xref: /freebsd/sys/kern/subr_vmem.c (revision f126890ac5386406dadf7c4cfa9566cbb56537c5)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c)2006,2007,2008,2009 YAMAMOTO Takashi,
5  * Copyright (c) 2013 EMC Corp.
6  * All rights reserved.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 /*
31  * From:
32  *	$NetBSD: vmem_impl.h,v 1.2 2013/01/29 21:26:24 para Exp $
33  *	$NetBSD: subr_vmem.c,v 1.83 2013/03/06 11:20:10 yamt Exp $
34  */
35 
36 /*
37  * reference:
38  * -	Magazines and Vmem: Extending the Slab Allocator
39  *	to Many CPUs and Arbitrary Resources
40  *	http://www.usenix.org/event/usenix01/bonwick.html
41  */
42 
43 #include <sys/cdefs.h>
44 #include "opt_ddb.h"
45 
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kernel.h>
49 #include <sys/queue.h>
50 #include <sys/callout.h>
51 #include <sys/hash.h>
52 #include <sys/lock.h>
53 #include <sys/malloc.h>
54 #include <sys/mutex.h>
55 #include <sys/smp.h>
56 #include <sys/condvar.h>
57 #include <sys/sysctl.h>
58 #include <sys/taskqueue.h>
59 #include <sys/vmem.h>
60 #include <sys/vmmeter.h>
61 
62 #include "opt_vm.h"
63 
64 #include <vm/uma.h>
65 #include <vm/vm.h>
66 #include <vm/pmap.h>
67 #include <vm/vm_map.h>
68 #include <vm/vm_object.h>
69 #include <vm/vm_kern.h>
70 #include <vm/vm_extern.h>
71 #include <vm/vm_param.h>
72 #include <vm/vm_page.h>
73 #include <vm/vm_pageout.h>
74 #include <vm/vm_phys.h>
75 #include <vm/vm_pagequeue.h>
76 #include <vm/uma_int.h>
77 
78 #define	VMEM_OPTORDER		5
79 #define	VMEM_OPTVALUE		(1 << VMEM_OPTORDER)
80 #define	VMEM_MAXORDER						\
81     (VMEM_OPTVALUE - 1 + sizeof(vmem_size_t) * NBBY - VMEM_OPTORDER)
82 
83 #define	VMEM_HASHSIZE_MIN	16
84 #define	VMEM_HASHSIZE_MAX	131072
85 
86 #define	VMEM_QCACHE_IDX_MAX	16
87 
88 #define	VMEM_FITMASK	(M_BESTFIT | M_FIRSTFIT | M_NEXTFIT)
89 
90 #define	VMEM_FLAGS	(M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM |	\
91     M_BESTFIT | M_FIRSTFIT | M_NEXTFIT)
92 
93 #define	BT_FLAGS	(M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM)
94 
95 #define	QC_NAME_MAX	16
96 
97 /*
98  * Data structures private to vmem.
99  */
100 MALLOC_DEFINE(M_VMEM, "vmem", "vmem internal structures");
101 
102 typedef struct vmem_btag bt_t;
103 
104 TAILQ_HEAD(vmem_seglist, vmem_btag);
105 LIST_HEAD(vmem_freelist, vmem_btag);
106 LIST_HEAD(vmem_hashlist, vmem_btag);
107 
108 struct qcache {
109 	uma_zone_t	qc_cache;
110 	vmem_t 		*qc_vmem;
111 	vmem_size_t	qc_size;
112 	char		qc_name[QC_NAME_MAX];
113 };
114 typedef struct qcache qcache_t;
115 #define	QC_POOL_TO_QCACHE(pool)	((qcache_t *)(pool->pr_qcache))
116 
117 #define	VMEM_NAME_MAX	16
118 
119 /* boundary tag */
120 struct vmem_btag {
121 	TAILQ_ENTRY(vmem_btag) bt_seglist;
122 	union {
123 		LIST_ENTRY(vmem_btag) u_freelist; /* BT_TYPE_FREE */
124 		LIST_ENTRY(vmem_btag) u_hashlist; /* BT_TYPE_BUSY */
125 	} bt_u;
126 #define	bt_hashlist	bt_u.u_hashlist
127 #define	bt_freelist	bt_u.u_freelist
128 	vmem_addr_t	bt_start;
129 	vmem_size_t	bt_size;
130 	int		bt_type;
131 };
132 
133 /* vmem arena */
134 struct vmem {
135 	struct mtx_padalign	vm_lock;
136 	struct cv		vm_cv;
137 	char			vm_name[VMEM_NAME_MAX+1];
138 	LIST_ENTRY(vmem)	vm_alllist;
139 	struct vmem_hashlist	vm_hash0[VMEM_HASHSIZE_MIN];
140 	struct vmem_freelist	vm_freelist[VMEM_MAXORDER];
141 	struct vmem_seglist	vm_seglist;
142 	struct vmem_hashlist	*vm_hashlist;
143 	vmem_size_t		vm_hashsize;
144 
145 	/* Constant after init */
146 	vmem_size_t		vm_qcache_max;
147 	vmem_size_t		vm_quantum_mask;
148 	vmem_size_t		vm_import_quantum;
149 	int			vm_quantum_shift;
150 
151 	/* Written on alloc/free */
152 	LIST_HEAD(, vmem_btag)	vm_freetags;
153 	int			vm_nfreetags;
154 	int			vm_nbusytag;
155 	vmem_size_t		vm_inuse;
156 	vmem_size_t		vm_size;
157 	vmem_size_t		vm_limit;
158 	struct vmem_btag	vm_cursor;
159 
160 	/* Used on import. */
161 	vmem_import_t		*vm_importfn;
162 	vmem_release_t		*vm_releasefn;
163 	void			*vm_arg;
164 
165 	/* Space exhaustion callback. */
166 	vmem_reclaim_t		*vm_reclaimfn;
167 
168 	/* quantum cache */
169 	qcache_t		vm_qcache[VMEM_QCACHE_IDX_MAX];
170 };
171 
172 #define	BT_TYPE_SPAN		1	/* Allocated from importfn */
173 #define	BT_TYPE_SPAN_STATIC	2	/* vmem_add() or create. */
174 #define	BT_TYPE_FREE		3	/* Available space. */
175 #define	BT_TYPE_BUSY		4	/* Used space. */
176 #define	BT_TYPE_CURSOR		5	/* Cursor for nextfit allocations. */
177 #define	BT_ISSPAN_P(bt)	((bt)->bt_type <= BT_TYPE_SPAN_STATIC)
178 
179 #define	BT_END(bt)	((bt)->bt_start + (bt)->bt_size - 1)
180 
181 #if defined(DIAGNOSTIC)
182 static int enable_vmem_check = 0;
183 SYSCTL_INT(_debug, OID_AUTO, vmem_check, CTLFLAG_RWTUN,
184     &enable_vmem_check, 0, "Enable vmem check");
185 static void vmem_check(vmem_t *);
186 #endif
187 
188 static struct callout	vmem_periodic_ch;
189 static int		vmem_periodic_interval;
190 static struct task	vmem_periodic_wk;
191 
192 static struct mtx_padalign __exclusive_cache_line vmem_list_lock;
193 static LIST_HEAD(, vmem) vmem_list = LIST_HEAD_INITIALIZER(vmem_list);
194 static uma_zone_t vmem_zone;
195 
196 /* ---- misc */
197 #define	VMEM_CONDVAR_INIT(vm, wchan)	cv_init(&vm->vm_cv, wchan)
198 #define	VMEM_CONDVAR_DESTROY(vm)	cv_destroy(&vm->vm_cv)
199 #define	VMEM_CONDVAR_WAIT(vm)		cv_wait(&vm->vm_cv, &vm->vm_lock)
200 #define	VMEM_CONDVAR_BROADCAST(vm)	cv_broadcast(&vm->vm_cv)
201 
202 #define	VMEM_LOCK(vm)		mtx_lock(&vm->vm_lock)
203 #define	VMEM_TRYLOCK(vm)	mtx_trylock(&vm->vm_lock)
204 #define	VMEM_UNLOCK(vm)		mtx_unlock(&vm->vm_lock)
205 #define	VMEM_LOCK_INIT(vm, name) mtx_init(&vm->vm_lock, (name), NULL, MTX_DEF)
206 #define	VMEM_LOCK_DESTROY(vm)	mtx_destroy(&vm->vm_lock)
207 #define	VMEM_ASSERT_LOCKED(vm)	mtx_assert(&vm->vm_lock, MA_OWNED);
208 
209 #define	VMEM_ALIGNUP(addr, align)	(-(-(addr) & -(align)))
210 
211 #define	VMEM_CROSS_P(addr1, addr2, boundary) \
212 	((((addr1) ^ (addr2)) & -(boundary)) != 0)
213 
214 #define	ORDER2SIZE(order)	((order) < VMEM_OPTVALUE ? ((order) + 1) : \
215     (vmem_size_t)1 << ((order) - (VMEM_OPTVALUE - VMEM_OPTORDER - 1)))
216 #define	SIZE2ORDER(size)	((size) <= VMEM_OPTVALUE ? ((size) - 1) : \
217     (flsl(size) + (VMEM_OPTVALUE - VMEM_OPTORDER - 2)))
218 
219 /*
220  * Maximum number of boundary tags that may be required to satisfy an
221  * allocation.  Two may be required to import.  Another two may be
222  * required to clip edges.
223  */
224 #define	BT_MAXALLOC	4
225 
226 /*
227  * Max free limits the number of locally cached boundary tags.  We
228  * just want to avoid hitting the zone allocator for every call.
229  */
230 #define BT_MAXFREE	(BT_MAXALLOC * 8)
231 
232 /* Allocator for boundary tags. */
233 static uma_zone_t vmem_bt_zone;
234 
235 /* boot time arena storage. */
236 static struct vmem kernel_arena_storage;
237 static struct vmem buffer_arena_storage;
238 static struct vmem transient_arena_storage;
239 /* kernel and kmem arenas are aliased for backwards KPI compat. */
240 vmem_t *kernel_arena = &kernel_arena_storage;
241 vmem_t *kmem_arena = &kernel_arena_storage;
242 vmem_t *buffer_arena = &buffer_arena_storage;
243 vmem_t *transient_arena = &transient_arena_storage;
244 
245 #ifdef DEBUG_MEMGUARD
246 static struct vmem memguard_arena_storage;
247 vmem_t *memguard_arena = &memguard_arena_storage;
248 #endif
249 
250 static bool
251 bt_isbusy(bt_t *bt)
252 {
253 	return (bt->bt_type == BT_TYPE_BUSY);
254 }
255 
256 static bool
257 bt_isfree(bt_t *bt)
258 {
259 	return (bt->bt_type == BT_TYPE_FREE);
260 }
261 
262 /*
263  * Fill the vmem's boundary tag cache.  We guarantee that boundary tag
264  * allocation will not fail once bt_fill() passes.  To do so we cache
265  * at least the maximum possible tag allocations in the arena.
266  */
267 static __noinline int
268 _bt_fill(vmem_t *vm, int flags)
269 {
270 	bt_t *bt;
271 
272 	VMEM_ASSERT_LOCKED(vm);
273 
274 	/*
275 	 * Only allow the kernel arena and arenas derived from kernel arena to
276 	 * dip into reserve tags.  They are where new tags come from.
277 	 */
278 	flags &= BT_FLAGS;
279 	if (vm != kernel_arena && vm->vm_arg != kernel_arena)
280 		flags &= ~M_USE_RESERVE;
281 
282 	/*
283 	 * Loop until we meet the reserve.  To minimize the lock shuffle
284 	 * and prevent simultaneous fills we first try a NOWAIT regardless
285 	 * of the caller's flags.  Specify M_NOVM so we don't recurse while
286 	 * holding a vmem lock.
287 	 */
288 	while (vm->vm_nfreetags < BT_MAXALLOC) {
289 		bt = uma_zalloc(vmem_bt_zone,
290 		    (flags & M_USE_RESERVE) | M_NOWAIT | M_NOVM);
291 		if (bt == NULL) {
292 			VMEM_UNLOCK(vm);
293 			bt = uma_zalloc(vmem_bt_zone, flags);
294 			VMEM_LOCK(vm);
295 			if (bt == NULL)
296 				break;
297 		}
298 		LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
299 		vm->vm_nfreetags++;
300 	}
301 
302 	if (vm->vm_nfreetags < BT_MAXALLOC)
303 		return ENOMEM;
304 
305 	return 0;
306 }
307 
308 static inline int
309 bt_fill(vmem_t *vm, int flags)
310 {
311 	if (vm->vm_nfreetags >= BT_MAXALLOC)
312 		return (0);
313 	return (_bt_fill(vm, flags));
314 }
315 
316 /*
317  * Pop a tag off of the freetag stack.
318  */
319 static bt_t *
320 bt_alloc(vmem_t *vm)
321 {
322 	bt_t *bt;
323 
324 	VMEM_ASSERT_LOCKED(vm);
325 	bt = LIST_FIRST(&vm->vm_freetags);
326 	MPASS(bt != NULL);
327 	LIST_REMOVE(bt, bt_freelist);
328 	vm->vm_nfreetags--;
329 
330 	return bt;
331 }
332 
333 /*
334  * Trim the per-vmem free list.  Returns with the lock released to
335  * avoid allocator recursions.
336  */
337 static void
338 bt_freetrim(vmem_t *vm, int freelimit)
339 {
340 	LIST_HEAD(, vmem_btag) freetags;
341 	bt_t *bt;
342 
343 	LIST_INIT(&freetags);
344 	VMEM_ASSERT_LOCKED(vm);
345 	while (vm->vm_nfreetags > freelimit) {
346 		bt = LIST_FIRST(&vm->vm_freetags);
347 		LIST_REMOVE(bt, bt_freelist);
348 		vm->vm_nfreetags--;
349 		LIST_INSERT_HEAD(&freetags, bt, bt_freelist);
350 	}
351 	VMEM_UNLOCK(vm);
352 	while ((bt = LIST_FIRST(&freetags)) != NULL) {
353 		LIST_REMOVE(bt, bt_freelist);
354 		uma_zfree(vmem_bt_zone, bt);
355 	}
356 }
357 
358 static inline void
359 bt_free(vmem_t *vm, bt_t *bt)
360 {
361 
362 	VMEM_ASSERT_LOCKED(vm);
363 	MPASS(LIST_FIRST(&vm->vm_freetags) != bt);
364 	LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
365 	vm->vm_nfreetags++;
366 }
367 
368 /*
369  * Hide MAXALLOC tags before dropping the arena lock to ensure that a
370  * concurrent allocation attempt does not grab them.
371  */
372 static void
373 bt_save(vmem_t *vm)
374 {
375 	KASSERT(vm->vm_nfreetags >= BT_MAXALLOC,
376 	    ("%s: insufficient free tags %d", __func__, vm->vm_nfreetags));
377 	vm->vm_nfreetags -= BT_MAXALLOC;
378 }
379 
380 static void
381 bt_restore(vmem_t *vm)
382 {
383 	vm->vm_nfreetags += BT_MAXALLOC;
384 }
385 
386 /*
387  * freelist[0] ... [1, 1]
388  * freelist[1] ... [2, 2]
389  *  :
390  * freelist[29] ... [30, 30]
391  * freelist[30] ... [31, 31]
392  * freelist[31] ... [32, 63]
393  * freelist[33] ... [64, 127]
394  *  :
395  * freelist[n] ... [(1 << (n - 26)), (1 << (n - 25)) - 1]
396  *  :
397  */
398 
399 static struct vmem_freelist *
400 bt_freehead_tofree(vmem_t *vm, vmem_size_t size)
401 {
402 	const vmem_size_t qsize = size >> vm->vm_quantum_shift;
403 	const int idx = SIZE2ORDER(qsize);
404 
405 	MPASS(size != 0 && qsize != 0);
406 	MPASS((size & vm->vm_quantum_mask) == 0);
407 	MPASS(idx >= 0);
408 	MPASS(idx < VMEM_MAXORDER);
409 
410 	return &vm->vm_freelist[idx];
411 }
412 
413 /*
414  * bt_freehead_toalloc: return the freelist for the given size and allocation
415  * strategy.
416  *
417  * For M_FIRSTFIT, return the list in which any blocks are large enough
418  * for the requested size.  otherwise, return the list which can have blocks
419  * large enough for the requested size.
420  */
421 static struct vmem_freelist *
422 bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, int strat)
423 {
424 	const vmem_size_t qsize = size >> vm->vm_quantum_shift;
425 	int idx = SIZE2ORDER(qsize);
426 
427 	MPASS(size != 0 && qsize != 0);
428 	MPASS((size & vm->vm_quantum_mask) == 0);
429 
430 	if (strat == M_FIRSTFIT && ORDER2SIZE(idx) != qsize) {
431 		idx++;
432 		/* check too large request? */
433 	}
434 	MPASS(idx >= 0);
435 	MPASS(idx < VMEM_MAXORDER);
436 
437 	return &vm->vm_freelist[idx];
438 }
439 
440 /* ---- boundary tag hash */
441 
442 static struct vmem_hashlist *
443 bt_hashhead(vmem_t *vm, vmem_addr_t addr)
444 {
445 	struct vmem_hashlist *list;
446 	unsigned int hash;
447 
448 	hash = hash32_buf(&addr, sizeof(addr), 0);
449 	list = &vm->vm_hashlist[hash % vm->vm_hashsize];
450 
451 	return list;
452 }
453 
454 static bt_t *
455 bt_lookupbusy(vmem_t *vm, vmem_addr_t addr)
456 {
457 	struct vmem_hashlist *list;
458 	bt_t *bt;
459 
460 	VMEM_ASSERT_LOCKED(vm);
461 	list = bt_hashhead(vm, addr);
462 	LIST_FOREACH(bt, list, bt_hashlist) {
463 		if (bt->bt_start == addr) {
464 			break;
465 		}
466 	}
467 
468 	return bt;
469 }
470 
471 static void
472 bt_rembusy(vmem_t *vm, bt_t *bt)
473 {
474 
475 	VMEM_ASSERT_LOCKED(vm);
476 	MPASS(vm->vm_nbusytag > 0);
477 	vm->vm_inuse -= bt->bt_size;
478 	vm->vm_nbusytag--;
479 	LIST_REMOVE(bt, bt_hashlist);
480 }
481 
482 static void
483 bt_insbusy(vmem_t *vm, bt_t *bt)
484 {
485 	struct vmem_hashlist *list;
486 
487 	VMEM_ASSERT_LOCKED(vm);
488 	MPASS(bt->bt_type == BT_TYPE_BUSY);
489 
490 	list = bt_hashhead(vm, bt->bt_start);
491 	LIST_INSERT_HEAD(list, bt, bt_hashlist);
492 	vm->vm_nbusytag++;
493 	vm->vm_inuse += bt->bt_size;
494 }
495 
496 /* ---- boundary tag list */
497 
498 static void
499 bt_remseg(vmem_t *vm, bt_t *bt)
500 {
501 
502 	MPASS(bt->bt_type != BT_TYPE_CURSOR);
503 	TAILQ_REMOVE(&vm->vm_seglist, bt, bt_seglist);
504 	bt_free(vm, bt);
505 }
506 
507 static void
508 bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev)
509 {
510 
511 	TAILQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist);
512 }
513 
514 static void
515 bt_insseg_tail(vmem_t *vm, bt_t *bt)
516 {
517 
518 	TAILQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist);
519 }
520 
521 static void
522 bt_remfree(vmem_t *vm __unused, bt_t *bt)
523 {
524 
525 	MPASS(bt->bt_type == BT_TYPE_FREE);
526 
527 	LIST_REMOVE(bt, bt_freelist);
528 }
529 
530 static void
531 bt_insfree(vmem_t *vm, bt_t *bt)
532 {
533 	struct vmem_freelist *list;
534 
535 	list = bt_freehead_tofree(vm, bt->bt_size);
536 	LIST_INSERT_HEAD(list, bt, bt_freelist);
537 }
538 
539 /* ---- vmem internal functions */
540 
541 /*
542  * Import from the arena into the quantum cache in UMA.
543  *
544  * We use VMEM_ADDR_QCACHE_MIN instead of 0: uma_zalloc() returns 0 to indicate
545  * failure, so UMA can't be used to cache a resource with value 0.
546  */
547 static int
548 qc_import(void *arg, void **store, int cnt, int domain, int flags)
549 {
550 	qcache_t *qc;
551 	vmem_addr_t addr;
552 	int i;
553 
554 	KASSERT((flags & M_WAITOK) == 0, ("blocking allocation"));
555 
556 	qc = arg;
557 	for (i = 0; i < cnt; i++) {
558 		if (vmem_xalloc(qc->qc_vmem, qc->qc_size, 0, 0, 0,
559 		    VMEM_ADDR_QCACHE_MIN, VMEM_ADDR_MAX, flags, &addr) != 0)
560 			break;
561 		store[i] = (void *)addr;
562 	}
563 	return (i);
564 }
565 
566 /*
567  * Release memory from the UMA cache to the arena.
568  */
569 static void
570 qc_release(void *arg, void **store, int cnt)
571 {
572 	qcache_t *qc;
573 	int i;
574 
575 	qc = arg;
576 	for (i = 0; i < cnt; i++)
577 		vmem_xfree(qc->qc_vmem, (vmem_addr_t)store[i], qc->qc_size);
578 }
579 
580 static void
581 qc_init(vmem_t *vm, vmem_size_t qcache_max)
582 {
583 	qcache_t *qc;
584 	vmem_size_t size;
585 	int qcache_idx_max;
586 	int i;
587 
588 	MPASS((qcache_max & vm->vm_quantum_mask) == 0);
589 	qcache_idx_max = MIN(qcache_max >> vm->vm_quantum_shift,
590 	    VMEM_QCACHE_IDX_MAX);
591 	vm->vm_qcache_max = qcache_idx_max << vm->vm_quantum_shift;
592 	for (i = 0; i < qcache_idx_max; i++) {
593 		qc = &vm->vm_qcache[i];
594 		size = (i + 1) << vm->vm_quantum_shift;
595 		snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu",
596 		    vm->vm_name, size);
597 		qc->qc_vmem = vm;
598 		qc->qc_size = size;
599 		qc->qc_cache = uma_zcache_create(qc->qc_name, size,
600 		    NULL, NULL, NULL, NULL, qc_import, qc_release, qc, 0);
601 		MPASS(qc->qc_cache);
602 	}
603 }
604 
605 static void
606 qc_destroy(vmem_t *vm)
607 {
608 	int qcache_idx_max;
609 	int i;
610 
611 	qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
612 	for (i = 0; i < qcache_idx_max; i++)
613 		uma_zdestroy(vm->vm_qcache[i].qc_cache);
614 }
615 
616 static void
617 qc_drain(vmem_t *vm)
618 {
619 	int qcache_idx_max;
620 	int i;
621 
622 	qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
623 	for (i = 0; i < qcache_idx_max; i++)
624 		uma_zone_reclaim(vm->vm_qcache[i].qc_cache, UMA_RECLAIM_DRAIN);
625 }
626 
627 #ifndef UMA_MD_SMALL_ALLOC
628 
629 static struct mtx_padalign __exclusive_cache_line vmem_bt_lock;
630 
631 /*
632  * vmem_bt_alloc:  Allocate a new page of boundary tags.
633  *
634  * On architectures with uma_small_alloc there is no recursion; no address
635  * space need be allocated to allocate boundary tags.  For the others, we
636  * must handle recursion.  Boundary tags are necessary to allocate new
637  * boundary tags.
638  *
639  * UMA guarantees that enough tags are held in reserve to allocate a new
640  * page of kva.  We dip into this reserve by specifying M_USE_RESERVE only
641  * when allocating the page to hold new boundary tags.  In this way the
642  * reserve is automatically filled by the allocation that uses the reserve.
643  *
644  * We still have to guarantee that the new tags are allocated atomically since
645  * many threads may try concurrently.  The bt_lock provides this guarantee.
646  * We convert WAITOK allocations to NOWAIT and then handle the blocking here
647  * on failure.  It's ok to return NULL for a WAITOK allocation as UMA will
648  * loop again after checking to see if we lost the race to allocate.
649  *
650  * There is a small race between vmem_bt_alloc() returning the page and the
651  * zone lock being acquired to add the page to the zone.  For WAITOK
652  * allocations we just pause briefly.  NOWAIT may experience a transient
653  * failure.  To alleviate this we permit a small number of simultaneous
654  * fills to proceed concurrently so NOWAIT is less likely to fail unless
655  * we are really out of KVA.
656  */
657 static void *
658 vmem_bt_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
659     int wait)
660 {
661 	vmem_addr_t addr;
662 
663 	*pflag = UMA_SLAB_KERNEL;
664 
665 	/*
666 	 * Single thread boundary tag allocation so that the address space
667 	 * and memory are added in one atomic operation.
668 	 */
669 	mtx_lock(&vmem_bt_lock);
670 	if (vmem_xalloc(vm_dom[domain].vmd_kernel_arena, bytes, 0, 0, 0,
671 	    VMEM_ADDR_MIN, VMEM_ADDR_MAX,
672 	    M_NOWAIT | M_NOVM | M_USE_RESERVE | M_BESTFIT, &addr) == 0) {
673 		if (kmem_back_domain(domain, kernel_object, addr, bytes,
674 		    M_NOWAIT | M_USE_RESERVE) == 0) {
675 			mtx_unlock(&vmem_bt_lock);
676 			return ((void *)addr);
677 		}
678 		vmem_xfree(vm_dom[domain].vmd_kernel_arena, addr, bytes);
679 		mtx_unlock(&vmem_bt_lock);
680 		/*
681 		 * Out of memory, not address space.  This may not even be
682 		 * possible due to M_USE_RESERVE page allocation.
683 		 */
684 		if (wait & M_WAITOK)
685 			vm_wait_domain(domain);
686 		return (NULL);
687 	}
688 	mtx_unlock(&vmem_bt_lock);
689 	/*
690 	 * We're either out of address space or lost a fill race.
691 	 */
692 	if (wait & M_WAITOK)
693 		pause("btalloc", 1);
694 
695 	return (NULL);
696 }
697 #endif
698 
699 void
700 vmem_startup(void)
701 {
702 
703 	mtx_init(&vmem_list_lock, "vmem list lock", NULL, MTX_DEF);
704 	vmem_zone = uma_zcreate("vmem",
705 	    sizeof(struct vmem), NULL, NULL, NULL, NULL,
706 	    UMA_ALIGN_PTR, 0);
707 	vmem_bt_zone = uma_zcreate("vmem btag",
708 	    sizeof(struct vmem_btag), NULL, NULL, NULL, NULL,
709 	    UMA_ALIGN_PTR, UMA_ZONE_VM);
710 #ifndef UMA_MD_SMALL_ALLOC
711 	mtx_init(&vmem_bt_lock, "btag lock", NULL, MTX_DEF);
712 	uma_prealloc(vmem_bt_zone, BT_MAXALLOC);
713 	/*
714 	 * Reserve enough tags to allocate new tags.  We allow multiple
715 	 * CPUs to attempt to allocate new tags concurrently to limit
716 	 * false restarts in UMA.  vmem_bt_alloc() allocates from a per-domain
717 	 * arena, which may involve importing a range from the kernel arena,
718 	 * so we need to keep at least 2 * BT_MAXALLOC tags reserved.
719 	 */
720 	uma_zone_reserve(vmem_bt_zone, 2 * BT_MAXALLOC * mp_ncpus);
721 	uma_zone_set_allocf(vmem_bt_zone, vmem_bt_alloc);
722 #endif
723 }
724 
725 /* ---- rehash */
726 
727 static int
728 vmem_rehash(vmem_t *vm, vmem_size_t newhashsize)
729 {
730 	bt_t *bt;
731 	struct vmem_hashlist *newhashlist;
732 	struct vmem_hashlist *oldhashlist;
733 	vmem_size_t i, oldhashsize;
734 
735 	MPASS(newhashsize > 0);
736 
737 	newhashlist = malloc(sizeof(struct vmem_hashlist) * newhashsize,
738 	    M_VMEM, M_NOWAIT);
739 	if (newhashlist == NULL)
740 		return ENOMEM;
741 	for (i = 0; i < newhashsize; i++) {
742 		LIST_INIT(&newhashlist[i]);
743 	}
744 
745 	VMEM_LOCK(vm);
746 	oldhashlist = vm->vm_hashlist;
747 	oldhashsize = vm->vm_hashsize;
748 	vm->vm_hashlist = newhashlist;
749 	vm->vm_hashsize = newhashsize;
750 	if (oldhashlist == NULL) {
751 		VMEM_UNLOCK(vm);
752 		return 0;
753 	}
754 	for (i = 0; i < oldhashsize; i++) {
755 		while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) {
756 			bt_rembusy(vm, bt);
757 			bt_insbusy(vm, bt);
758 		}
759 	}
760 	VMEM_UNLOCK(vm);
761 
762 	if (oldhashlist != vm->vm_hash0)
763 		free(oldhashlist, M_VMEM);
764 
765 	return 0;
766 }
767 
768 static void
769 vmem_periodic_kick(void *dummy)
770 {
771 
772 	taskqueue_enqueue(taskqueue_thread, &vmem_periodic_wk);
773 }
774 
775 static void
776 vmem_periodic(void *unused, int pending)
777 {
778 	vmem_t *vm;
779 	vmem_size_t desired;
780 	vmem_size_t current;
781 
782 	mtx_lock(&vmem_list_lock);
783 	LIST_FOREACH(vm, &vmem_list, vm_alllist) {
784 #ifdef DIAGNOSTIC
785 		/* Convenient time to verify vmem state. */
786 		if (enable_vmem_check == 1) {
787 			VMEM_LOCK(vm);
788 			vmem_check(vm);
789 			VMEM_UNLOCK(vm);
790 		}
791 #endif
792 		desired = 1 << flsl(vm->vm_nbusytag);
793 		desired = MIN(MAX(desired, VMEM_HASHSIZE_MIN),
794 		    VMEM_HASHSIZE_MAX);
795 		current = vm->vm_hashsize;
796 
797 		/* Grow in powers of two.  Shrink less aggressively. */
798 		if (desired >= current * 2 || desired * 4 <= current)
799 			vmem_rehash(vm, desired);
800 
801 		/*
802 		 * Periodically wake up threads waiting for resources,
803 		 * so they could ask for reclamation again.
804 		 */
805 		VMEM_CONDVAR_BROADCAST(vm);
806 	}
807 	mtx_unlock(&vmem_list_lock);
808 
809 	callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
810 	    vmem_periodic_kick, NULL);
811 }
812 
813 static void
814 vmem_start_callout(void *unused)
815 {
816 
817 	TASK_INIT(&vmem_periodic_wk, 0, vmem_periodic, NULL);
818 	vmem_periodic_interval = hz * 10;
819 	callout_init(&vmem_periodic_ch, 1);
820 	callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
821 	    vmem_periodic_kick, NULL);
822 }
823 SYSINIT(vfs, SI_SUB_CONFIGURE, SI_ORDER_ANY, vmem_start_callout, NULL);
824 
825 static void
826 vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int type)
827 {
828 	bt_t *btfree, *btprev, *btspan;
829 
830 	VMEM_ASSERT_LOCKED(vm);
831 	MPASS(type == BT_TYPE_SPAN || type == BT_TYPE_SPAN_STATIC);
832 	MPASS((size & vm->vm_quantum_mask) == 0);
833 
834 	if (vm->vm_releasefn == NULL) {
835 		/*
836 		 * The new segment will never be released, so see if it is
837 		 * contiguous with respect to an existing segment.  In this case
838 		 * a span tag is not needed, and it may be possible now or in
839 		 * the future to coalesce the new segment with an existing free
840 		 * segment.
841 		 */
842 		btprev = TAILQ_LAST(&vm->vm_seglist, vmem_seglist);
843 		if ((!bt_isbusy(btprev) && !bt_isfree(btprev)) ||
844 		    btprev->bt_start + btprev->bt_size != addr)
845 			btprev = NULL;
846 	} else {
847 		btprev = NULL;
848 	}
849 
850 	if (btprev == NULL || bt_isbusy(btprev)) {
851 		if (btprev == NULL) {
852 			btspan = bt_alloc(vm);
853 			btspan->bt_type = type;
854 			btspan->bt_start = addr;
855 			btspan->bt_size = size;
856 			bt_insseg_tail(vm, btspan);
857 		}
858 
859 		btfree = bt_alloc(vm);
860 		btfree->bt_type = BT_TYPE_FREE;
861 		btfree->bt_start = addr;
862 		btfree->bt_size = size;
863 		bt_insseg_tail(vm, btfree);
864 		bt_insfree(vm, btfree);
865 	} else {
866 		bt_remfree(vm, btprev);
867 		btprev->bt_size += size;
868 		bt_insfree(vm, btprev);
869 	}
870 
871 	vm->vm_size += size;
872 }
873 
874 static void
875 vmem_destroy1(vmem_t *vm)
876 {
877 	bt_t *bt;
878 
879 	/*
880 	 * Drain per-cpu quantum caches.
881 	 */
882 	qc_destroy(vm);
883 
884 	/*
885 	 * The vmem should now only contain empty segments.
886 	 */
887 	VMEM_LOCK(vm);
888 	MPASS(vm->vm_nbusytag == 0);
889 
890 	TAILQ_REMOVE(&vm->vm_seglist, &vm->vm_cursor, bt_seglist);
891 	while ((bt = TAILQ_FIRST(&vm->vm_seglist)) != NULL)
892 		bt_remseg(vm, bt);
893 
894 	if (vm->vm_hashlist != NULL && vm->vm_hashlist != vm->vm_hash0)
895 		free(vm->vm_hashlist, M_VMEM);
896 
897 	bt_freetrim(vm, 0);
898 
899 	VMEM_CONDVAR_DESTROY(vm);
900 	VMEM_LOCK_DESTROY(vm);
901 	uma_zfree(vmem_zone, vm);
902 }
903 
904 static int
905 vmem_import(vmem_t *vm, vmem_size_t size, vmem_size_t align, int flags)
906 {
907 	vmem_addr_t addr;
908 	int error;
909 
910 	if (vm->vm_importfn == NULL)
911 		return (EINVAL);
912 
913 	/*
914 	 * To make sure we get a span that meets the alignment we double it
915 	 * and add the size to the tail.  This slightly overestimates.
916 	 */
917 	if (align != vm->vm_quantum_mask + 1)
918 		size = (align * 2) + size;
919 	size = roundup(size, vm->vm_import_quantum);
920 
921 	if (vm->vm_limit != 0 && vm->vm_limit < vm->vm_size + size)
922 		return (ENOMEM);
923 
924 	bt_save(vm);
925 	VMEM_UNLOCK(vm);
926 	error = (vm->vm_importfn)(vm->vm_arg, size, flags, &addr);
927 	VMEM_LOCK(vm);
928 	bt_restore(vm);
929 	if (error)
930 		return (ENOMEM);
931 
932 	vmem_add1(vm, addr, size, BT_TYPE_SPAN);
933 
934 	return 0;
935 }
936 
937 /*
938  * vmem_fit: check if a bt can satisfy the given restrictions.
939  *
940  * it's a caller's responsibility to ensure the region is big enough
941  * before calling us.
942  */
943 static int
944 vmem_fit(const bt_t *bt, vmem_size_t size, vmem_size_t align,
945     vmem_size_t phase, vmem_size_t nocross, vmem_addr_t minaddr,
946     vmem_addr_t maxaddr, vmem_addr_t *addrp)
947 {
948 	vmem_addr_t start;
949 	vmem_addr_t end;
950 
951 	MPASS(size > 0);
952 	MPASS(bt->bt_size >= size); /* caller's responsibility */
953 
954 	/*
955 	 * XXX assumption: vmem_addr_t and vmem_size_t are
956 	 * unsigned integer of the same size.
957 	 */
958 
959 	start = bt->bt_start;
960 	if (start < minaddr) {
961 		start = minaddr;
962 	}
963 	end = BT_END(bt);
964 	if (end > maxaddr)
965 		end = maxaddr;
966 	if (start > end)
967 		return (ENOMEM);
968 
969 	start = VMEM_ALIGNUP(start - phase, align) + phase;
970 	if (start < bt->bt_start)
971 		start += align;
972 	if (VMEM_CROSS_P(start, start + size - 1, nocross)) {
973 		MPASS(align < nocross);
974 		start = VMEM_ALIGNUP(start - phase, nocross) + phase;
975 	}
976 	if (start <= end && end - start >= size - 1) {
977 		MPASS((start & (align - 1)) == phase);
978 		MPASS(!VMEM_CROSS_P(start, start + size - 1, nocross));
979 		MPASS(minaddr <= start);
980 		MPASS(maxaddr == 0 || start + size - 1 <= maxaddr);
981 		MPASS(bt->bt_start <= start);
982 		MPASS(BT_END(bt) - start >= size - 1);
983 		*addrp = start;
984 
985 		return (0);
986 	}
987 	return (ENOMEM);
988 }
989 
990 /*
991  * vmem_clip:  Trim the boundary tag edges to the requested start and size.
992  */
993 static void
994 vmem_clip(vmem_t *vm, bt_t *bt, vmem_addr_t start, vmem_size_t size)
995 {
996 	bt_t *btnew;
997 	bt_t *btprev;
998 
999 	VMEM_ASSERT_LOCKED(vm);
1000 	MPASS(bt->bt_type == BT_TYPE_FREE);
1001 	MPASS(bt->bt_size >= size);
1002 	bt_remfree(vm, bt);
1003 	if (bt->bt_start != start) {
1004 		btprev = bt_alloc(vm);
1005 		btprev->bt_type = BT_TYPE_FREE;
1006 		btprev->bt_start = bt->bt_start;
1007 		btprev->bt_size = start - bt->bt_start;
1008 		bt->bt_start = start;
1009 		bt->bt_size -= btprev->bt_size;
1010 		bt_insfree(vm, btprev);
1011 		bt_insseg(vm, btprev,
1012 		    TAILQ_PREV(bt, vmem_seglist, bt_seglist));
1013 	}
1014 	MPASS(bt->bt_start == start);
1015 	if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) {
1016 		/* split */
1017 		btnew = bt_alloc(vm);
1018 		btnew->bt_type = BT_TYPE_BUSY;
1019 		btnew->bt_start = bt->bt_start;
1020 		btnew->bt_size = size;
1021 		bt->bt_start = bt->bt_start + size;
1022 		bt->bt_size -= size;
1023 		bt_insfree(vm, bt);
1024 		bt_insseg(vm, btnew,
1025 		    TAILQ_PREV(bt, vmem_seglist, bt_seglist));
1026 		bt_insbusy(vm, btnew);
1027 		bt = btnew;
1028 	} else {
1029 		bt->bt_type = BT_TYPE_BUSY;
1030 		bt_insbusy(vm, bt);
1031 	}
1032 	MPASS(bt->bt_size >= size);
1033 }
1034 
1035 static int
1036 vmem_try_fetch(vmem_t *vm, const vmem_size_t size, vmem_size_t align, int flags)
1037 {
1038 	vmem_size_t avail;
1039 
1040 	VMEM_ASSERT_LOCKED(vm);
1041 
1042 	/*
1043 	 * XXX it is possible to fail to meet xalloc constraints with the
1044 	 * imported region.  It is up to the user to specify the
1045 	 * import quantum such that it can satisfy any allocation.
1046 	 */
1047 	if (vmem_import(vm, size, align, flags) == 0)
1048 		return (1);
1049 
1050 	/*
1051 	 * Try to free some space from the quantum cache or reclaim
1052 	 * functions if available.
1053 	 */
1054 	if (vm->vm_qcache_max != 0 || vm->vm_reclaimfn != NULL) {
1055 		avail = vm->vm_size - vm->vm_inuse;
1056 		bt_save(vm);
1057 		VMEM_UNLOCK(vm);
1058 		if (vm->vm_qcache_max != 0)
1059 			qc_drain(vm);
1060 		if (vm->vm_reclaimfn != NULL)
1061 			vm->vm_reclaimfn(vm, flags);
1062 		VMEM_LOCK(vm);
1063 		bt_restore(vm);
1064 		/* If we were successful retry even NOWAIT. */
1065 		if (vm->vm_size - vm->vm_inuse > avail)
1066 			return (1);
1067 	}
1068 	if ((flags & M_NOWAIT) != 0)
1069 		return (0);
1070 	bt_save(vm);
1071 	VMEM_CONDVAR_WAIT(vm);
1072 	bt_restore(vm);
1073 	return (1);
1074 }
1075 
1076 static int
1077 vmem_try_release(vmem_t *vm, struct vmem_btag *bt, const bool remfree)
1078 {
1079 	struct vmem_btag *prev;
1080 
1081 	MPASS(bt->bt_type == BT_TYPE_FREE);
1082 
1083 	if (vm->vm_releasefn == NULL)
1084 		return (0);
1085 
1086 	prev = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1087 	MPASS(prev != NULL);
1088 	MPASS(prev->bt_type != BT_TYPE_FREE);
1089 
1090 	if (prev->bt_type == BT_TYPE_SPAN && prev->bt_size == bt->bt_size) {
1091 		vmem_addr_t spanaddr;
1092 		vmem_size_t spansize;
1093 
1094 		MPASS(prev->bt_start == bt->bt_start);
1095 		spanaddr = prev->bt_start;
1096 		spansize = prev->bt_size;
1097 		if (remfree)
1098 			bt_remfree(vm, bt);
1099 		bt_remseg(vm, bt);
1100 		bt_remseg(vm, prev);
1101 		vm->vm_size -= spansize;
1102 		VMEM_CONDVAR_BROADCAST(vm);
1103 		bt_freetrim(vm, BT_MAXFREE);
1104 		vm->vm_releasefn(vm->vm_arg, spanaddr, spansize);
1105 		return (1);
1106 	}
1107 	return (0);
1108 }
1109 
1110 static int
1111 vmem_xalloc_nextfit(vmem_t *vm, const vmem_size_t size, vmem_size_t align,
1112     const vmem_size_t phase, const vmem_size_t nocross, int flags,
1113     vmem_addr_t *addrp)
1114 {
1115 	struct vmem_btag *bt, *cursor, *next, *prev;
1116 	int error;
1117 
1118 	error = ENOMEM;
1119 	VMEM_LOCK(vm);
1120 
1121 	/*
1122 	 * Make sure we have enough tags to complete the operation.
1123 	 */
1124 	if (bt_fill(vm, flags) != 0)
1125 		goto out;
1126 
1127 retry:
1128 	/*
1129 	 * Find the next free tag meeting our constraints.  If one is found,
1130 	 * perform the allocation.
1131 	 */
1132 	for (cursor = &vm->vm_cursor, bt = TAILQ_NEXT(cursor, bt_seglist);
1133 	    bt != cursor; bt = TAILQ_NEXT(bt, bt_seglist)) {
1134 		if (bt == NULL)
1135 			bt = TAILQ_FIRST(&vm->vm_seglist);
1136 		if (bt->bt_type == BT_TYPE_FREE && bt->bt_size >= size &&
1137 		    (error = vmem_fit(bt, size, align, phase, nocross,
1138 		    VMEM_ADDR_MIN, VMEM_ADDR_MAX, addrp)) == 0) {
1139 			vmem_clip(vm, bt, *addrp, size);
1140 			break;
1141 		}
1142 	}
1143 
1144 	/*
1145 	 * Try to coalesce free segments around the cursor.  If we succeed, and
1146 	 * have not yet satisfied the allocation request, try again with the
1147 	 * newly coalesced segment.
1148 	 */
1149 	if ((next = TAILQ_NEXT(cursor, bt_seglist)) != NULL &&
1150 	    (prev = TAILQ_PREV(cursor, vmem_seglist, bt_seglist)) != NULL &&
1151 	    next->bt_type == BT_TYPE_FREE && prev->bt_type == BT_TYPE_FREE &&
1152 	    prev->bt_start + prev->bt_size == next->bt_start) {
1153 		prev->bt_size += next->bt_size;
1154 		bt_remfree(vm, next);
1155 		bt_remseg(vm, next);
1156 
1157 		/*
1158 		 * The coalesced segment might be able to satisfy our request.
1159 		 * If not, we might need to release it from the arena.
1160 		 */
1161 		if (error == ENOMEM && prev->bt_size >= size &&
1162 		    (error = vmem_fit(prev, size, align, phase, nocross,
1163 		    VMEM_ADDR_MIN, VMEM_ADDR_MAX, addrp)) == 0) {
1164 			vmem_clip(vm, prev, *addrp, size);
1165 			bt = prev;
1166 		} else
1167 			(void)vmem_try_release(vm, prev, true);
1168 	}
1169 
1170 	/*
1171 	 * If the allocation was successful, advance the cursor.
1172 	 */
1173 	if (error == 0) {
1174 		TAILQ_REMOVE(&vm->vm_seglist, cursor, bt_seglist);
1175 		for (; bt != NULL && bt->bt_start < *addrp + size;
1176 		    bt = TAILQ_NEXT(bt, bt_seglist))
1177 			;
1178 		if (bt != NULL)
1179 			TAILQ_INSERT_BEFORE(bt, cursor, bt_seglist);
1180 		else
1181 			TAILQ_INSERT_HEAD(&vm->vm_seglist, cursor, bt_seglist);
1182 	}
1183 
1184 	/*
1185 	 * Attempt to bring additional resources into the arena.  If that fails
1186 	 * and M_WAITOK is specified, sleep waiting for resources to be freed.
1187 	 */
1188 	if (error == ENOMEM && vmem_try_fetch(vm, size, align, flags))
1189 		goto retry;
1190 
1191 out:
1192 	VMEM_UNLOCK(vm);
1193 	return (error);
1194 }
1195 
1196 /* ---- vmem API */
1197 
1198 void
1199 vmem_set_import(vmem_t *vm, vmem_import_t *importfn,
1200      vmem_release_t *releasefn, void *arg, vmem_size_t import_quantum)
1201 {
1202 
1203 	VMEM_LOCK(vm);
1204 	KASSERT(vm->vm_size == 0, ("%s: arena is non-empty", __func__));
1205 	vm->vm_importfn = importfn;
1206 	vm->vm_releasefn = releasefn;
1207 	vm->vm_arg = arg;
1208 	vm->vm_import_quantum = import_quantum;
1209 	VMEM_UNLOCK(vm);
1210 }
1211 
1212 void
1213 vmem_set_limit(vmem_t *vm, vmem_size_t limit)
1214 {
1215 
1216 	VMEM_LOCK(vm);
1217 	vm->vm_limit = limit;
1218 	VMEM_UNLOCK(vm);
1219 }
1220 
1221 void
1222 vmem_set_reclaim(vmem_t *vm, vmem_reclaim_t *reclaimfn)
1223 {
1224 
1225 	VMEM_LOCK(vm);
1226 	vm->vm_reclaimfn = reclaimfn;
1227 	VMEM_UNLOCK(vm);
1228 }
1229 
1230 /*
1231  * vmem_init: Initializes vmem arena.
1232  */
1233 vmem_t *
1234 vmem_init(vmem_t *vm, const char *name, vmem_addr_t base, vmem_size_t size,
1235     vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1236 {
1237 	vmem_size_t i;
1238 
1239 	MPASS(quantum > 0);
1240 	MPASS((quantum & (quantum - 1)) == 0);
1241 
1242 	bzero(vm, sizeof(*vm));
1243 
1244 	VMEM_CONDVAR_INIT(vm, name);
1245 	VMEM_LOCK_INIT(vm, name);
1246 	vm->vm_nfreetags = 0;
1247 	LIST_INIT(&vm->vm_freetags);
1248 	strlcpy(vm->vm_name, name, sizeof(vm->vm_name));
1249 	vm->vm_quantum_mask = quantum - 1;
1250 	vm->vm_quantum_shift = flsl(quantum) - 1;
1251 	vm->vm_nbusytag = 0;
1252 	vm->vm_size = 0;
1253 	vm->vm_limit = 0;
1254 	vm->vm_inuse = 0;
1255 	qc_init(vm, qcache_max);
1256 
1257 	TAILQ_INIT(&vm->vm_seglist);
1258 	vm->vm_cursor.bt_start = vm->vm_cursor.bt_size = 0;
1259 	vm->vm_cursor.bt_type = BT_TYPE_CURSOR;
1260 	TAILQ_INSERT_TAIL(&vm->vm_seglist, &vm->vm_cursor, bt_seglist);
1261 
1262 	for (i = 0; i < VMEM_MAXORDER; i++)
1263 		LIST_INIT(&vm->vm_freelist[i]);
1264 
1265 	memset(&vm->vm_hash0, 0, sizeof(vm->vm_hash0));
1266 	vm->vm_hashsize = VMEM_HASHSIZE_MIN;
1267 	vm->vm_hashlist = vm->vm_hash0;
1268 
1269 	if (size != 0) {
1270 		if (vmem_add(vm, base, size, flags) != 0) {
1271 			vmem_destroy1(vm);
1272 			return NULL;
1273 		}
1274 	}
1275 
1276 	mtx_lock(&vmem_list_lock);
1277 	LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist);
1278 	mtx_unlock(&vmem_list_lock);
1279 
1280 	return vm;
1281 }
1282 
1283 /*
1284  * vmem_create: create an arena.
1285  */
1286 vmem_t *
1287 vmem_create(const char *name, vmem_addr_t base, vmem_size_t size,
1288     vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1289 {
1290 
1291 	vmem_t *vm;
1292 
1293 	vm = uma_zalloc(vmem_zone, flags & (M_WAITOK|M_NOWAIT));
1294 	if (vm == NULL)
1295 		return (NULL);
1296 	if (vmem_init(vm, name, base, size, quantum, qcache_max,
1297 	    flags) == NULL)
1298 		return (NULL);
1299 	return (vm);
1300 }
1301 
1302 void
1303 vmem_destroy(vmem_t *vm)
1304 {
1305 
1306 	mtx_lock(&vmem_list_lock);
1307 	LIST_REMOVE(vm, vm_alllist);
1308 	mtx_unlock(&vmem_list_lock);
1309 
1310 	vmem_destroy1(vm);
1311 }
1312 
1313 vmem_size_t
1314 vmem_roundup_size(vmem_t *vm, vmem_size_t size)
1315 {
1316 
1317 	return (size + vm->vm_quantum_mask) & ~vm->vm_quantum_mask;
1318 }
1319 
1320 /*
1321  * vmem_alloc: allocate resource from the arena.
1322  */
1323 int
1324 vmem_alloc(vmem_t *vm, vmem_size_t size, int flags, vmem_addr_t *addrp)
1325 {
1326 	const int strat __unused = flags & VMEM_FITMASK;
1327 	qcache_t *qc;
1328 
1329 	flags &= VMEM_FLAGS;
1330 	MPASS(size > 0);
1331 	MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT || strat == M_NEXTFIT);
1332 	if ((flags & M_NOWAIT) == 0)
1333 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_alloc");
1334 
1335 	if (size <= vm->vm_qcache_max) {
1336 		/*
1337 		 * Resource 0 cannot be cached, so avoid a blocking allocation
1338 		 * in qc_import() and give the vmem_xalloc() call below a chance
1339 		 * to return 0.
1340 		 */
1341 		qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1342 		*addrp = (vmem_addr_t)uma_zalloc(qc->qc_cache,
1343 		    (flags & ~M_WAITOK) | M_NOWAIT);
1344 		if (__predict_true(*addrp != 0))
1345 			return (0);
1346 	}
1347 
1348 	return (vmem_xalloc(vm, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
1349 	    flags, addrp));
1350 }
1351 
1352 int
1353 vmem_xalloc(vmem_t *vm, const vmem_size_t size0, vmem_size_t align,
1354     const vmem_size_t phase, const vmem_size_t nocross,
1355     const vmem_addr_t minaddr, const vmem_addr_t maxaddr, int flags,
1356     vmem_addr_t *addrp)
1357 {
1358 	const vmem_size_t size = vmem_roundup_size(vm, size0);
1359 	struct vmem_freelist *list;
1360 	struct vmem_freelist *first;
1361 	struct vmem_freelist *end;
1362 	bt_t *bt;
1363 	int error;
1364 	int strat;
1365 
1366 	flags &= VMEM_FLAGS;
1367 	strat = flags & VMEM_FITMASK;
1368 	MPASS(size0 > 0);
1369 	MPASS(size > 0);
1370 	MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT || strat == M_NEXTFIT);
1371 	MPASS((flags & (M_NOWAIT|M_WAITOK)) != (M_NOWAIT|M_WAITOK));
1372 	if ((flags & M_NOWAIT) == 0)
1373 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_xalloc");
1374 	MPASS((align & vm->vm_quantum_mask) == 0);
1375 	MPASS((align & (align - 1)) == 0);
1376 	MPASS((phase & vm->vm_quantum_mask) == 0);
1377 	MPASS((nocross & vm->vm_quantum_mask) == 0);
1378 	MPASS((nocross & (nocross - 1)) == 0);
1379 	MPASS((align == 0 && phase == 0) || phase < align);
1380 	MPASS(nocross == 0 || nocross >= size);
1381 	MPASS(minaddr <= maxaddr);
1382 	MPASS(!VMEM_CROSS_P(phase, phase + size - 1, nocross));
1383 	if (strat == M_NEXTFIT)
1384 		MPASS(minaddr == VMEM_ADDR_MIN && maxaddr == VMEM_ADDR_MAX);
1385 
1386 	if (align == 0)
1387 		align = vm->vm_quantum_mask + 1;
1388 	*addrp = 0;
1389 
1390 	/*
1391 	 * Next-fit allocations don't use the freelists.
1392 	 */
1393 	if (strat == M_NEXTFIT)
1394 		return (vmem_xalloc_nextfit(vm, size0, align, phase, nocross,
1395 		    flags, addrp));
1396 
1397 	end = &vm->vm_freelist[VMEM_MAXORDER];
1398 	/*
1399 	 * choose a free block from which we allocate.
1400 	 */
1401 	first = bt_freehead_toalloc(vm, size, strat);
1402 	VMEM_LOCK(vm);
1403 
1404 	/*
1405 	 * Make sure we have enough tags to complete the operation.
1406 	 */
1407 	error = bt_fill(vm, flags);
1408 	if (error != 0)
1409 		goto out;
1410 	for (;;) {
1411 		/*
1412 	 	 * Scan freelists looking for a tag that satisfies the
1413 		 * allocation.  If we're doing BESTFIT we may encounter
1414 		 * sizes below the request.  If we're doing FIRSTFIT we
1415 		 * inspect only the first element from each list.
1416 		 */
1417 		for (list = first; list < end; list++) {
1418 			LIST_FOREACH(bt, list, bt_freelist) {
1419 				if (bt->bt_size >= size) {
1420 					error = vmem_fit(bt, size, align, phase,
1421 					    nocross, minaddr, maxaddr, addrp);
1422 					if (error == 0) {
1423 						vmem_clip(vm, bt, *addrp, size);
1424 						goto out;
1425 					}
1426 				}
1427 				/* FIRST skips to the next list. */
1428 				if (strat == M_FIRSTFIT)
1429 					break;
1430 			}
1431 		}
1432 
1433 		/*
1434 		 * Retry if the fast algorithm failed.
1435 		 */
1436 		if (strat == M_FIRSTFIT) {
1437 			strat = M_BESTFIT;
1438 			first = bt_freehead_toalloc(vm, size, strat);
1439 			continue;
1440 		}
1441 
1442 		/*
1443 		 * Try a few measures to bring additional resources into the
1444 		 * arena.  If all else fails, we will sleep waiting for
1445 		 * resources to be freed.
1446 		 */
1447 		if (!vmem_try_fetch(vm, size, align, flags)) {
1448 			error = ENOMEM;
1449 			break;
1450 		}
1451 	}
1452 out:
1453 	VMEM_UNLOCK(vm);
1454 	if (error != 0 && (flags & M_NOWAIT) == 0)
1455 		panic("failed to allocate waiting allocation\n");
1456 
1457 	return (error);
1458 }
1459 
1460 /*
1461  * vmem_free: free the resource to the arena.
1462  */
1463 void
1464 vmem_free(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1465 {
1466 	qcache_t *qc;
1467 	MPASS(size > 0);
1468 
1469 	if (size <= vm->vm_qcache_max &&
1470 	    __predict_true(addr >= VMEM_ADDR_QCACHE_MIN)) {
1471 		qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1472 		uma_zfree(qc->qc_cache, (void *)addr);
1473 	} else
1474 		vmem_xfree(vm, addr, size);
1475 }
1476 
1477 void
1478 vmem_xfree(vmem_t *vm, vmem_addr_t addr, vmem_size_t size __unused)
1479 {
1480 	bt_t *bt;
1481 	bt_t *t;
1482 
1483 	MPASS(size > 0);
1484 
1485 	VMEM_LOCK(vm);
1486 	bt = bt_lookupbusy(vm, addr);
1487 	MPASS(bt != NULL);
1488 	MPASS(bt->bt_start == addr);
1489 	MPASS(bt->bt_size == vmem_roundup_size(vm, size) ||
1490 	    bt->bt_size - vmem_roundup_size(vm, size) <= vm->vm_quantum_mask);
1491 	MPASS(bt->bt_type == BT_TYPE_BUSY);
1492 	bt_rembusy(vm, bt);
1493 	bt->bt_type = BT_TYPE_FREE;
1494 
1495 	/* coalesce */
1496 	t = TAILQ_NEXT(bt, bt_seglist);
1497 	if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1498 		MPASS(BT_END(bt) < t->bt_start);	/* YYY */
1499 		bt->bt_size += t->bt_size;
1500 		bt_remfree(vm, t);
1501 		bt_remseg(vm, t);
1502 	}
1503 	t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1504 	if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1505 		MPASS(BT_END(t) < bt->bt_start);	/* YYY */
1506 		bt->bt_size += t->bt_size;
1507 		bt->bt_start = t->bt_start;
1508 		bt_remfree(vm, t);
1509 		bt_remseg(vm, t);
1510 	}
1511 
1512 	if (!vmem_try_release(vm, bt, false)) {
1513 		bt_insfree(vm, bt);
1514 		VMEM_CONDVAR_BROADCAST(vm);
1515 		bt_freetrim(vm, BT_MAXFREE);
1516 	}
1517 }
1518 
1519 /*
1520  * vmem_add:
1521  *
1522  */
1523 int
1524 vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int flags)
1525 {
1526 	int error;
1527 
1528 	flags &= VMEM_FLAGS;
1529 
1530 	VMEM_LOCK(vm);
1531 	error = bt_fill(vm, flags);
1532 	if (error == 0)
1533 		vmem_add1(vm, addr, size, BT_TYPE_SPAN_STATIC);
1534 	VMEM_UNLOCK(vm);
1535 
1536 	return (error);
1537 }
1538 
1539 /*
1540  * vmem_size: information about arenas size
1541  */
1542 vmem_size_t
1543 vmem_size(vmem_t *vm, int typemask)
1544 {
1545 	int i;
1546 
1547 	switch (typemask) {
1548 	case VMEM_ALLOC:
1549 		return vm->vm_inuse;
1550 	case VMEM_FREE:
1551 		return vm->vm_size - vm->vm_inuse;
1552 	case VMEM_FREE|VMEM_ALLOC:
1553 		return vm->vm_size;
1554 	case VMEM_MAXFREE:
1555 		VMEM_LOCK(vm);
1556 		for (i = VMEM_MAXORDER - 1; i >= 0; i--) {
1557 			if (LIST_EMPTY(&vm->vm_freelist[i]))
1558 				continue;
1559 			VMEM_UNLOCK(vm);
1560 			return ((vmem_size_t)ORDER2SIZE(i) <<
1561 			    vm->vm_quantum_shift);
1562 		}
1563 		VMEM_UNLOCK(vm);
1564 		return (0);
1565 	default:
1566 		panic("vmem_size");
1567 	}
1568 }
1569 
1570 /* ---- debug */
1571 
1572 #if defined(DDB) || defined(DIAGNOSTIC)
1573 
1574 static void bt_dump(const bt_t *, int (*)(const char *, ...)
1575     __printflike(1, 2));
1576 
1577 static const char *
1578 bt_type_string(int type)
1579 {
1580 
1581 	switch (type) {
1582 	case BT_TYPE_BUSY:
1583 		return "busy";
1584 	case BT_TYPE_FREE:
1585 		return "free";
1586 	case BT_TYPE_SPAN:
1587 		return "span";
1588 	case BT_TYPE_SPAN_STATIC:
1589 		return "static span";
1590 	case BT_TYPE_CURSOR:
1591 		return "cursor";
1592 	default:
1593 		break;
1594 	}
1595 	return "BOGUS";
1596 }
1597 
1598 static void
1599 bt_dump(const bt_t *bt, int (*pr)(const char *, ...))
1600 {
1601 
1602 	(*pr)("\t%p: %jx %jx, %d(%s)\n",
1603 	    bt, (intmax_t)bt->bt_start, (intmax_t)bt->bt_size,
1604 	    bt->bt_type, bt_type_string(bt->bt_type));
1605 }
1606 
1607 static void
1608 vmem_dump(const vmem_t *vm , int (*pr)(const char *, ...) __printflike(1, 2))
1609 {
1610 	const bt_t *bt;
1611 	int i;
1612 
1613 	(*pr)("vmem %p '%s'\n", vm, vm->vm_name);
1614 	TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1615 		bt_dump(bt, pr);
1616 	}
1617 
1618 	for (i = 0; i < VMEM_MAXORDER; i++) {
1619 		const struct vmem_freelist *fl = &vm->vm_freelist[i];
1620 
1621 		if (LIST_EMPTY(fl)) {
1622 			continue;
1623 		}
1624 
1625 		(*pr)("freelist[%d]\n", i);
1626 		LIST_FOREACH(bt, fl, bt_freelist) {
1627 			bt_dump(bt, pr);
1628 		}
1629 	}
1630 }
1631 
1632 #endif /* defined(DDB) || defined(DIAGNOSTIC) */
1633 
1634 #if defined(DDB)
1635 #include <ddb/ddb.h>
1636 
1637 static bt_t *
1638 vmem_whatis_lookup(vmem_t *vm, vmem_addr_t addr)
1639 {
1640 	bt_t *bt;
1641 
1642 	TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1643 		if (BT_ISSPAN_P(bt)) {
1644 			continue;
1645 		}
1646 		if (bt->bt_start <= addr && addr <= BT_END(bt)) {
1647 			return bt;
1648 		}
1649 	}
1650 
1651 	return NULL;
1652 }
1653 
1654 void
1655 vmem_whatis(vmem_addr_t addr, int (*pr)(const char *, ...))
1656 {
1657 	vmem_t *vm;
1658 
1659 	LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1660 		bt_t *bt;
1661 
1662 		bt = vmem_whatis_lookup(vm, addr);
1663 		if (bt == NULL) {
1664 			continue;
1665 		}
1666 		(*pr)("%p is %p+%zu in VMEM '%s' (%s)\n",
1667 		    (void *)addr, (void *)bt->bt_start,
1668 		    (vmem_size_t)(addr - bt->bt_start), vm->vm_name,
1669 		    (bt->bt_type == BT_TYPE_BUSY) ? "allocated" : "free");
1670 	}
1671 }
1672 
1673 void
1674 vmem_printall(const char *modif, int (*pr)(const char *, ...))
1675 {
1676 	const vmem_t *vm;
1677 
1678 	LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1679 		vmem_dump(vm, pr);
1680 	}
1681 }
1682 
1683 void
1684 vmem_print(vmem_addr_t addr, const char *modif, int (*pr)(const char *, ...))
1685 {
1686 	const vmem_t *vm = (const void *)addr;
1687 
1688 	vmem_dump(vm, pr);
1689 }
1690 
1691 DB_SHOW_COMMAND(vmemdump, vmemdump)
1692 {
1693 
1694 	if (!have_addr) {
1695 		db_printf("usage: show vmemdump <addr>\n");
1696 		return;
1697 	}
1698 
1699 	vmem_dump((const vmem_t *)addr, db_printf);
1700 }
1701 
1702 DB_SHOW_ALL_COMMAND(vmemdump, vmemdumpall)
1703 {
1704 	const vmem_t *vm;
1705 
1706 	LIST_FOREACH(vm, &vmem_list, vm_alllist)
1707 		vmem_dump(vm, db_printf);
1708 }
1709 
1710 DB_SHOW_COMMAND(vmem, vmem_summ)
1711 {
1712 	const vmem_t *vm = (const void *)addr;
1713 	const bt_t *bt;
1714 	size_t ft[VMEM_MAXORDER], ut[VMEM_MAXORDER];
1715 	size_t fs[VMEM_MAXORDER], us[VMEM_MAXORDER];
1716 	int ord;
1717 
1718 	if (!have_addr) {
1719 		db_printf("usage: show vmem <addr>\n");
1720 		return;
1721 	}
1722 
1723 	db_printf("vmem %p '%s'\n", vm, vm->vm_name);
1724 	db_printf("\tquantum:\t%zu\n", vm->vm_quantum_mask + 1);
1725 	db_printf("\tsize:\t%zu\n", vm->vm_size);
1726 	db_printf("\tinuse:\t%zu\n", vm->vm_inuse);
1727 	db_printf("\tfree:\t%zu\n", vm->vm_size - vm->vm_inuse);
1728 	db_printf("\tbusy tags:\t%d\n", vm->vm_nbusytag);
1729 	db_printf("\tfree tags:\t%d\n", vm->vm_nfreetags);
1730 
1731 	memset(&ft, 0, sizeof(ft));
1732 	memset(&ut, 0, sizeof(ut));
1733 	memset(&fs, 0, sizeof(fs));
1734 	memset(&us, 0, sizeof(us));
1735 	TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1736 		ord = SIZE2ORDER(bt->bt_size >> vm->vm_quantum_shift);
1737 		if (bt->bt_type == BT_TYPE_BUSY) {
1738 			ut[ord]++;
1739 			us[ord] += bt->bt_size;
1740 		} else if (bt->bt_type == BT_TYPE_FREE) {
1741 			ft[ord]++;
1742 			fs[ord] += bt->bt_size;
1743 		}
1744 	}
1745 	db_printf("\t\t\tinuse\tsize\t\tfree\tsize\n");
1746 	for (ord = 0; ord < VMEM_MAXORDER; ord++) {
1747 		if (ut[ord] == 0 && ft[ord] == 0)
1748 			continue;
1749 		db_printf("\t%-15zu %zu\t%-15zu %zu\t%-16zu\n",
1750 		    ORDER2SIZE(ord) << vm->vm_quantum_shift,
1751 		    ut[ord], us[ord], ft[ord], fs[ord]);
1752 	}
1753 }
1754 
1755 DB_SHOW_ALL_COMMAND(vmem, vmem_summall)
1756 {
1757 	const vmem_t *vm;
1758 
1759 	LIST_FOREACH(vm, &vmem_list, vm_alllist)
1760 		vmem_summ((db_expr_t)vm, TRUE, count, modif);
1761 }
1762 #endif /* defined(DDB) */
1763 
1764 #define vmem_printf printf
1765 
1766 #if defined(DIAGNOSTIC)
1767 
1768 static bool
1769 vmem_check_sanity(vmem_t *vm)
1770 {
1771 	const bt_t *bt, *bt2;
1772 
1773 	MPASS(vm != NULL);
1774 
1775 	TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1776 		if (bt->bt_start > BT_END(bt)) {
1777 			printf("corrupted tag\n");
1778 			bt_dump(bt, vmem_printf);
1779 			return false;
1780 		}
1781 	}
1782 	TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1783 		if (bt->bt_type == BT_TYPE_CURSOR) {
1784 			if (bt->bt_start != 0 || bt->bt_size != 0) {
1785 				printf("corrupted cursor\n");
1786 				return false;
1787 			}
1788 			continue;
1789 		}
1790 		TAILQ_FOREACH(bt2, &vm->vm_seglist, bt_seglist) {
1791 			if (bt == bt2) {
1792 				continue;
1793 			}
1794 			if (bt2->bt_type == BT_TYPE_CURSOR) {
1795 				continue;
1796 			}
1797 			if (BT_ISSPAN_P(bt) != BT_ISSPAN_P(bt2)) {
1798 				continue;
1799 			}
1800 			if (bt->bt_start <= BT_END(bt2) &&
1801 			    bt2->bt_start <= BT_END(bt)) {
1802 				printf("overwrapped tags\n");
1803 				bt_dump(bt, vmem_printf);
1804 				bt_dump(bt2, vmem_printf);
1805 				return false;
1806 			}
1807 		}
1808 	}
1809 
1810 	return true;
1811 }
1812 
1813 static void
1814 vmem_check(vmem_t *vm)
1815 {
1816 
1817 	if (!vmem_check_sanity(vm)) {
1818 		panic("insanity vmem %p", vm);
1819 	}
1820 }
1821 
1822 #endif /* defined(DIAGNOSTIC) */
1823