Re: [PATCH 2/2] virtio: document virtio_config_ops restrictions

On Thu, 3 Jan 2019 18:28:49 +0100
Halil Pasic [off-list ref] wrote:

On Thu,  3 Jan 2019 17:08:04 +0100
Cornelia Huck [off-list ref] wrote:

quoted

Some transports (e.g. virtio-ccw) implement virtio operations that
seem to be a simple read/write as something more involved that
cannot be done from an atomic context.

Give at least a hint about that.

Signed-off-by: Cornelia Huck <cohuck@redhat.com>
---
 include/linux/virtio_config.h | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/include/linux/virtio_config.h b/include/linux/virtio_config.h
index 7087ef946ba7..987b6491b946 100644
--- a/include/linux/virtio_config.h
+++ b/include/linux/virtio_config.h

@@ -12,6 +12,11 @@ struct irq_affinity;
 
 /**
  * virtio_config_ops - operations for configuring a virtio device
+ * Note: Do not assume that a transport implements all of the operations
+ *       getting/setting a value as a simple read/write! Generally speaking,
+ *       any of @get/@set, @get_status/@set_status, or @get_features/
+ *       @finalize_features are NOT safe to be called from an atomic
+ *       context.  

I think the only exception is @bus_name (and maybe @generation, I don't
know) because it does not have to 'speak' with the hypervisor. If a
transport operation has to 'speak' with the hypervisor, we do it by
making it interpret a channel program. That means not safe to be called
form atomic context. Or am I missing something?

I explicitly singled out the listed callbacks because they read/write a
value and there might be more to them than meets the eye. I would
assume that nobody expects calling e.g. find_vqs (allocating queues)
from an atomic context to be a good idea :) Maybe I should do
s/Generally speaking/In particular/ ?

That said, it's only a hint; we should add might_sleep as well to
interfaces where it makes sense.