|
| 1 | +--- |
| 2 | +title: Troubleshooting inotify Watcher Limit Problems |
| 3 | +description: Learn how to resolve problems related to the inotify Watcher Limit. |
| 4 | +--- |
| 5 | + |
| 6 | +The [`inotify` facility] allows programs to monitor files and directories for |
| 7 | +changes, so that they will receive an event immediately whenever a user or |
| 8 | +program modifies the file or directory. Many such programs can fall back to |
| 9 | +periodically checking files and directories for changes, also known as polling. |
| 10 | +Using polling avoids the watch limit, but may result in higher system loading |
| 11 | +and a longer interval for the program to detect changes. |
| 12 | + |
| 13 | +`inotify` requires kernel resources (memory and processor) for each file it |
| 14 | +tracks. As a result, the Linux kernel limits the number of file watchers that |
| 15 | +each user can register. The default settings vary according to the host system |
| 16 | +distribution; on Ubuntu 20.04 LTS, the default limit is 8,192 watches per instance |
| 17 | + |
| 18 | +[`inotify` facility]: https://en.wikipedia.org/wiki/Inotify |
| 19 | + |
| 20 | +With some applications and tools, including Webpack or [code-server], you may |
| 21 | +encounter an error similar to the following: |
| 22 | + |
| 23 | +> Watchpack Error (watcher): Error: ENOSPC: System limit for number of file |
| 24 | +> watchers reached, watch '/some/path' |
| 25 | +
|
| 26 | +[code-server]: https://github.com/cdr/code-server |
| 27 | + |
| 28 | +On a 64-bit system, each `inotify` watch that programs register will consume |
| 29 | +approximately 1 kB of kernel memory, which cannot be swapped to disk and is not |
| 30 | +counted against the environment memory limit setting. |
| 31 | + |
| 32 | +## Diagnosis |
| 33 | + |
| 34 | +If the total number of watchers is close to the `max_user_watches` setting, then |
| 35 | +it is likely that you are encountering this limit. |
| 36 | + |
| 37 | +### Check tunable settings |
| 38 | + |
| 39 | +There are three kernel tuning options related to the `inotify` system: |
| 40 | + |
| 41 | +- `fs.inotify.max_queued_events`, which defines an upper bound on the number of |
| 42 | + file notification events pending delivery to programs. |
| 43 | +- `fs.inotify.max_user_instances`, which determines the maximum number of |
| 44 | + `inotify` instances per user. Programs using `inotify` will typically create |
| 45 | + a single _instance_, so this limit is unlikely to cause issues. |
| 46 | +- `fs.inotify.max_user_watches`, which defines the maximum number of files and |
| 47 | + folders that programs can monitor for changes. |
| 48 | + |
| 49 | +For additional detail regarding the `inotify` system, please see |
| 50 | +[inotify(7)](https://man7.org/linux/man-pages/man7/inotify.7.html). |
| 51 | + |
| 52 | +To see the values for these settings applicable to your environment, run the |
| 53 | +following commands: |
| 54 | + |
| 55 | +```console |
| 56 | +$ sysctl fs.inotify.{max_queued_events,max_user_instances,max_user_watches} |
| 57 | +fs.inotify.max_queued_events = 16384 |
| 58 | +fs.inotify.max_user_instances = 128 |
| 59 | +fs.inotify.max_user_watches = 8192 |
| 60 | +``` |
| 61 | + |
| 62 | +Because these settings are not namespace-aware, the values will be the same |
| 63 | +regardless of whether you run the commands on the host system or inside a |
| 64 | +container running on that host. |
| 65 | + |
| 66 | +### Identify inotify consumers |
| 67 | + |
| 68 | +To identify the programs consuming `inotify` watches, you can use a script that |
| 69 | +summarizes the information available in the `/proc` filesystem, such as |
| 70 | +[`inotify-consumers`]. This script will show the names of programs along with |
| 71 | +the number of `inotify` watches registered with the kernel: |
| 72 | + |
| 73 | +```console |
| 74 | +$ ./inotify-consumers |
| 75 | + INOTIFY |
| 76 | + WATCHER |
| 77 | + COUNT PID USER COMMAND |
| 78 | +-------------------------------------- |
| 79 | + 269 254560 coder /opt/coder/code-server/lib/node /opt/coder/code-server/lib/vscode/out/bootstrap-fork --type=watcherService |
| 80 | + 5 1722 coder /opt/coder/code-server/lib/node /opt/coder/code-server/lib/vscode/out/vs/server/fork |
| 81 | + 2 254538 coder /opt/coder/code-server/lib/node /opt/coder/code-server/lib/vscode/out/bootstrap-fork --type=extensionHost |
| 82 | + 2 1507 coder gpg-agent --homedir /home/coder/.gnupg --use-standard-socket --daemon |
| 83 | + |
| 84 | + 278 WATCHERS TOTAL COUNT |
| 85 | +``` |
| 86 | + |
| 87 | +Please note that this is a third-party script published by an individual who is |
| 88 | +not affiliated with Coder, and as such, we cannot provide a warranty or support |
| 89 | +for its usage. |
| 90 | + |
| 91 | +[`inotify-consumers`]: https://github.com/fatso83/dotfiles/blob/master/utils/scripts/inotify-consumers |
| 92 | + |
| 93 | +To see the specific files and directories that the tools track for changes, you |
| 94 | +can use `strace` to monitor invocations of the `inotify_add_watch` system call: |
| 95 | + |
| 96 | +```console |
| 97 | +$ strace --follow-forks --trace='inotify_add_watch' inotifywait --quiet test |
| 98 | +inotify_add_watch(3, "test", IN_ACCESS|IN_MODIFY|IN_ATTRIB|IN_CLOSE_WRITE|IN_CLOSE_NOWRITE|IN_OPEN|IN_MOVED_FROM|IN_MOVED_TO|IN_CREATE|IN_DELETE|IN_DELETE_SELF|IN_MOVE_SELF) = 1 |
| 99 | +``` |
| 100 | + |
| 101 | +This shows that the `inotifywait` command is listening for notifications related |
| 102 | +to the `test` file. |
| 103 | + |
| 104 | +## Resolution |
| 105 | + |
| 106 | +If you encounter the file watcher limit, there are two potential resolutions: |
| 107 | +either reduce the number of file watcher registrations, or increase the maximum |
| 108 | +file watcher limit. We recommend attempting to reduce the file watcher |
| 109 | +registrations first, because increasing the number of file watches may result |
| 110 | +in high processor utilization. |
| 111 | + |
| 112 | +### Reduce watchers |
| 113 | + |
| 114 | +Many applications include files that change very rarely, and developers often |
| 115 | +do not consider these to be part of _their_ application. For example, this may |
| 116 | +include third-party dependencies stored in `node_modules`. Tools may watch for |
| 117 | +changes to these files and folders, consuming `inotify` watchers. |
| 118 | + |
| 119 | +These tools typically provide configuration settings to exclude certain files, |
| 120 | +paths, and patterns from file watching. For example, Visual Studio Code and |
| 121 | +`code-server` apply the following [user workspace setting] by default: |
| 122 | + |
| 123 | +```json |
| 124 | +"files.watcherExclude": { |
| 125 | + "**/.git/objects/**": true, |
| 126 | + "**/.git/subtree-cache/**": true, |
| 127 | + "**/node_modules/**": true, |
| 128 | + "**/.hg/store/**": true |
| 129 | +}, |
| 130 | +``` |
| 131 | + |
| 132 | +Consider adding other infrequently-changed files to this list, which will cause |
| 133 | +Visual Studio Code to poll for changes to those files periodically. For other |
| 134 | +software tools, please see the respective user manual. |
| 135 | + |
| 136 | +[user workspace setting]: https://code.visualstudio.com/docs/getstarted/settings |
| 137 | + |
| 138 | +### Increase the watch limit |
| 139 | + |
| 140 | +You can increase the kernel tunable option to increase the maximum number of |
| 141 | +`inotify` watches for each user. This is a global setting that applies to all |
| 142 | +users sharing the same system, or _Node_ in Kubernetes parlance. To do this, |
| 143 | +modify the `sysctl` configuration file, or apply a DaemonSet to the Kubernetes |
| 144 | +cluster to apply that change to all nodes automatically. |
| 145 | + |
| 146 | +For example, create a file called `/etc/sysctl.d/watches.conf` and include the |
| 147 | +following contents: |
| 148 | + |
| 149 | +```text |
| 150 | +fs.inotify.max_user_watches = 65536 |
| 151 | +``` |
| 152 | + |
| 153 | +Alternatively, you can use the following DaemonSet with `kubectl apply`: |
| 154 | + |
| 155 | +```yaml |
| 156 | +apiVersion: apps/v1 |
| 157 | +kind: DaemonSet |
| 158 | +metadata: |
| 159 | + name: more-fs-watchers |
| 160 | + namespace: kube-system |
| 161 | + labels: |
| 162 | + app: more-fs-watchers |
| 163 | + k8s-app: more-fs-watchers |
| 164 | +spec: |
| 165 | + selector: |
| 166 | + matchLabels: |
| 167 | + k8s-app: more-fs-watchers |
| 168 | + template: |
| 169 | + metadata: |
| 170 | + labels: |
| 171 | + name: more-fs-watchers |
| 172 | + k8s-app: more-fs-watchers |
| 173 | + annotations: |
| 174 | + seccomp.security.alpha.kubernetes.io/defaultProfileName: runtime/default |
| 175 | + apparmor.security.beta.kubernetes.io/defaultProfileName: runtime/default |
| 176 | + spec: |
| 177 | + nodeSelector: |
| 178 | + kubernetes.io/os: linux |
| 179 | + initContainers: |
| 180 | + - name: sysctl |
| 181 | + image: alpine:3 |
| 182 | + command: |
| 183 | + - sysctl |
| 184 | + - -w |
| 185 | + - fs.inotify.max_user_watches=16384 |
| 186 | + resources: |
| 187 | + requests: |
| 188 | + cpu: 10m |
| 189 | + memory: 1Mi |
| 190 | + limits: |
| 191 | + cpu: 100m |
| 192 | + memory: 5Mi |
| 193 | + securityContext: |
| 194 | + # We need to run as root in a privileged container to modify |
| 195 | + # /proc/sys on the host (for sysctl) |
| 196 | + runAsUser: 0 |
| 197 | + privileged: true |
| 198 | + readOnlyRootFilesystem: true |
| 199 | + capabilities: |
| 200 | + drop: |
| 201 | + - ALL |
| 202 | + containers: |
| 203 | + - name: pause |
| 204 | + image: k8s.gcr.io/pause:3.5 |
| 205 | + command: |
| 206 | + - /pause |
| 207 | + resources: |
| 208 | + requests: |
| 209 | + cpu: 10m |
| 210 | + memory: 1Mi |
| 211 | + limits: |
| 212 | + cpu: 100m |
| 213 | + memory: 5Mi |
| 214 | + securityContext: |
| 215 | + runAsNonRoot: true |
| 216 | + runAsUser: 65535 |
| 217 | + allowPrivilegeEscalation: false |
| 218 | + privileged: false |
| 219 | + readOnlyRootFilesystem: true |
| 220 | + capabilities: |
| 221 | + drop: |
| 222 | + - ALL |
| 223 | + terminationGracePeriodSeconds: 5 |
| 224 | +``` |
| 225 | +
|
| 226 | +This DaemonSet will ensure that the corresponding Pod runs on _every_ Linux node |
| 227 | +in the cluster. When new nodes join the cluster, such as when an autoscaling |
| 228 | +event occurs, the DaemonsSet will ensure that the pod runs on that node as well. |
| 229 | +You may delete the DaemonSet by running the following command: |
| 230 | +
|
| 231 | +```shell-session |
| 232 | +$ kubectl delete --namespace=kube-system daemonset more-fs-watchers |
| 233 | +daemonset.apps "more-fs-watchers" deleted |
| 234 | +``` |
| 235 | + |
| 236 | +However, note that the setting will persist until the node restarts or until |
| 237 | +another program sets the `fs.inotify.max_user_watches` setting. |
| 238 | + |
| 239 | +## See Also |
| 240 | + |
| 241 | +- Resources for Visual Studio Code and code-server: |
| 242 | + - [User and Workspace Settings](https://code.visualstudio.com/docs/getstarted/settings), |
| 243 | + in particular, the setting called `files.watcherExclude`. |
| 244 | + - [VS Code Setting: files.watcherExclude](https://youtu.be/WMNua0ob6Aw) |
| 245 | + (YouTube). |
| 246 | + - [My ultimate VSCode |
| 247 | + configuration](https://dev.to/vaidhyanathan93/ulitmate-vscode-configuration-4i2o), |
| 248 | + a blog post describing a user's preferred settings, including file exclusions. |
| 249 | +- [Filesystem notification, part 1: An overview of dnotify and inotify](https://lwn.net/Articles/604686/) |
| 250 | + and [Filesystem notification, part 2: A deeper investigation of inotify](https://lwn.net/Articles/605128/) |
| 251 | + examine the `inotify` mechanism in detail, along with its predecessor, |
| 252 | + `dnotify`. |
| 253 | +- Microsoft's Language Server Protocol (LSP) specification describes an approach |
| 254 | + for using file watch notifications: |
| 255 | + [DidChangeWatchedFiles Notification](https://microsoft.github.io/language-server-protocol/specification#workspace_didChangeWatchedFiles). |
| 256 | + Visual Studio Code and code-server, along with many other editors, uses this |
| 257 | + protocol for programming language support, and the same constraints and |
| 258 | + limitations apply to those tools. |
| 259 | +- Kubernetes provides a mechanism for [Setting Sysctls for a Pod](https://kubernetes.io/docs/tasks/administer-cluster/sysctl-cluster/#setting-sysctls-for-a-pod). |
| 260 | + However, the `inotify` tunable options cannot be set this way, so this is not |
| 261 | + applicable to resolving this issue. |
| 262 | +- [INotify watch limit](https://blog.passcod.name/2017/jun/25/inotify-watch-limit.html) |
| 263 | + provides additional context on the problem and its resolution. |
| 264 | +- [inotify(7)](https://man7.org/linux/man-pages/man7/inotify.7.html), the Linux |
| 265 | + manual page related to the `inotify` system call. |
| 266 | +- [Kernel Korner - Intro to inotify](https://www.linuxjournal.com/article/8478) |
0 commit comments