inotify watch 耗尽

每个 linux 进程可以持有多个 fd，每个 inotify 类型的 fd 可以 watch 多个目录，每个用户下所有进程 inotify 类型的 fd 可以 watch 的总目录数有个最大限制，这个限制可以通过内核参数配置: fs.inotify.max_user_watches

查看最大 inotify watch 数:

  1. $ cat /proc/sys/fs/inotify/max_user_watches
  2. 8192

使用下面的脚本查看当前有 inotify watch 类型 fd 的进程以及每个 fd watch 的目录数量，降序输出，带总数统计:

  1. #!/usr/bin/env bash
  2. #
  3. # Copyright 2019 (c) roc
  4. #
  5. # This script shows processes holding the inotify fd, alone with HOW MANY directories each inotify fd watches(0 will be ignored).
  6. total=0
  7. result="EXE PID FD-INFO INOTIFY-WATCHES\n"
  8. while read pid fd; do \
  9.   exe="$(readlink -f /proc/$pid/exe || echo n/a)"; \
  10.   fdinfo="/proc/$pid/fdinfo/$fd" ; \
  11.   count="$(grep -c inotify "$fdinfo" || true)"; \
  12.   if [ $((count)) != 0 ]; then
  13.     total=$((total+count)); \
  14.     result+="$exe $pid $fdinfo $count\n"; \
  15.   fi
  16. done <<< "$(lsof +c 0 -n -P -u root|awk '/inotify$/ { gsub(/[urw]$/,"",$4); print $2" "$4 }')" && echo "total $total inotify watches" && result="$(echo -e $result|column -t)\n" && echo -e "$result" | head -1 && echo -e "$result" | sed "1d" | sort -k 4rn;

示例输出:

  1. total 7882 inotify watches
  2. EXE                                         PID    FD-INFO                INOTIFY-WATCHES
  3. /usr/local/qcloud/YunJing/YDEyes/YDService  25813  /proc/25813/fdinfo/8   7077
  4. /usr/bin/kubelet                            1173   /proc/1173/fdinfo/22   665
  5. /usr/bin/ruby2.3                            13381  /proc/13381/fdinfo/14  54
  6. /usr/lib/policykit-1/polkitd                1458   /proc/1458/fdinfo/9    14
  7. /lib/systemd/systemd-udevd                  450    /proc/450/fdinfo/9     13
  8. /usr/sbin/nscd                              7935   /proc/7935/fdinfo/3    6
  9. /usr/bin/kubelet                            1173   /proc/1173/fdinfo/28   5
  10. /lib/systemd/systemd                        1      /proc/1/fdinfo/17      4
  11. /lib/systemd/systemd                        1      /proc/1/fdinfo/18      4
  12. /lib/systemd/systemd                        1      /proc/1/fdinfo/26      4
  13. /lib/systemd/systemd                        1      /proc/1/fdinfo/28      4
  14. /usr/lib/policykit-1/polkitd                1458   /proc/1458/fdinfo/8    4
  15. /usr/local/bin/sidecar-injector             4751   /proc/4751/fdinfo/3    3
  16. /usr/lib/accountsservice/accounts-daemon    1178   /proc/1178/fdinfo/7    2
  17. /usr/local/bin/galley                       8228   /proc/8228/fdinfo/10   2
  18. /usr/local/bin/galley                       8228   /proc/8228/fdinfo/9    2
  19. /lib/systemd/systemd                        1      /proc/1/fdinfo/11      1
  20. /sbin/agetty                                1437   /proc/1437/fdinfo/4    1
  21. /sbin/agetty                                1440   /proc/1440/fdinfo/4    1
  22. /usr/bin/kubelet                            1173   /proc/1173/fdinfo/10   1
  23. /usr/local/bin/envoy                        4859   /proc/4859/fdinfo/5    1
  24. /usr/local/bin/envoy                        5427   /proc/5427/fdinfo/5    1
  25. /usr/local/bin/envoy                        6058   /proc/6058/fdinfo/3    1
  26. /usr/local/bin/envoy                        6893   /proc/6893/fdinfo/3    1
  27. /usr/local/bin/envoy                        6950   /proc/6950/fdinfo/3    1
  28. /usr/local/bin/galley                       8228   /proc/8228/fdinfo/3    1
  29. /usr/local/bin/pilot-agent                  3819   /proc/3819/fdinfo/5    1
  30. /usr/local/bin/pilot-agent                  4244   /proc/4244/fdinfo/5    1
  31. /usr/local/bin/pilot-agent                  5901   /proc/5901/fdinfo/3    1
  32. /usr/local/bin/pilot-agent                  6789   /proc/6789/fdinfo/3    1
  33. /usr/local/bin/pilot-agent                  6808   /proc/6808/fdinfo/3    1
  34. /usr/local/bin/pilot-discovery              6231   /proc/6231/fdinfo/3    1
  35. /usr/local/bin/sidecar-injector             4751   /proc/4751/fdinfo/5    1
  36. /usr/sbin/acpid                             1166   /proc/1166/fdinfo/6    1
  37. /usr/sbin/dnsmasq                           7572   /proc/7572/fdinfo/8    1

如果看到总 watch 数比较大，接近最大限制，可以修改内核参数调高下这个限制。

临时调整:

  1. sudo sysctl fs.inotify.max_user_watches=524288

永久生效:

  1. echo "fs.inotify.max_user_watches=524288" >> /etc/sysctl.conf && sysctl -p

打开 inotify_add_watch 跟踪，进一步 debug inotify watch 耗尽的原因:

  1. echo 1 >> /sys/kernel/debug/tracing/events/syscalls/sys_exit_inotify_add_watch/enable