[PATCH libdrm] xf86drm: Parse the separate files to retrieve the vendor/device info

[Peter Wu]

On Tue, Nov 01, 2016 at 06:13:34PM +0000, Emil Velikov wrote:
>  sysfs file wakes up the device. The latter of which may
> be slow and isn't required to begin with.
> 
> Reading through config is/was required since the revision is not
> available by other means, although with a kernel patch in the way we can
> 'cheat' temporarily.
> 
> That should be fine, since no open-source project has ever used the
> value.
> 
> Cc: Michel Dänzer <michel.daenzer at amd.com>
> Cc: Mauro Santos <registo.mailling at gmail.com>
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=98502
> Signed-off-by: Emil Velikov <emil.velikov at collabora.com>

See below for one observation. Other than that, strace confirms that
the new sysfs files are being accessed.

Reviewed-and-tested-by: Peter Wu <peter at lekensteyn.nl>

This was tested with Linux 4.8.4 (unpatched) and libdrm 2.4.71 (+this
patch) with i915 + nouveau. Tested with running glxgears and glxinfo.
Note that glxinfo still accesses 'config' due to libpciaccess.

+ drm_intel_get_aperture_sizes
  + drm_intel_probe_agp_aperture_size
    + pci_system_init()
      + pci_system_linux_sysfs_create
        + populate_entries
          + pci_device_linux_sysfs_read() <-- offending function
    + pci_device_find_by_slot(0, 0, 2, 0)

That populate_entries function can probably use a fix similar to this
one.

> ---
> Mauro can you apply this against libdrm and rebuild it. You do _not_
> need to rebuild mesa afterwords.
> 
> Thanks
> ---
>  xf86drm.c | 50 +++++++++++++++++++++++++++++++++++---------------
>  1 file changed, 35 insertions(+), 15 deletions(-)
> 
> diff --git a/xf86drm.c b/xf86drm.c
> index 52add5e..5a5100c 100644
> --- a/xf86drm.c
> +++ b/xf86drm.c
> @@ -2950,25 +2950,45 @@ static int drmParsePciDeviceInfo(const char *d_name,
>                                   drmPciDeviceInfoPtr device)
>  {
>  #ifdef __linux__
> +#define ARRAY_SIZE(a) (sizeof(a) / sizeof((a)[0]))
> +    static const char *attrs[] = {
> +      "revision", /* XXX: make sure it's always first, see note below */
> +      "vendor",
> +      "device",
> +      "subsystem_vendor",
> +      "subsystem_device",
> +    };
>      char path[PATH_MAX + 1];

The size for snprintf already includes the nul-terminator, so strictly
speaking the +1 is not needed. It does not hurt either though.

> -    unsigned char config[64];
> -    int fd, ret;
> +    unsigned int data[ARRAY_SIZE(attrs)];
> +    FILE *fp;
> +    int ret;
>  
> -    snprintf(path, PATH_MAX, "/sys/class/drm/%s/device/config", d_name);
> -    fd = open(path, O_RDONLY);
> -    if (fd < 0)
> -        return -errno;
> +    for (unsigned i = 0; i < ARRAY_SIZE(attrs); i++) {
> +        snprintf(path, PATH_MAX, "/sys/class/drm/%s/device/%s",
> +                 d_name, attrs[i]);
> +        fp = fopen(path, "r");
> +        if (!fp) {
> +            /* Note: First we check the revision, since older kernels
> +             * may not have it. Default to zero in such cases. */
> +            if (i == 0) {
> +                data[i] = 0;
> +                continue;
> +            }
> +            return -errno;
> +        }
>  
> -    ret = read(fd, config, sizeof(config));
> -    close(fd);
> -    if (ret < 0)
> -        return -errno;
> +        ret = fscanf(fp, "%x", &data[i]);
> +        fclose(fp);
> +        if (ret != 1)
> +            return -errno;
> +
> +    }
>  
> -    device->vendor_id = config[0] | (config[1] << 8);
> -    device->device_id = config[2] | (config[3] << 8);
> -    device->revision_id = config[8];
> -    device->subvendor_id = config[44] | (config[45] << 8);
> -    device->subdevice_id = config[46] | (config[47] << 8);
> +    device->revision_id = data[0] & 0xff;
> +    device->vendor_id = data[1] & 0xffff;
> +    device->device_id = data[2] & 0xffff;
> +    device->subvendor_id = data[3] & 0xffff;
> +    device->subdevice_id = data[4] & 0xffff;
>  
>      return 0;
>  #else
> -- 
> 2.10.0