On 2024-10-28 17:40, Xiaogang.Chen wrote:
From: Xiaogang Chen <xiaogang.chen@xxxxxxx>
To allow user better understand the cause triggering runlist oversubscription.
No function change.
Signed-off-by: Xiaogang Chen Xiaogang.Chen@xxxxxxx
---
.../gpu/drm/amd/amdkfd/kfd_packet_manager.c | 55 ++++++++++++++-----
1 file changed, 42 insertions(+), 13 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c
index 37930629edc5..e22be6da23b7 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_packet_manager.c
@@ -28,6 +28,10 @@
#include "kfd_kernel_queue.h"
#include "kfd_priv.h"
+#define OVER_SUBSCRIPTION_PROCESS_COUNT 1 << 0
+#define OVER_SUBSCRIPTION_COMPUTE_QUEUE_COUNT 1 << 1
+#define OVER_SUBSCRIPTION_GWS_QUEUE_COUNT 1 << 2
+
static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
unsigned int buffer_size_bytes)
{
@@ -40,7 +44,7 @@ static inline void inc_wptr(unsigned int *wptr, unsigned int increment_bytes,
static void pm_calc_rlib_size(struct packet_manager *pm,
unsigned int *rlib_size,
- bool *over_subscription)
+ int *over_subscription)
{
unsigned int process_count, queue_count, compute_queue_count, gws_queue_count;
unsigned int map_queue_size;
@@ -58,17 +62,20 @@ static void pm_calc_rlib_size(struct packet_manager *pm,
* hws_max_conc_proc has been done in
* kgd2kfd_device_init().
*/
- *over_subscription = false;
+ *over_subscription = 0;
if (node->max_proc_per_quantum > 1)
max_proc_per_quantum = node->max_proc_per_quantum;
- if ((process_count > max_proc_per_quantum) ||
- compute_queue_count > get_cp_queues_num(pm->dqm) ||
- gws_queue_count > 1) {
- *over_subscription = true;
+ if (process_count > max_proc_per_quantum)
+ *over_subscription = *over_subscription || OVER_SUBSCRIPTION_PROCESS_COUNT;
+ if (compute_queue_count > get_cp_queues_num(pm->dqm))
+ *over_subscription = *over_subscription || OVER_SUBSCRIPTION_COMPUTE_QUEUE_COUNT;
+ if (gws_queue_count > 1)
+ *over_subscription = *over_subscription || OVER_SUBSCRIPTION_GWS_QUEUE_COUNT;
+
+ if (*over_subscription)
dev_dbg(dev, "Over subscribed runlist\n");
- }
map_queue_size = pm->pmf->map_queues_size;
/* calculate run list ib allocation size */
@@ -89,7 +96,7 @@ static int pm_allocate_runlist_ib(struct packet_manager *pm,
unsigned int **rl_buffer,
uint64_t *rl_gpu_buffer,
unsigned int *rl_buffer_size,
- bool *is_over_subscription)
+ int *is_over_subscription)
{
struct kfd_node *node = pm->dqm->dev;
struct device *dev = node->adev->dev;
@@ -134,7 +141,7 @@ static int pm_create_runlist_ib(struct packet_manager *pm,
struct qcm_process_device *qpd;
struct queue *q;
struct kernel_queue *kq;
- bool is_over_subscription;
+ int is_over_subscription;
rl_wptr = retval = processes_mapped = 0;
@@ -212,16 +219,38 @@ static int pm_create_runlist_ib(struct packet_manager *pm,
dev_dbg(dev, "Finished map process and queues to runlist\n");
if (is_over_subscription) {
- if (!pm->is_over_subscription)
- dev_warn(
+ if (!pm->is_over_subscription) {
+
+ if (is_over_subscription & OVER_SUBSCRIPTION_PROCESS_COUNT) {
+ dev_warn(
dev,
- "Runlist is getting oversubscribed. Expect reduced ROCm performance.\n");
+ "process number is more than maximum number of processes that"
+ " HWS can schedule concurrently. Runlist is getting"
+ " oversubscribed. Expect reduced ROCm performance.\n");
+ }
+
+ if (is_over_subscription & OVER_SUBSCRIPTION_COMPUTE_QUEUE_COUNT) {
+ dev_warn(
+ dev,
+ "compute queue number is more than assigned compute queues."
+ " Runlist is getting"
+ " oversubscribed. Expect reduced ROCm performance.\n");
+ }
+
+ if (is_over_subscription & OVER_SUBSCRIPTION_GWS_QUEUE_COUNT) {
+ dev_warn(
+ dev,
+ "compute queue for cooperative workgroup is more than allowed."
+ " Runlist is getting"
+ " oversubscribed. Expect reduced ROCm performance.\n");
+ }
I like the concept of showing the cause of oversubscription. Maybe we
should add "process isolation mode" as a special case of "process count".
The messages are overly verbose. There is a common part of the message
that could be printed if is_over_subscription is non-zero. Then just
print some extra info about the cause, e.g.:
if (is_over_subscription) {
dev_warn("Runlist is getting oversubscribed due to%s%s%s. Expect reduced ROCm performance.\n",
is_over_subscription & OVER_SUBSCRIPTION_PROCESS_COUNT ? " number-of-processes" : "",
is_over_subscription & OVER_SUBSCRIPTION_COMPUTE_QUEUE_COUNT ? " number-of-queues" : "",
is_over_subscription & OVER_SUBSCRIPTION_GWS_QUEUE_COUNT ? " cooperative-launch" : "");
}
Regards,
Felix
+ }
retval = pm->pmf->runlist(pm, &rl_buffer[rl_wptr],
*rl_gpu_addr,
alloc_size_bytes / sizeof(uint32_t),
true);
}
- pm->is_over_subscription = is_over_subscription;
+ pm->is_over_subscription = is_over_subscription ? true : false;
for (i = 0; i < alloc_size_bytes / sizeof(uint32_t); i++)
pr_debug("0x%2X ", rl_buffer[i]);
