Message-ID: <26f41284-6e52-58d3-6119-8ed34511e5dc@kernel.dk>
Date:   Fri, 19 Jun 2020 09:20:32 -0600
From:   Jens Axboe <axboe@...nel.dk>
To:     Matias Bjørling <mb@...htnvm.io>,
        "javier.gonz@...sung.com" <javier@...igon.com>,
        Damien Le Moal <Damien.LeMoal@....com>
Cc:     Kanchan Joshi <joshi.k@...sung.com>,
        "viro@...iv.linux.org.uk" <viro@...iv.linux.org.uk>,
        "bcrl@...ck.org" <bcrl@...ck.org>,
        "linux-fsdevel@...r.kernel.org" <linux-fsdevel@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        "linux-aio@...ck.org" <linux-aio@...ck.org>,
        "io-uring@...r.kernel.org" <io-uring@...r.kernel.org>,
        "linux-block@...r.kernel.org" <linux-block@...r.kernel.org>,
        "selvakuma.s1@...sung.com" <selvakuma.s1@...sung.com>,
        "nj.shetty@...sung.com" <nj.shetty@...sung.com>
Subject: Re: [PATCH 3/3] io_uring: add support for zone-append

On 6/19/20 9:14 AM, Matias Bjørling wrote:
> On 19/06/2020 16.18, Jens Axboe wrote:
>> On 6/19/20 5:15 AM, Matias Bjørling wrote:
>>> On 19/06/2020 11.41, javier.gonz@...sung.com wrote:
>>>> Jens,
>>>>
>>>> Would you have time to answer a question below in this thread?
>>>>
>>>> On 18.06.2020 11:11, javier.gonz@...sung.com wrote:
>>>>> On 18.06.2020 08:47, Damien Le Moal wrote:
>>>>>> On 2020/06/18 17:35, javier.gonz@...sung.com 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@...sung.com>
>>>>>>>>>
>>>>>>>>> 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@...sung.com>
>>>>>>>>> Signed-off-by: Kanchan Joshi <joshi.k@...sung.com>
>>>>>>>>> Signed-off-by: Nitesh Shetty <nj.shetty@...sung.com>
>>>>>>>>> Signed-off-by: Javier Gonzalez <javier.gonz@...sung.com>
>>>>>>>>> ---
>>>>>>>>> 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.
>>> I took a quick look at the code. No expert, but why not use the existing
>>> userdata variable? use the lowest bits (40 bits) for the Zone Starting
>>> LBA, and use the highest (24 bits) as index into the completion data
>>> structure?
>>>
>>> If you want to pass the memory address (same as what fio does) for the
>>> data structure used for completion, one may also play some tricks by
>>> using a relative memory address to the data structure. For example, the
>>> x86_64 architecture uses 48 address bits for its memory addresses. With
>>> 24 bit, one can allocate the completion entries in a 32MB memory range,
>>> and then use base_address + index to get back to the completion data
>>> structure specified in the sqe.
>> For any current request, sqe->user_data is just provided back as
>> cqe->user_data. This would make these requests behave differently
>> from everything else in that sense, which seems very confusing to me
>> if I was an application writer.
>>
>> But generally I do agree with you, there are lots of ways to make
>> < 64-bit work as a tag without losing anything or having to jump
>> through hoops to do so. The lack of consistency introduced by having
>> zone append work differently is ugly, though.
>>
> Yep, agree, and extending to three cachelines is big no-go. We could add 
> a flag that said the kernel has changes the userdata variable. That'll 
> make it very explicit.

Don't like that either, as it doesn't really change the fact that you're
now doing something very different with the user_data field, which is
just supposed to be passed in/out directly. Adding a random flag to
signal this behavior isn't very explicit either, imho. It's still some
out-of-band (ish) notification of behavior that is different from any
other command. This is very different from having a flag that says
"there's extra information in this other field", which is much cleaner.

-- 
Jens Axboe

Powered by blists - more mailing lists