The main point is to clearly recomend using req.grace for the most
common use case - using different grace time when the backend is
healthy.

Right now there are use cases that cannot be solved without req.ttl,
so we undeprecate it.

The req.grace variable can be set in vcl_recv to cap the grace
of objects in the cache, in the same way as in 3.0.x

The "keep" behavior changes with this patch. We now always go
to vcl_miss when the expired object is out of grace, or we go
to the waiting list. The result is that it is no longer
possible to deliver a "keep" object in vcl_hit.

Note that when we get to vcl_miss, we will still have the 304
candidate, but without the detour by vcl_hit.

This commit changes VCL, but only slightly, so we aim to back
port this to earlier versions of Varnish Cache.

Refs: #1799 and #2519

address family is neither of AF_INET or AF_INET6.

This means that UNKNOWN is sent in probes to UDS backends when
.proxy_header=1 is set.

Also verify that a UDS backend receives PROXY LOCAL in a probe when
.proxy_header=2 is set.

Fixes #2702
Closes #2726

Since we verify that header blocks are not interleaved, and we zero the
struct h2h_decode on every new header block, there is no need to malloc
a separate struct h2h_decode per stream.

For (real) backends, we used to assign the probe after adding the
director via VRT_AddDirector(). For an already warm vcl, this lead to
a VCL_WARM event missed because the event was issued by
VRT_AddDirector() when the probe did not exist yet.

We fix this by preparing all of the backend before adding the
director, undoing the work for a failure doing do.

In addition, if adding a backend while the vcl is cold, we need to set
the initial director state based on the probe configuration after
adding the director.

We avoid the overhead of the vcl temperature lock because no harm is
done: If, after the director was added, the temperature goes cold, we
just run an additional update, and if it goes warm, the update will
already have happened.

Likely related to https://github.com/nigoroll/libvmod-dynamic/issues/35

... to make the next commit more readable

To get out of a catch-22, see second next commit

... not in the caller

Future-proofing to avoid mistakenly introducing another race down the
line.

The current flow control code's use of h2->cond is racy.

h2->cond is already used for handing over a DATA frame to a stream
thread. In the event that we have both streams waiting on this condvar
for window updates and at the same time the rxthread gets signaled for a
DATA frame, we could end up waking up the wrong thread and the rxthread
gets stuck forever.

This commit addresses this by using a separate condvar for window
updates.

An alternative would be to always issue a broadcast on h2->cond instead
of signal, but I found this approach much cleaner.

Probably fixes: #2623

Turn the director destruction inside-out: When the vmod object goes
out of scope, we inform VRT that the director is to be
destroyed. Actual destruction is moved to the respective director
callback.

This allows directors to outlast vmod objects.

For vdir based directors, the director object is the vmod object; for
shard, the two are different instances (we might want to reconsider
changing this later).

Prepare for use of the shard director as a dynamic backend, where the
VCL_BACKEND can outlast the vmod object:

* the director priv now points to the struct sharddir instead of
  the vmod object

* move the default parameters pointer to struct sharddir

* index the PRIV_TASK by the struct sharddir pointer

For ban statistics, we updated VSC_C_main directly, so if we raced
Pool_Sumstat(), that could undo our changes.

This patch fixes statistics by using the per-worker statistics
cache except for the following remaining corner cases:

* bans_persisted_* counters receive absolut updates, which does
  not seem to fit the incremental updates via the per-worker stats.

  I've kept these cases untouched and marked with comments. Worst
  that should happen here are temporary inconsistencies until the
  next absolute update.

* For BAN_Reload(), my understanding is that it should only
  happen during init, so we continue to update VSC_C_main
  directly.

* For bans via the cli, we would need to grab the wstat lock,
  which, at the moment, is private to the worker implementation.

  Until we make a change here, we could miss a ban increment
  from the cli.

* for VCL bans from vcl_init / fini, we do not have access
  to the worker struct at the moment, so for now we also
  accept an inconsistency here.

Fixes #2716 for relevant cases

A bound socket will timeout instead of refusing the connection.
Should fix b00015.vtc under macos.

The previous commit made the assumption that END_STREAM is in the last
of the frames in a header block. This is not necessarily the case.

Previously we've been incorrectly transtitioning to CLOS_REM on
END_HEADERS, which prevents us from seeing if a client incorrectly
transmitted a DATA frame on a closed stream.

This slightly complicates things in that we can now be in state OPEN
with an inactive hpack decoding state, and we need to make sure on
cleanup if that has already been finalized.

This would be simpler if the h/2 spec had split the OPEN state in two
parts, with an extra state transition on END_HEADERS.

Again big thanks to @xcir for his help in diagnosing this.

Fixes: #2623

This approach opens up to a data race.

This reverts commit c8bfae70.

Should help with the ASAN builds in travis.

If a client mistakenly sent a DATA frame on a stream where it already
transmitted an END_STREAM, it would lead to the rxthread sitting around
indefinitely.

Big thanks to @xcir for his help in diagnosing this.

Fixes: #2623

Should address #2666 and #2711.

Tune down h2_rx_window_low_water to make sure we don't get a
window_update racing against the response frames.

Fixes: #2709

If we failed to schedule a thread for a stream and it's not cleaned up
prior to handling a request body, the rxthread would sit around waiting
in h2_rx_data indefinitely.

the backend might be going away, and since we cannot afford to hold
the lock over VTP_Open(), we have to pull the VBE_vsc knowledge
one level back up.

Overlooked by:	slink, phk
Spotted by:	Coverity

termination of the rxbuf.

Found by:	fgs

This adds -j support to the command

varnishadm ping -j

Tested on RFC4627/RFC7159/ECMA404 standards (all valid JSON)

Found by:	fgs & ASAN

Previous fix for #2285 (and the duplicate #2624) was missdiagnosed. The
problem stems from a wrong assumption that the number of bytes already
pipelined will be less than maxbytes, with maxbytes beeing the maximum
number of bytes the HTC_RxStuff may need to get a full work unit. That
assumption may fail during the H/1 to H/2 upgrade path where maxbytes
change with the context, or during runtime changing of parameters.

This patch makes HTC_RxStuff not assert if the pipelined data turned out
to exceed maxbytes, but return overflow if we run out of workspace.

(#2624 has received a workaround in the H/2 code that perhaps should be
reverted).