Jens Axboe <axboe@xxxxxxxxx> writes: > On 8/31/20 9:31 AM, Gabriel Krisman Bertazi wrote: >> According to Documentation/block/stat.rst, inflight should not include >> I/O requests that are in the queue but not yet dispatched to the device, >> but blk-mq identifies as inflight any request that has a tag allocated, >> which, for queues without elevator, happens at request allocation time >> and before it is queued in the ctx (default case in blk_mq_submit_bio). >> >> A more precise approach would be to only consider requests with state >> MQ_RQ_IN_FLIGHT. > > We've had some churn here, last change in this area was: Hi Jens, Thanks for the quick reply. I wasn't aware of that patch. I'm collecting statistics on the queue depth of a NCQ disk, and my end goal is to have a time_in_driver clock, which is the time spent by IOs in the driver, similarly to what is shown in diskstats, but the latter includes the block layer time. I stumbled this issue with inflight, where we noticed a difference between inflight and what was actually dispatched to the driver. The problem is the current behavior doesn't seem to match the documentation in Documentation/block/stat.sh. I went back to history.git and found that this difference in behavior from the documentation has always been there for legacy path, unless I am misreading the documentation. The patch I proposed consolidates inflight in favor of documentation instead of the legacy patch. The documentation reads: """ This value counts the number of I/O requests that have been issued to the device driver but have not yet completed. It does not include I/O requests that are in the queue but not yet issued to the device driver. """ Should I patch the documentation instead? Thinking about semantics, it seem more useful to have inflight include only the time between dispatching and completion, but if you think there is a concern about stable abi here, would you accept a new in_device file tracking this metric? -- Gabriel Krisman Bertazi
