Re: [PATCH v6 01/21] backends/iommufd: Introduce the iommufd object

[Cédric Le Goater]

On 11/17/23 14:29, Eric Auger wrote:

Hi Cédric,

On 11/17/23 12:39, Duan, Zhenzhong wrote:

Hi Cédric,

-----Original Message-----
From: Cédric Le Goater <c...@redhat.com>
Sent: Friday, November 17, 2023 7:10 PM
Subject: Re: [PATCH v6 01/21] backends/iommufd: Introduce the iommufd object

Hello,

+int iommufd_backend_map_dma(IOMMUFDBackend *be, uint32_t ioas_id,

hwaddr iova,

+                            ram_addr_t size, void *vaddr, bool readonly)
+{
+    int ret, fd = be->fd;
+    struct iommu_ioas_map map = {
+        .size = sizeof(map),
+        .flags = IOMMU_IOAS_MAP_READABLE |
+                 IOMMU_IOAS_MAP_FIXED_IOVA,
+        .ioas_id = ioas_id,
+        .__reserved = 0,
+        .user_va = (uintptr_t)vaddr,
+        .iova = iova,
+        .length = size,
+    };
+
+    if (!readonly) {
+        map.flags |= IOMMU_IOAS_MAP_WRITEABLE;
+    }
+
+    ret = ioctl(fd, IOMMU_IOAS_MAP, &map);
+    trace_iommufd_backend_map_dma(fd, ioas_id, iova, size,
+                                  vaddr, readonly, ret);
+    if (ret) {
+        ret = -errno;
+        error_report("IOMMU_IOAS_MAP failed: %m");
+    }
+    return ret;
+}

When using a UEFI guest, QEMU reports errors when mapping regions
in the top PCI space :

   iommufd_backend_map_dma  iommufd=10 ioas=2 iova=0x380000001000
size=0x3000 addr=0x7fce2c28b000 readonly=0 (-1)
   qemu-system-x86_64: IOMMU_IOAS_MAP failed: Invalid argument
   qemu-system-x86_64: vfio_container_dma_map(0x55a21b03a150,
0x380000001000, 0x3000, 0x7fce2c28b000) = -22 (Invalid argument)

   iommufd_backend_map_dma  iommufd=10 ioas=2 iova=0x380000004000
size=0x4000 addr=0x7fce2c980000 readonly=0 (-1)
   qemu-system-x86_64: IOMMU_IOAS_MAP failed: Invalid argument
   qemu-system-x86_64: vfio_container_dma_map(0x55a21b03a150,
0x380000004000, 0x4000, 0x7fce2c980000) = -22 (Invalid argument)

This is because IOMMUFD reserved IOVAs areas are :

  [ fee00000 - feefffff ]
  [ 8000000000 - ffffffffffffffff ] (39 bits address space)

which were allocated when the device was initially attached.
The topology is basic. Something is wrong.

        
Thanks for your report. This looks a hardware limit of
host IOMMU address width(39) < guest physical address width.

A similar issue with a fix submitted below, ccing related people.
https://lists.gnu.org/archive/html/qemu-devel/2023-11/msg02937.html
It looks the fix will not work for hotplug.

Or below qemu cmdline may help:
"-cpu host,host-phys-bits-limit=39"

don't you have the same issue with legacy VFIO code, you should?

I tend to be lazy and use seabios for guests on the command line.
I do see the error with legacy VFIO and uefi.

However, with the address space size work-around and iommufd, the
error is different, an EFAULT now. Some page pinning issue it seems.

Thanks,

C.