> On Sep 22, 2020, at 2:01 AM, Arnd Bergmann <arnd@xxxxxxxx> wrote: > > On Tue, Sep 22, 2020 at 9:59 AM Pavel Begunkov <asml.silence@xxxxxxxxx> wrote: >>> On 22/09/2020 10:23, Arnd Bergmann wrote: >>> On Tue, Sep 22, 2020 at 8:32 AM Pavel Begunkov <asml.silence@xxxxxxxxx> wrote: >>>> On 22/09/2020 03:58, Andy Lutomirski wrote: >>>>> On Mon, Sep 21, 2020 at 5:24 PM Pavel Begunkov <asml.silence@xxxxxxxxx> wrote: >>>>> I may be looking at a different kernel than you, but aren't you >>>>> preventing creating an io_uring regardless of whether SQPOLL is >>>>> requested? >>>> >>>> I diffed a not-saved file on a sleepy head, thanks for noticing. >>>> As you said, there should be an SQPOLL check. >>>> >>>> ... >>>> if (ctx->compat && (p->flags & IORING_SETUP_SQPOLL)) >>>> goto err; >>> >>> Wouldn't that mean that now 32-bit containers behave differently >>> between compat and native execution? >>> >>> I think if you want to prevent 32-bit applications from using SQPOLL, >>> it needs to be done the same way on both to be consistent: >> >> The intention was to disable only compat not native 32-bit. > > I'm not following why that would be considered a valid option, > as that clearly breaks existing users that update from a 32-bit > kernel to a 64-bit one. > > Taking away the features from users that are still on 32-bit kernels > already seems questionable to me, but being inconsistent > about it seems much worse, in particular when the regression > is on the upgrade path. > >>> Can we expect all existing and future user space to have a sane >>> fallback when IORING_SETUP_SQPOLL fails? >> >> SQPOLL has a few differences with non-SQPOLL modes, but it's easy >> to convert between them. Anyway, SQPOLL is a privileged special >> case that's here for performance/latency reasons, I don't think >> there will be any non-accidental users of it. > > Ok, so the behavior of 32-bit tasks would be the same as running > the same application as unprivileged 64-bit tasks, with applications > already having to implement that fallback, right? > > I don’t have any real preference wrt SQPOLL, and it may be that we have a problem even without SQPOLL when IO gets punted without one of the fixes discussed. But banning the mismatched io_uring and io_uring_enter seems like it may be worthwhile regardless.
