xref: /titanic_50/usr/src/lib/libmtmalloc/common/mtmalloc.c (revision 79d2f1e1d69dd8621f962e48584f1e2434b3c837)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <mtmalloc.h>
30 #include "mtmalloc_impl.h"
31 #include <unistd.h>
32 #include <synch.h>
33 #include <thread.h>
34 #include <pthread.h>
35 #include <stdio.h>
36 #include <limits.h>
37 #include <errno.h>
38 #include <string.h>
39 #include <strings.h>
40 #include <sys/param.h>
41 #include <sys/sysmacros.h>
42 
43 /*
44  * To turn on the asserts just compile -DDEBUG
45  */
46 
47 #ifndef	DEBUG
48 #define	NDEBUG
49 #endif
50 
51 #include <assert.h>
52 
53 /*
54  * The MT hot malloc implementation contained herein is designed to be
55  * plug-compatible with the libc version of malloc. It is not intended
56  * to replace that implementation until we decide that it is ok to break
57  * customer apps (Solaris 3.0).
58  *
59  * For requests up to 2^^16, the allocator initializes itself into NCPUS
60  * worth of chains of caches. When a memory request is made, the calling thread
61  * is vectored into one of NCPUS worth of caches.  The LWP id gives us a cheap,
62  * contention-reducing index to use, eventually, this should be replaced with
63  * the actual CPU sequence number, when an interface to get it is available.
64  *
65  * Once the thread is vectored into one of the list of caches the real
66  * allocation of the memory begins. The size is determined to figure out which
67  * bucket the allocation should be satisfied from. The management of free
68  * buckets is done via a bitmask. A free bucket is represented by a 1. The
69  * first free bit represents the first free bucket. The position of the bit,
70  * represents the position of the bucket in the arena.
71  *
72  * When the memory from the arena is handed out, the address of the cache
73  * control structure is written in the word preceeding the returned memory.
74  * This cache control address is used during free() to mark the buffer free
75  * in the cache control structure.
76  *
77  * When all available memory in a cache has been depleted, a new chunk of memory
78  * is allocated via sbrk(). The new cache is allocated from this chunk of memory
79  * and initialized in the function create_cache(). New caches are installed at
80  * the front of a singly linked list of the same size memory pools. This helps
81  * to ensure that there will tend to be available memory in the beginning of the
82  * list.
83  *
84  * Long linked lists hurt performance. To decrease this effect, there is a
85  * tunable, requestsize, that bumps up the sbrk allocation size and thus
86  * increases the number of available blocks within an arena.  We also keep
87  * a "hint" for each cache list, which is the last cache in the list allocated
88  * from.  This lowers the cost of searching if there are a lot of fully
89  * allocated blocks at the front of the list.
90  *
91  * For requests greater than 2^^16 (oversize allocations), there are two pieces
92  * of overhead. There is the OVERHEAD used to hold the cache addr
93  * (&oversize_list), plus an oversize_t structure to further describe the block.
94  *
95  * The oversize list is kept as defragmented as possible by coalescing
96  * freed oversized allocations with adjacent neighbors.
97  *
98  * Addresses handed out are stored in a hash table, and are aligned on
99  * MTMALLOC_MIN_ALIGN-byte boundaries at both ends. Request sizes are rounded-up
100  * where necessary in order to achieve this. This eases the implementation of
101  * MTDEBUGPATTERN and MTINITPATTERN, particularly where coalescing occurs.
102  *
103  * A memalign allocation takes memalign header overhead.  There's two
104  * types of memalign headers distinguished by MTMALLOC_MEMALIGN_MAGIC
105  * and MTMALLOC_MEMALIGN_MIN_MAGIC.  When the size of memory taken to
106  * get to the aligned address from malloc'ed address is the minimum size
107  * OVERHEAD, we create a header taking only one OVERHEAD space with magic
108  * number MTMALLOC_MEMALIGN_MIN_MAGIC, and we know by subtracting OVERHEAD
109  * from memaligned address, we can get to the malloc'ed address. Otherwise,
110  * we create a memalign header taking two OVERHEAD space, one stores
111  * MTMALLOC_MEMALIGN_MAGIC magic number, the other one points back to the
112  * malloc'ed address.
113  */
114 
115 #if defined(__i386) || defined(__amd64)
116 #include <arpa/inet.h>	/* for htonl() */
117 #endif
118 
119 static void * morecore(size_t);
120 static void create_cache(cache_t *, size_t bufsize, uint_t hunks);
121 static void * malloc_internal(size_t, percpu_t *);
122 static void * oversize(size_t);
123 static oversize_t *find_oversize(size_t);
124 static void add_oversize(oversize_t *);
125 static void copy_pattern(uint32_t, void *, size_t);
126 static void * verify_pattern(uint32_t, void *, size_t);
127 static void reinit_cpu_list(void);
128 static void reinit_cache(cache_t *);
129 static void free_oversize(oversize_t *);
130 static oversize_t *oversize_header_alloc(uintptr_t, size_t);
131 
132 /*
133  * oversize hash table stuff
134  */
135 #define	NUM_BUCKETS	67	/* must be prime */
136 #define	HASH_OVERSIZE(caddr)	((uintptr_t)(caddr) % NUM_BUCKETS)
137 oversize_t *ovsz_hashtab[NUM_BUCKETS];
138 
139 #define	ALIGN(x, a)	((((uintptr_t)(x) + ((uintptr_t)(a) - 1)) \
140 			& ~((uintptr_t)(a) - 1)))
141 
142 /* need this to deal with little endianess of x86 */
143 #if defined(__i386) || defined(__amd64)
144 #define	FLIP_EM(x)	htonl((x))
145 #else
146 #define	FLIP_EM(x)	(x)
147 #endif
148 
149 #define	INSERT_ONLY			0
150 #define	COALESCE_LEFT			0x00000001
151 #define	COALESCE_RIGHT			0x00000002
152 #define	COALESCE_WITH_BOTH_SIDES	(COALESCE_LEFT | COALESCE_RIGHT)
153 
154 #define	OVERHEAD	8	/* size needed to write cache addr */
155 #define	HUNKSIZE	8192	/* just a multiplier */
156 
157 #define	MAX_CACHED_SHIFT	16	/* 64K is the max cached size */
158 #define	MAX_CACHED		(1 << MAX_CACHED_SHIFT)
159 #define	MIN_CACHED_SHIFT	4	/* smaller requests rounded up */
160 #define	MTMALLOC_MIN_ALIGN	8	/* min guaranteed alignment */
161 
162 /* maximum size before overflow */
163 #define	MAX_MTMALLOC	(SIZE_MAX - (SIZE_MAX % MTMALLOC_MIN_ALIGN) \
164 			- OVSZ_HEADER_SIZE)
165 
166 #define	NUM_CACHES	(MAX_CACHED_SHIFT - MIN_CACHED_SHIFT + 1)
167 #define	CACHELIST_SIZE	ALIGN(NUM_CACHES * sizeof (cache_head_t), \
168     CACHE_COHERENCY_UNIT)
169 
170 #define	MINSIZE		9	/* for requestsize, tunable */
171 #define	MAXSIZE		256	/* arbitrary, big enough, for requestsize */
172 
173 #define	FREEPATTERN	0xdeadbeef /* debug fill pattern for free buf */
174 #define	INITPATTERN	0xbaddcafe /* debug fill pattern for new buf */
175 
176 #define	misaligned(p)	((unsigned)(p) & (sizeof (int) - 1))
177 #define	IS_OVERSIZE(x, y)	(((x) < (y)) && (((x) > MAX_CACHED)? 1 : 0))
178 
179 static long requestsize = MINSIZE; /* 9 pages per cache; tunable; 9 is min */
180 
181 static uint_t cpu_mask;
182 static curcpu_func curcpu;
183 
184 static int32_t debugopt;
185 static int32_t reinit;
186 
187 static percpu_t *cpu_list;
188 static oversize_t oversize_list;
189 static mutex_t oversize_lock = DEFAULTMUTEX;
190 
191 static int ncpus = 0;
192 
193 #define	MTMALLOC_OVERSIZE_MAGIC		((uintptr_t)&oversize_list)
194 #define	MTMALLOC_MEMALIGN_MAGIC		((uintptr_t)&oversize_list + 1)
195 #define	MTMALLOC_MEMALIGN_MIN_MAGIC	((uintptr_t)&oversize_list + 2)
196 
197 /*
198  * We require allocations handed out to be aligned on MTMALLOC_MIN_ALIGN-byte
199  * boundaries. We round up sizeof (oversize_t) (when necessary) to ensure that
200  * this is achieved.
201  */
202 #define	OVSZ_SIZE		(ALIGN(sizeof (oversize_t), MTMALLOC_MIN_ALIGN))
203 #define	OVSZ_HEADER_SIZE	(OVSZ_SIZE + OVERHEAD)
204 
205 /*
206  * memalign header takes 2 OVERHEAD space.  One for memalign magic, and the
207  * other one points back to the start address of originally allocated space.
208  */
209 #define	MEMALIGN_HEADER_SIZE	2 * OVERHEAD
210 #define	MEMALIGN_HEADER_ALLOC(x, shift, malloc_addr)\
211 	if (shift == OVERHEAD)\
212 		*((uintptr_t *)((caddr_t)x - OVERHEAD)) = \
213 			MTMALLOC_MEMALIGN_MIN_MAGIC; \
214 	else {\
215 		*((uintptr_t *)((caddr_t)x - OVERHEAD)) = \
216 			MTMALLOC_MEMALIGN_MAGIC; \
217 		*((uintptr_t *)((caddr_t)x - 2 * OVERHEAD)) = \
218 			(uintptr_t)malloc_addr; \
219 	}
220 
221 void *
222 malloc(size_t bytes)
223 {
224 	percpu_t *list_rotor;
225 	uint_t	list_index;
226 
227 	if (bytes > MAX_CACHED)
228 		return (oversize(bytes));
229 
230 	list_index = (curcpu() & cpu_mask);
231 
232 	list_rotor = &cpu_list[list_index];
233 
234 	return (malloc_internal(bytes, list_rotor));
235 }
236 
237 void *
238 realloc(void * ptr, size_t bytes)
239 {
240 	void *new, *data_ptr;
241 	cache_t *cacheptr;
242 	caddr_t mem;
243 	size_t shift = 0;
244 
245 	if (ptr == NULL)
246 		return (malloc(bytes));
247 
248 	if (bytes == 0) {
249 		free(ptr);
250 		return (NULL);
251 	}
252 
253 	data_ptr = ptr;
254 	mem = (caddr_t)ptr - OVERHEAD;
255 
256 	new = malloc(bytes);
257 
258 	if (new == NULL)
259 		return (NULL);
260 
261 	/*
262 	 * If new == ptr, ptr has previously been freed. Passing a freed pointer
263 	 * to realloc() is not allowed - unless the caller specifically states
264 	 * otherwise, in which case we must avoid freeing ptr (ie new) before we
265 	 * return new. There is (obviously) no requirement to memcpy() ptr to
266 	 * new before we return.
267 	 */
268 	if (new == ptr) {
269 		if (!(debugopt & MTDOUBLEFREE))
270 			abort();
271 		return (new);
272 	}
273 
274 	if (*(uintptr_t *)mem == MTMALLOC_MEMALIGN_MAGIC) {
275 		mem -= OVERHEAD;
276 		ptr = (void *)*(uintptr_t *)mem;
277 		mem = (caddr_t)ptr - OVERHEAD;
278 		shift = (size_t)((uintptr_t)data_ptr - (uintptr_t)ptr);
279 	} else if (*(uintptr_t *)mem == MTMALLOC_MEMALIGN_MIN_MAGIC) {
280 		ptr = (void *) mem;
281 		mem -= OVERHEAD;
282 		shift = OVERHEAD;
283 	}
284 
285 	if (*(uintptr_t *)mem == MTMALLOC_OVERSIZE_MAGIC) {
286 		oversize_t *old;
287 
288 		old = (oversize_t *)(mem - OVSZ_SIZE);
289 		(void) memcpy(new, data_ptr, MIN(bytes, old->size - shift));
290 		free(ptr);
291 		return (new);
292 	}
293 
294 	cacheptr = (cache_t *)*(uintptr_t *)mem;
295 
296 	(void) memcpy(new, data_ptr,
297 		MIN(cacheptr->mt_size - OVERHEAD - shift, bytes));
298 	free(ptr);
299 
300 	return (new);
301 }
302 
303 void *
304 calloc(size_t nelem, size_t bytes)
305 {
306 	void * ptr;
307 	size_t size = nelem * bytes;
308 
309 	ptr = malloc(size);
310 	if (ptr == NULL)
311 		return (NULL);
312 	(void) memset(ptr, 0, size);
313 
314 	return (ptr);
315 }
316 
317 void
318 free(void * ptr)
319 {
320 	cache_t *cacheptr;
321 	caddr_t mem;
322 	int32_t i;
323 	caddr_t freeblocks;
324 	uintptr_t offset;
325 	uchar_t mask;
326 	int32_t which_bit, num_bytes;
327 
328 	if (ptr == NULL)
329 		return;
330 
331 	mem = (caddr_t)ptr - OVERHEAD;
332 
333 	if (*(uintptr_t *)mem == MTMALLOC_MEMALIGN_MAGIC) {
334 		mem -= OVERHEAD;
335 		ptr = (void *)*(uintptr_t *)mem;
336 		mem = (caddr_t)ptr - OVERHEAD;
337 	} else if (*(uintptr_t *)mem == MTMALLOC_MEMALIGN_MIN_MAGIC) {
338 		ptr = (void *) mem;
339 		mem -= OVERHEAD;
340 	}
341 
342 	if (*(uintptr_t *)mem == MTMALLOC_OVERSIZE_MAGIC) {
343 		oversize_t *big, **opp;
344 		int bucket;
345 
346 		big = (oversize_t *)(mem - OVSZ_SIZE);
347 		(void) mutex_lock(&oversize_lock);
348 
349 		bucket = HASH_OVERSIZE(big->addr);
350 		for (opp = &ovsz_hashtab[bucket]; *opp != NULL;
351 		    opp = &(*opp)->hash_next)
352 			if (*opp == big)
353 				break;
354 
355 		if (*opp == NULL) {
356 			if (!(debugopt & MTDOUBLEFREE))
357 				abort();
358 			(void) mutex_unlock(&oversize_lock);
359 			return;
360 		}
361 
362 		*opp = big->hash_next;	/* remove big from the hash table */
363 		big->hash_next = NULL;
364 
365 		if (debugopt & MTDEBUGPATTERN)
366 			copy_pattern(FREEPATTERN, ptr, big->size);
367 		add_oversize(big);
368 		(void) mutex_unlock(&oversize_lock);
369 		return;
370 	}
371 
372 	cacheptr = (cache_t *)*(uintptr_t *)mem;
373 	freeblocks = cacheptr->mt_freelist;
374 
375 	/*
376 	 * This is the distance measured in bits into the arena.
377 	 * The value of offset is in bytes but there is a 1-1 correlation
378 	 * between distance into the arena and distance into the
379 	 * freelist bitmask.
380 	 */
381 	offset = mem - cacheptr->mt_arena;
382 
383 	/*
384 	 * i is total number of bits to offset into freelist bitmask.
385 	 */
386 
387 	i = offset / cacheptr->mt_size;
388 
389 	num_bytes = i >> 3;
390 
391 	/*
392 	 * which_bit is the bit offset into the byte in the freelist.
393 	 * if our freelist bitmask looks like 0xf3 and we are freeing
394 	 * block 5 (ie: the 6th block) our mask will be 0xf7 after
395 	 * the free. Things go left to right that's why the mask is 0x80
396 	 * and not 0x01.
397 	 */
398 	which_bit = i - (num_bytes << 3);
399 
400 	mask = 0x80 >> which_bit;
401 
402 	freeblocks += num_bytes;
403 
404 	if (debugopt & MTDEBUGPATTERN)
405 		copy_pattern(FREEPATTERN, ptr, cacheptr->mt_size - OVERHEAD);
406 
407 	(void) mutex_lock(&cacheptr->mt_cache_lock);
408 
409 	if (*freeblocks & mask) {
410 		if (!(debugopt & MTDOUBLEFREE))
411 			abort();
412 	} else {
413 		*freeblocks |= mask;
414 		cacheptr->mt_nfree++;
415 	}
416 
417 	(void) mutex_unlock(&cacheptr->mt_cache_lock);
418 }
419 
420 void *
421 memalign(size_t alignment, size_t size)
422 {
423 	size_t alloc_size;
424 	uintptr_t offset;
425 	void *alloc_buf;
426 	void *ret_buf;
427 
428 	if (size == 0 || alignment == 0 ||
429 		misaligned(alignment) ||
430 		(alignment & (alignment - 1)) != 0) {
431 		errno = EINVAL;
432 		return (NULL);
433 	}
434 
435 	/* <= MTMALLOC_MIN_ALIGN, malloc can provide directly */
436 	if (alignment <= MTMALLOC_MIN_ALIGN)
437 		return (malloc(size));
438 
439 	alloc_size = size + alignment - MTMALLOC_MIN_ALIGN;
440 
441 	if (alloc_size < size) { /* overflow */
442 		errno = ENOMEM;
443 		return (NULL);
444 	}
445 
446 	alloc_buf = malloc(alloc_size);
447 
448 	if (alloc_buf == NULL)
449 		/* malloc sets errno */
450 		return (NULL);
451 
452 	/*
453 	 * If alloc_size > MAX_CACHED, malloc() will have returned a multiple of
454 	 * MTMALLOC_MIN_ALIGN, having rounded-up alloc_size if necessary. Since
455 	 * we will use alloc_size to return the excess fragments to the free
456 	 * list, we also round-up alloc_size if necessary.
457 	 */
458 	if ((alloc_size > MAX_CACHED) &&
459 	    (alloc_size & (MTMALLOC_MIN_ALIGN - 1)))
460 		alloc_size = ALIGN(alloc_size, MTMALLOC_MIN_ALIGN);
461 
462 	if ((offset = (uintptr_t)alloc_buf & (alignment - 1)) == 0) {
463 		/* aligned correctly */
464 
465 		size_t frag_size = alloc_size -
466 			(size + MTMALLOC_MIN_ALIGN + OVSZ_HEADER_SIZE);
467 
468 		/*
469 		 * If the leftover piece of the memory > MAX_CACHED,
470 		 * split off the piece and return it back to the freelist.
471 		 */
472 		if (IS_OVERSIZE(frag_size, alloc_size)) {
473 			oversize_t *orig, *tail;
474 			uintptr_t taddr;
475 			size_t data_size;
476 			taddr = ALIGN((uintptr_t)alloc_buf + size,
477 					MTMALLOC_MIN_ALIGN);
478 			data_size = taddr - (uintptr_t)alloc_buf;
479 			orig = (oversize_t *)((uintptr_t)alloc_buf -
480 					OVSZ_HEADER_SIZE);
481 			frag_size = orig->size - data_size -
482 					OVSZ_HEADER_SIZE;
483 			orig->size = data_size;
484 			tail = oversize_header_alloc(taddr, frag_size);
485 			free_oversize(tail);
486 		}
487 		ret_buf = alloc_buf;
488 	} else {
489 		uchar_t	oversize_bits = 0;
490 		size_t	head_sz, data_sz, tail_sz;
491 		uintptr_t ret_addr, taddr, shift, tshift;
492 		oversize_t *orig, *tail;
493 		size_t tsize;
494 
495 		/* needs to be aligned */
496 		shift = alignment - offset;
497 
498 		assert(shift >= MTMALLOC_MIN_ALIGN);
499 
500 		ret_addr = ((uintptr_t)alloc_buf + shift);
501 		ret_buf = (void *)ret_addr;
502 
503 		if (alloc_size <= MAX_CACHED) {
504 			MEMALIGN_HEADER_ALLOC(ret_addr, shift, alloc_buf);
505 			return (ret_buf);
506 		}
507 
508 		/*
509 		 * Only check for the fragments when the memory is allocted
510 		 * from oversize_list.  Split off a fragment and return it
511 		 * to the oversize freelist when it's > MAX_CACHED.
512 		 */
513 
514 		head_sz = shift - MAX(MEMALIGN_HEADER_SIZE, OVSZ_HEADER_SIZE);
515 
516 		tail_sz = alloc_size -
517 			(shift + size + MTMALLOC_MIN_ALIGN + OVSZ_HEADER_SIZE);
518 
519 		oversize_bits |= IS_OVERSIZE(head_sz, alloc_size) |
520 				IS_OVERSIZE(size, alloc_size) << DATA_SHIFT |
521 				IS_OVERSIZE(tail_sz, alloc_size) << TAIL_SHIFT;
522 
523 		switch (oversize_bits) {
524 			case NONE_OVERSIZE:
525 			case DATA_OVERSIZE:
526 				MEMALIGN_HEADER_ALLOC(ret_addr, shift,
527 					alloc_buf);
528 				break;
529 			case HEAD_OVERSIZE:
530 				/*
531 				 * If we can extend data > MAX_CACHED and have
532 				 * head still > MAX_CACHED, we split head-end
533 				 * as the case of head-end and data oversized,
534 				 * otherwise just create memalign header.
535 				 */
536 				tsize = (shift + size) - (MAX_CACHED + 8 +
537 					MTMALLOC_MIN_ALIGN + OVSZ_HEADER_SIZE);
538 
539 				if (!IS_OVERSIZE(tsize, alloc_size)) {
540 					MEMALIGN_HEADER_ALLOC(ret_addr, shift,
541 						alloc_buf);
542 					break;
543 				} else {
544 					tsize += OVSZ_HEADER_SIZE;
545 					taddr = ALIGN((uintptr_t)alloc_buf +
546 						tsize, MTMALLOC_MIN_ALIGN);
547 					tshift = ret_addr - taddr;
548 					MEMALIGN_HEADER_ALLOC(ret_addr, tshift,
549 						taddr);
550 					ret_addr = taddr;
551 					shift = ret_addr - (uintptr_t)alloc_buf;
552 				}
553 				/* FALLTHROUGH */
554 			case HEAD_AND_DATA_OVERSIZE:
555 				/*
556 				 * Split off the head fragment and
557 				 * return it back to oversize freelist.
558 				 * Create oversize header for the piece
559 				 * of (data + tail fragment).
560 				 */
561 				orig = (oversize_t *)((uintptr_t)alloc_buf -
562 						OVSZ_HEADER_SIZE);
563 				(void) oversize_header_alloc(ret_addr -
564 						OVSZ_HEADER_SIZE,
565 						(orig->size - shift));
566 				orig->size = shift - OVSZ_HEADER_SIZE;
567 
568 				/* free up the head fragment */
569 				free_oversize(orig);
570 				break;
571 			case TAIL_OVERSIZE:
572 				/*
573 				 * If we can extend data > MAX_CACHED and have
574 				 * tail-end still > MAX_CACHED, we split tail
575 				 * end, otherwise just create memalign header.
576 				 */
577 				orig = (oversize_t *)((uintptr_t)alloc_buf -
578 						OVSZ_HEADER_SIZE);
579 				tsize =  orig->size - (MAX_CACHED + 8 +
580 					shift + OVSZ_HEADER_SIZE +
581 					MTMALLOC_MIN_ALIGN);
582 				if (!IS_OVERSIZE(tsize, alloc_size)) {
583 					MEMALIGN_HEADER_ALLOC(ret_addr, shift,
584 						alloc_buf);
585 					break;
586 				} else {
587 					size = MAX_CACHED + 8;
588 				}
589 				/* FALLTHROUGH */
590 			case DATA_AND_TAIL_OVERSIZE:
591 				/*
592 				 * Split off the tail fragment and
593 				 * return it back to oversize freelist.
594 				 * Create memalign header and adjust
595 				 * the size for the piece of
596 				 * (head fragment + data).
597 				 */
598 				taddr = ALIGN(ret_addr + size,
599 						MTMALLOC_MIN_ALIGN);
600 				data_sz = (size_t)(taddr -
601 						(uintptr_t)alloc_buf);
602 				orig = (oversize_t *)((uintptr_t)alloc_buf -
603 						OVSZ_HEADER_SIZE);
604 				tsize = orig->size - data_sz;
605 				orig->size = data_sz;
606 				MEMALIGN_HEADER_ALLOC(ret_buf, shift,
607 					alloc_buf);
608 				tsize -= OVSZ_HEADER_SIZE;
609 				tail = oversize_header_alloc(taddr,  tsize);
610 				free_oversize(tail);
611 				break;
612 			case HEAD_AND_TAIL_OVERSIZE:
613 				/*
614 				 * Split off the head fragment.
615 				 * We try to free up tail-end when we can
616 				 * extend data size to (MAX_CACHED + 8)
617 				 * and remain tail-end oversized.
618 				 * The bottom line is all split pieces
619 				 * should be oversize in size.
620 				 */
621 				orig = (oversize_t *)((uintptr_t)alloc_buf -
622 					OVSZ_HEADER_SIZE);
623 				tsize =  orig->size - (MAX_CACHED + 8 +
624 					OVSZ_HEADER_SIZE + shift +
625 					MTMALLOC_MIN_ALIGN);
626 
627 				if (!IS_OVERSIZE(tsize, alloc_size)) {
628 					/*
629 					 * If the chunk is not big enough
630 					 * to make both data and tail oversize
631 					 * we just keep them as one piece.
632 					 */
633 					(void) oversize_header_alloc(ret_addr -
634 						OVSZ_HEADER_SIZE,
635 						orig->size - shift);
636 					orig->size = shift -
637 						OVSZ_HEADER_SIZE;
638 					free_oversize(orig);
639 					break;
640 				} else {
641 					/*
642 					 * extend data size > MAX_CACHED
643 					 * and handle it as head, data, tail
644 					 * are all oversized.
645 					 */
646 					size = MAX_CACHED + 8;
647 				}
648 				/* FALLTHROUGH */
649 			case ALL_OVERSIZE:
650 				/*
651 				 * split off the head and tail fragments,
652 				 * return them back to the oversize freelist.
653 				 * Alloc oversize header for data seg.
654 				 */
655 				orig = (oversize_t *)((uintptr_t)alloc_buf -
656 					OVSZ_HEADER_SIZE);
657 				tsize = orig->size;
658 				orig->size = shift - OVSZ_HEADER_SIZE;
659 				free_oversize(orig);
660 
661 				taddr = ALIGN(ret_addr + size,
662 					MTMALLOC_MIN_ALIGN);
663 				data_sz = taddr - ret_addr;
664 				assert(tsize > (shift + data_sz +
665 					OVSZ_HEADER_SIZE));
666 				tail_sz = tsize -
667 					(shift + data_sz + OVSZ_HEADER_SIZE);
668 
669 				/* create oversize header for data seg */
670 				(void) oversize_header_alloc(ret_addr -
671 					OVSZ_HEADER_SIZE, data_sz);
672 
673 				/* create oversize header for tail fragment */
674 				tail = oversize_header_alloc(taddr, tail_sz);
675 				free_oversize(tail);
676 				break;
677 			default:
678 				/* should not reach here */
679 				assert(0);
680 		}
681 	}
682 	return (ret_buf);
683 }
684 
685 
686 void *
687 valloc(size_t size)
688 {
689 	static unsigned pagesize;
690 
691 	if (size == 0)
692 		return (NULL);
693 
694 	if (!pagesize)
695 		pagesize = sysconf(_SC_PAGESIZE);
696 
697 	return (memalign(pagesize, size));
698 }
699 
700 void
701 mallocctl(int cmd, long value)
702 {
703 	switch (cmd) {
704 
705 	case MTDEBUGPATTERN:
706 		/*
707 		 * Reinitialize free blocks in case malloc() is called prior
708 		 * to mallocctl().
709 		 */
710 		if (value && !(debugopt & cmd)) {
711 			reinit++;
712 			debugopt |= cmd;
713 			reinit_cpu_list();
714 		}
715 		/*FALLTHRU*/
716 	case MTDOUBLEFREE:
717 	case MTINITBUFFER:
718 		if (value)
719 			debugopt |= cmd;
720 		else
721 			debugopt &= ~cmd;
722 		break;
723 	case MTCHUNKSIZE:
724 		if (value >= MINSIZE && value <= MAXSIZE)
725 			requestsize = value;
726 		break;
727 	default:
728 		break;
729 	}
730 }
731 
732 /*
733  * Initialization function, called from the init section of the library.
734  * No locking is required here because we are single-threaded during
735  * library initialization.
736  */
737 static void
738 setup_caches(void)
739 {
740 	uintptr_t oldbrk;
741 	uintptr_t newbrk;
742 
743 	size_t cache_space_needed;
744 	size_t padding;
745 
746 	curcpu_func new_curcpu;
747 	uint_t new_cpu_mask;
748 	percpu_t *new_cpu_list;
749 
750 	uint_t i, j;
751 	uintptr_t list_addr;
752 
753 	/*
754 	 * Get a decent "current cpu identifier", to be used to reduce
755 	 * contention.  Eventually, this should be replaced by an interface
756 	 * to get the actual CPU sequence number in libthread/liblwp.
757 	 */
758 	new_curcpu = (curcpu_func)thr_self;
759 	if ((ncpus = 2 * sysconf(_SC_NPROCESSORS_CONF)) <= 0)
760 		ncpus = 4; /* decent default value */
761 
762 	/* round ncpus up to a power of 2 */
763 	while (ncpus & (ncpus - 1))
764 		ncpus++;
765 
766 	new_cpu_mask = ncpus - 1;	/* create the cpu mask */
767 
768 	/*
769 	 * We now do some magic with the brk.  What we want to get in the
770 	 * end is a bunch of well-aligned stuff in a big initial allocation.
771 	 * Along the way, we do sanity checks to make sure no one else has
772 	 * touched the brk (which shouldn't happen, but it's always good to
773 	 * check)
774 	 *
775 	 * First, make sure sbrk is sane, and store the current brk in oldbrk.
776 	 */
777 	oldbrk = (uintptr_t)sbrk(0);
778 	if ((void *)oldbrk == (void *)-1)
779 		abort();	/* sbrk is broken -- we're doomed. */
780 
781 	/*
782 	 * Now, align the brk to a multiple of CACHE_COHERENCY_UNIT, so that
783 	 * the percpu structures and cache lists will be properly aligned.
784 	 *
785 	 *   2.  All hunks will be page-aligned, assuming HUNKSIZE >= PAGESIZE,
786 	 *	so they can be paged out individually.
787 	 */
788 	newbrk = ALIGN(oldbrk, CACHE_COHERENCY_UNIT);
789 	if (newbrk != oldbrk && (uintptr_t)sbrk(newbrk - oldbrk) != oldbrk)
790 		abort();	/* sbrk is broken -- we're doomed. */
791 
792 	/*
793 	 * For each cpu, there is one percpu_t and a list of caches
794 	 */
795 	cache_space_needed = ncpus * (sizeof (percpu_t) + CACHELIST_SIZE);
796 
797 	new_cpu_list = (percpu_t *)sbrk(cache_space_needed);
798 
799 	if (new_cpu_list == (percpu_t *)-1 ||
800 	    (uintptr_t)new_cpu_list != newbrk)
801 		abort();	/* sbrk is broken -- we're doomed. */
802 
803 	/*
804 	 * Finally, align the brk to HUNKSIZE so that all hunks are
805 	 * page-aligned, to avoid edge-effects.
806 	 */
807 
808 	newbrk = (uintptr_t)new_cpu_list + cache_space_needed;
809 
810 	padding = ALIGN(newbrk, HUNKSIZE) - newbrk;
811 
812 	if (padding > 0 && (uintptr_t)sbrk(padding) != newbrk)
813 		abort();	/* sbrk is broken -- we're doomed. */
814 
815 	list_addr = ((uintptr_t)new_cpu_list + (sizeof (percpu_t) * ncpus));
816 
817 	/* initialize the percpu list */
818 	for (i = 0; i < ncpus; i++) {
819 		new_cpu_list[i].mt_caches = (cache_head_t *)list_addr;
820 		for (j = 0; j < NUM_CACHES; j++) {
821 			new_cpu_list[i].mt_caches[j].mt_cache = NULL;
822 			new_cpu_list[i].mt_caches[j].mt_hint = NULL;
823 		}
824 
825 		(void) mutex_init(&new_cpu_list[i].mt_parent_lock,
826 		    USYNC_THREAD, NULL);
827 
828 		/* get the correct cache list alignment */
829 		list_addr += CACHELIST_SIZE;
830 	}
831 
832 	/*
833 	 * Initialize oversize listhead
834 	 */
835 	oversize_list.next_bysize = &oversize_list;
836 	oversize_list.prev_bysize = &oversize_list;
837 	oversize_list.next_byaddr = &oversize_list;
838 	oversize_list.prev_byaddr = &oversize_list;
839 	oversize_list.addr = NULL;
840 	oversize_list.size = 0;		/* sentinal */
841 
842 	/*
843 	 * Now install the global variables.
844 	 */
845 	curcpu = new_curcpu;
846 	cpu_mask = new_cpu_mask;
847 	cpu_list = new_cpu_list;
848 }
849 
850 static void
851 create_cache(cache_t *cp, size_t size, uint_t chunksize)
852 {
853 	long nblocks;
854 
855 	(void) mutex_init(&cp->mt_cache_lock, USYNC_THREAD, NULL);
856 	cp->mt_size = size;
857 	cp->mt_freelist = ((caddr_t)cp + sizeof (cache_t));
858 	cp->mt_span = chunksize * HUNKSIZE - sizeof (cache_t);
859 	cp->mt_hunks = chunksize;
860 	/*
861 	 * rough calculation. We will need to adjust later.
862 	 */
863 	nblocks = cp->mt_span / cp->mt_size;
864 	nblocks >>= 3;
865 	if (nblocks == 0) { /* less than 8 free blocks in this pool */
866 		int32_t numblocks = 0;
867 		long i = cp->mt_span;
868 		size_t sub = cp->mt_size;
869 		uchar_t mask = 0;
870 
871 		while (i > sub) {
872 			numblocks++;
873 			i -= sub;
874 		}
875 		nblocks = numblocks;
876 		cp->mt_arena = (caddr_t)ALIGN(cp->mt_freelist + 8, 8);
877 		cp->mt_nfree = numblocks;
878 		while (numblocks--) {
879 			mask |= 0x80 >> numblocks;
880 		}
881 		*(cp->mt_freelist) = mask;
882 	} else {
883 		cp->mt_arena = (caddr_t)ALIGN((caddr_t)cp->mt_freelist +
884 			nblocks, 32);
885 		/* recompute nblocks */
886 		nblocks = (uintptr_t)((caddr_t)cp->mt_freelist +
887 			cp->mt_span - cp->mt_arena) / cp->mt_size;
888 		cp->mt_nfree = ((nblocks >> 3) << 3);
889 		/* Set everything to free */
890 		(void) memset(cp->mt_freelist, 0xff, nblocks >> 3);
891 	}
892 
893 	if (debugopt & MTDEBUGPATTERN)
894 		copy_pattern(FREEPATTERN, cp->mt_arena, cp->mt_size * nblocks);
895 
896 	cp->mt_next = NULL;
897 }
898 
899 static void
900 reinit_cpu_list(void)
901 {
902 	oversize_t *wp = oversize_list.next_bysize;
903 	percpu_t *cpuptr;
904 	cache_t *thiscache;
905 	cache_head_t *cachehead;
906 
907 	/* Reinitialize free oversize blocks. */
908 	(void) mutex_lock(&oversize_lock);
909 	if (debugopt & MTDEBUGPATTERN)
910 		for (; wp != &oversize_list; wp = wp->next_bysize)
911 			copy_pattern(FREEPATTERN, wp->addr, wp->size);
912 	(void) mutex_unlock(&oversize_lock);
913 
914 	/* Reinitialize free blocks. */
915 	for (cpuptr = &cpu_list[0]; cpuptr < &cpu_list[ncpus]; cpuptr++) {
916 		(void) mutex_lock(&cpuptr->mt_parent_lock);
917 		for (cachehead = &cpuptr->mt_caches[0]; cachehead <
918 			&cpuptr->mt_caches[NUM_CACHES]; cachehead++) {
919 			for (thiscache = cachehead->mt_cache; thiscache != NULL;
920 				thiscache = thiscache->mt_next) {
921 				(void) mutex_lock(&thiscache->mt_cache_lock);
922 				if (thiscache->mt_nfree == 0) {
923 					(void) mutex_unlock(
924 					    &thiscache->mt_cache_lock);
925 					continue;
926 				}
927 				if (thiscache != NULL)
928 					reinit_cache(thiscache);
929 				(void) mutex_unlock(&thiscache->mt_cache_lock);
930 			}
931 		}
932 		(void) mutex_unlock(&cpuptr->mt_parent_lock);
933 	}
934 	reinit = 0;
935 }
936 
937 static void
938 reinit_cache(cache_t *thiscache)
939 {
940 	uint32_t *freeblocks; /* not a uintptr_t on purpose */
941 	int32_t i, n;
942 	caddr_t ret;
943 
944 	freeblocks = (uint32_t *)thiscache->mt_freelist;
945 	while (freeblocks < (uint32_t *)thiscache->mt_arena) {
946 		if (*freeblocks & 0xffffffff) {
947 		    for (i = 0; i < 32; i++) {
948 			if (FLIP_EM(*freeblocks) & (0x80000000 >> i)) {
949 				n = (uintptr_t)(((freeblocks -
950 				    (uint32_t *)thiscache->mt_freelist) << 5)
951 				    + i) * thiscache->mt_size;
952 				ret = thiscache->mt_arena + n;
953 				ret += OVERHEAD;
954 				copy_pattern(FREEPATTERN, ret,
955 				    thiscache->mt_size);
956 			}
957 		    }
958 		}
959 		freeblocks++;
960 	}
961 }
962 
963 static void *
964 malloc_internal(size_t size, percpu_t *cpuptr)
965 {
966 	cache_head_t *cachehead;
967 	cache_t *thiscache, *hintcache;
968 	int32_t i, n, logsz, bucket;
969 	uint32_t index;
970 	uint32_t *freeblocks; /* not a uintptr_t on purpose */
971 	caddr_t ret;
972 
973 	logsz = MIN_CACHED_SHIFT;
974 
975 	while (size > (1 << logsz))
976 		logsz++;
977 
978 	bucket = logsz - MIN_CACHED_SHIFT;
979 
980 	(void) mutex_lock(&cpuptr->mt_parent_lock);
981 
982 	/*
983 	 * Find a cache of the appropriate size with free buffers.
984 	 *
985 	 * We don't need to lock each cache as we check their mt_nfree count,
986 	 * since:
987 	 *	1.  We are only looking for caches with mt_nfree > 0.  If a
988 	 *	   free happens during our search, it will increment mt_nfree,
989 	 *	   which will not effect the test.
990 	 *	2.  Allocations can decrement mt_nfree, but they can't happen
991 	 *	   as long as we hold mt_parent_lock.
992 	 */
993 
994 	cachehead = &cpuptr->mt_caches[bucket];
995 
996 	/* Search through the list, starting at the mt_hint */
997 	thiscache = cachehead->mt_hint;
998 
999 	while (thiscache != NULL && thiscache->mt_nfree == 0)
1000 		thiscache = thiscache->mt_next;
1001 
1002 	if (thiscache == NULL) {
1003 		/* wrap around -- search up to the hint */
1004 		thiscache = cachehead->mt_cache;
1005 		hintcache = cachehead->mt_hint;
1006 
1007 		while (thiscache != NULL && thiscache != hintcache &&
1008 		    thiscache->mt_nfree == 0)
1009 			thiscache = thiscache->mt_next;
1010 
1011 		if (thiscache == hintcache)
1012 			thiscache = NULL;
1013 	}
1014 
1015 
1016 	if (thiscache == NULL) { /* there are no free caches */
1017 		int32_t thisrequest = requestsize;
1018 		int32_t buffer_size = (1 << logsz) + OVERHEAD;
1019 
1020 		thiscache = (cache_t *)morecore(thisrequest * HUNKSIZE);
1021 
1022 		if (thiscache == (cache_t *)-1) {
1023 		    (void) mutex_unlock(&cpuptr->mt_parent_lock);
1024 		    errno = EAGAIN;
1025 		    return (NULL);
1026 		}
1027 		create_cache(thiscache, buffer_size, thisrequest);
1028 
1029 		/* link in the new block at the beginning of the list */
1030 		thiscache->mt_next = cachehead->mt_cache;
1031 		cachehead->mt_cache = thiscache;
1032 	}
1033 
1034 	/* update the hint to the cache we found or created */
1035 	cachehead->mt_hint = thiscache;
1036 
1037 	/* thiscache now points to a cache with available space */
1038 	(void) mutex_lock(&thiscache->mt_cache_lock);
1039 
1040 	freeblocks = (uint32_t *)thiscache->mt_freelist;
1041 	while (freeblocks < (uint32_t *)thiscache->mt_arena) {
1042 		if (*freeblocks & 0xffffffff)
1043 			break;
1044 		freeblocks++;
1045 		if (freeblocks < (uint32_t *)thiscache->mt_arena &&
1046 		    *freeblocks & 0xffffffff)
1047 			break;
1048 		freeblocks++;
1049 		if (freeblocks < (uint32_t *)thiscache->mt_arena &&
1050 		    *freeblocks & 0xffffffff)
1051 			break;
1052 		freeblocks++;
1053 		if (freeblocks < (uint32_t *)thiscache->mt_arena &&
1054 		    *freeblocks & 0xffffffff)
1055 			break;
1056 		freeblocks++;
1057 	}
1058 
1059 	/*
1060 	 * the offset from mt_freelist to freeblocks is the offset into
1061 	 * the arena. Be sure to include the offset into freeblocks
1062 	 * of the bitmask. n is the offset.
1063 	 */
1064 	for (i = 0; i < 32; ) {
1065 		if (FLIP_EM(*freeblocks) & (0x80000000 >> i++))
1066 			break;
1067 		if (FLIP_EM(*freeblocks) & (0x80000000 >> i++))
1068 			break;
1069 		if (FLIP_EM(*freeblocks) & (0x80000000 >> i++))
1070 			break;
1071 		if (FLIP_EM(*freeblocks) & (0x80000000 >> i++))
1072 			break;
1073 	}
1074 	index = 0x80000000 >> --i;
1075 
1076 
1077 	*freeblocks &= FLIP_EM(~index);
1078 
1079 	thiscache->mt_nfree--;
1080 
1081 	(void) mutex_unlock(&thiscache->mt_cache_lock);
1082 	(void) mutex_unlock(&cpuptr->mt_parent_lock);
1083 
1084 	n = (uintptr_t)(((freeblocks - (uint32_t *)thiscache->mt_freelist) << 5)
1085 		+ i) * thiscache->mt_size;
1086 	/*
1087 	 * Now you have the offset in n, you've changed the free mask
1088 	 * in the freelist. Nothing left to do but find the block
1089 	 * in the arena and put the value of thiscache in the word
1090 	 * ahead of the handed out address and return the memory
1091 	 * back to the user.
1092 	 */
1093 	ret = thiscache->mt_arena + n;
1094 
1095 	/* Store the cache addr for this buf. Makes free go fast. */
1096 	*(uintptr_t *)ret = (uintptr_t)thiscache;
1097 
1098 	/*
1099 	 * This assert makes sure we don't hand out memory that is not
1100 	 * owned by this cache.
1101 	 */
1102 	assert(ret + thiscache->mt_size <= thiscache->mt_freelist +
1103 		thiscache->mt_span);
1104 
1105 	ret += OVERHEAD;
1106 
1107 	assert(((uintptr_t)ret & 7) == 0); /* are we 8 byte aligned */
1108 
1109 	if (reinit == 0 && (debugopt & MTDEBUGPATTERN))
1110 		if (verify_pattern(FREEPATTERN, ret, size))
1111 			abort();	/* reference after free */
1112 
1113 	if (debugopt & MTINITBUFFER)
1114 		copy_pattern(INITPATTERN, ret, size);
1115 	return ((void *)ret);
1116 }
1117 
1118 static void *
1119 morecore(size_t bytes)
1120 {
1121 	void * ret;
1122 
1123 	if (bytes > LONG_MAX) {
1124 		intptr_t wad;
1125 		/*
1126 		 * The request size is too big. We need to do this in
1127 		 * chunks. Sbrk only takes an int for an arg.
1128 		 */
1129 		if (bytes == ULONG_MAX)
1130 			return ((void *)-1);
1131 
1132 		ret = sbrk(0);
1133 		wad = LONG_MAX;
1134 		while (wad > 0) {
1135 			if (sbrk(wad) == (void *)-1) {
1136 				if (ret != sbrk(0))
1137 					(void) sbrk(-LONG_MAX);
1138 				return ((void *)-1);
1139 			}
1140 			bytes -= LONG_MAX;
1141 			wad = bytes;
1142 		}
1143 	} else
1144 		ret = sbrk(bytes);
1145 
1146 	return (ret);
1147 }
1148 
1149 
1150 static void *
1151 oversize(size_t size)
1152 {
1153 	caddr_t ret;
1154 	oversize_t *big;
1155 	int bucket;
1156 
1157 	/* make sure we will not overflow */
1158 	if (size > MAX_MTMALLOC) {
1159 		errno = ENOMEM;
1160 		return (NULL);
1161 	}
1162 
1163 	/*
1164 	 * Since we ensure every address we hand back is
1165 	 * MTMALLOC_MIN_ALIGN-byte aligned, ALIGNing size ensures that the
1166 	 * memory handed out is MTMALLOC_MIN_ALIGN-byte aligned at both ends.
1167 	 * This eases the implementation of MTDEBUGPATTERN and MTINITPATTERN,
1168 	 * particularly where coalescing occurs.
1169 	 */
1170 	size = ALIGN(size, MTMALLOC_MIN_ALIGN);
1171 
1172 	/*
1173 	 * The idea with the global lock is that we are sure to
1174 	 * block in the kernel anyway since given an oversize alloc
1175 	 * we are sure to have to call morecore();
1176 	 */
1177 	(void) mutex_lock(&oversize_lock);
1178 
1179 	if ((big = find_oversize(size)) != NULL) {
1180 		if (reinit == 0 && (debugopt & MTDEBUGPATTERN))
1181 			if (verify_pattern(FREEPATTERN, big->addr, size))
1182 				abort();	/* reference after free */
1183 	} else {
1184 		/* Get more 8-byte aligned memory from heap */
1185 		ret = morecore(size + OVSZ_HEADER_SIZE);
1186 		if (ret == (caddr_t)-1) {
1187 			(void) mutex_unlock(&oversize_lock);
1188 			errno = ENOMEM;
1189 			return (NULL);
1190 		}
1191 		big = oversize_header_alloc((uintptr_t)ret, size);
1192 	}
1193 	ret = big->addr;
1194 
1195 	/* Add big to the hash table at the head of the relevant bucket. */
1196 	bucket = HASH_OVERSIZE(ret);
1197 	big->hash_next = ovsz_hashtab[bucket];
1198 	ovsz_hashtab[bucket] = big;
1199 
1200 	if (debugopt & MTINITBUFFER)
1201 		copy_pattern(INITPATTERN, ret, size);
1202 
1203 	(void) mutex_unlock(&oversize_lock);
1204 	assert(((uintptr_t)ret & 7) == 0); /* are we 8 byte aligned */
1205 	return ((void *)ret);
1206 }
1207 
1208 static void
1209 insert_oversize(oversize_t *op, oversize_t *nx)
1210 {
1211 	oversize_t *sp;
1212 
1213 	/* locate correct insertion point in size-ordered list */
1214 	for (sp = oversize_list.next_bysize;
1215 	    sp != &oversize_list && (op->size > sp->size);
1216 	    sp = sp->next_bysize)
1217 		;
1218 
1219 	/* link into size-ordered list */
1220 	op->next_bysize = sp;
1221 	op->prev_bysize = sp->prev_bysize;
1222 	op->prev_bysize->next_bysize = op;
1223 	op->next_bysize->prev_bysize = op;
1224 
1225 	/*
1226 	 * link item into address-ordered list
1227 	 * (caller provides insertion point as an optimization)
1228 	 */
1229 	op->next_byaddr = nx;
1230 	op->prev_byaddr = nx->prev_byaddr;
1231 	op->prev_byaddr->next_byaddr = op;
1232 	op->next_byaddr->prev_byaddr = op;
1233 
1234 }
1235 
1236 static void
1237 unlink_oversize(oversize_t *lp)
1238 {
1239 	/* unlink from address list */
1240 	lp->prev_byaddr->next_byaddr = lp->next_byaddr;
1241 	lp->next_byaddr->prev_byaddr = lp->prev_byaddr;
1242 
1243 	/* unlink from size list */
1244 	lp->prev_bysize->next_bysize = lp->next_bysize;
1245 	lp->next_bysize->prev_bysize = lp->prev_bysize;
1246 }
1247 
1248 static void
1249 position_oversize_by_size(oversize_t *op)
1250 {
1251 	oversize_t *sp;
1252 
1253 	if (op->size > op->next_bysize->size ||
1254 	    op->size < op->prev_bysize->size) {
1255 
1256 		/* unlink from size list */
1257 		op->prev_bysize->next_bysize = op->next_bysize;
1258 		op->next_bysize->prev_bysize = op->prev_bysize;
1259 
1260 		/* locate correct insertion point in size-ordered list */
1261 		for (sp = oversize_list.next_bysize;
1262 		    sp != &oversize_list && (op->size > sp->size);
1263 		    sp = sp->next_bysize)
1264 			;
1265 
1266 		/* link into size-ordered list */
1267 		op->next_bysize = sp;
1268 		op->prev_bysize = sp->prev_bysize;
1269 		op->prev_bysize->next_bysize = op;
1270 		op->next_bysize->prev_bysize = op;
1271 	}
1272 }
1273 
1274 static void
1275 add_oversize(oversize_t *lp)
1276 {
1277 	int merge_flags = INSERT_ONLY;
1278 	oversize_t *nx;  	/* ptr to item right of insertion point */
1279 	oversize_t *pv;  	/* ptr to item left of insertion point */
1280 	uint_t size_lp, size_pv, size_nx;
1281 	uintptr_t endp_lp, endp_pv, endp_nx;
1282 
1283 	/*
1284 	 * Locate insertion point in address-ordered list
1285 	 */
1286 
1287 	for (nx = oversize_list.next_byaddr;
1288 	    nx != &oversize_list && (lp->addr > nx->addr);
1289 	    nx = nx->next_byaddr)
1290 		;
1291 
1292 	/*
1293 	 * Determine how to add chunk to oversize freelist
1294 	 */
1295 
1296 	size_lp = OVSZ_HEADER_SIZE + lp->size;
1297 	endp_lp = ALIGN((uintptr_t)lp + size_lp, MTMALLOC_MIN_ALIGN);
1298 	size_lp = endp_lp - (uintptr_t)lp;
1299 
1300 	pv = nx->prev_byaddr;
1301 
1302 	if (pv->size) {
1303 
1304 		size_pv = OVSZ_HEADER_SIZE + pv->size;
1305 		endp_pv = ALIGN((uintptr_t)pv + size_pv,
1306 		    MTMALLOC_MIN_ALIGN);
1307 		size_pv = endp_pv - (uintptr_t)pv;
1308 
1309 		/* Check for adjacency with left chunk */
1310 		if ((uintptr_t)lp == endp_pv)
1311 			merge_flags |= COALESCE_LEFT;
1312 	}
1313 
1314 	if (nx->size) {
1315 
1316 	    /* Check for adjacency with right chunk */
1317 	    if ((uintptr_t)nx == endp_lp) {
1318 		size_nx = OVSZ_HEADER_SIZE + nx->size;
1319 		endp_nx = ALIGN((uintptr_t)nx + size_nx,
1320 		    MTMALLOC_MIN_ALIGN);
1321 		size_nx = endp_nx - (uintptr_t)nx;
1322 		merge_flags |= COALESCE_RIGHT;
1323 	    }
1324 	}
1325 
1326 	/*
1327 	 * If MTDEBUGPATTERN==1, lp->addr will have been overwritten with
1328 	 * FREEPATTERN for lp->size bytes. If we can merge, the oversize
1329 	 * header(s) that will also become part of the memory available for
1330 	 * reallocation (ie lp and/or nx) must also be overwritten with
1331 	 * FREEPATTERN or we will SIGABRT when this memory is next reallocated.
1332 	 */
1333 	switch (merge_flags) {
1334 
1335 	case INSERT_ONLY:		/* Coalescing not possible */
1336 		insert_oversize(lp, nx);
1337 		break;
1338 	case COALESCE_LEFT:
1339 		pv->size += size_lp;
1340 		position_oversize_by_size(pv);
1341 		if (debugopt & MTDEBUGPATTERN)
1342 			copy_pattern(FREEPATTERN, lp, OVSZ_HEADER_SIZE);
1343 		break;
1344 	case COALESCE_RIGHT:
1345 		unlink_oversize(nx);
1346 		lp->size += size_nx;
1347 		insert_oversize(lp, pv->next_byaddr);
1348 		if (debugopt & MTDEBUGPATTERN)
1349 			copy_pattern(FREEPATTERN, nx, OVSZ_HEADER_SIZE);
1350 		break;
1351 	case COALESCE_WITH_BOTH_SIDES:	/* Merge (with right) to the left */
1352 		pv->size += size_lp + size_nx;
1353 		unlink_oversize(nx);
1354 		position_oversize_by_size(pv);
1355 		if (debugopt & MTDEBUGPATTERN) {
1356 			copy_pattern(FREEPATTERN, lp, OVSZ_HEADER_SIZE);
1357 			copy_pattern(FREEPATTERN, nx, OVSZ_HEADER_SIZE);
1358 		}
1359 		break;
1360 	}
1361 }
1362 
1363 /*
1364  * Find memory on our list that is at least size big. If we find a block that is
1365  * big enough, we break it up and return the associated oversize_t struct back
1366  * to the calling client. Any leftover piece of that block is returned to the
1367  * freelist.
1368  */
1369 static oversize_t *
1370 find_oversize(size_t size)
1371 {
1372 	oversize_t *wp = oversize_list.next_bysize;
1373 	while (wp != &oversize_list && size > wp->size)
1374 		wp = wp->next_bysize;
1375 
1376 	if (wp == &oversize_list) /* empty list or nothing big enough */
1377 		return (NULL);
1378 	/* breaking up a chunk of memory */
1379 	if ((long)((wp->size - (size + OVSZ_HEADER_SIZE + MTMALLOC_MIN_ALIGN)))
1380 	    > MAX_CACHED) {
1381 		caddr_t off;
1382 		oversize_t *np;
1383 		size_t osize;
1384 		off = (caddr_t)ALIGN(wp->addr + size,
1385 		    MTMALLOC_MIN_ALIGN);
1386 		osize = wp->size;
1387 		wp->size = (size_t)(off - wp->addr);
1388 		np = oversize_header_alloc((uintptr_t)off,
1389 		    osize - (wp->size + OVSZ_HEADER_SIZE));
1390 		if ((long)np->size < 0)
1391 			abort();
1392 		unlink_oversize(wp);
1393 		add_oversize(np);
1394 	} else {
1395 		unlink_oversize(wp);
1396 	}
1397 	return (wp);
1398 }
1399 
1400 static void
1401 copy_pattern(uint32_t pattern, void *buf_arg, size_t size)
1402 {
1403 	uint32_t *bufend = (uint32_t *)((char *)buf_arg + size);
1404 	uint32_t *buf = buf_arg;
1405 
1406 	while (buf < bufend - 3) {
1407 		buf[3] = buf[2] = buf[1] = buf[0] = pattern;
1408 		buf += 4;
1409 	}
1410 	while (buf < bufend)
1411 		*buf++ = pattern;
1412 }
1413 
1414 static void *
1415 verify_pattern(uint32_t pattern, void *buf_arg, size_t size)
1416 {
1417 	uint32_t *bufend = (uint32_t *)((char *)buf_arg + size);
1418 	uint32_t *buf;
1419 
1420 	for (buf = buf_arg; buf < bufend; buf++)
1421 		if (*buf != pattern)
1422 			return (buf);
1423 	return (NULL);
1424 }
1425 
1426 static void
1427 free_oversize(oversize_t *ovp)
1428 {
1429 	assert(((uintptr_t)ovp->addr & 7) == 0); /* are we 8 byte aligned */
1430 	assert(ovp->size > MAX_CACHED);
1431 
1432 	ovp->next_bysize = ovp->prev_bysize = NULL;
1433 	ovp->next_byaddr = ovp->prev_byaddr = NULL;
1434 	(void) mutex_lock(&oversize_lock);
1435 	add_oversize(ovp);
1436 	(void) mutex_unlock(&oversize_lock);
1437 }
1438 
1439 static oversize_t *
1440 oversize_header_alloc(uintptr_t mem, size_t size)
1441 {
1442 	oversize_t *ovsz_hdr;
1443 
1444 	assert(size > MAX_CACHED);
1445 
1446 	ovsz_hdr = (oversize_t *)mem;
1447 	ovsz_hdr->prev_bysize = NULL;
1448 	ovsz_hdr->next_bysize = NULL;
1449 	ovsz_hdr->prev_byaddr = NULL;
1450 	ovsz_hdr->next_byaddr = NULL;
1451 	ovsz_hdr->hash_next = NULL;
1452 	ovsz_hdr->size = size;
1453 	mem += OVSZ_SIZE;
1454 	*(uintptr_t *)mem = MTMALLOC_OVERSIZE_MAGIC;
1455 	mem += OVERHEAD;
1456 	assert(((uintptr_t)mem & 7) == 0); /* are we 8 byte aligned */
1457 	ovsz_hdr->addr = (caddr_t)mem;
1458 	return (ovsz_hdr);
1459 }
1460 
1461 static void
1462 malloc_prepare()
1463 {
1464 	percpu_t *cpuptr;
1465 	cache_head_t *cachehead;
1466 	cache_t *thiscache;
1467 
1468 	(void) mutex_lock(&oversize_lock);
1469 	for (cpuptr = &cpu_list[0]; cpuptr < &cpu_list[ncpus]; cpuptr++) {
1470 		(void) mutex_lock(&cpuptr->mt_parent_lock);
1471 		for (cachehead = &cpuptr->mt_caches[0];
1472 		    cachehead < &cpuptr->mt_caches[NUM_CACHES];
1473 		    cachehead++) {
1474 			for (thiscache = cachehead->mt_cache;
1475 			    thiscache != NULL;
1476 			    thiscache = thiscache->mt_next) {
1477 				(void) mutex_lock(
1478 				    &thiscache->mt_cache_lock);
1479 			}
1480 		}
1481 	}
1482 }
1483 
1484 static void
1485 malloc_release()
1486 {
1487 	percpu_t *cpuptr;
1488 	cache_head_t *cachehead;
1489 	cache_t *thiscache;
1490 
1491 	for (cpuptr = &cpu_list[ncpus - 1]; cpuptr >= &cpu_list[0]; cpuptr--) {
1492 		for (cachehead = &cpuptr->mt_caches[NUM_CACHES - 1];
1493 		    cachehead >= &cpuptr->mt_caches[0];
1494 		    cachehead--) {
1495 			for (thiscache = cachehead->mt_cache;
1496 			    thiscache != NULL;
1497 			    thiscache = thiscache->mt_next) {
1498 				(void) mutex_unlock(
1499 				    &thiscache->mt_cache_lock);
1500 			}
1501 		}
1502 		(void) mutex_unlock(&cpuptr->mt_parent_lock);
1503 	}
1504 	(void) mutex_unlock(&oversize_lock);
1505 }
1506 
1507 #pragma init(malloc_init)
1508 static void
1509 malloc_init(void)
1510 {
1511 	/*
1512 	 * This works in the init section for this library
1513 	 * because setup_caches() doesn't call anything in libc
1514 	 * that calls malloc().  If it did, disaster would ensue.
1515 	 *
1516 	 * For this to work properly, this library must be the first
1517 	 * one to have its init section called (after libc) by the
1518 	 * dynamic linker.  If some other library's init section
1519 	 * ran first and called malloc(), disaster would ensue.
1520 	 * Because this is an interposer library for malloc(), the
1521 	 * dynamic linker arranges for its init section to run first.
1522 	 */
1523 	(void) setup_caches();
1524 
1525 	(void) pthread_atfork(malloc_prepare, malloc_release, malloc_release);
1526 }
1527