Re: [PATCH v2 rcu/dev 1/2] rcu/tree: Reduce wake up for synchronize_rcu() common case

[Joel Fernandes]

On 3/19/2024 1:26 PM, Uladzislau Rezki wrote:
> On Tue, Mar 19, 2024 at 12:11:28PM -0400, Joel Fernandes wrote:
>> On Tue, Mar 19, 2024 at 12:02 PM Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
>>>
>>> On Tue, Mar 19, 2024 at 03:48:46PM +0100, Uladzislau Rezki wrote:
>>>> On Tue, Mar 19, 2024 at 10:29:59AM -0400, Joel Fernandes wrote:
>>>>>
>>>>>
>>>>>> On Mar 19, 2024, at 5:53 AM, Uladzislau Rezki <urezki@xxxxxxxxx> wrote:
>>>>>>
>>>>>> On Mon, Mar 18, 2024 at 05:05:31PM -0400, Joel Fernandes wrote:
>>>>>>>
>>>>>>>
>>>>>>>>> On Mar 18, 2024, at 2:58 PM, Uladzislau Rezki <urezki@xxxxxxxxx> wrote:
>>>>>>>>
>>>>>>>> Hello, Joel!
>>>>>>>>
>>>>>>>> Sorry for late checking, see below few comments:
>>>>>>>>
>>>>>>>>> In the synchronize_rcu() common case, we will have less than
>>>>>>>>> SR_MAX_USERS_WAKE_FROM_GP number of users per GP. Waking up the kworker
>>>>>>>>> is pointless just to free the last injected wait head since at that point,
>>>>>>>>> all the users have already been awakened.
>>>>>>>>>
>>>>>>>>> Introduce a new counter to track this and prevent the wakeup in the
>>>>>>>>> common case.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
>>>>>>>>> ---
>>>>>>>>> Rebased on paul/dev of today.
>>>>>>>>>
>>>>>>>>> kernel/rcu/tree.c | 36 +++++++++++++++++++++++++++++++-----
>>>>>>>>> kernel/rcu/tree.h |  1 +
>>>>>>>>> 2 files changed, 32 insertions(+), 5 deletions(-)
>>>>>>>>>
>>>>>>>>> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
>>>>>>>>> index 9fbb5ab57c84..bd29fe3c76bf 100644
>>>>>>>>> --- a/kernel/rcu/tree.c
>>>>>>>>> +++ b/kernel/rcu/tree.c
>>>>>>>>> @@ -96,6 +96,7 @@ static struct rcu_state rcu_state = {
>>>>>>>>>   .ofl_lock = __ARCH_SPIN_LOCK_UNLOCKED,
>>>>>>>>>   .srs_cleanup_work = __WORK_INITIALIZER(rcu_state.srs_cleanup_work,
>>>>>>>>>       rcu_sr_normal_gp_cleanup_work),
>>>>>>>>> +    .srs_cleanups_pending = ATOMIC_INIT(0),
>>>>>>>>> };
>>>>>>>>>
>>>>>>>>> /* Dump rcu_node combining tree at boot to verify correct setup. */
>>>>>>>>> @@ -1642,8 +1643,11 @@ static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work)
>>>>>>>>>    * the done tail list manipulations are protected here.
>>>>>>>>>    */
>>>>>>>>>   done = smp_load_acquire(&rcu_state.srs_done_tail);
>>>>>>>>> -    if (!done)
>>>>>>>>> +    if (!done) {
>>>>>>>>> +        /* See comments below. */
>>>>>>>>> +        atomic_dec_return_release(&rcu_state.srs_cleanups_pending);
>>>>>>>>>       return;
>>>>>>>>> +    }
>>>>>>>>>
>>>>>>>>>   WARN_ON_ONCE(!rcu_sr_is_wait_head(done));
>>>>>>>>>   head = done->next;
>>>>>>>>> @@ -1666,6 +1670,9 @@ static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work)
>>>>>>>>>
>>>>>>>>>       rcu_sr_put_wait_head(rcu);
>>>>>>>>>   }
>>>>>>>>> +
>>>>>>>>> +    /* Order list manipulations with atomic access. */
>>>>>>>>> +    atomic_dec_return_release(&rcu_state.srs_cleanups_pending);
>>>>>>>>> }
>>>>>>>>>
>>>>>>>>> /*
>>>>>>>>> @@ -1673,7 +1680,7 @@ static void rcu_sr_normal_gp_cleanup_work(struct work_struct *work)
>>>>>>>>> */
>>>>>>>>> static void rcu_sr_normal_gp_cleanup(void)
>>>>>>>>> {
>>>>>>>>> -    struct llist_node *wait_tail, *next, *rcu;
>>>>>>>>> +    struct llist_node *wait_tail, *next = NULL, *rcu = NULL;
>>>>>>>>>   int done = 0;
>>>>>>>>>
>>>>>>>>>   wait_tail = rcu_state.srs_wait_tail;
>>>>>>>>> @@ -1699,16 +1706,35 @@ static void rcu_sr_normal_gp_cleanup(void)
>>>>>>>>>           break;
>>>>>>>>>   }
>>>>>>>>>
>>>>>>>>> -    // concurrent sr_normal_gp_cleanup work might observe this update.
>>>>>>>>> -    smp_store_release(&rcu_state.srs_done_tail, wait_tail);
>>>>>>>>> +    /*
>>>>>>>>> +     * Fast path, no more users to process. Remove the last wait head
>>>>>>>>> +     * if no inflight-workers. If there are in-flight workers, let them
>>>>>>>>> +     * remove the last wait head.
>>>>>>>>> +     */
>>>>>>>>> +    WARN_ON_ONCE(!rcu);
>>>>>>>>>
>>>>>>>> This assumption is not correct. An "rcu" can be NULL in fact.
>>>>>>>
>>>>>>> Hmm I could never trigger that. Are you saying that is true after Neeraj recent patch or something else?
>>>>>>> Note, after Neeraj patch to handle the lack of heads availability, it could be true so I requested
>>>>>>> him to rebase his patch on top of this one.
>>>>>>>
>>>>>>> However I will revisit my patch and look for if it could occur but please let me know if you knew of a sequence of events to make it NULL.
>>>>>>>>
>>>>>> I think we should agree on your patch first otherwise it becomes a bit
>>>>>> messy or go with a Neeraj as first step and then work on youth. So, i
>>>>>> reviewed this patch based on latest Paul's dev branch. I see that Neeraj
>>>>>> needs further work.
>>>>>
>>>>> You are right. So the only change is to drop the warning and those braces. Agreed?
>>>>>
>>>> Let me check a bit. Looks like correct but just in case.
>>>>
>>>
>>> Thanks. I was also considering improving it for the rcu == NULL case, as
>>> below. I will test it more before re-sending.
>>>
>>> On top of my patch:
>>>
>>> ---8<-----------------------
>>>
>>> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
>>> index 0df659a878ee..a5ef844835d4 100644
>>> --- a/kernel/rcu/tree.c
>>> +++ b/kernel/rcu/tree.c
>>> @@ -1706,15 +1706,18 @@ static void rcu_sr_normal_gp_cleanup(void)
>>>                         break;
>>>         }
>>>
>>> +
>>> +       /* Last head stays. No more processing to do. */
>>> +       if (!rcu)
>>> +               return;
>>> +
>>
>> Ugh, should be "if (!wait_head->next)"  instead of "if (!rcu)".  But
>> in any case, the original patch except the warning should hold.
>> Still, I am testing the above diff now.
>>
>>  - Joel
>>
> Just in case, it is based on your patch:
> 
> <snip>
> diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> index bd29fe3c76bf..98546afe7c21 100644
> --- a/kernel/rcu/tree.c
> +++ b/kernel/rcu/tree.c
> @@ -1711,29 +1711,25 @@ static void rcu_sr_normal_gp_cleanup(void)
>  	 * if no inflight-workers. If there are in-flight workers, let them
>  	 * remove the last wait head.
>  	 */
> -	WARN_ON_ONCE(!rcu);
> -	ASSERT_EXCLUSIVE_WRITER(rcu_state.srs_done_tail);
> -
> -	if (rcu && rcu_sr_is_wait_head(rcu) && rcu->next == NULL &&
> -		/* Order atomic access with list manipulation. */
> -		!atomic_read_acquire(&rcu_state.srs_cleanups_pending)) {
> +	if (wait_tail->next && rcu_sr_is_wait_head(wait_tail->next) && !wait_tail->next->next &&
> +			!atomic_read_acquire(&rcu_state.srs_cleanups_pending)) {


Yes this also works. But also if wait_tail->next == NULL, then you do not need
to queue worker for that case as well. I sent this as v3.

If you want to add that and resend my patch with the above diff, that would also
be fine. Or I can do that, let me know.  Thanks!

 - Joel