Michal Hocko wrote: > On Fri 20-03-15 23:02:09, Tetsuo Handa wrote: > > Michal Hocko wrote: > > > On Fri 20-03-15 22:34:21, Tetsuo Handa wrote: > > > > Huang Ying wrote: > > > > > > > BTW: the test is run on 32 bit system. > > > > > > > > > > > > That sounds like the cause of your problem. The system might be out > > > > > > of > > > > > > address space available for the kernel (only 1GB if x86_32). You > > > > > > should > > > > > > try running tests on 64 bit systems. > > > > > > > > > > We run test on 32 bit and 64 bit systems. Try to catch problems on > > > > > both > > > > > platforms. I think we still need to support 32 bit systems? > > > > > > > > Yes, testing on both platforms is good. But please read > > > > http://lwn.net/Articles/627419/ , http://lwn.net/Articles/635354/ and > > > > http://lwn.net/Articles/636017/ . Then please add __GFP_NORETRY to > > > > memory > > > > allocations in btrfs code if it is appropriate. > > > > > > I guess you meant __GFP_NOFAIL? > > > > > No. btrfs's selftest (which is not using __GFP_NOFAIL) is already looping > > forever. If we want to avoid btrfs's selftest from looping forever, btrfs > > needs __GFP_NORETRY than __GFP_NOFAIL (until we establish a way to safely > > allow small allocations to fail). > > Sigh. If the code is using GFP_NOFS allocation (which seem to be the > case because it worked with the 9879de7373fc) and the proper fix for > this IMO is to simply not retry endlessly for these allocations.
We can avoid looping forever by passing __GFP_NORETRY (from the caller side) or by using sysctl_nr_alloc_retry == 1 (from the callee side). But > We > have to sort some other issues before we can make NOFS allocations fail > but let's not pile more workarounds on top in the meantime. But if btrfs > people really think __GFP_NORETRY then I do not really care much. https://lkml.org/lkml/2015/3/19/221 suggests that changing each caller to use either __GFP_NOFAIL or __GFP_NORETRY is the safer way to allow small allocations to fail than using sysctl_nr_alloc_retry, for we don't want to add __GFP_NOFAIL to allocations by page fault. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
