-- it works fine for phy's that are carrier-up, but crashes if they are
carrier-down.
0: /home/pim/src/vpp/src/vnet/interface_funcs.h:46 (vnet_get_hw_interface) assertion `! pool_is_free (vnm->interface_main.hw_interfaces, _e)' fails
at /home/pim/src/vpp/src/vppinfra/error.c:143
ns=0x7fff98774e80 "dataplane", host_sw_if_indexp=0x0) at /home/pim/src/vpp/src/plugins/lcpng/lcpng_interface.c:998
at /home/pim/src/vpp/src/plugins/lcpng/lcpng_if_cli.c:96
parent_command_index=371) at /home/pim/src/vpp/src/vlib/cli.c:591
parent_command_index=0) at /home/pim/src/vpp/src/vlib/cli.c:548
at /home/pim/src/vpp/src/vlib/cli.c:694
- move tap_set_carrier() upstream to lcp_itf_set_link_state()
- refuse to set admin-up on sub-int if parent is down
- no need to switch namespaces, lcp_itf_set_link_state() already does
- in change_mtu and change_admin_state, if the interface is a sub,
we only have to sync that one interface. Otherwise, walk the parent
interface and all sub-ints with lcp_itf_pair_sync_state_hw() and
make note of this in the (DBG) log
This is super complicated work, taken mostly verbatim from the upstream
linux-cp Gerrit, with due credit mgsmith@netgate.comneale@grafiant.com
First, add main handler lcp_nl_route_add() and lcp_nl_route_del()
Introduce two FIB sources: one for manual routes, one for dynamic
routes. See lcp_nl_proto_fib_source() fo details.
Add a bunch of helpers that translate Netlink message data into VPP
primitives:
- lcp_nl_mk_addr46()
converts a Netlink nl_addr to a VPP ip46_address_t.
- lcp_nl_mk_route_prefix()
converts a Netlink rtnl_route to a VPP fib_prefix_t.
- lcp_nl_mk_route_mprefix()
converts a Netlink rtnl_route to a VPP mfib_prefix_t.
- lcp_nl_proto_fib_source()
selects the most appropciate fib_src by looking at the rt_proto
(see /etc/iproute2/rt_protos for a hint). Anything RTPROT_STATIC or
better is 'fib_src', while anything above that becomes fib_src_dynamic.
- lcp_nl_mk_route_entry_flags()
generates fib_entry_flag_t from the Netlink route type,
table and proto metadata.
- lcp_nl_route_path_parse()
converts a Netlink rtnl_nexthop to VPP fib_route_path_t and adds
that to a growing list of paths.
- lcp_nl_route_path_add_special()
adds a blackhole/unreach/prohibit route to the list of paths, in
the special-case there is not yet a path for the destination.
Now we're ready to insert FIB entries:
- lcp_nl_table_find()
selects the matching table-id,protocol(v4/v6) from a hash of tables.
- lcp_nl_table_add_or_lock()
if at table-id,protocol(v4/v6) hasn't been used yet, create one,
otherwise increment a table reference counter so we know how many
FIB entries we have in this table. Then, return it.
- lcp_nl_table_unlock()
Decrease the refcount on a table, and if no more prefixes are in
the table, remove it from VPP.
- lcp_nl_route_del()
Remove a route from the given table-id/protocol. Do this by applying
rtnl_route_foreach_nexthop() to the list of Netlink nexthops,
converting them into VPP paths in a lcp_nl_route_path_parse_t
structure. If the route is for unreachable/blackhole/prohibit in
Linux, add that path too.
Then, remove the VPP paths from the FIB and reduce refcnt or
remove the table if it's empty using table_unlock().
- lcp_nl_route_add()
Not all routes are relevant for VPP. Those in table 255 are 'local'
routes, already set up by ip[46]_address_add(), and some other route
types are invalid, skip those. Link-local IPv6 and IPv6 multicast is
also skipped. Then, construct lcp_nl_route_path_parse_t by walking
the Netlink nexthops, and optionally add a special (in case the
route was for unreachable/blackhole/prohibit in Linux -- those won't
have a nexthop).
Then, insert the VPP paths found in the Netlink message into the FIB
or the multicast FIB, respectively.
And with that, Bird shoots to life. Both IPv4 and IPv6 OSPF interior
gateway protocol and BGP full tables can be consumed, on my bench in
about 9 seconds:
- A batch of 2048 Netlink messages is handled in 9-11ms, so we can do
approx 200K messages/sec at peak (and this will consume 50% CPU due
to the yielding logic in lcp_nl_process() (see the 'case
NL_EVENT_READ' block that adds a cooldown period of
LCP_NL_PROCESS_WAIT milliseconds between batches.
- With 3 route reflectors and 2 full BGP peers, at peak I could see
309K messages left in the producer queue.
- All IPv4 and IPv6 prefixes made their way into the FIB
pim@hippo:~/src/lcpng$ echo -n "IPv6: "; vppctl sh ip6 fib summary | awk '$1~/[0-9]+/ { total += $2 } END { print total }'
IPv6: 132506
pim@hippo:~/src/lcpng$ echo -n "IPv4: "; vppctl sh ip fib summary | awk '$1~/[0-9]+/ { total += $2 } END { print total }'
IPv4: 869966
- Compared to Bird2's view:
pim@hippo:~/src/lcpng$ birdc show route count
BIRD 2.0.7 ready.
3477845 of 3477845 routes for 869942 networks in table master4
527887 of 527887 routes for 132484 networks in table master6
Total: 4005732 of 4005732 routes for 1002426 networks in 2 tables
- Flipping one of the full feeds to another, forcing a reconvergence
of every prefix in the FIB took about 8 seconds, peaking at 242K
messages in the queue, with again an average consumption of 2048
messages per 9-10ms.
- All of this was done while iperf'ing 6Gbps to and from the
controlplane.
---
Because handling full BGP table is O(1M) messages, I will have to make
some changes in the logging:
- all neigh/route messages become DBG/INFO at best
- all addr/link messages become INFO/NOTICE at best
- when we overflow time/msgs, turn process_msgs into a WARN, otherwise
keep it at INFO so as not to spam.
In lcpng_interface.c:
- Log NOTICE for pair_add() and pair_del() call;
- Log NOTICE for set_interface_addr() call;
With this approach, setting the logging level of the linux-cp/nl plugin
to 'notice' hits the sweet spot: with things that the operator has
~explicitly done, leaving implicit actions (BGP route adds/dels, ARP/ND)
to stay below the NOTICE level.
Introduce lcp_main.lcp_sync, which determines if state changes made
to interfaces in VPP do or don't propagate into Linux.
- Add a startup.conf directive 'lcp-sync' to enable at startup time.
- Add CLI.short_help = "lcp lcp-sync [on|enable|off|disable]",
- Show the current value in "show lcp".
Gate changes in mtu, state and address on lcp_lcp_sync().
When the operator issues 'lcp lcp-sync on', it is prudent to do a
one-off sync of all interface attributes from VPP into Linux.
For this, add a lcp_itf_pair_sync_state_all() function.
In preparation of another feature 'netlink-auto-subint', rename
lcp_main's field to "lcp_auto_subint".
Add CLI .short_help = "lcp lcp-auto-subint [on|enable|off|disable]"
Show status of the field on "lcp show" output.
I was looking at the hw interface list, which makes sense for ethernet
devices but not for other devices, notably BondEthernets.
In addition, creating two separate interfaces with the same outer
(for example Gi3/0/0 dot1ad 2345 inner-dot1q 1000 AND the same on
Gi3/0/1) would yield an erratic match and a crash.
Switch to walking the sw interface list instead, and search for
the sup_sw_if_index that has the desired outer. Result:
BondEthernet0.{1234,1235,1236,1237} can be created and are functional.
Update the pair_config() parser to follow suite.
When the configuration 'lcp-auto-subint' is set, and the interface at
hand is a subinterface, in lcp_itf_interface_add_del():
- if it's a deletion and we're a sub-int, and we have a LIP: delete it.
- if it's a creation and we're a sub-int, and our parent has a LIP, create one.
Fix a few logging consistency issues (pair_del), and in
pair_delete_by_index() ensure that the right namespace is selected.
Due to this quirk with lip->lip_namespace not wanting to be a vec_dup()
string, rewrite them all to be strdup/free instead.
This first preparation moves lcp_itf_phy_add() to lcpng_if_sync.c
and renames it lcp_itf_interface_add_del().
It does all the pre-flight checks to validate that a new device, given
by sw_if_index, can have a LIP created:
- must be a sub-int
- must have a sw_sup_if_index, which itself has a LIP
However, I realize that I cannot create an interface from within an
interface add callback, so I'll have to schedule the child LIP to be
created by a process, after the callback returns.
I'll do that in the next commit.
I've made a few cosmetic adjustments:
- introduce debug, info, notice, warn and err loggers
- remove redundant logging, and set correct (conservative) log levels
- turn the sync-state error into a warning
And a little debt paydown:
- refactor sync-state into its own function, to be called instead of
all the spot fixes elsewhere. It's going to be the case that
sync-state is "the reconsiliation function".
- Fix a bug in lip->lip_namespace copy: vec_dup() there doesn't seem
to work for me, use strdup() instead and make a mental note to
reviist.
The plugin now works with 'lcpng default netns dataplane' in its
startup.conf; and with 'lcp default netns dataplane' as its first
command. A few of these fixes should go upstream, too, which I'll
do next.
I have been very careless in using the correct network namespace when
manipulating LCP host devices. Around any/every netlink write operation,
we must first clib_setns() into the correct namespace. So, wrap every
call of vnet_netlink_*() in all places.
For consistency, use the convention 'curr_ns_fd' (for the one we are
coming from) and 'vif_ns_fd' (to signal the one that the netlink VIF
is in).
Be careful as well to enter and exit everywhere without losing file
descriptors.
There are three ways in which IP addresses will want to be copied
from VPP into the companion Linux devices:
1) set interface ip address ... adds an IPv4 or IPv6 address
- this is handled by lcp_itf_ip[46]_add_del_interface_addr() which
is a callback installed in lcp_itf_pair_init()
2) set interface ip address del ... removes them
- also handled by lcp_itf_ip[46]_add_del_interface_addr() but
curiously there is no upstream vnet_netlink_del_ip[46]_addr() so
I wrote them inline here - I will try to get them upstreamed, as
they appear to be obvious companions in vnet/device/netlink.h
3) Upon LIP creation, it could be that there are L3 addresses already
on the VPP interface. If so, set them with lcp_itf_set_interface_addr()
This means that now, at any time a new LIP is created, its state from
VPP is fully copied over (MTU, Link state, IPv4/IPv6 addresses)!
At runtime, new addresses can be set/removed as well.
This is the first in a series of functions that aims to copy forward
interface changes in the VPP dataplane into the linux interfaces.
Capture link state changes (set interface state ...) and apply them
to Linux.
There's an important dissonance here:
- When Linux sets a parent interface up, all children also go up.
ip link set enp66s0f1 down
ip link add link enp66s0f1 name foo type vlan id 1234
ip link set foo down
ip link | grep enp66s0f1
9: enp66s0f1: <BROADCAST,MULTICAST> mtu 9000 qdisc mq state DOWN mode DEFAULT group default qlen 1000
61: foo@enp66s0f1: <BROADCAST,MULTICAST,M-DOWN> mtu 9000 qdisc noop state DOWN mode DEFAULT group default qlen 1000
ip link set enp66s0f1 up
ip link | grep s0f1
9: enp66s0f1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 9000 qdisc mq state UP mode DEFAULT group default qlen 1000
61: foo@enp66s0f1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 9000 qdisc noqueue state UP mode DEFAULT group default qlen 1000
While in VPP this is not so, there each individual interface and
sub-interface stands for itself. I think the proper fix here is to walk
all sub-interfaces when a phy changes, and force a sync of those from
VPP to LCP as well. I'll do that in a followup commit so it's easier to
roll back.
When I started in my copy, I removed a bunch of code and options that I
felt were distracting. I also renamed lots of elements like 'linux-cp'
and 'Linux CP' and 'Linux-CP' to just be 'lcpng'.
Now, rename all of this back, and make it ready for upstreaming.
The only diffs between my repo and upstream now are the includes and the
lcpng_interface.[ch] code changes, which is good.
In this situation, Linux and VPP really diverge. In VPP, any
sub-interface can carry arbitrary configuration, they can be dot1q,
dot1ad, with or without an inner dot1q. So the following is valid in
VPP:
vppctl create sub TenGigabitEthernet3/0/0 10 dot1ad 100 inner-dot1q 200 exact-match
In Linux however, double tagged interfaces have to be created as a chain
of two interfaces, first with the outer and then with the inner tag. So
there is no equivalent of the above command in Linux, where we must:
ip link add link e0 name e0.100 type vlan id 100 proto 802.1ad
ip link add link e0.100 name e0.100.200 type vlan id 200 proto 802.1q
So in order to create Q-in-Q sub-interfaces, for Linux their intermediary
parent must exist, while in VPP this is not true.
I have to make a compromise, so I'll be a bit more explicit and allow
this type of LCP to be created under these conditions:
* A sub-int exists with the intermediary (in this case, `dot1ad 100
exact-match`)
* That sub-int itself has an LCP, with a Linux interface device that
we can spawn the inner-dot1q 200 interface off of
Creation of qinq and qinad interfaces becomes thus:
vppctl create sub TenGigabitEthernet3/0/0 10 dot1ad 100 exact-match
vppctl create sub TenGigabitEthernet3/0/0 11 dot1ad 100 inner-dot1q 200 exact-match
vppctl lcpng create TenGigabitEthernet3/0/0 host-if e0
vppctl lcpng create TenGigabitEthernet3/0/0.10 host-if e0.10
vppctl lcpng create TenGigabitEthernet3/0/0.11 host-if e0.11
And the resulting situation in Linux:
pim@hippo:~/src/lcpng$ ip link | grep e0
397: e0: <BROADCAST,MULTICAST> mtu 9000 qdisc mq state DOWN mode DEFAULT group default qlen 1000
398: e0.10@e0: <BROADCAST,MULTICAST,M-DOWN> mtu 9000 qdisc noop state DOWN mode DEFAULT group default qlen 1000
399: e0.11@e0.10: <BROADCAST,MULTICAST,M-DOWN> mtu 9000 qdisc noop state DOWN mode DEFAULT group default qlen 1000
If there are no LCPs defined yet, and I try to create an LCP for a sub-int,
there is a NULL deref in lcp_itf_pair_get() because the pool hasn't been
initiated yet. Fix this by returning NULL if the pool isn't initialized,
and catching the invalid `lip` returning an error.
While I was here, I took a look at a few possible crashes:
- if the pair_create() is called with an invalid phy_sw_if_index, catch
this and return an error.
- if a pair_add_sub() is called but the parent cannot be found, catch
this and return an error.
DBGvpp# set interface state TenGigabitEthernet3/0/0 down
DBGvpp# lcpng create TenGigabitEthernet3/0/0 host-if e0
DBGvpp# set interface state TenGigabitEthernet3/0/1 up
DBGvpp# lcpng create TenGigabitEthernet3/0/1 host-if e1
Yields:
pim@hippo:~/src/lcpng$ ip link show e0
304: e0: <BROADCAST,MULTICAST> mtu 9000 qdisc mq state DOWN mode DEFAULT group default qlen 1000
link/ether 68:05:ca:32:46:14 brd ff:ff:ff:ff:ff:ff
pim@hippo:~/src/lcpng$ ip link show e1
305: e1: <BROADCAST,MULTICAST,UP,LOWER_UP> mtu 9000 qdisc mq state UNKNOWN mode DEFAULT group default qlen 1000
link/ether 68:05:ca:32:46:15 brd ff:ff:ff:ff:ff:ff
This is to debug why creating QinQ subinterfaces is failing. Consider
the following:
vppctl create sub TenGigabitEthernet3/0/0 1234
vppctl create sub TenGigabitEthernet3/0/0 1235 dot1q 1234 inner-dot1q 1000
vppctl lcpng create TenGigabitEthernet3/0/0 host-if e0
vppctl lcpng create TenGigabitEthernet3/0/0.1234 host-if e0.1234
vppctl lcpng create TenGigabitEthernet3/0/0.1235 host-if e0.1235
Creating the sub-interfaces works, the first is a normal .1q with
tag 1234. Creating its LCP works well, the parent is looked up,
a netlink VLAN link is created as a child, and it gets tag 1234.
Now the second one: it's also operating on parent Te3/0/0 which is
looked up, but now a netlink VLAN link is created as a child, again
with dot1q 1234: this interface already exists, so that's a no-go
and an error is thrown.
-- Thoughts on a fix for this:
I think the fix is probably retrieving the correct lip, ie not
the lip of the phy parent interface (e0) but the lip of the pair that
has the outer vlan 1234 already (e0.1234), and then asking netlink to
create a child interface with vlan 1235 on that e0.1234, rather than
the phy e0.
Although creating a dot1ad.dot1q or a dot1q.dot1q interface in VPP
is strictly valid, we will not be able to succeed without the
intermediate interface in the Linux model, so we return an error
in that case.
Remove the functionality that allows for a configuration of pairs in the
startup.conf -- I am intending on creating interfaces for each/any phy
and sub and tunnel interfaceb that is created.
Instead of injecting events and having a process listening for them,
simply use the callback to create LCP interfaces. This change removes
the old code and sets a helpful logging entry on line 870, in which
future automatic creations and removals of LCP taps will occur.
This way, three desirable usability properties are obtained:
1) on startup, all physical interfaces will be copied into LCP
2) on sub-interface or tunnel creation (or phy insertion), new
interfaces will be created.
3) Deletions and removals will allow for auto-cleanup of the LCP.
Still use a sensible default of 9216, but if the L3 packet size is set
on the VPP interface, copy it forward (just as we do in the 'host'
interface of the TAP itself, ie the interface created in the linux
namespace). Now they will all line up initially.
The TAP interface does copy forward the host MTU based on the VPP
interface's L3 MTU, but it should also ensure that the VPP tap
interface has an MTU that is greater-or-equal to those. Considering
users can set the interfaces at runtime (set interface mtu packet ...)
ensure that the tap MTU is large enough.
Ideally, a callback updates the MTU to the same value as the L3 MTU of
the VPP interface, or the L3 MTU of the host interface, if either of
those change. For now, it's a safe bet to take jumbo 9216.
This gives a lot of operational problems later. It's definitely reasonable
to be able to create tap interfaces in other namespaces, and this is
still possible (see below for syntax).
However, changing the runtime netns makes the netlink listener much more
complicated because it will have to listen on not just one netns, but all
of them, for netlink updates.
So, for now, let's remove the ability to set the namespace in the API.
Still possible:
- set at startup.conf in lcpng { netns <x> }
- force creating in 'lcpng create ... netns <x>'
This will nudge folks to create one singular namespace (say,
'dataplane', in the startup.conf), and then handle all netlink messages
in that namespace only.
