Hi Jean-Philippe,
On 19/01/18 17:21, Jean-Philippe Brucker wrote:
> On 16/01/18 23:26, Auger Eric wrote:
> [...]
>>> + switch (mem->subtype) {
>>> + case VIRTIO_IOMMU_RESV_MEM_T_MSI:
>>> + region = iommu_alloc_resv_region(addr, size, prot,
>>> + IOMMU_RESV_MSI);
>> if (!region)
>> return -ENOMEM;
>>> + break;
>>> + case VIRTIO_IOMMU_RESV_MEM_T_RESERVED:
>>> + default:
>>> + region = iommu_alloc_resv_region(addr, size, 0,
>>> + IOMMU_RESV_RESERVED);
>> same.
>
> I'll add them, thanks
>
>> There is another issue related to the exclusion of iovas belonging to
>> reserved regions. Typically on x86, when attempting to run
>> virtio-blk-pci with iommu I eventually saw the driver using iova
>> belonging to the IOAPIC regions to map phys addr and this stalled qemu
>> with a drown trace:
>>
>> "virtio: bogus descriptor or out of resources"
>>
>> those regions need to be excluded from the iova allocator. This was
>> resolved by adding
>> if (iommu_dma_init_domain(domain,
>> vdev->viommu->geometry.aperture_start,
>> vdev->viommu->geometry.aperture_end,
>> dev))
>> in viommu_attach_dev()
>
> The most recent hack for x86 [1] does call iommu_dma_init_domain() in
> attach_dev(). Is it buggy?
Oh OK. Actually I have never used that branch and in my version the last
arg of iommu_dma_init_domain was NULL hence the issue. But I was
thinking more generally that care must be taken to exclude all those
regions.
Thanks
Eric
>
> We probably shouldn't call iommu_dma_init_domain() unconditionally
> (outside of CONFIG_X86 that is), since it's normally done by the arch
> (arch/arm64/mm/dma-mapping.c)
>
> Thanks,
> Jean
>
> [1]
> http://www.linux-arm.org/git?p=linux-jpb.git;a=commitdiff;h=e910e224b58712151dda06df595a53ff07edef63
> on branch virtio-iommu/v0.5-x86
>
