Simulate Linux Kernel Faults

This document describes how to use KernelChaos to simulate Linux kernel faults. This feature injects I/O-based or memory-based faults into the specified kernel paths using BPF.

Although you can set the injection target of KernelChaos to one or several Pods, the performance of other Pods on the host will be affected, because all Pods share the same kernel.

danger

The simulation of Linux kernel faults is disabled by default. Do not use this feature in a production environment.

Prerequisites

• Linux kernel version >= 4.18.
• The Linux kernel configuration CONFIG_BPF_KPROBE_OVERRIDE is enabled.
• The bpfki.create configuration value in values.yaml is true.

Configuration file

A simple KernelChaos configuration file is as follows:

apiVersion: chaos-mesh.org/v1alpha1
kind: KernelChaos
metadata:
  name: kernel-chaos-example
  namespace: chaos-mesh
spec:
  mode: one
  selector:
    namespaces:
      - chaos-mount
  failKernRequest:
    callchain:
      - funcname: '__x64_sys_mount'
    failtype: 0

For more configuration examples, refer to examples. You can modify these configuration examples as needed.

Configuration description:

• mode specifies how the experiment runs. The options are as follows:

  • one: randomly selects an eligible Pod.
  • all: selects all eligible Pods.
  • fixed: selects a specified number of eligible Pods.
  • fixed-percent: selects a specified percentage of eligible Pods.
  • random-max-percent: selects the maximum percentage of eligible Pods.

• selector specifies the target Pod for fault injection.

• FailedkernRequest specifies the fault mode (such as kmallo and bio). It also specifies a specific call chain path and the optional filtering conditions. The configuration items are as follows:

  • Failtype specifies the fault type. The value options are as follows:

    • '0': injects the slab allocation error should_failslab.
    • '1': injects the memory page allocation error should_fail_alloc_page.
    • '2': injects the bio error should_fail_bio.

    For more information on these three fault types, refer to fault-injection and inject_example.

  • Callchain specifies a specific call chain. For example:

    ext4_mount
    -> mount_subtree
    -> ...
        -> should_failslab

    You can also use the function parameters as filtering rules to inject more fine-grained faults. Refer to call chain and predicate examples for more information. If no call chain is specified, keep the callchain field empty, indicating that faults will be injected to any path on which slab alloc is called (for example, kmallo).

    The call chain type is a frame array, consisting of the following three parts:

    • funcname, which can be found from the kernel source code or from /proc/kallsyms, such as ext4_mount.
    • parameters, which is used for filtering. If you want to inject a slab error on the d_alloc_parallel(struct dentry *parent, const struct qstr *name) with a special name bananas path, you need to set the parameters to struct dentry *parent, const struct qstr *name. Otherwise, omit this configuration.
    • predicate, which is used to access the parameters of the frame array. Taking parameters as an example, you can set it to STRNCMP(name->name, "bananas", 8) to control the path of fault injection, or you can leave it empty for all call paths that execute d_allo_parallel receive the slab fault injection.

  • headers specifies the kernel header file you need. For example, "linux/mmzone.h" and "linux/blkdev.h".

  • probability specifies the probability of faults. If you want the probability of 1%, set to '1'.

  • times specifies the maximum number of times a fault is triggered.

Create an experiment using kubectl

Use kubectl to create an experiment:

kubectl apply -f KernelChaos

The KernelChaos feature is similar to inject.py. For more information, refer to input_example.txt.

A simple example is as follows:

#include <sys/mount.h>
#include <stdio.h>
#include <string.h>
#include <errno.h>
#include <unistd.h>

int main(void) {
  int ret;
  while (1) {
    ret = mount("/dev/sdc", "/mnt", "ext4",
          MS_MGC_VAL | MS_RDONLY | MS_NOSUID, "");
    if (ret < 0)
      fprintf(stderr, "%s\n", strerror(errno));
    sleep(1);
    ret = umount("/mnt");
    if (ret < 0)
      fprintf(stderr, "%s\n", strerror(errno));
  }
}

During the fault injection, the output is as follows:

> Cannot allocate memory
> Invalid argument
> Cannot allocate memory
> Invalid argument
> Cannot allocate memory
> Invalid argument
> Cannot allocate memory
> Invalid argument
> Cannot allocate memory
> Invalid argument

Usage restriction

You can use container_id to limit the scope of the fault injection, but some paths trigger system-level behaviors. For example:

When failtype is 1, it means that the physical page allocation fails. If this event is frequently triggered within a short period of time (for example, while (1) {memset(malloc(1M), '1', 1M)}), the oom-killer system call is triggered to recycle memory.