[Nouveau] [PATCH] kmmio/mmiotrace: fix double free of kmmio_fault_pages

[Marcin Slusarz]

On Sat, Jun 05, 2010 at 06:49:42PM +0200, Marcin Slusarz wrote:
> After every iounmap mmiotrace has to free kmmio_fault_pages, but it
> can't do it directly, so it defers freeing by RCU.
> 
> It usually works, but when mmiotraced code calls ioremap-iounmap
> multiple times without sleeping between (so RCU won't kick in and
> start freeing) it can be given the same virtual address, so at
> every iounmap mmiotrace will schedule the same pages for release.
> Obviously it will explode on second free.
> 
> Fix it by marking kmmio_fault_pages which are scheduled for release
> and not adding them second time.
> 

Attached patch for mmiotrace testing module allows to reliably reproduce
the bug. It can be folded into the main patch.

---
diff --git a/arch/x86/mm/testmmiotrace.c b/arch/x86/mm/testmmiotrace.c
index 8565d94..5f0937b 100644
--- a/arch/x86/mm/testmmiotrace.c
+++ b/arch/x86/mm/testmmiotrace.c
@@ -90,6 +90,19 @@ static void do_test(unsigned long size)
 	iounmap(p);
 }
 
+static void do_test2(void)
+{
+	void __iomem *p;
+	int i;
+	
+	for (i = 0; i < 10; ++i) {
+		p = ioremap_nocache(mmio_address, 4096);
+		if (p)
+			iounmap(p);
+	}
+	synchronize_rcu(); /* will freeing work? */
+}
+
 static int __init init(void)
 {
 	unsigned long size = (read_far) ? (8 << 20) : (16 << 10);
@@ -104,6 +117,7 @@ static int __init init(void)
 		   "and writing 16 kB of rubbish in there.\n",
 		   size >> 10, mmio_address);
 	do_test(size);
+	do_test2();
 	pr_info("All done.\n");
 	return 0;
 }