Re: [PATCH 3/4] udp: Rework how we divide queued datagrams between sending methods

[David Gibson <david@gibson.dropbear.id.au>]

[-- Attachment #1: Type: text/plain, Size: 7617 bytes --]

On Thu, Jun 13, 2024 at 08:21:12PM +0200, Stefano Brivio wrote:
> On Wed,  5 Jun 2024 11:39:02 +1000
> David Gibson <david@gibson.dropbear.id.au> wrote:
> 
> > udp_sock_handler() takes a number of datagrams from sockets that depending
> > on their addresses could be forwarded either to the L2 interface ("tap")
> > or to another socket ("spliced").  In the latter case we can also only
> > send packets together if they have the same source port, and therefore
> > are sent via the same socket.
> > 
> > To reduce the total number of system calls we gather contiguous batches of
> > datagrams with the same destination interface and socket where applicable.
> > The determination of what the target is is made by udp_mmh_splice_port().
> > It returns the source port for splice packets and -1 for "tap" packets.
> > We find batches by looking ahead in our queue until we find a datagram
> > whose "splicefrom" port doesn't match the first in our current batch.
> > 
> > udp_mmh_splice_port() is moderately expensive, since it must examine IPv6
> > addresses.  But unfortunately we can call it twice on the same datagram:
> > once as the (last + 1) entry in one batch (showing that it's not in that
> > match, then again as the first entry in the next batch.
> 
> This paragraph took me an embarrassingly long time to grasp, if you
> re-spin it would be nice to fix it:
> 
> - "And unfortunately [...]", I guess: otherwise it looks like we're
>   lucky that udp_mmh_splice_port() is expensive or something like that
>   (because of the "But" implying contrast).
> 
>   I initially assumed "unfortunately" was a typo and tried to
>   understand why it was a good thing we'd call udp_mmh_splice_port()
>   twice on the same datagram (faster than calling it on two
>   datagrams!), then started reading the change and got even more
>   confused...
> 
> - "(to check that it's not that batch)" ?

Yeah, didn't help that there were some other typos in there.  Updated
referring to these suggestions.

> > Avoid this by keeping track of the "splice port" in the metadata structure,
> > and filling it in one entry ahead of the one we're currently considering.
> > This is a bit subtle, but not that hard.  It will also generalise better
> > when we have more complex possibilities based on the flow table.
> 
> I guess this is the actual, main reason for this change. :) I should
> have read this paragraph first.

Yes.  In the flow table we'll replace this splicesrc array with an
array of flow sidxs, updated in the same way.

> > Signed-off-by: David Gibson <david@gibson.dropbear.id.au>
> > ---
> >  udp.c | 147 ++++++++++++++++++++++++++++++++++------------------------
> >  1 file changed, 86 insertions(+), 61 deletions(-)
> > 
> > diff --git a/udp.c b/udp.c
> > index 7487d2b2..757c10ab 100644
> > --- a/udp.c
> > +++ b/udp.c
> > @@ -198,6 +198,7 @@ static struct ethhdr udp6_eth_hdr;
> >   * @ip4h:	Pre-filled IPv4 header (except for tot_len and saddr)
> >   * @taph:	Tap backend specific header
> >   * @s_in:	Source socket address, filled in by recvmmsg()
> > + * @splicesrc:	Source port for splicing, or -1 if not spliceable
> >   */
> >  static struct udp_meta_t {
> >  	struct ipv6hdr ip6h;
> > @@ -205,6 +206,7 @@ static struct udp_meta_t {
> >  	struct tap_hdr taph;
> >  
> >  	union sockaddr_inany s_in;
> > +	int splicesrc;
> >  }
> >  #ifdef __AVX2__
> >  __attribute__ ((aligned(32)))
> > @@ -492,28 +494,32 @@ static int udp_mmh_splice_port(union udp_epoll_ref uref,
> >  }
> >  
> >  /**
> > - * udp_splice_sendfrom() - Send datagrams from given port to given port
> > + * udp_splice_send() - Send datagrams from socket to socket
> >   * @c:		Execution context
> >   * @start:	Index of first datagram in udp[46]_l2_buf
> > - * @n:		Number of datagrams to send
> > - * @src:	Datagrams will be sent from this port (on origin side)
> > - * @dst:	Datagrams will be send to this port (on destination side)
> > - * @from_pif:	pif from which the packet originated
> > - * @v6:		Send as IPv6?
> > - * @allow_new:	If true create sending socket if needed, if false discard
> > - *              if no sending socket is available
> > + * @n:		Total number of datagrams in udp[46]_l2_buf pool
> > + * @dst:	Datagrams will be sent to this port (on destination side)
> > + * @uref:	UDP epoll reference for origin socket
> >   * @now:	Timestamp
> > + *
> > + * This consumes as many frames as are sendable via a single socket.  It
> 
> s/frames/datagrams/ ...or messages.

Good point, adjusted.

> > + * requires that udp_meta[@start].splicesrc is initialised, and will initialise
> > + * udp_meta[].splicesrc for each frame it consumes *and one more* (if present).
> > + *
> > + * Return: Number of frames sent
> 
> I'd say it's rather the number of datagrams (not frames) we tried to
> send.
> 
> In some sense, it's also the number of frames sent _by us_ (well, after
> calling sendmmsg(), messages were sent), but we call sendmmsg()
> ignoring the result, so this comment might look a bit misleading.

Right... I've gone with "Number of datagrams forwarded" how's that?

> >   */
> > -static void udp_splice_sendfrom(const struct ctx *c, unsigned start, unsigned n,
> > -				in_port_t src, in_port_t dst, uint8_t from_pif,
> > -				bool v6, bool allow_new,
> > +static unsigned udp_splice_send(const struct ctx *c, size_t start, size_t n,
> > +				in_port_t dst, union udp_epoll_ref uref,
> >  				const struct timespec *now)
> >  {
> > +	in_port_t src = udp_meta[start].splicesrc;
> >  	struct mmsghdr *mmh_recv, *mmh_send;
> > -	unsigned int i;
> > +	unsigned int i = start;
> >  	int s;
> >  
> > -	if (v6) {
> > +	ASSERT(udp_meta[start].splicesrc >= 0);
> > +
> > +	if (uref.v6) {
> >  		mmh_recv = udp6_l2_mh_sock;
> >  		mmh_send = udp6_mh_splice;
> >  	} else {
> > @@ -521,40 +527,48 @@ static void udp_splice_sendfrom(const struct ctx *c, unsigned start, unsigned n,
> >  		mmh_send = udp4_mh_splice;
> >  	}
> >  
> > -	if (from_pif == PIF_SPLICE) {
> > +	do {
> > +		mmh_send[i].msg_hdr.msg_iov->iov_len = mmh_recv[i].msg_len;
> > +
> > +		if (++i >= n)
> > +			break;
> > +
> > +		udp_meta[i].splicesrc = udp_mmh_splice_port(uref, &mmh_recv[i]);
> > +	} while (udp_meta[i].splicesrc == src);
> 
> I don't have a strong preference, but a for loop like this:
> 
> 	for (; i < n && udp_meta[i].splicesrc == src; i++) {
> 		mmh_send[i].msg_hdr.msg_iov->iov_len = mmh_recv[i].msg_len;
> 		udp_meta[i].splicesrc = udp_mmh_splice_port(uref, &mmh_recv[i]);

This needs to update udp_meta[i+1], not udp_meta[i], and therefore
also be conditional on i+1 < n.

> 	}
> 
> 	if (i++ < n)	/* Set splicesrc for first mismatching entry, too */

This needs to be ++i, not i++.

> 		udp_meta[i].splicesrc = udp_mmh_splice_port(uref, &mmh_recv[i]);
> 
> looks a bit more readable to me. Same for udp_tap_send().

At which point I'm not sure it's more readable after all.  It also
redundantly checks that udp_meta[start].splicesrc == src.

Like I said, subtle.  Putting the increment in the middle of the loop
body, rather than beginning or end, while unusual, was the sanest way
I could see to do this.  Well, other than havine one pass to set
splicesrc[], then another using it, which works but I'm concerned
about the effect on cache locality.

-- 
David Gibson			| I'll have my music baroque, and my code
david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
				| _way_ _around_!
http://www.ozlabs.org/~dgibson

[-- Attachment #2: signature.asc --]
[-- Type: application/pgp-signature, Size: 833 bytes --]