rwsem-rt: Make rwsem rt closer to mainline

From: "Steven Rostedt (Red Hat)" <[email protected]>

The readers of mainline rwsems are not allowed to nest, the rwsems in the PREEMPT_RT kernel should not nest either.

Signed-off-by Steven Rostedt <[email protected]>

From: "Steven Rostedt (Red Hat)" <[email protected]>

The readers of mainline rwsems are not allowed to nest, the rwsems in the PREEMPT_RT kernel should not nest either.

Quoted 4 lines · * Steven Rostedt | 2014-04-08 22:47:01 [-0400]:

From: "Steven Rostedt (Red Hat)" <[email protected]>

The readers of mainline rwsems are not allowed to nest, the rwsems in the PREEMPT_RT kernel should not nest either.

I applied this and this is the reason why cpufreq isn't working. What I see in cpufreq is:

|         test.sh-788   [004] .......    61.416288: store: down_read_try
|         test.sh-788   [004] .......    61.416296: cpufreq_cpu_get: down_read_try
|         test.sh-788   [004] .......    61.416301: cpufreq_cpu_put.part.6: up_read
|         test.sh-788   [004] .......    61.416332: store: up_read

as you see, one code path takes the read path of rw_sema twice.

Looking at the generic implementation, we have:

|#define RWSEM_UNLOCKED_VALUE            0x00000000L
|#define RWSEM_ACTIVE_BIAS               0x00000001L
|#define RWSEM_WAITING_BIAS              (-RWSEM_ACTIVE_MASK-1)

| static inline int __down_read_trylock(struct rw_semaphore *sem)
| {
|         long tmp;
| 
|         while ((tmp = sem->count) >= 0) {
|                 if (tmp == cmpxchg(&sem->count, tmp,
|                                    tmp + RWSEM_ACTIVE_READ_BIAS)) {
|                         return 1;
|                 }
|         }
|         return 0;
| }

While sem->count is >= 0 we loop and take the semaphore. So we can have five readers at once. The first writer would set count to a negative value resulting in trylock failure.

|static inline void __down_read(struct rw_semaphore *sem)
|{
|        if (unlikely(atomic_long_inc_return((atomic_long_t*)&sem->count) <= 0))
|                rwsem_down_read_failed(sem);
|}

Here the same thing but without cmpxchg(). _If_ after an increment the value is negative then we take slowpath. Otherwise we have the lock.

I think I'm going to revert this patch. Where is it written that multiple readers of a RW-semaphore can not nest? According to the code we can even have multiple readers without nesting (two+ processes may take a reader lock).

Quoted 44 lines · On Wed, 18 Feb 2015 20:57:10 +0100 Sebastian Andrzej Siewior…

Quoted 4 lines · * Steven Rostedt | 2014-04-08 22:47:01 [-0400]:

From: "Steven Rostedt (Red Hat)" <[email protected]>

The readers of mainline rwsems are not allowed to nest, the rwsems in the PREEMPT_RT kernel should not nest either.

I applied this and this is the reason why cpufreq isn't working. What I see in cpufreq is:

|         test.sh-788   [004] .......    61.416288: store: down_read_try
|         test.sh-788   [004] .......    61.416296: cpufreq_cpu_get: down_read_try
|         test.sh-788   [004] .......    61.416301: cpufreq_cpu_put.part.6: up_read
|         test.sh-788   [004] .......    61.416332: store: up_read

as you see, one code path takes the read path of rw_sema twice.

Looking at the generic implementation, we have:

|#define RWSEM_UNLOCKED_VALUE            0x00000000L
|#define RWSEM_ACTIVE_BIAS               0x00000001L
|#define RWSEM_WAITING_BIAS              (-RWSEM_ACTIVE_MASK-1)

| static inline int __down_read_trylock(struct rw_semaphore *sem)
| {
|         long tmp;
|
|         while ((tmp = sem->count) >= 0) {
|                 if (tmp == cmpxchg(&sem->count, tmp,
|                                    tmp + RWSEM_ACTIVE_READ_BIAS)) {
|                         return 1;
|                 }
|         }
|         return 0;
| }

While sem->count is >= 0 we loop and take the semaphore. So we can have five readers at once. The first writer would set count to a negative value resulting in trylock failure.

|static inline void __down_read(struct rw_semaphore *sem)
|{
|        if (unlikely(atomic_long_inc_return((atomic_long_t*)&sem->count) <= 0))
|                rwsem_down_read_failed(sem);
|}

Quoted 2 lines

Here the same thing but without cmpxchg(). _If_ after an increment the value is negative then we take slowpath. Otherwise we have the lock.

OK, so I need to make it so it can nest with trylock. I have to look at the patch again because it has been a while.

Quoted 2 lines

Here the same thing but without cmpxchg(). _If_ after an increment the value is negative then we take slowpath. Otherwise we have the lock.

OK, so I need to make it so it can nest with trylock. I have to look at the patch again because it has been a while.

An RW sem must not do two down_read()s on the same lock (it's fine for a trylock if it has a fail safe for it). The reason is, the second down_read() will block if there's a writer waiting. Thus you are guaranteed a deadlock if you have the lock for read, a write comes in and waits, and you grab the RW sem again, because it will block, and the writer is waiting for the reader to release. Thus you have a deadlock.

I'll have to revisit this. I also need to revisit the multi readers (although Thomas hates it, but he even admitted there's a better way to do it. Now only if I could remember what that was ;-)

Thanks,

-- Steve

From: "Steven Rostedt (Red Hat)" <[email protected]>

The rt_mutex_init() macro is the same whether or not CONFIG_DEBUG_RT_MUTEX is set. Remove the duplicate.

Signed-off-by Steven Rostedt <[email protected]>

Quoted 6 lines · * Steven Rostedt | 2014-04-08 22:47:02 [-0400]:

From: "Steven Rostedt (Red Hat)" <[email protected]>

The rt_mutex_init() macro is the same whether or not CONFIG_DEBUG_RT_MUTEX is set. Remove the duplicate.

Signed-off-by Steven Rostedt <[email protected]>

This does not apply cleanly on v3.14-rt

Quoted 7 lines

include/linux/rtmutex.h +0 -0

@@ -62,25 +62,19 @@ struct hrtimer_sleeper;
 # define __DEBUG_RT_MUTEX_INITIALIZER(mutexname) \
 	, .name = #mutexname, .file = __FILE__, .line = __LINE__
 

include/linux/rtmutex.h struct hrtimer_sleeper; · line 70

Quoted 6 lines

include/linux/rtmutex.h +0 -6

-# define rt_mutex_init(mutex)					\
-	do {							\
-		raw_spin_lock_init(&(mutex)->wait_lock);	\
-		__rt_mutex_init(mutex, #mutex);			\
-	} while (0)
-

The macro is the same in CONFIG_DEBUG_RT_MUTEXES and the else path. Shouldn't you remove both and define it before the ifdef?

There are some
users of this:
|drivers/i2c/i2c-core.c: rt_mutex_init(&adap->bus_lock); |drivers/media/usb/em28xx/em28xx-cards.c: rt_mutex_init(&dev->i2c_bus_lock);

Sebastian

Quoted 1 line

This does not apply cleanly on v3.14-rt

That's because I sent this out before 3.14-rt was released :-)

Quoted 6 lines

Quoted 6 lines

include/linux/rtmutex.h +0 -0

@@ -62,25 +62,19 @@ struct hrtimer_sleeper;
 # define __DEBUG_RT_MUTEX_INITIALIZER(mutexname) \
 	, .name = #mutexname, .file = __FILE__, .line = __LINE__
 

Quoted 9 lines

include/linux/rtmutex.h struct hrtimer_sleeper; · line 70

Quoted 6 lines

include/linux/rtmutex.h +0 -6

-# define rt_mutex_init(mutex)					\
-	do {							\
-		raw_spin_lock_init(&(mutex)->wait_lock);	\
-		__rt_mutex_init(mutex, #mutex);			\
-	} while (0)
-

The macro is the same in CONFIG_DEBUG_RT_MUTEXES and the else path. Shouldn't you remove both and define it before the ifdef?

That's exactly what the patch did! Except that it defined it *after* the ifdef.

Here's the patch again (as it was cut in the reply):

include/linux/rtmutex.h +3 -9

@@ -62,25 +62,19 @@ struct hrtimer_sleeper;
 # define __DEBUG_RT_MUTEX_INITIALIZER(mutexname) \
 	, .name = #mutexname, .file = __FILE__, .line = __LINE__
 
-# define rt_mutex_init(mutex)					\
-	do {							\
-		raw_spin_lock_init(&(mutex)->wait_lock);	\
-		__rt_mutex_init(mutex, #mutex);			\
-	} while (0)
-
  extern void rt_mutex_debug_task_free(struct task_struct *tsk);
 #else
 # define __DEBUG_RT_MUTEX_INITIALIZER(mutexname)
 
+# define rt_mutex_debug_task_free(t)			do { } while (0)
+#endif
+
 # define rt_mutex_init(mutex)					\
 	do {							\
 		raw_spin_lock_init(&(mutex)->wait_lock);	\
 		__rt_mutex_init(mutex, #mutex);			\
 	} while (0)
 
-# define rt_mutex_debug_task_free(t)			do { } while (0)
-#endif
-
 #define __RT_MUTEX_INITIALIZER_PLAIN(mutexname) \
 	.wait_lock = __RAW_SPIN_LOCK_UNLOCKED(mutexname.wait_lock) \
 	, .wait_list = PLIST_HEAD_INIT(mutexname.wait_list) \

Take special note of the movement of the "#endif".

-- Steve