Re: [RFC] fixed files

[Jens Axboe <axboe@kernel.dk>]

On 2/9/20 5:18 AM, Pavel Begunkov wrote:
> On 2/8/2020 11:15 PM, Jens Axboe wrote:
>> On 2/8/20 6:28 AM, Pavel Begunkov wrote:
>>> Hi,
>>>
>>> As you remember, splice(2) needs two fds, and it's a bit of a pain
>>> finding a place for the second REQ_F_FIXED_FILE flag. So, I was
>>> thinking, can we use the last (i.e. sign) bit to mark an fd as fixed? A
>>> lot of userspace programs consider any negative result of open() as an
>>> error, so it's more or less safe to reuse it.
>>>
>>> e.g.
>>> fill_sqe(fd) // is not fixed
>>> fill_sqe(buf_idx | LAST_BIT) // fixed file
>>
>> Right now we only support 1024 fixed buffers anyway, so we do have some
>> space there. If we steal a bit, it'll still allow us to expand to 32K of
>> fixed buffers in the future.
>>
>> It's a bit iffy, but like you, I don't immediately see a better way to
>> do this that doesn't include stealing an IOSQE bit or adding a special
>> splice flag for it. Might still prefer the latter, to be honest...
> 
> "fixed" is clearly a per-{fd,buffer} attribute. If I'd now design it
> from the scratch, I would store fixed-resource index in the same field
> as fds and addr (but not separate @buf_index), and have per-resource
> switch-flag somewhere. And then I see 2 convenient ways:
> 
> 1. encode the fixed bit into addr and fd, as supposed above.
> 
> 2. Add N generic IOSQE_FIXED bits (i.e. IOSQE_FIXED_RESOURSE{1,2,...}),
> which correspond to resources (fd, buffer, etc) in order of occurrence
> in an sqe. I wouldn't expect having more than 3-4 flags.
> 
> And then IORING_OP_{READ,WRITE}_FIXED would have been the same opcode as
> the corresponding non-fixed version. But backward-compatibility is a pain.

It's always much easier looking back, hindsight is much clearer. I'd also
expand the sqe flags bits to 16 at least, but oh well.

I do think that for this particular case we add a SPLICE_F_FD1_FIXED and
ditto for fd2, and just have the direct splice/vmsplice syscalls reject
them as invalid. Both splice and vmsplice -EINVAL for unknown flags,
which makes this possible.

That seems cleaner to me than trying to shoe-horn this information into
the sqe itself, and it can easily be done as a prep patch to adding
splice support.

-- 
Jens Axboe