In DMA mode we have a maximum transfer size, past that the driver falls back to PIO (see the check at the top of pxa2xx_spi_transfer_one). Falling back to PIO for big transfers defeats the point of a dma engine, hence set the max transfer size to inform spi clients that they need to do something smarter.
This was uncovered by the drm_mipi_dbi spi panel code, which does large spi transfers, but stopped splitting them after:
commit e143364b4c1774f68e923a5a0bb0fca28ac25888 Author: Noralf Trønnes noralf@tronnes.org Date: Fri Jul 19 17:59:10 2019 +0200
drm/tinydrm: Remove tinydrm_spi_max_transfer_size()
After this commit the code relied on the spi core to split transfers into max dma-able blocks, which also papered over the PIO fallback issue.
Fix this by setting the overall max transfer size to the DMA limit, but only when the controller runs in DMA mode.
Fixes: e143364b4c17 ("drm/tinydrm: Remove tinydrm_spi_max_transfer_size()") Cc: Sam Ravnborg sam@ravnborg.org Cc: Noralf Trønnes noralf@tronnes.org Cc: Andy Shevchenko andriy.shevchenko@intel.com Reported-and-tested-by: Andy Shevchenko andriy.shevchenko@intel.com Cc: Daniel Mack daniel@zonque.org Cc: Haojian Zhuang haojian.zhuang@gmail.com Cc: Robert Jarzmik robert.jarzmik@free.fr Cc: Mark Brown broonie@kernel.org Cc: linux-arm-kernel@lists.infradead.org Cc: linux-spi@vger.kernel.org Signed-off-by: Daniel Vetter daniel.vetter@ffwll.ch --- drivers/spi/spi-pxa2xx.c | 1 + 1 file changed, 1 insertion(+)
diff --git a/drivers/spi/spi-pxa2xx.c b/drivers/spi/spi-pxa2xx.c index bb6a14d1ab0f..f77201915033 100644 --- a/drivers/spi/spi-pxa2xx.c +++ b/drivers/spi/spi-pxa2xx.c @@ -1707,6 +1707,7 @@ static int pxa2xx_spi_probe(struct platform_device *pdev) } else { controller->can_dma = pxa2xx_spi_can_dma; controller->max_dma_len = MAX_DMA_LEN; + controller->max_transfer_size = MAX_DMA_LEN; } }
Den 16.10.2019 21.57, skrev Daniel Vetter:
In DMA mode we have a maximum transfer size, past that the driver falls back to PIO (see the check at the top of pxa2xx_spi_transfer_one). Falling back to PIO for big transfers defeats the point of a dma engine, hence set the max transfer size to inform spi clients that they need to do something smarter.
This was uncovered by the drm_mipi_dbi spi panel code, which does large spi transfers, but stopped splitting them after:
commit e143364b4c1774f68e923a5a0bb0fca28ac25888 Author: Noralf Trønnes noralf@tronnes.org Date: Fri Jul 19 17:59:10 2019 +0200
drm/tinydrm: Remove tinydrm_spi_max_transfer_size()
After this commit the code relied on the spi core to split transfers into max dma-able blocks, which also papered over the PIO fallback issue.
Fix this by setting the overall max transfer size to the DMA limit, but only when the controller runs in DMA mode.
Fixes: e143364b4c17 ("drm/tinydrm: Remove tinydrm_spi_max_transfer_size()") Cc: Sam Ravnborg sam@ravnborg.org Cc: Noralf Trønnes noralf@tronnes.org Cc: Andy Shevchenko andriy.shevchenko@intel.com Reported-and-tested-by: Andy Shevchenko andriy.shevchenko@intel.com Cc: Daniel Mack daniel@zonque.org Cc: Haojian Zhuang haojian.zhuang@gmail.com Cc: Robert Jarzmik robert.jarzmik@free.fr Cc: Mark Brown broonie@kernel.org Cc: linux-arm-kernel@lists.infradead.org Cc: linux-spi@vger.kernel.org Signed-off-by: Daniel Vetter daniel.vetter@ffwll.ch
drivers/spi/spi-pxa2xx.c | 1 + 1 file changed, 1 insertion(+)
diff --git a/drivers/spi/spi-pxa2xx.c b/drivers/spi/spi-pxa2xx.c index bb6a14d1ab0f..f77201915033 100644 --- a/drivers/spi/spi-pxa2xx.c +++ b/drivers/spi/spi-pxa2xx.c @@ -1707,6 +1707,7 @@ static int pxa2xx_spi_probe(struct platform_device *pdev) } else { controller->can_dma = pxa2xx_spi_can_dma; controller->max_dma_len = MAX_DMA_LEN;
controller->max_transfer_size = MAX_DMA_LEN;
As Andy mentioned, ->max_transfer_size is a callback:
struct spi_controller { /* * on some hardware transfer / message size may be constrained * the limit may depend on device transfer settings */ size_t (*max_transfer_size)(struct spi_device *spi);
(I looked at its users and they all return a static or probe time value so not sure why it's a callback.)
Noralf.
}
}
On Wed, Oct 16, 2019 at 11:30:51PM +0200, Noralf Trønnes wrote:
As Andy mentioned, ->max_transfer_size is a callback:
struct spi_controller { /* * on some hardware transfer / message size may be constrained * the limit may depend on device transfer settings */ size_t (*max_transfer_size)(struct spi_device *spi);
(I looked at its users and they all return a static or probe time value so not sure why it's a callback.)
Huh, I wonder why we did that... I can't find any clear record, and like you say nothing is using it so we should probably fix it - however that's best done separately to this :/ Sorry for the confusion, I didn't register what Andy was saying as being pointing out the current state of the code.
In DMA mode we have a maximum transfer size, past that the driver falls back to PIO (see the check at the top of pxa2xx_spi_transfer_one). Falling back to PIO for big transfers defeats the point of a dma engine, hence set the max transfer size to inform spi clients that they need to do something smarter.
This was uncovered by the drm_mipi_dbi spi panel code, which does large spi transfers, but stopped splitting them after:
commit e143364b4c1774f68e923a5a0bb0fca28ac25888 Author: Noralf Trønnes noralf@tronnes.org Date: Fri Jul 19 17:59:10 2019 +0200
drm/tinydrm: Remove tinydrm_spi_max_transfer_size()
After this commit the code relied on the spi core to split transfers into max dma-able blocks, which also papered over the PIO fallback issue.
Fix this by setting the overall max transfer size to the DMA limit, but only when the controller runs in DMA mode.
Fixes: e143364b4c17 ("drm/tinydrm: Remove tinydrm_spi_max_transfer_size()") Cc: Sam Ravnborg sam@ravnborg.org Cc: Noralf Trønnes noralf@tronnes.org Cc: Andy Shevchenko andriy.shevchenko@intel.com Reported-and-tested-by: Andy Shevchenko andriy.shevchenko@intel.com Cc: Daniel Mack daniel@zonque.org Cc: Haojian Zhuang haojian.zhuang@gmail.com Cc: Robert Jarzmik robert.jarzmik@free.fr Cc: Mark Brown broonie@kernel.org Cc: linux-arm-kernel@lists.infradead.org Cc: linux-spi@vger.kernel.org Signed-off-by: Daniel Vetter daniel.vetter@ffwll.ch -- v2: Brown paper bag edition, it needs to be a function ... --- drivers/spi/spi-pxa2xx.c | 7 +++++++ 1 file changed, 7 insertions(+)
diff --git a/drivers/spi/spi-pxa2xx.c b/drivers/spi/spi-pxa2xx.c index bb6a14d1ab0f..068c21037679 100644 --- a/drivers/spi/spi-pxa2xx.c +++ b/drivers/spi/spi-pxa2xx.c @@ -1602,6 +1602,11 @@ static int pxa2xx_spi_fw_translate_cs(struct spi_controller *controller, return cs; }
+static size_t pxa2xx_spi_max_dma_transfer_size(struct spi_device *spi) +{ + return MAX_DMA_LEN; +} + static int pxa2xx_spi_probe(struct platform_device *pdev) { struct device *dev = &pdev->dev; @@ -1707,6 +1712,8 @@ static int pxa2xx_spi_probe(struct platform_device *pdev) } else { controller->can_dma = pxa2xx_spi_can_dma; controller->max_dma_len = MAX_DMA_LEN; + controller->max_transfer_size = + pxa2xx_spi_max_dma_transfer_size; } }
On Thu, Oct 17, 2019 at 08:44:26AM +0200, Daniel Vetter wrote:
In DMA mode we have a maximum transfer size, past that the driver falls back to PIO (see the check at the top of pxa2xx_spi_transfer_one). Falling back to PIO for big transfers defeats the point of a dma engine, hence set the max transfer size to inform spi clients that they need to do something smarter.
This was uncovered by the drm_mipi_dbi spi panel code, which does large spi transfers, but stopped splitting them after:
commit e143364b4c1774f68e923a5a0bb0fca28ac25888 Author: Noralf Trønnes noralf@tronnes.org Date: Fri Jul 19 17:59:10 2019 +0200
drm/tinydrm: Remove tinydrm_spi_max_transfer_size()
After this commit the code relied on the spi core to split transfers into max dma-able blocks, which also papered over the PIO fallback issue.
Fix this by setting the overall max transfer size to the DMA limit, but only when the controller runs in DMA mode.
Thank you, Daniel!
Fixes: e143364b4c17 ("drm/tinydrm: Remove tinydrm_spi_max_transfer_size()") Cc: Sam Ravnborg sam@ravnborg.org Cc: Noralf Trønnes noralf@tronnes.org Cc: Andy Shevchenko andriy.shevchenko@intel.com Reported-and-tested-by: Andy Shevchenko andriy.shevchenko@intel.com Cc: Daniel Mack daniel@zonque.org Cc: Haojian Zhuang haojian.zhuang@gmail.com Cc: Robert Jarzmik robert.jarzmik@free.fr Cc: Mark Brown broonie@kernel.org Cc: linux-arm-kernel@lists.infradead.org Cc: linux-spi@vger.kernel.org Signed-off-by: Daniel Vetter daniel.vetter@ffwll.ch -- v2: Brown paper bag edition, it needs to be a function ...
drivers/spi/spi-pxa2xx.c | 7 +++++++ 1 file changed, 7 insertions(+)
diff --git a/drivers/spi/spi-pxa2xx.c b/drivers/spi/spi-pxa2xx.c index bb6a14d1ab0f..068c21037679 100644 --- a/drivers/spi/spi-pxa2xx.c +++ b/drivers/spi/spi-pxa2xx.c @@ -1602,6 +1602,11 @@ static int pxa2xx_spi_fw_translate_cs(struct spi_controller *controller, return cs; }
+static size_t pxa2xx_spi_max_dma_transfer_size(struct spi_device *spi) +{
- return MAX_DMA_LEN;
+}
static int pxa2xx_spi_probe(struct platform_device *pdev) { struct device *dev = &pdev->dev; @@ -1707,6 +1712,8 @@ static int pxa2xx_spi_probe(struct platform_device *pdev) } else { controller->can_dma = pxa2xx_spi_can_dma; controller->max_dma_len = MAX_DMA_LEN;
controller->max_transfer_size =
} }pxa2xx_spi_max_dma_transfer_size;
-- 2.23.0
On Thu, Oct 17, 2019 at 10:06:38AM +0300, Andy Shevchenko wrote:
On Thu, Oct 17, 2019 at 08:44:26AM +0200, Daniel Vetter wrote:
In DMA mode we have a maximum transfer size, past that the driver falls back to PIO (see the check at the top of pxa2xx_spi_transfer_one). Falling back to PIO for big transfers defeats the point of a dma engine, hence set the max transfer size to inform spi clients that they need to do something smarter.
This was uncovered by the drm_mipi_dbi spi panel code, which does large spi transfers, but stopped splitting them after:
commit e143364b4c1774f68e923a5a0bb0fca28ac25888 Author: Noralf Trønnes noralf@tronnes.org Date: Fri Jul 19 17:59:10 2019 +0200
drm/tinydrm: Remove tinydrm_spi_max_transfer_size()
After this commit the code relied on the spi core to split transfers into max dma-able blocks, which also papered over the PIO fallback issue.
Fix this by setting the overall max transfer size to the DMA limit, but only when the controller runs in DMA mode.
Thank you, Daniel!
Mark, can be this applied?
On Mon, Oct 21, 2019 at 02:08:57PM +0300, Andy Shevchenko wrote:
Mark, can be this applied?
b2662a164f9dc48d
Please don't send content free pings and please allow a reasonable time for review. People get busy, go on holiday, attend conferences and so on so unless there is some reason for urgency (like critical bug fixes) please allow at least a couple of weeks for review. If there have been review comments then people may be waiting for those to be addressed.
Sending content free pings adds to the mail volume (if they are seen at all) which is often the problem and since they can't be reviewed directly if something has gone wrong you'll have to resend the patches anyway, so sending again is generally a better approach though there are some other maintainers who like them - if in doubt look at how patches for the subsystem are normally handled.
dri-devel@lists.freedesktop.org