In 0.3.2.1-alpha, we've added notify_networkstatus_changed() in order to have
a way to notify other subsystems that the consensus just changed. The old and
new consensus are passed to it.
Before this patch, this was done _before_ the new consensus was set globally
(thus NOT accessible by getting the latest consensus). The scheduler
notification was assuming that it was set and select_scheduler() is looking at
the latest consensus to get the parameters it might needs. This was very wrong
because at that point it is still the old consensus set globally.
This commit changes the notify_networkstatus_changed() to be the "before"
function and adds an "after" notification from which the scheduler subsystem
is notified.
Fixes#24975
This is the quick fix that is keeping the channel in PENDING state so if we
ever try to reschedule the same channel, it won't happened.
Fixes#24700
Signed-off-by: David Goulet <dgoulet@torproject.org>
With extra_space negative, it means that the "notsent" queue is quite large so
we must consider that value with the current computed tcp_space. If we end up
to have negative space, we should not add more data to the kernel since the
notsent queue is just too filled up.
Fixes#24665
Signed-off-by: David Goulet <dgoulet@torproject.org>
Instead of using INT_MAX as a write limit for KISTLite, use the lower layer
limit which is using the specialized num_cells_writeable() of the channel that
will down the line check the connection's outbuf and limit it to 32KB
(OR_CONN_HIGHWATER).
That way we don't take the chance of bloating the connection's outbuf and we
keep the cells in the circuit queue which our OOM handler can take care of,
not the outbuf.
Finally, this commit adds a log_debug() in the update socket information
function of KIST so we can get the socket information in debug.
Fixes#24671
Signed-off-by: David Goulet <dgoulet@torproject.org>
In KIST, we could have a small congestion window value than the unacked
packets leading to a integer overflow which leaves the tcp_space value to be
humongous.
This has no security implications but it results in KIST scheduler allowing to
send cells on a potentially saturated connection.
Found by #24423. Fixes#24590.
Signed-off-by: David Goulet <dgoulet@torproject.org>
In the KIST main loop, if the channel happens to be not opened, set its state
to IDLE so we can release it properly later on. Prior to this fix, the channel
was in PENDING state, removed from the channel pending list and then kept in
that state because it is not opened.
This bug was introduced in commit dcabf801e5 for
which we made the scheduler loop not consider unopened channel.
This has no consequences on tor except for an annoying but harmless BUG()
warning.
Fixes#24502
Signed-off-by: David Goulet <dgoulet@torproject.org>
Some platforms don't have good monotonic time support so don't warn when the
diff between the last run of the scheduler time and now is negative. The
scheduler recovers properly from this so no need to be noisy.
Fixes#23696
Signed-off-by: David Goulet <dgoulet@torproject.org>
The flush cells process can close a channel if the connection write fails but
still return that it flushed at least one cell. This is due because the error
is not propagated up the call stack so there is no way of knowing if the flush
actually was successful or not.
Because this would require an important refactoring touching multiple
subsystems, this patch is a bandaid to avoid the KIST scheduler to handle
closed channel in its loop.
Bandaid on #23751.
Signed-off-by: David Goulet <dgoulet@torproject.org>
When a BUG() occurs, this macro will print extra information about the state
of the scheduler and the given channel if any. This will help us greatly to
fix future bugs in the scheduler especially when they occur rarely.
Fixes#23753
Signed-off-by: David Goulet <dgoulet@torproject.org>
Because our monotonic time interface doesn't play well with value set to 0,
always initialize to now() the scheduler_last_run at init() of the KIST
scheduler.
Fixes#23696
Signed-off-by: David Goulet <dgoulet@torproject.org>
When a channel is scheduled and flush cells returns 0 that is no cells to
flush, we flag it back in waiting for cells so it doesn't get stuck in a
possible infinite loop.
It has been observed on moria1 where a closed channel end up in the scheduler
where the flush process returned 0 cells but it was ultimately kept in the
scheduling loop forever. We suspect that this is due to a more deeper problem
in tor where the channel_more_to_flush() is actually looking at the wrong
queue and was returning 1 for an empty channel thus putting the channel in the
"Case 4" of the scheduler which is to go back in pending state thus
re-considered at the next iteration.
This is a fix that allows the KIST scheduler to recover properly from a not
entirelly diagnosed problem in tor.
Fixes#23676
Signed-off-by: David Goulet <dgoulet@torproject.org>
When the KIST schedule() is called, it computes a diff value between the last
scheduler run and the current monotonic time. If tha value is below the run
interval, the libevent even is updated else the event is run.
It turned out that casting to int32_t the returned int64_t value for the very
first scheduler run (which is set to 0) was creating an overflow on the 32 bit
value leading to adding the event with a gigantic usec value. The scheduler
was simply never running for a while.
First of all, a BUG() is added for a diff value < 0 because if the time is
really monotonic, we should never have a now time that is lower than the last
scheduler run time. And we will try to recover with a diff value to 0.
Second, the diff value is changed to int64_t so we avoid this "bootstrap
overflow" and future casting overflow problems.
Fixes#23558
Signed-off-by: David Goulet <dgoulet@torproject.org>
Otherwise integer overflows can happen. Remember, doing a i32xi32
multiply doesn't actually produce a 64-bit output. You need to do
i64xi32 or i64xi64.
Coverity found this as CID 1417753
Each type of scheduler implements its own static scheduler_t object and
returns a reference to it.
This commit also makes it a const pointer that is it can only change inside
the scheduler type subsystem but not outside for extra protection.
Signed-off-by: David Goulet <dgoulet@torproject.org>
