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

[Uladzislau Rezki]

On Tue, Mar 19, 2024 at 01:33:11PM -0400, Joel Fernandes wrote:
> 
> 
> On 3/19/2024 1:29 PM, Joel Fernandes wrote:
> > 
> > 
> > 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.
> > 
> Sorry, I see you did add that later in the patch ;-). I think we have converged
> on the final patch then, give or take the use of 'rcu' versus 'wait_tail->next'.
> 
Just combine all parts into one place and resend :)

Thanks!

--
Uladzislau Rezki