/*
* Initiate IO on given buffer.
*/
int
xfs_buf_iorequest(struct xfs_buf *bp)
{
bp->b_flags &= ~(B_INVAL|B_DONE);
bp->b_ioflags &= ~BIO_ERROR;
if (bp->b_flags & B_ASYNC)
BUF_KERNPROC(bp);
if (bp->b_vp == NULL) {
if (bp->b_iocmd == BIO_WRITE) {
bp->b_flags &= ~(B_DELWRI | B_DEFERRED);
bufobj_wref(bp->b_bufobj);
}
bp->b_iooffset = (bp->b_blkno << BBSHIFT);
bstrategy(bp);
} else {
if (bp->b_iocmd == BIO_WRITE) {
/* Mark the buffer clean */
bundirty(bp);
bufobj_wref(bp->b_bufobj);
vfs_busy_pages(bp, 1);
} else if (bp->b_iocmd == BIO_READ) {
vfs_busy_pages(bp, 0);
}
bp->b_iooffset = dbtob(bp->b_blkno);
bstrategy(bp);
}
return 0;
}
/*
* Release blocks associated with the inode ip and stored in the indirect
* block bn. Blocks are free'd in LIFO order up to (but not including)
* lastbn. If level is greater than SINGLE, the block is an indirect block
* and recursive calls to indirtrunc must be used to cleanse other indirect
* blocks.
*
* NB: triple indirect blocks are untested.
*/
static int
ffs_indirtrunc(struct inode *ip, ufs_daddr_t lbn, ufs_daddr_t dbn,
ufs_daddr_t lastbn, int level, long *countp)
{
int i;
struct buf *bp;
struct fs *fs = ip->i_fs;
ufs_daddr_t *bap;
struct vnode *vp;
ufs_daddr_t *copy = NULL, nb, nlbn, last;
long blkcount, factor;
int nblocks, blocksreleased = 0;
int error = 0, allerror = 0;
/*
* Calculate index in current block of last
* block to be kept. -1 indicates the entire
* block so we need not calculate the index.
*/
factor = 1;
for (i = SINGLE; i < level; i++)
factor *= NINDIR(fs);
last = lastbn;
if (lastbn > 0)
last /= factor;
nblocks = btodb(fs->fs_bsize);
/*
* Get buffer of block pointers, zero those entries corresponding
* to blocks to be free'd, and update on disk copy first. Since
* double(triple) indirect before single(double) indirect, calls
* to bmap on these blocks will fail. However, we already have
* the on disk address, so we have to set the bio_offset field
* explicitly instead of letting bread do everything for us.
*/
vp = ITOV(ip);
bp = getblk(vp, lblktodoff(fs, lbn), (int)fs->fs_bsize, 0, 0);
if ((bp->b_flags & B_CACHE) == 0) {
bp->b_flags &= ~(B_ERROR|B_INVAL);
bp->b_cmd = BUF_CMD_READ;
if (bp->b_bcount > bp->b_bufsize)
panic("ffs_indirtrunc: bad buffer size");
/*
* BIO is bio2 which chains back to bio1. We wait
* on bio1.
*/
bp->b_bio2.bio_offset = dbtodoff(fs, dbn);
bp->b_bio1.bio_done = biodone_sync;
bp->b_bio1.bio_flags |= BIO_SYNC;
vfs_busy_pages(vp, bp);
/*
* Access the block device layer using the device vnode
* and the translated block number (bio2) instead of the
* file vnode (vp) and logical block number (bio1).
*
* Even though we are bypassing the vnode layer, we still
* want the vnode state to indicate that an I/O on its behalf
* is in progress.
*/
bio_start_transaction(&bp->b_bio1, &vp->v_track_read);
vn_strategy(ip->i_devvp, &bp->b_bio2);
error = biowait(&bp->b_bio1, "biord");
}
if (error) {
brelse(bp);
*countp = 0;
return (error);
}
bap = (ufs_daddr_t *)bp->b_data;
if (lastbn != -1) {
copy = kmalloc(fs->fs_bsize, M_TEMP, M_WAITOK);
bcopy((caddr_t)bap, (caddr_t)copy, (uint)fs->fs_bsize);
bzero((caddr_t)&bap[last + 1],
(uint)(NINDIR(fs) - (last + 1)) * sizeof (ufs_daddr_t));
if (DOINGASYNC(vp)) {
bawrite(bp);
} else {
error = bwrite(bp);
if (error)
allerror = error;
}
bap = copy;
}
/*
* Recursively free totally unused blocks.
*/
for (i = NINDIR(fs) - 1, nlbn = lbn + 1 - i * factor; i > last;
i--, nlbn += factor) {
nb = bap[i];
if (nb == 0)
continue;
if (level > SINGLE) {
if ((error = ffs_indirtrunc(ip, nlbn, fsbtodb(fs, nb),
(ufs_daddr_t)-1, level - 1, &blkcount)) != 0)
allerror = error;
blocksreleased += blkcount;
}
ffs_blkfree(ip, nb, fs->fs_bsize);
blocksreleased += nblocks;
}
/*
* Recursively free last partial block.
*/
if (level > SINGLE && lastbn >= 0) {
last = lastbn % factor;
nb = bap[i];
if (nb != 0) {
error = ffs_indirtrunc(ip, nlbn, fsbtodb(fs, nb),
last, level - 1, &blkcount);
if (error)
allerror = error;
blocksreleased += blkcount;
}
}
if (copy != NULL) {
kfree(copy, M_TEMP);
} else {
bp->b_flags |= B_INVAL | B_NOCACHE;
brelse(bp);
}
*countp = blocksreleased;
return (allerror);
}
int
ext2_bmaparray(struct vnode *vp, daddr_t bn, daddr_t *bnp, int *runp, int *runb)
{
struct inode *ip;
struct buf *bp;
struct ext2mount *ump;
struct mount *mp;
struct vnode *devvp;
struct indir a[NIADDR+1], *ap;
daddr_t daddr;
e2fs_lbn_t metalbn;
int error, num, maxrun = 0, bsize;
int *nump;
ap = NULL;
ip = VTOI(vp);
mp = vp->v_mount;
ump = VFSTOEXT2(mp);
devvp = ump->um_devvp;
bsize = EXT2_BLOCK_SIZE(ump->um_e2fs);
if (runp) {
maxrun = mp->mnt_iosize_max / bsize - 1;
*runp = 0;
}
if (runb) {
*runb = 0;
}
ap = a;
nump = #
error = ext2_getlbns(vp, bn, ap, nump);
if (error)
return (error);
num = *nump;
if (num == 0) {
*bnp = blkptrtodb(ump, ip->i_db[bn]);
if (*bnp == 0) {
*bnp = -1;
} else if (runp) {
daddr_t bnb = bn;
for (++bn; bn < NDADDR && *runp < maxrun &&
is_sequential(ump, ip->i_db[bn - 1], ip->i_db[bn]);
++bn, ++*runp);
bn = bnb;
if (runb && (bn > 0)) {
for (--bn; (bn >= 0) && (*runb < maxrun) &&
is_sequential(ump, ip->i_db[bn],
ip->i_db[bn + 1]);
--bn, ++*runb);
}
}
return (0);
}
/* Get disk address out of indirect block array */
daddr = ip->i_ib[ap->in_off];
for (bp = NULL, ++ap; --num; ++ap) {
/*
* Exit the loop if there is no disk address assigned yet and
* the indirect block isn't in the cache, or if we were
* looking for an indirect block and we've found it.
*/
metalbn = ap->in_lbn;
if ((daddr == 0 && !incore(&vp->v_bufobj, metalbn)) || metalbn == bn)
break;
/*
* If we get here, we've either got the block in the cache
* or we have a disk address for it, go fetch it.
*/
if (bp)
bqrelse(bp);
bp = getblk(vp, metalbn, bsize, 0, 0, 0);
if ((bp->b_flags & B_CACHE) == 0) {
#ifdef INVARIANTS
if (!daddr)
panic("ext2_bmaparray: indirect block not in cache");
#endif
bp->b_blkno = blkptrtodb(ump, daddr);
bp->b_iocmd = BIO_READ;
bp->b_flags &= ~B_INVAL;
bp->b_ioflags &= ~BIO_ERROR;
vfs_busy_pages(bp, 0);
bp->b_iooffset = dbtob(bp->b_blkno);
bstrategy(bp);
curthread->td_ru.ru_inblock++;
error = bufwait(bp);
if (error) {
brelse(bp);
return (error);
}
}
daddr = ((e2fs_daddr_t *)bp->b_data)[ap->in_off];
if (num == 1 && daddr && runp) {
for (bn = ap->in_off + 1;
bn < MNINDIR(ump) && *runp < maxrun &&
is_sequential(ump,
((e2fs_daddr_t *)bp->b_data)[bn - 1],
((e2fs_daddr_t *)bp->b_data)[bn]);
++bn, ++*runp);
bn = ap->in_off;
if (runb && bn) {
for (--bn; bn >= 0 && *runb < maxrun &&
is_sequential(ump,
((e2fs_daddr_t *)bp->b_data)[bn],
((e2fs_daddr_t *)bp->b_data)[bn + 1]);
--bn, ++*runb);
}
}
}
if (bp)
bqrelse(bp);
/*
* Since this is FFS independent code, we are out of scope for the
* definitions of BLK_NOCOPY and BLK_SNAP, but we do know that they
* will fall in the range 1..um_seqinc, so we use that test and
* return a request for a zeroed out buffer if attempts are made
* to read a BLK_NOCOPY or BLK_SNAP block.
*/
if ((ip->i_flags & SF_SNAPSHOT) && daddr > 0 && daddr < ump->um_seqinc){
*bnp = -1;
return (0);
}
*bnp = blkptrtodb(ump, daddr);
if (*bnp == 0) {
*bnp = -1;
}
return (0);
}
static int
fuse_write_biobackend(struct vnode *vp, struct uio *uio,
struct ucred *cred, struct fuse_filehandle *fufh, int ioflag)
{
struct fuse_vnode_data *fvdat = VTOFUD(vp);
struct buf *bp;
daddr_t lbn;
int bcount;
int n, on, err = 0;
const int biosize = fuse_iosize(vp);
KASSERT(uio->uio_rw == UIO_WRITE, ("ncl_write mode"));
FS_DEBUG("resid=%zx offset=%jx fsize=%jx\n",
uio->uio_resid, uio->uio_offset, fvdat->filesize);
if (vp->v_type != VREG)
return (EIO);
if (uio->uio_offset < 0)
return (EINVAL);
if (uio->uio_resid == 0)
return (0);
if (ioflag & IO_APPEND)
uio_setoffset(uio, fvdat->filesize);
/*
* Find all of this file's B_NEEDCOMMIT buffers. If our writes
* would exceed the local maximum per-file write commit size when
* combined with those, we must decide whether to flush,
* go synchronous, or return err. We don't bother checking
* IO_UNIT -- we just make all writes atomic anyway, as there's
* no point optimizing for something that really won't ever happen.
*/
do {
if (fuse_isdeadfs(vp)) {
err = ENXIO;
break;
}
lbn = uio->uio_offset / biosize;
on = uio->uio_offset & (biosize - 1);
n = MIN((unsigned)(biosize - on), uio->uio_resid);
FS_DEBUG2G("lbn %ju, on %d, n %d, uio offset %ju, uio resid %zd\n",
(uintmax_t)lbn, on, n,
(uintmax_t)uio->uio_offset, uio->uio_resid);
again:
/*
* Handle direct append and file extension cases, calculate
* unaligned buffer size.
*/
if (uio->uio_offset == fvdat->filesize && n) {
/*
* Get the buffer (in its pre-append state to maintain
* B_CACHE if it was previously set). Resize the
* nfsnode after we have locked the buffer to prevent
* readers from reading garbage.
*/
bcount = on;
FS_DEBUG("getting block from OS, bcount %d\n", bcount);
bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
if (bp != NULL) {
long save;
err = fuse_vnode_setsize(vp, cred,
uio->uio_offset + n);
if (err) {
brelse(bp);
break;
}
save = bp->b_flags & B_CACHE;
bcount += n;
allocbuf(bp, bcount);
bp->b_flags |= save;
}
} else {
/*
* Obtain the locked cache block first, and then
* adjust the file's size as appropriate.
*/
bcount = on + n;
if ((off_t)lbn * biosize + bcount < fvdat->filesize) {
if ((off_t)(lbn + 1) * biosize < fvdat->filesize)
bcount = biosize;
else
bcount = fvdat->filesize -
(off_t)lbn *biosize;
}
FS_DEBUG("getting block from OS, bcount %d\n", bcount);
bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
if (bp && uio->uio_offset + n > fvdat->filesize) {
err = fuse_vnode_setsize(vp, cred,
uio->uio_offset + n);
if (err) {
brelse(bp);
break;
}
}
}
if (!bp) {
err = EINTR;
break;
}
/*
* Issue a READ if B_CACHE is not set. In special-append
* mode, B_CACHE is based on the buffer prior to the write
* op and is typically set, avoiding the read. If a read
* is required in special append mode, the server will
* probably send us a short-read since we extended the file
* on our end, resulting in b_resid == 0 and, thusly,
* B_CACHE getting set.
*
* We can also avoid issuing the read if the write covers
* the entire buffer. We have to make sure the buffer state
* is reasonable in this case since we will not be initiating
* I/O. See the comments in kern/vfs_bio.c's getblk() for
* more information.
*
* B_CACHE may also be set due to the buffer being cached
* normally.
*/
if (on == 0 && n == bcount) {
bp->b_flags |= B_CACHE;
bp->b_flags &= ~B_INVAL;
bp->b_ioflags &= ~BIO_ERROR;
}
if ((bp->b_flags & B_CACHE) == 0) {
bp->b_iocmd = BIO_READ;
vfs_busy_pages(bp, 0);
fuse_io_strategy(vp, bp);
if ((err = bp->b_error)) {
brelse(bp);
break;
}
}
if (bp->b_wcred == NOCRED)
bp->b_wcred = crhold(cred);
/*
* If dirtyend exceeds file size, chop it down. This should
* not normally occur but there is an append race where it
* might occur XXX, so we log it.
*
* If the chopping creates a reverse-indexed or degenerate
* situation with dirtyoff/end, we 0 both of them.
*/
if (bp->b_dirtyend > bcount) {
FS_DEBUG("FUSE append race @%lx:%d\n",
(long)bp->b_blkno * biosize,
bp->b_dirtyend - bcount);
bp->b_dirtyend = bcount;
}
if (bp->b_dirtyoff >= bp->b_dirtyend)
bp->b_dirtyoff = bp->b_dirtyend = 0;
/*
* If the new write will leave a contiguous dirty
* area, just update the b_dirtyoff and b_dirtyend,
* otherwise force a write rpc of the old dirty area.
*
* While it is possible to merge discontiguous writes due to
* our having a B_CACHE buffer ( and thus valid read data
* for the hole), we don't because it could lead to
* significant cache coherency problems with multiple clients,
* especially if locking is implemented later on.
*
* as an optimization we could theoretically maintain
* a linked list of discontinuous areas, but we would still
* have to commit them separately so there isn't much
* advantage to it except perhaps a bit of asynchronization.
*/
if (bp->b_dirtyend > 0 &&
(on > bp->b_dirtyend || (on + n) < bp->b_dirtyoff)) {
/*
* Yes, we mean it. Write out everything to "storage"
* immediately, without hesitation. (Apart from other
* reasons: the only way to know if a write is valid
* if its actually written out.)
*/
bwrite(bp);
if (bp->b_error == EINTR) {
err = EINTR;
break;
}
goto again;
}
err = uiomove((char *)bp->b_data + on, n, uio);
/*
* Since this block is being modified, it must be written
* again and not just committed. Since write clustering does
* not work for the stage 1 data write, only the stage 2
* commit rpc, we have to clear B_CLUSTEROK as well.
*/
bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
if (err) {
bp->b_ioflags |= BIO_ERROR;
bp->b_error = err;
brelse(bp);
break;
}
/*
* Only update dirtyoff/dirtyend if not a degenerate
* condition.
*/
if (n) {
if (bp->b_dirtyend > 0) {
bp->b_dirtyoff = MIN(on, bp->b_dirtyoff);
bp->b_dirtyend = MAX((on + n), bp->b_dirtyend);
} else {
bp->b_dirtyoff = on;
bp->b_dirtyend = on + n;
}
vfs_bio_set_valid(bp, on, n);
}
err = bwrite(bp);
if (err)
break;
} while (uio->uio_resid > 0 && n > 0);
if (fuse_sync_resize && (fvdat->flag & FN_SIZECHANGE) != 0)
fuse_vnode_savesize(vp, cred);
return (err);
}
static int
fuse_read_biobackend(struct vnode *vp, struct uio *uio,
struct ucred *cred, struct fuse_filehandle *fufh)
{
struct buf *bp;
daddr_t lbn;
int bcount;
int err = 0, n = 0, on = 0;
off_t filesize;
const int biosize = fuse_iosize(vp);
FS_DEBUG("resid=%zx offset=%jx fsize=%jx\n",
uio->uio_resid, uio->uio_offset, VTOFUD(vp)->filesize);
if (uio->uio_resid == 0)
return (0);
if (uio->uio_offset < 0)
return (EINVAL);
bcount = MIN(MAXBSIZE, biosize);
filesize = VTOFUD(vp)->filesize;
do {
if (fuse_isdeadfs(vp)) {
err = ENXIO;
break;
}
lbn = uio->uio_offset / biosize;
on = uio->uio_offset & (biosize - 1);
FS_DEBUG2G("biosize %d, lbn %d, on %d\n", biosize, (int)lbn, on);
/*
* Obtain the buffer cache block. Figure out the buffer size
* when we are at EOF. If we are modifying the size of the
* buffer based on an EOF condition we need to hold
* nfs_rslock() through obtaining the buffer to prevent
* a potential writer-appender from messing with n_size.
* Otherwise we may accidentally truncate the buffer and
* lose dirty data.
*
* Note that bcount is *not* DEV_BSIZE aligned.
*/
if ((off_t)lbn * biosize >= filesize) {
bcount = 0;
} else if ((off_t)(lbn + 1) * biosize > filesize) {
bcount = filesize - (off_t)lbn *biosize;
}
bp = getblk(vp, lbn, bcount, PCATCH, 0, 0);
if (!bp)
return (EINTR);
/*
* If B_CACHE is not set, we must issue the read. If this
* fails, we return an error.
*/
if ((bp->b_flags & B_CACHE) == 0) {
bp->b_iocmd = BIO_READ;
vfs_busy_pages(bp, 0);
err = fuse_io_strategy(vp, bp);
if (err) {
brelse(bp);
return (err);
}
}
/*
* on is the offset into the current bp. Figure out how many
* bytes we can copy out of the bp. Note that bcount is
* NOT DEV_BSIZE aligned.
*
* Then figure out how many bytes we can copy into the uio.
*/
n = 0;
if (on < bcount)
n = MIN((unsigned)(bcount - on), uio->uio_resid);
if (n > 0) {
FS_DEBUG2G("feeding buffeater with %d bytes of buffer %p,"
" saying %d was asked for\n",
n, bp->b_data + on, n + (int)bp->b_resid);
err = uiomove(bp->b_data + on, n, uio);
}
brelse(bp);
FS_DEBUG2G("end of turn, err %d, uio->uio_resid %zd, n %d\n",
err, uio->uio_resid, n);
} while (err == 0 && uio->uio_resid > 0 && n > 0);
return (err);
}
/*
* Read data to a buf, including read-ahead if we find this to be beneficial.
* cluster_read replaces bread.
*/
int
cluster_read(struct vnode *vp, u_quad_t filesize, daddr_t lblkno, long size,
struct ucred *cred, long totread, int seqcount, int gbflags,
struct buf **bpp)
{
struct buf *bp, *rbp, *reqbp;
struct bufobj *bo;
daddr_t blkno, origblkno;
int maxra, racluster;
int error, ncontig;
int i;
error = 0;
bo = &vp->v_bufobj;
if (!unmapped_buf_allowed)
gbflags &= ~GB_UNMAPPED;
/*
* Try to limit the amount of read-ahead by a few
* ad-hoc parameters. This needs work!!!
*/
racluster = vp->v_mount->mnt_iosize_max / size;
maxra = seqcount;
maxra = min(read_max, maxra);
maxra = min(nbuf/8, maxra);
if (((u_quad_t)(lblkno + maxra + 1) * size) > filesize)
maxra = (filesize / size) - lblkno;
/*
* get the requested block
*/
*bpp = reqbp = bp = getblk(vp, lblkno, size, 0, 0, gbflags);
if (bp == NULL)
return (EBUSY);
origblkno = lblkno;
/*
* if it is in the cache, then check to see if the reads have been
* sequential. If they have, then try some read-ahead, otherwise
* back-off on prospective read-aheads.
*/
if (bp->b_flags & B_CACHE) {
if (!seqcount) {
return 0;
} else if ((bp->b_flags & B_RAM) == 0) {
return 0;
} else {
bp->b_flags &= ~B_RAM;
BO_RLOCK(bo);
for (i = 1; i < maxra; i++) {
/*
* Stop if the buffer does not exist or it
* is invalid (about to go away?)
*/
rbp = gbincore(&vp->v_bufobj, lblkno+i);
if (rbp == NULL || (rbp->b_flags & B_INVAL))
break;
/*
* Set another read-ahead mark so we know
* to check again. (If we can lock the
* buffer without waiting)
*/
if ((((i % racluster) == (racluster - 1)) ||
(i == (maxra - 1)))
&& (0 == BUF_LOCK(rbp,
LK_EXCLUSIVE | LK_NOWAIT, NULL))) {
rbp->b_flags |= B_RAM;
BUF_UNLOCK(rbp);
}
}
BO_RUNLOCK(bo);
if (i >= maxra) {
return 0;
}
lblkno += i;
}
reqbp = bp = NULL;
/*
* If it isn't in the cache, then get a chunk from
* disk if sequential, otherwise just get the block.
*/
} else {
off_t firstread = bp->b_offset;
int nblks;
long minread;
KASSERT(bp->b_offset != NOOFFSET,
("cluster_read: no buffer offset"));
ncontig = 0;
/*
* Adjust totread if needed
*/
minread = read_min * size;
if (minread > totread)
totread = minread;
/*
* Compute the total number of blocks that we should read
* synchronously.
*/
if (firstread + totread > filesize)
totread = filesize - firstread;
nblks = howmany(totread, size);
if (nblks > racluster)
nblks = racluster;
/*
* Now compute the number of contiguous blocks.
*/
if (nblks > 1) {
error = VOP_BMAP(vp, lblkno, NULL,
&blkno, &ncontig, NULL);
/*
* If this failed to map just do the original block.
*/
if (error || blkno == -1)
ncontig = 0;
}
/*
* If we have contiguous data available do a cluster
* otherwise just read the requested block.
*/
if (ncontig) {
/* Account for our first block. */
ncontig = min(ncontig + 1, nblks);
if (ncontig < nblks)
nblks = ncontig;
bp = cluster_rbuild(vp, filesize, lblkno,
blkno, size, nblks, gbflags, bp);
lblkno += (bp->b_bufsize / size);
} else {
bp->b_flags |= B_RAM;
bp->b_iocmd = BIO_READ;
lblkno += 1;
}
}
/*
* handle the synchronous read so that it is available ASAP.
*/
if (bp) {
if ((bp->b_flags & B_CLUSTER) == 0) {
vfs_busy_pages(bp, 0);
}
bp->b_flags &= ~B_INVAL;
bp->b_ioflags &= ~BIO_ERROR;
if ((bp->b_flags & B_ASYNC) || bp->b_iodone != NULL)
BUF_KERNPROC(bp);
bp->b_iooffset = dbtob(bp->b_blkno);
bstrategy(bp);
#ifdef RACCT
if (racct_enable) {
PROC_LOCK(curproc);
racct_add_buf(curproc, bp, 0);
PROC_UNLOCK(curproc);
}
#endif /* RACCT */
curthread->td_ru.ru_inblock++;
}
/*
* If we have been doing sequential I/O, then do some read-ahead.
*/
while (lblkno < (origblkno + maxra)) {
error = VOP_BMAP(vp, lblkno, NULL, &blkno, &ncontig, NULL);
if (error)
break;
if (blkno == -1)
break;
/*
* We could throttle ncontig here by maxra but we might as
* well read the data if it is contiguous. We're throttled
* by racluster anyway.
*/
if (ncontig) {
ncontig = min(ncontig + 1, racluster);
rbp = cluster_rbuild(vp, filesize, lblkno, blkno,
size, ncontig, gbflags, NULL);
lblkno += (rbp->b_bufsize / size);
if (rbp->b_flags & B_DELWRI) {
bqrelse(rbp);
continue;
}
} else {
rbp = getblk(vp, lblkno, size, 0, 0, gbflags);
lblkno += 1;
if (rbp->b_flags & B_DELWRI) {
bqrelse(rbp);
continue;
}
rbp->b_flags |= B_ASYNC | B_RAM;
rbp->b_iocmd = BIO_READ;
rbp->b_blkno = blkno;
}
if (rbp->b_flags & B_CACHE) {
rbp->b_flags &= ~B_ASYNC;
bqrelse(rbp);
continue;
}
if ((rbp->b_flags & B_CLUSTER) == 0) {
vfs_busy_pages(rbp, 0);
}
rbp->b_flags &= ~B_INVAL;
rbp->b_ioflags &= ~BIO_ERROR;
if ((rbp->b_flags & B_ASYNC) || rbp->b_iodone != NULL)
BUF_KERNPROC(rbp);
rbp->b_iooffset = dbtob(rbp->b_blkno);
bstrategy(rbp);
#ifdef RACCT
if (racct_enable) {
PROC_LOCK(curproc);
racct_add_buf(curproc, rbp, 0);
PROC_UNLOCK(curproc);
}
#endif /* RACCT */
curthread->td_ru.ru_inblock++;
}
if (reqbp) {
/*
* Like bread, always brelse() the buffer when
* returning an error.
*/
error = bufwait(reqbp);
if (error != 0) {
brelse(reqbp);
*bpp = NULL;
}
}
return (error);
}
int
puffs_biowrite(struct vnode *vp, struct uio *uio, int ioflag,
struct ucred *cred)
{
int biosize = vp->v_mount->mnt_stat.f_iosize;
struct buf *bp;
struct vattr vattr;
off_t loffset, fsize;
int boff, bytes;
int error = 0;
int bcount;
int trivial;
KKASSERT(uio->uio_rw == UIO_WRITE);
KKASSERT(vp->v_type == VREG);
if (uio->uio_offset < 0)
return EINVAL;
if (uio->uio_resid == 0)
return 0;
/*
* If IO_APPEND then load uio_offset. We restart here if we cannot
* get the append lock.
*
* We need to obtain exclusize lock if we intend to modify file size
* in order to guarentee the append point with multiple contending
* writers.
*/
if (ioflag & IO_APPEND) {
/* XXXDF relock if necessary */
KKASSERT(vn_islocked(vp) == LK_EXCLUSIVE);
error = VOP_GETATTR(vp, &vattr);
if (error)
return error;
uio->uio_offset = puffs_meta_getsize(vp);
}
do {
boff = uio->uio_offset & (biosize-1);
loffset = uio->uio_offset - boff;
bytes = (int)szmin((unsigned)(biosize - boff), uio->uio_resid);
again:
/*
* Handle direct append and file extension cases, calculate
* unaligned buffer size. When extending B_CACHE will be
* set if possible. See UIO_NOCOPY note below.
*/
fsize = puffs_meta_getsize(vp);
if (uio->uio_offset + bytes > fsize) {
trivial = (uio->uio_segflg != UIO_NOCOPY &&
uio->uio_offset <= fsize);
puffs_meta_setsize(vp, uio->uio_offset + bytes,
trivial);
}
bp = getblk(vp, loffset, biosize, 0, 0);
if (bp == NULL) {
error = EINTR;
break;
}
/*
* Actual bytes in buffer which we care about
*/
if (loffset + biosize < fsize)
bcount = biosize;
else
bcount = (int)(fsize - loffset);
/*
* Avoid a read by setting B_CACHE where the data we
* intend to write covers the entire buffer. Note
* that the buffer may have been set to B_CACHE by
* puffs_meta_setsize() above or otherwise inherited the
* flag, but if B_CACHE isn't set the buffer may be
* uninitialized and must be zero'd to accomodate
* future seek+write's.
*
* See the comments in kern/vfs_bio.c's getblk() for
* more information.
*
* When doing a UIO_NOCOPY write the buffer is not
* overwritten and we cannot just set B_CACHE unconditionally
* for full-block writes.
*/
if (boff == 0 && bytes == biosize &&
uio->uio_segflg != UIO_NOCOPY) {
bp->b_flags |= B_CACHE;
bp->b_flags &= ~(B_ERROR | B_INVAL);
}
/*
* b_resid may be set due to file EOF if we extended out.
* The NFS bio code will zero the difference anyway so
* just acknowledged the fact and set b_resid to 0.
*/
if ((bp->b_flags & B_CACHE) == 0) {
bp->b_cmd = BUF_CMD_READ;
bp->b_bio2.bio_done = puffs_iodone;
bp->b_bio2.bio_flags |= BIO_SYNC;
vfs_busy_pages(vp, bp);
error = puffs_doio(vp, &bp->b_bio2, uio->uio_td);
if (error) {
brelse(bp);
break;
}
bp->b_resid = 0;
}
/*
* If dirtyend exceeds file size, chop it down. This should
* not normally occur but there is an append race where it
* might occur XXX, so we log it.
*
* If the chopping creates a reverse-indexed or degenerate
* situation with dirtyoff/end, we 0 both of them.
*/
if (bp->b_dirtyend > bcount) {
kprintf("PUFFS append race @%08llx:%d\n",
(long long)bp->b_bio2.bio_offset,
bp->b_dirtyend - bcount);
bp->b_dirtyend = bcount;
}
if (bp->b_dirtyoff >= bp->b_dirtyend)
bp->b_dirtyoff = bp->b_dirtyend = 0;
/*
* If the new write will leave a contiguous dirty
* area, just update the b_dirtyoff and b_dirtyend,
* otherwise force a write rpc of the old dirty area.
*
* While it is possible to merge discontiguous writes due to
* our having a B_CACHE buffer ( and thus valid read data
* for the hole), we don't because it could lead to
* significant cache coherency problems with multiple clients,
* especially if locking is implemented later on.
*
* as an optimization we could theoretically maintain
* a linked list of discontinuous areas, but we would still
* have to commit them separately so there isn't much
* advantage to it except perhaps a bit of asynchronization.
*/
if (bp->b_dirtyend > 0 &&
(boff > bp->b_dirtyend ||
(boff + bytes) < bp->b_dirtyoff)
) {
if (bwrite(bp) == EINTR) {
error = EINTR;
break;
}
goto again;
}
error = uiomove(bp->b_data + boff, bytes, uio);
/*
* Since this block is being modified, it must be written
* again and not just committed. Since write clustering does
* not work for the stage 1 data write, only the stage 2
* commit rpc, we have to clear B_CLUSTEROK as well.
*/
bp->b_flags &= ~(B_NEEDCOMMIT | B_CLUSTEROK);
if (error) {
brelse(bp);
break;
}
/*
* Only update dirtyoff/dirtyend if not a degenerate
* condition.
*
* The underlying VM pages have been marked valid by
* virtue of acquiring the bp. Because the entire buffer
* is marked dirty we do not have to worry about cleaning
* out the related dirty bits (and wouldn't really know
* how to deal with byte ranges anyway)
*/
if (bytes) {
if (bp->b_dirtyend > 0) {
bp->b_dirtyoff = imin(boff, bp->b_dirtyoff);
bp->b_dirtyend = imax(boff + bytes,
bp->b_dirtyend);
} else {
bp->b_dirtyoff = boff;
bp->b_dirtyend = boff + bytes;
}
}
if (ioflag & IO_SYNC) {
if (ioflag & IO_INVAL)
bp->b_flags |= B_NOCACHE;
error = bwrite(bp);
if (error)
break;
} else {
bdwrite(bp);
}
} while (uio->uio_resid > 0 && bytes > 0);
return error;
}
int
puffs_bioread(struct vnode *vp, struct uio *uio, int ioflag,
struct ucred *cred)
{
int biosize = vp->v_mount->mnt_stat.f_iosize;
struct buf *bp;
struct vattr vattr;
off_t lbn, loffset, fsize;
size_t n;
int boff;
int error = 0;
KKASSERT(uio->uio_rw == UIO_READ);
KKASSERT(vp->v_type == VREG);
if (uio->uio_offset < 0)
return EINVAL;
if (uio->uio_resid == 0)
return 0;
/*
* Cache consistency can only be maintained approximately.
*
* GETATTR is called to synchronize the file size.
*
* NOTE: In the normal case the attribute cache is not
* cleared which means GETATTR may use cached data and
* not immediately detect changes made on the server.
*/
error = VOP_GETATTR(vp, &vattr);
if (error)
return error;
/*
* Loop until uio exhausted or we hit EOF
*/
do {
bp = NULL;
lbn = uio->uio_offset / biosize;
boff = uio->uio_offset & (biosize - 1);
loffset = lbn * biosize;
fsize = puffs_meta_getsize(vp);
if (loffset + boff >= fsize) {
n = 0;
break;
}
bp = getblk(vp, loffset, biosize, 0, 0);
if (bp == NULL)
return EINTR;
/*
* If B_CACHE is not set, we must issue the read. If this
* fails, we return an error.
*/
if ((bp->b_flags & B_CACHE) == 0) {
bp->b_cmd = BUF_CMD_READ;
bp->b_bio2.bio_done = puffs_iodone;
bp->b_bio2.bio_flags |= BIO_SYNC;
vfs_busy_pages(vp, bp);
error = puffs_doio(vp, &bp->b_bio2, uio->uio_td);
if (error) {
brelse(bp);
return error;
}
}
/*
* on is the offset into the current bp. Figure out how many
* bytes we can copy out of the bp. Note that bcount is
* NOT DEV_BSIZE aligned.
*
* Then figure out how many bytes we can copy into the uio.
*/
n = biosize - boff;
if (n > uio->uio_resid)
n = uio->uio_resid;
if (loffset + boff + n > fsize)
n = fsize - loffset - boff;
if (n > 0)
error = uiomove(bp->b_data + boff, n, uio);
if (bp)
brelse(bp);
} while (error == 0 && uio->uio_resid > 0 && n > 0);
return error;
}
static int
ext2_indirtrunc(struct inode *ip, daddr_t lbn, daddr_t dbn,
daddr_t lastbn, int level, e4fs_daddr_t *countp)
{
struct buf *bp;
struct m_ext2fs *fs = ip->i_e2fs;
struct vnode *vp;
e2fs_daddr_t *bap, *copy;
int i, nblocks, error = 0, allerror = 0;
e2fs_lbn_t nb, nlbn, last;
e4fs_daddr_t blkcount, factor, blocksreleased = 0;
/*
* Calculate index in current block of last
* block to be kept. -1 indicates the entire
* block so we need not calculate the index.
*/
factor = 1;
for (i = SINGLE; i < level; i++)
factor *= NINDIR(fs);
last = lastbn;
if (lastbn > 0)
last /= factor;
nblocks = btodb(fs->e2fs_bsize);
/*
* Get buffer of block pointers, zero those entries corresponding
* to blocks to be free'd, and update on disk copy first. Since
* double(triple) indirect before single(double) indirect, calls
* to bmap on these blocks will fail. However, we already have
* the on disk address, so we have to set the b_blkno field
* explicitly instead of letting bread do everything for us.
*/
vp = ITOV(ip);
bp = getblk(vp, lbn, (int)fs->e2fs_bsize, 0, 0, 0);
if ((bp->b_flags & (B_DONE | B_DELWRI)) == 0) {
bp->b_iocmd = BIO_READ;
if (bp->b_bcount > bp->b_bufsize)
panic("ext2_indirtrunc: bad buffer size");
bp->b_blkno = dbn;
vfs_busy_pages(bp, 0);
bp->b_iooffset = dbtob(bp->b_blkno);
bstrategy(bp);
error = bufwait(bp);
}
if (error) {
brelse(bp);
*countp = 0;
return (error);
}
bap = (e2fs_daddr_t *)bp->b_data;
copy = malloc(fs->e2fs_bsize, M_TEMP, M_WAITOK);
bcopy((caddr_t)bap, (caddr_t)copy, (u_int)fs->e2fs_bsize);
bzero((caddr_t)&bap[last + 1],
(NINDIR(fs) - (last + 1)) * sizeof(e2fs_daddr_t));
if (last == -1)
bp->b_flags |= B_INVAL;
if (DOINGASYNC(vp)) {
bdwrite(bp);
} else {
error = bwrite(bp);
if (error)
allerror = error;
}
bap = copy;
/*
* Recursively free totally unused blocks.
*/
for (i = NINDIR(fs) - 1, nlbn = lbn + 1 - i * factor; i > last;
i--, nlbn += factor) {
nb = bap[i];
if (nb == 0)
continue;
if (level > SINGLE) {
if ((error = ext2_indirtrunc(ip, nlbn,
fsbtodb(fs, nb), (int32_t)-1, level - 1, &blkcount)) != 0)
allerror = error;
blocksreleased += blkcount;
}
ext2_blkfree(ip, nb, fs->e2fs_bsize);
blocksreleased += nblocks;
}
/*
* Recursively free last partial block.
*/
if (level > SINGLE && lastbn >= 0) {
last = lastbn % factor;
nb = bap[i];
if (nb != 0) {
if ((error = ext2_indirtrunc(ip, nlbn, fsbtodb(fs, nb),
last, level - 1, &blkcount)) != 0)
allerror = error;
blocksreleased += blkcount;
}
}
free(copy, M_TEMP);
*countp = blocksreleased;
return (allerror);
}
static int
ext2_indirtrunc(struct inode *ip, daddr_t lbn, off_t doffset, daddr_t lastbn,
int level, long *countp)
{
int i;
struct buf *bp;
struct ext2_sb_info *fs = ip->i_e2fs;
daddr_t *bap;
struct vnode *vp;
daddr_t *copy, nb, nlbn, last;
long blkcount, factor;
int nblocks, blocksreleased = 0;
int error = 0, allerror = 0;
/*
* Calculate index in current block of last
* block to be kept. -1 indicates the entire
* block so we need not calculate the index.
*/
factor = 1;
for (i = SINGLE; i < level; i++)
factor *= NINDIR(fs);
last = lastbn;
if (lastbn > 0)
last /= factor;
nblocks = btodb(fs->s_blocksize);
/*
* Get buffer of block pointers, zero those entries corresponding
* to blocks to be free'd, and update on disk copy first. Since
* double(triple) indirect before single(double) indirect, calls
* to bmap on these blocks will fail. However, we already have
* the on disk address, so we have to set the bio_offset field
* explicitly instead of letting bread do everything for us.
*/
vp = ITOV(ip);
bp = getblk(vp, lblktodoff(fs, lbn), (int)fs->s_blocksize, 0, 0);
if ((bp->b_flags & B_CACHE) == 0) {
bp->b_flags &= ~(B_ERROR | B_INVAL);
bp->b_cmd = BUF_CMD_READ;
if (bp->b_bcount > bp->b_bufsize)
panic("ext2_indirtrunc: bad buffer size");
bp->b_bio2.bio_offset = doffset;
bp->b_bio1.bio_done = biodone_sync;
bp->b_bio1.bio_flags |= BIO_SYNC;
vfs_busy_pages(bp->b_vp, bp);
vn_strategy(vp, &bp->b_bio1);
error = biowait(&bp->b_bio1, "biord");
}
if (error) {
brelse(bp);
*countp = 0;
return (error);
}
bap = (daddr_t *)bp->b_data;
MALLOC(copy, daddr_t *, fs->s_blocksize, M_TEMP, M_WAITOK);
bcopy((caddr_t)bap, (caddr_t)copy, (u_int)fs->s_blocksize);
bzero((caddr_t)&bap[last + 1],
(u_int)(NINDIR(fs) - (last + 1)) * sizeof (daddr_t));
if (last == -1)
bp->b_flags |= B_INVAL;
error = bwrite(bp);
if (error)
allerror = error;
bap = copy;
/*
* Recursively free totally unused blocks.
*/
for (i = NINDIR(fs) - 1, nlbn = lbn + 1 - i * factor; i > last;
i--, nlbn += factor) {
nb = bap[i];
if (nb == 0)
continue;
if (level > SINGLE) {
if ((error = ext2_indirtrunc(ip, nlbn,
fsbtodoff(fs, nb), (daddr_t)-1, level - 1, &blkcount)) != 0)
allerror = error;
blocksreleased += blkcount;
}
ext2_blkfree(ip, nb, fs->s_blocksize);
blocksreleased += nblocks;
}
/*
* Recursively free last partial block.
*/
if (level > SINGLE && lastbn >= 0) {
last = lastbn % factor;
nb = bap[i];
if (nb != 0) {
error = ext2_indirtrunc(ip, nlbn, fsbtodoff(fs, nb),
last, level - 1, &blkcount);
if (error)
allerror = error;
blocksreleased += blkcount;
}
}
FREE(copy, M_TEMP);
*countp = blocksreleased;
return (allerror);
}
/*
* Indirect blocks are now on the vnode for the file. They are given negative
* logical block numbers. Indirect blocks are addressed by the negative
* address of the first data block to which they point. Double indirect blocks
* are addressed by one less than the address of the first indirect block to
* which they point. Triple indirect blocks are addressed by one less than
* the address of the first double indirect block to which they point.
*
* ext2_bmaparray does the bmap conversion, and if requested returns the
* array of logical blocks which must be traversed to get to a block.
* Each entry contains the offset into that block that gets you to the
* next block and the disk address of the block (if it is assigned).
*/
static
int
ext2_bmaparray(struct vnode *vp, ext2_daddr_t bn, ext2_daddr_t *bnp,
struct indir *ap, int *nump, int *runp, int *runb)
{
struct inode *ip;
struct buf *bp;
struct ext2_mount *ump;
struct mount *mp;
struct ext2_sb_info *fs;
struct indir a[NIADDR+1], *xap;
ext2_daddr_t daddr;
long metalbn;
int error, maxrun, num;
ip = VTOI(vp);
mp = vp->v_mount;
ump = VFSTOEXT2(mp);
fs = ip->i_e2fs;
#ifdef DIAGNOSTIC
if ((ap != NULL && nump == NULL) || (ap == NULL && nump != NULL))
panic("ext2_bmaparray: invalid arguments");
#endif
if (runp) {
*runp = 0;
}
if (runb) {
*runb = 0;
}
maxrun = mp->mnt_iosize_max / mp->mnt_stat.f_iosize - 1;
xap = ap == NULL ? a : ap;
if (!nump)
nump = #
error = ext2_getlbns(vp, bn, xap, nump);
if (error)
return (error);
num = *nump;
if (num == 0) {
*bnp = blkptrtodb(ump, ip->i_db[bn]);
if (*bnp == 0)
*bnp = -1;
else if (runp) {
daddr_t bnb = bn;
for (++bn; bn < NDADDR && *runp < maxrun &&
is_sequential(ump, ip->i_db[bn - 1], ip->i_db[bn]);
++bn, ++*runp);
bn = bnb;
if (runb && (bn > 0)) {
for (--bn; (bn >= 0) && (*runb < maxrun) &&
is_sequential(ump, ip->i_db[bn],
ip->i_db[bn+1]);
--bn, ++*runb);
}
}
return (0);
}
/* Get disk address out of indirect block array */
daddr = ip->i_ib[xap->in_off];
for (bp = NULL, ++xap; --num; ++xap) {
/*
* Exit the loop if there is no disk address assigned yet and
* the indirect block isn't in the cache, or if we were
* looking for an indirect block and we've found it.
*/
metalbn = xap->in_lbn;
if ((daddr == 0 &&
!findblk(vp, dbtodoff(fs, metalbn), FINDBLK_TEST)) ||
metalbn == bn) {
break;
}
/*
* If we get here, we've either got the block in the cache
* or we have a disk address for it, go fetch it.
*/
if (bp)
bqrelse(bp);
xap->in_exists = 1;
bp = getblk(vp, lblktodoff(fs, metalbn),
mp->mnt_stat.f_iosize, 0, 0);
if ((bp->b_flags & B_CACHE) == 0) {
#ifdef DIAGNOSTIC
if (!daddr)
panic("ext2_bmaparray: indirect block not in cache");
#endif
/*
* This runs through ext2_strategy using bio2 to
* cache the disk offset, then comes back through
* bio1. So we want to wait on bio1
*/
bp->b_bio1.bio_done = biodone_sync;
bp->b_bio1.bio_flags |= BIO_SYNC;
bp->b_bio2.bio_offset = fsbtodoff(fs, daddr);
bp->b_flags &= ~(B_INVAL|B_ERROR);
bp->b_cmd = BUF_CMD_READ;
vfs_busy_pages(bp->b_vp, bp);
vn_strategy(bp->b_vp, &bp->b_bio1);
error = biowait(&bp->b_bio1, "biord");
if (error) {
brelse(bp);
return (error);
}
}
daddr = ((ext2_daddr_t *)bp->b_data)[xap->in_off];
if (num == 1 && daddr && runp) {
for (bn = xap->in_off + 1;
bn < MNINDIR(ump) && *runp < maxrun &&
is_sequential(ump,
((ext2_daddr_t *)bp->b_data)[bn - 1],
((ext2_daddr_t *)bp->b_data)[bn]);
++bn, ++*runp);
bn = xap->in_off;
if (runb && bn) {
for(--bn; bn >= 0 && *runb < maxrun &&
is_sequential(ump, ((daddr_t *)bp->b_data)[bn],
((daddr_t *)bp->b_data)[bn+1]);
--bn, ++*runb);
}
}
}
if (bp)
bqrelse(bp);
daddr = blkptrtodb(ump, daddr);
*bnp = daddr == 0 ? -1 : daddr;
return (0);
}
