Re: [PATCH v3 4/4] io_uring: add support for zone-append

[Kanchan Joshi]

On Mon, Jul 06, 2020 at 03:32:08PM +0100, Matthew Wilcox wrote:
> On Mon, Jul 06, 2020 at 08:27:17AM -0600, Jens Axboe wrote:
> > On 7/6/20 8:10 AM, Matthew Wilcox wrote:
> > > On Sun, Jul 05, 2020 at 03:12:50PM -0600, Jens Axboe wrote:
> > >> On 7/5/20 3:09 PM, Matthew Wilcox wrote:
> > >>> On Sun, Jul 05, 2020 at 03:00:47PM -0600, Jens Axboe wrote:
> > >>>> On 7/5/20 12:47 PM, Kanchan Joshi wrote:
> > >>>>> From: Selvakumar S <selvakuma.s1@xxxxxxxxxxx>
> > >>>>>
> > >>>>> For zone-append, block-layer will return zone-relative offset via ret2
> > >>>>> of ki_complete interface. Make changes to collect it, and send to
> > >>>>> user-space using cqe->flags.
> > >
> > >>> I'm surprised you aren't more upset by the abuse of cqe->flags for the
> > >>> address.

Documentation (https://kernel.dk/io_uring.pdf) mentioned cqe->flags can carry
the metadata for the operation. I wonder if this should be called abuse.

> > >> Yeah, it's not great either, but we have less leeway there in terms of
> > >> how much space is available to pass back extra data.
> > >>
> > >>> What do you think to my idea of interpreting the user_data as being a
> > >>> pointer to somewhere to store the address?  Obviously other things
> > >>> can be stored after the address in the user_data.
> > >>
> > >> I don't like that at all, as all other commands just pass user_data
> > >> through. This means the application would have to treat this very
> > >> differently, and potentially not have a way to store any data for
> > >> locating the original command on the user side.
> > >
> > > I think you misunderstood me.  You seem to have thought I meant
> > > "use the user_data field to return the address" when I actually meant
> > > "interpret the user_data field as a pointer to where userspace
> > > wants the address stored".
> >
> > It's still somewhat weird to have user_data have special meaning, you're
> > now having the kernel interpret it while every other command it's just
> > an opaque that is passed through.
> >
> > But it could of course work, and the app could embed the necessary
> > u32/u64 in some other structure that's persistent across IO. If it
> > doesn't have that, then it'd need to now have one allocated and freed
> > across the lifetime of the IO.
> >
> > If we're going that route, it'd be better to define the write such that
> > you're passing in the necessary information upfront. In syscall terms,
> > then that'd be something ala:
> >
> > ssize_t my_append_write(int fd, const struct iovec *iov, int iovcnt,
> > 			off_t *offset, int flags);
> >
> > where *offset is copied out when the write completes. That removes the
> > need to abuse user_data, with just providing the storage pointer for the
> > offset upfront.
>
> That works for me!  In io_uring terms, would you like to see that done
> as adding:
>
>        union {
>                __u64   off;    /* offset into file */
> +		__u64   *offp;	/* appending writes */
>                __u64   addr2;
>        };
But there are peformance implications of this approach?
If I got it right, the workflow is: 
- Application allocates 64bit of space, writes "off" into it and pass it
 in the sqe->addr2
- Kernel first reads sqe->addr2, reads the value to know the intended
 write-location, and stores the address somewhere (?) to be used during
 completion. Storing this address seems tricky as this may add one more
 cacheline (in io_kiocb->rw)?
- During completion cqe res/flags are written as before, but extra step
 to copy the append-completion-result into that user-space address.

Extra steps are due to the pointer indirection.
And it seems application needs to be careful about managing this 64bit of
space for a cluster of writes, especially if it wants to reuse the sqe
before the completion.
New one can handle 64bit result cleanly, but seems slower than current
one.
It will be good to have the tradeoff cleared before we take things to V4.