The error state is purposefully racy as we expect it to be called at any
time and so have avoided any locking whilst capturing the crash dump.
However, with multi-engine GPUs and multiple CPUs, those races can
manifest into OOPSes as we attempt to chase dangling pointers freed on
other CPUs. Under discussion are lots of ways to slow down normal
operation in order to protect the post-mortem error capture, but what it
we take the opposite approach and freeze the machine whilst the error
capture runs (note the GPU may still running, but as long as we don't
process any of the results the driver's bookkeeping will be static).

Note that by of itself, this is not a complete fix. It also depends on
the compiler barriers in list_add/list_del to prevent traversing the
lists into the void. We also depend that we only require state from
carefully controlled sources - i.e. all the state we require for
post-mortem debugging should be reachable from the request itself so
that we only have to worry about retrieving the request carefully. Once
we have the request, we know that all pointers from it are intact.

v2: Avoid drm_clflush_pages() inside stop_machine() as it may use
stop_machine() itself for its wbinvd fallback.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/20161012090522.367-3-chris@chris-wilson.co.uk