[PATCH 9/9] habanalabs: use queue pi/ci in order to determine queue occupancy
From: Oded Gabbay
Date: Sun Jul 05 2020 - 09:13:32 EST
From: Ofir Bitton <obitton@xxxxxxxxx>
Instead of using the free slots amount on the compute CQ to determine
whether we can submit work to queues, use the queues pi/ci.
This is needed in future ASICs where we don't have CQ per queue.
Signed-off-by: Ofir Bitton <obitton@xxxxxxxxx>
Signed-off-by: Oded Gabbay <oded.gabbay@xxxxxxxxx>
---
drivers/misc/habanalabs/device.c | 17 +++---
drivers/misc/habanalabs/habanalabs.h | 2 +-
drivers/misc/habanalabs/hw_queue.c | 82 +++++++++-------------------
drivers/misc/habanalabs/irq.c | 7 +--
4 files changed, 39 insertions(+), 69 deletions(-)
diff --git a/drivers/misc/habanalabs/device.c b/drivers/misc/habanalabs/device.c
index 2b38a119704c..65a5a5c52a48 100644
--- a/drivers/misc/habanalabs/device.c
+++ b/drivers/misc/habanalabs/device.c
@@ -1144,14 +1144,17 @@ int hl_device_init(struct hl_device *hdev, struct class *hclass)
* because there the addresses of the completion queues are being
* passed as arguments to request_irq
*/
- hdev->completion_queue = kcalloc(cq_cnt,
- sizeof(*hdev->completion_queue),
- GFP_KERNEL);
+ if (cq_cnt) {
+ hdev->completion_queue = kcalloc(cq_cnt,
+ sizeof(*hdev->completion_queue),
+ GFP_KERNEL);
- if (!hdev->completion_queue) {
- dev_err(hdev->dev, "failed to allocate completion queues\n");
- rc = -ENOMEM;
- goto hw_queues_destroy;
+ if (!hdev->completion_queue) {
+ dev_err(hdev->dev,
+ "failed to allocate completion queues\n");
+ rc = -ENOMEM;
+ goto hw_queues_destroy;
+ }
}
for (i = 0, cq_ready_cnt = 0 ; i < cq_cnt ; i++, cq_ready_cnt++) {
diff --git a/drivers/misc/habanalabs/habanalabs.h b/drivers/misc/habanalabs/habanalabs.h
index 9213d107b533..a61aab09778c 100644
--- a/drivers/misc/habanalabs/habanalabs.h
+++ b/drivers/misc/habanalabs/habanalabs.h
@@ -461,7 +461,7 @@ struct hl_hw_queue {
u64 kernel_address;
dma_addr_t bus_address;
u32 pi;
- u32 ci;
+ atomic_t ci;
u32 hw_queue_id;
u32 cq_id;
u32 msi_vec;
diff --git a/drivers/misc/habanalabs/hw_queue.c b/drivers/misc/habanalabs/hw_queue.c
index 7965551587fc..474a0e8a7797 100644
--- a/drivers/misc/habanalabs/hw_queue.c
+++ b/drivers/misc/habanalabs/hw_queue.c
@@ -23,10 +23,14 @@ inline u32 hl_hw_queue_add_ptr(u32 ptr, u16 val)
ptr &= ((HL_QUEUE_LENGTH << 1) - 1);
return ptr;
}
+static inline int queue_ci_get(atomic_t *ci, u32 queue_len)
+{
+ return atomic_read(ci) & ((queue_len << 1) - 1);
+}
static inline int queue_free_slots(struct hl_hw_queue *q, u32 queue_len)
{
- int delta = (q->pi - q->ci);
+ int delta = (q->pi - queue_ci_get(&q->ci, queue_len));
if (delta >= 0)
return (queue_len - delta);
@@ -40,21 +44,14 @@ void hl_int_hw_queue_update_ci(struct hl_cs *cs)
struct hl_hw_queue *q;
int i;
- hdev->asic_funcs->hw_queues_lock(hdev);
-
if (hdev->disabled)
- goto out;
+ return;
q = &hdev->kernel_queues[0];
for (i = 0 ; i < hdev->asic_prop.max_queues ; i++, q++) {
- if (q->queue_type == QUEUE_TYPE_INT) {
- q->ci += cs->jobs_in_queue_cnt[i];
- q->ci &= ((q->int_queue_len << 1) - 1);
- }
+ if (q->queue_type == QUEUE_TYPE_INT)
+ atomic_add(cs->jobs_in_queue_cnt[i], &q->ci);
}
-
-out:
- hdev->asic_funcs->hw_queues_unlock(hdev);
}
/*
@@ -174,38 +171,26 @@ static int int_queue_sanity_checks(struct hl_device *hdev,
}
/*
- * hw_queue_sanity_checks() - Perform some sanity checks on a H/W queue.
+ * hw_queue_sanity_checks() - Make sure we have enough space in the h/w queue
* @hdev: Pointer to hl_device structure.
* @q: Pointer to hl_hw_queue structure.
* @num_of_entries: How many entries to check for space.
*
- * Perform the following:
- * - Make sure we have enough space in the completion queue.
- * This check also ensures that there is enough space in the h/w queue, as
- * both queues are of the same size.
- * - Reserve space in the completion queue (needs to be reversed if there
- * is a failure down the road before the actual submission of work).
+ * Notice: We do not reserve queue entries so this function mustn't be called
+ * more than once per CS for the same queue
*
- * Both operations are done using the "free_slots_cnt" field of the completion
- * queue. The CI counters of the queue and the completion queue are not
- * needed/used for the H/W queue type.
*/
static int hw_queue_sanity_checks(struct hl_device *hdev, struct hl_hw_queue *q,
int num_of_entries)
{
- atomic_t *free_slots =
- &hdev->completion_queue[q->cq_id].free_slots_cnt;
+ int free_slots_cnt;
- /*
- * Check we have enough space in the completion queue.
- * Add -1 to counter (decrement) unless counter was already 0.
- * In that case, CQ is full so we can't submit a new CB.
- * atomic_add_unless will return 0 if counter was already 0.
- */
- if (atomic_add_negative(num_of_entries * -1, free_slots)) {
- dev_dbg(hdev->dev, "No space for %d entries on CQ %d\n",
- num_of_entries, q->hw_queue_id);
- atomic_add(num_of_entries, free_slots);
+ /* Check we have enough space in the queue */
+ free_slots_cnt = queue_free_slots(q, HL_QUEUE_LENGTH);
+
+ if (free_slots_cnt < num_of_entries) {
+ dev_dbg(hdev->dev, "Queue %d doesn't have room for %d CBs\n",
+ q->hw_queue_id, num_of_entries);
return -EAGAIN;
}
@@ -366,7 +351,6 @@ static void hw_queue_schedule_job(struct hl_cs_job *job)
{
struct hl_device *hdev = job->cs->ctx->hdev;
struct hl_hw_queue *q = &hdev->kernel_queues[job->hw_queue_id];
- struct hl_cq *cq;
u64 ptr;
u32 offset, ctl, len;
@@ -395,17 +379,6 @@ static void hw_queue_schedule_job(struct hl_cs_job *job)
else
ptr = (u64) (uintptr_t) job->user_cb;
- /*
- * No need to protect pi_offset because scheduling to the
- * H/W queues is done under the scheduler mutex
- *
- * No need to check if CQ is full because it was already
- * checked in hw_queue_sanity_checks
- */
- cq = &hdev->completion_queue[q->cq_id];
-
- cq->pi = hl_cq_inc_ptr(cq->pi);
-
ext_and_hw_queue_submit_bd(hdev, q, ctl, len, ptr);
}
@@ -552,8 +525,7 @@ int hl_hw_queue_schedule_cs(struct hl_cs *cs)
goto unroll_cq_resv;
}
- if (q->queue_type == QUEUE_TYPE_EXT ||
- q->queue_type == QUEUE_TYPE_HW)
+ if (q->queue_type == QUEUE_TYPE_EXT)
cq_cnt++;
}
}
@@ -605,9 +577,8 @@ int hl_hw_queue_schedule_cs(struct hl_cs *cs)
unroll_cq_resv:
q = &hdev->kernel_queues[0];
for (i = 0 ; (i < max_queues) && (cq_cnt > 0) ; i++, q++) {
- if ((q->queue_type == QUEUE_TYPE_EXT ||
- q->queue_type == QUEUE_TYPE_HW) &&
- cs->jobs_in_queue_cnt[i]) {
+ if ((q->queue_type == QUEUE_TYPE_EXT) &&
+ (cs->jobs_in_queue_cnt[i])) {
atomic_t *free_slots =
&hdev->completion_queue[i].free_slots_cnt;
atomic_add(cs->jobs_in_queue_cnt[i], free_slots);
@@ -631,7 +602,7 @@ void hl_hw_queue_inc_ci_kernel(struct hl_device *hdev, u32 hw_queue_id)
{
struct hl_hw_queue *q = &hdev->kernel_queues[hw_queue_id];
- q->ci = hl_queue_inc_ptr(q->ci);
+ atomic_inc(&q->ci);
}
static int ext_and_cpu_queue_init(struct hl_device *hdev, struct hl_hw_queue *q,
@@ -666,7 +637,7 @@ static int ext_and_cpu_queue_init(struct hl_device *hdev, struct hl_hw_queue *q,
}
/* Make sure read/write pointers are initialized to start of queue */
- q->ci = 0;
+ atomic_set(&q->ci, 0);
q->pi = 0;
return 0;
@@ -700,7 +671,7 @@ static int int_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
q->kernel_address = (u64) (uintptr_t) p;
q->pi = 0;
- q->ci = 0;
+ atomic_set(&q->ci, 0);
return 0;
}
@@ -729,7 +700,7 @@ static int hw_queue_init(struct hl_device *hdev, struct hl_hw_queue *q)
q->kernel_address = (u64) (uintptr_t) p;
/* Make sure read/write pointers are initialized to start of queue */
- q->ci = 0;
+ atomic_set(&q->ci, 0);
q->pi = 0;
return 0;
@@ -931,7 +902,8 @@ void hl_hw_queue_reset(struct hl_device *hdev, bool hard_reset)
if ((!q->valid) ||
((!hard_reset) && (q->queue_type == QUEUE_TYPE_CPU)))
continue;
- q->pi = q->ci = 0;
+ q->pi = 0;
+ atomic_set(&q->ci, 0);
if (q->supports_sync_stream)
sync_stream_queue_reset(hdev, q->hw_queue_id);
diff --git a/drivers/misc/habanalabs/irq.c b/drivers/misc/habanalabs/irq.c
index 7a4878edb1a3..195a5ecba0e8 100644
--- a/drivers/misc/habanalabs/irq.c
+++ b/drivers/misc/habanalabs/irq.c
@@ -122,12 +122,7 @@ irqreturn_t hl_irq_handler_cq(int irq, void *arg)
queue_work(hdev->cq_wq, &job->finish_work);
}
- /* Update ci of the context's queue. There is no
- * need to protect it with spinlock because this update is
- * done only inside IRQ and there is a different IRQ per
- * queue
- */
- queue->ci = hl_queue_inc_ptr(queue->ci);
+ atomic_inc(&queue->ci);
/* Clear CQ entry ready bit */
cq_entry->data = cpu_to_le32(le32_to_cpu(cq_entry->data) &
--
2.17.1