Skip to contentSkip to conversation

rtmutex: Add acquire semantics for rtmutex lock acquisition slow path

PATCH In review Sebastian Andrzej Siewior 3 participants 9 messages Apr 18, 2023 – Apr 22, 2023

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 rtmutex: Add acquire semantics for rtmutex lock acquisition slow path

From: Mel Gorman <[email protected]>

commit 1c0908d8e441631f5b8ba433523cf39339ee2ba0 upstream.

Jan Kara reported the following bug triggering on 6.0.5-rt14 running dbench on XFS on arm64.

kernel BUG at fs/inode.c:625!

 Internal error: Oops - BUG: 0 [#1] PREEMPT_RT SMP
 CPU: 11 PID: 6611 Comm: dbench Tainted: G            E   6.0.0-rt14-rt+ #1
 pc : clear_inode+0xa0/0xc0
 lr : clear_inode+0x38/0xc0
 Call trace:
  clear_inode+0xa0/0xc0
  evict+0x160/0x180
  iput+0x154/0x240
  do_unlinkat+0x184/0x300
  __arm64_sys_unlinkat+0x48/0xc0
  el0_svc_common.constprop.4+0xe4/0x2c0
  do_el0_svc+0xac/0x100
  el0_svc+0x78/0x200
  el0t_64_sync_handler+0x9c/0xc0
  el0t_64_sync+0x19c/0x1a0

It also affects 6.1-rc7-rt5 and affects a preempt-rt fork of 5.14 so this is likely a bug that existed forever and only became visible when ARM support was added to preempt-rt. The same problem does not occur on x86-64 and he also reported that converting sb->s_inode_wblist_lock to raw_spinlock_t makes the problem disappear indicating that the RT spinlock variant is the problem.

Which in turn means that RT mutexes on ARM64 and any other weakly ordered architecture are affected by this independent of RT.

Will Deacon observed:

"I'd be more inclined to be suspicious of the slowpath tbh, as we need to
make sure that we have acquire semantics on all paths where the lock can be taken. Looking at the rtmutex code, this really isn't obvious to me -- for example, try_to_take_rt_mutex() appears to be able to return via the 'takeit' label without acquire semantics and it looks like we might be relying on the caller's subsequent _unlock_ of the wait_lock for ordering, but that will give us release semantics which aren't correct."

Sebastian Andrzej Siewior prototyped a fix that does work based on that comment but it was a little bit overkill and added some fences that should not be necessary.

The lock owner is updated with an IRQ-safe raw spinlock held, but the spin_unlock does not provide acquire semantics which are needed when acquiring a mutex.

Adds the necessary acquire semantics for lock owner updates in the slow path acquisition and the waiter bit logic.

It successfully completed 10 iterations of the dbench workload while the vanilla kernel fails on the first iteration.

[ [email protected]: Initial prototype fix ]

Fixes 700318d1d7b38 ("locking/rtmutex: Use acquire/release semantics")
Fixes 23f78d4a03c5 ("[PATCH] pi-futex: rt mutex core")

---

Could this be please backported to 5.15 and earlier? It is already part of the 6.X kernels. I asked about this by the end of January and I'm kindly asking again ;)
This patch applies against v5.15. Should it not apply to earlier versions, please let me know an I kindly provide a backport.

I received reports that this fixes "mysterious" crashes and that is how I noticed that it is not part of the earlier kernels.

 kernel/locking/rtmutex.c     | 55 ++++++++++++++++++++++++++++++------
 kernel/locking/rtmutex_api.c |  6 ++--
 2 files changed, 49 insertions(+), 12 deletions(-)
kernel/locking/rtmutex.c +46 -9
index ea5a701ab2408..c9b21fd30bed5 100644
@@ -87,15 +87,31 @@ static inline int __ww_mutex_check_kill(struct rt_mutex *lock,
  * set this bit before looking at the lock.
  */
 
-static __always_inline void
-rt_mutex_set_owner(struct rt_mutex_base *lock, struct task_struct *owner)
+static __always_inline struct task_struct *
+rt_mutex_owner_encode(struct rt_mutex_base *lock, struct task_struct *owner)
 {
 	unsigned long val = (unsigned long)owner;
 
 	if (rt_mutex_has_waiters(lock))
 		val |= RT_MUTEX_HAS_WAITERS;
 
-	WRITE_ONCE(lock->owner, (struct task_struct *)val);
+	return (struct task_struct *)val;
+}
+
+static __always_inline void
+rt_mutex_set_owner(struct rt_mutex_base *lock, struct task_struct *owner)
+{
+	/*
+	 * lock->wait_lock is held but explicit acquire semantics are needed
+	 * for a new lock owner so WRITE_ONCE is insufficient.
+	 */
+	xchg_acquire(&lock->owner, rt_mutex_owner_encode(lock, owner));
+}
+
+static __always_inline void rt_mutex_clear_owner(struct rt_mutex_base *lock)
+{
+	/* lock->wait_lock is held so the unlock provides release semantics. */
+	WRITE_ONCE(lock->owner, rt_mutex_owner_encode(lock, NULL));
 }
 
 static __always_inline void clear_rt_mutex_waiters(struct rt_mutex_base *lock)
@@ -104,7 +120,8 @@ static __always_inline void clear_rt_mutex_waiters(struct rt_mutex_base *lock)
 			((unsigned long)lock->owner & ~RT_MUTEX_HAS_WAITERS);
 }
 
-static __always_inline void fixup_rt_mutex_waiters(struct rt_mutex_base *lock)
+static __always_inline void
+fixup_rt_mutex_waiters(struct rt_mutex_base *lock, bool acquire_lock)
 {
 	unsigned long owner, *p = (unsigned long *) &lock->owner;
 
@@ -170,8 +187,21 @@ static __always_inline void fixup_rt_mutex_waiters(struct rt_mutex_base *lock)
 	 * still set.
 	 */
 	owner = READ_ONCE(*p);
-	if (owner & RT_MUTEX_HAS_WAITERS)
-		WRITE_ONCE(*p, owner & ~RT_MUTEX_HAS_WAITERS);
+	if (owner & RT_MUTEX_HAS_WAITERS) {
+		/*
+		 * See rt_mutex_set_owner() and rt_mutex_clear_owner() on
+		 * why xchg_acquire() is used for updating owner for
+		 * locking and WRITE_ONCE() for unlocking.
+		 *
+		 * WRITE_ONCE() would work for the acquire case too, but
+		 * in case that the lock acquisition failed it might
+		 * force other lockers into the slow path unnecessarily.
+		 */
+		if (acquire_lock)
+			xchg_acquire(p, owner & ~RT_MUTEX_HAS_WAITERS);
+		else
+			WRITE_ONCE(*p, owner & ~RT_MUTEX_HAS_WAITERS);
+	}
 }
 
 /*
@@ -206,6 +236,13 @@ static __always_inline void mark_rt_mutex_waiters(struct rt_mutex_base *lock)
 		owner = *p;
 	} while (cmpxchg_relaxed(p, owner,
 				 owner | RT_MUTEX_HAS_WAITERS) != owner);
+
+	/*
+	 * The cmpxchg loop above is relaxed to avoid back-to-back ACQUIRE
+	 * operations in the event of contention. Ensure the successful
+	 * cmpxchg is visible.
+	 */
+	smp_mb__after_atomic();
 }
 
 /*
@@ -1231,7 +1268,7 @@ static int __sched __rt_mutex_slowtrylock(struct rt_mutex_base *lock)
 	 * try_to_take_rt_mutex() sets the lock waiters bit
 	 * unconditionally. Clean this up.
 	 */
-	fixup_rt_mutex_waiters(lock);
+	fixup_rt_mutex_waiters(lock, true);
 
 	return ret;
 }
@@ -1591,7 +1628,7 @@ static int __sched __rt_mutex_slowlock(struct rt_mutex_base *lock,
 	 * try_to_take_rt_mutex() sets the waiter bit
 	 * unconditionally. We might have to fix that up.
 	 */
-	fixup_rt_mutex_waiters(lock);
+	fixup_rt_mutex_waiters(lock, true);
 	return ret;
 }
 
@@ -1701,7 +1738,7 @@ static void __sched rtlock_slowlock_locked(struct rt_mutex_base *lock)
 	 * try_to_take_rt_mutex() sets the waiter bit unconditionally.
 	 * We might have to fix that up:
 	 */
-	fixup_rt_mutex_waiters(lock);
+	fixup_rt_mutex_waiters(lock, true);
 	debug_rt_mutex_free_waiter(&waiter);
 }
 
kernel/locking/rtmutex_api.c +3 -3
index 5c9299aaabae1..a461be2f873db 100644
@@ -245,7 +245,7 @@ void __sched rt_mutex_init_proxy_locked(struct rt_mutex_base *lock,
 void __sched rt_mutex_proxy_unlock(struct rt_mutex_base *lock)
 {
 	debug_rt_mutex_proxy_unlock(lock);
-	rt_mutex_set_owner(lock, NULL);
+	rt_mutex_clear_owner(lock);
 }
 
 /**
@@ -360,7 +360,7 @@ int __sched rt_mutex_wait_proxy_lock(struct rt_mutex_base *lock,
 	 * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
 	 * have to fix that up.
 	 */
-	fixup_rt_mutex_waiters(lock);
+	fixup_rt_mutex_waiters(lock, true);
 	raw_spin_unlock_irq(&lock->wait_lock);
 
 	return ret;
@@ -416,7 +416,7 @@ bool __sched rt_mutex_cleanup_proxy_lock(struct rt_mutex_base *lock,
 	 * try_to_take_rt_mutex() sets the waiter bit unconditionally. We might
 	 * have to fix that up.
 	 */
-	fixup_rt_mutex_waiters(lock);
+	fixup_rt_mutex_waiters(lock, false);
 
 	raw_spin_unlock_irq(&lock->wait_lock);
 
Signature · 1 line
2.39.1
maintainer
#1 lore raw
Quoted 68 lines · Sebastian Andrzej Siewior

From: Mel Gorman <[email protected]>

commit 1c0908d8e441631f5b8ba433523cf39339ee2ba0 upstream.

Jan Kara reported the following bug triggering on 6.0.5-rt14 running dbench on XFS on arm64.

kernel BUG at fs/inode.c:625!

 Internal error: Oops - BUG: 0 [#1] PREEMPT_RT SMP
 CPU: 11 PID: 6611 Comm: dbench Tainted: G            E   6.0.0-rt14-rt+ #1
 pc : clear_inode+0xa0/0xc0
 lr : clear_inode+0x38/0xc0
 Call trace:
  clear_inode+0xa0/0xc0
  evict+0x160/0x180
  iput+0x154/0x240
  do_unlinkat+0x184/0x300
  __arm64_sys_unlinkat+0x48/0xc0
  el0_svc_common.constprop.4+0xe4/0x2c0
  do_el0_svc+0xac/0x100
  el0_svc+0x78/0x200
  el0t_64_sync_handler+0x9c/0xc0
  el0t_64_sync+0x19c/0x1a0

It also affects 6.1-rc7-rt5 and affects a preempt-rt fork of 5.14 so this is likely a bug that existed forever and only became visible when ARM support was added to preempt-rt. The same problem does not occur on x86-64 and he also reported that converting sb->s_inode_wblist_lock to raw_spinlock_t makes the problem disappear indicating that the RT spinlock variant is the problem.

Which in turn means that RT mutexes on ARM64 and any other weakly ordered architecture are affected by this independent of RT.

Will Deacon observed:

"I'd be more inclined to be suspicious of the slowpath tbh, as we need to
make sure that we have acquire semantics on all paths where the lock can be taken. Looking at the rtmutex code, this really isn't obvious to me -- for example, try_to_take_rt_mutex() appears to be able to return via the 'takeit' label without acquire semantics and it looks like we might be relying on the caller's subsequent _unlock_ of the wait_lock for ordering, but that will give us release semantics which aren't correct."

Sebastian Andrzej Siewior prototyped a fix that does work based on that comment but it was a little bit overkill and added some fences that should not be necessary.

The lock owner is updated with an IRQ-safe raw spinlock held, but the spin_unlock does not provide acquire semantics which are needed when acquiring a mutex.

Adds the necessary acquire semantics for lock owner updates in the slow path acquisition and the waiter bit logic.

It successfully completed 10 iterations of the dbench workload while the vanilla kernel fails on the first iteration.

[ [email protected]: Initial prototype fix ]

Fixes 700318d1d7b38 ("locking/rtmutex: Use acquire/release semantics")
Fixes 23f78d4a03c5 ("[PATCH] pi-futex: rt mutex core")
Quoted 5 lines

---

Could this be please backported to 5.15 and earlier? It is already part of the 6.X kernels. I asked about this by the end of January and I'm kindly asking again ;)

I thought this was only an issues when using the out-of-tree RT patches with these kernels, right? Or is it relevant for 5.15.y from kernel.org without anything else?

Quoted 2 lines

This patch applies against v5.15. Should it not apply to earlier versions, please let me know an I kindly provide a backport.

How far back should it go?

Quoted 2 lines

I received reports that this fixes "mysterious" crashes and that is how I noticed that it is not part of the earlier kernels.

Again, isn't this only needed for -rt kernels?

thanks,

greg k-h

#2 lore raw
Quoted 7 lines · Greg KH
Quoted 3 lines

Could this be please backported to 5.15 and earlier? It is already part of the 6.X kernels. I asked about this by the end of January and I'm kindly asking again ;)

I thought this was only an issues when using the out-of-tree RT patches with these kernels, right? Or is it relevant for 5.15.y from kernel.org without anything else?

The out-of-tree RT patches make extensive use of the code. Since it is upstream code, I assumed it should go via the official stable trees. Without RT, the code is limited the rt_mutex_lock() used by I2C and the RCU booster-mutex.
Since this might not be enough to trigger the problem, let me route this via rt-stable trees.

Quoted 1 line

greg k-h

Sebastian

#3 lore raw
Quoted 5 lines · Sebastian Andrzej Siewior
Quoted 4 lines · Greg KH
Quoted 3 lines

Could this be please backported to 5.15 and earlier? It is already part of the 6.X kernels. I asked about this by the end of January and I'm kindly asking again ;)

Quoted 8 lines
Quoted 3 lines

I thought this was only an issues when using the out-of-tree RT patches with these kernels, right? Or is it relevant for 5.15.y from kernel.org without anything else?

The out-of-tree RT patches make extensive use of the code. Since it is upstream code, I assumed it should go via the official stable trees. Without RT, the code is limited the rt_mutex_lock() used by I2C and the RCU booster-mutex.

Which is a reason to route it through the upstream stable trees, no?

Thanks,

tglx

maintainer
#4 lore raw
Quoted 10 lines · Thomas Gleixner
Quoted 9 lines · Sebastian Andrzej Siewior
Quoted 7 lines · Greg KH
Quoted 3 lines

Could this be please backported to 5.15 and earlier? It is already part of the 6.X kernels. I asked about this by the end of January and I'm kindly asking again ;)

I thought this was only an issues when using the out-of-tree RT patches with these kernels, right? Or is it relevant for 5.15.y from kernel.org without anything else?

Quoted 6 lines
Quoted 4 lines

The out-of-tree RT patches make extensive use of the code. Since it is upstream code, I assumed it should go via the official stable trees. Without RT, the code is limited the rt_mutex_lock() used by I2C and the RCU booster-mutex.

Which is a reason to route it through the upstream stable trees, no?

I do not understand. Why would we take a patch in the stable tree because an out-of-tree change requires it?

confused,

greg k-h

#5 lore raw
Quoted 16 lines · Greg KH
Quoted 15 lines · Thomas Gleixner
Quoted 13 lines · Sebastian Andrzej Siewior
Quoted 7 lines · Greg KH
Quoted 3 lines

Could this be please backported to 5.15 and earlier? It is already part of the 6.X kernels. I asked about this by the end of January and I'm kindly asking again ;)

I thought this was only an issues when using the out-of-tree RT patches with these kernels, right? Or is it relevant for 5.15.y from kernel.org without anything else?

The out-of-tree RT patches make extensive use of the code. Since it is upstream code, I assumed it should go via the official stable trees. Without RT, the code is limited the rt_mutex_lock() used by I2C and the RCU booster-mutex.

Quoted 4 lines
Quoted 1 line

Which is a reason to route it through the upstream stable trees, no?

I do not understand. Why would we take a patch in the stable tree because an out-of-tree change requires it?

The change is to the rtmutex core which _IS_ used in tree by futex, RCU and some drivers.

Just because the problem was observed on RT it does not make the mainline usage of RTMUTEXes magically unaffected. The missing acquire semantics are required there too.

Thanks,

tglx

#6 lore raw
Quoted 5 lines · Thomas Gleixner
Quoted 5 lines
Quoted 5 lines
Quoted 4 lines

The out-of-tree RT patches make extensive use of the code. Since it is upstream code, I assumed it should go via the official stable trees. Without RT, the code is limited the rt_mutex_lock() used by I2C and the RCU booster-mutex.

Quoted 7 lines
Quoted 4 lines
Quoted 1 line

Which is a reason to route it through the upstream stable trees, no?

I do not understand. Why would we take a patch in the stable tree because an out-of-tree change requires it?

The change is to the rtmutex core which _IS_ used in tree by futex, RCU and some drivers.

not back stab but to clarify: futex does not use the annotation (it does not use the fastpath) but RCU-boosting _and_ I2C-bus code does use it.

So both mainline users users of the rtmutex code are affected.

Quoted 7 lines

Just because the problem was observed on RT it does not make the mainline usage of RTMUTEXes magically unaffected. The missing acquire semantics are required there too.

Thanks,

tglx

Sebastian

#7 lore raw
Quoted 8 lines · Sebastian Andrzej Siewior
Quoted 7 lines · Thomas Gleixner
Quoted 7 lines
Quoted 6 lines
Quoted 4 lines

The out-of-tree RT patches make extensive use of the code. Since it is upstream code, I assumed it should go via the official stable trees. Without RT, the code is limited the rt_mutex_lock() used by I2C and the RCU booster-mutex.

Which is a reason to route it through the upstream stable trees, no?

Quoted 8 lines
Quoted 5 lines
Quoted 2 lines

I do not understand. Why would we take a patch in the stable tree because an out-of-tree change requires it?

The change is to the rtmutex core which _IS_ used in tree by futex, RCU and some drivers.

not back stab but to clarify: futex does not use the annotation (it does not use the fastpath) but RCU-boosting _and_ I2C-bus code does use it.

Futex requires it too, really. The patch is about the slowpath, no?

maintainer
#8 lore raw
Quoted 12 lines · Thomas Gleixner
Quoted 11 lines · Sebastian Andrzej Siewior
Quoted 10 lines · Thomas Gleixner
Quoted 9 lines
Quoted 6 lines
Quoted 4 lines

The out-of-tree RT patches make extensive use of the code. Since it is upstream code, I assumed it should go via the official stable trees. Without RT, the code is limited the rt_mutex_lock() used by I2C and the RCU booster-mutex.

Which is a reason to route it through the upstream stable trees, no?

I do not understand. Why would we take a patch in the stable tree because an out-of-tree change requires it?

Quoted 9 lines
Quoted 5 lines
Quoted 2 lines

The change is to the rtmutex core which _IS_ used in tree by futex, RCU and some drivers.

not back stab but to clarify: futex does not use the annotation (it does not use the fastpath) but RCU-boosting _and_ I2C-bus code does use it.

Futex requires it too, really. The patch is about the slowpath, no?

Ok, sorry, I was confused, now queued up.

greg k-h