On Thu, May 18, 2023 at 11:48:55AM +0200, Konrad Dybcio wrote: > On 18.05.2023 11:39, Stephan Gerhold wrote: > > While SDHCI claims to support 64-bit DMA on MSM8916 it does not seem to > > be properly functional. It is not immediately obvious because SDHCI is > > usually used with IOMMU bypassed on this SoC, and all physical memory > > has 32-bit addresses. But when trying to enable the IOMMU it quickly > > fails with an error such as the following: > > > > arm-smmu 1e00000.iommu: Unhandled context fault: > > fsr=0x402, iova=0xfffff200, fsynr=0xe0000, cbfrsynra=0x140, cb=3 > > mmc1: ADMA error: 0x02000000 > > mmc1: sdhci: ============ SDHCI REGISTER DUMP =========== > > mmc1: sdhci: Sys addr: 0x00000000 | Version: 0x00002e02 > > mmc1: sdhci: Blk size: 0x00000008 | Blk cnt: 0x00000000 > > mmc1: sdhci: Argument: 0x00000000 | Trn mode: 0x00000013 > > mmc1: sdhci: Present: 0x03f80206 | Host ctl: 0x00000019 > > mmc1: sdhci: Power: 0x0000000f | Blk gap: 0x00000000 > > mmc1: sdhci: Wake-up: 0x00000000 | Clock: 0x00000007 > > mmc1: sdhci: Timeout: 0x0000000a | Int stat: 0x00000001 > > mmc1: sdhci: Int enab: 0x03ff900b | Sig enab: 0x03ff100b > > mmc1: sdhci: ACmd stat: 0x00000000 | Slot int: 0x00000000 > > mmc1: sdhci: Caps: 0x322dc8b2 | Caps_1: 0x00008007 > > mmc1: sdhci: Cmd: 0x0000333a | Max curr: 0x00000000 > > mmc1: sdhci: Resp[0]: 0x00000920 | Resp[1]: 0x5b590000 > > mmc1: sdhci: Resp[2]: 0xe6487f80 | Resp[3]: 0x0a404094 > > mmc1: sdhci: Host ctl2: 0x00000008 > > mmc1: sdhci: ADMA Err: 0x00000001 | ADMA Ptr: 0x0000000ffffff224 > > mmc1: sdhci_msm: ----------- VENDOR REGISTER DUMP ----------- > > mmc1: sdhci_msm: DLL sts: 0x00000000 | DLL cfg: 0x60006400 | DLL cfg2: 0x00000000 > > mmc1: sdhci_msm: DLL cfg3: 0x00000000 | DLL usr ctl: 0x00000000 | DDR cfg: 0x00000000 > > mmc1: sdhci_msm: Vndr func: 0x00018a9c | Vndr func2 : 0xf88018a8 Vndr func3: 0x00000000 > > mmc1: sdhci: ============================================ > > mmc1: sdhci: fffffffff200: DMA 0x0000ffffffffe100, LEN 0x0008, Attr=0x21 > > mmc1: sdhci: fffffffff20c: DMA 0x0000000000000000, LEN 0x0000, Attr=0x03 > > > > Looking closely it's obvious that only the 32-bit part of the address > > (0xfffff200) arrives at the SMMU, the higher 16-bit (0xffff...) get > > lost somewhere. This might not be a limitation of the SDHCI itself but > > perhaps the bus/interconnect it is connected to, or even the connection > > to the SMMU. > > > > Work around this by setting SDHCI_QUIRK2_BROKEN_64_BIT_DMA to avoid > > using 64-bit addresses. > > > > Signed-off-by: Stephan Gerhold <stephan@xxxxxxxxxxx> > > --- > Would using 64bit address translation not require you to use (dma-)ranges > with a greater-than-default size, like we do on newer platforms? Did you > test that by chance? > No, there is no "dma-ranges" in msm8916.dtsi. It seems to use 64-bit (virtual) addresses by default for SDHCI, limited to 48-bit by the SMMU. I tried limiting dma-ranges to 32-bit in msm8916.dtsi /soc@0 like dma-ranges = <0 0 0 0xffffffff>; This seems to work as well as an alternative to this patch, although it causes several annoying "Invalid size 0xffffffff for dma-range(s)" warnings. I know the 0xffffffff isn't valid but I don't see how I would actually specify 2^32 there without rewriting the entire msm8916.dtsi to use #address/size-cells = <2> instead of <1>. It's not entirely clear to me where the actual limitation is coming from here. It could be specific to the SDHCI (in which case this patch makes more sense) or a general limitation of the bus/interconnect (in which case dma-ranges might make more sense). In any case 64-bit DMA is broken for SDHCI on MSM8916 so I would say this patch is still the right thing to do. It's pointless to waste extra cycles and memory to manage 64-bit pointers for SDHCI ADMA 64-bit, when they can never work successfully on this SoC. The dma-ranges approach doesn't change that, while SDHCI_QUIRK2_BROKEN_64_BIT_DMA actually makes it use ADMA in 32-bit mode. Thanks, Stephan