drm/amdkfd: add debug set and clear address watch points operation

Shader read, write and atomic memory operations can be alerted to the
debugger as an address watch exception.

Allow the debugger to pass in a watch point to a particular memory
address per device.

Note that there exists only 4 watch points per devices to date, so have
the KFD keep track of what watch points are allocated or not.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c

@@ -118,6 +118,55 @@ static uint32_t kgd_aldebaran_set_wave_launch_mode(struct amdgpu_device *adev,
 	return data;
 }
 
+#define TCP_WATCH_STRIDE (regTCP_WATCH1_ADDR_H - regTCP_WATCH0_ADDR_H)
+static uint32_t kgd_gfx_aldebaran_set_address_watch(
+					struct amdgpu_device *adev,
+					uint64_t watch_address,
+					uint32_t watch_address_mask,
+					uint32_t watch_id,
+					uint32_t watch_mode,
+					uint32_t debug_vmid)
+{
+	uint32_t watch_address_high;
+	uint32_t watch_address_low;
+	uint32_t watch_address_cntl;
+
+	watch_address_cntl = 0;
+	watch_address_low = lower_32_bits(watch_address);
+	watch_address_high = upper_32_bits(watch_address) & 0xffff;
+
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			MODE,
+			watch_mode);
+
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			MASK,
+			watch_address_mask >> 6);
+
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			VALID,
+			1);
+
+	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, regTCP_WATCH0_ADDR_H) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_high);
+
+	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, regTCP_WATCH0_ADDR_L) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_low);
+
+	return watch_address_cntl;
+}
+
+uint32_t kgd_gfx_aldebaran_clear_address_watch(struct amdgpu_device *adev,
+					uint32_t watch_id)
+{
+	return 0;
+}
+
 const struct kfd2kgd_calls aldebaran_kfd2kgd = {
 	.program_sh_mem_settings = kgd_gfx_v9_program_sh_mem_settings,
 	.set_pasid_vmid_mapping = kgd_gfx_v9_set_pasid_vmid_mapping,
@@ -141,6 +190,8 @@ const struct kfd2kgd_calls aldebaran_kfd2kgd = {
 	.validate_trap_override_request = kgd_aldebaran_validate_trap_override_request,
 	.set_wave_launch_trap_override = kgd_aldebaran_set_wave_launch_trap_override,
 	.set_wave_launch_mode = kgd_aldebaran_set_wave_launch_mode,
+	.set_address_watch = kgd_gfx_aldebaran_set_address_watch,
+	.clear_address_watch = kgd_gfx_aldebaran_clear_address_watch,
 	.get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times,
 	.build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info,
 	.program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c

@@ -413,6 +413,8 @@ const struct kfd2kgd_calls arcturus_kfd2kgd = {
 	.validate_trap_override_request = kgd_gfx_v9_validate_trap_override_request,
 	.set_wave_launch_trap_override = kgd_gfx_v9_set_wave_launch_trap_override,
 	.set_wave_launch_mode = kgd_gfx_v9_set_wave_launch_mode,
+	.set_address_watch = kgd_gfx_v9_set_address_watch,
+	.clear_address_watch = kgd_gfx_v9_clear_address_watch,
 	.get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times,
 	.build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info,
 	.get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c

@@ -880,6 +880,82 @@ uint32_t kgd_gfx_v10_set_wave_launch_mode(struct amdgpu_device *adev,
 	return 0;
 }
 
+#define TCP_WATCH_STRIDE (mmTCP_WATCH1_ADDR_H - mmTCP_WATCH0_ADDR_H)
+uint32_t kgd_gfx_v10_set_address_watch(struct amdgpu_device *adev,
+					uint64_t watch_address,
+					uint32_t watch_address_mask,
+					uint32_t watch_id,
+					uint32_t watch_mode,
+					uint32_t debug_vmid)
+{
+	uint32_t watch_address_high;
+	uint32_t watch_address_low;
+	uint32_t watch_address_cntl;
+
+	watch_address_cntl = 0;
+
+	watch_address_low = lower_32_bits(watch_address);
+	watch_address_high = upper_32_bits(watch_address) & 0xffff;
+
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			VMID,
+			debug_vmid);
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			MODE,
+			watch_mode);
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			MASK,
+			watch_address_mask >> 7);
+
+	/* Turning off this watch point until we set all the registers */
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			VALID,
+			0);
+
+	WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_cntl);
+
+	WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_high);
+
+	WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_L) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_low);
+
+	/* Enable the watch point */
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			VALID,
+			1);
+
+	WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_cntl);
+
+	return 0;
+}
+
+uint32_t kgd_gfx_v10_clear_address_watch(struct amdgpu_device *adev,
+					uint32_t watch_id)
+{
+	uint32_t watch_address_cntl;
+
+	watch_address_cntl = 0;
+
+	WREG32((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_cntl);
+
+	return 0;
+}
+
+
 /* kgd_gfx_v10_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values
  * The values read are:
  *     ib_offload_wait_time     -- Wait Count for Indirect Buffer Offloads.
@@ -969,6 +1045,8 @@ const struct kfd2kgd_calls gfx_v10_kfd2kgd = {
 	.validate_trap_override_request = kgd_gfx_v10_validate_trap_override_request,
 	.set_wave_launch_trap_override = kgd_gfx_v10_set_wave_launch_trap_override,
 	.set_wave_launch_mode = kgd_gfx_v10_set_wave_launch_mode,
+	.set_address_watch = kgd_gfx_v10_set_address_watch,
+	.clear_address_watch = kgd_gfx_v10_clear_address_watch,
 	.get_iq_wait_times = kgd_gfx_v10_get_iq_wait_times,
 	.build_grace_period_packet_info = kgd_gfx_v10_build_grace_period_packet_info,
 	.program_trap_handler_settings = program_trap_handler_settings,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h

@@ -39,6 +39,14 @@ uint32_t kgd_gfx_v10_set_wave_launch_trap_override(struct amdgpu_device *adev,
 uint32_t kgd_gfx_v10_set_wave_launch_mode(struct amdgpu_device *adev,
 					 uint8_t wave_launch_mode,
 					 uint32_t vmid);
+uint32_t kgd_gfx_v10_set_address_watch(struct amdgpu_device *adev,
+					uint64_t watch_address,
+					uint32_t watch_address_mask,
+					uint32_t watch_id,
+					uint32_t watch_mode,
+					uint32_t debug_vmid);
+uint32_t kgd_gfx_v10_clear_address_watch(struct amdgpu_device *adev,
+					uint32_t watch_id);
 void kgd_gfx_v10_get_iq_wait_times(struct amdgpu_device *adev, uint32_t *wait_times);
 void kgd_gfx_v10_build_grace_period_packet_info(struct amdgpu_device *adev,
 					       uint32_t wait_times,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c

@@ -678,6 +678,7 @@ const struct kfd2kgd_calls gfx_v10_3_kfd2kgd = {
 	.disable_debug_trap = kgd_gfx_v10_disable_debug_trap,
 	.validate_trap_override_request = kgd_gfx_v10_validate_trap_override_request,
 	.set_wave_launch_trap_override = kgd_gfx_v10_set_wave_launch_trap_override,
-	.set_wave_launch_mode = kgd_gfx_v10_set_wave_launch_mode
-
+	.set_wave_launch_mode = kgd_gfx_v10_set_wave_launch_mode,
+	.set_address_watch = kgd_gfx_v10_set_address_watch,
+	.clear_address_watch = kgd_gfx_v10_clear_address_watch
 };

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c

@@ -737,6 +737,54 @@ static uint32_t kgd_gfx_v11_set_wave_launch_mode(struct amdgpu_device *adev,
 	return data;
 }
 
+#define TCP_WATCH_STRIDE (regTCP_WATCH1_ADDR_H - regTCP_WATCH0_ADDR_H)
+static uint32_t kgd_gfx_v11_set_address_watch(struct amdgpu_device *adev,
+					uint64_t watch_address,
+					uint32_t watch_address_mask,
+					uint32_t watch_id,
+					uint32_t watch_mode,
+					uint32_t debug_vmid)
+{
+	uint32_t watch_address_high;
+	uint32_t watch_address_low;
+	uint32_t watch_address_cntl;
+
+	watch_address_cntl = 0;
+	watch_address_low = lower_32_bits(watch_address);
+	watch_address_high = upper_32_bits(watch_address) & 0xffff;
+
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			MODE,
+			watch_mode);
+
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			MASK,
+			watch_address_mask >> 7);
+
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			VALID,
+			1);
+
+	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, regTCP_WATCH0_ADDR_H) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_high);
+
+	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, regTCP_WATCH0_ADDR_L) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_low);
+
+	return watch_address_cntl;
+}
+
+uint32_t kgd_gfx_v11_clear_address_watch(struct amdgpu_device *adev,
+					uint32_t watch_id)
+{
+	return 0;
+}
+
 const struct kfd2kgd_calls gfx_v11_kfd2kgd = {
 	.program_sh_mem_settings = program_sh_mem_settings_v11,
 	.set_pasid_vmid_mapping = set_pasid_vmid_mapping_v11,
@@ -757,5 +805,7 @@ const struct kfd2kgd_calls gfx_v11_kfd2kgd = {
 	.disable_debug_trap = kgd_gfx_v11_disable_debug_trap,
 	.validate_trap_override_request = kgd_gfx_v11_validate_trap_override_request,
 	.set_wave_launch_trap_override = kgd_gfx_v11_set_wave_launch_trap_override,
-	.set_wave_launch_mode = kgd_gfx_v11_set_wave_launch_mode
+	.set_wave_launch_mode = kgd_gfx_v11_set_wave_launch_mode,
+	.set_address_watch = kgd_gfx_v11_set_address_watch,
+	.clear_address_watch = kgd_gfx_v11_clear_address_watch
 };

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c

@@ -816,6 +816,81 @@ uint32_t kgd_gfx_v9_set_wave_launch_mode(struct amdgpu_device *adev,
 	return 0;
 }
 
+#define TCP_WATCH_STRIDE (mmTCP_WATCH1_ADDR_H - mmTCP_WATCH0_ADDR_H)
+uint32_t kgd_gfx_v9_set_address_watch(struct amdgpu_device *adev,
+					uint64_t watch_address,
+					uint32_t watch_address_mask,
+					uint32_t watch_id,
+					uint32_t watch_mode,
+					uint32_t debug_vmid)
+{
+	uint32_t watch_address_high;
+	uint32_t watch_address_low;
+	uint32_t watch_address_cntl;
+
+	watch_address_cntl = 0;
+
+	watch_address_low = lower_32_bits(watch_address);
+	watch_address_high = upper_32_bits(watch_address) & 0xffff;
+
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			VMID,
+			debug_vmid);
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			MODE,
+			watch_mode);
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			MASK,
+			watch_address_mask >> 6);
+
+	/* Turning off this watch point until we set all the registers */
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			VALID,
+			0);
+
+	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_cntl);
+
+	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_H) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_high);
+
+	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_ADDR_L) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_low);
+
+	/* Enable the watch point */
+	watch_address_cntl = REG_SET_FIELD(watch_address_cntl,
+			TCP_WATCH0_CNTL,
+			VALID,
+			1);
+
+	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_cntl);
+
+	return 0;
+}
+
+uint32_t kgd_gfx_v9_clear_address_watch(struct amdgpu_device *adev,
+					uint32_t watch_id)
+{
+	uint32_t watch_address_cntl;
+
+	watch_address_cntl = 0;
+
+	WREG32_RLC((SOC15_REG_OFFSET(GC, 0, mmTCP_WATCH0_CNTL) +
+			(watch_id * TCP_WATCH_STRIDE)),
+			watch_address_cntl);
+
+	return 0;
+}
+
 /* kgd_gfx_v9_get_iq_wait_times: Returns the mmCP_IQ_WAIT_TIME1/2 values
  * The values read are:
  *     ib_offload_wait_time     -- Wait Count for Indirect Buffer Offloads.
@@ -1090,6 +1165,8 @@ const struct kfd2kgd_calls gfx_v9_kfd2kgd = {
 	.validate_trap_override_request = kgd_gfx_v9_validate_trap_override_request,
 	.set_wave_launch_trap_override = kgd_gfx_v9_set_wave_launch_trap_override,
 	.set_wave_launch_mode = kgd_gfx_v9_set_wave_launch_mode,
+	.set_address_watch = kgd_gfx_v9_set_address_watch,
+	.clear_address_watch = kgd_gfx_v9_clear_address_watch,
 	.get_iq_wait_times = kgd_gfx_v9_get_iq_wait_times,
 	.build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info,
 	.get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy,

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h

@@ -84,6 +84,14 @@ uint32_t kgd_gfx_v9_set_wave_launch_trap_override(struct amdgpu_device *adev,
 					     uint32_t trap_mask_request,
 					     uint32_t *trap_mask_prev,
 					     uint32_t kfd_dbg_trap_cntl_prev);
+uint32_t kgd_gfx_v9_set_address_watch(struct amdgpu_device *adev,
+					uint64_t watch_address,
+					uint32_t watch_address_mask,
+					uint32_t watch_id,
+					uint32_t watch_mode,
+					uint32_t debug_vmid);
+uint32_t kgd_gfx_v9_clear_address_watch(struct amdgpu_device *adev,
+					uint32_t watch_id);
 void kgd_gfx_v9_get_iq_wait_times(struct amdgpu_device *adev, uint32_t *wait_times);
 void kgd_gfx_v9_build_grace_period_packet_info(struct amdgpu_device *adev,
 					       uint32_t wait_times,

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

@@ -2880,6 +2880,7 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v
 	struct mm_struct *mm = NULL;
 	struct pid *pid = NULL;
 	struct kfd_process *target = NULL;
+	struct kfd_process_device *pdd = NULL;
 	int r = 0;
 
 	if (sched_policy == KFD_SCHED_POLICY_NO_HWS) {
@@ -2957,6 +2958,20 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v
 		goto unlock_out;
 	}
 
+	if (args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ||
+	    args->op == KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH) {
+		int user_gpu_id = kfd_process_get_user_gpu_id(target,
+				args->op == KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH ?
+					args->set_node_address_watch.gpu_id :
+					args->clear_node_address_watch.gpu_id);
+
+		pdd = kfd_process_device_data_by_id(target, user_gpu_id);
+		if (user_gpu_id == -EINVAL || !pdd) {
+			r = -ENODEV;
+			goto unlock_out;
+		}
+	}
+
 	switch (args->op) {
 	case KFD_IOC_DBG_TRAP_ENABLE:
 		if (target != p)
@@ -3009,7 +3024,16 @@ static int kfd_ioctl_set_debug_trap(struct file *filep, struct kfd_process *p, v
 				(uint32_t *)args->resume_queues.queue_array_ptr);
 		break;
 	case KFD_IOC_DBG_TRAP_SET_NODE_ADDRESS_WATCH:
+		r = kfd_dbg_trap_set_dev_address_watch(pdd,
+				args->set_node_address_watch.address,
+				args->set_node_address_watch.mask,
+				&args->set_node_address_watch.id,
+				args->set_node_address_watch.mode);
+		break;
 	case KFD_IOC_DBG_TRAP_CLEAR_NODE_ADDRESS_WATCH:
+		r = kfd_dbg_trap_clear_dev_address_watch(pdd,
+				args->clear_node_address_watch.id);
+		break;
 	case KFD_IOC_DBG_TRAP_SET_FLAGS:
 	case KFD_IOC_DBG_TRAP_QUERY_DEBUG_EVENT:
 	case KFD_IOC_DBG_TRAP_QUERY_EXCEPTION_INFO:

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

@@ -24,6 +24,8 @@
 #include "kfd_device_queue_manager.h"
 #include <linux/file.h>
 
+#define MAX_WATCH_ADDRESSES	4
+
 void debug_event_write_work_handler(struct work_struct *work)
 {
 	struct kfd_process *process;
@@ -289,6 +291,139 @@ int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd)
 						pdd->watch_points, flags);
 }
 
+#define KFD_DEBUGGER_INVALID_WATCH_POINT_ID -1
+static int kfd_dbg_get_dev_watch_id(struct kfd_process_device *pdd, int *watch_id)
+{
+	int i;
+
+	*watch_id = KFD_DEBUGGER_INVALID_WATCH_POINT_ID;
+
+	spin_lock(&pdd->dev->kfd->watch_points_lock);
+
+	for (i = 0; i < MAX_WATCH_ADDRESSES; i++) {
+		/* device watchpoint in use so skip */
+		if ((pdd->dev->kfd->alloc_watch_ids >> i) & 0x1)
+			continue;
+
+		pdd->alloc_watch_ids |= 0x1 << i;
+		pdd->dev->kfd->alloc_watch_ids |= 0x1 << i;
+		*watch_id = i;
+		spin_unlock(&pdd->dev->kfd->watch_points_lock);
+		return 0;
+	}
+
+	spin_unlock(&pdd->dev->kfd->watch_points_lock);
+
+	return -ENOMEM;
+}
+
+static void kfd_dbg_clear_dev_watch_id(struct kfd_process_device *pdd, int watch_id)
+{
+	spin_lock(&pdd->dev->kfd->watch_points_lock);
+
+	/* process owns device watch point so safe to clear */
+	if ((pdd->alloc_watch_ids >> watch_id) & 0x1) {
+		pdd->alloc_watch_ids &= ~(0x1 << watch_id);
+		pdd->dev->kfd->alloc_watch_ids &= ~(0x1 << watch_id);
+	}
+
+	spin_unlock(&pdd->dev->kfd->watch_points_lock);
+}
+
+static bool kfd_dbg_owns_dev_watch_id(struct kfd_process_device *pdd, int watch_id)
+{
+	bool owns_watch_id = false;
+
+	spin_lock(&pdd->dev->kfd->watch_points_lock);
+	owns_watch_id = watch_id < MAX_WATCH_ADDRESSES &&
+			((pdd->alloc_watch_ids >> watch_id) & 0x1);
+
+	spin_unlock(&pdd->dev->kfd->watch_points_lock);
+
+	return owns_watch_id;
+}
+
+int kfd_dbg_trap_clear_dev_address_watch(struct kfd_process_device *pdd,
+					uint32_t watch_id)
+{
+	int r;
+
+	if (!kfd_dbg_owns_dev_watch_id(pdd, watch_id))
+		return -EINVAL;
+
+	if (!pdd->dev->kfd->shared_resources.enable_mes) {
+		r = debug_lock_and_unmap(pdd->dev->dqm);
+		if (r)
+			return r;
+	}
+
+	amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
+	pdd->watch_points[watch_id] = pdd->dev->kfd2kgd->clear_address_watch(
+							pdd->dev->adev,
+							watch_id);
+	amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
+
+	if (!pdd->dev->kfd->shared_resources.enable_mes)
+		r = debug_map_and_unlock(pdd->dev->dqm);
+	else
+		r = kfd_dbg_set_mes_debug_mode(pdd);
+
+	kfd_dbg_clear_dev_watch_id(pdd, watch_id);
+
+	return r;
+}
+
+int kfd_dbg_trap_set_dev_address_watch(struct kfd_process_device *pdd,
+					uint64_t watch_address,
+					uint32_t watch_address_mask,
+					uint32_t *watch_id,
+					uint32_t watch_mode)
+{
+	int r = kfd_dbg_get_dev_watch_id(pdd, watch_id);
+
+	if (r)
+		return r;
+
+	if (!pdd->dev->kfd->shared_resources.enable_mes) {
+		r = debug_lock_and_unmap(pdd->dev->dqm);
+		if (r) {
+			kfd_dbg_clear_dev_watch_id(pdd, *watch_id);
+			return r;
+		}
+	}
+
+	amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
+	pdd->watch_points[*watch_id] = pdd->dev->kfd2kgd->set_address_watch(
+				pdd->dev->adev,
+				watch_address,
+				watch_address_mask,
+				*watch_id,
+				watch_mode,
+				pdd->dev->vm_info.last_vmid_kfd);
+	amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
+
+	if (!pdd->dev->kfd->shared_resources.enable_mes)
+		r = debug_map_and_unlock(pdd->dev->dqm);
+	else
+		r = kfd_dbg_set_mes_debug_mode(pdd);
+
+	/* HWS is broken so no point in HW rollback but release the watchpoint anyways */
+	if (r)
+		kfd_dbg_clear_dev_watch_id(pdd, *watch_id);
+
+	return 0;
+}
+
+static void kfd_dbg_clear_process_address_watch(struct kfd_process *target)
+{
+	int i, j;
+
+	for (i = 0; i < target->n_pdds; i++)
+		for (j = 0; j < MAX_WATCH_ADDRESSES; j++)
+			kfd_dbg_trap_clear_dev_address_watch(target->pdds[i], j);
+}
+
+
 /* kfd_dbg_trap_deactivate:
  *	target: target process
  *	unwind: If this is unwinding a failed kfd_dbg_trap_enable()
@@ -303,6 +438,7 @@ void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind
 
 	if (!unwind) {
 		cancel_work_sync(&target->debug_event_workarea);
+		kfd_dbg_clear_process_address_watch(target);
 		kfd_dbg_trap_set_wave_launch_mode(target, 0);
 	}
 

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

@@ -50,7 +50,13 @@ int kfd_dbg_trap_set_wave_launch_override(struct kfd_process *target,
 					uint32_t *trap_mask_supported);
 int kfd_dbg_trap_set_wave_launch_mode(struct kfd_process *target,
 					uint8_t wave_launch_mode);
-
+int kfd_dbg_trap_clear_dev_address_watch(struct kfd_process_device *pdd,
+					uint32_t watch_id);
+int kfd_dbg_trap_set_dev_address_watch(struct kfd_process_device *pdd,
+					uint64_t watch_address,
+					uint32_t watch_address_mask,
+					uint32_t *watch_id,
+					uint32_t watch_mode);
 int kfd_dbg_send_exception_to_runtime(struct kfd_process *p,
 					unsigned int dev_id,
 					unsigned int queue_id,

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

@@ -811,6 +811,8 @@ bool kgd2kfd_device_init(struct kfd_dev *kfd,
 	if (kfd_resume_iommu(kfd))
 		goto kfd_resume_iommu_error;
 
+	spin_lock_init(&kfd->watch_points_lock);
+
 	kfd->init_complete = true;
 	dev_info(kfd_device, "added device %x:%x\n", kfd->adev->pdev->vendor,
 		 kfd->adev->pdev->device);

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

@@ -381,6 +381,10 @@ struct kfd_dev {
 
 	struct kfd_node *nodes[MAX_KFD_NODES];
 	unsigned int num_nodes;
+
+	/* Track per device allocated watch points */
+	uint32_t alloc_watch_ids;
+	spinlock_t watch_points_lock;
 };
 
 enum kfd_mempool {
@@ -833,6 +837,7 @@ struct kfd_process_device {
 	uint32_t spi_dbg_override;
 	uint32_t spi_dbg_launch_mode;
 	uint32_t watch_points[4];
+	uint32_t alloc_watch_ids;
 
 	/*
 	 * If this process has been checkpointed before, then the user
@@ -989,7 +994,6 @@ struct kfd_process {
 	struct semaphore runtime_enable_sema;
 	bool is_runtime_retry;
 	struct kfd_runtime_info runtime_info;
-
 };
 
 #define KFD_PROCESS_TABLE_SIZE 5 /* bits: 32 entries */