Re: PCIe device paththrough via vfio issue

[Alex Williamson <alex.williamson@redhat.com>]

On Tue, 14 Jan 2020 19:02:41 +0300
yurij <lnkgyv@gmail.com> wrote:

> On 1/14/20 5:04 PM, Alex Williamson wrote:
> > On Tue, 14 Jan 2020 17:14:33 +1100
> > Alexey Kardashevskiy <aik@ozlabs.ru> wrote:
> >   
> >> On 14/01/2020 03:28, Alex Williamson wrote:  
> >>> On Mon, 13 Jan 2020 18:49:21 +0300
> >>> yurij <lnkgyv@gmail.com> wrote:
> >>>      
> >>>> Hello everybody!
> >>>>
> >>>> I have a specific PCIe device (sorry, but I can't tell about what is it
> >>>> and what it does) but PCI configuration space consists of 4 BARs (lspci
> >>>> output brief):
> >>>>
> >>>> lspci -s 84:00.00 -vvv
> >>>>
> >>>> . . .
> >>>> Region 0: Memory at fa000000 (64-bit, non-prefetchable) [size=16M]
> >>>> 	Region 2: Memory at fb001000 (32-bit, non-prefetchable) [size=4K]
> >>>> 	Region 3: Memory at fb000000 (32-bit, non-prefetchable) [size=4K]
> >>>> 	Region 4: Memory at f9000000 (64-bit, non-prefetchable) [size=16M]
> >>>> . . .
> >>>> Kernel driver in use: vfio-pci
> >>>> . . .
> >>>>
> >>>> BAR0 merged with BAR1, BAR4 merged with BAR5 so they are 64 bit width.
> >>>>
> >>>> I put this PCIe device in virtual machine via vfio:
> >>>>
> >>>> -device vfio-pci,host=84:00.0,id=hostdev0,bus=pci.6,addr=0x0
> >>>>
> >>>> Virtual machine successfully boot. PCI configuration space in virtual
> >>>> environment looks OK (lspci output brief):
> >>>>
> >>>> lspci -s 06:00.0 -vvv
> >>>>
> >>>> . . .
> >>>> Region 0: Memory at f8000000 (64-bit, non-prefetchable) [size=16M]
> >>>> 	Region 2: Memory at fa000000 (32-bit, non-prefetchable) [size=4K]
> >>>> 	Region 3: Memory at fa001000 (32-bit, non-prefetchable) [size=4K]
> >>>> 	Region 4: Memory at f9000000 (64-bit, non-prefetchable) [size=16M]
> >>>> . . .
> >>>> Kernel driver in use: custom_driver
> >>>>
> >>>> BAR0 merged with BAR1 and BAR4 merged with BAR5 and so they are also 64
> >>>> bit width.
> >>>>
> >>>> The main problem in 4K HOLE in REGION 0 in virtual environment. So some
> >>>> device features don't work.
> >>>>
> >>>> I have enabled iommu trace in host system (trace_event=iommu) and
> >>>> display all events (for i in $(find
> >>>> /sys/kernel/debug/tracing/events/iommu/ -name enable);do echo 1 > $i;
> >>>> done). I saw next events during virtual machine booting:
> >>>>
> >>>> # cat /sys/kernel/debug/tracing/trace
> >>>> . . .
> >>>>          CPU 0/KVM-3046  [051] .... 63113.338894: map: IOMMU:
> >>>> iova=0x00000000f8000000 paddr=0x00000000fa000000 size=24576
> >>>>          CPU 0/KVM-3046  [051] .... 63113.339177: map: IOMMU:
> >>>> iova=0x00000000f8007000 paddr=0x00000000fa007000 size=16748544
> >>>>          CPU 0/KVM-3046  [051] .... 63113.339444: map: IOMMU:
> >>>> iova=0x00000000fa000000 paddr=0x00000000fb001000 size=4096
> >>>>          CPU 0/KVM-3046  [051] .... 63113.339697: map: IOMMU:
> >>>> iova=0x00000000fa001000 paddr=0x00000000fb000000 size=4096
> >>>>          CPU 0/KVM-3046  [051] .... 63113.340209: map: IOMMU:
> >>>> iova=0x00000000f9000000 paddr=0x00000000f9000000 size=16777216
> >>>> . . .
> >>>>
> >>>> I have enabled qemu trace(-trace events=/root/qemu/trace_events). Trace
> >>>> file consists of the falling functions:
> >>>> vfio_region_mmap
> >>>> vfio_get_dev_region
> >>>> vfio_pci_size_rom
> >>>> vfio_pci_read_config
> >>>> vfio_pci_write_config
> >>>> vfio_iommu_map_notify
> >>>> vfio_listener_region_add_iommu
> >>>> vfio_listener_region_add_ram
> >>>>
> >>>> Some important brief from qemu trace:
> >>>> . . .
> >>>> янв 13 18:17:24 VM qemu-system-x86_64[7131]: vfio_region_mmap Region
> >>>> 0000:84:00.0 BAR 0 mmaps[0] [0x0 - 0xffffff]
> >>>> янв 13 18:17:24 VM qemu-system-x86_64[7131]: vfio_region_mmap Region
> >>>> 0000:84:00.0 BAR 2 mmaps[0] [0x0 - 0xfff]
> >>>> янв 13 18:17:24 VM qemu-system-x86_64[7131]: vfio_region_mmap Region
> >>>> 0000:84:00.0 BAR 3 mmaps[0] [0x0 - 0xfff]
> >>>> янв 13 18:17:24 VM qemu-system-x86_64[7131]: vfio_region_mmap Region
> >>>> 0000:84:00.0 BAR 4 mmaps[0] [0x0 - 0xffffff]
> >>>> . . .
> >>>> янв 13 18:17:37 VM qemu-system-x86_64[7131]:
> >>>> vfio_listener_region_add_ram region_add [ram] 0xf8000000 - 0xf8005fff
> >>>> [0x7f691e800000]
> >>>> янв 13 18:17:37 VM qemu-system-x86_64[7131]:
> >>>> vfio_listener_region_add_ram region_add [ram] 0xf8007000 - 0xf8ffffff
> >>>> [0x7f691e807000]
> >>>> янв 13 18:17:37 VM qemu-system-x86_64[7131]:
> >>>> vfio_listener_region_add_ram region_add [ram] 0xfa000000 - 0xfa000fff
> >>>> [0x7f6b5de37000]
> >>>> янв 13 18:17:37 VM qemu-system-x86_64[7131]:
> >>>> vfio_listener_region_add_ram region_add [ram] 0xfa001000 - 0xfa001fff
> >>>> [0x7f6b58004000]
> >>>> янв 13 18:17:37 VM qemu-system-x86_64[7131]:
> >>>> vfio_listener_region_add_ram region_add [ram] 0xf9000000 - 0xf9ffffff
> >>>> [0x7f691d800000]
> >>>>
> >>>> I use qemu 4.0.0 which I rebuild for tracing support
> >>>> (--enable-trace-backends=syslog).
> >>>>
> >>>> Please, help me solve this issue. Thank you!  
> >>>
> >>> Something has probably created a QEMU MemoryRegion overlapping the BAR,
> >>> we do this for quirks where we want to intercept a range of MMIO for
> >>> emulation, but the offset 0x6000 on BAR0 doesn't sound familiar to me.
> >>> Run the VM with a monitor and see if 'info mtree' provides any info on
> >>> the handling of that overlap.  Thanks,  
> >>
> >>
> >> Could not it be an MSIX region? 'info mtree -f' should tell exactly what
> >> is going on.  
> > 
> > Oh, good call, that's probably it.  The PCI spec specifically
> > recommends against placing non-MSIX related registers within the same
> > 4K page as the vector table to avoid such things:
> > 
> >   If a Base Address register that maps address space for the MSI-X Table
> >   or MSI-X PBA also maps other usable address space that is not
> >   associated with MSI-X structures, locations (e.g., for CSRs) used in
> >   the other address space must not share any naturally aligned 4-KB
> >   address range with one where either MSI-X structure resides. This
> >   allows system software where applicable to use different processor
> >   attributes for MSI-X structures and the other address space.
> > 
> > We have the following QEMU vfio-pci device option to relocate the BAR
> > elsewhere for hardware that violates that recommendation or for where
> > the PCI spec recommended alignment isn't sufficient:
> > 
> >   x-msix-relocation=<OffAutoPCIBAR> - off/auto/bar0/bar1/bar2/bar3/bar4/bar5
> > 
> > In this case I'd probably recommend bar2 or bar3 as those BARs would
> > only be extended to 8K versus bar0/4 would be extended to 32M.  Thanks,
> > 
> > Alex
> >   
> 
>  >   x-msix-relocation=<OffAutoPCIBAR> -   
> off/auto/bar0/bar1/bar2/bar3/bar4/bar5
> 
> I have used successfully 'x-msix-relocation' option:
> -device 
> vfio-pci,host=84:00.0,id=hostdev0,bus=pci.6,addr=0x0,x-msix-relocation=bar2
> 
> Now,  IOMMU trace looks like:
> . . .
>         CPU 0/KVM-4237  [055] ....  4750.918416: map: IOMMU: 
> iova=0x00000000f8000000 paddr=0x00000000fa000000 size=16777216
>         CPU 0/KVM-4237  [055] ....  4750.918740: map: IOMMU: 
> iova=0x00000000fa000000 paddr=0x00000000fb001000 size=4096
>         CPU 0/KVM-4237  [055] ....  4750.919069: map: IOMMU: 
> iova=0x00000000fa002000 paddr=0x00000000fb000000 size=4096
>         CPU 0/KVM-4237  [055] ....  4750.919698: map: IOMMU: 
> iova=0x00000000f9000000 paddr=0x00000000f9000000 size=16777216
> . . .
> 
> All seems to be OK.
> 
> Thank you very much!

Glad it worked, but please also tell your hardware developers to follow
the PCI spec recommendations for alignment of MSI-X related data
structures and ideally use a BAR dedicated to MSI-X for independence
from the processor page size.  If this is a device under development,
it would avoid this headache for future users.  Thanks,

Alex