xref: /illumos-gate/usr/src/lib/libumem/common/vmem.c (revision b6805bf78d2bbbeeaea8909a05623587b42d58b3)
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  * Copyright 2012 Joyent, Inc. All rights reserved.
26  */
27 
28 /*
29  * For a more complete description of the main ideas, see:
30  *
31  *	Jeff Bonwick and Jonathan Adams,
32  *
33  *	Magazines and vmem: Extending the Slab Allocator to Many CPUs and
34  *	Arbitrary Resources.
35  *
36  *	Proceedings of the 2001 Usenix Conference.
37  *	Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
38  *
39  * For the "Big Theory Statement", see usr/src/uts/common/os/vmem.c
40  *
41  * 1. Overview of changes
42  * ------------------------------
43  * There have been a few changes to vmem in order to support umem.  The
44  * main areas are:
45  *
46  *	* VM_SLEEP unsupported
47  *
48  *	* Reaping changes
49  *
50  *	* initialization changes
51  *
52  *	* _vmem_extend_alloc
53  *
54  *
55  * 2. VM_SLEEP Removed
56  * -------------------
57  * Since VM_SLEEP allocations can hold locks (in vmem_populate()) for
58  * possibly infinite amounts of time, they are not supported in this
59  * version of vmem.  Sleep-like behavior can be achieved through
60  * UMEM_NOFAIL umem allocations.
61  *
62  *
63  * 3. Reaping changes
64  * ------------------
65  * Unlike kmem_reap(), which just asynchronously schedules work, umem_reap()
66  * can do allocations and frees synchronously.  This is a problem if it
67  * occurs during a vmem_populate() allocation.
68  *
69  * Instead, we delay reaps while populates are active.
70  *
71  *
72  * 4. Initialization changes
73  * -------------------------
74  * In the kernel, vmem_init() allows you to create a single, top-level arena,
75  * which has vmem_internal_arena as a child.  For umem, we want to be able
76  * to extend arenas dynamically.  It is much easier to support this if we
77  * allow a two-level "heap" arena:
78  *
79  *	+----------+
80  *	|  "fake"  |
81  *	+----------+
82  *	      |
83  *	+----------+
84  *	|  "heap"  |
85  *	+----------+
86  *	  |    \ \
87  *	  |     +-+-- ... <other children>
88  *	  |
89  *	+---------------+
90  *	| vmem_internal |
91  *	+---------------+
92  *	    | | | |
93  *	   <children>
94  *
95  * The new vmem_init() allows you to specify a "parent" of the heap, along
96  * with allocation functions.
97  *
98  *
99  * 5. _vmem_extend_alloc
100  * ---------------------
101  * The other part of extending is _vmem_extend_alloc.  This function allows
102  * you to extend (expand current spans, if possible) an arena and allocate
103  * a chunk of the newly extened span atomically.  This is needed to support
104  * extending the heap while vmem_populate()ing it.
105  *
106  * In order to increase the usefulness of extending, non-imported spans are
107  * sorted in address order.
108  */
109 
110 #include <sys/vmem_impl_user.h>
111 #include <alloca.h>
112 #include <sys/sysmacros.h>
113 #include <stdio.h>
114 #include <strings.h>
115 #include <atomic.h>
116 
117 #include "vmem_base.h"
118 #include "umem_base.h"
119 
120 #define	VMEM_INITIAL		6	/* early vmem arenas */
121 #define	VMEM_SEG_INITIAL	100	/* early segments */
122 
123 /*
124  * Adding a new span to an arena requires two segment structures: one to
125  * represent the span, and one to represent the free segment it contains.
126  */
127 #define	VMEM_SEGS_PER_SPAN_CREATE	2
128 
129 /*
130  * Allocating a piece of an existing segment requires 0-2 segment structures
131  * depending on how much of the segment we're allocating.
132  *
133  * To allocate the entire segment, no new segment structures are needed; we
134  * simply move the existing segment structure from the freelist to the
135  * allocation hash table.
136  *
137  * To allocate a piece from the left or right end of the segment, we must
138  * split the segment into two pieces (allocated part and remainder), so we
139  * need one new segment structure to represent the remainder.
140  *
141  * To allocate from the middle of a segment, we need two new segment strucures
142  * to represent the remainders on either side of the allocated part.
143  */
144 #define	VMEM_SEGS_PER_EXACT_ALLOC	0
145 #define	VMEM_SEGS_PER_LEFT_ALLOC	1
146 #define	VMEM_SEGS_PER_RIGHT_ALLOC	1
147 #define	VMEM_SEGS_PER_MIDDLE_ALLOC	2
148 
149 /*
150  * vmem_populate() preallocates segment structures for vmem to do its work.
151  * It must preallocate enough for the worst case, which is when we must import
152  * a new span and then allocate from the middle of it.
153  */
154 #define	VMEM_SEGS_PER_ALLOC_MAX		\
155 	(VMEM_SEGS_PER_SPAN_CREATE + VMEM_SEGS_PER_MIDDLE_ALLOC)
156 
157 /*
158  * The segment structures themselves are allocated from vmem_seg_arena, so
159  * we have a recursion problem when vmem_seg_arena needs to populate itself.
160  * We address this by working out the maximum number of segment structures
161  * this act will require, and multiplying by the maximum number of threads
162  * that we'll allow to do it simultaneously.
163  *
164  * The worst-case segment consumption to populate vmem_seg_arena is as
165  * follows (depicted as a stack trace to indicate why events are occurring):
166  *
167  * vmem_alloc(vmem_seg_arena)		-> 2 segs (span create + exact alloc)
168  *  vmem_alloc(vmem_internal_arena)	-> 2 segs (span create + exact alloc)
169  *   heap_alloc(heap_arena)
170  *    vmem_alloc(heap_arena)		-> 4 seg (span create + alloc)
171  *     parent_alloc(parent_arena)
172  *	_vmem_extend_alloc(parent_arena) -> 3 seg (span create + left alloc)
173  *
174  * Note:  The reservation for heap_arena must be 4, since vmem_xalloc()
175  * is overly pessimistic on allocations where parent_arena has a stricter
176  * alignment than heap_arena.
177  *
178  * The worst-case consumption for any arena is 4 segment structures.
179  * For now, we only support VM_NOSLEEP allocations, so as long as we
180  * serialize all vmem_populates, a 4-seg reserve is sufficient.
181  */
182 #define	VMEM_POPULATE_SEGS_PER_ARENA	4
183 #define	VMEM_POPULATE_LOCKS		1
184 
185 #define	VMEM_POPULATE_RESERVE		\
186 	(VMEM_POPULATE_SEGS_PER_ARENA * VMEM_POPULATE_LOCKS)
187 
188 /*
189  * vmem_populate() ensures that each arena has VMEM_MINFREE seg structures
190  * so that it can satisfy the worst-case allocation *and* participate in
191  * worst-case allocation from vmem_seg_arena.
192  */
193 #define	VMEM_MINFREE	(VMEM_POPULATE_RESERVE + VMEM_SEGS_PER_ALLOC_MAX)
194 
195 /* Don't assume new statics are zeroed - see vmem_startup() */
196 static vmem_t vmem0[VMEM_INITIAL];
197 static vmem_t *vmem_populator[VMEM_INITIAL];
198 static uint32_t vmem_id;
199 static uint32_t vmem_populators;
200 static vmem_seg_t vmem_seg0[VMEM_SEG_INITIAL];
201 static vmem_seg_t *vmem_segfree;
202 static mutex_t vmem_list_lock;
203 static mutex_t vmem_segfree_lock;
204 static vmem_populate_lock_t vmem_nosleep_lock;
205 #define	IN_POPULATE()	(vmem_nosleep_lock.vmpl_thr == thr_self())
206 static vmem_t *vmem_list;
207 static vmem_t *vmem_internal_arena;
208 static vmem_t *vmem_seg_arena;
209 static vmem_t *vmem_hash_arena;
210 static vmem_t *vmem_vmem_arena;
211 
212 vmem_t *vmem_heap;
213 vmem_alloc_t *vmem_heap_alloc;
214 vmem_free_t *vmem_heap_free;
215 
216 uint32_t vmem_mtbf;		/* mean time between failures [default: off] */
217 size_t vmem_seg_size = sizeof (vmem_seg_t);
218 
219 /*
220  * Insert/delete from arena list (type 'a') or next-of-kin list (type 'k').
221  */
222 #define	VMEM_INSERT(vprev, vsp, type)					\
223 {									\
224 	vmem_seg_t *vnext = (vprev)->vs_##type##next;			\
225 	(vsp)->vs_##type##next = (vnext);				\
226 	(vsp)->vs_##type##prev = (vprev);				\
227 	(vprev)->vs_##type##next = (vsp);				\
228 	(vnext)->vs_##type##prev = (vsp);				\
229 }
230 
231 #define	VMEM_DELETE(vsp, type)						\
232 {									\
233 	vmem_seg_t *vprev = (vsp)->vs_##type##prev;			\
234 	vmem_seg_t *vnext = (vsp)->vs_##type##next;			\
235 	(vprev)->vs_##type##next = (vnext);				\
236 	(vnext)->vs_##type##prev = (vprev);				\
237 }
238 
239 /*
240  * Get a vmem_seg_t from the global segfree list.
241  */
242 static vmem_seg_t *
243 vmem_getseg_global(void)
244 {
245 	vmem_seg_t *vsp;
246 
247 	(void) mutex_lock(&vmem_segfree_lock);
248 	if ((vsp = vmem_segfree) != NULL)
249 		vmem_segfree = vsp->vs_knext;
250 	(void) mutex_unlock(&vmem_segfree_lock);
251 
252 	return (vsp);
253 }
254 
255 /*
256  * Put a vmem_seg_t on the global segfree list.
257  */
258 static void
259 vmem_putseg_global(vmem_seg_t *vsp)
260 {
261 	(void) mutex_lock(&vmem_segfree_lock);
262 	vsp->vs_knext = vmem_segfree;
263 	vmem_segfree = vsp;
264 	(void) mutex_unlock(&vmem_segfree_lock);
265 }
266 
267 /*
268  * Get a vmem_seg_t from vmp's segfree list.
269  */
270 static vmem_seg_t *
271 vmem_getseg(vmem_t *vmp)
272 {
273 	vmem_seg_t *vsp;
274 
275 	ASSERT(vmp->vm_nsegfree > 0);
276 
277 	vsp = vmp->vm_segfree;
278 	vmp->vm_segfree = vsp->vs_knext;
279 	vmp->vm_nsegfree--;
280 
281 	return (vsp);
282 }
283 
284 /*
285  * Put a vmem_seg_t on vmp's segfree list.
286  */
287 static void
288 vmem_putseg(vmem_t *vmp, vmem_seg_t *vsp)
289 {
290 	vsp->vs_knext = vmp->vm_segfree;
291 	vmp->vm_segfree = vsp;
292 	vmp->vm_nsegfree++;
293 }
294 
295 /*
296  * Add vsp to the appropriate freelist.
297  */
298 static void
299 vmem_freelist_insert(vmem_t *vmp, vmem_seg_t *vsp)
300 {
301 	vmem_seg_t *vprev;
302 
303 	ASSERT(*VMEM_HASH(vmp, vsp->vs_start) != vsp);
304 
305 	vprev = (vmem_seg_t *)&vmp->vm_freelist[highbit(VS_SIZE(vsp)) - 1];
306 	vsp->vs_type = VMEM_FREE;
307 	vmp->vm_freemap |= VS_SIZE(vprev);
308 	VMEM_INSERT(vprev, vsp, k);
309 
310 	(void) cond_broadcast(&vmp->vm_cv);
311 }
312 
313 /*
314  * Take vsp from the freelist.
315  */
316 static void
317 vmem_freelist_delete(vmem_t *vmp, vmem_seg_t *vsp)
318 {
319 	ASSERT(*VMEM_HASH(vmp, vsp->vs_start) != vsp);
320 	ASSERT(vsp->vs_type == VMEM_FREE);
321 
322 	if (vsp->vs_knext->vs_start == 0 && vsp->vs_kprev->vs_start == 0) {
323 		/*
324 		 * The segments on both sides of 'vsp' are freelist heads,
325 		 * so taking vsp leaves the freelist at vsp->vs_kprev empty.
326 		 */
327 		ASSERT(vmp->vm_freemap & VS_SIZE(vsp->vs_kprev));
328 		vmp->vm_freemap ^= VS_SIZE(vsp->vs_kprev);
329 	}
330 	VMEM_DELETE(vsp, k);
331 }
332 
333 /*
334  * Add vsp to the allocated-segment hash table and update kstats.
335  */
336 static void
337 vmem_hash_insert(vmem_t *vmp, vmem_seg_t *vsp)
338 {
339 	vmem_seg_t **bucket;
340 
341 	vsp->vs_type = VMEM_ALLOC;
342 	bucket = VMEM_HASH(vmp, vsp->vs_start);
343 	vsp->vs_knext = *bucket;
344 	*bucket = vsp;
345 
346 	if (vmem_seg_size == sizeof (vmem_seg_t)) {
347 		vsp->vs_depth = (uint8_t)getpcstack(vsp->vs_stack,
348 		    VMEM_STACK_DEPTH, 0);
349 		vsp->vs_thread = thr_self();
350 		vsp->vs_timestamp = gethrtime();
351 	} else {
352 		vsp->vs_depth = 0;
353 	}
354 
355 	vmp->vm_kstat.vk_alloc++;
356 	vmp->vm_kstat.vk_mem_inuse += VS_SIZE(vsp);
357 }
358 
359 /*
360  * Remove vsp from the allocated-segment hash table and update kstats.
361  */
362 static vmem_seg_t *
363 vmem_hash_delete(vmem_t *vmp, uintptr_t addr, size_t size)
364 {
365 	vmem_seg_t *vsp, **prev_vspp;
366 
367 	prev_vspp = VMEM_HASH(vmp, addr);
368 	while ((vsp = *prev_vspp) != NULL) {
369 		if (vsp->vs_start == addr) {
370 			*prev_vspp = vsp->vs_knext;
371 			break;
372 		}
373 		vmp->vm_kstat.vk_lookup++;
374 		prev_vspp = &vsp->vs_knext;
375 	}
376 
377 	if (vsp == NULL) {
378 		umem_panic("vmem_hash_delete(%p, %lx, %lu): bad free",
379 		    vmp, addr, size);
380 	}
381 	if (VS_SIZE(vsp) != size) {
382 		umem_panic("vmem_hash_delete(%p, %lx, %lu): wrong size "
383 		    "(expect %lu)", vmp, addr, size, VS_SIZE(vsp));
384 	}
385 
386 	vmp->vm_kstat.vk_free++;
387 	vmp->vm_kstat.vk_mem_inuse -= size;
388 
389 	return (vsp);
390 }
391 
392 /*
393  * Create a segment spanning the range [start, end) and add it to the arena.
394  */
395 static vmem_seg_t *
396 vmem_seg_create(vmem_t *vmp, vmem_seg_t *vprev, uintptr_t start, uintptr_t end)
397 {
398 	vmem_seg_t *newseg = vmem_getseg(vmp);
399 
400 	newseg->vs_start = start;
401 	newseg->vs_end = end;
402 	newseg->vs_type = 0;
403 	newseg->vs_import = 0;
404 
405 	VMEM_INSERT(vprev, newseg, a);
406 
407 	return (newseg);
408 }
409 
410 /*
411  * Remove segment vsp from the arena.
412  */
413 static void
414 vmem_seg_destroy(vmem_t *vmp, vmem_seg_t *vsp)
415 {
416 	ASSERT(vsp->vs_type != VMEM_ROTOR);
417 	VMEM_DELETE(vsp, a);
418 
419 	vmem_putseg(vmp, vsp);
420 }
421 
422 /*
423  * Add the span [vaddr, vaddr + size) to vmp and update kstats.
424  */
425 static vmem_seg_t *
426 vmem_span_create(vmem_t *vmp, void *vaddr, size_t size, uint8_t import)
427 {
428 	vmem_seg_t *knext;
429 	vmem_seg_t *newseg, *span;
430 	uintptr_t start = (uintptr_t)vaddr;
431 	uintptr_t end = start + size;
432 
433 	knext = &vmp->vm_seg0;
434 	if (!import && vmp->vm_source_alloc == NULL) {
435 		vmem_seg_t *kend, *kprev;
436 		/*
437 		 * non-imported spans are sorted in address order.  This
438 		 * makes vmem_extend_unlocked() much more effective.
439 		 *
440 		 * We search in reverse order, since new spans are
441 		 * generally at higher addresses.
442 		 */
443 		kend = &vmp->vm_seg0;
444 		for (kprev = kend->vs_kprev; kprev != kend;
445 		    kprev = kprev->vs_kprev) {
446 			if (!kprev->vs_import && (kprev->vs_end - 1) < start)
447 				break;
448 		}
449 		knext = kprev->vs_knext;
450 	}
451 
452 	ASSERT(MUTEX_HELD(&vmp->vm_lock));
453 
454 	if ((start | end) & (vmp->vm_quantum - 1)) {
455 		umem_panic("vmem_span_create(%p, %p, %lu): misaligned",
456 		    vmp, vaddr, size);
457 	}
458 
459 	span = vmem_seg_create(vmp, knext->vs_aprev, start, end);
460 	span->vs_type = VMEM_SPAN;
461 	VMEM_INSERT(knext->vs_kprev, span, k);
462 
463 	newseg = vmem_seg_create(vmp, span, start, end);
464 	vmem_freelist_insert(vmp, newseg);
465 
466 	newseg->vs_import = import;
467 	if (import)
468 		vmp->vm_kstat.vk_mem_import += size;
469 	vmp->vm_kstat.vk_mem_total += size;
470 
471 	return (newseg);
472 }
473 
474 /*
475  * Remove span vsp from vmp and update kstats.
476  */
477 static void
478 vmem_span_destroy(vmem_t *vmp, vmem_seg_t *vsp)
479 {
480 	vmem_seg_t *span = vsp->vs_aprev;
481 	size_t size = VS_SIZE(vsp);
482 
483 	ASSERT(MUTEX_HELD(&vmp->vm_lock));
484 	ASSERT(span->vs_type == VMEM_SPAN);
485 
486 	if (vsp->vs_import)
487 		vmp->vm_kstat.vk_mem_import -= size;
488 	vmp->vm_kstat.vk_mem_total -= size;
489 
490 	VMEM_DELETE(span, k);
491 
492 	vmem_seg_destroy(vmp, vsp);
493 	vmem_seg_destroy(vmp, span);
494 }
495 
496 /*
497  * Allocate the subrange [addr, addr + size) from segment vsp.
498  * If there are leftovers on either side, place them on the freelist.
499  * Returns a pointer to the segment representing [addr, addr + size).
500  */
501 static vmem_seg_t *
502 vmem_seg_alloc(vmem_t *vmp, vmem_seg_t *vsp, uintptr_t addr, size_t size)
503 {
504 	uintptr_t vs_start = vsp->vs_start;
505 	uintptr_t vs_end = vsp->vs_end;
506 	size_t vs_size = vs_end - vs_start;
507 	size_t realsize = P2ROUNDUP(size, vmp->vm_quantum);
508 	uintptr_t addr_end = addr + realsize;
509 
510 	ASSERT(P2PHASE(vs_start, vmp->vm_quantum) == 0);
511 	ASSERT(P2PHASE(addr, vmp->vm_quantum) == 0);
512 	ASSERT(vsp->vs_type == VMEM_FREE);
513 	ASSERT(addr >= vs_start && addr_end - 1 <= vs_end - 1);
514 	ASSERT(addr - 1 <= addr_end - 1);
515 
516 	/*
517 	 * If we're allocating from the start of the segment, and the
518 	 * remainder will be on the same freelist, we can save quite
519 	 * a bit of work.
520 	 */
521 	if (P2SAMEHIGHBIT(vs_size, vs_size - realsize) && addr == vs_start) {
522 		ASSERT(highbit(vs_size) == highbit(vs_size - realsize));
523 		vsp->vs_start = addr_end;
524 		vsp = vmem_seg_create(vmp, vsp->vs_aprev, addr, addr + size);
525 		vmem_hash_insert(vmp, vsp);
526 		return (vsp);
527 	}
528 
529 	vmem_freelist_delete(vmp, vsp);
530 
531 	if (vs_end != addr_end)
532 		vmem_freelist_insert(vmp,
533 		    vmem_seg_create(vmp, vsp, addr_end, vs_end));
534 
535 	if (vs_start != addr)
536 		vmem_freelist_insert(vmp,
537 		    vmem_seg_create(vmp, vsp->vs_aprev, vs_start, addr));
538 
539 	vsp->vs_start = addr;
540 	vsp->vs_end = addr + size;
541 
542 	vmem_hash_insert(vmp, vsp);
543 	return (vsp);
544 }
545 
546 /*
547  * We cannot reap if we are in the middle of a vmem_populate().
548  */
549 void
550 vmem_reap(void)
551 {
552 	if (!IN_POPULATE())
553 		umem_reap();
554 }
555 
556 /*
557  * Populate vmp's segfree list with VMEM_MINFREE vmem_seg_t structures.
558  */
559 static int
560 vmem_populate(vmem_t *vmp, int vmflag)
561 {
562 	char *p;
563 	vmem_seg_t *vsp;
564 	ssize_t nseg;
565 	size_t size;
566 	vmem_populate_lock_t *lp;
567 	int i;
568 
569 	while (vmp->vm_nsegfree < VMEM_MINFREE &&
570 	    (vsp = vmem_getseg_global()) != NULL)
571 		vmem_putseg(vmp, vsp);
572 
573 	if (vmp->vm_nsegfree >= VMEM_MINFREE)
574 		return (1);
575 
576 	/*
577 	 * If we're already populating, tap the reserve.
578 	 */
579 	if (vmem_nosleep_lock.vmpl_thr == thr_self()) {
580 		ASSERT(vmp->vm_cflags & VMC_POPULATOR);
581 		return (1);
582 	}
583 
584 	(void) mutex_unlock(&vmp->vm_lock);
585 
586 	ASSERT(vmflag & VM_NOSLEEP);	/* we do not allow sleep allocations */
587 	lp = &vmem_nosleep_lock;
588 
589 	/*
590 	 * Cannot be just a mutex_lock(), since that has no effect if
591 	 * libthread is not linked.
592 	 */
593 	(void) mutex_lock(&lp->vmpl_mutex);
594 	ASSERT(lp->vmpl_thr == 0);
595 	lp->vmpl_thr = thr_self();
596 
597 	nseg = VMEM_MINFREE + vmem_populators * VMEM_POPULATE_RESERVE;
598 	size = P2ROUNDUP(nseg * vmem_seg_size, vmem_seg_arena->vm_quantum);
599 	nseg = size / vmem_seg_size;
600 
601 	/*
602 	 * The following vmem_alloc() may need to populate vmem_seg_arena
603 	 * and all the things it imports from.  When doing so, it will tap
604 	 * each arena's reserve to prevent recursion (see the block comment
605 	 * above the definition of VMEM_POPULATE_RESERVE).
606 	 *
607 	 * During this allocation, vmem_reap() is a no-op.  If the allocation
608 	 * fails, we call vmem_reap() after dropping the population lock.
609 	 */
610 	p = vmem_alloc(vmem_seg_arena, size, vmflag & VM_UMFLAGS);
611 	if (p == NULL) {
612 		lp->vmpl_thr = 0;
613 		(void) mutex_unlock(&lp->vmpl_mutex);
614 		vmem_reap();
615 
616 		(void) mutex_lock(&vmp->vm_lock);
617 		vmp->vm_kstat.vk_populate_fail++;
618 		return (0);
619 	}
620 	/*
621 	 * Restock the arenas that may have been depleted during population.
622 	 */
623 	for (i = 0; i < vmem_populators; i++) {
624 		(void) mutex_lock(&vmem_populator[i]->vm_lock);
625 		while (vmem_populator[i]->vm_nsegfree < VMEM_POPULATE_RESERVE)
626 			vmem_putseg(vmem_populator[i],
627 			    (vmem_seg_t *)(p + --nseg * vmem_seg_size));
628 		(void) mutex_unlock(&vmem_populator[i]->vm_lock);
629 	}
630 
631 	lp->vmpl_thr = 0;
632 	(void) mutex_unlock(&lp->vmpl_mutex);
633 	(void) mutex_lock(&vmp->vm_lock);
634 
635 	/*
636 	 * Now take our own segments.
637 	 */
638 	ASSERT(nseg >= VMEM_MINFREE);
639 	while (vmp->vm_nsegfree < VMEM_MINFREE)
640 		vmem_putseg(vmp, (vmem_seg_t *)(p + --nseg * vmem_seg_size));
641 
642 	/*
643 	 * Give the remainder to charity.
644 	 */
645 	while (nseg > 0)
646 		vmem_putseg_global((vmem_seg_t *)(p + --nseg * vmem_seg_size));
647 
648 	return (1);
649 }
650 
651 /*
652  * Advance a walker from its previous position to 'afterme'.
653  * Note: may drop and reacquire vmp->vm_lock.
654  */
655 static void
656 vmem_advance(vmem_t *vmp, vmem_seg_t *walker, vmem_seg_t *afterme)
657 {
658 	vmem_seg_t *vprev = walker->vs_aprev;
659 	vmem_seg_t *vnext = walker->vs_anext;
660 	vmem_seg_t *vsp = NULL;
661 
662 	VMEM_DELETE(walker, a);
663 
664 	if (afterme != NULL)
665 		VMEM_INSERT(afterme, walker, a);
666 
667 	/*
668 	 * The walker segment's presence may have prevented its neighbors
669 	 * from coalescing.  If so, coalesce them now.
670 	 */
671 	if (vprev->vs_type == VMEM_FREE) {
672 		if (vnext->vs_type == VMEM_FREE) {
673 			ASSERT(vprev->vs_end == vnext->vs_start);
674 			vmem_freelist_delete(vmp, vnext);
675 			vmem_freelist_delete(vmp, vprev);
676 			vprev->vs_end = vnext->vs_end;
677 			vmem_freelist_insert(vmp, vprev);
678 			vmem_seg_destroy(vmp, vnext);
679 		}
680 		vsp = vprev;
681 	} else if (vnext->vs_type == VMEM_FREE) {
682 		vsp = vnext;
683 	}
684 
685 	/*
686 	 * vsp could represent a complete imported span,
687 	 * in which case we must return it to the source.
688 	 */
689 	if (vsp != NULL && vsp->vs_import && vmp->vm_source_free != NULL &&
690 	    vsp->vs_aprev->vs_type == VMEM_SPAN &&
691 	    vsp->vs_anext->vs_type == VMEM_SPAN) {
692 		void *vaddr = (void *)vsp->vs_start;
693 		size_t size = VS_SIZE(vsp);
694 		ASSERT(size == VS_SIZE(vsp->vs_aprev));
695 		vmem_freelist_delete(vmp, vsp);
696 		vmem_span_destroy(vmp, vsp);
697 		(void) mutex_unlock(&vmp->vm_lock);
698 		vmp->vm_source_free(vmp->vm_source, vaddr, size);
699 		(void) mutex_lock(&vmp->vm_lock);
700 	}
701 }
702 
703 /*
704  * VM_NEXTFIT allocations deliberately cycle through all virtual addresses
705  * in an arena, so that we avoid reusing addresses for as long as possible.
706  * This helps to catch used-after-freed bugs.  It's also the perfect policy
707  * for allocating things like process IDs, where we want to cycle through
708  * all values in order.
709  */
710 static void *
711 vmem_nextfit_alloc(vmem_t *vmp, size_t size, int vmflag)
712 {
713 	vmem_seg_t *vsp, *rotor;
714 	uintptr_t addr;
715 	size_t realsize = P2ROUNDUP(size, vmp->vm_quantum);
716 	size_t vs_size;
717 
718 	(void) mutex_lock(&vmp->vm_lock);
719 
720 	if (vmp->vm_nsegfree < VMEM_MINFREE && !vmem_populate(vmp, vmflag)) {
721 		(void) mutex_unlock(&vmp->vm_lock);
722 		return (NULL);
723 	}
724 
725 	/*
726 	 * The common case is that the segment right after the rotor is free,
727 	 * and large enough that extracting 'size' bytes won't change which
728 	 * freelist it's on.  In this case we can avoid a *lot* of work.
729 	 * Instead of the normal vmem_seg_alloc(), we just advance the start
730 	 * address of the victim segment.  Instead of moving the rotor, we
731 	 * create the new segment structure *behind the rotor*, which has
732 	 * the same effect.  And finally, we know we don't have to coalesce
733 	 * the rotor's neighbors because the new segment lies between them.
734 	 */
735 	rotor = &vmp->vm_rotor;
736 	vsp = rotor->vs_anext;
737 	if (vsp->vs_type == VMEM_FREE && (vs_size = VS_SIZE(vsp)) > realsize &&
738 	    P2SAMEHIGHBIT(vs_size, vs_size - realsize)) {
739 		ASSERT(highbit(vs_size) == highbit(vs_size - realsize));
740 		addr = vsp->vs_start;
741 		vsp->vs_start = addr + realsize;
742 		vmem_hash_insert(vmp,
743 		    vmem_seg_create(vmp, rotor->vs_aprev, addr, addr + size));
744 		(void) mutex_unlock(&vmp->vm_lock);
745 		return ((void *)addr);
746 	}
747 
748 	/*
749 	 * Starting at the rotor, look for a segment large enough to
750 	 * satisfy the allocation.
751 	 */
752 	for (;;) {
753 		vmp->vm_kstat.vk_search++;
754 		if (vsp->vs_type == VMEM_FREE && VS_SIZE(vsp) >= size)
755 			break;
756 		vsp = vsp->vs_anext;
757 		if (vsp == rotor) {
758 			int cancel_state;
759 
760 			/*
761 			 * We've come full circle.  One possibility is that the
762 			 * there's actually enough space, but the rotor itself
763 			 * is preventing the allocation from succeeding because
764 			 * it's sitting between two free segments.  Therefore,
765 			 * we advance the rotor and see if that liberates a
766 			 * suitable segment.
767 			 */
768 			vmem_advance(vmp, rotor, rotor->vs_anext);
769 			vsp = rotor->vs_aprev;
770 			if (vsp->vs_type == VMEM_FREE && VS_SIZE(vsp) >= size)
771 				break;
772 			/*
773 			 * If there's a lower arena we can import from, or it's
774 			 * a VM_NOSLEEP allocation, let vmem_xalloc() handle it.
775 			 * Otherwise, wait until another thread frees something.
776 			 */
777 			if (vmp->vm_source_alloc != NULL ||
778 			    (vmflag & VM_NOSLEEP)) {
779 				(void) mutex_unlock(&vmp->vm_lock);
780 				return (vmem_xalloc(vmp, size, vmp->vm_quantum,
781 				    0, 0, NULL, NULL, vmflag & VM_UMFLAGS));
782 			}
783 			vmp->vm_kstat.vk_wait++;
784 			(void) pthread_setcancelstate(PTHREAD_CANCEL_DISABLE,
785 			    &cancel_state);
786 			(void) cond_wait(&vmp->vm_cv, &vmp->vm_lock);
787 			(void) pthread_setcancelstate(cancel_state, NULL);
788 			vsp = rotor->vs_anext;
789 		}
790 	}
791 
792 	/*
793 	 * We found a segment.  Extract enough space to satisfy the allocation.
794 	 */
795 	addr = vsp->vs_start;
796 	vsp = vmem_seg_alloc(vmp, vsp, addr, size);
797 	ASSERT(vsp->vs_type == VMEM_ALLOC &&
798 	    vsp->vs_start == addr && vsp->vs_end == addr + size);
799 
800 	/*
801 	 * Advance the rotor to right after the newly-allocated segment.
802 	 * That's where the next VM_NEXTFIT allocation will begin searching.
803 	 */
804 	vmem_advance(vmp, rotor, vsp);
805 	(void) mutex_unlock(&vmp->vm_lock);
806 	return ((void *)addr);
807 }
808 
809 /*
810  * Allocate size bytes at offset phase from an align boundary such that the
811  * resulting segment [addr, addr + size) is a subset of [minaddr, maxaddr)
812  * that does not straddle a nocross-aligned boundary.
813  */
814 void *
815 vmem_xalloc(vmem_t *vmp, size_t size, size_t align, size_t phase,
816 	size_t nocross, void *minaddr, void *maxaddr, int vmflag)
817 {
818 	vmem_seg_t *vsp;
819 	vmem_seg_t *vbest = NULL;
820 	uintptr_t addr, taddr, start, end;
821 	void *vaddr;
822 	int hb, flist, resv;
823 	uint32_t mtbf;
824 
825 	if (phase > 0 && phase >= align)
826 		umem_panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
827 		    "invalid phase",
828 		    (void *)vmp, size, align, phase, nocross,
829 		    minaddr, maxaddr, vmflag);
830 
831 	if (align == 0)
832 		align = vmp->vm_quantum;
833 
834 	if ((align | phase | nocross) & (vmp->vm_quantum - 1)) {
835 		umem_panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
836 		    "parameters not vm_quantum aligned",
837 		    (void *)vmp, size, align, phase, nocross,
838 		    minaddr, maxaddr, vmflag);
839 	}
840 
841 	if (nocross != 0 &&
842 	    (align > nocross || P2ROUNDUP(phase + size, align) > nocross)) {
843 		umem_panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
844 		    "overconstrained allocation",
845 		    (void *)vmp, size, align, phase, nocross,
846 		    minaddr, maxaddr, vmflag);
847 	}
848 
849 	if ((mtbf = vmem_mtbf | vmp->vm_mtbf) != 0 && gethrtime() % mtbf == 0 &&
850 	    (vmflag & (VM_NOSLEEP | VM_PANIC)) == VM_NOSLEEP)
851 		return (NULL);
852 
853 	(void) mutex_lock(&vmp->vm_lock);
854 	for (;;) {
855 		int cancel_state;
856 
857 		if (vmp->vm_nsegfree < VMEM_MINFREE &&
858 		    !vmem_populate(vmp, vmflag))
859 			break;
860 
861 		/*
862 		 * highbit() returns the highest bit + 1, which is exactly
863 		 * what we want: we want to search the first freelist whose
864 		 * members are *definitely* large enough to satisfy our
865 		 * allocation.  However, there are certain cases in which we
866 		 * want to look at the next-smallest freelist (which *might*
867 		 * be able to satisfy the allocation):
868 		 *
869 		 * (1)	The size is exactly a power of 2, in which case
870 		 *	the smaller freelist is always big enough;
871 		 *
872 		 * (2)	All other freelists are empty;
873 		 *
874 		 * (3)	We're in the highest possible freelist, which is
875 		 *	always empty (e.g. the 4GB freelist on 32-bit systems);
876 		 *
877 		 * (4)	We're doing a best-fit or first-fit allocation.
878 		 */
879 		if ((size & (size - 1)) == 0) {
880 			flist = lowbit(P2ALIGN(vmp->vm_freemap, size));
881 		} else {
882 			hb = highbit(size);
883 			if ((vmp->vm_freemap >> hb) == 0 ||
884 			    hb == VMEM_FREELISTS ||
885 			    (vmflag & (VM_BESTFIT | VM_FIRSTFIT)))
886 				hb--;
887 			flist = lowbit(P2ALIGN(vmp->vm_freemap, 1UL << hb));
888 		}
889 
890 		for (vbest = NULL, vsp = (flist == 0) ? NULL :
891 		    vmp->vm_freelist[flist - 1].vs_knext;
892 		    vsp != NULL; vsp = vsp->vs_knext) {
893 			vmp->vm_kstat.vk_search++;
894 			if (vsp->vs_start == 0) {
895 				/*
896 				 * We're moving up to a larger freelist,
897 				 * so if we've already found a candidate,
898 				 * the fit can't possibly get any better.
899 				 */
900 				if (vbest != NULL)
901 					break;
902 				/*
903 				 * Find the next non-empty freelist.
904 				 */
905 				flist = lowbit(P2ALIGN(vmp->vm_freemap,
906 				    VS_SIZE(vsp)));
907 				if (flist-- == 0)
908 					break;
909 				vsp = (vmem_seg_t *)&vmp->vm_freelist[flist];
910 				ASSERT(vsp->vs_knext->vs_type == VMEM_FREE);
911 				continue;
912 			}
913 			if (vsp->vs_end - 1 < (uintptr_t)minaddr)
914 				continue;
915 			if (vsp->vs_start > (uintptr_t)maxaddr - 1)
916 				continue;
917 			start = MAX(vsp->vs_start, (uintptr_t)minaddr);
918 			end = MIN(vsp->vs_end - 1, (uintptr_t)maxaddr - 1) + 1;
919 			taddr = P2PHASEUP(start, align, phase);
920 			if (P2BOUNDARY(taddr, size, nocross))
921 				taddr +=
922 				    P2ROUNDUP(P2NPHASE(taddr, nocross), align);
923 			if ((taddr - start) + size > end - start ||
924 			    (vbest != NULL && VS_SIZE(vsp) >= VS_SIZE(vbest)))
925 				continue;
926 			vbest = vsp;
927 			addr = taddr;
928 			if (!(vmflag & VM_BESTFIT) || VS_SIZE(vbest) == size)
929 				break;
930 		}
931 		if (vbest != NULL)
932 			break;
933 		if (size == 0)
934 			umem_panic("vmem_xalloc(): size == 0");
935 		if (vmp->vm_source_alloc != NULL && nocross == 0 &&
936 		    minaddr == NULL && maxaddr == NULL) {
937 			size_t asize = P2ROUNDUP(size + phase,
938 			    MAX(align, vmp->vm_source->vm_quantum));
939 			if (asize < size) {		/* overflow */
940 				(void) mutex_unlock(&vmp->vm_lock);
941 				if (vmflag & VM_NOSLEEP)
942 					return (NULL);
943 
944 				umem_panic("vmem_xalloc(): "
945 				    "overflow on VM_SLEEP allocation");
946 			}
947 			/*
948 			 * Determine how many segment structures we'll consume.
949 			 * The calculation must be presise because if we're
950 			 * here on behalf of vmem_populate(), we are taking
951 			 * segments from a very limited reserve.
952 			 */
953 			resv = (size == asize) ?
954 			    VMEM_SEGS_PER_SPAN_CREATE +
955 			    VMEM_SEGS_PER_EXACT_ALLOC :
956 			    VMEM_SEGS_PER_ALLOC_MAX;
957 			ASSERT(vmp->vm_nsegfree >= resv);
958 			vmp->vm_nsegfree -= resv;	/* reserve our segs */
959 			(void) mutex_unlock(&vmp->vm_lock);
960 			vaddr = vmp->vm_source_alloc(vmp->vm_source, asize,
961 			    vmflag & VM_UMFLAGS);
962 			(void) mutex_lock(&vmp->vm_lock);
963 			vmp->vm_nsegfree += resv;	/* claim reservation */
964 			if (vaddr != NULL) {
965 				vbest = vmem_span_create(vmp, vaddr, asize, 1);
966 				addr = P2PHASEUP(vbest->vs_start, align, phase);
967 				break;
968 			}
969 		}
970 		(void) mutex_unlock(&vmp->vm_lock);
971 		vmem_reap();
972 		(void) mutex_lock(&vmp->vm_lock);
973 		if (vmflag & VM_NOSLEEP)
974 			break;
975 		vmp->vm_kstat.vk_wait++;
976 		(void) pthread_setcancelstate(PTHREAD_CANCEL_DISABLE,
977 		    &cancel_state);
978 		(void) cond_wait(&vmp->vm_cv, &vmp->vm_lock);
979 		(void) pthread_setcancelstate(cancel_state, NULL);
980 	}
981 	if (vbest != NULL) {
982 		ASSERT(vbest->vs_type == VMEM_FREE);
983 		ASSERT(vbest->vs_knext != vbest);
984 		(void) vmem_seg_alloc(vmp, vbest, addr, size);
985 		(void) mutex_unlock(&vmp->vm_lock);
986 		ASSERT(P2PHASE(addr, align) == phase);
987 		ASSERT(!P2BOUNDARY(addr, size, nocross));
988 		ASSERT(addr >= (uintptr_t)minaddr);
989 		ASSERT(addr + size - 1 <= (uintptr_t)maxaddr - 1);
990 		return ((void *)addr);
991 	}
992 	vmp->vm_kstat.vk_fail++;
993 	(void) mutex_unlock(&vmp->vm_lock);
994 	if (vmflag & VM_PANIC)
995 		umem_panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
996 		    "cannot satisfy mandatory allocation",
997 		    (void *)vmp, size, align, phase, nocross,
998 		    minaddr, maxaddr, vmflag);
999 	return (NULL);
1000 }
1001 
1002 /*
1003  * Free the segment [vaddr, vaddr + size), where vaddr was a constrained
1004  * allocation.  vmem_xalloc() and vmem_xfree() must always be paired because
1005  * both routines bypass the quantum caches.
1006  */
1007 void
1008 vmem_xfree(vmem_t *vmp, void *vaddr, size_t size)
1009 {
1010 	vmem_seg_t *vsp, *vnext, *vprev;
1011 
1012 	(void) mutex_lock(&vmp->vm_lock);
1013 
1014 	vsp = vmem_hash_delete(vmp, (uintptr_t)vaddr, size);
1015 	vsp->vs_end = P2ROUNDUP(vsp->vs_end, vmp->vm_quantum);
1016 
1017 	/*
1018 	 * Attempt to coalesce with the next segment.
1019 	 */
1020 	vnext = vsp->vs_anext;
1021 	if (vnext->vs_type == VMEM_FREE) {
1022 		ASSERT(vsp->vs_end == vnext->vs_start);
1023 		vmem_freelist_delete(vmp, vnext);
1024 		vsp->vs_end = vnext->vs_end;
1025 		vmem_seg_destroy(vmp, vnext);
1026 	}
1027 
1028 	/*
1029 	 * Attempt to coalesce with the previous segment.
1030 	 */
1031 	vprev = vsp->vs_aprev;
1032 	if (vprev->vs_type == VMEM_FREE) {
1033 		ASSERT(vprev->vs_end == vsp->vs_start);
1034 		vmem_freelist_delete(vmp, vprev);
1035 		vprev->vs_end = vsp->vs_end;
1036 		vmem_seg_destroy(vmp, vsp);
1037 		vsp = vprev;
1038 	}
1039 
1040 	/*
1041 	 * If the entire span is free, return it to the source.
1042 	 */
1043 	if (vsp->vs_import && vmp->vm_source_free != NULL &&
1044 	    vsp->vs_aprev->vs_type == VMEM_SPAN &&
1045 	    vsp->vs_anext->vs_type == VMEM_SPAN) {
1046 		vaddr = (void *)vsp->vs_start;
1047 		size = VS_SIZE(vsp);
1048 		ASSERT(size == VS_SIZE(vsp->vs_aprev));
1049 		vmem_span_destroy(vmp, vsp);
1050 		(void) mutex_unlock(&vmp->vm_lock);
1051 		vmp->vm_source_free(vmp->vm_source, vaddr, size);
1052 	} else {
1053 		vmem_freelist_insert(vmp, vsp);
1054 		(void) mutex_unlock(&vmp->vm_lock);
1055 	}
1056 }
1057 
1058 /*
1059  * Allocate size bytes from arena vmp.  Returns the allocated address
1060  * on success, NULL on failure.  vmflag specifies VM_SLEEP or VM_NOSLEEP,
1061  * and may also specify best-fit, first-fit, or next-fit allocation policy
1062  * instead of the default instant-fit policy.  VM_SLEEP allocations are
1063  * guaranteed to succeed.
1064  */
1065 void *
1066 vmem_alloc(vmem_t *vmp, size_t size, int vmflag)
1067 {
1068 	vmem_seg_t *vsp;
1069 	uintptr_t addr;
1070 	int hb;
1071 	int flist = 0;
1072 	uint32_t mtbf;
1073 	vmflag |= vmem_allocator;
1074 
1075 	if (size - 1 < vmp->vm_qcache_max) {
1076 		ASSERT(vmflag & VM_NOSLEEP);
1077 		return (_umem_cache_alloc(vmp->vm_qcache[(size - 1) >>
1078 		    vmp->vm_qshift], UMEM_DEFAULT));
1079 	}
1080 
1081 	if ((mtbf = vmem_mtbf | vmp->vm_mtbf) != 0 && gethrtime() % mtbf == 0 &&
1082 	    (vmflag & (VM_NOSLEEP | VM_PANIC)) == VM_NOSLEEP)
1083 		return (NULL);
1084 
1085 	if (vmflag & VM_NEXTFIT)
1086 		return (vmem_nextfit_alloc(vmp, size, vmflag));
1087 
1088 	if (vmflag & (VM_BESTFIT | VM_FIRSTFIT))
1089 		return (vmem_xalloc(vmp, size, vmp->vm_quantum, 0, 0,
1090 		    NULL, NULL, vmflag));
1091 
1092 	/*
1093 	 * Unconstrained instant-fit allocation from the segment list.
1094 	 */
1095 	(void) mutex_lock(&vmp->vm_lock);
1096 
1097 	if (vmp->vm_nsegfree >= VMEM_MINFREE || vmem_populate(vmp, vmflag)) {
1098 		if ((size & (size - 1)) == 0)
1099 			flist = lowbit(P2ALIGN(vmp->vm_freemap, size));
1100 		else if ((hb = highbit(size)) < VMEM_FREELISTS)
1101 			flist = lowbit(P2ALIGN(vmp->vm_freemap, 1UL << hb));
1102 	}
1103 
1104 	if (flist-- == 0) {
1105 		(void) mutex_unlock(&vmp->vm_lock);
1106 		return (vmem_xalloc(vmp, size, vmp->vm_quantum,
1107 		    0, 0, NULL, NULL, vmflag));
1108 	}
1109 
1110 	ASSERT(size <= (1UL << flist));
1111 	vsp = vmp->vm_freelist[flist].vs_knext;
1112 	addr = vsp->vs_start;
1113 	(void) vmem_seg_alloc(vmp, vsp, addr, size);
1114 	(void) mutex_unlock(&vmp->vm_lock);
1115 	return ((void *)addr);
1116 }
1117 
1118 /*
1119  * Free the segment [vaddr, vaddr + size).
1120  */
1121 void
1122 vmem_free(vmem_t *vmp, void *vaddr, size_t size)
1123 {
1124 	if (size - 1 < vmp->vm_qcache_max)
1125 		_umem_cache_free(vmp->vm_qcache[(size - 1) >> vmp->vm_qshift],
1126 		    vaddr);
1127 	else
1128 		vmem_xfree(vmp, vaddr, size);
1129 }
1130 
1131 /*
1132  * Determine whether arena vmp contains the segment [vaddr, vaddr + size).
1133  */
1134 int
1135 vmem_contains(vmem_t *vmp, void *vaddr, size_t size)
1136 {
1137 	uintptr_t start = (uintptr_t)vaddr;
1138 	uintptr_t end = start + size;
1139 	vmem_seg_t *vsp;
1140 	vmem_seg_t *seg0 = &vmp->vm_seg0;
1141 
1142 	(void) mutex_lock(&vmp->vm_lock);
1143 	vmp->vm_kstat.vk_contains++;
1144 	for (vsp = seg0->vs_knext; vsp != seg0; vsp = vsp->vs_knext) {
1145 		vmp->vm_kstat.vk_contains_search++;
1146 		ASSERT(vsp->vs_type == VMEM_SPAN);
1147 		if (start >= vsp->vs_start && end - 1 <= vsp->vs_end - 1)
1148 			break;
1149 	}
1150 	(void) mutex_unlock(&vmp->vm_lock);
1151 	return (vsp != seg0);
1152 }
1153 
1154 /*
1155  * Add the span [vaddr, vaddr + size) to arena vmp.
1156  */
1157 void *
1158 vmem_add(vmem_t *vmp, void *vaddr, size_t size, int vmflag)
1159 {
1160 	if (vaddr == NULL || size == 0) {
1161 		umem_panic("vmem_add(%p, %p, %lu): bad arguments",
1162 		    vmp, vaddr, size);
1163 	}
1164 
1165 	ASSERT(!vmem_contains(vmp, vaddr, size));
1166 
1167 	(void) mutex_lock(&vmp->vm_lock);
1168 	if (vmem_populate(vmp, vmflag))
1169 		(void) vmem_span_create(vmp, vaddr, size, 0);
1170 	else
1171 		vaddr = NULL;
1172 	(void) cond_broadcast(&vmp->vm_cv);
1173 	(void) mutex_unlock(&vmp->vm_lock);
1174 	return (vaddr);
1175 }
1176 
1177 /*
1178  * Adds the address range [addr, endaddr) to arena vmp, by either:
1179  *   1. joining two existing spans, [x, addr), and [endaddr, y) (which
1180  *      are in that order) into a single [x, y) span,
1181  *   2. expanding an existing [x, addr) span to [x, endaddr),
1182  *   3. expanding an existing [endaddr, x) span to [addr, x), or
1183  *   4. creating a new [addr, endaddr) span.
1184  *
1185  * Called with vmp->vm_lock held, and a successful vmem_populate() completed.
1186  * Cannot fail.  Returns the new segment.
1187  *
1188  * NOTE:  this algorithm is linear-time in the number of spans, but is
1189  *      constant-time when you are extending the last (highest-addressed)
1190  *      span.
1191  */
1192 static vmem_seg_t *
1193 vmem_extend_unlocked(vmem_t *vmp, uintptr_t addr, uintptr_t endaddr)
1194 {
1195 	vmem_seg_t *span;
1196 	vmem_seg_t *vsp;
1197 
1198 	vmem_seg_t *end = &vmp->vm_seg0;
1199 
1200 	ASSERT(MUTEX_HELD(&vmp->vm_lock));
1201 
1202 	/*
1203 	 * the second "if" clause below relies on the direction of this search
1204 	 */
1205 	for (span = end->vs_kprev; span != end; span = span->vs_kprev) {
1206 		if (span->vs_end == addr || span->vs_start == endaddr)
1207 			break;
1208 	}
1209 
1210 	if (span == end)
1211 		return (vmem_span_create(vmp, (void *)addr, endaddr - addr, 0));
1212 	if (span->vs_kprev->vs_end == addr && span->vs_start == endaddr) {
1213 		vmem_seg_t *prevspan = span->vs_kprev;
1214 		vmem_seg_t *nextseg = span->vs_anext;
1215 		vmem_seg_t *prevseg = span->vs_aprev;
1216 
1217 		/*
1218 		 * prevspan becomes the span marker for the full range
1219 		 */
1220 		prevspan->vs_end = span->vs_end;
1221 
1222 		/*
1223 		 * Notionally, span becomes a free segment representing
1224 		 * [addr, endaddr).
1225 		 *
1226 		 * However, if either of its neighbors are free, we coalesce
1227 		 * by destroying span and changing the free segment.
1228 		 */
1229 		if (prevseg->vs_type == VMEM_FREE &&
1230 		    nextseg->vs_type == VMEM_FREE) {
1231 			/*
1232 			 * coalesce both ways
1233 			 */
1234 			ASSERT(prevseg->vs_end == addr &&
1235 			    nextseg->vs_start == endaddr);
1236 
1237 			vmem_freelist_delete(vmp, prevseg);
1238 			prevseg->vs_end = nextseg->vs_end;
1239 
1240 			vmem_freelist_delete(vmp, nextseg);
1241 			VMEM_DELETE(span, k);
1242 			vmem_seg_destroy(vmp, nextseg);
1243 			vmem_seg_destroy(vmp, span);
1244 
1245 			vsp = prevseg;
1246 		} else if (prevseg->vs_type == VMEM_FREE) {
1247 			/*
1248 			 * coalesce left
1249 			 */
1250 			ASSERT(prevseg->vs_end == addr);
1251 
1252 			VMEM_DELETE(span, k);
1253 			vmem_seg_destroy(vmp, span);
1254 
1255 			vmem_freelist_delete(vmp, prevseg);
1256 			prevseg->vs_end = endaddr;
1257 
1258 			vsp = prevseg;
1259 		} else if (nextseg->vs_type == VMEM_FREE) {
1260 			/*
1261 			 * coalesce right
1262 			 */
1263 			ASSERT(nextseg->vs_start == endaddr);
1264 
1265 			VMEM_DELETE(span, k);
1266 			vmem_seg_destroy(vmp, span);
1267 
1268 			vmem_freelist_delete(vmp, nextseg);
1269 			nextseg->vs_start = addr;
1270 
1271 			vsp = nextseg;
1272 		} else {
1273 			/*
1274 			 * cannnot coalesce
1275 			 */
1276 			VMEM_DELETE(span, k);
1277 			span->vs_start = addr;
1278 			span->vs_end = endaddr;
1279 
1280 			vsp = span;
1281 		}
1282 	} else if (span->vs_end == addr) {
1283 		vmem_seg_t *oldseg = span->vs_knext->vs_aprev;
1284 		span->vs_end = endaddr;
1285 
1286 		ASSERT(oldseg->vs_type != VMEM_SPAN);
1287 		if (oldseg->vs_type == VMEM_FREE) {
1288 			ASSERT(oldseg->vs_end == addr);
1289 			vmem_freelist_delete(vmp, oldseg);
1290 			oldseg->vs_end = endaddr;
1291 			vsp = oldseg;
1292 		} else
1293 			vsp = vmem_seg_create(vmp, oldseg, addr, endaddr);
1294 	} else {
1295 		vmem_seg_t *oldseg = span->vs_anext;
1296 		ASSERT(span->vs_start == endaddr);
1297 		span->vs_start = addr;
1298 
1299 		ASSERT(oldseg->vs_type != VMEM_SPAN);
1300 		if (oldseg->vs_type == VMEM_FREE) {
1301 			ASSERT(oldseg->vs_start == endaddr);
1302 			vmem_freelist_delete(vmp, oldseg);
1303 			oldseg->vs_start = addr;
1304 			vsp = oldseg;
1305 		} else
1306 			vsp = vmem_seg_create(vmp, span, addr, endaddr);
1307 	}
1308 	vmem_freelist_insert(vmp, vsp);
1309 	vmp->vm_kstat.vk_mem_total += (endaddr - addr);
1310 	return (vsp);
1311 }
1312 
1313 /*
1314  * Does some error checking, calls vmem_extend_unlocked to add
1315  * [vaddr, vaddr+size) to vmp, then allocates alloc bytes from the
1316  * newly merged segment.
1317  */
1318 void *
1319 _vmem_extend_alloc(vmem_t *vmp, void *vaddr, size_t size, size_t alloc,
1320     int vmflag)
1321 {
1322 	uintptr_t addr = (uintptr_t)vaddr;
1323 	uintptr_t endaddr = addr + size;
1324 	vmem_seg_t *vsp;
1325 
1326 	ASSERT(vaddr != NULL && size != 0 && endaddr > addr);
1327 	ASSERT(alloc <= size && alloc != 0);
1328 	ASSERT(((addr | size | alloc) & (vmp->vm_quantum - 1)) == 0);
1329 
1330 	ASSERT(!vmem_contains(vmp, vaddr, size));
1331 
1332 	(void) mutex_lock(&vmp->vm_lock);
1333 	if (!vmem_populate(vmp, vmflag)) {
1334 		(void) mutex_unlock(&vmp->vm_lock);
1335 		return (NULL);
1336 	}
1337 	/*
1338 	 * if there is a source, we can't mess with the spans
1339 	 */
1340 	if (vmp->vm_source_alloc != NULL)
1341 		vsp = vmem_span_create(vmp, vaddr, size, 0);
1342 	else
1343 		vsp = vmem_extend_unlocked(vmp, addr, endaddr);
1344 
1345 	ASSERT(VS_SIZE(vsp) >= alloc);
1346 
1347 	addr = vsp->vs_start;
1348 	(void) vmem_seg_alloc(vmp, vsp, addr, alloc);
1349 	vaddr = (void *)addr;
1350 
1351 	(void) cond_broadcast(&vmp->vm_cv);
1352 	(void) mutex_unlock(&vmp->vm_lock);
1353 
1354 	return (vaddr);
1355 }
1356 
1357 /*
1358  * Walk the vmp arena, applying func to each segment matching typemask.
1359  * If VMEM_REENTRANT is specified, the arena lock is dropped across each
1360  * call to func(); otherwise, it is held for the duration of vmem_walk()
1361  * to ensure a consistent snapshot.  Note that VMEM_REENTRANT callbacks
1362  * are *not* necessarily consistent, so they may only be used when a hint
1363  * is adequate.
1364  */
1365 void
1366 vmem_walk(vmem_t *vmp, int typemask,
1367 	void (*func)(void *, void *, size_t), void *arg)
1368 {
1369 	vmem_seg_t *vsp;
1370 	vmem_seg_t *seg0 = &vmp->vm_seg0;
1371 	vmem_seg_t walker;
1372 
1373 	if (typemask & VMEM_WALKER)
1374 		return;
1375 
1376 	bzero(&walker, sizeof (walker));
1377 	walker.vs_type = VMEM_WALKER;
1378 
1379 	(void) mutex_lock(&vmp->vm_lock);
1380 	VMEM_INSERT(seg0, &walker, a);
1381 	for (vsp = seg0->vs_anext; vsp != seg0; vsp = vsp->vs_anext) {
1382 		if (vsp->vs_type & typemask) {
1383 			void *start = (void *)vsp->vs_start;
1384 			size_t size = VS_SIZE(vsp);
1385 			if (typemask & VMEM_REENTRANT) {
1386 				vmem_advance(vmp, &walker, vsp);
1387 				(void) mutex_unlock(&vmp->vm_lock);
1388 				func(arg, start, size);
1389 				(void) mutex_lock(&vmp->vm_lock);
1390 				vsp = &walker;
1391 			} else {
1392 				func(arg, start, size);
1393 			}
1394 		}
1395 	}
1396 	vmem_advance(vmp, &walker, NULL);
1397 	(void) mutex_unlock(&vmp->vm_lock);
1398 }
1399 
1400 /*
1401  * Return the total amount of memory whose type matches typemask.  Thus:
1402  *
1403  *	typemask VMEM_ALLOC yields total memory allocated (in use).
1404  *	typemask VMEM_FREE yields total memory free (available).
1405  *	typemask (VMEM_ALLOC | VMEM_FREE) yields total arena size.
1406  */
1407 size_t
1408 vmem_size(vmem_t *vmp, int typemask)
1409 {
1410 	uint64_t size = 0;
1411 
1412 	if (typemask & VMEM_ALLOC)
1413 		size += vmp->vm_kstat.vk_mem_inuse;
1414 	if (typemask & VMEM_FREE)
1415 		size += vmp->vm_kstat.vk_mem_total -
1416 		    vmp->vm_kstat.vk_mem_inuse;
1417 	return ((size_t)size);
1418 }
1419 
1420 /*
1421  * Create an arena called name whose initial span is [base, base + size).
1422  * The arena's natural unit of currency is quantum, so vmem_alloc()
1423  * guarantees quantum-aligned results.  The arena may import new spans
1424  * by invoking afunc() on source, and may return those spans by invoking
1425  * ffunc() on source.  To make small allocations fast and scalable,
1426  * the arena offers high-performance caching for each integer multiple
1427  * of quantum up to qcache_max.
1428  */
1429 vmem_t *
1430 vmem_create(const char *name, void *base, size_t size, size_t quantum,
1431 	vmem_alloc_t *afunc, vmem_free_t *ffunc, vmem_t *source,
1432 	size_t qcache_max, int vmflag)
1433 {
1434 	int i;
1435 	size_t nqcache;
1436 	vmem_t *vmp, *cur, **vmpp;
1437 	vmem_seg_t *vsp;
1438 	vmem_freelist_t *vfp;
1439 	uint32_t id = atomic_add_32_nv(&vmem_id, 1);
1440 
1441 	if (vmem_vmem_arena != NULL) {
1442 		vmp = vmem_alloc(vmem_vmem_arena, sizeof (vmem_t),
1443 		    vmflag & VM_UMFLAGS);
1444 	} else {
1445 		ASSERT(id <= VMEM_INITIAL);
1446 		vmp = &vmem0[id - 1];
1447 	}
1448 
1449 	if (vmp == NULL)
1450 		return (NULL);
1451 	bzero(vmp, sizeof (vmem_t));
1452 
1453 	(void) snprintf(vmp->vm_name, VMEM_NAMELEN, "%s", name);
1454 	(void) mutex_init(&vmp->vm_lock, USYNC_THREAD, NULL);
1455 	(void) cond_init(&vmp->vm_cv, USYNC_THREAD, NULL);
1456 	vmp->vm_cflags = vmflag;
1457 	vmflag &= VM_UMFLAGS;
1458 
1459 	vmp->vm_quantum = quantum;
1460 	vmp->vm_qshift = highbit(quantum) - 1;
1461 	nqcache = MIN(qcache_max >> vmp->vm_qshift, VMEM_NQCACHE_MAX);
1462 
1463 	for (i = 0; i <= VMEM_FREELISTS; i++) {
1464 		vfp = &vmp->vm_freelist[i];
1465 		vfp->vs_end = 1UL << i;
1466 		vfp->vs_knext = (vmem_seg_t *)(vfp + 1);
1467 		vfp->vs_kprev = (vmem_seg_t *)(vfp - 1);
1468 	}
1469 
1470 	vmp->vm_freelist[0].vs_kprev = NULL;
1471 	vmp->vm_freelist[VMEM_FREELISTS].vs_knext = NULL;
1472 	vmp->vm_freelist[VMEM_FREELISTS].vs_end = 0;
1473 	vmp->vm_hash_table = vmp->vm_hash0;
1474 	vmp->vm_hash_mask = VMEM_HASH_INITIAL - 1;
1475 	vmp->vm_hash_shift = highbit(vmp->vm_hash_mask);
1476 
1477 	vsp = &vmp->vm_seg0;
1478 	vsp->vs_anext = vsp;
1479 	vsp->vs_aprev = vsp;
1480 	vsp->vs_knext = vsp;
1481 	vsp->vs_kprev = vsp;
1482 	vsp->vs_type = VMEM_SPAN;
1483 
1484 	vsp = &vmp->vm_rotor;
1485 	vsp->vs_type = VMEM_ROTOR;
1486 	VMEM_INSERT(&vmp->vm_seg0, vsp, a);
1487 
1488 	vmp->vm_id = id;
1489 	if (source != NULL)
1490 		vmp->vm_kstat.vk_source_id = source->vm_id;
1491 	vmp->vm_source = source;
1492 	vmp->vm_source_alloc = afunc;
1493 	vmp->vm_source_free = ffunc;
1494 
1495 	if (nqcache != 0) {
1496 		vmp->vm_qcache_max = nqcache << vmp->vm_qshift;
1497 		for (i = 0; i < nqcache; i++) {
1498 			char buf[VMEM_NAMELEN + 21];
1499 			(void) snprintf(buf, sizeof (buf), "%s_%lu",
1500 			    vmp->vm_name, (long)((i + 1) * quantum));
1501 			vmp->vm_qcache[i] = umem_cache_create(buf,
1502 			    (i + 1) * quantum, quantum, NULL, NULL, NULL,
1503 			    NULL, vmp, UMC_QCACHE | UMC_NOTOUCH);
1504 			if (vmp->vm_qcache[i] == NULL) {
1505 				vmp->vm_qcache_max = i * quantum;
1506 				break;
1507 			}
1508 		}
1509 	}
1510 
1511 	(void) mutex_lock(&vmem_list_lock);
1512 	vmpp = &vmem_list;
1513 	while ((cur = *vmpp) != NULL)
1514 		vmpp = &cur->vm_next;
1515 	*vmpp = vmp;
1516 	(void) mutex_unlock(&vmem_list_lock);
1517 
1518 	if (vmp->vm_cflags & VMC_POPULATOR) {
1519 		uint_t pop_id = atomic_add_32_nv(&vmem_populators, 1);
1520 		ASSERT(pop_id <= VMEM_INITIAL);
1521 		vmem_populator[pop_id - 1] = vmp;
1522 		(void) mutex_lock(&vmp->vm_lock);
1523 		(void) vmem_populate(vmp, vmflag | VM_PANIC);
1524 		(void) mutex_unlock(&vmp->vm_lock);
1525 	}
1526 
1527 	if ((base || size) && vmem_add(vmp, base, size, vmflag) == NULL) {
1528 		vmem_destroy(vmp);
1529 		return (NULL);
1530 	}
1531 
1532 	return (vmp);
1533 }
1534 
1535 /*
1536  * Destroy arena vmp.
1537  */
1538 void
1539 vmem_destroy(vmem_t *vmp)
1540 {
1541 	vmem_t *cur, **vmpp;
1542 	vmem_seg_t *seg0 = &vmp->vm_seg0;
1543 	vmem_seg_t *vsp;
1544 	size_t leaked;
1545 	int i;
1546 
1547 	(void) mutex_lock(&vmem_list_lock);
1548 	vmpp = &vmem_list;
1549 	while ((cur = *vmpp) != vmp)
1550 		vmpp = &cur->vm_next;
1551 	*vmpp = vmp->vm_next;
1552 	(void) mutex_unlock(&vmem_list_lock);
1553 
1554 	for (i = 0; i < VMEM_NQCACHE_MAX; i++)
1555 		if (vmp->vm_qcache[i])
1556 			umem_cache_destroy(vmp->vm_qcache[i]);
1557 
1558 	leaked = vmem_size(vmp, VMEM_ALLOC);
1559 	if (leaked != 0)
1560 		umem_printf("vmem_destroy('%s'): leaked %lu bytes",
1561 		    vmp->vm_name, leaked);
1562 
1563 	if (vmp->vm_hash_table != vmp->vm_hash0)
1564 		vmem_free(vmem_hash_arena, vmp->vm_hash_table,
1565 		    (vmp->vm_hash_mask + 1) * sizeof (void *));
1566 
1567 	/*
1568 	 * Give back the segment structures for anything that's left in the
1569 	 * arena, e.g. the primary spans and their free segments.
1570 	 */
1571 	VMEM_DELETE(&vmp->vm_rotor, a);
1572 	for (vsp = seg0->vs_anext; vsp != seg0; vsp = vsp->vs_anext)
1573 		vmem_putseg_global(vsp);
1574 
1575 	while (vmp->vm_nsegfree > 0)
1576 		vmem_putseg_global(vmem_getseg(vmp));
1577 
1578 	(void) mutex_destroy(&vmp->vm_lock);
1579 	(void) cond_destroy(&vmp->vm_cv);
1580 	vmem_free(vmem_vmem_arena, vmp, sizeof (vmem_t));
1581 }
1582 
1583 /*
1584  * Resize vmp's hash table to keep the average lookup depth near 1.0.
1585  */
1586 static void
1587 vmem_hash_rescale(vmem_t *vmp)
1588 {
1589 	vmem_seg_t **old_table, **new_table, *vsp;
1590 	size_t old_size, new_size, h, nseg;
1591 
1592 	nseg = (size_t)(vmp->vm_kstat.vk_alloc - vmp->vm_kstat.vk_free);
1593 
1594 	new_size = MAX(VMEM_HASH_INITIAL, 1 << (highbit(3 * nseg + 4) - 2));
1595 	old_size = vmp->vm_hash_mask + 1;
1596 
1597 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
1598 		return;
1599 
1600 	new_table = vmem_alloc(vmem_hash_arena, new_size * sizeof (void *),
1601 	    VM_NOSLEEP);
1602 	if (new_table == NULL)
1603 		return;
1604 	bzero(new_table, new_size * sizeof (void *));
1605 
1606 	(void) mutex_lock(&vmp->vm_lock);
1607 
1608 	old_size = vmp->vm_hash_mask + 1;
1609 	old_table = vmp->vm_hash_table;
1610 
1611 	vmp->vm_hash_mask = new_size - 1;
1612 	vmp->vm_hash_table = new_table;
1613 	vmp->vm_hash_shift = highbit(vmp->vm_hash_mask);
1614 
1615 	for (h = 0; h < old_size; h++) {
1616 		vsp = old_table[h];
1617 		while (vsp != NULL) {
1618 			uintptr_t addr = vsp->vs_start;
1619 			vmem_seg_t *next_vsp = vsp->vs_knext;
1620 			vmem_seg_t **hash_bucket = VMEM_HASH(vmp, addr);
1621 			vsp->vs_knext = *hash_bucket;
1622 			*hash_bucket = vsp;
1623 			vsp = next_vsp;
1624 		}
1625 	}
1626 
1627 	(void) mutex_unlock(&vmp->vm_lock);
1628 
1629 	if (old_table != vmp->vm_hash0)
1630 		vmem_free(vmem_hash_arena, old_table,
1631 		    old_size * sizeof (void *));
1632 }
1633 
1634 /*
1635  * Perform periodic maintenance on all vmem arenas.
1636  */
1637 /*ARGSUSED*/
1638 void
1639 vmem_update(void *dummy)
1640 {
1641 	vmem_t *vmp;
1642 
1643 	(void) mutex_lock(&vmem_list_lock);
1644 	for (vmp = vmem_list; vmp != NULL; vmp = vmp->vm_next) {
1645 		/*
1646 		 * If threads are waiting for resources, wake them up
1647 		 * periodically so they can issue another vmem_reap()
1648 		 * to reclaim resources cached by the slab allocator.
1649 		 */
1650 		(void) cond_broadcast(&vmp->vm_cv);
1651 
1652 		/*
1653 		 * Rescale the hash table to keep the hash chains short.
1654 		 */
1655 		vmem_hash_rescale(vmp);
1656 	}
1657 	(void) mutex_unlock(&vmem_list_lock);
1658 }
1659 
1660 /*
1661  * If vmem_init is called again, we need to be able to reset the world.
1662  * That includes resetting the statics back to their original values.
1663  */
1664 void
1665 vmem_startup(void)
1666 {
1667 #ifdef UMEM_STANDALONE
1668 	vmem_id = 0;
1669 	vmem_populators = 0;
1670 	vmem_segfree = NULL;
1671 	vmem_list = NULL;
1672 	vmem_internal_arena = NULL;
1673 	vmem_seg_arena = NULL;
1674 	vmem_hash_arena = NULL;
1675 	vmem_vmem_arena = NULL;
1676 	vmem_heap = NULL;
1677 	vmem_heap_alloc = NULL;
1678 	vmem_heap_free = NULL;
1679 
1680 	bzero(vmem0, sizeof (vmem0));
1681 	bzero(vmem_populator, sizeof (vmem_populator));
1682 	bzero(vmem_seg0, sizeof (vmem_seg0));
1683 #endif
1684 }
1685 
1686 /*
1687  * Prepare vmem for use.
1688  */
1689 vmem_t *
1690 vmem_init(const char *parent_name, size_t parent_quantum,
1691     vmem_alloc_t *parent_alloc, vmem_free_t *parent_free,
1692     const char *heap_name, void *heap_start, size_t heap_size,
1693     size_t heap_quantum, vmem_alloc_t *heap_alloc, vmem_free_t *heap_free)
1694 {
1695 	uint32_t id;
1696 	int nseg = VMEM_SEG_INITIAL;
1697 	vmem_t *parent, *heap;
1698 
1699 	ASSERT(vmem_internal_arena == NULL);
1700 
1701 	while (--nseg >= 0)
1702 		vmem_putseg_global(&vmem_seg0[nseg]);
1703 
1704 	if (parent_name != NULL) {
1705 		parent = vmem_create(parent_name,
1706 		    heap_start, heap_size, parent_quantum,
1707 		    NULL, NULL, NULL, 0,
1708 		    VM_SLEEP | VMC_POPULATOR);
1709 		heap_start = NULL;
1710 		heap_size = 0;
1711 	} else {
1712 		ASSERT(parent_alloc == NULL && parent_free == NULL);
1713 		parent = NULL;
1714 	}
1715 
1716 	heap = vmem_create(heap_name,
1717 	    heap_start, heap_size, heap_quantum,
1718 	    parent_alloc, parent_free, parent, 0,
1719 	    VM_SLEEP | VMC_POPULATOR);
1720 
1721 	vmem_heap = heap;
1722 	vmem_heap_alloc = heap_alloc;
1723 	vmem_heap_free = heap_free;
1724 
1725 	vmem_internal_arena = vmem_create("vmem_internal",
1726 	    NULL, 0, heap_quantum,
1727 	    heap_alloc, heap_free, heap, 0,
1728 	    VM_SLEEP | VMC_POPULATOR);
1729 
1730 	vmem_seg_arena = vmem_create("vmem_seg",
1731 	    NULL, 0, heap_quantum,
1732 	    vmem_alloc, vmem_free, vmem_internal_arena, 0,
1733 	    VM_SLEEP | VMC_POPULATOR);
1734 
1735 	vmem_hash_arena = vmem_create("vmem_hash",
1736 	    NULL, 0, 8,
1737 	    vmem_alloc, vmem_free, vmem_internal_arena, 0,
1738 	    VM_SLEEP);
1739 
1740 	vmem_vmem_arena = vmem_create("vmem_vmem",
1741 	    vmem0, sizeof (vmem0), 1,
1742 	    vmem_alloc, vmem_free, vmem_internal_arena, 0,
1743 	    VM_SLEEP);
1744 
1745 	for (id = 0; id < vmem_id; id++)
1746 		(void) vmem_xalloc(vmem_vmem_arena, sizeof (vmem_t),
1747 		    1, 0, 0, &vmem0[id], &vmem0[id + 1],
1748 		    VM_NOSLEEP | VM_BESTFIT | VM_PANIC);
1749 
1750 	return (heap);
1751 }
1752 
1753 void
1754 vmem_no_debug(void)
1755 {
1756 	/*
1757 	 * This size must be a multiple of the minimum required alignment,
1758 	 * since vmem_populate allocates them compactly.
1759 	 */
1760 	vmem_seg_size = P2ROUNDUP(offsetof(vmem_seg_t, vs_thread),
1761 	    sizeof (hrtime_t));
1762 }
1763 
1764 /*
1765  * Lockup and release, for fork1(2) handling.
1766  */
1767 void
1768 vmem_lockup(void)
1769 {
1770 	vmem_t *cur;
1771 
1772 	(void) mutex_lock(&vmem_list_lock);
1773 	(void) mutex_lock(&vmem_nosleep_lock.vmpl_mutex);
1774 
1775 	/*
1776 	 * Lock up and broadcast all arenas.
1777 	 */
1778 	for (cur = vmem_list; cur != NULL; cur = cur->vm_next) {
1779 		(void) mutex_lock(&cur->vm_lock);
1780 		(void) cond_broadcast(&cur->vm_cv);
1781 	}
1782 
1783 	(void) mutex_lock(&vmem_segfree_lock);
1784 }
1785 
1786 void
1787 vmem_release(void)
1788 {
1789 	vmem_t *cur;
1790 
1791 	(void) mutex_unlock(&vmem_nosleep_lock.vmpl_mutex);
1792 
1793 	for (cur = vmem_list; cur != NULL; cur = cur->vm_next)
1794 		(void) mutex_unlock(&cur->vm_lock);
1795 
1796 	(void) mutex_unlock(&vmem_segfree_lock);
1797 	(void) mutex_unlock(&vmem_list_lock);
1798 }
1799