xref: /illumos-gate/usr/src/lib/libumem/common/vmem.c (revision d48be21240dfd051b689384ce2b23479d757f2d8)
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 2019, Joyent, Inc.
26  * Copyright (c) 2017 by Delphix. All rights reserved.
27  */
28 
29 /*
30  * For a more complete description of the main ideas, see:
31  *
32  *	Jeff Bonwick and Jonathan Adams,
33  *
34  *	Magazines and vmem: Extending the Slab Allocator to Many CPUs and
35  *	Arbitrary Resources.
36  *
37  *	Proceedings of the 2001 Usenix Conference.
38  *	Available as /shared/sac/PSARC/2000/550/materials/vmem.pdf.
39  *
40  * For the "Big Theory Statement", see usr/src/uts/common/os/vmem.c
41  *
42  * 1. Overview of changes
43  * ------------------------------
44  * There have been a few changes to vmem in order to support umem.  The
45  * main areas are:
46  *
47  *	* VM_SLEEP unsupported
48  *
49  *	* Reaping changes
50  *
51  *	* initialization changes
52  *
53  *	* _vmem_extend_alloc
54  *
55  *
56  * 2. VM_SLEEP Removed
57  * -------------------
58  * Since VM_SLEEP allocations can hold locks (in vmem_populate()) for
59  * possibly infinite amounts of time, they are not supported in this
60  * version of vmem.  Sleep-like behavior can be achieved through
61  * UMEM_NOFAIL umem allocations.
62  *
63  *
64  * 3. Reaping changes
65  * ------------------
66  * Unlike kmem_reap(), which just asynchronously schedules work, umem_reap()
67  * can do allocations and frees synchronously.  This is a problem if it
68  * occurs during a vmem_populate() allocation.
69  *
70  * Instead, we delay reaps while populates are active.
71  *
72  *
73  * 4. Initialization changes
74  * -------------------------
75  * In the kernel, vmem_init() allows you to create a single, top-level arena,
76  * which has vmem_internal_arena as a child.  For umem, we want to be able
77  * to extend arenas dynamically.  It is much easier to support this if we
78  * allow a two-level "heap" arena:
79  *
80  *	+----------+
81  *	|  "fake"  |
82  *	+----------+
83  *	      |
84  *	+----------+
85  *	|  "heap"  |
86  *	+----------+
87  *	  |    \ \
88  *	  |     +-+-- ... <other children>
89  *	  |
90  *	+---------------+
91  *	| vmem_internal |
92  *	+---------------+
93  *	    | | | |
94  *	   <children>
95  *
96  * The new vmem_init() allows you to specify a "parent" of the heap, along
97  * with allocation functions.
98  *
99  *
100  * 5. _vmem_extend_alloc
101  * ---------------------
102  * The other part of extending is _vmem_extend_alloc.  This function allows
103  * you to extend (expand current spans, if possible) an arena and allocate
104  * a chunk of the newly extened span atomically.  This is needed to support
105  * extending the heap while vmem_populate()ing it.
106  *
107  * In order to increase the usefulness of extending, non-imported spans are
108  * sorted in address order.
109  */
110 
111 #include <sys/vmem_impl_user.h>
112 #include <alloca.h>
113 #include <sys/sysmacros.h>
114 #include <stdio.h>
115 #include <strings.h>
116 #include <atomic.h>
117 
118 #include "vmem_base.h"
119 #include "umem_base.h"
120 
121 #define	VMEM_INITIAL		6	/* early vmem arenas */
122 #define	VMEM_SEG_INITIAL	100	/* early segments */
123 
124 /*
125  * Adding a new span to an arena requires two segment structures: one to
126  * represent the span, and one to represent the free segment it contains.
127  */
128 #define	VMEM_SEGS_PER_SPAN_CREATE	2
129 
130 /*
131  * Allocating a piece of an existing segment requires 0-2 segment structures
132  * depending on how much of the segment we're allocating.
133  *
134  * To allocate the entire segment, no new segment structures are needed; we
135  * simply move the existing segment structure from the freelist to the
136  * allocation hash table.
137  *
138  * To allocate a piece from the left or right end of the segment, we must
139  * split the segment into two pieces (allocated part and remainder), so we
140  * need one new segment structure to represent the remainder.
141  *
142  * To allocate from the middle of a segment, we need two new segment strucures
143  * to represent the remainders on either side of the allocated part.
144  */
145 #define	VMEM_SEGS_PER_EXACT_ALLOC	0
146 #define	VMEM_SEGS_PER_LEFT_ALLOC	1
147 #define	VMEM_SEGS_PER_RIGHT_ALLOC	1
148 #define	VMEM_SEGS_PER_MIDDLE_ALLOC	2
149 
150 /*
151  * vmem_populate() preallocates segment structures for vmem to do its work.
152  * It must preallocate enough for the worst case, which is when we must import
153  * a new span and then allocate from the middle of it.
154  */
155 #define	VMEM_SEGS_PER_ALLOC_MAX		\
156 	(VMEM_SEGS_PER_SPAN_CREATE + VMEM_SEGS_PER_MIDDLE_ALLOC)
157 
158 /*
159  * The segment structures themselves are allocated from vmem_seg_arena, so
160  * we have a recursion problem when vmem_seg_arena needs to populate itself.
161  * We address this by working out the maximum number of segment structures
162  * this act will require, and multiplying by the maximum number of threads
163  * that we'll allow to do it simultaneously.
164  *
165  * The worst-case segment consumption to populate vmem_seg_arena is as
166  * follows (depicted as a stack trace to indicate why events are occurring):
167  *
168  * vmem_alloc(vmem_seg_arena)		-> 2 segs (span create + exact alloc)
169  *  vmem_alloc(vmem_internal_arena)	-> 2 segs (span create + exact alloc)
170  *   heap_alloc(heap_arena)
171  *    vmem_alloc(heap_arena)		-> 4 seg (span create + alloc)
172  *     parent_alloc(parent_arena)
173  *	_vmem_extend_alloc(parent_arena) -> 3 seg (span create + left alloc)
174  *
175  * Note:  The reservation for heap_arena must be 4, since vmem_xalloc()
176  * is overly pessimistic on allocations where parent_arena has a stricter
177  * alignment than heap_arena.
178  *
179  * The worst-case consumption for any arena is 4 segment structures.
180  * For now, we only support VM_NOSLEEP allocations, so as long as we
181  * serialize all vmem_populates, a 4-seg reserve is sufficient.
182  */
183 #define	VMEM_POPULATE_SEGS_PER_ARENA	4
184 #define	VMEM_POPULATE_LOCKS		1
185 
186 #define	VMEM_POPULATE_RESERVE		\
187 	(VMEM_POPULATE_SEGS_PER_ARENA * VMEM_POPULATE_LOCKS)
188 
189 /*
190  * vmem_populate() ensures that each arena has VMEM_MINFREE seg structures
191  * so that it can satisfy the worst-case allocation *and* participate in
192  * worst-case allocation from vmem_seg_arena.
193  */
194 #define	VMEM_MINFREE	(VMEM_POPULATE_RESERVE + VMEM_SEGS_PER_ALLOC_MAX)
195 
196 /* Don't assume new statics are zeroed - see vmem_startup() */
197 static vmem_t vmem0[VMEM_INITIAL];
198 static vmem_t *vmem_populator[VMEM_INITIAL];
199 static uint32_t vmem_id;
200 static uint32_t vmem_populators;
201 static vmem_seg_t vmem_seg0[VMEM_SEG_INITIAL];
202 static vmem_seg_t *vmem_segfree;
203 static mutex_t vmem_list_lock;
204 static mutex_t vmem_segfree_lock;
205 static vmem_populate_lock_t vmem_nosleep_lock;
206 #define	IN_POPULATE()	(vmem_nosleep_lock.vmpl_thr == thr_self())
207 static vmem_t *vmem_list;
208 static vmem_t *vmem_internal_arena;
209 static vmem_t *vmem_seg_arena;
210 static vmem_t *vmem_hash_arena;
211 static vmem_t *vmem_vmem_arena;
212 
213 vmem_t *vmem_heap;
214 vmem_alloc_t *vmem_heap_alloc;
215 vmem_free_t *vmem_heap_free;
216 
217 uint32_t vmem_mtbf;		/* mean time between failures [default: off] */
218 size_t vmem_seg_size = sizeof (vmem_seg_t);
219 
220 /*
221  * Insert/delete from arena list (type 'a') or next-of-kin list (type 'k').
222  */
223 #define	VMEM_INSERT(vprev, vsp, type)					\
224 {									\
225 	vmem_seg_t *vnext = (vprev)->vs_##type##next;			\
226 	(vsp)->vs_##type##next = (vnext);				\
227 	(vsp)->vs_##type##prev = (vprev);				\
228 	(vprev)->vs_##type##next = (vsp);				\
229 	(vnext)->vs_##type##prev = (vsp);				\
230 }
231 
232 #define	VMEM_DELETE(vsp, type)						\
233 {									\
234 	vmem_seg_t *vprev = (vsp)->vs_##type##prev;			\
235 	vmem_seg_t *vnext = (vsp)->vs_##type##next;			\
236 	(vprev)->vs_##type##next = (vnext);				\
237 	(vnext)->vs_##type##prev = (vprev);				\
238 }
239 
240 /*
241  * Get a vmem_seg_t from the global segfree list.
242  */
243 static vmem_seg_t *
244 vmem_getseg_global(void)
245 {
246 	vmem_seg_t *vsp;
247 
248 	(void) mutex_lock(&vmem_segfree_lock);
249 	if ((vsp = vmem_segfree) != NULL)
250 		vmem_segfree = vsp->vs_knext;
251 	(void) mutex_unlock(&vmem_segfree_lock);
252 
253 	return (vsp);
254 }
255 
256 /*
257  * Put a vmem_seg_t on the global segfree list.
258  */
259 static void
260 vmem_putseg_global(vmem_seg_t *vsp)
261 {
262 	(void) mutex_lock(&vmem_segfree_lock);
263 	vsp->vs_knext = vmem_segfree;
264 	vmem_segfree = vsp;
265 	(void) mutex_unlock(&vmem_segfree_lock);
266 }
267 
268 /*
269  * Get a vmem_seg_t from vmp's segfree list.
270  */
271 static vmem_seg_t *
272 vmem_getseg(vmem_t *vmp)
273 {
274 	vmem_seg_t *vsp;
275 
276 	ASSERT(vmp->vm_nsegfree > 0);
277 
278 	vsp = vmp->vm_segfree;
279 	vmp->vm_segfree = vsp->vs_knext;
280 	vmp->vm_nsegfree--;
281 
282 	return (vsp);
283 }
284 
285 /*
286  * Put a vmem_seg_t on vmp's segfree list.
287  */
288 static void
289 vmem_putseg(vmem_t *vmp, vmem_seg_t *vsp)
290 {
291 	vsp->vs_knext = vmp->vm_segfree;
292 	vmp->vm_segfree = vsp;
293 	vmp->vm_nsegfree++;
294 }
295 
296 /*
297  * Add vsp to the appropriate freelist.
298  */
299 static void
300 vmem_freelist_insert(vmem_t *vmp, vmem_seg_t *vsp)
301 {
302 	vmem_seg_t *vprev;
303 
304 	ASSERT(*VMEM_HASH(vmp, vsp->vs_start) != vsp);
305 
306 	vprev = (vmem_seg_t *)&vmp->vm_freelist[highbit(VS_SIZE(vsp)) - 1];
307 	vsp->vs_type = VMEM_FREE;
308 	vmp->vm_freemap |= VS_SIZE(vprev);
309 	VMEM_INSERT(vprev, vsp, k);
310 
311 	(void) cond_broadcast(&vmp->vm_cv);
312 }
313 
314 /*
315  * Take vsp from the freelist.
316  */
317 static void
318 vmem_freelist_delete(vmem_t *vmp, vmem_seg_t *vsp)
319 {
320 	ASSERT(*VMEM_HASH(vmp, vsp->vs_start) != vsp);
321 	ASSERT(vsp->vs_type == VMEM_FREE);
322 
323 	if (vsp->vs_knext->vs_start == 0 && vsp->vs_kprev->vs_start == 0) {
324 		/*
325 		 * The segments on both sides of 'vsp' are freelist heads,
326 		 * so taking vsp leaves the freelist at vsp->vs_kprev empty.
327 		 */
328 		ASSERT(vmp->vm_freemap & VS_SIZE(vsp->vs_kprev));
329 		vmp->vm_freemap ^= VS_SIZE(vsp->vs_kprev);
330 	}
331 	VMEM_DELETE(vsp, k);
332 }
333 
334 /*
335  * Add vsp to the allocated-segment hash table and update kstats.
336  */
337 static void
338 vmem_hash_insert(vmem_t *vmp, vmem_seg_t *vsp)
339 {
340 	vmem_seg_t **bucket;
341 
342 	vsp->vs_type = VMEM_ALLOC;
343 	bucket = VMEM_HASH(vmp, vsp->vs_start);
344 	vsp->vs_knext = *bucket;
345 	*bucket = vsp;
346 
347 	if (vmem_seg_size == sizeof (vmem_seg_t)) {
348 		vsp->vs_depth = (uint8_t)getpcstack(vsp->vs_stack,
349 		    VMEM_STACK_DEPTH, 0);
350 		vsp->vs_thread = thr_self();
351 		vsp->vs_timestamp = gethrtime();
352 	} else {
353 		vsp->vs_depth = 0;
354 	}
355 
356 	vmp->vm_kstat.vk_alloc++;
357 	vmp->vm_kstat.vk_mem_inuse += VS_SIZE(vsp);
358 }
359 
360 /*
361  * Remove vsp from the allocated-segment hash table and update kstats.
362  */
363 static vmem_seg_t *
364 vmem_hash_delete(vmem_t *vmp, uintptr_t addr, size_t size)
365 {
366 	vmem_seg_t *vsp, **prev_vspp;
367 
368 	prev_vspp = VMEM_HASH(vmp, addr);
369 	while ((vsp = *prev_vspp) != NULL) {
370 		if (vsp->vs_start == addr) {
371 			*prev_vspp = vsp->vs_knext;
372 			break;
373 		}
374 		vmp->vm_kstat.vk_lookup++;
375 		prev_vspp = &vsp->vs_knext;
376 	}
377 
378 	if (vsp == NULL) {
379 		umem_panic("vmem_hash_delete(%p, %lx, %lu): bad free",
380 		    vmp, addr, size);
381 	}
382 	if (VS_SIZE(vsp) != size) {
383 		umem_panic("vmem_hash_delete(%p, %lx, %lu): wrong size "
384 		    "(expect %lu)", vmp, addr, size, VS_SIZE(vsp));
385 	}
386 
387 	vmp->vm_kstat.vk_free++;
388 	vmp->vm_kstat.vk_mem_inuse -= size;
389 
390 	return (vsp);
391 }
392 
393 /*
394  * Create a segment spanning the range [start, end) and add it to the arena.
395  */
396 static vmem_seg_t *
397 vmem_seg_create(vmem_t *vmp, vmem_seg_t *vprev, uintptr_t start, uintptr_t end)
398 {
399 	vmem_seg_t *newseg = vmem_getseg(vmp);
400 
401 	newseg->vs_start = start;
402 	newseg->vs_end = end;
403 	newseg->vs_type = 0;
404 	newseg->vs_import = 0;
405 
406 	VMEM_INSERT(vprev, newseg, a);
407 
408 	return (newseg);
409 }
410 
411 /*
412  * Remove segment vsp from the arena.
413  */
414 static void
415 vmem_seg_destroy(vmem_t *vmp, vmem_seg_t *vsp)
416 {
417 	ASSERT(vsp->vs_type != VMEM_ROTOR);
418 	VMEM_DELETE(vsp, a);
419 
420 	vmem_putseg(vmp, vsp);
421 }
422 
423 /*
424  * Add the span [vaddr, vaddr + size) to vmp and update kstats.
425  */
426 static vmem_seg_t *
427 vmem_span_create(vmem_t *vmp, void *vaddr, size_t size, uint8_t import)
428 {
429 	vmem_seg_t *knext;
430 	vmem_seg_t *newseg, *span;
431 	uintptr_t start = (uintptr_t)vaddr;
432 	uintptr_t end = start + size;
433 
434 	knext = &vmp->vm_seg0;
435 	if (!import && vmp->vm_source_alloc == NULL) {
436 		vmem_seg_t *kend, *kprev;
437 		/*
438 		 * non-imported spans are sorted in address order.  This
439 		 * makes vmem_extend_unlocked() much more effective.
440 		 *
441 		 * We search in reverse order, since new spans are
442 		 * generally at higher addresses.
443 		 */
444 		kend = &vmp->vm_seg0;
445 		for (kprev = kend->vs_kprev; kprev != kend;
446 		    kprev = kprev->vs_kprev) {
447 			if (!kprev->vs_import && (kprev->vs_end - 1) < start)
448 				break;
449 		}
450 		knext = kprev->vs_knext;
451 	}
452 
453 	ASSERT(MUTEX_HELD(&vmp->vm_lock));
454 
455 	if ((start | end) & (vmp->vm_quantum - 1)) {
456 		umem_panic("vmem_span_create(%p, %p, %lu): misaligned",
457 		    vmp, vaddr, size);
458 	}
459 
460 	span = vmem_seg_create(vmp, knext->vs_aprev, start, end);
461 	span->vs_type = VMEM_SPAN;
462 	span->vs_import = import;
463 	VMEM_INSERT(knext->vs_kprev, span, k);
464 
465 	newseg = vmem_seg_create(vmp, span, start, end);
466 	vmem_freelist_insert(vmp, newseg);
467 
468 	if (import)
469 		vmp->vm_kstat.vk_mem_import += size;
470 	vmp->vm_kstat.vk_mem_total += size;
471 
472 	return (newseg);
473 }
474 
475 /*
476  * Remove span vsp from vmp and update kstats.
477  */
478 static void
479 vmem_span_destroy(vmem_t *vmp, vmem_seg_t *vsp)
480 {
481 	vmem_seg_t *span = vsp->vs_aprev;
482 	size_t size = VS_SIZE(vsp);
483 
484 	ASSERT(MUTEX_HELD(&vmp->vm_lock));
485 	ASSERT(span->vs_type == VMEM_SPAN);
486 
487 	if (span->vs_import)
488 		vmp->vm_kstat.vk_mem_import -= size;
489 	vmp->vm_kstat.vk_mem_total -= size;
490 
491 	VMEM_DELETE(span, k);
492 
493 	vmem_seg_destroy(vmp, vsp);
494 	vmem_seg_destroy(vmp, span);
495 }
496 
497 /*
498  * Allocate the subrange [addr, addr + size) from segment vsp.
499  * If there are leftovers on either side, place them on the freelist.
500  * Returns a pointer to the segment representing [addr, addr + size).
501  */
502 static vmem_seg_t *
503 vmem_seg_alloc(vmem_t *vmp, vmem_seg_t *vsp, uintptr_t addr, size_t size)
504 {
505 	uintptr_t vs_start = vsp->vs_start;
506 	uintptr_t vs_end = vsp->vs_end;
507 	size_t vs_size = vs_end - vs_start;
508 	size_t realsize = P2ROUNDUP(size, vmp->vm_quantum);
509 	uintptr_t addr_end = addr + realsize;
510 
511 	ASSERT(P2PHASE(vs_start, vmp->vm_quantum) == 0);
512 	ASSERT(P2PHASE(addr, vmp->vm_quantum) == 0);
513 	ASSERT(vsp->vs_type == VMEM_FREE);
514 	ASSERT(addr >= vs_start && addr_end - 1 <= vs_end - 1);
515 	ASSERT(addr - 1 <= addr_end - 1);
516 
517 	/*
518 	 * If we're allocating from the start of the segment, and the
519 	 * remainder will be on the same freelist, we can save quite
520 	 * a bit of work.
521 	 */
522 	if (P2SAMEHIGHBIT(vs_size, vs_size - realsize) && addr == vs_start) {
523 		ASSERT(highbit(vs_size) == highbit(vs_size - realsize));
524 		vsp->vs_start = addr_end;
525 		vsp = vmem_seg_create(vmp, vsp->vs_aprev, addr, addr + size);
526 		vmem_hash_insert(vmp, vsp);
527 		return (vsp);
528 	}
529 
530 	vmem_freelist_delete(vmp, vsp);
531 
532 	if (vs_end != addr_end)
533 		vmem_freelist_insert(vmp,
534 		    vmem_seg_create(vmp, vsp, addr_end, vs_end));
535 
536 	if (vs_start != addr)
537 		vmem_freelist_insert(vmp,
538 		    vmem_seg_create(vmp, vsp->vs_aprev, vs_start, addr));
539 
540 	vsp->vs_start = addr;
541 	vsp->vs_end = addr + size;
542 
543 	vmem_hash_insert(vmp, vsp);
544 	return (vsp);
545 }
546 
547 /*
548  * We cannot reap if we are in the middle of a vmem_populate().
549  */
550 void
551 vmem_reap(void)
552 {
553 	if (!IN_POPULATE())
554 		umem_reap();
555 }
556 
557 /*
558  * Populate vmp's segfree list with VMEM_MINFREE vmem_seg_t structures.
559  */
560 static int
561 vmem_populate(vmem_t *vmp, int vmflag)
562 {
563 	char *p;
564 	vmem_seg_t *vsp;
565 	ssize_t nseg;
566 	size_t size;
567 	vmem_populate_lock_t *lp;
568 	int i;
569 
570 	while (vmp->vm_nsegfree < VMEM_MINFREE &&
571 	    (vsp = vmem_getseg_global()) != NULL)
572 		vmem_putseg(vmp, vsp);
573 
574 	if (vmp->vm_nsegfree >= VMEM_MINFREE)
575 		return (1);
576 
577 	/*
578 	 * If we're already populating, tap the reserve.
579 	 */
580 	if (vmem_nosleep_lock.vmpl_thr == thr_self()) {
581 		ASSERT(vmp->vm_cflags & VMC_POPULATOR);
582 		return (1);
583 	}
584 
585 	(void) mutex_unlock(&vmp->vm_lock);
586 
587 	ASSERT(vmflag & VM_NOSLEEP);	/* we do not allow sleep allocations */
588 	lp = &vmem_nosleep_lock;
589 
590 	/*
591 	 * Cannot be just a mutex_lock(), since that has no effect if
592 	 * libthread is not linked.
593 	 */
594 	(void) mutex_lock(&lp->vmpl_mutex);
595 	ASSERT(lp->vmpl_thr == 0);
596 	lp->vmpl_thr = thr_self();
597 
598 	nseg = VMEM_MINFREE + vmem_populators * VMEM_POPULATE_RESERVE;
599 	size = P2ROUNDUP(nseg * vmem_seg_size, vmem_seg_arena->vm_quantum);
600 	nseg = size / vmem_seg_size;
601 
602 	/*
603 	 * The following vmem_alloc() may need to populate vmem_seg_arena
604 	 * and all the things it imports from.  When doing so, it will tap
605 	 * each arena's reserve to prevent recursion (see the block comment
606 	 * above the definition of VMEM_POPULATE_RESERVE).
607 	 *
608 	 * During this allocation, vmem_reap() is a no-op.  If the allocation
609 	 * fails, we call vmem_reap() after dropping the population lock.
610 	 */
611 	p = vmem_alloc(vmem_seg_arena, size, vmflag & VM_UMFLAGS);
612 	if (p == NULL) {
613 		lp->vmpl_thr = 0;
614 		(void) mutex_unlock(&lp->vmpl_mutex);
615 		vmem_reap();
616 
617 		(void) mutex_lock(&vmp->vm_lock);
618 		vmp->vm_kstat.vk_populate_fail++;
619 		return (0);
620 	}
621 	/*
622 	 * Restock the arenas that may have been depleted during population.
623 	 */
624 	for (i = 0; i < vmem_populators; i++) {
625 		(void) mutex_lock(&vmem_populator[i]->vm_lock);
626 		while (vmem_populator[i]->vm_nsegfree < VMEM_POPULATE_RESERVE)
627 			vmem_putseg(vmem_populator[i],
628 			    (vmem_seg_t *)(p + --nseg * vmem_seg_size));
629 		(void) mutex_unlock(&vmem_populator[i]->vm_lock);
630 	}
631 
632 	lp->vmpl_thr = 0;
633 	(void) mutex_unlock(&lp->vmpl_mutex);
634 	(void) mutex_lock(&vmp->vm_lock);
635 
636 	/*
637 	 * Now take our own segments.
638 	 */
639 	ASSERT(nseg >= VMEM_MINFREE);
640 	while (vmp->vm_nsegfree < VMEM_MINFREE)
641 		vmem_putseg(vmp, (vmem_seg_t *)(p + --nseg * vmem_seg_size));
642 
643 	/*
644 	 * Give the remainder to charity.
645 	 */
646 	while (nseg > 0)
647 		vmem_putseg_global((vmem_seg_t *)(p + --nseg * vmem_seg_size));
648 
649 	return (1);
650 }
651 
652 /*
653  * Advance a walker from its previous position to 'afterme'.
654  * Note: may drop and reacquire vmp->vm_lock.
655  */
656 static void
657 vmem_advance(vmem_t *vmp, vmem_seg_t *walker, vmem_seg_t *afterme)
658 {
659 	vmem_seg_t *vprev = walker->vs_aprev;
660 	vmem_seg_t *vnext = walker->vs_anext;
661 	vmem_seg_t *vsp = NULL;
662 
663 	VMEM_DELETE(walker, a);
664 
665 	if (afterme != NULL)
666 		VMEM_INSERT(afterme, walker, a);
667 
668 	/*
669 	 * The walker segment's presence may have prevented its neighbors
670 	 * from coalescing.  If so, coalesce them now.
671 	 */
672 	if (vprev->vs_type == VMEM_FREE) {
673 		if (vnext->vs_type == VMEM_FREE) {
674 			ASSERT(vprev->vs_end == vnext->vs_start);
675 			vmem_freelist_delete(vmp, vnext);
676 			vmem_freelist_delete(vmp, vprev);
677 			vprev->vs_end = vnext->vs_end;
678 			vmem_freelist_insert(vmp, vprev);
679 			vmem_seg_destroy(vmp, vnext);
680 		}
681 		vsp = vprev;
682 	} else if (vnext->vs_type == VMEM_FREE) {
683 		vsp = vnext;
684 	}
685 
686 	/*
687 	 * vsp could represent a complete imported span,
688 	 * in which case we must return it to the source.
689 	 */
690 	if (vsp != NULL && vsp->vs_aprev->vs_import &&
691 	    vmp->vm_source_free != NULL &&
692 	    vsp->vs_aprev->vs_type == VMEM_SPAN &&
693 	    vsp->vs_anext->vs_type == VMEM_SPAN) {
694 		void *vaddr = (void *)vsp->vs_start;
695 		size_t size = VS_SIZE(vsp);
696 		ASSERT(size == VS_SIZE(vsp->vs_aprev));
697 		vmem_freelist_delete(vmp, vsp);
698 		vmem_span_destroy(vmp, vsp);
699 		(void) mutex_unlock(&vmp->vm_lock);
700 		vmp->vm_source_free(vmp->vm_source, vaddr, size);
701 		(void) mutex_lock(&vmp->vm_lock);
702 	}
703 }
704 
705 /*
706  * VM_NEXTFIT allocations deliberately cycle through all virtual addresses
707  * in an arena, so that we avoid reusing addresses for as long as possible.
708  * This helps to catch used-after-freed bugs.  It's also the perfect policy
709  * for allocating things like process IDs, where we want to cycle through
710  * all values in order.
711  */
712 static void *
713 vmem_nextfit_alloc(vmem_t *vmp, size_t size, int vmflag)
714 {
715 	vmem_seg_t *vsp, *rotor;
716 	uintptr_t addr;
717 	size_t realsize = P2ROUNDUP(size, vmp->vm_quantum);
718 	size_t vs_size;
719 
720 	(void) mutex_lock(&vmp->vm_lock);
721 
722 	if (vmp->vm_nsegfree < VMEM_MINFREE && !vmem_populate(vmp, vmflag)) {
723 		(void) mutex_unlock(&vmp->vm_lock);
724 		return (NULL);
725 	}
726 
727 	/*
728 	 * The common case is that the segment right after the rotor is free,
729 	 * and large enough that extracting 'size' bytes won't change which
730 	 * freelist it's on.  In this case we can avoid a *lot* of work.
731 	 * Instead of the normal vmem_seg_alloc(), we just advance the start
732 	 * address of the victim segment.  Instead of moving the rotor, we
733 	 * create the new segment structure *behind the rotor*, which has
734 	 * the same effect.  And finally, we know we don't have to coalesce
735 	 * the rotor's neighbors because the new segment lies between them.
736 	 */
737 	rotor = &vmp->vm_rotor;
738 	vsp = rotor->vs_anext;
739 	if (vsp->vs_type == VMEM_FREE && (vs_size = VS_SIZE(vsp)) > realsize &&
740 	    P2SAMEHIGHBIT(vs_size, vs_size - realsize)) {
741 		ASSERT(highbit(vs_size) == highbit(vs_size - realsize));
742 		addr = vsp->vs_start;
743 		vsp->vs_start = addr + realsize;
744 		vmem_hash_insert(vmp,
745 		    vmem_seg_create(vmp, rotor->vs_aprev, addr, addr + size));
746 		(void) mutex_unlock(&vmp->vm_lock);
747 		return ((void *)addr);
748 	}
749 
750 	/*
751 	 * Starting at the rotor, look for a segment large enough to
752 	 * satisfy the allocation.
753 	 */
754 	for (;;) {
755 		vmp->vm_kstat.vk_search++;
756 		if (vsp->vs_type == VMEM_FREE && VS_SIZE(vsp) >= size)
757 			break;
758 		vsp = vsp->vs_anext;
759 		if (vsp == rotor) {
760 			int cancel_state;
761 
762 			/*
763 			 * We've come full circle.  One possibility is that the
764 			 * there's actually enough space, but the rotor itself
765 			 * is preventing the allocation from succeeding because
766 			 * it's sitting between two free segments.  Therefore,
767 			 * we advance the rotor and see if that liberates a
768 			 * suitable segment.
769 			 */
770 			vmem_advance(vmp, rotor, rotor->vs_anext);
771 			vsp = rotor->vs_aprev;
772 			if (vsp->vs_type == VMEM_FREE && VS_SIZE(vsp) >= size)
773 				break;
774 			/*
775 			 * If there's a lower arena we can import from, or it's
776 			 * a VM_NOSLEEP allocation, let vmem_xalloc() handle it.
777 			 * Otherwise, wait until another thread frees something.
778 			 */
779 			if (vmp->vm_source_alloc != NULL ||
780 			    (vmflag & VM_NOSLEEP)) {
781 				(void) mutex_unlock(&vmp->vm_lock);
782 				return (vmem_xalloc(vmp, size, vmp->vm_quantum,
783 				    0, 0, NULL, NULL, vmflag & VM_UMFLAGS));
784 			}
785 			vmp->vm_kstat.vk_wait++;
786 			(void) pthread_setcancelstate(PTHREAD_CANCEL_DISABLE,
787 			    &cancel_state);
788 			(void) cond_wait(&vmp->vm_cv, &vmp->vm_lock);
789 			(void) pthread_setcancelstate(cancel_state, NULL);
790 			vsp = rotor->vs_anext;
791 		}
792 	}
793 
794 	/*
795 	 * We found a segment.  Extract enough space to satisfy the allocation.
796 	 */
797 	addr = vsp->vs_start;
798 	vsp = vmem_seg_alloc(vmp, vsp, addr, size);
799 	ASSERT(vsp->vs_type == VMEM_ALLOC &&
800 	    vsp->vs_start == addr && vsp->vs_end == addr + size);
801 
802 	/*
803 	 * Advance the rotor to right after the newly-allocated segment.
804 	 * That's where the next VM_NEXTFIT allocation will begin searching.
805 	 */
806 	vmem_advance(vmp, rotor, vsp);
807 	(void) mutex_unlock(&vmp->vm_lock);
808 	return ((void *)addr);
809 }
810 
811 /*
812  * Allocate size bytes at offset phase from an align boundary such that the
813  * resulting segment [addr, addr + size) is a subset of [minaddr, maxaddr)
814  * that does not straddle a nocross-aligned boundary.
815  */
816 void *
817 vmem_xalloc(vmem_t *vmp, size_t size, size_t align, size_t phase,
818     size_t nocross, void *minaddr, void *maxaddr, int vmflag)
819 {
820 	vmem_seg_t *vsp;
821 	vmem_seg_t *vbest = NULL;
822 	uintptr_t addr = 0, taddr, start, end;
823 	void *vaddr;
824 	int hb, flist, resv;
825 	uint32_t mtbf;
826 
827 	if (phase > 0 && phase >= align)
828 		umem_panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
829 		    "invalid phase",
830 		    (void *)vmp, size, align, phase, nocross,
831 		    minaddr, maxaddr, vmflag);
832 
833 	if (align == 0)
834 		align = vmp->vm_quantum;
835 
836 	if ((align | phase | nocross) & (vmp->vm_quantum - 1)) {
837 		umem_panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
838 		    "parameters not vm_quantum aligned",
839 		    (void *)vmp, size, align, phase, nocross,
840 		    minaddr, maxaddr, vmflag);
841 	}
842 
843 	if (nocross != 0 &&
844 	    (align > nocross || P2ROUNDUP(phase + size, align) > nocross)) {
845 		umem_panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
846 		    "overconstrained allocation",
847 		    (void *)vmp, size, align, phase, nocross,
848 		    minaddr, maxaddr, vmflag);
849 	}
850 
851 	if ((mtbf = vmem_mtbf | vmp->vm_mtbf) != 0 && gethrtime() % mtbf == 0 &&
852 	    (vmflag & (VM_NOSLEEP | VM_PANIC)) == VM_NOSLEEP)
853 		return (NULL);
854 
855 	(void) mutex_lock(&vmp->vm_lock);
856 	for (;;) {
857 		int cancel_state;
858 
859 		if (vmp->vm_nsegfree < VMEM_MINFREE &&
860 		    !vmem_populate(vmp, vmflag))
861 			break;
862 
863 		/*
864 		 * highbit() returns the highest bit + 1, which is exactly
865 		 * what we want: we want to search the first freelist whose
866 		 * members are *definitely* large enough to satisfy our
867 		 * allocation.  However, there are certain cases in which we
868 		 * want to look at the next-smallest freelist (which *might*
869 		 * be able to satisfy the allocation):
870 		 *
871 		 * (1)	The size is exactly a power of 2, in which case
872 		 *	the smaller freelist is always big enough;
873 		 *
874 		 * (2)	All other freelists are empty;
875 		 *
876 		 * (3)	We're in the highest possible freelist, which is
877 		 *	always empty (e.g. the 4GB freelist on 32-bit systems);
878 		 *
879 		 * (4)	We're doing a best-fit or first-fit allocation.
880 		 */
881 		if ((size & (size - 1)) == 0) {
882 			flist = lowbit(P2ALIGN(vmp->vm_freemap, size));
883 		} else {
884 			hb = highbit(size);
885 			if ((vmp->vm_freemap >> hb) == 0 ||
886 			    hb == VMEM_FREELISTS ||
887 			    (vmflag & (VM_BESTFIT | VM_FIRSTFIT)))
888 				hb--;
889 			flist = lowbit(P2ALIGN(vmp->vm_freemap, 1UL << hb));
890 		}
891 
892 		for (vbest = NULL, vsp = (flist == 0) ? NULL :
893 		    vmp->vm_freelist[flist - 1].vs_knext;
894 		    vsp != NULL; vsp = vsp->vs_knext) {
895 			vmp->vm_kstat.vk_search++;
896 			if (vsp->vs_start == 0) {
897 				/*
898 				 * We're moving up to a larger freelist,
899 				 * so if we've already found a candidate,
900 				 * the fit can't possibly get any better.
901 				 */
902 				if (vbest != NULL)
903 					break;
904 				/*
905 				 * Find the next non-empty freelist.
906 				 */
907 				flist = lowbit(P2ALIGN(vmp->vm_freemap,
908 				    VS_SIZE(vsp)));
909 				if (flist-- == 0)
910 					break;
911 				vsp = (vmem_seg_t *)&vmp->vm_freelist[flist];
912 				ASSERT(vsp->vs_knext->vs_type == VMEM_FREE);
913 				continue;
914 			}
915 			if (vsp->vs_end - 1 < (uintptr_t)minaddr)
916 				continue;
917 			if (vsp->vs_start > (uintptr_t)maxaddr - 1)
918 				continue;
919 			start = MAX(vsp->vs_start, (uintptr_t)minaddr);
920 			end = MIN(vsp->vs_end - 1, (uintptr_t)maxaddr - 1) + 1;
921 			taddr = P2PHASEUP(start, align, phase);
922 			if (P2BOUNDARY(taddr, size, nocross))
923 				taddr +=
924 				    P2ROUNDUP(P2NPHASE(taddr, nocross), align);
925 			if ((taddr - start) + size > end - start ||
926 			    (vbest != NULL && VS_SIZE(vsp) >= VS_SIZE(vbest)))
927 				continue;
928 			vbest = vsp;
929 			addr = taddr;
930 			if (!(vmflag & VM_BESTFIT) || VS_SIZE(vbest) == size)
931 				break;
932 		}
933 		if (vbest != NULL)
934 			break;
935 		if (size == 0)
936 			umem_panic("vmem_xalloc(): size == 0");
937 		if (vmp->vm_source_alloc != NULL && nocross == 0 &&
938 		    minaddr == NULL && maxaddr == NULL) {
939 			size_t asize = P2ROUNDUP(size + phase,
940 			    MAX(align, vmp->vm_source->vm_quantum));
941 			if (asize < size) {		/* overflow */
942 				(void) mutex_unlock(&vmp->vm_lock);
943 				if (vmflag & VM_NOSLEEP)
944 					return (NULL);
945 
946 				umem_panic("vmem_xalloc(): "
947 				    "overflow on VM_SLEEP allocation");
948 			}
949 			/*
950 			 * Determine how many segment structures we'll consume.
951 			 * The calculation must be presise because if we're
952 			 * here on behalf of vmem_populate(), we are taking
953 			 * segments from a very limited reserve.
954 			 */
955 			resv = (size == asize) ?
956 			    VMEM_SEGS_PER_SPAN_CREATE +
957 			    VMEM_SEGS_PER_EXACT_ALLOC :
958 			    VMEM_SEGS_PER_ALLOC_MAX;
959 			ASSERT(vmp->vm_nsegfree >= resv);
960 			vmp->vm_nsegfree -= resv;	/* reserve our segs */
961 			(void) mutex_unlock(&vmp->vm_lock);
962 			vaddr = vmp->vm_source_alloc(vmp->vm_source, asize,
963 			    vmflag & VM_UMFLAGS);
964 			(void) mutex_lock(&vmp->vm_lock);
965 			vmp->vm_nsegfree += resv;	/* claim reservation */
966 			if (vaddr != NULL) {
967 				vbest = vmem_span_create(vmp, vaddr, asize, 1);
968 				addr = P2PHASEUP(vbest->vs_start, align, phase);
969 				break;
970 			}
971 		}
972 		(void) mutex_unlock(&vmp->vm_lock);
973 		vmem_reap();
974 		(void) mutex_lock(&vmp->vm_lock);
975 		if (vmflag & VM_NOSLEEP)
976 			break;
977 		vmp->vm_kstat.vk_wait++;
978 		(void) pthread_setcancelstate(PTHREAD_CANCEL_DISABLE,
979 		    &cancel_state);
980 		(void) cond_wait(&vmp->vm_cv, &vmp->vm_lock);
981 		(void) pthread_setcancelstate(cancel_state, NULL);
982 	}
983 	if (vbest != NULL) {
984 		if (addr == 0) {
985 			umem_panic("vmem_xalloc(): addr == 0");
986 		}
987 		ASSERT(vbest->vs_type == VMEM_FREE);
988 		ASSERT(vbest->vs_knext != vbest);
989 		(void) vmem_seg_alloc(vmp, vbest, addr, size);
990 		(void) mutex_unlock(&vmp->vm_lock);
991 		ASSERT(P2PHASE(addr, align) == phase);
992 		ASSERT(!P2BOUNDARY(addr, size, nocross));
993 		ASSERT(addr >= (uintptr_t)minaddr);
994 		ASSERT(addr + size - 1 <= (uintptr_t)maxaddr - 1);
995 		return ((void *)addr);
996 	}
997 	vmp->vm_kstat.vk_fail++;
998 	(void) mutex_unlock(&vmp->vm_lock);
999 	if (vmflag & VM_PANIC)
1000 		umem_panic("vmem_xalloc(%p, %lu, %lu, %lu, %lu, %p, %p, %x): "
1001 		    "cannot satisfy mandatory allocation",
1002 		    (void *)vmp, size, align, phase, nocross,
1003 		    minaddr, maxaddr, vmflag);
1004 	return (NULL);
1005 }
1006 
1007 /*
1008  * Free the segment [vaddr, vaddr + size), where vaddr was a constrained
1009  * allocation.  vmem_xalloc() and vmem_xfree() must always be paired because
1010  * both routines bypass the quantum caches.
1011  */
1012 void
1013 vmem_xfree(vmem_t *vmp, void *vaddr, size_t size)
1014 {
1015 	vmem_seg_t *vsp, *vnext, *vprev;
1016 
1017 	(void) mutex_lock(&vmp->vm_lock);
1018 
1019 	vsp = vmem_hash_delete(vmp, (uintptr_t)vaddr, size);
1020 	vsp->vs_end = P2ROUNDUP(vsp->vs_end, vmp->vm_quantum);
1021 
1022 	/*
1023 	 * Attempt to coalesce with the next segment.
1024 	 */
1025 	vnext = vsp->vs_anext;
1026 	if (vnext->vs_type == VMEM_FREE) {
1027 		ASSERT(vsp->vs_end == vnext->vs_start);
1028 		vmem_freelist_delete(vmp, vnext);
1029 		vsp->vs_end = vnext->vs_end;
1030 		vmem_seg_destroy(vmp, vnext);
1031 	}
1032 
1033 	/*
1034 	 * Attempt to coalesce with the previous segment.
1035 	 */
1036 	vprev = vsp->vs_aprev;
1037 	if (vprev->vs_type == VMEM_FREE) {
1038 		ASSERT(vprev->vs_end == vsp->vs_start);
1039 		vmem_freelist_delete(vmp, vprev);
1040 		vprev->vs_end = vsp->vs_end;
1041 		vmem_seg_destroy(vmp, vsp);
1042 		vsp = vprev;
1043 	}
1044 
1045 	/*
1046 	 * If the entire span is free, return it to the source.
1047 	 */
1048 	if (vsp->vs_aprev->vs_import && vmp->vm_source_free != NULL &&
1049 	    vsp->vs_aprev->vs_type == VMEM_SPAN &&
1050 	    vsp->vs_anext->vs_type == VMEM_SPAN) {
1051 		vaddr = (void *)vsp->vs_start;
1052 		size = VS_SIZE(vsp);
1053 		ASSERT(size == VS_SIZE(vsp->vs_aprev));
1054 		vmem_span_destroy(vmp, vsp);
1055 		(void) mutex_unlock(&vmp->vm_lock);
1056 		vmp->vm_source_free(vmp->vm_source, vaddr, size);
1057 	} else {
1058 		vmem_freelist_insert(vmp, vsp);
1059 		(void) mutex_unlock(&vmp->vm_lock);
1060 	}
1061 }
1062 
1063 /*
1064  * Allocate size bytes from arena vmp.  Returns the allocated address
1065  * on success, NULL on failure.  vmflag specifies VM_SLEEP or VM_NOSLEEP,
1066  * and may also specify best-fit, first-fit, or next-fit allocation policy
1067  * instead of the default instant-fit policy.  VM_SLEEP allocations are
1068  * guaranteed to succeed.
1069  */
1070 void *
1071 vmem_alloc(vmem_t *vmp, size_t size, int vmflag)
1072 {
1073 	vmem_seg_t *vsp;
1074 	uintptr_t addr;
1075 	int hb;
1076 	int flist = 0;
1077 	uint32_t mtbf;
1078 	vmflag |= vmem_allocator;
1079 
1080 	if (size - 1 < vmp->vm_qcache_max) {
1081 		ASSERT(vmflag & VM_NOSLEEP);
1082 		return (_umem_cache_alloc(vmp->vm_qcache[(size - 1) >>
1083 		    vmp->vm_qshift], UMEM_DEFAULT));
1084 	}
1085 
1086 	if ((mtbf = vmem_mtbf | vmp->vm_mtbf) != 0 && gethrtime() % mtbf == 0 &&
1087 	    (vmflag & (VM_NOSLEEP | VM_PANIC)) == VM_NOSLEEP)
1088 		return (NULL);
1089 
1090 	if (vmflag & VM_NEXTFIT)
1091 		return (vmem_nextfit_alloc(vmp, size, vmflag));
1092 
1093 	if (vmflag & (VM_BESTFIT | VM_FIRSTFIT))
1094 		return (vmem_xalloc(vmp, size, vmp->vm_quantum, 0, 0,
1095 		    NULL, NULL, vmflag));
1096 
1097 	/*
1098 	 * Unconstrained instant-fit allocation from the segment list.
1099 	 */
1100 	(void) mutex_lock(&vmp->vm_lock);
1101 
1102 	if (vmp->vm_nsegfree >= VMEM_MINFREE || vmem_populate(vmp, vmflag)) {
1103 		if ((size & (size - 1)) == 0)
1104 			flist = lowbit(P2ALIGN(vmp->vm_freemap, size));
1105 		else if ((hb = highbit(size)) < VMEM_FREELISTS)
1106 			flist = lowbit(P2ALIGN(vmp->vm_freemap, 1UL << hb));
1107 	}
1108 
1109 	if (flist-- == 0) {
1110 		(void) mutex_unlock(&vmp->vm_lock);
1111 		return (vmem_xalloc(vmp, size, vmp->vm_quantum,
1112 		    0, 0, NULL, NULL, vmflag));
1113 	}
1114 
1115 	ASSERT(size <= (1UL << flist));
1116 	vsp = vmp->vm_freelist[flist].vs_knext;
1117 	addr = vsp->vs_start;
1118 	(void) vmem_seg_alloc(vmp, vsp, addr, size);
1119 	(void) mutex_unlock(&vmp->vm_lock);
1120 	return ((void *)addr);
1121 }
1122 
1123 /*
1124  * Free the segment [vaddr, vaddr + size).
1125  */
1126 void
1127 vmem_free(vmem_t *vmp, void *vaddr, size_t size)
1128 {
1129 	if (size - 1 < vmp->vm_qcache_max)
1130 		_umem_cache_free(vmp->vm_qcache[(size - 1) >> vmp->vm_qshift],
1131 		    vaddr);
1132 	else
1133 		vmem_xfree(vmp, vaddr, size);
1134 }
1135 
1136 /*
1137  * Determine whether arena vmp contains the segment [vaddr, vaddr + size).
1138  */
1139 int
1140 vmem_contains(vmem_t *vmp, void *vaddr, size_t size)
1141 {
1142 	uintptr_t start = (uintptr_t)vaddr;
1143 	uintptr_t end = start + size;
1144 	vmem_seg_t *vsp;
1145 	vmem_seg_t *seg0 = &vmp->vm_seg0;
1146 
1147 	(void) mutex_lock(&vmp->vm_lock);
1148 	vmp->vm_kstat.vk_contains++;
1149 	for (vsp = seg0->vs_knext; vsp != seg0; vsp = vsp->vs_knext) {
1150 		vmp->vm_kstat.vk_contains_search++;
1151 		ASSERT(vsp->vs_type == VMEM_SPAN);
1152 		if (start >= vsp->vs_start && end - 1 <= vsp->vs_end - 1)
1153 			break;
1154 	}
1155 	(void) mutex_unlock(&vmp->vm_lock);
1156 	return (vsp != seg0);
1157 }
1158 
1159 /*
1160  * Add the span [vaddr, vaddr + size) to arena vmp.
1161  */
1162 void *
1163 vmem_add(vmem_t *vmp, void *vaddr, size_t size, int vmflag)
1164 {
1165 	if (vaddr == NULL || size == 0) {
1166 		umem_panic("vmem_add(%p, %p, %lu): bad arguments",
1167 		    vmp, vaddr, size);
1168 	}
1169 
1170 	ASSERT(!vmem_contains(vmp, vaddr, size));
1171 
1172 	(void) mutex_lock(&vmp->vm_lock);
1173 	if (vmem_populate(vmp, vmflag))
1174 		(void) vmem_span_create(vmp, vaddr, size, 0);
1175 	else
1176 		vaddr = NULL;
1177 	(void) cond_broadcast(&vmp->vm_cv);
1178 	(void) mutex_unlock(&vmp->vm_lock);
1179 	return (vaddr);
1180 }
1181 
1182 /*
1183  * Adds the address range [addr, endaddr) to arena vmp, by either:
1184  *   1. joining two existing spans, [x, addr), and [endaddr, y) (which
1185  *      are in that order) into a single [x, y) span,
1186  *   2. expanding an existing [x, addr) span to [x, endaddr),
1187  *   3. expanding an existing [endaddr, x) span to [addr, x), or
1188  *   4. creating a new [addr, endaddr) span.
1189  *
1190  * Called with vmp->vm_lock held, and a successful vmem_populate() completed.
1191  * Cannot fail.  Returns the new segment.
1192  *
1193  * NOTE:  this algorithm is linear-time in the number of spans, but is
1194  *      constant-time when you are extending the last (highest-addressed)
1195  *      span.
1196  */
1197 static vmem_seg_t *
1198 vmem_extend_unlocked(vmem_t *vmp, uintptr_t addr, uintptr_t endaddr)
1199 {
1200 	vmem_seg_t *span;
1201 	vmem_seg_t *vsp;
1202 
1203 	vmem_seg_t *end = &vmp->vm_seg0;
1204 
1205 	ASSERT(MUTEX_HELD(&vmp->vm_lock));
1206 
1207 	/*
1208 	 * the second "if" clause below relies on the direction of this search
1209 	 */
1210 	for (span = end->vs_kprev; span != end; span = span->vs_kprev) {
1211 		if (span->vs_end == addr || span->vs_start == endaddr)
1212 			break;
1213 	}
1214 
1215 	if (span == end)
1216 		return (vmem_span_create(vmp, (void *)addr, endaddr - addr, 0));
1217 	if (span->vs_kprev->vs_end == addr && span->vs_start == endaddr) {
1218 		vmem_seg_t *prevspan = span->vs_kprev;
1219 		vmem_seg_t *nextseg = span->vs_anext;
1220 		vmem_seg_t *prevseg = span->vs_aprev;
1221 
1222 		/*
1223 		 * prevspan becomes the span marker for the full range
1224 		 */
1225 		prevspan->vs_end = span->vs_end;
1226 
1227 		/*
1228 		 * Notionally, span becomes a free segment representing
1229 		 * [addr, endaddr).
1230 		 *
1231 		 * However, if either of its neighbors are free, we coalesce
1232 		 * by destroying span and changing the free segment.
1233 		 */
1234 		if (prevseg->vs_type == VMEM_FREE &&
1235 		    nextseg->vs_type == VMEM_FREE) {
1236 			/*
1237 			 * coalesce both ways
1238 			 */
1239 			ASSERT(prevseg->vs_end == addr &&
1240 			    nextseg->vs_start == endaddr);
1241 
1242 			vmem_freelist_delete(vmp, prevseg);
1243 			prevseg->vs_end = nextseg->vs_end;
1244 
1245 			vmem_freelist_delete(vmp, nextseg);
1246 			VMEM_DELETE(span, k);
1247 			vmem_seg_destroy(vmp, nextseg);
1248 			vmem_seg_destroy(vmp, span);
1249 
1250 			vsp = prevseg;
1251 		} else if (prevseg->vs_type == VMEM_FREE) {
1252 			/*
1253 			 * coalesce left
1254 			 */
1255 			ASSERT(prevseg->vs_end == addr);
1256 
1257 			VMEM_DELETE(span, k);
1258 			vmem_seg_destroy(vmp, span);
1259 
1260 			vmem_freelist_delete(vmp, prevseg);
1261 			prevseg->vs_end = endaddr;
1262 
1263 			vsp = prevseg;
1264 		} else if (nextseg->vs_type == VMEM_FREE) {
1265 			/*
1266 			 * coalesce right
1267 			 */
1268 			ASSERT(nextseg->vs_start == endaddr);
1269 
1270 			VMEM_DELETE(span, k);
1271 			vmem_seg_destroy(vmp, span);
1272 
1273 			vmem_freelist_delete(vmp, nextseg);
1274 			nextseg->vs_start = addr;
1275 
1276 			vsp = nextseg;
1277 		} else {
1278 			/*
1279 			 * cannnot coalesce
1280 			 */
1281 			VMEM_DELETE(span, k);
1282 			span->vs_start = addr;
1283 			span->vs_end = endaddr;
1284 
1285 			vsp = span;
1286 		}
1287 	} else if (span->vs_end == addr) {
1288 		vmem_seg_t *oldseg = span->vs_knext->vs_aprev;
1289 		span->vs_end = endaddr;
1290 
1291 		ASSERT(oldseg->vs_type != VMEM_SPAN);
1292 		if (oldseg->vs_type == VMEM_FREE) {
1293 			ASSERT(oldseg->vs_end == addr);
1294 			vmem_freelist_delete(vmp, oldseg);
1295 			oldseg->vs_end = endaddr;
1296 			vsp = oldseg;
1297 		} else
1298 			vsp = vmem_seg_create(vmp, oldseg, addr, endaddr);
1299 	} else {
1300 		vmem_seg_t *oldseg = span->vs_anext;
1301 		ASSERT(span->vs_start == endaddr);
1302 		span->vs_start = addr;
1303 
1304 		ASSERT(oldseg->vs_type != VMEM_SPAN);
1305 		if (oldseg->vs_type == VMEM_FREE) {
1306 			ASSERT(oldseg->vs_start == endaddr);
1307 			vmem_freelist_delete(vmp, oldseg);
1308 			oldseg->vs_start = addr;
1309 			vsp = oldseg;
1310 		} else
1311 			vsp = vmem_seg_create(vmp, span, addr, endaddr);
1312 	}
1313 	vmem_freelist_insert(vmp, vsp);
1314 	vmp->vm_kstat.vk_mem_total += (endaddr - addr);
1315 	return (vsp);
1316 }
1317 
1318 /*
1319  * Does some error checking, calls vmem_extend_unlocked to add
1320  * [vaddr, vaddr+size) to vmp, then allocates alloc bytes from the
1321  * newly merged segment.
1322  */
1323 void *
1324 _vmem_extend_alloc(vmem_t *vmp, void *vaddr, size_t size, size_t alloc,
1325     int vmflag)
1326 {
1327 	uintptr_t addr = (uintptr_t)vaddr;
1328 	uintptr_t endaddr = addr + size;
1329 	vmem_seg_t *vsp;
1330 
1331 	ASSERT(vaddr != NULL && size != 0 && endaddr > addr);
1332 	ASSERT(alloc <= size && alloc != 0);
1333 	ASSERT(((addr | size | alloc) & (vmp->vm_quantum - 1)) == 0);
1334 
1335 	ASSERT(!vmem_contains(vmp, vaddr, size));
1336 
1337 	(void) mutex_lock(&vmp->vm_lock);
1338 	if (!vmem_populate(vmp, vmflag)) {
1339 		(void) mutex_unlock(&vmp->vm_lock);
1340 		return (NULL);
1341 	}
1342 	/*
1343 	 * if there is a source, we can't mess with the spans
1344 	 */
1345 	if (vmp->vm_source_alloc != NULL)
1346 		vsp = vmem_span_create(vmp, vaddr, size, 0);
1347 	else
1348 		vsp = vmem_extend_unlocked(vmp, addr, endaddr);
1349 
1350 	ASSERT(VS_SIZE(vsp) >= alloc);
1351 
1352 	addr = vsp->vs_start;
1353 	(void) vmem_seg_alloc(vmp, vsp, addr, alloc);
1354 	vaddr = (void *)addr;
1355 
1356 	(void) cond_broadcast(&vmp->vm_cv);
1357 	(void) mutex_unlock(&vmp->vm_lock);
1358 
1359 	return (vaddr);
1360 }
1361 
1362 /*
1363  * Walk the vmp arena, applying func to each segment matching typemask.
1364  * If VMEM_REENTRANT is specified, the arena lock is dropped across each
1365  * call to func(); otherwise, it is held for the duration of vmem_walk()
1366  * to ensure a consistent snapshot.  Note that VMEM_REENTRANT callbacks
1367  * are *not* necessarily consistent, so they may only be used when a hint
1368  * is adequate.
1369  */
1370 void
1371 vmem_walk(vmem_t *vmp, int typemask,
1372     void (*func)(void *, void *, size_t), void *arg)
1373 {
1374 	vmem_seg_t *vsp;
1375 	vmem_seg_t *seg0 = &vmp->vm_seg0;
1376 	vmem_seg_t walker;
1377 
1378 	if (typemask & VMEM_WALKER)
1379 		return;
1380 
1381 	bzero(&walker, sizeof (walker));
1382 	walker.vs_type = VMEM_WALKER;
1383 
1384 	(void) mutex_lock(&vmp->vm_lock);
1385 	VMEM_INSERT(seg0, &walker, a);
1386 	for (vsp = seg0->vs_anext; vsp != seg0; vsp = vsp->vs_anext) {
1387 		if (vsp->vs_type & typemask) {
1388 			void *start = (void *)vsp->vs_start;
1389 			size_t size = VS_SIZE(vsp);
1390 			if (typemask & VMEM_REENTRANT) {
1391 				vmem_advance(vmp, &walker, vsp);
1392 				(void) mutex_unlock(&vmp->vm_lock);
1393 				func(arg, start, size);
1394 				(void) mutex_lock(&vmp->vm_lock);
1395 				vsp = &walker;
1396 			} else {
1397 				func(arg, start, size);
1398 			}
1399 		}
1400 	}
1401 	vmem_advance(vmp, &walker, NULL);
1402 	(void) mutex_unlock(&vmp->vm_lock);
1403 }
1404 
1405 /*
1406  * Return the total amount of memory whose type matches typemask.  Thus:
1407  *
1408  *	typemask VMEM_ALLOC yields total memory allocated (in use).
1409  *	typemask VMEM_FREE yields total memory free (available).
1410  *	typemask (VMEM_ALLOC | VMEM_FREE) yields total arena size.
1411  */
1412 size_t
1413 vmem_size(vmem_t *vmp, int typemask)
1414 {
1415 	uint64_t size = 0;
1416 
1417 	if (typemask & VMEM_ALLOC)
1418 		size += vmp->vm_kstat.vk_mem_inuse;
1419 	if (typemask & VMEM_FREE)
1420 		size += vmp->vm_kstat.vk_mem_total -
1421 		    vmp->vm_kstat.vk_mem_inuse;
1422 	return ((size_t)size);
1423 }
1424 
1425 /*
1426  * Create an arena called name whose initial span is [base, base + size).
1427  * The arena's natural unit of currency is quantum, so vmem_alloc()
1428  * guarantees quantum-aligned results.  The arena may import new spans
1429  * by invoking afunc() on source, and may return those spans by invoking
1430  * ffunc() on source.  To make small allocations fast and scalable,
1431  * the arena offers high-performance caching for each integer multiple
1432  * of quantum up to qcache_max.
1433  */
1434 vmem_t *
1435 vmem_create(const char *name, void *base, size_t size, size_t quantum,
1436     vmem_alloc_t *afunc, vmem_free_t *ffunc, vmem_t *source,
1437     size_t qcache_max, int vmflag)
1438 {
1439 	int i;
1440 	size_t nqcache;
1441 	vmem_t *vmp, *cur, **vmpp;
1442 	vmem_seg_t *vsp;
1443 	vmem_freelist_t *vfp;
1444 	uint32_t id = atomic_add_32_nv(&vmem_id, 1);
1445 
1446 	if (vmem_vmem_arena != NULL) {
1447 		vmp = vmem_alloc(vmem_vmem_arena, sizeof (vmem_t),
1448 		    vmflag & VM_UMFLAGS);
1449 	} else {
1450 		ASSERT(id <= VMEM_INITIAL);
1451 		vmp = &vmem0[id - 1];
1452 	}
1453 
1454 	if (vmp == NULL)
1455 		return (NULL);
1456 	bzero(vmp, sizeof (vmem_t));
1457 
1458 	(void) snprintf(vmp->vm_name, VMEM_NAMELEN, "%s", name);
1459 	(void) mutex_init(&vmp->vm_lock, USYNC_THREAD, NULL);
1460 	(void) cond_init(&vmp->vm_cv, USYNC_THREAD, NULL);
1461 	vmp->vm_cflags = vmflag;
1462 	vmflag &= VM_UMFLAGS;
1463 
1464 	vmp->vm_quantum = quantum;
1465 	vmp->vm_qshift = highbit(quantum) - 1;
1466 	nqcache = MIN(qcache_max >> vmp->vm_qshift, VMEM_NQCACHE_MAX);
1467 
1468 	for (i = 0; i <= VMEM_FREELISTS; i++) {
1469 		vfp = &vmp->vm_freelist[i];
1470 		vfp->vs_end = 1UL << i;
1471 		vfp->vs_knext = (vmem_seg_t *)(vfp + 1);
1472 		vfp->vs_kprev = (vmem_seg_t *)(vfp - 1);
1473 	}
1474 
1475 	vmp->vm_freelist[0].vs_kprev = NULL;
1476 	vmp->vm_freelist[VMEM_FREELISTS].vs_knext = NULL;
1477 	vmp->vm_freelist[VMEM_FREELISTS].vs_end = 0;
1478 	vmp->vm_hash_table = vmp->vm_hash0;
1479 	vmp->vm_hash_mask = VMEM_HASH_INITIAL - 1;
1480 	vmp->vm_hash_shift = highbit(vmp->vm_hash_mask);
1481 
1482 	vsp = &vmp->vm_seg0;
1483 	vsp->vs_anext = vsp;
1484 	vsp->vs_aprev = vsp;
1485 	vsp->vs_knext = vsp;
1486 	vsp->vs_kprev = vsp;
1487 	vsp->vs_type = VMEM_SPAN;
1488 
1489 	vsp = &vmp->vm_rotor;
1490 	vsp->vs_type = VMEM_ROTOR;
1491 	VMEM_INSERT(&vmp->vm_seg0, vsp, a);
1492 
1493 	vmp->vm_id = id;
1494 	if (source != NULL)
1495 		vmp->vm_kstat.vk_source_id = source->vm_id;
1496 	vmp->vm_source = source;
1497 	vmp->vm_source_alloc = afunc;
1498 	vmp->vm_source_free = ffunc;
1499 
1500 	if (nqcache != 0) {
1501 		vmp->vm_qcache_max = nqcache << vmp->vm_qshift;
1502 		for (i = 0; i < nqcache; i++) {
1503 			char buf[VMEM_NAMELEN + 21];
1504 			(void) snprintf(buf, sizeof (buf), "%s_%lu",
1505 			    vmp->vm_name, (long)((i + 1) * quantum));
1506 			vmp->vm_qcache[i] = umem_cache_create(buf,
1507 			    (i + 1) * quantum, quantum, NULL, NULL, NULL,
1508 			    NULL, vmp, UMC_QCACHE | UMC_NOTOUCH);
1509 			if (vmp->vm_qcache[i] == NULL) {
1510 				vmp->vm_qcache_max = i * quantum;
1511 				break;
1512 			}
1513 		}
1514 	}
1515 
1516 	(void) mutex_lock(&vmem_list_lock);
1517 	vmpp = &vmem_list;
1518 	while ((cur = *vmpp) != NULL)
1519 		vmpp = &cur->vm_next;
1520 	*vmpp = vmp;
1521 	(void) mutex_unlock(&vmem_list_lock);
1522 
1523 	if (vmp->vm_cflags & VMC_POPULATOR) {
1524 		uint_t pop_id = atomic_add_32_nv(&vmem_populators, 1);
1525 		ASSERT(pop_id <= VMEM_INITIAL);
1526 		vmem_populator[pop_id - 1] = vmp;
1527 		(void) mutex_lock(&vmp->vm_lock);
1528 		(void) vmem_populate(vmp, vmflag | VM_PANIC);
1529 		(void) mutex_unlock(&vmp->vm_lock);
1530 	}
1531 
1532 	if ((base || size) && vmem_add(vmp, base, size, vmflag) == NULL) {
1533 		vmem_destroy(vmp);
1534 		return (NULL);
1535 	}
1536 
1537 	return (vmp);
1538 }
1539 
1540 /*
1541  * Destroy arena vmp.
1542  */
1543 void
1544 vmem_destroy(vmem_t *vmp)
1545 {
1546 	vmem_t *cur, **vmpp;
1547 	vmem_seg_t *seg0 = &vmp->vm_seg0;
1548 	vmem_seg_t *vsp;
1549 	size_t leaked;
1550 	int i;
1551 
1552 	(void) mutex_lock(&vmem_list_lock);
1553 	vmpp = &vmem_list;
1554 	while ((cur = *vmpp) != vmp)
1555 		vmpp = &cur->vm_next;
1556 	*vmpp = vmp->vm_next;
1557 	(void) mutex_unlock(&vmem_list_lock);
1558 
1559 	for (i = 0; i < VMEM_NQCACHE_MAX; i++)
1560 		if (vmp->vm_qcache[i])
1561 			umem_cache_destroy(vmp->vm_qcache[i]);
1562 
1563 	leaked = vmem_size(vmp, VMEM_ALLOC);
1564 	if (leaked != 0)
1565 		umem_printf("vmem_destroy('%s'): leaked %lu bytes",
1566 		    vmp->vm_name, leaked);
1567 
1568 	if (vmp->vm_hash_table != vmp->vm_hash0)
1569 		vmem_free(vmem_hash_arena, vmp->vm_hash_table,
1570 		    (vmp->vm_hash_mask + 1) * sizeof (void *));
1571 
1572 	/*
1573 	 * Give back the segment structures for anything that's left in the
1574 	 * arena, e.g. the primary spans and their free segments.
1575 	 */
1576 	VMEM_DELETE(&vmp->vm_rotor, a);
1577 	for (vsp = seg0->vs_anext; vsp != seg0; vsp = vsp->vs_anext)
1578 		vmem_putseg_global(vsp);
1579 
1580 	while (vmp->vm_nsegfree > 0)
1581 		vmem_putseg_global(vmem_getseg(vmp));
1582 
1583 	(void) mutex_destroy(&vmp->vm_lock);
1584 	(void) cond_destroy(&vmp->vm_cv);
1585 	vmem_free(vmem_vmem_arena, vmp, sizeof (vmem_t));
1586 }
1587 
1588 /*
1589  * Resize vmp's hash table to keep the average lookup depth near 1.0.
1590  */
1591 static void
1592 vmem_hash_rescale(vmem_t *vmp)
1593 {
1594 	vmem_seg_t **old_table, **new_table, *vsp;
1595 	size_t old_size, new_size, h, nseg;
1596 
1597 	nseg = (size_t)(vmp->vm_kstat.vk_alloc - vmp->vm_kstat.vk_free);
1598 
1599 	new_size = MAX(VMEM_HASH_INITIAL, 1 << (highbit(3 * nseg + 4) - 2));
1600 	old_size = vmp->vm_hash_mask + 1;
1601 
1602 	if ((old_size >> 1) <= new_size && new_size <= (old_size << 1))
1603 		return;
1604 
1605 	new_table = vmem_alloc(vmem_hash_arena, new_size * sizeof (void *),
1606 	    VM_NOSLEEP);
1607 	if (new_table == NULL)
1608 		return;
1609 	bzero(new_table, new_size * sizeof (void *));
1610 
1611 	(void) mutex_lock(&vmp->vm_lock);
1612 
1613 	old_size = vmp->vm_hash_mask + 1;
1614 	old_table = vmp->vm_hash_table;
1615 
1616 	vmp->vm_hash_mask = new_size - 1;
1617 	vmp->vm_hash_table = new_table;
1618 	vmp->vm_hash_shift = highbit(vmp->vm_hash_mask);
1619 
1620 	for (h = 0; h < old_size; h++) {
1621 		vsp = old_table[h];
1622 		while (vsp != NULL) {
1623 			uintptr_t addr = vsp->vs_start;
1624 			vmem_seg_t *next_vsp = vsp->vs_knext;
1625 			vmem_seg_t **hash_bucket = VMEM_HASH(vmp, addr);
1626 			vsp->vs_knext = *hash_bucket;
1627 			*hash_bucket = vsp;
1628 			vsp = next_vsp;
1629 		}
1630 	}
1631 
1632 	(void) mutex_unlock(&vmp->vm_lock);
1633 
1634 	if (old_table != vmp->vm_hash0)
1635 		vmem_free(vmem_hash_arena, old_table,
1636 		    old_size * sizeof (void *));
1637 }
1638 
1639 /*
1640  * Perform periodic maintenance on all vmem arenas.
1641  */
1642 /*ARGSUSED*/
1643 void
1644 vmem_update(void *dummy)
1645 {
1646 	vmem_t *vmp;
1647 
1648 	(void) mutex_lock(&vmem_list_lock);
1649 	for (vmp = vmem_list; vmp != NULL; vmp = vmp->vm_next) {
1650 		/*
1651 		 * If threads are waiting for resources, wake them up
1652 		 * periodically so they can issue another vmem_reap()
1653 		 * to reclaim resources cached by the slab allocator.
1654 		 */
1655 		(void) cond_broadcast(&vmp->vm_cv);
1656 
1657 		/*
1658 		 * Rescale the hash table to keep the hash chains short.
1659 		 */
1660 		vmem_hash_rescale(vmp);
1661 	}
1662 	(void) mutex_unlock(&vmem_list_lock);
1663 }
1664 
1665 /*
1666  * If vmem_init is called again, we need to be able to reset the world.
1667  * That includes resetting the statics back to their original values.
1668  */
1669 void
1670 vmem_startup(void)
1671 {
1672 #ifdef UMEM_STANDALONE
1673 	vmem_id = 0;
1674 	vmem_populators = 0;
1675 	vmem_segfree = NULL;
1676 	vmem_list = NULL;
1677 	vmem_internal_arena = NULL;
1678 	vmem_seg_arena = NULL;
1679 	vmem_hash_arena = NULL;
1680 	vmem_vmem_arena = NULL;
1681 	vmem_heap = NULL;
1682 	vmem_heap_alloc = NULL;
1683 	vmem_heap_free = NULL;
1684 
1685 	bzero(vmem0, sizeof (vmem0));
1686 	bzero(vmem_populator, sizeof (vmem_populator));
1687 	bzero(vmem_seg0, sizeof (vmem_seg0));
1688 #endif
1689 }
1690 
1691 /*
1692  * Prepare vmem for use.
1693  */
1694 vmem_t *
1695 vmem_init(const char *parent_name, size_t parent_quantum,
1696     vmem_alloc_t *parent_alloc, vmem_free_t *parent_free,
1697     const char *heap_name, void *heap_start, size_t heap_size,
1698     size_t heap_quantum, vmem_alloc_t *heap_alloc, vmem_free_t *heap_free)
1699 {
1700 	uint32_t id;
1701 	int nseg = VMEM_SEG_INITIAL;
1702 	vmem_t *parent, *heap;
1703 
1704 	ASSERT(vmem_internal_arena == NULL);
1705 
1706 	while (--nseg >= 0)
1707 		vmem_putseg_global(&vmem_seg0[nseg]);
1708 
1709 	if (parent_name != NULL) {
1710 		parent = vmem_create(parent_name,
1711 		    heap_start, heap_size, parent_quantum,
1712 		    NULL, NULL, NULL, 0,
1713 		    VM_SLEEP | VMC_POPULATOR);
1714 		heap_start = NULL;
1715 		heap_size = 0;
1716 	} else {
1717 		ASSERT(parent_alloc == NULL && parent_free == NULL);
1718 		parent = NULL;
1719 	}
1720 
1721 	heap = vmem_create(heap_name,
1722 	    heap_start, heap_size, heap_quantum,
1723 	    parent_alloc, parent_free, parent, 0,
1724 	    VM_SLEEP | VMC_POPULATOR);
1725 
1726 	vmem_heap = heap;
1727 	vmem_heap_alloc = heap_alloc;
1728 	vmem_heap_free = heap_free;
1729 
1730 	vmem_internal_arena = vmem_create("vmem_internal",
1731 	    NULL, 0, heap_quantum,
1732 	    heap_alloc, heap_free, heap, 0,
1733 	    VM_SLEEP | VMC_POPULATOR);
1734 
1735 	vmem_seg_arena = vmem_create("vmem_seg",
1736 	    NULL, 0, heap_quantum,
1737 	    vmem_alloc, vmem_free, vmem_internal_arena, 0,
1738 	    VM_SLEEP | VMC_POPULATOR);
1739 
1740 	vmem_hash_arena = vmem_create("vmem_hash",
1741 	    NULL, 0, 8,
1742 	    vmem_alloc, vmem_free, vmem_internal_arena, 0,
1743 	    VM_SLEEP);
1744 
1745 	vmem_vmem_arena = vmem_create("vmem_vmem",
1746 	    vmem0, sizeof (vmem0), 1,
1747 	    vmem_alloc, vmem_free, vmem_internal_arena, 0,
1748 	    VM_SLEEP);
1749 
1750 	for (id = 0; id < vmem_id; id++)
1751 		(void) vmem_xalloc(vmem_vmem_arena, sizeof (vmem_t),
1752 		    1, 0, 0, &vmem0[id], &vmem0[id + 1],
1753 		    VM_NOSLEEP | VM_BESTFIT | VM_PANIC);
1754 
1755 	return (heap);
1756 }
1757 
1758 void
1759 vmem_no_debug(void)
1760 {
1761 	/*
1762 	 * This size must be a multiple of the minimum required alignment,
1763 	 * since vmem_populate allocates them compactly.
1764 	 */
1765 	vmem_seg_size = P2ROUNDUP(offsetof(vmem_seg_t, vs_thread),
1766 	    sizeof (hrtime_t));
1767 }
1768 
1769 /*
1770  * Lockup and release, for fork1(2) handling.
1771  */
1772 void
1773 vmem_lockup(void)
1774 {
1775 	vmem_t *cur;
1776 
1777 	(void) mutex_lock(&vmem_list_lock);
1778 	(void) mutex_lock(&vmem_nosleep_lock.vmpl_mutex);
1779 
1780 	/*
1781 	 * Lock up and broadcast all arenas.
1782 	 */
1783 	for (cur = vmem_list; cur != NULL; cur = cur->vm_next) {
1784 		(void) mutex_lock(&cur->vm_lock);
1785 		(void) cond_broadcast(&cur->vm_cv);
1786 	}
1787 
1788 	(void) mutex_lock(&vmem_segfree_lock);
1789 }
1790 
1791 void
1792 vmem_release(void)
1793 {
1794 	vmem_t *cur;
1795 
1796 	(void) mutex_unlock(&vmem_nosleep_lock.vmpl_mutex);
1797 
1798 	for (cur = vmem_list; cur != NULL; cur = cur->vm_next)
1799 		(void) mutex_unlock(&cur->vm_lock);
1800 
1801 	(void) mutex_unlock(&vmem_segfree_lock);
1802 	(void) mutex_unlock(&vmem_list_lock);
1803 }
1804