On Fri, Mar 31, 2023 at 07:13:21PM +0000, Bernd Schubert wrote: > On 3/30/23 13:36, Ming Lei wrote: > > Apply io_uring fused command for supporting zero copy: > > > > 1) init the fused cmd buffer(io_mapped_buf) in ublk_map_io(), and deinit it > > in ublk_unmap_io(), and this buffer is immutable, so it is just fine to > > retrieve it from concurrent fused command. > > > > 1) add sub-command opcode of UBLK_IO_FUSED_SUBMIT_IO for retrieving this > > fused cmd(zero copy) buffer > > > > 2) call io_fused_cmd_start_secondary_req() to provide buffer to secondary > > request and submit secondary request; meantime setup complete callback via > > this API, once secondary request is completed, the complete callback is > > called for freeing the buffer and completing the fused command > > > > Also request reference is held during fused command lifetime, and this way > > guarantees that request buffer won't be freed until all inflight fused > > commands are completed. > > > > userspace(only implement sqe128 fused command): > > > > https://github.com/ming1/ubdsrv/tree/fused-cmd-zc-for-v6 > > > > liburing test(only implement normal sqe fused command: two 64byte SQEs) > > > > https://github.com/ming1/liburing/tree/fused_cmd_miniublk_for_v6 > > > > Signed-off-by: Ming Lei <ming.lei@xxxxxxxxxx> > > --- > > Documentation/block/ublk.rst | 126 ++++++++++++++++++++-- > > drivers/block/ublk_drv.c | 192 ++++++++++++++++++++++++++++++++-- > > include/uapi/linux/ublk_cmd.h | 6 +- > > 3 files changed, 303 insertions(+), 21 deletions(-) > > > > diff --git a/Documentation/block/ublk.rst b/Documentation/block/ublk.rst > > index 1713b2890abb..7b7aa24e9729 100644 > > --- a/Documentation/block/ublk.rst > > +++ b/Documentation/block/ublk.rst > > @@ -297,18 +297,126 @@ with specified IO tag in the command data: > > ``UBLK_IO_COMMIT_AND_FETCH_REQ`` to the server, ublkdrv needs to copy > > the server buffer (pages) read to the IO request pages. > > > > -Future development > > -================== > > +- ``UBLK_IO_FUSED_SUBMIT_IO`` > > + > > + Used for implementing zero copy feature. > > + > > + It has to been the primary command of io_uring fused command. This command > > s/been/be/ ? Yeah. > > > + submits the generic secondary IO request with io buffer provided by our primary > > + command, and won't be completed until the secondary request is done. > > + > > + The provided buffer is represented as ``io_uring_bvec_buf``, which is > > + actually ublk request buffer's reference, and the reference is shared & > > + read-only, so the generic secondary request can retrieve any part of the buffer > > + by passing buffer offset & length. > > > > Zero copy > > ---------- > > +========= > > + > > +What is zero copy? > > +------------------ > > + > > +When application submits IO to ``/dev/ublkb*``, userspace buffer(direct io) > > +or page cache buffer(buffered io) or kernel buffer(meta io often) is used > > +for submitting data to ublk driver, and all kinds of these buffers are > > +represented by bio/bvecs(ublk request buffer) finally. Before supporting > > +zero copy, data in these buffers has to be copied to ublk server userspace > > +buffer before handling WRITE IO, or after handing READ IO, so that ublk > > +server can handle IO for ``/dev/ublkb*`` with the copied data. > > + > > +The extra copy between ublk request buffer and ublk server userspace buffer > > +not only increases CPU utilization(such as pinning pages, copy data), but > > +also consumes memory bandwidth, and the cost could be very big when IO size > > +is big. It is observed that ublk-null IOPS may be increased to ~5X if the > > +extra copy can be avoided. > > + > > +So zero copy is very important for supporting high performance block device > > +in userspace. > > + > > +Technical requirements > > +---------------------- > > + > > +- ublk request buffer use > > + > > +ublk request buffer is represented by bio/bvec, which is immutable, so do > > +not try to change bvec via buffer reference; data can be read from or > > +written to the buffer according to buffer direction, but bvec can't be > > +changed > > + > > +- buffer lifetime > > + > > +Ublk server borrows ublk request buffer for handling ublk IO, ublk request > > +buffer reference is used. Reference can't outlive the referent buffer. That > > +means all request buffer references have to be released by ublk server > > +before ublk driver completes this request, when request buffer ownership > > +is transferred to upper layer(FS, application, ...). > > + > > +Also after ublk request is completed, any page belonging to this ublk > > +request can not be written or read any more from ublk server since it is > > +one block device from kernel viewpoint. > > + > > +- buffer direction > > + > > +For ublk WRITE request, ublk request buffer should only be accessed as data > > +source, and the buffer can't be written by ublk server > > + > > +For ublk READ request, ublk request buffer should only be accessed as data > > +destination, and the buffer can't be read by ublk server, otherwise kernel > > +data is leaked to ublk server, which can be unprivileged application. > > + > > +- arbitrary size sub-buffer needs to be retrieved from ublk server > > + > > +ublk is one generic framework for implementing block device in userspace, > > +and typical requirements include logical volume manager(mirror, stripped, ...), > > +distributed network storage, compressed target, ... > > + > > +ublk server needs to retrieve arbitrary size sub-buffer of ublk request, and > > +ublk server needs to submit IOs with these sub-buffer(s). That also means > > +arbitrary size sub-buffer(s) can be used to submit IO multiple times. > > + > > +Any sub-buffer is actually one reference of ublk request buffer, which > > +ownership can't be transferred to upper layer if any reference is held > > +by ublk server. > > + > > +Why slice isn't good for ublk zero copy > > slice Good catch. > > > +--------------------------------------- > > + > > +- spliced page from ->splice_read() can't be written > > + > > +ublk READ request can't be handled because spliced page can't be written to, and > > +extending splice for ublk zero copy isn't one good solution [#splice_extend]_ > > + > > +- it is very hard to meet above requirements wrt. request buffer lifetime > > + > > +splice/pipe focuses on page reference lifetime, but ublk zero copy pays more > > +attention to ublk request buffer lifetime. If is very inefficient to respect > > +request buffer lifetime by using all pipe buffer's ->release() which requires > > +all pipe buffers and pipe to be kept when ublk server handles IO. That means > > +one single dedicated ``pipe_inode_info`` has to be allocated runtime for each > > +provided buffer, and the pipe needs to be populated with pages in ublk request > > +buffer. > > + > > + > > +io_uring fused command based zero copy > > +-------------------------------------- > > + > > +io_uring fused command includes one primary command(uring command) and one > > +generic secondary request. The primary command is responsible for submitting > > +secondary request with provided buffer from ublk request, and primary command > > +won't be completed until the secondary request is completed. > > + > > +Typical ublk IO handling includes network and FS IO, so it is usual enough > > +for io_uring net & fs to support IO with provided buffer from primary command. > > > > -Zero copy is a generic requirement for nbd, fuse or similar drivers. A > > Maybe replace 'requirement' with "very useful"? It works without zc, > just with limited performance. OK, also it has been observed that the extra copy affects performance a log for big IO size(64K+). > > > -problem [#xiaoguang]_ Xiaoguang mentioned is that pages mapped to userspace > > -can't be remapped any more in kernel with existing mm interfaces. This can > > -occurs when destining direct IO to ``/dev/ublkb*``. Also, he reported that > > -big requests (IO size >= 256 KB) may benefit a lot from zero copy. > > +Once primary command is submitted successfully, ublk driver guarantees that > > +the ublk request buffer won't be gone away since secondary request actually > > +grabs the buffer's reference. This way also guarantees that multiple > > +concurrent fused commands associated with same request buffer works fine, > > +as the provided buffer reference is shared & read-only. > > > > +Also buffer usage direction flag is passed to primary command from userspace, > > +so ublk driver can validate if it is legal to use buffer with requested > > +direction. > > > > References > > ========== > > @@ -323,4 +431,4 @@ References > > > > .. [#stefan] https://lore.kernel.org/linux-block/YoOr6jBfgVm8GvWg@stefanha-x1.localdomain/ > > > > -.. [#xiaoguang] https://lore.kernel.org/linux-block/YoOr6jBfgVm8GvWg@stefanha-x1.localdomain/ > > +.. [#splice_extend] https://lore.kernel.org/linux-block/CAHk-=wgJsi7t7YYpuo6ewXGnHz2nmj67iWR6KPGoz5TBu34mWQ@xxxxxxxxxxxxxx/ > > > Any chance you could add a few code path into this document? For example > as in Documentation/filesystems/fuse.rst, > section "Kernel - userspace interface" OK, just it isn't handy to describe the flow by pure text. BTW, you can also see the code path in the link: https://github.com/ming1/ubdsrv/blob/master/doc/ublk_intro.pdf thanks, Ming
