> Looking at the previous patch, get_token_bucket_throttle() blocks > waiting for the requested tokens to become available. This is not > going to work because ceph_osdc_start_request() is called from all > sorts of places, including the messenger worker threads. > > ceph_osdc_start_request() should never fail. Treat it as a void > function -- the return value is ignored in new code and the signature > will eventually be updated to *req -> void. > > Thanks, > > Ilya Hi, llya, thanks for your review:-) We've already run this code in one of our cluster. It does work. ceph_osdc_start_request() will not fail, because we treat tb->max being 0 as shutting down the controller, and made sure that, before setting tb->max to non-zero value, token bucket timer work is already started and tb->throughput is not zero too. So if tb->max is 0, the controller is shut down, and ceph_osdc_start_request() would behave just like the code without the controller; if tb->max is not 0, the token will be filled when the token bucket timer fires, in which case ceph_osdc_start_request() would not fail, either. Am i right? Thanks:-)
