Re: [PATCH v4] udp: support traceroute

[Stefano Brivio <sbrivio@redhat.com>]

Jon, I wasn't actually suggesting that you would drop *all* the Cc:'s.
:) I just saw no reason to specifically spam Laurent with this.

On Fri, 4 Apr 2025 10:31:29 +1100
David Gibson <david@gibson.dropbear.id.au> wrote:

> On Thu, Apr 03, 2025 at 04:27:12PM -0400, Jon Maloy wrote:
> > 
> > 
> > On 2025-04-03 11:48, Stefano Brivio wrote:  
> > > The implementation looks solid to me, a list of nits (or a bit
> > > more) below.
> > > 
> > > By the way, I don't think you need to Cc: people who are already on
> > > this list unless you specifically want their attention.
> > > 
> > > On Wed,  2 Apr 2025 22:22:29 -0400
> > > Jon Maloy <jmaloy@redhat.com> wrote:
> > >   
> > > > Now that ICMP pass-through from socket-to-tap is in place, it is
> > > > easy to support UDP based traceroute functionality in direction
> > > > tap-to-socket.
> > > > 
> > > > We fix that  in this commit.
> > > > 
> > > > Signed-off-by: Jon Maloy <jmaloy@redhat.com>  
> > > 
> > > This fixes https://bugs.passt.top/show_bug.cgi?id=64 ("Link:" tag) if I
> > > understood correctly.
> > >   
> > > > ---
> > > > v2: - Using ancillary data instead of setsockopt to transfer outgoing
> > > >        TTL.
> > > >      - Support IPv6
> > > > v3: - Storing ttl per packet instead of per flow. This may not be
> > > >        elegant, but much less intrusive than changing the flow  
> > 
> > [...]
> >   
> > > > @@ -11,6 +11,8 @@
> > > >   /* Maximum size of a single packet stored in pool, including headers */
> > > >   #define PACKET_MAX_LEN	((size_t)UINT16_MAX)
> > > > +#define DEFAULT_TTL 64  
> > > 
> > > If I understood correctly, David's comment to this on v3:
> > > 
> > >    https://archives.passt.top/passt-dev/Z-om3Ey-HR1Hj8UH@zatzit/
> > > 
> > > was meant to imply that, as the default value can be changed via
> > > sysctl, the value set via sysctl could be read at start-up. I'm fine
> > > with 64 as well, by the way, with a slight preference for reading the
> > > value via sysctl.  
> > 
> > I don't think the local host/container setting will have any effect
> > if the sending guest is a VM.  
> 
> That's true, but..
> 
> > The benefit is of this is dubious.  
> 
> .. uflow->ttl[] isn't so much representing what the guest set, as a
> cache of what the socket is sending and that *does* depend on the host
> value.

Right, my concern is that now we'll use the host value (whatever it is)
if the value from the container / guest is 64.

So:

- guest uses 63, host has 255 configured: we use 63

- guest uses 64, host has 64 configured: we use 64

- ...but: guest uses 64, host has 255 configured: we use 255

...and this might actually break traceroute itself in some extreme
cases.

Let's say we have 255 configured on the host and you're in the middle
of a traceroute:

- guest sends TTL 62, goes out with 62 -> 62nd hop replies
- guest sends TTL 63, goes out with 63 -> 63rd hop replies
- guest sends TTL 64, goes out with 255 -> destination replies
- guest sends TTL 65, goes out with 65 -> 65th hop replies, traceroute broken

See also the comment below.

> > > All this might go away, though, please read the comment to
> > > udp_flow_new() below, first.
> > >   
> > > > +
> > > >   /**
> > > >    * struct pool - Generic pool of packets stored in a buffer
> > > >    * @buf:	Buffer storing packet descriptors,
> > > > diff --git a/tap.c b/tap.c
> > > > index 3a6fcbe..e65d592 100644
> > > > --- a/tap.c
> > > > +++ b/tap.c
> > > > @@ -563,6 +563,7 @@ PACKET_POOL_DECL(pool_l4, UIO_MAXIOV, pkt_buf);
> > > >    * @dest:	Destination port
> > > >    * @saddr:	Source address
> > > >    * @daddr:	Destination address
> > > > + * @ttl:	Time to live
> > > >    * @msg:	Array of messages that can be handled in a single call
> > > >    */
> > > >   static struct tap4_l4_t {
> > > > @@ -574,6 +575,8 @@ static struct tap4_l4_t {
> > > >   	struct in_addr saddr;
> > > >   	struct in_addr daddr;
> > > > +	uint8_t ttl;  
> > > 
> > > If you move this after 'protocol' you save 4 or 8 bytes depending on
> > > the architecture and, perhaps more importantly, with 64-byte cachelines,
> > > you can fit the set of fields involved in the L4_MATCH() comparison
> > > four times instead of three. If you have a look with pahole(1):
> > >   
> > Good point. I  didn't notice.
> > 
> > 
> > [...]  
> > > >   	const struct flowside *toside;
> > > >   	struct mmsghdr mm[UIO_MAXIOV];
> > > > @@ -938,6 +940,19 @@ int udp_tap_handler(const struct ctx *c, uint8_t pif,
> > > >   		mm[i].msg_hdr.msg_controllen = 0;
> > > >   		mm[i].msg_hdr.msg_flags = 0;
> > > > +		if (ttl != uflow->ttl[tosidx.sidei]) {
> > > > +			uflow->ttl[tosidx.sidei] = ttl;
> > > > +			if (af == AF_INET) {
> > > > +				if (setsockopt(s, IPPROTO_IP, IP_TTL,
> > > > +					       &ttl, sizeof(ttl)) < 0)
> > > > +					perror("setsockopt (IP_TTL)");  
> > > 
> > > This would print to file descriptor 2 even if it's a socket. It should
> > > be err_perror() instead, but now we also have flow_perror() which
> > > prints flow index and type, given 'uflow' here, say:
> > > 
> > > 					flow_perror(uflow, "IP_TTL setsockopt");
> > >   
> > > > +			} else {
> > > > +				if (setsockopt(s, IPPROTO_IPV6, IPV6_HOPLIMIT,
> > > > +					       &ttl, sizeof(ttl)) < 0)
> > > > +					perror("setsockopt (IP_TTL)");  
> > > 
> > > ...and this is IPV6_HOPLIMIT, not IP_TTL, so perhaps:
> > > 
> > > 					flow_perror(uflow,
> > > 						    "setsockopt IPV6_HOPLIMIT");
> > >   
> > Ok.
> >   
> > > > +			}
> > > > +		}
> > > > +
> > > >   		count++;
> > > >   	}
> > > > diff --git a/udp.h b/udp.h
> > > > index de2df6d..041fad4 100644
> > > > --- a/udp.h
> > > > +++ b/udp.h
> > > > @@ -15,7 +15,8 @@ void udp_reply_sock_handler(const struct ctx *c, union epoll_ref ref,
> > > >   			    uint32_t events, const struct timespec *now);
> > > >   int udp_tap_handler(const struct ctx *c, uint8_t pif,
> > > >   		    sa_family_t af, const void *saddr, const void *daddr,
> > > > -		    const struct pool *p, int idx, const struct timespec *now);
> > > > +		    uint8_t  ttl, const struct pool *p, int idx,  
> > > 
> > > Excess whitespace beetween 'uint8_t' and 'ttl'.
> > >   
> > > > +		    const struct timespec *now);
> > > >   int udp_sock_init(const struct ctx *c, int ns, const union inany_addr *addr,
> > > >   		  const char *ifname, in_port_t port);
> > > >   int udp_init(struct ctx *c);
> > > > diff --git a/udp_flow.c b/udp_flow.c
> > > > index bf4b896..39372c2 100644
> > > > --- a/udp_flow.c
> > > > +++ b/udp_flow.c
> > > > @@ -137,6 +137,7 @@ static flow_sidx_t udp_flow_new(const struct ctx *c, union flow *flow,
> > > >   	uflow = FLOW_SET_TYPE(flow, FLOW_UDP, udp);
> > > >   	uflow->ts = now->tv_sec;
> > > >   	uflow->s[INISIDE] = uflow->s[TGTSIDE] = -1;
> > > > +	uflow->ttl[INISIDE] = uflow->ttl[TGTSIDE] = DEFAULT_TTL;  
> > > 
> > > By the way, instead of using a default value, what about fetching the
> > > current value with getsockopt()?
> > > 
> > > One additional system call per UDP flow doesn't feel like a lot of
> > > overhead, and we can be sure it's correct, no matter if the user
> > > configures a different value before or after we start.
> > >   
> > This patch fixes UDP messaging tap->socket, and TTL may have any
> > value in the first arriving packet. Reading it from the socket here only
> > makes sense when I add the same support in direction socket->tap.
> > That is my next project.

Well, wait, the getsockopt() will not tell you the value the socket is
receiving. It tells you the value that the socket would send, at least
according to the documentation:

       IP_TTL (since Linux 1.0)
              Set or retrieve the current time-to-live field that is
              used in every packet sent from this socket.

and that's what makes it relevant: this is the value that we would
normally use, unless you issue the setsockopt().

But... there's a plot twist: this is just for IPv4. For IPv6:

       IPV6_RTHDR, IPV6_AUTHHDR, IPV6_DSTOPTS, IPV6_HOPOPTS, IPV6_FLOWINFO,
       IPV6_HOPLIMIT
              Set delivery of control messages for incoming datagrams
              containing extension headers from the received packet.

              [...]

              IPV6_HOPLIMIT delivers an integer containing the hop count of
              the packet.

so I wonder: is it correct to use IPV6_HOPLIMIT at all, even for the
setsockopt() you're adding?

I haven't tested this (at least not yet), but from the documentation
that seems to apply to *received* packets. No idea what the
setsockopt() would do, at this point.

Could it be that we should use IP_TTL for *sent* IPv4 packets as well?

I'll try to test this specific part in a bit, unless you already did.

-- 
Stefano