Hi, Willy, after upgraded to haproxy-ss-20131122, enable and disable servers via socket will crash haproxy, there's no this issue in haproxy-ss-20131031.
Bests, -Igor On Thu, Nov 21, 2013 at 10:42 PM, Willy Tarreau <w...@1wt.eu> wrote: > Hi Igor, > > On Thu, Nov 21, 2013 at 09:03:05PM +0800, Igor wrote: >> Thanks Willy, because I use snapshot haproxy in production, so I have >> no change to do more investigation, glad you could reproduce the bug >> :) > > now you'll have something a bit more reliable to work with. I've just > committed the two following fixes : > > e7b7348 BUG/MEDIUM: checks: fix slow start regression after fix attempt > 004e045 BUG/MAJOR: server: weight calculation fails for map-based algorithms > > The first one is a crash when using slowstart without checks. It's not your > case but you'll probably want to fix it in case you happen to switch to > slowstart. I encountered it while trying to reproduce your bug based on > Lukas' findings. > > The second one fixes the issue you're facing which is also what Lukas > noticed (wrong weight after a set weight on a map-based algorithm). > I can confirm that it resulted in the total backend weight to be larger > than the table, causing out of bounds accesses after a weight change > as soon as the map indx went far enough (depending on your load and > the total initial weights, it could take a few seconds to a few minutes). > > The fix is quite large because I wanted to get rid of all places where > the computations were hazardous (and there were quite a few). From my > opinion and my tests, it now correcty covers all situations. > > Thanks for reporting it! > > Willy >
