On Thu, Mar 09, 2023 at 07:58:31PM -0700, Jens Axboe wrote: > On 3/9/23 6:38?PM, Ming Lei wrote: > > On Thu, Mar 09, 2023 at 08:48:08AM -0500, Stefan Hajnoczi wrote: > >> Hi, > >> For block I/O an application can queue excess SQEs in userspace when the > >> SQ ring becomes full. For network and IPC operations that is not > >> possible because deadlocks can occur when socket, pipe, and eventfd SQEs > >> cannot be submitted. > > > > Can you explain a bit the deadlock in case of network application? io_uring > > does support to queue many network SQEs via IOSQE_IO_LINK, at least for > > send. > > > >> > >> Sometimes the application does not know how many SQEs/CQEs are needed upfront > >> and that's when we face this challenge. > > > > When running out of SQEs, the application can call io_uring_enter() to submit > > queued SQEs immediately without waiting for get events, then once > > io_uring_enter() returns, you get free SQEs for moving one. > > > >> > >> A simple solution is to call io_uring_setup(2) with a higher entries > >> value than you'll ever need. However, if that value is exceeded then > >> we're back to the deadlock scenario and that worries me. > > > > Can you please explain the deadlock scenario? > > I'm also curious of what these deadlocks are. As Ming says, you > generally never run out of SQEs as you can always just submit what you > have pending and now you have a full queue size worth of them available > again. > In my limited io_uring experiments it was convenient to know I could *always* get+prepare N number of concurrent SQEs before having to submit. I was working with a set of files I needed to first discover the quantity of, so I would start with a bootstrap ring size sufficient for the discovery process. Then once known, I'd resize the ring to accomodate the maximum width of SEQs N files could produce for the given operation. The convenience was it made the dispatch functions logically atomic units. In the sense that they didn't need to be able to handle running out of SQEs, submitting, and resuming in a continuation style. They could just be coded simply in a single loop iterating across the N files getting+preparing SQEs, confident they wouldn't "deadlock" from exhaustion. Perhaps that's a similar "deadlock" scenario to Ming's. But I should note that in my experiments I was always operating under the assumption that I'd never have N so large it couldn't possibly exceed the maximum SQ size I could allocate. And that probably isn't a safe assumption for a real production program, I was just experimenting after all. Also I was able to "resize" by just quiescing the ring, destroying it, and recreating it with the new size. It wasn't a perf sensitive thing, just startup rigamarole. I do recall being a little surprised I had to ad-hoc implement the resize at the time though... Regards, Vito Caputo
