[PATCH 13/13] drm/amdkfd: Generate xGMI direct iolink
shaoyunl
Shaoyun.Liu at amd.com
Wed Sep 5 15:32:50 UTC 2018
From: Shaoyun Liu <Shaoyun.Liu at amd.com>
Generate xGMI iolink for upper level usage
Change-Id: I37bc29fee45cb10d1da849956055c59d823f6f5d
Signed-off-by: Shaoyun Liu <Shaoyun.Liu at amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling at amd.com>
---
drivers/gpu/drm/amd/amdkfd/kfd_crat.c | 78 ++++++++++++++++++++++++++++++-----
1 file changed, 68 insertions(+), 10 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
index ee49960..130db4d 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_crat.c
@@ -346,7 +346,7 @@ static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink,
struct list_head *device_list)
{
struct kfd_iolink_properties *props = NULL, *props2;
- struct kfd_topology_device *dev, *cpu_dev;
+ struct kfd_topology_device *dev, *to_dev;
uint32_t id_from;
uint32_t id_to;
@@ -369,6 +369,8 @@ static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink,
if (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS)
props->weight = 20;
+ else if (props->iolink_type == CRAT_IOLINK_TYPE_XGMI)
+ props->weight = 15;
else
props->weight = node_distance(id_from, id_to);
@@ -390,19 +392,22 @@ static int kfd_parse_subtype_iolink(struct crat_subtype_iolink *iolink,
* links are not built at that time. If a PCIe type is discovered, it
* means a GPU is detected and we are adding GPU->CPU to the topology.
* At this time, also add the corresponded CPU->GPU link.
+ * For xGMI, we only added the link with one direction in the crat
+ * table, add corresponded reversed direction link now.
*/
- if (props && props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS) {
- cpu_dev = kfd_topology_device_by_proximity_domain(id_to);
- if (!cpu_dev)
+ if (props && (props->iolink_type == CRAT_IOLINK_TYPE_PCIEXPRESS ||
+ props->iolink_type == CRAT_IOLINK_TYPE_XGMI)) {
+ to_dev = kfd_topology_device_by_proximity_domain(id_to);
+ if (!to_dev)
return -ENODEV;
/* same everything but the other direction */
props2 = kmemdup(props, sizeof(*props2), GFP_KERNEL);
props2->node_from = id_to;
props2->node_to = id_from;
props2->kobj = NULL;
- cpu_dev->io_link_count++;
- cpu_dev->node_props.io_links_count++;
- list_add_tail(&props2->list, &cpu_dev->io_link_props);
+ to_dev->io_link_count++;
+ to_dev->node_props.io_links_count++;
+ list_add_tail(&props2->list, &to_dev->io_link_props);
}
return 0;
@@ -1037,7 +1042,7 @@ static int kfd_fill_gpu_memory_affinity(int *avail_size,
*
* Return 0 if successful else return -ve value
*/
-static int kfd_fill_gpu_direct_io_link(int *avail_size,
+static int kfd_fill_gpu_direct_io_link_to_cpu(int *avail_size,
struct kfd_dev *kdev,
struct crat_subtype_iolink *sub_type_hdr,
uint32_t proximity_domain)
@@ -1069,6 +1074,28 @@ static int kfd_fill_gpu_direct_io_link(int *avail_size,
return 0;
}
+static int kfd_fill_gpu_xgmi_link_to_gpu(int *avail_size,
+ struct kfd_dev *kdev,
+ struct crat_subtype_iolink *sub_type_hdr,
+ uint32_t proximity_domain_from,
+ uint32_t proximity_domain_to)
+{
+ *avail_size -= sizeof(struct crat_subtype_iolink);
+ if (*avail_size < 0)
+ return -ENOMEM;
+
+ memset((void *)sub_type_hdr, 0, sizeof(struct crat_subtype_iolink));
+
+ sub_type_hdr->type = CRAT_SUBTYPE_IOLINK_AFFINITY;
+ sub_type_hdr->length = sizeof(struct crat_subtype_iolink);
+ sub_type_hdr->flags |= CRAT_SUBTYPE_FLAGS_ENABLED;
+
+ sub_type_hdr->io_interface_type = CRAT_IOLINK_TYPE_XGMI;
+ sub_type_hdr->proximity_domain_from = proximity_domain_from;
+ sub_type_hdr->proximity_domain_to = proximity_domain_to;
+ return 0;
+}
+
/* kfd_create_vcrat_image_gpu - Create Virtual CRAT for CPU
*
* @pcrat_image: Fill in VCRAT for GPU
@@ -1081,14 +1108,16 @@ static int kfd_create_vcrat_image_gpu(void *pcrat_image,
{
struct crat_header *crat_table = (struct crat_header *)pcrat_image;
struct crat_subtype_generic *sub_type_hdr;
+ struct kfd_local_mem_info local_mem_info;
+ struct kfd_topology_device *peer_dev;
struct crat_subtype_computeunit *cu;
struct kfd_cu_info cu_info;
int avail_size = *size;
uint32_t total_num_of_cu;
int num_of_cache_entries = 0;
int cache_mem_filled = 0;
+ uint32_t nid = 0;
int ret = 0;
- struct kfd_local_mem_info local_mem_info;
if (!pcrat_image || avail_size < VCRAT_SIZE_FOR_GPU)
return -EINVAL;
@@ -1212,7 +1241,7 @@ static int kfd_create_vcrat_image_gpu(void *pcrat_image,
*/
sub_type_hdr = (typeof(sub_type_hdr))((char *)sub_type_hdr +
cache_mem_filled);
- ret = kfd_fill_gpu_direct_io_link(&avail_size, kdev,
+ ret = kfd_fill_gpu_direct_io_link_to_cpu(&avail_size, kdev,
(struct crat_subtype_iolink *)sub_type_hdr, proximity_domain);
if (ret < 0)
@@ -1221,6 +1250,35 @@ static int kfd_create_vcrat_image_gpu(void *pcrat_image,
crat_table->length += sub_type_hdr->length;
crat_table->total_entries++;
+
+ /* Fill in Subtype: IO_LINKS
+ * Direct links from GPU to other GPUs through xGMI.
+ * We will loop GPUs that already be processed (with lower value
+ * of proximity_domain), add the link for the GPUs with same
+ * hive id (from this GPU to other GPU) . The reversed iolink
+ * (from other GPU to this GPU) will be added
+ * in kfd_parse_subtype_iolink.
+ */
+ if (kdev->hive_id) {
+ for (nid = 0; nid < proximity_domain; ++nid) {
+ peer_dev = kfd_topology_device_by_proximity_domain(nid);
+ if (!peer_dev->gpu)
+ continue;
+ if (peer_dev->gpu->hive_id != kdev->hive_id)
+ continue;
+ sub_type_hdr = (typeof(sub_type_hdr))(
+ (char *)sub_type_hdr +
+ sizeof(struct crat_subtype_iolink));
+ ret = kfd_fill_gpu_xgmi_link_to_gpu(
+ &avail_size, kdev,
+ (struct crat_subtype_iolink *)sub_type_hdr,
+ proximity_domain, nid);
+ if (ret < 0)
+ return ret;
+ crat_table->length += sub_type_hdr->length;
+ crat_table->total_entries++;
+ }
+ }
*size = crat_table->length;
pr_info("Virtual CRAT table created for GPU\n");
--
2.7.4
More information about the amd-gfx
mailing list