Thanks scoof

instead of the number of incomplete transactions.

Session grouping mode could grab an unintended amount of memory
because the transaction limit was counted towards the incomplete
rather that not yet reported transactions. A long lasting session
would only count as 1 towards the limit even if it had thousands of
complete children that should put it above the limit.

VSL_Dispatch would return with (1) when the cursor reports more data
available, causing the processing of ready transaction not to be
run. Would cause 'varnishlog -d' not to report any data in some
situations.

Add some comments about the different return points in VSL_Dispatch.

If this option is set, the cursor will return EOF when reaching the log tail.

evaluation.

Some options were defined in the incorrect file. Move them where they
should be.

Now EXP_Rearm() can take the timers to change.

Close a race where we could have an object be both born and killed
in the same slot on the inbox.

Also don't flush on SIGINT exit

If the option isn't set, batch records are never returned from the cursor.

If the option is set, batch records are reported, and the next call to
the cursor will return the first record after the batch. Calling
application will then need to know how to use VSL_NEXT pointer macros
to extract the records in the batch.

Change VSLQ_Dispatch to always return after processing one record or
one batch of records. It will return 1 if it should be called again
without sleeping.

This fixes the situation where a calling program is lagging behind,
and VSLQ_Dispatch never returns because there is a constant stream of
new log records.

connection.

Highlights:

* Improved buffer handling. A vtx can now have a mix of buffered and
  shm chunks. And as shm refs are used and later buffered, they can be
  reused on the same vtx. This benefits long running transactions.

* Synth records are now returned by the cursor at the position the
  cursor was at when it was created. This solves a problem when
  writing to a file and a vtx times out, the reading of that file
  would also cause a time out delay.

* Begin records are now strictly honored, and any records before a
  Begin record are now ignored.

* More asserts on the structures

* More comments, should be easier to understand now

* Various other changes

Log a little bit more debugging information for now.

In general new/changed/killed objects are posted in the exp_thread
mailbox, and suitably handled by the exp_thread at its convenience.

The binheap is only manipulated by exp_thread, and it doesn't
need to hold any locks to do so.

The exp_mtx only protects the inbox to the exp_thread.

oc->flags and the lru-list are protected by the lru->mtx.

EXP_NukeLRU() is temporarily disabled.

When LRU-Nuking to make space, we mark the object as dying, so HSH
will ignore it and steal the storage from it right away.  We post the
rest of the object to exp_thread to clean up later.

The LRU change makes the 3-4 LRU test-cases unstable, (c00044, c00045,
r01140...) because they were used to the full object being reaped,
rather than only the storage segments.  I'm to tired to fix/tune those
cases tonight.

But apart from that, it seems to run the test-cases.

This should make object creation faster.

As a sideeffect, we loose the $expiry_sleep parameter:  Now we always
sleep optimally.

multiple instances at a later date.

Put the "LRU_DONTMOVE" flag on the lru, rather than the objcore.

32 and 64 bit architectures.

The good news is that this is necessary because our obj overhead
decreases.

to check:  The req might get rescheduled before we get there.