Re: Poor thread performance on Linux vs. Solaris

On Tue, Sep 09, 2003 at 03:38:57PM +0800, Hu, Boris wrote:
> Try the futex_q_lock-0.2 patch. It is also against linux-2.6.0-test4. 
> 
> It does the following things:
> * Remove the global futex_lock as the previous futex_q_lock patch did.
> * Add bucket spinlock recursively check as Jakub mentioned.
> * Move vcache_lock out of lock/unlock_futex_mm() and only to protect the actual vcache operations.
> * Shrink some lock/unlock_futex_mm() scopes.

> @@ -231,26 +244,38 @@
>  	if (!page2)
>  		goto out;
>  
> -	head1 = hash_futex(page1, offset1);
> -	head2 = hash_futex(page2, offset2);
> +	unlock_futex_mm();
>  
> +	bh1 = hash_futex(page1, offset1);
> +	bh2 = hash_futex(page2, offset2);
> +	spin_lock(&bh1->lock);
> +	if (bh1 != bh2)
> +		spin_lock(&bh2->lock);
> +	head1 = &bh1->chain;
> +	head2 = &bh2->chain;
> +       

But the AB BA deadlock is still there unless I'm missing something.
I'd write:

+	bh1 = hash_futex(page1, offset1);
+	bh2 = hash_futex(page2, offset2);
+	if (bh1 < bh2) {
+		spin_lock(&bh1->lock);
+		spin_lock(&bh2->lock);
+	} else {
+		spin_lock(&bh2->lock);
+		if (bh1 > bh2)
+			spin_lock(&bh1->lock);
+	}
+	head1 = &bh1->chain;
+	head2 = &bh2->chain;

> @@ -260,7 +285,9 @@
>  	}
>  
>  out:
> -	unlock_futex_mm();
> +	if (bh1 != bh2)
> +		spin_unlock(&bh2->lock);
> +	spin_unlock(&bh1->lock);

+	if (bh1 < bh2) {
+		spin_unlock(&bh2->lock);
+		spin_unlock(&bh1->lock);
+	} else {
+		if (bh1 > bh2)
+			spin_unlock(&bh1->lock);
+		spin_unlock(&bh2->lock);
+	}

	Jakub