Skip to contentSkip to conversation

drm/amdgpu/mes: add front end for detect and reset hung queue

PATCH New [email protected] 1 participant 12 messages Aug 1, 2025

Subsystems (2)
lore
Share & tools
lore

[email protected]

author maintainer
#0 lore raw

PATCH v6 1/12 drm/amdgpu/mes: add front end for detect and reset hung queue

From: Alex Deucher <[email protected]>

Helper function to detect and reset hung queues. MES will return an array of doorbell indices of which queues are hung and were optionally reset.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c | 62 +++++++++++++++++++++++++
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h | 19 ++++++++
 2 files changed, 81 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +62 -0
index 135598502c8d..64c5cac9ad5b 100644
@@ -191,6 +191,20 @@ int amdgpu_mes_init(struct amdgpu_device *adev)
 	if (r)
 		goto error_doorbell;
 
+	if (adev->mes.hung_queue_db_array_size) {
+		r = amdgpu_bo_create_kernel(adev,
+					    adev->mes.hung_queue_db_array_size * sizeof(u32),
+					    PAGE_SIZE,
+					    AMDGPU_GEM_DOMAIN_GTT,
+					    &adev->mes.hung_queue_db_array_gpu_obj,
+					    &adev->mes.hung_queue_db_array_gpu_addr,
+					    &adev->mes.hung_queue_db_array_cpu_addr);
+		if (r) {
+			dev_warn(adev->dev, "failed to create MES hung db array buffer (%d)", r);
+			goto error_doorbell;
+		}
+	}
+
 	return 0;
 
 error_doorbell:
@@ -216,6 +230,10 @@ void amdgpu_mes_fini(struct amdgpu_device *adev)
 {
 	int i;
 
+	amdgpu_bo_free_kernel(&adev->mes.hung_queue_db_array_gpu_obj,
+			      &adev->mes.hung_queue_db_array_gpu_addr,
+			      &adev->mes.hung_queue_db_array_cpu_addr);
+
 	amdgpu_bo_free_kernel(&adev->mes.event_log_gpu_obj,
 			      &adev->mes.event_log_gpu_addr,
 			      &adev->mes.event_log_cpu_addr);
@@ -366,6 +384,50 @@ int amdgpu_mes_reset_legacy_queue(struct amdgpu_device *adev,
 	return r;
 }
 
+int amdgpu_mes_get_hung_queue_db_array_size(struct amdgpu_device *adev)
+{
+	return adev->mes.hung_queue_db_array_size;
+}
+
+int amdgpu_mes_detect_and_reset_hung_queues(struct amdgpu_device *adev,
+					    int queue_type,
+					    bool detect_only,
+					    unsigned int *hung_db_num,
+					    u32 *hung_db_array)
+
+{
+	struct mes_detect_and_reset_queue_input input;
+	u32 *db_array = adev->mes.hung_queue_db_array_cpu_addr;
+	int r, i;
+
+	if (!hung_db_num || !hung_db_array)
+		return -EINVAL;
+
+	if ((queue_type != AMDGPU_RING_TYPE_GFX) &&
+	    (queue_type != AMDGPU_RING_TYPE_COMPUTE) &&
+	    (queue_type != AMDGPU_RING_TYPE_SDMA))
+		return -EINVAL;
+
+	input.queue_type = queue_type;
+	input.detect_only = detect_only;
+
+	r = adev->mes.funcs->detect_and_reset_hung_queues(&adev->mes,
+							  &input);
+	if (r) {
+		dev_err(adev->dev, "failed to detect and reset\n");
+	} else {
+		*hung_db_num = 0;
+		for (i = 0; i < adev->mes.hung_queue_db_array_size; i++) {
+			if (db_array[i] != AMDGPU_MES_INVALID_DB_OFFSET) {
+				hung_db_array[i] = db_array[i];
+				*hung_db_num += 1;
+			}
+		}
+	}
+
+	return r;
+}
+
 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg)
 {
 	struct mes_misc_op_input op_input;
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +19 -0
index c0d2c195fe2e..2c4568951edb 100644
@@ -41,6 +41,7 @@
 #define AMDGPU_MES_API_VERSION_MASK	0x00fff000
 #define AMDGPU_MES_FEAT_VERSION_MASK	0xff000000
 #define AMDGPU_MES_MSCRATCH_SIZE	0x40000
+#define AMDGPU_MES_INVALID_DB_OFFSET	0xffffffff
 
 enum amdgpu_mes_priority_level {
 	AMDGPU_MES_PRIORITY_LEVEL_LOW       = 0,
@@ -147,6 +148,10 @@ struct amdgpu_mes {
 	uint64_t            resource_1_gpu_addr[AMDGPU_MAX_MES_PIPES];
 	void                *resource_1_addr[AMDGPU_MAX_MES_PIPES];
 
+	int				hung_queue_db_array_size;
+	struct amdgpu_bo		*hung_queue_db_array_gpu_obj;
+	uint64_t			hung_queue_db_array_gpu_addr;
+	void				*hung_queue_db_array_cpu_addr;
 };
 
 struct amdgpu_mes_gang {
@@ -280,6 +285,11 @@ struct mes_reset_queue_input {
 	bool                               is_kq;
 };
 
+struct mes_detect_and_reset_queue_input {
+	uint32_t                           queue_type;
+	bool                               detect_only;
+};
+
 enum mes_misc_opcode {
 	MES_MISC_OP_WRITE_REG,
 	MES_MISC_OP_READ_REG,
@@ -367,6 +377,8 @@ struct amdgpu_mes_funcs {
 
 	int (*reset_hw_queue)(struct amdgpu_mes *mes,
 			      struct mes_reset_queue_input *input);
+	int (*detect_and_reset_hung_queues)(struct amdgpu_mes *mes,
+					    struct mes_detect_and_reset_queue_input *input);
 };
 
 #define amdgpu_mes_kiq_hw_init(adev) (adev)->mes.kiq_hw_init((adev))
@@ -390,6 +402,13 @@ int amdgpu_mes_reset_legacy_queue(struct amdgpu_device *adev,
 				  unsigned int vmid,
 				  bool use_mmio);
 
+int amdgpu_mes_get_hung_queue_db_array_size(struct amdgpu_device *adev);
+int amdgpu_mes_detect_and_reset_hung_queues(struct amdgpu_device *adev,
+					    int queue_type,
+					    bool detect_only,
+					    unsigned int *hung_db_num,
+					    u32 *hung_db_array);
+
 uint32_t amdgpu_mes_rreg(struct amdgpu_device *adev, uint32_t reg);
 int amdgpu_mes_wreg(struct amdgpu_device *adev,
 		    uint32_t reg, uint32_t val);
Signature · 1 line
2.49.0
author maintainer
#1 lore raw

PATCH v6 2/12 drm/amdgpu/mes11: implement detect and reset callback

From: Alex Deucher <[email protected]>

Implement support for the hung queue detect and reset functionality.

---
 drivers/gpu/drm/amd/amdgpu/mes_v11_0.c | 31 ++++++++++++++++++++++++++
 1 file changed, 31 insertions(+)
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +31 -0
index 28eb846280dd..ed6a7f8af544 100644
@@ -66,6 +66,8 @@ static int mes_v11_0_kiq_hw_fini(struct amdgpu_device *adev);
 #define GFX_MES_DRAM_SIZE	0x80000
 #define MES11_HW_RESOURCE_1_SIZE (128 * AMDGPU_GPU_PAGE_SIZE)
 
+#define MES11_HUNG_DB_OFFSET_ARRAY_SIZE 4
+
 static void mes_v11_0_ring_set_wptr(struct amdgpu_ring *ring)
 {
 	struct amdgpu_device *adev = ring->adev;
@@ -783,6 +785,32 @@ static int mes_v11_0_reset_hw_queue(struct amdgpu_mes *mes,
 			offsetof(union MESAPI__RESET, api_status));
 }
 
+static int mes_v11_0_detect_and_reset_hung_queues(struct amdgpu_mes *mes,
+						  struct mes_detect_and_reset_queue_input *input)
+{
+	union MESAPI__RESET mes_reset_queue_pkt;
+
+	memset(&mes_reset_queue_pkt, 0, sizeof(mes_reset_queue_pkt));
+
+	mes_reset_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
+	mes_reset_queue_pkt.header.opcode = MES_SCH_API_RESET;
+	mes_reset_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
+
+	mes_reset_queue_pkt.queue_type =
+		convert_to_mes_queue_type(input->queue_type);
+	mes_reset_queue_pkt.doorbell_offset_addr =
+		mes->hung_queue_db_array_gpu_addr;
+
+	if (input->detect_only)
+		mes_reset_queue_pkt.hang_detect_only = 1;
+	else
+		mes_reset_queue_pkt.hang_detect_then_reset = 1;
+
+	return mes_v11_0_submit_pkt_and_poll_completion(mes,
+			&mes_reset_queue_pkt, sizeof(mes_reset_queue_pkt),
+			offsetof(union MESAPI__RESET, api_status));
+}
+
 static const struct amdgpu_mes_funcs mes_v11_0_funcs = {
 	.add_hw_queue = mes_v11_0_add_hw_queue,
 	.remove_hw_queue = mes_v11_0_remove_hw_queue,
@@ -792,6 +820,7 @@ static const struct amdgpu_mes_funcs mes_v11_0_funcs = {
 	.resume_gang = mes_v11_0_resume_gang,
 	.misc_op = mes_v11_0_misc_op,
 	.reset_hw_queue = mes_v11_0_reset_hw_queue,
+	.detect_and_reset_hung_queues = mes_v11_0_detect_and_reset_hung_queues,
 };
 
 static int mes_v11_0_allocate_ucode_buffer(struct amdgpu_device *adev,
@@ -1684,6 +1713,8 @@ static int mes_v11_0_early_init(struct amdgpu_ip_block *ip_block)
 	struct amdgpu_device *adev = ip_block->adev;
 	int pipe, r;
 
+	adev->mes.hung_queue_db_array_size =
+		MES11_HUNG_DB_OFFSET_ARRAY_SIZE;
 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
 		if (!adev->enable_mes_kiq && pipe == AMDGPU_MES_KIQ_PIPE)
 			continue;
Signature · 1 line
2.49.0
author maintainer
#2 lore raw

PATCH v6 3/12 drm/amdgpu/mes12: implement detect and reset callback

In reply to [email protected]

From: Alex Deucher <[email protected]>

Implement support for the hung queue detect and reset functionality.

---
 drivers/gpu/drm/amd/amdgpu/mes_v12_0.c | 37 ++++++++++++++++++++++++++
 1 file changed, 37 insertions(+)
drivers/gpu/drm/amd/amdgpu/mes_v12_0.c +37 -0
index 6b222630f3fa..29d38aa1897e 100644
@@ -47,6 +47,8 @@ static int mes_v12_0_kiq_hw_fini(struct amdgpu_device *adev);
 
 #define MES_EOP_SIZE   2048
 
+#define MES12_HUNG_DB_OFFSET_ARRAY_SIZE 4
+
 static void mes_v12_0_ring_set_wptr(struct amdgpu_ring *ring)
 {
 	struct amdgpu_device *adev = ring->adev;
@@ -879,6 +881,38 @@ static int mes_v12_0_reset_hw_queue(struct amdgpu_mes *mes,
 			offsetof(union MESAPI__RESET, api_status));
 }
 
+static int mes_v12_0_detect_and_reset_hung_queues(struct amdgpu_mes *mes,
+						  struct mes_detect_and_reset_queue_input *input)
+{
+	union MESAPI__RESET mes_reset_queue_pkt;
+	int pipe;
+
+	memset(&mes_reset_queue_pkt, 0, sizeof(mes_reset_queue_pkt));
+
+	mes_reset_queue_pkt.header.type = MES_API_TYPE_SCHEDULER;
+	mes_reset_queue_pkt.header.opcode = MES_SCH_API_RESET;
+	mes_reset_queue_pkt.header.dwsize = API_FRAME_SIZE_IN_DWORDS;
+
+	mes_reset_queue_pkt.queue_type =
+		convert_to_mes_queue_type(input->queue_type);
+	mes_reset_queue_pkt.doorbell_offset_addr =
+		mes->hung_queue_db_array_gpu_addr;
+
+	if (input->detect_only)
+		mes_reset_queue_pkt.hang_detect_only = 1;
+	else
+		mes_reset_queue_pkt.hang_detect_then_reset = 1;
+
+	if (mes->adev->enable_uni_mes)
+		pipe = AMDGPU_MES_KIQ_PIPE;
+	else
+		pipe = AMDGPU_MES_SCHED_PIPE;
+
+	return mes_v12_0_submit_pkt_and_poll_completion(mes, pipe,
+			&mes_reset_queue_pkt, sizeof(mes_reset_queue_pkt),
+			offsetof(union MESAPI__RESET, api_status));
+}
+
 static const struct amdgpu_mes_funcs mes_v12_0_funcs = {
 	.add_hw_queue = mes_v12_0_add_hw_queue,
 	.remove_hw_queue = mes_v12_0_remove_hw_queue,
@@ -888,6 +922,7 @@ static const struct amdgpu_mes_funcs mes_v12_0_funcs = {
 	.resume_gang = mes_v12_0_resume_gang,
 	.misc_op = mes_v12_0_misc_op,
 	.reset_hw_queue = mes_v12_0_reset_hw_queue,
+	.detect_and_reset_hung_queues = mes_v12_0_detect_and_reset_hung_queues,
 };
 
 static int mes_v12_0_allocate_ucode_buffer(struct amdgpu_device *adev,
@@ -1793,6 +1828,8 @@ static int mes_v12_0_early_init(struct amdgpu_ip_block *ip_block)
 	struct amdgpu_device *adev = ip_block->adev;
 	int pipe, r;
 
+	adev->mes.hung_queue_db_array_size =
+		MES12_HUNG_DB_OFFSET_ARRAY_SIZE;
 	for (pipe = 0; pipe < AMDGPU_MAX_MES_PIPES; pipe++) {
 		r = amdgpu_mes_init_microcode(adev, pipe);
 		if (r)
Signature · 1 line
2.49.0
author maintainer
#3 lore raw

PATCH v6 4/12 drm/amdgpu: Implement active VMID detection in MES11 queue reset for GFX

In reply to [email protected]

From: "Jesse.Zhang" <[email protected]>

MES queue reset functionality for GFX queues. The changes include:

1. Added detection of active VMIDs by reading CP_CNTX_STAT and CP_VMID registers to properly identify contexts that need resetting

2. Implemented fallback to HPD status method when no active VMIDs are found, checking both pipe 0 and pipe 1 queues

3. Extended the MES reset packet with:
- active_vmids bitmap
- connected_queue_index for pipe 0
- connected_queue_index_p1 for pipe 1

---
 drivers/gpu/drm/amd/amdgpu/mes_v11_0.c        | 51 +++++++++++++++++++
 drivers/gpu/drm/amd/include/mes_v11_api_def.h | 13 ++++-
 2 files changed, 63 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +51 -0
index ed6a7f8af544..1422bc59cd40 100644
@@ -789,6 +789,12 @@ static int mes_v11_0_detect_and_reset_hung_queues(struct amdgpu_mes *mes,
 						  struct mes_detect_and_reset_queue_input *input)
 {
 	union MESAPI__RESET mes_reset_queue_pkt;
+	struct amdgpu_device *adev = mes->adev;
+	uint32_t active_vmids = 0;
+	uint32_t connected_queue_index = 0;
+	uint32_t queue_status = 0;
+	uint32_t connected_queue_index_p1 = 0;
+	uint32_t queue_status_p1 = 0;
 
 	memset(&mes_reset_queue_pkt, 0, sizeof(mes_reset_queue_pkt));
 
@@ -801,6 +807,51 @@ static int mes_v11_0_detect_and_reset_hung_queues(struct amdgpu_mes *mes,
 	mes_reset_queue_pkt.doorbell_offset_addr =
 		mes->hung_queue_db_array_gpu_addr;
 
+	/* Add VMID detection for GFX queues */
+	if (input->queue_type == AMDGPU_RING_TYPE_GFX) {
+		uint32_t cp_cntx_stat = RREG32_SOC15(GC, 0, regCP_CNTX_STAT);
+		uint32_t cp_vmid, grbm_gfx_cntl;
+
+		/* Check active contexts in CP_CNTX_STAT */
+		for (uint32_t i = 0; i < 8; i++) {
+			if ((cp_cntx_stat >> (0x14 + i)) & 0x1) {
+				grbm_gfx_cntl = (i << 11);
+				WREG32_SOC15(GC, 0, regGRBM_GFX_CNTL, grbm_gfx_cntl);
+				cp_vmid = RREG32_SOC15(GC, 0, regCP_VMID);
+				active_vmids |= (1 << cp_vmid);
+			}
+		}
+
+		/* Fallback to HPD status if no active VMIDs found */
+		if (active_vmids == 0) {
+			uint32_t hpd_status;
+
+			/* Pipe 0 */
+			WREG32_SOC15(GC, 0, regGRBM_GFX_CNTL, 0);
+			hpd_status = RREG32_SOC15(GC, 0, regCP_GFX_HPD_STATUS0);
+			queue_status = hpd_status & 0x1F;
+			connected_queue_index = (hpd_status & 0xE0) >> 5;
+
+			/* Pipe 1 */
+			WREG32_SOC15(GC, 0, regGRBM_GFX_CNTL, (1 << 6));
+			hpd_status = RREG32_SOC15(GC, 0, regCP_GFX_HPD_STATUS0);
+			queue_status_p1 = hpd_status & 0x1F;
+			connected_queue_index_p1 = (hpd_status & 0xE0) >> 5;
+		}
+
+		mes_reset_queue_pkt.active_vmids = active_vmids;
+		if (active_vmids == 0) {
+			if (queue_status != 0) {
+				mes_reset_queue_pkt.use_connected_queue_index = 1;
+				mes_reset_queue_pkt.connected_queue_index = connected_queue_index;
+			}
+			if (queue_status_p1 != 0) {
+				mes_reset_queue_pkt.use_connected_queue_index_p1 = 1;
+				mes_reset_queue_pkt.connected_queue_index_p1 = connected_queue_index_p1;
+			}
+		}
+	}
+
 	if (input->detect_only)
 		mes_reset_queue_pkt.hang_detect_only = 1;
 	else
drivers/gpu/drm/amd/include/mes_v11_api_def.h +12 -1
index 15680c3f4970..62ad4f0337eb 100644
@@ -460,7 +460,11 @@ union MESAPI__RESET {
 			uint32_t                hang_detect_only : 1;
 			/* Rest HP and LP kernel queues not managed by MES */
 			uint32_t                reset_legacy_gfx : 1;
-			uint32_t                reserved : 28;
+			/* Fallback to use conneceted queue index when CP_CNTX_STAT method fails (gfx pipe 0) */
+			uint32_t		use_connected_queue_index : 1;
+			/* For gfx pipe 1 */
+			uint32_t		use_connected_queue_index_p1 : 1;
+			uint32_t                reserved : 26;
 		};
 
 		uint64_t			gang_context_addr;
@@ -488,6 +492,13 @@ union MESAPI__RESET {
 		uint64_t			wptr_addr_hp;
 
 		struct MES_API_STATUS		api_status;
+		uint32_t			active_vmids;
+		uint64_t			timestamp;
+
+		uint32_t			gang_context_array_index;
+
+		uint32_t			connected_queue_index;
+		uint32_t			connected_queue_index_p1;
 	};
 
 	uint32_t	max_dwords_in_api[API_FRAME_SIZE_IN_DWORDS];
Signature · 1 line
2.49.0
author maintainer
#4 lore raw

PATCH v6 5/12 drm/amdgpu: Implement active VMID detection in MES12 queue reset for GFX

In reply to [email protected]

From: "Jesse.Zhang" <[email protected]>

MES queue reset functionality for GFX queues. The changes include:

1. Added detection of active VMIDs by reading CP_CNTX_STAT and CP_VMID registers to properly identify contexts that need resetting

2. Implemented fallback to HPD status method when no active VMIDs are found, checking both pipe 0 and pipe 1 queues

3. Extended the MES reset packet with:
- active_vmids bitmap
- connected_queue_index for pipe 0
- connected_queue_index_p1 for pipe 1

---
 drivers/gpu/drm/amd/amdgpu/mes_v12_0.c | 51 ++++++++++++++++++++++++++
 1 file changed, 51 insertions(+)
drivers/gpu/drm/amd/amdgpu/mes_v12_0.c +51 -0
index 29d38aa1897e..579720695e9e 100644
@@ -885,6 +885,12 @@ static int mes_v12_0_detect_and_reset_hung_queues(struct amdgpu_mes *mes,
 						  struct mes_detect_and_reset_queue_input *input)
 {
 	union MESAPI__RESET mes_reset_queue_pkt;
+	struct amdgpu_device *adev = mes->adev;
+	uint32_t active_vmids = 0;
+	uint32_t connected_queue_index = 0;
+	uint32_t queue_status = 0;
+	uint32_t connected_queue_index_p1 = 0;
+	uint32_t queue_status_p1 = 0;
 	int pipe;
 
 	memset(&mes_reset_queue_pkt, 0, sizeof(mes_reset_queue_pkt));
@@ -898,6 +904,51 @@ static int mes_v12_0_detect_and_reset_hung_queues(struct amdgpu_mes *mes,
 	mes_reset_queue_pkt.doorbell_offset_addr =
 		mes->hung_queue_db_array_gpu_addr;
 
+	/* Add VMID detection for GFX queues */
+	if (input->queue_type == AMDGPU_RING_TYPE_GFX) {
+		uint32_t cp_cntx_stat = RREG32_SOC15(GC, 0, regCP_CNTX_STAT);
+		uint32_t cp_vmid, grbm_gfx_cntl;
+
+		/* Check active contexts in CP_CNTX_STAT */
+		for (uint32_t i = 0; i < 8; i++) {
+			if ((cp_cntx_stat >> (0x14 + i)) & 0x1) {
+				grbm_gfx_cntl = (i << 11);
+				WREG32_SOC15(GC, 0, regGRBM_GFX_CNTL, grbm_gfx_cntl);
+				cp_vmid = RREG32_SOC15(GC, 0, regCP_CP_VMID);
+				active_vmids |= (1 << cp_vmid);
+			}
+		}
+
+		/* Fallback to HPD status if no active VMIDs found */
+		if (active_vmids == 0) {
+			uint32_t hpd_status;
+
+			/* Pipe 0 */
+			WREG32_SOC15(GC, 0, regGRBM_GFX_CNTL, 0);
+			hpd_status = RREG32_SOC15(GC, 0, regCP_GFX_HPD_STATUS0);
+			queue_status = hpd_status & 0x1F;
+			connected_queue_index = (hpd_status & 0xE0) >> 5;
+
+			/* Pipe 1 */
+			WREG32_SOC15(GC, 0, regGRBM_GFX_CNTL, (1 << 6));
+			hpd_status = RREG32_SOC15(GC, 0, regCP_GFX_HPD_STATUS0);
+			queue_status_p1 = hpd_status & 0x1F;
+			connected_queue_index_p1 = (hpd_status & 0xE0) >> 5;
+		}
+
+		mes_reset_queue_pkt.active_vmids = active_vmids;
+		if (active_vmids == 0) {
+			if (queue_status != 0) {
+				mes_reset_queue_pkt.use_connected_queue_index = 1;
+				mes_reset_queue_pkt.connected_queue_index = connected_queue_index;
+			}
+			if (queue_status_p1 != 0) {
+				mes_reset_queue_pkt.use_connected_queue_index_p1 = 1;
+				mes_reset_queue_pkt.connected_queue_index_p1 = connected_queue_index_p1;
+			}
+		}
+	}
+
 	if (input->detect_only)
 		mes_reset_queue_pkt.hang_detect_only = 1;
 	else
Signature · 1 line
2.49.0
author maintainer
#5 lore raw

PATCH v6 6/12 drm/amdgpu/userq: add a detect and reset callback

In reply to [email protected]

From: Alex Deucher <[email protected]>

Add a detect and reset callback and add the implementation for mes. The callback will detect all hung queues of a particular ip type (e.g., GFX or compute or SDMA) and reset them.

v2: increase reset counter and set fence force completion
v3: Removed userq_mutex in mes_userq_detect_and_reset since the driver holds it when calling

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h  |  3 ++
 drivers/gpu/drm/amd/amdgpu/mes_userqueue.c | 49 ++++++++++++++++++++++
 2 files changed, 52 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h +3 -0
index ec040c2fd6c9..0335ff03f65f 100644
@@ -77,6 +77,9 @@ struct amdgpu_userq_funcs {
 		     struct amdgpu_usermode_queue *queue);
 	int (*map)(struct amdgpu_userq_mgr *uq_mgr,
 		   struct amdgpu_usermode_queue *queue);
+	int (*detect_and_reset)(struct amdgpu_device *adev,
+				int queue_type);
+
 };
 
 /* Usermode queues for gfx */
drivers/gpu/drm/amd/amdgpu/mes_userqueue.c +49 -0

65 lines not on this page · raw message

Signature · 1 line
2.49.0
author maintainer
#6 lore raw

PATCH v6 7/12 drm/amd: Add preempt and restore callbacks to userq funcs

In reply to [email protected]

From: Alex Deucher <[email protected]>

Add two new function pointers to struct amdgpu_userq_funcs:
- preempt: To handle preemption of user mode queues
- restore: To restore preempted user mode queues

These callbacks will allow the driver to properly manage queue preemption and restoration when needed, such as during context switching or priority changes.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h | 4 ++++
 1 file changed, 4 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h +4 -0
index 0335ff03f65f..68e46d01bed2 100644
@@ -77,6 +77,10 @@ struct amdgpu_userq_funcs {
 		     struct amdgpu_usermode_queue *queue);
 	int (*map)(struct amdgpu_userq_mgr *uq_mgr,
 		   struct amdgpu_usermode_queue *queue);
+	int (*preempt)(struct amdgpu_userq_mgr *uq_mgr,
+		   struct amdgpu_usermode_queue *queue);
+	int (*restore)(struct amdgpu_userq_mgr *uq_mgr,
+		   struct amdgpu_usermode_queue *queue);
 	int (*detect_and_reset)(struct amdgpu_device *adev,
 				int queue_type);
 
Signature · 1 line
2.49.0
author maintainer
#7 lore raw

PATCH v6 8/12 drm/amdgpu: adjust MES API used for suspend and resume

In reply to [email protected]

From: Alex Deucher <[email protected]>

Use the suspend and resume API rather than remove queue and add queue API. The former just preempts the queue while the latter remove it from the scheduler completely. There is no need to do that, we only need preemption in this case.

V2: remove queue_active

---
 drivers/gpu/drm/amd/amdgpu/mes_userqueue.c | 42 ++++++++++++++++++++++
 1 file changed, 42 insertions(+)
drivers/gpu/drm/amd/amdgpu/mes_userqueue.c +42 -0
index a871bac71e1e..edd4b1757b31 100644
@@ -395,10 +395,52 @@ mes_userq_mqd_destroy(struct amdgpu_userq_mgr *uq_mgr,
 	amdgpu_userq_destroy_object(uq_mgr, &queue->mqd);
 }
 
+static int mes_userq_preempt(struct amdgpu_userq_mgr *uq_mgr,
+			     struct amdgpu_usermode_queue *queue)
+{
+	struct amdgpu_device *adev = uq_mgr->adev;
+	struct mes_suspend_gang_input queue_input;
+	struct amdgpu_userq_obj *ctx = &queue->fw_obj;
+	int r;
+
+
+	memset(&queue_input, 0x0, sizeof(struct mes_suspend_gang_input));
+	queue_input.gang_context_addr = ctx->gpu_addr + AMDGPU_USERQ_PROC_CTX_SZ;
+
+	amdgpu_mes_lock(&adev->mes);
+	r = adev->mes.funcs->suspend_gang(&adev->mes, &queue_input);
+	amdgpu_mes_unlock(&adev->mes);
+	if (r)
+		dev_err(adev->dev, "Failed to suspend queue, err (%d)\n", r);
+	return r;
+}
+
+static int mes_userq_restore(struct amdgpu_userq_mgr *uq_mgr,
+			    struct amdgpu_usermode_queue *queue)
+{
+	struct amdgpu_device *adev = uq_mgr->adev;
+	struct mes_resume_gang_input queue_input;
+	struct amdgpu_userq_obj *ctx = &queue->fw_obj;
+	int r;
+
+
+	memset(&queue_input, 0x0, sizeof(struct mes_resume_gang_input));
+	queue_input.gang_context_addr = ctx->gpu_addr + AMDGPU_USERQ_PROC_CTX_SZ;
+
+	amdgpu_mes_lock(&adev->mes);
+	r = adev->mes.funcs->resume_gang(&adev->mes, &queue_input);
+	amdgpu_mes_unlock(&adev->mes);
+	if (r)
+		dev_err(adev->dev, "Failed to resume queue, err (%d)\n", r);
+	return r;
+ }
+
 const struct amdgpu_userq_funcs userq_mes_funcs = {
 	.mqd_create = mes_userq_mqd_create,
 	.mqd_destroy = mes_userq_mqd_destroy,
 	.unmap = mes_userq_unmap,
 	.map = mes_userq_map,
+	.preempt = mes_userq_preempt,
+	.restore = mes_userq_restore,
 	.detect_and_reset = mes_userq_detect_and_reset,
 };
Signature · 1 line
2.49.0
author maintainer
#8 lore raw

PATCH v6 9/12 drm/amdgpu: add user queue reset source

In reply to [email protected]

From: Alex Deucher <[email protected]>

Track resets from user queues.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c | 3 +++
 drivers/gpu/drm/amd/amdgpu/amdgpu_reset.h | 1 +
 2 files changed, 4 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_reset.c +3 -0
index dabfbdf6f1ce..28c4ad62f50e 100644
@@ -340,6 +340,9 @@ void amdgpu_reset_get_desc(struct amdgpu_reset_context *rst_ctxt, char *buf,
 	case AMDGPU_RESET_SRC_USER:
 		strscpy(buf, "user trigger", len);
 		break;
+	case AMDGPU_RESET_SRC_USERQ:
+		strscpy(buf, "user queue trigger", len);
+		break;
 	default:
 		strscpy(buf, "unknown", len);
 	}
drivers/gpu/drm/amd/amdgpu/amdgpu_reset.h +1 -0
index 4d9b9701139b..ebcea44dd743 100644
@@ -43,6 +43,7 @@ enum AMDGPU_RESET_SRCS {
 	AMDGPU_RESET_SRC_MES,
 	AMDGPU_RESET_SRC_HWS,
 	AMDGPU_RESET_SRC_USER,
+	AMDGPU_RESET_SRC_USERQ,
 };
 
 struct amdgpu_reset_context {
Signature · 1 line
2.49.0
author maintainer
#9 lore raw

PATCH v6 10/12 drm/amdgpu/userq: add force completion helpers

In reply to [email protected]

From: Alex Deucher <[email protected]>

Add support for forcing completion of userq fences. This is needed for userq resets and asic resets so that we can set the error on the fence and force completion.

---
 .../gpu/drm/amd/amdgpu/amdgpu_userq_fence.c   | 42 +++++++++++++++++++
 .../gpu/drm/amd/amdgpu/amdgpu_userq_fence.h   |  1 +
 2 files changed, 43 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c +42 -0
index a86616c6deef..01688bbf3f56 100644
@@ -67,6 +67,14 @@ static u64 amdgpu_userq_fence_read(struct amdgpu_userq_fence_driver *fence_drv)
 	return le64_to_cpu(*fence_drv->cpu_addr);
 }
 
+static void
+amdgpu_userq_fence_write(struct amdgpu_userq_fence_driver *fence_drv,
+			 u64 seq)
+{
+	if (fence_drv->cpu_addr)
+		*fence_drv->cpu_addr = cpu_to_le64(seq);
+}
+
 int amdgpu_userq_fence_driver_alloc(struct amdgpu_device *adev,
 				    struct amdgpu_usermode_queue *userq)
 {
@@ -409,6 +417,40 @@ static void amdgpu_userq_fence_cleanup(struct dma_fence *fence)
 	dma_fence_put(fence);
 }
 
+static void
+amdgpu_userq_fence_driver_set_error(struct amdgpu_userq_fence *fence,
+				    int error)
+{
+	struct amdgpu_userq_fence_driver *fence_drv = fence->fence_drv;
+	unsigned long flags;
+	struct dma_fence *f;
+
+	spin_lock_irqsave(&fence_drv->fence_list_lock, flags);
+
+	f = rcu_dereference_protected(&fence->base,
+				      lockdep_is_held(&fence_drv->fence_list_lock));
+	if (f && !dma_fence_is_signaled_locked(f))
+		dma_fence_set_error(f, error);
+	spin_unlock_irqrestore(&fence_drv->fence_list_lock, flags);
+}
+
+void
+amdgpu_userq_fence_driver_force_completion(struct amdgpu_usermode_queue *userq)
+{
+	struct dma_fence *f = userq->last_fence;
+
+	if (f) {
+		struct amdgpu_userq_fence *fence = to_amdgpu_userq_fence(f);
+		struct amdgpu_userq_fence_driver *fence_drv = fence->fence_drv;
+		u64 wptr = fence->base.seqno;
+
+		amdgpu_userq_fence_driver_set_error(fence, -ECANCELED);
+		amdgpu_userq_fence_write(fence_drv, wptr);
+		amdgpu_userq_fence_driver_process(fence_drv);
+
+	}
+}
+
 int amdgpu_userq_signal_ioctl(struct drm_device *dev, void *data,
 			      struct drm_file *filp)
 {
drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.h +1 -0
index 97a125ab8a78..d76add2afc77 100644
@@ -67,6 +67,7 @@ int amdgpu_userq_fence_driver_alloc(struct amdgpu_device *adev,
 				    struct amdgpu_usermode_queue *userq);
 void amdgpu_userq_fence_driver_free(struct amdgpu_usermode_queue *userq);
 void amdgpu_userq_fence_driver_process(struct amdgpu_userq_fence_driver *fence_drv);
+void amdgpu_userq_fence_driver_force_completion(struct amdgpu_usermode_queue *userq);
 void amdgpu_userq_fence_driver_destroy(struct kref *ref);
 int amdgpu_userq_signal_ioctl(struct drm_device *dev, void *data,
 			      struct drm_file *filp);
Signature · 1 line
2.49.0
author maintainer
#10 lore raw

PATCH v6 11/12 drm/amdgpu: Add GPU reset handling for user mode queues

In reply to [email protected]

From: Alex Deucher <[email protected]>

This patch introduces GPU reset support for userq.
It adds detection and recovery logic for gfx, compute, and SDMA queues when errors occur during queue operations (e.g., map/unmap failures).

Key changes include:
- New function amdgpu_userq_detect_and_reset_queues() to scan active queues and attempt recovery via queue-specific handlers.
- Integration of userq reset logic into suspend/resume, eviction, and reset paths.
- New pre-reset and post-reset hooks (amdgpu_userq_pre_reset() and amdgpu_userq_post_reset()) for cleaning up and reinitializing user queues during full GPU reset.
- A new work_struct (userq_reset_work) is introduced and used in amdgpu_userq_gpu_reset() to trigger GPU recovery using the standard GPU reset framework.

This enhances user queue resiliency and ensures graceful recovery in case of individual queue failure without immediately requiring full GPU reset.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu.h        |   1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c |   8 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c  | 200 +++++++++++++++++++--
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h  |   5 +
 4 files changed, 203 insertions(+), 11 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +1 -0
index ef3af170dda4..9db05cdc7304 100644
@@ -1302,6 +1302,7 @@ struct amdgpu_device {
 	struct list_head		userq_mgr_list;
 	struct mutex                    userq_mutex;
 	bool                            userq_halt_for_enforce_isolation;
+	struct work_struct		userq_reset_work;
 };
 
 static inline uint32_t amdgpu_ip_version(const struct amdgpu_device *adev,
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +8 -0
index 3757634613c3..1dc88b0055dd 100644
@@ -4475,6 +4475,7 @@ int amdgpu_device_init(struct amdgpu_device *adev,
 	}
 
 	INIT_WORK(&adev->xgmi_reset_work, amdgpu_device_xgmi_reset_func);
+	INIT_WORK(&adev->userq_reset_work, amdgpu_userq_reset_work);
 
 	adev->gfx.gfx_off_req_count = 1;
 	adev->gfx.gfx_off_residency = 0;
@@ -5880,6 +5881,10 @@ int amdgpu_device_reinit_after_reset(struct amdgpu_reset_context *reset_context)
 				if (r)
 					goto out;
 
+				r = amdgpu_userq_post_reset(tmp_adev, vram_lost);
+				if (r)
+					goto out;
+
 				drm_client_dev_resume(adev_to_drm(tmp_adev), false);
 
 				/*
@@ -6102,6 +6107,7 @@ static inline void amdgpu_device_stop_pending_resets(struct amdgpu_device *adev)
 	if (!amdgpu_sriov_vf(adev))
 		cancel_work(&adev->reset_work);
 #endif
+	cancel_work(&adev->userq_reset_work);
 
 	if (adev->kfd.dev)
 		cancel_work(&adev->kfd.reset_work);
@@ -6232,6 +6238,8 @@ static void amdgpu_device_halt_activities(struct amdgpu_device *adev,
 		      amdgpu_device_ip_need_full_reset(tmp_adev))
 			amdgpu_ras_suspend(tmp_adev);
 
+		amdgpu_userq_pre_reset(tmp_adev);
+
 		for (i = 0; i < AMDGPU_MAX_RINGS; ++i) {
 			struct amdgpu_ring *ring = tmp_adev->rings[i];
 
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +189 -11

294 lines not on this page · raw message

drivers/gpu/drm/amd/amdgpu/amdgpu_userq.h +5 -0
index 68e46d01bed2..c136d7e7a763 100644
@@ -139,4 +139,9 @@ int amdgpu_userq_stop_sched_for_enforce_isolation(struct amdgpu_device *adev,
 int amdgpu_userq_start_sched_for_enforce_isolation(struct amdgpu_device *adev,
 						   u32 idx);
 
+void amdgpu_userq_reset_work(struct work_struct *work);
+
+void amdgpu_userq_pre_reset(struct amdgpu_device *adev);
+int amdgpu_userq_post_reset(struct amdgpu_device *adev, bool vram_lost);
+
 #endif
Signature · 1 line
2.49.0
author maintainer
#11 lore raw

PATCH v6 12/12 drm/amdgpu: Implement queue preemption using suspend/resume API

In reply to [email protected]

From: Alex Deucher <[email protected]>

Replace the queue remove/add approach with suspend/resume semantics for user queue preemption. This change:

1. Maintains queue scheduling registration while only preempting execution
- Previously used remove_queue/add_queue would fully deregister queues
- New suspend/resume approach keeps scheduler state while preempting

2. Introduces proper preemption helpers:
- amdgpu_userqueue_preempt_helper(): Suspends queue execution
- Transitions MAPPED→UNMAPPED state on success
- Marks as HUNG and triggers reset on failure
- amdgpu_userqueue_restore_helper(): Resumes queue execution
- Transitions UNMAPPED→MAPPED state on success
- Triggers GPU reset on failure

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 53 +++++++++++++++++++++--
 1 file changed, 50 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c +50 -3
index 0c91302162fa..af0ac4b73ddf 100644
@@ -185,6 +185,54 @@ amdgpu_userq_map_helper(struct amdgpu_userq_mgr *uq_mgr,
 	return r;
 }
 
+static int
+amdgpu_userqueue_preempt_helper(struct amdgpu_userq_mgr *uq_mgr,
+			  struct amdgpu_usermode_queue *queue)
+{
+	struct amdgpu_device *adev = uq_mgr->adev;
+	const struct amdgpu_userq_funcs *userq_funcs =
+		adev->userq_funcs[queue->queue_type];
+	int r = 0;
+
+	if (queue->state == AMDGPU_USERQ_STATE_MAPPED) {
+		r = userq_funcs->preempt(uq_mgr, queue);
+		if (r) {
+			amdgpu_userq_detect_and_reset_queues(uq_mgr);
+			queue->state = AMDGPU_USERQ_STATE_HUNG;
+		} else {
+			queue->state = AMDGPU_USERQ_STATE_UNMAPPED;
+		}
+	}
+
+	return r;
+}
+
+static int
+amdgpu_userqueue_restore_helper(struct amdgpu_userq_mgr *uq_mgr,
+			struct amdgpu_usermode_queue *queue)
+{
+	struct amdgpu_device *adev = uq_mgr->adev;
+	const struct amdgpu_userq_funcs *userq_funcs =
+		adev->userq_funcs[queue->queue_type];
+	bool gpu_reset = false;
+	int r = 0;
+
+	if (queue->state == AMDGPU_USERQ_STATE_UNMAPPED) {
+		r = userq_funcs->restore(uq_mgr, queue);
+		if (r) {
+			queue->state = AMDGPU_USERQ_STATE_HUNG;
+			gpu_reset = true;
+		} else {
+			queue->state = AMDGPU_USERQ_STATE_MAPPED;
+		}
+	}
+
+	if (gpu_reset)
+		amdgpu_userq_gpu_reset(adev);
+
+	return r;
+}
+
 static void
 amdgpu_userq_wait_for_last_fence(struct amdgpu_userq_mgr *uq_mgr,
 				 struct amdgpu_usermode_queue *queue)
@@ -639,7 +687,7 @@ amdgpu_userq_restore_all(struct amdgpu_userq_mgr *uq_mgr)
 
 	/* Resume all the queues for this process */
 	idr_for_each_entry(&uq_mgr->userq_idr, queue, queue_id) {
-		r = amdgpu_userq_map_helper(uq_mgr, queue);
+		r = amdgpu_userqueue_restore_helper(uq_mgr, queue);
 		if (r)
 			ret = r;
 	}
@@ -794,10 +842,9 @@ amdgpu_userq_evict_all(struct amdgpu_userq_mgr *uq_mgr)
 	int queue_id;
 	int ret = 0, r;
 
-	amdgpu_userq_detect_and_reset_queues(uq_mgr);
 	/* Try to unmap all the queues in this process ctx */
 	idr_for_each_entry(&uq_mgr->userq_idr, queue, queue_id) {
-		r = amdgpu_userq_unmap_helper(uq_mgr, queue);
+		r = amdgpu_userqueue_preempt_helper(uq_mgr, queue);
 		if (r)
 			ret = r;
 	}
Signature · 1 line
2.49.0