Re: Things I wish I'd known about Inotify

[Michael Kerrisk (man-pages)]

On 04/07/2014 11:31 AM, Jan Kara wrote:
> On Sun 06-04-14 11:00:29, Michael Kerrisk (man-pages) wrote:
>> On 04/04/2014 02:43 PM, Jan Kara wrote:
>>> On Fri 04-04-14 09:35:50, Michael Kerrisk (man-pages) wrote:
>>>> On 04/03/2014 10:52 PM, Jan Kara wrote:
>>>>> On Thu 03-04-14 08:34:44, Michael Kerrisk (man-pages) wrote:
>>
>> [...]
>>
>>>>>>    Dealing with rename() events
>>>>>>        The  IN_MOVED_FROM  and  IN_MOVED_TO events that are generated by
>>>>>>        rename(2) are usually available as consecutive events when  readâ
>>>>>>        ing from the inotify file descriptor.  However, this is not guarâ
>>>>>>        anteed.  If multiple processes are triggering  events  for  moniâ
>>>>>>        tored  objects,  then  (on rare occasions) an arbitrary number of
>>>>>>        other events may appear between the IN_MOVED_FROM and IN_MOVED_TO
>>>>>>        events.
>>>>>>
>>>>>>        Matching  up  the IN_MOVED_FROM and IN_MOVED_TO event pair generâ
>>>>>>        ated by rename(2) is thus inherently racy.  (Don't forget that if
>>>>>>        an  object is renamed outside of a monitored directory, there may
>>>>>>        not even be an IN_MOVED_TO event.)  Heuristic  approaches  (e.g.,
>>>>>>        assume the events are always consecutive) can be used to ensure a
>>>>>>        match in most cases, but will inevitably miss some cases, causing
>>>>>>        the  application  to  perceive  the IN_MOVED_FROM and IN_MOVED_TO
>>>>>>        events as being unrelated.  If watch  descriptors  are  destroyed
>>>>>>        and  re-created as a result, then those watch descriptors will be
>>>>>>        inconsistent with the watch descriptors in  any  pending  events.
>>>>>>        (Re-creating the inotify file descriptor and rebuilding the cache
>>>>>>        may be useful to deal with this scenario.)
>>>>>   Well, but there's 'cookie' value meant exactly for matching up
>>>>> IN_MOVED_FROM and IN_MOVED_TO events. And 'cookie' is guaranteed to be
>>>>> unique at least within the inotify instance (in fact currently it is unique
>>>>> within the whole system but I don't think we want to give that promise).
>>>>
>>>> Yes, that's already assumed by my discussion above (its described elsewhere
>>>> in the page). But your comment makes me think I should add a few words to
>>>> remind the reader of that fact. I'll do that.
>>>   Yes, that would be good.
>>>
>>>> But, the point is that even with the cookie, matching the events is 
>>>> nontrivial, since:
>>>>
>>>> * There may not even be an IN_MOVED_FROM event
>>>> * There may be an arbitrary number of other events in between the 
>>>>   IN_MOVED_FROM and the IN_MOVED_TO.
>>>>
>>>> Therefore, one has to use heuristic approaches such as "allow at least
>>>> N millisconds" or "check the next N events" to see if there is an
>>>> IN_MOVED_FROM that matches the IN_MOVED_TO. I can't see any way around
>>>> that being inherently racy. (It's unfortunate that the kernel can't 
>>>> provide a guarantee that the two events are always consecutive, since
>>>> that would simply user space's life considerably.)
>>
>>>   Yeah, it's unpleasant but doing that would be quite costly/complex at the
>>> kernel side. 
>>
>> Yep, I imagined that was probably the reason.
>   I had a look into that code again and it's all designed around the fact
> that there's a single inode to notify. If you liked to have atomic rename
> notifications, you'd have to rewrite that to work with two inodes, finding
> out whether these two inodes are actually watched by the same group or
> not... Doable but complex. Alternatively you could just lock down the whole
> notification subsystem while generating rename events. But that's rather
> costly. Just that we have the complications written down somewhere in case
> someone wants to look into this in future.
> 
>>> And the race would in the worst case lead to application
>>> thinking there's been file moved outside of watched area & a file moved
>>> somewhere else inside the watched area. So the application will have to
>>> possibly inspect that file. That doesn't seem too bad.
>>
>> It's actually very bad. See the text above. The point is that one likely
>> treatment on an IN_MOVED_FROM event that has no IN_MOVED_TO is to remove
>> the watches for the moved out subtree. If it turns out that this really
>> was just a rename(), then on the IN_MOVED_TO, the watches will be recreated
>> *with different watch descriptors*, thus invalidating the watch descriptors
>> in any queued but as yet unprocessed inotify events. See what I mean? 
>> That's quite painful for user space.

Sorry for the late follow-up....

>   But if I understand it right, you loose only the information for recreated
> watches. So you effectively loose all the information about what has
> happened inside the subtree of moved directory (or what has happened with
> the moved file). But since you think it's a file / dir moved from outside
> of watched area, you have to fully rescan that file / dir anyway.  

Ack on you summary there.

> Sure
> that's costly but if your heuristics for detecting rename works 99.9% of
> time it should be OK, shouldn't it? And you have to have that code handling
> caching file / dir written anyway for handling real moves from outside of
> watched hierarchy.

And ack on that.

> Don't get me wrong, I understand it would be easier for userspace to get
> atomic rename notifications, I'm just trying to understand what exactly is
> painful so that I can compare the cost at the kernel side with the cost at
> the userspace side...

Yes, I was probably a little too strong in my statement. My perspective
is that I'd tried to write an (experimental) application that would track 
*all* events for a file tree (modulo queue overflow), and then I 
encountered the wall of "rename() events are not consecutive", which 
basically rendered that task impossible because of the races involved. 

All that you say above also fits with my understanding. I was just
(perhaps overly) disappointed to find that I couldn't (perfectly) 
achieve the tracking task that I'd attempted. (And furthermore, of 
course, the code became a bit more complicated to handle the 
possibility that some queued events may be for watch descriptors 
that are no longer valid.)

Thanks for your response.

Cheers,

Michael




-- 
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/