xref: /netbsd-src/sys/kern/subr_vmem.c (revision 9aa0541bdf64142d9a27c2cf274394d60182818f)
1 /*	$NetBSD: subr_vmem.c,v 1.62 2011/10/02 21:32:48 rmind Exp $	*/
2 
3 /*-
4  * Copyright (c)2006,2007,2008,2009 YAMAMOTO Takashi,
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions and the following disclaimer.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
20  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  */
28 
29 /*
30  * reference:
31  * -	Magazines and Vmem: Extending the Slab Allocator
32  *	to Many CPUs and Arbitrary Resources
33  *	http://www.usenix.org/event/usenix01/bonwick.html
34  *
35  * todo:
36  * -	decide how to import segments for vmem_xalloc.
37  * -	don't rely on malloc(9).
38  */
39 
40 #include <sys/cdefs.h>
41 __KERNEL_RCSID(0, "$NetBSD: subr_vmem.c,v 1.62 2011/10/02 21:32:48 rmind Exp $");
42 
43 #if defined(_KERNEL)
44 #include "opt_ddb.h"
45 #define	QCACHE
46 #endif /* defined(_KERNEL) */
47 
48 #include <sys/param.h>
49 #include <sys/hash.h>
50 #include <sys/queue.h>
51 #include <sys/bitops.h>
52 
53 #if defined(_KERNEL)
54 #include <sys/systm.h>
55 #include <sys/kernel.h>	/* hz */
56 #include <sys/callout.h>
57 #include <sys/malloc.h>
58 #include <sys/once.h>
59 #include <sys/pool.h>
60 #include <sys/vmem.h>
61 #include <sys/workqueue.h>
62 #else /* defined(_KERNEL) */
63 #include "../sys/vmem.h"
64 #endif /* defined(_KERNEL) */
65 
66 #if defined(_KERNEL)
67 #define	LOCK_DECL(name)		\
68     kmutex_t name; char lockpad[COHERENCY_UNIT - sizeof(kmutex_t)]
69 #else /* defined(_KERNEL) */
70 #include <errno.h>
71 #include <assert.h>
72 #include <stdlib.h>
73 
74 #define	UNITTEST
75 #define	KASSERT(a)		assert(a)
76 #define	LOCK_DECL(name)		/* nothing */
77 #define	mutex_init(a, b, c)	/* nothing */
78 #define	mutex_destroy(a)	/* nothing */
79 #define	mutex_enter(a)		/* nothing */
80 #define	mutex_tryenter(a)	true
81 #define	mutex_exit(a)		/* nothing */
82 #define	mutex_owned(a)		/* nothing */
83 #define	ASSERT_SLEEPABLE()	/* nothing */
84 #define	panic(...)		printf(__VA_ARGS__); abort()
85 #endif /* defined(_KERNEL) */
86 
87 struct vmem;
88 struct vmem_btag;
89 
90 #if defined(VMEM_SANITY)
91 static void vmem_check(vmem_t *);
92 #else /* defined(VMEM_SANITY) */
93 #define vmem_check(vm)	/* nothing */
94 #endif /* defined(VMEM_SANITY) */
95 
96 #define	VMEM_MAXORDER		(sizeof(vmem_size_t) * CHAR_BIT)
97 
98 #define	VMEM_HASHSIZE_MIN	1	/* XXX */
99 #define	VMEM_HASHSIZE_MAX	65536	/* XXX */
100 #define	VMEM_HASHSIZE_INIT	128
101 
102 #define	VM_FITMASK	(VM_BESTFIT | VM_INSTANTFIT)
103 
104 CIRCLEQ_HEAD(vmem_seglist, vmem_btag);
105 LIST_HEAD(vmem_freelist, vmem_btag);
106 LIST_HEAD(vmem_hashlist, vmem_btag);
107 
108 #if defined(QCACHE)
109 #define	VMEM_QCACHE_IDX_MAX	32
110 
111 #define	QC_NAME_MAX	16
112 
113 struct qcache {
114 	pool_cache_t qc_cache;
115 	vmem_t *qc_vmem;
116 	char qc_name[QC_NAME_MAX];
117 };
118 typedef struct qcache qcache_t;
119 #define	QC_POOL_TO_QCACHE(pool)	((qcache_t *)(pool->pr_qcache))
120 #endif /* defined(QCACHE) */
121 
122 /* vmem arena */
123 struct vmem {
124 	LOCK_DECL(vm_lock);
125 	int (*vm_importfn)(void *, vmem_size_t, vmem_size_t *,
126 	    vm_flag_t, vmem_addr_t *);
127 	void (*vm_releasefn)(void *, vmem_addr_t, vmem_size_t);
128 	vmem_t *vm_source;
129 	void *vm_arg;
130 	struct vmem_seglist vm_seglist;
131 	struct vmem_freelist vm_freelist[VMEM_MAXORDER];
132 	size_t vm_hashsize;
133 	size_t vm_nbusytag;
134 	struct vmem_hashlist *vm_hashlist;
135 	size_t vm_quantum_mask;
136 	int vm_quantum_shift;
137 	const char *vm_name;
138 	LIST_ENTRY(vmem) vm_alllist;
139 
140 #if defined(QCACHE)
141 	/* quantum cache */
142 	size_t vm_qcache_max;
143 	struct pool_allocator vm_qcache_allocator;
144 	qcache_t vm_qcache_store[VMEM_QCACHE_IDX_MAX];
145 	qcache_t *vm_qcache[VMEM_QCACHE_IDX_MAX];
146 #endif /* defined(QCACHE) */
147 };
148 
149 #define	VMEM_LOCK(vm)		mutex_enter(&vm->vm_lock)
150 #define	VMEM_TRYLOCK(vm)	mutex_tryenter(&vm->vm_lock)
151 #define	VMEM_UNLOCK(vm)		mutex_exit(&vm->vm_lock)
152 #define	VMEM_LOCK_INIT(vm, ipl)	mutex_init(&vm->vm_lock, MUTEX_DEFAULT, ipl)
153 #define	VMEM_LOCK_DESTROY(vm)	mutex_destroy(&vm->vm_lock)
154 #define	VMEM_ASSERT_LOCKED(vm)	KASSERT(mutex_owned(&vm->vm_lock))
155 
156 /* boundary tag */
157 struct vmem_btag {
158 	CIRCLEQ_ENTRY(vmem_btag) bt_seglist;
159 	union {
160 		LIST_ENTRY(vmem_btag) u_freelist; /* BT_TYPE_FREE */
161 		LIST_ENTRY(vmem_btag) u_hashlist; /* BT_TYPE_BUSY */
162 	} bt_u;
163 #define	bt_hashlist	bt_u.u_hashlist
164 #define	bt_freelist	bt_u.u_freelist
165 	vmem_addr_t bt_start;
166 	vmem_size_t bt_size;
167 	int bt_type;
168 };
169 
170 #define	BT_TYPE_SPAN		1
171 #define	BT_TYPE_SPAN_STATIC	2
172 #define	BT_TYPE_FREE		3
173 #define	BT_TYPE_BUSY		4
174 #define	BT_ISSPAN_P(bt)	((bt)->bt_type <= BT_TYPE_SPAN_STATIC)
175 
176 #define	BT_END(bt)	((bt)->bt_start + (bt)->bt_size - 1)
177 
178 typedef struct vmem_btag bt_t;
179 
180 /* ---- misc */
181 
182 #define	VMEM_ALIGNUP(addr, align) \
183 	(-(-(addr) & -(align)))
184 
185 #define	VMEM_CROSS_P(addr1, addr2, boundary) \
186 	((((addr1) ^ (addr2)) & -(boundary)) != 0)
187 
188 #define	ORDER2SIZE(order)	((vmem_size_t)1 << (order))
189 #define	SIZE2ORDER(size)	((int)ilog2(size))
190 
191 #if !defined(_KERNEL)
192 #define	xmalloc(sz, flags)	malloc(sz)
193 #define	xfree(p)		free(p)
194 #define	bt_alloc(vm, flags)	malloc(sizeof(bt_t))
195 #define	bt_free(vm, bt)		free(bt)
196 #else	/* !defined(_KERNEL) */
197 
198 static MALLOC_DEFINE(M_VMEM, "vmem", "vmem");
199 
200 static inline void *
201 xmalloc(size_t sz, vm_flag_t flags)
202 {
203 	return malloc(sz, M_VMEM,
204 	    M_CANFAIL | ((flags & VM_SLEEP) ? M_WAITOK : M_NOWAIT));
205 }
206 
207 static inline void
208 xfree(void *p)
209 {
210 	return free(p, M_VMEM);
211 }
212 
213 /* ---- boundary tag */
214 
215 static struct pool_cache bt_cache;
216 
217 static inline bt_t *
218 bt_alloc(vmem_t *vm, vm_flag_t flags)
219 {
220 	return pool_cache_get(&bt_cache,
221 	    (flags & VM_SLEEP) ? PR_WAITOK : PR_NOWAIT);
222 }
223 
224 static inline void
225 bt_free(vmem_t *vm, bt_t *bt)
226 {
227 	pool_cache_put(&bt_cache, bt);
228 }
229 
230 #endif	/* !defined(_KERNEL) */
231 
232 /*
233  * freelist[0] ... [1, 1]
234  * freelist[1] ... [2, 3]
235  * freelist[2] ... [4, 7]
236  * freelist[3] ... [8, 15]
237  *  :
238  * freelist[n] ... [(1 << n), (1 << (n + 1)) - 1]
239  *  :
240  */
241 
242 static struct vmem_freelist *
243 bt_freehead_tofree(vmem_t *vm, vmem_size_t size)
244 {
245 	const vmem_size_t qsize = size >> vm->vm_quantum_shift;
246 	const int idx = SIZE2ORDER(qsize);
247 
248 	KASSERT(size != 0 && qsize != 0);
249 	KASSERT((size & vm->vm_quantum_mask) == 0);
250 	KASSERT(idx >= 0);
251 	KASSERT(idx < VMEM_MAXORDER);
252 
253 	return &vm->vm_freelist[idx];
254 }
255 
256 /*
257  * bt_freehead_toalloc: return the freelist for the given size and allocation
258  * strategy.
259  *
260  * for VM_INSTANTFIT, return the list in which any blocks are large enough
261  * for the requested size.  otherwise, return the list which can have blocks
262  * large enough for the requested size.
263  */
264 
265 static struct vmem_freelist *
266 bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, vm_flag_t strat)
267 {
268 	const vmem_size_t qsize = size >> vm->vm_quantum_shift;
269 	int idx = SIZE2ORDER(qsize);
270 
271 	KASSERT(size != 0 && qsize != 0);
272 	KASSERT((size & vm->vm_quantum_mask) == 0);
273 
274 	if (strat == VM_INSTANTFIT && ORDER2SIZE(idx) != qsize) {
275 		idx++;
276 		/* check too large request? */
277 	}
278 	KASSERT(idx >= 0);
279 	KASSERT(idx < VMEM_MAXORDER);
280 
281 	return &vm->vm_freelist[idx];
282 }
283 
284 /* ---- boundary tag hash */
285 
286 static struct vmem_hashlist *
287 bt_hashhead(vmem_t *vm, vmem_addr_t addr)
288 {
289 	struct vmem_hashlist *list;
290 	unsigned int hash;
291 
292 	hash = hash32_buf(&addr, sizeof(addr), HASH32_BUF_INIT);
293 	list = &vm->vm_hashlist[hash % vm->vm_hashsize];
294 
295 	return list;
296 }
297 
298 static bt_t *
299 bt_lookupbusy(vmem_t *vm, vmem_addr_t addr)
300 {
301 	struct vmem_hashlist *list;
302 	bt_t *bt;
303 
304 	list = bt_hashhead(vm, addr);
305 	LIST_FOREACH(bt, list, bt_hashlist) {
306 		if (bt->bt_start == addr) {
307 			break;
308 		}
309 	}
310 
311 	return bt;
312 }
313 
314 static void
315 bt_rembusy(vmem_t *vm, bt_t *bt)
316 {
317 
318 	KASSERT(vm->vm_nbusytag > 0);
319 	vm->vm_nbusytag--;
320 	LIST_REMOVE(bt, bt_hashlist);
321 }
322 
323 static void
324 bt_insbusy(vmem_t *vm, bt_t *bt)
325 {
326 	struct vmem_hashlist *list;
327 
328 	KASSERT(bt->bt_type == BT_TYPE_BUSY);
329 
330 	list = bt_hashhead(vm, bt->bt_start);
331 	LIST_INSERT_HEAD(list, bt, bt_hashlist);
332 	vm->vm_nbusytag++;
333 }
334 
335 /* ---- boundary tag list */
336 
337 static void
338 bt_remseg(vmem_t *vm, bt_t *bt)
339 {
340 
341 	CIRCLEQ_REMOVE(&vm->vm_seglist, bt, bt_seglist);
342 }
343 
344 static void
345 bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev)
346 {
347 
348 	CIRCLEQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist);
349 }
350 
351 static void
352 bt_insseg_tail(vmem_t *vm, bt_t *bt)
353 {
354 
355 	CIRCLEQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist);
356 }
357 
358 static void
359 bt_remfree(vmem_t *vm, bt_t *bt)
360 {
361 
362 	KASSERT(bt->bt_type == BT_TYPE_FREE);
363 
364 	LIST_REMOVE(bt, bt_freelist);
365 }
366 
367 static void
368 bt_insfree(vmem_t *vm, bt_t *bt)
369 {
370 	struct vmem_freelist *list;
371 
372 	list = bt_freehead_tofree(vm, bt->bt_size);
373 	LIST_INSERT_HEAD(list, bt, bt_freelist);
374 }
375 
376 /* ---- vmem internal functions */
377 
378 #if defined(_KERNEL)
379 static kmutex_t vmem_list_lock;
380 static LIST_HEAD(, vmem) vmem_list = LIST_HEAD_INITIALIZER(vmem_list);
381 #endif /* defined(_KERNEL) */
382 
383 #if defined(QCACHE)
384 static inline vm_flag_t
385 prf_to_vmf(int prflags)
386 {
387 	vm_flag_t vmflags;
388 
389 	KASSERT((prflags & ~(PR_LIMITFAIL | PR_WAITOK | PR_NOWAIT)) == 0);
390 	if ((prflags & PR_WAITOK) != 0) {
391 		vmflags = VM_SLEEP;
392 	} else {
393 		vmflags = VM_NOSLEEP;
394 	}
395 	return vmflags;
396 }
397 
398 static inline int
399 vmf_to_prf(vm_flag_t vmflags)
400 {
401 	int prflags;
402 
403 	if ((vmflags & VM_SLEEP) != 0) {
404 		prflags = PR_WAITOK;
405 	} else {
406 		prflags = PR_NOWAIT;
407 	}
408 	return prflags;
409 }
410 
411 static size_t
412 qc_poolpage_size(size_t qcache_max)
413 {
414 	int i;
415 
416 	for (i = 0; ORDER2SIZE(i) <= qcache_max * 3; i++) {
417 		/* nothing */
418 	}
419 	return ORDER2SIZE(i);
420 }
421 
422 static void *
423 qc_poolpage_alloc(struct pool *pool, int prflags)
424 {
425 	qcache_t *qc = QC_POOL_TO_QCACHE(pool);
426 	vmem_t *vm = qc->qc_vmem;
427 	vmem_addr_t addr;
428 
429 	if (vmem_alloc(vm, pool->pr_alloc->pa_pagesz,
430 	    prf_to_vmf(prflags) | VM_INSTANTFIT, &addr) != 0)
431 		return NULL;
432 	return (void *)addr;
433 }
434 
435 static void
436 qc_poolpage_free(struct pool *pool, void *addr)
437 {
438 	qcache_t *qc = QC_POOL_TO_QCACHE(pool);
439 	vmem_t *vm = qc->qc_vmem;
440 
441 	vmem_free(vm, (vmem_addr_t)addr, pool->pr_alloc->pa_pagesz);
442 }
443 
444 static void
445 qc_init(vmem_t *vm, size_t qcache_max, int ipl)
446 {
447 	qcache_t *prevqc;
448 	struct pool_allocator *pa;
449 	int qcache_idx_max;
450 	int i;
451 
452 	KASSERT((qcache_max & vm->vm_quantum_mask) == 0);
453 	if (qcache_max > (VMEM_QCACHE_IDX_MAX << vm->vm_quantum_shift)) {
454 		qcache_max = VMEM_QCACHE_IDX_MAX << vm->vm_quantum_shift;
455 	}
456 	vm->vm_qcache_max = qcache_max;
457 	pa = &vm->vm_qcache_allocator;
458 	memset(pa, 0, sizeof(*pa));
459 	pa->pa_alloc = qc_poolpage_alloc;
460 	pa->pa_free = qc_poolpage_free;
461 	pa->pa_pagesz = qc_poolpage_size(qcache_max);
462 
463 	qcache_idx_max = qcache_max >> vm->vm_quantum_shift;
464 	prevqc = NULL;
465 	for (i = qcache_idx_max; i > 0; i--) {
466 		qcache_t *qc = &vm->vm_qcache_store[i - 1];
467 		size_t size = i << vm->vm_quantum_shift;
468 
469 		qc->qc_vmem = vm;
470 		snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu",
471 		    vm->vm_name, size);
472 		qc->qc_cache = pool_cache_init(size,
473 		    ORDER2SIZE(vm->vm_quantum_shift), 0,
474 		    PR_NOALIGN | PR_NOTOUCH /* XXX */,
475 		    qc->qc_name, pa, ipl, NULL, NULL, NULL);
476 		KASSERT(qc->qc_cache != NULL);	/* XXX */
477 		if (prevqc != NULL &&
478 		    qc->qc_cache->pc_pool.pr_itemsperpage ==
479 		    prevqc->qc_cache->pc_pool.pr_itemsperpage) {
480 			pool_cache_destroy(qc->qc_cache);
481 			vm->vm_qcache[i - 1] = prevqc;
482 			continue;
483 		}
484 		qc->qc_cache->pc_pool.pr_qcache = qc;
485 		vm->vm_qcache[i - 1] = qc;
486 		prevqc = qc;
487 	}
488 }
489 
490 static void
491 qc_destroy(vmem_t *vm)
492 {
493 	const qcache_t *prevqc;
494 	int i;
495 	int qcache_idx_max;
496 
497 	qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
498 	prevqc = NULL;
499 	for (i = 0; i < qcache_idx_max; i++) {
500 		qcache_t *qc = vm->vm_qcache[i];
501 
502 		if (prevqc == qc) {
503 			continue;
504 		}
505 		pool_cache_destroy(qc->qc_cache);
506 		prevqc = qc;
507 	}
508 }
509 
510 static bool
511 qc_reap(vmem_t *vm)
512 {
513 	const qcache_t *prevqc;
514 	int i;
515 	int qcache_idx_max;
516 	bool didsomething = false;
517 
518 	qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
519 	prevqc = NULL;
520 	for (i = 0; i < qcache_idx_max; i++) {
521 		qcache_t *qc = vm->vm_qcache[i];
522 
523 		if (prevqc == qc) {
524 			continue;
525 		}
526 		if (pool_cache_reclaim(qc->qc_cache) != 0) {
527 			didsomething = true;
528 		}
529 		prevqc = qc;
530 	}
531 
532 	return didsomething;
533 }
534 #endif /* defined(QCACHE) */
535 
536 #if defined(_KERNEL)
537 static int
538 vmem_init(void)
539 {
540 
541 	mutex_init(&vmem_list_lock, MUTEX_DEFAULT, IPL_NONE);
542 	pool_cache_bootstrap(&bt_cache, sizeof(bt_t), 0, 0, 0, "vmembt",
543 	    NULL, IPL_VM, NULL, NULL, NULL);
544 	return 0;
545 }
546 #endif /* defined(_KERNEL) */
547 
548 static int
549 vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, vm_flag_t flags,
550     int spanbttype)
551 {
552 	bt_t *btspan;
553 	bt_t *btfree;
554 
555 	KASSERT((flags & (VM_SLEEP|VM_NOSLEEP)) != 0);
556 	KASSERT((~flags & (VM_SLEEP|VM_NOSLEEP)) != 0);
557 	KASSERT(spanbttype == BT_TYPE_SPAN ||
558 	    spanbttype == BT_TYPE_SPAN_STATIC);
559 
560 	btspan = bt_alloc(vm, flags);
561 	if (btspan == NULL) {
562 		return ENOMEM;
563 	}
564 	btfree = bt_alloc(vm, flags);
565 	if (btfree == NULL) {
566 		bt_free(vm, btspan);
567 		return ENOMEM;
568 	}
569 
570 	btspan->bt_type = spanbttype;
571 	btspan->bt_start = addr;
572 	btspan->bt_size = size;
573 
574 	btfree->bt_type = BT_TYPE_FREE;
575 	btfree->bt_start = addr;
576 	btfree->bt_size = size;
577 
578 	VMEM_LOCK(vm);
579 	bt_insseg_tail(vm, btspan);
580 	bt_insseg(vm, btfree, btspan);
581 	bt_insfree(vm, btfree);
582 	VMEM_UNLOCK(vm);
583 
584 	return 0;
585 }
586 
587 static void
588 vmem_destroy1(vmem_t *vm)
589 {
590 
591 #if defined(QCACHE)
592 	qc_destroy(vm);
593 #endif /* defined(QCACHE) */
594 	if (vm->vm_hashlist != NULL) {
595 		int i;
596 
597 		for (i = 0; i < vm->vm_hashsize; i++) {
598 			bt_t *bt;
599 
600 			while ((bt = LIST_FIRST(&vm->vm_hashlist[i])) != NULL) {
601 				KASSERT(bt->bt_type == BT_TYPE_SPAN_STATIC);
602 				bt_free(vm, bt);
603 			}
604 		}
605 		xfree(vm->vm_hashlist);
606 	}
607 	VMEM_LOCK_DESTROY(vm);
608 	xfree(vm);
609 }
610 
611 static int
612 vmem_import(vmem_t *vm, vmem_size_t size, vm_flag_t flags)
613 {
614 	vmem_addr_t addr;
615 	int rc;
616 
617 	if (vm->vm_importfn == NULL) {
618 		return EINVAL;
619 	}
620 
621 	rc = (*vm->vm_importfn)(vm->vm_arg, size, &size, flags, &addr);
622 	if (rc != 0) {
623 		return ENOMEM;
624 	}
625 
626 	if (vmem_add1(vm, addr, size, flags, BT_TYPE_SPAN) != 0) {
627 		(*vm->vm_releasefn)(vm->vm_arg, addr, size);
628 		return ENOMEM;
629 	}
630 
631 	return 0;
632 }
633 
634 static int
635 vmem_rehash(vmem_t *vm, size_t newhashsize, vm_flag_t flags)
636 {
637 	bt_t *bt;
638 	int i;
639 	struct vmem_hashlist *newhashlist;
640 	struct vmem_hashlist *oldhashlist;
641 	size_t oldhashsize;
642 
643 	KASSERT(newhashsize > 0);
644 
645 	newhashlist =
646 	    xmalloc(sizeof(struct vmem_hashlist *) * newhashsize, flags);
647 	if (newhashlist == NULL) {
648 		return ENOMEM;
649 	}
650 	for (i = 0; i < newhashsize; i++) {
651 		LIST_INIT(&newhashlist[i]);
652 	}
653 
654 	if (!VMEM_TRYLOCK(vm)) {
655 		xfree(newhashlist);
656 		return EBUSY;
657 	}
658 	oldhashlist = vm->vm_hashlist;
659 	oldhashsize = vm->vm_hashsize;
660 	vm->vm_hashlist = newhashlist;
661 	vm->vm_hashsize = newhashsize;
662 	if (oldhashlist == NULL) {
663 		VMEM_UNLOCK(vm);
664 		return 0;
665 	}
666 	for (i = 0; i < oldhashsize; i++) {
667 		while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) {
668 			bt_rembusy(vm, bt); /* XXX */
669 			bt_insbusy(vm, bt);
670 		}
671 	}
672 	VMEM_UNLOCK(vm);
673 
674 	xfree(oldhashlist);
675 
676 	return 0;
677 }
678 
679 /*
680  * vmem_fit: check if a bt can satisfy the given restrictions.
681  *
682  * it's a caller's responsibility to ensure the region is big enough
683  * before calling us.
684  */
685 
686 static int
687 vmem_fit(const bt_t const *bt, vmem_size_t size, vmem_size_t align,
688     vmem_size_t phase, vmem_size_t nocross,
689     vmem_addr_t minaddr, vmem_addr_t maxaddr, vmem_addr_t *addrp)
690 {
691 	vmem_addr_t start;
692 	vmem_addr_t end;
693 
694 	KASSERT(size > 0);
695 	KASSERT(bt->bt_size >= size); /* caller's responsibility */
696 
697 	/*
698 	 * XXX assumption: vmem_addr_t and vmem_size_t are
699 	 * unsigned integer of the same size.
700 	 */
701 
702 	start = bt->bt_start;
703 	if (start < minaddr) {
704 		start = minaddr;
705 	}
706 	end = BT_END(bt);
707 	if (end > maxaddr) {
708 		end = maxaddr;
709 	}
710 	if (start > end) {
711 		return ENOMEM;
712 	}
713 
714 	start = VMEM_ALIGNUP(start - phase, align) + phase;
715 	if (start < bt->bt_start) {
716 		start += align;
717 	}
718 	if (VMEM_CROSS_P(start, start + size - 1, nocross)) {
719 		KASSERT(align < nocross);
720 		start = VMEM_ALIGNUP(start - phase, nocross) + phase;
721 	}
722 	if (start <= end && end - start >= size - 1) {
723 		KASSERT((start & (align - 1)) == phase);
724 		KASSERT(!VMEM_CROSS_P(start, start + size - 1, nocross));
725 		KASSERT(minaddr <= start);
726 		KASSERT(maxaddr == 0 || start + size - 1 <= maxaddr);
727 		KASSERT(bt->bt_start <= start);
728 		KASSERT(BT_END(bt) - start >= size - 1);
729 		*addrp = start;
730 		return 0;
731 	}
732 	return ENOMEM;
733 }
734 
735 /* ---- vmem API */
736 
737 /*
738  * vmem_create: create an arena.
739  *
740  * => must not be called from interrupt context.
741  */
742 
743 vmem_t *
744 vmem_create(const char *name, vmem_addr_t base, vmem_size_t size,
745     vmem_size_t quantum,
746     int (*importfn)(void *, vmem_size_t, vmem_size_t *, vm_flag_t,
747         vmem_addr_t *),
748     void (*releasefn)(void *, vmem_addr_t, vmem_size_t),
749     void *arg, vmem_size_t qcache_max, vm_flag_t flags, int ipl)
750 {
751 	vmem_t *vm;
752 	int i;
753 #if defined(_KERNEL)
754 	static ONCE_DECL(control);
755 #endif /* defined(_KERNEL) */
756 
757 	KASSERT((flags & (VM_SLEEP|VM_NOSLEEP)) != 0);
758 	KASSERT((~flags & (VM_SLEEP|VM_NOSLEEP)) != 0);
759 	KASSERT(quantum > 0);
760 
761 #if defined(_KERNEL)
762 	if (RUN_ONCE(&control, vmem_init)) {
763 		return NULL;
764 	}
765 #endif /* defined(_KERNEL) */
766 	vm = xmalloc(sizeof(*vm), flags);
767 	if (vm == NULL) {
768 		return NULL;
769 	}
770 
771 	VMEM_LOCK_INIT(vm, ipl);
772 	vm->vm_name = name;
773 	vm->vm_quantum_mask = quantum - 1;
774 	vm->vm_quantum_shift = SIZE2ORDER(quantum);
775 	KASSERT(ORDER2SIZE(vm->vm_quantum_shift) == quantum);
776 	vm->vm_importfn = importfn;
777 	vm->vm_releasefn = releasefn;
778 	vm->vm_arg = arg;
779 	vm->vm_nbusytag = 0;
780 #if defined(QCACHE)
781 	qc_init(vm, qcache_max, ipl);
782 #endif /* defined(QCACHE) */
783 
784 	CIRCLEQ_INIT(&vm->vm_seglist);
785 	for (i = 0; i < VMEM_MAXORDER; i++) {
786 		LIST_INIT(&vm->vm_freelist[i]);
787 	}
788 	vm->vm_hashlist = NULL;
789 	if (vmem_rehash(vm, VMEM_HASHSIZE_INIT, flags)) {
790 		vmem_destroy1(vm);
791 		return NULL;
792 	}
793 
794 	if (size != 0) {
795 		if (vmem_add(vm, base, size, flags) != 0) {
796 			vmem_destroy1(vm);
797 			return NULL;
798 		}
799 	}
800 
801 #if defined(_KERNEL)
802 	mutex_enter(&vmem_list_lock);
803 	LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist);
804 	mutex_exit(&vmem_list_lock);
805 #endif /* defined(_KERNEL) */
806 
807 	return vm;
808 }
809 
810 void
811 vmem_destroy(vmem_t *vm)
812 {
813 
814 #if defined(_KERNEL)
815 	mutex_enter(&vmem_list_lock);
816 	LIST_REMOVE(vm, vm_alllist);
817 	mutex_exit(&vmem_list_lock);
818 #endif /* defined(_KERNEL) */
819 
820 	vmem_destroy1(vm);
821 }
822 
823 vmem_size_t
824 vmem_roundup_size(vmem_t *vm, vmem_size_t size)
825 {
826 
827 	return (size + vm->vm_quantum_mask) & ~vm->vm_quantum_mask;
828 }
829 
830 /*
831  * vmem_alloc:
832  *
833  * => caller must ensure appropriate spl,
834  *    if the arena can be accessed from interrupt context.
835  */
836 
837 int
838 vmem_alloc(vmem_t *vm, vmem_size_t size, vm_flag_t flags, vmem_addr_t *addrp)
839 {
840 	const vm_flag_t strat __unused = flags & VM_FITMASK;
841 
842 	KASSERT((flags & (VM_SLEEP|VM_NOSLEEP)) != 0);
843 	KASSERT((~flags & (VM_SLEEP|VM_NOSLEEP)) != 0);
844 
845 	KASSERT(size > 0);
846 	KASSERT(strat == VM_BESTFIT || strat == VM_INSTANTFIT);
847 	if ((flags & VM_SLEEP) != 0) {
848 		ASSERT_SLEEPABLE();
849 	}
850 
851 #if defined(QCACHE)
852 	if (size <= vm->vm_qcache_max) {
853 		void *p;
854 		int qidx = (size + vm->vm_quantum_mask) >> vm->vm_quantum_shift;
855 		qcache_t *qc = vm->vm_qcache[qidx - 1];
856 
857 		p = pool_cache_get(qc->qc_cache, vmf_to_prf(flags));
858 		if (addrp != NULL)
859 			*addrp = (vmem_addr_t)p;
860 		return (p == NULL) ? ENOMEM : 0;
861 	}
862 #endif /* defined(QCACHE) */
863 
864 	return vmem_xalloc(vm, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
865 	    flags, addrp);
866 }
867 
868 int
869 vmem_xalloc(vmem_t *vm, const vmem_size_t size0, vmem_size_t align,
870     const vmem_size_t phase, const vmem_size_t nocross,
871     const vmem_addr_t minaddr, const vmem_addr_t maxaddr, const vm_flag_t flags,
872     vmem_addr_t *addrp)
873 {
874 	struct vmem_freelist *list;
875 	struct vmem_freelist *first;
876 	struct vmem_freelist *end;
877 	bt_t *bt;
878 	bt_t *btnew;
879 	bt_t *btnew2;
880 	const vmem_size_t size = vmem_roundup_size(vm, size0);
881 	vm_flag_t strat = flags & VM_FITMASK;
882 	vmem_addr_t start;
883 	int rc;
884 
885 	KASSERT(size0 > 0);
886 	KASSERT(size > 0);
887 	KASSERT(strat == VM_BESTFIT || strat == VM_INSTANTFIT);
888 	if ((flags & VM_SLEEP) != 0) {
889 		ASSERT_SLEEPABLE();
890 	}
891 	KASSERT((align & vm->vm_quantum_mask) == 0);
892 	KASSERT((align & (align - 1)) == 0);
893 	KASSERT((phase & vm->vm_quantum_mask) == 0);
894 	KASSERT((nocross & vm->vm_quantum_mask) == 0);
895 	KASSERT((nocross & (nocross - 1)) == 0);
896 	KASSERT((align == 0 && phase == 0) || phase < align);
897 	KASSERT(nocross == 0 || nocross >= size);
898 	KASSERT(minaddr <= maxaddr);
899 	KASSERT(!VMEM_CROSS_P(phase, phase + size - 1, nocross));
900 
901 	if (align == 0) {
902 		align = vm->vm_quantum_mask + 1;
903 	}
904 
905 	/*
906 	 * allocate boundary tags before acquiring the vmem lock.
907 	 */
908 	btnew = bt_alloc(vm, flags);
909 	if (btnew == NULL) {
910 		return ENOMEM;
911 	}
912 	btnew2 = bt_alloc(vm, flags); /* XXX not necessary if no restrictions */
913 	if (btnew2 == NULL) {
914 		bt_free(vm, btnew);
915 		return ENOMEM;
916 	}
917 
918 	/*
919 	 * choose a free block from which we allocate.
920 	 */
921 retry_strat:
922 	first = bt_freehead_toalloc(vm, size, strat);
923 	end = &vm->vm_freelist[VMEM_MAXORDER];
924 retry:
925 	bt = NULL;
926 	VMEM_LOCK(vm);
927 	vmem_check(vm);
928 	if (strat == VM_INSTANTFIT) {
929 		/*
930 		 * just choose the first block which satisfies our restrictions.
931 		 *
932 		 * note that we don't need to check the size of the blocks
933 		 * because any blocks found on these list should be larger than
934 		 * the given size.
935 		 */
936 		for (list = first; list < end; list++) {
937 			bt = LIST_FIRST(list);
938 			if (bt != NULL) {
939 				rc = vmem_fit(bt, size, align, phase,
940 				    nocross, minaddr, maxaddr, &start);
941 				if (rc == 0) {
942 					goto gotit;
943 				}
944 				/*
945 				 * don't bother to follow the bt_freelist link
946 				 * here.  the list can be very long and we are
947 				 * told to run fast.  blocks from the later free
948 				 * lists are larger and have better chances to
949 				 * satisfy our restrictions.
950 				 */
951 			}
952 		}
953 	} else { /* VM_BESTFIT */
954 		/*
955 		 * we assume that, for space efficiency, it's better to
956 		 * allocate from a smaller block.  thus we will start searching
957 		 * from the lower-order list than VM_INSTANTFIT.
958 		 * however, don't bother to find the smallest block in a free
959 		 * list because the list can be very long.  we can revisit it
960 		 * if/when it turns out to be a problem.
961 		 *
962 		 * note that the 'first' list can contain blocks smaller than
963 		 * the requested size.  thus we need to check bt_size.
964 		 */
965 		for (list = first; list < end; list++) {
966 			LIST_FOREACH(bt, list, bt_freelist) {
967 				if (bt->bt_size >= size) {
968 					rc = vmem_fit(bt, size, align, phase,
969 					    nocross, minaddr, maxaddr, &start);
970 					if (rc == 0) {
971 						goto gotit;
972 					}
973 				}
974 			}
975 		}
976 	}
977 	VMEM_UNLOCK(vm);
978 #if 1
979 	if (strat == VM_INSTANTFIT) {
980 		strat = VM_BESTFIT;
981 		goto retry_strat;
982 	}
983 #endif
984 	if (align != vm->vm_quantum_mask + 1 || phase != 0 ||
985 	    nocross != 0) {
986 
987 		/*
988 		 * XXX should try to import a region large enough to
989 		 * satisfy restrictions?
990 		 */
991 
992 		goto fail;
993 	}
994 	/* XXX eeek, minaddr & maxaddr not respected */
995 	if (vmem_import(vm, size, flags) == 0) {
996 		goto retry;
997 	}
998 	/* XXX */
999 fail:
1000 	bt_free(vm, btnew);
1001 	bt_free(vm, btnew2);
1002 	return ENOMEM;
1003 
1004 gotit:
1005 	KASSERT(bt->bt_type == BT_TYPE_FREE);
1006 	KASSERT(bt->bt_size >= size);
1007 	bt_remfree(vm, bt);
1008 	vmem_check(vm);
1009 	if (bt->bt_start != start) {
1010 		btnew2->bt_type = BT_TYPE_FREE;
1011 		btnew2->bt_start = bt->bt_start;
1012 		btnew2->bt_size = start - bt->bt_start;
1013 		bt->bt_start = start;
1014 		bt->bt_size -= btnew2->bt_size;
1015 		bt_insfree(vm, btnew2);
1016 		bt_insseg(vm, btnew2, CIRCLEQ_PREV(bt, bt_seglist));
1017 		btnew2 = NULL;
1018 		vmem_check(vm);
1019 	}
1020 	KASSERT(bt->bt_start == start);
1021 	if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) {
1022 		/* split */
1023 		btnew->bt_type = BT_TYPE_BUSY;
1024 		btnew->bt_start = bt->bt_start;
1025 		btnew->bt_size = size;
1026 		bt->bt_start = bt->bt_start + size;
1027 		bt->bt_size -= size;
1028 		bt_insfree(vm, bt);
1029 		bt_insseg(vm, btnew, CIRCLEQ_PREV(bt, bt_seglist));
1030 		bt_insbusy(vm, btnew);
1031 		vmem_check(vm);
1032 		VMEM_UNLOCK(vm);
1033 	} else {
1034 		bt->bt_type = BT_TYPE_BUSY;
1035 		bt_insbusy(vm, bt);
1036 		vmem_check(vm);
1037 		VMEM_UNLOCK(vm);
1038 		bt_free(vm, btnew);
1039 		btnew = bt;
1040 	}
1041 	if (btnew2 != NULL) {
1042 		bt_free(vm, btnew2);
1043 	}
1044 	KASSERT(btnew->bt_size >= size);
1045 	btnew->bt_type = BT_TYPE_BUSY;
1046 
1047 	if (addrp != NULL)
1048 		*addrp = btnew->bt_start;
1049 	return 0;
1050 }
1051 
1052 /*
1053  * vmem_free:
1054  *
1055  * => caller must ensure appropriate spl,
1056  *    if the arena can be accessed from interrupt context.
1057  */
1058 
1059 void
1060 vmem_free(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1061 {
1062 
1063 	KASSERT(size > 0);
1064 
1065 #if defined(QCACHE)
1066 	if (size <= vm->vm_qcache_max) {
1067 		int qidx = (size + vm->vm_quantum_mask) >> vm->vm_quantum_shift;
1068 		qcache_t *qc = vm->vm_qcache[qidx - 1];
1069 
1070 		return pool_cache_put(qc->qc_cache, (void *)addr);
1071 	}
1072 #endif /* defined(QCACHE) */
1073 
1074 	vmem_xfree(vm, addr, size);
1075 }
1076 
1077 void
1078 vmem_xfree(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1079 {
1080 	bt_t *bt;
1081 	bt_t *t;
1082 
1083 	KASSERT(size > 0);
1084 
1085 	VMEM_LOCK(vm);
1086 
1087 	bt = bt_lookupbusy(vm, addr);
1088 	KASSERT(bt != NULL);
1089 	KASSERT(bt->bt_start == addr);
1090 	KASSERT(bt->bt_size == vmem_roundup_size(vm, size) ||
1091 	    bt->bt_size - vmem_roundup_size(vm, size) <= vm->vm_quantum_mask);
1092 	KASSERT(bt->bt_type == BT_TYPE_BUSY);
1093 	bt_rembusy(vm, bt);
1094 	bt->bt_type = BT_TYPE_FREE;
1095 
1096 	/* coalesce */
1097 	t = CIRCLEQ_NEXT(bt, bt_seglist);
1098 	if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1099 		KASSERT(BT_END(bt) < t->bt_start);	/* YYY */
1100 		bt_remfree(vm, t);
1101 		bt_remseg(vm, t);
1102 		bt->bt_size += t->bt_size;
1103 		bt_free(vm, t);
1104 	}
1105 	t = CIRCLEQ_PREV(bt, bt_seglist);
1106 	if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1107 		KASSERT(BT_END(t) < bt->bt_start);	/* YYY */
1108 		bt_remfree(vm, t);
1109 		bt_remseg(vm, t);
1110 		bt->bt_size += t->bt_size;
1111 		bt->bt_start = t->bt_start;
1112 		bt_free(vm, t);
1113 	}
1114 
1115 	t = CIRCLEQ_PREV(bt, bt_seglist);
1116 	KASSERT(t != NULL);
1117 	KASSERT(BT_ISSPAN_P(t) || t->bt_type == BT_TYPE_BUSY);
1118 	if (vm->vm_releasefn != NULL && t->bt_type == BT_TYPE_SPAN &&
1119 	    t->bt_size == bt->bt_size) {
1120 		vmem_addr_t spanaddr;
1121 		vmem_size_t spansize;
1122 
1123 		KASSERT(t->bt_start == bt->bt_start);
1124 		spanaddr = bt->bt_start;
1125 		spansize = bt->bt_size;
1126 		bt_remseg(vm, bt);
1127 		bt_free(vm, bt);
1128 		bt_remseg(vm, t);
1129 		bt_free(vm, t);
1130 		VMEM_UNLOCK(vm);
1131 		(*vm->vm_releasefn)(vm->vm_arg, spanaddr, spansize);
1132 	} else {
1133 		bt_insfree(vm, bt);
1134 		VMEM_UNLOCK(vm);
1135 	}
1136 }
1137 
1138 /*
1139  * vmem_add:
1140  *
1141  * => caller must ensure appropriate spl,
1142  *    if the arena can be accessed from interrupt context.
1143  */
1144 
1145 int
1146 vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, vm_flag_t flags)
1147 {
1148 
1149 	return vmem_add1(vm, addr, size, flags, BT_TYPE_SPAN_STATIC);
1150 }
1151 
1152 /*
1153  * vmem_reap: reap unused resources.
1154  *
1155  * => return true if we successfully reaped something.
1156  */
1157 
1158 bool
1159 vmem_reap(vmem_t *vm)
1160 {
1161 	bool didsomething = false;
1162 
1163 #if defined(QCACHE)
1164 	didsomething = qc_reap(vm);
1165 #endif /* defined(QCACHE) */
1166 	return didsomething;
1167 }
1168 
1169 /* ---- rehash */
1170 
1171 #if defined(_KERNEL)
1172 static struct callout vmem_rehash_ch;
1173 static int vmem_rehash_interval;
1174 static struct workqueue *vmem_rehash_wq;
1175 static struct work vmem_rehash_wk;
1176 
1177 static void
1178 vmem_rehash_all(struct work *wk, void *dummy)
1179 {
1180 	vmem_t *vm;
1181 
1182 	KASSERT(wk == &vmem_rehash_wk);
1183 	mutex_enter(&vmem_list_lock);
1184 	LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1185 		size_t desired;
1186 		size_t current;
1187 
1188 		if (!VMEM_TRYLOCK(vm)) {
1189 			continue;
1190 		}
1191 		desired = vm->vm_nbusytag;
1192 		current = vm->vm_hashsize;
1193 		VMEM_UNLOCK(vm);
1194 
1195 		if (desired > VMEM_HASHSIZE_MAX) {
1196 			desired = VMEM_HASHSIZE_MAX;
1197 		} else if (desired < VMEM_HASHSIZE_MIN) {
1198 			desired = VMEM_HASHSIZE_MIN;
1199 		}
1200 		if (desired > current * 2 || desired * 2 < current) {
1201 			vmem_rehash(vm, desired, VM_NOSLEEP);
1202 		}
1203 	}
1204 	mutex_exit(&vmem_list_lock);
1205 
1206 	callout_schedule(&vmem_rehash_ch, vmem_rehash_interval);
1207 }
1208 
1209 static void
1210 vmem_rehash_all_kick(void *dummy)
1211 {
1212 
1213 	workqueue_enqueue(vmem_rehash_wq, &vmem_rehash_wk, NULL);
1214 }
1215 
1216 void
1217 vmem_rehash_start(void)
1218 {
1219 	int error;
1220 
1221 	error = workqueue_create(&vmem_rehash_wq, "vmem_rehash",
1222 	    vmem_rehash_all, NULL, PRI_VM, IPL_SOFTCLOCK, WQ_MPSAFE);
1223 	if (error) {
1224 		panic("%s: workqueue_create %d\n", __func__, error);
1225 	}
1226 	callout_init(&vmem_rehash_ch, CALLOUT_MPSAFE);
1227 	callout_setfunc(&vmem_rehash_ch, vmem_rehash_all_kick, NULL);
1228 
1229 	vmem_rehash_interval = hz * 10;
1230 	callout_schedule(&vmem_rehash_ch, vmem_rehash_interval);
1231 }
1232 #endif /* defined(_KERNEL) */
1233 
1234 /* ---- debug */
1235 
1236 #if defined(DDB) || defined(UNITTEST) || defined(VMEM_SANITY)
1237 
1238 static void bt_dump(const bt_t *, void (*)(const char *, ...));
1239 
1240 static const char *
1241 bt_type_string(int type)
1242 {
1243 	static const char * const table[] = {
1244 		[BT_TYPE_BUSY] = "busy",
1245 		[BT_TYPE_FREE] = "free",
1246 		[BT_TYPE_SPAN] = "span",
1247 		[BT_TYPE_SPAN_STATIC] = "static span",
1248 	};
1249 
1250 	if (type >= __arraycount(table)) {
1251 		return "BOGUS";
1252 	}
1253 	return table[type];
1254 }
1255 
1256 static void
1257 bt_dump(const bt_t *bt, void (*pr)(const char *, ...))
1258 {
1259 
1260 	(*pr)("\t%p: %" PRIu64 ", %" PRIu64 ", %d(%s)\n",
1261 	    bt, (uint64_t)bt->bt_start, (uint64_t)bt->bt_size,
1262 	    bt->bt_type, bt_type_string(bt->bt_type));
1263 }
1264 
1265 static void
1266 vmem_dump(const vmem_t *vm , void (*pr)(const char *, ...))
1267 {
1268 	const bt_t *bt;
1269 	int i;
1270 
1271 	(*pr)("vmem %p '%s'\n", vm, vm->vm_name);
1272 	CIRCLEQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1273 		bt_dump(bt, pr);
1274 	}
1275 
1276 	for (i = 0; i < VMEM_MAXORDER; i++) {
1277 		const struct vmem_freelist *fl = &vm->vm_freelist[i];
1278 
1279 		if (LIST_EMPTY(fl)) {
1280 			continue;
1281 		}
1282 
1283 		(*pr)("freelist[%d]\n", i);
1284 		LIST_FOREACH(bt, fl, bt_freelist) {
1285 			bt_dump(bt, pr);
1286 		}
1287 	}
1288 }
1289 
1290 #endif /* defined(DDB) || defined(UNITTEST) || defined(VMEM_SANITY) */
1291 
1292 #if defined(DDB)
1293 static bt_t *
1294 vmem_whatis_lookup(vmem_t *vm, uintptr_t addr)
1295 {
1296 	bt_t *bt;
1297 
1298 	CIRCLEQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1299 		if (BT_ISSPAN_P(bt)) {
1300 			continue;
1301 		}
1302 		if (bt->bt_start <= addr && addr <= BT_END(bt)) {
1303 			return bt;
1304 		}
1305 	}
1306 
1307 	return NULL;
1308 }
1309 
1310 void
1311 vmem_whatis(uintptr_t addr, void (*pr)(const char *, ...))
1312 {
1313 	vmem_t *vm;
1314 
1315 	LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1316 		bt_t *bt;
1317 
1318 		bt = vmem_whatis_lookup(vm, addr);
1319 		if (bt == NULL) {
1320 			continue;
1321 		}
1322 		(*pr)("%p is %p+%zu in VMEM '%s' (%s)\n",
1323 		    (void *)addr, (void *)bt->bt_start,
1324 		    (size_t)(addr - bt->bt_start), vm->vm_name,
1325 		    (bt->bt_type == BT_TYPE_BUSY) ? "allocated" : "free");
1326 	}
1327 }
1328 
1329 void
1330 vmem_printall(const char *modif, void (*pr)(const char *, ...))
1331 {
1332 	const vmem_t *vm;
1333 
1334 	LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1335 		vmem_dump(vm, pr);
1336 	}
1337 }
1338 
1339 void
1340 vmem_print(uintptr_t addr, const char *modif, void (*pr)(const char *, ...))
1341 {
1342 	const vmem_t *vm = (const void *)addr;
1343 
1344 	vmem_dump(vm, pr);
1345 }
1346 #endif /* defined(DDB) */
1347 
1348 #if defined(_KERNEL)
1349 #define vmem_printf printf
1350 #else
1351 #include <stdio.h>
1352 #include <stdarg.h>
1353 
1354 static void
1355 vmem_printf(const char *fmt, ...)
1356 {
1357 	va_list ap;
1358 	va_start(ap, fmt);
1359 	vprintf(fmt, ap);
1360 	va_end(ap);
1361 }
1362 #endif
1363 
1364 #if defined(VMEM_SANITY)
1365 
1366 static bool
1367 vmem_check_sanity(vmem_t *vm)
1368 {
1369 	const bt_t *bt, *bt2;
1370 
1371 	KASSERT(vm != NULL);
1372 
1373 	CIRCLEQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1374 		if (bt->bt_start > BT_END(bt)) {
1375 			printf("corrupted tag\n");
1376 			bt_dump(bt, vmem_printf);
1377 			return false;
1378 		}
1379 	}
1380 	CIRCLEQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1381 		CIRCLEQ_FOREACH(bt2, &vm->vm_seglist, bt_seglist) {
1382 			if (bt == bt2) {
1383 				continue;
1384 			}
1385 			if (BT_ISSPAN_P(bt) != BT_ISSPAN_P(bt2)) {
1386 				continue;
1387 			}
1388 			if (bt->bt_start <= BT_END(bt2) &&
1389 			    bt2->bt_start <= BT_END(bt)) {
1390 				printf("overwrapped tags\n");
1391 				bt_dump(bt, vmem_printf);
1392 				bt_dump(bt2, vmem_printf);
1393 				return false;
1394 			}
1395 		}
1396 	}
1397 
1398 	return true;
1399 }
1400 
1401 static void
1402 vmem_check(vmem_t *vm)
1403 {
1404 
1405 	if (!vmem_check_sanity(vm)) {
1406 		panic("insanity vmem %p", vm);
1407 	}
1408 }
1409 
1410 #endif /* defined(VMEM_SANITY) */
1411 
1412 #if defined(UNITTEST)
1413 int
1414 main(void)
1415 {
1416 	int rc;
1417 	vmem_t *vm;
1418 	vmem_addr_t p;
1419 	struct reg {
1420 		vmem_addr_t p;
1421 		vmem_size_t sz;
1422 		bool x;
1423 	} *reg = NULL;
1424 	int nreg = 0;
1425 	int nalloc = 0;
1426 	int nfree = 0;
1427 	vmem_size_t total = 0;
1428 #if 1
1429 	vm_flag_t strat = VM_INSTANTFIT;
1430 #else
1431 	vm_flag_t strat = VM_BESTFIT;
1432 #endif
1433 
1434 	vm = vmem_create("test", 0, 0, 1, NULL, NULL, NULL, 0, VM_SLEEP,
1435 #ifdef _KERNEL
1436 	    IPL_NONE
1437 #else
1438 	    0
1439 #endif
1440 	    );
1441 	if (vm == NULL) {
1442 		printf("vmem_create\n");
1443 		exit(EXIT_FAILURE);
1444 	}
1445 	vmem_dump(vm, vmem_printf);
1446 
1447 	rc = vmem_add(vm, 0, 50, VM_SLEEP);
1448 	assert(rc == 0);
1449 	rc = vmem_add(vm, 100, 200, VM_SLEEP);
1450 	assert(rc == 0);
1451 	rc = vmem_add(vm, 2000, 1, VM_SLEEP);
1452 	assert(rc == 0);
1453 	rc = vmem_add(vm, 40000, 65536, VM_SLEEP);
1454 	assert(rc == 0);
1455 	rc = vmem_add(vm, 10000, 10000, VM_SLEEP);
1456 	assert(rc == 0);
1457 	rc = vmem_add(vm, 500, 1000, VM_SLEEP);
1458 	assert(rc == 0);
1459 	rc = vmem_add(vm, 0xffffff00, 0x100, VM_SLEEP);
1460 	assert(rc == 0);
1461 	rc = vmem_xalloc(vm, 0x101, 0, 0, 0,
1462 	    0xffffff00, 0xffffffff, strat|VM_SLEEP, &p);
1463 	assert(rc != 0);
1464 	rc = vmem_xalloc(vm, 50, 0, 0, 0, 0, 49, strat|VM_SLEEP, &p);
1465 	assert(rc == 0 && p == 0);
1466 	vmem_xfree(vm, p, 50);
1467 	rc = vmem_xalloc(vm, 25, 0, 0, 0, 0, 24, strat|VM_SLEEP, &p);
1468 	assert(rc == 0 && p == 0);
1469 	rc = vmem_xalloc(vm, 0x100, 0, 0, 0,
1470 	    0xffffff01, 0xffffffff, strat|VM_SLEEP, &p);
1471 	assert(rc != 0);
1472 	rc = vmem_xalloc(vm, 0x100, 0, 0, 0,
1473 	    0xffffff00, 0xfffffffe, strat|VM_SLEEP, &p);
1474 	assert(rc != 0);
1475 	rc = vmem_xalloc(vm, 0x100, 0, 0, 0,
1476 	    0xffffff00, 0xffffffff, strat|VM_SLEEP, &p);
1477 	assert(rc == 0);
1478 	vmem_dump(vm, vmem_printf);
1479 	for (;;) {
1480 		struct reg *r;
1481 		int t = rand() % 100;
1482 
1483 		if (t > 45) {
1484 			/* alloc */
1485 			vmem_size_t sz = rand() % 500 + 1;
1486 			bool x;
1487 			vmem_size_t align, phase, nocross;
1488 			vmem_addr_t minaddr, maxaddr;
1489 
1490 			if (t > 70) {
1491 				x = true;
1492 				/* XXX */
1493 				align = 1 << (rand() % 15);
1494 				phase = rand() % 65536;
1495 				nocross = 1 << (rand() % 15);
1496 				if (align <= phase) {
1497 					phase = 0;
1498 				}
1499 				if (VMEM_CROSS_P(phase, phase + sz - 1,
1500 				    nocross)) {
1501 					nocross = 0;
1502 				}
1503 				do {
1504 					minaddr = rand() % 50000;
1505 					maxaddr = rand() % 70000;
1506 				} while (minaddr > maxaddr);
1507 				printf("=== xalloc %" PRIu64
1508 				    " align=%" PRIu64 ", phase=%" PRIu64
1509 				    ", nocross=%" PRIu64 ", min=%" PRIu64
1510 				    ", max=%" PRIu64 "\n",
1511 				    (uint64_t)sz,
1512 				    (uint64_t)align,
1513 				    (uint64_t)phase,
1514 				    (uint64_t)nocross,
1515 				    (uint64_t)minaddr,
1516 				    (uint64_t)maxaddr);
1517 				rc = vmem_xalloc(vm, sz, align, phase, nocross,
1518 				    minaddr, maxaddr, strat|VM_SLEEP, &p);
1519 			} else {
1520 				x = false;
1521 				printf("=== alloc %" PRIu64 "\n", (uint64_t)sz);
1522 				rc = vmem_alloc(vm, sz, strat|VM_SLEEP, &p);
1523 			}
1524 			printf("-> %" PRIu64 "\n", (uint64_t)p);
1525 			vmem_dump(vm, vmem_printf);
1526 			if (rc != 0) {
1527 				if (x) {
1528 					continue;
1529 				}
1530 				break;
1531 			}
1532 			nreg++;
1533 			reg = realloc(reg, sizeof(*reg) * nreg);
1534 			r = &reg[nreg - 1];
1535 			r->p = p;
1536 			r->sz = sz;
1537 			r->x = x;
1538 			total += sz;
1539 			nalloc++;
1540 		} else if (nreg != 0) {
1541 			/* free */
1542 			r = &reg[rand() % nreg];
1543 			printf("=== free %" PRIu64 ", %" PRIu64 "\n",
1544 			    (uint64_t)r->p, (uint64_t)r->sz);
1545 			if (r->x) {
1546 				vmem_xfree(vm, r->p, r->sz);
1547 			} else {
1548 				vmem_free(vm, r->p, r->sz);
1549 			}
1550 			total -= r->sz;
1551 			vmem_dump(vm, vmem_printf);
1552 			*r = reg[nreg - 1];
1553 			nreg--;
1554 			nfree++;
1555 		}
1556 		printf("total=%" PRIu64 "\n", (uint64_t)total);
1557 	}
1558 	fprintf(stderr, "total=%" PRIu64 ", nalloc=%d, nfree=%d\n",
1559 	    (uint64_t)total, nalloc, nfree);
1560 	exit(EXIT_SUCCESS);
1561 }
1562 #endif /* defined(UNITTEST) */
1563