Merge pull request #38 from xdp-project/vestas06_tc_qdisc

TC policy example of overriding netstack TXQ

Author: netoptimizer

tc-policy/Makefile

@@ -0,0 +1,13 @@
+# SPDX-License-Identifier: (GPL-2.0 OR BSD-2-Clause)
+
+USER_TARGETS   := tc_txq_policy
+BPF_TARGETS    := tc_txq_policy_kern
+
+# Depend on bpftool for auto generating
+# skeleton header file tc_txq_policy_kern.skel.h
+#
+BPF_SKEL_OBJ := tc_txq_policy_kern.o
+
+LIB_DIR = ../lib
+
+include $(LIB_DIR)/common.mk

tc-policy/README.org

@@ -0,0 +1,118 @@
+#+Title: Controlling TC qdisc TXQ selection via BPF
+
+* Use-case
+
+As a policy we don't want any traffic generated by the Linux networking stack,
+to use transmit queue *zero*.
+
+This use-case is connected with =AF_XDP=.  The example
+[[file:../AF_XDP-interaction/]] is sending important Real-Time traffic on XDP-socket
+queue zero.  Some HW and NIC drivers (e.g. igb and igc) don't have enough
+hardware TX-queues to allocate seperate queues for XDP.  Thus, these queues are
+shared between XDP and network stack, and there is a potential lock-contention
+and also HW queue usage contention.
+
+* Example
+
+The BPF code in this example is rather simple:
+ - See: [[file:tc_txq_policy_kern.c]]
+
+This BPF program is meant to be loaded in the TC *egress* hook.
+
+** TC-BPF loader
+
+The =tc= cmdline tool is notorious difficult to use, and have issues (mounting
+BPF file-system) on Yocto build.
+
+Thus, [[file:tc_txq_policy.c]] contains a C-code loader, that attach the BPF-prog to
+the TC-hook, without depending on =tc= command util. Furthermore, the loader
+uses =bpftool= skeleton feature (to generate a header file) allowing to create a
+binary that contains the BPF-object itself, making it self-contained.
+
+* Gotchas: XPS
+
+For TXQ (=queue_mapping=) overwrite to work, you need to *disable* XPS (Transmit
+Packet Steering), as XSP will have higher precedence than our BPF change to
+=queue_mapping=. This is done by writing 0 into each =/sys/class/net/= tx-queue
+file =/sys/class/net/DEV/queues/tx-*/xps_cpus=.
+
+A script for configuring and disabling XPS is provided here: [[file:xps_setup_ash.sh]].
+
+Script command line to disable XPS:
+#+begin_src sh
+ sudo ./xps_setup_ash.sh --dev DEVICE --default --disable
+#+end_src
+
+* Different ways to view queue_mapping
+
+Notice that =queue_mapping= set in BPF-prog is like RX-recorded number
+(=skb_rx_queue_recorded=).  When reaching TX-layer it will have been decremented
+by one (by =skb_get_rx_queue()=) at the TX netstack processing stage (in
+=__dev_queue_xmit()=).
+
+** perf probe
+
+The perf tool can be used for recording and inspecting the =skb->queue_mapping=.
+
+Remember: BPF-prog =queue_mapping= setting have been decremented by one at this
+TX netstack processing stage.
+
+#+begin_src sh
+perf probe -a 'dev_hard_start_xmit skb->dev->name:string skb->queue_mapping skb->hash'
+Added new event:
+  probe:dev_hard_start_xmit (on dev_hard_start_xmit with name=skb->dev->name:string queue_mapping=skb->queue_mapping hash=skb->hash)
+
+You can now use it in all perf tools, such as:
+	perf record -e probe:dev_hard_start_xmit -aR sleep 1
+#+end_src
+
+Afterwards run =perf script= and see results.
+
+** bpftrace
+
+It is also possible to monitor TXQ usage via a =bpftrace= script.
+ * see [[file:monitor_txq_usage.bt]].
+
+The main part of the script is:
+#+begin_src sh
+ tracepoint:net:net_dev_start_xmit {
+	$qm = args->queue_mapping;
+	$dev = str(args->name, 15);
+
+	@stat_txq_usage[$dev] = lhist($qm, 0,32,1);
+ }
+#+end_src
+
+Or as oneliner:
+#+begin_src sh
+ bpftrace -e 't:net:net_dev_start_xmit {@txq[str(args->name, 15)]=lhist(args->queue_mapping, 0,32,1)}'
+#+end_src
+
+* Inspecting loaded BPF
+
+How do you see if these BPF TC-hook programs are loaded?
+
+** bpftool
+
+The cmdline =bpftool net= can list any network related BPF program:
+
+#+begin_example
+ root@main-ctrl2:~ # bpftool net
+ xdp:
+ eth1(5) driver id 59
+
+ tc:
+ eth1(5) clsact/egress not_txq_zero:[17] id 17
+#+end_example
+
+There we see both the *XDP* BPF-program used by AF_XDP to redirect frames, and
+the *TC* hook BPF-prog loaded and attached.
+
+** tc egress
+
+The tc command need to be longer and more explicit:
+#+begin_example
+ root@main-ctrl2:~ # tc filter show dev eth1 egress
+ filter protocol all pref 49199 bpf chain 0
+ filter protocol all pref 49199 bpf chain 0 handle 0x1 not_txq_zero:[17] direct-action not_in_hw id 17 tag a761e11074b78959 jited
+#+end_example

tc-policy/adv_monitor_txq_usage.bt

@@ -0,0 +1,42 @@
+#!/usr/bin/bpftrace
+
+//BEGIN {
+//	printf("Monitor TXQ usage\n");
+//	printf(" - Remember: BPF set queue_mapping is one-less here (zero-indexed)\n");
+//}
+
+tracepoint:net:net_dev_start_xmit {
+	$qm = args->queue_mapping;
+	$dev = str(args->name, 16);
+
+	@stat_txq_usage[$dev] = lhist($qm, 0,32,1);
+}
+
+/*
+ * More precisely we actually want to see what netdev_pick_tx() is
+ * selecting, as sockets can possibly return another queue_id.
+ */
+
+kprobe:netdev_pick_tx {
+	$dev = ((struct net_device *)arg0)->name;
+	@record[cpu] = $dev;
+}
+
+kretprobe:netdev_pick_tx {
+	$dev = @record[cpu];
+	@netdev_pick_tx[$dev] = lhist(retval, 0,32,1);
+}
+
+/* Periodically print stats */
+interval:s:3
+{
+	printf("\nPeriodic show stats - time: ");
+        time();
+	print(@stat_txq_usage);
+	print(@netdev_pick_tx);
+}
+
+/* Default bpftrace will print all remaining maps at END */
+//END {
+//	printf("END stats:\n");
+//}

tc-policy/monitor_txq_usage.bt

@@ -0,0 +1,26 @@
+#!/usr/bin/bpftrace
+
+//BEGIN {
+//	printf("Monitor TXQ usage\n");
+//	printf(" - Remember: BPF set queue_mapping is one-less here (zero-indexed)\n");
+//}
+
+tracepoint:net:net_dev_start_xmit {
+	$qm = args->queue_mapping;
+	$dev = str(args->name, 16);
+
+	@stat_txq_usage[$dev] = lhist($qm, 0,32,1);
+}
+
+/* Periodically print stats */
+interval:s:3
+{
+	printf("\nPeriodic show stats - time: ");
+        time();
+	print(@stat_txq_usage);
+}
+
+/* Default bpftrace will print all remaining maps at END */
+//END {
+//	printf("END stats:\n");
+//}