On Mon, Jan 7, 2019 at 10:05 PM Nikolaus Rath <Nikolaus@xxxxxxxx> wrote: > > On Jan 07 2019, Miklos Szeredi <miklos@xxxxxxxxxx> wrote: > > On Wed, Dec 26, 2018 at 10:44 PM Nikolaus Rath <Nikolaus@xxxxxxxx> wrote: > >> > >> Hi, > >> > >> I am seeing relatively regular occurences of > >> > >> $ sudo dmesg | tail > >> [21929.138815] fuse: trying to steal weird page > >> [21929.138821] page=00000000a7dd2617 index=64 flags=17fffc0000000ad, > >> count=1, mapcount=0, mapping= (null) > >> [21930.647338] fuse: trying to steal weird page > >> [21930.647345] page=00000000a07f32af index=2848 > >> flags=17fffc0000000ad, count=1, mapcount=0, mapping= (null) > >> [21932.338873] fuse: trying to steal weird page > >> [21932.338879] page=0000000067e3a012 index=64 flags=17fffc0000000ad, > >> count=1, mapcount=0, mapping= (null) > >> [21933.930703] fuse: trying to steal weird page > >> [21933.930710] page=00000000046feb25 index=845 > >> flags=17fffc0000000ad, count=1, mapcount=0, mapping= (null) > >> [21936.163174] fuse: trying to steal weird page > >> [21936.163180] page=00000000fb80fe27 index=0 flags=17fffc0000000ad, > >> count=1, mapcount=0, mapping= (null) > > > > The page has the PG_dity and PG_waiters flags set which are > > incompatible with stealing. page_cache_pipe_buf_steal() does > > apparently filter out dirty ones, so it's not a regular file that we > > are trying to streal the page from. So the question is: what is the > > source of the splice()? > > Hmm. I think it has to be a regular file. But as I mentioned in my other > email, I did have a race condition where fd's were closed > incorrectly. Is it possible that this also triggered the above, > i.e. that the fd was closed sometime during splice? Close during a syscall that uses the fd is not an issue, because a ref to the file is acquired. So the race is between the close() and the internal fget(); if the close() wins then fget() will fail and the syscall will return EBADF. If the fget() wins, then the syscall can run normally despite the fact that the fd was closed. Can you tell me what filesystem is the regular file (the one being spliced into fuse) is on? It actually has to be a regular file, since AFAIK nothing else has dirty pages. It could be using something other than page_cache_pipe_buf_steal(), or there's some other mechanism that lets the page be dirtied after being unmapped, though that looks impossible... Thanks, Miklos
