| Message ID | 20250106095613.847700-2-hch@lst.de (mailing list archive) |
|---|---|
| State | New |
| Headers | show |
| Series | [01/15] xfs: fix a double completion for buffers on in-memory targets | expand |
On Mon, Jan 06, 2025 at 10:54:38AM +0100, Christoph Hellwig wrote: > __xfs_buf_submit calls xfs_buf_ioend when b_io_remaining hits zero. For > in-memory buftargs b_io_remaining is never incremented from it's initial > value of 1, so this always happens. Thus the extra call to xfs_buf_ioend > in _xfs_buf_ioapply causes a double completion. Fortunately > __xfs_buf_submit is only used for synchronous reads on in-memory buftargs > due to the peculiarities of how they work, so this is mostly harmless and > just causes a little extra work to be done. Tempted to add: Cc: <stable@vger.kernel.org> # v6.9 though I think backporting isn't strictly necessary because in-memory buffers don't have log items, right? If so, then we don't need to cc stable. Reviewed-by: "Darrick J. Wong" <djwong@kernel.org> --D > Fixes: 5076a6040ca1 ("xfs: support in-memory buffer cache targets") > Signed-off-by: Christoph Hellwig <hch@lst.de> > --- > fs/xfs/xfs_buf.c | 4 +--- > 1 file changed, 1 insertion(+), 3 deletions(-) > > diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c > index aa63b8efd782..787caf0c3254 100644 > --- a/fs/xfs/xfs_buf.c > +++ b/fs/xfs/xfs_buf.c > @@ -1656,10 +1656,8 @@ _xfs_buf_ioapply( > op |= REQ_META; > > /* in-memory targets are directly mapped, no IO required. */ > - if (xfs_buftarg_is_mem(bp->b_target)) { > - xfs_buf_ioend(bp); > + if (xfs_buftarg_is_mem(bp->b_target)) > return; > - } > > /* > * Walk all the vectors issuing IO on them. Set up the initial offset > -- > 2.45.2 > >
On Mon, Jan 06, 2025 at 06:00:35PM -0800, Darrick J. Wong wrote: > Cc: <stable@vger.kernel.org> # v6.9 > > though I think backporting isn't strictly necessary because in-memory > buffers don't have log items, right? If so, then we don't need to cc > stable. The stable CC is fine with me. But as I tried to explain I think it's nor really causing problems right now, just a double call to read_verify. It is a bit of a landmine, though.
diff --git a/fs/xfs/xfs_buf.c b/fs/xfs/xfs_buf.c index aa63b8efd782..787caf0c3254 100644 --- a/fs/xfs/xfs_buf.c +++ b/fs/xfs/xfs_buf.c @@ -1656,10 +1656,8 @@ _xfs_buf_ioapply( op |= REQ_META; /* in-memory targets are directly mapped, no IO required. */ - if (xfs_buftarg_is_mem(bp->b_target)) { - xfs_buf_ioend(bp); + if (xfs_buftarg_is_mem(bp->b_target)) return; - } /* * Walk all the vectors issuing IO on them. Set up the initial offset
__xfs_buf_submit calls xfs_buf_ioend when b_io_remaining hits zero. For in-memory buftargs b_io_remaining is never incremented from it's initial value of 1, so this always happens. Thus the extra call to xfs_buf_ioend in _xfs_buf_ioapply causes a double completion. Fortunately __xfs_buf_submit is only used for synchronous reads on in-memory buftargs due to the peculiarities of how they work, so this is mostly harmless and just causes a little extra work to be done. Fixes: 5076a6040ca1 ("xfs: support in-memory buffer cache targets") Signed-off-by: Christoph Hellwig <hch@lst.de> --- fs/xfs/xfs_buf.c | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-)