Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
🔒 Limit exponential memory usage to parse uid-set
The UID sets in UIDPlusData are stored as arrays of UIDs. In common scenarios, copying between one and a few hundred emails at a time, this is barely noticable. But the memory use expands _exponentially_. This should not be an issue for _trusted_ servers, and (I assume) compromised servers will be more interested in evading detection and stealing your credentials and your email than in causing client Denial of Service. Nevertheless, this is a very simple DoS attack against clients connecting to untrusted servers (for example, a service that connects to user-specified servers). For example, assuming a 64-bit architecture, considering only the data in the two arrays, assuming the arrays' internal capacity is no more than needed, and ignoring the fixed cost of the response structs: * 32 bytes expands to ~160KB (about 5000 times more): `"* OK [COPYUID 1 1:9999 1:9999]\r\n"` * 40 bytes expands to ~1.6GB (about 50 million times more): `"* OK [COPYUID 1 1:99999999 1:99999999]\r\n"` * In the worst scenario (uint32 max), 44 bytes expands to 64GiB in memory, using over 1.5 billion times more to store than to send: `"* OK [COPYUID 1 1:4294967295 1:4294967295]\r\n"` ---- The preferred fix is to store `uid-set` as a SequenceSet, not an array. Unfortunately, this is not fully backwards compatible. For v0.4 and v0.5, use `Config#parser_use_deprecated_uidplus_data` to false to use AppendUIDData and CopyUIDData instead of UIDPlusData. Unless you are _using_ UIDPLUS, this is completely safe. v0.6 will drop UIDPlusData. ---- The simplest _partial_ fix (preserving full backward compatibility) is to raise an error when the number of UIDs goes over some threshold, and continue using arrays inside UIDPlusData. For v0.3.x (and in this commit) the maximum count is hard-coded to 10,000. This is high enough that it should almost never be triggered by normal usage, and low enough to be a less extreme problem. For v0.4 and v0.5, the next commit will make the maximum array size configurable, with a much lower default: 1000 for 0.4 and 100 for 0.5. These are low enough that they are _unlikely_ to cause a problem, but 0.4 and 0.5 can also use the newer AppendUIDData and CopyUIDData classes. However, because unhandled responses are stored on the `#responses` hash, this can still be a problem. A malicious server could repeatedly use 160Kb of client memory by sending only 32 bytes in a loop. To fully solve this problem, a response handler must be added to prune excessive APPENDUID/COPYUID responses as they are received. Because unhandled responses have always been retained, managing unhandled responses is already documented as necessary for long-lived connections.
- Loading branch information