Re: [PATCH 3/3] io_uring: add splice(2) support

[Jens Axboe <axboe@kernel.dk>]

On 1/21/20 8:16 PM, Pavel Begunkov wrote:
> On 22/01/2020 05:47, Jens Axboe wrote:
>> On 1/21/20 7:40 PM, Pavel Begunkov wrote:
>>>>> @@ -719,6 +730,11 @@ static const struct io_op_def io_op_defs[] = {
>>>>>  		.needs_file		= 1,
>>>>>  		.fd_non_neg		= 1,
>>>>>  	},
>>>>> +	[IORING_OP_SPLICE] = {
>>>>> +		.needs_file		= 1,
>>>>> +		.hash_reg_file		= 1,
>>>>> +		.unbound_nonreg_file	= 1,
>>>>> +	}
>>>>>  };
>>>>>  
>>>>>  static void io_wq_submit_work(struct io_wq_work **workptr);
>>>>
>>>> I probably want to queue up a reservation for the EPOLL_CTL that I
>>>> haven't included yet, but which has been tested. But that's easily
>>>> manageable, so no biggy on my end.
>>>
>>> I didn't quite get it. Do you mean collision of opcode numbers?
>>
>> Yeah that's all I meant, sorry wasn't too clear. But you can disregard,
>> I'll just pop a reservation in front if/when this is ready to go in if
>> it's before EPOLL_CTL op.
>>
>>>>> diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
>>>>> index 57d05cc5e271..f234b13e7ed3 100644
>>>>> --- a/include/uapi/linux/io_uring.h
>>>>> +++ b/include/uapi/linux/io_uring.h
>>>>> @@ -23,8 +23,14 @@ struct io_uring_sqe {
>>>>>  		__u64	off;	/* offset into file */
>>>>>  		__u64	addr2;
>>>>>  	};
>>>>> -	__u64	addr;		/* pointer to buffer or iovecs */
>>>>> -	__u32	len;		/* buffer size or number of iovecs */
>>>>> +	union {
>>>>> +		__u64	addr;		/* pointer to buffer or iovecs */
>>>>> +		__u64	off_out;
>>>>> +	};
>>>>> +	union {
>>>>> +		__u32	len;	/* buffer size or number of iovecs */
>>>>> +		__s32	fd_out;
>>>>> +	};
>>>>>  	union {
>>>>>  		__kernel_rwf_t	rw_flags;
>>>>>  		__u32		fsync_flags;
>>>>> @@ -37,10 +43,12 @@ struct io_uring_sqe {
>>>>>  		__u32		open_flags;
>>>>>  		__u32		statx_flags;
>>>>>  		__u32		fadvise_advice;
>>>>> +		__u32		splice_flags;
>>>>>  	};
>>>>>  	__u64	user_data;	/* data to be passed back at completion time */
>>>>>  	union {
>>>>>  		__u16	buf_index;	/* index into fixed buffers, if used */
>>>>> +		__u64	splice_len;
>>>>>  		__u64	__pad2[3];
>>>>>  	};
>>>>>  };
>>>>
>>>> Not a huge fan of this, also mean splice can't ever used fixed buffers.
>>>> Hmm...
>>>
>>> But it's not like splice() ever uses user buffers. Isn't it? vmsplice
>>> does, but that's another opcode.
>>
>> I guess that's true, I had vmsplice on my mind for this as well. But
>> won't be a problem there, since it doesn't take 6 arguments like splice
>> does.
>>
>> Another option is to do an indirect for splice, stuff the arguments in a
>> struct that's passed in as a pointer in ->addr. A bit slower, but
>> probably not a huge deal.
>>
>>>>> @@ -67,6 +75,9 @@ enum {
>>>>>  /* always go async */
>>>>>  #define IOSQE_ASYNC		(1U << IOSQE_ASYNC_BIT)
>>>>>  
>>>>> +/* op custom flags */
>>>>> +#define IOSQE_SPLICE_FIXED_OUT	(1U << 16)
>>>>> +
>>>>
>>>> I don't think it's unreasonable to say that if you specify
>>>> IOSQE_FIXED_FILE, then both are fixed. If not, then none of them are.
>>>> What do you think?
>>>>
>>>
>>> It's plausible to register only one end for splicing, e.g. splice from
>>> short-lived sockets to pre-registered buffers-pipes. And it's clearer
>>> do it now.
>>
>> You're probably right, though it's a bit nasty to add an unrelated flag
>> in the splice flag space... We should probably reserve it in splice
>> instead, and just not have it available from the regular system call.
>>
> Agree, it looks bad. I don't want to add it into sqe->splice_flags to
> not clash with splice(2) in the future, but could have a separate
> field in @sqe...  or can leave in in sqe->flags, as it's done in the
> patch, but that's like a portion of bits would be opcode specific and
> we would need to set rules for their use.

It won't clash with splice(2), just make that flag illegal if done
through splice(2) directly. Honestly I think that's (by FAR) the best
way to do it, having a private io_uring flag that acts as a splice flag
is really confusing and prone to breakage. Not that it's a huge issue
with splice as the flags have been stable for years, so don't really see
a high risk of collision. But we should still do it right, which means
adding SPLICE_F_OUT_FIXED or whatever you want to call it. Do that as a
prep patch, make do_splice() into __do_splice(), and have io_uring call
__do_splice(). Currently splice(2) is permissive in terms of flags, so
maybe just mask it in do_splice() to be on the safe side. Then we know
only internal users will set SPLICE_F_OUT_FIXED, and we'll never run
into the risk of having a collision as it's part of the flag space
anyway.

The sqe->flags space is very tight, so adding a splice specific opcode
there would be bad.

-- 
Jens Axboe