Message-ID: <CAD=FV=U5xmz2wOnx315amtdQ5D615Pb1spiHbYfidb9nq9yxzg@mail.gmail.com>
Date:   Fri, 17 Apr 2020 10:17:21 -0700
From:   Doug Anderson <dianders@...omium.org>
To:     Maulik Shah <mkshah@...eaurora.org>
Cc:     Greg Kroah-Hartman <gregkh@...uxfoundation.org>,
        "Rafael J. Wysocki" <rafael.j.wysocki@...el.com>,
        Andy Gross <agross@...nel.org>,
        Bjorn Andersson <bjorn.andersson@...aro.org>,
        Evan Green <evgreen@...omium.org>,
        Stephen Boyd <swboyd@...omium.org>,
        Matthias Kaehlcke <mka@...omium.org>,
        linux-arm-msm <linux-arm-msm@...r.kernel.org>,
        LKML <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v2 2/2] soc: qcom: rpmh-rsc: Remove the pm_lock

Hi,

On Fri, Apr 17, 2020 at 1:07 AM Maulik Shah <mkshah@...eaurora.org> wrote:
>
> Hi,
>
> On 4/17/2020 6:02 AM, Douglas Anderson wrote:
> > It has been postulated that the pm_lock is bad for performance because
> > a CPU currently running rpmh_flush() could block other CPUs from
> > coming out of idle.  Similarly CPUs coming out of / going into idle
> > all need to contend with each other for the spinlock just to update
> > the variable tracking who's in PM.
> >
> > Let's optimize this a bit.  Specifically:
> >
> > - Use a count rather than a bitmask.  This is faster to access and
> >    also means we can use the atomic_inc_return() function to really
> >    detect who the last one to enter PM was.
> > - Accept that it's OK if we race and are doing the flush (because we
> >    think we're last) while another CPU is coming out of idle.  As long
> >    as we block that CPU if/when it tries to do an active-only transfer
> >    we're OK.
> >
> > Signed-off-by: Douglas Anderson <dianders@...omium.org>
> > ---
> >
> > Changes in v2:
> > - Always grab drv->lock first to ensure lock ordering.
> > - Grab the cache_lock in rpmh_flush().
> > - Comments about why num_online_cpus() is OK.
> > - Return NOTIFY_DONE for things we don't care about.
> > - Use trylock to avoid spinning in CPU_PM code.
> > - !rpmh_flush() should have been rpmh_flush(), so we were alwys failing.
> > - Account for CPU_PM_ENTER_FAILED not being called if we return NOTIFY_BAD.
> >
> >   drivers/soc/qcom/rpmh-internal.h | 13 +++---
> >   drivers/soc/qcom/rpmh-rsc.c      | 75 ++++++++++++++++++++++----------
> >   drivers/soc/qcom/rpmh.c          | 22 +++++++---
> >   3 files changed, 73 insertions(+), 37 deletions(-)
> >
> > diff --git a/drivers/soc/qcom/rpmh-internal.h b/drivers/soc/qcom/rpmh-internal.h
> > index dba8510c0669..f378c0f5e66d 100644
> > --- a/drivers/soc/qcom/rpmh-internal.h
> > +++ b/drivers/soc/qcom/rpmh-internal.h
> > @@ -97,7 +97,7 @@ struct rpmh_ctrlr {
> >    * @num_tcs:            Number of TCSes in this DRV.
> >    * @rsc_pm:             CPU PM notifier for controller.
> >    *                      Used when solver mode is not present.
> > - * @cpus_entered_pm:    CPU mask for cpus in idle power collapse.
> > + * @cpus_in_pm:         Number of CPUs not in idle power collapse.
> >    *                      Used when solver mode is not present.
> >    * @tcs:                TCS groups.
> >    * @tcs_in_use:         S/W state of the TCS; only set for ACTIVE_ONLY
> > @@ -109,10 +109,10 @@ struct rpmh_ctrlr {
> >    *                      (aka the drv->lock) to mark one freed.
> >    * @lock:               Synchronize state of the controller.  If you will be
> >    *                      grabbing this lock and a tcs_lock at the same time,
> > - *                      grab the tcs_lock first so we always have a
> > - *                      consistent lock ordering.
> > - * @pm_lock:            Synchronize during PM notifications.
> > - *                      Used when solver mode is not present.
> > + *                      grab the this lock first so we always have a
> > + *                      consistent lock ordering.  If RPMH's cache lock will
> > + *                      also be held, the order is: drv->lock, cache_lock,
> > + *                      tcs_lock.
> >    * @client:             Handle to the DRV's client.
> >    */
> >   struct rsc_drv {
> > @@ -121,11 +121,10 @@ struct rsc_drv {
> >       int id;
> >       int num_tcs;
> >       struct notifier_block rsc_pm;
> > -     struct cpumask cpus_entered_pm;
> > +     atomic_t cpus_in_pm;
> >       struct tcs_group tcs[TCS_TYPE_NR];
> >       DECLARE_BITMAP(tcs_in_use, MAX_TCS_NR);
> >       spinlock_t lock;
> > -     spinlock_t pm_lock;
> >       struct rpmh_ctrlr client;
> >   };
> >
> > diff --git a/drivers/soc/qcom/rpmh-rsc.c b/drivers/soc/qcom/rpmh-rsc.c
> > index a9e15699f55f..5ef4dd9c72ae 100644
> > --- a/drivers/soc/qcom/rpmh-rsc.c
> > +++ b/drivers/soc/qcom/rpmh-rsc.c
> > @@ -581,8 +581,9 @@ static int tcs_write(struct rsc_drv *drv, const struct tcs_request *msg)
> >       if (IS_ERR(tcs))
> >               return PTR_ERR(tcs);
> >
> > -     spin_lock_irqsave(&tcs->lock, flags);
> > +     local_irq_save(flags);
> >       spin_lock(&drv->lock);
> > +     spin_lock(&tcs->lock);
> >       /*
> >        * The h/w does not like if we send a request to the same address,
> >        * when one is already in-flight or being processed.
> > @@ -612,13 +613,20 @@ static int tcs_write(struct rsc_drv *drv, const struct tcs_request *msg)
> >               write_tcs_reg_sync(drv, RSC_DRV_CMD_WAIT_FOR_CMPL, tcs_id, 0);
> >               enable_tcs_irq(drv, tcs_id, true);
> >       }
> > +
> > +     /*
> > +      * Though we grabbed drv->lock first (to keep lock ordering consistent),
> > +      * we release it first.  This is fine as long as we don't try to grab
> > +      * it again until we release our tcs->lock.
> > +      */
>
> In [1]  we removed tcs->lock and are only using drv->lock , this should
> help avoid ABBA/deadlocks since we are operating on single lock.
>
> Although it needs rebase on latest tip. Let me know i can resend the
> patch/if you want to club in this change (removing tcs->lock part) i am ok.

Sure.  Looking at the places where tcs_lock() is used:

rpmh_rsc_write_ctrl_data() - only called from flushing routines which
have to have their own exclusion anyway
tcs_write() - holds drv->lock for almost the same amount of time

...so yeah, I can make that change as part of my series and it seems
sane.  Looking at it closely I think in tcs_write() I can still drop
the "drv->lock" before the __tcs_buffer_write() and
__tcs_set_trigger() calls.  That's because I already claimed the TCS
(under lock) by setting "tcs_in_use".  Once I've done that I know
nobody else could be writing to the TCS, right?


> >       spin_unlock(&drv->lock);
> >
> >       __tcs_buffer_write(drv, tcs_id, 0, msg);
> >       __tcs_set_trigger(drv, tcs_id, true);
> >
> >   done_write:
> > -     spin_unlock_irqrestore(&tcs->lock, flags);
> > +     spin_unlock(&tcs->lock);
> > +     local_irq_restore(flags);
> >       return ret;
> >   }
> >
> > @@ -741,6 +749,8 @@ int rpmh_rsc_write_ctrl_data(struct rsc_drv *drv, const struct tcs_request *msg)
> >    * SLEEP and WAKE sets. If AMCs are busy, controller can not enter
> >    * power collapse, so deny from the last cpu's pm notification.
> >    *
> > + * Context: Must be called with the drv->lock held.
> > + *
> >    * Return:
> >    * * False          - AMCs are idle
> >    * * True           - AMCs are busy
> > @@ -755,9 +765,6 @@ static bool rpmh_rsc_ctrlr_is_busy(struct rsc_drv *drv)
> >        * dedicated TCS for active state use, then re-purposed wake TCSes
> >        * should be checked for not busy, because we used wake TCSes for
> >        * active requests in this case.
> > -      *
> > -      * Since this is called from the last cpu, need not take drv or tcs
> > -      * lock before checking tcs_is_free().
> >        */
> >       if (!tcs->num_tcs)
> >               tcs = &drv->tcs[WAKE_TCS];
> > @@ -792,36 +799,57 @@ static int rpmh_rsc_cpu_pm_callback(struct notifier_block *nfb,
> >   {
> >       struct rsc_drv *drv = container_of(nfb, struct rsc_drv, rsc_pm);
> >       int ret = NOTIFY_OK;
> > -
> > -     spin_lock(&drv->pm_lock);
> > +     int cpus_in_pm;
> >
> >       switch (action) {
> >       case CPU_PM_ENTER:
> > -             cpumask_set_cpu(smp_processor_id(), &drv->cpus_entered_pm);
> > -
> > -             if (!cpumask_equal(&drv->cpus_entered_pm, cpu_online_mask))
> > -                     goto exit;
> > +             cpus_in_pm = atomic_inc_return(&drv->cpus_in_pm);
> > +             /*
> > +              * NOTE: comments for num_online_cpus() point out that it's
> > +              * only a snapshot so we need to be careful. It should be OK
> > +              * for us to use, though.  It's important for us not to miss
> > +              * if we're the last CPU going down so it would only be a
> > +              * problem if a CPU went offline right after we did the check
> > +              * AND that CPU was not idle AND that CPU was the last non-idle
> > +              * CPU. That can't happen. CPUs would have to come out of idle
> > +              * before the CPU could go offline.
> > +              */
> > +             if (cpus_in_pm < num_online_cpus())
> > +                     return NOTIFY_OK;
> >               break;
> >       case CPU_PM_ENTER_FAILED:
> >       case CPU_PM_EXIT:
> > -             cpumask_clear_cpu(smp_processor_id(), &drv->cpus_entered_pm);
> > -             goto exit;
> > +             atomic_dec(&drv->cpus_in_pm);
> > +             return NOTIFY_OK;
> > +     default:
> > +             return NOTIFY_DONE;
> >       }
> >
> > -     ret = rpmh_rsc_ctrlr_is_busy(drv);
> > -     if (ret) {
> > +     /*
> > +      * It's likely we're on the last CPU. Grab the drv->lock and write
> > +      * out the sleep/wake commands to RPMH hardware. Grabbing the lock
> > +      * means that if we race with another CPU coming up we are still
> > +      * guaranteed to be safe. If another CPU came up just after we checked
> > +      * and has grabbed thelock or started an active transfer then we'll
> thelock, the lock. one space inbetween.

Thanks.  The spacebar on my work-from-home keyboard isn't perfect.
:-P  I'll fix the typo.


> > +      * notice we're busy and abort. If another CPU comes up after we start
> > +      * flushing it will be blocked from starting an active transfer until
> > +      * we're done flushing. If another CPU starts an active transfer after
> > +      * we release the lock we're still OK because we're no longer the last
> > +      * CPU.
> > +      */
> > +     if (spin_trylock(&drv->lock)) {
> > +             if (rpmh_rsc_ctrlr_is_busy(drv) || rpmh_flush(&drv->client))
> > +                     ret = NOTIFY_BAD;
> if (rpmh_rsc_ctrlr_is_busy(drv) || rpmh_flush(&drv->client)) {
>          if (cpus_in_pm < num_online_cpus())
>              ret = NOTIFY_OK;
>          else
>              ret = NOTIFY_BAD;
>
> }
>
> if rpmh_flush() failed because another CPU woken up and started rpmh
> transfer, we are ok to power down current cpu by returning NOTIFY_OK.
>
> Can change this to return error only when last cpu failed to flush, like
> above.

I thought about doing stuff like this but decided against it.  It's
adding extra code to solve a problem that (I'll argue) isn't a
problem.  Specifically I believe that I believe it will be exceedingly
rare for _any_ of the "NOTIFY_BAD" cases to hit.  In that super rare
case we do hit one we will still be correct (we won't crash), one CPU
just won't go idle.  I would also argue that (perhaps) if it looked
like the system was idle enough for all the CPUs to go down and then
suddenly one CPU woke up then perhaps it's actually the correct thing
for this CPU not to go idle and to force the system to try again.

To answer whether adding this code would be worth it (and would help
or hurt), it would be interesting to test what's happening on real
systems.  Are you able to hit these "NOTIFY_BAD" cases?  If so, which
ones do you actually hit?  Do you _ever_ hit the case where one of the
spinlocks fails to lock?


I'll await your response here before sending v3.


-Doug

>
> > +             spin_unlock(&drv->lock);
> > +     } else {
> > +             /* Another CPU must be up */
> >               ret = NOTIFY_BAD;
> > -             goto exit;
> >       }
> >
> > -     ret = rpmh_flush(&drv->client);
> > -     if (ret)
> > -             ret = NOTIFY_BAD;
> > -     else
> > -             ret = NOTIFY_OK;
> > +     /* If we reject we won't be called w/ CPU_PM_ENTER_FAILED */
> > +     if (ret == NOTIFY_BAD)
> > +             atomic_dec(&drv->cpus_in_pm);
> >
> > -exit:
> > -     spin_unlock(&drv->pm_lock);
> >       return ret;
> >   }
> >
> > @@ -965,7 +993,6 @@ static int rpmh_rsc_probe(struct platform_device *pdev)
> >       solver_config = solver_config >> DRV_HW_SOLVER_SHIFT;
> >       if (!solver_config) {
> >               drv->rsc_pm.notifier_call = rpmh_rsc_cpu_pm_callback;
> > -             spin_lock_init(&drv->pm_lock);
> >               cpu_pm_register_notifier(&drv->rsc_pm);
> >       }
> >
> > diff --git a/drivers/soc/qcom/rpmh.c b/drivers/soc/qcom/rpmh.c
> > index 3abbb08cd6e1..3ba08bd14fa2 100644
> > --- a/drivers/soc/qcom/rpmh.c
> > +++ b/drivers/soc/qcom/rpmh.c
> > @@ -445,13 +445,21 @@ static int send_single(struct rpmh_ctrlr *ctrlr, enum rpmh_state state,
> >   int rpmh_flush(struct rpmh_ctrlr *ctrlr)
> >   {
> >       struct cache_req *p;
> > -     int ret;
> > +     int ret = 0;
> >
> >       lockdep_assert_irqs_disabled();
> >
> > +     /*
> > +      * Currently rpmh_flush() is only called when we think we're running
> > +      * on the last processor.  If the lock is busy it means another
> > +      * processor is up and it's better to abort than spin.
> > +      */
> > +     if (!spin_trylock(&ctrlr->cache_lock))
> > +             return -EBUSY;
> > +
> >       if (!ctrlr->dirty) {
> >               pr_debug("Skipping flush, TCS has latest data.\n");
> > -             return 0;
> > +             goto exit;
> >       }
> >
> >       /* Invalidate the TCSes first to avoid stale data */
> > @@ -460,7 +468,7 @@ int rpmh_flush(struct rpmh_ctrlr *ctrlr)
> >       /* First flush the cached batch requests */
> >       ret = flush_batch(ctrlr);
> >       if (ret)
> > -             return ret;
> > +             goto exit;
> >
> >       list_for_each_entry(p, &ctrlr->cache, list) {
> >               if (!is_req_valid(p)) {
> > @@ -471,16 +479,18 @@ int rpmh_flush(struct rpmh_ctrlr *ctrlr)
> >               ret = send_single(ctrlr, RPMH_SLEEP_STATE, p->addr,
> >                                 p->sleep_val);
> >               if (ret)
> > -                     return ret;
> > +                     goto exit;
> >               ret = send_single(ctrlr, RPMH_WAKE_ONLY_STATE, p->addr,
> >                                 p->wake_val);
> >               if (ret)
> > -                     return ret;
> > +                     goto exit;
> >       }
> >
> >       ctrlr->dirty = false;
> >
> > -     return 0;
> > +exit:
> > +     spin_unlock(&ctrlr->cache_lock);
> > +     return ret;
> >   }
> >
> >   /**
> [1]https://lore.kernel.org/patchwork/patch/1044697/
>
> --
> QUALCOMM INDIA, on behalf of Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, hosted by The Linux Foundation

Powered by blists - more mailing lists