CVSROOT: /cvs Module name: src Changes by: d...@cvs.openbsd.org 2015/01/20 17:15:50
Modified files: lib/libssl/src/ssl: d1_pkt.c Log message: Fix DTLS memory leak (CVE-2015-0206). There were four bugs fixed by this patch: * dtls1_buffer_record() now frees rdata->rbuf.buf on error. Since s->s3->rbuf was memset, rdata->rbuf is the only pointer left which points to the old rbuf. On error, rdata is freed so there will not be any way of freeing this memory unless we do it here. * Changed the return code of dtls1_buffer_record() to differentiate between queue full (0) and error (-1). See below as this differs from upstream. * Handle errors if calls to dtls1_buffer_record() fail with -1. Previously, it did not check the return value. * Changed the way receipts are recorded. Previously, it was recorded when processed successfully (whether buffered or not) in dtls1_process_record(). Now, it records when it is handled in dtls1_get_record(): either when it is entered into the queue to buffer for the next epoch or when it is processed directly. Processing buffered records does not add a receipt because it needed one in order to get into the queue. The above bugs combined contributed to an eventual DoS through memory exhaustion. The memory leak came from dtls1_buffer_record()'s error handling. The error handling can be triggered by a duplicate record or malloc failure. It was possible to add duplicate records because they were not being dropped. The faulty receipts logic did not detect replays when dealing with records for the next epoch. Additionally, dtls1_buffer_record()'s return value was not checked so an attacker could send repeated replay records for the next epoch. Reported to OpenSSL by Chris Mueller. Patch based on OpenSSL commit 103b171d8fc282ef435f8de9afbf7782e312961f and BoringSSL commit 44e2709cd65fbd2172b9516c79e56f1875f60300. Our patch matches BoringSSL's commit. OpenSSL returns 0 when the queue is full or when malloc() or pitem_new() fails. They return -1 on error including !ssl3_setup_buffers() which is another failure to allocate memory. BoringSSL and LibreSSL changed the return code for dtls1_buffer_record() to be 1 on success, 0 when the queue is full and -1 on error. input + ok bcook@, jsing@