Re: amdgpu progress

matthew green <[email protected]> wrote:
>Robert Swindells writes:
>> 
>> matthew green <[email protected]> wrote:
>> > the pkgsrc amdgpu issue is odd.  the amdgpu xf86 driver should
>> > be telling mesa to load "radeonsi", and this seems to work with
>> > xsrc:
>> >
>> > xf86-video-amdgpu/dist/src/amdgpu_version.h:39:#define SI_DRIVER_NAME       "radeonsi"
>> > xf86-video-amdgpu/dist/src/amdgpu_dri2.c:1308:      dri2_info.driverName = SI_DRIVER_NAME;
>>
>> There seem to be a lot more PCI IDs in:
>>
>> xsrc/external/mit/MesaLib.old/dist/include/pci_ids/radeonsi_pci_ids.h
>>
>> than in:
>>
>> xsrc/external/mit/MesaLib/dist/include/pci_ids/radeonsi_pci_ids.h
>
> huh.  did you find why this was removed?  i don't see them
> added anywhere instead.. and all those chipsets AFAIK, still
> want to use radeonsi driver -- weird.

If USE_DRMCONF is defined then I don't think it matters whether the IDs
are there or not.

There is the function below in loader.c, I think the call to
loader_get_pci_driver() should be before the one to
loader_get_dri_config_driver().

char *
loader_get_driver_for_fd(int fd)
{
   char *driver;

   /* Allow an environment variable to force choosing a different driver
    * binary.  If that driver binary can't survive on this FD, that's the
    * user's problem, but this allows vc4 simulator to run on an i965 host,
    * and may be useful for some touch testing of i915 on an i965 host.
    */
   if (geteuid() == getuid()) {
      driver = getenv("MESA_LOADER_DRIVER_OVERRIDE");
      if (driver)
         return strdup(driver);
   }

#if defined(HAVE_LIBDRM) && defined(USE_DRICONF)
   driver = loader_get_dri_config_driver(fd);
   if (driver)
      return driver;
#endif

   driver = loader_get_pci_driver(fd);
   if (!driver)
      driver = loader_get_kernel_driver_name(fd);

   return driver;
}

For evbarm, we would want both config methods to work.

> with either new or old xsrc Mesa i still have problems.  with
> old, it eventually crashes inside glamoregl portion of the
> X server.. with new it hangs, and fails reset, leaving with
> a weird broken palette window etc.

I'm wondering if the version of DRM/KMS code in the kernel is supposed
to work for amdgpu, or if it was incomplete at that point.