Re: [RESEND PATCH 1/3] mmc: host: Introduce the request_atomic() for the host

From: Baolin Wang
Date: Mon Mar 16 2020 - 23:36:46 EST


On Mon, Mar 16, 2020 at 9:09 PM Adrian Hunter <adrian.hunter@xxxxxxxxx> wrote:
>
> On 4/03/20 9:42 am, Baolin Wang wrote:
> > The SD host controller can process one request in the atomic context if
> > the card is nonremovable, which means we can submit next request in the
> > irq hard handler when using the MMC software queue to reduce the latency.
> > Thus this patch adds a new API request_atomic() for the host controller
> > and implement it for the SD host controller.
> >
> > Suggested-by: Adrian Hunter <adrian.hunter@xxxxxxxxx>
> > Signed-off-by: Baolin Wang <baolin.wang7@xxxxxxxxx>
> > ---
> > drivers/mmc/host/sdhci.c | 27 +++++++++++++++++++--------
> > drivers/mmc/host/sdhci.h | 1 +
> > include/linux/mmc/host.h | 3 +++
> > 3 files changed, 23 insertions(+), 8 deletions(-)
> >
> > diff --git a/drivers/mmc/host/sdhci.c b/drivers/mmc/host/sdhci.c
> > index 9c37451..4febbcb 100644
> > --- a/drivers/mmc/host/sdhci.c
> > +++ b/drivers/mmc/host/sdhci.c
> > @@ -2016,17 +2016,12 @@ void sdhci_set_power(struct sdhci_host *host, unsigned char mode,
> > * *
> > \*****************************************************************************/
> >
> > -void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq)
> > +static void sdhci_start_request(struct mmc_host *mmc, struct mmc_request *mrq,
> > + int present)
> > {
> > - struct sdhci_host *host;
> > - int present;
> > + struct sdhci_host *host = mmc_priv(mmc);
> > unsigned long flags;
> >
> > - host = mmc_priv(mmc);
> > -
> > - /* Firstly check card presence */
> > - present = mmc->ops->get_cd(mmc);
> > -
> > spin_lock_irqsave(&host->lock, flags);
> >
> > sdhci_led_activate(host);
> > @@ -2043,6 +2038,22 @@ void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq)
> >
> > spin_unlock_irqrestore(&host->lock, flags);
> > }
> > +
> > +void sdhci_request_atomic(struct mmc_host *mmc, struct mmc_request *mrq)
> > +{
> > + sdhci_start_request(mmc, mrq, 1);
> > +}
> > +EXPORT_SYMBOL_GPL(sdhci_request_atomic);
> > +
> > +void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq)
> > +{
> > + int present;
> > +
> > + /* Firstly check card presence */
> > + present = mmc->ops->get_cd(mmc);
> > +
> > + sdhci_start_request(mmc, mrq, present);
> > +}
> > EXPORT_SYMBOL_GPL(sdhci_request);
> >
> > void sdhci_set_bus_width(struct sdhci_host *host, int width)
> > diff --git a/drivers/mmc/host/sdhci.h b/drivers/mmc/host/sdhci.h
> > index cac2d97..5507a73 100644
> > --- a/drivers/mmc/host/sdhci.h
> > +++ b/drivers/mmc/host/sdhci.h
> > @@ -775,6 +775,7 @@ void sdhci_set_power(struct sdhci_host *host, unsigned char mode,
> > void sdhci_set_power_noreg(struct sdhci_host *host, unsigned char mode,
> > unsigned short vdd);
> > void sdhci_request(struct mmc_host *mmc, struct mmc_request *mrq);
> > +void sdhci_request_atomic(struct mmc_host *mmc, struct mmc_request *mrq);
> > void sdhci_set_bus_width(struct sdhci_host *host, int width);
> > void sdhci_reset(struct sdhci_host *host, u8 mask);
> > void sdhci_set_uhs_signaling(struct sdhci_host *host, unsigned timing);
> > diff --git a/include/linux/mmc/host.h b/include/linux/mmc/host.h
> > index 562ed06..db5e59c 100644
> > --- a/include/linux/mmc/host.h
> > +++ b/include/linux/mmc/host.h
> > @@ -92,6 +92,9 @@ struct mmc_host_ops {
> > int err);
> > void (*pre_req)(struct mmc_host *host, struct mmc_request *req);
> > void (*request)(struct mmc_host *host, struct mmc_request *req);
> > + /* Submit one request to host in atomic context. */
> > + void (*request_atomic)(struct mmc_host *host,
> > + struct mmc_request *req);
>
> This doesn't have the flexibility to return "busy". For example,
> sdhci_send_command() will potentially wait quite some time if the inhibit
> bits are set. That is not good in interrupt context. It would be better to
> return immediately in that case and have the caller fall back to a
> non-atomic context. Thoughts?

Yes, I unserstood your concern. But the sdhci_send_command() is
already under the spin_lock_irqsave() protection, which will also
disable the interrupt for some time if the inhibit bits are set. That
is same with moving it in interrupt context.

Moreover, if the previous command complete interrupt and transfer
complete interrupt are normal, we should not meet this issue of
polling inhibit bits (I have not met this issue on my platform). So I
think we can remove the polling here? If the inhibit bits are set, I
think the command complete interrupt or the transfer complete
interrupt have been abnormal, so we can just return the error here.
What do you think? Thanks.

--
Baolin Wang