Forcing the DTLS role to passive only makes sense for the backwards
(torwards the offerer) direction anyway, as towards the answerer (in an
offer) it's not allowed by the RFC, and in an answer it's not possible
as the active DTLS connection would already be in progress.
Convert DTLS= processing into a list as multiple flags might be
required.
Change-Id: I54239280460e1bbce1966bc17a384489eb978618
Amends: 78ee0451 / Ib1d07831
Related: #2156
Related: #2145
(cherry picked from commit 032bd398b9)
Using the FILLED flag to decide how to resolve actpass is not reliable,
as during a reinvite FILLED would be set for both stream directions.
Instead, immediately clear the ACTIVE flag during the offer phase
(dtls_logic) and assume a strict passive role. Then if the answer
indicates that the connection needs to be active instead, a reset occurs
via dtls_connection_init.
Closes#2156
Change-Id: I0ded94620bcaa481eea920ae0c412c7bb86bff1f
(cherry picked from commit 91d0669499)
DTLS connections created during an offer are passive by default. If the
answer requires a role change to active, don't reset the TLS ID, as we
have no way to communicate this change until a re-invite, which would
then lead to an unexpected reset of the connection.
Related: #2156
Change-Id: I49dcd93d63840a223c41a59c74e2f270d0a08f62
(cherry picked from commit 001d8fe4c6)
The primary pair selection already checks this, but a retransmit can
still trigger a check on a socket that has since been closed.
Closes#2154
Change-Id: I8308693e7269bfa9dcfcb24daf6863669c065139
(cherry picked from commit c22d724aa9)
6.12.101+ has removed this as well. Use the macro itself to determine
whether this struct member exists.
Change-Id: Ib43157567a78afc664398b9f22ee0a27b0facfa0
(cherry picked from commit 83b05e9a11)
On some 32-bit architectures, such as armhf, the CPU does not have a
64-bit division instruction, and the compiler injects an intrinsic
function reference (such as __aeabi_ldivmod) that is provided for
example by libgcc.
When compiling Linux kernel code, we are not using libgcc, so such
references end up being undefined, failing the link. Instead we need to
use support provided by the kernel itself. In this case the do_div()
macro.
Change-Id: I5ed60aee487e66d6fcfc2293644d773b6820c178
(cherry picked from commit b688988363)
We are compressing the documentation explicitly with gzip, but were not
passing -n to get reproducible output.
Change-Id: I7819701ef33d34472eb2a63458ae16ef69ac352e
(cherry picked from commit 9b74a7fd12)
This is checked for race conditions but is not actually increased
anywhere.
Fixes#2148
Change-Id: Ia46ea32ffea6abf267f7ae872368f4f9c4446f29
(cherry picked from commit 5daae51b5b)
If a SSRC object is still in use (refcount is >1), keep it in the list,
even if the list is too large.
Otherwise we may end up recreating a new SSRC object for the same SSRC
that still has an object in use elsewhere, making locks unreliable.
Ref #2148
Change-Id: I2afd277696e05c356fc9cf5aea59f51e1fba0784
(cherry picked from commit 26fee2610a)
Per RFC 4867 section 8.1, absent fmtp parameters imply defaults (bandwidth-efficient mode, no CRC,
no robust-sorting, no interleaving). Previously, codec matching was rejected when either side had no
``a=fmtp`` attribute for the AMR payload type, causing AMR to be dropped as a "stray answer codec".
This patch changes this behavior to instead "parse" fmtp even when absent so that the format struct
is populated with zero-initialized defaults and fmtp_parsed is set.
Closes#2141
Change-Id: I1371cb1fdb9354746a4019477845d777e619a2f6
(cherry picked from commit 702ccec969)
To not come across the sink's mutation (e.g. `endpoint.address.family`)
just lock way earlier, before to start processing and
translating endpoint to the kernel related structs.
Otherwise the other signaling path may mutate something
which will make the sink unactual and provide
to the `__re_address_translate_ep()` NULLed or garbage structs.
Change-Id: I88d87ae88abe4e37aaa5d3cdda76ed0ca82e4842
(cherry picked from commit e16a2309e0)
16de9688d6 resets the `socket.family` to NULL,
that makes the direct appeal to the corresponding
object unsafe.
E.g. `ps->selected_sfd->socket.family->name` when
filling the streams.
Add a simply guard and log `none` in case
the ip family is NULL.
Change-Id: I7ad7399e32e5d0ffb2fcc919822d05760fdc7421
(cherry picked from commit 05817bad13)
We use `family == NULL` as test for the socket being open in several
places, and expect it to match the familt of the addresses. Set it to
zero to make it consistent.
Change-Id: Ide2b41e84fc27b5f47f579dcf4ab2ec2f6b0a37e
(cherry picked from commit 16de9688d6)
ICE drops packets early in media_demux_protocols() (via
ice_peer_address_known()), which means in media_packet_address_check()
we can immediately set any socket as confirmed that has received
anything.
Take receiving socket into account to determine whether the endpoint is
a match.
Disable endpoint learning with ICE. ICE is in charge of determining
endpoint address and local socket.
`media-handover` and `strict-source` handling remains in place. With
`media-handover` set, we allow changing endpoint and local socket based
on what is actually received, overriding the nomination from ICE.
Change-Id: Ieac2893355df6ee560b497867181e6f47d646f2e
(cherry picked from commit d3af013494)
Don't just check for the selected_sfd to be non null, but also for the
socket to actually be open, as this is needed for the translate_ep set
of functions in particular.
Change-Id: I78cb0f7b8c9e4359c9e037485b29f63e0cd9476b
(cherry picked from commit 9cb902319c)
A stream that needs to be unkernelized that wasn't actually kernelized
due to lack of support or information still needs to have the flag reset
so that it can retried.
Change-Id: I7de17236d382d4f4dbfc3225c37688769cc9c8ad
(cherry picked from commit ab8b7fb228)
Pull the kernel in/out checks down below the "out" label, and use the
existing flags from the STUN connect event instead of directly calling
the relevant function.
Change-Id: Ide1ebf0d18446427535d9e63dadf7fdbbddc6c94
(cherry picked from commit 209e5a38dc)
Set the stream as kernelized even if it couldn't be, so that we don't
keep trying.
Change-Id: Icf05c6cab2f52d918ba04e7cb2089347b3f72378
(cherry picked from commit 46e8d6c305)
If SRTP/DTLS isn't yet established, do fully bail out of the kernelize
function, instead of just logging a message and blindly continuing, as
it would eventually fail due to missing or invalid SRTP parameters.
Fix-up for f2d57de078
Change-Id: I3521f54fe2001d3c67118a3502807d8cb6ada7ee
(cherry picked from commit 7ab5cf3ebf)
We can bail right there instead of having to rely on the
`(!reti->local.family)` condition
Change-Id: I584e90dd1a705cfbe1cd3e3d849e1fa88f511448
(cherry picked from commit dbcd657908)
Previously, the ports used for a monologue subscription would leak upon re-subscribe.
``__assign_stream_fds`` cleared the stream fd queue without decrementing the ref count. This
prevented ``monologue_destroy`` from releasing the ports on unsubscribe.
Closes#2125
Change-Id: Ib86290d6a35e2388e34098f2847f1e34b3db2c5d
(cherry picked from commit 250df2af5e)
Similar in spirit to f2d57de078, but for
the encrypt case.
Closes#2127
Change-Id: I06c82f4a1d768a863922e2240aebb6c1d243e269
(cherry picked from commit e6469c006b)
In case of the reverse direction of a subscription, no stream handler will be found. Previously, the
decrypt struct was left zeroed, the kernel module rejected this with ``EINVAL`` and the daemon logged:
```
[core] Failed to push relay stream to kernel: Invalid argument
```
This patch sets ``REC_NULL``/``REH_NULL`` and marks the target as non-forwarding, matching what is
expected by ``validate_srtp``.
Closes#2126
Change-Id: I5aa2b490c65e5d97dc9b2ddd24c09f9ae7b63892
(cherry picked from commit 47371520ee)
Allow invalid payload types for rejected RTP streams.
Allow empty codec store for zero-port streams.
Has some side effects on tests.
Closes#2122
Change-Id: I28c63a3e5b7309084e7899a19ff4550e8f56aac5
(cherry picked from commit 71cceeed5a)
ffmpeg's CNG expects a fixed frame size, or at least a maximum frame
size, as it allocates internal buffers on construction based on that
frame size. Calling the decoder with a larger frame size therefore leads
to an internal buffer overflow. Calling it with a smaller frame size is
fine.
Retain the simple one-shot CNG for the most common case (frame size
smaller than the max). For larger frames, call the CNG in a loop and
manually construct a return frame.
Change-Id: I0ebe0b7a9ad8a08bbabd0d8dea264b08cb44feee
(cherry picked from commit dfcb9a8322)
Richard Fuchs
Guillem Jover
Gustavo Almeida
Alexander Bakker
Donat Zenichev
Idan Freiberg