Re: [PATC 7/9] Skipping drm build, unsupported

["Philippe Mathieu-Daudé" <philmd@redhat.com>]

Hi Gerd,

On 7/1/20 5:15 PM, Gerd Hoffmann wrote:
>>> Ah, that is the problem.  Yes, DT_CHR is an non-posix optimization which
>>> allows to get the file type directly, without another round-trip to the
>>> kernel.  If that isn't available you can stat() the file and check
>>> ((st_mode & S_IFMT) == S_IFCHR) instead.
>>
>> Even when d_type and DT_CHR is available, there are filesystems where the
>> Linux kernel reports d_type of DT_UNKNOWN, and where you are best having
>> that code also falling back to an fstat().
> 
> Given this isn't perforance critical at all it is probably simplest to
> avoid non-portable d_type altogether and just to the fstat
> unconditionally.
> 
> David, does that work for haiku?
> 
> take care,
>   Gerd
> 
> diff --git a/util/drm.c b/util/drm.c
> index a23ff2453826..a1d3520d00f2 100644
> --- a/util/drm.c
> +++ b/util/drm.c
> @@ -24,6 +24,7 @@ int qemu_drm_rendernode_open(const char *rendernode)
>  {
>      DIR *dir;
>      struct dirent *e;
> +    struct stat st;
>      int r, fd;
>      char *p;
>  
> @@ -38,10 +39,6 @@ int qemu_drm_rendernode_open(const char *rendernode)
>  
>      fd = -1;
>      while ((e = readdir(dir))) {
> -        if (e->d_type != DT_CHR) {
> -            continue;
> -        }
> -
>          if (strncmp(e->d_name, "renderD", 7)) {
>              continue;
>          }
> @@ -53,6 +50,12 @@ int qemu_drm_rendernode_open(const char *rendernode)
>              g_free(p);
>              continue;
>          }
> +        fstat(r, &st);

While preparing the formal patch, can you add a comment here explaining
we deliberately use this way for portability (not checking DT_CHR /
DT_UNKNOWN ...)?

Thanks!

> +        if ((st.st_mode & S_IFMT) != S_IFCHR) {
> +            close(r);
> +            g_free(p);
> +            continue;
> +        }
>          fd = r;
>          g_free(p);
>          break;
>