Skip to contentSkip to conversation

Robust Futex update

PATCH Applied David Singleton 3 participants 4 messages Oct 13, 2005 – Oct 14, 2005 Futex

lore
Share & tools
lore
#0 lore raw

Ingo,
here 's a patch for 2.6.14-rc4-rt1 that fixes two things:

1) Deregister futex returns -EBUSY instead of -EINVAL if a thread tries to deregister a pthread_mutex that another thread has locked.

2) Make the fast path robust. If a pthread_mutex is only locked in user space we need to clear the owner when the thread dies. This makes both the 'rt_mutex is locked in the kernel' and the 'pthread_mutex is only locked in user space' paths robust.

David

kernel/futex.c +17 -10
Index: linux-2.6.13/kernel/futex.c
===================================================================
@@ -1354,18 +1354,25 @@ static void find_owned_futex(struct vm_a
 		if (uaddr == 0)
 			continue;
 
+		up(&mapping->robust_sem);
+		up_read(&current->mm->mmap_sem);
+		value = futex_get_user(uaddr);
+		if (this->futex_mutex.mutex_attr & FUTEX_ATTR_ROBUST)
+			value |= FUTEX_OWNER_DIED;
 		if (rt_mutex_owned_by(&this->futex_mutex, ti)) {
-			up(&mapping->robust_sem);
-			up_read(&current->mm->mmap_sem);
-			if (this->futex_mutex.mutex_attr & FUTEX_ATTR_ROBUST) {
-	 			value = futex_get_user(uaddr);
-				value |= FUTEX_OWNER_DIED;
- 				futex_put_user(value, uaddr);
-			}
+			futex_put_user(value, uaddr);
 			up_futex(&this->futex_mutex);
-			down(&mapping->robust_sem);
-			down_read(&current->mm->mmap_sem);
+		} else if ((value & FUTEX_PID) == current->pid) {
+			/*
+			 * this bit is for the fast path.  If the lock is only
+			 * locked in user space wee need to unlock it
+			 * for the exiting thread.
+			 */
+			value &= ~FUTEX_PID;
+			futex_put_user(value, uaddr);
 		}
+		down(&mapping->robust_sem);
+		down_read(&current->mm->mmap_sem);
 	}
 
 	up(&mapping->robust_sem);
@@ -1524,7 +1531,7 @@ static int futex_deregister(unsigned lon
 			if (rt_mutex_owned_by(&this->futex_mutex, ti)) {
 				up_futex(&this->futex_mutex);
 			} else if (rt_mutex_owner(&this->futex_mutex) != NULL) {
-				ret = -EINVAL;
+				ret = -EBUSY;
 				break;
 			}
 			list_del(&this->list);
include/linux/futex.h +0 -2
Index: linux-2.6.13/include/linux/futex.h
===================================================================
@@ -1,8 +1,6 @@
 #ifndef _LINUX_FUTEX_H
 #define _LINUX_FUTEX_H
 
-#include <linux/fs.h>
-
 /* Second argument to futex syscall */
 
 
kernel/rt.c +1 -0
Index: linux-2.6.13/kernel/rt.c
===================================================================
@@ -47,6 +47,7 @@
 #include <linux/syscalls.h>
 #include <linux/interrupt.h>
 #include <linux/plist.h>
+#include <linux/fs.h>
 #include <linux/futex.h>
 
 #define CAPTURE_LOCK
Applied
#1 lore raw
Quoted 10 lines · David Singleton

Ingo,
here 's a patch for 2.6.14-rc4-rt1 that fixes two things:

1) Deregister futex returns -EBUSY instead of -EINVAL if a thread tries to deregister a pthread_mutex that another thread has locked.

2) Make the fast path robust. If a pthread_mutex is only locked in user space we need to clear the owner when the thread dies. This makes both the 'rt_mutex is locked in the kernel' and the 'pthread_mutex is only locked in user space' paths robust.

thanks, applied.

Ingo

#2 lore raw
In reply to David Singleton

include/linux/futex.h line 5

Quoted 12 lines · David Singleton

Index: linux-2.6.13/include/linux/futex.h
===================================================================

include/linux/futex.h +0 -2
@@ -1,8 +1,6 @@
 #ifndef _LINUX_FUTEX_H
 #define _LINUX_FUTEX_H
 
-#include <linux/fs.h>
-
 /* Second argument to futex syscall */
 

this chunk broke the build, so i added the #include back. Really, the robust mutex glibc patches should _NOT_ automatically include the kernel's futex.h file. If they do so and rely on an installed kernel-headers package then they are broken. Just copy the file into the glibc tree and remove the #include line.

Ingo

#3 lore raw

Hi Ingo

As we pass same flags from glibc to kernel we thought it will be good to have defines in one place to avoid going out of sync in future.
glibc already has dependency on kernel headers. So were looking for ways to fix futex.h in kernel so that glibc can use it. But if it is not feasible then we can define these flags inside glibc headers. Only additional burden will be that these headers will be required to be synced whenever there is any change pertaining to them(mostly the defines.)

Thanks

Khem

Quoted 32 lines · Ingo Molnar
Quoted 6 lines · David Singleton

Index: linux-2.6.13/include/linux/futex.h
===================================================================

include/linux/futex.h line 6

Quoted 8 lines
include/linux/futex.h +0 -2
@@ -1,8 +1,6 @@
 #ifndef _LINUX_FUTEX_H
 #define _LINUX_FUTEX_H
 
-#include <linux/fs.h>
-
 /* Second argument to futex syscall */
 
 

this chunk broke the build, so i added the #include back. Really, the robust mutex glibc patches should _NOT_ automatically include the kernel's futex.h file. If they do so and rely on an installed kernel-headers package then they are broken. Just copy the file into the glibc tree and remove the #include line.

Ingo

------------------------------------------------------------------------

_______________________________________________
robustmutexes mailing list
[email protected]
https://lists.osdl.org/mailman/listinfo/robustmutexes

Signature · 3 lines
Khem Raj <[email protected]>
MontaVista Software, Inc.
www.mvista.com