1 =========================================================
2 Cluster-wide Power-up/power-down race avoidance algorithm
3 =========================================================
5 This file documents the algorithm which is used to coordinate CPU and
6 cluster setup and teardown operations and to manage hardware coherency
9 The section "Rationale" explains what the algorithm is for and why it is
10 needed. "Basic model" explains general concepts using a simplified view
11 of the system. The other sections explain the actual details of the
18 In a system containing multiple CPUs, it is desirable to have the
19 ability to turn off individual CPUs when the system is idle, reducing
20 power consumption and thermal dissipation.
22 In a system containing multiple clusters of CPUs, it is also desirable
23 to have the ability to turn off entire clusters.
25 Turning entire clusters off and on is a risky business, because it
26 involves performing potentially destructive operations affecting a group
27 of independently running CPUs, while the OS continues to run. This
28 means that we need some coordination in order to ensure that critical
29 cluster-level operations are only performed when it is truly safe to do
32 Simple locking may not be sufficient to solve this problem, because
33 mechanisms like Linux spinlocks may rely on coherency mechanisms which
34 are not immediately enabled when a cluster powers up. Since enabling or
35 disabling those mechanisms may itself be a non-atomic operation (such as
36 writing some hardware registers and invalidating large caches), other
37 methods of coordination are required in order to guarantee safe
38 power-down and power-up at the cluster level.
40 The mechanism presented in this document describes a coherent memory
41 based protocol for performing the needed coordination. It aims to be as
42 lightweight as possible, while providing the required safety properties.
48 Each cluster and CPU is assigned a state, as follows:
57 +---------> UP ----------+
63 +--------- DOWN <--------+
67 The CPU or cluster is not coherent, and is either powered off or
68 suspended, or is ready to be powered off or suspended.
71 The CPU or cluster has committed to moving to the UP state.
72 It may be part way through the process of initialisation and
76 The CPU or cluster is active and coherent at the hardware
77 level. A CPU in this state is not necessarily being used
78 actively by the kernel.
81 The CPU or cluster has committed to moving to the DOWN
82 state. It may be part way through the process of teardown and
86 Each CPU has one of these states assigned to it at any point in time.
87 The CPU states are described in the "CPU state" section, below.
89 Each cluster is also assigned a state, but it is necessary to split the
90 state value into two parts (the "cluster" state and "inbound" state) and
91 to introduce additional states in order to avoid races between different
92 CPUs in the cluster simultaneously modifying the state. The cluster-
93 level states are described in the "Cluster state" section.
95 To help distinguish the CPU states from cluster states in this
96 discussion, the state names are given a `CPU_` prefix for the CPU states,
97 and a `CLUSTER_` or `INBOUND_` prefix for the cluster states.
103 In this algorithm, each individual core in a multi-core processor is
104 referred to as a "CPU". CPUs are assumed to be single-threaded:
105 therefore, a CPU can only be doing one thing at a single point in time.
107 This means that CPUs fit the basic model closely.
109 The algorithm defines the following states for each CPU in the system:
119 CPU setup complete policy decision
120 +-----------> CPU_UP ------------+
123 CPU_COMING_UP CPU_GOING_DOWN
126 +----------- CPU_DOWN <----------+
127 policy decision CPU teardown complete
131 The definitions of the four states correspond closely to the states of
134 Transitions between states occur as follows.
136 A trigger event (spontaneous) means that the CPU can transition to the
137 next state as a result of making local progress only, with no
138 requirement for any external event to happen.
142 A CPU reaches the CPU_DOWN state when it is ready for
143 power-down. On reaching this state, the CPU will typically
144 power itself down or suspend itself, via a WFI instruction or a
153 a) an explicit hardware power-up operation, resulting
154 from a policy decision on another CPU;
156 b) a hardware event, such as an interrupt.
160 A CPU cannot start participating in hardware coherency until the
161 cluster is set up and coherent. If the cluster is not ready,
162 then the CPU will wait in the CPU_COMING_UP state until the
163 cluster has been set up.
168 The CPU's parent cluster must be in CLUSTER_UP.
170 Transition of the parent cluster to CLUSTER_UP.
172 Refer to the "Cluster state" section for a description of the
177 When a CPU reaches the CPU_UP state, it is safe for the CPU to
178 start participating in local coherency.
180 This is done by jumping to the kernel's CPU resume code.
182 Note that the definition of this state is slightly different
183 from the basic model definition: CPU_UP does not mean that the
184 CPU is coherent yet, but it does mean that it is safe to resume
185 the kernel. The kernel handles the rest of the resume
186 procedure, so the remaining steps are not visible as part of the
187 race avoidance algorithm.
189 The CPU remains in this state until an explicit policy decision
190 is made to shut down or suspend the CPU.
197 explicit policy decision
201 While in this state, the CPU exits coherency, including any
202 operations required to achieve this (such as cleaning data
208 local CPU teardown complete
216 A cluster is a group of connected CPUs with some common resources.
217 Because a cluster contains multiple CPUs, it can be doing multiple
218 things at the same time. This has some implications. In particular, a
219 CPU can start up while another CPU is tearing the cluster down.
221 In this discussion, the "outbound side" is the view of the cluster state
222 as seen by a CPU tearing the cluster down. The "inbound side" is the
223 view of the cluster state as seen by a CPU setting the CPU up.
225 In order to enable safe coordination in such situations, it is important
226 that a CPU which is setting up the cluster can advertise its state
227 independently of the CPU which is tearing down the cluster. For this
228 reason, the cluster state is split into two parts:
230 "cluster" state: The global state of the cluster; or the state
231 on the outbound side:
237 "inbound" state: The state of the cluster on the inbound side.
239 - INBOUND_NOT_COMING_UP
243 The different pairings of these states results in six possible
244 states for the cluster as a whole::
247 +==========> INBOUND_NOT_COMING_UP -------------+
251 INBOUND_COMING_UP | v
253 ^ CLUSTER_GOING_DOWN CLUSTER_GOING_DOWN
254 # INBOUND_COMING_UP <=== INBOUND_NOT_COMING_UP
257 INBOUND_COMING_UP <----+ |
260 +=========== CLUSTER_DOWN <------------+
261 INBOUND_NOT_COMING_UP
263 Transitions -----> can only be made by the outbound CPU, and
264 only involve changes to the "cluster" state.
266 Transitions ===##> can only be made by the inbound CPU, and only
267 involve changes to the "inbound" state, except where there is no
268 further transition possible on the outbound side (i.e., the
269 outbound CPU has put the cluster into the CLUSTER_DOWN state).
271 The race avoidance algorithm does not provide a way to determine
272 which exact CPUs within the cluster play these roles. This must
273 be decided in advance by some other means. Refer to the section
274 "Last man and first man selection" for more explanation.
277 CLUSTER_DOWN/INBOUND_NOT_COMING_UP is the only state where the
278 cluster can actually be powered down.
280 The parallelism of the inbound and outbound CPUs is observed by
281 the existence of two different paths from CLUSTER_GOING_DOWN/
282 INBOUND_NOT_COMING_UP (corresponding to GOING_DOWN in the basic
283 model) to CLUSTER_DOWN/INBOUND_COMING_UP (corresponding to
284 COMING_UP in the basic model). The second path avoids cluster
287 CLUSTER_UP/INBOUND_COMING_UP is equivalent to UP in the basic
288 model. The final transition to CLUSTER_UP/INBOUND_NOT_COMING_UP
289 is trivial and merely resets the state machine ready for the
292 Details of the allowable transitions follow.
294 The next state in each case is notated
296 <cluster state>/<inbound state> (<transitioner>)
298 where the <transitioner> is the side on which the transition
299 can occur; either the inbound or the outbound side.
302 CLUSTER_DOWN/INBOUND_NOT_COMING_UP:
304 CLUSTER_DOWN/INBOUND_COMING_UP (inbound)
309 a) an explicit hardware power-up operation, resulting
310 from a policy decision on another CPU;
312 b) a hardware event, such as an interrupt.
315 CLUSTER_DOWN/INBOUND_COMING_UP:
317 In this state, an inbound CPU sets up the cluster, including
318 enabling of hardware coherency at the cluster level and any
319 other operations (such as cache invalidation) which are required
320 in order to achieve this.
322 The purpose of this state is to do sufficient cluster-level
323 setup to enable other CPUs in the cluster to enter coherency
327 CLUSTER_UP/INBOUND_COMING_UP (inbound)
329 cluster-level setup and hardware coherency complete
334 CLUSTER_UP/INBOUND_COMING_UP:
336 Cluster-level setup is complete and hardware coherency is
337 enabled for the cluster. Other CPUs in the cluster can safely
340 This is a transient state, leading immediately to
341 CLUSTER_UP/INBOUND_NOT_COMING_UP. All other CPUs on the cluster
342 should consider treat these two states as equivalent.
345 CLUSTER_UP/INBOUND_NOT_COMING_UP (inbound)
352 CLUSTER_UP/INBOUND_NOT_COMING_UP:
354 Cluster-level setup is complete and hardware coherency is
355 enabled for the cluster. Other CPUs in the cluster can safely
358 The cluster will remain in this state until a policy decision is
359 made to power the cluster down.
362 CLUSTER_GOING_DOWN/INBOUND_NOT_COMING_UP (outbound)
366 policy decision to power down the cluster
369 CLUSTER_GOING_DOWN/INBOUND_NOT_COMING_UP:
371 An outbound CPU is tearing the cluster down. The selected CPU
372 must wait in this state until all CPUs in the cluster are in the
375 When all CPUs are in the CPU_DOWN state, the cluster can be torn
376 down, for example by cleaning data caches and exiting
377 cluster-level coherency.
379 To avoid wasteful unnecessary teardown operations, the outbound
380 should check the inbound cluster state for asynchronous
381 transitions to INBOUND_COMING_UP. Alternatively, individual
382 CPUs can be checked for entry into CPU_COMING_UP or CPU_UP.
387 CLUSTER_DOWN/INBOUND_NOT_COMING_UP (outbound)
389 cluster torn down and ready to power off
393 CLUSTER_GOING_DOWN/INBOUND_COMING_UP (inbound)
398 a) an explicit hardware power-up operation,
399 resulting from a policy decision on another
402 b) a hardware event, such as an interrupt.
405 CLUSTER_GOING_DOWN/INBOUND_COMING_UP:
407 The cluster is (or was) being torn down, but another CPU has
408 come online in the meantime and is trying to set up the cluster
411 If the outbound CPU observes this state, it has two choices:
413 a) back out of teardown, restoring the cluster to the
416 b) finish tearing the cluster down and put the cluster
417 in the CLUSTER_DOWN state; the inbound CPU will
418 set up the cluster again from there.
420 Choice (a) permits the removal of some latency by avoiding
421 unnecessary teardown and setup operations in situations where
422 the cluster is not really going to be powered down.
427 CLUSTER_UP/INBOUND_COMING_UP (outbound)
429 cluster-level setup and hardware
435 CLUSTER_DOWN/INBOUND_COMING_UP (outbound)
437 cluster torn down and ready to power off
443 Last man and First man selection
444 --------------------------------
446 The CPU which performs cluster tear-down operations on the outbound side
447 is commonly referred to as the "last man".
449 The CPU which performs cluster setup on the inbound side is commonly
450 referred to as the "first man".
452 The race avoidance algorithm documented above does not provide a
453 mechanism to choose which CPUs should play these roles.
458 When shutting down the cluster, all the CPUs involved are initially
459 executing Linux and hence coherent. Therefore, ordinary spinlocks can
460 be used to select a last man safely, before the CPUs become
466 Because CPUs may power up asynchronously in response to external wake-up
467 events, a dynamic mechanism is needed to make sure that only one CPU
468 attempts to play the first man role and do the cluster-level
469 initialisation: any other CPUs must wait for this to complete before
472 Cluster-level initialisation may involve actions such as configuring
473 coherency controls in the bus fabric.
475 The current implementation in mcpm_head.S uses a separate mutual exclusion
476 mechanism to do this arbitration. This mechanism is documented in
477 detail in vlocks.txt.
480 Features and Limitations
481 ------------------------
485 The current ARM-based implementation is split between
486 arch/arm/common/mcpm_head.S (low-level inbound CPU operations) and
487 arch/arm/common/mcpm_entry.c (everything else):
489 __mcpm_cpu_going_down() signals the transition of a CPU to the
490 CPU_GOING_DOWN state.
492 __mcpm_cpu_down() signals the transition of a CPU to the CPU_DOWN
495 A CPU transitions to CPU_COMING_UP and then to CPU_UP via the
496 low-level power-up code in mcpm_head.S. This could
497 involve CPU-specific setup code, but in the current
498 implementation it does not.
500 __mcpm_outbound_enter_critical() and __mcpm_outbound_leave_critical()
501 handle transitions from CLUSTER_UP to CLUSTER_GOING_DOWN
502 and from there to CLUSTER_DOWN or back to CLUSTER_UP (in
503 the case of an aborted cluster power-down).
505 These functions are more complex than the __mcpm_cpu_*()
506 functions due to the extra inter-CPU coordination which
507 is needed for safe transitions at the cluster level.
509 A cluster transitions from CLUSTER_DOWN back to CLUSTER_UP via
510 the low-level power-up code in mcpm_head.S. This
511 typically involves platform-specific setup code,
512 provided by the platform-specific power_up_setup
513 function registered via mcpm_sync_init.
517 As currently described and implemented, the algorithm does not
518 support CPU topologies involving more than two levels (i.e.,
519 clusters of clusters are not supported). The algorithm could be
520 extended by replicating the cluster-level states for the
521 additional topological levels, and modifying the transition
522 rules for the intermediate (non-outermost) cluster levels.
528 Originally created and documented by Dave Martin for Linaro Limited, in
529 collaboration with Nicolas Pitre and Achin Gupta.
531 Copyright (C) 2012-2013 Linaro Limited
532 Distributed under the terms of Version 2 of the GNU General Public
533 License, as defined in linux/COPYING.