Commit 7a78d4d4 authored by Oded Gabbay's avatar Oded Gabbay
Browse files

habanalabs: fix race between wait and irq 



There is a race in the user interrupts code, where between checking
the target value and adding the new pend to the list, there is a chance
the interrupt happened.

In that case, no one will complete the node, and we will get a timeout
on it.

Signed-off-by: default avatarOded Gabbay <ogabbay@kernel.org>
parent 54faa560

drivers/misc/habanalabs/common/command_submission.c

+5 −1
Original line number Diff line number Diff line
@@ -2892,16 +2892,21 @@ static int _hl_interrupt_wait_ioctl(struct hl_device *hdev, struct hl_ctx *ctx, 
	pend->cq_kernel_addr = (u64 *) cb->kernel_address + cq_counters_offset;

	pend->cq_target_value = target_value;



	spin_lock_irqsave(&interrupt->wait_list_lock, flags);



	/* We check for completion value as interrupt could have been received

	 * before we added the node to the wait list

	 */

	if (*pend->cq_kernel_addr >= target_value) {

		spin_unlock_irqrestore(&interrupt->wait_list_lock, flags);



		*status = HL_WAIT_CS_STATUS_COMPLETED;

		/* There was no interrupt, we assume the completion is now. */

		pend->fence.timestamp = ktime_get();

		goto set_timestamp;



	} else if (!timeout_us) {

		spin_unlock_irqrestore(&interrupt->wait_list_lock, flags);

		*status = HL_WAIT_CS_STATUS_BUSY;

		pend->fence.timestamp = ktime_get();

		goto set_timestamp;
@@ -2910,7 +2915,6 @@ static int _hl_interrupt_wait_ioctl(struct hl_device *hdev, struct hl_ctx *ctx, 
	/* Add pending user interrupt to relevant list for the interrupt

	 * handler to monitor

	 */

	spin_lock_irqsave(&interrupt->wait_list_lock, flags);

	list_add_tail(&pend->wait_list_node, &interrupt->wait_list_head);

	spin_unlock_irqrestore(&interrupt->wait_list_lock, flags);