Message-ID: <MWHPR04MB3758DC08EA17780E498E9EC0E74E0@MWHPR04MB3758.namprd04.prod.outlook.com>
Date:   Fri, 31 Jul 2020 06:42:10 +0000
From:   Damien Le Moal <Damien.LeMoal@....com>
To:     Kanchan Joshi <joshiiitr@...il.com>, Jens Axboe <axboe@...nel.dk>
CC:     Pavel Begunkov <asml.silence@...il.com>,
        Kanchan Joshi <joshi.k@...sung.com>,
        "viro@...iv.linux.org.uk" <viro@...iv.linux.org.uk>,
        "bcrl@...ck.org" <bcrl@...ck.org>,
        Matthew Wilcox <willy@...radead.org>,
        "hch@...radead.org" <hch@...radead.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>,
        "linux-api@...r.kernel.org" <linux-api@...r.kernel.org>,
        SelvaKumar S <selvakuma.s1@...sung.com>,
        Nitesh Shetty <nj.shetty@...sung.com>,
        Javier Gonzalez <javier.gonz@...sung.com>
Subject: Re: [PATCH v4 6/6] io_uring: add support for zone-append

On 2020/07/31 3:26, Kanchan Joshi wrote:
> On Thu, Jul 30, 2020 at 11:24 PM Jens Axboe <axboe@...nel.dk> wrote:
>>
>> On 7/30/20 11:51 AM, Kanchan Joshi wrote:
>>> On Thu, Jul 30, 2020 at 11:10 PM Pavel Begunkov <asml.silence@...il.com> wrote:
>>>>
>>>> On 30/07/2020 20:16, Jens Axboe wrote:
>>>>> On 7/30/20 10:26 AM, Pavel Begunkov wrote:
>>>>>> On 30/07/2020 19:13, Jens Axboe wrote:
>>>>>>> On 7/30/20 10:08 AM, Pavel Begunkov wrote:
>>>>>>>> On 27/07/2020 23:34, Jens Axboe wrote:
>>>>>>>>> On 7/27/20 1:16 PM, Kanchan Joshi wrote:
>>>>>>>>>> On Fri, Jul 24, 2020 at 10:00 PM Jens Axboe <axboe@...nel.dk> wrote:
>>>>>>>>>>>
>>>>>>>>>>> On 7/24/20 9:49 AM, Kanchan Joshi wrote:
>>>>>>>>>>>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>>>>>>>>>>>> index 7809ab2..6510cf5 100644
>>>>>>>>>>>> --- a/fs/io_uring.c
>>>>>>>>>>>> +++ b/fs/io_uring.c
>>>>>>>>>>>> @@ -1284,8 +1301,15 @@ static void __io_cqring_fill_event(struct io_kiocb *req, long res, long cflags)
>>>>>>>>>>>>       cqe = io_get_cqring(ctx);
>>>>>>>>>>>>       if (likely(cqe)) {
>>>>>>>>>>>>               WRITE_ONCE(cqe->user_data, req->user_data);
>>>>>>>>>>>> -             WRITE_ONCE(cqe->res, res);
>>>>>>>>>>>> -             WRITE_ONCE(cqe->flags, cflags);
>>>>>>>>>>>> +             if (unlikely(req->flags & REQ_F_ZONE_APPEND)) {
>>>>>>>>>>>> +                     if (likely(res > 0))
>>>>>>>>>>>> +                             WRITE_ONCE(cqe->res64, req->rw.append_offset);
>>>>>>>>>>>> +                     else
>>>>>>>>>>>> +                             WRITE_ONCE(cqe->res64, res);
>>>>>>>>>>>> +             } else {
>>>>>>>>>>>> +                     WRITE_ONCE(cqe->res, res);
>>>>>>>>>>>> +                     WRITE_ONCE(cqe->flags, cflags);
>>>>>>>>>>>> +             }
>>>>>>>>>>>
>>>>>>>>>>> This would be nice to keep out of the fast path, if possible.
>>>>>>>>>>
>>>>>>>>>> I was thinking of keeping a function-pointer (in io_kiocb) during
>>>>>>>>>> submission. That would have avoided this check......but argument count
>>>>>>>>>> differs, so it did not add up.
>>>>>>>>>
>>>>>>>>> But that'd grow the io_kiocb just for this use case, which is arguably
>>>>>>>>> even worse. Unless you can keep it in the per-request private data,
>>>>>>>>> but there's no more room there for the regular read/write side.
>>>>>>>>>
>>>>>>>>>>>> diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
>>>>>>>>>>>> index 92c2269..2580d93 100644
>>>>>>>>>>>> --- a/include/uapi/linux/io_uring.h
>>>>>>>>>>>> +++ b/include/uapi/linux/io_uring.h
>>>>>>>>>>>> @@ -156,8 +156,13 @@ enum {
>>>>>>>>>>>>   */
>>>>>>>>>>>>  struct io_uring_cqe {
>>>>>>>>>>>>       __u64   user_data;      /* sqe->data submission passed back */
>>>>>>>>>>>> -     __s32   res;            /* result code for this event */
>>>>>>>>>>>> -     __u32   flags;
>>>>>>>>>>>> +     union {
>>>>>>>>>>>> +             struct {
>>>>>>>>>>>> +                     __s32   res;    /* result code for this event */
>>>>>>>>>>>> +                     __u32   flags;
>>>>>>>>>>>> +             };
>>>>>>>>>>>> +             __s64   res64;  /* appending offset for zone append */
>>>>>>>>>>>> +     };
>>>>>>>>>>>>  };
>>>>>>>>>>>
>>>>>>>>>>> Is this a compatible change, both for now but also going forward? You
>>>>>>>>>>> could randomly have IORING_CQE_F_BUFFER set, or any other future flags.
>>>>>>>>>>
>>>>>>>>>> Sorry, I didn't quite understand the concern. CQE_F_BUFFER is not
>>>>>>>>>> used/set for write currently, so it looked compatible at this point.
>>>>>>>>>
>>>>>>>>> Not worried about that, since we won't ever use that for writes. But it
>>>>>>>>> is a potential headache down the line for other flags, if they apply to
>>>>>>>>> normal writes.
>>>>>>>>>
>>>>>>>>>> Yes, no room for future flags for this operation.
>>>>>>>>>> Do you see any other way to enable this support in io-uring?
>>>>>>>>>
>>>>>>>>> Honestly I think the only viable option is as we discussed previously,
>>>>>>>>> pass in a pointer to a 64-bit type where we can copy the additional
>>>>>>>>> completion information to.
>>>>>>>>
>>>>>>>> TBH, I hate the idea of such overhead/latency at times when SSDs can
>>>>>>>> serve writes in less than 10ms. Any chance you measured how long does it
>>>>>>>
>>>>>>> 10us? :-)
>>>>>>
>>>>>> Hah, 10us indeed :)
>>>>>>
>>>>>>>
>>>>>>>> take to drag through task_work?
>>>>>>>
>>>>>>> A 64-bit value copy is really not a lot of overhead... But yes, we'd
>>>>>>> need to push the completion through task_work at that point, as we can't
>>>>>>> do it from the completion side. That's not a lot of overhead, and most
>>>>>>> notably, it's overhead that only affects this particular type.
>>>>>>>
>>>>>>> That's not a bad starting point, and something that can always be
>>>>>>> optimized later if need be. But I seriously doubt it'd be anything to
>>>>>>> worry about.
>>>>>>
>>>>>> I probably need to look myself how it's really scheduled, but if you don't
>>>>>> mind, here is a quick question: if we do work_add(task) when the task is
>>>>>> running in the userspace, wouldn't the work execution wait until the next
>>>>>> syscall/allotted time ends up?
>>>>>
>>>>> It'll get the task to enter the kernel, just like signal delivery. The only
>>>>> tricky part is really if we have a dependency waiting in the kernel, like
>>>>> the recent eventfd fix.
>>>>
>>>> I see, thanks for sorting this out!
>>>
>>> Few more doubts about this (please mark me wrong if that is the case):
>>>
>>> - Task-work makes me feel like N completions waiting to be served by
>>> single task.
>>> Currently completions keep arriving and CQEs would be updated with
>>> result, but the user-space (submitter task) would not be poked.
>>>
>>> - Completion-code will set the task-work. But post that it cannot go
>>> immediately to its regular business of picking cqe and updating
>>> res/flags, as we cannot afford user-space to see the cqe before the
>>> pointer update. So it seems completion-code needs to spawn another
>>> work which will allocate/update cqe after waiting for pointer-update
>>> from task-work?
>>
>> The task work would post the completion CQE for the request after
>> writing the offset.
> 
> Got it, thank you for making it simple.
> Overall if I try to put the tradeoffs of moving to indirect-offset
> (compared to current scheme)–
> 
> Upside:
> - cqe res/flags would be intact, avoids future-headaches as you mentioned
> - short-write cases do not have to be failed in lower-layers (as
> cqe->res is there to report bytes-copied)

I personally think it is a super bad idea to allow short asynchronous append
writes. The interface should allow the async zone append write to proceed only
and only if it can be stuffed entirely into a single BIO which necessarilly will
be a single request on the device side. Otherwise, the application would have no
guarantees as to where a split may happen, and since this is zone append, the
next async append will not leave any hole to complete a previous short write.
This will wreak the structure of the application data.

For the sync case, this is fine. The application can just issue a new append
write with the remaining unwritten data from the previous append write. But in
the async case, if one write == one data record (e.g. a key-value tuple for an
SSTable in an LSM tree), then allowing a short write will destroy the record:
the partial write will be garbage data that will need garbage collection...

> 
> Downside:
> - We may not be able to use RWF_APPEND, and need exposing a new
> type/flag (RWF_INDIRECT_OFFSET etc.) user-space. Not sure if this
> sounds outrageous, but is it OK to have uring-only flag which can be
> combined with RWF_APPEND?

Why ? Where is the problem ? O_APPEND/RWF_APPEND is currently meaningless for
raw block device accesses. We could certainly define a meaning for these in the
context of zoned block devices.

I already commented on the need for first defining an interface (flags etc) and
its semantic (e.g. do we allow short zone append or not ? What happens for
regular files ? etc). Did you read my comment ? We really need to first agree on
something to clarify what needs to be done.


> -  Expensive compared to sending results in cqe itself. But I agree
> that this may not be major, and only for one type of write.
> 
> 


-- 
Damien Le Moal
Western Digital Research

Powered by blists - more mailing lists