On Fri, Feb 10, 2023 at 10:37 AM Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote: > > On Fri, Feb 10, 2023 at 9:57 AM Andy Lutomirski <luto@xxxxxxxxxx> wrote: > > I'm not convinced your suggestion of extending io_uring with new > primitives is any better in practice, though. I don't know if I'm really suggesting new primitives. I think I'm making two change suggestions that go together. First, let splice() and IORING_OP_SPLICE copy (or zero-copy) data from a file to a socket. Second, either make splice more strict or add a new "strict splice" variant. Strict splice only completes when it can promise that writes to the source that start after strict splice's completion won't change what gets written to the destination. I think that strict splice fixes Stefan's use case. It's also easier to reason about than regular splice. The major caveat here is that zero-copy strict splice is fundamentally a potentially long-running operation in a way that zero-copy splice() isn't right now. So the combination of O_NONBLOCK and strict splice() (the syscall, not necessarily the io_uring operation) to something like a TCP socket requires complicated locking or change tracking to make sense. This means that a splice() syscall providing strict semantics to a TCP socket may just need to do a copy, at least in many cases. But maybe that's fine -- very-high-performance networking is moving pretty aggressively to io_uring anyway. And my possibly-quite-out-there claim is that, if Linux implements strict splice, maybe non-strict splice could get replaced in a user ABI-compatible manner with a much simpler non-zero-copy implementation. And strict splice from a file to a pipe could be implemented as a copy -- high performance users can, if needed, start strict-splicing from a file directly to a socket.
