When reading a packet from a host-to-guest ring buffer, there is no
memory barrier between reading the write index (to see if there is
a packet to read) and reading the contents of the packet. The Hyper-V
host uses store-release when updating the write index to ensure that
writes of the packet data are completed first. On the guest side,
the processor can reorder and read the packet data before the write
index, and sometimes get stale packet data. Getting such stale packet
data has been observed in a reproducible case in a VM on ARM64.
Fix this by using virt_load_acquire() to read the write index,
ensuring that reads of the packet data cannot be reordered
before it. Preventing such reordering is logically correct, and
with this change, getting stale data can no longer be reproduced.
Signed-off-by: Michael Kelley <mikelley@xxxxxxxxxxxxx>
---
drivers/hv/ring_buffer.c | 11 ++++++++++-
1 file changed, 10 insertions(+), 1 deletion(-)
diff --git a/drivers/hv/ring_buffer.c b/drivers/hv/ring_buffer.c
index 71efacb..3d215d9 100644
--- a/drivers/hv/ring_buffer.c
+++ b/drivers/hv/ring_buffer.c
@@ -439,7 +439,16 @@ int hv_ringbuffer_read(struct vmbus_channel *channel,
static u32 hv_pkt_iter_avail(const struct hv_ring_buffer_info *rbi)
{
u32 priv_read_loc = rbi->priv_read_index;
- u32 write_loc = READ_ONCE(rbi->ring_buffer->write_index);
+ u32 write_loc;
+
+ /*
+ * The Hyper-V host writes the packet data, then uses
+ * store_release() to update the write_index. Use load_acquire()
+ * here to prevent loads of the packet data from being re-ordered
+ * before the read of the write_index and potentially getting
+ * stale data.
+ */
+ write_loc = virt_load_acquire(&rbi->ring_buffer->write_index);
if (write_loc >= priv_read_loc)
return write_loc - priv_read_loc;
--
1.8.3.1
