Fixes previous commit. Sigh.

Broken in 9bb8f962.
Committed solution proposed by phk@, discussed with phk@ and martin@.

Fixes #1845.

into the macro-table.

Instead of counting the references, the VCL keeps track of them with up
to 31 characters in the description. It is the VMOD's responsibility to
keep track of the opaque struct vclref * and provide a meaningful user-
friendly description.

By the time we decide to switch to a VCL, it must be warm and usable.
The deprecated VCL_EVENT_USE should not get in the way.

Don't update the state of the VCL to warm if it failed, and don't start
the child if the active VCL failed to warm up.

The implementation is similar to the load/discard dance when a load
fails. New VGC functions are introduced iff the VCL has at least one
VMOD handling events.

The generated code looks like this:

	static unsigned vgc_inistep;
	static unsigned vgc_warmupstep;

	...

	static int
	VGC_Load(VRT_CTX)
	{
		...
	}

	static int
	VGC_Discard(VRT_CTX)
	{
		...
	}

	static int
	VGC_Warmup(VRT_CTX, enum vcl_event_e ev)
	{

		vgc_warmupstep = 0;

		/* 4 */
		if (Vmod_debug_Func._event(ctx, &vmod_priv_debug, ev))
			return (1);
		vgc_warmupstep = 4;

		return (0);
	}

	static int
	VGC_Use(VRT_CTX, enum vcl_event_e ev)
	{

		/* 4 */
		if (Vmod_debug_Func._event(ctx, &vmod_priv_debug, ev))
			return (1);

		return (0);
	}

	static int
	VGC_Cooldown(VRT_CTX, enum vcl_event_e ev)
	{
		int retval = 0;

		/* 4 */
		if (vgc_warmupstep >= 4 &&
		    Vmod_debug_Func._event(ctx, &vmod_priv_debug, ev) != 0)
			retval = 1;

		return (retval);
	}

	static int
	VGC_Event(VRT_CTX, enum vcl_event_e ev)
	{
		if (ev == VCL_EVENT_LOAD)
			return(VGC_Load(ctx));
		if (ev == VCL_EVENT_WARM)
			return(VGC_Warmup(ctx, ev));
		if (ev == VCL_EVENT_USE)
			return(VGC_Use(ctx, ev));
		if (ev == VCL_EVENT_COLD)
			return(VGC_Cooldown(ctx, ev));
		if (ev == VCL_EVENT_DISCARD)
			return(VGC_Discard(ctx));

		return (1);
	}

However, if there are no VMODs handling events, no new functions shall
be generated, leading to code looking like this:

	static unsigned vgc_inistep;
	static unsigned vgc_warmupstep;

	...

	static int
	VGC_Load(VRT_CTX)
	{
		...
	}

	static int
	VGC_Discard(VRT_CTX)
	{
		...
	}

	static int
	VGC_Event(VRT_CTX, enum vcl_event_e ev)
	{
		if (ev == VCL_EVENT_LOAD)
			return(VGC_Load(ctx));
		if (ev == VCL_EVENT_DISCARD)
			return(VGC_Discard(ctx));

		(void)vgc_warmupstep;
		return (0);
	}

It is also possible to convey a message to the CLI. In case of a failure
the VCL is cooled down. VCL_Load may now fail with either CLIS_PARAM or
CLIS_CANT.

The two functions are separated in order to have self-documenting
signatures because one class of events is (or will be) allowed to
fail while the other MUST be failsafe.

Make a clear distinction between (struct vclprog).warm that is used as a
boolean, unlike the second parameter of mgt_vcl_setstate. We don't use
an actual C enum but const char[] to make debugging easier.

Discussed with mithrandir@.

concurrent -C tests don't collided.

Problem spotted by:	fgs

Fixes:	#1802

In 5c826806 the checks are relaxed,
but this commit (or part of) was not brought into 4.1.

Fixes: #1842

try to delete a objhead while it still has a waiting list,
by forcing the last ref holder to rush the WL.

Since the hasher owns the refcounts on objhead, we cannot just
mingle req and objcore refcounts.

Fortunately we don't need to add another refcounter, because all
we really care about is the wl being empty when we drop the last
ref.

The wl/hsh_rush() mechanism will work differently with different
thread-scheduling schenarios, and I cannot definitively rule out
that we can drop the last ref on an oh, while there are still req's
on the waiting list.

Given that, and the existence proof in ticket #1823's race, this
might have been the indicated memory-trampler.

Conflicts:
	bin/varnishd/cache/cache_hash.c

The latter takes precedence as per RFC 7232.  Also if we have an I-N-M
but no entity don't check I-M-S, just do a non-conditional request.

Conflicts:
	include/vrt.h

This paves the way towards failing VCL_EVENT_WARM events in VMODs. If
setting the temperature fails, we may need to carry a message back to
the CLI and other bits of context.

	import vmod from "/some/dir/"

and have the default vmod shlib filename appended.

know the filename of the -f argument, it should be possible to
include "./" relative to it.

The footnote is that the include is resolved at lower privlevel
than the -f file is read, so in really obscure cases you could
have really obscure error messages.

VCL file, and illegal everywhere else.

This makes it easier to distribute "VCL-packages"

Test-case written by:	Kacper

the VCL fileopening logic in mgt so we have it only one place.

away const.