Re: [PATCH v4 14/16] locking/rwsem: Guard against making count negative

[Waiman Long <llong@redhat.com>]

On 4/23/19 10:17 AM, Peter Zijlstra wrote:
> On Sun, Apr 21, 2019 at 05:07:56PM -0400, Waiman Long wrote:
>
>> How about the following chunks to disable preemption temporarily for the
>> increment-check-decrement sequence?
>>
>> diff --git a/include/linux/preempt.h b/include/linux/preempt.h
>> index dd92b1a93919..4cc03ac66e13 100644
>> --- a/include/linux/preempt.h
>> +++ b/include/linux/preempt.h
>> @@ -250,6 +250,8 @@ do { \
>>  #define preempt_enable_notrace()               barrier()
>>  #define preemptible()                          0
>>  
>> +#define __preempt_disable_nop  /* preempt_disable() is nop */
>> +
>>  #endif /* CONFIG_PREEMPT_COUNT */
>>  
>>  #ifdef MODULE
>> diff --git a/kernel/locking/rwsem.c b/kernel/locking/rwsem.c
>> index 043fd29b7534..54029e6af17b 100644
>> --- a/kernel/locking/rwsem.c
>> +++ b/kernel/locking/rwsem.c
>> @@ -256,11 +256,64 @@ static inline struct task_struct
>> *rwsem_get_owner(struct r
>>         return (struct task_struct *) (cowner
>>                 ? cowner | (sowner & RWSEM_NONSPINNABLE) : sowner);
>>  }
>> +
>> +/*
>> + * If __preempt_disable_nop is defined, calling preempt_disable() and
>> + * preempt_enable() directly is the most efficient way. Otherwise, it may
>> + * be more efficient to disable and enable interrupt instead for disabling
>> + * preemption tempoarily.
>> + */
>> +#ifdef __preempt_disable_nop
>> +#define disable_preemption()   preempt_disable()
>> +#define enable_preemption()    preempt_enable()
>> +#else
>> +#define disable_preemption()   local_irq_disable()
>> +#define enable_preemption()    local_irq_enable()
>> +#endif
> I'm not aware of an architecture where disabling interrupts is faster
> than disabling preemption.

I have actually done some performance test measuring the effects of
disabling interrupt and preemption on readers (on x86-64 system).

  Threads    Before patch    Disable irq    Disable preemption
  -------    ------------    -----------    ------------------
     1          9,088          8,766           9,172
     2          9,296          9,169           8,707
     4         11,192         11,205          10,712
     8         11,329         11,332          11,213

For uncontended case, disable interrupt is slower. The slowdown is gone
once the rwsem becomes contended. So it may not be a good idea to
disable interrupt as a proxy of disabling preemption.

BTW, preemption count is not enabled in typical distro production
kernels like RHEL. So preempt_disable() is just a barrier. It is turned
on in the debug kernel, though.


>> +/*
>> + * When the owner task structure pointer is merged into couunt, less bits
>> + * will be available for readers. Therefore, there is a very slight chance
>> + * that the reader count may overflow. We try to prevent that from
>> happening
>> + * by checking for the MS bit of the count and failing the trylock attempt
>> + * if this bit is set.
>> + *
>> + * With preemption enabled, there is a remote possibility that preemption
>> + * can happen in the narrow timing window between incrementing and
>> + * decrementing the reader count and the task is put to sleep for a
>> + * considerable amount of time. If sufficient number of such unfortunate
>> + * sequence of events happen, we may still overflow the reader count.
>> + * To avoid such possibility, we have to disable preemption for the
>> + * whole increment-check-decrement sequence.
>> + *
>> + * The function returns true if there are too many readers and the count
>> + * has already been properly decremented so the reader must go directly
>> + * into the wait list.
>> + */
>> +static inline bool rwsem_read_trylock(struct rw_semaphore *sem, long *cnt)
>> +{
>> +       bool wait = false;      /* Wait now flag */
>> +
>> +       disable_preemption();
>> +       *cnt = atomic_long_fetch_add_acquire(RWSEM_READER_BIAS,
>> &sem->count);
>> +       if (unlikely(*cnt < 0)) {
>> +               atomic_long_add(-RWSEM_READER_BIAS, &sem->count);
>> +               wait = true;
>> +       }
>> +       enable_preemption();
>> +       return wait;
>> +}
>>  #else /* !CONFIG_RWSEM_OWNER_COUNT */
> This also means you have to ensure CONFIG_NR_CPUS < 32K for
> RWSEM_OWNER_COUNT.


Yes, that can be done.


>
>>  static inline struct task_struct *rwsem_get_owner(struct rw_semaphore *sem)
>>  {
>>         return READ_ONCE(sem->owner);
>>  }
>> +
>> +static inline bool rwsem_read_trylock(struct rw_semaphore *sem, long *cnt)
>> +{
>> +       *cnt = atomic_long_fetch_add_acquire(RWSEM_READER_BIAS,
>> &sem->count);
>> +       return false;
>> +}
>>  #endif /* CONFIG_RWSEM_OWNER_COUNT */
>>  
>>  /*
>> @@ -981,32 +1034,18 @@ static inline void clear_wr_nonspinnable(struct
>> rw_semaph
>>   * Wait for the read lock to be granted
>>   */
>>  static struct rw_semaphore __sched *
>> -rwsem_down_read_slowpath(struct rw_semaphore *sem, int state, long count)
>> +rwsem_down_read_slowpath(struct rw_semaphore *sem, int state, const
>> bool wait)
>>  {
>> -       long adjustment = -RWSEM_READER_BIAS;
>> +       long count, adjustment = -RWSEM_READER_BIAS;
>>         bool wake = false;
>>         struct rwsem_waiter waiter;
>>         DEFINE_WAKE_Q(wake_q);
>>  
>> -       if (unlikely(count < 0)) {
>> +       if (unlikely(wait)) {
>>                 /*
>> -                * The sign bit has been set meaning that too many active
>> -                * readers are present. We need to decrement reader count &
>> -                * enter wait queue immediately to avoid overflowing the
>> -                * reader count.
>> -                *
>> -                * As preemption is not disabled, there is a remote
>> -                * possibility that preemption can happen in the narrow
>> -                * timing window between incrementing and decrementing
>> -                * the reader count and the task is put to sleep for a
>> -                * considerable amount of time. If sufficient number
>> -                * of such unfortunate sequence of events happen, we
>> -                * may still overflow the reader count. It is extremely
>> -                * unlikey, though. If this is a concern, we should consider
>> -                * disable preemption during this timing window to make
>> -                * sure that such unfortunate event will not happen.
>> +                * The reader count has already been decremented and the
>> +                * reader should go directly into the wait list now.
>>                  */
>> -               atomic_long_add(-RWSEM_READER_BIAS, &sem->count);
>>                 adjustment = 0;
>>                 goto queue;
>>         }
>> @@ -1358,11 +1397,12 @@ static struct rw_semaphore
>> *rwsem_downgrade_wake(struct
>>   */
>>  inline void __down_read(struct rw_semaphore *sem)
>>  {
>> -       long tmp = atomic_long_fetch_add_acquire(RWSEM_READER_BIAS,
>> -                                                &sem->count);
>> +       long tmp;
>> +       bool wait;
>>  
>> +       wait = rwsem_read_trylock(sem, &tmp);
>>         if (unlikely(tmp & RWSEM_READ_FAILED_MASK)) {
>> -               rwsem_down_read_slowpath(sem, TASK_UNINTERRUPTIBLE, tmp);
>> +               rwsem_down_read_slowpath(sem, TASK_UNINTERRUPTIBLE, wait);
>>                 DEBUG_RWSEMS_WARN_ON(!is_rwsem_reader_owned(sem), sem);
>>         } else {
>>                 rwsem_set_reader_owned(sem);
> I think I prefer that function returning/taking the bias/adjustment
> value instead of a bool, if it is all the same.

Sure, I can do that.

Cheers,
Longman