The name idle_send_timeout does not seem to describe well where the timeout
applies (individual write()/writev() calls for arbitraty amounts of data).

The fact that the explanation is complicated might be an indication that a
different timeout implementation be more appropriate.

HTTP2 should probably implement send_timeout somewhen.

Spotted by Coverity:

    /lib/libvgz/deflate.c: 605 in deflateParams()
    599     }
    600     if (s->level != level) {
    601         if (s->level == 0 && s->matches != 0) {
    602             if (s->matches == 1)
    603                 slide_hash(s);
    604             else
    >>> CID 1401053:  Control flow issues  (NESTING_INDENT_MISMATCH)
    >>> The macro on this line expands into multiple statements, only
    >>> the first of which is nested within the preceding parent while
    >>> the rest are not.
    605                 CLEAR_HASH(s);
    606             s->matches = 0;
    607         }
    608         s->level = level;
    609         s->max_lazy_match   = configuration_table[level].max_lazy;
    610         s->good_match       = configuration_table[level].good_length;

We now process the inclusive globs first, and then the exclusive
ones.  The former are accumulative, while the latter will filter
the output from the first.

Discussed in detail with and final diff from @mbgrydeland.
Tests tweaking by yours truly.

Fixes #2320.

If an object has OA_VARY, we should succeed in getting it.

This might help to narrow down #2319

A variable length attribute (OA_HEADERS and OA_VARY) are allowed to be
overwritten if the new data has the exact same length as the previous
data. To check this, its length is recorded. But the length wasn't
recorded, and the code would treat it as a new attribute.

The reason this hasn't caused issues is that the code looking at
OA_HEADERS and OA_VARY ignores the stored attribute length, and that
these aren't ever overwritten.

Retune SHA2 code for improved performance on CPUs with more ILP and
a preference for memory load instructions over large code footprints
with embedded immediate variables.

On amd64 CPUs from 2007-2008 there is not a significant change, but
amd64 CPUs from 2009-2010 get roughly 10% more throughput with this
code; amd64 CPUs from 2011-2012 get roughly 15% more throughput; and
AMD64 CPUs from 2013-2015 get 20-25% more throughput.  The Raspberry
Pi 2 increases its throughput by 6-8%.

Sponsored by:   Tarsnap Backup Inc.

The generated code needs a larger stack when optimizations are
disabled, i.e. with --enable-debugging-symbols.

(One of the thing you'd think ISO-C had solved long time ago...)

put a pid-file in the workdir.

Improve the collision error message for -P by including the other pid.

Because of Transient storage being a special-case stevedore, the
STV_Foreach logic would only work after the child's startup once
all storage backends are defined. So during the setup, collisions
would only be detected for the Transient storage.

The underlying STV__iter function is now pretty much like a plain
VTAILQ_FOREACH loop with an additional CHECK_OBJ_ORNULL step on all
iterations.

As a result, the transient storage is appended to the list at the
end of the manager's setup and is now reported last in the CLI's
storage.list command.

Error messages related to the -s option are slightly more helpful.

Fixes #2321

Inspired by:

https://git.alpinelinux.org/cgit/aports/plain/main/varnish/musl-mode_t.patch?h=3.5-stable

From upstream:

Avoid left shift of a negative value in flush rank calculation.
The C standard permits an undefined result for a left shift of a
negative value.