| Message ID | 20200224215650.39624-1-chris@chris-wilson.co.uk (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | [i-g-t] i915/gem_ctx_persistence: Check precision of hostile cancellation | expand |
On 24/02/2020 21:56, Chris Wilson wrote: > Check that if we have to remove a hostile request from a non-persistent > context, we do so without harming any other concurrent users. > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > --- > tests/i915/gem_ctx_persistence.c | 63 ++++++++++++++++++++++++++++++++ > 1 file changed, 63 insertions(+) > > diff --git a/tests/i915/gem_ctx_persistence.c b/tests/i915/gem_ctx_persistence.c > index 20007f5c4..cd174d263 100644 > --- a/tests/i915/gem_ctx_persistence.c > +++ b/tests/i915/gem_ctx_persistence.c > @@ -613,6 +613,62 @@ static void test_process_mixed(int pfd, unsigned int engine) > gem_quiescent_gpu(pfd); > } > > +static void > +test_saturated_hostile(int i915, const struct intel_execution_engine2 *engine) > +{ > + const struct intel_execution_engine2 *other; > + igt_spin_t *spin; > + uint32_t ctx; > + int fence = -1; > + > + /* > + * Check that if we have to remove a hostile request from a > + * non-persistent context, we do so without harming any other > + * concurrent users. > + */ > + > + __for_each_physical_engine(i915, other) { > + if (other->flags == engine->flags) > + continue; > + > + spin = igt_spin_new(i915, > + .engine = other->flags, > + .flags = (IGT_SPIN_NO_PREEMPTION | > + IGT_SPIN_FENCE_OUT)); > + > + if (fence < 0) { > + fence = spin->out_fence; > + } else { > + int tmp; > + > + tmp = sync_fence_merge(fence, spin->out_fence); > + close(fence); > + close(spin->out_fence); > + > + fence = tmp; > + } > + spin->out_fence = -1; > + } > + > + ctx = gem_context_clone_with_engines(i915, 0); > + gem_context_set_persistence(i915, ctx, false); > + spin = igt_spin_new(i915, ctx, > + .engine = engine->flags, > + .flags = (IGT_SPIN_NO_PREEMPTION | > + IGT_SPIN_POLL_RUN | > + IGT_SPIN_FENCE_OUT)); > + igt_spin_busywait_until_started(spin); > + gem_context_destroy(i915, ctx); > + > + igt_assert_eq(sync_fence_wait(spin->out_fence, reset_timeout_ms), 0); > + igt_assert_eq(sync_fence_status(spin->out_fence), -EIO); > + > + /* All other spinners should be left unharmed */ > + gem_quiescent_gpu(i915); > + igt_assert_eq(sync_fence_wait(fence, reset_timeout_ms), 0); > + igt_assert_eq(sync_fence_status(fence), 1); I don't quite get this test. Why would other spinners be unharmed? They are non-preemptible as well. And non-persistent spinner is alone on the engine. So what aspect you wanted to test? Regards, Tvrtko > +} > + > static void test_processes(int i915) > { > struct { > @@ -1041,6 +1097,13 @@ igt_main > } > } > > + igt_subtest_with_dynamic_f("saturated-hostile") { > + __for_each_physical_engine(i915, e) { > + igt_dynamic_f("%s", e->name) > + test_saturated_hostile(i915, e); > + } > + } > + > igt_subtest("smoketest") > smoketest(i915); > } >
Quoting Tvrtko Ursulin (2020-02-25 18:08:14) > > On 24/02/2020 21:56, Chris Wilson wrote: > > Check that if we have to remove a hostile request from a non-persistent > > context, we do so without harming any other concurrent users. > > > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > > --- > > + /* All other spinners should be left unharmed */ > > + gem_quiescent_gpu(i915); > > + igt_assert_eq(sync_fence_wait(fence, reset_timeout_ms), 0); > > + igt_assert_eq(sync_fence_status(fence), 1); > > I don't quite get this test. Why would other spinners be unharmed? They > are non-preemptible as well. And non-persistent spinner is alone on the > engine. So what aspect you wanted to test? Per-engine reset. Termination of the non-persistent context should be clean and precise, we don't allow creation of non-persistent contexts unless we have that level of surgical precision. Otherwise it becomes a new attack vector. -Chris
On 25/02/2020 18:11, Chris Wilson wrote: > Quoting Tvrtko Ursulin (2020-02-25 18:08:14) >> >> On 24/02/2020 21:56, Chris Wilson wrote: >>> Check that if we have to remove a hostile request from a non-persistent >>> context, we do so without harming any other concurrent users. >>> >>> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> >>> --- >>> + /* All other spinners should be left unharmed */ >>> + gem_quiescent_gpu(i915); >>> + igt_assert_eq(sync_fence_wait(fence, reset_timeout_ms), 0); >>> + igt_assert_eq(sync_fence_status(fence), 1); >> >> I don't quite get this test. Why would other spinners be unharmed? They >> are non-preemptible as well. And non-persistent spinner is alone on the >> engine. So what aspect you wanted to test? > > Per-engine reset. Termination of the non-persistent context should be > clean and precise, we don't allow creation of non-persistent contexts > unless we have that level of surgical precision. Otherwise it becomes a > new attack vector. If it is just engine reset then it does what it says on the tin. Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> Regards, Tvrtko
Quoting Tvrtko Ursulin (2020-02-25 18:18:12) > > On 25/02/2020 18:11, Chris Wilson wrote: > > Quoting Tvrtko Ursulin (2020-02-25 18:08:14) > >> > >> On 24/02/2020 21:56, Chris Wilson wrote: > >>> Check that if we have to remove a hostile request from a non-persistent > >>> context, we do so without harming any other concurrent users. > >>> > >>> Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> > >>> --- > >>> + /* All other spinners should be left unharmed */ > >>> + gem_quiescent_gpu(i915); > >>> + igt_assert_eq(sync_fence_wait(fence, reset_timeout_ms), 0); > >>> + igt_assert_eq(sync_fence_status(fence), 1); > >> > >> I don't quite get this test. Why would other spinners be unharmed? They > >> are non-preemptible as well. And non-persistent spinner is alone on the > >> engine. So what aspect you wanted to test? > > > > Per-engine reset. Termination of the non-persistent context should be > > clean and precise, we don't allow creation of non-persistent contexts > > unless we have that level of surgical precision. Otherwise it becomes a > > new attack vector. > > If it is just engine reset then it does what it says on the tin. I shall update the description to clarify that then :) -Chris
diff --git a/tests/i915/gem_ctx_persistence.c b/tests/i915/gem_ctx_persistence.c index 20007f5c4..cd174d263 100644 --- a/tests/i915/gem_ctx_persistence.c +++ b/tests/i915/gem_ctx_persistence.c @@ -613,6 +613,62 @@ static void test_process_mixed(int pfd, unsigned int engine) gem_quiescent_gpu(pfd); } +static void +test_saturated_hostile(int i915, const struct intel_execution_engine2 *engine) +{ + const struct intel_execution_engine2 *other; + igt_spin_t *spin; + uint32_t ctx; + int fence = -1; + + /* + * Check that if we have to remove a hostile request from a + * non-persistent context, we do so without harming any other + * concurrent users. + */ + + __for_each_physical_engine(i915, other) { + if (other->flags == engine->flags) + continue; + + spin = igt_spin_new(i915, + .engine = other->flags, + .flags = (IGT_SPIN_NO_PREEMPTION | + IGT_SPIN_FENCE_OUT)); + + if (fence < 0) { + fence = spin->out_fence; + } else { + int tmp; + + tmp = sync_fence_merge(fence, spin->out_fence); + close(fence); + close(spin->out_fence); + + fence = tmp; + } + spin->out_fence = -1; + } + + ctx = gem_context_clone_with_engines(i915, 0); + gem_context_set_persistence(i915, ctx, false); + spin = igt_spin_new(i915, ctx, + .engine = engine->flags, + .flags = (IGT_SPIN_NO_PREEMPTION | + IGT_SPIN_POLL_RUN | + IGT_SPIN_FENCE_OUT)); + igt_spin_busywait_until_started(spin); + gem_context_destroy(i915, ctx); + + igt_assert_eq(sync_fence_wait(spin->out_fence, reset_timeout_ms), 0); + igt_assert_eq(sync_fence_status(spin->out_fence), -EIO); + + /* All other spinners should be left unharmed */ + gem_quiescent_gpu(i915); + igt_assert_eq(sync_fence_wait(fence, reset_timeout_ms), 0); + igt_assert_eq(sync_fence_status(fence), 1); +} + static void test_processes(int i915) { struct { @@ -1041,6 +1097,13 @@ igt_main } } + igt_subtest_with_dynamic_f("saturated-hostile") { + __for_each_physical_engine(i915, e) { + igt_dynamic_f("%s", e->name) + test_saturated_hostile(i915, e); + } + } + igt_subtest("smoketest") smoketest(i915); }
Check that if we have to remove a hostile request from a non-persistent context, we do so without harming any other concurrent users. Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> --- tests/i915/gem_ctx_persistence.c | 63 ++++++++++++++++++++++++++++++++ 1 file changed, 63 insertions(+)