the SMA to run out of space eventually.

If we were panic-ing in multiple threads at the same time, the panic
buffer would be partially overwritten.  Prevent this with a mutex

When we pushed the object allocation into the stevedores for
-spersistent, we did not add LRU eviction to that allocation
path.

Then we added the Transient storage as a fallback for objects
we could not allocate, and they all went there.

Change the way object allocation works as follows:

        If VCL set a stevedore hint and it is valid, we stick
        with it, and LRU that stevedore attempting to make space.

        If no valid hint is given, try all stevedores in turn,
        then LRU one of them to make space.

Fixes	#953

an object.

Fixes	#985

are willing to evict per storage allocation attempt.

Previously this was hardcoded at 50, make it 10 instead.

The ABI we give vmods consist of the Varnish version number and the
git commit ID meaning we can break ABI at will.

Output a warning if we can't determine git commit ID

data, the thread would go into a spin.

Fixes	#942

Found & Fixed by:	Martin

Fixes	#978

Found & Fixed by:	Martin

Fixes	#979

	we transform (gzip/gunzip)
	we stream and backend didn't send c-l.

Fixes	#980

Testcase by:	Martin

non-busy.  Not sure what I thought when I changed it last, but it
was clearly not smart thinking.

Spotted by:	Martin
Test case by:	Martin

Fixes	#963

If the first (policy-chosen) backend fails to get us a connection,
retry a random backend (still according to their weight) until
retries are exhausted.

Kristian sent a proof of concept patch, I just cleaned it up and made
it compile.

Thanks to:	Kristian

Fixes	#977

32 bit systems.

Spotted by:	Kristian

didn't get it rigth the first time.
Thanks to scoof for pointing it out.

Explained the semantics quickly so people understand why not all objects survive a crash