*/
#include "gt/intel_migrate.h"
+#include "gt/intel_gpu_commands.h"
#include "gem/i915_gem_ttm_move.h"
+#include "i915_deps.h"
+
+#include "selftests/igt_spinner.h"
+
static int igt_fill_check_buffer(struct drm_i915_gem_object *obj,
bool fill)
{
}
static int lmem_pages_migrate_one(struct i915_gem_ww_ctx *ww,
- struct drm_i915_gem_object *obj)
+ struct drm_i915_gem_object *obj,
+ struct i915_vma *vma)
{
int err;
if (err)
return err;
+ if (vma) {
+ err = i915_vma_pin_ww(vma, ww, obj->base.size, 0,
+ 0UL | PIN_OFFSET_FIXED |
+ PIN_USER);
+ if (err) {
+ if (err != -EINTR && err != ERESTARTSYS &&
+ err != -EDEADLK)
+ pr_err("Failed to pin vma.\n");
+ return err;
+ }
+
+ i915_vma_unpin(vma);
+ }
+
+ /*
+ * Migration will implicitly unbind (asynchronously) any bound
+ * vmas.
+ */
if (i915_gem_object_is_lmem(obj)) {
err = i915_gem_object_migrate(obj, ww, INTEL_REGION_SMEM);
if (err) {
return err;
}
-static int igt_lmem_pages_migrate(void *arg)
+static int __igt_lmem_pages_migrate(struct intel_gt *gt,
+ struct i915_address_space *vm,
+ struct i915_deps *deps,
+ struct igt_spinner *spin,
+ struct dma_fence *spin_fence)
{
- struct intel_gt *gt = arg;
struct drm_i915_private *i915 = gt->i915;
struct drm_i915_gem_object *obj;
+ struct i915_vma *vma = NULL;
struct i915_gem_ww_ctx ww;
struct i915_request *rq;
int err;
if (IS_ERR(obj))
return PTR_ERR(obj);
+ if (vm) {
+ vma = i915_vma_instance(obj, vm, NULL);
+ if (IS_ERR(vma)) {
+ err = PTR_ERR(vma);
+ goto out_put;
+ }
+ }
+
/* Initial GPU fill, sync, CPU initialization. */
for_i915_gem_ww(&ww, err, true) {
err = i915_gem_object_lock(obj, &ww);
if (err)
continue;
- err = intel_migrate_clear(>->migrate, &ww, NULL,
+ err = intel_migrate_clear(>->migrate, &ww, deps,
obj->mm.pages->sgl, obj->cache_level,
i915_gem_object_is_lmem(obj),
0xdeadbeaf, &rq);
if (rq) {
dma_resv_add_excl_fence(obj->base.resv, &rq->fence);
+ i915_gem_object_set_moving_fence(obj, &rq->fence);
i915_request_put(rq);
}
if (err)
continue;
- err = i915_gem_object_wait(obj, I915_WAIT_INTERRUPTIBLE,
- 5 * HZ);
- if (err)
- continue;
-
- err = igt_fill_check_buffer(obj, true);
- if (err)
- continue;
+ if (!vma) {
+ err = igt_fill_check_buffer(obj, true);
+ if (err)
+ continue;
+ }
}
if (err)
goto out_put;
*/
for (i = 1; i <= 5; ++i) {
for_i915_gem_ww(&ww, err, true)
- err = lmem_pages_migrate_one(&ww, obj);
+ err = lmem_pages_migrate_one(&ww, obj, vma);
if (err)
goto out_put;
}
if (err)
goto out_put;
+ if (spin) {
+ if (dma_fence_is_signaled(spin_fence)) {
+ pr_err("Spinner was terminated by hangcheck.\n");
+ err = -EBUSY;
+ goto out_unlock;
+ }
+ igt_spinner_end(spin);
+ }
+
/* Finally sync migration and check content. */
err = i915_gem_object_wait_migration(obj, true);
if (err)
goto out_unlock;
- err = igt_fill_check_buffer(obj, false);
+ if (vma) {
+ err = i915_vma_wait_for_bind(vma);
+ if (err)
+ goto out_unlock;
+ } else {
+ err = igt_fill_check_buffer(obj, false);
+ }
out_unlock:
i915_gem_object_unlock(obj);
static int igt_lmem_pages_failsafe_migrate(void *arg)
{
int fail_gpu, fail_alloc, ret;
+ struct intel_gt *gt = arg;
for (fail_gpu = 0; fail_gpu < 2; ++fail_gpu) {
for (fail_alloc = 0; fail_alloc < 2; ++fail_alloc) {
fail_gpu, fail_alloc);
i915_ttm_migrate_set_failure_modes(fail_gpu,
fail_alloc);
- ret = igt_lmem_pages_migrate(arg);
+ ret = __igt_lmem_pages_migrate(gt, NULL, NULL, NULL, NULL);
+ if (ret)
+ goto out_err;
+ }
+ }
+
+out_err:
+ i915_ttm_migrate_set_failure_modes(false, false);
+ return ret;
+}
+
+/*
+ * This subtest tests that unbinding at migration is indeed performed
+ * async. We launch a spinner and a number of migrations depending on
+ * that spinner to have terminated. Before each migration we bind a
+ * vma, which should then be async unbound by the migration operation.
+ * If we are able to schedule migrations without blocking while the
+ * spinner is still running, those unbinds are indeed async and non-
+ * blocking.
+ *
+ * Note that each async bind operation is awaiting the previous migration
+ * due to the moving fence resulting from the migration.
+ */
+static int igt_async_migrate(struct intel_gt *gt)
+{
+ struct intel_engine_cs *engine;
+ enum intel_engine_id id;
+ struct i915_ppgtt *ppgtt;
+ struct igt_spinner spin;
+ int err;
+
+ ppgtt = i915_ppgtt_create(gt, 0);
+ if (IS_ERR(ppgtt))
+ return PTR_ERR(ppgtt);
+
+ if (igt_spinner_init(&spin, gt)) {
+ err = -ENOMEM;
+ goto out_spin;
+ }
+
+ for_each_engine(engine, gt, id) {
+ struct ttm_operation_ctx ctx = {
+ .interruptible = true
+ };
+ struct dma_fence *spin_fence;
+ struct intel_context *ce;
+ struct i915_request *rq;
+ struct i915_deps deps;
+
+ ce = intel_context_create(engine);
+ if (IS_ERR(ce)) {
+ err = PTR_ERR(ce);
+ goto out_ce;
+ }
+
+ /*
+ * Use MI_NOOP, making the spinner non-preemptible. If there
+ * is a code path where we fail async operation due to the
+ * running spinner, we will block and fail to end the
+ * spinner resulting in a deadlock. But with a non-
+ * preemptible spinner, hangcheck will terminate the spinner
+ * for us, and we will later detect that and fail the test.
+ */
+ rq = igt_spinner_create_request(&spin, ce, MI_NOOP);
+ intel_context_put(ce);
+ if (IS_ERR(rq)) {
+ err = PTR_ERR(rq);
+ goto out_ce;
+ }
+
+ i915_deps_init(&deps, GFP_KERNEL);
+ err = i915_deps_add_dependency(&deps, &rq->fence, &ctx);
+ spin_fence = dma_fence_get(&rq->fence);
+ i915_request_add(rq);
+ if (err)
+ goto out_ce;
+
+ err = __igt_lmem_pages_migrate(gt, &ppgtt->vm, &deps, &spin,
+ spin_fence);
+ i915_deps_fini(&deps);
+ dma_fence_put(spin_fence);
+ if (err)
+ goto out_ce;
+ }
+
+out_ce:
+ igt_spinner_fini(&spin);
+out_spin:
+ i915_vm_put(&ppgtt->vm);
+
+ return err;
+}
+
+/*
+ * Setting ASYNC_FAIL_ALLOC to 2 will simulate memory allocation failure while
+ * arming the migration error check and block async migration. This
+ * will cause us to deadlock and hangcheck will terminate the spinner
+ * causing the test to fail.
+ */
+#define ASYNC_FAIL_ALLOC 1
+static int igt_lmem_async_migrate(void *arg)
+{
+ int fail_gpu, fail_alloc, ret;
+ struct intel_gt *gt = arg;
+
+ for (fail_gpu = 0; fail_gpu < 2; ++fail_gpu) {
+ for (fail_alloc = 0; fail_alloc < ASYNC_FAIL_ALLOC; ++fail_alloc) {
+ pr_info("Simulated failure modes: gpu: %d, alloc: %d\n",
+ fail_gpu, fail_alloc);
+ i915_ttm_migrate_set_failure_modes(fail_gpu,
+ fail_alloc);
+ ret = igt_async_migrate(gt);
if (ret)
goto out_err;
}
SUBTEST(igt_lmem_create_migrate),
SUBTEST(igt_same_create_migrate),
SUBTEST(igt_lmem_pages_failsafe_migrate),
+ SUBTEST(igt_lmem_async_migrate),
};
if (!HAS_LMEM(i915))