drm/amdkfd: Add wave control operation to debugger

The wave control operation supports several command types executed upon
existing wave fronts that belong to the currently debugged process.

The available commands are:

HALT   - Freeze wave front(s) execution
RESUME - Resume freezed wave front(s) execution
KILL   - Kill existing wave front(s)
Signed-off-by: Yair Shachar <yair.shachar@amd.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>

drivers/gpu/drm/amd/amdkfd/kfd_dbgdev.c

@@ -47,6 +47,125 @@ static void dbgdev_address_watch_disable_nodiq(struct kfd_dev *dev)
 	dev->kfd2kgd->address_watch_disable(dev->kgd);
 }
 
+static int dbgdev_diq_submit_ib(struct kfd_dbgdev *dbgdev,
+				unsigned int pasid, uint64_t vmid0_address,
+				uint32_t *packet_buff, size_t size_in_bytes)
+{
+	struct pm4__release_mem *rm_packet;
+	struct pm4__indirect_buffer_pasid *ib_packet;
+	struct kfd_mem_obj *mem_obj;
+	size_t pq_packets_size_in_bytes;
+	union ULARGE_INTEGER *largep;
+	union ULARGE_INTEGER addr;
+	struct kernel_queue *kq;
+	uint64_t *rm_state;
+	unsigned int *ib_packet_buff;
+	int status;
+
+	BUG_ON(!dbgdev || !dbgdev->kq || !packet_buff || !size_in_bytes);
+
+	kq = dbgdev->kq;
+
+	pq_packets_size_in_bytes = sizeof(struct pm4__release_mem) +
+				sizeof(struct pm4__indirect_buffer_pasid);
+
+	/*
+	 * We acquire a buffer from DIQ
+	 * The receive packet buff will be sitting on the Indirect Buffer
+	 * and in the PQ we put the IB packet + sync packet(s).
+	 */
+	status = kq->ops.acquire_packet_buffer(kq,
+				pq_packets_size_in_bytes / sizeof(uint32_t),
+				&ib_packet_buff);
+	if (status != 0) {
+		pr_err("amdkfd: acquire_packet_buffer failed\n");
+		return status;
+	}
+
+	memset(ib_packet_buff, 0, pq_packets_size_in_bytes);
+
+	ib_packet = (struct pm4__indirect_buffer_pasid *) (ib_packet_buff);
+
+	ib_packet->header.count = 3;
+	ib_packet->header.opcode = IT_INDIRECT_BUFFER_PASID;
+	ib_packet->header.type = PM4_TYPE_3;
+
+	largep = (union ULARGE_INTEGER *) &vmid0_address;
+
+	ib_packet->bitfields2.ib_base_lo = largep->u.low_part >> 2;
+	ib_packet->bitfields3.ib_base_hi = largep->u.high_part;
+
+	ib_packet->control = (1 << 23) | (1 << 31) |
+			((size_in_bytes / sizeof(uint32_t)) & 0xfffff);
+
+	ib_packet->bitfields5.pasid = pasid;
+
+	/*
+	 * for now we use release mem for GPU-CPU synchronization
+	 * Consider WaitRegMem + WriteData as a better alternative
+	 * we get a GART allocations ( gpu/cpu mapping),
+	 * for the sync variable, and wait until:
+	 * (a) Sync with HW
+	 * (b) Sync var is written by CP to mem.
+	 */
+	rm_packet = (struct pm4__release_mem *) (ib_packet_buff +
+			(sizeof(struct pm4__indirect_buffer_pasid) /
+					sizeof(unsigned int)));
+
+	status = kfd_gtt_sa_allocate(dbgdev->dev, sizeof(uint64_t),
+					&mem_obj);
+
+	if (status != 0) {
+		pr_err("amdkfd: Failed to allocate GART memory\n");
+		kq->ops.rollback_packet(kq);
+		return status;
+	}
+
+	rm_state = (uint64_t *) mem_obj->cpu_ptr;
+
+	*rm_state = QUEUESTATE__ACTIVE_COMPLETION_PENDING;
+
+	rm_packet->header.opcode = IT_RELEASE_MEM;
+	rm_packet->header.type = PM4_TYPE_3;
+	rm_packet->header.count = sizeof(struct pm4__release_mem) /
+					sizeof(unsigned int) - 2;
+
+	rm_packet->bitfields2.event_type = CACHE_FLUSH_AND_INV_TS_EVENT;
+	rm_packet->bitfields2.event_index =
+				event_index___release_mem__end_of_pipe;
+
+	rm_packet->bitfields2.cache_policy = cache_policy___release_mem__lru;
+	rm_packet->bitfields2.atc = 0;
+	rm_packet->bitfields2.tc_wb_action_ena = 1;
+
+	addr.quad_part = mem_obj->gpu_addr;
+
+	rm_packet->bitfields4.address_lo_32b = addr.u.low_part >> 2;
+	rm_packet->address_hi = addr.u.high_part;
+
+	rm_packet->bitfields3.data_sel =
+				data_sel___release_mem__send_64_bit_data;
+
+	rm_packet->bitfields3.int_sel =
+			int_sel___release_mem__send_data_after_write_confirm;
+
+	rm_packet->bitfields3.dst_sel =
+			dst_sel___release_mem__memory_controller;
+
+	rm_packet->data_lo = QUEUESTATE__ACTIVE;
+
+	kq->ops.submit_packet(kq);
+
+	/* Wait till CP writes sync code: */
+	status = amdkfd_fence_wait_timeout(
+			(unsigned int *) rm_state,
+			QUEUESTATE__ACTIVE, 1500);
+
+	kfd_gtt_sa_free(dbgdev->dev, mem_obj);
+
+	return status;
+}
+
 static int dbgdev_register_nodiq(struct kfd_dbgdev *dbgdev)
 {
 	BUG_ON(!dbgdev);
@@ -117,6 +236,290 @@ static int dbgdev_unregister_diq(struct kfd_dbgdev *dbgdev)
 	return status;
 }
 
+static int dbgdev_wave_control_set_registers(
+				struct dbg_wave_control_info *wac_info,
+				union SQ_CMD_BITS *in_reg_sq_cmd,
+				union GRBM_GFX_INDEX_BITS *in_reg_gfx_index)
+{
+	int status;
+	union SQ_CMD_BITS reg_sq_cmd;
+	union GRBM_GFX_INDEX_BITS reg_gfx_index;
+	struct HsaDbgWaveMsgAMDGen2 *pMsg;
+
+	BUG_ON(!wac_info || !in_reg_sq_cmd || !in_reg_gfx_index);
+
+	reg_sq_cmd.u32All = 0;
+	reg_gfx_index.u32All = 0;
+	pMsg = &wac_info->dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2;
+
+	switch (wac_info->mode) {
+	/* Send command to single wave */
+	case HSA_DBG_WAVEMODE_SINGLE:
+		/*
+		 * Limit access to the process waves only,
+		 * by setting vmid check
+		 */
+		reg_sq_cmd.bits.check_vmid = 1;
+		reg_sq_cmd.bits.simd_id = pMsg->ui32.SIMD;
+		reg_sq_cmd.bits.wave_id = pMsg->ui32.WaveId;
+		reg_sq_cmd.bits.mode = SQ_IND_CMD_MODE_SINGLE;
+
+		reg_gfx_index.bits.sh_index = pMsg->ui32.ShaderArray;
+		reg_gfx_index.bits.se_index = pMsg->ui32.ShaderEngine;
+		reg_gfx_index.bits.instance_index = pMsg->ui32.HSACU;
+
+		break;
+
+	/* Send command to all waves with matching VMID */
+	case HSA_DBG_WAVEMODE_BROADCAST_PROCESS:
+
+		reg_gfx_index.bits.sh_broadcast_writes = 1;
+		reg_gfx_index.bits.se_broadcast_writes = 1;
+		reg_gfx_index.bits.instance_broadcast_writes = 1;
+
+		reg_sq_cmd.bits.mode = SQ_IND_CMD_MODE_BROADCAST;
+
+		break;
+
+	/* Send command to all CU waves with matching VMID */
+	case HSA_DBG_WAVEMODE_BROADCAST_PROCESS_CU:
+
+		reg_sq_cmd.bits.check_vmid = 1;
+		reg_sq_cmd.bits.mode = SQ_IND_CMD_MODE_BROADCAST;
+
+		reg_gfx_index.bits.sh_index = pMsg->ui32.ShaderArray;
+		reg_gfx_index.bits.se_index = pMsg->ui32.ShaderEngine;
+		reg_gfx_index.bits.instance_index = pMsg->ui32.HSACU;
+
+		break;
+
+	default:
+		return -EINVAL;
+	}
+
+	switch (wac_info->operand) {
+	case HSA_DBG_WAVEOP_HALT:
+		reg_sq_cmd.bits.cmd = SQ_IND_CMD_CMD_HALT;
+		break;
+
+	case HSA_DBG_WAVEOP_RESUME:
+		reg_sq_cmd.bits.cmd = SQ_IND_CMD_CMD_RESUME;
+		break;
+
+	case HSA_DBG_WAVEOP_KILL:
+		reg_sq_cmd.bits.cmd = SQ_IND_CMD_CMD_KILL;
+		break;
+
+	case HSA_DBG_WAVEOP_DEBUG:
+		reg_sq_cmd.bits.cmd = SQ_IND_CMD_CMD_DEBUG;
+		break;
+
+	case HSA_DBG_WAVEOP_TRAP:
+		if (wac_info->trapId < MAX_TRAPID) {
+			reg_sq_cmd.bits.cmd = SQ_IND_CMD_CMD_TRAP;
+			reg_sq_cmd.bits.trap_id = wac_info->trapId;
+		} else {
+			status = -EINVAL;
+		}
+		break;
+
+	default:
+		status = -EINVAL;
+		break;
+	}
+
+	if (status == 0) {
+		*in_reg_sq_cmd = reg_sq_cmd;
+		*in_reg_gfx_index = reg_gfx_index;
+	}
+
+	return status;
+}
+
+static int dbgdev_wave_control_diq(struct kfd_dbgdev *dbgdev,
+					struct dbg_wave_control_info *wac_info)
+{
+
+	int status;
+	union SQ_CMD_BITS reg_sq_cmd;
+	union GRBM_GFX_INDEX_BITS reg_gfx_index;
+	struct kfd_mem_obj *mem_obj;
+	uint32_t *packet_buff_uint;
+	struct pm4__set_config_reg *packets_vec;
+	size_t ib_size = sizeof(struct pm4__set_config_reg) * 3;
+
+	BUG_ON(!dbgdev || !wac_info);
+
+	reg_sq_cmd.u32All = 0;
+
+	status = dbgdev_wave_control_set_registers(wac_info, &reg_sq_cmd,
+							&reg_gfx_index);
+	if (status) {
+		pr_err("amdkfd: Failed to set wave control registers\n");
+		return status;
+	}
+
+	/* we do not control the VMID in DIQ,so reset it to a known value */
+	reg_sq_cmd.bits.vm_id = 0;
+
+	pr_debug("\t\t %30s\n", "* * * * * * * * * * * * * * * * * *");
+
+	pr_debug("\t\t mode      is: %u\n", wac_info->mode);
+	pr_debug("\t\t operand   is: %u\n", wac_info->operand);
+	pr_debug("\t\t trap id   is: %u\n", wac_info->trapId);
+	pr_debug("\t\t msg value is: %u\n",
+			wac_info->dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value);
+	pr_debug("\t\t vmid      is: N/A\n");
+
+	pr_debug("\t\t chk_vmid  is : %u\n", reg_sq_cmd.bitfields.check_vmid);
+	pr_debug("\t\t command   is : %u\n", reg_sq_cmd.bitfields.cmd);
+	pr_debug("\t\t queue id  is : %u\n", reg_sq_cmd.bitfields.queue_id);
+	pr_debug("\t\t simd id   is : %u\n", reg_sq_cmd.bitfields.simd_id);
+	pr_debug("\t\t mode      is : %u\n", reg_sq_cmd.bitfields.mode);
+	pr_debug("\t\t vm_id     is : %u\n", reg_sq_cmd.bitfields.vm_id);
+	pr_debug("\t\t wave_id   is : %u\n", reg_sq_cmd.bitfields.wave_id);
+
+	pr_debug("\t\t ibw       is : %u\n",
+			reg_gfx_index.bitfields.instance_broadcast_writes);
+	pr_debug("\t\t ii        is : %u\n",
+			reg_gfx_index.bitfields.instance_index);
+	pr_debug("\t\t sebw      is : %u\n",
+			reg_gfx_index.bitfields.se_broadcast_writes);
+	pr_debug("\t\t se_ind    is : %u\n", reg_gfx_index.bitfields.se_index);
+	pr_debug("\t\t sh_ind    is : %u\n", reg_gfx_index.bitfields.sh_index);
+	pr_debug("\t\t sbw       is : %u\n",
+			reg_gfx_index.bitfields.sh_broadcast_writes);
+
+	pr_debug("\t\t %30s\n", "* * * * * * * * * * * * * * * * * *");
+
+	status = kfd_gtt_sa_allocate(dbgdev->dev, ib_size, &mem_obj);
+
+	if (status != 0) {
+		pr_err("amdkfd: Failed to allocate GART memory\n");
+		return status;
+	}
+
+	packet_buff_uint = mem_obj->cpu_ptr;
+
+	memset(packet_buff_uint, 0, ib_size);
+
+	packets_vec =  (struct pm4__set_config_reg *) packet_buff_uint;
+	packets_vec[0].header.count = 1;
+	packets_vec[0].header.opcode = IT_SET_UCONFIG_REG;
+	packets_vec[0].header.type = PM4_TYPE_3;
+	packets_vec[0].bitfields2.reg_offset =
+			GRBM_GFX_INDEX / (sizeof(uint32_t)) -
+				USERCONFIG_REG_BASE;
+
+	packets_vec[0].bitfields2.insert_vmid = 0;
+	packets_vec[0].reg_data[0] = reg_gfx_index.u32All;
+
+	packets_vec[1].header.count = 1;
+	packets_vec[1].header.opcode = IT_SET_CONFIG_REG;
+	packets_vec[1].header.type = PM4_TYPE_3;
+	packets_vec[1].bitfields2.reg_offset = SQ_CMD / (sizeof(uint32_t)) -
+						CONFIG_REG_BASE;
+
+	packets_vec[1].bitfields2.vmid_shift = SQ_CMD_VMID_OFFSET;
+	packets_vec[1].bitfields2.insert_vmid = 1;
+	packets_vec[1].reg_data[0] = reg_sq_cmd.u32All;
+
+	/* Restore the GRBM_GFX_INDEX register */
+
+	reg_gfx_index.u32All = 0;
+	reg_gfx_index.bits.sh_broadcast_writes = 1;
+	reg_gfx_index.bits.instance_broadcast_writes = 1;
+	reg_gfx_index.bits.se_broadcast_writes = 1;
+
+
+	packets_vec[2].ordinal1 = packets_vec[0].ordinal1;
+	packets_vec[2].bitfields2.reg_offset =
+				GRBM_GFX_INDEX / (sizeof(uint32_t)) -
+					USERCONFIG_REG_BASE;
+
+	packets_vec[2].bitfields2.insert_vmid = 0;
+	packets_vec[2].reg_data[0] = reg_gfx_index.u32All;
+
+	status = dbgdev_diq_submit_ib(
+			dbgdev,
+			wac_info->process->pasid,
+			mem_obj->gpu_addr,
+			packet_buff_uint,
+			ib_size);
+
+	if (status != 0)
+		pr_err("amdkfd: Failed to submit IB to DIQ\n");
+
+	kfd_gtt_sa_free(dbgdev->dev, mem_obj);
+
+	return status;
+}
+
+static int dbgdev_wave_control_nodiq(struct kfd_dbgdev *dbgdev,
+					struct dbg_wave_control_info *wac_info)
+{
+	int status;
+	union SQ_CMD_BITS reg_sq_cmd;
+	union GRBM_GFX_INDEX_BITS reg_gfx_index;
+	struct kfd_process_device *pdd;
+
+	BUG_ON(!dbgdev || !dbgdev->dev || !wac_info);
+
+	reg_sq_cmd.u32All = 0;
+
+	/* taking the VMID for that process on the safe way using PDD */
+	pdd = kfd_get_process_device_data(dbgdev->dev, wac_info->process);
+
+	if (!pdd) {
+		pr_err("amdkfd: Failed to get pdd for wave control no DIQ\n");
+		return -EFAULT;
+	}
+	status = dbgdev_wave_control_set_registers(wac_info, &reg_sq_cmd,
+							&reg_gfx_index);
+	if (status) {
+		pr_err("amdkfd: Failed to set wave control registers\n");
+		return status;
+	}
+
+	/* for non DIQ we need to patch the VMID: */
+
+	reg_sq_cmd.bits.vm_id = pdd->qpd.vmid;
+
+	pr_debug("\t\t %30s\n", "* * * * * * * * * * * * * * * * * *");
+
+	pr_debug("\t\t mode      is: %u\n", wac_info->mode);
+	pr_debug("\t\t operand   is: %u\n", wac_info->operand);
+	pr_debug("\t\t trap id   is: %u\n", wac_info->trapId);
+	pr_debug("\t\t msg value is: %u\n",
+			wac_info->dbgWave_msg.DbgWaveMsg.WaveMsgInfoGen2.Value);
+	pr_debug("\t\t vmid      is: %u\n", pdd->qpd.vmid);
+
+	pr_debug("\t\t chk_vmid  is : %u\n", reg_sq_cmd.bitfields.check_vmid);
+	pr_debug("\t\t command   is : %u\n", reg_sq_cmd.bitfields.cmd);
+	pr_debug("\t\t queue id  is : %u\n", reg_sq_cmd.bitfields.queue_id);
+	pr_debug("\t\t simd id   is : %u\n", reg_sq_cmd.bitfields.simd_id);
+	pr_debug("\t\t mode      is : %u\n", reg_sq_cmd.bitfields.mode);
+	pr_debug("\t\t vm_id     is : %u\n", reg_sq_cmd.bitfields.vm_id);
+	pr_debug("\t\t wave_id   is : %u\n", reg_sq_cmd.bitfields.wave_id);
+
+	pr_debug("\t\t ibw       is : %u\n",
+			reg_gfx_index.bitfields.instance_broadcast_writes);
+	pr_debug("\t\t ii        is : %u\n",
+			reg_gfx_index.bitfields.instance_index);
+	pr_debug("\t\t sebw      is : %u\n",
+			reg_gfx_index.bitfields.se_broadcast_writes);
+	pr_debug("\t\t se_ind    is : %u\n", reg_gfx_index.bitfields.se_index);
+	pr_debug("\t\t sh_ind    is : %u\n", reg_gfx_index.bitfields.sh_index);
+	pr_debug("\t\t sbw       is : %u\n",
+			reg_gfx_index.bitfields.sh_broadcast_writes);
+
+	pr_debug("\t\t %30s\n", "* * * * * * * * * * * * * * * * * *");
+
+	return dbgdev->dev->kfd2kgd->wave_control_execute(dbgdev->dev->kgd,
+							reg_gfx_index.u32All,
+							reg_sq_cmd.u32All);
+}
+
 void kfd_dbgdev_init(struct kfd_dbgdev *pdbgdev, struct kfd_dev *pdev,
 			enum DBGDEV_TYPE type)
 {
@@ -131,11 +534,13 @@ void kfd_dbgdev_init(struct kfd_dbgdev *pdbgdev, struct kfd_dev *pdev,
 	case DBGDEV_TYPE_NODIQ:
 		pdbgdev->dbgdev_register = dbgdev_register_nodiq;
 		pdbgdev->dbgdev_unregister = dbgdev_unregister_nodiq;
+		pdbgdev->dbgdev_wave_control = dbgdev_wave_control_nodiq;
 		break;
 	case DBGDEV_TYPE_DIQ:
 	default:
 		pdbgdev->dbgdev_register = dbgdev_register_diq;
 		pdbgdev->dbgdev_unregister = dbgdev_unregister_diq;
+		pdbgdev->dbgdev_wave_control =  dbgdev_wave_control_diq;
 		break;
 	}
 

drivers/gpu/drm/amd/amdkfd/kfd_dbgmgr.c

@@ -133,3 +133,19 @@ long kfd_dbgmgr_unregister(struct kfd_dbgmgr *pmgr, struct kfd_process *p)
 
 	return 0;
 }
+
+long kfd_dbgmgr_wave_control(struct kfd_dbgmgr *pmgr,
+				struct dbg_wave_control_info *wac_info)
+{
+	BUG_ON(!pmgr || !pmgr->dbgdev || !wac_info);
+
+	/* Is the requests coming from the already registered process? */
+	if (pmgr->pasid != wac_info->process->pasid) {
+		pr_debug("H/W debugger support was not registered for requester pasid %d\n",
+				wac_info->process->pasid);
+		return -EINVAL;
+	}
+
+	return (long) pmgr->dbgdev->dbgdev_wave_control(pmgr->dbgdev, wac_info);
+}
+

drivers/gpu/drm/amd/amdkfd/kfd_dbgmgr.h

@@ -268,6 +268,8 @@ struct kfd_dbgdev {
 	/* virtualized function pointers to device dbg */
 	int (*dbgdev_register)(struct kfd_dbgdev *dbgdev);
 	int (*dbgdev_unregister)(struct kfd_dbgdev *dbgdev);
+	int (*dbgdev_wave_control)(struct kfd_dbgdev *dbgdev,
+				struct dbg_wave_control_info *wac_info);
 
 };
 
@@ -283,5 +285,6 @@ void kfd_dbgmgr_destroy(struct kfd_dbgmgr *pmgr);
 bool kfd_dbgmgr_create(struct kfd_dbgmgr **ppmgr, struct kfd_dev *pdev);
 long kfd_dbgmgr_register(struct kfd_dbgmgr *pmgr, struct kfd_process *p);
 long kfd_dbgmgr_unregister(struct kfd_dbgmgr *pmgr, struct kfd_process *p);
-
+long kfd_dbgmgr_wave_control(struct kfd_dbgmgr *pmgr,
+				struct dbg_wave_control_info *wac_info);
 #endif /* KFD_DBGMGR_H_ */

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c

@@ -915,7 +915,7 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
 	return retval;
 }
 
-static int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
+int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
 				unsigned int fence_value,
 				unsigned long timeout)
 {

drivers/gpu/drm/amd/amdkfd/kfd_priv.h

@@ -656,6 +656,10 @@ int pqm_update_queue(struct process_queue_manager *pqm, unsigned int qid,
 struct kernel_queue *pqm_get_kernel_queue(struct process_queue_manager *pqm,
 						unsigned int qid);
 
+int amdkfd_fence_wait_timeout(unsigned int *fence_addr,
+				unsigned int fence_value,
+				unsigned long timeout);
+
 /* Packet Manager */
 
 #define KFD_HIQ_TIMEOUT (500)