The partition suspend sequence as specified in the platform
architecture requires that all active processor threads call
H_JOIN, which:

- suspends the calling thread until it is the target of
  an H_PROD; or
- immediately returns H_CONTINUE, if the calling thread is the last to
  call H_JOIN. This thread is expected to call ibm,suspend-me to
  completely suspend the partition.

Upon returning from ibm,suspend-me the calling thread must wake all
others using H_PROD.

rtas_ibm_suspend_me_unsafe() uses on_each_cpu() to implement this
protocol, but because of its synchronizing nature this is susceptible
to deadlock versus users of stop_machine() or other callers of
on_each_cpu().

Not only is stop_machine() intended for use cases like this, it
handles error propagation and allows us to keep the data shared
between CPUs minimal: a single atomic counter which ensures exactly
one CPU will wake the others from their joined states.

Switch the migration code to use stop_machine() and a less complex
local implementation of the H_JOIN/ibm,suspend-me logic, which
carries additional benefits:

- more informative error reporting, appropriately ratelimited
- resets the lockup detector / watchdog on resume to prevent lockup
  warnings when the OS has been suspended for a time exceeding the
  threshold.

Fixes: 91dc182c ("[PATCH] powerpc: special-case ibm,suspend-me RTAS call")
Signed-off-by: Nathan Lynch <nathanl@linux.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20201207215200.1785968-13-nathanl@linux.ibm.com