xref: /illumos-gate/usr/src/lib/libmtmalloc/common/mtmalloc.c (revision 581cede61ac9c14d8d4ea452562a567189eead78)
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 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #include <mtmalloc.h>
28 #include "mtmalloc_impl.h"
29 #include <unistd.h>
30 #include <synch.h>
31 #include <thread.h>
32 #include <pthread.h>
33 #include <stdio.h>
34 #include <limits.h>
35 #include <errno.h>
36 #include <string.h>
37 #include <strings.h>
38 #include <sys/param.h>
39 #include <sys/sysmacros.h>
40 
41 /*
42  * To turn on the asserts just compile -DDEBUG
43  */
44 
45 #ifndef	DEBUG
46 #define	NDEBUG
47 #endif
48 
49 #include <assert.h>
50 
51 /*
52  * The MT hot malloc implementation contained herein is designed to be
53  * plug-compatible with the libc version of malloc. It is not intended
54  * to replace that implementation until we decide that it is ok to break
55  * customer apps (Solaris 3.0).
56  *
57  * For requests up to 2^^16, the allocator initializes itself into NCPUS
58  * worth of chains of caches. When a memory request is made, the calling thread
59  * is vectored into one of NCPUS worth of caches.  The LWP id gives us a cheap,
60  * contention-reducing index to use, eventually, this should be replaced with
61  * the actual CPU sequence number, when an interface to get it is available.
62  *
63  * Once the thread is vectored into one of the list of caches the real
64  * allocation of the memory begins. The size is determined to figure out which
65  * bucket the allocation should be satisfied from. The management of free
66  * buckets is done via a bitmask. A free bucket is represented by a 1. The
67  * first free bit represents the first free bucket. The position of the bit,
68  * represents the position of the bucket in the arena.
69  *
70  * When the memory from the arena is handed out, the address of the cache
71  * control structure is written in the word preceeding the returned memory.
72  * This cache control address is used during free() to mark the buffer free
73  * in the cache control structure.
74  *
75  * When all available memory in a cache has been depleted, a new chunk of memory
76  * is allocated via sbrk(). The new cache is allocated from this chunk of memory
77  * and initialized in the function create_cache(). New caches are installed at
78  * the front of a singly linked list of the same size memory pools. This helps
79  * to ensure that there will tend to be available memory in the beginning of the
80  * list.
81  *
82  * Long linked lists hurt performance. To decrease this effect, there is a
83  * tunable, requestsize, that bumps up the sbrk allocation size and thus
84  * increases the number of available blocks within an arena.  We also keep
85  * a "hint" for each cache list, which is the last cache in the list allocated
86  * from.  This lowers the cost of searching if there are a lot of fully
87  * allocated blocks at the front of the list.
88  *
89  * For requests greater than 2^^16 (oversize allocations), there are two pieces
90  * of overhead. There is the OVERHEAD used to hold the cache addr
91  * (&oversize_list), plus an oversize_t structure to further describe the block.
92  *
93  * The oversize list is kept as defragmented as possible by coalescing
94  * freed oversized allocations with adjacent neighbors.
95  *
96  * Addresses handed out are stored in a hash table, and are aligned on
97  * MTMALLOC_MIN_ALIGN-byte boundaries at both ends. Request sizes are rounded-up
98  * where necessary in order to achieve this. This eases the implementation of
99  * MTDEBUGPATTERN and MTINITPATTERN, particularly where coalescing occurs.
100  *
101  * A memalign allocation takes memalign header overhead.  There's two
102  * types of memalign headers distinguished by MTMALLOC_MEMALIGN_MAGIC
103  * and MTMALLOC_MEMALIGN_MIN_MAGIC.  When the size of memory taken to
104  * get to the aligned address from malloc'ed address is the minimum size
105  * OVERHEAD, we create a header taking only one OVERHEAD space with magic
106  * number MTMALLOC_MEMALIGN_MIN_MAGIC, and we know by subtracting OVERHEAD
107  * from memaligned address, we can get to the malloc'ed address. Otherwise,
108  * we create a memalign header taking two OVERHEAD space, one stores
109  * MTMALLOC_MEMALIGN_MAGIC magic number, the other one points back to the
110  * malloc'ed address.
111  */
112 
113 #if defined(__i386) || defined(__amd64)
114 #include <arpa/inet.h>	/* for htonl() */
115 #endif
116 
117 static void * morecore(size_t);
118 static void create_cache(cache_t *, size_t bufsize, uint_t hunks);
119 static void * malloc_internal(size_t, percpu_t *);
120 static void * oversize(size_t);
121 static oversize_t *find_oversize(size_t);
122 static void add_oversize(oversize_t *);
123 static void copy_pattern(uint32_t, void *, size_t);
124 static void * verify_pattern(uint32_t, void *, size_t);
125 static void reinit_cpu_list(void);
126 static void reinit_cache(cache_t *);
127 static void free_oversize(oversize_t *);
128 static oversize_t *oversize_header_alloc(uintptr_t, size_t);
129 
130 /*
131  * oversize hash table stuff
132  */
133 #define	NUM_BUCKETS	67	/* must be prime */
134 #define	HASH_OVERSIZE(caddr)	((uintptr_t)(caddr) % NUM_BUCKETS)
135 oversize_t *ovsz_hashtab[NUM_BUCKETS];
136 
137 #define	ALIGN(x, a)	((((uintptr_t)(x) + ((uintptr_t)(a) - 1)) \
138 			& ~((uintptr_t)(a) - 1)))
139 
140 /* need this to deal with little endianess of x86 */
141 #if defined(__i386) || defined(__amd64)
142 #define	FLIP_EM(x)	htonl((x))
143 #else
144 #define	FLIP_EM(x)	(x)
145 #endif
146 
147 #define	INSERT_ONLY			0
148 #define	COALESCE_LEFT			0x00000001
149 #define	COALESCE_RIGHT			0x00000002
150 #define	COALESCE_WITH_BOTH_SIDES	(COALESCE_LEFT | COALESCE_RIGHT)
151 
152 #define	OVERHEAD	8	/* size needed to write cache addr */
153 #define	HUNKSIZE	8192	/* just a multiplier */
154 
155 #define	MAX_CACHED_SHIFT	16	/* 64K is the max cached size */
156 #define	MAX_CACHED		(1 << MAX_CACHED_SHIFT)
157 #define	MIN_CACHED_SHIFT	4	/* smaller requests rounded up */
158 #define	MTMALLOC_MIN_ALIGN	8	/* min guaranteed alignment */
159 
160 /* maximum size before overflow */
161 #define	MAX_MTMALLOC	(SIZE_MAX - (SIZE_MAX % MTMALLOC_MIN_ALIGN) \
162 			- OVSZ_HEADER_SIZE)
163 
164 #define	NUM_CACHES	(MAX_CACHED_SHIFT - MIN_CACHED_SHIFT + 1)
165 #define	CACHELIST_SIZE	ALIGN(NUM_CACHES * sizeof (cache_head_t), \
166     CACHE_COHERENCY_UNIT)
167 
168 #define	MINSIZE		9	/* for requestsize, tunable */
169 #define	MAXSIZE		256	/* arbitrary, big enough, for requestsize */
170 
171 #define	FREEPATTERN	0xdeadbeef /* debug fill pattern for free buf */
172 #define	INITPATTERN	0xbaddcafe /* debug fill pattern for new buf */
173 
174 #define	misaligned(p)	((unsigned)(p) & (sizeof (int) - 1))
175 #define	IS_OVERSIZE(x, y)	(((x) < (y)) && (((x) > MAX_CACHED)? 1 : 0))
176 
177 static long requestsize = MINSIZE; /* 9 pages per cache; tunable; 9 is min */
178 
179 static uint_t cpu_mask;
180 static curcpu_func curcpu;
181 
182 static int32_t debugopt;
183 static int32_t reinit;
184 
185 static percpu_t *cpu_list;
186 static oversize_t oversize_list;
187 static mutex_t oversize_lock = DEFAULTMUTEX;
188 
189 static int ncpus = 0;
190 
191 #define	MTMALLOC_OVERSIZE_MAGIC		((uintptr_t)&oversize_list)
192 #define	MTMALLOC_MEMALIGN_MAGIC		((uintptr_t)&oversize_list + 1)
193 #define	MTMALLOC_MEMALIGN_MIN_MAGIC	((uintptr_t)&oversize_list + 2)
194 
195 /*
196  * We require allocations handed out to be aligned on MTMALLOC_MIN_ALIGN-byte
197  * boundaries. We round up sizeof (oversize_t) (when necessary) to ensure that
198  * this is achieved.
199  */
200 #define	OVSZ_SIZE		(ALIGN(sizeof (oversize_t), MTMALLOC_MIN_ALIGN))
201 #define	OVSZ_HEADER_SIZE	(OVSZ_SIZE + OVERHEAD)
202 
203 /*
204  * memalign header takes 2 OVERHEAD space.  One for memalign magic, and the
205  * other one points back to the start address of originally allocated space.
206  */
207 #define	MEMALIGN_HEADER_SIZE	2 * OVERHEAD
208 #define	MEMALIGN_HEADER_ALLOC(x, shift, malloc_addr)\
209 	if (shift == OVERHEAD)\
210 		*((uintptr_t *)((caddr_t)x - OVERHEAD)) = \
211 			MTMALLOC_MEMALIGN_MIN_MAGIC; \
212 	else {\
213 		*((uintptr_t *)((caddr_t)x - OVERHEAD)) = \
214 			MTMALLOC_MEMALIGN_MAGIC; \
215 		*((uintptr_t *)((caddr_t)x - 2 * OVERHEAD)) = \
216 			(uintptr_t)malloc_addr; \
217 	}
218 
219 /*
220  * Add big to the oversize hash table at the head of the relevant bucket.
221  */
222 static void
223 insert_hash(oversize_t *big)
224 {
225 	caddr_t ret = big->addr;
226 	int bucket = HASH_OVERSIZE(ret);
227 
228 	assert(MUTEX_HELD(&oversize_lock));
229 	big->hash_next = ovsz_hashtab[bucket];
230 	ovsz_hashtab[bucket] = big;
231 }
232 
233 void *
234 malloc(size_t bytes)
235 {
236 	percpu_t *list_rotor;
237 	uint_t	list_index;
238 
239 	if (bytes > MAX_CACHED)
240 		return (oversize(bytes));
241 
242 	list_index = (curcpu() & cpu_mask);
243 
244 	list_rotor = &cpu_list[list_index];
245 
246 	return (malloc_internal(bytes, list_rotor));
247 }
248 
249 void *
250 realloc(void * ptr, size_t bytes)
251 {
252 	void *new, *data_ptr;
253 	cache_t *cacheptr;
254 	caddr_t mem;
255 	size_t shift = 0;
256 
257 	if (ptr == NULL)
258 		return (malloc(bytes));
259 
260 	if (bytes == 0) {
261 		free(ptr);
262 		return (NULL);
263 	}
264 
265 	data_ptr = ptr;
266 	mem = (caddr_t)ptr - OVERHEAD;
267 
268 	/*
269 	 * Optimization possibility :
270 	 *	p = malloc(64);
271 	 *	q = realloc(p, 64);
272 	 * q can be same as p.
273 	 * Apply this optimization for the normal
274 	 * sized caches for now.
275 	 */
276 	if (*(uintptr_t *)mem < MTMALLOC_OVERSIZE_MAGIC ||
277 	    *(uintptr_t *)mem > MTMALLOC_MEMALIGN_MIN_MAGIC) {
278 		cacheptr = (cache_t *)*(uintptr_t *)mem;
279 		if (bytes <= (cacheptr->mt_size - OVERHEAD))
280 			return (ptr);
281 	}
282 
283 	new = malloc(bytes);
284 
285 	if (new == NULL)
286 		return (NULL);
287 
288 	/*
289 	 * If new == ptr, ptr has previously been freed. Passing a freed pointer
290 	 * to realloc() is not allowed - unless the caller specifically states
291 	 * otherwise, in which case we must avoid freeing ptr (ie new) before we
292 	 * return new. There is (obviously) no requirement to memcpy() ptr to
293 	 * new before we return.
294 	 */
295 	if (new == ptr) {
296 		if (!(debugopt & MTDOUBLEFREE))
297 			abort();
298 		return (new);
299 	}
300 
301 	if (*(uintptr_t *)mem == MTMALLOC_MEMALIGN_MAGIC) {
302 		mem -= OVERHEAD;
303 		ptr = (void *)*(uintptr_t *)mem;
304 		mem = (caddr_t)ptr - OVERHEAD;
305 		shift = (size_t)((uintptr_t)data_ptr - (uintptr_t)ptr);
306 	} else if (*(uintptr_t *)mem == MTMALLOC_MEMALIGN_MIN_MAGIC) {
307 		ptr = (void *) mem;
308 		mem -= OVERHEAD;
309 		shift = OVERHEAD;
310 	}
311 
312 	if (*(uintptr_t *)mem == MTMALLOC_OVERSIZE_MAGIC) {
313 		oversize_t *old;
314 
315 		old = (oversize_t *)(mem - OVSZ_SIZE);
316 		(void) memcpy(new, data_ptr, MIN(bytes, old->size - shift));
317 		free(ptr);
318 		return (new);
319 	}
320 
321 	cacheptr = (cache_t *)*(uintptr_t *)mem;
322 
323 	(void) memcpy(new, data_ptr,
324 	    MIN(cacheptr->mt_size - OVERHEAD - shift, bytes));
325 	free(ptr);
326 
327 	return (new);
328 }
329 
330 void *
331 calloc(size_t nelem, size_t bytes)
332 {
333 	void * ptr;
334 	size_t size = nelem * bytes;
335 
336 	ptr = malloc(size);
337 	if (ptr == NULL)
338 		return (NULL);
339 	(void) memset(ptr, 0, size);
340 
341 	return (ptr);
342 }
343 
344 void
345 free(void * ptr)
346 {
347 	cache_t *cacheptr;
348 	caddr_t mem;
349 	int32_t i;
350 	caddr_t freeblocks;
351 	uintptr_t offset;
352 	uchar_t mask;
353 	int32_t which_bit, num_bytes;
354 
355 	if (ptr == NULL)
356 		return;
357 
358 	mem = (caddr_t)ptr - OVERHEAD;
359 
360 	if (*(uintptr_t *)mem == MTMALLOC_MEMALIGN_MAGIC) {
361 		mem -= OVERHEAD;
362 		ptr = (void *)*(uintptr_t *)mem;
363 		mem = (caddr_t)ptr - OVERHEAD;
364 	} else if (*(uintptr_t *)mem == MTMALLOC_MEMALIGN_MIN_MAGIC) {
365 		ptr = (void *) mem;
366 		mem -= OVERHEAD;
367 	}
368 
369 	if (*(uintptr_t *)mem == MTMALLOC_OVERSIZE_MAGIC) {
370 		oversize_t *big, **opp;
371 		int bucket;
372 
373 		big = (oversize_t *)(mem - OVSZ_SIZE);
374 		(void) mutex_lock(&oversize_lock);
375 
376 		bucket = HASH_OVERSIZE(big->addr);
377 		for (opp = &ovsz_hashtab[bucket]; *opp != NULL;
378 		    opp = &(*opp)->hash_next)
379 			if (*opp == big)
380 				break;
381 
382 		if (*opp == NULL) {
383 			if (!(debugopt & MTDOUBLEFREE))
384 				abort();
385 			(void) mutex_unlock(&oversize_lock);
386 			return;
387 		}
388 
389 		*opp = big->hash_next;	/* remove big from the hash table */
390 		big->hash_next = NULL;
391 
392 		if (debugopt & MTDEBUGPATTERN)
393 			copy_pattern(FREEPATTERN, ptr, big->size);
394 		add_oversize(big);
395 		(void) mutex_unlock(&oversize_lock);
396 		return;
397 	}
398 
399 	cacheptr = (cache_t *)*(uintptr_t *)mem;
400 	freeblocks = cacheptr->mt_freelist;
401 
402 	/*
403 	 * This is the distance measured in bits into the arena.
404 	 * The value of offset is in bytes but there is a 1-1 correlation
405 	 * between distance into the arena and distance into the
406 	 * freelist bitmask.
407 	 */
408 	offset = mem - cacheptr->mt_arena;
409 
410 	/*
411 	 * i is total number of bits to offset into freelist bitmask.
412 	 */
413 
414 	i = offset / cacheptr->mt_size;
415 
416 	num_bytes = i >> 3;
417 
418 	/*
419 	 * which_bit is the bit offset into the byte in the freelist.
420 	 * if our freelist bitmask looks like 0xf3 and we are freeing
421 	 * block 5 (ie: the 6th block) our mask will be 0xf7 after
422 	 * the free. Things go left to right that's why the mask is 0x80
423 	 * and not 0x01.
424 	 */
425 	which_bit = i - (num_bytes << 3);
426 
427 	mask = 0x80 >> which_bit;
428 
429 	freeblocks += num_bytes;
430 
431 	if (debugopt & MTDEBUGPATTERN)
432 		copy_pattern(FREEPATTERN, ptr, cacheptr->mt_size - OVERHEAD);
433 
434 	(void) mutex_lock(&cacheptr->mt_cache_lock);
435 
436 	if (*freeblocks & mask) {
437 		if (!(debugopt & MTDOUBLEFREE))
438 			abort();
439 	} else {
440 		*freeblocks |= mask;
441 		cacheptr->mt_nfree++;
442 	}
443 
444 	(void) mutex_unlock(&cacheptr->mt_cache_lock);
445 }
446 
447 void *
448 memalign(size_t alignment, size_t size)
449 {
450 	size_t alloc_size;
451 	uintptr_t offset;
452 	void *alloc_buf;
453 	void *ret_buf;
454 
455 	if (size == 0 || alignment == 0 || misaligned(alignment) ||
456 	    (alignment & (alignment - 1)) != 0) {
457 		errno = EINVAL;
458 		return (NULL);
459 	}
460 
461 	/* <= MTMALLOC_MIN_ALIGN, malloc can provide directly */
462 	if (alignment <= MTMALLOC_MIN_ALIGN)
463 		return (malloc(size));
464 
465 	alloc_size = size + alignment - MTMALLOC_MIN_ALIGN;
466 
467 	if (alloc_size < size) { /* overflow */
468 		errno = ENOMEM;
469 		return (NULL);
470 	}
471 
472 	alloc_buf = malloc(alloc_size);
473 
474 	if (alloc_buf == NULL)
475 		/* malloc sets errno */
476 		return (NULL);
477 
478 	/*
479 	 * If alloc_size > MAX_CACHED, malloc() will have returned a multiple of
480 	 * MTMALLOC_MIN_ALIGN, having rounded-up alloc_size if necessary. Since
481 	 * we will use alloc_size to return the excess fragments to the free
482 	 * list, we also round-up alloc_size if necessary.
483 	 */
484 	if ((alloc_size > MAX_CACHED) &&
485 	    (alloc_size & (MTMALLOC_MIN_ALIGN - 1)))
486 		alloc_size = ALIGN(alloc_size, MTMALLOC_MIN_ALIGN);
487 
488 	if ((offset = (uintptr_t)alloc_buf & (alignment - 1)) == 0) {
489 		/* aligned correctly */
490 
491 		size_t frag_size = alloc_size -
492 		    (size + MTMALLOC_MIN_ALIGN + OVSZ_HEADER_SIZE);
493 
494 		/*
495 		 * If the leftover piece of the memory > MAX_CACHED,
496 		 * split off the piece and return it back to the freelist.
497 		 */
498 		if (IS_OVERSIZE(frag_size, alloc_size)) {
499 			oversize_t *orig, *tail;
500 			uintptr_t taddr;
501 			size_t data_size;
502 			taddr = ALIGN((uintptr_t)alloc_buf + size,
503 			    MTMALLOC_MIN_ALIGN);
504 			data_size = taddr - (uintptr_t)alloc_buf;
505 			orig = (oversize_t *)((uintptr_t)alloc_buf -
506 			    OVSZ_HEADER_SIZE);
507 			frag_size = orig->size - data_size -
508 			    OVSZ_HEADER_SIZE;
509 			orig->size = data_size;
510 			tail = oversize_header_alloc(taddr, frag_size);
511 			free_oversize(tail);
512 		}
513 		ret_buf = alloc_buf;
514 	} else {
515 		uchar_t	oversize_bits = 0;
516 		size_t	head_sz, data_sz, tail_sz;
517 		uintptr_t ret_addr, taddr, shift, tshift;
518 		oversize_t *orig, *tail, *big;
519 		size_t tsize;
520 
521 		/* needs to be aligned */
522 		shift = alignment - offset;
523 
524 		assert(shift >= MTMALLOC_MIN_ALIGN);
525 
526 		ret_addr = ((uintptr_t)alloc_buf + shift);
527 		ret_buf = (void *)ret_addr;
528 
529 		if (alloc_size <= MAX_CACHED) {
530 			MEMALIGN_HEADER_ALLOC(ret_addr, shift, alloc_buf);
531 			return (ret_buf);
532 		}
533 
534 		/*
535 		 * Only check for the fragments when the memory is allocted
536 		 * from oversize_list.  Split off a fragment and return it
537 		 * to the oversize freelist when it's > MAX_CACHED.
538 		 */
539 
540 		head_sz = shift - MAX(MEMALIGN_HEADER_SIZE, OVSZ_HEADER_SIZE);
541 
542 		tail_sz = alloc_size -
543 		    (shift + size + MTMALLOC_MIN_ALIGN + OVSZ_HEADER_SIZE);
544 
545 		oversize_bits |= IS_OVERSIZE(head_sz, alloc_size) |
546 		    IS_OVERSIZE(size, alloc_size) << DATA_SHIFT |
547 		    IS_OVERSIZE(tail_sz, alloc_size) << TAIL_SHIFT;
548 
549 		switch (oversize_bits) {
550 			case NONE_OVERSIZE:
551 			case DATA_OVERSIZE:
552 				MEMALIGN_HEADER_ALLOC(ret_addr, shift,
553 				    alloc_buf);
554 				break;
555 			case HEAD_OVERSIZE:
556 				/*
557 				 * If we can extend data > MAX_CACHED and have
558 				 * head still > MAX_CACHED, we split head-end
559 				 * as the case of head-end and data oversized,
560 				 * otherwise just create memalign header.
561 				 */
562 				tsize = (shift + size) - (MAX_CACHED + 8 +
563 				    MTMALLOC_MIN_ALIGN + OVSZ_HEADER_SIZE);
564 
565 				if (!IS_OVERSIZE(tsize, alloc_size)) {
566 					MEMALIGN_HEADER_ALLOC(ret_addr, shift,
567 					    alloc_buf);
568 					break;
569 				} else {
570 					tsize += OVSZ_HEADER_SIZE;
571 					taddr = ALIGN((uintptr_t)alloc_buf +
572 					    tsize, MTMALLOC_MIN_ALIGN);
573 					tshift = ret_addr - taddr;
574 					MEMALIGN_HEADER_ALLOC(ret_addr, tshift,
575 					    taddr);
576 					ret_addr = taddr;
577 					shift = ret_addr - (uintptr_t)alloc_buf;
578 				}
579 				/* FALLTHROUGH */
580 			case HEAD_AND_DATA_OVERSIZE:
581 				/*
582 				 * Split off the head fragment and
583 				 * return it back to oversize freelist.
584 				 * Create oversize header for the piece
585 				 * of (data + tail fragment).
586 				 */
587 				orig = (oversize_t *)((uintptr_t)alloc_buf -
588 				    OVSZ_HEADER_SIZE);
589 				big = oversize_header_alloc(ret_addr -
590 				    OVSZ_HEADER_SIZE, (orig->size - shift));
591 				(void) mutex_lock(&oversize_lock);
592 				insert_hash(big);
593 				(void) mutex_unlock(&oversize_lock);
594 				orig->size = shift - OVSZ_HEADER_SIZE;
595 
596 				/* free up the head fragment */
597 				free_oversize(orig);
598 				break;
599 			case TAIL_OVERSIZE:
600 				/*
601 				 * If we can extend data > MAX_CACHED and have
602 				 * tail-end still > MAX_CACHED, we split tail
603 				 * end, otherwise just create memalign header.
604 				 */
605 				orig = (oversize_t *)((uintptr_t)alloc_buf -
606 				    OVSZ_HEADER_SIZE);
607 				tsize =  orig->size - (MAX_CACHED + 8 +
608 				    shift + OVSZ_HEADER_SIZE +
609 				    MTMALLOC_MIN_ALIGN);
610 				if (!IS_OVERSIZE(tsize, alloc_size)) {
611 					MEMALIGN_HEADER_ALLOC(ret_addr, shift,
612 					    alloc_buf);
613 					break;
614 				} else {
615 					size = MAX_CACHED + 8;
616 				}
617 				/* FALLTHROUGH */
618 			case DATA_AND_TAIL_OVERSIZE:
619 				/*
620 				 * Split off the tail fragment and
621 				 * return it back to oversize freelist.
622 				 * Create memalign header and adjust
623 				 * the size for the piece of
624 				 * (head fragment + data).
625 				 */
626 				taddr = ALIGN(ret_addr + size,
627 				    MTMALLOC_MIN_ALIGN);
628 				data_sz = (size_t)(taddr -
629 				    (uintptr_t)alloc_buf);
630 				orig = (oversize_t *)((uintptr_t)alloc_buf -
631 				    OVSZ_HEADER_SIZE);
632 				tsize = orig->size - data_sz;
633 				orig->size = data_sz;
634 				MEMALIGN_HEADER_ALLOC(ret_buf, shift,
635 				    alloc_buf);
636 				tsize -= OVSZ_HEADER_SIZE;
637 				tail = oversize_header_alloc(taddr,  tsize);
638 				free_oversize(tail);
639 				break;
640 			case HEAD_AND_TAIL_OVERSIZE:
641 				/*
642 				 * Split off the head fragment.
643 				 * We try to free up tail-end when we can
644 				 * extend data size to (MAX_CACHED + 8)
645 				 * and remain tail-end oversized.
646 				 * The bottom line is all split pieces
647 				 * should be oversize in size.
648 				 */
649 				orig = (oversize_t *)((uintptr_t)alloc_buf -
650 				    OVSZ_HEADER_SIZE);
651 				tsize =  orig->size - (MAX_CACHED + 8 +
652 				    OVSZ_HEADER_SIZE + shift +
653 				    MTMALLOC_MIN_ALIGN);
654 
655 				if (!IS_OVERSIZE(tsize, alloc_size)) {
656 					/*
657 					 * If the chunk is not big enough
658 					 * to make both data and tail oversize
659 					 * we just keep them as one piece.
660 					 */
661 					big = oversize_header_alloc(ret_addr -
662 					    OVSZ_HEADER_SIZE,
663 					    orig->size - shift);
664 					(void) mutex_lock(&oversize_lock);
665 					insert_hash(big);
666 					(void) mutex_unlock(&oversize_lock);
667 					orig->size = shift - OVSZ_HEADER_SIZE;
668 					free_oversize(orig);
669 					break;
670 				} else {
671 					/*
672 					 * extend data size > MAX_CACHED
673 					 * and handle it as head, data, tail
674 					 * are all oversized.
675 					 */
676 					size = MAX_CACHED + 8;
677 				}
678 				/* FALLTHROUGH */
679 			case ALL_OVERSIZE:
680 				/*
681 				 * split off the head and tail fragments,
682 				 * return them back to the oversize freelist.
683 				 * Alloc oversize header for data seg.
684 				 */
685 				orig = (oversize_t *)((uintptr_t)alloc_buf -
686 				    OVSZ_HEADER_SIZE);
687 				tsize = orig->size;
688 				orig->size = shift - OVSZ_HEADER_SIZE;
689 				free_oversize(orig);
690 
691 				taddr = ALIGN(ret_addr + size,
692 				    MTMALLOC_MIN_ALIGN);
693 				data_sz = taddr - ret_addr;
694 				assert(tsize > (shift + data_sz +
695 				    OVSZ_HEADER_SIZE));
696 				tail_sz = tsize -
697 				    (shift + data_sz + OVSZ_HEADER_SIZE);
698 
699 				/* create oversize header for data seg */
700 				big = oversize_header_alloc(ret_addr -
701 				    OVSZ_HEADER_SIZE, data_sz);
702 				(void) mutex_lock(&oversize_lock);
703 				insert_hash(big);
704 				(void) mutex_unlock(&oversize_lock);
705 
706 				/* create oversize header for tail fragment */
707 				tail = oversize_header_alloc(taddr, tail_sz);
708 				free_oversize(tail);
709 				break;
710 			default:
711 				/* should not reach here */
712 				assert(0);
713 		}
714 	}
715 	return (ret_buf);
716 }
717 
718 
719 void *
720 valloc(size_t size)
721 {
722 	static unsigned pagesize;
723 
724 	if (size == 0)
725 		return (NULL);
726 
727 	if (!pagesize)
728 		pagesize = sysconf(_SC_PAGESIZE);
729 
730 	return (memalign(pagesize, size));
731 }
732 
733 void
734 mallocctl(int cmd, long value)
735 {
736 	switch (cmd) {
737 
738 	case MTDEBUGPATTERN:
739 		/*
740 		 * Reinitialize free blocks in case malloc() is called prior
741 		 * to mallocctl().
742 		 */
743 		if (value && !(debugopt & cmd)) {
744 			reinit++;
745 			debugopt |= cmd;
746 			reinit_cpu_list();
747 		}
748 		/*FALLTHRU*/
749 	case MTDOUBLEFREE:
750 	case MTINITBUFFER:
751 		if (value)
752 			debugopt |= cmd;
753 		else
754 			debugopt &= ~cmd;
755 		break;
756 	case MTCHUNKSIZE:
757 		if (value >= MINSIZE && value <= MAXSIZE)
758 			requestsize = value;
759 		break;
760 	default:
761 		break;
762 	}
763 }
764 
765 /*
766  * Initialization function, called from the init section of the library.
767  * No locking is required here because we are single-threaded during
768  * library initialization.
769  */
770 static void
771 setup_caches(void)
772 {
773 	uintptr_t oldbrk;
774 	uintptr_t newbrk;
775 
776 	size_t cache_space_needed;
777 	size_t padding;
778 
779 	curcpu_func new_curcpu;
780 	uint_t new_cpu_mask;
781 	percpu_t *new_cpu_list;
782 
783 	uint_t i, j;
784 	uintptr_t list_addr;
785 
786 	/*
787 	 * Get a decent "current cpu identifier", to be used to reduce
788 	 * contention.  Eventually, this should be replaced by an interface
789 	 * to get the actual CPU sequence number in libthread/liblwp.
790 	 */
791 	new_curcpu = (curcpu_func)thr_self;
792 	if ((ncpus = 2 * sysconf(_SC_NPROCESSORS_CONF)) <= 0)
793 		ncpus = 4; /* decent default value */
794 
795 	/* round ncpus up to a power of 2 */
796 	while (ncpus & (ncpus - 1))
797 		ncpus++;
798 
799 	new_cpu_mask = ncpus - 1;	/* create the cpu mask */
800 
801 	/*
802 	 * We now do some magic with the brk.  What we want to get in the
803 	 * end is a bunch of well-aligned stuff in a big initial allocation.
804 	 * Along the way, we do sanity checks to make sure no one else has
805 	 * touched the brk (which shouldn't happen, but it's always good to
806 	 * check)
807 	 *
808 	 * First, make sure sbrk is sane, and store the current brk in oldbrk.
809 	 */
810 	oldbrk = (uintptr_t)sbrk(0);
811 	if ((void *)oldbrk == (void *)-1)
812 		abort();	/* sbrk is broken -- we're doomed. */
813 
814 	/*
815 	 * Now, align the brk to a multiple of CACHE_COHERENCY_UNIT, so that
816 	 * the percpu structures and cache lists will be properly aligned.
817 	 *
818 	 *   2.  All hunks will be page-aligned, assuming HUNKSIZE >= PAGESIZE,
819 	 *	so they can be paged out individually.
820 	 */
821 	newbrk = ALIGN(oldbrk, CACHE_COHERENCY_UNIT);
822 	if (newbrk != oldbrk && (uintptr_t)sbrk(newbrk - oldbrk) != oldbrk)
823 		abort();	/* sbrk is broken -- we're doomed. */
824 
825 	/*
826 	 * For each cpu, there is one percpu_t and a list of caches
827 	 */
828 	cache_space_needed = ncpus * (sizeof (percpu_t) + CACHELIST_SIZE);
829 
830 	new_cpu_list = (percpu_t *)sbrk(cache_space_needed);
831 
832 	if (new_cpu_list == (percpu_t *)-1 ||
833 	    (uintptr_t)new_cpu_list != newbrk)
834 		abort();	/* sbrk is broken -- we're doomed. */
835 
836 	/*
837 	 * Finally, align the brk to HUNKSIZE so that all hunks are
838 	 * page-aligned, to avoid edge-effects.
839 	 */
840 
841 	newbrk = (uintptr_t)new_cpu_list + cache_space_needed;
842 
843 	padding = ALIGN(newbrk, HUNKSIZE) - newbrk;
844 
845 	if (padding > 0 && (uintptr_t)sbrk(padding) != newbrk)
846 		abort();	/* sbrk is broken -- we're doomed. */
847 
848 	list_addr = ((uintptr_t)new_cpu_list + (sizeof (percpu_t) * ncpus));
849 
850 	/* initialize the percpu list */
851 	for (i = 0; i < ncpus; i++) {
852 		new_cpu_list[i].mt_caches = (cache_head_t *)list_addr;
853 		for (j = 0; j < NUM_CACHES; j++) {
854 			new_cpu_list[i].mt_caches[j].mt_cache = NULL;
855 			new_cpu_list[i].mt_caches[j].mt_hint = NULL;
856 		}
857 
858 		(void) mutex_init(&new_cpu_list[i].mt_parent_lock,
859 		    USYNC_THREAD, NULL);
860 
861 		/* get the correct cache list alignment */
862 		list_addr += CACHELIST_SIZE;
863 	}
864 
865 	/*
866 	 * Initialize oversize listhead
867 	 */
868 	oversize_list.next_bysize = &oversize_list;
869 	oversize_list.prev_bysize = &oversize_list;
870 	oversize_list.next_byaddr = &oversize_list;
871 	oversize_list.prev_byaddr = &oversize_list;
872 	oversize_list.addr = NULL;
873 	oversize_list.size = 0;		/* sentinal */
874 
875 	/*
876 	 * Now install the global variables.
877 	 */
878 	curcpu = new_curcpu;
879 	cpu_mask = new_cpu_mask;
880 	cpu_list = new_cpu_list;
881 }
882 
883 static void
884 create_cache(cache_t *cp, size_t size, uint_t chunksize)
885 {
886 	long nblocks;
887 
888 	(void) mutex_init(&cp->mt_cache_lock, USYNC_THREAD, NULL);
889 	cp->mt_size = size;
890 	cp->mt_freelist = ((caddr_t)cp + sizeof (cache_t));
891 	cp->mt_span = chunksize * HUNKSIZE - sizeof (cache_t);
892 	cp->mt_hunks = chunksize;
893 	/*
894 	 * rough calculation. We will need to adjust later.
895 	 */
896 	nblocks = cp->mt_span / cp->mt_size;
897 	nblocks >>= 3;
898 	if (nblocks == 0) { /* less than 8 free blocks in this pool */
899 		int32_t numblocks = 0;
900 		long i = cp->mt_span;
901 		size_t sub = cp->mt_size;
902 		uchar_t mask = 0;
903 
904 		while (i > sub) {
905 			numblocks++;
906 			i -= sub;
907 		}
908 		nblocks = numblocks;
909 		cp->mt_arena = (caddr_t)ALIGN(cp->mt_freelist + 8, 8);
910 		cp->mt_nfree = numblocks;
911 		while (numblocks--) {
912 			mask |= 0x80 >> numblocks;
913 		}
914 		*(cp->mt_freelist) = mask;
915 	} else {
916 		cp->mt_arena = (caddr_t)ALIGN((caddr_t)cp->mt_freelist +
917 		    nblocks, 32);
918 		/* recompute nblocks */
919 		nblocks = (uintptr_t)((caddr_t)cp->mt_freelist +
920 		    cp->mt_span - cp->mt_arena) / cp->mt_size;
921 		cp->mt_nfree = ((nblocks >> 3) << 3);
922 		/* Set everything to free */
923 		(void) memset(cp->mt_freelist, 0xff, nblocks >> 3);
924 	}
925 
926 	if (debugopt & MTDEBUGPATTERN)
927 		copy_pattern(FREEPATTERN, cp->mt_arena, cp->mt_size * nblocks);
928 
929 	cp->mt_next = NULL;
930 }
931 
932 static void
933 reinit_cpu_list(void)
934 {
935 	oversize_t *wp = oversize_list.next_bysize;
936 	percpu_t *cpuptr;
937 	cache_t *thiscache;
938 	cache_head_t *cachehead;
939 
940 	/* Reinitialize free oversize blocks. */
941 	(void) mutex_lock(&oversize_lock);
942 	if (debugopt & MTDEBUGPATTERN)
943 		for (; wp != &oversize_list; wp = wp->next_bysize)
944 			copy_pattern(FREEPATTERN, wp->addr, wp->size);
945 	(void) mutex_unlock(&oversize_lock);
946 
947 	/* Reinitialize free blocks. */
948 	for (cpuptr = &cpu_list[0]; cpuptr < &cpu_list[ncpus]; cpuptr++) {
949 		(void) mutex_lock(&cpuptr->mt_parent_lock);
950 		for (cachehead = &cpuptr->mt_caches[0]; cachehead <
951 		    &cpuptr->mt_caches[NUM_CACHES]; cachehead++) {
952 			for (thiscache = cachehead->mt_cache; thiscache != NULL;
953 			    thiscache = thiscache->mt_next) {
954 				(void) mutex_lock(&thiscache->mt_cache_lock);
955 				if (thiscache->mt_nfree == 0) {
956 					(void) mutex_unlock(
957 					    &thiscache->mt_cache_lock);
958 					continue;
959 				}
960 				if (thiscache != NULL)
961 					reinit_cache(thiscache);
962 				(void) mutex_unlock(&thiscache->mt_cache_lock);
963 			}
964 		}
965 		(void) mutex_unlock(&cpuptr->mt_parent_lock);
966 	}
967 	reinit = 0;
968 }
969 
970 static void
971 reinit_cache(cache_t *thiscache)
972 {
973 	uint32_t *freeblocks; /* not a uintptr_t on purpose */
974 	int32_t i, n;
975 	caddr_t ret;
976 
977 	freeblocks = (uint32_t *)thiscache->mt_freelist;
978 	while (freeblocks < (uint32_t *)thiscache->mt_arena) {
979 		if (*freeblocks & 0xffffffff) {
980 			for (i = 0; i < 32; i++) {
981 				if (FLIP_EM(*freeblocks) & (0x80000000 >> i)) {
982 					n = (uintptr_t)(((freeblocks -
983 					    (uint32_t *)thiscache->mt_freelist)
984 					    << 5) + i) * thiscache->mt_size;
985 					ret = thiscache->mt_arena + n;
986 					ret += OVERHEAD;
987 					copy_pattern(FREEPATTERN, ret,
988 					    thiscache->mt_size);
989 				}
990 			}
991 		}
992 		freeblocks++;
993 	}
994 }
995 
996 static void *
997 malloc_internal(size_t size, percpu_t *cpuptr)
998 {
999 	cache_head_t *cachehead;
1000 	cache_t *thiscache, *hintcache;
1001 	int32_t i, n, logsz, bucket;
1002 	uint32_t index;
1003 	uint32_t *freeblocks; /* not a uintptr_t on purpose */
1004 	caddr_t ret;
1005 
1006 	logsz = MIN_CACHED_SHIFT;
1007 
1008 	while (size > (1 << logsz))
1009 		logsz++;
1010 
1011 	bucket = logsz - MIN_CACHED_SHIFT;
1012 
1013 	(void) mutex_lock(&cpuptr->mt_parent_lock);
1014 
1015 	/*
1016 	 * Find a cache of the appropriate size with free buffers.
1017 	 *
1018 	 * We don't need to lock each cache as we check their mt_nfree count,
1019 	 * since:
1020 	 *	1.  We are only looking for caches with mt_nfree > 0.  If a
1021 	 *	   free happens during our search, it will increment mt_nfree,
1022 	 *	   which will not effect the test.
1023 	 *	2.  Allocations can decrement mt_nfree, but they can't happen
1024 	 *	   as long as we hold mt_parent_lock.
1025 	 */
1026 
1027 	cachehead = &cpuptr->mt_caches[bucket];
1028 
1029 	/* Search through the list, starting at the mt_hint */
1030 	thiscache = cachehead->mt_hint;
1031 
1032 	while (thiscache != NULL && thiscache->mt_nfree == 0)
1033 		thiscache = thiscache->mt_next;
1034 
1035 	if (thiscache == NULL) {
1036 		/* wrap around -- search up to the hint */
1037 		thiscache = cachehead->mt_cache;
1038 		hintcache = cachehead->mt_hint;
1039 
1040 		while (thiscache != NULL && thiscache != hintcache &&
1041 		    thiscache->mt_nfree == 0)
1042 			thiscache = thiscache->mt_next;
1043 
1044 		if (thiscache == hintcache)
1045 			thiscache = NULL;
1046 	}
1047 
1048 
1049 	if (thiscache == NULL) { /* there are no free caches */
1050 		int32_t thisrequest = requestsize;
1051 		int32_t buffer_size = (1 << logsz) + OVERHEAD;
1052 
1053 		thiscache = (cache_t *)morecore(thisrequest * HUNKSIZE);
1054 
1055 		if (thiscache == (cache_t *)-1) {
1056 			(void) mutex_unlock(&cpuptr->mt_parent_lock);
1057 			errno = EAGAIN;
1058 			return (NULL);
1059 		}
1060 		create_cache(thiscache, buffer_size, thisrequest);
1061 
1062 		/* link in the new block at the beginning of the list */
1063 		thiscache->mt_next = cachehead->mt_cache;
1064 		cachehead->mt_cache = thiscache;
1065 	}
1066 
1067 	/* update the hint to the cache we found or created */
1068 	cachehead->mt_hint = thiscache;
1069 
1070 	/* thiscache now points to a cache with available space */
1071 	(void) mutex_lock(&thiscache->mt_cache_lock);
1072 
1073 	freeblocks = (uint32_t *)thiscache->mt_freelist;
1074 	while (freeblocks < (uint32_t *)thiscache->mt_arena) {
1075 		if (*freeblocks & 0xffffffff)
1076 			break;
1077 		freeblocks++;
1078 		if (freeblocks < (uint32_t *)thiscache->mt_arena &&
1079 		    *freeblocks & 0xffffffff)
1080 			break;
1081 		freeblocks++;
1082 		if (freeblocks < (uint32_t *)thiscache->mt_arena &&
1083 		    *freeblocks & 0xffffffff)
1084 			break;
1085 		freeblocks++;
1086 		if (freeblocks < (uint32_t *)thiscache->mt_arena &&
1087 		    *freeblocks & 0xffffffff)
1088 			break;
1089 		freeblocks++;
1090 	}
1091 
1092 	/*
1093 	 * the offset from mt_freelist to freeblocks is the offset into
1094 	 * the arena. Be sure to include the offset into freeblocks
1095 	 * of the bitmask. n is the offset.
1096 	 */
1097 	for (i = 0; i < 32; ) {
1098 		if (FLIP_EM(*freeblocks) & (0x80000000 >> i++))
1099 			break;
1100 		if (FLIP_EM(*freeblocks) & (0x80000000 >> i++))
1101 			break;
1102 		if (FLIP_EM(*freeblocks) & (0x80000000 >> i++))
1103 			break;
1104 		if (FLIP_EM(*freeblocks) & (0x80000000 >> i++))
1105 			break;
1106 	}
1107 	index = 0x80000000 >> --i;
1108 
1109 
1110 	*freeblocks &= FLIP_EM(~index);
1111 
1112 	thiscache->mt_nfree--;
1113 
1114 	(void) mutex_unlock(&thiscache->mt_cache_lock);
1115 	(void) mutex_unlock(&cpuptr->mt_parent_lock);
1116 
1117 	n = (uintptr_t)(((freeblocks - (uint32_t *)thiscache->mt_freelist) << 5)
1118 	    + i) * thiscache->mt_size;
1119 	/*
1120 	 * Now you have the offset in n, you've changed the free mask
1121 	 * in the freelist. Nothing left to do but find the block
1122 	 * in the arena and put the value of thiscache in the word
1123 	 * ahead of the handed out address and return the memory
1124 	 * back to the user.
1125 	 */
1126 	ret = thiscache->mt_arena + n;
1127 
1128 	/* Store the cache addr for this buf. Makes free go fast. */
1129 	*(uintptr_t *)ret = (uintptr_t)thiscache;
1130 
1131 	/*
1132 	 * This assert makes sure we don't hand out memory that is not
1133 	 * owned by this cache.
1134 	 */
1135 	assert(ret + thiscache->mt_size <= thiscache->mt_freelist +
1136 	    thiscache->mt_span);
1137 
1138 	ret += OVERHEAD;
1139 
1140 	assert(((uintptr_t)ret & 7) == 0); /* are we 8 byte aligned */
1141 
1142 	if (reinit == 0 && (debugopt & MTDEBUGPATTERN))
1143 		if (verify_pattern(FREEPATTERN, ret, size))
1144 			abort();	/* reference after free */
1145 
1146 	if (debugopt & MTINITBUFFER)
1147 		copy_pattern(INITPATTERN, ret, size);
1148 	return ((void *)ret);
1149 }
1150 
1151 static void *
1152 morecore(size_t bytes)
1153 {
1154 	void * ret;
1155 
1156 	if (bytes > LONG_MAX) {
1157 		intptr_t wad;
1158 		/*
1159 		 * The request size is too big. We need to do this in
1160 		 * chunks. Sbrk only takes an int for an arg.
1161 		 */
1162 		if (bytes == ULONG_MAX)
1163 			return ((void *)-1);
1164 
1165 		ret = sbrk(0);
1166 		wad = LONG_MAX;
1167 		while (wad > 0) {
1168 			if (sbrk(wad) == (void *)-1) {
1169 				if (ret != sbrk(0))
1170 					(void) sbrk(-LONG_MAX);
1171 				return ((void *)-1);
1172 			}
1173 			bytes -= LONG_MAX;
1174 			wad = bytes;
1175 		}
1176 	} else
1177 		ret = sbrk(bytes);
1178 
1179 	return (ret);
1180 }
1181 
1182 
1183 static void *
1184 oversize(size_t size)
1185 {
1186 	caddr_t ret;
1187 	oversize_t *big;
1188 
1189 	/* make sure we will not overflow */
1190 	if (size > MAX_MTMALLOC) {
1191 		errno = ENOMEM;
1192 		return (NULL);
1193 	}
1194 
1195 	/*
1196 	 * Since we ensure every address we hand back is
1197 	 * MTMALLOC_MIN_ALIGN-byte aligned, ALIGNing size ensures that the
1198 	 * memory handed out is MTMALLOC_MIN_ALIGN-byte aligned at both ends.
1199 	 * This eases the implementation of MTDEBUGPATTERN and MTINITPATTERN,
1200 	 * particularly where coalescing occurs.
1201 	 */
1202 	size = ALIGN(size, MTMALLOC_MIN_ALIGN);
1203 
1204 	/*
1205 	 * The idea with the global lock is that we are sure to
1206 	 * block in the kernel anyway since given an oversize alloc
1207 	 * we are sure to have to call morecore();
1208 	 */
1209 	(void) mutex_lock(&oversize_lock);
1210 
1211 	if ((big = find_oversize(size)) != NULL) {
1212 		if (reinit == 0 && (debugopt & MTDEBUGPATTERN))
1213 			if (verify_pattern(FREEPATTERN, big->addr, size))
1214 				abort();	/* reference after free */
1215 	} else {
1216 		/* Get more 8-byte aligned memory from heap */
1217 		ret = morecore(size + OVSZ_HEADER_SIZE);
1218 		if (ret == (caddr_t)-1) {
1219 			(void) mutex_unlock(&oversize_lock);
1220 			errno = ENOMEM;
1221 			return (NULL);
1222 		}
1223 		big = oversize_header_alloc((uintptr_t)ret, size);
1224 	}
1225 	ret = big->addr;
1226 
1227 	insert_hash(big);
1228 
1229 	if (debugopt & MTINITBUFFER)
1230 		copy_pattern(INITPATTERN, ret, size);
1231 
1232 	(void) mutex_unlock(&oversize_lock);
1233 	assert(((uintptr_t)ret & 7) == 0); /* are we 8 byte aligned */
1234 	return ((void *)ret);
1235 }
1236 
1237 static void
1238 insert_oversize(oversize_t *op, oversize_t *nx)
1239 {
1240 	oversize_t *sp;
1241 
1242 	/* locate correct insertion point in size-ordered list */
1243 	for (sp = oversize_list.next_bysize;
1244 	    sp != &oversize_list && (op->size > sp->size);
1245 	    sp = sp->next_bysize)
1246 		;
1247 
1248 	/* link into size-ordered list */
1249 	op->next_bysize = sp;
1250 	op->prev_bysize = sp->prev_bysize;
1251 	op->prev_bysize->next_bysize = op;
1252 	op->next_bysize->prev_bysize = op;
1253 
1254 	/*
1255 	 * link item into address-ordered list
1256 	 * (caller provides insertion point as an optimization)
1257 	 */
1258 	op->next_byaddr = nx;
1259 	op->prev_byaddr = nx->prev_byaddr;
1260 	op->prev_byaddr->next_byaddr = op;
1261 	op->next_byaddr->prev_byaddr = op;
1262 
1263 }
1264 
1265 static void
1266 unlink_oversize(oversize_t *lp)
1267 {
1268 	/* unlink from address list */
1269 	lp->prev_byaddr->next_byaddr = lp->next_byaddr;
1270 	lp->next_byaddr->prev_byaddr = lp->prev_byaddr;
1271 
1272 	/* unlink from size list */
1273 	lp->prev_bysize->next_bysize = lp->next_bysize;
1274 	lp->next_bysize->prev_bysize = lp->prev_bysize;
1275 }
1276 
1277 static void
1278 position_oversize_by_size(oversize_t *op)
1279 {
1280 	oversize_t *sp;
1281 
1282 	if (op->size > op->next_bysize->size ||
1283 	    op->size < op->prev_bysize->size) {
1284 
1285 		/* unlink from size list */
1286 		op->prev_bysize->next_bysize = op->next_bysize;
1287 		op->next_bysize->prev_bysize = op->prev_bysize;
1288 
1289 		/* locate correct insertion point in size-ordered list */
1290 		for (sp = oversize_list.next_bysize;
1291 		    sp != &oversize_list && (op->size > sp->size);
1292 		    sp = sp->next_bysize)
1293 			;
1294 
1295 		/* link into size-ordered list */
1296 		op->next_bysize = sp;
1297 		op->prev_bysize = sp->prev_bysize;
1298 		op->prev_bysize->next_bysize = op;
1299 		op->next_bysize->prev_bysize = op;
1300 	}
1301 }
1302 
1303 static void
1304 add_oversize(oversize_t *lp)
1305 {
1306 	int merge_flags = INSERT_ONLY;
1307 	oversize_t *nx;  	/* ptr to item right of insertion point */
1308 	oversize_t *pv;  	/* ptr to item left of insertion point */
1309 	uint_t size_lp, size_pv, size_nx;
1310 	uintptr_t endp_lp, endp_pv, endp_nx;
1311 
1312 	/*
1313 	 * Locate insertion point in address-ordered list
1314 	 */
1315 
1316 	for (nx = oversize_list.next_byaddr;
1317 	    nx != &oversize_list && (lp->addr > nx->addr);
1318 	    nx = nx->next_byaddr)
1319 		;
1320 
1321 	/*
1322 	 * Determine how to add chunk to oversize freelist
1323 	 */
1324 
1325 	size_lp = OVSZ_HEADER_SIZE + lp->size;
1326 	endp_lp = ALIGN((uintptr_t)lp + size_lp, MTMALLOC_MIN_ALIGN);
1327 	size_lp = endp_lp - (uintptr_t)lp;
1328 
1329 	pv = nx->prev_byaddr;
1330 
1331 	if (pv->size) {
1332 
1333 		size_pv = OVSZ_HEADER_SIZE + pv->size;
1334 		endp_pv = ALIGN((uintptr_t)pv + size_pv,
1335 		    MTMALLOC_MIN_ALIGN);
1336 		size_pv = endp_pv - (uintptr_t)pv;
1337 
1338 		/* Check for adjacency with left chunk */
1339 		if ((uintptr_t)lp == endp_pv)
1340 			merge_flags |= COALESCE_LEFT;
1341 	}
1342 
1343 	if (nx->size) {
1344 
1345 		/* Check for adjacency with right chunk */
1346 		if ((uintptr_t)nx == endp_lp) {
1347 			size_nx = OVSZ_HEADER_SIZE + nx->size;
1348 			endp_nx = ALIGN((uintptr_t)nx + size_nx,
1349 			    MTMALLOC_MIN_ALIGN);
1350 			size_nx = endp_nx - (uintptr_t)nx;
1351 			merge_flags |= COALESCE_RIGHT;
1352 		}
1353 	}
1354 
1355 	/*
1356 	 * If MTDEBUGPATTERN==1, lp->addr will have been overwritten with
1357 	 * FREEPATTERN for lp->size bytes. If we can merge, the oversize
1358 	 * header(s) that will also become part of the memory available for
1359 	 * reallocation (ie lp and/or nx) must also be overwritten with
1360 	 * FREEPATTERN or we will SIGABRT when this memory is next reallocated.
1361 	 */
1362 	switch (merge_flags) {
1363 
1364 	case INSERT_ONLY:		/* Coalescing not possible */
1365 		insert_oversize(lp, nx);
1366 		break;
1367 	case COALESCE_LEFT:
1368 		pv->size += size_lp;
1369 		position_oversize_by_size(pv);
1370 		if (debugopt & MTDEBUGPATTERN)
1371 			copy_pattern(FREEPATTERN, lp, OVSZ_HEADER_SIZE);
1372 		break;
1373 	case COALESCE_RIGHT:
1374 		unlink_oversize(nx);
1375 		lp->size += size_nx;
1376 		insert_oversize(lp, pv->next_byaddr);
1377 		if (debugopt & MTDEBUGPATTERN)
1378 			copy_pattern(FREEPATTERN, nx, OVSZ_HEADER_SIZE);
1379 		break;
1380 	case COALESCE_WITH_BOTH_SIDES:	/* Merge (with right) to the left */
1381 		pv->size += size_lp + size_nx;
1382 		unlink_oversize(nx);
1383 		position_oversize_by_size(pv);
1384 		if (debugopt & MTDEBUGPATTERN) {
1385 			copy_pattern(FREEPATTERN, lp, OVSZ_HEADER_SIZE);
1386 			copy_pattern(FREEPATTERN, nx, OVSZ_HEADER_SIZE);
1387 		}
1388 		break;
1389 	}
1390 }
1391 
1392 /*
1393  * Find memory on our list that is at least size big. If we find a block that is
1394  * big enough, we break it up and return the associated oversize_t struct back
1395  * to the calling client. Any leftover piece of that block is returned to the
1396  * freelist.
1397  */
1398 static oversize_t *
1399 find_oversize(size_t size)
1400 {
1401 	oversize_t *wp = oversize_list.next_bysize;
1402 	while (wp != &oversize_list && size > wp->size)
1403 		wp = wp->next_bysize;
1404 
1405 	if (wp == &oversize_list) /* empty list or nothing big enough */
1406 		return (NULL);
1407 	/* breaking up a chunk of memory */
1408 	if ((long)((wp->size - (size + OVSZ_HEADER_SIZE + MTMALLOC_MIN_ALIGN)))
1409 	    > MAX_CACHED) {
1410 		caddr_t off;
1411 		oversize_t *np;
1412 		size_t osize;
1413 		off = (caddr_t)ALIGN(wp->addr + size,
1414 		    MTMALLOC_MIN_ALIGN);
1415 		osize = wp->size;
1416 		wp->size = (size_t)(off - wp->addr);
1417 		np = oversize_header_alloc((uintptr_t)off,
1418 		    osize - (wp->size + OVSZ_HEADER_SIZE));
1419 		if ((long)np->size < 0)
1420 			abort();
1421 		unlink_oversize(wp);
1422 		add_oversize(np);
1423 	} else {
1424 		unlink_oversize(wp);
1425 	}
1426 	return (wp);
1427 }
1428 
1429 static void
1430 copy_pattern(uint32_t pattern, void *buf_arg, size_t size)
1431 {
1432 	uint32_t *bufend = (uint32_t *)((char *)buf_arg + size);
1433 	uint32_t *buf = buf_arg;
1434 
1435 	while (buf < bufend - 3) {
1436 		buf[3] = buf[2] = buf[1] = buf[0] = pattern;
1437 		buf += 4;
1438 	}
1439 	while (buf < bufend)
1440 		*buf++ = pattern;
1441 }
1442 
1443 static void *
1444 verify_pattern(uint32_t pattern, void *buf_arg, size_t size)
1445 {
1446 	uint32_t *bufend = (uint32_t *)((char *)buf_arg + size);
1447 	uint32_t *buf;
1448 
1449 	for (buf = buf_arg; buf < bufend; buf++)
1450 		if (*buf != pattern)
1451 			return (buf);
1452 	return (NULL);
1453 }
1454 
1455 static void
1456 free_oversize(oversize_t *ovp)
1457 {
1458 	assert(((uintptr_t)ovp->addr & 7) == 0); /* are we 8 byte aligned */
1459 	assert(ovp->size > MAX_CACHED);
1460 
1461 	ovp->next_bysize = ovp->prev_bysize = NULL;
1462 	ovp->next_byaddr = ovp->prev_byaddr = NULL;
1463 	(void) mutex_lock(&oversize_lock);
1464 	add_oversize(ovp);
1465 	(void) mutex_unlock(&oversize_lock);
1466 }
1467 
1468 static oversize_t *
1469 oversize_header_alloc(uintptr_t mem, size_t size)
1470 {
1471 	oversize_t *ovsz_hdr;
1472 
1473 	assert(size > MAX_CACHED);
1474 
1475 	ovsz_hdr = (oversize_t *)mem;
1476 	ovsz_hdr->prev_bysize = NULL;
1477 	ovsz_hdr->next_bysize = NULL;
1478 	ovsz_hdr->prev_byaddr = NULL;
1479 	ovsz_hdr->next_byaddr = NULL;
1480 	ovsz_hdr->hash_next = NULL;
1481 	ovsz_hdr->size = size;
1482 	mem += OVSZ_SIZE;
1483 	*(uintptr_t *)mem = MTMALLOC_OVERSIZE_MAGIC;
1484 	mem += OVERHEAD;
1485 	assert(((uintptr_t)mem & 7) == 0); /* are we 8 byte aligned */
1486 	ovsz_hdr->addr = (caddr_t)mem;
1487 	return (ovsz_hdr);
1488 }
1489 
1490 static void
1491 malloc_prepare()
1492 {
1493 	percpu_t *cpuptr;
1494 	cache_head_t *cachehead;
1495 	cache_t *thiscache;
1496 
1497 	(void) mutex_lock(&oversize_lock);
1498 	for (cpuptr = &cpu_list[0]; cpuptr < &cpu_list[ncpus]; cpuptr++) {
1499 		(void) mutex_lock(&cpuptr->mt_parent_lock);
1500 		for (cachehead = &cpuptr->mt_caches[0];
1501 		    cachehead < &cpuptr->mt_caches[NUM_CACHES];
1502 		    cachehead++) {
1503 			for (thiscache = cachehead->mt_cache;
1504 			    thiscache != NULL;
1505 			    thiscache = thiscache->mt_next) {
1506 				(void) mutex_lock(
1507 				    &thiscache->mt_cache_lock);
1508 			}
1509 		}
1510 	}
1511 }
1512 
1513 static void
1514 malloc_release()
1515 {
1516 	percpu_t *cpuptr;
1517 	cache_head_t *cachehead;
1518 	cache_t *thiscache;
1519 
1520 	for (cpuptr = &cpu_list[ncpus - 1]; cpuptr >= &cpu_list[0]; cpuptr--) {
1521 		for (cachehead = &cpuptr->mt_caches[NUM_CACHES - 1];
1522 		    cachehead >= &cpuptr->mt_caches[0];
1523 		    cachehead--) {
1524 			for (thiscache = cachehead->mt_cache;
1525 			    thiscache != NULL;
1526 			    thiscache = thiscache->mt_next) {
1527 				(void) mutex_unlock(
1528 				    &thiscache->mt_cache_lock);
1529 			}
1530 		}
1531 		(void) mutex_unlock(&cpuptr->mt_parent_lock);
1532 	}
1533 	(void) mutex_unlock(&oversize_lock);
1534 }
1535 
1536 #pragma init(malloc_init)
1537 static void
1538 malloc_init(void)
1539 {
1540 	/*
1541 	 * This works in the init section for this library
1542 	 * because setup_caches() doesn't call anything in libc
1543 	 * that calls malloc().  If it did, disaster would ensue.
1544 	 *
1545 	 * For this to work properly, this library must be the first
1546 	 * one to have its init section called (after libc) by the
1547 	 * dynamic linker.  If some other library's init section
1548 	 * ran first and called malloc(), disaster would ensue.
1549 	 * Because this is an interposer library for malloc(), the
1550 	 * dynamic linker arranges for its init section to run first.
1551 	 */
1552 	(void) setup_caches();
1553 
1554 	(void) pthread_atfork(malloc_prepare, malloc_release, malloc_release);
1555 }
1556