tetragon: Add docs for tetragon performance stats

Adding details on 'tetra debug progs' and overhead metrics.

Signed-off-by: Jiri Olsa <[email protected]>

docs/content/en/docs/concepts/performance-stats.md

@@ -0,0 +1,172 @@
+---
+title: "BPF programs statistics"
+weight: 5
+description: "Monitor bpf programs statistics"
+---
+
+This page shows you how to monitor bpf programs statistics.
+
+## Concept
+
+The BPF subsystem provides performance data for each loaded program and tetragon
+exports that in metrics or display that in terminal in top like tool.
+
+## In terminal
+
+The tetra command allows to display loaded BPF programs in terminal with:
+
+```
+# tetra debug progs
+```
+
+The default output shows tetragon programs only and looks like:
+
+```
+2024-10-31 11:12:45.94715546 +0000 UTC m=+8.038098448
+
+Ovh(%)  Id      Cnt     Time    Name                            Pin
+  0.00  22201   0       0       event_execve                    /sys/fs/bpf/tetragon/__base__/event_execve/prog
+  0.00  22198   0       0       event_exit_acct_process         /sys/fs/bpf/tetragon/__base__/event_exit/prog
+  0.00  22200   0       0       event_wake_up_new_task          /sys/fs/bpf/tetragon/__base__/kprobe_pid_clear/prog
+  0.00  22207   0       0       tg_cgroup_rmdir                 /sys/fs/bpf/tetragon/__base__/tg_cgroup_rmdir/prog
+  0.00  22206   0       0       tg_kp_bprm_committing_creds     /sys/fs/bpf/tetragon/__base__/tg_kp_bprm_committing_creds/prog
+  0.00  22221   0       0       generic_kprobe_event            /sys/fs/bpf/tetragon/syswritefollowfdpsswd/generic_kprobe/__x64_sys_close/prog
+  0.00  22225   0       0       generic_kprobe_event            /sys/fs/bpf/tetragon/syswritefollowfdpsswd/generic_kprobe/__x64_sys_write/prog
+  0.00  22211   0       0       generic_kprobe_event            /sys/fs/bpf/tetragon/syswritefollowfdpsswd/generic_kprobe/fd_install/prog
+```
+
+The fields have following meaning:
+
+- `Ovh` is system wide overhead of the BPF program
+- `Id` is global BPF ID of the program (as shown by `bpftool prog`)
+- `Cnt` is count with number of BPF program executions
+- `Time` is sum of the time of all BPF program executions
+- `Pin` is BPF program pin path in bpfffs
+
+It's possible to display all BPF programs with `--all`:
+
+```
+# tetra debug progs --all
+```
+
+That has following output:
+
+```
+2024-10-31 11:19:37.720137195 +0000 UTC m=+7.165535117
+
+Ovh(%)  Id      Cnt     Time    Name            Pin
+  0.00  159     2       82620   event_execve    -
+  0.00  171     68      18564   iter            -
+  0.00  158     2       10170   event_wake_up_n -
+  0.00  164     2       4254    tg_kp_bprm_comm -
+  0.00  157     2       3868    event_exit_acct -
+  0.00  97      2       1680                    -
+  0.00  35      2       1442                    -
+  0.00  83      0       0       sd_devices      -
+  0.00  9       0       0                       -
+  0.00  7       0       0                       -
+  0.00  8       0       0                       -
+  0.00  87      0       0       sd_devices      -
+...
+```
+
+Above commands should run properly on top of the tetragon sources.
+At the moment to run it properly under K8S you need to specify extra
+directory flags:
+
+```
+# kubectl exec -ti -n kube-system tetragon-66rk4 -c tetragon -- tetra debug progs --bpf-dir /run/cilium/bpffs/tetragon/ --all --bpf-lib /var/lib/tetragon/
+```
+
+Note that there are other options to customize the behaviour:
+
+```
+# tetra debug progs --help
+Retrieve information about BPF programs on the host.
+
+Examples:
+- tetragon BPF programs top style
+  # tetra debug progs
+- all BPF programs top style
+  # tetra debug progs --all
+- one shot mode (displays one interval data)
+  # tetra debug progs --once
+- change interval to 10 seconds
+  # tetra debug progs  --timeout 10
+- change interval to 10 seconds in one shot mode
+  # tetra debug progs --once --timeout 10
+
+Usage:
+  tetra debug progs [flags]
+
+Aliases:
+  progs, top
+
+Flags:
+      --all              Get all programs
+      --bpf-dir string   Location of bpffs tetragon directory (default "/sys/fs/bpf/tetragon")
+      --bpf-lib string   Location of Tetragon libs (btf and bpf files) (default "bpf/objs/")
+  -h, --help             help for progs
+      --no-clear         Do not clear screen between rounds
+      --once             Run in one shot mode
+      --timeout int      Interval in seconds (delay in one shot mode) (default 1)
+```
+
+## Metrics
+
+The BPF subsystem provides performance data for each loaded program
+and tetragon exports that in metrics.
+
+For each loaded BPF program we get:
+- `run count` which counts how many times the BPF program was executed
+- `run time` which sums the time BPF program spent in all its executions
+
+
+Hence for each loaded BPF program we export 2 related metrics:
+
+- `tetragon_overhead_time_program_total[namespace,policy,sensor,attach]`
+- `tetragon_overhead_cnt_program_total[namespace,policy,sensor,attach]`
+
+
+Each loaded program is identified by labels:
+
+- `namespace` is policy K8S namespace
+- `policy` is policy name
+- `sensor` is sensor name
+- `attach` is program attachment name
+
+
+If we have `generic_kprobe` sensor attached on `__x64_sys_close` kernel function
+under `syswritefollowfdpsswd` policy, the related metrics will look like:
+
+```
+tetragon_overhead_program_runs_total{attach="__x64_sys_close",policy="syswritefollowfdpsswd",policy_namespace="",sensor="generic_kprobe"} 15894
+tetragon_overhead_program_seconds_total{attach="__x64_sys_close",policy="syswritefollowfdpsswd",policy_namespace="",sensor="generic_kprobe"} 1.03908217e+08
+```
+
+
+##  Limitations
+
+Note that the BPF programs statistics are not enabled by default, because they introduce extra overhead,
+so it's necessary to enable them manually.
+
+- Either with `sysctl`:
+
+  ```
+  # sysctl kernel.bpf_stats_enabled=1
+  ```
+
+  and make sure you disable the stats when it's no longer needed:
+
+  ```
+  # sysctl kernel.bpf_stats_enabled=0
+  ```
+
+- Or with following `tetra` command:
+
+  ```
+  # tetra debug enable-stats
+  ^C
+  ```
+
+  where the stats are enabled as long as the command is running (sleeping really).