Fixes	#1518

Add a bit more detail while cleaning up the language.

Reformat for tabs instead of 7 spaces.

Using multiple names for the instanciateed directors in this document
just makes it more confusing, since .add_backend() and .backend() is
mostly identical for all of them.

Fixes: #1519

keep them in sync.  If you want a custom reason field, set status
first, then reason.

With this in place, HTTP1_Write() doesn't need to worry about them
being OK, and we can polish that code a fair bit.

Turn a couple of boolean arguments into more readable magics.

to follow shortly.

The rst2man utility exists only in "newer" versions of python-docutils.

Require at least 0.7, released July 2010.

the pthread implementation in *every* test case.

Move policy decisions out where they belong and let the http_Filter
function send the first like through unmolested.

Rather than value specific functions, make a generic
http_ForceField() function, because the name is cool.

"response/reason" confusion for good.

one function which does what you'd expect.

Start allowing internal VCL use of *resp.status > 999.  The idea
is that we ignore the higher bits but it still has to end up >=
100, so 10404 becomes 404 and 20099 is illegal.   (This change
is not yet complete it will panic if you use it yet.)

More http->failed stuff.

Usually, (probably always) this means failure to get workspace for
modifications of the http.

The intent is that this flag will have the same "latching" behaviour
as error handling in VSB's:  Once set, it stays set and nobody
reads or writes the struct http any more.

Setting the flag causes a SLT_Error message.

Rewrite http_CollectHdr() to respect failed and optimize it
slightly while were here.

Fixes: #1514

Other minor cleanups

This adds %I and %O formatters to varnishncsa, which are total bytes
received and total bytes sent. These are similar to the same
formatters in apache (with mod_logio). Useful for bandwidth accounting.

Patch by: David Robertson

Verify that s_resp_bodybytes is indeed incremented the correct number of bytes on partial content responses.

Pointed out by:	Dridi Boukelmoune

Fixes #1510

is a contention point and scales negatively with number of thread-pools.

Sum into a per-pool dstat structure instead, and have the first
idle thread summ that into the global counters without blocking the
pool mutex.

Fixes	#1495

(I hope)

also takes account of fetch jobs.