Re: [PATCH 1/3] udp: Merge udp[46]_mh_recv arrays

[Stefano Brivio <sbrivio@redhat.com>]

On Tue, 27 Aug 2024 11:12:46 +1000
David Gibson <david@gibson.dropbear.id.au> wrote:

> On Mon, Aug 26, 2024 at 09:32:55PM +0200, Stefano Brivio wrote:
> > On Mon, 26 Aug 2024 19:37:14 +1000
> > David Gibson <david@gibson.dropbear.id.au> wrote:
> >   
> > > We've already gotten rid of most of the IPv4/IPv6 specific data structures
> > > in udp.c by merging them with each other.  One significant one remains:
> > > udp[46]_mh_recv.  This was a bit awkward to remove because of a subtle
> > > interaction.  We initialise the msg_namelen fields to represent the total
> > > size we have for a socket address, but when we receive into the arrays
> > > those are modified to the actual length of the sockaddr we received.
> > > 
> > > That meant that naively merging the arrays meant that if we received IPv4
> > > datagrams, then IPv6 datagrams, the addresses for the latter would be
> > > truncated.  In this patch address that by resetting the received
> > > msg_namelen as soon as we've found a flow for the datagram.  Finding the
> > > flow is the only thing that might use the actual sockaddr length, although
> > > we in fact don't need it for the time being.
> > > 
> > > This also removes the last use of the 'v6' field from udp_listen_epoll_ref,
> > > so remove that as well.
> > > 
> > > Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
> > > ---
> > >  udp.c | 57 ++++++++++++++++++++-------------------------------------
> > >  udp.h |  2 --
> > >  2 files changed, 20 insertions(+), 39 deletions(-)
> > > 
> > > diff --git a/udp.c b/udp.c
> > > index 8a93aad6..6638c22b 100644
> > > --- a/udp.c
> > > +++ b/udp.c
> > > @@ -178,8 +178,7 @@ enum udp_iov_idx {
> > >  
> > >  /* IOVs and msghdr arrays for receiving datagrams from sockets */
> > >  static struct iovec	udp_iov_recv		[UDP_MAX_FRAMES];
> > > -static struct mmsghdr	udp4_mh_recv		[UDP_MAX_FRAMES];
> > > -static struct mmsghdr	udp6_mh_recv		[UDP_MAX_FRAMES];
> > > +static struct mmsghdr	udp_mh_recv		[UDP_MAX_FRAMES];
> > >  
> > >  /* IOVs and msghdr arrays for sending "spliced" datagrams to sockets */
> > >  static union sockaddr_inany udp_splice_to;
> > > @@ -222,6 +221,7 @@ void udp_update_l2_buf(const unsigned char *eth_d, const unsigned char *eth_s)
> > >  static void udp_iov_init_one(const struct ctx *c, size_t i)
> > >  {
> > >  	struct udp_payload_t *payload = &udp_payload[i];
> > > +	struct msghdr *mh = &udp_mh_recv[i].msg_hdr;
> > >  	struct udp_meta_t *meta = &udp_meta[i];
> > >  	struct iovec *siov = &udp_iov_recv[i];
> > >  	struct iovec *tiov = udp_l2_iov[i];
> > > @@ -236,27 +236,10 @@ static void udp_iov_init_one(const struct ctx *c, size_t i)
> > >  	tiov[UDP_IOV_TAP] = tap_hdr_iov(c, &meta->taph);
> > >  	tiov[UDP_IOV_PAYLOAD].iov_base = payload;
> > >  
> > > -	/* It's useful to have separate msghdr arrays for receiving.  Otherwise,
> > > -	 * an IPv4 recv() will alter msg_namelen, so we'd have to reset it every
> > > -	 * time or risk truncating the address on future IPv6 recv()s.
> > > -	 */
> > > -	if (c->ifi4) {
> > > -		struct msghdr *mh = &udp4_mh_recv[i].msg_hdr;
> > > -
> > > -		mh->msg_name	= &meta->s_in;
> > > -		mh->msg_namelen	= sizeof(struct sockaddr_in);
> > > -		mh->msg_iov	= siov;
> > > -		mh->msg_iovlen	= 1;
> > > -	}
> > > -
> > > -	if (c->ifi6) {
> > > -		struct msghdr *mh = &udp6_mh_recv[i].msg_hdr;
> > > -
> > > -		mh->msg_name	= &meta->s_in;
> > > -		mh->msg_namelen	= sizeof(struct sockaddr_in6);
> > > -		mh->msg_iov	= siov;
> > > -		mh->msg_iovlen	= 1;
> > > -	}
> > > +	mh->msg_name	= &meta->s_in;
> > > +	mh->msg_namelen	= sizeof(meta->s_in);
> > > +	mh->msg_iov	= siov;
> > > +	mh->msg_iovlen	= 1;
> > >  }
> > >  
> > >  /**
> > > @@ -506,10 +489,10 @@ static int udp_sock_recv(const struct ctx *c, int s, uint32_t events,
> > >  void udp_listen_sock_handler(const struct ctx *c, union epoll_ref ref,
> > >  			     uint32_t events, const struct timespec *now)
> > >  {
> > > -	struct mmsghdr *mmh_recv = ref.udp.v6 ? udp6_mh_recv : udp4_mh_recv;
> > > +	const socklen_t sasize = sizeof(udp_meta[0].s_in);
> > >  	int n, i;
> > >  
> > > -	if ((n = udp_sock_recv(c, ref.fd, events, mmh_recv)) <= 0)
> > > +	if ((n = udp_sock_recv(c, ref.fd, events, udp_mh_recv)) <= 0)
> > >  		return;
> > >  
> > >  	/* We divide datagrams into batches based on how we need to send them,
> > > @@ -518,6 +501,7 @@ void udp_listen_sock_handler(const struct ctx *c, union epoll_ref ref,
> > >  	 * populate it one entry *ahead* of the loop counter.
> > >  	 */
> > >  	udp_meta[0].tosidx = udp_flow_from_sock(c, ref, &udp_meta[0].s_in, now);
> > > +	udp_mh_recv[0].msg_hdr.msg_namelen = sasize;  
> > 
> > I don't understand why you need this assignment. To me it looks
> > redundant with:
> > 
> >   udp_mh_recv[i].msg_hdr.msg_namelen = sizeof(udp_meta[i].s_in);  
> 
> It's not redundant per se, because the later assignment only occurs
> for i > 0, so the first one is for slot 0.

I still don't see how: the second assignment (out of three) is done
before i is incremented, so that should cover i == 0 as well, right?

> It would, however, be
> possible to move to a single assignment in the loop body before i is
> incremented.
> 
> I did it this way, because I found it easier to reason about.  At
> least theoretically the value of msg_namelen written by recvmmsg()
> could be important, although we don't use yet (we rely on the
> sa_family field instead).  But because of that it felt wrong to
> overwrite that value before we've "consumed" it.  Logically that
> happens in udp_flow_from_sock() which is what takes the address in
> msg_name / msg_namelen and converts it into the long-term form (as
> part of the flowside).  Hence, clearing msg_namelen immediately after
> each call to udp_flow_from_sock() made sense to me.
> 
> I did consider changing udp_flow_from_sock() to take a socklen_t *
> which it clears after using.  That seemed slightly abstraction
> violationy to me: clearing msg_namelen only makes sense because the
> address is part of a re-used mmsghdr array, and that's not something
> udp_flow_from_sock() "knows".
> 
> That was my reasoning, anyway.  I'm happy enough to change it if you
> have a preferred approach.

No, no, this all makes sense. But you add three assignments here, and I
don't understand why #1 is needed if we have #2 and #3, or why #2 is
needed if we have #1 and #3.

> > later (because n > 0), and:
> >   
> > >  	for (i = 0; i < n; ) {
> > >  		flow_sidx_t batchsidx = udp_meta[i].tosidx;
> > >  		uint8_t batchpif = pif_at_sidx(batchsidx);
> > > @@ -525,18 +509,22 @@ void udp_listen_sock_handler(const struct ctx *c, union epoll_ref ref,
> > >  
> > >  		do {
> > >  			if (pif_is_socket(batchpif)) {
> > > -				udp_splice_prepare(mmh_recv, i);
> > > +				udp_splice_prepare(udp_mh_recv, i);
> > >  			} else if (batchpif == PIF_TAP) {
> > > -				udp_tap_prepare(mmh_recv, i,
> > > +				udp_tap_prepare(udp_mh_recv, i,
> > >  						flowside_at_sidx(batchsidx));
> > >  			}
> > >  
> > > +			/* Restore sockaddr length clobbered by recvmsg() */
> > > +			udp_mh_recv[i].msg_hdr.msg_namelen = sizeof(udp_meta[i].s_in);  
> > 
> > what is the difference between assigning sizeof(udp_meta[i].s_in); and
> > sasize? I thought it would be the same quantity.  
> 
> It is.  The only purpose of sasize is to avoid some over-long lines.

Right, but why do you use it just twice out of three assignments? What
is special with the one immediately above here?

> > > +
> > >  			if (++i >= n)
> > >  				break;
> > >  
> > >  			udp_meta[i].tosidx = udp_flow_from_sock(c, ref,
> > >  								&udp_meta[i].s_in,
> > >  								now);
> > > +			udp_mh_recv[i].msg_hdr.msg_namelen = sasize;
> > >  		} while (flow_sidx_eq(udp_meta[i].tosidx, batchsidx));
> > >  
> > >  		if (pif_is_socket(batchpif)) {  

-- 
Stefano