On 11/3/23 14:15, Laszlo Ersek wrote: > On 11/3/23 06:15, Vivek Kasireddy wrote: >> If the user specified a size for the PCI MMIO window via the option: >> -fw_cfg name=opt/ovmf/X-PciMmio64Mb,string=32768 >> then this patch ensures that the mmio window is not resized again. >> >> Essentially, this prevents the change introduced in the following >> patch from taking effect: >> commit ecb778d0ac62560aa172786ba19521f27bc3f650 >> Author: Gerd Hoffmann <kra...@redhat.com> >> Date: Tue Oct 4 15:47:27 2022 +0200 >> >> OvmfPkg/PlatformInitLib: dynamic mmio window size >> >> In case we have a reliable PhysMemAddressWidth use that to dynamically >> size the 64bit address window. Allocate 1/8 of the physical address >> space and place the window at the upper end of the address space. >> >> The problem this patch is trying to solve is the VFIO mapping failures: >> VFIO_MAP_DMA failed: Invalid argument >> vfio_dma_map(0x557b2f2736d0, 0x380000000000, 0x1000000, 0x7f98ac400000) = >> -22 (Invalid argument) >> that occur when we try to passthrough the graphics device to the guest: >> qemu-system-x86_64 -m 4096 -enable-kvm -cpu host -smp >> cores=4,threads=2,sockets=1 >> -device vfio-pci,host=0000:00:02.0 -bios OVMF.fd -nographic >> >> The above failures seem to occur because of a mismatch between the >> PhysMemAddressWidth and the Host IOMMU address width. More specifically, >> if the PhysMemAddressWidth is bigger than the IOMMU address width, >> VFIO fails to map the MMIO regions as the IOVAs would be larger >> than the IOMMU aperture regions. When tested on modern Intel platforms >> such as ADL, MTL, etc, OVMF determines PhysMemAddressWidth = 46 which >> matches the Host address width but the IOMMU address width seems to >> range anywhere from 38 to 48 depending on the IOMMU hardware >> capabilities, version, etc. >> >> One way to address this issue is if we ensure that PhysMemAddressWidth >> matches IOMMU address width: >> -cpu host,host-phys-bits=on,host-phys-bits-limit=<IOMMU address width> >> However, this requires the user to figure out the IOMMU address width; >> which can be determined by looking at the 16-21 bits of the cap value: >> cat /sys/devices/virtual/iommu/dmar0/intel-iommu/cap >> or by reading the DMAR_CAP_REG register. But this does not seem like >> a reasonable approach to solve this problem. > > Very nice problem description, already outlining the solution as well. > >> >> Therefore, this problem requires an OVMF specific solution to retain >> the prior behavior. To this end, this patch reuses the X-PciMmio64Mb >> option to opt-out of the behavior introduced in the above commit >> instead of adding a new option or mechanism. > > No, the right solution is to enhance QEMU to query the host IOMMU > address width. Then the following options arise: > > - either pass *both* the host CPU address width *and* the host IOMMU > address width down to OVMF, and teach OVMF to pick the stricter > limitation, for dynamically sizing the MMIO window > > - or let QEMU calulate the stricter width internally, and pass that > (sole, scalar) piece of information down to OVMF. Teach OVMF to query > this new piece of information, and size the MMIO window accordingly. > > Basically the QEMU command line-based workaround that you describe is > what we need to automate (except we need a new information channel for > it, because presenting the strict host IOMMU address width as the VCPU > address width (via CPUID) to the guest is smelly). > > I agree that the proposed patch can function as a stop-gap, but the QEMU > command line hack is already a stop-gap. And for the long term, this > patch is not good enough. We should enhance the dynamic sizing, now that > Gerd has put it into place.
... I do agree however that the current behavior is strange -- the user specifies an explicit fw_cfg knob for OVMF, and OVMF ignores it (for whatever reason). I'd like to know what Gerd thinks of this. Personally I'd ACK the *code* in this patch, just to restore the correct priority between the dynamic sizing and the explicit fw_cfg knob, if: (a) the commit message referred to that exactly (i.e., the to the proper priority between these two configuration avenues), and (b) there were a promise to enhance QEMU and OVMF as I suggest above. I don't want the fw_cfg knob to stick around as a permanent excuse for not improving the dynamic sizing -- now that we *do have* dynamic sizing. BTW: I'd suggest renaming "QemuFwCfgSizeSpecified" to "PciMmio64MbOverride" or something like that. The important trait is that the value is an override (or direct setting) from the user; fw_cfg is incidental / irrelevant where the value is consumed. Thanks! Laszlo -=-=-=-=-=-=-=-=-=-=-=- Groups.io Links: You receive all messages sent to this group. View/Reply Online (#110624): https://edk2.groups.io/g/devel/message/110624 Mute This Topic: https://groups.io/mt/102359124/21656 Group Owner: devel+ow...@edk2.groups.io Unsubscribe: https://edk2.groups.io/g/devel/leave/9847357/21656/1706620634/xyzzy [arch...@mail-archive.com] -=-=-=-=-=-=-=-=-=-=-=-