Re: [RFC PATCH v3 15/30] io: Add a pwritev/preadv version that takes a discontiguous iovec

[Peter Xu <peterx@redhat.com>]

On Tue, Jan 16, 2024 at 03:15:50PM -0300, Fabiano Rosas wrote:
> Peter Xu <peterx@redhat.com> writes:
> 
> > On Mon, Nov 27, 2023 at 05:25:57PM -0300, Fabiano Rosas wrote:
> >> For the upcoming support to fixed-ram migration with multifd, we need
> >> to be able to accept an iovec array with non-contiguous data.
> >> 
> >> Add a pwritev and preadv version that splits the array into contiguous
> >> segments before writing. With that we can have the ram code continue
> >> to add pages in any order and the multifd code continue to send large
> >> arrays for reading and writing.
> >> 
> >> Signed-off-by: Fabiano Rosas <farosas@suse.de>
> >> ---
> >> - split the API that was merged into a single function
> >> - use uintptr_t for compatibility with 32-bit
> >> ---
> >>  include/io/channel.h | 26 ++++++++++++++++
> >>  io/channel.c         | 70 ++++++++++++++++++++++++++++++++++++++++++++
> >>  2 files changed, 96 insertions(+)
> >> 
> >> diff --git a/include/io/channel.h b/include/io/channel.h
> >> index 7986c49c71..25383db5aa 100644
> >> --- a/include/io/channel.h
> >> +++ b/include/io/channel.h
> >> @@ -559,6 +559,19 @@ int qio_channel_close(QIOChannel *ioc,
> >>  ssize_t qio_channel_pwritev(QIOChannel *ioc, const struct iovec *iov,
> >>                              size_t niov, off_t offset, Error **errp);
> >>  
> >> +/**
> >> + * qio_channel_pwritev_all:
> >> + * @ioc: the channel object
> >> + * @iov: the array of memory regions to write data from
> >> + * @niov: the length of the @iov array
> >> + * @offset: the iovec offset in the file where to write the data
> >> + * @errp: pointer to a NULL-initialized error object
> >> + *
> >> + * Returns: 0 if all bytes were written, or -1 on error
> >> + */
> >> +int qio_channel_pwritev_all(QIOChannel *ioc, const struct iovec *iov,
> >> +                            size_t niov, off_t offset, Error **errp);
> >> +
> >>  /**
> >>   * qio_channel_pwrite
> >>   * @ioc: the channel object
> >> @@ -595,6 +608,19 @@ ssize_t qio_channel_pwrite(QIOChannel *ioc, char *buf, size_t buflen,
> >>  ssize_t qio_channel_preadv(QIOChannel *ioc, const struct iovec *iov,
> >>                             size_t niov, off_t offset, Error **errp);
> >>  
> >> +/**
> >> + * qio_channel_preadv_all:
> >> + * @ioc: the channel object
> >> + * @iov: the array of memory regions to read data to
> >> + * @niov: the length of the @iov array
> >> + * @offset: the iovec offset in the file from where to read the data
> >> + * @errp: pointer to a NULL-initialized error object
> >> + *
> >> + * Returns: 0 if all bytes were read, or -1 on error
> >> + */
> >> +int qio_channel_preadv_all(QIOChannel *ioc, const struct iovec *iov,
> >> +                           size_t niov, off_t offset, Error **errp);
> >> +
> >>  /**
> >>   * qio_channel_pread
> >>   * @ioc: the channel object
> >> diff --git a/io/channel.c b/io/channel.c
> >> index a1f12f8e90..2f1745d052 100644
> >> --- a/io/channel.c
> >> +++ b/io/channel.c
> >> @@ -472,6 +472,69 @@ ssize_t qio_channel_pwritev(QIOChannel *ioc, const struct iovec *iov,
> >>      return klass->io_pwritev(ioc, iov, niov, offset, errp);
> >>  }
> >>  
> >> +static int qio_channel_preadv_pwritev_contiguous(QIOChannel *ioc,
> >> +                                                 const struct iovec *iov,
> >> +                                                 size_t niov, off_t offset,
> >> +                                                 bool is_write, Error **errp)
> >> +{
> >> +    ssize_t ret = -1;
> >> +    int i, slice_idx, slice_num;
> >> +    uintptr_t base, next, file_offset;
> >> +    size_t len;
> >> +
> >> +    slice_idx = 0;
> >> +    slice_num = 1;
> >> +
> >> +    /*
> >> +     * If the iov array doesn't have contiguous elements, we need to
> >> +     * split it in slices because we only have one (file) 'offset' for
> >> +     * the whole iov. Do this here so callers don't need to break the
> >> +     * iov array themselves.
> >> +     */
> >> +    for (i = 0; i < niov; i++, slice_num++) {
> >> +        base = (uintptr_t) iov[i].iov_base;
> >> +
> >> +        if (i != niov - 1) {
> >> +            len = iov[i].iov_len;
> >> +            next = (uintptr_t) iov[i + 1].iov_base;
> >> +
> >> +            if (base + len == next) {
> >> +                continue;
> >> +            }
> >> +        }
> >> +
> >> +        /*
> >> +         * Use the offset of the first element of the segment that
> >> +         * we're sending.
> >> +         */
> >> +        file_offset = offset + (uintptr_t) iov[slice_idx].iov_base;
> >> +
> >> +        if (is_write) {
> >> +            ret = qio_channel_pwritev(ioc, &iov[slice_idx], slice_num,
> >> +                                      file_offset, errp);
> >> +        } else {
> >> +            ret = qio_channel_preadv(ioc, &iov[slice_idx], slice_num,
> >> +                                     file_offset, errp);
> >> +        }
> >> +
> >> +        if (ret < 0) {
> >> +            break;
> >> +        }
> >> +
> >> +        slice_idx += slice_num;
> >> +        slice_num = 0;
> >> +    }
> >> +
> >> +    return (ret < 0) ? -1 : 0;
> >> +}
> >> +
> >> +int qio_channel_pwritev_all(QIOChannel *ioc, const struct iovec *iov,
> >> +                            size_t niov, off_t offset, Error **errp)
> >> +{
> >> +    return qio_channel_preadv_pwritev_contiguous(ioc, iov, niov,
> >> +                                                 offset, true, errp);
> >> +}
> >
> > I'm not sure how Dan thinks about this, but I don't think this is pretty..
> >
> > With this implementation, iochannels' preadv/pwritev is completely not
> > compatible with most OSes now, afaiu.
> 
> This is internal QEMU code. I hope no one is expecting qio_channel_foo()
> to behave like some OS's foo() system call. We cannot guarantee that
> compatibility save for the simplest of wrappers.

I was expecting that when I started to read. :)

https://man.freebsd.org/cgi/man.cgi?query=pwritev
https://linux.die.net/man/2/pwritev

It's not "some OSes", it's mostly all.  I can understand you prefer such
approach, but even if so, shall we still try to avoid using pwritev/preadv
as the names?

> 
> >
> > The definition of offset in preadv/pwritev of current iochannel is hard to
> > understand.. if I read it right it'll later be set to:
> >       
> >                 /*
> >                  * If we subtract the host page now, we don't need to
> >                  * pass it into qio_channel_pwritev_all() below.
> >                  */
> >                 write_base = p->pages->block->pages_offset -
> >                     (uintptr_t)p->pages->block->host;
> >
> > Which I cannot easily tell what it is.. besides being an unsigned int.
> 
> This description was unfortunately dropped along the way:
> 
> "Since iovs can be non contiguous, we'd need a separate array on the
> side to carry an extra file offset for each of them, so I'm relying on
> the fact that iovs are all within a same host page and passing in an
> encoded offset that takes the host page into account."
> 
> > IIUC it's also based on the assumption that the host address of each iov
> > entry is linear to its offset in the file, but it may not be true for
> > future iochannel users of such interface called as pwritev/preadv.  So
> > error prone.
> 
> Yes, but it's also our choice whether to make this a generic API. We may
> have good reasons to consider a migration-specific function here.
> 
> > Would it be possible we keep using the offset array (p->pages->offset[x])?
> > We have it already anyway, right?  Wouldn't that be clearer?
> >
> 
> We'd have to make a copy of the array because p->pages is expected to
> change while the IO happens.

Hmm, I don't see why p->pages can change. IIUC p->pages will be there solid
at least until all IO syscalls are completed, then the next call to, e.g.,
multifd_send_pages() will swap that with multifd_send_state->pages.  But I
think I get your point, with below.

> And while we already have a copy in
> p->normal, my intention for multifd was to eliminate p->normal in the
> future, so it would be nice if we could avoid it.
> 
> Also, we cannot use p->pages->offset alone because we still need the
> pages_offset, i.e. the file offset where that ramblocks's pages begin.
> So that means also adding that to each element of the new array.
> 
> It would probably be overall clearer and less wasteful to pass in the
> host page address instead of an array of offsets. I don't see an issue
> with restricting the iovs to the same host page. The migration code is
> the only user for this code and AFAIK we don't have plans to change that
> invariant.

So I think I get your point now, the only concern (besides naming..) is,
I still want to avoid an interface that contains a field that is hard to
understand like write_base.

How about this?

  /**
   * multifd_write_ramblock_iov: Write IO vector (of ramblock) to channel
   *
   * @ioc: The iochannel to write to. The IOC must have pwritev/preadv
   *       interface must be implemented.
   * @iov: The IO vector to write.  All addresses must be within the
   *       ramblock host address range.
   * @iov_len: The IO vector size
   * @ramblock: The ramblock that covers all buffers in this IO vector
   */
  int multifd_write_ramblock_iov(ioc, iov, iov_len, ramblock);

-- 
Peter Xu