Re: [PATCH i-g-t] lib/intel_pat: Cache PAT config for unprivileged processes

["Dixit, Ashutosh" <ashutosh.dixit@intel.com>]

On Sun, 01 Mar 2026 07:35:53 -0800, himanshu.girotra@intel.com wrote:
>
> diff --git a/lib/intel_pat.c b/lib/intel_pat.c
> index 8660a2515..16dfc0217 100644
> --- a/lib/intel_pat.c
> +++ b/lib/intel_pat.c
> @@ -7,6 +7,16 @@
>  #include "igt.h"
>  #include "intel_pat.h"
>
> +/*
> + * Global PAT cache.  PAT configuration is a hardware/driver property that
> + * does not change over the lifetime of a device, so we cache the result of
> + * the first successful query.  This allows unprivileged processes (e.g.
> + * after igt_drop_root()) to use the cached values without needing debugfs
> + * access.
> + */
> +static struct intel_pat_cache pat_cache;
> +static bool pat_cached;

This is not correct. The cache should be maintained as part of 'struct
xe_device'.  All tests, include xe_oa, already call xe_device_get(). The
code in lib/intel_pat.c can also call xe_device_get() to check if pat
entries are cached.

> diff --git a/tests/intel/xe_oa.c b/tests/intel/xe_oa.c
> index 927f3f4f2..37449c87c 100644
> --- a/tests/intel/xe_oa.c
> +++ b/tests/intel/xe_oa.c
> @@ -27,6 +27,7 @@
>  #include "xe/xe_ioctl.h"
>  #include "xe/xe_query.h"
>  #include "xe/xe_oa.h"
> +#include "intel_pat.h"
>
>  /**
>   * TEST: perf
> @@ -5094,6 +5095,15 @@ int igt_main_args("b:t", long_options, help_str, opt_handler, NULL)
>		write_u64_file("/proc/sys/dev/xe/observation_paranoid", 1);
>
>		render_copy = igt_get_render_copyfunc(drm_fd);
> +
> +		/*
> +		 * Pre-cache PAT configuration while we still have root
> +		 * privileges.  Several subtests fork a child that drops
> +		 * root before using bufops/intel_bb, which internally
> +		 * needs the PAT WB index.  Without this, the child would
> +		 * fail to read debugfs after dropping privileges.
> +		 */
> +		intel_pat_precache(drm_fd);

As mentioned above, this is not needed if the cache is maintained as part
of 'struct xe_device' since xe_device_get() is being called above.