I hit a strange problem with v2022.10: Sometimes my tftp transfer
would seemingly just hang. It only happened for some files. Moreover,
changing tftpblocksize from 65464 to 65460 or 65000 made it work again
for all the files I tried. So I started suspecting it had something to
do with the file sizes and in particular the way the tftp blocks get
fragmented and reassembled.
v2022.01 showed no problems with any of the files or any value of
tftpblocksize.
Looking at what had changed in net.c or tftp.c since January showed
only one remotely interesting thing, b85d130ea0.
So I fired up wireshark on my host to see if somehow one of the
packets would be too small. But no, with both v2022.01 and v2022.10,
the exact same sequence of packets were sent, all but the last of size
1500, and the last being 1280 bytes.
But then it struck me that 1280 is 5*256, so one of the two bytes
on-the-wire is 0 and the other is 5, and when then looking at the code
again the lack of endianness conversion becomes obvious. [ntohs is
both applied to ip->ip_off just above, as well as to ip->ip_len just a
little further down when the "len" is actually computed].
IOWs the current code would falsely reject any packet which happens to
be a multiple of 256 bytes in size, breaking tftp transfers somewhat
randomly, and if it did get one of those "malicious" packets with
ip_len set to, say, 27, it would be seen by this check as being 6912
and hence not rejected.
====
Now, just adding the missing ntohs() would make my initial problem go
away, in that I can now download the file where the last fragment ends
up being 1280 bytes. But there's another bug in the code and/or
analysis: The right-hand side is too strict, in that it is ok for the
last fragment not to have a multiple of 8 bytes as payload - it really
must be ok, because nothing in the IP spec says that IP datagrams must
have a multiple of 8 bytes as payload. And comments in the code also
mention this.
To fix that, replace the comparison with <= IP_HDR_SIZE and add
another check that len is actually a multiple of 8 when the "more
fragments" bit is set - which it necessarily is for the case where
offset8 ends up being 0, since we're only called when
(ip_off & (IP_OFFS | IP_FLAGS_MFRAG)).
====
So, does this fix CVE-2022-30790 for real? It certainly correctly
rejects the POC code which relies on sending a packet of size 27 with
the MFRAG flag set. Can the attack be carried out with a size 27
packet that doesn't set MFRAG (hence must set a non-zero fragment
offset)? I dunno. If we get a packet without MFRAG, we update
h->last_byte in the hole we've found to be start+len, hence we'd enter
one of
if ((h >= thisfrag) && (h->last_byte <= start + len)) {
or
} else if (h->last_byte <= start + len) {
and thus won't reach any of the
/* overlaps with initial part of the hole: move this hole */
newh = thisfrag + (len / 8);
/* fragment sits in the middle: split the hole */
newh = thisfrag + (len / 8);
IOW these division are now guaranteed to be exact, and thus I think
the scenario in CVE-2022-30790 cannot happen anymore.
====
However, there's a big elephant in the room, which has always been
spelled out in the comments, and which makes me believe that one can
still cause mayhem even with packets whose payloads are all 8-byte
aligned:
This code doesn't deal with a fragment that overlaps with two
different holes (thus being a superset of a previously-received
fragment).
Suppose each character below represents 8 bytes, with D being already
received data, H being a hole descriptor (struct hole), h being
non-populated chunks, and P representing where the payload of a just
received packet should go:
DDDHhhhhDDDDHhhhDDDD
PPPPPPPPP
I'm pretty sure in this case we'd end up with h being the first hole,
enter the simple
} else if (h->last_byte <= start + len) {
/* overlaps with final part of the hole: shorten this hole */
h->last_byte = start;
case, and thus in the memcpy happily overwrite the second H with our
chosen payload. This is probably worth fixing...
Signed-off-by: Rasmus Villemoes <rasmus.villemoes@prevas.dk>
1817c3824a