Hi, On 3/18/22 23:28, David Laight wrote: > From: Juergen Gross >> Sent: 18 March 2022 16:56 >> >> On 18.03.22 16:22, David Laight wrote: >>> From: Juergen Gross >>>> Sent: 18 March 2022 15:10 >>>> >>>> The dcdbas driver is used to call SMI handlers for both, dcdbas and >>>> dell-smbios-smm. Both drivers allocate a buffer for communicating >>>> with the SMI handler. The physical buffer address is then passed to >>>> the called SMI handler via %ebx. >>>> >>>> Unfortunately this doesn't work when running in Xen dom0, as the >>>> physical address obtained via virt_to_phys() is only a guest physical >>>> address, and not a machine physical address as needed by SMI. >>> >>> The physical address from virt_to_phy() is always wrong. >>> That is the physical address the cpu has for the memory. >>> What you want is the address the dma master interface needs to use. >>> That can be different for a physical system - no need for virtualisation. >>> >>> On x86 they do usually match, but anything with a full iommu >>> will need completely different addresses. >> >> Yes, thanks for reminding me of that. >> >> The SMI handler is running on the cpu, right? So using the DMA >> address is wrong in case of an IOMMU. I really need the machine >> physical address. > > That ought to be handled by the 'dev' parameter to dma_alloc_coherent(). > > David I must admit that I'm not too familiar with all the intricate details of the DMA API here. So does this mean that the patch in its original form is good as is and should be merged? An Acked-by or Reviewed-by from someone more familiar with the DMA APIs would be helpful. Regards, Hans
