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). -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
