Age | Commit message (Collapse) | Author | Files | Lines |
|
Adds tx pacing infrastructure for transport protocols that want to use
it. Particularly useful for connections with non-negligible rtt and
constrained network throughput as it avoids large tx bursts that lead to
local interface tx or network drops.
By default the pacer is disabled. To enabled it for tcp, add tx-pacing
to tcp's startup conf. We are still slightly inefficient in the handling
of incoming packets in established state so the pacer slightly affect
maximum throughput in low lacency scenarios.
Change-Id: Id445b2ffcd64cce015f75b773f7d722faa0f7ca9
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
frame->frame_flags
Change-Id: I56b573b5da04a27766bcbcafbd5438555424f2e7
Signed-off-by: Kingwel Xie <kingwel.xie@ericsson.com>
|
|
According to the RFC 4096 (section 3.3.1) the DATA chunk needs to be
padded to a boundary of 4 bytes with zeros. This patch addresses that
requirement.
At the same time, this patch takes care of adding some hardening for
corner-cases where the transmitted tag could be wrong.
Change-Id: I3b653926e9933d0d3d46bc5f37eaceefd932e874
Signed-off-by: Marco Varlese <marco.varlese@suse.com>
|
|
Change-Id: I09cf6ee7b4874d71f293f8f1a426d3a5e9651749
Signed-off-by: Neale Ranns <nranns@cisco.com>
|
|
Change-Id: I1f58f441c65fbca101bee2e864bfa6ae2306b475
Signed-off-by: Neale Ranns <nranns@cisco.com>
|
|
Change-Id: I8bf411adc6c5f4caa349d161174b544d2de3ad1d
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
Patch is too large to be ported to 18.10 just days before release.
- handle fast retransmits outside of established node and limit the
retransmit burst size to avoid tx losses and worsening congestion.
- in the absance of a tx pacer, use slow start after fast retransmit
exists
- add fast retransmit heuristic that re-retries sending the first
segment if everything else fails
- fine tuning
Change-Id: I84a2ab8fbba8b97f1d2b26584dc11a1e2c33c8d2
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
Change-Id: I7030951215f0cf7d00c037892ac92d9c304cb5a1
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
if udp length is specified, then use this length to calculate udp checksum
otherwise, use length from vlib_buffer
Change-Id: I5304a60d5d429993d0524b864b65ec503775412d
Signed-off-by: Kingwel Xie <kingwel.xie@ericsson.com>
|
|
Change-Id: Ied34720ca5a6e6e717eea4e86003e854031b6eab
Signed-off-by: Dave Barach <dave@barachs.net>
|
|
Change-Id: I5105b688ef3df2c949ba09e1e90c1b8913502388
Signed-off-by: Klement Sekera <ksekera@cisco.com>
|
|
Change-Id: I70bc5af646894811d373456ec66aa83f2d75a477
Signed-off-by: Klement Sekera <ksekera@cisco.com>
|
|
Change-Id: Idcda9ae55fa2efb0b2e928bac3e8e86ff8d19eba
Signed-off-by: Juraj Sloboda <jsloboda@cisco.com>
|
|
Change-Id: Ic6b27659f1fe9e8df39e80a0441305e4e952195a
Signed-off-by: Klement Sekera <ksekera@cisco.com>
|
|
Change-Id: I7d0930a19d927bbd7ba3fc879d5a0c8064827629
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
Change-Id: I3ac136e2a10796d8fa86ddb6f0d6cabe5fa749f8
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
pipe-rx node to match that of ethernet-input node
Since pipe-rx is a sibling node of ethernet-input, it ought to perform similarly: set l2/l3 header offsets,
and l2.l2_len value if the interface is in the l2 mode.
The use cases of pipes do not assume the tagged traffic, so
assume the simple ethernet header.
Change-Id: I7c9b5f4f2b1402cfbd10513f76cdd59b2db7a7a6
Signed-off-by: Andrew Yourtchenko <ayourtch@gmail.com>
|
|
This is first part of addition of atomic macros with only macros for
__sync builtins.
- Based on earlier patch by Damjan (https://gerrit.fd.io/r/#/c/10729/)
Additionally
- clib_atomic_release macro added and used in the absence
of any memory barrier.
- clib_atomic_bool_cmp_and_swap added
Change-Id: Ie4e48c1e184a652018d1d0d87c4be80ddd180a3b
Original-patch-by: Damjan Marion <damarion@cisco.com>
Signed-off-by: Sirshak Das <sirshak.das@arm.com>
Reviewed-by: Honnappa Nagarahalli <honnappa.nagarahalli@arm.com>
Reviewed-by: Ola Liljedahl <ola.liljedahl@arm.com>
Reviewed-by: Steve Capper <steve.capper@arm.com>
|
|
Change-Id: Ie46b3a81de4ed39b7b40e3879436f7e5a2908d98
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
Change-Id: I0bc4062c1fd3202ee201acb36a2bb14fc6ee1543
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
- fix delete of connection in syn-received
- fix delete of half-open connection
Change-Id: I72ff4b60406a2762d998328c52f41adea40d2c1b
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
Change-Id: I104f2b4e0b1c354d3671962be3f6002f2721388d
Signed-off-by: Neale Ranns <nranns@cisco.com>
|
|
Change-Id: I3ef0725684bcb8ea526abe0ce62562b35a0070f5
Signed-off-by: Eyal Bari <ebari@cisco.com>
(cherry picked from commit 0d87894bf279a4678cfca6cc438583090b166f85)
|
|
Change-Id: I0bfa4c9a3bb22155d7862cc13019cb172e77407c
Signed-off-by: Neale Ranns <nranns@cisco.com>
|
|
Change-Id: I1be7c59df7b48875f81ebeebf5f39ed15a43d2d8
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
Change-Id: If81ee34e1f1e929de1a5b758ddb9aede4002e858
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
Break up bond tx function into multiple small workloads:
1. parse the packet header and hash it based on the configured algorithm
2. optionally, trace the packet
3. convert the hash value from (1) to the slave port
4. update the buffers with the slave sw_if_index
5. Add the buffers to the queues
6. Create and send the frames
old numbers
-----------
Time 5.3, average vectors/node 223.74, last 128 main loops 40.00 per node 222.61
vector rates in 3.3627e6, out 6.6574e6, drop 3.3964e4, punt 0.0000e0
Name State Calls Vectors Suspends Clocks Vectors/Call
BondEthernet0-output active 68998 17662979 0 1.89e1 255.99
BondEthernet0-tx active 68998 17662979 0 2.60e1 255.99
TenGigabitEthernet3/0/1-output active 68998 8797416 0 1.03e1 127.50
TenGigabitEthernet3/0/1-tx active 68998 8797416 0 7.85e1 127.50
TenGigabitEthernet7/0/1-output active 68996 8865563 0 1.02e1 128.49
TenGigabitEthernet7/0/1-tx active 68996 8865563 0 7.65e1 128.49
new numbers
-----------
BondEthernet0-output active 304064 77840384 0 2.29e1 256.00
BondEthernet0-tx active 304064 77840384 0 2.47e1 256.00
TenGigabitEthernet3/0/1-output active 304064 38765525 0 1.03e1 127.49
TenGigabitEthernet3/0/1-tx active 304064 38765525 0 7.66e1 127.49
TenGigabitEthernet7/0/1-output active 304064 39074859 0 1.01e1 128.51
Change-Id: I3ef9a52bfe235559dae09d055c03c5612c08a0f7
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Change-Id: I4ba0aeb65219596475345e42b8cd34019f5594c6
Signed-off-by: mu.duojiao <mu.duojiao@zte.com.cn>
|
|
- ensure session enqueue epoch does not wrap between two enqueues
- use 3 states for echo clients app, to distinguish between starting and
closing phases
- force tcp fin retransmit if out of buffers while sending a fin
Change-Id: I6f2cab46affd1148aba2a33fb6d58bcc54f32805
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
Change-Id: I9052202b8cbcf656e61d635253d515f0f3a8d145
Signed-off-by: Neale Ranns <nranns@cisco.com>
|
|
Change-Id: Icd48de31302f62c59961c573699a1dd0474b8acb
Signed-off-by: Neale Ranns <nranns@cisco.com>
|
|
- add epoll dequeued events beyond maxevents to unhandled
- filter multiple epoll rx events
Change-Id: I618f5f02b19581473de891b3b59bb6a0faad10b5
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
keep the number of buckets in the load-balanced fixed. If a
path goes dwon fill its buckets with those from the next
available up path.
Change-Id: I15603ccb899fa9b77556b898c99136379cf32eae
Signed-off-by: Neale Ranns <nranns@cisco.com>
|
|
This commit adds two new nodes in the L2
datapath in input and output direction respectively.
These nodes fork the traffic into three feature
arcs: ip4, ip6 and nonip, which later join
to continue the usual L2 processing.
The vnet_l2_feature_enable_disable() function
with the same signature as
vnet_feature_enable_disable() takes care of
enabling the L2 datapath feature bits as needed, when
the features are enabled/disabled.
Thus, L2 features may use the similar plumbing as
the L3 features enjoy.
Change-Id: I76877b3a92d794c492bff1622bb26acba05705b2
Signed-off-by: Andrew Yourtchenko <ayourtch@gmail.com>
|
|
Change-Id: Idfed8243643780d3f52dfe6e6ec621c440daa6ae
Signed-off-by: mu.duojiao <mu.duojiao@zte.com.cn>
|
|
Change-Id: I70fb7394f85b26f7e632d74fc31ef83597efdd16
Signed-off-by: Eyal Bari <ebari@cisco.com>
|
|
Change-Id: I913ffc963928b74cf4753b6b9a0eed4436887e04
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
Change-Id: I544b24d2b2c4a09829773cf180d1747f4b087d4c
Signed-off-by: Florin Coras <fcoras@cisco.com>
|
|
fastpass case (VPP-1444)
20e6d36b has moved the calculation of the l3_hdr_offset into the determine_next_node()
function, with the assumption that the current_data in the buffer is at
the L3 header. This is not the case for the single loop fastpath,
where the vlib_buffer_advance() call is made after the call to
determine_next_node(), as a day1 behavior. As a result - that path
incorrectly sets the l3_hdr_offset.
Solution: move the vlib_buffer_advance() call to before determine_next_node()
Change-Id: Id5eaa084c43fb6564f8239df4a0b3dc0412b15de
Signed-off-by: Andrew Yourtchenko <ayourtch@gmail.com>
(cherry picked from commit 376414f4c3f53af44da4e82ee5d0b1843b291f8e)
|
|
Change-Id: I3cb89dbfb7174b9913a8c4ad9b3b1dc9f6ed6326
Signed-off-by: Matus Fabian <matfabia@cisco.com>
|
|
when multiple session creating script is ran (via exec) only the first
one actually starts
Change-Id: I0fc36f65795c8921cf180e0b555c446e5a80be45
Signed-off-by: Eyal Bari <ebari@cisco.com>
|
|
Change-Id: I5495c37da2fb8ff48c4af14ccba021d64eac52b6
Signed-off-by: Neale Ranns <nranns@cisco.com>
|
|
Change-Id: If3abeab2b304e2df41139707a4c8d045fab6ef99
Signed-off-by: Juraj Sloboda <jsloboda@cisco.com>
|
|
Change-Id: If178dd38e7920f35588f5d821ff097168b078026
Signed-off-by: Neale Ranns <nranns@cisco.com>
|
|
The issue surfaced when developing the tap GSO code, with
an iteration where output path is reliant on
vnet_buffer (b0)->l3_hdr_offset being set correctly in
the input path, during performance testing.
Adding a workaround in the TX path shows that
the issue surfaces only for relatively few packets
during the test (about 100 out of 600000).
Analysis shows the issue arises if the ethernet-input
is handling two untagged packets with different sw_if_index
values - then the accelerated path punts to slow path,
before the setting of the l2.l2_len values is done,
thus resulting in them being 0, and l3_hdr_offset being
the same as l2_hdr_offset, wreaking havoc on TX path.
The solution is to move the l2_hdr_offset calculation
into a place where it is done for all the packets,
and move the l3_hdr_offset calculation into
the determine_next_node() function - as that function is
also the one setting the special-case l2.l2_len value for
tagged packets and moving the current_data for the L2 case.
Change-Id: If728c7715e011930c1887691188c98055bddde67
Signed-off-by: Andrew Yourtchenko <ayourtch@gmail.com>
|
|
Change-Id: I3c84daf046dbad972b36e48fa2548bbe20c7b338
Signed-off-by: Vijayabhaskar Katamreddy <vkatamre@cisco.com>
|
|
Change-Id: I785ecadbf30812a500629870aa717e64f4cf0cdd
Signed-off-by: Neale Ranns <nranns@cisco.com>
|
|
active-backup mode is using l2 load balance algo. It should be using
active-backup. Also notice that the output is missing a character.
vpp# create bond mode active-backup
create bond mode active-backup
vpp# sh bond
sh bond
interface name sw_if_index mode load balance active slaves slaves
BondEthernet0 6 xor l34 2 2
BondEthernet1 9 xor l34 1 1
BondEthernet2 10 active-backu l2 0 0
vpp#
Change-Id: If5ed0cc6c25f6c2ddabec15ff6188b34923d38e3
Signed-off-by: Steven <sluong@cisco.com>
|
|
Introduce bond_tx_inline which takes lb as a constant for gcc to do the optimization
The number appears a tad better for 256 bytes frame.
with the patch
--------------
Thread 2 vpp_wk_1 (lcore 3)
Time 4.3, average vectors/node 224.00, last 128 main loops 40.00 per node 222.61
vector rates in 8.4836e6, out 1.6967e7, drop 0.0000e0, punt 0.0000e0
Name State Calls Vectors Suspends Clocks Vectors/Call
BondEthernet0-output active 141054 36109824 0 2.51e1 256.00
BondEthernet0-tx active 141054 36109824 0 2.55e1 256.00
TenGigabitEthernet6/0/0-output active 141054 18055469 0 9.43e0 128.00
TenGigabitEthernet6/0/0-tx active 141054 18055469 0 6.97e1 128.00
TenGigabitEthernet6/0/1-output active 141054 18054355 0 9.54e0 127.99
TenGigabitEthernet6/0/1-tx active 141054 18054355 0 7.05e1 127.99
bond-input active 141054 36109824 0 1.76e1 256.00
dpdk-input polling 70527 36109824 0 5.03e1 512.00
ethernet-input active 141054 36109824 0 6.12e1 256.00
ip4-input active 141054 36109824 0 3.26e1 256.00
ip4-lookup active 141054 36109824 0 2.94e1 256.00
ip4-rewrite active 141054 36109824 0 3.27e1 256.00
without the patch
-----------------
Thread 2 vpp_wk_1 (lcore 3)
Time 4.3, average vectors/node 224.00, last 128 main loops 40.00 per node 222.61
vector rates in 8.4443e6, out 1.6889e7, drop 0.0000e0, punt 0.0000e0
Name State Calls Vectors Suspends Clocks Vectors/Call
BondEthernet0-output active 142744 36542464 0 2.51e1 256.00
BondEthernet0-tx active 142744 36542464 0 2.67e1 256.00
TenGigabitEthernet6/0/0-output active 142744 18270813 0 9.19e0 127.99
TenGigabitEthernet6/0/0-tx active 142744 18270813 0 6.98e1 127.99
TenGigabitEthernet6/0/1-output active 142744 18271651 0 9.43e0 128.00
TenGigabitEthernet6/0/1-tx active 142744 18271651 0 7.02e1 128.00
bond-input active 142744 36542464 0 1.76e1 256.00
dpdk-input polling 71372 36542464 0 5.08e1 512.00
ethernet-input active 142744 36542464 0 6.15e1 256.00
ip4-input active 142744 36542464 0 3.23e1 256.00
ip4-lookup active 142744 36542464 0 2.96e1 256.00
ip4-rewrite active 142744 36542464 0 3.28e1 256.00
Change-Id: I9fd43eda3c735cbff680ac6d2f01ecdae81f0eda
Signed-off-by: Damjan Marion <damarion@cisco.com>
|
|
Change-Id: I0d42a0c71fea7dd669fb1fe5ded7e6e944245c7d
Signed-off-by: Florin Coras <fcoras@cisco.com>
|