Re: [PATCH 2/2] Fix possible leakage of blocks in UDF

[Andrew Morton <akpm@linux-foundation.org>]

On Fri, 1 Jun 2007 20:49:26 +0400 Cyrill Gorcunov <gorcunov@gmail.com> wrote:

> [Eric Sandeen - Thu, May 31, 2007 at 12:46:15PM -0500]
> | Cyrill Gorcunov wrote:
> | 
> | >Eric, could you please try the following:
> | >
> | >1) declare the spinlock in the top of inode.c as
> | >
> | >	DEFINE_SPINLOCK(udf_drop_lock);
> | >
> | >2) replace in udf_drop_inode()
> | >
> | >	kernel_lock -> spin_lock(&udf_drop_lock);
> | >	kernel_unlock -> spin_unlock(&udf_drop_lock);
> | >
> | >I'm not sure if it help but you may try ;)
> | >
> | >		Cyrill
> | >
> | 
> | I'm sure it'll avoid the deadlock but....
> | 
> | Any sense of what the BKL is actually trying to protect in this case?
> | 
> | Is it really only trying to prevent concurrent prealloc-discarders, or 
> | is there more?
> | 
> | -Eric
> | 
> 
> Hi Eric,
> it seems BKL only trying to protect from concurrent discard_prealloc.
> Moreover, a lot of UDF code does call iput with BKL held, so the only
> solution I see is to add spinlocks to udf_drop_inode... I'm making patch
> soon. Any comments?
> 

Recursive lock_kernel() is OK.

spin_lock() insode lock_kernel() is OK.

lock_kernel() inside spin_lock() is not OK, but if this was happening you'd
only rarely hit a deadlock and I think this locks up every time.

We don't know what's causing this hang, do we?