drm/amdkfd: Fix a bug in SDMA RLC queue counting under HWS mode
Subsystems (3)
PATCH 1/1 drm/amdkfd: Fix a bug in SDMA RLC queue counting under HWS mode
The sdma_queue_count increment should be done before execute_queues_cpsch(), which calls pm_calc_rlib_size() where sdma_queue_count is used to calculate whether over_subscription is triggered.
With the previous code, when a SDMA queue is created, compute_queue_count in pm_calc_rlib_size() is one more than the actual compute queue number, because the queue_count has been incremented while sdma_queue_count has not. This patch fixes that.
--- drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +6 -4
@@ -1237,16 +1237,18 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,list_add (&q->list, &qpd->queues_list); qpd->queue_count++; + +if (q->properties.type ==KFD_QUEUE_TYPE_SDMA ) + dqm->sdma_queue_count++; +else if (q->properties.type ==KFD_QUEUE_TYPE_SDMA_XGMI ) + dqm->xgmi_sdma_queue_count++; +if (q->properties.is_active) { dqm->queue_count++; retval =execute_queues_cpsch (dqm,KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES ,0 ); } -if (q->properties.type ==KFD_QUEUE_TYPE_SDMA ) - dqm->sdma_queue_count++; -else if (q->properties.type ==KFD_QUEUE_TYPE_SDMA_XGMI ) - dqm->xgmi_sdma_queue_count++;/* * Unconditionally increment this counter, regardless of the queue's * type or whether the queue is active.
Signature · 6 lines
2.17.1 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Quoted 5 lines · Yong Zhao
The
sdma_queue_countincrement should be done beforeexecute_queues_cpsch(), which callspm_calc_rlib_size()wheresdma_queue_countis used to calculate whetherover_subscriptionis triggered.
Quoted 7 lines
With the previous code, when a SDMA queue is created,
compute_queue_countinpm_calc_rlib_size()is one more than the actual compute queue number, because thequeue_counthas been incremented whilesdma_queue_counthas not. This patch fixes that.Change-Id I20353e657efd505353d0dd9f7eb2fab5085e7202Signed-off-by Yong Zhao <[email protected]>
But I took a look at pm_calc_rlib_size. I don't think subtracting dqm->sdma_queue_count from dqm->queue_count is not quite correct, because sdma_queue_count counts all SDMA queues, while queue_count only counts active queues. So an application that creates inactive SDMA queues will also create errors here. We probably need to count active compute and active SDMA queues separately in DQM to fix this properly.
Regards,
Felix
Quoted 21 lines
--- drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +6
-0@@ -1237,16 +1237,18 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q,list_add (&q->list, &qpd->queues_list); qpd->queue_count++; + +if (q->properties.type ==KFD_QUEUE_TYPE_SDMA ) + dqm->sdma_queue_count++; +else if (q->properties.type ==KFD_QUEUE_TYPE_SDMA_XGMI ) + dqm->xgmi_sdma_queue_count++; +if (q->properties.is_active) { dqm->queue_count++; retval = execute_queues_cpsch(dqm,
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, · line 1254
Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0
-4KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0); } -if (q->properties.type ==KFD_QUEUE_TYPE_SDMA ) - dqm->sdma_queue_count++; -else if (q->properties.type ==KFD_QUEUE_TYPE_SDMA_XGMI ) - dqm->xgmi_sdma_queue_count++;/* * Unconditionally increment this counter, regardless of the queue's * type or whether the queue is active.
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
True. It is a bug too. I am looking into it.
Yong
Quoted 44 lines · Felix Kuehling
Quoted 5 lines · Yong Zhao
The
sdma_queue_countincrement should be done beforeexecute_queues_cpsch(), which callspm_calc_rlib_size()wheresdma_queue_countis used to calculate whetherover_subscriptionis triggered.Quoted 7 lines
With the previous code, when a SDMA queue is created,
compute_queue_countinpm_calc_rlib_size()is one more than the actual compute queue number, because thequeue_counthas been incremented whilesdma_queue_counthas not. This patch fixes that.Change-Id I20353e657efd505353d0dd9f7eb2fab5085e7202Signed-off-by Yong Zhao <[email protected]>Reviewed-by Felix Kuehling <[email protected]>But I took a look at
pm_calc_rlib_size. I don't think subtractingdqm->sdma_queue_countfromdqm->queue_countis not quite correct, becausesdma_queue_countcounts all SDMA queues, whilequeue_countonly counts active queues. So an application that creates inactive SDMA queues will also create errors here. We probably need to count active compute and active SDMA queues separately in DQM to fix this properly.Regards,
FelixQuoted 22 lines
---
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c| 10 ++++++----1 file changed, 6 insertions(+), 4 deletions(-)drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +6
-0@@ -1237,16 +1237,18 @@ static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, list_add(&q->list, &qpd->queues_list); qpd->queue_count++; + + if (q->properties.type == KFD_QUEUE_TYPE_SDMA) + dqm->sdma_queue_count++; + else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) + dqm->xgmi_sdma_queue_count++; + if (q->properties.is_active) { dqm->queue_count++; retval = execute_queues_cpsch(dqm, KFD_UNMAP_QUEUES_FILTER_DYNAMIC_QUEUES, 0);
Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c static int create_queue_cpsch(struct device_queue_manager *dqm, struct queue *q, · line 1254
Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c +0
-3} - if (q->properties.type == KFD_QUEUE_TYPE_SDMA) - dqm->sdma_queue_count++; - else if (q->properties.type == KFD_QUEUE_TYPE_SDMA_XGMI) - dqm->xgmi_sdma_queue_count++; /** Unconditionally increment this counter, regardless of the queue's * type or whether the queue is active.
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx