lkml.org 
[lkml]   [2020]   [Jul]   [10]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
Patch in this message
/
From
Subject[PATCH 2/4] habanalabs: Assign each CQ with its own work queue
Date
From: Ofir Bitton <obitton@habana.ai>

We identified a possible race during job completion when working
with a single multi-threaded work queue. In order to overcome this
race we suggest using a single threaded work queue per completion
queue, hence we guarantee jobs completion in order.

Signed-off-by: Ofir Bitton <obitton@habana.ai>
Reviewed-by: Oded Gabbay <oded.gabbay@gmail.com>
Signed-off-by: Oded Gabbay <oded.gabbay@gmail.com>
---
drivers/misc/habanalabs/command_submission.c | 4 +-
drivers/misc/habanalabs/device.c | 39 ++++++++++++++++----
drivers/misc/habanalabs/habanalabs.h | 7 +++-
drivers/misc/habanalabs/irq.c | 2 +-
4 files changed, 40 insertions(+), 12 deletions(-)

diff --git a/drivers/misc/habanalabs/command_submission.c b/drivers/misc/habanalabs/command_submission.c
index 1ba937b9a22e..54f2f5afdd2a 100644
--- a/drivers/misc/habanalabs/command_submission.c
+++ b/drivers/misc/habanalabs/command_submission.c
@@ -487,10 +487,12 @@ static void cs_rollback(struct hl_device *hdev, struct hl_cs *cs)

void hl_cs_rollback_all(struct hl_device *hdev)
{
+ int i;
struct hl_cs *cs, *tmp;

/* flush all completions */
- flush_workqueue(hdev->cq_wq);
+ for (i = 0 ; i < hdev->asic_prop.completion_queues_count ; i++)
+ flush_workqueue(hdev->cq_wq[i]);

/* Make sure we don't have leftovers in the H/W queues mirror list */
list_for_each_entry_safe(cs, tmp, &hdev->hw_queues_mirror_list,
diff --git a/drivers/misc/habanalabs/device.c b/drivers/misc/habanalabs/device.c
index df709767c7ea..84800efec10d 100644
--- a/drivers/misc/habanalabs/device.c
+++ b/drivers/misc/habanalabs/device.c
@@ -249,7 +249,8 @@ static void device_cdev_sysfs_del(struct hl_device *hdev)
*/
static int device_early_init(struct hl_device *hdev)
{
- int rc;
+ int i, rc;
+ char workq_name[32];

switch (hdev->asic_type) {
case ASIC_GOYA:
@@ -274,11 +275,24 @@ static int device_early_init(struct hl_device *hdev)
if (rc)
goto early_fini;

- hdev->cq_wq = alloc_workqueue("hl-free-jobs", WQ_UNBOUND, 0);
- if (hdev->cq_wq == NULL) {
- dev_err(hdev->dev, "Failed to allocate CQ workqueue\n");
- rc = -ENOMEM;
- goto asid_fini;
+ if (hdev->asic_prop.completion_queues_count) {
+ hdev->cq_wq = kcalloc(hdev->asic_prop.completion_queues_count,
+ sizeof(*hdev->cq_wq),
+ GFP_ATOMIC);
+ if (!hdev->cq_wq) {
+ rc = -ENOMEM;
+ goto asid_fini;
+ }
+ }
+
+ for (i = 0 ; i < hdev->asic_prop.completion_queues_count ; i++) {
+ snprintf(workq_name, 32, "hl-free-jobs-%u", i);
+ hdev->cq_wq[i] = create_singlethread_workqueue(workq_name);
+ if (hdev->cq_wq == NULL) {
+ dev_err(hdev->dev, "Failed to allocate CQ workqueue\n");
+ rc = -ENOMEM;
+ goto free_cq_wq;
+ }
}

hdev->eq_wq = alloc_workqueue("hl-events", WQ_UNBOUND, 0);
@@ -321,7 +335,10 @@ static int device_early_init(struct hl_device *hdev)
free_eq_wq:
destroy_workqueue(hdev->eq_wq);
free_cq_wq:
- destroy_workqueue(hdev->cq_wq);
+ for (i = 0 ; i < hdev->asic_prop.completion_queues_count ; i++)
+ if (hdev->cq_wq[i])
+ destroy_workqueue(hdev->cq_wq[i]);
+ kfree(hdev->cq_wq);
asid_fini:
hl_asid_fini(hdev);
early_fini:
@@ -339,6 +356,8 @@ static int device_early_init(struct hl_device *hdev)
*/
static void device_early_fini(struct hl_device *hdev)
{
+ int i;
+
mutex_destroy(&hdev->mmu_cache_lock);
mutex_destroy(&hdev->debug_lock);
mutex_destroy(&hdev->send_cpu_message_lock);
@@ -351,7 +370,10 @@ static void device_early_fini(struct hl_device *hdev)
kfree(hdev->hl_chip_info);

destroy_workqueue(hdev->eq_wq);
- destroy_workqueue(hdev->cq_wq);
+
+ for (i = 0 ; i < hdev->asic_prop.completion_queues_count ; i++)
+ destroy_workqueue(hdev->cq_wq[i]);
+ kfree(hdev->cq_wq);

hl_asid_fini(hdev);

@@ -1181,6 +1203,7 @@ int hl_device_init(struct hl_device *hdev, struct class *hclass)
"failed to initialize completion queue\n");
goto cq_fini;
}
+ hdev->completion_queue[i].cq_idx = i;
}

/*
diff --git a/drivers/misc/habanalabs/habanalabs.h b/drivers/misc/habanalabs/habanalabs.h
index ea0fd178accb..01fb45887a5a 100644
--- a/drivers/misc/habanalabs/habanalabs.h
+++ b/drivers/misc/habanalabs/habanalabs.h
@@ -479,6 +479,7 @@ struct hl_hw_queue {
* @hdev: pointer to the device structure
* @kernel_address: holds the queue's kernel virtual address
* @bus_address: holds the queue's DMA address
+ * @cq_idx: completion queue index in array
* @hw_queue_id: the id of the matching H/W queue
* @ci: ci inside the queue
* @pi: pi inside the queue
@@ -488,6 +489,7 @@ struct hl_cq {
struct hl_device *hdev;
u64 kernel_address;
dma_addr_t bus_address;
+ u32 cq_idx;
u32 hw_queue_id;
u32 ci;
u32 pi;
@@ -1396,7 +1398,8 @@ struct hl_device_idle_busy_ts {
* @asic_name: ASIC specific nmae.
* @asic_type: ASIC specific type.
* @completion_queue: array of hl_cq.
- * @cq_wq: work queue of completion queues for executing work in process context
+ * @cq_wq: work queues of completion queues for executing work in process
+ * context.
* @eq_wq: work queue of event queue for executing work in process context.
* @kernel_ctx: Kernel driver context structure.
* @kernel_queues: array of hl_hw_queue.
@@ -1492,7 +1495,7 @@ struct hl_device {
char asic_name[16];
enum hl_asic_type asic_type;
struct hl_cq *completion_queue;
- struct workqueue_struct *cq_wq;
+ struct workqueue_struct **cq_wq;
struct workqueue_struct *eq_wq;
struct hl_ctx *kernel_ctx;
struct hl_hw_queue *kernel_queues;
diff --git a/drivers/misc/habanalabs/irq.c b/drivers/misc/habanalabs/irq.c
index 195a5ecba0e8..c8db717023f5 100644
--- a/drivers/misc/habanalabs/irq.c
+++ b/drivers/misc/habanalabs/irq.c
@@ -119,7 +119,7 @@ irqreturn_t hl_irq_handler_cq(int irq, void *arg)

if ((shadow_index_valid) && (!hdev->disabled)) {
job = queue->shadow_queue[hl_pi_2_offset(shadow_index)];
- queue_work(hdev->cq_wq, &job->finish_work);
+ queue_work(hdev->cq_wq[cq->cq_idx], &job->finish_work);
}

atomic_inc(&queue->ci);
--
2.17.1
\
 
 \ /
  Last update: 2020-07-10 19:37    [W:0.136 / U:0.232 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site