Re: [dm-devel] Direct syslog calls from multipathd

[christophe varoqui <christophe varoqui free fr>]

On mar, 2005-03-29 at 16:05 -0500, goggin, edward wrote:
> On Tue, 29 Mar 2005 00:48:55 +0200
> christophe varoqui <christophe varoqui free fr> wrote:
> > A new snapshot should have poped up on the ftp (0.4.4-pre4).
> > It contains an async logger thread for the daemon.
> > 
> > Edward, can you test if it solves the issues with direct 
> > syslog call you
> > spotted ?
> 
> Looks good.
> 
> After fixing my multipath.conf file device node blacklist to
> specify "hdc" and "hdd" instead of just "hd", [:((]
> I had three consecutive NDU test runs without hanging up on
> synchronous syslog calls from multipathd.  Considering I
> couldn't get one of these test runs to work before, I'd
> consider the syslog deadlock problem fixed.
> 
> The bad news is that the first run hung up on the inode
> deadlock involving bdget_disk() calls in the dm map
> reload code path (__set_size, __lock_fs, or __unlock_fs)
> one an inode which is already locked.  Also, the
> multipathd died on the last of the 3 runs.  No idea what
> was wrong.
> 
> I noticed you have changed multipath to allow concurrent
> execution.  Was this done in order to not lose recovery
> from i/o failure events in the waitevent and path test
> failures in the checkerloop pthreads of multipathd?
> 
That was the idea, yes.
In early design, multipath was an all-or-nothing thing.

As it evolved to trim its scope to specific multipaths, I thought it was
not necessary to keep it from running in parralel on different
multipaths.

I eyeballed the code and didn't see a justification to keep it running
alone.

Did I miss something important there ?

> --
-- 
christophe varoqui <christophe varoqui free fr>