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_page.h>
74 #include <vm/vm_pageout.h>
76 #define VMEM_OPTORDER 5
77 #define VMEM_OPTVALUE (1 << VMEM_OPTORDER)
78 #define VMEM_MAXORDER \
79 (VMEM_OPTVALUE - 1 + sizeof(vmem_size_t) * NBBY - VMEM_OPTORDER)
81 #define VMEM_HASHSIZE_MIN 16
82 #define VMEM_HASHSIZE_MAX 131072
84 #define VMEM_QCACHE_IDX_MAX 16
86 #define VMEM_FITMASK (M_BESTFIT | M_FIRSTFIT)
89 (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM | M_BESTFIT | M_FIRSTFIT)
91 #define BT_FLAGS (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM)
93 #define QC_NAME_MAX 16
96 * Data structures private to vmem.
98 MALLOC_DEFINE(M_VMEM, "vmem", "vmem internal structures");
100 typedef struct vmem_btag bt_t;
102 TAILQ_HEAD(vmem_seglist, vmem_btag);
103 LIST_HEAD(vmem_freelist, vmem_btag);
104 LIST_HEAD(vmem_hashlist, vmem_btag);
110 char qc_name[QC_NAME_MAX];
112 typedef struct qcache qcache_t;
113 #define QC_POOL_TO_QCACHE(pool) ((qcache_t *)(pool->pr_qcache))
115 #define VMEM_NAME_MAX 16
119 struct mtx_padalign vm_lock;
121 char vm_name[VMEM_NAME_MAX+1];
122 LIST_ENTRY(vmem) vm_alllist;
123 struct vmem_hashlist vm_hash0[VMEM_HASHSIZE_MIN];
124 struct vmem_freelist vm_freelist[VMEM_MAXORDER];
125 struct vmem_seglist vm_seglist;
126 struct vmem_hashlist *vm_hashlist;
127 vmem_size_t vm_hashsize;
129 /* Constant after init */
130 vmem_size_t vm_qcache_max;
131 vmem_size_t vm_quantum_mask;
132 vmem_size_t vm_import_quantum;
133 int vm_quantum_shift;
135 /* Written on alloc/free */
136 LIST_HEAD(, vmem_btag) vm_freetags;
139 vmem_size_t vm_inuse;
141 vmem_size_t vm_limit;
143 /* Used on import. */
144 vmem_import_t *vm_importfn;
145 vmem_release_t *vm_releasefn;
148 /* Space exhaustion callback. */
149 vmem_reclaim_t *vm_reclaimfn;
152 qcache_t vm_qcache[VMEM_QCACHE_IDX_MAX];
157 TAILQ_ENTRY(vmem_btag) bt_seglist;
159 LIST_ENTRY(vmem_btag) u_freelist; /* BT_TYPE_FREE */
160 LIST_ENTRY(vmem_btag) u_hashlist; /* BT_TYPE_BUSY */
162 #define bt_hashlist bt_u.u_hashlist
163 #define bt_freelist bt_u.u_freelist
164 vmem_addr_t bt_start;
169 #define BT_TYPE_SPAN 1 /* Allocated from importfn */
170 #define BT_TYPE_SPAN_STATIC 2 /* vmem_add() or create. */
171 #define BT_TYPE_FREE 3 /* Available space. */
172 #define BT_TYPE_BUSY 4 /* Used space. */
173 #define BT_ISSPAN_P(bt) ((bt)->bt_type <= BT_TYPE_SPAN_STATIC)
175 #define BT_END(bt) ((bt)->bt_start + (bt)->bt_size - 1)
177 #if defined(DIAGNOSTIC)
178 static int enable_vmem_check = 1;
179 SYSCTL_INT(_debug, OID_AUTO, vmem_check, CTLFLAG_RWTUN,
180 &enable_vmem_check, 0, "Enable vmem check");
181 static void vmem_check(vmem_t *);
184 static struct callout vmem_periodic_ch;
185 static int vmem_periodic_interval;
186 static struct task vmem_periodic_wk;
188 static struct mtx_padalign __exclusive_cache_line vmem_list_lock;
189 static LIST_HEAD(, vmem) vmem_list = LIST_HEAD_INITIALIZER(vmem_list);
190 static uma_zone_t vmem_zone;
193 #define VMEM_CONDVAR_INIT(vm, wchan) cv_init(&vm->vm_cv, wchan)
194 #define VMEM_CONDVAR_DESTROY(vm) cv_destroy(&vm->vm_cv)
195 #define VMEM_CONDVAR_WAIT(vm) cv_wait(&vm->vm_cv, &vm->vm_lock)
196 #define VMEM_CONDVAR_BROADCAST(vm) cv_broadcast(&vm->vm_cv)
199 #define VMEM_LOCK(vm) mtx_lock(&vm->vm_lock)
200 #define VMEM_TRYLOCK(vm) mtx_trylock(&vm->vm_lock)
201 #define VMEM_UNLOCK(vm) mtx_unlock(&vm->vm_lock)
202 #define VMEM_LOCK_INIT(vm, name) mtx_init(&vm->vm_lock, (name), NULL, MTX_DEF)
203 #define VMEM_LOCK_DESTROY(vm) mtx_destroy(&vm->vm_lock)
204 #define VMEM_ASSERT_LOCKED(vm) mtx_assert(&vm->vm_lock, MA_OWNED);
206 #define VMEM_ALIGNUP(addr, align) (-(-(addr) & -(align)))
208 #define VMEM_CROSS_P(addr1, addr2, boundary) \
209 ((((addr1) ^ (addr2)) & -(boundary)) != 0)
211 #define ORDER2SIZE(order) ((order) < VMEM_OPTVALUE ? ((order) + 1) : \
212 (vmem_size_t)1 << ((order) - (VMEM_OPTVALUE - VMEM_OPTORDER - 1)))
213 #define SIZE2ORDER(size) ((size) <= VMEM_OPTVALUE ? ((size) - 1) : \
214 (flsl(size) + (VMEM_OPTVALUE - VMEM_OPTORDER - 2)))
217 * Maximum number of boundary tags that may be required to satisfy an
218 * allocation. Two may be required to import. Another two may be
219 * required to clip edges.
221 #define BT_MAXALLOC 4
224 * Max free limits the number of locally cached boundary tags. We
225 * just want to avoid hitting the zone allocator for every call.
227 #define BT_MAXFREE (BT_MAXALLOC * 8)
229 /* Allocator for boundary tags. */
230 static uma_zone_t vmem_bt_zone;
232 /* boot time arena storage. */
233 static struct vmem kernel_arena_storage;
234 static struct vmem buffer_arena_storage;
235 static struct vmem transient_arena_storage;
236 /* kernel and kmem arenas are aliased for backwards KPI compat. */
237 vmem_t *kernel_arena = &kernel_arena_storage;
238 vmem_t *kmem_arena = &kernel_arena_storage;
239 vmem_t *buffer_arena = &buffer_arena_storage;
240 vmem_t *transient_arena = &transient_arena_storage;
242 #ifdef DEBUG_MEMGUARD
243 static struct vmem memguard_arena_storage;
244 vmem_t *memguard_arena = &memguard_arena_storage;
248 * Fill the vmem's boundary tag cache. We guarantee that boundary tag
249 * allocation will not fail once bt_fill() passes. To do so we cache
250 * at least the maximum possible tag allocations in the arena.
253 bt_fill(vmem_t *vm, int flags)
257 VMEM_ASSERT_LOCKED(vm);
260 * Only allow the kernel arena and arenas derived from kernel arena to
261 * dip into reserve tags. They are where new tags come from.
264 if (vm != kernel_arena && vm->vm_arg != kernel_arena)
265 flags &= ~M_USE_RESERVE;
268 * Loop until we meet the reserve. To minimize the lock shuffle
269 * and prevent simultaneous fills we first try a NOWAIT regardless
270 * of the caller's flags. Specify M_NOVM so we don't recurse while
271 * holding a vmem lock.
273 while (vm->vm_nfreetags < BT_MAXALLOC) {
274 bt = uma_zalloc(vmem_bt_zone,
275 (flags & M_USE_RESERVE) | M_NOWAIT | M_NOVM);
278 bt = uma_zalloc(vmem_bt_zone, flags);
280 if (bt == NULL && (flags & M_NOWAIT) != 0)
283 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
287 if (vm->vm_nfreetags < BT_MAXALLOC)
294 * Pop a tag off of the freetag stack.
301 VMEM_ASSERT_LOCKED(vm);
302 bt = LIST_FIRST(&vm->vm_freetags);
304 LIST_REMOVE(bt, bt_freelist);
311 * Trim the per-vmem free list. Returns with the lock released to
312 * avoid allocator recursions.
315 bt_freetrim(vmem_t *vm, int freelimit)
317 LIST_HEAD(, vmem_btag) freetags;
320 LIST_INIT(&freetags);
321 VMEM_ASSERT_LOCKED(vm);
322 while (vm->vm_nfreetags > freelimit) {
323 bt = LIST_FIRST(&vm->vm_freetags);
324 LIST_REMOVE(bt, bt_freelist);
326 LIST_INSERT_HEAD(&freetags, bt, bt_freelist);
329 while ((bt = LIST_FIRST(&freetags)) != NULL) {
330 LIST_REMOVE(bt, bt_freelist);
331 uma_zfree(vmem_bt_zone, bt);
336 bt_free(vmem_t *vm, bt_t *bt)
339 VMEM_ASSERT_LOCKED(vm);
340 MPASS(LIST_FIRST(&vm->vm_freetags) != bt);
341 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
346 * freelist[0] ... [1, 1]
347 * freelist[1] ... [2, 2]
349 * freelist[29] ... [30, 30]
350 * freelist[30] ... [31, 31]
351 * freelist[31] ... [32, 63]
352 * freelist[33] ... [64, 127]
354 * freelist[n] ... [(1 << (n - 26)), (1 << (n - 25)) - 1]
358 static struct vmem_freelist *
359 bt_freehead_tofree(vmem_t *vm, vmem_size_t size)
361 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
362 const int idx = SIZE2ORDER(qsize);
364 MPASS(size != 0 && qsize != 0);
365 MPASS((size & vm->vm_quantum_mask) == 0);
367 MPASS(idx < VMEM_MAXORDER);
369 return &vm->vm_freelist[idx];
373 * bt_freehead_toalloc: return the freelist for the given size and allocation
376 * For M_FIRSTFIT, return the list in which any blocks are large enough
377 * for the requested size. otherwise, return the list which can have blocks
378 * large enough for the requested size.
380 static struct vmem_freelist *
381 bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, int strat)
383 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
384 int idx = SIZE2ORDER(qsize);
386 MPASS(size != 0 && qsize != 0);
387 MPASS((size & vm->vm_quantum_mask) == 0);
389 if (strat == M_FIRSTFIT && ORDER2SIZE(idx) != qsize) {
391 /* check too large request? */
394 MPASS(idx < VMEM_MAXORDER);
396 return &vm->vm_freelist[idx];
399 /* ---- boundary tag hash */
401 static struct vmem_hashlist *
402 bt_hashhead(vmem_t *vm, vmem_addr_t addr)
404 struct vmem_hashlist *list;
407 hash = hash32_buf(&addr, sizeof(addr), 0);
408 list = &vm->vm_hashlist[hash % vm->vm_hashsize];
414 bt_lookupbusy(vmem_t *vm, vmem_addr_t addr)
416 struct vmem_hashlist *list;
419 VMEM_ASSERT_LOCKED(vm);
420 list = bt_hashhead(vm, addr);
421 LIST_FOREACH(bt, list, bt_hashlist) {
422 if (bt->bt_start == addr) {
431 bt_rembusy(vmem_t *vm, bt_t *bt)
434 VMEM_ASSERT_LOCKED(vm);
435 MPASS(vm->vm_nbusytag > 0);
436 vm->vm_inuse -= bt->bt_size;
438 LIST_REMOVE(bt, bt_hashlist);
442 bt_insbusy(vmem_t *vm, bt_t *bt)
444 struct vmem_hashlist *list;
446 VMEM_ASSERT_LOCKED(vm);
447 MPASS(bt->bt_type == BT_TYPE_BUSY);
449 list = bt_hashhead(vm, bt->bt_start);
450 LIST_INSERT_HEAD(list, bt, bt_hashlist);
452 vm->vm_inuse += bt->bt_size;
455 /* ---- boundary tag list */
458 bt_remseg(vmem_t *vm, bt_t *bt)
461 TAILQ_REMOVE(&vm->vm_seglist, bt, bt_seglist);
466 bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev)
469 TAILQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist);
473 bt_insseg_tail(vmem_t *vm, bt_t *bt)
476 TAILQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist);
480 bt_remfree(vmem_t *vm, bt_t *bt)
483 MPASS(bt->bt_type == BT_TYPE_FREE);
485 LIST_REMOVE(bt, bt_freelist);
489 bt_insfree(vmem_t *vm, bt_t *bt)
491 struct vmem_freelist *list;
493 list = bt_freehead_tofree(vm, bt->bt_size);
494 LIST_INSERT_HEAD(list, bt, bt_freelist);
497 /* ---- vmem internal functions */
500 * Import from the arena into the quantum cache in UMA.
503 qc_import(void *arg, void **store, int cnt, int domain, int flags)
510 if ((flags & VMEM_FITMASK) == 0)
512 for (i = 0; i < cnt; i++) {
513 if (vmem_xalloc(qc->qc_vmem, qc->qc_size, 0, 0, 0,
514 VMEM_ADDR_MIN, VMEM_ADDR_MAX, flags, &addr) != 0)
516 store[i] = (void *)addr;
517 /* Only guarantee one allocation. */
525 * Release memory from the UMA cache to the arena.
528 qc_release(void *arg, void **store, int cnt)
534 for (i = 0; i < cnt; i++)
535 vmem_xfree(qc->qc_vmem, (vmem_addr_t)store[i], qc->qc_size);
539 qc_init(vmem_t *vm, vmem_size_t qcache_max)
546 MPASS((qcache_max & vm->vm_quantum_mask) == 0);
547 qcache_idx_max = MIN(qcache_max >> vm->vm_quantum_shift,
548 VMEM_QCACHE_IDX_MAX);
549 vm->vm_qcache_max = qcache_idx_max << vm->vm_quantum_shift;
550 for (i = 0; i < qcache_idx_max; i++) {
551 qc = &vm->vm_qcache[i];
552 size = (i + 1) << vm->vm_quantum_shift;
553 snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu",
557 qc->qc_cache = uma_zcache_create(qc->qc_name, size,
558 NULL, NULL, NULL, NULL, qc_import, qc_release, qc,
565 qc_destroy(vmem_t *vm)
570 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
571 for (i = 0; i < qcache_idx_max; i++)
572 uma_zdestroy(vm->vm_qcache[i].qc_cache);
581 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
582 for (i = 0; i < qcache_idx_max; i++)
583 zone_drain(vm->vm_qcache[i].qc_cache);
586 #ifndef UMA_MD_SMALL_ALLOC
588 static struct mtx_padalign __exclusive_cache_line vmem_bt_lock;
591 * vmem_bt_alloc: Allocate a new page of boundary tags.
593 * On architectures with uma_small_alloc there is no recursion; no address
594 * space need be allocated to allocate boundary tags. For the others, we
595 * must handle recursion. Boundary tags are necessary to allocate new
598 * UMA guarantees that enough tags are held in reserve to allocate a new
599 * page of kva. We dip into this reserve by specifying M_USE_RESERVE only
600 * when allocating the page to hold new boundary tags. In this way the
601 * reserve is automatically filled by the allocation that uses the reserve.
603 * We still have to guarantee that the new tags are allocated atomically since
604 * many threads may try concurrently. The bt_lock provides this guarantee.
605 * We convert WAITOK allocations to NOWAIT and then handle the blocking here
606 * on failure. It's ok to return NULL for a WAITOK allocation as UMA will
607 * loop again after checking to see if we lost the race to allocate.
609 * There is a small race between vmem_bt_alloc() returning the page and the
610 * zone lock being acquired to add the page to the zone. For WAITOK
611 * allocations we just pause briefly. NOWAIT may experience a transient
612 * failure. To alleviate this we permit a small number of simultaneous
613 * fills to proceed concurrently so NOWAIT is less likely to fail unless
614 * we are really out of KVA.
617 vmem_bt_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
622 *pflag = UMA_SLAB_KERNEL;
625 * Single thread boundary tag allocation so that the address space
626 * and memory are added in one atomic operation.
628 mtx_lock(&vmem_bt_lock);
629 if (vmem_xalloc(vm_dom[domain].vmd_kernel_arena, bytes, 0, 0, 0,
630 VMEM_ADDR_MIN, VMEM_ADDR_MAX,
631 M_NOWAIT | M_NOVM | M_USE_RESERVE | M_BESTFIT, &addr) == 0) {
632 if (kmem_back_domain(domain, kernel_object, addr, bytes,
633 M_NOWAIT | M_USE_RESERVE) == 0) {
634 mtx_unlock(&vmem_bt_lock);
635 return ((void *)addr);
637 vmem_xfree(vm_dom[domain].vmd_kernel_arena, addr, bytes);
638 mtx_unlock(&vmem_bt_lock);
640 * Out of memory, not address space. This may not even be
641 * possible due to M_USE_RESERVE page allocation.
647 mtx_unlock(&vmem_bt_lock);
649 * We're either out of address space or lost a fill race.
662 mtx_init(&vmem_list_lock, "vmem list lock", NULL, MTX_DEF);
663 vmem_zone = uma_zcreate("vmem",
664 sizeof(struct vmem), NULL, NULL, NULL, NULL,
665 UMA_ALIGN_PTR, UMA_ZONE_VM);
666 vmem_bt_zone = uma_zcreate("vmem btag",
667 sizeof(struct vmem_btag), NULL, NULL, NULL, NULL,
668 UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE);
669 #ifndef UMA_MD_SMALL_ALLOC
670 mtx_init(&vmem_bt_lock, "btag lock", NULL, MTX_DEF);
671 uma_prealloc(vmem_bt_zone, BT_MAXALLOC);
673 * Reserve enough tags to allocate new tags. We allow multiple
674 * CPUs to attempt to allocate new tags concurrently to limit
675 * false restarts in UMA.
677 uma_zone_reserve(vmem_bt_zone, BT_MAXALLOC * (mp_ncpus + 1) / 2);
678 uma_zone_set_allocf(vmem_bt_zone, vmem_bt_alloc);
685 vmem_rehash(vmem_t *vm, vmem_size_t newhashsize)
689 struct vmem_hashlist *newhashlist;
690 struct vmem_hashlist *oldhashlist;
691 vmem_size_t oldhashsize;
693 MPASS(newhashsize > 0);
695 newhashlist = malloc(sizeof(struct vmem_hashlist) * newhashsize,
697 if (newhashlist == NULL)
699 for (i = 0; i < newhashsize; i++) {
700 LIST_INIT(&newhashlist[i]);
704 oldhashlist = vm->vm_hashlist;
705 oldhashsize = vm->vm_hashsize;
706 vm->vm_hashlist = newhashlist;
707 vm->vm_hashsize = newhashsize;
708 if (oldhashlist == NULL) {
712 for (i = 0; i < oldhashsize; i++) {
713 while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) {
720 if (oldhashlist != vm->vm_hash0) {
721 free(oldhashlist, M_VMEM);
728 vmem_periodic_kick(void *dummy)
731 taskqueue_enqueue(taskqueue_thread, &vmem_periodic_wk);
735 vmem_periodic(void *unused, int pending)
741 mtx_lock(&vmem_list_lock);
742 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
744 /* Convenient time to verify vmem state. */
745 if (enable_vmem_check == 1) {
751 desired = 1 << flsl(vm->vm_nbusytag);
752 desired = MIN(MAX(desired, VMEM_HASHSIZE_MIN),
754 current = vm->vm_hashsize;
756 /* Grow in powers of two. Shrink less aggressively. */
757 if (desired >= current * 2 || desired * 4 <= current)
758 vmem_rehash(vm, desired);
761 * Periodically wake up threads waiting for resources,
762 * so they could ask for reclamation again.
764 VMEM_CONDVAR_BROADCAST(vm);
766 mtx_unlock(&vmem_list_lock);
768 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
769 vmem_periodic_kick, NULL);
773 vmem_start_callout(void *unused)
776 TASK_INIT(&vmem_periodic_wk, 0, vmem_periodic, NULL);
777 vmem_periodic_interval = hz * 10;
778 callout_init(&vmem_periodic_ch, 1);
779 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
780 vmem_periodic_kick, NULL);
782 SYSINIT(vfs, SI_SUB_CONFIGURE, SI_ORDER_ANY, vmem_start_callout, NULL);
785 vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int type)
790 MPASS(type == BT_TYPE_SPAN || type == BT_TYPE_SPAN_STATIC);
791 MPASS((size & vm->vm_quantum_mask) == 0);
793 btspan = bt_alloc(vm);
794 btspan->bt_type = type;
795 btspan->bt_start = addr;
796 btspan->bt_size = size;
797 bt_insseg_tail(vm, btspan);
799 btfree = bt_alloc(vm);
800 btfree->bt_type = BT_TYPE_FREE;
801 btfree->bt_start = addr;
802 btfree->bt_size = size;
803 bt_insseg(vm, btfree, btspan);
804 bt_insfree(vm, btfree);
810 vmem_destroy1(vmem_t *vm)
815 * Drain per-cpu quantum caches.
820 * The vmem should now only contain empty segments.
823 MPASS(vm->vm_nbusytag == 0);
825 while ((bt = TAILQ_FIRST(&vm->vm_seglist)) != NULL)
828 if (vm->vm_hashlist != NULL && vm->vm_hashlist != vm->vm_hash0)
829 free(vm->vm_hashlist, M_VMEM);
833 VMEM_CONDVAR_DESTROY(vm);
834 VMEM_LOCK_DESTROY(vm);
835 uma_zfree(vmem_zone, vm);
839 vmem_import(vmem_t *vm, vmem_size_t size, vmem_size_t align, int flags)
844 if (vm->vm_importfn == NULL)
848 * To make sure we get a span that meets the alignment we double it
849 * and add the size to the tail. This slightly overestimates.
851 if (align != vm->vm_quantum_mask + 1)
852 size = (align * 2) + size;
853 size = roundup(size, vm->vm_import_quantum);
855 if (vm->vm_limit != 0 && vm->vm_limit < vm->vm_size + size)
859 * Hide MAXALLOC tags so we're guaranteed to be able to add this
860 * span and the tag we want to allocate from it.
862 MPASS(vm->vm_nfreetags >= BT_MAXALLOC);
863 vm->vm_nfreetags -= BT_MAXALLOC;
865 error = (vm->vm_importfn)(vm->vm_arg, size, flags, &addr);
867 vm->vm_nfreetags += BT_MAXALLOC;
871 vmem_add1(vm, addr, size, BT_TYPE_SPAN);
877 * vmem_fit: check if a bt can satisfy the given restrictions.
879 * it's a caller's responsibility to ensure the region is big enough
883 vmem_fit(const bt_t *bt, vmem_size_t size, vmem_size_t align,
884 vmem_size_t phase, vmem_size_t nocross, vmem_addr_t minaddr,
885 vmem_addr_t maxaddr, vmem_addr_t *addrp)
891 MPASS(bt->bt_size >= size); /* caller's responsibility */
894 * XXX assumption: vmem_addr_t and vmem_size_t are
895 * unsigned integer of the same size.
898 start = bt->bt_start;
899 if (start < minaddr) {
908 start = VMEM_ALIGNUP(start - phase, align) + phase;
909 if (start < bt->bt_start)
911 if (VMEM_CROSS_P(start, start + size - 1, nocross)) {
912 MPASS(align < nocross);
913 start = VMEM_ALIGNUP(start - phase, nocross) + phase;
915 if (start <= end && end - start >= size - 1) {
916 MPASS((start & (align - 1)) == phase);
917 MPASS(!VMEM_CROSS_P(start, start + size - 1, nocross));
918 MPASS(minaddr <= start);
919 MPASS(maxaddr == 0 || start + size - 1 <= maxaddr);
920 MPASS(bt->bt_start <= start);
921 MPASS(BT_END(bt) - start >= size - 1);
930 * vmem_clip: Trim the boundary tag edges to the requested start and size.
933 vmem_clip(vmem_t *vm, bt_t *bt, vmem_addr_t start, vmem_size_t size)
938 VMEM_ASSERT_LOCKED(vm);
939 MPASS(bt->bt_type == BT_TYPE_FREE);
940 MPASS(bt->bt_size >= size);
942 if (bt->bt_start != start) {
943 btprev = bt_alloc(vm);
944 btprev->bt_type = BT_TYPE_FREE;
945 btprev->bt_start = bt->bt_start;
946 btprev->bt_size = start - bt->bt_start;
947 bt->bt_start = start;
948 bt->bt_size -= btprev->bt_size;
949 bt_insfree(vm, btprev);
950 bt_insseg(vm, btprev,
951 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
953 MPASS(bt->bt_start == start);
954 if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) {
956 btnew = bt_alloc(vm);
957 btnew->bt_type = BT_TYPE_BUSY;
958 btnew->bt_start = bt->bt_start;
959 btnew->bt_size = size;
960 bt->bt_start = bt->bt_start + size;
964 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
965 bt_insbusy(vm, btnew);
968 bt->bt_type = BT_TYPE_BUSY;
971 MPASS(bt->bt_size >= size);
972 bt->bt_type = BT_TYPE_BUSY;
978 vmem_set_import(vmem_t *vm, vmem_import_t *importfn,
979 vmem_release_t *releasefn, void *arg, vmem_size_t import_quantum)
983 vm->vm_importfn = importfn;
984 vm->vm_releasefn = releasefn;
986 vm->vm_import_quantum = import_quantum;
991 vmem_set_limit(vmem_t *vm, vmem_size_t limit)
995 vm->vm_limit = limit;
1000 vmem_set_reclaim(vmem_t *vm, vmem_reclaim_t *reclaimfn)
1004 vm->vm_reclaimfn = reclaimfn;
1009 * vmem_init: Initializes vmem arena.
1012 vmem_init(vmem_t *vm, const char *name, vmem_addr_t base, vmem_size_t size,
1013 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1018 MPASS((quantum & (quantum - 1)) == 0);
1020 bzero(vm, sizeof(*vm));
1022 VMEM_CONDVAR_INIT(vm, name);
1023 VMEM_LOCK_INIT(vm, name);
1024 vm->vm_nfreetags = 0;
1025 LIST_INIT(&vm->vm_freetags);
1026 strlcpy(vm->vm_name, name, sizeof(vm->vm_name));
1027 vm->vm_quantum_mask = quantum - 1;
1028 vm->vm_quantum_shift = flsl(quantum) - 1;
1029 vm->vm_nbusytag = 0;
1033 qc_init(vm, qcache_max);
1035 TAILQ_INIT(&vm->vm_seglist);
1036 for (i = 0; i < VMEM_MAXORDER; i++) {
1037 LIST_INIT(&vm->vm_freelist[i]);
1039 memset(&vm->vm_hash0, 0, sizeof(vm->vm_hash0));
1040 vm->vm_hashsize = VMEM_HASHSIZE_MIN;
1041 vm->vm_hashlist = vm->vm_hash0;
1044 if (vmem_add(vm, base, size, flags) != 0) {
1050 mtx_lock(&vmem_list_lock);
1051 LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist);
1052 mtx_unlock(&vmem_list_lock);
1058 * vmem_create: create an arena.
1061 vmem_create(const char *name, vmem_addr_t base, vmem_size_t size,
1062 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1067 vm = uma_zalloc(vmem_zone, flags & (M_WAITOK|M_NOWAIT));
1070 if (vmem_init(vm, name, base, size, quantum, qcache_max,
1077 vmem_destroy(vmem_t *vm)
1080 mtx_lock(&vmem_list_lock);
1081 LIST_REMOVE(vm, vm_alllist);
1082 mtx_unlock(&vmem_list_lock);
1088 vmem_roundup_size(vmem_t *vm, vmem_size_t size)
1091 return (size + vm->vm_quantum_mask) & ~vm->vm_quantum_mask;
1095 * vmem_alloc: allocate resource from the arena.
1098 vmem_alloc(vmem_t *vm, vmem_size_t size, int flags, vmem_addr_t *addrp)
1100 const int strat __unused = flags & VMEM_FITMASK;
1103 flags &= VMEM_FLAGS;
1105 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT);
1106 if ((flags & M_NOWAIT) == 0)
1107 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_alloc");
1109 if (size <= vm->vm_qcache_max) {
1110 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1111 *addrp = (vmem_addr_t)uma_zalloc(qc->qc_cache, flags);
1117 return vmem_xalloc(vm, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
1122 vmem_xalloc(vmem_t *vm, const vmem_size_t size0, vmem_size_t align,
1123 const vmem_size_t phase, const vmem_size_t nocross,
1124 const vmem_addr_t minaddr, const vmem_addr_t maxaddr, int flags,
1127 const vmem_size_t size = vmem_roundup_size(vm, size0);
1128 struct vmem_freelist *list;
1129 struct vmem_freelist *first;
1130 struct vmem_freelist *end;
1136 flags &= VMEM_FLAGS;
1137 strat = flags & VMEM_FITMASK;
1140 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT);
1141 MPASS((flags & (M_NOWAIT|M_WAITOK)) != (M_NOWAIT|M_WAITOK));
1142 if ((flags & M_NOWAIT) == 0)
1143 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_xalloc");
1144 MPASS((align & vm->vm_quantum_mask) == 0);
1145 MPASS((align & (align - 1)) == 0);
1146 MPASS((phase & vm->vm_quantum_mask) == 0);
1147 MPASS((nocross & vm->vm_quantum_mask) == 0);
1148 MPASS((nocross & (nocross - 1)) == 0);
1149 MPASS((align == 0 && phase == 0) || phase < align);
1150 MPASS(nocross == 0 || nocross >= size);
1151 MPASS(minaddr <= maxaddr);
1152 MPASS(!VMEM_CROSS_P(phase, phase + size - 1, nocross));
1155 align = vm->vm_quantum_mask + 1;
1158 end = &vm->vm_freelist[VMEM_MAXORDER];
1160 * choose a free block from which we allocate.
1162 first = bt_freehead_toalloc(vm, size, strat);
1166 * Make sure we have enough tags to complete the
1169 if (vm->vm_nfreetags < BT_MAXALLOC &&
1170 bt_fill(vm, flags) != 0) {
1175 * Scan freelists looking for a tag that satisfies the
1176 * allocation. If we're doing BESTFIT we may encounter
1177 * sizes below the request. If we're doing FIRSTFIT we
1178 * inspect only the first element from each list.
1180 for (list = first; list < end; list++) {
1181 LIST_FOREACH(bt, list, bt_freelist) {
1182 if (bt->bt_size >= size) {
1183 error = vmem_fit(bt, size, align, phase,
1184 nocross, minaddr, maxaddr, addrp);
1186 vmem_clip(vm, bt, *addrp, size);
1190 /* FIRST skips to the next list. */
1191 if (strat == M_FIRSTFIT)
1196 * Retry if the fast algorithm failed.
1198 if (strat == M_FIRSTFIT) {
1200 first = bt_freehead_toalloc(vm, size, strat);
1204 * XXX it is possible to fail to meet restrictions with the
1205 * imported region. It is up to the user to specify the
1206 * import quantum such that it can satisfy any allocation.
1208 if (vmem_import(vm, size, align, flags) == 0)
1212 * Try to free some space from the quantum cache or reclaim
1213 * functions if available.
1215 if (vm->vm_qcache_max != 0 || vm->vm_reclaimfn != NULL) {
1216 avail = vm->vm_size - vm->vm_inuse;
1218 if (vm->vm_qcache_max != 0)
1220 if (vm->vm_reclaimfn != NULL)
1221 vm->vm_reclaimfn(vm, flags);
1223 /* If we were successful retry even NOWAIT. */
1224 if (vm->vm_size - vm->vm_inuse > avail)
1227 if ((flags & M_NOWAIT) != 0) {
1231 VMEM_CONDVAR_WAIT(vm);
1235 if (error != 0 && (flags & M_NOWAIT) == 0)
1236 panic("failed to allocate waiting allocation\n");
1242 * vmem_free: free the resource to the arena.
1245 vmem_free(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1250 if (size <= vm->vm_qcache_max) {
1251 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1252 uma_zfree(qc->qc_cache, (void *)addr);
1254 vmem_xfree(vm, addr, size);
1258 vmem_xfree(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1266 bt = bt_lookupbusy(vm, addr);
1268 MPASS(bt->bt_start == addr);
1269 MPASS(bt->bt_size == vmem_roundup_size(vm, size) ||
1270 bt->bt_size - vmem_roundup_size(vm, size) <= vm->vm_quantum_mask);
1271 MPASS(bt->bt_type == BT_TYPE_BUSY);
1273 bt->bt_type = BT_TYPE_FREE;
1276 t = TAILQ_NEXT(bt, bt_seglist);
1277 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1278 MPASS(BT_END(bt) < t->bt_start); /* YYY */
1279 bt->bt_size += t->bt_size;
1283 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1284 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1285 MPASS(BT_END(t) < bt->bt_start); /* YYY */
1286 bt->bt_size += t->bt_size;
1287 bt->bt_start = t->bt_start;
1292 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1294 MPASS(BT_ISSPAN_P(t) || t->bt_type == BT_TYPE_BUSY);
1295 if (vm->vm_releasefn != NULL && t->bt_type == BT_TYPE_SPAN &&
1296 t->bt_size == bt->bt_size) {
1297 vmem_addr_t spanaddr;
1298 vmem_size_t spansize;
1300 MPASS(t->bt_start == bt->bt_start);
1301 spanaddr = bt->bt_start;
1302 spansize = bt->bt_size;
1305 vm->vm_size -= spansize;
1306 VMEM_CONDVAR_BROADCAST(vm);
1307 bt_freetrim(vm, BT_MAXFREE);
1308 (*vm->vm_releasefn)(vm->vm_arg, spanaddr, spansize);
1311 VMEM_CONDVAR_BROADCAST(vm);
1312 bt_freetrim(vm, BT_MAXFREE);
1321 vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int flags)
1326 flags &= VMEM_FLAGS;
1328 if (vm->vm_nfreetags >= BT_MAXALLOC || bt_fill(vm, flags) == 0)
1329 vmem_add1(vm, addr, size, BT_TYPE_SPAN_STATIC);
1338 * vmem_size: information about arenas size
1341 vmem_size(vmem_t *vm, int typemask)
1347 return vm->vm_inuse;
1349 return vm->vm_size - vm->vm_inuse;
1350 case VMEM_FREE|VMEM_ALLOC:
1354 for (i = VMEM_MAXORDER - 1; i >= 0; i--) {
1355 if (LIST_EMPTY(&vm->vm_freelist[i]))
1358 return ((vmem_size_t)ORDER2SIZE(i) <<
1359 vm->vm_quantum_shift);
1370 #if defined(DDB) || defined(DIAGNOSTIC)
1372 static void bt_dump(const bt_t *, int (*)(const char *, ...)
1373 __printflike(1, 2));
1376 bt_type_string(int type)
1386 case BT_TYPE_SPAN_STATIC:
1387 return "static span";
1395 bt_dump(const bt_t *bt, int (*pr)(const char *, ...))
1398 (*pr)("\t%p: %jx %jx, %d(%s)\n",
1399 bt, (intmax_t)bt->bt_start, (intmax_t)bt->bt_size,
1400 bt->bt_type, bt_type_string(bt->bt_type));
1404 vmem_dump(const vmem_t *vm , int (*pr)(const char *, ...) __printflike(1, 2))
1409 (*pr)("vmem %p '%s'\n", vm, vm->vm_name);
1410 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1414 for (i = 0; i < VMEM_MAXORDER; i++) {
1415 const struct vmem_freelist *fl = &vm->vm_freelist[i];
1417 if (LIST_EMPTY(fl)) {
1421 (*pr)("freelist[%d]\n", i);
1422 LIST_FOREACH(bt, fl, bt_freelist) {
1428 #endif /* defined(DDB) || defined(DIAGNOSTIC) */
1431 #include <ddb/ddb.h>
1434 vmem_whatis_lookup(vmem_t *vm, vmem_addr_t addr)
1438 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1439 if (BT_ISSPAN_P(bt)) {
1442 if (bt->bt_start <= addr && addr <= BT_END(bt)) {
1451 vmem_whatis(vmem_addr_t addr, int (*pr)(const char *, ...))
1455 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1458 bt = vmem_whatis_lookup(vm, addr);
1462 (*pr)("%p is %p+%zu in VMEM '%s' (%s)\n",
1463 (void *)addr, (void *)bt->bt_start,
1464 (vmem_size_t)(addr - bt->bt_start), vm->vm_name,
1465 (bt->bt_type == BT_TYPE_BUSY) ? "allocated" : "free");
1470 vmem_printall(const char *modif, int (*pr)(const char *, ...))
1474 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1480 vmem_print(vmem_addr_t addr, const char *modif, int (*pr)(const char *, ...))
1482 const vmem_t *vm = (const void *)addr;
1487 DB_SHOW_COMMAND(vmemdump, vmemdump)
1491 db_printf("usage: show vmemdump <addr>\n");
1495 vmem_dump((const vmem_t *)addr, db_printf);
1498 DB_SHOW_ALL_COMMAND(vmemdump, vmemdumpall)
1502 LIST_FOREACH(vm, &vmem_list, vm_alllist)
1503 vmem_dump(vm, db_printf);
1506 DB_SHOW_COMMAND(vmem, vmem_summ)
1508 const vmem_t *vm = (const void *)addr;
1510 size_t ft[VMEM_MAXORDER], ut[VMEM_MAXORDER];
1511 size_t fs[VMEM_MAXORDER], us[VMEM_MAXORDER];
1515 db_printf("usage: show vmem <addr>\n");
1519 db_printf("vmem %p '%s'\n", vm, vm->vm_name);
1520 db_printf("\tquantum:\t%zu\n", vm->vm_quantum_mask + 1);
1521 db_printf("\tsize:\t%zu\n", vm->vm_size);
1522 db_printf("\tinuse:\t%zu\n", vm->vm_inuse);
1523 db_printf("\tfree:\t%zu\n", vm->vm_size - vm->vm_inuse);
1524 db_printf("\tbusy tags:\t%d\n", vm->vm_nbusytag);
1525 db_printf("\tfree tags:\t%d\n", vm->vm_nfreetags);
1527 memset(&ft, 0, sizeof(ft));
1528 memset(&ut, 0, sizeof(ut));
1529 memset(&fs, 0, sizeof(fs));
1530 memset(&us, 0, sizeof(us));
1531 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1532 ord = SIZE2ORDER(bt->bt_size >> vm->vm_quantum_shift);
1533 if (bt->bt_type == BT_TYPE_BUSY) {
1535 us[ord] += bt->bt_size;
1536 } else if (bt->bt_type == BT_TYPE_FREE) {
1538 fs[ord] += bt->bt_size;
1541 db_printf("\t\t\tinuse\tsize\t\tfree\tsize\n");
1542 for (ord = 0; ord < VMEM_MAXORDER; ord++) {
1543 if (ut[ord] == 0 && ft[ord] == 0)
1545 db_printf("\t%-15zu %zu\t%-15zu %zu\t%-16zu\n",
1546 ORDER2SIZE(ord) << vm->vm_quantum_shift,
1547 ut[ord], us[ord], ft[ord], fs[ord]);
1551 DB_SHOW_ALL_COMMAND(vmem, vmem_summall)
1555 LIST_FOREACH(vm, &vmem_list, vm_alllist)
1556 vmem_summ((db_expr_t)vm, TRUE, count, modif);
1558 #endif /* defined(DDB) */
1560 #define vmem_printf printf
1562 #if defined(DIAGNOSTIC)
1565 vmem_check_sanity(vmem_t *vm)
1567 const bt_t *bt, *bt2;
1571 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1572 if (bt->bt_start > BT_END(bt)) {
1573 printf("corrupted tag\n");
1574 bt_dump(bt, vmem_printf);
1578 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1579 TAILQ_FOREACH(bt2, &vm->vm_seglist, bt_seglist) {
1583 if (BT_ISSPAN_P(bt) != BT_ISSPAN_P(bt2)) {
1586 if (bt->bt_start <= BT_END(bt2) &&
1587 bt2->bt_start <= BT_END(bt)) {
1588 printf("overwrapped tags\n");
1589 bt_dump(bt, vmem_printf);
1590 bt_dump(bt2, vmem_printf);
1600 vmem_check(vmem_t *vm)
1603 if (!vmem_check_sanity(vm)) {
1604 panic("insanity vmem %p", vm);
1608 #endif /* defined(DIAGNOSTIC) */