it does not need many worker threads and runs for >30s. This helps being
able to run this test with more concurrency to test #3362

... thinks this statement is confusing.

Ref: 15f94d33

VRT_new_backend_clustered() is the only caller of VBP_Insert().

There, VBP_Update_Backend() needs to be called after the director is
created in order to update the director status for a cold VCL.

For a warm VCL, VBP_Update_Backend() is called via
VRT_AddDirector() -> VDI_Event() -> vbe_dir_event() -> VBP_Control()

Thus, the additonal call before the director is added to the backend,
does not do anything but update the poll bits.

Somehow related to #3362

We don't need to assert the integrity of the VSB after every single
byte, it can become very expensive depending on the workload.

Ref 001485e6

Going back to lowercase, and aligning code blocks to 4 spaces to match
the rest of the code snippets.

Ref #3470

Ref #3470

Fixes #3478

When resolve requests race, we were not guaranteed to consider all
backends because we updated a shared nxt variable.

Fixes #3474

... also (C)ompleted bans.

The previous code made a lot of sense in light of the performance issue
due to too many bans, but would not work to prune a long ban list due to
(R)equest bans at the tail, as in this example:

Present bans:
1607346873.538879 354304 C
1607346873.532313     0 C
... lots of (C)ompleted bans
1607083561.980118     0 -  req.http.Host ~ foo
1607083561.972629    15 C

The documentation does not mention the previous behavior (that only
"active" bans are being counted), so this changes aligns code to
documentation.

This introduces a `PAN_dump_struct` function to do the
standard up-front checks for NULL, bad magic & already dumped.

    vtc_haproxy: VTCP_listen_on 127.0.0.1 instead of localhost

    Always bind vtc_haproxy with 127.0.0.1 instead of binding on ::1 or
    127.0.0.1 depending on the environment.

    This is related to bug https://github.com/haproxy/haproxy/issues/937
    here the ipv6 and ipv4 are randomly mixed which can't work with socks4.

And wrap the logic in a macro to reduce duplication.

It's more an action than housekeeping.

Make it possible to compare VSCs, to create a synchronization point
based on internal events (especially with custom VSCs).

For example:

    # block until baz catches up
    varnish v1 -expect FOO.bar == FOO.baz

This is still subject to roughly 5s before timing out.

... unless we f*up, which is the sole purpose of its existence

this needs to live within the exclusive circle of cache_vcl.h includers, so I
went for cache_vrt_vcl.c just because of the name.

This is the refactoring we agreed on to enable an alternative
implementation of #3454. This PR does not yet introduce the copy
callback needed to add the functionality suggested in #3454.

We replace the .free pointer of struct vmod_priv with a pointer to a
methods struct with callbacks. As of this commit, it only contains
the .free callback renamed to .fini. The purpose of the refactoring is
to allow addition of more callbacks later.

The new struct vmod_priv_methods also contains a .type member pointing
to a string to contain an arbitrary description of the type of data any
priv holds which uses these methods.

Implementation:

relevant changes are in cache_vrt_priv.c and vrt.h, other changes are to
the documentation and bundled vmods.

The implementation is a simple refactoring for indirection of the call
to the .fini callback via the methods structure.

This change also exposed #3385

See #3470 for a suggestion how to get out of this

This reverts commit 5cbb3586.

Flexelint complains that this truncation is suspicious, which is fair.
The pdiff() function allows for a more expressive, and hopefully less
suspicious alternative with a bit of hardening in the process.

This way it can be used almost anywhere in the code base and out of tree,
not just in the cache process. The function was slightly polished after
prior discussions with phk and slink.

stdbool.h gets pulled in indirectly, which defines bool to _Bool.
Since we are not using the bool type here, just undefine it.

Better diff with the --word-diff --ignore-all-space options.

Having now the guarantee that VCL sources are always null-terminated we
can use standard string.h functions to detect long strings. We can also
generalize vcc_decstr() to work with all kinds of strings.

Instead of duplicating the code in 3 locations with slight variations,
encode the specificities of each type of string into a struct and have
one location to parse them.

Bonus deduplication, inline C parsing is identical.

This is loosely inspired by vcl_fixed_token() and we could further
reduce the size of the loop in vcc_Lexer() by extracting more logic.

And with that we know for sure that we always have null-terminated
VCL sources.

For a first byte timeout on a new connection, we emit

	FetchError: first byte timeout

For a first byte timeout being reached on a recycled connection, we
should keep the error message similar.

Ref eecd409d