On 07/10/2014 04:09 PM, Andrew Morton wrote:
On Tue, 8 Jul 2014 20:37:56 -0700 Guenter Roeck <li...@roeck-us.net> wrote:
The existing mechanisms have a number of drawbacks. Typically only one scheme
to restart the system is supported (at least if arm_pm_restart is used).
At least in theory there can be mutliple means to restart the system, some of
which may be less desirable (for example one mechanism may only reset the CPU,
while another may reset the entire system).
So the callbacks need to be prioritized.
Using arm_pm_restart can also be
racy if the function pointer is set from a driver, as the driver may be in
the process of being unloaded when arm_pm_restart is called.
Using the reboot notifier is always racy, as it is unknown if and when
other functions using the reboot notifier have completed execution
by the time the watchdog fires.
To solve the problem, introduce a system restart notifier. This notifier
is expected to be called from the architecture specific machine_restart()
function. Drivers providing system restart functionality (such as the watchdog
drivers mentioned above) are expected to register with this notifier.
It's worth mentioning here that the notifier_block priority scheme is
used to address the problem which was identified in the previous
paragraph.
Ok.
If this scheme is to be successful we will need to set in place some
protocol for specifying how the priorities are managed. If someone
sits down and writes a new restart handler, how is that person to
decide how to prioritize it against other handlers, both present and
future?
Also, looking at the patches, you don't appear to have actually *used*
the prioritization - everything is left at zero. So we'll end up using
the most-recently-registered handler to restart the system. The
patches don't actually solve the problem which was identified in the
above paragraph?
The primary goal of this patch set was to provide a generic scheme for
registering restart handlers, and the ability to load and unload notifiers
without race conditions. Support for multiple restart handlers with
different priorities was a secondary objective. The conversions I did
so far are expected to be mutually exclusive, ie provide the one and
only means on a given architecture to restart the system. So I guess
you do have a point - the priorities for those notifiers should
probably be higher. Error on my part - I thought lower numbers would
have higher priority, but after looking into the code again that
is wrong.
To avoid making things too complicated, maybe it would make sense to
specify guidelines for notifier priorities, such as
0 - restart notifier of last resort, with least reset capabilities
128 - default; use if no other notifier is expected to be available
and/or if restart functionality is acceptable
255 - highest priority notifier which _must_ be used
Would that make sense and be acceptable ? In this context, I would then
set the notifier priorities for the callers in the patch set to 128.
Thanks,
Guenter
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
