On Sun, 2007-05-13 at 23:46 +0200, Jan-Frode Myklebust wrote:
On Mon, May 14, 2007 at 12:20:43AM +0300, Timo Sirainen wrote:
I haven't even started doing the index code cleanups. But I did write a small summary about it: http://dovecot.org/list/dovecot/2007-May/022591.html
Which got me thinking.. Do you think changing locking method might help with these assertion failures ? Currently we're using default lock_method, but maybe one of the others are more appropriate for shared file-systems ? ... maybe even just to change code paths if this is a race we're seeing. Any thoughts ?
Code paths between fcntl and flock are pretty much the same. Unless there's a bug in GPFS it shouldn't make a difference which one you use. You can always try of course. Changing to dotlock would make it use a bit different code paths, but it also would make it slower.
The biggest difference is between mmap_disable=yes and =no, but unless GPFS supports shared mmaps it's probably not a good idea to set that to "no".