[PATCH 2/2] drm/amdkfd: Check preemption status on all XCDs
Joshi, Mukul
Mukul.Joshi at amd.com
Thu Mar 14 19:28:18 UTC 2024
[AMD Official Use Only - General]
From: Kuehling, Felix <Felix.Kuehling at amd.com>
Sent: Thursday, March 14, 2024 2:39 PM
To: Joshi, Mukul <Mukul.Joshi at amd.com>; amd-gfx at lists.freedesktop.org
Cc: Cornwall, Jay <Jay.Cornwall at amd.com>
Subject: Re: [PATCH 2/2] drm/amdkfd: Check preemption status on all XCDs
On 2024-03-14 12:00, Mukul Joshi wrote:
This patch adds the following functionality:
- Check the queue preemption status on all XCDs in a partition
for GFX 9.4.3.
- Update the queue preemption debug message to print the queue
doorbell id for which preemption failed.
- Change the signature of check preemption failed function to
return a bool instead of uint32_t and pass the MQD manager
as an argument.
Suggested-by: Jay Cornwall <jay.cornwall at amd.com><mailto:jay.cornwall at amd.com>
Signed-off-by: Mukul Joshi <mukul.joshi at amd.com><mailto:mukul.joshi at amd.com>
---
.../drm/amd/amdkfd/kfd_device_queue_manager.c | 3 +--
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c | 18 +++++++++++++
drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h | 4 ++-
.../gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c | 4 +--
.../gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c | 4 +--
.../gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c | 4 +--
.../gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c | 25 ++++++++++++++++---
.../gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c | 4 +--
8 files changed, 52 insertions(+), 14 deletions(-)
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 1ce398ab0b3d..151fabf84040 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_device_queue_manager.c
@@ -1997,8 +1997,7 @@ static int unmap_queues_cpsch(struct device_queue_manager *dqm,
* check those fields
*/
mqd_mgr = dqm->mqd_mgrs[KFD_MQD_TYPE_HIQ];
- if (mqd_mgr->check_preemption_failed(dqm->packet_mgr.priv_queue->queue->mqd)) {
- dev_err(dev, "HIQ MQD's queue_doorbell_id0 is not 0, Queue preemption time out\n");
+ if (mqd_mgr->check_preemption_failed(mqd_mgr, dqm->packet_mgr.priv_queue->queue->mqd)) {
while (halt_if_hws_hang)
schedule();
return -ETIME;
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c
index 050a6936ff84..cbec8c87c984 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.c
@@ -290,3 +290,21 @@ uint64_t kfd_mqd_stride(struct mqd_manager *mm,
{
return mm->mqd_size;
}
+
+bool kfd_check_hiq_mqd_doorbell_id(struct kfd_node *node, uint32_t doorbell_id,
+ uint32_t inst)
+{
+ if (doorbell_id) {
+ struct device *dev = node->adev->dev;
+
+ if (KFD_GC_VERSION(node) == IP_VERSION(9, 4, 3))
Could this be made more generic? E.g.:
if (node->adev->xcp_mgr && node->adev->xcp_mgr->num_xcps > 0)
Yes this seems much better. I will make this change before submitting.
Thanks,
Mukul
Other than that, the series is
Reviewed-by: Felix Kuehling <felix.kuehling at amd.com><mailto:felix.kuehling at amd.com>
+ dev_err(dev, "XCC %d: Queue preemption failed for queue with doorbell_id: %x\n",
+ inst, doorbell_id);
+ else
+ dev_err(dev, "Queue preemption failed for queue with doorbell_id: %x\n",
+ doorbell_id);
+ return true;
+ }
+
+ return false;
+}
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h
index ba3eebb2ca6d..17cc1f25c8d0 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager.h
@@ -119,7 +119,7 @@ struct mqd_manager {
#if defined(CONFIG_DEBUG_FS)
int (*debugfs_show_mqd)(struct seq_file *m, void *data);
#endif
- uint32_t (*check_preemption_failed)(void *mqd);
+ bool (*check_preemption_failed)(struct mqd_manager *mm, void *mqd);
uint64_t (*mqd_stride)(struct mqd_manager *mm,
struct queue_properties *p);
@@ -198,4 +198,6 @@ void kfd_get_hiq_xcc_mqd(struct kfd_node *dev,
uint64_t kfd_hiq_mqd_stride(struct kfd_node *dev);
uint64_t kfd_mqd_stride(struct mqd_manager *mm,
struct queue_properties *q);
+bool kfd_check_hiq_mqd_doorbell_id(struct kfd_node *node, uint32_t doorbell_id,
+ uint32_t inst);
#endif /* KFD_MQD_MANAGER_H_ */
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c
index 8f9f56f7a8b0..05f3ac2eaef9 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_cik.c
@@ -206,11 +206,11 @@ static void __update_mqd(struct mqd_manager *mm, void *mqd,
q->is_active = QUEUE_IS_ACTIVE(*q);
}
-static uint32_t check_preemption_failed(void *mqd)
+static bool check_preemption_failed(struct mqd_manager *mm, void *mqd)
{
struct cik_mqd *m = (struct cik_mqd *)mqd;
- return m->queue_doorbell_id0;
+ return kfd_check_hiq_mqd_doorbell_id(mm->dev, m->queue_doorbell_id0, 0);
}
static void update_mqd(struct mqd_manager *mm, void *mqd,
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c
index d4cf7d845928..2eff37aaf827 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v10.c
@@ -224,11 +224,11 @@ static void update_mqd(struct mqd_manager *mm, void *mqd,
q->is_active = QUEUE_IS_ACTIVE(*q);
}
-static uint32_t check_preemption_failed(void *mqd)
+static bool check_preemption_failed(struct mqd_manager *mm, void *mqd)
{
struct v10_compute_mqd *m = (struct v10_compute_mqd *)mqd;
- return m->queue_doorbell_id0;
+ return kfd_check_hiq_mqd_doorbell_id(mm->dev, m->queue_doorbell_id0, 0);
}
static int get_wave_state(struct mqd_manager *mm, void *mqd,
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c
index 2b9f57c267eb..68dbc0399c87 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v11.c
@@ -278,11 +278,11 @@ static void update_mqd(struct mqd_manager *mm, void *mqd,
q->is_active = QUEUE_IS_ACTIVE(*q);
}
-static uint32_t check_preemption_failed(void *mqd)
+static bool check_preemption_failed(struct mqd_manager *mm, void *mqd)
{
struct v11_compute_mqd *m = (struct v11_compute_mqd *)mqd;
- return m->queue_doorbell_id0;
+ return kfd_check_hiq_mqd_doorbell_id(mm->dev, m->queue_doorbell_id0, 0);
}
static int get_wave_state(struct mqd_manager *mm, void *mqd,
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c
index 7c93a0932677..6bddc16808d7 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_v9.c
@@ -316,11 +316,11 @@ static void update_mqd(struct mqd_manager *mm, void *mqd,
}
-static uint32_t check_preemption_failed(void *mqd)
+static bool check_preemption_failed(struct mqd_manager *mm, void *mqd)
{
struct v9_mqd *m = (struct v9_mqd *)mqd;
- return m->queue_doorbell_id0;
+ return kfd_check_hiq_mqd_doorbell_id(mm->dev, m->queue_doorbell_id0, 0);
}
static int get_wave_state(struct mqd_manager *mm, void *mqd,
@@ -607,6 +607,24 @@ static int destroy_hiq_mqd_v9_4_3(struct mqd_manager *mm, void *mqd,
return err;
}
+static bool check_preemption_failed_v9_4_3(struct mqd_manager *mm, void *mqd)
+{
+ uint64_t hiq_mqd_size = kfd_hiq_mqd_stride(mm->dev);
+ uint32_t xcc_mask = mm->dev->xcc_mask;
+ int inst = 0, xcc_id;
+ struct v9_mqd *m;
+ bool ret = false;
+
+ for_each_inst(xcc_id, xcc_mask) {
+ m = get_mqd(mqd + hiq_mqd_size * inst);
+ ret |= kfd_check_hiq_mqd_doorbell_id(mm->dev,
+ m->queue_doorbell_id0, inst);
+ ++inst;
+ }
+
+ return ret;
+}
+
static void get_xcc_mqd(struct kfd_mem_obj *mqd_mem_obj,
struct kfd_mem_obj *xcc_mqd_mem_obj,
uint64_t offset)
@@ -881,15 +899,16 @@ struct mqd_manager *mqd_manager_init_v9(enum KFD_MQD_TYPE type,
#if defined(CONFIG_DEBUG_FS)
mqd->debugfs_show_mqd = debugfs_show_mqd;
#endif
- mqd->check_preemption_failed = check_preemption_failed;
if (KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 3)) {
mqd->init_mqd = init_mqd_hiq_v9_4_3;
mqd->load_mqd = hiq_load_mqd_kiq_v9_4_3;
mqd->destroy_mqd = destroy_hiq_mqd_v9_4_3;
+ mqd->check_preemption_failed = check_preemption_failed_v9_4_3;
} else {
mqd->init_mqd = init_mqd_hiq;
mqd->load_mqd = kfd_hiq_load_mqd_kiq;
mqd->destroy_mqd = destroy_hiq_mqd;
+ mqd->check_preemption_failed = check_preemption_failed;
}
break;
case KFD_MQD_TYPE_DIQ:
diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
index dbc868e0363f..c1fafc502515 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_mqd_manager_vi.c
@@ -237,11 +237,11 @@ static void __update_mqd(struct mqd_manager *mm, void *mqd,
q->is_active = QUEUE_IS_ACTIVE(*q);
}
-static uint32_t check_preemption_failed(void *mqd)
+static bool check_preemption_failed(struct mqd_manager *mm, void *mqd)
{
struct vi_mqd *m = (struct vi_mqd *)mqd;
- return m->queue_doorbell_id0;
+ return kfd_check_hiq_mqd_doorbell_id(mm->dev, m->queue_doorbell_id0, 0);
}
static void update_mqd(struct mqd_manager *mm, void *mqd,
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.freedesktop.org/archives/amd-gfx/attachments/20240314/43122eec/attachment-0001.htm>
More information about the amd-gfx
mailing list