Re: [PATCH v1 2/3] io: Add zerocopy and errqueue

[Peter Xu]

On Wed, Sep 01, 2021 at 09:50:56AM +0100, Daniel P. Berrangé wrote:
> On Tue, Aug 31, 2021 at 04:27:04PM -0400, Peter Xu wrote:
> > On Tue, Aug 31, 2021 at 01:57:33PM +0100, Daniel P. Berrangé wrote:
> > > On Tue, Aug 31, 2021 at 08:02:38AM -0300, Leonardo Bras wrote:
> > > > MSG_ZEROCOPY is a feature that enables copy avoidance in TCP/UDP socket
> > > > send calls. It does so by avoiding copying user data into kernel 
> > > > buffers.
> > > > 
> > > > To make it work, three steps are needed:
> > > > 1 - A setsockopt() system call, enabling SO_ZEROCOPY
> > > > 2 - Passing down the MSG_ZEROCOPY flag for each send*() syscall
> > > > 3 - Process the socket's error queue, dealing with any error
> > > 
> > > AFAICT, this is missing the single most critical aspect of MSG_ZEROCOPY.
> > > 
> > > It is non-obvious, but setting the MSG_ZEROCOPY flag turns sendmsg()
> > > from a synchronous call to an asynchronous call.
> > > 
> > > It is forbidden to overwrite/reuse/free the buffer passed to sendmsg
> > > until an asynchronous completion notification has been received from
> > > the socket error queue. These notifications are not required to
> > > arrive in-order, even for a TCP stream, because the kernel hangs on
> > > to the buffer if a re-transmit is needed.
> > > 
> > > https://www.kernel.org/doc/html/v5.4/networking/msg_zerocopy.html
> > > 
> > >   "Page pinning also changes system call semantics. It temporarily 
> > >    shares the buffer between process and network stack. Unlike with
> > >    copying, the process cannot immediately overwrite the buffer 
> > >    after system call return without possibly modifying the data in 
> > >    flight. Kernel integrity is not affected, but a buggy program
> > >    can possibly corrupt its own data stream."
> > > 
> > > AFAICT, the design added in this patch does not provide any way
> > > to honour these requirements around buffer lifetime.
> > > 
> > > I can't see how we can introduce MSG_ZEROCOPY in any seemless
> > > way. The buffer lifetime requirements imply need for an API
> > > design that is fundamentally different for asynchronous usage,
> > > with a callback to notify when the write has finished/failed.
> > 
> > Regarding buffer reuse - it indeed has a very deep implication on the buffer
> > being available and it's not obvious at all.  Just to mention that the 
> > initial
> > user of this work will make sure all zero copy buffers will be guest pages 
> > only
> > (as it's only used in multi-fd), so they should always be there during the
> > process.
> 
> That is not the case when migration is using TLS, because the buffers
> transmitted are the ciphertext buffer, not the plaintext guest page.

Agreed.

> 
> > In short, we may just want to at least having a way to make sure all zero
> > copied buffers are finished using and they're sent after some function 
> > returns
> > (e.g., qio_channel_flush()).  That may require us to do some accounting on 
> > when
> > we called sendmsg(MSG_ZEROCOPY), meanwhile we should need to read out the
> > ee_data field within SO_EE_ORIGIN_ZEROCOPY msg when we do recvmsg() for the
> > error queue and keep those information somewhere too.
> > 
> > Some other side notes that reached my mind..
> > 
> > The qio_channel_writev_full() may not be suitable for async operations, as 
> > the
> > name "full" implies synchronous to me.  So maybe we can add a new helper for
> > zero copy on the channel?
> 
> All the APIs that exist today are fundamentally only suitable for sync
> operations. Supporting async correctly will definitely a brand new APIs
> separate from what exists today.

Yes.  What I wanted to say is maybe we can still keep the io_writev() interface
untouched, but just add a new interface at qio_channel_writev_full() level.

IOW, we could comment on io_writev() that it can be either sync or async now,
just like sendmsg() has that implication too with different flag passed in.
When calling io_writev() with different upper helpers, QIO channel could
identify what flag to pass over to io_writev().

-- 
Peter Xu