Spotted by:	Dridi

previous commit.

Fixes: #1755

connection count.

Instead we mark the dynamic backends unhealthy with the status "deleted",
hide it from CLI displays, prevent new connections from being opened
and clean the backend up when the connection count goes to zero.

Radically different from patch from: Dridi

conversions between enum and text.

Based on patch from:  Dridi

There is potential for improvement in this whole
documentation build business of ours.

The amount of output does (did) not correspond with the importance of
this task. Now at least we'll see the warnings clearly.

I think it makes sense to ship a built HTML copy in
the tarball, as well as the RST sources.

The old cleanup code was a tad eager and made phk miss
his svg. Hopefully this way is a bit more future proof.

automake is our friend.

We might have to update some packaging files later on.
(and perhaps the git push hook.)

Partially from minusf via github

I have not added any other lock-order reversals.

Give the backend tcp-pools their own lock class instead of mixing
them up with the backend lock class.

held over Wait_Enter() processing, and that makes it impossible
for the code using the waiter to avoid one deadlock or another.

Depending on what the call-back code does, this may increase or
decrease contests for the waiters mutex.

This output was very wide and hard to get any sense out of without
actually reading it closely.

The things you are interested in, typically counter name, description
and value, are now on separate lines.

In addition the output should now fit nicely on any VT100 terminals in
80 character mode, I'm sure there are some still in use somewhere.

JSON document layout (key names, hierarchy) is the same as before.

(implicit) CLI lock.

Various white-space fixes

Do not assert if the transient stevedore fails to provide storage for
the synth body.

Log an SLT_Error record when failing on out of storage.

Document that synth objects are created on transient storage.

Fix an objcore leak that would come if we failed to create the object.

Fixes: #1740

We intentionally provoke an early wrap of our segment_n unsigned int.
The previous code is fine according to
http://www.open-std.org/jtc1/sc22/wg14/www/docs/n1570.pdf
"A computation involving unsigned operands can never overflow...",
but nevertheless we don't want to trigger Flexelint warnings.

Thanks to Dag Haavi Finstad for digging the ISO standards url.

Might be relevant to #1753

Strengthen another assert for increased safety.

Fixes: #1749

Safe a couple of needless calls to vslc_vsm_check when we know that
there are no new records and this no need to check shm references /
the next cursor.

vsl sequence and segment updates didn't happen atomically, so
vslc_vsm_check could report spurious overruns.

Replace sequence and segment index with a counter (segment_n), which
always increments (with overflow after UINT_MAX). The actual segment
index will be segment_n % VSL_SEGMENTS. Overrun detection by
calculation of the difference between two segment numbers becomes
simple and safe because we only ever access/update a single integer.

Update the shared memory log head.

(struct VSLC_ptr).priv is now the the equivalent of segment_n from the
reader side. It gets initialized once and is maintained independently.

Patch prepared in collaboration with Martin Blix Grydeland
<martin@varnish-software.com>

Fixes: #1747