xref: /titanic_41/usr/src/uts/common/os/kmem.c (revision 9f1b636ab7ec84b6553f369641dc9f8414ec0e7e)
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  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 /*
29  * Kernel memory allocator, as described in the following two papers:
30  *
31  * Jeff Bonwick,
32  * The Slab Allocator: An Object-Caching Kernel Memory Allocator.
33  * Proceedings of the Summer 1994 Usenix Conference.
34  * Available as /shared/sac/PSARC/1994/028/materials/kmem.pdf.
35  *
36  * Jeff Bonwick and Jonathan Adams,
37  * Magazines and vmem: Extending the Slab Allocator to Many CPUs and
38  * Arbitrary Resources.
39  * Proceedings of the 2001 Usenix Conference.
40  * Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
41  */
42 
43 #include <sys/kmem_impl.h>
44 #include <sys/vmem_impl.h>
45 #include <sys/param.h>
46 #include <sys/sysmacros.h>
47 #include <sys/vm.h>
48 #include <sys/proc.h>
49 #include <sys/tuneable.h>
50 #include <sys/systm.h>
51 #include <sys/cmn_err.h>
52 #include <sys/debug.h>
53 #include <sys/mutex.h>
54 #include <sys/bitmap.h>
55 #include <sys/atomic.h>
56 #include <sys/kobj.h>
57 #include <sys/disp.h>
58 #include <vm/seg_kmem.h>
59 #include <sys/log.h>
60 #include <sys/callb.h>
61 #include <sys/taskq.h>
62 #include <sys/modctl.h>
63 #include <sys/reboot.h>
64 #include <sys/id32.h>
65 #include <sys/zone.h>
66 #include <sys/netstack.h>
67 
68 extern void streams_msg_init(void);
69 extern int segkp_fromheap;
70 extern void segkp_cache_free(void);
71 
72 struct kmem_cache_kstat {
73 	kstat_named_t	kmc_buf_size;
74 	kstat_named_t	kmc_align;
75 	kstat_named_t	kmc_chunk_size;
76 	kstat_named_t	kmc_slab_size;
77 	kstat_named_t	kmc_alloc;
78 	kstat_named_t	kmc_alloc_fail;
79 	kstat_named_t	kmc_free;
80 	kstat_named_t	kmc_depot_alloc;
81 	kstat_named_t	kmc_depot_free;
82 	kstat_named_t	kmc_depot_contention;
83 	kstat_named_t	kmc_slab_alloc;
84 	kstat_named_t	kmc_slab_free;
85 	kstat_named_t	kmc_buf_constructed;
86 	kstat_named_t	kmc_buf_avail;
87 	kstat_named_t	kmc_buf_inuse;
88 	kstat_named_t	kmc_buf_total;
89 	kstat_named_t	kmc_buf_max;
90 	kstat_named_t	kmc_slab_create;
91 	kstat_named_t	kmc_slab_destroy;
92 	kstat_named_t	kmc_vmem_source;
93 	kstat_named_t	kmc_hash_size;
94 	kstat_named_t	kmc_hash_lookup_depth;
95 	kstat_named_t	kmc_hash_rescale;
96 	kstat_named_t	kmc_full_magazines;
97 	kstat_named_t	kmc_empty_magazines;
98 	kstat_named_t	kmc_magazine_size;
99 } kmem_cache_kstat = {
100 	{ "buf_size",		KSTAT_DATA_UINT64 },
101 	{ "align",		KSTAT_DATA_UINT64 },
102 	{ "chunk_size",		KSTAT_DATA_UINT64 },
103 	{ "slab_size",		KSTAT_DATA_UINT64 },
104 	{ "alloc",		KSTAT_DATA_UINT64 },
105 	{ "alloc_fail",		KSTAT_DATA_UINT64 },
106 	{ "free",		KSTAT_DATA_UINT64 },
107 	{ "depot_alloc",	KSTAT_DATA_UINT64 },
108 	{ "depot_free",		KSTAT_DATA_UINT64 },
109 	{ "depot_contention",	KSTAT_DATA_UINT64 },
110 	{ "slab_alloc",		KSTAT_DATA_UINT64 },
111 	{ "slab_free",		KSTAT_DATA_UINT64 },
112 	{ "buf_constructed",	KSTAT_DATA_UINT64 },
113 	{ "buf_avail",		KSTAT_DATA_UINT64 },
114 	{ "buf_inuse",		KSTAT_DATA_UINT64 },
115 	{ "buf_total",		KSTAT_DATA_UINT64 },
116 	{ "buf_max",		KSTAT_DATA_UINT64 },
117 	{ "slab_create",	KSTAT_DATA_UINT64 },
118 	{ "slab_destroy",	KSTAT_DATA_UINT64 },
119 	{ "vmem_source",	KSTAT_DATA_UINT64 },
120 	{ "hash_size",		KSTAT_DATA_UINT64 },
121 	{ "hash_lookup_depth",	KSTAT_DATA_UINT64 },
122 	{ "hash_rescale",	KSTAT_DATA_UINT64 },
123 	{ "full_magazines",	KSTAT_DATA_UINT64 },
124 	{ "empty_magazines",	KSTAT_DATA_UINT64 },
125 	{ "magazine_size",	KSTAT_DATA_UINT64 },
126 };
127 
128 static kmutex_t kmem_cache_kstat_lock;
129 
130 /*
131  * The default set of caches to back kmem_alloc().
132  * These sizes should be reevaluated periodically.
133  *
134  * We want allocations that are multiples of the coherency granularity
135  * (64 bytes) to be satisfied from a cache which is a multiple of 64
136  * bytes, so that it will be 64-byte aligned.  For all multiples of 64,
137  * the next kmem_cache_size greater than or equal to it must be a
138  * multiple of 64.
139  */
140 static const int kmem_alloc_sizes[] = {
141 	1 * 8,
142 	2 * 8,
143 	3 * 8,
144 	4 * 8,		5 * 8,		6 * 8,		7 * 8,
145 	4 * 16,		5 * 16,		6 * 16,		7 * 16,
146 	4 * 32,		5 * 32,		6 * 32,		7 * 32,
147 	4 * 64,		5 * 64,		6 * 64,		7 * 64,
148 	4 * 128,	5 * 128,	6 * 128,	7 * 128,
149 	P2ALIGN(8192 / 7, 64),
150 	P2ALIGN(8192 / 6, 64),
151 	P2ALIGN(8192 / 5, 64),
152 	P2ALIGN(8192 / 4, 64),
153 	P2ALIGN(8192 / 3, 64),
154 	P2ALIGN(8192 / 2, 64),
155 	P2ALIGN(8192 / 1, 64),
156 	4096 * 3,
157 	8192 * 2,
158 	8192 * 3,
159 	8192 * 4,
160 };
161 
162 #define	KMEM_MAXBUF	32768
163 
164 static kmem_cache_t *kmem_alloc_table[KMEM_MAXBUF >> KMEM_ALIGN_SHIFT];
165 
166 static kmem_magtype_t kmem_magtype[] = {
167 	{ 1,	8,	3200,	65536	},
168 	{ 3,	16,	256,	32768	},
169 	{ 7,	32,	64,	16384	},
170 	{ 15,	64,	0,	8192	},
171 	{ 31,	64,	0,	4096	},
172 	{ 47,	64,	0,	2048	},
173 	{ 63,	64,	0,	1024	},
174 	{ 95,	64,	0,	512	},
175 	{ 143,	64,	0,	0	},
176 };
177 
178 static uint32_t kmem_reaping;
179 static uint32_t kmem_reaping_idspace;
180 
181 /*
182  * kmem tunables
183  */
184 clock_t kmem_reap_interval;	/* cache reaping rate [15 * HZ ticks] */
185 int kmem_depot_contention = 3;	/* max failed tryenters per real interval */
186 pgcnt_t kmem_reapahead = 0;	/* start reaping N pages before pageout */
187 int kmem_panic = 1;		/* whether to panic on error */
188 int kmem_logging = 1;		/* kmem_log_enter() override */
189 uint32_t kmem_mtbf = 0;		/* mean time between failures [default: off] */
190 size_t kmem_transaction_log_size; /* transaction log size [2% of memory] */
191 size_t kmem_content_log_size;	/* content log size [2% of memory] */
192 size_t kmem_failure_log_size;	/* failure log [4 pages per CPU] */
193 size_t kmem_slab_log_size;	/* slab create log [4 pages per CPU] */
194 size_t kmem_content_maxsave = 256; /* KMF_CONTENTS max bytes to log */
195 size_t kmem_lite_minsize = 0;	/* minimum buffer size for KMF_LITE */
196 size_t kmem_lite_maxalign = 1024; /* maximum buffer alignment for KMF_LITE */
197 int kmem_lite_pcs = 4;		/* number of PCs to store in KMF_LITE mode */
198 size_t kmem_maxverify;		/* maximum bytes to inspect in debug routines */
199 size_t kmem_minfirewall;	/* hardware-enforced redzone threshold */
200 
201 #ifdef DEBUG
202 int kmem_flags = KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE | KMF_CONTENTS;
203 #else
204 int kmem_flags = 0;
205 #endif
206 int kmem_ready;
207 
208 static kmem_cache_t	*kmem_slab_cache;
209 static kmem_cache_t	*kmem_bufctl_cache;
210 static kmem_cache_t	*kmem_bufctl_audit_cache;
211 
212 static kmutex_t		kmem_cache_lock;	/* inter-cache linkage only */
213 kmem_cache_t		kmem_null_cache;
214 
215 static taskq_t		*kmem_taskq;
216 static kmutex_t		kmem_flags_lock;
217 static vmem_t		*kmem_metadata_arena;
218 static vmem_t		*kmem_msb_arena;	/* arena for metadata caches */
219 static vmem_t		*kmem_cache_arena;
220 static vmem_t		*kmem_hash_arena;
221 static vmem_t		*kmem_log_arena;
222 static vmem_t		*kmem_oversize_arena;
223 static vmem_t		*kmem_va_arena;
224 static vmem_t		*kmem_default_arena;
225 static vmem_t		*kmem_firewall_va_arena;
226 static vmem_t		*kmem_firewall_arena;
227 
228 kmem_log_header_t	*kmem_transaction_log;
229 kmem_log_header_t	*kmem_content_log;
230 kmem_log_header_t	*kmem_failure_log;
231 kmem_log_header_t	*kmem_slab_log;
232 
233 static int		kmem_lite_count; /* # of PCs in kmem_buftag_lite_t */
234 
235 #define	KMEM_BUFTAG_LITE_ENTER(bt, count, caller)			\
236 	if ((count) > 0) {						\
237 		pc_t *_s = ((kmem_buftag_lite_t *)(bt))->bt_history;	\
238 		pc_t *_e;						\
239 		/* memmove() the old entries down one notch */		\
240 		for (_e = &_s[(count) - 1]; _e > _s; _e--)		\
241 			*_e = *(_e - 1);				\
242 		*_s = (uintptr_t)(caller);				\
243 	}
244 
245 #define	KMERR_MODIFIED	0	/* buffer modified while on freelist */
246 #define	KMERR_REDZONE	1	/* redzone violation (write past end of buf) */
247 #define	KMERR_DUPFREE	2	/* freed a buffer twice */
248 #define	KMERR_BADADDR	3	/* freed a bad (unallocated) address */
249 #define	KMERR_BADBUFTAG	4	/* buftag corrupted */
250 #define	KMERR_BADBUFCTL	5	/* bufctl corrupted */
251 #define	KMERR_BADCACHE	6	/* freed a buffer to the wrong cache */
252 #define	KMERR_BADSIZE	7	/* alloc size != free size */
253 #define	KMERR_BADBASE	8	/* buffer base address wrong */
254 
255 struct {
256 	hrtime_t	kmp_timestamp;	/* timestamp of panic */
257 	int		kmp_error;	/* type of kmem error */
258 	void		*kmp_buffer;	/* buffer that induced panic */
259 	void		*kmp_realbuf;	/* real start address for buffer */
260 	kmem_cache_t	*kmp_cache;	/* buffer's cache according to client */
261 	kmem_cache_t	*kmp_realcache;	/* actual cache containing buffer */
262 	kmem_slab_t	*kmp_slab;	/* slab accoring to kmem_findslab() */
263 	kmem_bufctl_t	*kmp_bufctl;	/* bufctl */
264 } kmem_panic_info;
265 
266 
267 static void
268 copy_pattern(uint64_t pattern, void *buf_arg, size_t size)
269 {
270 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
271 	uint64_t *buf = buf_arg;
272 
273 	while (buf < bufend)
274 		*buf++ = pattern;
275 }
276 
277 static void *
278 verify_pattern(uint64_t pattern, void *buf_arg, size_t size)
279 {
280 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
281 	uint64_t *buf;
282 
283 	for (buf = buf_arg; buf < bufend; buf++)
284 		if (*buf != pattern)
285 			return (buf);
286 	return (NULL);
287 }
288 
289 static void *
290 verify_and_copy_pattern(uint64_t old, uint64_t new, void *buf_arg, size_t size)
291 {
292 	uint64_t *bufend = (uint64_t *)((char *)buf_arg + size);
293 	uint64_t *buf;
294 
295 	for (buf = buf_arg; buf < bufend; buf++) {
296 		if (*buf != old) {
297 			copy_pattern(old, buf_arg,
298 			    (char *)buf - (char *)buf_arg);
299 			return (buf);
300 		}
301 		*buf = new;
302 	}
303 
304 	return (NULL);
305 }
306 
307 static void
308 kmem_cache_applyall(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
309 {
310 	kmem_cache_t *cp;
311 
312 	mutex_enter(&kmem_cache_lock);
313 	for (cp = kmem_null_cache.cache_next; cp != &kmem_null_cache;
314 	    cp = cp->cache_next)
315 		if (tq != NULL)
316 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
317 			    tqflag);
318 		else
319 			func(cp);
320 	mutex_exit(&kmem_cache_lock);
321 }
322 
323 static void
324 kmem_cache_applyall_id(void (*func)(kmem_cache_t *), taskq_t *tq, int tqflag)
325 {
326 	kmem_cache_t *cp;
327 
328 	mutex_enter(&kmem_cache_lock);
329 	for (cp = kmem_null_cache.cache_next; cp != &kmem_null_cache;
330 	    cp = cp->cache_next) {
331 		if (!(cp->cache_cflags & KMC_IDENTIFIER))
332 			continue;
333 		if (tq != NULL)
334 			(void) taskq_dispatch(tq, (task_func_t *)func, cp,
335 			    tqflag);
336 		else
337 			func(cp);
338 	}
339 	mutex_exit(&kmem_cache_lock);
340 }
341 
342 /*
343  * Debugging support.  Given a buffer address, find its slab.
344  */
345 static kmem_slab_t *
346 kmem_findslab(kmem_cache_t *cp, void *buf)
347 {
348 	kmem_slab_t *sp;
349 
350 	mutex_enter(&cp->cache_lock);
351 	for (sp = cp->cache_nullslab.slab_next;
352 	    sp != &cp->cache_nullslab; sp = sp->slab_next) {
353 		if (KMEM_SLAB_MEMBER(sp, buf)) {
354 			mutex_exit(&cp->cache_lock);
355 			return (sp);
356 		}
357 	}
358 	mutex_exit(&cp->cache_lock);
359 
360 	return (NULL);
361 }
362 
363 static void
364 kmem_error(int error, kmem_cache_t *cparg, void *bufarg)
365 {
366 	kmem_buftag_t *btp = NULL;
367 	kmem_bufctl_t *bcp = NULL;
368 	kmem_cache_t *cp = cparg;
369 	kmem_slab_t *sp;
370 	uint64_t *off;
371 	void *buf = bufarg;
372 
373 	kmem_logging = 0;	/* stop logging when a bad thing happens */
374 
375 	kmem_panic_info.kmp_timestamp = gethrtime();
376 
377 	sp = kmem_findslab(cp, buf);
378 	if (sp == NULL) {
379 		for (cp = kmem_null_cache.cache_prev; cp != &kmem_null_cache;
380 		    cp = cp->cache_prev) {
381 			if ((sp = kmem_findslab(cp, buf)) != NULL)
382 				break;
383 		}
384 	}
385 
386 	if (sp == NULL) {
387 		cp = NULL;
388 		error = KMERR_BADADDR;
389 	} else {
390 		if (cp != cparg)
391 			error = KMERR_BADCACHE;
392 		else
393 			buf = (char *)bufarg - ((uintptr_t)bufarg -
394 			    (uintptr_t)sp->slab_base) % cp->cache_chunksize;
395 		if (buf != bufarg)
396 			error = KMERR_BADBASE;
397 		if (cp->cache_flags & KMF_BUFTAG)
398 			btp = KMEM_BUFTAG(cp, buf);
399 		if (cp->cache_flags & KMF_HASH) {
400 			mutex_enter(&cp->cache_lock);
401 			for (bcp = *KMEM_HASH(cp, buf); bcp; bcp = bcp->bc_next)
402 				if (bcp->bc_addr == buf)
403 					break;
404 			mutex_exit(&cp->cache_lock);
405 			if (bcp == NULL && btp != NULL)
406 				bcp = btp->bt_bufctl;
407 			if (kmem_findslab(cp->cache_bufctl_cache, bcp) ==
408 			    NULL || P2PHASE((uintptr_t)bcp, KMEM_ALIGN) ||
409 			    bcp->bc_addr != buf) {
410 				error = KMERR_BADBUFCTL;
411 				bcp = NULL;
412 			}
413 		}
414 	}
415 
416 	kmem_panic_info.kmp_error = error;
417 	kmem_panic_info.kmp_buffer = bufarg;
418 	kmem_panic_info.kmp_realbuf = buf;
419 	kmem_panic_info.kmp_cache = cparg;
420 	kmem_panic_info.kmp_realcache = cp;
421 	kmem_panic_info.kmp_slab = sp;
422 	kmem_panic_info.kmp_bufctl = bcp;
423 
424 	printf("kernel memory allocator: ");
425 
426 	switch (error) {
427 
428 	case KMERR_MODIFIED:
429 		printf("buffer modified after being freed\n");
430 		off = verify_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
431 		if (off == NULL)	/* shouldn't happen */
432 			off = buf;
433 		printf("modification occurred at offset 0x%lx "
434 		    "(0x%llx replaced by 0x%llx)\n",
435 		    (uintptr_t)off - (uintptr_t)buf,
436 		    (longlong_t)KMEM_FREE_PATTERN, (longlong_t)*off);
437 		break;
438 
439 	case KMERR_REDZONE:
440 		printf("redzone violation: write past end of buffer\n");
441 		break;
442 
443 	case KMERR_BADADDR:
444 		printf("invalid free: buffer not in cache\n");
445 		break;
446 
447 	case KMERR_DUPFREE:
448 		printf("duplicate free: buffer freed twice\n");
449 		break;
450 
451 	case KMERR_BADBUFTAG:
452 		printf("boundary tag corrupted\n");
453 		printf("bcp ^ bxstat = %lx, should be %lx\n",
454 		    (intptr_t)btp->bt_bufctl ^ btp->bt_bxstat,
455 		    KMEM_BUFTAG_FREE);
456 		break;
457 
458 	case KMERR_BADBUFCTL:
459 		printf("bufctl corrupted\n");
460 		break;
461 
462 	case KMERR_BADCACHE:
463 		printf("buffer freed to wrong cache\n");
464 		printf("buffer was allocated from %s,\n", cp->cache_name);
465 		printf("caller attempting free to %s.\n", cparg->cache_name);
466 		break;
467 
468 	case KMERR_BADSIZE:
469 		printf("bad free: free size (%u) != alloc size (%u)\n",
470 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[0]),
471 		    KMEM_SIZE_DECODE(((uint32_t *)btp)[1]));
472 		break;
473 
474 	case KMERR_BADBASE:
475 		printf("bad free: free address (%p) != alloc address (%p)\n",
476 		    bufarg, buf);
477 		break;
478 	}
479 
480 	printf("buffer=%p  bufctl=%p  cache: %s\n",
481 	    bufarg, (void *)bcp, cparg->cache_name);
482 
483 	if (bcp != NULL && (cp->cache_flags & KMF_AUDIT) &&
484 	    error != KMERR_BADBUFCTL) {
485 		int d;
486 		timestruc_t ts;
487 		kmem_bufctl_audit_t *bcap = (kmem_bufctl_audit_t *)bcp;
488 
489 		hrt2ts(kmem_panic_info.kmp_timestamp - bcap->bc_timestamp, &ts);
490 		printf("previous transaction on buffer %p:\n", buf);
491 		printf("thread=%p  time=T-%ld.%09ld  slab=%p  cache: %s\n",
492 		    (void *)bcap->bc_thread, ts.tv_sec, ts.tv_nsec,
493 		    (void *)sp, cp->cache_name);
494 		for (d = 0; d < MIN(bcap->bc_depth, KMEM_STACK_DEPTH); d++) {
495 			ulong_t off;
496 			char *sym = kobj_getsymname(bcap->bc_stack[d], &off);
497 			printf("%s+%lx\n", sym ? sym : "?", off);
498 		}
499 	}
500 	if (kmem_panic > 0)
501 		panic("kernel heap corruption detected");
502 	if (kmem_panic == 0)
503 		debug_enter(NULL);
504 	kmem_logging = 1;	/* resume logging */
505 }
506 
507 static kmem_log_header_t *
508 kmem_log_init(size_t logsize)
509 {
510 	kmem_log_header_t *lhp;
511 	int nchunks = 4 * max_ncpus;
512 	size_t lhsize = (size_t)&((kmem_log_header_t *)0)->lh_cpu[max_ncpus];
513 	int i;
514 
515 	/*
516 	 * Make sure that lhp->lh_cpu[] is nicely aligned
517 	 * to prevent false sharing of cache lines.
518 	 */
519 	lhsize = P2ROUNDUP(lhsize, KMEM_ALIGN);
520 	lhp = vmem_xalloc(kmem_log_arena, lhsize, 64, P2NPHASE(lhsize, 64), 0,
521 	    NULL, NULL, VM_SLEEP);
522 	bzero(lhp, lhsize);
523 
524 	mutex_init(&lhp->lh_lock, NULL, MUTEX_DEFAULT, NULL);
525 	lhp->lh_nchunks = nchunks;
526 	lhp->lh_chunksize = P2ROUNDUP(logsize / nchunks + 1, PAGESIZE);
527 	lhp->lh_base = vmem_alloc(kmem_log_arena,
528 	    lhp->lh_chunksize * nchunks, VM_SLEEP);
529 	lhp->lh_free = vmem_alloc(kmem_log_arena,
530 	    nchunks * sizeof (int), VM_SLEEP);
531 	bzero(lhp->lh_base, lhp->lh_chunksize * nchunks);
532 
533 	for (i = 0; i < max_ncpus; i++) {
534 		kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[i];
535 		mutex_init(&clhp->clh_lock, NULL, MUTEX_DEFAULT, NULL);
536 		clhp->clh_chunk = i;
537 	}
538 
539 	for (i = max_ncpus; i < nchunks; i++)
540 		lhp->lh_free[i] = i;
541 
542 	lhp->lh_head = max_ncpus;
543 	lhp->lh_tail = 0;
544 
545 	return (lhp);
546 }
547 
548 static void *
549 kmem_log_enter(kmem_log_header_t *lhp, void *data, size_t size)
550 {
551 	void *logspace;
552 	kmem_cpu_log_header_t *clhp = &lhp->lh_cpu[CPU->cpu_seqid];
553 
554 	if (lhp == NULL || kmem_logging == 0 || panicstr)
555 		return (NULL);
556 
557 	mutex_enter(&clhp->clh_lock);
558 	clhp->clh_hits++;
559 	if (size > clhp->clh_avail) {
560 		mutex_enter(&lhp->lh_lock);
561 		lhp->lh_hits++;
562 		lhp->lh_free[lhp->lh_tail] = clhp->clh_chunk;
563 		lhp->lh_tail = (lhp->lh_tail + 1) % lhp->lh_nchunks;
564 		clhp->clh_chunk = lhp->lh_free[lhp->lh_head];
565 		lhp->lh_head = (lhp->lh_head + 1) % lhp->lh_nchunks;
566 		clhp->clh_current = lhp->lh_base +
567 		    clhp->clh_chunk * lhp->lh_chunksize;
568 		clhp->clh_avail = lhp->lh_chunksize;
569 		if (size > lhp->lh_chunksize)
570 			size = lhp->lh_chunksize;
571 		mutex_exit(&lhp->lh_lock);
572 	}
573 	logspace = clhp->clh_current;
574 	clhp->clh_current += size;
575 	clhp->clh_avail -= size;
576 	bcopy(data, logspace, size);
577 	mutex_exit(&clhp->clh_lock);
578 	return (logspace);
579 }
580 
581 #define	KMEM_AUDIT(lp, cp, bcp)						\
582 {									\
583 	kmem_bufctl_audit_t *_bcp = (kmem_bufctl_audit_t *)(bcp);	\
584 	_bcp->bc_timestamp = gethrtime();				\
585 	_bcp->bc_thread = curthread;					\
586 	_bcp->bc_depth = getpcstack(_bcp->bc_stack, KMEM_STACK_DEPTH);	\
587 	_bcp->bc_lastlog = kmem_log_enter((lp), _bcp, sizeof (*_bcp));	\
588 }
589 
590 static void
591 kmem_log_event(kmem_log_header_t *lp, kmem_cache_t *cp,
592 	kmem_slab_t *sp, void *addr)
593 {
594 	kmem_bufctl_audit_t bca;
595 
596 	bzero(&bca, sizeof (kmem_bufctl_audit_t));
597 	bca.bc_addr = addr;
598 	bca.bc_slab = sp;
599 	bca.bc_cache = cp;
600 	KMEM_AUDIT(lp, cp, &bca);
601 }
602 
603 /*
604  * Create a new slab for cache cp.
605  */
606 static kmem_slab_t *
607 kmem_slab_create(kmem_cache_t *cp, int kmflag)
608 {
609 	size_t slabsize = cp->cache_slabsize;
610 	size_t chunksize = cp->cache_chunksize;
611 	int cache_flags = cp->cache_flags;
612 	size_t color, chunks;
613 	char *buf, *slab;
614 	kmem_slab_t *sp;
615 	kmem_bufctl_t *bcp;
616 	vmem_t *vmp = cp->cache_arena;
617 
618 	color = cp->cache_color + cp->cache_align;
619 	if (color > cp->cache_maxcolor)
620 		color = cp->cache_mincolor;
621 	cp->cache_color = color;
622 
623 	slab = vmem_alloc(vmp, slabsize, kmflag & KM_VMFLAGS);
624 
625 	if (slab == NULL)
626 		goto vmem_alloc_failure;
627 
628 	ASSERT(P2PHASE((uintptr_t)slab, vmp->vm_quantum) == 0);
629 
630 	if (!(cp->cache_cflags & KMC_NOTOUCH))
631 		copy_pattern(KMEM_UNINITIALIZED_PATTERN, slab, slabsize);
632 
633 	if (cache_flags & KMF_HASH) {
634 		if ((sp = kmem_cache_alloc(kmem_slab_cache, kmflag)) == NULL)
635 			goto slab_alloc_failure;
636 		chunks = (slabsize - color) / chunksize;
637 	} else {
638 		sp = KMEM_SLAB(cp, slab);
639 		chunks = (slabsize - sizeof (kmem_slab_t) - color) / chunksize;
640 	}
641 
642 	sp->slab_cache	= cp;
643 	sp->slab_head	= NULL;
644 	sp->slab_refcnt	= 0;
645 	sp->slab_base	= buf = slab + color;
646 	sp->slab_chunks	= chunks;
647 
648 	ASSERT(chunks > 0);
649 	while (chunks-- != 0) {
650 		if (cache_flags & KMF_HASH) {
651 			bcp = kmem_cache_alloc(cp->cache_bufctl_cache, kmflag);
652 			if (bcp == NULL)
653 				goto bufctl_alloc_failure;
654 			if (cache_flags & KMF_AUDIT) {
655 				kmem_bufctl_audit_t *bcap =
656 				    (kmem_bufctl_audit_t *)bcp;
657 				bzero(bcap, sizeof (kmem_bufctl_audit_t));
658 				bcap->bc_cache = cp;
659 			}
660 			bcp->bc_addr = buf;
661 			bcp->bc_slab = sp;
662 		} else {
663 			bcp = KMEM_BUFCTL(cp, buf);
664 		}
665 		if (cache_flags & KMF_BUFTAG) {
666 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
667 			btp->bt_redzone = KMEM_REDZONE_PATTERN;
668 			btp->bt_bufctl = bcp;
669 			btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
670 			if (cache_flags & KMF_DEADBEEF) {
671 				copy_pattern(KMEM_FREE_PATTERN, buf,
672 				    cp->cache_verify);
673 			}
674 		}
675 		bcp->bc_next = sp->slab_head;
676 		sp->slab_head = bcp;
677 		buf += chunksize;
678 	}
679 
680 	kmem_log_event(kmem_slab_log, cp, sp, slab);
681 
682 	return (sp);
683 
684 bufctl_alloc_failure:
685 
686 	while ((bcp = sp->slab_head) != NULL) {
687 		sp->slab_head = bcp->bc_next;
688 		kmem_cache_free(cp->cache_bufctl_cache, bcp);
689 	}
690 	kmem_cache_free(kmem_slab_cache, sp);
691 
692 slab_alloc_failure:
693 
694 	vmem_free(vmp, slab, slabsize);
695 
696 vmem_alloc_failure:
697 
698 	kmem_log_event(kmem_failure_log, cp, NULL, NULL);
699 	atomic_add_64(&cp->cache_alloc_fail, 1);
700 
701 	return (NULL);
702 }
703 
704 /*
705  * Destroy a slab.
706  */
707 static void
708 kmem_slab_destroy(kmem_cache_t *cp, kmem_slab_t *sp)
709 {
710 	vmem_t *vmp = cp->cache_arena;
711 	void *slab = (void *)P2ALIGN((uintptr_t)sp->slab_base, vmp->vm_quantum);
712 
713 	if (cp->cache_flags & KMF_HASH) {
714 		kmem_bufctl_t *bcp;
715 		while ((bcp = sp->slab_head) != NULL) {
716 			sp->slab_head = bcp->bc_next;
717 			kmem_cache_free(cp->cache_bufctl_cache, bcp);
718 		}
719 		kmem_cache_free(kmem_slab_cache, sp);
720 	}
721 	vmem_free(vmp, slab, cp->cache_slabsize);
722 }
723 
724 /*
725  * Allocate a raw (unconstructed) buffer from cp's slab layer.
726  */
727 static void *
728 kmem_slab_alloc(kmem_cache_t *cp, int kmflag)
729 {
730 	kmem_bufctl_t *bcp, **hash_bucket;
731 	kmem_slab_t *sp;
732 	void *buf;
733 
734 	mutex_enter(&cp->cache_lock);
735 	cp->cache_slab_alloc++;
736 	sp = cp->cache_freelist;
737 	ASSERT(sp->slab_cache == cp);
738 	if (sp->slab_head == NULL) {
739 		ASSERT(cp->cache_bufslab == 0);
740 
741 		/*
742 		 * The freelist is empty.  Create a new slab.
743 		 */
744 		mutex_exit(&cp->cache_lock);
745 		if ((sp = kmem_slab_create(cp, kmflag)) == NULL)
746 			return (NULL);
747 		mutex_enter(&cp->cache_lock);
748 		cp->cache_slab_create++;
749 		if ((cp->cache_buftotal += sp->slab_chunks) > cp->cache_bufmax)
750 			cp->cache_bufmax = cp->cache_buftotal;
751 		cp->cache_bufslab += sp->slab_chunks;
752 		sp->slab_next = cp->cache_freelist;
753 		sp->slab_prev = cp->cache_freelist->slab_prev;
754 		sp->slab_next->slab_prev = sp;
755 		sp->slab_prev->slab_next = sp;
756 		cp->cache_freelist = sp;
757 	}
758 
759 	cp->cache_bufslab--;
760 	sp->slab_refcnt++;
761 	ASSERT(sp->slab_refcnt <= sp->slab_chunks);
762 
763 	/*
764 	 * If we're taking the last buffer in the slab,
765 	 * remove the slab from the cache's freelist.
766 	 */
767 	bcp = sp->slab_head;
768 	if ((sp->slab_head = bcp->bc_next) == NULL) {
769 		cp->cache_freelist = sp->slab_next;
770 		ASSERT(sp->slab_refcnt == sp->slab_chunks);
771 	}
772 
773 	if (cp->cache_flags & KMF_HASH) {
774 		/*
775 		 * Add buffer to allocated-address hash table.
776 		 */
777 		buf = bcp->bc_addr;
778 		hash_bucket = KMEM_HASH(cp, buf);
779 		bcp->bc_next = *hash_bucket;
780 		*hash_bucket = bcp;
781 		if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
782 			KMEM_AUDIT(kmem_transaction_log, cp, bcp);
783 		}
784 	} else {
785 		buf = KMEM_BUF(cp, bcp);
786 	}
787 
788 	ASSERT(KMEM_SLAB_MEMBER(sp, buf));
789 
790 	mutex_exit(&cp->cache_lock);
791 
792 	return (buf);
793 }
794 
795 /*
796  * Free a raw (unconstructed) buffer to cp's slab layer.
797  */
798 static void
799 kmem_slab_free(kmem_cache_t *cp, void *buf)
800 {
801 	kmem_slab_t *sp;
802 	kmem_bufctl_t *bcp, **prev_bcpp;
803 
804 	ASSERT(buf != NULL);
805 
806 	mutex_enter(&cp->cache_lock);
807 	cp->cache_slab_free++;
808 
809 	if (cp->cache_flags & KMF_HASH) {
810 		/*
811 		 * Look up buffer in allocated-address hash table.
812 		 */
813 		prev_bcpp = KMEM_HASH(cp, buf);
814 		while ((bcp = *prev_bcpp) != NULL) {
815 			if (bcp->bc_addr == buf) {
816 				*prev_bcpp = bcp->bc_next;
817 				sp = bcp->bc_slab;
818 				break;
819 			}
820 			cp->cache_lookup_depth++;
821 			prev_bcpp = &bcp->bc_next;
822 		}
823 	} else {
824 		bcp = KMEM_BUFCTL(cp, buf);
825 		sp = KMEM_SLAB(cp, buf);
826 	}
827 
828 	if (bcp == NULL || sp->slab_cache != cp || !KMEM_SLAB_MEMBER(sp, buf)) {
829 		mutex_exit(&cp->cache_lock);
830 		kmem_error(KMERR_BADADDR, cp, buf);
831 		return;
832 	}
833 
834 	if ((cp->cache_flags & (KMF_AUDIT | KMF_BUFTAG)) == KMF_AUDIT) {
835 		if (cp->cache_flags & KMF_CONTENTS)
836 			((kmem_bufctl_audit_t *)bcp)->bc_contents =
837 			    kmem_log_enter(kmem_content_log, buf,
838 			    cp->cache_contents);
839 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
840 	}
841 
842 	/*
843 	 * If this slab isn't currently on the freelist, put it there.
844 	 */
845 	if (sp->slab_head == NULL) {
846 		ASSERT(sp->slab_refcnt == sp->slab_chunks);
847 		ASSERT(cp->cache_freelist != sp);
848 		sp->slab_next->slab_prev = sp->slab_prev;
849 		sp->slab_prev->slab_next = sp->slab_next;
850 		sp->slab_next = cp->cache_freelist;
851 		sp->slab_prev = cp->cache_freelist->slab_prev;
852 		sp->slab_next->slab_prev = sp;
853 		sp->slab_prev->slab_next = sp;
854 		cp->cache_freelist = sp;
855 	}
856 
857 	bcp->bc_next = sp->slab_head;
858 	sp->slab_head = bcp;
859 
860 	cp->cache_bufslab++;
861 	ASSERT(sp->slab_refcnt >= 1);
862 	if (--sp->slab_refcnt == 0) {
863 		/*
864 		 * There are no outstanding allocations from this slab,
865 		 * so we can reclaim the memory.
866 		 */
867 		sp->slab_next->slab_prev = sp->slab_prev;
868 		sp->slab_prev->slab_next = sp->slab_next;
869 		if (sp == cp->cache_freelist)
870 			cp->cache_freelist = sp->slab_next;
871 		cp->cache_slab_destroy++;
872 		cp->cache_buftotal -= sp->slab_chunks;
873 		cp->cache_bufslab -= sp->slab_chunks;
874 		mutex_exit(&cp->cache_lock);
875 		kmem_slab_destroy(cp, sp);
876 		return;
877 	}
878 	mutex_exit(&cp->cache_lock);
879 }
880 
881 static int
882 kmem_cache_alloc_debug(kmem_cache_t *cp, void *buf, int kmflag, int construct,
883     caddr_t caller)
884 {
885 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
886 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
887 	uint32_t mtbf;
888 
889 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
890 		kmem_error(KMERR_BADBUFTAG, cp, buf);
891 		return (-1);
892 	}
893 
894 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_ALLOC;
895 
896 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
897 		kmem_error(KMERR_BADBUFCTL, cp, buf);
898 		return (-1);
899 	}
900 
901 	if (cp->cache_flags & KMF_DEADBEEF) {
902 		if (!construct && (cp->cache_flags & KMF_LITE)) {
903 			if (*(uint64_t *)buf != KMEM_FREE_PATTERN) {
904 				kmem_error(KMERR_MODIFIED, cp, buf);
905 				return (-1);
906 			}
907 			if (cp->cache_constructor != NULL)
908 				*(uint64_t *)buf = btp->bt_redzone;
909 			else
910 				*(uint64_t *)buf = KMEM_UNINITIALIZED_PATTERN;
911 		} else {
912 			construct = 1;
913 			if (verify_and_copy_pattern(KMEM_FREE_PATTERN,
914 			    KMEM_UNINITIALIZED_PATTERN, buf,
915 			    cp->cache_verify)) {
916 				kmem_error(KMERR_MODIFIED, cp, buf);
917 				return (-1);
918 			}
919 		}
920 	}
921 	btp->bt_redzone = KMEM_REDZONE_PATTERN;
922 
923 	if ((mtbf = kmem_mtbf | cp->cache_mtbf) != 0 &&
924 	    gethrtime() % mtbf == 0 &&
925 	    (kmflag & (KM_NOSLEEP | KM_PANIC)) == KM_NOSLEEP) {
926 		kmem_log_event(kmem_failure_log, cp, NULL, NULL);
927 		if (!construct && cp->cache_destructor != NULL)
928 			cp->cache_destructor(buf, cp->cache_private);
929 	} else {
930 		mtbf = 0;
931 	}
932 
933 	if (mtbf || (construct && cp->cache_constructor != NULL &&
934 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0)) {
935 		atomic_add_64(&cp->cache_alloc_fail, 1);
936 		btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
937 		if (cp->cache_flags & KMF_DEADBEEF)
938 			copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
939 		kmem_slab_free(cp, buf);
940 		return (-1);
941 	}
942 
943 	if (cp->cache_flags & KMF_AUDIT) {
944 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
945 	}
946 
947 	if ((cp->cache_flags & KMF_LITE) &&
948 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
949 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
950 	}
951 
952 	return (0);
953 }
954 
955 static int
956 kmem_cache_free_debug(kmem_cache_t *cp, void *buf, caddr_t caller)
957 {
958 	kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
959 	kmem_bufctl_audit_t *bcp = (kmem_bufctl_audit_t *)btp->bt_bufctl;
960 	kmem_slab_t *sp;
961 
962 	if (btp->bt_bxstat != ((intptr_t)bcp ^ KMEM_BUFTAG_ALLOC)) {
963 		if (btp->bt_bxstat == ((intptr_t)bcp ^ KMEM_BUFTAG_FREE)) {
964 			kmem_error(KMERR_DUPFREE, cp, buf);
965 			return (-1);
966 		}
967 		sp = kmem_findslab(cp, buf);
968 		if (sp == NULL || sp->slab_cache != cp)
969 			kmem_error(KMERR_BADADDR, cp, buf);
970 		else
971 			kmem_error(KMERR_REDZONE, cp, buf);
972 		return (-1);
973 	}
974 
975 	btp->bt_bxstat = (intptr_t)bcp ^ KMEM_BUFTAG_FREE;
976 
977 	if ((cp->cache_flags & KMF_HASH) && bcp->bc_addr != buf) {
978 		kmem_error(KMERR_BADBUFCTL, cp, buf);
979 		return (-1);
980 	}
981 
982 	if (btp->bt_redzone != KMEM_REDZONE_PATTERN) {
983 		kmem_error(KMERR_REDZONE, cp, buf);
984 		return (-1);
985 	}
986 
987 	if (cp->cache_flags & KMF_AUDIT) {
988 		if (cp->cache_flags & KMF_CONTENTS)
989 			bcp->bc_contents = kmem_log_enter(kmem_content_log,
990 			    buf, cp->cache_contents);
991 		KMEM_AUDIT(kmem_transaction_log, cp, bcp);
992 	}
993 
994 	if ((cp->cache_flags & KMF_LITE) &&
995 	    !(cp->cache_cflags & KMC_KMEM_ALLOC)) {
996 		KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count, caller);
997 	}
998 
999 	if (cp->cache_flags & KMF_DEADBEEF) {
1000 		if (cp->cache_flags & KMF_LITE)
1001 			btp->bt_redzone = *(uint64_t *)buf;
1002 		else if (cp->cache_destructor != NULL)
1003 			cp->cache_destructor(buf, cp->cache_private);
1004 
1005 		copy_pattern(KMEM_FREE_PATTERN, buf, cp->cache_verify);
1006 	}
1007 
1008 	return (0);
1009 }
1010 
1011 /*
1012  * Free each object in magazine mp to cp's slab layer, and free mp itself.
1013  */
1014 static void
1015 kmem_magazine_destroy(kmem_cache_t *cp, kmem_magazine_t *mp, int nrounds)
1016 {
1017 	int round;
1018 
1019 	ASSERT(cp->cache_next == NULL || taskq_member(kmem_taskq, curthread));
1020 
1021 	for (round = 0; round < nrounds; round++) {
1022 		void *buf = mp->mag_round[round];
1023 
1024 		if (cp->cache_flags & KMF_DEADBEEF) {
1025 			if (verify_pattern(KMEM_FREE_PATTERN, buf,
1026 			    cp->cache_verify) != NULL) {
1027 				kmem_error(KMERR_MODIFIED, cp, buf);
1028 				continue;
1029 			}
1030 			if ((cp->cache_flags & KMF_LITE) &&
1031 			    cp->cache_destructor != NULL) {
1032 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1033 				*(uint64_t *)buf = btp->bt_redzone;
1034 				cp->cache_destructor(buf, cp->cache_private);
1035 				*(uint64_t *)buf = KMEM_FREE_PATTERN;
1036 			}
1037 		} else if (cp->cache_destructor != NULL) {
1038 			cp->cache_destructor(buf, cp->cache_private);
1039 		}
1040 
1041 		kmem_slab_free(cp, buf);
1042 	}
1043 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
1044 	kmem_cache_free(cp->cache_magtype->mt_cache, mp);
1045 }
1046 
1047 /*
1048  * Allocate a magazine from the depot.
1049  */
1050 static kmem_magazine_t *
1051 kmem_depot_alloc(kmem_cache_t *cp, kmem_maglist_t *mlp)
1052 {
1053 	kmem_magazine_t *mp;
1054 
1055 	/*
1056 	 * If we can't get the depot lock without contention,
1057 	 * update our contention count.  We use the depot
1058 	 * contention rate to determine whether we need to
1059 	 * increase the magazine size for better scalability.
1060 	 */
1061 	if (!mutex_tryenter(&cp->cache_depot_lock)) {
1062 		mutex_enter(&cp->cache_depot_lock);
1063 		cp->cache_depot_contention++;
1064 	}
1065 
1066 	if ((mp = mlp->ml_list) != NULL) {
1067 		ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
1068 		mlp->ml_list = mp->mag_next;
1069 		if (--mlp->ml_total < mlp->ml_min)
1070 			mlp->ml_min = mlp->ml_total;
1071 		mlp->ml_alloc++;
1072 	}
1073 
1074 	mutex_exit(&cp->cache_depot_lock);
1075 
1076 	return (mp);
1077 }
1078 
1079 /*
1080  * Free a magazine to the depot.
1081  */
1082 static void
1083 kmem_depot_free(kmem_cache_t *cp, kmem_maglist_t *mlp, kmem_magazine_t *mp)
1084 {
1085 	mutex_enter(&cp->cache_depot_lock);
1086 	ASSERT(KMEM_MAGAZINE_VALID(cp, mp));
1087 	mp->mag_next = mlp->ml_list;
1088 	mlp->ml_list = mp;
1089 	mlp->ml_total++;
1090 	mutex_exit(&cp->cache_depot_lock);
1091 }
1092 
1093 /*
1094  * Update the working set statistics for cp's depot.
1095  */
1096 static void
1097 kmem_depot_ws_update(kmem_cache_t *cp)
1098 {
1099 	mutex_enter(&cp->cache_depot_lock);
1100 	cp->cache_full.ml_reaplimit = cp->cache_full.ml_min;
1101 	cp->cache_full.ml_min = cp->cache_full.ml_total;
1102 	cp->cache_empty.ml_reaplimit = cp->cache_empty.ml_min;
1103 	cp->cache_empty.ml_min = cp->cache_empty.ml_total;
1104 	mutex_exit(&cp->cache_depot_lock);
1105 }
1106 
1107 /*
1108  * Reap all magazines that have fallen out of the depot's working set.
1109  */
1110 static void
1111 kmem_depot_ws_reap(kmem_cache_t *cp)
1112 {
1113 	long reap;
1114 	kmem_magazine_t *mp;
1115 
1116 	ASSERT(cp->cache_next == NULL || taskq_member(kmem_taskq, curthread));
1117 
1118 	reap = MIN(cp->cache_full.ml_reaplimit, cp->cache_full.ml_min);
1119 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_full)) != NULL)
1120 		kmem_magazine_destroy(cp, mp, cp->cache_magtype->mt_magsize);
1121 
1122 	reap = MIN(cp->cache_empty.ml_reaplimit, cp->cache_empty.ml_min);
1123 	while (reap-- && (mp = kmem_depot_alloc(cp, &cp->cache_empty)) != NULL)
1124 		kmem_magazine_destroy(cp, mp, 0);
1125 }
1126 
1127 static void
1128 kmem_cpu_reload(kmem_cpu_cache_t *ccp, kmem_magazine_t *mp, int rounds)
1129 {
1130 	ASSERT((ccp->cc_loaded == NULL && ccp->cc_rounds == -1) ||
1131 	    (ccp->cc_loaded && ccp->cc_rounds + rounds == ccp->cc_magsize));
1132 	ASSERT(ccp->cc_magsize > 0);
1133 
1134 	ccp->cc_ploaded = ccp->cc_loaded;
1135 	ccp->cc_prounds = ccp->cc_rounds;
1136 	ccp->cc_loaded = mp;
1137 	ccp->cc_rounds = rounds;
1138 }
1139 
1140 /*
1141  * Allocate a constructed object from cache cp.
1142  */
1143 void *
1144 kmem_cache_alloc(kmem_cache_t *cp, int kmflag)
1145 {
1146 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
1147 	kmem_magazine_t *fmp;
1148 	void *buf;
1149 
1150 	mutex_enter(&ccp->cc_lock);
1151 	for (;;) {
1152 		/*
1153 		 * If there's an object available in the current CPU's
1154 		 * loaded magazine, just take it and return.
1155 		 */
1156 		if (ccp->cc_rounds > 0) {
1157 			buf = ccp->cc_loaded->mag_round[--ccp->cc_rounds];
1158 			ccp->cc_alloc++;
1159 			mutex_exit(&ccp->cc_lock);
1160 			if ((ccp->cc_flags & KMF_BUFTAG) &&
1161 			    kmem_cache_alloc_debug(cp, buf, kmflag, 0,
1162 			    caller()) == -1) {
1163 				if (kmflag & KM_NOSLEEP)
1164 					return (NULL);
1165 				mutex_enter(&ccp->cc_lock);
1166 				continue;
1167 			}
1168 			return (buf);
1169 		}
1170 
1171 		/*
1172 		 * The loaded magazine is empty.  If the previously loaded
1173 		 * magazine was full, exchange them and try again.
1174 		 */
1175 		if (ccp->cc_prounds > 0) {
1176 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
1177 			continue;
1178 		}
1179 
1180 		/*
1181 		 * If the magazine layer is disabled, break out now.
1182 		 */
1183 		if (ccp->cc_magsize == 0)
1184 			break;
1185 
1186 		/*
1187 		 * Try to get a full magazine from the depot.
1188 		 */
1189 		fmp = kmem_depot_alloc(cp, &cp->cache_full);
1190 		if (fmp != NULL) {
1191 			if (ccp->cc_ploaded != NULL)
1192 				kmem_depot_free(cp, &cp->cache_empty,
1193 				    ccp->cc_ploaded);
1194 			kmem_cpu_reload(ccp, fmp, ccp->cc_magsize);
1195 			continue;
1196 		}
1197 
1198 		/*
1199 		 * There are no full magazines in the depot,
1200 		 * so fall through to the slab layer.
1201 		 */
1202 		break;
1203 	}
1204 	mutex_exit(&ccp->cc_lock);
1205 
1206 	/*
1207 	 * We couldn't allocate a constructed object from the magazine layer,
1208 	 * so get a raw buffer from the slab layer and apply its constructor.
1209 	 */
1210 	buf = kmem_slab_alloc(cp, kmflag);
1211 
1212 	if (buf == NULL)
1213 		return (NULL);
1214 
1215 	if (cp->cache_flags & KMF_BUFTAG) {
1216 		/*
1217 		 * Make kmem_cache_alloc_debug() apply the constructor for us.
1218 		 */
1219 		if (kmem_cache_alloc_debug(cp, buf, kmflag, 1,
1220 		    caller()) == -1) {
1221 			if (kmflag & KM_NOSLEEP)
1222 				return (NULL);
1223 			/*
1224 			 * kmem_cache_alloc_debug() detected corruption
1225 			 * but didn't panic (kmem_panic <= 0).  Try again.
1226 			 */
1227 			return (kmem_cache_alloc(cp, kmflag));
1228 		}
1229 		return (buf);
1230 	}
1231 
1232 	if (cp->cache_constructor != NULL &&
1233 	    cp->cache_constructor(buf, cp->cache_private, kmflag) != 0) {
1234 		atomic_add_64(&cp->cache_alloc_fail, 1);
1235 		kmem_slab_free(cp, buf);
1236 		return (NULL);
1237 	}
1238 
1239 	return (buf);
1240 }
1241 
1242 /*
1243  * Free a constructed object to cache cp.
1244  */
1245 void
1246 kmem_cache_free(kmem_cache_t *cp, void *buf)
1247 {
1248 	kmem_cpu_cache_t *ccp = KMEM_CPU_CACHE(cp);
1249 	kmem_magazine_t *emp;
1250 	kmem_magtype_t *mtp;
1251 
1252 	if (ccp->cc_flags & KMF_BUFTAG)
1253 		if (kmem_cache_free_debug(cp, buf, caller()) == -1)
1254 			return;
1255 
1256 	mutex_enter(&ccp->cc_lock);
1257 	for (;;) {
1258 		/*
1259 		 * If there's a slot available in the current CPU's
1260 		 * loaded magazine, just put the object there and return.
1261 		 */
1262 		if ((uint_t)ccp->cc_rounds < ccp->cc_magsize) {
1263 			ccp->cc_loaded->mag_round[ccp->cc_rounds++] = buf;
1264 			ccp->cc_free++;
1265 			mutex_exit(&ccp->cc_lock);
1266 			return;
1267 		}
1268 
1269 		/*
1270 		 * The loaded magazine is full.  If the previously loaded
1271 		 * magazine was empty, exchange them and try again.
1272 		 */
1273 		if (ccp->cc_prounds == 0) {
1274 			kmem_cpu_reload(ccp, ccp->cc_ploaded, ccp->cc_prounds);
1275 			continue;
1276 		}
1277 
1278 		/*
1279 		 * If the magazine layer is disabled, break out now.
1280 		 */
1281 		if (ccp->cc_magsize == 0)
1282 			break;
1283 
1284 		/*
1285 		 * Try to get an empty magazine from the depot.
1286 		 */
1287 		emp = kmem_depot_alloc(cp, &cp->cache_empty);
1288 		if (emp != NULL) {
1289 			if (ccp->cc_ploaded != NULL)
1290 				kmem_depot_free(cp, &cp->cache_full,
1291 				    ccp->cc_ploaded);
1292 			kmem_cpu_reload(ccp, emp, 0);
1293 			continue;
1294 		}
1295 
1296 		/*
1297 		 * There are no empty magazines in the depot,
1298 		 * so try to allocate a new one.  We must drop all locks
1299 		 * across kmem_cache_alloc() because lower layers may
1300 		 * attempt to allocate from this cache.
1301 		 */
1302 		mtp = cp->cache_magtype;
1303 		mutex_exit(&ccp->cc_lock);
1304 		emp = kmem_cache_alloc(mtp->mt_cache, KM_NOSLEEP);
1305 		mutex_enter(&ccp->cc_lock);
1306 
1307 		if (emp != NULL) {
1308 			/*
1309 			 * We successfully allocated an empty magazine.
1310 			 * However, we had to drop ccp->cc_lock to do it,
1311 			 * so the cache's magazine size may have changed.
1312 			 * If so, free the magazine and try again.
1313 			 */
1314 			if (ccp->cc_magsize != mtp->mt_magsize) {
1315 				mutex_exit(&ccp->cc_lock);
1316 				kmem_cache_free(mtp->mt_cache, emp);
1317 				mutex_enter(&ccp->cc_lock);
1318 				continue;
1319 			}
1320 
1321 			/*
1322 			 * We got a magazine of the right size.  Add it to
1323 			 * the depot and try the whole dance again.
1324 			 */
1325 			kmem_depot_free(cp, &cp->cache_empty, emp);
1326 			continue;
1327 		}
1328 
1329 		/*
1330 		 * We couldn't allocate an empty magazine,
1331 		 * so fall through to the slab layer.
1332 		 */
1333 		break;
1334 	}
1335 	mutex_exit(&ccp->cc_lock);
1336 
1337 	/*
1338 	 * We couldn't free our constructed object to the magazine layer,
1339 	 * so apply its destructor and free it to the slab layer.
1340 	 * Note that if KMF_DEADBEEF is in effect and KMF_LITE is not,
1341 	 * kmem_cache_free_debug() will have already applied the destructor.
1342 	 */
1343 	if ((cp->cache_flags & (KMF_DEADBEEF | KMF_LITE)) != KMF_DEADBEEF &&
1344 	    cp->cache_destructor != NULL) {
1345 		if (cp->cache_flags & KMF_DEADBEEF) {	/* KMF_LITE implied */
1346 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1347 			*(uint64_t *)buf = btp->bt_redzone;
1348 			cp->cache_destructor(buf, cp->cache_private);
1349 			*(uint64_t *)buf = KMEM_FREE_PATTERN;
1350 		} else {
1351 			cp->cache_destructor(buf, cp->cache_private);
1352 		}
1353 	}
1354 
1355 	kmem_slab_free(cp, buf);
1356 }
1357 
1358 void *
1359 kmem_zalloc(size_t size, int kmflag)
1360 {
1361 	size_t index = (size - 1) >> KMEM_ALIGN_SHIFT;
1362 	void *buf;
1363 
1364 	if (index < KMEM_MAXBUF >> KMEM_ALIGN_SHIFT) {
1365 		kmem_cache_t *cp = kmem_alloc_table[index];
1366 		buf = kmem_cache_alloc(cp, kmflag);
1367 		if (buf != NULL) {
1368 			if (cp->cache_flags & KMF_BUFTAG) {
1369 				kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1370 				((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
1371 				((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
1372 
1373 				if (cp->cache_flags & KMF_LITE) {
1374 					KMEM_BUFTAG_LITE_ENTER(btp,
1375 					    kmem_lite_count, caller());
1376 				}
1377 			}
1378 			bzero(buf, size);
1379 		}
1380 	} else {
1381 		buf = kmem_alloc(size, kmflag);
1382 		if (buf != NULL)
1383 			bzero(buf, size);
1384 	}
1385 	return (buf);
1386 }
1387 
1388 void *
1389 kmem_alloc(size_t size, int kmflag)
1390 {
1391 	size_t index = (size - 1) >> KMEM_ALIGN_SHIFT;
1392 	void *buf;
1393 
1394 	if (index < KMEM_MAXBUF >> KMEM_ALIGN_SHIFT) {
1395 		kmem_cache_t *cp = kmem_alloc_table[index];
1396 		buf = kmem_cache_alloc(cp, kmflag);
1397 		if ((cp->cache_flags & KMF_BUFTAG) && buf != NULL) {
1398 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1399 			((uint8_t *)buf)[size] = KMEM_REDZONE_BYTE;
1400 			((uint32_t *)btp)[1] = KMEM_SIZE_ENCODE(size);
1401 
1402 			if (cp->cache_flags & KMF_LITE) {
1403 				KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
1404 				    caller());
1405 			}
1406 		}
1407 		return (buf);
1408 	}
1409 	if (size == 0)
1410 		return (NULL);
1411 	buf = vmem_alloc(kmem_oversize_arena, size, kmflag & KM_VMFLAGS);
1412 	if (buf == NULL)
1413 		kmem_log_event(kmem_failure_log, NULL, NULL, (void *)size);
1414 	return (buf);
1415 }
1416 
1417 void
1418 kmem_free(void *buf, size_t size)
1419 {
1420 	size_t index = (size - 1) >> KMEM_ALIGN_SHIFT;
1421 
1422 	if (index < KMEM_MAXBUF >> KMEM_ALIGN_SHIFT) {
1423 		kmem_cache_t *cp = kmem_alloc_table[index];
1424 		if (cp->cache_flags & KMF_BUFTAG) {
1425 			kmem_buftag_t *btp = KMEM_BUFTAG(cp, buf);
1426 			uint32_t *ip = (uint32_t *)btp;
1427 			if (ip[1] != KMEM_SIZE_ENCODE(size)) {
1428 				if (*(uint64_t *)buf == KMEM_FREE_PATTERN) {
1429 					kmem_error(KMERR_DUPFREE, cp, buf);
1430 					return;
1431 				}
1432 				if (KMEM_SIZE_VALID(ip[1])) {
1433 					ip[0] = KMEM_SIZE_ENCODE(size);
1434 					kmem_error(KMERR_BADSIZE, cp, buf);
1435 				} else {
1436 					kmem_error(KMERR_REDZONE, cp, buf);
1437 				}
1438 				return;
1439 			}
1440 			if (((uint8_t *)buf)[size] != KMEM_REDZONE_BYTE) {
1441 				kmem_error(KMERR_REDZONE, cp, buf);
1442 				return;
1443 			}
1444 			btp->bt_redzone = KMEM_REDZONE_PATTERN;
1445 			if (cp->cache_flags & KMF_LITE) {
1446 				KMEM_BUFTAG_LITE_ENTER(btp, kmem_lite_count,
1447 				    caller());
1448 			}
1449 		}
1450 		kmem_cache_free(cp, buf);
1451 	} else {
1452 		if (buf == NULL && size == 0)
1453 			return;
1454 		vmem_free(kmem_oversize_arena, buf, size);
1455 	}
1456 }
1457 
1458 void *
1459 kmem_firewall_va_alloc(vmem_t *vmp, size_t size, int vmflag)
1460 {
1461 	size_t realsize = size + vmp->vm_quantum;
1462 	void *addr;
1463 
1464 	/*
1465 	 * Annoying edge case: if 'size' is just shy of ULONG_MAX, adding
1466 	 * vm_quantum will cause integer wraparound.  Check for this, and
1467 	 * blow off the firewall page in this case.  Note that such a
1468 	 * giant allocation (the entire kernel address space) can never
1469 	 * be satisfied, so it will either fail immediately (VM_NOSLEEP)
1470 	 * or sleep forever (VM_SLEEP).  Thus, there is no need for a
1471 	 * corresponding check in kmem_firewall_va_free().
1472 	 */
1473 	if (realsize < size)
1474 		realsize = size;
1475 
1476 	/*
1477 	 * While boot still owns resource management, make sure that this
1478 	 * redzone virtual address allocation is properly accounted for in
1479 	 * OBPs "virtual-memory" "available" lists because we're
1480 	 * effectively claiming them for a red zone.  If we don't do this,
1481 	 * the available lists become too fragmented and too large for the
1482 	 * current boot/kernel memory list interface.
1483 	 */
1484 	addr = vmem_alloc(vmp, realsize, vmflag | VM_NEXTFIT);
1485 
1486 	if (addr != NULL && kvseg.s_base == NULL && realsize != size)
1487 		(void) boot_virt_alloc((char *)addr + size, vmp->vm_quantum);
1488 
1489 	return (addr);
1490 }
1491 
1492 void
1493 kmem_firewall_va_free(vmem_t *vmp, void *addr, size_t size)
1494 {
1495 	ASSERT((kvseg.s_base == NULL ?
1496 	    va_to_pfn((char *)addr + size) :
1497 	    hat_getpfnum(kas.a_hat, (caddr_t)addr + size)) == PFN_INVALID);
1498 
1499 	vmem_free(vmp, addr, size + vmp->vm_quantum);
1500 }
1501 
1502 /*
1503  * Try to allocate at least `size' bytes of memory without sleeping or
1504  * panicking. Return actual allocated size in `asize'. If allocation failed,
1505  * try final allocation with sleep or panic allowed.
1506  */
1507 void *
1508 kmem_alloc_tryhard(size_t size, size_t *asize, int kmflag)
1509 {
1510 	void *p;
1511 
1512 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
1513 	do {
1514 		p = kmem_alloc(*asize, (kmflag | KM_NOSLEEP) & ~KM_PANIC);
1515 		if (p != NULL)
1516 			return (p);
1517 		*asize += KMEM_ALIGN;
1518 	} while (*asize <= PAGESIZE);
1519 
1520 	*asize = P2ROUNDUP(size, KMEM_ALIGN);
1521 	return (kmem_alloc(*asize, kmflag));
1522 }
1523 
1524 /*
1525  * Reclaim all unused memory from a cache.
1526  */
1527 static void
1528 kmem_cache_reap(kmem_cache_t *cp)
1529 {
1530 	/*
1531 	 * Ask the cache's owner to free some memory if possible.
1532 	 * The idea is to handle things like the inode cache, which
1533 	 * typically sits on a bunch of memory that it doesn't truly
1534 	 * *need*.  Reclaim policy is entirely up to the owner; this
1535 	 * callback is just an advisory plea for help.
1536 	 */
1537 	if (cp->cache_reclaim != NULL)
1538 		cp->cache_reclaim(cp->cache_private);
1539 
1540 	kmem_depot_ws_reap(cp);
1541 }
1542 
1543 static void
1544 kmem_reap_timeout(void *flag_arg)
1545 {
1546 	uint32_t *flag = (uint32_t *)flag_arg;
1547 
1548 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
1549 	*flag = 0;
1550 }
1551 
1552 static void
1553 kmem_reap_done(void *flag)
1554 {
1555 	(void) timeout(kmem_reap_timeout, flag, kmem_reap_interval);
1556 }
1557 
1558 static void
1559 kmem_reap_start(void *flag)
1560 {
1561 	ASSERT(flag == &kmem_reaping || flag == &kmem_reaping_idspace);
1562 
1563 	if (flag == &kmem_reaping) {
1564 		kmem_cache_applyall(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
1565 		/*
1566 		 * if we have segkp under heap, reap segkp cache.
1567 		 */
1568 		if (segkp_fromheap)
1569 			segkp_cache_free();
1570 	}
1571 	else
1572 		kmem_cache_applyall_id(kmem_cache_reap, kmem_taskq, TQ_NOSLEEP);
1573 
1574 	/*
1575 	 * We use taskq_dispatch() to schedule a timeout to clear
1576 	 * the flag so that kmem_reap() becomes self-throttling:
1577 	 * we won't reap again until the current reap completes *and*
1578 	 * at least kmem_reap_interval ticks have elapsed.
1579 	 */
1580 	if (!taskq_dispatch(kmem_taskq, kmem_reap_done, flag, TQ_NOSLEEP))
1581 		kmem_reap_done(flag);
1582 }
1583 
1584 static void
1585 kmem_reap_common(void *flag_arg)
1586 {
1587 	uint32_t *flag = (uint32_t *)flag_arg;
1588 
1589 	if (MUTEX_HELD(&kmem_cache_lock) || kmem_taskq == NULL ||
1590 	    cas32(flag, 0, 1) != 0)
1591 		return;
1592 
1593 	/*
1594 	 * It may not be kosher to do memory allocation when a reap is called
1595 	 * is called (for example, if vmem_populate() is in the call chain).
1596 	 * So we start the reap going with a TQ_NOALLOC dispatch.  If the
1597 	 * dispatch fails, we reset the flag, and the next reap will try again.
1598 	 */
1599 	if (!taskq_dispatch(kmem_taskq, kmem_reap_start, flag, TQ_NOALLOC))
1600 		*flag = 0;
1601 }
1602 
1603 /*
1604  * Reclaim all unused memory from all caches.  Called from the VM system
1605  * when memory gets tight.
1606  */
1607 void
1608 kmem_reap(void)
1609 {
1610 	kmem_reap_common(&kmem_reaping);
1611 }
1612 
1613 /*
1614  * Reclaim all unused memory from identifier arenas, called when a vmem
1615  * arena not back by memory is exhausted.  Since reaping memory-backed caches
1616  * cannot help with identifier exhaustion, we avoid both a large amount of
1617  * work and unwanted side-effects from reclaim callbacks.
1618  */
1619 void
1620 kmem_reap_idspace(void)
1621 {
1622 	kmem_reap_common(&kmem_reaping_idspace);
1623 }
1624 
1625 /*
1626  * Purge all magazines from a cache and set its magazine limit to zero.
1627  * All calls are serialized by the kmem_taskq lock, except for the final
1628  * call from kmem_cache_destroy().
1629  */
1630 static void
1631 kmem_cache_magazine_purge(kmem_cache_t *cp)
1632 {
1633 	kmem_cpu_cache_t *ccp;
1634 	kmem_magazine_t *mp, *pmp;
1635 	int rounds, prounds, cpu_seqid;
1636 
1637 	ASSERT(cp->cache_next == NULL || taskq_member(kmem_taskq, curthread));
1638 	ASSERT(MUTEX_NOT_HELD(&cp->cache_lock));
1639 
1640 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
1641 		ccp = &cp->cache_cpu[cpu_seqid];
1642 
1643 		mutex_enter(&ccp->cc_lock);
1644 		mp = ccp->cc_loaded;
1645 		pmp = ccp->cc_ploaded;
1646 		rounds = ccp->cc_rounds;
1647 		prounds = ccp->cc_prounds;
1648 		ccp->cc_loaded = NULL;
1649 		ccp->cc_ploaded = NULL;
1650 		ccp->cc_rounds = -1;
1651 		ccp->cc_prounds = -1;
1652 		ccp->cc_magsize = 0;
1653 		mutex_exit(&ccp->cc_lock);
1654 
1655 		if (mp)
1656 			kmem_magazine_destroy(cp, mp, rounds);
1657 		if (pmp)
1658 			kmem_magazine_destroy(cp, pmp, prounds);
1659 	}
1660 
1661 	/*
1662 	 * Updating the working set statistics twice in a row has the
1663 	 * effect of setting the working set size to zero, so everything
1664 	 * is eligible for reaping.
1665 	 */
1666 	kmem_depot_ws_update(cp);
1667 	kmem_depot_ws_update(cp);
1668 
1669 	kmem_depot_ws_reap(cp);
1670 }
1671 
1672 /*
1673  * Enable per-cpu magazines on a cache.
1674  */
1675 static void
1676 kmem_cache_magazine_enable(kmem_cache_t *cp)
1677 {
1678 	int cpu_seqid;
1679 
1680 	if (cp->cache_flags & KMF_NOMAGAZINE)
1681 		return;
1682 
1683 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
1684 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
1685 		mutex_enter(&ccp->cc_lock);
1686 		ccp->cc_magsize = cp->cache_magtype->mt_magsize;
1687 		mutex_exit(&ccp->cc_lock);
1688 	}
1689 
1690 }
1691 
1692 /*
1693  * Reap (almost) everything right now.  See kmem_cache_magazine_purge()
1694  * for explanation of the back-to-back kmem_depot_ws_update() calls.
1695  */
1696 void
1697 kmem_cache_reap_now(kmem_cache_t *cp)
1698 {
1699 	kmem_depot_ws_update(cp);
1700 	kmem_depot_ws_update(cp);
1701 
1702 	(void) taskq_dispatch(kmem_taskq,
1703 	    (task_func_t *)kmem_depot_ws_reap, cp, TQ_SLEEP);
1704 	taskq_wait(kmem_taskq);
1705 }
1706 
1707 /*
1708  * Recompute a cache's magazine size.  The trade-off is that larger magazines
1709  * provide a higher transfer rate with the depot, while smaller magazines
1710  * reduce memory consumption.  Magazine resizing is an expensive operation;
1711  * it should not be done frequently.
1712  *
1713  * Changes to the magazine size are serialized by the kmem_taskq lock.
1714  *
1715  * Note: at present this only grows the magazine size.  It might be useful
1716  * to allow shrinkage too.
1717  */
1718 static void
1719 kmem_cache_magazine_resize(kmem_cache_t *cp)
1720 {
1721 	kmem_magtype_t *mtp = cp->cache_magtype;
1722 
1723 	ASSERT(taskq_member(kmem_taskq, curthread));
1724 
1725 	if (cp->cache_chunksize < mtp->mt_maxbuf) {
1726 		kmem_cache_magazine_purge(cp);
1727 		mutex_enter(&cp->cache_depot_lock);
1728 		cp->cache_magtype = ++mtp;
1729 		cp->cache_depot_contention_prev =
1730 		    cp->cache_depot_contention + INT_MAX;
1731 		mutex_exit(&cp->cache_depot_lock);
1732 		kmem_cache_magazine_enable(cp);
1733 	}
1734 }
1735 
1736 /*
1737  * Rescale a cache's hash table, so that the table size is roughly the
1738  * cache size.  We want the average lookup time to be extremely small.
1739  */
1740 static void
1741 kmem_hash_rescale(kmem_cache_t *cp)
1742 {
1743 	kmem_bufctl_t **old_table, **new_table, *bcp;
1744 	size_t old_size, new_size, h;
1745 
1746 	ASSERT(taskq_member(kmem_taskq, curthread));
1747 
1748 	new_size = MAX(KMEM_HASH_INITIAL,
1749 	    1 << (highbit(3 * cp->cache_buftotal + 4) - 2));
1750 	old_size = cp->cache_hash_mask + 1;
1751 
1752 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
1753 		return;
1754 
1755 	new_table = vmem_alloc(kmem_hash_arena, new_size * sizeof (void *),
1756 	    VM_NOSLEEP);
1757 	if (new_table == NULL)
1758 		return;
1759 	bzero(new_table, new_size * sizeof (void *));
1760 
1761 	mutex_enter(&cp->cache_lock);
1762 
1763 	old_size = cp->cache_hash_mask + 1;
1764 	old_table = cp->cache_hash_table;
1765 
1766 	cp->cache_hash_mask = new_size - 1;
1767 	cp->cache_hash_table = new_table;
1768 	cp->cache_rescale++;
1769 
1770 	for (h = 0; h < old_size; h++) {
1771 		bcp = old_table[h];
1772 		while (bcp != NULL) {
1773 			void *addr = bcp->bc_addr;
1774 			kmem_bufctl_t *next_bcp = bcp->bc_next;
1775 			kmem_bufctl_t **hash_bucket = KMEM_HASH(cp, addr);
1776 			bcp->bc_next = *hash_bucket;
1777 			*hash_bucket = bcp;
1778 			bcp = next_bcp;
1779 		}
1780 	}
1781 
1782 	mutex_exit(&cp->cache_lock);
1783 
1784 	vmem_free(kmem_hash_arena, old_table, old_size * sizeof (void *));
1785 }
1786 
1787 /*
1788  * Perform periodic maintenance on a cache: hash rescaling,
1789  * depot working-set update, and magazine resizing.
1790  */
1791 static void
1792 kmem_cache_update(kmem_cache_t *cp)
1793 {
1794 	int need_hash_rescale = 0;
1795 	int need_magazine_resize = 0;
1796 
1797 	ASSERT(MUTEX_HELD(&kmem_cache_lock));
1798 
1799 	/*
1800 	 * If the cache has become much larger or smaller than its hash table,
1801 	 * fire off a request to rescale the hash table.
1802 	 */
1803 	mutex_enter(&cp->cache_lock);
1804 
1805 	if ((cp->cache_flags & KMF_HASH) &&
1806 	    (cp->cache_buftotal > (cp->cache_hash_mask << 1) ||
1807 	    (cp->cache_buftotal < (cp->cache_hash_mask >> 1) &&
1808 	    cp->cache_hash_mask > KMEM_HASH_INITIAL)))
1809 		need_hash_rescale = 1;
1810 
1811 	mutex_exit(&cp->cache_lock);
1812 
1813 	/*
1814 	 * Update the depot working set statistics.
1815 	 */
1816 	kmem_depot_ws_update(cp);
1817 
1818 	/*
1819 	 * If there's a lot of contention in the depot,
1820 	 * increase the magazine size.
1821 	 */
1822 	mutex_enter(&cp->cache_depot_lock);
1823 
1824 	if (cp->cache_chunksize < cp->cache_magtype->mt_maxbuf &&
1825 	    (int)(cp->cache_depot_contention -
1826 	    cp->cache_depot_contention_prev) > kmem_depot_contention)
1827 		need_magazine_resize = 1;
1828 
1829 	cp->cache_depot_contention_prev = cp->cache_depot_contention;
1830 
1831 	mutex_exit(&cp->cache_depot_lock);
1832 
1833 	if (need_hash_rescale)
1834 		(void) taskq_dispatch(kmem_taskq,
1835 		    (task_func_t *)kmem_hash_rescale, cp, TQ_NOSLEEP);
1836 
1837 	if (need_magazine_resize)
1838 		(void) taskq_dispatch(kmem_taskq,
1839 		    (task_func_t *)kmem_cache_magazine_resize, cp, TQ_NOSLEEP);
1840 }
1841 
1842 static void
1843 kmem_update_timeout(void *dummy)
1844 {
1845 	static void kmem_update(void *);
1846 
1847 	(void) timeout(kmem_update, dummy, kmem_reap_interval);
1848 }
1849 
1850 static void
1851 kmem_update(void *dummy)
1852 {
1853 	kmem_cache_applyall(kmem_cache_update, NULL, TQ_NOSLEEP);
1854 
1855 	/*
1856 	 * We use taskq_dispatch() to reschedule the timeout so that
1857 	 * kmem_update() becomes self-throttling: it won't schedule
1858 	 * new tasks until all previous tasks have completed.
1859 	 */
1860 	if (!taskq_dispatch(kmem_taskq, kmem_update_timeout, dummy, TQ_NOSLEEP))
1861 		kmem_update_timeout(NULL);
1862 }
1863 
1864 static int
1865 kmem_cache_kstat_update(kstat_t *ksp, int rw)
1866 {
1867 	struct kmem_cache_kstat *kmcp = &kmem_cache_kstat;
1868 	kmem_cache_t *cp = ksp->ks_private;
1869 	uint64_t cpu_buf_avail;
1870 	uint64_t buf_avail = 0;
1871 	int cpu_seqid;
1872 
1873 	ASSERT(MUTEX_HELD(&kmem_cache_kstat_lock));
1874 
1875 	if (rw == KSTAT_WRITE)
1876 		return (EACCES);
1877 
1878 	mutex_enter(&cp->cache_lock);
1879 
1880 	kmcp->kmc_alloc_fail.value.ui64		= cp->cache_alloc_fail;
1881 	kmcp->kmc_alloc.value.ui64		= cp->cache_slab_alloc;
1882 	kmcp->kmc_free.value.ui64		= cp->cache_slab_free;
1883 	kmcp->kmc_slab_alloc.value.ui64		= cp->cache_slab_alloc;
1884 	kmcp->kmc_slab_free.value.ui64		= cp->cache_slab_free;
1885 
1886 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
1887 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
1888 
1889 		mutex_enter(&ccp->cc_lock);
1890 
1891 		cpu_buf_avail = 0;
1892 		if (ccp->cc_rounds > 0)
1893 			cpu_buf_avail += ccp->cc_rounds;
1894 		if (ccp->cc_prounds > 0)
1895 			cpu_buf_avail += ccp->cc_prounds;
1896 
1897 		kmcp->kmc_alloc.value.ui64	+= ccp->cc_alloc;
1898 		kmcp->kmc_free.value.ui64	+= ccp->cc_free;
1899 		buf_avail			+= cpu_buf_avail;
1900 
1901 		mutex_exit(&ccp->cc_lock);
1902 	}
1903 
1904 	mutex_enter(&cp->cache_depot_lock);
1905 
1906 	kmcp->kmc_depot_alloc.value.ui64	= cp->cache_full.ml_alloc;
1907 	kmcp->kmc_depot_free.value.ui64		= cp->cache_empty.ml_alloc;
1908 	kmcp->kmc_depot_contention.value.ui64	= cp->cache_depot_contention;
1909 	kmcp->kmc_full_magazines.value.ui64	= cp->cache_full.ml_total;
1910 	kmcp->kmc_empty_magazines.value.ui64	= cp->cache_empty.ml_total;
1911 	kmcp->kmc_magazine_size.value.ui64	=
1912 	    (cp->cache_flags & KMF_NOMAGAZINE) ?
1913 	    0 : cp->cache_magtype->mt_magsize;
1914 
1915 	kmcp->kmc_alloc.value.ui64		+= cp->cache_full.ml_alloc;
1916 	kmcp->kmc_free.value.ui64		+= cp->cache_empty.ml_alloc;
1917 	buf_avail += cp->cache_full.ml_total * cp->cache_magtype->mt_magsize;
1918 
1919 	mutex_exit(&cp->cache_depot_lock);
1920 
1921 	kmcp->kmc_buf_size.value.ui64	= cp->cache_bufsize;
1922 	kmcp->kmc_align.value.ui64	= cp->cache_align;
1923 	kmcp->kmc_chunk_size.value.ui64	= cp->cache_chunksize;
1924 	kmcp->kmc_slab_size.value.ui64	= cp->cache_slabsize;
1925 	kmcp->kmc_buf_constructed.value.ui64 = buf_avail;
1926 	buf_avail += cp->cache_bufslab;
1927 	kmcp->kmc_buf_avail.value.ui64	= buf_avail;
1928 	kmcp->kmc_buf_inuse.value.ui64	= cp->cache_buftotal - buf_avail;
1929 	kmcp->kmc_buf_total.value.ui64	= cp->cache_buftotal;
1930 	kmcp->kmc_buf_max.value.ui64	= cp->cache_bufmax;
1931 	kmcp->kmc_slab_create.value.ui64	= cp->cache_slab_create;
1932 	kmcp->kmc_slab_destroy.value.ui64	= cp->cache_slab_destroy;
1933 	kmcp->kmc_hash_size.value.ui64	= (cp->cache_flags & KMF_HASH) ?
1934 	    cp->cache_hash_mask + 1 : 0;
1935 	kmcp->kmc_hash_lookup_depth.value.ui64	= cp->cache_lookup_depth;
1936 	kmcp->kmc_hash_rescale.value.ui64	= cp->cache_rescale;
1937 	kmcp->kmc_vmem_source.value.ui64	= cp->cache_arena->vm_id;
1938 
1939 	mutex_exit(&cp->cache_lock);
1940 	return (0);
1941 }
1942 
1943 /*
1944  * Return a named statistic about a particular cache.
1945  * This shouldn't be called very often, so it's currently designed for
1946  * simplicity (leverages existing kstat support) rather than efficiency.
1947  */
1948 uint64_t
1949 kmem_cache_stat(kmem_cache_t *cp, char *name)
1950 {
1951 	int i;
1952 	kstat_t *ksp = cp->cache_kstat;
1953 	kstat_named_t *knp = (kstat_named_t *)&kmem_cache_kstat;
1954 	uint64_t value = 0;
1955 
1956 	if (ksp != NULL) {
1957 		mutex_enter(&kmem_cache_kstat_lock);
1958 		(void) kmem_cache_kstat_update(ksp, KSTAT_READ);
1959 		for (i = 0; i < ksp->ks_ndata; i++) {
1960 			if (strcmp(knp[i].name, name) == 0) {
1961 				value = knp[i].value.ui64;
1962 				break;
1963 			}
1964 		}
1965 		mutex_exit(&kmem_cache_kstat_lock);
1966 	}
1967 	return (value);
1968 }
1969 
1970 /*
1971  * Return an estimate of currently available kernel heap memory.
1972  * On 32-bit systems, physical memory may exceed virtual memory,
1973  * we just truncate the result at 1GB.
1974  */
1975 size_t
1976 kmem_avail(void)
1977 {
1978 	spgcnt_t rmem = availrmem - tune.t_minarmem;
1979 	spgcnt_t fmem = freemem - minfree;
1980 
1981 	return ((size_t)ptob(MIN(MAX(MIN(rmem, fmem), 0),
1982 	    1 << (30 - PAGESHIFT))));
1983 }
1984 
1985 /*
1986  * Return the maximum amount of memory that is (in theory) allocatable
1987  * from the heap. This may be used as an estimate only since there
1988  * is no guarentee this space will still be available when an allocation
1989  * request is made, nor that the space may be allocated in one big request
1990  * due to kernel heap fragmentation.
1991  */
1992 size_t
1993 kmem_maxavail(void)
1994 {
1995 	spgcnt_t pmem = availrmem - tune.t_minarmem;
1996 	spgcnt_t vmem = btop(vmem_size(heap_arena, VMEM_FREE));
1997 
1998 	return ((size_t)ptob(MAX(MIN(pmem, vmem), 0)));
1999 }
2000 
2001 /*
2002  * Indicate whether memory-intensive kmem debugging is enabled.
2003  */
2004 int
2005 kmem_debugging(void)
2006 {
2007 	return (kmem_flags & (KMF_AUDIT | KMF_REDZONE));
2008 }
2009 
2010 kmem_cache_t *
2011 kmem_cache_create(
2012 	char *name,		/* descriptive name for this cache */
2013 	size_t bufsize,		/* size of the objects it manages */
2014 	size_t align,		/* required object alignment */
2015 	int (*constructor)(void *, void *, int), /* object constructor */
2016 	void (*destructor)(void *, void *),	/* object destructor */
2017 	void (*reclaim)(void *), /* memory reclaim callback */
2018 	void *private,		/* pass-thru arg for constr/destr/reclaim */
2019 	vmem_t *vmp,		/* vmem source for slab allocation */
2020 	int cflags)		/* cache creation flags */
2021 {
2022 	int cpu_seqid;
2023 	size_t chunksize;
2024 	kmem_cache_t *cp, *cnext, *cprev;
2025 	kmem_magtype_t *mtp;
2026 	size_t csize = KMEM_CACHE_SIZE(max_ncpus);
2027 
2028 #ifdef	DEBUG
2029 	/*
2030 	 * Cache names should conform to the rules for valid C identifiers
2031 	 */
2032 	if (!strident_valid(name)) {
2033 		cmn_err(CE_CONT,
2034 		    "kmem_cache_create: '%s' is an invalid cache name\n"
2035 		    "cache names must conform to the rules for "
2036 		    "C identifiers\n", name);
2037 	}
2038 #endif	/* DEBUG */
2039 
2040 	if (vmp == NULL)
2041 		vmp = kmem_default_arena;
2042 
2043 	/*
2044 	 * If this kmem cache has an identifier vmem arena as its source, mark
2045 	 * it such to allow kmem_reap_idspace().
2046 	 */
2047 	ASSERT(!(cflags & KMC_IDENTIFIER));   /* consumer should not set this */
2048 	if (vmp->vm_cflags & VMC_IDENTIFIER)
2049 		cflags |= KMC_IDENTIFIER;
2050 
2051 	/*
2052 	 * Get a kmem_cache structure.  We arrange that cp->cache_cpu[]
2053 	 * is aligned on a KMEM_CPU_CACHE_SIZE boundary to prevent
2054 	 * false sharing of per-CPU data.
2055 	 */
2056 	cp = vmem_xalloc(kmem_cache_arena, csize, KMEM_CPU_CACHE_SIZE,
2057 	    P2NPHASE(csize, KMEM_CPU_CACHE_SIZE), 0, NULL, NULL, VM_SLEEP);
2058 	bzero(cp, csize);
2059 
2060 	if (align == 0)
2061 		align = KMEM_ALIGN;
2062 
2063 	/*
2064 	 * If we're not at least KMEM_ALIGN aligned, we can't use free
2065 	 * memory to hold bufctl information (because we can't safely
2066 	 * perform word loads and stores on it).
2067 	 */
2068 	if (align < KMEM_ALIGN)
2069 		cflags |= KMC_NOTOUCH;
2070 
2071 	if ((align & (align - 1)) != 0 || align > vmp->vm_quantum)
2072 		panic("kmem_cache_create: bad alignment %lu", align);
2073 
2074 	mutex_enter(&kmem_flags_lock);
2075 	if (kmem_flags & KMF_RANDOMIZE)
2076 		kmem_flags = (((kmem_flags | ~KMF_RANDOM) + 1) & KMF_RANDOM) |
2077 		    KMF_RANDOMIZE;
2078 	cp->cache_flags = (kmem_flags | cflags) & KMF_DEBUG;
2079 	mutex_exit(&kmem_flags_lock);
2080 
2081 	/*
2082 	 * Make sure all the various flags are reasonable.
2083 	 */
2084 	ASSERT(!(cflags & KMC_NOHASH) || !(cflags & KMC_NOTOUCH));
2085 
2086 	if (cp->cache_flags & KMF_LITE) {
2087 		if (bufsize >= kmem_lite_minsize &&
2088 		    align <= kmem_lite_maxalign &&
2089 		    P2PHASE(bufsize, kmem_lite_maxalign) != 0) {
2090 			cp->cache_flags |= KMF_BUFTAG;
2091 			cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
2092 		} else {
2093 			cp->cache_flags &= ~KMF_DEBUG;
2094 		}
2095 	}
2096 
2097 	if (cp->cache_flags & KMF_DEADBEEF)
2098 		cp->cache_flags |= KMF_REDZONE;
2099 
2100 	if ((cflags & KMC_QCACHE) && (cp->cache_flags & KMF_AUDIT))
2101 		cp->cache_flags |= KMF_NOMAGAZINE;
2102 
2103 	if (cflags & KMC_NODEBUG)
2104 		cp->cache_flags &= ~KMF_DEBUG;
2105 
2106 	if (cflags & KMC_NOTOUCH)
2107 		cp->cache_flags &= ~KMF_TOUCH;
2108 
2109 	if (cflags & KMC_NOHASH)
2110 		cp->cache_flags &= ~(KMF_AUDIT | KMF_FIREWALL);
2111 
2112 	if (cflags & KMC_NOMAGAZINE)
2113 		cp->cache_flags |= KMF_NOMAGAZINE;
2114 
2115 	if ((cp->cache_flags & KMF_AUDIT) && !(cflags & KMC_NOTOUCH))
2116 		cp->cache_flags |= KMF_REDZONE;
2117 
2118 	if (!(cp->cache_flags & KMF_AUDIT))
2119 		cp->cache_flags &= ~KMF_CONTENTS;
2120 
2121 	if ((cp->cache_flags & KMF_BUFTAG) && bufsize >= kmem_minfirewall &&
2122 	    !(cp->cache_flags & KMF_LITE) && !(cflags & KMC_NOHASH))
2123 		cp->cache_flags |= KMF_FIREWALL;
2124 
2125 	if (vmp != kmem_default_arena || kmem_firewall_arena == NULL)
2126 		cp->cache_flags &= ~KMF_FIREWALL;
2127 
2128 	if (cp->cache_flags & KMF_FIREWALL) {
2129 		cp->cache_flags &= ~KMF_BUFTAG;
2130 		cp->cache_flags |= KMF_NOMAGAZINE;
2131 		ASSERT(vmp == kmem_default_arena);
2132 		vmp = kmem_firewall_arena;
2133 	}
2134 
2135 	/*
2136 	 * Set cache properties.
2137 	 */
2138 	(void) strncpy(cp->cache_name, name, KMEM_CACHE_NAMELEN);
2139 	strident_canon(cp->cache_name, KMEM_CACHE_NAMELEN);
2140 	cp->cache_bufsize = bufsize;
2141 	cp->cache_align = align;
2142 	cp->cache_constructor = constructor;
2143 	cp->cache_destructor = destructor;
2144 	cp->cache_reclaim = reclaim;
2145 	cp->cache_private = private;
2146 	cp->cache_arena = vmp;
2147 	cp->cache_cflags = cflags;
2148 
2149 	/*
2150 	 * Determine the chunk size.
2151 	 */
2152 	chunksize = bufsize;
2153 
2154 	if (align >= KMEM_ALIGN) {
2155 		chunksize = P2ROUNDUP(chunksize, KMEM_ALIGN);
2156 		cp->cache_bufctl = chunksize - KMEM_ALIGN;
2157 	}
2158 
2159 	if (cp->cache_flags & KMF_BUFTAG) {
2160 		cp->cache_bufctl = chunksize;
2161 		cp->cache_buftag = chunksize;
2162 		if (cp->cache_flags & KMF_LITE)
2163 			chunksize += KMEM_BUFTAG_LITE_SIZE(kmem_lite_count);
2164 		else
2165 			chunksize += sizeof (kmem_buftag_t);
2166 	}
2167 
2168 	if (cp->cache_flags & KMF_DEADBEEF) {
2169 		cp->cache_verify = MIN(cp->cache_buftag, kmem_maxverify);
2170 		if (cp->cache_flags & KMF_LITE)
2171 			cp->cache_verify = sizeof (uint64_t);
2172 	}
2173 
2174 	cp->cache_contents = MIN(cp->cache_bufctl, kmem_content_maxsave);
2175 
2176 	cp->cache_chunksize = chunksize = P2ROUNDUP(chunksize, align);
2177 
2178 	/*
2179 	 * Now that we know the chunk size, determine the optimal slab size.
2180 	 */
2181 	if (vmp == kmem_firewall_arena) {
2182 		cp->cache_slabsize = P2ROUNDUP(chunksize, vmp->vm_quantum);
2183 		cp->cache_mincolor = cp->cache_slabsize - chunksize;
2184 		cp->cache_maxcolor = cp->cache_mincolor;
2185 		cp->cache_flags |= KMF_HASH;
2186 		ASSERT(!(cp->cache_flags & KMF_BUFTAG));
2187 	} else if ((cflags & KMC_NOHASH) || (!(cflags & KMC_NOTOUCH) &&
2188 	    !(cp->cache_flags & KMF_AUDIT) &&
2189 	    chunksize < vmp->vm_quantum / KMEM_VOID_FRACTION)) {
2190 		cp->cache_slabsize = vmp->vm_quantum;
2191 		cp->cache_mincolor = 0;
2192 		cp->cache_maxcolor =
2193 		    (cp->cache_slabsize - sizeof (kmem_slab_t)) % chunksize;
2194 		ASSERT(chunksize + sizeof (kmem_slab_t) <= cp->cache_slabsize);
2195 		ASSERT(!(cp->cache_flags & KMF_AUDIT));
2196 	} else {
2197 		size_t chunks, bestfit, waste, slabsize;
2198 		size_t minwaste = LONG_MAX;
2199 
2200 		for (chunks = 1; chunks <= KMEM_VOID_FRACTION; chunks++) {
2201 			slabsize = P2ROUNDUP(chunksize * chunks,
2202 			    vmp->vm_quantum);
2203 			chunks = slabsize / chunksize;
2204 			waste = (slabsize % chunksize) / chunks;
2205 			if (waste < minwaste) {
2206 				minwaste = waste;
2207 				bestfit = slabsize;
2208 			}
2209 		}
2210 		if (cflags & KMC_QCACHE)
2211 			bestfit = VMEM_QCACHE_SLABSIZE(vmp->vm_qcache_max);
2212 		cp->cache_slabsize = bestfit;
2213 		cp->cache_mincolor = 0;
2214 		cp->cache_maxcolor = bestfit % chunksize;
2215 		cp->cache_flags |= KMF_HASH;
2216 	}
2217 
2218 	if (cp->cache_flags & KMF_HASH) {
2219 		ASSERT(!(cflags & KMC_NOHASH));
2220 		cp->cache_bufctl_cache = (cp->cache_flags & KMF_AUDIT) ?
2221 		    kmem_bufctl_audit_cache : kmem_bufctl_cache;
2222 	}
2223 
2224 	if (cp->cache_maxcolor >= vmp->vm_quantum)
2225 		cp->cache_maxcolor = vmp->vm_quantum - 1;
2226 
2227 	cp->cache_color = cp->cache_mincolor;
2228 
2229 	/*
2230 	 * Initialize the rest of the slab layer.
2231 	 */
2232 	mutex_init(&cp->cache_lock, NULL, MUTEX_DEFAULT, NULL);
2233 
2234 	cp->cache_freelist = &cp->cache_nullslab;
2235 	cp->cache_nullslab.slab_cache = cp;
2236 	cp->cache_nullslab.slab_refcnt = -1;
2237 	cp->cache_nullslab.slab_next = &cp->cache_nullslab;
2238 	cp->cache_nullslab.slab_prev = &cp->cache_nullslab;
2239 
2240 	if (cp->cache_flags & KMF_HASH) {
2241 		cp->cache_hash_table = vmem_alloc(kmem_hash_arena,
2242 		    KMEM_HASH_INITIAL * sizeof (void *), VM_SLEEP);
2243 		bzero(cp->cache_hash_table,
2244 		    KMEM_HASH_INITIAL * sizeof (void *));
2245 		cp->cache_hash_mask = KMEM_HASH_INITIAL - 1;
2246 		cp->cache_hash_shift = highbit((ulong_t)chunksize) - 1;
2247 	}
2248 
2249 	/*
2250 	 * Initialize the depot.
2251 	 */
2252 	mutex_init(&cp->cache_depot_lock, NULL, MUTEX_DEFAULT, NULL);
2253 
2254 	for (mtp = kmem_magtype; chunksize <= mtp->mt_minbuf; mtp++)
2255 		continue;
2256 
2257 	cp->cache_magtype = mtp;
2258 
2259 	/*
2260 	 * Initialize the CPU layer.
2261 	 */
2262 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++) {
2263 		kmem_cpu_cache_t *ccp = &cp->cache_cpu[cpu_seqid];
2264 		mutex_init(&ccp->cc_lock, NULL, MUTEX_DEFAULT, NULL);
2265 		ccp->cc_flags = cp->cache_flags;
2266 		ccp->cc_rounds = -1;
2267 		ccp->cc_prounds = -1;
2268 	}
2269 
2270 	/*
2271 	 * Create the cache's kstats.
2272 	 */
2273 	if ((cp->cache_kstat = kstat_create("unix", 0, cp->cache_name,
2274 	    "kmem_cache", KSTAT_TYPE_NAMED,
2275 	    sizeof (kmem_cache_kstat) / sizeof (kstat_named_t),
2276 	    KSTAT_FLAG_VIRTUAL)) != NULL) {
2277 		cp->cache_kstat->ks_data = &kmem_cache_kstat;
2278 		cp->cache_kstat->ks_update = kmem_cache_kstat_update;
2279 		cp->cache_kstat->ks_private = cp;
2280 		cp->cache_kstat->ks_lock = &kmem_cache_kstat_lock;
2281 		kstat_install(cp->cache_kstat);
2282 	}
2283 
2284 	/*
2285 	 * Add the cache to the global list.  This makes it visible
2286 	 * to kmem_update(), so the cache must be ready for business.
2287 	 */
2288 	mutex_enter(&kmem_cache_lock);
2289 	cp->cache_next = cnext = &kmem_null_cache;
2290 	cp->cache_prev = cprev = kmem_null_cache.cache_prev;
2291 	cnext->cache_prev = cp;
2292 	cprev->cache_next = cp;
2293 	mutex_exit(&kmem_cache_lock);
2294 
2295 	if (kmem_ready)
2296 		kmem_cache_magazine_enable(cp);
2297 
2298 	return (cp);
2299 }
2300 
2301 void
2302 kmem_cache_destroy(kmem_cache_t *cp)
2303 {
2304 	int cpu_seqid;
2305 
2306 	/*
2307 	 * Remove the cache from the global cache list so that no one else
2308 	 * can schedule tasks on its behalf, wait for any pending tasks to
2309 	 * complete, purge the cache, and then destroy it.
2310 	 */
2311 	mutex_enter(&kmem_cache_lock);
2312 	cp->cache_prev->cache_next = cp->cache_next;
2313 	cp->cache_next->cache_prev = cp->cache_prev;
2314 	cp->cache_prev = cp->cache_next = NULL;
2315 	mutex_exit(&kmem_cache_lock);
2316 
2317 	if (kmem_taskq != NULL)
2318 		taskq_wait(kmem_taskq);
2319 
2320 	kmem_cache_magazine_purge(cp);
2321 
2322 	mutex_enter(&cp->cache_lock);
2323 	if (cp->cache_buftotal != 0)
2324 		cmn_err(CE_WARN, "kmem_cache_destroy: '%s' (%p) not empty",
2325 		    cp->cache_name, (void *)cp);
2326 	cp->cache_reclaim = NULL;
2327 	/*
2328 	 * The cache is now dead.  There should be no further activity.
2329 	 * We enforce this by setting land mines in the constructor and
2330 	 * destructor routines that induce a kernel text fault if invoked.
2331 	 */
2332 	cp->cache_constructor = (int (*)(void *, void *, int))1;
2333 	cp->cache_destructor = (void (*)(void *, void *))2;
2334 	mutex_exit(&cp->cache_lock);
2335 
2336 	kstat_delete(cp->cache_kstat);
2337 
2338 	if (cp->cache_hash_table != NULL)
2339 		vmem_free(kmem_hash_arena, cp->cache_hash_table,
2340 		    (cp->cache_hash_mask + 1) * sizeof (void *));
2341 
2342 	for (cpu_seqid = 0; cpu_seqid < max_ncpus; cpu_seqid++)
2343 		mutex_destroy(&cp->cache_cpu[cpu_seqid].cc_lock);
2344 
2345 	mutex_destroy(&cp->cache_depot_lock);
2346 	mutex_destroy(&cp->cache_lock);
2347 
2348 	vmem_free(kmem_cache_arena, cp, KMEM_CACHE_SIZE(max_ncpus));
2349 }
2350 
2351 /*ARGSUSED*/
2352 static int
2353 kmem_cpu_setup(cpu_setup_t what, int id, void *arg)
2354 {
2355 	ASSERT(MUTEX_HELD(&cpu_lock));
2356 	if (what == CPU_UNCONFIG) {
2357 		kmem_cache_applyall(kmem_cache_magazine_purge,
2358 		    kmem_taskq, TQ_SLEEP);
2359 		kmem_cache_applyall(kmem_cache_magazine_enable,
2360 		    kmem_taskq, TQ_SLEEP);
2361 	}
2362 	return (0);
2363 }
2364 
2365 static void
2366 kmem_cache_init(int pass, int use_large_pages)
2367 {
2368 	int i;
2369 	size_t size;
2370 	kmem_cache_t *cp;
2371 	kmem_magtype_t *mtp;
2372 	char name[KMEM_CACHE_NAMELEN + 1];
2373 
2374 	for (i = 0; i < sizeof (kmem_magtype) / sizeof (*mtp); i++) {
2375 		mtp = &kmem_magtype[i];
2376 		(void) sprintf(name, "kmem_magazine_%d", mtp->mt_magsize);
2377 		mtp->mt_cache = kmem_cache_create(name,
2378 		    (mtp->mt_magsize + 1) * sizeof (void *),
2379 		    mtp->mt_align, NULL, NULL, NULL, NULL,
2380 		    kmem_msb_arena, KMC_NOHASH);
2381 	}
2382 
2383 	kmem_slab_cache = kmem_cache_create("kmem_slab_cache",
2384 	    sizeof (kmem_slab_t), 0, NULL, NULL, NULL, NULL,
2385 	    kmem_msb_arena, KMC_NOHASH);
2386 
2387 	kmem_bufctl_cache = kmem_cache_create("kmem_bufctl_cache",
2388 	    sizeof (kmem_bufctl_t), 0, NULL, NULL, NULL, NULL,
2389 	    kmem_msb_arena, KMC_NOHASH);
2390 
2391 	kmem_bufctl_audit_cache = kmem_cache_create("kmem_bufctl_audit_cache",
2392 	    sizeof (kmem_bufctl_audit_t), 0, NULL, NULL, NULL, NULL,
2393 	    kmem_msb_arena, KMC_NOHASH);
2394 
2395 	if (pass == 2) {
2396 		kmem_va_arena = vmem_create("kmem_va",
2397 		    NULL, 0, PAGESIZE,
2398 		    vmem_alloc, vmem_free, heap_arena,
2399 		    8 * PAGESIZE, VM_SLEEP);
2400 
2401 		if (use_large_pages) {
2402 			kmem_default_arena = vmem_xcreate("kmem_default",
2403 			    NULL, 0, PAGESIZE,
2404 			    segkmem_alloc_lp, segkmem_free_lp, kmem_va_arena,
2405 			    0, VM_SLEEP);
2406 		} else {
2407 			kmem_default_arena = vmem_create("kmem_default",
2408 			    NULL, 0, PAGESIZE,
2409 			    segkmem_alloc, segkmem_free, kmem_va_arena,
2410 			    0, VM_SLEEP);
2411 		}
2412 	} else {
2413 		/*
2414 		 * During the first pass, the kmem_alloc_* caches
2415 		 * are treated as metadata.
2416 		 */
2417 		kmem_default_arena = kmem_msb_arena;
2418 	}
2419 
2420 	/*
2421 	 * Set up the default caches to back kmem_alloc()
2422 	 */
2423 	size = KMEM_ALIGN;
2424 	for (i = 0; i < sizeof (kmem_alloc_sizes) / sizeof (int); i++) {
2425 		size_t align = KMEM_ALIGN;
2426 		size_t cache_size = kmem_alloc_sizes[i];
2427 		/*
2428 		 * If they allocate a multiple of the coherency granularity,
2429 		 * they get a coherency-granularity-aligned address.
2430 		 */
2431 		if (IS_P2ALIGNED(cache_size, 64))
2432 			align = 64;
2433 		if (IS_P2ALIGNED(cache_size, PAGESIZE))
2434 			align = PAGESIZE;
2435 		(void) sprintf(name, "kmem_alloc_%lu", cache_size);
2436 		cp = kmem_cache_create(name, cache_size, align,
2437 		    NULL, NULL, NULL, NULL, NULL, KMC_KMEM_ALLOC);
2438 		while (size <= cache_size) {
2439 			kmem_alloc_table[(size - 1) >> KMEM_ALIGN_SHIFT] = cp;
2440 			size += KMEM_ALIGN;
2441 		}
2442 	}
2443 }
2444 
2445 void
2446 kmem_init(void)
2447 {
2448 	kmem_cache_t *cp;
2449 	int old_kmem_flags = kmem_flags;
2450 	int use_large_pages = 0;
2451 	size_t maxverify, minfirewall;
2452 
2453 	kstat_init();
2454 
2455 	/*
2456 	 * Small-memory systems (< 24 MB) can't handle kmem_flags overhead.
2457 	 */
2458 	if (physmem < btop(24 << 20) && !(old_kmem_flags & KMF_STICKY))
2459 		kmem_flags = 0;
2460 
2461 	/*
2462 	 * Don't do firewalled allocations if the heap is less than 1TB
2463 	 * (i.e. on a 32-bit kernel)
2464 	 * The resulting VM_NEXTFIT allocations would create too much
2465 	 * fragmentation in a small heap.
2466 	 */
2467 #if defined(_LP64)
2468 	maxverify = minfirewall = PAGESIZE / 2;
2469 #else
2470 	maxverify = minfirewall = ULONG_MAX;
2471 #endif
2472 
2473 	/* LINTED */
2474 	ASSERT(sizeof (kmem_cpu_cache_t) == KMEM_CPU_CACHE_SIZE);
2475 
2476 	kmem_null_cache.cache_next = &kmem_null_cache;
2477 	kmem_null_cache.cache_prev = &kmem_null_cache;
2478 
2479 	kmem_metadata_arena = vmem_create("kmem_metadata", NULL, 0, PAGESIZE,
2480 	    vmem_alloc, vmem_free, heap_arena, 8 * PAGESIZE,
2481 	    VM_SLEEP | VMC_NO_QCACHE);
2482 
2483 	kmem_msb_arena = vmem_create("kmem_msb", NULL, 0,
2484 	    PAGESIZE, segkmem_alloc, segkmem_free, kmem_metadata_arena, 0,
2485 	    VM_SLEEP);
2486 
2487 	kmem_cache_arena = vmem_create("kmem_cache", NULL, 0, KMEM_ALIGN,
2488 	    segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
2489 
2490 	kmem_hash_arena = vmem_create("kmem_hash", NULL, 0, KMEM_ALIGN,
2491 	    segkmem_alloc, segkmem_free, kmem_metadata_arena, 0, VM_SLEEP);
2492 
2493 	kmem_log_arena = vmem_create("kmem_log", NULL, 0, KMEM_ALIGN,
2494 	    segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
2495 
2496 	kmem_firewall_va_arena = vmem_create("kmem_firewall_va",
2497 	    NULL, 0, PAGESIZE,
2498 	    kmem_firewall_va_alloc, kmem_firewall_va_free, heap_arena,
2499 	    0, VM_SLEEP);
2500 
2501 	kmem_firewall_arena = vmem_create("kmem_firewall", NULL, 0, PAGESIZE,
2502 	    segkmem_alloc, segkmem_free, kmem_firewall_va_arena, 0, VM_SLEEP);
2503 
2504 	/* temporary oversize arena for mod_read_system_file */
2505 	kmem_oversize_arena = vmem_create("kmem_oversize", NULL, 0, PAGESIZE,
2506 	    segkmem_alloc, segkmem_free, heap_arena, 0, VM_SLEEP);
2507 
2508 	kmem_null_cache.cache_next = &kmem_null_cache;
2509 	kmem_null_cache.cache_prev = &kmem_null_cache;
2510 
2511 	kmem_reap_interval = 15 * hz;
2512 
2513 	/*
2514 	 * Read /etc/system.  This is a chicken-and-egg problem because
2515 	 * kmem_flags may be set in /etc/system, but mod_read_system_file()
2516 	 * needs to use the allocator.  The simplest solution is to create
2517 	 * all the standard kmem caches, read /etc/system, destroy all the
2518 	 * caches we just created, and then create them all again in light
2519 	 * of the (possibly) new kmem_flags and other kmem tunables.
2520 	 */
2521 	kmem_cache_init(1, 0);
2522 
2523 	mod_read_system_file(boothowto & RB_ASKNAME);
2524 
2525 	while ((cp = kmem_null_cache.cache_prev) != &kmem_null_cache)
2526 		kmem_cache_destroy(cp);
2527 
2528 	vmem_destroy(kmem_oversize_arena);
2529 
2530 	if (old_kmem_flags & KMF_STICKY)
2531 		kmem_flags = old_kmem_flags;
2532 
2533 	if (!(kmem_flags & KMF_AUDIT))
2534 		vmem_seg_size = offsetof(vmem_seg_t, vs_thread);
2535 
2536 	if (kmem_maxverify == 0)
2537 		kmem_maxverify = maxverify;
2538 
2539 	if (kmem_minfirewall == 0)
2540 		kmem_minfirewall = minfirewall;
2541 
2542 	/*
2543 	 * give segkmem a chance to figure out if we are using large pages
2544 	 * for the kernel heap
2545 	 */
2546 	use_large_pages = segkmem_lpsetup();
2547 
2548 	/*
2549 	 * To protect against corruption, we keep the actual number of callers
2550 	 * KMF_LITE records seperate from the tunable.  We arbitrarily clamp
2551 	 * to 16, since the overhead for small buffers quickly gets out of
2552 	 * hand.
2553 	 *
2554 	 * The real limit would depend on the needs of the largest KMC_NOHASH
2555 	 * cache.
2556 	 */
2557 	kmem_lite_count = MIN(MAX(0, kmem_lite_pcs), 16);
2558 	kmem_lite_pcs = kmem_lite_count;
2559 
2560 	/*
2561 	 * Normally, we firewall oversized allocations when possible, but
2562 	 * if we are using large pages for kernel memory, and we don't have
2563 	 * any non-LITE debugging flags set, we want to allocate oversized
2564 	 * buffers from large pages, and so skip the firewalling.
2565 	 */
2566 	if (use_large_pages &&
2567 	    ((kmem_flags & KMF_LITE) || !(kmem_flags & KMF_DEBUG))) {
2568 		kmem_oversize_arena = vmem_xcreate("kmem_oversize", NULL, 0,
2569 		    PAGESIZE, segkmem_alloc_lp, segkmem_free_lp, heap_arena,
2570 		    0, VM_SLEEP);
2571 	} else {
2572 		kmem_oversize_arena = vmem_create("kmem_oversize",
2573 		    NULL, 0, PAGESIZE,
2574 		    segkmem_alloc, segkmem_free, kmem_minfirewall < ULONG_MAX?
2575 		    kmem_firewall_va_arena : heap_arena, 0, VM_SLEEP);
2576 	}
2577 
2578 	kmem_cache_init(2, use_large_pages);
2579 
2580 	if (kmem_flags & (KMF_AUDIT | KMF_RANDOMIZE)) {
2581 		if (kmem_transaction_log_size == 0)
2582 			kmem_transaction_log_size = kmem_maxavail() / 50;
2583 		kmem_transaction_log = kmem_log_init(kmem_transaction_log_size);
2584 	}
2585 
2586 	if (kmem_flags & (KMF_CONTENTS | KMF_RANDOMIZE)) {
2587 		if (kmem_content_log_size == 0)
2588 			kmem_content_log_size = kmem_maxavail() / 50;
2589 		kmem_content_log = kmem_log_init(kmem_content_log_size);
2590 	}
2591 
2592 	kmem_failure_log = kmem_log_init(kmem_failure_log_size);
2593 
2594 	kmem_slab_log = kmem_log_init(kmem_slab_log_size);
2595 
2596 	/*
2597 	 * Initialize STREAMS message caches so allocb() is available.
2598 	 * This allows us to initialize the logging framework (cmn_err(9F),
2599 	 * strlog(9F), etc) so we can start recording messages.
2600 	 */
2601 	streams_msg_init();
2602 
2603 	/*
2604 	 * Initialize the ZSD framework in Zones so modules loaded henceforth
2605 	 * can register their callbacks.
2606 	 */
2607 	zone_zsd_init();
2608 
2609 	log_init();
2610 	taskq_init();
2611 
2612 	/*
2613 	 * Warn about invalid or dangerous values of kmem_flags.
2614 	 * Always warn about unsupported values.
2615 	 */
2616 	if (((kmem_flags & ~(KMF_AUDIT | KMF_DEADBEEF | KMF_REDZONE |
2617 	    KMF_CONTENTS | KMF_LITE)) != 0) ||
2618 	    ((kmem_flags & KMF_LITE) && kmem_flags != KMF_LITE))
2619 		cmn_err(CE_WARN, "kmem_flags set to unsupported value 0x%x. "
2620 		    "See the Solaris Tunable Parameters Reference Manual.",
2621 		    kmem_flags);
2622 
2623 #ifdef DEBUG
2624 	if ((kmem_flags & KMF_DEBUG) == 0)
2625 		cmn_err(CE_NOTE, "kmem debugging disabled.");
2626 #else
2627 	/*
2628 	 * For non-debug kernels, the only "normal" flags are 0, KMF_LITE,
2629 	 * KMF_REDZONE, and KMF_CONTENTS (the last because it is only enabled
2630 	 * if KMF_AUDIT is set). We should warn the user about the performance
2631 	 * penalty of KMF_AUDIT or KMF_DEADBEEF if they are set and KMF_LITE
2632 	 * isn't set (since that disables AUDIT).
2633 	 */
2634 	if (!(kmem_flags & KMF_LITE) &&
2635 	    (kmem_flags & (KMF_AUDIT | KMF_DEADBEEF)) != 0)
2636 		cmn_err(CE_WARN, "High-overhead kmem debugging features "
2637 		    "enabled (kmem_flags = 0x%x).  Performance degradation "
2638 		    "and large memory overhead possible. See the Solaris "
2639 		    "Tunable Parameters Reference Manual.", kmem_flags);
2640 #endif /* not DEBUG */
2641 
2642 	kmem_cache_applyall(kmem_cache_magazine_enable, NULL, TQ_SLEEP);
2643 
2644 	kmem_ready = 1;
2645 
2646 	/*
2647 	 * Initialize the platform-specific aligned/DMA memory allocator.
2648 	 */
2649 	ka_init();
2650 
2651 	/*
2652 	 * Initialize 32-bit ID cache.
2653 	 */
2654 	id32_init();
2655 
2656 	/*
2657 	 * Initialize the networking stack so modules loaded can
2658 	 * register their callbacks.
2659 	 */
2660 	netstack_init();
2661 }
2662 
2663 void
2664 kmem_thread_init(void)
2665 {
2666 	kmem_taskq = taskq_create_instance("kmem_taskq", 0, 1, minclsyspri,
2667 	    300, INT_MAX, TASKQ_PREPOPULATE);
2668 }
2669 
2670 void
2671 kmem_mp_init(void)
2672 {
2673 	mutex_enter(&cpu_lock);
2674 	register_cpu_setup_func(kmem_cpu_setup, NULL);
2675 	mutex_exit(&cpu_lock);
2676 
2677 	kmem_update_timeout(NULL);
2678 }
2679