On Sat 31-08-19 10:03:18, Tetsuo Handa wrote: > On 2019/08/30 19:35, Michal Hocko wrote: > > On Fri 30-08-19 19:04:53, Tetsuo Handa wrote: > >> If /proc/sys/vm/oom_dump_tasks != 0, dump_header() can become very slow > >> because dump_tasks() synchronously reports all OOM victim candidates, and > >> as a result ratelimit test for dump_header() cannot work as expected. > >> > >> This patch defers dump_tasks() till oom_mutex is released. As a result of > >> this patch, the latency between out_of_memory() is called and SIGKILL is > >> sent (and the OOM reaper starts reclaiming memory) will be significantly > >> reduced. > > > > This is adding a lot of code for something that might be simply worked > > around by disabling dump_tasks. Unless there is a real world workload > > that suffers from the latency and depends on the eligible task list then > > I do not think this is mergeable. > > > > People had to use /proc/sys/vm/oom_dump_tasks == 0 (and give up obtaining some > clue) because they worried stalls caused by /proc/sys/vm/oom_dump_tasks != 0 > while they have to use /proc/sys/vm/panic_on_oom == 0 because they don't want the > down time caused by rebooting. The main qustion is whether disabling that information is actually causing any real problems. > This patch avoids stalls (and gives them some clue). > This patch also helps mitigating __ratelimit(&oom_rs) == "always true" problem. > A straightforward improvement. This is a wrong approach to mitigate that problem. Ratelimiting doesn't really work for any operation that takes a longer time. Solving that problem sounds usef in a generic way. > If there are objections we can't apply this change, reasons would be something > like "This change breaks existing userspace scripts that parse OOM messages". No, not really. There is another aspect of inclusion criterion - maintainability and code complexity. This patch doesn't help neither. -- Michal Hocko SUSE Labs
