On 09/29 01:34, Alex Williamson wrote: > On Mon, 28 Sep 2020 15:05:05 -0500 > Wei Huang <wei.hua...@amd.com> wrote: > > > Add support to sync the IOVA-to-GPA translation at the time of IOMMU > > page invalidation. This function is called when two IOMMU commands, > > AMDVI_CMD_INVAL_AMDVI_PAGES and AMDVI_CMD_INVAL_AMDVI_ALL, are > > intercepted. Address space notifiers are called accordingly. > > > > Co-developed-by: Wei Huang <wei.hua...@amd.com> > > Signed-off-by: Suravee Suthikulpanit <suravee.suthikulpa...@amd.com> > > --- > > hw/i386/amd_iommu.c | 177 ++++++++++++++++++++++++++++++++++++++++++++ > > hw/i386/amd_iommu.h | 10 +++ > > hw/vfio/common.c | 3 +- > > 3 files changed, 189 insertions(+), 1 deletion(-) > ... > > diff --git a/hw/vfio/common.c b/hw/vfio/common.c > > index 13471ae29436..243216499ce0 100644 > > --- a/hw/vfio/common.c > > +++ b/hw/vfio/common.c > > @@ -346,7 +346,8 @@ static int vfio_dma_map(VFIOContainer *container, > > hwaddr iova, > > * the VGA ROM space. > > */ > > if (ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0 || > > - (errno == EBUSY && vfio_dma_unmap(container, iova, size) == 0 && > > + ((errno == EEXIST || errno == EBUSY) && > > + vfio_dma_unmap(container, iova, size) == 0 && > > ioctl(container->fd, VFIO_IOMMU_MAP_DMA, &map) == 0)) { > > return 0; > > } > > > This seems like it should be a separate patch. AFAICT the commit log > doesn't even hint at why this change is necessary. I think the -EBUSY > error pre-dates vIOMMU as well. Responding the same for an -EEXIST > almost suggests a coherency issue between QEMU and the kernel, or a > direct mapping replacement without an invalidation, which doesn't seem > to be what this patch is implementing. Thanks,
I went back to check it. Removing this checking code (original code) didn't trigger any issues with Intel 10G passthru NIC. I think this was from the residual debugging code when we started to implement it. Sorry for the confusion. I will remove this code in V2 with more tests. -Wei > > Alex >