Re: bug in tty ldisc and friends

[Joerg Roedel <joerg.roedel@amd.com>]

On Tue, Jun 16, 2009 at 12:36:06PM +0200, Ingo Molnar wrote:
> 
> * Ingo Molnar <mingo@elte.hu> wrote:
> 
> > This is something we noticed recently: dma-debug uses a lot of 
> > lock classes and thus creates a really large lock-graph, depleting 
> > the reserves quickly.
> 
> yep, i think dma-debug is the culprit. It has:
> 
>  #define HASH_SIZE       1024ULL
> 
> each hash entry has a separate lock.
> 
> But this should be easily to solve: a special lock-class for these 
> locks. Patch below (warning: utterly untested!).
> 
> (Btw., that SPIN_LOCK_UNLOCKED init method should really be 
> deprecated for real.)
> 
> 	Ingo
> 
> diff --git a/lib/dma-debug.c b/lib/dma-debug.c
> index 3b93129..012c240 100644
> --- a/lib/dma-debug.c
> +++ b/lib/dma-debug.c
> @@ -62,6 +62,8 @@ struct dma_debug_entry {
>  #endif
>  };
>  
> +static struct lock_class_key hash_bucket_class;
> +
>  struct hash_bucket {
>  	struct list_head list;
>  	spinlock_t lock;
> @@ -716,7 +718,8 @@ void dma_debug_init(u32 num_entries)
>  
>  	for (i = 0; i < HASH_SIZE; ++i) {
>  		INIT_LIST_HEAD(&dma_entry_hash[i].list);
> -		dma_entry_hash[i].lock = SPIN_LOCK_UNLOCKED;
> +		spin_lock_init(&dma_entry_hash[i].lock);
> +		lockdep_set_lock_class(&dma_entry_hash[i].lock, &hash_bucket_class);
                ^^^^^^^^^^^^^^^^^^^^^^
Ok, this function is called lockdep_set_class. Otherwise the test went
fine. I queued the patch.

	Joerg

-- 
           | Advanced Micro Devices GmbH
 Operating | Karl-Hammerschmidt-Str. 34, 85609 Dornach bei München
 System    | 
 Research  | Geschäftsführer: Thomas M. McCoy, Giuliano Meroni
 Center    | Sitz: Dornach, Gemeinde Aschheim, Landkreis München
           | Registergericht München, HRB Nr. 43632