On Wed, Feb 07, 2018 at 08:03:27PM +0100, Andrew Jones wrote:
> Before 031755db ("arm: enable vmalloc") we were allocating the
> queue with two pages of zeroed memory using memalign(), but
> afterwards with only one uninitialized page using alloc_pages().
> We can keep alloc_pages(), but we need two pages, and they need
> to be clean, otherwise QEMU gets angry when we attempt to migrate
> a unit test as the used vring index is corrupted by the page
> allocator's next page link.
>
> Signed-off-by: Andrew Jones <drjones@xxxxxxxxxx>
> ---
> lib/virtio-mmio.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/lib/virtio-mmio.c b/lib/virtio-mmio.c
> index e5e8f660b5cd..cbc9e6217bbe 100644
> --- a/lib/virtio-mmio.c
> +++ b/lib/virtio-mmio.c
> @@ -55,7 +55,8 @@ static struct virtqueue *vm_setup_vq(struct virtio_device *vdev,
>
> vq = calloc(1, sizeof(*vq));
> assert(VIRTIO_MMIO_QUEUE_SIZE_MIN <= 2*PAGE_SIZE);
> - queue = alloc_pages(1);
> + queue = alloc_pages(2);
Whoops. As usual my brain engages right after posting patches.
The 1 was ok here because alloc_pages() takes an order, not a
number. The memset below is still key to avoid the migration
bug though.
I'll send a v2 of this patch - mostly to fix the commit message.
drew
> + memset(queue, 0, 2*PAGE_SIZE);
> assert(vq && queue);
>
> writel(index, vm_dev->base + VIRTIO_MMIO_QUEUE_SEL);
> --
> 2.13.6
>
