gmirror crash writing to disk? Or is it su+j crash?

[Ian Lepore]

On Fri, 2013-08-30 at 21:50 +0200, Edward Tomasz Napierała wrote:
> Wiadomość napisana przez Zaphod Beeblebrox <zbeeble at gmail.com> w dniu 29 sie 2013, o godz. 23:35:
> > So I have a system running:
> > 
> > FreeBSD walk.dclg.ca 9.2-RC3 FreeBSD 9.2-RC3 # r254952: Wed Aug 28 03:02:55
> > EDT 2013     root at walk.dclg.ca:/usr/obj/usr/src/sys/STRIKE  i386
> > 
> > and it has two 2T SATA disks.  To keep this post short, the crash.txt is
> > here.
> > 
> > https://uk.eicat.ca/owncloud/public.php?service=files&t=fea9d25579fe0c4afb808859e80e1493
> 
> Login error.
> 
> > now curiously, while running a "make -j4 buildkernel" ... almost every time
> > ... it crashes with:
> > 
> > g_vfs_done():mirror/walke[WRITE(offset=516764794880, length=65536)]error =
> > 11
> > /usr: got error 11 while accessing filesystem
> > panic: softdep_deallocate_dependencies: unrecovered I/O error
> 
> This is softupdates panic caused by write operation returning error 11, which,
> according to 'man errno', is EDEADLK.
> 
> To be honest, I have no idea why gmirror might be returning this error.
> 
> > ... no error report from the hard drives, simply an error report from the
> > mirror.
> 
> Note that ahci(4) does not log errors unless you're running with bootverbose.
> 
> > The filesystem is ufs with su+j... but I'm not sure this matters here.
> 
> It does, kind of - without soft updates/SUJ, the error would be non-fatal - it
> wouldn't panic the box, but it would (probably) cause data corruption.

One of the few places in the kernel that uses EDEADLK is in geom_io.c
(line 642 in -current) in g_io_transient_map_bio()...

		g_io_deliver(bp, EDEADLK/* XXXKIB */);

-- Ian