[Intel-gfx] [PATCH 03/52] drm: add managed resources tied to drm_device
Daniel Vetter
daniel.vetter at ffwll.ch
Wed Feb 19 16:53:59 UTC 2020
On Wed, Feb 19, 2020 at 5:46 PM Laurent Pinchart
<laurent.pinchart at ideasonboard.com> wrote:
>
> Hi Daniel,
>
> On Wed, Feb 19, 2020 at 05:22:38PM +0100, Daniel Vetter wrote:
> > On Wed, Feb 19, 2020 at 5:09 PM Emil Velikov wrote:
> > > On Wed, 19 Feb 2020 at 14:23, Daniel Vetter wrote:
> > >> On Wed, Feb 19, 2020 at 2:33 PM Greg Kroah-Hartman wrote:
> > >>> On Wed, Feb 19, 2020 at 03:28:47PM +0200, Laurent Pinchart wrote:
> > >>>> 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.
>
> If we wanted to rename drm_driver to align with the rest of the kernel,
> it should probably be drm_device_ops, with the non-ops fields being
> moved to a separate structure.
>
> I don't mind churn (but I agree it may not be worth it), but even if we
> don't rename the structure, I think it would be very useful to remove
> the non-const fields, in order to allow storing the structure as a
> global static const struct. Function pointers in non-const memory can be
> a security issue. As far as I can tell, the only blocker is the
> legacy_dev_list field.
Oh man ... we could make the legacy_dev_list depend on
CONFIG_DRM_LEGACY and the INIT_LIST_HEAD also depend upon
DRIVER_LEGACY and then at least all the new drivers could make their
drm_driver structure const. Or something along those lines.
Properly ditching legacy_dev_list is probably not worth it, since
those drivers tend to be all root exploits anyway :-)
Cheers, Daniel
> > 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 so they're tied to the lifetime of the struct device that models the
> userspace interface. Shame they're both called device :-)
>
> > 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.
> >
> > > WRT type safety, with the embedded work sorted, one could introduce
> > > trivial helpers for drmm_connector_init and friends.
> > >
> > > In another email you've also raised the question of API diversity and
> > > reviews, I believe. IMHO one could start with a bare minimum set and
> > > extend as needed.
> > > Based on the prompt response from Greg, I suspect review won't be an issue.
> >
> > The drmm_ stuff in here is the bare minimum we need to get started. I
> > expect lots of stuff will be added, but those are all just going to be
> > convenience functions on top of the drmm_add_action primitive.
> >
> > > If people agree with my analysis and considering the size/complexity
> > > of drm_device <> drm_driver reshuffle, we could add a TODO task.
> > > I suspect the underlying work will be larger than the current 52 patch
> > > set, so doing it in one go will be PITA.
> >
> > I'm not following what you want to shuffle. drm_driver is entirely
> > static and kinda global, drm_device is the per-instance structure we
> > have. And here we mean per-userspace uapi interface instance. So I
> > guess I'm confused what you want to do?
> >
> > > * Based on the following quick greps
> > > $git grep -W "struct [a-zA-Z0-9-]*_driver {" -- include/ | grep -w
> > > "struct device_driver\>.*;" | wc -l
> > > 56
> > > $git cgrep "struct [a-zA-Z0-9-]*_driver {" -- include/ | wc -l
> > > 71
>
> --
> Regards,
>
> Laurent Pinchart
--
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
More information about the Intel-gfx
mailing list