Message ID | 1390845177-2626-1-git-send-email-lauraa@codeaurora.org (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
On Mon, Jan 27, 2014 at 09:52:57AM -0800, Laura Abbott wrote: > The dma_mask for a device structure is a pointer. This pointer > needs to be set up before the dma mask can actually be set. Most > frameworks in the kernel take care of setting this up properly but > platform devices that don't follow a regular bus structure may not > ever have this set. As a result, checks such as dma_capable will > always return false on a raw platform device and dma_set_mask will > always return -EIO. Fix this by adding a dma_mask in the > platform_device archdata and setting it to be the dma_mask. Devices > used in other frameworks can change this as needed. > > Cc: Will Deacon <will.deacon@arm.com> > Cc: Catalin Marinas <catalin.marinas@arm.com> > Suggested-by: Kumar Gala <galak@codeaurora.org> > Signed-off-by: Laura Abbott <lauraa@codeaurora.org> Hello, there is another non-platform dependant approach available, that might be worth to evaluate: http://mid.gmane.org/1390817152-30898-1-git-send-email-ydroneaud@opteya.com > --- > arch/arm64/include/asm/device.h | 1 + > arch/arm64/kernel/setup.c | 7 +++++++ > 2 files changed, 8 insertions(+), 0 deletions(-) > > diff --git a/arch/arm64/include/asm/device.h b/arch/arm64/include/asm/device.h > index cf98b36..209d40c 100644 > --- a/arch/arm64/include/asm/device.h > +++ b/arch/arm64/include/asm/device.h > @@ -24,6 +24,7 @@ struct dev_archdata { > }; > > struct pdev_archdata { > + u64 dma_mask; > }; > > #endif > diff --git a/arch/arm64/kernel/setup.c b/arch/arm64/kernel/setup.c > index bd9bbd0..f164347 100644 > --- a/arch/arm64/kernel/setup.c > +++ b/arch/arm64/kernel/setup.c > @@ -41,6 +41,7 @@ > #include <linux/memblock.h> > #include <linux/of_fdt.h> > #include <linux/of_platform.h> > +#include <linux/dma-mapping.h> > > #include <asm/cputype.h> > #include <asm/elf.h> > @@ -337,3 +338,9 @@ const struct seq_operations cpuinfo_op = { > .stop = c_stop, > .show = c_show > }; > + > +void arch_setup_pdev_archdata(struct platform_device *pdev) > +{ > + pdev->archdata.dma_mask = DMA_BIT_MASK(32); > + pdev->dev.dma_mask = &pdev->archdata.dma_mask; > +} Is it save to assume a default of DMA_BIT_MASK(32)? Best regards Uwe
On 1/27/2014 10:18 AM, Uwe Kleine-König wrote: > On Mon, Jan 27, 2014 at 09:52:57AM -0800, Laura Abbott wrote: >> The dma_mask for a device structure is a pointer. This pointer >> needs to be set up before the dma mask can actually be set. Most >> frameworks in the kernel take care of setting this up properly but >> platform devices that don't follow a regular bus structure may not >> ever have this set. As a result, checks such as dma_capable will >> always return false on a raw platform device and dma_set_mask will >> always return -EIO. Fix this by adding a dma_mask in the >> platform_device archdata and setting it to be the dma_mask. Devices >> used in other frameworks can change this as needed. >> >> Cc: Will Deacon <will.deacon@arm.com> >> Cc: Catalin Marinas <catalin.marinas@arm.com> >> Suggested-by: Kumar Gala <galak@codeaurora.org> >> Signed-off-by: Laura Abbott <lauraa@codeaurora.org> > Hello, > > there is another non-platform dependant approach available, that might > be worth to evaluate: > > http://mid.gmane.org/1390817152-30898-1-git-send-email-ydroneaud@opteya.com > That covers dynamically allocated devices but it doesn't look like it covers devices setup with just platform_device_register like arch_setup_pdev_archdata does. >> --- >> arch/arm64/include/asm/device.h | 1 + >> arch/arm64/kernel/setup.c | 7 +++++++ >> 2 files changed, 8 insertions(+), 0 deletions(-) >> >> diff --git a/arch/arm64/include/asm/device.h b/arch/arm64/include/asm/device.h >> index cf98b36..209d40c 100644 >> --- a/arch/arm64/include/asm/device.h >> +++ b/arch/arm64/include/asm/device.h >> @@ -24,6 +24,7 @@ struct dev_archdata { >> }; >> >> struct pdev_archdata { >> + u64 dma_mask; >> }; >> >> #endif >> diff --git a/arch/arm64/kernel/setup.c b/arch/arm64/kernel/setup.c >> index bd9bbd0..f164347 100644 >> --- a/arch/arm64/kernel/setup.c >> +++ b/arch/arm64/kernel/setup.c >> @@ -41,6 +41,7 @@ >> #include <linux/memblock.h> >> #include <linux/of_fdt.h> >> #include <linux/of_platform.h> >> +#include <linux/dma-mapping.h> >> >> #include <asm/cputype.h> >> #include <asm/elf.h> >> @@ -337,3 +338,9 @@ const struct seq_operations cpuinfo_op = { >> .stop = c_stop, >> .show = c_show >> }; >> + >> +void arch_setup_pdev_archdata(struct platform_device *pdev) >> +{ >> + pdev->archdata.dma_mask = DMA_BIT_MASK(32); >> + pdev->dev.dma_mask = &pdev->archdata.dma_mask; >> +} > Is it save to assume a default of DMA_BIT_MASK(32)? > This seemed like a reasonable default and matches what powerpc does. Any device who wants to really guarantee a DMA mask should be setting the dma mask explicitly and not relying on a default. > Best regards > Uwe > Thanks, Laura
On Mon, Jan 27, 2014 at 09:52:57AM -0800, Laura Abbott wrote: > The dma_mask for a device structure is a pointer. This pointer > needs to be set up before the dma mask can actually be set. Most > frameworks in the kernel take care of setting this up properly but > platform devices that don't follow a regular bus structure may not > ever have this set. As a result, checks such as dma_capable will > always return false on a raw platform device and dma_set_mask will > always return -EIO. Fix this by adding a dma_mask in the > platform_device archdata and setting it to be the dma_mask. Devices > used in other frameworks can change this as needed. You shouldn't need to do this. I went through a lot of the drivers we currently have, fixing them up in various manners. The basic rules for this stuff are: - It is the responsibility of the code creating the device to set a reasonable default for the dma mask according to the bus and whether DMA is supportable. - It is the responsibility of the driver _always_ to make a call to dma_set_mask() and/or dma_set_coherent_mask() according to the driver's needs if the driver is going to be using DMA. As a work-around for the buggy situation we have in the kernel with DT, various buggy workarounds have been incorporated into drivers which involve writing directly to the DMA masks, and other such games. None of that is necessary when the dma_coerce_*() functions are used - but these are a stop-gap until the DT code gets fixed. The real answer here is to make DT conform to the first point above and not add yet another different hack to the kernel.
Hi, Le lundi 27 janvier 2014 à 19:18 +0100, Uwe Kleine-König a écrit : > On Mon, Jan 27, 2014 at 09:52:57AM -0800, Laura Abbott wrote: > > The dma_mask for a device structure is a pointer. This pointer > > needs to be set up before the dma mask can actually be set. Most > > frameworks in the kernel take care of setting this up properly but > > platform devices that don't follow a regular bus structure may not > > ever have this set. As a result, checks such as dma_capable will > > always return false on a raw platform device and dma_set_mask will > > always return -EIO. Fix this by adding a dma_mask in the > > platform_device archdata and setting it to be the dma_mask. Devices > > used in other frameworks can change this as needed. > > > > Cc: Will Deacon <will.deacon@arm.com> > > Cc: Catalin Marinas <catalin.marinas@arm.com> > > Suggested-by: Kumar Gala <galak@codeaurora.org> > > Signed-off-by: Laura Abbott <lauraa@codeaurora.org> > Hello, > > there is another non-platform dependant approach available, that might > be worth to evaluate: > > http://mid.gmane.org/1390817152-30898-1-git-send-email-ydroneaud@opteya.com > ARM, even AAAAARGH64 [1], doesn't need a special treatement regarding the infamous dma_mask pointer. So perhaps my solution is better. This solution (adding dma_mask in pdev_archdata) is already in use in powerpc architecture. See arch/powerpc/kernel/setup-common.c The advantage of this solution is that it makes a dma_mask placeholder available to statically allocated platform_device struct, while mine only address the problem for platform_device struct allocated with platform_device_alloc(). A possible drawback of adding dma_mask in pdev_archdata and setting the pointer in arch_setup_pdev_archdata() may be that a dma_mask setup prior calling platform_device_register() got overwritten. For example at91rm9200_usbh_device in arch/arm/mach-at91/at91rm9200_devices.c (it's the first occurence returned by cscope). So I guess there's not yet a perfect solution (add dma_mask to struct platform_device instead of hidding it either in struct pdev_archdata or struct platform_object ?). BTW, I've started to convert some drivers just to try my first option. I'm also considering using dma_set_mask_and_coherent() in platform_device_register_full() and drivers using platform_device_alloc(). But I'm not sure about it : will it break something with the additionals check on the mask, just as the failure reported in the message ("dma_set_mask will always return -EIO.") ? [1] http://lkml.kernel.org/r/20140124154002.GF31570@twins.programming.kicks-ass.net Regards.
On Mon, Jan 27, 2014 at 09:25:31PM +0100, Yann Droneaud wrote: > ARM, even AAAAARGH64 [1], doesn't need a special treatement regarding > the infamous dma_mask pointer. So perhaps my solution is better. > > This solution (adding dma_mask in pdev_archdata) is already in use in > powerpc architecture. See arch/powerpc/kernel/setup-common.c > > The advantage of this solution is that it makes a dma_mask placeholder > available to statically allocated platform_device struct, while mine > only address the problem for platform_device struct allocated with > platform_device_alloc(). As I've already said in this thread, the basic problem comes from DT's platform device creation. It's the responsibility of the device creator to set the dma_mask pointer appropriately, and DT doesn't do that. So, DT needs to be fixed rather than everyone introducing their own workarounds for this. > I'm also considering using dma_set_mask_and_coherent() in > platform_device_register_full() and drivers using > platform_device_alloc(). As the one who introduced dma_set_mask_and_coherent, consider this a strong NAK on that. The reason is dma_set_mask_and_coherent() is for drivers to set their requirements, not for the bus requirements to be set in the first place. It also means that drivers which need no DMA support are subjected to DMA restrictions (in that dma_set_mask_and_coherent can error out if the platform can't support the DMA mask.)
On 1/27/2014 11:31 AM, Russell King - ARM Linux wrote: > On Mon, Jan 27, 2014 at 09:52:57AM -0800, Laura Abbott wrote: >> The dma_mask for a device structure is a pointer. This pointer >> needs to be set up before the dma mask can actually be set. Most >> frameworks in the kernel take care of setting this up properly but >> platform devices that don't follow a regular bus structure may not >> ever have this set. As a result, checks such as dma_capable will >> always return false on a raw platform device and dma_set_mask will >> always return -EIO. Fix this by adding a dma_mask in the >> platform_device archdata and setting it to be the dma_mask. Devices >> used in other frameworks can change this as needed. > > You shouldn't need to do this. I went through a lot of the drivers we > currently have, fixing them up in various manners. The basic rules > for this stuff are: > > - It is the responsibility of the code creating the device to set a > reasonable default for the dma mask according to the bus and whether > DMA is supportable. > > - It is the responsibility of the driver _always_ to make a call to > dma_set_mask() and/or dma_set_coherent_mask() according to the > driver's needs if the driver is going to be using DMA. > > As a work-around for the buggy situation we have in the kernel with DT, > various buggy workarounds have been incorporated into drivers which > involve writing directly to the DMA masks, and other such games. None > of that is necessary when the dma_coerce_*() functions are used - but > these are a stop-gap until the DT code gets fixed. > > The real answer here is to make DT conform to the first point above > and not add yet another different hack to the kernel. > powerpc ran into this exact problem before and fixed it using this method (a77ce8167cc1d0370fcb1d79b367d62e050cb2b0 "driver core: Add ability for arch code to setup pdev_archdata" and 314b02f503c2c219fde0fcf6f086fda415f8a847 "powerpc: implement arch_setup_pdev_archdata") so there is at least some precedent for this method. Are there patches/discussion somewhere else on what a proper solution would be in the DT? Thanks, Laura
On Tue, Jan 28, 2014 at 01:42:07AM +0000, Laura Abbott wrote: > On 1/27/2014 11:31 AM, Russell King - ARM Linux wrote: > > On Mon, Jan 27, 2014 at 09:52:57AM -0800, Laura Abbott wrote: > >> The dma_mask for a device structure is a pointer. This pointer > >> needs to be set up before the dma mask can actually be set. Most > >> frameworks in the kernel take care of setting this up properly but > >> platform devices that don't follow a regular bus structure may not > >> ever have this set. As a result, checks such as dma_capable will > >> always return false on a raw platform device and dma_set_mask will > >> always return -EIO. Fix this by adding a dma_mask in the > >> platform_device archdata and setting it to be the dma_mask. Devices > >> used in other frameworks can change this as needed. > > > > You shouldn't need to do this. I went through a lot of the drivers we > > currently have, fixing them up in various manners. The basic rules > > for this stuff are: > > > > - It is the responsibility of the code creating the device to set a > > reasonable default for the dma mask according to the bus and whether > > DMA is supportable. > > > > - It is the responsibility of the driver _always_ to make a call to > > dma_set_mask() and/or dma_set_coherent_mask() according to the > > driver's needs if the driver is going to be using DMA. > > > > As a work-around for the buggy situation we have in the kernel with DT, > > various buggy workarounds have been incorporated into drivers which > > involve writing directly to the DMA masks, and other such games. None > > of that is necessary when the dma_coerce_*() functions are used - but > > these are a stop-gap until the DT code gets fixed. > > > > The real answer here is to make DT conform to the first point above > > and not add yet another different hack to the kernel. > > > > powerpc ran into this exact problem before and fixed it using this > method (a77ce8167cc1d0370fcb1d79b367d62e050cb2b0 > "driver core: Add ability for arch code to setup pdev_archdata" and > 314b02f503c2c219fde0fcf6f086fda415f8a847 "powerpc: implement > arch_setup_pdev_archdata") so there is at least some precedent for this > method. > > Are there patches/discussion somewhere else on what a proper solution > would be in the DT? I guess this comes under the system topology discussion we briefly touched at the kernel summit. Dave Martin (cc'ed) is looking into this now, so at some point we'll have some longer discussions on the list.
On Mon, Jan 27, 2014 at 05:42:07PM -0800, Laura Abbott wrote: > On 1/27/2014 11:31 AM, Russell King - ARM Linux wrote: >> On Mon, Jan 27, 2014 at 09:52:57AM -0800, Laura Abbott wrote: >>> The dma_mask for a device structure is a pointer. This pointer >>> needs to be set up before the dma mask can actually be set. Most >>> frameworks in the kernel take care of setting this up properly but >>> platform devices that don't follow a regular bus structure may not >>> ever have this set. As a result, checks such as dma_capable will >>> always return false on a raw platform device and dma_set_mask will >>> always return -EIO. Fix this by adding a dma_mask in the >>> platform_device archdata and setting it to be the dma_mask. Devices >>> used in other frameworks can change this as needed. >> >> You shouldn't need to do this. I went through a lot of the drivers we >> currently have, fixing them up in various manners. The basic rules >> for this stuff are: >> >> - It is the responsibility of the code creating the device to set a >> reasonable default for the dma mask according to the bus and whether >> DMA is supportable. >> >> - It is the responsibility of the driver _always_ to make a call to >> dma_set_mask() and/or dma_set_coherent_mask() according to the >> driver's needs if the driver is going to be using DMA. >> >> As a work-around for the buggy situation we have in the kernel with DT, >> various buggy workarounds have been incorporated into drivers which >> involve writing directly to the DMA masks, and other such games. None >> of that is necessary when the dma_coerce_*() functions are used - but >> these are a stop-gap until the DT code gets fixed. >> >> The real answer here is to make DT conform to the first point above >> and not add yet another different hack to the kernel. >> > > powerpc ran into this exact problem before and fixed it using this > method (a77ce8167cc1d0370fcb1d79b367d62e050cb2b0 > "driver core: Add ability for arch code to setup pdev_archdata" and > 314b02f503c2c219fde0fcf6f086fda415f8a847 "powerpc: implement > arch_setup_pdev_archdata") so there is at least some precedent for this > method. The result being that all platform devices get their DMA masks setup whether they like it or not - what they did was nothing more than a work-around the problem. That doesn't get away from what I stated above, which is what's expected of drivers, and by drivers. Drivers which perform DMA _must_ without fail call the appropriate DMA mask setting functions. Code declaring devices must set the DMA masks to a reasonable default if the driver is to perform DMA with that device. As the ARM architecture moves forward, and we start seeing 64-bit DMA controllers, this becomes extremely important, because the first part of that is part of the negotiation whether memory outside of the 4GB DMA range can be passed to the DMA engine. We're also going to hit problems where people have been lazy, and don't map/unmap/allocate DMA memory against the DMA engine device, but against the DMA client device. ASoC is going to be one such area of pain because the DMA engine backend that was merged totally disregarded this point which was present in my backend.
diff --git a/arch/arm64/include/asm/device.h b/arch/arm64/include/asm/device.h index cf98b36..209d40c 100644 --- a/arch/arm64/include/asm/device.h +++ b/arch/arm64/include/asm/device.h @@ -24,6 +24,7 @@ struct dev_archdata { }; struct pdev_archdata { + u64 dma_mask; }; #endif diff --git a/arch/arm64/kernel/setup.c b/arch/arm64/kernel/setup.c index bd9bbd0..f164347 100644 --- a/arch/arm64/kernel/setup.c +++ b/arch/arm64/kernel/setup.c @@ -41,6 +41,7 @@ #include <linux/memblock.h> #include <linux/of_fdt.h> #include <linux/of_platform.h> +#include <linux/dma-mapping.h> #include <asm/cputype.h> #include <asm/elf.h> @@ -337,3 +338,9 @@ const struct seq_operations cpuinfo_op = { .stop = c_stop, .show = c_show }; + +void arch_setup_pdev_archdata(struct platform_device *pdev) +{ + pdev->archdata.dma_mask = DMA_BIT_MASK(32); + pdev->dev.dma_mask = &pdev->archdata.dma_mask; +}
The dma_mask for a device structure is a pointer. This pointer needs to be set up before the dma mask can actually be set. Most frameworks in the kernel take care of setting this up properly but platform devices that don't follow a regular bus structure may not ever have this set. As a result, checks such as dma_capable will always return false on a raw platform device and dma_set_mask will always return -EIO. Fix this by adding a dma_mask in the platform_device archdata and setting it to be the dma_mask. Devices used in other frameworks can change this as needed. Cc: Will Deacon <will.deacon@arm.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Suggested-by: Kumar Gala <galak@codeaurora.org> Signed-off-by: Laura Abbott <lauraa@codeaurora.org> --- arch/arm64/include/asm/device.h | 1 + arch/arm64/kernel/setup.c | 7 +++++++ 2 files changed, 8 insertions(+), 0 deletions(-)