Skip to contentSkip to conversation

cachefiles: Fix excess dput() after end_removing()

PATCH Superseded David Howells 3 participants 7 messages Mar 24 – Mar 26

Subsystems (3)
lore
Share & tools
lore

David Howells

author maintainer
#0 lore raw

PATCH 1/1 cachefiles: Fix excess dput() after end_removing()

When cachefiles_cull() calls cachefiles_bury_object(), the latter eats the former's ref on the victim dentry that it obtained from cachefiles_lookup_for_cull(). However, commit 7bb1eb45e43c left the dput of the victim in place, resulting in occasional:

  WARNING: fs/dcache.c:829 at dput.part.0+0xf5/0x110, CPU#7: cachefilesd/11831
  cachefiles_cull+0x8c/0xe0 [cachefiles]
  cachefiles_daemon_cull+0xcd/0x120 [cachefiles]
  cachefiles_daemon_write+0x14e/0x1d0 [cachefiles]
  vfs_write+0xc3/0x480
  ...

reports.

Fix this by removing the dput().

Fixes 7bb1eb45e43c ("VFS: introduce start_removing_dentry()")
cc NeilBrown <[email protected]>
cc Paulo Alcantara <[email protected]>
---
 fs/cachefiles/namei.c |    1 -
 1 file changed, 1 deletion(-)
fs/cachefiles/namei.c +0 -1
index bdac2f33edf3..e2023e78e4df 100644
@@ -795,7 +795,6 @@ int cachefiles_cull(struct cachefiles_cache *cache, struct dentry *dir,
 
 	ret = cachefiles_bury_object(cache, NULL, dir, victim,
 				     FSCACHE_OBJECT_WAS_CULLED);
-	dput(victim);
 	if (ret < 0)
 		goto error;
 
author maintainer
#1 lore raw
Quoted 1 line · David Howells
Fixes 7bb1eb45e43c ("VFS: introduce start_removing_dentry()")

Actually, this should probably be:

Fixes bd6ede8a06e8 ("VFS/nfsd/cachefiles/ovl: introduce start_removing() and end_removing()")
#2 lore raw
Quoted 8 lines · David Howells
Quoted 1 line · David Howells
Fixes 7bb1eb45e43c ("VFS: introduce start_removing_dentry()")

Actually, this should probably be:

Fixes bd6ede8a06e8 ("VFS/nfsd/cachefiles/ovl: introduce start_removing() and end_removing()")

I think it is the correct Fixes tag, but I'm not sure that this is actually the right fix. 7bb1eb45e43c switched other callers of cachefiles_bury_object to use start_removing_dentry, which gets an additional ref, and removed the extra dget from cachefiles_bury_object. In the cachefiles_cull case however, the dentry is from start_removing and has a single ref on entry to cachefiles_bury_object, which is an issue as "rep" may be used there after end_removing may have put the last ref. So the correct is probably for cachefiles_cull to add a dget() before the call to cachefiles_bury_object.

Marc

author maintainer
#3 lore raw
Quoted 10 lines · Marc Dionne

I think it is the correct Fixes tag, but I'm not sure that this is actually the right fix. 7bb1eb45e43c switched other callers of cachefiles_bury_object to use start_removing_dentry, which gets an additional ref, and removed the extra dget from cachefiles_bury_object. In the cachefiles_cull case however, the dentry is from start_removing and has a single ref on entry to cachefiles_bury_object, which is an issue as "rep" may be used there after end_removing may have put the last ref. So the correct is probably for cachefiles_cull to add a dget() before the call to cachefiles_bury_object.

Ugh. You're right.

The problem is that we're calling start_removing() without knowing whether we can just unlink the object. I wonder if I need to do the lookup in cachefiles_lookup_for_cull() and only then call start_removing_dentry() if it's not a directory (directories get moved to the graveyard for cachefilesd to tear down).

I think the right solution is actually to move start_removing_dentry() down into cachefiles_bury_object() and make it contingent on the dentry being a non-dir.

David

#4 lore raw
Quoted 21 lines · David Howells
Quoted 10 lines · Marc Dionne

I think it is the correct Fixes tag, but I'm not sure that this is actually the right fix. 7bb1eb45e43c switched other callers of cachefiles_bury_object to use start_removing_dentry, which gets an additional ref, and removed the extra dget from cachefiles_bury_object. In the cachefiles_cull case however, the dentry is from start_removing and has a single ref on entry to cachefiles_bury_object, which is an issue as "rep" may be used there after end_removing may have put the last ref. So the correct is probably for cachefiles_cull to add a dget() before the call to cachefiles_bury_object.

Ugh. You're right.

The problem is that we're calling start_removing() without knowing whether we can just unlink the object. I wonder if I need to do the lookup in cachefiles_lookup_for_cull() and only then call start_removing_dentry() if it's not a directory (directories get moved to the graveyard for cachefilesd to tear down).

Quoted 7 lines

I think the right solution is actually to move start_removing_dentry() down into cachefiles_bury_object() and make it contingent on the dentry being a non-dir.

David

cachesfiles_bury_object() has a comment saying:

* On entry there must be at least 2 refs on rep, one will be dropped on exit.

and this is consistent with the code in that function.

It is called from 3 places.
- cachefiles_invalidate_cookie(), cachesfiles_look_up_object(), and cachefiles_acquire_volume() all precede it with a start_removing_dentry() which results in 2 references to the dentry (the original and and extra which it takes) - so that fits with the comment.

- cachesfiles_cull() preceeds it with cachesfiles_lookup_for_cull() which uses start_removing() which returns with 1 reference to the dentry. As the dentry didn't pre-exist, there is only one ref. So this is incorrect.

cachesfiles_cull() needs to take an extra reference to victim so that when cachefiles_busy_object() calls end_removing, it still has a valid reference.

So I think

fs/cachefiles/namei.c +1 -1
@@ -781,7 +781,7 @@ int cachefiles_cull(struct cachefiles_cache *cache, struct dentry *dir,
 	if (ret < 0)
 		goto error_unlock;
 
-	ret = cachefiles_bury_object(cache, NULL, dir, victim,
+	ret = cachefiles_bury_object(cache, NULL, dir, dget(victim),
 				     FSCACHE_OBJECT_WAS_CULLED);
 	dput(victim);
 	if (ret < 0)

would be a correct fix.

If you agree I can post a properly formated patch which explanation.

Thanks,
NeilBrown

author maintainer
#5 lore raw

fs/cachefiles/namei.c int cachefiles_cull(struct cachefiles_cache *cache, struct dentry *dir, · line 784

Quoted 2 lines · NeilBrown
fs/cachefiles/namei.c +1 -1
-	ret = cachefiles_bury_object(cache, NULL, dir, victim,
+	ret = cachefiles_bury_object(cache, NULL, dir, dget(victim),

I would prefer the dget() to be on a line on its own before this one to make it easier to spot.

Quoted 1 line

If you agree I can post a properly formated patch which explanation.

That would be great!

Thanks,
David

author maintainer
#6 lore raw

PATCH v2 1/1 cachefiles: Fix excess dput() after end_removing()

In reply to David Howells

When cachefiles_cull() calls cachefiles_bury_object(), the latter eats the former's ref on the victim dentry that it obtained from cachefiles_lookup_for_cull(). However, commit 7bb1eb45e43c left the dput of the victim in place, resulting in occasional:

  WARNING: fs/dcache.c:829 at dput.part.0+0xf5/0x110, CPU#7: cachefilesd/11831
  cachefiles_cull+0x8c/0xe0 [cachefiles]
  cachefiles_daemon_cull+0xcd/0x120 [cachefiles]
  cachefiles_daemon_write+0x14e/0x1d0 [cachefiles]
  vfs_write+0xc3/0x480
  ...

reports.

Actually, it's worse than that: cachefiles_bury_object() eats the ref it was given - and then may continue to access the now-unref'd dentry it if it turns out to be a directory. So simply removing the aberrant dput() is not sufficient.

Fix this by making cachefiles_bury_object() retain the ref itself around end_removing() if it needs to keep it and then drop the ref before returning.

Fixes bd6ede8a06e8 ("VFS/nfsd/cachefiles/ovl: introduce start_removing() and end_removing()")
cc NeilBrown <[email protected]>
cc Paulo Alcantara <[email protected]>
---
 fs/cachefiles/namei.c |   36 +++++++++++++++++++++---------------
 1 file changed, 21 insertions(+), 15 deletions(-)
fs/cachefiles/namei.c +21 -15
index e5ec90dccc27..20138309733f 100644
@@ -287,14 +287,14 @@ int cachefiles_bury_object(struct cachefiles_cache *cache,
 	if (!d_is_dir(rep)) {
 		ret = cachefiles_unlink(cache, object, dir, rep, why);
 		end_removing(rep);
-
 		_leave(" = %d", ret);
 		return ret;
 	}
 
 	/* directories have to be moved to the graveyard */
 	_debug("move stale object to graveyard");
-	end_removing(rep);
+	dget(rep);
+	end_removing(rep); /* Drops ref on rep */
 
 try_again:
 	/* first step is to make up a grave dentry in the graveyard */
@@ -304,8 +304,10 @@ int cachefiles_bury_object(struct cachefiles_cache *cache,
 
 	/* do the multiway lock magic */
 	trap = lock_rename(cache->graveyard, dir);
-	if (IS_ERR(trap))
-		return PTR_ERR(trap);
+	if (IS_ERR(trap)) {
+		ret = PTR_ERR(trap);
+		goto out;
+	}
 
 	/* do some checks before getting the grave dentry */
 	if (rep->d_parent != dir || IS_DEADDIR(d_inode(rep))) {
@@ -313,25 +315,27 @@ int cachefiles_bury_object(struct cachefiles_cache *cache,
 		 * lock */
 		unlock_rename(cache->graveyard, dir);
 		_leave(" = 0 [culled?]");
-		return 0;
+		ret = 0;
+		goto out;
 	}
 
+	ret = -EIO;
 	if (!d_can_lookup(cache->graveyard)) {
 		unlock_rename(cache->graveyard, dir);
 		cachefiles_io_error(cache, "Graveyard no longer a directory");
-		return -EIO;
+		goto out;
 	}
 
 	if (trap == rep) {
 		unlock_rename(cache->graveyard, dir);
 		cachefiles_io_error(cache, "May not make directory loop");
-		return -EIO;
+		goto out;
 	}
 
 	if (d_mountpoint(rep)) {
 		unlock_rename(cache->graveyard, dir);
 		cachefiles_io_error(cache, "Mountpoint in cache");
-		return -EIO;
+		goto out;
 	}
 
 	grave = lookup_one(&nop_mnt_idmap, &QSTR(nbuffer), cache->graveyard);
@@ -343,11 +347,12 @@ int cachefiles_bury_object(struct cachefiles_cache *cache,
 
 		if (PTR_ERR(grave) == -ENOMEM) {
 			_leave(" = -ENOMEM");
-			return -ENOMEM;
+			ret = -ENOMEM;
+			goto out;
 		}
 
 		cachefiles_io_error(cache, "Lookup error %ld", PTR_ERR(grave));
-		return -EIO;
+		goto out;
 	}
 
 	if (d_is_positive(grave)) {
@@ -362,7 +367,7 @@ int cachefiles_bury_object(struct cachefiles_cache *cache,
 		unlock_rename(cache->graveyard, dir);
 		dput(grave);
 		cachefiles_io_error(cache, "Mountpoint in graveyard");
-		return -EIO;
+		goto out;
 	}
 
 	/* target should not be an ancestor of source */
@@ -370,7 +375,7 @@ int cachefiles_bury_object(struct cachefiles_cache *cache,
 		unlock_rename(cache->graveyard, dir);
 		dput(grave);
 		cachefiles_io_error(cache, "May not make directory loop");
-		return -EIO;
+		goto out;
 	}
 
 	/* attempt the rename */
@@ -404,8 +409,10 @@ int cachefiles_bury_object(struct cachefiles_cache *cache,
 	__cachefiles_unmark_inode_in_use(object, d_inode(rep));
 	unlock_rename(cache->graveyard, dir);
 	dput(grave);
-	_leave(" = 0");
-	return 0;
+	_leave(" = %d", ret);
+out:
+	dput(rep);
+	return ret;
 }
 
 /*
@@ -812,7 +819,6 @@ int cachefiles_cull(struct cachefiles_cache *cache, struct dentry *dir,
 
 	ret = cachefiles_bury_object(cache, NULL, dir, victim,
 				     FSCACHE_OBJECT_WAS_CULLED);
-	dput(victim);
 	if (ret < 0)
 		goto error;