On 5/29/21 1:33 AM, Andres Freund wrote: > Hi, > > I started to see buffer registration randomly failing with ENOMEM on > 5.13. Registering buffer or two often succeeds, but more than that > rarely. Running the same program as root succeeds - but the user has a high > rlimit. > > The issue is that io_sqe_buffer_register() doesn't initialize > imu. io_buffer_account_pin() does imu->acct_pages++, before calling > io_account_mem(ctx, imu->acct_pages); > > Which means that a random amount of memory is being accounted for. On the first > few allocations this sometimes fails to fail because the memory is zero, but > after a bit of reuse... Makes sense, thanks for digging in. I've just sent a patch, would be great if you can test it or send your own. > It only doesn't fail as root because the rlimit doesn't apply. > > This is caused by > > commit 41edf1a5ec967bf4bddedb83c48e02dfea8315b4 > Author: Pavel Begunkov <asml.silence@xxxxxxxxx> > Date: 2021-04-25 14:32:23 +0100 > > io_uring: keep table of pointers to ubufs > > Greetings, > > Andres Freund > -- Pavel Begunkov