/*
- * Copyright (c) 2000-2007 Apple Inc. All rights reserved.
+ * Copyright (c) 2000-2010 Apple Inc. All rights reserved.
*
* @APPLE_OSREFERENCE_LICENSE_HEADER_START@
*
#include <kern/sched_prim.h> /* thread_block() */
#include <vm/vm_kern.h>
+#include <vm/vm_pageout.h>
#include <sys/kdebug.h>
#include <libkern/OSAtomic.h>
+#include <libkern/OSDebug.h>
#include <sys/ubc_internal.h>
#include <sys/sdt.h>
+
#if BALANCE_QUEUES
static __inline__ void bufqinc(int q);
static __inline__ void bufqdec(int q);
#endif
-static int bcleanbuf(buf_t bp);
+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 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);
+
__private_extern__ int bdwrite_internal(buf_t, int);
int blaundrycnt = 0;
static int boot_nbuf_headers = 0;
+static TAILQ_HEAD(delayqueue, buf) delaybufqueue;
static TAILQ_HEAD(ioqueue, buf) iobufqueue;
static TAILQ_HEAD(bqueues, buf) bufqueues[BQUEUES];
static lck_mtx_t *iobuffer_mtxp;
static lck_mtx_t *buf_mtxp;
+static int buf_busycount;
+
static __inline__ int
buf_timestamp(void)
{
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
*bp->b_hash.le_prev = (bp)->b_hash.le_next;
}
+/*
+ * buf_mtxp held.
+ */
+static __inline__ void
+bmovelaundry(buf_t bp)
+{
+ bp->b_whichq = BQ_LAUNDRY;
+ bp->b_timestamp = buf_timestamp();
+ binstailfree(bp, &bufqueues[BQ_LAUNDRY], BQ_LAUNDRY);
+ blaundrycnt++;
+}
+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);
+ }
+}
int
if (!ISSET(bp->b_flags, B_DELWRI)) {
SET(bp->b_flags, B_DELWRI);
- OSAddAtomic(1, &nbdwrite);
+ OSAddAtomicLong(1, &nbdwrite);
buf_reassign(bp, bp->b_vp);
}
SET(bp->b_flags, B_DONE);
}
+void
+buf_markclean(buf_t bp) {
+
+ if (ISSET(bp->b_flags, B_DELWRI)) {
+ CLR(bp->b_flags, B_DELWRI);
+
+ OSAddAtomicLong(-1, &nbdwrite);
+ buf_reassign(bp, bp->b_vp);
+ }
+}
+
void
buf_markeintr(buf_t bp) {
SET(bp->b_flags, B_FUA);
}
+#ifdef CONFIG_PROTECT
+void *
+buf_getcpaddr(buf_t bp) {
+ return bp->b_cpentry;
+}
+
+void
+buf_setcpaddr(buf_t bp, void *cp_entry_addr) {
+ bp->b_cpentry = (struct cprotect *) cp_entry_addr;
+}
+
+#else
+void *
+buf_getcpaddr(buf_t bp __unused) {
+ return NULL;
+}
+
+void
+buf_setcpaddr(buf_t bp __unused, void *cp_entry_addr __unused) {
+ return;
+}
+#endif /* CONFIG_PROTECT */
+
errno_t
buf_error(buf_t 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, void **old_transaction)
+ void (**old_iodone)(buf_t, void *), void **old_transaction)
{
- if (old_iodone)
- *old_iodone = (void *)(bp->b_iodone);
+ if (old_iodone)
+ *old_iodone = bp->b_iodone;
if (old_transaction)
- *old_transaction = (void *)(bp->b_transaction);
+ *old_transaction = bp->b_transaction;
bp->b_transaction = transaction;
bp->b_iodone = filter;
bp->b_fsprivate = fsprivate;
}
-ucred_t
+kauth_cred_t
buf_rcred(buf_t bp) {
return (bp->b_rcred);
}
-ucred_t
+kauth_cred_t
buf_wcred(buf_t bp) {
return (bp->b_wcred);
buf_map(buf_t bp, caddr_t *io_addr)
{
buf_t real_bp;
- vm_offset_t vaddr;
+ vm_offset_t vaddr;
kern_return_t kret;
if ( !(bp->b_flags & B_CLUSTER)) {
bp->b_resid = 0;
}
-
-
/*
* Read or write a buffer that is not contiguous on disk.
* buffer is marked done/error at the conclusion
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 phisical
+ * 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");
* end up issuing the I/O...
*/
bp->b_dev = devvp->v_rdev;
- DTRACE_IO1(start, buf_t, bp);
if (bp->b_flags & B_READ)
bmap_flags = VNODE_READ;
* 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) {
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);
}
- if (bp->b_blkno == -1)
+
+ 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)
+ }
+ 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 */
+
/*
* we can issue the I/O because...
* either B_CLUSTER is set which
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 (;;) {
goto try_dirty_list;
}
while (!LIST_EMPTY(&local_iterblkhd)) {
+
bp = LIST_FIRST(&local_iterblkhd);
LIST_REMOVE(bp, b_vnbufs);
if ((flags & BUF_SKIP_META) && (bp->b_lblkno < 0 || ISSET(bp->b_flags, B_META)))
continue;
- if ( (error = (int)buf_acquire_locked(bp, BAC_REMOVE | BAC_SKIP_LOCKED, slpflag, slptimeo)) ) {
+ 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...
}
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);
if ((flags & BUF_SKIP_META) && (bp->b_lblkno < 0 || ISSET(bp->b_flags, B_META)))
continue;
- if ( (error = (int)buf_acquire_locked(bp, BAC_REMOVE | BAC_SKIP_LOCKED, slpflag, slptimeo)) ) {
+ 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...
}
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))
{
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
*
* NB: This makes an assumption about how tailq's are implemented.
*/
- whichq = bp->b_whichq;
-
if (bp->b_freelist.tqe_next == NULL) {
dp = &bufqueues[whichq];
bp->b_whichq = -1;
bp->b_timestamp = 0;
+ bp->b_shadow = 0;
}
/*
static void
buf_reassign(buf_t bp, vnode_t newvp)
{
- register struct buflists *listheadp;
+ struct buflists *listheadp;
if (newvp == NULL) {
printf("buf_reassign: NULL");
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++;
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);
/* 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");
+ }
+
#if BALANCE_QUEUES
{
static void bufq_balance_thread_init(void) __attribute__((section("__TEXT, initcode")));
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
static struct buf *
-bio_doread(vnode_t vp, daddr64_t blkno, int size, ucred_t cred, int async, int queuetype)
+bio_doread(vnode_t vp, daddr64_t blkno, int size, kauth_cred_t cred, int async, int queuetype)
{
buf_t bp;
/* Pay for the read. */
if (p && p->p_stats)
- OSIncrementAtomic(&p->p_stats->p_ru.ru_inblock); /* XXX */
+ OSIncrementAtomicLong(&p->p_stats->p_ru.ru_inblock); /* XXX */
if (async) {
/*
*/
static errno_t
do_breadn_for_type(vnode_t vp, daddr64_t blkno, int size, daddr64_t *rablks, int *rasizes,
- int nrablks, ucred_t cred, buf_t *bpp, int queuetype)
+ int nrablks, kauth_cred_t cred, buf_t *bpp, int queuetype)
{
buf_t bp;
int i;
* This algorithm described in Bach (p.54).
*/
errno_t
-buf_bread(vnode_t vp, daddr64_t blkno, int size, ucred_t cred, buf_t *bpp)
+buf_bread(vnode_t vp, daddr64_t blkno, int size, kauth_cred_t cred, buf_t *bpp)
{
buf_t bp;
* This algorithm described in Bach (p.54).
*/
errno_t
-buf_meta_bread(vnode_t vp, daddr64_t blkno, int size, ucred_t cred, buf_t *bpp)
+buf_meta_bread(vnode_t vp, daddr64_t blkno, int size, kauth_cred_t cred, buf_t *bpp)
{
buf_t 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, ucred_t cred, buf_t *bpp)
+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));
}
* [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, ucred_t cred, buf_t *bpp)
+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));
}
CLR(bp->b_flags, (B_READ | B_DONE | B_ERROR | B_DELWRI));
if (wasdelayed)
- OSAddAtomic(-1, &nbdwrite);
+ OSAddAtomicLong(-1, &nbdwrite);
if (!sync) {
/*
buf_reassign(bp, vp);
else
if (p && p->p_stats)
- OSIncrementAtomic(&p->p_stats->p_ru.ru_oublock); /* XXX */
+ OSIncrementAtomicLong(&p->p_stats->p_ru.ru_oublock); /* XXX */
}
trace(TR_BUFWRITE, pack(vp, bp->b_bcount), bp->b_lblkno);
buf_reassign(bp, vp);
else
if (p && p->p_stats)
- OSIncrementAtomic(&p->p_stats->p_ru.ru_oublock); /* XXX */
+ OSIncrementAtomicLong(&p->p_stats->p_ru.ru_oublock); /* XXX */
/* Release the buffer. */
// XXXdbg - only if the unused bit is set
*
* 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
* buf_bdwrite()s can create situation where the kernel can create
* buffers faster than the disks can service. Doing a buf_bawrite() in
- * cases were we have "too many" outstanding buf_bdwrite()s avoids that.
+ * cases where we have "too many" outstanding buf_bdwrite()s avoids that.
*/
__private_extern__ int
bdwrite_internal(buf_t bp, int return_error)
if (!ISSET(bp->b_flags, B_DELWRI)) {
SET(bp->b_flags, B_DELWRI);
if (p && p->p_stats)
- OSIncrementAtomic(&p->p_stats->p_ru.ru_oublock); /* XXX */
- OSAddAtomic(1, &nbdwrite);
+ OSIncrementAtomicLong(&p->p_stats->p_ru.ru_oublock); /* XXX */
+ OSAddAtomicLong(1, &nbdwrite);
buf_reassign(bp, vp);
}
- /* If this is a tape block, write it the block now. */
- if (ISSET(bp->b_flags, B_TAPE)) {
- VNOP_BWRITE(bp);
- return (0);
- }
-
/*
* if we're not LOCKED, but the total number of delayed writes
* has climbed above 75% of the total buffers in the system
}
+
+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)
+{
+ 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
+ lck_mtx_lock_spin(buf_mtxp);
+
+ 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--;
+ /*
+ * 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
+ */
+ 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;
+ }
+ }
+ lck_mtx_unlock(buf_mtxp);
+#ifdef BUF_MAKE_PRIVATE
+ if (bp == bp_data && data_ref == 0)
+ buf_free_meta_store(bp);
+
+ 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).
panic("buf_brelse: bad buffer = %p\n", bp);
#ifdef JOE_DEBUG
- bp->b_stackbrelse[0] = (int)__builtin_return_address(0);
- bp->b_stackbrelse[1] = (int)__builtin_return_address(1);
- bp->b_stackbrelse[2] = (int)__builtin_return_address(2);
- bp->b_stackbrelse[3] = (int)__builtin_return_address(3);
- bp->b_stackbrelse[4] = (int)__builtin_return_address(4);
- bp->b_stackbrelse[5] = (int)__builtin_return_address(5);
+ (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_datap,
+ 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);
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 = (void *)bp->b_transaction;
+ void *arg = bp->b_transaction;
CLR(bp->b_flags, B_FILTER); /* but note callout done */
bp->b_iodone = NULL;
kern_return_t kret;
int upl_flags;
- if ( (upl == NULL) ) {
+ 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),
if (kret != KERN_SUCCESS)
panic("brelse: Failed to create UPL");
-#ifdef UPL_DEBUG
- upl_ubc_alias_set(upl, bp, 5);
+#if UPL_DEBUG
+ upl_ubc_alias_set(upl, (uintptr_t) bp, (uintptr_t) 5);
#endif /* UPL_DEBUG */
}
} else {
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)) {
+ 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 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 (ISSET(bp->b_flags, B_DELWRI))
- OSAddAtomic(-1, &nbdwrite);
+ OSAddAtomicLong(-1, &nbdwrite);
if (ISSET(bp->b_flags, B_META)) {
- 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;
- }
+ if (bp->b_shadow_ref)
+ delayed_buf_free_meta_store = TRUE;
+ else
+ buf_free_meta_store(bp);
}
/*
* nuke any credentials we were holding
*/
- 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);
+ buf_release_credentials(bp);
+
+ lck_mtx_lock_spin(buf_mtxp);
+
+ if (bp->b_shadow_ref) {
+ SET(bp->b_lflags, BL_WAITSHADOW);
+
+ lck_mtx_unlock(buf_mtxp);
+
+ return;
}
- CLR(bp->b_flags, (B_META | B_ZALLOC | B_DELWRI | B_LOCKED | B_AGE | B_ASYNC | B_NOCACHE | B_FUA));
+ if (delayed_buf_free_meta_store == TRUE) {
- bufq = &bufqueues[BQ_EMPTY];
- bp->b_whichq = BQ_EMPTY;
+ lck_mtx_unlock(buf_mtxp);
+finish_shadow_master:
+ buf_free_meta_store(bp);
- lck_mtx_lock_spin(buf_mtxp);
+ 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);
BLISTNONE(bp);
binshash(bp, &invalhash);
- binsheadfree(bp, bufq, BQ_EMPTY);
+ 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.
whichq = BQ_LRU; /* valid data */
bufq = &bufqueues[whichq];
- CLR(bp->b_flags, (B_AGE | B_ASYNC | B_NOCACHE));
- bp->b_whichq = whichq;
bp->b_timestamp = buf_timestamp();
- lck_mtx_lock_spin(buf_mtxp);
-
- binstailfree(bp, bufq, whichq);
+ 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) {
/*
* Unlock the buffer.
*/
CLR(bp->b_lflags, (BL_BUSY | BL_WANTED));
+ buf_busycount--;
lck_mtx_unlock(buf_mtxp);
wakeup(bp);
}
KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 388)) | DBG_FUNC_END,
- (int)bp, (int)bp->b_datap, bp->b_flags, 0, 0);
+ bp, bp->b_datap, bp->b_flags, 0, 0);
}
/*
struct bufhashhdr *dp;
KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 386)) | DBG_FUNC_START,
- (int)(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);
-start_locked:
+
if ((bp = incore_locked(vp, blkno, dp))) {
/*
* Found in the Buffer Cache
SET(bp->b_lflags, BL_WANTED);
bufstats.bufs_busyincore++;
- lck_mtx_convert_spin(buf_mtxp);
/*
* 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);
/*
*/
SET(bp->b_lflags, BL_BUSY);
SET(bp->b_flags, B_CACHE);
+ buf_busycount++;
bremfree_locked(bp);
bufstats.bufs_incore++;
} else
CLR(bp->b_flags, (B_DONE | B_CACHE | B_WASDIRTY | B_DELWRI));
- kret = ubc_upl_map(upl, (vm_address_t *)&(bp->b_datap));
+ kret = ubc_upl_map(upl, (vm_offset_t*)&(bp->b_datap));
if (kret != KERN_SUCCESS)
panic("getblk: ubc_upl_map() failed with (%d)", kret);
lck_mtx_unlock(buf_mtxp);
return (NULL);
}
- lck_mtx_convert_spin(buf_mtxp);
-
if ((vnode_isreg(vp) == 0) || (UBCINFOEXISTS(vp) == 0) /*|| (vnode_issystem(vp) == 1)*/)
operation = BLK_META;
if ((bp = getnewbuf(slpflag, slptimeo, &queue)) == NULL)
- goto start_locked;
+ goto start;
/*
* getnewbuf may block for a number of different reasons...
* in bufstats are protected with either
* buf_mtxp or iobuffer_mtxp
*/
- OSAddAtomic(1, &bufstats.bufs_miss);
+ OSAddAtomicLong(1, &bufstats.bufs_miss);
break;
case BLK_WRITE:
if (kret != KERN_SUCCESS)
panic("Failed to create UPL");
-#ifdef UPL_DEBUG
- upl_ubc_alias_set(upl, bp, 4);
+#if UPL_DEBUG
+ upl_ubc_alias_set(upl, (uintptr_t) bp, (uintptr_t) 4);
#endif /* UPL_DEBUG */
bp->b_upl = upl;
SET(bp->b_flags, B_CACHE | B_DONE);
- OSAddAtomic(1, &bufstats.bufs_vmhits);
+ OSAddAtomicLong(1, &bufstats.bufs_vmhits);
bp->b_validoff = 0;
bp->b_dirtyoff = 0;
if ((long)contig_bytes < bp->b_bcount)
bp->b_blkno = bp->b_lblkno;
} else {
- OSAddAtomic(1, &bufstats.bufs_miss);
+ OSAddAtomicLong(1, &bufstats.bufs_miss);
}
- kret = ubc_upl_map(upl, (vm_address_t *)&(bp->b_datap));
+ kret = ubc_upl_map(upl, (vm_offset_t *)&(bp->b_datap));
if (kret != KERN_SUCCESS)
panic("getblk: ubc_upl_map() failed with (%d)", kret);
}
}
KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 386)) | DBG_FUNC_END,
- (int)bp, (int)bp->b_datap, bp->b_flags, 3, 0);
+ bp, bp->b_datap, bp->b_flags, 3, 0);
#ifdef JOE_DEBUG
- bp->b_stackgetblk[0] = (int)__builtin_return_address(0);
- bp->b_stackgetblk[1] = (int)__builtin_return_address(1);
- bp->b_stackgetblk[2] = (int)__builtin_return_address(2);
- bp->b_stackgetblk[3] = (int)__builtin_return_address(3);
- bp->b_stackgetblk[4] = (int)__builtin_return_address(4);
- bp->b_stackgetblk[5] = (int)__builtin_return_address(5);
+ (void) OSBacktrace(&bp->b_stackgetblk[0], 6);
#endif
return (bp);
}
buf_t
buf_geteblk(int size)
{
- buf_t bp;
+ buf_t bp = NULL;
int queue = BQ_EMPTY;
- lck_mtx_lock(buf_mtxp);
+ 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
return (bp);
}
+uint32_t
+buf_redundancy_flags(buf_t bp)
+{
+ return bp->b_redundancy_flags;
+}
+
+void
+buf_set_redundancy_flags(buf_t bp, uint32_t flags)
+{
+ SET(bp->b_redundancy_flags, flags);
+}
+
+void
+buf_clear_redundancy_flags(buf_t bp, uint32_t flags)
+{
+ CLR(bp->b_redundancy_flags, flags);
+}
/*
* With UBC, there is no need to expand / shrink the file data
*(void **)(&bp->b_datap) = zalloc(z);
} else {
bp->b_datap = (uintptr_t)NULL;
- kmem_alloc_wired(kernel_map, (vm_offset_t *)&bp->b_datap, desired_size);
+ kmem_alloc_kobject(kernel_map, (vm_offset_t *)&bp->b_datap, desired_size);
CLR(bp->b_flags, B_ZALLOC);
}
bcopy((void *)elem, (caddr_t)bp->b_datap, bp->b_bufsize);
if ((vm_size_t)bp->b_bufsize < desired_size) {
/* reallocate to a bigger size */
bp->b_datap = (uintptr_t)NULL;
- kmem_alloc_wired(kernel_map, (vm_offset_t *)&bp->b_datap, desired_size);
+ kmem_alloc_kobject(kernel_map, (vm_offset_t *)&bp->b_datap, desired_size);
bcopy((const void *)elem, (caddr_t)bp->b_datap, bp->b_bufsize);
kmem_free(kernel_map, elem, bp->b_bufsize);
} else {
*(void **)(&bp->b_datap) = zalloc(z);
SET(bp->b_flags, B_ZALLOC);
} else
- kmem_alloc_wired(kernel_map, (vm_offset_t *)&bp->b_datap, desired_size);
+ kmem_alloc_kobject(kernel_map, (vm_offset_t *)&bp->b_datap, desired_size);
}
if (bp->b_datap == 0)
* on which it was found.
*
* buf_mtxp is held upon entry
- * returns with buf_mtxp locked
+ * returns with buf_mtxp locked if new buf available
+ * returns with buf_mtxp UNlocked if new buf NOT available
*/
static buf_t
SET(bp->b_flags, B_HDRALLOC);
*queue = BQ_EMPTY;
}
- lck_mtx_lock(buf_mtxp);
+ lck_mtx_lock_spin(buf_mtxp);
if (bp) {
binshash(bp, &invalhash);
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|(PRIBIO+1), "getnewbuf", &ts);
+
+ msleep(&needbuffer, buf_mtxp, slpflag | PDROP | (PRIBIO+1), "getnewbuf", &ts);
return (NULL);
}
}
found:
if (ISSET(bp->b_flags, B_LOCKED) || ISSET(bp->b_lflags, BL_BUSY))
- panic("getnewbuf: bp @ %p is LOCKED or BUSY! (flags 0x%lx)\n", bp, bp->b_flags);
+ panic("getnewbuf: bp @ %p is LOCKED or BUSY! (flags 0x%x)\n", bp, bp->b_flags);
/* Clean it */
- if (bcleanbuf(bp)) {
+ if (bcleanbuf(bp, FALSE)) {
/*
* moved to the laundry thread, buffer not ready
*/
/*
* Clean a buffer.
- * Returns 0 is buffer is ready to use,
+ * 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(buf_t bp)
+int
+bcleanbuf(buf_t bp, boolean_t discard)
{
/* Remove from the queue */
bremfree_locked(bp);
* it on the LAUNDRY queue, and return 1
*/
if (ISSET(bp->b_flags, B_DELWRI)) {
- bp->b_whichq = BQ_LAUNDRY;
- bp->b_timestamp = buf_timestamp();
- binstailfree(bp, &bufqueues[BQ_LAUNDRY], BQ_LAUNDRY);
- blaundrycnt++;
+ if (discard) {
+ SET(bp->b_lflags, BL_WANTDEALLOC);
+ }
+
+ bmovelaundry(bp);
lck_mtx_unlock(buf_mtxp);
*/
(void)thread_block(THREAD_CONTINUE_NULL);
- lck_mtx_lock(buf_mtxp);
+ lck_mtx_lock_spin(buf_mtxp);
return (1);
}
* Buffer is no longer on any free list... we own it
*/
SET(bp->b_lflags, BL_BUSY);
-
+ buf_busycount++;
+
bremhash(bp);
/*
BLISTNONE(bp);
- if (ISSET(bp->b_flags, B_META)) {
- vm_offset_t elem;
-
- elem = (vm_offset_t)bp->b_datap;
- bp->b_datap = (uintptr_t)0xdeadbeef;
-
- if (ISSET(bp->b_flags, B_ZALLOC)) {
- zone_t z;
-
- z = getbufzone(bp->b_bufsize);
- zfree(z, (void *)elem);
- } else
- kmem_free(kernel_map, elem, bp->b_bufsize);
- }
+ if (ISSET(bp->b_flags, B_META))
+ buf_free_meta_store(bp);
trace(TR_BRELSE, pack(bp->b_vp, bp->b_bufsize), bp->b_lblkno);
- /* clear out various other fields */
- bp->b_bufsize = 0;
- bp->b_datap = (uintptr_t)NULL;
- bp->b_upl = (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...
- */
+ 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;
+ /*
+ * 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;
+ 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_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;
+#ifdef CONFIG_PROTECT
+ bp->b_cpentry = 0;
#endif
- bp->b_lflags = BL_BUSY;
- bp->b_flags = (bp->b_flags & B_HDRALLOC);
- 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;
- /* nuke any credentials we were holding */
- 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);
+ lck_mtx_lock_spin(buf_mtxp);
}
- lck_mtx_lock(buf_mtxp);
-
return (0);
}
dp = BUFHASH(vp, lblkno);
- lck_mtx_lock(buf_mtxp);
relook:
+ lck_mtx_lock_spin(buf_mtxp);
+
if ((bp = incore_locked(vp, lblkno, dp)) == (struct buf *)0) {
lck_mtx_unlock(buf_mtxp);
return (0);
}
SET(bp->b_lflags, BL_WANTED);
- error = msleep((caddr_t)bp, buf_mtxp, (PRIBIO + 1), "buf_invalblkno", NULL);
+ error = msleep((caddr_t)bp, buf_mtxp, PDROP | (PRIBIO + 1), "buf_invalblkno", NULL);
if (error) {
- lck_mtx_unlock(buf_mtxp);
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;
* Unlock the buffer.
*/
CLR(bp->b_lflags, (BL_BUSY | BL_WANTED));
+ buf_busycount--;
lck_mtx_unlock(buf_mtxp);
buf_acquire(buf_t bp, int flags, int slpflag, int slptimeo) {
errno_t error;
- lck_mtx_lock(buf_mtxp);
+ lck_mtx_lock_spin(buf_mtxp);
error = buf_acquire_locked(bp, flags, slpflag, slptimeo);
}
if (ISSET(bp->b_lflags, BL_BUSY)) {
/*
- * since the mutex_lock may block, the buffer
+ * since the lck_mtx_lock may block, the buffer
* may become BUSY, so we need to
* recheck for a NOWAIT request
*/
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;
errno_t
buf_biowait(buf_t bp)
{
- lck_mtx_lock(buf_mtxp);
+ while (!ISSET(bp->b_flags, B_DONE)) {
- DTRACE_IO1(wait__start, buf_t, bp);
- while (!ISSET(bp->b_flags, B_DONE))
- (void) msleep(bp, buf_mtxp, (PRIBIO+1), "buf_biowait", NULL);
- DTRACE_IO1(wait__done, buf_t, bp);
+ lck_mtx_lock_spin(buf_mtxp);
- lck_mtx_unlock(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);
return (0);
}
-/*
- * Wait for the callback operation on a B_CALL buffer to complete.
- */
-void
-buf_biowait_callback(buf_t bp)
-{
- lck_mtx_lock(buf_mtxp);
-
- DTRACE_IO1(wait__start, buf_t, bp);
- while (!ISSET(bp->b_lflags, BL_CALLDONE))
- (void) msleep(bp, buf_mtxp, (PRIBIO+1), "buf_biowait", NULL);
- DTRACE_IO1(wait__done, buf_t, bp);
-
- lck_mtx_unlock(buf_mtxp);
-}
/*
* Mark I/O complete on a buffer.
void
buf_biodone(buf_t bp)
{
+ mount_t mp;
+
KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 387)) | DBG_FUNC_START,
- (int)bp, (int)bp->b_datap, bp->b_flags, 0, 0);
+ bp, bp->b_datap, bp->b_flags, 0, 0);
if (ISSET(bp->b_flags, B_DONE))
panic("biodone already");
fslog_io_error(bp);
}
- if (bp->b_vp && bp->b_vp->v_mount && (bp->b_flags & B_READ) == 0) {
- update_last_io_time(bp->b_vp->v_mount);
+ if (bp->b_vp && bp->b_vp->v_mount) {
+ mp = bp->b_vp->v_mount;
+ } else {
+ mp = NULL;
+ }
+
+ 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);
}
if (kdebug_enable) {
else if (bp->b_flags & B_PAGEIO)
code |= DKIO_PAGING;
+ if (bp->b_flags & B_THROTTLED_IO)
+ code |= DKIO_THROTTLE;
+ else if (bp->b_flags & B_PASSIVE)
+ code |= DKIO_PASSIVE;
+
KERNEL_DEBUG_CONSTANT(FSDBG_CODE(DBG_DKRW, code) | DBG_FUNC_NONE,
- (unsigned int)bp, (unsigned int)bp->b_vp,
+ bp, (uintptr_t)bp->b_vp,
bp->b_resid, bp->b_error, 0);
}
if ((bp->b_vp != NULLVP) &&
- ((bp->b_flags & (B_PAGEIO | B_READ)) == (B_PAGEIO | B_READ)) &&
+ ((bp->b_flags & (B_IOSTREAMING | B_PAGEIO | B_READ)) == (B_PAGEIO | B_READ)) &&
(bp->b_vp->v_mount->mnt_kern_flag & MNTK_ROOTDEV)) {
microuptime(&priority_IO_timestamp_for_root);
hard_throttle_on_root = 0;
}
+
/*
* 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_THROTTLED_IO | B_PASSIVE));
DTRACE_IO1(done, buf_t, bp);
if (!ISSET(bp->b_flags, B_READ) && !ISSET(bp->b_flags, B_RAW))
if (ISSET(bp->b_flags, (B_CALL | B_FILTER))) { /* if necessary, call out */
void (*iodone_func)(struct buf *, void *) = bp->b_iodone;
- void *arg = (void *)bp->b_transaction;
+ 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 (iodone_func == NULL) {
- panic("biodone: bp @ %p has NULL b_iodone!\n", bp);
- } else {
- if (callout)
- SET(bp->b_flags, B_DONE); /* note that it's done */
- (*iodone_func)(bp, arg);
- }
- if (callout) {
- int need_wakeup = 0;
+ 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
- * BL_WANTED indicates that we've decided to wait on the
- * completion of this I/O in a synchronous manner... we
- * still call the callback function, but in addition we
- * will do a wakeup... BL_CALLDONE indicates that the callback
- * routine has completed and its ok for the waiter to take
- * 'ownership' of this bp back
*/
- lck_mtx_lock_spin(buf_mtxp);
-
- if (bp->b_lflags & BL_WANTED) {
- CLR(bp->b_lflags, BL_WANTED);
- need_wakeup = 1;
- }
- SET(bp->b_lflags, BL_CALLDONE);
-
- lck_mtx_unlock(buf_mtxp);
-
- if (need_wakeup)
- wakeup(bp);
-
goto biodone_done;
}
/*
}
biodone_done:
KERNEL_DEBUG((FSDBG_CODE(DBG_FSRW, 387)) | DBG_FUNC_END,
- (int)bp, (int)bp->b_datap, bp->b_flags, 0, 0);
+ (uintptr_t)bp, (uintptr_t)bp->b_datap, bp->b_flags, 0, 0);
}
/*
buf_t bp;
int n = 0;
- lck_mtx_lock(buf_mtxp);
+ lck_mtx_lock_spin(buf_mtxp);
for (bp = bufqueues[BQ_LOCKED].tqh_first; bp;
bp = bp->b_freelist.tqe_next)
int
count_busy_buffers(void)
{
- buf_t bp;
- int nbusy = 0;
-
- lck_mtx_lock(buf_mtxp);
- for (bp = &buf_headers[boot_nbuf_headers]; --bp >= buf_headers; )
- if (!ISSET(bp->b_flags, B_INVAL) && ISSET(bp->b_lflags, BL_BUSY))
- nbusy++;
- lck_mtx_unlock(buf_mtxp);
-
- return (nbusy);
+ return buf_busycount + bufstats.bufs_iobufinuse;
}
#if DIAGNOSTIC
vfs_bufstats()
{
int i, j, count;
- register struct buf *bp;
- register struct bqueues *dp;
+ struct buf *bp;
+ struct bqueues *dp;
int counts[MAXBSIZE/CLBYTES+1];
static char *bname[BQUEUES] =
{ "LOCKED", "LRU", "AGE", "EMPTY", "META", "LAUNDRY" };
}
#endif /* DIAGNOSTIC */
-#define NRESERVEDIOBUFS 64
+#define NRESERVEDIOBUFS 128
buf_t
{
buf_t bp;
- lck_mtx_lock(iobuffer_mtxp);
+ lck_mtx_lock_spin(iobuffer_mtxp);
while (((niobuf_headers - NRESERVEDIOBUFS < bufstats.bufs_iobufinuse) && !priv) ||
(bp = iobufqueue.tqh_first) == NULL) {
bufstats.bufs_iobufsleeps++;
need_iobuffer = 1;
- (void) msleep(&need_iobuffer, iobuffer_mtxp, (PRIBIO+1), (const char *)"alloc_io_buf", NULL);
+ (void) msleep(&need_iobuffer, iobuffer_mtxp, PSPIN | (PRIBIO+1), (const char *)"alloc_io_buf", NULL);
}
TAILQ_REMOVE(&iobufqueue, bp, b_freelist);
bp->b_datap = 0;
bp->b_flags = 0;
bp->b_lflags = BL_BUSY | BL_IOBUF;
+ bp->b_redundancy_flags = 0;
bp->b_blkno = bp->b_lblkno = 0;
#ifdef JOE_DEBUG
bp->b_owner = current_thread();
bp->b_bufsize = 0;
bp->b_upl = NULL;
bp->b_vp = vp;
+#ifdef CONFIG_PROTECT
+ bp->b_cpentry = 0;
+#endif
if (vp && (vp->v_type == VBLK || vp->v_type == VCHR))
bp->b_dev = vp->v_rdev;
need_iobuffer = 0;
need_wakeup = 1;
}
+ if (bufstats.bufs_iobufinuse <= 0)
+ panic("free_io_buf: bp(%p) - bufstats.bufs_iobufinuse < 0", bp);
+
bufstats.bufs_iobufinuse--;
lck_mtx_unlock(iobuffer_mtxp);
void
buf_list_lock(void)
{
- lck_mtx_lock(buf_mtxp);
+ lck_mtx_lock_spin(buf_mtxp);
}
void
static void
bcleanbuf_thread_init(void)
{
+ thread_t thread = THREAD_NULL;
+
/* create worker thread */
- kernel_thread(kernel_task, bcleanbuf_thread);
+ kernel_thread_start((thread_continue_t)bcleanbuf_thread, NULL, &thread);
+ thread_deallocate(thread);
}
+typedef int (*bcleanbufcontinuation)(int);
+
static void
bcleanbuf_thread(void)
{
int loopcnt = 0;
for (;;) {
- lck_mtx_lock(buf_mtxp);
-
- while ( (bp = TAILQ_FIRST(&bufqueues[BQ_LAUNDRY])) == NULL)
- (void)msleep((void *)&bufqueues[BQ_LAUNDRY], buf_mtxp, PRIBIO, "blaundry", NULL);
+ lck_mtx_lock_spin(buf_mtxp);
+ 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
*/
* Buffer is no longer on any free list
*/
SET(bp->b_lflags, BL_BUSY);
+ buf_busycount++;
#ifdef JOE_DEBUG
bp->b_owner = current_thread();
binstailfree(bp, &bufqueues[BQ_LAUNDRY], BQ_LAUNDRY);
blaundrycnt++;
- /* we never leave a busy page on the laundary queue */
+ /* 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
lck_mtx_unlock(buf_mtxp);
-
- if (loopcnt > 10) {
- (void)tsleep((void *)&bufqueues[BQ_LAUNDRY], PRIBIO, "blaundry", 1);
+
+ 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 {
- (void)thread_block(THREAD_CONTINUE_NULL);
+ /* give other threads a chance to run */
+ (void)thread_block(THREAD_CONTINUE_NULL);
loopcnt++;
}
}
}
bp->b_upl = upl;
- kret = ubc_upl_map(upl, (vm_address_t *)&(bp->b_datap));
+ 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(0);
}
+boolean_t
+buffer_cache_gc(int all)
+{
+ buf_t bp;
+ boolean_t did_large_zfree = FALSE;
+ boolean_t need_wakeup = FALSE;
+ int now = buf_timestamp();
+ uint32_t found = 0, total_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()).
+ * No more than 1024 buffers total, and only those not accessed within the
+ * last 30s. We will also only examine 128 buffers during a single grab
+ * of the lock in order to limit lock hold time.
+ */
+ lck_mtx_lock(buf_mtxp);
+ do {
+ found = 0;
+ TAILQ_INIT(&privq);
+ need_wakeup = FALSE;
+
+ while (((bp = TAILQ_FIRST(&bufqueues[BQ_META]))) &&
+ (now > bp->b_timestamp) &&
+ (now - bp->b_timestamp > thresh_hold) &&
+ (found < BUF_MAX_GC_BATCH_SIZE)) {
+
+ /* Remove from free list */
+ bremfree_locked(bp);
+ found++;
+
+#ifdef JOE_DEBUG
+ bp->b_owner = current_thread();
+ bp->b_tag = 12;
+#endif
+
+ /* If dirty, move to laundry queue and remember to do wakeup */
+ if (ISSET(bp->b_flags, B_DELWRI)) {
+ SET(bp->b_lflags, BL_WANTDEALLOC);
+
+ 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);
+ }
+
+ 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);
+ total_found += found;
+
+ } while ((all || (total_found < BUF_MAX_GC_COUNT)) && (found == BUF_MAX_GC_BATCH_SIZE));
+
+ lck_mtx_unlock(buf_mtxp);
+
+ return did_large_zfree;
+}
/*
bp->b_tag = 7;
#endif
SET(bp->b_lflags, BL_BUSY);
+ buf_busycount++;
+
flush_table[buf_count] = bp;
buf_count++;
total_writes++;
static void
bufq_balance_thread_init(void)
{
+ thread_t thread = THREAD_NULL;
if (bufqscanwait++ == 0) {
}
/* create worker thread */
- kernel_thread(kernel_task, bufqscan_thread);
+ kernel_thread_start((thread_continue_t)bufqscan_thread, NULL, &thread);
+ thread_deallocate(thread);
}
/* The workloop for the buffer balancing thread */
/* check if it's stale */
if ((t - bp->b_timestamp) > bufqlim[q].bl_stale) {
- if (bcleanbuf(bp)) {
+ if (bcleanbuf(bp, FALSE)) {
/* buf_bawrite() issued, bp not ready */
moretodo = 1;
} else {
projects / apple / xnu.git / blobdiff