2 * Copyright (c)2006,2007,2008,2009 YAMAMOTO Takashi,
3 * Copyright (c) 2013 EMC Corp.
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. Redistributions in binary form must reproduce the above copyright
12 * notice, this list of conditions and the following disclaimer in the
13 * documentation and/or other materials provided with the distribution.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
30 * $NetBSD: vmem_impl.h,v 1.2 2013/01/29 21:26:24 para Exp $
31 * $NetBSD: subr_vmem.c,v 1.83 2013/03/06 11:20:10 yamt Exp $
36 * - Magazines and Vmem: Extending the Slab Allocator
37 * to Many CPUs and Arbitrary Resources
38 * http://www.usenix.org/event/usenix01/bonwick.html
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
46 #include <sys/param.h>
47 #include <sys/systm.h>
48 #include <sys/kernel.h>
49 #include <sys/queue.h>
50 #include <sys/callout.h>
53 #include <sys/malloc.h>
54 #include <sys/mutex.h>
56 #include <sys/condvar.h>
57 #include <sys/taskqueue.h>
65 #include <vm/vm_map.h>
66 #include <vm/vm_object.h>
67 #include <vm/vm_kern.h>
68 #include <vm/vm_extern.h>
69 #include <vm/vm_param.h>
70 #include <vm/vm_pageout.h>
72 #define VMEM_MAXORDER (sizeof(vmem_size_t) * NBBY)
74 #define VMEM_HASHSIZE_MIN 16
75 #define VMEM_HASHSIZE_MAX 131072
77 #define VMEM_QCACHE_IDX_MAX 16
79 #define VMEM_FITMASK (M_BESTFIT | M_FIRSTFIT)
82 (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM | M_BESTFIT | M_FIRSTFIT)
84 #define BT_FLAGS (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM)
86 #define QC_NAME_MAX 16
89 * Data structures private to vmem.
91 MALLOC_DEFINE(M_VMEM, "vmem", "vmem internal structures");
93 typedef struct vmem_btag bt_t;
95 TAILQ_HEAD(vmem_seglist, vmem_btag);
96 LIST_HEAD(vmem_freelist, vmem_btag);
97 LIST_HEAD(vmem_hashlist, vmem_btag);
103 char qc_name[QC_NAME_MAX];
105 typedef struct qcache qcache_t;
106 #define QC_POOL_TO_QCACHE(pool) ((qcache_t *)(pool->pr_qcache))
108 #define VMEM_NAME_MAX 16
112 struct mtx_padalign vm_lock;
114 char vm_name[VMEM_NAME_MAX+1];
115 LIST_ENTRY(vmem) vm_alllist;
116 struct vmem_hashlist vm_hash0[VMEM_HASHSIZE_MIN];
117 struct vmem_freelist vm_freelist[VMEM_MAXORDER];
118 struct vmem_seglist vm_seglist;
119 struct vmem_hashlist *vm_hashlist;
120 vmem_size_t vm_hashsize;
122 /* Constant after init */
123 vmem_size_t vm_qcache_max;
124 vmem_size_t vm_quantum_mask;
125 vmem_size_t vm_import_quantum;
126 int vm_quantum_shift;
128 /* Written on alloc/free */
129 LIST_HEAD(, vmem_btag) vm_freetags;
132 vmem_size_t vm_inuse;
135 /* Used on import. */
136 vmem_import_t *vm_importfn;
137 vmem_release_t *vm_releasefn;
140 /* Space exhaustion callback. */
141 vmem_reclaim_t *vm_reclaimfn;
144 qcache_t vm_qcache[VMEM_QCACHE_IDX_MAX];
149 TAILQ_ENTRY(vmem_btag) bt_seglist;
151 LIST_ENTRY(vmem_btag) u_freelist; /* BT_TYPE_FREE */
152 LIST_ENTRY(vmem_btag) u_hashlist; /* BT_TYPE_BUSY */
154 #define bt_hashlist bt_u.u_hashlist
155 #define bt_freelist bt_u.u_freelist
156 vmem_addr_t bt_start;
161 #define BT_TYPE_SPAN 1 /* Allocated from importfn */
162 #define BT_TYPE_SPAN_STATIC 2 /* vmem_add() or create. */
163 #define BT_TYPE_FREE 3 /* Available space. */
164 #define BT_TYPE_BUSY 4 /* Used space. */
165 #define BT_ISSPAN_P(bt) ((bt)->bt_type <= BT_TYPE_SPAN_STATIC)
167 #define BT_END(bt) ((bt)->bt_start + (bt)->bt_size - 1)
169 #if defined(DIAGNOSTIC)
170 static void vmem_check(vmem_t *);
173 static struct callout vmem_periodic_ch;
174 static int vmem_periodic_interval;
175 static struct task vmem_periodic_wk;
177 static struct mtx_padalign vmem_list_lock;
178 static LIST_HEAD(, vmem) vmem_list = LIST_HEAD_INITIALIZER(vmem_list);
181 #define VMEM_CONDVAR_INIT(vm, wchan) cv_init(&vm->vm_cv, wchan)
182 #define VMEM_CONDVAR_DESTROY(vm) cv_destroy(&vm->vm_cv)
183 #define VMEM_CONDVAR_WAIT(vm) cv_wait(&vm->vm_cv, &vm->vm_lock)
184 #define VMEM_CONDVAR_BROADCAST(vm) cv_broadcast(&vm->vm_cv)
187 #define VMEM_LOCK(vm) mtx_lock(&vm->vm_lock)
188 #define VMEM_TRYLOCK(vm) mtx_trylock(&vm->vm_lock)
189 #define VMEM_UNLOCK(vm) mtx_unlock(&vm->vm_lock)
190 #define VMEM_LOCK_INIT(vm, name) mtx_init(&vm->vm_lock, (name), NULL, MTX_DEF)
191 #define VMEM_LOCK_DESTROY(vm) mtx_destroy(&vm->vm_lock)
192 #define VMEM_ASSERT_LOCKED(vm) mtx_assert(&vm->vm_lock, MA_OWNED);
194 #define VMEM_ALIGNUP(addr, align) (-(-(addr) & -(align)))
196 #define VMEM_CROSS_P(addr1, addr2, boundary) \
197 ((((addr1) ^ (addr2)) & -(boundary)) != 0)
199 #define ORDER2SIZE(order) ((vmem_size_t)1 << (order))
200 #define SIZE2ORDER(size) ((int)flsl(size) - 1)
203 * Maximum number of boundary tags that may be required to satisfy an
204 * allocation. Two may be required to import. Another two may be
205 * required to clip edges.
207 #define BT_MAXALLOC 4
210 * Max free limits the number of locally cached boundary tags. We
211 * just want to avoid hitting the zone allocator for every call.
213 #define BT_MAXFREE (BT_MAXALLOC * 8)
215 /* Allocator for boundary tags. */
216 static uma_zone_t vmem_bt_zone;
218 /* boot time arena storage. */
219 static struct vmem kernel_arena_storage;
220 static struct vmem kmem_arena_storage;
221 static struct vmem buffer_arena_storage;
222 static struct vmem transient_arena_storage;
223 vmem_t *kernel_arena = &kernel_arena_storage;
224 vmem_t *kmem_arena = &kmem_arena_storage;
225 vmem_t *buffer_arena = &buffer_arena_storage;
226 vmem_t *transient_arena = &transient_arena_storage;
228 #ifdef DEBUG_MEMGUARD
229 static struct vmem memguard_arena_storage;
230 vmem_t *memguard_arena = &memguard_arena_storage;
234 * Fill the vmem's boundary tag cache. We guarantee that boundary tag
235 * allocation will not fail once bt_fill() passes. To do so we cache
236 * at least the maximum possible tag allocations in the arena.
239 bt_fill(vmem_t *vm, int flags)
243 VMEM_ASSERT_LOCKED(vm);
246 * Only allow the kmem arena to dip into reserve tags. It is the
247 * vmem where new tags come from.
250 if (vm != kmem_arena)
251 flags &= ~M_USE_RESERVE;
254 * Loop until we meet the reserve. To minimize the lock shuffle
255 * and prevent simultaneous fills we first try a NOWAIT regardless
256 * of the caller's flags. Specify M_NOVM so we don't recurse while
257 * holding a vmem lock.
259 while (vm->vm_nfreetags < BT_MAXALLOC) {
260 bt = uma_zalloc(vmem_bt_zone,
261 (flags & M_USE_RESERVE) | M_NOWAIT | M_NOVM);
264 bt = uma_zalloc(vmem_bt_zone, flags);
266 if (bt == NULL && (flags & M_NOWAIT) != 0)
269 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
273 if (vm->vm_nfreetags < BT_MAXALLOC)
280 * Pop a tag off of the freetag stack.
287 VMEM_ASSERT_LOCKED(vm);
288 bt = LIST_FIRST(&vm->vm_freetags);
290 LIST_REMOVE(bt, bt_freelist);
297 * Trim the per-vmem free list. Returns with the lock released to
298 * avoid allocator recursions.
301 bt_freetrim(vmem_t *vm, int freelimit)
303 LIST_HEAD(, vmem_btag) freetags;
306 LIST_INIT(&freetags);
307 VMEM_ASSERT_LOCKED(vm);
308 while (vm->vm_nfreetags > freelimit) {
309 bt = LIST_FIRST(&vm->vm_freetags);
310 LIST_REMOVE(bt, bt_freelist);
312 LIST_INSERT_HEAD(&freetags, bt, bt_freelist);
315 while ((bt = LIST_FIRST(&freetags)) != NULL) {
316 LIST_REMOVE(bt, bt_freelist);
317 uma_zfree(vmem_bt_zone, bt);
322 bt_free(vmem_t *vm, bt_t *bt)
325 VMEM_ASSERT_LOCKED(vm);
326 MPASS(LIST_FIRST(&vm->vm_freetags) != bt);
327 LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist);
332 * freelist[0] ... [1, 1]
333 * freelist[1] ... [2, 3]
334 * freelist[2] ... [4, 7]
335 * freelist[3] ... [8, 15]
337 * freelist[n] ... [(1 << n), (1 << (n + 1)) - 1]
341 static struct vmem_freelist *
342 bt_freehead_tofree(vmem_t *vm, vmem_size_t size)
344 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
345 const int idx = SIZE2ORDER(qsize);
347 MPASS(size != 0 && qsize != 0);
348 MPASS((size & vm->vm_quantum_mask) == 0);
350 MPASS(idx < VMEM_MAXORDER);
352 return &vm->vm_freelist[idx];
356 * bt_freehead_toalloc: return the freelist for the given size and allocation
359 * For M_FIRSTFIT, return the list in which any blocks are large enough
360 * for the requested size. otherwise, return the list which can have blocks
361 * large enough for the requested size.
363 static struct vmem_freelist *
364 bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, int strat)
366 const vmem_size_t qsize = size >> vm->vm_quantum_shift;
367 int idx = SIZE2ORDER(qsize);
369 MPASS(size != 0 && qsize != 0);
370 MPASS((size & vm->vm_quantum_mask) == 0);
372 if (strat == M_FIRSTFIT && ORDER2SIZE(idx) != qsize) {
374 /* check too large request? */
377 MPASS(idx < VMEM_MAXORDER);
379 return &vm->vm_freelist[idx];
382 /* ---- boundary tag hash */
384 static struct vmem_hashlist *
385 bt_hashhead(vmem_t *vm, vmem_addr_t addr)
387 struct vmem_hashlist *list;
390 hash = hash32_buf(&addr, sizeof(addr), 0);
391 list = &vm->vm_hashlist[hash % vm->vm_hashsize];
397 bt_lookupbusy(vmem_t *vm, vmem_addr_t addr)
399 struct vmem_hashlist *list;
402 VMEM_ASSERT_LOCKED(vm);
403 list = bt_hashhead(vm, addr);
404 LIST_FOREACH(bt, list, bt_hashlist) {
405 if (bt->bt_start == addr) {
414 bt_rembusy(vmem_t *vm, bt_t *bt)
417 VMEM_ASSERT_LOCKED(vm);
418 MPASS(vm->vm_nbusytag > 0);
419 vm->vm_inuse -= bt->bt_size;
421 LIST_REMOVE(bt, bt_hashlist);
425 bt_insbusy(vmem_t *vm, bt_t *bt)
427 struct vmem_hashlist *list;
429 VMEM_ASSERT_LOCKED(vm);
430 MPASS(bt->bt_type == BT_TYPE_BUSY);
432 list = bt_hashhead(vm, bt->bt_start);
433 LIST_INSERT_HEAD(list, bt, bt_hashlist);
435 vm->vm_inuse += bt->bt_size;
438 /* ---- boundary tag list */
441 bt_remseg(vmem_t *vm, bt_t *bt)
444 TAILQ_REMOVE(&vm->vm_seglist, bt, bt_seglist);
449 bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev)
452 TAILQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist);
456 bt_insseg_tail(vmem_t *vm, bt_t *bt)
459 TAILQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist);
463 bt_remfree(vmem_t *vm, bt_t *bt)
466 MPASS(bt->bt_type == BT_TYPE_FREE);
468 LIST_REMOVE(bt, bt_freelist);
472 bt_insfree(vmem_t *vm, bt_t *bt)
474 struct vmem_freelist *list;
476 list = bt_freehead_tofree(vm, bt->bt_size);
477 LIST_INSERT_HEAD(list, bt, bt_freelist);
480 /* ---- vmem internal functions */
483 * Import from the arena into the quantum cache in UMA.
486 qc_import(void *arg, void **store, int cnt, int flags)
494 for (i = 0; i < cnt; i++) {
495 if (vmem_xalloc(qc->qc_vmem, qc->qc_size, 0, 0, 0,
496 VMEM_ADDR_MIN, VMEM_ADDR_MAX, flags, &addr) != 0)
498 store[i] = (void *)addr;
499 /* Only guarantee one allocation. */
507 * Release memory from the UMA cache to the arena.
510 qc_release(void *arg, void **store, int cnt)
516 for (i = 0; i < cnt; i++)
517 vmem_xfree(qc->qc_vmem, (vmem_addr_t)store[i], qc->qc_size);
521 qc_init(vmem_t *vm, vmem_size_t qcache_max)
528 MPASS((qcache_max & vm->vm_quantum_mask) == 0);
529 qcache_idx_max = MIN(qcache_max >> vm->vm_quantum_shift,
530 VMEM_QCACHE_IDX_MAX);
531 vm->vm_qcache_max = qcache_idx_max << vm->vm_quantum_shift;
532 for (i = 0; i < qcache_idx_max; i++) {
533 qc = &vm->vm_qcache[i];
534 size = (i + 1) << vm->vm_quantum_shift;
535 snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu",
539 qc->qc_cache = uma_zcache_create(qc->qc_name, size,
540 NULL, NULL, NULL, NULL, qc_import, qc_release, qc,
547 qc_destroy(vmem_t *vm)
552 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
553 for (i = 0; i < qcache_idx_max; i++)
554 uma_zdestroy(vm->vm_qcache[i].qc_cache);
563 qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift;
564 for (i = 0; i < qcache_idx_max; i++)
565 zone_drain(vm->vm_qcache[i].qc_cache);
568 #ifndef UMA_MD_SMALL_ALLOC
570 static struct mtx_padalign vmem_bt_lock;
573 * vmem_bt_alloc: Allocate a new page of boundary tags.
575 * On architectures with uma_small_alloc there is no recursion; no address
576 * space need be allocated to allocate boundary tags. For the others, we
577 * must handle recursion. Boundary tags are necessary to allocate new
580 * UMA guarantees that enough tags are held in reserve to allocate a new
581 * page of kva. We dip into this reserve by specifying M_USE_RESERVE only
582 * when allocating the page to hold new boundary tags. In this way the
583 * reserve is automatically filled by the allocation that uses the reserve.
585 * We still have to guarantee that the new tags are allocated atomically since
586 * many threads may try concurrently. The bt_lock provides this guarantee.
587 * We convert WAITOK allocations to NOWAIT and then handle the blocking here
588 * on failure. It's ok to return NULL for a WAITOK allocation as UMA will
589 * loop again after checking to see if we lost the race to allocate.
591 * There is a small race between vmem_bt_alloc() returning the page and the
592 * zone lock being acquired to add the page to the zone. For WAITOK
593 * allocations we just pause briefly. NOWAIT may experience a transient
594 * failure. To alleviate this we permit a small number of simultaneous
595 * fills to proceed concurrently so NOWAIT is less likely to fail unless
596 * we are really out of KVA.
599 vmem_bt_alloc(uma_zone_t zone, int bytes, uint8_t *pflag, int wait)
603 *pflag = UMA_SLAB_KMEM;
606 * Single thread boundary tag allocation so that the address space
607 * and memory are added in one atomic operation.
609 mtx_lock(&vmem_bt_lock);
610 if (vmem_xalloc(kmem_arena, bytes, 0, 0, 0, VMEM_ADDR_MIN,
611 VMEM_ADDR_MAX, M_NOWAIT | M_NOVM | M_USE_RESERVE | M_BESTFIT,
613 if (kmem_back(kmem_object, addr, bytes,
614 M_NOWAIT | M_USE_RESERVE) == 0) {
615 mtx_unlock(&vmem_bt_lock);
616 return ((void *)addr);
618 vmem_xfree(kmem_arena, addr, bytes);
619 mtx_unlock(&vmem_bt_lock);
621 * Out of memory, not address space. This may not even be
622 * possible due to M_USE_RESERVE page allocation.
628 mtx_unlock(&vmem_bt_lock);
630 * We're either out of address space or lost a fill race.
643 mtx_init(&vmem_list_lock, "vmem list lock", NULL, MTX_DEF);
644 vmem_bt_zone = uma_zcreate("vmem btag",
645 sizeof(struct vmem_btag), NULL, NULL, NULL, NULL,
646 UMA_ALIGN_PTR, UMA_ZONE_VM);
647 #ifndef UMA_MD_SMALL_ALLOC
648 mtx_init(&vmem_bt_lock, "btag lock", NULL, MTX_DEF);
649 uma_prealloc(vmem_bt_zone, BT_MAXALLOC);
651 * Reserve enough tags to allocate new tags. We allow multiple
652 * CPUs to attempt to allocate new tags concurrently to limit
653 * false restarts in UMA.
655 uma_zone_reserve(vmem_bt_zone, BT_MAXALLOC * (mp_ncpus + 1) / 2);
656 uma_zone_set_allocf(vmem_bt_zone, vmem_bt_alloc);
663 vmem_rehash(vmem_t *vm, vmem_size_t newhashsize)
667 struct vmem_hashlist *newhashlist;
668 struct vmem_hashlist *oldhashlist;
669 vmem_size_t oldhashsize;
671 MPASS(newhashsize > 0);
673 newhashlist = malloc(sizeof(struct vmem_hashlist) * newhashsize,
675 if (newhashlist == NULL)
677 for (i = 0; i < newhashsize; i++) {
678 LIST_INIT(&newhashlist[i]);
682 oldhashlist = vm->vm_hashlist;
683 oldhashsize = vm->vm_hashsize;
684 vm->vm_hashlist = newhashlist;
685 vm->vm_hashsize = newhashsize;
686 if (oldhashlist == NULL) {
690 for (i = 0; i < oldhashsize; i++) {
691 while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) {
698 if (oldhashlist != vm->vm_hash0) {
699 free(oldhashlist, M_VMEM);
706 vmem_periodic_kick(void *dummy)
709 taskqueue_enqueue(taskqueue_thread, &vmem_periodic_wk);
713 vmem_periodic(void *unused, int pending)
719 mtx_lock(&vmem_list_lock);
720 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
722 /* Convenient time to verify vmem state. */
727 desired = 1 << flsl(vm->vm_nbusytag);
728 desired = MIN(MAX(desired, VMEM_HASHSIZE_MIN),
730 current = vm->vm_hashsize;
732 /* Grow in powers of two. Shrink less aggressively. */
733 if (desired >= current * 2 || desired * 4 <= current)
734 vmem_rehash(vm, desired);
736 mtx_unlock(&vmem_list_lock);
738 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
739 vmem_periodic_kick, NULL);
743 vmem_start_callout(void *unused)
746 TASK_INIT(&vmem_periodic_wk, 0, vmem_periodic, NULL);
747 vmem_periodic_interval = hz * 10;
748 callout_init(&vmem_periodic_ch, CALLOUT_MPSAFE);
749 callout_reset(&vmem_periodic_ch, vmem_periodic_interval,
750 vmem_periodic_kick, NULL);
752 SYSINIT(vfs, SI_SUB_CONFIGURE, SI_ORDER_ANY, vmem_start_callout, NULL);
755 vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int type)
760 MPASS(type == BT_TYPE_SPAN || type == BT_TYPE_SPAN_STATIC);
761 MPASS((size & vm->vm_quantum_mask) == 0);
763 btspan = bt_alloc(vm);
764 btspan->bt_type = type;
765 btspan->bt_start = addr;
766 btspan->bt_size = size;
767 bt_insseg_tail(vm, btspan);
769 btfree = bt_alloc(vm);
770 btfree->bt_type = BT_TYPE_FREE;
771 btfree->bt_start = addr;
772 btfree->bt_size = size;
773 bt_insseg(vm, btfree, btspan);
774 bt_insfree(vm, btfree);
780 vmem_destroy1(vmem_t *vm)
785 * Drain per-cpu quantum caches.
790 * The vmem should now only contain empty segments.
793 MPASS(vm->vm_nbusytag == 0);
795 while ((bt = TAILQ_FIRST(&vm->vm_seglist)) != NULL)
798 if (vm->vm_hashlist != NULL && vm->vm_hashlist != vm->vm_hash0)
799 free(vm->vm_hashlist, M_VMEM);
803 VMEM_CONDVAR_DESTROY(vm);
804 VMEM_LOCK_DESTROY(vm);
809 vmem_import(vmem_t *vm, vmem_size_t size, vmem_size_t align, int flags)
814 if (vm->vm_importfn == NULL)
818 * To make sure we get a span that meets the alignment we double it
819 * and add the size to the tail. This slightly overestimates.
821 if (align != vm->vm_quantum_mask + 1)
822 size = (align * 2) + size;
823 size = roundup(size, vm->vm_import_quantum);
826 * Hide MAXALLOC tags so we're guaranteed to be able to add this
827 * span and the tag we want to allocate from it.
829 MPASS(vm->vm_nfreetags >= BT_MAXALLOC);
830 vm->vm_nfreetags -= BT_MAXALLOC;
832 error = (vm->vm_importfn)(vm->vm_arg, size, flags, &addr);
834 vm->vm_nfreetags += BT_MAXALLOC;
838 vmem_add1(vm, addr, size, BT_TYPE_SPAN);
844 * vmem_fit: check if a bt can satisfy the given restrictions.
846 * it's a caller's responsibility to ensure the region is big enough
850 vmem_fit(const bt_t *bt, vmem_size_t size, vmem_size_t align,
851 vmem_size_t phase, vmem_size_t nocross, vmem_addr_t minaddr,
852 vmem_addr_t maxaddr, vmem_addr_t *addrp)
858 MPASS(bt->bt_size >= size); /* caller's responsibility */
861 * XXX assumption: vmem_addr_t and vmem_size_t are
862 * unsigned integer of the same size.
865 start = bt->bt_start;
866 if (start < minaddr) {
875 start = VMEM_ALIGNUP(start - phase, align) + phase;
876 if (start < bt->bt_start)
878 if (VMEM_CROSS_P(start, start + size - 1, nocross)) {
879 MPASS(align < nocross);
880 start = VMEM_ALIGNUP(start - phase, nocross) + phase;
882 if (start <= end && end - start >= size - 1) {
883 MPASS((start & (align - 1)) == phase);
884 MPASS(!VMEM_CROSS_P(start, start + size - 1, nocross));
885 MPASS(minaddr <= start);
886 MPASS(maxaddr == 0 || start + size - 1 <= maxaddr);
887 MPASS(bt->bt_start <= start);
888 MPASS(BT_END(bt) - start >= size - 1);
897 * vmem_clip: Trim the boundary tag edges to the requested start and size.
900 vmem_clip(vmem_t *vm, bt_t *bt, vmem_addr_t start, vmem_size_t size)
905 VMEM_ASSERT_LOCKED(vm);
906 MPASS(bt->bt_type == BT_TYPE_FREE);
907 MPASS(bt->bt_size >= size);
909 if (bt->bt_start != start) {
910 btprev = bt_alloc(vm);
911 btprev->bt_type = BT_TYPE_FREE;
912 btprev->bt_start = bt->bt_start;
913 btprev->bt_size = start - bt->bt_start;
914 bt->bt_start = start;
915 bt->bt_size -= btprev->bt_size;
916 bt_insfree(vm, btprev);
917 bt_insseg(vm, btprev,
918 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
920 MPASS(bt->bt_start == start);
921 if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) {
923 btnew = bt_alloc(vm);
924 btnew->bt_type = BT_TYPE_BUSY;
925 btnew->bt_start = bt->bt_start;
926 btnew->bt_size = size;
927 bt->bt_start = bt->bt_start + size;
931 TAILQ_PREV(bt, vmem_seglist, bt_seglist));
932 bt_insbusy(vm, btnew);
935 bt->bt_type = BT_TYPE_BUSY;
938 MPASS(bt->bt_size >= size);
939 bt->bt_type = BT_TYPE_BUSY;
945 vmem_set_import(vmem_t *vm, vmem_import_t *importfn,
946 vmem_release_t *releasefn, void *arg, vmem_size_t import_quantum)
950 vm->vm_importfn = importfn;
951 vm->vm_releasefn = releasefn;
953 vm->vm_import_quantum = import_quantum;
958 vmem_set_reclaim(vmem_t *vm, vmem_reclaim_t *reclaimfn)
962 vm->vm_reclaimfn = reclaimfn;
967 * vmem_init: Initializes vmem arena.
970 vmem_init(vmem_t *vm, const char *name, vmem_addr_t base, vmem_size_t size,
971 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
977 bzero(vm, sizeof(*vm));
979 VMEM_CONDVAR_INIT(vm, name);
980 VMEM_LOCK_INIT(vm, name);
981 vm->vm_nfreetags = 0;
982 LIST_INIT(&vm->vm_freetags);
983 strlcpy(vm->vm_name, name, sizeof(vm->vm_name));
984 vm->vm_quantum_mask = quantum - 1;
985 vm->vm_quantum_shift = SIZE2ORDER(quantum);
986 MPASS(ORDER2SIZE(vm->vm_quantum_shift) == quantum);
990 qc_init(vm, qcache_max);
992 TAILQ_INIT(&vm->vm_seglist);
993 for (i = 0; i < VMEM_MAXORDER; i++) {
994 LIST_INIT(&vm->vm_freelist[i]);
996 memset(&vm->vm_hash0, 0, sizeof(vm->vm_hash0));
997 vm->vm_hashsize = VMEM_HASHSIZE_MIN;
998 vm->vm_hashlist = vm->vm_hash0;
1001 if (vmem_add(vm, base, size, flags) != 0) {
1007 mtx_lock(&vmem_list_lock);
1008 LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist);
1009 mtx_unlock(&vmem_list_lock);
1015 * vmem_create: create an arena.
1018 vmem_create(const char *name, vmem_addr_t base, vmem_size_t size,
1019 vmem_size_t quantum, vmem_size_t qcache_max, int flags)
1024 vm = malloc(sizeof(*vm), M_VMEM, flags & (M_WAITOK|M_NOWAIT));
1027 if (vmem_init(vm, name, base, size, quantum, qcache_max,
1036 vmem_destroy(vmem_t *vm)
1039 mtx_lock(&vmem_list_lock);
1040 LIST_REMOVE(vm, vm_alllist);
1041 mtx_unlock(&vmem_list_lock);
1047 vmem_roundup_size(vmem_t *vm, vmem_size_t size)
1050 return (size + vm->vm_quantum_mask) & ~vm->vm_quantum_mask;
1054 * vmem_alloc: allocate resource from the arena.
1057 vmem_alloc(vmem_t *vm, vmem_size_t size, int flags, vmem_addr_t *addrp)
1059 const int strat __unused = flags & VMEM_FITMASK;
1062 flags &= VMEM_FLAGS;
1064 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT);
1065 if ((flags & M_NOWAIT) == 0)
1066 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_alloc");
1068 if (size <= vm->vm_qcache_max) {
1069 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1070 *addrp = (vmem_addr_t)uma_zalloc(qc->qc_cache, flags);
1076 return vmem_xalloc(vm, size, 0, 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX,
1081 vmem_xalloc(vmem_t *vm, const vmem_size_t size0, vmem_size_t align,
1082 const vmem_size_t phase, const vmem_size_t nocross,
1083 const vmem_addr_t minaddr, const vmem_addr_t maxaddr, int flags,
1086 const vmem_size_t size = vmem_roundup_size(vm, size0);
1087 struct vmem_freelist *list;
1088 struct vmem_freelist *first;
1089 struct vmem_freelist *end;
1095 flags &= VMEM_FLAGS;
1096 strat = flags & VMEM_FITMASK;
1099 MPASS(strat == M_BESTFIT || strat == M_FIRSTFIT);
1100 MPASS((flags & (M_NOWAIT|M_WAITOK)) != (M_NOWAIT|M_WAITOK));
1101 if ((flags & M_NOWAIT) == 0)
1102 WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, "vmem_xalloc");
1103 MPASS((align & vm->vm_quantum_mask) == 0);
1104 MPASS((align & (align - 1)) == 0);
1105 MPASS((phase & vm->vm_quantum_mask) == 0);
1106 MPASS((nocross & vm->vm_quantum_mask) == 0);
1107 MPASS((nocross & (nocross - 1)) == 0);
1108 MPASS((align == 0 && phase == 0) || phase < align);
1109 MPASS(nocross == 0 || nocross >= size);
1110 MPASS(minaddr <= maxaddr);
1111 MPASS(!VMEM_CROSS_P(phase, phase + size - 1, nocross));
1114 align = vm->vm_quantum_mask + 1;
1117 end = &vm->vm_freelist[VMEM_MAXORDER];
1119 * choose a free block from which we allocate.
1121 first = bt_freehead_toalloc(vm, size, strat);
1125 * Make sure we have enough tags to complete the
1128 if (vm->vm_nfreetags < BT_MAXALLOC &&
1129 bt_fill(vm, flags) != 0) {
1134 * Scan freelists looking for a tag that satisfies the
1135 * allocation. If we're doing BESTFIT we may encounter
1136 * sizes below the request. If we're doing FIRSTFIT we
1137 * inspect only the first element from each list.
1139 for (list = first; list < end; list++) {
1140 LIST_FOREACH(bt, list, bt_freelist) {
1141 if (bt->bt_size >= size) {
1142 error = vmem_fit(bt, size, align, phase,
1143 nocross, minaddr, maxaddr, addrp);
1145 vmem_clip(vm, bt, *addrp, size);
1149 /* FIRST skips to the next list. */
1150 if (strat == M_FIRSTFIT)
1155 * Retry if the fast algorithm failed.
1157 if (strat == M_FIRSTFIT) {
1159 first = bt_freehead_toalloc(vm, size, strat);
1163 * XXX it is possible to fail to meet restrictions with the
1164 * imported region. It is up to the user to specify the
1165 * import quantum such that it can satisfy any allocation.
1167 if (vmem_import(vm, size, align, flags) == 0)
1171 * Try to free some space from the quantum cache or reclaim
1172 * functions if available.
1174 if (vm->vm_qcache_max != 0 || vm->vm_reclaimfn != NULL) {
1175 avail = vm->vm_size - vm->vm_inuse;
1177 if (vm->vm_qcache_max != 0)
1179 if (vm->vm_reclaimfn != NULL)
1180 vm->vm_reclaimfn(vm, flags);
1182 /* If we were successful retry even NOWAIT. */
1183 if (vm->vm_size - vm->vm_inuse > avail)
1186 if ((flags & M_NOWAIT) != 0) {
1190 VMEM_CONDVAR_WAIT(vm);
1194 if (error != 0 && (flags & M_NOWAIT) == 0)
1195 panic("failed to allocate waiting allocation\n");
1201 * vmem_free: free the resource to the arena.
1204 vmem_free(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1209 if (size <= vm->vm_qcache_max) {
1210 qc = &vm->vm_qcache[(size - 1) >> vm->vm_quantum_shift];
1211 uma_zfree(qc->qc_cache, (void *)addr);
1213 vmem_xfree(vm, addr, size);
1217 vmem_xfree(vmem_t *vm, vmem_addr_t addr, vmem_size_t size)
1225 bt = bt_lookupbusy(vm, addr);
1227 MPASS(bt->bt_start == addr);
1228 MPASS(bt->bt_size == vmem_roundup_size(vm, size) ||
1229 bt->bt_size - vmem_roundup_size(vm, size) <= vm->vm_quantum_mask);
1230 MPASS(bt->bt_type == BT_TYPE_BUSY);
1232 bt->bt_type = BT_TYPE_FREE;
1235 t = TAILQ_NEXT(bt, bt_seglist);
1236 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1237 MPASS(BT_END(bt) < t->bt_start); /* YYY */
1238 bt->bt_size += t->bt_size;
1242 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1243 if (t != NULL && t->bt_type == BT_TYPE_FREE) {
1244 MPASS(BT_END(t) < bt->bt_start); /* YYY */
1245 bt->bt_size += t->bt_size;
1246 bt->bt_start = t->bt_start;
1251 t = TAILQ_PREV(bt, vmem_seglist, bt_seglist);
1253 MPASS(BT_ISSPAN_P(t) || t->bt_type == BT_TYPE_BUSY);
1254 if (vm->vm_releasefn != NULL && t->bt_type == BT_TYPE_SPAN &&
1255 t->bt_size == bt->bt_size) {
1256 vmem_addr_t spanaddr;
1257 vmem_size_t spansize;
1259 MPASS(t->bt_start == bt->bt_start);
1260 spanaddr = bt->bt_start;
1261 spansize = bt->bt_size;
1264 vm->vm_size -= spansize;
1265 VMEM_CONDVAR_BROADCAST(vm);
1266 bt_freetrim(vm, BT_MAXFREE);
1267 (*vm->vm_releasefn)(vm->vm_arg, spanaddr, spansize);
1270 VMEM_CONDVAR_BROADCAST(vm);
1271 bt_freetrim(vm, BT_MAXFREE);
1280 vmem_add(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int flags)
1285 flags &= VMEM_FLAGS;
1287 if (vm->vm_nfreetags >= BT_MAXALLOC || bt_fill(vm, flags) == 0)
1288 vmem_add1(vm, addr, size, BT_TYPE_SPAN_STATIC);
1297 * vmem_size: information about arenas size
1300 vmem_size(vmem_t *vm, int typemask)
1305 return vm->vm_inuse;
1307 return vm->vm_size - vm->vm_inuse;
1308 case VMEM_FREE|VMEM_ALLOC:
1317 #if defined(DDB) || defined(DIAGNOSTIC)
1319 static void bt_dump(const bt_t *, int (*)(const char *, ...)
1320 __printflike(1, 2));
1323 bt_type_string(int type)
1333 case BT_TYPE_SPAN_STATIC:
1334 return "static span";
1342 bt_dump(const bt_t *bt, int (*pr)(const char *, ...))
1345 (*pr)("\t%p: %jx %jx, %d(%s)\n",
1346 bt, (intmax_t)bt->bt_start, (intmax_t)bt->bt_size,
1347 bt->bt_type, bt_type_string(bt->bt_type));
1351 vmem_dump(const vmem_t *vm , int (*pr)(const char *, ...) __printflike(1, 2))
1356 (*pr)("vmem %p '%s'\n", vm, vm->vm_name);
1357 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1361 for (i = 0; i < VMEM_MAXORDER; i++) {
1362 const struct vmem_freelist *fl = &vm->vm_freelist[i];
1364 if (LIST_EMPTY(fl)) {
1368 (*pr)("freelist[%d]\n", i);
1369 LIST_FOREACH(bt, fl, bt_freelist) {
1375 #endif /* defined(DDB) || defined(DIAGNOSTIC) */
1379 vmem_whatis_lookup(vmem_t *vm, vmem_addr_t addr)
1383 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1384 if (BT_ISSPAN_P(bt)) {
1387 if (bt->bt_start <= addr && addr <= BT_END(bt)) {
1396 vmem_whatis(vmem_addr_t addr, int (*pr)(const char *, ...))
1400 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1403 bt = vmem_whatis_lookup(vm, addr);
1407 (*pr)("%p is %p+%zu in VMEM '%s' (%s)\n",
1408 (void *)addr, (void *)bt->bt_start,
1409 (vmem_size_t)(addr - bt->bt_start), vm->vm_name,
1410 (bt->bt_type == BT_TYPE_BUSY) ? "allocated" : "free");
1415 vmem_printall(const char *modif, int (*pr)(const char *, ...))
1419 LIST_FOREACH(vm, &vmem_list, vm_alllist) {
1425 vmem_print(vmem_addr_t addr, const char *modif, int (*pr)(const char *, ...))
1427 const vmem_t *vm = (const void *)addr;
1431 #endif /* defined(DDB) */
1433 #define vmem_printf printf
1435 #if defined(DIAGNOSTIC)
1438 vmem_check_sanity(vmem_t *vm)
1440 const bt_t *bt, *bt2;
1444 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1445 if (bt->bt_start > BT_END(bt)) {
1446 printf("corrupted tag\n");
1447 bt_dump(bt, vmem_printf);
1451 TAILQ_FOREACH(bt, &vm->vm_seglist, bt_seglist) {
1452 TAILQ_FOREACH(bt2, &vm->vm_seglist, bt_seglist) {
1456 if (BT_ISSPAN_P(bt) != BT_ISSPAN_P(bt2)) {
1459 if (bt->bt_start <= BT_END(bt2) &&
1460 bt2->bt_start <= BT_END(bt)) {
1461 printf("overwrapped tags\n");
1462 bt_dump(bt, vmem_printf);
1463 bt_dump(bt2, vmem_printf);
1473 vmem_check(vmem_t *vm)
1476 if (!vmem_check_sanity(vm)) {
1477 panic("insanity vmem %p", vm);
1481 #endif /* defined(DIAGNOSTIC) */