Skip to contentSkip to conversation

Ignore mask handling in fanotify_group_event_mask()

DISCUSSION Jan Kara 2 participants 6 messages May 21, 2020 – May 25, 2020 VFS

lore
Share & tools
lore

Jan Kara

author reviewer
#0 lore raw

Hello Amir!

I was looking into backporting of commit 55bf882c7f13dd "fanotify: fix merging marks masks with FAN_ONDIR" and realized one oddity in fanotify_group_event_mask(). The thing is: Even if the mark mask is such that current event shouldn't trigger on the mark, we still have to take mark's ignore mask into account.

The most realistic example that would demonstrate the issue that comes to my mind is:

mount mark watching for FAN_OPEN | FAN_ONDIR.
inode mark on a directory with mask == 0 and ignore_mask == FAN_OPEN.

I'd expect the group will not get any event for opening the dir but the code in fanotify_group_event_mask() would not prevent event generation. Now as I've tested the event currently actually does not get generated because there is a rough test in send_to_group() that actually finds out that there shouldn't be anything to report and so fanotify handler is actually never called in such case. But I don't think it's good to have an inconsistent test in fanotify_group_event_mask(). What do you think?

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
#1 lore raw
Quoted 15 lines · Jan Kara

Hello Amir!

I was looking into backporting of commit 55bf882c7f13dd "fanotify: fix merging marks masks with FAN_ONDIR" and realized one oddity in fanotify_group_event_mask(). The thing is: Even if the mark mask is such that current event shouldn't trigger on the mark, we still have to take mark's ignore mask into account.

The most realistic example that would demonstrate the issue that comes to my mind is:

mount mark watching for FAN_OPEN | FAN_ONDIR.
inode mark on a directory with mask == 0 and ignore_mask == FAN_OPEN.

Quoted 8 lines

I'd expect the group will not get any event for opening the dir but the code in fanotify_group_event_mask() would not prevent event generation. Now as I've tested the event currently actually does not get generated because there is a rough test in send_to_group() that actually finds out that there shouldn't be anything to report and so fanotify handler is actually never called in such case. But I don't think it's good to have an inconsistent test in fanotify_group_event_mask(). What do you think?

I agree this is not perfect.
I think that moving the marks_ignored_mask line
To the top of the foreach loop should fix the broken logic. It will not make the code any less complicated to follow though. Perhaps with a comment along the lines of:

             /* Ignore mask is applied regardless of ISDIR and ON_CHILD flags */
             marks_ignored_mask |= mark->ignored_mask;

Now is there a real bug here?
Probably not because send_to_group() always applied an ignore mask that is greater or equal to that of fanotify_group_event_mask().

should fanotify_group_event_mask() re-apply the same generic logic already applied in send_to_group()? Maybe there is no point. After all, fanotify_group_event_mask() also does not handle FSNOTIFY_MARK_FLAG_IGNORED_SURV_MODIFY, so the assumption that send_to_group() is doing some of the logic is already there.

Not sure if I helped answering your question.

Thanks,
Amir.

#2 lore raw
Quoted 32 lines · Amir Goldstein
Quoted 15 lines · Jan Kara

Hello Amir!

I was looking into backporting of commit 55bf882c7f13dd "fanotify: fix merging marks masks with FAN_ONDIR" and realized one oddity in fanotify_group_event_mask(). The thing is: Even if the mark mask is such that current event shouldn't trigger on the mark, we still have to take mark's ignore mask into account.

The most realistic example that would demonstrate the issue that comes to my mind is:

mount mark watching for FAN_OPEN | FAN_ONDIR.
inode mark on a directory with mask == 0 and ignore_mask == FAN_OPEN.

Quoted 8 lines

I'd expect the group will not get any event for opening the dir but the code in fanotify_group_event_mask() would not prevent event generation. Now as I've tested the event currently actually does not get generated because there is a rough test in send_to_group() that actually finds out that there shouldn't be anything to report and so fanotify handler is actually never called in such case. But I don't think it's good to have an inconsistent test in fanotify_group_event_mask(). What do you think?

I agree this is not perfect.
I think that moving the marks_ignored_mask line
To the top of the foreach loop should fix the broken logic. It will not make the code any less complicated to follow though. Perhaps with a comment along the lines of:

Quoted 7 lines
             /* Ignore mask is applied regardless of ISDIR and ON_CHILD flags */
             marks_ignored_mask |= mark->ignored_mask;

Now is there a real bug here?
Probably not because send_to_group() always applied an ignore mask that is greater or equal to that of fanotify_group_event_mask().

That's a wrong statement of course.
We do need to re-apply the ignore mask when narrowing the event mask.

Exposing the bug requires a "compound" event.

The only case of compound event I could think of is this:

mount mark with mask == 0 and ignore_mask == FAN_OPEN. inode mark on a directory with mask == FAN_EXEC | FAN_EVENT_ON_CHILD.

The event: FAN_OPEN | FAN_EXEC | FAN_EVENT_ON_CHILD would be reported to group with the FAN_OPEN flag despite the fact that FAN_OPEN is in ignore mask of mount mark because the mark iteration loop skips over non-inode marks for events on child.

I'll try to work that case into the relevant LTP test to prove it and post a fix.

Thanks,
Amir.

author reviewer
#3 lore raw
Quoted 51 lines · Amir Goldstein
Quoted 32 lines · Amir Goldstein
Quoted 15 lines · Jan Kara

Hello Amir!

I was looking into backporting of commit 55bf882c7f13dd "fanotify: fix merging marks masks with FAN_ONDIR" and realized one oddity in fanotify_group_event_mask(). The thing is: Even if the mark mask is such that current event shouldn't trigger on the mark, we still have to take mark's ignore mask into account.

The most realistic example that would demonstrate the issue that comes to my mind is:

mount mark watching for FAN_OPEN | FAN_ONDIR.
inode mark on a directory with mask == 0 and ignore_mask == FAN_OPEN.

Quoted 8 lines

I'd expect the group will not get any event for opening the dir but the code in fanotify_group_event_mask() would not prevent event generation. Now as I've tested the event currently actually does not get generated because there is a rough test in send_to_group() that actually finds out that there shouldn't be anything to report and so fanotify handler is actually never called in such case. But I don't think it's good to have an inconsistent test in fanotify_group_event_mask(). What do you think?

I agree this is not perfect.
I think that moving the marks_ignored_mask line
To the top of the foreach loop should fix the broken logic. It will not make the code any less complicated to follow though. Perhaps with a comment along the lines of:

Quoted 7 lines
             /* Ignore mask is applied regardless of ISDIR and ON_CHILD flags */
             marks_ignored_mask |= mark->ignored_mask;

Now is there a real bug here?
Probably not because send_to_group() always applied an ignore mask that is greater or equal to that of fanotify_group_event_mask().

That's a wrong statement of course.
We do need to re-apply the ignore mask when narrowing the event mask.

Exposing the bug requires a "compound" event.

The only case of compound event I could think of is this:

mount mark with mask == 0 and ignore_mask == FAN_OPEN. inode mark on a directory with mask == FAN_EXEC | FAN_EVENT_ON_CHILD.

Quoted 8 lines

The event: FAN_OPEN | FAN_EXEC | FAN_EVENT_ON_CHILD would be reported to group with the FAN_OPEN flag despite the fact that FAN_OPEN is in ignore mask of mount mark because the mark iteration loop skips over non-inode marks for events on child.

I'll try to work that case into the relevant LTP test to prove it and post a fix.

Ha, that's clever. But FAN_EXEC does not exist in current fanotify. We only have FAN_OPEN_EXEC... And I don't think we have any compound events.

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
#4 lore raw
Quoted 59 lines · Jan Kara
Quoted 57 lines · Amir Goldstein
Quoted 32 lines · Amir Goldstein
Quoted 15 lines · Jan Kara

Hello Amir!

I was looking into backporting of commit 55bf882c7f13dd "fanotify: fix merging marks masks with FAN_ONDIR" and realized one oddity in fanotify_group_event_mask(). The thing is: Even if the mark mask is such that current event shouldn't trigger on the mark, we still have to take mark's ignore mask into account.

The most realistic example that would demonstrate the issue that comes to my mind is:

mount mark watching for FAN_OPEN | FAN_ONDIR.
inode mark on a directory with mask == 0 and ignore_mask == FAN_OPEN.

Quoted 8 lines

I'd expect the group will not get any event for opening the dir but the code in fanotify_group_event_mask() would not prevent event generation. Now as I've tested the event currently actually does not get generated because there is a rough test in send_to_group() that actually finds out that there shouldn't be anything to report and so fanotify handler is actually never called in such case. But I don't think it's good to have an inconsistent test in fanotify_group_event_mask(). What do you think?

I agree this is not perfect.
I think that moving the marks_ignored_mask line
To the top of the foreach loop should fix the broken logic. It will not make the code any less complicated to follow though. Perhaps with a comment along the lines of:

Quoted 7 lines
             /* Ignore mask is applied regardless of ISDIR and ON_CHILD flags */
             marks_ignored_mask |= mark->ignored_mask;

Now is there a real bug here?
Probably not because send_to_group() always applied an ignore mask that is greater or equal to that of fanotify_group_event_mask().

That's a wrong statement of course.
We do need to re-apply the ignore mask when narrowing the event mask.

Exposing the bug requires a "compound" event.

The only case of compound event I could think of is this:

mount mark with mask == 0 and ignore_mask == FAN_OPEN. inode mark on a directory with mask == FAN_EXEC | FAN_EVENT_ON_CHILD.

The event: FAN_OPEN | FAN_EXEC | FAN_EVENT_ON_CHILD would be reported to group with the FAN_OPEN flag despite the fact that FAN_OPEN is in ignore mask of mount mark because the mark iteration loop skips over non-inode marks for events on child.

Quoted 6 lines
Quoted 2 lines

I'll try to work that case into the relevant LTP test to prove it and post a fix.

Ha, that's clever. But FAN_EXEC does not exist in current fanotify. We only have FAN_OPEN_EXEC... And I don't think we have any compound events.

Typo. I meant FAN_OPEN_EXEC and you can see from LTP test we do have at least this one compound event.

We could also split it if we wanted to, but no reason to do it now.

Thanks,
Amir.

author reviewer
#5 lore raw
Quoted 63 lines · Amir Goldstein
Quoted 62 lines · Jan Kara
Quoted 59 lines · Amir Goldstein
Quoted 32 lines · Amir Goldstein
Quoted 15 lines · Jan Kara

Hello Amir!

I was looking into backporting of commit 55bf882c7f13dd "fanotify: fix merging marks masks with FAN_ONDIR" and realized one oddity in fanotify_group_event_mask(). The thing is: Even if the mark mask is such that current event shouldn't trigger on the mark, we still have to take mark's ignore mask into account.

The most realistic example that would demonstrate the issue that comes to my mind is:

mount mark watching for FAN_OPEN | FAN_ONDIR.
inode mark on a directory with mask == 0 and ignore_mask == FAN_OPEN.

Quoted 8 lines

I'd expect the group will not get any event for opening the dir but the code in fanotify_group_event_mask() would not prevent event generation. Now as I've tested the event currently actually does not get generated because there is a rough test in send_to_group() that actually finds out that there shouldn't be anything to report and so fanotify handler is actually never called in such case. But I don't think it's good to have an inconsistent test in fanotify_group_event_mask(). What do you think?

I agree this is not perfect.
I think that moving the marks_ignored_mask line
To the top of the foreach loop should fix the broken logic. It will not make the code any less complicated to follow though. Perhaps with a comment along the lines of:

Quoted 7 lines
             /* Ignore mask is applied regardless of ISDIR and ON_CHILD flags */
             marks_ignored_mask |= mark->ignored_mask;

Now is there a real bug here?
Probably not because send_to_group() always applied an ignore mask that is greater or equal to that of fanotify_group_event_mask().

That's a wrong statement of course.
We do need to re-apply the ignore mask when narrowing the event mask.

Exposing the bug requires a "compound" event.

The only case of compound event I could think of is this:

mount mark with mask == 0 and ignore_mask == FAN_OPEN. inode mark on a directory with mask == FAN_EXEC | FAN_EVENT_ON_CHILD.

The event: FAN_OPEN | FAN_EXEC | FAN_EVENT_ON_CHILD would be reported to group with the FAN_OPEN flag despite the fact that FAN_OPEN is in ignore mask of mount mark because the mark iteration loop skips over non-inode marks for events on child.

I'll try to work that case into the relevant LTP test to prove it and post a fix.

Quoted 6 lines
Quoted 3 lines

Ha, that's clever. But FAN_EXEC does not exist in current fanotify. We only have FAN_OPEN_EXEC... And I don't think we have any compound events.

Typo. I meant FAN_OPEN_EXEC and you can see from LTP test we do have at least this one compound event.

Yeah, I understood what you meant once I saw the changelog of your patch. Thanks for it.

Quoted 1 line

We could also split it if we wanted to, but no reason to do it now.

Agreed.

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR