On 11/09/2014 09:27 AM, Pavel Machek wrote: > Hi! > >>>> Oh and did I ask in this thread for /proc/zoneinfo yet? :) >>> >>> Using that same kernel[1], got again into a race, gathered a few more data. >>> >>> This time, I had 1x "urpmq" process [2] hung at 100% CPU , when "kwin" got >>> apparently blocked (100% CPU, too) trying to resize a GUI window. I suppose >>> the resizing operation would mean heavy memory alloc/free. >>> >>> The rest of the system was responsive, I could easily get a console, login, >>> gather the files.. Then, I have *killed* -9 the "urpmq" process, which solved >>> the race and my system is still alive! "kwin" is still running, returned to >>> regular CPU load. >>> >>> Attached is traces from SysRq+l (pressed a few times, wanted to "snapshot" the >>> stack) and /proc/zoneinfo + /proc/vmstat >>> >>> Bisection is not yet meaningful, IMHO, because I cannot be sure that "good" >>> points are really free from this issue. I'd estimate that each test would take >>> +3days, unless I really find a deterministic way to reproduce the issue . >> >> Hi, >> >> I think I finally found the cause by staring into the code... CCing >> people from all 4 separate threads I know about this issue. >> The problem with finding the cause was that the first report I got from >> Markus was about isolate_freepages_block() overhead, and later Norbert >> reported that reverting a patch for isolate_freepages* helped. But the >> problem seems to be that although the loop in isolate_migratepages exits >> because the scanners almost meet (they are within same pageblock), they >> don't truly meet, therefore compact_finished() decides to continue, but >> isolate_migratepages() exits immediately... boom! But indeed e14c720efdd7 >> made this situation possible, as free scaner pfn can now point to a >> middle of pageblock. > > Ok, it seems it happened second time now, again shortly after > resume. I guess I should apply your patch after all. Thanks. > (Or... instead it should go to Linus ASAP -- it fixes known problem > that is affected people, and we want it in soon in case it is not > complete fix.) I don't want to send untested fix, and wasn't able to reproduce the bug myself. I think Norbert could do it rather quickly so I hope he can tell us soon. > Dmesg is in the attachment, perhaps it helps. > Pavel It looks the same as before, so no surprises there, which is good. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>