diff mbox series

mmc: core: Lower max_seg_size if too high for DMA

Message ID 20181031155738.18367-1-tony@atomide.com (mailing list archive)
State New, archived
Headers show
Series mmc: core: Lower max_seg_size if too high for DMA | expand

Commit Message

Tony Lindgren Oct. 31, 2018, 3:57 p.m. UTC
With CONFIG_DMA_API_DEBUG_SG a device may produce the following warning:

"DMA-API: mapping sg segment longer than device claims to support"

We default to 64KiB if a DMA engine driver does not initialize dma_parms
and call dma_set_max_seg_size(). This may be lower that what many MMC
drivers do with mmc->max_seg_size = mmc->max_blk_size * mmc->max_blk_count.

Let's do a sanity check for max_seg_size being higher than what DMA
supports in mmc_add_host() and lower it as needed.

Cc: Kishon Vijay Abraham I <kishon@ti.com>
Cc: Peter Ujfalusi <peter.ujfalusi@ti.com>
Cc: Russell King <rmk+kernel@armlinux.org.uk>
Reported-by: Russell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: Tony Lindgren <tony@atomide.com>
---
 drivers/mmc/core/host.c | 16 ++++++++++++++++
 1 file changed, 16 insertions(+)

Comments

Peter Ujfalusi Nov. 12, 2018, 4:48 p.m. UTC | #1
Tony,

On 2018-10-31 17:57, Tony Lindgren wrote:
> With CONFIG_DMA_API_DEBUG_SG a device may produce the following warning:
> 
> "DMA-API: mapping sg segment longer than device claims to support"
> 
> We default to 64KiB if a DMA engine driver does not initialize dma_parms
> and call dma_set_max_seg_size(). This may be lower that what many MMC
> drivers do with mmc->max_seg_size = mmc->max_blk_size * mmc->max_blk_count.
> 
> Let's do a sanity check for max_seg_size being higher than what DMA
> supports in mmc_add_host() and lower it as needed.

I tried to address it with:

https://patchwork.kernel.org/patch/9948967/
https://patchwork.kernel.org/patch/9948989/
https://patchwork.kernel.org/patch/9948975/

but it is back in the drawing board for now.

The problem is that with eDMA/SDMA the size limit is not bytes, but the
number of bursts. Different addr_width and burst combination would
result different max_seg_size and I can not guess before the user
provide the  dma_slave_config via dmaengine_slave_config().

I have experimental patches to modify the max_seg_size of the _channel_
proactively in response slave_config and then we should have an API to
query it from the DMA _channel_, not from the DMA device as each channel
can have different limit on the max_seg_size.

> 
> Cc: Kishon Vijay Abraham I <kishon@ti.com>
> Cc: Peter Ujfalusi <peter.ujfalusi@ti.com>
> Cc: Russell King <rmk+kernel@armlinux.org.uk>
> Reported-by: Russell King <rmk+kernel@armlinux.org.uk>
> Signed-off-by: Tony Lindgren <tony@atomide.com>

Reviewed-by: Peter Ujfalusi <peter.ujfalusi@ti.com>

> ---
>  drivers/mmc/core/host.c | 16 ++++++++++++++++
>  1 file changed, 16 insertions(+)
> 
> diff --git a/drivers/mmc/core/host.c b/drivers/mmc/core/host.c
> --- a/drivers/mmc/core/host.c
> +++ b/drivers/mmc/core/host.c
> @@ -13,6 +13,7 @@
>   */
>  
>  #include <linux/device.h>
> +#include <linux/dma-mapping.h>
>  #include <linux/err.h>
>  #include <linux/idr.h>
>  #include <linux/of.h>
> @@ -415,6 +416,19 @@ struct mmc_host *mmc_alloc_host(int extra, struct device *dev)
>  
>  EXPORT_SYMBOL(mmc_alloc_host);
>  
> +static void mmc_check_max_seg_size(struct mmc_host *host)
> +{
> +	unsigned int max_seg_size = dma_get_max_seg_size(mmc_dev(host));
> +
> +	if (host->max_seg_size <= max_seg_size)
> +		return;
> +
> +	dev_info(mmc_dev(host), "Lowering max_seg_size for DMA: %u vs %u\n",
> +		 host->max_seg_size, max_seg_size);
> +
> +	host->max_seg_size = max_seg_size;
> +}
> +
>  /**
>   *	mmc_add_host - initialise host hardware
>   *	@host: mmc host
> @@ -430,6 +444,8 @@ int mmc_add_host(struct mmc_host *host)
>  	WARN_ON((host->caps & MMC_CAP_SDIO_IRQ) &&
>  		!host->ops->enable_sdio_irq);
>  
> +	mmc_check_max_seg_size(host);
> +
>  	err = device_add(&host->class_dev);
>  	if (err)
>  		return err;
> 

- Péter

Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki.
Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki
Ulf Hansson Nov. 19, 2018, 12:08 p.m. UTC | #2
On 31 October 2018 at 16:57, Tony Lindgren <tony@atomide.com> wrote:
> With CONFIG_DMA_API_DEBUG_SG a device may produce the following warning:
>
> "DMA-API: mapping sg segment longer than device claims to support"
>
> We default to 64KiB if a DMA engine driver does not initialize dma_parms
> and call dma_set_max_seg_size(). This may be lower that what many MMC
> drivers do with mmc->max_seg_size = mmc->max_blk_size * mmc->max_blk_count.
>
> Let's do a sanity check for max_seg_size being higher than what DMA
> supports in mmc_add_host() and lower it as needed.
>
> Cc: Kishon Vijay Abraham I <kishon@ti.com>
> Cc: Peter Ujfalusi <peter.ujfalusi@ti.com>
> Cc: Russell King <rmk+kernel@armlinux.org.uk>
> Reported-by: Russell King <rmk+kernel@armlinux.org.uk>
> Signed-off-by: Tony Lindgren <tony@atomide.com>
> ---
>  drivers/mmc/core/host.c | 16 ++++++++++++++++
>  1 file changed, 16 insertions(+)
>
> diff --git a/drivers/mmc/core/host.c b/drivers/mmc/core/host.c
> --- a/drivers/mmc/core/host.c
> +++ b/drivers/mmc/core/host.c
> @@ -13,6 +13,7 @@
>   */
>
>  #include <linux/device.h>
> +#include <linux/dma-mapping.h>
>  #include <linux/err.h>
>  #include <linux/idr.h>
>  #include <linux/of.h>
> @@ -415,6 +416,19 @@ struct mmc_host *mmc_alloc_host(int extra, struct device *dev)
>
>  EXPORT_SYMBOL(mmc_alloc_host);
>
> +static void mmc_check_max_seg_size(struct mmc_host *host)
> +{
> +       unsigned int max_seg_size = dma_get_max_seg_size(mmc_dev(host));

Is dma_get_max_seg_size() really intended to be called for any struct
device (representing the mmc controller) like this?

My understanding is that the dma_get_max_seg_size() is supposed to be
called by using the DMA engine device, no?

> +
> +       if (host->max_seg_size <= max_seg_size)
> +               return;
> +
> +       dev_info(mmc_dev(host), "Lowering max_seg_size for DMA: %u vs %u\n",
> +                host->max_seg_size, max_seg_size);
> +
> +       host->max_seg_size = max_seg_size;
> +}
> +
>  /**
>   *     mmc_add_host - initialise host hardware
>   *     @host: mmc host
> @@ -430,6 +444,8 @@ int mmc_add_host(struct mmc_host *host)
>         WARN_ON((host->caps & MMC_CAP_SDIO_IRQ) &&
>                 !host->ops->enable_sdio_irq);
>
> +       mmc_check_max_seg_size(host);
> +
>         err = device_add(&host->class_dev);
>         if (err)
>                 return err;
> --
> 2.19.1

Kind regards
Uffe
Tony Lindgren Nov. 29, 2018, 7:13 p.m. UTC | #3
* Ulf Hansson <ulf.hansson@linaro.org> [181119 12:09]:
> On 31 October 2018 at 16:57, Tony Lindgren <tony@atomide.com> wrote:
> > With CONFIG_DMA_API_DEBUG_SG a device may produce the following warning:
> >
> > "DMA-API: mapping sg segment longer than device claims to support"
> >
> > We default to 64KiB if a DMA engine driver does not initialize dma_parms
> > and call dma_set_max_seg_size(). This may be lower that what many MMC
> > drivers do with mmc->max_seg_size = mmc->max_blk_size * mmc->max_blk_count.
> >
> > Let's do a sanity check for max_seg_size being higher than what DMA
> > supports in mmc_add_host() and lower it as needed.
> >
> > Cc: Kishon Vijay Abraham I <kishon@ti.com>
> > Cc: Peter Ujfalusi <peter.ujfalusi@ti.com>
> > Cc: Russell King <rmk+kernel@armlinux.org.uk>
> > Reported-by: Russell King <rmk+kernel@armlinux.org.uk>
> > Signed-off-by: Tony Lindgren <tony@atomide.com>
> > ---
> >  drivers/mmc/core/host.c | 16 ++++++++++++++++
> >  1 file changed, 16 insertions(+)
> >
> > diff --git a/drivers/mmc/core/host.c b/drivers/mmc/core/host.c
> > --- a/drivers/mmc/core/host.c
> > +++ b/drivers/mmc/core/host.c
> > @@ -13,6 +13,7 @@
> >   */
> >
> >  #include <linux/device.h>
> > +#include <linux/dma-mapping.h>
> >  #include <linux/err.h>
> >  #include <linux/idr.h>
> >  #include <linux/of.h>
> > @@ -415,6 +416,19 @@ struct mmc_host *mmc_alloc_host(int extra, struct device *dev)
> >
> >  EXPORT_SYMBOL(mmc_alloc_host);
> >
> > +static void mmc_check_max_seg_size(struct mmc_host *host)
> > +{
> > +       unsigned int max_seg_size = dma_get_max_seg_size(mmc_dev(host));
> 
> Is dma_get_max_seg_size() really intended to be called for any struct
> device (representing the mmc controller) like this?
> 
> My understanding is that the dma_get_max_seg_size() is supposed to be
> called by using the DMA engine device, no?

Oh good catch sounds like I'm calling it for the wrong device,
need to check. In that case sounds like this can't be generic?

Regards,

Tony
Ulf Hansson Nov. 29, 2018, 8:11 p.m. UTC | #4
On Thu, 29 Nov 2018 at 20:13, Tony Lindgren <tony@atomide.com> wrote:
>
> * Ulf Hansson <ulf.hansson@linaro.org> [181119 12:09]:
> > On 31 October 2018 at 16:57, Tony Lindgren <tony@atomide.com> wrote:
> > > With CONFIG_DMA_API_DEBUG_SG a device may produce the following warning:
> > >
> > > "DMA-API: mapping sg segment longer than device claims to support"
> > >
> > > We default to 64KiB if a DMA engine driver does not initialize dma_parms
> > > and call dma_set_max_seg_size(). This may be lower that what many MMC
> > > drivers do with mmc->max_seg_size = mmc->max_blk_size * mmc->max_blk_count.
> > >
> > > Let's do a sanity check for max_seg_size being higher than what DMA
> > > supports in mmc_add_host() and lower it as needed.
> > >
> > > Cc: Kishon Vijay Abraham I <kishon@ti.com>
> > > Cc: Peter Ujfalusi <peter.ujfalusi@ti.com>
> > > Cc: Russell King <rmk+kernel@armlinux.org.uk>
> > > Reported-by: Russell King <rmk+kernel@armlinux.org.uk>
> > > Signed-off-by: Tony Lindgren <tony@atomide.com>
> > > ---
> > >  drivers/mmc/core/host.c | 16 ++++++++++++++++
> > >  1 file changed, 16 insertions(+)
> > >
> > > diff --git a/drivers/mmc/core/host.c b/drivers/mmc/core/host.c
> > > --- a/drivers/mmc/core/host.c
> > > +++ b/drivers/mmc/core/host.c
> > > @@ -13,6 +13,7 @@
> > >   */
> > >
> > >  #include <linux/device.h>
> > > +#include <linux/dma-mapping.h>
> > >  #include <linux/err.h>
> > >  #include <linux/idr.h>
> > >  #include <linux/of.h>
> > > @@ -415,6 +416,19 @@ struct mmc_host *mmc_alloc_host(int extra, struct device *dev)
> > >
> > >  EXPORT_SYMBOL(mmc_alloc_host);
> > >
> > > +static void mmc_check_max_seg_size(struct mmc_host *host)
> > > +{
> > > +       unsigned int max_seg_size = dma_get_max_seg_size(mmc_dev(host));
> >
> > Is dma_get_max_seg_size() really intended to be called for any struct
> > device (representing the mmc controller) like this?
> >
> > My understanding is that the dma_get_max_seg_size() is supposed to be
> > called by using the DMA engine device, no?
>
> Oh good catch sounds like I'm calling it for the wrong device,
> need to check. In that case sounds like this can't be generic?

No, I don't think so as it's only the mmc host driver that knows about
the DMA engine device.

Kind regards
Uffe
Russell King (Oracle) Dec. 11, 2018, 1:13 p.m. UTC | #5
On Thu, Nov 29, 2018 at 09:11:17PM +0100, Ulf Hansson wrote:
> On Thu, 29 Nov 2018 at 20:13, Tony Lindgren <tony@atomide.com> wrote:
> >
> > * Ulf Hansson <ulf.hansson@linaro.org> [181119 12:09]:
> > > On 31 October 2018 at 16:57, Tony Lindgren <tony@atomide.com> wrote:
> > > > With CONFIG_DMA_API_DEBUG_SG a device may produce the following warning:
> > > >
> > > > "DMA-API: mapping sg segment longer than device claims to support"
> > > >
> > > > We default to 64KiB if a DMA engine driver does not initialize dma_parms
> > > > and call dma_set_max_seg_size(). This may be lower that what many MMC
> > > > drivers do with mmc->max_seg_size = mmc->max_blk_size * mmc->max_blk_count.
> > > >
> > > > Let's do a sanity check for max_seg_size being higher than what DMA
> > > > supports in mmc_add_host() and lower it as needed.
> > > >
> > > > Cc: Kishon Vijay Abraham I <kishon@ti.com>
> > > > Cc: Peter Ujfalusi <peter.ujfalusi@ti.com>
> > > > Cc: Russell King <rmk+kernel@armlinux.org.uk>
> > > > Reported-by: Russell King <rmk+kernel@armlinux.org.uk>
> > > > Signed-off-by: Tony Lindgren <tony@atomide.com>
> > > > ---
> > > >  drivers/mmc/core/host.c | 16 ++++++++++++++++
> > > >  1 file changed, 16 insertions(+)
> > > >
> > > > diff --git a/drivers/mmc/core/host.c b/drivers/mmc/core/host.c
> > > > --- a/drivers/mmc/core/host.c
> > > > +++ b/drivers/mmc/core/host.c
> > > > @@ -13,6 +13,7 @@
> > > >   */
> > > >
> > > >  #include <linux/device.h>
> > > > +#include <linux/dma-mapping.h>
> > > >  #include <linux/err.h>
> > > >  #include <linux/idr.h>
> > > >  #include <linux/of.h>
> > > > @@ -415,6 +416,19 @@ struct mmc_host *mmc_alloc_host(int extra, struct device *dev)
> > > >
> > > >  EXPORT_SYMBOL(mmc_alloc_host);
> > > >
> > > > +static void mmc_check_max_seg_size(struct mmc_host *host)
> > > > +{
> > > > +       unsigned int max_seg_size = dma_get_max_seg_size(mmc_dev(host));
> > >
> > > Is dma_get_max_seg_size() really intended to be called for any struct
> > > device (representing the mmc controller) like this?
> > >
> > > My understanding is that the dma_get_max_seg_size() is supposed to be
> > > called by using the DMA engine device, no?
> >
> > Oh good catch sounds like I'm calling it for the wrong device,
> > need to check. In that case sounds like this can't be generic?
> 
> No, I don't think so as it's only the mmc host driver that knows about
> the DMA engine device.

We're nearing the merge window, and this is a regression that is yet
to be solved.  It causes a kernel warning with backtrace, so it's
very annoying.

The error is:

omap-dma-engine 4a056000.dma-controller: DMA-API: mapping sg segment longer than device claims to support [len=69632] [max=65536]

which indicates that we have a SG segment length that exceeds thte
published maximum segment size for a device - in this case the
DMA engine device.  The maximum segment size for the DMA engine comes
from the default per-device setting, in linux/dma-mapping.h, which is
64K.

However, omap_hsmmc sets:

        mmc->max_blk_size = 512;       /* Block Length at max can be 1024 */
        mmc->max_blk_count = 0xFFFF;    /* No. of Blocks is 16 bits */
        mmc->max_req_size = mmc->max_blk_size * mmc->max_blk_count;
        mmc->max_seg_size = mmc->max_req_size;

which ends up telling the block layer that we support a maximum segment
size of 65535*512, so of course it _will_ pass SG lists where a segment
is longer than 64K.

The problem here is that the HSMMC driver doesn't take account of the
DMA engine device's capabilities.

We have something of an odd situation in that the omap-dma device's
maximum SG size depends on the "address width" - it's 64K transfers
of whatever unit "address width" is, so the current implementation of
per-device parameters doesn't exactly work.  That means the default
64K limit at the device-level is reasonable.

The only thing I can think of doing is adding into omap_hsmmc:

	mmc->max_seg_size = min(mmc->max_req_size,
				min(dma_get_max_seg_size(host->rx_chan->device->dev),
				    dma_get_max_seg_size(host->tx_chan->device->dev)));

to limit the maximum segment size to that of the device _and_ dma
engine's capabilities.

Doing this solves the problem for me.
Ulf Hansson Dec. 11, 2018, 1:39 p.m. UTC | #6
On Tue, 11 Dec 2018 at 14:13, Russell King - ARM Linux
<linux@armlinux.org.uk> wrote:
>
> On Thu, Nov 29, 2018 at 09:11:17PM +0100, Ulf Hansson wrote:
> > On Thu, 29 Nov 2018 at 20:13, Tony Lindgren <tony@atomide.com> wrote:
> > >
> > > * Ulf Hansson <ulf.hansson@linaro.org> [181119 12:09]:
> > > > On 31 October 2018 at 16:57, Tony Lindgren <tony@atomide.com> wrote:
> > > > > With CONFIG_DMA_API_DEBUG_SG a device may produce the following warning:
> > > > >
> > > > > "DMA-API: mapping sg segment longer than device claims to support"
> > > > >
> > > > > We default to 64KiB if a DMA engine driver does not initialize dma_parms
> > > > > and call dma_set_max_seg_size(). This may be lower that what many MMC
> > > > > drivers do with mmc->max_seg_size = mmc->max_blk_size * mmc->max_blk_count.
> > > > >
> > > > > Let's do a sanity check for max_seg_size being higher than what DMA
> > > > > supports in mmc_add_host() and lower it as needed.
> > > > >
> > > > > Cc: Kishon Vijay Abraham I <kishon@ti.com>
> > > > > Cc: Peter Ujfalusi <peter.ujfalusi@ti.com>
> > > > > Cc: Russell King <rmk+kernel@armlinux.org.uk>
> > > > > Reported-by: Russell King <rmk+kernel@armlinux.org.uk>
> > > > > Signed-off-by: Tony Lindgren <tony@atomide.com>
> > > > > ---
> > > > >  drivers/mmc/core/host.c | 16 ++++++++++++++++
> > > > >  1 file changed, 16 insertions(+)
> > > > >
> > > > > diff --git a/drivers/mmc/core/host.c b/drivers/mmc/core/host.c
> > > > > --- a/drivers/mmc/core/host.c
> > > > > +++ b/drivers/mmc/core/host.c
> > > > > @@ -13,6 +13,7 @@
> > > > >   */
> > > > >
> > > > >  #include <linux/device.h>
> > > > > +#include <linux/dma-mapping.h>
> > > > >  #include <linux/err.h>
> > > > >  #include <linux/idr.h>
> > > > >  #include <linux/of.h>
> > > > > @@ -415,6 +416,19 @@ struct mmc_host *mmc_alloc_host(int extra, struct device *dev)
> > > > >
> > > > >  EXPORT_SYMBOL(mmc_alloc_host);
> > > > >
> > > > > +static void mmc_check_max_seg_size(struct mmc_host *host)
> > > > > +{
> > > > > +       unsigned int max_seg_size = dma_get_max_seg_size(mmc_dev(host));
> > > >
> > > > Is dma_get_max_seg_size() really intended to be called for any struct
> > > > device (representing the mmc controller) like this?
> > > >
> > > > My understanding is that the dma_get_max_seg_size() is supposed to be
> > > > called by using the DMA engine device, no?
> > >
> > > Oh good catch sounds like I'm calling it for the wrong device,
> > > need to check. In that case sounds like this can't be generic?
> >
> > No, I don't think so as it's only the mmc host driver that knows about
> > the DMA engine device.
>
> We're nearing the merge window, and this is a regression that is yet
> to be solved.  It causes a kernel warning with backtrace, so it's
> very annoying.
>
> The error is:
>
> omap-dma-engine 4a056000.dma-controller: DMA-API: mapping sg segment longer than device claims to support [len=69632] [max=65536]
>
> which indicates that we have a SG segment length that exceeds thte
> published maximum segment size for a device - in this case the
> DMA engine device.  The maximum segment size for the DMA engine comes
> from the default per-device setting, in linux/dma-mapping.h, which is
> 64K.
>
> However, omap_hsmmc sets:
>
>         mmc->max_blk_size = 512;       /* Block Length at max can be 1024 */
>         mmc->max_blk_count = 0xFFFF;    /* No. of Blocks is 16 bits */
>         mmc->max_req_size = mmc->max_blk_size * mmc->max_blk_count;
>         mmc->max_seg_size = mmc->max_req_size;
>
> which ends up telling the block layer that we support a maximum segment
> size of 65535*512, so of course it _will_ pass SG lists where a segment
> is longer than 64K.
>
> The problem here is that the HSMMC driver doesn't take account of the
> DMA engine device's capabilities.
>
> We have something of an odd situation in that the omap-dma device's
> maximum SG size depends on the "address width" - it's 64K transfers
> of whatever unit "address width" is, so the current implementation of
> per-device parameters doesn't exactly work.  That means the default
> 64K limit at the device-level is reasonable.
>
> The only thing I can think of doing is adding into omap_hsmmc:
>
>         mmc->max_seg_size = min(mmc->max_req_size,
>                                 min(dma_get_max_seg_size(host->rx_chan->device->dev),
>                                     dma_get_max_seg_size(host->tx_chan->device->dev)));
>
> to limit the maximum segment size to that of the device _and_ dma
> engine's capabilities.
>
> Doing this solves the problem for me.

Thanks for the suggestion - it sounds like a reasonable way to fix the
problem, at least for now.

Do you want to send to patch or do you expect someone else to do it?

Kind regards
Uffe
Peter Ujfalusi Dec. 11, 2018, 1:49 p.m. UTC | #7
Russell,

On 11/12/2018 15.13, Russell King - ARM Linux wrote:
> 
> We're nearing the merge window, and this is a regression that is yet
> to be solved.  It causes a kernel warning with backtrace, so it's
> very annoying.
> 
> The error is:
> 
> omap-dma-engine 4a056000.dma-controller: DMA-API: mapping sg segment longer than device claims to support [len=69632] [max=65536]
> 
> which indicates that we have a SG segment length that exceeds thte
> published maximum segment size for a device - in this case the
> DMA engine device.  The maximum segment size for the DMA engine comes
> from the default per-device setting, in linux/dma-mapping.h, which is
> 64K.
> 
> However, omap_hsmmc sets:
> 
>         mmc->max_blk_size = 512;       /* Block Length at max can be 1024 */
>         mmc->max_blk_count = 0xFFFF;    /* No. of Blocks is 16 bits */
>         mmc->max_req_size = mmc->max_blk_size * mmc->max_blk_count;
>         mmc->max_seg_size = mmc->max_req_size;
> 
> which ends up telling the block layer that we support a maximum segment
> size of 65535*512, so of course it _will_ pass SG lists where a segment
> is longer than 64K.
> 
> The problem here is that the HSMMC driver doesn't take account of the
> DMA engine device's capabilities.
> 
> We have something of an odd situation in that the omap-dma device's
> maximum SG size depends on the "address width" - it's 64K transfers
> of whatever unit "address width" is, so the current implementation of
> per-device parameters doesn't exactly work.  That means the default
> 64K limit at the device-level is reasonable.
> 
> The only thing I can think of doing is adding into omap_hsmmc:
> 
> 	mmc->max_seg_size = min(mmc->max_req_size,
> 				min(dma_get_max_seg_size(host->rx_chan->device->dev),
> 				    dma_get_max_seg_size(host->tx_chan->device->dev)));
> 
> to limit the maximum segment size to that of the device _and_ dma
> engine's capabilities.

Make sense.

> Doing this solves the problem for me.
> 

- Péter

Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki.
Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki
Russell King (Oracle) Dec. 11, 2018, 1:49 p.m. UTC | #8
On Tue, Dec 11, 2018 at 02:39:11PM +0100, Ulf Hansson wrote:
> On Tue, 11 Dec 2018 at 14:13, Russell King - ARM Linux
> <linux@armlinux.org.uk> wrote:
> >
> > On Thu, Nov 29, 2018 at 09:11:17PM +0100, Ulf Hansson wrote:
> > > On Thu, 29 Nov 2018 at 20:13, Tony Lindgren <tony@atomide.com> wrote:
> > > >
> > > > * Ulf Hansson <ulf.hansson@linaro.org> [181119 12:09]:
> > > > > On 31 October 2018 at 16:57, Tony Lindgren <tony@atomide.com> wrote:
> > > > > > With CONFIG_DMA_API_DEBUG_SG a device may produce the following warning:
> > > > > >
> > > > > > "DMA-API: mapping sg segment longer than device claims to support"
> > > > > >
> > > > > > We default to 64KiB if a DMA engine driver does not initialize dma_parms
> > > > > > and call dma_set_max_seg_size(). This may be lower that what many MMC
> > > > > > drivers do with mmc->max_seg_size = mmc->max_blk_size * mmc->max_blk_count.
> > > > > >
> > > > > > Let's do a sanity check for max_seg_size being higher than what DMA
> > > > > > supports in mmc_add_host() and lower it as needed.
> > > > > >
> > > > > > Cc: Kishon Vijay Abraham I <kishon@ti.com>
> > > > > > Cc: Peter Ujfalusi <peter.ujfalusi@ti.com>
> > > > > > Cc: Russell King <rmk+kernel@armlinux.org.uk>
> > > > > > Reported-by: Russell King <rmk+kernel@armlinux.org.uk>
> > > > > > Signed-off-by: Tony Lindgren <tony@atomide.com>
> > > > > > ---
> > > > > >  drivers/mmc/core/host.c | 16 ++++++++++++++++
> > > > > >  1 file changed, 16 insertions(+)
> > > > > >
> > > > > > diff --git a/drivers/mmc/core/host.c b/drivers/mmc/core/host.c
> > > > > > --- a/drivers/mmc/core/host.c
> > > > > > +++ b/drivers/mmc/core/host.c
> > > > > > @@ -13,6 +13,7 @@
> > > > > >   */
> > > > > >
> > > > > >  #include <linux/device.h>
> > > > > > +#include <linux/dma-mapping.h>
> > > > > >  #include <linux/err.h>
> > > > > >  #include <linux/idr.h>
> > > > > >  #include <linux/of.h>
> > > > > > @@ -415,6 +416,19 @@ struct mmc_host *mmc_alloc_host(int extra, struct device *dev)
> > > > > >
> > > > > >  EXPORT_SYMBOL(mmc_alloc_host);
> > > > > >
> > > > > > +static void mmc_check_max_seg_size(struct mmc_host *host)
> > > > > > +{
> > > > > > +       unsigned int max_seg_size = dma_get_max_seg_size(mmc_dev(host));
> > > > >
> > > > > Is dma_get_max_seg_size() really intended to be called for any struct
> > > > > device (representing the mmc controller) like this?
> > > > >
> > > > > My understanding is that the dma_get_max_seg_size() is supposed to be
> > > > > called by using the DMA engine device, no?
> > > >
> > > > Oh good catch sounds like I'm calling it for the wrong device,
> > > > need to check. In that case sounds like this can't be generic?
> > >
> > > No, I don't think so as it's only the mmc host driver that knows about
> > > the DMA engine device.
> >
> > We're nearing the merge window, and this is a regression that is yet
> > to be solved.  It causes a kernel warning with backtrace, so it's
> > very annoying.
> >
> > The error is:
> >
> > omap-dma-engine 4a056000.dma-controller: DMA-API: mapping sg segment longer than device claims to support [len=69632] [max=65536]
> >
> > which indicates that we have a SG segment length that exceeds thte
> > published maximum segment size for a device - in this case the
> > DMA engine device.  The maximum segment size for the DMA engine comes
> > from the default per-device setting, in linux/dma-mapping.h, which is
> > 64K.
> >
> > However, omap_hsmmc sets:
> >
> >         mmc->max_blk_size = 512;       /* Block Length at max can be 1024 */
> >         mmc->max_blk_count = 0xFFFF;    /* No. of Blocks is 16 bits */
> >         mmc->max_req_size = mmc->max_blk_size * mmc->max_blk_count;
> >         mmc->max_seg_size = mmc->max_req_size;
> >
> > which ends up telling the block layer that we support a maximum segment
> > size of 65535*512, so of course it _will_ pass SG lists where a segment
> > is longer than 64K.
> >
> > The problem here is that the HSMMC driver doesn't take account of the
> > DMA engine device's capabilities.
> >
> > We have something of an odd situation in that the omap-dma device's
> > maximum SG size depends on the "address width" - it's 64K transfers
> > of whatever unit "address width" is, so the current implementation of
> > per-device parameters doesn't exactly work.  That means the default
> > 64K limit at the device-level is reasonable.
> >
> > The only thing I can think of doing is adding into omap_hsmmc:
> >
> >         mmc->max_seg_size = min(mmc->max_req_size,
> >                                 min(dma_get_max_seg_size(host->rx_chan->device->dev),
> >                                     dma_get_max_seg_size(host->tx_chan->device->dev)));
> >
> > to limit the maximum segment size to that of the device _and_ dma
> > engine's capabilities.
> >
> > Doing this solves the problem for me.
> 
> Thanks for the suggestion - it sounds like a reasonable way to fix the
> problem, at least for now.

I don't think there's any "at least for now" about this approach - it's
not something that the MMC core can know about, because whether a driver
uses DMA engine or not, and how many channels it uses is completely
driver specific.

The only questionable thing is around the dma_get_max_seg_size()
interface, but the only way that's going to get solved is to eliminate
it entirely as it isn't a per-device property with some DMA engines
such as omap-dma - it's a per-request property.  That also means
killing the check in the DMA debug code, which isn't going to go down
very well.

> Do you want to send to patch or do you expect someone else to do it?

I'll send a patch once I've checked the corner cases, and whether
we should go further and include other DMA parameters from the
dma engine.
Tony Lindgren Dec. 11, 2018, 2:23 p.m. UTC | #9
* Peter Ujfalusi <peter.ujfalusi@ti.com> [181211 13:47]:
> On 11/12/2018 15.13, Russell King - ARM Linux wrote:
> > The only thing I can think of doing is adding into omap_hsmmc:
> > 
> > 	mmc->max_seg_size = min(mmc->max_req_size,
> > 				min(dma_get_max_seg_size(host->rx_chan->device->dev),
> > 				    dma_get_max_seg_size(host->tx_chan->device->dev)));
> > 
> > to limit the maximum segment size to that of the device _and_ dma
> > engine's capabilities.
> 
> Make sense.
> 
> > Doing this solves the problem for me.

OK sounds good to me too.

Thanks,

Tony
Peter Ujfalusi Dec. 11, 2018, 2:33 p.m. UTC | #10
Russell,

On 11/12/2018 15.49, Russell King - ARM Linux wrote:
>>> The only thing I can think of doing is adding into omap_hsmmc:
>>>
>>>         mmc->max_seg_size = min(mmc->max_req_size,
>>>                                 min(dma_get_max_seg_size(host->rx_chan->device->dev),
>>>                                     dma_get_max_seg_size(host->tx_chan->device->dev)));
>>>
>>> to limit the maximum segment size to that of the device _and_ dma
>>> engine's capabilities.
>>>
>>> Doing this solves the problem for me.
>>
>> Thanks for the suggestion - it sounds like a reasonable way to fix the
>> problem, at least for now.
> 
> I don't think there's any "at least for now" about this approach - it's
> not something that the MMC core can know about, because whether a driver
> uses DMA engine or not, and how many channels it uses is completely
> driver specific.

I agree.

> The only questionable thing is around the dma_get_max_seg_size()
> interface, but the only way that's going to get solved is to eliminate
> it entirely as it isn't a per-device property with some DMA engines
> such as omap-dma - it's a per-request property.  That also means
> killing the check in the DMA debug code, which isn't going to go down
> very well.
> 
>> Do you want to send to patch or do you expect someone else to do it?
> 
> I'll send a patch once I've checked the corner cases, and whether
> we should go further and include other DMA parameters from the
> dma engine.

There is one thing which I'm trying to figure out:
sDMA (and EDMA also) have limitation on the length of each sg segment.

To make things a bit more interesting the limit is not in bytes, but in
number of bursts:

65535 * burst * addr_width = max_sg_segment_len in bytes.

We can not set this to a static value as depending on the addr_width and
burst it can be different for each transfer.

- Péter

Texas Instruments Finland Oy, Porkkalankatu 22, 00180 Helsinki.
Y-tunnus/Business ID: 0615521-4. Kotipaikka/Domicile: Helsinki
Russell King (Oracle) Dec. 11, 2018, 2:42 p.m. UTC | #11
On Tue, Dec 11, 2018 at 04:33:26PM +0200, Peter Ujfalusi wrote:
> Russell,
> 
> On 11/12/2018 15.49, Russell King - ARM Linux wrote:
> >>> The only thing I can think of doing is adding into omap_hsmmc:
> >>>
> >>>         mmc->max_seg_size = min(mmc->max_req_size,
> >>>                                 min(dma_get_max_seg_size(host->rx_chan->device->dev),
> >>>                                     dma_get_max_seg_size(host->tx_chan->device->dev)));
> >>>
> >>> to limit the maximum segment size to that of the device _and_ dma
> >>> engine's capabilities.
> >>>
> >>> Doing this solves the problem for me.
> >>
> >> Thanks for the suggestion - it sounds like a reasonable way to fix the
> >> problem, at least for now.
> > 
> > I don't think there's any "at least for now" about this approach - it's
> > not something that the MMC core can know about, because whether a driver
> > uses DMA engine or not, and how many channels it uses is completely
> > driver specific.
> 
> I agree.
> 
> > The only questionable thing is around the dma_get_max_seg_size()
> > interface, but the only way that's going to get solved is to eliminate
> > it entirely as it isn't a per-device property with some DMA engines
> > such as omap-dma - it's a per-request property.  That also means
> > killing the check in the DMA debug code, which isn't going to go down
> > very well.
> > 
> >> Do you want to send to patch or do you expect someone else to do it?
> > 
> > I'll send a patch once I've checked the corner cases, and whether
> > we should go further and include other DMA parameters from the
> > dma engine.
> 
> There is one thing which I'm trying to figure out:
> sDMA (and EDMA also) have limitation on the length of each sg segment.
> 
> To make things a bit more interesting the limit is not in bytes, but in
> number of bursts:
> 
> 65535 * burst * addr_width = max_sg_segment_len in bytes.
> 
> We can not set this to a static value as depending on the addr_width and
> burst it can be different for each transfer.

However, it does have to be a static value as it's at device level,
rather than per-channel or per-request.  If it is set dynamically
by a request submission, it could vary wildly depending on the burst
and address width of the last request submitted to DMA engine, and
so the parameter becomes meaningless.

Consider what would happen to the value if we had one channel being
used with burst=32 addr_width=4 and another channel also being used
simultaneously with burst=8 addr_width=1...
diff mbox series

Patch

diff --git a/drivers/mmc/core/host.c b/drivers/mmc/core/host.c
--- a/drivers/mmc/core/host.c
+++ b/drivers/mmc/core/host.c
@@ -13,6 +13,7 @@ 
  */
 
 #include <linux/device.h>
+#include <linux/dma-mapping.h>
 #include <linux/err.h>
 #include <linux/idr.h>
 #include <linux/of.h>
@@ -415,6 +416,19 @@  struct mmc_host *mmc_alloc_host(int extra, struct device *dev)
 
 EXPORT_SYMBOL(mmc_alloc_host);
 
+static void mmc_check_max_seg_size(struct mmc_host *host)
+{
+	unsigned int max_seg_size = dma_get_max_seg_size(mmc_dev(host));
+
+	if (host->max_seg_size <= max_seg_size)
+		return;
+
+	dev_info(mmc_dev(host), "Lowering max_seg_size for DMA: %u vs %u\n",
+		 host->max_seg_size, max_seg_size);
+
+	host->max_seg_size = max_seg_size;
+}
+
 /**
  *	mmc_add_host - initialise host hardware
  *	@host: mmc host
@@ -430,6 +444,8 @@  int mmc_add_host(struct mmc_host *host)
 	WARN_ON((host->caps & MMC_CAP_SDIO_IRQ) &&
 		!host->ops->enable_sdio_irq);
 
+	mmc_check_max_seg_size(host);
+
 	err = device_add(&host->class_dev);
 	if (err)
 		return err;