On Wed, Mar 21, 2012 at 4:56 PM, Bernardo Dal Seno <[email protected]> wrote: > On Wed, Mar 21, 2012 at 17:51, Guido Trotter <[email protected]> wrote: >> On Wed, Mar 21, 2012 at 4:49 PM, Iustin Pop <[email protected]> wrote: >>> On Wed, Mar 21, 2012 at 04:43:36PM +0000, Guido Trotter wrote: >>>> On Wed, Mar 21, 2012 at 4:39 PM, Iustin Pop <[email protected]> wrote: >>>> > On Wed, Mar 21, 2012 at 05:38:15PM +0100, Bernardo Dal Seno wrote: >>>> >> On Wed, Mar 21, 2012 at 16:59, Michael Hanselmann <[email protected]> >>>> >> wrote: >>>> >> > Instead of acquiring the BGL in exclusive mode (which blocks all other >>>> >> > operations), we acquire all node and instance locks in shared mode >>>> >> > before verifying the configuration. >>>> >> >>>> >> Actually, also group locks are acquired. >>>> > >>>> > Can you guys describe what relation this patch has to the previous one? >>>> > What was the result of the offline discussion? >>>> > >>>> >>>> While the previous one fixes the fact that queries currently acquire >>>> no locks *except* the BGL (which is good because we already had given >>>> up on locks on queries and the BGL was an exception, and still the >>>> only "potential" problems are on node-add and such, with that change) >>>> it's still good for cluster verify not to acquire the BGL and conflict >>>> with other jobs which could go on at the same time, especially if they >>>> are read only. In general we shouldn't have added a new LU requiring >>>> the BGL, and went in the opposite direction (of removing the BGL if >>>> possible) so this needed to happen anyway. >>> >>> So the conclusion was to: >>> >>> - keep queries not acquiring the BGL at all (prev patch) >>> - fix c-v to not use exclusive BGL >>> >>> Correct? >> >> Yes. > > Except that apparently there are more locks in some queries. So we > probably need to postpone the fix for LUXI. >
Ah I see... well, then I agree, let's revert that patch and schedule a better fix for 2.6 rather than such an abrupt change for 2.5 at post-rc stage. Guido
