It was seen on a real pcap trace (sctp & gsmtap_log) that the Linux
kernel stack may decide to kill the connection (sending an ABORT) if
it fails to transmit some data after a while:
ABORT Cause code: "Protocol violation (0x000d)",
Cause Information: "Association exceeded its max_retrans count".
When this occurs, the kernel sends the
MSG_NOTIFICATION,SCTP_ASSOC_CHANGE,SCTP_COMM_LOST notification when
reading from the socket with sctp_recvmsg(). This basically signals that
the socket conn is dead, and subsequent writes to it will result in
send() failures (and receive SCTP_SEND_FAILED notification upon follow
up reads).
It's important to notice that after those events, there's no other sort
of different event like SHUTDOWN coming in, so that's the time at which
we must tell the user to close the socket.
Hence, let's signal the caller that the socket is dead by returning 0,
to comply with usual recv() API.
Related: SYS#6113
Change-Id: If94d44f25b76a96a5ea402fec9fc14c4e6296ba3
It could be that the user reuses the msgb passed to
osmo_stream_srv_recv() all the time, hence we need to reset the flags,
otherwise it may end up being set forever.
Change-Id: Id358140db82b018e3973111e530834589c0b7224
This allows the user to properly check the size of the content in case
the struct ABI changes.
-EAGAIN is still returned to avoid older users of the API reading SCTP
notifications as user data.
Change-Id: I95e2457498fd8e0d790d221cb97695ace0dd673e
The user may want to check the flags in order to know if the content
pointed at by msgb is an sctp_notification structure, and not in-band
data.
This is useful for the user in order to gain connection state such as
SCTP RESET notification, which means the client reconnected reusing the
same socket, loosing all higher layers state.
The MSG_NOTIFICATION from netinet/sctp.h is not reused, and instead an
osmocom specific flag (and bitmask) is used. This is done in order to
fit the flags in one byte, since for instance MSG_NOTIFICATION requires
2 bytes in my system (0x8000).
Related: SYS#6113
Change-Id: I0ee94846a15a23950b9d70eaaef1251267296bdd
The fd is not valid anymore after close() call, so let's set it to a
clearly invalid value, to avoid confusion on closed_cb() users
attempting to use the fd get info about the socket at that time.
Change-Id: I82d9bdfb38cf5e9f689eca0d9a4c19ddd5541af7
Found out by following gcc warning:
"""
libosmo-netif/src/stream.c:147:11: warning: argument to variable-length array may be too large due to conversion from ‘int’ to ‘unsigned int’ [-Wvla-larger-than=]
147 | uint8_t buf[sctp_sockopt_event_subscribe_size];
| ^~~
"""
Change-Id: I181ae5c0480788fc96abd2bb1edf003244884c8f
This makes it easier to follow the general path selection based on
protocol type. It will also make it easier when we add new paths based
on socket domain.
Change-Id: Ia3e0f4407e00a2dac9ee885fe1cc1cb4b463898a
In the case of SCTP, an association supports now containing both IPv4
and v6 addresses.
Depends: libosmocore.git Change-Id I36d8ab85d92bba4d6adb83bc1875eb61094ed2ef
Change-Id: Ie6bb17a9af6ca21d5e350f9c9d2d74c97c5a00af
Some previous commit added some code to workaround some kernel ABI
unstabilities but forgot to add the code inside a conditional
preprocessor block.
Fixes: 157ad6d707
Change-Id: Ibee01222d8d11cf2520cdbd4d0d5bed6b72d2107
New define is available since libosmocore 1.1.0, and we already require
1.3.0, so no need to update dependenices.
Let's change it to avoid people re-using old BSC_FD_* symbols when
copy-pasting somewhere else.
Change-Id: I269690c1c9e4d19b5b69eef206b95e71f7931188
It makes a lot more sense calling it this way since it matches the state
of the stream at that point.
Change-Id: Ic02aec3f7f095e0e0e1f940425f577be5048e98f
It's not really needed right now from logic point of view, since we
reused NONE for that. But it makes logging and logic clearer, and will
make it easier if we decide to move it to FSMs at a later point in time.
Other state value_string names are also modified with its whitespace
removed since anyway we'd need to change them to match WAIT_RECONNECT
length. Let's drop the space because imho it's not that useful and
anyway if we move to FSMs at some point then we won't have them anyway.
Change-Id: I7b9a6da87081c418b0d14bab5f34369c5eca6fe8
With previous state, osmo_stream_cli_close() could be called from
osmo_stream_cli_open()(), and in that case state was kept as NONE while
ending up with an associated fd being registered in the select loop.
As a result, osmo_stream_cli_fd_cb() could be called while in state
NONE, which was not expected and would simply return without modifying
fd state flags, causing it to be called again and again.
Related: OS#4378
Change-Id: Ie3342f882893a71ad1538c17ad9ee9fa4433eaa4
It should not happen anyway because no fd should be active if state is
NONE, but still it's an extra check.
Change-Id: I6d58762b7d10078eb8d0981c13d35cb6f85cfe86
Similar to what we do in libosmocore already, we want to
deterministically enable or disable support for the feature without
having into account if the system has a libsctp. If libsctp is missing
and support is enabled, then fail.
Extra checks are also added:
* Check netinet/sctp.h header
* Check libosmocore was built with libsctp support (API
osmo_sock_init2_multiaddr() we require).
* In stream.c make sure it can be built without HAVE_LIBSCTP, and that
set_addrs() fails for more than 1 address (since that feature is only
supported through osmo_sock_init2_multiaddrs()).
Change-Id: I4b3e1f1894f13ac1175a71a5139c02a2633be26d
This API will be later used to set multiple addresses for SCTP sockets.
Depends: libosmocore.git Ic8681d9e093216c99c6bca4be81c31ef83688ed1
Related: OS#3608
Change-Id: I09f0d989f2309abdeb304fe570355abed2cd107b
This API will be later used to set multiple addresses for SCTP sockets.
Depends: libosmocore.git Ic8681d9e093216c99c6bca4be81c31ef83688ed1
Related: OS#3608
Change-Id: I0fe62f518e195db4e34f3b0ad1762bb57ba9d92a
If messages are sent using osmo_stream_cli_send() while the stream
is still (re)connecting, they won't have a chance to be sent until the
stream is connected, and hence they are queued until
CONNECTING->CONNECTED is done. However, at that time
(osmo_stream_cli_fd_cb), the WRITE flag was dropped unconditionally,
which meant already queued packets didn't have the opportunity to be
sent by the same callback until first message is enqueued and WRITE flag
is set (again by osmo_stream_cli_send()).
Let's make them be sent as soon as possible once the connection is
available.
Related: OS#4188
Change-Id: I289495f9aad6389c5f2623fb072d676235b7d24c
This supposed to be variant of osmo_stream_cli_open() with explicit
control over reconnection logic but it's plain broken: doxygen docs
contradict the code, actual reconnection logic is affected by timeout
parameter directly which is set in different function.
It seems like we haven't been affected by this so far because we always
use it in auto-reconnection mode which is triggered by default due to
positive reconnection timeout value (5 sec) automatically used in the
absense of explicitly set timeout.
Looking at commit history, this function already been source of
confusion in the past. Instead of trying to fix this mess, let's just
deprecate it entirely and properly document use of
osmo_stream_cli_set_reconnect_timeout() to control reconnection logic.
The only known user is libosmo-sccp which won't use it as of
0a93a683f3cb8e5977eb4a666ab207db6e7d7af9 commit.
Change-Id: Id988ed0274b363db049f59cbf6a193727c8c3c8a
Previously closing the client did not alter its state, so we might
end-up with a client without any file descriptors, but being in state
STREAM_CLI_STATE_CONNECTED. Fix this inconsistency by setting
appropriate state.
Related issue is that reconnect function, which is always (at least in
the library and examples) called when some problem with the connection
is detected, closed the connection only after checking whether
reconnection is enabled. This might result in another inconsistency
fixed in this patch by moving the check below connection cleanup.
While at it, also move connection close logging to appropriate place:
it's confusing to see logs about connection being closed while in
reality it wasn't even established.
Change-Id: If41ed60bd625488c283d1e8a2b078e640f04c78e
Add functions to get the description of a server link or client
connection which examine data on corresponding socket.
Those functions use static buffers and intended for single use in
log/printf statements as illustarted by corresponding example changes.
Change-Id: If9a8e211da85956781479862a63c4fc6e53ed6be
Introduce logging macro wrapper to properly log current client state and
function to aid in debugging.
Change-Id: Ie22a80dcec95998cce0b25053fdf74f23eab6e53
While we are processing a read event, the connection's
callback might free the connection. Check for this and don't
attempt to process further events on an already freed connection.
Change-Id: I0a9c7d8e3263c73440f7084dbb1792a4ca5038f0
Related: OS#3685
Depends: g#11704 (for libosmo-sccp)
When establishing a client-side stream connection via libosmo-netif,
we must using non-blocking connect if we want to avoid blocking/stalling
the entire process. The libosmocore socket API provides the
OSMO_SOCK_F_NONBLOCK flag for this. Make use of it!
Change-Id: I9bfcb39b5801a36ef32ca0d1f3eb8236687d7ed6
Related: OS#3383
Documentation of osmo_sock_init2 doesn't provide information of any
specific value of errno set/expected after running the function. It is
incorrect to expect a specific value of errno and looking at the
implementation it is actually not a good idea to check it.
If reconnect flag is set, let's reconnect always instead of looking at
errno to decide.
Change-Id: I25b33f4cdc496ae31ff240d445b9b2805091845c
Pau Espin
Max
Harald Welte
Vadim Yanitskiy
Sylvain Munaut