wingologA mostly dorky weblog by Andy Wingo2018-05-21T14:29:32Ztekutihttps://wingolog.org/feed/atomAndy Wingohttps://wingolog.org/correct or inotify: pick onehttps://wingolog.org/2018/05/21/correct-or-inotify-pick-one2018-05-21T14:29:32Z2018-05-21T14:29:32Z

Let's say you decide that you'd like to see what some other processes on your system are doing to a subtree of the file system. You don't want to have to change how those processes work -- you just want to see what files those processes create and delete.

One approach would be to just scan the file-system tree periodically, enumerating its contents. But when the file system tree is large and the change rate is low, that's not an optimal thing to do.

Fortunately, Linux provides an API to allow a process to receive notifications on file-system change events, called inotify. So you open up the inotify(7) manual page, and are greeted with this:

With careful programming, an application can use inotify to efficiently monitor and cache the state of a set of filesystem objects. However, robust applications should allow for the fact that bugs in the monitoring logic or races of the kind described below may leave the cache inconsistent with the filesystem state. It is probably wise to do some consistency checking, and rebuild the cache when inconsistencies are detected.

It's not exactly reassuring is it? I mean, "you had one job" and all.

Reading down a bit farther, I thought that with some "careful programming", I could get by. After a day of trying, I am now certain that it is impossible to build a correct recursive directory monitor with inotify, and I am not even sure that "good enough" solutions exist.

pitfall the first: buffer overflow

Fundamentally, inotify races the monitoring process with all other processes on the system. Events are delivered to the monitoring process via a fixed-size buffer that can overflow, and the monitoring process provides no back-pressure on the system's rate of filesystem modifications. With inotify, you have to be ready to lose events.

This I think is probably the easiest limitation to work around. The kernel can let you know when the buffer overflows, and you can tweak the buffer size. Still, it's a first indication that perfect is not possible.

pitfall the second: now you see it, now you don't

This one is the real kicker. Say you get an event that says that a file "frenemies.txt" has been created in the directory "/contacts/". You go to open the file -- but is it still there? By the time you get around to looking for it, it could have been deleted, or renamed, or maybe even created again or replaced! This is a TOCTTOU race, built-in to the inotify API. It is literally impossible to use inotify without this class of error.

The canonical solution to this kind of issue in the kernel is to use file descriptors instead. Instead of or possibly in addition to getting a name with the file change event, you get a descriptor to a (possibly-unlinked) open file, which you would then be responsible for closing. But that's not what inotify does. Oh well!

pitfall the third: race conditions between inotify instances

When you inotify a directory, you get change notifications for just that directory. If you want to get change notifications for subdirectories, you need to open more inotify instances and poll on them all. However now you have N2 problems: as poll and the like return an unordered set of readable file descriptors, each with their own ordering, you no longer have access to a linear order in which changes occurred.

It is impossible to build a recursive directory watcher that definitively says "ok, first /contacts/frenemies.txt was created, then /contacts was renamed to /peeps, ..." because you have no ordering between the different watches. You don't know that there was ever even a time that /contacts/frenemies.txt was an accessible file name; it could have been only ever openable as /peeps/frenemies.txt.

Of course, this is the most basic ordering problem. If you are building a monitoring tool that actually wants to open files -- good luck bubster! It literally cannot be correct. (It might work well enough, of course.)

reflections

As far as I am aware, inotify came out to address the needs of desktop search tools like the belated Beagle (11/10 good pupper just trying to get his pup on). Especially in the days of spinning metal, grovelling over the whole hard-drive was a real non-starter, especially if the search database should to be up-to-date.

But after looking into inotify, I start to see why someone at Google said that desktop search was in some ways harder than web search -- I mean we all struggle to find files on our own machines, even now, 15 years after the whole dnotify/inotify thing started. Part of it is that the given the choice between supporting reliable, fool-proof file system indexes on the one hand, and overclocking the IOPS benchmarks on the other, the kernel gave us inotify. I understand it, but inotify still sucks.

I dunno about you all but whenever I've had to document such an egregious uncorrectable failure mode as any of the ones in the inotify manual, I have rewritten the software instead. In that spirit, I hope that some day we shall send inotify to the pet cemetery, to rest in peace beside Beagle.