On Fri, Dec 11, 2020 at 09:05:26AM -0700, Jens Axboe wrote: > On 12/10/20 7:45 PM, Al Viro wrote: > > So how hard are your "we don't want to block here" requirements? Because > > the stuff you do after complete_walk() can easily be much longer than > > everything else. > > Ideally it'd extend a bit beyond the RCU lookup, as things like proc > resolution will still fail with the proposed patch. But that's not a > huge deal to me, I consider the dentry lookup to be Good Enough. FWIW, /proc/$pid always falls back to REF walks. Here's a patch from one of my colleagues that aims to fix that. https://lore.kernel.org/linux-fsdevel/20201204000212.773032-1-stephen.s.brennan@xxxxxxxxxx/ Maybe you had one of the other parts of /proc in mind?
