On 12/01/2021 21:33, Bijan Mottahedeh wrote:
> v5:
>
> - call io_get_fixed_rsrc_ref for buffers
> - make percpu_ref_release names consistent
> - rebase on for-5.12/io_uring
To reduce the burden I'll take the generalisation patches from that,
review and resend to Jens with small changes leaving your "from:".
I hope you don't mind, that should be faster.
I'll remove your signed-off and will need it back by you replying
on this coming resend.
>
> v4:
>
> - address v3 comments (TBD REGISTER_BUFFERS)
> - rebase
>
> v3:
>
> - batch file->rsrc renames into a signle patch when possible
> - fix other review changes from v2
> - fix checkpatch warnings
>
> v2:
>
> - drop readv/writev with fixed buffers patch
> - handle ref_nodes both both files/buffers with a single ref_list
> - make file/buffer handling more unified
>
> This patchset implements a set of enhancements to buffer registration
> consistent with existing file registration functionality:
>
> - buffer registration updates IORING_REGISTER_BUFFERS_UPDATE
> IORING_OP_BUFFERS_UPDATE
>
> - buffer registration sharing IORING_SETUP_SHARE_BUF
> IORING_SETUP_ATTACH_BUF
>
> Patch 1-5 generalize fixed_file functionality to fixed_rsrc.
>
> Patch 6 applies fixed_rsrc functionality for fixed buffers support.
>
> Patch 7-8 generalize files_update functionality to rsrc_update.
>
> Patch 9 implements buffer registration update, and introduces
> IORING_REGISTER_BUFFERS_UPDATE and IORING_OP_BUFFERS_UPDATE, consistent
> with file registration update.
>
> Patch 10 generalizes fixed resource allocation
>
> Patch 11 renames percpu release routines for consistency
>
> Patch 12 calls io_get_fixed_rsrc_ref() for buffers as well as files
>
> Patch 13 implements buffer sharing among multiple rings; it works as follows:
>
> - A new ring, A, is setup. Since no buffers have been registered, the
> registered buffer state is an empty set, Z. That's different from the
> NULL state in current implementation.
>
> - Ring B is setup, attaching to Ring A. It's also attaching to it's
> buffer registrations, now we have two references to the same empty
> set, Z.
>
> - Ring A registers buffers into set Z, which is no longer empty.
>
> - Ring B sees this immediately, since it's already sharing that set.
>
> Testing
>
> I have used liburing file-{register,update} tests as models for
> buffer-{register,update,share}, tests and they run ok.
>
> TBD
>
> - Need a patch from Pavel to address a race between fixed IO from async
> context and buffer unregister, or force buffer registration ops to do
> full quiesce.
>
> - Need to still address Pavel's comments about unkillable deadlocks. It
> seems that we should no long hange unkillably in io_rsrc_ref_quiesce()
> because of Pavel's changes.
>
> - I tried to use a single opcode for files/buffers but ran into an
> issue since work_flags is different for files/buffers. This should
> be ok for the most part since req->work.flags is ultimately examined;
> however, there are place where io_op_defs[opcode].work_flags is examined
> directly, and I wasn't sure what would the best way to handle that.
>
> Bijan Mottahedeh (13):
> io_uring: rename file related variables to rsrc
> io_uring: generalize io_queue_rsrc_removal
> io_uring: separate ref_list from fixed_rsrc_data
> io_uring: split alloc_fixed_file_ref_node
> io_uring: add rsrc_ref locking routines
> io_uring: implement fixed buffers registration similar to fixed files
> io_uring: create common fixed_rsrc_ref_node handling routines
> io_uring: generalize files_update functionlity to rsrc_update
> io_uring: support buffer registration updates
> io_uring: create common fixed_rsrc_data allocation routines
> io_uring: make percpu_ref_release names consistent
> io_uring: call io_get_fixed_rsrc_ref for buffers
> io_uring: support buffer registration sharing
>
> fs/io_uring.c | 803 ++++++++++++++++++++++++++++++++----------
> include/uapi/linux/io_uring.h | 13 +-
> 2 files changed, 626 insertions(+), 190 deletions(-)
>
--
Pavel Begunkov
