i915 render node discovery buggy?

i915 render node discovery buggy?

[Russell King - ARM Linux]

In passing, while reading the Intel DDX driver code, I've noticed
that the Intel driver contains code which assumes that the master and
render devices are related by minor device number, eg:

        /* Are we a render-node ourselves? */
        if (is_render_node(fd, &master))
                return NULL;

        sprintf(buf, "/dev/dri/renderD%d", (int)((master.st_rdev | 0x80) & 0xbf));
        if (stat(buf, &render) == 0 &&
            master.st_mode == render.st_mode &&
            render.st_rdev == ((master.st_rdev | 0x80) & 0xbf))
                return strdup(buf);

There's also code doing the reverse as well.

From my observations, the assumption that this code is built upon is
false.  I have an ARM platform here (non-Intel graphics) which shows
the problem - we have a KMS-only DRM driver (card0) and a GPU-only
DRM driver (card1).  This populates the /dev/dri subdirectory as
follows:

crw-rw----+ 1 root video 226,   0 Oct 27 04:59 card0
crw-rw----+ 1 root video 226,   1 Oct 26 20:40 card1
crw-rw----  1 root video 226,  64 Oct 26 20:40 controlD64
crw-rw----  1 root video 226, 128 Oct 26 20:40 renderD128

and if I look at /sys/class/drm, you can then see who owns which devices:

lrwxrwxrwx 1 root root    0 Oct 26 20:40 card0 -> ../../devices/platform/armada-drm/drm/card0
lrwxrwxrwx 1 root root    0 Oct 26 20:40 card0-HDMI-A-1 -> ../../devices/platform/armada-drm/drm/card0/card0-HDMI-A-1
lrwxrwxrwx 1 root root    0 Oct 26 20:40 card1 -> ../../devices/platform/etnaviv/drm/card1
lrwxrwxrwx 1 root root    0 Oct 26 20:40 controlD64 -> ../../devices/platform/armada-drm/drm/controlD64
lrwxrwxrwx 1 root root    0 Oct 26 20:40 renderD128 -> ../../devices/platform/etnaviv/drm/renderD128

So, renderD128 is card1's render node, which does not conform to the
assumption which the Intel DDX driver makes - (1 | 0x80) & 0xbf is
not 128.  The same thing can happen if there's ever a case on Intel
hardware where a KMS DRM driver registers prior to the i915 driver.

I think the only way to properly determine which render nodes
correspond with which master node is to actually open the device and
check the device names, or parse sysfs - maybe reading the links of
/sys/class/drm, and checking which link dirname corresponds with the
master node.

Any comments?

-- 
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.
_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/intel-gfx

Re: i915 render node discovery buggy?

[Martin Peres]

On 27/10/15 12:24, Russell King - ARM Linux wrote:
> In passing, while reading the Intel DDX driver code, I've noticed
> that the Intel driver contains code which assumes that the master and
> render devices are related by minor device number, eg:
>
>          /* Are we a render-node ourselves? */
>          if (is_render_node(fd, &master))
>                  return NULL;
>
>          sprintf(buf, "/dev/dri/renderD%d", (int)((master.st_rdev | 0x80) & 0xbf));
>          if (stat(buf, &render) == 0 &&
>              master.st_mode == render.st_mode &&
>              render.st_rdev == ((master.st_rdev | 0x80) & 0xbf))
>                  return strdup(buf);
>
> There's also code doing the reverse as well.
>
>  From my observations, the assumption that this code is built upon is
> false.  I have an ARM platform here (non-Intel graphics) which shows
> the problem - we have a KMS-only DRM driver (card0) and a GPU-only
> DRM driver (card1).  This populates the /dev/dri subdirectory as
> follows:
>
> crw-rw----+ 1 root video 226,   0 Oct 27 04:59 card0
> crw-rw----+ 1 root video 226,   1 Oct 26 20:40 card1
> crw-rw----  1 root video 226,  64 Oct 26 20:40 controlD64
> crw-rw----  1 root video 226, 128 Oct 26 20:40 renderD128
>
> and if I look at /sys/class/drm, you can then see who owns which devices:
>
> lrwxrwxrwx 1 root root    0 Oct 26 20:40 card0 -> ../../devices/platform/armada-drm/drm/card0
> lrwxrwxrwx 1 root root    0 Oct 26 20:40 card0-HDMI-A-1 -> ../../devices/platform/armada-drm/drm/card0/card0-HDMI-A-1
> lrwxrwxrwx 1 root root    0 Oct 26 20:40 card1 -> ../../devices/platform/etnaviv/drm/card1
> lrwxrwxrwx 1 root root    0 Oct 26 20:40 controlD64 -> ../../devices/platform/armada-drm/drm/controlD64
> lrwxrwxrwx 1 root root    0 Oct 26 20:40 renderD128 -> ../../devices/platform/etnaviv/drm/renderD128
>
> So, renderD128 is card1's render node, which does not conform to the
> assumption which the Intel DDX driver makes - (1 | 0x80) & 0xbf is
> not 128.  The same thing can happen if there's ever a case on Intel
> hardware where a KMS DRM driver registers prior to the i915 driver.
>
> I think the only way to properly determine which render nodes
> correspond with which master node is to actually open the device and
> check the device names, or parse sysfs - maybe reading the links of
> /sys/class/drm, and checking which link dirname corresponds with the
> master node.
>
> Any comments?

libdrm already has the right function for that:|char 
*drmGetRenderDeviceNameFromFd(int fd);


|

_______________________________________________
Intel-gfx mailing list
Intel-gfx@lists.freedesktop.org
http://lists.freedesktop.org/mailman/listinfo/intel-gfx