On 9/15/21 1:26 PM, Linus Torvalds wrote: > On Wed, Sep 15, 2021 at 11:46 AM Jens Axboe <axboe@xxxxxxxxx> wrote: >> >> The usual tests >> do end up hitting the -EAGAIN path quite easily for certain device >> types, but not the short read/write. > > No way to do something like "read in file to make sure it's cached, > then invalidate caches from position X with POSIX_FADV_DONTNEED, then > do a read that crosses that cached/uncached boundary"? > > To at least verify that "partly synchronous, but partly punted to > async" case? > > Or were you talking about some other situation? No that covers some of it, and that happens naturally with buffered IO. The typical case is -EAGAIN on the first try, then you get a partial or all of it the next loop, and then done or continue. I tend to run fio verification workloads for that, as you get all the flexibility of fio with the data verification. And there are tests in there that run DONTNEED in parallel with buffered IO, exactly to catch some of these csaes. But they don't verify the data, generally. In that sense buffered is a lot easier than O_DIRECT, as it's easier to provoke these cases. And that does hit all the save/restore parts and looping, and if you do it with registered buffers then you get to work with bvec iter as well. O_DIRECT may get you -EAGAIN for low queue depth devices, but it'll never do a short read/write after that. But that's not in the regressions tests. I'll write a test case that can go with the liburing regressions for it. -- Jens Axboe
