> -----Original Message----- > From: Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> > Sent: Tuesday, July 28, 2020 9:28 AM > To: Daniel Vetter <daniel@xxxxxxxx>; Dave Airlie <airlied@xxxxxxxxxx> > Cc: intel-gfx <intel-gfx@xxxxxxxxxxxxxxxxxxxxx>; stable > <stable@xxxxxxxxxxxxxxx>; Gustavo Padovan > <gustavo.padovan@xxxxxxxxxxxxx>; Tang, CQ <cq.tang@xxxxxxxxx>; dri- > devel <dri-devel@xxxxxxxxxxxxxxxxxxxxx>; Vetter, Daniel > <daniel.vetter@xxxxxxxxx> > Subject: Re: [PATCH 1/3] drm: Restore driver.preclose() for all to use > > Quoting Daniel Vetter (2020-07-27 20:32:45) > > On Thu, Jul 23, 2020 at 7:21 PM Chris Wilson <chris@xxxxxxxxxxxxxxxxxx> > wrote: > > > > > > An unfortunate sequence of events, but it turns out there is a valid > > > usecase for being able to free/decouple the driver objects before > > > they are freed by the DRM core. In particular, if we have a pointer > > > into a drm core object from inside a driver object, that pointer > > > needs to be nerfed *before* it is freed so that concurrent access > > > (e.g. debugfs) does not following the dangling pointer. > > > > > > The legacy marker was adding in the code movement from drp_fops.c to > > > drm_file.c > > > > I might fumble a lot, but not this one: > > > > commit 45c3d213a400c952ab7119f394c5293bb6877e6b > > Author: Daniel Vetter <daniel.vetter@xxxxxxxx> > > Date: Mon May 8 10:26:33 2017 +0200 > > > > drm: Nerf the preclose callback for modern drivers > > Gah, when I going through the history it looked like it appeared out of > nowhere. > > > Also looking at the debugfs hook that has some rather adventurous > > stuff going on I think, feels a bit like a kitchensink with batteries > > included. If that's really all needed I'd say iterate the contexts by > > first going over files, then the ctx (which arent shared anyway) and > > the problem should also be gone. > > Or we could cut out the middlelayer and put the release under the driver > control with a call to the drm_release() when the driver is ready. Chiris, can explain this idea, or post a patch ? --CQ > -Chris
