xref: /illumos-gate/usr/src/lib/libumem/common/umem.c (revision 4f364e7c)
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 /*
28  * Copyright (c) 2014 Joyent, Inc.  All rights reserved.
29  */
30 
31 /*
32  * based on usr/src/uts/common/os/kmem.c r1.64 from 2001/12/18
33  *
34  * The slab allocator, as described in the following two papers:
35  *
36  *	Jeff Bonwick,
37  *	The Slab Allocator: An Object-Caching Kernel Memory Allocator.
38  *	Proceedings of the Summer 1994 Usenix Conference.
39  *	Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
40  *
41  *	Jeff Bonwick and Jonathan Adams,
42  *	Magazines and vmem: Extending the Slab Allocator to Many CPUs and
43  *	Arbitrary Resources.
44  *	Proceedings of the 2001 Usenix Conference.
45  *	Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
46  *
47  * 1. Overview
48  * -----------
49  * umem is very close to kmem in implementation.  There are seven major
50  * areas of divergence:
51  *
52  *	* Initialization
53  *
54  *	* CPU handling
55  *
56  *	* umem_update()
57  *
58  *	* KM_SLEEP v.s. UMEM_NOFAIL
59  *
60  *	* lock ordering
61  *
62  *	* changing UMEM_MAXBUF
63  *
64  *	* Per-thread caching for malloc/free
65  *
66  * 2. Initialization
67  * -----------------
68  * kmem is initialized early on in boot, and knows that no one will call
69  * into it before it is ready.  umem does not have these luxuries. Instead,
70  * initialization is divided into two phases:
71  *
72  *	* library initialization, and
73  *
74  *	* first use
75  *
76  * umem's full initialization happens at the time of the first allocation
77  * request (via malloc() and friends, umem_alloc(), or umem_zalloc()),
78  * or the first call to umem_cache_create().
79  *
80  * umem_free(), and umem_cache_alloc() do not require special handling,
81  * since the only way to get valid arguments for them is to successfully
82  * call a function from the first group.
83  *
84  * 2.1. Library Initialization: umem_startup()
85  * -------------------------------------------
86  * umem_startup() is libumem.so's .init section.  It calls pthread_atfork()
87  * to install the handlers necessary for umem's Fork1-Safety.  Because of
88  * race condition issues, all other pre-umem_init() initialization is done
89  * statically (i.e. by the dynamic linker).
90  *
91  * For standalone use, umem_startup() returns everything to its initial
92  * state.
93  *
94  * 2.2. First use: umem_init()
95  * ------------------------------
96  * The first time any memory allocation function is used, we have to
97  * create the backing caches and vmem arenas which are needed for it.
98  * umem_init() is the central point for that task.  When it completes,
99  * umem_ready is either UMEM_READY (all set) or UMEM_READY_INIT_FAILED (unable
100  * to initialize, probably due to lack of memory).
101  *
102  * There are four different paths from which umem_init() is called:
103  *
104  *	* from umem_alloc() or umem_zalloc(), with 0 < size < UMEM_MAXBUF,
105  *
106  *	* from umem_alloc() or umem_zalloc(), with size > UMEM_MAXBUF,
107  *
108  *	* from umem_cache_create(), and
109  *
110  *	* from memalign(), with align > UMEM_ALIGN.
111  *
112  * The last three just check if umem is initialized, and call umem_init()
113  * if it is not.  For performance reasons, the first case is more complicated.
114  *
115  * 2.2.1. umem_alloc()/umem_zalloc(), with 0 < size < UMEM_MAXBUF
116  * -----------------------------------------------------------------
117  * In this case, umem_cache_alloc(&umem_null_cache, ...) is called.
118  * There is special case code in which causes any allocation on
119  * &umem_null_cache to fail by returning (NULL), regardless of the
120  * flags argument.
121  *
122  * So umem_cache_alloc() returns NULL, and umem_alloc()/umem_zalloc() call
123  * umem_alloc_retry().  umem_alloc_retry() sees that the allocation
124  * was agains &umem_null_cache, and calls umem_init().
125  *
126  * If initialization is successful, umem_alloc_retry() returns 1, which
127  * causes umem_alloc()/umem_zalloc() to start over, which causes it to load
128  * the (now valid) cache pointer from umem_alloc_table.
129  *
130  * 2.2.2. Dealing with race conditions
131  * -----------------------------------
132  * There are a couple race conditions resulting from the initialization
133  * code that we have to guard against:
134  *
135  *	* In umem_cache_create(), there is a special UMC_INTERNAL cflag
136  *	that is passed for caches created during initialization.  It
137  *	is illegal for a user to try to create a UMC_INTERNAL cache.
138  *	This allows initialization to proceed, but any other
139  *	umem_cache_create()s will block by calling umem_init().
140  *
141  *	* Since umem_null_cache has a 1-element cache_cpu, it's cache_cpu_mask
142  *	is always zero.  umem_cache_alloc uses cp->cache_cpu_mask to
143  *	mask the cpu number.  This prevents a race between grabbing a
144  *	cache pointer out of umem_alloc_table and growing the cpu array.
145  *
146  *
147  * 3. CPU handling
148  * ---------------
149  * kmem uses the CPU's sequence number to determine which "cpu cache" to
150  * use for an allocation.  Currently, there is no way to get the sequence
151  * number in userspace.
152  *
153  * umem keeps track of cpu information in umem_cpus, an array of umem_max_ncpus
154  * umem_cpu_t structures.  CURCPU() is a a "hint" function, which we then mask
155  * with either umem_cpu_mask or cp->cache_cpu_mask to find the actual "cpu" id.
156  * The mechanics of this is all in the CPU(mask) macro.
157  *
158  * Currently, umem uses _lwp_self() as its hint.
159  *
160  *
161  * 4. The update thread
162  * --------------------
163  * kmem uses a task queue, kmem_taskq, to do periodic maintenance on
164  * every kmem cache.  vmem has a periodic timeout for hash table resizing.
165  * The kmem_taskq also provides a separate context for kmem_cache_reap()'s
166  * to be done in, avoiding issues of the context of kmem_reap() callers.
167  *
168  * Instead, umem has the concept of "updates", which are asynchronous requests
169  * for work attached to single caches.  All caches with pending work are
170  * on a doubly linked list rooted at the umem_null_cache.  All update state
171  * is protected by the umem_update_lock mutex, and the umem_update_cv is used
172  * for notification between threads.
173  *
174  * 4.1. Cache states with regards to updates
175  * -----------------------------------------
176  * A given cache is in one of three states:
177  *
178  * Inactive		cache_uflags is zero, cache_u{next,prev} are NULL
179  *
180  * Work Requested	cache_uflags is non-zero (but UMU_ACTIVE is not set),
181  *			cache_u{next,prev} link the cache onto the global
182  *			update list
183  *
184  * Active		cache_uflags has UMU_ACTIVE set, cache_u{next,prev}
185  *			are NULL, and either umem_update_thr or
186  *			umem_st_update_thr are actively doing work on the
187  *			cache.
188  *
189  * An update can be added to any cache in any state -- if the cache is
190  * Inactive, it transitions to being Work Requested.  If the cache is
191  * Active, the worker will notice the new update and act on it before
192  * transitioning the cache to the Inactive state.
193  *
194  * If a cache is in the Active state, UMU_NOTIFY can be set, which asks
195  * the worker to broadcast the umem_update_cv when it has finished.
196  *
197  * 4.2. Update interface
198  * ---------------------
199  * umem_add_update() adds an update to a particular cache.
200  * umem_updateall() adds an update to all caches.
201  * umem_remove_updates() returns a cache to the Inactive state.
202  *
203  * umem_process_updates() process all caches in the Work Requested state.
204  *
205  * 4.3. Reaping
206  * ------------
207  * When umem_reap() is called (at the time of heap growth), it schedule
208  * UMU_REAP updates on every cache.  It then checks to see if the update
209  * thread exists (umem_update_thr != 0).  If it is, it broadcasts
210  * the umem_update_cv to wake the update thread up, and returns.
211  *
212  * If the update thread does not exist (umem_update_thr == 0), and the
213  * program currently has multiple threads, umem_reap() attempts to create
214  * a new update thread.
215  *
216  * If the process is not multithreaded, or the creation fails, umem_reap()
217  * calls umem_st_update() to do an inline update.
218  *
219  * 4.4. The update thread
220  * ----------------------
221  * The update thread spends most of its time in cond_timedwait() on the
222  * umem_update_cv.  It wakes up under two conditions:
223  *
224  *	* The timedwait times out, in which case it needs to run a global
225  *	update, or
226  *
227  *	* someone cond_broadcast(3THR)s the umem_update_cv, in which case
228  *	it needs to check if there are any caches in the Work Requested
229  *	state.
230  *
231  * When it is time for another global update, umem calls umem_cache_update()
232  * on every cache, then calls vmem_update(), which tunes the vmem structures.
233  * umem_cache_update() can request further work using umem_add_update().
234  *
235  * After any work from the global update completes, the update timer is
236  * reset to umem_reap_interval seconds in the future.  This makes the
237  * updates self-throttling.
238  *
239  * Reaps are similarly self-throttling.  After a UMU_REAP update has
240  * been scheduled on all caches, umem_reap() sets a flag and wakes up the
241  * update thread.  The update thread notices the flag, and resets the
242  * reap state.
243  *
244  * 4.5. Inline updates
245  * -------------------
246  * If the update thread is not running, umem_st_update() is used instead.  It
247  * immediately does a global update (as above), then calls
248  * umem_process_updates() to process both the reaps that umem_reap() added and
249  * any work generated by the global update.  Afterwards, it resets the reap
250  * state.
251  *
252  * While the umem_st_update() is running, umem_st_update_thr holds the thread
253  * id of the thread performing the update.
254  *
255  * 4.6. Updates and fork1()
256  * ------------------------
257  * umem has fork1() pre- and post-handlers which lock up (and release) every
258  * mutex in every cache.  They also lock up the umem_update_lock.  Since
259  * fork1() only copies over a single lwp, other threads (including the update
260  * thread) could have been actively using a cache in the parent.  This
261  * can lead to inconsistencies in the child process.
262  *
263  * Because we locked all of the mutexes, the only possible inconsistancies are:
264  *
265  *	* a umem_cache_alloc() could leak its buffer.
266  *
267  *	* a caller of umem_depot_alloc() could leak a magazine, and all the
268  *	buffers contained in it.
269  *
270  *	* a cache could be in the Active update state.  In the child, there
271  *	would be no thread actually working on it.
272  *
273  *	* a umem_hash_rescale() could leak the new hash table.
274  *
275  *	* a umem_magazine_resize() could be in progress.
276  *
277  *	* a umem_reap() could be in progress.
278  *
279  * The memory leaks we can't do anything about.  umem_release_child() resets
280  * the update state, moves any caches in the Active state to the Work Requested
281  * state.  This might cause some updates to be re-run, but UMU_REAP and
282  * UMU_HASH_RESCALE are effectively idempotent, and the worst that can
283  * happen from umem_magazine_resize() is resizing the magazine twice in close
284  * succession.
285  *
286  * Much of the cleanup in umem_release_child() is skipped if
287  * umem_st_update_thr == thr_self().  This is so that applications which call
288  * fork1() from a cache callback does not break.  Needless to say, any such
289  * application is tremendously broken.
290  *
291  *
292  * 5. KM_SLEEP v.s. UMEM_NOFAIL
293  * ----------------------------
294  * Allocations against kmem and vmem have two basic modes:  SLEEP and
295  * NOSLEEP.  A sleeping allocation is will go to sleep (waiting for
296  * more memory) instead of failing (returning NULL).
297  *
298  * SLEEP allocations presume an extremely multithreaded model, with
299  * a lot of allocation and deallocation activity.  umem cannot presume
300  * that its clients have any particular type of behavior.  Instead,
301  * it provides two types of allocations:
302  *
303  *	* UMEM_DEFAULT, equivalent to KM_NOSLEEP (i.e. return NULL on
304  *	failure)
305  *
306  *	* UMEM_NOFAIL, which, on failure, calls an optional callback
307  *	(registered with umem_nofail_callback()).
308  *
309  * The callback is invoked with no locks held, and can do an arbitrary
310  * amount of work.  It then has a choice between:
311  *
312  *	* Returning UMEM_CALLBACK_RETRY, which will cause the allocation
313  *	to be restarted.
314  *
315  *	* Returning UMEM_CALLBACK_EXIT(status), which will cause exit(2)
316  *	to be invoked with status.  If multiple threads attempt to do
317  *	this simultaneously, only one will call exit(2).
318  *
319  *	* Doing some kind of non-local exit (thr_exit(3thr), longjmp(3C),
320  *	etc.)
321  *
322  * The default callback returns UMEM_CALLBACK_EXIT(255).
323  *
324  * To have these callbacks without risk of state corruption (in the case of
325  * a non-local exit), we have to ensure that the callbacks get invoked
326  * close to the original allocation, with no inconsistent state or held
327  * locks.  The following steps are taken:
328  *
329  *	* All invocations of vmem are VM_NOSLEEP.
330  *
331  *	* All constructor callbacks (which can themselves to allocations)
332  *	are passed UMEM_DEFAULT as their required allocation argument.  This
333  *	way, the constructor will fail, allowing the highest-level allocation
334  *	invoke the nofail callback.
335  *
336  *	If a constructor callback _does_ do a UMEM_NOFAIL allocation, and
337  *	the nofail callback does a non-local exit, we will leak the
338  *	partially-constructed buffer.
339  *
340  *
341  * 6. Lock Ordering
342  * ----------------
343  * umem has a few more locks than kmem does, mostly in the update path.  The
344  * overall lock ordering (earlier locks must be acquired first) is:
345  *
346  *	umem_init_lock
347  *
348  *	vmem_list_lock
349  *	vmem_nosleep_lock.vmpl_mutex
350  *	vmem_t's:
351  *		vm_lock
352  *	sbrk_lock
353  *
354  *	umem_cache_lock
355  *	umem_update_lock
356  *	umem_flags_lock
357  *	umem_cache_t's:
358  *		cache_cpu[*].cc_lock
359  *		cache_depot_lock
360  *		cache_lock
361  *	umem_log_header_t's:
362  *		lh_cpu[*].clh_lock
363  *		lh_lock
364  *
365  * 7. Changing UMEM_MAXBUF
366  * -----------------------
367  *
368  * When changing UMEM_MAXBUF extra care has to be taken. It is not sufficient to
369  * simply increase this number. First, one must update the umem_alloc_table to
370  * have the appropriate number of entires based upon the new size. If this is
371  * not done, this will lead to libumem blowing an assertion.
372  *
373  * The second place to update, which is not required, is the umem_alloc_sizes.
374  * These determine the default cache sizes that we're going to support.
375  *
376  * 8. Per-thread caching for malloc/free
377  * -------------------------------------
378  *
379  * "Time is an illusion. Lunchtime doubly so." -- Douglas Adams
380  *
381  * Time may be an illusion, but CPU cycles aren't.  While libumem is designed
382  * to be a highly scalable allocator, that scalability comes with a fixed cycle
383  * penalty even in the absence of contention: libumem must acquire (and release
384  * a per-CPU lock for each allocation.  When contention is low and malloc(3C)
385  * frequency is high, this overhead can dominate execution time.  To alleviate
386  * this, we allow for per-thread caching, a lock-free means of caching recent
387  * deallocations on a per-thread basis for use in satisfying subsequent calls
388  *
389  * In addition to improving performance, we also want to:
390  *	* Minimize fragmentation
391  *	* Not add additional memory overhead (no larger malloc tags)
392  *
393  * In the ulwp_t of each thread there is a private data structure called a
394  * umem_t that looks like:
395  *
396  * typedef struct {
397  * 	size_t	tm_size;
398  * 	void	*tm_roots[NTMEMBASE];  (Currently 16)
399  * } tmem_t;
400  *
401  * Each of the roots is treated as the head of a linked list. Each entry in the
402  * list can be thought of as a void ** which points to the next entry, until one
403  * of them points to NULL. If the head points to NULL, the list is empty.
404  *
405  * Each head corresponds to a umem_cache. Currently there is a linear mapping
406  * where the first root corresponds to the first cache, second root to the
407  * second cache, etc. This works because every allocation that malloc makes to
408  * umem_alloc that can be satisified by a umem_cache will actually return a
409  * number of bytes equal to the size of that cache. Because of this property and
410  * a one to one mapping between caches and roots we can guarantee that every
411  * entry in a given root's list will be able to satisfy the same requests as the
412  * corresponding cache.
413  *
414  * The choice of sixteen roots is based on where we believe we get the biggest
415  * bang for our buck. The per-thread caches will cache up to 256 byte and 448
416  * byte allocations on ILP32 and LP64 respectively. Generally applications plan
417  * more carefully how they do larger allocations than smaller ones. Therefore
418  * sixteen roots is a reasonable compromise between the amount of additional
419  * overhead per thread, and the likelihood of a program to benefit from it.
420  *
421  * The maximum amount of memory that can be cached in each thread is determined
422  * by the perthread_cache UMEM_OPTION. It corresponds to the umem_ptc_size
423  * value. The default value for this is currently 1 MB. Once umem_init() has
424  * finished this cannot be directly tuned without directly modifying the
425  * instruction text. If, upon calling free(3C), the amount cached would exceed
426  * this maximum, we instead actually return the buffer to the umem_cache instead
427  * of holding onto it in the thread.
428  *
429  * When a thread calls malloc(3C) it first determines which umem_cache it
430  * would be serviced by. If the allocation is not covered by ptcumem it goes to
431  * the normal malloc instead.  Next, it checks if the tmem_root's list is empty
432  * or not. If it is empty, we instead go and allocate the memory from
433  * umem_alloc. If it is not empty, we remove the head of the list, set the
434  * appropriate malloc tags, and return that buffer.
435  *
436  * When a thread calls free(3C) it first looks at the malloc tag and if it is
437  * invalid or the allocation exceeds the largest cache in ptcumem and sends it
438  * off to the original free() to handle and clean up appropriately. Next, it
439  * checks if the allocation size is covered by one of the per-thread roots and
440  * if it isn't, it passes it off to the original free() to be released. Finally,
441  * before it inserts this buffer as the head, it checks if adding this buffer
442  * would put the thread over its maximum cache size. If it would, it frees the
443  * buffer back to the umem_cache. Otherwise it increments the threads total
444  * cached amount and makes the buffer the new head of the appropriate tm_root.
445  *
446  * When a thread exits, all of the buffers that it has in its per-thread cache
447  * will be passed to umem_free() and returned to the appropriate umem_cache.
448  *
449  * 8.1 Handling addition and removal of umem_caches
450  * ------------------------------------------------
451  *
452  * The set of umem_caches that are used to back calls to umem_alloc() and
453  * ultimately malloc() are determined at program execution time. The default set
454  * of caches is defined below in umem_alloc_sizes[]. Various umem_options exist
455  * that modify the set of caches: size_add, size_clear, and size_remove. Because
456  * the set of caches can only be determined once umem_init() has been called and
457  * we have the additional goals of minimizing additional fragmentation and
458  * metadata space overhead in the malloc tags, this forces our hand to go down a
459  * slightly different path: the one tread by fasttrap and trapstat.
460  *
461  * During umem_init we're going to dynamically construct a new version of
462  * malloc(3C) and free(3C) that utilizes the known cache sizes and then ensure
463  * that ptcmalloc and ptcfree replace malloc and free as entries in the plt. If
464  * ptcmalloc and ptcfree cannot handle a request, they simply jump to the
465  * original libumem implementations.
466  *
467  * After creating all of the umem_caches, but before making them visible,
468  * umem_cache_init checks that umem_genasm_supported is non-zero. This value is
469  * set by each architecture in $ARCH/umem_genasm.c to indicate whether or not
470  * they support this. If the value is zero, then this process is skipped.
471  * Similarly, if the cache size has been tuned to zero by UMEM_OPTIONS, then
472  * this is also skipped.
473  *
474  * In umem_genasm.c, each architecture's implementation implements a single
475  * function called umem_genasm() that is responsible for generating the
476  * appropriate versions of ptcmalloc() and ptcfree(), placing them in the
477  * appropriate memory location, and finally doing the switch from malloc() and
478  * free() to ptcmalloc() and ptcfree().  Once the change has been made, there is
479  * no way to switch back, short of restarting the program or modifying program
480  * text with mdb.
481  *
482  * 8.2 Modifying the Procedure Linkage Table (PLT)
483  * -----------------------------------------------
484  *
485  * The last piece of this puzzle is how we actually jam ptcmalloc() into the
486  * PLT.  To handle this, we have defined two functions, _malloc and _free and
487  * used a special mapfile directive to place them into the a readable,
488  * writeable, and executable segment.  Next we use a standard #pragma weak for
489  * malloc and free and direct them to those symbols. By default, those symbols
490  * have text defined as nops for our generated functions and when they're
491  * invoked, they jump to the default malloc and free functions.
492  *
493  * When umem_genasm() is called, it goes through and generates new malloc() and
494  * free() functions in the text provided for by _malloc and _free just after the
495  * jump. Once both have been successfully generated, umem_genasm() nops over the
496  * original jump so that we now call into the genasm versions of these
497  * functions.
498  *
499  * 8.3 umem_genasm()
500  * -----------------
501  *
502  * umem_genasm() is currently implemented for i386 and amd64. This section
503  * describes the theory behind the construction. For specific byte code to
504  * assembly instructions and niceish C and asm versions of ptcmalloc and
505  * ptcfree, see the individual umem_genasm.c files. The layout consists of the
506  * following sections:
507  *
508  *	o. function-specfic prologue
509  *	o. function-generic cache-selecting elements
510  *	o. function-specific epilogue
511  *
512  * There are three different generic cache elements that exist:
513  *
514  *	o. the last or only cache
515  *	o. the intermediary caches if more than two
516  *	o. the first one if more than one cache
517  *
518  * The malloc and free prologues and epilogues mimic the necessary portions of
519  * libumem's malloc and free. This includes things like checking for size
520  * overflow, setting and verifying the malloc tags.
521  *
522  * It is an important constraint that these functions do not make use of the
523  * call instruction. The only jmp outside of the individual functions is to the
524  * original libumem malloc and free respectively. Because doing things like
525  * setting errno or raising an internal umem error on improper malloc tags would
526  * require using calls into the PLT, whenever we encounter one of those cases we
527  * just jump to the original malloc and free functions reusing the same stack
528  * frame.
529  *
530  * Each of the above sections, the three caches, and the malloc and free
531  * prologue and epilogue are implemented as blocks of machine code with the
532  * corresponding assembly in comments. There are known offsets into each block
533  * that corresponds to locations of data and addresses that we only know at run
534  * time. These blocks are copied as necessary and the blanks filled in
535  * appropriately.
536  *
537  * As mentioned in section 8.2, the trampoline library uses specifically named
538  * variables to communicate the buffers and size to use. These variables are:
539  *
540  *	o. umem_genasm_mptr: The buffer for ptcmalloc
541  *	o. umem_genasm_msize: The size in bytes of the above buffer
542  *	o. umem_genasm_fptr: The buffer for ptcfree
543  *	o. umem_genasm_fsize: The size in bytes of the above buffer
544  *
545  * Finally, to enable the generated assembly we need to remove the previous jump
546  * to the actual malloc that exists at the start of these buffers. On x86, this
547  * is a five byte region. We could zero out the jump offset to be a jmp +0, but
548  * using nops can be faster. We specifically use a single five byte nop on x86
549  * as it is faster. When porting ptcumem to other architectures, the various
550  * opcode changes and options should be analyzed.
551  *
552  * 8.4 Interface with libc.so
553  * --------------------------
554  *
555  * The tmem_t structure as described in the beginning of section 8, is part of a
556  * private interface with libc. There are three functions that exist to cover
557  * this. They are not documented in man pages or header files. They are in the
558  * SUNWprivate part of libc's mapfile.
559  *
560  *	o. _tmem_get_base(void)
561  *
562  * 	Returns the offset from the ulwp_t (curthread) to the tmem_t structure.
563  * 	This is a constant for all threads and is effectively a way to to do
564  * 	::offsetof ulwp_t ul_tmem without having to know the specifics of the
565  * 	structure outside of libc.
566  *
567  *	o. _tmem_get_nentries(void)
568  *
569  *	Returns the number of roots that exist in the tmem_t. This is one part
570  *	of the cap on the number of umem_caches that we can back with tmem.
571  *
572  *	o. _tmem_set_cleanup(void (*)(void *, int))
573  *
574  *	This sets a clean up handler that gets called back when a thread exits.
575  *	There is one call per buffer, the void * is a pointer to the buffer on
576  *	the list, the int is the index into the roots array for this buffer.
577  *
578  * 8.5 Tuning and disabling per-thread caching
579  * -------------------------------------------
580  *
581  * There is only one tunable for per-thread caching:  the amount of memory each
582  * thread should be able to cache.  This is specified via the perthread_cache
583  * UMEM_OPTION option.  No attempt is made to to sanity check the specified
584  * value; the limit is simply the maximum value of a size_t.
585  *
586  * If the perthread_cache UMEM_OPTION is set to zero, nomagazines was requested,
587  * or UMEM_DEBUG has been turned on then we will never call into umem_genasm;
588  * however, the trampoline audit library and jump will still be in place.
589  *
590  * 8.6 Observing efficacy of per-thread caching
591  * --------------------------------------------
592  *
593  * To understand the efficacy of per-thread caching, use the ::umastat dcmd
594  * to see the percentage of capacity consumed on a per-thread basis, the
595  * degree to which each umem cache contributes to per-thread cache consumption,
596  * and the number of buffers in per-thread caches on a per-umem cache basis.
597  * If more detail is required, the specific buffers in a per-thread cache can
598  * be iterated over with the umem_ptc_* walkers. (These walkers allow an
599  * optional ulwp_t to be specified to iterate only over a particular thread's
600  * cache.)
601  */
602 
603 #include <umem_impl.h>
604 #include <sys/vmem_impl_user.h>
605 #include "umem_base.h"
606 #include "vmem_base.h"
607 
608 #include <sys/processor.h>
609 #include <sys/sysmacros.h>
610 
611 #include <alloca.h>
612 #include <errno.h>
613 #include <limits.h>
614 #include <stdio.h>
615 #include <stdlib.h>
616 #include <string.h>
617 #include <strings.h>
618 #include <signal.h>
619 #include <unistd.h>
620 #include <atomic.h>
621 
622 #include "misc.h"
623 
624 #define	UMEM_VMFLAGS(umflag)	(VM_NOSLEEP)
625 
626 size_t pagesize;
627 
628 /*
629  * The default set of caches to back umem_alloc().
630  * These sizes should be reevaluated periodically.
631  *
632  * We want allocations that are multiples of the coherency granularity
633  * (64 bytes) to be satisfied from a cache which is a multiple of 64
634  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
635  * the next kmem_cache_size greater than or equal to it must be a
636  * multiple of 64.
637  *
638  * This table must be in sorted order, from smallest to highest.  The
639  * highest slot must be UMEM_MAXBUF, and every slot afterwards must be
640  * zero.
641  */
642 static int umem_alloc_sizes[] = {
643 #ifdef _LP64
644 	1 * 8,
645 	1 * 16,
646 	2 * 16,
647 	3 * 16,
648 #else
649 	1 * 8,
650 	2 * 8,
651 	3 * 8,
652 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
653 #endif
654 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
655 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
656 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
657 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
658 	P2ALIGN(8192 / 7, 64),
659 	P2ALIGN(8192 / 6, 64),
660 	P2ALIGN(8192 / 5, 64),
661 	P2ALIGN(8192 / 4, 64), 2304,
662 	P2ALIGN(8192 / 3, 64),
663 	P2ALIGN(8192 / 2, 64), 4544,
664 	P2ALIGN(8192 / 1, 64), 9216,
665 	4096 * 3,
666 	8192 * 2,				/* = 8192 * 2 */
667 	24576, 32768, 40960, 49152, 57344, 65536, 73728, 81920,
668 	90112, 98304, 106496, 114688, 122880, UMEM_MAXBUF, /* 128k */
669 	/* 24 slots for user expansion */
670 	0, 0, 0, 0, 0, 0, 0, 0,
671 	0, 0, 0, 0, 0, 0, 0, 0,
672 	0, 0, 0, 0, 0, 0, 0, 0,
673 };
674 #define	NUM_ALLOC_SIZES (sizeof (umem_alloc_sizes) / sizeof (*umem_alloc_sizes))
675 
676 static umem_magtype_t umem_magtype[] = {
677 	{ 1,	8,	3200,	65536	},
678 	{ 3,	16,	256,	32768	},
679 	{ 7,	32,	64,	16384	},
680 	{ 15,	64,	0,	8192	},
681 	{ 31,	64,	0,	4096	},
682 	{ 47,	64,	0,	2048	},
683 	{ 63,	64,	0,	1024	},
684 	{ 95,	64,	0,	512	},
685 	{ 143,	64,	0,	0	},
686 };
687 
688 /*
689  * umem tunables
690  */
691 uint32_t umem_max_ncpus;	/* # of CPU caches. */
692 
693 uint32_t umem_stack_depth = 15; /* # stack frames in a bufctl_audit */
694 uint32_t umem_reap_interval = 10; /* max reaping rate (seconds) */
695 uint_t umem_depot_contention = 2; /* max failed trylocks per real interval */
696 uint_t umem_abort = 1;		/* whether to abort on error */
697 uint_t umem_output = 0;		/* whether to write to standard error */
698 uint_t umem_logging = 0;	/* umem_log_enter() override */
699 uint32_t umem_mtbf = 0;		/* mean time between failures [default: off] */
700 size_t umem_transaction_log_size; /* size of transaction log */
701 size_t umem_content_log_size;	/* size of content log */
702 size_t umem_failure_log_size;	/* failure log [4 pages per CPU] */
703 size_t umem_slab_log_size;	/* slab create log [4 pages per CPU] */
704 size_t umem_content_maxsave = 256; /* UMF_CONTENTS max bytes to log */
705 size_t umem_lite_minsize = 0;	/* minimum buffer size for UMF_LITE */
706 size_t umem_lite_maxalign = 1024; /* maximum buffer alignment for UMF_LITE */
707 size_t umem_maxverify;		/* maximum bytes to inspect in debug routines */
708 size_t umem_minfirewall;	/* hardware-enforced redzone threshold */
709 size_t umem_ptc_size = 1048576;	/* size of per-thread cache (in bytes) */
710 
711 uint_t umem_flags = 0;
712 uintptr_t umem_tmem_off;
713 
714 mutex_t			umem_init_lock;		/* locks initialization */
715 cond_t			umem_init_cv;		/* initialization CV */
716 thread_t		umem_init_thr;		/* thread initializing */
717 int			umem_init_env_ready;	/* environ pre-initted */
718 int			umem_ready = UMEM_READY_STARTUP;
719 
720 int			umem_ptc_enabled;	/* per-thread caching enabled */
721 
722 static umem_nofail_callback_t *nofail_callback;
723 static mutex_t		umem_nofail_exit_lock;
724 static thread_t		umem_nofail_exit_thr;
725 
726 static umem_cache_t	*umem_slab_cache;
727 static umem_cache_t	*umem_bufctl_cache;
728 static umem_cache_t	*umem_bufctl_audit_cache;
729 
730 mutex_t			umem_flags_lock;
731 
732 static vmem_t		*heap_arena;
733 static vmem_alloc_t	*heap_alloc;
734 static vmem_free_t	*heap_free;
735 
736 static vmem_t		*umem_internal_arena;
737 static vmem_t		*umem_cache_arena;
738 static vmem_t		*umem_hash_arena;
739 static vmem_t		*umem_log_arena;
740 static vmem_t		*umem_oversize_arena;
741 static vmem_t		*umem_va_arena;
742 static vmem_t		*umem_default_arena;
743 static vmem_t		*umem_firewall_va_arena;
744 static vmem_t		*umem_firewall_arena;
745 
746 vmem_t			*umem_memalign_arena;
747 
748 umem_log_header_t *umem_transaction_log;
749 umem_log_header_t *umem_content_log;
750 umem_log_header_t *umem_failure_log;
751 umem_log_header_t *umem_slab_log;
752 
753 #define	CPUHINT()		(thr_self())
754 #define	CPUHINT_MAX()		INT_MAX
755 
756 #define	CPU(mask)		(umem_cpus + (CPUHINT() & (mask)))
757 static umem_cpu_t umem_startup_cpu = {	/* initial, single, cpu */
758 	UMEM_CACHE_SIZE(0),
759 	0
760 };
761 
762 static uint32_t umem_cpu_mask = 0;			/* global cpu mask */
763 static umem_cpu_t *umem_cpus = &umem_startup_cpu;	/* cpu list */
764 
765 volatile uint32_t umem_reaping;
766 
767 thread_t		umem_update_thr;
768 struct timeval		umem_update_next;	/* timeofday of next update */
769 volatile thread_t	umem_st_update_thr;	/* only used when single-thd */
770 
771 #define	IN_UPDATE()	(thr_self() == umem_update_thr || \
772 			    thr_self() == umem_st_update_thr)
773 #define	IN_REAP()	IN_UPDATE()
774 
775 mutex_t			umem_update_lock;	/* cache_u{next,prev,flags} */
776 cond_t			umem_update_cv;
777 
778 volatile hrtime_t umem_reap_next;	/* min hrtime of next reap */
779 
780 mutex_t			umem_cache_lock;	/* inter-cache linkage only */
781 
782 #ifdef UMEM_STANDALONE
783 umem_cache_t		umem_null_cache;
784 static const umem_cache_t umem_null_cache_template = {
785 #else
786 umem_cache_t		umem_null_cache = {
787 #endif
788 	0, 0, 0, 0, 0,
789 	0, 0,
790 	0, 0,
791 	0, 0,
792 	"invalid_cache",
793 	0, 0,
794 	NULL, NULL, NULL, NULL,
795 	NULL,
796 	0, 0, 0, 0,
797 	&umem_null_cache, &umem_null_cache,
798 	&umem_null_cache, &umem_null_cache,
799 	0,
800 	DEFAULTMUTEX,				/* start of slab layer */
801 	0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
802 	&umem_null_cache.cache_nullslab,
803 	{
804 		&umem_null_cache,
805 		NULL,
806 		&umem_null_cache.cache_nullslab,
807 		&umem_null_cache.cache_nullslab,
808 		NULL,
809 		-1,
810 		0
811 	},
812 	NULL,
813 	NULL,
814 	DEFAULTMUTEX,				/* start of depot layer */
815 	NULL, {
816 		NULL, 0, 0, 0, 0
817 	}, {
818 		NULL, 0, 0, 0, 0
819 	}, {
820 		{
821 			DEFAULTMUTEX,		/* start of CPU cache */
822 			0, 0, NULL, NULL, -1, -1, 0
823 		}
824 	}
825 };
826 
827 #define	ALLOC_TABLE_4 \
828 	&umem_null_cache, &umem_null_cache, &umem_null_cache, &umem_null_cache
829 
830 #define	ALLOC_TABLE_64 \
831 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
832 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
833 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, \
834 	ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4, ALLOC_TABLE_4
835 
836 #define	ALLOC_TABLE_1024 \
837 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
838 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
839 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, \
840 	ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64, ALLOC_TABLE_64
841 
842 static umem_cache_t *umem_alloc_table[UMEM_MAXBUF >> UMEM_ALIGN_SHIFT] = {
843 	ALLOC_TABLE_1024,
844 	ALLOC_TABLE_1024,
845 	ALLOC_TABLE_1024,
846 	ALLOC_TABLE_1024,
847 	ALLOC_TABLE_1024,
848 	ALLOC_TABLE_1024,
849 	ALLOC_TABLE_1024,
850 	ALLOC_TABLE_1024,
851 	ALLOC_TABLE_1024,
852 	ALLOC_TABLE_1024,
853 	ALLOC_TABLE_1024,
854 	ALLOC_TABLE_1024,
855 	ALLOC_TABLE_1024,
856 	ALLOC_TABLE_1024,
857 	ALLOC_TABLE_1024,
858 	ALLOC_TABLE_1024
859 };
860 
861 
862 /* Used to constrain audit-log stack traces */
863 caddr_t			umem_min_stack;
864 caddr_t			umem_max_stack;
865 
866 
867 #define	UMERR_MODIFIED	0	/* buffer modified while on freelist */
868 #define	UMERR_REDZONE	1	/* redzone violation (write past end of buf) */
869 #define	UMERR_DUPFREE	2	/* freed a buffer twice */
870 #define	UMERR_BADADDR	3	/* freed a bad (unallocated) address */
871 #define	UMERR_BADBUFTAG	4	/* buftag corrupted */
872 #define	UMERR_BADBUFCTL	5	/* bufctl corrupted */
873 #define	UMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
874 #define	UMERR_BADSIZE	7	/* alloc size != free size */
875 #define	UMERR_BADBASE	8	/* buffer base address wrong */
876 
877 struct {
878 	hrtime_t	ump_timestamp;	/* timestamp of error */
879 	int		ump_error;	/* type of umem error (UMERR_*) */
880 	void		*ump_buffer;	/* buffer that induced abort */
881 	void		*ump_realbuf;	/* real start address for buffer */
882 	umem_cache_t	*ump_cache;	/* buffer's cache according to client */
883 	umem_cache_t	*ump_realcache;	/* actual cache containing buffer */
884 	umem_slab_t	*ump_slab;	/* slab accoring to umem_findslab() */
885 	umem_bufctl_t	*ump_bufctl;	/* bufctl */
886 } umem_abort_info;
887 
888 static void
889 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
890 {
891 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
892 	uint64_t *buf = buf_arg;
893 
894 	while (buf < bufend)
895 		*buf++ = pattern;
896 }
897 
898 static void *
899 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
900 {
901 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
902 	uint64_t *buf;
903 
904 	for (buf = buf_arg; buf < bufend; buf++)
905 		if (*buf != pattern)
906 			return (buf);
907 	return (NULL);
908 }
909 
910 static void *
911 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
912 {
913 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
914 	uint64_t *buf;
915 
916 	for (buf = buf_arg; buf < bufend; buf++) {
917 		if (*buf != old) {
918 			copy_pattern(old, buf_arg,
919 			    (char *)buf - (char *)buf_arg);
920 			return (buf);
921 		}
922 		*buf = new;
923 	}
924 
925 	return (NULL);
926 }
927 
928 void
929 umem_cache_applyall(void (*func)(umem_cache_t *))
930 {
931 	umem_cache_t *cp;
932 
933 	(void) mutex_lock(&umem_cache_lock);
934 	for (cp = umem_null_cache.cache_next; cp != &umem_null_cache;
935 	    cp = cp->cache_next)
936 		func(cp);
937 	(void) mutex_unlock(&umem_cache_lock);
938 }
939 
940 static void
941 umem_add_update_unlocked(umem_cache_t *cp, int flags)
942 {
943 	umem_cache_t *cnext, *cprev;
944 
945 	flags &= ~UMU_ACTIVE;
946 
947 	if (!flags)
948 		return;
949 
950 	if (cp->cache_uflags & UMU_ACTIVE) {
951 		cp->cache_uflags |= flags;
952 	} else {
953 		if (cp->cache_unext != NULL) {
954 			ASSERT(cp->cache_uflags != 0);
955 			cp->cache_uflags |= flags;
956 		} else {
957 			ASSERT(cp->cache_uflags == 0);
958 			cp->cache_uflags = flags;
959 			cp->cache_unext = cnext = &umem_null_cache;
960 			cp->cache_uprev = cprev = umem_null_cache.cache_uprev;
961 			cnext->cache_uprev = cp;
962 			cprev->cache_unext = cp;
963 		}
964 	}
965 }
966 
967 static void
968 umem_add_update(umem_cache_t *cp, int flags)
969 {
970 	(void) mutex_lock(&umem_update_lock);
971 
972 	umem_add_update_unlocked(cp, flags);
973 
974 	if (!IN_UPDATE())
975 		(void) cond_broadcast(&umem_update_cv);
976 
977 	(void) mutex_unlock(&umem_update_lock);
978 }
979 
980 /*
981  * Remove a cache from the update list, waiting for any in-progress work to
982  * complete first.
983  */
984 static void
985 umem_remove_updates(umem_cache_t *cp)
986 {
987 	(void) mutex_lock(&umem_update_lock);
988 
989 	/*
990 	 * Get it out of the active state
991 	 */
992 	while (cp->cache_uflags & UMU_ACTIVE) {
993 		int cancel_state;
994 
995 		ASSERT(cp->cache_unext == NULL);
996 
997 		cp->cache_uflags |= UMU_NOTIFY;
998 
999 		/*
1000 		 * Make sure the update state is sane, before we wait
1001 		 */
1002 		ASSERT(umem_update_thr != 0 || umem_st_update_thr != 0);
1003 		ASSERT(umem_update_thr != thr_self() &&
1004 		    umem_st_update_thr != thr_self());
1005 
1006 		(void) pthread_setcancelstate(PTHREAD_CANCEL_DISABLE,
1007 		    &cancel_state);
1008 		(void) cond_wait(&umem_update_cv, &umem_update_lock);
1009 		(void) pthread_setcancelstate(cancel_state, NULL);
1010 	}
1011 	/*
1012 	 * Get it out of the Work Requested state
1013 	 */
1014 	if (cp->cache_unext != NULL) {
1015 		cp->cache_uprev->cache_unext = cp->cache_unext;
1016 		cp->cache_unext->cache_uprev = cp->cache_uprev;
1017 		cp->cache_uprev = cp->cache_unext = NULL;
1018 		cp->cache_uflags = 0;
1019 	}
1020 	/*
1021 	 * Make sure it is in the Inactive state
1022 	 */
1023 	ASSERT(cp->cache_unext == NULL && cp->cache_uflags == 0);
1024 	(void) mutex_unlock(&umem_update_lock);
1025 }
1026 
1027 static void
1028 umem_updateall(int flags)
1029 {
1030 	umem_cache_t *cp;
1031 
1032 	/*
1033 	 * NOTE:  To prevent deadlock, umem_cache_lock is always acquired first.
1034 	 *
1035 	 * (umem_add_update is called from things run via umem_cache_applyall)
1036 	 */
1037 	(void) mutex_lock(&umem_cache_lock);
1038 	(void) mutex_lock(&umem_update_lock);
1039 
1040 	for (cp = umem_null_cache.cache_next; cp != &umem_null_cache;
1041 	    cp = cp->cache_next)
1042 		umem_add_update_unlocked(cp, flags);
1043 
1044 	if (!IN_UPDATE())
1045 		(void) cond_broadcast(&umem_update_cv);
1046 
1047 	(void) mutex_unlock(&umem_update_lock);
1048 	(void) mutex_unlock(&umem_cache_lock);
1049 }
1050 
1051 /*
1052  * Debugging support.  Given a buffer address, find its slab.
1053  */
1054 static umem_slab_t *
1055 umem_findslab(umem_cache_t *cp, void *buf)
1056 {
1057 	umem_slab_t *sp;
1058 
1059 	(void) mutex_lock(&cp->cache_lock);
1060 	for (sp = cp->cache_nullslab.slab_next;
1061 	    sp != &cp->cache_nullslab; sp = sp->slab_next) {
1062 		if (UMEM_SLAB_MEMBER(sp, buf)) {
1063 			(void) mutex_unlock(&cp->cache_lock);
1064 			return (sp);
1065 		}
1066 	}
1067 	(void) mutex_unlock(&cp->cache_lock);
1068 
1069 	return (NULL);
1070 }
1071 
1072 static void
1073 umem_error(int error, umem_cache_t *cparg, void *bufarg)
1074 {
1075 	umem_buftag_t *btp = NULL;
1076 	umem_bufctl_t *bcp = NULL;
1077 	umem_cache_t *cp = cparg;
1078 	umem_slab_t *sp;
1079 	uint64_t *off;
1080 	void *buf = bufarg;
1081 
1082 	int old_logging = umem_logging;
1083 
1084 	umem_logging = 0;	/* stop logging when a bad thing happens */
1085 
1086 	umem_abort_info.ump_timestamp = gethrtime();
1087 
1088 	sp = umem_findslab(cp, buf);
1089 	if (sp == NULL) {
1090 		for (cp = umem_null_cache.cache_prev; cp != &umem_null_cache;
1091 		    cp = cp->cache_prev) {
1092 			if ((sp = umem_findslab(cp, buf)) != NULL)
1093 				break;
1094 		}
1095 	}
1096 
1097 	if (sp == NULL) {
1098 		cp = NULL;
1099 		error = UMERR_BADADDR;
1100 	} else {
1101 		if (cp != cparg)
1102 			error = UMERR_BADCACHE;
1103 		else
1104 			buf = (char *)bufarg - ((uintptr_t)bufarg -
1105 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
1106 		if (buf != bufarg)
1107 			error = UMERR_BADBASE;
1108 		if (cp->cache_flags & UMF_BUFTAG)
1109 			btp = UMEM_BUFTAG(cp, buf);
1110 		if (cp->cache_flags & UMF_HASH) {
1111 			(void) mutex_lock(&cp->cache_lock);
1112 			for (bcp = *UMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
1113 				if (bcp->bc_addr == buf)
1114 					break;
1115 			(void) mutex_unlock(&cp->cache_lock);
1116 			if (bcp == NULL && btp != NULL)
1117 				bcp = btp->bt_bufctl;
1118 			if (umem_findslab(cp->cache_bufctl_cache, bcp) ==
1119 			    NULL || P2PHASE((uintptr_t)bcp, UMEM_ALIGN) ||
1120 			    bcp->bc_addr != buf) {
1121 				error = UMERR_BADBUFCTL;
1122 				bcp = NULL;
1123 			}
1124 		}
1125 	}
1126 
1127 	umem_abort_info.ump_error = error;
1128 	umem_abort_info.ump_buffer = bufarg;
1129 	umem_abort_info.ump_realbuf = buf;
1130 	umem_abort_info.ump_cache = cparg;
1131 	umem_abort_info.ump_realcache = cp;
1132 	umem_abort_info.ump_slab = sp;
1133 	umem_abort_info.ump_bufctl = bcp;
1134 
1135 	umem_printf("umem allocator: ");
1136 
1137 	switch (error) {
1138 
1139 	case UMERR_MODIFIED:
1140 		umem_printf("buffer modified after being freed\n");
1141 		off = verify_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
1142 		if (off == NULL)	/* shouldn't happen */
1143 			off = buf;
1144 		umem_printf("modification occurred at offset 0x%lx "
1145 		    "(0x%llx replaced by 0x%llx)\n",
1146 		    (uintptr_t)off - (uintptr_t)buf,
1147 		    (longlong_t)UMEM_FREE_PATTERN, (longlong_t)*off);
1148 		break;
1149 
1150 	case UMERR_REDZONE:
1151 		umem_printf("redzone violation: write past end of buffer\n");
1152 		break;
1153 
1154 	case UMERR_BADADDR:
1155 		umem_printf("invalid free: buffer not in cache\n");
1156 		break;
1157 
1158 	case UMERR_DUPFREE:
1159 		umem_printf("duplicate free: buffer freed twice\n");
1160 		break;
1161 
1162 	case UMERR_BADBUFTAG:
1163 		umem_printf("boundary tag corrupted\n");
1164 		umem_printf("bcp ^ bxstat = %lx, should be %lx\n",
1165 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
1166 		    UMEM_BUFTAG_FREE);
1167 		break;
1168 
1169 	case UMERR_BADBUFCTL:
1170 		umem_printf("bufctl corrupted\n");
1171 		break;
1172 
1173 	case UMERR_BADCACHE:
1174 		umem_printf("buffer freed to wrong cache\n");
1175 		umem_printf("buffer was allocated from %s,\n", cp->cache_name);
1176 		umem_printf("caller attempting free to %s.\n",
1177 		    cparg->cache_name);
1178 		break;
1179 
1180 	case UMERR_BADSIZE:
1181 		umem_printf("bad free: free size (%u) != alloc size (%u)\n",
1182 		    UMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
1183 		    UMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
1184 		break;
1185 
1186 	case UMERR_BADBASE:
1187 		umem_printf("bad free: free address (%p) != alloc address "
1188 		    "(%p)\n", bufarg, buf);
1189 		break;
1190 	}
1191 
1192 	umem_printf("buffer=%p  bufctl=%p  cache: %s\n",
1193 	    bufarg, (void *)bcp, cparg->cache_name);
1194 
1195 	if (bcp != NULL && (cp->cache_flags & UMF_AUDIT) &&
1196 	    error != UMERR_BADBUFCTL) {
1197 		int d;
1198 		timespec_t ts;
1199 		hrtime_t diff;
1200 		umem_bufctl_audit_t *bcap = (umem_bufctl_audit_t *)bcp;
1201 
1202 		diff = umem_abort_info.ump_timestamp - bcap->bc_timestamp;
1203 		ts.tv_sec = diff / NANOSEC;
1204 		ts.tv_nsec = diff % NANOSEC;
1205 
1206 		umem_printf("previous transaction on buffer %p:\n", buf);
1207 		umem_printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
1208 		    (void *)(intptr_t)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
1209 		    (void *)sp, cp->cache_name);
1210 		for (d = 0; d < MIN(bcap->bc_depth, umem_stack_depth); d++) {
1211 			(void) print_sym((void *)bcap->bc_stack[d]);
1212 			umem_printf("\n");
1213 		}
1214 	}
1215 
1216 	umem_err_recoverable("umem: heap corruption detected");
1217 
1218 	umem_logging = old_logging;	/* resume logging */
1219 }
1220 
1221 void
1222 umem_nofail_callback(umem_nofail_callback_t *cb)
1223 {
1224 	nofail_callback = cb;
1225 }
1226 
1227 static int
1228 umem_alloc_retry(umem_cache_t *cp, int umflag)
1229 {
1230 	if (cp == &umem_null_cache) {
1231 		if (umem_init())
1232 			return (1);				/* retry */
1233 		/*
1234 		 * Initialization failed.  Do normal failure processing.
1235 		 */
1236 	}
1237 	if (umflag & UMEM_NOFAIL) {
1238 		int def_result = UMEM_CALLBACK_EXIT(255);
1239 		int result = def_result;
1240 		umem_nofail_callback_t *callback = nofail_callback;
1241 
1242 		if (callback != NULL)
1243 			result = callback();
1244 
1245 		if (result == UMEM_CALLBACK_RETRY)
1246 			return (1);
1247 
1248 		if ((result & ~0xFF) != UMEM_CALLBACK_EXIT(0)) {
1249 			log_message("nofail callback returned %x\n", result);
1250 			result = def_result;
1251 		}
1252 
1253 		/*
1254 		 * only one thread will call exit
1255 		 */
1256 		if (umem_nofail_exit_thr == thr_self())
1257 			umem_panic("recursive UMEM_CALLBACK_EXIT()\n");
1258 
1259 		(void) mutex_lock(&umem_nofail_exit_lock);
1260 		umem_nofail_exit_thr = thr_self();
1261 		exit(result & 0xFF);
1262 		/*NOTREACHED*/
1263 	}
1264 	return (0);
1265 }
1266 
1267 static umem_log_header_t *
1268 umem_log_init(size_t logsize)
1269 {
1270 	umem_log_header_t *lhp;
1271 	int nchunks = 4 * umem_max_ncpus;
1272 	size_t lhsize = offsetof(umem_log_header_t, lh_cpu[umem_max_ncpus]);
1273 	int i;
1274 
1275 	if (logsize == 0)
1276 		return (NULL);
1277 
1278 	/*
1279 	 * Make sure that lhp->lh_cpu[] is nicely aligned
1280 	 * to prevent false sharing of cache lines.
1281 	 */
1282 	lhsize = P2ROUNDUP(lhsize, UMEM_ALIGN);
1283 	lhp = vmem_xalloc(umem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
1284 	    NULL, NULL, VM_NOSLEEP);
1285 	if (lhp == NULL)
1286 		goto fail;
1287 
1288 	bzero(lhp, lhsize);
1289 
1290 	(void) mutex_init(&lhp->lh_lock, USYNC_THREAD, NULL);
1291 	lhp->lh_nchunks = nchunks;
1292 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks, PAGESIZE);
1293 	if (lhp->lh_chunksize == 0)
1294 		lhp->lh_chunksize = PAGESIZE;
1295 
1296 	lhp->lh_base = vmem_alloc(umem_log_arena,
1297 	    lhp->lh_chunksize * nchunks, VM_NOSLEEP);
1298 	if (lhp->lh_base == NULL)
1299 		goto fail;
1300 
1301 	lhp->lh_free = vmem_alloc(umem_log_arena,
1302 	    nchunks * sizeof (int), VM_NOSLEEP);
1303 	if (lhp->lh_free == NULL)
1304 		goto fail;
1305 
1306 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
1307 
1308 	for (i = 0; i < umem_max_ncpus; i++) {
1309 		umem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
1310 		(void) mutex_init(&clhp->clh_lock, USYNC_THREAD, NULL);
1311 		clhp->clh_chunk = i;
1312 	}
1313 
1314 	for (i = umem_max_ncpus; i < nchunks; i++)
1315 		lhp->lh_free[i] = i;
1316 
1317 	lhp->lh_head = umem_max_ncpus;
1318 	lhp->lh_tail = 0;
1319 
1320 	return (lhp);
1321 
1322 fail:
1323 	if (lhp != NULL) {
1324 		if (lhp->lh_base != NULL)
1325 			vmem_free(umem_log_arena, lhp->lh_base,
1326 			    lhp->lh_chunksize * nchunks);
1327 
1328 		vmem_xfree(umem_log_arena, lhp, lhsize);
1329 	}
1330 	return (NULL);
1331 }
1332 
1333 static void *
1334 umem_log_enter(umem_log_header_t *lhp, void *data, size_t size)
1335 {
1336 	void *logspace;
1337 	umem_cpu_log_header_t *clhp =
1338 	    &lhp->lh_cpu[CPU(umem_cpu_mask)->cpu_number];
1339 
1340 	if (lhp == NULL || umem_logging == 0)
1341 		return (NULL);
1342 
1343 	(void) mutex_lock(&clhp->clh_lock);
1344 	clhp->clh_hits++;
1345 	if (size > clhp->clh_avail) {
1346 		(void) mutex_lock(&lhp->lh_lock);
1347 		lhp->lh_hits++;
1348 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
1349 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
1350 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
1351 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
1352 		clhp->clh_current = lhp->lh_base +
1353 		    clhp->clh_chunk * lhp->lh_chunksize;
1354 		clhp->clh_avail = lhp->lh_chunksize;
1355 		if (size > lhp->lh_chunksize)
1356 			size = lhp->lh_chunksize;
1357 		(void) mutex_unlock(&lhp->lh_lock);
1358 	}
1359 	logspace = clhp->clh_current;
1360 	clhp->clh_current += size;
1361 	clhp->clh_avail -= size;
1362 	bcopy(data, logspace, size);
1363 	(void) mutex_unlock(&clhp->clh_lock);
1364 	return (logspace);
1365 }
1366 
1367 #define	UMEM_AUDIT(lp, cp, bcp)						\
1368 {									\
1369 	umem_bufctl_audit_t *_bcp = (umem_bufctl_audit_t *)(bcp);	\
1370 	_bcp->bc_timestamp = gethrtime();				\
1371 	_bcp->bc_thread = thr_self();					\
1372 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, umem_stack_depth,	\
1373 	    (cp != NULL) && (cp->cache_flags & UMF_CHECKSIGNAL));	\
1374 	_bcp->bc_lastlog = umem_log_enter((lp), _bcp,			\
1375 	    UMEM_BUFCTL_AUDIT_SIZE);					\
1376 }
1377 
1378 static void
1379 umem_log_event(umem_log_header_t *lp, umem_cache_t *cp,
1380 	umem_slab_t *sp, void *addr)
1381 {
1382 	umem_bufctl_audit_t *bcp;
1383 	UMEM_LOCAL_BUFCTL_AUDIT(&bcp);
1384 
1385 	bzero(bcp, UMEM_BUFCTL_AUDIT_SIZE);
1386 	bcp->bc_addr = addr;
1387 	bcp->bc_slab = sp;
1388 	bcp->bc_cache = cp;
1389 	UMEM_AUDIT(lp, cp, bcp);
1390 }
1391 
1392 /*
1393  * Create a new slab for cache cp.
1394  */
1395 static umem_slab_t *
1396 umem_slab_create(umem_cache_t *cp, int umflag)
1397 {
1398 	size_t slabsize = cp->cache_slabsize;
1399 	size_t chunksize = cp->cache_chunksize;
1400 	int cache_flags = cp->cache_flags;
1401 	size_t color, chunks;
1402 	char *buf, *slab;
1403 	umem_slab_t *sp;
1404 	umem_bufctl_t *bcp;
1405 	vmem_t *vmp = cp->cache_arena;
1406 
1407 	color = cp->cache_color + cp->cache_align;
1408 	if (color > cp->cache_maxcolor)
1409 		color = cp->cache_mincolor;
1410 	cp->cache_color = color;
1411 
1412 	slab = vmem_alloc(vmp, slabsize, UMEM_VMFLAGS(umflag));
1413 
1414 	if (slab == NULL)
1415 		goto vmem_alloc_failure;
1416 
1417 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
1418 
1419 	if (!(cp->cache_cflags & UMC_NOTOUCH) &&
1420 	    (cp->cache_flags & UMF_DEADBEEF))
1421 		copy_pattern(UMEM_UNINITIALIZED_PATTERN, slab, slabsize);
1422 
1423 	if (cache_flags & UMF_HASH) {
1424 		if ((sp = _umem_cache_alloc(umem_slab_cache, umflag)) == NULL)
1425 			goto slab_alloc_failure;
1426 		chunks = (slabsize - color) / chunksize;
1427 	} else {
1428 		sp = UMEM_SLAB(cp, slab);
1429 		chunks = (slabsize - sizeof (umem_slab_t) - color) / chunksize;
1430 	}
1431 
1432 	sp->slab_cache	= cp;
1433 	sp->slab_head	= NULL;
1434 	sp->slab_refcnt	= 0;
1435 	sp->slab_base	= buf = slab + color;
1436 	sp->slab_chunks	= chunks;
1437 
1438 	ASSERT(chunks > 0);
1439 	while (chunks-- != 0) {
1440 		if (cache_flags & UMF_HASH) {
1441 			bcp = _umem_cache_alloc(cp->cache_bufctl_cache, umflag);
1442 			if (bcp == NULL)
1443 				goto bufctl_alloc_failure;
1444 			if (cache_flags & UMF_AUDIT) {
1445 				umem_bufctl_audit_t *bcap =
1446 				    (umem_bufctl_audit_t *)bcp;
1447 				bzero(bcap, UMEM_BUFCTL_AUDIT_SIZE);
1448 				bcap->bc_cache = cp;
1449 			}
1450 			bcp->bc_addr = buf;
1451 			bcp->bc_slab = sp;
1452 		} else {
1453 			bcp = UMEM_BUFCTL(cp, buf);
1454 		}
1455 		if (cache_flags & UMF_BUFTAG) {
1456 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1457 			btp->bt_redzone = UMEM_REDZONE_PATTERN;
1458 			btp->bt_bufctl = bcp;
1459 			btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
1460 			if (cache_flags & UMF_DEADBEEF) {
1461 				copy_pattern(UMEM_FREE_PATTERN, buf,
1462 				    cp->cache_verify);
1463 			}
1464 		}
1465 		bcp->bc_next = sp->slab_head;
1466 		sp->slab_head = bcp;
1467 		buf += chunksize;
1468 	}
1469 
1470 	umem_log_event(umem_slab_log, cp, sp, slab);
1471 
1472 	return (sp);
1473 
1474 bufctl_alloc_failure:
1475 
1476 	while ((bcp = sp->slab_head) != NULL) {
1477 		sp->slab_head = bcp->bc_next;
1478 		_umem_cache_free(cp->cache_bufctl_cache, bcp);
1479 	}
1480 	_umem_cache_free(umem_slab_cache, sp);
1481 
1482 slab_alloc_failure:
1483 
1484 	vmem_free(vmp, slab, slabsize);
1485 
1486 vmem_alloc_failure:
1487 
1488 	umem_log_event(umem_failure_log, cp, NULL, NULL);
1489 	atomic_add_64(&cp->cache_alloc_fail, 1);
1490 
1491 	return (NULL);
1492 }
1493 
1494 /*
1495  * Destroy a slab.
1496  */
1497 static void
1498 umem_slab_destroy(umem_cache_t *cp, umem_slab_t *sp)
1499 {
1500 	vmem_t *vmp = cp->cache_arena;
1501 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
1502 
1503 	if (cp->cache_flags & UMF_HASH) {
1504 		umem_bufctl_t *bcp;
1505 		while ((bcp = sp->slab_head) != NULL) {
1506 			sp->slab_head = bcp->bc_next;
1507 			_umem_cache_free(cp->cache_bufctl_cache, bcp);
1508 		}
1509 		_umem_cache_free(umem_slab_cache, sp);
1510 	}
1511 	vmem_free(vmp, slab, cp->cache_slabsize);
1512 }
1513 
1514 /*
1515  * Allocate a raw (unconstructed) buffer from cp's slab layer.
1516  */
1517 static void *
1518 umem_slab_alloc(umem_cache_t *cp, int umflag)
1519 {
1520 	umem_bufctl_t *bcp, **hash_bucket;
1521 	umem_slab_t *sp;
1522 	void *buf;
1523 
1524 	(void) mutex_lock(&cp->cache_lock);
1525 	cp->cache_slab_alloc++;
1526 	sp = cp->cache_freelist;
1527 	ASSERT(sp->slab_cache == cp);
1528 	if (sp->slab_head == NULL) {
1529 		/*
1530 		 * The freelist is empty.  Create a new slab.
1531 		 */
1532 		(void) mutex_unlock(&cp->cache_lock);
1533 		if (cp == &umem_null_cache)
1534 			return (NULL);
1535 		if ((sp = umem_slab_create(cp, umflag)) == NULL)
1536 			return (NULL);
1537 		(void) mutex_lock(&cp->cache_lock);
1538 		cp->cache_slab_create++;
1539 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
1540 			cp->cache_bufmax = cp->cache_buftotal;
1541 		sp->slab_next = cp->cache_freelist;
1542 		sp->slab_prev = cp->cache_freelist->slab_prev;
1543 		sp->slab_next->slab_prev = sp;
1544 		sp->slab_prev->slab_next = sp;
1545 		cp->cache_freelist = sp;
1546 	}
1547 
1548 	sp->slab_refcnt++;
1549 	ASSERT(sp->slab_refcnt <= sp->slab_chunks);
1550 
1551 	/*
1552 	 * If we're taking the last buffer in the slab,
1553 	 * remove the slab from the cache's freelist.
1554 	 */
1555 	bcp = sp->slab_head;
1556 	if ((sp->slab_head = bcp->bc_next) == NULL) {
1557 		cp->cache_freelist = sp->slab_next;
1558 		ASSERT(sp->slab_refcnt == sp->slab_chunks);
1559 	}
1560 
1561 	if (cp->cache_flags & UMF_HASH) {
1562 		/*
1563 		 * Add buffer to allocated-address hash table.
1564 		 */
1565 		buf = bcp->bc_addr;
1566 		hash_bucket = UMEM_HASH(cp, buf);
1567 		bcp->bc_next = *hash_bucket;
1568 		*hash_bucket = bcp;
1569 		if ((cp->cache_flags & (UMF_AUDIT | UMF_BUFTAG)) == UMF_AUDIT) {
1570 			UMEM_AUDIT(umem_transaction_log, cp, bcp);
1571 		}
1572 	} else {
1573 		buf = UMEM_BUF(cp, bcp);
1574 	}
1575 
1576 	ASSERT(UMEM_SLAB_MEMBER(sp, buf));
1577 
1578 	(void) mutex_unlock(&cp->cache_lock);
1579 
1580 	return (buf);
1581 }
1582 
1583 /*
1584  * Free a raw (unconstructed) buffer to cp's slab layer.
1585  */
1586 static void
1587 umem_slab_free(umem_cache_t *cp, void *buf)
1588 {
1589 	umem_slab_t *sp;
1590 	umem_bufctl_t *bcp, **prev_bcpp;
1591 
1592 	ASSERT(buf != NULL);
1593 
1594 	(void) mutex_lock(&cp->cache_lock);
1595 	cp->cache_slab_free++;
1596 
1597 	if (cp->cache_flags & UMF_HASH) {
1598 		/*
1599 		 * Look up buffer in allocated-address hash table.
1600 		 */
1601 		prev_bcpp = UMEM_HASH(cp, buf);
1602 		while ((bcp = *prev_bcpp) != NULL) {
1603 			if (bcp->bc_addr == buf) {
1604 				*prev_bcpp = bcp->bc_next;
1605 				sp = bcp->bc_slab;
1606 				break;
1607 			}
1608 			cp->cache_lookup_depth++;
1609 			prev_bcpp = &bcp->bc_next;
1610 		}
1611 	} else {
1612 		bcp = UMEM_BUFCTL(cp, buf);
1613 		sp = UMEM_SLAB(cp, buf);
1614 	}
1615 
1616 	if (bcp == NULL || sp->slab_cache != cp || !UMEM_SLAB_MEMBER(sp, buf)) {
1617 		(void) mutex_unlock(&cp->cache_lock);
1618 		umem_error(UMERR_BADADDR, cp, buf);
1619 		return;
1620 	}
1621 
1622 	if ((cp->cache_flags & (UMF_AUDIT | UMF_BUFTAG)) == UMF_AUDIT) {
1623 		if (cp->cache_flags & UMF_CONTENTS)
1624 			((umem_bufctl_audit_t *)bcp)->bc_contents =
1625 			    umem_log_enter(umem_content_log, buf,
1626 			    cp->cache_contents);
1627 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
1628 	}
1629 
1630 	/*
1631 	 * If this slab isn't currently on the freelist, put it there.
1632 	 */
1633 	if (sp->slab_head == NULL) {
1634 		ASSERT(sp->slab_refcnt == sp->slab_chunks);
1635 		ASSERT(cp->cache_freelist != sp);
1636 		sp->slab_next->slab_prev = sp->slab_prev;
1637 		sp->slab_prev->slab_next = sp->slab_next;
1638 		sp->slab_next = cp->cache_freelist;
1639 		sp->slab_prev = cp->cache_freelist->slab_prev;
1640 		sp->slab_next->slab_prev = sp;
1641 		sp->slab_prev->slab_next = sp;
1642 		cp->cache_freelist = sp;
1643 	}
1644 
1645 	bcp->bc_next = sp->slab_head;
1646 	sp->slab_head = bcp;
1647 
1648 	ASSERT(sp->slab_refcnt >= 1);
1649 	if (--sp->slab_refcnt == 0) {
1650 		/*
1651 		 * There are no outstanding allocations from this slab,
1652 		 * so we can reclaim the memory.
1653 		 */
1654 		sp->slab_next->slab_prev = sp->slab_prev;
1655 		sp->slab_prev->slab_next = sp->slab_next;
1656 		if (sp == cp->cache_freelist)
1657 			cp->cache_freelist = sp->slab_next;
1658 		cp->cache_slab_destroy++;
1659 		cp->cache_buftotal -= sp->slab_chunks;
1660 		(void) mutex_unlock(&cp->cache_lock);
1661 		umem_slab_destroy(cp, sp);
1662 		return;
1663 	}
1664 	(void) mutex_unlock(&cp->cache_lock);
1665 }
1666 
1667 static int
1668 umem_cache_alloc_debug(umem_cache_t *cp, void *buf, int umflag)
1669 {
1670 	umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1671 	umem_bufctl_audit_t *bcp = (umem_bufctl_audit_t *)btp->bt_bufctl;
1672 	uint32_t mtbf;
1673 	int flags_nfatal;
1674 
1675 	if (btp->bt_bxstat != ((intptr_t)bcp ^ UMEM_BUFTAG_FREE)) {
1676 		umem_error(UMERR_BADBUFTAG, cp, buf);
1677 		return (-1);
1678 	}
1679 
1680 	btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_ALLOC;
1681 
1682 	if ((cp->cache_flags & UMF_HASH) && bcp->bc_addr != buf) {
1683 		umem_error(UMERR_BADBUFCTL, cp, buf);
1684 		return (-1);
1685 	}
1686 
1687 	btp->bt_redzone = UMEM_REDZONE_PATTERN;
1688 
1689 	if (cp->cache_flags & UMF_DEADBEEF) {
1690 		if (verify_and_copy_pattern(UMEM_FREE_PATTERN,
1691 		    UMEM_UNINITIALIZED_PATTERN, buf, cp->cache_verify)) {
1692 			umem_error(UMERR_MODIFIED, cp, buf);
1693 			return (-1);
1694 		}
1695 	}
1696 
1697 	if ((mtbf = umem_mtbf | cp->cache_mtbf) != 0 &&
1698 	    gethrtime() % mtbf == 0 &&
1699 	    (umflag & (UMEM_FATAL_FLAGS)) == 0) {
1700 		umem_log_event(umem_failure_log, cp, NULL, NULL);
1701 	} else {
1702 		mtbf = 0;
1703 	}
1704 
1705 	/*
1706 	 * We do not pass fatal flags on to the constructor.  This prevents
1707 	 * leaking buffers in the event of a subordinate constructor failing.
1708 	 */
1709 	flags_nfatal = UMEM_DEFAULT;
1710 	if (mtbf || (cp->cache_constructor != NULL &&
1711 	    cp->cache_constructor(buf, cp->cache_private, flags_nfatal) != 0)) {
1712 		atomic_add_64(&cp->cache_alloc_fail, 1);
1713 		btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
1714 		copy_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
1715 		umem_slab_free(cp, buf);
1716 		return (-1);
1717 	}
1718 
1719 	if (cp->cache_flags & UMF_AUDIT) {
1720 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
1721 	}
1722 
1723 	return (0);
1724 }
1725 
1726 static int
1727 umem_cache_free_debug(umem_cache_t *cp, void *buf)
1728 {
1729 	umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
1730 	umem_bufctl_audit_t *bcp = (umem_bufctl_audit_t *)btp->bt_bufctl;
1731 	umem_slab_t *sp;
1732 
1733 	if (btp->bt_bxstat != ((intptr_t)bcp ^ UMEM_BUFTAG_ALLOC)) {
1734 		if (btp->bt_bxstat == ((intptr_t)bcp ^ UMEM_BUFTAG_FREE)) {
1735 			umem_error(UMERR_DUPFREE, cp, buf);
1736 			return (-1);
1737 		}
1738 		sp = umem_findslab(cp, buf);
1739 		if (sp == NULL || sp->slab_cache != cp)
1740 			umem_error(UMERR_BADADDR, cp, buf);
1741 		else
1742 			umem_error(UMERR_REDZONE, cp, buf);
1743 		return (-1);
1744 	}
1745 
1746 	btp->bt_bxstat = (intptr_t)bcp ^ UMEM_BUFTAG_FREE;
1747 
1748 	if ((cp->cache_flags & UMF_HASH) && bcp->bc_addr != buf) {
1749 		umem_error(UMERR_BADBUFCTL, cp, buf);
1750 		return (-1);
1751 	}
1752 
1753 	if (btp->bt_redzone != UMEM_REDZONE_PATTERN) {
1754 		umem_error(UMERR_REDZONE, cp, buf);
1755 		return (-1);
1756 	}
1757 
1758 	if (cp->cache_flags & UMF_AUDIT) {
1759 		if (cp->cache_flags & UMF_CONTENTS)
1760 			bcp->bc_contents = umem_log_enter(umem_content_log,
1761 			    buf, cp->cache_contents);
1762 		UMEM_AUDIT(umem_transaction_log, cp, bcp);
1763 	}
1764 
1765 	if (cp->cache_destructor != NULL)
1766 		cp->cache_destructor(buf, cp->cache_private);
1767 
1768 	if (cp->cache_flags & UMF_DEADBEEF)
1769 		copy_pattern(UMEM_FREE_PATTERN, buf, cp->cache_verify);
1770 
1771 	return (0);
1772 }
1773 
1774 /*
1775  * Free each object in magazine mp to cp's slab layer, and free mp itself.
1776  */
1777 static void
1778 umem_magazine_destroy(umem_cache_t *cp, umem_magazine_t *mp, int nrounds)
1779 {
1780 	int round;
1781 
1782 	ASSERT(cp->cache_next == NULL || IN_UPDATE());
1783 
1784 	for (round = 0; round < nrounds; round++) {
1785 		void *buf = mp->mag_round[round];
1786 
1787 		if ((cp->cache_flags & UMF_DEADBEEF) &&
1788 		    verify_pattern(UMEM_FREE_PATTERN, buf,
1789 		    cp->cache_verify) != NULL) {
1790 			umem_error(UMERR_MODIFIED, cp, buf);
1791 			continue;
1792 		}
1793 
1794 		if (!(cp->cache_flags & UMF_BUFTAG) &&
1795 		    cp->cache_destructor != NULL)
1796 			cp->cache_destructor(buf, cp->cache_private);
1797 
1798 		umem_slab_free(cp, buf);
1799 	}
1800 	ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
1801 	_umem_cache_free(cp->cache_magtype->mt_cache, mp);
1802 }
1803 
1804 /*
1805  * Allocate a magazine from the depot.
1806  */
1807 static umem_magazine_t *
1808 umem_depot_alloc(umem_cache_t *cp, umem_maglist_t *mlp)
1809 {
1810 	umem_magazine_t *mp;
1811 
1812 	/*
1813 	 * If we can't get the depot lock without contention,
1814 	 * update our contention count.  We use the depot
1815 	 * contention rate to determine whether we need to
1816 	 * increase the magazine size for better scalability.
1817 	 */
1818 	if (mutex_trylock(&cp->cache_depot_lock) != 0) {
1819 		(void) mutex_lock(&cp->cache_depot_lock);
1820 		cp->cache_depot_contention++;
1821 	}
1822 
1823 	if ((mp = mlp->ml_list) != NULL) {
1824 		ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
1825 		mlp->ml_list = mp->mag_next;
1826 		if (--mlp->ml_total < mlp->ml_min)
1827 			mlp->ml_min = mlp->ml_total;
1828 		mlp->ml_alloc++;
1829 	}
1830 
1831 	(void) mutex_unlock(&cp->cache_depot_lock);
1832 
1833 	return (mp);
1834 }
1835 
1836 /*
1837  * Free a magazine to the depot.
1838  */
1839 static void
1840 umem_depot_free(umem_cache_t *cp, umem_maglist_t *mlp, umem_magazine_t *mp)
1841 {
1842 	(void) mutex_lock(&cp->cache_depot_lock);
1843 	ASSERT(UMEM_MAGAZINE_VALID(cp, mp));
1844 	mp->mag_next = mlp->ml_list;
1845 	mlp->ml_list = mp;
1846 	mlp->ml_total++;
1847 	(void) mutex_unlock(&cp->cache_depot_lock);
1848 }
1849 
1850 /*
1851  * Update the working set statistics for cp's depot.
1852  */
1853 static void
1854 umem_depot_ws_update(umem_cache_t *cp)
1855 {
1856 	(void) mutex_lock(&cp->cache_depot_lock);
1857 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
1858 	cp->cache_full.ml_min = cp->cache_full.ml_total;
1859 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
1860 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
1861 	(void) mutex_unlock(&cp->cache_depot_lock);
1862 }
1863 
1864 /*
1865  * Reap all magazines that have fallen out of the depot's working set.
1866  */
1867 static void
1868 umem_depot_ws_reap(umem_cache_t *cp)
1869 {
1870 	long reap;
1871 	umem_magazine_t *mp;
1872 
1873 	ASSERT(cp->cache_next == NULL || IN_REAP());
1874 
1875 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
1876 	while (reap-- && (mp = umem_depot_alloc(cp, &cp->cache_full)) != NULL)
1877 		umem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
1878 
1879 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
1880 	while (reap-- && (mp = umem_depot_alloc(cp, &cp->cache_empty)) != NULL)
1881 		umem_magazine_destroy(cp, mp, 0);
1882 }
1883 
1884 static void
1885 umem_cpu_reload(umem_cpu_cache_t *ccp, umem_magazine_t *mp, int rounds)
1886 {
1887 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
1888 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
1889 	ASSERT(ccp->cc_magsize > 0);
1890 
1891 	ccp->cc_ploaded = ccp->cc_loaded;
1892 	ccp->cc_prounds = ccp->cc_rounds;
1893 	ccp->cc_loaded = mp;
1894 	ccp->cc_rounds = rounds;
1895 }
1896 
1897 /*
1898  * Allocate a constructed object from cache cp.
1899  */
1900 #pragma weak umem_cache_alloc = _umem_cache_alloc
1901 void *
1902 _umem_cache_alloc(umem_cache_t *cp, int umflag)
1903 {
1904 	umem_cpu_cache_t *ccp;
1905 	umem_magazine_t *fmp;
1906 	void *buf;
1907 	int flags_nfatal;
1908 
1909 retry:
1910 	ccp = UMEM_CPU_CACHE(cp, CPU(cp->cache_cpu_mask));
1911 	(void) mutex_lock(&ccp->cc_lock);
1912 	for (;;) {
1913 		/*
1914 		 * If there's an object available in the current CPU's
1915 		 * loaded magazine, just take it and return.
1916 		 */
1917 		if (ccp->cc_rounds > 0) {
1918 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
1919 			ccp->cc_alloc++;
1920 			(void) mutex_unlock(&ccp->cc_lock);
1921 			if ((ccp->cc_flags & UMF_BUFTAG) &&
1922 			    umem_cache_alloc_debug(cp, buf, umflag) == -1) {
1923 				if (umem_alloc_retry(cp, umflag)) {
1924 					goto retry;
1925 				}
1926 
1927 				return (NULL);
1928 			}
1929 			return (buf);
1930 		}
1931 
1932 		/*
1933 		 * The loaded magazine is empty.  If the previously loaded
1934 		 * magazine was full, exchange them and try again.
1935 		 */
1936 		if (ccp->cc_prounds > 0) {
1937 			umem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
1938 			continue;
1939 		}
1940 
1941 		/*
1942 		 * If the magazine layer is disabled, break out now.
1943 		 */
1944 		if (ccp->cc_magsize == 0)
1945 			break;
1946 
1947 		/*
1948 		 * Try to get a full magazine from the depot.
1949 		 */
1950 		fmp = umem_depot_alloc(cp, &cp->cache_full);
1951 		if (fmp != NULL) {
1952 			if (ccp->cc_ploaded != NULL)
1953 				umem_depot_free(cp, &cp->cache_empty,
1954 				    ccp->cc_ploaded);
1955 			umem_cpu_reload(ccp, fmp, ccp->cc_magsize);
1956 			continue;
1957 		}
1958 
1959 		/*
1960 		 * There are no full magazines in the depot,
1961 		 * so fall through to the slab layer.
1962 		 */
1963 		break;
1964 	}
1965 	(void) mutex_unlock(&ccp->cc_lock);
1966 
1967 	/*
1968 	 * We couldn't allocate a constructed object from the magazine layer,
1969 	 * so get a raw buffer from the slab layer and apply its constructor.
1970 	 */
1971 	buf = umem_slab_alloc(cp, umflag);
1972 
1973 	if (buf == NULL) {
1974 		if (cp == &umem_null_cache)
1975 			return (NULL);
1976 		if (umem_alloc_retry(cp, umflag)) {
1977 			goto retry;
1978 		}
1979 
1980 		return (NULL);
1981 	}
1982 
1983 	if (cp->cache_flags & UMF_BUFTAG) {
1984 		/*
1985 		 * Let umem_cache_alloc_debug() apply the constructor for us.
1986 		 */
1987 		if (umem_cache_alloc_debug(cp, buf, umflag) == -1) {
1988 			if (umem_alloc_retry(cp, umflag)) {
1989 				goto retry;
1990 			}
1991 			return (NULL);
1992 		}
1993 		return (buf);
1994 	}
1995 
1996 	/*
1997 	 * We do not pass fatal flags on to the constructor.  This prevents
1998 	 * leaking buffers in the event of a subordinate constructor failing.
1999 	 */
2000 	flags_nfatal = UMEM_DEFAULT;
2001 	if (cp->cache_constructor != NULL &&
2002 	    cp->cache_constructor(buf, cp->cache_private, flags_nfatal) != 0) {
2003 		atomic_add_64(&cp->cache_alloc_fail, 1);
2004 		umem_slab_free(cp, buf);
2005 
2006 		if (umem_alloc_retry(cp, umflag)) {
2007 			goto retry;
2008 		}
2009 		return (NULL);
2010 	}
2011 
2012 	return (buf);
2013 }
2014 
2015 /*
2016  * Free a constructed object to cache cp.
2017  */
2018 #pragma weak umem_cache_free = _umem_cache_free
2019 void
2020 _umem_cache_free(umem_cache_t *cp, void *buf)
2021 {
2022 	umem_cpu_cache_t *ccp = UMEM_CPU_CACHE(cp, CPU(cp->cache_cpu_mask));
2023 	umem_magazine_t *emp;
2024 	umem_magtype_t *mtp;
2025 
2026 	if (ccp->cc_flags & UMF_BUFTAG)
2027 		if (umem_cache_free_debug(cp, buf) == -1)
2028 			return;
2029 
2030 	(void) mutex_lock(&ccp->cc_lock);
2031 	for (;;) {
2032 		/*
2033 		 * If there's a slot available in the current CPU's
2034 		 * loaded magazine, just put the object there and return.
2035 		 */
2036 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
2037 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
2038 			ccp->cc_free++;
2039 			(void) mutex_unlock(&ccp->cc_lock);
2040 			return;
2041 		}
2042 
2043 		/*
2044 		 * The loaded magazine is full.  If the previously loaded
2045 		 * magazine was empty, exchange them and try again.
2046 		 */
2047 		if (ccp->cc_prounds == 0) {
2048 			umem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
2049 			continue;
2050 		}
2051 
2052 		/*
2053 		 * If the magazine layer is disabled, break out now.
2054 		 */
2055 		if (ccp->cc_magsize == 0)
2056 			break;
2057 
2058 		/*
2059 		 * Try to get an empty magazine from the depot.
2060 		 */
2061 		emp = umem_depot_alloc(cp, &cp->cache_empty);
2062 		if (emp != NULL) {
2063 			if (ccp->cc_ploaded != NULL)
2064 				umem_depot_free(cp, &cp->cache_full,
2065 				    ccp->cc_ploaded);
2066 			umem_cpu_reload(ccp, emp, 0);
2067 			continue;
2068 		}
2069 
2070 		/*
2071 		 * There are no empty magazines in the depot,
2072 		 * so try to allocate a new one.  We must drop all locks
2073 		 * across umem_cache_alloc() because lower layers may
2074 		 * attempt to allocate from this cache.
2075 		 */
2076 		mtp = cp->cache_magtype;
2077 		(void) mutex_unlock(&ccp->cc_lock);
2078 		emp = _umem_cache_alloc(mtp->mt_cache, UMEM_DEFAULT);
2079 		(void) mutex_lock(&ccp->cc_lock);
2080 
2081 		if (emp != NULL) {
2082 			/*
2083 			 * We successfully allocated an empty magazine.
2084 			 * However, we had to drop ccp->cc_lock to do it,
2085 			 * so the cache's magazine size may have changed.
2086 			 * If so, free the magazine and try again.
2087 			 */
2088 			if (ccp->cc_magsize != mtp->mt_magsize) {
2089 				(void) mutex_unlock(&ccp->cc_lock);
2090 				_umem_cache_free(mtp->mt_cache, emp);
2091 				(void) mutex_lock(&ccp->cc_lock);
2092 				continue;
2093 			}
2094 
2095 			/*
2096 			 * We got a magazine of the right size.  Add it to
2097 			 * the depot and try the whole dance again.
2098 			 */
2099 			umem_depot_free(cp, &cp->cache_empty, emp);
2100 			continue;
2101 		}
2102 
2103 		/*
2104 		 * We couldn't allocate an empty magazine,
2105 		 * so fall through to the slab layer.
2106 		 */
2107 		break;
2108 	}
2109 	(void) mutex_unlock(&ccp->cc_lock);
2110 
2111 	/*
2112 	 * We couldn't free our constructed object to the magazine layer,
2113 	 * so apply its destructor and free it to the slab layer.
2114 	 * Note that if UMF_BUFTAG is in effect, umem_cache_free_debug()
2115 	 * will have already applied the destructor.
2116 	 */
2117 	if (!(cp->cache_flags & UMF_BUFTAG) && cp->cache_destructor != NULL)
2118 		cp->cache_destructor(buf, cp->cache_private);
2119 
2120 	umem_slab_free(cp, buf);
2121 }
2122 
2123 #pragma weak umem_zalloc = _umem_zalloc
2124 void *
2125 _umem_zalloc(size_t size, int umflag)
2126 {
2127 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
2128 	void *buf;
2129 
2130 retry:
2131 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
2132 		umem_cache_t *cp = umem_alloc_table[index];
2133 		buf = _umem_cache_alloc(cp, umflag);
2134 		if (buf != NULL) {
2135 			if (cp->cache_flags & UMF_BUFTAG) {
2136 				umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
2137 				((uint8_t *)buf)[size] = UMEM_REDZONE_BYTE;
2138 				((uint32_t *)btp)[1] = UMEM_SIZE_ENCODE(size);
2139 			}
2140 			bzero(buf, size);
2141 		} else if (umem_alloc_retry(cp, umflag))
2142 			goto retry;
2143 	} else {
2144 		buf = _umem_alloc(size, umflag);	/* handles failure */
2145 		if (buf != NULL)
2146 			bzero(buf, size);
2147 	}
2148 	return (buf);
2149 }
2150 
2151 #pragma weak umem_alloc = _umem_alloc
2152 void *
2153 _umem_alloc(size_t size, int umflag)
2154 {
2155 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
2156 	void *buf;
2157 umem_alloc_retry:
2158 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
2159 		umem_cache_t *cp = umem_alloc_table[index];
2160 		buf = _umem_cache_alloc(cp, umflag);
2161 		if ((cp->cache_flags & UMF_BUFTAG) && buf != NULL) {
2162 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
2163 			((uint8_t *)buf)[size] = UMEM_REDZONE_BYTE;
2164 			((uint32_t *)btp)[1] = UMEM_SIZE_ENCODE(size);
2165 		}
2166 		if (buf == NULL && umem_alloc_retry(cp, umflag))
2167 			goto umem_alloc_retry;
2168 		return (buf);
2169 	}
2170 	if (size == 0)
2171 		return (NULL);
2172 	if (umem_oversize_arena == NULL) {
2173 		if (umem_init())
2174 			ASSERT(umem_oversize_arena != NULL);
2175 		else
2176 			return (NULL);
2177 	}
2178 	buf = vmem_alloc(umem_oversize_arena, size, UMEM_VMFLAGS(umflag));
2179 	if (buf == NULL) {
2180 		umem_log_event(umem_failure_log, NULL, NULL, (void *)size);
2181 		if (umem_alloc_retry(NULL, umflag))
2182 			goto umem_alloc_retry;
2183 	}
2184 	return (buf);
2185 }
2186 
2187 #pragma weak umem_alloc_align = _umem_alloc_align
2188 void *
2189 _umem_alloc_align(size_t size, size_t align, int umflag)
2190 {
2191 	void *buf;
2192 
2193 	if (size == 0)
2194 		return (NULL);
2195 	if ((align & (align - 1)) != 0)
2196 		return (NULL);
2197 	if (align < UMEM_ALIGN)
2198 		align = UMEM_ALIGN;
2199 
2200 umem_alloc_align_retry:
2201 	if (umem_memalign_arena == NULL) {
2202 		if (umem_init())
2203 			ASSERT(umem_oversize_arena != NULL);
2204 		else
2205 			return (NULL);
2206 	}
2207 	buf = vmem_xalloc(umem_memalign_arena, size, align, 0, 0, NULL, NULL,
2208 	    UMEM_VMFLAGS(umflag));
2209 	if (buf == NULL) {
2210 		umem_log_event(umem_failure_log, NULL, NULL, (void *)size);
2211 		if (umem_alloc_retry(NULL, umflag))
2212 			goto umem_alloc_align_retry;
2213 	}
2214 	return (buf);
2215 }
2216 
2217 #pragma weak umem_free = _umem_free
2218 void
2219 _umem_free(void *buf, size_t size)
2220 {
2221 	size_t index = (size - 1) >> UMEM_ALIGN_SHIFT;
2222 
2223 	if (index < UMEM_MAXBUF >> UMEM_ALIGN_SHIFT) {
2224 		umem_cache_t *cp = umem_alloc_table[index];
2225 		if (cp->cache_flags & UMF_BUFTAG) {
2226 			umem_buftag_t *btp = UMEM_BUFTAG(cp, buf);
2227 			uint32_t *ip = (uint32_t *)btp;
2228 			if (ip[1] != UMEM_SIZE_ENCODE(size)) {
2229 				if (*(uint64_t *)buf == UMEM_FREE_PATTERN) {
2230 					umem_error(UMERR_DUPFREE, cp, buf);
2231 					return;
2232 				}
2233 				if (UMEM_SIZE_VALID(ip[1])) {
2234 					ip[0] = UMEM_SIZE_ENCODE(size);
2235 					umem_error(UMERR_BADSIZE, cp, buf);
2236 				} else {
2237 					umem_error(UMERR_REDZONE, cp, buf);
2238 				}
2239 				return;
2240 			}
2241 			if (((uint8_t *)buf)[size] != UMEM_REDZONE_BYTE) {
2242 				umem_error(UMERR_REDZONE, cp, buf);
2243 				return;
2244 			}
2245 			btp->bt_redzone = UMEM_REDZONE_PATTERN;
2246 		}
2247 		_umem_cache_free(cp, buf);
2248 	} else {
2249 		if (buf == NULL && size == 0)
2250 			return;
2251 		vmem_free(umem_oversize_arena, buf, size);
2252 	}
2253 }
2254 
2255 #pragma weak umem_free_align = _umem_free_align
2256 void
2257 _umem_free_align(void *buf, size_t size)
2258 {
2259 	if (buf == NULL && size == 0)
2260 		return;
2261 	vmem_xfree(umem_memalign_arena, buf, size);
2262 }
2263 
2264 static void *
2265 umem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
2266 {
2267 	size_t realsize = size + vmp->vm_quantum;
2268 
2269 	/*
2270 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
2271 	 * vm_quantum will cause integer wraparound.  Check for this, and
2272 	 * blow off the firewall page in this case.  Note that such a
2273 	 * giant allocation (the entire address space) can never be
2274 	 * satisfied, so it will either fail immediately (VM_NOSLEEP)
2275 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
2276 	 * corresponding check in umem_firewall_va_free().
2277 	 */
2278 	if (realsize < size)
2279 		realsize = size;
2280 
2281 	return (vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT));
2282 }
2283 
2284 static void
2285 umem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
2286 {
2287 	vmem_free(vmp, addr, size + vmp->vm_quantum);
2288 }
2289 
2290 /*
2291  * Reclaim all unused memory from a cache.
2292  */
2293 static void
2294 umem_cache_reap(umem_cache_t *cp)
2295 {
2296 	/*
2297 	 * Ask the cache's owner to free some memory if possible.
2298 	 * The idea is to handle things like the inode cache, which
2299 	 * typically sits on a bunch of memory that it doesn't truly
2300 	 * *need*.  Reclaim policy is entirely up to the owner; this
2301 	 * callback is just an advisory plea for help.
2302 	 */
2303 	if (cp->cache_reclaim != NULL)
2304 		cp->cache_reclaim(cp->cache_private);
2305 
2306 	umem_depot_ws_reap(cp);
2307 }
2308 
2309 /*
2310  * Purge all magazines from a cache and set its magazine limit to zero.
2311  * All calls are serialized by being done by the update thread, except for
2312  * the final call from umem_cache_destroy().
2313  */
2314 static void
2315 umem_cache_magazine_purge(umem_cache_t *cp)
2316 {
2317 	umem_cpu_cache_t *ccp;
2318 	umem_magazine_t *mp, *pmp;
2319 	int rounds, prounds, cpu_seqid;
2320 
2321 	ASSERT(cp->cache_next == NULL || IN_UPDATE());
2322 
2323 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
2324 		ccp = &cp->cache_cpu[cpu_seqid];
2325 
2326 		(void) mutex_lock(&ccp->cc_lock);
2327 		mp = ccp->cc_loaded;
2328 		pmp = ccp->cc_ploaded;
2329 		rounds = ccp->cc_rounds;
2330 		prounds = ccp->cc_prounds;
2331 		ccp->cc_loaded = NULL;
2332 		ccp->cc_ploaded = NULL;
2333 		ccp->cc_rounds = -1;
2334 		ccp->cc_prounds = -1;
2335 		ccp->cc_magsize = 0;
2336 		(void) mutex_unlock(&ccp->cc_lock);
2337 
2338 		if (mp)
2339 			umem_magazine_destroy(cp, mp, rounds);
2340 		if (pmp)
2341 			umem_magazine_destroy(cp, pmp, prounds);
2342 	}
2343 
2344 	/*
2345 	 * Updating the working set statistics twice in a row has the
2346 	 * effect of setting the working set size to zero, so everything
2347 	 * is eligible for reaping.
2348 	 */
2349 	umem_depot_ws_update(cp);
2350 	umem_depot_ws_update(cp);
2351 
2352 	umem_depot_ws_reap(cp);
2353 }
2354 
2355 /*
2356  * Enable per-cpu magazines on a cache.
2357  */
2358 static void
2359 umem_cache_magazine_enable(umem_cache_t *cp)
2360 {
2361 	int cpu_seqid;
2362 
2363 	if (cp->cache_flags & UMF_NOMAGAZINE)
2364 		return;
2365 
2366 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
2367 		umem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
2368 		(void) mutex_lock(&ccp->cc_lock);
2369 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
2370 		(void) mutex_unlock(&ccp->cc_lock);
2371 	}
2372 
2373 }
2374 
2375 /*
2376  * Recompute a cache's magazine size.  The trade-off is that larger magazines
2377  * provide a higher transfer rate with the depot, while smaller magazines
2378  * reduce memory consumption.  Magazine resizing is an expensive operation;
2379  * it should not be done frequently.
2380  *
2381  * Changes to the magazine size are serialized by only having one thread
2382  * doing updates. (the update thread)
2383  *
2384  * Note: at present this only grows the magazine size.  It might be useful
2385  * to allow shrinkage too.
2386  */
2387 static void
2388 umem_cache_magazine_resize(umem_cache_t *cp)
2389 {
2390 	umem_magtype_t *mtp = cp->cache_magtype;
2391 
2392 	ASSERT(IN_UPDATE());
2393 
2394 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
2395 		umem_cache_magazine_purge(cp);
2396 		(void) mutex_lock(&cp->cache_depot_lock);
2397 		cp->cache_magtype = ++mtp;
2398 		cp->cache_depot_contention_prev =
2399 		    cp->cache_depot_contention + INT_MAX;
2400 		(void) mutex_unlock(&cp->cache_depot_lock);
2401 		umem_cache_magazine_enable(cp);
2402 	}
2403 }
2404 
2405 /*
2406  * Rescale a cache's hash table, so that the table size is roughly the
2407  * cache size.  We want the average lookup time to be extremely small.
2408  */
2409 static void
2410 umem_hash_rescale(umem_cache_t *cp)
2411 {
2412 	umem_bufctl_t **old_table, **new_table, *bcp;
2413 	size_t old_size, new_size, h;
2414 
2415 	ASSERT(IN_UPDATE());
2416 
2417 	new_size = MAX(UMEM_HASH_INITIAL,
2418 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
2419 	old_size = cp->cache_hash_mask + 1;
2420 
2421 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
2422 		return;
2423 
2424 	new_table = vmem_alloc(umem_hash_arena, new_size * sizeof (void *),
2425 	    VM_NOSLEEP);
2426 	if (new_table == NULL)
2427 		return;
2428 	bzero(new_table, new_size * sizeof (void *));
2429 
2430 	(void) mutex_lock(&cp->cache_lock);
2431 
2432 	old_size = cp->cache_hash_mask + 1;
2433 	old_table = cp->cache_hash_table;
2434 
2435 	cp->cache_hash_mask = new_size - 1;
2436 	cp->cache_hash_table = new_table;
2437 	cp->cache_rescale++;
2438 
2439 	for (h = 0; h < old_size; h++) {
2440 		bcp = old_table[h];
2441 		while (bcp != NULL) {
2442 			void *addr = bcp->bc_addr;
2443 			umem_bufctl_t *next_bcp = bcp->bc_next;
2444 			umem_bufctl_t **hash_bucket = UMEM_HASH(cp, addr);
2445 			bcp->bc_next = *hash_bucket;
2446 			*hash_bucket = bcp;
2447 			bcp = next_bcp;
2448 		}
2449 	}
2450 
2451 	(void) mutex_unlock(&cp->cache_lock);
2452 
2453 	vmem_free(umem_hash_arena, old_table, old_size * sizeof (void *));
2454 }
2455 
2456 /*
2457  * Perform periodic maintenance on a cache: hash rescaling,
2458  * depot working-set update, and magazine resizing.
2459  */
2460 void
2461 umem_cache_update(umem_cache_t *cp)
2462 {
2463 	int update_flags = 0;
2464 
2465 	ASSERT(MUTEX_HELD(&umem_cache_lock));
2466 
2467 	/*
2468 	 * If the cache has become much larger or smaller than its hash table,
2469 	 * fire off a request to rescale the hash table.
2470 	 */
2471 	(void) mutex_lock(&cp->cache_lock);
2472 
2473 	if ((cp->cache_flags & UMF_HASH) &&
2474 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
2475 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
2476 	    cp->cache_hash_mask > UMEM_HASH_INITIAL)))
2477 		update_flags |= UMU_HASH_RESCALE;
2478 
2479 	(void) mutex_unlock(&cp->cache_lock);
2480 
2481 	/*
2482 	 * Update the depot working set statistics.
2483 	 */
2484 	umem_depot_ws_update(cp);
2485 
2486 	/*
2487 	 * If there's a lot of contention in the depot,
2488 	 * increase the magazine size.
2489 	 */
2490 	(void) mutex_lock(&cp->cache_depot_lock);
2491 
2492 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
2493 	    (int)(cp->cache_depot_contention -
2494 	    cp->cache_depot_contention_prev) > umem_depot_contention)
2495 		update_flags |= UMU_MAGAZINE_RESIZE;
2496 
2497 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
2498 
2499 	(void) mutex_unlock(&cp->cache_depot_lock);
2500 
2501 	if (update_flags)
2502 		umem_add_update(cp, update_flags);
2503 }
2504 
2505 /*
2506  * Runs all pending updates.
2507  *
2508  * The update lock must be held on entrance, and will be held on exit.
2509  */
2510 void
2511 umem_process_updates(void)
2512 {
2513 	ASSERT(MUTEX_HELD(&umem_update_lock));
2514 
2515 	while (umem_null_cache.cache_unext != &umem_null_cache) {
2516 		int notify = 0;
2517 		umem_cache_t *cp = umem_null_cache.cache_unext;
2518 
2519 		cp->cache_uprev->cache_unext = cp->cache_unext;
2520 		cp->cache_unext->cache_uprev = cp->cache_uprev;
2521 		cp->cache_uprev = cp->cache_unext = NULL;
2522 
2523 		ASSERT(!(cp->cache_uflags & UMU_ACTIVE));
2524 
2525 		while (cp->cache_uflags) {
2526 			int uflags = (cp->cache_uflags |= UMU_ACTIVE);
2527 			(void) mutex_unlock(&umem_update_lock);
2528 
2529 			/*
2530 			 * The order here is important.  Each step can speed up
2531 			 * later steps.
2532 			 */
2533 
2534 			if (uflags & UMU_HASH_RESCALE)
2535 				umem_hash_rescale(cp);
2536 
2537 			if (uflags & UMU_MAGAZINE_RESIZE)
2538 				umem_cache_magazine_resize(cp);
2539 
2540 			if (uflags & UMU_REAP)
2541 				umem_cache_reap(cp);
2542 
2543 			(void) mutex_lock(&umem_update_lock);
2544 
2545 			/*
2546 			 * check if anyone has requested notification
2547 			 */
2548 			if (cp->cache_uflags & UMU_NOTIFY) {
2549 				uflags |= UMU_NOTIFY;
2550 				notify = 1;
2551 			}
2552 			cp->cache_uflags &= ~uflags;
2553 		}
2554 		if (notify)
2555 			(void) cond_broadcast(&umem_update_cv);
2556 	}
2557 }
2558 
2559 #ifndef UMEM_STANDALONE
2560 static void
2561 umem_st_update(void)
2562 {
2563 	ASSERT(MUTEX_HELD(&umem_update_lock));
2564 	ASSERT(umem_update_thr == 0 && umem_st_update_thr == 0);
2565 
2566 	umem_st_update_thr = thr_self();
2567 
2568 	(void) mutex_unlock(&umem_update_lock);
2569 
2570 	vmem_update(NULL);
2571 	umem_cache_applyall(umem_cache_update);
2572 
2573 	(void) mutex_lock(&umem_update_lock);
2574 
2575 	umem_process_updates();	/* does all of the requested work */
2576 
2577 	umem_reap_next = gethrtime() +
2578 	    (hrtime_t)umem_reap_interval * NANOSEC;
2579 
2580 	umem_reaping = UMEM_REAP_DONE;
2581 
2582 	umem_st_update_thr = 0;
2583 }
2584 #endif
2585 
2586 /*
2587  * Reclaim all unused memory from all caches.  Called from vmem when memory
2588  * gets tight.  Must be called with no locks held.
2589  *
2590  * This just requests a reap on all caches, and notifies the update thread.
2591  */
2592 void
2593 umem_reap(void)
2594 {
2595 #ifndef UMEM_STANDALONE
2596 	extern int __nthreads(void);
2597 #endif
2598 
2599 	if (umem_ready != UMEM_READY || umem_reaping != UMEM_REAP_DONE ||
2600 	    gethrtime() < umem_reap_next)
2601 		return;
2602 
2603 	(void) mutex_lock(&umem_update_lock);
2604 
2605 	if (umem_reaping != UMEM_REAP_DONE || gethrtime() < umem_reap_next) {
2606 		(void) mutex_unlock(&umem_update_lock);
2607 		return;
2608 	}
2609 	umem_reaping = UMEM_REAP_ADDING;	/* lock out other reaps */
2610 
2611 	(void) mutex_unlock(&umem_update_lock);
2612 
2613 	umem_updateall(UMU_REAP);
2614 
2615 	(void) mutex_lock(&umem_update_lock);
2616 
2617 	umem_reaping = UMEM_REAP_ACTIVE;
2618 
2619 	/* Standalone is single-threaded */
2620 #ifndef UMEM_STANDALONE
2621 	if (umem_update_thr == 0) {
2622 		/*
2623 		 * The update thread does not exist.  If the process is
2624 		 * multi-threaded, create it.  If not, or the creation fails,
2625 		 * do the update processing inline.
2626 		 */
2627 		ASSERT(umem_st_update_thr == 0);
2628 
2629 		if (__nthreads() <= 1 || umem_create_update_thread() == 0)
2630 			umem_st_update();
2631 	}
2632 
2633 	(void) cond_broadcast(&umem_update_cv);	/* wake up the update thread */
2634 #endif
2635 
2636 	(void) mutex_unlock(&umem_update_lock);
2637 }
2638 
2639 umem_cache_t *
2640 umem_cache_create(
2641 	char *name,		/* descriptive name for this cache */
2642 	size_t bufsize,		/* size of the objects it manages */
2643 	size_t align,		/* required object alignment */
2644 	umem_constructor_t *constructor, /* object constructor */
2645 	umem_destructor_t *destructor, /* object destructor */
2646 	umem_reclaim_t *reclaim, /* memory reclaim callback */
2647 	void *private,		/* pass-thru arg for constr/destr/reclaim */
2648 	vmem_t *vmp,		/* vmem source for slab allocation */
2649 	int cflags)		/* cache creation flags */
2650 {
2651 	int cpu_seqid;
2652 	size_t chunksize;
2653 	umem_cache_t *cp, *cnext, *cprev;
2654 	umem_magtype_t *mtp;
2655 	size_t csize;
2656 	size_t phase;
2657 
2658 	/*
2659 	 * The init thread is allowed to create internal and quantum caches.
2660 	 *
2661 	 * Other threads must wait until until initialization is complete.
2662 	 */
2663 	if (umem_init_thr == thr_self())
2664 		ASSERT((cflags & (UMC_INTERNAL | UMC_QCACHE)) != 0);
2665 	else {
2666 		ASSERT(!(cflags & UMC_INTERNAL));
2667 		if (umem_ready != UMEM_READY && umem_init() == 0) {
2668 			errno = EAGAIN;
2669 			return (NULL);
2670 		}
2671 	}
2672 
2673 	csize = UMEM_CACHE_SIZE(umem_max_ncpus);
2674 	phase = P2NPHASE(csize, UMEM_CPU_CACHE_SIZE);
2675 
2676 	if (vmp == NULL)
2677 		vmp = umem_default_arena;
2678 
2679 	ASSERT(P2PHASE(phase, UMEM_ALIGN) == 0);
2680 
2681 	/*
2682 	 * Check that the arguments are reasonable
2683 	 */
2684 	if ((align & (align - 1)) != 0 || align > vmp->vm_quantum ||
2685 	    ((cflags & UMC_NOHASH) && (cflags & UMC_NOTOUCH)) ||
2686 	    name == NULL || bufsize == 0) {
2687 		errno = EINVAL;
2688 		return (NULL);
2689 	}
2690 
2691 	/*
2692 	 * If align == 0, we set it to the minimum required alignment.
2693 	 *
2694 	 * If align < UMEM_ALIGN, we round it up to UMEM_ALIGN, unless
2695 	 * UMC_NOTOUCH was passed.
2696 	 */
2697 	if (align == 0) {
2698 		if (P2ROUNDUP(bufsize, UMEM_ALIGN) >= UMEM_SECOND_ALIGN)
2699 			align = UMEM_SECOND_ALIGN;
2700 		else
2701 			align = UMEM_ALIGN;
2702 	} else if (align < UMEM_ALIGN && (cflags & UMC_NOTOUCH) == 0)
2703 		align = UMEM_ALIGN;
2704 
2705 
2706 	/*
2707 	 * Get a umem_cache structure.  We arrange that cp->cache_cpu[]
2708 	 * is aligned on a UMEM_CPU_CACHE_SIZE boundary to prevent
2709 	 * false sharing of per-CPU data.
2710 	 */
2711 	cp = vmem_xalloc(umem_cache_arena, csize, UMEM_CPU_CACHE_SIZE, phase,
2712 	    0, NULL, NULL, VM_NOSLEEP);
2713 
2714 	if (cp == NULL) {
2715 		errno = EAGAIN;
2716 		return (NULL);
2717 	}
2718 
2719 	bzero(cp, csize);
2720 
2721 	(void) mutex_lock(&umem_flags_lock);
2722 	if (umem_flags & UMF_RANDOMIZE)
2723 		umem_flags = (((umem_flags | ~UMF_RANDOM) + 1) & UMF_RANDOM) |
2724 		    UMF_RANDOMIZE;
2725 	cp->cache_flags = umem_flags | (cflags & UMF_DEBUG);
2726 	(void) mutex_unlock(&umem_flags_lock);
2727 
2728 	/*
2729 	 * Make sure all the various flags are reasonable.
2730 	 */
2731 	if (cp->cache_flags & UMF_LITE) {
2732 		if (bufsize >= umem_lite_minsize &&
2733 		    align <= umem_lite_maxalign &&
2734 		    P2PHASE(bufsize, umem_lite_maxalign) != 0) {
2735 			cp->cache_flags |= UMF_BUFTAG;
2736 			cp->cache_flags &= ~(UMF_AUDIT | UMF_FIREWALL);
2737 		} else {
2738 			cp->cache_flags &= ~UMF_DEBUG;
2739 		}
2740 	}
2741 
2742 	if ((cflags & UMC_QCACHE) && (cp->cache_flags & UMF_AUDIT))
2743 		cp->cache_flags |= UMF_NOMAGAZINE;
2744 
2745 	if (cflags & UMC_NODEBUG)
2746 		cp->cache_flags &= ~UMF_DEBUG;
2747 
2748 	if (cflags & UMC_NOTOUCH)
2749 		cp->cache_flags &= ~UMF_TOUCH;
2750 
2751 	if (cflags & UMC_NOHASH)
2752 		cp->cache_flags &= ~(UMF_AUDIT | UMF_FIREWALL);
2753 
2754 	if (cflags & UMC_NOMAGAZINE)
2755 		cp->cache_flags |= UMF_NOMAGAZINE;
2756 
2757 	if ((cp->cache_flags & UMF_AUDIT) && !(cflags & UMC_NOTOUCH))
2758 		cp->cache_flags |= UMF_REDZONE;
2759 
2760 	if ((cp->cache_flags & UMF_BUFTAG) && bufsize >= umem_minfirewall &&
2761 	    !(cp->cache_flags & UMF_LITE) && !(cflags & UMC_NOHASH))
2762 		cp->cache_flags |= UMF_FIREWALL;
2763 
2764 	if (vmp != umem_default_arena || umem_firewall_arena == NULL)
2765 		cp->cache_flags &= ~UMF_FIREWALL;
2766 
2767 	if (cp->cache_flags & UMF_FIREWALL) {
2768 		cp->cache_flags &= ~UMF_BUFTAG;
2769 		cp->cache_flags |= UMF_NOMAGAZINE;
2770 		ASSERT(vmp == umem_default_arena);
2771 		vmp = umem_firewall_arena;
2772 	}
2773 
2774 	/*
2775 	 * Set cache properties.
2776 	 */
2777 	(void) strncpy(cp->cache_name, name, sizeof (cp->cache_name) - 1);
2778 	cp->cache_bufsize = bufsize;
2779 	cp->cache_align = align;
2780 	cp->cache_constructor = constructor;
2781 	cp->cache_destructor = destructor;
2782 	cp->cache_reclaim = reclaim;
2783 	cp->cache_private = private;
2784 	cp->cache_arena = vmp;
2785 	cp->cache_cflags = cflags;
2786 	cp->cache_cpu_mask = umem_cpu_mask;
2787 
2788 	/*
2789 	 * Determine the chunk size.
2790 	 */
2791 	chunksize = bufsize;
2792 
2793 	if (align >= UMEM_ALIGN) {
2794 		chunksize = P2ROUNDUP(chunksize, UMEM_ALIGN);
2795 		cp->cache_bufctl = chunksize - UMEM_ALIGN;
2796 	}
2797 
2798 	if (cp->cache_flags & UMF_BUFTAG) {
2799 		cp->cache_bufctl = chunksize;
2800 		cp->cache_buftag = chunksize;
2801 		chunksize += sizeof (umem_buftag_t);
2802 	}
2803 
2804 	if (cp->cache_flags & UMF_DEADBEEF) {
2805 		cp->cache_verify = MIN(cp->cache_buftag, umem_maxverify);
2806 		if (cp->cache_flags & UMF_LITE)
2807 			cp->cache_verify = MIN(cp->cache_verify, UMEM_ALIGN);
2808 	}
2809 
2810 	cp->cache_contents = MIN(cp->cache_bufctl, umem_content_maxsave);
2811 
2812 	cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
2813 
2814 	if (chunksize < bufsize) {
2815 		errno = ENOMEM;
2816 		goto fail;
2817 	}
2818 
2819 	/*
2820 	 * Now that we know the chunk size, determine the optimal slab size.
2821 	 */
2822 	if (vmp == umem_firewall_arena) {
2823 		cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
2824 		cp->cache_mincolor = cp->cache_slabsize - chunksize;
2825 		cp->cache_maxcolor = cp->cache_mincolor;
2826 		cp->cache_flags |= UMF_HASH;
2827 		ASSERT(!(cp->cache_flags & UMF_BUFTAG));
2828 	} else if ((cflags & UMC_NOHASH) || (!(cflags & UMC_NOTOUCH) &&
2829 	    !(cp->cache_flags & UMF_AUDIT) &&
2830 	    chunksize < vmp->vm_quantum / UMEM_VOID_FRACTION)) {
2831 		cp->cache_slabsize = vmp->vm_quantum;
2832 		cp->cache_mincolor = 0;
2833 		cp->cache_maxcolor =
2834 		    (cp->cache_slabsize - sizeof (umem_slab_t)) % chunksize;
2835 
2836 		if (chunksize + sizeof (umem_slab_t) > cp->cache_slabsize) {
2837 			errno = EINVAL;
2838 			goto fail;
2839 		}
2840 		ASSERT(!(cp->cache_flags & UMF_AUDIT));
2841 	} else {
2842 		size_t chunks, bestfit, waste, slabsize;
2843 		size_t minwaste = LONG_MAX;
2844 
2845 		for (chunks = 1; chunks <= UMEM_VOID_FRACTION; chunks++) {
2846 			slabsize = P2ROUNDUP(chunksize * chunks,
2847 			    vmp->vm_quantum);
2848 			/*
2849 			 * check for overflow
2850 			 */
2851 			if ((slabsize / chunks) < chunksize) {
2852 				errno = ENOMEM;
2853 				goto fail;
2854 			}
2855 			chunks = slabsize / chunksize;
2856 			waste = (slabsize % chunksize) / chunks;
2857 			if (waste < minwaste) {
2858 				minwaste = waste;
2859 				bestfit = slabsize;
2860 			}
2861 		}
2862 		if (cflags & UMC_QCACHE)
2863 			bestfit = MAX(1 << highbit(3 * vmp->vm_qcache_max), 64);
2864 		cp->cache_slabsize = bestfit;
2865 		cp->cache_mincolor = 0;
2866 		cp->cache_maxcolor = bestfit % chunksize;
2867 		cp->cache_flags |= UMF_HASH;
2868 	}
2869 
2870 	if (cp->cache_flags & UMF_HASH) {
2871 		ASSERT(!(cflags & UMC_NOHASH));
2872 		cp->cache_bufctl_cache = (cp->cache_flags & UMF_AUDIT) ?
2873 		    umem_bufctl_audit_cache : umem_bufctl_cache;
2874 	}
2875 
2876 	if (cp->cache_maxcolor >= vmp->vm_quantum)
2877 		cp->cache_maxcolor = vmp->vm_quantum - 1;
2878 
2879 	cp->cache_color = cp->cache_mincolor;
2880 
2881 	/*
2882 	 * Initialize the rest of the slab layer.
2883 	 */
2884 	(void) mutex_init(&cp->cache_lock, USYNC_THREAD, NULL);
2885 
2886 	cp->cache_freelist = &cp->cache_nullslab;
2887 	cp->cache_nullslab.slab_cache = cp;
2888 	cp->cache_nullslab.slab_refcnt = -1;
2889 	cp->cache_nullslab.slab_next = &cp->cache_nullslab;
2890 	cp->cache_nullslab.slab_prev = &cp->cache_nullslab;
2891 
2892 	if (cp->cache_flags & UMF_HASH) {
2893 		cp->cache_hash_table = vmem_alloc(umem_hash_arena,
2894 		    UMEM_HASH_INITIAL * sizeof (void *), VM_NOSLEEP);
2895 		if (cp->cache_hash_table == NULL) {
2896 			errno = EAGAIN;
2897 			goto fail_lock;
2898 		}
2899 		bzero(cp->cache_hash_table,
2900 		    UMEM_HASH_INITIAL * sizeof (void *));
2901 		cp->cache_hash_mask = UMEM_HASH_INITIAL - 1;
2902 		cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
2903 	}
2904 
2905 	/*
2906 	 * Initialize the depot.
2907 	 */
2908 	(void) mutex_init(&cp->cache_depot_lock, USYNC_THREAD, NULL);
2909 
2910 	for (mtp = umem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
2911 		continue;
2912 
2913 	cp->cache_magtype = mtp;
2914 
2915 	/*
2916 	 * Initialize the CPU layer.
2917 	 */
2918 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++) {
2919 		umem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
2920 		(void) mutex_init(&ccp->cc_lock, USYNC_THREAD, NULL);
2921 		ccp->cc_flags = cp->cache_flags;
2922 		ccp->cc_rounds = -1;
2923 		ccp->cc_prounds = -1;
2924 	}
2925 
2926 	/*
2927 	 * Add the cache to the global list.  This makes it visible
2928 	 * to umem_update(), so the cache must be ready for business.
2929 	 */
2930 	(void) mutex_lock(&umem_cache_lock);
2931 	cp->cache_next = cnext = &umem_null_cache;
2932 	cp->cache_prev = cprev = umem_null_cache.cache_prev;
2933 	cnext->cache_prev = cp;
2934 	cprev->cache_next = cp;
2935 	(void) mutex_unlock(&umem_cache_lock);
2936 
2937 	if (umem_ready == UMEM_READY)
2938 		umem_cache_magazine_enable(cp);
2939 
2940 	return (cp);
2941 
2942 fail_lock:
2943 	(void) mutex_destroy(&cp->cache_lock);
2944 fail:
2945 	vmem_xfree(umem_cache_arena, cp, csize);
2946 	return (NULL);
2947 }
2948 
2949 void
2950 umem_cache_destroy(umem_cache_t *cp)
2951 {
2952 	int cpu_seqid;
2953 
2954 	/*
2955 	 * Remove the cache from the global cache list so that no new updates
2956 	 * will be scheduled on its behalf, wait for any pending tasks to
2957 	 * complete, purge the cache, and then destroy it.
2958 	 */
2959 	(void) mutex_lock(&umem_cache_lock);
2960 	cp->cache_prev->cache_next = cp->cache_next;
2961 	cp->cache_next->cache_prev = cp->cache_prev;
2962 	cp->cache_prev = cp->cache_next = NULL;
2963 	(void) mutex_unlock(&umem_cache_lock);
2964 
2965 	umem_remove_updates(cp);
2966 
2967 	umem_cache_magazine_purge(cp);
2968 
2969 	(void) mutex_lock(&cp->cache_lock);
2970 	if (cp->cache_buftotal != 0)
2971 		log_message("umem_cache_destroy: '%s' (%p) not empty\n",
2972 		    cp->cache_name, (void *)cp);
2973 	cp->cache_reclaim = NULL;
2974 	/*
2975 	 * The cache is now dead.  There should be no further activity.
2976 	 * We enforce this by setting land mines in the constructor and
2977 	 * destructor routines that induce a segmentation fault if invoked.
2978 	 */
2979 	cp->cache_constructor = (umem_constructor_t *)1;
2980 	cp->cache_destructor = (umem_destructor_t *)2;
2981 	(void) mutex_unlock(&cp->cache_lock);
2982 
2983 	if (cp->cache_hash_table != NULL)
2984 		vmem_free(umem_hash_arena, cp->cache_hash_table,
2985 		    (cp->cache_hash_mask + 1) * sizeof (void *));
2986 
2987 	for (cpu_seqid = 0; cpu_seqid < umem_max_ncpus; cpu_seqid++)
2988 		(void) mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
2989 
2990 	(void) mutex_destroy(&cp->cache_depot_lock);
2991 	(void) mutex_destroy(&cp->cache_lock);
2992 
2993 	vmem_free(umem_cache_arena, cp, UMEM_CACHE_SIZE(umem_max_ncpus));
2994 }
2995 
2996 void
2997 umem_alloc_sizes_clear(void)
2998 {
2999 	int i;
3000 
3001 	umem_alloc_sizes[0] = UMEM_MAXBUF;
3002 	for (i = 1; i < NUM_ALLOC_SIZES; i++)
3003 		umem_alloc_sizes[i] = 0;
3004 }
3005 
3006 void
3007 umem_alloc_sizes_add(size_t size_arg)
3008 {
3009 	int i, j;
3010 	size_t size = size_arg;
3011 
3012 	if (size == 0) {
3013 		log_message("size_add: cannot add zero-sized cache\n",
3014 		    size, UMEM_MAXBUF);
3015 		return;
3016 	}
3017 
3018 	if (size > UMEM_MAXBUF) {
3019 		log_message("size_add: %ld > %d, cannot add\n", size,
3020 		    UMEM_MAXBUF);
3021 		return;
3022 	}
3023 
3024 	if (umem_alloc_sizes[NUM_ALLOC_SIZES - 1] != 0) {
3025 		log_message("size_add: no space in alloc_table for %d\n",
3026 		    size);
3027 		return;
3028 	}
3029 
3030 	if (P2PHASE(size, UMEM_ALIGN) != 0) {
3031 		size = P2ROUNDUP(size, UMEM_ALIGN);
3032 		log_message("size_add: rounding %d up to %d\n", size_arg,
3033 		    size);
3034 	}
3035 
3036 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
3037 		int cur = umem_alloc_sizes[i];
3038 		if (cur == size) {
3039 			log_message("size_add: %ld already in table\n",
3040 			    size);
3041 			return;
3042 		}
3043 		if (cur > size)
3044 			break;
3045 	}
3046 
3047 	for (j = NUM_ALLOC_SIZES - 1; j > i; j--)
3048 		umem_alloc_sizes[j] = umem_alloc_sizes[j-1];
3049 	umem_alloc_sizes[i] = size;
3050 }
3051 
3052 void
3053 umem_alloc_sizes_remove(size_t size)
3054 {
3055 	int i;
3056 
3057 	if (size == UMEM_MAXBUF) {
3058 		log_message("size_remove: cannot remove %ld\n", size);
3059 		return;
3060 	}
3061 
3062 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
3063 		int cur = umem_alloc_sizes[i];
3064 		if (cur == size)
3065 			break;
3066 		else if (cur > size || cur == 0) {
3067 			log_message("size_remove: %ld not found in table\n",
3068 			    size);
3069 			return;
3070 		}
3071 	}
3072 
3073 	for (; i + 1 < NUM_ALLOC_SIZES; i++)
3074 		umem_alloc_sizes[i] = umem_alloc_sizes[i+1];
3075 	umem_alloc_sizes[i] = 0;
3076 }
3077 
3078 /*
3079  * We've been called back from libc to indicate that thread is terminating and
3080  * that it needs to release the per-thread memory that it has. We get to know
3081  * which entry in the thread's tmem array the allocation came from. Currently
3082  * this refers to first n umem_caches which makes this a pretty simple indexing
3083  * job.
3084  */
3085 static void
3086 umem_cache_tmem_cleanup(void *buf, int entry)
3087 {
3088 	size_t size;
3089 	umem_cache_t *cp;
3090 
3091 	size = umem_alloc_sizes[entry];
3092 	cp = umem_alloc_table[(size - 1) >> UMEM_ALIGN_SHIFT];
3093 	_umem_cache_free(cp, buf);
3094 }
3095 
3096 static int
3097 umem_cache_init(void)
3098 {
3099 	int i;
3100 	size_t size, max_size;
3101 	umem_cache_t *cp;
3102 	umem_magtype_t *mtp;
3103 	char name[UMEM_CACHE_NAMELEN + 1];
3104 	umem_cache_t *umem_alloc_caches[NUM_ALLOC_SIZES];
3105 
3106 	for (i = 0; i < sizeof (umem_magtype) / sizeof (*mtp); i++) {
3107 		mtp = &umem_magtype[i];
3108 		(void) snprintf(name, sizeof (name), "umem_magazine_%d",
3109 		    mtp->mt_magsize);
3110 		mtp->mt_cache = umem_cache_create(name,
3111 		    (mtp->mt_magsize + 1) * sizeof (void *),
3112 		    mtp->mt_align, NULL, NULL, NULL, NULL,
3113 		    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
3114 		if (mtp->mt_cache == NULL)
3115 			return (0);
3116 	}
3117 
3118 	umem_slab_cache = umem_cache_create("umem_slab_cache",
3119 	    sizeof (umem_slab_t), 0, NULL, NULL, NULL, NULL,
3120 	    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
3121 
3122 	if (umem_slab_cache == NULL)
3123 		return (0);
3124 
3125 	umem_bufctl_cache = umem_cache_create("umem_bufctl_cache",
3126 	    sizeof (umem_bufctl_t), 0, NULL, NULL, NULL, NULL,
3127 	    umem_internal_arena, UMC_NOHASH | UMC_INTERNAL);
3128 
3129 	if (umem_bufctl_cache == NULL)
3130 		return (0);
3131 
3132 	/*
3133 	 * The size of the umem_bufctl_audit structure depends upon
3134 	 * umem_stack_depth.   See umem_impl.h for details on the size
3135 	 * restrictions.
3136 	 */
3137 
3138 	size = UMEM_BUFCTL_AUDIT_SIZE_DEPTH(umem_stack_depth);
3139 	max_size = UMEM_BUFCTL_AUDIT_MAX_SIZE;
3140 
3141 	if (size > max_size) {			/* too large -- truncate */
3142 		int max_frames = UMEM_MAX_STACK_DEPTH;
3143 
3144 		ASSERT(UMEM_BUFCTL_AUDIT_SIZE_DEPTH(max_frames) <= max_size);
3145 
3146 		umem_stack_depth = max_frames;
3147 		size = UMEM_BUFCTL_AUDIT_SIZE_DEPTH(umem_stack_depth);
3148 	}
3149 
3150 	umem_bufctl_audit_cache = umem_cache_create("umem_bufctl_audit_cache",
3151 	    size, 0, NULL, NULL, NULL, NULL, umem_internal_arena,
3152 	    UMC_NOHASH | UMC_INTERNAL);
3153 
3154 	if (umem_bufctl_audit_cache == NULL)
3155 		return (0);
3156 
3157 	if (vmem_backend & VMEM_BACKEND_MMAP)
3158 		umem_va_arena = vmem_create("umem_va",
3159 		    NULL, 0, pagesize,
3160 		    vmem_alloc, vmem_free, heap_arena,
3161 		    8 * pagesize, VM_NOSLEEP);
3162 	else
3163 		umem_va_arena = heap_arena;
3164 
3165 	if (umem_va_arena == NULL)
3166 		return (0);
3167 
3168 	umem_default_arena = vmem_create("umem_default",
3169 	    NULL, 0, pagesize,
3170 	    heap_alloc, heap_free, umem_va_arena,
3171 	    0, VM_NOSLEEP);
3172 
3173 	if (umem_default_arena == NULL)
3174 		return (0);
3175 
3176 	/*
3177 	 * make sure the umem_alloc table initializer is correct
3178 	 */
3179 	i = sizeof (umem_alloc_table) / sizeof (*umem_alloc_table);
3180 	ASSERT(umem_alloc_table[i - 1] == &umem_null_cache);
3181 
3182 	/*
3183 	 * Create the default caches to back umem_alloc()
3184 	 */
3185 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
3186 		size_t cache_size = umem_alloc_sizes[i];
3187 		size_t align = 0;
3188 
3189 		if (cache_size == 0)
3190 			break;		/* 0 terminates the list */
3191 
3192 		/*
3193 		 * If they allocate a multiple of the coherency granularity,
3194 		 * they get a coherency-granularity-aligned address.
3195 		 */
3196 		if (IS_P2ALIGNED(cache_size, 64))
3197 			align = 64;
3198 		if (IS_P2ALIGNED(cache_size, pagesize))
3199 			align = pagesize;
3200 		(void) snprintf(name, sizeof (name), "umem_alloc_%lu",
3201 		    (long)cache_size);
3202 
3203 		cp = umem_cache_create(name, cache_size, align,
3204 		    NULL, NULL, NULL, NULL, NULL, UMC_INTERNAL);
3205 		if (cp == NULL)
3206 			return (0);
3207 
3208 		umem_alloc_caches[i] = cp;
3209 	}
3210 
3211 	umem_tmem_off = _tmem_get_base();
3212 	_tmem_set_cleanup(umem_cache_tmem_cleanup);
3213 
3214 	if (umem_genasm_supported && !(umem_flags & UMF_DEBUG) &&
3215 	    !(umem_flags & UMF_NOMAGAZINE) &&
3216 	    umem_ptc_size > 0) {
3217 		umem_ptc_enabled = umem_genasm(umem_alloc_sizes,
3218 		    umem_alloc_caches, i) == 0 ? 1 : 0;
3219 	}
3220 
3221 	/*
3222 	 * Initialization cannot fail at this point.  Make the caches
3223 	 * visible to umem_alloc() and friends.
3224 	 */
3225 	size = UMEM_ALIGN;
3226 	for (i = 0; i < NUM_ALLOC_SIZES; i++) {
3227 		size_t cache_size = umem_alloc_sizes[i];
3228 
3229 		if (cache_size == 0)
3230 			break;		/* 0 terminates the list */
3231 
3232 		cp = umem_alloc_caches[i];
3233 
3234 		while (size <= cache_size) {
3235 			umem_alloc_table[(size - 1) >> UMEM_ALIGN_SHIFT] = cp;
3236 			size += UMEM_ALIGN;
3237 		}
3238 	}
3239 	ASSERT(size - UMEM_ALIGN == UMEM_MAXBUF);
3240 	return (1);
3241 }
3242 
3243 /*
3244  * umem_startup() is called early on, and must be called explicitly if we're
3245  * the standalone version.
3246  */
3247 #ifdef UMEM_STANDALONE
3248 void
3249 #else
3250 #pragma init(umem_startup)
3251 static void
3252 #endif
3253 umem_startup(caddr_t start, size_t len, size_t pagesize, caddr_t minstack,
3254     caddr_t maxstack)
3255 {
3256 #ifdef UMEM_STANDALONE
3257 	int idx;
3258 	/* Standalone doesn't fork */
3259 #else
3260 	umem_forkhandler_init(); /* register the fork handler */
3261 #endif
3262 
3263 #ifdef __lint
3264 	/* make lint happy */
3265 	minstack = maxstack;
3266 #endif
3267 
3268 #ifdef UMEM_STANDALONE
3269 	umem_ready = UMEM_READY_STARTUP;
3270 	umem_init_env_ready = 0;
3271 
3272 	umem_min_stack = minstack;
3273 	umem_max_stack = maxstack;
3274 
3275 	nofail_callback = NULL;
3276 	umem_slab_cache = NULL;
3277 	umem_bufctl_cache = NULL;
3278 	umem_bufctl_audit_cache = NULL;
3279 	heap_arena = NULL;
3280 	heap_alloc = NULL;
3281 	heap_free = NULL;
3282 	umem_internal_arena = NULL;
3283 	umem_cache_arena = NULL;
3284 	umem_hash_arena = NULL;
3285 	umem_log_arena = NULL;
3286 	umem_oversize_arena = NULL;
3287 	umem_va_arena = NULL;
3288 	umem_default_arena = NULL;
3289 	umem_firewall_va_arena = NULL;
3290 	umem_firewall_arena = NULL;
3291 	umem_memalign_arena = NULL;
3292 	umem_transaction_log = NULL;
3293 	umem_content_log = NULL;
3294 	umem_failure_log = NULL;
3295 	umem_slab_log = NULL;
3296 	umem_cpu_mask = 0;
3297 
3298 	umem_cpus = &umem_startup_cpu;
3299 	umem_startup_cpu.cpu_cache_offset = UMEM_CACHE_SIZE(0);
3300 	umem_startup_cpu.cpu_number = 0;
3301 
3302 	bcopy(&umem_null_cache_template, &umem_null_cache,
3303 	    sizeof (umem_cache_t));
3304 
3305 	for (idx = 0; idx < (UMEM_MAXBUF >> UMEM_ALIGN_SHIFT); idx++)
3306 		umem_alloc_table[idx] = &umem_null_cache;
3307 #endif
3308 
3309 	/*
3310 	 * Perform initialization specific to the way we've been compiled
3311 	 * (library or standalone)
3312 	 */
3313 	umem_type_init(start, len, pagesize);
3314 
3315 	vmem_startup();
3316 }
3317 
3318 int
3319 umem_init(void)
3320 {
3321 	size_t maxverify, minfirewall;
3322 	size_t size;
3323 	int idx;
3324 	umem_cpu_t *new_cpus;
3325 
3326 	vmem_t *memalign_arena, *oversize_arena;
3327 
3328 	if (thr_self() != umem_init_thr) {
3329 		/*
3330 		 * The usual case -- non-recursive invocation of umem_init().
3331 		 */
3332 		(void) mutex_lock(&umem_init_lock);
3333 		if (umem_ready != UMEM_READY_STARTUP) {
3334 			/*
3335 			 * someone else beat us to initializing umem.  Wait
3336 			 * for them to complete, then return.
3337 			 */
3338 			while (umem_ready == UMEM_READY_INITING) {
3339 				int cancel_state;
3340 
3341 				(void) pthread_setcancelstate(
3342 				    PTHREAD_CANCEL_DISABLE, &cancel_state);
3343 				(void) cond_wait(&umem_init_cv,
3344 				    &umem_init_lock);
3345 				(void) pthread_setcancelstate(
3346 				    cancel_state, NULL);
3347 			}
3348 			ASSERT(umem_ready == UMEM_READY ||
3349 			    umem_ready == UMEM_READY_INIT_FAILED);
3350 			(void) mutex_unlock(&umem_init_lock);
3351 			return (umem_ready == UMEM_READY);
3352 		}
3353 
3354 		ASSERT(umem_ready == UMEM_READY_STARTUP);
3355 		ASSERT(umem_init_env_ready == 0);
3356 
3357 		umem_ready = UMEM_READY_INITING;
3358 		umem_init_thr = thr_self();
3359 
3360 		(void) mutex_unlock(&umem_init_lock);
3361 		umem_setup_envvars(0);		/* can recurse -- see below */
3362 		if (umem_init_env_ready) {
3363 			/*
3364 			 * initialization was completed already
3365 			 */
3366 			ASSERT(umem_ready == UMEM_READY ||
3367 			    umem_ready == UMEM_READY_INIT_FAILED);
3368 			ASSERT(umem_init_thr == 0);
3369 			return (umem_ready == UMEM_READY);
3370 		}
3371 	} else if (!umem_init_env_ready) {
3372 		/*
3373 		 * The umem_setup_envvars() call (above) makes calls into
3374 		 * the dynamic linker and directly into user-supplied code.
3375 		 * Since we cannot know what that code will do, we could be
3376 		 * recursively invoked (by, say, a malloc() call in the code
3377 		 * itself, or in a (C++) _init section it causes to be fired).
3378 		 *
3379 		 * This code is where we end up if such recursion occurs.  We
3380 		 * first clean up any partial results in the envvar code, then
3381 		 * proceed to finish initialization processing in the recursive
3382 		 * call.  The original call will notice this, and return
3383 		 * immediately.
3384 		 */
3385 		umem_setup_envvars(1);		/* clean up any partial state */
3386 	} else {
3387 		umem_panic(
3388 		    "recursive allocation while initializing umem\n");
3389 	}
3390 	umem_init_env_ready = 1;
3391 
3392 	/*
3393 	 * From this point until we finish, recursion into umem_init() will
3394 	 * cause a umem_panic().
3395 	 */
3396 	maxverify = minfirewall = ULONG_MAX;
3397 
3398 	/* LINTED constant condition */
3399 	if (sizeof (umem_cpu_cache_t) != UMEM_CPU_CACHE_SIZE) {
3400 		umem_panic("sizeof (umem_cpu_cache_t) = %d, should be %d\n",
3401 		    sizeof (umem_cpu_cache_t), UMEM_CPU_CACHE_SIZE);
3402 	}
3403 
3404 	umem_max_ncpus = umem_get_max_ncpus();
3405 
3406 	/*
3407 	 * load tunables from environment
3408 	 */
3409 	umem_process_envvars();
3410 
3411 	if (issetugid())
3412 		umem_mtbf = 0;
3413 
3414 	/*
3415 	 * set up vmem
3416 	 */
3417 	if (!(umem_flags & UMF_AUDIT))
3418 		vmem_no_debug();
3419 
3420 	heap_arena = vmem_heap_arena(&heap_alloc, &heap_free);
3421 
3422 	pagesize = heap_arena->vm_quantum;
3423 
3424 	umem_internal_arena = vmem_create("umem_internal", NULL, 0, pagesize,
3425 	    heap_alloc, heap_free, heap_arena, 0, VM_NOSLEEP);
3426 
3427 	umem_default_arena = umem_internal_arena;
3428 
3429 	if (umem_internal_arena == NULL)
3430 		goto fail;
3431 
3432 	umem_cache_arena = vmem_create("umem_cache", NULL, 0, UMEM_ALIGN,
3433 	    vmem_alloc, vmem_free, umem_internal_arena, 0, VM_NOSLEEP);
3434 
3435 	umem_hash_arena = vmem_create("umem_hash", NULL, 0, UMEM_ALIGN,
3436 	    vmem_alloc, vmem_free, umem_internal_arena, 0, VM_NOSLEEP);
3437 
3438 	umem_log_arena = vmem_create("umem_log", NULL, 0, UMEM_ALIGN,
3439 	    heap_alloc, heap_free, heap_arena, 0, VM_NOSLEEP);
3440 
3441 	umem_firewall_va_arena = vmem_create("umem_firewall_va",
3442 	    NULL, 0, pagesize,
3443 	    umem_firewall_va_alloc, umem_firewall_va_free, heap_arena,
3444 	    0, VM_NOSLEEP);
3445 
3446 	if (umem_cache_arena == NULL || umem_hash_arena == NULL ||
3447 	    umem_log_arena == NULL || umem_firewall_va_arena == NULL)
3448 		goto fail;
3449 
3450 	umem_firewall_arena = vmem_create("umem_firewall", NULL, 0, pagesize,
3451 	    heap_alloc, heap_free, umem_firewall_va_arena, 0,
3452 	    VM_NOSLEEP);
3453 
3454 	if (umem_firewall_arena == NULL)
3455 		goto fail;
3456 
3457 	oversize_arena = vmem_create("umem_oversize", NULL, 0, pagesize,
3458 	    heap_alloc, heap_free, minfirewall < ULONG_MAX ?
3459 	    umem_firewall_va_arena : heap_arena, 0, VM_NOSLEEP);
3460 
3461 	memalign_arena = vmem_create("umem_memalign", NULL, 0, UMEM_ALIGN,
3462 	    heap_alloc, heap_free, minfirewall < ULONG_MAX ?
3463 	    umem_firewall_va_arena : heap_arena, 0, VM_NOSLEEP);
3464 
3465 	if (oversize_arena == NULL || memalign_arena == NULL)
3466 		goto fail;
3467 
3468 	if (umem_max_ncpus > CPUHINT_MAX())
3469 		umem_max_ncpus = CPUHINT_MAX();
3470 
3471 	while ((umem_max_ncpus & (umem_max_ncpus - 1)) != 0)
3472 		umem_max_ncpus++;
3473 
3474 	if (umem_max_ncpus == 0)
3475 		umem_max_ncpus = 1;
3476 
3477 	size = umem_max_ncpus * sizeof (umem_cpu_t);
3478 	new_cpus = vmem_alloc(umem_internal_arena, size, VM_NOSLEEP);
3479 	if (new_cpus == NULL)
3480 		goto fail;
3481 
3482 	bzero(new_cpus, size);
3483 	for (idx = 0; idx < umem_max_ncpus; idx++) {
3484 		new_cpus[idx].cpu_number = idx;
3485 		new_cpus[idx].cpu_cache_offset = UMEM_CACHE_SIZE(idx);
3486 	}
3487 	umem_cpus = new_cpus;
3488 	umem_cpu_mask = (umem_max_ncpus - 1);
3489 
3490 	if (umem_maxverify == 0)
3491 		umem_maxverify = maxverify;
3492 
3493 	if (umem_minfirewall == 0)
3494 		umem_minfirewall = minfirewall;
3495 
3496 	/*
3497 	 * Set up updating and reaping
3498 	 */
3499 	umem_reap_next = gethrtime() + NANOSEC;
3500 
3501 #ifndef UMEM_STANDALONE
3502 	(void) gettimeofday(&umem_update_next, NULL);
3503 #endif
3504 
3505 	/*
3506 	 * Set up logging -- failure here is okay, since it will just disable
3507 	 * the logs
3508 	 */
3509 	if (umem_logging) {
3510 		umem_transaction_log = umem_log_init(umem_transaction_log_size);
3511 		umem_content_log = umem_log_init(umem_content_log_size);
3512 		umem_failure_log = umem_log_init(umem_failure_log_size);
3513 		umem_slab_log = umem_log_init(umem_slab_log_size);
3514 	}
3515 
3516 	/*
3517 	 * Set up caches -- if successful, initialization cannot fail, since
3518 	 * allocations from other threads can now succeed.
3519 	 */
3520 	if (umem_cache_init() == 0) {
3521 		log_message("unable to create initial caches\n");
3522 		goto fail;
3523 	}
3524 	umem_oversize_arena = oversize_arena;
3525 	umem_memalign_arena = memalign_arena;
3526 
3527 	umem_cache_applyall(umem_cache_magazine_enable);
3528 
3529 	/*
3530 	 * initialization done, ready to go
3531 	 */
3532 	(void) mutex_lock(&umem_init_lock);
3533 	umem_ready = UMEM_READY;
3534 	umem_init_thr = 0;
3535 	(void) cond_broadcast(&umem_init_cv);
3536 	(void) mutex_unlock(&umem_init_lock);
3537 	return (1);
3538 
3539 fail:
3540 	log_message("umem initialization failed\n");
3541 
3542 	(void) mutex_lock(&umem_init_lock);
3543 	umem_ready = UMEM_READY_INIT_FAILED;
3544 	umem_init_thr = 0;
3545 	(void) cond_broadcast(&umem_init_cv);
3546 	(void) mutex_unlock(&umem_init_lock);
3547 	return (0);
3548 }
3549