xref: /illumos-gate/usr/src/lib/libumem/common/umem.c (revision 7257d1b4)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 /*
30  * based on usr/src/uts/common/os/kmem.c r1.64 from 2001/12/18
31  *
32  * The slab allocator, as described in the following two papers:
33  *
34  *	Jeff Bonwick,
35  *	The Slab Allocator: An Object-Caching Kernel Memory Allocator.
36  *	Proceedings of the Summer 1994 Usenix Conference.
37  *	Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
38  *
39  *	Jeff Bonwick and Jonathan Adams,
40  *	Magazines and vmem: Extending the Slab Allocator to Many CPUs and
41  *	Arbitrary Resources.
42  *	Proceedings of the 2001 Usenix Conference.
43  *	Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
44  *
45  * 1. Overview
46  * -----------
47  * umem is very close to kmem in implementation.  There are four major
48  * areas of divergence:
49  *
50  *	* Initialization
51  *
52  *	* CPU handling
53  *
54  *	* umem_update()
55  *
56  *	* KM_SLEEP v.s. UMEM_NOFAIL
57  *
58  *	* lock ordering
59  *
60  * 2. Initialization
61  * -----------------
62  * kmem is initialized early on in boot, and knows that no one will call
63  * into it before it is ready.  umem does not have these luxuries. Instead,
64  * initialization is divided into two phases:
65  *
66  *	* library initialization, and
67  *
68  *	* first use
69  *
70  * umem's full initialization happens at the time of the first allocation
71  * request (via malloc() and friends, umem_alloc(), or umem_zalloc()),
72  * or the first call to umem_cache_create().
73  *
74  * umem_free(), and umem_cache_alloc() do not require special handling,
75  * since the only way to get valid arguments for them is to successfully
76  * call a function from the first group.
77  *
78  * 2.1. Library Initialization: umem_startup()
79  * -------------------------------------------
80  * umem_startup() is libumem.so's .init section.  It calls pthread_atfork()
81  * to install the handlers necessary for umem's Fork1-Safety.  Because of
82  * race condition issues, all other pre-umem_init() initialization is done
83  * statically (i.e. by the dynamic linker).
84  *
85  * For standalone use, umem_startup() returns everything to its initial
86  * state.
87  *
88  * 2.2. First use: umem_init()
89  * ------------------------------
90  * The first time any memory allocation function is used, we have to
91  * create the backing caches and vmem arenas which are needed for it.
92  * umem_init() is the central point for that task.  When it completes,
93  * umem_ready is either UMEM_READY (all set) or UMEM_READY_INIT_FAILED (unable
94  * to initialize, probably due to lack of memory).
95  *
96  * There are four different paths from which umem_init() is called:
97  *
98  *	* from umem_alloc() or umem_zalloc(), with 0 < size < UMEM_MAXBUF,
99  *
100  *	* from umem_alloc() or umem_zalloc(), with size > UMEM_MAXBUF,
101  *
102  *	* from umem_cache_create(), and
103  *
104  *	* from memalign(), with align > UMEM_ALIGN.
105  *
106  * The last three just check if umem is initialized, and call umem_init()
107  * if it is not.  For performance reasons, the first case is more complicated.
108  *
109  * 2.2.1. umem_alloc()/umem_zalloc(), with 0 < size < UMEM_MAXBUF
110  * -----------------------------------------------------------------
111  * In this case, umem_cache_alloc(&umem_null_cache, ...) is called.
112  * There is special case code in which causes any allocation on
113  * &umem_null_cache to fail by returning (NULL), regardless of the
114  * flags argument.
115  *
116  * So umem_cache_alloc() returns NULL, and umem_alloc()/umem_zalloc() call
117  * umem_alloc_retry().  umem_alloc_retry() sees that the allocation
118  * was agains &umem_null_cache, and calls umem_init().
119  *
120  * If initialization is successful, umem_alloc_retry() returns 1, which
121  * causes umem_alloc()/umem_zalloc() to start over, which causes it to load
122  * the (now valid) cache pointer from umem_alloc_table.
123  *
124  * 2.2.2. Dealing with race conditions
125  * -----------------------------------
126  * There are a couple race conditions resulting from the initialization
127  * code that we have to guard against:
128  *
129  *	* In umem_cache_create(), there is a special UMC_INTERNAL cflag
130  *	that is passed for caches created during initialization.  It
131  *	is illegal for a user to try to create a UMC_INTERNAL cache.
132  *	This allows initialization to proceed, but any other
133  *	umem_cache_create()s will block by calling umem_init().
134  *
135  *	* Since umem_null_cache has a 1-element cache_cpu, it's cache_cpu_mask
136  *	is always zero.  umem_cache_alloc uses cp->cache_cpu_mask to
137  *	mask the cpu number.  This prevents a race between grabbing a
138  *	cache pointer out of umem_alloc_table and growing the cpu array.
139  *
140  *
141  * 3. CPU handling
142  * ---------------
143  * kmem uses the CPU's sequence number to determine which "cpu cache" to
144  * use for an allocation.  Currently, there is no way to get the sequence
145  * number in userspace.
146  *
147  * umem keeps track of cpu information in umem_cpus, an array of umem_max_ncpus
148  * umem_cpu_t structures.  CURCPU() is a a "hint" function, which we then mask
149  * with either umem_cpu_mask or cp->cache_cpu_mask to find the actual "cpu" id.
150  * The mechanics of this is all in the CPU(mask) macro.
151  *
152  * Currently, umem uses _lwp_self() as its hint.
153  *
154  *
155  * 4. The update thread
156  * --------------------
157  * kmem uses a task queue, kmem_taskq, to do periodic maintenance on
158  * every kmem cache.  vmem has a periodic timeout for hash table resizing.
159  * The kmem_taskq also provides a separate context for kmem_cache_reap()'s
160  * to be done in, avoiding issues of the context of kmem_reap() callers.
161  *
162  * Instead, umem has the concept of "updates", which are asynchronous requests
163  * for work attached to single caches.  All caches with pending work are
164  * on a doubly linked list rooted at the umem_null_cache.  All update state
165  * is protected by the umem_update_lock mutex, and the umem_update_cv is used
166  * for notification between threads.
167  *
168  * 4.1. Cache states with regards to updates
169  * -----------------------------------------
170  * A given cache is in one of three states:
171  *
172  * Inactive		cache_uflags is zero, cache_u{next,prev} are NULL
173  *
174  * Work Requested	cache_uflags is non-zero (but UMU_ACTIVE is not set),
175  *			cache_u{next,prev} link the cache onto the global
176  *			update list
177  *
178  * Active		cache_uflags has UMU_ACTIVE set, cache_u{next,prev}
179  *			are NULL, and either umem_update_thr or
180  *			umem_st_update_thr are actively doing work on the
181  *			cache.
182  *
183  * An update can be added to any cache in any state -- if the cache is
184  * Inactive, it transitions to being Work Requested.  If the cache is
185  * Active, the worker will notice the new update and act on it before
186  * transitioning the cache to the Inactive state.
187  *
188  * If a cache is in the Active state, UMU_NOTIFY can be set, which asks
189  * the worker to broadcast the umem_update_cv when it has finished.
190  *
191  * 4.2. Update interface
192  * ---------------------
193  * umem_add_update() adds an update to a particular cache.
194  * umem_updateall() adds an update to all caches.
195  * umem_remove_updates() returns a cache to the Inactive state.
196  *
197  * umem_process_updates() process all caches in the Work Requested state.
198  *
199  * 4.3. Reaping
200  * ------------
201  * When umem_reap() is called (at the time of heap growth), it schedule
202  * UMU_REAP updates on every cache.  It then checks to see if the update
203  * thread exists (umem_update_thr != 0).  If it is, it broadcasts
204  * the umem_update_cv to wake the update thread up, and returns.
205  *
206  * If the update thread does not exist (umem_update_thr == 0), and the
207  * program currently has multiple threads, umem_reap() attempts to create
208  * a new update thread.
209  *
210  * If the process is not multithreaded, or the creation fails, umem_reap()
211  * calls umem_st_update() to do an inline update.
212  *
213  * 4.4. The update thread
214  * ----------------------
215  * The update thread spends most of its time in cond_timedwait() on the
216  * umem_update_cv.  It wakes up under two conditions:
217  *
218  *	* The timedwait times out, in which case it needs to run a global
219  *	update, or
220  *
221  *	* someone cond_broadcast(3THR)s the umem_update_cv, in which case
222  *	it needs to check if there are any caches in the Work Requested
223  *	state.
224  *
225  * When it is time for another global update, umem calls umem_cache_update()
226  * on every cache, then calls vmem_update(), which tunes the vmem structures.
227  * umem_cache_update() can request further work using umem_add_update().
228  *
229  * After any work from the global update completes, the update timer is
230  * reset to umem_reap_interval seconds in the future.  This makes the
231  * updates self-throttling.
232  *
233  * Reaps are similarly self-throttling.  After a UMU_REAP update has
234  * been scheduled on all caches, umem_reap() sets a flag and wakes up the
235  * update thread.  The update thread notices the flag, and resets the
236  * reap state.
237  *
238  * 4.5. Inline updates
239  * -------------------
240  * If the update thread is not running, umem_st_update() is used instead.  It
241  * immediately does a global update (as above), then calls
242  * umem_process_updates() to process both the reaps that umem_reap() added and
243  * any work generated by the global update.  Afterwards, it resets the reap
244  * state.
245  *
246  * While the umem_st_update() is running, umem_st_update_thr holds the thread
247  * id of the thread performing the update.
248  *
249  * 4.6. Updates and fork1()
250  * ------------------------
251  * umem has fork1() pre- and post-handlers which lock up (and release) every
252  * mutex in every cache.  They also lock up the umem_update_lock.  Since
253  * fork1() only copies over a single lwp, other threads (including the update
254  * thread) could have been actively using a cache in the parent.  This
255  * can lead to inconsistencies in the child process.
256  *
257  * Because we locked all of the mutexes, the only possible inconsistancies are:
258  *
259  *	* a umem_cache_alloc() could leak its buffer.
260  *
261  *	* a caller of umem_depot_alloc() could leak a magazine, and all the
262  *	buffers contained in it.
263  *
264  *	* a cache could be in the Active update state.  In the child, there
265  *	would be no thread actually working on it.
266  *
267  *	* a umem_hash_rescale() could leak the new hash table.
268  *
269  *	* a umem_magazine_resize() could be in progress.
270  *
271  *	* a umem_reap() could be in progress.
272  *
273  * The memory leaks we can't do anything about.  umem_release_child() resets
274  * the update state, moves any caches in the Active state to the Work Requested
275  * state.  This might cause some updates to be re-run, but UMU_REAP and
276  * UMU_HASH_RESCALE are effectively idempotent, and the worst that can
277  * happen from umem_magazine_resize() is resizing the magazine twice in close
278  * succession.
279  *
280  * Much of the cleanup in umem_release_child() is skipped if
281  * umem_st_update_thr == thr_self().  This is so that applications which call
282  * fork1() from a cache callback does not break.  Needless to say, any such
283  * application is tremendously broken.
284  *
285  *
286  * 5. KM_SLEEP v.s. UMEM_NOFAIL
287  * ----------------------------
288  * Allocations against kmem and vmem have two basic modes:  SLEEP and
289  * NOSLEEP.  A sleeping allocation is will go to sleep (waiting for
290  * more memory) instead of failing (returning NULL).
291  *
292  * SLEEP allocations presume an extremely multithreaded model, with
293  * a lot of allocation and deallocation activity.  umem cannot presume
294  * that its clients have any particular type of behavior.  Instead,
295  * it provides two types of allocations:
296  *
297  *	* UMEM_DEFAULT, equivalent to KM_NOSLEEP (i.e. return NULL on
298  *	failure)
299  *
300  *	* UMEM_NOFAIL, which, on failure, calls an optional callback
301  *	(registered with umem_nofail_callback()).
302  *
303  * The callback is invoked with no locks held, and can do an arbitrary
304  * amount of work.  It then has a choice between:
305  *
306  *	* Returning UMEM_CALLBACK_RETRY, which will cause the allocation
307  *	to be restarted.
308  *
309  *	* Returning UMEM_CALLBACK_EXIT(status), which will cause exit(2)
310  *	to be invoked with status.  If multiple threads attempt to do
311  *	this simultaneously, only one will call exit(2).
312  *
313  *	* Doing some kind of non-local exit (thr_exit(3thr), longjmp(3C),
314  *	etc.)
315  *
316  * The default callback returns UMEM_CALLBACK_EXIT(255).
317  *
318  * To have these callbacks without risk of state corruption (in the case of
319  * a non-local exit), we have to ensure that the callbacks get invoked
320  * close to the original allocation, with no inconsistent state or held
321  * locks.  The following steps are taken:
322  *
323  *	* All invocations of vmem are VM_NOSLEEP.
324  *
325  *	* All constructor callbacks (which can themselves to allocations)
326  *	are passed UMEM_DEFAULT as their required allocation argument.  This
327  *	way, the constructor will fail, allowing the highest-level allocation
328  *	invoke the nofail callback.
329  *
330  *	If a constructor callback _does_ do a UMEM_NOFAIL allocation, and
331  *	the nofail callback does a non-local exit, we will leak the
332  *	partially-constructed buffer.
333  *
334  *
335  * 6. Lock Ordering
336  * ----------------
337  * umem has a few more locks than kmem does, mostly in the update path.  The
338  * overall lock ordering (earlier locks must be acquired first) is:
339  *
340  *	umem_init_lock
341  *
342  *	vmem_list_lock
343  *	vmem_nosleep_lock.vmpl_mutex
344  *	vmem_t's:
345  *		vm_lock
346  *	sbrk_lock
347  *
348  *	umem_cache_lock
349  *	umem_update_lock
350  *	umem_flags_lock
351  *	umem_cache_t's:
352  *		cache_cpu[*].cc_lock
353  *		cache_depot_lock
354  *		cache_lock
355  *	umem_log_header_t's:
356  *		lh_cpu[*].clh_lock
357  *		lh_lock
358  */
359 
360 #include <umem_impl.h>
361 #include <sys/vmem_impl_user.h>
362 #include "umem_base.h"
363 #include "vmem_base.h"
364 
365 #include <sys/processor.h>
366 #include <sys/sysmacros.h>
367 
368 #include <alloca.h>
369 #include <errno.h>
370 #include <limits.h>
371 #include <stdio.h>
372 #include <stdlib.h>
373 #include <string.h>
374 #include <strings.h>
375 #include <signal.h>
376 #include <unistd.h>
377 #include <atomic.h>
378 
379 #include "misc.h"
380 
381 #define	UMEM_VMFLAGS(umflag)	(VM_NOSLEEP)
382 
383 size_t pagesize;
384 
385 /*
386  * The default set of caches to back umem_alloc().
387  * These sizes should be reevaluated periodically.
388  *
389  * We want allocations that are multiples of the coherency granularity
390  * (64 bytes) to be satisfied from a cache which is a multiple of 64
391  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
392  * the next kmem_cache_size greater than or equal to it must be a
393  * multiple of 64.
394  *
395  * This table must be in sorted order, from smallest to highest.  The
396  * highest slot must be UMEM_MAXBUF, and every slot afterwards must be
397  * zero.
398  */
399 static int umem_alloc_sizes[] = {
400 #ifdef _LP64
401 	1 * 8,
402 	1 * 16,
403 	2 * 16,
404 	3 * 16,
405 #else
406 	1 * 8,
407 	2 * 8,
408 	3 * 8,
409 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
410 #endif
411 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
412 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
413 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
414 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
415 	P2ALIGN(8192 / 7, 64),
416 	P2ALIGN(8192 / 6, 64),
417 	P2ALIGN(8192 / 5, 64),
418 	P2ALIGN(8192 / 4, 64), 2304,
419 	P2ALIGN(8192 / 3, 64),
420 	P2ALIGN(8192 / 2, 64), 4544,
421 	P2ALIGN(8192 / 1, 64), 9216,
422 	4096 * 3,
423 	UMEM_MAXBUF,				/* = 8192 * 2 */
424 	/* 24 slots for user expansion */
425 	0, 0, 0, 0, 0, 0, 0, 0,
426 	0, 0, 0, 0, 0, 0, 0, 0,
427 	0, 0, 0, 0, 0, 0, 0, 0,
428 };
429 #define	NUM_ALLOC_SIZES (sizeof (umem_alloc_sizes) / sizeof (*umem_alloc_sizes))
430 
431 static umem_magtype_t umem_magtype[] = {
432 	{ 1,	8,	3200,	65536	},
433 	{ 3,	16,	256,	32768	},
434 	{ 7,	32,	64,	16384	},
435 	{ 15,	64,	0,	8192	},
436 	{ 31,	64,	0,	4096	},
437 	{ 47,	64,	0,	2048	},
438 	{ 63,	64,	0,	1024	},
439 	{ 95,	64,	0,	512	},
440 	{ 143,	64,	0,	0	},
441 };
442 
443 /*
444  * umem tunables
445  */
446 uint32_t umem_max_ncpus;	/* # of CPU caches. */
447 
448 uint32_t umem_stack_depth = 15; /* # stack frames in a bufctl_audit */
449 uint32_t umem_reap_interval = 10; /* max reaping rate (seconds) */
450 uint_t umem_depot_contention = 2; /* max failed trylocks per real interval */
451 uint_t umem_abort = 1;		/* whether to abort on error */
452 uint_t umem_output = 0;		/* whether to write to standard error */
453 uint_t umem_logging = 0;	/* umem_log_enter() override */
454 uint32_t umem_mtbf = 0;		/* mean time between failures [default: off] */
455 size_t umem_transaction_log_size; /* size of transaction log */
456 size_t umem_content_log_size;	/* size of content log */
457 size_t umem_failure_log_size;	/* failure log [4 pages per CPU] */
458 size_t umem_slab_log_size;	/* slab create log [4 pages per CPU] */
459 size_t umem_content_maxsave = 256; /* UMF_CONTENTS max bytes to log */
460 size_t umem_lite_minsize = 0;	/* minimum buffer size for UMF_LITE */
461 size_t umem_lite_maxalign = 1024; /* maximum buffer alignment for UMF_LITE */
462 size_t umem_maxverify;		/* maximum bytes to inspect in debug routines */
463 size_t umem_minfirewall;	/* hardware-enforced redzone threshold */
464 
465 uint_t umem_flags = 0;
466 
467 mutex_t			umem_init_lock;		/* locks initialization */
468 cond_t			umem_init_cv;		/* initialization CV */
469 thread_t		umem_init_thr;		/* thread initializing */
470 int			umem_init_env_ready;	/* environ pre-initted */
471 int			umem_ready = UMEM_READY_STARTUP;
472 
473 static umem_nofail_callback_t *nofail_callback;
474 static mutex_t		umem_nofail_exit_lock;
475 static thread_t		umem_nofail_exit_thr;
476 
477 static umem_cache_t	*umem_slab_cache;
478 static umem_cache_t	*umem_bufctl_cache;
479 static umem_cache_t	*umem_bufctl_audit_cache;
480 
481 mutex_t			umem_flags_lock;
482 
483 static vmem_t		*heap_arena;
484 static vmem_alloc_t	*heap_alloc;
485 static vmem_free_t	*heap_free;
486 
487 static vmem_t		*umem_internal_arena;
488 static vmem_t		*umem_cache_arena;
489 static vmem_t		*umem_hash_arena;
490 static vmem_t		*umem_log_arena;
491 static vmem_t		*umem_oversize_arena;
492 static vmem_t		*umem_va_arena;
493 static vmem_t		*umem_default_arena;
494 static vmem_t		*umem_firewall_va_arena;
495 static vmem_t		*umem_firewall_arena;
496 
497 vmem_t			*umem_memalign_arena;
498 
499 umem_log_header_t *umem_transaction_log;
500 umem_log_header_t *umem_content_log;
501 umem_log_header_t *umem_failure_log;
502 umem_log_header_t *umem_slab_log;
503 
504 #define	CPUHINT()		(thr_self())
505 #define	CPUHINT_MAX()		INT_MAX
506 
507 #define	CPU(mask)		(umem_cpus + (CPUHINT() & (mask)))
508 static umem_cpu_t umem_startup_cpu = {	/* initial, single, cpu */
509 	UMEM_CACHE_SIZE(0),
510 	0
511 };
512 
513 static uint32_t umem_cpu_mask = 0;			/* global cpu mask */
514 static umem_cpu_t *umem_cpus = &umem_startup_cpu;	/* cpu list */
515 
516 volatile uint32_t umem_reaping;
517 
518 thread_t		umem_update_thr;
519 struct timeval		umem_update_next;	/* timeofday of next update */
520 volatile thread_t	umem_st_update_thr;	/* only used when single-thd */
521 
522 #define	IN_UPDATE()	(thr_self() == umem_update_thr || \
523 			    thr_self() == umem_st_update_thr)
524 #define	IN_REAP()	IN_UPDATE()
525 
526 mutex_t			umem_update_lock;	/* cache_u{next,prev,flags} */
527 cond_t			umem_update_cv;
528 
529 volatile hrtime_t umem_reap_next;	/* min hrtime of next reap */
530 
531 mutex_t			umem_cache_lock;	/* inter-cache linkage only */
532 
533 #ifdef UMEM_STANDALONE
534 umem_cache_t		umem_null_cache;
535 static const umem_cache_t umem_null_cache_template = {
536 #else
537 umem_cache_t		umem_null_cache = {
538 #endif
539 	0, 0, 0, 0, 0,
540 	0, 0,
541 	0, 0,
542 	0, 0,
543 	"invalid_cache",
544 	0, 0,
545 	NULL, NULL, NULL, NULL,
546 	NULL,
547 	0, 0, 0, 0,
548 	&umem_null_cache, &umem_null_cache,
549 	&umem_null_cache, &umem_null_cache,
550 	0,
551 	DEFAULTMUTEX,				/* start of slab layer */
552 	0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
553 	&umem_null_cache.cache_nullslab,
554 	{
555 		&umem_null_cache,
556 		NULL,
557 		&umem_null_cache.cache_nullslab,
558 		&umem_null_cache.cache_nullslab,
559 		NULL,
560 		-1,
561 		0
562 	},
563 	NULL,
564 	NULL,
565 	DEFAULTMUTEX,				/* start of depot layer */
566 	NULL, {
567 		NULL, 0, 0, 0, 0
568 	}, {
569 		NULL, 0, 0, 0, 0
570 	}, {
571 		{
572 			DEFAULTMUTEX,		/* start of CPU cache */
573 			0, 0, NULL, NULL, -1, -1, 0
574 		}
575 	}
576 };
577 
578 #define	ALLOC_TABLE_4 \
579 	&umem_null_cache, &umem_null_cache, &umem_null_cache, &umem_null_cache
580 
581 #define	ALLOC_TABLE_64 \
582 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
583 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
584 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
585 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4
586 
587 #define	ALLOC_TABLE_1024 \
588 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
589 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
590 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
591 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64
592 
593 static umem_cache_t *umem_alloc_table[UMEM_MAXBUF >> UMEM_ALIGN_SHIFT] = {
594 	ALLOC_TABLE_1024,
595 	ALLOC_TABLE_1024
596 };
597 
598 
599 /* Used to constrain audit-log stack traces */
600 caddr_t			umem_min_stack;
601 caddr_t			umem_max_stack;
602 
603 
604 #define	UMERR_MODIFIED	0	/* buffer modified while on freelist */
605 #define	UMERR_REDZONE	1	/* redzone violation (write past end of buf) */
606 #define	UMERR_DUPFREE	2	/* freed a buffer twice */
607 #define	UMERR_BADADDR	3	/* freed a bad (unallocated) address */
608 #define	UMERR_BADBUFTAG	4	/* buftag corrupted */
609 #define	UMERR_BADBUFCTL	5	/* bufctl corrupted */
610 #define	UMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
611 #define	UMERR_BADSIZE	7	/* alloc size != free size */
612 #define	UMERR_BADBASE	8	/* buffer base address wrong */
613 
614 struct {
615 	hrtime_t	ump_timestamp;	/* timestamp of error */
616 	int		ump_error;	/* type of umem error (UMERR_*) */
617 	void		*ump_buffer;	/* buffer that induced abort */
618 	void		*ump_realbuf;	/* real start address for buffer */
619 	umem_cache_t	*ump_cache;	/* buffer's cache according to client */
620 	umem_cache_t	*ump_realcache;	/* actual cache containing buffer */
621 	umem_slab_t	*ump_slab;	/* slab accoring to umem_findslab() */
622 	umem_bufctl_t	*ump_bufctl;	/* bufctl */
623 } umem_abort_info;
624 
625 static void
626 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
627 {
628 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
629 	uint64_t *buf = buf_arg;
630 
631 	while (buf < bufend)
632 		*buf++ = pattern;
633 }
634 
635 static void *
636 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
637 {
638 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
639 	uint64_t *buf;
640 
641 	for (buf = buf_arg; buf < bufend; buf++)
642 		if (*buf != pattern)
643 			return (buf);
644 	return (NULL);
645 }
646 
647 static void *
648 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
649 {
650 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
651 	uint64_t *buf;
652 
653 	for (buf = buf_arg; buf < bufend; buf++) {
654 		if (*buf != old) {
655 			copy_pattern(old, buf_arg,
656 			    (char *)buf - (char *)buf_arg);
657 			return (buf);
658 		}
659 		*buf = new;
660 	}
661 
662 	return (NULL);
663 }
664 
665 void
666 umem_cache_applyall(void (*func)(umem_cache_t *))
667 {
668 	umem_cache_t *cp;
669 
670 	(void) mutex_lock(&umem_cache_lock);
671 	for (cp = umem_null_cache.cache_next; cp != &umem_null_cache;
672 	    cp = cp->cache_next)
673 		func(cp);
674 	(void) mutex_unlock(&umem_cache_lock);
675 }
676 
677 static void
678 umem_add_update_unlocked(umem_cache_t *cp, int flags)
679 {
680 	umem_cache_t *cnext, *cprev;
681 
682 	flags &= ~UMU_ACTIVE;
683 
684 	if (!flags)
685 		return;
686 
687 	if (cp->cache_uflags & UMU_ACTIVE) {
688 		cp->cache_uflags |= flags;
689 	} else {
690 		if (cp->cache_unext != NULL) {
691 			ASSERT(cp->cache_uflags != 0);
692 			cp->cache_uflags |= flags;
693 		} else {
694 			ASSERT(cp->cache_uflags == 0);
695 			cp->cache_uflags = flags;
696 			cp->cache_unext = cnext = &umem_null_cache;
697 			cp->cache_uprev = cprev = umem_null_cache.cache_uprev;
698 			cnext->cache_uprev = cp;
699 			cprev->cache_unext = cp;
700 		}
701 	}
702 }
703 
704 static void
705 umem_add_update(umem_cache_t *cp, int flags)
706 {
707 	(void) mutex_lock(&umem_update_lock);
708 
709 	umem_add_update_unlocked(cp, flags);
710 
711 	if (!IN_UPDATE())
712 		(void) cond_broadcast(&umem_update_cv);
713 
714 	(void) mutex_unlock(&umem_update_lock);
715 }
716 
717 /*
718  * Remove a cache from the update list, waiting for any in-progress work to
719  * complete first.
720  */
721 static void
722 umem_remove_updates(umem_cache_t *cp)
723 {
724 	(void) mutex_lock(&umem_update_lock);
725 
726 	/*
727 	 * Get it out of the active state
728 	 */
729 	while (cp->cache_uflags & UMU_ACTIVE) {
730 		int cancel_state;
731 
732 		ASSERT(cp->cache_unext == NULL);
733 
734 		cp->cache_uflags |= UMU_NOTIFY;
735 
736 		/*
737 		 * Make sure the update state is sane, before we wait
738 		 */
739 		ASSERT(umem_update_thr != 0 || umem_st_update_thr != 0);
740 		ASSERT(umem_update_thr != thr_self() &&
741 		    umem_st_update_thr != thr_self());
742 
743 		(void) pthread_setcancelstate(PTHREAD_CANCEL_DISABLE,
744 		    &cancel_state);
745 		(void) cond_wait(&umem_update_cv, &umem_update_lock);
746 		(void) pthread_setcancelstate(cancel_state, NULL);
747 	}
748 	/*
749 	 * Get it out of the Work Requested state
750 	 */
751 	if (cp->cache_unext != NULL) {
752 		cp->cache_uprev->cache_unext = cp->cache_unext;
753 		cp->cache_unext->cache_uprev = cp->cache_uprev;
754 		cp->cache_uprev = cp->cache_unext = NULL;
755 		cp->cache_uflags = 0;
756 	}
757 	/*
758 	 * Make sure it is in the Inactive state
759 	 */
760 	ASSERT(cp->cache_unext == NULL && cp->cache_uflags == 0);
761 	(void) mutex_unlock(&umem_update_lock);
762 }
763 
764 static void
765 umem_updateall(int flags)
766 {
767 	umem_cache_t *cp;
768 
769 	/*
770 	 * NOTE:  To prevent deadlock, umem_cache_lock is always acquired first.
771 	 *
772 	 * (umem_add_update is called from things run via umem_cache_applyall)
773 	 */
774 	(void) mutex_lock(&umem_cache_lock);
775 	(void) mutex_lock(&umem_update_lock);
776 
777 	for (cp = umem_null_cache.cache_next; cp != &umem_null_cache;
778 	    cp = cp->cache_next)
779 		umem_add_update_unlocked(cp, flags);
780 
781 	if (!IN_UPDATE())
782 		(void) cond_broadcast(&umem_update_cv);
783 
784 	(void) mutex_unlock(&umem_update_lock);
785 	(void) mutex_unlock(&umem_cache_lock);
786 }
787 
788 /*
789  * Debugging support.  Given a buffer address, find its slab.
790  */
791 static umem_slab_t *
792 umem_findslab(umem_cache_t *cp, void *buf)
793 {
794 	umem_slab_t *sp;
795 
796 	(void) mutex_lock(&cp->cache_lock);
797 	for (sp = cp->cache_nullslab.slab_next;
798 	    sp != &cp->cache_nullslab; sp = sp->slab_next) {
799 		if (UMEM_SLAB_MEMBER(sp, buf)) {
800 			(void) mutex_unlock(&cp->cache_lock);
801 			return (sp);
802 		}
803 	}
804 	(void) mutex_unlock(&cp->cache_lock);
805 
806 	return (NULL);
807 }
808 
809 static void
810 umem_error(int error, umem_cache_t *cparg, void *bufarg)
811 {
812 	umem_buftag_t *btp = NULL;
813 	umem_bufctl_t *bcp = NULL;
814 	umem_cache_t *cp = cparg;
815 	umem_slab_t *sp;
816 	uint64_t *off;
817 	void *buf = bufarg;
818 
819 	int old_logging = umem_logging;
820 
821 	umem_logging = 0;	/* stop logging when a bad thing happens */
822 
823 	umem_abort_info.ump_timestamp = gethrtime();
824 
825 	sp = umem_findslab(cp, buf);
826 	if (sp == NULL) {
827 		for (cp = umem_null_cache.cache_prev; cp != &umem_null_cache;
828 		    cp = cp->cache_prev) {
829 			if ((sp = umem_findslab(cp, buf)) != NULL)
830 				break;
831 		}
832 	}
833 
834 	if (sp == NULL) {
835 		cp = NULL;
836 		error = UMERR_BADADDR;
837 	} else {
838 		if (cp != cparg)
839 			error = UMERR_BADCACHE;
840 		else
841 			buf = (char *)bufarg - ((uintptr_t)bufarg -
842 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
843 		if (buf != bufarg)
844 			error = UMERR_BADBASE;
845 		if (cp->cache_flags & UMF_BUFTAG)
846 			btp = UMEM_BUFTAG(cp, buf);
847 		if (cp->cache_flags & UMF_HASH) {
848 			(void) mutex_lock(&cp->cache_lock);
849 			for (bcp = *UMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
850 				if (bcp->bc_addr == buf)
851 					break;
852 			(void) mutex_unlock(&cp->cache_lock);
853 			if (bcp == NULL && btp != NULL)
854 				bcp = btp->bt_bufctl;
855 			if (umem_findslab(cp->cache_bufctl_cache, bcp) ==
856 			    NULL || P2PHASE((uintptr_t)bcp, UMEM_ALIGN) ||
857 			    bcp->bc_addr != buf) {
858 				error = UMERR_BADBUFCTL;
859 				bcp = NULL;
860 			}
861 		}
862 	}
863 
864 	umem_abort_info.ump_error = error;
865 	umem_abort_info.ump_buffer = bufarg;
866 	umem_abort_info.ump_realbuf = buf;
867 	umem_abort_info.ump_cache = cparg;
868 	umem_abort_info.ump_realcache = cp;
869 	umem_abort_info.ump_slab = sp;
870 	umem_abort_info.ump_bufctl = bcp;
871 
872 	umem_printf("umem allocator: ");
873 
874 	switch (error) {
875 
876 	case UMERR_MODIFIED:
877 		umem_printf("buffer modified after being freed\n");
878 		off = verify_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
879 		if (off == NULL)	/* shouldn't happen */
880 			off = buf;
881 		umem_printf("modification occurred at offset 0x%lx "
882 		    "(0x%llx replaced by 0x%llx)\n",
883 		    (uintptr_t)off - (uintptr_t)buf,
884 		    (longlong_t)UMEM_FREE_PATTERN, (longlong_t)*off);
885 		break;
886 
887 	case UMERR_REDZONE:
888 		umem_printf("redzone violation: write past end of buffer\n");
889 		break;
890 
891 	case UMERR_BADADDR:
892 		umem_printf("invalid free: buffer not in cache\n");
893 		break;
894 
895 	case UMERR_DUPFREE:
896 		umem_printf("duplicate free: buffer freed twice\n");
897 		break;
898 
899 	case UMERR_BADBUFTAG:
900 		umem_printf("boundary tag corrupted\n");
901 		umem_printf("bcp ^ bxstat = %lx, should be %lx\n",
902 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
903 		    UMEM_BUFTAG_FREE);
904 		break;
905 
906 	case UMERR_BADBUFCTL:
907 		umem_printf("bufctl corrupted\n");
908 		break;
909 
910 	case UMERR_BADCACHE:
911 		umem_printf("buffer freed to wrong cache\n");
912 		umem_printf("buffer was allocated from %s,\n", cp->cache_name);
913 		umem_printf("caller attempting free to %s.\n",
914 		    cparg->cache_name);
915 		break;
916 
917 	case UMERR_BADSIZE:
918 		umem_printf("bad free: free size (%u) != alloc size (%u)\n",
919 		    UMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
920 		    UMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
921 		break;
922 
923 	case UMERR_BADBASE:
924 		umem_printf("bad free: free address (%p) != alloc address "
925 		    "(%p)\n", bufarg, buf);
926 		break;
927 	}
928 
929 	umem_printf("buffer=%p  bufctl=%p  cache: %s\n",
930 	    bufarg, (void *)bcp, cparg->cache_name);
931 
932 	if (bcp != NULL && (cp->cache_flags & UMF_AUDIT) &&
933 	    error != UMERR_BADBUFCTL) {
934 		int d;
935 		timespec_t ts;
936 		hrtime_t diff;
937 		umem_bufctl_audit_t *bcap = (umem_bufctl_audit_t *)bcp;
938 
939 		diff = umem_abort_info.ump_timestamp - bcap->bc_timestamp;
940 		ts.tv_sec = diff / NANOSEC;
941 		ts.tv_nsec = diff % NANOSEC;
942 
943 		umem_printf("previous transaction on buffer %p:\n", buf);
944 		umem_printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
945 		    (void *)(intptr_t)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
946 		    (void *)sp, cp->cache_name);
947 		for (d = 0; d < MIN(bcap->bc_depth, umem_stack_depth); d++) {
948 			(void) print_sym((void *)bcap->bc_stack[d]);
949 			umem_printf("\n");
950 		}
951 	}
952 
953 	umem_err_recoverable("umem: heap corruption detected");
954 
955 	umem_logging = old_logging;	/* resume logging */
956 }
957 
958 void
959 umem_nofail_callback(umem_nofail_callback_t *cb)
960 {
961 	nofail_callback = cb;
962 }
963 
964 static int
965 umem_alloc_retry(umem_cache_t *cp, int umflag)
966 {
967 	if (cp == &umem_null_cache) {
968 		if (umem_init())
969 			return (1);				/* retry */
970 		/*
971 		 * Initialization failed.  Do normal failure processing.
972 		 */
973 	}
974 	if (umflag & UMEM_NOFAIL) {
975 		int def_result = UMEM_CALLBACK_EXIT(255);
976 		int result = def_result;
977 		umem_nofail_callback_t *callback = nofail_callback;
978 
979 		if (callback != NULL)
980 			result = callback();
981 
982 		if (result == UMEM_CALLBACK_RETRY)
983 			return (1);
984 
985 		if ((result & ~0xFF) != UMEM_CALLBACK_EXIT(0)) {
986 			log_message("nofail callback returned %x\n", result);
987 			result = def_result;
988 		}
989 
990 		/*
991 		 * only one thread will call exit
992 		 */
993 		if (umem_nofail_exit_thr == thr_self())
994 			umem_panic("recursive UMEM_CALLBACK_EXIT()\n");
995 
996 		(void) mutex_lock(&umem_nofail_exit_lock);
997 		umem_nofail_exit_thr = thr_self();
998 		exit(result & 0xFF);
999 		/*NOTREACHED*/
1000 	}
1001 	return (0);
1002 }
1003 
1004 static umem_log_header_t *
1005 umem_log_init(size_t logsize)
1006 {
1007 	umem_log_header_t *lhp;
1008 	int nchunks = 4 * umem_max_ncpus;
1009 	size_t lhsize = offsetof(umem_log_header_t, lh_cpu[umem_max_ncpus]);
1010 	int i;
1011 
1012 	if (logsize == 0)
1013 		return (NULL);
1014 
1015 	/*
1016 	 * Make sure that lhp->lh_cpu[] is nicely aligned
1017 	 * to prevent false sharing of cache lines.
1018 	 */
1019 	lhsize = P2ROUNDUP(lhsize, UMEM_ALIGN);
1020 	lhp = vmem_xalloc(umem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1021 	    NULL, NULL, VM_NOSLEEP);
1022 	if (lhp == NULL)
1023 		goto fail;
1024 
1025 	bzero(lhp, lhsize);
1026 
1027 	(void) mutex_init(&lhp->lh_lock, USYNC_THREAD, NULL);
1028 	lhp->lh_nchunks = nchunks;
1029 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks, PAGESIZE);
1030 	if (lhp->lh_chunksize == 0)
1031 		lhp->lh_chunksize = PAGESIZE;
1032 
1033 	lhp->lh_base = vmem_alloc(umem_log_arena,
1034 	    lhp->lh_chunksize * nchunks, VM_NOSLEEP);
1035 	if (lhp->lh_base == NULL)
1036 		goto fail;
1037 
1038 	lhp->lh_free = vmem_alloc(umem_log_arena,
1039 	    nchunks * sizeof (int), VM_NOSLEEP);
1040 	if (lhp->lh_free == NULL)
1041 		goto fail;
1042 
1043 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1044 
1045 	for (i = 0; i < umem_max_ncpus; i++) {
1046 		umem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1047 		(void) mutex_init(&clhp->clh_lock, USYNC_THREAD, NULL);
1048 		clhp->clh_chunk = i;
1049 	}
1050 
1051 	for (i = umem_max_ncpus; i < nchunks; i++)
1052 		lhp->lh_free[i] = i;
1053 
1054 	lhp->lh_head = umem_max_ncpus;
1055 	lhp->lh_tail = 0;
1056 
1057 	return (lhp);
1058 
1059 fail:
1060 	if (lhp != NULL) {
1061 		if (lhp->lh_base != NULL)
1062 			vmem_free(umem_log_arena, lhp->lh_base,
1063 			    lhp->lh_chunksize * nchunks);
1064 
1065 		vmem_xfree(umem_log_arena, lhp, lhsize);
1066 	}
1067 	return (NULL);
1068 }
1069 
1070 static void *
1071 umem_log_enter(umem_log_header_t *lhp, void *data, size_t size)
1072 {
1073 	void *logspace;
1074 	umem_cpu_log_header_t *clhp =
1075 	    &lhp->lh_cpu[CPU(umem_cpu_mask)->cpu_number];
1076 
1077 	if (lhp == NULL || umem_logging == 0)
1078 		return (NULL);
1079 
1080 	(void) mutex_lock(&clhp->clh_lock);
1081 	clhp->clh_hits++;
1082 	if (size > clhp->clh_avail) {
1083 		(void) mutex_lock(&lhp->lh_lock);
1084 		lhp->lh_hits++;
1085 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1086 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1087 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1088 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1089 		clhp->clh_current = lhp->lh_base +
1090 		    clhp->clh_chunk * lhp->lh_chunksize;
1091 		clhp->clh_avail = lhp->lh_chunksize;
1092 		if (size > lhp->lh_chunksize)
1093 			size = lhp->lh_chunksize;
1094 		(void) mutex_unlock(&lhp->lh_lock);
1095 	}
1096 	logspace = clhp->clh_current;
1097 	clhp->clh_current += size;
1098 	clhp->clh_avail -= size;
1099 	bcopy(data, logspace, size);
1100 	(void) mutex_unlock(&clhp->clh_lock);
1101 	return (logspace);
1102 }
1103 
1104 #define	UMEM_AUDIT(lp, cp, bcp)						\
1105 {									\
1106 	umem_bufctl_audit_t *_bcp = (umem_bufctl_audit_t *)(bcp);	\
1107 	_bcp->bc_timestamp = gethrtime();				\
1108 	_bcp->bc_thread = thr_self();					\
1109 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, umem_stack_depth,	\
1110 	    (cp != NULL) && (cp->cache_flags & UMF_CHECKSIGNAL));	\
1111 	_bcp->bc_lastlog = umem_log_enter((lp), _bcp,			\
1112 	    UMEM_BUFCTL_AUDIT_SIZE);					\
1113 }
1114 
1115 static void
1116 umem_log_event(umem_log_header_t *lp, umem_cache_t *cp,
1117 	umem_slab_t *sp, void *addr)
1118 {
1119 	umem_bufctl_audit_t *bcp;
1120 	UMEM_LOCAL_BUFCTL_AUDIT(&bcp);
1121 
1122 	bzero(bcp, UMEM_BUFCTL_AUDIT_SIZE);
1123 	bcp->bc_addr = addr;
1124 	bcp->bc_slab = sp;
1125 	bcp->bc_cache = cp;
1126 	UMEM_AUDIT(lp, cp, bcp);
1127 }
1128 
1129 /*
1130  * Create a new slab for cache cp.
1131  */
1132 static umem_slab_t *
1133 umem_slab_create(umem_cache_t *cp, int umflag)
1134 {
1135 	size_t slabsize = cp->cache_slabsize;
1136 	size_t chunksize = cp->cache_chunksize;
1137 	int cache_flags = cp->cache_flags;
1138 	size_t color, chunks;
1139 	char *buf, *slab;
1140 	umem_slab_t *sp;
1141 	umem_bufctl_t *bcp;
1142 	vmem_t *vmp = cp->cache_arena;
1143 
1144 	color = cp->cache_color + cp->cache_align;
1145 	if (color > cp->cache_maxcolor)
1146 		color = cp->cache_mincolor;
1147 	cp->cache_color = color;
1148 
1149 	slab = vmem_alloc(vmp, slabsize, UMEM_VMFLAGS(umflag));
1150 
1151 	if (slab == NULL)
1152 		goto vmem_alloc_failure;
1153 
1154 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1155 
1156 	if (!(cp->cache_cflags & UMC_NOTOUCH) &&
1157 	    (cp->cache_flags & UMF_DEADBEEF))
1158 		copy_pattern(UMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1159 
1160 	if (cache_flags & UMF_HASH) {
1161 		if ((sp = _umem_cache_alloc(umem_slab_cache, umflag)) == NULL)
1162 			goto slab_alloc_failure;
1163 		chunks = (slabsize - color) / chunksize;
1164 	} else {
1165 		sp = UMEM_SLAB(cp, slab);
1166 		chunks = (slabsize - sizeof (umem_slab_t) - color) / chunksize;
1167 	}
1168 
1169 	sp->slab_cache	= cp;
1170 	sp->slab_head	= NULL;
1171 	sp->slab_refcnt	= 0;
1172 	sp->slab_base	= buf = slab + color;
1173 	sp->slab_chunks	= chunks;
1174 
1175 	ASSERT(chunks > 0);
1176 	while (chunks-- != 0) {
1177 		if (cache_flags & UMF_HASH) {
1178 			bcp = _umem_cache_alloc(cp->cache_bufctl_cache, umflag);
1179 			if (bcp == NULL)
1180 				goto bufctl_alloc_failure;
1181 			if (cache_flags & UMF_AUDIT) {
1182 				umem_bufctl_audit_t *bcap =
1183 				    (umem_bufctl_audit_t *)bcp;
1184 				bzero(bcap, UMEM_BUFCTL_AUDIT_SIZE);
1185 				bcap->bc_cache = cp;
1186 			}
1187 			bcp->bc_addr = buf;
1188 			bcp->bc_slab = sp;
1189 		} else {
1190 			bcp = UMEM_BUFCTL(cp, buf);
1191 		}
1192 		if (cache_flags & UMF_BUFTAG) {
1193 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1194 			btp->bt_redzone = UMEM_REDZONE_PATTERN;
1195 			btp->bt_bufctl = bcp;
1196 			btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
1197 			if (cache_flags & UMF_DEADBEEF) {
1198 				copy_pattern(UMEM_FREE_PATTERN, buf,
1199 				    cp->cache_verify);
1200 			}
1201 		}
1202 		bcp->bc_next = sp->slab_head;
1203 		sp->slab_head = bcp;
1204 		buf += chunksize;
1205 	}
1206 
1207 	umem_log_event(umem_slab_log, cp, sp, slab);
1208 
1209 	return (sp);
1210 
1211 bufctl_alloc_failure:
1212 
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 slab_alloc_failure:
1220 
1221 	vmem_free(vmp, slab, slabsize);
1222 
1223 vmem_alloc_failure:
1224 
1225 	umem_log_event(umem_failure_log, cp, NULL, NULL);
1226 	atomic_add_64(&cp->cache_alloc_fail, 1);
1227 
1228 	return (NULL);
1229 }
1230 
1231 /*
1232  * Destroy a slab.
1233  */
1234 static void
1235 umem_slab_destroy(umem_cache_t *cp, umem_slab_t *sp)
1236 {
1237 	vmem_t *vmp = cp->cache_arena;
1238 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1239 
1240 	if (cp->cache_flags & UMF_HASH) {
1241 		umem_bufctl_t *bcp;
1242 		while ((bcp = sp->slab_head) != NULL) {
1243 			sp->slab_head = bcp->bc_next;
1244 			_umem_cache_free(cp->cache_bufctl_cache, bcp);
1245 		}
1246 		_umem_cache_free(umem_slab_cache, sp);
1247 	}
1248 	vmem_free(vmp, slab, cp->cache_slabsize);
1249 }
1250 
1251 /*
1252  * Allocate a raw (unconstructed) buffer from cp's slab layer.
1253  */
1254 static void *
1255 umem_slab_alloc(umem_cache_t *cp, int umflag)
1256 {
1257 	umem_bufctl_t *bcp, **hash_bucket;
1258 	umem_slab_t *sp;
1259 	void *buf;
1260 
1261 	(void) mutex_lock(&cp->cache_lock);
1262 	cp->cache_slab_alloc++;
1263 	sp = cp->cache_freelist;
1264 	ASSERT(sp->slab_cache == cp);
1265 	if (sp->slab_head == NULL) {
1266 		/*
1267 		 * The freelist is empty.  Create a new slab.
1268 		 */
1269 		(void) mutex_unlock(&cp->cache_lock);
1270 		if (cp == &umem_null_cache)
1271 			return (NULL);
1272 		if ((sp = umem_slab_create(cp, umflag)) == NULL)
1273 			return (NULL);
1274 		(void) mutex_lock(&cp->cache_lock);
1275 		cp->cache_slab_create++;
1276 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1277 			cp->cache_bufmax = cp->cache_buftotal;
1278 		sp->slab_next = cp->cache_freelist;
1279 		sp->slab_prev = cp->cache_freelist->slab_prev;
1280 		sp->slab_next->slab_prev = sp;
1281 		sp->slab_prev->slab_next = sp;
1282 		cp->cache_freelist = sp;
1283 	}
1284 
1285 	sp->slab_refcnt++;
1286 	ASSERT(sp->slab_refcnt <= sp->slab_chunks);
1287 
1288 	/*
1289 	 * If we're taking the last buffer in the slab,
1290 	 * remove the slab from the cache's freelist.
1291 	 */
1292 	bcp = sp->slab_head;
1293 	if ((sp->slab_head = bcp->bc_next) == NULL) {
1294 		cp->cache_freelist = sp->slab_next;
1295 		ASSERT(sp->slab_refcnt == sp->slab_chunks);
1296 	}
1297 
1298 	if (cp->cache_flags & UMF_HASH) {
1299 		/*
1300 		 * Add buffer to allocated-address hash table.
1301 		 */
1302 		buf = bcp->bc_addr;
1303 		hash_bucket = UMEM_HASH(cp, buf);
1304 		bcp->bc_next = *hash_bucket;
1305 		*hash_bucket = bcp;
1306 		if ((cp->cache_flags & (UMF_AUDIT | UMF_BUFTAG)) == UMF_AUDIT) {
1307 			UMEM_AUDIT(umem_transaction_log, cp, bcp);
1308 		}
1309 	} else {
1310 		buf = UMEM_BUF(cp, bcp);
1311 	}
1312 
1313 	ASSERT(UMEM_SLAB_MEMBER(sp, buf));
1314 
1315 	(void) mutex_unlock(&cp->cache_lock);
1316 
1317 	return (buf);
1318 }
1319 
1320 /*
1321  * Free a raw (unconstructed) buffer to cp's slab layer.
1322  */
1323 static void
1324 umem_slab_free(umem_cache_t *cp, void *buf)
1325 {
1326 	umem_slab_t *sp;
1327 	umem_bufctl_t *bcp, **prev_bcpp;
1328 
1329 	ASSERT(buf != NULL);
1330 
1331 	(void) mutex_lock(&cp->cache_lock);
1332 	cp->cache_slab_free++;
1333 
1334 	if (cp->cache_flags & UMF_HASH) {
1335 		/*
1336 		 * Look up buffer in allocated-address hash table.
1337 		 */
1338 		prev_bcpp = UMEM_HASH(cp, buf);
1339 		while ((bcp = *prev_bcpp) != NULL) {
1340 			if (bcp->bc_addr == buf) {
1341 				*prev_bcpp = bcp->bc_next;
1342 				sp = bcp->bc_slab;
1343 				break;
1344 			}
1345 			cp->cache_lookup_depth++;
1346 			prev_bcpp = &bcp->bc_next;
1347 		}
1348 	} else {
1349 		bcp = UMEM_BUFCTL(cp, buf);
1350 		sp = UMEM_SLAB(cp, buf);
1351 	}
1352 
1353 	if (bcp == NULL || sp->slab_cache != cp || !UMEM_SLAB_MEMBER(sp, buf)) {
1354 		(void) mutex_unlock(&cp->cache_lock);
1355 		umem_error(UMERR_BADADDR, cp, buf);
1356 		return;
1357 	}
1358 
1359 	if ((cp->cache_flags & (UMF_AUDIT | UMF_BUFTAG)) == UMF_AUDIT) {
1360 		if (cp->cache_flags & UMF_CONTENTS)
1361 			((umem_bufctl_audit_t *)bcp)->bc_contents =
1362 			    umem_log_enter(umem_content_log, buf,
1363 			    cp->cache_contents);
1364 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
1365 	}
1366 
1367 	/*
1368 	 * If this slab isn't currently on the freelist, put it there.
1369 	 */
1370 	if (sp->slab_head == NULL) {
1371 		ASSERT(sp->slab_refcnt == sp->slab_chunks);
1372 		ASSERT(cp->cache_freelist != sp);
1373 		sp->slab_next->slab_prev = sp->slab_prev;
1374 		sp->slab_prev->slab_next = sp->slab_next;
1375 		sp->slab_next = cp->cache_freelist;
1376 		sp->slab_prev = cp->cache_freelist->slab_prev;
1377 		sp->slab_next->slab_prev = sp;
1378 		sp->slab_prev->slab_next = sp;
1379 		cp->cache_freelist = sp;
1380 	}
1381 
1382 	bcp->bc_next = sp->slab_head;
1383 	sp->slab_head = bcp;
1384 
1385 	ASSERT(sp->slab_refcnt >= 1);
1386 	if (--sp->slab_refcnt == 0) {
1387 		/*
1388 		 * There are no outstanding allocations from this slab,
1389 		 * so we can reclaim the memory.
1390 		 */
1391 		sp->slab_next->slab_prev = sp->slab_prev;
1392 		sp->slab_prev->slab_next = sp->slab_next;
1393 		if (sp == cp->cache_freelist)
1394 			cp->cache_freelist = sp->slab_next;
1395 		cp->cache_slab_destroy++;
1396 		cp->cache_buftotal -= sp->slab_chunks;
1397 		(void) mutex_unlock(&cp->cache_lock);
1398 		umem_slab_destroy(cp, sp);
1399 		return;
1400 	}
1401 	(void) mutex_unlock(&cp->cache_lock);
1402 }
1403 
1404 static int
1405 umem_cache_alloc_debug(umem_cache_t *cp, void *buf, int umflag)
1406 {
1407 	umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1408 	umem_bufctl_audit_t *bcp = (umem_bufctl_audit_t *)btp->bt_bufctl;
1409 	uint32_t mtbf;
1410 	int flags_nfatal;
1411 
1412 	if (btp->bt_bxstat != ((intptr_t)bcp ^ UMEM_BUFTAG_FREE)) {
1413 		umem_error(UMERR_BADBUFTAG, cp, buf);
1414 		return (-1);
1415 	}
1416 
1417 	btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_ALLOC;
1418 
1419 	if ((cp->cache_flags & UMF_HASH) && bcp->bc_addr != buf) {
1420 		umem_error(UMERR_BADBUFCTL, cp, buf);
1421 		return (-1);
1422 	}
1423 
1424 	btp->bt_redzone = UMEM_REDZONE_PATTERN;
1425 
1426 	if (cp->cache_flags & UMF_DEADBEEF) {
1427 		if (verify_and_copy_pattern(UMEM_FREE_PATTERN,
1428 		    UMEM_UNINITIALIZED_PATTERN, buf, cp->cache_verify)) {
1429 			umem_error(UMERR_MODIFIED, cp, buf);
1430 			return (-1);
1431 		}
1432 	}
1433 
1434 	if ((mtbf = umem_mtbf | cp->cache_mtbf) != 0 &&
1435 	    gethrtime() % mtbf == 0 &&
1436 	    (umflag & (UMEM_FATAL_FLAGS)) == 0) {
1437 		umem_log_event(umem_failure_log, cp, NULL, NULL);
1438 	} else {
1439 		mtbf = 0;
1440 	}
1441 
1442 	/*
1443 	 * We do not pass fatal flags on to the constructor.  This prevents
1444 	 * leaking buffers in the event of a subordinate constructor failing.
1445 	 */
1446 	flags_nfatal = UMEM_DEFAULT;
1447 	if (mtbf || (cp->cache_constructor != NULL &&
1448 	    cp->cache_constructor(buf, cp->cache_private, flags_nfatal) != 0)) {
1449 		atomic_add_64(&cp->cache_alloc_fail, 1);
1450 		btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
1451 		copy_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
1452 		umem_slab_free(cp, buf);
1453 		return (-1);
1454 	}
1455 
1456 	if (cp->cache_flags & UMF_AUDIT) {
1457 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
1458 	}
1459 
1460 	return (0);
1461 }
1462 
1463 static int
1464 umem_cache_free_debug(umem_cache_t *cp, void *buf)
1465 {
1466 	umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1467 	umem_bufctl_audit_t *bcp = (umem_bufctl_audit_t *)btp->bt_bufctl;
1468 	umem_slab_t *sp;
1469 
1470 	if (btp->bt_bxstat != ((intptr_t)bcp ^ UMEM_BUFTAG_ALLOC)) {
1471 		if (btp->bt_bxstat == ((intptr_t)bcp ^ UMEM_BUFTAG_FREE)) {
1472 			umem_error(UMERR_DUPFREE, cp, buf);
1473 			return (-1);
1474 		}
1475 		sp = umem_findslab(cp, buf);
1476 		if (sp == NULL || sp->slab_cache != cp)
1477 			umem_error(UMERR_BADADDR, cp, buf);
1478 		else
1479 			umem_error(UMERR_REDZONE, cp, buf);
1480 		return (-1);
1481 	}
1482 
1483 	btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
1484 
1485 	if ((cp->cache_flags & UMF_HASH) && bcp->bc_addr != buf) {
1486 		umem_error(UMERR_BADBUFCTL, cp, buf);
1487 		return (-1);
1488 	}
1489 
1490 	if (btp->bt_redzone != UMEM_REDZONE_PATTERN) {
1491 		umem_error(UMERR_REDZONE, cp, buf);
1492 		return (-1);
1493 	}
1494 
1495 	if (cp->cache_flags & UMF_AUDIT) {
1496 		if (cp->cache_flags & UMF_CONTENTS)
1497 			bcp->bc_contents = umem_log_enter(umem_content_log,
1498 			    buf, cp->cache_contents);
1499 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
1500 	}
1501 
1502 	if (cp->cache_destructor != NULL)
1503 		cp->cache_destructor(buf, cp->cache_private);
1504 
1505 	if (cp->cache_flags & UMF_DEADBEEF)
1506 		copy_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
1507 
1508 	return (0);
1509 }
1510 
1511 /*
1512  * Free each object in magazine mp to cp's slab layer, and free mp itself.
1513  */
1514 static void
1515 umem_magazine_destroy(umem_cache_t *cp, umem_magazine_t *mp, int nrounds)
1516 {
1517 	int round;
1518 
1519 	ASSERT(cp->cache_next == NULL || IN_UPDATE());
1520 
1521 	for (round = 0; round < nrounds; round++) {
1522 		void *buf = mp->mag_round[round];
1523 
1524 		if ((cp->cache_flags & UMF_DEADBEEF) &&
1525 		    verify_pattern(UMEM_FREE_PATTERN, buf,
1526 		    cp->cache_verify) != NULL) {
1527 			umem_error(UMERR_MODIFIED, cp, buf);
1528 			continue;
1529 		}
1530 
1531 		if (!(cp->cache_flags & UMF_BUFTAG) &&
1532 		    cp->cache_destructor != NULL)
1533 			cp->cache_destructor(buf, cp->cache_private);
1534 
1535 		umem_slab_free(cp, buf);
1536 	}
1537 	ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
1538 	_umem_cache_free(cp->cache_magtype->mt_cache, mp);
1539 }
1540 
1541 /*
1542  * Allocate a magazine from the depot.
1543  */
1544 static umem_magazine_t *
1545 umem_depot_alloc(umem_cache_t *cp, umem_maglist_t *mlp)
1546 {
1547 	umem_magazine_t *mp;
1548 
1549 	/*
1550 	 * If we can't get the depot lock without contention,
1551 	 * update our contention count.  We use the depot
1552 	 * contention rate to determine whether we need to
1553 	 * increase the magazine size for better scalability.
1554 	 */
1555 	if (mutex_trylock(&cp->cache_depot_lock) != 0) {
1556 		(void) mutex_lock(&cp->cache_depot_lock);
1557 		cp->cache_depot_contention++;
1558 	}
1559 
1560 	if ((mp = mlp->ml_list) != NULL) {
1561 		ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
1562 		mlp->ml_list = mp->mag_next;
1563 		if (--mlp->ml_total < mlp->ml_min)
1564 			mlp->ml_min = mlp->ml_total;
1565 		mlp->ml_alloc++;
1566 	}
1567 
1568 	(void) mutex_unlock(&cp->cache_depot_lock);
1569 
1570 	return (mp);
1571 }
1572 
1573 /*
1574  * Free a magazine to the depot.
1575  */
1576 static void
1577 umem_depot_free(umem_cache_t *cp, umem_maglist_t *mlp, umem_magazine_t *mp)
1578 {
1579 	(void) mutex_lock(&cp->cache_depot_lock);
1580 	ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
1581 	mp->mag_next = mlp->ml_list;
1582 	mlp->ml_list = mp;
1583 	mlp->ml_total++;
1584 	(void) mutex_unlock(&cp->cache_depot_lock);
1585 }
1586 
1587 /*
1588  * Update the working set statistics for cp's depot.
1589  */
1590 static void
1591 umem_depot_ws_update(umem_cache_t *cp)
1592 {
1593 	(void) mutex_lock(&cp->cache_depot_lock);
1594 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
1595 	cp->cache_full.ml_min = cp->cache_full.ml_total;
1596 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
1597 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
1598 	(void) mutex_unlock(&cp->cache_depot_lock);
1599 }
1600 
1601 /*
1602  * Reap all magazines that have fallen out of the depot's working set.
1603  */
1604 static void
1605 umem_depot_ws_reap(umem_cache_t *cp)
1606 {
1607 	long reap;
1608 	umem_magazine_t *mp;
1609 
1610 	ASSERT(cp->cache_next == NULL || IN_REAP());
1611 
1612 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
1613 	while (reap-- && (mp = umem_depot_alloc(cp, &cp->cache_full)) != NULL)
1614 		umem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
1615 
1616 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
1617 	while (reap-- && (mp = umem_depot_alloc(cp, &cp->cache_empty)) != NULL)
1618 		umem_magazine_destroy(cp, mp, 0);
1619 }
1620 
1621 static void
1622 umem_cpu_reload(umem_cpu_cache_t *ccp, umem_magazine_t *mp, int rounds)
1623 {
1624 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
1625 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
1626 	ASSERT(ccp->cc_magsize > 0);
1627 
1628 	ccp->cc_ploaded = ccp->cc_loaded;
1629 	ccp->cc_prounds = ccp->cc_rounds;
1630 	ccp->cc_loaded = mp;
1631 	ccp->cc_rounds = rounds;
1632 }
1633 
1634 /*
1635  * Allocate a constructed object from cache cp.
1636  */
1637 #pragma weak umem_cache_alloc = _umem_cache_alloc
1638 void *
1639 _umem_cache_alloc(umem_cache_t *cp, int umflag)
1640 {
1641 	umem_cpu_cache_t *ccp;
1642 	umem_magazine_t *fmp;
1643 	void *buf;
1644 	int flags_nfatal;
1645 
1646 retry:
1647 	ccp = UMEM_CPU_CACHE(cp, CPU(cp->cache_cpu_mask));
1648 	(void) mutex_lock(&ccp->cc_lock);
1649 	for (;;) {
1650 		/*
1651 		 * If there's an object available in the current CPU's
1652 		 * loaded magazine, just take it and return.
1653 		 */
1654 		if (ccp->cc_rounds > 0) {
1655 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
1656 			ccp->cc_alloc++;
1657 			(void) mutex_unlock(&ccp->cc_lock);
1658 			if ((ccp->cc_flags & UMF_BUFTAG) &&
1659 			    umem_cache_alloc_debug(cp, buf, umflag) == -1) {
1660 				if (umem_alloc_retry(cp, umflag)) {
1661 					goto retry;
1662 				}
1663 
1664 				return (NULL);
1665 			}
1666 			return (buf);
1667 		}
1668 
1669 		/*
1670 		 * The loaded magazine is empty.  If the previously loaded
1671 		 * magazine was full, exchange them and try again.
1672 		 */
1673 		if (ccp->cc_prounds > 0) {
1674 			umem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
1675 			continue;
1676 		}
1677 
1678 		/*
1679 		 * If the magazine layer is disabled, break out now.
1680 		 */
1681 		if (ccp->cc_magsize == 0)
1682 			break;
1683 
1684 		/*
1685 		 * Try to get a full magazine from the depot.
1686 		 */
1687 		fmp = umem_depot_alloc(cp, &cp->cache_full);
1688 		if (fmp != NULL) {
1689 			if (ccp->cc_ploaded != NULL)
1690 				umem_depot_free(cp, &cp->cache_empty,
1691 				    ccp->cc_ploaded);
1692 			umem_cpu_reload(ccp, fmp, ccp->cc_magsize);
1693 			continue;
1694 		}
1695 
1696 		/*
1697 		 * There are no full magazines in the depot,
1698 		 * so fall through to the slab layer.
1699 		 */
1700 		break;
1701 	}
1702 	(void) mutex_unlock(&ccp->cc_lock);
1703 
1704 	/*
1705 	 * We couldn't allocate a constructed object from the magazine layer,
1706 	 * so get a raw buffer from the slab layer and apply its constructor.
1707 	 */
1708 	buf = umem_slab_alloc(cp, umflag);
1709 
1710 	if (buf == NULL) {
1711 		if (cp == &umem_null_cache)
1712 			return (NULL);
1713 		if (umem_alloc_retry(cp, umflag)) {
1714 			goto retry;
1715 		}
1716 
1717 		return (NULL);
1718 	}
1719 
1720 	if (cp->cache_flags & UMF_BUFTAG) {
1721 		/*
1722 		 * Let umem_cache_alloc_debug() apply the constructor for us.
1723 		 */
1724 		if (umem_cache_alloc_debug(cp, buf, umflag) == -1) {
1725 			if (umem_alloc_retry(cp, umflag)) {
1726 				goto retry;
1727 			}
1728 			return (NULL);
1729 		}
1730 		return (buf);
1731 	}
1732 
1733 	/*
1734 	 * We do not pass fatal flags on to the constructor.  This prevents
1735 	 * leaking buffers in the event of a subordinate constructor failing.
1736 	 */
1737 	flags_nfatal = UMEM_DEFAULT;
1738 	if (cp->cache_constructor != NULL &&
1739 	    cp->cache_constructor(buf, cp->cache_private, flags_nfatal) != 0) {
1740 		atomic_add_64(&cp->cache_alloc_fail, 1);
1741 		umem_slab_free(cp, buf);
1742 
1743 		if (umem_alloc_retry(cp, umflag)) {
1744 			goto retry;
1745 		}
1746 		return (NULL);
1747 	}
1748 
1749 	return (buf);
1750 }
1751 
1752 /*
1753  * Free a constructed object to cache cp.
1754  */
1755 #pragma weak umem_cache_free = _umem_cache_free
1756 void
1757 _umem_cache_free(umem_cache_t *cp, void *buf)
1758 {
1759 	umem_cpu_cache_t *ccp = UMEM_CPU_CACHE(cp, CPU(cp->cache_cpu_mask));
1760 	umem_magazine_t *emp;
1761 	umem_magtype_t *mtp;
1762 
1763 	if (ccp->cc_flags & UMF_BUFTAG)
1764 		if (umem_cache_free_debug(cp, buf) == -1)
1765 			return;
1766 
1767 	(void) mutex_lock(&ccp->cc_lock);
1768 	for (;;) {
1769 		/*
1770 		 * If there's a slot available in the current CPU's
1771 		 * loaded magazine, just put the object there and return.
1772 		 */
1773 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
1774 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
1775 			ccp->cc_free++;
1776 			(void) mutex_unlock(&ccp->cc_lock);
1777 			return;
1778 		}
1779 
1780 		/*
1781 		 * The loaded magazine is full.  If the previously loaded
1782 		 * magazine was empty, exchange them and try again.
1783 		 */
1784 		if (ccp->cc_prounds == 0) {
1785 			umem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
1786 			continue;
1787 		}
1788 
1789 		/*
1790 		 * If the magazine layer is disabled, break out now.
1791 		 */
1792 		if (ccp->cc_magsize == 0)
1793 			break;
1794 
1795 		/*
1796 		 * Try to get an empty magazine from the depot.
1797 		 */
1798 		emp = umem_depot_alloc(cp, &cp->cache_empty);
1799 		if (emp != NULL) {
1800 			if (ccp->cc_ploaded != NULL)
1801 				umem_depot_free(cp, &cp->cache_full,
1802 				    ccp->cc_ploaded);
1803 			umem_cpu_reload(ccp, emp, 0);
1804 			continue;
1805 		}
1806 
1807 		/*
1808 		 * There are no empty magazines in the depot,
1809 		 * so try to allocate a new one.  We must drop all locks
1810 		 * across umem_cache_alloc() because lower layers may
1811 		 * attempt to allocate from this cache.
1812 		 */
1813 		mtp = cp->cache_magtype;
1814 		(void) mutex_unlock(&ccp->cc_lock);
1815 		emp = _umem_cache_alloc(mtp->mt_cache, UMEM_DEFAULT);
1816 		(void) mutex_lock(&ccp->cc_lock);
1817 
1818 		if (emp != NULL) {
1819 			/*
1820 			 * We successfully allocated an empty magazine.
1821 			 * However, we had to drop ccp->cc_lock to do it,
1822 			 * so the cache's magazine size may have changed.
1823 			 * If so, free the magazine and try again.
1824 			 */
1825 			if (ccp->cc_magsize != mtp->mt_magsize) {
1826 				(void) mutex_unlock(&ccp->cc_lock);
1827 				_umem_cache_free(mtp->mt_cache, emp);
1828 				(void) mutex_lock(&ccp->cc_lock);
1829 				continue;
1830 			}
1831 
1832 			/*
1833 			 * We got a magazine of the right size.  Add it to
1834 			 * the depot and try the whole dance again.
1835 			 */
1836 			umem_depot_free(cp, &cp->cache_empty, emp);
1837 			continue;
1838 		}
1839 
1840 		/*
1841 		 * We couldn't allocate an empty magazine,
1842 		 * so fall through to the slab layer.
1843 		 */
1844 		break;
1845 	}
1846 	(void) mutex_unlock(&ccp->cc_lock);
1847 
1848 	/*
1849 	 * We couldn't free our constructed object to the magazine layer,
1850 	 * so apply its destructor and free it to the slab layer.
1851 	 * Note that if UMF_BUFTAG is in effect, umem_cache_free_debug()
1852 	 * will have already applied the destructor.
1853 	 */
1854 	if (!(cp->cache_flags & UMF_BUFTAG) && cp->cache_destructor != NULL)
1855 		cp->cache_destructor(buf, cp->cache_private);
1856 
1857 	umem_slab_free(cp, buf);
1858 }
1859 
1860 #pragma weak umem_zalloc = _umem_zalloc
1861 void *
1862 _umem_zalloc(size_t size, int umflag)
1863 {
1864 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
1865 	void *buf;
1866 
1867 retry:
1868 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
1869 		umem_cache_t *cp = umem_alloc_table[index];
1870 		buf = _umem_cache_alloc(cp, umflag);
1871 		if (buf != NULL) {
1872 			if (cp->cache_flags & UMF_BUFTAG) {
1873 				umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1874 				((uint8_t *)buf)[size] = UMEM_REDZONE_BYTE;
1875 				((uint32_t *)btp)[1] = UMEM_SIZE_ENCODE(size);
1876 			}
1877 			bzero(buf, size);
1878 		} else if (umem_alloc_retry(cp, umflag))
1879 			goto retry;
1880 	} else {
1881 		buf = _umem_alloc(size, umflag);	/* handles failure */
1882 		if (buf != NULL)
1883 			bzero(buf, size);
1884 	}
1885 	return (buf);
1886 }
1887 
1888 #pragma weak umem_alloc = _umem_alloc
1889 void *
1890 _umem_alloc(size_t size, int umflag)
1891 {
1892 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
1893 	void *buf;
1894 umem_alloc_retry:
1895 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
1896 		umem_cache_t *cp = umem_alloc_table[index];
1897 		buf = _umem_cache_alloc(cp, umflag);
1898 		if ((cp->cache_flags & UMF_BUFTAG) && buf != NULL) {
1899 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1900 			((uint8_t *)buf)[size] = UMEM_REDZONE_BYTE;
1901 			((uint32_t *)btp)[1] = UMEM_SIZE_ENCODE(size);
1902 		}
1903 		if (buf == NULL && umem_alloc_retry(cp, umflag))
1904 			goto umem_alloc_retry;
1905 		return (buf);
1906 	}
1907 	if (size == 0)
1908 		return (NULL);
1909 	if (umem_oversize_arena == NULL) {
1910 		if (umem_init())
1911 			ASSERT(umem_oversize_arena != NULL);
1912 		else
1913 			return (NULL);
1914 	}
1915 	buf = vmem_alloc(umem_oversize_arena, size, UMEM_VMFLAGS(umflag));
1916 	if (buf == NULL) {
1917 		umem_log_event(umem_failure_log, NULL, NULL, (void *)size);
1918 		if (umem_alloc_retry(NULL, umflag))
1919 			goto umem_alloc_retry;
1920 	}
1921 	return (buf);
1922 }
1923 
1924 #pragma weak umem_alloc_align = _umem_alloc_align
1925 void *
1926 _umem_alloc_align(size_t size, size_t align, int umflag)
1927 {
1928 	void *buf;
1929 
1930 	if (size == 0)
1931 		return (NULL);
1932 	if ((align & (align - 1)) != 0)
1933 		return (NULL);
1934 	if (align < UMEM_ALIGN)
1935 		align = UMEM_ALIGN;
1936 
1937 umem_alloc_align_retry:
1938 	if (umem_memalign_arena == NULL) {
1939 		if (umem_init())
1940 			ASSERT(umem_oversize_arena != NULL);
1941 		else
1942 			return (NULL);
1943 	}
1944 	buf = vmem_xalloc(umem_memalign_arena, size, align, 0, 0, NULL, NULL,
1945 	    UMEM_VMFLAGS(umflag));
1946 	if (buf == NULL) {
1947 		umem_log_event(umem_failure_log, NULL, NULL, (void *)size);
1948 		if (umem_alloc_retry(NULL, umflag))
1949 			goto umem_alloc_align_retry;
1950 	}
1951 	return (buf);
1952 }
1953 
1954 #pragma weak umem_free = _umem_free
1955 void
1956 _umem_free(void *buf, size_t size)
1957 {
1958 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
1959 
1960 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
1961 		umem_cache_t *cp = umem_alloc_table[index];
1962 		if (cp->cache_flags & UMF_BUFTAG) {
1963 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1964 			uint32_t *ip = (uint32_t *)btp;
1965 			if (ip[1] != UMEM_SIZE_ENCODE(size)) {
1966 				if (*(uint64_t *)buf == UMEM_FREE_PATTERN) {
1967 					umem_error(UMERR_DUPFREE, cp, buf);
1968 					return;
1969 				}
1970 				if (UMEM_SIZE_VALID(ip[1])) {
1971 					ip[0] = UMEM_SIZE_ENCODE(size);
1972 					umem_error(UMERR_BADSIZE, cp, buf);
1973 				} else {
1974 					umem_error(UMERR_REDZONE, cp, buf);
1975 				}
1976 				return;
1977 			}
1978 			if (((uint8_t *)buf)[size] != UMEM_REDZONE_BYTE) {
1979 				umem_error(UMERR_REDZONE, cp, buf);
1980 				return;
1981 			}
1982 			btp->bt_redzone = UMEM_REDZONE_PATTERN;
1983 		}
1984 		_umem_cache_free(cp, buf);
1985 	} else {
1986 		if (buf == NULL && size == 0)
1987 			return;
1988 		vmem_free(umem_oversize_arena, buf, size);
1989 	}
1990 }
1991 
1992 #pragma weak umem_free_align = _umem_free_align
1993 void
1994 _umem_free_align(void *buf, size_t size)
1995 {
1996 	if (buf == NULL && size == 0)
1997 		return;
1998 	vmem_xfree(umem_memalign_arena, buf, size);
1999 }
2000 
2001 static void *
2002 umem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
2003 {
2004 	size_t realsize = size + vmp->vm_quantum;
2005 
2006 	/*
2007 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
2008 	 * vm_quantum will cause integer wraparound.  Check for this, and
2009 	 * blow off the firewall page in this case.  Note that such a
2010 	 * giant allocation (the entire address space) can never be
2011 	 * satisfied, so it will either fail immediately (VM_NOSLEEP)
2012 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
2013 	 * corresponding check in umem_firewall_va_free().
2014 	 */
2015 	if (realsize < size)
2016 		realsize = size;
2017 
2018 	return (vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT));
2019 }
2020 
2021 static void
2022 umem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
2023 {
2024 	vmem_free(vmp, addr, size + vmp->vm_quantum);
2025 }
2026 
2027 /*
2028  * Reclaim all unused memory from a cache.
2029  */
2030 static void
2031 umem_cache_reap(umem_cache_t *cp)
2032 {
2033 	/*
2034 	 * Ask the cache's owner to free some memory if possible.
2035 	 * The idea is to handle things like the inode cache, which
2036 	 * typically sits on a bunch of memory that it doesn't truly
2037 	 * *need*.  Reclaim policy is entirely up to the owner; this
2038 	 * callback is just an advisory plea for help.
2039 	 */
2040 	if (cp->cache_reclaim != NULL)
2041 		cp->cache_reclaim(cp->cache_private);
2042 
2043 	umem_depot_ws_reap(cp);
2044 }
2045 
2046 /*
2047  * Purge all magazines from a cache and set its magazine limit to zero.
2048  * All calls are serialized by being done by the update thread, except for
2049  * the final call from umem_cache_destroy().
2050  */
2051 static void
2052 umem_cache_magazine_purge(umem_cache_t *cp)
2053 {
2054 	umem_cpu_cache_t *ccp;
2055 	umem_magazine_t *mp, *pmp;
2056 	int rounds, prounds, cpu_seqid;
2057 
2058 	ASSERT(cp->cache_next == NULL || IN_UPDATE());
2059 
2060 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
2061 		ccp = &cp->cache_cpu[cpu_seqid];
2062 
2063 		(void) mutex_lock(&ccp->cc_lock);
2064 		mp = ccp->cc_loaded;
2065 		pmp = ccp->cc_ploaded;
2066 		rounds = ccp->cc_rounds;
2067 		prounds = ccp->cc_prounds;
2068 		ccp->cc_loaded = NULL;
2069 		ccp->cc_ploaded = NULL;
2070 		ccp->cc_rounds = -1;
2071 		ccp->cc_prounds = -1;
2072 		ccp->cc_magsize = 0;
2073 		(void) mutex_unlock(&ccp->cc_lock);
2074 
2075 		if (mp)
2076 			umem_magazine_destroy(cp, mp, rounds);
2077 		if (pmp)
2078 			umem_magazine_destroy(cp, pmp, prounds);
2079 	}
2080 
2081 	/*
2082 	 * Updating the working set statistics twice in a row has the
2083 	 * effect of setting the working set size to zero, so everything
2084 	 * is eligible for reaping.
2085 	 */
2086 	umem_depot_ws_update(cp);
2087 	umem_depot_ws_update(cp);
2088 
2089 	umem_depot_ws_reap(cp);
2090 }
2091 
2092 /*
2093  * Enable per-cpu magazines on a cache.
2094  */
2095 static void
2096 umem_cache_magazine_enable(umem_cache_t *cp)
2097 {
2098 	int cpu_seqid;
2099 
2100 	if (cp->cache_flags & UMF_NOMAGAZINE)
2101 		return;
2102 
2103 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
2104 		umem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
2105 		(void) mutex_lock(&ccp->cc_lock);
2106 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
2107 		(void) mutex_unlock(&ccp->cc_lock);
2108 	}
2109 
2110 }
2111 
2112 /*
2113  * Recompute a cache's magazine size.  The trade-off is that larger magazines
2114  * provide a higher transfer rate with the depot, while smaller magazines
2115  * reduce memory consumption.  Magazine resizing is an expensive operation;
2116  * it should not be done frequently.
2117  *
2118  * Changes to the magazine size are serialized by only having one thread
2119  * doing updates. (the update thread)
2120  *
2121  * Note: at present this only grows the magazine size.  It might be useful
2122  * to allow shrinkage too.
2123  */
2124 static void
2125 umem_cache_magazine_resize(umem_cache_t *cp)
2126 {
2127 	umem_magtype_t *mtp = cp->cache_magtype;
2128 
2129 	ASSERT(IN_UPDATE());
2130 
2131 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
2132 		umem_cache_magazine_purge(cp);
2133 		(void) mutex_lock(&cp->cache_depot_lock);
2134 		cp->cache_magtype = ++mtp;
2135 		cp->cache_depot_contention_prev =
2136 		    cp->cache_depot_contention + INT_MAX;
2137 		(void) mutex_unlock(&cp->cache_depot_lock);
2138 		umem_cache_magazine_enable(cp);
2139 	}
2140 }
2141 
2142 /*
2143  * Rescale a cache's hash table, so that the table size is roughly the
2144  * cache size.  We want the average lookup time to be extremely small.
2145  */
2146 static void
2147 umem_hash_rescale(umem_cache_t *cp)
2148 {
2149 	umem_bufctl_t **old_table, **new_table, *bcp;
2150 	size_t old_size, new_size, h;
2151 
2152 	ASSERT(IN_UPDATE());
2153 
2154 	new_size = MAX(UMEM_HASH_INITIAL,
2155 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
2156 	old_size = cp->cache_hash_mask + 1;
2157 
2158 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
2159 		return;
2160 
2161 	new_table = vmem_alloc(umem_hash_arena, new_size * sizeof (void *),
2162 	    VM_NOSLEEP);
2163 	if (new_table == NULL)
2164 		return;
2165 	bzero(new_table, new_size * sizeof (void *));
2166 
2167 	(void) mutex_lock(&cp->cache_lock);
2168 
2169 	old_size = cp->cache_hash_mask + 1;
2170 	old_table = cp->cache_hash_table;
2171 
2172 	cp->cache_hash_mask = new_size - 1;
2173 	cp->cache_hash_table = new_table;
2174 	cp->cache_rescale++;
2175 
2176 	for (h = 0; h < old_size; h++) {
2177 		bcp = old_table[h];
2178 		while (bcp != NULL) {
2179 			void *addr = bcp->bc_addr;
2180 			umem_bufctl_t *next_bcp = bcp->bc_next;
2181 			umem_bufctl_t **hash_bucket = UMEM_HASH(cp, addr);
2182 			bcp->bc_next = *hash_bucket;
2183 			*hash_bucket = bcp;
2184 			bcp = next_bcp;
2185 		}
2186 	}
2187 
2188 	(void) mutex_unlock(&cp->cache_lock);
2189 
2190 	vmem_free(umem_hash_arena, old_table, old_size * sizeof (void *));
2191 }
2192 
2193 /*
2194  * Perform periodic maintenance on a cache: hash rescaling,
2195  * depot working-set update, and magazine resizing.
2196  */
2197 void
2198 umem_cache_update(umem_cache_t *cp)
2199 {
2200 	int update_flags = 0;
2201 
2202 	ASSERT(MUTEX_HELD(&umem_cache_lock));
2203 
2204 	/*
2205 	 * If the cache has become much larger or smaller than its hash table,
2206 	 * fire off a request to rescale the hash table.
2207 	 */
2208 	(void) mutex_lock(&cp->cache_lock);
2209 
2210 	if ((cp->cache_flags & UMF_HASH) &&
2211 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
2212 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
2213 	    cp->cache_hash_mask > UMEM_HASH_INITIAL)))
2214 		update_flags |= UMU_HASH_RESCALE;
2215 
2216 	(void) mutex_unlock(&cp->cache_lock);
2217 
2218 	/*
2219 	 * Update the depot working set statistics.
2220 	 */
2221 	umem_depot_ws_update(cp);
2222 
2223 	/*
2224 	 * If there's a lot of contention in the depot,
2225 	 * increase the magazine size.
2226 	 */
2227 	(void) mutex_lock(&cp->cache_depot_lock);
2228 
2229 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
2230 	    (int)(cp->cache_depot_contention -
2231 	    cp->cache_depot_contention_prev) > umem_depot_contention)
2232 		update_flags |= UMU_MAGAZINE_RESIZE;
2233 
2234 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
2235 
2236 	(void) mutex_unlock(&cp->cache_depot_lock);
2237 
2238 	if (update_flags)
2239 		umem_add_update(cp, update_flags);
2240 }
2241 
2242 /*
2243  * Runs all pending updates.
2244  *
2245  * The update lock must be held on entrance, and will be held on exit.
2246  */
2247 void
2248 umem_process_updates(void)
2249 {
2250 	ASSERT(MUTEX_HELD(&umem_update_lock));
2251 
2252 	while (umem_null_cache.cache_unext != &umem_null_cache) {
2253 		int notify = 0;
2254 		umem_cache_t *cp = umem_null_cache.cache_unext;
2255 
2256 		cp->cache_uprev->cache_unext = cp->cache_unext;
2257 		cp->cache_unext->cache_uprev = cp->cache_uprev;
2258 		cp->cache_uprev = cp->cache_unext = NULL;
2259 
2260 		ASSERT(!(cp->cache_uflags & UMU_ACTIVE));
2261 
2262 		while (cp->cache_uflags) {
2263 			int uflags = (cp->cache_uflags |= UMU_ACTIVE);
2264 			(void) mutex_unlock(&umem_update_lock);
2265 
2266 			/*
2267 			 * The order here is important.  Each step can speed up
2268 			 * later steps.
2269 			 */
2270 
2271 			if (uflags & UMU_HASH_RESCALE)
2272 				umem_hash_rescale(cp);
2273 
2274 			if (uflags & UMU_MAGAZINE_RESIZE)
2275 				umem_cache_magazine_resize(cp);
2276 
2277 			if (uflags & UMU_REAP)
2278 				umem_cache_reap(cp);
2279 
2280 			(void) mutex_lock(&umem_update_lock);
2281 
2282 			/*
2283 			 * check if anyone has requested notification
2284 			 */
2285 			if (cp->cache_uflags & UMU_NOTIFY) {
2286 				uflags |= UMU_NOTIFY;
2287 				notify = 1;
2288 			}
2289 			cp->cache_uflags &= ~uflags;
2290 		}
2291 		if (notify)
2292 			(void) cond_broadcast(&umem_update_cv);
2293 	}
2294 }
2295 
2296 #ifndef UMEM_STANDALONE
2297 static void
2298 umem_st_update(void)
2299 {
2300 	ASSERT(MUTEX_HELD(&umem_update_lock));
2301 	ASSERT(umem_update_thr == 0 && umem_st_update_thr == 0);
2302 
2303 	umem_st_update_thr = thr_self();
2304 
2305 	(void) mutex_unlock(&umem_update_lock);
2306 
2307 	vmem_update(NULL);
2308 	umem_cache_applyall(umem_cache_update);
2309 
2310 	(void) mutex_lock(&umem_update_lock);
2311 
2312 	umem_process_updates();	/* does all of the requested work */
2313 
2314 	umem_reap_next = gethrtime() +
2315 	    (hrtime_t)umem_reap_interval * NANOSEC;
2316 
2317 	umem_reaping = UMEM_REAP_DONE;
2318 
2319 	umem_st_update_thr = 0;
2320 }
2321 #endif
2322 
2323 /*
2324  * Reclaim all unused memory from all caches.  Called from vmem when memory
2325  * gets tight.  Must be called with no locks held.
2326  *
2327  * This just requests a reap on all caches, and notifies the update thread.
2328  */
2329 void
2330 umem_reap(void)
2331 {
2332 #ifndef UMEM_STANDALONE
2333 	extern int __nthreads(void);
2334 #endif
2335 
2336 	if (umem_ready != UMEM_READY || umem_reaping != UMEM_REAP_DONE ||
2337 	    gethrtime() < umem_reap_next)
2338 		return;
2339 
2340 	(void) mutex_lock(&umem_update_lock);
2341 
2342 	if (umem_reaping != UMEM_REAP_DONE || gethrtime() < umem_reap_next) {
2343 		(void) mutex_unlock(&umem_update_lock);
2344 		return;
2345 	}
2346 	umem_reaping = UMEM_REAP_ADDING;	/* lock out other reaps */
2347 
2348 	(void) mutex_unlock(&umem_update_lock);
2349 
2350 	umem_updateall(UMU_REAP);
2351 
2352 	(void) mutex_lock(&umem_update_lock);
2353 
2354 	umem_reaping = UMEM_REAP_ACTIVE;
2355 
2356 	/* Standalone is single-threaded */
2357 #ifndef UMEM_STANDALONE
2358 	if (umem_update_thr == 0) {
2359 		/*
2360 		 * The update thread does not exist.  If the process is
2361 		 * multi-threaded, create it.  If not, or the creation fails,
2362 		 * do the update processing inline.
2363 		 */
2364 		ASSERT(umem_st_update_thr == 0);
2365 
2366 		if (__nthreads() <= 1 || umem_create_update_thread() == 0)
2367 			umem_st_update();
2368 	}
2369 
2370 	(void) cond_broadcast(&umem_update_cv);	/* wake up the update thread */
2371 #endif
2372 
2373 	(void) mutex_unlock(&umem_update_lock);
2374 }
2375 
2376 umem_cache_t *
2377 umem_cache_create(
2378 	char *name,		/* descriptive name for this cache */
2379 	size_t bufsize,		/* size of the objects it manages */
2380 	size_t align,		/* required object alignment */
2381 	umem_constructor_t *constructor, /* object constructor */
2382 	umem_destructor_t *destructor, /* object destructor */
2383 	umem_reclaim_t *reclaim, /* memory reclaim callback */
2384 	void *private,		/* pass-thru arg for constr/destr/reclaim */
2385 	vmem_t *vmp,		/* vmem source for slab allocation */
2386 	int cflags)		/* cache creation flags */
2387 {
2388 	int cpu_seqid;
2389 	size_t chunksize;
2390 	umem_cache_t *cp, *cnext, *cprev;
2391 	umem_magtype_t *mtp;
2392 	size_t csize;
2393 	size_t phase;
2394 
2395 	/*
2396 	 * The init thread is allowed to create internal and quantum caches.
2397 	 *
2398 	 * Other threads must wait until until initialization is complete.
2399 	 */
2400 	if (umem_init_thr == thr_self())
2401 		ASSERT((cflags & (UMC_INTERNAL | UMC_QCACHE)) != 0);
2402 	else {
2403 		ASSERT(!(cflags & UMC_INTERNAL));
2404 		if (umem_ready != UMEM_READY && umem_init() == 0) {
2405 			errno = EAGAIN;
2406 			return (NULL);
2407 		}
2408 	}
2409 
2410 	csize = UMEM_CACHE_SIZE(umem_max_ncpus);
2411 	phase = P2NPHASE(csize, UMEM_CPU_CACHE_SIZE);
2412 
2413 	if (vmp == NULL)
2414 		vmp = umem_default_arena;
2415 
2416 	ASSERT(P2PHASE(phase, UMEM_ALIGN) == 0);
2417 
2418 	/*
2419 	 * Check that the arguments are reasonable
2420 	 */
2421 	if ((align & (align - 1)) != 0 || align > vmp->vm_quantum ||
2422 	    ((cflags & UMC_NOHASH) && (cflags & UMC_NOTOUCH)) ||
2423 	    name == NULL || bufsize == 0) {
2424 		errno = EINVAL;
2425 		return (NULL);
2426 	}
2427 
2428 	/*
2429 	 * If align == 0, we set it to the minimum required alignment.
2430 	 *
2431 	 * If align < UMEM_ALIGN, we round it up to UMEM_ALIGN, unless
2432 	 * UMC_NOTOUCH was passed.
2433 	 */
2434 	if (align == 0) {
2435 		if (P2ROUNDUP(bufsize, UMEM_ALIGN) >= UMEM_SECOND_ALIGN)
2436 			align = UMEM_SECOND_ALIGN;
2437 		else
2438 			align = UMEM_ALIGN;
2439 	} else if (align < UMEM_ALIGN && (cflags & UMC_NOTOUCH) == 0)
2440 		align = UMEM_ALIGN;
2441 
2442 
2443 	/*
2444 	 * Get a umem_cache structure.  We arrange that cp->cache_cpu[]
2445 	 * is aligned on a UMEM_CPU_CACHE_SIZE boundary to prevent
2446 	 * false sharing of per-CPU data.
2447 	 */
2448 	cp = vmem_xalloc(umem_cache_arena, csize, UMEM_CPU_CACHE_SIZE, phase,
2449 	    0, NULL, NULL, VM_NOSLEEP);
2450 
2451 	if (cp == NULL) {
2452 		errno = EAGAIN;
2453 		return (NULL);
2454 	}
2455 
2456 	bzero(cp, csize);
2457 
2458 	(void) mutex_lock(&umem_flags_lock);
2459 	if (umem_flags & UMF_RANDOMIZE)
2460 		umem_flags = (((umem_flags | ~UMF_RANDOM) + 1) & UMF_RANDOM) |
2461 		    UMF_RANDOMIZE;
2462 	cp->cache_flags = umem_flags | (cflags & UMF_DEBUG);
2463 	(void) mutex_unlock(&umem_flags_lock);
2464 
2465 	/*
2466 	 * Make sure all the various flags are reasonable.
2467 	 */
2468 	if (cp->cache_flags & UMF_LITE) {
2469 		if (bufsize >= umem_lite_minsize &&
2470 		    align <= umem_lite_maxalign &&
2471 		    P2PHASE(bufsize, umem_lite_maxalign) != 0) {
2472 			cp->cache_flags |= UMF_BUFTAG;
2473 			cp->cache_flags &= ~(UMF_AUDIT | UMF_FIREWALL);
2474 		} else {
2475 			cp->cache_flags &= ~UMF_DEBUG;
2476 		}
2477 	}
2478 
2479 	if ((cflags & UMC_QCACHE) && (cp->cache_flags & UMF_AUDIT))
2480 		cp->cache_flags |= UMF_NOMAGAZINE;
2481 
2482 	if (cflags & UMC_NODEBUG)
2483 		cp->cache_flags &= ~UMF_DEBUG;
2484 
2485 	if (cflags & UMC_NOTOUCH)
2486 		cp->cache_flags &= ~UMF_TOUCH;
2487 
2488 	if (cflags & UMC_NOHASH)
2489 		cp->cache_flags &= ~(UMF_AUDIT | UMF_FIREWALL);
2490 
2491 	if (cflags & UMC_NOMAGAZINE)
2492 		cp->cache_flags |= UMF_NOMAGAZINE;
2493 
2494 	if ((cp->cache_flags & UMF_AUDIT) && !(cflags & UMC_NOTOUCH))
2495 		cp->cache_flags |= UMF_REDZONE;
2496 
2497 	if ((cp->cache_flags & UMF_BUFTAG) && bufsize >= umem_minfirewall &&
2498 	    !(cp->cache_flags & UMF_LITE) && !(cflags & UMC_NOHASH))
2499 		cp->cache_flags |= UMF_FIREWALL;
2500 
2501 	if (vmp != umem_default_arena || umem_firewall_arena == NULL)
2502 		cp->cache_flags &= ~UMF_FIREWALL;
2503 
2504 	if (cp->cache_flags & UMF_FIREWALL) {
2505 		cp->cache_flags &= ~UMF_BUFTAG;
2506 		cp->cache_flags |= UMF_NOMAGAZINE;
2507 		ASSERT(vmp == umem_default_arena);
2508 		vmp = umem_firewall_arena;
2509 	}
2510 
2511 	/*
2512 	 * Set cache properties.
2513 	 */
2514 	(void) strncpy(cp->cache_name, name, sizeof (cp->cache_name) - 1);
2515 	cp->cache_bufsize = bufsize;
2516 	cp->cache_align = align;
2517 	cp->cache_constructor = constructor;
2518 	cp->cache_destructor = destructor;
2519 	cp->cache_reclaim = reclaim;
2520 	cp->cache_private = private;
2521 	cp->cache_arena = vmp;
2522 	cp->cache_cflags = cflags;
2523 	cp->cache_cpu_mask = umem_cpu_mask;
2524 
2525 	/*
2526 	 * Determine the chunk size.
2527 	 */
2528 	chunksize = bufsize;
2529 
2530 	if (align >= UMEM_ALIGN) {
2531 		chunksize = P2ROUNDUP(chunksize, UMEM_ALIGN);
2532 		cp->cache_bufctl = chunksize - UMEM_ALIGN;
2533 	}
2534 
2535 	if (cp->cache_flags & UMF_BUFTAG) {
2536 		cp->cache_bufctl = chunksize;
2537 		cp->cache_buftag = chunksize;
2538 		chunksize += sizeof (umem_buftag_t);
2539 	}
2540 
2541 	if (cp->cache_flags & UMF_DEADBEEF) {
2542 		cp->cache_verify = MIN(cp->cache_buftag, umem_maxverify);
2543 		if (cp->cache_flags & UMF_LITE)
2544 			cp->cache_verify = MIN(cp->cache_verify, UMEM_ALIGN);
2545 	}
2546 
2547 	cp->cache_contents = MIN(cp->cache_bufctl, umem_content_maxsave);
2548 
2549 	cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
2550 
2551 	if (chunksize < bufsize) {
2552 		errno = ENOMEM;
2553 		goto fail;
2554 	}
2555 
2556 	/*
2557 	 * Now that we know the chunk size, determine the optimal slab size.
2558 	 */
2559 	if (vmp == umem_firewall_arena) {
2560 		cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
2561 		cp->cache_mincolor = cp->cache_slabsize - chunksize;
2562 		cp->cache_maxcolor = cp->cache_mincolor;
2563 		cp->cache_flags |= UMF_HASH;
2564 		ASSERT(!(cp->cache_flags & UMF_BUFTAG));
2565 	} else if ((cflags & UMC_NOHASH) || (!(cflags & UMC_NOTOUCH) &&
2566 	    !(cp->cache_flags & UMF_AUDIT) &&
2567 	    chunksize < vmp->vm_quantum / UMEM_VOID_FRACTION)) {
2568 		cp->cache_slabsize = vmp->vm_quantum;
2569 		cp->cache_mincolor = 0;
2570 		cp->cache_maxcolor =
2571 		    (cp->cache_slabsize - sizeof (umem_slab_t)) % chunksize;
2572 
2573 		if (chunksize + sizeof (umem_slab_t) > cp->cache_slabsize) {
2574 			errno = EINVAL;
2575 			goto fail;
2576 		}
2577 		ASSERT(!(cp->cache_flags & UMF_AUDIT));
2578 	} else {
2579 		size_t chunks, bestfit, waste, slabsize;
2580 		size_t minwaste = LONG_MAX;
2581 
2582 		for (chunks = 1; chunks <= UMEM_VOID_FRACTION; chunks++) {
2583 			slabsize = P2ROUNDUP(chunksize * chunks,
2584 			    vmp->vm_quantum);
2585 			/*
2586 			 * check for overflow
2587 			 */
2588 			if ((slabsize / chunks) < chunksize) {
2589 				errno = ENOMEM;
2590 				goto fail;
2591 			}
2592 			chunks = slabsize / chunksize;
2593 			waste = (slabsize % chunksize) / chunks;
2594 			if (waste < minwaste) {
2595 				minwaste = waste;
2596 				bestfit = slabsize;
2597 			}
2598 		}
2599 		if (cflags & UMC_QCACHE)
2600 			bestfit = MAX(1 << highbit(3 * vmp->vm_qcache_max), 64);
2601 		cp->cache_slabsize = bestfit;
2602 		cp->cache_mincolor = 0;
2603 		cp->cache_maxcolor = bestfit % chunksize;
2604 		cp->cache_flags |= UMF_HASH;
2605 	}
2606 
2607 	if (cp->cache_flags & UMF_HASH) {
2608 		ASSERT(!(cflags & UMC_NOHASH));
2609 		cp->cache_bufctl_cache = (cp->cache_flags & UMF_AUDIT) ?
2610 		    umem_bufctl_audit_cache : umem_bufctl_cache;
2611 	}
2612 
2613 	if (cp->cache_maxcolor >= vmp->vm_quantum)
2614 		cp->cache_maxcolor = vmp->vm_quantum - 1;
2615 
2616 	cp->cache_color = cp->cache_mincolor;
2617 
2618 	/*
2619 	 * Initialize the rest of the slab layer.
2620 	 */
2621 	(void) mutex_init(&cp->cache_lock, USYNC_THREAD, NULL);
2622 
2623 	cp->cache_freelist = &cp->cache_nullslab;
2624 	cp->cache_nullslab.slab_cache = cp;
2625 	cp->cache_nullslab.slab_refcnt = -1;
2626 	cp->cache_nullslab.slab_next = &cp->cache_nullslab;
2627 	cp->cache_nullslab.slab_prev = &cp->cache_nullslab;
2628 
2629 	if (cp->cache_flags & UMF_HASH) {
2630 		cp->cache_hash_table = vmem_alloc(umem_hash_arena,
2631 		    UMEM_HASH_INITIAL * sizeof (void *), VM_NOSLEEP);
2632 		if (cp->cache_hash_table == NULL) {
2633 			errno = EAGAIN;
2634 			goto fail_lock;
2635 		}
2636 		bzero(cp->cache_hash_table,
2637 		    UMEM_HASH_INITIAL * sizeof (void *));
2638 		cp->cache_hash_mask = UMEM_HASH_INITIAL - 1;
2639 		cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
2640 	}
2641 
2642 	/*
2643 	 * Initialize the depot.
2644 	 */
2645 	(void) mutex_init(&cp->cache_depot_lock, USYNC_THREAD, NULL);
2646 
2647 	for (mtp = umem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
2648 		continue;
2649 
2650 	cp->cache_magtype = mtp;
2651 
2652 	/*
2653 	 * Initialize the CPU layer.
2654 	 */
2655 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
2656 		umem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
2657 		(void) mutex_init(&ccp->cc_lock, USYNC_THREAD, NULL);
2658 		ccp->cc_flags = cp->cache_flags;
2659 		ccp->cc_rounds = -1;
2660 		ccp->cc_prounds = -1;
2661 	}
2662 
2663 	/*
2664 	 * Add the cache to the global list.  This makes it visible
2665 	 * to umem_update(), so the cache must be ready for business.
2666 	 */
2667 	(void) mutex_lock(&umem_cache_lock);
2668 	cp->cache_next = cnext = &umem_null_cache;
2669 	cp->cache_prev = cprev = umem_null_cache.cache_prev;
2670 	cnext->cache_prev = cp;
2671 	cprev->cache_next = cp;
2672 	(void) mutex_unlock(&umem_cache_lock);
2673 
2674 	if (umem_ready == UMEM_READY)
2675 		umem_cache_magazine_enable(cp);
2676 
2677 	return (cp);
2678 
2679 fail_lock:
2680 	(void) mutex_destroy(&cp->cache_lock);
2681 fail:
2682 	vmem_xfree(umem_cache_arena, cp, csize);
2683 	return (NULL);
2684 }
2685 
2686 void
2687 umem_cache_destroy(umem_cache_t *cp)
2688 {
2689 	int cpu_seqid;
2690 
2691 	/*
2692 	 * Remove the cache from the global cache list so that no new updates
2693 	 * will be scheduled on its behalf, wait for any pending tasks to
2694 	 * complete, purge the cache, and then destroy it.
2695 	 */
2696 	(void) mutex_lock(&umem_cache_lock);
2697 	cp->cache_prev->cache_next = cp->cache_next;
2698 	cp->cache_next->cache_prev = cp->cache_prev;
2699 	cp->cache_prev = cp->cache_next = NULL;
2700 	(void) mutex_unlock(&umem_cache_lock);
2701 
2702 	umem_remove_updates(cp);
2703 
2704 	umem_cache_magazine_purge(cp);
2705 
2706 	(void) mutex_lock(&cp->cache_lock);
2707 	if (cp->cache_buftotal != 0)
2708 		log_message("umem_cache_destroy: '%s' (%p) not empty\n",
2709 		    cp->cache_name, (void *)cp);
2710 	cp->cache_reclaim = NULL;
2711 	/*
2712 	 * The cache is now dead.  There should be no further activity.
2713 	 * We enforce this by setting land mines in the constructor and
2714 	 * destructor routines that induce a segmentation fault if invoked.
2715 	 */
2716 	cp->cache_constructor = (umem_constructor_t *)1;
2717 	cp->cache_destructor = (umem_destructor_t *)2;
2718 	(void) mutex_unlock(&cp->cache_lock);
2719 
2720 	if (cp->cache_hash_table != NULL)
2721 		vmem_free(umem_hash_arena, cp->cache_hash_table,
2722 		    (cp->cache_hash_mask + 1) * sizeof (void *));
2723 
2724 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++)
2725 		(void) mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
2726 
2727 	(void) mutex_destroy(&cp->cache_depot_lock);
2728 	(void) mutex_destroy(&cp->cache_lock);
2729 
2730 	vmem_free(umem_cache_arena, cp, UMEM_CACHE_SIZE(umem_max_ncpus));
2731 }
2732 
2733 void
2734 umem_alloc_sizes_clear(void)
2735 {
2736 	int i;
2737 
2738 	umem_alloc_sizes[0] = UMEM_MAXBUF;
2739 	for (i = 1; i < NUM_ALLOC_SIZES; i++)
2740 		umem_alloc_sizes[i] = 0;
2741 }
2742 
2743 void
2744 umem_alloc_sizes_add(size_t size_arg)
2745 {
2746 	int i, j;
2747 	size_t size = size_arg;
2748 
2749 	if (size == 0) {
2750 		log_message("size_add: cannot add zero-sized cache\n",
2751 		    size, UMEM_MAXBUF);
2752 		return;
2753 	}
2754 
2755 	if (size > UMEM_MAXBUF) {
2756 		log_message("size_add: %ld > %d, cannot add\n", size,
2757 		    UMEM_MAXBUF);
2758 		return;
2759 	}
2760 
2761 	if (umem_alloc_sizes[NUM_ALLOC_SIZES - 1] != 0) {
2762 		log_message("size_add: no space in alloc_table for %d\n",
2763 		    size);
2764 		return;
2765 	}
2766 
2767 	if (P2PHASE(size, UMEM_ALIGN) != 0) {
2768 		size = P2ROUNDUP(size, UMEM_ALIGN);
2769 		log_message("size_add: rounding %d up to %d\n", size_arg,
2770 		    size);
2771 	}
2772 
2773 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
2774 		int cur = umem_alloc_sizes[i];
2775 		if (cur == size) {
2776 			log_message("size_add: %ld already in table\n",
2777 			    size);
2778 			return;
2779 		}
2780 		if (cur > size)
2781 			break;
2782 	}
2783 
2784 	for (j = NUM_ALLOC_SIZES - 1; j > i; j--)
2785 		umem_alloc_sizes[j] = umem_alloc_sizes[j-1];
2786 	umem_alloc_sizes[i] = size;
2787 }
2788 
2789 void
2790 umem_alloc_sizes_remove(size_t size)
2791 {
2792 	int i;
2793 
2794 	if (size == UMEM_MAXBUF) {
2795 		log_message("size_remove: cannot remove %ld\n", size);
2796 		return;
2797 	}
2798 
2799 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
2800 		int cur = umem_alloc_sizes[i];
2801 		if (cur == size)
2802 			break;
2803 		else if (cur > size || cur == 0) {
2804 			log_message("size_remove: %ld not found in table\n",
2805 			    size);
2806 			return;
2807 		}
2808 	}
2809 
2810 	for (; i + 1 < NUM_ALLOC_SIZES; i++)
2811 		umem_alloc_sizes[i] = umem_alloc_sizes[i+1];
2812 	umem_alloc_sizes[i] = 0;
2813 }
2814 
2815 static int
2816 umem_cache_init(void)
2817 {
2818 	int i;
2819 	size_t size, max_size;
2820 	umem_cache_t *cp;
2821 	umem_magtype_t *mtp;
2822 	char name[UMEM_CACHE_NAMELEN + 1];
2823 	umem_cache_t *umem_alloc_caches[NUM_ALLOC_SIZES];
2824 
2825 	for (i = 0; i < sizeof (umem_magtype) / sizeof (*mtp); i++) {
2826 		mtp = &umem_magtype[i];
2827 		(void) snprintf(name, sizeof (name), "umem_magazine_%d",
2828 		    mtp->mt_magsize);
2829 		mtp->mt_cache = umem_cache_create(name,
2830 		    (mtp->mt_magsize + 1) * sizeof (void *),
2831 		    mtp->mt_align, NULL, NULL, NULL, NULL,
2832 		    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
2833 		if (mtp->mt_cache == NULL)
2834 			return (0);
2835 	}
2836 
2837 	umem_slab_cache = umem_cache_create("umem_slab_cache",
2838 	    sizeof (umem_slab_t), 0, NULL, NULL, NULL, NULL,
2839 	    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
2840 
2841 	if (umem_slab_cache == NULL)
2842 		return (0);
2843 
2844 	umem_bufctl_cache = umem_cache_create("umem_bufctl_cache",
2845 	    sizeof (umem_bufctl_t), 0, NULL, NULL, NULL, NULL,
2846 	    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
2847 
2848 	if (umem_bufctl_cache == NULL)
2849 		return (0);
2850 
2851 	/*
2852 	 * The size of the umem_bufctl_audit structure depends upon
2853 	 * umem_stack_depth.   See umem_impl.h for details on the size
2854 	 * restrictions.
2855 	 */
2856 
2857 	size = UMEM_BUFCTL_AUDIT_SIZE_DEPTH(umem_stack_depth);
2858 	max_size = UMEM_BUFCTL_AUDIT_MAX_SIZE;
2859 
2860 	if (size > max_size) {			/* too large -- truncate */
2861 		int max_frames = UMEM_MAX_STACK_DEPTH;
2862 
2863 		ASSERT(UMEM_BUFCTL_AUDIT_SIZE_DEPTH(max_frames) <= max_size);
2864 
2865 		umem_stack_depth = max_frames;
2866 		size = UMEM_BUFCTL_AUDIT_SIZE_DEPTH(umem_stack_depth);
2867 	}
2868 
2869 	umem_bufctl_audit_cache = umem_cache_create("umem_bufctl_audit_cache",
2870 	    size, 0, NULL, NULL, NULL, NULL, umem_internal_arena,
2871 	    UMC_NOHASH | UMC_INTERNAL);
2872 
2873 	if (umem_bufctl_audit_cache == NULL)
2874 		return (0);
2875 
2876 	if (vmem_backend & VMEM_BACKEND_MMAP)
2877 		umem_va_arena = vmem_create("umem_va",
2878 		    NULL, 0, pagesize,
2879 		    vmem_alloc, vmem_free, heap_arena,
2880 		    8 * pagesize, VM_NOSLEEP);
2881 	else
2882 		umem_va_arena = heap_arena;
2883 
2884 	if (umem_va_arena == NULL)
2885 		return (0);
2886 
2887 	umem_default_arena = vmem_create("umem_default",
2888 	    NULL, 0, pagesize,
2889 	    heap_alloc, heap_free, umem_va_arena,
2890 	    0, VM_NOSLEEP);
2891 
2892 	if (umem_default_arena == NULL)
2893 		return (0);
2894 
2895 	/*
2896 	 * make sure the umem_alloc table initializer is correct
2897 	 */
2898 	i = sizeof (umem_alloc_table) / sizeof (*umem_alloc_table);
2899 	ASSERT(umem_alloc_table[i - 1] == &umem_null_cache);
2900 
2901 	/*
2902 	 * Create the default caches to back umem_alloc()
2903 	 */
2904 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
2905 		size_t cache_size = umem_alloc_sizes[i];
2906 		size_t align = 0;
2907 
2908 		if (cache_size == 0)
2909 			break;		/* 0 terminates the list */
2910 
2911 		/*
2912 		 * If they allocate a multiple of the coherency granularity,
2913 		 * they get a coherency-granularity-aligned address.
2914 		 */
2915 		if (IS_P2ALIGNED(cache_size, 64))
2916 			align = 64;
2917 		if (IS_P2ALIGNED(cache_size, pagesize))
2918 			align = pagesize;
2919 		(void) snprintf(name, sizeof (name), "umem_alloc_%lu",
2920 		    (long)cache_size);
2921 
2922 		cp = umem_cache_create(name, cache_size, align,
2923 		    NULL, NULL, NULL, NULL, NULL, UMC_INTERNAL);
2924 		if (cp == NULL)
2925 			return (0);
2926 
2927 		umem_alloc_caches[i] = cp;
2928 	}
2929 
2930 	/*
2931 	 * Initialization cannot fail at this point.  Make the caches
2932 	 * visible to umem_alloc() and friends.
2933 	 */
2934 	size = UMEM_ALIGN;
2935 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
2936 		size_t cache_size = umem_alloc_sizes[i];
2937 
2938 		if (cache_size == 0)
2939 			break;		/* 0 terminates the list */
2940 
2941 		cp = umem_alloc_caches[i];
2942 
2943 		while (size <= cache_size) {
2944 			umem_alloc_table[(size - 1) >> UMEM_ALIGN_SHIFT] = cp;
2945 			size += UMEM_ALIGN;
2946 		}
2947 	}
2948 	ASSERT(size - UMEM_ALIGN == UMEM_MAXBUF);
2949 	return (1);
2950 }
2951 
2952 /*
2953  * umem_startup() is called early on, and must be called explicitly if we're
2954  * the standalone version.
2955  */
2956 #ifdef UMEM_STANDALONE
2957 void
2958 #else
2959 #pragma init(umem_startup)
2960 static void
2961 #endif
2962 umem_startup(caddr_t start, size_t len, size_t pagesize, caddr_t minstack,
2963     caddr_t maxstack)
2964 {
2965 #ifdef UMEM_STANDALONE
2966 	int idx;
2967 	/* Standalone doesn't fork */
2968 #else
2969 	umem_forkhandler_init(); /* register the fork handler */
2970 #endif
2971 
2972 #ifdef __lint
2973 	/* make lint happy */
2974 	minstack = maxstack;
2975 #endif
2976 
2977 #ifdef UMEM_STANDALONE
2978 	umem_ready = UMEM_READY_STARTUP;
2979 	umem_init_env_ready = 0;
2980 
2981 	umem_min_stack = minstack;
2982 	umem_max_stack = maxstack;
2983 
2984 	nofail_callback = NULL;
2985 	umem_slab_cache = NULL;
2986 	umem_bufctl_cache = NULL;
2987 	umem_bufctl_audit_cache = NULL;
2988 	heap_arena = NULL;
2989 	heap_alloc = NULL;
2990 	heap_free = NULL;
2991 	umem_internal_arena = NULL;
2992 	umem_cache_arena = NULL;
2993 	umem_hash_arena = NULL;
2994 	umem_log_arena = NULL;
2995 	umem_oversize_arena = NULL;
2996 	umem_va_arena = NULL;
2997 	umem_default_arena = NULL;
2998 	umem_firewall_va_arena = NULL;
2999 	umem_firewall_arena = NULL;
3000 	umem_memalign_arena = NULL;
3001 	umem_transaction_log = NULL;
3002 	umem_content_log = NULL;
3003 	umem_failure_log = NULL;
3004 	umem_slab_log = NULL;
3005 	umem_cpu_mask = 0;
3006 
3007 	umem_cpus = &umem_startup_cpu;
3008 	umem_startup_cpu.cpu_cache_offset = UMEM_CACHE_SIZE(0);
3009 	umem_startup_cpu.cpu_number = 0;
3010 
3011 	bcopy(&umem_null_cache_template, &umem_null_cache,
3012 	    sizeof (umem_cache_t));
3013 
3014 	for (idx = 0; idx < (UMEM_MAXBUF >> UMEM_ALIGN_SHIFT); idx++)
3015 		umem_alloc_table[idx] = &umem_null_cache;
3016 #endif
3017 
3018 	/*
3019 	 * Perform initialization specific to the way we've been compiled
3020 	 * (library or standalone)
3021 	 */
3022 	umem_type_init(start, len, pagesize);
3023 
3024 	vmem_startup();
3025 }
3026 
3027 int
3028 umem_init(void)
3029 {
3030 	size_t maxverify, minfirewall;
3031 	size_t size;
3032 	int idx;
3033 	umem_cpu_t *new_cpus;
3034 
3035 	vmem_t *memalign_arena, *oversize_arena;
3036 
3037 	if (thr_self() != umem_init_thr) {
3038 		/*
3039 		 * The usual case -- non-recursive invocation of umem_init().
3040 		 */
3041 		(void) mutex_lock(&umem_init_lock);
3042 		if (umem_ready != UMEM_READY_STARTUP) {
3043 			/*
3044 			 * someone else beat us to initializing umem.  Wait
3045 			 * for them to complete, then return.
3046 			 */
3047 			while (umem_ready == UMEM_READY_INITING) {
3048 				int cancel_state;
3049 
3050 				(void) pthread_setcancelstate(
3051 				    PTHREAD_CANCEL_DISABLE, &cancel_state);
3052 				(void) cond_wait(&umem_init_cv,
3053 				    &umem_init_lock);
3054 				(void) pthread_setcancelstate(
3055 				    cancel_state, NULL);
3056 			}
3057 			ASSERT(umem_ready == UMEM_READY ||
3058 			    umem_ready == UMEM_READY_INIT_FAILED);
3059 			(void) mutex_unlock(&umem_init_lock);
3060 			return (umem_ready == UMEM_READY);
3061 		}
3062 
3063 		ASSERT(umem_ready == UMEM_READY_STARTUP);
3064 		ASSERT(umem_init_env_ready == 0);
3065 
3066 		umem_ready = UMEM_READY_INITING;
3067 		umem_init_thr = thr_self();
3068 
3069 		(void) mutex_unlock(&umem_init_lock);
3070 		umem_setup_envvars(0);		/* can recurse -- see below */
3071 		if (umem_init_env_ready) {
3072 			/*
3073 			 * initialization was completed already
3074 			 */
3075 			ASSERT(umem_ready == UMEM_READY ||
3076 			    umem_ready == UMEM_READY_INIT_FAILED);
3077 			ASSERT(umem_init_thr == 0);
3078 			return (umem_ready == UMEM_READY);
3079 		}
3080 	} else if (!umem_init_env_ready) {
3081 		/*
3082 		 * The umem_setup_envvars() call (above) makes calls into
3083 		 * the dynamic linker and directly into user-supplied code.
3084 		 * Since we cannot know what that code will do, we could be
3085 		 * recursively invoked (by, say, a malloc() call in the code
3086 		 * itself, or in a (C++) _init section it causes to be fired).
3087 		 *
3088 		 * This code is where we end up if such recursion occurs.  We
3089 		 * first clean up any partial results in the envvar code, then
3090 		 * proceed to finish initialization processing in the recursive
3091 		 * call.  The original call will notice this, and return
3092 		 * immediately.
3093 		 */
3094 		umem_setup_envvars(1);		/* clean up any partial state */
3095 	} else {
3096 		umem_panic(
3097 		    "recursive allocation while initializing umem\n");
3098 	}
3099 	umem_init_env_ready = 1;
3100 
3101 	/*
3102 	 * From this point until we finish, recursion into umem_init() will
3103 	 * cause a umem_panic().
3104 	 */
3105 	maxverify = minfirewall = ULONG_MAX;
3106 
3107 	/* LINTED constant condition */
3108 	if (sizeof (umem_cpu_cache_t) != UMEM_CPU_CACHE_SIZE) {
3109 		umem_panic("sizeof (umem_cpu_cache_t) = %d, should be %d\n",
3110 		    sizeof (umem_cpu_cache_t), UMEM_CPU_CACHE_SIZE);
3111 	}
3112 
3113 	umem_max_ncpus = umem_get_max_ncpus();
3114 
3115 	/*
3116 	 * load tunables from environment
3117 	 */
3118 	umem_process_envvars();
3119 
3120 	if (issetugid())
3121 		umem_mtbf = 0;
3122 
3123 	/*
3124 	 * set up vmem
3125 	 */
3126 	if (!(umem_flags & UMF_AUDIT))
3127 		vmem_no_debug();
3128 
3129 	heap_arena = vmem_heap_arena(&heap_alloc, &heap_free);
3130 
3131 	pagesize = heap_arena->vm_quantum;
3132 
3133 	umem_internal_arena = vmem_create("umem_internal", NULL, 0, pagesize,
3134 	    heap_alloc, heap_free, heap_arena, 0, VM_NOSLEEP);
3135 
3136 	umem_default_arena = umem_internal_arena;
3137 
3138 	if (umem_internal_arena == NULL)
3139 		goto fail;
3140 
3141 	umem_cache_arena = vmem_create("umem_cache", NULL, 0, UMEM_ALIGN,
3142 	    vmem_alloc, vmem_free, umem_internal_arena, 0, VM_NOSLEEP);
3143 
3144 	umem_hash_arena = vmem_create("umem_hash", NULL, 0, UMEM_ALIGN,
3145 	    vmem_alloc, vmem_free, umem_internal_arena, 0, VM_NOSLEEP);
3146 
3147 	umem_log_arena = vmem_create("umem_log", NULL, 0, UMEM_ALIGN,
3148 	    heap_alloc, heap_free, heap_arena, 0, VM_NOSLEEP);
3149 
3150 	umem_firewall_va_arena = vmem_create("umem_firewall_va",
3151 	    NULL, 0, pagesize,
3152 	    umem_firewall_va_alloc, umem_firewall_va_free, heap_arena,
3153 	    0, VM_NOSLEEP);
3154 
3155 	if (umem_cache_arena == NULL || umem_hash_arena == NULL ||
3156 	    umem_log_arena == NULL || umem_firewall_va_arena == NULL)
3157 		goto fail;
3158 
3159 	umem_firewall_arena = vmem_create("umem_firewall", NULL, 0, pagesize,
3160 	    heap_alloc, heap_free, umem_firewall_va_arena, 0,
3161 	    VM_NOSLEEP);
3162 
3163 	if (umem_firewall_arena == NULL)
3164 		goto fail;
3165 
3166 	oversize_arena = vmem_create("umem_oversize", NULL, 0, pagesize,
3167 	    heap_alloc, heap_free, minfirewall < ULONG_MAX ?
3168 	    umem_firewall_va_arena : heap_arena, 0, VM_NOSLEEP);
3169 
3170 	memalign_arena = vmem_create("umem_memalign", NULL, 0, UMEM_ALIGN,
3171 	    heap_alloc, heap_free, minfirewall < ULONG_MAX ?
3172 	    umem_firewall_va_arena : heap_arena, 0, VM_NOSLEEP);
3173 
3174 	if (oversize_arena == NULL || memalign_arena == NULL)
3175 		goto fail;
3176 
3177 	if (umem_max_ncpus > CPUHINT_MAX())
3178 		umem_max_ncpus = CPUHINT_MAX();
3179 
3180 	while ((umem_max_ncpus & (umem_max_ncpus - 1)) != 0)
3181 		umem_max_ncpus++;
3182 
3183 	if (umem_max_ncpus == 0)
3184 		umem_max_ncpus = 1;
3185 
3186 	size = umem_max_ncpus * sizeof (umem_cpu_t);
3187 	new_cpus = vmem_alloc(umem_internal_arena, size, VM_NOSLEEP);
3188 	if (new_cpus == NULL)
3189 		goto fail;
3190 
3191 	bzero(new_cpus, size);
3192 	for (idx = 0; idx < umem_max_ncpus; idx++) {
3193 		new_cpus[idx].cpu_number = idx;
3194 		new_cpus[idx].cpu_cache_offset = UMEM_CACHE_SIZE(idx);
3195 	}
3196 	umem_cpus = new_cpus;
3197 	umem_cpu_mask = (umem_max_ncpus - 1);
3198 
3199 	if (umem_maxverify == 0)
3200 		umem_maxverify = maxverify;
3201 
3202 	if (umem_minfirewall == 0)
3203 		umem_minfirewall = minfirewall;
3204 
3205 	/*
3206 	 * Set up updating and reaping
3207 	 */
3208 	umem_reap_next = gethrtime() + NANOSEC;
3209 
3210 #ifndef UMEM_STANDALONE
3211 	(void) gettimeofday(&umem_update_next, NULL);
3212 #endif
3213 
3214 	/*
3215 	 * Set up logging -- failure here is okay, since it will just disable
3216 	 * the logs
3217 	 */
3218 	if (umem_logging) {
3219 		umem_transaction_log = umem_log_init(umem_transaction_log_size);
3220 		umem_content_log = umem_log_init(umem_content_log_size);
3221 		umem_failure_log = umem_log_init(umem_failure_log_size);
3222 		umem_slab_log = umem_log_init(umem_slab_log_size);
3223 	}
3224 
3225 	/*
3226 	 * Set up caches -- if successful, initialization cannot fail, since
3227 	 * allocations from other threads can now succeed.
3228 	 */
3229 	if (umem_cache_init() == 0) {
3230 		log_message("unable to create initial caches\n");
3231 		goto fail;
3232 	}
3233 	umem_oversize_arena = oversize_arena;
3234 	umem_memalign_arena = memalign_arena;
3235 
3236 	umem_cache_applyall(umem_cache_magazine_enable);
3237 
3238 	/*
3239 	 * initialization done, ready to go
3240 	 */
3241 	(void) mutex_lock(&umem_init_lock);
3242 	umem_ready = UMEM_READY;
3243 	umem_init_thr = 0;
3244 	(void) cond_broadcast(&umem_init_cv);
3245 	(void) mutex_unlock(&umem_init_lock);
3246 	return (1);
3247 
3248 fail:
3249 	log_message("umem initialization failed\n");
3250 
3251 	(void) mutex_lock(&umem_init_lock);
3252 	umem_ready = UMEM_READY_INIT_FAILED;
3253 	umem_init_thr = 0;
3254 	(void) cond_broadcast(&umem_init_cv);
3255 	(void) mutex_unlock(&umem_init_lock);
3256 	return (0);
3257 }
3258