Re: [PATCH v3 09/14] nbd/server: Initial support for extended headers

[Eric Blake]

On Wed, May 31, 2023 at 05:46:55PM +0300, Vladimir Sementsov-Ogievskiy wrote:
> On 15.05.23 22:53, Eric Blake wrote:
> > Time to support clients that request extended headers.  Now we can
> > finally reach the code added across several previous patches.
> > 
> > Even though the NBD spec has been altered to allow us to accept
> > NBD_CMD_READ larger than the max payload size (provided our response
> > is a hole or broken up over more than one data chunk), we are not
> > planning to take advantage of that, and continue to cap NBD_CMD_READ
> > to 32M regardless of header size.
> > 
> > For NBD_CMD_WRITE_ZEROES and NBD_CMD_TRIM, the block layer already
> > supports 64-bit operations without any effort on our part.  For
> > NBD_CMD_BLOCK_STATUS, the client's length is a hint, and the previous
> > patch took care of implementing the required
> > NBD_REPLY_TYPE_BLOCK_STATUS_EXT.
> > 
> > Signed-off-by: Eric Blake <eblake@redhat.com>
> > ---
> >   nbd/nbd-internal.h |   5 +-
> 
> [..]
> 
> > 
> >   static inline void set_be_simple_reply(NBDClient *client, struct iovec 
> > *iov,
> > -                                       uint64_t error, NBDRequest *request)
> > +                                       uint32_t error, NBDStructuredError 
> > *err,
> > +                                       NBDRequest *request)
> >   {
> > -    NBDSimpleReply *reply = iov->iov_base;
> > +    if (client->header_style >= NBD_HEADER_EXTENDED) {
> > +        NBDExtendedReplyChunk *chunk = iov->iov_base;
> > 
> > -    iov->iov_len = sizeof(*reply);
> > -    stl_be_p(&reply->magic, NBD_SIMPLE_REPLY_MAGIC);
> > -    stl_be_p(&reply->error, error);
> > -    stq_be_p(&reply->handle, request->handle);
> > +        iov->iov_len = sizeof(*chunk);
> > +        stl_be_p(&chunk->magic, NBD_EXTENDED_REPLY_MAGIC);
> > +        stw_be_p(&chunk->flags, NBD_REPLY_FLAG_DONE);
> > +        stq_be_p(&chunk->handle, request->handle);
> > +        stq_be_p(&chunk->offset, request->from);
> > +        if (error) {
> > +            assert(!iov[1].iov_base);
> > +            iov[1].iov_base = err;
> > +            iov[1].iov_len = sizeof(*err);
> > +            stw_be_p(&chunk->type, NBD_REPLY_TYPE_ERROR);
> > +            stq_be_p(&chunk->length, sizeof(*err));
> > +            stl_be_p(&err->error, error);
> > +            stw_be_p(&err->message_length, 0);
> > +        } else {
> > +            stw_be_p(&chunk->type, NBD_REPLY_TYPE_NONE);
> > +            stq_be_p(&chunk->length, 0);
> > +        }
> > +    } else {
> > +        NBDSimpleReply *reply = iov->iov_base;
> > +
> > +        iov->iov_len = sizeof(*reply);
> > +        stl_be_p(&reply->magic, NBD_SIMPLE_REPLY_MAGIC);
> > +        stl_be_p(&reply->error, error);
> > +        stq_be_p(&reply->handle, request->handle);
> > +    }
> >   }
> > 
> >   static int coroutine_fn nbd_co_send_simple_reply(NBDClient *client,
> > @@ -1906,30 +1966,44 @@ static int coroutine_fn 
> > nbd_co_send_simple_reply(NBDClient *client,
> 
> So, that's not _simple_ now.. The function should be renamed. As well as 
> set_be_simple_reply(). _simple_or_extended_ ? a bit too long. But continuing 
> to use "simple" is in bad relation with use of "simple" word in specification.

In fact, I added an assertion that set_be_simple_reply() can only be
reached when extended replies are not in use, so none of this
complexity here was needed after all.

> 
> Probably better to update callers? The only caller isi 
> nbd_send_generic_reply(). So, could we just add 
> nbd_co_send_single_extended_reply() to call from nbd_send_generic_reply() in 
> case of EXTENDED?
> 
> Also, transformation of set_be_simple_reply() do look like it should be two 
> separate functions.
> 
> >   {
> >       NBDReply hdr;
> >       int nbd_err = system_errno_to_nbd_errno(error);
> > +    NBDStructuredError err;
> >       struct iovec iov[] = {
> >           {.iov_base = &hdr},
> >           {.iov_base = data, .iov_len = len}
> >       };
> > 
> > +    assert(!len || !nbd_err);
> >       trace_nbd_co_send_simple_reply(request->handle, nbd_err,
> >                                      nbd_err_lookup(nbd_err), len);
> > -    set_be_simple_reply(client, &iov[0], nbd_err, request);
> > +    set_be_simple_reply(client, &iov[0], nbd_err, &err, request);
> > 
> > -    return nbd_co_send_iov(client, iov, len ? 2 : 1, errp);
> > +    return nbd_co_send_iov(client, iov, iov[1].iov_len ? 2 : 1, errp);

Not introduced in this patch, but it turns out that when
iov[1].iov_len == 0, blindly passing niov==2 to nbd_co_send_iov()
still does the right thing, so I can lose the conditional here.

-- 
Eric Blake, Principal Software Engineer
Red Hat, Inc.           +1-919-301-3266
Virtualization:  qemu.org | libvirt.org