Re: [PATCH 3/3] io_uring: add support for zone-append

[Pavel Begunkov]

On 19/06/2020 17:15, Jens Axboe wrote:
> On 6/19/20 3:41 AM, javier.gonz@xxxxxxxxxxx wrote:
>> Jens,
>>
>> Would you have time to answer a question below in this thread?
>>
>> On 18.06.2020 11:11, javier.gonz@xxxxxxxxxxx wrote:
>>> On 18.06.2020 08:47, Damien Le Moal wrote:
>>>> On 2020/06/18 17:35, javier.gonz@xxxxxxxxxxx wrote:
>>>>> On 18.06.2020 07:39, Damien Le Moal wrote:
>>>>>> On 2020/06/18 2:27, Kanchan Joshi wrote:
>>>>>>> From: Selvakumar S <selvakuma.s1@xxxxxxxxxxx>
>>>>>>>
>>>>>>> Introduce three new opcodes for zone-append -
>>>>>>>
>>>>>>>   IORING_OP_ZONE_APPEND     : non-vectord, similiar to IORING_OP_WRITE
>>>>>>>   IORING_OP_ZONE_APPENDV    : vectored, similar to IORING_OP_WRITEV
>>>>>>>   IORING_OP_ZONE_APPEND_FIXED : append using fixed-buffers
>>>>>>>
>>>>>>> Repurpose cqe->flags to return zone-relative offset.
>>>>>>>
>>>>>>> Signed-off-by: SelvaKumar S <selvakuma.s1@xxxxxxxxxxx>
>>>>>>> Signed-off-by: Kanchan Joshi <joshi.k@xxxxxxxxxxx>
>>>>>>> Signed-off-by: Nitesh Shetty <nj.shetty@xxxxxxxxxxx>
>>>>>>> Signed-off-by: Javier Gonzalez <javier.gonz@xxxxxxxxxxx>
>>>>>>> ---
>>>>>>> fs/io_uring.c                 | 72 +++++++++++++++++++++++++++++++++++++++++--
>>>>>>> include/uapi/linux/io_uring.h |  8 ++++-
>>>>>>> 2 files changed, 77 insertions(+), 3 deletions(-)
>>>>>>>
>>>>>>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>>>>>>> index 155f3d8..c14c873 100644
>>>>>>> --- a/fs/io_uring.c
>>>>>>> +++ b/fs/io_uring.c
>>>>>>> @@ -649,6 +649,10 @@ struct io_kiocb {
>>>>>>> 	unsigned long		fsize;
>>>>>>> 	u64			user_data;
>>>>>>> 	u32			result;
>>>>>>> +#ifdef CONFIG_BLK_DEV_ZONED
>>>>>>> +	/* zone-relative offset for append, in bytes */
>>>>>>> +	u32			append_offset;
>>>>>>
>>>>>> this can overflow. u64 is needed.
>>>>>
>>>>> We chose to do it this way to start with because struct io_uring_cqe
>>>>> only has space for u32 when we reuse the flags.
>>>>>
>>>>> We can of course create a new cqe structure, but that will come with
>>>>> larger changes to io_uring for supporting append.
>>>>>
>>>>> Do you believe this is a better approach?
>>>>
>>>> The problem is that zone size are 32 bits in the kernel, as a number
>>>> of sectors.  So any device that has a zone size smaller or equal to
>>>> 2^31 512B sectors can be accepted. Using a zone relative offset in
>>>> bytes for returning zone append result is OK-ish, but to match the
>>>> kernel supported range of possible zone size, you need 31+9 bits...
>>>> 32 does not cut it.
>>>
>>> Agree. Our initial assumption was that u32 would cover current zone size
>>> requirements, but if this is a no-go, we will take the longer path.
>>
>> Converting to u64 will require a new version of io_uring_cqe, where we
>> extend at least 32 bits. I believe this will need a whole new allocation
>> and probably ioctl().
>>
>> Is this an acceptable change for you? We will of course add support for
>> liburing when we agree on the right way to do this.
> 
> If you need 64-bit of return value, then it's not going to work. Even
> with the existing patches, reusing cqe->flags isn't going to fly, as
> it would conflict with eg doing zone append writes with automatic
> buffer selection.

Buffer selection is for reads/recv kind of requests, but appends
are writes. In theory they can co-exist using cqe->flags.


> 
> We're not changing the io_uring_cqe. It's important to keep it lean, and
> any other request type is generally fine with 64-bit tag + 32-bit result
> (and 32-bit flags on the side) for completions.
> 
> Only viable alternative I see would be to provide an area to store this
> information, and pass in a pointer to this at submission time through
> the sqe. One issue I do see with that is if we only have this
> information available at completion time, then we'd need some async punt
> to copy it to user space... Generally not ideal.
> 
> A hackier approach would be to play some tricks with cqe->res and
> cqe->flags, setting aside a flag to denote an extension of cqe->res.
> That would mean excluding zone append (etc) from using buffer selection,
> which probably isn't a huge deal. It'd be more problematic for any other
> future flags. But if you just need 40 bits, then it could certainly
> work. Rigth now, if cqe->flags & 1 is set, then (cqe->flags >> 16) is
> the buffer ID. You could define IORING_CQE_F_ZONE_FOO to be bit 1, so
> that:
> 
> 	uint64_t val = cqe->res; // assuming non-error here
> 
> 	if (cqe->flags & IORING_CQE_F_ZONE_FOO)
> 		val |= (cqe->flags >> 16) << 32ULL;
> 
> and hence use the upper 16 bits of cqe->flags for the upper bits of your
> (then) 48-bit total value.

How about returning offset in terms of 512-bytes chunks? NVMe is 512B
atomic/aligned. We'll lose an ability to do non-512 aligned appends, but
it won't hit media as such anyway (will be padded or cached), so
personally I don't see much benefit in having it.


-- 
Pavel Begunkov