On Tue, Aug 20, 2019 at 05:59:49PM +0200, Oleg Nesterov wrote: > On 08/19, Andrea Arcangeli wrote: > > > > The proposed fix looks correct, can you resend in a way that can be merged? > > OK, I'll send the same patch to lkml, the only change is s/xxx/still_valid/. Thanks! Actually I wasn't sure if I should send it myself to avoid delaying it to next week, but I see you already sent it so problem solved. > > It's a bit strange that the file that > > was opened by the ioctl() syscall gets released > > and this look like another bug we need to investigate, I did some more debugging in the meanwhile. The current theory is there are multiple uffd in the same mm and the uffd ctx of the page fault is not the same uffd ctx of the ioctl that triggers the copy user. I'll need to add some more bpftrace code to be sure. Thanks, Andrea
