on startup, it forks off a master dns handler, which forks off dns
slaves (like the apache model). slaves as spawned as load increases,
and then reused. excess slaves are not ever killed, currently.
implemented topics. each topic has a receive window in each direction
at each edge of the circuit, and sends sendme's at the data level, as
per before. each circuit also has receive windows in each direction at
each hop; an edge sends a circuit-level sendme as soon as enough data
cells have arrived (regardless of whether the data cells were flushed
to the exit conns). removed the 'connected' cell type, since it's now
a topic command within data cells.
at the edge of the circuit, there can be multiple connections associated
with a single circuit. you find them via the linked list conn->next_topic.
currently each new ap connection starts its own circuit, so we ought
to see comparable performance to what we had before. but that's only
because i haven't written the code to reattach to old circuits. please
try to break it as-is, and then i'll make it reuse the same circuit and
we'll try to break that.
svn:r152
Servers are allowed to send 100 cells initially, and can't send more until
they receive a 'sendme' cell from that direction, indicating that they
can send 10 more cells. As it currently stands, the exit node quickly
runs out of window, and sends bursts of 10 whenever a sendme cell gets
to him. This is much much much faster (and more flexible) than the old
"give each circuit 1 kB/s and hope nothing overflows" approach.
Also divided out the connection_watch_events into stop_reading,
start_writing, etc. That way we can control them separately.
svn:r54
Each socket reads at most 'bandwidth' bytes per second sustained, but
can handle bursts of up to 10*bandwidth bytes.
Cells are now sent out at evenly-spaced intervals, with padding sent
out otherwise. Set Linkpadding=0 in the rc file to send cells as soon
as they're available (and to never send padding cells).
Added license/copyrights statements at the top of most files.
router->min and router->max have been merged into a single 'bandwidth'
value. We should make the routerinfo_t reflect this (want to do that,
Mat?)
As the bandwidth increases, and we want to stop sleeping more and more
frequently to send a single cell, cpu usage goes up. At 128kB/s we're
pretty much calling poll with a timeout of 1ms or even 0ms. The current
code takes a timeout of 0-9ms and makes it 10ms. prepare_for_poll()
handles everything that should have happened in the past, so as long as
our buffers don't get too full in that 10ms, we're ok.
Speaking of too full, if you run three servers at 100kB/s with -l debug,
it spends too much time printing debugging messages to be able to keep
up with the cells. The outbuf ultimately fills up and it kills that
connection. If you run with -l err, it works fine up through 500kB/s and
probably beyond. Down the road we'll want to teach it to recognize when
an outbuf is getting full, and back off.
svn:r50
