Skip to contentSkip to conversation

Raven support for KFD

PATCH In review Felix Kuehling 3 participants 15 messages Jul 12, 2018 – Jul 13, 2018 8 messages not in the archive

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

cover Raven support for KFD

Raven refers to Ryzen APUs with integrated GFXv9 GPU. This patch series completes Raven support for KFD:

* fix up memory banks info from CRAT
* support different number of SDMA engines
* workaround IOMMUv2 PPR issues
* add device info

Yong Zhao (6):
drm/amdkfd: Consolidate duplicate memory banks info in topology drm/amdkfd: Make SDMA engine number an ASIC-dependent variable drm/amdkfd: Avoid flooding dmesg on Raven due to IOMMU issues drm/amdkfd: Workaround to accommodate Raven too many PPR issue drm/amdkfd: Optimize out some duplicated code in
kfd_signal_iommu_event()
drm/amdkfd: Enable Raven for KFD

 drivers/gpu/drm/amd/amdkfd/kfd_crat.c              | 57 +++++++++++++++++-----
 drivers/gpu/drm/amd/amdkfd/kfd_device.c            | 28 +++++++++++
 .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c  | 29 +++++++----
 .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h  |  6 +--
 drivers/gpu/drm/amd/amdkfd/kfd_events.c            | 47 ++++++++++--------
 drivers/gpu/drm/amd/amdkfd/kfd_iommu.c             |  2 +-
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h              |  1 +
 .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c |  3 +-
 8 files changed, 126 insertions(+), 47 deletions(-)
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1531430694-23966-1-git-send-email-Felix.Kuehling-5C7GfCeVMHo@public.gmane.org

#2 lore raw

PATCH 1/6 drm/amdkfd: Consolidate duplicate memory banks info in topology

From: Yong Zhao <[email protected]>

If there are several memory banks that has the same properties in CRAT, we aggregate them into one memory bank. This cleans up memory banks on APUs (e.g. Raven) where the CRAT reports each memory channel as a separate bank. This only confuses user mode, which only deals with virtual memory.

---
 drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 57 ++++++++++++++++++++++++++++-------
 1 file changed, 46 insertions(+), 11 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_crat.c +46 -11
index 296b3f2..ee49960 100644
@@ -189,6 +189,21 @@ static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu,
 	return 0;
 }
 
+static struct kfd_mem_properties *
+find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width,
+		struct kfd_topology_device *dev)
+{
+	struct kfd_mem_properties *props;
+
+	list_for_each_entry(props, &dev->mem_props, list) {
+		if (props->heap_type == heap_type
+				&& props->flags == flags
+				&& props->width == width)
+			return props;
+	}
+
+	return NULL;
+}
 /* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct
  * topology device present in the device_list
  */
@@ -197,36 +212,56 @@ static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem,
 {
 	struct kfd_mem_properties *props;
 	struct kfd_topology_device *dev;
+	uint32_t heap_type;
+	uint64_t size_in_bytes;
+	uint32_t flags = 0;
+	uint32_t width;
 
 	pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n",
 			mem->proximity_domain);
 	list_for_each_entry(dev, device_list, list) {
 		if (mem->proximity_domain == dev->proximity_domain) {
-			props = kfd_alloc_struct(props);
-			if (!props)
-				return -ENOMEM;
-
 			/* We're on GPU node */
 			if (dev->node_props.cpu_cores_count == 0) {
 				/* APU */
 				if (mem->visibility_type == 0)
-					props->heap_type =
+					heap_type =
 						HSA_MEM_HEAP_TYPE_FB_PRIVATE;
 				/* dGPU */
 				else
-					props->heap_type = mem->visibility_type;
+					heap_type = mem->visibility_type;
 			} else
-				props->heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
+				heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
 
 			if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
-				props->flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
+				flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
 			if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
-				props->flags |= HSA_MEM_FLAGS_NON_VOLATILE;
+				flags |= HSA_MEM_FLAGS_NON_VOLATILE;
 
-			props->size_in_bytes =
+			size_in_bytes =
 				((uint64_t)mem->length_high << 32) +
 							mem->length_low;
-			props->width = mem->width;
+			width = mem->width;
+
+			/* Multiple banks of the same type are aggregated into
+			 * one. User mode doesn't care about multiple physical
+			 * memory segments. It's managed as a single virtual
+			 * heap for user mode.
+			 */
+			props = find_subtype_mem(heap_type, flags, width, dev);
+			if (props) {
+				props->size_in_bytes += size_in_bytes;
+				break;
+			}
+
+			props = kfd_alloc_struct(props);
+			if (!props)
+				return -ENOMEM;
+
+			props->heap_type = heap_type;
+			props->flags = flags;
+			props->size_in_bytes = size_in_bytes;
+			props->width = width;
 
 			dev->node_props.mem_banks_count++;
 			list_add_tail(&props->list, &dev->mem_props);
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1531430694-23966-2-git-send-email-Felix.Kuehling-5C7GfCeVMHo@public.gmane.org

Acked-by
#4 lore raw

Re: [PATCH 1/6] drm/amdkfd: Consolidate duplicate memory banks info in topology

Quoted 11 lines · Felix Kuehling

From: Yong Zhao <[email protected]>

If there are several memory banks that has the same properties in CRAT, we aggregate them into one memory bank. This cleans up memory banks on APUs (e.g. Raven) where the CRAT reports each memory channel as a separate bank. This only confuses user mode, which only deals with virtual memory.

Quoted 101 lines
---
 drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 57 ++++++++++++++++++++++++++++-------
 1 file changed, 46 insertions(+), 11 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_crat.c +46 -12
index 296b3f2..ee49960 100644
@@ -189,6 +189,21 @@ static int kfd_parse_subtype_cu(struct crat_subtype_computeunit *cu,
        return 0;
 }
 
+static struct kfd_mem_properties *
+find_subtype_mem(uint32_t heap_type, uint32_t flags, uint32_t width,
+               struct kfd_topology_device *dev)
+{
+       struct kfd_mem_properties *props;
+
+       list_for_each_entry(props, &dev->mem_props, list) {
+               if (props->heap_type == heap_type
+                               && props->flags == flags
+                               && props->width == width)
+                       return props;
+       }
+
+       return NULL;
+}
 /* kfd_parse_subtype_mem - parse memory subtypes and attach it to correct
  * topology device present in the device_list
  */
@@ -197,36 +212,56 @@ static int kfd_parse_subtype_mem(struct crat_subtype_memory *mem,
 {
        struct kfd_mem_properties *props;
        struct kfd_topology_device *dev;
+       uint32_t heap_type;
+       uint64_t size_in_bytes;
+       uint32_t flags = 0;
+       uint32_t width;
 
        pr_debug("Found memory entry in CRAT table with proximity_domain=%d\n",
                        mem->proximity_domain);
        list_for_each_entry(dev, device_list, list) {
                if (mem->proximity_domain == dev->proximity_domain) {
-                       props = kfd_alloc_struct(props);
-                       if (!props)
-                               return -ENOMEM;
-
                        /* We're on GPU node */
                        if (dev->node_props.cpu_cores_count == 0) {
                                /* APU */
                                if (mem->visibility_type == 0)
-                                       props->heap_type =
+                                       heap_type =
                                                HSA_MEM_HEAP_TYPE_FB_PRIVATE;
                                /* dGPU */
                                else
-                                       props->heap_type = mem->visibility_type;
+                                       heap_type = mem->visibility_type;
                        } else
-                               props->heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
+                               heap_type = HSA_MEM_HEAP_TYPE_SYSTEM;
 
                        if (mem->flags & CRAT_MEM_FLAGS_HOT_PLUGGABLE)
-                               props->flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
+                               flags |= HSA_MEM_FLAGS_HOT_PLUGGABLE;
                        if (mem->flags & CRAT_MEM_FLAGS_NON_VOLATILE)
-                               props->flags |= HSA_MEM_FLAGS_NON_VOLATILE;
+                               flags |= HSA_MEM_FLAGS_NON_VOLATILE;
 
-                       props->size_in_bytes =
+                       size_in_bytes =
                                ((uint64_t)mem->length_high << 32) +
                                                        mem->length_low;
-                       props->width = mem->width;
+                       width = mem->width;
+
+                       /* Multiple banks of the same type are aggregated into
+                        * one. User mode doesn't care about multiple physical
+                        * memory segments. It's managed as a single virtual
+                        * heap for user mode.
+                        */
+                       props = find_subtype_mem(heap_type, flags, width, dev);
+                       if (props) {
+                               props->size_in_bytes += size_in_bytes;
+                               break;
+                       }
+
+                       props = kfd_alloc_struct(props);
+                       if (!props)
+                               return -ENOMEM;
+
+                       props->heap_type = heap_type;
+                       props->flags = flags;
+                       props->size_in_bytes = size_in_bytes;
+                       props->width = width;
 
                        dev->node_props.mem_banks_count++;
                        list_add_tail(&props->list, &dev->mem_props);
--

2.7.4

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#5 lore raw

PATCH 2/6 drm/amdkfd: Make SDMA engine number an ASIC-dependent variable

In reply to a message not in the archive

From: Yong Zhao <[email protected]>

On Raven there is only one SDMA engine instead of previously assumed two, so we need to adapt our code to this new scenario.

---
 drivers/gpu/drm/amd/amdkfd/kfd_device.c            | 12 +++++++++
 .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c  | 29 +++++++++++++++-------
 .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h  |  6 ++---
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h              |  1 +
 .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c |  3 +--
 5 files changed, 36 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +12 -0
index 8faa8db..572235c 100644
@@ -52,6 +52,7 @@ static const struct kfd_device_info kaveri_device_info = {
 	.supports_cwsr = false,
 	.needs_iommu_device = true,
 	.needs_pci_atomics = false,
+	.num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info carrizo_device_info = {
@@ -67,6 +68,7 @@ static const struct kfd_device_info carrizo_device_info = {
 	.supports_cwsr = true,
 	.needs_iommu_device = true,
 	.needs_pci_atomics = false,
+	.num_sdma_engines = 2,
 };
 #endif
 
@@ -83,6 +85,7 @@ static const struct kfd_device_info hawaii_device_info = {
 	.supports_cwsr = false,
 	.needs_iommu_device = false,
 	.needs_pci_atomics = false,
+	.num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info tonga_device_info = {
@@ -97,6 +100,7 @@ static const struct kfd_device_info tonga_device_info = {
 	.supports_cwsr = false,
 	.needs_iommu_device = false,
 	.needs_pci_atomics = true,
+	.num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info tonga_vf_device_info = {
@@ -111,6 +115,7 @@ static const struct kfd_device_info tonga_vf_device_info = {
 	.supports_cwsr = false,
 	.needs_iommu_device = false,
 	.needs_pci_atomics = false,
+	.num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info fiji_device_info = {
@@ -125,6 +130,7 @@ static const struct kfd_device_info fiji_device_info = {
 	.supports_cwsr = true,
 	.needs_iommu_device = false,
 	.needs_pci_atomics = true,
+	.num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info fiji_vf_device_info = {
@@ -139,6 +145,7 @@ static const struct kfd_device_info fiji_vf_device_info = {
 	.supports_cwsr = true,
 	.needs_iommu_device = false,
 	.needs_pci_atomics = false,
+	.num_sdma_engines = 2,
 };
 
 
@@ -154,6 +161,7 @@ static const struct kfd_device_info polaris10_device_info = {
 	.supports_cwsr = true,
 	.needs_iommu_device = false,
 	.needs_pci_atomics = true,
+	.num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info polaris10_vf_device_info = {
@@ -168,6 +176,7 @@ static const struct kfd_device_info polaris10_vf_device_info = {
 	.supports_cwsr = true,
 	.needs_iommu_device = false,
 	.needs_pci_atomics = false,
+	.num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info polaris11_device_info = {
@@ -182,6 +191,7 @@ static const struct kfd_device_info polaris11_device_info = {
 	.supports_cwsr = true,
 	.needs_iommu_device = false,
 	.needs_pci_atomics = true,
+	.num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info vega10_device_info = {
@@ -196,6 +206,7 @@ static const struct kfd_device_info vega10_device_info = {
 	.supports_cwsr = true,
 	.needs_iommu_device = false,
 	.needs_pci_atomics = false,
+	.num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info vega10_vf_device_info = {
@@ -210,6 +221,7 @@ static const struct kfd_device_info vega10_vf_device_info = {
 	.supports_cwsr = true,
 	.needs_iommu_device = false,
 	.needs_pci_atomics = false,
+	.num_sdma_engines = 2,
 };
 
 
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +20 -9
index 97c9f10..ace94d6 100644
@@ -101,6 +101,17 @@ unsigned int get_pipes_per_mec(struct device_queue_manager *dqm)
 	return dqm->dev->shared_resources.num_pipe_per_mec;
 }
 
+static unsigned int get_num_sdma_engines(struct device_queue_manager *dqm)
+{
+	return dqm->dev->device_info->num_sdma_engines;
+}
+
+unsigned int get_num_sdma_queues(struct device_queue_manager *dqm)
+{
+	return dqm->dev->device_info->num_sdma_engines
+			* KFD_SDMA_QUEUES_PER_ENGINE;
+}
+
 void program_sh_mem_settings(struct device_queue_manager *dqm,
 					struct qcm_process_device *qpd)
 {
@@ -855,7 +866,7 @@ static int initialize_nocpsch(struct device_queue_manager *dqm)
 	}
 
 	dqm->vmid_bitmap = (1 << dqm->dev->vm_info.vmid_num_kfd) - 1;
-	dqm->sdma_bitmap = (1 << CIK_SDMA_QUEUES) - 1;
+	dqm->sdma_bitmap = (1 << get_num_sdma_queues(dqm)) - 1;
 
 	return 0;
 }
@@ -903,7 +914,7 @@ static int allocate_sdma_queue(struct device_queue_manager *dqm,
 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
 				unsigned int sdma_queue_id)
 {
-	if (sdma_queue_id >= CIK_SDMA_QUEUES)
+	if (sdma_queue_id >= get_num_sdma_queues(dqm))
 		return;
 	dqm->sdma_bitmap |= (1 << sdma_queue_id);
 }
@@ -923,8 +934,8 @@ static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
 	if (retval)
 		return retval;
 
-	q->properties.sdma_queue_id = q->sdma_id / CIK_SDMA_QUEUES_PER_ENGINE;
-	q->properties.sdma_engine_id = q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
+	q->properties.sdma_queue_id = q->sdma_id / get_num_sdma_engines(dqm);
+	q->properties.sdma_engine_id = q->sdma_id % get_num_sdma_engines(dqm);
 
 	retval = allocate_doorbell(qpd, q);
 	if (retval)
@@ -1011,7 +1022,7 @@ static int initialize_cpsch(struct device_queue_manager *dqm)
 	dqm->queue_count = dqm->processes_count = 0;
 	dqm->sdma_queue_count = 0;
 	dqm->active_runlist = false;
-	dqm->sdma_bitmap = (1 << CIK_SDMA_QUEUES) - 1;
+	dqm->sdma_bitmap = (1 << get_num_sdma_queues(dqm)) - 1;
 
 	INIT_WORK(&dqm->hw_exception_work, kfd_process_hw_exception);
 
@@ -1142,9 +1153,9 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
 		if (retval)
 			goto out_unlock;
 		q->properties.sdma_queue_id =
-			q->sdma_id / CIK_SDMA_QUEUES_PER_ENGINE;
+			q->sdma_id / get_num_sdma_engines(dqm);
 		q->properties.sdma_engine_id =
-			q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
+			q->sdma_id % get_num_sdma_engines(dqm);
 	}
 
 	retval = allocate_doorbell(qpd, q);
@@ -1791,8 +1802,8 @@ int dqm_debugfs_hqds(struct seq_file *m, void *data)
 		}
 	}
 
-	for (pipe = 0; pipe < CIK_SDMA_ENGINE_NUM; pipe++) {
-		for (queue = 0; queue < CIK_SDMA_QUEUES_PER_ENGINE; queue++) {
+	for (pipe = 0; pipe < get_num_sdma_engines(dqm); pipe++) {
+		for (queue = 0; queue < KFD_SDMA_QUEUES_PER_ENGINE; queue++) {
 			r = dqm->dev->kfd2kgd->hqd_sdma_dump(
 				dqm->dev->kgd, pipe, queue, &dump, &n_regs);
 			if (r)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +2 -4
index 52e708c..00da316 100644
@@ -33,10 +33,7 @@
 
 #define KFD_UNMAP_LATENCY_MS			(4000)
 #define QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS (2 * KFD_UNMAP_LATENCY_MS + 1000)
-
-#define CIK_SDMA_QUEUES				(4)
-#define CIK_SDMA_QUEUES_PER_ENGINE		(2)
-#define CIK_SDMA_ENGINE_NUM			(2)
+#define KFD_SDMA_QUEUES_PER_ENGINE		(2)
 
 struct device_process_node {
 	struct qcm_process_device *qpd;
@@ -214,6 +211,7 @@ void program_sh_mem_settings(struct device_queue_manager *dqm,
 unsigned int get_queues_num(struct device_queue_manager *dqm);
 unsigned int get_queues_per_pipe(struct device_queue_manager *dqm);
 unsigned int get_pipes_per_mec(struct device_queue_manager *dqm);
+unsigned int get_num_sdma_queues(struct device_queue_manager *dqm);
 
 static inline unsigned int get_sh_mem_bases_32(struct kfd_process_device *pdd)
 {
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1 -0
index 37d179e..ca83254 100644
@@ -203,6 +203,7 @@ struct kfd_device_info {
 	bool supports_cwsr;
 	bool needs_iommu_device;
 	bool needs_pci_atomics;
+	unsigned int num_sdma_engines;
 };
 
 struct kfd_mem_obj {
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -2
index 1303b14..eb4e5fb 100644
@@ -186,8 +186,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
 
 	switch (type) {
 	case KFD_QUEUE_TYPE_SDMA:
-		if (dev->dqm->queue_count >=
-			CIK_SDMA_QUEUES_PER_ENGINE * CIK_SDMA_ENGINE_NUM) {
+		if (dev->dqm->queue_count >= get_num_sdma_queues(dev->dqm)) {
 			pr_err("Over-subscription is not allowed for SDMA.\n");
 			retval = -EPERM;
 			goto err_create_queue;
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1531430694-23966-3-git-send-email-Felix.Kuehling-5C7GfCeVMHo@public.gmane.org

Acked-by
#7 lore raw

Re: [PATCH 2/6] drm/amdkfd: Make SDMA engine number an ASIC-dependent variable

Quoted 8 lines · Felix Kuehling

From: Yong Zhao <[email protected]>

On Raven there is only one SDMA engine instead of previously assumed two, so we need to adapt our code to this new scenario.

Quoted 244 lines
---
 drivers/gpu/drm/amd/amdkfd/kfd_device.c            | 12 +++++++++
 .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c  | 29 +++++++++++++++-------
 .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h  |  6 ++---
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h              |  1 +
 .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c |  3 +--
 5 files changed, 36 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +12 -0
index 8faa8db..572235c 100644
@@ -52,6 +52,7 @@ static const struct kfd_device_info kaveri_device_info = {
        .supports_cwsr = false,
        .needs_iommu_device = true,
        .needs_pci_atomics = false,
+       .num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info carrizo_device_info = {
@@ -67,6 +68,7 @@ static const struct kfd_device_info carrizo_device_info = {
        .supports_cwsr = true,
        .needs_iommu_device = true,
        .needs_pci_atomics = false,
+       .num_sdma_engines = 2,
 };
 #endif
 
@@ -83,6 +85,7 @@ static const struct kfd_device_info hawaii_device_info = {
        .supports_cwsr = false,
        .needs_iommu_device = false,
        .needs_pci_atomics = false,
+       .num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info tonga_device_info = {
@@ -97,6 +100,7 @@ static const struct kfd_device_info tonga_device_info = {
        .supports_cwsr = false,
        .needs_iommu_device = false,
        .needs_pci_atomics = true,
+       .num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info tonga_vf_device_info = {
@@ -111,6 +115,7 @@ static const struct kfd_device_info tonga_vf_device_info = {
        .supports_cwsr = false,
        .needs_iommu_device = false,
        .needs_pci_atomics = false,
+       .num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info fiji_device_info = {
@@ -125,6 +130,7 @@ static const struct kfd_device_info fiji_device_info = {
        .supports_cwsr = true,
        .needs_iommu_device = false,
        .needs_pci_atomics = true,
+       .num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info fiji_vf_device_info = {
@@ -139,6 +145,7 @@ static const struct kfd_device_info fiji_vf_device_info = {
        .supports_cwsr = true,
        .needs_iommu_device = false,
        .needs_pci_atomics = false,
+       .num_sdma_engines = 2,
 };
 
 
@@ -154,6 +161,7 @@ static const struct kfd_device_info polaris10_device_info = {
        .supports_cwsr = true,
        .needs_iommu_device = false,
        .needs_pci_atomics = true,
+       .num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info polaris10_vf_device_info = {
@@ -168,6 +176,7 @@ static const struct kfd_device_info polaris10_vf_device_info = {
        .supports_cwsr = true,
        .needs_iommu_device = false,
        .needs_pci_atomics = false,
+       .num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info polaris11_device_info = {
@@ -182,6 +191,7 @@ static const struct kfd_device_info polaris11_device_info = {
        .supports_cwsr = true,
        .needs_iommu_device = false,
        .needs_pci_atomics = true,
+       .num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info vega10_device_info = {
@@ -196,6 +206,7 @@ static const struct kfd_device_info vega10_device_info = {
        .supports_cwsr = true,
        .needs_iommu_device = false,
        .needs_pci_atomics = false,
+       .num_sdma_engines = 2,
 };
 
 static const struct kfd_device_info vega10_vf_device_info = {
@@ -210,6 +221,7 @@ static const struct kfd_device_info vega10_vf_device_info = {
        .supports_cwsr = true,
        .needs_iommu_device = false,
        .needs_pci_atomics = false,
+       .num_sdma_engines = 2,
 };
 
 
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +20 -9
index 97c9f10..ace94d6 100644
@@ -101,6 +101,17 @@ unsigned int get_pipes_per_mec(struct device_queue_manager *dqm)
        return dqm->dev->shared_resources.num_pipe_per_mec;
 }
 
+static unsigned int get_num_sdma_engines(struct device_queue_manager *dqm)
+{
+       return dqm->dev->device_info->num_sdma_engines;
+}
+
+unsigned int get_num_sdma_queues(struct device_queue_manager *dqm)
+{
+       return dqm->dev->device_info->num_sdma_engines
+                       * KFD_SDMA_QUEUES_PER_ENGINE;
+}
+
 void program_sh_mem_settings(struct device_queue_manager *dqm,
                                        struct qcm_process_device *qpd)
 {
@@ -855,7 +866,7 @@ static int initialize_nocpsch(struct device_queue_manager *dqm)
        }
 
        dqm->vmid_bitmap = (1 << dqm->dev->vm_info.vmid_num_kfd) - 1;
-       dqm->sdma_bitmap = (1 << CIK_SDMA_QUEUES) - 1;
+       dqm->sdma_bitmap = (1 << get_num_sdma_queues(dqm)) - 1;
 
        return 0;
 }
@@ -903,7 +914,7 @@ static int allocate_sdma_queue(struct device_queue_manager *dqm,
 static void deallocate_sdma_queue(struct device_queue_manager *dqm,
                                unsigned int sdma_queue_id)
 {
-       if (sdma_queue_id >= CIK_SDMA_QUEUES)
+       if (sdma_queue_id >= get_num_sdma_queues(dqm))
                return;
        dqm->sdma_bitmap |= (1 << sdma_queue_id);
 }
@@ -923,8 +934,8 @@ static int create_sdma_queue_nocpsch(struct device_queue_manager *dqm,
        if (retval)
                return retval;
 
-       q->properties.sdma_queue_id = q->sdma_id / CIK_SDMA_QUEUES_PER_ENGINE;
-       q->properties.sdma_engine_id = q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
+       q->properties.sdma_queue_id = q->sdma_id / get_num_sdma_engines(dqm);
+       q->properties.sdma_engine_id = q->sdma_id % get_num_sdma_engines(dqm);
 
        retval = allocate_doorbell(qpd, q);
        if (retval)
@@ -1011,7 +1022,7 @@ static int initialize_cpsch(struct device_queue_manager *dqm)
        dqm->queue_count = dqm->processes_count = 0;
        dqm->sdma_queue_count = 0;
        dqm->active_runlist = false;
-       dqm->sdma_bitmap = (1 << CIK_SDMA_QUEUES) - 1;
+       dqm->sdma_bitmap = (1 << get_num_sdma_queues(dqm)) - 1;
 
        INIT_WORK(&dqm->hw_exception_work, kfd_process_hw_exception);
 
@@ -1142,9 +1153,9 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,
                if (retval)
                        goto out_unlock;
                q->properties.sdma_queue_id =
-                       q->sdma_id / CIK_SDMA_QUEUES_PER_ENGINE;
+                       q->sdma_id / get_num_sdma_engines(dqm);
                q->properties.sdma_engine_id =
-                       q->sdma_id % CIK_SDMA_QUEUES_PER_ENGINE;
+                       q->sdma_id % get_num_sdma_engines(dqm);
        }
 
        retval = allocate_doorbell(qpd, q);
@@ -1791,8 +1802,8 @@ int dqm_debugfs_hqds(struct seq_file *m, void *data)
                }
        }
 
-       for (pipe = 0; pipe < CIK_SDMA_ENGINE_NUM; pipe++) {
-               for (queue = 0; queue < CIK_SDMA_QUEUES_PER_ENGINE; queue++) {
+       for (pipe = 0; pipe < get_num_sdma_engines(dqm); pipe++) {
+               for (queue = 0; queue < KFD_SDMA_QUEUES_PER_ENGINE; queue++) {
                        r = dqm->dev->kfd2kgd->hqd_sdma_dump(
                                dqm->dev->kgd, pipe, queue, &dump, &n_regs);
                        if (r)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.h +2 -4
index 52e708c..00da316 100644
@@ -33,10 +33,7 @@
 
 #define KFD_UNMAP_LATENCY_MS                   (4000)
 #define QUEUE_PREEMPT_DEFAULT_TIMEOUT_MS (2 * KFD_UNMAP_LATENCY_MS + 1000)
-
-#define CIK_SDMA_QUEUES                                (4)
-#define CIK_SDMA_QUEUES_PER_ENGINE             (2)
-#define CIK_SDMA_ENGINE_NUM                    (2)
+#define KFD_SDMA_QUEUES_PER_ENGINE             (2)
 
 struct device_process_node {
        struct qcm_process_device *qpd;
@@ -214,6 +211,7 @@ void program_sh_mem_settings(struct device_queue_manager *dqm,
 unsigned int get_queues_num(struct device_queue_manager *dqm);
 unsigned int get_queues_per_pipe(struct device_queue_manager *dqm);
 unsigned int get_pipes_per_mec(struct device_queue_manager *dqm);
+unsigned int get_num_sdma_queues(struct device_queue_manager *dqm);
 
 static inline unsigned int get_sh_mem_bases_32(struct kfd_process_device *pdd)
 {
drivers/gpu/drm/amd/amdkfd/kfd_priv.h +1 -0
index 37d179e..ca83254 100644
@@ -203,6 +203,7 @@ struct kfd_device_info {
        bool supports_cwsr;
        bool needs_iommu_device;
        bool needs_pci_atomics;
+       unsigned int num_sdma_engines;
 };
 
 struct kfd_mem_obj {
drivers/gpu/drm/amd/amdkfd/kfd_process_queue_manager.c +1 -3
index 1303b14..eb4e5fb 100644
@@ -186,8 +186,7 @@ int pqm_create_queue(struct process_queue_manager *pqm,
 
        switch (type) {
        case KFD_QUEUE_TYPE_SDMA:
-               if (dev->dqm->queue_count >=
-                       CIK_SDMA_QUEUES_PER_ENGINE * CIK_SDMA_ENGINE_NUM) {
+               if (dev->dqm->queue_count >= get_num_sdma_queues(dev->dqm)) {
                        pr_err("Over-subscription is not allowed for SDMA.\n");
                        retval = -EPERM;
                        goto err_create_queue;
--

2.7.4

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#8 lore raw

PATCH 3/6 drm/amdkfd: Avoid flooding dmesg on Raven due to IOMMU issues

In reply to a message not in the archive

From: Yong Zhao <[email protected]>

On Raven Invalid PPRs can be reported because multiple PPRs can be still queued when memory is freed. Apply a rate limit to avoid flooding the log in this case.

---
 drivers/gpu/drm/amd/amdkfd/kfd_iommu.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_iommu.c +1 -1
index c718179..7a61f38 100644
@@ -190,7 +190,7 @@ static int iommu_invalid_ppr_cb(struct pci_dev *pdev, int pasid,
 {
 	struct kfd_dev *dev;
 
-	dev_warn(kfd_device,
+	dev_warn_ratelimited(kfd_device,
 			"Invalid PPR device %x:%x.%x pasid %d address 0x%lX flags 0x%X",
 			PCI_BUS_NUM(pdev->devfn),
 			PCI_SLOT(pdev->devfn),
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1531430694-23966-4-git-send-email-Felix.Kuehling-5C7GfCeVMHo@public.gmane.org

Acked-by
#10 lore raw

Re: [PATCH 3/6] drm/amdkfd: Avoid flooding dmesg on Raven due to IOMMU issues

Quoted 9 lines · Felix Kuehling

From: Yong Zhao <[email protected]>

On Raven Invalid PPRs can be reported because multiple PPRs can be still queued when memory is freed. Apply a rate limit to avoid flooding the log in this case.

Maybe define PPRs in the description. With that fixed:

Quoted 20 lines
---
 drivers/gpu/drm/amd/amdkfd/kfd_iommu.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_iommu.c +1 -2
index c718179..7a61f38 100644
@@ -190,7 +190,7 @@ static int iommu_invalid_ppr_cb(struct pci_dev *pdev, int pasid,
 {
        struct kfd_dev *dev;
 
-       dev_warn(kfd_device,
+       dev_warn_ratelimited(kfd_device,
                        "Invalid PPR device %x:%x.%x pasid %d address 0x%lX flags 0x%X",
                        PCI_BUS_NUM(pdev->devfn),
                        PCI_SLOT(pdev->devfn),
--

2.7.4

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#11 lore raw

PATCH 4/6 drm/amdkfd: Workaround to accommodate Raven too many PPR issue

In reply to a message not in the archive

From: Yong Zhao <[email protected]>

---
 drivers/gpu/drm/amd/amdkfd/kfd_events.c | 21 ++++++++++++++++-----
 1 file changed, 16 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_events.c +16 -5
index 820133c..4dcacce 100644
@@ -932,13 +932,24 @@ void kfd_signal_iommu_event(struct kfd_dev *dev, unsigned int pasid,
 	up_read(&mm->mmap_sem);
 	mmput(mm);
 
-	mutex_lock(&p->event_mutex);
+	pr_debug("notpresent %d, noexecute %d, readonly %d\n",
+			memory_exception_data.failure.NotPresent,
+			memory_exception_data.failure.NoExecute,
+			memory_exception_data.failure.ReadOnly);
 
-	/* Lookup events by type and signal them */
-	lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_MEMORY,
-			&memory_exception_data);
+	/* Workaround on Raven to not kill the process when memory is freed
+	 * before IOMMU is able to finish processing all the excessive PPRs
+	 */
+	if (dev->device_info->asic_family != CHIP_RAVEN) {
+		mutex_lock(&p->event_mutex);
+
+		/* Lookup events by type and signal them */
+		lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_MEMORY,
+				&memory_exception_data);
+
+		mutex_unlock(&p->event_mutex);
+	}
 
-	mutex_unlock(&p->event_mutex);
 	kfd_unref_process(p);
 }
 #endif /* KFD_SUPPORT_IOMMU_V2 */
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1531430694-23966-5-git-send-email-Felix.Kuehling-5C7GfCeVMHo@public.gmane.org

#13 lore raw

Re: [PATCH 4/6] drm/amdkfd: Workaround to accommodate Raven too many PPR issue

Quoted 1 line · Felix Kuehling

From: Yong Zhao <[email protected]>

Please add a patch description explaining why this is needed for Raven.

Alex

Quoted 45 lines
---
 drivers/gpu/drm/amd/amdkfd/kfd_events.c | 21 ++++++++++++++++-----
 1 file changed, 16 insertions(+), 5 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_events.c +16 -6
index 820133c..4dcacce 100644
@@ -932,13 +932,24 @@ void kfd_signal_iommu_event(struct kfd_dev *dev, unsigned int pasid,
        up_read(&mm->mmap_sem);
        mmput(mm);
 
-       mutex_lock(&p->event_mutex);
+       pr_debug("notpresent %d, noexecute %d, readonly %d\n",
+                       memory_exception_data.failure.NotPresent,
+                       memory_exception_data.failure.NoExecute,
+                       memory_exception_data.failure.ReadOnly);
 
-       /* Lookup events by type and signal them */
-       lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_MEMORY,
-                       &memory_exception_data);
+       /* Workaround on Raven to not kill the process when memory is freed
+        * before IOMMU is able to finish processing all the excessive PPRs
+        */
+       if (dev->device_info->asic_family != CHIP_RAVEN) {
+               mutex_lock(&p->event_mutex);
+
+               /* Lookup events by type and signal them */
+               lookup_events_by_type_and_signal(p, KFD_EVENT_TYPE_MEMORY,
+                               &memory_exception_data);
+
+               mutex_unlock(&p->event_mutex);
+       }
 
-       mutex_unlock(&p->event_mutex);
        kfd_unref_process(p);
 }
 #endif /* KFD_SUPPORT_IOMMU_V2 */
--

2.7.4

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#14 lore raw

PATCH 5/6 drm/amdkfd: Optimize out some duplicated code in kfd_signal_iommu_event()

In reply to a message not in the archive

From: Yong Zhao <[email protected]>

---
 drivers/gpu/drm/amd/amdkfd/kfd_events.c | 26 +++++++++++---------------
 1 file changed, 11 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_events.c +11 -15
index 4dcacce..e9f0e0a 100644
@@ -911,22 +911,18 @@ void kfd_signal_iommu_event(struct kfd_dev *dev, unsigned int pasid,
 	memory_exception_data.failure.NotPresent = 1;
 	memory_exception_data.failure.NoExecute = 0;
 	memory_exception_data.failure.ReadOnly = 0;
-	if (vma) {
-		if (vma->vm_start > address) {
-			memory_exception_data.failure.NotPresent = 1;
-			memory_exception_data.failure.NoExecute = 0;
+	if (vma && address >= vma->vm_start) {
+		memory_exception_data.failure.NotPresent = 0;
+
+		if (is_write_requested && !(vma->vm_flags & VM_WRITE))
+			memory_exception_data.failure.ReadOnly = 1;
+		else
 			memory_exception_data.failure.ReadOnly = 0;
-		} else {
-			memory_exception_data.failure.NotPresent = 0;
-			if (is_write_requested && !(vma->vm_flags & VM_WRITE))
-				memory_exception_data.failure.ReadOnly = 1;
-			else
-				memory_exception_data.failure.ReadOnly = 0;
-			if (is_execute_requested && !(vma->vm_flags & VM_EXEC))
-				memory_exception_data.failure.NoExecute = 1;
-			else
-				memory_exception_data.failure.NoExecute = 0;
-		}
+
+		if (is_execute_requested && !(vma->vm_flags & VM_EXEC))
+			memory_exception_data.failure.NoExecute = 1;
+		else
+			memory_exception_data.failure.NoExecute = 0;
 	}
 
 	up_read(&mm->mmap_sem);
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1531430694-23966-6-git-send-email-Felix.Kuehling-5C7GfCeVMHo@public.gmane.org

#16 lore raw

Re: [PATCH 5/6] drm/amdkfd: Optimize out some duplicated code in kfd_signal_iommu_event()

Quoted 1 line · Felix Kuehling

From: Yong Zhao <[email protected]>

Please add a patch description.

Alex

Quoted 49 lines
---
 drivers/gpu/drm/amd/amdkfd/kfd_events.c | 26 +++++++++++---------------
 1 file changed, 11 insertions(+), 15 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_events.c +11 -16
index 4dcacce..e9f0e0a 100644
@@ -911,22 +911,18 @@ void kfd_signal_iommu_event(struct kfd_dev *dev, unsigned int pasid,
        memory_exception_data.failure.NotPresent = 1;
        memory_exception_data.failure.NoExecute = 0;
        memory_exception_data.failure.ReadOnly = 0;
-       if (vma) {
-               if (vma->vm_start > address) {
-                       memory_exception_data.failure.NotPresent = 1;
-                       memory_exception_data.failure.NoExecute = 0;
+       if (vma && address >= vma->vm_start) {
+               memory_exception_data.failure.NotPresent = 0;
+
+               if (is_write_requested && !(vma->vm_flags & VM_WRITE))
+                       memory_exception_data.failure.ReadOnly = 1;
+               else
                        memory_exception_data.failure.ReadOnly = 0;
-               } else {
-                       memory_exception_data.failure.NotPresent = 0;
-                       if (is_write_requested && !(vma->vm_flags & VM_WRITE))
-                               memory_exception_data.failure.ReadOnly = 1;
-                       else
-                               memory_exception_data.failure.ReadOnly = 0;
-                       if (is_execute_requested && !(vma->vm_flags & VM_EXEC))
-                               memory_exception_data.failure.NoExecute = 1;
-                       else
-                               memory_exception_data.failure.NoExecute = 0;
-               }
+
+               if (is_execute_requested && !(vma->vm_flags & VM_EXEC))
+                       memory_exception_data.failure.NoExecute = 1;
+               else
+                       memory_exception_data.failure.NoExecute = 0;
        }
 
        up_read(&mm->mmap_sem);
--

2.7.4

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#17 lore raw

PATCH 6/6 drm/amdkfd: Enable Raven for KFD

In reply to a message not in the archive

From: Yong Zhao <[email protected]>

---
 drivers/gpu/drm/amd/amdkfd/kfd_device.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +16 -0
index 572235c..1b04871 100644
@@ -70,6 +70,21 @@ static const struct kfd_device_info carrizo_device_info = {
 	.needs_pci_atomics = false,
 	.num_sdma_engines = 2,
 };
+
+static const struct kfd_device_info raven_device_info = {
+	.asic_family = CHIP_RAVEN,
+	.max_pasid_bits = 16,
+	.max_no_of_hqd  = 24,
+	.doorbell_size  = 8,
+	.ih_ring_entry_size = 8 * sizeof(uint32_t),
+	.event_interrupt_class = &event_interrupt_class_v9,
+	.num_of_watch_points = 4,
+	.mqd_size_aligned = MQD_SIZE_ALIGNED,
+	.supports_cwsr = true,
+	.needs_iommu_device = true,
+	.needs_pci_atomics = true,
+	.num_sdma_engines = 1,
+};
 #endif
 
 static const struct kfd_device_info hawaii_device_info = {
@@ -259,6 +274,7 @@ static const struct kfd_deviceid supported_devices[] = {
 	{ 0x9875, &carrizo_device_info },	/* Carrizo */
 	{ 0x9876, &carrizo_device_info },	/* Carrizo */
 	{ 0x9877, &carrizo_device_info },	/* Carrizo */
+	{ 0x15DD, &raven_device_info },		/* Raven */
 #endif
 	{ 0x67A0, &hawaii_device_info },	/* Hawaii */
 	{ 0x67A1, &hawaii_device_info },	/* Hawaii */
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1531430694-23966-7-git-send-email-Felix.Kuehling-5C7GfCeVMHo@public.gmane.org

Acked-by
#19 lore raw

Re: [PATCH 6/6] drm/amdkfd: Enable Raven for KFD

Quoted 5 lines · Felix Kuehling

From: Yong Zhao <[email protected]>

Please add a patch description. With that fixed:

Quoted 41 lines
---
 drivers/gpu/drm/amd/amdkfd/kfd_device.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +16 -1
index 572235c..1b04871 100644
@@ -70,6 +70,21 @@ static const struct kfd_device_info carrizo_device_info = {
        .needs_pci_atomics = false,
        .num_sdma_engines = 2,
 };
+
+static const struct kfd_device_info raven_device_info = {
+       .asic_family = CHIP_RAVEN,
+       .max_pasid_bits = 16,
+       .max_no_of_hqd  = 24,
+       .doorbell_size  = 8,
+       .ih_ring_entry_size = 8 * sizeof(uint32_t),
+       .event_interrupt_class = &event_interrupt_class_v9,
+       .num_of_watch_points = 4,
+       .mqd_size_aligned = MQD_SIZE_ALIGNED,
+       .supports_cwsr = true,
+       .needs_iommu_device = true,
+       .needs_pci_atomics = true,
+       .num_sdma_engines = 1,
+};
 #endif
 
 static const struct kfd_device_info hawaii_device_info = {
@@ -259,6 +274,7 @@ static const struct kfd_deviceid supported_devices[] = {
        { 0x9875, &carrizo_device_info },       /* Carrizo */
        { 0x9876, &carrizo_device_info },       /* Carrizo */
        { 0x9877, &carrizo_device_info },       /* Carrizo */
+       { 0x15DD, &raven_device_info },         /* Raven */
 #endif
        { 0x67A0, &hawaii_device_info },        /* Hawaii */
        { 0x67A1, &hawaii_device_info },        /* Hawaii */
--

2.7.4

Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

#20 lore raw
In reply to a message not in the archive

Hi

I'm happy to test this out, just wondering what userspace I should pair it with

Cheers

Mike

Quoted 28 lines · On Thu, 12 Jul 2018 at 22:25 Felix Kuehling…

Raven refers to Ryzen APUs with integrated GFXv9 GPU. This patch series completes Raven support for KFD:

* fix up memory banks info from CRAT
* support different number of SDMA engines
* workaround IOMMUv2 PPR issues
* add device info

Yong Zhao (6):
drm/amdkfd: Consolidate duplicate memory banks info in topology drm/amdkfd: Make SDMA engine number an ASIC-dependent variable drm/amdkfd: Avoid flooding dmesg on Raven due to IOMMU issues drm/amdkfd: Workaround to accommodate Raven too many PPR issue drm/amdkfd: Optimize out some duplicated code in
kfd_signal_iommu_event()
drm/amdkfd: Enable Raven for KFD

 drivers/gpu/drm/amd/amdkfd/kfd_crat.c              | 57
+++++++++++++++++-----
 drivers/gpu/drm/amd/amdkfd/kfd_device.c            | 28 +++++++++++
 .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c  | 29 +++++++----
 .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h  |  6 +--
 drivers/gpu/drm/amd/amdkfd/kfd_events.c            | 47 ++++++++++--------
 drivers/gpu/drm/amd/amdkfd/kfd_iommu.c             |  2 +-
 drivers/gpu/drm/amd/amdkfd/kfd_priv.h              |  1 +
 .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c |  3 +-
 8 files changed, 126 insertions(+), 47 deletions(-)
Quoted 8 lines

--
2.7.4

_______________________________________________
amd-gfx mailing list [email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): CAHbf0-Fv9npKsGZmgrNLmx6tS7800YOP4ijFEWGjCa3EwfzftQ-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org

#22 lore raw

This patch series applies on top of my previous 25-patch series on top of Oded's amdkfd-next branch.

On the user mode side, you need my drm-next-wip branch in
https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface/tree/fxkamd/drm-next-wip.
The rest of ROCm user mode should work unmodified.

Regards,
  Felix

Quoted 41 lines · Mike Lothian

Hi

I'm happy to test this out, just wondering what userspace I should pair it with

Cheers

Mike

On Thu, 12 Jul 2018 at 22:25 Felix Kuehling <[email protected] <mailto:[email protected]>> wrote:

Raven refers to Ryzen APUs with integrated GFXv9 GPU. This patch series completes Raven support for KFD:

* fix up memory banks info from CRAT
* support different number of SDMA engines
* workaround IOMMUv2 PPR issues
* add device info

Yong Zhao (6):
  drm/amdkfd: Consolidate duplicate memory banks info in topology drm/amdkfd: Make SDMA engine number an ASIC-dependent variable drm/amdkfd: Avoid flooding dmesg on Raven due to IOMMU issues drm/amdkfd: Workaround to accommodate Raven too many PPR issue drm/amdkfd: Optimize out some duplicated code in kfd_signal_iommu_event()
  drm/amdkfd: Enable Raven for KFD

 drivers/gpu/drm/amd/amdkfd/kfd_crat.c              | 57
+++++++++++++++++-----
 drivers/gpu/drm/amd/amdkfd/kfd_device.c            | 28 +++++++++++ .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.c  | 29 +++++++---- .../gpu/drm/amd/amdkfd/kfd_device_queue_manager.h  |  6 +-- drivers/gpu/drm/amd/amdkfd/kfd_events.c            | 47
++++++++++--------
 drivers/gpu/drm/amd/amdkfd/kfd_iommu.c             |  2 +- drivers/gpu/drm/amd/amdkfd/kfd_priv.h              |  1 + .../gpu/drm/amd/amdkfd/kfd_process_queue_manager.c |  3 +- 8 files changed, 126 insertions(+), 47 deletions(-)

Quoted 8 lines

--
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected] <mailto:[email protected]>
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx