Ewan, > Delaying the queue re-run vs. a ramp down might negatively affect > performance. I'm not sure how accurate the ramp is at discovering the > optimum though. The optimum - as well as the actual limit - might change over time in a multi-initiator setup as other hosts grab resources on the device. I do think that the only way forward here, if we want to avoid counting outstanding commands for performance reasons, is to ensure that the BUSY/QUEUE_FULL/TASK_SET_FULL handling is relatively fast path and not something deep in the bowels of error handling. Which it actually isn't. But I do think we'll need to take a closer look. -- Martin K. Petersen Oracle Linux Engineering
