On Mon 17-12-18 04:25:46, Matthew Wilcox wrote: > On Mon, Dec 17, 2018 at 07:51:27PM +0900, Tetsuo Handa wrote: > > On 2018/12/17 18:33, Michal Hocko wrote: > > > On Sun 16-12-18 19:51:57, Matthew Wilcox wrote: > > > [...] > > >> Ah, yes, that makes perfect sense. Thank you for the explanation. > > >> > > >> I wonder if the correct fix, however, is not to move the check for > > >> GFP_NOFS in out_of_memory() down to below the check whether to kill > > >> the current task. That would solve your problem, and I don't _think_ > > >> it would cause any new ones. Michal, you touched this code last, what > > >> do you think? > > > > > > What do you mean exactly? Whether we kill a current task or something > > > else doesn't change much on the fact that NOFS is a reclaim restricted > > > context and we might kill too early. If the fs can do GFP_FS then it is > > > obviously a better thing to do because FS metadata can be reclaimed as > > > well and therefore there is potentially less memory pressure on > > > application data. > > > > > > > I interpreted "to move the check for GFP_NOFS in out_of_memory() down to > > below the check whether to kill the current task" as > > Too far; I meant one line earlier, before we try to select a different > process. We could still panic the system on pre-mature OOM. So it doesn't really seem good. > > @@ -1104,6 +1095,19 @@ bool out_of_memory(struct oom_control *oc) > > } > > > > select_bad_process(oc); > > + > > + /* > > + * The OOM killer does not compensate for IO-less reclaim. > > + * pagefault_out_of_memory lost its gfp context so we have to > > + * make sure exclude 0 mask - all other users should have at least > > + * ___GFP_DIRECT_RECLAIM to get here. > > + */ > > + if ((oc->gfp_mask && !(oc->gfp_mask & __GFP_FS)) && oc->chosen && > > + oc->chosen != (void *)-1UL && oc->chosen != current) { > > + put_task_struct(oc->chosen); > > + return true; > > + } > > + > > /* Found nothing?!?! */ > > if (!oc->chosen) { > > dump_header(oc, NULL); > > > > which is prefixed by "the correct fix is not". > > > > Behaving like sysctl_oom_kill_allocating_task == 1 if __GFP_FS is not used > > will not be the correct fix. But ... > > > > Hou Tao wrote: > > > There is no need to disable __GFP_FS in ->readpage: > > > * It's a read-only fs, so there will be no dirty/writeback page and > > > there will be no deadlock against the caller's locked page > > > > is read-only filesystem sufficient for safe to use __GFP_FS? > > > > Isn't "whether it is safe to use __GFP_FS" depends on "whether fs locks > > are held or not" rather than "whether fs has dirty/writeback page or not" ? > > It's worth noticing that squashfs _is_ in fact holding a page locked in > squashfs_copy_cache() when it calls grab_cache_page_nowait(). I'm not > sure if this will lead to trouble or not because I'm insufficiently > familiar with the reclaim path. Hmm, this is more interesting then. If there is any memcg accounted allocation down that path _and_ the squashfs writeout can lock more pages and mark them writeback before they are really sent to the storage then we have a problem. See [1] [1] http://lkml.kernel.org/r/20181213092221.27270-1-mhocko@xxxxxxxxxx -- Michal Hocko SUSE Labs