the req workspace un-reserved before we can call vcl_lookup{} and
we can't tell if we need the predictive Vary: until vcl_lookup{}
tells us.

The missing handler would end the object spec prematurily, hiding the
rest of the methods.

Fixes: #1332

pointer as the first element.

Fixes: #1287

Spotted by:	Federico Schwindt

where the rest of the compiler can get at it.

Make it clearer that vcl_hash is called on lookup, pass and pipe.

This does absolutely not stand for "suckaddr", and but it tries to
sanitize the sockaddr_* mess enough that we can simply pass around
struct sockaddr_storage without also tracking its length.

The cost is that we do know something about address families (IPv4, IPv6),
which strictly speaking should not be necessary, had the sockaddr
not been messed up by the IPv6 people.

Add a few macros that are useful when writing vmods.

Duplicate commit, already done by daghf in ebb66701

This reverts commit bca2da56.

The default number of processes/threads (ulimit -u) on rhel6 and clones
is 1024, which is too little for varnish.

This patch adds a variable NPROC to sysconfig/initscript suppport in the
rpm package. If no value is set in sysconfig, it will be set to 10000.

This looks like a bit of forgotten debugging, left by mistake.

Thanks to InvertedAcceleration <chris.magee@velocity42.com> for a
patch that pointed me at documentation that needed improving.

Fixes #1322

Fixes #1324

get at uncached req.body at the same time.

which will need this.

We may need to improve MPL or LCK to cope with this.

state.

but sequester it and mark it HTTP1 specific

turn the transaction into a "pass" and complain in VSL.

Previously we faulted the fetch, but that was overly draconian.

bereq.uncacheable is our *intention* to pass

beresp.uncacheable is the *consequence* if we pass.

This way it is possible to fetch as miss first time, and try for
a pass on retry.