When the send queue limit is reached, it is likely to not drain
any time soon. If we call close on the connection, it will stay
alive, waiting for the queue to drain before actually closing,
and will hit that check again and again. Since the queue size
limit is the reason we're closing in the first place, we call
shutdown directly.
If we reach the send queue size limit, we need to release the lock,
or we will deadlock and it will never drain.
If we reach that limit, it's likely there's another problem in the
first place though, so it will probably not drain in practice either,
unless some kind of transient network timeout.
Since connections from the ::connect method are now kept in
a deque to be able to cancel them on exit, this leaks both
memory and a file descriptor. Here, we clean those up after
30 seconds, to avoid this. 30 seconds is higher then the
5 second timeout used in the async code, so this should be
safe. However, this is an assumption which would break if
that async code was to start relying on longer timeouts.
When the boost ioservice is stopped, pending work notifications
will not happen. This includes deadline timers, which would
otherwise time out the now cancelled I/O operations. When this
happens just after starting a new connect operation, this can
leave that operations in a state where it won't receive either
the completion notification nor a timeout, causing a hang.
This is fixed by keeping a list of connections corresponding
to the connect operations, and cancelling them before stopping
the boost ioservice.
Note that the list of these connections can grow unbounded, as
they're never cleaned up. Cleaning them up would involve
working out which connections do not have any pending work,
and it's not quite clear yet how to go about this.
new update of the pr with network limits
more debug options:
discarding downloaded blocks all or after given height.
trying to trigger the locking errors.
debug levels polished/tuned to sane values.
debug/logging improved.
warning: this pr should be correct code, but it could make
an existing (in master version) locking error appear more often.
it's a race on the list (map) of peers, e.g. between closing/deleting
them versus working on them in net-limit sleep in sending chunk.
the bug is not in this code/this pr, but in the master version.
the locking problem of master will be fixed in other pr.
problem is ub, and in practice is seems to usually cause program abort
(tested on debian stable with updated gcc). see --help for option
to add sleep to trigger the error faster.
Update of the PR with network limits
works very well for all speeds
(but remember that low download speed can stop upload
because we then slow down downloading of blockchain
requests too)
more debug options
fixed pedantic warnings in our code
should work again on Mac OS X and FreeBSD
fixed warning about size_t
tested on Debian, Ubuntu, Windows(testing now)
TCP options and ToS (QoS) flag
FIXED peer number limit
FIXED some spikes in ingress/download
FIXED problems when other up and down limit
commands and options for network limiting
works very well e.g. for 50 KiB/sec up and down
ToS (QoS) flag
peer number limit
TODO some spikes in ingress/download
TODO problems when other up and down limit
added "otshell utils" - simple logging (with colors, text files channels)
moneromooo-monero
Riccardo Spagni
Howard Chu
Javier Smooth
rfree2monero
mydesktop
Antonio Juarez