Adds flags to the packet buffer to track the number of VLANs in
the current Ethernet frame. We use two bits to signify 0, 1 or
2 VLANs. The value 3 signififies an unknown quantity of VLANs,
which includes "three or more" which is not widely supported.
We place the bits in the vlib_buffer section; that is not the
opaque section, so that all subordinate nodes can use it.
For background, see the discussion thread at
https://lists.fd.io/pipermail/vpp-dev/2016-March/000354.html
The helper macro ethernet_buffer_header_size(buffer) uses
these bits stored in "buffer" to calculate the Ethernet header
size.
The macro ethernet_buffer_set_vlan_count(buffer, count) sets the
appropriate bit values based on the number in "count".
By current frame we are referring to the case where a packet
that arrives from the wire is carrying an encapsulated Ethernet
packet. Once decapsulated that inner packet becomes the current
frame.
There are two places where this value is set; For most Ethernet
frames this will be in the "ethernet-input" node when that node
parses the Ethernet header. The second place is whenever
vnet_update_l2_len() is used to update the layer 2 opaque data.
Typically this function is used by nodes just before they send
a packet into l2-input.
These bits are zeroed in vlib_buffer_init_for_free_list()
meaning that wherever the buffer comes from they have a reasonable
value (eg, if ip4/ip6 generates the packet.)
Primarily this VLAN counter is used by nodes below "ethernet-
input" and "l2-input" to determine where the start of the
current Ethernet header is. There is opaque data set by
"ethernet-input" storing the offset of the current Ethernet
header but, since this is opaque, it's not usable by downstream
nodes. Previously several nodes have made assumptions regarding
the location of the Ethernet header, including that it is always
at the start of the packet buffer (incorrect when we have
encapsulated packets) or that it is exactly
sizeof(ethernet_header_t) away (incorrect when we have VLAN tags.)
One notable case where this functionality is required is in
ip6_neighbor when it generates a response to a received neighbor
soliciation request; it reuses the incoming Ethernet header
in-situ and thus needs to reliably know where that header begins.
Also, at the suggestion of Dave Barach, this patch removes
definition of HGSHM bits in the buffer flags since they are
unused and unlikely to ever be.
Change-Id: I00e4b9ced5ef814a776020c395d1774aba6185b3
Signed-off-by: Chris Luke <chrisy@flirble.org>
Buffers output from CLI commands as it arrives and allows the terminal to
view this buffer one page at a time.
- View of the buffer is moved with space/enter/up/down/page-up/page-down/
home/end.
- At the end of the CLI command if less than one page of output was
generated then the pager is not engaged.
- 'q' to quit the pager, or scroll off the bottom.
- Pager prompt displays the current view aperture line numbers and total
number of lines buffered.
- Can be disabled at runtime with "no-cli-pager" in the unix configuration.
- The number of lines that will be stored in the buffer is limited to
prevent excessive memory use; this limit is configurable with
"cli-pager-buffer-limit" in the unix configuration.
- Both these options can also be set in the console with "set terminal
pager [on|off] [limit <lines>]".
Limitations:
- Does not yet implement a search function.
- Whilst the terminal size is detected and tracked, changing the terminal
size when the pager is being used will not cause a redraw of the page.
- Lines that wrap the right most column are not yet handled gracefully.
Change-Id: I69548c1464eff79c53e122668f25758266daf1c2
Signed-off-by: Chris Luke <chrisy@flirble.org>
- Remove the '?' mechanism that previously only worked on telnet
connections in favor of a more shell-like "history" command.
The '?' approach had strange side-effects, like executing what
was already in the command buffer.
Change-Id: I043086b7f400c66c332a32dbd06ef580ecb18ee8
Signed-off-by: Chris Luke <chrisy@flirble.org>
_______ _ _ _____ ___
__/ __/ _ \ (_)__ | | / / _ \/ _ \
_/ _// // / / / _ \ | |/ / ___/ ___/
/_/ /____(_)_/\___/ |___/_/ /_/
- For terminals that look like they support ANSI output the FD.io
part is colored red.
- This is only shown at the start of a debug CLI or a telnet CLI
session.
- This banner can be disabled with "cli-no-banner" in the "unix"
section of the startup config file.
Change-Id: I085b3780dcca3eae546859dbde6c1c34c8258b9f
Signed-off-by: Chris Luke <chrisy@flirble.org>
A fairly comprehensive re-work of the built-in debug and telnet CLI
to add various command line editing features and to add command history
to the debug CLI.
This may seem like a large change but a good amount of it is merely
reworking the existing CLI code (which changed its indent level).
The features this patch enables include:
- Enable history in the debug CLI.
- Put both stdin and telnet connections in char-by-char mode.
- Echo from the server, not the client, for more control.
- Add a mostly no-op but fairly complete Telnet protocol processor.
- Perform control code parsing on the input byte stream to match strings
of both control codes and ANSI/VT100 escape sequences.
- Up/down keys scroll through the history (like ^P/^N).
- Do CRLF output cooking (\n -> \r\n) for connections that need it.
- Left/right cursor movements, insert/erase at cursor.
- Home/end cursor jumps.
- Jump left/right word at a time (Ctrl-left/right).
- Negotiate the terminal type from Telnet clients. (well, the code doesn’t
really negotiate, it demands it, but the client is led to believe it
was a negotiation)
- Read terminal type from TERM variable for the local debug CLI.
- Delete from cursor to end of line (^K). Delete char-right (^D/Del).
- Clear screen (^L) and repaint prompt/current command (on non-ANSI
terminals it just newlines and repaints the line).
Change-Id: Id274b56ccfd4cc8c19ddc0f478890f21f284262a
Signed-off-by: Chris Luke <chrisy@flirble.org>
For using clang as a compiler it is enough
to specify CC=clang in the make command line
Change-Id: I06f1c1d418b68768f8119de5bdc8748c51f90c02
Signed-off-by: Damjan Marion <damarion@cisco.com>
When loopback interface is configured as BVI, instead of changing its
output node from loopN-output to l2-input, the loopN-output node is now
kept while its next tx node is changed from ethernet-input to l2-input.
The packet setup previously done in bvi_to_l2 as part of l2-input is now
performed in the loop output node.
This change adds an extra node in the BVI output path but provides the
following improvements:
1. IP address/route created on loopback prior to it being configured as
BVI will still work properly. The requirement to (re)configure IP/route
on loopback after it is configured as BVI is removed.
2. The output stats for loopback interfaces are always provided irrespective
of their BVI configuration.
3. The loopback-BVI output stats can be batch updated outside the packet
loop in output node, instead of per packet update in l2-input node,
making l2-input node more efficient for BVI packets.
4. Restore original node property as implemented in node.c function
vlib_node_add_next_with_slot() where next node indices stored in next
slots of each node will remain unique.
5. Packet trace for BVI output includes loopN output node which provides
useful packet data.
Change-Id: I7f5bc72ef953a367363a179088210596881f9e73
Signed-off-by: John Lo <loj@cisco.com>
gcc version 4.9.2 (Raspbian 4.9.2-10)
Tested on Linux raspberrypi 4.4.6-v7+ #875 SMP Tue Apr 12 16:33:02 BST 2016 armv7l GNU/Linux
CPUs may be little or big endian, detect with gcc flags, not the processor architecture
Add a new flag $(PLATFORM)_uses_openssl which allows to disable the link with openssl lib.
vlib/vlib/threads.c:
startup.conf must:
- specify the heapsize as we don't have hugepages on raspbian
cpu {
main-core 3
}
heapsize 64M
Corrects in various files the assumption uword == u64 and replaces 'u64' cast with 'pointer_to_uword' and 'uword_to_pointer' where appropriate.
256 CPUs may create an OOM when testing with small memory footprint ( heapsize 64M ), allows the number of VLIB_MAX_CPUS to be set in platforms/*.mk
vppinfra/vppinfra/longjmp.S:
ARM - copy r1 (1st parameter of the setjmp call) to r0 (return value)
vppinfra/vppinfra/time.h:
On ARMv7 in AArch32 mode, we can access to a 64bit register to retreive the cycles count.
gcc on rpi only declare ARM_ARCH 6. Override this info, and check if it is possible to use 'mrrc'.
/!\ the time function will NOT work without allowing the user mode access to the PMU.
You may download the source of the kmod here:
https://github.com/christophefontaine/arm_rdtsc
Change-Id: I8142606436d9671a184133b935398427f08a8bd2
Signed-off-by: Christophe Fontaine <christophe.fontaine@qosmos.com>
This is complete rework of DPDK PCI initialization. It drops
previous scheme where lspci/route/awk/sed are used and instead
sysfs is solely used for discovering Ethernet PCI devices. Criteria
for blacklisting device is changed from exsiting routing table entry
to simple interface state obtained by SIOCGIFFLAGS ioctl().
It checks for IFF_UP flag, so as long as interface is declared
up and even when carrier is down interface will be blacklisted.
Change-Id: I59961ddcf1c19c728934e7fe746f343983741bf1
Signed-off-by: Damjan Marion <damarion@cisco.com>
The current mechanism for setting up arp-input and ip6-discover-neighbor
output nodes for interfaces using their interface link up/down callback
function is inefficient and has potential timing issue, as observed for
bonded interface. Now both nodes will setup output interface sw_if_index
in the the sw_if_index[VLIB_TX] field of current packet buffer and then
use the interface-ouput node to tx the packet.
One side effect is that vlib_node_add_next_with_slot() needs to be
modified to allow the same output node-id to be put at the specified
slot, even if another slot contain that same node-id already exist. This
requirement is caused by BVI support where all loopback interfaces set
up as BVIs will have the same output node-id being l2-input while, for
output-interface node, the output slot must match the hw_if_index of the
interface.
Change-Id: I18bd1d4fe9bea047018796f7b8a4d4c20ee31d6e
Signed-off-by: John Lo <loj@cisco.com>
Without this, frames can be double-freed to nodes like "error-punt",
leading to buffer leaks and other problems.
Change-Id: Ie28a4f504254ee439f720dbaac7f12206cea753b
Signed-off-by: Todd Foggoa <tfoggoa@cisco.com>
I noticed while mucking about with lsof that vpp
was listening on port 5000.
telnet 0 5000 revealed that it was listening for
the cli on that port.
Digging into the code, it turns out that if you
do not configure cli-listen (Example:
unix {
cli-listen localhost:5002
}
)
Then vpp is listening on the first available port
starting at port 5000 anyway. This is a simple
patch to *not* listen unless configured to do so.
Change-Id: Id7f6f4d69e0a1642d2767849a90b21f38f21ecaa
Signed-off-by: Ed Warnicke <eaw@cisco.com>
Do not propagate flags into cloned vlib_next_frame_t's.
vlib_next_frame_init(...) sets nf->frame_index to ~0. If it turns out
that the original flags include VLIB_FRAME_IS_ALLOCATED, the wheels
fall off. And so on.
Change-Id: I8de18653acfcc8eb20cee36f4eb5b9e82234f21b
Signed-off-by: Dave Barach <dave@barachs.net>
This is 1st drop of VPP native driver for linux AF_PACKET.
New CLI:
create host-interface name <host-if-name> [hw-addr <mac-address>]
References:
- Documentation/networking/packet_mmap.txt in the Linux kernel tree
- man 7 packet
Known issues:
- attaching to linux bridge doesn't work
- it is not expected to work in multicore setup
Change-Id: I1cb1c3d305f349759e90e76e25696718b73bd73d
Signed-off-by: Damjan Marion <damarion@cisco.com>
The VLIB_FRAME_NO_FREE_AFTER_DISPATCH flag is not preserved when cloning
next_frames, as a result VLIB_FRAME_FREE_AFTER_DISPATCH can
erroneously be set for a frame (see vlib_get_next_frame_internal())
Change-Id: Ice1d7ddcb807e1168aa0c157d9474be492d102c2
Signed-off-by: Nikhil P Rao <nikhil.rao@intel.com>
It turns out that unix_physmem_init(...) has been effectively disabled
for a very long time. The vnet library supplied a weak symbol override
for the vlib_app_physmem_init(...) which returned 1, meaning "do
nothing." When we switched libvnet.a -> libvnet.so, the symbol
override stopped working.
Presto: unix_physmem_init(...) romps all over the data set up by
vlib_buffer_pool_create(...), leading to ASSERT failures and/or bus
errors, but only when using worker threads. Even then, the failure
depended in some complicated way on library dynamic load order.
We should remove .../vlib/vlib/unix/physmem.c entirely once we're sure
we'll never want it back.
Change-Id: I27747edbeb0de88d2f2d8728f7f8eb3135e7f0cf
Signed-off-by: Dave Barach <dave@barachs.net>
This avoids checking the buffer flags bit if tracing is not enabled.
Change-Id: I32e1a90b5fd10318254c611344488bc2a441c71e
Signed-off-by: Todd Foggoa (tfoggoa) <tfoggoa@cisco.com>
Turn of srp, mainly as an example of how to restructure a featurette
for selective disablement.
Change-Id: Id3364c58a8711b103939f4434adfa67177380f67
Signed-off-by: Dave Barach <dave@barachs.net>
This code provided inter-VM (2 cores per VM) throughput of
22Gbps using iperf through VPP (1 core) with 9k frames.
With the same setup and pktgen running on both sides, it
reached 5Mpps with no packets drop (Equivalent to before the patch).
During the tests the average vector length was about 1, which
likely means that VPP is not the bottleneck.
The patch also includes some generic functions for vlib buffers
allowing for chained buffer construction whether or not DPDK is enabled.
Change-Id: Icfd1803e84b2b4578f305ab730576211f6242d6a
Signed-off-by: Pierre Pfister <ppfister@cisco.com>
It also includes check to ensure that number of
per-cpu mheaps is not lower than number of cpus.
Change-Id: Ibc68b34dda130f922243f9ea15b03e44bbcac269
Signed-off-by: Damjan Marion <damarion@cisco.com>
does not close the socket. Resulting in the main thread being stuck
in a tight infinite loop polling on the erronous socket.
Change-Id: I630b84b97c059acce117d56e41cd201131db4cab
Signed-off-by: Ole Troan <ot@cisco.com>
The DPDK glue did not support cloned packets which do not
have a freelist handler. Add support for this case.
Change-Id: I8f17cd4952df97989d90d3f3e39792bc3739705c
Signed-off-by: Kevin Paul Herbert <kph@cisco.com>
Limit buffer tracing to 50 in order to limit large output, unless
the user over rides the max "sh trace max <number>".
Add trace filtering, to be able to only trace packets that were
processed by a specific node or exclude packets processed by a node.
Example, only include packets processed by error-drop:
# trace filter include error-drop 1
# trace add dpdk-input 1000000
<wait for packets, to come in>
# show trace
Change-Id: I5d9e15d2268ea55e6ef87b2b8756049c49b2791b
Signed-off-by: Todd Foggoa <tfoggoa@cisco.com>
Add some more ASSERTs to track down improper frees.
Change-Id: I2bd4b69fb14f522c82e6006131b6ad982f6f7e6b
Signed-off-by: Kevin Paul Herbert <kph@cisco.com>
This fixed performance issue in muti-threaded setup
due to sharing of the same cacheline between multiple threads
Change-Id: I930ee44c17a83d4da350d15b4b97b8bb4633a9b0
Signed-off-by: Damjan Marion <damarion@cisco.com>
