On 2018-08-16 12:45, Benjamin Herrenschmidt wrote:
On Thu, 2018-08-16 at 17:05 +1000, Benjamin Herrenschmidt wrote:
Those changes are utterly broken.
The basic premise of the design that we woudl do that unplug/replug
trick if and ONLY IF the driver doesn't have the appropriate
callbacks.
We are also now looking at replacing this with an ubind/re-bind
because
in practice, the unplugging is causing us all sort of problems. Sam
(CC) can elaborate.
Bjorn, we are the main authors of that spec (Linas wrote it under my
supervision) and created those callbacks for EEH. AER picked them up
only later. Those changes must be at the very least acked by us before
going upstream.
Also I had a quick look at the DPC spec, it looks like a subset of our
EEH capability.
Again, the code in Linux was written without concertation with us,
misunderstands/misuses the driver callbacks, does unplugs when it
shouldn't etc...
It's all very broken.
Please, at the very least revert the spec changes. They are utterly
wrong.
The driver MUST remain active during the recovery process *including*
fatal errors.
Only if the recovery fails and the driver gives us may you chose to
unplug the device (though there is little point).
What you have designed will work fine for network drivers but will not
work for storage.
Ben.
Hi Ben,
If you look at original DPC driver it was doing exactly the same as it
is doing now.
so with or without new changes DPC driver had the same behavior from
beginning.
commit 26e515713342b6f7c553aa3c66b21c6ab7cf82af
Regards,
Oza.
