@@ -472,6 +472,18 @@ static void amdgpu_disable_suites()
"compute ring bad dispatch test (set amdgpu.lockup_timeout=50)", CU_FALSE))
fprintf(stderr, "test deactivation failed - %s\n", CU_get_error_msg());
+ /* This test was ran on GFX9 only */
+ //if (family_id < AMDGPU_FAMILY_AI || family_id > AMDGPU_FAMILY_RV)
+ if (amdgpu_set_test_active(DEADLOCK_TESTS_STR,
+ "gfx ring bad slow dispatch test (set amdgpu.lockup_timeout=50)", CU_FALSE))
+ fprintf(stderr, "test deactivation failed - %s\n", CU_get_error_msg());
+
+ /* This test was ran on GFX9 only */
+ //if (family_id < AMDGPU_FAMILY_AI || family_id > AMDGPU_FAMILY_RV)
+ if (amdgpu_set_test_active(DEADLOCK_TESTS_STR,
+ "compute ring bad slow dispatch test (set amdgpu.lockup_timeout=50,50)", CU_FALSE))
+ fprintf(stderr, "test deactivation failed - %s\n", CU_get_error_msg());
+
if (amdgpu_set_test_active(BO_TESTS_STR, "Metadata", CU_FALSE))
fprintf(stderr, "test deactivation failed - %s\n", CU_get_error_msg());
@@ -242,6 +242,7 @@ CU_BOOL suite_syncobj_timeline_tests_enable(void);
extern CU_TestInfo syncobj_timeline_tests[];
void amdgpu_dispatch_hang_helper(amdgpu_device_handle device_handle, uint32_t ip_type);
+void amdgpu_dispatch_hang_slow_helper(amdgpu_device_handle device_handle, uint32_t ip_type);
/**
* Helper functions
@@ -312,7 +312,8 @@ static uint32_t shader_bin[] = {
enum cs_type {
CS_BUFFERCLEAR,
CS_BUFFERCOPY,
- CS_HANG
+ CS_HANG,
+ CS_HANG_SLOW
};
static const uint32_t bufferclear_cs_shader_gfx9[] = {
@@ -482,6 +483,37 @@ unsigned int memcpy_ps_hang[] = {
0xF800180F, 0x03020100, 0xBF810000
};
+struct amdgpu_test_shader {
+ uint32_t *shader;
+ uint32_t header_length;
+ uint32_t body_length;
+ uint32_t foot_length;
+};
+
+unsigned int memcpy_cs_hang_slow_ai_codes[] = {
+ 0xd1fd0000, 0x04010c08, 0xe00c2000, 0x80000100,
+ 0xbf8c0f70, 0xe01c2000, 0x80010100, 0xbf810000
+};
+
+struct amdgpu_test_shader memcpy_cs_hang_slow_ai = {
+ memcpy_cs_hang_slow_ai_codes,
+ 4,
+ 3,
+ 1
+};
+
+unsigned int memcpy_cs_hang_slow_rv_codes[] = {
+ 0x8e00860c, 0x32000000, 0xe00c2000, 0x80010100,
+ 0xbf8c0f70, 0xe01c2000, 0x80020100, 0xbf810000
+};
+
+struct amdgpu_test_shader memcpy_cs_hang_slow_rv = {
+ memcpy_cs_hang_slow_rv_codes,
+ 4,
+ 3,
+ 1
+};
+
int amdgpu_bo_alloc_and_map_raw(amdgpu_device_handle dev, unsigned size,
unsigned alignment, unsigned heap, uint64_t alloc_flags,
uint64_t mapping_flags, amdgpu_bo_handle *bo, void **cpu,
@@ -2183,6 +2215,37 @@ static void amdgpu_sync_dependency_test(void)
free(ibs_request.dependencies);
}
+static int amdgpu_dispatch_load_cs_shader_hang_slow(uint32_t *ptr, int family)
+{
+ struct amdgpu_test_shader *shader;
+ int i, loop = 0x10000;
+
+ switch (family) {
+ case AMDGPU_FAMILY_AI:
+ shader = &memcpy_cs_hang_slow_ai;
+ break;
+ case AMDGPU_FAMILY_RV:
+ shader = &memcpy_cs_hang_slow_rv;
+ break;
+ default:
+ return -1;
+ break;
+ }
+
+ memcpy(ptr, shader->shader, shader->header_length * sizeof(uint32_t));
+
+ for (i = 0; i < loop; i++)
+ memcpy(ptr + shader->header_length + shader->body_length * i,
+ shader->shader + shader->header_length,
+ shader->body_length * sizeof(uint32_t));
+
+ memcpy(ptr + shader->header_length + shader->body_length * loop,
+ shader->shader + shader->header_length + shader->body_length,
+ shader->foot_length * sizeof(uint32_t));
+
+ return 0;
+}
+
static int amdgpu_dispatch_load_cs_shader(uint8_t *ptr,
int cs_type)
{
@@ -2635,6 +2698,170 @@ void amdgpu_dispatch_hang_helper(amdgpu_device_handle device_handle, uint32_t ip
}
}
+static void amdgpu_memcpy_dispatch_hang_slow_test(amdgpu_device_handle device_handle,
+ uint32_t ip_type, uint32_t ring)
+{
+ amdgpu_context_handle context_handle;
+ amdgpu_bo_handle bo_src, bo_dst, bo_shader, bo_cmd, resources[4];
+ volatile unsigned char *ptr_dst;
+ void *ptr_shader;
+ unsigned char *ptr_src;
+ uint32_t *ptr_cmd;
+ uint64_t mc_address_src, mc_address_dst, mc_address_shader, mc_address_cmd;
+ amdgpu_va_handle va_src, va_dst, va_shader, va_cmd;
+ int i, r;
+ int bo_dst_size = 0x4000000;
+ int bo_shader_size = 0x400000;
+ int bo_cmd_size = 4096;
+ struct amdgpu_cs_request ibs_request = {0};
+ struct amdgpu_cs_ib_info ib_info= {0};
+ uint32_t hang_state, hangs, expired;
+ struct amdgpu_gpu_info gpu_info = {0};
+ amdgpu_bo_list_handle bo_list;
+ struct amdgpu_cs_fence fence_status = {0};
+
+ r = amdgpu_query_gpu_info(device_handle, &gpu_info);
+ CU_ASSERT_EQUAL(r, 0);
+
+ r = amdgpu_cs_ctx_create(device_handle, &context_handle);
+ CU_ASSERT_EQUAL(r, 0);
+
+ r = amdgpu_bo_alloc_and_map(device_handle, bo_cmd_size, 4096,
+ AMDGPU_GEM_DOMAIN_GTT, 0,
+ &bo_cmd, (void **)&ptr_cmd,
+ &mc_address_cmd, &va_cmd);
+ CU_ASSERT_EQUAL(r, 0);
+ memset(ptr_cmd, 0, bo_cmd_size);
+
+ r = amdgpu_bo_alloc_and_map(device_handle, bo_shader_size, 4096,
+ AMDGPU_GEM_DOMAIN_VRAM, 0,
+ &bo_shader, &ptr_shader,
+ &mc_address_shader, &va_shader);
+ CU_ASSERT_EQUAL(r, 0);
+ memset(ptr_shader, 0, bo_shader_size);
+
+ r = amdgpu_dispatch_load_cs_shader_hang_slow(ptr_shader, gpu_info.family_id);
+ CU_ASSERT_EQUAL(r, 0);
+
+ r = amdgpu_bo_alloc_and_map(device_handle, bo_dst_size, 4096,
+ AMDGPU_GEM_DOMAIN_VRAM, 0,
+ &bo_src, (void **)&ptr_src,
+ &mc_address_src, &va_src);
+ CU_ASSERT_EQUAL(r, 0);
+
+ r = amdgpu_bo_alloc_and_map(device_handle, bo_dst_size, 4096,
+ AMDGPU_GEM_DOMAIN_VRAM, 0,
+ &bo_dst, (void **)&ptr_dst,
+ &mc_address_dst, &va_dst);
+ CU_ASSERT_EQUAL(r, 0);
+
+ memset(ptr_src, 0x55, bo_dst_size);
+
+ i = 0;
+ i += amdgpu_dispatch_init(ptr_cmd + i, ip_type);
+
+ /* Issue commands to set cu mask used in current dispatch */
+ i += amdgpu_dispatch_write_cumask(ptr_cmd + i);
+
+ /* Writes shader state to HW */
+ i += amdgpu_dispatch_write2hw(ptr_cmd + i, mc_address_shader);
+
+ /* Write constant data */
+ /* Writes the texture resource constants data to the SGPRs */
+ ptr_cmd[i++] = PACKET3_COMPUTE(PKT3_SET_SH_REG, 4);
+ ptr_cmd[i++] = 0x240;
+ ptr_cmd[i++] = mc_address_src;
+ ptr_cmd[i++] = (mc_address_src >> 32) | 0x100000;
+ ptr_cmd[i++] = 0x400000;
+ ptr_cmd[i++] = 0x74fac;
+
+ /* Writes the UAV constant data to the SGPRs. */
+ ptr_cmd[i++] = PACKET3_COMPUTE(PKT3_SET_SH_REG, 4);
+ ptr_cmd[i++] = 0x244;
+ ptr_cmd[i++] = mc_address_dst;
+ ptr_cmd[i++] = (mc_address_dst >> 32) | 0x100000;
+ ptr_cmd[i++] = 0x400000;
+ ptr_cmd[i++] = 0x74fac;
+
+ /* dispatch direct command */
+ ptr_cmd[i++] = PACKET3_COMPUTE(PACKET3_DISPATCH_DIRECT, 3);
+ ptr_cmd[i++] = 0x10000;
+ ptr_cmd[i++] = 1;
+ ptr_cmd[i++] = 1;
+ ptr_cmd[i++] = 1;
+
+ while (i & 7)
+ ptr_cmd[i++] = 0xffff1000; /* type3 nop packet */
+
+ resources[0] = bo_shader;
+ resources[1] = bo_src;
+ resources[2] = bo_dst;
+ resources[3] = bo_cmd;
+ r = amdgpu_bo_list_create(device_handle, 4, resources, NULL, &bo_list);
+ CU_ASSERT_EQUAL(r, 0);
+
+ ib_info.ib_mc_address = mc_address_cmd;
+ ib_info.size = i;
+ ibs_request.ip_type = ip_type;
+ ibs_request.ring = ring;
+ ibs_request.resources = bo_list;
+ ibs_request.number_of_ibs = 1;
+ ibs_request.ibs = &ib_info;
+ ibs_request.fence_info.handle = NULL;
+ r = amdgpu_cs_submit(context_handle, 0, &ibs_request, 1);
+ CU_ASSERT_EQUAL(r, 0);
+
+ fence_status.ip_type = ip_type;
+ fence_status.ip_instance = 0;
+ fence_status.ring = ring;
+ fence_status.context = context_handle;
+ fence_status.fence = ibs_request.seq_no;
+
+ /* wait for IB accomplished */
+ r = amdgpu_cs_query_fence_status(&fence_status,
+ AMDGPU_TIMEOUT_INFINITE,
+ 0, &expired);
+
+ r = amdgpu_cs_query_reset_state(context_handle, &hang_state, &hangs);
+ CU_ASSERT_EQUAL(r, 0);
+ CU_ASSERT_EQUAL(hang_state, AMDGPU_CTX_UNKNOWN_RESET);
+
+ r = amdgpu_bo_list_destroy(bo_list);
+ CU_ASSERT_EQUAL(r, 0);
+
+ r = amdgpu_bo_unmap_and_free(bo_src, va_src, mc_address_src, bo_dst_size);
+ CU_ASSERT_EQUAL(r, 0);
+ r = amdgpu_bo_unmap_and_free(bo_dst, va_dst, mc_address_dst, bo_dst_size);
+ CU_ASSERT_EQUAL(r, 0);
+
+ r = amdgpu_bo_unmap_and_free(bo_cmd, va_cmd, mc_address_cmd, bo_cmd_size);
+ CU_ASSERT_EQUAL(r, 0);
+
+ r = amdgpu_bo_unmap_and_free(bo_shader, va_shader, mc_address_shader, bo_shader_size);
+ CU_ASSERT_EQUAL(r, 0);
+
+ r = amdgpu_cs_ctx_free(context_handle);
+ CU_ASSERT_EQUAL(r, 0);
+}
+
+void amdgpu_dispatch_hang_slow_helper(amdgpu_device_handle device_handle, uint32_t ip_type)
+{
+ int r;
+ struct drm_amdgpu_info_hw_ip info;
+ uint32_t ring_id;
+
+ r = amdgpu_query_hw_ip_info(device_handle, ip_type, 0, &info);
+ CU_ASSERT_EQUAL(r, 0);
+ if (!info.available_rings)
+ printf("SKIP ... as there's no ring for ip %d\n", ip_type);
+
+ for (ring_id = 0; (1 << ring_id) & info.available_rings; ring_id++) {
+ amdgpu_memcpy_dispatch_test(device_handle, ip_type, ring_id, 0);
+ amdgpu_memcpy_dispatch_hang_slow_test(device_handle, ip_type, ring_id);
+ amdgpu_memcpy_dispatch_test(device_handle, ip_type, ring_id, 0);
+ }
+}
+
static int amdgpu_draw_load_ps_shader(uint8_t *ptr, int ps_type)
{
int i;
@@ -116,6 +116,8 @@ static void amdgpu_illegal_mem_access();
static void amdgpu_deadlock_sdma(void);
static void amdgpu_dispatch_hang_gfx(void);
static void amdgpu_dispatch_hang_compute(void);
+static void amdgpu_dispatch_hang_slow_gfx(void);
+static void amdgpu_dispatch_hang_slow_compute(void);
CU_BOOL suite_deadlock_tests_enable(void)
{
@@ -192,6 +194,8 @@ CU_TestInfo deadlock_tests[] = {
{ "illegal mem access test (set amdgpu.vm_fault_stop=2)", amdgpu_illegal_mem_access },
{ "gfx ring bad dispatch test (set amdgpu.lockup_timeout=50)", amdgpu_dispatch_hang_gfx },
{ "compute ring bad dispatch test (set amdgpu.lockup_timeout=50,50)", amdgpu_dispatch_hang_compute },
+ { "gfx ring bad slow dispatch test (set amdgpu.lockup_timeout=50)", amdgpu_dispatch_hang_slow_gfx },
+ { "compute ring bad slow dispatch test (set amdgpu.lockup_timeout=50,50)", amdgpu_dispatch_hang_slow_compute },
CU_TEST_INFO_NULL,
};
@@ -502,3 +506,13 @@ static void amdgpu_dispatch_hang_compute(void)
{
amdgpu_dispatch_hang_helper(device_handle, AMDGPU_HW_IP_COMPUTE);
}
+
+static void amdgpu_dispatch_hang_slow_gfx(void)
+{
+ amdgpu_dispatch_hang_slow_helper(device_handle, AMDGPU_HW_IP_GFX);
+}
+
+static void amdgpu_dispatch_hang_slow_compute(void)
+{
+ amdgpu_dispatch_hang_slow_helper(device_handle, AMDGPU_HW_IP_COMPUTE);
+}
add gfx/compute bad slow dispatch test for gfx9 Signed-off-by: Flora Cui <flora.cui@amd.com> --- tests/amdgpu/amdgpu_test.c | 12 ++ tests/amdgpu/amdgpu_test.h | 1 + tests/amdgpu/basic_tests.c | 229 +++++++++++++++++++++++++++++++++- tests/amdgpu/deadlock_tests.c | 14 +++ 4 files changed, 255 insertions(+), 1 deletion(-)