by the operating system.

wasting partial pages with likely parameter values.

Introduce a new "backend workspace" which is used for the backend
transactions (bereq.*, beresp.*)

Together with the "client workspace" which holds the client transaction
(req.*, resp.*) this should make memory management much simpler to
understand.

The backend workspace lives in "busyobj" rather than the worker thread,
which therefore needs a lot less stack, allowing many more threads
in the same space.

Experiments with thread_pool_stack are encouraged, I wouldn't be
surprised if 32k is now enough.

The per-thread workspace is still around, known as the "auxiallary
workspace", but it is down to 256 bytes and it is only used to hand
the acceptor information off to worker threads.

Once the dust settles, I may remove it entirely, but I can see a
value of having a small amount of fast memory in a thread, so I may
also keep it.  No promises.

Ohh, and I removed a couple of small memory leaks which were only
present in -trunk, as a result of a git-mis-merge.

trust that to be the same throughout the request.

update the desired size of objects from the guard thread.

Rename the gzip_stack_buffer param to gzip_buffer.
Remove the gzip_tmp_space param.

loose the dedicated "pending" buffer and just use the front of the
input buffer.  This simplifies the code quite a bit.

It was a mistake to put the gunzip delivery buffers on the worker
stack, given that only a minority of clients (mostly spiders and robots)
don't grok "Content-Encoding: gzip".

Move them to malloc space.

Change
	return (hit_for_pass)
to:
	set beresp.do_pass = true

It was wrong to make hit_for_pass a return action, because it is not.
It is a flag on the cached object, and that is all it is.

With streaming becoming the default, having it as return-action would
make writing vcl_fetch{} more cumbersome.

CLI to control backend health, rather than rely on probes.

Submitted by:	Federico G. Schwindt

clearer for everybody.

Add restart in vcl_miss{} implementation.

Note that we get into vcl_error{} on 413 and 417 also.

Rearrange the order in which we do things in cnt_start, to be more
consistent and correct: Process Connection: and Expect: headers
before we snapshot and possibly head into vcl_error{}.

We should never come in with a busyobj or obj.

Always allocate obj from Transient storage.

was related to the entry parameters of cnt_error(), but I belive it
has long been OBE, and now gets in the way of streaming.

cruft now not needed.

cache_session.c

two pointers in workers.