The qualifiers were accidentally lost in 17a456bd, because barriers
became the template and were the only non-const occurrence.

A command like `barrier r1` would fail while holding the lock because of
the invalid name, but the reset of the test case would then deadlock. It
leads to a timeout like we see sometimes on VTEST so getting rid of them
may just require to review when `vtc_fatal` is called while locking.

In this very case we can only create barriers in the top thread so the
lock is useless.

It takes the barrier as a template and provides a consistent error
message logged with the caller's vtclog. The name is the first thing
checked when a new named entity is created, saving useless setup.

Fixes #2249

This is just what I could find with "Fixes" or "Closes" in
git log varnish-5.0.0..HEAD

Fix a stupid misstake where the segment counter in the VSL segment was
set just before the entire struct was cleared during
initialization. This didn't cause problems in the common case because
it would be set correctly on the very first log line produced. But it
did open a race where an unfortunate early log client would assume it
was lagging indefinetely behind varnishd on the log reading and report
overflow error. This was most commonly experienced with varnishtest.

Fixes: #2241

have to test with 16k frame sizes.

in What's New for 5.1.

This was the rollback of the fix for #1206 brought about in #2042.

Remains of copy/pasta made that v2 listened with PROXY protocol which is
not needed. It's a passive instance that only does health checking using
a probe.

not as smart as LLVM about const.

I guess if you can't say anything else good about it, H2 is forcing
me to invent new ways to implement table-driven object-oriented programming.

which per-frame flags only get sent on the last frame of a
continued frame-sequence.