is still deranged. (See: CERT secure coding POS04-C)

Also prepare a global condattr_monotime, which will soon be used.

Increment client_req_400 in more cases.

Spotted by: Dridi

Info 776 (Possible truncation of addition): I found no better way than
using an additional size_t argument. This triggered a subsequent
warning about an out-of-bounds access, which could be silenced with
additional assertions on nac to be large anough such that av[ac] was
not out of bounds.

Warning 429: Custodial pointer 'av' (line 60) has not been freed or
returned:

In other cases, the argument vector comes from VAV_Parse() which
flexelint apparently can't follow along, yet memory for a VAV is also
allocated and never freed (see also bottom of STV_Config()).

In this case I found no better way than to silence flexelint.

Declare anonymous vsmw_cluster and vsc_seg structs in VSC header
files. This is to silence errors when the header file is used in other
contexts that don't know about these structs.

This is to resolve the chicken and egg problem of the VSC headers being
generated only once bin/varnishd is built, while they are used in several
VMODs and libraries.

This moves the generation of the VSC .c and .h header files to a directory
in lib that can be built earlier than its dependencies.

Also Don't AC_SUBST the vsc files. The complete list of vsc files is known
and listed in the .am file when automake is run, so there is no need to
have configure.ac help with making lists.

Closes #3708

Conceptually, when HSH_Lookup() returns with HSH_BUSY, ownership of
the request has been ceded: Right after Lck_Unlock(&oh->mtx) in
HSH_Lookup(), it could have got picked up from the waitinglist by
another thread and, in the extreme, could already be done with when
HSH_Lookup() returns.

Ref #3707

Ref #3707

Ref #3707

for the objgetspace callback

This makes b17.vtc fail with fc1ce6da reverted.

To avoid code duplication, we only provide an init callback, which
copies all other callbacks from the malloc storage.

The size argument of ObjGetSpace() is only a hint and storage engines
are free to return whatever size.

If the transient storage returned less space than requested, the
session was closed (no response body sent) with SC_OVERLOAD and
"Error: Could not get storage" logged.

Fixes #3706

Closes #3692

Use of SML is supported by the storage API though the sml_* function
pointers.

Yet, like struct object, struct storage is part of the simple storage
(sml) implementation and does not belong to the storage API. Storage
not using SML (that is, code which does not include storage_simple.h)
should not be concerned with struct storage.

Even thought the first rxresp implies that the backend fetch must have
completed, it is still possible to observe a streaming hit.

Since resp.is_streaming merely looks at the presence of a boc, and not
whether its state is lower than FINISHED, waiting until we see the fetch
end in the logs and forcing a new client session should reduce the race
window in those test cases.

We could also try to settle on the value of resp.is_streaming when we
hit an object in the cache or resume from a backend fetch and change the
semantic to whether resp was streaming when VCL entered the subroutine.

By that I mean, with a proper check of the boc state.

Refs 02ada45c

Spotted by oss-fuzz.

This to silence errors on OSX where apparently int64_t isn't type
equivalent to intmax_t, causing printf-errors when using %jd.

Fixes: #3699

My bad.

fixes #3697

Going back to pretending 32bit bionic builds are fine until #3702 is
sorted out. We have 32bit coverage in our vtest CI anyway.

Reverts f8974400
Reverts c5ed00b6

Fixes: #3688

While at it, try to better formalize all the setup steps.

It operates on a pointer like FREE_OBJ() and clears it.