mbox series

[RFC,v2,0/2] New device-tree format and Opal based idle save-restore

Message ID 20181015091531.14126-1-akshay.adiga@linux.vnet.ibm.com
Headers show
Series New device-tree format and Opal based idle save-restore | expand

Message

Akshay Adiga Oct. 15, 2018, 9:15 a.m. UTC
Previously if a older kernel runs on a newer firmware, it may enable
all available states irrespective of its capability of handling it.
New device tree format adds a compatible flag, so that only kernel
which has the capability to handle the version of stop state will enable
it.

Older kernel will still see stop0 and stop0_lite in older format and we
will depricate it after some time.

1) Idea is to bump up the version string in firmware if we find a bug or
regression in stop states. A fix will be provided in linux which would
now know about the bumped up version of stop states, where as kernel
without fixes would ignore the states.

2) Slowly deprecate cpuidle/cpuhotplug threshold which is hard-coded
into cpuidle-powernv driver. Instead use compatible strings to indicate
if idle state is suitable for cpuidle and hotplug.

New idle state device tree format :
       power-mgt {
            ...
         ibm,enabled-stop-levels = <0xec000000>;
         ibm,cpu-idle-state-psscr-mask = <0x0 0x3003ff 0x0 0x3003ff>;
         ibm,cpu-idle-state-latencies-ns = <0x3e8 0x7d0>;
         ibm,cpu-idle-state-psscr = <0x0 0x330 0x0 0x300330>;
         ibm,cpu-idle-state-flags = <0x100000 0x101000>;
         ibm,cpu-idle-state-residency-ns = <0x2710 0x4e20>;
         ibm,idle-states {
                     stop4 {
                         flags = <0x207000>;
                         compatible = "ibm,state-v1",
				      "opal-support";
			 type = "cpuidle";
                         psscr-mask = <0x0 0x3003ff>;
                         handle = <0x102>;
                         latency-ns = <0x186a0>;
                         residency-ns = <0x989680>;
                         psscr = <0x0 0x300374>;
                  };
                    ...
                    stop11 {
                     ...
                         compatible = "ibm,state-v1",
				      "opal-support";
			 type = "cpuoffline";
                         ...
                  };
             };

High-level parsing algorithm :

Say Known version string = "ibm,state-v1"

for each stop state node in device tree:
	if (compatible has known version string)
		kernel takes care of stop-transitions
	else if (compatible has "opal-support")
		OPAL takes care of stop-transitions
	else
		Skip All deeper states

When a state does not have both version support and opal support,
Its possible to exit from a shallower state. Hence skipping all
deeper states.

OPAL support for idle states
----------------------------

With this patch series, all the states that loose hypervisor state
will be handled through opal_call.

Patch 3 adds support for Saving/restoring of SPRs and resync-timebase
in OPAL. Also all the decision making such as identifying first thread
in the core and taking locks before restoring, etc are implemented in
OPAL.

How does it work ?
-------------------

Consider a case that stop4 has a bug. We take the following steps to
mitigate the problem.

1) Change compatible string for stop4 in OPAL to "ibm-state-v2" and
remove "opal-supported". ship the new firmware.
The kernel ignores stop4 and all deeper states. But we will still have
shallower states. Prevents from completely disabling stop states.

2) Implement workaround in OPAL and add "opal-supported". Ship new firmware
The kernel uses opal for stop-transtion , which has workaround implemented.
We get stop4 and deeper states working without kernel changes and backports.
(and considerably less time)

3) Implement workaround in kernel and add "ibm-state-v2" as known versions
The kernel will now be able to handle stop4 and deeper states.

Changes from v1 :
 - Deep states will have only opal-support instead of both opal and kernel
 - Moved "cpuidle" and "cpuoffline" as a different property called
   "type"
 - All the decision making such as identifying first thread in the core
   and taking locks before restoring in such cases have also been moved
   to OPAL


Abhishek Goel (1):
  opal : Support spr save-restore using new

Akshay Adiga (1):
  SLW: Add new idle device-tree format

 asm/head.S              |   2 -
 hw/chiptod.c            |   2 +-
 hw/slw.c                | 306 ++++++++++++++++++++++++++++++++++++----
 include/cpu.h           |  28 ++++
 include/opal-api.h      |   4 +-
 include/opal-internal.h |   1 +
 include/processor.h     |  17 +++
 7 files changed, 329 insertions(+), 31 deletions(-)