On Mon, Feb 22, 2016 at 4:07 PM, Bjorn Andersson
<[email protected]> wrote:
> On Mon 22 Feb 02:03 PST 2016, Srinivas Kandagatla wrote:
>
>>
>>
>> On 22/02/16 05:32, Bjorn Andersson wrote:
>> >On certain platforms (e.g. ARM64) the dma_ops needs to be explicitly set
>> >to be able to do DMA allocations, so use the of_dma_configure() helper
>> >to populate the dma properties and assign an appropriate dma_ops.
We also hit the same issue with the dwc3 driver.
>> >
>> >Signed-off-by: Bjorn Andersson <[email protected]>
>> >---
>> > drivers/usb/chipidea/core.c | 4 ++++
>> > 1 file changed, 4 insertions(+)
>> >
>> >diff --git a/drivers/usb/chipidea/core.c b/drivers/usb/chipidea/core.c
>> >index 7404064b9bbc..047b9d4e67aa 100644
>> >--- a/drivers/usb/chipidea/core.c
>> >+++ b/drivers/usb/chipidea/core.c
>> >@@ -62,6 +62,7 @@
>> > #include <linux/usb/chipidea.h>
>> > #include <linux/usb/of.h>
>> > #include <linux/of.h>
>> >+#include <linux/of_device.h>
>> > #include <linux/phy.h>
>> > #include <linux/regulator/consumer.h>
>> > #include <linux/usb/ehci_def.h>
>> >@@ -834,6 +835,9 @@ struct platform_device *ci_hdrc_add_device(struct device *dev,
>> > pdev->dev.dma_parms = dev->dma_parms;
>> > dma_set_coherent_mask(&pdev->dev, dev->coherent_dma_mask);
>> >
>> >+ if (IS_ENABLED(CONFIG_OF) && dev->of_node)
>> >+ of_dma_configure(&pdev->dev, dev->of_node);
>> >+
>> Would we hit the same issue if we are on non Device tree platforms like ACPI
>> or platform device style itself?
>>
>
> As far as I can see, yes.
>
>>
>> > ret = platform_device_add_resources(pdev, res, nres);
>> > if (ret)
>> > goto err;
>> >
>>
>> I think this is the side effect of commit
>> 1dccb598df549d892b6450c261da54cdd7af44b4(arm64: simplify dma_get_ops)
>>
>
> I agree, before that we would have hit:
>
> __generic_dma_ops() {
> ..
> else if (acpi_disabled)
> return dma_ops;
> ...
> }
>
> with dma_ops being swiotlb_dma_ops from arm64_dma_init().
>
>
> But this would not have saved us in the ACPI case, i.e. the result would
> have been as with my suggested patch. Poking Arnd here to see if he has
> any input.
>
>> None of the drivers call of_dma_configure() explicitly, which makes me feel
>> that we are doing something wrong. TBH, this should be handled in more
>> generic way rather than driver like this having an explicit call to
>> of_dma_configure().
>>
>
> I agree, trying to figure out if it should be inherited or something.
I also agree. We need address it in a more generic way. I did a
search for platform_device_add()/platform_device_register() in the
kernel source code. I found a lot of them and many could be also
doing DMA. Looks like it is still too early to assume every device is
already getting dma_ops set through bus probe. Otherwise, many
drivers are potentially broken by this assumption.
Regards,
Leo
On Wed, Mar 2, 2016 at 4:59 PM, Li Yang <[email protected]> wrote:
> On Mon, Feb 22, 2016 at 4:07 PM, Bjorn Andersson
> <[email protected]> wrote:
>> On Mon 22 Feb 02:03 PST 2016, Srinivas Kandagatla wrote:
>>
>>>
>>>
>>> On 22/02/16 05:32, Bjorn Andersson wrote:
>>> >On certain platforms (e.g. ARM64) the dma_ops needs to be explicitly set
>>> >to be able to do DMA allocations, so use the of_dma_configure() helper
>>> >to populate the dma properties and assign an appropriate dma_ops.
>
> We also hit the same issue with the dwc3 driver.
>
>>> >
>>> >Signed-off-by: Bjorn Andersson <[email protected]>
>>> >---
>>> > drivers/usb/chipidea/core.c | 4 ++++
>>> > 1 file changed, 4 insertions(+)
>>> >
>>> >diff --git a/drivers/usb/chipidea/core.c b/drivers/usb/chipidea/core.c
>>> >index 7404064b9bbc..047b9d4e67aa 100644
>>> >--- a/drivers/usb/chipidea/core.c
>>> >+++ b/drivers/usb/chipidea/core.c
>>> >@@ -62,6 +62,7 @@
>>> > #include <linux/usb/chipidea.h>
>>> > #include <linux/usb/of.h>
>>> > #include <linux/of.h>
>>> >+#include <linux/of_device.h>
>>> > #include <linux/phy.h>
>>> > #include <linux/regulator/consumer.h>
>>> > #include <linux/usb/ehci_def.h>
>>> >@@ -834,6 +835,9 @@ struct platform_device *ci_hdrc_add_device(struct device *dev,
>>> > pdev->dev.dma_parms = dev->dma_parms;
>>> > dma_set_coherent_mask(&pdev->dev, dev->coherent_dma_mask);
>>> >
>>> >+ if (IS_ENABLED(CONFIG_OF) && dev->of_node)
>>> >+ of_dma_configure(&pdev->dev, dev->of_node);
>>> >+
>>> Would we hit the same issue if we are on non Device tree platforms like ACPI
>>> or platform device style itself?
>>>
>>
>> As far as I can see, yes.
>>
>>>
>>> > ret = platform_device_add_resources(pdev, res, nres);
>>> > if (ret)
>>> > goto err;
>>> >
>>>
>>> I think this is the side effect of commit
>>> 1dccb598df549d892b6450c261da54cdd7af44b4(arm64: simplify dma_get_ops)
>>>
>>
>> I agree, before that we would have hit:
>>
>> __generic_dma_ops() {
>> ..
>> else if (acpi_disabled)
>> return dma_ops;
>> ...
>> }
>>
>> with dma_ops being swiotlb_dma_ops from arm64_dma_init().
>>
>>
>> But this would not have saved us in the ACPI case, i.e. the result would
>> have been as with my suggested patch. Poking Arnd here to see if he has
>> any input.
>>
>>> None of the drivers call of_dma_configure() explicitly, which makes me feel
>>> that we are doing something wrong. TBH, this should be handled in more
>>> generic way rather than driver like this having an explicit call to
>>> of_dma_configure().
>>>
>>
>> I agree, trying to figure out if it should be inherited or something.
>
> I also agree. We need address it in a more generic way. I did a
> search for platform_device_add()/platform_device_register() in the
> kernel source code. I found a lot of them and many could be also
> doing DMA. Looks like it is still too early to assume every device is
> already getting dma_ops set through bus probe. Otherwise, many
> drivers are potentially broken by this assumption.
Any further comment on this topic? I added the linux-arm mailing list
which was missing from previous discussion.
Regards,
Leo
On Tue, Mar 8, 2016 at 11:52 AM, Li Yang <[email protected]> wrote:
> On Wed, Mar 2, 2016 at 4:59 PM, Li Yang <[email protected]> wrote:
>> On Mon, Feb 22, 2016 at 4:07 PM, Bjorn Andersson
>> <[email protected]> wrote:
>>> On Mon 22 Feb 02:03 PST 2016, Srinivas Kandagatla wrote:
>>>
>>>>
>>>>
>>>> On 22/02/16 05:32, Bjorn Andersson wrote:
>>>> >On certain platforms (e.g. ARM64) the dma_ops needs to be explicitly set
>>>> >to be able to do DMA allocations, so use the of_dma_configure() helper
>>>> >to populate the dma properties and assign an appropriate dma_ops.
[..]
>>>> None of the drivers call of_dma_configure() explicitly, which makes me feel
>>>> that we are doing something wrong. TBH, this should be handled in more
>>>> generic way rather than driver like this having an explicit call to
>>>> of_dma_configure().
>>>>
>>>
>>> I agree, trying to figure out if it should be inherited or something.
>>
>> I also agree. We need address it in a more generic way. I did a
>> search for platform_device_add()/platform_device_register() in the
>> kernel source code. I found a lot of them and many could be also
>> doing DMA. Looks like it is still too early to assume every device is
>> already getting dma_ops set through bus probe. Otherwise, many
>> drivers are potentially broken by this assumption.
>
> Any further comment on this topic? I added the linux-arm mailing list
> which was missing from previous discussion.
>
I had the chance to go through this with Arnd and the verdict is that
devices not described in DT should not do DMA (or allocate buffers for
doing DMA).
So I believe the solution is to fall back on Peter's description; the
chipidea driver is the core driver and the Qualcomm code should just
be a platform layer.
My suggestion is that we turn the chipidea core into a set of APIs
that can called by the platform specific pieces. That way we will have
the chipidea core be the device described in the DT.
I'll try to find some time to prototype this after Connect.
Regards,
Bjorn
On Tue, Mar 8, 2016 at 9:40 PM, Bjorn Andersson
<[email protected]> wrote:
> On Tue, Mar 8, 2016 at 11:52 AM, Li Yang <[email protected]> wrote:
>> On Wed, Mar 2, 2016 at 4:59 PM, Li Yang <[email protected]> wrote:
>>> On Mon, Feb 22, 2016 at 4:07 PM, Bjorn Andersson
>>> <[email protected]> wrote:
>>>> On Mon 22 Feb 02:03 PST 2016, Srinivas Kandagatla wrote:
>>>>
>>>>>
>>>>>
>>>>> On 22/02/16 05:32, Bjorn Andersson wrote:
>>>>> >On certain platforms (e.g. ARM64) the dma_ops needs to be explicitly set
>>>>> >to be able to do DMA allocations, so use the of_dma_configure() helper
>>>>> >to populate the dma properties and assign an appropriate dma_ops.
> [..]
>>>>> None of the drivers call of_dma_configure() explicitly, which makes me feel
>>>>> that we are doing something wrong. TBH, this should be handled in more
>>>>> generic way rather than driver like this having an explicit call to
>>>>> of_dma_configure().
>>>>>
>>>>
>>>> I agree, trying to figure out if it should be inherited or something.
>>>
>>> I also agree. We need address it in a more generic way. I did a
>>> search for platform_device_add()/platform_device_register() in the
>>> kernel source code. I found a lot of them and many could be also
>>> doing DMA. Looks like it is still too early to assume every device is
>>> already getting dma_ops set through bus probe. Otherwise, many
>>> drivers are potentially broken by this assumption.
>>
>> Any further comment on this topic? I added the linux-arm mailing list
>> which was missing from previous discussion.
>>
>
> I had the chance to go through this with Arnd and the verdict is that
> devices not described in DT should not do DMA (or allocate buffers for
> doing DMA).
>
> So I believe the solution is to fall back on Peter's description; the
> chipidea driver is the core driver and the Qualcomm code should just
> be a platform layer.
>
> My suggestion is that we turn the chipidea core into a set of APIs
> that can called by the platform specific pieces. That way we will have
> the chipidea core be the device described in the DT.
But like I said, this problem is not just existing for chipidea
driver. We already found that the dwc3 driver is also suffering from
the same issue. I don't know how many other drivers are impacted by
this change, but I suspect there will be some. A grep of
platform_device_add() in driver/ directory returns many possible
drivers to be impacted. As far as I know, the
drivers/net/ethernet/freescale/fman/mac.c is registering child
ethernet devices that definitely will do dma. If you want to do this
kind of rework to all these drivers, it will be a really big effort.
Regards,
Leo
On Wed, Mar 09, 2016 at 05:16:50PM -0600, Li Yang wrote:
> On Tue, Mar 8, 2016 at 9:40 PM, Bjorn Andersson
> <[email protected]> wrote:
> > On Tue, Mar 8, 2016 at 11:52 AM, Li Yang <[email protected]> wrote:
> >> On Wed, Mar 2, 2016 at 4:59 PM, Li Yang <[email protected]> wrote:
> >>> On Mon, Feb 22, 2016 at 4:07 PM, Bjorn Andersson
> >>> <[email protected]> wrote:
> >>>> On Mon 22 Feb 02:03 PST 2016, Srinivas Kandagatla wrote:
> >>>>
> >>>>>
> >>>>>
> >>>>> On 22/02/16 05:32, Bjorn Andersson wrote:
> >>>>> >On certain platforms (e.g. ARM64) the dma_ops needs to be explicitly set
> >>>>> >to be able to do DMA allocations, so use the of_dma_configure() helper
> >>>>> >to populate the dma properties and assign an appropriate dma_ops.
> > [..]
> >>>>> None of the drivers call of_dma_configure() explicitly, which makes me feel
> >>>>> that we are doing something wrong. TBH, this should be handled in more
> >>>>> generic way rather than driver like this having an explicit call to
> >>>>> of_dma_configure().
> >>>>>
> >>>>
> >>>> I agree, trying to figure out if it should be inherited or something.
> >>>
> >>> I also agree. We need address it in a more generic way. I did a
> >>> search for platform_device_add()/platform_device_register() in the
> >>> kernel source code. I found a lot of them and many could be also
> >>> doing DMA. Looks like it is still too early to assume every device is
> >>> already getting dma_ops set through bus probe. Otherwise, many
> >>> drivers are potentially broken by this assumption.
> >>
> >> Any further comment on this topic? I added the linux-arm mailing list
> >> which was missing from previous discussion.
> >>
> >
> > I had the chance to go through this with Arnd and the verdict is that
> > devices not described in DT should not do DMA (or allocate buffers for
> > doing DMA).
> >
> > So I believe the solution is to fall back on Peter's description; the
> > chipidea driver is the core driver and the Qualcomm code should just
> > be a platform layer.
> >
> > My suggestion is that we turn the chipidea core into a set of APIs
> > that can called by the platform specific pieces. That way we will have
> > the chipidea core be the device described in the DT.
>
> But like I said, this problem is not just existing for chipidea
> driver. We already found that the dwc3 driver is also suffering from
> the same issue. I don't know how many other drivers are impacted by
> this change, but I suspect there will be some. A grep of
> platform_device_add() in driver/ directory returns many possible
> drivers to be impacted. As far as I know, the
> drivers/net/ethernet/freescale/fman/mac.c is registering child
> ethernet devices that definitely will do dma. If you want to do this
> kind of rework to all these drivers, it will be a really big effort.
>
+1
Yes, I think this DMA things should be covered by driver core too.
--
Best Regards,
Peter Chen
On Monday 14 March 2016 18:51:08 Peter Chen wrote:
> On Wed, Mar 09, 2016 at 05:16:50PM -0600, Li Yang wrote:
> > On Tue, Mar 8, 2016 at 9:40 PM, Bjorn Andersson
> > <[email protected]> wrote:
> > > On Tue, Mar 8, 2016 at 11:52 AM, Li Yang <[email protected]> wrote:
> > >> On Wed, Mar 2, 2016 at 4:59 PM, Li Yang <[email protected]> wrote:
> > >>> On Mon, Feb 22, 2016 at 4:07 PM, Bjorn Andersson
> > >>> <[email protected]> wrote:
> > >>>> On Mon 22 Feb 02:03 PST 2016, Srinivas Kandagatla wrote:
> > >
> > > I had the chance to go through this with Arnd and the verdict is that
> > > devices not described in DT should not do DMA (or allocate buffers for
> > > doing DMA).
> > >
> > > So I believe the solution is to fall back on Peter's description; the
> > > chipidea driver is the core driver and the Qualcomm code should just
> > > be a platform layer.
> > >
> > > My suggestion is that we turn the chipidea core into a set of APIs
> > > that can called by the platform specific pieces. That way we will have
> > > the chipidea core be the device described in the DT.
> >
> > But like I said, this problem is not just existing for chipidea
> > driver. We already found that the dwc3 driver is also suffering from
> > the same issue. I don't know how many other drivers are impacted by
> > this change, but I suspect there will be some. A grep of
> > platform_device_add() in driver/ directory returns many possible
> > drivers to be impacted. As far as I know, the
> > drivers/net/ethernet/freescale/fman/mac.c is registering child
> > ethernet devices that definitely will do dma. If you want to do this
> > kind of rework to all these drivers, it will be a really big effort.
> >
>
> +1
>
> Yes, I think this DMA things should be covered by driver core too.
>
I don't think it's a very widespread problem, there are only very few
developers that intentionally use this method, and some use the
platform_device_register_full() call to create a device with a known
mask, which is generally ok for the limited case where the driver
is only ever going to run on a single platform, but not in the
more general case that of_dma_configure is designed to handle.
I think we should fix the drivers to consistently use the device
that was created by the platform (DT or ACPI or board file)
to pass that into the DMA API, anything else will just cause
more subtle bugs.
Arnd
On Thu, Mar 17, 2016 at 04:52:55PM +0100, Arnd Bergmann wrote:
> On Monday 14 March 2016 18:51:08 Peter Chen wrote:
> > On Wed, Mar 09, 2016 at 05:16:50PM -0600, Li Yang wrote:
> > > On Tue, Mar 8, 2016 at 9:40 PM, Bjorn Andersson
> > > <[email protected]> wrote:
> > > > On Tue, Mar 8, 2016 at 11:52 AM, Li Yang <[email protected]> wrote:
> > > >> On Wed, Mar 2, 2016 at 4:59 PM, Li Yang <[email protected]> wrote:
> > > >>> On Mon, Feb 22, 2016 at 4:07 PM, Bjorn Andersson
> > > >>> <[email protected]> wrote:
> > > >>>> On Mon 22 Feb 02:03 PST 2016, Srinivas Kandagatla wrote:
> > > >
> > > > I had the chance to go through this with Arnd and the verdict is that
> > > > devices not described in DT should not do DMA (or allocate buffers for
> > > > doing DMA).
> > > >
> > > > So I believe the solution is to fall back on Peter's description; the
> > > > chipidea driver is the core driver and the Qualcomm code should just
> > > > be a platform layer.
> > > >
> > > > My suggestion is that we turn the chipidea core into a set of APIs
> > > > that can called by the platform specific pieces. That way we will have
> > > > the chipidea core be the device described in the DT.
> > >
> > > But like I said, this problem is not just existing for chipidea
> > > driver. We already found that the dwc3 driver is also suffering from
> > > the same issue. I don't know how many other drivers are impacted by
> > > this change, but I suspect there will be some. A grep of
> > > platform_device_add() in driver/ directory returns many possible
> > > drivers to be impacted. As far as I know, the
> > > drivers/net/ethernet/freescale/fman/mac.c is registering child
> > > ethernet devices that definitely will do dma. If you want to do this
> > > kind of rework to all these drivers, it will be a really big effort.
> > >
> >
> > +1
> >
> > Yes, I think this DMA things should be covered by driver core too.
> >
>
> I don't think it's a very widespread problem, there are only very few
> developers that intentionally use this method, and some use the
> platform_device_register_full() call to create a device with a known
> mask, which is generally ok for the limited case where the driver
> is only ever going to run on a single platform, but not in the
> more general case that of_dma_configure is designed to handle.
Even only for qualcomm platforms, it may be possible have different
DMA masks at ARM64 platforms, so we may can't use a fixed value
at glue layer driver. So, using of_dma_configure is suitable choice
for DT platforms for this case, right?
>
> I think we should fix the drivers to consistently use the device
> that was created by the platform (DT or ACPI or board file)
> to pass that into the DMA API, anything else will just cause
> more subtle bugs.
>
Although I don't know what kinds of bugs it may have, it may be
met before, otherwise, why most of platform drivers need to call
dma_set_coherent_mask or dma_coerce_mask_and_coherent explicitly
--
Best Regards,
Peter Chen
> -----Original Message-----
> From: Peter Chen [mailto:[email protected]]
> Sent: Friday, March 18, 2016 7:24 AM
> To: Arnd Bergmann <[email protected]>
> Cc: Li Yang <[email protected]>; Bjorn Andersson <[email protected]>;
> Peter Chen <[email protected]>; Greg Kroah-Hartman
> <[email protected]>; Rajesh Bhagat <[email protected]>; linux-
> [email protected]; lkml <[email protected]>; Srinivas Kandagatla
> <[email protected]>; linux-arm-msm <linux-arm-
> [email protected]>; [email protected]
> Subject: Re: [PATCH] usb: chipidea: Configure DMA properties and ops from DT
>
> On Thu, Mar 17, 2016 at 04:52:55PM +0100, Arnd Bergmann wrote:
> > On Monday 14 March 2016 18:51:08 Peter Chen wrote:
> > > On Wed, Mar 09, 2016 at 05:16:50PM -0600, Li Yang wrote:
> > > > On Tue, Mar 8, 2016 at 9:40 PM, Bjorn Andersson
> > > > <[email protected]> wrote:
> > > > > On Tue, Mar 8, 2016 at 11:52 AM, Li Yang <[email protected]> wrote:
> > > > >> On Wed, Mar 2, 2016 at 4:59 PM, Li Yang <[email protected]> wrote:
> > > > >>> On Mon, Feb 22, 2016 at 4:07 PM, Bjorn Andersson
> > > > >>> <[email protected]> wrote:
> > > > >>>> On Mon 22 Feb 02:03 PST 2016, Srinivas Kandagatla wrote:
> > > > >
> > > > > I had the chance to go through this with Arnd and the verdict is
> > > > > that devices not described in DT should not do DMA (or allocate
> > > > > buffers for doing DMA).
> > > > >
> > > > > So I believe the solution is to fall back on Peter's
> > > > > description; the chipidea driver is the core driver and the
> > > > > Qualcomm code should just be a platform layer.
> > > > >
> > > > > My suggestion is that we turn the chipidea core into a set of
> > > > > APIs that can called by the platform specific pieces. That way
> > > > > we will have the chipidea core be the device described in the DT.
> > > >
> > > > But like I said, this problem is not just existing for chipidea
> > > > driver. We already found that the dwc3 driver is also suffering
> > > > from the same issue. I don't know how many other drivers are
> > > > impacted by this change, but I suspect there will be some. A grep
> > > > of
> > > > platform_device_add() in driver/ directory returns many possible
> > > > drivers to be impacted. As far as I know, the
> > > > drivers/net/ethernet/freescale/fman/mac.c is registering child
> > > > ethernet devices that definitely will do dma. If you want to do this
> > > > kind of rework to all these drivers, it will be a really big effort.
> > > >
> > >
> > > +1
> > >
> > > Yes, I think this DMA things should be covered by driver core too.
> > >
> >
> > I don't think it's a very widespread problem, there are only very few
> > developers that intentionally use this method, and some use the
> > platform_device_register_full() call to create a device with a known
> > mask, which is generally ok for the limited case where the driver is
> > only ever going to run on a single platform, but not in the more
> > general case that of_dma_configure is designed to handle.
>
> Even only for qualcomm platforms, it may be possible have different DMA masks at
> ARM64 platforms, so we may can't use a fixed value at glue layer driver. So, using
> of_dma_configure is suitable choice for DT platforms for this case, right?
>
> >
> > I think we should fix the drivers to consistently use the device that
> > was created by the platform (DT or ACPI or board file) to pass that
> > into the DMA API, anything else will just cause more subtle bugs.
> >
>
> Although I don't know what kinds of bugs it may have, it may be met before,
> otherwise, why most of platform drivers need to call dma_set_coherent_mask or
> dma_coerce_mask_and_coherent explicitly
>
> --
>
> Best Regards,
> Peter Chen
Though chipidea platform drivers are calling functions mentioned by you i.e. dma_set_coherent_mask
or dma_coerce_mask_and_coherent explicity e.g. in file drivers/usb/chipidea/ci_hdrc_imx.c.
Still the mentioned error is coming while calling ci_hdrc_add_device which lies in chipidea/core.c. similar is
the case with DWC3 driver.
Best Regards,
Rajesh Bhagat
On Friday 18 March 2016 09:54:14 Peter Chen wrote:
> >
> > I don't think it's a very widespread problem, there are only very few
> > developers that intentionally use this method, and some use the
> > platform_device_register_full() call to create a device with a known
> > mask, which is generally ok for the limited case where the driver
> > is only ever going to run on a single platform, but not in the
> > more general case that of_dma_configure is designed to handle.
>
> Even only for qualcomm platforms, it may be possible have different
> DMA masks at ARM64 platforms, so we may can't use a fixed value
> at glue layer driver. So, using of_dma_configure is suitable choice
> for DT platforms for this case, right?
Yes.
> > I think we should fix the drivers to consistently use the device
> > that was created by the platform (DT or ACPI or board file)
> > to pass that into the DMA API, anything else will just cause
> > more subtle bugs.
> >
>
> Although I don't know what kinds of bugs it may have, it may be
> met before, otherwise, why most of platform drivers need to call
> dma_set_coherent_mask or dma_coerce_mask_and_coherent explicitly
Any driver that wants to do 64-bit addressing on DMA should call
dma_set_mask()/dma_set_coherent_mask() on its device and check the
return code.
No driver should call dma_coerce_mask_and_coherent() on its own
device, it's basically always a bug and we named the function
to make that more obvious. The problem with dma_coerce_mask_and_coherent()
is that it just overrides whatever the platform knows about the
device when the driver thinks it knows better.
The reason for having those calls in a lot of drivers is that
traditionally, ARM platforms booting with DT did not set up any DMA
mask and the drivers worked around it by manually setting up a mask
that happened to work for them (almost all 32-bit ARM devices need
a 32-bit mask without coherency or offset or iommu, so that's easy).
We now call of_dma_configure() for all devices that get probed from
DT, so we should be removing those calls.
Arnd
On Fri, Mar 18, 2016 at 09:54:14AM +0800, Peter Chen wrote:
> Although I don't know what kinds of bugs it may have, it may be
> met before, otherwise, why most of platform drivers need to call
> dma_set_coherent_mask or dma_coerce_mask_and_coherent explicitly
See Documentation/DMA-API.txt, specifically the section starting
Part Ic - DMA addressing limitations
------------------------------------
and also Documentation/DMA-API-HOWTO.txt, the section on
DMA addressing limitations
which provides further information.
Drivers using DMA should be using dma_set_mask_and_coherent() _or_
one of dma_set_mask() and dma_set_coherent_mask() depending on which
types of DMA they wish to perform. Drivers should not use
dma_coerce_mask_and_coherent() except in exceptional circumstances:
that function is more a marker that they or some bus/platform code
is doing something wrong.
--
RMK's Patch system: http://www.arm.linux.org.uk/developer/patches/
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.
On Tue, Mar 08, 2016 at 07:40:08PM -0800, Bjorn Andersson wrote:
> On Tue, Mar 8, 2016 at 11:52 AM, Li Yang <[email protected]> wrote:
> > On Wed, Mar 2, 2016 at 4:59 PM, Li Yang <[email protected]> wrote:
> >> On Mon, Feb 22, 2016 at 4:07 PM, Bjorn Andersson
> >> <[email protected]> wrote:
> >>> On Mon 22 Feb 02:03 PST 2016, Srinivas Kandagatla wrote:
> >>>
> >>>>
> >>>>
> >>>> On 22/02/16 05:32, Bjorn Andersson wrote:
> >>>> >On certain platforms (e.g. ARM64) the dma_ops needs to be explicitly set
> >>>> >to be able to do DMA allocations, so use the of_dma_configure() helper
> >>>> >to populate the dma properties and assign an appropriate dma_ops.
> [..]
> >>>> None of the drivers call of_dma_configure() explicitly, which makes me feel
> >>>> that we are doing something wrong. TBH, this should be handled in more
> >>>> generic way rather than driver like this having an explicit call to
> >>>> of_dma_configure().
> >>>>
> >>>
>
> I had the chance to go through this with Arnd and the verdict is that
> devices not described in DT should not do DMA (or allocate buffers for
> doing DMA).
>
> So I believe the solution is to fall back on Peter's description; the
> chipidea driver is the core driver and the Qualcomm code should just
> be a platform layer.
>
> My suggestion is that we turn the chipidea core into a set of APIs
> that can called by the platform specific pieces. That way we will have
> the chipidea core be the device described in the DT.
>
Hi Bjorn,
After reading the DMA documentation Russell supplied and several related
DMA APIs, would you please try below patch on your ARM64 platform?
Since the core device has no device node at all, I don't know why
your patch can work, or am I missing something?
>From bcf7eaf694d29fb7557a9406fb6c89213216069c Mon Sep 17 00:00:00 2001
From: Peter Chen <[email protected]>
Date: Fri, 25 Mar 2016 11:54:21 +0800
Subject: [PATCH 1/1] usb: chipidea: add DMA mask configuration API
Signed-off-by: Peter Chen <[email protected]>
---
drivers/usb/chipidea/ci_hdrc_msm.c | 6 ++++++
drivers/usb/chipidea/core.c | 25 +++++++++++++++++++++++++
include/linux/usb/chipidea.h | 2 ++
3 files changed, 33 insertions(+)
diff --git a/drivers/usb/chipidea/ci_hdrc_msm.c b/drivers/usb/chipidea/ci_hdrc_msm.c
index 3889809..43ceb38 100644
--- a/drivers/usb/chipidea/ci_hdrc_msm.c
+++ b/drivers/usb/chipidea/ci_hdrc_msm.c
@@ -5,6 +5,7 @@
* only version 2 as published by the Free Software Foundation.
*/
+#include <linux/dma-mapping.h>
#include <linux/module.h>
#include <linux/platform_device.h>
#include <linux/pm_runtime.h>
@@ -56,6 +57,7 @@ static int ci_hdrc_msm_probe(struct platform_device *pdev)
{
struct platform_device *plat_ci;
struct usb_phy *phy;
+ int ret;
dev_dbg(&pdev->dev, "ci_hdrc_msm_probe\n");
@@ -70,6 +72,10 @@ static int ci_hdrc_msm_probe(struct platform_device *pdev)
ci_hdrc_msm_platdata.usb_phy = phy;
+ ret = ci_hdrc_set_dma_mask(&pdev->dev, DMA_BIT_MASK(64));
+ if (ret)
+ return ret;
+
plat_ci = ci_hdrc_add_device(&pdev->dev,
pdev->resource, pdev->num_resources,
&ci_hdrc_msm_platdata);
diff --git a/drivers/usb/chipidea/core.c b/drivers/usb/chipidea/core.c
index 69426e6..b8ca5e3 100644
--- a/drivers/usb/chipidea/core.c
+++ b/drivers/usb/chipidea/core.c
@@ -62,6 +62,7 @@
#include <linux/usb/chipidea.h>
#include <linux/usb/of.h>
#include <linux/of.h>
+#include <linux/of_device.h>
#include <linux/phy.h>
#include <linux/regulator/consumer.h>
#include <linux/usb/ehci_def.h>
@@ -811,6 +812,30 @@ static void ci_extcon_unregister(struct ci_hdrc *ci)
static DEFINE_IDA(ci_ida);
+/**
+ * ci_hdrc_set_dma_mask
+ *
+ * Set dma mask and coherent dma mask for glue layer device, and the core
+ * device will inherit these values. If the 'dma-ranges' is specified at
+ * DT, it will use this value for both dma mask and coherent dma mask.
+ *
+ * @dev: a pointer to the device struct of glue layer device
+ * @ci_coherent_dma_mask: the mask for both dma_mask and cohrent_dma_mask
+ */
+int ci_hdrc_set_dma_mask(struct device *dev, u64 ci_coherent_dma_mask)
+{
+ int ret = dma_set_mask_and_coherent(dev, ci_coherent_dma_mask);
+ if (ret) {
+ dev_err(dev, "dma_set_mask_and_coherent fails\n");
+ return ret;
+ }
+
+ if (dev_of_node(dev))
+ of_dma_configure(dev, dev->of_node);
+
+ return ret;
+}
+
struct platform_device *ci_hdrc_add_device(struct device *dev,
struct resource *res, int nres,
struct ci_hdrc_platform_data *platdata)
diff --git a/include/linux/usb/chipidea.h b/include/linux/usb/chipidea.h
index 5dd75fa..8649930 100644
--- a/include/linux/usb/chipidea.h
+++ b/include/linux/usb/chipidea.h
@@ -84,4 +84,6 @@ struct platform_device *ci_hdrc_add_device(struct device *dev,
/* Remove ci hdrc device */
void ci_hdrc_remove_device(struct platform_device *pdev);
+int ci_hdrc_set_dma_mask(struct device *dev, u64 ci_coherent_dma_mask);
+
#endif
--
Best Regards,
Peter Chen
Hi Peter,
[auto build test ERROR on v4.5-rc7]
[also build test ERROR on next-20160324]
[if your patch is applied to the wrong git tree, please drop us a note to help improving the system]
url: https://github.com/0day-ci/linux/commits/Peter-Chen/usb-chipidea-add-DMA-mask-configuration-API/20160325-120535
config: i386-randconfig-s1-201612 (attached as .config)
reproduce:
# save the attached .config to linux build tree
make ARCH=i386
All errors (new ones prefixed by >>):
>> ERROR: "ci_hdrc_set_dma_mask" [drivers/usb/chipidea/ci_hdrc_msm.ko] undefined!
---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation