On 7/4/19 4:09 AM, Michal Hocko wrote: > On Wed 03-07-19 16:54:35, Mike Kravetz wrote: >> On 7/3/19 2:43 AM, Mel Gorman wrote: >>> Indeed. I'm getting knocked offline shortly so I didn't give this the >>> time it deserves but it appears that part of this problem is >>> hugetlb-specific when one node is full and can enter into this continual >>> loop due to __GFP_RETRY_MAYFAIL requiring both nr_reclaimed and >>> nr_scanned to be zero. >> >> Yes, I am not aware of any other large order allocations consistently made >> with __GFP_RETRY_MAYFAIL. But, I did not look too closely. Michal believes >> that hugetlb pages allocations should use __GFP_RETRY_MAYFAIL. > > Yes. The argument is that this is controlable by an admin and failures > should be prevented as much as possible. I didn't get to understand > should_continue_reclaim part of the problem but I have a strong feeling > that __GFP_RETRY_MAYFAIL handling at that layer is not correct. What > happens if it is simply removed and we rely only on the retry mechanism > from the page allocator instead? Does the success rate is reduced > considerably? It certainly will be reduced. I 'think' it will be hard to predict how much it will be reduced as this will depend on the state of memory usage and fragmentation at the time of the attempt. I can try to measure this, but I will be a few days due to U.S. holiday. -- Mike Kravetz
