Hi, small GFP_NOFS, like GFP_KERNEL, allocations have not been not failing traditionally even though their reclaim capabilities are restricted because the VM code cannot recurse into filesystems to clean dirty pages. At the same time these allocation requests do not allow to trigger the OOM killer because that would lead to pre-mature OOM killing during heavy fs metadata workloads. This leaves the VM code in an unfortunate situation where GFP_NOFS requests is looping inside the allocator relying on somebody else to make a progress on its behalf. This is prone to deadlocks when the request is holding resources which are necessary for other task to make a progress and release memory (e.g. OOM victim is blocked on the lock held by the NONFS request). Another drawback is that the caller of the allocator cannot define any fallback strategy because the request doesn't fail. As the VM cannot do much about these requests we should face the reality and allow those allocations to fail. Johannes has already posted the patch which does that (http://marc.info/?l=linux-mm&m=142726428514236&w=2) but the discussion died pretty quickly. I was playing with this patch and xfs, ext[34] and btrfs for a while to see what is the effect under heavy memory pressure. As expected this led to some fallouts. My test consisted of a simple memory hog which allocates a lot of anonymous memory and writes to a fs mainly to trigger a fs activity on exit. In parallel there is a parallel fs metadata load (multiple tasks creating thousands of empty files and directories). All is running in a VM with small amount of memory to emulate an under provisioned system. The metadata load is triggering a sufficient load to invoke the direct reclaim even without the memory hog. The memory hog forks several tasks sharing the VM and OOM killer manages to kill it without locking up the system (this was based on the test case from Tetsuo Handa - http://www.spinics.net/lists/linux-fsdevel/msg82958.html - I just didn't want to kill my machine ;)). With all the patches applied none of the 4 filesystems gets aborted transactions and RO remount (well xfs didn't need any special treatment). This is obviously not sufficient to claim that failing GFP_NOFS is OK now but I think it is a good start for the further discussion. I would be grateful if FS people could have a look at those patches. I have simply used __GFP_NOFAIL in the critical paths. This might be not the best strategy but it sounds like a good first step. The first patch in the series also allows __GFP_NOFAIL allocations to access memory reserves when the system is OOM which should help those requests to make a forward progress - especially in combination with GFP_NOFS. The second patch tries to address a potential pre-mature OOM killer from the page fault path. I have posted it separately but it didn't get much traction. The third patch allows GFP_NOFS to fail and I believe it should see much more testing coverage. It would be really great if it could sit in the mmotm tree for few release cycles so that we can catch more fallouts. The rest are the FS specific patches to fortify allocations requests which are really needed to finish transactions without RO remounts. There might be more needed but my test case survives with these in place. They would obviously need some rewording if they are going to be applied even without Patch3 and I will do that if respective maintainers will take them. Ext3 and JBD are going away soon so they might be dropped but they have been in the tree while I was testing so I've kept them. Thoughts? Opinions? -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
