Date:	Sun, 06 Jan 2008 11:07:52 +0900
From:	Tejun Heo <htejun@...il.com>
To:	Al Viro <viro@...IV.linux.org.uk>
CC:	Gabor Gombas <gombasg@...aki.hu>,
	Dave Young <hidave.darkstar@...il.com>,
	linux-kernel@...r.kernel.org, bluez-devel@...ts.sf.net,
	Greg KH <greg@...ah.com>, ebiederm@...ssion.com
Subject: Re: [Bluez-devel] Oops involving RFCOMM and sysfs

Hello,

Al Viro wrote:
> On Sat, Jan 05, 2008 at 11:30:25PM +0900, Tejun Heo wrote:
>>> Assuming that this is what we get, everything looks explainable - we
>>> have sysfs_rename_dir() calling sysfs_get_dentry() while the parent
>>> gets evicted.  We don't have any exclusion, so while we are playing
>>> silly buggers with lookups in sysfs_get_dentry() we have parent become
>>> negative; the rest is obvious...
>> That part of code is walking down the sysfs tree from the s_root of
>> sysfs hierarchy and on each step parent is held using dget() while being
>> referenced, so I don't think they can turn negative there.
> 
> Turn?  Just what stops you from getting a negative (and unhashed) from
> lookup_one_noperm() and on the next iteration being buggered on mutex_lock()?

Right, I haven't thought about that.  When sysfs_get_dentry() is called,
@sd is always valid so unless there was existing negative dentry, lookup
is guaranteed to return positive dentry, but by populating dcache with
negative dentry before a node is created, things can go wrong.  I don't
think that's what's going on here tho.  If that was the case, the
while() loop looking up the next sd to lookup (@cur) should have blown
up as negative dentry will have NULL d_fsdata which doesn't match any sd.

I guess what's needed here is d_revalidate() as other distributed
filesystems do.  I'll test whether this can be actually triggered and
prepare a fix.  Thanks a lot for pointing out the problem.

>>> AFAICS, the locking here is quite broken and frankly, sysfs_get_dentry()
>>> and the way it plays with fs/namei.c are ucking fugly.
>> Can you elaborate a bit?  The locking in sysfs is unconventional but
>> that's mostly from necessity.  It has dual interface - vfs and driver
>> model && vfs data structures (dentry and inode) are too big to always
>> keep around, so it basically becomes a small distributed file system
>> where the backing data can change asynchronously.
> 
> ... with all fun that creates.  As it is, you have those async changers
> of backing data using VFS locking _under_ sysfs locks via lookup_one_noperm()
> and yet it needs sysfs_mutex inside sysfs_lookup().  So you can't have
> sysfs_get_dentry() under it.  So you don't have exclusion with arseloads
> of sysfs tree changes in there.  Joy...

There are two locks.  sysfs_rename_mutex and sysfs_mutex.
sysfs_rename_mutex is above VFS locks while sysfs_mutex is below VFS
locks.  sysfs_rename_mutex() protects against move/rename which can
change the ancestry of a held sysfs_dirent while sysfs_mutex protects
the sd hierarchy itself.  Locking can be wrong if sysfs_rename_mutex
locking is missing from the places where ancestry of a held sd can
change but I can't find one ATM.  If I'm missing your point again, feel
free to scream at me.  :-)

As it's unnecessarily unintuitive, there's a pending change to rename
sysfs_rename_mutex and use it to protect the whole tree structure to
make locking simpler while using sysfs_mutex to guard VFS access such
that the locking hierarchy plainly becomes sysfs_rename_mutex - VFS
locks - sysfs_mutex where all internal sysfs structure is protected by
the outer mutex and the inner one just protects VFS accesses.

> Frankly, with the current state of sysfs the last vestiges of arguments
> used to push it into the tree back then are dead and buried.  I'm not
> blaming you, BTW - the shitpile *did* grow past the point where its
> memory footprint became far too large and something needed to be done.
> Unfortunately, it happened too late for that something being "get rid
> of the entire mess" and now we are saddled with it for good.

Yeah, it's too late to get rid of sysfs and regardless implementation
ugliness, which BTW I think has improved a lot during last six or so
months, it's now pretty useful and important to drivers, so I guess the
only option is trying hard to make it better.

Oh, BTW, the ugly lookup_one_noperm() can be removed if LOOKUP_NOPERM
flag is added.  The only reason sysfs_lookup() uses the specialized
lookup is to avoid permission check.

Thanks.

-- 
tejun
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists