On Tue, Aug 01, 2023 at 03:13:59PM +0100, Pavel Begunkov wrote: > On 7/31/23 22:00, Jens Axboe wrote: > > On 7/31/23 6:53?AM, Pavel Begunkov wrote: > > > On 7/28/23 21:14, Keith Busch wrote: > > > > From: Keith Busch <kbusch@xxxxxxxxxx> > > > > > > > > Split the req initialization and link handling from the submit. This > > > > simplifies the submit path since everything that can fail is separate > > > > from it, and makes it easier to create batched submissions later. > > > > > > Keith, I don't think this prep patch does us any good, I'd rather > > > shove the link assembling code further out of the common path. I like > > > the first version more (see [1]). I'd suggest to merge it, and do > > > cleaning up after. > > > > > > I'll also say that IMHO the overhead is well justified. It's not only > > > about having multiple nvmes, the problem slows down cases mixing storage > > > with net and the rest of IO in a single ring. > > > > > > [1] https://lore.kernel.org/io-uring/20230504162427.1099469-1-kbusch@xxxxxxxx/ > > > > The downside of that one, to me, is that it just serializes all of it > > and we end up looping over the submission list twice. > > Right, and there is nothing can be done if we want to know about all > requests in advance, at least without changing uapi and/or adding > userspace hints. > > > With alloc+init > > split, at least we get some locality wins by grouping the setup side of > > the requests. > > I don't think I follow, what grouping do you mean? As far as I see, v1 > and v2 are essentially same with the difference of whether you have a > helper for setting up links or not, see io_setup_link() from v2. In both > cases it's executed in the same sequence: > > 1) init (generic init + opcode init + link setup) each request and put > into a temporary list. > 2) go go over the list and submit them one by one > > And after inlining they should look pretty close. The main difference in this one compared to the original version is that everything in the 2nd loop is just for the final dispatch. Anything that can fail, fallback, or defer to async happens in the first loop. I'm not sure that makes a difference in runtime, but having the 2nd loop handle only fast-path requests was what I set out to do for this version.
