The recent generic_file_write() deadlock fix caused
generic_file_buffered_write() to loop inifinitely when presented with a
zero-length iovec segment. Fix.
Note that this fix deliberately avoids calling ->prepare_write(),
->commit_write() etc with a zero-length write. This is because I don't trust
all filesystems to get that right.
This is a cautious approach, for 2.6.17.x. For 2.6.18 we should just go ahead
and call ->prepare_write() and ->commit_write() with the zero length and fix
any broken filesystems. So I'll make that change once this code is stabilised
and backported into 2.6.17.x.
The reason for preferring to call ->prepare_write() and ->commit_write() with
the zero-length segment: a zero-length segment _should_ be sufficiently
uncommon that this is the correct way of handling it. We don't want to
optimise for poorly-written userspace at the expense of well-written
userspace.
Cc: "Vladimir V. Saveliev" <[email protected]>
Cc: Neil Brown <[email protected]>
Cc: Martin Schwidefsky <[email protected]>
Cc: Chris Wright <[email protected]>
Cc: Greg KH <[email protected]>
Cc: <[email protected]>
Cc: walt <[email protected]>
Signed-off-by: Andrew Morton <[email protected]>
Signed-off-by: Linus Torvalds <[email protected]>
break;
}
+ if (unlikely(bytes == 0)) {
+ status = 0;
+ copied = 0;
+ goto zero_length_segment;
+ }
+
status = a_ops->prepare_write(file, page, offset, offset+bytes);
if (unlikely(status)) {
loff_t isize = i_size_read(inode);
page_cache_release(page);
continue;
}
- if (likely(copied > 0)) {
+zero_length_segment:
+ if (likely(copied >= 0)) {
if (!status)
status = copied;
const struct iovec *iov = *iovp;
size_t base = *basep;
- while (bytes) {
+ do {
int copy = min(bytes, iov->iov_len - base);
bytes -= copy;
iov++;
base = 0;
}
- }
+ } while (bytes);
*iovp = iov;
*basep = base;
}