This document is incomplete; it describes some mechanisms that Tor uses to avoid different kinds of denial-of-service attacks.
(See also tor-spec.txt, section 8.1.)
The Tor protocol requires clients, onion services, relays, and authorities to store various kind of information in buffers and caches. But an attacker can use these buffers and queues to queues to exhaust the memory of the a targeted Tor process, and force the operating system to kill that process.
Worse still, the ability to kill targeted Tor instances can be used to facilitate traffic analysis. (For example, see the "Sniper Attack" paper by Jansen, Tschorsch, Johnson, and Scheuermann.
With this in mind, any Tor implementation—especially one that runs as a relay or onion service—must take steps to prevent memory-based denial-of-service attacks.
The easiest way to notice you're out of memory would, in theory, be getting an error when you try to allocate more. Unfortunately, some systems (e.g. Linux) won't actually give you an "out of memory" error when you're low on memory. Instead, they overcommit and promise you memory that they can't actually provide… and then later on, they might kill processes that actually try to use more memory than they wish they'd given out.
So in practice, the mainline Tor implementation uses a different strategy. It uses a self-imposed "MaxMemInQueues" value as an upper bound for how much memory it's willing to allocate to certain kinds of queued usages. This value can either be set by the user, or derived from a fraction of the total amount of system RAM.
As of Tor 0.4.7.x, the MaxMemInQueues mechanism tracks the following kinds of allocation:
- Cells queued on circuits.
- Per-connection read or write buffers.
- On-the-fly compression or decompression state.
- Half-open stream records.
- Cached onion service descriptors (hsdir only).
- Cached DNS resolves (relay only).
- GEOIP-based usage activity statistics.
Note that directory caches aren't counted, since those are stored on disk and accessed via mmap.
If our allocations exceed MaxMemInQueues, then we take the following steps to reduce our memory allocation.
Freeing from caches: For each of our onion service descriptor cache, our DNS cache, and our GEOIP statistics cache, we check whether they account for greater than 20% of our total allocation. If they do, we free memory from the offending cache until the total remaining is no more than 10% of our total allocation.
When freeing entries from a cache, we aim to free (approximately) the oldest entries first.
Freeing from buffers: After freeing data from caches, we see whether allocations are still above 90% of MaxMemInQueues. If they are, we try to close circuits and connections until we are below 90% of MaxMemInQueues.
When deciding to what circuits to free, we sort them based on the age of the oldest data in their queues, and free the ones with the oldest data. (For example, a circuit on which a single cell has been queued for 5 minutes would be freed before a circuit where 100 cells have been queued for 5 seconds.) "Data queued on a circuit" includes all data that we could drop if the circuit were destroyed: not only the cells on the circuit's cell queue, but also any bytes queued in buffers associated with streams or half-stream records attached to the circuit.
We free non-tunneled directory connections according to a similar rule, according to the age of their oldest queued data.
Upon freeing a circuit, a "DESTROY cell" must be sent in both directions.
We define a "low threshold" equal to 3/4 of MaxMemInQueues. Every time our memory usage is above the low threshold, we record ourselves as being "under memory pressure".
(This is not currently reported.)