int
fuse_vnode_setsize(struct vnode *vp, struct ucred *cred, off_t newsize)
{
struct fuse_vnode_data *fvdat = VTOFUD(vp);
off_t oldsize;
int err = 0;
ASSERT_VOP_ELOCKED(vp, "fuse_vnode_setsize");
oldsize = fvdat->filesize;
fvdat->filesize = newsize;
fvdat->flag |= FN_SIZECHANGE;
if (newsize < oldsize) {
err = vtruncbuf(vp, cred, newsize, fuse_iosize(vp));
}
vnode_pager_setsize(vp, newsize);
return err;
}
int
fuse_vnode_setsize(struct vnode *vp, struct ucred *cred, off_t newsize)
{
struct fuse_vnode_data *fvdat = VTOFUD(vp);
off_t oldsize;
int err = 0;
FS_DEBUG("inode=%ju oldsize=%ju newsize=%ju\n",
(uintmax_t)VTOI(vp), (uintmax_t)fvdat->filesize,
(uintmax_t)newsize);
ASSERT_VOP_ELOCKED(vp, "fuse_vnode_setsize");
oldsize = fvdat->filesize;
fvdat->filesize = newsize;
fvdat->flag |= FN_SIZECHANGE;
if (newsize < oldsize) {
err = vtruncbuf(vp, cred, newsize, fuse_iosize(vp));
}
vnode_pager_setsize(vp, newsize);
fuse_invalidate_attr(vp);
return err;
}
int
fuse_io_strategy(struct vnode *vp, struct buf *bp)
{
struct fuse_filehandle *fufh;
struct fuse_vnode_data *fvdat = VTOFUD(vp);
struct ucred *cred;
struct uio *uiop;
struct uio uio;
struct iovec io;
int error = 0;
const int biosize = fuse_iosize(vp);
MPASS(vp->v_type == VREG || vp->v_type == VDIR);
MPASS(bp->b_iocmd == BIO_READ || bp->b_iocmd == BIO_WRITE);
FS_DEBUG("inode=%ju offset=%jd resid=%ld\n",
(uintmax_t)VTOI(vp), (intmax_t)(((off_t)bp->b_blkno) * biosize),
bp->b_bcount);
error = fuse_filehandle_getrw(vp,
(bp->b_iocmd == BIO_READ) ? FUFH_RDONLY : FUFH_WRONLY, &fufh);
if (error) {
printf("FUSE: strategy: filehandles are closed\n");
bp->b_ioflags |= BIO_ERROR;
bp->b_error = error;
return (error);
}
cred = bp->b_iocmd == BIO_READ ? bp->b_rcred : bp->b_wcred;
uiop = &uio;
uiop->uio_iov = &io;
uiop->uio_iovcnt = 1;
uiop->uio_segflg = UIO_SYSSPACE;
uiop->uio_td = curthread;
/*
* clear BIO_ERROR and B_INVAL state prior to initiating the I/O. We
* do this here so we do not have to do it in all the code that
* calls us.
*/
bp->b_flags &= ~B_INVAL;
bp->b_ioflags &= ~BIO_ERROR;
KASSERT(!(bp->b_flags & B_DONE),
("fuse_io_strategy: bp %p already marked done", bp));
if (bp->b_iocmd == BIO_READ) {
io.iov_len = uiop->uio_resid = bp->b_bcount;
io.iov_base = bp->b_data;
uiop->uio_rw = UIO_READ;
uiop->uio_offset = ((off_t)bp->b_blkno) * biosize;
error = fuse_read_directbackend(vp, uiop, cred, fufh);
if ((!error && uiop->uio_resid) ||
(fsess_opt_brokenio(vnode_mount(vp)) && error == EIO &&
uiop->uio_offset < fvdat->filesize && fvdat->filesize > 0 &&
uiop->uio_offset >= fvdat->cached_attrs.va_size)) {
/*
* If we had a short read with no error, we must have
* hit a file hole. We should zero-fill the remainder.
* This can also occur if the server hits the file EOF.
*
* Holes used to be able to occur due to pending
* writes, but that is not possible any longer.
*/
int nread = bp->b_bcount - uiop->uio_resid;
int left = uiop->uio_resid;
if (error != 0) {
printf("FUSE: Fix broken io: offset %ju, "
" resid %zd, file size %ju/%ju\n",
(uintmax_t)uiop->uio_offset,
uiop->uio_resid, fvdat->filesize,
fvdat->cached_attrs.va_size);
error = 0;
}
if (left > 0)
bzero((char *)bp->b_data + nread, left);
uiop->uio_resid = 0;
}
if (error) {
bp->b_ioflags |= BIO_ERROR;
bp->b_error = error;
}
} else {
/*
* If we only need to commit, try to commit
*/
if (bp->b_flags & B_NEEDCOMMIT) {
FS_DEBUG("write: B_NEEDCOMMIT flags set\n");
}
/*
* Setup for actual write
*/
if ((off_t)bp->b_blkno * biosize + bp->b_dirtyend >
fvdat->filesize)
bp->b_dirtyend = fvdat->filesize -
(off_t)bp->b_blkno * biosize;
if (bp->b_dirtyend > bp->b_dirtyoff) {
io.iov_len = uiop->uio_resid = bp->b_dirtyend
- bp->b_dirtyoff;
uiop->uio_offset = (off_t)bp->b_blkno * biosize
+ bp->b_dirtyoff;
io.iov_base = (char *)bp->b_data + bp->b_dirtyoff;
uiop->uio_rw = UIO_WRITE;
error = fuse_write_directbackend(vp, uiop, cred, fufh, 0);
if (error == EINTR || error == ETIMEDOUT
|| (!error && (bp->b_flags & B_NEEDCOMMIT))) {
bp->b_flags &= ~(B_INVAL | B_NOCACHE);
if ((bp->b_flags & B_PAGING) == 0) {
bdirty(bp);
bp->b_flags &= ~B_DONE;
}
if ((error == EINTR || error == ETIMEDOUT) &&
(bp->b_flags & B_ASYNC) == 0)
bp->b_flags |= B_EINTR;
} else {
if (error) {
bp->b_ioflags |= BIO_ERROR;
bp->b_flags |= B_INVAL;
bp->b_error = error;
}
bp->b_dirtyoff = bp->b_dirtyend = 0;
}
} else {
bp->b_resid = 0;
bufdone(bp);
return (0);
}
}
bp->b_resid = uiop->uio_resid;
bufdone(bp);
return (error);
}
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);
}