Re: [PATCH 3/4] io_uring: add IORING_OP_READ[WRITE]_SPLICE_BUF

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 2/10/23 8:32?AM, Ming Lei wrote:
> IORING_OP_READ_SPLICE_BUF: read to buffer which is built from
> ->read_splice() of specified fd, so user needs to provide (splice_fd, offset, len)
> for building buffer.
> 
> IORING_OP_WRITE_SPLICE_BUF: write from buffer which is built from
> ->read_splice() of specified fd, so user needs to provide (splice_fd, offset, len)
> for building buffer.
> 
> The typical use case is for supporting ublk/fuse io_uring zero copy,
> and READ/WRITE OP retrieves ublk/fuse request buffer via direct pipe
> from device->read_splice(), then READ/WRITE can be done to/from this
> buffer directly.

Main question here - would this be better not plumbed up through the rw
path? Might be cleaner, even if it either requires a bit of helper
refactoring or accepting a bit of duplication. But would still be better
than polluting the rw fast path imho.

Also seems like this should be separately testable. We can't add new
opcodes that don't have a feature test at least, and should also have
various corner case tests. A bit of commenting outside of this below.

> diff --git a/io_uring/opdef.c b/io_uring/opdef.c
> index 5238ecd7af6a..91e8d8f96134 100644
> --- a/io_uring/opdef.c
> +++ b/io_uring/opdef.c
> @@ -427,6 +427,31 @@ const struct io_issue_def io_issue_defs[] = {
>  		.prep			= io_eopnotsupp_prep,
>  #endif
>  	},
> +	[IORING_OP_READ_SPLICE_BUF] = {
> +		.needs_file		= 1,
> +		.unbound_nonreg_file	= 1,
> +		.pollin			= 1,
> +		.plug			= 1,
> +		.audit_skip		= 1,
> +		.ioprio			= 1,
> +		.iopoll			= 1,
> +		.iopoll_queue		= 1,
> +		.prep			= io_prep_rw,
> +		.issue			= io_read,
> +	},
> +	[IORING_OP_WRITE_SPLICE_BUF] = {
> +		.needs_file		= 1,
> +		.hash_reg_file		= 1,
> +		.unbound_nonreg_file	= 1,
> +		.pollout		= 1,
> +		.plug			= 1,
> +		.audit_skip		= 1,
> +		.ioprio			= 1,
> +		.iopoll			= 1,
> +		.iopoll_queue		= 1,
> +		.prep			= io_prep_rw,
> +		.issue			= io_write,
> +	},

Are these really safe with iopoll?

> +static int io_prep_rw_splice_buf(struct io_kiocb *req,
> +				 const struct io_uring_sqe *sqe)
> +{
> +	struct io_rw *rw = io_kiocb_to_cmd(req, struct io_rw);
> +	unsigned nr_pages = io_rw_splice_buf_nr_bvecs(rw->len);
> +	loff_t splice_off = READ_ONCE(sqe->splice_off_in);
> +	struct io_rw_splice_buf_data data;
> +	struct io_mapped_ubuf *imu;
> +	struct fd splice_fd;
> +	int ret;
> +
> +	splice_fd = fdget(READ_ONCE(sqe->splice_fd_in));
> +	if (!splice_fd.file)
> +		return -EBADF;

Seems like this should check for SPLICE_F_FD_IN_FIXED, and also use
io_file_get_normal() for the non-fixed case in case someone passed in an
io_uring fd.

> +	data.imu = &imu;
> +
> +	rw->addr = 0;
> +	req->flags |= REQ_F_NEED_CLEANUP;
> +
> +	ret = __io_prep_rw_splice_buf(req, &data, splice_fd.file, rw->len,
> +			splice_off);
> +	imu = *data.imu;
> +	imu->acct_pages = 0;
> +	imu->ubuf = 0;
> +	imu->ubuf_end = data.total;
> +	rw->len = data.total;
> +	req->imu = imu;
> +	if (!data.total) {
> +		io_rw_cleanup_splice_buf(req);
> +	} else  {
> +		ret = 0;
> +	}
> +out_put_fd:
> +	if (splice_fd.file)
> +		fdput(splice_fd);
> +
> +	return ret;
> +}

If the operation is done, clear NEED_CLEANUP and do the cleanup here?
That'll be faster.

-- 
Jens Axboe




[Index of Archives]     [Linux RAID]     [Linux SCSI]     [Linux ATA RAID]     [IDE]     [Linux Wireless]     [Linux Kernel]     [ATH6KL]     [Linux Bluetooth]     [Linux Netdev]     [Kernel Newbies]     [Security]     [Git]     [Netfilter]     [Bugtraq]     [Yosemite News]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Device Mapper]

  Powered by Linux