In mark_remove_from_mask() the mark is destroyed regardless of whether the
event mask or ignore mask is cleared. We should only destroy the mark if the
event mask is cleared.
Signed-off-by: Lino Sanfilippo <[email protected]>
---
fs/notify/fanotify/fanotify_user.c | 2 +-
1 files changed, 1 insertions(+), 1 deletions(-)
Please apply this after patch "Dont allow a mask of 0 if setting or removing a
mark" which i sent today.
diff --git a/fs/notify/fanotify/fanotify_user.c b/fs/notify/fanotify/fanotify_user.c
index 207cdeb..c9143a0 100644
--- a/fs/notify/fanotify/fanotify_user.c
+++ b/fs/notify/fanotify/fanotify_user.c
@@ -518,7 +518,7 @@ static __u32 fanotify_mark_remove_from_mask(struct fsnotify_mark *fsn_mark,
}
spin_unlock(&fsn_mark->lock);
- if (!(oldmask & ~mask))
+ if (!(flags & FAN_MARK_IGNORED_MASK) && !(oldmask & ~mask))
fsnotify_destroy_mark(fsn_mark);
return mask & oldmask;
--
1.5.6.5
On Mon, 2010-11-22 at 18:52 +0100, Lino Sanfilippo wrote:
> In mark_remove_from_mask() the mark is destroyed regardless of whether the
> event mask or ignore mask is cleared. We should only destroy the mark if the
> event mask is cleared.
>
> Signed-off-by: Lino Sanfilippo <[email protected]>
Hmmmm, really I'm not sure if that is right either (but it's certainly
closer) What about something like:
diff --git a/fs/notify/fanotify/fanotify_user.c b/fs/notify/fanotify/fanotify_user.c
index 81df3ad..29fbf17 100644
--- a/fs/notify/fanotify/fanotify_user.c
+++ b/fs/notify/fanotify/fanotify_user.c
@@ -527,7 +527,7 @@ static __u32 fanotify_mark_remove_from_mask(struct fsnotify_mark *fsn_mark,
}
spin_unlock(&fsn_mark->lock);
- if (!(oldmask & ~mask))
+ if (!fsn_mark->mask && !fsn_mark->ignored_mask)
fsnotify_destroy_mark(fsn_mark);
return mask & oldmask;
On Tue, Nov 23, 2010 at 02:51:19PM -0500, Eric Paris wrote:
>
> Hmmmm, really I'm not sure if that is right either (but it's certainly
> closer) What about something like:
>
> diff --git a/fs/notify/fanotify/fanotify_user.c b/fs/notify/fanotify/fanotify_user.c
> index 81df3ad..29fbf17 100644
> --- a/fs/notify/fanotify/fanotify_user.c
> +++ b/fs/notify/fanotify/fanotify_user.c
> @@ -527,7 +527,7 @@ static __u32 fanotify_mark_remove_from_mask(struct fsnotify_mark *fsn_mark,
> }
> spin_unlock(&fsn_mark->lock);
>
> - if (!(oldmask & ~mask))
> + if (!fsn_mark->mask && !fsn_mark->ignored_mask)
> fsnotify_destroy_mark(fsn_mark);
>
> return mask & oldmask;
>
>
Yep, youre right, we should also check the ignore mask before we destroy the mark.
BUT:
1. There is this flag FAN_MARK_ONDIR which is set implicitly in the ignore mask
whenever it has not explicitly been set by the user (see fanotify_mark_add_to_mask()).
If that flag has been set the ignore mask will never get cleared unless the
user does something like
fanotify_mark(fd, FAN_MARK_REMOVE | FAN_MARK_IGNORED_MASK, FAN_MARK_ONDIR,...)
which presumes that he knows that this flag has been set.
Btw.:
every time a user _forgets_ to explicitly set FAN_MARK_ONDIR, it will be set
in the ignored mask and thus events on dirs are skipped. Thus calls like
fanotify_mark(fd, FAN_MARK_ADD, FAN_MARK_ONDIR, ...) /* get dir events */
fanotify_mark(fd, FAN_MARK_ADD, ...) /* add mark for some kind of event */
will result in dir events being ignored by the second call to fanotify_mark(),
although the user has requested those events in his first call.
This is very likely not what the user expected. Is there a reason why ONDIR is
set implicitly in the ignore mask?
Otherwise i would suggest to not set it implicitly in mark->ignored_mask,
but to set it in mark->mask if requested by the user. Then we could ignore dir
events as long as the flag has not been set there.
2. I just realized that we cant simply call destroy_mark() if the masks are 0.
There may be one or more concurrent processes calling fsnotify_find_inode_mark()
(see fanotify_add_inode_mark()) and get the mark we are about to destroy at the
same time.
I will take a closer look at it, but it seems to be difficult to me to safely
call destroy_mark() as long as we are not in the context of fanotify_release() (in
which we dont have to deal with concurrency like that any more).
On Wed, 2010-11-24 at 13:31 +0100, Lino Sanfilippo wrote:
> 2. I just realized that we cant simply call destroy_mark() if the masks are 0.
> There may be one or more concurrent processes calling fsnotify_find_inode_mark()
> (see fanotify_add_inode_mark()) and get the mark we are about to destroy at the
> same time.
>
> I will take a closer look at it, but it seems to be difficult to me to safely
> call destroy_mark() as long as we are not in the context of fanotify_release() (in
> which we dont have to deal with concurrency like that any more).
I guess it is a question of safe vs racy. Yes it is safe, nothing will
explode or panic. But we might have a race between one task removing an
event type causing the mask to go to 0 and we should destroy the mark
and another task adding an event type. If it raced just right we might
destroy the mark after the second task added to it. I guess we really
need to serialize fsnotify_mark() per group to solve the race...
Do you want to take a stab at fixing these things or should I?
-Eric
On Tue, Nov 30, 2010 at 01:16:35PM +0100, Lino Sanfilippo wrote:
> I guess it is a question of safe vs racy. Yes it is safe, nothing will
> explode or panic. But we might have a race between one task removing an
> event type causing the mask to go to 0 and we should destroy the mark
> and another task adding an event type. If it raced just right we might
> destroy the mark after the second task added to it. I guess we really
> need to serialize fsnotify_mark() per group to solve the race...
>
> Do you want to take a stab at fixing these things or should I?
>
> -Eric
IMHO the right thing to serialize this would be to do
LOCK(groups->mark_lock)
- get the inode mark
- set the marks mask
- possibly destroy the mask
UNLOCK(groups->mark_lock)
But we cant do this since setting the marks mask requires the lock of the mark
- which would mean an incorrect lock order according to fsnotify_add_mark():
mark->lock
group->mark_lock
inode->i_lock
What we could do very easily is use another mutex instead (use an existing one like the
groups notification_mutex, or a completely new one) which is responsible for synchronising
add_mark()/remove_mark().
If that solution is ok I'll prepare the patches for it. Otherwise i am not sure how to solve this...
Lino
On Tue, 2010-11-30 at 16:59 +0100, Lino Sanfilippo wrote:
> On Tue, Nov 30, 2010 at 01:16:35PM +0100, Lino Sanfilippo wrote:
>
> > I guess it is a question of safe vs racy. Yes it is safe, nothing will
> > explode or panic. But we might have a race between one task removing an
> > event type causing the mask to go to 0 and we should destroy the mark
> > and another task adding an event type. If it raced just right we might
> > destroy the mark after the second task added to it. I guess we really
> > need to serialize fsnotify_mark() per group to solve the race...
> >
> > Do you want to take a stab at fixing these things or should I?
> >
> > -Eric
>
> IMHO the right thing to serialize this would be to do
>
> LOCK(groups->mark_lock)
> - get the inode mark
> - set the marks mask
> - possibly destroy the mask
> UNLOCK(groups->mark_lock)
>
> But we cant do this since setting the marks mask requires the lock of the mark
> - which would mean an incorrect lock order according to fsnotify_add_mark():
>
> mark->lock
> group->mark_lock
> inode->i_lock
>
> What we could do very easily is use another mutex instead (use an existing one like the
> groups notification_mutex, or a completely new one) which is responsible for synchronising
> add_mark()/remove_mark().
I'd think a new per group mutex would be the right way to go. I'm not
sure how I feel about notification_mutex. I guess you can go ahead and
overload it and we can split it off later if someone finds it to be a
performance blocker.
-Eric