]> git.baikalelectronics.ru Git - kernel.git/commitdiff
drm/i915/selftests: Basic stress test for rapid context switching
authorChris Wilson <chris@chris-wilson.co.uk>
Thu, 20 Sep 2018 10:58:09 +0000 (11:58 +0100)
committerChris Wilson <chris@chris-wilson.co.uk>
Thu, 20 Sep 2018 11:49:25 +0000 (12:49 +0100)
We need to exercise the HW and submission paths for switching contexts
rapidly to check that features such as execlists' wa_tail are adequate.
Plus it's an interesting baseline latency metric.

v2: Check the initial request for allocation errors
v3: Use finite waits for more robust handling of broken code

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180920105809.1872-1-chris@chris-wilson.co.uk
drivers/gpu/drm/i915/selftests/i915_gem_context.c

index 07d700b7cb9440dbaa2e43e6825b71f6f99049bc..441a51d4aa54a7c30129f4541ad5611bf478767c 100644 (file)
@@ -22,6 +22,8 @@
  *
  */
 
+#include <linux/prime_numbers.h>
+
 #include "../i915_selftest.h"
 #include "i915_random.h"
 #include "igt_flush_test.h"
 
 #define DW_PER_PAGE (PAGE_SIZE / sizeof(u32))
 
+struct live_test {
+       struct drm_i915_private *i915;
+       const char *func;
+       const char *name;
+
+       unsigned int reset_count;
+};
+
+static int begin_live_test(struct live_test *t,
+                          struct drm_i915_private *i915,
+                          const char *func,
+                          const char *name)
+{
+       int err;
+
+       t->i915 = i915;
+       t->func = func;
+       t->name = name;
+
+       err = i915_gem_wait_for_idle(i915,
+                                    I915_WAIT_LOCKED,
+                                    MAX_SCHEDULE_TIMEOUT);
+       if (err) {
+               pr_err("%s(%s): failed to idle before, with err=%d!",
+                      func, name, err);
+               return err;
+       }
+
+       i915->gpu_error.missed_irq_rings = 0;
+       t->reset_count = i915_reset_count(&i915->gpu_error);
+
+       return 0;
+}
+
+static int end_live_test(struct live_test *t)
+{
+       struct drm_i915_private *i915 = t->i915;
+
+       if (igt_flush_test(i915, I915_WAIT_LOCKED))
+               return -EIO;
+
+       if (t->reset_count != i915_reset_count(&i915->gpu_error)) {
+               pr_err("%s(%s): GPU was reset %d times!\n",
+                      t->func, t->name,
+                      i915_reset_count(&i915->gpu_error) - t->reset_count);
+               return -EIO;
+       }
+
+       if (i915->gpu_error.missed_irq_rings) {
+               pr_err("%s(%s): Missed interrupts on engines %lx\n",
+                      t->func, t->name, i915->gpu_error.missed_irq_rings);
+               return -EIO;
+       }
+
+       return 0;
+}
+
+static int live_nop_switch(void *arg)
+{
+       const unsigned int nctx = 1024;
+       struct drm_i915_private *i915 = arg;
+       struct intel_engine_cs *engine;
+       struct i915_gem_context **ctx;
+       enum intel_engine_id id;
+       struct drm_file *file;
+       struct live_test t;
+       unsigned long n;
+       int err = -ENODEV;
+
+       /*
+        * Create as many contexts as we can feasibly get away with
+        * and check we can switch between them rapidly.
+        *
+        * Serves as very simple stress test for submission and HW switching
+        * between contexts.
+        */
+
+       if (!DRIVER_CAPS(i915)->has_logical_contexts)
+               return 0;
+
+       file = mock_file(i915);
+       if (IS_ERR(file))
+               return PTR_ERR(file);
+
+       mutex_lock(&i915->drm.struct_mutex);
+
+       ctx = kcalloc(nctx, sizeof(*ctx), GFP_KERNEL);
+       if (!ctx) {
+               err = -ENOMEM;
+               goto out_unlock;
+       }
+
+       for (n = 0; n < nctx; n++) {
+               ctx[n] = i915_gem_create_context(i915, file->driver_priv);
+               if (IS_ERR(ctx[n])) {
+                       err = PTR_ERR(ctx[n]);
+                       goto out_unlock;
+               }
+       }
+
+       for_each_engine(engine, i915, id) {
+               struct i915_request *rq;
+               unsigned long end_time, prime;
+               ktime_t times[2] = {};
+
+               times[0] = ktime_get_raw();
+               for (n = 0; n < nctx; n++) {
+                       rq = i915_request_alloc(engine, ctx[n]);
+                       if (IS_ERR(rq)) {
+                               err = PTR_ERR(rq);
+                               goto out_unlock;
+                       }
+                       i915_request_add(rq);
+               }
+               if (i915_request_wait(rq,
+                                     I915_WAIT_LOCKED,
+                                     HZ / 5) < 0) {
+                       pr_err("Failed to populated %d contexts\n", nctx);
+                       i915_gem_set_wedged(i915);
+                       err = -EIO;
+                       goto out_unlock;
+               }
+
+               times[1] = ktime_get_raw();
+
+               pr_info("Populated %d contexts on %s in %lluns\n",
+                       nctx, engine->name, ktime_to_ns(times[1] - times[0]));
+
+               err = begin_live_test(&t, i915, __func__, engine->name);
+               if (err)
+                       goto out_unlock;
+
+               end_time = jiffies + i915_selftest.timeout_jiffies;
+               for_each_prime_number_from(prime, 2, 8192) {
+                       times[1] = ktime_get_raw();
+
+                       for (n = 0; n < prime; n++) {
+                               rq = i915_request_alloc(engine, ctx[n % nctx]);
+                               if (IS_ERR(rq)) {
+                                       err = PTR_ERR(rq);
+                                       goto out_unlock;
+                               }
+
+                               /*
+                                * This space is left intentionally blank.
+                                *
+                                * We do not actually want to perform any
+                                * action with this request, we just want
+                                * to measure the latency in allocation
+                                * and submission of our breadcrumbs -
+                                * ensuring that the bare request is sufficient
+                                * for the system to work (i.e. proper HEAD
+                                * tracking of the rings, interrupt handling,
+                                * etc). It also gives us the lowest bounds
+                                * for latency.
+                                */
+
+                               i915_request_add(rq);
+                       }
+                       if (i915_request_wait(rq,
+                                             I915_WAIT_LOCKED,
+                                             HZ / 5) < 0) {
+                               pr_err("Switching between %ld contexts timed out\n",
+                                      prime);
+                               i915_gem_set_wedged(i915);
+                               break;
+                       }
+
+                       times[1] = ktime_sub(ktime_get_raw(), times[1]);
+                       if (prime == 2)
+                               times[0] = times[1];
+
+                       if (__igt_timeout(end_time, NULL))
+                               break;
+               }
+
+               err = end_live_test(&t);
+               if (err)
+                       goto out_unlock;
+
+               pr_info("Switch latencies on %s: 1 = %lluns, %lu = %lluns\n",
+                       engine->name,
+                       ktime_to_ns(times[0]),
+                       prime - 1, div64_u64(ktime_to_ns(times[1]), prime - 1));
+       }
+
+out_unlock:
+       mutex_unlock(&i915->drm.struct_mutex);
+       mock_file_free(i915, file);
+       return err;
+}
+
 static struct i915_vma *
 gpu_fill_dw(struct i915_vma *vma, u64 offset, unsigned long count, u32 value)
 {
@@ -714,6 +908,7 @@ int i915_gem_context_live_selftests(struct drm_i915_private *dev_priv)
 {
        static const struct i915_subtest tests[] = {
                SUBTEST(igt_switch_to_kernel_context),
+               SUBTEST(live_nop_switch),
                SUBTEST(igt_ctx_exec),
                SUBTEST(igt_ctx_readonly),
        };