rpmsg: virito: limit the buffer allocate from shared memory pool

[xiaoxiang781216]

rpmsg_virtio_get_tx_buffer shouldn't allocate the number of
buffer bigger than the virtio ring length of sending
report here: #162 (comment)

[arnopo]

This issue seems quite old and related to the first introduction of the zero copy that introduce index (b16ca55)...

Here I would be agree with @kr-satish #162 (comment):
"all buffers should be created at the time of initialization".

I think in a first step we have to figure out if adding new buffer in TX for master, if no more buffer is available, is a good strategy...
I had a look to Linux implementation and i don't see such approach...

@wjliang @MarekNovakNXP and @eanjum: I would like to require your memory for this topic...

• As you have the legacy, do you know the reason of such implementation?
• Do you think that it is still needed, or could we cleanup this implementation?

[xiaoxiang781216]

No, Linux use the similar strategy:
https://github.com/torvalds/linux/blob/master/drivers/rpmsg/virtio_rpmsg_bus.c#L447-L468
The difference is how to determine whether to allocate memory from the pool. Linux use last_sbuf, OpenAMP depend on the returned value from rpmsg_virtio_shm_pool_get_buffer before, now use vq_free_cnt.

[arnopo]

I re-parsed the code to better understood the impact of not limiting the allocation. I think you are right, allocating more buffer than the number of entry of the vrings is a bug.
And i don't see any code that try to extend the TX vrings...
So LGTM

Thanks,

[edmooring]

This looks good. We really shouldn't be trying to allocate more buffers than there are entries in the ring.