Re: [RFC V1 02/12] iommufd: no DMA to BARs

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

On Sat, 20 Jul 2024 12:15:27 -0700
Steve Sistare <steven.sistare@oracle.com> wrote:

> Do not map VFIO PCI BARs for DMA.  This stops a raft of warnings of the
> following form at QEMU start time when using -object iommufd:
> 
> qemu-kvm: warning: IOMMU_IOAS_MAP failed: Bad address, PCI BAR?
> qemu-kvm: vfio_container_dma_map(0x555558282db0, 0x8800010000, 0x4000, 0x7ffff7ff0000) = -14 (Bad address)

NAK.  These mappings are required for P2P DMA between devices.  This is
currently a gap in IOMMUFD support that it doesn't have parity to legacy
vfio containers for these mappings.  Thanks,

Alex

> Signed-off-by: Steve Sistare <steven.sistare@oracle.com>
> ---
>  hw/vfio/common.c      | 3 ++-
>  hw/vfio/helpers.c     | 1 +
>  include/exec/memory.h | 1 +
>  3 files changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/hw/vfio/common.c b/hw/vfio/common.c
> index da2e0ec..403d45a 100644
> --- a/hw/vfio/common.c
> +++ b/hw/vfio/common.c
> @@ -248,7 +248,8 @@ static bool vfio_listener_skipped_section(MemoryRegionSection *section)
>              * are never accessed by the CPU and beyond the address width of
>              * some IOMMU hardware.  TODO: VFIO should tell us the IOMMU width.
>              */
> -           section->offset_within_address_space & (1ULL << 63);
> +           section->offset_within_address_space & (1ULL << 63) ||
> +           section->mr->no_dma;
>  }
>  
>  /* Called with rcu_read_lock held.  */
> diff --git a/hw/vfio/helpers.c b/hw/vfio/helpers.c
> index b14edd4..e4cfdd2 100644
> --- a/hw/vfio/helpers.c
> +++ b/hw/vfio/helpers.c
> @@ -435,6 +435,7 @@ int vfio_region_mmap(VFIORegion *region)
>                                            memory_region_owner(region->mem),
>                                            name, region->mmaps[i].size,
>                                            region->mmaps[i].mmap);
> +        region->mmaps[i].mem.no_dma = true;
>          g_free(name);
>          memory_region_add_subregion(region->mem, region->mmaps[i].offset,
>                                      &region->mmaps[i].mem);
> diff --git a/include/exec/memory.h b/include/exec/memory.h
> index ea03ef2..850cc8c 100644
> --- a/include/exec/memory.h
> +++ b/include/exec/memory.h
> @@ -794,6 +794,7 @@ struct MemoryRegion {
>      bool unmergeable;
>      uint8_t dirty_log_mask;
>      bool is_iommu;
> +    bool no_dma;
>      RAMBlock *ram_block;
>      Object *owner;
>      /* owner as TYPE_DEVICE. Used for re-entrancy checks in MR access hotpath */