Brad Spencer <[email protected]> writes: >> Interesting that you are still having lockups on current. It has a lot >> of locking fixes (also in 11, not in 10). > > The previous kernel wasn't all that new, in fact, it was pretty old. I > noted the locking fixes as they rolled by in source-change, but until > now didn't have time to try anything newer. It is *entirely* possible > that there is more than one thing doing on here. It is also true that I > don't know if I could make a unmodified -current hang up. I was mostly > just pleased to see the nice improvement.
First, I want to be clear that I welcome Sotiris's efforts to make tunables observable and controllable. which I think is good regardless of hang issues. If you are willing, I suggest trying to run stock current in terms of the ARC etc. settings. I think it's reasonably likely things will be ok, and whether they are or aren't is a really useful datapoint. ZFS doctrine is to use a lot for the ARC, and to free it on memory pressure. This works fine in Solaris and FreeBSD (idk about Linux where ZFS is not first class). In NetBSD we have had deadlocks -- or maybe just missing wakeups -- under memory pressure. >> My guess is that reducing arc is a workaround, not a fix, and is >> avoiding the locking bugs by leaving more free. Still, not crashing is >> good.... I don't mean to say we shouldn't merge the patch, or that it >> isn't helpful to be able to set/observe -- only that there is deeper >> trouble. > > In as much as I was able to tell, the hangs I saw mostly had to do with > running out of some sort of memory in a way that the system could not > recover from or could not prevent. This was the first time in a very > long time that I was able to push the build system as hard as I wanted > to. Agreed, but I would like to know if it's reduced ARC size or the fixes.
