X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/9bccf70c0258c7cac2dcb80011b2a964d884c552..eee3565979933af707c711411001ba11fe406a3c:/bsd/vfs/vfs_bio.c?ds=inline diff --git a/bsd/vfs/vfs_bio.c b/bsd/vfs/vfs_bio.c index c11c03bea..f54c98779 100644 --- a/bsd/vfs/vfs_bio.c +++ b/bsd/vfs/vfs_bio.c @@ -1,23 +1,29 @@ /* - * Copyright (c) 2000-2002 Apple Computer, Inc. All rights reserved. + * Copyright (c) 2000-2016 Apple Inc. All rights reserved. * - * @APPLE_LICENSE_HEADER_START@ + * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ * - * The contents of this file constitute Original Code as defined in and - * are subject to the Apple Public Source License Version 1.1 (the - * "License"). You may not use this file except in compliance with the - * License. Please obtain a copy of the License at - * http://www.apple.com/publicsource and read it before using this file. + * This file contains Original Code and/or Modifications of Original Code + * as defined in and that are subject to the Apple Public Source License + * Version 2.0 (the 'License'). You may not use this file except in + * compliance with the License. The rights granted to you under the License + * may not be used to create, or enable the creation or redistribution of, + * unlawful or unlicensed copies of an Apple operating system, or to + * circumvent, violate, or enable the circumvention or violation of, any + * terms of an Apple operating system software license agreement. * - * This Original Code and all software distributed under the License are - * distributed on an "AS IS" basis, WITHOUT WARRANTY OF ANY KIND, EITHER + * Please obtain a copy of the License at + * http://www.opensource.apple.com/apsl/ and read it before using this file. + * + * The Original Code and all software distributed under the License are + * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES, * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE OR NON-INFRINGEMENT. Please see the - * License for the specific language governing rights and limitations - * under the License. + * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT. + * Please see the License for the specific language governing rights and + * limitations under the License. * - * @APPLE_LICENSE_HEADER_END@ + * @APPLE_OSREFERENCE_LICENSE_HEADER_END@ */ /* Copyright (c) 1995 NeXT Computer, Inc. All Rights Reserved */ /*- @@ -58,9 +64,6 @@ * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * - * The NEXTSTEP Software License Agreement specifies the terms - * and conditions for redistribution. - * * @(#)vfs_bio.c 8.6 (Berkeley) 1/11/94 */ @@ -73,45 +76,70 @@ #include #include -#include -#include -#include -#include +#include +#include +#include +#include #include #include #include #include #include -#include +#include #if DIAGNOSTIC #include #endif /* DIAGNOSTIC */ #include #include +#include +#include + +#include /* fslog_io_error() */ +#include /* dk_error_description_t */ + +#include +#include +#include /* thread_block() */ + +#include +#include #include -#include -static __inline__ void bufqinc(int q); -static __inline__ void bufqdec(int q); +#include +#include +#include + +#include + +int bcleanbuf(buf_t bp, boolean_t discard); +static int brecover_data(buf_t bp); +static boolean_t incore(vnode_t vp, daddr64_t blkno); +/* timeout is in msecs */ +static buf_t getnewbuf(int slpflag, int slptimeo, int *queue); +static void bremfree_locked(buf_t bp); +static void buf_reassign(buf_t bp, vnode_t newvp); +static errno_t buf_acquire_locked(buf_t bp, int flags, int slpflag, int slptimeo); +static int buf_iterprepare(vnode_t vp, struct buflists *, int flags); +static void buf_itercomplete(vnode_t vp, struct buflists *, int flags); +static boolean_t buffer_cache_gc(int); +static buf_t buf_brelse_shadow(buf_t bp); +static void buf_free_meta_store(buf_t bp); + +static buf_t buf_create_shadow_internal(buf_t bp, boolean_t force_copy, + uintptr_t external_storage, void (*iodone)(buf_t, void *), void *arg, int priv); -static struct buf *getnewbuf(int slpflag, int slptimeo, int *queue); -static int bcleanbuf(struct buf *bp); -extern void vwakeup(); -extern int niobuf; /* The number of IO buffer headers for cluster IO */ -int blaundrycnt; +int bdwrite_internal(buf_t, int); /* zone allocated buffer headers */ -static zone_t buf_hdr_zone; -static int buf_hdr_count; +static void bufzoneinit(void); +static void bcleanbuf_thread_init(void); +static void bcleanbuf_thread(void); + +static zone_t buf_hdr_zone; +static int buf_hdr_count; -#if TRACE -struct proc *traceproc; -int tracewhich, tracebuf[TRCSIZ]; -u_int tracex; -char traceflags[TR_NFLAGS]; -#endif /* TRACE */ /* * Definitions for the buffer hash lists. @@ -121,46 +149,53 @@ char traceflags[TR_NFLAGS]; LIST_HEAD(bufhashhdr, buf) *bufhashtbl, invalhash; u_long bufhash; +static buf_t incore_locked(vnode_t vp, daddr64_t blkno, struct bufhashhdr *dp); + /* Definitions for the buffer stats. */ struct bufstats bufstats; /* Number of delayed write buffers */ -int nbdwrite = 0; - -/* - * Insq/Remq for the buffer hash lists. - */ -#if 0 -#define binshash(bp, dp) LIST_INSERT_HEAD(dp, bp, b_hash) -#define bremhash(bp) LIST_REMOVE(bp, b_hash) -#endif /* 0 */ +long nbdwrite = 0; +int blaundrycnt = 0; +static int boot_nbuf_headers = 0; +static TAILQ_HEAD(delayqueue, buf) delaybufqueue; -TAILQ_HEAD(ioqueue, buf) iobufqueue; -TAILQ_HEAD(bqueues, buf) bufqueues[BQUEUES]; +static TAILQ_HEAD(ioqueue, buf) iobufqueue; +static TAILQ_HEAD(bqueues, buf) bufqueues[BQUEUES]; static int needbuffer; static int need_iobuffer; +static lck_grp_t *buf_mtx_grp; +static lck_attr_t *buf_mtx_attr; +static lck_grp_attr_t *buf_mtx_grp_attr; +static lck_mtx_t *iobuffer_mtxp; +static lck_mtx_t *buf_mtxp; + +static int buf_busycount; + +static __inline__ int +buf_timestamp(void) +{ + struct timeval t; + microuptime(&t); + return (t.tv_sec); +} + /* * Insq/Remq for the buffer free lists. */ #define binsheadfree(bp, dp, whichq) do { \ TAILQ_INSERT_HEAD(dp, bp, b_freelist); \ - bufqinc((whichq)); \ - (bp)->b_whichq = whichq; \ - (bp)->b_timestamp = time.tv_sec; \ } while (0) #define binstailfree(bp, dp, whichq) do { \ TAILQ_INSERT_TAIL(dp, bp, b_freelist); \ - bufqinc((whichq)); \ - (bp)->b_whichq = whichq; \ - (bp)->b_timestamp = time.tv_sec; \ } while (0) #define BHASHENTCHECK(bp) \ if ((bp)->b_hash.le_prev != (struct buf **)0xdeadbeef) \ - panic("%x: b_hash.le_prev is not deadbeef", (bp)); + panic("%p: b_hash.le_prev is not deadbeef", (bp)); #define BLISTNONE(bp) \ (bp)->b_hash.le_next = (struct buf *)0; \ @@ -175,11 +210,6 @@ static int need_iobuffer; (bp)->b_vnbufs.le_next = NOLIST; \ } -simple_lock_data_t bufhashlist_slock; /* lock on buffer hash list */ - -/* number of per vnode, "in flight" buffer writes */ -#define BUFWRITE_THROTTLE 9 - /* * Time in seconds before a buffer on a list is * considered as a stale buffer @@ -192,9 +222,11 @@ int lru_is_stale = LRU_IS_STALE; int age_is_stale = AGE_IS_STALE; int meta_is_stale = META_IS_STALE; +#define MAXLAUNDRY 10 + /* LIST_INSERT_HEAD() with assertions */ static __inline__ void -blistenterhead(struct bufhashhdr * head, struct buf * bp) +blistenterhead(struct bufhashhdr * head, buf_t bp) { if ((bp->b_hash.le_next = (head)->lh_first) != NULL) (head)->lh_first->b_hash.le_prev = &(bp)->b_hash.le_next; @@ -205,33 +237,28 @@ blistenterhead(struct bufhashhdr * head, struct buf * bp) } static __inline__ void -binshash(struct buf *bp, struct bufhashhdr *dp) +binshash(buf_t bp, struct bufhashhdr *dp) { - struct buf *nbp; - - simple_lock(&bufhashlist_slock); - -#if 0 - if(incore(bp->b_vp, bp->b_lblkno)) - panic("binshash: already incore"); -#endif /* 0 */ +#if DIAGNOSTIC + buf_t nbp; +#endif /* DIAGNOSTIC */ BHASHENTCHECK(bp); +#if DIAGNOSTIC nbp = dp->lh_first; for(; nbp != NULL; nbp = nbp->b_hash.le_next) { if(nbp == bp) panic("buf already in hashlist"); } +#endif /* DIAGNOSTIC */ blistenterhead(dp, bp); - simple_unlock(&bufhashlist_slock); } static __inline__ void -bremhash(struct buf *bp) +bremhash(buf_t bp) { - simple_lock(&bufhashlist_slock); if (bp->b_hash.le_prev == (struct buf **)0xdeadbeef) panic("bremhash le_prev is deadbeef"); if (bp->b_hash.le_next == bp) @@ -240,352 +267,1974 @@ bremhash(struct buf *bp) if (bp->b_hash.le_next != NULL) bp->b_hash.le_next->b_hash.le_prev = bp->b_hash.le_prev; *bp->b_hash.le_prev = (bp)->b_hash.le_next; - simple_unlock(&bufhashlist_slock); } /* - * Remove a buffer from the free list it's on + * buf_mtxp held. */ -void -bremfree(bp) - struct buf *bp; +static __inline__ void +bmovelaundry(buf_t bp) { - struct bqueues *dp = NULL; - int whichq = -1; + bp->b_whichq = BQ_LAUNDRY; + bp->b_timestamp = buf_timestamp(); + binstailfree(bp, &bufqueues[BQ_LAUNDRY], BQ_LAUNDRY); + blaundrycnt++; +} - /* - * We only calculate the head of the freelist when removing - * the last element of the list as that is the only time that - * it is needed (e.g. to reset the tail pointer). - * - * NB: This makes an assumption about how tailq's are implemented. - */ - if (bp->b_freelist.tqe_next == NULL) { - for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++) - if (dp->tqh_last == &bp->b_freelist.tqe_next) - break; - if (dp == &bufqueues[BQUEUES]) - panic("bremfree: lost tail"); +static __inline__ void +buf_release_credentials(buf_t bp) +{ + if (IS_VALID_CRED(bp->b_rcred)) { + kauth_cred_unref(&bp->b_rcred); + } + if (IS_VALID_CRED(bp->b_wcred)) { + kauth_cred_unref(&bp->b_wcred); } - TAILQ_REMOVE(dp, bp, b_freelist); - whichq = bp->b_whichq; - bufqdec(whichq); - bp->b_whichq = -1; - bp->b_timestamp = 0; } -/* - * Associate a buffer with a vnode. - */ -static void -bgetvp(vp, bp) - register struct vnode *vp; - register struct buf *bp; -{ - if (bp->b_vp != vp) - panic("bgetvp: not free"); - VHOLD(vp); - bp->b_vp = vp; - if (vp->v_type == VBLK || vp->v_type == VCHR) - bp->b_dev = vp->v_rdev; - else - bp->b_dev = NODEV; - /* - * Insert onto list for new vnode. - */ - bufinsvn(bp, &vp->v_cleanblkhd); +int +buf_valid(buf_t bp) { + + if ( (bp->b_flags & (B_DONE | B_DELWRI)) ) + return 1; + return 0; } -/* - * Disassociate a buffer from a vnode. - */ -static void -brelvp(bp) - register struct buf *bp; -{ - struct vnode *vp; +int +buf_fromcache(buf_t bp) { - if (bp->b_vp == (struct vnode *) 0) - panic("brelvp: NULL vp"); - /* - * Delete from old vnode list, if on one. - */ - if (bp->b_vnbufs.le_next != NOLIST) - bufremvn(bp); - vp = bp->b_vp; - bp->b_vp = (struct vnode *) 0; - HOLDRELE(vp); + if ( (bp->b_flags & B_CACHE) ) + return 1; + return 0; } -/* - * Reassign a buffer from one vnode to another. - * Used to assign file specific control information - * (indirect blocks) to the vnode to which they belong. - */ void -reassignbuf(bp, newvp) - register struct buf *bp; - register struct vnode *newvp; -{ - register struct buflists *listheadp; +buf_markinvalid(buf_t bp) { + + SET(bp->b_flags, B_INVAL); +} - if (newvp == NULL) { - printf("reassignbuf: NULL"); - return; +void +buf_markdelayed(buf_t bp) { + + if (!ISSET(bp->b_flags, B_DELWRI)) { + SET(bp->b_flags, B_DELWRI); + + OSAddAtomicLong(1, &nbdwrite); + buf_reassign(bp, bp->b_vp); } - /* - * Delete from old vnode list, if on one. - */ - if (bp->b_vnbufs.le_next != NOLIST) - bufremvn(bp); - /* - * If dirty, put on list of dirty buffers; - * otherwise insert onto list of clean buffers. - */ - if (ISSET(bp->b_flags, B_DELWRI)) - listheadp = &newvp->v_dirtyblkhd; - else - listheadp = &newvp->v_cleanblkhd; - bufinsvn(bp, listheadp); + SET(bp->b_flags, B_DONE); } -static __inline__ void -bufhdrinit(struct buf *bp) -{ - bzero((char *)bp, sizeof *bp); - bp->b_dev = NODEV; - bp->b_rcred = NOCRED; - bp->b_wcred = NOCRED; - bp->b_vnbufs.le_next = NOLIST; - bp->b_flags = B_INVAL; +void +buf_markclean(buf_t bp) { - return; + if (ISSET(bp->b_flags, B_DELWRI)) { + CLR(bp->b_flags, B_DELWRI); + + OSAddAtomicLong(-1, &nbdwrite); + buf_reassign(bp, bp->b_vp); + } } -/* - * Initialize buffers and hash links for buffers. - */ -__private_extern__ void -bufinit() -{ - register struct buf *bp; - register struct bqueues *dp; - register int i; - int metabuf; - long whichq; - static void bufzoneinit(); - static void bcleanbuf_thread_init(); +void +buf_markeintr(buf_t bp) { + + SET(bp->b_flags, B_EINTR); +} - /* Initialize the buffer queues ('freelists') and the hash table */ - for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++) - TAILQ_INIT(dp); - bufhashtbl = hashinit(nbuf, M_CACHE, &bufhash); - simple_lock_init(&bufhashlist_slock ); +void +buf_markaged(buf_t bp) { + + SET(bp->b_flags, B_AGE); +} - metabuf = nbuf/8; /* reserved for meta buf */ +int +buf_fua(buf_t bp) { - /* Initialize the buffer headers */ - for (i = 0; i < nbuf; i++) { - bp = &buf[i]; - bufhdrinit(bp); + if ((bp->b_flags & B_FUA) == B_FUA) + return 1; + return 0; +} - /* - * metabuf buffer headers on the meta-data list and - * rest of the buffer headers on the empty list - */ - if (--metabuf) - whichq = BQ_META; - else - whichq = BQ_EMPTY; +void +buf_markfua(buf_t bp) { - BLISTNONE(bp); - dp = &bufqueues[whichq]; - binsheadfree(bp, dp, whichq); - binshash(bp, &invalhash); - } + SET(bp->b_flags, B_FUA); +} - for (; i < nbuf + niobuf; i++) { - bp = &buf[i]; - bufhdrinit(bp); - binsheadfree(bp, &iobufqueue, -1); - } +#if CONFIG_PROTECT +cpx_t bufattr_cpx(bufattr_t bap) +{ + return bap->ba_cpx; +} - printf("using %d buffer headers and %d cluster IO buffer headers\n", - nbuf, niobuf); +void bufattr_setcpx(bufattr_t bap, cpx_t cpx) +{ + bap->ba_cpx = cpx; +} - /* Set up zones used by the buffer cache */ - bufzoneinit(); +void +buf_setcpoff (buf_t bp, uint64_t foffset) { + bp->b_attr.ba_cp_file_off = foffset; +} - /* start the bcleanbuf() thread */ - bcleanbuf_thread_init(); +uint64_t +bufattr_cpoff(bufattr_t bap) { + return bap->ba_cp_file_off; +} -#if 0 /* notyet */ - { - static void bufq_balance_thread_init(); - /* create a thread to do dynamic buffer queue balancing */ - bufq_balance_thread_init(); - } -#endif /* notyet */ +void +bufattr_setcpoff(bufattr_t bap, uint64_t foffset) { + bap->ba_cp_file_off = foffset; } -static struct buf * -bio_doread(vp, blkno, size, cred, async, queuetype) - struct vnode *vp; - daddr_t blkno; - int size; - struct ucred *cred; - int async; - int queuetype; +#else // !CONTECT_PROTECT + +uint64_t +bufattr_cpoff(bufattr_t bap __unused) { + return 0; +} + +void +bufattr_setcpoff(__unused bufattr_t bap, __unused uint64_t foffset) { + return; +} + +struct cpx *bufattr_cpx(__unused bufattr_t bap) { - register struct buf *bp; - struct proc *p = current_proc(); + return NULL; +} - bp = getblk(vp, blkno, size, 0, 0, queuetype); +void bufattr_setcpx(__unused bufattr_t bap, __unused struct cpx *cpx) +{ +} - /* - * If buffer does not have data valid, start a read. - * Note that if buffer is B_INVAL, getblk() won't return it. - * Therefore, it's valid if it's I/O has completed or been delayed. - */ - if (!ISSET(bp->b_flags, (B_DONE | B_DELWRI))) { - /* Start I/O for the buffer (keeping credentials). */ - SET(bp->b_flags, B_READ | async); - if (cred != NOCRED && bp->b_rcred == NOCRED) { - /* - * NFS has embedded ucred. - * Can not crhold() here as that causes zone corruption - */ - bp->b_rcred = crdup(cred); - } - VOP_STRATEGY(bp); +#endif /* !CONFIG_PROTECT */ - trace(TR_BREADMISS, pack(vp, size), blkno); +bufattr_t +bufattr_alloc() { + bufattr_t bap; + MALLOC(bap, bufattr_t, sizeof(struct bufattr), M_TEMP, M_WAITOK); + if (bap == NULL) + return NULL; - /* Pay for the read. */ - if (p && p->p_stats) - p->p_stats->p_ru.ru_inblock++; /* XXX */ - } else if (async) { - brelse(bp); - } + bzero(bap, sizeof(struct bufattr)); + return bap; +} - trace(TR_BREADHIT, pack(vp, size), blkno); +void +bufattr_free(bufattr_t bap) { + if (bap) + FREE(bap, M_TEMP); +} - return (bp); +bufattr_t +bufattr_dup(bufattr_t bap) { + bufattr_t new_bufattr; + MALLOC(new_bufattr, bufattr_t, sizeof(struct bufattr), M_TEMP, M_WAITOK); + if (new_bufattr == NULL) + return NULL; + + /* Copy the provided one into the new copy */ + memcpy (new_bufattr, bap, sizeof(struct bufattr)); + return new_bufattr; } -/* - * Read a disk block. - * This algorithm described in Bach (p.54). - */ + int -bread(vp, blkno, size, cred, bpp) - struct vnode *vp; - daddr_t blkno; - int size; - struct ucred *cred; - struct buf **bpp; -{ - register struct buf *bp; +bufattr_rawencrypted(bufattr_t bap) { + if ( (bap->ba_flags & BA_RAW_ENCRYPTED_IO) ) + return 1; + return 0; +} - /* Get buffer for block. */ - bp = *bpp = bio_doread(vp, blkno, size, cred, 0, BLK_READ); +int +bufattr_throttled(bufattr_t bap) { + return (GET_BUFATTR_IO_TIER(bap)); +} - /* Wait for the read to complete, and return result. */ - return (biowait(bp)); +int +bufattr_passive(bufattr_t bap) { + if ( (bap->ba_flags & BA_PASSIVE) ) + return 1; + return 0; } -/* - * Read a disk block. [bread() for meta-data] - * This algorithm described in Bach (p.54). - */ int -meta_bread(vp, blkno, size, cred, bpp) - struct vnode *vp; - daddr_t blkno; - int size; - struct ucred *cred; - struct buf **bpp; -{ - register struct buf *bp; +bufattr_nocache(bufattr_t bap) { + if ( (bap->ba_flags & BA_NOCACHE) ) + return 1; + return 0; +} - /* Get buffer for block. */ - bp = *bpp = bio_doread(vp, blkno, size, cred, 0, BLK_META); +int +bufattr_meta(bufattr_t bap) { + if ( (bap->ba_flags & BA_META) ) + return 1; + return 0; +} - /* Wait for the read to complete, and return result. */ - return (biowait(bp)); +void +bufattr_markmeta(bufattr_t bap) { + SET(bap->ba_flags, BA_META); } -/* - * Read-ahead multiple disk blocks. The first is sync, the rest async. - * Trivial modification to the breada algorithm presented in Bach (p.55). - */ int -breadn(vp, blkno, size, rablks, rasizes, nrablks, cred, bpp) - struct vnode *vp; - daddr_t blkno; int size; - daddr_t rablks[]; int rasizes[]; - int nrablks; - struct ucred *cred; - struct buf **bpp; -{ - register struct buf *bp; - int i; +bufattr_delayidlesleep(bufattr_t bap) +{ + if ( (bap->ba_flags & BA_DELAYIDLESLEEP) ) + return 1; + return 0; +} - bp = *bpp = bio_doread(vp, blkno, size, cred, 0, BLK_READ); +bufattr_t +buf_attr(buf_t bp) { + return &bp->b_attr; +} - /* - * For each of the read-ahead blocks, start a read, if necessary. - */ - for (i = 0; i < nrablks; i++) { - /* If it's in the cache, just go on to next one. */ - if (incore(vp, rablks[i])) - continue; +void +buf_markstatic(buf_t bp __unused) { + SET(bp->b_flags, B_STATICCONTENT); +} - /* Get a buffer for the read-ahead block */ - (void) bio_doread(vp, rablks[i], rasizes[i], cred, B_ASYNC, BLK_READ); - } +int +buf_static(buf_t bp) { + if ( (bp->b_flags & B_STATICCONTENT) ) + return 1; + return 0; +} - /* Otherwise, we had to start a read for it; wait until it's valid. */ - return (biowait(bp)); +void +bufattr_markgreedymode(bufattr_t bap) { + SET(bap->ba_flags, BA_GREEDY_MODE); } -/* - * Read with single-block read-ahead. Defined in Bach (p.55), but - * implemented as a call to breadn(). - * XXX for compatibility with old file systems. - */ int -breada(vp, blkno, size, rablkno, rabsize, cred, bpp) - struct vnode *vp; - daddr_t blkno; int size; - daddr_t rablkno; int rabsize; - struct ucred *cred; - struct buf **bpp; -{ +bufattr_greedymode(bufattr_t bap) { + if ( (bap->ba_flags & BA_GREEDY_MODE) ) + return 1; + return 0; +} - return (breadn(vp, blkno, size, &rablkno, &rabsize, 1, cred, bpp)); +void +bufattr_markisochronous(bufattr_t bap) { + SET(bap->ba_flags, BA_ISOCHRONOUS); } -/* - * Block write. Described in Bach (p.56) - */ int -bwrite(bp) - struct buf *bp; -{ - int rv, sync, wasdelayed; - struct proc *p = current_proc(); - struct vnode *vp = bp->b_vp; +bufattr_isochronous(bufattr_t bap) { + if ( (bap->ba_flags & BA_ISOCHRONOUS) ) + return 1; + return 0; +} - /* Remember buffer type, to switch on it later. */ - sync = !ISSET(bp->b_flags, B_ASYNC); - wasdelayed = ISSET(bp->b_flags, B_DELWRI); - CLR(bp->b_flags, (B_READ | B_DONE | B_ERROR | B_DELWRI)); - if (wasdelayed) { - nbdwrite--; - wakeup((caddr_t)&nbdwrite); +void +bufattr_markquickcomplete(bufattr_t bap) { + SET(bap->ba_flags, BA_QUICK_COMPLETE); +} + +int +bufattr_quickcomplete(bufattr_t bap) { + if ( (bap->ba_flags & BA_QUICK_COMPLETE) ) + return 1; + return 0; +} + +errno_t +buf_error(buf_t bp) { + + return (bp->b_error); +} + +void +buf_seterror(buf_t bp, errno_t error) { + + if ((bp->b_error = error)) + SET(bp->b_flags, B_ERROR); + else + CLR(bp->b_flags, B_ERROR); +} + +void +buf_setflags(buf_t bp, int32_t flags) { + + SET(bp->b_flags, (flags & BUF_X_WRFLAGS)); +} + +void +buf_clearflags(buf_t bp, int32_t flags) { + + CLR(bp->b_flags, (flags & BUF_X_WRFLAGS)); +} + +int32_t +buf_flags(buf_t bp) { + + return ((bp->b_flags & BUF_X_RDFLAGS)); +} + +void +buf_reset(buf_t bp, int32_t io_flags) { + + CLR(bp->b_flags, (B_READ | B_WRITE | B_ERROR | B_DONE | B_INVAL | B_ASYNC | B_NOCACHE | B_FUA)); + SET(bp->b_flags, (io_flags & (B_ASYNC | B_READ | B_WRITE | B_NOCACHE))); + + bp->b_error = 0; +} + +uint32_t +buf_count(buf_t bp) { + + return (bp->b_bcount); +} + +void +buf_setcount(buf_t bp, uint32_t bcount) { + + bp->b_bcount = bcount; +} + +uint32_t +buf_size(buf_t bp) { + + return (bp->b_bufsize); +} + +void +buf_setsize(buf_t bp, uint32_t bufsize) { + + bp->b_bufsize = bufsize; +} + +uint32_t +buf_resid(buf_t bp) { + + return (bp->b_resid); +} + +void +buf_setresid(buf_t bp, uint32_t resid) { + + bp->b_resid = resid; +} + +uint32_t +buf_dirtyoff(buf_t bp) { + + return (bp->b_dirtyoff); +} + +uint32_t +buf_dirtyend(buf_t bp) { + + return (bp->b_dirtyend); +} + +void +buf_setdirtyoff(buf_t bp, uint32_t dirtyoff) { + + bp->b_dirtyoff = dirtyoff; +} + +void +buf_setdirtyend(buf_t bp, uint32_t dirtyend) { + + bp->b_dirtyend = dirtyend; +} + +uintptr_t +buf_dataptr(buf_t bp) { + + return (bp->b_datap); +} + +void +buf_setdataptr(buf_t bp, uintptr_t data) { + + bp->b_datap = data; +} + +vnode_t +buf_vnode(buf_t bp) { + + return (bp->b_vp); +} + +void +buf_setvnode(buf_t bp, vnode_t vp) { + + bp->b_vp = vp; +} + + +void * +buf_callback(buf_t bp) +{ + if ( !(bp->b_flags & B_CALL) ) + return ((void *) NULL); + + return ((void *)bp->b_iodone); +} + + +errno_t +buf_setcallback(buf_t bp, void (*callback)(buf_t, void *), void *transaction) +{ + assert(!ISSET(bp->b_flags, B_FILTER) && ISSET(bp->b_lflags, BL_BUSY)); + + if (callback) + bp->b_flags |= (B_CALL | B_ASYNC); + else + bp->b_flags &= ~B_CALL; + bp->b_transaction = transaction; + bp->b_iodone = callback; + + return (0); +} + +errno_t +buf_setupl(buf_t bp, upl_t upl, uint32_t offset) +{ + + if ( !(bp->b_lflags & BL_IOBUF) ) + return (EINVAL); + + if (upl) + bp->b_flags |= B_CLUSTER; + else + bp->b_flags &= ~B_CLUSTER; + bp->b_upl = upl; + bp->b_uploffset = offset; + + return (0); +} + +buf_t +buf_clone(buf_t bp, int io_offset, int io_size, void (*iodone)(buf_t, void *), void *arg) +{ + buf_t io_bp; + + if (io_offset < 0 || io_size < 0) + return (NULL); + + if ((unsigned)(io_offset + io_size) > (unsigned)bp->b_bcount) + return (NULL); + + if (bp->b_flags & B_CLUSTER) { + if (io_offset && ((bp->b_uploffset + io_offset) & PAGE_MASK)) + return (NULL); + + if (((bp->b_uploffset + io_offset + io_size) & PAGE_MASK) && ((io_offset + io_size) < bp->b_bcount)) + return (NULL); + } + io_bp = alloc_io_buf(bp->b_vp, 0); + + io_bp->b_flags = bp->b_flags & (B_COMMIT_UPL | B_META | B_PAGEIO | B_CLUSTER | B_PHYS | B_RAW | B_ASYNC | B_READ | B_FUA); + + if (iodone) { + io_bp->b_transaction = arg; + io_bp->b_iodone = iodone; + io_bp->b_flags |= B_CALL; + } + if (bp->b_flags & B_CLUSTER) { + io_bp->b_upl = bp->b_upl; + io_bp->b_uploffset = bp->b_uploffset + io_offset; + } else { + io_bp->b_datap = (uintptr_t)(((char *)bp->b_datap) + io_offset); + } + io_bp->b_bcount = io_size; + + return (io_bp); +} + + +int +buf_shadow(buf_t bp) +{ + if (bp->b_lflags & BL_SHADOW) + return 1; + return 0; +} + + +buf_t +buf_create_shadow_priv(buf_t bp, boolean_t force_copy, uintptr_t external_storage, void (*iodone)(buf_t, void *), void *arg) +{ + return (buf_create_shadow_internal(bp, force_copy, external_storage, iodone, arg, 1)); +} + +buf_t +buf_create_shadow(buf_t bp, boolean_t force_copy, uintptr_t external_storage, void (*iodone)(buf_t, void *), void *arg) +{ + return (buf_create_shadow_internal(bp, force_copy, external_storage, iodone, arg, 0)); +} + + +static buf_t +buf_create_shadow_internal(buf_t bp, boolean_t force_copy, uintptr_t external_storage, void (*iodone)(buf_t, void *), void *arg, int priv) +{ + buf_t io_bp; + + KERNEL_DEBUG(0xbbbbc000 | DBG_FUNC_START, bp, 0, 0, 0, 0); + + if ( !(bp->b_flags & B_META) || (bp->b_lflags & BL_IOBUF)) { + + KERNEL_DEBUG(0xbbbbc000 | DBG_FUNC_END, bp, 0, 0, 0, 0); + return (NULL); + } +#ifdef BUF_MAKE_PRIVATE + if (bp->b_shadow_ref && bp->b_data_ref == 0 && external_storage == 0) + panic("buf_create_shadow: %p is in the private state (%d, %d)", bp, bp->b_shadow_ref, bp->b_data_ref); +#endif + io_bp = alloc_io_buf(bp->b_vp, priv); + + io_bp->b_flags = bp->b_flags & (B_META | B_ZALLOC | B_ASYNC | B_READ | B_FUA); + io_bp->b_blkno = bp->b_blkno; + io_bp->b_lblkno = bp->b_lblkno; + + if (iodone) { + io_bp->b_transaction = arg; + io_bp->b_iodone = iodone; + io_bp->b_flags |= B_CALL; + } + if (force_copy == FALSE) { + io_bp->b_bcount = bp->b_bcount; + io_bp->b_bufsize = bp->b_bufsize; + + if (external_storage) { + io_bp->b_datap = external_storage; +#ifdef BUF_MAKE_PRIVATE + io_bp->b_data_store = NULL; +#endif + } else { + io_bp->b_datap = bp->b_datap; +#ifdef BUF_MAKE_PRIVATE + io_bp->b_data_store = bp; +#endif + } + *(buf_t *)(&io_bp->b_orig) = bp; + + lck_mtx_lock_spin(buf_mtxp); + + io_bp->b_lflags |= BL_SHADOW; + io_bp->b_shadow = bp->b_shadow; + bp->b_shadow = io_bp; + bp->b_shadow_ref++; + +#ifdef BUF_MAKE_PRIVATE + if (external_storage) + io_bp->b_lflags |= BL_EXTERNAL; + else + bp->b_data_ref++; +#endif + lck_mtx_unlock(buf_mtxp); + } else { + if (external_storage) { +#ifdef BUF_MAKE_PRIVATE + io_bp->b_lflags |= BL_EXTERNAL; +#endif + io_bp->b_bcount = bp->b_bcount; + io_bp->b_bufsize = bp->b_bufsize; + io_bp->b_datap = external_storage; + } else { + allocbuf(io_bp, bp->b_bcount); + + io_bp->b_lflags |= BL_IOBUF_ALLOC; + } + bcopy((caddr_t)bp->b_datap, (caddr_t)io_bp->b_datap, bp->b_bcount); + +#ifdef BUF_MAKE_PRIVATE + io_bp->b_data_store = NULL; +#endif + } + KERNEL_DEBUG(0xbbbbc000 | DBG_FUNC_END, bp, bp->b_shadow_ref, 0, io_bp, 0); + + return (io_bp); +} + + +#ifdef BUF_MAKE_PRIVATE +errno_t +buf_make_private(buf_t bp) +{ + buf_t ds_bp; + buf_t t_bp; + struct buf my_buf; + + KERNEL_DEBUG(0xbbbbc004 | DBG_FUNC_START, bp, bp->b_shadow_ref, 0, 0, 0); + + if (bp->b_shadow_ref == 0 || bp->b_data_ref == 0 || ISSET(bp->b_lflags, BL_SHADOW)) { + + KERNEL_DEBUG(0xbbbbc004 | DBG_FUNC_END, bp, bp->b_shadow_ref, 0, EINVAL, 0); + return (EINVAL); + } + my_buf.b_flags = B_META; + my_buf.b_datap = (uintptr_t)NULL; + allocbuf(&my_buf, bp->b_bcount); + + bcopy((caddr_t)bp->b_datap, (caddr_t)my_buf.b_datap, bp->b_bcount); + + lck_mtx_lock_spin(buf_mtxp); + + for (t_bp = bp->b_shadow; t_bp; t_bp = t_bp->b_shadow) { + if ( !ISSET(bp->b_lflags, BL_EXTERNAL)) + break; + } + ds_bp = t_bp; + + if (ds_bp == NULL && bp->b_data_ref) + panic("buf_make_private: b_data_ref != 0 && ds_bp == NULL"); + + if (ds_bp && (bp->b_data_ref == 0 || bp->b_shadow_ref == 0)) + panic("buf_make_private: ref_count == 0 && ds_bp != NULL"); + + if (ds_bp == NULL) { + lck_mtx_unlock(buf_mtxp); + + buf_free_meta_store(&my_buf); + + KERNEL_DEBUG(0xbbbbc004 | DBG_FUNC_END, bp, bp->b_shadow_ref, 0, EINVAL, 0); + return (EINVAL); + } + for (t_bp = bp->b_shadow; t_bp; t_bp = t_bp->b_shadow) { + if ( !ISSET(t_bp->b_lflags, BL_EXTERNAL)) + t_bp->b_data_store = ds_bp; + } + ds_bp->b_data_ref = bp->b_data_ref; + + bp->b_data_ref = 0; + bp->b_datap = my_buf.b_datap; + + lck_mtx_unlock(buf_mtxp); + + KERNEL_DEBUG(0xbbbbc004 | DBG_FUNC_END, bp, bp->b_shadow_ref, 0, 0, 0); + return (0); +} +#endif + + +void +buf_setfilter(buf_t bp, void (*filter)(buf_t, void *), void *transaction, + void (**old_iodone)(buf_t, void *), void **old_transaction) +{ + assert(ISSET(bp->b_lflags, BL_BUSY)); + + if (old_iodone) + *old_iodone = bp->b_iodone; + if (old_transaction) + *old_transaction = bp->b_transaction; + + bp->b_transaction = transaction; + bp->b_iodone = filter; + if (filter) + bp->b_flags |= B_FILTER; + else + bp->b_flags &= ~B_FILTER; +} + + +daddr64_t +buf_blkno(buf_t bp) { + + return (bp->b_blkno); +} + +daddr64_t +buf_lblkno(buf_t bp) { + + return (bp->b_lblkno); +} + +void +buf_setblkno(buf_t bp, daddr64_t blkno) { + + bp->b_blkno = blkno; +} + +void +buf_setlblkno(buf_t bp, daddr64_t lblkno) { + + bp->b_lblkno = lblkno; +} + +dev_t +buf_device(buf_t bp) { + + return (bp->b_dev); +} + +errno_t +buf_setdevice(buf_t bp, vnode_t vp) { + + if ((vp->v_type != VBLK) && (vp->v_type != VCHR)) + return EINVAL; + bp->b_dev = vp->v_rdev; + + return 0; +} + + +void * +buf_drvdata(buf_t bp) { + + return (bp->b_drvdata); +} + +void +buf_setdrvdata(buf_t bp, void *drvdata) { + + bp->b_drvdata = drvdata; +} + +void * +buf_fsprivate(buf_t bp) { + + return (bp->b_fsprivate); +} + +void +buf_setfsprivate(buf_t bp, void *fsprivate) { + + bp->b_fsprivate = fsprivate; +} + +kauth_cred_t +buf_rcred(buf_t bp) { + + return (bp->b_rcred); +} + +kauth_cred_t +buf_wcred(buf_t bp) { + + return (bp->b_wcred); +} + +void * +buf_upl(buf_t bp) { + + return (bp->b_upl); +} + +uint32_t +buf_uploffset(buf_t bp) { + + return ((uint32_t)(bp->b_uploffset)); +} + +proc_t +buf_proc(buf_t bp) { + + return (bp->b_proc); +} + + +errno_t +buf_map(buf_t bp, caddr_t *io_addr) +{ + buf_t real_bp; + vm_offset_t vaddr; + kern_return_t kret; + + if ( !(bp->b_flags & B_CLUSTER)) { + *io_addr = (caddr_t)bp->b_datap; + return (0); + } + real_bp = (buf_t)(bp->b_real_bp); + + if (real_bp && real_bp->b_datap) { + /* + * b_real_bp is only valid if B_CLUSTER is SET + * if it's non-zero, than someone did a cluster_bp call + * if the backing physical pages were already mapped + * in before the call to cluster_bp (non-zero b_datap), + * than we just use that mapping + */ + *io_addr = (caddr_t)real_bp->b_datap; + return (0); + } + kret = ubc_upl_map(bp->b_upl, &vaddr); /* Map it in */ + + if (kret != KERN_SUCCESS) { + *io_addr = NULL; + + return(ENOMEM); + } + vaddr += bp->b_uploffset; + + *io_addr = (caddr_t)vaddr; + + return (0); +} + +errno_t +buf_unmap(buf_t bp) +{ + buf_t real_bp; + kern_return_t kret; + + if ( !(bp->b_flags & B_CLUSTER)) + return (0); + /* + * see buf_map for the explanation + */ + real_bp = (buf_t)(bp->b_real_bp); + + if (real_bp && real_bp->b_datap) + return (0); + + if ((bp->b_lflags & BL_IOBUF) && + ((bp->b_flags & (B_PAGEIO | B_READ)) != (B_PAGEIO | B_READ))) { + /* + * ignore pageins... the 'right' thing will + * happen due to the way we handle speculative + * clusters... + * + * when we commit these pages, we'll hit + * it with UPL_COMMIT_INACTIVE which + * will clear the reference bit that got + * turned on when we touched the mapping + */ + bp->b_flags |= B_AGE; + } + kret = ubc_upl_unmap(bp->b_upl); + + if (kret != KERN_SUCCESS) + return (EINVAL); + return (0); +} + + +void +buf_clear(buf_t bp) { + caddr_t baddr; + + if (buf_map(bp, &baddr) == 0) { + bzero(baddr, bp->b_bcount); + buf_unmap(bp); + } + bp->b_resid = 0; +} + +/* + * Read or write a buffer that is not contiguous on disk. + * buffer is marked done/error at the conclusion + */ +static int +buf_strategy_fragmented(vnode_t devvp, buf_t bp, off_t f_offset, size_t contig_bytes) +{ + vnode_t vp = buf_vnode(bp); + buf_t io_bp; /* For reading or writing a single block */ + int io_direction; + int io_resid; + size_t io_contig_bytes; + daddr64_t io_blkno; + int error = 0; + int bmap_flags; + + /* + * save our starting point... the bp was already mapped + * in buf_strategy before we got called + * no sense doing it again. + */ + io_blkno = bp->b_blkno; + /* + * Make sure we redo this mapping for the next I/O + * i.e. this can never be a 'permanent' mapping + */ + bp->b_blkno = bp->b_lblkno; + + /* + * Get an io buffer to do the deblocking + */ + io_bp = alloc_io_buf(devvp, 0); + + io_bp->b_lblkno = bp->b_lblkno; + io_bp->b_datap = bp->b_datap; + io_resid = bp->b_bcount; + io_direction = bp->b_flags & B_READ; + io_contig_bytes = contig_bytes; + + if (bp->b_flags & B_READ) + bmap_flags = VNODE_READ; + else + bmap_flags = VNODE_WRITE; + + for (;;) { + if (io_blkno == -1) + /* + * this is unexepected, but we'll allow for it + */ + bzero((caddr_t)io_bp->b_datap, (int)io_contig_bytes); + else { + io_bp->b_bcount = io_contig_bytes; + io_bp->b_bufsize = io_contig_bytes; + io_bp->b_resid = io_contig_bytes; + io_bp->b_blkno = io_blkno; + + buf_reset(io_bp, io_direction); + + /* + * Call the device to do the I/O and wait for it. Make sure the appropriate party is charged for write + */ + + if (!ISSET(bp->b_flags, B_READ)) + OSAddAtomic(1, &devvp->v_numoutput); + + if ((error = VNOP_STRATEGY(io_bp))) + break; + if ((error = (int)buf_biowait(io_bp))) + break; + if (io_bp->b_resid) { + io_resid -= (io_contig_bytes - io_bp->b_resid); + break; + } + } + if ((io_resid -= io_contig_bytes) == 0) + break; + f_offset += io_contig_bytes; + io_bp->b_datap += io_contig_bytes; + + /* + * Map the current position to a physical block number + */ + if ((error = VNOP_BLOCKMAP(vp, f_offset, io_resid, &io_blkno, &io_contig_bytes, NULL, bmap_flags, NULL))) + break; + } + buf_free(io_bp); + + if (error) + buf_seterror(bp, error); + bp->b_resid = io_resid; + /* + * This I/O is now complete + */ + buf_biodone(bp); + + return error; +} + + +/* + * struct vnop_strategy_args { + * struct buf *a_bp; + * } *ap; + */ +errno_t +buf_strategy(vnode_t devvp, void *ap) +{ + buf_t bp = ((struct vnop_strategy_args *)ap)->a_bp; + vnode_t vp = bp->b_vp; + int bmap_flags; + errno_t error; +#if CONFIG_DTRACE + int dtrace_io_start_flag = 0; /* We only want to trip the io:::start + * probe once, with the true physical + * block in place (b_blkno) + */ + +#endif + + if (vp == NULL || vp->v_type == VCHR || vp->v_type == VBLK) + panic("buf_strategy: b_vp == NULL || vtype == VCHR | VBLK\n"); + /* + * associate the physical device with + * with this buf_t even if we don't + * end up issuing the I/O... + */ + bp->b_dev = devvp->v_rdev; + + if (bp->b_flags & B_READ) + bmap_flags = VNODE_READ; + else + bmap_flags = VNODE_WRITE; + + if ( !(bp->b_flags & B_CLUSTER)) { + + if ( (bp->b_upl) ) { + /* + * we have a UPL associated with this bp + * go through cluster_bp which knows how + * to deal with filesystem block sizes + * that aren't equal to the page size + */ + DTRACE_IO1(start, buf_t, bp); + return (cluster_bp(bp)); + } + if (bp->b_blkno == bp->b_lblkno) { + off_t f_offset; + size_t contig_bytes; + + if ((error = VNOP_BLKTOOFF(vp, bp->b_lblkno, &f_offset))) { + DTRACE_IO1(start, buf_t, bp); + buf_seterror(bp, error); + buf_biodone(bp); + + return (error); + } + + if ((error = VNOP_BLOCKMAP(vp, f_offset, bp->b_bcount, &bp->b_blkno, &contig_bytes, NULL, bmap_flags, NULL))) { + DTRACE_IO1(start, buf_t, bp); + buf_seterror(bp, error); + buf_biodone(bp); + + return (error); + } + + DTRACE_IO1(start, buf_t, bp); +#if CONFIG_DTRACE + dtrace_io_start_flag = 1; +#endif /* CONFIG_DTRACE */ + + if ((bp->b_blkno == -1) || (contig_bytes == 0)) { + /* Set block number to force biodone later */ + bp->b_blkno = -1; + buf_clear(bp); + } + else if ((long)contig_bytes < bp->b_bcount) { + return (buf_strategy_fragmented(devvp, bp, f_offset, contig_bytes)); + } + } + +#if CONFIG_DTRACE + if (dtrace_io_start_flag == 0) { + DTRACE_IO1(start, buf_t, bp); + dtrace_io_start_flag = 1; + } +#endif /* CONFIG_DTRACE */ + + if (bp->b_blkno == -1) { + buf_biodone(bp); + return (0); + } + } + +#if CONFIG_DTRACE + if (dtrace_io_start_flag == 0) + DTRACE_IO1(start, buf_t, bp); +#endif /* CONFIG_DTRACE */ + +#if CONFIG_PROTECT + /* Capture f_offset in the bufattr*/ + cpx_t cpx = bufattr_cpx(buf_attr(bp)); + if (cpx) { + /* No need to go here for older EAs */ + if(cpx_use_offset_for_iv(cpx) && !cpx_synthetic_offset_for_iv(cpx)) { + off_t f_offset; + if ((error = VNOP_BLKTOOFF(bp->b_vp, bp->b_lblkno, &f_offset))) + return error; + + /* + * Attach the file offset to this buffer. The + * bufattr attributes will be passed down the stack + * until they reach IOFlashStorage. IOFlashStorage + * will retain the offset in a local variable when it + * issues its I/Os to the NAND controller. + * + * Note that LwVM may end up splitting this I/O + * into sub-I/Os if it crosses a chunk boundary. In this + * case, LwVM will update this field when it dispatches + * each I/O to IOFlashStorage. But from our perspective + * we have only issued a single I/O. + */ + buf_setcpoff(bp, f_offset); + CP_DEBUG((CPDBG_OFFSET_IO | DBG_FUNC_NONE), (uint32_t) f_offset, (uint32_t) bp->b_lblkno, (uint32_t) bp->b_blkno, (uint32_t) bp->b_bcount, 0); + } + } +#endif + + /* + * we can issue the I/O because... + * either B_CLUSTER is set which + * means that the I/O is properly set + * up to be a multiple of the page size, or + * we were able to successfully set up the + * physical block mapping + */ + error = VOCALL(devvp->v_op, VOFFSET(vnop_strategy), ap); + DTRACE_FSINFO(strategy, vnode_t, vp); + return (error); +} + + + +buf_t +buf_alloc(vnode_t vp) +{ + return(alloc_io_buf(vp, is_vm_privileged())); +} + +void +buf_free(buf_t bp) { + + free_io_buf(bp); +} + + +/* + * iterate buffers for the specified vp. + * if BUF_SCAN_DIRTY is set, do the dirty list + * if BUF_SCAN_CLEAN is set, do the clean list + * if neither flag is set, default to BUF_SCAN_DIRTY + * if BUF_NOTIFY_BUSY is set, call the callout function using a NULL bp for busy pages + */ + +struct buf_iterate_info_t { + int flag; + struct buflists *listhead; +}; + +void +buf_iterate(vnode_t vp, int (*callout)(buf_t, void *), int flags, void *arg) +{ + buf_t bp; + int retval; + struct buflists local_iterblkhd; + int lock_flags = BAC_NOWAIT | BAC_REMOVE; + int notify_busy = flags & BUF_NOTIFY_BUSY; + struct buf_iterate_info_t list[2]; + int num_lists, i; + + if (flags & BUF_SKIP_LOCKED) + lock_flags |= BAC_SKIP_LOCKED; + if (flags & BUF_SKIP_NONLOCKED) + lock_flags |= BAC_SKIP_NONLOCKED; + + if ( !(flags & (BUF_SCAN_DIRTY | BUF_SCAN_CLEAN))) + flags |= BUF_SCAN_DIRTY; + + num_lists = 0; + + if (flags & BUF_SCAN_DIRTY) { + list[num_lists].flag = VBI_DIRTY; + list[num_lists].listhead = &vp->v_dirtyblkhd; + num_lists++; + } + if (flags & BUF_SCAN_CLEAN) { + list[num_lists].flag = VBI_CLEAN; + list[num_lists].listhead = &vp->v_cleanblkhd; + num_lists++; + } + + for (i = 0; i < num_lists; i++) { + lck_mtx_lock(buf_mtxp); + + if (buf_iterprepare(vp, &local_iterblkhd, list[i].flag)) { + lck_mtx_unlock(buf_mtxp); + continue; + } + while (!LIST_EMPTY(&local_iterblkhd)) { + bp = LIST_FIRST(&local_iterblkhd); + LIST_REMOVE(bp, b_vnbufs); + LIST_INSERT_HEAD(list[i].listhead, bp, b_vnbufs); + + if (buf_acquire_locked(bp, lock_flags, 0, 0)) { + if (notify_busy) { + bp = NULL; + } else { + continue; + } + } + + lck_mtx_unlock(buf_mtxp); + + retval = callout(bp, arg); + + switch (retval) { + case BUF_RETURNED: + if (bp) + buf_brelse(bp); + break; + case BUF_CLAIMED: + break; + case BUF_RETURNED_DONE: + if (bp) + buf_brelse(bp); + lck_mtx_lock(buf_mtxp); + goto out; + case BUF_CLAIMED_DONE: + lck_mtx_lock(buf_mtxp); + goto out; + } + lck_mtx_lock(buf_mtxp); + } /* while list has more nodes */ + out: + buf_itercomplete(vp, &local_iterblkhd, list[i].flag); + lck_mtx_unlock(buf_mtxp); + } /* for each list */ +} /* buf_iterate */ + + +/* + * Flush out and invalidate all buffers associated with a vnode. + */ +int +buf_invalidateblks(vnode_t vp, int flags, int slpflag, int slptimeo) +{ + buf_t bp; + int aflags; + int error = 0; + int must_rescan = 1; + struct buflists local_iterblkhd; + + + if (LIST_EMPTY(&vp->v_cleanblkhd) && LIST_EMPTY(&vp->v_dirtyblkhd)) + return (0); + + lck_mtx_lock(buf_mtxp); + + for (;;) { + if (must_rescan == 0) + /* + * the lists may not be empty, but all that's left at this + * point are metadata or B_LOCKED buffers which are being + * skipped... we know this because we made it through both + * the clean and dirty lists without dropping buf_mtxp... + * each time we drop buf_mtxp we bump "must_rescan" + */ + break; + if (LIST_EMPTY(&vp->v_cleanblkhd) && LIST_EMPTY(&vp->v_dirtyblkhd)) + break; + must_rescan = 0; + /* + * iterate the clean list + */ + if (buf_iterprepare(vp, &local_iterblkhd, VBI_CLEAN)) { + goto try_dirty_list; + } + while (!LIST_EMPTY(&local_iterblkhd)) { + + bp = LIST_FIRST(&local_iterblkhd); + + LIST_REMOVE(bp, b_vnbufs); + LIST_INSERT_HEAD(&vp->v_cleanblkhd, bp, b_vnbufs); + + /* + * some filesystems distinguish meta data blocks with a negative logical block # + */ + if ((flags & BUF_SKIP_META) && (bp->b_lblkno < 0 || ISSET(bp->b_flags, B_META))) + continue; + + aflags = BAC_REMOVE; + + if ( !(flags & BUF_INVALIDATE_LOCKED) ) + aflags |= BAC_SKIP_LOCKED; + + if ( (error = (int)buf_acquire_locked(bp, aflags, slpflag, slptimeo)) ) { + if (error == EDEADLK) + /* + * this buffer was marked B_LOCKED... + * we didn't drop buf_mtxp, so we + * we don't need to rescan + */ + continue; + if (error == EAGAIN) { + /* + * found a busy buffer... we blocked and + * dropped buf_mtxp, so we're going to + * need to rescan after this pass is completed + */ + must_rescan++; + continue; + } + /* + * got some kind of 'real' error out of the msleep + * in buf_acquire_locked, terminate the scan and return the error + */ + buf_itercomplete(vp, &local_iterblkhd, VBI_CLEAN); + + lck_mtx_unlock(buf_mtxp); + return (error); + } + lck_mtx_unlock(buf_mtxp); + + if (bp->b_flags & B_LOCKED) + KERNEL_DEBUG(0xbbbbc038, bp, 0, 0, 0, 0); + + CLR(bp->b_flags, B_LOCKED); + SET(bp->b_flags, B_INVAL); + buf_brelse(bp); + + lck_mtx_lock(buf_mtxp); + + /* + * by dropping buf_mtxp, we allow new + * buffers to be added to the vnode list(s) + * we'll have to rescan at least once more + * if the queues aren't empty + */ + must_rescan++; + } + buf_itercomplete(vp, &local_iterblkhd, VBI_CLEAN); + +try_dirty_list: + /* + * Now iterate on dirty blks + */ + if (buf_iterprepare(vp, &local_iterblkhd, VBI_DIRTY)) { + continue; + } + while (!LIST_EMPTY(&local_iterblkhd)) { + bp = LIST_FIRST(&local_iterblkhd); + + LIST_REMOVE(bp, b_vnbufs); + LIST_INSERT_HEAD(&vp->v_dirtyblkhd, bp, b_vnbufs); + + /* + * some filesystems distinguish meta data blocks with a negative logical block # + */ + if ((flags & BUF_SKIP_META) && (bp->b_lblkno < 0 || ISSET(bp->b_flags, B_META))) + continue; + + aflags = BAC_REMOVE; + + if ( !(flags & BUF_INVALIDATE_LOCKED) ) + aflags |= BAC_SKIP_LOCKED; + + if ( (error = (int)buf_acquire_locked(bp, aflags, slpflag, slptimeo)) ) { + if (error == EDEADLK) + /* + * this buffer was marked B_LOCKED... + * we didn't drop buf_mtxp, so we + * we don't need to rescan + */ + continue; + if (error == EAGAIN) { + /* + * found a busy buffer... we blocked and + * dropped buf_mtxp, so we're going to + * need to rescan after this pass is completed + */ + must_rescan++; + continue; + } + /* + * got some kind of 'real' error out of the msleep + * in buf_acquire_locked, terminate the scan and return the error + */ + buf_itercomplete(vp, &local_iterblkhd, VBI_DIRTY); + + lck_mtx_unlock(buf_mtxp); + return (error); + } + lck_mtx_unlock(buf_mtxp); + + if (bp->b_flags & B_LOCKED) + KERNEL_DEBUG(0xbbbbc038, bp, 0, 0, 1, 0); + + CLR(bp->b_flags, B_LOCKED); + SET(bp->b_flags, B_INVAL); + + if (ISSET(bp->b_flags, B_DELWRI) && (flags & BUF_WRITE_DATA)) + (void) VNOP_BWRITE(bp); + else + buf_brelse(bp); + + lck_mtx_lock(buf_mtxp); + /* + * by dropping buf_mtxp, we allow new + * buffers to be added to the vnode list(s) + * we'll have to rescan at least once more + * if the queues aren't empty + */ + must_rescan++; + } + buf_itercomplete(vp, &local_iterblkhd, VBI_DIRTY); + } + lck_mtx_unlock(buf_mtxp); + + return (0); +} + +void +buf_flushdirtyblks(vnode_t vp, int wait, int flags, const char *msg) { + + (void) buf_flushdirtyblks_skipinfo(vp, wait, flags, msg); + return; +} + +int +buf_flushdirtyblks_skipinfo(vnode_t vp, int wait, int flags, const char *msg) { + buf_t bp; + int writes_issued = 0; + errno_t error; + int busy = 0; + struct buflists local_iterblkhd; + int lock_flags = BAC_NOWAIT | BAC_REMOVE; + int any_locked = 0; + + if (flags & BUF_SKIP_LOCKED) + lock_flags |= BAC_SKIP_LOCKED; + if (flags & BUF_SKIP_NONLOCKED) + lock_flags |= BAC_SKIP_NONLOCKED; +loop: + lck_mtx_lock(buf_mtxp); + + if (buf_iterprepare(vp, &local_iterblkhd, VBI_DIRTY) == 0) { + while (!LIST_EMPTY(&local_iterblkhd)) { + bp = LIST_FIRST(&local_iterblkhd); + LIST_REMOVE(bp, b_vnbufs); + LIST_INSERT_HEAD(&vp->v_dirtyblkhd, bp, b_vnbufs); + + if ((error = buf_acquire_locked(bp, lock_flags, 0, 0)) == EBUSY) { + busy++; + } + if (error) { + /* + * If we passed in BUF_SKIP_LOCKED or BUF_SKIP_NONLOCKED, + * we may want to do somethign differently if a locked or unlocked + * buffer was encountered (depending on the arg specified). + * In this case, we know that one of those two was set, and the + * buf acquisition failed above. + * + * If it failed with EDEADLK, then save state which can be emitted + * later on to the caller. Most callers should not care. + */ + if (error == EDEADLK) { + any_locked++; + } + continue; + } + lck_mtx_unlock(buf_mtxp); + + bp->b_flags &= ~B_LOCKED; + + /* + * Wait for I/O associated with indirect blocks to complete, + * since there is no way to quickly wait for them below. + */ + if ((bp->b_vp == vp) || (wait == 0)) + (void) buf_bawrite(bp); + else + (void) VNOP_BWRITE(bp); + writes_issued++; + + lck_mtx_lock(buf_mtxp); + } + buf_itercomplete(vp, &local_iterblkhd, VBI_DIRTY); + } + lck_mtx_unlock(buf_mtxp); + + if (wait) { + (void)vnode_waitforwrites(vp, 0, 0, 0, msg); + + if (vp->v_dirtyblkhd.lh_first && busy) { + /* + * we had one or more BUSY buffers on + * the dirtyblock list... most likely + * these are due to delayed writes that + * were moved to the bclean queue but + * have not yet been 'written'. + * if we issued some writes on the + * previous pass, we try again immediately + * if we didn't, we'll sleep for some time + * to allow the state to change... + */ + if (writes_issued == 0) { + (void)tsleep((caddr_t)&vp->v_numoutput, + PRIBIO + 1, "vnode_flushdirtyblks", hz/20); + } + writes_issued = 0; + busy = 0; + + goto loop; + } + } + + return any_locked; +} + + +/* + * called with buf_mtxp held... + * this lock protects the queue manipulation + */ +static int +buf_iterprepare(vnode_t vp, struct buflists *iterheadp, int flags) +{ + struct buflists * listheadp; + + if (flags & VBI_DIRTY) + listheadp = &vp->v_dirtyblkhd; + else + listheadp = &vp->v_cleanblkhd; + + while (vp->v_iterblkflags & VBI_ITER) { + vp->v_iterblkflags |= VBI_ITERWANT; + msleep(&vp->v_iterblkflags, buf_mtxp, 0, "buf_iterprepare", NULL); + } + if (LIST_EMPTY(listheadp)) { + LIST_INIT(iterheadp); + return(EINVAL); + } + vp->v_iterblkflags |= VBI_ITER; + + iterheadp->lh_first = listheadp->lh_first; + listheadp->lh_first->b_vnbufs.le_prev = &iterheadp->lh_first; + LIST_INIT(listheadp); + + return(0); +} + +/* + * called with buf_mtxp held... + * this lock protects the queue manipulation + */ +static void +buf_itercomplete(vnode_t vp, struct buflists *iterheadp, int flags) +{ + struct buflists * listheadp; + buf_t bp; + + if (flags & VBI_DIRTY) + listheadp = &vp->v_dirtyblkhd; + else + listheadp = &vp->v_cleanblkhd; + + while (!LIST_EMPTY(iterheadp)) { + bp = LIST_FIRST(iterheadp); + LIST_REMOVE(bp, b_vnbufs); + LIST_INSERT_HEAD(listheadp, bp, b_vnbufs); + } + vp->v_iterblkflags &= ~VBI_ITER; + + if (vp->v_iterblkflags & VBI_ITERWANT) { + vp->v_iterblkflags &= ~VBI_ITERWANT; + wakeup(&vp->v_iterblkflags); + } +} + + +static void +bremfree_locked(buf_t bp) +{ + struct bqueues *dp = NULL; + int whichq; + + whichq = bp->b_whichq; + + if (whichq == -1) { + if (bp->b_shadow_ref == 0) + panic("bremfree_locked: %p not on freelist", bp); + /* + * there are clones pointing to 'bp'... + * therefore, it was not put on a freelist + * when buf_brelse was last called on 'bp' + */ + return; + } + /* + * We only calculate the head of the freelist when removing + * the last element of the list as that is the only time that + * it is needed (e.g. to reset the tail pointer). + * + * NB: This makes an assumption about how tailq's are implemented. + */ + if (bp->b_freelist.tqe_next == NULL) { + dp = &bufqueues[whichq]; + + if (dp->tqh_last != &bp->b_freelist.tqe_next) + panic("bremfree: lost tail"); + } + TAILQ_REMOVE(dp, bp, b_freelist); + + if (whichq == BQ_LAUNDRY) + blaundrycnt--; + + bp->b_whichq = -1; + bp->b_timestamp = 0; + bp->b_shadow = 0; +} + +/* + * Associate a buffer with a vnode. + * buf_mtxp must be locked on entry + */ +static void +bgetvp_locked(vnode_t vp, buf_t bp) +{ + + if (bp->b_vp != vp) + panic("bgetvp_locked: not free"); + + if (vp->v_type == VBLK || vp->v_type == VCHR) + bp->b_dev = vp->v_rdev; + else + bp->b_dev = NODEV; + /* + * Insert onto list for new vnode. + */ + bufinsvn(bp, &vp->v_cleanblkhd); +} + +/* + * Disassociate a buffer from a vnode. + * buf_mtxp must be locked on entry + */ +static void +brelvp_locked(buf_t bp) +{ + /* + * Delete from old vnode list, if on one. + */ + if (bp->b_vnbufs.le_next != NOLIST) + bufremvn(bp); + + bp->b_vp = (vnode_t)NULL; +} + +/* + * Reassign a buffer from one vnode to another. + * Used to assign file specific control information + * (indirect blocks) to the vnode to which they belong. + */ +static void +buf_reassign(buf_t bp, vnode_t newvp) +{ + struct buflists *listheadp; + + if (newvp == NULL) { + printf("buf_reassign: NULL"); + return; + } + lck_mtx_lock_spin(buf_mtxp); + + /* + * Delete from old vnode list, if on one. + */ + if (bp->b_vnbufs.le_next != NOLIST) + bufremvn(bp); + /* + * If dirty, put on list of dirty buffers; + * otherwise insert onto list of clean buffers. + */ + if (ISSET(bp->b_flags, B_DELWRI)) + listheadp = &newvp->v_dirtyblkhd; + else + listheadp = &newvp->v_cleanblkhd; + bufinsvn(bp, listheadp); + + lck_mtx_unlock(buf_mtxp); +} + +static __inline__ void +bufhdrinit(buf_t bp) +{ + bzero((char *)bp, sizeof *bp); + bp->b_dev = NODEV; + bp->b_rcred = NOCRED; + bp->b_wcred = NOCRED; + bp->b_vnbufs.le_next = NOLIST; + bp->b_flags = B_INVAL; + + return; +} + +/* + * Initialize buffers and hash links for buffers. + */ +__private_extern__ void +bufinit(void) +{ + buf_t bp; + struct bqueues *dp; + int i; + + nbuf_headers = 0; + /* Initialize the buffer queues ('freelists') and the hash table */ + for (dp = bufqueues; dp < &bufqueues[BQUEUES]; dp++) + TAILQ_INIT(dp); + bufhashtbl = hashinit(nbuf_hashelements, M_CACHE, &bufhash); + + buf_busycount = 0; + + /* Initialize the buffer headers */ + for (i = 0; i < max_nbuf_headers; i++) { + nbuf_headers++; + bp = &buf_headers[i]; + bufhdrinit(bp); + + BLISTNONE(bp); + dp = &bufqueues[BQ_EMPTY]; + bp->b_whichq = BQ_EMPTY; + bp->b_timestamp = buf_timestamp(); + binsheadfree(bp, dp, BQ_EMPTY); + binshash(bp, &invalhash); + } + boot_nbuf_headers = nbuf_headers; + + TAILQ_INIT(&iobufqueue); + TAILQ_INIT(&delaybufqueue); + + for (; i < nbuf_headers + niobuf_headers; i++) { + bp = &buf_headers[i]; + bufhdrinit(bp); + bp->b_whichq = -1; + binsheadfree(bp, &iobufqueue, -1); + } + + /* + * allocate lock group attribute and group + */ + buf_mtx_grp_attr = lck_grp_attr_alloc_init(); + buf_mtx_grp = lck_grp_alloc_init("buffer cache", buf_mtx_grp_attr); + + /* + * allocate the lock attribute + */ + buf_mtx_attr = lck_attr_alloc_init(); + + /* + * allocate and initialize mutex's for the buffer and iobuffer pools + */ + buf_mtxp = lck_mtx_alloc_init(buf_mtx_grp, buf_mtx_attr); + iobuffer_mtxp = lck_mtx_alloc_init(buf_mtx_grp, buf_mtx_attr); + + if (iobuffer_mtxp == NULL) + panic("couldn't create iobuffer mutex"); + + if (buf_mtxp == NULL) + panic("couldn't create buf mutex"); + + /* + * allocate and initialize cluster specific global locks... + */ + cluster_init(); + + printf("using %d buffer headers and %d cluster IO buffer headers\n", + nbuf_headers, niobuf_headers); + + /* Set up zones used by the buffer cache */ + bufzoneinit(); + + /* start the bcleanbuf() thread */ + bcleanbuf_thread_init(); + + /* Register a callout for relieving vm pressure */ + if (vm_set_buffer_cleanup_callout(buffer_cache_gc) != KERN_SUCCESS) { + panic("Couldn't register buffer cache callout for vm pressure!\n"); + } + +} + +/* + * Zones for the meta data buffers + */ + +#define MINMETA 512 +#define MAXMETA 8192 + +struct meta_zone_entry { + zone_t mz_zone; + vm_size_t mz_size; + vm_size_t mz_max; + const char *mz_name; +}; + +struct meta_zone_entry meta_zones[] = { + {NULL, (MINMETA * 1), 128 * (MINMETA * 1), "buf.512" }, + {NULL, (MINMETA * 2), 64 * (MINMETA * 2), "buf.1024" }, + {NULL, (MINMETA * 4), 16 * (MINMETA * 4), "buf.2048" }, + {NULL, (MINMETA * 8), 512 * (MINMETA * 8), "buf.4096" }, + {NULL, (MINMETA * 16), 512 * (MINMETA * 16), "buf.8192" }, + {NULL, 0, 0, "" } /* End */ +}; + +/* + * Initialize the meta data zones + */ +static void +bufzoneinit(void) +{ + int i; + + for (i = 0; meta_zones[i].mz_size != 0; i++) { + meta_zones[i].mz_zone = + zinit(meta_zones[i].mz_size, + meta_zones[i].mz_max, + PAGE_SIZE, + meta_zones[i].mz_name); + zone_change(meta_zones[i].mz_zone, Z_CALLERACCT, FALSE); + } + buf_hdr_zone = zinit(sizeof(struct buf), 32, PAGE_SIZE, "buf headers"); + zone_change(buf_hdr_zone, Z_CALLERACCT, FALSE); +} + +static __inline__ zone_t +getbufzone(size_t size) +{ + int i; + + if ((size % 512) || (size < MINMETA) || (size > MAXMETA)) + panic("getbufzone: incorect size = %lu", size); + + for (i = 0; meta_zones[i].mz_size != 0; i++) { + if (meta_zones[i].mz_size >= size) + break; + } + + return (meta_zones[i].mz_zone); +} + + + +static struct buf * +bio_doread(vnode_t vp, daddr64_t blkno, int size, kauth_cred_t cred, int async, int queuetype) +{ + buf_t bp; + + bp = buf_getblk(vp, blkno, size, 0, 0, queuetype); + + /* + * If buffer does not have data valid, start a read. + * Note that if buffer is B_INVAL, buf_getblk() won't return it. + * Therefore, it's valid if it's I/O has completed or been delayed. + */ + if (!ISSET(bp->b_flags, (B_DONE | B_DELWRI))) { + struct proc *p; + + p = current_proc(); + + /* Start I/O for the buffer (keeping credentials). */ + SET(bp->b_flags, B_READ | async); + if (IS_VALID_CRED(cred) && !IS_VALID_CRED(bp->b_rcred)) { + kauth_cred_ref(cred); + bp->b_rcred = cred; + } + + VNOP_STRATEGY(bp); + + trace(TR_BREADMISS, pack(vp, size), blkno); + + /* Pay for the read. */ + if (p && p->p_stats) { + OSIncrementAtomicLong(&p->p_stats->p_ru.ru_inblock); /* XXX */ + } + + if (async) { + /* + * since we asked for an ASYNC I/O + * the biodone will do the brelse + * we don't want to pass back a bp + * that we don't 'own' + */ + bp = NULL; + } + } else if (async) { + buf_brelse(bp); + bp = NULL; + } + + trace(TR_BREADHIT, pack(vp, size), blkno); + + return (bp); +} + +/* + * Perform the reads for buf_breadn() and buf_meta_breadn(). + * Trivial modification to the breada algorithm presented in Bach (p.55). + */ +static errno_t +do_breadn_for_type(vnode_t vp, daddr64_t blkno, int size, daddr64_t *rablks, int *rasizes, + int nrablks, kauth_cred_t cred, buf_t *bpp, int queuetype) +{ + buf_t bp; + int i; + + bp = *bpp = bio_doread(vp, blkno, size, cred, 0, queuetype); + + /* + * For each of the read-ahead blocks, start a read, if necessary. + */ + for (i = 0; i < nrablks; i++) { + /* If it's in the cache, just go on to next one. */ + if (incore(vp, rablks[i])) + continue; + + /* Get a buffer for the read-ahead block */ + (void) bio_doread(vp, rablks[i], rasizes[i], cred, B_ASYNC, queuetype); + } + + /* Otherwise, we had to start a read for it; wait until it's valid. */ + return (buf_biowait(bp)); +} + + +/* + * Read a disk block. + * This algorithm described in Bach (p.54). + */ +errno_t +buf_bread(vnode_t vp, daddr64_t blkno, int size, kauth_cred_t cred, buf_t *bpp) +{ + buf_t bp; + + /* Get buffer for block. */ + bp = *bpp = bio_doread(vp, blkno, size, cred, 0, BLK_READ); + + /* Wait for the read to complete, and return result. */ + return (buf_biowait(bp)); +} + +/* + * Read a disk block. [bread() for meta-data] + * This algorithm described in Bach (p.54). + */ +errno_t +buf_meta_bread(vnode_t vp, daddr64_t blkno, int size, kauth_cred_t cred, buf_t *bpp) +{ + buf_t bp; + + /* Get buffer for block. */ + bp = *bpp = bio_doread(vp, blkno, size, cred, 0, BLK_META); + + /* Wait for the read to complete, and return result. */ + return (buf_biowait(bp)); +} + +/* + * Read-ahead multiple disk blocks. The first is sync, the rest async. + */ +errno_t +buf_breadn(vnode_t vp, daddr64_t blkno, int size, daddr64_t *rablks, int *rasizes, int nrablks, kauth_cred_t cred, buf_t *bpp) +{ + return (do_breadn_for_type(vp, blkno, size, rablks, rasizes, nrablks, cred, bpp, BLK_READ)); +} + +/* + * Read-ahead multiple disk blocks. The first is sync, the rest async. + * [buf_breadn() for meta-data] + */ +errno_t +buf_meta_breadn(vnode_t vp, daddr64_t blkno, int size, daddr64_t *rablks, int *rasizes, int nrablks, kauth_cred_t cred, buf_t *bpp) +{ + return (do_breadn_for_type(vp, blkno, size, rablks, rasizes, nrablks, cred, bpp, BLK_META)); +} + +/* + * Block write. Described in Bach (p.56) + */ +errno_t +buf_bwrite(buf_t bp) +{ + int sync, wasdelayed; + errno_t rv; + proc_t p = current_proc(); + vnode_t vp = bp->b_vp; + + if (bp->b_datap == 0) { + if (brecover_data(bp) == 0) + return (0); } + /* Remember buffer type, to switch on it later. */ + sync = !ISSET(bp->b_flags, B_ASYNC); + wasdelayed = ISSET(bp->b_flags, B_DELWRI); + CLR(bp->b_flags, (B_READ | B_DONE | B_ERROR | B_DELWRI)); + + if (wasdelayed) + OSAddAtomicLong(-1, &nbdwrite); if (!sync) { /* @@ -595,25 +2244,25 @@ bwrite(bp) * be properly notified that its I/O has completed. */ if (wasdelayed) - reassignbuf(bp, vp); - else - if (p && p->p_stats) - p->p_stats->p_ru.ru_oublock++; /* XXX */ + buf_reassign(bp, vp); + else + if (p && p->p_stats) { + OSIncrementAtomicLong(&p->p_stats->p_ru.ru_oublock); /* XXX */ + } } - trace(TR_BUFWRITE, pack(vp, bp->b_bcount), bp->b_lblkno); /* Initiate disk write. Make sure the appropriate party is charged. */ - SET(bp->b_flags, B_WRITEINPROG); - vp->v_numoutput++; + + OSAddAtomic(1, &vp->v_numoutput); - VOP_STRATEGY(bp); + VNOP_STRATEGY(bp); if (sync) { /* * If I/O was synchronous, wait for it to complete. */ - rv = biowait(bp); + rv = buf_biowait(bp); /* * Pay for the I/O operation, if it's not been paid for, and @@ -621,13 +2270,14 @@ bwrite(bp) * were payed for above.) */ if (wasdelayed) - reassignbuf(bp, vp); + buf_reassign(bp, vp); else - if (p && p->p_stats) - p->p_stats->p_ru.ru_oublock++; /* XXX */ + if (p && p->p_stats) { + OSIncrementAtomicLong(&p->p_stats->p_ru.ru_oublock); /* XXX */ + } /* Release the buffer. */ - brelse(bp); + buf_brelse(bp); return (rv); } else { @@ -636,10 +2286,9 @@ bwrite(bp) } int -vn_bwrite(ap) - struct vop_bwrite_args *ap; +vn_bwrite(struct vnop_bwrite_args *ap) { - return (bwrite(ap->a_bp)); + return (buf_bwrite(ap->a_bp)); } /* @@ -655,19 +2304,17 @@ vn_bwrite(ap) * * Described in Leffler, et al. (pp. 208-213). * - * Note: With the abilitty to allocate additional buffer + * Note: With the ability to allocate additional buffer * headers, we can get in to the situation where "too" many - * bdwrite()s can create situation where the kernel can create - * buffers faster than the disks can service. Doing a bawrite() in - * cases were we have "too many" outstanding bdwrite()s avoids that. + * buf_bdwrite()s can create situation where the kernel can create + * buffers faster than the disks can service. Doing a buf_bawrite() in + * cases where we have "too many" outstanding buf_bdwrite()s avoids that. */ -__private_extern__ int -bdwrite_internal(bp, return_error) - struct buf *bp; - int return_error; +int +bdwrite_internal(buf_t bp, int return_error) { - struct proc *p = current_proc(); - struct vnode *vp = bp->b_vp; + proc_t p = current_proc(); + vnode_t vp = bp->b_vp; /* * If the block hasn't been seen before: @@ -677,242 +2324,403 @@ bdwrite_internal(bp, return_error) */ if (!ISSET(bp->b_flags, B_DELWRI)) { SET(bp->b_flags, B_DELWRI); - if (p && p->p_stats) - p->p_stats->p_ru.ru_oublock++; /* XXX */ - nbdwrite ++; - reassignbuf(bp, vp); - } - - /* If this is a tape block, write it the block now. */ - if (ISSET(bp->b_flags, B_TAPE)) { - /* bwrite(bp); */ - VOP_BWRITE(bp); - return (0); - } - - /* - * If the vnode has "too many" write operations in progress - * wait for them to finish the IO - */ - while (vp->v_numoutput >= BUFWRITE_THROTTLE) { - vp->v_flag |= VTHROTTLED; - (void)tsleep((caddr_t)&vp->v_numoutput, PRIBIO + 1, "bdwrite", 0); + if (p && p->p_stats) { + OSIncrementAtomicLong(&p->p_stats->p_ru.ru_oublock); /* XXX */ + } + OSAddAtomicLong(1, &nbdwrite); + buf_reassign(bp, vp); } /* - * If we have too many delayed write buffers, - * more than we can "safely" handle, just fall back to - * doing the async write + * if we're not LOCKED, but the total number of delayed writes + * has climbed above 75% of the total buffers in the system + * return an error if the caller has indicated that it can + * handle one in this case, otherwise schedule the I/O now + * this is done to prevent us from allocating tons of extra + * buffers when dealing with virtual disks (i.e. DiskImages), + * because additional buffers are dynamically allocated to prevent + * deadlocks from occurring + * + * however, can't do a buf_bawrite() if the LOCKED bit is set because the + * buffer is part of a transaction and can't go to disk until + * the LOCKED bit is cleared. */ - if (nbdwrite < 0) - panic("bdwrite: Negative nbdwrite"); - - if (nbdwrite > ((nbuf/4)*3)) { + if (!ISSET(bp->b_flags, B_LOCKED) && nbdwrite > ((nbuf_headers/4)*3)) { if (return_error) return (EAGAIN); - else - bawrite(bp); - return (0); + /* + * If the vnode has "too many" write operations in progress + * wait for them to finish the IO + */ + (void)vnode_waitforwrites(vp, VNODE_ASYNC_THROTTLE, 0, 0, "buf_bdwrite"); + + return (buf_bawrite(bp)); } /* Otherwise, the "write" is done, so mark and release the buffer. */ SET(bp->b_flags, B_DONE); - brelse(bp); + buf_brelse(bp); return (0); } -void -bdwrite(bp) - struct buf *bp; +errno_t +buf_bdwrite(buf_t bp) { - (void) bdwrite_internal(bp, 0); + return (bdwrite_internal(bp, 0)); } /* - * Asynchronous block write; just an asynchronous bwrite(). + * Asynchronous block write; just an asynchronous buf_bwrite(). * * Note: With the abilitty to allocate additional buffer * headers, we can get in to the situation where "too" many - * bawrite()s can create situation where the kernel can create + * buf_bawrite()s can create situation where the kernel can create * buffers faster than the disks can service. * We limit the number of "in flight" writes a vnode can have to * avoid this. */ static int -bawrite_internal(bp, throttle) - struct buf *bp; - int throttle; +bawrite_internal(buf_t bp, int throttle) +{ + vnode_t vp = bp->b_vp; + + if (vp) { + if (throttle) + /* + * If the vnode has "too many" write operations in progress + * wait for them to finish the IO + */ + (void)vnode_waitforwrites(vp, VNODE_ASYNC_THROTTLE, 0, 0, (const char *)"buf_bawrite"); + else if (vp->v_numoutput >= VNODE_ASYNC_THROTTLE) + /* + * return to the caller and + * let him decide what to do + */ + return (EWOULDBLOCK); + } + SET(bp->b_flags, B_ASYNC); + + return (VNOP_BWRITE(bp)); +} + +errno_t +buf_bawrite(buf_t bp) +{ + return (bawrite_internal(bp, 1)); +} + + + +static void +buf_free_meta_store(buf_t bp) +{ + if (bp->b_bufsize) { + if (ISSET(bp->b_flags, B_ZALLOC)) { + zone_t z; + + z = getbufzone(bp->b_bufsize); + zfree(z, (void *)bp->b_datap); + } else + kmem_free(kernel_map, bp->b_datap, bp->b_bufsize); + + bp->b_datap = (uintptr_t)NULL; + bp->b_bufsize = 0; + } +} + + +static buf_t +buf_brelse_shadow(buf_t bp) { - struct vnode *vp = bp->b_vp; + buf_t bp_head; + buf_t bp_temp; + buf_t bp_return = NULL; +#ifdef BUF_MAKE_PRIVATE + buf_t bp_data; + int data_ref = 0; +#endif + int need_wakeup = 0; - if (vp) { + lck_mtx_lock_spin(buf_mtxp); + + __IGNORE_WCASTALIGN(bp_head = (buf_t)bp->b_orig); + + if (bp_head->b_whichq != -1) + panic("buf_brelse_shadow: bp_head on freelist %d\n", bp_head->b_whichq); + +#ifdef BUF_MAKE_PRIVATE + if (bp_data = bp->b_data_store) { + bp_data->b_data_ref--; /* - * If the vnode has "too many" write operations in progress - * wait for them to finish the IO + * snapshot the ref count so that we can check it + * outside of the lock... we only want the guy going + * from 1 -> 0 to try and release the storage */ - while (vp->v_numoutput >= BUFWRITE_THROTTLE) { - if (throttle) { - vp->v_flag |= VTHROTTLED; - (void)tsleep((caddr_t)&vp->v_numoutput, - PRIBIO + 1, "bawrite", 0); - } else - return (EWOULDBLOCK); + data_ref = bp_data->b_data_ref; + } +#endif + KERNEL_DEBUG(0xbbbbc008 | DBG_FUNC_START, bp, bp_head, bp_head->b_shadow_ref, 0, 0); + + bp_head->b_shadow_ref--; + + for (bp_temp = bp_head; bp_temp && bp != bp_temp->b_shadow; bp_temp = bp_temp->b_shadow); + + if (bp_temp == NULL) + panic("buf_brelse_shadow: bp not on list %p", bp_head); + + bp_temp->b_shadow = bp_temp->b_shadow->b_shadow; + +#ifdef BUF_MAKE_PRIVATE + /* + * we're about to free the current 'owner' of the data buffer and + * there is at least one other shadow buf_t still pointing at it + * so transfer it to the first shadow buf left in the chain + */ + if (bp == bp_data && data_ref) { + if ((bp_data = bp_head->b_shadow) == NULL) + panic("buf_brelse_shadow: data_ref mismatch bp(%p)", bp); + + for (bp_temp = bp_data; bp_temp; bp_temp = bp_temp->b_shadow) + bp_temp->b_data_store = bp_data; + bp_data->b_data_ref = data_ref; + } +#endif + if (bp_head->b_shadow_ref == 0 && bp_head->b_shadow) + panic("buf_relse_shadow: b_shadow != NULL && b_shadow_ref == 0 bp(%p)", bp); + if (bp_head->b_shadow_ref && bp_head->b_shadow == 0) + panic("buf_relse_shadow: b_shadow == NULL && b_shadow_ref != 0 bp(%p)", bp); + + if (bp_head->b_shadow_ref == 0) { + if (!ISSET(bp_head->b_lflags, BL_BUSY)) { + + CLR(bp_head->b_flags, B_AGE); + bp_head->b_timestamp = buf_timestamp(); + + if (ISSET(bp_head->b_flags, B_LOCKED)) { + bp_head->b_whichq = BQ_LOCKED; + binstailfree(bp_head, &bufqueues[BQ_LOCKED], BQ_LOCKED); + } else { + bp_head->b_whichq = BQ_META; + binstailfree(bp_head, &bufqueues[BQ_META], BQ_META); + } + } else if (ISSET(bp_head->b_lflags, BL_WAITSHADOW)) { + CLR(bp_head->b_lflags, BL_WAITSHADOW); + + bp_return = bp_head; + } + if (ISSET(bp_head->b_lflags, BL_WANTED_REF)) { + CLR(bp_head->b_lflags, BL_WANTED_REF); + need_wakeup = 1; } } + lck_mtx_unlock(buf_mtxp); - SET(bp->b_flags, B_ASYNC); - VOP_BWRITE(bp); - return (0); -} + if (need_wakeup) + wakeup(bp_head); -void -bawrite(bp) - struct buf *bp; -{ - (void) bawrite_internal(bp, 1); -} +#ifdef BUF_MAKE_PRIVATE + if (bp == bp_data && data_ref == 0) + buf_free_meta_store(bp); -/* - * bwillwrite: - * - * Called prior to the locking of any vnodes when we are expecting to - * write. We do not want to starve the buffer cache with too many - * dirty buffers so we block here. By blocking prior to the locking - * of any vnodes we attempt to avoid the situation where a locked vnode - * prevents the various system daemons from flushing related buffers. - */ - -void -bwillwrite(void) -{ - /* XXX To be implemented later */ + bp->b_data_store = NULL; +#endif + KERNEL_DEBUG(0xbbbbc008 | DBG_FUNC_END, bp, 0, 0, 0, 0); + + return (bp_return); } + /* * Release a buffer on to the free lists. * Described in Bach (p. 46). */ void -brelse(bp) - struct buf *bp; +buf_brelse(buf_t bp) { struct bqueues *bufq; - int s; - long whichq; + long whichq; + upl_t upl; + int need_wakeup = 0; + int need_bp_wakeup = 0; + + + if (bp->b_whichq != -1 || !(bp->b_lflags & BL_BUSY)) + panic("buf_brelse: bad buffer = %p\n", bp); + +#ifdef JOE_DEBUG + (void) OSBacktrace(&bp->b_stackbrelse[0], 6); + + bp->b_lastbrelse = current_thread(); + bp->b_tag = 0; +#endif + if (bp->b_lflags & BL_IOBUF) { + buf_t shadow_master_bp = NULL; + + if (ISSET(bp->b_lflags, BL_SHADOW)) + shadow_master_bp = buf_brelse_shadow(bp); + else if (ISSET(bp->b_lflags, BL_IOBUF_ALLOC)) + buf_free_meta_store(bp); + free_io_buf(bp); + + if (shadow_master_bp) { + bp = shadow_master_bp; + goto finish_shadow_master; + } + return; + } KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 388)) | DBG_FUNC_START, - bp->b_lblkno * PAGE_SIZE, (int)bp, (int)bp->b_data, + bp->b_lblkno * PAGE_SIZE, bp, bp->b_datap, bp->b_flags, 0); trace(TR_BRELSE, pack(bp->b_vp, bp->b_bufsize), bp->b_lblkno); - /* IO is done. Cleanup the UPL state */ - if (!ISSET(bp->b_flags, B_META) - && UBCINFOEXISTS(bp->b_vp) && bp->b_bufsize) { + /* + * if we're invalidating a buffer that has the B_FILTER bit + * set then call the b_iodone function so it gets cleaned + * up properly. + * + * the HFS journal code depends on this + */ + if (ISSET(bp->b_flags, B_META) && ISSET(bp->b_flags, B_INVAL)) { + if (ISSET(bp->b_flags, B_FILTER)) { /* if necessary, call out */ + void (*iodone_func)(struct buf *, void *) = bp->b_iodone; + void *arg = bp->b_transaction; + + CLR(bp->b_flags, B_FILTER); /* but note callout done */ + bp->b_iodone = NULL; + bp->b_transaction = NULL; + + if (iodone_func == NULL) { + panic("brelse: bp @ %p has NULL b_iodone!\n", bp); + } + (*iodone_func)(bp, arg); + } + } + /* + * I/O is done. Cleanup the UPL state + */ + upl = bp->b_upl; + + if ( !ISSET(bp->b_flags, B_META) && UBCINFOEXISTS(bp->b_vp) && bp->b_bufsize) { kern_return_t kret; - upl_t upl; int upl_flags; - if ( !ISSET(bp->b_flags, B_PAGELIST)) { + if (upl == NULL) { if ( !ISSET(bp->b_flags, B_INVAL)) { kret = ubc_create_upl(bp->b_vp, - ubc_blktooff(bp->b_vp, bp->b_lblkno), - bp->b_bufsize, - &upl, - NULL, - UPL_PRECIOUS); + ubc_blktooff(bp->b_vp, bp->b_lblkno), + bp->b_bufsize, + &upl, + NULL, + UPL_PRECIOUS); + if (kret != KERN_SUCCESS) - panic("brelse: Failed to get pagelists"); -#ifdef UBC_DEBUG - upl_ubc_alias_set(upl, bp, 5); -#endif /* UBC_DEBUG */ - } else - upl = (upl_t) 0; + panic("brelse: Failed to create UPL"); +#if UPL_DEBUG + upl_ubc_alias_set(upl, (uintptr_t) bp, (uintptr_t) 5); +#endif /* UPL_DEBUG */ + } } else { - upl = bp->b_pagelist; - kret = ubc_upl_unmap(upl); + if (bp->b_datap) { + kret = ubc_upl_unmap(upl); - if (kret != KERN_SUCCESS) - panic("kernel_upl_unmap failed"); - bp->b_data = 0; + if (kret != KERN_SUCCESS) + panic("ubc_upl_unmap failed"); + bp->b_datap = (uintptr_t)NULL; + } } if (upl) { if (bp->b_flags & (B_ERROR | B_INVAL)) { - if (bp->b_flags & (B_READ | B_INVAL)) + if (bp->b_flags & (B_READ | B_INVAL)) upl_flags = UPL_ABORT_DUMP_PAGES; else upl_flags = 0; + ubc_upl_abort(upl, upl_flags); } else { - if (ISSET(bp->b_flags, B_NEEDCOMMIT)) - upl_flags = UPL_COMMIT_CLEAR_DIRTY ; - else if (ISSET(bp->b_flags, B_DELWRI | B_WASDIRTY)) - upl_flags = UPL_COMMIT_SET_DIRTY ; + if (ISSET(bp->b_flags, B_DELWRI | B_WASDIRTY)) + upl_flags = UPL_COMMIT_SET_DIRTY ; else - upl_flags = UPL_COMMIT_CLEAR_DIRTY ; + upl_flags = UPL_COMMIT_CLEAR_DIRTY ; + ubc_upl_commit_range(upl, 0, bp->b_bufsize, upl_flags | - UPL_COMMIT_INACTIVATE | UPL_COMMIT_FREE_ON_EMPTY); + UPL_COMMIT_INACTIVATE | UPL_COMMIT_FREE_ON_EMPTY); } - s = splbio(); - CLR(bp->b_flags, B_PAGELIST); - bp->b_pagelist = 0; - splx(s); + bp->b_upl = NULL; } } else { - if(ISSET(bp->b_flags, B_PAGELIST)) - panic("brelse: pagelist set for non VREG; vp=%x", bp->b_vp); + if ( (upl) ) + panic("brelse: UPL set for non VREG; vp=%p", bp->b_vp); } - /* Wake up any processes waiting for any buffer to become free. */ - if (needbuffer) { - needbuffer = 0; - wakeup(&needbuffer); - } - - /* Wake up any proceeses waiting for _this_ buffer to become free. */ - if (ISSET(bp->b_flags, B_WANTED)) { - CLR(bp->b_flags, B_WANTED); - wakeup(bp); - } - - /* Block disk interrupts. */ - s = splbio(); - /* - * Determine which queue the buffer should be on, then put it there. + * If it's locked, don't report an error; try again later. */ - - /* If it's locked, don't report an error; try again later. */ if (ISSET(bp->b_flags, (B_LOCKED|B_ERROR)) == (B_LOCKED|B_ERROR)) CLR(bp->b_flags, B_ERROR); - - /* If it's not cacheable, or an error, mark it invalid. */ + /* + * If it's not cacheable, or an error, mark it invalid. + */ if (ISSET(bp->b_flags, (B_NOCACHE|B_ERROR))) SET(bp->b_flags, B_INVAL); + + if ((bp->b_bufsize <= 0) || + ISSET(bp->b_flags, B_INVAL) || + (ISSET(bp->b_lflags, BL_WANTDEALLOC) && !ISSET(bp->b_flags, B_DELWRI))) { + + boolean_t delayed_buf_free_meta_store = FALSE; - if ((bp->b_bufsize <= 0) || ISSET(bp->b_flags, B_INVAL)) { /* - * If it's invalid or empty, dissociate it from its vnode - * and put on the head of the appropriate queue. + * If it's invalid or empty, dissociate it from its vnode, + * release its storage if B_META, and + * clean it up a bit and put it on the EMPTY queue */ - if (bp->b_vp) - brelvp(bp); - if (ISSET(bp->b_flags, B_DELWRI)) { - CLR(bp->b_flags, B_DELWRI); - nbdwrite--; - wakeup((caddr_t)&nbdwrite); + if (ISSET(bp->b_flags, B_DELWRI)) + OSAddAtomicLong(-1, &nbdwrite); + + if (ISSET(bp->b_flags, B_META)) { + if (bp->b_shadow_ref) + delayed_buf_free_meta_store = TRUE; + else + buf_free_meta_store(bp); } - if (bp->b_bufsize <= 0) - whichq = BQ_EMPTY; /* no data */ - else if (ISSET(bp->b_flags, B_META)) - whichq = BQ_META; /* meta-data */ - else - whichq = BQ_AGE; /* invalid data */ + /* + * nuke any credentials we were holding + */ + buf_release_credentials(bp); - bufq = &bufqueues[whichq]; - binsheadfree(bp, bufq, whichq); + lck_mtx_lock_spin(buf_mtxp); + + if (bp->b_shadow_ref) { + SET(bp->b_lflags, BL_WAITSHADOW); + + lck_mtx_unlock(buf_mtxp); + + return; + } + if (delayed_buf_free_meta_store == TRUE) { + + lck_mtx_unlock(buf_mtxp); +finish_shadow_master: + buf_free_meta_store(bp); + + lck_mtx_lock_spin(buf_mtxp); + } + CLR(bp->b_flags, (B_META | B_ZALLOC | B_DELWRI | B_LOCKED | B_AGE | B_ASYNC | B_NOCACHE | B_FUA)); + + if (bp->b_vp) + brelvp_locked(bp); + + bremhash(bp); + BLISTNONE(bp); + binshash(bp, &invalhash); + + bp->b_whichq = BQ_EMPTY; + binsheadfree(bp, &bufqueues[BQ_EMPTY], BQ_EMPTY); } else { + /* * It has valid data. Put it on the end of the appropriate * queue, so that it'll stick around for as long as possible. @@ -925,19 +2733,74 @@ brelse(bp) whichq = BQ_AGE; /* stale but valid data */ else whichq = BQ_LRU; /* valid data */ - bufq = &bufqueues[whichq]; - binstailfree(bp, bufq, whichq); - } - /* Unlock the buffer. */ - CLR(bp->b_flags, (B_AGE | B_ASYNC | B_BUSY | B_NOCACHE)); + bp->b_timestamp = buf_timestamp(); + + lck_mtx_lock_spin(buf_mtxp); + + /* + * the buf_brelse_shadow routine doesn't take 'ownership' + * of the parent buf_t... it updates state that is protected by + * the buf_mtxp, and checks for BL_BUSY to determine whether to + * put the buf_t back on a free list. b_shadow_ref is protected + * by the lock, and since we have not yet cleared B_BUSY, we need + * to check it while holding the lock to insure that one of us + * puts this buf_t back on a free list when it is safe to do so + */ + if (bp->b_shadow_ref == 0) { + CLR(bp->b_flags, (B_AGE | B_ASYNC | B_NOCACHE)); + bp->b_whichq = whichq; + binstailfree(bp, bufq, whichq); + } else { + /* + * there are still cloned buf_t's pointing + * at this guy... need to keep it off the + * freelists until a buf_brelse is done on + * the last clone + */ + CLR(bp->b_flags, (B_ASYNC | B_NOCACHE)); + } + } + if (needbuffer) { + /* + * needbuffer is a global + * we're currently using buf_mtxp to protect it + * delay doing the actual wakeup until after + * we drop buf_mtxp + */ + needbuffer = 0; + need_wakeup = 1; + } + if (ISSET(bp->b_lflags, BL_WANTED)) { + /* + * delay the actual wakeup until after we + * clear BL_BUSY and we've dropped buf_mtxp + */ + need_bp_wakeup = 1; + } + /* + * Unlock the buffer. + */ + CLR(bp->b_lflags, (BL_BUSY | BL_WANTED)); + buf_busycount--; - /* Allow disk interrupts. */ - splx(s); + lck_mtx_unlock(buf_mtxp); + if (need_wakeup) { + /* + * Wake up any processes waiting for any buffer to become free. + */ + wakeup(&needbuffer); + } + if (need_bp_wakeup) { + /* + * Wake up any proceeses waiting for _this_ buffer to become free. + */ + wakeup(bp); + } KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 388)) | DBG_FUNC_END, - (int)bp, (int)bp->b_data, bp->b_flags, 0, 0); + bp, bp->b_datap, bp->b_flags, 0, 0); } /* @@ -947,26 +2810,66 @@ brelse(bp) * we normally don't return the buffer, unless the caller explicitly * wants us to. */ -struct buf * -incore(vp, blkno) - struct vnode *vp; - daddr_t blkno; +static boolean_t +incore(vnode_t vp, daddr64_t blkno) { - struct buf *bp; + boolean_t retval; + struct bufhashhdr *dp; + + dp = BUFHASH(vp, blkno); - bp = BUFHASH(vp, blkno)->lh_first; + lck_mtx_lock_spin(buf_mtxp); + + if (incore_locked(vp, blkno, dp)) + retval = TRUE; + else + retval = FALSE; + lck_mtx_unlock(buf_mtxp); + + return (retval); +} + + +static buf_t +incore_locked(vnode_t vp, daddr64_t blkno, struct bufhashhdr *dp) +{ + struct buf *bp; /* Search hash chain */ - for (; bp != NULL; bp = bp->b_hash.le_next) { + for (bp = dp->lh_first; bp != NULL; bp = bp->b_hash.le_next) { if (bp->b_lblkno == blkno && bp->b_vp == vp && - !ISSET(bp->b_flags, B_INVAL)) + !ISSET(bp->b_flags, B_INVAL)) { return (bp); + } } - - return (0); + return (NULL); } +void +buf_wait_for_shadow_io(vnode_t vp, daddr64_t blkno) +{ + buf_t bp; + struct bufhashhdr *dp; + + dp = BUFHASH(vp, blkno); + + lck_mtx_lock_spin(buf_mtxp); + + for (;;) { + if ((bp = incore_locked(vp, blkno, dp)) == NULL) + break; + + if (bp->b_shadow_ref == 0) + break; + + SET(bp->b_lflags, BL_WANTED_REF); + + (void) msleep(bp, buf_mtxp, PSPIN | (PRIBIO+1), "buf_wait_for_shadow", NULL); + } + lck_mtx_unlock(buf_mtxp); +} + /* XXX FIXME -- Update the comment to reflect the UBC changes (please) -- */ /* * Get a block of requested size that is associated with @@ -976,112 +2879,164 @@ incore(vp, blkno) * correct size. It is up to the caller to insure that the * cached blocks be of the correct size. */ -struct buf * -getblk(vp, blkno, size, slpflag, slptimeo, operation) - register struct vnode *vp; - daddr_t blkno; - int size, slpflag, slptimeo, operation; +buf_t +buf_getblk(vnode_t vp, daddr64_t blkno, int size, int slpflag, int slptimeo, int operation) { - struct buf *bp; - int s, err; + buf_t bp; + int err; upl_t upl; upl_page_info_t *pl; kern_return_t kret; - int error=0; - int pagedirty = 0; + int ret_only_valid; + struct timespec ts; + int upl_flags; + struct bufhashhdr *dp; KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 386)) | DBG_FUNC_START, - blkno * PAGE_SIZE, size, operation, 0, 0); + (uintptr_t)(blkno * PAGE_SIZE), size, operation, 0, 0); + + ret_only_valid = operation & BLK_ONLYVALID; + operation &= ~BLK_ONLYVALID; + dp = BUFHASH(vp, blkno); start: + lck_mtx_lock_spin(buf_mtxp); - s = splbio(); - if ((bp = incore(vp, blkno))) { - /* Found in the Buffer Cache */ - if (ISSET(bp->b_flags, B_BUSY)) { - /* but is busy */ + if ((bp = incore_locked(vp, blkno, dp))) { + /* + * Found in the Buffer Cache + */ + if (ISSET(bp->b_lflags, BL_BUSY)) { + /* + * but is busy + */ switch (operation) { case BLK_READ: case BLK_WRITE: case BLK_META: - SET(bp->b_flags, B_WANTED); + SET(bp->b_lflags, BL_WANTED); bufstats.bufs_busyincore++; - err = tsleep(bp, slpflag | (PRIBIO + 1), "getblk", - slptimeo); - splx(s); + + /* + * don't retake the mutex after being awakened... + * the time out is in msecs + */ + ts.tv_sec = (slptimeo/1000); + ts.tv_nsec = (slptimeo % 1000) * 10 * NSEC_PER_USEC * 1000; + + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 396)) | DBG_FUNC_NONE, + (uintptr_t)blkno, size, operation, 0, 0); + + err = msleep(bp, buf_mtxp, slpflag | PDROP | (PRIBIO + 1), "buf_getblk", &ts); + /* * Callers who call with PCATCH or timeout are * willing to deal with the NULL pointer */ - if (err && ((slpflag & PCATCH) || - ((err == EWOULDBLOCK) && slptimeo))) + if (err && ((slpflag & PCATCH) || ((err == EWOULDBLOCK) && slptimeo))) return (NULL); goto start; /*NOTREACHED*/ - break; - - case BLK_PAGEIN: - /* pagein operation must not use getblk */ - panic("getblk: pagein for incore busy buffer"); - splx(s); - /*NOTREACHED*/ - break; - - case BLK_PAGEOUT: - /* pageout operation must not use getblk */ - panic("getblk: pageout for incore busy buffer"); - splx(s); - /*NOTREACHED*/ - break; default: - panic("getblk: %d unknown operation 1", operation); + /* + * unknown operation requested + */ + panic("getblk: paging or unknown operation for incore busy buffer - %x\n", operation); /*NOTREACHED*/ break; } } else { - /* not busy */ - SET(bp->b_flags, (B_BUSY | B_CACHE)); - bremfree(bp); + int clear_bdone; + + /* + * buffer in core and not busy + */ + SET(bp->b_lflags, BL_BUSY); + SET(bp->b_flags, B_CACHE); + buf_busycount++; + + bremfree_locked(bp); bufstats.bufs_incore++; - splx(s); + + lck_mtx_unlock(buf_mtxp); +#ifdef JOE_DEBUG + bp->b_owner = current_thread(); + bp->b_tag = 1; +#endif + if ( (bp->b_upl) ) + panic("buffer has UPL, but not marked BUSY: %p", bp); + + clear_bdone = FALSE; + if (!ret_only_valid) { + /* + * If the number bytes that are valid is going + * to increase (even if we end up not doing a + * reallocation through allocbuf) we have to read + * the new size first. + * + * This is required in cases where we doing a read + * modify write of a already valid data on disk but + * in cases where the data on disk beyond (blkno + b_bcount) + * is invalid, we may end up doing extra I/O. + */ + if (operation == BLK_META && bp->b_bcount < size) { + /* + * Since we are going to read in the whole size first + * we first have to ensure that any pending delayed write + * is flushed to disk first. + */ + if (ISSET(bp->b_flags, B_DELWRI)) { + CLR(bp->b_flags, B_CACHE); + buf_bwrite(bp); + goto start; + } + /* + * clear B_DONE before returning from + * this function so that the caller can + * can issue a read for the new size. + */ + clear_bdone = TRUE; + } - allocbuf(bp, size); - if (ISSET(bp->b_flags, B_PAGELIST)) - panic("pagelist buffer is not busy"); + if (bp->b_bufsize != size) + allocbuf(bp, size); + } + upl_flags = 0; switch (operation) { - case BLK_READ: case BLK_WRITE: - if (UBCISVALID(bp->b_vp) && bp->b_bufsize) { + /* + * "write" operation: let the UPL subsystem + * know that we intend to modify the buffer + * cache pages we're gathering. + */ + upl_flags |= UPL_WILL_MODIFY; + case BLK_READ: + upl_flags |= UPL_PRECIOUS; + if (UBCINFOEXISTS(bp->b_vp) && bp->b_bufsize) { kret = ubc_create_upl(vp, - ubc_blktooff(vp, bp->b_lblkno), - bp->b_bufsize, - &upl, - &pl, - UPL_PRECIOUS); + ubc_blktooff(vp, bp->b_lblkno), + bp->b_bufsize, + &upl, + &pl, + upl_flags); if (kret != KERN_SUCCESS) - panic("Failed to get pagelists"); + panic("Failed to create UPL"); - SET(bp->b_flags, B_PAGELIST); - bp->b_pagelist = upl; + bp->b_upl = upl; - if (!upl_valid_page(pl, 0)) { - if (vp->v_tag != VT_NFS) - panic("getblk: incore buffer without valid page"); - CLR(bp->b_flags, B_CACHE); - } + if (upl_valid_page(pl, 0)) { + if (upl_dirty_page(pl, 0)) + SET(bp->b_flags, B_WASDIRTY); + else + CLR(bp->b_flags, B_WASDIRTY); + } else + CLR(bp->b_flags, (B_DONE | B_CACHE | B_WASDIRTY | B_DELWRI)); - if (upl_dirty_page(pl, 0)) - SET(bp->b_flags, B_WASDIRTY); - else - CLR(bp->b_flags, B_WASDIRTY); + kret = ubc_upl_map(upl, (vm_offset_t*)&(bp->b_datap)); - kret = ubc_upl_map(upl, (vm_address_t *)&(bp->b_data)); if (kret != KERN_SUCCESS) - panic("getblk: ubc_upl_map() failed with (%d)", - kret); - if (bp->b_data == 0) - panic("ubc_upl_map mapped 0"); + panic("getblk: ubc_upl_map() failed with (%d)", kret); } break; @@ -1090,48 +3045,57 @@ start: * VM is not involved in IO for the meta data * buffer already has valid data */ - if(bp->b_data == 0) - panic("bp->b_data null incore buf=%x", bp); - break; - - case BLK_PAGEIN: - case BLK_PAGEOUT: - panic("getblk: paging operation 1"); break; default: - panic("getblk: %d unknown operation 2", operation); + panic("getblk: paging or unknown operation for incore buffer- %d\n", operation); /*NOTREACHED*/ break; } + + if (clear_bdone) + CLR(bp->b_flags, B_DONE); } } else { /* not incore() */ int queue = BQ_EMPTY; /* Start with no preference */ - splx(s); - if ((operation == BLK_META) || (UBCINVALID(vp)) || - !(UBCINFOEXISTS(vp))) { - operation = BLK_META; + if (ret_only_valid) { + lck_mtx_unlock(buf_mtxp); + return (NULL); } + if ((vnode_isreg(vp) == 0) || (UBCINFOEXISTS(vp) == 0) /*|| (vnode_issystem(vp) == 1)*/) + operation = BLK_META; + if ((bp = getnewbuf(slpflag, slptimeo, &queue)) == NULL) goto start; - if (incore(vp, blkno)) { + + /* + * getnewbuf may block for a number of different reasons... + * if it does, it's then possible for someone else to + * create a buffer for the same block and insert it into + * the hash... if we see it incore at this point we dump + * the buffer we were working on and start over + */ + if (incore_locked(vp, blkno, dp)) { SET(bp->b_flags, B_INVAL); binshash(bp, &invalhash); - brelse(bp); + + lck_mtx_unlock(buf_mtxp); + + buf_brelse(bp); goto start; } + /* + * NOTE: YOU CAN NOT BLOCK UNTIL binshash() HAS BEEN + * CALLED! BE CAREFUL. + */ /* - * if it is meta, the queue may be set to other - * type so reset as well as mark it to be B_META + * mark the buffer as B_META if indicated * so that when buffer is released it will goto META queue - * Also, if the vnode is not VREG, then it is META */ - if (operation == BLK_META) { - SET(bp->b_flags, B_META); - queue = BQ_META; - } + if (operation == BLK_META) + SET(bp->b_flags, B_META); bp->b_blkno = bp->b_lblkno = blkno; bp->b_vp = vp; @@ -1141,161 +3105,156 @@ start: */ binshash(bp, BUFHASH(vp, blkno)); - s = splbio(); - bgetvp(vp, bp); - splx(s); + bgetvp_locked(vp, bp); + + lck_mtx_unlock(buf_mtxp); allocbuf(bp, size); + upl_flags = 0; switch (operation) { case BLK_META: - /* buffer data is invalid */ - - if(bp->b_data == 0) - panic("bp->b_data is null %x",bp); - - bufstats.bufs_miss++; - - /* wakeup the buffer */ - CLR(bp->b_flags, B_WANTED); - wakeup(bp); + /* + * buffer data is invalid... + * + * I don't want to have to retake buf_mtxp, + * so the miss and vmhits counters are done + * with Atomic updates... all other counters + * in bufstats are protected with either + * buf_mtxp or iobuffer_mtxp + */ + OSAddAtomicLong(1, &bufstats.bufs_miss); break; - case BLK_READ: case BLK_WRITE: + /* + * "write" operation: let the UPL subsystem know + * that we intend to modify the buffer cache pages + * we're gathering. + */ + upl_flags |= UPL_WILL_MODIFY; + case BLK_READ: + { off_t f_offset; + size_t contig_bytes; + int bmap_flags; - if (ISSET(bp->b_flags, B_PAGELIST)) - panic("B_PAGELIST in bp=%x",bp); - - kret = ubc_create_upl(vp, - ubc_blktooff(vp, blkno), - bp->b_bufsize, - &upl, - &pl, - UPL_PRECIOUS); - if (kret != KERN_SUCCESS) - panic("Failed to get pagelists"); +#if DEVELOPMENT || DEBUG + /* + * Apple implemented file systems use UBC excludively; they should + * not call in here." + */ + const char* excldfs[] = {"hfs", "afpfs", "smbfs", "acfs", + "exfat", "msdos", "webdav", NULL}; + + for (int i = 0; excldfs[i] != NULL; i++) { + if (vp->v_mount && + !strcmp(vp->v_mount->mnt_vfsstat.f_fstypename, + excldfs[i])) { + panic("%s %s calls buf_getblk", + excldfs[i], + operation == BLK_READ ? "BLK_READ" : "BLK_WRITE"); + } + } +#endif -#ifdef UBC_DEBUG - upl_ubc_alias_set(upl, bp, 4); -#endif /* UBC_DEBUG */ - bp->b_pagelist = upl; + if ( (bp->b_upl) ) + panic("bp already has UPL: %p",bp); - SET(bp->b_flags, B_PAGELIST); + f_offset = ubc_blktooff(vp, blkno); - if (upl_valid_page(pl, 0)) { - SET(bp->b_flags, B_CACHE | B_DONE); - bufstats.bufs_vmhits++; + upl_flags |= UPL_PRECIOUS; + kret = ubc_create_upl(vp, + f_offset, + bp->b_bufsize, + &upl, + &pl, + upl_flags); - pagedirty = upl_dirty_page(pl, 0); + if (kret != KERN_SUCCESS) + panic("Failed to create UPL"); +#if UPL_DEBUG + upl_ubc_alias_set(upl, (uintptr_t) bp, (uintptr_t) 4); +#endif /* UPL_DEBUG */ + bp->b_upl = upl; - if (pagedirty) - SET(bp->b_flags, B_WASDIRTY); + if (upl_valid_page(pl, 0)) { - if (vp->v_tag == VT_NFS) { - off_t f_offset; - int valid_size; + if (operation == BLK_READ) + bmap_flags = VNODE_READ; + else + bmap_flags = VNODE_WRITE; - bp->b_validoff = 0; - bp->b_dirtyoff = 0; + SET(bp->b_flags, B_CACHE | B_DONE); - f_offset = ubc_blktooff(vp, blkno); + OSAddAtomicLong(1, &bufstats.bufs_vmhits); - if (f_offset > vp->v_ubcinfo->ui_size) { - CLR(bp->b_flags, (B_CACHE|B_DONE|B_WASDIRTY)); - bp->b_validend = 0; - bp->b_dirtyend = 0; - } else { - valid_size = min(((unsigned int)(vp->v_ubcinfo->ui_size - f_offset)), PAGE_SIZE); - bp->b_validend = valid_size; + bp->b_validoff = 0; + bp->b_dirtyoff = 0; - if (pagedirty) - bp->b_dirtyend = valid_size; - else - bp->b_dirtyend = 0; + if (upl_dirty_page(pl, 0)) { + /* page is dirty */ + SET(bp->b_flags, B_WASDIRTY); - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 386)) | DBG_FUNC_NONE, - bp->b_validend, bp->b_dirtyend, - (int)vp->v_ubcinfo->ui_size, 0, 0); - } + bp->b_validend = bp->b_bcount; + bp->b_dirtyend = bp->b_bcount; } else { - bp->b_validoff = 0; - bp->b_dirtyoff = 0; - - if (pagedirty) { - /* page is dirty */ - bp->b_validend = bp->b_bcount; - bp->b_dirtyend = bp->b_bcount; - } else { - /* page is clean */ - bp->b_validend = bp->b_bcount; - bp->b_dirtyend = 0; - } - } - error = VOP_BMAP(vp, bp->b_lblkno, NULL, &bp->b_blkno, NULL); - if(error) { - panic("getblk: VOP_BMAP failed"); - /*NOTREACHED*/ - /* - * XXX: We probably should invalidate the VM Page - */ - bp->b_error = error; - SET(bp->b_flags, (B_ERROR | B_INVAL)); - /* undo B_DONE that was set before upl_commit() */ - CLR(bp->b_flags, B_DONE); - brelse(bp); - return (0); + /* page is clean */ + bp->b_validend = bp->b_bcount; + bp->b_dirtyend = 0; } + /* + * try to recreate the physical block number associated with + * this buffer... + */ + if (VNOP_BLOCKMAP(vp, f_offset, bp->b_bcount, &bp->b_blkno, &contig_bytes, NULL, bmap_flags, NULL)) + panic("getblk: VNOP_BLOCKMAP failed"); + /* + * if the extent represented by this buffer + * is not completely physically contiguous on + * disk, than we can't cache the physical mapping + * in the buffer header + */ + if ((long)contig_bytes < bp->b_bcount) + bp->b_blkno = bp->b_lblkno; } else { - bufstats.bufs_miss++; - } - kret = ubc_upl_map(upl, (vm_address_t *)&(bp->b_data)); - if (kret != KERN_SUCCESS) { - panic("getblk: ubc_upl_map() " - "failed with (%d)", kret); + OSAddAtomicLong(1, &bufstats.bufs_miss); } - if (bp->b_data == 0) - panic("kernel_upl_map mapped 0"); - - break; + kret = ubc_upl_map(upl, (vm_offset_t *)&(bp->b_datap)); - case BLK_PAGEIN: - case BLK_PAGEOUT: - panic("getblk: paging operation 2"); + if (kret != KERN_SUCCESS) + panic("getblk: ubc_upl_map() failed with (%d)", kret); break; + } default: - panic("getblk: %d unknown operation 3", operation); + panic("getblk: paging or unknown operation - %x", operation); /*NOTREACHED*/ break; } } - - if (bp->b_data == NULL) - panic("getblk: bp->b_addr is null"); - - if (bp->b_bufsize & 0xfff) { - if (ISSET(bp->b_flags, B_META) && (bp->b_bufsize & 0x1ff)) - panic("getblk: bp->b_bufsize = %d", bp->b_bufsize); - } - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 386)) | DBG_FUNC_END, - (int)bp, (int)bp->b_data, bp->b_flags, 3, 0); + bp, bp->b_datap, bp->b_flags, 3, 0); +#ifdef JOE_DEBUG + (void) OSBacktrace(&bp->b_stackgetblk[0], 6); +#endif return (bp); } /* * Get an empty, disassociated buffer of given size. */ -struct buf * -geteblk(size) - int size; +buf_t +buf_geteblk(int size) { - struct buf *bp; - int queue = BQ_EMPTY; + buf_t bp = NULL; + int queue = BQ_EMPTY; + + do { + lck_mtx_lock_spin(buf_mtxp); + + bp = getnewbuf(0, 0, &queue); + } while (bp == NULL); - while ((bp = getnewbuf(0, 0, &queue)) == 0) - ; SET(bp->b_flags, (B_META|B_INVAL)); #if DIAGNOSTIC @@ -1304,66 +3263,105 @@ geteblk(size) /* XXX need to implement logic to deal with other queues */ binshash(bp, &invalhash); - allocbuf(bp, size); bufstats.bufs_eblk++; + lck_mtx_unlock(buf_mtxp); + + allocbuf(bp, size); + return (bp); } -/* - * Zones for the meta data buffers - */ +uint32_t +buf_redundancy_flags(buf_t bp) +{ + return bp->b_redundancy_flags; +} -#define MINMETA 512 -#define MAXMETA 4096 +void +buf_set_redundancy_flags(buf_t bp, uint32_t flags) +{ + SET(bp->b_redundancy_flags, flags); +} -struct meta_zone_entry { - zone_t mz_zone; - vm_size_t mz_size; - vm_size_t mz_max; - char *mz_name; -}; +void +buf_clear_redundancy_flags(buf_t bp, uint32_t flags) +{ + CLR(bp->b_redundancy_flags, flags); +} -struct meta_zone_entry meta_zones[] = { - {NULL, (MINMETA * 1), 128 * (MINMETA * 1), "buf.512" }, - {NULL, (MINMETA * 2), 64 * (MINMETA * 2), "buf.1024" }, - {NULL, (MINMETA * 4), 16 * (MINMETA * 4), "buf.2048" }, - {NULL, (MINMETA * 8), 512 * (MINMETA * 8), "buf.4096" }, - {NULL, 0, 0, "" } /* End */ -}; -/* - * Initialize the meta data zones - */ -static void -bufzoneinit(void) + +static void * +recycle_buf_from_pool(int nsize) { - int i; + buf_t bp; + void *ptr = NULL; - for (i = 0; meta_zones[i].mz_size != 0; i++) { - meta_zones[i].mz_zone = - zinit(meta_zones[i].mz_size, - meta_zones[i].mz_max, - PAGE_SIZE, - meta_zones[i].mz_name); + lck_mtx_lock_spin(buf_mtxp); + + TAILQ_FOREACH(bp, &bufqueues[BQ_META], b_freelist) { + if (ISSET(bp->b_flags, B_DELWRI) || bp->b_bufsize != nsize) + continue; + ptr = (void *)bp->b_datap; + bp->b_bufsize = 0; + + bcleanbuf(bp, TRUE); + break; } - buf_hdr_zone = zinit(sizeof(struct buf), 32, PAGE_SIZE, "buf headers"); + lck_mtx_unlock(buf_mtxp); + + return (ptr); } -static __inline__ zone_t -getbufzone(size_t size) + + +int zalloc_nopagewait_failed = 0; +int recycle_buf_failed = 0; + +static void * +grab_memory_for_meta_buf(int nsize) { - int i; + zone_t z; + void *ptr; + boolean_t was_vmpriv; - if ((size % 512) || (size < MINMETA) || (size > MAXMETA)) - panic("getbufzone: incorect size = %d", size); + z = getbufzone(nsize); - for (i = 0; meta_zones[i].mz_size != 0; i++) { - if (meta_zones[i].mz_size >= size) - break; - } + /* + * make sure we're NOT priviliged so that + * if a vm_page_grab is needed, it won't + * block if we're out of free pages... if + * it blocks, then we can't honor the + * nopagewait request + */ + was_vmpriv = set_vm_privilege(FALSE); - return (meta_zones[i].mz_zone); + ptr = zalloc_nopagewait(z); + + if (was_vmpriv == TRUE) + set_vm_privilege(TRUE); + + if (ptr == NULL) { + + zalloc_nopagewait_failed++; + + ptr = recycle_buf_from_pool(nsize); + + if (ptr == NULL) { + + recycle_buf_failed++; + + if (was_vmpriv == FALSE) + set_vm_privilege(TRUE); + + ptr = zalloc(z); + + if (was_vmpriv == FALSE) + set_vm_privilege(FALSE); + } + } + return (ptr); } /* @@ -1379,82 +3377,75 @@ getbufzone(size_t size) */ int -allocbuf(bp, size) - struct buf *bp; - int size; +allocbuf(buf_t bp, int size) { vm_size_t desired_size; desired_size = roundup(size, CLBYTES); - if(desired_size < PAGE_SIZE) + if (desired_size < PAGE_SIZE) desired_size = PAGE_SIZE; if (desired_size > MAXBSIZE) panic("allocbuf: buffer larger than MAXBSIZE requested"); if (ISSET(bp->b_flags, B_META)) { - kern_return_t kret; - zone_t zprev, z; - size_t nsize = roundup(size, MINMETA); - - if (bp->b_data) { - vm_offset_t elem = (vm_offset_t)bp->b_data; - - if (ISSET(bp->b_flags, B_ZALLOC)) - if (bp->b_bufsize <= MAXMETA) { - if (bp->b_bufsize < nsize) { - /* reallocate to a bigger size */ - desired_size = nsize; - - zprev = getbufzone(bp->b_bufsize); - z = getbufzone(nsize); - bp->b_data = (caddr_t)zalloc(z); - if(bp->b_data == 0) - panic("allocbuf: zalloc() returned NULL"); - bcopy(elem, bp->b_data, bp->b_bufsize); - zfree(zprev, elem); + int nsize = roundup(size, MINMETA); + + if (bp->b_datap) { + vm_offset_t elem = (vm_offset_t)bp->b_datap; + + if (ISSET(bp->b_flags, B_ZALLOC)) { + if (bp->b_bufsize < nsize) { + zone_t zprev; + + /* reallocate to a bigger size */ + + zprev = getbufzone(bp->b_bufsize); + if (nsize <= MAXMETA) { + desired_size = nsize; + + /* b_datap not really a ptr */ + *(void **)(&bp->b_datap) = grab_memory_for_meta_buf(nsize); } else { - desired_size = bp->b_bufsize; + bp->b_datap = (uintptr_t)NULL; + kmem_alloc_kobject(kernel_map, (vm_offset_t *)&bp->b_datap, desired_size, VM_KERN_MEMORY_FILE); + CLR(bp->b_flags, B_ZALLOC); } - } else - panic("allocbuf: B_ZALLOC set incorrectly"); - else - if (bp->b_bufsize < desired_size) { + bcopy((void *)elem, (caddr_t)bp->b_datap, bp->b_bufsize); + zfree(zprev, (void *)elem); + } else { + desired_size = bp->b_bufsize; + } + + } else { + if ((vm_size_t)bp->b_bufsize < desired_size) { /* reallocate to a bigger size */ - kret = kmem_alloc(kernel_map, &bp->b_data, desired_size); - if (kret != KERN_SUCCESS) - panic("allocbuf: kmem_alloc() returned %d", kret); - if(bp->b_data == 0) - panic("allocbuf: null b_data"); - bcopy(elem, bp->b_data, bp->b_bufsize); + bp->b_datap = (uintptr_t)NULL; + kmem_alloc_kobject(kernel_map, (vm_offset_t *)&bp->b_datap, desired_size, VM_KERN_MEMORY_FILE); + bcopy((const void *)elem, (caddr_t)bp->b_datap, bp->b_bufsize); kmem_free(kernel_map, elem, bp->b_bufsize); } else { desired_size = bp->b_bufsize; } + } } else { /* new allocation */ if (nsize <= MAXMETA) { desired_size = nsize; - z = getbufzone(nsize); - bp->b_data = (caddr_t)zalloc(z); - if(bp->b_data == 0) - panic("allocbuf: zalloc() returned NULL 2"); + + /* b_datap not really a ptr */ + *(void **)(&bp->b_datap) = grab_memory_for_meta_buf(nsize); SET(bp->b_flags, B_ZALLOC); - } else { - kret = kmem_alloc(kernel_map, &bp->b_data, desired_size); - if (kret != KERN_SUCCESS) - panic("allocbuf: kmem_alloc() 2 returned %d", kret); - if(bp->b_data == 0) - panic("allocbuf: null b_data 2"); - } + } else + kmem_alloc_kobject(kernel_map, (vm_offset_t *)&bp->b_datap, desired_size, VM_KERN_MEMORY_FILE); } - } - - if (ISSET(bp->b_flags, B_META) && (bp->b_data == 0)) - panic("allocbuf: bp->b_data is NULL"); + if (bp->b_datap == 0) + panic("allocbuf: NULL b_datap"); + } bp->b_bufsize = desired_size; bp->b_bcount = size; + return (0); } @@ -1474,36 +3465,51 @@ allocbuf(bp, size) * Initialize the fields and disassociate the buffer from the vnode. * Remove the buffer from the hash. Return the buffer and the queue * on which it was found. + * + * buf_mtxp is held upon entry + * returns with buf_mtxp locked if new buf available + * returns with buf_mtxp UNlocked if new buf NOT available */ -static struct buf * -getnewbuf(slpflag, slptimeo, queue) - int slpflag, slptimeo; - int *queue; -{ - register struct buf *bp; - register struct buf *lru_bp; - register struct buf *age_bp; - register struct buf *meta_bp; - register int age_time, lru_time, bp_time, meta_time; - int s; - int req = *queue; /* save it for restarts */ +static buf_t +getnewbuf(int slpflag, int slptimeo, int * queue) +{ + buf_t bp; + buf_t lru_bp; + buf_t age_bp; + buf_t meta_bp; + int age_time, lru_time, bp_time, meta_time; + int req = *queue; /* save it for restarts */ + struct timespec ts; start: - s = splbio(); - - /* invalid request gets empty queue */ - if ((*queue > BQUEUES) || (*queue < 0) + /* + * invalid request gets empty queue + */ + if ((*queue >= BQUEUES) || (*queue < 0) || (*queue == BQ_LAUNDRY) || (*queue == BQ_LOCKED)) *queue = BQ_EMPTY; - /* (*queue == BQUEUES) means no preference */ - if (*queue != BQUEUES) { - /* Try for the requested queue first */ - bp = bufqueues[*queue].tqh_first; - if (bp) - goto found; + + if (*queue == BQ_EMPTY && (bp = bufqueues[*queue].tqh_first)) + goto found; + + /* + * need to grow number of bufs, add another one rather than recycling + */ + if (nbuf_headers < max_nbuf_headers) { + /* + * Increment count now as lock + * is dropped for allocation. + * That avoids over commits + */ + nbuf_headers++; + goto add_newbufs; } + /* Try for the requested queue first */ + bp = bufqueues[*queue].tqh_first; + if (bp) + goto found; /* Unable to use requested queue */ age_bp = bufqueues[BQ_AGE].tqh_first; @@ -1520,30 +3526,48 @@ start: *queue = BQ_EMPTY; goto found; } + /* + * We have seen is this is hard to trigger. + * This is an overcommit of nbufs but needed + * in some scenarios with diskiamges + */ + +add_newbufs: + lck_mtx_unlock(buf_mtxp); - /* Create a new temparory buffer header */ + /* Create a new temporary buffer header */ bp = (struct buf *)zalloc(buf_hdr_zone); - + if (bp) { bufhdrinit(bp); + bp->b_whichq = BQ_EMPTY; + bp->b_timestamp = buf_timestamp(); BLISTNONE(bp); - binshash(bp, &invalhash); SET(bp->b_flags, B_HDRALLOC); *queue = BQ_EMPTY; + } + lck_mtx_lock_spin(buf_mtxp); + + if (bp) { + binshash(bp, &invalhash); binsheadfree(bp, &bufqueues[BQ_EMPTY], BQ_EMPTY); buf_hdr_count++; goto found; } + /* subtract already accounted bufcount */ + nbuf_headers--; - /* Log this error condition */ - printf("getnewbuf: No useful buffers"); + bufstats.bufs_sleeps++; /* wait for a free buffer of any kind */ needbuffer = 1; - bufstats.bufs_sleeps++; - tsleep(&needbuffer, slpflag|(PRIBIO+1), "getnewbuf", slptimeo); - splx(s); - return (0); + /* hz value is 100 */ + ts.tv_sec = (slptimeo/1000); + /* the hz value is 100; which leads to 10ms */ + ts.tv_nsec = (slptimeo % 1000) * NSEC_PER_USEC * 1000 * 10; + + msleep(&needbuffer, buf_mtxp, slpflag | PDROP | (PRIBIO+1), "getnewbuf", &ts); + return (NULL); } /* Buffer available either on AGE or LRU or META */ @@ -1558,8 +3582,10 @@ start: bp = age_bp; *queue = BQ_AGE; } else { /* buffer available on both AGE and LRU */ - age_time = time.tv_sec - age_bp->b_timestamp; - lru_time = time.tv_sec - lru_bp->b_timestamp; + int t = buf_timestamp(); + + age_time = t - age_bp->b_timestamp; + lru_time = t - lru_bp->b_timestamp; if ((age_time < 0) || (lru_time < 0)) { /* time set backwards */ bp = age_bp; *queue = BQ_AGE; @@ -1582,8 +3608,10 @@ start: bp = meta_bp; *queue = BQ_META; } else if (meta_bp) { - bp_time = time.tv_sec - bp->b_timestamp; - meta_time = time.tv_sec - meta_bp->b_timestamp; + int t = buf_timestamp(); + + bp_time = t - bp->b_timestamp; + meta_time = t - meta_bp->b_timestamp; if (!(bp_time < 0) && !(meta_time < 0)) { /* time not set backwards */ @@ -1598,134 +3626,273 @@ start: } } } - - if (bp == NULL) - panic("getnewbuf: null bp"); - found: - if (bp->b_hash.le_prev == (struct buf **)0xdeadbeef) - panic("getnewbuf: le_prev is deadbeef"); - - if(ISSET(bp->b_flags, B_BUSY)) - panic("getnewbuf reusing BUSY buf"); + if (ISSET(bp->b_flags, B_LOCKED) || ISSET(bp->b_lflags, BL_BUSY)) + panic("getnewbuf: bp @ %p is LOCKED or BUSY! (flags 0x%x)\n", bp, bp->b_flags); /* Clean it */ - if (bcleanbuf(bp)) { - /* bawrite() issued, buffer not ready */ - splx(s); + if (bcleanbuf(bp, FALSE)) { + /* + * moved to the laundry thread, buffer not ready + */ *queue = req; goto start; } - splx(s); return (bp); } -#include -#include -#include /* * Clean a buffer. - * Returns 0 is buffer is ready to use, - * Returns 1 if issued a bawrite() to indicate + * Returns 0 if buffer is ready to use, + * Returns 1 if issued a buf_bawrite() to indicate * that the buffer is not ready. + * + * buf_mtxp is held upon entry + * returns with buf_mtxp locked */ -static int -bcleanbuf(struct buf *bp) +int +bcleanbuf(buf_t bp, boolean_t discard) { - int s; - struct ucred *cred; - int hdralloc = 0; - - s = splbio(); - /* Remove from the queue */ - bremfree(bp); - - /* Buffer is no longer on free lists. */ - SET(bp->b_flags, B_BUSY); - - /* Check whether the buffer header was "allocated" */ - if (ISSET(bp->b_flags, B_HDRALLOC)) - hdralloc = 1; - - if (bp->b_hash.le_prev == (struct buf **)0xdeadbeef) - panic("bcleanbuf: le_prev is deadbeef"); + bremfree_locked(bp); +#ifdef JOE_DEBUG + bp->b_owner = current_thread(); + bp->b_tag = 2; +#endif /* * If buffer was a delayed write, start the IO by queuing * it on the LAUNDRY queue, and return 1 */ if (ISSET(bp->b_flags, B_DELWRI)) { - splx(s); - binstailfree(bp, &bufqueues[BQ_LAUNDRY], BQ_LAUNDRY); - blaundrycnt++; - wakeup(&blaundrycnt); - /* and give it a chance to run */ + if (discard) { + SET(bp->b_lflags, BL_WANTDEALLOC); + } + + bmovelaundry(bp); + + lck_mtx_unlock(buf_mtxp); + + wakeup(&bufqueues[BQ_LAUNDRY]); + /* + * and give it a chance to run + */ (void)thread_block(THREAD_CONTINUE_NULL); + + lck_mtx_lock_spin(buf_mtxp); + return (1); } +#ifdef JOE_DEBUG + bp->b_owner = current_thread(); + bp->b_tag = 8; +#endif + /* + * Buffer is no longer on any free list... we own it + */ + SET(bp->b_lflags, BL_BUSY); + buf_busycount++; + + bremhash(bp); + /* + * disassociate us from our vnode, if we had one... + */ if (bp->b_vp) - brelvp(bp); - bremhash(bp); + brelvp_locked(bp); + + lck_mtx_unlock(buf_mtxp); + BLISTNONE(bp); - splx(s); + if (ISSET(bp->b_flags, B_META)) + buf_free_meta_store(bp); - if (ISSET(bp->b_flags, B_META)) { - vm_offset_t elem = (vm_offset_t)bp->b_data; - if (elem == 0) - panic("bcleanbuf: NULL bp->b_data B_META buffer"); + trace(TR_BRELSE, pack(bp->b_vp, bp->b_bufsize), bp->b_lblkno); - if (ISSET(bp->b_flags, B_ZALLOC)) { - if (bp->b_bufsize <= MAXMETA) { - zone_t z; + buf_release_credentials(bp); + + /* If discarding, just move to the empty queue */ + if (discard) { + lck_mtx_lock_spin(buf_mtxp); + CLR(bp->b_flags, (B_META | B_ZALLOC | B_DELWRI | B_LOCKED | B_AGE | B_ASYNC | B_NOCACHE | B_FUA)); + bp->b_whichq = BQ_EMPTY; + binshash(bp, &invalhash); + binsheadfree(bp, &bufqueues[BQ_EMPTY], BQ_EMPTY); + CLR(bp->b_lflags, BL_BUSY); + buf_busycount--; + } else { + /* Not discarding: clean up and prepare for reuse */ + bp->b_bufsize = 0; + bp->b_datap = (uintptr_t)NULL; + bp->b_upl = (void *)NULL; + bp->b_fsprivate = (void *)NULL; + /* + * preserve the state of whether this buffer + * was allocated on the fly or not... + * the only other flag that should be set at + * this point is BL_BUSY... + */ +#ifdef JOE_DEBUG + bp->b_owner = current_thread(); + bp->b_tag = 3; +#endif + bp->b_lflags = BL_BUSY; + bp->b_flags = (bp->b_flags & B_HDRALLOC); + bp->b_redundancy_flags = 0; + bp->b_dev = NODEV; + bp->b_blkno = bp->b_lblkno = 0; + bp->b_iodone = NULL; + bp->b_error = 0; + bp->b_resid = 0; + bp->b_bcount = 0; + bp->b_dirtyoff = bp->b_dirtyend = 0; + bp->b_validoff = bp->b_validend = 0; + bzero(&bp->b_attr, sizeof(struct bufattr)); + + lck_mtx_lock_spin(buf_mtxp); + } + return (0); +} - z = getbufzone(bp->b_bufsize); - bp->b_data = (caddr_t)0xdeadbeef; - zfree(z, elem); - CLR(bp->b_flags, B_ZALLOC); - } else - panic("bcleanbuf: B_ZALLOC set incorrectly"); - } else { - bp->b_data = (caddr_t)0xdeadbeef; - kmem_free(kernel_map, elem, bp->b_bufsize); + + +errno_t +buf_invalblkno(vnode_t vp, daddr64_t lblkno, int flags) +{ + buf_t bp; + errno_t error; + struct bufhashhdr *dp; + + dp = BUFHASH(vp, lblkno); + +relook: + lck_mtx_lock_spin(buf_mtxp); + + if ((bp = incore_locked(vp, lblkno, dp)) == (struct buf *)0) { + lck_mtx_unlock(buf_mtxp); + return (0); + } + if (ISSET(bp->b_lflags, BL_BUSY)) { + if ( !ISSET(flags, BUF_WAIT)) { + lck_mtx_unlock(buf_mtxp); + return (EBUSY); + } + SET(bp->b_lflags, BL_WANTED); + + error = msleep((caddr_t)bp, buf_mtxp, PDROP | (PRIBIO + 1), "buf_invalblkno", NULL); + + if (error) { + return (error); } + goto relook; } + bremfree_locked(bp); + SET(bp->b_lflags, BL_BUSY); + SET(bp->b_flags, B_INVAL); + buf_busycount++; +#ifdef JOE_DEBUG + bp->b_owner = current_thread(); + bp->b_tag = 4; +#endif + lck_mtx_unlock(buf_mtxp); + buf_brelse(bp); - trace(TR_BRELSE, pack(bp->b_vp, bp->b_bufsize), bp->b_lblkno); + return (0); +} - /* disassociate us from our vnode, if we had one... */ - s = splbio(); - /* clear out various other fields */ - bp->b_bufsize = 0; - bp->b_data = 0; - bp->b_flags = B_BUSY; - if (hdralloc) - SET(bp->b_flags, B_HDRALLOC); - bp->b_dev = NODEV; - bp->b_blkno = bp->b_lblkno = 0; - bp->b_iodone = 0; - bp->b_error = 0; - bp->b_resid = 0; - bp->b_bcount = 0; - bp->b_dirtyoff = bp->b_dirtyend = 0; - bp->b_validoff = bp->b_validend = 0; +void +buf_drop(buf_t bp) +{ + int need_wakeup = 0; + + lck_mtx_lock_spin(buf_mtxp); + + if (ISSET(bp->b_lflags, BL_WANTED)) { + /* + * delay the actual wakeup until after we + * clear BL_BUSY and we've dropped buf_mtxp + */ + need_wakeup = 1; + } +#ifdef JOE_DEBUG + bp->b_owner = current_thread(); + bp->b_tag = 9; +#endif + /* + * Unlock the buffer. + */ + CLR(bp->b_lflags, (BL_BUSY | BL_WANTED)); + buf_busycount--; + + lck_mtx_unlock(buf_mtxp); + + if (need_wakeup) { + /* + * Wake up any proceeses waiting for _this_ buffer to become free. + */ + wakeup(bp); + } +} + + +errno_t +buf_acquire(buf_t bp, int flags, int slpflag, int slptimeo) { + errno_t error; + + lck_mtx_lock_spin(buf_mtxp); - /* nuke any credentials we were holding */ - cred = bp->b_rcred; - if (cred != NOCRED) { - bp->b_rcred = NOCRED; - crfree(cred); + error = buf_acquire_locked(bp, flags, slpflag, slptimeo); + + lck_mtx_unlock(buf_mtxp); + + return (error); +} + + +static errno_t +buf_acquire_locked(buf_t bp, int flags, int slpflag, int slptimeo) +{ + errno_t error; + struct timespec ts; + + if (ISSET(bp->b_flags, B_LOCKED)) { + if ((flags & BAC_SKIP_LOCKED)) + return (EDEADLK); + } else { + if ((flags & BAC_SKIP_NONLOCKED)) + return (EDEADLK); } - cred = bp->b_wcred; - if (cred != NOCRED) { - bp->b_wcred = NOCRED; - crfree(cred); + if (ISSET(bp->b_lflags, BL_BUSY)) { + /* + * since the lck_mtx_lock may block, the buffer + * may become BUSY, so we need to + * recheck for a NOWAIT request + */ + if (flags & BAC_NOWAIT) + return (EBUSY); + SET(bp->b_lflags, BL_WANTED); + + /* the hz value is 100; which leads to 10ms */ + ts.tv_sec = (slptimeo/100); + ts.tv_nsec = (slptimeo % 100) * 10 * NSEC_PER_USEC * 1000; + error = msleep((caddr_t)bp, buf_mtxp, slpflag | (PRIBIO + 1), "buf_acquire", &ts); + + if (error) + return (error); + return (EAGAIN); } - splx(s); + if (flags & BAC_REMOVE) + bremfree_locked(bp); + SET(bp->b_lflags, BL_BUSY); + buf_busycount++; + +#ifdef JOE_DEBUG + bp->b_owner = current_thread(); + bp->b_tag = 5; +#endif return (0); } @@ -1734,17 +3901,20 @@ bcleanbuf(struct buf *bp) * Wait for operations on the buffer to complete. * When they do, extract and return the I/O's error value. */ -int -biowait(bp) - struct buf *bp; +errno_t +buf_biowait(buf_t bp) { - int s; + while (!ISSET(bp->b_flags, B_DONE)) { - s = splbio(); - while (!ISSET(bp->b_flags, B_DONE)) - tsleep(bp, PRIBIO + 1, "biowait", 0); - splx(s); - + lck_mtx_lock_spin(buf_mtxp); + + if (!ISSET(bp->b_flags, B_DONE)) { + DTRACE_IO1(wait__start, buf_t, bp); + (void) msleep(bp, buf_mtxp, PDROP | (PRIBIO+1), "buf_biowait", NULL); + DTRACE_IO1(wait__done, buf_t, bp); + } else + lck_mtx_unlock(buf_mtxp); + } /* check for interruption of I/O (e.g. via NFS), then errors. */ if (ISSET(bp->b_flags, B_EINTR)) { CLR(bp->b_flags, B_EINTR); @@ -1755,6 +3925,7 @@ biowait(bp) return (0); } + /* * Mark I/O complete on a buffer. * @@ -1767,104 +3938,208 @@ biowait(bp) * process, invokes a procedure specified in the buffer structure" ] * * In real life, the pagedaemon (or other system processes) wants - * to do async stuff to, and doesn't want the buffer brelse()'d. + * to do async stuff to, and doesn't want the buffer buf_brelse()'d. * (for swap pager, that puts swap buffers on the free lists (!!!), * for the vn device, that puts malloc'd buffers on the free lists!) */ + void -biodone(bp) - struct buf *bp; +buf_biodone(buf_t bp) { - boolean_t funnel_state; - struct vnode *vp; + mount_t mp; + struct bufattr *bap; + + KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 387)) | DBG_FUNC_START, + bp, bp->b_datap, bp->b_flags, 0, 0); + + if (ISSET(bp->b_flags, B_DONE)) + panic("biodone already"); + + bap = &bp->b_attr; + + if (bp->b_vp && bp->b_vp->v_mount) { + mp = bp->b_vp->v_mount; + } else { + mp = NULL; + } + + if (ISSET(bp->b_flags, B_ERROR)) { + if (mp && (MNT_ROOTFS & mp->mnt_flag)) { + dk_error_description_t desc; + bzero(&desc, sizeof(desc)); + desc.description = panic_disk_error_description; + desc.description_size = panic_disk_error_description_size; + VNOP_IOCTL(mp->mnt_devvp, DKIOCGETERRORDESCRIPTION, (caddr_t)&desc, 0, vfs_context_kernel()); + } + } + + if (mp && (bp->b_flags & B_READ) == 0) { + update_last_io_time(mp); + INCR_PENDING_IO(-(pending_io_t)buf_count(bp), mp->mnt_pending_write_size); + } else if (mp) { + INCR_PENDING_IO(-(pending_io_t)buf_count(bp), mp->mnt_pending_read_size); + } - funnel_state = thread_funnel_set(kernel_flock, TRUE); + throttle_info_end_io(bp); - KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 387)) | DBG_FUNC_START, - (int)bp, (int)bp->b_data, bp->b_flags, 0, 0); + if (kdebug_enable) { + int code = DKIO_DONE; + int io_tier = GET_BUFATTR_IO_TIER(bap); + + if (bp->b_flags & B_READ) + code |= DKIO_READ; + if (bp->b_flags & B_ASYNC) + code |= DKIO_ASYNC; + + if (bp->b_flags & B_META) + code |= DKIO_META; + else if (bp->b_flags & B_PAGEIO) + code |= DKIO_PAGING; + + if (io_tier != 0) + code |= DKIO_THROTTLE; + + code |= ((io_tier << DKIO_TIER_SHIFT) & DKIO_TIER_MASK); + + if (bp->b_flags & B_PASSIVE) + code |= DKIO_PASSIVE; + + if (bap->ba_flags & BA_NOCACHE) + code |= DKIO_NOCACHE; + + if (bap->ba_flags & BA_IO_TIER_UPGRADE) { + code |= DKIO_TIER_UPGRADE; + } + + KERNEL_DEBUG_CONSTANT_IST(KDEBUG_COMMON, FSDBG_CODE(DBG_DKRW, code) | DBG_FUNC_NONE, + buf_kernel_addrperm_addr(bp), (uintptr_t)VM_KERNEL_ADDRPERM(bp->b_vp), bp->b_resid, bp->b_error, 0); + } - if (ISSET(bp->b_flags, B_DONE)) - panic("biodone already"); - SET(bp->b_flags, B_DONE); /* note that it's done */ /* * I/O was done, so don't believe - * the DIRTY state from VM anymore + * the DIRTY state from VM anymore... + * and we need to reset the THROTTLED/PASSIVE + * indicators */ - CLR(bp->b_flags, B_WASDIRTY); + CLR(bp->b_flags, (B_WASDIRTY | B_PASSIVE)); + CLR(bap->ba_flags, (BA_META | BA_NOCACHE | BA_DELAYIDLESLEEP | BA_IO_TIER_UPGRADE)); + + SET_BUFATTR_IO_TIER(bap, 0); + + DTRACE_IO1(done, buf_t, bp); if (!ISSET(bp->b_flags, B_READ) && !ISSET(bp->b_flags, B_RAW)) - vwakeup(bp); /* wake up reader */ - - if (kdebug_enable) { - int code = DKIO_DONE; - - if (bp->b_flags & B_READ) - code |= DKIO_READ; - if (bp->b_flags & B_ASYNC) - code |= DKIO_ASYNC; - - if (bp->b_flags & B_META) - code |= DKIO_META; - else if (bp->b_flags & (B_PGIN | B_PAGEOUT)) - code |= DKIO_PAGING; - - KERNEL_DEBUG_CONSTANT(FSDBG_CODE(DBG_DKRW, code) | DBG_FUNC_NONE, - bp, bp->b_vp, bp->b_resid, bp->b_error, 0); - } - - /* Wakeup the throttled write operations as needed */ - vp = bp->b_vp; - if (vp - && (vp->v_flag & VTHROTTLED) - && (vp->v_numoutput <= (BUFWRITE_THROTTLE / 3))) { - vp->v_flag &= ~VTHROTTLED; - wakeup((caddr_t)&vp->v_numoutput); - } - - if (ISSET(bp->b_flags, B_CALL)) { /* if necessary, call out */ - CLR(bp->b_flags, B_CALL); /* but note callout done */ - (*bp->b_iodone)(bp); - } else if (ISSET(bp->b_flags, B_ASYNC)) /* if async, release it */ - brelse(bp); - else { /* or just wakeup the buffer */ - CLR(bp->b_flags, B_WANTED); - wakeup(bp); + /* + * wake up any writer's blocked + * on throttle or waiting for I/O + * to drain + */ + vnode_writedone(bp->b_vp); + + if (ISSET(bp->b_flags, (B_CALL | B_FILTER))) { /* if necessary, call out */ + void (*iodone_func)(struct buf *, void *) = bp->b_iodone; + void *arg = bp->b_transaction; + int callout = ISSET(bp->b_flags, B_CALL); + + if (iodone_func == NULL) + panic("biodone: bp @ %p has NULL b_iodone!\n", bp); + + CLR(bp->b_flags, (B_CALL | B_FILTER)); /* filters and callouts are one-shot */ + bp->b_iodone = NULL; + bp->b_transaction = NULL; + + if (callout) + SET(bp->b_flags, B_DONE); /* note that it's done */ + + (*iodone_func)(bp, arg); + + if (callout) { + /* + * assumes that the callback function takes + * ownership of the bp and deals with releasing it if necessary + */ + goto biodone_done; + } + /* + * in this case the call back function is acting + * strictly as a filter... it does not take + * ownership of the bp and is expecting us + * to finish cleaning up... this is currently used + * by the HFS journaling code + */ } + if (ISSET(bp->b_flags, B_ASYNC)) { /* if async, release it */ + SET(bp->b_flags, B_DONE); /* note that it's done */ + + buf_brelse(bp); + } else { /* or just wakeup the buffer */ + /* + * by taking the mutex, we serialize + * the buf owner calling buf_biowait so that we'll + * only see him in one of 2 states... + * state 1: B_DONE wasn't set and he's + * blocked in msleep + * state 2: he's blocked trying to take the + * mutex before looking at B_DONE + * BL_WANTED is cleared in case anyone else + * is blocked waiting for the buffer... note + * that we haven't cleared B_BUSY yet, so if + * they do get to run, their going to re-set + * BL_WANTED and go back to sleep + */ + lck_mtx_lock_spin(buf_mtxp); + CLR(bp->b_lflags, BL_WANTED); + SET(bp->b_flags, B_DONE); /* note that it's done */ + + lck_mtx_unlock(buf_mtxp); + + wakeup(bp); + } +biodone_done: KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 387)) | DBG_FUNC_END, - (int)bp, (int)bp->b_data, bp->b_flags, 0, 0); + (uintptr_t)bp, (uintptr_t)bp->b_datap, bp->b_flags, 0, 0); +} - thread_funnel_set(kernel_flock, funnel_state); +/* + * Obfuscate buf pointers. + */ +vm_offset_t +buf_kernel_addrperm_addr(void * addr) +{ + if ((vm_offset_t)addr == 0) + return 0; + else + return ((vm_offset_t)addr + buf_kernel_addrperm); } /* * Return a count of buffers on the "locked" queue. */ int -count_lock_queue() +count_lock_queue(void) { - register struct buf *bp; - register int n = 0; + buf_t bp; + int n = 0; + + lck_mtx_lock_spin(buf_mtxp); for (bp = bufqueues[BQ_LOCKED].tqh_first; bp; bp = bp->b_freelist.tqe_next) n++; + lck_mtx_unlock(buf_mtxp); + return (n); } /* * Return a count of 'busy' buffers. Used at the time of shutdown. + * note: This is also called from the mach side in debug context in kdp.c */ int -count_busy_buffers() +count_busy_buffers(void) { - register struct buf *bp; - register int nbusy = 0; - - for (bp = &buf[nbuf]; --bp >= buf; ) - if ((bp->b_flags & (B_BUSY|B_INVAL)) == B_BUSY) - nbusy++; - return (nbusy); + return buf_busycount + bufstats.bufs_iobufinuse; } #if DIAGNOSTIC @@ -1876,9 +4151,9 @@ count_busy_buffers() void vfs_bufstats() { - int s, i, j, count; - register struct buf *bp; - register struct bqueues *dp; + int i, j, count; + struct buf *bp; + struct bqueues *dp; int counts[MAXBSIZE/CLBYTES+1]; static char *bname[BQUEUES] = { "LOCKED", "LRU", "AGE", "EMPTY", "META", "LAUNDRY" }; @@ -1887,12 +4162,15 @@ vfs_bufstats() count = 0; for (j = 0; j <= MAXBSIZE/CLBYTES; j++) counts[j] = 0; - s = splbio(); + + lck_mtx_lock(buf_mtxp); + for (bp = dp->tqh_first; bp; bp = bp->b_freelist.tqe_next) { counts[bp->b_bufsize/CLBYTES]++; count++; } - splx(s); + lck_mtx_unlock(buf_mtxp); + printf("%s: total-%d", bname[i], count); for (j = 0; j <= MAXBSIZE/CLBYTES; j++) if (counts[j] != 0) @@ -1902,445 +4180,539 @@ vfs_bufstats() } #endif /* DIAGNOSTIC */ -#define NRESERVEDIOBUFS 64 +#define NRESERVEDIOBUFS 128 -__private_extern__ struct buf * -alloc_io_buf(vp, priv) - struct vnode *vp; - int priv; +#define MNT_VIRTUALDEV_MAX_IOBUFS 16 +#define VIRTUALDEV_MAX_IOBUFS ((40*niobuf_headers)/100) + +buf_t +alloc_io_buf(vnode_t vp, int priv) { - register struct buf *bp; - int s; + buf_t bp; + mount_t mp = NULL; + int alloc_for_virtualdev = FALSE; - s = splbio(); + lck_mtx_lock_spin(iobuffer_mtxp); - while (niobuf - NRESERVEDIOBUFS < bufstats.bufs_iobufinuse && !priv) { - need_iobuffer = 1; - bufstats.bufs_iobufsleeps++; - (void) tsleep(&need_iobuffer, (PRIBIO+1), "alloc_io_buf", 0); + /* + * We subject iobuf requests for diskimages to additional restrictions. + * + * a) A single diskimage mount cannot use up more than + * MNT_VIRTUALDEV_MAX_IOBUFS. However,vm privileged (pageout) requests + * are not subject to this restriction. + * b) iobuf headers used by all diskimage headers by all mount + * points cannot exceed VIRTUALDEV_MAX_IOBUFS. + */ + if (vp && ((mp = vp->v_mount)) && mp != dead_mountp && + mp->mnt_kern_flag & MNTK_VIRTUALDEV) { + alloc_for_virtualdev = TRUE; + while ((!priv && mp->mnt_iobufinuse > MNT_VIRTUALDEV_MAX_IOBUFS) || + bufstats.bufs_iobufinuse_vdev > VIRTUALDEV_MAX_IOBUFS) { + bufstats.bufs_iobufsleeps++; + + need_iobuffer = 1; + (void)msleep(&need_iobuffer, iobuffer_mtxp, + PSPIN | (PRIBIO+1), (const char *)"alloc_io_buf (1)", + NULL); + } } - while ((bp = iobufqueue.tqh_first) == NULL) { - need_iobuffer = 1; + while (((niobuf_headers - NRESERVEDIOBUFS < bufstats.bufs_iobufinuse) && !priv) || + (bp = iobufqueue.tqh_first) == NULL) { bufstats.bufs_iobufsleeps++; - (void) tsleep(&need_iobuffer, (PRIBIO+1), "alloc_io_buf1", 0); - } + need_iobuffer = 1; + (void)msleep(&need_iobuffer, iobuffer_mtxp, PSPIN | (PRIBIO+1), + (const char *)"alloc_io_buf (2)", NULL); + } TAILQ_REMOVE(&iobufqueue, bp, b_freelist); - bp->b_timestamp = 0; - /* clear out various fields */ - bp->b_flags = B_BUSY; + bufstats.bufs_iobufinuse++; + if (bufstats.bufs_iobufinuse > bufstats.bufs_iobufmax) + bufstats.bufs_iobufmax = bufstats.bufs_iobufinuse; + + if (alloc_for_virtualdev) { + mp->mnt_iobufinuse++; + bufstats.bufs_iobufinuse_vdev++; + } + + lck_mtx_unlock(iobuffer_mtxp); + + /* + * initialize various fields + * we don't need to hold the mutex since the buffer + * is now private... the vp should have a reference + * on it and is not protected by this mutex in any event + */ + bp->b_timestamp = 0; + bp->b_proc = NULL; + + bp->b_datap = 0; + bp->b_flags = 0; + bp->b_lflags = BL_BUSY | BL_IOBUF; + if (alloc_for_virtualdev) + bp->b_lflags |= BL_IOBUF_VDEV; + bp->b_redundancy_flags = 0; bp->b_blkno = bp->b_lblkno = 0; - bp->b_iodone = 0; +#ifdef JOE_DEBUG + bp->b_owner = current_thread(); + bp->b_tag = 6; +#endif + bp->b_iodone = NULL; bp->b_error = 0; bp->b_resid = 0; bp->b_bcount = 0; bp->b_bufsize = 0; + bp->b_upl = NULL; + bp->b_fsprivate = (void *)NULL; bp->b_vp = vp; + bzero(&bp->b_attr, sizeof(struct bufattr)); - if (vp->v_type == VBLK || vp->v_type == VCHR) + if (vp && (vp->v_type == VBLK || vp->v_type == VCHR)) bp->b_dev = vp->v_rdev; else bp->b_dev = NODEV; - bufstats.bufs_iobufinuse++; - if (bufstats.bufs_iobufinuse > bufstats.bufs_iobufmax) - bufstats.bufs_iobufmax = bufstats.bufs_iobufinuse; - splx(s); return (bp); } -__private_extern__ void -free_io_buf(bp) - struct buf *bp; + +void +free_io_buf(buf_t bp) { - int s; + int need_wakeup = 0; + int free_for_virtualdev = FALSE; + mount_t mp = NULL; + + /* Was this iobuf for a diskimage ? */ + if (bp->b_lflags & BL_IOBUF_VDEV) { + free_for_virtualdev = TRUE; + if (bp->b_vp) + mp = bp->b_vp->v_mount; + } - s = splbio(); - /* put buffer back on the head of the iobufqueue */ + /* + * put buffer back on the head of the iobufqueue + */ bp->b_vp = NULL; bp->b_flags = B_INVAL; + /* Zero out the bufattr and its flags before relinquishing this iobuf */ + bzero (&bp->b_attr, sizeof(struct bufattr)); + + lck_mtx_lock_spin(iobuffer_mtxp); + binsheadfree(bp, &iobufqueue, -1); - /* Wake up any processes waiting for any buffer to become free. */ if (need_iobuffer) { + /* + * Wake up any processes waiting because they need an io buffer + * + * do the wakeup after we drop the mutex... it's possible that the + * wakeup will be superfluous if need_iobuffer gets set again and + * another thread runs this path, but it's highly unlikely, doesn't + * hurt, and it means we don't hold up I/O progress if the wakeup blocks + * trying to grab a task related lock... + */ need_iobuffer = 0; - wakeup(&need_iobuffer); + need_wakeup = 1; } - bufstats.bufs_iobufinuse--; - splx(s); -} + if (bufstats.bufs_iobufinuse <= 0) + panic("free_io_buf: bp(%p) - bufstats.bufs_iobufinuse < 0", bp); -/* disabled for now */ + bufstats.bufs_iobufinuse--; -/* XXX move this to a separate file */ -/* - * Dynamic Scaling of the Buffer Queues - */ + if (free_for_virtualdev) { + bufstats.bufs_iobufinuse_vdev--; + if (mp && mp != dead_mountp) + mp->mnt_iobufinuse--; + } -typedef long long blsize_t; + lck_mtx_unlock(iobuffer_mtxp); -blsize_t MAXNBUF; /* initialize to (mem_size / PAGE_SIZE) */ -/* Global tunable limits */ -blsize_t nbufh; /* number of buffer headers */ -blsize_t nbuflow; /* minimum number of buffer headers required */ -blsize_t nbufhigh; /* maximum number of buffer headers allowed */ -blsize_t nbuftarget; /* preferred number of buffer headers */ + if (need_wakeup) + wakeup(&need_iobuffer); +} -/* - * assertions: - * - * 1. 0 < nbuflow <= nbufh <= nbufhigh - * 2. nbufhigh <= MAXNBUF - * 3. 0 < nbuflow <= nbuftarget <= nbufhigh - * 4. nbufh can not be set by sysctl(). - */ -/* Per queue tunable limits */ +void +buf_list_lock(void) +{ + lck_mtx_lock_spin(buf_mtxp); +} -struct bufqlim { - blsize_t bl_nlow; /* minimum number of buffer headers required */ - blsize_t bl_num; /* number of buffer headers on the queue */ - blsize_t bl_nlhigh; /* maximum number of buffer headers allowed */ - blsize_t bl_target; /* preferred number of buffer headers */ - long bl_stale; /* Seconds after which a buffer is considered stale */ -} bufqlim[BQUEUES]; +void +buf_list_unlock(void) +{ + lck_mtx_unlock(buf_mtxp); +} /* - * assertions: - * - * 1. 0 <= bl_nlow <= bl_num <= bl_nlhigh - * 2. bl_nlhigh <= MAXNBUF - * 3. bufqlim[BQ_META].bl_nlow != 0 - * 4. bufqlim[BQ_META].bl_nlow > (number of possible concurrent - * file system IO operations) - * 5. bl_num can not be set by sysctl(). - * 6. bl_nhigh <= nbufhigh + * If getnewbuf() calls bcleanbuf() on the same thread + * there is a potential for stack overrun and deadlocks. + * So we always handoff the work to a worker thread for completion */ -/* - * Rationale: - * ---------- - * Defining it blsize_t as long permits 2^31 buffer headers per queue. - * Which can describe (2^31 * PAGE_SIZE) memory per queue. - * - * These limits are exported to by means of sysctl(). - * It was decided to define blsize_t as a 64 bit quantity. - * This will make sure that we will not be required to change it - * as long as we do not exceed 64 bit address space for the kernel. - * - * low and high numbers parameters initialized at compile time - * and boot arguments can be used to override them. sysctl() - * would not change the value. sysctl() can get all the values - * but can set only target. num is the current level. - * - * Advantages of having a "bufqscan" thread doing the balancing are, - * Keep enough bufs on BQ_EMPTY. - * getnewbuf() by default will always select a buffer from the BQ_EMPTY. - * getnewbuf() perfoms best if a buffer was found there. - * Also this minimizes the possibility of starting IO - * from getnewbuf(). That's a performance win, too. - * - * Localize complex logic [balancing as well as time aging] - * to balancebufq(). - * - * Simplify getnewbuf() logic by elimination of time aging code. - */ -/* - * Algorithm: - * ----------- - * The goal of the dynamic scaling of the buffer queues to to keep - * the size of the LRU close to bl_target. Buffers on a queue would - * be time aged. - * - * There would be a thread which will be responsible for "balancing" - * the buffer cache queues. - * - * The scan order would be: AGE, LRU, META, EMPTY. - */ +static void +bcleanbuf_thread_init(void) +{ + thread_t thread = THREAD_NULL; -long bufqscanwait = 0; + /* create worker thread */ + kernel_thread_start((thread_continue_t)bcleanbuf_thread, NULL, &thread); + thread_deallocate(thread); +} -static void bufqscan_thread(); -static int balancebufq(int q); -static int btrimempty(int n); -static __inline__ int initbufqscan(void); -static __inline__ int nextbufq(int q); -static void buqlimprt(int all); +typedef int (*bcleanbufcontinuation)(int); +__attribute__((noreturn)) static void -bufq_balance_thread_init() +bcleanbuf_thread(void) { + struct buf *bp; + int error = 0; + int loopcnt = 0; - if (bufqscanwait++ == 0) { + for (;;) { + lck_mtx_lock_spin(buf_mtxp); - /* Initalize globals */ - MAXNBUF = (mem_size / PAGE_SIZE); - nbufh = nbuf; - nbuflow = min(nbufh, 100); - nbufhigh = min(MAXNBUF, max(nbufh, 2048)); - nbuftarget = (mem_size >> 5) / PAGE_SIZE; - nbuftarget = max(nbuflow, nbuftarget); - nbuftarget = min(nbufhigh, nbuftarget); + while ( (bp = TAILQ_FIRST(&bufqueues[BQ_LAUNDRY])) == NULL) { + (void)msleep0(&bufqueues[BQ_LAUNDRY], buf_mtxp, PRIBIO|PDROP, "blaundry", 0, (bcleanbufcontinuation)bcleanbuf_thread); + } + + /* + * Remove from the queue + */ + bremfree_locked(bp); /* - * Initialize the bufqlim - */ - - /* LOCKED queue */ - bufqlim[BQ_LOCKED].bl_nlow = 0; - bufqlim[BQ_LOCKED].bl_nlhigh = 32; - bufqlim[BQ_LOCKED].bl_target = 0; - bufqlim[BQ_LOCKED].bl_stale = 30; - - /* LRU queue */ - bufqlim[BQ_LRU].bl_nlow = 0; - bufqlim[BQ_LRU].bl_nlhigh = nbufhigh/4; - bufqlim[BQ_LRU].bl_target = nbuftarget/4; - bufqlim[BQ_LRU].bl_stale = LRU_IS_STALE; - - /* AGE queue */ - bufqlim[BQ_AGE].bl_nlow = 0; - bufqlim[BQ_AGE].bl_nlhigh = nbufhigh/4; - bufqlim[BQ_AGE].bl_target = nbuftarget/4; - bufqlim[BQ_AGE].bl_stale = AGE_IS_STALE; - - /* EMPTY queue */ - bufqlim[BQ_EMPTY].bl_nlow = 0; - bufqlim[BQ_EMPTY].bl_nlhigh = nbufhigh/4; - bufqlim[BQ_EMPTY].bl_target = nbuftarget/4; - bufqlim[BQ_EMPTY].bl_stale = 600000; - - /* META queue */ - bufqlim[BQ_META].bl_nlow = 0; - bufqlim[BQ_META].bl_nlhigh = nbufhigh/4; - bufqlim[BQ_META].bl_target = nbuftarget/4; - bufqlim[BQ_META].bl_stale = META_IS_STALE; - - /* LAUNDRY queue */ - bufqlim[BQ_LOCKED].bl_nlow = 0; - bufqlim[BQ_LOCKED].bl_nlhigh = 32; - bufqlim[BQ_LOCKED].bl_target = 0; - bufqlim[BQ_LOCKED].bl_stale = 30; - - buqlimprt(1); - } + * Buffer is no longer on any free list + */ + SET(bp->b_lflags, BL_BUSY); + buf_busycount++; - /* create worker thread */ - kernel_thread(kernel_task, bufqscan_thread); -} +#ifdef JOE_DEBUG + bp->b_owner = current_thread(); + bp->b_tag = 10; +#endif -/* The workloop for the buffer balancing thread */ -static void -bufqscan_thread() -{ - boolean_t funnel_state; - int moretodo = 0; + lck_mtx_unlock(buf_mtxp); + /* + * do the IO + */ + error = bawrite_internal(bp, 0); - funnel_state = thread_funnel_set(kernel_flock, TRUE); + if (error) { + bp->b_whichq = BQ_LAUNDRY; + bp->b_timestamp = buf_timestamp(); - for(;;) { - do { - int q; /* buffer queue to process */ - - q = initbufqscan(); - for (; q; ) { - moretodo |= balancebufq(q); - q = nextbufq(q); - } - } while (moretodo); + lck_mtx_lock_spin(buf_mtxp); -#if DIAGNOSTIC - vfs_bufstats(); - buqlimprt(0); -#endif - (void)tsleep((void *)&bufqscanwait, PRIBIO, "bufqscanwait", 60 * hz); - moretodo = 0; - } + binstailfree(bp, &bufqueues[BQ_LAUNDRY], BQ_LAUNDRY); + blaundrycnt++; - (void) thread_funnel_set(kernel_flock, FALSE); -} + /* we never leave a busy page on the laundry queue */ + CLR(bp->b_lflags, BL_BUSY); + buf_busycount--; +#ifdef JOE_DEBUG + bp->b_owner = current_thread(); + bp->b_tag = 11; +#endif -/* Seed for the buffer queue balancing */ -static __inline__ int -initbufqscan() -{ - /* Start with AGE queue */ - return (BQ_AGE); + lck_mtx_unlock(buf_mtxp); + + if (loopcnt > MAXLAUNDRY) { + /* + * bawrite_internal() can return errors if we're throttled. If we've + * done several I/Os and failed, give the system some time to unthrottle + * the vnode + */ + (void)tsleep((void *)&bufqueues[BQ_LAUNDRY], PRIBIO, "blaundry", 1); + loopcnt = 0; + } else { + /* give other threads a chance to run */ + (void)thread_block(THREAD_CONTINUE_NULL); + loopcnt++; + } + } + } } -/* Pick next buffer queue to balance */ -static __inline__ int -nextbufq(int q) -{ - int order[] = { BQ_AGE, BQ_LRU, BQ_META, BQ_EMPTY, 0 }; - - q++; - q %= sizeof(order); - return (order[q]); -} -/* function to balance the buffer queues */ static int -balancebufq(int q) +brecover_data(buf_t bp) { - int moretodo = 0; - int s = splbio(); - int n; - - /* reject invalid q */ - if ((q < 0) || (q >= BQUEUES)) - goto out; - - /* LOCKED or LAUNDRY queue MUST not be balanced */ - if ((q == BQ_LOCKED) || (q == BQ_LAUNDRY)) - goto out; - - n = (bufqlim[q].bl_num - bufqlim[q].bl_target); + int upl_offset; + upl_t upl; + upl_page_info_t *pl; + kern_return_t kret; + vnode_t vp = bp->b_vp; + int upl_flags; - /* If queue has less than target nothing more to do */ - if (n < 0) - goto out; - if ( n > 8 ) { - /* Balance only a small amount (12.5%) at a time */ - n >>= 3; - } + if ( !UBCINFOEXISTS(vp) || bp->b_bufsize == 0) + goto dump_buffer; - /* EMPTY queue needs special handling */ - if (q == BQ_EMPTY) { - moretodo |= btrimempty(n); - goto out; + upl_flags = UPL_PRECIOUS; + if (! (buf_flags(bp) & B_READ)) { + /* + * "write" operation: let the UPL subsystem know + * that we intend to modify the buffer cache pages we're + * gathering. + */ + upl_flags |= UPL_WILL_MODIFY; } - - for (; n > 0; n--) { - struct buf *bp = bufqueues[q].tqh_first; - if (!bp) - break; - /* check if it's stale */ - if ((time.tv_sec - bp->b_timestamp) > bufqlim[q].bl_stale) { - if (bcleanbuf(bp)) { - /* bawrite() issued, bp not ready */ - moretodo = 1; - } else { - /* release the cleaned buffer to BQ_EMPTY */ - SET(bp->b_flags, B_INVAL); - brelse(bp); - } - } else - break; + kret = ubc_create_upl(vp, + ubc_blktooff(vp, bp->b_lblkno), + bp->b_bufsize, + &upl, + &pl, + upl_flags); + if (kret != KERN_SUCCESS) + panic("Failed to create UPL"); + + for (upl_offset = 0; upl_offset < bp->b_bufsize; upl_offset += PAGE_SIZE) { + + if (!upl_valid_page(pl, upl_offset / PAGE_SIZE) || !upl_dirty_page(pl, upl_offset / PAGE_SIZE)) { + ubc_upl_abort(upl, 0); + goto dump_buffer; + } } + bp->b_upl = upl; + + kret = ubc_upl_map(upl, (vm_offset_t *)&(bp->b_datap)); + + if (kret != KERN_SUCCESS) + panic("getblk: ubc_upl_map() failed with (%d)", kret); + return (1); + +dump_buffer: + bp->b_bufsize = 0; + SET(bp->b_flags, B_INVAL); + buf_brelse(bp); -out: - splx(s); - return (moretodo); + return(0); } -static int -btrimempty(int n) +boolean_t +buffer_cache_gc(int all) { - /* - * When struct buf are allocated dynamically, this would - * reclaim upto 'n' struct buf from the empty queue. + buf_t bp; + boolean_t did_large_zfree = FALSE; + boolean_t need_wakeup = FALSE; + int now = buf_timestamp(); + uint32_t found = 0; + struct bqueues privq; + int thresh_hold = BUF_STALE_THRESHHOLD; + + if (all) + thresh_hold = 0; + /* + * We only care about metadata (incore storage comes from zalloc()). + * Unless "all" is set (used to evict meta data buffers in preparation + * for deep sleep), we only evict up to BUF_MAX_GC_BATCH_SIZE buffers + * that have not been accessed in the last BUF_STALE_THRESHOLD seconds. + * BUF_MAX_GC_BATCH_SIZE controls both the hold time of the global lock + * "buf_mtxp" and the length of time we spend compute bound in the GC + * thread which calls this function */ - - return (0); -} + lck_mtx_lock(buf_mtxp); -static __inline__ void -bufqinc(int q) -{ - if ((q < 0) || (q >= BQUEUES)) - return; + do { + found = 0; + TAILQ_INIT(&privq); + need_wakeup = FALSE; - bufqlim[q].bl_num++; - return; -} + while (((bp = TAILQ_FIRST(&bufqueues[BQ_META]))) && + (now > bp->b_timestamp) && + (now - bp->b_timestamp > thresh_hold) && + (found < BUF_MAX_GC_BATCH_SIZE)) { -static __inline__ void -bufqdec(int q) -{ - if ((q < 0) || (q >= BQUEUES)) - return; + /* Remove from free list */ + bremfree_locked(bp); + found++; - bufqlim[q].bl_num--; - return; -} +#ifdef JOE_DEBUG + bp->b_owner = current_thread(); + bp->b_tag = 12; +#endif -static void -buqlimprt(int all) -{ - int i; - static char *bname[BQUEUES] = - { "LOCKED", "LRU", "AGE", "EMPTY", "META", "LAUNDRY" }; + /* If dirty, move to laundry queue and remember to do wakeup */ + if (ISSET(bp->b_flags, B_DELWRI)) { + SET(bp->b_lflags, BL_WANTDEALLOC); - if (all) - for (i = 0; i < BQUEUES; i++) { - printf("%s : ", bname[i]); - printf("min = %ld, ", (long)bufqlim[i].bl_nlow); - printf("cur = %ld, ", (long)bufqlim[i].bl_num); - printf("max = %ld, ", (long)bufqlim[i].bl_nlhigh); - printf("target = %ld, ", (long)bufqlim[i].bl_target); - printf("stale after %ld seconds\n", bufqlim[i].bl_stale); + bmovelaundry(bp); + need_wakeup = TRUE; + + continue; + } + + /* + * Mark busy and put on private list. We could technically get + * away without setting BL_BUSY here. + */ + SET(bp->b_lflags, BL_BUSY); + buf_busycount++; + + /* + * Remove from hash and dissociate from vp. + */ + bremhash(bp); + if (bp->b_vp) { + brelvp_locked(bp); + } + + TAILQ_INSERT_TAIL(&privq, bp, b_freelist); } - else - for (i = 0; i < BQUEUES; i++) { - printf("%s : ", bname[i]); - printf("cur = %ld, ", (long)bufqlim[i].bl_num); + + if (found == 0) { + break; + } + + /* Drop lock for batch processing */ + lck_mtx_unlock(buf_mtxp); + + /* Wakeup and yield for laundry if need be */ + if (need_wakeup) { + wakeup(&bufqueues[BQ_LAUNDRY]); + (void)thread_block(THREAD_CONTINUE_NULL); } + + /* Clean up every buffer on private list */ + TAILQ_FOREACH(bp, &privq, b_freelist) { + /* Take note if we've definitely freed at least a page to a zone */ + if ((ISSET(bp->b_flags, B_ZALLOC)) && (buf_size(bp) >= PAGE_SIZE)) { + did_large_zfree = TRUE; + } + + trace(TR_BRELSE, pack(bp->b_vp, bp->b_bufsize), bp->b_lblkno); + + /* Free Storage */ + buf_free_meta_store(bp); + + /* Release credentials */ + buf_release_credentials(bp); + + /* Prepare for moving to empty queue */ + CLR(bp->b_flags, (B_META | B_ZALLOC | B_DELWRI | B_LOCKED + | B_AGE | B_ASYNC | B_NOCACHE | B_FUA)); + bp->b_whichq = BQ_EMPTY; + BLISTNONE(bp); + } + lck_mtx_lock(buf_mtxp); + + /* Back under lock, move them all to invalid hash and clear busy */ + TAILQ_FOREACH(bp, &privq, b_freelist) { + binshash(bp, &invalhash); + CLR(bp->b_lflags, BL_BUSY); + buf_busycount--; + +#ifdef JOE_DEBUG + if (bp->b_owner != current_thread()) { + panic("Buffer stolen from buffer_cache_gc()"); + } + bp->b_owner = current_thread(); + bp->b_tag = 13; +#endif + } + + /* And do a big bulk move to the empty queue */ + TAILQ_CONCAT(&bufqueues[BQ_EMPTY], &privq, b_freelist); + + } while (all && (found == BUF_MAX_GC_BATCH_SIZE)); + + lck_mtx_unlock(buf_mtxp); + + return did_large_zfree; } + /* - * If the getnewbuf() calls bcleanbuf() on the same thread - * there is a potential for stack overrun and deadlocks. - * So we always handoff the work to worker thread for completion + * disabled for now */ -static void -bcleanbuf_thread_init() +#if FLUSH_QUEUES + +#define NFLUSH 32 + +static int +bp_cmp(void *a, void *b) { - static void bcleanbuf_thread(); + buf_t *bp_a = *(buf_t **)a, + *bp_b = *(buf_t **)b; + daddr64_t res; - /* create worker thread */ - kernel_thread(kernel_task, bcleanbuf_thread); + // don't have to worry about negative block + // numbers so this is ok to do. + // + res = (bp_a->b_blkno - bp_b->b_blkno); + + return (int)res; } -static void -bcleanbuf_thread() + +int +bflushq(int whichq, mount_t mp) { - boolean_t funnel_state; - struct buf *bp; - int error = 0; - int loopcnt = 0; + buf_t bp, next; + int i, buf_count; + int total_writes = 0; + static buf_t flush_table[NFLUSH]; + + if (whichq < 0 || whichq >= BQUEUES) { + return (0); + } - funnel_state = thread_funnel_set(kernel_flock, TRUE); + restart: + lck_mtx_lock(buf_mtxp); -doit: - while (blaundrycnt == 0) - (void)tsleep((void *)&blaundrycnt, PRIBIO, "blaundry", 60 * hz); - bp = TAILQ_FIRST(&bufqueues[BQ_LAUNDRY]); - /* Remove from the queue */ - bremfree(bp); - blaundrycnt--; - /* do the IO */ - error = bawrite_internal(bp, 0); - if (error) { - binstailfree(bp, &bufqueues[BQ_LAUNDRY], BQ_LAUNDRY); - blaundrycnt++; - if (loopcnt > 10) { - (void)tsleep((void *)&blaundrycnt, PRIBIO, "blaundry", 1); - loopcnt = 0; - } else { - (void)thread_block(THREAD_CONTINUE_NULL); - loopcnt++; + bp = TAILQ_FIRST(&bufqueues[whichq]); + + for (buf_count = 0; bp; bp = next) { + next = bp->b_freelist.tqe_next; + + if (bp->b_vp == NULL || bp->b_vp->v_mount != mp) { + continue; + } + + if (ISSET(bp->b_flags, B_DELWRI) && !ISSET(bp->b_lflags, BL_BUSY)) { + + bremfree_locked(bp); +#ifdef JOE_DEBUG + bp->b_owner = current_thread(); + bp->b_tag = 7; +#endif + SET(bp->b_lflags, BL_BUSY); + buf_busycount++; + + flush_table[buf_count] = bp; + buf_count++; + total_writes++; + + if (buf_count >= NFLUSH) { + lck_mtx_unlock(buf_mtxp); + + qsort(flush_table, buf_count, sizeof(struct buf *), bp_cmp); + + for (i = 0; i < buf_count; i++) { + buf_bawrite(flush_table[i]); + } + goto restart; } + } } - /* start again */ - goto doit; + lck_mtx_unlock(buf_mtxp); - (void) thread_funnel_set(kernel_flock, funnel_state); + if (buf_count > 0) { + qsort(flush_table, buf_count, sizeof(struct buf *), bp_cmp); + + for (i = 0; i < buf_count; i++) { + buf_bawrite(flush_table[i]); + } + } + + return (total_writes); } +#endif