Skip to contentSkip to conversation

drm/amdkfd: support per-queue reset on gfx9

PATCH In review Jonathan Kim 3 participants 9 messages Jul 26, 2024 – Jul 30, 2024

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/2 drm/amdkfd: support per-queue reset on gfx9

Support per-queue reset for GFX9. The recommendation is for the driver to target reset the HW queue via a SPI MMIO register write.

Since this requires pipe and HW queue info and MEC FW is limited to doorbell reports of hung queues after an unmap failure, scan the HW queue slots defined by SET_RESOURCES first to identify the user queue candidates to reset.

Only signal reset events to processes that have had a queue reset.

If queue reset fails, fall back to GPU reset.

v2: move reset queue flag for house keeping to process device. split detect and reset into separate functions.
make reset call safe during power saving modes. clean up some other nitpicks.

---
 .../drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c  |   2 +
 .../drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c   |   4 +-
 .../drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c   |   4 +-
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c    |  16 ++
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h    |   9 +
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c  |   4 +-
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c    |  18 +-
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c |  85 +++++++++
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h |   9 +
 .../drm/amd/amdkfd/kfd_device_queue_manager.c | 172 +++++++++++++++++-
 .../drm/amd/amdkfd/kfd_device_queue_manager.h |  12 ++
 drivers/gpu/drm/amd/amdkfd/kfd_events.c       |  21 +++
 .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c   |   6 +-
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |   3 +
 drivers/gpu/drm/amd/amdkfd/kfd_process.c      |   2 +
 .../gpu/drm/amd/include/kgd_kfd_interface.h   |   6 +
 16 files changed, 360 insertions(+), 13 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c +2 -0
index aff08321e976..8dfdb18197c4 100644
@@ -191,4 +191,6 @@ const struct kfd2kgd_calls aldebaran_kfd2kgd = {
 	.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,
+	.hqd_get_pq_addr = kgd_gfx_v9_hqd_get_pq_addr,
+	.hqd_reset = kgd_gfx_v9_hqd_reset,
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c +3 -1
index 3a3f3ce09f00..017e8a3013aa 100644
@@ -418,5 +418,7 @@ const struct kfd2kgd_calls arcturus_kfd2kgd = {
 	.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,
-	.program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings
+	.program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings,
+	.hqd_get_pq_addr = kgd_gfx_v9_hqd_get_pq_addr,
+	.hqd_reset = kgd_gfx_v9_hqd_reset
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c +3 -1
index a5c7259cf2a3..e2ae714a700f 100644
@@ -541,5 +541,7 @@ const struct kfd2kgd_calls gc_9_4_3_kfd2kgd = {
 			kgd_gfx_v9_4_3_set_wave_launch_trap_override,
 	.set_wave_launch_mode = kgd_aldebaran_set_wave_launch_mode,
 	.set_address_watch = kgd_gfx_v9_4_3_set_address_watch,
-	.clear_address_watch = kgd_gfx_v9_4_3_clear_address_watch
+	.clear_address_watch = kgd_gfx_v9_4_3_clear_address_watch,
+	.hqd_get_pq_addr = kgd_gfx_v9_hqd_get_pq_addr,
+	.hqd_reset = kgd_gfx_v9_hqd_reset
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c +16 -0
index 3ab6c3aa0ad1..62176d607bef 100644
@@ -1070,6 +1070,20 @@ static void program_trap_handler_settings(struct amdgpu_device *adev,
 	unlock_srbm(adev);
 }
 
+uint64_t kgd_gfx_v10_hqd_get_pq_addr(struct amdgpu_device *adev,
+				     uint32_t pipe_id, uint32_t queue_id,
+				     uint32_t inst)
+{
+	return 0;
+}
+
+uint64_t kgd_gfx_v10_hqd_reset(struct amdgpu_device *adev,
+			       uint32_t pipe_id, uint32_t queue_id,
+			       uint32_t inst, unsigned int utimeout)
+{
+	return 0;
+}
+
 const struct kfd2kgd_calls gfx_v10_kfd2kgd = {
 	.program_sh_mem_settings = kgd_program_sh_mem_settings,
 	.set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
@@ -1097,4 +1111,6 @@ const struct kfd2kgd_calls gfx_v10_kfd2kgd = {
 	.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,
+	.hqd_get_pq_addr = kgd_gfx_v10_hqd_get_pq_addr,
+	.hqd_reset = kgd_gfx_v10_hqd_reset
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h +9 -0
index 67bcaa3d4226..9efd2dd4fdd7 100644
@@ -56,3 +56,12 @@ void kgd_gfx_v10_build_grace_period_packet_info(struct amdgpu_device *adev,
 					       uint32_t grace_period,
 					       uint32_t *reg_offset,
 					       uint32_t *reg_data);
+uint64_t kgd_gfx_v10_hqd_get_pq_addr(struct amdgpu_device *adev,
+				    uint32_t pipe_id,
+				    uint32_t queue_id,
+				    uint32_t inst);
+uint64_t kgd_gfx_v10_hqd_reset(struct amdgpu_device *adev,
+			      uint32_t pipe_id,
+			      uint32_t queue_id,
+			      uint32_t inst,
+			      unsigned int utimeout);
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c +3 -1
index 8c8437a4383f..c718bedda0ca 100644
@@ -680,5 +680,7 @@ const struct kfd2kgd_calls gfx_v10_3_kfd2kgd = {
 	.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
+	.clear_address_watch = kgd_gfx_v10_clear_address_watch,
+	.hqd_get_pq_addr = kgd_gfx_v10_hqd_get_pq_addr,
+	.hqd_reset = kgd_gfx_v10_hqd_reset
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c +17 -1
index b61a32d6af4b..a4ba49cb22db 100644
@@ -786,6 +786,20 @@ static uint32_t kgd_gfx_v11_clear_address_watch(struct amdgpu_device *adev,
 	return 0;
 }
 
+static uint64_t kgd_gfx_v11_hqd_get_pq_addr(struct amdgpu_device *adev,
+					    uint32_t pipe_id, uint32_t queue_id,
+					    uint32_t inst)
+{
+	return 0;
+}
+
+static uint64_t kgd_gfx_v11_hqd_reset(struct amdgpu_device *adev,
+				      uint32_t pipe_id, uint32_t queue_id,
+				      uint32_t inst, unsigned int utimeout)
+{
+	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,
@@ -808,5 +822,7 @@ const struct kfd2kgd_calls gfx_v11_kfd2kgd = {
 	.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_address_watch = kgd_gfx_v11_set_address_watch,
-	.clear_address_watch = kgd_gfx_v11_clear_address_watch
+	.clear_address_watch = kgd_gfx_v11_clear_address_watch,
+	.hqd_get_pq_addr = kgd_gfx_v11_hqd_get_pq_addr,
+	.hqd_reset = kgd_gfx_v11_hqd_reset
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c +85 -0

95 lines not on this page · raw message

drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h +9 -0
index ce424615f59b..988c50ac3be0 100644
@@ -101,3 +101,12 @@ void kgd_gfx_v9_build_grace_period_packet_info(struct amdgpu_device *adev,
 					       uint32_t grace_period,
 					       uint32_t *reg_offset,
 					       uint32_t *reg_data);
+uint64_t kgd_gfx_v9_hqd_get_pq_addr(struct amdgpu_device *adev,
+				    uint32_t pipe_id,
+				    uint32_t queue_id,
+				    uint32_t inst);
+uint64_t kgd_gfx_v9_hqd_reset(struct amdgpu_device *adev,
+			      uint32_t pipe_id,
+			      uint32_t queue_id,
+			      uint32_t inst,
+			      unsigned int utimeout);
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +164 -8

220 lines not on this page · raw message

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +12 -0
index 3b9b8eabaacc..dfb36a246637 100644
@@ -210,6 +210,13 @@ struct device_queue_manager_asic_ops {
 				 struct kfd_node *dev);
 };
 
+struct dqm_detect_hang_info {
+	int pipe_id;
+	int queue_id;
+	int xcc_id;
+	uint64_t queue_address;
+};
+
 /**
  * struct device_queue_manager
  *
@@ -264,6 +271,11 @@ struct device_queue_manager {
 	uint32_t		wait_times;
 
 	wait_queue_head_t	destroy_wait;
+
+	/* for per-queue reset support */
+	struct dqm_detect_hang_info *detect_hang_info;
+	size_t detect_hang_info_size;
+	int detect_hang_count;
 };
 
 void device_queue_manager_init_cik(
drivers/gpu/drm/amd/amdkfd/kfd_events.c +21 -0
index 9b33d9d2c9ad..3a6e7a4c8895 100644
@@ -31,6 +31,7 @@
 #include <linux/memory.h>
 #include "kfd_priv.h"
 #include "kfd_events.h"
+#include "kfd_device_queue_manager.h"
 #include <linux/device.h>
 
 /*
@@ -1244,12 +1245,32 @@ void kfd_signal_reset_event(struct kfd_node *dev)
 	idx = srcu_read_lock(&kfd_processes_srcu);
 	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
 		int user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
+		struct kfd_process_device *pdd = kfd_get_process_device_data(dev, p);
 
 		if (unlikely(user_gpu_id == -EINVAL)) {
 			WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
 			continue;
 		}
 
+		if (unlikely(!pdd)) {
+			WARN_ONCE(1, "Could not get device data from pasid:0x%x\n", p->pasid);
+			continue;
+		}
+
+		if (dev->dqm->detect_hang_count && !pdd->has_reset_queue)
+			continue;
+
+		if (dev->dqm->detect_hang_count) {
+			struct amdgpu_task_info *ti;
+
+			ti = amdgpu_vm_get_task_info_pasid(dev->adev, p->pasid);
+			if (ti) {
+				DRM_ERROR("Process info: process %s tid %d thread %s pid %d\n",
+				ti->process_name, ti->tgid, ti->task_name, ti->pid);
+				amdgpu_vm_put_task_info(ti);
+			}
+		}
+
 		rcu_read_lock();
 
 		id = KFD_FIRST_NONSIGNAL_EVENT_ID;
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +5 -1
index 66c73825c0a0..84e8ea3a8a0c 100644
@@ -321,8 +321,11 @@ static void update_mqd(struct mqd_manager *mm, void *mqd,
 static bool check_preemption_failed(struct mqd_manager *mm, void *mqd)
 {
 	struct v9_mqd *m = (struct v9_mqd *)mqd;
+	uint32_t doorbell_id = m->queue_doorbell_id0;
 
-	return kfd_check_hiq_mqd_doorbell_id(mm->dev, m->queue_doorbell_id0, 0);
+	m->queue_doorbell_id0 = 0;
+
+	return kfd_check_hiq_mqd_doorbell_id(mm->dev, doorbell_id, 0);
 }
 
 static int get_wave_state(struct mqd_manager *mm, void *mqd,
@@ -624,6 +627,7 @@ static bool check_preemption_failed_v9_4_3(struct mqd_manager *mm, void *mqd)
 		m = get_mqd(mqd + hiq_mqd_size * inst);
 		ret |= kfd_check_hiq_mqd_doorbell_id(mm->dev,
 					m->queue_doorbell_id0, inst);
+		m->queue_doorbell_id0 = 0;
 		++inst;
 	}
 
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +3 -0
index b5cae48dff66..892a85408c09 100644
@@ -843,6 +843,9 @@ struct kfd_process_device {
 	void *proc_ctx_bo;
 	uint64_t proc_ctx_gpu_addr;
 	void *proc_ctx_cpu_ptr;
+
+	/* Tracks queue reset status */
+	bool has_reset_queue;
 };
 
 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
drivers/gpu/drm/amd/amdkfd/kfd_process.c +2 -0
index 9e29b92eb523..a902950cc060 100644
@@ -1851,6 +1851,8 @@ int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger)
 			goto fail;
 		}
 		n_evicted++;
+
+		pdd->dev->dqm->is_hws_hang = false;
 	}
 
 	return r;
drivers/gpu/drm/amd/include/kgd_kfd_interface.h +6 -0
index 6d094cf3587d..7744ca3ef4b1 100644
@@ -318,6 +318,12 @@ struct kfd2kgd_calls {
 	void (*program_trap_handler_settings)(struct amdgpu_device *adev,
 			uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr,
 			uint32_t inst);
+	uint64_t (*hqd_get_pq_addr)(struct amdgpu_device *adev,
+				    uint32_t pipe_id, uint32_t queue_id,
+				    uint32_t inst);
+	uint64_t (*hqd_reset)(struct amdgpu_device *adev,
+			      uint32_t pipe_id, uint32_t queue_id,
+			      uint32_t inst, unsigned int utimeout);
 };
 
 #endif	/* KGD_KFD_INTERFACE_H_INCLUDED */
Signature · 1 line
2.34.1
#1 lore raw

PATCH 2/2 drm/amdkfd: support the debugger during per-queue reset

In order to allow ROCm GDB to handle reset queues, raise an EC_QUEUE_RESET exception so that the debugger can subscribe and query this exception.

Reset queues should still be considered suspendable with a status flag of KFD_DBG_QUEUE_RESET_MASK.
However they should not be resumable since user space will no longer be able to access reset queues.

v2: move per-queue reset flag to this patch
rebase based on patch 1 changes

---
 .../drm/amd/amdkfd/kfd_device_queue_manager.c | 31 ++++++++++++++++---
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |  1 +
 include/uapi/linux/kfd_ioctl.h                |  4 +++
 3 files changed, 31 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +26 -5
index e335703eff84..cb7b5bbf5c40 100644
@@ -164,6 +164,10 @@ static void kfd_hws_hang(struct device_queue_manager *dqm)
 			struct kfd_process_device *pdd = qpd_to_pdd(qpd);
 
 			pdd->has_reset_queue = true;
+			q->properties.is_reset = true;
+			kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_RESET),
+					 q->process, q->device, q->doorbell_id,
+					 false, NULL, 0);
 		}
 	}
 
@@ -986,7 +990,7 @@ static int suspend_single_queue(struct device_queue_manager *dqm,
 {
 	bool is_new;
 
-	if (q->properties.is_suspended)
+	if (q->properties.is_suspended || q->properties.is_reset)
 		return 0;
 
 	pr_debug("Suspending PASID %u queue [%i]\n",
@@ -1007,6 +1011,9 @@ static int suspend_single_queue(struct device_queue_manager *dqm,
 		if (dqm->dev->kfd->shared_resources.enable_mes) {
 			int r = remove_queue_mes(dqm, q, &pdd->qpd);
 
+			if (q->properties.is_reset)
+				return 0;
+
 			if (r)
 				return r;
 		}
@@ -1967,10 +1974,14 @@ static void set_queue_as_reset(struct device_queue_manager *dqm, struct queue *q
 	       q->properties.queue_id, q->process->pasid);
 
 	pdd->has_reset_queue = true;
+	q->properties.is_reset = true;
 	if (q->properties.is_active) {
 		q->properties.is_active = false;
 		decrement_queue_count(dqm, qpd, q);
 	}
+
+	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_RESET), q->process, q->device,
+			 q->doorbell_id, false, NULL, 0);
 }
 
 static int detect_queue_hang(struct device_queue_manager *dqm)
@@ -3037,7 +3048,8 @@ int resume_queues(struct kfd_process *p,
 						queue_ids[q_idx] &=
 							~KFD_DBG_QUEUE_INVALID_MASK;
 					} else {
-						queue_ids[q_idx] |=
+						queue_ids[q_idx] |= q->properties.is_reset ?
+							KFD_DBG_QUEUE_RESET_MASK :
 							KFD_DBG_QUEUE_ERROR_MASK;
 						break;
 					}
@@ -3072,7 +3084,7 @@ int resume_queues(struct kfd_process *p,
 							queue_ids);
 
 					/* mask queue as error on resume fail */
-					if (q_idx != QUEUE_NOT_FOUND)
+					if (q_idx != QUEUE_NOT_FOUND && !q->properties.is_reset)
 						queue_ids[q_idx] |=
 							KFD_DBG_QUEUE_ERROR_MASK;
 				}
@@ -3119,6 +3131,7 @@ int suspend_queues(struct kfd_process *p,
 		struct qcm_process_device *qpd = &pdd->qpd;
 		struct queue *q;
 		int r, per_device_suspended = 0;
+		bool has_queue_reset_fail = false;
 
 		mutex_lock(&p->event_mutex);
 		dqm_lock(dqm);
@@ -3135,6 +3148,9 @@ int suspend_queues(struct kfd_process *p,
 
 				if (!err) {
 					queue_ids[q_idx] &= ~KFD_DBG_QUEUE_INVALID_MASK;
+					if (q->properties.is_reset)
+						queue_ids[q_idx] |= KFD_DBG_QUEUE_RESET_MASK;
+
 					if (exception_clear_mask && is_mes)
 						q->properties.exception_status &=
 							~exception_clear_mask;
@@ -3176,13 +3192,18 @@ int suspend_queues(struct kfd_process *p,
 				continue;
 
 			/* mask queue as error on suspend fail */
-			if (r)
+			if (r && !q->properties.is_reset) {
+				has_queue_reset_fail = true;
 				queue_ids[q_idx] |= KFD_DBG_QUEUE_ERROR_MASK;
-			else if (exception_clear_mask)
+			} else if (exception_clear_mask) {
 				q->properties.exception_status &=
 							~exception_clear_mask;
+			}
 		}
 
+		if (!has_queue_reset_fail)
+			total_suspended += per_device_suspended;
+
 		dqm_unlock(dqm);
 		mutex_unlock(&p->event_mutex);
 		amdgpu_device_flush_hdp(dqm->dev->adev, NULL);
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1 -0
index 892a85408c09..192e3102c152 100644
@@ -504,6 +504,7 @@ struct queue_properties {
 	bool is_being_destroyed;
 	bool is_active;
 	bool is_gws;
+	bool is_reset;
 	uint32_t pm4_target_xcc;
 	bool is_dbg_wa;
 	bool is_user_cu_masked;
include/uapi/linux/kfd_ioctl.h +4 -0
index 285a36601dc9..4713f9a6796e 100644
@@ -875,6 +875,7 @@ enum kfd_dbg_trap_exception_code {
 	EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID = 21,
 	EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID = 22,
 	EC_QUEUE_PACKET_VENDOR_UNSUPPORTED = 23,
+	EC_QUEUE_RESET = 29,
 	EC_QUEUE_PREEMPTION_ERROR = 30,
 	EC_QUEUE_NEW = 31,
 	/* per device */
@@ -907,6 +908,7 @@ enum kfd_dbg_trap_exception_code {
 				 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID) |	\
 				 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) |	\
 				 KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED)	|	\
+				 KFD_EC_MASK(EC_QUEUE_RESET)	|	\
 				 KFD_EC_MASK(EC_QUEUE_PREEMPTION_ERROR)	|	\
 				 KFD_EC_MASK(EC_QUEUE_NEW))
 #define KFD_EC_MASK_DEVICE	(KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE) |		\
@@ -997,8 +999,10 @@ struct kfd_queue_snapshot_entry {
 };
 
 /* Queue status return for suspend/resume */
+#define KFD_DBG_QUEUE_RESET_BIT		29
 #define KFD_DBG_QUEUE_ERROR_BIT		30
 #define KFD_DBG_QUEUE_INVALID_BIT	31
+#define KFD_DBG_QUEUE_RESET_MASK	(1 << KFD_DBG_QUEUE_RESET_BIT)
 #define KFD_DBG_QUEUE_ERROR_MASK	(1 << KFD_DBG_QUEUE_ERROR_BIT)
 #define KFD_DBG_QUEUE_INVALID_MASK	(1 << KFD_DBG_QUEUE_INVALID_BIT)
 
Signature · 1 line
2.34.1
#2 lore raw

Re: [PATCH 2/2] drm/amdkfd: support the debugger during per-queue reset

Quoted 9 lines · Jonathan Kim

In order to allow ROCm GDB to handle reset queues, raise an EC_QUEUE_RESET exception so that the debugger can subscribe and query this exception.

Reset queues should still be considered suspendable with a status flag of KFD_DBG_QUEUE_RESET_MASK.
However they should not be resumable since user space will no longer be able to access reset queues.

Do you have a link to the proposed debugger code which uses this?

Alex

Quoted 164 lines

v2: move per-queue reset flag to this patch
rebase based on patch 1 changes

---
 .../drm/amd/amdkfd/kfd_device_queue_manager.c | 31 ++++++++++++++++---
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |  1 +
 include/uapi/linux/kfd_ioctl.h                |  4 +++
 3 files changed, 31 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +26 -5
index e335703eff84..cb7b5bbf5c40 100644
@@ -164,6 +164,10 @@ static void kfd_hws_hang(struct device_queue_manager *dqm)
                        struct kfd_process_device *pdd = qpd_to_pdd(qpd);
 
                        pdd->has_reset_queue = true;
+                       q->properties.is_reset = true;
+                       kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_RESET),
+                                        q->process, q->device, q->doorbell_id,
+                                        false, NULL, 0);
                }
        }
 
@@ -986,7 +990,7 @@ static int suspend_single_queue(struct device_queue_manager *dqm,
 {
        bool is_new;
 
-       if (q->properties.is_suspended)
+       if (q->properties.is_suspended || q->properties.is_reset)
                return 0;
 
        pr_debug("Suspending PASID %u queue [%i]\n",
@@ -1007,6 +1011,9 @@ static int suspend_single_queue(struct device_queue_manager *dqm,
                if (dqm->dev->kfd->shared_resources.enable_mes) {
                        int r = remove_queue_mes(dqm, q, &pdd->qpd);
 
+                       if (q->properties.is_reset)
+                               return 0;
+
                        if (r)
                                return r;
                }
@@ -1967,10 +1974,14 @@ static void set_queue_as_reset(struct device_queue_manager *dqm, struct queue *q
               q->properties.queue_id, q->process->pasid);
 
        pdd->has_reset_queue = true;
+       q->properties.is_reset = true;
        if (q->properties.is_active) {
                q->properties.is_active = false;
                decrement_queue_count(dqm, qpd, q);
        }
+
+       kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_RESET), q->process, q->device,
+                        q->doorbell_id, false, NULL, 0);
 }
 
 static int detect_queue_hang(struct device_queue_manager *dqm)
@@ -3037,7 +3048,8 @@ int resume_queues(struct kfd_process *p,
                                                queue_ids[q_idx] &=
                                                        ~KFD_DBG_QUEUE_INVALID_MASK;
                                        } else {
-                                               queue_ids[q_idx] |=
+                                               queue_ids[q_idx] |= q->properties.is_reset ?
+                                                       KFD_DBG_QUEUE_RESET_MASK :
                                                        KFD_DBG_QUEUE_ERROR_MASK;
                                                break;
                                        }
@@ -3072,7 +3084,7 @@ int resume_queues(struct kfd_process *p,
                                                        queue_ids);
 
                                        /* mask queue as error on resume fail */
-                                       if (q_idx != QUEUE_NOT_FOUND)
+                                       if (q_idx != QUEUE_NOT_FOUND && !q->properties.is_reset)
                                                queue_ids[q_idx] |=
                                                        KFD_DBG_QUEUE_ERROR_MASK;
                                }
@@ -3119,6 +3131,7 @@ int suspend_queues(struct kfd_process *p,
                struct qcm_process_device *qpd = &pdd->qpd;
                struct queue *q;
                int r, per_device_suspended = 0;
+               bool has_queue_reset_fail = false;
 
                mutex_lock(&p->event_mutex);
                dqm_lock(dqm);
@@ -3135,6 +3148,9 @@ int suspend_queues(struct kfd_process *p,
 
                                if (!err) {
                                        queue_ids[q_idx] &= ~KFD_DBG_QUEUE_INVALID_MASK;
+                                       if (q->properties.is_reset)
+                                               queue_ids[q_idx] |= KFD_DBG_QUEUE_RESET_MASK;
+
                                        if (exception_clear_mask && is_mes)
                                                q->properties.exception_status &=
                                                        ~exception_clear_mask;
@@ -3176,13 +3192,18 @@ int suspend_queues(struct kfd_process *p,
                                continue;
 
                        /* mask queue as error on suspend fail */
-                       if (r)
+                       if (r && !q->properties.is_reset) {
+                               has_queue_reset_fail = true;
                                queue_ids[q_idx] |= KFD_DBG_QUEUE_ERROR_MASK;
-                       else if (exception_clear_mask)
+                       } else if (exception_clear_mask) {
                                q->properties.exception_status &=
                                                        ~exception_clear_mask;
+                       }
                }
 
+               if (!has_queue_reset_fail)
+                       total_suspended += per_device_suspended;
+
                dqm_unlock(dqm);
                mutex_unlock(&p->event_mutex);
                amdgpu_device_flush_hdp(dqm->dev->adev, NULL);
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1 -0
index 892a85408c09..192e3102c152 100644
@@ -504,6 +504,7 @@ struct queue_properties {
        bool is_being_destroyed;
        bool is_active;
        bool is_gws;
+       bool is_reset;
        uint32_t pm4_target_xcc;
        bool is_dbg_wa;
        bool is_user_cu_masked;
include/uapi/linux/kfd_ioctl.h +4 -1
index 285a36601dc9..4713f9a6796e 100644
@@ -875,6 +875,7 @@ enum kfd_dbg_trap_exception_code {
        EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID = 21,
        EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID = 22,
        EC_QUEUE_PACKET_VENDOR_UNSUPPORTED = 23,
+       EC_QUEUE_RESET = 29,
        EC_QUEUE_PREEMPTION_ERROR = 30,
        EC_QUEUE_NEW = 31,
        /* per device */
@@ -907,6 +908,7 @@ enum kfd_dbg_trap_exception_code {
                                 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID) |        \
                                 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) |       \
                                 KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED)        |       \
+                                KFD_EC_MASK(EC_QUEUE_RESET)    |       \
                                 KFD_EC_MASK(EC_QUEUE_PREEMPTION_ERROR) |       \
                                 KFD_EC_MASK(EC_QUEUE_NEW))
 #define KFD_EC_MASK_DEVICE     (KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE) |          \
@@ -997,8 +999,10 @@ struct kfd_queue_snapshot_entry {
 };
 
 /* Queue status return for suspend/resume */
+#define KFD_DBG_QUEUE_RESET_BIT                29
 #define KFD_DBG_QUEUE_ERROR_BIT                30
 #define KFD_DBG_QUEUE_INVALID_BIT      31
+#define KFD_DBG_QUEUE_RESET_MASK       (1 << KFD_DBG_QUEUE_RESET_BIT)
 #define KFD_DBG_QUEUE_ERROR_MASK       (1 << KFD_DBG_QUEUE_ERROR_BIT)
 #define KFD_DBG_QUEUE_INVALID_MASK     (1 << KFD_DBG_QUEUE_INVALID_BIT)
 
--

2.34.1

#3 lore raw

RE: [PATCH 2/2] drm/amdkfd: support the debugger during per-queue reset

[Public]

Quoted 21 lines

-----Original Message-----
From: Alex Deucher <[email protected]>
Sent: Friday, July 26, 2024 2:57 PM
To: Kim, Jonathan <[email protected]>
Cc: [email protected]; Kuehling, Felix <[email protected]>; Deucher, Alexander <[email protected]>
Subject: Re: [PATCH 2/2] drm/amdkfd: support the debugger during per- queue reset

Caution: This message originated from an External Source. Use proper caution when opening attachments, clicking links, or responding.

Quoted 5 lines · Jonathan Kim

In order to allow ROCm GDB to handle reset queues, raise an EC_QUEUE_RESET exception so that the debugger can subscribe and query this exception.

Quoted 6 lines
Quoted 4 lines

Reset queues should still be considered suspendable with a status flag of KFD_DBG_QUEUE_RESET_MASK.
However they should not be resumable since user space will no longer be able to access reset queues.

Do you have a link to the proposed debugger code which uses this?

Not yet. + Laurent/Lancelot for awareness.

Jon

Quoted 186 lines

Alex

Quoted 7 lines

v2: move per-queue reset flag to this patch
rebase based on patch 1 changes

---
 .../drm/amd/amdkfd/kfd_device_queue_manager.c | 31

++++++++++++++++---

Review

Quoted 5 lines
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |  1 +
 include/uapi/linux/kfd_ioctl.h                |  4 +++
 3 files changed, 31 insertions(+), 5 deletions(-)
(quoted diff) +0 -0

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

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static void kfd_hws_hang(struct

Quoted 4 lines

index e335703eff84..cb7b5bbf5c40 100644

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
@@ -164,6 +164,10 @@ static void kfd_hws_hang(struct

device_queue_manager *dqm)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int suspend_single_queue(struct

Quoted 11 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +4 -0
                        struct kfd_process_device *pdd = qpd_to_pdd(qpd);
 
                        pdd->has_reset_queue = true;
+                       q->properties.is_reset = true;
+                       kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_RESET),
+                                        q->process, q->device, q->doorbell_id,
+                                        false, NULL, 0);
                }
        }
 
@@ -986,7 +990,7 @@ static int suspend_single_queue(struct

device_queue_manager *dqm,

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int suspend_single_queue(struct

Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -1
 {
        bool is_new;
 
-       if (q->properties.is_suspended)
+       if (q->properties.is_suspended || q->properties.is_reset)
                return 0;
 
        pr_debug("Suspending PASID %u queue [%i]\n",
@@ -1007,6 +1011,9 @@ static int suspend_single_queue(struct

device_queue_manager *dqm,

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static void set_queue_as_reset(struct

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +3 -0
                if (dqm->dev->kfd->shared_resources.enable_mes) {
                        int r = remove_queue_mes(dqm, q, &pdd->qpd);
 
+                       if (q->properties.is_reset)
+                               return 0;
+
                        if (r)
                                return r;
                }
@@ -1967,10 +1974,14 @@ static void set_queue_as_reset(struct

device_queue_manager *dqm, struct queue *q

Quoted 44 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +4 -0
               q->properties.queue_id, q->process->pasid);
 
        pdd->has_reset_queue = true;
+       q->properties.is_reset = true;
        if (q->properties.is_active) {
                q->properties.is_active = false;
                decrement_queue_count(dqm, qpd, q);
        }
+
+       kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_RESET), q->process, q-
 device,
+                        q->doorbell_id, false, NULL, 0);
 }
 

static int detect_queue_hang(struct device_queue_manager *dqm)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +3 -2
@@ -3037,7 +3048,8 @@ int resume_queues(struct kfd_process *p,
                                                queue_ids[q_idx] &=
                                                        ~KFD_DBG_QUEUE_INVALID_MASK;
                                        } else {
-                                               queue_ids[q_idx] |=
+                                               queue_ids[q_idx] |= q->properties.is_reset ?
+                                                       KFD_DBG_QUEUE_RESET_MASK :
                                                        KFD_DBG_QUEUE_ERROR_MASK;
                                                break;
                                        }
@@ -3072,7 +3084,7 @@ int resume_queues(struct kfd_process *p,
                                                        queue_ids);
 
                                        /* mask queue as error on resume fail */
-                                       if (q_idx != QUEUE_NOT_FOUND)
+                                       if (q_idx != QUEUE_NOT_FOUND && !q-
 properties.is_reset)
                                                queue_ids[q_idx] |=
                                                        KFD_DBG_QUEUE_ERROR_MASK;
                                }
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -0
@@ -3119,6 +3131,7 @@ int suspend_queues(struct kfd_process *p,
                struct qcm_process_device *qpd = &pdd->qpd;
                struct queue *q;
                int r, per_device_suspended = 0;
+               bool has_queue_reset_fail = false;
 
                mutex_lock(&p->event_mutex);
                dqm_lock(dqm);

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c int suspend_queues(struct kfd_process *p, · line 3150

Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
@@ -3135,6 +3148,9 @@ int suspend_queues(struct kfd_process *p,
 
                                if (!err) {
                                        queue_ids[q_idx] &=

~KFD_DBG_QUEUE_INVALID_MASK;

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c int suspend_queues(struct kfd_process *p, · line 3152

Quoted 2 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -0
+                                       if (q->properties.is_reset)
+                                               queue_ids[q_idx] |=

KFD_DBG_QUEUE_RESET_MASK;

Review

Quoted 26 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +8 -2
+
                                        if (exception_clear_mask && is_mes)
                                                q->properties.exception_status &=
                                                        ~exception_clear_mask;
@@ -3176,13 +3192,18 @@ int suspend_queues(struct kfd_process *p,
                                continue;
 
                        /* mask queue as error on suspend fail */
-                       if (r)
+                       if (r && !q->properties.is_reset) {
+                               has_queue_reset_fail = true;
                                queue_ids[q_idx] |= KFD_DBG_QUEUE_ERROR_MASK;
-                       else if (exception_clear_mask)
+                       } else if (exception_clear_mask) {
                                q->properties.exception_status &=
                                                        ~exception_clear_mask;
+                       }
                }
 
+               if (!has_queue_reset_fail)
+                       total_suspended += per_device_suspended;
+
                dqm_unlock(dqm);
                mutex_unlock(&p->event_mutex);
                amdgpu_device_flush_hdp(dqm->dev->adev, NULL);
(quoted diff) +0 -0

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

Quoted 24 lines

index 892a85408c09..192e3102c152 100644

drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1 -0
@@ -504,6 +504,7 @@ struct queue_properties {
        bool is_being_destroyed;
        bool is_active;
        bool is_gws;
+       bool is_reset;
        uint32_t pm4_target_xcc;
        bool is_dbg_wa;
        bool is_user_cu_masked;
include/uapi/linux/kfd_ioctl.h +1 -0
index 285a36601dc9..4713f9a6796e 100644
@@ -875,6 +875,7 @@ enum kfd_dbg_trap_exception_code {
        EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID = 21,
        EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID = 22,
        EC_QUEUE_PACKET_VENDOR_UNSUPPORTED = 23,
+       EC_QUEUE_RESET = 29,
        EC_QUEUE_PREEMPTION_ERROR = 30,
        EC_QUEUE_NEW = 31,
        /* per device */

include/uapi/linux/kfd_ioctl.h enum kfd_dbg_trap_exception_code { · line 908

Quoted 1 line
include/uapi/linux/kfd_ioctl.h +0 -0
@@ -907,6 +908,7 @@ enum kfd_dbg_trap_exception_code {
 

KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID
) | \

Quoted 1 line

KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) | \

Quoted 1 line

KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED) | \

Quoted 4 lines
+                                KFD_EC_MASK(EC_QUEUE_RESET)    |       \
                                 KFD_EC_MASK(EC_QUEUE_PREEMPTION_ERROR) |       \
                                 KFD_EC_MASK(EC_QUEUE_NEW))

#define KFD_EC_MASK_DEVICE

(KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE) | \

include/uapi/linux/kfd_ioctl.h struct kfd_queue_snapshot_entry { · line 1005

Quoted 8 lines
include/uapi/linux/kfd_ioctl.h +2 -0
@@ -997,8 +999,10 @@ struct kfd_queue_snapshot_entry {
 };
 
 /* Queue status return for suspend/resume */
+#define KFD_DBG_QUEUE_RESET_BIT                29
 #define KFD_DBG_QUEUE_ERROR_BIT                30
 #define KFD_DBG_QUEUE_INVALID_BIT      31
+#define KFD_DBG_QUEUE_RESET_MASK       (1 <<

KFD_DBG_QUEUE_RESET_BIT)

include/uapi/linux/kfd_ioctl.h struct kfd_queue_snapshot_entry { · line 1006

Quoted 1 line
include/uapi/linux/kfd_ioctl.h +0 -0
 #define KFD_DBG_QUEUE_ERROR_MASK       (1 <<

KFD_DBG_QUEUE_ERROR_BIT)

include/uapi/linux/kfd_ioctl.h struct kfd_queue_snapshot_entry { · line 1007

Quoted 1 line
include/uapi/linux/kfd_ioctl.h +0 -0
 #define KFD_DBG_QUEUE_INVALID_MASK     (1 <<

KFD_DBG_QUEUE_INVALID_BIT)

Quoted 4 lines
include/uapi/linux/kfd_ioctl.h +0 -1
--

2.34.1

maintainer
#4 lore raw

Re: [PATCH 2/2] drm/amdkfd: support the debugger during per-queue reset

In reply to Jonathan Kim
Quoted 143 lines · Jonathan Kim

In order to allow ROCm GDB to handle reset queues, raise an EC_QUEUE_RESET exception so that the debugger can subscribe and query this exception.

Reset queues should still be considered suspendable with a status flag of KFD_DBG_QUEUE_RESET_MASK.
However they should not be resumable since user space will no longer be able to access reset queues.

v2: move per-queue reset flag to this patch
rebase based on patch 1 changes

---
 .../drm/amd/amdkfd/kfd_device_queue_manager.c | 31 ++++++++++++++++---
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |  1 +
 include/uapi/linux/kfd_ioctl.h                |  4 +++
 3 files changed, 31 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +26 -5
index e335703eff84..cb7b5bbf5c40 100644
@@ -164,6 +164,10 @@ static void kfd_hws_hang(struct device_queue_manager *dqm)
 			struct kfd_process_device *pdd = qpd_to_pdd(qpd);
 
 			pdd->has_reset_queue = true;
+			q->properties.is_reset = true;
+			kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_RESET),
+					 q->process, q->device, q->doorbell_id,
+					 false, NULL, 0);
 		}
 	}
 
@@ -986,7 +990,7 @@ static int suspend_single_queue(struct device_queue_manager *dqm,
 {
 	bool is_new;
 
-	if (q->properties.is_suspended)
+	if (q->properties.is_suspended || q->properties.is_reset)
 		return 0;
 
 	pr_debug("Suspending PASID %u queue [%i]\n",
@@ -1007,6 +1011,9 @@ static int suspend_single_queue(struct device_queue_manager *dqm,
 		if (dqm->dev->kfd->shared_resources.enable_mes) {
 			int r = remove_queue_mes(dqm, q, &pdd->qpd);
 
+			if (q->properties.is_reset)
+				return 0;
+
 			if (r)
 				return r;
 		}
@@ -1967,10 +1974,14 @@ static void set_queue_as_reset(struct device_queue_manager *dqm, struct queue *q
 	       q->properties.queue_id, q->process->pasid);
 
 	pdd->has_reset_queue = true;
+	q->properties.is_reset = true;
 	if (q->properties.is_active) {
 		q->properties.is_active = false;
 		decrement_queue_count(dqm, qpd, q);
 	}
+
+	kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_RESET), q->process, q->device,
+			 q->doorbell_id, false, NULL, 0);
 }
 
 static int detect_queue_hang(struct device_queue_manager *dqm)
@@ -3037,7 +3048,8 @@ int resume_queues(struct kfd_process *p,
 						queue_ids[q_idx] &=
 							~KFD_DBG_QUEUE_INVALID_MASK;
 					} else {
-						queue_ids[q_idx] |=
+						queue_ids[q_idx] |= q->properties.is_reset ?
+							KFD_DBG_QUEUE_RESET_MASK :
 							KFD_DBG_QUEUE_ERROR_MASK;
 						break;
 					}
@@ -3072,7 +3084,7 @@ int resume_queues(struct kfd_process *p,
 							queue_ids);
 
 					/* mask queue as error on resume fail */
-					if (q_idx != QUEUE_NOT_FOUND)
+					if (q_idx != QUEUE_NOT_FOUND && !q->properties.is_reset)
 						queue_ids[q_idx] |=
 							KFD_DBG_QUEUE_ERROR_MASK;
 				}
@@ -3119,6 +3131,7 @@ int suspend_queues(struct kfd_process *p,
 		struct qcm_process_device *qpd = &pdd->qpd;
 		struct queue *q;
 		int r, per_device_suspended = 0;
+		bool has_queue_reset_fail = false;
 
 		mutex_lock(&p->event_mutex);
 		dqm_lock(dqm);
@@ -3135,6 +3148,9 @@ int suspend_queues(struct kfd_process *p,
 
 				if (!err) {
 					queue_ids[q_idx] &= ~KFD_DBG_QUEUE_INVALID_MASK;
+					if (q->properties.is_reset)
+						queue_ids[q_idx] |= KFD_DBG_QUEUE_RESET_MASK;
+
 					if (exception_clear_mask && is_mes)
 						q->properties.exception_status &=
 							~exception_clear_mask;
@@ -3176,13 +3192,18 @@ int suspend_queues(struct kfd_process *p,
 				continue;
 
 			/* mask queue as error on suspend fail */
-			if (r)
+			if (r && !q->properties.is_reset) {
+				has_queue_reset_fail = true;
 				queue_ids[q_idx] |= KFD_DBG_QUEUE_ERROR_MASK;
-			else if (exception_clear_mask)
+			} else if (exception_clear_mask) {
 				q->properties.exception_status &=
 							~exception_clear_mask;
+			}
 		}
 
+		if (!has_queue_reset_fail)
+			total_suspended += per_device_suspended;
+
 		dqm_unlock(dqm);
 		mutex_unlock(&p->event_mutex);
 		amdgpu_device_flush_hdp(dqm->dev->adev, NULL);
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1 -0
index 892a85408c09..192e3102c152 100644
@@ -504,6 +504,7 @@ struct queue_properties {
 	bool is_being_destroyed;
 	bool is_active;
 	bool is_gws;
+	bool is_reset;
 	uint32_t pm4_target_xcc;
 	bool is_dbg_wa;
 	bool is_user_cu_masked;
include/uapi/linux/kfd_ioctl.h +0 -0
index 285a36601dc9..4713f9a6796e 100644

include/uapi/linux/kfd_ioctl.h enum kfd_dbg_trap_exception_code { · line 879

Quoted 5 lines
include/uapi/linux/kfd_ioctl.h +1 -0
@@ -875,6 +875,7 @@ enum kfd_dbg_trap_exception_code {
 	EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID = 21,
 	EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID = 22,
 	EC_QUEUE_PACKET_VENDOR_UNSUPPORTED = 23,
+	EC_QUEUE_RESET = 29,
 	EC_QUEUE_PREEMPTION_ERROR = 30,

Do we really need to distinguish between preemption errors and resets? I mean, both are caused by preemption errors. The only difference is, that in the "reset" case we succeeded in resetting only the affected queues.

Quoted 11 lines
include/uapi/linux/kfd_ioctl.h +1 -0
 	EC_QUEUE_NEW = 31,
 	/* per device */
@@ -907,6 +908,7 @@ enum kfd_dbg_trap_exception_code {
 				 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID) |	\
 				 KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) |	\
 				 KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED)	|	\
+				 KFD_EC_MASK(EC_QUEUE_RESET)	|	\
 				 KFD_EC_MASK(EC_QUEUE_PREEMPTION_ERROR)	|	\
 				 KFD_EC_MASK(EC_QUEUE_NEW))
 #define KFD_EC_MASK_DEVICE	(KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE) |		\

include/uapi/linux/kfd_ioctl.h struct kfd_queue_snapshot_entry { · line 1005

Quoted 7 lines
include/uapi/linux/kfd_ioctl.h +2 -0
@@ -997,8 +999,10 @@ struct kfd_queue_snapshot_entry {
 };
 
 /* Queue status return for suspend/resume */
+#define KFD_DBG_QUEUE_RESET_BIT		29
 #define KFD_DBG_QUEUE_ERROR_BIT		30
 #define KFD_DBG_QUEUE_INVALID_BIT	31
+#define KFD_DBG_QUEUE_RESET_MASK	(1 << KFD_DBG_QUEUE_RESET_BIT)

Same as above. If the distinction is not necessary, we don't need to change the user mode API. Just report queue resets as generic preemption errors.

Regards,
Felix

Quoted 3 lines
include/uapi/linux/kfd_ioctl.h +0 -0
 #define KFD_DBG_QUEUE_ERROR_MASK	(1 << KFD_DBG_QUEUE_ERROR_BIT)
 #define KFD_DBG_QUEUE_INVALID_MASK	(1 << KFD_DBG_QUEUE_INVALID_BIT)
 
#5 lore raw

RE: [PATCH 2/2] drm/amdkfd: support the debugger during per-queue reset

[Public]

Quoted 213 lines

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Tuesday, July 30, 2024 6:17 PM
To: Kim, Jonathan <[email protected]>; [email protected]
Cc: Deucher, Alexander <[email protected]>
Subject: Re: [PATCH 2/2] drm/amdkfd: support the debugger during per- queue reset

Quoted 9 lines · Jonathan Kim

In order to allow ROCm GDB to handle reset queues, raise an EC_QUEUE_RESET exception so that the debugger can subscribe and query this exception.

Reset queues should still be considered suspendable with a status flag of KFD_DBG_QUEUE_RESET_MASK.
However they should not be resumable since user space will no longer be able to access reset queues.

Quoted 6 lines

v2: move per-queue reset flag to this patch
rebase based on patch 1 changes

---
 .../drm/amd/amdkfd/kfd_device_queue_manager.c | 31

++++++++++++++++---

Review

Quoted 5 lines
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |  1 +
 include/uapi/linux/kfd_ioctl.h                |  4 +++
 3 files changed, 31 insertions(+), 5 deletions(-)
(quoted diff) +0 -0

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

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static void kfd_hws_hang(struct

Quoted 4 lines

index e335703eff84..cb7b5bbf5c40 100644

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
@@ -164,6 +164,10 @@ static void kfd_hws_hang(struct

device_queue_manager *dqm)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int suspend_single_queue(struct

Quoted 12 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +4 -0
                    struct kfd_process_device *pdd = qpd_to_pdd(qpd);
 
                    pdd->has_reset_queue = true;
+                   q->properties.is_reset = true;
+                   kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_RESET),
+                                    q->process, q->device, q-
 doorbell_id,
+                                    false, NULL, 0);
            }
    }
@@ -986,7 +990,7 @@ static int suspend_single_queue(struct

device_queue_manager *dqm,

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int suspend_single_queue(struct

Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -1
 {
    bool is_new;
 
-   if (q->properties.is_suspended)
+   if (q->properties.is_suspended || q->properties.is_reset)
            return 0;
 
    pr_debug("Suspending PASID %u queue [%i]\n",
@@ -1007,6 +1011,9 @@ static int suspend_single_queue(struct

device_queue_manager *dqm,

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static void set_queue_as_reset(struct

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +3 -0
            if (dqm->dev->kfd->shared_resources.enable_mes) {
                    int r = remove_queue_mes(dqm, q, &pdd->qpd);
 
+                   if (q->properties.is_reset)
+                           return 0;
+
                    if (r)
                            return r;
            }
@@ -1967,10 +1974,14 @@ static void set_queue_as_reset(struct

device_queue_manager *dqm, struct queue *q

Quoted 16 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +4 -0
           q->properties.queue_id, q->process->pasid);
 
    pdd->has_reset_queue = true;
+   q->properties.is_reset = true;
    if (q->properties.is_active) {
            q->properties.is_active = false;
            decrement_queue_count(dqm, qpd, q);
    }
+
+   kfd_dbg_ev_raise(KFD_EC_MASK(EC_QUEUE_RESET), q->process, q-
 device,
+                    q->doorbell_id, false, NULL, 0);
 }
 

static int detect_queue_hang(struct device_queue_manager *dqm)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c int resume_queues(struct kfd_process *p, · line 3049

Quoted 2 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
@@ -3037,7 +3048,8 @@ int resume_queues(struct kfd_process *p,
                                            queue_ids[q_idx] &=
 
      ~KFD_DBG_QUEUE_INVALID_MASK;
Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -1
                                    } else {
-                                           queue_ids[q_idx] |=
+                                           queue_ids[q_idx] |= q-

properties.is_reset ?
+

KFD_DBG_QUEUE_RESET_MASK :

Quoted 1 line
      KFD_DBG_QUEUE_ERROR_MASK;

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c int resume_queues(struct kfd_process *p, · line 3086

Quoted 6 lines
                                            break;
                                    }
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
@@ -3072,7 +3084,7 @@ int resume_queues(struct kfd_process *p,
                                                    queue_ids);
 
                                    /* mask queue as error on resume fail

*/

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c int resume_queues(struct kfd_process *p, · line 3087

Quoted 2 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -1
-                                   if (q_idx != QUEUE_NOT_FOUND)
+                                   if (q_idx != QUEUE_NOT_FOUND

&& !q->properties.is_reset)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c int resume_queues(struct kfd_process *p, · line 3089

Quoted 2 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
                                            queue_ids[q_idx] |=
 
      KFD_DBG_QUEUE_ERROR_MASK;

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c int suspend_queues(struct kfd_process *p, · line 3150

Quoted 13 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -0
                            }
@@ -3119,6 +3131,7 @@ int suspend_queues(struct kfd_process *p,
            struct qcm_process_device *qpd = &pdd->qpd;
            struct queue *q;
            int r, per_device_suspended = 0;
+           bool has_queue_reset_fail = false;
 
            mutex_lock(&p->event_mutex);
            dqm_lock(dqm);
@@ -3135,6 +3148,9 @@ int suspend_queues(struct kfd_process *p,
 
                            if (!err) {
                                    queue_ids[q_idx] &=

~KFD_DBG_QUEUE_INVALID_MASK;

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c int suspend_queues(struct kfd_process *p, · line 3152

Quoted 2 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -0
+                                   if (q->properties.is_reset)
+                                           queue_ids[q_idx] |=

KFD_DBG_QUEUE_RESET_MASK;

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -0
+
                                    if (exception_clear_mask && is_mes)
                                            q-

properties.exception_status &=

      ~exception_clear_mask;

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c int suspend_queues(struct kfd_process *p, · line 3197

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -1
@@ -3176,13 +3192,18 @@ int suspend_queues(struct kfd_process *p,
                            continue;
 
                    /* mask queue as error on suspend fail */
-                   if (r)
+                   if (r && !q->properties.is_reset) {
+                           has_queue_reset_fail = true;
                            queue_ids[q_idx] |=

KFD_DBG_QUEUE_ERROR_MASK;

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c int suspend_queues(struct kfd_process *p, · line 3203

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -1
-                   else if (exception_clear_mask)
+                   } else if (exception_clear_mask) {
                            q->properties.exception_status &=
 
      ~exception_clear_mask;

Review

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +4 -0
+                   }
            }
 
+           if (!has_queue_reset_fail)
+                   total_suspended += per_device_suspended;
+
            dqm_unlock(dqm);
            mutex_unlock(&p->event_mutex);
            amdgpu_device_flush_hdp(dqm->dev->adev, NULL);
(quoted diff) +0 -0

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

Quoted 16 lines

index 892a85408c09..192e3102c152 100644

drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1 -0
@@ -504,6 +504,7 @@ struct queue_properties {
    bool is_being_destroyed;
    bool is_active;
    bool is_gws;
+   bool is_reset;
    uint32_t pm4_target_xcc;
    bool is_dbg_wa;
    bool is_user_cu_masked;
include/uapi/linux/kfd_ioctl.h +0 -0
index 285a36601dc9..4713f9a6796e 100644

include/uapi/linux/kfd_ioctl.h enum kfd_dbg_trap_exception_code { · line 879

Quoted 5 lines
include/uapi/linux/kfd_ioctl.h +1 -0
@@ -875,6 +875,7 @@ enum kfd_dbg_trap_exception_code {
    EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID = 21,
    EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID = 22,
    EC_QUEUE_PACKET_VENDOR_UNSUPPORTED = 23,
+   EC_QUEUE_RESET = 29,
    EC_QUEUE_PREEMPTION_ERROR = 30,

Do we really need to distinguish between preemption errors and resets? I mean, both are caused by preemption errors. The only difference is, that in the "reset" case we succeeded in resetting only the affected queues.

include/uapi/linux/kfd_ioctl.h enum kfd_dbg_trap_exception_code { · line 908

Quoted 4 lines
include/uapi/linux/kfd_ioctl.h +0 -0
    EC_QUEUE_NEW = 31,
    /* per device */
@@ -907,6 +908,7 @@ enum kfd_dbg_trap_exception_code {
 

KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_WORK_GROUP_SIZE_INVALID
) | \

Quoted 1 line

KFD_EC_MASK(EC_QUEUE_PACKET_DISPATCH_REGISTER_INVALID) | \

Quoted 1 line

KFD_EC_MASK(EC_QUEUE_PACKET_VENDOR_UNSUPPORTED) | \

Quoted 1 line

+ KFD_EC_MASK(EC_QUEUE_RESET) |

      \
Quoted 1 line

KFD_EC_MASK(EC_QUEUE_PREEMPTION_ERROR) | \

Quoted 2 lines
                             KFD_EC_MASK(EC_QUEUE_NEW))

#define KFD_EC_MASK_DEVICE

      (KFD_EC_MASK(EC_DEVICE_QUEUE_DELETE) |          \

include/uapi/linux/kfd_ioctl.h struct kfd_queue_snapshot_entry { · line 1005

Quoted 8 lines
include/uapi/linux/kfd_ioctl.h +2 -0
@@ -997,8 +999,10 @@ struct kfd_queue_snapshot_entry {
 };
 
 /* Queue status return for suspend/resume */
+#define KFD_DBG_QUEUE_RESET_BIT            29
 #define KFD_DBG_QUEUE_ERROR_BIT            30
 #define KFD_DBG_QUEUE_INVALID_BIT  31
+#define KFD_DBG_QUEUE_RESET_MASK   (1 <<

KFD_DBG_QUEUE_RESET_BIT)

Quoted 2 lines

Same as above. If the distinction is not necessary, we don't need to change the user mode API. Just report queue resets as generic preemption errors.

Ok. Let me think more about this. Right now preemption errors don't get counted as suspended queues so we'd end up changing the behaviour of the existing API anyways because the request is for reset queues to be counted as suspended so that the debugger knows to inspect them. That being said, maybe there's a need to distinguish generic errors when a gpu reset fails versus when things actual reset whether per-adapter or per-queue.

Thanks,

Jon

Quoted 9 lines

Regards,
Felix

include/uapi/linux/kfd_ioctl.h struct kfd_queue_snapshot_entry { · line 1006

Quoted 1 line
include/uapi/linux/kfd_ioctl.h +0 -0
 #define KFD_DBG_QUEUE_ERROR_MASK   (1 <<

KFD_DBG_QUEUE_ERROR_BIT)

include/uapi/linux/kfd_ioctl.h struct kfd_queue_snapshot_entry { · line 1007

Quoted 1 line
include/uapi/linux/kfd_ioctl.h +0 -0
 #define KFD_DBG_QUEUE_INVALID_MASK (1 <<

KFD_DBG_QUEUE_INVALID_BIT)

Quoted 1 line
Acked-by on v1 1/2
#6 lore raw
In reply to Jonathan Kim
Quoted 580 lines · Jonathan Kim

Support per-queue reset for GFX9. The recommendation is for the driver to target reset the HW queue via a SPI MMIO register write.

Since this requires pipe and HW queue info and MEC FW is limited to doorbell reports of hung queues after an unmap failure, scan the HW queue slots defined by SET_RESOURCES first to identify the user queue candidates to reset.

Only signal reset events to processes that have had a queue reset.

If queue reset fails, fall back to GPU reset.

v2: move reset queue flag for house keeping to process device. split detect and reset into separate functions.
make reset call safe during power saving modes. clean up some other nitpicks.

---
 .../drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c  |   2 +
 .../drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c   |   4 +-
 .../drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c   |   4 +-
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c    |  16 ++
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h    |   9 +
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c  |   4 +-
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c    |  18 +-
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c |  85 +++++++++
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h |   9 +
 .../drm/amd/amdkfd/kfd_device_queue_manager.c | 172 +++++++++++++++++-
 .../drm/amd/amdkfd/kfd_device_queue_manager.h |  12 ++
 drivers/gpu/drm/amd/amdkfd/kfd_events.c       |  21 +++
 .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c   |   6 +-
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |   3 +
 drivers/gpu/drm/amd/amdkfd/kfd_process.c      |   2 +
 .../gpu/drm/amd/include/kgd_kfd_interface.h   |   6 +
 16 files changed, 360 insertions(+), 13 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c +2 -0
index aff08321e976..8dfdb18197c4 100644
@@ -191,4 +191,6 @@ const struct kfd2kgd_calls aldebaran_kfd2kgd = {
        .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,
+       .hqd_get_pq_addr = kgd_gfx_v9_hqd_get_pq_addr,
+       .hqd_reset = kgd_gfx_v9_hqd_reset,
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c +3 -1
index 3a3f3ce09f00..017e8a3013aa 100644
@@ -418,5 +418,7 @@ const struct kfd2kgd_calls arcturus_kfd2kgd = {
        .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,
-       .program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings
+       .program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings,
+       .hqd_get_pq_addr = kgd_gfx_v9_hqd_get_pq_addr,
+       .hqd_reset = kgd_gfx_v9_hqd_reset
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c +3 -1
index a5c7259cf2a3..e2ae714a700f 100644
@@ -541,5 +541,7 @@ const struct kfd2kgd_calls gc_9_4_3_kfd2kgd = {
                        kgd_gfx_v9_4_3_set_wave_launch_trap_override,
        .set_wave_launch_mode = kgd_aldebaran_set_wave_launch_mode,
        .set_address_watch = kgd_gfx_v9_4_3_set_address_watch,
-       .clear_address_watch = kgd_gfx_v9_4_3_clear_address_watch
+       .clear_address_watch = kgd_gfx_v9_4_3_clear_address_watch,
+       .hqd_get_pq_addr = kgd_gfx_v9_hqd_get_pq_addr,
+       .hqd_reset = kgd_gfx_v9_hqd_reset
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c +16 -0
index 3ab6c3aa0ad1..62176d607bef 100644
@@ -1070,6 +1070,20 @@ static void program_trap_handler_settings(struct amdgpu_device *adev,
        unlock_srbm(adev);
 }
 
+uint64_t kgd_gfx_v10_hqd_get_pq_addr(struct amdgpu_device *adev,
+                                    uint32_t pipe_id, uint32_t queue_id,
+                                    uint32_t inst)
+{
+       return 0;
+}
+
+uint64_t kgd_gfx_v10_hqd_reset(struct amdgpu_device *adev,
+                              uint32_t pipe_id, uint32_t queue_id,
+                              uint32_t inst, unsigned int utimeout)
+{
+       return 0;
+}
+
 const struct kfd2kgd_calls gfx_v10_kfd2kgd = {
        .program_sh_mem_settings = kgd_program_sh_mem_settings,
        .set_pasid_vmid_mapping = kgd_set_pasid_vmid_mapping,
@@ -1097,4 +1111,6 @@ const struct kfd2kgd_calls gfx_v10_kfd2kgd = {
        .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,
+       .hqd_get_pq_addr = kgd_gfx_v10_hqd_get_pq_addr,
+       .hqd_reset = kgd_gfx_v10_hqd_reset
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h +9 -0
index 67bcaa3d4226..9efd2dd4fdd7 100644
@@ -56,3 +56,12 @@ void kgd_gfx_v10_build_grace_period_packet_info(struct amdgpu_device *adev,
                                               uint32_t grace_period,
                                               uint32_t *reg_offset,
                                               uint32_t *reg_data);
+uint64_t kgd_gfx_v10_hqd_get_pq_addr(struct amdgpu_device *adev,
+                                   uint32_t pipe_id,
+                                   uint32_t queue_id,
+                                   uint32_t inst);
+uint64_t kgd_gfx_v10_hqd_reset(struct amdgpu_device *adev,
+                             uint32_t pipe_id,
+                             uint32_t queue_id,
+                             uint32_t inst,
+                             unsigned int utimeout);
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c +3 -1
index 8c8437a4383f..c718bedda0ca 100644
@@ -680,5 +680,7 @@ const struct kfd2kgd_calls gfx_v10_3_kfd2kgd = {
        .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
+       .clear_address_watch = kgd_gfx_v10_clear_address_watch,
+       .hqd_get_pq_addr = kgd_gfx_v10_hqd_get_pq_addr,
+       .hqd_reset = kgd_gfx_v10_hqd_reset
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c +17 -1
index b61a32d6af4b..a4ba49cb22db 100644
@@ -786,6 +786,20 @@ static uint32_t kgd_gfx_v11_clear_address_watch(struct amdgpu_device *adev,
        return 0;
 }
 
+static uint64_t kgd_gfx_v11_hqd_get_pq_addr(struct amdgpu_device *adev,
+                                           uint32_t pipe_id, uint32_t queue_id,
+                                           uint32_t inst)
+{
+       return 0;
+}
+
+static uint64_t kgd_gfx_v11_hqd_reset(struct amdgpu_device *adev,
+                                     uint32_t pipe_id, uint32_t queue_id,
+                                     uint32_t inst, unsigned int utimeout)
+{
+       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,
@@ -808,5 +822,7 @@ const struct kfd2kgd_calls gfx_v11_kfd2kgd = {
        .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_address_watch = kgd_gfx_v11_set_address_watch,
-       .clear_address_watch = kgd_gfx_v11_clear_address_watch
+       .clear_address_watch = kgd_gfx_v11_clear_address_watch,
+       .hqd_get_pq_addr = kgd_gfx_v11_hqd_get_pq_addr,
+       .hqd_reset = kgd_gfx_v11_hqd_reset
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c +85 -0
index 5a35a8ca8922..32f28c12077b 100644
@@ -1144,6 +1144,89 @@ void kgd_gfx_v9_program_trap_handler_settings(struct amdgpu_device *adev,
        kgd_gfx_v9_unlock_srbm(adev, inst);
 }
 
+uint64_t kgd_gfx_v9_hqd_get_pq_addr(struct amdgpu_device *adev,
+                                   uint32_t pipe_id, uint32_t queue_id,
+                                   uint32_t inst)
+{
+       uint32_t low, high;
+       uint64_t queue_addr = 0;
+
+       kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst);
+       amdgpu_gfx_rlc_enter_safe_mode(adev, inst);
+
+       if (!RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_ACTIVE))
+               goto unlock_out;
+
+       low = RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_PQ_BASE);
+       high = RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_PQ_BASE_HI);
+
+       /* only concerned with user queues. */
+       if (!high)
+               goto unlock_out;
+
+       queue_addr = (((queue_addr | high) << 32) | low) << 8;
+
+unlock_out:
+       amdgpu_gfx_rlc_exit_safe_mode(adev, inst);
+       kgd_gfx_v9_release_queue(adev, inst);
+
+       return queue_addr;
+}
+
+uint64_t kgd_gfx_v9_hqd_reset(struct amdgpu_device *adev,
+                             uint32_t pipe_id, uint32_t queue_id,
+                             uint32_t inst, unsigned int utimeout)
+{
+       uint32_t low, high, temp;
+       unsigned long end_jiffies;
+       uint64_t queue_addr = 0;
+
+       kgd_gfx_v9_acquire_queue(adev, pipe_id, queue_id, inst);
+       amdgpu_gfx_rlc_enter_safe_mode(adev, inst);
+
+       if (!RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_ACTIVE))
+               goto unlock_out;
+
+       low = RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_PQ_BASE);
+       high = RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_PQ_BASE_HI);
+
+       /* only concerned with user queues. */
+       if (!high)
+               goto unlock_out;
+
+       queue_addr = (((queue_addr | high) << 32) | low) << 8;
+
+       pr_debug("Attempting queue reset on XCC %i pipe id %i queue id %i\n",
+                inst, pipe_id, queue_id);
+
+       /* assume previous dequeue request issued will take affect after reset */
+       WREG32_SOC15(GC, GET_INST(GC, inst), mmSPI_COMPUTE_QUEUE_RESET, 0x1);
+
+       end_jiffies = (utimeout * HZ / 1000) + jiffies;
+       while (true) {
+               temp = RREG32_SOC15(GC, GET_INST(GC, inst), mmCP_HQD_ACTIVE);
+
+               if (!(temp & CP_HQD_ACTIVE__ACTIVE_MASK))
+                       break;
+
+               if (time_after(jiffies, end_jiffies)) {
+                       queue_addr = 0;
+                       break;
+               }
+
+               usleep_range(500, 1000);
+       }
+
+       pr_debug("queue reset on XCC %i pipe id %i queue id %i %s\n",
+                inst, pipe_id, queue_id, !!queue_addr ? "succeeded!" : "failed!");
+
+unlock_out:
+       amdgpu_gfx_rlc_exit_safe_mode(adev, inst);
+       kgd_gfx_v9_release_queue(adev, inst);
+
+       return queue_addr;
+}
+
 const struct kfd2kgd_calls gfx_v9_kfd2kgd = {
        .program_sh_mem_settings = kgd_gfx_v9_program_sh_mem_settings,
        .set_pasid_vmid_mapping = kgd_gfx_v9_set_pasid_vmid_mapping,
@@ -1172,4 +1255,6 @@ const struct kfd2kgd_calls gfx_v9_kfd2kgd = {
        .build_grace_period_packet_info = kgd_gfx_v9_build_grace_period_packet_info,
        .get_cu_occupancy = kgd_gfx_v9_get_cu_occupancy,
        .program_trap_handler_settings = kgd_gfx_v9_program_trap_handler_settings,
+       .hqd_get_pq_addr = kgd_gfx_v9_hqd_get_pq_addr,
+       .hqd_reset = kgd_gfx_v9_hqd_reset
 };
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h +9 -0
index ce424615f59b..988c50ac3be0 100644
@@ -101,3 +101,12 @@ void kgd_gfx_v9_build_grace_period_packet_info(struct amdgpu_device *adev,
                                               uint32_t grace_period,
                                               uint32_t *reg_offset,
                                               uint32_t *reg_data);
+uint64_t kgd_gfx_v9_hqd_get_pq_addr(struct amdgpu_device *adev,
+                                   uint32_t pipe_id,
+                                   uint32_t queue_id,
+                                   uint32_t inst);
+uint64_t kgd_gfx_v9_hqd_reset(struct amdgpu_device *adev,
+                             uint32_t pipe_id,
+                             uint32_t queue_id,
+                             uint32_t inst,
+                             unsigned int utimeout);
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +164 -8
index fdc76c24b2e7..e335703eff84 100644
@@ -153,6 +153,20 @@ void program_sh_mem_settings(struct device_queue_manager *dqm,
 
 static void kfd_hws_hang(struct device_queue_manager *dqm)
 {
+       struct device_process_node *cur;
+       struct qcm_process_device *qpd;
+       struct queue *q;
+
+       /* Mark all device queues as reset. */
+       list_for_each_entry(cur, &dqm->queues, list) {
+               qpd = cur->qpd;
+               list_for_each_entry(q, &qpd->queues_list, list) {
+                       struct kfd_process_device *pdd = qpd_to_pdd(qpd);
+
+                       pdd->has_reset_queue = true;
+               }
+       }
+
        /*
         * Issue a GPU reset if HWS is unresponsive
         */
@@ -878,6 +892,12 @@ static int update_queue(struct device_queue_manager *dqm, struct queue *q,
                else if (prev_active)
                        retval = remove_queue_mes(dqm, q, &pdd->qpd);
 
+               /* queue is reset so inaccessable  */
+               if (pdd->has_reset_queue) {
+                       retval = -EACCES;
+                       goto out_unlock;
+               }
+
                if (retval) {
                        dev_err(dev, "unmap queue failed\n");
                        goto out_unlock;
@@ -1627,7 +1647,7 @@ static int initialize_cpsch(struct device_queue_manager *dqm)
 static int start_cpsch(struct device_queue_manager *dqm)
 {
        struct device *dev = dqm->dev->adev->dev;
-       int retval;
+       int retval, num_hw_queue_slots;
 
        retval = 0;
 
@@ -1680,6 +1700,14 @@ static int start_cpsch(struct device_queue_manager *dqm)
                                        &dqm->wait_times);
        }
 
+       /* setup per-queue reset detection buffer  */
+       num_hw_queue_slots =  dqm->dev->kfd->shared_resources.num_queue_per_pipe *
+                             dqm->dev->kfd->shared_resources.num_pipe_per_mec *
+                             NUM_XCC(dqm->dev->xcc_mask);
+
+       dqm->detect_hang_info_size = num_hw_queue_slots * sizeof(struct dqm_detect_hang_info);
+       dqm->detect_hang_info = kzalloc(dqm->detect_hang_info_size, GFP_KERNEL);
+
        dqm_unlock(dqm);
 
        return 0;
@@ -1713,6 +1741,7 @@ static int stop_cpsch(struct device_queue_manager *dqm)
        kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
        if (!dqm->dev->kfd->shared_resources.enable_mes)
                pm_uninit(&dqm->packet_mgr);
+       kfree(dqm->detect_hang_info);
        dqm_unlock(dqm);
 
        return 0;
@@ -1929,6 +1958,131 @@ static int map_queues_cpsch(struct device_queue_manager *dqm)
        return retval;
 }
 
+static void set_queue_as_reset(struct device_queue_manager *dqm, struct queue *q,
+                              struct qcm_process_device *qpd)
+{
+       struct kfd_process_device *pdd = qpd_to_pdd(qpd);
+
+       pr_err("queue id 0x%0x at pasid 0x%0x is reset\n",
+              q->properties.queue_id, q->process->pasid);
+
+       pdd->has_reset_queue = true;
+       if (q->properties.is_active) {
+               q->properties.is_active = false;
+               decrement_queue_count(dqm, qpd, q);
+       }
+}
+
+static int detect_queue_hang(struct device_queue_manager *dqm)
+{
+       int i;
+
+       memset(dqm->detect_hang_info, 0, dqm->detect_hang_info_size);
+
+       for (i = 0; i < AMDGPU_MAX_QUEUES; ++i) {
+               uint32_t mec, pipe, queue;
+               int xcc_id;
+
+               mec = (i / dqm->dev->kfd->shared_resources.num_queue_per_pipe)
+                       / dqm->dev->kfd->shared_resources.num_pipe_per_mec;
+
+               if (mec || !test_bit(i, dqm->dev->kfd->shared_resources.cp_queue_bitmap))
+                       continue;
+
+               amdgpu_queue_mask_bit_to_mec_queue(dqm->dev->adev, i, &mec, &pipe, &queue);
+
+               for_each_inst(xcc_id, dqm->dev->xcc_mask) {
+                       uint64_t queue_addr = dqm->dev->kfd2kgd->hqd_get_pq_addr(
+                                               dqm->dev->adev, pipe, queue, xcc_id);
+                       struct dqm_detect_hang_info hang_info;
+
+                       if (!queue_addr)
+                               continue;
+
+                       hang_info.pipe_id = pipe;
+                       hang_info.queue_id = queue;
+                       hang_info.xcc_id = xcc_id;
+                       hang_info.queue_address = queue_addr;
+
+                       dqm->detect_hang_info[dqm->detect_hang_count] = hang_info;
+                       dqm->detect_hang_count++;
+               }
+       }
+
+       return dqm->detect_hang_count;
+}
+
+static struct queue *find_queue_by_address(struct device_queue_manager *dqm, uint64_t queue_address)
+{
+       struct device_process_node *cur;
+       struct qcm_process_device *qpd;
+       struct queue *q;
+
+       list_for_each_entry(cur, &dqm->queues, list) {
+               qpd = cur->qpd;
+               list_for_each_entry(q, &qpd->queues_list, list) {
+                       if (queue_address == q->properties.queue_address)
+                               return q;
+               }
+       }
+
+       return NULL;
+}
+
+/* only for compute queue */
+static int reset_queues_on_hws_hang(struct device_queue_manager *dqm)
+{
+       int r = 0, reset_count = 0, i;
+
+       if (!dqm->detect_hang_info || dqm->is_hws_hang)
+               return -EIO;
+
+       /* assume dqm locked. */
+       if (!detect_queue_hang(dqm))
+               return -ENOTRECOVERABLE;
+
+       for (i = 0; i < dqm->detect_hang_count; i++) {
+               struct dqm_detect_hang_info hang_info = dqm->detect_hang_info[i];
+               struct queue *q = find_queue_by_address(dqm, hang_info.queue_address);
+               struct kfd_process_device *pdd;
+               uint64_t queue_addr = 0;
+
+               if (!q) {
+                       r = -ENOTRECOVERABLE;
+                       goto reset_fail;
+               }
+
+               pdd = kfd_get_process_device_data(dqm->dev, q->process);
+               if (!pdd) {
+                       r = -ENOTRECOVERABLE;
+                       goto reset_fail;
+               }
+
+               queue_addr = dqm->dev->kfd2kgd->hqd_reset(dqm->dev->adev,
+                               hang_info.pipe_id, hang_info.queue_id, hang_info.xcc_id,
+                               KFD_UNMAP_LATENCY_MS);
+
+               /* either reset failed or we reset an unexpected queue. */
+               if (queue_addr != q->properties.queue_address) {
+                       r = -ENOTRECOVERABLE;
+                       goto reset_fail;
+               }
+
+               set_queue_as_reset(dqm, q, &pdd->qpd);
+               reset_count++;
+       }
+
+       if (reset_count == dqm->detect_hang_count)
+               kfd_signal_reset_event(dqm->dev);
+       else
+               r = -ENOTRECOVERABLE;
+
+reset_fail:
+       dqm->detect_hang_count = 0;
+
+       return r;
+}
+
 /* dqm->lock mutex has to be locked before calling this function */
 static int unmap_queues_cpsch(struct device_queue_manager *dqm,
                                enum kfd_unmap_queues_filter filter,
@@ -1979,11 +2133,14 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm,
         */
        mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
        if (mqd_mgr->check_preemption_failed(mqd_mgr, dqm->packet_mgr.priv_queue->queue->mqd)) {
-               while (halt_if_hws_hang)
-                       schedule();
-               kfd_hws_hang(dqm);
-               retval = -ETIME;
-               goto out;
+               if (reset_queues_on_hws_hang(dqm)) {
+                       while (halt_if_hws_hang)
+                               schedule();
+                       dqm->is_hws_hang = true;
+                       kfd_hws_hang(dqm);
+                       retval = -ETIME;
+                       goto out;
+               }
        }
 
        /* We need to reset the grace period value for this device */
@@ -2002,8 +2159,7 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm,
 }
 
 /* only for compute queue */
-static int reset_queues_cpsch(struct device_queue_manager *dqm,
-                       uint16_t pasid)
+static int reset_queues_cpsch(struct device_queue_manager *dqm, uint16_t pasid)
 {
        int retval;
 
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +12 -0
index 3b9b8eabaacc..dfb36a246637 100644
@@ -210,6 +210,13 @@ struct device_queue_manager_asic_ops {
                                 struct kfd_node *dev);
 };
 
+struct dqm_detect_hang_info {
+       int pipe_id;
+       int queue_id;
+       int xcc_id;
+       uint64_t queue_address;
+};
+
 /**
  * struct device_queue_manager
  *
@@ -264,6 +271,11 @@ struct device_queue_manager {
        uint32_t                wait_times;
 
        wait_queue_head_t       destroy_wait;
+
+       /* for per-queue reset support */
+       struct dqm_detect_hang_info *detect_hang_info;
+       size_t detect_hang_info_size;
+       int detect_hang_count;
 };
 
 void device_queue_manager_init_cik(
drivers/gpu/drm/amd/amdkfd/kfd_events.c +8 -0
index 9b33d9d2c9ad..3a6e7a4c8895 100644
@@ -31,6 +31,7 @@
 #include <linux/memory.h>
 #include "kfd_priv.h"
 #include "kfd_events.h"
+#include "kfd_device_queue_manager.h"
 #include <linux/device.h>
 
 /*
@@ -1244,12 +1245,32 @@ void kfd_signal_reset_event(struct kfd_node *dev)
        idx = srcu_read_lock(&kfd_processes_srcu);
        hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
                int user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
+               struct kfd_process_device *pdd = kfd_get_process_device_data(dev, p);
 
                if (unlikely(user_gpu_id == -EINVAL)) {
                        WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
                        continue;
                }
 
+               if (unlikely(!pdd)) {
+                       WARN_ONCE(1, "Could not get device data from pasid:0x%x\n", p->pasid);
+                       continue;
+               }
+
+               if (dev->dqm->detect_hang_count && !pdd->has_reset_queue)

drivers/gpu/drm/amd/amdkfd/kfd_events.c void kfd_signal_reset_event(struct kfd_node *dev) · line 1270

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_events.c +10 -0
+                       continue;
+
+               if (dev->dqm->detect_hang_count) {
+                       struct amdgpu_task_info *ti;
+
+                       ti = amdgpu_vm_get_task_info_pasid(dev->adev, p->pasid);
+                       if (ti) {
+                               DRM_ERROR("Process info: process %s tid %d thread %s pid %d\n",
+                               ti->process_name, ti->tgid, ti->task_name, ti->pid);
+                               amdgpu_vm_put_task_info(ti);

Use dev_err() here so we know which device issued the error on multi-GPU systems. With that fixed, the patch looks good to me. I'm not a KFD expert, so:

Alex

Quoted 78 lines
drivers/gpu/drm/amd/amdkfd/kfd_events.c +3 -0
+                       }
+               }
+
                rcu_read_lock();
 
                id = KFD_FIRST_NONSIGNAL_EVENT_ID;
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +5 -1
index 66c73825c0a0..84e8ea3a8a0c 100644
@@ -321,8 +321,11 @@ static void update_mqd(struct mqd_manager *mm, void *mqd,
 static bool check_preemption_failed(struct mqd_manager *mm, void *mqd)
 {
        struct v9_mqd *m = (struct v9_mqd *)mqd;
+       uint32_t doorbell_id = m->queue_doorbell_id0;
 
-       return kfd_check_hiq_mqd_doorbell_id(mm->dev, m->queue_doorbell_id0, 0);
+       m->queue_doorbell_id0 = 0;
+
+       return kfd_check_hiq_mqd_doorbell_id(mm->dev, doorbell_id, 0);
 }
 
 static int get_wave_state(struct mqd_manager *mm, void *mqd,
@@ -624,6 +627,7 @@ static bool check_preemption_failed_v9_4_3(struct mqd_manager *mm, void *mqd)
                m = get_mqd(mqd + hiq_mqd_size * inst);
                ret |= kfd_check_hiq_mqd_doorbell_id(mm->dev,
                                        m->queue_doorbell_id0, inst);
+               m->queue_doorbell_id0 = 0;
                ++inst;
        }
 
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +3 -0
index b5cae48dff66..892a85408c09 100644
@@ -843,6 +843,9 @@ struct kfd_process_device {
        void *proc_ctx_bo;
        uint64_t proc_ctx_gpu_addr;
        void *proc_ctx_cpu_ptr;
+
+       /* Tracks queue reset status */
+       bool has_reset_queue;
 };
 
 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
drivers/gpu/drm/amd/amdkfd/kfd_process.c +2 -0
index 9e29b92eb523..a902950cc060 100644
@@ -1851,6 +1851,8 @@ int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger)
                        goto fail;
                }
                n_evicted++;
+
+               pdd->dev->dqm->is_hws_hang = false;
        }
 
        return r;
drivers/gpu/drm/amd/include/kgd_kfd_interface.h +6 -1
index 6d094cf3587d..7744ca3ef4b1 100644
@@ -318,6 +318,12 @@ struct kfd2kgd_calls {
        void (*program_trap_handler_settings)(struct amdgpu_device *adev,
                        uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr,
                        uint32_t inst);
+       uint64_t (*hqd_get_pq_addr)(struct amdgpu_device *adev,
+                                   uint32_t pipe_id, uint32_t queue_id,
+                                   uint32_t inst);
+       uint64_t (*hqd_reset)(struct amdgpu_device *adev,
+                             uint32_t pipe_id, uint32_t queue_id,
+                             uint32_t inst, unsigned int utimeout);
 };
 
 #endif /* KGD_KFD_INTERFACE_H_INCLUDED */
--

2.34.1

maintainer
#7 lore raw
In reply to Jonathan Kim
Quoted 8 lines · Jonathan Kim

Support per-queue reset for GFX9. The recommendation is for the driver to target reset the HW queue via a SPI MMIO register write.

Since this requires pipe and HW queue info and MEC FW is limited to doorbell reports of hung queues after an unmap failure, scan the HW queue slots defined by SET_RESOURCES first to identify the user queue candidates to reset.

Quoted 8 lines

Only signal reset events to processes that have had a queue reset.

If queue reset fails, fall back to GPU reset.

v2: move reset queue flag for house keeping to process device. split detect and reset into separate functions.
make reset call safe during power saving modes. clean up some other nitpicks.

Some more nit-picks inline.

Quoted 21 lines
---
 .../drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c  |   2 +
 .../drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c   |   4 +-
 .../drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c   |   4 +-
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c    |  16 ++
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h    |   9 +
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c  |   4 +-
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c    |  18 +-
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c |  85 +++++++++
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h |   9 +
 .../drm/amd/amdkfd/kfd_device_queue_manager.c | 172 +++++++++++++++++-
 .../drm/amd/amdkfd/kfd_device_queue_manager.h |  12 ++
 drivers/gpu/drm/amd/amdkfd/kfd_events.c       |  21 +++
 .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c   |   6 +-
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |   3 +
 drivers/gpu/drm/amd/amdkfd/kfd_process.c      |   2 +
 .../gpu/drm/amd/include/kgd_kfd_interface.h   |   6 +
 16 files changed, 360 insertions(+), 13 deletions(-)

[snip]

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int start_cpsch(struct device_queue_manager *dqm) · line 1709

Quoted 11 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +7 -0
@@ -1680,6 +1700,14 @@ static int start_cpsch(struct device_queue_manager *dqm)
 					&dqm->wait_times);
 	}
 
+	/* setup per-queue reset detection buffer  */
+	num_hw_queue_slots =  dqm->dev->kfd->shared_resources.num_queue_per_pipe *
+			      dqm->dev->kfd->shared_resources.num_pipe_per_mec *
+			      NUM_XCC(dqm->dev->xcc_mask);
+
+	dqm->detect_hang_info_size = num_hw_queue_slots * sizeof(struct dqm_detect_hang_info);
+	dqm->detect_hang_info = kzalloc(dqm->detect_hang_info_size, GFP_KERNEL);

You need to check the return value and handle allocation failures.

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int stop_cpsch(struct device_queue_manager *dqm) · line 1744

Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -0
+
 	dqm_unlock(dqm);
 
 	return 0;
@@ -1713,6 +1741,7 @@ static int stop_cpsch(struct device_queue_manager *dqm)
 	kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
 	if (!dqm->dev->kfd->shared_resources.enable_mes)
 		pm_uninit(&dqm->packet_mgr);
+	kfree(dqm->detect_hang_info);

Reset dqm->detect_hang_info to NULL to avoid a dangling pointer.

Quoted 17 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +10 -0
 	dqm_unlock(dqm);
 
 	return 0;
@@ -1929,6 +1958,131 @@ static int map_queues_cpsch(struct device_queue_manager *dqm)
 	return retval;
 }
 
+static void set_queue_as_reset(struct device_queue_manager *dqm, struct queue *q,
+			       struct qcm_process_device *qpd)
+{
+	struct kfd_process_device *pdd = qpd_to_pdd(qpd);
+
+	pr_err("queue id 0x%0x at pasid 0x%0x is reset\n",
+	       q->properties.queue_id, q->process->pasid);
+
+	pdd->has_reset_queue = true;
+	if (q->properties.is_active) {

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int map_queues_cpsch(struct device_queue_manager *dqm) · line 1980

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +10 -0
+		q->properties.is_active = false;
+		decrement_queue_count(dqm, qpd, q);
+	}
+}
+
+static int detect_queue_hang(struct device_queue_manager *dqm)
+{
+	int i;
+
+	memset(dqm->detect_hang_info, 0, dqm->detect_hang_info_size);

Set dqm->detect_hang_count to 0 to avoid overflows in case multiple hand detections get kicked off. Or if that's not possible, just print a WARN_ON if detect_hang_count is not 0 and return.

Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +3 -0
+
+	for (i = 0; i < AMDGPU_MAX_QUEUES; ++i) {
+		uint32_t mec, pipe, queue;

[snip]

Quoted 16 lines
drivers/gpu/drm/amd/amdkfd/kfd_events.c +6 -0
@@ -1244,12 +1245,32 @@ void kfd_signal_reset_event(struct kfd_node *dev)
 	idx = srcu_read_lock(&kfd_processes_srcu);
 	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
 		int user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
+		struct kfd_process_device *pdd = kfd_get_process_device_data(dev, p);
 
 		if (unlikely(user_gpu_id == -EINVAL)) {
 			WARN_ONCE(1, "Could not get user_gpu_id from dev->id:%x\n", dev->id);
 			continue;
 		}
 
+		if (unlikely(!pdd)) {
+			WARN_ONCE(1, "Could not get device data from pasid:0x%x\n", p->pasid);
+			continue;
+		}
+

drivers/gpu/drm/amd/amdkfd/kfd_events.c void kfd_signal_reset_event(struct kfd_node *dev) · line 1269

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_events.c +10 -0
+		if (dev->dqm->detect_hang_count && !pdd->has_reset_queue)
+			continue;
+
+		if (dev->dqm->detect_hang_count) {
+			struct amdgpu_task_info *ti;
+
+			ti = amdgpu_vm_get_task_info_pasid(dev->adev, p->pasid);
+			if (ti) {
+				DRM_ERROR("Process info: process %s tid %d thread %s pid %d\n",
+				ti->process_name, ti->tgid, ti->task_name, ti->pid);

This looks like debug code that's not meant to be checked in. If you mean to check it in, the message needs more information. Use dev_err or dev_info to print the affected GPU and put something about a queue reset into the message.

Regards,
Felix

Quoted 76 lines
drivers/gpu/drm/amd/amdkfd/kfd_events.c +4 -0
+				amdgpu_vm_put_task_info(ti);
+			}
+		}
+
 		rcu_read_lock();
 
 		id = KFD_FIRST_NONSIGNAL_EVENT_ID;
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +5 -1
index 66c73825c0a0..84e8ea3a8a0c 100644
@@ -321,8 +321,11 @@ static void update_mqd(struct mqd_manager *mm, void *mqd,
 static bool check_preemption_failed(struct mqd_manager *mm, void *mqd)
 {
 	struct v9_mqd *m = (struct v9_mqd *)mqd;
+	uint32_t doorbell_id = m->queue_doorbell_id0;
 
-	return kfd_check_hiq_mqd_doorbell_id(mm->dev, m->queue_doorbell_id0, 0);
+	m->queue_doorbell_id0 = 0;
+
+	return kfd_check_hiq_mqd_doorbell_id(mm->dev, doorbell_id, 0);
 }
 
 static int get_wave_state(struct mqd_manager *mm, void *mqd,
@@ -624,6 +627,7 @@ static bool check_preemption_failed_v9_4_3(struct mqd_manager *mm, void *mqd)
 		m = get_mqd(mqd + hiq_mqd_size * inst);
 		ret |= kfd_check_hiq_mqd_doorbell_id(mm->dev,
 					m->queue_doorbell_id0, inst);
+		m->queue_doorbell_id0 = 0;
 		++inst;
 	}
 
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +3 -0
index b5cae48dff66..892a85408c09 100644
@@ -843,6 +843,9 @@ struct kfd_process_device {
 	void *proc_ctx_bo;
 	uint64_t proc_ctx_gpu_addr;
 	void *proc_ctx_cpu_ptr;
+
+	/* Tracks queue reset status */
+	bool has_reset_queue;
 };
 
 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
drivers/gpu/drm/amd/amdkfd/kfd_process.c +2 -0
index 9e29b92eb523..a902950cc060 100644
@@ -1851,6 +1851,8 @@ int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger)
 			goto fail;
 		}
 		n_evicted++;
+
+		pdd->dev->dqm->is_hws_hang = false;
 	}
 
 	return r;
drivers/gpu/drm/amd/include/kgd_kfd_interface.h +6 -0
index 6d094cf3587d..7744ca3ef4b1 100644
@@ -318,6 +318,12 @@ struct kfd2kgd_calls {
 	void (*program_trap_handler_settings)(struct amdgpu_device *adev,
 			uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr,
 			uint32_t inst);
+	uint64_t (*hqd_get_pq_addr)(struct amdgpu_device *adev,
+				    uint32_t pipe_id, uint32_t queue_id,
+				    uint32_t inst);
+	uint64_t (*hqd_reset)(struct amdgpu_device *adev,
+			      uint32_t pipe_id, uint32_t queue_id,
+			      uint32_t inst, unsigned int utimeout);
 };
 
 #endif	/* KGD_KFD_INTERFACE_H_INCLUDED */
#8 lore raw

[Public]

Quoted 153 lines

-----Original Message-----
From: Kuehling, Felix <[email protected]>
Sent: Tuesday, July 30, 2024 6:07 PM
To: Kim, Jonathan <[email protected]>; [email protected]
Cc: Deucher, Alexander <[email protected]>
Subject: Re: [PATCH 1/2] drm/amdkfd: support per-queue reset on gfx9

Quoted 8 lines · Jonathan Kim

Support per-queue reset for GFX9. The recommendation is for the driver to target reset the HW queue via a SPI MMIO register write.

Since this requires pipe and HW queue info and MEC FW is limited to doorbell reports of hung queues after an unmap failure, scan the HW queue slots defined by SET_RESOURCES first to identify the user queue candidates to reset.

Quoted 8 lines

Only signal reset events to processes that have had a queue reset.

If queue reset fails, fall back to GPU reset.

v2: move reset queue flag for house keeping to process device. split detect and reset into separate functions.
make reset call safe during power saving modes. clean up some other nitpicks.

Some more nit-picks inline.

Quoted 13 lines
---
 .../drm/amd/amdgpu/amdgpu_amdkfd_aldebaran.c  |   2 +
 .../drm/amd/amdgpu/amdgpu_amdkfd_arcturus.c   |   4 +-
 .../drm/amd/amdgpu/amdgpu_amdkfd_gc_9_4_3.c   |   4 +-
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.c    |  16 ++
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10.h    |   9 +
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v10_3.c  |   4 +-
 .../drm/amd/amdgpu/amdgpu_amdkfd_gfx_v11.c    |  18 +-
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.c |  85 +++++++++
 .../gpu/drm/amd/amdgpu/amdgpu_amdkfd_gfx_v9.h |   9 +
 .../drm/amd/amdkfd/kfd_device_queue_manager.c | 172

+++++++++++++++++-

Quoted 8 lines
 .../drm/amd/amdkfd/kfd_device_queue_manager.h |  12 ++
 drivers/gpu/drm/amd/amdkfd/kfd_events.c       |  21 +++
 .../gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c   |   6 +-
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h         |   3 +
 drivers/gpu/drm/amd/amdkfd/kfd_process.c      |   2 +
 .../gpu/drm/amd/include/kgd_kfd_interface.h   |   6 +
 16 files changed, 360 insertions(+), 13 deletions(-)

[snip]

Review static int start_cpsch(struct

Quoted 1 line
(quoted diff) +0 -0
@@ -1680,6 +1700,14 @@ static int start_cpsch(struct

device_queue_manager *dqm)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int start_cpsch(struct device_queue_manager *dqm)

Quoted 11 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -0
                                    &dqm->wait_times);
    }
 
+   /* setup per-queue reset detection buffer  */
+   num_hw_queue_slots =  dqm->dev->kfd-

shared_resources.num_queue_per_pipe *
+ dqm->dev->kfd-
shared_resources.num_pipe_per_mec *

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +3 -0
+                         NUM_XCC(dqm->dev->xcc_mask);
+
+   dqm->detect_hang_info_size = num_hw_queue_slots * sizeof(struct

dqm_detect_hang_info);

Quoted 1 line

+ dqm->detect_hang_info = kzalloc(dqm->detect_hang_info_size,

GFP_KERNEL);

You need to check the return value and handle allocation failures.

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int stop_cpsch(struct

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -0
+
    dqm_unlock(dqm);
 
    return 0;
@@ -1713,6 +1741,7 @@ static int stop_cpsch(struct

device_queue_manager *dqm)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int stop_cpsch(struct device_queue_manager *dqm) · line 1744

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -0
    kfd_gtt_sa_free(dqm->dev, dqm->fence_mem);
    if (!dqm->dev->kfd->shared_resources.enable_mes)
            pm_uninit(&dqm->packet_mgr);
+   kfree(dqm->detect_hang_info);

Reset dqm->detect_hang_info to NULL to avoid a dangling pointer.

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int map_queues_cpsch(struct

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
    dqm_unlock(dqm);
 
    return 0;
@@ -1929,6 +1958,131 @@ static int map_queues_cpsch(struct

device_queue_manager *dqm)

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int map_queues_cpsch(struct device_queue_manager *dqm)

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -0
    return retval;
 }
 
+static void set_queue_as_reset(struct device_queue_manager *dqm, struct

queue *q,

Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +9 -0
+                          struct qcm_process_device *qpd)
+{
+   struct kfd_process_device *pdd = qpd_to_pdd(qpd);
+
+   pr_err("queue id 0x%0x at pasid 0x%0x is reset\n",
+          q->properties.queue_id, q->process->pasid);
+
+   pdd->has_reset_queue = true;
+   if (q->properties.is_active) {

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int map_queues_cpsch(struct device_queue_manager *dqm) · line 1980

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +10 -0
+           q->properties.is_active = false;
+           decrement_queue_count(dqm, qpd, q);
+   }
+}
+
+static int detect_queue_hang(struct device_queue_manager *dqm)
+{
+   int i;
+
+   memset(dqm->detect_hang_info, 0, dqm->detect_hang_info_size);

Set dqm->detect_hang_count to 0 to avoid overflows in case multiple hand detections get kicked off. Or if that's not possible, just print a WARN_ON if detect_hang_count is not 0 and return.

Quoted 3 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +3 -0
+
+   for (i = 0; i < AMDGPU_MAX_QUEUES; ++i) {
+           uint32_t mec, pipe, queue;

[snip]

Review void kfd_signal_reset_event(struct kfd_node

Quoted 1 line
(quoted diff) +0 -0
@@ -1244,12 +1245,32 @@ void kfd_signal_reset_event(struct kfd_node

*dev)

drivers/gpu/drm/amd/amdkfd/kfd_events.c void kfd_signal_reset_event(struct kfd_node *dev)

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_events.c +1 -0
    idx = srcu_read_lock(&kfd_processes_srcu);
    hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
            int user_gpu_id = kfd_process_get_user_gpu_id(p, dev->id);
+           struct kfd_process_device *pdd =

kfd_get_process_device_data(dev, p);

Review void kfd_signal_reset_event(struct kfd_node · line 1250

Quoted 3 lines
(quoted diff) +0 -0
            if (unlikely(user_gpu_id == -EINVAL)) {
                    WARN_ONCE(1, "Could not get user_gpu_id from

dev->id:%x\n", dev->id);

drivers/gpu/drm/amd/amdkfd/kfd_events.c void kfd_signal_reset_event(struct kfd_node *dev)

Quoted 5 lines
drivers/gpu/drm/amd/amdkfd/kfd_events.c +2 -0
                    continue;
            }
 
+           if (unlikely(!pdd)) {
+                   WARN_ONCE(1, "Could not get device data from

pasid:0x%x\n", p->pasid);

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_events.c +4 -0
+                   continue;
+           }
+
+           if (dev->dqm->detect_hang_count && !pdd-

drivers/gpu/drm/amd/amdkfd/kfd_events.c void kfd_signal_reset_event(struct kfd_node *dev)

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_events.c +8 -0
 has_reset_queue)
+                   continue;
+
+           if (dev->dqm->detect_hang_count) {
+                   struct amdgpu_task_info *ti;
+
+                   ti = amdgpu_vm_get_task_info_pasid(dev->adev, p-
 pasid);
+                   if (ti) {
+                           DRM_ERROR("Process info: process %s tid %d

thread %s pid %d\n",

Quoted 7 lines

Review void kfd_signal_reset_event(struct kfd_node · line 1271

Quoted 2 lines
(quoted diff) +1 -0
+                           ti->process_name, ti->tgid, ti->task_name, ti-
 pid);

This looks like debug code that's not meant to be checked in. If you mean to check it in, the message needs more information. Use dev_err or dev_info to print the affected GPU and put something about a queue reset into the message.

Thanks for the review. Yes this was meant to be in as it's an ask to flag offending applications in dmesgs. I'll update it with your suggestions.

Jon

Quoted 89 lines

Regards,
Felix

Review

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_events.c +4 -0
+                           amdgpu_vm_put_task_info(ti);
+                   }
+           }
+
            rcu_read_lock();
            id = KFD_FIRST_NONSIGNAL_EVENT_ID;
(quoted diff) +0 -0

b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c

drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c static void update_mqd(struct mqd_manager

Quoted 4 lines

index 66c73825c0a0..84e8ea3a8a0c 100644

drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +0 -0
@@ -321,8 +321,11 @@ static void update_mqd(struct mqd_manager

*mm, void *mqd,

drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c static bool check_preemption_failed_v9_4_3(struct

Quoted 14 lines
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +4 -1
 static bool check_preemption_failed(struct mqd_manager *mm, void *mqd)
 {
    struct v9_mqd *m = (struct v9_mqd *)mqd;
+   uint32_t doorbell_id = m->queue_doorbell_id0;
 
-   return kfd_check_hiq_mqd_doorbell_id(mm->dev, m-
 queue_doorbell_id0, 0);
+   m->queue_doorbell_id0 = 0;
+
+   return kfd_check_hiq_mqd_doorbell_id(mm->dev, doorbell_id, 0);
 }
 

static int get_wave_state(struct mqd_manager *mm, void *mqd,

drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +0 -0
@@ -624,6 +627,7 @@ static bool check_preemption_failed_v9_4_3(struct

mqd_manager *mm, void *mqd)

Review

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c +1 -0
            m = get_mqd(mqd + hiq_mqd_size * inst);
            ret |= kfd_check_hiq_mqd_doorbell_id(mm->dev,
                                    m->queue_doorbell_id0, inst);
+           m->queue_doorbell_id0 = 0;
            ++inst;
    }
 
(quoted diff) +0 -0

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

Review

Quoted 14 lines

index b5cae48dff66..892a85408c09 100644

drivers/gpu/drm/amd/amdkfd/kfd_priv.h +3 -0
@@ -843,6 +843,9 @@ struct kfd_process_device {
    void *proc_ctx_bo;
    uint64_t proc_ctx_gpu_addr;
    void *proc_ctx_cpu_ptr;
+
+   /* Tracks queue reset status */
+   bool has_reset_queue;
 };
 
 #define qpd_to_pdd(x) container_of(x, struct kfd_process_device, qpd)
(quoted diff) +0 -0

b/drivers/gpu/drm/amd/amdkfd/kfd_process.c

drivers/gpu/drm/amd/amdkfd/kfd_process.c int kfd_process_evict_queues(struct kfd_process

Quoted 4 lines

index 9e29b92eb523..a902950cc060 100644

drivers/gpu/drm/amd/amdkfd/kfd_process.c +0 -0
@@ -1851,6 +1851,8 @@ int kfd_process_evict_queues(struct kfd_process

*p, uint32_t trigger)

Review

Quoted 9 lines
drivers/gpu/drm/amd/amdkfd/kfd_process.c +2 -0
                    goto fail;
            }
            n_evicted++;
+
+           pdd->dev->dqm->is_hws_hang = false;
    }
 
    return r;
(quoted diff) +0 -0

b/drivers/gpu/drm/amd/include/kgd_kfd_interface.h

Quoted 16 lines

index 6d094cf3587d..7744ca3ef4b1 100644

drivers/gpu/drm/amd/include/kgd_kfd_interface.h +6 -0
@@ -318,6 +318,12 @@ struct kfd2kgd_calls {
    void (*program_trap_handler_settings)(struct amdgpu_device *adev,
                    uint32_t vmid, uint64_t tba_addr, uint64_t tma_addr,
                    uint32_t inst);
+   uint64_t (*hqd_get_pq_addr)(struct amdgpu_device *adev,
+                               uint32_t pipe_id, uint32_t queue_id,
+                               uint32_t inst);
+   uint64_t (*hqd_reset)(struct amdgpu_device *adev,
+                         uint32_t pipe_id, uint32_t queue_id,
+                         uint32_t inst, unsigned int utimeout);
 };
 
 #endif     /* KGD_KFD_INTERFACE_H_INCLUDED */