Date: Fri, 28 Jun 2013 00:36:29 -1000 (HST) From: Jeff Roberson <jroberson@jroberson.net> To: Adrian Chadd <adrian@freebsd.org> Cc: svn-src-head@freebsd.org, Jeff Roberson <jeff@freebsd.org>, src-committers@freebsd.org, svn-src-all@freebsd.org Subject: Re: svn commit: r252330 - in head/sys: conf geom kern sys vm Message-ID: <alpine.BSF.2.00.1306280032160.95115@desktop> In-Reply-To: <CAJ-VmokRJrdKbZnNEDRfWahsWL-WNReHozV=_t3sw=6OrbcKNQ@mail.gmail.com> References: <201306280351.r5S3pLAm098083@svn.freebsd.org> <CAJ-VmokRJrdKbZnNEDRfWahsWL-WNReHozV=_t3sw=6OrbcKNQ@mail.gmail.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Fri, 28 Jun 2013, Adrian Chadd wrote: > Hi, > > Do we really need another allocator / resource manager just for this? > No, however; I have a follow-up patch to replace kmem with this. And then we will use it for NUMA allocations in the kernel. After that it is likely that we could replace several other less efficient allocators with this. Solaris uses it for pids, tids, device unit numbers. etc. We could easily do the same. The existing allocators have failure modes, big O cost, and allocation requirements that are not tolerable for use in the vm. This also has a very nice feature that works with UMA to provide per-cpu caches of arbitrary number ranges. So it is more scalable as well as providing for less fragmentation. Thanks, Jeff > > > > Adrian > > On 27 June 2013 20:51, Jeff Roberson <jeff@freebsd.org> wrote: >> Author: jeff >> Date: Fri Jun 28 03:51:20 2013 >> New Revision: 252330 >> URL: http://svnweb.freebsd.org/changeset/base/252330 >> >> Log: >> - Add a general purpose resource allocator, vmem, from NetBSD. It was >> originally inspired by the Solaris vmem detailed in the proceedings >> of usenix 2001. The NetBSD version was heavily refactored for bugs >> and simplicity. >> - Use this resource allocator to allocate the buffer and transient maps. >> Buffer cache defrags are reduced by 25% when used by filesystems with >> mixed block sizes. Ultimately this may permit dynamic buffer cache >> sizing on low KVA machines. >> >> Discussed with: alc, kib, attilio >> Tested by: pho >> Sponsored by: EMC / Isilon Storage Division >> >> Added: >> head/sys/kern/subr_vmem.c (contents, props changed) >> head/sys/sys/vmem.h (contents, props changed) >> Modified: >> head/sys/conf/files >> head/sys/geom/geom_io.c >> head/sys/kern/vfs_bio.c >> head/sys/sys/malloc.h >> head/sys/vm/vm.h >> head/sys/vm/vm_init.c >> head/sys/vm/vm_kern.c >> head/sys/vm/vm_kern.h >> head/sys/vm/vm_object.c >> head/sys/vm/vm_pager.c >> head/sys/vm/vm_pager.h >> >> Modified: head/sys/conf/files >> ============================================================================== >> --- head/sys/conf/files Fri Jun 28 03:41:23 2013 (r252329) >> +++ head/sys/conf/files Fri Jun 28 03:51:20 2013 (r252330) >> @@ -2797,6 +2797,7 @@ kern/subr_trap.c standard >> kern/subr_turnstile.c standard >> kern/subr_uio.c standard >> kern/subr_unit.c standard >> +kern/subr_vmem.c standard >> kern/subr_witness.c optional witness >> kern/sys_capability.c standard >> kern/sys_generic.c standard >> >> Modified: head/sys/geom/geom_io.c >> ============================================================================== >> --- head/sys/geom/geom_io.c Fri Jun 28 03:41:23 2013 (r252329) >> +++ head/sys/geom/geom_io.c Fri Jun 28 03:51:20 2013 (r252330) >> @@ -49,6 +49,7 @@ __FBSDID("$FreeBSD$"); >> #include <sys/proc.h> >> #include <sys/stack.h> >> #include <sys/sysctl.h> >> +#include <sys/vmem.h> >> >> #include <sys/errno.h> >> #include <geom/geom.h> >> @@ -626,7 +627,6 @@ g_io_transient_map_bio(struct bio *bp) >> vm_offset_t addr; >> long size; >> u_int retried; >> - int rv; >> >> KASSERT(unmapped_buf_allowed, ("unmapped disabled")); >> >> @@ -636,10 +636,7 @@ g_io_transient_map_bio(struct bio *bp) >> retried = 0; >> atomic_add_long(&transient_maps, 1); >> retry: >> - vm_map_lock(bio_transient_map); >> - if (vm_map_findspace(bio_transient_map, vm_map_min(bio_transient_map), >> - size, &addr)) { >> - vm_map_unlock(bio_transient_map); >> + if (vmem_alloc(transient_arena, size, M_BESTFIT | M_NOWAIT, &addr)) { >> if (transient_map_retries != 0 && >> retried >= transient_map_retries) { >> g_io_deliver(bp, EDEADLK/* XXXKIB */); >> @@ -651,7 +648,7 @@ retry: >> /* >> * Naive attempt to quisce the I/O to get more >> * in-flight requests completed and defragment >> - * the bio_transient_map. >> + * the transient_arena. >> */ >> CTR3(KTR_GEOM, "g_down retrymap bp %p provider %s r %d", >> bp, bp->bio_to->name, retried); >> @@ -661,12 +658,6 @@ retry: >> goto retry; >> } >> } >> - rv = vm_map_insert(bio_transient_map, NULL, 0, addr, addr + size, >> - VM_PROT_RW, VM_PROT_RW, MAP_NOFAULT); >> - KASSERT(rv == KERN_SUCCESS, >> - ("vm_map_insert(bio_transient_map) rv %d %jx %lx", >> - rv, (uintmax_t)addr, size)); >> - vm_map_unlock(bio_transient_map); >> atomic_add_int(&inflight_transient_maps, 1); >> pmap_qenter((vm_offset_t)addr, bp->bio_ma, OFF_TO_IDX(size)); >> bp->bio_data = (caddr_t)addr + bp->bio_ma_offset; >> >> Added: head/sys/kern/subr_vmem.c >> ============================================================================== >> --- /dev/null 00:00:00 1970 (empty, because file is newly added) >> +++ head/sys/kern/subr_vmem.c Fri Jun 28 03:51:20 2013 (r252330) >> @@ -0,0 +1,1372 @@ >> +/*- >> + * Copyright (c)2006,2007,2008,2009 YAMAMOTO Takashi, >> + * Copyright (c) 2013 EMC Corp. >> + * All rights reserved. >> + * >> + * Redistribution and use in source and binary forms, with or without >> + * modification, are permitted provided that the following conditions >> + * are met: >> + * 1. Redistributions of source code must retain the above copyright >> + * notice, this list of conditions and the following disclaimer. >> + * 2. Redistributions in binary form must reproduce the above copyright >> + * notice, this list of conditions and the following disclaimer in the >> + * documentation and/or other materials provided with the distribution. >> + * >> + * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND >> + * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE >> + * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE >> + * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE >> + * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL >> + * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS >> + * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) >> + * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT >> + * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY >> + * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF >> + * SUCH DAMAGE. >> + */ >> + >> +/* >> + * From: >> + * $NetBSD: vmem_impl.h,v 1.2 2013/01/29 21:26:24 para Exp $ >> + * $NetBSD: subr_vmem.c,v 1.83 2013/03/06 11:20:10 yamt Exp $ >> + */ >> + >> +/* >> + * reference: >> + * - Magazines and Vmem: Extending the Slab Allocator >> + * to Many CPUs and Arbitrary Resources >> + * http://www.usenix.org/event/usenix01/bonwick.html >> + */ >> + >> +#include <sys/cdefs.h> >> +__FBSDID("$FreeBSD$"); >> + >> +#include "opt_ddb.h" >> + >> +#include <sys/param.h> >> +#include <sys/systm.h> >> +#include <sys/kernel.h> >> +#include <sys/queue.h> >> +#include <sys/callout.h> >> +#include <sys/hash.h> >> +#include <sys/lock.h> >> +#include <sys/malloc.h> >> +#include <sys/mutex.h> >> +#include <sys/smp.h> >> +#include <sys/condvar.h> >> +#include <sys/taskqueue.h> >> +#include <sys/vmem.h> >> + >> +#include <vm/uma.h> >> +#include <vm/vm.h> >> +#include <vm/pmap.h> >> +#include <vm/vm_map.h> >> +#include <vm/vm_kern.h> >> +#include <vm/vm_extern.h> >> +#include <vm/vm_param.h> >> +#include <vm/vm_pageout.h> >> + >> +#define VMEM_MAXORDER (sizeof(vmem_size_t) * NBBY) >> + >> +#define VMEM_HASHSIZE_MIN 16 >> +#define VMEM_HASHSIZE_MAX 131072 >> + >> +#define VMEM_QCACHE_IDX_MAX 16 >> + >> +#define VMEM_FITMASK (M_BESTFIT | M_FIRSTFIT) >> + >> +#define VMEM_FLAGS \ >> + (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM | M_BESTFIT | M_FIRSTFIT) >> + >> +#define BT_FLAGS (M_NOWAIT | M_WAITOK | M_USE_RESERVE | M_NOVM) >> + >> +#define QC_NAME_MAX 16 >> + >> +/* >> + * Data structures private to vmem. >> + */ >> +MALLOC_DEFINE(M_VMEM, "vmem", "vmem internal structures"); >> + >> +typedef struct vmem_btag bt_t; >> + >> +TAILQ_HEAD(vmem_seglist, vmem_btag); >> +LIST_HEAD(vmem_freelist, vmem_btag); >> +LIST_HEAD(vmem_hashlist, vmem_btag); >> + >> +struct qcache { >> + uma_zone_t qc_cache; >> + vmem_t *qc_vmem; >> + vmem_size_t qc_size; >> + char qc_name[QC_NAME_MAX]; >> +}; >> +typedef struct qcache qcache_t; >> +#define QC_POOL_TO_QCACHE(pool) ((qcache_t *)(pool->pr_qcache)) >> + >> +#define VMEM_NAME_MAX 16 >> + >> +/* vmem arena */ >> +struct vmem { >> + struct mtx_padalign vm_lock; >> + struct cv vm_cv; >> + char vm_name[VMEM_NAME_MAX+1]; >> + LIST_ENTRY(vmem) vm_alllist; >> + struct vmem_hashlist vm_hash0[VMEM_HASHSIZE_MIN]; >> + struct vmem_freelist vm_freelist[VMEM_MAXORDER]; >> + struct vmem_seglist vm_seglist; >> + struct vmem_hashlist *vm_hashlist; >> + vmem_size_t vm_hashsize; >> + >> + /* Constant after init */ >> + vmem_size_t vm_qcache_max; >> + vmem_size_t vm_quantum_mask; >> + vmem_size_t vm_import_quantum; >> + int vm_quantum_shift; >> + >> + /* Written on alloc/free */ >> + LIST_HEAD(, vmem_btag) vm_freetags; >> + int vm_nfreetags; >> + int vm_nbusytag; >> + vmem_size_t vm_inuse; >> + vmem_size_t vm_size; >> + >> + /* Used on import. */ >> + vmem_import_t *vm_importfn; >> + vmem_release_t *vm_releasefn; >> + void *vm_arg; >> + >> + /* Space exhaustion callback. */ >> + vmem_reclaim_t *vm_reclaimfn; >> + >> + /* quantum cache */ >> + qcache_t vm_qcache[VMEM_QCACHE_IDX_MAX]; >> +}; >> + >> +/* boundary tag */ >> +struct vmem_btag { >> + TAILQ_ENTRY(vmem_btag) bt_seglist; >> + union { >> + LIST_ENTRY(vmem_btag) u_freelist; /* BT_TYPE_FREE */ >> + LIST_ENTRY(vmem_btag) u_hashlist; /* BT_TYPE_BUSY */ >> + } bt_u; >> +#define bt_hashlist bt_u.u_hashlist >> +#define bt_freelist bt_u.u_freelist >> + vmem_addr_t bt_start; >> + vmem_size_t bt_size; >> + int bt_type; >> +}; >> + >> +#define BT_TYPE_SPAN 1 /* Allocated from importfn */ >> +#define BT_TYPE_SPAN_STATIC 2 /* vmem_add() or create. */ >> +#define BT_TYPE_FREE 3 /* Available space. */ >> +#define BT_TYPE_BUSY 4 /* Used space. */ >> +#define BT_ISSPAN_P(bt) ((bt)->bt_type <= BT_TYPE_SPAN_STATIC) >> + >> +#define BT_END(bt) ((bt)->bt_start + (bt)->bt_size - 1) >> + >> +#if defined(DIAGNOSTIC) >> +static void vmem_check(vmem_t *); >> +#endif >> + >> +static struct callout vmem_periodic_ch; >> +static int vmem_periodic_interval; >> +static struct task vmem_periodic_wk; >> + >> +static struct mtx_padalign vmem_list_lock; >> +static LIST_HEAD(, vmem) vmem_list = LIST_HEAD_INITIALIZER(vmem_list); >> + >> +/* ---- misc */ >> +#define VMEM_CONDVAR_INIT(vm, wchan) cv_init(&vm->vm_cv, wchan) >> +#define VMEM_CONDVAR_DESTROY(vm) cv_destroy(&vm->vm_cv) >> +#define VMEM_CONDVAR_WAIT(vm) cv_wait(&vm->vm_cv, &vm->vm_lock) >> +#define VMEM_CONDVAR_BROADCAST(vm) cv_broadcast(&vm->vm_cv) >> + >> + >> +#define VMEM_LOCK(vm) mtx_lock(&vm->vm_lock) >> +#define VMEM_TRYLOCK(vm) mtx_trylock(&vm->vm_lock) >> +#define VMEM_UNLOCK(vm) mtx_unlock(&vm->vm_lock) >> +#define VMEM_LOCK_INIT(vm, name) mtx_init(&vm->vm_lock, (name), NULL, MTX_DEF) >> +#define VMEM_LOCK_DESTROY(vm) mtx_destroy(&vm->vm_lock) >> +#define VMEM_ASSERT_LOCKED(vm) mtx_assert(&vm->vm_lock, MA_OWNED); >> + >> +#define VMEM_ALIGNUP(addr, align) (-(-(addr) & -(align))) >> + >> +#define VMEM_CROSS_P(addr1, addr2, boundary) \ >> + ((((addr1) ^ (addr2)) & -(boundary)) != 0) >> + >> +#define ORDER2SIZE(order) ((vmem_size_t)1 << (order)) >> +#define SIZE2ORDER(size) ((int)flsl(size) - 1) >> + >> +/* >> + * Maximum number of boundary tags that may be required to satisfy an >> + * allocation. Two may be required to import. Another two may be >> + * required to clip edges. >> + */ >> +#define BT_MAXALLOC 4 >> + >> +/* >> + * Max free limits the number of locally cached boundary tags. We >> + * just want to avoid hitting the zone allocator for every call. >> + */ >> +#define BT_MAXFREE (BT_MAXALLOC * 8) >> + >> +/* Allocator for boundary tags. */ >> +static uma_zone_t vmem_bt_zone; >> + >> +/* boot time arena storage. */ >> +static struct vmem buffer_arena_storage; >> +static struct vmem transient_arena_storage; >> +vmem_t *buffer_arena = &buffer_arena_storage; >> +vmem_t *transient_arena = &transient_arena_storage; >> + >> +/* >> + * Fill the vmem's boundary tag cache. We guarantee that boundary tag >> + * allocation will not fail once bt_fill() passes. To do so we cache >> + * at least the maximum possible tag allocations in the arena. >> + */ >> +static int >> +bt_fill(vmem_t *vm, int flags) >> +{ >> + bt_t *bt; >> + >> + VMEM_ASSERT_LOCKED(vm); >> + >> + /* >> + * Loop until we meet the reserve. To minimize the lock shuffle >> + * and prevent simultaneous fills we first try a NOWAIT regardless >> + * of the caller's flags. Specify M_NOVM so we don't recurse while >> + * holding a vmem lock. >> + */ >> + while (vm->vm_nfreetags < BT_MAXALLOC) { >> + bt = uma_zalloc(vmem_bt_zone, >> + (flags & M_USE_RESERVE) | M_NOWAIT | M_NOVM); >> + if (bt == NULL) { >> + VMEM_UNLOCK(vm); >> + bt = uma_zalloc(vmem_bt_zone, flags); >> + VMEM_LOCK(vm); >> + if (bt == NULL && (flags & M_NOWAIT) != 0) >> + break; >> + } >> + LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist); >> + vm->vm_nfreetags++; >> + } >> + >> + if (vm->vm_nfreetags < BT_MAXALLOC) >> + return ENOMEM; >> + >> + return 0; >> +} >> + >> +/* >> + * Pop a tag off of the freetag stack. >> + */ >> +static bt_t * >> +bt_alloc(vmem_t *vm) >> +{ >> + bt_t *bt; >> + >> + VMEM_ASSERT_LOCKED(vm); >> + bt = LIST_FIRST(&vm->vm_freetags); >> + MPASS(bt != NULL); >> + LIST_REMOVE(bt, bt_freelist); >> + vm->vm_nfreetags--; >> + >> + return bt; >> +} >> + >> +/* >> + * Trim the per-vmem free list. Returns with the lock released to >> + * avoid allocator recursions. >> + */ >> +static void >> +bt_freetrim(vmem_t *vm, int freelimit) >> +{ >> + LIST_HEAD(, vmem_btag) freetags; >> + bt_t *bt; >> + >> + LIST_INIT(&freetags); >> + VMEM_ASSERT_LOCKED(vm); >> + while (vm->vm_nfreetags > freelimit) { >> + bt = LIST_FIRST(&vm->vm_freetags); >> + LIST_REMOVE(bt, bt_freelist); >> + vm->vm_nfreetags--; >> + LIST_INSERT_HEAD(&freetags, bt, bt_freelist); >> + } >> + VMEM_UNLOCK(vm); >> + while ((bt = LIST_FIRST(&freetags)) != NULL) { >> + LIST_REMOVE(bt, bt_freelist); >> + uma_zfree(vmem_bt_zone, bt); >> + } >> +} >> + >> +static inline void >> +bt_free(vmem_t *vm, bt_t *bt) >> +{ >> + >> + VMEM_ASSERT_LOCKED(vm); >> + MPASS(LIST_FIRST(&vm->vm_freetags) != bt); >> + LIST_INSERT_HEAD(&vm->vm_freetags, bt, bt_freelist); >> + vm->vm_nfreetags++; >> +} >> + >> +/* >> + * freelist[0] ... [1, 1] >> + * freelist[1] ... [2, 3] >> + * freelist[2] ... [4, 7] >> + * freelist[3] ... [8, 15] >> + * : >> + * freelist[n] ... [(1 << n), (1 << (n + 1)) - 1] >> + * : >> + */ >> + >> +static struct vmem_freelist * >> +bt_freehead_tofree(vmem_t *vm, vmem_size_t size) >> +{ >> + const vmem_size_t qsize = size >> vm->vm_quantum_shift; >> + const int idx = SIZE2ORDER(qsize); >> + >> + MPASS(size != 0 && qsize != 0); >> + MPASS((size & vm->vm_quantum_mask) == 0); >> + MPASS(idx >= 0); >> + MPASS(idx < VMEM_MAXORDER); >> + >> + return &vm->vm_freelist[idx]; >> +} >> + >> +/* >> + * bt_freehead_toalloc: return the freelist for the given size and allocation >> + * strategy. >> + * >> + * For M_FIRSTFIT, return the list in which any blocks are large enough >> + * for the requested size. otherwise, return the list which can have blocks >> + * large enough for the requested size. >> + */ >> +static struct vmem_freelist * >> +bt_freehead_toalloc(vmem_t *vm, vmem_size_t size, int strat) >> +{ >> + const vmem_size_t qsize = size >> vm->vm_quantum_shift; >> + int idx = SIZE2ORDER(qsize); >> + >> + MPASS(size != 0 && qsize != 0); >> + MPASS((size & vm->vm_quantum_mask) == 0); >> + >> + if (strat == M_FIRSTFIT && ORDER2SIZE(idx) != qsize) { >> + idx++; >> + /* check too large request? */ >> + } >> + MPASS(idx >= 0); >> + MPASS(idx < VMEM_MAXORDER); >> + >> + return &vm->vm_freelist[idx]; >> +} >> + >> +/* ---- boundary tag hash */ >> + >> +static struct vmem_hashlist * >> +bt_hashhead(vmem_t *vm, vmem_addr_t addr) >> +{ >> + struct vmem_hashlist *list; >> + unsigned int hash; >> + >> + hash = hash32_buf(&addr, sizeof(addr), 0); >> + list = &vm->vm_hashlist[hash % vm->vm_hashsize]; >> + >> + return list; >> +} >> + >> +static bt_t * >> +bt_lookupbusy(vmem_t *vm, vmem_addr_t addr) >> +{ >> + struct vmem_hashlist *list; >> + bt_t *bt; >> + >> + VMEM_ASSERT_LOCKED(vm); >> + list = bt_hashhead(vm, addr); >> + LIST_FOREACH(bt, list, bt_hashlist) { >> + if (bt->bt_start == addr) { >> + break; >> + } >> + } >> + >> + return bt; >> +} >> + >> +static void >> +bt_rembusy(vmem_t *vm, bt_t *bt) >> +{ >> + >> + VMEM_ASSERT_LOCKED(vm); >> + MPASS(vm->vm_nbusytag > 0); >> + vm->vm_inuse -= bt->bt_size; >> + vm->vm_nbusytag--; >> + LIST_REMOVE(bt, bt_hashlist); >> +} >> + >> +static void >> +bt_insbusy(vmem_t *vm, bt_t *bt) >> +{ >> + struct vmem_hashlist *list; >> + >> + VMEM_ASSERT_LOCKED(vm); >> + MPASS(bt->bt_type == BT_TYPE_BUSY); >> + >> + list = bt_hashhead(vm, bt->bt_start); >> + LIST_INSERT_HEAD(list, bt, bt_hashlist); >> + vm->vm_nbusytag++; >> + vm->vm_inuse += bt->bt_size; >> +} >> + >> +/* ---- boundary tag list */ >> + >> +static void >> +bt_remseg(vmem_t *vm, bt_t *bt) >> +{ >> + >> + TAILQ_REMOVE(&vm->vm_seglist, bt, bt_seglist); >> + bt_free(vm, bt); >> +} >> + >> +static void >> +bt_insseg(vmem_t *vm, bt_t *bt, bt_t *prev) >> +{ >> + >> + TAILQ_INSERT_AFTER(&vm->vm_seglist, prev, bt, bt_seglist); >> +} >> + >> +static void >> +bt_insseg_tail(vmem_t *vm, bt_t *bt) >> +{ >> + >> + TAILQ_INSERT_TAIL(&vm->vm_seglist, bt, bt_seglist); >> +} >> + >> +static void >> +bt_remfree(vmem_t *vm, bt_t *bt) >> +{ >> + >> + MPASS(bt->bt_type == BT_TYPE_FREE); >> + >> + LIST_REMOVE(bt, bt_freelist); >> +} >> + >> +static void >> +bt_insfree(vmem_t *vm, bt_t *bt) >> +{ >> + struct vmem_freelist *list; >> + >> + list = bt_freehead_tofree(vm, bt->bt_size); >> + LIST_INSERT_HEAD(list, bt, bt_freelist); >> +} >> + >> +/* ---- vmem internal functions */ >> + >> +/* >> + * Import from the arena into the quantum cache in UMA. >> + */ >> +static int >> +qc_import(void *arg, void **store, int cnt, int flags) >> +{ >> + qcache_t *qc; >> + vmem_addr_t addr; >> + int i; >> + >> + qc = arg; >> + flags |= M_BESTFIT; >> + for (i = 0; i < cnt; i++) { >> + if (vmem_xalloc(qc->qc_vmem, qc->qc_size, 0, 0, 0, >> + VMEM_ADDR_MIN, VMEM_ADDR_MAX, flags, &addr) != 0) >> + break; >> + store[i] = (void *)addr; >> + /* Only guarantee one allocation. */ >> + flags &= ~M_WAITOK; >> + flags |= M_NOWAIT; >> + } >> + return i; >> +} >> + >> +/* >> + * Release memory from the UMA cache to the arena. >> + */ >> +static void >> +qc_release(void *arg, void **store, int cnt) >> +{ >> + qcache_t *qc; >> + int i; >> + >> + qc = arg; >> + for (i = 0; i < cnt; i++) >> + vmem_xfree(qc->qc_vmem, (vmem_addr_t)store[i], qc->qc_size); >> +} >> + >> +static void >> +qc_init(vmem_t *vm, vmem_size_t qcache_max) >> +{ >> + qcache_t *qc; >> + vmem_size_t size; >> + int qcache_idx_max; >> + int i; >> + >> + MPASS((qcache_max & vm->vm_quantum_mask) == 0); >> + qcache_idx_max = MIN(qcache_max >> vm->vm_quantum_shift, >> + VMEM_QCACHE_IDX_MAX); >> + vm->vm_qcache_max = qcache_idx_max << vm->vm_quantum_shift; >> + for (i = 0; i < qcache_idx_max; i++) { >> + qc = &vm->vm_qcache[i]; >> + size = (i + 1) << vm->vm_quantum_shift; >> + snprintf(qc->qc_name, sizeof(qc->qc_name), "%s-%zu", >> + vm->vm_name, size); >> + qc->qc_vmem = vm; >> + qc->qc_size = size; >> + qc->qc_cache = uma_zcache_create(qc->qc_name, size, >> + NULL, NULL, NULL, NULL, qc_import, qc_release, qc, >> + UMA_ZONE_VM); >> + MPASS(qc->qc_cache); >> + } >> +} >> + >> +static void >> +qc_destroy(vmem_t *vm) >> +{ >> + int qcache_idx_max; >> + int i; >> + >> + qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift; >> + for (i = 0; i < qcache_idx_max; i++) >> + uma_zdestroy(vm->vm_qcache[i].qc_cache); >> +} >> + >> +static void >> +qc_drain(vmem_t *vm) >> +{ >> + int qcache_idx_max; >> + int i; >> + >> + qcache_idx_max = vm->vm_qcache_max >> vm->vm_quantum_shift; >> + for (i = 0; i < qcache_idx_max; i++) >> + zone_drain(vm->vm_qcache[i].qc_cache); >> +} >> + >> +void >> +vmem_startup(void) >> +{ >> + >> + mtx_init(&vmem_list_lock, "vmem list lock", NULL, MTX_DEF); >> + vmem_bt_zone = uma_zcreate("vmem btag", >> + sizeof(struct vmem_btag), NULL, NULL, NULL, NULL, >> + UMA_ALIGN_PTR, UMA_ZONE_VM); >> +} >> + >> +/* ---- rehash */ >> + >> +static int >> +vmem_rehash(vmem_t *vm, vmem_size_t newhashsize) >> +{ >> + bt_t *bt; >> + int i; >> + struct vmem_hashlist *newhashlist; >> + struct vmem_hashlist *oldhashlist; >> + vmem_size_t oldhashsize; >> + >> + MPASS(newhashsize > 0); >> + >> + newhashlist = malloc(sizeof(struct vmem_hashlist) * newhashsize, >> + M_VMEM, M_NOWAIT); >> + if (newhashlist == NULL) >> + return ENOMEM; >> + for (i = 0; i < newhashsize; i++) { >> + LIST_INIT(&newhashlist[i]); >> + } >> + >> + VMEM_LOCK(vm); >> + oldhashlist = vm->vm_hashlist; >> + oldhashsize = vm->vm_hashsize; >> + vm->vm_hashlist = newhashlist; >> + vm->vm_hashsize = newhashsize; >> + if (oldhashlist == NULL) { >> + VMEM_UNLOCK(vm); >> + return 0; >> + } >> + for (i = 0; i < oldhashsize; i++) { >> + while ((bt = LIST_FIRST(&oldhashlist[i])) != NULL) { >> + bt_rembusy(vm, bt); >> + bt_insbusy(vm, bt); >> + } >> + } >> + VMEM_UNLOCK(vm); >> + >> + if (oldhashlist != vm->vm_hash0) { >> + free(oldhashlist, M_VMEM); >> + } >> + >> + return 0; >> +} >> + >> +static void >> +vmem_periodic_kick(void *dummy) >> +{ >> + >> + taskqueue_enqueue(taskqueue_thread, &vmem_periodic_wk); >> +} >> + >> +static void >> +vmem_periodic(void *unused, int pending) >> +{ >> + vmem_t *vm; >> + vmem_size_t desired; >> + vmem_size_t current; >> + >> + mtx_lock(&vmem_list_lock); >> + LIST_FOREACH(vm, &vmem_list, vm_alllist) { >> +#ifdef DIAGNOSTIC >> + /* Convenient time to verify vmem state. */ >> + VMEM_LOCK(vm); >> + vmem_check(vm); >> + VMEM_UNLOCK(vm); >> +#endif >> + desired = 1 << flsl(vm->vm_nbusytag); >> + desired = MIN(MAX(desired, VMEM_HASHSIZE_MIN), >> + VMEM_HASHSIZE_MAX); >> + current = vm->vm_hashsize; >> + >> + /* Grow in powers of two. Shrink less aggressively. */ >> + if (desired >= current * 2 || desired * 4 <= current) >> + vmem_rehash(vm, desired); >> + } >> + mtx_unlock(&vmem_list_lock); >> + >> + callout_reset(&vmem_periodic_ch, vmem_periodic_interval, >> + vmem_periodic_kick, NULL); >> +} >> + >> +static void >> +vmem_start_callout(void *unused) >> +{ >> + >> + TASK_INIT(&vmem_periodic_wk, 0, vmem_periodic, NULL); >> + vmem_periodic_interval = hz * 10; >> + callout_init(&vmem_periodic_ch, CALLOUT_MPSAFE); >> + callout_reset(&vmem_periodic_ch, vmem_periodic_interval, >> + vmem_periodic_kick, NULL); >> +} >> +SYSINIT(vfs, SI_SUB_CONFIGURE, SI_ORDER_ANY, vmem_start_callout, NULL); >> + >> +static void >> +vmem_add1(vmem_t *vm, vmem_addr_t addr, vmem_size_t size, int flags, int type) >> +{ >> + bt_t *btspan; >> + bt_t *btfree; >> + >> + MPASS(type == BT_TYPE_SPAN || type == BT_TYPE_SPAN_STATIC); >> + >> + btspan = bt_alloc(vm); >> + btspan->bt_type = type; >> + btspan->bt_start = addr; >> + btspan->bt_size = size; >> + >> + btfree = bt_alloc(vm); >> + btfree->bt_type = BT_TYPE_FREE; >> + btfree->bt_start = addr; >> + btfree->bt_size = size; >> + >> + bt_insseg_tail(vm, btspan); >> + bt_insseg(vm, btfree, btspan); >> + bt_insfree(vm, btfree); >> + vm->vm_size += size; >> +} >> + >> +static void >> +vmem_destroy1(vmem_t *vm) >> +{ >> + bt_t *bt; >> + >> + /* >> + * Drain per-cpu quantum caches. >> + */ >> + qc_destroy(vm); >> + >> + /* >> + * The vmem should now only contain empty segments. >> + */ >> + VMEM_LOCK(vm); >> + MPASS(vm->vm_nbusytag == 0); >> + >> + while ((bt = TAILQ_FIRST(&vm->vm_seglist)) != NULL) >> + bt_remseg(vm, bt); >> + >> + if (vm->vm_hashlist != NULL && vm->vm_hashlist != vm->vm_hash0) >> + free(vm->vm_hashlist, M_VMEM); >> + >> + bt_freetrim(vm, 0); >> + >> + VMEM_CONDVAR_DESTROY(vm); >> + VMEM_LOCK_DESTROY(vm); >> + free(vm, M_VMEM); >> +} >> + >> +static int >> +vmem_import(vmem_t *vm, vmem_size_t size, int flags) >> +{ >> + vmem_addr_t addr; >> + int error; >> + >> + if (vm->vm_importfn == NULL) >> + return EINVAL; >> + >> + size = roundup(size, vm->vm_import_quantum); >> + >> + /* >> + * Hide MAXALLOC tags so we're guaranteed to be able to add this >> + * span and the tag we want to allocate from it. >> + */ >> + MPASS(vm->vm_nfreetags >= BT_MAXALLOC); >> + vm->vm_nfreetags -= BT_MAXALLOC; >> + VMEM_UNLOCK(vm); >> + error = (vm->vm_importfn)(vm->vm_arg, size, flags, &addr); >> + VMEM_LOCK(vm); >> + vm->vm_nfreetags += BT_MAXALLOC; >> + if (error) >> + return ENOMEM; >> + >> + vmem_add1(vm, addr, size, flags, BT_TYPE_SPAN); >> + >> + return 0; >> +} >> + >> +/* >> + * vmem_fit: check if a bt can satisfy the given restrictions. >> + * >> + * it's a caller's responsibility to ensure the region is big enough >> + * before calling us. >> + */ >> +static int >> +vmem_fit(const bt_t *bt, vmem_size_t size, vmem_size_t align, >> + vmem_size_t phase, vmem_size_t nocross, vmem_addr_t minaddr, >> + vmem_addr_t maxaddr, vmem_addr_t *addrp) >> +{ >> + vmem_addr_t start; >> + vmem_addr_t end; >> + >> + MPASS(size > 0); >> + MPASS(bt->bt_size >= size); /* caller's responsibility */ >> + >> + /* >> + * XXX assumption: vmem_addr_t and vmem_size_t are >> + * unsigned integer of the same size. >> + */ >> + >> + start = bt->bt_start; >> + if (start < minaddr) { >> + start = minaddr; >> + } >> + end = BT_END(bt); >> + if (end > maxaddr) >> + end = maxaddr; >> + if (start > end) >> + return (ENOMEM); >> + >> + start = VMEM_ALIGNUP(start - phase, align) + phase; >> + if (start < bt->bt_start) >> + start += align; >> + if (VMEM_CROSS_P(start, start + size - 1, nocross)) { >> + MPASS(align < nocross); >> + start = VMEM_ALIGNUP(start - phase, nocross) + phase; >> + } >> + if (start <= end && end - start >= size - 1) { >> + MPASS((start & (align - 1)) == phase); >> + MPASS(!VMEM_CROSS_P(start, start + size - 1, nocross)); >> + MPASS(minaddr <= start); >> + MPASS(maxaddr == 0 || start + size - 1 <= maxaddr); >> + MPASS(bt->bt_start <= start); >> + MPASS(BT_END(bt) - start >= size - 1); >> + *addrp = start; >> + >> + return (0); >> + } >> + return (ENOMEM); >> +} >> + >> +/* >> + * vmem_clip: Trim the boundary tag edges to the requested start and size. >> + */ >> +static void >> +vmem_clip(vmem_t *vm, bt_t *bt, vmem_addr_t start, vmem_size_t size) >> +{ >> + bt_t *btnew; >> + bt_t *btprev; >> + >> + VMEM_ASSERT_LOCKED(vm); >> + MPASS(bt->bt_type == BT_TYPE_FREE); >> + MPASS(bt->bt_size >= size); >> + bt_remfree(vm, bt); >> + if (bt->bt_start != start) { >> + btprev = bt_alloc(vm); >> + btprev->bt_type = BT_TYPE_FREE; >> + btprev->bt_start = bt->bt_start; >> + btprev->bt_size = start - bt->bt_start; >> + bt->bt_start = start; >> + bt->bt_size -= btprev->bt_size; >> + bt_insfree(vm, btprev); >> + bt_insseg(vm, btprev, >> + TAILQ_PREV(bt, vmem_seglist, bt_seglist)); >> + } >> + MPASS(bt->bt_start == start); >> + if (bt->bt_size != size && bt->bt_size - size > vm->vm_quantum_mask) { >> + /* split */ >> + btnew = bt_alloc(vm); >> + btnew->bt_type = BT_TYPE_BUSY; >> + btnew->bt_start = bt->bt_start; >> + btnew->bt_size = size; >> + bt->bt_start = bt->bt_start + size; >> + bt->bt_size -= size; >> + bt_insfree(vm, bt); >> + bt_insseg(vm, btnew, >> + TAILQ_PREV(bt, vmem_seglist, bt_seglist)); >> + bt_insbusy(vm, btnew); >> + bt = btnew; >> + } else { >> + bt->bt_type = BT_TYPE_BUSY; >> + bt_insbusy(vm, bt); >> + } >> + MPASS(bt->bt_size >= size); >> + bt->bt_type = BT_TYPE_BUSY; >> +} >> + >> +/* ---- vmem API */ >> + >> +void >> +vmem_set_import(vmem_t *vm, vmem_import_t *importfn, >> + vmem_release_t *releasefn, void *arg, vmem_size_t import_quantum) >> +{ >> + >> + VMEM_LOCK(vm); >> + vm->vm_importfn = importfn; >> + vm->vm_releasefn = releasefn; >> + vm->vm_arg = arg; >> + vm->vm_import_quantum = import_quantum; >> + VMEM_UNLOCK(vm); >> +} >> + >> +void >> +vmem_set_reclaim(vmem_t *vm, vmem_reclaim_t *reclaimfn) >> +{ >> + >> + VMEM_LOCK(vm); >> + vm->vm_reclaimfn = reclaimfn; >> + VMEM_UNLOCK(vm); >> +} >> + >> +/* >> + * vmem_init: Initializes vmem arena. >> + */ >> +vmem_t * >> +vmem_init(vmem_t *vm, const char *name, vmem_addr_t base, vmem_size_t size, >> + vmem_size_t quantum, vmem_size_t qcache_max, int flags) >> +{ >> + int i; >> + >> + MPASS(quantum > 0); >> + >> + bzero(vm, sizeof(*vm)); >> + >> + VMEM_CONDVAR_INIT(vm, name); >> + VMEM_LOCK_INIT(vm, name); >> + vm->vm_nfreetags = 0; >> + LIST_INIT(&vm->vm_freetags); >> + strlcpy(vm->vm_name, name, sizeof(vm->vm_name)); >> + vm->vm_quantum_mask = quantum - 1; >> + vm->vm_quantum_shift = SIZE2ORDER(quantum); >> + MPASS(ORDER2SIZE(vm->vm_quantum_shift) == quantum); >> + vm->vm_nbusytag = 0; >> + vm->vm_size = 0; >> + vm->vm_inuse = 0; >> + qc_init(vm, qcache_max); >> + >> + TAILQ_INIT(&vm->vm_seglist); >> + for (i = 0; i < VMEM_MAXORDER; i++) { >> + LIST_INIT(&vm->vm_freelist[i]); >> + } >> + memset(&vm->vm_hash0, 0, sizeof(vm->vm_hash0)); >> + vm->vm_hashsize = VMEM_HASHSIZE_MIN; >> + vm->vm_hashlist = vm->vm_hash0; >> + >> + if (size != 0) { >> + if (vmem_add(vm, base, size, flags) != 0) { >> + vmem_destroy1(vm); >> + return NULL; >> + } >> + } >> + >> + mtx_lock(&vmem_list_lock); >> + LIST_INSERT_HEAD(&vmem_list, vm, vm_alllist); >> + mtx_unlock(&vmem_list_lock); >> + >> + return vm; >> +} >> + >> +/* >> + * vmem_create: create an arena. >> + */ >> +vmem_t * >> +vmem_create(const char *name, vmem_addr_t base, vmem_size_t size, >> + vmem_size_t quantum, vmem_size_t qcache_max, int flags) >> +{ >> + >> + vmem_t *vm; >> + >> + vm = malloc(sizeof(*vm), M_VMEM, flags & (M_WAITOK|M_NOWAIT)); >> + if (vm == NULL) >> + return (NULL); >> + if (vmem_init(vm, name, base, size, quantum, qcache_max, >> + flags) == NULL) { >> + free(vm, M_VMEM); >> + return (NULL); >> + } >> + return (vm); >> +} >> + >> +void >> +vmem_destroy(vmem_t *vm) >> +{ >> + >> + mtx_lock(&vmem_list_lock); >> + LIST_REMOVE(vm, vm_alllist); >> + mtx_unlock(&vmem_list_lock); >> + >> + vmem_destroy1(vm); >> +} >> >> *** DIFF OUTPUT TRUNCATED AT 1000 LINES *** >
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?alpine.BSF.2.00.1306280032160.95115>