Fixes #4088

This is not the case for example on Alpine Linux.

But it is not required when building with libunwind support.

This facility allows intermediate layers to set a prefix without
meddling with the provided format string.

This is done in a VSB to match what the panic code is currently doing.

When libunwind is used but fails to produce a back-trace, there is now
a fallback to the original libexecinfo-based implementation (suggested
by @asadsa92).

I don't understand why they didn't maintain the unversioned centos:stream
tag as a kind of latest tag after dropping the actual centos:latest tag.

The plain "stream" image has not been updated for a year, and instead of
being the equivalent of a "latest" tag it just appears to be frozen.

Since there isn't an automatic centos-stream upgrade path, let's target
version 9 explicitly for now.

veto'ed by phk.

It was my intention to make a helpful improvement along the way, but I have
no capacity at the moment for a longer process.

This reverts commit d6aef69a.
This reverts commit 9b76b9a4.
This reverts commit 8fc1b2c3.

Code after vtc_fatal() is unreachable

it repeatedly timed out on arm

I misread the code and did not try running varnishtest as root :|

Clients to the Object API need to know not only the current extension
(new length) of streaming objects, but also the streaming state - in
particular BOS_FINISHED and BOS_FAILED. The latter for obvious
reasons, and the former to call the delivery function with
OBJ_ITER_END, which then likely results in VDP_END sent down the
delivery pipeline.

Background:

It is important for efficient delivery to not receive an additional
VDP_END with a null buffer, but rather combined with the last chunk of
data, so, consequently, it is important to reliably send OBJ_INTER_END
also with the last chunk of data.

Consequent to all of this, ObjWaitExtend() callers need to know when
BOS_FINISHED has been reached for some extension.

The current API, however, does not provide a consistent view of the
streaming state, which is only available from within the critical region
of ObjWaitExtend().

Thus, we add the streaming state as an optional return value.

With this commit, we also remove a superfluous line to set rv again:
Because boc->fetched_so_far must only be updated while holding the boc
mutex, reading the value again provides no benefit.

note: use of isnan() is for forward compatibility, at the moment, we use
-1 to denone NAN for backends.

Processes created by the worker might require a TMPDIR for temporary
files (think: filters). Create a dedicated directory for this purpose,
which is otherwise hard to achieve in a universal manner.

The problem is that a MAP_FAILED segment ends up in its vsm_set::segs
list, at which point NULL checks will determine what to do with the
segment, and MAP_FAILED != NULL.

This can happen when winning (or losing?) the race with a VSC segment
that goes away after being found in the index, but before being mapped.

One surefire result is a crash in vsm_unmapseg() tripping on assertions
that other fields are also not NULL. There may be other unidentified ill
effects.

It should be resolved like the argument of a -n option.

Refs 1d7cdadd

Putting the new fallback next to the historical fallback.

Refs #4105

The host name default referenced by VUT_OPT_n is long gone since
879c9b37

Refer to varnishd(1) to avoid the documentation getting out of sync.

Ref #4105

with greetings from the rabbit-hole opened by #4105

to anchor patches

Maybe it should gain a boc_state_e * argument to return a consistent
state value?

It turns out we still need it in a couple places.