2 * Copyright (c)2006,2007,2008,2009 YAMAMOTO Takashi,
3 * Copyright (c) 2013 EMC Corp.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * $NetBSD: vmem_impl.h,v 1.2 2013/01/29 21:26:24 para Exp $
31 * $NetBSD: subr_vmem.c,v 1.83 2013/03/06 11:20:10 yamt Exp $
36 * - Magazines and Vmem: Extending the Slab Allocator
37 * to Many CPUs and Arbitrary Resources
38 * http://www.usenix.org/event/usenix01/bonwick.html
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
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>
53 #include <sys/malloc.h>
54 #include <sys/mutex.h>
56 #include <sys/condvar.h>
57 #include <sys/sysctl.h>
58 #include <sys/taskqueue.h>
66 #include <vm/vm_map.h>
67 #include <vm/vm_object.h>
68 #include <vm/vm_kern.h>
69 #include <vm/vm_extern.h>
70 #include <vm/vm_param.h>
71 #include <vm/vm_pageout.h>
73 #define VMEM_OPTORDER 5
74 #define VMEM_OPTVALUE (1 << VMEM_OPTORDER)
75 #define VMEM_MAXORDER \
76 (VMEM_OPTVALUE - 1 + sizeof(vmem_size_t) * NBBY - VMEM_OPTORDER)
78 #define VMEM_HASHSIZE_MIN 16
79 #define VMEM_HASHSIZE_MAX 131072
81 #define VMEM_QCACHE_IDX_MAX 16
83 #define VMEM_FITMASK (M_BESTFIT | M_FIRSTFIT)
86 (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM | M_BESTFIT | M_FIRSTFIT)
88 #define BT_FLAGS (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM)
90 #define QC_NAME_MAX 16
93 * Data structures private to vmem.
95 MALLOC_DEFINE(M_VMEM, "vmem", "vmem internal structures");
97 typedef struct vmem_btag bt_t;
99 TAILQ_HEAD(vmem_seglist, vmem_btag);
100 LIST_HEAD(vmem_freelist, vmem_btag);
101 LIST_HEAD(vmem_hashlist, vmem_btag);
107 char qc_name[QC_NAME_MAX];
109 typedef struct qcache qcache_t;
110 #define QC_POOL_TO_QCACHE(pool) ((qcache_t *)(pool->pr_qcache))
112 #define VMEM_NAME_MAX 16
116 struct mtx_padalign vm_lock;
118 char vm_name[VMEM_NAME_MAX+1];
119 LIST_ENTRY(vmem) vm_alllist;
120 struct vmem_hashlist vm_hash0[VMEM_HASHSIZE_MIN];
121 struct vmem_freelist vm_freelist[VMEM_MAXORDER];
122 struct vmem_seglist vm_seglist;
123 struct vmem_hashlist *vm_hashlist;
124 vmem_size_t vm_hashsize;
126 /* Constant after init */
127 vmem_size_t vm_qcache_max;
128 vmem_size_t vm_quantum_mask;
129 vmem_size_t vm_import_quantum;
130 int vm_quantum_shift;
132 /* Written on alloc/free */
133 LIST_HEAD(, vmem_btag) vm_freetags;
136 vmem_size_t vm_inuse;
139 /* Used on import. */
140 vmem_import_t *vm_importfn;
141 vmem_release_t *vm_releasefn;
144 /* Space exhaustion callback. */
145 vmem_reclaim_t *vm_reclaimfn;
148 qcache_t vm_qcache[VMEM_QCACHE_IDX_MAX];
153 TAILQ_ENTRY(vmem_btag) bt_seglist;
155 LIST_ENTRY(vmem_btag) u_freelist; /* BT_TYPE_FREE */
156 LIST_ENTRY(vmem_btag) u_hashlist; /* BT_TYPE_BUSY */
158 #define bt_hashlist bt_u.u_hashlist
159 #define bt_freelist bt_u.u_freelist
160 vmem_addr_t bt_start;
165 #define BT_TYPE_SPAN 1 /* Allocated from importfn */
166 #define BT_TYPE_SPAN_STATIC 2 /* vmem_add() or create. */
167 #define BT_TYPE_FREE 3 /* Available space. */
168 #define BT_TYPE_BUSY 4 /* Used space. */
169 #define BT_ISSPAN_P(bt) ((bt)->bt_type <= BT_TYPE_SPAN_STATIC)
171 #define BT_END(bt) ((bt)->bt_start + (bt)->bt_size - 1)
173 #if defined(DIAGNOSTIC)
174 static int enable_vmem_check = 1;
175 SYSCTL_INT(_debug, OID_AUTO, vmem_check, CTLFLAG_RWTUN,
176 &enable_vmem_check, 0, "Enable vmem check");
177 static void vmem_check(vmem_t *);
180 static struct callout vmem_periodic_ch;
181 static int vmem_periodic_interval;
182 static struct task vmem_periodic_wk;
184 static struct mtx_padalign vmem_list_lock;
185 static LIST_HEAD(, vmem) vmem_list = LIST_HEAD_INITIALIZER(vmem_list);
188 #define VMEM_CONDVAR_INIT(vm, wchan) cv_init(&vm->vm_cv, wchan)
189 #define VMEM_CONDVAR_DESTROY(vm) cv_destroy(&vm->vm_cv)
190 #define VMEM_CONDVAR_WAIT(vm) cv_wait(&vm->vm_cv, &vm->vm_lock)
191 #define VMEM_CONDVAR_BROADCAST(vm) cv_broadcast(&vm->vm_cv)
194 #define VMEM_LOCK(vm) mtx_lock(&vm->vm_lock)
195 #define VMEM_TRYLOCK(vm) mtx_trylock(&vm->vm_lock)
196 #define VMEM_UNLOCK(vm) mtx_unlock(&vm->vm_lock)
197 #define VMEM_LOCK_INIT(vm, name) mtx_init(&vm->vm_lock, (name), NULL, MTX_DEF)
198 #define VMEM_LOCK_DESTROY(vm) mtx_destroy(&vm->vm_lock)
199 #define VMEM_ASSERT_LOCKED(vm) mtx_assert(&vm->vm_lock, MA_OWNED);
201 #define VMEM_ALIGNUP(addr, align) (-(-(addr) & -(align)))
203 #define VMEM_CROSS_P(addr1, addr2, boundary) \
204 ((((addr1) ^ (addr2)) & -(boundary)) != 0)
206 #define ORDER2SIZE(order) ((order) < VMEM_OPTVALUE ? ((order) + 1) : \
207 (vmem_size_t)1 << ((order) - (VMEM_OPTVALUE - VMEM_OPTORDER - 1)))
208 #define SIZE2ORDER(size) ((size) <= VMEM_OPTVALUE ? ((size) - 1) : \
209 (flsl(size) + (VMEM_OPTVALUE - VMEM_OPTORDER - 2)))
212 * Maximum number of boundary tags that may be required to satisfy an
213 * allocation. Two may be required to import. Another two may be
214 * required to clip edges.
216 #define BT_MAXALLOC 4
219 * Max free limits the number of locally cached boundary tags. We
220 * just want to avoid hitting the zone allocator for every call.
222 #define BT_MAXFREE (BT_MAXALLOC * 8)
224 /* Allocator for boundary tags. */
225 static uma_zone_t vmem_bt_zone;
227 /* boot time arena storage. */
228 static struct vmem kernel_arena_storage;
229 static struct vmem kmem_arena_storage;
230 static struct vmem buffer_arena_storage;
231 static struct vmem transient_arena_storage;
232 vmem_t *kernel_arena = &kernel_arena_storage;
233 vmem_t *kmem_arena = &kmem_arena_storage;
234 vmem_t *buffer_arena = &buffer_arena_storage;
235 vmem_t *transient_arena = &transient_arena_storage;
237 #ifdef DEBUG_MEMGUARD
238 static struct vmem memguard_arena_storage;
239 vmem_t *memguard_arena = &memguard_arena_storage;
243 * Fill the vmem's boundary tag cache. We guarantee that boundary tag
244 * allocation will not fail once bt_fill() passes. To do so we cache
245 * at least the maximum possible tag allocations in the arena.
248 bt_fill(vmem_t *vm, int flags)
252 VMEM_ASSERT_LOCKED(vm);
255 * Only allow the kmem arena to dip into reserve tags. It is the
256 * vmem where new tags come from.
259 if (vm != kmem_arena)
260 flags &= ~M_USE_RESERVE;
263 * Loop until we meet the reserve. To minimize the lock shuffle
264 * and prevent simultaneous fills we first try a NOWAIT regardless
265 * of the caller's flags. Specify M_NOVM so we don't recurse while
266 * holding a vmem lock.
268 while (vm->vm_nfreetags < BT_MAXALLOC) {
269 bt = uma_zalloc(vmem_bt_zone,
270 (flags & M_USE_RESERVE) | M_NOWAIT | M_NOVM);
273 bt = uma_zalloc(vmem_bt_zone, flags);
275 if (bt == NULL && (flags & M_NOWAIT) != 0)
278 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
282 if (vm->vm_nfreetags < BT_MAXALLOC)
289 * Pop a tag off of the freetag stack.
296 VMEM_ASSERT_LOCKED(vm);
297 bt = LIST_FIRST(&vm->vm_freetags);
299 LIST_REMOVE(bt, bt_freelist);
306 * Trim the per-vmem free list. Returns with the lock released to
307 * avoid allocator recursions.
310 bt_freetrim(vmem_t *vm, int freelimit)
312 LIST_HEAD(, vmem_btag) freetags;
315 LIST_INIT(&freetags);
316 VMEM_ASSERT_LOCKED(vm);
317 while (vm->vm_nfreetags > freelimit) {
318 bt = LIST_FIRST(&vm->vm_freetags);
319 LIST_REMOVE(bt, bt_freelist);
321 LIST_INSERT_HEAD(&freetags, bt, bt_freelist);
324 while ((bt = LIST_FIRST(&freetags)) != NULL) {
325 LIST_REMOVE(bt, bt_freelist);
326 uma_zfree(vmem_bt_zone, bt);
331 bt_free(vmem_t *vm, bt_t *bt)
334 VMEM_ASSERT_LOCKED(vm);
335 MPASS(LIST_FIRST(&vm->vm_freetags) != bt);
336 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
341 * freelist[0] ... [1, 1]
342 * freelist[1] ... [2, 2]
344 * freelist[29] ... [30, 30]
345 * freelist[30] ... [31, 31]
346 * freelist[31] ... [32, 63]
347 * freelist[33] ... [64, 127]
349 * freelist[n] ... [(1 << (n - 26)), (1 << (n - 25)) - 1]
353 static struct vmem_freelist *
354 bt_freehead_tofree(vmem_t *vm, vmem_size_t size)
356 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
357 const int idx = SIZE2ORDER(qsize);
359 MPASS(size != 0 && qsize != 0);
360 MPASS((size & vm->vm_quantum_mask) == 0);
362 MPASS(idx < VMEM_MAXORDER);
364 return &vm->vm_freelist[idx];
368 * bt_freehead_toalloc: return the freelist for the given size and allocation
371 * For M_FIRSTFIT, return the list in which any blocks are large enough
372 * for the requested size. otherwise, return the list which can have blocks
373 * large enough for the requested size.
375 static struct vmem_freelist *
376 bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, int strat)
378 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
379 int idx = SIZE2ORDER(qsize);
381 MPASS(size != 0 && qsize != 0);
382 MPASS((size & vm->vm_quantum_mask) == 0);
384 if (strat == M_FIRSTFIT && ORDER2SIZE(idx) != qsize) {
386 /* check too large request? */
389 MPASS(idx < VMEM_MAXORDER);
391 return &vm->vm_freelist[idx];
394 /* ---- boundary tag hash */
396 static struct vmem_hashlist *
397 bt_hashhead(vmem_t *vm, vmem_addr_t addr)
399 struct vmem_hashlist *list;
402 hash = hash32_buf(&addr, sizeof(addr), 0);
403 list = &vm->vm_hashlist[hash % vm->vm_hashsize];
409 bt_lookupbusy(vmem_t *vm, vmem_addr_t addr)
411 struct vmem_hashlist *list;
414 VMEM_ASSERT_LOCKED(vm);
415 list = bt_hashhead(vm, addr);
416 LIST_FOREACH(bt, list, bt_hashlist) {
417 if (bt->bt_start == addr) {
426 bt_rembusy(vmem_t *vm, bt_t *bt)
429 VMEM_ASSERT_LOCKED(vm);
430 MPASS(vm->vm_nbusytag > 0);
431 vm->vm_inuse -= bt->bt_size;
433 LIST_REMOVE(bt, bt_hashlist);
437 bt_insbusy(vmem_t *vm, bt_t *bt)
439 struct vmem_hashlist *list;
441 VMEM_ASSERT_LOCKED(vm);
442 MPASS(bt->bt_type == BT_TYPE_BUSY);
444 list = bt_hashhead(vm, bt->bt_start);
445 LIST_INSERT_HEAD(list, bt, bt_hashlist);
447 vm->vm_inuse += bt->bt_size;
450 /* ---- boundary tag list */
453 bt_remseg(vmem_t *vm, bt_t *bt)
456 TAILQ_REMOVE(&vm->vm_seglist, bt, bt_seglist);
461 bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev)
464 TAILQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist);
468 bt_insseg_tail(vmem_t *vm, bt_t *bt)
471 TAILQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist);
475 bt_remfree(vmem_t *vm, bt_t *bt)
478 MPASS(bt->bt_type == BT_TYPE_FREE);
480 LIST_REMOVE(bt, bt_freelist);
484 bt_insfree(vmem_t *vm, bt_t *bt)
486 struct vmem_freelist *list;
488 list = bt_freehead_tofree(vm, bt->bt_size);
489 LIST_INSERT_HEAD(list, bt, bt_freelist);
492 /* ---- vmem internal functions */
495 * Import from the arena into the quantum cache in UMA.
498 qc_import(void *arg, void **store, int cnt, int flags)
505 if ((flags & VMEM_FITMASK) == 0)
507 for (i = 0; i < cnt; i++) {
508 if (vmem_xalloc(qc->qc_vmem, qc->qc_size, 0, 0, 0,
509 VMEM_ADDR_MIN, VMEM_ADDR_MAX, flags, &addr) != 0)
511 store[i] = (void *)addr;
512 /* Only guarantee one allocation. */
520 * Release memory from the UMA cache to the arena.
523 qc_release(void *arg, void **store, int cnt)
529 for (i = 0; i < cnt; i++)
530 vmem_xfree(qc->qc_vmem, (vmem_addr_t)store[i], qc->qc_size);
534 qc_init(vmem_t *vm, vmem_size_t qcache_max)
541 MPASS((qcache_max & vm->vm_quantum_mask) == 0);
542 qcache_idx_max = MIN(qcache_max >> vm->vm_quantum_shift,
543 VMEM_QCACHE_IDX_MAX);
544 vm->vm_qcache_max = qcache_idx_max << vm->vm_quantum_shift;
545 for (i = 0; i < qcache_idx_max; i++) {
546 qc = &vm->vm_qcache[i];
547 size = (i + 1) << vm->vm_quantum_shift;
548 snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu",
552 qc->qc_cache = uma_zcache_create(qc->qc_name, size,
553 NULL, NULL, NULL, NULL, qc_import, qc_release, qc,
560 qc_destroy(vmem_t *vm)
565 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
566 for (i = 0; i < qcache_idx_max; i++)
567 uma_zdestroy(vm->vm_qcache[i].qc_cache);
576 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
577 for (i = 0; i < qcache_idx_max; i++)
578 zone_drain(vm->vm_qcache[i].qc_cache);
581 #ifndef UMA_MD_SMALL_ALLOC
583 static struct mtx_padalign vmem_bt_lock;
586 * vmem_bt_alloc: Allocate a new page of boundary tags.
588 * On architectures with uma_small_alloc there is no recursion; no address
589 * space need be allocated to allocate boundary tags. For the others, we
590 * must handle recursion. Boundary tags are necessary to allocate new
593 * UMA guarantees that enough tags are held in reserve to allocate a new
594 * page of kva. We dip into this reserve by specifying M_USE_RESERVE only
595 * when allocating the page to hold new boundary tags. In this way the
596 * reserve is automatically filled by the allocation that uses the reserve.
598 * We still have to guarantee that the new tags are allocated atomically since
599 * many threads may try concurrently. The bt_lock provides this guarantee.
600 * We convert WAITOK allocations to NOWAIT and then handle the blocking here
601 * on failure. It's ok to return NULL for a WAITOK allocation as UMA will
602 * loop again after checking to see if we lost the race to allocate.
604 * There is a small race between vmem_bt_alloc() returning the page and the
605 * zone lock being acquired to add the page to the zone. For WAITOK
606 * allocations we just pause briefly. NOWAIT may experience a transient
607 * failure. To alleviate this we permit a small number of simultaneous
608 * fills to proceed concurrently so NOWAIT is less likely to fail unless
609 * we are really out of KVA.
612 vmem_bt_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
616 *pflag = UMA_SLAB_KMEM;
619 * Single thread boundary tag allocation so that the address space
620 * and memory are added in one atomic operation.
622 mtx_lock(&vmem_bt_lock);
623 if (vmem_xalloc(kmem_arena, bytes, 0, 0, 0, VMEM_ADDR_MIN,
624 VMEM_ADDR_MAX, M_NOWAIT | M_NOVM | M_USE_RESERVE | M_BESTFIT,
626 if (kmem_back(kmem_object, addr, bytes,
627 M_NOWAIT | M_USE_RESERVE) == 0) {
628 mtx_unlock(&vmem_bt_lock);
629 return ((void *)addr);
631 vmem_xfree(kmem_arena, addr, bytes);
632 mtx_unlock(&vmem_bt_lock);
634 * Out of memory, not address space. This may not even be
635 * possible due to M_USE_RESERVE page allocation.
641 mtx_unlock(&vmem_bt_lock);
643 * We're either out of address space or lost a fill race.
656 mtx_init(&vmem_list_lock, "vmem list lock", NULL, MTX_DEF);
657 vmem_bt_zone = uma_zcreate("vmem btag",
658 sizeof(struct vmem_btag), NULL, NULL, NULL, NULL,
659 UMA_ALIGN_PTR, UMA_ZONE_VM);
660 #ifndef UMA_MD_SMALL_ALLOC
661 mtx_init(&vmem_bt_lock, "btag lock", NULL, MTX_DEF);
662 uma_prealloc(vmem_bt_zone, BT_MAXALLOC);
664 * Reserve enough tags to allocate new tags. We allow multiple
665 * CPUs to attempt to allocate new tags concurrently to limit
666 * false restarts in UMA.
668 uma_zone_reserve(vmem_bt_zone, BT_MAXALLOC * (mp_ncpus + 1) / 2);
669 uma_zone_set_allocf(vmem_bt_zone, vmem_bt_alloc);
676 vmem_rehash(vmem_t *vm, vmem_size_t newhashsize)
680 struct vmem_hashlist *newhashlist;
681 struct vmem_hashlist *oldhashlist;
682 vmem_size_t oldhashsize;
684 MPASS(newhashsize > 0);
686 newhashlist = malloc(sizeof(struct vmem_hashlist) * newhashsize,
688 if (newhashlist == NULL)
690 for (i = 0; i < newhashsize; i++) {
691 LIST_INIT(&newhashlist[i]);
695 oldhashlist = vm->vm_hashlist;
696 oldhashsize = vm->vm_hashsize;
697 vm->vm_hashlist = newhashlist;
698 vm->vm_hashsize = newhashsize;
699 if (oldhashlist == NULL) {
703 for (i = 0; i < oldhashsize; i++) {
704 while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) {
711 if (oldhashlist != vm->vm_hash0) {
712 free(oldhashlist, M_VMEM);
719 vmem_periodic_kick(void *dummy)
722 taskqueue_enqueue(taskqueue_thread, &vmem_periodic_wk);
726 vmem_periodic(void *unused, int pending)
732 mtx_lock(&vmem_list_lock);
733 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
735 /* Convenient time to verify vmem state. */
736 if (enable_vmem_check == 1) {
742 desired = 1 << flsl(vm->vm_nbusytag);
743 desired = MIN(MAX(desired, VMEM_HASHSIZE_MIN),
745 current = vm->vm_hashsize;
747 /* Grow in powers of two. Shrink less aggressively. */
748 if (desired >= current * 2 || desired * 4 <= current)
749 vmem_rehash(vm, desired);
752 * Periodically wake up threads waiting for resources,
753 * so they could ask for reclamation again.
755 VMEM_CONDVAR_BROADCAST(vm);
757 mtx_unlock(&vmem_list_lock);
759 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
760 vmem_periodic_kick, NULL);
764 vmem_start_callout(void *unused)
767 TASK_INIT(&vmem_periodic_wk, 0, vmem_periodic, NULL);
768 vmem_periodic_interval = hz * 10;
769 callout_init(&vmem_periodic_ch, 1);
770 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
771 vmem_periodic_kick, NULL);
773 SYSINIT(vfs, SI_SUB_CONFIGURE, SI_ORDER_ANY, vmem_start_callout, NULL);
776 vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int type)
781 MPASS(type == BT_TYPE_SPAN || type == BT_TYPE_SPAN_STATIC);
782 MPASS((size & vm->vm_quantum_mask) == 0);
784 btspan = bt_alloc(vm);
785 btspan->bt_type = type;
786 btspan->bt_start = addr;
787 btspan->bt_size = size;
788 bt_insseg_tail(vm, btspan);
790 btfree = bt_alloc(vm);
791 btfree->bt_type = BT_TYPE_FREE;
792 btfree->bt_start = addr;
793 btfree->bt_size = size;
794 bt_insseg(vm, btfree, btspan);
795 bt_insfree(vm, btfree);
801 vmem_destroy1(vmem_t *vm)
806 * Drain per-cpu quantum caches.
811 * The vmem should now only contain empty segments.
814 MPASS(vm->vm_nbusytag == 0);
816 while ((bt = TAILQ_FIRST(&vm->vm_seglist)) != NULL)
819 if (vm->vm_hashlist != NULL && vm->vm_hashlist != vm->vm_hash0)
820 free(vm->vm_hashlist, M_VMEM);
824 VMEM_CONDVAR_DESTROY(vm);
825 VMEM_LOCK_DESTROY(vm);
830 vmem_import(vmem_t *vm, vmem_size_t size, vmem_size_t align, int flags)
835 if (vm->vm_importfn == NULL)
839 * To make sure we get a span that meets the alignment we double it
840 * and add the size to the tail. This slightly overestimates.
842 if (align != vm->vm_quantum_mask + 1)
843 size = (align * 2) + size;
844 size = roundup(size, vm->vm_import_quantum);
847 * Hide MAXALLOC tags so we're guaranteed to be able to add this
848 * span and the tag we want to allocate from it.
850 MPASS(vm->vm_nfreetags >= BT_MAXALLOC);
851 vm->vm_nfreetags -= BT_MAXALLOC;
853 error = (vm->vm_importfn)(vm->vm_arg, size, flags, &addr);
855 vm->vm_nfreetags += BT_MAXALLOC;
859 vmem_add1(vm, addr, size, BT_TYPE_SPAN);
865 * vmem_fit: check if a bt can satisfy the given restrictions.
867 * it's a caller's responsibility to ensure the region is big enough
871 vmem_fit(const bt_t *bt, vmem_size_t size, vmem_size_t align,
872 vmem_size_t phase, vmem_size_t nocross, vmem_addr_t minaddr,
873 vmem_addr_t maxaddr, vmem_addr_t *addrp)
879 MPASS(bt->bt_size >= size); /* caller's responsibility */
882 * XXX assumption: vmem_addr_t and vmem_size_t are
883 * unsigned integer of the same size.
886 start = bt->bt_start;
887 if (start < minaddr) {
896 start = VMEM_ALIGNUP(start - phase, align) + phase;
897 if (start < bt->bt_start)
899 if (VMEM_CROSS_P(start, start + size - 1, nocross)) {
900 MPASS(align < nocross);
901 start = VMEM_ALIGNUP(start - phase, nocross) + phase;
903 if (start <= end && end - start >= size - 1) {
904 MPASS((start & (align - 1)) == phase);
905 MPASS(!VMEM_CROSS_P(start, start + size - 1, nocross));
906 MPASS(minaddr <= start);
907 MPASS(maxaddr == 0 || start + size - 1 <= maxaddr);
908 MPASS(bt->bt_start <= start);
909 MPASS(BT_END(bt) - start >= size - 1);
918 * vmem_clip: Trim the boundary tag edges to the requested start and size.
921 vmem_clip(vmem_t *vm, bt_t *bt, vmem_addr_t start, vmem_size_t size)
926 VMEM_ASSERT_LOCKED(vm);
927 MPASS(bt->bt_type == BT_TYPE_FREE);
928 MPASS(bt->bt_size >= size);
930 if (bt->bt_start != start) {
931 btprev = bt_alloc(vm);
932 btprev->bt_type = BT_TYPE_FREE;
933 btprev->bt_start = bt->bt_start;
934 btprev->bt_size = start - bt->bt_start;
935 bt->bt_start = start;
936 bt->bt_size -= btprev->bt_size;
937 bt_insfree(vm, btprev);
938 bt_insseg(vm, btprev,
939 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
941 MPASS(bt->bt_start == start);
942 if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) {
944 btnew = bt_alloc(vm);
945 btnew->bt_type = BT_TYPE_BUSY;
946 btnew->bt_start = bt->bt_start;
947 btnew->bt_size = size;
948 bt->bt_start = bt->bt_start + size;
952 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
953 bt_insbusy(vm, btnew);
956 bt->bt_type = BT_TYPE_BUSY;
959 MPASS(bt->bt_size >= size);
960 bt->bt_type = BT_TYPE_BUSY;
966 vmem_set_import(vmem_t *vm, vmem_import_t *importfn,
967 vmem_release_t *releasefn, void *arg, vmem_size_t import_quantum)
971 vm->vm_importfn = importfn;
972 vm->vm_releasefn = releasefn;
974 vm->vm_import_quantum = import_quantum;
979 vmem_set_reclaim(vmem_t *vm, vmem_reclaim_t *reclaimfn)
983 vm->vm_reclaimfn = reclaimfn;
988 * vmem_init: Initializes vmem arena.
991 vmem_init(vmem_t *vm, const char *name, vmem_addr_t base, vmem_size_t size,
992 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
997 MPASS((quantum & (quantum - 1)) == 0);
999 bzero(vm, sizeof(*vm));
1001 VMEM_CONDVAR_INIT(vm, name);
1002 VMEM_LOCK_INIT(vm, name);
1003 vm->vm_nfreetags = 0;
1004 LIST_INIT(&vm->vm_freetags);
1005 strlcpy(vm->vm_name, name, sizeof(vm->vm_name));
1006 vm->vm_quantum_mask = quantum - 1;
1007 vm->vm_quantum_shift = flsl(quantum) - 1;
1008 vm->vm_nbusytag = 0;
1011 qc_init(vm, qcache_max);
1013 TAILQ_INIT(&vm->vm_seglist);
1014 for (i = 0; i < VMEM_MAXORDER; i++) {
1015 LIST_INIT(&vm->vm_freelist[i]);
1017 memset(&vm->vm_hash0, 0, sizeof(vm->vm_hash0));
1018 vm->vm_hashsize = VMEM_HASHSIZE_MIN;
1019 vm->vm_hashlist = vm->vm_hash0;
1022 if (vmem_add(vm, base, size, flags) != 0) {
1028 mtx_lock(&vmem_list_lock);
1029 LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist);
1030 mtx_unlock(&vmem_list_lock);
1036 * vmem_create: create an arena.
1039 vmem_create(const char *name, vmem_addr_t base, vmem_size_t size,
1040 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1045 vm = malloc(sizeof(*vm), M_VMEM, flags & (M_WAITOK|M_NOWAIT));
1048 if (vmem_init(vm, name, base, size, quantum, qcache_max,
1055 vmem_destroy(vmem_t *vm)
1058 mtx_lock(&vmem_list_lock);
1059 LIST_REMOVE(vm, vm_alllist);
1060 mtx_unlock(&vmem_list_lock);
1066 vmem_roundup_size(vmem_t *vm, vmem_size_t size)
1069 return (size + vm->vm_quantum_mask) & ~vm->vm_quantum_mask;
1073 * vmem_alloc: allocate resource from the arena.
1076 vmem_alloc(vmem_t *vm, vmem_size_t size, int flags, vmem_addr_t *addrp)
1078 const int strat __unused = flags & VMEM_FITMASK;
1081 flags &= VMEM_FLAGS;
1083 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT);
1084 if ((flags & M_NOWAIT) == 0)
1085 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_alloc");
1087 if (size <= vm->vm_qcache_max) {
1088 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1089 *addrp = (vmem_addr_t)uma_zalloc(qc->qc_cache, flags);
1095 return vmem_xalloc(vm, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
1100 vmem_xalloc(vmem_t *vm, const vmem_size_t size0, vmem_size_t align,
1101 const vmem_size_t phase, const vmem_size_t nocross,
1102 const vmem_addr_t minaddr, const vmem_addr_t maxaddr, int flags,
1105 const vmem_size_t size = vmem_roundup_size(vm, size0);
1106 struct vmem_freelist *list;
1107 struct vmem_freelist *first;
1108 struct vmem_freelist *end;
1114 flags &= VMEM_FLAGS;
1115 strat = flags & VMEM_FITMASK;
1118 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT);
1119 MPASS((flags & (M_NOWAIT|M_WAITOK)) != (M_NOWAIT|M_WAITOK));
1120 if ((flags & M_NOWAIT) == 0)
1121 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_xalloc");
1122 MPASS((align & vm->vm_quantum_mask) == 0);
1123 MPASS((align & (align - 1)) == 0);
1124 MPASS((phase & vm->vm_quantum_mask) == 0);
1125 MPASS((nocross & vm->vm_quantum_mask) == 0);
1126 MPASS((nocross & (nocross - 1)) == 0);
1127 MPASS((align == 0 && phase == 0) || phase < align);
1128 MPASS(nocross == 0 || nocross >= size);
1129 MPASS(minaddr <= maxaddr);
1130 MPASS(!VMEM_CROSS_P(phase, phase + size - 1, nocross));
1133 align = vm->vm_quantum_mask + 1;
1136 end = &vm->vm_freelist[VMEM_MAXORDER];
1138 * choose a free block from which we allocate.
1140 first = bt_freehead_toalloc(vm, size, strat);
1144 * Make sure we have enough tags to complete the
1147 if (vm->vm_nfreetags < BT_MAXALLOC &&
1148 bt_fill(vm, flags) != 0) {
1153 * Scan freelists looking for a tag that satisfies the
1154 * allocation. If we're doing BESTFIT we may encounter
1155 * sizes below the request. If we're doing FIRSTFIT we
1156 * inspect only the first element from each list.
1158 for (list = first; list < end; list++) {
1159 LIST_FOREACH(bt, list, bt_freelist) {
1160 if (bt->bt_size >= size) {
1161 error = vmem_fit(bt, size, align, phase,
1162 nocross, minaddr, maxaddr, addrp);
1164 vmem_clip(vm, bt, *addrp, size);
1168 /* FIRST skips to the next list. */
1169 if (strat == M_FIRSTFIT)
1174 * Retry if the fast algorithm failed.
1176 if (strat == M_FIRSTFIT) {
1178 first = bt_freehead_toalloc(vm, size, strat);
1182 * XXX it is possible to fail to meet restrictions with the
1183 * imported region. It is up to the user to specify the
1184 * import quantum such that it can satisfy any allocation.
1186 if (vmem_import(vm, size, align, flags) == 0)
1190 * Try to free some space from the quantum cache or reclaim
1191 * functions if available.
1193 if (vm->vm_qcache_max != 0 || vm->vm_reclaimfn != NULL) {
1194 avail = vm->vm_size - vm->vm_inuse;
1196 if (vm->vm_qcache_max != 0)
1198 if (vm->vm_reclaimfn != NULL)
1199 vm->vm_reclaimfn(vm, flags);
1201 /* If we were successful retry even NOWAIT. */
1202 if (vm->vm_size - vm->vm_inuse > avail)
1205 if ((flags & M_NOWAIT) != 0) {
1209 VMEM_CONDVAR_WAIT(vm);
1213 if (error != 0 && (flags & M_NOWAIT) == 0)
1214 panic("failed to allocate waiting allocation\n");
1220 * vmem_free: free the resource to the arena.
1223 vmem_free(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1228 if (size <= vm->vm_qcache_max) {
1229 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1230 uma_zfree(qc->qc_cache, (void *)addr);
1232 vmem_xfree(vm, addr, size);
1236 vmem_xfree(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1244 bt = bt_lookupbusy(vm, addr);
1246 MPASS(bt->bt_start == addr);
1247 MPASS(bt->bt_size == vmem_roundup_size(vm, size) ||
1248 bt->bt_size - vmem_roundup_size(vm, size) <= vm->vm_quantum_mask);
1249 MPASS(bt->bt_type == BT_TYPE_BUSY);
1251 bt->bt_type = BT_TYPE_FREE;
1254 t = TAILQ_NEXT(bt, bt_seglist);
1255 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1256 MPASS(BT_END(bt) < t->bt_start); /* YYY */
1257 bt->bt_size += t->bt_size;
1261 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1262 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1263 MPASS(BT_END(t) < bt->bt_start); /* YYY */
1264 bt->bt_size += t->bt_size;
1265 bt->bt_start = t->bt_start;
1270 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1272 MPASS(BT_ISSPAN_P(t) || t->bt_type == BT_TYPE_BUSY);
1273 if (vm->vm_releasefn != NULL && t->bt_type == BT_TYPE_SPAN &&
1274 t->bt_size == bt->bt_size) {
1275 vmem_addr_t spanaddr;
1276 vmem_size_t spansize;
1278 MPASS(t->bt_start == bt->bt_start);
1279 spanaddr = bt->bt_start;
1280 spansize = bt->bt_size;
1283 vm->vm_size -= spansize;
1284 VMEM_CONDVAR_BROADCAST(vm);
1285 bt_freetrim(vm, BT_MAXFREE);
1286 (*vm->vm_releasefn)(vm->vm_arg, spanaddr, spansize);
1289 VMEM_CONDVAR_BROADCAST(vm);
1290 bt_freetrim(vm, BT_MAXFREE);
1299 vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int flags)
1304 flags &= VMEM_FLAGS;
1306 if (vm->vm_nfreetags >= BT_MAXALLOC || bt_fill(vm, flags) == 0)
1307 vmem_add1(vm, addr, size, BT_TYPE_SPAN_STATIC);
1316 * vmem_size: information about arenas size
1319 vmem_size(vmem_t *vm, int typemask)
1325 return vm->vm_inuse;
1327 return vm->vm_size - vm->vm_inuse;
1328 case VMEM_FREE|VMEM_ALLOC:
1332 for (i = VMEM_MAXORDER - 1; i >= 0; i--) {
1333 if (LIST_EMPTY(&vm->vm_freelist[i]))
1336 return ((vmem_size_t)ORDER2SIZE(i) <<
1337 vm->vm_quantum_shift);
1348 #if defined(DDB) || defined(DIAGNOSTIC)
1350 static void bt_dump(const bt_t *, int (*)(const char *, ...)
1351 __printflike(1, 2));
1354 bt_type_string(int type)
1364 case BT_TYPE_SPAN_STATIC:
1365 return "static span";
1373 bt_dump(const bt_t *bt, int (*pr)(const char *, ...))
1376 (*pr)("\t%p: %jx %jx, %d(%s)\n",
1377 bt, (intmax_t)bt->bt_start, (intmax_t)bt->bt_size,
1378 bt->bt_type, bt_type_string(bt->bt_type));
1382 vmem_dump(const vmem_t *vm , int (*pr)(const char *, ...) __printflike(1, 2))
1387 (*pr)("vmem %p '%s'\n", vm, vm->vm_name);
1388 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1392 for (i = 0; i < VMEM_MAXORDER; i++) {
1393 const struct vmem_freelist *fl = &vm->vm_freelist[i];
1395 if (LIST_EMPTY(fl)) {
1399 (*pr)("freelist[%d]\n", i);
1400 LIST_FOREACH(bt, fl, bt_freelist) {
1406 #endif /* defined(DDB) || defined(DIAGNOSTIC) */
1409 #include <ddb/ddb.h>
1412 vmem_whatis_lookup(vmem_t *vm, vmem_addr_t addr)
1416 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1417 if (BT_ISSPAN_P(bt)) {
1420 if (bt->bt_start <= addr && addr <= BT_END(bt)) {
1429 vmem_whatis(vmem_addr_t addr, int (*pr)(const char *, ...))
1433 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1436 bt = vmem_whatis_lookup(vm, addr);
1440 (*pr)("%p is %p+%zu in VMEM '%s' (%s)\n",
1441 (void *)addr, (void *)bt->bt_start,
1442 (vmem_size_t)(addr - bt->bt_start), vm->vm_name,
1443 (bt->bt_type == BT_TYPE_BUSY) ? "allocated" : "free");
1448 vmem_printall(const char *modif, int (*pr)(const char *, ...))
1452 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1458 vmem_print(vmem_addr_t addr, const char *modif, int (*pr)(const char *, ...))
1460 const vmem_t *vm = (const void *)addr;
1465 DB_SHOW_COMMAND(vmemdump, vmemdump)
1469 db_printf("usage: show vmemdump <addr>\n");
1473 vmem_dump((const vmem_t *)addr, db_printf);
1476 DB_SHOW_ALL_COMMAND(vmemdump, vmemdumpall)
1480 LIST_FOREACH(vm, &vmem_list, vm_alllist)
1481 vmem_dump(vm, db_printf);
1484 DB_SHOW_COMMAND(vmem, vmem_summ)
1486 const vmem_t *vm = (const void *)addr;
1488 size_t ft[VMEM_MAXORDER], ut[VMEM_MAXORDER];
1489 size_t fs[VMEM_MAXORDER], us[VMEM_MAXORDER];
1493 db_printf("usage: show vmem <addr>\n");
1497 db_printf("vmem %p '%s'\n", vm, vm->vm_name);
1498 db_printf("\tquantum:\t%zu\n", vm->vm_quantum_mask + 1);
1499 db_printf("\tsize:\t%zu\n", vm->vm_size);
1500 db_printf("\tinuse:\t%zu\n", vm->vm_inuse);
1501 db_printf("\tfree:\t%zu\n", vm->vm_size - vm->vm_inuse);
1502 db_printf("\tbusy tags:\t%d\n", vm->vm_nbusytag);
1503 db_printf("\tfree tags:\t%d\n", vm->vm_nfreetags);
1505 memset(&ft, 0, sizeof(ft));
1506 memset(&ut, 0, sizeof(ut));
1507 memset(&fs, 0, sizeof(fs));
1508 memset(&us, 0, sizeof(us));
1509 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1510 ord = SIZE2ORDER(bt->bt_size >> vm->vm_quantum_shift);
1511 if (bt->bt_type == BT_TYPE_BUSY) {
1513 us[ord] += bt->bt_size;
1514 } else if (bt->bt_type == BT_TYPE_FREE) {
1516 fs[ord] += bt->bt_size;
1519 db_printf("\t\t\tinuse\tsize\t\tfree\tsize\n");
1520 for (ord = 0; ord < VMEM_MAXORDER; ord++) {
1521 if (ut[ord] == 0 && ft[ord] == 0)
1523 db_printf("\t%-15zu %zu\t%-15zu %zu\t%-16zu\n",
1524 ORDER2SIZE(ord) << vm->vm_quantum_shift,
1525 ut[ord], us[ord], ft[ord], fs[ord]);
1529 DB_SHOW_ALL_COMMAND(vmem, vmem_summall)
1533 LIST_FOREACH(vm, &vmem_list, vm_alllist)
1534 vmem_summ((db_expr_t)vm, TRUE, count, modif);
1536 #endif /* defined(DDB) */
1538 #define vmem_printf printf
1540 #if defined(DIAGNOSTIC)
1543 vmem_check_sanity(vmem_t *vm)
1545 const bt_t *bt, *bt2;
1549 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1550 if (bt->bt_start > BT_END(bt)) {
1551 printf("corrupted tag\n");
1552 bt_dump(bt, vmem_printf);
1556 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1557 TAILQ_FOREACH(bt2, &vm->vm_seglist, bt_seglist) {
1561 if (BT_ISSPAN_P(bt) != BT_ISSPAN_P(bt2)) {
1564 if (bt->bt_start <= BT_END(bt2) &&
1565 bt2->bt_start <= BT_END(bt)) {
1566 printf("overwrapped tags\n");
1567 bt_dump(bt, vmem_printf);
1568 bt_dump(bt2, vmem_printf);
1578 vmem_check(vmem_t *vm)
1581 if (!vmem_check_sanity(vm)) {
1582 panic("insanity vmem %p", vm);
1586 #endif /* defined(DIAGNOSTIC) */