Re: [PATCH] locking/rwsem: Prevent non-first waiter from spinning in down_write() slowpath

[Hillf Danton <hdanton@sina.com>]

On 11 Oct 2022 18:46:20 +0530 Mukesh Ojha <quic_mojha@quicinc.com>
>On 10/11/2022 4:16 PM, Hillf Danton wrote:
>> On 10/10/22 06:24 Mukesh Ojha <quic_mojha@quicinc.com>
>>> Hi Waiman,
>>>
>>> On 9/29/2022 11:36 PM, Waiman Long wrote:
>>>> On 9/29/22 14:04, Waiman Long wrote:
>>>>> A non-first waiter can potentially spin in the for loop of
>>>>> rwsem_down_write_slowpath() without sleeping but fail to acquire the
>>>>> lock even if the rwsem is free if the following sequence happens:
>>>>>
>>>>>     Non-first waiter       First waiter      Lock holder
>>>>>     ----------------       ------------      -----------
>>>>>     Acquire wait_lock
>>>>>     rwsem_try_write_lock():
>>>>>       Set handoff bit if RT or
>>>>>         wait too long
>>>>>       Set waiter->handoff_set
>>>>>     Release wait_lock
>>>>>                            Acquire wait_lock
>>>>>                            Inherit waiter->handoff_set
>>>>>                            Release wait_lock
>>>>>                         Clear owner
>>>>>                                              Release lock
>>>>>     if (waiter.handoff_set) {
>>>>>       rwsem_spin_on_owner(();
>>>>>       if (OWNER_NULL)
>>>>>         goto trylock_again;
>>>>>     }
>>>>>     trylock_again:
>>>>>     Acquire wait_lock
>>>>>     rwsem_try_write_lock():
>>>>>        if (first->handoff_set && (waiter != first))
>>>>>            return false;
>>>>>     Release wait_lock
>>>>>
>>>>> It is especially problematic if the non-first waiter is an RT task and
>>>>> it is running on the same CPU as the first waiter as this can lead to
>>>>> live lock.
>>>>>
>>>>> Fixes: d257cc8cb8d5 ("locking/rwsem: Make handoff bit handling more
>>>>> consistent")
>>>>> Signed-off-by: Waiman Long <longman@redhat.com>
>>>>> ---
>>>>>    kernel/locking/rwsem.c | 13 ++++++++++---
>>>>>    1 file changed, 10 insertions(+), 3 deletions(-)
>>>>
>>>> Mukesh, can you test if this patch can fix the RT task lockup problem?
>>>>
>>>
>>> Looks like, There is still a window for a race.
>>>
>>> There is a chance when a reader who came first added it's BIAS and
>>> goes to slowpath and before it gets added to wait list it got
>>> preempted by RT task which  goes to slowpath as well and being the
>>> first waiter gets its hand-off bit set and not able to get the lock
>>> due to following condition in rwsem_try_write_lock()
>
>[]
>
>>>
>>>   630                 if (count & RWSEM_LOCK_MASK) {  ==> reader has
>>> sets its bias
>>> ..
>>> ...
>>>
>>>   634
>>>   635                         new |= RWSEM_FLAG_HANDOFF;
>>>   636                 } else {
>>>   637                         new |= RWSEM_WRITER_LOCKED;
>>>
>>>
>>> ---------------------->----------------------->-------------------------
>>>
>>> First reader (1)          writer(2) RT task             Lock holder(3)
>>>
>>> It sets
>>> RWSEM_READER_BIAS.
>>> while it is going to
>>> slowpath(as the lock
>>> was held by (3)) and
>>> before it got added
>>> to the waiters list
>>> it got preempted
>>> by (2).
>>>                         RT task also takes
>>>                          the slowpath and add              release the
>>>                          itself into waiting list          rwsem lock
>>>              and since it is the first         clear the
>>>                          it is the next one to get         owner.
>>>                          the lock but it can not
>>>                          get the lock as (count &
>>>                          RWSEM_LOCK_MASK) is set
>>>                          as (1) has added it but
>>>                          not able to remove its
>>>              adjustment.
>
>[]
>
>>>
>> Hey Mukesh,
>> 
>> Can you test the diff if it makes sense to you?
>> 
>> It simply prevents the first waiter from spinning any longer after detecting
>> it barely makes any progress to spin without lock owner.
>> 
>> Hillf
>> 
>> --- mainline/kernel/locking/rwsem.c
>> +++ b/kernel/locking/rwsem.c
>> @@ -611,26 +611,15 @@ static inline bool rwsem_try_write_lock(
>>   	long count, new;
>>   
>>   	lockdep_assert_held(&sem->wait_lock);
>> +	waiter->handoff_set = false;
>>   
>>   	count = atomic_long_read(&sem->count);
>>   	do {
>>   		bool has_handoff = !!(count & RWSEM_FLAG_HANDOFF);
>>   
>>   		if (has_handoff) {
>> -			/*
>> -			 * Honor handoff bit and yield only when the first
>> -			 * waiter is the one that set it. Otherwisee, we
>> -			 * still try to acquire the rwsem.
>> -			 */
>> -			if (first->handoff_set && (waiter != first))
>> +			if (waiter != first)
>>   				return false;
>
> you mean, you want to check and change waiter->handoff_set on every run 
> rwsem_try_write_lock().
>
Yes, with RWSEM_FLAG_HANDOFF set, it is too late for non first waiters to
spin, and with both RWSEM_LOCK_MASK and RWSEM_FLAG_HANDOFF set, the rivals
in the RWSEM_LOCK_MASK have an uphand over the first waiter wrt acquiring
the lock, and it is not a bad option for the first waiter to take a step
back off.

		if (count & RWSEM_LOCK_MASK) {
			if (has_handoff || (!rt_task(waiter->task) &&
					    !time_after(jiffies, waiter->timeout)))
				return false;

			new |= RWSEM_FLAG_HANDOFF;
		} else {

> But does it break optimistic spinning ? @waiman ?

Waiters spin for acquiring lock instead of lockup and your report shows
spinning too much makes trouble. The key is stop spinning neither too
late nor too early. My proposal is a simple one with as few heuristics
added as possible.

Hillf