Re: [PATCH 05/11] mmc: core: Clarify sdio_irq_pending flag for MMC_CAP2_SDIO_IRQ_NOTHREAD

[Ulf Hansson]

On Fri, 6 Sep 2019 at 23:30, Doug Anderson <dianders@xxxxxxxxxxxx> wrote:
>
> Hi,
>
> On Fri, Sep 6, 2019 at 2:20 AM Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
> >
> > On Fri, 6 Sep 2019 at 01:47, Doug Anderson <dianders@xxxxxxxxxxxx> wrote:
> > >
> > > Hi,
> > >
> > > On Tue, Sep 3, 2019 at 7:22 AM Ulf Hansson <ulf.hansson@xxxxxxxxxx> wrote:
> > > >
> > > > In the single SDIO IRQ handler case, the sdio_irq_pending flag is used to
> > > > avoid reading the SDIO_CCCR_INTx register and instead immediately call the
> > > > SDIO func's >irq_handler() callback.
> > > >
> > > > To clarify the use behind the flag for the MMC_CAP2_SDIO_IRQ_NOTHREAD case,
> > > > let's set the flag from inside sdio_signal_irq(), rather from
> > > > sdio_run_irqs().
> > >
> > > I'm having a hard time parsing the above statement...  Can you reword
> > > and maybe I'll understand?
> >
> > Sure, I admit, it's not very good. :-) How about the below.
> >
> > The sdio_irq_pending flag is used to let host drivers indicate that it
> > has signaled an IRQ. If that is the case and we only have a single
> > SDIO func that have claimed an SDIO IRQ, our assumption is that we can
> > avoid reading the SDIO_CCCR_INTx register and just call the SDIO func
> > irq handler immediately. This makes sense, but the flag is set/cleared
> > in a somewhat messy order, let's fix that up according to below.
> >
> > First, the flag is currently set in sdio_run_irqs(), which is executed
> > as a work that was scheduled from sdio_signal_irq(). To make it more
> > implicit that the host have signaled an IRQ, let's instead immediately
> > set the flag in sdio_signal_irq(). This also makes the behavior
> > consistent with host drivers that uses the legacy,
> > mmc_signal_sdio_irq() API. This have no functional impact, because we
> > don't expect host drivers to call sdio_signal_irq() until after the
> > work (sdio_run_irqs()) have been executed anyways.
> >
> > Second, currently we never clears the flag when using the
> > sdio_run_irqs() work, but only when using the sdio_irq_thread(). Let
> > make the behavior consistent, by moving the flag to be cleared inside
> > the common process_sdio_pending_irqs() function. Additionally, tweak
> > the behavior of the flag slightly, by avoiding to clear it unless we
> > processed the SDIO IRQ. The purpose with this at this point, is to
> > keep the information about whether there have been an SDIO IRQ
> > signaled by the host, so at system resume we can decide to process it
> > without reading the SDIO_CCCR_INTx register.
> >
> > >
> > >
> > > > Moreover, let's also reset the flag when the SDIO IRQ have
> > > > been properly processed.
> > > >
> > > > Signed-off-by: Ulf Hansson <ulf.hansson@xxxxxxxxxx>
> > > > ---
> > > >  drivers/mmc/core/sdio_irq.c | 9 ++++++---
> > > >  1 file changed, 6 insertions(+), 3 deletions(-)
> > >
> > > Nice!  This looks like it addresses some of the things that came up in
> > > the previous discussion [1] and should be a nice improvement.  From
> > > re-reading that discussion that will probably change the behvaior
> > > slightly (hopefully for the better) in the single-function case where
> > > we might actually poll CCCR_INTx sometimes now.
> >
> > Correct!
> >
> > >
> > >
> > > > diff --git a/drivers/mmc/core/sdio_irq.c b/drivers/mmc/core/sdio_irq.c
> > > > index f75043266984..0962a4357d54 100644
> > > > --- a/drivers/mmc/core/sdio_irq.c
> > > > +++ b/drivers/mmc/core/sdio_irq.c
> > > > @@ -59,6 +59,7 @@ static int process_sdio_pending_irqs(struct mmc_host *host)
> > > >  {
> > > >         struct mmc_card *card = host->card;
> > > >         int i, ret, count;
> > > > +       bool sdio_irq_pending = host->sdio_irq_pending;
> > > >         unsigned char pending;
> > > >         struct sdio_func *func;
> > > >
> > > > @@ -66,13 +67,16 @@ static int process_sdio_pending_irqs(struct mmc_host *host)
> > > >         if (mmc_card_suspended(card))
> > > >                 return 0;
> > > >
> > > > +       /* Clear the flag to indicate that we have processed the IRQ. */
> > > > +       host->sdio_irq_pending = false;
> > > > +
> > > >         /*
> > > >          * Optimization, if there is only 1 function interrupt registered
> > > >          * and we know an IRQ was signaled then call irq handler directly.
> > > >          * Otherwise do the full probe.
> > > >          */
> > > >         func = card->sdio_single_irq;
> > > > -       if (func && host->sdio_irq_pending) {
> > > > +       if (func && sdio_irq_pending) {
> > > >                 func->irq_handler(func);
> > > >                 return 1;
> > > >         }
> > > > @@ -110,7 +114,6 @@ static void sdio_run_irqs(struct mmc_host *host)
> > > >  {
> > > >         mmc_claim_host(host);
> > > >         if (host->sdio_irqs) {
> > > > -               host->sdio_irq_pending = true;
> > > >                 process_sdio_pending_irqs(host);
> > > >                 if (host->ops->ack_sdio_irq)
> > > >                         host->ops->ack_sdio_irq(host);
> > > > @@ -128,6 +131,7 @@ void sdio_irq_work(struct work_struct *work)
> > > >
> > > >  void sdio_signal_irq(struct mmc_host *host)
> > > >  {
> > > > +       host->sdio_irq_pending = true;
> > >
> > > Is this safe to do without claiming the host or any other type of
> > > locking?  sdio_signal_irq() is called directly from the interrupt
> > > handler on dw_mmc with no locks held at all.  Could we have races /
> > > problems with weakly ordered memory?
> >
> > At this point, for $subject patch and @subject series, I don't see any
> > issues. But perhaps when we go forward and start using the flag
> > slightly differently.
>
> Lockless concurrency always makes my head hurt (especially when I try
> to consider weakly ordered memory) and I've learned that the only way
> I can reason about it and have any belief that I got it right is to
> always make sure I access values in a context where things are locked.
> :-P
>
> Let's see if I can figure out any actual problem, though...
>
> Certainly the queue_delayed_work() would act as a barrier so you don't
> have to worry about the worker not seeing the "= true".
>
> I suppose it's definitely possible that (if the worker is already
> running) that our "= true" will get clobbered by an "= false" from a
> previous instance of the worker running.  I guess that's unlikely
> because we can't get a second IRQ signaled until the "->ack_sdio_irq"
> ran and presumably there's enough stuff after the "= false" that one
> of them would have a barrier that made sure that the "= false" didn't
> affect us in a delayed way.
>
> So I guess we're fine...

Yeah, the trick is simply that we don't expect another IRQ being
signaled via sdio_signal_irq(), until the ->ack_sdio_irq() has been
invoked.

And even if that happens, the works case scenario would be that we
would skip the 1-func optimized pat and end up reading the poll
CCCR_INTx. This should be fine.

Kind regards
Uffe