test/i915/gem_exec_reloc: Restore interclient testings

These tests were removed under a blatantly false statement to hide a
regression. Denying that userspace can deadlock other clients, does not
stop the violations. Introducing such deadlocks deliberately is a cause
for concern.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

1 files changed, 219 insertions, 0 deletions

diff --git a/tests/i915/gem_exec_reloc.c b/tests/i915/gem_exec_reloc.c
index 8dcb24a6..cc9b8cd6 100644
--- a/tests/i915/gem_exec_reloc.c
+++ b/tests/i915/gem_exec_reloc.c
@@ -424,6 +424,76 @@ static void many_active(int i915, unsigned engine)
 	}
 }
 
+static void __wide_active(int i915, unsigned engine, unsigned long count)
+{
+	struct drm_i915_gem_relocation_entry *reloc =
+		calloc(count, sizeof(*reloc));
+	struct drm_i915_gem_exec_object2 *obj =
+		calloc(count + 1, sizeof(*obj));
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(obj),
+		.buffer_count = count + 1,
+		.flags = engine | I915_EXEC_HANDLE_LUT,
+	};
+	igt_spin_t *spin;
+
+	for (unsigned long i = 0; i < count; i++) {
+		obj[i].handle = gem_create(i915, 4096);
+		obj[i].flags = EXEC_OBJECT_WRITE;
+		obj[i].flags |= EXEC_OBJECT_SUPPORTS_48B_ADDRESS;
+	}
+
+	spin = __igt_spin_new(i915,
+			      .engine = engine,
+			      .flags = (IGT_SPIN_FENCE_OUT |
+					IGT_SPIN_NO_PREEMPTION));
+	obj[count] = spin->obj[1];
+	gem_execbuf(i915, &execbuf); /* mark all the objects as active */
+
+	for (unsigned long i = 0; i < count; i++) {
+		reloc[i].target_handle = i;
+		reloc[i].presumed_offset = ~0ull;
+		obj[i].relocs_ptr = to_user_pointer(&reloc[i]);
+		obj[i].relocation_count = 1;
+	}
+	gem_execbuf(i915, &execbuf); /* relocation onto active objects */
+
+	igt_assert_eq(sync_fence_status(spin->out_fence), 0);
+	igt_spin_free(i915, spin);
+
+	for (unsigned long i = 0; i < count; i++) {
+		uint64_t addr;
+
+		gem_read(i915, obj[i].handle, 0, &addr, sizeof(addr));
+		igt_assert_eq_u64(addr, obj[i].offset);
+
+		gem_close(i915, obj[i].handle);
+	}
+	free(obj);
+	free(reloc);
+}
+
+static void wide_active(int i915, unsigned engine)
+{
+	const uint64_t max = gem_aperture_size(i915) / 4096 / 2;
+	unsigned long count = 256;
+
+	igt_until_timeout(2) {
+		uint64_t required, total;
+
+		if (!__intel_check_memory(count, 4096, CHECK_RAM,
+					  &required, &total))
+			break;
+
+		igt_debug("Testing count:%lu\n", count);
+		__wide_active(i915, engine, count);
+
+		count <<= 1;
+		if (count >= max)
+			break;
+	}
+}
+
 static unsigned int offset_in_page(void *addr)
 {
 	return (uintptr_t)addr & 4095;
@@ -867,6 +937,145 @@ static void basic_softpin(int fd)
 	gem_close(fd, obj[1].handle);
 }
 
+static struct drm_i915_gem_relocation_entry *
+parallel_relocs(int count, unsigned long *out)
+{
+	struct drm_i915_gem_relocation_entry *reloc;
+	unsigned long sz;
+	int i;
+
+	sz = count * sizeof(*reloc);
+	sz = ALIGN(sz, 4096);
+
+	reloc = mmap(0, sz, PROT_WRITE, MAP_PRIVATE | MAP_ANON, -1, 0);
+	igt_assert(reloc != MAP_FAILED);
+	for (i = 0; i < count; i++) {
+		reloc[i].target_handle = 0;
+		reloc[i].presumed_offset = ~0ull;
+		reloc[i].offset = 8 * i;
+		reloc[i].delta = i;
+		reloc[i].read_domains = I915_GEM_DOMAIN_INSTRUCTION;
+		reloc[i].write_domain = 0;
+	}
+	mprotect(reloc, sz, PROT_READ);
+
+	*out = sz;
+	return reloc;
+}
+
+static int __execbuf(int i915, struct drm_i915_gem_execbuffer2 *execbuf)
+{
+	int err;
+
+	err = 0;
+	if (ioctl(i915, DRM_IOCTL_I915_GEM_EXECBUFFER2, execbuf)) {
+		err = -errno;
+		igt_assume(err);
+	}
+
+	errno = 0;
+	return err;
+}
+
+static int stop;
+static void sighandler(int sig)
+{
+	stop = 1;
+}
+
+static void parallel_child(int i915,
+			   const struct intel_execution_engine2 *engine,
+			   struct drm_i915_gem_relocation_entry *reloc,
+			   uint32_t common)
+{
+	igt_spin_t *spin = __igt_spin_new(i915, .engine = engine->flags);
+	struct drm_i915_gem_exec_object2 reloc_target = {
+		.handle = gem_create(i915, 32 * 1024 * 8),
+		.relocation_count = 32 * 1024,
+		.relocs_ptr = to_user_pointer(reloc),
+	};
+	struct drm_i915_gem_exec_object2 obj[3] = {
+		reloc_target,
+		{ .handle = common },
+		spin->obj[1],
+	};
+	struct drm_i915_gem_execbuffer2 execbuf = {
+		.buffers_ptr = to_user_pointer(obj),
+		.buffer_count = ARRAY_SIZE(obj),
+		.flags = engine->flags | I915_EXEC_HANDLE_LUT,
+	};
+	struct sigaction act = {
+		.sa_handler = sighandler,
+	};
+	unsigned long count = 0;
+
+	sigaction(SIGINT, &act, NULL);
+	while (!READ_ONCE(stop)) {
+		int err = __execbuf(i915, &execbuf);
+		if (err == -EINTR)
+			break;
+
+		igt_assert_eq(err, 0);
+		count++;
+	}
+
+	igt_info("%s: count %lu\n", engine->name, count);
+	igt_spin_free(i915, spin);
+}
+
+static void kill_children(int sig)
+{
+	signal(sig, SIG_IGN);
+	kill(-getpgrp(), SIGINT);
+	signal(sig, SIG_DFL);
+}
+
+static void parallel(int i915)
+{
+	const struct intel_execution_engine2 *e;
+	struct drm_i915_gem_relocation_entry *reloc;
+	uint32_t common = gem_create(i915, 4096);
+	uint32_t batch = batch_create(i915);
+	unsigned long reloc_sz;
+
+	reloc = parallel_relocs(32 * 1024, &reloc_sz);
+
+	stop = 0;
+	__for_each_physical_engine(i915, e) {
+		igt_fork(child, 1)
+			parallel_child(i915, e, reloc, common);
+	}
+	sleep(2);
+
+	if (gem_scheduler_has_preemption(i915)) {
+		uint32_t ctx = gem_context_clone_with_engines(i915, 0);
+
+		__for_each_physical_engine(i915, e) {
+			struct drm_i915_gem_exec_object2 obj[2] = {
+				{ .handle = common },
+				{ .handle = batch },
+			};
+			struct drm_i915_gem_execbuffer2 execbuf = {
+				.buffers_ptr = to_user_pointer(obj),
+				.buffer_count = ARRAY_SIZE(obj),
+				.flags = e->flags,
+				.rsvd1 = ctx,
+			};
+			gem_execbuf(i915, &execbuf);
+		}
+
+		gem_context_destroy(i915, ctx);
+	}
+	gem_sync(i915, batch);
+	gem_close(i915, batch);
+
+	kill_children(SIGINT);
+	igt_waitchildren();
+
+	gem_close(i915, common);
+	munmap(reloc, reloc_sz);
+}
+
 #define CONCURRENT 1024
 
 static uint64_t concurrent_relocs(int i915, int idx, int count)
@@ -1284,6 +1493,16 @@ igt_main
 		}
 	}
 
+	igt_subtest_with_dynamic("basic-wide-active") {
+		__for_each_physical_engine(fd, e) {
+			igt_dynamic_f("%s", e->name)
+				wide_active(fd, e->flags);
+		}
+	}
+
+	igt_subtest("basic-parallel")
+		parallel(fd);
+
 	igt_subtest("basic-concurrent0")
 		concurrent(fd, 0);
 	igt_subtest("basic-concurrent16")