tools: add filtering by mount namespace
In previous patches, I added the option --cgroupmap to filter events
belonging to a set of cgroup-v2. Although this approach works fine with
systemd services and containers when cgroup-v2 is enabled, it does not
work with containers when only cgroup-v1 is enabled because
bpf_get_current_cgroup_id() only works with cgroup-v2. It also requires
Linux 4.18 to get this bpf helper function.
This patch adds an additional way to filter by containers, using mount
namespaces.
Note that this does not help with systemd services since they normally
don't create a new mount namespace (unless you set some options like
'ReadOnlyPaths=', see "man 5 systemd.exec").
My goal with this patch is to filter Kubernetes pods, even on
distributions with an older kernel (<4.18) or without cgroup-v2 enabled.
- This is only implemented for tools that already support filtering by
cgroup id (bindsnoop, capable, execsnoop, profile, tcpaccept, tcpconnect,
tcptop and tcptracer).
- I picked the mount namespace because the other namespaces could be
disabled in Kubernetes (e.g. HostNetwork, HostPID, HostIPC).
It can be tested by following the example in docs/special_filtering added
in this commit, to avoid compiling locally the following command can be used
```
sudo bpftool map create /sys/fs/bpf/mnt_ns_set type hash key 8 value 4 \
entries 128 name mnt_ns_set flags 0
docker run -ti --rm --privileged \
-v /usr/src:/usr/src -v /lib/modules:/lib/modules \
-v /sys/fs/bpf:/sys/fs/bpf --pid=host kinvolk/bcc:alban-containers-filters \
/usr/share/bcc/tools/execsnoop --mntnsmap /sys/fs/bpf/mnt_ns_set
```
Co-authored-by: Alban Crequy <alban@kinvolk.io>
Co-authored-by: Mauricio Vásquez <mauricio@kinvolk.io>
diff --git a/tools/profile.py b/tools/profile.py
index 2067933..dd6f65f 100755
--- a/tools/profile.py
+++ b/tools/profile.py
@@ -24,10 +24,11 @@
#
# 15-Jul-2016 Brendan Gregg Created this.
# 20-Oct-2016 " " Switched to use the new 4.9 support.
-# 26-Jan-2019 " " Changed to exclude CPU idle by default.
+# 26-Jan-2019 " " Changed to exclude CPU idle by default.
from __future__ import print_function
from bcc import BPF, PerfType, PerfSWConfig
+from bcc.containers import filter_by_containers
from sys import stderr
from time import sleep
import argparse
@@ -72,7 +73,8 @@
./profile -L 185 # only profile thread with TID 185
./profile -U # only show user space stacks (no kernel)
./profile -K # only show kernel space stacks (no user)
- ./profile --cgroupmap ./mappath # only trace cgroups in this BPF map
+ ./profile --cgroupmap mappath # only trace cgroups in this BPF map
+ ./profile --mntnsmap mappath # only trace mount namespaces in the map
"""
parser = argparse.ArgumentParser(
description="Profile CPU stack traces at a timed interval",
@@ -115,6 +117,8 @@
help=argparse.SUPPRESS)
parser.add_argument("--cgroupmap",
help="trace cgroups in this BPF map only")
+parser.add_argument("--mntnsmap",
+ help="trace mount namespaces in this BPF map only")
# option logic
args = parser.parse_args()
@@ -146,10 +150,6 @@
BPF_HASH(counts, struct key_t);
BPF_STACK_TRACE(stack_traces, STACK_STORAGE_SIZE);
-#if CGROUPSET
-BPF_TABLE_PINNED("hash", u64, u64, cgroupset, 1024, "CGROUPPATH");
-#endif
-
// This code gets a bit complex. Probably not suitable for casual hacking.
int do_perf_event(struct bpf_perf_event_data *ctx) {
@@ -163,12 +163,9 @@
if (!(THREAD_FILTER))
return 0;
-#if CGROUPSET
- u64 cgroupid = bpf_get_current_cgroup_id();
- if (cgroupset.lookup(&cgroupid) == NULL) {
+ if (container_should_be_filtered()) {
return 0;
}
-#endif
// create map key
struct key_t key = {.pid = tgid};
@@ -246,11 +243,7 @@
stack_context = "user + kernel"
bpf_text = bpf_text.replace('USER_STACK_GET', user_stack_get)
bpf_text = bpf_text.replace('KERNEL_STACK_GET', kernel_stack_get)
-if args.cgroupmap:
- bpf_text = bpf_text.replace('CGROUPSET', '1')
- bpf_text = bpf_text.replace('CGROUPPATH', args.cgroupmap)
-else:
- bpf_text = bpf_text.replace('CGROUPSET', '0')
+bpf_text = filter_by_containers(args) + bpf_text
sample_freq = 0
sample_period = 0