On 3/28/19 4:25 AM, Marek Vasut wrote: > On 3/19/19 12:25 AM, Marek Vasut wrote: >> On 3/18/19 2:14 PM, Robin Murphy wrote: >>> On 17/03/2019 23:36, Marek Vasut wrote: >>>> On 3/17/19 11:29 AM, Geert Uytterhoeven wrote: >>>>> Hi Marek, >>>> >>>> Hi, >>>> >>>>> On Sun, Mar 17, 2019 at 12:04 AM Marek Vasut <marek.vasut@xxxxxxxxx> >>>>> wrote: >>>>>> On 3/16/19 10:25 PM, Marek Vasut wrote: >>>>>>> On 3/13/19 7:30 PM, Christoph Hellwig wrote: >>>>>>>> On Sat, Mar 09, 2019 at 12:23:15AM +0100, Marek Vasut wrote: >>>>>>>>> On 3/8/19 8:18 AM, Christoph Hellwig wrote: >>>>>>>>>> On Thu, Mar 07, 2019 at 12:14:06PM +0100, Marek Vasut wrote: >>>>>>>>>>>> Right, but whoever *interprets* the device masks after the >>>>>>>>>>>> driver has >>>>>>>>>>>> overridden them should be taking the (smaller) bus mask into >>>>>>>>>>>> account as >>>>>>>>>>>> well, so the question is where is *that* not being done >>>>>>>>>>>> correctly? >>>>>>>>>>> >>>>>>>>>>> Do you have a hint where I should look for that ? >>>>>>>>>> >>>>>>>>>> If this a 32-bit ARM platform it might the complete lack of support >>>>>>>>>> for bus_dma_mask in arch/arm/mm/dma-mapping.c.. >>>>>>>>> >>>>>>>>> It's an ARM 64bit platform, just the PCIe controller is limited >>>>>>>>> to 32bit >>>>>>>>> address range, so the devices on the PCIe bus cannot read the host's >>>>>>>>> DRAM above the 32bit limit. >>>>>>>> >>>>>>>> arm64 should take the mask into account both for the swiotlb and >>>>>>>> iommu case. What are the exact symptoms you see? >>>>>>> >>>>>>> With the nvme, the device is recognized, but cannot be used. >>>>>>> It boils down to PCI BAR access being possible, since that's all below >>>>>>> the 32bit boundary, but when the device tries to do any sort of DMA, >>>>>>> that transfer returns nonsense data. >>>>>>> >>>>>>> But when I call dma_set_mask_and_coherent(dev->dev, >>>>>>> DMA_BIT_MASK(32) in >>>>>>> the affected driver (thus far I tried this nvme, xhci-pci and ahci-pci >>>>>>> drivers), it all starts to work fine. >>>>>>> >>>>>>> Could it be that the driver overwrites the (coherent_)dma_mask and >>>>>>> that's why the swiotlb/iommu code cannot take this into account ? >>>>>>> >>>>>>>> Does it involve >>>>>>>> swiotlb not kicking in, or iommu issues? >>>>>>> >>>>>>> How can I check ? I added printks into arch/arm64/mm/dma-mapping.c and >>>>>>> drivers/iommu/dma-iommu.c , but I suspect I need to look elsewhere. >>>>>> >>>>>> Digging further ... >>>>>> >>>>>> drivers/nvme/host/pci.c nvme_map_data() calls dma_map_sg_attrs() and >>>>>> the >>>>>> resulting sglist contains entry with >32bit PA. This is because >>>>>> dma_map_sg_attrs() calls dma_direct_map_sg(), which in turn calls >>>>>> dma_direct_map_sg(), then dma_direct_map_page() and that's where it >>>>>> goes >>>>>> weird. >>>>>> >>>>>> dma_direct_map_page() does a dma_direct_possible() check before >>>>>> triggering swiotlb_map(). The check succeeds, so the later isn't >>>>>> executed. >>>>>> >>>>>> dma_direct_possible() calls dma_capable() with dev->dma_mask = >>>>>> DMA_BIT_MASK(64) and dev->dma_bus_mask = 0, so >>>>>> min_not_zero(*dev->dma_mask, dev->bus_dma_mask) returns >>>>>> DMA_BIT_MASK(64). >>>>>> >>>>>> Surely enough, if I hack dma_direct_possible() to return 0, >>>>>> swiotlb_map() kicks in and the nvme driver starts working fine. >>>>>> >>>>>> I presume the question here is, why is dev->bus_dma_mask = 0 ? >>>>> >>>>> Because that's the default, and almost no code overrides that? >>>> >>>> But shouldn't drivers/of/device.c set that for the PCIe controller ? >>> >>> Urgh, I really should have spotted the significance of "NVMe", but >>> somehow it failed to click :( >> >> Good thing it did now :-) >> >>> Of course the existing code works fine for everything *except* PCI >>> devices on DT-based systems... That's because of_dma_get_range() has >>> never been made to work correctly with the trick we play of passing the >>> host bridge of_node through of_dma_configure(). I've got at least 2 or 3 >>> half-finished attempts at improving that, but they keep getting >>> sidetracked into trying to clean up the various new of_dma_configure() >>> hacks I find in drivers and/or falling down the rabbit-hole of starting >>> to redesign the whole dma_pfn_offset machinery entirely. Let me dig one >>> up and try to constrain it to solve just this most common "one single >>> limited range" condition for the sake of making actual progress... >> >> That'd be nice, thank you. I'm happy to test it on various devices here. > > Just curious, no stress, did you get anywhere with this patch(set) yet? Bump ? -- Best regards, Marek Vasut