave);
2519 
2520 	cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
2521 
2522 	if (chunksize < bufsize) {
2523 		errno = ENOMEM;
2524 		goto fail;
2525 	}
2526 
2527 	/*
2528 	 * Now that we know the chunk size, determine the optimal slab size.
2529 	 */
2530 	if (vmp == umem_firewall_arena) {
2531 		cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
2532 		cp->cache_mincolor = cp->cache_slabsize - chunksize;
2533 		cp->cache_maxcolor = cp->cache_mincolor;
2534 		cp->cache_flags |= UMF_HASH;
2535 		ASSERT(!(cp->cache_flags & UMF_BUFTAG));
2536 	} else if ((cflags & UMC_NOHASH) || (!(cflags & UMC_NOTOUCH) &&
2537 	    !(cp->cache_flags & UMF_AUDIT) &&
2538 	    chunksize < vmp->vm_quantum / UMEM_VOID_FRACTION)) {
2539 		cp->cache_slabsize = vmp->vm_quantum;
2540 		cp->cache_mincolor = 0;
2541 		cp->cache_maxcolor =
2542 		    (cp->cache_slabsize - sizeof (umem_slab_t)) % chunksize;
2543 
2544 		if (chunksize + sizeof (umem_slab_t) > cp->cache_slabsize) {
2545 			errno = EINVAL;
2546 			goto fail;
2547 		}
2548 		ASSERT(!(cp->cache_flags & UMF_AUDIT));
2549 	} else {
2550 		size_t chunks, bestfit, waste, slabsize;
2551 		size_t minwaste = LONG_MAX;
2552 
2553 		for (chunks = 1; chunks <= UMEM_VOID_FRACTION; chunks++) {
2554 			slabsize = P2ROUNDUP(chunksize * chunks,
2555 			    vmp->vm_quantum);
2556 			/*
2557 			 * check for overflow
2558 			 */
2559 			if ((slabsize / chunks) < chunksize) {
2560 				errno = ENOMEM;
2561 				goto fail;
2562 			}
2563 			chunks = slabsize / chunksize;
2564 			waste = (slabsize % chunksize) / chunks;
2565 			if (waste < minwaste) {
2566 				minwaste = waste;
2567 				bestfit = slabsize;
2568 			}
2569 		}
2570 		if (cflags & UMC_QCACHE)
2571 			bestfit = MAX(1 << highbit(3 * vmp->vm_qcache_max), 64);
2572 		cp->cache_slabsize = bestfit;
2573 		cp->cache_mincolor = 0;
2574 		cp->cache_maxcolor = bestfit % chunksize;
2575 		cp->cache_flags |= UMF_HASH;
2576 	}
2577 
2578 	if (cp->cache_flags & UMF_HASH) {
2579 		ASSERT(!(cflags & UMC_NOHASH));
2580 		cp->cache_bufctl_cache = (cp->cache_flags & UMF_AUDIT) ?
2581 		    umem_bufctl_audit_cache : umem_bufctl_cache;
2582 	}
2583 
2584 	if (cp->cache_maxcolor >= vmp->vm_quantum)
2585 		cp->cache_maxcolor = vmp->vm_quantum - 1;
2586 
2587 	cp->cache_color = cp->cache_mincolor;
2588 
2589 	/*
2590 	 * Initialize the rest of the slab layer.
2591 	 */
2592 	(void) mutex_init(&cp->cache_lock, USYNC_THREAD, NULL);
2593 
2594 	cp->cache_freelist = &cp->cache_nullslab;
2595 	cp->cache_nullslab.slab_cache = cp;
2596 	cp->cache_nullslab.slab_refcnt = -1;
2597 	cp->cache_nullslab.slab_next = &cp->cache_nullslab;
2598 	cp->cache_nullslab.slab_prev = &cp->cache_nullslab;
2599 
2600 	if (cp->cache_flags & UMF_HASH) {
2601 		cp->cache_hash_table = vmem_alloc(umem_hash_arena,
2602 		    UMEM_HASH_INITIAL * sizeof (void *), VM_NOSLEEP);
2603 		if (cp->cache_hash_table == NULL) {
2604 			errno = EAGAIN;
2605 			goto fail_lock;
2606 		}
2607 		bzero(cp->cache_hash_table,
2608 		    UMEM_HASH_INITIAL * sizeof (void *));
2609 		cp->cache_hash_mask = UMEM_HASH_INITIAL - 1;
2610 		cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
2611 	}
2612 
2613 	/*
2614 	 * Initialize the depot.
2615 	 */
2616 	(void) mutex_init(&cp->cache_depot_lock, USYNC_THREAD, NULL);
2617 
2618 	for (mtp = umem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
2619 		continue;
2620 
2621 	cp->cache_magtype = mtp;
2622 
2623 	/*
2624 	 * Initialize the CPU layer.
2625 	 */
2626 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
2627 		umem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
2628 		(void) mutex_init(&ccp->cc_lock, USYNC_THREAD, NULL);
2629 		ccp->cc_flags = cp->cache_flags;
2630 		ccp->cc_rounds = -1;
2631 		ccp->cc_prounds = -1;
2632 	}
2633 
2634 	/*
2635 	 * Add the cache to the global list.  This makes it visible
2636 	 * to umem_update(), so the cache must be ready for business.
2637 	 */
2638 	(void) mutex_lock(&umem_cache_lock);
2639 	cp->cache_next = cnext = &umem_null_cache;
2640 	cp->cache_prev = cprev = umem_null_cache.cache_prev;
2641 	cnext->cache_prev = cp;
2642 	cprev->cache_next = cp;
2643 	(void) mutex_unlock(&umem_cache_lock);
2644 
2645 	if (umem_ready == UMEM_READY)
2646 		umem_cache_magazine_enable(cp);
2647 
2648 	return (cp);
2649 
2650 fail_lock:
2651 	(void) mutex_destroy(&cp->cache_lock);
2652 fail:
2653 	vmem_xfree(umem_cache_arena, cp, csize);
2654 	return (NULL);
2655 }
2656 
2657 void
2658 umem_cache_destroy(umem_cache_t *cp)
2659 {
2660 	int cpu_seqid;
2661 
2662 	/*
2663 	 * Remove the cache from the global cache list so that no new updates
2664 	 * will be scheduled on its behalf, wait for any pending tasks to
2665 	 * complete, purge the cache, and then destroy it.
2666 	 */
2667 	(void) mutex_lock(&umem_cache_lock);
2668 	cp->cache_prev->cache_next = cp->cache_next;
2669 	cp->cache_next->cache_prev = cp->cache_prev;
2670 	cp->cache_prev = cp->cache_next = NULL;
2671 	(void) mutex_unlock(&umem_cache_lock);
2672 
2673 	umem_remove_updates(cp);
2674 
2675 	umem_cache_magazine_purge(cp);
2676 
2677 	(void) mutex_lock(&cp->cache_lock);
2678 	if (cp->cache_buftotal != 0)
2679 		log_message("umem_cache_destroy: '%s' (%p) not empty\n",
2680 		    cp->cache_name, (void *)cp);
2681 	cp->cache_reclaim = NULL;
2682 	/*
2683 	 * The cache is now dead.  There should be no further activity.
2684 	 * We enforce this by setting land mines in the constructor and
2685 	 * destructor routines that induce a segmentation fault if invoked.
2686 	 */
2687 	cp->cache_constructor = (umem_constructor_t *)1;
2688 	cp->cache_destructor = (umem_destructor_t *)2;
2689 	(void) mutex_unlock(&cp->cache_lock);
2690 
2691 	if (cp->cache_hash_table != NULL)
2692 		vmem_free(umem_hash_arena, cp->cache_hash_table,
2693 		    (cp->cache_hash_mask + 1) * sizeof (void *));
2694 
2695 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++)
2696 		(void) mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
2697 
2698 	(void) mutex_destroy(&cp->cache_depot_lock);
2699 	(void) mutex_destroy(&cp->cache_lock);
2700 
2701 	vmem_free(umem_cache_arena, cp, UMEM_CACHE_SIZE(umem_max_ncpus));
2702 }
2703 
2704 static int
2705 umem_cache_init(void)
2706 {
2707 	int i;
2708 	size_t size, max_size;
2709 	umem_cache_t *cp;
2710 	umem_magtype_t *mtp;
2711 	char name[UMEM_CACHE_NAMELEN + 1];
2712 	umem_cache_t *umem_alloc_caches[NUM_ALLOC_SIZES];
2713 
2714 	for (i = 0; i < sizeof (umem_magtype) / sizeof (*mtp); i++) {
2715 		mtp = &umem_magtype[i];
2716 		(void) snprintf(name, sizeof (name), "umem_magazine_%d",
2717 		    mtp->mt_magsize);
2718 		mtp->mt_cache = umem_cache_create(name,
2719 		    (mtp->mt_magsize + 1) * sizeof (void *),
2720 		    mtp->mt_align, NULL, NULL, NULL, NULL,
2721 		    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
2722 		if (mtp->mt_cache == NULL)
2723 			return (0);
2724 	}
2725 
2726 	umem_slab_cache = umem_cache_create("umem_slab_cache",
2727 	    sizeof (umem_slab_t), 0, NULL, NULL, NULL, NULL,
2728 	    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
2729 
2730 	if (umem_slab_cache == NULL)
2731 		return (0);
2732 
2733 	umem_bufctl_cache = umem_cache_create("umem_bufctl_cache",
2734 	    sizeof (umem_bufctl_t), 0, NULL, NULL, NULL, NULL,
2735 	    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
2736 
2737 	if (umem_bufctl_cache == NULL)
2738 		return (0);
2739 
2740 	/*
2741 	 * The size of the umem_bufctl_audit structure depends upon
2742 	 * umem_stack_depth.   See umem_impl.h for details on the size
2743 	 * restrictions.
2744 	 */
2745 
2746 	size = UMEM_BUFCTL_AUDIT_SIZE_DEPTH(umem_stack_depth);
2747 	max_size = UMEM_BUFCTL_AUDIT_MAX_SIZE;
2748 
2749 	if (size > max_size) {			/* too large -- truncate */
2750 		int max_frames = UMEM_MAX_STACK_DEPTH;
2751 
2752 		ASSERT(UMEM_BUFCTL_AUDIT_SIZE_DEPTH(max_frames) <= max_size);
2753 
2754 		umem_stack_depth = max_frames;
2755 		size = UMEM_BUFCTL_AUDIT_SIZE_DEPTH(umem_stack_depth);
2756 	}
2757 
2758 	umem_bufctl_audit_cache = umem_cache_create("umem_bufctl_audit_cache",
2759 	    size, 0, NULL, NULL, NULL, NULL, umem_internal_arena,
2760 	    UMC_NOHASH | UMC_INTERNAL);
2761 
2762 	if (umem_bufctl_audit_cache == NULL)
2763 		return (0);
2764 
2765 	if (vmem_backend & VMEM_BACKEND_MMAP)
2766 		umem_va_arena = vmem_create("umem_va",
2767 		    NULL, 0, pagesize,
2768 		    vmem_alloc, vmem_free, heap_arena,
2769 		    8 * pagesize, VM_NOSLEEP);
2770 	else
2771 		umem_va_arena = heap_arena;
2772 
2773 	if (umem_va_arena == NULL)
2774 		return (0);
2775 
2776 	umem_default_arena = vmem_create("umem_default",
2777 	    NULL, 0, pagesize,
2778 	    heap_alloc, heap_free, umem_va_arena,
2779 	    0, VM_NOSLEEP);
2780 
2781 	if (umem_default_arena == NULL)
2782 		return (0);
2783 
2784 	/*
2785 	 * make sure the umem_alloc table initializer is correct
2786 	 */
2787 	i = sizeof (umem_alloc_table) / sizeof (*umem_alloc_table);
2788 	ASSERT(umem_alloc_table[i - 1] == &umem_null_cache);
2789 
2790 	/*
2791 	 * Create the default caches to back umem_alloc()
2792 	 */
2793 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
2794 		size_t cache_size = umem_alloc_sizes[i];
2795 		size_t align = 0;
2796 		/*
2797 		 * If they allocate a multiple of the coherency granularity,
2798 		 * they get a coherency-granularity-aligned address.
2799 		 */
2800 		if (IS_P2ALIGNED(cache_size, 64))
2801 			align = 64;
2802 		if (IS_P2ALIGNED(cache_size, pagesize))
2803 			align = pagesize;
2804 		(void) snprintf(name, sizeof (name), "umem_alloc_%lu",
2805 		    (long)cache_size);
2806 
2807 		cp = umem_cache_create(name, cache_size, align,
2808 		    NULL, NULL, NULL, NULL, NULL, UMC_INTERNAL);
2809 		if (cp == NULL)
2810 			return (0);
2811 
2812 		umem_alloc_caches[i] = cp;
2813 	}
2814 
2815 	/*
2816 	 * Initialization cannot fail at this point.  Make the caches
2817 	 * visible to umem_alloc() and friends.
2818 	 */
2819 	size = UMEM_ALIGN;
2820 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
2821 		size_t cache_size = umem_alloc_sizes[i];
2822 
2823 		cp = umem_alloc_caches[i];
2824 
2825 		while (size <= cache_size) {
2826 			umem_alloc_table[(size - 1) >> UMEM_ALIGN_SHIFT] = cp;
2827 			size += UMEM_ALIGN;
2828 		}
2829 	}
2830 	return (1);
2831 }
2832 
2833 /*
2834  * umem_startup() is called early on, and must be called explicitly if we're
2835  * the standalone version.
2836  */
2837 #ifdef UMEM_STANDALONE
2838 void
2839 #else
2840 #pragma init(umem_startup)
2841 static void
2842 #endif
2843 umem_startup(caddr_t start, size_t len, size_t pagesize, caddr_t minstack,
2844     caddr_t maxstack)
2845 {
2846 #ifdef UMEM_STANDALONE
2847 	int idx;
2848 	/* Standalone doesn't fork */
2849 #else
2850 	umem_forkhandler_init(); /* register the fork handler */
2851 #endif
2852 
2853 #ifdef __lint
2854 	/* make lint happy */
2855 	minstack = maxstack;
2856 #endif
2857 
2858 #ifdef UMEM_STANDALONE
2859 	umem_ready = UMEM_READY_STARTUP;
2860 	umem_init_env_ready = 0;
2861 
2862 	umem_min_stack = minstack;
2863 	umem_max_stack = maxstack;
2864 
2865 	nofail_callback = NULL;
2866 	umem_slab_cache = NULL;
2867 	umem_bufctl_cache = NULL;
2868 	umem_bufctl_audit_cache = NULL;
2869 	heap_arena = NULL;
2870 	heap_alloc = NULL;
2871 	heap_free = NULL;
2872 	umem_internal_arena = NULL;
2873 	umem_cache_arena = NULL;
2874 	umem_hash_arena = NULL;
2875 	umem_log_arena = NULL;
2876 	umem_oversize_arena = NULL;
2877 	umem_va_arena = NULL;
2878 	umem_default_arena = NULL;
2879 	umem_firewall_va_arena = NULL;
2880 	umem_firewall_arena = NULL;
2881 	umem_memalign_arena = NULL;
2882 	umem_transaction_log = NULL;
2883 	umem_content_log = NULL;
2884 	umem_failure_log = NULL;
2885 	umem_slab_log = NULL;
2886 	umem_cpu_mask = 0;
2887 
2888 	umem_cpus = &umem_startup_cpu;
2889 	umem_startup_cpu.cpu_cache_offset = UMEM_CACHE_SIZE(0);
2890 	umem_startup_cpu.cpu_number = 0;
2891 
2892 	bcopy(&umem_null_cache_template, &umem_null_cache,
2893 	    sizeof (umem_cache_t));
2894 
2895 	for (idx = 0; idx < (UMEM_MAXBUF >> UMEM_ALIGN_SHIFT); idx++)
2896 		umem_alloc_table[idx] = &umem_null_cache;
2897 #endif
2898 
2899 	/*
2900 	 * Perform initialization specific to the way we've been compiled
2901 	 * (library or standalone)
2902 	 */
2903 	umem_type_init(start, len, pagesize);
2904 
2905 	vmem_startup();
2906 }
2907 
2908 int
2909 umem_init(void)
2910 {
2911 	size_t maxverify, minfirewall;
2912 	size_t size;
2913 	int idx;
2914 	umem_cpu_t *new_cpus;
2915 
2916 	vmem_t *memalign_arena, *oversize_arena;
2917 
2918 	if (thr_self() != umem_init_thr) {
2919 		/*
2920 		 * The usual case -- non-recursive invocation of umem_init().
2921 		 */
2922 		(void) mutex_lock(&umem_init_lock);
2923 		if (umem_ready != UMEM_READY_STARTUP) {
2924 			/*
2925 			 * someone else beat us to initializing umem.  Wait
2926 			 * for them to complete, then return.
2927 			 */
2928 			while (umem_ready == UMEM_READY_INITING)
2929 				(void) _cond_wait(&umem_init_cv,
2930 				    &umem_init_lock);
2931 			ASSERT(umem_ready == UMEM_READY ||
2932 			    umem_ready == UMEM_READY_INIT_FAILED);
2933 			(void) mutex_unlock(&umem_init_lock);
2934 			return (umem_ready == UMEM_READY);
2935 		}
2936 
2937 		ASSERT(umem_ready == UMEM_READY_STARTUP);
2938 		ASSERT(umem_init_env_ready == 0);
2939 
2940 		umem_ready = UMEM_READY_INITING;
2941 		umem_init_thr = thr_self();
2942 
2943 		(void) mutex_unlock(&umem_init_lock);
2944 		umem_setup_envvars(0);		/* can recurse -- see below */
2945 		if (umem_init_env_ready) {
2946 			/*
2947 			 * initialization was completed already
2948 			 */
2949 			ASSERT(umem_ready == UMEM_READY ||
2950 			    umem_ready == UMEM_READY_INIT_FAILED);
2951 			ASSERT(umem_init_thr == 0);
2952 			return (umem_ready == UMEM_READY);
2953 		}
2954 	} else if (!umem_init_env_ready) {
2955 		/*
2956 		 * The umem_setup_envvars() call (above) makes calls into
2957 		 * the dynamic linker and directly into user-supplied code.
2958 		 * Since we cannot know what that code will do, we could be
2959 		 * recursively invoked (by, say, a malloc() call in the code
2960 		 * itself, or in a (C++) _init section it causes to be fired).
2961 		 *
2962 		 * This code is where we end up if such recursion occurs.  We
2963 		 * first clean up any partial results in the envvar code, then
2964 		 * proceed to finish initialization processing in the recursive
2965 		 * call.  The original call will notice this, and return
2966 		 * immediately.
2967 		 */
2968 		umem_setup_envvars(1);		/* clean up any partial state */
2969 	} else {
2970 		umem_panic(
2971 		    "recursive allocation while initializing umem\n");
2972 	}
2973 	umem_init_env_ready = 1;
2974 
2975 	/*
2976 	 * From this point until we finish, recursion into umem_init() will
2977 	 * cause a umem_panic().
2978 	 */
2979 	maxverify = minfirewall = ULONG_MAX;
2980 
2981 	/* LINTED constant condition */
2982 	if (sizeof (umem_cpu_cache_t) != UMEM_CPU_CACHE_SIZE) {
2983 		umem_panic("sizeof (umem_cpu_cache_t) = %d, should be %d\n",
2984 		    sizeof (umem_cpu_cache_t), UMEM_CPU_CACHE_SIZE);
2985 	}
2986 
2987 	umem_max_ncpus = umem_get_max_ncpus();
2988 
2989 	/*
2990 	 * load tunables from environment
2991 	 */
2992 	umem_process_envvars();
2993 
2994 	if (issetugid())
2995 		umem_mtbf = 0;
2996 
2997 	/*
2998 	 * set up vmem
2999 	 */
3000 	if (!(umem_flags & UMF_AUDIT))
3001 		vmem_no_debug();
3002 
3003 	heap_arena = vmem_heap_arena(&heap_alloc, &heap_free);
3004 
3005 	pagesize = heap_arena->vm_quantum;
3006 
3007 	umem_internal_arena = vmem_create("umem_internal", NULL, 0, pagesize,
3008 	    heap_alloc, heap_free, heap_arena, 0, VM_NOSLEEP);
3009 
3010 	umem_default_arena = umem_internal_arena;
3011 
3012 	if (umem_internal_arena == NULL)
3013 		goto fail;
3014 
3015 	umem_cache_arena = vmem_create("umem_cache", NULL, 0, UMEM_ALIGN,
3016 	    vmem_alloc, vmem_free, umem_internal_arena, 0, VM_NOSLEEP);
3017 
3018 	umem_hash_arena = vmem_create("umem_hash", NULL, 0, UMEM_ALIGN,
3019 	    vmem_alloc, vmem_free, umem_internal_arena, 0, VM_NOSLEEP);
3020 
3021 	umem_log_arena = vmem_create("umem_log", NULL, 0, UMEM_ALIGN,
3022 	    heap_alloc, heap_free, heap_arena, 0, VM_NOSLEEP);
3023 
3024 	umem_firewall_va_arena = vmem_create("umem_firewall_va",
3025 	    NULL, 0, pagesize,
3026 	    umem_firewall_va_alloc, umem_firewall_va_free, heap_arena,
3027 	    0, VM_NOSLEEP);
3028 
3029 	if (umem_cache_arena == NULL || umem_hash_arena == NULL ||
3030 	    umem_log_arena == NULL || umem_firewall_va_arena == NULL)
3031 		goto fail;
3032 
3033 	umem_firewall_arena = vmem_create("umem_firewall", NULL, 0, pagesize,
3034 	    heap_alloc, heap_free, umem_firewall_va_arena, 0,
3035 	    VM_NOSLEEP);
3036 
3037 	if (umem_firewall_arena == NULL)
3038 		goto fail;
3039 
3040 	oversize_arena = vmem_create("umem_oversize", NULL, 0, pagesize,
3041 	    heap_alloc, heap_free, minfirewall < ULONG_MAX ?
3042 	    umem_firewall_va_arena : heap_arena, 0, VM_NOSLEEP);
3043 
3044 	memalign_arena = vmem_create("umem_memalign", NULL, 0, UMEM_ALIGN,
3045 	    heap_alloc, heap_free, minfirewall < ULONG_MAX ?
3046 	    umem_firewall_va_arena : heap_arena, 0, VM_NOSLEEP);
3047 
3048 	if (oversize_arena == NULL || memalign_arena == NULL)
3049 		goto fail;
3050 
3051 	if (umem_max_ncpus > CPUHINT_MAX())
3052 		umem_max_ncpus = CPUHINT_MAX();
3053 
3054 	while ((umem_max_ncpus & (umem_max_ncpus - 1)) != 0)
3055 		umem_max_ncpus++;
3056 
3057 	if (umem_max_ncpus == 0)
3058 		umem_max_ncpus = 1;
3059 
3060 	size = umem_max_ncpus * sizeof (umem_cpu_t);
3061 	new_cpus = vmem_alloc(umem_internal_arena, size, VM_NOSLEEP);
3062 	if (new_cpus == NULL)
3063 		goto fail;
3064 
3065 	bzero(new_cpus, size);
3066 	for (idx = 0; idx < umem_max_ncpus; idx++) {
3067 		new_cpus[idx].cpu_number = idx;
3068 		new_cpus[idx].cpu_cache_offset = UMEM_CACHE_SIZE(idx);
3069 	}
3070 	umem_cpus = new_cpus;
3071 	umem_cpu_mask = (umem_max_ncpus - 1);
3072 
3073 	if (umem_maxverify == 0)
3074 		umem_maxverify = maxverify;
3075 
3076 	if (umem_minfirewall == 0)
3077 		umem_minfirewall = minfirewall;
3078 
3079 	/*
3080 	 * Set up updating and reaping
3081 	 */
3082 	umem_reap_next = gethrtime() + NANOSEC;
3083 
3084 #ifndef UMEM_STANDALONE
3085 	(void) gettimeofday(&umem_update_next, NULL);
3086 #endif
3087 
3088 	/*
3089 	 * Set up logging -- failure here is okay, since it will just disable
3090 	 * the logs
3091 	 */
3092 	if (umem_logging) {
3093 		umem_transaction_log = umem_log_init(umem_transaction_log_size);
3094 		umem_content_log = umem_log_init(umem_content_log_size);
3095 		umem_failure_log = umem_log_init(umem_failure_log_size);
3096 		umem_slab_log = umem_log_init(umem_slab_log_size);
3097 	}
3098 
3099 	/*
3100 	 * Set up caches -- if successful, initialization cannot fail, since
3101 	 * allocations from other threads can now succeed.
3102 	 */
3103 	if (umem_cache_init() == 0) {
3104 		log_message("unable to create initial caches\n");
3105 		goto fail;
3106 	}
3107 	umem_oversize_arena = oversize_arena;
3108 	umem_memalign_arena = memalign_arena;
3109 
3110 	umem_cache_applyall(umem_cache_magazine_enable);
3111 
3112 	/*
3113 	 * initialization done, ready to go
3114 	 */
3115 	(void) mutex_lock(&umem_init_lock);
3116 	umem_ready = UMEM_READY;
3117 	umem_init_thr = 0;
3118 	(void) cond_broadcast(&umem_init_cv);
3119 	(void) mutex_unlock(&umem_init_lock);
3120 	return (1);
3121 
3122 fail:
3123 	log_message("umem initialization failed\n");
3124 
3125 	(void) mutex_lock(&umem_init_lock);
3126 	umem_ready = UMEM_READY_INIT_FAILED;
3127 	umem_init_thr = 0;
3128 	(void) cond_broadcast(&umem_init_cv);
3129 	(void) mutex_unlock(&umem_init_lock);
3130 	return (0);
3131 }
3132