Kirill Kolyshkin:
> It happens in Docker's aufs graph driver (which does not use mount.aufs
> but rather calls mount() syscall directly, and does the same for umount
> except it calls auplink flush beforehand). Previously, we had a global lock
:::
So your scenario is
- taskA:
+ assume the given dir is /somewhere/mntpntA
+ issue mount(2) for the dir
+ run auplink for the dir
+ and then issue umount(2) for the dir
- run taskA repeatedly
- run the same task repeatedly in parallel, with giving the dir
dedicated for each task.
- it is 100% sure that mount(2) succeeded before auplink (for the same
dir) starts.
But /proc/mounts sometimes doesn't show the mntpnt correctly, and
auplilnk cannot get it.
Hmm, it sounds like a problem in /proc/mounts kernel-space. But I
cannot believe easily that such bug exists. Won't you post your kernel
info (version, patches, etc)? And can we reproduce the problem easily
using a shell script like this (which uses tmpfs instead of aufs)?
----------------------------------------
#!/bin/sh
stopme=/tmp/stopme
mntpnt="$(seq -f '/tmp/mntpnt%g' 256)"
mkdir -p $mntpnt
rm -f $stopme
sudo true
f()
{
while test ! -e $stopme
do
if sudo mount -t tmpfs none $1
then
fgrep -q $1 /proc/mounts ||
{
> $stopme
return 1
}
sudo umount $1
fi
done
}
for i in $mntpnt
do
f $i &
done
wait
----------------------------------------
> By the way, this /proc/mounts issue was not the only one; I also discovered
> that two parallel mount syscalls sometimes clash on something related to
> xino
> handling (I guess they concurrently try to access the file without
> locking... but
> I have yet to look into it).
That is really interesting.
If you can, please post the clash message (kernel log).
Just for your information, for Docker users especially for the very
short lived aufs mount, I'd recommend you to try "ephemeral" option.
This aufs mount option is handled by /sbin/mount.aufs, and converted
into some other options including "noplink,noxino."
These options will be helpful for you, as long as you don't want the
plink and xino feature.
For the details of the option, please refer to aufs manual.
J. R. Okajima
