[CRIU] [PATCH] perf: add ability to trace system calls with perf

[Andrey Vagin]

First of all perf can trace parasite code. strace can't because ptrace
doesn't allow to attach to one task twice

perf doesn't affect performance so significant as strace and
it allows to trace parasite code. But perf isn't so verbose. For example
it can't show arguments which are tranfered by pointers.

All processes are executed in a new perf_evets cgroup and
"perf record" collects events only for processes from this group.

On dump perf is executed from a network-lock action script, it's a first
point where processes are frozed.
And perf is stopped from post-dump and post-restore, it's the last point
where processes are frozed.

Example of output:
    18.651 ( 0.001 ms): criu/9243 close(fd: 44                                                          ) = 0
    18.655 ( 0.001 ms): criu/9243 gettimeofday(tv: 0x7fff32fdd7f0                                       ) = 0
    18.670 ( 0.013 ms): criu/9243 write(fd: 1023, buf: 0x6b6a00, count: 53                              ) = 53
    18.710 ( 1.078 ms): env00/9226 recvfrom(fd: 3, ubuf: 0x7f963bb95998, size: 12, flags: WAITALL        ) = 12
    18.726 ( 0.014 ms): env00/9226 write(fd: 4, buf: 0x7f963bb95878, count: 26                           ) = 26
    18.761 ( 0.034 ms): env00/9226 mprotect(start: 0x7f963b771000, len: 2093056, prot: READ              ) = 0

Signed-off-by: Andrey Vagin <avagin at openvz.org>
---
 test/perf-script.sh | 16 ++++++++++++++++
 test/zdtm.sh        | 26 +++++++++++++++++++++++++-
 2 files changed, 41 insertions(+), 1 deletion(-)
 create mode 100755 test/perf-script.sh

diff --git a/test/perf-script.sh b/test/perf-script.sh
new file mode 100755
index 0000000..ad56187
--- /dev/null
+++ b/test/perf-script.sh
@@ -0,0 +1,16 @@
+#!/bin/sh
+
+set -e
+[ "$CRTOOLS_SCRIPT_ACTION" == network-lock ] && {
+	echo 0 > /sys/fs/cgroup/perf_event/tasks
+	perf record -o perf.dump -e raw_syscalls:sys_exit,raw_syscalls:sys_enter --cgroup zdtm,zdtm -a &
+	echo $! > perf.pid
+	sleep 1
+	exit 0
+}
+[ "$CRTOOLS_SCRIPT_ACTION" == post-dump -o "$CRTOOLS_SCRIPT_ACTION" == post-restore ] && {
+	kill `cat perf.pid`
+	exit 0
+}
+
+exit 0
diff --git a/test/zdtm.sh b/test/zdtm.sh
index 9a2c188..1d0f6c3 100755
--- a/test/zdtm.sh
+++ b/test/zdtm.sh
@@ -255,6 +255,7 @@ COMPILE_ONLY=0
 START_ONLY=0
 BATCH_TEST=0
 SPECIFIED_NAME_USED=0
+PERF=""
 
 zdtm_sep()
 { (
@@ -547,6 +548,8 @@ EOF
 		gen_args="$gen_args --force-irmap"
 	fi
 
+	[ -n "$PERF" ] && gen_args="$gen_args --action-script $SCRIPTDIR/perf-script.sh"
+
 	for i in `seq $ITERATIONS`; do
 		local cpt_args=
 		local dump_only=
@@ -585,6 +588,12 @@ EOF
 		setsid $CRIU_CPT $dump_cmd -D $ddump -o dump.log -v4 -t $PID $gen_args $cpt_args
 		retcode=$?
 
+		[ -n "$PERF" ] && {
+			echo Use the following command to analize perf data:
+			echo perf trace report -i $ddump/perf.dump
+			[ "$retcode" -ne 0 ] && kill `cat $ddump/perf.pid`
+		}
+
 		#
 		# Here we may have two cases: either checkpoint is failed
 		# with some error code, or checkpoint is complete but return
@@ -645,7 +654,16 @@ EOF
 			rm -f $TPID || true
 
 			echo Restore
-			setsid $CRIU restore -D $ddump -o restore.log -v4 -d $gen_args || return 2
+			[ -n "$PERF" ] &&  (
+				echo 0 > /sys/fs/cgroup/perf_event/tasks
+				perf record -o $ddump/perf.restore -e raw_syscalls:sys_exit,raw_syscalls:sys_enter --cgroup zdtm,zdtm -a &
+				echo $! > $ddump/perf.pid
+				sleep 1
+			)
+			setsid $CRIU restore -D $ddump -o restore.log -v4 -d $gen_args || {
+				[ -n "$PERF" ] && kill `cat $ddump/perf.pid`
+				return 2
+			}
 
 			[ -n "$PIDNS" ] && PID=`cat $TPID`
 			for i in `seq 5`; do
@@ -911,6 +929,12 @@ while :; do
 		}
 		shift
 		;;
+	  --perf)
+		PERF=1
+		test -d /sys/fs/cgroup/perf_event/zdtm || mkdir /sys/fs/cgroup/perf_event/zdtm || exit 1
+		echo $$ > /sys/fs/cgroup/perf_event/zdtm/tasks
+		shift
+		;;
 	  -*)
 		echo "Unrecognized option $1, aborting!" 1>&2
 		usage
-- 
1.9.3