On 9/03/2020 16:18, Theodore Y. Ts'o wrote:
Did the panic happen immediately, or did things hang until the storage
recovered, and*then* it rebooted. Or did the hard reset and reboot
happened before the storage network connection was restored?
The panic (well it was just frozen, no stacktrace or automatic reboot)
did happen *after* storage came back online.
So nothing happens while the storage is offline, even if we wait until
the scsi timeout is exceeded (180s * 6).
It's only when the storage returns that the filesystem goes read-only /
panic (depending on the error setting).
Fundamentally I think what's going on is that even though there is an
I/O error reported back to the OS, but in some cases, the outstanding
I/O actually happens. So in the error=panic case, we do update the
superblock saying that the file system contains inconsistencies. And
then we reboot. But it appears that even though host rebooted, the
storage area network*did* manage to send the I/O to the device.
It seems that by updating the superblock to state that filesystem
contains errors, things are made worse.
At the moment it does this, the storage is already accessible again, so
it seems logic the I/O is written.
I'm not sure what we can really do here, other than simply making the
SCSI timeout infinite. The problem is that storage area networks are
flaky. Sometimes I/O's make it through, and even though we get an
error, it's an error from the local SCSI layer --- and it's possible
that I/O will make it through. In other cases, even though the
storage area network was disconnected at the time we sent the I/O
saying the file system has problems, and then rebooted, the I/O
actually makes it through. Given that, assuming that if we're not
sure, forcing an full file system check is better part of valor.
If we do reset the VM before storage is back, the filesystem check just
goes fine in automatic mode.
So I think we should (in some cases) not try to update the superblock
anymore on I/O errors, but just go read-only/panic.
Cause it seems like updating the superblock makes things worse.
Or changes could be made to e2fsck to allow automatic repair of this
kind of error for example?
And if it hangs forever, and we do a hard reset reboot, I don't know
*what* to trust from the storage area network. Ideally, there would
be some way to do a hard reset of the storage area network so that all
outstanding I/O's from the host that we are about to reset will get
forgotten before we do actually the hard reset.
- Ted
