| Commit message (Collapse) | Author | Age | Files | Lines |
... | |
|
|
|
|
|
|
|
|
|
|
|
| |
The mb-ecn policy has a couple of divisions in the math, which I
wanted to avoid. Now it measures the number of bytes sent in a window,
and updates the next window with AIMD logic. If the number of bytes in
the window is reached, the call blocks. To avoid long packet bursts,
the window size continually scales to contain between CA_MINPS (8) and
CA_MAXPS (64) packets.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
The dt component bypasses the flow allocator on the receiver side, and
may try to update congestion context when the flow has already been
deallocated by the receiver. I will fix this bypass and always pass
through the flow allocator sometime soon; for now, I added a check in
the flow allocator call to avoid the SEGV.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
The enrollment procedure was not passing the policy for congestion
avoidance.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This adds congestion avoidance policies to the unicast IPCP. The
default policy is a multi-bit explicit congestion avoidance algorithm
based on data-center TCP congestion avoidance (DCTCP) to relay
information about the maximum queue depth that packets experienced to
the receiver. There's also a "nop" policy to disable congestion
avoidance for testing and benchmarking purposes.
The (initial) API for congestion avoidance policies is:
void * (* ctx_create)(void);
void (* ctx_destroy)(void * ctx);
These calls create / and or destroy a context for congestion control
for a specific flow. Thread-safety of the context is the
responsability of the flow allocator (operations on the ctx should be
performed under a lock).
ca_wnd_t (* ctx_update_snd)(void * ctx,
size_t len);
This is the sender call to update the context, and should be called
for every packet that is sent on the flow. The len parameter in this
API is the packet length, which allows calculating the bandwidth. It
returns an opaque union type that is used for the call to check/wait
if the congestion window is open or closed (and allowing to release
locks before waiting).
bool (* ctx_update_rcv)(void * ctx,
size_t len,
uint8_t ecn,
uint16_t * ece);
This is the call to update the flow congestion context on the receiver
side. It should be called for every received packet. It gets the ecn
value from the packet and its length, and returns the ECE (explicit
congestion experienced) value to be sent to the sender in case of
congestion. The boolean returned signals whether or not a congestion
update needs to be sent.
void (* ctx_update_ece)(void * ctx,
uint16_t ece);
This is the call for the sending side top update the context when it
receives an ECE update from the receiver.
void (* wnd_wait)(ca_wnd_t wnd);
This is a (blocking) call that waits for the congestion window to
clear. It should be stateless (to avoid waiting under locks). This may
change later on if passing the context is needed for different algorithms.
uint8_t (* calc_ecn)(int fd,
size_t len);
This is the call that intermediate IPCPs(routers) should use to update
the ECN field on passing packets.
The multi-bit ECN policy bases the value for the ECN field on the
depth of the rbuff queue packets will be sent on. I created another
call to grab the queue depth as fccntl is write-locking the
application. We can further optimize this to avoid most locking on the
rbuff.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
|
| |
The ocbr client was spinning the CPU by default, which made sense on
lab servers with dual xeons, but not so much for average users. Now
sleeping becomes the default. Busy waiting can be enabled using --spin
if needed.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
| |
The ocbr server was using non-blocking reads (probably because we
didn't have read timeouts when we wrote it) and was using a whole CPU
core per thread.
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]>
|
|
|
|
|
|
|
|
|
|
|
| |
The flow information in the main loop is passed as a direct pointer to
an irm_flow object in the flow database. This was (probably) not
really an issue due to how the flow allocation operations work, but
the thread sanitizer was barfing a lot of (correct) data race errors
when running bigger tests, so now makes a safe copy of the data.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
I mistakenly set the default to the (buggy) lockless rbuff
implementation instead of the pthread one in commit 3aec660e.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
This reverts commit 978266fe4beba21292daad2d341fe5ff22e08aba.
We were incorrectly unmounting the directory under normal conditions.
Signed-off-by: Sander Vrijders <[email protected]>
Signed-off-by: Dimitri Staessens <[email protected]>
|
|
|
|
|
|
|
| |
The flow stats had quite a lot of duplication.
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]>
|
|
|
|
|
|
|
|
|
|
|
| |
If the sending window for flow control is closed, the sending
application will now block until the window opens. Beware that until
the rendez-vous mechanism is implemented, shutting down a server while
the client is sending (with non-timed-out blocking write) will cause
the client to hang indefinitely because its window will close.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
| |
Refactor flow_write cleanup.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
| |
This adds sending and receiving window updates for flow control. I
used the 8 pad bits as part of the window update field, so it's 24
bits, allowing for ~16 million packets in flight.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
| |
The function was returning under a cleanup handler, which is not
allowed. We don't do anything with the return value if the write
thread ends, so just stopping the thread is fine.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
The condition variable was not initialized correctly and using the
wrong clock for pthread_cond_timedwait.
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]>
|
|
|
|
|
|
|
|
| |
Flows should be locked when moving the timerwheel. For frcti_snd, a
rdlock is enough.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
| |
A flow_set is thread-safe and doesn't need to be protected by a lock.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
| |
Fix assignment instead of comparison operator.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
| |
There was a dealloc() call in oping server under mutex, which could
leave that mutex locked when the thread was cancelled, causing oping
to hang on exit. This avoids calling dealloc under lock.
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 the logic to send a pure acknowledgment packet without any
data to send. This needed the event filter for the fqueue, as these
non-data packets should not trigger application PKT events. The
default timeout is now 10ms, until we have FRCP tuning as part of
fccntl.
Karn's algorithm seems to be very unstable with low (sub-ms) RTT
estimates. Doubling RTO (every RTO) seems still too slow to prevent
rtx storms when the measured rtt suddenly spikes several orders of
magnitude. Just assuming the ACK'd packet is the last one transmitted
seems to be a lot more stable. It can lead to temporary
underestimation, but this is not a throughput-killer in FRCP.
Changes most time units to nanoseconds for faster computation.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
The sanitize function in the rdrbuff should only be compiled if robust
mutexes are present on the system.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
There were some issues identified by the Clang static analyzer that
are now fixed.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
GCC 10 static analyzer found that the wrong index was used in the fail
path of psched_create, causing double (multiple) frees.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
| |
GCC 10 defaults to -fno-common, so some variables that were defined in
the headers needed to be declared "extern". The GCC 10 static analyzer
can now be invoked using the DebugAnalyzer build option.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
|
| |
This is a small refactor of FRCT because I found some things a bit
hard to read. I tried to refactor frcti_rcv to always queue the
packet, but that causes unnecessarily retaking the lock when calling
queued_pdu and thus returning idx is a tiny bit faster.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
| |
The compiler spotted some variables that weren't really used.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
GCC 9.3.0 started complaining despite the /* FALLTHRU */
comments. Apparently this changed level.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
|
| |
The retransmission was always disabling the DRF flag. This caused
problems with the loss of the first packet, which of course needs a
DRF flag set. The retransmitted packet will now contain a the original
DRF flag and an updated ack number.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
| |
The single retransmission wheel caused locking headaches as the calls
for different flows could block on the same rxmwheel. This stabilizes
the stack, but if the rdrbuff gets full there can now be big delays.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
On a bad write, the writer thread would shutdown, leaving the
client hanging.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
The equal-cost multipath option wasn't mentioned in the Ouroboros man
page.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
|
| |
Fixes infinite rescheduling with RTO getting lower than the timerwheel
resolution. For very low RTO values we'd need a big packet buffer with
the current memory allocator implementation (rdrbuff). Setting a
(configurable) minimum RTO (250 us) reduces this need.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
If Ouroboros crashed, the RIB directory might still be mounted. This
checks if this is the case, then unmounts it.
Signed-off-by: Sander Vrijders <[email protected]>
Signed-off-by: Dimitri Staessens <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
There were a bunch of bugs in FRCP that urgently needed fixing. Now
data QoS is usable even with heavy packet loss (within some
parameters). The current RTT estimator is the IETF one. It should be
updated to the improved one used in the Linux kernel once the A-timer
(ACKs without data) and graceful shutdown are implemented.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
The allocation response was always containing an ECDHE key, which is
not needed if the client doesn't request an encrypted flow.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
| |
The shm_flowset destroy was using the irmd pid, resulting in wrong
unlinks. The irmd was not cleaning up the process table, resulting in
shm leaks if there were still running processes on exit.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
| |
The thread pool manager wasn't counting working threads when deciding
to create new ones, resulting in constant starting of new threads when
threads were busy.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
| |
This is more in line with the write() system call and prepares for
partial writes. Partial writes are disabled by default (and not yet
implemented).
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
The return type was still an int, but since it returns the number of
events, it should be an ssize_t.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This revises the naming API to treat names (or reg_name in the source)
as first-class citizens of the architecture. This is more in line with
the way they are described in the article.
Operations have been added to create/destroy names independently of
registering. This was previously done only as part of register, and
there was no way to delete a name from the IRMd. The create call now
allows specifying a policy for load-balancing incoming flows for a
name. The default is the new round-robin load-balancer, the previous
behaviour is still available as a spillover load-balancer.
The register calls will still create a name if it doesn't exist, with
the default round-robin load-balancer.
The tools now have a "name" section, so the format is now
irm name <operation> <name> ...
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
| |
The service was called "irmd", but it makes a bit more sense to give
it the system name. Only the service name is changed, the irmd binary
remains irmd.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
There were updates under rdlock instead of wrlock, causing data races
and trouble. Also speeds up shutdown a bit.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
| |
There was a rare deadlock upon destruction of the threadpool manager
because the threads were cancelled/joined under lock.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
|
|
| |
The initial implementation for the ECDHE key exchange was doing the
key exchange after a flow was established. The public keys are now
sent allowg on the flow allocation messages, so that an encrypted
tunnel can be created within 1 RTT. The flow allocation steps had to
be extended to pass the opaque data ('piggybacking').
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
| |
This adds tests for LFA and ECMP to the graph_test routine.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|
|
|
|
|
|
|
|
|
| |
There were bugs introduced in the LFA algorithm refactor causing
infinite recursion and SEGV. The infinite recursion check was added as
an explicit compiler flag to the build.
Signed-off-by: Dimitri Staessens <[email protected]>
Signed-off-by: Sander Vrijders <[email protected]>
|