On Fri, Sep 10, 2021 at 03:15:35AM +0000, Al Viro wrote: > On Thu, Sep 09, 2021 at 09:06:58PM -0600, Jens Axboe wrote: > > On 9/9/21 8:48 PM, Al Viro wrote: > > > On Thu, Sep 09, 2021 at 07:35:13PM -0600, Jens Axboe wrote: > > > > > >> Yep ok I follow you now. And yes, if we get a partial one but one that > > >> has more consumed than what was returned, that would not work well. I'm > > >> guessing that a) we've never seen that, or b) we always end up with > > >> either correctly advanced OR fully advanced, and the fully advanced case > > >> would then just return 0 next time and we'd just get a short IO back to > > >> userspace. > > >> > > >> The safer way here would likely be to import the iovec again. We're > > >> still in the context of the original submission, and the sqe hasn't been > > >> consumed in the ring yet, so that can be done safely. > > > > > > ... until you end up with something assuming that you've got the same > > > iovec from userland the second time around. > > > > > > IOW, generally it's a bad idea to do that kind of re-imports. > > > > That's really no different than having one thread do the issue, and > > another modify the iovec while it happens. It's only an issue if you > > don't validate it, just like you did the first time you imported. No > > assumptions need to be made here. > > It's not "need to be made", it's "will be mistakenly made by > somebody several years down the road"... E.g. somebody blindly assuming that the amount of data read the last time around will not exceed the size of reimported iov_iter. What I'm saying is that there's a plenty of ways to fuck up in that direction, and they will *not* be caught by normal fuzzers. I'm not arguing in favour of an uncoditional copy, BTW - I would like to see something resembling profiling data, but it's obviously not a pretty solution.
