ENOSPC is used with non-volatile (disk) storage, while ENOMEM is
customarily used for RAM allocation failures.
Change-Id: Ia4c16d8278dc30c7cc69169b18428cda5272738d
Avoid code duplication between three different drivers by sharing
the "core" of the bind initialization in a new, shared ns2_bind_alloc().
Change-Id: I535fc68e94fcd695de827dd922706adc1c5a2cb7
When the FR code decides to drop a packet (ENOBUFS from the AF_PACKET
socket for non-signaling packet), let's report that back via the
frame_relay code into the generic NS2 code.
This way the generic NS2 code always knows if a packet was actually
successfully transmitted, or if it was dropped for some reason.
Change-Id: I4bb517fd04af69dbe6da628b132d57994ab3e5a4
AF_PACKET sockets cannot be written-to using select(), as they
will always return "writable" but then still fail with ENOBUFS.
This also means that we cannot use osmo_wqueue() as it assumes
that a writable socket can actually be written to.
As there's no way to figure out when exactly we can again perform
a successful write, we have no other option but to start a timer
and re-try at a later time.
We will scale that timer based on the estimated duration of transmission
for the just-failed PDU on the line rate of a 31TS E1 Line.
Furthermore, with this patch, we stop queueing anything but signaling
traffic (NS-BVCI=0) or Q.933 LMI. User data (NS-BVCI != 0) will
instead be dropped in order to avoid buffer-bloat.
Change-Id: I4f79a246236c94175ceffa5f3f556c8931c6bc62
Closes: OS#4995
All public enum should have the prefix GPRS_NS2_.
API change which must be synchronized with osmo-pcu,
osmo-gbproxy, osmo-sgsn.
Change-Id: I548ff12f7277cbb7e1a630a3dc02b738ce89be72
All functions which are exposed by gprs_ns2_internal.h should not contain
the public prefix gprs_. Internal function should only contain ns2_ prefix.
Change-Id: Icecc5a918902cd10efac72bbac20780d39aab272
Let's avoid open-coding the printing of log context and rather rely on
log macros to prefix each log line with the relevant context. This
helps log readability, log post processing whether by grep or more
sophisticated tools.
Change-Id: I946c0e77686d91efc5afb62031e1ac1033a9a586
Usually talloc_free() and other free functions in osmocom allows
to be called with NULL which is then ignored.
Change-Id: If7b0c6916a29d4611d0a40c388414076eb83e6b5
Add a OSMO_ASSERT to all bind calls which doesn't
check if the bind is from the expected type.
The only exception is rx and tx functions (hot path).
Change-Id: Ia4f8932263c60618c7f0dfc32d50ba5a8d57602b
gprs_sn2_fr_connect2() is the same as gprs_ns2_fr_connect() with
the lookup-and-create-on-demand of the NSE first.
Rather than copy+paste, they should simply invoke each other.
Change-Id: If835bf138f213e7f58205018e7efe3ecb772c624
When a dahdi device hasn't been set up yet, ioctl IF_GET_PROTO fails
with invalid argument.
Also fix the device check to skip ioctl's if the device is also in the
correct state.
Change-Id: I398d056546e35465a2944e1b4a86a8c93b3e5f7a
A hdlc can be used in different modes. Also a FR device can be used
with lmi and certain settings as without it.
ns2 will use FR with no lmi in the kernel.
Related: SYS#5169
Change-Id: I04786d2b864860b08c2e1afdb199470f4b80cc3b
This use of strncpy() fails to account for the terminating nul
character. Use OSMO_STRLCPY_ARRAY() instead.
(Interestingly my compiler doesn't complain about this one, though it
failed on another similar use of strncpy().)
Change-Id: Id53e940c7a39ab154966548f4173a179c5bc9151
My gcc (Debian 8.3.0-6) 8.3.0 refuses to build this strncpy() use: it
issues the buffer length as n and thus potentially fails to account for
the terminating nul. The line after that fixes the problem, so it's not
an actual bug. Anyway, we have a policy to never use strncpy(), and have
osmo_strlcpy() and OSMO_STRLCPY_ARRAY() for this.
This strncpy() was introduced last month during first addition of
gprs_ns2_fr.c:
commit 841817ec52
ns2: add support for frame relay
Change-Id Id3b49f93d33c271f77cd9c9db03cde6b727a4d30
Change-Id: I494a6fb7ccd7938a39e8956f73ec4282da38d7fb
When a frame relay interface doesn't exist gprs_ns2_fr_bind() would
detect this but still return a success.
Change-Id: I815b6ef5c3df780ac94461a05975a2b70898b01e
Add gprs_ns2_fr_connect2() and change gprs_ns2_fr_connect() to
be similar to gprs_ns2_ip_connect() and gprs_ns2_connect2().
This is an API break but there wasn't yet a release with NS2.
Change-Id: I4e1374b0e979b3293302c5ed46a91a58f3a5a916
Every bind will have a unique name. Add a name argument
to all bind creating functions and require them to be unique.
This is an API break but there wasn't yet a release with NS2.
Change-Id: I8f1d66b7b3b12da12db8b5e6bd08c1beff085b3e
A NS dialect describes how the NS Entity interacts with
different virtual circuits. E.g. ipaccess use reset/block on udp
and is a dynamic connection.
A single NS Entity can only support one dialect. This can be later
used to protect a NS Entity against dynamic NS virtual circuits of a
different type.
It further allows a bind to support multiple dialects at the same time.
Change-Id: Ia118bb6f994845d84db09de7a94856f5ca573404
When opening the socket, use ETH_P_HLDC to restrict the socket to
packet received on HLDC interfaces. This avoids packets from random
other (ethernet, ...) interfaces to appear before we can bind()
it to the actual hdlc-net-device we're interested in.
We still are racing against other HLDC net-devices, but those have
lower PPS and throughput ratese as 1G/10G or even higher speed ethernet
devices that might exist on the same machine.
Change-Id: I6a556e6e2d012c17a2777cc8b30fed0f318db178
An AF_PACKET socket will immediately receive packets of _all_ interfaces
until it is bound to one specific interface. This introduces a race
condition between the socket() and the bind() syscall.
Let's use the ifindex passed for each packet in recvmsg() to drop
any packets received for other interfaces.
Change-Id: I8f708ba4f9b7f76525acce17b24a8f7b125a1c1c
Related: SYS#5245
It's not needed because net/if.h already include
the required parts. Furthermore the linux/if.h generates
a compiler error in combination with net/if.h on older systems
(e.g. debian jessie)
Change-Id: I53650e9c55bb9dd98ba60269025e72673e9f82c1
We use the newly-introduced libmnl integration of libosmocore in order
to receive netlink events from the kernel on link state changes.
If one of "our" interfaces changed link state, we report this in the log
and also store it within the "bind".
Change-Id: I779556991bfc88b7751b2be17bb81c329cfb9e01
It's not needed because net/if.h already include
the required parts. Furthermore the linux/if.h generates
a compiler error in combination with net/if.h on older systems
(e.g. debian jessie)
Change-Id: Iccf931207d8bea19969fbfc4225f622a1a709972
Even it was in theory possible to mix NS-VC ll types within
a NSE. This is an unrealistic configuration.
Further more to select the correct load sharing mechanism
the NSE must know the correct link layer.
Change-Id: I18dfd40a2429cd61b7c4a3dad5f226c64296f7d8
Add support for frame relay over dahdi hdlc device.
It's supporting lmi by q933 and supports both
SGSN and BSS.
Change-Id: Id3b49f93d33c271f77cd9c9db03cde6b727a4d30
Harald Welte
Alexander Couzens
Vadim Yanitskiy