On Sat, Oct 07, 2017 at 06:36:57AM -0400, Jeff Layton wrote:
> On Sat, 2017-10-07 at 17:55 +0800, Jia-Ju Bai wrote:
> > The kernel may sleep under a spinlock, and the function call paths are:
> > afs_do_unlk (acquire the spinlock)
> > posix_lock_file
> > posix_lock_inode (fs/locks.c)
> > locks_get_lock_context
> > kmem_cache_alloc(GFP_KERNEL) --> may sleep
> >
> > afs_do_setlk (acquire the spinlock)
> > posix_lock_file
> > posix_lock_inode (fs/locks.c)
> > locks_get_lock_context
> > kmem_cache_alloc(GFP_KERNEL) --> may sleep
> >
> > To fix them, GFP_KERNEL is replaced with GFP_ATOMIC.
> > These bugs are found by my static analysis tool and my code review.
> >
> > Signed-off-by: Jia-Ju Bai <baijiaju1990@xxxxxxx>
> > ---
> > fs/locks.c | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> >
> > diff --git a/fs/locks.c b/fs/locks.c
> > index 1bd71c4..975cc62 100644
> > --- a/fs/locks.c
> > +++ b/fs/locks.c
> > @@ -222,7 +222,7 @@ struct file_lock_list_struct {
> > if (likely(ctx) || type == F_UNLCK)
> > goto out;
> >
> > - ctx = kmem_cache_alloc(flctx_cache, GFP_KERNEL);
> > + ctx = kmem_cache_alloc(flctx_cache, GFP_ATOMIC);
> > if (!ctx)
> > goto out;
> >
>
> NAK
>
> This needs to be fixed in the AFS code. It should not be calling these
> functions with a spinlock held.
Agreed.
>From a quick look at afs_do_setlk: am I misreading something, or is it
actually trying to do an rpc call to the server while holding i_lock?
I wonder if this is the fault of the BKL conversion: 72f98e72551f
"locks: turn lock_flocks into a spinlock" claims "nothing depends on
lock_flocks using the BKL any more, so we can do the switch over to a
private spinlock." But this code, with lots of blockers, was under
lock_flocks(). Does that mean nobody's tested fcntl locking over afs
since that change in 2010?
--b.
