method avoided all race conditions, but suffered from
sometimes running out of buffer space if enough clients
were piled up at the same time.
Now, the client pushes the link descriptor, one end of a
socketpair() and the ppp version via sendmsg() at the
server. The server replies with a pid. The client then
transfers any link lock with uu_lock_txfr() and writev()s
the actual link contents. The socketpair is now the only
place we need to have large socket buffers and the bind()ed
socket can keep the default 4k buffer while still handling
around 90 racing clients.
Previously, ppp attempted to bind() to a local domain tcp socket
based on the peer authname & enddisc. If it succeeded, it listen()ed
and became MP server. If it failed, it connect()ed and became MP
client. The server then select()ed on the descriptor, accept()ed
it and wrote its pid to it then read the link data & link file descriptor,
and finally sent an ack (``!''). The client would read() the server
pid, transfer the link lock to that pid, send the link data & descriptor
and read the ack. It would then close the descriptor and clean up.
There was a race between the bind() and listen() where someone could
attempt to connect() and fail.
This change removes the race. Now ppp makes the RCVBUF big enough on a
socket descriptor and attempts to bind() to a local domain *udp* socket
(same name as before). If it succeeds, it becomes MP server. If it
fails, it sets the SNDBUF and connect()s, becoming MP client. The server
select()s on the descriptor and recvmsg()s the message, insisting on at
least two descriptors (plus the link data). It uses the second descriptor
to write() its pid then read()s an ack (``!''). The client creates a
socketpair() and sendmsg()s the link data, link descriptor and one of
the socketpair descriptors. It then read()s the server pid from the
other socketpair descriptor, transfers any locks and write()s an ack.
Now, there can be no race, and a connect() failure indicates a stale
socket file.
This also fixes MP ppp over ethernet, where the struct msghdr was being
misconstructed when transferring the control socket descriptor.
Also, if we fail to send the link, don't hang around in a ``session
owner'' state, just do the setsid() and fork() if it's required to
disown a tty.
UDP idea suggested by: Chris Bennet from Mindspring at FreeBSDCon
(LCP/CCP/IPCP), one for urgent IP traffic and one for
everything else.
o Add the ``set urgent'' command for adjusting the list of
urgent port numbers. The default urgent ports are 21, 22,
23, 513, 514, 543 and 544 (Ports 80 and 81 have been
removed from the default priority list).
o Increase the buffered packet threshold from 20 to 30.
o Report the number of packets in the IP output queue and the
list of urgent ports under ``show ipcp''.
that ppp stays in the foreground.
o Add the -quiet switch to quieten ppps startup
o Add the -nat flag and discourage the use of the -alias flag. Both do
the same thing.
o Correct some nat usage strings.
o Change the internal ``alias'' command to ``nat''.
o If we're using RADIUS and the RADIUS mtu is less than our
peers mru/mrru, reduce our mtu to this value for NetBSD too.
o Make struct throughput's sample period dynamic and tweak the ppp
version number to reflect the extra stuff being passed through
the local domain socket as a result (MP mode).
o Measure the current throughput based on the number of samples actually
taken rather than on the full sample period.
o Keep the throughput statisics persistent while being passed to
another ppp invocation through the local domain socket.
o When showing throughput statistics after the timer has stopped, use
the stopped time for overall calculations, not the current time.
Also show the stopped time and how long the current throughput has
been sampled for.
o Use time() consistently in throughput.c
o Tighten up the ``show bundle'' output.
o Introduce the ``set bandwidth'' command.
o Rewrite the ``set autoload'' command. It now takes three arguments
and works based on a rolling bundle throughput average compared against
the theoretical bundle bandwidth over a given period (read: it's now
functional).
having different speed links in a bundle. This would manifest itself
by having the link occasionally hang, but revive when a new connection
is made....
Make ``show mp'' a bit prettier.
o Show more information about missing MP fragments in ``show mp''.
o Do away with mbuf_Log(). It was showing mbuf stats twice on
receipt of LCP/CCP/IPCP packets.... ???!!?
o Pre-allocate a bit extra when creating LQR packets to avoid having
to allocate another mbuf in mbuf_Prepend().
being the same as the previous (still supported) ``host:port''
syntax for tcp socket devices.
A udp device uses synchronous ppp rather than async, and avoids
the double-retransmit overhead that comes with ppp over tcp (it's
usually a bad idea to transport IP over a reliable transport that
itself is using an unreliable transport). PPP over UDP provides
througput of ** 1.5Mb per second ** with all compression disabled,
maxing out a PPro/200 when running ppp twice, back-to-back.
This proves that PPPoE is plausable in userland....
This change adds a few more handler functions to struct device and
allows derivations of struct device (which may contain their own
data etc) to pass themselves through the unix domain socket for MP.
** At last **, struct physical has lost all the tty crud !
iov2physical() is now smart enough to restore the correct stack of
layers so that MP servers will work again.
The version number has bumped as our MP link transfer contents have
changed (they now may contain a `struct device').
Don't extract the protocol twice in MP mode (resulting in protocol
rejects for every MP packet). This was broken with my original
layering changes.
Add ``Physical'' and ``Sync'' log levels for logging the relevent
raw packets and add protocol-tracking LogDEBUG stuff in various
LayerPush & LayerPull functions.
Assign our physical device name for incoming tcp connections by
calling getpeername().
Assign our physical device name for incoming udp connections from
the address retrieved by the first recvfrom().
the layering.
We now ``stack'' layers as soon as we open the device (when we figure
out what we're dealing with). A static set of `dispatch' routines are
also declared for dealing with incoming packets after they've been
`pulled' up through the stacked layers.
Physical devices are now assigned handlers based on the device type
when they're opened. For the moment there are three device types;
ttys, execs and tcps.
o Increment version number to 2.2
o Make an entry in [uw]tmp for non-tty -direct invocations (after
pap/chap authentication).
o Make throughput counters quad_t's
o Account for the absolute number of mbuf malloc()s and free()s in
``show mem''.
o ``show modem'' becomes ``show physical''.
details. Compiling with -DNORADIUS (the default for `release')
removes support.
TODO: The functionality in libradius::rad_send_request() needs
to be supplied as a set of routines so that ppp doesn't
have to wait indefinitely for the radius server(s). Instead,
we need to get a descriptor back, select() on the descriptor,
and ask libradius to service it when necessary.
For now, ppp blocks SIGALRM while in rad_send_request(), so
it misses PAP/CHAP retries & timeouts if they occur.
Only PAP is functional. When CHAP is attempted, libradius
complains that no User-Password has been specified... rfc2138
says that it *mustn't* be used for CHAP :-(
Sponsored by: Internet Business Solutions Ltd., Switzerland
shortseq, authname and authkey.
o Auth{name,key} may additionally be set in PHASE_ESTABLISH.
o The others may be set in PHASE_ESTABLISH as long as no links
have yet reached DATALINK_LCP.
anything for two mintues (see ``set choked'' and ``show
bundle''), nuke the ip, mp and link level buffer queues.
This should fix problems where ``ppp -auto'' seems to stop
responding after failing to connect to the peer a few times.
(see the new ``set callback'' and ``set cbcp'' commands)
o Add a ``cbcp'' log level and mbuf type.
o Don't dump core when \T is given in ``set login'' or
``set hangup''.
o Allow ``*'' and blanks as placeholders in ppp.secret and
allow a fifth field for specifying auth/cbcp dialback
parameters.
o Remove a few extraneous #includes
o Define the default number of REQs (restart counter) in defs.h
rather than hardcoding ``5'' all over the place.
o Fix a few man page inconsistencies.
o If we've denied and disabled all compression protocols, stay
in ST_INITIAL and do an LCP protocol reject if we receive any
CCP packets.
o If we've disabled all compression protocols, go to ST_STOPPED
and wait for the other side to ask for something.
o If we've got anything enabled, start REQing as soon as the auth
layer is up.
o If we're in multilink mode, than the link level CCP goes
straight to ST_STOPPED irrespective of what's configured so that
we never try to compress compressed stuff by default.
end up writing zero bytes, sleep for 1/10 of a second so that
we don't end up using up too much cpu.
This should only ever happen on systems that wrongly report a
descriptor as writable despite the tty buffer being full.
Discussed with: Jeff Evarts
o Do an initial run-time check to see if select() alters the passed
timeval. This knowledge isn't yet used, but will be soon.
o Bring the static ``ttystate'' into struct prompt so that
the tilde context is per prompt and not global.
o Comment the remaining static variables so that it's
clear why they're static.
o Add some XXX comments suggesting that our interface list
and our hostname should be re-generated after a signal
(say SIGUSR1) so that a machine with PCCARDs has a chance.
transferring session rights with them. Instead, create two
`/bin/cat' processes. A new child is spawned and disassociated from
the terminal and the parent, which continues with the rest of the ppp
process. Meanwhile, the parent spawns another child, and both the
parent and child exec the `/bin/cat' commands with the appropriate
descriptors. This way, the session is owned by the parent, and the
tty is held open.
o Close LCPs that have done a TLF and are now in ST_STOPPED before
calling Down. This prevents them from trying to come back up again
after the peer has shut them down (it seems a bit strange that the
rfc says that a Down in ST_STOPPED will cause a TLS etc).
o Don't try to set the physical link name pointer when we're receiving
and renaming a datalink. The physical hasn't been created yet, and as
it happens, the garbage physical pointer happens to be the value of another
physical - so we're pointing that other physical name at ourselves.
yeuck.
o Re-arrange the order of things in main (DoLoop()). We now handle
signals only after the select and not before the UpdateSet. It's
possible that either a signal (FSM timeout) or a descriptor_Read()
brings a link down, after which we'd better tidy up any dead direct
and 1off descriptors before calling UpdateSet() again.
o Mention when we detect a PPP packet when we see one before the link
is up (then start LCP as before).
It's now dealt with by the `server' object. This simplifies
things as we only have one list of prompt descriptors and
the log_ routines check prompt::logactive to determine
whether it should be used for output.
o Include the MP socket UpdateSet() result in bundle::UpdateSet().
o Don't select on the tun device unless we're in NETWORK
phase or AUTO mode.
o Stop the idle timer when we go to DEAD phase. We may
have transferred a link and not had a chance to kill
it.
o Don't fail when trying to unlink our transferred datalink
from our descriptor lists just before the transfer.
o Add our link descriptor to the write set if we got a short
write the last time (physical::out is set).
o Log the connection source address when a connection is closed.
o Remove descriptor::next field. Descriptor lists are not required
any more.
incoming fragments when a link goes down.
o Don't use the minimum sequence numbers of links that aren't open.
o Understand sequence number wrapping when determining the minimum
sequence number.
o Add & adjust a few comments.
the bundle has the opportunity to go PHASE_DEAD and cleanup
the interface (if it's the last link).
o Regnerate our phys_type value when we transfer the link.
o Always clean up the interface when destroying our bundle in case
we're abending.
o Always clean up our interface when the last link is gone rather than
delaying things 'till exit time in the -direct case (the interface
is useless anyway). Do this *after* slamming down our NCPs (if
they're still around).
o Our MP server descriptor now clears the relevent device descriptor
from our descriptor [fd]sets when a datalink is on death-row (to
be transferred to another running ppp), thus avoiding the possibility
of passing a bum descriptor to select() and having ppp abend.
o Handle the MP socket descriptor functions from within the bundle
descriptor functions. Now we ensure that the MP socket descriptor
functions see the descriptor sets *after* they've been seen by our
datalinks.
o Add/fix a few more comments.
into the ST_STOPPED state.
o Allow an optional ccp|lcp argument to `down'. The default is
still lcp (as before). You can now call down with no context
in multilink mode, in which case it'll down the multilink ccp
or the entire bundle (*very* rude).
o Allow an optional `!' after `close ccp' (close ccp!) to tell
ccp to stay in the CLOSED state after the terminate ACK. The
default is now to re-enter STOPPED so that the peer can bring
the layer back up if desired.
o Always handle proto-compressed packets, even if we've agreed
(in LCP) that the peer will not send us 1 byte protocols.
If the peer violates the LCP agreement, log it to the HDLC
log.
o Fix some comments.
o Log FD_SET()s in LogTIMER.
o Identify the descriptor that causes an EBADF from select()
if LogTIMER is enabled (then exit).
o Call the MP server UpdateSet() function after calling
the UpdateSet() for all links - the link may enter
PHASE_TERMINATE and bring down the MP server - breaking
the imminent select().
LQM and HDLC timer diagnostics come out with the correct name.
o Don't send an LQR immediately upon reviving a datalink. Leave
it 'till the next timeout.
o Add the link name to some more LQR diagnostics.
o Break out of the main loop when a descriptor exception is seen
in select().
o Remove the evil nointr_[u]sleep() functions. Timers should be
(and are) used instead.
o Treat a read() of 0 bytes as an error that's fatal to the link
on which the read() is done. We should never read() 0 after
select() says there's something there - not unless the link
has been closed by the other side.
o Write the data seen before a HDLC header to the terminal in
`term' mode, *not* back to the modem :-/
o Initialise our transmitted file descriptor before starting any
timers.
o Only send data links that have *no* pending output data. This
means that our final ACK will be written rather than being
nuked with the datalink transmission.
according to SHORTSEQ negotiation.
o Don't forget to attach incoming fragments with a
number greater than everything else in the queue
(rather than leaking memory).
o Output the link name with the ``other'' hdlc
diagnostic message.
o Correct a VJ diagnostic (`COMPPROTO', not `proto').
throughput measurement is enabled).
o Load balance the links based on weight *and* on a round-robin
basis. This makes things fairly even on an output basis. We
don't try to allow for the peer sending all his data down one
link (and try to send ours back up the other).
o Show the number of pending input buffers that can't be processed
in ``show mp''.
o Fix a typo in the man page.
using the scatter/gather array to transfer the link
information. The whole link is now passed in one message.
This is far better than the two `/bin/cat' processes per additional
link :-) I remember years ago thinking that file descriptor
transferral would be a really amazing facility !
Suggested by: Garrett Wollman <wollman@khavrinen.lcs.mit.edu>
and: Eivind Eklund <eivind@yes.no>
log debug'' without filling our filesystem/screen with
junk that we don't really want to see.
o change PHYS_STDIN to PHYS_DIRECT - we can handle incoming
connections that aren't on STDIN_FILENO now.
o Allow return values from our FSM LayerUp functions. If
LayerUp() fails, the FSM does an immediate FsmDown() without
calling the fsm_parent's Layer{Up,Down} functions.
o Clear the close-on-exec flag of file descriptor 3 when executing
chat programs so that our documented ability to communicate with
/dev/tty via that descriptor works. Also document it as
descriptor 3, not 4 :-O
o Allow a ``rm'' command as an alias for ``remove''.
o Fix the bind()/connect()/accept() calls made by the MP server.
o Create bundle_SendDatalink() and bundle_ReceiveDatalink().
This allows `struct datalink's to flatten themselves, pass
through a pipe (read: the eye of a needle !) and come alive
at the other end. The donator then fork()s & exec()s pppmpipe,
``passing'' the connection to another ppp instance.
*** PPP NOW TALKS MULTILINK :-))) ***
Our link utilization is hideous, and lots of code needs
tidying still. It's also probably riddled with bugs !
It's been tested against itself only, and has hung once,
so confidence isn't high....
o Create struct mpserver as part of struct mp.
mpserver creates a unix-domain socket based on the
peers auth name and endpoint discriminator. If it
already exists, ppp will ``pass the link'' over to
the owner of the socket, joining it into the bundle
of another ppp invocation, otherwise ppp waits for
other invocations to pass it links through this
socket.
The final piece of code will be the code that flattens
our datalink info and passes it down this channel
(not yet implemented).
