svn commit: r280971 - in head: contrib/ipfilter/tools share/man/man4 sys/contrib/ipfilter/netinet sys/netinet sys/netipsec sys/netpfil/pf

[Mateusz Guzik]

On Thu, Apr 02, 2015 at 03:35:22PM +0300, Gleb Smirnoff wrote:
> On Wed, Apr 01, 2015 at 05:07:56PM -0600, Ian Lepore wrote:
> I> > Author: glebius
> I> > Date: Wed Apr  1 22:26:39 2015
> I> > New Revision: 280971
> I> > URL: https://svnweb.freebsd.org/changeset/base/280971
> I> > 
> I> > Log:
> I> >   o Use new function ip_fillid() in all places throughout the kernel,
> I> >     where we want to create a new IP datagram.
> I> >   o Add support for RFC6864, which allows to set IP ID for atomic IP
> I> >     datagrams to any value, to improve performance. The behaviour is
> I> >     controlled by net.inet.ip.rfc6864 sysctl knob, which is enabled by
> I> >     default.
> I> >   o In case if we generate IP ID, use counter(9) to improve performance.
> I> >   o Gather all code related to IP ID into ip_id.c.
> I> >   
> I> >   Differential Revision:		https://reviews.freebsd.org/D2177
> I> >   Reviewed by:			adrian, cy, rpaulo
> I> >   Tested by:			Emeric POUPON <emeric.poupon stormshield.eu>
> I> >   Sponsored by:			Netflix
> I> >   Sponsored by:			Nginx, Inc.
> I> >   Relnotes:			yes
> I> > 
> I> [...]
> I> > +void
> I> > +ip_fillid(struct ip *ip)
> I> > +{
> I> > +
> I> > +	/*
> I> > +	 * Per RFC6864 Section 4
> I> > +	 *
> I> > +	 * o  Atomic datagrams: (DF==1) && (MF==0) && (frag_offset==0)
> I> > +	 * o  Non-atomic datagrams: (DF==0) || (MF==1) || (frag_offset>0)
> I> > +	 */
> I> > +	if (V_ip_rfc6864 && (ip->ip_off & htons(IP_DF)) == htons(IP_DF))
> I> > +		ip->ip_id = 0;
> I> > +	else if (V_ip_do_randomid)
> I> > +		ip->ip_id = ip_randomid();
> I> > +	else {
> I> > +		counter_u64_add(V_ip_id, 1);
> I> > +		ip->ip_id = htons((*(uint64_t *)zpcpu_get(V_ip_id)) & 0xffff);
> I> > +	}
> I> > +}
> I> > +
> I> 
> I> This is completely bogus.  It's a big opacity violation (it relies on
> I> what should be opaque private internal implementation details of
> I> counter(9)).  The fact that the counter api doesn't provide a function
> I> for retrieving one cpu's counter value should be a big clue there -- the
> I> fact that you know the internals doesn't make it okay to reach behind
> I> the counter and grab a value like that.  It may not even be safe to do
> I> so on any given architecture; it certainly isn't safe on arm, and that
> I> line of code above will work only by accident because you're throwing
> I> way all but 16 bits.
> 
> I though about providing that API, but since it isn't safe in general,
> I decided to not do that.
> 
> I> But even more importantly, this WILL result in multiple threads using
> I> the same value at the same time...
> I>  
> I>  - Thread A on CPU 1 and thread B on CPU 2 both begin executing here at
> I> the same time, and both get through counter_u64_add().
> I>  - Thread A keeps running and uses CPU 1's new value, call it 27.
> I>  - Thread B gets prempted between counter_u64_add() and zpcpu_get().
> I> When it resumes it's now on CPU 1, so it retrieves value 27 as well.
> 
> This was already discussed in this thread:
> 
> https://lists.freebsd.org/pipermail/svn-src-head/2015-March/069864.html
> 

For this particular use-case you never care what CPU you are executing
on, you only want to obtain a unique number.

per-cpu counters can serve this purpose no problem, just provide an
operation which guarantees to return the new value of the counter it
incremented. Should be easily achieved with e.g. just pinning curthread
to the cpu it executes on for the duration of inc + fetch.

-- 
Mateusz Guzik <mjguzik gmail.com>