Re: Deadlock on the mm->mmap_sem

[Manfred Spraul <manfred@colorfullife.com>]

[-- Attachment #1: Type: text/plain, Size: 876 bytes --]

Linus Torvalds wrote:
> 
> If the choice is between a hack to do strange and incomprehensible things
> for a special case, and just making the semaphores do the same thing
> rw-spinlocks do and make the problem go away naturally, Ill take #2 any
> day. The patches already exist, after all.
>

I've attached a recursive semaphore patch against 2.4.10-pre11 - but it
makes mmap io unusable:

Testcase:
* file, mmaped, 300 MB, 128 MB Ram
* 2 threads: touch random pages 
* third thread: calls mprotect(0xFFFF00000,0x1000, PAGE_READ)

Result:
mprotect hangs forever (minutes) with recursive semaphores.

With fair semaphores, mprotect returns after ~80 milliseconds with 5
worker threads, after ~380 milliseconds with 20 worker threads (slow IDE
disk).

One alternative to David's patch would be moving the locking into the
coredump handlers - would you prefer that?

--
	Manfred

[-- Attachment #2: patch-recursive --]
[-- Type: text/plain, Size: 835 bytes --]

--- 2.4/lib/rwsem-spinlock.c	Sat Apr 28 10:37:27 2001
+++ build-2.4/lib/rwsem-spinlock.c	Wed Sep 19 15:03:28 2001
@@ -115,7 +115,7 @@
 
 	spin_lock(&sem->wait_lock);
 
-	if (sem->activity>=0 && list_empty(&sem->wait_list)) {
+	if (sem->activity>=0) {
 		/* granted */
 		sem->activity++;
 		spin_unlock(&sem->wait_lock);
--- 2.4/arch/i386/config.in	Wed Sep 19 14:36:35 2001
+++ build-2.4/arch/i386/config.in	Wed Sep 19 14:48:06 2001
@@ -59,8 +59,8 @@
    define_bool CONFIG_X86_XADD y
    define_bool CONFIG_X86_BSWAP y
    define_bool CONFIG_X86_POPAD_OK y
-   define_bool CONFIG_RWSEM_GENERIC_SPINLOCK n
-   define_bool CONFIG_RWSEM_XCHGADD_ALGORITHM y
+   define_bool CONFIG_RWSEM_GENERIC_SPINLOCK y
+   define_bool CONFIG_RWSEM_XCHGADD_ALGORITHM n
 fi
 if [ "$CONFIG_M486" = "y" ]; then
    define_int  CONFIG_X86_L1_CACHE_SHIFT 4