inotify_rm_watch() user-space safety requirements?

inotify_rm_watch() user-space safety requirements?

[Jeff Smith]

inotify's behavior concerning events from removed watches (they do
happen) and watch descriptor reuse (beyond my knowledge) is currently
undocumented.

Although it mimics a standard multiplexing interface in most regards,
writing a robust user-space handler is comparatively more complex due
to the atypical delivery of "stale" wd events preceding an IN_IGNORE
event and a lack of guarantees about how quickly a wd can be reused
via inotify_add_watch(). Not being familiar with inotify/fsnotify
internals, it's not trivially obvious to me how the fsnotify_group
management is being done. Up to the present, I've maintained queues of
"dead" wd wrappers (or at least a counter) to filter stale events, but
I am clueless whether or not this is overkill.

If removed descriptors are reserved until the IN_IGNORE event is
drained from the read queue, could that be formally guaranteed? If
it's not, is it functionality that could ever reasonably be expected
to be added, short of some other form of new (optional?)
queue-filter-on-rm functionality? It's my experience that the
asynchronous handling of watch removals is a cost that seldom serves
much user benefit.

Regards,
Jeff

Re: inotify_rm_watch() user-space safety requirements?

[Heinrich Schuchardt]

On 27.05.2014 19:25, Jeff Smith wrote:
> inotify's behavior concerning events from removed watches (they do
> happen) and watch descriptor reuse (beyond my knowledge) is currently
> undocumented.
>
> Although it mimics a standard multiplexing interface in most regards,
> writing a robust user-space handler is comparatively more complex due
> to the atypical delivery of "stale" wd events preceding an IN_IGNORE
> event and a lack of guarantees about how quickly a wd can be reused
> via inotify_add_watch(). Not being familiar with inotify/fsnotify
> internals, it's not trivially obvious to me how the fsnotify_group
> management is being done. Up to the present, I've maintained queues of
> "dead" wd wrappers (or at least a counter) to filter stale events, but
> I am clueless whether or not this is overkill.
>
> If removed descriptors are reserved until the IN_IGNORE event is
> drained from the read queue, could that be formally guaranteed? If
> it's not, is it functionality that could ever reasonably be expected
> to be added, short of some other form of new (optional?)
> queue-filter-on-rm functionality? It's my experience that the
> asynchronous handling of watch removals is a cost that seldom serves
> much user benefit.
>
 > Regards,
 > Jeff

Hello Jeff,

I tried to dive a bit into the code. This is what I understand:

Function inotify_ignored_and_remove_idr is called after the mark has 
been removed. This function puts an IN_IGNORED event onto the inotify 
queue and removes the watch descriptor from the list of used watch 
descriptors using function idr_remove.

With a test program I could receive the IN_IGNORED event. This behavior 
is currently not documented in the manpages (inotify.7 and 
inotify_rm_watch.2).

When inotify_add_watch is called it uses function idr_alloc_cyclic to 
assign a watch descriptor ID. This function starts looking for an unused 
id starting with the id after the last assigned watch descriptor.

This implies that in most cases inotify_add_watch will return a watch 
descriptor different to the one released by a prior call to 
inotify_rm_watch. But there is no guarantee.

I consider this a bug.

I CCed the maintainers of the inotify interface hoping that they can 
provide a better solution.

Until such a solution is provided I suggest you use the following 
workaround. After calling inotify_rm_watch read from the inotify file 
descriptor until you reach the matching IN_IGNORED event.

Only thereafter you can safely call inotify_add_watch again.

Best regards

Heinrich Schuchardt

Re: inotify_rm_watch() user-space safety requirements?

[Jeff Smith]

Hi Heinrich,

Thanks for looking into this. I've never encountered the aliasing via
wraparound or otherwise, but I've tried to code against unexpected
descriptor reuse anyway through more flexible workarounds. From a user
standpoint, denser descriptor reassignment would be preferable to
facilitate simple dense linear mappings of cache objects, but it's
easy to see why things were designed as they were given the queue
entry persistence. Getting current behavior accurately documented is
my first hope, and fixing any gotcha cases my second, but being able
to have inotify_add_watch/read look and feel more like
open/epoll_ctl(EPOLL_CTL_ADD))/epoll_wait would be my ultimate wish.
The asynchronous-feeling handle removal is not a great interface in
general to deal with.

Best regards,
Jeff

On Tue, May 27, 2014 at 2:32 PM, Heinrich Schuchardt <xypron.glpk@gmx.de> wrote:
> On 27.05.2014 19:25, Jeff Smith wrote:
>>
>> inotify's behavior concerning events from removed watches (they do
>> happen) and watch descriptor reuse (beyond my knowledge) is currently
>> undocumented.
>>
>> Although it mimics a standard multiplexing interface in most regards,
>> writing a robust user-space handler is comparatively more complex due
>> to the atypical delivery of "stale" wd events preceding an IN_IGNORE
>> event and a lack of guarantees about how quickly a wd can be reused
>> via inotify_add_watch(). Not being familiar with inotify/fsnotify
>> internals, it's not trivially obvious to me how the fsnotify_group
>> management is being done. Up to the present, I've maintained queues of
>> "dead" wd wrappers (or at least a counter) to filter stale events, but
>> I am clueless whether or not this is overkill.
>>
>> If removed descriptors are reserved until the IN_IGNORE event is
>> drained from the read queue, could that be formally guaranteed? If
>> it's not, is it functionality that could ever reasonably be expected
>> to be added, short of some other form of new (optional?)
>> queue-filter-on-rm functionality? It's my experience that the
>> asynchronous handling of watch removals is a cost that seldom serves
>> much user benefit.
>>
>> Regards,
>> Jeff
>
> Hello Jeff,
>
> I tried to dive a bit into the code. This is what I understand:
>
> Function inotify_ignored_and_remove_idr is called after the mark has been
> removed. This function puts an IN_IGNORED event onto the inotify queue and
> removes the watch descriptor from the list of used watch descriptors using
> function idr_remove.
>
> With a test program I could receive the IN_IGNORED event. This behavior is
> currently not documented in the manpages (inotify.7 and inotify_rm_watch.2).
>
> When inotify_add_watch is called it uses function idr_alloc_cyclic to assign
> a watch descriptor ID. This function starts looking for an unused id
> starting with the id after the last assigned watch descriptor.
>
> This implies that in most cases inotify_add_watch will return a watch
> descriptor different to the one released by a prior call to
> inotify_rm_watch. But there is no guarantee.
>
> I consider this a bug.
>
> I CCed the maintainers of the inotify interface hoping that they can provide
> a better solution.
>
> Until such a solution is provided I suggest you use the following
> workaround. After calling inotify_rm_watch read from the inotify file
> descriptor until you reach the matching IN_IGNORED event.
>
> Only thereafter you can safely call inotify_add_watch again.
>
> Best regards
>
> Heinrich Schuchardt
>

Re: inotify_rm_watch() user-space safety requirements?

[Michael Kerrisk (man-pages)]

Hi Heinrich,

On 05/27/2014 09:32 PM, Heinrich Schuchardt wrote:
> On 27.05.2014 19:25, Jeff Smith wrote:
>> inotify's behavior concerning events from removed watches (they do
>> happen) and watch descriptor reuse (beyond my knowledge) is currently
>> undocumented.
>>
>> Although it mimics a standard multiplexing interface in most regards,
>> writing a robust user-space handler is comparatively more complex due
>> to the atypical delivery of "stale" wd events preceding an IN_IGNORE
>> event and a lack of guarantees about how quickly a wd can be reused
>> via inotify_add_watch(). Not being familiar with inotify/fsnotify
>> internals, it's not trivially obvious to me how the fsnotify_group
>> management is being done. Up to the present, I've maintained queues of
>> "dead" wd wrappers (or at least a counter) to filter stale events, but
>> I am clueless whether or not this is overkill.
>>
>> If removed descriptors are reserved until the IN_IGNORE event is
>> drained from the read queue, could that be formally guaranteed? If
>> it's not, is it functionality that could ever reasonably be expected
>> to be added, short of some other form of new (optional?)
>> queue-filter-on-rm functionality? It's my experience that the
>> asynchronous handling of watch removals is a cost that seldom serves
>> much user benefit.
>>
>  > Regards,
>  > Jeff
> 
> Hello Jeff,
> 
> I tried to dive a bit into the code. This is what I understand:
> 
> Function inotify_ignored_and_remove_idr is called after the mark has 
> been removed. This function puts an IN_IGNORED event onto the inotify 
> queue and removes the watch descriptor from the list of used watch 
> descriptors using function idr_remove.
> 
> With a test program I could receive the IN_IGNORED event. This behavior 
> is currently not documented in the manpages (inotify.7 and 
> inotify_rm_watch.2).

I'm a little unclear here... What behavior is not documented?

> When inotify_add_watch is called it uses function idr_alloc_cyclic to 
> assign a watch descriptor ID. This function starts looking for an unused 
> id starting with the id after the last assigned watch descriptor.
> 
> This implies that in most cases inotify_add_watch will return a watch 
> descriptor different to the one released by a prior call to 
> inotify_rm_watch. But there is no guarantee.
> 
> I consider this a bug.
> 
> I CCed the maintainers of the inotify interface hoping that they can 
> provide a better solution.
> Until such a solution is provided I suggest you use the following 
> workaround. After calling inotify_rm_watch read from the inotify file 
> descriptor until you reach the matching IN_IGNORED event.
> 
> Only thereafter you can safely call inotify_add_watch again.

As I wrote elsewhere (http://marc.info/?l=linux-man&m=140145559605117&w=2), 
it's not clear to me if this is really anything more than a theoretical 
problem. Quoting my earlier text:

Background is bugs https://bugzilla.kernel.org/show_bug.cgi?id=76851
and https://bugzilla.kernel.org/show_bug.cgi?id=77111 . The point is:

1. When an inotify watch descriptor is removed, pending unread 
   events remain pending.
2. When allocating a new watch descriptor, a past WD may 
   be recycled.
3. In theory, it could happen that events left over at 1 could
   be interpreted as though they belonged to the filesystem
   object watch in step 2.

But, in order to trigger this bug you need to

0. Remove your watch descriptor (wd1),
1. Leave some unread events for wd1 on the queue. and in the meantime,
2. Cycle through INT_MAX watch descriptors until you reuse wd1.

Unless I've missed something, the chances of that happening are 
pretty remote. (On the other hand, user-space applications can sometimes
be creatively zany ;_).)

Cheers,

Michael

-- 
Michael Kerrisk
Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
Linux/UNIX System Programming Training: http://man7.org/training/