In `srtp_unprotect_rtcp()` we are not validating that the packet
length is as long as the minimum required. This would cause
`enc_octet_len` to underflow, which would cause us to try to decrypt
data past the end of the packet in memory -- a buffer over-read and
buffer overflow.
In `srtp_protect_rtcp()`, we were similarly not validating the packet
length. Here we were also polluting the address of the SRTCP
encrypted flag and index (the `trailer`), causing us to write one word
to a bogus memory address before getting to the encryption where we
would also overflow.
In this commit we add checks to appropriately validate the RTCP/SRTCP
packet lengths.
`srtp_unprotect_rtcp_aead()` (but not protect) did correctly validate
the packet length; this check would now be redundant as the check in
`srtcp_unprotect_rtcp()` will also run first, so it has been removed.
In the defined AEAD modes, SRTP packets must always be encrypted and
authenticated, but SRTCP packets may be only authenticated. It's
possible, therefore, for us to end up in `srtp_protect_aead()` without
the `sec_serv_conf` bit being set. We should just ignore this and
encrypt the RTP packet anyway.
What we are doing instead is encrypting the packet anyway, but setting
`enc_start` to NULL first. This causes `aad_len` to underflow which
will cause us to over-read in `cipher_set_aad()`.
If we could get past that, we would try to read and write memory
starting at 0x0 down in `cipher_encrypt()`.
This commit causes us to not check the `sec_serv_conf` bit and never
set `enc_start` to NULL in `srtp_protect_aead()`.
`srtp_unprotect_aead()` does not contain a similar error.
When computing the start address of the RTP data to encrypt or SRTP
data to decrypt (`enc_start`), we are using `hdr->cc` (the CSRC
count), which is untrusted data from the packet, and the length field
of an RTP header extension, which is also untrusted and unchecked data
from the packet.
This value then pollutes our calculation of how much data we'll be
encrypting or decrypting (`enc_octet_len`), possibly causing us to
underflow.
We'll then call `cipher_encrypt()` or `cipher_decrypt()` with these
two values, causing us to read from and write to arbitrary addresses
in memory.
(In the AEAD functions, we'd also pollute `aad_len`, which would cause
us to read undefined memory in `cipher_set_aad`.)
This commit adds checks to verify that the `enc_start` we calculate is
sane based on the actual packet length.
We're checking whether we've hit the warning threshold before checking
whether we should just end the call. This causes an off-by-one error
where we take one SRTP error more than intended.
This commit reverses the order of the tests.
In a carrier interop we saw the call get killed for SRTP failures
during a reinvite. We're wondering if the SRTP errors may have been
transitory and if it may have recovered after a few more packets.
It's debatable whether we should kill calls at all for SRTP auth
failures; semantically the right thing to do when a MAC fails is to
ignore the packet completely. So raising this limit to 100 packets
shouldn't do any harm. With this change we still warn at 10 errors
and every 10 errors thereafter.
We hangup the channel after receiving 10 SRTP packets in a row with a
bad auth tag or that are replayed. Prior to this commit we were
indicating a normal clearing. When doing interop and looking first at
packet traces, this made freeswitch's behavior look surprising. With
this commit we'll indicate more loudly what's happening.
switch_rtp_set_invalid_handler has been misspelled as
switch_rtp_set_invald_handler going all the way back to the
beginning. So while it's possible that someone somewhere could be
relying on this misspelling, I think it's more likely that no one has
used it much and that's why it wasn't spotted. We don't even use it
ourselves anywhere anymore.
Introduced in commit: 828e03715f
mod_sofia's parameter shutdown-on-fail now accepts the value
"reincarnate-now". This will cause the switch to exit immediately
with a non-zero exit code so that the supervisor can recover the
switch. For this to work you have to pass in -reincarnate or
-reincarnate-reexec to freeswitch.
This is the result of auditing each mod_sofia profile parameter to
ensure that it can be unset or reset after being set. One use-case
for this being done correctly is so a later parameter in a
configuration file can reliably override an earlier one, which is
useful for setups with layered include files.
This commit allows you to set a `log-file` string parameter in a
format_cdr profile. This string is a template that may (and should!)
contain variables. This template will be expanded and used as the
file name of the CDR to be written. This parameter should contain
only the template for the file name itself; the path is relative to
the `log-dir`.
Previously if send-display-update was set to false we would also
remove UPDATE from our Allow: headers. This is unnecessary. The
UPDATE message is useful in SIP transactions even if we're not sending
display updates.
With this commit, we add a new boolean profile flag, allow-update. If
set to true we'll send Allow: UPDATE. If set to false, we will not.
If there is a conflict with another setting that requires UPDATE
support, the allow-update parameter will win and a warning will be
printed.
ref: RFC 3311
Anthony Minessale
Chris Rienzo
jfigus
Travis Cross
Steve Underwood
Brian West
Seven Du
William King
Steve Underwood
Eliot Gable