Hi Dave, Dave Chinner <david@xxxxxxxxxxxxx> writes: > On Mon, Oct 28, 2024 at 11:44:00PM +0530, Ritesh Harjani wrote: >> >> Hi Dave, >> >> Dave Chinner <david@xxxxxxxxxxxxx> writes: >> >> > On Mon, Oct 28, 2024 at 06:39:36AM +0530, Ritesh Harjani wrote: >> >> >> >> Hi Dave, >> >> >> >> Dave Chinner <david@xxxxxxxxxxxxx> writes: >> >> >> >> > On Fri, Oct 25, 2024 at 09:15:53AM +0530, Ritesh Harjani (IBM) wrote: >> >> >> iomap will not return -ENOTBLK in case of dio atomic writes. But let's >> >> >> also add a WARN_ON_ONCE and return -EIO as a safety net. >> >> >> >> >> >> Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@xxxxxxxxx> >> >> >> --- >> >> >> fs/ext4/file.c | 10 +++++++++- >> >> >> 1 file changed, 9 insertions(+), 1 deletion(-) >> >> >> >> >> >> diff --git a/fs/ext4/file.c b/fs/ext4/file.c >> >> >> index f9516121a036..af6ebd0ac0d6 100644 >> >> >> --- a/fs/ext4/file.c >> >> >> +++ b/fs/ext4/file.c >> >> >> @@ -576,8 +576,16 @@ static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from) >> >> >> iomap_ops = &ext4_iomap_overwrite_ops; >> >> >> ret = iomap_dio_rw(iocb, from, iomap_ops, &ext4_dio_write_ops, >> >> >> dio_flags, NULL, 0); >> >> >> - if (ret == -ENOTBLK) >> >> >> + if (ret == -ENOTBLK) { >> >> >> ret = 0; >> >> >> + /* >> >> >> + * iomap will never return -ENOTBLK if write fails for atomic >> >> >> + * write. But let's just add a safety net. >> >> >> + */ >> >> >> + if (WARN_ON_ONCE(iocb->ki_flags & IOCB_ATOMIC)) >> >> >> + ret = -EIO; >> >> >> + } >> >> > >> >> > Why can't the iomap code return EIO in this case for IOCB_ATOMIC? >> >> > That way we don't have to put this logic into every filesystem. >> >> >> >> This was origially intended as a safety net hence the WARN_ON_ONCE. >> >> Later Darrick pointed out that we still might have an unconverted >> >> condition in iomap which can return ENOTBLK for DIO atomic writes (page >> >> cache invalidation). >> > >> > Yes. That's my point - iomap knows that it's an atomic write, it >> > knows that invalidation failed, and it knows that there is no such >> > thing as buffered atomic writes. So there is no possible fallback >> > here, and it should be returning EIO in the page cache invalidation >> > failure case and not ENOTBLK. >> > >> >> So the iomap DIO can return following as return values which can make >> some filesystems fallback to buffered-io (if they implement fallback >> logic) - >> (1) -ENOTBLK -> this is only returned for pagecache invalidation failure. >> (2) 0 or partial write size -> This can never happen for atomic writes >> (since we are only allowing for single fsblock as of now). > > Even when we allow multi-FSB atomic writes, the definition of > atomic write is still "all or nothing". There is no scope for "short > writes" when IOCB_ATOMIC is set - any condition that means we can't > write the entire IO as a single bio, we need to abort and return > EINVAL. yes. As long as it is a single bio, I agree even the short write condition should not hit based on the current iomap code. > > Hence -ENOTBLK should never be returned by iomap for atomic DIO > writes - we need to say -EINVAL if the write could not be issued > atomically for whatever reason it may be so the application knows > that atomic IO submission was not possible for that IO. > Agreed Dave. That is what iomap is doing today for atomic write code. (Except maybe one minor difference where it returns -EAGAIN in case of page cache invalidation assuming the failure maybe transient and the request could be tried again). >> Now looking at XFS, it never fallsback to buffered-io ever except just 2 >> cases - >> 1. When pagecache invalidation fails in iomap (can never happen for >> atomic writes) > > Why can't this happen for atomic DIO writes? It's the same failure > cases as for normal DIO writes, isn't it? (i.e. race with mmap > writes) > I meant after the patch which adds atomic write support in iomap code from John, make sure we don't return -ENOTBLK in case of atomic write request. > My point is that if it's an atomic write, this failure should get > turned into -EINVAL by the iomap code. We do not want a fallback to > buffered IO when this situation happens for atomic IO. > >> 2. On unaligned DIO writes to reflinked CoW (not possible for atomic writes) > > This path doesn't ever go through iomap - XFS catches that case > before it calls into iomap, so it's not relevant to how iomap > behaves w.r.t atomic IO. > Right. >> So it anyways should never happen that XFS ever fallback to buffered-io >> for DIO atomic writes. Even today it does not fallback to buffered-io >> for non-atomic short DIO writes. >> >> >> You pointed it right that it should be fixed in iomap. However do you >> >> think filesystems can still keep this as safety net (maybe no need of >> >> WARN_ON_ONCE). >> > >> > I don't see any point in adding "impossible to hit" checks into >> > filesystems just in case some core infrastructure has a bug >> > introduced.... >> >> Yes, that is true for XFS. EXT4 however can return -ENOTBLK for short >> writes, though it should not happen for current atomic write case where >> we are only allowing for 1 fsblock. > > Yes, but the -ENOTBLK error returned from ext4_iomap_end() if > nothing was written does not get returned to ext4 from > __iomap_dio_rw(). It is consumed by the iomap code: > > /* magic error code to fall back to buffered I/O */ > if (ret == -ENOTBLK) { > wait_for_completion = true; > ret = 0; > } > > This means that all the IO that was issued gets completed before > returning to the caller and that's how the short write comes about. > > -ENOTBLK is *not returned to the caller* on a short write - yes. That's my understanding too of the short write case handling in iomap. > iomap_dio_rw will return 0 (success). The caller then has to look > at the iov_iter state to determine if the write was fully completed. > This is exactly what the ext4 code currently does for all DIO > writes, not just those that return -ENOTBLK. > yes. Agreed. >> I would still like to go with a WARN_ON_ONCE where we are calling ext4 >> buffered-io handling for DIO fallback writes. This is to catch any bugs >> even in future when we move to multi-fsblock case (until we have atomic >> write support for buffered-io). > > Your choice, but please realise that it is not going to catch short > atomic writes at all. > Thanks Dave. Yes, I would like to maybe keep a WARN_ON_ONCE since ext4 has a fallback handling logic where a short DIO or -ENOTBLK case could be later handled by buffered-io logic (though I agree iomap won't let it happen for atomic write case). But a WARN_ON_ONCE just before buffered-io fallback handling logic in ext4 DIO path would be my preferred choice only to make sure we could catch any unwanted bugs in future too. So I was thinking of this change instead - diff --git a/fs/ext4/file.c b/fs/ext4/file.c index 8116bd78910b..61787a37e9d4 100644 --- a/fs/ext4/file.c +++ b/fs/ext4/file.c @@ -599,6 +599,13 @@ static ssize_t ext4_dio_write_iter(struct kiocb *iocb, struct iov_iter *from) ssize_t err; loff_t endbyte; + /* + * There is no support for atomic writes on buffered-io yet, + * we should never fallback to buffered-io for DIO atomic + * writes. + */ + WARN_ON_ONCE(iocb->ki_flags & IOCB_ATOMIC); + offset = iocb->ki_pos; err = ext4_buffered_write_iter(iocb, from); if (err < 0) diff --git a/fs/ext4/inode.c b/fs/ext4/inode.c index fcdee27b9aa2..26b3c84d7f64 100644 --- a/fs/ext4/inode.c +++ b/fs/ext4/inode.c @@ -3449,12 +3449,16 @@ static int ext4_iomap_end(struct inode *inode, loff_t offset, loff_t length, { /* * Check to see whether an error occurred while writing out the data to - * the allocated blocks. If so, return the magic error code so that we - * fallback to buffered I/O and attempt to complete the remainder of - * the I/O. Any blocks that may have been allocated in preparation for - * the direct I/O will be reused during buffered I/O. + * the allocated blocks. If so, return the magic error code for + * non-atomic write so that we fallback to buffered I/O and attempt to + * complete the remainder of the I/O. + * For atomic writes we will simply fail the I/O request if we coudn't + * write anything. For non-atomic writes, any blocks that may have been + * allocated in preparation for the direct I/O will be reused during + * buffered I/O. */ - if (flags & (IOMAP_WRITE | IOMAP_DIRECT) && written == 0) + if (!(flags & IOMAP_ATOMIC) && (flags & (IOMAP_WRITE | IOMAP_DIRECT)) + && written == 0) return -ENOTBLK; return 0; > -Dave. > -- > Dave Chinner > david@xxxxxxxxxxxxx Thanks a lot for the review! -ritesh