Thanks a lot to Andreas Plesner Jacobsen for writing those up.

DocWilcos proposed fix in #956

XXX: the 64k fd limitation should probably be fixed.

dlopen typically only fails here if the child process does not have
access to the build directory and the user runs varnishtest as root
(meaning the child setuids to nobody).  Report the dlerror and give a
hopefully helpful hint to help diagnose the error.

Fixes: #959

dereferencing and destroying a busy objcore.

Free the busyobj in the case that the worker already has one
(potential race condition from the expiry/ban-lurker threads?)

Reminded by:	Tollef
Related to:	#953

Thanks a lot to David for minimised test case showing the bug.

Fixes: #965

to avoid dropping an object on out of storage conditions.

I belive this...

Fixes	#953

Otherwise please reopen.

Ensure that stats are correct if allocations fail.

Don't rely on the VSC counter for checking the limit.

It seems most RPM-based distros has the -p switch to killproc and
status those days, so only blacklist the ones we know about that does
not have it.

Fixes: #969

Make sure we compensate for sizeof(int) and the stuff we have already
read.

Fixes: #848

LLVM C-compiler.

You have no idea how much or how long (20+ years!) I have missed having
an alternative to GCC.

Remember:

  Mono cultures are particularly bad for you, when they are good for you

Kudos to:       The LLVM team.

Patch by:	scoof
Fixes		#961

as per discussions in httpbis WG

different functions.

I have not managed to write a vtc case for this one, but I am pretty
sure this:

Fixes:	#945

a21746d2.

Fixes: #941

of source code.

This is substantially DocWilcos fix, but instead of adding yet another
unique numbering variable, I have collapsed the once we have into
a single one.

Fixes #948

Use it to hold the derived vary string for the request (if any) and
skip busy objects with non-matching Vary during hash lookup.

Idea by:	sky

Not detected by:  FlexeLint 9.00f
Detected by:  FlexeLint 9.00g

to prevent old junk on the workspace from being taken for valid.

cache_hash to cache_center.

do not trusting) that all objects under an objhdr have the same
Vary string.

Build a vary string for the request on the wrk->ws as we do the
comparison and use it as a cache for subsequent comparisons.

This will also be first step of sky's "pre-vary"

white space.

Repack struct object, reducing a couple of field widths along the way.

This shaves 16 bytes of struct object.