xref: /freebsd/sys/kern/subr_vmem.c (revision cc68614da8232d8baaca0ae0d0dd8f890f06623e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
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 __FBSDID("$FreeBSD$");
45 
46 #include "opt_ddb.h"
47 
48 #include <sys/param.h>
49 #include <sys/systm.h>
50 #include <sys/kernel.h>
51 #include <sys/queue.h>
52 #include <sys/callout.h>
53 #include <sys/hash.h>
54 #include <sys/lock.h>
55 #include <sys/malloc.h>
56 #include <sys/mutex.h>
57 #include <sys/smp.h>
58 #include <sys/condvar.h>
59 #include <sys/sysctl.h>
60 #include <sys/taskqueue.h>
61 #include <sys/vmem.h>
62 #include <sys/vmmeter.h>
63 
64 #include "opt_vm.h"
65 
66 #include <vm/uma.h>
67 #include <vm/vm.h>
68 #include <vm/pmap.h>
69 #include <vm/vm_map.h>
70 #include <vm/vm_object.h>
71 #include <vm/vm_kern.h>
72 #include <vm/vm_extern.h>
73 #include <vm/vm_param.h>
74 #include <vm/vm_page.h>
75 #include <vm/vm_pageout.h>
76 #include <vm/vm_phys.h>
77 #include <vm/vm_pagequeue.h>
78 #include <vm/uma_int.h>
79 
80 #define	VMEM_OPTORDER		5
81 #define	VMEM_OPTVALUE		(1 << VMEM_OPTORDER)
82 #define	VMEM_MAXORDER						\
83     (VMEM_OPTVALUE - 1 + sizeof(vmem_size_t) * NBBY - VMEM_OPTORDER)
84 
85 #define	VMEM_HASHSIZE_MIN	16
86 #define	VMEM_HASHSIZE_MAX	131072
87 
88 #define	VMEM_QCACHE_IDX_MAX	16
89 
90 #define	VMEM_FITMASK	(M_BESTFIT | M_FIRSTFIT | M_NEXTFIT)
91 
92 #define	VMEM_FLAGS	(M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM |	\
93     M_BESTFIT | M_FIRSTFIT | M_NEXTFIT)
94 
95 #define	BT_FLAGS	(M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM)
96 
97 #define	QC_NAME_MAX	16
98 
99 /*
100  * Data structures private to vmem.
101  */
102 MALLOC_DEFINE(M_VMEM, "vmem", "vmem internal structures");
103 
104 typedef struct vmem_btag bt_t;
105 
106 TAILQ_HEAD(vmem_seglist, vmem_btag);
107 LIST_HEAD(vmem_freelist, vmem_btag);
108 LIST_HEAD(vmem_hashlist, vmem_btag);
109 
110 struct qcache {
111 	uma_zone_t	qc_cache;
112 	vmem_t 		*qc_vmem;
113 	vmem_size_t	qc_size;
114 	char		qc_name[QC_NAME_MAX];
115 };
116 typedef struct qcache qcache_t;
117 #define	QC_POOL_TO_QCACHE(pool)	((qcache_t *)(pool->pr_qcache))
118 
119 #define	VMEM_NAME_MAX	16
120 
121 /* boundary tag */
122 struct vmem_btag {
123 	TAILQ_ENTRY(vmem_btag) bt_seglist;
124 	union {
125 		LIST_ENTRY(vmem_btag) u_freelist; /* BT_TYPE_FREE */
126 		LIST_ENTRY(vmem_btag) u_hashlist; /* BT_TYPE_BUSY */
127 	} bt_u;
128 #define	bt_hashlist	bt_u.u_hashlist
129 #define	bt_freelist	bt_u.u_freelist
130 	vmem_addr_t	bt_start;
131 	vmem_size_t	bt_size;
132 	int		bt_type;
133 };
134 
135 /* vmem arena */
136 struct vmem {
137 	struct mtx_padalign	vm_lock;
138 	struct cv		vm_cv;
139 	char			vm_name[VMEM_NAME_MAX+1];
140 	LIST_ENTRY(vmem)	vm_alllist;
141 	struct vmem_hashlist	vm_hash0[VMEM_HASHSIZE_MIN];
142 	struct vmem_freelist	vm_freelist[VMEM_MAXORDER];
143 	struct vmem_seglist	vm_seglist;
144 	struct vmem_hashlist	*vm_hashlist;
145 	vmem_size_t		vm_hashsize;
146 
147 	/* Constant after init */
148 	vmem_size_t		vm_qcache_max;
149 	vmem_size_t		vm_quantum_mask;
150 	vmem_size_t		vm_import_quantum;
151 	int			vm_quantum_shift;
152 
153 	/* Written on alloc/free */
154 	LIST_HEAD(, vmem_btag)	vm_freetags;
155 	int			vm_nfreetags;
156 	int			vm_nbusytag;
157 	vmem_size_t		vm_inuse;
158 	vmem_size_t		vm_size;
159 	vmem_size_t		vm_limit;
160 	struct vmem_btag	vm_cursor;
161 
162 	/* Used on import. */
163 	vmem_import_t		*vm_importfn;
164 	vmem_release_t		*vm_releasefn;
165 	void			*vm_arg;
166 
167 	/* Space exhaustion callback. */
168 	vmem_reclaim_t		*vm_reclaimfn;
169 
170 	/* quantum cache */
171 	qcache_t		vm_qcache[VMEM_QCACHE_IDX_MAX];
172 };
173 
174 #define	BT_TYPE_SPAN		1	/* Allocated from importfn */
175 #define	BT_TYPE_SPAN_STATIC	2	/* vmem_add() or create. */
176 #define	BT_TYPE_FREE		3	/* Available space. */
177 #define	BT_TYPE_BUSY		4	/* Used space. */
178 #define	BT_TYPE_CURSOR		5	/* Cursor for nextfit allocations. */
179 #define	BT_ISSPAN_P(bt)	((bt)->bt_type <= BT_TYPE_SPAN_STATIC)
180 
181 #define	BT_END(bt)	((bt)->bt_start + (bt)->bt_size - 1)
182 
183 #if defined(DIAGNOSTIC)
184 static int enable_vmem_check = 0;
185 SYSCTL_INT(_debug, OID_AUTO, vmem_check, CTLFLAG_RWTUN,
186     &enable_vmem_check, 0, "Enable vmem check");
187 static void vmem_check(vmem_t *);
188 #endif
189 
190 static struct callout	vmem_periodic_ch;
191 static int		vmem_periodic_interval;
192 static struct task	vmem_periodic_wk;
193 
194 static struct mtx_padalign __exclusive_cache_line vmem_list_lock;
195 static LIST_HEAD(, vmem) vmem_list = LIST_HEAD_INITIALIZER(vmem_list);
196 static uma_zone_t vmem_zone;
197 
198 /* ---- misc */
199 #define	VMEM_CONDVAR_INIT(vm, wchan)	cv_init(&vm->vm_cv, wchan)
200 #define	VMEM_CONDVAR_DESTROY(vm)	cv_destroy(&vm->vm_cv)
201 #define	VMEM_CONDVAR_WAIT(vm)		cv_wait(&vm->vm_cv, &vm->vm_lock)
202 #define	VMEM_CONDVAR_BROADCAST(vm)	cv_broadcast(&vm->vm_cv)
203 
204 #define	VMEM_LOCK(vm)		mtx_lock(&vm->vm_lock)
205 #define	VMEM_TRYLOCK(vm)	mtx_trylock(&vm->vm_lock)
206 #define	VMEM_UNLOCK(vm)		mtx_unlock(&vm->vm_lock)
207 #define	VMEM_LOCK_INIT(vm, name) mtx_init(&vm->vm_lock, (name), NULL, MTX_DEF)
208 #define	VMEM_LOCK_DESTROY(vm)	mtx_destroy(&vm->vm_lock)
209 #define	VMEM_ASSERT_LOCKED(vm)	mtx_assert(&vm->vm_lock, MA_OWNED);
210 
211 #define	VMEM_ALIGNUP(addr, align)	(-(-(addr) & -(align)))
212 
213 #define	VMEM_CROSS_P(addr1, addr2, boundary) \
214 	((((addr1) ^ (addr2)) & -(boundary)) != 0)
215 
216 #define	ORDER2SIZE(order)	((order) < VMEM_OPTVALUE ? ((order) + 1) : \
217     (vmem_size_t)1 << ((order) - (VMEM_OPTVALUE - VMEM_OPTORDER - 1)))
218 #define	SIZE2ORDER(size)	((size) <= VMEM_OPTVALUE ? ((size) - 1) : \
219     (flsl(size) + (VMEM_OPTVALUE - VMEM_OPTORDER - 2)))
220 
221 /*
222  * Maximum number of boundary tags that may be required to satisfy an
223  * allocation.  Two may be required to import.  Another two may be
224  * required to clip edges.
225  */
226 #define	BT_MAXALLOC	4
227 
228 /*
229  * Max free limits the number of locally cached boundary tags.  We
230  * just want to avoid hitting the zone allocator for every call.
231  */
232 #define BT_MAXFREE	(BT_MAXALLOC * 8)
233 
234 /* Allocator for boundary tags. */
235 static uma_zone_t vmem_bt_zone;
236 
237 /* boot time arena storage. */
238 static struct vmem kernel_arena_storage;
239 static struct vmem buffer_arena_storage;
240 static struct vmem transient_arena_storage;
241 /* kernel and kmem arenas are aliased for backwards KPI compat. */
242 vmem_t *kernel_arena = &kernel_arena_storage;
243 vmem_t *kmem_arena = &kernel_arena_storage;
244 vmem_t *buffer_arena = &buffer_arena_storage;
245 vmem_t *transient_arena = &transient_arena_storage;
246 
247 #ifdef DEBUG_MEMGUARD
248 static struct vmem memguard_arena_storage;
249 vmem_t *memguard_arena = &memguard_arena_storage;
250 #endif
251 
252 static bool
253 bt_isbusy(bt_t *bt)
254 {
255 	return (bt->bt_type == BT_TYPE_BUSY);
256 }
257 
258 static bool
259 bt_isfree(bt_t *bt)
260 {
261 	return (bt->bt_type == BT_TYPE_FREE);
262 }
263 
264 /*
265  * Fill the vmem's boundary tag cache.  We guarantee that boundary tag
266  * allocation will not fail once bt_fill() passes.  To do so we cache
267  * at least the maximum possible tag allocations in the arena.
268  */
269 static __noinline int
270 _bt_fill(vmem_t *vm, int flags)
271 {
272 	bt_t *bt;
273 
274 	VMEM_ASSERT_LOCKED(vm);
275 
276 	/*
277 	 * Only allow the kernel arena and arenas derived from kernel arena to
278 	 * dip into reserve tags.  They are where new tags come from.
279 	 */
280 	flags &= BT_FLAGS;
281 	if (vm != kernel_arena && vm->vm_arg != kernel_arena)
282 		flags &= ~M_USE_RESERVE;
283 
284 	/*
285 	 * Loop until we meet the reserve.  To minimize the lock shuffle
286 	 * and prevent simultaneous fills we first try a NOWAIT regardless
287 	 * of the caller's flags.  Specify M_NOVM so we don't recurse while
288 	 * holding a vmem lock.
289 	 */
290 	while (vm->vm_nfreetags < BT_MAXALLOC) {
291 		bt = uma_zalloc(vmem_bt_zone,
292 		    (flags & M_USE_RESERVE) | M_NOWAIT | M_NOVM);
293 		if (bt == NULL) {
294 			VMEM_UNLOCK(vm);
295 			bt = uma_zalloc(vmem_bt_zone, flags);
296 			VMEM_LOCK(vm);
297 			if (bt == NULL)
298 				break;
299 		}
300 		LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
301 		vm->vm_nfreetags++;
302 	}
303 
304 	if (vm->vm_nfreetags < BT_MAXALLOC)
305 		return ENOMEM;
306 
307 	return 0;
308 }
309 
310 static inline int
311 bt_fill(vmem_t *vm, int flags)
312 {
313 	if (vm->vm_nfreetags >= BT_MAXALLOC)
314 		return (0);
315 	return (_bt_fill(vm, flags));
316 }
317 
318 /*
319  * Pop a tag off of the freetag stack.
320  */
321 static bt_t *
322 bt_alloc(vmem_t *vm)
323 {
324 	bt_t *bt;
325 
326 	VMEM_ASSERT_LOCKED(vm);
327 	bt = LIST_FIRST(&vm->vm_freetags);
328 	MPASS(bt != NULL);
329 	LIST_REMOVE(bt, bt_freelist);
330 	vm->vm_nfreetags--;
331 
332 	return bt;
333 }
334 
335 /*
336  * Trim the per-vmem free list.  Returns with the lock released to
337  * avoid allocator recursions.
338  */
339 static void
340 bt_freetrim(vmem_t *vm, int freelimit)
341 {
342 	LIST_HEAD(, vmem_btag) freetags;
343 	bt_t *bt;
344 
345 	LIST_INIT(&freetags);
346 	VMEM_ASSERT_LOCKED(vm);
347 	while (vm->vm_nfreetags > freelimit) {
348 		bt = LIST_FIRST(&vm->vm_freetags);
349 		LIST_REMOVE(bt, bt_freelist);
350 		vm->vm_nfreetags--;
351 		LIST_INSERT_HEAD(&freetags, bt, bt_freelist);
352 	}
353 	VMEM_UNLOCK(vm);
354 	while ((bt = LIST_FIRST(&freetags)) != NULL) {
355 		LIST_REMOVE(bt, bt_freelist);
356 		uma_zfree(vmem_bt_zone, bt);
357 	}
358 }
359 
360 static inline void
361 bt_free(vmem_t *vm, bt_t *bt)
362 {
363 
364 	VMEM_ASSERT_LOCKED(vm);
365 	MPASS(LIST_FIRST(&vm->vm_freetags) != bt);
366 	LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
367 	vm->vm_nfreetags++;
368 }
369 
370 /*
371  * Hide MAXALLOC tags before dropping the arena lock to ensure that a
372  * concurrent allocation attempt does not grab them.
373  */
374 static void
375 bt_save(vmem_t *vm)
376 {
377 	KASSERT(vm->vm_nfreetags >= BT_MAXALLOC,
378 	    ("%s: insufficient free tags %d", __func__, vm->vm_nfreetags));
379 	vm->vm_nfreetags -= BT_MAXALLOC;
380 }
381 
382 static void
383 bt_restore(vmem_t *vm)
384 {
385 	vm->vm_nfreetags += BT_MAXALLOC;
386 }
387 
388 /*
389  * freelist[0] ... [1, 1]
390  * freelist[1] ... [2, 2]
391  *  :
392  * freelist[29] ... [30, 30]
393  * freelist[30] ... [31, 31]
394  * freelist[31] ... [32, 63]
395  * freelist[33] ... [64, 127]
396  *  :
397  * freelist[n] ... [(1 << (n - 26)), (1 << (n - 25)) - 1]
398  *  :
399  */
400 
401 static struct vmem_freelist *
402 bt_freehead_tofree(vmem_t *vm, vmem_size_t size)
403 {
404 	const vmem_size_t qsize = size >> vm->vm_quantum_shift;
405 	const int idx = SIZE2ORDER(qsize);
406 
407 	MPASS(size != 0 && qsize != 0);
408 	MPASS((size & vm->vm_quantum_mask) == 0);
409 	MPASS(idx >= 0);
410 	MPASS(idx < VMEM_MAXORDER);
411 
412 	return &vm->vm_freelist[idx];
413 }
414 
415 /*
416  * bt_freehead_toalloc: return the freelist for the given size and allocation
417  * strategy.
418  *
419  * For M_FIRSTFIT, return the list in which any blocks are large enough
420  * for the requested size.  otherwise, return the list which can have blocks
421  * large enough for the requested size.
422  */
423 static struct vmem_freelist *
424 bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, int strat)
425 {
426 	const vmem_size_t qsize = size >> vm->vm_quantum_shift;
427 	int idx = SIZE2ORDER(qsize);
428 
429 	MPASS(size != 0 && qsize != 0);
430 	MPASS((size & vm->vm_quantum_mask) == 0);
431 
432 	if (strat == M_FIRSTFIT && ORDER2SIZE(idx) != qsize) {
433 		idx++;
434 		/* check too large request? */
435 	}
436 	MPASS(idx >= 0);
437 	MPASS(idx < VMEM_MAXORDER);
438 
439 	return &vm->vm_freelist[idx];
440 }
441 
442 /* ---- boundary tag hash */
443 
444 static struct vmem_hashlist *
445 bt_hashhead(vmem_t *vm, vmem_addr_t addr)
446 {
447 	struct vmem_hashlist *list;
448 	unsigned int hash;
449 
450 	hash = hash32_buf(&addr, sizeof(addr), 0);
451 	list = &vm->vm_hashlist[hash % vm->vm_hashsize];
452 
453 	return list;
454 }
455 
456 static bt_t *
457 bt_lookupbusy(vmem_t *vm, vmem_addr_t addr)
458 {
459 	struct vmem_hashlist *list;
460 	bt_t *bt;
461 
462 	VMEM_ASSERT_LOCKED(vm);
463 	list = bt_hashhead(vm, addr);
464 	LIST_FOREACH(bt, list, bt_hashlist) {
465 		if (bt->bt_start == addr) {
466 			break;
467 		}
468 	}
469 
470 	return bt;
471 }
472 
473 static void
474 bt_rembusy(vmem_t *vm, bt_t *bt)
475 {
476 
477 	VMEM_ASSERT_LOCKED(vm);
478 	MPASS(vm->vm_nbusytag > 0);
479 	vm->vm_inuse -= bt->bt_size;
480 	vm->vm_nbusytag--;
481 	LIST_REMOVE(bt, bt_hashlist);
482 }
483 
484 static void
485 bt_insbusy(vmem_t *vm, bt_t *bt)
486 {
487 	struct vmem_hashlist *list;
488 
489 	VMEM_ASSERT_LOCKED(vm);
490 	MPASS(bt->bt_type == BT_TYPE_BUSY);
491 
492 	list = bt_hashhead(vm, bt->bt_start);
493 	LIST_INSERT_HEAD(list, bt, bt_hashlist);
494 	vm->vm_nbusytag++;
495 	vm->vm_inuse += bt->bt_size;
496 }
497 
498 /* ---- boundary tag list */
499 
500 static void
501 bt_remseg(vmem_t *vm, bt_t *bt)
502 {
503 
504 	MPASS(bt->bt_type != BT_TYPE_CURSOR);
505 	TAILQ_REMOVE(&vm->vm_seglist, bt, bt_seglist);
506 	bt_free(vm, bt);
507 }
508 
509 static void
510 bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev)
511 {
512 
513 	TAILQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist);
514 }
515 
516 static void
517 bt_insseg_tail(vmem_t *vm, bt_t *bt)
518 {
519 
520 	TAILQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist);
521 }
522 
523 static void
524 bt_remfree(vmem_t *vm __unused, bt_t *bt)
525 {
526 
527 	MPASS(bt->bt_type == BT_TYPE_FREE);
528 
529 	LIST_REMOVE(bt, bt_freelist);
530 }
531 
532 static void
533 bt_insfree(vmem_t *vm, bt_t *bt)
534 {
535 	struct vmem_freelist *list;
536 
537 	list = bt_freehead_tofree(vm, bt->bt_size);
538 	LIST_INSERT_HEAD(list, bt, bt_freelist);
539 }
540 
541 /* ---- vmem internal functions */
542 
543 /*
544  * Import from the arena into the quantum cache in UMA.
545  *
546  * We use VMEM_ADDR_QCACHE_MIN instead of 0: uma_zalloc() returns 0 to indicate
547  * failure, so UMA can't be used to cache a resource with value 0.
548  */
549 static int
550 qc_import(void *arg, void **store, int cnt, int domain, int flags)
551 {
552 	qcache_t *qc;
553 	vmem_addr_t addr;
554 	int i;
555 
556 	KASSERT((flags & M_WAITOK) == 0, ("blocking allocation"));
557 
558 	qc = arg;
559 	for (i = 0; i < cnt; i++) {
560 		if (vmem_xalloc(qc->qc_vmem, qc->qc_size, 0, 0, 0,
561 		    VMEM_ADDR_QCACHE_MIN, VMEM_ADDR_MAX, flags, &addr) != 0)
562 			break;
563 		store[i] = (void *)addr;
564 	}
565 	return (i);
566 }
567 
568 /*
569  * Release memory from the UMA cache to the arena.
570  */
571 static void
572 qc_release(void *arg, void **store, int cnt)
573 {
574 	qcache_t *qc;
575 	int i;
576 
577 	qc = arg;
578 	for (i = 0; i < cnt; i++)
579 		vmem_xfree(qc->qc_vmem, (vmem_addr_t)store[i], qc->qc_size);
580 }
581 
582 static void
583 qc_init(vmem_t *vm, vmem_size_t qcache_max)
584 {
585 	qcache_t *qc;
586 	vmem_size_t size;
587 	int qcache_idx_max;
588 	int i;
589 
590 	MPASS((qcache_max & vm->vm_quantum_mask) == 0);
591 	qcache_idx_max = MIN(qcache_max >> vm->vm_quantum_shift,
592 	    VMEM_QCACHE_IDX_MAX);
593 	vm->vm_qcache_max = qcache_idx_max << vm->vm_quantum_shift;
594 	for (i = 0; i < qcache_idx_max; i++) {
595 		qc = &vm->vm_qcache[i];
596 		size = (i + 1) << vm->vm_quantum_shift;
597 		snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu",
598 		    vm->vm_name, size);
599 		qc->qc_vmem = vm;
600 		qc->qc_size = size;
601 		qc->qc_cache = uma_zcache_create(qc->qc_name, size,
602 		    NULL, NULL, NULL, NULL, qc_import, qc_release, qc, 0);
603 		MPASS(qc->qc_cache);
604 	}
605 }
606 
607 static void
608 qc_destroy(vmem_t *vm)
609 {
610 	int qcache_idx_max;
611 	int i;
612 
613 	qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
614 	for (i = 0; i < qcache_idx_max; i++)
615 		uma_zdestroy(vm->vm_qcache[i].qc_cache);
616 }
617 
618 static void
619 qc_drain(vmem_t *vm)
620 {
621 	int qcache_idx_max;
622 	int i;
623 
624 	qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
625 	for (i = 0; i < qcache_idx_max; i++)
626 		uma_zone_reclaim(vm->vm_qcache[i].qc_cache, UMA_RECLAIM_DRAIN);
627 }
628 
629 #ifndef UMA_MD_SMALL_ALLOC
630 
631 static struct mtx_padalign __exclusive_cache_line vmem_bt_lock;
632 
633 /*
634  * vmem_bt_alloc:  Allocate a new page of boundary tags.
635  *
636  * On architectures with uma_small_alloc there is no recursion; no address
637  * space need be allocated to allocate boundary tags.  For the others, we
638  * must handle recursion.  Boundary tags are necessary to allocate new
639  * boundary tags.
640  *
641  * UMA guarantees that enough tags are held in reserve to allocate a new
642  * page of kva.  We dip into this reserve by specifying M_USE_RESERVE only
643  * when allocating the page to hold new boundary tags.  In this way the
644  * reserve is automatically filled by the allocation that uses the reserve.
645  *
646  * We still have to guarantee that the new tags are allocated atomically since
647  * many threads may try concurrently.  The bt_lock provides this guarantee.
648  * We convert WAITOK allocations to NOWAIT and then handle the blocking here
649  * on failure.  It's ok to return NULL for a WAITOK allocation as UMA will
650  * loop again after checking to see if we lost the race to allocate.
651  *
652  * There is a small race between vmem_bt_alloc() returning the page and the
653  * zone lock being acquired to add the page to the zone.  For WAITOK
654  * allocations we just pause briefly.  NOWAIT may experience a transient
655  * failure.  To alleviate this we permit a small number of simultaneous
656  * fills to proceed concurrently so NOWAIT is less likely to fail unless
657  * we are really out of KVA.
658  */
659 static void *
660 vmem_bt_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
661     int wait)
662 {
663 	vmem_addr_t addr;
664 
665 	*pflag = UMA_SLAB_KERNEL;
666 
667 	/*
668 	 * Single thread boundary tag allocation so that the address space
669 	 * and memory are added in one atomic operation.
670 	 */
671 	mtx_lock(&vmem_bt_lock);
672 	if (vmem_xalloc(vm_dom[domain].vmd_kernel_arena, bytes, 0, 0, 0,
673 	    VMEM_ADDR_MIN, VMEM_ADDR_MAX,
674 	    M_NOWAIT | M_NOVM | M_USE_RESERVE | M_BESTFIT, &addr) == 0) {
675 		if (kmem_back_domain(domain, kernel_object, addr, bytes,
676 		    M_NOWAIT | M_USE_RESERVE) == 0) {
677 			mtx_unlock(&vmem_bt_lock);
678 			return ((void *)addr);
679 		}
680 		vmem_xfree(vm_dom[domain].vmd_kernel_arena, addr, bytes);
681 		mtx_unlock(&vmem_bt_lock);
682 		/*
683 		 * Out of memory, not address space.  This may not even be
684 		 * possible due to M_USE_RESERVE page allocation.
685 		 */
686 		if (wait & M_WAITOK)
687 			vm_wait_domain(domain);
688 		return (NULL);
689 	}
690 	mtx_unlock(&vmem_bt_lock);
691 	/*
692 	 * We're either out of address space or lost a fill race.
693 	 */
694 	if (wait & M_WAITOK)
695 		pause("btalloc", 1);
696 
697 	return (NULL);
698 }
699 #endif
700 
701 void
702 vmem_startup(void)
703 {
704 
705 	mtx_init(&vmem_list_lock, "vmem list lock", NULL, MTX_DEF);
706 	vmem_zone = uma_zcreate("vmem",
707 	    sizeof(struct vmem), NULL, NULL, NULL, NULL,
708 	    UMA_ALIGN_PTR, 0);
709 	vmem_bt_zone = uma_zcreate("vmem btag",
710 	    sizeof(struct vmem_btag), NULL, NULL, NULL, NULL,
711 	    UMA_ALIGN_PTR, UMA_ZONE_VM);
712 #ifndef UMA_MD_SMALL_ALLOC
713 	mtx_init(&vmem_bt_lock, "btag lock", NULL, MTX_DEF);
714 	uma_prealloc(vmem_bt_zone, BT_MAXALLOC);
715 	/*
716 	 * Reserve enough tags to allocate new tags.  We allow multiple
717 	 * CPUs to attempt to allocate new tags concurrently to limit
718 	 * false restarts in UMA.  vmem_bt_alloc() allocates from a per-domain
719 	 * arena, which may involve importing a range from the kernel arena,
720 	 * so we need to keep at least 2 * BT_MAXALLOC tags reserved.
721 	 */
722 	uma_zone_reserve(vmem_bt_zone, 2 * BT_MAXALLOC * mp_ncpus);
723 	uma_zone_set_allocf(vmem_bt_zone, vmem_bt_alloc);
724 #endif
725 }
726 
727 /* ---- rehash */
728 
729 static int
730 vmem_rehash(vmem_t *vm, vmem_size_t newhashsize)
731 {
732 	bt_t *bt;
733 	struct vmem_hashlist *newhashlist;
734 	struct vmem_hashlist *oldhashlist;
735 	vmem_size_t i, oldhashsize;
736 
737 	MPASS(newhashsize > 0);
738 
739 	newhashlist = malloc(sizeof(struct vmem_hashlist) * newhashsize,
740 	    M_VMEM, M_NOWAIT);
741 	if (newhashlist == NULL)
742 		return ENOMEM;
743 	for (i = 0; i < newhashsize; i++) {
744 		LIST_INIT(&newhashlist[i]);
745 	}
746 
747 	VMEM_LOCK(vm);
748 	oldhashlist = vm->vm_hashlist;
749 	oldhashsize = vm->vm_hashsize;
750 	vm->vm_hashlist = newhashlist;
751 	vm->vm_hashsize = newhashsize;
752 	if (oldhashlist == NULL) {
753 		VMEM_UNLOCK(vm);
754 		return 0;
755 	}
756 	for (i = 0; i < oldhashsize; i++) {
757 		while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) {
758 			bt_rembusy(vm, bt);
759 			bt_insbusy(vm, bt);
760 		}
761 	}
762 	VMEM_UNLOCK(vm);
763 
764 	if (oldhashlist != vm->vm_hash0)
765 		free(oldhashlist, M_VMEM);
766 
767 	return 0;
768 }
769 
770 static void
771 vmem_periodic_kick(void *dummy)
772 {
773 
774 	taskqueue_enqueue(taskqueue_thread, &vmem_periodic_wk);
775 }
776 
777 static void
778 vmem_periodic(void *unused, int pending)
779 {
780 	vmem_t *vm;
781 	vmem_size_t desired;
782 	vmem_size_t current;
783 
784 	mtx_lock(&vmem_list_lock);
785 	LIST_FOREACH(vm, &vmem_list, vm_alllist) {
786 #ifdef DIAGNOSTIC
787 		/* Convenient time to verify vmem state. */
788 		if (enable_vmem_check == 1) {
789 			VMEM_LOCK(vm);
790 			vmem_check(vm);
791 			VMEM_UNLOCK(vm);
792 		}
793 #endif
794 		desired = 1 << flsl(vm->vm_nbusytag);
795 		desired = MIN(MAX(desired, VMEM_HASHSIZE_MIN),
796 		    VMEM_HASHSIZE_MAX);
797 		current = vm->vm_hashsize;
798 
799 		/* Grow in powers of two.  Shrink less aggressively. */
800 		if (desired >= current * 2 || desired * 4 <= current)
801 			vmem_rehash(vm, desired);
802 
803 		/*
804 		 * Periodically wake up threads waiting for resources,
805 		 * so they could ask for reclamation again.
806 		 */
807 		VMEM_CONDVAR_BROADCAST(vm);
808 	}
809 	mtx_unlock(&vmem_list_lock);
810 
811 	callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
812 	    vmem_periodic_kick, NULL);
813 }
814 
815 static void
816 vmem_start_callout(void *unused)
817 {
818 
819 	TASK_INIT(&vmem_periodic_wk, 0, vmem_periodic, NULL);
820 	vmem_periodic_interval = hz * 10;
821 	callout_init(&vmem_periodic_ch, 1);
822 	callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
823 	    vmem_periodic_kick, NULL);
824 }
825 SYSINIT(vfs, SI_SUB_CONFIGURE, SI_ORDER_ANY, vmem_start_callout, NULL);
826 
827 static void
828 vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int type)
829 {
830 	bt_t *btfree, *btprev, *btspan;
831 
832 	VMEM_ASSERT_LOCKED(vm);
833 	MPASS(type == BT_TYPE_SPAN || type == BT_TYPE_SPAN_STATIC);
834 	MPASS((size & vm->vm_quantum_mask) == 0);
835 
836 	if (vm->vm_releasefn == NULL) {
837 		/*
838 		 * The new segment will never be released, so see if it is
839 		 * contiguous with respect to an existing segment.  In this case
840 		 * a span tag is not needed, and it may be possible now or in
841 		 * the future to coalesce the new segment with an existing free
842 		 * segment.
843 		 */
844 		btprev = TAILQ_LAST(&vm->vm_seglist, vmem_seglist);
845 		if ((!bt_isbusy(btprev) && !bt_isfree(btprev)) ||
846 		    btprev->bt_start + btprev->bt_size != addr)
847 			btprev = NULL;
848 	} else {
849 		btprev = NULL;
850 	}
851 
852 	if (btprev == NULL || bt_isbusy(btprev)) {
853 		if (btprev == NULL) {
854 			btspan = bt_alloc(vm);
855 			btspan->bt_type = type;
856 			btspan->bt_start = addr;
857 			btspan->bt_size = size;
858 			bt_insseg_tail(vm, btspan);
859 		}
860 
861 		btfree = bt_alloc(vm);
862 		btfree->bt_type = BT_TYPE_FREE;
863 		btfree->bt_start = addr;
864 		btfree->bt_size = size;
865 		bt_insseg_tail(vm, btfree);
866 		bt_insfree(vm, btfree);
867 	} else {
868 		bt_remfree(vm, btprev);
869 		btprev->bt_size += size;
870 		bt_insfree(vm, btprev);
871 	}
872 
873 	vm->vm_size += size;
874 }
875 
876 static void
877 vmem_destroy1(vmem_t *vm)
878 {
879 	bt_t *bt;
880 
881 	/*
882 	 * Drain per-cpu quantum caches.
883 	 */
884 	qc_destroy(vm);
885 
886 	/*
887 	 * The vmem should now only contain empty segments.
888 	 */
889 	VMEM_LOCK(vm);
890 	MPASS(vm->vm_nbusytag == 0);
891 
892 	TAILQ_REMOVE(&vm->vm_seglist, &vm->vm_cursor, bt_seglist);
893 	while ((bt = TAILQ_FIRST(&vm->vm_seglist)) != NULL)
894 		bt_remseg(vm, bt);
895 
896 	if (vm->vm_hashlist != NULL && vm->vm_hashlist != vm->vm_hash0)
897 		free(vm->vm_hashlist, M_VMEM);
898 
899 	bt_freetrim(vm, 0);
900 
901 	VMEM_CONDVAR_DESTROY(vm);
902 	VMEM_LOCK_DESTROY(vm);
903 	uma_zfree(vmem_zone, vm);
904 }
905 
906 static int
907 vmem_import(vmem_t *vm, vmem_size_t size, vmem_size_t align, int flags)
908 {
909 	vmem_addr_t addr;
910 	int error;
911 
912 	if (vm->vm_importfn == NULL)
913 		return (EINVAL);
914 
915 	/*
916 	 * To make sure we get a span that meets the alignment we double it
917 	 * and add the size to the tail.  This slightly overestimates.
918 	 */
919 	if (align != vm->vm_quantum_mask + 1)
920 		size = (align * 2) + size;
921 	size = roundup(size, vm->vm_import_quantum);
922 
923 	if (vm->vm_limit != 0 && vm->vm_limit < vm->vm_size + size)
924 		return (ENOMEM);
925 
926 	bt_save(vm);
927 	VMEM_UNLOCK(vm);
928 	error = (vm->vm_importfn)(vm->vm_arg, size, flags, &addr);
929 	VMEM_LOCK(vm);
930 	bt_restore(vm);
931 	if (error)
932 		return (ENOMEM);
933 
934 	vmem_add1(vm, addr, size, BT_TYPE_SPAN);
935 
936 	return 0;
937 }
938 
939 /*
940  * vmem_fit: check if a bt can satisfy the given restrictions.
941  *
942  * it's a caller's responsibility to ensure the region is big enough
943  * before calling us.
944  */
945 static int
946 vmem_fit(const bt_t *bt, vmem_size_t size, vmem_size_t align,
947     vmem_size_t phase, vmem_size_t nocross, vmem_addr_t minaddr,
948     vmem_addr_t maxaddr, vmem_addr_t *addrp)
949 {
950 	vmem_addr_t start;
951 	vmem_addr_t end;
952 
953 	MPASS(size > 0);
954 	MPASS(bt->bt_size >= size); /* caller's responsibility */
955 
956 	/*
957 	 * XXX assumption: vmem_addr_t and vmem_size_t are
958 	 * unsigned integer of the same size.
959 	 */
960 
961 	start = bt->bt_start;
962 	if (start < minaddr) {
963 		start = minaddr;
964 	}
965 	end = BT_END(bt);
966 	if (end > maxaddr)
967 		end = maxaddr;
968 	if (start > end)
969 		return (ENOMEM);
970 
971 	start = VMEM_ALIGNUP(start - phase, align) + phase;
972 	if (start < bt->bt_start)
973 		start += align;
974 	if (VMEM_CROSS_P(start, start + size - 1, nocross)) {
975 		MPASS(align < nocross);
976 		start = VMEM_ALIGNUP(start - phase, nocross) + phase;
977 	}
978 	if (start <= end && end - start >= size - 1) {
979 		MPASS((start & (align - 1)) == phase);
980 		MPASS(!VMEM_CROSS_P(start, start + size - 1, nocross));
981 		MPASS(minaddr <= start);
982 		MPASS(maxaddr == 0 || start + size - 1 <= maxaddr);
983 		MPASS(bt->bt_start <= start);
984 		MPASS(BT_END(bt) - start >= size - 1);
985 		*addrp = start;
986 
987 		return (0);
988 	}
989 	return (ENOMEM);
990 }
991 
992 /*
993  * vmem_clip:  Trim the boundary tag edges to the requested start and size.
994  */
995 static void
996 vmem_clip(vmem_t *vm, bt_t *bt, vmem_addr_t start, vmem_size_t size)
997 {
998 	bt_t *btnew;
999 	bt_t *btprev;
1000 
1001 	VMEM_ASSERT_LOCKED(vm);
1002 	MPASS(bt->bt_type == BT_TYPE_FREE);
1003 	MPASS(bt->bt_size >= size);
1004 	bt_remfree(vm, bt);
1005 	if (bt->bt_start != start) {
1006 		btprev = bt_alloc(vm);
1007 		btprev->bt_type = BT_TYPE_FREE;
1008 		btprev->bt_start = bt->bt_start;
1009 		btprev->bt_size = start - bt->bt_start;
1010 		bt->bt_start = start;
1011 		bt->bt_size -= btprev->bt_size;
1012 		bt_insfree(vm, btprev);
1013 		bt_insseg(vm, btprev,
1014 		    TAILQ_PREV(bt, vmem_seglist, bt_seglist));
1015 	}
1016 	MPASS(bt->bt_start == start);
1017 	if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) {
1018 		/* split */
1019 		btnew = bt_alloc(vm);
1020 		btnew->bt_type = BT_TYPE_BUSY;
1021 		btnew->bt_start = bt->bt_start;
1022 		btnew->bt_size = size;
1023 		bt->bt_start = bt->bt_start + size;
1024 		bt->bt_size -= size;
1025 		bt_insfree(vm, bt);
1026 		bt_insseg(vm, btnew,
1027 		    TAILQ_PREV(bt, vmem_seglist, bt_seglist));
1028 		bt_insbusy(vm, btnew);
1029 		bt = btnew;
1030 	} else {
1031 		bt->bt_type = BT_TYPE_BUSY;
1032 		bt_insbusy(vm, bt);
1033 	}
1034 	MPASS(bt->bt_size >= size);
1035 }
1036 
1037 static int
1038 vmem_try_fetch(vmem_t *vm, const vmem_size_t size, vmem_size_t align, int flags)
1039 {
1040 	vmem_size_t avail;
1041 
1042 	VMEM_ASSERT_LOCKED(vm);
1043 
1044 	/*
1045 	 * XXX it is possible to fail to meet xalloc constraints with the
1046 	 * imported region.  It is up to the user to specify the
1047 	 * import quantum such that it can satisfy any allocation.
1048 	 */
1049 	if (vmem_import(vm, size, align, flags) == 0)
1050 		return (1);
1051 
1052 	/*
1053 	 * Try to free some space from the quantum cache or reclaim
1054 	 * functions if available.
1055 	 */
1056 	if (vm->vm_qcache_max != 0 || vm->vm_reclaimfn != NULL) {
1057 		avail = vm->vm_size - vm->vm_inuse;
1058 		bt_save(vm);
1059 		VMEM_UNLOCK(vm);
1060 		if (vm->vm_qcache_max != 0)
1061 			qc_drain(vm);
1062 		if (vm->vm_reclaimfn != NULL)
1063 			vm->vm_reclaimfn(vm, flags);
1064 		VMEM_LOCK(vm);
1065 		bt_restore(vm);
1066 		/* If we were successful retry even NOWAIT. */
1067 		if (vm->vm_size - vm->vm_inuse > avail)
1068 			return (1);
1069 	}
1070 	if ((flags & M_NOWAIT) != 0)
1071 		return (0);
1072 	bt_save(vm);
1073 	VMEM_CONDVAR_WAIT(vm);
1074 	bt_restore(vm);
1075 	return (1);
1076 }
1077 
1078 static int
1079 vmem_try_release(vmem_t *vm, struct vmem_btag *bt, const bool remfree)
1080 {
1081 	struct vmem_btag *prev;
1082 
1083 	MPASS(bt->bt_type == BT_TYPE_FREE);
1084 
1085 	if (vm->vm_releasefn == NULL)
1086 		return (0);
1087 
1088 	prev = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1089 	MPASS(prev != NULL);
1090 	MPASS(prev->bt_type != BT_TYPE_FREE);
1091 
1092 	if (prev->bt_type == BT_TYPE_SPAN && prev->bt_size == bt->bt_size) {
1093 		vmem_addr_t spanaddr;
1094 		vmem_size_t spansize;
1095 
1096 		MPASS(prev->bt_start == bt->bt_start);
1097 		spanaddr = prev->bt_start;
1098 		spansize = prev->bt_size;
1099 		if (remfree)
1100 			bt_remfree(vm, bt);
1101 		bt_remseg(vm, bt);
1102 		bt_remseg(vm, prev);
1103 		vm->vm_size -= spansize;
1104 		VMEM_CONDVAR_BROADCAST(vm);
1105 		bt_freetrim(vm, BT_MAXFREE);
1106 		vm->vm_releasefn(vm->vm_arg, spanaddr, spansize);
1107 		return (1);
1108 	}
1109 	return (0);
1110 }
1111 
1112 static int
1113 vmem_xalloc_nextfit(vmem_t *vm, const vmem_size_t size, vmem_size_t align,
1114     const vmem_size_t phase, const vmem_size_t nocross, int flags,
1115     vmem_addr_t *addrp)
1116 {
1117 	struct vmem_btag *bt, *cursor, *next, *prev;
1118 	int error;
1119 
1120 	error = ENOMEM;
1121 	VMEM_LOCK(vm);
1122 
1123 	/*
1124 	 * Make sure we have enough tags to complete the operation.
1125 	 */
1126 	if (bt_fill(vm, flags) != 0)
1127 		goto out;
1128 
1129 retry:
1130 	/*
1131 	 * Find the next free tag meeting our constraints.  If one is found,
1132 	 * perform the allocation.
1133 	 */
1134 	for (cursor = &vm->vm_cursor, bt = TAILQ_NEXT(cursor, bt_seglist);
1135 	    bt != cursor; bt = TAILQ_NEXT(bt, bt_seglist)) {
1136 		if (bt == NULL)
1137 			bt = TAILQ_FIRST(&vm->vm_seglist);
1138 		if (bt->bt_type == BT_TYPE_FREE && bt->bt_size >= size &&
1139 		    (error = vmem_fit(bt, size, align, phase, nocross,
1140 		    VMEM_ADDR_MIN, VMEM_ADDR_MAX, addrp)) == 0) {
1141 			vmem_clip(vm, bt, *addrp, size);
1142 			break;
1143 		}
1144 	}
1145 
1146 	/*
1147 	 * Try to coalesce free segments around the cursor.  If we succeed, and
1148 	 * have not yet satisfied the allocation request, try again with the
1149 	 * newly coalesced segment.
1150 	 */
1151 	if ((next = TAILQ_NEXT(cursor, bt_seglist)) != NULL &&
1152 	    (prev = TAILQ_PREV(cursor, vmem_seglist, bt_seglist)) != NULL &&
1153 	    next->bt_type == BT_TYPE_FREE && prev->bt_type == BT_TYPE_FREE &&
1154 	    prev->bt_start + prev->bt_size == next->bt_start) {
1155 		prev->bt_size += next->bt_size;
1156 		bt_remfree(vm, next);
1157 		bt_remseg(vm, next);
1158 
1159 		/*
1160 		 * The coalesced segment might be able to satisfy our request.
1161 		 * If not, we might need to release it from the arena.
1162 		 */
1163 		if (error == ENOMEM && prev->bt_size >= size &&
1164 		    (error = vmem_fit(prev, size, align, phase, nocross,
1165 		    VMEM_ADDR_MIN, VMEM_ADDR_MAX, addrp)) == 0) {
1166 			vmem_clip(vm, prev, *addrp, size);
1167 			bt = prev;
1168 		} else
1169 			(void)vmem_try_release(vm, prev, true);
1170 	}
1171 
1172 	/*
1173 	 * If the allocation was successful, advance the cursor.
1174 	 */
1175 	if (error == 0) {
1176 		TAILQ_REMOVE(&vm->vm_seglist, cursor, bt_seglist);
1177 		for (; bt != NULL && bt->bt_start < *addrp + size;
1178 		    bt = TAILQ_NEXT(bt, bt_seglist))
1179 			;
1180 		if (bt != NULL)
1181 			TAILQ_INSERT_BEFORE(bt, cursor, bt_seglist);
1182 		else
1183 			TAILQ_INSERT_HEAD(&vm->vm_seglist, cursor, bt_seglist);
1184 	}
1185 
1186 	/*
1187 	 * Attempt to bring additional resources into the arena.  If that fails
1188 	 * and M_WAITOK is specified, sleep waiting for resources to be freed.
1189 	 */
1190 	if (error == ENOMEM && vmem_try_fetch(vm, size, align, flags))
1191 		goto retry;
1192 
1193 out:
1194 	VMEM_UNLOCK(vm);
1195 	return (error);
1196 }
1197 
1198 /* ---- vmem API */
1199 
1200 void
1201 vmem_set_import(vmem_t *vm, vmem_import_t *importfn,
1202      vmem_release_t *releasefn, void *arg, vmem_size_t import_quantum)
1203 {
1204 
1205 	VMEM_LOCK(vm);
1206 	KASSERT(vm->vm_size == 0, ("%s: arena is non-empty", __func__));
1207 	vm->vm_importfn = importfn;
1208 	vm->vm_releasefn = releasefn;
1209 	vm->vm_arg = arg;
1210 	vm->vm_import_quantum = import_quantum;
1211 	VMEM_UNLOCK(vm);
1212 }
1213 
1214 void
1215 vmem_set_limit(vmem_t *vm, vmem_size_t limit)
1216 {
1217 
1218 	VMEM_LOCK(vm);
1219 	vm->vm_limit = limit;
1220 	VMEM_UNLOCK(vm);
1221 }
1222 
1223 void
1224 vmem_set_reclaim(vmem_t *vm, vmem_reclaim_t *reclaimfn)
1225 {
1226 
1227 	VMEM_LOCK(vm);
1228 	vm->vm_reclaimfn = reclaimfn;
1229 	VMEM_UNLOCK(vm);
1230 }
1231 
1232 /*
1233  * vmem_init: Initializes vmem arena.
1234  */
1235 vmem_t *
1236 vmem_init(vmem_t *vm, const char *name, vmem_addr_t base, vmem_size_t size,
1237     vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1238 {
1239 	vmem_size_t i;
1240 
1241 	MPASS(quantum > 0);
1242 	MPASS((quantum & (quantum - 1)) == 0);
1243 
1244 	bzero(vm, sizeof(*vm));
1245 
1246 	VMEM_CONDVAR_INIT(vm, name);
1247 	VMEM_LOCK_INIT(vm, name);
1248 	vm->vm_nfreetags = 0;
1249 	LIST_INIT(&vm->vm_freetags);
1250 	strlcpy(vm->vm_name, name, sizeof(vm->vm_name));
1251 	vm->vm_quantum_mask = quantum - 1;
1252 	vm->vm_quantum_shift = flsl(quantum) - 1;
1253 	vm->vm_nbusytag = 0;
1254 	vm->vm_size = 0;
1255 	vm->vm_limit = 0;
1256 	vm->vm_inuse = 0;
1257 	qc_init(vm, qcache_max);
1258 
1259 	TAILQ_INIT(&vm->vm_seglist);
1260 	vm->vm_cursor.bt_start = vm->vm_cursor.bt_size = 0;
1261 	vm->vm_cursor.bt_type = BT_TYPE_CURSOR;
1262 	TAILQ_INSERT_TAIL(&vm->vm_seglist, &vm->vm_cursor, bt_seglist);
1263 
1264 	for (i = 0; i < VMEM_MAXORDER; i++)
1265 		LIST_INIT(&vm->vm_freelist[i]);
1266 
1267 	memset(&vm->vm_hash0, 0, sizeof(vm->vm_hash0));
1268 	vm->vm_hashsize = VMEM_HASHSIZE_MIN;
1269 	vm->vm_hashlist = vm->vm_hash0;
1270 
1271 	if (size != 0) {
1272 		if (vmem_add(vm, base, size, flags) != 0) {
1273 			vmem_destroy1(vm);
1274 			return NULL;
1275 		}
1276 	}
1277 
1278 	mtx_lock(&vmem_list_lock);
1279 	LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist);
1280 	mtx_unlock(&vmem_list_lock);
1281 
1282 	return vm;
1283 }
1284 
1285 /*
1286  * vmem_create: create an arena.
1287  */
1288 vmem_t *
1289 vmem_create(const char *name, vmem_addr_t base, vmem_size_t size,
1290     vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1291 {
1292 
1293 	vmem_t *vm;
1294 
1295 	vm = uma_zalloc(vmem_zone, flags & (M_WAITOK|M_NOWAIT));
1296 	if (vm == NULL)
1297 		return (NULL);
1298 	if (vmem_init(vm, name, base, size, quantum, qcache_max,
1299 	    flags) == NULL)
1300 		return (NULL);
1301 	return (vm);
1302 }
1303 
1304 void
1305 vmem_destroy(vmem_t *vm)
1306 {
1307 
1308 	mtx_lock(&vmem_list_lock);
1309 	LIST_REMOVE(vm, vm_alllist);
1310 	mtx_unlock(&vmem_list_lock);
1311 
1312 	vmem_destroy1(vm);
1313 }
1314 
1315 vmem_size_t
1316 vmem_roundup_size(vmem_t *vm, vmem_size_t size)
1317 {
1318 
1319 	return (size + vm->vm_quantum_mask) & ~vm->vm_quantum_mask;
1320 }
1321 
1322 /*
1323  * vmem_alloc: allocate resource from the arena.
1324  */
1325 int
1326 vmem_alloc(vmem_t *vm, vmem_size_t size, int flags, vmem_addr_t *addrp)
1327 {
1328 	const int strat __unused = flags & VMEM_FITMASK;
1329 	qcache_t *qc;
1330 
1331 	flags &= VMEM_FLAGS;
1332 	MPASS(size > 0);
1333 	MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT || strat == M_NEXTFIT);
1334 	if ((flags & M_NOWAIT) == 0)
1335 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_alloc");
1336 
1337 	if (size <= vm->vm_qcache_max) {
1338 		/*
1339 		 * Resource 0 cannot be cached, so avoid a blocking allocation
1340 		 * in qc_import() and give the vmem_xalloc() call below a chance
1341 		 * to return 0.
1342 		 */
1343 		qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1344 		*addrp = (vmem_addr_t)uma_zalloc(qc->qc_cache,
1345 		    (flags & ~M_WAITOK) | M_NOWAIT);
1346 		if (__predict_true(*addrp != 0))
1347 			return (0);
1348 	}
1349 
1350 	return (vmem_xalloc(vm, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
1351 	    flags, addrp));
1352 }
1353 
1354 int
1355 vmem_xalloc(vmem_t *vm, const vmem_size_t size0, vmem_size_t align,
1356     const vmem_size_t phase, const vmem_size_t nocross,
1357     const vmem_addr_t minaddr, const vmem_addr_t maxaddr, int flags,
1358     vmem_addr_t *addrp)
1359 {
1360 	const vmem_size_t size = vmem_roundup_size(vm, size0);
1361 	struct vmem_freelist *list;
1362 	struct vmem_freelist *first;
1363 	struct vmem_freelist *end;
1364 	bt_t *bt;
1365 	int error;
1366 	int strat;
1367 
1368 	flags &= VMEM_FLAGS;
1369 	strat = flags & VMEM_FITMASK;
1370 	MPASS(size0 > 0);
1371 	MPASS(size > 0);
1372 	MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT || strat == M_NEXTFIT);
1373 	MPASS((flags & (M_NOWAIT|M_WAITOK)) != (M_NOWAIT|M_WAITOK));
1374 	if ((flags & M_NOWAIT) == 0)
1375 		WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_xalloc");
1376 	MPASS((align & vm->vm_quantum_mask) == 0);
1377 	MPASS((align & (align - 1)) == 0);
1378 	MPASS((phase & vm->vm_quantum_mask) == 0);
1379 	MPASS((nocross & vm->vm_quantum_mask) == 0);
1380 	MPASS((nocross & (nocross - 1)) == 0);
1381 	MPASS((align == 0 && phase == 0) || phase < align);
1382 	MPASS(nocross == 0 || nocross >= size);
1383 	MPASS(minaddr <= maxaddr);
1384 	MPASS(!VMEM_CROSS_P(phase, phase + size - 1, nocross));
1385 	if (strat == M_NEXTFIT)
1386 		MPASS(minaddr == VMEM_ADDR_MIN && maxaddr == VMEM_ADDR_MAX);
1387 
1388 	if (align == 0)
1389 		align = vm->vm_quantum_mask + 1;
1390 	*addrp = 0;
1391 
1392 	/*
1393 	 * Next-fit allocations don't use the freelists.
1394 	 */
1395 	if (strat == M_NEXTFIT)
1396 		return (vmem_xalloc_nextfit(vm, size0, align, phase, nocross,
1397 		    flags, addrp));
1398 
1399 	end = &vm->vm_freelist[VMEM_MAXORDER];
1400 	/*
1401 	 * choose a free block from which we allocate.
1402 	 */
1403 	first = bt_freehead_toalloc(vm, size, strat);
1404 	VMEM_LOCK(vm);
1405 
1406 	/*
1407 	 * Make sure we have enough tags to complete the operation.
1408 	 */
1409 	error = bt_fill(vm, flags);
1410 	if (error != 0)
1411 		goto out;
1412 	for (;;) {
1413 		/*
1414 	 	 * Scan freelists looking for a tag that satisfies the
1415 		 * allocation.  If we're doing BESTFIT we may encounter
1416 		 * sizes below the request.  If we're doing FIRSTFIT we
1417 		 * inspect only the first element from each list.
1418 		 */
1419 		for (list = first; list < end; list++) {
1420 			LIST_FOREACH(bt, list, bt_freelist) {
1421 				if (bt->bt_size >= size) {
1422 					error = vmem_fit(bt, size, align, phase,
1423 					    nocross, minaddr, maxaddr, addrp);
1424 					if (error == 0) {
1425 						vmem_clip(vm, bt, *addrp, size);
1426 						goto out;
1427 					}
1428 				}
1429 				/* FIRST skips to the next list. */
1430 				if (strat == M_FIRSTFIT)
1431 					break;
1432 			}
1433 		}
1434 
1435 		/*
1436 		 * Retry if the fast algorithm failed.
1437 		 */
1438 		if (strat == M_FIRSTFIT) {
1439 			strat = M_BESTFIT;
1440 			first = bt_freehead_toalloc(vm, size, strat);
1441 			continue;
1442 		}
1443 
1444 		/*
1445 		 * Try a few measures to bring additional resources into the
1446 		 * arena.  If all else fails, we will sleep waiting for
1447 		 * resources to be freed.
1448 		 */
1449 		if (!vmem_try_fetch(vm, size, align, flags)) {
1450 			error = ENOMEM;
1451 			break;
1452 		}
1453 	}
1454 out:
1455 	VMEM_UNLOCK(vm);
1456 	if (error != 0 && (flags & M_NOWAIT) == 0)
1457 		panic("failed to allocate waiting allocation\n");
1458 
1459 	return (error);
1460 }
1461 
1462 /*
1463  * vmem_free: free the resource to the arena.
1464  */
1465 void
1466 vmem_free(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1467 {
1468 	qcache_t *qc;
1469 	MPASS(size > 0);
1470 
1471 	if (size <= vm->vm_qcache_max &&
1472 	    __predict_true(addr >= VMEM_ADDR_QCACHE_MIN)) {
1473 		qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1474 		uma_zfree(qc->qc_cache, (void *)addr);
1475 	} else
1476 		vmem_xfree(vm, addr, size);
1477 }
1478 
1479 void
1480 vmem_xfree(vmem_t *vm, vmem_addr_t addr, vmem_size_t size __unused)
1481 {
1482 	bt_t *bt;
1483 	bt_t *t;
1484 
1485 	MPASS(size > 0);
1486 
1487 	VMEM_LOCK(vm);
1488 	bt = bt_lookupbusy(vm, addr);
1489 	MPASS(bt != NULL);
1490 	MPASS(bt->bt_start == addr);
1491 	MPASS(bt->bt_size == vmem_roundup_size(vm, size) ||
1492 	    bt->bt_size - vmem_roundup_size(vm, size) <= vm->vm_quantum_mask);
1493 	MPASS(bt->bt_type == BT_TYPE_BUSY);
1494 	bt_rembusy(vm, bt);
1495 	bt->bt_type = BT_TYPE_FREE;
1496 
1497 	/* coalesce */
1498 	t = TAILQ_NEXT(bt, bt_seglist);
1499 	if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1500 		MPASS(BT_END(bt) < t->bt_start);	/* YYY */
1501 		bt->bt_size += t->bt_size;
1502 		bt_remfree(vm, t);
1503 		bt_remseg(vm, t);
1504 	}
1505 	t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1506 	if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1507 		MPASS(BT_END(t) < bt->bt_start);	/* YYY */
1508 		bt->bt_size += t->bt_size;
1509 		bt->bt_start = t->bt_start;
1510 		bt_remfree(vm, t);
1511 		bt_remseg(vm, t);
1512 	}
1513 
1514 	if (!vmem_try_release(vm, bt, false)) {
1515 		bt_insfree(vm, bt);
1516 		VMEM_CONDVAR_BROADCAST(vm);
1517 		bt_freetrim(vm, BT_MAXFREE);
1518 	}
1519 }
1520 
1521 /*
1522  * vmem_add:
1523  *
1524  */
1525 int
1526 vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int flags)
1527 {
1528 	int error;
1529 
1530 	flags &= VMEM_FLAGS;
1531 
1532 	VMEM_LOCK(vm);
1533 	error = bt_fill(vm, flags);
1534 	if (error == 0)
1535 		vmem_add1(vm, addr, size, BT_TYPE_SPAN_STATIC);
1536 	VMEM_UNLOCK(vm);
1537 
1538 	return (error);
1539 }
1540 
1541 /*
1542  * vmem_size: information about arenas size
1543  */
1544 vmem_size_t
1545 vmem_size(vmem_t *vm, int typemask)
1546 {
1547 	int i;
1548 
1549 	switch (typemask) {
1550 	case VMEM_ALLOC:
1551 		return vm->vm_inuse;
1552 	case VMEM_FREE:
1553 		return vm->vm_size - vm->vm_inuse;
1554 	case VMEM_FREE|VMEM_ALLOC:
1555 		return vm->vm_size;
1556 	case VMEM_MAXFREE:
1557 		VMEM_LOCK(vm);
1558 		for (i = VMEM_MAXORDER - 1; i >= 0; i--) {
1559 			if (LIST_EMPTY(&vm->vm_freelist[i]))
1560 				continue;
1561 			VMEM_UNLOCK(vm);
1562 			return ((vmem_size_t)ORDER2SIZE(i) <<
1563 			    vm->vm_quantum_shift);
1564 		}
1565 		VMEM_UNLOCK(vm);
1566 		return (0);
1567 	default:
1568 		panic("vmem_size");
1569 	}
1570 }
1571 
1572 /* ---- debug */
1573 
1574 #if defined(DDB) || defined(DIAGNOSTIC)
1575 
1576 static void bt_dump(const bt_t *, int (*)(const char *, ...)
1577     __printflike(1, 2));
1578 
1579 static const char *
1580 bt_type_string(int type)
1581 {
1582 
1583 	switch (type) {
1584 	case BT_TYPE_BUSY:
1585 		return "busy";
1586 	case BT_TYPE_FREE:
1587 		return "free";
1588 	case BT_TYPE_SPAN:
1589 		return "span";
1590 	case BT_TYPE_SPAN_STATIC:
1591 		return "static span";
1592 	case BT_TYPE_CURSOR:
1593 		return "cursor";
1594 	default:
1595 		break;
1596 	}
1597 	return "BOGUS";
1598 }
1599 
1600 static void
1601 bt_dump(const bt_t *bt, int (*pr)(const char *, ...))
1602 {
1603 
1604 	(*pr)("\t%p: %jx %jx, %d(%s)\n",
1605 	    bt, (intmax_t)bt->bt_start, (intmax_t)bt->bt_size,
1606 	    bt->bt_type, bt_type_string(bt->bt_type));
1607 }
1608 
1609 static void
1610 vmem_dump(const vmem_t *vm , int (*pr)(const char *, ...) __printflike(1, 2))
1611 {
1612 	const bt_t *bt;
1613 	int i;
1614 
1615 	(*pr)("vmem %p '%s'\n", vm, vm->vm_name);
1616 	TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1617 		bt_dump(bt, pr);
1618 	}
1619 
1620 	for (i = 0; i < VMEM_MAXORDER; i++) {
1621 		const struct vmem_freelist *fl = &vm->vm_freelist[i];
1622 
1623 		if (LIST_EMPTY(fl)) {
1624 			continue;
1625 		}
1626 
1627 		(*pr)("freelist[%d]\n", i);
1628 		LIST_FOREACH(bt, fl, bt_freelist) {
1629 			bt_dump(bt, pr);
1630 		}
1631 	}
1632 }
1633 
1634 #endif /* defined(DDB) || defined(DIAGNOSTIC) */
1635 
1636 #if defined(DDB)
1637 #include <ddb/ddb.h>
1638 
1639 static bt_t *
1640 vmem_whatis_lookup(vmem_t *vm, vmem_addr_t addr)
1641 {
1642 	bt_t *bt;
1643 
1644 	TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1645 		if (BT_ISSPAN_P(bt)) {
1646 			continue;
1647 		}
1648 		if (bt->bt_start <= addr && addr <= BT_END(bt)) {
1649 			return bt;
1650 		}
1651 	}
1652 
1653 	return NULL;
1654 }
1655 
1656 void
1657 vmem_whatis(vmem_addr_t addr, int (*pr)(const char *, ...))
1658 {
1659 	vmem_t *vm;
1660 
1661 	LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1662 		bt_t *bt;
1663 
1664 		bt = vmem_whatis_lookup(vm, addr);
1665 		if (bt == NULL) {
1666 			continue;
1667 		}
1668 		(*pr)("%p is %p+%zu in VMEM '%s' (%s)\n",
1669 		    (void *)addr, (void *)bt->bt_start,
1670 		    (vmem_size_t)(addr - bt->bt_start), vm->vm_name,
1671 		    (bt->bt_type == BT_TYPE_BUSY) ? "allocated" : "free");
1672 	}
1673 }
1674 
1675 void
1676 vmem_printall(const char *modif, int (*pr)(const char *, ...))
1677 {
1678 	const vmem_t *vm;
1679 
1680 	LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1681 		vmem_dump(vm, pr);
1682 	}
1683 }
1684 
1685 void
1686 vmem_print(vmem_addr_t addr, const char *modif, int (*pr)(const char *, ...))
1687 {
1688 	const vmem_t *vm = (const void *)addr;
1689 
1690 	vmem_dump(vm, pr);
1691 }
1692 
1693 DB_SHOW_COMMAND(vmemdump, vmemdump)
1694 {
1695 
1696 	if (!have_addr) {
1697 		db_printf("usage: show vmemdump <addr>\n");
1698 		return;
1699 	}
1700 
1701 	vmem_dump((const vmem_t *)addr, db_printf);
1702 }
1703 
1704 DB_SHOW_ALL_COMMAND(vmemdump, vmemdumpall)
1705 {
1706 	const vmem_t *vm;
1707 
1708 	LIST_FOREACH(vm, &vmem_list, vm_alllist)
1709 		vmem_dump(vm, db_printf);
1710 }
1711 
1712 DB_SHOW_COMMAND(vmem, vmem_summ)
1713 {
1714 	const vmem_t *vm = (const void *)addr;
1715 	const bt_t *bt;
1716 	size_t ft[VMEM_MAXORDER], ut[VMEM_MAXORDER];
1717 	size_t fs[VMEM_MAXORDER], us[VMEM_MAXORDER];
1718 	int ord;
1719 
1720 	if (!have_addr) {
1721 		db_printf("usage: show vmem <addr>\n");
1722 		return;
1723 	}
1724 
1725 	db_printf("vmem %p '%s'\n", vm, vm->vm_name);
1726 	db_printf("\tquantum:\t%zu\n", vm->vm_quantum_mask + 1);
1727 	db_printf("\tsize:\t%zu\n", vm->vm_size);
1728 	db_printf("\tinuse:\t%zu\n", vm->vm_inuse);
1729 	db_printf("\tfree:\t%zu\n", vm->vm_size - vm->vm_inuse);
1730 	db_printf("\tbusy tags:\t%d\n", vm->vm_nbusytag);
1731 	db_printf("\tfree tags:\t%d\n", vm->vm_nfreetags);
1732 
1733 	memset(&ft, 0, sizeof(ft));
1734 	memset(&ut, 0, sizeof(ut));
1735 	memset(&fs, 0, sizeof(fs));
1736 	memset(&us, 0, sizeof(us));
1737 	TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1738 		ord = SIZE2ORDER(bt->bt_size >> vm->vm_quantum_shift);
1739 		if (bt->bt_type == BT_TYPE_BUSY) {
1740 			ut[ord]++;
1741 			us[ord] += bt->bt_size;
1742 		} else if (bt->bt_type == BT_TYPE_FREE) {
1743 			ft[ord]++;
1744 			fs[ord] += bt->bt_size;
1745 		}
1746 	}
1747 	db_printf("\t\t\tinuse\tsize\t\tfree\tsize\n");
1748 	for (ord = 0; ord < VMEM_MAXORDER; ord++) {
1749 		if (ut[ord] == 0 && ft[ord] == 0)
1750 			continue;
1751 		db_printf("\t%-15zu %zu\t%-15zu %zu\t%-16zu\n",
1752 		    ORDER2SIZE(ord) << vm->vm_quantum_shift,
1753 		    ut[ord], us[ord], ft[ord], fs[ord]);
1754 	}
1755 }
1756 
1757 DB_SHOW_ALL_COMMAND(vmem, vmem_summall)
1758 {
1759 	const vmem_t *vm;
1760 
1761 	LIST_FOREACH(vm, &vmem_list, vm_alllist)
1762 		vmem_summ((db_expr_t)vm, TRUE, count, modif);
1763 }
1764 #endif /* defined(DDB) */
1765 
1766 #define vmem_printf printf
1767 
1768 #if defined(DIAGNOSTIC)
1769 
1770 static bool
1771 vmem_check_sanity(vmem_t *vm)
1772 {
1773 	const bt_t *bt, *bt2;
1774 
1775 	MPASS(vm != NULL);
1776 
1777 	TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1778 		if (bt->bt_start > BT_END(bt)) {
1779 			printf("corrupted tag\n");
1780 			bt_dump(bt, vmem_printf);
1781 			return false;
1782 		}
1783 	}
1784 	TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1785 		if (bt->bt_type == BT_TYPE_CURSOR) {
1786 			if (bt->bt_start != 0 || bt->bt_size != 0) {
1787 				printf("corrupted cursor\n");
1788 				return false;
1789 			}
1790 			continue;
1791 		}
1792 		TAILQ_FOREACH(bt2, &vm->vm_seglist, bt_seglist) {
1793 			if (bt == bt2) {
1794 				continue;
1795 			}
1796 			if (bt2->bt_type == BT_TYPE_CURSOR) {
1797 				continue;
1798 			}
1799 			if (BT_ISSPAN_P(bt) != BT_ISSPAN_P(bt2)) {
1800 				continue;
1801 			}
1802 			if (bt->bt_start <= BT_END(bt2) &&
1803 			    bt2->bt_start <= BT_END(bt)) {
1804 				printf("overwrapped tags\n");
1805 				bt_dump(bt, vmem_printf);
1806 				bt_dump(bt2, vmem_printf);
1807 				return false;
1808 			}
1809 		}
1810 	}
1811 
1812 	return true;
1813 }
1814 
1815 static void
1816 vmem_check(vmem_t *vm)
1817 {
1818 
1819 	if (!vmem_check_sanity(vm)) {
1820 		panic("insanity vmem %p", vm);
1821 	}
1822 }
1823 
1824 #endif /* defined(DIAGNOSTIC) */
1825