According to the virtio v1.x "entropy device" specification, a virtio-rng
device is supposed to always return at least one byte of entropy.
However the virtio v0.9 spec does not mention such a requirement.

The Arm Fixed Virtual Platform (FVP) implementation of virtio-rng always
returns 8 bytes less of entropy than requested. If 8 bytes or less are
requested, it will return 0 bytes.
This behaviour makes U-Boot's virtio_rng_read() implementation go into an
endless loop, hanging the system.

Work around this problem by always requesting 8 bytes more than needed,
but only if a previous call to virtqueue_get_buf() returned 0 bytes.

This should never trigger on a v1.x spec compliant implementation, but
fixes the hang on the Arm FVP.

Signed-off-by: Andre Przywara <andre.przyw...@arm.com>
Reported-by: Peter Hoyes <peter.ho...@arm.com>
---
 drivers/virtio/virtio_rng.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/drivers/virtio/virtio_rng.c b/drivers/virtio/virtio_rng.c
index b85545c2ee5..786359a6e36 100644
--- a/drivers/virtio/virtio_rng.c
+++ b/drivers/virtio/virtio_rng.c
@@ -20,7 +20,7 @@ struct virtio_rng_priv {
 static int virtio_rng_read(struct udevice *dev, void *data, size_t len)
 {
        int ret;
-       unsigned int rsize;
+       unsigned int rsize = 1;
        unsigned char buf[BUFFER_SIZE] __aligned(4);
        unsigned char *ptr = data;
        struct virtio_sg sg;
@@ -29,7 +29,12 @@ static int virtio_rng_read(struct udevice *dev, void *data, 
size_t len)
 
        while (len) {
                sg.addr = buf;
-               sg.length = min(len, sizeof(buf));
+               /*
+                * Work around implementations which always return 8 bytes
+                * less than requested, down to 0 bytes, which would
+                * cause an endless loop otherwise.
+                */
+               sg.length = min(rsize ? len : len + 8, sizeof(buf));
                sgs[0] = &sg;
 
                ret = virtqueue_add(priv->rng_vq, sgs, 0, 1);
-- 
2.25.1

Reply via email to