| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
| |
2022 was a rather slow year...
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
| |
This makes it clear that we are scheduling a potential delayed
acknowledgment instead of acknowledging a packet scheduled for
retransmission. Also some small cosmetic fixes.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
This fixes the RTO doubling on timeout according to Karn/Partridge.
Exponentially increasing RTO when it times out (e.g. doubling)
fixes the problem that a sudden increase in real RTT starves the sRTT
updates by never getting out of backoff as retransmitted packets can't
update RTT.
Added an parameter to make it less aggressive, default is doubling.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
| |
Growing pains.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
Delayed ACKs are now sent after twice the internal tick time. Fixes
initial ACK record (rcv_cr.seqno) being uninitialized (0) when the
first ACK was to be sent. Adds some FRCT metrics for number of
received delayed (bare) ACKs and the RTT estimator.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
Unidirectional traffic has one of the peers only send bare FRCT
packets. These never set a DRF, since they have no sequence number.
At the receiver, all these ACKs and window updates were always dropped
as the receiver connection record was timed out.
Also fixes a SEGV if flow control kicks in (passing NULL timeout to
pthread_cond_timedwait).
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This is the first step moving away from scheduling the FRCT and flow
monitoring functions as part of the IPC calls (flow_read / flow_write
/ fevent) and towards the more scalable (and far less complicated)
implementation to take care of these functions in separate threads.
If a process creates the first flow that requires FRCT, it will spin
up a thread to process events on the timerwheel (retransmissions and
delayed ACKs). This single thread lives until the last flow with FRCT
is deallocated.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Reading packets from the rbuff and checking their validity (non-zero
size, pass crc check, pass decryption) is now extracted into a
function.
Also adds a function to get the length of an sdu_du_buff instead of
subtracting the tail and head pointers.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
It doesn't really make sense to manually and one-sidedly configure the
timeout of delayed acknowledgements, as setting it too high upsets the
peer's sRTT estimates. Even worse, it also causes a lot of spurious
retransmissions if it exceeds the sRTT mean deviation calculated by
the receiver. Compensating on bare acknowledgment for the ack delay
could improve the RTT estimate deviation, but not the spurious
retransmissions if it was set too high. This sets the delayed ack to
wait for a single RTT mean deviation. Probably needs more tweaking to
further reduce differences between the RTT estimates at the sender and
receiver, e.g. compensate the RTT estimate for delayed acks, or
increase the RTO to add 8 mdevs to sRTT instead of 4. However, it
looks like the mdev estimate is the trickiest one to get to sync, not
the RTT average. Linux reduces the sample weight for mdev from 1/4 to
1/32 in some cases, will give that a shot some day too to see if that
further align sRTT estimates. In any case, this patch already improves
things a lot.
Also fixes a bug where the sender was sending acknowlegments on the
first packets in flight for the 0 sequence number. The receiver
activity was measured in seconds but compared to a timeout value in
nanoseconds.
There's still a lot of spurious retransmissions that start after
actual packet loss occurs, I'm still investigating what causes it.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
This exposes some additional metrics relating to FRCT / Flow control:
the number of duplicate packets received, number of packets received
out of the flow control window and / or reordering queue, and the
number of rendez-vous messages sent.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
There still were a couple of bugs in the timerwheel. If the future
schedule was coinciding with the slot currently being processed
(i.e. exactly RXMQ_SLOTS in the future), the list_add_tail caused an
infinite loop. Another bug was causing the slots at higher levels to
be processed too soon.
Retransmissions should now schedule correctly.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The timerwheel was retransmitting packets and the error check for
negative values of the rbuff allocation was instead checking for
non-zero values, causing a buffer allocation to succeed but the
program to continue down the unhappy path leaving that packet stuck in
the buffer unattended.
Also fixes wrongly scheduled retransmissions that cause packet storms.
FRCP is much more stable now. Still needs some work for high
bandwidth-delay products (fast-retransmit).
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
| |
The checked condition can't happen.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
| |
It was taking a write lock when a read lock was sufficient.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
This add an ouroboros/pthread.h header that wraps the
pthread_..._unlock() functions for cleanup using
pthread_cleanup_push() as this casting is not safe (and there were
definitely bad casts in the code). The close() function is now also
wrapped for cleanup in ouroboros/sockets.h.
This allows enabling more compiler checks.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
| |
The ugent email addresses are shut down, updated to Ouroboros mail
addresses.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
| |
Happy New Year, Ouroboros!
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
| |
The timerwheel is checked during IPC calls (fevent, flow_read),
causing huge load on CPU consumption in IPCPs, since they have a lot
of fevent() threads for QoS. The timerwheel will need further
optimization), but for now I reduced the default tick time to 5 ms and
added a boolean to check that the wheel is actually used.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
This adds the rendez-vous mechanism to handle the case where the
sending window is closed and window updates get lost. If the sending
window is closed, the sender side will send an RDVS every DELT_RDV
time (100ms), and give up after MAX_RDV time (1 second). Upon
reception of a RDVS packet, a window update is sent immediately. We
can make this much more configurable later on (build options for
defaults, fccntl for runtime tuning).
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
| |
This allows configuring some parameters for FRCP at compile time, such
as default values for Delta-t and configuration of the timerwheel. The
timerwheel will now reschedule when it fails to create a packet,
instead of setting the flow down immediately. Some new things added
are options to store packets for retransmission on the heap, and using
non-blocking calls for retransmission. The defaults do not change the
current behaviour.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This completes the retransmission (automated repeat-request, ARQ)
logic, sending (delayed) ACK messages when needed.
On deallocation, flows will ACK try to retransmit any remaining
unacknowledged messages (unless the FRCTFLINGER flag is turned off;
this is on by default). Applications can safely shut down as soon as
everything is ACK'd (i.e. the current Delta-t run is done). The
activity timeout is now passed to the IPCP for it to sleep before
completing deallocation (and releasing the flow_id). That should be
moved to the IRMd in due time.
The timerwheel is revised to be multi-level to reduce memory
consumption. The resolution bumps by a factor of 1 << RXMQ_BUMP (16)
and each level has RXMQ_SLOTS (1 << 8) slots. The lowest level has a
resolution of (1 << RXMQ_RES) (20) ns, which is roughly a
millisecond. Currently, 3 levels are defined, so the largest delay we
can schedule at each level is:
Level 0: 256ms
Level 1: 4s
Level 2: about a minute.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This adds rudimentary support for sending and processing
acknowledgments and doing retransmission.
It replaces the generic timerwheel with a specific one for
retransmission. This is currently a fixed wheel allowing
retransmissions to be scheduled up to about 32 seconds into the
future. It currently has an 8ms resolution. This could be made
configurable in the future. Failures of the flow (i.e. rtx not
working) are indicated by the rxmwheel_move() function returning a fd.
This is currently not yet handled (maybe just setting the state of the
flow to FLOWDOWN is a better solution).
The shm_rdrbuff tracks the number of users of a du_buff. One user is
the full stack, each retransmission will increment the refs counter
(which effectively acts as a semaphore). The refs counter is
decremented when a packet is acked. The du_buff is only allowed to be
removed if there is only one user left (the "stack").
When a packet is retransmitted, it is copied in the rdrbuff. This is
to ensure integrity of the packet when multiple layers do
retransmission and it is passed down the stack again.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
|
|
| |
The previous commit only fixed the issue for release builds. This
fixes it for debug builds as well.
Signed-off-by: Sander Vrijders <[email protected]>
Signed-off-by: Dimitri Staessens <[email protected]>
|
| |
|
|
|
|
|
|
| |
A comparison was done in the timerwheel between an unsigned value and
a time_t.
Signed-off-by: Sander Vrijders <[email protected]>
Signed-off-by: Dimitri Staessens <[email protected]>
|
| |
|
|
|
|
|
| |
Happy New Year, Ouroboros.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
| |
|
|
|
|
| |
This turns the timerwheel into a passive component since it is used by
application using the library. The user of the timerwheel now has to
call timerwheel_move to advance the timerwheel.
|
| |
|
|
|
|
|
|
|
|
| |
This revises the build system to have configuration per system
component. System settings can now be set using cmake.
The standard compliance defines were removed from configuration header
and are set in the sources where needed. Also some small code
refactors, such as moving the data for shims out of the ipcp structure
to the respective shims were performed.
|
|
|
This adds the basic FRCT mechanisms to the library. Upon flow alloc or
accept an FRCT instance is now created and used when reading or
writing to the flow. The timerwheel has been refactored to allow
recharging timers and removing them and is now part of the
library. The first SDU sent over the connection has the DRF set and
this initializes the connection. Sender and receiver inactivity timers
are added.
|
