- 03 Mar, 2022 5 commits
-
-
Christophe JAILLET authored
Use ida_alloc_xxx()/ida_free() instead to ida_simple_get()/ida_simple_remove(). The latter is deprecated and more verbose. Signed-off-by:
Christophe JAILLET <christophe.jaillet@wanadoo.fr> Signed-off-by:
Simon Horman <simon.horman@corigine.com> Link: https://lore.kernel.org/r/20220301131212.26348-1-simon.horman@corigine.comSigned-off-by:
Jakub Kicinski <kuba@kernel.org>
-
Russell King (Oracle) authored
Convert sfp to use %pe for printing error codes, which can print them as errno symbols rather than numbers. Signed-off-by:
Russell King (Oracle) <rmk+kernel@armlinux.org.uk> Reviewed-by:
Andrew Lunn <andrew@lunn.ch> Link: https://lore.kernel.org/r/E1nOyEN-00BuuE-OB@rmk-PC.armlinux.org.ukSigned-off-by:
-
Russell King (Oracle) authored
Convert phylink to use %pe for printing error codes, which can print them as errno symbols rather than numbers. Signed-off-by:
-
Harold Huang authored
In patch [1], tun_msg_ctl was added to allow pass batched xdp buffers to tun_sendmsg. Although we donot use msg_controllen in this path, we should check msg_controllen to make sure the caller pass a valid msg_ctl. [1]: https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=fe8dd45bb7556246c6b76277b1ba4296c91c2505Reported-by:
Eric Dumazet <eric.dumazet@gmail.com> Suggested-by:
Jason Wang <jasowang@redhat.com> Signed-off-by:
Harold Huang <baymaxhuang@gmail.com> Acked-by:
-
git://git.open-mesh.org/linux-mergeJakub Kicinski authored
Simon Wunderlich says: ==================== This cleanup patchset includes the following patches: - bump version strings, by Simon Wunderlich - Remove redundant 'flush_workqueue()' calls, by Christophe JAILLET - Migrate to linux/container_of.h, by Sven Eckelmann - Demote batadv-on-batadv skip error message, by Sven Eckelmann * tag 'batadv-next-pullrequest-20220302' of git://git.open-mesh.org/linux-merge: batman-adv: Demote batadv-on-batadv skip error message batman-adv: Migrate to linux/container_of.h batman-adv: Remove redundant 'flush_workqueue()' calls batman-adv: Start new development cycle ==================== Link: https://lore.kernel.org/r/20220302163522.102842-1-sw@simonwunderlich.deSigned-off-by:
-
- 02 Mar, 2022 13 commits
-
-
Wang Qing authored
add missing ")" which caused by previous commit. Fixes: 61c4fb9c ("net: hamradio: use time_is_after_jiffies() instead of open coding it") Link: https://lore.kernel.org/all/1646018012-61129-1-git-send-email-wangqing@vivo.com/Signed-off-by:
Wang Qing <wangqing@vivo.com> Link: https://lore.kernel.org/r/1646203277-83159-1-git-send-email-wangqing@vivo.comSigned-off-by:
-
Sven Eckelmann authored
The error message "Cannot find parent device" was shown for users of macvtap (on batadv devices) whenever the macvtap was moved to a different netns. This happens because macvtap doesn't provide an implementation for rtnl_link_ops->get_link_net. The situation for which this message is printed is actually not an error but just a warning that the optional sanity check was skipped. So demote the message from error to warning and adjust the text to better explain what happened. Reported-by:
Leonardo Mörlein <freifunk@irrelefant.net> Signed-off-by:
Sven Eckelmann <sven@narfation.org> Signed-off-by:
Simon Wunderlich <sw@simonwunderlich.de>
-
Sven Eckelmann authored
The commit d2a8ebbf ("kernel.h: split out container_of() and typeof_member() macros") introduced a new header for the container_of related macros from (previously) linux/kernel.h. Signed-off-by:
-
Jakub Kicinski authored
Daniel Braunwarth says: ==================== if_ether.h: add industrial fieldbus Ethertypes This set of patches adds the Ethertypes for PROFINET and EtherCAT. The defines should be used by iproute2 to extend the list of available link layer protocols. ==================== Link: https://lore.kernel.org/r/20220228133029.100913-1-daniel@braunwarth.devSigned-off-by:
-
Daniel Braunwarth authored
Add the Ethertype for EtherCAT protocol. Signed-off-by:
Daniel Braunwarth <daniel@braunwarth.dev> Signed-off-by:
-
Daniel Braunwarth authored
Add the Ethertype for PROFINET protocol. Signed-off-by:
-
Sven Eckelmann authored
Assign rtnl_link_ops->get_link_net() callback so that IFLA_LINK_NETNSID is added to rtnetlink messages. This fixes iproute2 which otherwise resolved the link interface to an interface in the wrong namespace. Test commands: ip netns add nst ip link add dummy0 type dummy ip link add link macvtap0 link dummy0 type macvtap ip link set macvtap0 netns nst ip -netns nst link show macvtap0 Before: 10: macvtap0@gre0: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state DOWN mode DEFAULT group default qlen 500 link/ether 5e:8f:ae:1d:60:50 brd ff:ff:ff:ff:ff:ff After: 10: macvtap0@if2: <BROADCAST,MULTICAST> mtu 1500 qdisc noop state DOWN mode DEFAULT group default qlen 500 link/ether 5e:8f:ae:1d:60:50 brd ff:ff:ff:ff:ff:ff link-netnsid 0 Reported-by: -
Wan Jiabing authored
Fix the following coccicheck warning: ./drivers/net/ethernet/netronome/nfp/flower/qos_conf.c:750:7-55: WARNING avoid newline at end of message in NL_SET_ERR_MSG_MOD Signed-off-by:
Wan Jiabing <wanjiabing@vivo.com> Reviewed-by:
-
Harold Huang authored
In tun, NAPI is supported and we can also use NAPI in the path of batched XDP buffs to accelerate packet processing. What is more, after we use NAPI, GRO is also supported. The iperf shows that the throughput of single stream could be improved from 4.5Gbps to 9.2Gbps. Additionally, 9.2 Gbps nearly reachs the line speed of the phy nic and there is still about 15% idle cpu core remaining on the vhost thread. Test topology: [iperf server]<--->tap<--->dpdk testpmd<--->phy nic<--->[iperf client] Iperf stream: iperf3 -c 10.0.0.2 -i 1 -t 10 Before: ... [ 5] 5.00-6.00 sec 558 MBytes 4.68 Gbits/sec 0 1.50 MBytes [ 5] 6.00-7.00 sec 556 MBytes 4.67 Gbits/sec 1 1.35 MBytes [ 5] 7.00-8.00 sec 556 MBytes 4.67 Gbits/sec 2 1.18 MBytes [ 5] 8.00-9.00 sec 559 MBytes 4.69 Gbits/sec 0 1.48 MBytes [ 5] 9.00-10.00 sec 556 MBytes 4.67 Gbits/sec 1 1.33 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 5.39 GBytes 4.63 Gbits/sec 72 sender [ 5] 0.00-10.04 sec 5.39 GBytes 4.61 Gbits/sec receiver After: ... [ 5] 5.00-6.00 sec 1.07 GBytes 9.19 Gbits/sec 0 1.55 MBytes [ 5] 6.00-7.00 sec 1.08 GBytes 9.30 Gbits/sec 0 1.63 MBytes [ 5] 7.00-8.00 sec 1.08 GBytes 9.25 Gbits/sec 0 1.72 MBytes [ 5] 8.00-9.00 sec 1.08 GBytes 9.25 Gbits/sec 77 1.31 MBytes [ 5] 9.00-10.00 sec 1.08 GBytes 9.24 Gbits/sec 0 1.48 MBytes - - - - - - - - - - - - - - - - - - - - - - - - - [ ID] Interval Transfer Bitrate Retr [ 5] 0.00-10.00 sec 10.8 GBytes 9.28 Gbits/sec 166 sender [ 5] 0.00-10.04 sec 10.8 GBytes 9.24 Gbits/sec receiver Reported-at: https://lore.kernel.org/all/CACGkMEvTLG0Ayg+TtbN4q4pPW-ycgCCs3sC3-TF8cuRTf7Pp1A@mail.gmail.comSigned-off-by:
-
Jakub Kicinski authored
Íñigo Huguet says: ==================== sfc: optimize RXQs count and affinities In sfc driver one RX queue per physical core was allocated by default. Later on, IRQ affinities were set spreading the IRQs in all NUMA local CPUs. However, with that default configuration it result in a non very optimal configuration in many modern systems. Specifically, in systems with hyper threading and 2 NUMA nodes, affinities are set in a way that IRQs are handled by all logical cores of one same NUMA node. Handling IRQs from both hyper threading siblings has no benefit, and setting affinities to one queue per physical core is neither a very good idea because there is a performance penalty for moving data across nodes (I was able to check it with some XDP tests using pktgen). This patches reduce the default number of channels to one per physical core in the local NUMA node. Then, they set IRQ affinities to CPUs in the local NUMA node only. This way we save hardware resources since channels are limited resources. We also leave more room for XDP_TX channels without hitting driver's limit of 32 channels per interface. Running performance tests using iperf with a SFC9140 device showed no performance penalty for reducing the number of channels. RX XDP tests showed that performance can go down to less than half if the IRQ is handled by a CPU in a different NUMA node, which doesn't happen with the new defaults from this patches. ==================== Link: https://lore.kernel.org/r/20220228132254.25787-1-ihuguet@redhat.comSigned-off-by:
-
Íñigo Huguet authored
Affinity hints were being set to CPUs in local NUMA node first, and then in other CPUs. This was creating 2 unintended issues: 1. Channels created to be assigned each to a different physical core were assigned to hyperthreading siblings because of being in same NUMA node. Since the patch previous to this one, this did not longer happen with default rss_cpus modparam because less channels are created. 2. XDP channels could be assigned to CPUs in different NUMA nodes, decreasing performance too much (to less than half in some of my tests). This patch sets the affinity hints spreading the channels only in local NUMA node's CPUs. A fallback for the case that no CPU in local NUMA node is online has been added too. Example of CPUs being assigned in a non optimal way before this and the previous patch (note: in this system, xdp-8 to xdp-15 are created because num_possible_cpus == 64, but num_present_cpus == 32 so they're never used): $ lscpu | grep -i numa NUMA node(s): 2 NUMA node0 CPU(s): 0-7,16-23 NUMA node1 CPU(s): 8-15,24-31 $ grep -H . /proc/irq/*/0000:07:00.0*/../smp_affinity_list /proc/irq/141/0000:07:00.0-0/../smp_affinity_list:0 /proc/irq/142/0000:07:00.0-1/../smp_affinity_list:1 /proc/irq/143/0000:07:00.0-2/../smp_affinity_list:2 /proc/irq/144/0000:07:00.0-3/../smp_affinity_list:3 /proc/irq/145/0000:07:00.0-4/../smp_affinity_list:4 /proc/irq/146/0000:07:00.0-5/../smp_affinity_list:5 /proc/irq/147/0000:07:00.0-6/../smp_affinity_list:6 /proc/irq/148/0000:07:00.0-7/../smp_affinity_list:7 /proc/irq/149/0000:07:00.0-8/../smp_affinity_list:16 /proc/irq/150/0000:07:00.0-9/../smp_affinity_list:17 /proc/irq/151/0000:07:00.0-10/../smp_affinity_list:18 /proc/irq/152/0000:07:00.0-11/../smp_affinity_list:19 /proc/irq/153/0000:07:00.0-12/../smp_affinity_list:20 /proc/irq/154/0000:07:00.0-13/../smp_affinity_list:21 /proc/irq/155/0000:07:00.0-14/../smp_affinity_list:22 /proc/irq/156/0000:07:00.0-15/../smp_affinity_list:23 /proc/irq/157/0000:07:00.0-xdp-0/../smp_affinity_list:8 /proc/irq/158/0000:07:00.0-xdp-1/../smp_affinity_list:9 /proc/irq/159/0000:07:00.0-xdp-2/../smp_affinity_list:10 /proc/irq/160/0000:07:00.0-xdp-3/../smp_affinity_list:11 /proc/irq/161/0000:07:00.0-xdp-4/../smp_affinity_list:12 /proc/irq/162/0000:07:00.0-xdp-5/../smp_affinity_list:13 /proc/irq/163/0000:07:00.0-xdp-6/../smp_affinity_list:14 /proc/irq/164/0000:07:00.0-xdp-7/../smp_affinity_list:15 /proc/irq/165/0000:07:00.0-xdp-8/../smp_affinity_list:24 /proc/irq/166/0000:07:00.0-xdp-9/../smp_affinity_list:25 /proc/irq/167/0000:07:00.0-xdp-10/../smp_affinity_list:26 /proc/irq/168/0000:07:00.0-xdp-11/../smp_affinity_list:27 /proc/irq/169/0000:07:00.0-xdp-12/../smp_affinity_list:28 /proc/irq/170/0000:07:00.0-xdp-13/../smp_affinity_list:29 /proc/irq/171/0000:07:00.0-xdp-14/../smp_affinity_list:30 /proc/irq/172/0000:07:00.0-xdp-15/../smp_affinity_list:31 CPUs assignments after this and previous patch, so normal channels created only one per core in NUMA node and affinities set only to local NUMA node: $ grep -H . /proc/irq/*/0000:07:00.0*/../smp_affinity_list /proc/irq/116/0000:07:00.0-0/../smp_affinity_list:0 /proc/irq/117/0000:07:00.0-1/../smp_affinity_list:1 /proc/irq/118/0000:07:00.0-2/../smp_affinity_list:2 /proc/irq/119/0000:07:00.0-3/../smp_affinity_list:3 /proc/irq/120/0000:07:00.0-4/../smp_affinity_list:4 /proc/irq/121/0000:07:00.0-5/../smp_affinity_list:5 /proc/irq/122/0000:07:00.0-6/../smp_affinity_list:6 /proc/irq/123/0000:07:00.0-7/../smp_affinity_list:7 /proc/irq/124/0000:07:00.0-xdp-0/../smp_affinity_list:16 /proc/irq/125/0000:07:00.0-xdp-1/../smp_affinity_list:17 /proc/irq/126/0000:07:00.0-xdp-2/../smp_affinity_list:18 /proc/irq/127/0000:07:00.0-xdp-3/../smp_affinity_list:19 /proc/irq/128/0000:07:00.0-xdp-4/../smp_affinity_list:20 /proc/irq/129/0000:07:00.0-xdp-5/../smp_affinity_list:21 /proc/irq/130/0000:07:00.0-xdp-6/../smp_affinity_list:22 /proc/irq/131/0000:07:00.0-xdp-7/../smp_affinity_list:23 /proc/irq/132/0000:07:00.0-xdp-8/../smp_affinity_list:0 /proc/irq/133/0000:07:00.0-xdp-9/../smp_affinity_list:1 /proc/irq/134/0000:07:00.0-xdp-10/../smp_affinity_list:2 /proc/irq/135/0000:07:00.0-xdp-11/../smp_affinity_list:3 /proc/irq/136/0000:07:00.0-xdp-12/../smp_affinity_list:4 /proc/irq/137/0000:07:00.0-xdp-13/../smp_affinity_list:5 /proc/irq/138/0000:07:00.0-xdp-14/../smp_affinity_list:6 /proc/irq/139/0000:07:00.0-xdp-15/../smp_affinity_list:7 Signed-off-by:
Íñigo Huguet <ihuguet@redhat.com> Acked-by:
Martin Habets <habetsm.xilinx@gmail.com> Signed-off-by:
-
Íñigo Huguet authored
Handling channels from CPUs in different NUMA node can penalize performance, so better configure only one channel per core in the same NUMA node than the NIC, and not per each core in the system. Fallback to all other online cores if there are not online CPUs in local NUMA node. Signed-off-by:
-
Jakub Kicinski authored
Fix build: include/linux/minmax.h:45:25: note: in expansion of macro ‘__careful_cmp’ 45 | #define min(x, y) __careful_cmp(x, y, <) | ^~~~~~~~~~~~~ net/smc/smc_tx.c:150:24: note: in expansion of macro ‘min’ 150 | corking_size = min(sock_net(&smc->sk)->smc.sysctl_autocorking_size, | ^~~ Fixes: 12bbb0d1 ("net/smc: add sysctl for autocorking") Link: https://lore.kernel.org/r/20220301222446.1271127-1-kuba@kernel.orgSigned-off-by:
-
- 01 Mar, 2022 22 commits
-
-
David S. Miller authored
Dust Li says: ==================== net/smc: some datapath performance optimizations This series tries to improve the performance of SMC in datapath. - patch #1, add sysctl interface to support tuning the behaviour of SMC in container environment. - patch #2/#3, add autocorking support which is very efficient for small messages without trade-off for latency. - patch #4, send directly on setting TCP_NODELAY, without wake up the TX worker, this make it consistent with clearing TCP_CORK. - patch #5, this correct the setting of RMB window update limit, so we don't send CDC messages to update peer's RMB window too frequently in some cases. - patch #6, implemented something like NAPI in SMC, decrease the number of hardirq when busy. - patch #7, this moves TX work doing in the BH to the user context when sock_lock is hold by user. With this patchset applied, we can get a good performance gain: - qperf tcp_bw test has shown a great improvement. Other benchmarks like 'netperf TCP_STREAM' or 'sockperf throughput' has similar result. - In my testing environment, running qperf tcp_bw and tcp_lat, SMC behaves better then TCP in most all message size. Here are some test results with the following testing command: client: smc_run taskset -c 1 qperf smc-server -oo msg_size:1:64K:*2 \ -t 30 -vu tcp_{bw|lat} server: smc_run taskset -c 1 qperf ==== Bandwidth ==== MsgSize Origin SMC TCP SMC with patches 1 0.578 MB/s 2.392 MB/s(313.57%) 2.561 MB/s(342.83%) 2 1.159 MB/s 4.780 MB/s(312.53%) 5.162 MB/s(345.46%) 4 2.283 MB/s 10.266 MB/s(349.77%) 10.122 MB/s(343.46%) 8 4.668 MB/s 19.040 MB/s(307.86%) 20.521 MB/s(339.59%) 16 9.147 MB/s 38.904 MB/s(325.31%) 40.823 MB/s(346.29%) 32 18.369 MB/s 79.587 MB/s(333.25%) 80.535 MB/s(338.42%) 64 36.562 MB/s 148.668 MB/s(306.61%) 158.170 MB/s(332.60%) 128 72.961 MB/s 274.913 MB/s(276.80%) 316.217 MB/s(333.41%) 256 144.705 MB/s 512.059 MB/s(253.86%) 626.019 MB/s(332.62%) 512 288.873 MB/s 884.977 MB/s(206.35%) 1221.596 MB/s(322.88%) 1024 574.180 MB/s 1337.736 MB/s(132.98%) 2203.156 MB/s(283.70%) 2048 1095.192 MB/s 1865.952 MB/s( 70.38%) 3036.448 MB/s(177.25%) 4096 2066.157 MB/s 2380.337 MB/s( 15.21%) 3834.271 MB/s( 85.58%) 8192 3717.198 MB/s 2733.073 MB/s(-26.47%) 4904.910 MB/s( 31.95%) 16384 4742.221 MB/s 2958.693 MB/s(-37.61%) 5220.272 MB/s( 10.08%) 32768 5349.550 MB/s 3061.285 MB/s(-42.77%) 5321.865 MB/s( -0.52%) 65536 5162.919 MB/s 3731.408 MB/s(-27.73%) 5245.021 MB/s( 1.59%) ==== Latency ==== MsgSize Origin SMC TCP SMC with patches 1 10.540 us 11.938 us( 13.26%) 10.356 us( -1.75%) 2 10.996 us 11.992 us( 9.06%) 10.073 us( -8.39%) 4 10.229 us 11.687 us( 14.25%) 9.996 us( -2.28%) 8 10.203 us 11.653 us( 14.21%) 10.063 us( -1.37%) 16 10.530 us 11.313 us( 7.44%) 10.013 us( -4.91%) 32 10.241 us 11.586 us( 13.13%) 10.081 us( -1.56%) 64 10.693 us 11.652 us( 8.97%) 9.986 us( -6.61%) 128 10.597 us 11.579 us( 9.27%) 10.262 us( -3.16%) 256 10.409 us 11.957 us( 14.87%) 10.148 us( -2.51%) 512 11.088 us 12.505 us( 12.78%) 10.206 us( -7.95%) 1024 11.240 us 12.255 us( 9.03%) 10.631 us( -5.42%) 2048 11.485 us 16.970 us( 47.76%) 10.981 us( -4.39%) 4096 12.077 us 13.948 us( 15.49%) 11.847 us( -1.90%) 8192 13.683 us 16.693 us( 22.00%) 13.336 us( -2.54%) 16384 16.470 us 23.615 us( 43.38%) 16.519 us( 0.30%) 32768 22.540 us 40.966 us( 81.75%) 22.452 us( -0.39%) 65536 34.192 us 73.003 us(113.51%) 33.916 us( -0.81%) ------------ Test environment notes: 1. Testing is run on 2 VMs within the same physical host 2. The NIC is ConnectX-4Lx, using SRIOV, and passing through 2 VFs to the 2 VMs respectively. 3. To decrease jitter, VM's vCPU are binded to each physical CPU, and those physical CPUs are all isolated using boot parameter `isolcpus=xxx` 4. The queue number are set to 1, and interrupt from the queue is binded to CPU0 in the guest ==================== Signed-off-by:David S. Miller <davem@davemloft.net>
-
Dust Li authored
Send data all the way down to the RDMA device is a time consuming operation(get a new slot, maybe do RDMA Write and send a CDC, etc). Moving those operations from BH to user context is good for performance. If the sock_lock is hold by user, we don't try to send data out in the BH context, but just mark we should send. Since the user will release the sock_lock soon, we can do the sending there. Add smc_release_cb() which will be called in release_sock() and try send in the callback if needed. This patch moves the sending part out from BH if sock lock is hold by user. In my testing environment, this saves about 20% softirq in the qperf 4K tcp_bw test in the sender side with no noticeable throughput drop. Signed-off-by:
Dust Li <dust.li@linux.alibaba.com> Signed-off-by:
-
Dust Li authored
When we are handling softirq workload, enable hardirq may again interrupt the current routine of softirq, and then try to raise softirq again. This only wastes CPU cycles and won't have any real gain. Since IB_CQ_REPORT_MISSED_EVENTS already make sure if ib_req_notify_cq() returns 0, it is safe to wait for the next event, with no need to poll the CQ again in this case. This patch disables hardirq during the processing of softirq, and re-arm the CQ after softirq is done. Somehow like NAPI. Co-developed-by:
Guangguan Wang <guangguan.wang@linux.alibaba.com> Signed-off-by:
-
Dust Li authored
rmbe_update_limit is used to limit announcing receive window updating too frequently. RFC7609 request a minimal increase in the window size of 10% of the receive buffer space. But current implementation used: min_t(int, rmbe_size / 10, SOCK_MIN_SNDBUF / 2) and SOCK_MIN_SNDBUF / 2 == 2304 Bytes, which is almost always less then 10% of the receive buffer space. This causes the receiver always sending CDC message to update its consumer cursor when it consumes more then 2K of data. And as a result, we may encounter something like "TCP silly window syndrome" when sending 2.5~8K message. This patch fixes this using max(rmbe_size / 10, SOCK_MIN_SNDBUF / 2). With this patch and SMC autocorking enabled, qperf 2K/4K/8K tcp_bw test shows 45%/75%/40% increase in throughput respectively. Signed-off-by:
-
Dust Li authored
In commit ea785a1a("net/smc: Send directly when TCP_CORK is cleared"), we don't use delayed work to implement cork. This patch use the same algorithm, removes the delayed work when setting TCP_NODELAY and send directly in setsockopt(). This also makes the TCP_NODELAY the same as TCP. Cc: Tony Lu <tonylu@linux.alibaba.com> Signed-off-by:
-
Dust Li authored
This add a new sysctl: net.smc.autocorking_size We can dynamically change the behaviour of autocorking by change the value of autocorking_size. Setting to 0 disables autocorking in SMC Signed-off-by:
-
Dust Li authored
This patch adds autocorking support for SMC which could improve throughput for small message by x3+. The main idea is borrowed from TCP autocorking with some RDMA specific modification: 1. The first message should never cork to make sure we won't bring extra latency 2. If we have posted any Tx WRs to the NIC that have not completed, cork the new messages until: a) Receive CQE for the last Tx WR b) We have corked enough message on the connection 3. Try to push the corked data out when we receive CQE of the last Tx WR to prevent the corked messages hang in the send queue. Both SMC autocorking and TCP autocorking check the TX completion to decide whether we should cork or not. The difference is when we got a SMC Tx WR completion, the data have been confirmed by the RNIC while TCP TX completion just tells us the data have been sent out by the local NIC. Add an atomic variable tx_pushing in smc_connection to make sure only one can send to let it cork more and save CDC slot. SMC autocorking should not bring extra latency since the first message will always been sent out immediately. The qperf tcp_bw test shows more than x4 increase under small message size with Mellanox connectX4-Lx, same result with other throughput benchmarks like sockperf/netperf. The qperf tcp_lat test shows SMC autocorking has not increase any ping-pong latency. Test command: client: smc_run taskset -c 1 qperf smc-server -oo msg_size:1:64K:*2 \ -t 30 -vu tcp_{bw|lat} server: smc_run taskset -c 1 qperf === Bandwidth ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 0.578 MB/s 2.392 MB/s(313.57%) 2.647 MB/s(357.72%) 2 1.159 MB/s 4.780 MB/s(312.53%) 5.153 MB/s(344.71%) 4 2.283 MB/s 10.266 MB/s(349.77%) 10.363 MB/s(354.02%) 8 4.668 MB/s 19.040 MB/s(307.86%) 21.215 MB/s(354.45%) 16 9.147 MB/s 38.904 MB/s(325.31%) 41.740 MB/s(356.32%) 32 18.369 MB/s 79.587 MB/s(333.25%) 82.392 MB/s(348.52%) 64 36.562 MB/s 148.668 MB/s(306.61%) 161.564 MB/s(341.89%) 128 72.961 MB/s 274.913 MB/s(276.80%) 325.363 MB/s(345.94%) 256 144.705 MB/s 512.059 MB/s(253.86%) 633.743 MB/s(337.96%) 512 288.873 MB/s 884.977 MB/s(206.35%) 1250.681 MB/s(332.95%) 1024 574.180 MB/s 1337.736 MB/s(132.98%) 2246.121 MB/s(291.19%) 2048 1095.192 MB/s 1865.952 MB/s( 70.38%) 2057.767 MB/s( 87.89%) 4096 2066.157 MB/s 2380.337 MB/s( 15.21%) 2173.983 MB/s( 5.22%) 8192 3717.198 MB/s 2733.073 MB/s(-26.47%) 3491.223 MB/s( -6.08%) 16384 4742.221 MB/s 2958.693 MB/s(-37.61%) 4637.692 MB/s( -2.20%) 32768 5349.550 MB/s 3061.285 MB/s(-42.77%) 5385.796 MB/s( 0.68%) 65536 5162.919 MB/s 3731.408 MB/s(-27.73%) 5223.890 MB/s( 1.18%) ==== Latency ==== MsgSize(Bytes) SMC-NoCork TCP SMC-AutoCorking 1 10.540 us 11.938 us( 13.26%) 10.573 us( 0.31%) 2 10.996 us 11.992 us( 9.06%) 10.269 us( -6.61%) 4 10.229 us 11.687 us( 14.25%) 10.240 us( 0.11%) 8 10.203 us 11.653 us( 14.21%) 10.402 us( 1.95%) 16 10.530 us 11.313 us( 7.44%) 10.599 us( 0.66%) 32 10.241 us 11.586 us( 13.13%) 10.223 us( -0.18%) 64 10.693 us 11.652 us( 8.97%) 10.251 us( -4.13%) 128 10.597 us 11.579 us( 9.27%) 10.494 us( -0.97%) 256 10.409 us 11.957 us( 14.87%) 10.710 us( 2.89%) 512 11.088 us 12.505 us( 12.78%) 10.547 us( -4.88%) 1024 11.240 us 12.255 us( 9.03%) 10.787 us( -4.03%) 2048 11.485 us 16.970 us( 47.76%) 11.256 us( -1.99%) 4096 12.077 us 13.948 us( 15.49%) 12.230 us( 1.27%) 8192 13.683 us 16.693 us( 22.00%) 13.786 us( 0.75%) 16384 16.470 us 23.615 us( 43.38%) 16.459 us( -0.07%) 32768 22.540 us 40.966 us( 81.75%) 23.284 us( 3.30%) 65536 34.192 us 73.003 us(113.51%) 34.233 us( 0.12%) With SMC autocorking support, we can archive better throughput than TCP in most message sizes without any latency trade-off. Signed-off-by: -
Dust Li authored
This patch add sysctl interface to support container environment for SMC as we talk in the mail list. Link: https://lore.kernel.org/netdev/20220224020253.GF5443@linux.alibaba.comCo-developed-by:
Tony Lu <tonylu@linux.alibaba.com> Signed-off-by:
-
David S. Miller authored
Roopa Prabhu says: ==================== vxlan metadata device vnifiltering support This series adds vnifiltering support to vxlan collect metadata device. Motivation: You can only use a single vxlan collect metadata device for a given vxlan udp port in the system today. The vxlan collect metadata device terminates all received vxlan packets. As shown in the below diagram, there are use-cases where you need to support multiple such vxlan devices in independent bridge domains. Each vxlan device must terminate the vni's it is configured for. Example usecase: In a service provider network a service provider typically supports multiple bridge domains with overlapping vlans. One bridge domain per customer. Vlans in each bridge domain are mapped to globally unique vxlan ranges assigned to each customer. This series adds vnifiltering support to collect metadata devices to terminate only configured vnis. This is similar to vlan filtering in bridge driver. The vni filtering capability is provided by a new flag on collect metadata device. In the below pic: - customer1 is mapped to br1 bridge domain - customer2 is mapped to br2 bridge domain - customer1 vlan 10-11 is mapped to vni 1001-1002 - customer2 vlan 10-11 is mapped to vni 2001-2002 - br1 and br2 are vlan filtering bridges - vxlan1 and vxlan2 are collect metadata devices with vnifiltering enabled ┌──────────────────────────────────────────────────────────────────┐ │ switch │ │ │ │ ┌───────────┐ ┌───────────┐ │ │ │ │ │ │ │ │ │ br1 │ │ br2 │ │ │ └┬─────────┬┘ └──┬───────┬┘ │ │ vlans│ │ vlans │ │ │ │ 10,11│ │ 10,11│ │ │ │ │ vlanvnimap: │ vlanvnimap: │ │ │ 10-1001,11-1002 │ 10-2001,11-2002 │ │ │ │ │ │ │ │ ┌──────┴┐ ┌──┴─────────┐ ┌───┴────┐ │ │ │ │ swp1 │ │vxlan1 │ │ swp2 │ ┌┴─────────────┐ │ │ │ │ │ vnifilter:│ │ │ │vxlan2 │ │ │ └───┬───┘ │ 1001,1002│ └───┬────┘ │ vnifilter: │ │ │ │ └────────────┘ │ │ 2001,2002 │ │ │ │ │ └──────────────┘ │ │ │ │ │ └───────┼──────────────────────────────────┼───────────────────────┘ │ │ │ │ ┌─────┴───────┐ │ │ customer1 │ ┌─────┴──────┐ │ host/VM │ │customer2 │ └─────────────┘ │ host/VM │ └────────────┘ v2: - remove stale xstats declarations pointed out by Nikolay Aleksandrov - squash selinux patch with the tunnel api patch as pointed out by benjamin poirier - Fix various build issues: Reported-by:kernel test robot <lkp@intel.com> v3: - incorporate review feedback from Jakub - move rhashtable declarations to c file - define and use netlink policy for top level vxlan filter api - fix unused stats function warning - pass vninode from vnifilter lookup into stats count function to avoid another lookup (only applicable to vxlan_rcv) - fix missing vxlan vni delete notifications in vnifilter uninit function - misc cleanups - remote dev check for multicast groups added via vnifiltering api ==================== Signed-off-by:
-
Nikolay Aleksandrov authored
Add support for VXLAN vni filter entries' stats dumping Signed-off-by:
Nikolay Aleksandrov <nikolay@nvidia.com> Signed-off-by:
-
Nikolay Aleksandrov authored
Add per-vni statistics for vni filter mode. Counting Rx/Tx bytes/packets/drops/errors at the appropriate places. This patch changes vxlan_vs_find_vni to also return the vxlan_vni_node in cases where the vni belongs to a vni filtering vxlan device Signed-off-by:
Roopa Prabhu <roopa@nvidia.com> Signed-off-by:
-
Roopa Prabhu authored
This patch adds a new test script test_vxlan_vnifiltering.sh with tests for vni filtering api, various datapath tests. Also has a test with a mix of traditional, metadata and vni filtering devices inuse at the same time. Signed-off-by:
-
Roopa Prabhu authored
This patch adds vnifiltering support to collect metadata device. Motivation: You can only use a single vxlan collect metadata device for a given vxlan udp port in the system today. The vxlan collect metadata device terminates all received vxlan packets. As shown in the below diagram, there are use-cases where you need to support multiple such vxlan devices in independent bridge domains. Each vxlan device must terminate the vni's it is configured for. Example usecase: In a service provider network a service provider typically supports multiple bridge domains with overlapping vlans. One bridge domain per customer. Vlans in each bridge domain are mapped to globally unique vxlan ranges assigned to each customer. vnifiltering support in collect metadata devices terminates only configured vnis. This is similar to vlan filtering in bridge driver. The vni filtering capability is provided by a new flag on collect metadata device. In the below pic: - customer1 is mapped to br1 bridge domain - customer2 is mapped to br2 bridge domain - customer1 vlan 10-11 is mapped to vni 1001-1002 - customer2 vlan 10-11 is mapped to vni 2001-2002 - br1 and br2 are vlan filtering bridges - vxlan1 and vxlan2 are collect metadata devices with vnifiltering enabled ┌──────────────────────────────────────────────────────────────────┐ │ switch │ │ │ │ ┌───────────┐ ┌───────────┐ │ │ │ │ │ │ │ │ │ br1 │ │ br2 │ │ │ └┬─────────┬┘ └──┬───────┬┘ │ │ vlans│ │ vlans │ │ │ │ 10,11│ │ 10,11│ │ │ │ │ vlanvnimap: │ vlanvnimap: │ │ │ 10-1001,11-1002 │ 10-2001,11-2002 │ │ │ │ │ │ │ │ ┌──────┴┐ ┌──┴─────────┐ ┌───┴────┐ │ │ │ │ swp1 │ │vxlan1 │ │ swp2 │ ┌┴─────────────┐ │ │ │ │ │ vnifilter:│ │ │ │vxlan2 │ │ │ └───┬───┘ │ 1001,1002│ └───┬────┘ │ vnifilter: │ │ │ │ └────────────┘ │ │ 2001,2002 │ │ │ │ │ └──────────────┘ │ │ │ │ │ └───────┼──────────────────────────────────┼───────────────────────┘ │ │ │ │ ┌─────┴───────┐ │ │ customer1 │ ┌─────┴──────┐ │ host/VM │ │customer2 │ └─────────────┘ │ host/VM │ └────────────┘ With this implementation, vxlan dst metadata device can be associated with range of vnis. struct vxlan_vni_node is introduced to represent a configured vni. We start with vni and its associated remote_ip in this structure. This structure can be extended to bring in other per vni attributes if there are usecases for it. A vni inherits an attribute from the base vxlan device if there is no per vni attributes defined. struct vxlan_dev gets a new rhashtable for vnis called vxlan_vni_group. vxlan_vnifilter.c implements the necessary netlink api, notifications and helper functions to process and manage lifecycle of vxlan_vni_node. This patch also adds new helper functions in vxlan_multicast.c to handle per vni remote_ip multicast groups which are part of vxlan_vni_group. Fix build problems: Reported-by: -
Roopa Prabhu authored
subsequent patches will add more helpers. Signed-off-by:
-
Roopa Prabhu authored
This patch adds new rtm tunnel msg and api for tunnel id filtering in dst_metadata devices. First dst_metadata device to use the api is vxlan driver with AF_BRIDGE family. This and later changes add ability in vxlan driver to do tunnel id filtering (or vni filtering) on dst_metadata devices. This is similar to vlan api in the vlan filtering bridge. this patch includes selinux nlmsg_route_perms support for RTM_*TUNNEL api from Benjamin Poirier. Signed-off-by:
-
Roopa Prabhu authored
more users in follow up patches Signed-off-by:
-
Roopa Prabhu authored
This patch changes multicast helpers to take rip and ifindex as input. This is needed in future patches where rip can come from a pervni structure while the ifindex can come from the vxlan device. Signed-off-by:
-
Roopa Prabhu authored
This patch moves some fdb helpers to non-static for use in later patches. Ideally, all fdb code could move into its own file vxlan_fdb.c. This can be done as a subsequent patch and is out of scope of this series. Signed-off-by:
-
Roopa Prabhu authored
This patch moves common structures and global declarations to a shared private headerfile vxlan_private.h. Subsequent patches use this header file as a common header file for additional shared declarations. Signed-off-by:
-
Roopa Prabhu authored
Fix the below build warnings reported by kernel test robot: - initialize vni in vxlan_xmit_one - wrap label in ipv6 enabled checks in vxlan_xmit_one warnings: static drivers/net/vxlan/vxlan_core.c:2437:14: warning: variable 'label' set but not used [-Wunused-but-set-variable] __be32 vni, label; ^ >> drivers/net/vxlan/vxlan_core.c:2483:7: warning: variable 'vni' is used uninitialized whenever 'if' condition is true [-Wsometimes-uninitialized] Reported-by: -
Roopa Prabhu authored
vxlan.c has grown too long. This patch moves it to its own directory. subsequent patches add new functionality in new files. Signed-off-by:
-
git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linuxJakub Kicinski authored
Saeed Mahameed says: ==================== mlx5-next 2022-22-02 The following PR includes updates to mlx5-next branch: Headlines: ========== 1) Jakub cleans up unused static inline functions 2) I did some low level firmware command interface return status changes to provide the caller with full visibility on the error/status returned by the Firmware. 3) Use the new command interface in RDMA DEVX usecases to avoid flooding dmesg with some "expected" user error prone use cases. 4) Moshe also uses the new command interface to grab the specific error code from MFRL register command to provide the exact error reason for why SW reset couldn't perform internally in FW. 5) From Mark Bloch: Lag, drop packets in hardware when possible In active-backup mode the inactive interface's packets are dropped by the bond device. In switchdev where TC rules are offloaded to the FDB this can lead to packets being hit in the FDB where without offload they would have been dropped before reaching TC rules in the kernel. Create a drop rule to make sure packets on inactive ports are dropped before reaching the FDB. Listen on NETDEV_CHANGEUPPER / NETDEV_CHANGEINFODATA events and record the inactive state and offload accordingly. * 'mlx5-next' of git://git.kernel.org/pub/scm/linux/kernel/git/mellanox/linux: net/mlx5: Add clarification on sync reset failure net/mlx5: Add reset_state field to MFRL register RDMA/mlx5: Use new command interface API net/mlx5: cmdif, Refactor error handling and reporting of async commands net/mlx5: Use mlx5_cmd_do() in core create_{cq,dct} net/mlx5: cmdif, Add new api for command execution net/mlx5: cmdif, cmd_check refactoring net/mlx5: cmdif, Return value improvements net/mlx5: Lag, offload active-backup drops to hardware net/mlx5: Lag, record inactive state of bond device net/mlx5: Lag, don't use magic numbers for ports net/mlx5: Lag, use local variable already defined to access E-Switch net/mlx5: E-switch, add drop rule support to ingress ACL net/mlx5: E-switch, remove special uplink ingress ACL handling net/mlx5: E-Switch, reserve and use same uplink metadata across ports net/mlx5: Add ability to insert to specific flow group mlx5: remove unused static inlines ==================== Link: https://lore.kernel.org/r/20220223233930.319301-1-saeed@kernel.orgSigned-off-by:
-
