[RFC PATCH 0/4] Asynchronous passthrough ioctl

[Jens Axboe]

On 1/28/21 5:04 AM, Kanchan Joshi wrote:
> On Wed, Jan 27, 2021 at 9:32 PM Pavel Begunkov <asml.silence at gmail.com> wrote:
>>
>> On 27/01/2021 15:42, Pavel Begunkov wrote:
>>> On 27/01/2021 15:00, Kanchan Joshi wrote:
>>>> This RFC patchset adds asynchronous ioctl capability for NVMe devices.
>>>> Purpose of RFC is to get the feedback and optimize the path.
>>>>
>>>> At the uppermost io-uring layer, a new opcode IORING_OP_IOCTL_PT is
>>>> presented to user-space applications. Like regular-ioctl, it takes
>>>> ioctl opcode and an optional argument (ioctl-specific input/output
>>>> parameter). Unlike regular-ioctl, it is made to skip the block-layer
>>>> and reach directly to the underlying driver (nvme in the case of this
>>>> patchset). This path between io-uring and nvme is via a newly
>>>> introduced block-device operation "async_ioctl". This operation
>>>> expects io-uring to supply a callback function which can be used to
>>>> report completion at later stage.
>>>>
>>>> For a regular ioctl, NVMe driver submits the command to the device and
>>>> the submitter (task) is made to wait until completion arrives. For
>>>> async-ioctl, completion is decoupled from submission. Submitter goes
>>>> back to its business without waiting for nvme-completion. When
>>>> nvme-completion arrives, it informs io-uring via the registered
>>>> completion-handler. But some ioctls may require updating certain
>>>> ioctl-specific fields which can be accessed only in context of the
>>>> submitter task. For that reason, NVMe driver uses task-work infra for
>>>> that ioctl-specific update. Since task-work is not exported, it cannot
>>>> be referenced when nvme is compiled as a module. Therefore, one of the
>>>> patch exports task-work API.
>>>>
>>>> Here goes example of usage (pseudo-code).
>>>> Actual nvme-cli source, modified to issue all ioctls via this opcode
>>>> is present at-
>>>> https://github.com/joshkan/nvme-cli/commit/a008a733f24ab5593e7874cfbc69ee04e88068c5
>>>
>>> see https://git.kernel.dk/cgit/linux-block/log/?h=io_uring-fops
>>>
>>> Looks like good time to bring that branch/discussion back
>>
>> a bit more context:
>> https://github.com/axboe/liburing/issues/270
> 
> Thanks, it looked good. It seems key differences (compared to
> uring-patch that I posted) are -
> 1. using file-operation instead of block-dev operation.

Right, it's meant to span wider than just block devices.

> 2. repurpose the sqe memory for ioctl-cmd. If an application does
> ioctl with <=40 bytes of cmd, it does not have to allocate ioctl-cmd.
> That's nifty. We still need to support passing larger-cmd (e.g.
> nvme-passthru ioctl takes 72 bytes) but that shouldn't get too
> difficult I suppose.

It's actually 48 bytes in the as-posted version, and I've bumped it to
56 bytes in the latest branch. So not quite enough for everything,
nothing ever will be, but should work for a lot of cases without
requiring per-command allocations just for the actual command.

> And for some ioctls, driver may still need to use task-work to update
> the user-space pointers (embedded in uring/ioctl cmd) during
> completion.
> 
> @Jens - will it be fine if I start looking at plumbing nvme-part of
> this series on top of your work?

Sure, go ahead. Just beware that things are still changing, so you might
have to adapt it a few times. It's still early days, but I do think
that's the way forward in providing controlled access to what is
basically async ioctls.

-- 
Jens Axboe