Re: [git pull] vfs pile 1

[Al Viro <viro@ZenIV.linux.org.uk>]

On Sun, Jan 08, 2012 at 03:04:30PM -0800, Linus Torvalds wrote:
> Hmm.
> 
> Al, Miklos, I just triggered the new inc_link() warning three times in
> very close succession by just doing a simple
> 
>               perf record -f -e cycles:pp make -j
> 
> (I don't think the "perf record" part was actually required or
> relevant, although it may have been instrumental in triggering just
> the right timing)
> 
> This was in a fully built tree, I'm just checking
> 
> The warning trace looks like this:
> 
>   [ 6933.446125] ------------[ cut here ]------------
>   [ 6933.446131] WARNING: at fs/inode.c:349 inc_nlink+0x30/0x40()
>   [ 6933.446133] Hardware name: System Product Name
>   [ 6933.446135] Pid: 22955, comm: perf Not tainted
> 3.2.0-03190-g972b2c719990 #72
>   [ 6933.446137] Call Trace:
>   [ 6933.446143]  [<ffffffff8102ee15>] warn_slowpath_common+0x75/0xb0
>   [ 6933.446145]  [<ffffffff8102ef15>] warn_slowpath_null+0x15/0x20
>   [ 6933.446147]  [<ffffffff810f1450>] inc_nlink+0x30/0x40
>   [ 6933.446151]  [<ffffffff8116f333>] ext4_symlink+0x113/0x2c0
>   [ 6933.446154]  [<ffffffff810e7b01>] vfs_symlink+0xa1/0xf0
>   [ 6933.446157]  [<ffffffff810e7c03>] sys_symlinkat+0xb3/0xd0
>   [ 6933.446159]  [<ffffffff810e7c31>] sys_symlink+0x11/0x20
>   [ 6933.446162]  [<ffffffff816916a2>] system_call_fastpath+0x16/0x1b
>   [ 6933.446164] ---[ end trace 550d4b2eb6b8fbc9 ]---
> 
> and the three warnings all triggered within 2/100th of a second of
> that one (the last one had a timestamp of 6933.465678), so they are
> probably all related to the same thing happening.

WTF is ext4_symlink() doing in case of long symlinks?  Look:
                drop_nlink(inode);
                err = ext4_orphan_add(handle, inode);
                ext4_journal_stop(handle);
[write symlink body]
                inc_nlink(inode);
                err = ext4_orphan_del(handle, inode);
oh, I see...  The comment above that re deadlocks and inability to do that
in a single transaction ;-/

OK, try this; that's equivalent to what they are doing and will not WARN_ON();
I hadn't checked other filesystems for similar tricks yet, so this has a good
chance of being incomplete.

diff --git a/fs/ext4/namei.c b/fs/ext4/namei.c
index 86edc45..2043f48 100644
--- a/fs/ext4/namei.c
+++ b/fs/ext4/namei.c
@@ -2315,7 +2315,7 @@ retry:
 			err = PTR_ERR(handle);
 			goto err_drop_inode;
 		}
-		inc_nlink(inode);
+		set_nlink(inode, 1);
 		err = ext4_orphan_del(handle, inode);
 		if (err) {
 			ext4_journal_stop(handle);