I shouldn't rely on auto-complete too much, especially for only two
characters...

Spotted by @fgsch

It also applies for the 4.1 and 4.0 series.

Refs #1890

Fixes #1889.

Raise a ParseError exception if not.

Fixes #1885.

When ESI delivering a gzip'ed object, emit the GZIP header+tail
precisely if no parent ESI has done so.

Fixes:	#1878
Testcase by:	fgs

When varnishtest creates a socket barrier, it will bind a socket and
listen to incoming connections. Once the number of expected connections
is open, connections are closed.

Barrier users only need to connect to the socket, read "nothing" and
block until the connection is closed. It allows virtually any process
to sync with varnishtest. The barrier will provide macros with its
socket information.

This is similar ASSERT_CLI, in order to make socket barriers creation
thread-safe.

All usage of cyclic semaphores has been replaced by non-cylic barriers,
except in the disabled test r01252 that absolutely needs one. It still
passes when run with the proper requirements.

This is useful in HTTP loops where it's impossible to predict and use an
array of barriers. All cycles expect the same number of waiters, unlike
semaphores.

They work like semaphores, except that they need explicit initialization
and can optionally be shared between processes.

This fixes a problem when a backend replies with 304 Not Modified
(after a http conditional request from varnish) and does not supply a
header that was duplicate in the cached object.

Before this patch, varnish would only supply (by copying from the
expired object) the first instance of a duplicate header.

This reverts commit e7779bcb.

Add "last updated" timestamp for backend health in varnishadm

OK'd by phk@.

Fixes #1883.

Spotted by Coverity

Fixes a bug where we wait for an event to fire on the backend socket
prior to passing the request body.

The VBT_Wait call is now handled in vbe_dir_finish.

Test case by thomaslc

Fixes: #1806

This reverts commit eaac258d.

The benefits in terms of readability are uncertain and YMMV, but
compilers seem to find it harder to optimize too:

   cmpl   $0x0,-0x205c(%rbp)
   sete   %al
   movzbl %al,%eax
   or     %rax,%rdx

Before the change it would translate to just:

   or     $0x1,%rax
   mov    %rax,%rdx

Spotted by @fgsch

This patch was created using Coccinelle and the following steps:

   $ cat >take.cocci <<EOF
   @@
   expression t, f;
   constant m;
   @@

   - AN(f);
   - t = *f;
   - *f = NULL;
   - CHECK_OBJ_NOTNULL(t, m);
   + TAKE_OBJ_NOTNULL(t, f, m);
   EOF
   $ spatch --sp-file take.cocci --dir . --in-place

There's a recurring pattern throughout the code base of taking over
a reference (by nulling the original) essentially when resources are
freed. Instead of repeating this anti dangling pointers measure, it
can be wrapped in a documenting utility macro.