Message ID | 20170119114158.17941-15-chris@chris-wilson.co.uk (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On to, 2017-01-19 at 11:41 +0000, Chris Wilson wrote: > Just create several batches of requests and expect it to not fall over! > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> <SNIP> > +static int live_nop_request(void *arg) > +{ <SNIP> > + for_each_prime_number_from(prime, 1, 8192) { #define MAGIC_PRIME_2 8192 > + times[1] = ktime_get_raw(); > + > + for (n = 0; n < prime; n++) { > + request = i915_gem_request_alloc(engine, > + i915->kernel_context); > + if (IS_ERR(request)) { > + err = PTR_ERR(request); > + goto out_unlock; > + } Emitting even a single MI_NOOP or making a comment that the request will contain instructions even if we do nothing here, might help a newcomer. > + > + i915_add_request(request); > + } > + i915_wait_request(request, > + I915_WAIT_LOCKED, > + MAX_SCHEDULE_TIMEOUT); > + > + times[1] = ktime_sub(ktime_get_raw(), times[1]); > + if (prime == 1) > + times[0] = times[1]; Having this as an array is just hiding names and gaining nothing, how about calling them just time_first, time_last? Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> Regards, Joonas
On Wed, Feb 01, 2017 at 10:14:28AM +0200, Joonas Lahtinen wrote: > On to, 2017-01-19 at 11:41 +0000, Chris Wilson wrote: > > Just create several batches of requests and expect it to not fall over! > > > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > > <SNIP> > > > +static int live_nop_request(void *arg) > > +{ > > <SNIP> > > > + for_each_prime_number_from(prime, 1, 8192) { > > #define MAGIC_PRIME_2 8192 I'm sorry, but that is worse than having a very clear arbitrary number. > > + > > + i915_add_request(request); > > + } > > + i915_wait_request(request, > > + I915_WAIT_LOCKED, > > + MAX_SCHEDULE_TIMEOUT); > > + > > + times[1] = ktime_sub(ktime_get_raw(), times[1]); > > + if (prime == 1) > > + times[0] = times[1]; > > Having this as an array is just hiding names and gaining nothing, how > about calling them just time_first, time_last? time_0, time_N? The array is because I did/am considering that the graph may be interesting. There should be a plateau at the point the ring is full - not world shattering, just mildly interesting. Still, probably want the times for cold submit_1 vs warm submit_1. -Chris
diff --git a/drivers/gpu/drm/i915/selftests/i915_gem_request.c b/drivers/gpu/drm/i915/selftests/i915_gem_request.c index 20bf10dd85ed..19103d87a4c3 100644 --- a/drivers/gpu/drm/i915/selftests/i915_gem_request.c +++ b/drivers/gpu/drm/i915/selftests/i915_gem_request.c @@ -22,6 +22,8 @@ * */ +#include <linux/prime_numbers.h> + #include "i915_selftest.h" #include "mock_gem_device.h" @@ -161,3 +163,96 @@ int i915_gem_request_mock_selftests(void) return err; } + +static int live_nop_request(void *arg) +{ + struct drm_i915_private *i915 = arg; + struct intel_engine_cs *engine; + unsigned int id; + int err; + + /* Submit various sized batches of empty requests, to each engine + * (individually), and wait for the batch to complete. We can check + * the overhead of submitting requests to the hardware. + */ + + mutex_lock(&i915->drm.struct_mutex); + + for_each_engine(engine, i915, id) { + IGT_TIMEOUT(end_time); + struct drm_i915_gem_request *request; + unsigned int reset_count; + unsigned long n, prime; + ktime_t times[2] = {}; + + err = i915_gem_wait_for_idle(i915, I915_WAIT_LOCKED); + if (err) { + pr_err("Failed to idle GPU before %s(%s)\n", + __func__, engine->name); + goto out_unlock; + } + + i915->gpu_error.missed_irq_rings = 0; + reset_count = i915_reset_count(&i915->gpu_error); + + for_each_prime_number_from(prime, 1, 8192) { + times[1] = ktime_get_raw(); + + for (n = 0; n < prime; n++) { + request = i915_gem_request_alloc(engine, + i915->kernel_context); + if (IS_ERR(request)) { + err = PTR_ERR(request); + goto out_unlock; + } + + i915_add_request(request); + } + i915_wait_request(request, + I915_WAIT_LOCKED, + MAX_SCHEDULE_TIMEOUT); + + times[1] = ktime_sub(ktime_get_raw(), times[1]); + if (prime == 1) + times[0] = times[1]; + + if (igt_timeout(end_time, + "%s(%s) timed out: last batch size %lu\n", + __func__, engine->name, prime)) + break; + } + + if (reset_count != i915_reset_count(&i915->gpu_error)) { + pr_err("%s(%s): GPU was reset %d times!\n", + __func__, engine->name, + i915_reset_count(&i915->gpu_error) - reset_count); + err = -EIO; + goto out_unlock; + } + + if (i915->gpu_error.missed_irq_rings) { + pr_err("%s(%s): Missed interrupts on rings %lx\n", + __func__, engine->name, + i915->gpu_error.missed_irq_rings); + err = -EIO; + goto out_unlock; + } + + pr_info("Request latencies on %s: 1 = %lluns, %lu = %lluns\n", + engine->name, + ktime_to_ns(times[0]), + prime, div64_u64(ktime_to_ns(times[1]), prime)); + } + +out_unlock: + mutex_unlock(&i915->drm.struct_mutex); + return err; +} + +int i915_gem_request_live_selftests(struct drm_i915_private *i915) +{ + static const struct i915_subtest tests[] = { + SUBTEST(live_nop_request), + }; + return i915_subtests(tests, i915); +} diff --git a/drivers/gpu/drm/i915/selftests/i915_live_selftests.h b/drivers/gpu/drm/i915/selftests/i915_live_selftests.h index f3e17cb10e05..09bf538826df 100644 --- a/drivers/gpu/drm/i915/selftests/i915_live_selftests.h +++ b/drivers/gpu/drm/i915/selftests/i915_live_selftests.h @@ -9,3 +9,4 @@ * Tests are executed in order by igt/drv_selftest */ selftest(sanitycheck, i915_live_sanitycheck) /* keep first (igt selfcheck) */ +selftest(requests, i915_gem_request_live_selftests)
Just create several batches of requests and expect it to not fall over! Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> --- drivers/gpu/drm/i915/selftests/i915_gem_request.c | 95 ++++++++++++++++++++++ .../gpu/drm/i915/selftests/i915_live_selftests.h | 1 + 2 files changed, 96 insertions(+)