reordered in transit instead of dropping them altogether.
It uses sequence numbers of PPtPGRE packets.
A set of new sysctl(8) added to control this ability or disable it:
net.graph.pptpgre.reorder_max (1) defines maximum length of node's
private reorder queue used to keep data waiting for late packets.
Zero value disables reordering. Default value 1 allows the node to restore
the order for two packets swapped in transit. Greater values allow the node
to deliver packets being late after more packets in sequence
at cost of increased kernel memory usage.
net.graph.pptpgre.reorder_timeout (1) defines time value in miliseconds
used to wait for late packets. It may be useful to increase this
if reordering spot is distant.
MFC after: 1 month
of pptpgre and ksocket nodes for all calls between two peers. This patch
modifies node's API by adding new "session_%04x" hook names support, while
keeping backward compatibility.
Together with appropriate user-level support (by latest mpd5) it gives
huge performance benefits for case of multiple active calls between
two peers because of avoiding data duplication and extra socket processing.
On my benchmarks I have got more then 10 times speedup for the 200
simultaneous PPTP calls between two peers.
In conclusion, it allows now to build effective "clients <=> PAC <=> PNS"
setups.
Formulas described in RFC require high precision of floating point.
Formulas of integer math implemented in ng_pptpgre give mistake in range
of +0-7ms on RTT and +0-3ms on deviation. This leads to significant
underestimation of real packet RTT.
I have made a very simple patch to reduce mistake to +4-3ms on RTT and
+2-1ms on deviation. Mistake in RTT is not good, but gets covered by
deviation. To cover worst possible negative mistake in deviation I have
added 2ms to it. Also this 2 ms cover the case when measured deviation
is so small (about zero) that it can interfere with process scheduling
delays or weather on Mars.
My tests show decreasing of packet losses on 20ms RTT link from 2.5% to
0.3% while speed increased un 1/3.
Reviewed by: archie
First, mutexed callouts are incompatible with netgraph nodes, because
netgraph(4) can guarantee that the function will be called with mutex
held.
Second, nodes should not send data to their neighbor holding their
mutex. A node does not know what stack can it enter sending data in
some direction. May be executing will encounter a place to sleep.
New locking:
- ng_pptpgre_recv() and ng_pptpgre_xmit() must be entered with mutex held.
- ng_pptpgre_recv() and ng_pptpgre_xmit() unlock mutex before
sending data and then return unlocked.
- callout routines acquire mutex themselves.
parallel ng_pptp_rcvdata():
- Add a per-node mutex.
- Acquire mutex during all ng_pptp_rcvdata() method.
- Make callouts protected by mutex. Now callouts count as
netgraph writers, but there are plans to allow reader callouts
for nodes, that have internal locking.
- Acquire mutex in ng_pptp_reset(), which can be triggered
by a message or node shutdown.
PR: kern/80035
Tested by: Deomid Ryabkov <myself rojer.pp.ru>
Reviewed by: Deomid Ryabkov <myself rojer.pp.ru>
loss links, and 1 second appeared to be too small for high latency links.
If we will receive more complaints, we should make this parameter configurable.
PR: kern/69536
Approved by: archie, julian (mentor)
MFC after: 3 days
In practice it seems that in situations of high packet loss the ACK
timeout seems to hit this maximum (perhaps inappropriately, but the
estimation algorithm is not perfect, so apparently it happens). In
any case, 10 seconds is way too high a value so lower to 1 second.
MFC after: 3 days
the mbuf allocator flags {M_TRYWAIT, M_DONTWAIT}.
o Fix a bpf_compat issue where malloc() was defined to just call
bpf_alloc() and pass the 'canwait' flag(s) along. It's been changed
to call bpf_alloc() but pass the corresponding M_TRYWAIT or M_DONTWAIT
flag (and only one of those two).
Submitted by: Hiten Pandya <hiten@unixdaemons.com> (hiten->commit_count++)
exhausting the kernel timeout table. Perform the usual gymnastics to
avoid race conditions between node shutdown and timeouts occurring.
Also fix a bug in handling ack delays < PPTP_MIN_ACK_DELAY. Before,
we were ack'ing immediately. Instead, just impose a minimum ack delay
time, like the name of the macro implies.
MFC after: 1 week
Change a prototype.
Add a function version of ng_ref_node() when debugging so
a breakpoint can be set on it.
ng_base.c:
add 'node' as an argument to ng_apply_item so that it is up
to the caller to take over and release the item's reference on
the node. If the release reports back that the node went away
due to the reference going to 0, the caller should cease referencing
the now defunct node. (e.g. the item was a 'kill node' message).
Alter ng_unref_node to report back the residual references as a result.
ng_pptpgre.c:
Don't reference a node after we dropped a reference to it.
(What if it was the last?)
Fixes a node leak reported by Harti Brandt <brandt@fokus.gmd.de>
which was due to an incorrect earlier attempt to fix the
"accessing node after dropping the last reference" problem.
This version is functional and is aproaching solid..
notice I said APROACHING. There are many node types I cannot test
I have tested: echo hole ppp socket vjc iface tee bpf async tty
The rest compile and "Look" right. More changes to follow.
DEBUGGING is enabled in this code to help if people have problems.
format version number. (userland programs should not need to be
recompiled when the netgraph kernel internal ABI is changed.
Also fix modules that don;t handle the fact that a caller may not supply
a return message pointer. (benign at the moment because the calling code
checks, but that will change)
This clears out my outstanding netgraph changes.
There is a netgraph change of design in the offing and this is to some
extent a superset of soem of the new functionality and some of the old
functionality that may be removed.
This code works as before, but allows some new features that I want to
work with and evaluate. It is the basis for a version of netgraph
with integral locking for SMP use.
This is running on my test machine with no new problems :-)
- It's worthwhile to use untimeout(9), even though we must still protect
against "false" timeouts, because most of the time it saves having to
handle a dummy timeout event.
- Slight tweaks to the delayed ACK algorithm paramters.
- Fix slowness when operating over fast connections, where the timeout(9)
granularity is on the same order of magnitude as the round trip time.
timeout(9) can happen up to 1 tick early, which was causing receive
ack timeouts to happen too early, causing bogus "lost" packets.
- Increase the local time counter to 64 bits to avoid roll-over.
- Keep statistics on memory allocation failures.
- Add a new option to always include the ack when sending data packets.
Might be useful in high packet loss situations. Might not.
instead of bumping the recvAck counter by one, pretend that
all outstanding xmit packets are acknowleged, and restart
transmitting anew, with an empty (but halved) transmit window.
Put a lower bound on the adaptive timeout value.
world seems to interpret the spec this way
- Initialize transmit window to two instead of one; helps get things
going initially when the first packet may get dropped
- Really fix the shutdown + timeout race condition this time
- Properly handle 32 bit sequence numbers when they wrap around
- Don't drop GRE packets with stale ACK numbers, just ignore the ACK
- Close race between node being shutdown and timer going off
Also add support for lots of statistics, and control message ASCIIfication
