[Intel-gfx] [PATCH 03/52] drm: add managed resources tied to drm_device
Emil Velikov
emil.l.velikov at gmail.com
Wed Feb 19 16:41:54 UTC 2020
On Wed, 19 Feb 2020 at 16:22, Daniel Vetter <daniel.vetter at ffwll.ch> wrote:
>
> On Wed, Feb 19, 2020 at 5:09 PM Emil Velikov <emil.l.velikov at gmail.com> wrote:
> >
> > On Wed, 19 Feb 2020 at 14:23, Daniel Vetter <daniel.vetter at ffwll.ch> wrote:
> > >
> > > On Wed, Feb 19, 2020 at 2:33 PM Greg Kroah-Hartman
> > > <gregkh at linuxfoundation.org> wrote:
> > > >
> > > > On Wed, Feb 19, 2020 at 03:28:47PM +0200, Laurent Pinchart wrote:
> > > > > Hi Daniel,
> > > > >
> > > > > Thank you for the patch.
> > > > >
> > > > > On Wed, Feb 19, 2020 at 11:20:33AM +0100, Daniel Vetter wrote:
> > > > > > We have lots of these. And the cleanup code tends to be of dubious
> > > > > > quality. The biggest wrong pattern is that developers use devm_, which
> > > > > > ties the release action to the underlying struct device, whereas
> > > > > > all the userspace visible stuff attached to a drm_device can long
> > > > > > outlive that one (e.g. after a hotunplug while userspace has open
> > > > > > files and mmap'ed buffers). Give people what they want, but with more
> > > > > > correctness.
> > > > > >
> > > > > > Mostly copied from devres.c, with types adjusted to fit drm_device and
> > > > > > a few simplifications - I didn't (yet) copy over everything. Since
> > > > > > the types don't match code sharing looked like a hopeless endeavour.
> > > > > >
> > > > > > For now it's only super simplified, no groups, you can't remove
> > > > > > actions (but kfree exists, we'll need that soon). Plus all specific to
> > > > > > drm_device ofc, including the logging. Which I didn't bother to make
> > > > > > compile-time optional, since none of the other drm logging is compile
> > > > > > time optional either.
> > > > > >
> > > > > > One tricky bit here is the chicken&egg between allocating your
> > > > > > drm_device structure and initiliazing it with drm_dev_init. For
> > > > > > perfect onion unwinding we'd need to have the action to kfree the
> > > > > > allocation registered before drm_dev_init registers any of its own
> > > > > > release handlers. But drm_dev_init doesn't know where exactly the
> > > > > > drm_device is emebedded into the overall structure, and by the time it
> > > > > > returns it'll all be too late. And forcing drivers to be able clean up
> > > > > > everything except the one kzalloc is silly.
> > > > > >
> > > > > > Work around this by having a very special final_kfree pointer. This
> > > > > > also avoids troubles with the list head possibly disappearing from
> > > > > > underneath us when we release all resources attached to the
> > > > > > drm_device.
> > > > >
> > > > > This is all a very good idea ! Many subsystems are plagged by drivers
> > > > > using devm_k*alloc to allocate data accessible by userspace. Since the
> > > > > introduction of devm_*, we've likely reduced the number of memory leaks,
> > > > > but I'm pretty sure we've increased the risk of crashes as I've seen
> > > > > some drivers that used .release() callbacks correctly being naively
> > > > > converted to incorrect devm_* usage :-(
> > > > >
> > > > > This leads me to a question: if other subsystems have the same problem,
> > > > > could we turn this implementation into something more generic ? It
> > > > > doesn't have to be done right away and shouldn't block merging this
> > > > > series, but I think it would be very useful.
> > > >
> > > > It shouldn't be that hard to tie this into a drv_m() type of a thing
> > > > (driver_memory?)
> > > >
> > > > And yes, I think it's much better than devm_* for the obvious reasons of
> > > > this being needed here.
> > >
> > > There's two reasons I went with copypasta instead of trying to share code:
> > > - Type checking, I definitely don't want people to mix up devm_ with
> > > drmm_. But even if we do a drv_m that subsystems could embed we do
> > > have quite a few different types of component drivers (and with
> > > drm_panel and drm_bridge even standardized), and I don't want people
> > > to be able to pass the wrong kind of struct to e.g. a managed
> > > drmm_connector_init - it really needs to be the drm_device, not a
> > > panel or bridge or something else.
> > >
> > > - We could still share the code as a kind of implementation/backend
> > > library. But it's not much, and with embedding I could use the drm
> > > device logging stuff which is kinda nice. But if there's more demand
> > > for this I can definitely see the point in sharing this, as Laurent
> > > pointed out with the tiny optimization with not allocating a NULL void
> > > * that I've done (and screwed up) it's not entirely trivial code.
> > >
> >
> > My 2c as they say, although closer to a brain dump :-)
> >
> > On one hand the drm_device has an embedded struct device. On the other
> > drm_device preserves state which outlives the embedded struct device.
> >
> > Would it make sense to keep drm_device better related to the
> > underlying device? Effectively moving the $misc state to drm_driver.
> > This idea does raise another question - struct drm_driver unlike many
> > other struct $foo_driver, does not embedded device_driver :-(
> > So if one is to cover the above two, then the embedding concerns will
> > be elevated.
>
> drm_driver isn't a bus device driver in the linux driver model sense,
> but an uapi thing that sits on top of some underlying device. So maybe
> better to rename drm_driver to drm_interface_driver, and drm_device to
> drm_interface. But that would be giantic churn and probably lots of
> confusion. We do require a link between drm_device->struct device
> nowadays, but that's just to guarantee userspace can find the
> drm_device in sysfs somewhere and make sense of what it actually
> drives.
>
> That's also why the lifetimes for the two things are totally
> different. The device driver an all it's resources are tied to the
> underlying physical device, and resources can be released when that
> driver<->device link is broken (either unbind or hotunplug). That's
> what devm_ does. The drm_driver/drm_device otoh is tied to the
> userspace api, and can only disappear once all the userspace handles
> have been cleaned up and released. And we have an enormous amount of
> those, with all the mmaps, and shared fd for dma-buf, sync_file,
> synobj and whatever else. The drm_device can only be cleaned up once
> userspace has closed all these things, or we'll go boom somewhere. The
> only connection is that the userspace interface drives the underlying
> hw (as long as it's still there) and the hw side holds a reference on
> the uapi side (drm_dev_get/put) to make sure the userspace side
> doesn't go poof and disappear when no one has the /dev node open :-)
>
> But aside from these links they're completely separate worlds, and
> mixing up the lifetimes results in all kinds of bad things happening.
> Ofc normally these two things exist at the same time, but hotunplug
> makes things very interesting here. And traditionally we've handled it
> badly, if at all in drm.
>
Seems like my drm_device/drm_driver definitions were off.
Thanks a lot for clarifying.
-Emil
More information about the dri-devel
mailing list