Autofs vs shrink_dcache_for_umount() problem
Subsystems (2)
A message that is not in the archive (sent off-list or not yet received): [email protected]
Quoted 2 lines · Andrew Morton
But why was yum (or an RPM script) dinking with loop and mounting/unmounting things?
It may have restarted the automounter... this may cause mounts and unmounts...
I've done a bit of probing, and when both ext3 and NFS go into the superblock squidger with a tree that only has a root dentry, that root dentry has a usage count of 1:
### Shrink root c629b088{1}
-->shrink_dcache_for_umount_subtree(c629b088)
- AT c629b088{0}
- consume c629b088{0}
### Shrunk
### Shrink root c621d710{1}
-->shrink_dcache_for_umount_subtree(c621d710)
- AT c621d710{0}
- consume c621d710{0}
### ShrunkBut when autofs4 goes in there with a minimal tree:
### Shrink root c6b714b0{2}
-->shrink_dcache_for_umount_subtree(c6b714b0)
- AT c6b714b0{1}
- consume c6b714b0{1}
BUG: Dentry c6b714b0{i=1a80,n=/} still in use (1) [unmount of autofs autofs]
------------[ cut here ]------------
kernel BUG at fs/dcache.c:618!It's got a usage count of *2*. That means the root dentry is still in use somewhere beyond the superblock s_root reference.
I think this in autofs4_fill_super() is the problem:
/* * Take a reference to the root dentry so we get a chance to * clean up the dentry tree on umount. * See autofs4_force_release. */ sbi->root = dget(root);
autofs4 violates the assumptions that I was told I could make. autofs4 expects to clean up its dentries in autofs4_put_super(), when it should perhaps really be overloading the kill_sb() op and doing it before calling kill_anon_super().
I've attached the pair of debugging patches I used to generate the above.
David
fs/dcache.c +14 -1
@@ -597,7 +597,20 @@ static void shrink_dcache_for_umount_sub do { struct inode *inode; - BUG_ON(atomic_read(&dentry->d_count) != 0); + if (atomic_read(&dentry->d_count) != 0) { + printk(KERN_ERR + "BUG: Dentry %p{i=%lx,n=%s}" + " still in use (%d)" + " [unmount of %s %s]\n", + dentry, + dentry->d_inode ? + dentry->d_inode->i_ino : 0UL, + dentry->d_name.name, + atomic_read(&dentry->d_count), + dentry->d_sb->s_type->name, + dentry->d_sb->s_id); + BUG(); + } parent = dentry->d_parent; if (parent == dentry)
fs/dcache.c +11 -0
@@ -558,6 +558,8 @@ static void shrink_dcache_for_umount_subBUG_ON (!IS_ROOT (dentry)); +printk ("-->shrink_dcache_for_umount_subtree(%p)\n" , dentry); +/* detach this root from the system */ spin_lock (&dcache_lock);if (!list_empty (&dentry->d_lru)) {
@@ -568,6 +570,8 @@ static void shrink_dcache_for_umount_subspin_unlock (&dcache_lock);for (;;) { +printk ("- AT %p{%d}\n" , dentry,atomic_read (&dentry->d_count)); +/* descend to the first leaf in the current subtree */ while (!list_empty (&dentry->d_subdirs)) {struct dentry *loop;
@@ -590,6 +594,8 @@ static void shrink_dcache_for_umount_sub/* move to the first child */ dentry =list_entry (dentry->d_subdirs.next, struct dentry, d_u.d_child); + +printk ("- descend %p{%d}\n" , dentry,atomic_read (&dentry->d_count)); }/* consume the dentries from this leaf up through its parents
@@ -643,6 +649,7 @@ #endifreturn ; dentry = parent; +printk ("- ascend %p{%d}\n" , dentry,atomic_read (&dentry->d_count)); }while (list_empty (&dentry->d_subdirs));
@@ -671,13 +678,17 @@ void shrink_dcache_for_umount(struct sup dentry = sb->s_root; sb->s_root =NULL ; +printk ("### Shrink root %p{%d}\n" , dentry,atomic_read (&dentry->d_count));atomic_dec (&dentry->d_count);shrink_dcache_for_umount_subtree (dentry);while (!hlist_empty (&sb->s_anon)) { +printk ("### Shrink anon\n" ); dentry = hlist_entry(sb->s_anon.first, struct dentry, d_hash);shrink_dcache_for_umount_subtree (dentry); } + +printk ("### Shrunk\n" ); }/*
Ian Kent
maintainerQuoted 46 lines · David Howells
Quoted 2 lines · Andrew Morton
But why was yum (or an RPM script) dinking with loop and mounting/unmounting things?
It may have restarted the automounter... this may cause mounts and unmounts...
I've done a bit of probing, and when both ext3 and NFS go into the superblock squidger with a tree that only has a root dentry, that root dentry has a usage count of 1:
### Shrink root c629b088{1} -->shrink_dcache_for_umount_subtree(c629b088) - AT c629b088{0} - consume c629b088{0} ### Shrunk ### Shrink root c621d710{1} -->shrink_dcache_for_umount_subtree(c621d710) - AT c621d710{0} - consume c621d710{0} ### ShrunkBut when autofs4 goes in there with a minimal tree:
### Shrink root c6b714b0{2} -->shrink_dcache_for_umount_subtree(c6b714b0) - AT c6b714b0{1} - consume c6b714b0{1} BUG: Dentry c6b714b0{i=1a80,n=/} still in use (1) [unmount of autofs autofs] ------------[ cut here ]------------ kernel BUG at fs/dcache.c:618!It's got a usage count of *2*. That means the root dentry is still in use somewhere beyond the superblock
s_rootreference.I think this in
autofs4_fill_super()is the problem:/* * Take a reference to the root dentry so we get a chance to * clean up the dentry tree on umount. * See autofs4_force_release. */ sbi->root = dget(root);
Quoted 4 lines
autofs4 violates the assumptions that I was told I could make. autofs4 expects to clean up its dentries in
autofs4_put_super(), when it should perhaps really be overloading thekill_sb()op and doing it before callingkill_anon_super().
Right you are. Sorry.
Looks like I've got a bit of work to do then.
The main reason this exists is that if the daemon is summarily killed dentrys with a non-zero reference count can be left in the tree so they need to be culled.
I wish I'd had time to pay more attention to the work your doing David. A super brief overview would be really helpful if you have the time.
Sounds like all I need to do is move the function. Overloading kill_sb sounds good. I'll investigate.
I really need to try and motivate myself to do some work on my house this weekend so it may be a little while before I get back.
Ian