On Wed, Aug 18, 2021 at 5:26 AM Viresh Kumar viresh.kumar@linaro.org wrote:
On 18-08-21, 10:38, Jie Deng wrote:
From the perspective of specification，I think we can allow zero-length buffers in virtio.
we can use the len of descriptor to see if it is a zero-length buffer.
But for a specific implementation, I don't think "NULL" can be passed to this API.
There is a check "BUG_ON(!virt_addr_valid(buf))" in it.
I tried to look at implementations of virt_addr_valid() and it doesn't check for NULL specifically (for the ones I looked at). I haven't tested it though.
Though I am not sure what's better here, remove the need of sending buffer altogether, the way this patch proposed initially or what Michael has suggested. And all that to prevent just a single bit to be used in flags field, which will likely be used for more things later on.
I'd prefer your earlier approach.
My feeling is that changing the virtqueue code to allow zero-length buffers is more fragile than having something in the virtio-i2c code that has a special case for leaving out both read_buf and write_buf.
Even if we fix the linux virtqueue code to deal with zero-length NULL buffers, there are other implementations that will require similar workarounds for existing sanity checks.
Arnd