On Mon 18-06-18 18:11:26, Mikulas Patocka wrote: [...] > I grepped the kernel for __GFP_NORETRY and triaged them. I found 16 cases > without a fallback - those are bugs that make various functions randomly > return -ENOMEM. Well, maybe those are just optimistic attempts to allocate memory and have a fallback somewhere else. So I would be careful calling them outright bugs. But maybe you are right. > Most of the callers provide callback. > > There is another strange flag - __GFP_RETRY_MAYFAIL - it provides two > different functions - if the allocation is larger than > PAGE_ALLOC_COSTLY_ORDER, it retries the allocation as if it were smaller. > If the allocations is smaller than PAGE_ALLOC_COSTLY_ORDER, > __GFP_RETRY_MAYFAIL will avoid the oom killer (larger order allocations > don't trigger the oom killer at all). Well, the primary purpose of this flag is to provide a consistent failure behavior for all requests regardless of the size. > So, perhaps __GFP_RETRY_MAYFAIL could be used instead of __GFP_NORETRY in > the cases where the caller wants to avoid trigerring the oom killer (the > problem is that __GFP_NORETRY causes random failure even in no-oom > situations but __GFP_RETRY_MAYFAIL doesn't). myabe yes. > So my suggestion is - fix these obvious bugs when someone allocates memory > with __GFP_NORETRY without any fallback - and then, __GFP_NORETRY could be > just changed to return NULL instead of sleeping. No real objection to fixing wrong __GFP_NORETRY usage. But __GFP_NORETRY can sleep. Nothing will really change in that regards. It does a reclaim and that _might_ sleep. But seriously, isn't the best way around the throttling issue to use PF_LESS_THROTTLE? -- Michal Hocko SUSE Labs
