Pretty much an RFC for executing BPF requests in io_uring to start a discussion, so any ideas and wish lists are welcom. Some technical but not much important for the discussion detais are omitted in the patchset, i.e. good cancellation system, optimisations and some BPF specific parts. Some of major missing points: 1. We need to pass an CQE/result of a previous linked request to BPF. 2. Whether to keep to a new IORING_OP_BPF opcode, or do it in a common path for each request, e.g. on CQE completion. The former looks saner, but is not nicely aligned with (1.). 3. Allow BPF programs to generate CQEs not binded to a request. A problem can be with supporting overflowed CQEs, it's either to always kmalloc()'ing storage for them instead of using req's memory or piling up on top. Eventually we will need it anyway to be able to post several CQEs for a single requests. Pavel Begunkov (4): bpf: add IOURING program type io_uring: implement bpf prog registration io_uring: add IORING_OP_BPF io_uring: enable BPF to submit SQEs fs/io_uring.c | 259 +++++++++++++++++++++++++++++++++- include/linux/bpf_types.h | 2 + include/uapi/linux/bpf.h | 2 + include/uapi/linux/io_uring.h | 3 + kernel/bpf/syscall.c | 1 + kernel/bpf/verifier.c | 3 + 6 files changed, 264 insertions(+), 6 deletions(-) -- 2.24.0
