Reduce default queue size to twice the vlib frame size. This throttles
the PP from transferring more packets than are necessary for the next
round of graph execution on a core. When combined with DMA'ing into L3
cache, Mpps for 64B packets increases 35%.
Change-Id: I0afe426d23560f262fdc56accb6302b099bd0076
Signed-off-by: Brian Brooks <brian.brooks@arm.com>
Needed a spinlock to protect the data vector. Cleaned up debug cli so
the output makes sense, and so that various parameters exist in one
place. Removed a nonsense memset-to-zero which led to ultra-confusing
results.
Change-Id: I91cd14ce7fe84fd2eceab86e016b5ee001993be4
Signed-off-by: Dave Barach <dbarach@cisco.com>
Three separate implementations, which vary by nearly a factor of two
in performance. Most of the performance difference is due to swapping
the src/dst mac addresses with an avx2 vector shuffle instruction.
Change-Id: Ieb36546d6074e4ac720d452a99d013c698135c57
Signed-off-by: Dave Barach <dave@barachs.net>
calling thread cancel on the timeout thread whilst it was
sleep on condwait and then send the cond signal did not reliably
wake up the thread.
instead don;t cancel the thread, use a loop variable to terminate it.
Change-Id: Ibc8ab6f21db7e4a98266bdf88b8b208b887820dd
Signed-off-by: Neale Ranns <neale.ranns@cisco.com>
Wrong vl_msg_id set in vl_api_stats_get_poller_delay_t_handler while reporting stats delay.
Change-Id: I7d6ad14359e41c717b976390c58e254e3602e0ff
Signed-off-by: Lukasz Majczak <lma@semihalf.com>
- assign rx thread on interface create, not on admin up
- keep interrupts unmasked all the time as they are needed for writeback
- forward interrupt to graph dispatcher only if queue is in the interrupt mode
Change-Id: I5cc5afe24960143bef08f12d59fbd80ae0c6ccf8
Signed-off-by: Damjan Marion <damarion@cisco.com>
in case of relative path, create subdirs inside runtime dir.
in case of absolute path, if folder does not exist return error.
Change-Id: I01fe018102c69deb105160cfa18b741541d76c82
Signed-off-by: Jakub Grajciar <jgrajcia@cisco.com>
avf_create_reply returns software index for the new interface
Change-Id: I8a6b1a1985b072efafa24eb258b1f2cb1bea1110
Signed-off-by: Jakub Grajciar <jgrajcia@cisco.com>
Move the tcp unit test to the plugin
Add a bihash unit test and a "make test" program to call it
Adjust framework.py to load the plugin, which is disabled by default
Change-Id: Ic229d386a56a9d28dbd54974f231149053ca8f93
Signed-off-by: Dave Barach <dave@barachs.net>
One handler is good enough when engine sends out a retry status
Thus this patch will just go one branch
Change-Id: Id81cb3fa67d2b322b0fe1b2f62cd866cf3491eb4
Signed-off-by: Ping Yu <ping.yu@intel.com>
It's not clear that fine-graind pipelining will ever produce optimal
performance. That having been written, it's an easier coding model for
folks interesting in decent results without having to manually
implement 2 or 4 way fine-grained parallelism.
Change-Id: Ida877292832eaf759ffbb6d2e16012f8fbd427e5
Signed-off-by: Dave Barach <dave@barachs.net>
Can happen if code bypasses the per-interface output node, and
dispatches packets directly to the tx node.
Switch to vlib_get_buffers(...) ... vlib_buffer_enqueue_to_next (...),
quad/single loop coding pattern.
Change-Id: Ic0e5d3b9748230f4e545a54186e6e64e7a782bb1
Signed-off-by: Dave Barach <dave@barachs.net>
The dev build ('make build') was failing because
ip4_rewrite_bcast() and ip6_rewrite_bcast() are declared static
and were not referenced.
The node functions that were set for directed broadcast rewrite:
ip4_rewrite_bcast_node - ip4_rewrite()
ip6_rewrite_bcast_node - ip6_rewrite()
Changed to:
ip4_rewrite_bcast_node - ip4_rewrite_bcast()
ip6_rewrite_bcast_node - ip6_rewrite_bcast()
The release build ('make build-release') succeeds because there is
a VLIB_NODE_FUNCTION_MULTIARCH() call referencing the function. That
macro is empty for dev builds.
Change-Id: I2a05f00ca43d0eae8ff2e3026991e69917003fe1
Signed-off-by: Matthew Smith <mgsmith@netgate.com>
This prevents deadlock in case when worker A sends to B and worker B
sends to A
Change-Id: Id9436960f932c58325fe4f5ef8ec67b50031aeda
Signed-off-by: Damjan Marion <damarion@cisco.com>
Add a bucket-level lock bit. Use a spinlock only when actually
allocating, freeing, or splitting a bucket. Should improve
multi-thread add/del performance.
Change-Id: I3e40e2a8371685457f340d6584dea14e3207f2b0
Signed-off-by: Dave Barach <dave@barachs.net>
with ip direct broadcast enable a packet to the interface's
subnet broadcast address with be sent L2 broadcast on the
interface. dissabled, it will be dropped. it is disabled by
default, which preserves current behaviour
Change-Id: If154cb92e64834e97a541b32624354348a0eafb3
Signed-off-by: Neale Ranns <nranns@cisco.com>
DBGvpp# set interface rx-mode AVF0/4/2/0 interrupt
set interface rx-mode AVF0/4/2/0 interrupt
DBGvpp# sh int rx
sh int rx
Thread 0 (vpp_main):
node dpdk-input:
FortyGigabitEthernet6/0/0 queue 0 (polling)
node avf-input:
AVF0/4/2/0 queue 0 (interrupt)
DBGvpp# sh int addr
sh int addr
AVF0/4/2/0 (up):
L3 133.1.1.10/24
FortyGigabitEthernet6/0/0 (dn):
VirtualEthernet0/0/0 (up):
L3 192.168.42.1/24
DBGvpp# clear run
clear run
DBGvpp# clear interfaces
clear interfaces
DBGvpp# sh int rx
sh int rx
Thread 0 (vpp_main):
node dpdk-input:
FortyGigabitEthernet6/0/0 queue 0 (polling)
node avf-input:
AVF0/4/2/0 queue 0 (interrupt)
DBGvpp# ping 133.1.1.3 repeat 1
ping 133.1.1.3 repeat 1
64 bytes from 133.1.1.3: icmp_seq=1 ttl=64 time=.6237 ms
Statistics: 1 sent, 1 received, 0% packet loss
DBGvpp# sh int
sh int
Name Idx State MTU (L3/IP4/IP6/MPLS) Counter Count
AVF0/4/2/0 5 up 9000/0/0/0 rx packets 1
rx bytes 98
tx packets 1
tx bytes 98
drops 1
ip4 1
FortyGigabitEthernet6/0/0 1 down 9000/0/0/0
VirtualEthernet0/0/0 2 up 9000/0/0/0
local0 0 down 9000/0/0/0
DBGvpp# sh run
sh run
Time 41.9, average vectors/node 1.00, last 128 main loops 0.00 per node 0.00
vector rates in 2.3858e-2, out 2.3858e-2, drop 2.3858e-2, punt 0.0000e0
Name State Calls Vectors Suspends Clocks Vectors/Call
AVF0/4/2/0-output active 1 1 0 1.24e4 1.00
AVF0/4/2/0-tx active 1 1 0 4.28e4 1.00
api-rx-from-ring any wait 0 0 2 6.94e4 0.00
avf-input interrupt wa 1 1 0 1.09e4 1.00
avf-process any wait 0 0 47 1.01e4 0.00
...
DBGvpp#
DBGvpp# set interface rx-mode AVF0/4/2/0 polling
set interface rx-mode AVF0/4/2/0 polling
DBGvpp# sh int rx
sh int rx
Thread 0 (vpp_main):
node dpdk-input:
FortyGigabitEthernet6/0/0 queue 0 (polling)
node avf-input:
AVF0/4/2/0 queue 0 (polling)
DBGvpp# ping 133.1.1.3 repeat 1
ping 133.1.1.3 repeat 1
64 bytes from 133.1.1.3: icmp_seq=1 ttl=64 time=.2038 ms
Statistics: 1 sent, 1 received, 0% packet loss
DBGvpp#
Change-Id: Ib7d2b505ae4bb74a052263af766ca37d9df7bfe4
Signed-off-by: Steven <sluong@cisco.com>
Of the form "if ((foo & 0x10) == 1)", as opposed to
"if (foo & 0x10)"
Change-Id: I6a6d276aeed4af7c1c6c78546ee68d598d54d7db
Signed-off-by: Dave Barach <dave@barachs.net>
Both vpp-devel and vpp-dpdk-devel have /usr/include/dpdk
in their files list. CentOS 7 won't allow them to both be
installed at the same time as a result.
Change vpp-dpdk-devel to /usr/include/dpdk/* so both can
be installed. It is useful to have both installed on a
development system.
Change-Id: I94b6cca299dfbd991638e43c41006db0d8f7ee8b
Signed-off-by: Matthew Smith <mgsmith@netgate.com>
Fix a typo from previous patch. Change 0x104 to 0x1004.
Change-Id: I82230a8a0ec01567eb1d4bc12ac02062c2a98347
Signed-off-by: Matthew Smith <mgsmith@netgate.com>
The java command fails with missing class error,
when some of the JARs given by -cp are missing,
which may be missleading.
This patch fixes that by adding os.path.isfile check to test_jvpp.py.
Change-Id: I3a0b6ef338c7f70cfd0ba78ee0888efe0b0957b3
Signed-off-by: Marek Gradzki <mgradzki@cisco.com>
Thus when engine buffer is full during a burst in performance
tesing, this code will help VPP handle retry machansim.
Change-Id: I0f9fc05d3dba8a54d34dca4c6137700d6c80f714
Signed-off-by: Ping Yu <ping.yu@intel.com>
- removed handoff-dispatch node
- removed some unused buffer metadata fields
- enqueue to thread logic moved to inline function
Change-Id: I7361e1d88f8cce74cd4fcec90d172eade1855cbd
Signed-off-by: Damjan Marion <damarion@cisco.com>
handle the case where qos record was enabled more than once
Change-Id: Ic901b2f5b400751e82d57655983d10457ec62243
Signed-off-by: Eyal Bari <ebari@cisco.com>
each edge/arc from these nodes must be the same.
Change-Id: Id5dace61bca0af71ad1df98583425226e81fd0de
Signed-off-by: Neale Ranns <neale.ranns@cisco.com>