I am seeing missing inotify IN_ATTRIB events in the following situation:
1. "touch foo"
2. Make inotify watch "foo"
3. "ln foo bar"
--> Link count changed so I should have gotten an IN_ATTRIB.
4. "rm foo"
--> Link count changed so I should have gotten an IN_ATTRIB. (Or
IN_DELETE_SELF;
I don't care which.)
5. "ln bar foo && rm bar"
--> Still no events.
6. "mv foo bar"
--> I get IN_MOVED_SELF. Good!
7. "mv bar foo"
--> I get IN_MOVED_SELF. Good!
3+4 is pretty much the same as 6, so I really ought to be told that my
file has changed
name. I don't really care much about getting notified about 3, but
for completeness
it ought to be handled.
As far as I can see, the only way to be told about 4 is to put a watch
on the directory in
which foo resides. That is inelegant and has an inherent race condition.
This is with "Linux version 2.6.22.12-0.1-default" (SuSE 10.3)
Looking at current source, fs/namei.c, I notice that vfs_rename has a
fsnotify_move call
(which notified directory as well as files) whereas sys_link only has
a fsnotify_create call
(which notified the directory only).
Morten
Hi,
> I am seeing missing inotify IN_ATTRIB events in the following situation:
>
> 1. "touch foo"
>
> 2. Make inotify watch "foo"
>
> 3. "ln foo bar"
> --> Link count changed so I should have gotten an IN_ATTRIB.
>
> 4. "rm foo"
> --> Link count changed so I should have gotten an IN_ATTRIB. (Or
> IN_DELETE_SELF;
> I don't care which.)
>
> 5. "ln bar foo && rm bar"
> --> Still no events.
>
> 6. "mv foo bar"
> --> I get IN_MOVED_SELF. Good!
>
> 7. "mv bar foo"
> --> I get IN_MOVED_SELF. Good!
>
>
> 3+4 is pretty much the same as 6, so I really ought to be told that my
> file has changed
> name. I don't really care much about getting notified about 3, but
> for completeness
> it ought to be handled.
>
> As far as I can see, the only way to be told about 4 is to put a watch
> on the directory in
> which foo resides. That is inelegant and has an inherent race condition.
It looks sensible what you ask for ;) Wanna try the patch below?
Anybody has some objection to it?
Honza
--
Jan Kara <[email protected]>
SuSE CR Labs
---
Send inotify events to the inode itself when its link count has changed.
Signed-off-by: Jan Kara <[email protected]>
diff --git a/fs/namei.c b/fs/namei.c
index 3b993db..9a91130 100644
--- a/fs/namei.c
+++ b/fs/namei.c
@@ -2188,6 +2188,7 @@ int vfs_unlink(struct inode *dir, struct
/* We don't d_delete() NFS sillyrenamed files--they still exist. */
if (!error && !(dentry->d_flags & DCACHE_NFSFS_RENAMED)) {
+ fsnotify_link_count(dentry->d_inode);
d_delete(dentry);
}
@@ -2360,7 +2361,7 @@ int vfs_link(struct dentry *old_dentry,
error = dir->i_op->link(old_dentry, dir, new_dentry);
mutex_unlock(&old_dentry->d_inode->i_mutex);
if (!error)
- fsnotify_create(dir, new_dentry);
+ fsnotify_link(dir, new_dentry);
return error;
}
diff --git a/include/linux/fsnotify.h b/include/linux/fsnotify.h
index 2bd31fa..b21b818 100644
--- a/include/linux/fsnotify.h
+++ b/include/linux/fsnotify.h
@@ -92,6 +92,14 @@ static inline void fsnotify_inoderemove(
}
/*
+ * fsnotify_link_count - inode's link count changed
+ */
+static inline void fsnotify_link_count(struct inode *inode)
+{
+ inotify_inode_queue_event(inode, IN_ATTRIB, 0, NULL, NULL);
+}
+
+/*
* fsnotify_create - 'name' was linked in
*/
static inline void fsnotify_create(struct inode *inode, struct dentry *dentry)
@@ -103,6 +111,18 @@ static inline void fsnotify_create(struc
}
/*
+ * fsnotify_link - new hardlink in 'inode' directory
+ */
+static inline void fsnotify_link(struct inode *inode, struct dentry *new_dentry)
+{
+ inode_dir_notify(inode, DN_CREATE);
+ inotify_inode_queue_event(inode, IN_CREATE, 0, new_dentry->d_name.name,
+ new_dentry->d_inode);
+ fsnotify_link_count(new_dentry->d_inode);
+ audit_inode_child(new_dentry->d_name.name, new_dentry, inode);
+}
+
+/*
* fsnotify_mkdir - directory 'name' was created
*/
static inline void fsnotify_mkdir(struct inode *inode, struct dentry *dentry)
> Wanna try the patch below?
With this patch I am seeing a endless stream of IN_IGNORED events
for a removed watch. I don't see a reason that user space should
ever see any IN_IGNORED, but an endless steam is not good.
Utterly unrelated, inotify does not work in /proc/. The list archives
suggest that it isn't likely to start working anytime soon, but shouldn't
inotify_add_watch when fail with ENOSYS instead of pretending
it worked?
Morten
Failed to create inotify watch for /home/welinder/hi: No such file or directory
Created inotify watch 1 with mask 0x000007c0 for /home/welinder
# "touch hi" here
Got event 00000100 for 1
Created inotify watch 2 with mask 0x00000fc6 for /home/welinder/hi
Removing notify watch 1
Got event 00008000 for 1
Got event 00000004 for 2
# "rm hi" here
Got event 00000400 for 2
Removing notify watch 2
Got event 00008000 for 2
Failed to create inotify watch for /home/welinder/hi: No such file or directory
Created inotify watch 3 with mask 0x000007c0 for /home/welinder
Got event 00000200 for 3
Failed to create inotify watch for /home/welinder/hi: No such file or directory
Created inotify watch 3 with mask 0x000007c0 for /home/welinder
# "touch hi" here
Got event 00000100 for 3
Created inotify watch 4 with mask 0x00000fc6 for /home/welinder/hi
Removing notify watch 3
Got event 00008000 for 3
Got event 00000004 for 4
Got event 00000004 for 4
# "rm hi" here
Got event 00000400 for 4
Removing notify watch 4
Failed to create inotify watch for /home/welinder/hi: No such file or directory
Created inotify watch 5 with mask 0x000007c0 for /home/welinder
Got event 00008000 for 4
Got event 00008000 for 4
Got event 00008000 for 4
Got event 00008000 for 4
...
> With this patch I am seeing a endless stream of IN_IGNORED events
> for a removed watch.
Sorry, that problem was on the chair. side of the keyboard -- too much
turkey, I fear. It just happened to be easier to trigger with the patch's
double event on deletion.
I'd still love to see ENOSYS when inotify can't handle a given virtual
file.
I'm not sure "signed-off-by" if appropriate for me, but at least...
Acked-by: Morten Welinder <[email protected]>
Morten
This looks bad, though:
include/linux/fsnotify.h:121: warning: passing argument 2 of
'audit_inode_child' from incompatible pointer type
Missing "->d_inode"?
M.
> This looks bad, though:
>
> include/linux/fsnotify.h:121: warning: passing argument 2 of
> 'audit_inode_child' from incompatible pointer type
>
> Missing "->d_inode"?
That's the difference between 2.6.22 and 2.6.24-git against which I
wrote the patch :).
Honza
--
Jan Kara <[email protected]>
SuSE CR Labs