2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c)2006,2007,2008,2009 YAMAMOTO Takashi,
5 * Copyright (c) 2013 EMC Corp.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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.
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
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 $
38 * - Magazines and Vmem: Extending the Slab Allocator
39 * to Many CPUs and Arbitrary Resources
40 * http://www.usenix.org/event/usenix01/bonwick.html
43 #include <sys/cdefs.h>
44 __FBSDID("$FreeBSD$");
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>
55 #include <sys/malloc.h>
56 #include <sys/mutex.h>
58 #include <sys/condvar.h>
59 #include <sys/sysctl.h>
60 #include <sys/taskqueue.h>
68 #include <vm/vm_map.h>
69 #include <vm/vm_object.h>
70 #include <vm/vm_kern.h>
71 #include <vm/vm_extern.h>
72 #include <vm/vm_param.h>
73 #include <vm/vm_pageout.h>
75 #define VMEM_OPTORDER 5
76 #define VMEM_OPTVALUE (1 << VMEM_OPTORDER)
77 #define VMEM_MAXORDER \
78 (VMEM_OPTVALUE - 1 + sizeof(vmem_size_t) * NBBY - VMEM_OPTORDER)
80 #define VMEM_HASHSIZE_MIN 16
81 #define VMEM_HASHSIZE_MAX 131072
83 #define VMEM_QCACHE_IDX_MAX 16
85 #define VMEM_FITMASK (M_BESTFIT | M_FIRSTFIT)
88 (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM | M_BESTFIT | M_FIRSTFIT)
90 #define BT_FLAGS (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM)
92 #define QC_NAME_MAX 16
95 * Data structures private to vmem.
97 MALLOC_DEFINE(M_VMEM, "vmem", "vmem internal structures");
99 typedef struct vmem_btag bt_t;
101 TAILQ_HEAD(vmem_seglist, vmem_btag);
102 LIST_HEAD(vmem_freelist, vmem_btag);
103 LIST_HEAD(vmem_hashlist, vmem_btag);
109 char qc_name[QC_NAME_MAX];
111 typedef struct qcache qcache_t;
112 #define QC_POOL_TO_QCACHE(pool) ((qcache_t *)(pool->pr_qcache))
114 #define VMEM_NAME_MAX 16
118 struct mtx_padalign vm_lock;
120 char vm_name[VMEM_NAME_MAX+1];
121 LIST_ENTRY(vmem) vm_alllist;
122 struct vmem_hashlist vm_hash0[VMEM_HASHSIZE_MIN];
123 struct vmem_freelist vm_freelist[VMEM_MAXORDER];
124 struct vmem_seglist vm_seglist;
125 struct vmem_hashlist *vm_hashlist;
126 vmem_size_t vm_hashsize;
128 /* Constant after init */
129 vmem_size_t vm_qcache_max;
130 vmem_size_t vm_quantum_mask;
131 vmem_size_t vm_import_quantum;
132 int vm_quantum_shift;
134 /* Written on alloc/free */
135 LIST_HEAD(, vmem_btag) vm_freetags;
138 vmem_size_t vm_inuse;
140 vmem_size_t vm_limit;
142 /* Used on import. */
143 vmem_import_t *vm_importfn;
144 vmem_release_t *vm_releasefn;
147 /* Space exhaustion callback. */
148 vmem_reclaim_t *vm_reclaimfn;
151 qcache_t vm_qcache[VMEM_QCACHE_IDX_MAX];
156 TAILQ_ENTRY(vmem_btag) bt_seglist;
158 LIST_ENTRY(vmem_btag) u_freelist; /* BT_TYPE_FREE */
159 LIST_ENTRY(vmem_btag) u_hashlist; /* BT_TYPE_BUSY */
161 #define bt_hashlist bt_u.u_hashlist
162 #define bt_freelist bt_u.u_freelist
163 vmem_addr_t bt_start;
168 #define BT_TYPE_SPAN 1 /* Allocated from importfn */
169 #define BT_TYPE_SPAN_STATIC 2 /* vmem_add() or create. */
170 #define BT_TYPE_FREE 3 /* Available space. */
171 #define BT_TYPE_BUSY 4 /* Used space. */
172 #define BT_ISSPAN_P(bt) ((bt)->bt_type <= BT_TYPE_SPAN_STATIC)
174 #define BT_END(bt) ((bt)->bt_start + (bt)->bt_size - 1)
176 #if defined(DIAGNOSTIC)
177 static int enable_vmem_check = 1;
178 SYSCTL_INT(_debug, OID_AUTO, vmem_check, CTLFLAG_RWTUN,
179 &enable_vmem_check, 0, "Enable vmem check");
180 static void vmem_check(vmem_t *);
183 static struct callout vmem_periodic_ch;
184 static int vmem_periodic_interval;
185 static struct task vmem_periodic_wk;
187 static struct mtx_padalign __exclusive_cache_line vmem_list_lock;
188 static LIST_HEAD(, vmem) vmem_list = LIST_HEAD_INITIALIZER(vmem_list);
191 #define VMEM_CONDVAR_INIT(vm, wchan) cv_init(&vm->vm_cv, wchan)
192 #define VMEM_CONDVAR_DESTROY(vm) cv_destroy(&vm->vm_cv)
193 #define VMEM_CONDVAR_WAIT(vm) cv_wait(&vm->vm_cv, &vm->vm_lock)
194 #define VMEM_CONDVAR_BROADCAST(vm) cv_broadcast(&vm->vm_cv)
197 #define VMEM_LOCK(vm) mtx_lock(&vm->vm_lock)
198 #define VMEM_TRYLOCK(vm) mtx_trylock(&vm->vm_lock)
199 #define VMEM_UNLOCK(vm) mtx_unlock(&vm->vm_lock)
200 #define VMEM_LOCK_INIT(vm, name) mtx_init(&vm->vm_lock, (name), NULL, MTX_DEF)
201 #define VMEM_LOCK_DESTROY(vm) mtx_destroy(&vm->vm_lock)
202 #define VMEM_ASSERT_LOCKED(vm) mtx_assert(&vm->vm_lock, MA_OWNED);
204 #define VMEM_ALIGNUP(addr, align) (-(-(addr) & -(align)))
206 #define VMEM_CROSS_P(addr1, addr2, boundary) \
207 ((((addr1) ^ (addr2)) & -(boundary)) != 0)
209 #define ORDER2SIZE(order) ((order) < VMEM_OPTVALUE ? ((order) + 1) : \
210 (vmem_size_t)1 << ((order) - (VMEM_OPTVALUE - VMEM_OPTORDER - 1)))
211 #define SIZE2ORDER(size) ((size) <= VMEM_OPTVALUE ? ((size) - 1) : \
212 (flsl(size) + (VMEM_OPTVALUE - VMEM_OPTORDER - 2)))
215 * Maximum number of boundary tags that may be required to satisfy an
216 * allocation. Two may be required to import. Another two may be
217 * required to clip edges.
219 #define BT_MAXALLOC 4
222 * Max free limits the number of locally cached boundary tags. We
223 * just want to avoid hitting the zone allocator for every call.
225 #define BT_MAXFREE (BT_MAXALLOC * 8)
227 /* Allocator for boundary tags. */
228 static uma_zone_t vmem_bt_zone;
230 /* boot time arena storage. */
231 static struct vmem kernel_arena_storage;
232 static struct vmem buffer_arena_storage;
233 static struct vmem transient_arena_storage;
234 /* kernel and kmem arenas are aliased for backwards KPI compat. */
235 vmem_t *kernel_arena = &kernel_arena_storage;
236 vmem_t *kmem_arena = &kernel_arena_storage;
237 vmem_t *buffer_arena = &buffer_arena_storage;
238 vmem_t *transient_arena = &transient_arena_storage;
240 #ifdef DEBUG_MEMGUARD
241 static struct vmem memguard_arena_storage;
242 vmem_t *memguard_arena = &memguard_arena_storage;
246 * Fill the vmem's boundary tag cache. We guarantee that boundary tag
247 * allocation will not fail once bt_fill() passes. To do so we cache
248 * at least the maximum possible tag allocations in the arena.
251 bt_fill(vmem_t *vm, int flags)
255 VMEM_ASSERT_LOCKED(vm);
258 * Only allow the kernel arena to dip into reserve tags. It is the
259 * vmem where new tags come from.
262 if (vm != kernel_arena)
263 flags &= ~M_USE_RESERVE;
266 * Loop until we meet the reserve. To minimize the lock shuffle
267 * and prevent simultaneous fills we first try a NOWAIT regardless
268 * of the caller's flags. Specify M_NOVM so we don't recurse while
269 * holding a vmem lock.
271 while (vm->vm_nfreetags < BT_MAXALLOC) {
272 bt = uma_zalloc(vmem_bt_zone,
273 (flags & M_USE_RESERVE) | M_NOWAIT | M_NOVM);
276 bt = uma_zalloc(vmem_bt_zone, flags);
278 if (bt == NULL && (flags & M_NOWAIT) != 0)
281 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
285 if (vm->vm_nfreetags < BT_MAXALLOC)
292 * Pop a tag off of the freetag stack.
299 VMEM_ASSERT_LOCKED(vm);
300 bt = LIST_FIRST(&vm->vm_freetags);
302 LIST_REMOVE(bt, bt_freelist);
309 * Trim the per-vmem free list. Returns with the lock released to
310 * avoid allocator recursions.
313 bt_freetrim(vmem_t *vm, int freelimit)
315 LIST_HEAD(, vmem_btag) freetags;
318 LIST_INIT(&freetags);
319 VMEM_ASSERT_LOCKED(vm);
320 while (vm->vm_nfreetags > freelimit) {
321 bt = LIST_FIRST(&vm->vm_freetags);
322 LIST_REMOVE(bt, bt_freelist);
324 LIST_INSERT_HEAD(&freetags, bt, bt_freelist);
327 while ((bt = LIST_FIRST(&freetags)) != NULL) {
328 LIST_REMOVE(bt, bt_freelist);
329 uma_zfree(vmem_bt_zone, bt);
334 bt_free(vmem_t *vm, bt_t *bt)
337 VMEM_ASSERT_LOCKED(vm);
338 MPASS(LIST_FIRST(&vm->vm_freetags) != bt);
339 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
344 * freelist[0] ... [1, 1]
345 * freelist[1] ... [2, 2]
347 * freelist[29] ... [30, 30]
348 * freelist[30] ... [31, 31]
349 * freelist[31] ... [32, 63]
350 * freelist[33] ... [64, 127]
352 * freelist[n] ... [(1 << (n - 26)), (1 << (n - 25)) - 1]
356 static struct vmem_freelist *
357 bt_freehead_tofree(vmem_t *vm, vmem_size_t size)
359 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
360 const int idx = SIZE2ORDER(qsize);
362 MPASS(size != 0 && qsize != 0);
363 MPASS((size & vm->vm_quantum_mask) == 0);
365 MPASS(idx < VMEM_MAXORDER);
367 return &vm->vm_freelist[idx];
371 * bt_freehead_toalloc: return the freelist for the given size and allocation
374 * For M_FIRSTFIT, return the list in which any blocks are large enough
375 * for the requested size. otherwise, return the list which can have blocks
376 * large enough for the requested size.
378 static struct vmem_freelist *
379 bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, int strat)
381 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
382 int idx = SIZE2ORDER(qsize);
384 MPASS(size != 0 && qsize != 0);
385 MPASS((size & vm->vm_quantum_mask) == 0);
387 if (strat == M_FIRSTFIT && ORDER2SIZE(idx) != qsize) {
389 /* check too large request? */
392 MPASS(idx < VMEM_MAXORDER);
394 return &vm->vm_freelist[idx];
397 /* ---- boundary tag hash */
399 static struct vmem_hashlist *
400 bt_hashhead(vmem_t *vm, vmem_addr_t addr)
402 struct vmem_hashlist *list;
405 hash = hash32_buf(&addr, sizeof(addr), 0);
406 list = &vm->vm_hashlist[hash % vm->vm_hashsize];
412 bt_lookupbusy(vmem_t *vm, vmem_addr_t addr)
414 struct vmem_hashlist *list;
417 VMEM_ASSERT_LOCKED(vm);
418 list = bt_hashhead(vm, addr);
419 LIST_FOREACH(bt, list, bt_hashlist) {
420 if (bt->bt_start == addr) {
429 bt_rembusy(vmem_t *vm, bt_t *bt)
432 VMEM_ASSERT_LOCKED(vm);
433 MPASS(vm->vm_nbusytag > 0);
434 vm->vm_inuse -= bt->bt_size;
436 LIST_REMOVE(bt, bt_hashlist);
440 bt_insbusy(vmem_t *vm, bt_t *bt)
442 struct vmem_hashlist *list;
444 VMEM_ASSERT_LOCKED(vm);
445 MPASS(bt->bt_type == BT_TYPE_BUSY);
447 list = bt_hashhead(vm, bt->bt_start);
448 LIST_INSERT_HEAD(list, bt, bt_hashlist);
450 vm->vm_inuse += bt->bt_size;
453 /* ---- boundary tag list */
456 bt_remseg(vmem_t *vm, bt_t *bt)
459 TAILQ_REMOVE(&vm->vm_seglist, bt, bt_seglist);
464 bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev)
467 TAILQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist);
471 bt_insseg_tail(vmem_t *vm, bt_t *bt)
474 TAILQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist);
478 bt_remfree(vmem_t *vm, bt_t *bt)
481 MPASS(bt->bt_type == BT_TYPE_FREE);
483 LIST_REMOVE(bt, bt_freelist);
487 bt_insfree(vmem_t *vm, bt_t *bt)
489 struct vmem_freelist *list;
491 list = bt_freehead_tofree(vm, bt->bt_size);
492 LIST_INSERT_HEAD(list, bt, bt_freelist);
495 /* ---- vmem internal functions */
498 * Import from the arena into the quantum cache in UMA.
501 qc_import(void *arg, void **store, int cnt, int flags)
508 if ((flags & VMEM_FITMASK) == 0)
510 for (i = 0; i < cnt; i++) {
511 if (vmem_xalloc(qc->qc_vmem, qc->qc_size, 0, 0, 0,
512 VMEM_ADDR_MIN, VMEM_ADDR_MAX, flags, &addr) != 0)
514 store[i] = (void *)addr;
515 /* Only guarantee one allocation. */
523 * Release memory from the UMA cache to the arena.
526 qc_release(void *arg, void **store, int cnt)
532 for (i = 0; i < cnt; i++)
533 vmem_xfree(qc->qc_vmem, (vmem_addr_t)store[i], qc->qc_size);
537 qc_init(vmem_t *vm, vmem_size_t qcache_max)
544 MPASS((qcache_max & vm->vm_quantum_mask) == 0);
545 qcache_idx_max = MIN(qcache_max >> vm->vm_quantum_shift,
546 VMEM_QCACHE_IDX_MAX);
547 vm->vm_qcache_max = qcache_idx_max << vm->vm_quantum_shift;
548 for (i = 0; i < qcache_idx_max; i++) {
549 qc = &vm->vm_qcache[i];
550 size = (i + 1) << vm->vm_quantum_shift;
551 snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu",
555 qc->qc_cache = uma_zcache_create(qc->qc_name, size,
556 NULL, NULL, NULL, NULL, qc_import, qc_release, qc,
563 qc_destroy(vmem_t *vm)
568 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
569 for (i = 0; i < qcache_idx_max; i++)
570 uma_zdestroy(vm->vm_qcache[i].qc_cache);
579 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
580 for (i = 0; i < qcache_idx_max; i++)
581 zone_drain(vm->vm_qcache[i].qc_cache);
584 #ifndef UMA_MD_SMALL_ALLOC
586 static struct mtx_padalign __exclusive_cache_line vmem_bt_lock;
589 * vmem_bt_alloc: Allocate a new page of boundary tags.
591 * On architectures with uma_small_alloc there is no recursion; no address
592 * space need be allocated to allocate boundary tags. For the others, we
593 * must handle recursion. Boundary tags are necessary to allocate new
596 * UMA guarantees that enough tags are held in reserve to allocate a new
597 * page of kva. We dip into this reserve by specifying M_USE_RESERVE only
598 * when allocating the page to hold new boundary tags. In this way the
599 * reserve is automatically filled by the allocation that uses the reserve.
601 * We still have to guarantee that the new tags are allocated atomically since
602 * many threads may try concurrently. The bt_lock provides this guarantee.
603 * We convert WAITOK allocations to NOWAIT and then handle the blocking here
604 * on failure. It's ok to return NULL for a WAITOK allocation as UMA will
605 * loop again after checking to see if we lost the race to allocate.
607 * There is a small race between vmem_bt_alloc() returning the page and the
608 * zone lock being acquired to add the page to the zone. For WAITOK
609 * allocations we just pause briefly. NOWAIT may experience a transient
610 * failure. To alleviate this we permit a small number of simultaneous
611 * fills to proceed concurrently so NOWAIT is less likely to fail unless
612 * we are really out of KVA.
615 vmem_bt_alloc(uma_zone_t zone, vm_size_t bytes, uint8_t *pflag, int wait)
619 *pflag = UMA_SLAB_KERNEL;
622 * Single thread boundary tag allocation so that the address space
623 * and memory are added in one atomic operation.
625 mtx_lock(&vmem_bt_lock);
626 if (vmem_xalloc(kernel_arena, bytes, 0, 0, 0, VMEM_ADDR_MIN,
627 VMEM_ADDR_MAX, M_NOWAIT | M_NOVM | M_USE_RESERVE | M_BESTFIT,
629 if (kmem_back(kernel_object, addr, bytes,
630 M_NOWAIT | M_USE_RESERVE) == 0) {
631 mtx_unlock(&vmem_bt_lock);
632 return ((void *)addr);
634 vmem_xfree(kernel_arena, addr, bytes);
635 mtx_unlock(&vmem_bt_lock);
637 * Out of memory, not address space. This may not even be
638 * possible due to M_USE_RESERVE page allocation.
644 mtx_unlock(&vmem_bt_lock);
646 * We're either out of address space or lost a fill race.
659 mtx_init(&vmem_list_lock, "vmem list lock", NULL, MTX_DEF);
660 vmem_bt_zone = uma_zcreate("vmem btag",
661 sizeof(struct vmem_btag), NULL, NULL, NULL, NULL,
662 UMA_ALIGN_PTR, UMA_ZONE_VM);
663 #ifndef UMA_MD_SMALL_ALLOC
664 mtx_init(&vmem_bt_lock, "btag lock", NULL, MTX_DEF);
665 uma_prealloc(vmem_bt_zone, BT_MAXALLOC);
667 * Reserve enough tags to allocate new tags. We allow multiple
668 * CPUs to attempt to allocate new tags concurrently to limit
669 * false restarts in UMA.
671 uma_zone_reserve(vmem_bt_zone, BT_MAXALLOC * (mp_ncpus + 1) / 2);
672 uma_zone_set_allocf(vmem_bt_zone, vmem_bt_alloc);
679 vmem_rehash(vmem_t *vm, vmem_size_t newhashsize)
683 struct vmem_hashlist *newhashlist;
684 struct vmem_hashlist *oldhashlist;
685 vmem_size_t oldhashsize;
687 MPASS(newhashsize > 0);
689 newhashlist = malloc(sizeof(struct vmem_hashlist) * newhashsize,
691 if (newhashlist == NULL)
693 for (i = 0; i < newhashsize; i++) {
694 LIST_INIT(&newhashlist[i]);
698 oldhashlist = vm->vm_hashlist;
699 oldhashsize = vm->vm_hashsize;
700 vm->vm_hashlist = newhashlist;
701 vm->vm_hashsize = newhashsize;
702 if (oldhashlist == NULL) {
706 for (i = 0; i < oldhashsize; i++) {
707 while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) {
714 if (oldhashlist != vm->vm_hash0) {
715 free(oldhashlist, M_VMEM);
722 vmem_periodic_kick(void *dummy)
725 taskqueue_enqueue(taskqueue_thread, &vmem_periodic_wk);
729 vmem_periodic(void *unused, int pending)
735 mtx_lock(&vmem_list_lock);
736 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
738 /* Convenient time to verify vmem state. */
739 if (enable_vmem_check == 1) {
745 desired = 1 << flsl(vm->vm_nbusytag);
746 desired = MIN(MAX(desired, VMEM_HASHSIZE_MIN),
748 current = vm->vm_hashsize;
750 /* Grow in powers of two. Shrink less aggressively. */
751 if (desired >= current * 2 || desired * 4 <= current)
752 vmem_rehash(vm, desired);
755 * Periodically wake up threads waiting for resources,
756 * so they could ask for reclamation again.
758 VMEM_CONDVAR_BROADCAST(vm);
760 mtx_unlock(&vmem_list_lock);
762 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
763 vmem_periodic_kick, NULL);
767 vmem_start_callout(void *unused)
770 TASK_INIT(&vmem_periodic_wk, 0, vmem_periodic, NULL);
771 vmem_periodic_interval = hz * 10;
772 callout_init(&vmem_periodic_ch, 1);
773 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
774 vmem_periodic_kick, NULL);
776 SYSINIT(vfs, SI_SUB_CONFIGURE, SI_ORDER_ANY, vmem_start_callout, NULL);
779 vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int type)
784 MPASS(type == BT_TYPE_SPAN || type == BT_TYPE_SPAN_STATIC);
785 MPASS((size & vm->vm_quantum_mask) == 0);
787 btspan = bt_alloc(vm);
788 btspan->bt_type = type;
789 btspan->bt_start = addr;
790 btspan->bt_size = size;
791 bt_insseg_tail(vm, btspan);
793 btfree = bt_alloc(vm);
794 btfree->bt_type = BT_TYPE_FREE;
795 btfree->bt_start = addr;
796 btfree->bt_size = size;
797 bt_insseg(vm, btfree, btspan);
798 bt_insfree(vm, btfree);
804 vmem_destroy1(vmem_t *vm)
809 * Drain per-cpu quantum caches.
814 * The vmem should now only contain empty segments.
817 MPASS(vm->vm_nbusytag == 0);
819 while ((bt = TAILQ_FIRST(&vm->vm_seglist)) != NULL)
822 if (vm->vm_hashlist != NULL && vm->vm_hashlist != vm->vm_hash0)
823 free(vm->vm_hashlist, M_VMEM);
827 VMEM_CONDVAR_DESTROY(vm);
828 VMEM_LOCK_DESTROY(vm);
833 vmem_import(vmem_t *vm, vmem_size_t size, vmem_size_t align, int flags)
838 if (vm->vm_importfn == NULL)
842 * To make sure we get a span that meets the alignment we double it
843 * and add the size to the tail. This slightly overestimates.
845 if (align != vm->vm_quantum_mask + 1)
846 size = (align * 2) + size;
847 size = roundup(size, vm->vm_import_quantum);
849 if (vm->vm_limit != 0 && vm->vm_limit < vm->vm_size + size)
853 * Hide MAXALLOC tags so we're guaranteed to be able to add this
854 * span and the tag we want to allocate from it.
856 MPASS(vm->vm_nfreetags >= BT_MAXALLOC);
857 vm->vm_nfreetags -= BT_MAXALLOC;
859 error = (vm->vm_importfn)(vm->vm_arg, size, flags, &addr);
861 vm->vm_nfreetags += BT_MAXALLOC;
865 vmem_add1(vm, addr, size, BT_TYPE_SPAN);
871 * vmem_fit: check if a bt can satisfy the given restrictions.
873 * it's a caller's responsibility to ensure the region is big enough
877 vmem_fit(const bt_t *bt, vmem_size_t size, vmem_size_t align,
878 vmem_size_t phase, vmem_size_t nocross, vmem_addr_t minaddr,
879 vmem_addr_t maxaddr, vmem_addr_t *addrp)
885 MPASS(bt->bt_size >= size); /* caller's responsibility */
888 * XXX assumption: vmem_addr_t and vmem_size_t are
889 * unsigned integer of the same size.
892 start = bt->bt_start;
893 if (start < minaddr) {
902 start = VMEM_ALIGNUP(start - phase, align) + phase;
903 if (start < bt->bt_start)
905 if (VMEM_CROSS_P(start, start + size - 1, nocross)) {
906 MPASS(align < nocross);
907 start = VMEM_ALIGNUP(start - phase, nocross) + phase;
909 if (start <= end && end - start >= size - 1) {
910 MPASS((start & (align - 1)) == phase);
911 MPASS(!VMEM_CROSS_P(start, start + size - 1, nocross));
912 MPASS(minaddr <= start);
913 MPASS(maxaddr == 0 || start + size - 1 <= maxaddr);
914 MPASS(bt->bt_start <= start);
915 MPASS(BT_END(bt) - start >= size - 1);
924 * vmem_clip: Trim the boundary tag edges to the requested start and size.
927 vmem_clip(vmem_t *vm, bt_t *bt, vmem_addr_t start, vmem_size_t size)
932 VMEM_ASSERT_LOCKED(vm);
933 MPASS(bt->bt_type == BT_TYPE_FREE);
934 MPASS(bt->bt_size >= size);
936 if (bt->bt_start != start) {
937 btprev = bt_alloc(vm);
938 btprev->bt_type = BT_TYPE_FREE;
939 btprev->bt_start = bt->bt_start;
940 btprev->bt_size = start - bt->bt_start;
941 bt->bt_start = start;
942 bt->bt_size -= btprev->bt_size;
943 bt_insfree(vm, btprev);
944 bt_insseg(vm, btprev,
945 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
947 MPASS(bt->bt_start == start);
948 if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) {
950 btnew = bt_alloc(vm);
951 btnew->bt_type = BT_TYPE_BUSY;
952 btnew->bt_start = bt->bt_start;
953 btnew->bt_size = size;
954 bt->bt_start = bt->bt_start + size;
958 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
959 bt_insbusy(vm, btnew);
962 bt->bt_type = BT_TYPE_BUSY;
965 MPASS(bt->bt_size >= size);
966 bt->bt_type = BT_TYPE_BUSY;
972 vmem_set_import(vmem_t *vm, vmem_import_t *importfn,
973 vmem_release_t *releasefn, void *arg, vmem_size_t import_quantum)
977 vm->vm_importfn = importfn;
978 vm->vm_releasefn = releasefn;
980 vm->vm_import_quantum = import_quantum;
985 vmem_set_limit(vmem_t *vm, vmem_size_t limit)
989 vm->vm_limit = limit;
994 vmem_set_reclaim(vmem_t *vm, vmem_reclaim_t *reclaimfn)
998 vm->vm_reclaimfn = reclaimfn;
1003 * vmem_init: Initializes vmem arena.
1006 vmem_init(vmem_t *vm, const char *name, vmem_addr_t base, vmem_size_t size,
1007 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1012 MPASS((quantum & (quantum - 1)) == 0);
1014 bzero(vm, sizeof(*vm));
1016 VMEM_CONDVAR_INIT(vm, name);
1017 VMEM_LOCK_INIT(vm, name);
1018 vm->vm_nfreetags = 0;
1019 LIST_INIT(&vm->vm_freetags);
1020 strlcpy(vm->vm_name, name, sizeof(vm->vm_name));
1021 vm->vm_quantum_mask = quantum - 1;
1022 vm->vm_quantum_shift = flsl(quantum) - 1;
1023 vm->vm_nbusytag = 0;
1027 qc_init(vm, qcache_max);
1029 TAILQ_INIT(&vm->vm_seglist);
1030 for (i = 0; i < VMEM_MAXORDER; i++) {
1031 LIST_INIT(&vm->vm_freelist[i]);
1033 memset(&vm->vm_hash0, 0, sizeof(vm->vm_hash0));
1034 vm->vm_hashsize = VMEM_HASHSIZE_MIN;
1035 vm->vm_hashlist = vm->vm_hash0;
1038 if (vmem_add(vm, base, size, flags) != 0) {
1044 mtx_lock(&vmem_list_lock);
1045 LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist);
1046 mtx_unlock(&vmem_list_lock);
1052 * vmem_create: create an arena.
1055 vmem_create(const char *name, vmem_addr_t base, vmem_size_t size,
1056 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1061 vm = malloc(sizeof(*vm), M_VMEM, flags & (M_WAITOK|M_NOWAIT));
1064 if (vmem_init(vm, name, base, size, quantum, qcache_max,
1071 vmem_destroy(vmem_t *vm)
1074 mtx_lock(&vmem_list_lock);
1075 LIST_REMOVE(vm, vm_alllist);
1076 mtx_unlock(&vmem_list_lock);
1082 vmem_roundup_size(vmem_t *vm, vmem_size_t size)
1085 return (size + vm->vm_quantum_mask) & ~vm->vm_quantum_mask;
1089 * vmem_alloc: allocate resource from the arena.
1092 vmem_alloc(vmem_t *vm, vmem_size_t size, int flags, vmem_addr_t *addrp)
1094 const int strat __unused = flags & VMEM_FITMASK;
1097 flags &= VMEM_FLAGS;
1099 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT);
1100 if ((flags & M_NOWAIT) == 0)
1101 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_alloc");
1103 if (size <= vm->vm_qcache_max) {
1104 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1105 *addrp = (vmem_addr_t)uma_zalloc(qc->qc_cache, flags);
1111 return vmem_xalloc(vm, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
1116 vmem_xalloc(vmem_t *vm, const vmem_size_t size0, vmem_size_t align,
1117 const vmem_size_t phase, const vmem_size_t nocross,
1118 const vmem_addr_t minaddr, const vmem_addr_t maxaddr, int flags,
1121 const vmem_size_t size = vmem_roundup_size(vm, size0);
1122 struct vmem_freelist *list;
1123 struct vmem_freelist *first;
1124 struct vmem_freelist *end;
1130 flags &= VMEM_FLAGS;
1131 strat = flags & VMEM_FITMASK;
1134 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT);
1135 MPASS((flags & (M_NOWAIT|M_WAITOK)) != (M_NOWAIT|M_WAITOK));
1136 if ((flags & M_NOWAIT) == 0)
1137 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_xalloc");
1138 MPASS((align & vm->vm_quantum_mask) == 0);
1139 MPASS((align & (align - 1)) == 0);
1140 MPASS((phase & vm->vm_quantum_mask) == 0);
1141 MPASS((nocross & vm->vm_quantum_mask) == 0);
1142 MPASS((nocross & (nocross - 1)) == 0);
1143 MPASS((align == 0 && phase == 0) || phase < align);
1144 MPASS(nocross == 0 || nocross >= size);
1145 MPASS(minaddr <= maxaddr);
1146 MPASS(!VMEM_CROSS_P(phase, phase + size - 1, nocross));
1149 align = vm->vm_quantum_mask + 1;
1152 end = &vm->vm_freelist[VMEM_MAXORDER];
1154 * choose a free block from which we allocate.
1156 first = bt_freehead_toalloc(vm, size, strat);
1160 * Make sure we have enough tags to complete the
1163 if (vm->vm_nfreetags < BT_MAXALLOC &&
1164 bt_fill(vm, flags) != 0) {
1169 * Scan freelists looking for a tag that satisfies the
1170 * allocation. If we're doing BESTFIT we may encounter
1171 * sizes below the request. If we're doing FIRSTFIT we
1172 * inspect only the first element from each list.
1174 for (list = first; list < end; list++) {
1175 LIST_FOREACH(bt, list, bt_freelist) {
1176 if (bt->bt_size >= size) {
1177 error = vmem_fit(bt, size, align, phase,
1178 nocross, minaddr, maxaddr, addrp);
1180 vmem_clip(vm, bt, *addrp, size);
1184 /* FIRST skips to the next list. */
1185 if (strat == M_FIRSTFIT)
1190 * Retry if the fast algorithm failed.
1192 if (strat == M_FIRSTFIT) {
1194 first = bt_freehead_toalloc(vm, size, strat);
1198 * XXX it is possible to fail to meet restrictions with the
1199 * imported region. It is up to the user to specify the
1200 * import quantum such that it can satisfy any allocation.
1202 if (vmem_import(vm, size, align, flags) == 0)
1206 * Try to free some space from the quantum cache or reclaim
1207 * functions if available.
1209 if (vm->vm_qcache_max != 0 || vm->vm_reclaimfn != NULL) {
1210 avail = vm->vm_size - vm->vm_inuse;
1212 if (vm->vm_qcache_max != 0)
1214 if (vm->vm_reclaimfn != NULL)
1215 vm->vm_reclaimfn(vm, flags);
1217 /* If we were successful retry even NOWAIT. */
1218 if (vm->vm_size - vm->vm_inuse > avail)
1221 if ((flags & M_NOWAIT) != 0) {
1225 VMEM_CONDVAR_WAIT(vm);
1229 if (error != 0 && (flags & M_NOWAIT) == 0)
1230 panic("failed to allocate waiting allocation\n");
1236 * vmem_free: free the resource to the arena.
1239 vmem_free(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1244 if (size <= vm->vm_qcache_max) {
1245 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1246 uma_zfree(qc->qc_cache, (void *)addr);
1248 vmem_xfree(vm, addr, size);
1252 vmem_xfree(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1260 bt = bt_lookupbusy(vm, addr);
1262 MPASS(bt->bt_start == addr);
1263 MPASS(bt->bt_size == vmem_roundup_size(vm, size) ||
1264 bt->bt_size - vmem_roundup_size(vm, size) <= vm->vm_quantum_mask);
1265 MPASS(bt->bt_type == BT_TYPE_BUSY);
1267 bt->bt_type = BT_TYPE_FREE;
1270 t = TAILQ_NEXT(bt, bt_seglist);
1271 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1272 MPASS(BT_END(bt) < t->bt_start); /* YYY */
1273 bt->bt_size += t->bt_size;
1277 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1278 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1279 MPASS(BT_END(t) < bt->bt_start); /* YYY */
1280 bt->bt_size += t->bt_size;
1281 bt->bt_start = t->bt_start;
1286 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1288 MPASS(BT_ISSPAN_P(t) || t->bt_type == BT_TYPE_BUSY);
1289 if (vm->vm_releasefn != NULL && t->bt_type == BT_TYPE_SPAN &&
1290 t->bt_size == bt->bt_size) {
1291 vmem_addr_t spanaddr;
1292 vmem_size_t spansize;
1294 MPASS(t->bt_start == bt->bt_start);
1295 spanaddr = bt->bt_start;
1296 spansize = bt->bt_size;
1299 vm->vm_size -= spansize;
1300 VMEM_CONDVAR_BROADCAST(vm);
1301 bt_freetrim(vm, BT_MAXFREE);
1302 (*vm->vm_releasefn)(vm->vm_arg, spanaddr, spansize);
1305 VMEM_CONDVAR_BROADCAST(vm);
1306 bt_freetrim(vm, BT_MAXFREE);
1315 vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int flags)
1320 flags &= VMEM_FLAGS;
1322 if (vm->vm_nfreetags >= BT_MAXALLOC || bt_fill(vm, flags) == 0)
1323 vmem_add1(vm, addr, size, BT_TYPE_SPAN_STATIC);
1332 * vmem_size: information about arenas size
1335 vmem_size(vmem_t *vm, int typemask)
1341 return vm->vm_inuse;
1343 return vm->vm_size - vm->vm_inuse;
1344 case VMEM_FREE|VMEM_ALLOC:
1348 for (i = VMEM_MAXORDER - 1; i >= 0; i--) {
1349 if (LIST_EMPTY(&vm->vm_freelist[i]))
1352 return ((vmem_size_t)ORDER2SIZE(i) <<
1353 vm->vm_quantum_shift);
1364 #if defined(DDB) || defined(DIAGNOSTIC)
1366 static void bt_dump(const bt_t *, int (*)(const char *, ...)
1367 __printflike(1, 2));
1370 bt_type_string(int type)
1380 case BT_TYPE_SPAN_STATIC:
1381 return "static span";
1389 bt_dump(const bt_t *bt, int (*pr)(const char *, ...))
1392 (*pr)("\t%p: %jx %jx, %d(%s)\n",
1393 bt, (intmax_t)bt->bt_start, (intmax_t)bt->bt_size,
1394 bt->bt_type, bt_type_string(bt->bt_type));
1398 vmem_dump(const vmem_t *vm , int (*pr)(const char *, ...) __printflike(1, 2))
1403 (*pr)("vmem %p '%s'\n", vm, vm->vm_name);
1404 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1408 for (i = 0; i < VMEM_MAXORDER; i++) {
1409 const struct vmem_freelist *fl = &vm->vm_freelist[i];
1411 if (LIST_EMPTY(fl)) {
1415 (*pr)("freelist[%d]\n", i);
1416 LIST_FOREACH(bt, fl, bt_freelist) {
1422 #endif /* defined(DDB) || defined(DIAGNOSTIC) */
1425 #include <ddb/ddb.h>
1428 vmem_whatis_lookup(vmem_t *vm, vmem_addr_t addr)
1432 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1433 if (BT_ISSPAN_P(bt)) {
1436 if (bt->bt_start <= addr && addr <= BT_END(bt)) {
1445 vmem_whatis(vmem_addr_t addr, int (*pr)(const char *, ...))
1449 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1452 bt = vmem_whatis_lookup(vm, addr);
1456 (*pr)("%p is %p+%zu in VMEM '%s' (%s)\n",
1457 (void *)addr, (void *)bt->bt_start,
1458 (vmem_size_t)(addr - bt->bt_start), vm->vm_name,
1459 (bt->bt_type == BT_TYPE_BUSY) ? "allocated" : "free");
1464 vmem_printall(const char *modif, int (*pr)(const char *, ...))
1468 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1474 vmem_print(vmem_addr_t addr, const char *modif, int (*pr)(const char *, ...))
1476 const vmem_t *vm = (const void *)addr;
1481 DB_SHOW_COMMAND(vmemdump, vmemdump)
1485 db_printf("usage: show vmemdump <addr>\n");
1489 vmem_dump((const vmem_t *)addr, db_printf);
1492 DB_SHOW_ALL_COMMAND(vmemdump, vmemdumpall)
1496 LIST_FOREACH(vm, &vmem_list, vm_alllist)
1497 vmem_dump(vm, db_printf);
1500 DB_SHOW_COMMAND(vmem, vmem_summ)
1502 const vmem_t *vm = (const void *)addr;
1504 size_t ft[VMEM_MAXORDER], ut[VMEM_MAXORDER];
1505 size_t fs[VMEM_MAXORDER], us[VMEM_MAXORDER];
1509 db_printf("usage: show vmem <addr>\n");
1513 db_printf("vmem %p '%s'\n", vm, vm->vm_name);
1514 db_printf("\tquantum:\t%zu\n", vm->vm_quantum_mask + 1);
1515 db_printf("\tsize:\t%zu\n", vm->vm_size);
1516 db_printf("\tinuse:\t%zu\n", vm->vm_inuse);
1517 db_printf("\tfree:\t%zu\n", vm->vm_size - vm->vm_inuse);
1518 db_printf("\tbusy tags:\t%d\n", vm->vm_nbusytag);
1519 db_printf("\tfree tags:\t%d\n", vm->vm_nfreetags);
1521 memset(&ft, 0, sizeof(ft));
1522 memset(&ut, 0, sizeof(ut));
1523 memset(&fs, 0, sizeof(fs));
1524 memset(&us, 0, sizeof(us));
1525 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1526 ord = SIZE2ORDER(bt->bt_size >> vm->vm_quantum_shift);
1527 if (bt->bt_type == BT_TYPE_BUSY) {
1529 us[ord] += bt->bt_size;
1530 } else if (bt->bt_type == BT_TYPE_FREE) {
1532 fs[ord] += bt->bt_size;
1535 db_printf("\t\t\tinuse\tsize\t\tfree\tsize\n");
1536 for (ord = 0; ord < VMEM_MAXORDER; ord++) {
1537 if (ut[ord] == 0 && ft[ord] == 0)
1539 db_printf("\t%-15zu %zu\t%-15zu %zu\t%-16zu\n",
1540 ORDER2SIZE(ord) << vm->vm_quantum_shift,
1541 ut[ord], us[ord], ft[ord], fs[ord]);
1545 DB_SHOW_ALL_COMMAND(vmem, vmem_summall)
1549 LIST_FOREACH(vm, &vmem_list, vm_alllist)
1550 vmem_summ((db_expr_t)vm, TRUE, count, modif);
1552 #endif /* defined(DDB) */
1554 #define vmem_printf printf
1556 #if defined(DIAGNOSTIC)
1559 vmem_check_sanity(vmem_t *vm)
1561 const bt_t *bt, *bt2;
1565 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1566 if (bt->bt_start > BT_END(bt)) {
1567 printf("corrupted tag\n");
1568 bt_dump(bt, vmem_printf);
1572 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1573 TAILQ_FOREACH(bt2, &vm->vm_seglist, bt_seglist) {
1577 if (BT_ISSPAN_P(bt) != BT_ISSPAN_P(bt2)) {
1580 if (bt->bt_start <= BT_END(bt2) &&
1581 bt2->bt_start <= BT_END(bt)) {
1582 printf("overwrapped tags\n");
1583 bt_dump(bt, vmem_printf);
1584 bt_dump(bt2, vmem_printf);
1594 vmem_check(vmem_t *vm)
1597 if (!vmem_check_sanity(vm)) {
1598 panic("insanity vmem %p", vm);
1602 #endif /* defined(DIAGNOSTIC) */