Messages in this thread Patch in this message | | | From | Oded Gabbay <> | Subject | [PATCH 05/12] habanalabs: fix race condition in multi CS completion | Date | Sun, 28 Nov 2021 21:34:28 +0200 |
| |
From: Dani Liberman <dliberman@habana.ai>
Race example scenario: 1. User have 2 threads that waits on multi CS: - thread_0 waits on QID 0 and uses multi CS context 0. - thread_1 waits on QID 1 and uses multi CS context 1. 2. thread_1 got completion and release multi CS context 1. 3. CS related to multi CS of thread_0 starts executing complete_multi_cs function, the first iteration of the loop completes the multi CS of thread_0, hence multi CS context 0 is released. 4. thread_1 waits on QID 1 and uses multi CS context 0. 5. thread_0 waits on QID 0 and uses multi CS context 1. 6. The second iterattion of the loop (from step 3) starts, which means, start checking multi CS context 1: - multi CS contetxt is being used by thread_0 waiting on QID 0. - The fence of the CS (still CS from step 3) has QID map the same as the multi CS context 1. - multi CS context 1 (thread_0) gets completion on CS that triggered already thread_0 (with multi CS context 0) and is no longer being waited on.
Fixed by exiting the loop in complete_multi_cs after getting completion
Signed-off-by: Dani Liberman <dliberman@habana.ai> Reviewed-by: Oded Gabbay <ogabbay@kernel.org> Signed-off-by: Oded Gabbay <ogabbay@kernel.org> --- drivers/misc/habanalabs/common/command_submission.c | 7 +++++++ 1 file changed, 7 insertions(+)
diff --git a/drivers/misc/habanalabs/common/command_submission.c b/drivers/misc/habanalabs/common/command_submission.c index c1fd4ba14c60..4e893364a3cc 100644 --- a/drivers/misc/habanalabs/common/command_submission.c +++ b/drivers/misc/habanalabs/common/command_submission.c @@ -545,6 +545,13 @@ static void complete_multi_cs(struct hl_device *hdev, struct hl_cs *cs) * mcs fences. */ fence->mcs_handling_done = true; + /* + * Since CS (and its related fence) can be associated with only one + * multi CS context, once it triggered multi CS completion no need to + * continue checking other multi CS contexts. + */ + spin_unlock(&mcs_compl->lock); + break; } spin_unlock(&mcs_compl->lock); -- 2.25.1
| |