Re: [RFC PATCH] libfrog: make cmn_err() emit each message atomically to avoid torn output

On Thu, Jul 16, 2026 at 09:44:28AM +0200, Avinesh Kumar wrote:
> Hello Darrick,
> 
> 
> On 7/15/26 9:09 PM, Darrick J. Wong wrote:
> > On Wed, Jul 15, 2026 at 07:28:21PM +0200, Avinesh Kumar wrote:
> > > From: Avinesh Kumar <[email protected]>
> > > 
> > > fstests xfs/033 fails sporadically with a spurious blank line in the
> > > xfs_repair output:
> > > 
> > > - output mismatch (see /opt/xfstests/results//xfs/033.out.bad)
> > >      --- tests/xfs/033.out	2026-06-24 15:52:51.000000000 -0400
> > >      +++ /opt/xfstests/results//xfs/033.out.bad	2026-07-14 18:54:46.582495041 -0400
> > >      @@ -103,6 +103,7 @@
> > >       Phase 3 - for each AG...
> > >               - scan and clear agi unlinked lists...
> > >               - process known inodes and perform inode discovery...
> > >      +
> > >       bad magic number 0xffff on inode INO
> > >       bad version number 0xffffffff on inode INO
> > >       inode identifier 18446744073709551615 mismatch on inode INO
> > > 
> > > Root cause is in cmn_err() (libfrog/util.c), which emits a message
> > > and its newline as two separate writes to unbuffered stderr.
> > > xfs_repair's threads all share stderr. If one is preempted between the
> > > two writes, another thread's line lands in between:
> > > 
> > > (snips from `cat -A 033.raw`) -
> > > Phase 3 - for each AG...$
> > >          - scan and clear agi unlinked lists...$
> > >          - process known inodes and perform inode discovery...$
> > > Metadata corruption detected at 0x445dd3, xfs_inode block 0x80/0x4000        - agno = 0$
> > > $
> > > bad CRC for inode 128$
> > > bad magic number 0x0 on inode 128$
> > > 
> > > which should be like -
> > > 
> > > Phase 3 - for each AG...$
> > >          - scan and clear agi unlinked lists...$
> > >          - process known inodes and perform inode discovery...$
> > > Metadata corruption detected at 0x445dd3, xfs_inode block 0x80/0x4000$
> > >          - agno = 0$
> > > bad CRC for inode 130$
> > > bad magic number 0x0 on inode 130$
> > > 
> > > _filter_repair() strips the noise line it glued onto but not the lone
> > > newline, which then fails the golden diff.
> > > 
> > > Make the two writes atomic.
> > > 
> > > Signed-off-by: Avinesh Kumar <[email protected]>
> > > ---
> > > 
> > > Hi, please give feedback if this should be fixed in a different way.
> > > Thanks!
> > > 
> > >   libfrog/util.c | 2 ++
> > >   1 file changed, 2 insertions(+)
> > > 
> > > diff --git a/libfrog/util.c b/libfrog/util.c
> > > index 5bae5bab..3b6df917 100644
> > > --- a/libfrog/util.c
> > > +++ b/libfrog/util.c
> > > @@ -116,8 +116,10 @@ cmn_err(int level, char *fmt, ...)
> > >   	va_list	ap;
> > >   	va_start(ap, fmt);
> > > +	flockfile(stderr);
> > >   	vfprintf(stderr, fmt, ap);
> > >   	fputs("\n", stderr);
> > > +	funlockfile(stderr);
> > 
> > I think do_error() in xfs_repair.c is going to need the same treatment
> > but otherwise this chunk looks ok to me.
> > 
> Thanks for the review. I can send a follow up patch for do_error(),
> or do you want these to go in the same patch?

They probably ought to be in the same patch.

--D

> Regards,
> Avinesh
> 
> > --D
> > 
> > >   	va_end(ap);
> > >   }
> > > -- 
> > > 2.55.0
> > > 
> > > 
> 
>