IFINDEX_DELETED is not necessary anymore as we moved from a global
list of interfaces to a list of interfaces per VRF.
This reverts commit 84361d615.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
This is an important optimization for users running FRR on systems with
a large number of interfaces (e.g. thousands of tunnels). Red-black
trees scale much better than sorted linked-lists and also store the
elements in an ordered way (contrary to hash tables).
This is a big patch but the interesting bits are all in lib/if.[ch].
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
size is not used for further parsing. Keep it updated but tell
to the compiler that we know it is not used just in case one
needs to extend the parsing somedays.
Signed-off-by: Vincent Jardin <vincent.jardin@6wind.com>
Certain compilers cannot recognize that rt is
actually being init'ed, but let's set it to
NULL 'till we get them updated.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
Doing a "ssh user@node 'ldpd -d'" was making the SSH session hang. In
the original OpenBSD's ldpd(8) daemon, the daemon function takes care
of connecting stdin/stdout/stderr to /dev/null. In the FRR port, this
only happens in the frr_run() function, after all children have been
forked. Ideally we could try to rearrange libfrr.c and ldpd.c in a way
that start_child() is called only after the parent connects the standard
I/O streams to /dev/null. But since this issue needs an immediate
fix, let's do this workaround for now. Note: even when running on the
foreground, all log messages from the child processes are sent to the
parent process, which then prints the messages to stdout/stderr and/or
to a log file.
Reported-by: Martin Winter <mwinter@opensourcerouting.org>
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
The -n option was not aligned with the other ones:
% ldpd --help
[snip]
-A, --vty_addr Set vty's bind address
-P, --vty_port Set vty's port number
--ctl_socket Override ctl socket path
-n, --instance Instance id
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
Once ldpd allocated label 48 for a given FEC, all subsequent requests
for a new label would return the same value (48). The problem is that
we were left shifting an uint32_t value up to 64 times, losing important
information.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
Some differences compared to the old API:
* Now the redistributed routes are sent using address-family
independent messages (ZEBRA_REDISTRIBUTE_ROUTE_ADD and
ZEBRA_REDISTRIBUTE_ROUTE_DEL). This allows us to unify the ipv4/ipv6
zclient callbacks in the client daemons and thus remove a lot of
duplicate code;
* Now zebra sends all nexthops of the redistributed routes to the client
daemons, not only the first one. This shouldn't have any noticeable
performance implications and will allow us to remove an ugly exception
we had for ldpd (which needs to know all nexthops of the redistributed
routes). The other client daemons can simply ignore the nexthops if
they want or consult just the first one (e.g. ospfd/ospf6d/ripd/ripngd).
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
The simple queue implementation in OpenBSD and FreeBSD are called diferently,
standardize in the use of the FreeBSD version and map the missing names only
if we compile on OpenBSD.
Signed-off-by: Jorge Boncompte <jbonor@gmail.com>
Start creating a counterpart to frr_init and frr_late_init.
Unfortunately, some daemons don't do any exit handling, this doesn't
change that just yet.
Signed-off-by: David Lamparter <equinox@opensourcerouting.org>
- VTY_NEWLINE is out
- thread API changed with the extra arg
- struct rib got renamed to struct route_entry
- MPLS_NO_LABEL was removed
- RB-tree implementation has some extra args
Signed-off-by: David Lamparter <equinox@opensourcerouting.org>
In some scenarios, it's possible to send a Label Withdraw to a neighbor
and not receive a corresponding Label Release right away. This can happen
during reconvergence after a network failure or configuration change.
When this happens, the list of upstream mappings of a given FEC might
not be empty even after sending a Label Withdraw to all neighbors. This
situation holds until all neighbors either send a Label Release or are
torn down (e.g. keepalive timeout).
With that said, we shouldn't check for 'RB_EMPTY(&fn->upstream)'
in lde_kernel_update() because it can prevent ldpd from sending label
mappings in such circumstances. This check was introduced to avoid sending
the same label mapping more than once to the same neighbor, but we need
to remove this optimization for now until we find a better solution (which
probably involves refactoring the whole zebra<->ldpd communication).
While here, add a new debug message in lde_send_labelmapping() which
can aid in troubleshooting label problems in the future.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
If we receive a notification from zebra indicating that the installation
of a pseudowire has failed (e.g. no reachability), send a PW Status
notification to the remote peer (or a Label Withdraw if the remote peer
doesn't support the PW Status TLV).
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
Base framework for supporting MPLS pseudowires in FRR.
A consistent zserv interface is provided so that any client daemon
(e.g. ldpd, bgpd) can install/uninstall pseudowires in a standard
way. Static pseudowires can also be implemented by using the same
interface.
When zebra receives a request to install a pseudowire and the installation
in the kernel or hardware fails, a notification is sent back to the
client daemon and a new install attempt is made every 60 seconds (until
it succeeds).
Support for external dataplanes is provided by the use of hooks to
install/uninstall pseudowires.
Signed-off-by: ßingen <bingen@voltanet.io>
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
We were assuming that a neighbor can be deleted only when all of its
adjacencies are dead. This is not the case for dual-stack neighbors. If
the transport-preference is IPv4 and all adjacencies are IPv6 (or
vice-versa), then it should be deleted and everything cleaned-up
accordingly.
Bug exposed by the new RB tree implementation on master, but the fix
also applies to stable/3.0.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
This adds "@tcp" as new choice on the -z option present in zebra and the
protocol daemons. The --enable-tcp-zebra option on configure is no
longer needed, both UNIX and TCP socket support is always available.
Note that @tcp should not be used by default (e.g. in an init script),
and --enable-tcp-zebra should never have been in any distro package
builds, because
**** TCP-ZEBRA IS A SECURITY PROBLEM ****
It allows arbitrary local users to mess with the routing table and
inject bogus data -- and also ZAPI is not designed to be robust against
attacks.
Signed-off-by: David Lamparter <equinox@opensourcerouting.org>
This also fixes a build problem where using #include
"ldpd/ldp_vty_cmds_clippy.c" results in the Makefile dependency tracking
having both
ldp_vty_cmds.c: ldp_vty_cmds_clippy.c
ldp_vty_cmds.c: ../ldpd/ldp_vty_cmds_clippy.c
(because, if it's included as "ldpd/..", it uses the "-I.." include path
in gcc, so the gcc -MD dependency output is "../ldpd/...")
... all of which causes the build to try to build it twice (at the same
time) and fail rather stupidly.
With a non-recursive build, the two paths are identical and everything
just works.
Signed-off-by: David Lamparter <equinox@opensourcerouting.org>
This splits off privs_preinit(), which does the lookups for user and
group IDs. This is so the init code can create state directories while
still running as root.
Signed-off-by: David Lamparter <equinox@opensourcerouting.org>
Split the "no" version of some commands into a different DEFUN so that
DEFUN_NOSH doesn't apply to them.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
Signed-off-by: Daniel Walton <dwalton@cumulusnetworks.com>
This allows frr-reload.py (or anything else that scripts via vtysh)
to know if the vtysh command worked or hit an error.
Add support for naming pthreads. Also, note that we don't have any
records yet if that's the case.
Signed-off-by: Quentin Young <qlyoung@cumulusnetworks.com>
The xml2cli.pl script was useful years ago when the vty code was very
rudimentary. This is not the case anymore, so convert all ldpd CLI
commands to use DEFUNs directly and get rid of the XML interface.
The benefits are:
* Consistency with the other daemons;
* One less build dependency (the LibXML perl module);
* Easier to add new commands.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
Switch the RB tree implementation completely to the new dlg@'s version
that uses pre-declared functions instead of macros for tree functions.
Original e-mail/diff:
https://marc.info/?l=openbsd-tech&m=147087487111068&w=2
Pros:
* Reduces the amount of code that the usage of those macros generate
* Allows the compiler to do a better compile-time check job
* Might have better i-cache utilization since the tree code is shared
Con:
* dlg@ benchmarks shows it has 'very slightly slower' insertions
* imported RB_* code must adapt the following calls:
RB_INIT(), RB_GENERATE(), RB_ROOT(), RB_EMPTY(), make compare
functions use 'const' (if not already) and maybe others.
Handling configuration changes from single-stack mode to dual-stack mode
(and vice-versa) is tricky. This patch attempts to solve all issues that
might happen on such circumstances.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
In OpenBSD pledge is a mitigation mechanism used to restrict the syscalls
a program can use, enforcing its correct behavior.
In this port of OpenBSD's ldpd(8), it's hard to run under the same
tight pledge promises because of libfrr and additional components we
introduced, like a zclient in lde. Since ldpd is already privsep'ed,
removing the pledge calls shouldn't be a big compromise security-wise.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
Bingen discovered a bug in the pseudowire control-word negotiation that
might happen when the "control-word exclude" command is used. Under some
very specific conditions, ldpd might ignore a PWID label mapping when
it shouldn't.
This patch removes a wrong optimization that was preventing ldpd to call
l2vpn_pw_reset() every time we change the configuration of a pseudowire.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
When the transport address is changed, all interfaces and targeted
neighbors are temporary disabled in the ldpe process until new sockets
bound to the new transport address are received from the parent.
This patch fixes a problem in which adjacencies weren't being removed
after the associated targeted neighbors were disabled. This was causing
ldpd not to set some MD5 sockoptions for new neighbors are thus preventing
MD5-protected sessions to come up after a change in the transport-address.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
It's doesn't make sense to enforce that a targeted-hello is received
on an LDP-enabled interface. It should be possible, for example, to use
LDP only to signal pseudowires and other another protocol (e.g. RSVP-TE)
to create end-to-end LSPs.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
Once we send a Label Withdraw, we can't send a Label Mapping for the
same FEC until we receive a Label Release from the peer. This is due to
some limitations in the LDP algorithms described in Appendix A. ("LDP
Label Distribution Procedures") of RFC 5036.
To workaround this issue, make it possible to schedule the sending of
a Label Mapping as soon as a Label Release is received for the same FEC.
The easiest way to test this patch is by typing the "label local advertise
explicit-null" command. ldpd will withdraw all null labels using a
Wildcard FEC and then send new Label Mappings as soon the corresponding
Label Releases are received.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
On unstable networks, routes can be lost and relearned very often. If
we deallocate the input label every time a route is lost and allocate
a new one when the route is relearned, a lot of changes are made in vain.
This patch introduces a logic in which labels are preserved for at least
five minutes before being deallocated by the LIB garbage collector. This
is consistent with what other implementations do.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
If we don't do this, we'll never trigger the backoff exponential timer
since it's impossible to distinguish between Initialization NAK's and
general errors.
Also:
* Implement some missing bits from RFC 5036;
* remove superfluous log message in session_shutdown()
(send_notification() logs that we're sending a fatal notification).
Regression introduced by commit 8819fc3.
Fixes the following ANVL LDP regressions: 6.19 and 6.21.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
This is necessary to guarantee that all log messages sent from the child
processes are received in the parent process right away.
Without this patch, when a child process calls fatal() or fatalx(),
the log messages don't make it to the parent because the child doesn't
have a chance to flush its buffers before exiting.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
When ldpd fails to start for some reason, like failing to create a pid
file, the child processes call their shutdown functions without being
completely initialized. This patch adds some protections to prevent a
segmentation fault on such circumstances.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
In order to have separate ASLR/cookies per process, ldpd calls exec()
in the child processes after fork() (this is also known as the fork+exec
model).
This is an important security feature but it makes the initialization
of the child processes a bit more complicated as they're not a copy of
the parent anymore, so all parameters given via command line are lost.
To solve this problem, we were creating an argv array by hand with all
necessary parameters and providing it to the exec() syscall. This works
but it's a very ugly solution. This patch introduces a different approach
to solve the problem: send an IMSG_INIT message to the child processes
with all parameters they need in order to initialize properly. This
makes adding additional initialization parameters much more convenient
and less error prone.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
Conflicts:
ldpd/ldpd.c
ldpd/ldpd.h
The log_warn() and log_warnx() functions indicate non-critical warnings
and errors, so use LOG_ERR instead of LOG_CRIT.
Keep using LOG_CRIT only in fatal() and fatalx() since these functions
indicate critical errors (when the program needs to exit).
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
This patch attempts to make the code easier to read by removing a
global variable and changing a few other things. Also, ldpd now calls
merge_config() only after reading the whole initial configuration at
startup, instead of doing that for every command in the configuration
file.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
Bingen discovered a bug in the pseudowire control-word negotiation that
might happen when the "control-word exclude" command is used. Under some
very specific conditions, ldpd might ignore a PWID label mapping when
it shouldn't.
This patch removes a wrong optimization that was preventing ldpd to call
l2vpn_pw_reset() every time we change the configuration of a pseudowire.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
We only needed to add/change the vrf callbacks when we initialize
the vrf subsystem. As such it is not necessary to handle the callbacks
in any other way than through the init function.
Signed-off-by: Donald Sharp <sharpd@cumulusnetworks.com>
The FSF's address changed, and we had a mixture of comment styles for
the GPL file header. (The style with * at the beginning won out with
580 to 141 in existing files.)
Note: I've intentionally left intact other "variations" of the copyright
header, e.g. whether it says "Zebra", "Quagga", "FRR", or nothing.
Signed-off-by: David Lamparter <equinox@opensourcerouting.org>
To avoid blocking zebra when it's acting as a proxy for an external
label manager.
Besides:
Fix get chunk reconnection. Socket was still being destroyed on failure,
so next attempt would never work.
Filter out unwanted messages in lm sync sock.
Until LDE client sends ZEBRA_LABEL_MANAGER_CONNECT message, zserv
doesn't know which kind of client it is, so it might enqueue unwanted
messages like interface add, interface up, etc. Changes in this commit
discard those messages in the client side in case they arrive before the
expected response.
Change function name for zclient_connect in label manager to avoid
confusion with zclient one.
Signed-off-by: ßingen <bingen@voltanet.io>
Pass pointer to pointer instead of assigning by return value. See
previous commit message.
To ensure that the behavior stays functionally correct, any assignments
with the result of a thread_add* function have been transformed to set
the pointer to null before passing it. These can be removed wherever the
pointer is known to already be null.
Signed-off-by: Quentin Young <qlyoung@cumulusnetworks.com>
The way thread.c is written, a caller who wishes to be able to cancel a
thread or avoid scheduling it twice must keep a reference to the thread.
Typically this is done with a long lived pointer whose value is checked
for null in order to know if the thread is currently scheduled. The
check-and-schedule idiom is so common that several wrapper macros in
thread.h existed solely to provide it.
This patch removes those macros and adds a new parameter to all
thread_add_* functions which is a pointer to the struct thread * to
store the result of a scheduling call. If the value passed is non-null,
the thread will only be scheduled if the value is null. This helps with
consistency.
A Coccinelle spatch has been used to transform code of the form:
if (t == NULL)
t = thread_add_* (...)
to the form
thread_add_* (..., &t)
The THREAD_ON macros have also been transformed to the underlying
thread.c calls.
Signed-off-by: Quentin Young <qlyoung@cumulusnetworks.com>
- If we use the option "--enable-tcp-zebra " to configure, the zclient will use the tcp socket instead of unix domain socket.
- The ldpd will pass the "-z path" options to `frr_opt` and that will execute zclient_serv_path_set to check the domain socket.
- Add the define to ldpd, if the "HAVE_TCP_ZEBRA" has been define, don't pass the "-z path" to `frr_opt`.
Signed-off-by: Hung-Weic Chiu <sppsorrg@gmail.com>
We shouldn't check the operational status of an interface in ldpd if
it's configured with "no link-detect" in zebra. That's what all the
other routing daemons do.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
When the transport address is changed, all interfaces and targeted
neighbors are temporary disabled in the ldpe process until new sockets
bound to the new transport address are received from the parent.
This patch fixes a problem in which adjacencies weren't being removed
after the associated targeted neighbors were disabled. This was causing
ldpd not to set some MD5 sockoptions for new neighbors are thus preventing
MD5-protected sessions to come up after a change in the transport-address.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
It's doesn't make sense to enforce that a targeted-hello is received
on an LDP-enabled interface. It should be possible, for example, to use
LDP only to signal pseudowires and other another protocol (e.g. RSVP-TE)
to create end-to-end LSPs.
Signed-off-by: Renato Westphal <renato@opensourcerouting.org>
