On Wed, 7 Jul 2021 at 22:25, NeilBrown <neilb@xxxxxxx> wrote: > > On Wed, 07 Jul 2021, Daire Byrne wrote: > > I stripped out all my patches so it's just this one on top of 5.13-rc7 > > and I can still reproduce it. > > > > I can only trigger it by mounting the same export (RHEL7 server) using > > two different namespaces and performing a heavy IO benchmark to either > > mount (leaving one idle). Part of the benchmark walks thousands of > > dirs with files (hence the readdirs). > > > > If I mount the same server twice with no (same) namespaces, even with > > the patch applied, it works fine without any crash. > > That's pretty solid evidence! > > I just realized that the stack trace you reported mentions > "kfree_const()". > My latest patch doesn't include that, and nfs doesn't use it at all. > Might you still be using the older patch? > > NeilBrown Oh... the last stack trace, the readdir one? I don't see kfree_const myself but I may have a case of word blindness. The first one I reported definitely has kfree_const but after your latest patch, this last one around readdir doesn't seem to? I'm pretty sure I have your latest patch (with kfree instead of kfree_const) correctly applied. Though, I will double check that the correct kernel and modules were then installed properly on my test VM. Daire
