On Thu, Aug 13, 2020 at 09:46:11AM +0200, Florian Westphal wrote: > syzkaller reports splat: > ------------[ cut here ]------------ > Buffer overflow detected (80 < 137)! > Call Trace: > do_ebt_get_ctl+0x2b4/0x790 net/bridge/netfilter/ebtables.c:2317 > nf_getsockopt+0x72/0xd0 net/netfilter/nf_sockopt.c:116 > ip_getsockopt net/ipv4/ip_sockglue.c:1778 [inline] > > caused by a copy-to-user with a too-large "*len" value. > This adds a argument check on *len just like in the non-compat version > of the handler. > > Before the "Fixes" commit, the reproducer fails with -EINVAL as > expected: > 1. core calls the "compat" getsockopt version > 2. compat getsockopt version detects the *len value is possibly > in 64-bit layout (*len != compat_len) > 3. compat getsockopt version delegates everything to native getsockopt > version > 4. native getsockopt rejects invalid *len > > -> compat handler only sees len == sizeof(compat_struct) for GET_ENTRIES. > > After the refactor, event sequence is: > 1. getsockopt calls "compat" version (len != native_len) > 2. compat version attempts to copy *len bytes, where *len is random > value from userspace Applied, thanks.
