On 12/07/2020 23:32, Jens Axboe wrote: > On 7/12/20 11:34 AM, Pavel Begunkov wrote: >> On 12/07/2020 18:59, Jens Axboe wrote: >>> On 7/12/20 3:41 AM, Pavel Begunkov wrote: >>>> Make io_kiocb slimmer by 24 bytes mainly by revising lists usage. The >>>> drawback is adding extra kmalloc in draining path, but that's a slow >>>> path, so meh. It also frees some space for the deferred completion path >>>> if would be needed in the future, but the main idea here is to shrink it >>>> to 3 cachelines in the end. >>>> >>>> I'm not happy yet with a few details, so that's not final, but it would >>>> be lovely to hear some feedback. >>> >>> I think it looks pretty good, most of the changes are straight forward. >>> Adding a completion entry that shares the submit space is a good idea, >>> and really helps bring it together. >>> >>> From a quick look, the only part I'm not super crazy about is patch #3. >> >> Thanks! >> >>> I'd probably rather use a generic list name and not unionize the tw >>> lists. >> >> I don't care much, but without compiler's help always have troubles >> finding and distinguishing something as generic as "list". > > To me, it's easier to verify that we're doing the right thing when they > use the same list member. Otherwise you have to cross reference two > different names, easier to shoot yourself in the foot that way. So I'd > prefer just retaining it as 'list' or something generic. If you don't have objections, I'll just leave it "inflight_entry". This one is easy to grep. >> BTW, I thought out how to bring it down to 3 cache lines, but that would >> require taking io_wq_work out of io_kiocb and kmalloc'ing it on demand. >> And there should also be a bunch of nice side effects like improving apoll. > > How would this work with the current use of io_wq_work as storage for > whatever bits we're hanging on to? I guess it could work with a prep > series first more cleanly separating it, though I do feel like we've > been getting closer to that already. It's definitely not a single patch. I'm going to prepare a series for discussion later, and then we'll see whether it worths it. > Definitely always interested in shrinking io_kiocb, just need to keep > an eye out for the fast(er) paths not needing two allocations (and > frees) for a single request. -- Pavel Begunkov
