Converting from osmo_fd to osmo_io allows us to switch to the new
io_uring backend and benefit from related performance benefits.
In a benchmark running 200 concurrent bi-directional voice calls with
GSM-EFR codec, I am observing:
* the code before this patch uses 40..42% of a single core on a
Ryzen 5950X at 200 calls (=> 200 endpoints with each two connections)
* no increase in CPU utilization before/after this patch, i.e. the
osmo_io overhead for the osmo_fd backend is insignificant compared
to the direct osmo_fd mode before
* an almost exactly 50% reduction of CPU utilization when running the
same osmo-mgw build with LIBOSMO_IO_BACKEND=IO_URING - top shows
19..21% for the same workload instead of 40..42% with the OSMO_FD
default backend.
* An increase of about 4 Megabytes in both RSS and VIRT size when
enabling the OSMO_IO backend. This is likely the memory-mapped rings.
No memory leakage is observed when using either of the backends.
Change-Id: I8471960d5d8088a70cf105f2f40dfa5d5458169a
The old approach was: rtp_data_net() reads a msgb from the incomging
socket, calls through whatever function chain and in the end free's it.
So none of the intermediate functions was permitted to take msgb
ownership.
This was a good choice as all processing would happen synchronously,
up to the point where that msgb was written on the output RTP socket.
Let's change this from passing msgb ownership throug the whole call
chain, through rx_rtp() to the various *_dispatch_rtp() functions.
This is required for upcoming migration to osmo_io, as in that case the
write (sendto) calls are asynchronous and hence msgb ownership needs
to be transferred.
Change-Id: I6a331f3c6b2eb51ea312ac6ef8c357185ddb79cf
the processing call-back is working with a raw buffer + length,
while we actually work with struct msgb. Let's simply pass the msgb
into the call-back, and the call-back can then do what they want with
the contents of that msgb.
Change-Id: I002624f9008726e3d754d48aa2282c38e3b42953
The existing support preparing the mgw for transcoding (which doesn't exist)
has some kind of method where the transcoding function might be called
multiple times in a row. However, as it is not used, it is not entirely
clear how it was intended to work. Let's remove this unused looping
feature which makes it hard to understand how upcoming osmo_io should
deal with it.
Change-Id: Ie1a629fd31c5ab806fc929d1e6b279c4be5b8246
The entire mgw has no transcoding support. So printing that message is
useless to begin with. And printing it for *every RTP packet* is even
more useless. Let's remove it.
Change-Id: If0ee2607404afc3a00665a5cf22a9e0eb62eb476
Before this patch, when an CRCX+MDCX wants to set a codec list that has
no match with the codecs for the other conn of that same endpoint,
osmo-mgw returns an MGCP "FAIL" response.
When a client wants to change the codec, it has to do that one RTP port
at a time. So osmo-mgw *must* allow to configure an MGCP conn with a
codec choice that mismatches the other conn.
This is crucial to allow codec negotiation in osmo-msc: if MO has
already assigned a specific codec, and later wants to re-assign to the
codec that MT has chosen, the codec needs to be changed at osmo-mgw.
This patch is the minimal fix required to get re-assignment to a
different codec to work (via osmo-msc). There is more work to be done
about this bit of code in osmo-mgw, but keep that to a separate patch.
In detail, before this patch, we fail both
- when a side has no codecs,
- or when there is no single match between codecs of the two sides of
the endpoint.
Remove only the second condition; after this patch, still fail when a
side has no codecs -- this allows mgcp_test.c to still pass.
Related: OS#6293
Related: osmo-msc I8760feaa8598047369ef8c3ab2673013bac8ac8a
Change-Id: I3d1163fe622bdd7dc42a485f796072524ab39db9
The new osmo_io framework means that we can [optionally] make use
of the io_uring backend, which greatly reduces the syscall load
compared to the legacy osmo_wqueue + osmo_select_main + read/write.
We only use features already present in the intiial osmo_io support
of libosmocore 1.9.0, so no entry in TODO-RELEASE is needed.
Closes: OS#5754
Related: OS#5755
Change-Id: I766224da4691695c023d4d08d042a4bbeba05e47
We already did a lookup from conn_src[i] and found a matching
codec_conn_dst, no need to do another reverse lookup to end up at the
same conn_src[i] codec.
Change-Id: Iecc7f22c551fd17b23db434fdb177266407d2621
There is considerable code complexity in place for this ancient hack.
It dates back to 5ea1bc77a3
"
mgcp: Allow to freely control the a=fmtp line for experiments
In case of AMR one can specify the available codecs out-of-band. Allow
to configure this line statically in the configuration file.
"
Looking in mgcp_test.c output, the fmtp-extra tests do not even make
sense: they result in fmtp for pt=126 being added, even though there is
no payload type 126 listed in the SDP...
Related: OS#6313
Change-Id: Icee0cd1f5a751fa760d5a9deca29089e78e7eb93
It seems to be a remnant from early openbsc_mgcp. There is only the
default implementation for this callback and it simply returns two
pointers. Simplify that.
Change-Id: I18dfd44c931540caf4ac360c08ed10e5f65b2165
Removing the duality of codecs[] and ptmap[] in structs mgcp_msg,
mgcp_response and mgcp_conn_peer has removed the need to "map" from
codec type enum to payload type number. They are stored together now.
Remove functions that are no longer used.
None of our osmocom users of libosmo-mgcp-client call these functions.
Change-Id: I84e5285831397c992af59deee12dea8458d16cc6
So far we allow only MGCP_MAX_CODECS-1 entries, because the parsing exit
condition hits only after the array size check. Instead, check the array
size a bit later, just before actually adding a valid entry.
This is verified to work as expected in upcoming patch
I842ce65a9a70f313570857b7df53727cc572b9e6 that adds a new
mgcp_client_test.c section for this.
Change-Id: I9a28da85e437f118026ea71a5a708e5758fff623
codecs[] is an array of enum osmo_mgcp_codecs.
ptmap[] is an array of { enum osmo_mgcp_codecs, unsigned int ptmap }.
MGCP lists first a bunch of payload type numbers and then specifies them
again for details, like the numbers 112, 96, 3 in this example:
m=audio <port> RTP/AVP 112 96 3
a=rtpmap:112 AMR/8000
a=rtpmap:96 VND.3GPP.IUFP/16000
a=rtpmap:3 GSM-FR/8000
So far we keep these lists in two separate arrays:
- codecs[], codecs_len stores the 'm=audio' list
- ptmap[], ptmap_len stores the 'a=rtpmap' list (and may omit some
elements present in codecs[])
This applies to both struct mgcp_response and struct mgcp_msg.
These are semantically identical, and the separation leads to checks,
conversions and dear hopes of correct ordering.
So let's keep only one list with both codec and payload type number in
it. The 'm=audio' list establishes the order of the pt numbers, and the
'a=rtpmap' list adds codec information to the established entries.
In the internal API structs mgcp_msg and mgcp_response, just drop the
codecs[] entirely.
In public API struct mgcp_conn_peer, keep the codecs[] array, mark it
deprecated, and provide a backwards compat conversion: if any caller
invokes mgcp_conn_create() or mgcp_conn_modify() with codecs[] still
present in the verb_info arg, then move codecs[] entries over to the
ptmap[] array in a sensible way.
(BTW, even mgcp_conn_create() and mgcp_conn_modify() are never called
from outside of libosmo-mgcp-client in any of our osmo-cni programs;
users call osmo_mgcpc_ep_ci_add() and osmo_mgcpc_ep_ci_request(), which
in turn may pass user-provided codecs[] lists on to mgcp_conn_create() or
mgcp_conn_modify().)
Tests for parsing the MGCP response are mostly missing. They will be
added in upcoming patch I842ce65a9a70f313570857b7df53727cc572b9e6,
because they will be using only the new ptmap API.
Related: OS#6171
Change-Id: I798e02c6663376d3d52f4a74fc4b32411ce95bed
We need to be able to terminate the received string in case it was
not nul-terminated by the sender (see mgcp_msg_terminate_nul()).
Change-Id: Icc878af7f671213bb516af62cb601914d86ff808
Fixes: CID#272990
There is an upcoming modification of structs mgcp_msg and mgcp_response.
When it is public API, it needs to be kept backwards compatible.
But no libosmo-mgcp-client caller (in the osmocom-cni world) has used
these structs in years and years. Everyone should use the higher level
osmo_mgcpc_* API only.
Move this legacy API to a private header, so we no longer need to worry
about compatibility there.
Related: OS#6171
Related: I798e02c6663376d3d52f4a74fc4b32411ce95bed
Change-Id: I6d8bdda5c2ffa236e94a0b6111de4dbfac923187
end.codec may be NULL, depending on whether a single matching codec
could be found or not, i.e. based on external input. So let's not crash
if an RTP packet arrives while end.codec == NULL.
Change-Id: I9bfb55a343b3f2b1459e0aba4ee71a6133b992b3
A client may know the IP address during CRCX but not yet the port, or it
may simply want to hint an initial IP address so that the MGW can better
guess when allocating a local IP address.
Related: SYS#6657
Change-Id: I30165dbac5e484011d0acf46af36f105954a501d
Do not refuse IuUP Initialization messages coming in on an RTP port if
the remote port is not yet known.
If an IUFP conn is not yet configured (pre-Initialization), allow rx
from any address or port.
An osmo-mgw client (eg. osmo-hnbgw) may wish to initially set a remote
IP address as a hint during CRCX, hence the IP address may already be
set while the port may be unset.
Related: SYS#6657
Change-Id: Idd833997abce46886e9664505b2776fa5dadc8db
We have proper IuUP support and everything about this legacy hack should
be purged.
The purpose of this function is to validate that RTP is coming from the
expected address and port. To allow that legacy IuUP hack, which is no
longer needed, we punched a hole into this validation, by adding this
special case for loopback mode (suddenly we don't care who or what sends
RTP and bounce it back to anyone). So let's get rid of this hole that
was only needed for very early 3G voice hacking.
Instead, we permit RTP for IuUP Initialization regardless of the RTP
loopback/send/recv mode since I6c365559a7bd197349f0ea99f7a13b56a4bb580b
Related: SYS#6657
Change-Id: I158dd046fdfcb10392cde3de8cc88dd095a05b40
Do not refuse IuUP Initialization messages coming in on an RTP port if
the remote IP address is not yet known.
If an IUFP conn is not yet configured (pre-Initialization), allow rx
from any remote address if the remote IP address is not yet known.
If we refuse the IuUP Initialization, a 3G RNC may fail to set up a RAB.
We will know the remote address only *after* assigning a RAB succeeded.
So the IuUP Initialization must be allowed before knowing all addresses.
At the time of writing, CRCX for IUFP are sent to osmo-mgw in either
LOOPBACK or in RECVONLY mode:
- current osmo-msc: recvonly
- osmo-msc <= v1.10.0: loopback
- osmo-hnbgw: loopback
IuUP Initialization should work regardless of that.
See also next patch I158dd046fdfcb10392cde3de8cc88dd095a05b40
IuUP is one layer below the loopback/send/recv decision for RTP; IuUP is
always terminated at the MGW, while the AMR payload carries through.
Related: alternative patch Idd833997abce46886e9664505b2776fa5dadc8db
Related: SYS#6657
Change-Id: I6c365559a7bd197349f0ea99f7a13b56a4bb580b
A user reports crashes of osmo-bsc upon EV_MDCX. It turns out that there
is a lot of error reporting and a distinct possibility to get a NULL
return value because of external input. Terminate the FSM instead.
FSM termination is the proper way to report a bad error, it signals the
parent_term_evt to the FSM parent, which will then be able to act on the
failed MGCP operation.
Related: SYS#6632
Change-Id: Ia5d8a9aff565399a85a5b116d7029fedcab234e0
Before:
CONN: (1226/rtp, id:0xD94316AD, ip:127.0.0.2, rtp:2344 rtcp:2345)
After:
CONN: (1226/rtp C:D94316AD r=127.0.0.2:2344<->l=127.0.0.1:4002)
While changing that string, also include these changes for consistency
and readability:
- use the same r:...<->l:... format as osmo_sock_get_name().
- Instead of 'id:0x' use the actual MGCP format 'C: 9B686BE3'.
- drop the commas
- drop RTCP port: it is always RTP+1 and always an odd number.
Rationale:
The CONN pairs associated with each endpoint show remote RTP
ports. When osmo-mgw is being used by both BSC and MSC, one
side of the pair is showing the internal loop connection inside
osmo-mgw, while my intuition suggested this connection pair
is showing me the RTP port tuple of a single RTP stream. Adding
the local port to the display makes it more clear, IMHO.
Seeing the local port can also help to correlate the MGW vty
dump with a capture of RTP.
Implementation:
I first tried directly using osmo_sock_get_name_buf() on
conn->u.rtp.end.rtp.fd, but that might hide already known information
when the fd is not actively used yet (before SDP): the local address and
port would then be shown from the fd, not from
conn->u.rtp.end.local_addr/_port == hidden before the fd is set up.
Patch-By: whytek, nhofmeyr
Change-Id: Ib89a6779e1d68c6600f00699d4303f6c0ee07132
The mgcp client may first configure the connection to use RTP-AMR, but
after setting up another call leg may find out that both legs are IuUP
and hence want to forward the IuUP between the 2 connections instead.
In that case, an MDCX with codec VND.3GPP.IUFP would be set.
Until now, osmo-mgw didn't take that scenario into account, and it was
only upgrading the rtp conn to iuup internally during CRCX.
As a result, in the mentioned scenario osmo-mgw would continue to
output RTP instead of IuUP after the MDCX with VND.3GPP.IUFP, which is
wrong.
Related: SYS#6578
Change-Id: Ic94bf90f54d8ba3e65a2cd52734867847f3a60c2
For each RTP packet that is received from a connection, the mode is
checked whether receiving is allowed or not. If not it is discarded.
In case of "confecho" mode, the RTP is also sent by the receiving
connection.
Then a loop is used to send RTP to all sending endpoints except the one
that received the packet.
Because we have a loop that allows to have 1..n connections, we have no
maximum number of allowed connections anymore.
Change-Id: Ic99a55ab5a3a6170e940403fadd52697e99f2f3a
Related: OS#4853
The new mode "confecho" is similar to "sendrecv", except that it also
echoes back RTP towards the sender. This is required for voice group or
broadcast calls. Talker and listeners use the same timeslot, so that
audio must be echoed from the talker to the listeners.
It is different from "loopback", because a loopback only echoes back RTP
towards the sender, but does not forward audio through the endpoint to
the other connections. Also it does not forward RTP from senders of
other connections.
The current implementation of MGW does not support transcoding and
mixing. This means that a sending connection must not send RTP that has
been received by multiple receiving connections. The application that
uses the MGW has to set the connection modes, so that only one
connection receives RTP in case of a conference.
Change-Id: I0639c663e119d85bef1010c7aa45e2f133a9daf0
Related: OS#4853
Instead of calling strlen() for every loop iteration, just use strchr()
to find the position of '/'.
Change-Id: Ifc7302b6c5f9288a622e33c3e8b5fe0e7037dbdc
In the "L: a:" header, read the first codec name only up to the ";"
separator, and ignore the rest.
According to RFC-2705, the "L: a:" header may include multiple codecs
like "GSM-EFR;GSM" in:
L: p:20, a:GSM-EFR;GSM, nt:IN
osmo-mgw can handle only a single codec here. Since recently, osmo-msc
is our first client that may actually send more than one codec. This
uncovered a bug that leads to failing voice calls:
* osmo-mgw parses the entire list "GSM-EFR;GSM" as a single codec name,
* puts that into the ptmap without scrutiny,
* and even sends it back in the OK response, in the *SDP* part, as a
single "GSM-EFR;GSM" codec entry.
We do not care very much about the "a:" codec list, because we always
establish codecs via SDP later. So all we need to fix this is: parse the
first codec done correctly, and ignore the rest.
Related: OS#6081
Change-Id: I0342e85b32ed89f3a1fdb6131c3c8ded8f47a455
Abort with error log when map_str_to_codec is -1, instead of writing it
as unsigned integer into ptmap[i].codec.
Related: OS#6074
Change-Id: I08b91c849d810fe3cdb72c0f6f2a558b7377deab
This way if keepalive becomes disabled for an MGW, it can be selected
again (otherwise it would become non-selectable forever, and we already
have a "blocked" config for that).
Fixes: 563386e8bb
Related: SYS#6481
Change-Id: I263d23885a1a967f844033f9810b96691b8e3b30
In the previous logic, if "keepalive request-interval" was disabled
(value 0, default), then if the client was configured to send a DLCX on
startup it would end up in state UP or DOWN depending on whether the MGW
answered to that request.
As a result, an MGW that wouldn't answer would be left forever in DOWN
state since it would have been selected no more and there would be a
keepalive configured to re-mark it as UP.
Fixes: 563386e8bb
Related: SYS#6481
Change-Id: I290f7436f48418ee25179951359c76208796e279
This way the user ends up picking a working MGW instance instead of one
which is not reachable around the time.
Related: SYS#6481
Change-Id: Ia3f451d3cd97851f65074408812b1ddc68f67056
The `keepalive` feature in libosmo-mgcp-client allows scheduling periodical
queries on the MGCP layer in order to make sure it is reachable and hence
obtain information on the state of the MGCP link.
This patch only uses it to print the status on the VTY, but it will be used
too in a follow-up commit by the MGW Pool when picking an MGW from the pool:
MGWs whose link is considered to be DOWN are skipped.
The feature consists of:
- A `keepalive request-interval` which will trigger a transmission of an MGCP
AuditEndpoint command targeting endpoint with name `keepalive request-endpoint`.
This interval is updated every time any message is transmitted in the MGCP
link, meaning the MGCP AuditEndpoint message is only triggered if no message
has been transmitted since `keepalive request-interval` seconds ago.
- A `keepalive timeout` considering the MGW to be non-reachable (link DOWN) if
no message is received over that amount of time.
The `keepalive` parameters are to be preferrably configured so that
"keepalive request-interval" * 2 < "keepalive timeout".
Example VTY configuration of `keepalive` feature in libosmo-mgcp-client:
----
mgw 0
...
keepalive request-interval 20 <1>
keepalive request-endpoint null <2>
keepalive timeout 50 <3>
----
<1> Transmit an MGCP AuditEndpoint message to the MGW if no message has been
sent to it over last 10 seconds
<2> The MGCP AuditEndpoint targets the `null` endpoint. This is a special
endpoint available at OsmoMGW for those purposes, but any available
endpoint can be configured and used instead.
<3> Consider the MGCP link to be DOWN if no message is received from the
MGW over the last 50 seconds
NOTE: The `keepalive` feature is disabled by default, and must be explicitly
configured in order to enable it.
Related: SYS#6481
Change-Id: I3dc74c78548d017f272da863d5282dc5e0020ca3
This is a special endpoint which can always be audited. This is useful
for clients who wish to submit requests to osmo-mgw periodically to find
out whether the MGW is still reachable. This endpoint will be used by
libomso-mgcp-client as default target endpoint to implement such
feature.
This "null" Endpoint is osmo-mgw specific, not described in MGCP specs.
Related: SYS#6481
Change-Id: Ia409b16e9211e6261e2e0f21288544289d6f3733
This is a preparation commit for follow-up one, where some of these
functions need to be used in other static functions.
Related: SYS#6481
Change-Id: I3a00d8c47ec773390d9626364c4c75ca579f1508
So far, the users of the old non-pooled API were in charge of allocating
the struct mgcp_client_conf by themselves, then init them using
mgcp_client_conf_init(). This causes a major problem, since it makes it
difficult to extend the struct mgcp_client_conf structure to add new
features, which may happen frequently.
The MGW pool API doesn't have this problem, because the struct
mgcp_client_conf is allocated as parts/fields of private structs defined
in internal headers. Only pointers to it are used in public headers.
Since it still has to internally initialize the conf fields, we still
need the API to initialize it internally, and hence why is it marked as
DEPRECTED_OUTSIDE instead of DEPRECATED.
While some programs already moved to the new MGW pool infrastructure,
they still use the old APIs to accomodate for old config files in order
to be back-compatible, hence most users of libosmo-mgcp-client are
affected.
Introduce an API to allocate the conf struct internally, which, while
still breaking the ABI, allows for a more relaxed update path where it's
possible to extend the struct mgcp_client_conf at the end.
Eventually the non pooled API should be gone and the struct
mgcp_client_conf can then be moved to a private header, but for now
let's add this small feature to avoid major ABI breakage.
Change-Id: Iba0853ed099a32cf1dde78c17e1b34343db41cfc
Harald Welte
Neels Hofmeyr
Vadim Yanitskiy
Pau Espin
Keith Whyte
Oliver Smith
Andreas Eversberg