drm/i915: Add dev.i915.perf_stream_paranoid sysctl option
authorRobert Bragg <robert@sixbynine.org>
Mon, 7 Nov 2016 19:49:54 +0000 (19:49 +0000)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Tue, 22 Nov 2016 13:39:00 +0000 (14:39 +0100)
Consistent with the kernel.perf_event_paranoid sysctl option that can
allow non-root users to access system wide cpu metrics, this can
optionally allow non-root users to access system wide OA counter metrics
from Gen graphics hardware.

Signed-off-by: Robert Bragg <robert@sixbynine.org>
Reviewed-by: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: Sourab Gupta <sourab.gupta@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/20161107194957.3385-9-robert@sixbynine.org
drivers/gpu/drm/i915/i915_drv.h
drivers/gpu/drm/i915/i915_perf.c

index 8da752304a9196cc815649eff125005aab920cc8..970e50bf98840e6bc0ca5aa8ce88999c55b3131a 100644 (file)
@@ -2194,6 +2194,7 @@ struct drm_i915_private {
                bool initialized;
 
                struct kobject *metrics_kobj;
+               struct ctl_table_header *sysctl_header;
 
                struct mutex lock;
                struct list_head streams;
index c427cd8cbe1bdb1d6d002153ba0461b793c43f51..e51c1d82255ff5ab54286a22d3db3dcb9f59a63d 100644 (file)
 #define POLL_FREQUENCY 200
 #define POLL_PERIOD (NSEC_PER_SEC / POLL_FREQUENCY)
 
+/* for sysctl proc_dointvec_minmax of dev.i915.perf_stream_paranoid */
+static int zero;
+static int one = 1;
+static u32 i915_perf_stream_paranoid = true;
+
 /* The maximum exponent the hardware accepts is 63 (essentially it selects one
  * of the 64bit timestamp bits to trigger reports from) but there's currently
  * no known use case for sampling as infrequently as once per 47 thousand years.
@@ -1207,7 +1212,13 @@ i915_perf_open_ioctl_locked(struct drm_i915_private *dev_priv,
                }
        }
 
-       if (!specific_ctx && !capable(CAP_SYS_ADMIN)) {
+       /* Similar to perf's kernel.perf_paranoid_cpu sysctl option
+        * we check a dev.i915.perf_stream_paranoid sysctl option
+        * to determine if it's ok to access system wide OA counters
+        * without CAP_SYS_ADMIN privileges.
+        */
+       if (!specific_ctx &&
+           i915_perf_stream_paranoid && !capable(CAP_SYS_ADMIN)) {
                DRM_ERROR("Insufficient privileges to open system-wide i915 perf stream\n");
                ret = -EACCES;
                goto err_ctx;
@@ -1460,6 +1471,39 @@ void i915_perf_unregister(struct drm_i915_private *dev_priv)
        dev_priv->perf.metrics_kobj = NULL;
 }
 
+static struct ctl_table oa_table[] = {
+       {
+        .procname = "perf_stream_paranoid",
+        .data = &i915_perf_stream_paranoid,
+        .maxlen = sizeof(i915_perf_stream_paranoid),
+        .mode = 0644,
+        .proc_handler = proc_dointvec_minmax,
+        .extra1 = &zero,
+        .extra2 = &one,
+        },
+       {}
+};
+
+static struct ctl_table i915_root[] = {
+       {
+        .procname = "i915",
+        .maxlen = 0,
+        .mode = 0555,
+        .child = oa_table,
+        },
+       {}
+};
+
+static struct ctl_table dev_root[] = {
+       {
+        .procname = "dev",
+        .maxlen = 0,
+        .mode = 0555,
+        .child = i915_root,
+        },
+       {}
+};
+
 void i915_perf_init(struct drm_i915_private *dev_priv)
 {
        if (!IS_HASWELL(dev_priv))
@@ -1490,6 +1534,8 @@ void i915_perf_init(struct drm_i915_private *dev_priv)
        dev_priv->perf.oa.n_builtin_sets =
                i915_oa_n_builtin_metric_sets_hsw;
 
+       dev_priv->perf.sysctl_header = register_sysctl_table(dev_root);
+
        dev_priv->perf.initialized = true;
 }
 
@@ -1498,6 +1544,8 @@ void i915_perf_fini(struct drm_i915_private *dev_priv)
        if (!dev_priv->perf.initialized)
                return;
 
+       unregister_sysctl_table(dev_priv->perf.sysctl_header);
+
        memset(&dev_priv->perf.oa.ops, 0, sizeof(dev_priv->perf.oa.ops));
        dev_priv->perf.initialized = false;
 }