Kevin Corry wrote:
On Friday 11 June 2004 12:23 pm, Kevin Corry wrote:
On Friday 11 June 2004 07:09, christophe.varoqui@xxxxxxx wrote:
Hello,
please consider the following log, illustrating the bug that hits
lk2.6.7+
I've been playing with the dm-mirror code this week and I know for certain
that I've seen the event-nr increment correctly when the mirrors finish
syncing. So the core event counting code should be working fine.
Mike, did one of your recent patches possibly change the event counting in
dm-multipath? I'll take a look once I get to work this morning.
In the "mv-path-code-to-ps" patch (originally posted to dm-devel on May 25),
the fail_path() routine is removed from dm-mpath.c. This routine called
schedule_work(&m->trigger_event) (which eventually leads to a call to
dm_table_event()). However, this schedule_work() call was not added anywhere
else in the code, so it looks like dm-multipath is no longer triggering an
event after a path-failure.
That's right. I messed up.
The problem is that the path-selector was supposed to be where vendors
add their code. As a result - as it is in udm today - it will eventually
need to get detailed error values to decide if they want to fail the
path. This is not possible today as SCSI cannot pass sense data up to
DM, and it is not possible today for SCSI to accurately decode the
vendor specific sense into a generic block layer error value.
Also putting the vendor specifics in the selector has problems wrt each
vendor having to duplicate the path-selection algorithm.
As both problems are not solved, and even my current patch set that was
posted the other day solves the latter problem but is still a work in
progress, I do not know if it is worth it to keep my patches in udm. <-
I cc'd Alasdair for this.