With this change a SPI controller can be added without having a IRQ
associated, and causing all transfers to be polled. For SPI controllers
without DMA, this can significantly improve performance by less
interrupt handling overhead.
Reviewed-by: Alexandre Belloni <[email protected]>
Signed-off-by: Lars Povlsen <[email protected]>
---
drivers/spi/spi-dw.c | 21 +++++++++++++--------
1 file changed, 13 insertions(+), 8 deletions(-)
diff --git a/drivers/spi/spi-dw.c b/drivers/spi/spi-dw.c
index 31e3f866d11a7..e572eb34a3c1a 100644
--- a/drivers/spi/spi-dw.c
+++ b/drivers/spi/spi-dw.c
@@ -19,6 +19,8 @@
#include <linux/debugfs.h>
#endif
+#define VALID_IRQ(i) (i >= 0)
+
/* Slave spi_dev related */
struct chip_data {
u8 tmode; /* TR/TO/RO/EEPROM */
@@ -359,7 +361,7 @@ static int dw_spi_transfer_one(struct spi_controller *master,
spi_enable_chip(dws, 1);
return ret;
}
- } else if (!chip->poll_mode) {
+ } else if (!chip->poll_mode && VALID_IRQ(dws->irq)) {
txlevel = min_t(u16, dws->fifo_len / 2, dws->len / dws->n_bytes);
dw_writel(dws, DW_SPI_TXFLTR, txlevel);
@@ -379,7 +381,7 @@ static int dw_spi_transfer_one(struct spi_controller *master,
return ret;
}
- if (chip->poll_mode)
+ if (chip->poll_mode || !VALID_IRQ(dws->irq))
return poll_transfer(dws);
return 1;
@@ -487,11 +489,13 @@ int dw_spi_add_host(struct device *dev, struct dw_spi *dws)
spi_controller_set_devdata(master, dws);
- ret = request_irq(dws->irq, dw_spi_irq, IRQF_SHARED, dev_name(dev),
- master);
- if (ret < 0) {
- dev_err(dev, "can not get IRQ\n");
- goto err_free_master;
+ if (VALID_IRQ(dws->irq)) {
+ ret = request_irq(dws->irq, dw_spi_irq, IRQF_SHARED,
+ dev_name(dev), master);
+ if (ret < 0) {
+ dev_err(dev, "can not get IRQ\n");
+ goto err_free_master;
+ }
}
master->use_gpio_descriptors = true;
@@ -539,7 +543,8 @@ int dw_spi_add_host(struct device *dev, struct dw_spi *dws)
if (dws->dma_ops && dws->dma_ops->dma_exit)
dws->dma_ops->dma_exit(dws);
spi_enable_chip(dws, 0);
- free_irq(dws->irq, master);
+ if (VALID_IRQ(dws->irq))
+ free_irq(dws->irq, master);
err_free_master:
spi_controller_put(master);
return ret;
--
2.26.2
On Wed, May 13, 2020 at 5:03 PM Lars Povlsen <[email protected]> wrote:
>
> With this change a SPI controller can be added without having a IRQ
> associated, and causing all transfers to be polled. For SPI controllers
> without DMA, this can significantly improve performance by less
> interrupt handling overhead.
...
> +#define VALID_IRQ(i) (i >= 0)
drivers/rtc/rtc-cmos.c:95:#define is_valid_irq(n) ((n) > 0)
Candidate to be in include/linux/irq.h ?
...
> + if (VALID_IRQ(dws->irq))
> + free_irq(dws->irq, master);
Isn't free_irq() aware of invalid ones (not found IRQ in the tree or
any other backend container won't do anything)?
> err_free_master:
> spi_controller_put(master);
> return ret;
> --
> 2.26.2
--
With Best Regards,
Andy Shevchenko
Hi Mark
On Wed, May 13, 2020 at 03:20:50PM +0100, Mark Brown wrote:
> On Wed, May 13, 2020 at 04:00:22PM +0200, Lars Povlsen wrote:
> > With this change a SPI controller can be added without having a IRQ
> > associated, and causing all transfers to be polled. For SPI controllers
> > without DMA, this can significantly improve performance by less
> > interrupt handling overhead.
>
> This overlaps substantially with some work that Serge Semin (CCed) has
> in progress, please coordinate with him.
Thanks for copying me these mails. I haven't been Cc'ed in the series and
hasn't been subscribed to the SPI mailing list, so I would have definitely
missed that.
I would like to coordinate my efforts with Lars. I'll have the patchset reviewed
soon in addition providing my comments/suggestions of how to make it useful for
both mine and Lars solution.
One thing I can tell about the mem_ops he implemented, is that they aren't
mem_ops, but dirmap (as you remember it's also implemented in my code, but with
alignment specific), and the exec_mem_op partly consists of a code, which belong
to the supports_op() callback. The rest of my comments will be inlined in the
patches.
-Sergey
Serge Semin writes:
> Hi Mark
>
> On Wed, May 13, 2020 at 03:20:50PM +0100, Mark Brown wrote:
>> On Wed, May 13, 2020 at 04:00:22PM +0200, Lars Povlsen wrote:
>> > With this change a SPI controller can be added without having a IRQ
>> > associated, and causing all transfers to be polled. For SPI controllers
>> > without DMA, this can significantly improve performance by less
>> > interrupt handling overhead.
>>
>> This overlaps substantially with some work that Serge Semin (CCed) has
>> in progress, please coordinate with him.
>
> Thanks for copying me these mails. I haven't been Cc'ed in the series and
> hasn't been subscribed to the SPI mailing list, so I would have definitely
> missed that.
>
> I would like to coordinate my efforts with Lars. I'll have the patchset reviewed
> soon in addition providing my comments/suggestions of how to make it useful for
> both mine and Lars solution.
Serge - thanks for taking on this.
Note that my primary concern now is to get Sparx5 upstreamed. The
mem_ops (or dirmap) and polled mode are both performance enhancements,
which can be pulled from my series if it creates too much noise. I can
then add the necessary on top of your work/current kernel at a later
time.
> One thing I can tell about the mem_ops he implemented, is that they aren't
> mem_ops, but dirmap (as you remember it's also implemented in my code, but with
> alignment specific), and the exec_mem_op partly consists of a code, which belong
> to the supports_op() callback. The rest of my comments will be inlined in the
> patches.
>
> -Sergey
--
Lars Povlsen,
Microchip
On 13/05/20 15:37, Mark Brown wrote:
> Date: Wed, 13 May 2020 15:37:53 +0100
> From: Mark Brown <[email protected]>
> To: Lars Povlsen <[email protected]>
> Cc: SoC Team <[email protected]>, Microchip Linux Driver Support
> <[email protected]>, [email protected],
> [email protected], [email protected],
> [email protected], Alexandre Belloni
> <[email protected]>
> Subject: Re: [PATCH 01/10] spi: dw: Add support for polled operation via no
> IRQ specified in DT
> User-Agent: Mutt/1.10.1 (2018-07-13)
>
> On Wed, May 13, 2020 at 04:00:22PM +0200, Lars Povlsen wrote:
>
> > +#define VALID_IRQ(i) (i >= 0)
> > +
>
> This isn't something that should be defined by an individual driver, it
> should be in a generic header.
Thanks, I will work with Serge on getting this integrated right.
---Lars
Andy Shevchenko writes:
> On Wed, May 13, 2020 at 5:03 PM Lars Povlsen <[email protected]> wrote:
>>
>> With this change a SPI controller can be added without having a IRQ
>> associated, and causing all transfers to be polled. For SPI controllers
>> without DMA, this can significantly improve performance by less
>> interrupt handling overhead.
>
> ...
>
>> +#define VALID_IRQ(i) (i >= 0)
>
> drivers/rtc/rtc-cmos.c:95:#define is_valid_irq(n) ((n) > 0)
>
> Candidate to be in include/linux/irq.h ?
>
> ...
>
>> + if (VALID_IRQ(dws->irq))
>> + free_irq(dws->irq, master);
>
> Isn't free_irq() aware of invalid ones (not found IRQ in the tree or
> any other backend container won't do anything)?
>
>
>> err_free_master:
>> spi_controller_put(master);
>> return ret;
>> --
>> 2.26.2
I'll rework this with Serge.
Thank you!
--
Lars Povlsen,
Microchip
On Wed, May 13, 2020 at 04:00:22PM +0200, Lars Povlsen wrote:
> With this change a SPI controller can be added without having a IRQ
> associated, and causing all transfers to be polled. For SPI controllers
> without DMA, this can significantly improve performance by less
> interrupt handling overhead.
>
> Reviewed-by: Alexandre Belloni <[email protected]>
> Signed-off-by: Lars Povlsen <[email protected]>
> ---
> drivers/spi/spi-dw.c | 21 +++++++++++++--------
> 1 file changed, 13 insertions(+), 8 deletions(-)
>
> diff --git a/drivers/spi/spi-dw.c b/drivers/spi/spi-dw.c
> index 31e3f866d11a7..e572eb34a3c1a 100644
> --- a/drivers/spi/spi-dw.c
> +++ b/drivers/spi/spi-dw.c
> @@ -19,6 +19,8 @@
> #include <linux/debugfs.h>
> #endif
>
> +#define VALID_IRQ(i) (i >= 0)
Mark and Andy are right. It is a good candidate to be in a generic IRQ-related
code as Anyd suggested:
> > drivers/rtc/rtc-cmos.c:95:#define is_valid_irq(n) ((n) > 0)
> > Candidate to be in include/linux/irq.h ?
So if you feel like to author additional useful patch integrated into the
kernel, this one is a good chance for it.
> +
> /* Slave spi_dev related */
> struct chip_data {
> u8 tmode; /* TR/TO/RO/EEPROM */
> @@ -359,7 +361,7 @@ static int dw_spi_transfer_one(struct spi_controller *master,
> spi_enable_chip(dws, 1);
> return ret;
> }
> - } else if (!chip->poll_mode) {
> + } else if (!chip->poll_mode && VALID_IRQ(dws->irq)) {
> txlevel = min_t(u16, dws->fifo_len / 2, dws->len / dws->n_bytes);
> dw_writel(dws, DW_SPI_TXFLTR, txlevel);
>
> @@ -379,7 +381,7 @@ static int dw_spi_transfer_one(struct spi_controller *master,
> return ret;
> }
>
> - if (chip->poll_mode)
> + if (chip->poll_mode || !VALID_IRQ(dws->irq))
> return poll_transfer(dws);
Please note. The chip->poll and the poll_transfer() methods've been discarded
from the driver, since commit 1ceb09717e98 ("spi: dw: remove cs_control and
poll_mode members from chip_data"). So you gonna have to get the
poll_transfer-like method back.
-Sergey
>
> return 1;
> @@ -487,11 +489,13 @@ int dw_spi_add_host(struct device *dev, struct dw_spi *dws)
>
> spi_controller_set_devdata(master, dws);
>
> - ret = request_irq(dws->irq, dw_spi_irq, IRQF_SHARED, dev_name(dev),
> - master);
> - if (ret < 0) {
> - dev_err(dev, "can not get IRQ\n");
> - goto err_free_master;
> + if (VALID_IRQ(dws->irq)) {
> + ret = request_irq(dws->irq, dw_spi_irq, IRQF_SHARED,
> + dev_name(dev), master);
> + if (ret < 0) {
> + dev_err(dev, "can not get IRQ\n");
> + goto err_free_master;
> + }
> }
>
> master->use_gpio_descriptors = true;
> @@ -539,7 +543,8 @@ int dw_spi_add_host(struct device *dev, struct dw_spi *dws)
> if (dws->dma_ops && dws->dma_ops->dma_exit)
> dws->dma_ops->dma_exit(dws);
> spi_enable_chip(dws, 0);
> - free_irq(dws->irq, master);
> + if (VALID_IRQ(dws->irq))
> + free_irq(dws->irq, master);
> err_free_master:
> spi_controller_put(master);
> return ret;
> --
> 2.26.2
>
> _______________________________________________
> linux-arm-kernel mailing list
> [email protected]
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel