On Wed 17-10-18 19:06:22, Tetsuo Handa wrote:
> syzbot is hitting RCU stall at shmem_fault() [1].
> This is because memcg-OOM events with no eligible task (current thread
> is marked as OOM-unkillable) continued calling dump_header() from
> out_of_memory() enabled by commit 3100dab2aa09dc6e ("mm: memcontrol:
> print proper OOM header when no eligible victim left.").
>
> Michal proposed ratelimiting dump_header() [2]. But I don't think that
> that patch is appropriate because that patch does not ratelimit
>
> "%s invoked oom-killer: gfp_mask=%#x(%pGg), nodemask=%*pbl, order=%d, oom_score_adj=%hd\n"
> "Out of memory and no killable processes...\n"
>
> messages which can be printed for every few milliseconds (i.e. effectively
> denial of service for console users) until the OOM situation is solved.
>
> Let's make sure that next dump_header() waits for at least 60 seconds from
> previous "Out of memory and no killable processes..." message. Michal is
> thinking that any interval is meaningless without knowing the printk()
> throughput. But since printk() is synchronous unless handed over to
> somebody else by commit dbdda842fe96f893 ("printk: Add console owner and
> waiter logic to load balance console writes"), it is likely that all OOM
> messages from this out_of_memory() request is already flushed to consoles
> when pr_warn("Out of memory and no killable processes...\n") returned.
> Thus, we will be able to allow console users to do what they need to do.
>
> To summarize, this patch allows threads in requested memcg to complete
> memory allocation requests for doing recovery operation, and also allows
> administrators to manually do recovery operation from console if
> OOM-unkillable thread is failing to solve the OOM situation automatically.
Could you explain why this is any better than using a well established
ratelimit approach?
--
Michal Hocko
SUSE Labs
