On Mon, Oct 04, 2021 at 10:17:56AM -0300, Jason Gunthorpe wrote:

> Shutdown is supposed to quiet the HW so it is not doing DMAs any
> more. This is basically an 'emergency' kind of path, the HW should be
> violently stopped if available - ie clearing the bus master bits on
> PCI, for instance.

> When something like kexec happens we need the machine to be in a state
> where random DMA's are not corrupting memory.

That's all well and good but there's no point in implementing something
half baked that's opening up a whole bunch of opportunities to crash the
system if more work comes in after it's half broken the device setup.  

> Due to the emergency sort of nature it is not appropriate to do
> locking complicated sorts of things like struct device unregistrations
> here.

That's just not what's actually implemented in a bunch of places, nor
something one would infer from the documentation ("Called at shut-down
to quiesce the device", no mention of emergency cases which I'd guess
would just be kdump) - there's a bunch of locks in shutdown paths, and
drivers on sleeping buses with shutdown callbacks.  Never mind the few
of them that use a shutdown callback to power the system down, though
that's a different thing and definitely abusing the API.  I would guess
that a good proportion of people implementing it are more worried about
clean system shutdown than they are about kdump.