habanalabs: signal/wait change sync object reset flow

Currently the SOB reset was in fence release function which happens
only at the CS wraparound during the CS allocation time.

In order to support the new encapsulated signals reservation feature,
we need to move the SOB reset to an earlier phase because this SOB
could reach it's max value very fast using the signal reservation.

Signed-off-by: farah kassabri <fkassabri@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>

1 files changed, 26 insertions, 8 deletions

diff --git a/drivers/misc/habanalabs/gaudi/gaudi.c b/drivers/misc/habanalabs/gaudi/gaudi.c
index a4b33b0b17d4..5b7a5692cd21 100644
--- a/drivers/misc/habanalabs/gaudi/gaudi.c
+++ b/drivers/misc/habanalabs/gaudi/gaudi.c
@@ -1239,7 +1239,7 @@ static void gaudi_collective_slave_init_job(struct hl_device *hdev,
 			prop->collective_sob_id, cb_size, false);
 }
 
-static void gaudi_collective_wait_init_cs(struct hl_cs *cs)
+static int gaudi_collective_wait_init_cs(struct hl_cs *cs)
 {
 	struct hl_cs_compl *signal_cs_cmpl =
 		container_of(cs->signal_fence, struct hl_cs_compl, base_fence);
@@ -1261,6 +1261,29 @@ static void gaudi_collective_wait_init_cs(struct hl_cs *cs)
 	cs_cmpl->hw_sob = signal_cs_cmpl->hw_sob;
 	cs_cmpl->sob_val = signal_cs_cmpl->sob_val;
 
+	/* check again if the signal cs already completed.
+	 * if yes then don't send any wait cs since the hw_sob
+	 * could be in reset already. if signal is not completed
+	 * then get refcount to hw_sob to prevent resetting the sob
+	 * while wait cs is not submitted.
+	 * note that this check is protected by two locks,
+	 * hw queue lock and completion object lock,
+	 * and the same completion object lock also protects
+	 * the hw_sob reset handler function.
+	 * The hw_queue lock prevent out of sync of hw_sob
+	 * refcount value, changed by signal/wait flows.
+	 */
+	spin_lock(&signal_cs_cmpl->lock);
+
+	if (completion_done(&cs->signal_fence->completion)) {
+		spin_unlock(&signal_cs_cmpl->lock);
+		return -EINVAL;
+	}
+	/* Increment kref since all slave queues are now waiting on it */
+	kref_get(&cs_cmpl->hw_sob->kref);
+
+	spin_unlock(&signal_cs_cmpl->lock);
+
 	/* Calculate the stream from collective master queue (1st job) */
 	job = list_first_entry(&cs->job_list, struct hl_cs_job, cs_node);
 	stream = job->hw_queue_id % 4;
@@ -1304,16 +1327,11 @@ static void gaudi_collective_wait_init_cs(struct hl_cs *cs)
 				cprop->curr_sob_group_idx[stream], stream);
 	}
 
-	/* Increment kref since all slave queues are now waiting on it */
-	kref_get(&cs_cmpl->hw_sob->kref);
-	/*
-	 * Must put the signal fence after the SOB refcnt increment so
-	 * the SOB refcnt won't turn 0 and reset the SOB before the
-	 * wait CS was submitted.
-	 */
 	mb();
 	hl_fence_put(cs->signal_fence);
 	cs->signal_fence = NULL;
+
+	return 0;
 }
 
 static int gaudi_collective_wait_create_job(struct hl_device *hdev,