Igt-dev Archive on lore.kernel.org
 help / color / mirror / Atom feed
* [PATCH i-g-t 1/2] tests/amdgpu/amd_deadlock: add gfx user-queue clean-hang reset test
@ 2026-07-20  9:55 Jesse Zhang
  2026-07-20  9:55 ` [PATCH i-g-t 2/2] tests/amdgpu/amd_deadlock: add gfx user-queue priv-fault " Jesse Zhang
                   ` (4 more replies)
  0 siblings, 5 replies; 6+ messages in thread
From: Jesse Zhang @ 2026-07-20  9:55 UTC (permalink / raw)
  To: igt-dev
  Cc: Vitaly Prosyak, Alex Deucher, Christian Koenig, Jesse Zhang,
	Jesse Zhang

Submit a gfx user-queue job whose only packet is a WAIT_REG_MEM that never
completes (polls a resident BO for a value that never arrives). The queue
hangs with no fault, so the driver recovers it with a per-queue reset. Add
it as the amdgpu-deadlock-gfx-umq subtest, gated on AMDGPU_ENABLE_USERQTEST
and per-queue reset capability.

Signed-off-by: Jesse Zhang <jesse.zhang@amd.com>
---
 lib/amdgpu/amd_deadlock_helpers.c | 91 +++++++++++++++++++++++++++++++
 lib/amdgpu/amd_deadlock_helpers.h |  4 ++
 tests/amdgpu/amd_deadlock.c       |  9 +++
 3 files changed, 104 insertions(+)

diff --git a/lib/amdgpu/amd_deadlock_helpers.c b/lib/amdgpu/amd_deadlock_helpers.c
index 06c577085..c679e60c7 100644
--- a/lib/amdgpu/amd_deadlock_helpers.c
+++ b/lib/amdgpu/amd_deadlock_helpers.c
@@ -412,6 +412,97 @@ bad_access_helper(amdgpu_device_handle device_handle, unsigned int cmd_error,
 	}
 }
 
+/*
+ * Emit a WAIT_REG_MEM that waits forever: poll bo_mc (initialised to 0) for a
+ * value != 0. This is a valid packet that never completes, i.e. a clean hang
+ * with no page fault, so it can be recovered by a per-queue reset.
+ */
+static void gfx_ring_emit_wait_reg_mem_hang(struct amdgpu_ring_context *ring_context)
+{
+	uint32_t i = 0;
+
+	ring_context->pm4[i++] = PACKET3(PACKET3_WAIT_REG_MEM, 5);
+	ring_context->pm4[i++] = WAIT_REG_MEM_MEM_SPACE(1)  | /* memory */
+				 WAIT_REG_MEM_FUNCTION(4)  | /* != */
+				 WAIT_REG_MEM_ENGINE(0);     /* me */
+	ring_context->pm4[i++] = ring_context->bo_mc & 0xfffffffc;
+	ring_context->pm4[i++] = (ring_context->bo_mc >> 32) & 0xffffffff;
+	ring_context->pm4[i++] = 0;          /* reference value */
+	ring_context->pm4[i++] = 0xffffffff; /* and mask */
+	ring_context->pm4[i++] = 0x00000004; /* poll interval */
+	ring_context->pm4_dw = i;
+}
+
+/*
+ * Submit a single job that hangs the queue on a WAIT_REG_MEM (no fault) and
+ * let per-queue reset recover it. Uses the normal (synchronised) submit path
+ * so the CP actually runs the packet.
+ */
+void amdgpu_hang_ring_helper(amdgpu_device_handle device_handle, unsigned int ip_type,
+			     struct pci_addr *pci, bool user_queue)
+{
+	const struct amdgpu_ip_block_version *ip_block;
+	const int write_length = 128;
+	const int pm4_dw = 256;
+	struct amdgpu_ring_context *ring_context;
+	int r = 0;
+
+	ip_block = get_ip_block(device_handle, ip_type);
+	ring_context = calloc(1, sizeof(*ring_context));
+	igt_assert(ring_context);
+
+	if (user_queue) {
+		ip_block->funcs->userq_create(device_handle, ring_context, ip_type);
+	} else {
+		r = amdgpu_cs_ctx_create(device_handle, &ring_context->context_handle);
+		igt_assert_eq(r, 0);
+	}
+
+	ring_context->write_length = write_length;
+	ring_context->pm4 = calloc(pm4_dw, sizeof(*ring_context->pm4));
+	ring_context->pm4_size = pm4_dw;
+	ring_context->res_cnt = 1;
+	ring_context->ring_id = 0;
+	ring_context->user_queue = user_queue;
+	igt_assert(ring_context->pm4);
+
+	r = amdgpu_bo_alloc_and_map_sync(device_handle,
+				    ring_context->write_length * sizeof(uint32_t),
+				    4096, AMDGPU_GEM_DOMAIN_GTT,
+				    AMDGPU_GEM_CREATE_CPU_GTT_USWC,
+				    AMDGPU_VM_MTYPE_UC,
+				    &ring_context->bo,
+				    (void **)&ring_context->bo_cpu,
+				    &ring_context->bo_mc,
+				    &ring_context->va_handle,
+				    ring_context->timeline_syncobj_handle,
+				    ++ring_context->point, user_queue);
+	igt_assert_eq(r, 0);
+	if (user_queue) {
+		r = amdgpu_timeline_syncobj_wait(device_handle,
+			ring_context->timeline_syncobj_handle,
+			ring_context->point);
+		igt_assert_eq(r, 0);
+	}
+
+	/* the memory the queue polls stays 0, so the queue hangs */
+	memset((void *)ring_context->bo_cpu, 0, ring_context->write_length * sizeof(uint32_t));
+	ring_context->resources[0] = ring_context->bo;
+
+	gfx_ring_emit_wait_reg_mem_hang(ring_context);
+
+	amdgpu_test_exec_cs_helper(device_handle, ip_block->type, ring_context, 0);
+
+	amdgpu_bo_unmap_and_free(ring_context->bo, ring_context->va_handle, ring_context->bo_mc,
+				 ring_context->write_length * sizeof(uint32_t));
+	if (user_queue) {
+		ip_block->funcs->userq_destroy(device_handle, ring_context, ip_type);
+	} else {
+		free(ring_context->pm4);
+		free(ring_context);
+	}
+}
+
 #define MAX_DMABUF_COUNT 0x20000
 #define MAX_DWORD_COUNT 256
 
diff --git a/lib/amdgpu/amd_deadlock_helpers.h b/lib/amdgpu/amd_deadlock_helpers.h
index c46ebb67b..accbf2e41 100644
--- a/lib/amdgpu/amd_deadlock_helpers.h
+++ b/lib/amdgpu/amd_deadlock_helpers.h
@@ -33,5 +33,9 @@ bad_access_ring_helper(amdgpu_device_handle device_handle, unsigned int cmd_erro
 
 void
 amdgpu_hang_sdma_ring_helper(amdgpu_device_handle device_handle, uint8_t hang_type, struct pci_addr *pci);
+
+void
+amdgpu_hang_ring_helper(amdgpu_device_handle device_handle, unsigned int ip_type,
+			struct pci_addr *pci, bool user_queue);
 #endif
 
diff --git a/tests/amdgpu/amd_deadlock.c b/tests/amdgpu/amd_deadlock.c
index 5325c0ade..118948992 100644
--- a/tests/amdgpu/amd_deadlock.c
+++ b/tests/amdgpu/amd_deadlock.c
@@ -246,6 +246,15 @@ int igt_main()
 			amdgpu_wait_memory_helper(device, AMDGPU_HW_IP_DMA, &pci, true);
 		}
 	}
+
+	igt_describe("Test-per-queue-reset-of-a-cleanly-hung-gfx-user-queue");
+	igt_subtest_with_dynamic("amdgpu-deadlock-gfx-umq") {
+		if (enable_test && userq_arr_cap[AMD_IP_GFX] &&
+			is_reset_enable(AMD_IP_GFX, AMDGPU_RESET_TYPE_PER_QUEUE, &pci)) {
+			igt_dynamic_f("amdgpu-deadlock-gfx-umq")
+			amdgpu_hang_ring_helper(device, AMDGPU_HW_IP_GFX, &pci, true);
+		}
+	}
 #endif
 
 	igt_fixture() {
-- 
2.49.0


^ permalink raw reply related	[flat|nested] 6+ messages in thread
end of thread, other threads:[~2026-07-21 14:29 UTC | newest]

Thread overview: 6+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-07-20  9:55 [PATCH i-g-t 1/2] tests/amdgpu/amd_deadlock: add gfx user-queue clean-hang reset test Jesse Zhang
2026-07-20  9:55 ` [PATCH i-g-t 2/2] tests/amdgpu/amd_deadlock: add gfx user-queue priv-fault " Jesse Zhang
2026-07-21  1:46 ` ✓ Xe.CI.BAT: success for series starting with [i-g-t,1/2] tests/amdgpu/amd_deadlock: add gfx user-queue clean-hang " Patchwork
2026-07-21  2:03 ` ✓ i915.CI.BAT: " Patchwork
2026-07-21  9:47 ` ✗ Xe.CI.FULL: failure " Patchwork
2026-07-21 14:29 ` ✗ i915.CI.Full: " Patchwork

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox