Skip to contentSkip to conversation

Revert "cgroup: remove redundant variable in cgroup_mount()"

PATCH Tejun Heo 2 participants 4 messages Sep 26, 2014 1 message not in the archive

Subsystems (2)
lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

Tejun Heo

author maintainer Applied
#1 lore raw
Quoted 10 lines · Zefan Li

This reverts commit 0c7bf3e8cab7900e17ce7f97104c39927d835469.

If there are child cgroups in the cgroupfs and then we umount it, the superblock will be destroyed but the cgroup_root will be kept around. When we mount it again, cgroup_mount() will find this cgroup_root and allocate a new sb for it.

So with this commit we will be trapped in a dead loop in the case described above, because kernfs_pin_sb() keeps returning NULL.

Quoted 6 lines

Currently I don't see how we can avoid using both pinned_sb and new_sb, so just revert it.

Applied to cgroup/for-3.18.

Thanks.

Signature · 1 line
tejun
#2 lore raw
Quoted 11 lines · Tejun Heo
Quoted 10 lines · Zefan Li

This reverts commit 0c7bf3e8cab7900e17ce7f97104c39927d835469.

If there are child cgroups in the cgroupfs and then we umount it, the superblock will be destroyed but the cgroup_root will be kept around. When we mount it again, cgroup_mount() will find this cgroup_root and allocate a new sb for it.

So with this commit we will be trapped in a dead loop in the case described above, because kernfs_pin_sb() keeps returning NULL.

Quoted 8 lines
Quoted 6 lines

Currently I don't see how we can avoid using both pinned_sb and new_sb, so just revert it.

Applied to cgroup/for-3.18.

Er? It's a clear regression since 3.16; shouldn't it go into mainline before 3.17-final? Another thing: AFAICS, if kernfs_fill_super() fails, you get unbalanced kernfs_put() on the cleanup path - kernfs_kill_sb() will be called in those cases as well. IOW, we'd better move that

        kernfs_get(info->root->kn);

in kernfs_fill_super() all way up to the point before kernfs_get_inode(). Something like this:

---

fs/kernfs/mount.c +1 -1
index f973ae9..e5e92e3 100644
@@ -74,6 +74,7 @@ static int kernfs_fill_super(struct super_block *sb, unsigned long magic)
 	sb->s_magic = magic;
 	sb->s_op = &kernfs_sops;
 	sb->s_time_gran = 1;
+	kernfs_get(info->root->kn);
 
 	/* get root inode, initialize and unlock it */
 	mutex_lock(&kernfs_mutex);
@@ -90,7 +91,6 @@ static int kernfs_fill_super(struct super_block *sb, unsigned long magic)
 		pr_debug("%s: could not get root dentry!\n", __func__);
 		return -ENOMEM;
 	}
-	kernfs_get(info->root->kn);
 	root->d_fsdata = info->root->kn;
 	sb->s_root = root;
 	sb->s_d_op = &kernfs_dops;
author maintainer
#3 lore raw
Quoted 3 lines · Al Viro
Quoted 2 lines · Tejun Heo
Quoted 1 line · Zefan Li

This reverts commit 0c7bf3e8cab7900e17ce7f97104c39927d835469.

...

Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

Applied to cgroup/for-3.18.

Er? It's a clear regression since 3.16; shouldn't it go into mainline before 3.17-final? Another thing: AFAICS, if kernfs_fill_super() fails,

Hmmm? This is reverting 0c7bf3e8cab7 which was committed six days ago to cgroup/for-3.18.

Quoted 21 lines

you get unbalanced kernfs_put() on the cleanup path - kernfs_kill_sb() will be called in those cases as well. IOW, we'd better move that

        kernfs_get(info->root->kn);

in kernfs_fill_super() all way up to the point before kernfs_get_inode(). Something like this:

---

fs/kernfs/mount.c +1 -0
index f973ae9..e5e92e3 100644
@@ -74,6 +74,7 @@ static int kernfs_fill_super(struct super_block *sb, unsigned long magic)
 	sb->s_magic = magic;
 	sb->s_op = &kernfs_sops;
 	sb->s_time_gran = 1;
+	kernfs_get(info->root->kn);
 
 	/* get root inode, initialize and unlock it */
 	mutex_lock(&kernfs_mutex);

fs/kernfs/mount.c static int kernfs_fill_super(struct super_block *sb, unsigned long magic) · line 93

Quoted 4 lines
fs/kernfs/mount.c +0 -1
@@ -90,7 +91,6 @@ static int kernfs_fill_super(struct super_block *sb, unsigned long magic)
 		pr_debug("%s: could not get root dentry!\n", __func__);
 		return -ENOMEM;
 	}
-	kernfs_get(info->root->kn);

Ooh, right. Will test & send the patch to Greg.

Thanks.

Signature · 1 line
tejun
#4 lore raw
Quoted 9 lines · Tejun Heo
Quoted 3 lines · Al Viro
Quoted 2 lines · Tejun Heo
Quoted 1 line · Zefan Li

This reverts commit 0c7bf3e8cab7900e17ce7f97104c39927d835469.

...

Quoted 4 lines
Quoted 4 lines
Quoted 3 lines
Quoted 7 lines
Quoted 4 lines
Quoted 1 line

Applied to cgroup/for-3.18.

Er? It's a clear regression since 3.16; shouldn't it go into mainline before 3.17-final? Another thing: AFAICS, if kernfs_fill_super() fails,

Hmmm? This is reverting 0c7bf3e8cab7 which was committed six days ago to cgroup/for-3.18.

*blink*
Right you are, I forgot that the tree I'm currently in gets -next as one of remotes, so seeing that commit there doesn't imply it being in mainline. My apologies. I plead going near-blind from digging through shitloads of callchains leading to __d_move() and __d_rehash() since yesterday ;-/ Oh, well - looks like either I'll go completely insane or fs/dcache.c will shrink quite a bit come next cycle...

Quoted 18 lines
Quoted 18 lines

you get unbalanced kernfs_put() on the cleanup path - kernfs_kill_sb() will be called in those cases as well. IOW, we'd better move that

        kernfs_get(info->root->kn);

in kernfs_fill_super() all way up to the point before kernfs_get_inode(). Something like this:

---

fs/kernfs/mount.c +1 -0
index f973ae9..e5e92e3 100644
@@ -74,6 +74,7 @@ static int kernfs_fill_super(struct super_block *sb, unsigned long magic)
 	sb->s_magic = magic;
 	sb->s_op = &kernfs_sops;
 	sb->s_time_gran = 1;
+	kernfs_get(info->root->kn);
 
Quoted 9 lines

fs/kernfs/mount.c static int kernfs_fill_super(struct super_block *sb, unsigned long magic) · line 93

Quoted 7 lines
fs/kernfs/mount.c +0 -1
 	/* get root inode, initialize and unlock it */
 	mutex_lock(&kernfs_mutex);
@@ -90,7 +91,6 @@ static int kernfs_fill_super(struct super_block *sb, unsigned long magic)
 		pr_debug("%s: could not get root dentry!\n", __func__);
 		return -ENOMEM;
 	}
-	kernfs_get(info->root->kn);

Ooh, right. Will test & send the patch to Greg.

OK... You'll need some way to force those failures to be able to test it - they are normally triggered when you are very low on memory, i.e. the failure exits won't trigger on normal testing (and near OOM you could easily fail not reaching that point at all).

Normally I'm doing such testing by something like slapping

	if (!strcmp(current->comm, "bugger"))
		return -ENOMEM;

in there, followed by cp /bin/mount /tmp/bugger; /tmp/bugger -t cgroup <...> Hell knows, maybe it would be worth to add a debugging option that would check some sysctl and if it's non-zero, looked through the call chain in kmem_cache_alloc() in search of such address, failing allocation if found. Never got around to doing it properly, might or might not be useful...