Modular filesystem using drop_inode would need inode_lock

Modular filesystem using drop_inode would need inode_lock

[Luck, Tony]

This is mostly a logical inconsistency at the moment (since the
only filesystem that has a "drop_inode" function defined in its
super_operations is hugetlbfs, and it is unlikely to move out of
the kernel and into a module).  But the ->drop_inode() function
is called with inode_lock held, and it is expected to drop the
lock ... which would be impossible for a module as the lock is
not exported.

Signed-off-by: <tony.luck@intel.com>



--- old/fs/inode.c	2004-07-09 14:05:16.986225857 -0700
+++ new/fs/inode.c	2004-07-09 14:05:42.198492986 -0700
@@ -81,6 +81,7 @@
  * the i_state of an inode while it is in use..
  */
 spinlock_t inode_lock = SPIN_LOCK_UNLOCKED;
+EXPORT_SYMBOL(inode_lock);
 
 /*
  * iprune_sem provides exclusion between the kswapd or
try_to_free_pages

Re: Modular filesystem using drop_inode would need inode_lock

[Christoph Hellwig]

On Fri, Jul 09, 2004 at 02:43:32PM -0700, Luck, Tony wrote:
> This is mostly a logical inconsistency at the moment (since the
> only filesystem that has a "drop_inode" function defined in its
> super_operations is hugetlbfs, and it is unlikely to move out of
> the kernel and into a module).  But the ->drop_inode() function
> is called with inode_lock held, and it is expected to drop the
> lock ... which would be impossible for a module as the lock is
> not exported.

->drop_inode is a bad idea anyway.  Please send a pointer to your
filesystem so we can fix it.

Re: Modular filesystem using drop_inode would need inode_lock

[Christoph Hellwig]

On Fri, Jul 09, 2004 at 02:43:32PM -0700, Luck, Tony wrote:
> This is mostly a logical inconsistency at the moment (since the
> only filesystem that has a "drop_inode" function defined in its
> super_operations is hugetlbfs, and it is unlikely to move out of

And btw, ->drop_inode usage in hugetlbfs is also a really bad idea,
it's duplicating large parts of fs/inode.c and is already missing
all kinds of updates.

Re: Modular filesystem using drop_inode would need inode_lock

[William Lee Irwin III]

On Fri, Jul 09, 2004 at 02:43:32PM -0700, Luck, Tony wrote:
>> This is mostly a logical inconsistency at the moment (since the
>> only filesystem that has a "drop_inode" function defined in its
>> super_operations is hugetlbfs, and it is unlikely to move out of

On Fri, Jul 09, 2004 at 11:06:44PM +0100, Christoph Hellwig wrote:
> And btw, ->drop_inode usage in hugetlbfs is also a really bad idea,
> it's duplicating large parts of fs/inode.c and is already missing
> all kinds of updates.

I suppose this reply is a bit late, but...

I'd be happy to hear of alternative methods of dealing with hugetlbfs'
operational constraints. The unusability of the core VM function
truncate_inode_pages() on hugetlb pagecache was IIRC be the primary
reason for it, but there also appears to be some inode management going
on there now that it's been disturbed by someone else.


-- wli