Skip to contentSkip to conversation

drm/amdkfd: fix trap handling work around for debugging

PATCH In review Jonathan Kim 3 participants 6 messages Jul 14, 2023 – Jul 19, 2023

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

PATCH 1/2 drm/amdkfd: fix trap handling work around for debugging

Update the list of devices that require the cwsr trap handling workaround for debugging use cases.

---
 drivers/gpu/drm/amd/amdkfd/kfd_debug.c                | 5 ++---
 drivers/gpu/drm/amd/amdkfd/kfd_debug.h                | 6 ++++++
 drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 6 ++----
 3 files changed, 10 insertions(+), 7 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +2 -3
index 190b03efe5ff..ccfc81f085ce 100644
@@ -302,8 +302,7 @@ static int kfd_dbg_set_queue_workaround(struct queue *q, bool enable)
 	if (!q)
 		return 0;
 
-	if (KFD_GC_VERSION(q->device) < IP_VERSION(11, 0, 0) ||
-	    KFD_GC_VERSION(q->device) >= IP_VERSION(12, 0, 0))
+	if (!kfd_dbg_has_cwsr_workaround(q->device))
 		return 0;
 
 	if (enable && q->properties.is_user_cu_masked)
@@ -349,7 +348,7 @@ int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd)
 {
 	uint32_t spi_dbg_cntl = pdd->spi_dbg_override | pdd->spi_dbg_launch_mode;
 	uint32_t flags = pdd->process->dbg_flags;
-	bool sq_trap_en = !!spi_dbg_cntl;
+	bool sq_trap_en = !!spi_dbg_cntl || !kfd_dbg_has_cwsr_workaround(pdd->dev);
 
 	if (!kfd_dbg_is_per_vmid_supported(pdd->dev))
 		return 0;
drivers/gpu/drm/amd/amdkfd/kfd_debug.h +6 -0
index ba616ed17dee..586d7f886712 100644
@@ -101,6 +101,12 @@ static inline bool kfd_dbg_is_rlc_restore_supported(struct kfd_node *dev)
 		 KFD_GC_VERSION(dev) == IP_VERSION(10, 1, 1));
 }
 
+static inline bool kfd_dbg_has_cwsr_workaround(struct kfd_node *dev)
+{
+	return KFD_GC_VERSION(dev) >= IP_VERSION(11, 0, 0) &&
+	       KFD_GC_VERSION(dev) <= IP_VERSION(11, 0, 3);
+}
+
 static inline bool kfd_dbg_has_gws_support(struct kfd_node *dev)
 {
 	if ((KFD_GC_VERSION(dev) == IP_VERSION(9, 0, 1)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -4
index 31cac1fd0d58..761963ad6154 100644
@@ -226,8 +226,7 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
 	queue_input.paging = false;
 	queue_input.tba_addr = qpd->tba_addr;
 	queue_input.tma_addr = qpd->tma_addr;
-	queue_input.trap_en = KFD_GC_VERSION(q->device) < IP_VERSION(11, 0, 0) ||
-			      KFD_GC_VERSION(q->device) > IP_VERSION(11, 0, 3);
+	queue_input.trap_en = !kfd_dbg_has_cwsr_workaround(q->device);
 	queue_input.skip_process_ctx_clear = qpd->pqm->process->debug_trap_enabled;
 
 	queue_type = convert_to_mes_queue_type(q->properties.type);
@@ -1827,8 +1826,7 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
 	 */
 	q->properties.is_evicted = !!qpd->evicted;
 	q->properties.is_dbg_wa = qpd->pqm->process->debug_trap_enabled &&
-			KFD_GC_VERSION(q->device) >= IP_VERSION(11, 0, 0) &&
-			KFD_GC_VERSION(q->device) <= IP_VERSION(11, 0, 3);
+				  kfd_dbg_has_cwsr_workaround(q->device);
 
 	if (qd)
 		mqd_mgr->restore_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, &q->gart_mqd_addr,
Signature · 1 line
2.25.1
#1 lore raw

PATCH 2/2 drm/amdkfd: enable cooperative groups for gfx11

MES can concurrently schedule queues on the device that require exclusive device access if marked exclusively_scheduled without the requirement of GWS. Similar to the F32 HWS, MES will manage quality of service for these queues.
Use this for cooperative groups since cooperative groups are device occupancy limited.

Since some GFX11 devices can only be debugged with partial CUs, do not allow the debugging of cooperative groups on these devices as the CU occupancy limit will change on attach.

In addition, zero initialize the MES add queue submission vector for MES initialization tests as we do not want these to be cooperative dispatches.

NOTE: FIXME MES FW enablement checks are a placeholder at the moment and will be updated when the binary revision number is finalized.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c               |  2 +-
 drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h               |  1 +
 drivers/gpu/drm/amd/amdgpu/mes_v11_0.c                |  2 ++
 drivers/gpu/drm/amd/amdkfd/kfd_chardev.c              |  3 ++-
 drivers/gpu/drm/amd/amdkfd/kfd_debug.c                |  3 ++-
 drivers/gpu/drm/amd/amdkfd/kfd_device.c               |  6 +++++-
 drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c |  9 ++++-----
 .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c    | 11 +++++++----
 drivers/gpu/drm/amd/include/mes_v11_api_def.h         |  4 +++-
 9 files changed, 27 insertions(+), 14 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +1 -1
index e9091ebfe230..8d13623389d8 100644
@@ -638,7 +638,7 @@ int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
 {
 	struct amdgpu_mes_queue *queue;
 	struct amdgpu_mes_gang *gang;
-	struct mes_add_queue_input queue_input;
+	struct mes_add_queue_input queue_input = {0};
 	unsigned long flags;
 	int r;
 
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +1 -0
index 2d6ac30b7135..2053954a235c 100644
@@ -224,6 +224,7 @@ struct mes_add_queue_input {
 	uint32_t	is_kfd_process;
 	uint32_t	is_aql_queue;
 	uint32_t	queue_size;
+	uint32_t	exclusively_scheduled;
 };
 
 struct mes_remove_queue_input {
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +2 -0
index 1bdaa00c0b46..8e67e965f7ea 100644
@@ -214,6 +214,8 @@ static int mes_v11_0_add_hw_queue(struct amdgpu_mes *mes,
 	mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
 	mes_add_queue_pkt.gds_size = input->queue_size;
 
+	mes_add_queue_pkt.exclusively_scheduled = input->exclusively_scheduled;
+
 	return mes_v11_0_submit_pkt_and_poll_completion(mes,
 			&mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
 			offsetof(union MESAPI__ADD_QUEUE, api_status));
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2 -1
index 40ac093b5035..e18401811956 100644
@@ -1487,7 +1487,8 @@ static int kfd_ioctl_alloc_queue_gws(struct file *filep,
 		goto out_unlock;
 	}
 
-	if (!kfd_dbg_has_gws_support(dev) && p->debug_trap_enabled) {
+	if (p->debug_trap_enabled && (!kfd_dbg_has_gws_support(dev) ||
+				       kfd_dbg_has_cwsr_workaround(dev))) {
 		retval = -EBUSY;
 		goto out_unlock;
 	}
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +2 -1
index ccfc81f085ce..895e7f690fd0 100644
@@ -753,7 +753,8 @@ int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd,
 		if (!KFD_IS_SOC15(pdd->dev))
 			return -ENODEV;
 
-		if (!kfd_dbg_has_gws_support(pdd->dev) && pdd->qpd.num_gws)
+		if (pdd->qpd.num_gws && (!kfd_dbg_has_gws_support(pdd->dev) ||
+					  kfd_dbg_has_cwsr_workaround(pdd->dev)))
 			return -EBUSY;
 	}
 
drivers/gpu/drm/amd/amdkfd/kfd_device.c +5 -1
index 0b3dc754e06b..9f4f75fd2fb2 100644
@@ -508,6 +508,7 @@ static int kfd_gws_init(struct kfd_node *node)
 {
 	int ret = 0;
 	struct kfd_dev *kfd = node->kfd;
+	uint32_t mes_rev = node->adev->mes.sched_version & AMDGPU_MES_VERSION_MASK;
 
 	if (node->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS)
 		return 0;
@@ -524,7 +525,10 @@ static int kfd_gws_init(struct kfd_node *node)
 		(KFD_GC_VERSION(node) == IP_VERSION(9, 4, 3)) ||
 		(KFD_GC_VERSION(node) >= IP_VERSION(10, 3, 0)
 			&& KFD_GC_VERSION(node) < IP_VERSION(11, 0, 0)
-			&& kfd->mec2_fw_version >= 0x6b))))
+			&& kfd->mec2_fw_version >= 0x6b) ||
+		(KFD_GC_VERSION(node) >= IP_VERSION(11, 0, 0)
+			&& KFD_GC_VERSION(node) < IP_VERSION(12, 0, 0)
+			&& mes_rev >= 68)))) /* FIXME: Placeholder version */
 		ret = amdgpu_amdkfd_alloc_gws(node->adev,
 				node->adev->gds.gws_size, &node->gws);
 
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +4 -5
index 761963ad6154..7e8bc7328a79 100644
@@ -237,10 +237,9 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
 	}
 	queue_input.queue_type = (uint32_t)queue_type;
 
-	if (q->gws) {
-		queue_input.gws_base = 0;
-		queue_input.gws_size = qpd->num_gws;
-	}
+	queue_input.exclusively_scheduled = q->properties.is_gws;
+	if (q->properties.is_gws && dqm->gws_queue_count > 1)
+		pr_warn("Runlist is getting oversubscibed.  Expect reduced ROCm performance.\n");
 
 	amdgpu_mes_lock(&adev->mes);
 	r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
@@ -250,7 +249,7 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
 			q->properties.doorbell_off);
 		pr_err("MES might be in unrecoverable state, issue a GPU reset\n");
 		kfd_hws_hang(dqm);
-}
+	}
 
 	return r;
 }
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +7 -4
index ba9d69054119..139c6b58bb7e 100644
@@ -123,7 +123,7 @@ int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
 	if (!gws && pdd->qpd.num_gws == 0)
 		return -EINVAL;
 
-	if (KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 3)) {
+	if (KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 3) && !dev->kfd->shared_resources.enable_mes) {
 		if (gws)
 			ret = amdgpu_amdkfd_add_gws_to_process(pdd->process->kgd_process_info,
 				gws, &mem);
@@ -136,7 +136,8 @@ int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
 	} else {
 		/*
 		 * Intentionally set GWS to a non-NULL value
-		 * for GFX 9.4.3.
+		 * for devices that do not require GWS but require the formality
+		 * of setting GWS for cooperative groups.
 		 */
 		pqn->q->gws = gws ? ERR_PTR(-ENOMEM) : NULL;
 	}
@@ -173,7 +174,8 @@ void pqm_uninit(struct process_queue_manager *pqm)
 
 	list_for_each_entry_safe(pqn, next, &pqm->queues, process_queue_list) {
 		if (pqn->q && pqn->q->gws &&
-		    KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 3))
+		    KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 3) &&
+				!pqn->q->device->kfd->shared_resources.enable_mes)
 			amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info,
 				pqn->q->gws);
 		kfd_procfs_del_queue(pqn->q);
@@ -455,7 +457,8 @@ int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
 		}
 
 		if (pqn->q->gws) {
-			if (KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 3))
+			if (KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 3) &&
+					!dev->kfd->shared_resources.enable_mes)
 				amdgpu_amdkfd_remove_gws_from_process(
 						pqm->process->kgd_process_info,
 						pqn->q->gws);
drivers/gpu/drm/amd/include/mes_v11_api_def.h +3 -1
index 0997e999416a..b1db2b190187 100644
@@ -275,7 +275,9 @@ union MESAPI__ADD_QUEUE {
 			uint32_t trap_en		: 1;
 			uint32_t is_aql_queue		: 1;
 			uint32_t skip_process_ctx_clear : 1;
-			uint32_t reserved		: 19;
+			uint32_t map_legacy_kq		: 1;
+			uint32_t exclusively_scheduled	: 1;
+			uint32_t reserved		: 17;
 		};
 		struct MES_API_STATUS		api_status;
 		uint64_t                        tma_addr;
Signature · 1 line
2.25.1
maintainer
#2 lore raw

Re: [PATCH 2/2] drm/amdkfd: enable cooperative groups for gfx11

Am 2023-07-14 um 05:37 schrieb Jonathan Kim:

Quoted 11 lines

MES can concurrently schedule queues on the device that require exclusive device access if marked exclusively_scheduled without the requirement of GWS. Similar to the F32 HWS, MES will manage quality of service for these queues.
Use this for cooperative groups since cooperative groups are device occupancy limited.

Since some GFX11 devices can only be debugged with partial CUs, do not allow the debugging of cooperative groups on these devices as the CU occupancy limit will change on attach.

Quoted 8 lines

In addition, zero initialize the MES add queue submission vector for MES initialization tests as we do not want these to be cooperative dispatches.

NOTE: FIXME MES FW enablement checks are a placeholder at the moment and will be updated when the binary revision number is finalized.

Some nit-picks inline. Looks good to me otherwise.

Quoted 17 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c               |  2 +-
  drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h               |  1 +
  drivers/gpu/drm/amd/amdgpu/mes_v11_0.c                |  2 ++
  drivers/gpu/drm/amd/amdkfd/kfd_chardev.c              |  3 ++-
  drivers/gpu/drm/amd/amdkfd/kfd_debug.c                |  3 ++-
  drivers/gpu/drm/amd/amdkfd/kfd_device.c               |  6 +++++-
  drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c |  9 ++++-----
  .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c    | 11 +++++++----
  drivers/gpu/drm/amd/include/mes_v11_api_def.h         |  4 +++-
  9 files changed, 27 insertions(+), 14 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +0 -0
index e9091ebfe230..8d13623389d8 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id, · line 641

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +1 -1
@@ -638,7 +638,7 @@ int amdgpu_mes_add_hw_queue(struct amdgpu_device *adev, int gang_id,
  {
  	struct amdgpu_mes_queue *queue;
  	struct amdgpu_mes_gang *gang;
-	struct mes_add_queue_input queue_input;
+	struct mes_add_queue_input queue_input = {0};

 Generally, it is preferred to use memset to initialize structures on the stack because that also initializes padding.

Quoted 33 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.c +0 -0
  	unsigned long flags;
  	int r;
 
drivers/gpu/drm/amd/amdgpu/amdgpu_mes.h +1 -0
index 2d6ac30b7135..2053954a235c 100644
@@ -224,6 +224,7 @@ struct mes_add_queue_input {
  	uint32_t	is_kfd_process;
  	uint32_t	is_aql_queue;
  	uint32_t	queue_size;
+	uint32_t	exclusively_scheduled;
  };
 
  struct mes_remove_queue_input {
drivers/gpu/drm/amd/amdgpu/mes_v11_0.c +2 -0
index 1bdaa00c0b46..8e67e965f7ea 100644
@@ -214,6 +214,8 @@ static int mes_v11_0_add_hw_queue(struct amdgpu_mes *mes,
  	mes_add_queue_pkt.is_aql_queue = input->is_aql_queue;
  	mes_add_queue_pkt.gds_size = input->queue_size;
 
+	mes_add_queue_pkt.exclusively_scheduled = input->exclusively_scheduled;
+
  	return mes_v11_0_submit_pkt_and_poll_completion(mes,
  			&mes_add_queue_pkt, sizeof(mes_add_queue_pkt),
  			offsetof(union MESAPI__ADD_QUEUE, api_status));
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
index 40ac093b5035..e18401811956 100644

drivers/gpu/drm/amd/amdkfd/kfd_chardev.c static int kfd_ioctl_alloc_queue_gws(struct file *filep, · line 1491

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +2 -1
@@ -1487,7 +1487,8 @@ static int kfd_ioctl_alloc_queue_gws(struct file *filep,
  		goto out_unlock;
  	}
 
-	if (!kfd_dbg_has_gws_support(dev) && p->debug_trap_enabled) {
+	if (p->debug_trap_enabled && (!kfd_dbg_has_gws_support(dev) ||
+				       kfd_dbg_has_cwsr_workaround(dev))) {

Indentation looks off. kfd_dbg_has_cwsr_workaround should be indented one less space. Otherwise you may be incorrectly implying that the ! applies to it.

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_chardev.c +0 -0
  		retval = -EBUSY;
  		goto out_unlock;
  	}
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +0 -0
index ccfc81f085ce..895e7f690fd0 100644

drivers/gpu/drm/amd/amdkfd/kfd_debug.c int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, · line 757

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +2 -1
@@ -753,7 +753,8 @@ int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd,
  		if (!KFD_IS_SOC15(pdd->dev))
  			return -ENODEV;
 
-		if (!kfd_dbg_has_gws_support(pdd->dev) && pdd->qpd.num_gws)
+		if (pdd->qpd.num_gws && (!kfd_dbg_has_gws_support(pdd->dev) ||
+					  kfd_dbg_has_cwsr_workaround(pdd->dev)))

Same as above.

Quoted 16 lines
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +0 -0
  			return -EBUSY;
  	}
 
drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -0
index 0b3dc754e06b..9f4f75fd2fb2 100644
@@ -508,6 +508,7 @@ static int kfd_gws_init(struct kfd_node *node)
  {
  	int ret = 0;
  	struct kfd_dev *kfd = node->kfd;
+	uint32_t mes_rev = node->adev->mes.sched_version & AMDGPU_MES_VERSION_MASK;
 
  	if (node->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS)
  		return 0;

drivers/gpu/drm/amd/amdkfd/kfd_device.c static int kfd_gws_init(struct kfd_node *node) · line 531

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_device.c +4 -1
@@ -524,7 +525,10 @@ static int kfd_gws_init(struct kfd_node *node)
  		(KFD_GC_VERSION(node) == IP_VERSION(9, 4, 3)) ||
  		(KFD_GC_VERSION(node) >= IP_VERSION(10, 3, 0)
  			&& KFD_GC_VERSION(node) < IP_VERSION(11, 0, 0)
-			&& kfd->mec2_fw_version >= 0x6b))))
+			&& kfd->mec2_fw_version >= 0x6b) ||
+		(KFD_GC_VERSION(node) >= IP_VERSION(11, 0, 0)
+			&& KFD_GC_VERSION(node) < IP_VERSION(12, 0, 0)
+			&& mes_rev >= 68)))) /* FIXME: Placeholder version */

Can this comment be removed? You should know the correct MES version before submitting this patch.

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_device.c +0 -0
  		ret = amdgpu_amdkfd_alloc_gws(node->adev,
  				node->adev->gds.gws_size, &node->gws);
 
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
index 761963ad6154..7e8bc7328a79 100644

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q, · line 242

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +3 -4
@@ -237,10 +237,9 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
  	}
  	queue_input.queue_type = (uint32_t)queue_type;
 
-	if (q->gws) {
-		queue_input.gws_base = 0;
-		queue_input.gws_size = qpd->num_gws;
-	}
+	queue_input.exclusively_scheduled = q->properties.is_gws;
+	if (q->properties.is_gws && dqm->gws_queue_count > 1)
+		pr_warn("Runlist is getting oversubscibed.  Expect reduced ROCm performance.\n");

I'm not sure this message makes sense for MES. There isn't a "runlist" in the same sense as with the old HWS. And the scheduling algorithm of MES should perform better, so maybe we don't need to warn about poor performance.

Quoted 26 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -1
  	amdgpu_mes_lock(&adev->mes);
  	r = adev->mes.funcs->add_hw_queue(&adev->mes, &queue_input);
@@ -250,7 +249,7 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
  			q->properties.doorbell_off);
  		pr_err("MES might be in unrecoverable state, issue a GPU reset\n");
  		kfd_hws_hang(dqm);
-}
+	}
 
  	return r;
  }
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -1
index ba9d69054119..139c6b58bb7e 100644
@@ -123,7 +123,7 @@ int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
  	if (!gws && pdd->qpd.num_gws == 0)
  		return -EINVAL;
 
-	if (KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 3)) {
+	if (KFD_GC_VERSION(dev) != IP_VERSION(9, 4, 3) && !dev->kfd->shared_resources.enable_mes) {
  		if (gws)
  			ret = amdgpu_amdkfd_add_gws_to_process(pdd->process->kgd_process_info,
  				gws, &mem);

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid, · line 140

Quoted 6 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +2 -1
@@ -136,7 +136,8 @@ int pqm_set_gws(struct process_queue_manager *pqm, unsigned int qid,
  	} else {
  		/*
  		 * Intentionally set GWS to a non-NULL value
-		 * for GFX 9.4.3.
+		 * for devices that do not require GWS but require the formality
+		 * of setting GWS for cooperative groups.

Change "devices that do not require GWS" to "devices that do not use GWS for global wave synchronization".

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c void pqm_uninit(struct process_queue_manager *pqm) · line 178

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +2 -1
  		 */
  		pqn->q->gws = gws ? ERR_PTR(-ENOMEM) : NULL;
  	}
@@ -173,7 +174,8 @@ void pqm_uninit(struct process_queue_manager *pqm)
 
  	list_for_each_entry_safe(pqn, next, &pqm->queues, process_queue_list) {
  		if (pqn->q && pqn->q->gws &&
-		    KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 3))
+		    KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 3) &&
+				!pqn->q->device->kfd->shared_resources.enable_mes)

The indentation looks off here.

drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid) · line 461

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +2 -1
  			amdgpu_amdkfd_remove_gws_from_process(pqm->process->kgd_process_info,
  				pqn->q->gws);
  		kfd_procfs_del_queue(pqn->q);
@@ -455,7 +457,8 @@ int pqm_destroy_queue(struct process_queue_manager *pqm, unsigned int qid)
  		}
 
  		if (pqn->q->gws) {
-			if (KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 3))
+			if (KFD_GC_VERSION(pqn->q->device) != IP_VERSION(9, 4, 3) &&
+					!dev->kfd->shared_resources.enable_mes)

Indentation.

Regards,
  Felix

Quoted 18 lines
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +0 -0
  				amdgpu_amdkfd_remove_gws_from_process(
  						pqm->process->kgd_process_info,
  						pqn->q->gws);
drivers/gpu/drm/amd/include/mes_v11_api_def.h +3 -1
index 0997e999416a..b1db2b190187 100644
@@ -275,7 +275,9 @@ union MESAPI__ADD_QUEUE {
  			uint32_t trap_en		: 1;
  			uint32_t is_aql_queue		: 1;
  			uint32_t skip_process_ctx_clear : 1;
-			uint32_t reserved		: 19;
+			uint32_t map_legacy_kq		: 1;
+			uint32_t exclusively_scheduled	: 1;
+			uint32_t reserved		: 17;
  		};
  		struct MES_API_STATUS		api_status;
  		uint64_t                        tma_addr;
author Reviewed-by Alex Deucher on v1 1/2
#3 lore raw
In reply to Jonathan Kim

[Public]

+ Ruiji Li as this is a follow up to

commit 52223c7e74d124bea47beec467e59fdfc77559fc

Author: Ruili Ji <[email protected]>
Date:   Tue Jun 6 14:06:01 2023 +0800

drm/amdkfd: To enable traps for GC_11_0_4 and up

Flag trap_en should be enabled for trap handler.

To ensure debugger is consistent with other checks.

Thanks,

Jon

Quoted 101 lines

-----Original Message-----
From: Kim, Jonathan <[email protected]>
Sent: Friday, July 14, 2023 5:38 AM
To: [email protected]
Cc: Kuehling, Felix <[email protected]>; Kim, Jonathan <[email protected]>
Subject: [PATCH 1/2] drm/amdkfd: fix trap handling work around for debugging

Update the list of devices that require the cwsr trap handling workaround for debugging use cases.

---
 drivers/gpu/drm/amd/amdkfd/kfd_debug.c                | 5 ++---
 drivers/gpu/drm/amd/amdkfd/kfd_debug.h                | 6 ++++++
 drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 6 ++----
 3 files changed, 10 insertions(+), 7 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +1 -2
index 190b03efe5ff..ccfc81f085ce 100644
@@ -302,8 +302,7 @@ static int kfd_dbg_set_queue_workaround(struct
 queue *q, bool enable)
      if (!q)
              return 0;
 
-     if (KFD_GC_VERSION(q->device) < IP_VERSION(11, 0, 0) ||
-         KFD_GC_VERSION(q->device) >= IP_VERSION(12, 0, 0))
+     if (!kfd_dbg_has_cwsr_workaround(q->device))
              return 0;
 
      if (enable && q->properties.is_user_cu_masked)
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +1 -1
@@ -349,7 +348,7 @@ int kfd_dbg_set_mes_debug_mode(struct
 kfd_process_device *pdd)
 {
      uint32_t spi_dbg_cntl = pdd->spi_dbg_override | pdd-
 >spi_dbg_launch_mode;
      uint32_t flags = pdd->process->dbg_flags;
-     bool sq_trap_en = !!spi_dbg_cntl;
+     bool sq_trap_en = !!spi_dbg_cntl ||
 !kfd_dbg_has_cwsr_workaround(pdd->dev);
      if (!kfd_dbg_is_per_vmid_supported(pdd->dev))
              return 0;
drivers/gpu/drm/amd/amdkfd/kfd_debug.h +6 -0
index ba616ed17dee..586d7f886712 100644
@@ -101,6 +101,12 @@ static inline bool
 kfd_dbg_is_rlc_restore_supported(struct kfd_node *dev)
               KFD_GC_VERSION(dev) == IP_VERSION(10, 1, 1));
 }
 
+static inline bool kfd_dbg_has_cwsr_workaround(struct kfd_node *dev)
+{
+     return KFD_GC_VERSION(dev) >= IP_VERSION(11, 0, 0) &&
+            KFD_GC_VERSION(dev) <= IP_VERSION(11, 0, 3);
+}
+
 static inline bool kfd_dbg_has_gws_support(struct kfd_node *dev)
 {
      if ((KFD_GC_VERSION(dev) == IP_VERSION(9, 0, 1)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -1
index 31cac1fd0d58..761963ad6154 100644
@@ -226,8 +226,7 @@ static int add_queue_mes(struct
 device_queue_manager *dqm, struct queue *q,
      queue_input.paging = false;
      queue_input.tba_addr = qpd->tba_addr;
      queue_input.tma_addr = qpd->tma_addr;
-     queue_input.trap_en = KFD_GC_VERSION(q->device) <

IP_VERSION(11, 0, 0) ||
- KFD_GC_VERSION(q->device) > IP_VERSION(11, 0,
3);

+     queue_input.trap_en = !kfd_dbg_has_cwsr_workaround(q->device);
      queue_input.skip_process_ctx_clear = qpd->pqm->process-
Quoted 1 line

debug_trap_enabled;

      queue_type = convert_to_mes_queue_type(q->properties.type);
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
@@ -1827,8 +1826,7 @@ static int create_queue_cpsch(struct
 device_queue_manager *dqm, struct queue *q,
       */
      q->properties.is_evicted = !!qpd->evicted;
      q->properties.is_dbg_wa = qpd->pqm->process-
Quoted 1 line

debug_trap_enabled &&

- KFD_GC_VERSION(q->device) >= IP_VERSION(11, 0, 0)
&&

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +1 -1
-                     KFD_GC_VERSION(q->device) <= IP_VERSION(11, 0, 3);
+                               kfd_dbg_has_cwsr_workaround(q->device);
 
      if (qd)
              mqd_mgr->restore_mqd(mqd_mgr, &q->mqd, q-
Quoted 1 line

mqd_mem_obj, &q->gart_mqd_addr,

--
2.25.1

Reviewed-by Alex Deucher
#4 lore raw

[Public]

Hi Jonathan,

Your change looks fine.

Thanks,
Ruili
-----Original Message-----
From: Kim, Jonathan <[email protected]>
Sent: Wednesday, July 19, 2023 6:13 AM
To: [email protected]
Cc: Kuehling, Felix <[email protected]>; Ji, Ruili <[email protected]>
Subject: RE: [PATCH 1/2] drm/amdkfd: fix trap handling work around for debugging

[Public]

+ Ruiji Li as this is a follow up to

commit 52223c7e74d124bea47beec467e59fdfc77559fc

Author: Ruili Ji <[email protected]>
Date:   Tue Jun 6 14:06:01 2023 +0800

drm/amdkfd: To enable traps for GC_11_0_4 and up

Flag trap_en should be enabled for trap handler.

To ensure debugger is consistent with other checks.

Thanks,

Jon

Quoted 97 lines

-----Original Message-----
From: Kim, Jonathan <[email protected]>
Sent: Friday, July 14, 2023 5:38 AM
To: [email protected]
Cc: Kuehling, Felix <[email protected]>; Kim, Jonathan <[email protected]>
Subject: [PATCH 1/2] drm/amdkfd: fix trap handling work around for debugging

Update the list of devices that require the cwsr trap handling workaround for debugging use cases.

---
 drivers/gpu/drm/amd/amdkfd/kfd_debug.c                | 5 ++---
 drivers/gpu/drm/amd/amdkfd/kfd_debug.h                | 6 ++++++
 drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 6 ++----
 3 files changed, 10 insertions(+), 7 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +1 -2
index 190b03efe5ff..ccfc81f085ce 100644
@@ -302,8 +302,7 @@ static int kfd_dbg_set_queue_workaround(struct
 queue *q, bool enable)
      if (!q)
              return 0;
 
-     if (KFD_GC_VERSION(q->device) < IP_VERSION(11, 0, 0) ||
-         KFD_GC_VERSION(q->device) >= IP_VERSION(12, 0, 0))
+     if (!kfd_dbg_has_cwsr_workaround(q->device))
              return 0;
 
      if (enable && q->properties.is_user_cu_masked) @@ -349,7 +348,7
@@ int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd)  {
      uint32_t spi_dbg_cntl = pdd->spi_dbg_override | pdd-
Quoted 1 line

spi_dbg_launch_mode;

      uint32_t flags = pdd->process->dbg_flags;
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +1 -1
-     bool sq_trap_en = !!spi_dbg_cntl;
+     bool sq_trap_en = !!spi_dbg_cntl ||

!kfd_dbg_has_cwsr_workaround(pdd->dev);

      if (!kfd_dbg_is_per_vmid_supported(pdd->dev))
              return 0;
drivers/gpu/drm/amd/amdkfd/kfd_debug.h +5 -0
index ba616ed17dee..586d7f886712 100644
@@ -101,6 +101,12 @@ static inline bool
 kfd_dbg_is_rlc_restore_supported(struct kfd_node *dev)
               KFD_GC_VERSION(dev) == IP_VERSION(10, 1, 1));  }
 
+static inline bool kfd_dbg_has_cwsr_workaround(struct kfd_node *dev)
+{
+     return KFD_GC_VERSION(dev) >= IP_VERSION(11, 0, 0) &&
+            KFD_GC_VERSION(dev) <= IP_VERSION(11, 0, 3); }
+
 static inline bool kfd_dbg_has_gws_support(struct kfd_node *dev)  {
      if ((KFD_GC_VERSION(dev) == IP_VERSION(9, 0, 1) diff --git

a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c index 31cac1fd0d58..761963ad6154 100644

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -1
@@ -226,8 +226,7 @@ static int add_queue_mes(struct
 device_queue_manager *dqm, struct queue *q,
      queue_input.paging = false;
      queue_input.tba_addr = qpd->tba_addr;
      queue_input.tma_addr = qpd->tma_addr;
-     queue_input.trap_en = KFD_GC_VERSION(q->device) <

IP_VERSION(11, 0, 0) ||
- KFD_GC_VERSION(q->device) > IP_VERSION(11, 0,
3);

+     queue_input.trap_en = !kfd_dbg_has_cwsr_workaround(q->device);
      queue_input.skip_process_ctx_clear = qpd->pqm->process-
Quoted 1 line

debug_trap_enabled;

      queue_type = convert_to_mes_queue_type(q->properties.type);
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0 -0
@@ -1827,8 +1826,7 @@ static int create_queue_cpsch(struct
 device_queue_manager *dqm, struct queue *q,
       */
      q->properties.is_evicted = !!qpd->evicted;
      q->properties.is_dbg_wa = qpd->pqm->process-
Quoted 1 line

debug_trap_enabled &&

- KFD_GC_VERSION(q->device) >= IP_VERSION(11, 0, 0)
&&

drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -1
-                     KFD_GC_VERSION(q->device) <= IP_VERSION(11, 0, 3);
+
+ kfd_dbg_has_cwsr_workaround(q->device);
 
      if (qd)
              mqd_mgr->restore_mqd(mqd_mgr, &q->mqd, q-
Quoted 1 line

mqd_mem_obj, &q->gart_mqd_addr,

--
2.25.1

maintainer Reviewed-by on v1 1/2
#5 lore raw
In reply to Jonathan Kim

Am 2023-07-14 um 05:37 schrieb Jonathan Kim:

Quoted 4 lines

Update the list of devices that require the cwsr trap handling workaround for debugging use cases.

This patch is

Quoted 70 lines
---
  drivers/gpu/drm/amd/amdkfd/kfd_debug.c                | 5 ++---
  drivers/gpu/drm/amd/amdkfd/kfd_debug.h                | 6 ++++++
  drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 6 ++----
  3 files changed, 10 insertions(+), 7 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_debug.c +2 -3
index 190b03efe5ff..ccfc81f085ce 100644
@@ -302,8 +302,7 @@ static int kfd_dbg_set_queue_workaround(struct queue *q, bool enable)
  	if (!q)
  		return 0;
 
-	if (KFD_GC_VERSION(q->device) < IP_VERSION(11, 0, 0) ||
-	    KFD_GC_VERSION(q->device) >= IP_VERSION(12, 0, 0))
+	if (!kfd_dbg_has_cwsr_workaround(q->device))
  		return 0;
 
  	if (enable && q->properties.is_user_cu_masked)
@@ -349,7 +348,7 @@ int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd)
  {
  	uint32_t spi_dbg_cntl = pdd->spi_dbg_override | pdd->spi_dbg_launch_mode;
  	uint32_t flags = pdd->process->dbg_flags;
-	bool sq_trap_en = !!spi_dbg_cntl;
+	bool sq_trap_en = !!spi_dbg_cntl || !kfd_dbg_has_cwsr_workaround(pdd->dev);
 
  	if (!kfd_dbg_is_per_vmid_supported(pdd->dev))
  		return 0;
drivers/gpu/drm/amd/amdkfd/kfd_debug.h +6 -0
index ba616ed17dee..586d7f886712 100644
@@ -101,6 +101,12 @@ static inline bool kfd_dbg_is_rlc_restore_supported(struct kfd_node *dev)
  		 KFD_GC_VERSION(dev) == IP_VERSION(10, 1, 1));
  }
 
+static inline bool kfd_dbg_has_cwsr_workaround(struct kfd_node *dev)
+{
+	return KFD_GC_VERSION(dev) >= IP_VERSION(11, 0, 0) &&
+	       KFD_GC_VERSION(dev) <= IP_VERSION(11, 0, 3);
+}
+
  static inline bool kfd_dbg_has_gws_support(struct kfd_node *dev)
  {
  	if ((KFD_GC_VERSION(dev) == IP_VERSION(9, 0, 1)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +2 -4
index 31cac1fd0d58..761963ad6154 100644
@@ -226,8 +226,7 @@ static int add_queue_mes(struct device_queue_manager *dqm, struct queue *q,
  	queue_input.paging = false;
  	queue_input.tba_addr = qpd->tba_addr;
  	queue_input.tma_addr = qpd->tma_addr;
-	queue_input.trap_en = KFD_GC_VERSION(q->device) < IP_VERSION(11, 0, 0) ||
-			      KFD_GC_VERSION(q->device) > IP_VERSION(11, 0, 3);
+	queue_input.trap_en = !kfd_dbg_has_cwsr_workaround(q->device);
  	queue_input.skip_process_ctx_clear = qpd->pqm->process->debug_trap_enabled;
 
  	queue_type = convert_to_mes_queue_type(q->properties.type);
@@ -1827,8 +1826,7 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
  	 */
  	q->properties.is_evicted = !!qpd->evicted;
  	q->properties.is_dbg_wa = qpd->pqm->process->debug_trap_enabled &&
-			KFD_GC_VERSION(q->device) >= IP_VERSION(11, 0, 0) &&
-			KFD_GC_VERSION(q->device) <= IP_VERSION(11, 0, 3);
+				  kfd_dbg_has_cwsr_workaround(q->device);
 
  	if (qd)
  		mqd_mgr->restore_mqd(mqd_mgr, &q->mqd, q->mqd_mem_obj, &q->gart_mqd_addr,