Re: System freezes after OOM

[Michal Hocko <mhocko@kernel.org>]

On Thu 14-07-16 19:07:52, Ondrej Kozina wrote:
> On 07/14/2016 05:31 PM, Michal Hocko wrote:
> > On Thu 14-07-16 16:08:28, Ondrej Kozina wrote:
> > [...]
> > > As Mikulas pointed out, this doesn't work. The system froze as well with the
> > > patch above. Will try to tweak the patch with Mikulas's suggestion...
> > 
> > Thank you for testing! Do you happen to have traces of the frozen
> > processes? Does the flusher still gets throttled because the bias it
> > gets is not sufficient. Or does it get throttled at a different place?
> > 
> 
> Sure. Here it is (including sysrq+t and sysrq+w output): https://okozina.fedorapeople.org/bugs/swap_on_dmcrypt/4.7.0-rc7+/1/4.7.0-rc7+.log

Thanks a lot! This is helpful.
[  162.716376] active_anon:107874 inactive_anon:108176 isolated_anon:64
[  162.716376]  active_file:1086 inactive_file:1103 isolated_file:0
[  162.716376]  unevictable:0 dirty:0 writeback:69824 unstable:0
[  162.716376]  slab_reclaimable:3119 slab_unreclaimable:24124
[  162.716376]  mapped:2165 shmem:57 pagetables:1509 bounce:0
[  162.716376]  free:701 free_pcp:0 free_cma:0

No surprise that PF_LESS_THROTTLE didn't help. It gives some bias but
considering how many pages are under writeback it cannot possibly help
to prevent from sleeping in throttle_vm_writeout. I suppose adding
the following on top of the memalloc patch helps, right?
It is an alternative to what you were suggesting in other email but
it doesn't affect current_may_throttle paths which I would rather not
touch.
---
diff --git a/mm/page-writeback.c b/mm/page-writeback.c
index 7fbb2d008078..a37661f1a11b 100644
--- a/mm/page-writeback.c
+++ b/mm/page-writeback.c
@@ -1971,6 +1971,9 @@ void throttle_vm_writeout(gfp_t gfp_mask)
 	unsigned long background_thresh;
 	unsigned long dirty_thresh;
 
+	if (current->flags & PF_LESS_THROTTLE)
+		return;
+
         for ( ; ; ) {
 		global_dirty_limits(&background_thresh, &dirty_thresh);
 		dirty_thresh = hard_dirty_limit(&global_wb_domain, dirty_thresh);
-- 
Michal Hocko
SUSE Labs