This commit introduces the consensus parameter "circ_max_cell_queue_size"
which controls the maximum number of cells a circuit queue should have.
The default value is currently 50000 cells which is above what should be
expected but keeps us a margin of error for padding cells.
Related to this is #9072. Back in 0.2.4.14-alpha, we've removed that limit due
to a Guard discovery attack. Ticket #25226 details why we are putting back the
limit due to the memory pressure issue on relays.
Fixes#25226
Signed-off-by: David Goulet <dgoulet@torproject.org>
Both header and code file had some indentation issues after mass renaming.
No code behavior change.
Signed-off-by: David Goulet <dgoulet@torproject.org>
Really, the uint32_t is only an optimization; any kind of unit
should work fine. Some users might want to use time_t or
monotime_coarse_t or something like that.
Begin by creating a lowest-level triple of the types needed to
implement a token bucket: a configuration, a timestamp, and the raw
bucket itself.
Note that for low-level buckets, the units of the timestamp and the
bucket itself are unspecified: each user can use a different type.
(This patch breaks check-spaces; a later patch will fix it)
Add to the Denial of Service section of the man page an explanation about the
three different mitigation Tor has.
Fixes#25248.
Signed-off-by: David Goulet <dgoulet@torproject.org>
This is a simple search-and-replace to rename the token bucket type
to indicate that it contains both a read and a write bucket, bundled
with their configuration. It's preliminary to refactoring the
bucket type.
This test works by having two post-loop events activate one another
in a tight loop. If the "post-loop" mechanism didn't work, this
would be enough to starve all other events.
A linked connection_t is one that gets its I/O, not from the
network, but from another connection_t. When such a connection has
something to write, we want the corresponding connection to run its
read callback ... but not immediately, to avoid infinite recursion
and/or event loop starvation.
Previously we handled this case by activating the read events
outside the event loop. Now we use the "postloop event" logic.
This lets us simplify do_main_loop_once() a little.
We've been labeling some events as happening "outside the event
loop", to avoid Libevent starvation. This patch provides a cleaner
mechanism to avoid that starvation.
For background, the problem here is that Libevent only scans for new
events once it has run all its active callbacks. So if the
callbacks keep activating new callbacks, they could potentially
starve Libevent indefinitely and keep it from ever checking for
timed, socket, or signal events.
To solve this, we add the ability to label some events as
"post-loop". The rule for a "post-loop" event is that any events
_it_ activates can only be run after libevent has re-scanned for new
events at least once.
This differs from our previous token bucket abstraction in a few
ways:
1) It is an abstraction, and not a collection of fields.
2) It is meant to be used with monotonic timestamps, which should
produce better results than calling gettimeofday over and over.
In d1874b4339, we adjusted this check so that we insist on
using routerinfos for bridges. That's almost correct... but if we
have a bridge that is also a regular relay, then we should use
insist on its routerinfo when connecting to it as a bridge
(directly), and be willing to use its microdescriptor when
connecting to it elsewhere in our circuits.
This bug is a likely cause of some (all?) of the (exit_ei == NULL)
failures we've been seeing.
Fixes bug 25691; bugfix on 0.3.3.4-alpha
When size_t is 32 bits, the unit tests can't fit anything more than
4GB-1 into a size_t.
Additionally, tt_int_op() uses "long" -- we need tt_u64_op() to
safely test uint64_t values for equality.
Bug caused by tests for #24782 fix; not in any released Tor.
