This repository demonstrates a Linux kernel /proc bug.
In particular, when many mounts are added and removed in parallel, reading
/proc/self/mounts (or mountinfo, or mountstats) is unreliable.
The bug existed in the Linux kernel for about 10+ years, and is recently fixed in the kernel 5.8 by commit 9f6c61f96f2d97. For older kernels, the only known workaround is to re-read the whole file.
To run this on your Linux system:
git clone https://github.com/kolyshkin/procfs-test
cd procfs-test
make test
Note you need to either run make test as root, or have sudo working
to run a binary that does mount/umount (sudo might ask you for a password).
In case of no bugs, the output should look like
bash ./test
i.e. no errors of any kind.
When the bug is present, you'll see something like
bash ./test
./mnt-umnt: can't find /home/user/git/procfs-test/testdata/8
./mnt-umnt: can't find /home/user/git/procfs-test/testdata/75
....
This test runs a number of short-lived processes in parallel. Each of these processes does the following (see mnt-umnt.c for details):
- mounts a tmpfs to a directory provided
- finds that mount in
/proc/self/mounts - unmounts that tmpfs
This was found while trying to optimize aufs graph driver in Docker engine, by making its locking more fine-grained. Previously, there was a single global lock between Get and Put (i.e. mount and unmount), so only one mount or unmount operation was possible at a time. With finer-grained locking, multiple mounts and unmounts can happen in parallel, and that revealed this very bug.