Submitted by:   Frederic Lecaille <flecaille@haproxy.com>

any more.

It turns out people are really attached to their varnish.params file.

See `git grep PROXY -- bin/varnishd/mgt/mgt_main.c`.

that should not happen.

would give subsequent esi-includes the wrong vcl.

not need it (acceptor, h2_sess)

store more stuff in it.

This reverts commit 5b6a4e56.

The moment we unlock the oh, another thread could rush req

Partially undoes f35b86bc

Huge thank you to @mbgrydeland for spotting my mistake

and polish the output formatting by reordering

Ref #1799

Ref #2743

see 11 lines down

req's on waiting list do indeed hold an objhead ref (req->hash_objhead),
so the condition for the while loop in hsh_deref_objhead_unlock will
always evaluate to false.

This gets rid of the needless rushing code and replaces it with an
assert.

avoid the LONG_MAX check.

Simplify the !busy cases

Inspired by:	#2856

Use req->hash_always_miss directly

Inspired by:    #2856

Refactor the hit-cases

Inspired by:	#2856

Move successful loop termination out of the loop.

Inspired by:	#2856

Make use of the new hsh_deref_objhead_unlock() in HSH_Lookup() to reduce
the number of times the objhead mutex is taken during cache hits.

Split hsh_deref_objhead() into two parts, with the new
hsh_deref_objhead_unlock() function expecting a locked objhead reference
on input.

Make this function lower case for consistency.

This changes the hash slingers deref function to take a locked objhead on
input, and unlocking it before returning. This saves a lock/unlock
sequence for critbit (which is the critical one) on HSH_DerefObjHead.

Move the clean up call to HSH_DeleteObjHead inside of the hash deref
methods that require it, instead of having the outside (single) caller do
it based off the return value. This just cleans up and makes the logic
more transparent.