On Mon 30-07-18 18:54:24, Georgi Nikolov wrote: [...] > No i was wrong. The regression starts actually with 0537250fdc6c8. > - old code, which opencodes kvmalloc, is masking error but error is there > - kvmalloc without GFP_NORETRY works fine, but probably can consume a > lot of memory - commit: eacd86ca3b036 > - kvmalloc with GFP_NORETRY shows error - commit: 0537250fdc6c8 OK. > >> What is correct way to fix it. > >> - inside xt_alloc_table_info remove GFP_NORETRY from kvmalloc or add > >> this flag only for sizes bigger than some threshold > > This would reintroduce issue fixed by 0537250fdc6c8. Note that > > kvmalloc(GFP_KERNEL | __GFP_NORETRY) is more or less equivalent to the > > original code (well, except for __GFP_NOWARN). > > So probably we should pass GFP_NORETRY only for large requests (above > some threshold). What would be the treshold? This is not really my area so I just wanted to keep the original code semantic. > >> - inside kvmalloc_node remove GFP_NORETRY from > >> __vmalloc_node_flags_caller (i don't know if it honors this flag, or > >> the problem is elsewhere) > > No, not really. This is basically equivalent to kvmalloc(GFP_KERNEL). > > > > I strongly suspect that this is not a regression in this code but rather > > a side effect of larger memory fragmentation caused by something else. > > In any case do you see this failure also without artificial test case > > with a standard workload? > > Yes i can see failures with standard workload, in fact it was hard to > reproduce it. > Here is the error from production servers where allocation is smaller: > iptables: vmalloc: allocation failure, allocated 131072 of 225280 bytes, > mode:0x14010c0(GFP_KERNEL|__GFP_NORETRY), nodemask=(null) > > I didn't understand if vmalloc honors GFP_NORETRY. 0537250fdc6c8 changelog tries to explain. kvmalloc doesn't really support the GFP_NORETRY remantic because that would imply the request wouldn't trigger the oom killer but in rare cases this might happen (e.g. when page tables are allocated because those are hardcoded GFP_KERNEL). That being said, I have no objection to use GFP_KERNEL if it helps real workloads but we probably need some cap... -- Michal Hocko SUSE Labs
