max inotify user watches in the kernel should be increased

[Mads Chr. Olesen]

Binary package hint: beagle

As described by Joe Shaw in http://thomas.apestaart.org/log/?p=482#comment-2699 beagle will fall-back to continously indexing a user's files, if it fails to inotfy watch all directories for changes. The described solution is increasing /proc/sys/fs/inotify/max_user_watches to some higher value. Increasing this number should not have any adverse effects, unless the watches are actually used.

On my system I increased the number to 32768, as i have almost that many sub-directories in my homedir:
find . -type d | wc -l

I increased the value by adding "fs.inotify.max_user_watches = 32768" to /etc/sysctl.conf, not sure if it can be done in some way only if beagle has been installed?

[Mads Chr. Olesen (shiyee)]

This is possibly also the cause for bug #77106

[Joe Shaw (joeshaw)]

I added procps to this as an affected package, because I think the right place to fix this is there. This can happen with any program that will set up lots of inotify watches, not just Beagle, although it's the most likely culprit at this point.

I'd suggest setting it to at least 65536. A higher number of watches will take additional kernel memory, but only if those watches are used. So for the vast majority of users who fit under the 8192 default limit, increasing the value will have no adverse effects.

Bug #77106 is not caused by this, that is another bug that was fixed in Beagle 0.2.14, I believe.

[Andreas Zitzelsberger (az82)]

I think max_user_watches should be set according on the system's RAM in the desktop distribution and kept at the low default for the server distribution.
Does anybody actually know how much memory an inotify watch consumes?

[Joe Shaw (joeshaw)]

If an inotify watch isn't used, it uses no memory. So simply increasing the limit has no overhead on the system.

It's only when they're used that they actually consume kernel memory. I don't know exactly how much is used, but I believe there's no additional overhead other than the inotify structure in the kernel.

[Kees Cook (kees)]

This was addressed in procps 1:3.2.7-3ubuntu6

[jay armstrong (jayarmstrong)]

As an end user, i'm curious -- how was this addressed in procps 1:3.2.7-3ubuntu6? I'm still running into this hard limit in jaunty.