| Message ID | 20160427135859.GC20646@e104818-lin.cambridge.arm.com |
|---|---|
| State | New |
| Headers | show |
On Wednesday 27 April 2016 14:59:00 Catalin Marinas wrote: > > I would be in favour of a dma_inherit() function as well. We could hack > something up in the arch code (like below) but I would rather prefer an > explicit dma_inherit() call by drivers creating such devices. > > diff --git a/arch/arm64/include/asm/dma-mapping.h b/arch/arm64/include/asm/dma-mapping.h > index ba437f090a74..ea6fb9b0e8fa 100644 > --- a/arch/arm64/include/asm/dma-mapping.h > +++ b/arch/arm64/include/asm/dma-mapping.h > @@ -29,8 +29,11 @@ extern struct dma_map_ops dummy_dma_ops; > > static inline struct dma_map_ops *__generic_dma_ops(struct device *dev) > { > - if (dev && dev->archdata.dma_ops) > - return dev->archdata.dma_ops; > + while (dev) { > + if (dev->archdata.dma_ops) > + return dev->archdata.dma_ops; > + dev = dev->parent; > + } I think this would be a very bad idea: we don't want to have random devices be able to perform DMA just because their parent devices have been set up that way. Arnd
On 04/27/2016 04:59 PM, Catalin Marinas wrote: > On Wed, Apr 27, 2016 at 08:41:06AM +0300, Felipe Balbi wrote: >> Grygorii Strashko <grygorii.strashko@ti.com> writes: >>> On 04/26/2016 09:17 AM, Felipe Balbi wrote: >>>> Grygorii Strashko <grygorii.strashko@ti.com> writes: >>>>> Now not all DMA paremters configured properly for "xhci-hcd" platform >>>>> device which is created manually. For example: dma_pfn_offset, dam_ops >>>>> and iommu configuration will not corresponds "dwc3" devices >>>>> configuration. As result, this will cause problems like wrong DMA >>>>> addresses translation on platforms with LPAE enabled like Keystone 2. >>>>> >>>>> When platform is using DT boot mode the DMA configuration will be >>>>> parsed and applied from DT, so, to fix this issue, reuse >>>>> of_dma_configure() API and retrieve DMA configuartion for "xhci-hcd" >>>>> from DWC3 device node. >>>> >>>> patch is incomplete. You left out non-DT users which might suffer from >>>> the same problem. >>> >>> Honestly, I don't know how to fix it gracefully for non-DT case. >>> I can update commit message to mention that this is fix for DT case only. >> >> no, that won't do :-) There are other users for this driver and they are >> all "out-of-compliance" when it comes to DMA usage. Apparently, the >> desired behavior is to pass correct device to DMA API which the gadget >> side is already doing (see below). For the host side, the fix has to be >> more involved. >> >> Frankly, I'd prefer that DMA setup could be inherited from parent >> device, then it wouldn't really matter and a bunch of this could be >> simplified. Some sort of dma_inherit(struct device *dev, struct device >> *parent) would go a long way, IMHO. > > I would be in favour of a dma_inherit() function as well. We could hack > something up in the arch code (like below) but I would rather prefer an > explicit dma_inherit() call by drivers creating such devices. > > diff --git a/arch/arm64/include/asm/dma-mapping.h b/arch/arm64/include/asm/dma-mapping.h > index ba437f090a74..ea6fb9b0e8fa 100644 > --- a/arch/arm64/include/asm/dma-mapping.h > +++ b/arch/arm64/include/asm/dma-mapping.h > @@ -29,8 +29,11 @@ extern struct dma_map_ops dummy_dma_ops; > > static inline struct dma_map_ops *__generic_dma_ops(struct device *dev) > { > - if (dev && dev->archdata.dma_ops) > - return dev->archdata.dma_ops; > + while (dev) { > + if (dev->archdata.dma_ops) > + return dev->archdata.dma_ops; > + dev = dev->parent; > + } > > /* > * We expect no ISA devices, and all other DMA masters are expected to > It's no enough to W/A just dma_ops :( dma_inherit()... FYI: http://www.spinics.net/lists/arm-kernel/msg384012.html Maybe you'll be able to find the way to make it acceptable. -- regards, -grygorii
On Wed, Apr 27, 2016 at 04:11:17PM +0200, Arnd Bergmann wrote: > On Wednesday 27 April 2016 14:59:00 Catalin Marinas wrote: > > > > I would be in favour of a dma_inherit() function as well. We could hack > > something up in the arch code (like below) but I would rather prefer an > > explicit dma_inherit() call by drivers creating such devices. > > > > diff --git a/arch/arm64/include/asm/dma-mapping.h b/arch/arm64/include/asm/dma-mapping.h > > index ba437f090a74..ea6fb9b0e8fa 100644 > > --- a/arch/arm64/include/asm/dma-mapping.h > > +++ b/arch/arm64/include/asm/dma-mapping.h > > @@ -29,8 +29,11 @@ extern struct dma_map_ops dummy_dma_ops; > > > > static inline struct dma_map_ops *__generic_dma_ops(struct device *dev) > > { > > - if (dev && dev->archdata.dma_ops) > > - return dev->archdata.dma_ops; > > + while (dev) { > > + if (dev->archdata.dma_ops) > > + return dev->archdata.dma_ops; > > + dev = dev->parent; > > + } > > I think this would be a very bad idea: we don't want to have random > devices be able to perform DMA just because their parent devices > have been set up that way. I agree, it's a big hack. It would be nice to have a simpler way to do this in driver code rather than explicitly calling of_dma_configure/arch_setup_dma_ops as per the original patch in this thread. -- Catalin
On Wednesday 27 April 2016 16:50:19 Catalin Marinas wrote: > On Wed, Apr 27, 2016 at 04:11:17PM +0200, Arnd Bergmann wrote: > > On Wednesday 27 April 2016 14:59:00 Catalin Marinas wrote: > > > > > > I would be in favour of a dma_inherit() function as well. We could hack > > > something up in the arch code (like below) but I would rather prefer an > > > explicit dma_inherit() call by drivers creating such devices. > > > > > > diff --git a/arch/arm64/include/asm/dma-mapping.h b/arch/arm64/include/asm/dma-mapping.h > > > index ba437f090a74..ea6fb9b0e8fa 100644 > > > --- a/arch/arm64/include/asm/dma-mapping.h > > > +++ b/arch/arm64/include/asm/dma-mapping.h > > > @@ -29,8 +29,11 @@ extern struct dma_map_ops dummy_dma_ops; > > > > > > static inline struct dma_map_ops *__generic_dma_ops(struct device *dev) > > > { > > > - if (dev && dev->archdata.dma_ops) > > > - return dev->archdata.dma_ops; > > > + while (dev) { > > > + if (dev->archdata.dma_ops) > > > + return dev->archdata.dma_ops; > > > + dev = dev->parent; > > > + } > > > > I think this would be a very bad idea: we don't want to have random > > devices be able to perform DMA just because their parent devices > > have been set up that way. > > I agree, it's a big hack. It would be nice to have a simpler way to do > this in driver code rather than explicitly calling > of_dma_configure/arch_setup_dma_ops as per the original patch in this > thread. > I haven't followed the entire discussion, but what's wrong with passing around a pointer to a 'struct device *hwdev' that represents the physical device that does the DMA? Arnd
Hi, Arnd Bergmann <arnd@arndb.de> writes: > On Wednesday 27 April 2016 16:50:19 Catalin Marinas wrote: >> On Wed, Apr 27, 2016 at 04:11:17PM +0200, Arnd Bergmann wrote: >> > On Wednesday 27 April 2016 14:59:00 Catalin Marinas wrote: >> > > >> > > I would be in favour of a dma_inherit() function as well. We could hack >> > > something up in the arch code (like below) but I would rather prefer an >> > > explicit dma_inherit() call by drivers creating such devices. >> > > >> > > diff --git a/arch/arm64/include/asm/dma-mapping.h b/arch/arm64/include/asm/dma-mapping.h >> > > index ba437f090a74..ea6fb9b0e8fa 100644 >> > > --- a/arch/arm64/include/asm/dma-mapping.h >> > > +++ b/arch/arm64/include/asm/dma-mapping.h >> > > @@ -29,8 +29,11 @@ extern struct dma_map_ops dummy_dma_ops; >> > > >> > > static inline struct dma_map_ops *__generic_dma_ops(struct device *dev) >> > > { >> > > - if (dev && dev->archdata.dma_ops) >> > > - return dev->archdata.dma_ops; >> > > + while (dev) { >> > > + if (dev->archdata.dma_ops) >> > > + return dev->archdata.dma_ops; >> > > + dev = dev->parent; >> > > + } >> > >> > I think this would be a very bad idea: we don't want to have random >> > devices be able to perform DMA just because their parent devices >> > have been set up that way. >> >> I agree, it's a big hack. It would be nice to have a simpler way to do >> this in driver code rather than explicitly calling >> of_dma_configure/arch_setup_dma_ops as per the original patch in this >> thread. > > I haven't followed the entire discussion, but what's wrong with passing > around a pointer to a 'struct device *hwdev' that represents the physical > device that does the DMA? that will likely create duplicated solutions in several drivers and it'll be a pain to maintain. There's another complication, dwc3 can be integrated in many different ways. See the device child-parent tree representations below: a) with a parent PCI device: pci_bus_type - dwc3-pci - dwc3 - xhci-plat b) with a parent platform_device (OF): platform_bus_type - dwc3-${omap,st,of-simple,exynos,keystone} - dwc3 - xhci-plat c) without a parent at all (thanks Grygorii): platform_bus_type - dwc3 - xhci-plat (a) and (b) above are the common cases. The DMA-capable device is clearly dwc3-${pci,omap,st,of-simple,exynos,keystone} with dwc3 only having proper DMA configuration in OF platforms (because of the unconditional of_dma_configure() during OF device creation) and xhci-plat not knowing about DMA at all and hardcoding some crappy defaults. (c) is the uncommon case which creates some problems. In this case, dwc3 itself is the DMA-capable device and dwc3->dev->parent is the platform_bus_type itself. Now consider the problem this creates: i. the patch that I wrote [1] becomes invalid for (c), thanks to Grygorii for pointing this out before it was too late. ii. xhci-plat can also be described directly in DT (and is in some cases). This means that assuming xhci-plat's parent's parent to be the DMA-capable device is also an invalid assumption. iii. one might argue that for DT-based platforms *with* a glue layer ((b) above), OF already "copies" some sensible DMA defaults during device creation. PCI-based systems just don't have the luxury of creating random PCI devices like that :-) I say it copies because I can pass *any* struct device_node pointer and it'll just copy that to the struct device argument. Here's of_dma_configure() to make your life easier: void of_dma_configure(struct device *dev, struct device_node *np) { u64 dma_addr, paddr, size; int ret; bool coherent; unsigned long offset; struct iommu_ops *iommu; /* * Set default coherent_dma_mask to 32 bit. Drivers are expected to * setup the correct supported mask. */ if (!dev->coherent_dma_mask) dev->coherent_dma_mask = DMA_BIT_MASK(32); /* * Set it to coherent_dma_mask by default if the architecture * code has not set it. */ if (!dev->dma_mask) dev->dma_mask = &dev->coherent_dma_mask; ret = of_dma_get_range(np, &dma_addr, &paddr, &size); if (ret < 0) { dma_addr = offset = 0; size = dev->coherent_dma_mask + 1; } else { offset = PFN_DOWN(paddr - dma_addr); /* * Add a work around to treat the size as mask + 1 in case * it is defined in DT as a mask. */ if (size & 1) { dev_warn(dev, "Invalid size 0x%llx for dma-range\n", size); size = size + 1; } if (!size) { dev_err(dev, "Adjusted size 0x%llx invalid\n", size); return; } dev_dbg(dev, "dma_pfn_offset(%#08lx)\n", offset); } dev->dma_pfn_offset = offset; /* * Limit coherent and dma mask based on size and default mask * set by the driver. */ dev->coherent_dma_mask = min(dev->coherent_dma_mask, DMA_BIT_MASK(ilog2(dma_addr + size))); *dev->dma_mask = min((*dev->dma_mask), DMA_BIT_MASK(ilog2(dma_addr + size))); coherent = of_dma_is_coherent(np); dev_dbg(dev, "device is%sdma coherent\n", coherent ? " " : " not "); iommu = of_iommu_configure(dev, np); dev_dbg(dev, "device is%sbehind an iommu\n", iommu ? " " : " not "); arch_setup_dma_ops(dev, dma_addr, size, iommu, coherent); } The only clean way to fix this up is with a dma_inherit()-like API which would allow dwc3's glue layers ((a) and (b) above) to initialize child's (dwc3) DMA configuration during child's creation. Something like below: that's all I'm asking for :-) dma_inherit() should, probably, be arch-specific to handle details like IOMMU (which today sits under archdata). without something like this, we're gonna have to resort to tons of checks trying to find the correct DMA configuration to use in all possible usage scenarios of dwc3. Note, however, that I'm using dwc3 as an example, but anywhere you see manual platform_device creation, there's a potential problem WRT DMA and/or IOMMU. [1] https://git.kernel.org/cgit/linux/kernel/git/balbi/usb.git/commit/?h=testing/next&id=2725d6f974c4c268ae5fb746f8e3b33b76135aa8 -- balbidiff --git a/drivers/usb/dwc3/dwc3-pci.c b/drivers/usb/dwc3/dwc3-pci.c index adc1e8a624cb..74b599269e2c 100644 --- a/drivers/usb/dwc3/dwc3-pci.c +++ b/drivers/usb/dwc3/dwc3-pci.c @@ -152,6 +152,8 @@ static int dwc3_pci_probe(struct pci_dev *pci, return -ENOMEM; } + dma_inherit(&dwc->dev, dev); + memset(res, 0x00, sizeof(struct resource) * ARRAY_SIZE(res)); res[0].start = pci_resource_start(pci, 0);
diff --git a/arch/arm64/include/asm/dma-mapping.h b/arch/arm64/include/asm/dma-mapping.h index ba437f090a74..ea6fb9b0e8fa 100644 --- a/arch/arm64/include/asm/dma-mapping.h +++ b/arch/arm64/include/asm/dma-mapping.h @@ -29,8 +29,11 @@ extern struct dma_map_ops dummy_dma_ops; static inline struct dma_map_ops *__generic_dma_ops(struct device *dev) { - if (dev && dev->archdata.dma_ops) - return dev->archdata.dma_ops; + while (dev) { + if (dev->archdata.dma_ops) + return dev->archdata.dma_ops; + dev = dev->parent; + } /* * We expect no ISA devices, and all other DMA masters are expected to