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>
62 #include <sys/vmmeter.h>
69 #include <vm/vm_map.h>
70 #include <vm/vm_object.h>
71 #include <vm/vm_kern.h>
72 #include <vm/vm_extern.h>
73 #include <vm/vm_param.h>
74 #include <vm/vm_page.h>
75 #include <vm/vm_pageout.h>
76 #include <vm/vm_phys.h>
77 #include <vm/vm_pagequeue.h>
78 #include <vm/uma_int.h>
80 int vmem_startup_count(void);
82 #define VMEM_OPTORDER 5
83 #define VMEM_OPTVALUE (1 << VMEM_OPTORDER)
84 #define VMEM_MAXORDER \
85 (VMEM_OPTVALUE - 1 + sizeof(vmem_size_t) * NBBY - VMEM_OPTORDER)
87 #define VMEM_HASHSIZE_MIN 16
88 #define VMEM_HASHSIZE_MAX 131072
90 #define VMEM_QCACHE_IDX_MAX 16
92 #define VMEM_FITMASK (M_BESTFIT | M_FIRSTFIT)
95 (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM | M_BESTFIT | M_FIRSTFIT)
97 #define BT_FLAGS (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM)
99 #define QC_NAME_MAX 16
102 * Data structures private to vmem.
104 MALLOC_DEFINE(M_VMEM, "vmem", "vmem internal structures");
106 typedef struct vmem_btag bt_t;
108 TAILQ_HEAD(vmem_seglist, vmem_btag);
109 LIST_HEAD(vmem_freelist, vmem_btag);
110 LIST_HEAD(vmem_hashlist, vmem_btag);
116 char qc_name[QC_NAME_MAX];
118 typedef struct qcache qcache_t;
119 #define QC_POOL_TO_QCACHE(pool) ((qcache_t *)(pool->pr_qcache))
121 #define VMEM_NAME_MAX 16
125 struct mtx_padalign vm_lock;
127 char vm_name[VMEM_NAME_MAX+1];
128 LIST_ENTRY(vmem) vm_alllist;
129 struct vmem_hashlist vm_hash0[VMEM_HASHSIZE_MIN];
130 struct vmem_freelist vm_freelist[VMEM_MAXORDER];
131 struct vmem_seglist vm_seglist;
132 struct vmem_hashlist *vm_hashlist;
133 vmem_size_t vm_hashsize;
135 /* Constant after init */
136 vmem_size_t vm_qcache_max;
137 vmem_size_t vm_quantum_mask;
138 vmem_size_t vm_import_quantum;
139 int vm_quantum_shift;
141 /* Written on alloc/free */
142 LIST_HEAD(, vmem_btag) vm_freetags;
145 vmem_size_t vm_inuse;
147 vmem_size_t vm_limit;
149 /* Used on import. */
150 vmem_import_t *vm_importfn;
151 vmem_release_t *vm_releasefn;
154 /* Space exhaustion callback. */
155 vmem_reclaim_t *vm_reclaimfn;
158 qcache_t vm_qcache[VMEM_QCACHE_IDX_MAX];
163 TAILQ_ENTRY(vmem_btag) bt_seglist;
165 LIST_ENTRY(vmem_btag) u_freelist; /* BT_TYPE_FREE */
166 LIST_ENTRY(vmem_btag) u_hashlist; /* BT_TYPE_BUSY */
168 #define bt_hashlist bt_u.u_hashlist
169 #define bt_freelist bt_u.u_freelist
170 vmem_addr_t bt_start;
175 #define BT_TYPE_SPAN 1 /* Allocated from importfn */
176 #define BT_TYPE_SPAN_STATIC 2 /* vmem_add() or create. */
177 #define BT_TYPE_FREE 3 /* Available space. */
178 #define BT_TYPE_BUSY 4 /* Used space. */
179 #define BT_ISSPAN_P(bt) ((bt)->bt_type <= BT_TYPE_SPAN_STATIC)
181 #define BT_END(bt) ((bt)->bt_start + (bt)->bt_size - 1)
183 #if defined(DIAGNOSTIC)
184 static int enable_vmem_check = 1;
185 SYSCTL_INT(_debug, OID_AUTO, vmem_check, CTLFLAG_RWTUN,
186 &enable_vmem_check, 0, "Enable vmem check");
187 static void vmem_check(vmem_t *);
190 static struct callout vmem_periodic_ch;
191 static int vmem_periodic_interval;
192 static struct task vmem_periodic_wk;
194 static struct mtx_padalign __exclusive_cache_line vmem_list_lock;
195 static LIST_HEAD(, vmem) vmem_list = LIST_HEAD_INITIALIZER(vmem_list);
196 static uma_zone_t vmem_zone;
199 #define VMEM_CONDVAR_INIT(vm, wchan) cv_init(&vm->vm_cv, wchan)
200 #define VMEM_CONDVAR_DESTROY(vm) cv_destroy(&vm->vm_cv)
201 #define VMEM_CONDVAR_WAIT(vm) cv_wait(&vm->vm_cv, &vm->vm_lock)
202 #define VMEM_CONDVAR_BROADCAST(vm) cv_broadcast(&vm->vm_cv)
205 #define VMEM_LOCK(vm) mtx_lock(&vm->vm_lock)
206 #define VMEM_TRYLOCK(vm) mtx_trylock(&vm->vm_lock)
207 #define VMEM_UNLOCK(vm) mtx_unlock(&vm->vm_lock)
208 #define VMEM_LOCK_INIT(vm, name) mtx_init(&vm->vm_lock, (name), NULL, MTX_DEF)
209 #define VMEM_LOCK_DESTROY(vm) mtx_destroy(&vm->vm_lock)
210 #define VMEM_ASSERT_LOCKED(vm) mtx_assert(&vm->vm_lock, MA_OWNED);
212 #define VMEM_ALIGNUP(addr, align) (-(-(addr) & -(align)))
214 #define VMEM_CROSS_P(addr1, addr2, boundary) \
215 ((((addr1) ^ (addr2)) & -(boundary)) != 0)
217 #define ORDER2SIZE(order) ((order) < VMEM_OPTVALUE ? ((order) + 1) : \
218 (vmem_size_t)1 << ((order) - (VMEM_OPTVALUE - VMEM_OPTORDER - 1)))
219 #define SIZE2ORDER(size) ((size) <= VMEM_OPTVALUE ? ((size) - 1) : \
220 (flsl(size) + (VMEM_OPTVALUE - VMEM_OPTORDER - 2)))
223 * Maximum number of boundary tags that may be required to satisfy an
224 * allocation. Two may be required to import. Another two may be
225 * required to clip edges.
227 #define BT_MAXALLOC 4
230 * Max free limits the number of locally cached boundary tags. We
231 * just want to avoid hitting the zone allocator for every call.
233 #define BT_MAXFREE (BT_MAXALLOC * 8)
235 /* Allocator for boundary tags. */
236 static uma_zone_t vmem_bt_zone;
238 /* boot time arena storage. */
239 static struct vmem kernel_arena_storage;
240 static struct vmem buffer_arena_storage;
241 static struct vmem transient_arena_storage;
242 /* kernel and kmem arenas are aliased for backwards KPI compat. */
243 vmem_t *kernel_arena = &kernel_arena_storage;
244 vmem_t *kmem_arena = &kernel_arena_storage;
245 vmem_t *buffer_arena = &buffer_arena_storage;
246 vmem_t *transient_arena = &transient_arena_storage;
248 #ifdef DEBUG_MEMGUARD
249 static struct vmem memguard_arena_storage;
250 vmem_t *memguard_arena = &memguard_arena_storage;
254 * Fill the vmem's boundary tag cache. We guarantee that boundary tag
255 * allocation will not fail once bt_fill() passes. To do so we cache
256 * at least the maximum possible tag allocations in the arena.
259 bt_fill(vmem_t *vm, int flags)
263 VMEM_ASSERT_LOCKED(vm);
266 * Only allow the kernel arena and arenas derived from kernel arena to
267 * dip into reserve tags. They are where new tags come from.
270 if (vm != kernel_arena && vm->vm_arg != kernel_arena)
271 flags &= ~M_USE_RESERVE;
274 * Loop until we meet the reserve. To minimize the lock shuffle
275 * and prevent simultaneous fills we first try a NOWAIT regardless
276 * of the caller's flags. Specify M_NOVM so we don't recurse while
277 * holding a vmem lock.
279 while (vm->vm_nfreetags < BT_MAXALLOC) {
280 bt = uma_zalloc(vmem_bt_zone,
281 (flags & M_USE_RESERVE) | M_NOWAIT | M_NOVM);
284 bt = uma_zalloc(vmem_bt_zone, flags);
289 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
293 if (vm->vm_nfreetags < BT_MAXALLOC)
300 * Pop a tag off of the freetag stack.
307 VMEM_ASSERT_LOCKED(vm);
308 bt = LIST_FIRST(&vm->vm_freetags);
310 LIST_REMOVE(bt, bt_freelist);
317 * Trim the per-vmem free list. Returns with the lock released to
318 * avoid allocator recursions.
321 bt_freetrim(vmem_t *vm, int freelimit)
323 LIST_HEAD(, vmem_btag) freetags;
326 LIST_INIT(&freetags);
327 VMEM_ASSERT_LOCKED(vm);
328 while (vm->vm_nfreetags > freelimit) {
329 bt = LIST_FIRST(&vm->vm_freetags);
330 LIST_REMOVE(bt, bt_freelist);
332 LIST_INSERT_HEAD(&freetags, bt, bt_freelist);
335 while ((bt = LIST_FIRST(&freetags)) != NULL) {
336 LIST_REMOVE(bt, bt_freelist);
337 uma_zfree(vmem_bt_zone, bt);
342 bt_free(vmem_t *vm, bt_t *bt)
345 VMEM_ASSERT_LOCKED(vm);
346 MPASS(LIST_FIRST(&vm->vm_freetags) != bt);
347 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
352 * freelist[0] ... [1, 1]
353 * freelist[1] ... [2, 2]
355 * freelist[29] ... [30, 30]
356 * freelist[30] ... [31, 31]
357 * freelist[31] ... [32, 63]
358 * freelist[33] ... [64, 127]
360 * freelist[n] ... [(1 << (n - 26)), (1 << (n - 25)) - 1]
364 static struct vmem_freelist *
365 bt_freehead_tofree(vmem_t *vm, vmem_size_t size)
367 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
368 const int idx = SIZE2ORDER(qsize);
370 MPASS(size != 0 && qsize != 0);
371 MPASS((size & vm->vm_quantum_mask) == 0);
373 MPASS(idx < VMEM_MAXORDER);
375 return &vm->vm_freelist[idx];
379 * bt_freehead_toalloc: return the freelist for the given size and allocation
382 * For M_FIRSTFIT, return the list in which any blocks are large enough
383 * for the requested size. otherwise, return the list which can have blocks
384 * large enough for the requested size.
386 static struct vmem_freelist *
387 bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, int strat)
389 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
390 int idx = SIZE2ORDER(qsize);
392 MPASS(size != 0 && qsize != 0);
393 MPASS((size & vm->vm_quantum_mask) == 0);
395 if (strat == M_FIRSTFIT && ORDER2SIZE(idx) != qsize) {
397 /* check too large request? */
400 MPASS(idx < VMEM_MAXORDER);
402 return &vm->vm_freelist[idx];
405 /* ---- boundary tag hash */
407 static struct vmem_hashlist *
408 bt_hashhead(vmem_t *vm, vmem_addr_t addr)
410 struct vmem_hashlist *list;
413 hash = hash32_buf(&addr, sizeof(addr), 0);
414 list = &vm->vm_hashlist[hash % vm->vm_hashsize];
420 bt_lookupbusy(vmem_t *vm, vmem_addr_t addr)
422 struct vmem_hashlist *list;
425 VMEM_ASSERT_LOCKED(vm);
426 list = bt_hashhead(vm, addr);
427 LIST_FOREACH(bt, list, bt_hashlist) {
428 if (bt->bt_start == addr) {
437 bt_rembusy(vmem_t *vm, bt_t *bt)
440 VMEM_ASSERT_LOCKED(vm);
441 MPASS(vm->vm_nbusytag > 0);
442 vm->vm_inuse -= bt->bt_size;
444 LIST_REMOVE(bt, bt_hashlist);
448 bt_insbusy(vmem_t *vm, bt_t *bt)
450 struct vmem_hashlist *list;
452 VMEM_ASSERT_LOCKED(vm);
453 MPASS(bt->bt_type == BT_TYPE_BUSY);
455 list = bt_hashhead(vm, bt->bt_start);
456 LIST_INSERT_HEAD(list, bt, bt_hashlist);
458 vm->vm_inuse += bt->bt_size;
461 /* ---- boundary tag list */
464 bt_remseg(vmem_t *vm, bt_t *bt)
467 TAILQ_REMOVE(&vm->vm_seglist, bt, bt_seglist);
472 bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev)
475 TAILQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist);
479 bt_insseg_tail(vmem_t *vm, bt_t *bt)
482 TAILQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist);
486 bt_remfree(vmem_t *vm, bt_t *bt)
489 MPASS(bt->bt_type == BT_TYPE_FREE);
491 LIST_REMOVE(bt, bt_freelist);
495 bt_insfree(vmem_t *vm, bt_t *bt)
497 struct vmem_freelist *list;
499 list = bt_freehead_tofree(vm, bt->bt_size);
500 LIST_INSERT_HEAD(list, bt, bt_freelist);
503 /* ---- vmem internal functions */
506 * Import from the arena into the quantum cache in UMA.
508 * We use VMEM_ADDR_QCACHE_MIN instead of 0: uma_zalloc() returns 0 to indicate
509 * failure, so UMA can't be used to cache a resource with value 0.
512 qc_import(void *arg, void **store, int cnt, int domain, int flags)
518 KASSERT((flags & M_WAITOK) == 0, ("blocking allocation"));
521 for (i = 0; i < cnt; i++) {
522 if (vmem_xalloc(qc->qc_vmem, qc->qc_size, 0, 0, 0,
523 VMEM_ADDR_QCACHE_MIN, VMEM_ADDR_MAX, flags, &addr) != 0)
525 store[i] = (void *)addr;
531 * Release memory from the UMA cache to the arena.
534 qc_release(void *arg, void **store, int cnt)
540 for (i = 0; i < cnt; i++)
541 vmem_xfree(qc->qc_vmem, (vmem_addr_t)store[i], qc->qc_size);
545 qc_init(vmem_t *vm, vmem_size_t qcache_max)
552 MPASS((qcache_max & vm->vm_quantum_mask) == 0);
553 qcache_idx_max = MIN(qcache_max >> vm->vm_quantum_shift,
554 VMEM_QCACHE_IDX_MAX);
555 vm->vm_qcache_max = qcache_idx_max << vm->vm_quantum_shift;
556 for (i = 0; i < qcache_idx_max; i++) {
557 qc = &vm->vm_qcache[i];
558 size = (i + 1) << vm->vm_quantum_shift;
559 snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu",
563 qc->qc_cache = uma_zcache_create(qc->qc_name, size,
564 NULL, NULL, NULL, NULL, qc_import, qc_release, qc,
571 qc_destroy(vmem_t *vm)
576 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
577 for (i = 0; i < qcache_idx_max; i++)
578 uma_zdestroy(vm->vm_qcache[i].qc_cache);
587 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
588 for (i = 0; i < qcache_idx_max; i++)
589 zone_drain(vm->vm_qcache[i].qc_cache);
592 #ifndef UMA_MD_SMALL_ALLOC
594 static struct mtx_padalign __exclusive_cache_line vmem_bt_lock;
597 * vmem_bt_alloc: Allocate a new page of boundary tags.
599 * On architectures with uma_small_alloc there is no recursion; no address
600 * space need be allocated to allocate boundary tags. For the others, we
601 * must handle recursion. Boundary tags are necessary to allocate new
604 * UMA guarantees that enough tags are held in reserve to allocate a new
605 * page of kva. We dip into this reserve by specifying M_USE_RESERVE only
606 * when allocating the page to hold new boundary tags. In this way the
607 * reserve is automatically filled by the allocation that uses the reserve.
609 * We still have to guarantee that the new tags are allocated atomically since
610 * many threads may try concurrently. The bt_lock provides this guarantee.
611 * We convert WAITOK allocations to NOWAIT and then handle the blocking here
612 * on failure. It's ok to return NULL for a WAITOK allocation as UMA will
613 * loop again after checking to see if we lost the race to allocate.
615 * There is a small race between vmem_bt_alloc() returning the page and the
616 * zone lock being acquired to add the page to the zone. For WAITOK
617 * allocations we just pause briefly. NOWAIT may experience a transient
618 * failure. To alleviate this we permit a small number of simultaneous
619 * fills to proceed concurrently so NOWAIT is less likely to fail unless
620 * we are really out of KVA.
623 vmem_bt_alloc(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *pflag,
628 *pflag = UMA_SLAB_KERNEL;
631 * Single thread boundary tag allocation so that the address space
632 * and memory are added in one atomic operation.
634 mtx_lock(&vmem_bt_lock);
635 if (vmem_xalloc(vm_dom[domain].vmd_kernel_arena, bytes, 0, 0, 0,
636 VMEM_ADDR_MIN, VMEM_ADDR_MAX,
637 M_NOWAIT | M_NOVM | M_USE_RESERVE | M_BESTFIT, &addr) == 0) {
638 if (kmem_back_domain(domain, kernel_object, addr, bytes,
639 M_NOWAIT | M_USE_RESERVE) == 0) {
640 mtx_unlock(&vmem_bt_lock);
641 return ((void *)addr);
643 vmem_xfree(vm_dom[domain].vmd_kernel_arena, addr, bytes);
644 mtx_unlock(&vmem_bt_lock);
646 * Out of memory, not address space. This may not even be
647 * possible due to M_USE_RESERVE page allocation.
650 vm_wait_domain(domain);
653 mtx_unlock(&vmem_bt_lock);
655 * We're either out of address space or lost a fill race.
664 * How many pages do we need to startup_alloc.
667 vmem_startup_count(void)
670 return (howmany(BT_MAXALLOC,
671 UMA_SLAB_SPACE / sizeof(struct vmem_btag)));
679 mtx_init(&vmem_list_lock, "vmem list lock", NULL, MTX_DEF);
680 vmem_zone = uma_zcreate("vmem",
681 sizeof(struct vmem), NULL, NULL, NULL, NULL,
682 UMA_ALIGN_PTR, UMA_ZONE_VM);
683 vmem_bt_zone = uma_zcreate("vmem btag",
684 sizeof(struct vmem_btag), NULL, NULL, NULL, NULL,
685 UMA_ALIGN_PTR, UMA_ZONE_VM | UMA_ZONE_NOFREE);
686 #ifndef UMA_MD_SMALL_ALLOC
687 mtx_init(&vmem_bt_lock, "btag lock", NULL, MTX_DEF);
688 uma_prealloc(vmem_bt_zone, BT_MAXALLOC);
690 * Reserve enough tags to allocate new tags. We allow multiple
691 * CPUs to attempt to allocate new tags concurrently to limit
692 * false restarts in UMA. vmem_bt_alloc() allocates from a per-domain
693 * arena, which may involve importing a range from the kernel arena,
694 * so we need to keep at least 2 * BT_MAXALLOC tags reserved.
696 uma_zone_reserve(vmem_bt_zone, 2 * BT_MAXALLOC * mp_ncpus);
697 uma_zone_set_allocf(vmem_bt_zone, vmem_bt_alloc);
704 vmem_rehash(vmem_t *vm, vmem_size_t newhashsize)
708 struct vmem_hashlist *newhashlist;
709 struct vmem_hashlist *oldhashlist;
710 vmem_size_t oldhashsize;
712 MPASS(newhashsize > 0);
714 newhashlist = malloc(sizeof(struct vmem_hashlist) * newhashsize,
716 if (newhashlist == NULL)
718 for (i = 0; i < newhashsize; i++) {
719 LIST_INIT(&newhashlist[i]);
723 oldhashlist = vm->vm_hashlist;
724 oldhashsize = vm->vm_hashsize;
725 vm->vm_hashlist = newhashlist;
726 vm->vm_hashsize = newhashsize;
727 if (oldhashlist == NULL) {
731 for (i = 0; i < oldhashsize; i++) {
732 while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) {
739 if (oldhashlist != vm->vm_hash0) {
740 free(oldhashlist, M_VMEM);
747 vmem_periodic_kick(void *dummy)
750 taskqueue_enqueue(taskqueue_thread, &vmem_periodic_wk);
754 vmem_periodic(void *unused, int pending)
760 mtx_lock(&vmem_list_lock);
761 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
763 /* Convenient time to verify vmem state. */
764 if (enable_vmem_check == 1) {
770 desired = 1 << flsl(vm->vm_nbusytag);
771 desired = MIN(MAX(desired, VMEM_HASHSIZE_MIN),
773 current = vm->vm_hashsize;
775 /* Grow in powers of two. Shrink less aggressively. */
776 if (desired >= current * 2 || desired * 4 <= current)
777 vmem_rehash(vm, desired);
780 * Periodically wake up threads waiting for resources,
781 * so they could ask for reclamation again.
783 VMEM_CONDVAR_BROADCAST(vm);
785 mtx_unlock(&vmem_list_lock);
787 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
788 vmem_periodic_kick, NULL);
792 vmem_start_callout(void *unused)
795 TASK_INIT(&vmem_periodic_wk, 0, vmem_periodic, NULL);
796 vmem_periodic_interval = hz * 10;
797 callout_init(&vmem_periodic_ch, 1);
798 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
799 vmem_periodic_kick, NULL);
801 SYSINIT(vfs, SI_SUB_CONFIGURE, SI_ORDER_ANY, vmem_start_callout, NULL);
804 vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int type)
809 MPASS(type == BT_TYPE_SPAN || type == BT_TYPE_SPAN_STATIC);
810 MPASS((size & vm->vm_quantum_mask) == 0);
812 btspan = bt_alloc(vm);
813 btspan->bt_type = type;
814 btspan->bt_start = addr;
815 btspan->bt_size = size;
816 bt_insseg_tail(vm, btspan);
818 btfree = bt_alloc(vm);
819 btfree->bt_type = BT_TYPE_FREE;
820 btfree->bt_start = addr;
821 btfree->bt_size = size;
822 bt_insseg(vm, btfree, btspan);
823 bt_insfree(vm, btfree);
829 vmem_destroy1(vmem_t *vm)
834 * Drain per-cpu quantum caches.
839 * The vmem should now only contain empty segments.
842 MPASS(vm->vm_nbusytag == 0);
844 while ((bt = TAILQ_FIRST(&vm->vm_seglist)) != NULL)
847 if (vm->vm_hashlist != NULL && vm->vm_hashlist != vm->vm_hash0)
848 free(vm->vm_hashlist, M_VMEM);
852 VMEM_CONDVAR_DESTROY(vm);
853 VMEM_LOCK_DESTROY(vm);
854 uma_zfree(vmem_zone, vm);
858 vmem_import(vmem_t *vm, vmem_size_t size, vmem_size_t align, int flags)
863 if (vm->vm_importfn == NULL)
867 * To make sure we get a span that meets the alignment we double it
868 * and add the size to the tail. This slightly overestimates.
870 if (align != vm->vm_quantum_mask + 1)
871 size = (align * 2) + size;
872 size = roundup(size, vm->vm_import_quantum);
874 if (vm->vm_limit != 0 && vm->vm_limit < vm->vm_size + size)
878 * Hide MAXALLOC tags so we're guaranteed to be able to add this
879 * span and the tag we want to allocate from it.
881 MPASS(vm->vm_nfreetags >= BT_MAXALLOC);
882 vm->vm_nfreetags -= BT_MAXALLOC;
884 error = (vm->vm_importfn)(vm->vm_arg, size, flags, &addr);
886 vm->vm_nfreetags += BT_MAXALLOC;
890 vmem_add1(vm, addr, size, BT_TYPE_SPAN);
896 * vmem_fit: check if a bt can satisfy the given restrictions.
898 * it's a caller's responsibility to ensure the region is big enough
902 vmem_fit(const bt_t *bt, vmem_size_t size, vmem_size_t align,
903 vmem_size_t phase, vmem_size_t nocross, vmem_addr_t minaddr,
904 vmem_addr_t maxaddr, vmem_addr_t *addrp)
910 MPASS(bt->bt_size >= size); /* caller's responsibility */
913 * XXX assumption: vmem_addr_t and vmem_size_t are
914 * unsigned integer of the same size.
917 start = bt->bt_start;
918 if (start < minaddr) {
927 start = VMEM_ALIGNUP(start - phase, align) + phase;
928 if (start < bt->bt_start)
930 if (VMEM_CROSS_P(start, start + size - 1, nocross)) {
931 MPASS(align < nocross);
932 start = VMEM_ALIGNUP(start - phase, nocross) + phase;
934 if (start <= end && end - start >= size - 1) {
935 MPASS((start & (align - 1)) == phase);
936 MPASS(!VMEM_CROSS_P(start, start + size - 1, nocross));
937 MPASS(minaddr <= start);
938 MPASS(maxaddr == 0 || start + size - 1 <= maxaddr);
939 MPASS(bt->bt_start <= start);
940 MPASS(BT_END(bt) - start >= size - 1);
949 * vmem_clip: Trim the boundary tag edges to the requested start and size.
952 vmem_clip(vmem_t *vm, bt_t *bt, vmem_addr_t start, vmem_size_t size)
957 VMEM_ASSERT_LOCKED(vm);
958 MPASS(bt->bt_type == BT_TYPE_FREE);
959 MPASS(bt->bt_size >= size);
961 if (bt->bt_start != start) {
962 btprev = bt_alloc(vm);
963 btprev->bt_type = BT_TYPE_FREE;
964 btprev->bt_start = bt->bt_start;
965 btprev->bt_size = start - bt->bt_start;
966 bt->bt_start = start;
967 bt->bt_size -= btprev->bt_size;
968 bt_insfree(vm, btprev);
969 bt_insseg(vm, btprev,
970 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
972 MPASS(bt->bt_start == start);
973 if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) {
975 btnew = bt_alloc(vm);
976 btnew->bt_type = BT_TYPE_BUSY;
977 btnew->bt_start = bt->bt_start;
978 btnew->bt_size = size;
979 bt->bt_start = bt->bt_start + size;
983 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
984 bt_insbusy(vm, btnew);
987 bt->bt_type = BT_TYPE_BUSY;
990 MPASS(bt->bt_size >= size);
996 vmem_set_import(vmem_t *vm, vmem_import_t *importfn,
997 vmem_release_t *releasefn, void *arg, vmem_size_t import_quantum)
1001 vm->vm_importfn = importfn;
1002 vm->vm_releasefn = releasefn;
1004 vm->vm_import_quantum = import_quantum;
1009 vmem_set_limit(vmem_t *vm, vmem_size_t limit)
1013 vm->vm_limit = limit;
1018 vmem_set_reclaim(vmem_t *vm, vmem_reclaim_t *reclaimfn)
1022 vm->vm_reclaimfn = reclaimfn;
1027 * vmem_init: Initializes vmem arena.
1030 vmem_init(vmem_t *vm, const char *name, vmem_addr_t base, vmem_size_t size,
1031 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1036 MPASS((quantum & (quantum - 1)) == 0);
1038 bzero(vm, sizeof(*vm));
1040 VMEM_CONDVAR_INIT(vm, name);
1041 VMEM_LOCK_INIT(vm, name);
1042 vm->vm_nfreetags = 0;
1043 LIST_INIT(&vm->vm_freetags);
1044 strlcpy(vm->vm_name, name, sizeof(vm->vm_name));
1045 vm->vm_quantum_mask = quantum - 1;
1046 vm->vm_quantum_shift = flsl(quantum) - 1;
1047 vm->vm_nbusytag = 0;
1051 qc_init(vm, qcache_max);
1053 TAILQ_INIT(&vm->vm_seglist);
1054 for (i = 0; i < VMEM_MAXORDER; i++) {
1055 LIST_INIT(&vm->vm_freelist[i]);
1057 memset(&vm->vm_hash0, 0, sizeof(vm->vm_hash0));
1058 vm->vm_hashsize = VMEM_HASHSIZE_MIN;
1059 vm->vm_hashlist = vm->vm_hash0;
1062 if (vmem_add(vm, base, size, flags) != 0) {
1068 mtx_lock(&vmem_list_lock);
1069 LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist);
1070 mtx_unlock(&vmem_list_lock);
1076 * vmem_create: create an arena.
1079 vmem_create(const char *name, vmem_addr_t base, vmem_size_t size,
1080 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1085 vm = uma_zalloc(vmem_zone, flags & (M_WAITOK|M_NOWAIT));
1088 if (vmem_init(vm, name, base, size, quantum, qcache_max,
1095 vmem_destroy(vmem_t *vm)
1098 mtx_lock(&vmem_list_lock);
1099 LIST_REMOVE(vm, vm_alllist);
1100 mtx_unlock(&vmem_list_lock);
1106 vmem_roundup_size(vmem_t *vm, vmem_size_t size)
1109 return (size + vm->vm_quantum_mask) & ~vm->vm_quantum_mask;
1113 * vmem_alloc: allocate resource from the arena.
1116 vmem_alloc(vmem_t *vm, vmem_size_t size, int flags, vmem_addr_t *addrp)
1118 const int strat __unused = flags & VMEM_FITMASK;
1121 flags &= VMEM_FLAGS;
1123 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT);
1124 if ((flags & M_NOWAIT) == 0)
1125 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_alloc");
1127 if (size <= vm->vm_qcache_max) {
1129 * Resource 0 cannot be cached, so avoid a blocking allocation
1130 * in qc_import() and give the vmem_xalloc() call below a chance
1133 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1134 *addrp = (vmem_addr_t)uma_zalloc(qc->qc_cache,
1135 (flags & ~M_WAITOK) | M_NOWAIT);
1136 if (__predict_true(*addrp != 0))
1140 return (vmem_xalloc(vm, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
1145 vmem_xalloc(vmem_t *vm, const vmem_size_t size0, vmem_size_t align,
1146 const vmem_size_t phase, const vmem_size_t nocross,
1147 const vmem_addr_t minaddr, const vmem_addr_t maxaddr, int flags,
1150 const vmem_size_t size = vmem_roundup_size(vm, size0);
1151 struct vmem_freelist *list;
1152 struct vmem_freelist *first;
1153 struct vmem_freelist *end;
1159 flags &= VMEM_FLAGS;
1160 strat = flags & VMEM_FITMASK;
1163 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT);
1164 MPASS((flags & (M_NOWAIT|M_WAITOK)) != (M_NOWAIT|M_WAITOK));
1165 if ((flags & M_NOWAIT) == 0)
1166 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_xalloc");
1167 MPASS((align & vm->vm_quantum_mask) == 0);
1168 MPASS((align & (align - 1)) == 0);
1169 MPASS((phase & vm->vm_quantum_mask) == 0);
1170 MPASS((nocross & vm->vm_quantum_mask) == 0);
1171 MPASS((nocross & (nocross - 1)) == 0);
1172 MPASS((align == 0 && phase == 0) || phase < align);
1173 MPASS(nocross == 0 || nocross >= size);
1174 MPASS(minaddr <= maxaddr);
1175 MPASS(!VMEM_CROSS_P(phase, phase + size - 1, nocross));
1178 align = vm->vm_quantum_mask + 1;
1181 end = &vm->vm_freelist[VMEM_MAXORDER];
1183 * choose a free block from which we allocate.
1185 first = bt_freehead_toalloc(vm, size, strat);
1189 * Make sure we have enough tags to complete the
1192 if (vm->vm_nfreetags < BT_MAXALLOC &&
1193 bt_fill(vm, flags) != 0) {
1198 * Scan freelists looking for a tag that satisfies the
1199 * allocation. If we're doing BESTFIT we may encounter
1200 * sizes below the request. If we're doing FIRSTFIT we
1201 * inspect only the first element from each list.
1203 for (list = first; list < end; list++) {
1204 LIST_FOREACH(bt, list, bt_freelist) {
1205 if (bt->bt_size >= size) {
1206 error = vmem_fit(bt, size, align, phase,
1207 nocross, minaddr, maxaddr, addrp);
1209 vmem_clip(vm, bt, *addrp, size);
1213 /* FIRST skips to the next list. */
1214 if (strat == M_FIRSTFIT)
1219 * Retry if the fast algorithm failed.
1221 if (strat == M_FIRSTFIT) {
1223 first = bt_freehead_toalloc(vm, size, strat);
1227 * XXX it is possible to fail to meet restrictions with the
1228 * imported region. It is up to the user to specify the
1229 * import quantum such that it can satisfy any allocation.
1231 if (vmem_import(vm, size, align, flags) == 0)
1235 * Try to free some space from the quantum cache or reclaim
1236 * functions if available.
1238 if (vm->vm_qcache_max != 0 || vm->vm_reclaimfn != NULL) {
1239 avail = vm->vm_size - vm->vm_inuse;
1241 if (vm->vm_qcache_max != 0)
1243 if (vm->vm_reclaimfn != NULL)
1244 vm->vm_reclaimfn(vm, flags);
1246 /* If we were successful retry even NOWAIT. */
1247 if (vm->vm_size - vm->vm_inuse > avail)
1250 if ((flags & M_NOWAIT) != 0) {
1254 VMEM_CONDVAR_WAIT(vm);
1258 if (error != 0 && (flags & M_NOWAIT) == 0)
1259 panic("failed to allocate waiting allocation\n");
1265 * vmem_free: free the resource to the arena.
1268 vmem_free(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1273 if (size <= vm->vm_qcache_max &&
1274 __predict_true(addr >= VMEM_ADDR_QCACHE_MIN)) {
1275 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1276 uma_zfree(qc->qc_cache, (void *)addr);
1278 vmem_xfree(vm, addr, size);
1282 vmem_xfree(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1290 bt = bt_lookupbusy(vm, addr);
1292 MPASS(bt->bt_start == addr);
1293 MPASS(bt->bt_size == vmem_roundup_size(vm, size) ||
1294 bt->bt_size - vmem_roundup_size(vm, size) <= vm->vm_quantum_mask);
1295 MPASS(bt->bt_type == BT_TYPE_BUSY);
1297 bt->bt_type = BT_TYPE_FREE;
1300 t = TAILQ_NEXT(bt, bt_seglist);
1301 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1302 MPASS(BT_END(bt) < t->bt_start); /* YYY */
1303 bt->bt_size += t->bt_size;
1307 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1308 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1309 MPASS(BT_END(t) < bt->bt_start); /* YYY */
1310 bt->bt_size += t->bt_size;
1311 bt->bt_start = t->bt_start;
1316 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1318 MPASS(BT_ISSPAN_P(t) || t->bt_type == BT_TYPE_BUSY);
1319 if (vm->vm_releasefn != NULL && t->bt_type == BT_TYPE_SPAN &&
1320 t->bt_size == bt->bt_size) {
1321 vmem_addr_t spanaddr;
1322 vmem_size_t spansize;
1324 MPASS(t->bt_start == bt->bt_start);
1325 spanaddr = bt->bt_start;
1326 spansize = bt->bt_size;
1329 vm->vm_size -= spansize;
1330 VMEM_CONDVAR_BROADCAST(vm);
1331 bt_freetrim(vm, BT_MAXFREE);
1332 (*vm->vm_releasefn)(vm->vm_arg, spanaddr, spansize);
1335 VMEM_CONDVAR_BROADCAST(vm);
1336 bt_freetrim(vm, BT_MAXFREE);
1345 vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int flags)
1350 flags &= VMEM_FLAGS;
1352 if (vm->vm_nfreetags >= BT_MAXALLOC || bt_fill(vm, flags) == 0)
1353 vmem_add1(vm, addr, size, BT_TYPE_SPAN_STATIC);
1362 * vmem_size: information about arenas size
1365 vmem_size(vmem_t *vm, int typemask)
1371 return vm->vm_inuse;
1373 return vm->vm_size - vm->vm_inuse;
1374 case VMEM_FREE|VMEM_ALLOC:
1378 for (i = VMEM_MAXORDER - 1; i >= 0; i--) {
1379 if (LIST_EMPTY(&vm->vm_freelist[i]))
1382 return ((vmem_size_t)ORDER2SIZE(i) <<
1383 vm->vm_quantum_shift);
1394 #if defined(DDB) || defined(DIAGNOSTIC)
1396 static void bt_dump(const bt_t *, int (*)(const char *, ...)
1397 __printflike(1, 2));
1400 bt_type_string(int type)
1410 case BT_TYPE_SPAN_STATIC:
1411 return "static span";
1419 bt_dump(const bt_t *bt, int (*pr)(const char *, ...))
1422 (*pr)("\t%p: %jx %jx, %d(%s)\n",
1423 bt, (intmax_t)bt->bt_start, (intmax_t)bt->bt_size,
1424 bt->bt_type, bt_type_string(bt->bt_type));
1428 vmem_dump(const vmem_t *vm , int (*pr)(const char *, ...) __printflike(1, 2))
1433 (*pr)("vmem %p '%s'\n", vm, vm->vm_name);
1434 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1438 for (i = 0; i < VMEM_MAXORDER; i++) {
1439 const struct vmem_freelist *fl = &vm->vm_freelist[i];
1441 if (LIST_EMPTY(fl)) {
1445 (*pr)("freelist[%d]\n", i);
1446 LIST_FOREACH(bt, fl, bt_freelist) {
1452 #endif /* defined(DDB) || defined(DIAGNOSTIC) */
1455 #include <ddb/ddb.h>
1458 vmem_whatis_lookup(vmem_t *vm, vmem_addr_t addr)
1462 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1463 if (BT_ISSPAN_P(bt)) {
1466 if (bt->bt_start <= addr && addr <= BT_END(bt)) {
1475 vmem_whatis(vmem_addr_t addr, int (*pr)(const char *, ...))
1479 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1482 bt = vmem_whatis_lookup(vm, addr);
1486 (*pr)("%p is %p+%zu in VMEM '%s' (%s)\n",
1487 (void *)addr, (void *)bt->bt_start,
1488 (vmem_size_t)(addr - bt->bt_start), vm->vm_name,
1489 (bt->bt_type == BT_TYPE_BUSY) ? "allocated" : "free");
1494 vmem_printall(const char *modif, int (*pr)(const char *, ...))
1498 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1504 vmem_print(vmem_addr_t addr, const char *modif, int (*pr)(const char *, ...))
1506 const vmem_t *vm = (const void *)addr;
1511 DB_SHOW_COMMAND(vmemdump, vmemdump)
1515 db_printf("usage: show vmemdump <addr>\n");
1519 vmem_dump((const vmem_t *)addr, db_printf);
1522 DB_SHOW_ALL_COMMAND(vmemdump, vmemdumpall)
1526 LIST_FOREACH(vm, &vmem_list, vm_alllist)
1527 vmem_dump(vm, db_printf);
1530 DB_SHOW_COMMAND(vmem, vmem_summ)
1532 const vmem_t *vm = (const void *)addr;
1534 size_t ft[VMEM_MAXORDER], ut[VMEM_MAXORDER];
1535 size_t fs[VMEM_MAXORDER], us[VMEM_MAXORDER];
1539 db_printf("usage: show vmem <addr>\n");
1543 db_printf("vmem %p '%s'\n", vm, vm->vm_name);
1544 db_printf("\tquantum:\t%zu\n", vm->vm_quantum_mask + 1);
1545 db_printf("\tsize:\t%zu\n", vm->vm_size);
1546 db_printf("\tinuse:\t%zu\n", vm->vm_inuse);
1547 db_printf("\tfree:\t%zu\n", vm->vm_size - vm->vm_inuse);
1548 db_printf("\tbusy tags:\t%d\n", vm->vm_nbusytag);
1549 db_printf("\tfree tags:\t%d\n", vm->vm_nfreetags);
1551 memset(&ft, 0, sizeof(ft));
1552 memset(&ut, 0, sizeof(ut));
1553 memset(&fs, 0, sizeof(fs));
1554 memset(&us, 0, sizeof(us));
1555 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1556 ord = SIZE2ORDER(bt->bt_size >> vm->vm_quantum_shift);
1557 if (bt->bt_type == BT_TYPE_BUSY) {
1559 us[ord] += bt->bt_size;
1560 } else if (bt->bt_type == BT_TYPE_FREE) {
1562 fs[ord] += bt->bt_size;
1565 db_printf("\t\t\tinuse\tsize\t\tfree\tsize\n");
1566 for (ord = 0; ord < VMEM_MAXORDER; ord++) {
1567 if (ut[ord] == 0 && ft[ord] == 0)
1569 db_printf("\t%-15zu %zu\t%-15zu %zu\t%-16zu\n",
1570 ORDER2SIZE(ord) << vm->vm_quantum_shift,
1571 ut[ord], us[ord], ft[ord], fs[ord]);
1575 DB_SHOW_ALL_COMMAND(vmem, vmem_summall)
1579 LIST_FOREACH(vm, &vmem_list, vm_alllist)
1580 vmem_summ((db_expr_t)vm, TRUE, count, modif);
1582 #endif /* defined(DDB) */
1584 #define vmem_printf printf
1586 #if defined(DIAGNOSTIC)
1589 vmem_check_sanity(vmem_t *vm)
1591 const bt_t *bt, *bt2;
1595 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1596 if (bt->bt_start > BT_END(bt)) {
1597 printf("corrupted tag\n");
1598 bt_dump(bt, vmem_printf);
1602 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1603 TAILQ_FOREACH(bt2, &vm->vm_seglist, bt_seglist) {
1607 if (BT_ISSPAN_P(bt) != BT_ISSPAN_P(bt2)) {
1610 if (bt->bt_start <= BT_END(bt2) &&
1611 bt2->bt_start <= BT_END(bt)) {
1612 printf("overwrapped tags\n");
1613 bt_dump(bt, vmem_printf);
1614 bt_dump(bt2, vmem_printf);
1624 vmem_check(vmem_t *vm)
1627 if (!vmem_check_sanity(vm)) {
1628 panic("insanity vmem %p", vm);
1632 #endif /* defined(DIAGNOSTIC) */