See comment on how we could auto-generate the documentation from the
include - any less hackish ideas welcome

Fixes #2685

Fixes #2647

The h/2 request body VFP would drop data when the input buffer was too
small to fit the data in the received frame.

With this fix we have the VFP code call us again with a fresh buffer
when we run out.

Fixes: #2679

Ref #2683

HSH_Unbusy() calls BAN_NewObjCore() not holding the objhead
lock, so the ban lurker may race and grab the ban mtx just
after the new oc has been inserted, but the busy flag not
yet cleared.

While it would be correct to call BAN_NewObjCore() with the
objhead mtx held, doing so would increase the pressure on the
combined ban & objhead mtx.

If the ban lurker encounters a busy object, we know that there
must be an unbusy in progress and it would be wiser to rather
back off in favor of the it.

Fixes #2681

The check is added to the builtin logic for now.

Fixes #2631.

Conflicts:
	doc/changes.rst

or should we use "%" PRIu64 in the format ?

reported by Travis via @fgsch

The blind caching is OBE for six years now, see
344a709c

the schema intended for JSON output.

Conflicts:
	bin/varnishd/cache/cache_backend.c
	bin/varnishd/cache/cache_backend_probe.c
	bin/varnishd/cache/cache_director.c
	bin/varnishtest/tests/d00005.vtc
	bin/varnishtest/tests/v00014.vtc

Only help -j is fixed, the backend changes are left out.

to avoid micro-managing memory

Follow up cf20e04e

required after the constification in
8f83087f

I have no idea why the umem authors did not make this argument
a const, they call strncpy on it

A show stopper used to be that the top request's workspace
was reserved for delivery during ESI processing, but since
beeaa19c the path is clear.

spotted by @lkarsten

Ref #2672

After recent failures discovered by travis-ci it should be safe to
have this back on.

rely on h2_errcheck() calls anyway.