Please sign in to comment.
net: sctp: fix remote memory pressure from excessive queueing
This scenario is not limited to ASCONF, just taken as one example triggering the issue. When receiving ASCONF probes in the form of ... -------------- INIT[ASCONF; ASCONF_ACK] -------------> <----------- INIT-ACK[ASCONF; ASCONF_ACK] ------------ -------------------- COOKIE-ECHO --------------------> <-------------------- COOKIE-ACK --------------------- ---- ASCONF_a; [ASCONF_b; ...; ASCONF_n;] JUNK ------> [...] ---- ASCONF_m; [ASCONF_o; ...; ASCONF_z;] JUNK ------> ... where ASCONF_a, ASCONF_b, ..., ASCONF_z are good-formed ASCONFs and have increasing serial numbers, we process such ASCONF chunk(s) marked with !end_of_packet and !singleton, since we have not yet reached the SCTP packet end. SCTP does only do verification on a chunk by chunk basis, as an SCTP packet is nothing more than just a container of a stream of chunks which it eats up one by one. We could run into the case that we receive a packet with a malformed tail, above marked as trailing JUNK. All previous chunks are here goodformed, so the stack will eat up all previous chunks up to this point. In case JUNK does not fit into a chunk header and there are no more other chunks in the input queue, or in case JUNK contains a garbage chunk header, but the encoded chunk length would exceed the skb tail, or we came here from an entirely different scenario and the chunk has pdiscard=1 mark (without having had a flush point), it will happen, that we will excessively queue up the association's output queue (a correct final chunk may then turn it into a response flood when flushing the queue ;)): I ran a simple script with incremental ASCONF serial numbers and could see the server side consuming excessive amount of RAM [before/after: up to 2GB and more]. The issue at heart is that the chunk train basically ends with !end_of_packet and !singleton markers and since commit 2e3216c ("sctp: Follow security requirement of responding with 1 packet") therefore preventing an output queue flush point in sctp_do_sm() -> sctp_cmd_interpreter() on the input chunk (chunk = event_arg) even though local_cork is set, but its precedence has changed since then. In the normal case, the last chunk with end_of_packet=1 would trigger the queue flush to accommodate possible outgoing bundling. In the input queue, sctp_inq_pop() seems to do the right thing in terms of discarding invalid chunks. So, above JUNK will not enter the state machine and instead be released and exit the sctp_assoc_bh_rcv() chunk processing loop. It's simply the flush point being missing at loop exit. Adding a try-flush approach on the output queue might not work as the underlying infrastructure might be long gone at this point due to the side-effect interpreter run. One possibility, albeit a bit of a kludge, would be to defer invalid chunk freeing into the state machine in order to possibly trigger packet discards and thus indirectly a queue flush on error. It would surely be better to discard chunks as in the current, perhaps better controlled environment, but going back and forth, it's simply architecturally not possible. I tried various trailing JUNK attack cases and it seems to look good now. Joint work with Vlad Yasevich. Fixes: 2e3216c ("sctp: Follow security requirement of responding with 1 packet") Signed-off-by: Daniel Borkmann <firstname.lastname@example.org> Signed-off-by: Vlad Yasevich <email@example.com> Signed-off-by: David S. Miller <firstname.lastname@example.org>
- Loading branch information...
Showing with 10 additions and 26 deletions.