Hi guys,
Above all, yes the current approach looks a bit funky.
Given the constrains (cannot use ioctl and libudev) it's rather reasonable.
That said, ideas for improvements are always welcome.
On 26 June 2018 at 13:03, Mark Kettenis <mark.kettenis@xxxxxxxxx> wrote:
>> Date: Tue, 26 Jun 2018 20:58:18 +1000
>> From: Jonathan Gray <jsg@xxxxxxxxx>
>>
>> On Tue, Jun 26, 2018 at 11:38:20AM +0100, Emil Velikov wrote:
>> > On 21 June 2018 at 16:32, Jonathan Gray <jsg@xxxxxxxxx> wrote:
>> > > On Thu, Jun 21, 2018 at 03:24:49PM +0100, Emil Velikov wrote:
>> > >> Hi Jonathan,
>> > >>
>> > >> On 1 December 2016 at 04:18, Jonathan Gray <jsg@xxxxxxxxx> wrote:
>> > >>
>> > >> > --- a/xf86drm.c
>> > >> > +++ b/xf86drm.c
>> > >> > @@ -3248,6 +3248,67 @@ drm_device_validate_flags(uint32_t flags)
>> > >> > */
>> > >> > int drmGetDevice2(int fd, uint32_t flags, drmDevicePtr *device)
>> > >> > {
>> > >> > +#ifdef __OpenBSD__
>> > >> > + /*
>> > >> > + * DRI device nodes on OpenBSD are not in their own directory, they reside
>> > >> > + * in /dev along with a large number of statically generated /dev nodes.
>> > >> > + * Avoid stat'ing all of /dev needlessly by implementing this custom path.
>> > >> > + */
>> > >> I was in the area, fixing the odd bug and doing some cleanups and a
>> > >> question came to mind:
>> > >>
>> > >> Is there some obstacle of placing the drm nodes under /dev/dri/? It
>> > >> would involve a check/update through the OpenBSD tree, yet no obvious
>> > >> downsides comes to mind.
>> > >> I think it would make things more distinct and obvious. IIRC the
>> > >> OpenBSD xserver does some checking which /dev node opened, the
>> > >> suggestion should help there.
>> > >>
>> > >> What do you think?
>> > >> Emil
>> > >
>> > > There are no other devices under a sub-directory besides /dev/fd/.
>> > > I don't think anyone is going to be onboard with changing things for
>> > > drm nodes. Though drm device nodes names will have to be revisted
>> > > when/if render nodes etc get supported. drmR drmC etc have not
>> > > been proposed yet.
>> >
>> > I see, that's enlighening.
>> >
>> > Out of curiosity: personally, do you see any technical issues with a
>> > sub-directory approach?
>>
>> I get that you want a single path but it seems these functions were
>> really designed assuming the approach was going to be a sub-directory.
Right, I should have said "Ignoring the actual implementation, do you
see any technical issues..."
> The bigger picture here is that this code is primarily used to query
> for information about an open drm file descriptor. The Linux
> implementation does a lot of filesystem groveling to achieve that.
> Especially the bits that descend into /sys are never going to work on
> OpenBSD. If a more OS-agnostic approach is wanted, I would say an
> ioctl to return the relevant information would be more appropriate.
> This is actually what we did for OpenBSD where we implemented
> DRM_IOCTL_GET_PCIINFO to implement drmParsePciBusInfo.
An ioctl approach has two issues:
- linux aims to keep the uabi 'forever' as such adding extra ioctl's
is fairly hard
This particular instance was discussed and rejected with linux devs.
- on linux the device (even the bus it's on) can be powered off
Issuing an ioctl will wake up the device (which can be slow) even if
you don't end up using that device.
Admittedly you'd want the power-off machinery in !linux at some point.
GPUs power hungry beasts, even when they're not used ;-)
As I mentioned libudev (yes I know you're not using it), it's also a
no-go since apps ship with their own version of it, causing all sorts
of grief.
We tried that in Mesa and after over a year of various fixes, I nuked
it in favour of drmDevice.
> Such an
> approach avoids the issue of mapping the file descriptor back to a
> file system path. Also note that mapping and open file descriptor to
> a file system path is fundamentally impossible on Unix without
> cheating. Although cheating is fairly easy for devices.
>
Agreed, I've skimmed through the code of lsof and ouch...
> If mapping a file descriptor back to a filesystem path is necessary,
> OpenBSD actually implements devname(3) which uses a database to map
> device major/minor pairs back to a device name. This is actually a
> function that was introduced in 4.4BSD, and as far as I can tell it is
> still present in FreeBSD and NetBSD as well. We could change the
> OpenBSD implementation of drmGetDevice2() to use that, but it wouldn't
> really make the code simpler.
>
> All in all, I think it is best to keep the Linux and OpenBSD code
> separate here.
While I can see the current approach seems foobar, my question was orthogonal.
Namely: is there a thread/documentation covering the technical reasons
behind the lack of sub-directories in /dev/?
Thanks
Emil
_______________________________________________
dri-devel mailing list
dri-devel@xxxxxxxxxxxxxxxxxxxxx
https://lists.freedesktop.org/mailman/listinfo/dri-devel
