kfdtest: add hit trap event test

Check that a jump to trap event can be picked up by the debugger. Signed-off-by: Jonathan Kim <jonathan.kim@amd.com> Change-Id: Iad5f87092f2b82d5018013bba548979122a9bd02 [ROCm/ROCR-Runtime commit: b77189cf83]
2022-04-25 11:19:09 -04:00
@@ -237,3 +237,78 @@ exit:
    LOG() << std::endl;
    TEST_END
 }
+
+TEST_F(KFDDBGTest, HitTrapEvent) {
+    TEST_START(TESTPROFILE_RUNALL)
+    if (m_FamilyId >= FAMILY_AI) {
+        int defaultGPUNode = m_NodeInfo.HsaDefaultGPUNode();
+
+        if (hsaKmtCheckRuntimeDebugSupport()) {
+            LOG() << "Skip test as debug API not supported";
+            goto exit;
+        }
+
+        ASSERT_GE(defaultGPUNode, 0) << "failed to get default GPU Node";
+
+        // create shader and trap bufs then enable 2nd level trap
+        HsaMemoryBuffer isaBuf(PAGE_SIZE, defaultGPUNode, true, false, true);
+        HsaMemoryBuffer trapStatusBuf(PAGE_SIZE, defaultGPUNode, true, false, false);
+
+        HsaMemoryBuffer trap(PAGE_SIZE*2, defaultGPUNode, true, false, true);
+        HsaMemoryBuffer tmaBuf(PAGE_SIZE, defaultGPUNode, false, false, false);
+
+        ASSERT_SUCCESS(hsaKmtSetTrapHandler(defaultGPUNode,
+                                            trap.As<void *>(),
+                                            0x1000,
+                                            tmaBuf.As<void*>(),
+                                            0x1000));
+
+        // compile and dispatch shader
+        ASSERT_SUCCESS(m_pAsm->RunAssembleBuf(jump_to_trap_gfx, isaBuf.As<char*>()));
+        ASSERT_SUCCESS(m_pAsm->RunAssembleBuf(trap_handler_gfx, trap.As<char*>()));
+
+        uint32_t rDebug;
+        ASSERT_SUCCESS(hsaKmtRuntimeEnable(&rDebug, true));
+
+        BaseDebug *debug = new BaseDebug();
+        struct kfd_runtime_info r_info = {0};
+        ASSERT_SUCCESS(debug->Attach(&r_info, sizeof(r_info), getpid(), 0));
+        ASSERT_EQ(r_info.runtime_state, DEBUG_RUNTIME_STATE_ENABLED);
+
+        PM4Queue queue;
+        HsaQueueResource *qResources;
+        ASSERT_SUCCESS(queue.Create(defaultGPUNode));
+
+        unsigned int* trapStatus = trapStatusBuf.As<unsigned int*>();
+        trapStatus[0] = 0;
+        Dispatch *dispatch;
+        dispatch = new Dispatch(isaBuf);
+        dispatch->SetArgs(&trapStatus[0], NULL);
+        dispatch->SetDim(1, 1, 1);
+
+        /* Subscribe to trap events and submit the queue */
+        uint64_t trapMask = KFD_EC_MASK(EC_QUEUE_WAVE_TRAP);
+        debug->SetExceptionsEnabled(trapMask);
+        dispatch->Submit(queue);
+
+        /* Wait for trap event */
+        uint32_t QueueId = -1;
+        ASSERT_SUCCESS(debug->QueryDebugEvent(&trapMask, NULL, &QueueId, 5000));
+        ASSERT_NE(QueueId, -1);
+        ASSERT_EQ(trapMask, KFD_EC_MASK(EC_QUEUE_WAVE_TRAP) | KFD_EC_MASK(EC_QUEUE_NEW));
+
+        dispatch->Sync();
+        EXPECT_SUCCESS(queue.Destroy());
+
+        ASSERT_NE(trapStatus[0], 0);
+
+        debug->Detach();
+        hsaKmtRuntimeDisable();
+    } else {
+        LOG() << "Skipping test: Test not supported on family ID 0x"
+              << m_FamilyId << "." << std::endl;
+    }
+exit:
+    LOG() << std::endl;
+    TEST_END
+}
@@ -80,6 +80,13 @@ const std::vector<const char*> ShaderList = {
    "       .else\n"\
    "           v_cmp_lt_u32        vcc, \\src0, \\vsrc1\n"\
    "       .endif\n"\
+    "   .endm\n"\
+    "   .macro V_CMP_EQ_U32 src0, vsrc1\n"\
+    "       .if (.amdgcn.gfx_generation_number >= 10)\n"\
+    "           v_cmp_eq_u32        vcc_lo, \\src0, \\vsrc1\n"\
+    "       .else\n"\
+    "           v_cmp_eq_u32        vcc, \\src0, \\vsrc1\n"\
+    "       .endif\n"\
    "   .endm\n"

 /* Macros for portable flat load/store/atomic instructions.
@@ -893,3 +900,74 @@ const char *GwsAtomicIncreaseIsa =
        s_waitcnt 0
        s_endpgm
 )";
+
+const char *jump_to_trap_gfx =
+    SHADER_START
+    SHADER_MACROS_U32
+    R"(
+        /*copy the parameters from scalar registers to vector registers*/
+        v_mov_b32 v4, 0
+        v_mov_b32 v0, s0
+        v_mov_b32 v1, s1
+        s_trap 1
+        EXIT_LOOP:
+        V_CMP_EQ_U32 v4, 0
+        s_cbranch_vccnz EXIT_LOOP
+        flat_store_dword v[0:1], v4
+        s_waitcnt vmcnt(0)&lgkmcnt(0)
+        s_endpgm
+)";
+
+const char *trap_handler_gfx =
+    SHADER_START
+    R"(
+        CHECK_VMFAULT:
+        /*if trap jumped to by vmfault, restore skip m0 signalling*/
+        s_getreg_b32 ttmp14, hwreg(HW_REG_TRAPSTS)
+        s_and_b32 ttmp2, ttmp14, 0x800
+        s_cbranch_scc1 RESTORE_AND_EXIT
+        GET_DOORBELL:
+        .if .amdgcn.gfx_generation_number < 11
+            s_mov_b32 ttmp2, exec_lo
+            s_mov_b32 ttmp3, exec_hi
+            s_mov_b32 exec_lo, 0x80000000
+            s_sendmsg 10
+            WAIT_SENDMSG:
+            /*wait until msb is cleared (i.e. doorbell fetched)*/
+            s_nop 7
+            s_bitcmp0_b32 exec_lo, 0x1F
+            s_cbranch_scc0 WAIT_SENDMSG
+            /* restore exec */
+            s_mov_b32 exec_hi, ttmp3
+            s_and_b32 exec_lo, exec_lo, 0xfff
+            s_mov_b32 ttmp3, exec_lo
+            s_mov_b32 exec_lo, ttmp2
+        .else
+            s_sendmsg_rtn_b32 ttmp3, sendmsg(MSG_RTN_GET_DOORBELL)
+            s_waitcnt lgkmcnt(0)
+            s_and_b32 ttmp3, ttmp3, 0x3ff
+        .endif
+        s_mov_b32 ttmp2, m0
+        s_or_b32 ttmp3, ttmp3, 0x800
+        /* set m0, send interrupt and restore m0 and exit trap*/
+        s_mov_b32 m0, ttmp3
+        s_nop 0x0
+        s_sendmsg sendmsg(MSG_INTERRUPT)
+        s_waitcnt lgkmcnt(0)
+        s_mov_b32 m0, ttmp2
+        v_mov_b32 v4, ttmp1
+        /* restore and increment program counter to skip shader trap jump*/
+        s_add_u32 ttmp0, ttmp0, 4
+        s_addc_u32 ttmp1, ttmp1, 0
+        s_and_b32 ttmp1, ttmp1, 0xffff
+        RESTORE_AND_EXIT:
+        /* restore SQ_WAVE_IB_STS */
+        s_lshr_b32 ttmp2, ttmp11, (26 - 15)
+        s_and_b32 ttmp2, ttmp2, (0x8000 | 0x1F0000)
+        s_setreg_b32 hwreg(HW_REG_IB_STS), ttmp2
+        /* restore SQ_WAVE_STATUS */
+        s_and_b64 exec, exec, exec
+        s_and_b64 vcc, vcc, vcc
+        s_setreg_b32 hwreg(HW_REG_STATUS), ttmp12
+        s_rfe_b64 [ttmp0, ttmp1]
+)";
@@ -62,4 +62,8 @@ extern const char *ReadMemoryIsa;
 extern const char *GwsInitIsa;
 extern const char *GwsAtomicIncreaseIsa;

+/* HitTrapEvent */
+extern const char *jump_to_trap_gfx;
+extern const char *trap_handler_gfx;
+
 #endif  // _SHADERSTORE_H_