docs: inotify: add initial documentation

We're introducing a workaround for enabling users to utilize inotify on
mounts that are backed by virtiofs. Let's add some documentation on how
this work.

Signed-off-by: Eric Ernst <eric_ernst@apple.com>

docs/design/README.md

@@ -8,4 +8,5 @@ Kata Containers design documents:
 - [VSocks](VSocks.md)
 - [VCPU handling](vcpu-handling.md)
 - [Host cgroups](host-cgroups.md)
+- [`Inotify` support](inotify.md)
 - [Metrics(Kata 2.0)](kata-2-0-metrics.md)

docs/design/inotify.md

@@ -0,0 +1,30 @@
+# Kata Containers support for `inotify`
+
+## Background on `inotify` usage
+
+A common pattern in Kubernetes is to watch for changes to files/directories passed in as `ConfigMaps`
+or `Secrets`. Sidecar's normally use `inotify` to watch for changes and then signal the primary container to reload
+the updated configuration. Kata Containers typically will pass these host files into the guest using `virtiofs`, which
+does not support `inotify` today. While we work to enable this use case in `virtiofs`, we introduced a workaround in Kata Containers.
+This document describes how Kata Containers implements this workaround.
+
+### Detecting a `watchable` mount
+
+Kubernetes creates `secrets` and `ConfigMap` mounts at very specific locations on the host filesystem. For container mounts,
+the `Kata Containers` runtime will check the source of the mount to identify these special cases. For these use cases, only a single file
+or very few would typically need to be watched. To avoid excessive overheads in making a mount watchable,
+we enforce a limit of eight files per mount. If a `secret` or `ConfigMap` mount contains more than 8 files, it will not be
+considered watchable. We similarly enforce a limit of 1 MB per mount to be considered watchable. Non-watchable mounts will
+continue to propagate changes from the mount on the host to the container workload, but these updates will not trigger an
+`inotify` event.
+
+If at any point a mount grows beyond the eight file or 1MB limit, it will no longer be `watchable.`
+
+### Presenting a `watchable` mount to the workload
+
+For mounts that are considered `watchable`, inside the guest, the `kata-agent` will poll the mount presented from
+the host through `virtiofs` and copy any changed files to a `tmpfs` mount that is presented to the container. In this way,
+for `watchable` mounts, Kata will do the polling on behalf of the workload and existing workloads needn't change their usage
+of `inotify`.
+
+![drawing](arch-images/inotify-workaround.png)