ARM: mcpm: introduce helpers for platform coherency exit/setup

This provides helper methods to coordinate between CPUs coming down
and CPUs going up, as well as documentation on the used algorithms,
so that cluster teardown and setup
operations are not done for a cluster simultaneously.

For use in the power_down() implementation:
  * __mcpm_cpu_going_down(unsigned int cluster, unsigned int cpu)
  * __mcpm_outbound_enter_critical(unsigned int cluster)
  * __mcpm_outbound_leave_critical(unsigned int cluster)
  * __mcpm_cpu_down(unsigned int cluster, unsigned int cpu)

The power_up_setup() helper should do platform-specific setup in
preparation for turning the CPU on, such as invalidating local caches
or entering coherency.  It must be assembler for now, since it must
run before the MMU can be switched on.  It is passed the affinity level
for which initialization should be performed.

Because the mcpm_sync_struct content is looked-up and modified
with the cache enabled or disabled depending on the code path, it is
crucial to always ensure proper cache maintenance to update main memory
right away.  The sync_cache_*() helpers are used to that end.

Also, in order to prevent a cached writer from interfering with an
adjacent non-cached writer, we ensure each state variable is located to
a separate cache line.

Thanks to Nicolas Pitre and Achin Gupta for the help with this
patch.

Signed-off-by: Dave Martin <dave.martin@linaro.org>
Signed-off-by: Nicolas Pitre <nico@linaro.org>
Reviewed-by: Will Deacon <will.deacon@arm.com>

5 files changed, 828 insertions, 2 deletions

diff --git a/Documentation/arm/cluster-pm-race-avoidance.txt b/Documentation/arm/cluster-pm-race-avoidance.txt
new file mode 100644
index 000000000000..750b6fc24af9
--- /dev/null
+++ b/Documentation/arm/cluster-pm-race-avoidance.txt
@@ -0,0 +1,498 @@
+Cluster-wide Power-up/power-down race avoidance algorithm
+=========================================================
+
+This file documents the algorithm which is used to coordinate CPU and
+cluster setup and teardown operations and to manage hardware coherency
+controls safely.
+
+The section "Rationale" explains what the algorithm is for and why it is
+needed.  "Basic model" explains general concepts using a simplified view
+of the system.  The other sections explain the actual details of the
+algorithm in use.
+
+
+Rationale
+---------
+
+In a system containing multiple CPUs, it is desirable to have the
+ability to turn off individual CPUs when the system is idle, reducing
+power consumption and thermal dissipation.
+
+In a system containing multiple clusters of CPUs, it is also desirable
+to have the ability to turn off entire clusters.
+
+Turning entire clusters off and on is a risky business, because it
+involves performing potentially destructive operations affecting a group
+of independently running CPUs, while the OS continues to run.  This
+means that we need some coordination in order to ensure that critical
+cluster-level operations are only performed when it is truly safe to do
+so.
+
+Simple locking may not be sufficient to solve this problem, because
+mechanisms like Linux spinlocks may rely on coherency mechanisms which
+are not immediately enabled when a cluster powers up.  Since enabling or
+disabling those mechanisms may itself be a non-atomic operation (such as
+writing some hardware registers and invalidating large caches), other
+methods of coordination are required in order to guarantee safe
+power-down and power-up at the cluster level.
+
+The mechanism presented in this document describes a coherent memory
+based protocol for performing the needed coordination.  It aims to be as
+lightweight as possible, while providing the required safety properties.
+
+
+Basic model
+-----------
+
+Each cluster and CPU is assigned a state, as follows:
+
+        DOWN
+        COMING_UP
+        UP
+        GOING_DOWN
+
+            +---------> UP ----------+
+            |                        v
+
+        COMING_UP                GOING_DOWN
+
+            ^                        |
+            +--------- DOWN <--------+
+
+
+DOWN:   The CPU or cluster is not coherent, and is either powered off or
+        suspended, or is ready to be powered off or suspended.
+
+COMING_UP: The CPU or cluster has committed to moving to the UP state.
+        It may be part way through the process of initialisation and
+        enabling coherency.
+
+UP:     The CPU or cluster is active and coherent at the hardware
+        level.  A CPU in this state is not necessarily being used
+        actively by the kernel.
+
+GOING_DOWN: The CPU or cluster has committed to moving to the DOWN
+        state.  It may be part way through the process of teardown and
+        coherency exit.
+
+
+Each CPU has one of these states assigned to it at any point in time.
+The CPU states are described in the "CPU state" section, below.
+
+Each cluster is also assigned a state, but it is necessary to split the
+state value into two parts (the "cluster" state and "inbound" state) and
+to introduce additional states in order to avoid races between different
+CPUs in the cluster simultaneously modifying the state.  The cluster-
+level states are described in the "Cluster state" section.
+
+To help distinguish the CPU states from cluster states in this
+discussion, the state names are given a CPU_ prefix for the CPU states,
+and a CLUSTER_ or INBOUND_ prefix for the cluster states.
+
+
+CPU state
+---------
+
+In this algorithm, each individual core in a multi-core processor is
+referred to as a "CPU".  CPUs are assumed to be single-threaded:
+therefore, a CPU can only be doing one thing at a single point in time.
+
+This means that CPUs fit the basic model closely.
+
+The algorithm defines the following states for each CPU in the system:
+
+        CPU_DOWN
+        CPU_COMING_UP
+        CPU_UP
+        CPU_GOING_DOWN
+
+         cluster setup and
+        CPU setup complete          policy decision
+              +-----------> CPU_UP ------------+
+              |                                v
+
+        CPU_COMING_UP                   CPU_GOING_DOWN
+
+              ^                                |
+              +----------- CPU_DOWN <----------+
+         policy decision           CPU teardown complete
+        or hardware event
+
+
+The definitions of the four states correspond closely to the states of
+the basic model.
+
+Transitions between states occur as follows.
+
+A trigger event (spontaneous) means that the CPU can transition to the
+next state as a result of making local progress only, with no
+requirement for any external event to happen.
+
+
+CPU_DOWN:
+
+        A CPU reaches the CPU_DOWN state when it is ready for
+        power-down.  On reaching this state, the CPU will typically
+        power itself down or suspend itself, via a WFI instruction or a
+        firmware call.
+
+        Next state:     CPU_COMING_UP
+        Conditions:     none
+
+        Trigger events:
+
+                a) an explicit hardware power-up operation, resulting
+                   from a policy decision on another CPU;
+
+                b) a hardware event, such as an interrupt.
+
+
+CPU_COMING_UP:
+
+        A CPU cannot start participating in hardware coherency until the
+        cluster is set up and coherent.  If the cluster is not ready,
+        then the CPU will wait in the CPU_COMING_UP state until the
+        cluster has been set up.
+
+        Next state:     CPU_UP
+        Conditions:     The CPU's parent cluster must be in CLUSTER_UP.
+        Trigger events: Transition of the parent cluster to CLUSTER_UP.
+
+        Refer to the "Cluster state" section for a description of the
+        CLUSTER_UP state.
+
+
+CPU_UP:
+        When a CPU reaches the CPU_UP state, it is safe for the CPU to
+        start participating in local coherency.
+
+        This is done by jumping to the kernel's CPU resume code.
+
+        Note that the definition of this state is slightly different
+        from the basic model definition: CPU_UP does not mean that the
+        CPU is coherent yet, but it does mean that it is safe to resume
+        the kernel.  The kernel handles the rest of the resume
+        procedure, so the remaining steps are not visible as part of the
+        race avoidance algorithm.
+
+        The CPU remains in this state until an explicit policy decision
+        is made to shut down or suspend the CPU.
+
+        Next state:     CPU_GOING_DOWN
+        Conditions:     none
+        Trigger events: explicit policy decision
+
+
+CPU_GOING_DOWN:
+
+        While in this state, the CPU exits coherency, including any
+        operations required to achieve this (such as cleaning data
+        caches).
+
+        Next state:     CPU_DOWN
+        Conditions:     local CPU teardown complete
+        Trigger events: (spontaneous)
+
+
+Cluster state
+-------------
+
+A cluster is a group of connected CPUs with some common resources.
+Because a cluster contains multiple CPUs, it can be doing multiple
+things at the same time.  This has some implications.  In particular, a
+CPU can start up while another CPU is tearing the cluster down.
+
+In this discussion, the "outbound side" is the view of the cluster state
+as seen by a CPU tearing the cluster down.  The "inbound side" is the
+view of the cluster state as seen by a CPU setting the CPU up.
+
+In order to enable safe coordination in such situations, it is important
+that a CPU which is setting up the cluster can advertise its state
+independently of the CPU which is tearing down the cluster.  For this
+reason, the cluster state is split into two parts:
+
+        "cluster" state: The global state of the cluster; or the state
+                on the outbound side:
+
+                CLUSTER_DOWN
+                CLUSTER_UP
+                CLUSTER_GOING_DOWN
+
+        "inbound" state: The state of the cluster on the inbound side.
+
+                INBOUND_NOT_COMING_UP
+                INBOUND_COMING_UP
+
+
+        The different pairings of these states results in six possible
+        states for the cluster as a whole:
+
+                                    CLUSTER_UP
+                  +==========> INBOUND_NOT_COMING_UP -------------+
+                  #                                               |
+                                                                  |
+             CLUSTER_UP     <----+                                |
+          INBOUND_COMING_UP      |                                v
+
+                  ^             CLUSTER_GOING_DOWN       CLUSTER_GOING_DOWN
+                  #              INBOUND_COMING_UP <=== INBOUND_NOT_COMING_UP
+
+            CLUSTER_DOWN         |                                |
+          INBOUND_COMING_UP <----+                                |
+                                                                  |
+                  ^                                               |
+                  +===========     CLUSTER_DOWN      <------------+
+                               INBOUND_NOT_COMING_UP
+
+        Transitions -----> can only be made by the outbound CPU, and
+        only involve changes to the "cluster" state.
+
+        Transitions ===##> can only be made by the inbound CPU, and only
+        involve changes to the "inbound" state, except where there is no
+        further transition possible on the outbound side (i.e., the
+        outbound CPU has put the cluster into the CLUSTER_DOWN state).
+
+        The race avoidance algorithm does not provide a way to determine
+        which exact CPUs within the cluster play these roles.  This must
+        be decided in advance by some other means.  Refer to the section
+        "Last man and first man selection" for more explanation.
+
+
+        CLUSTER_DOWN/INBOUND_NOT_COMING_UP is the only state where the
+        cluster can actually be powered down.
+
+        The parallelism of the inbound and outbound CPUs is observed by
+        the existence of two different paths from CLUSTER_GOING_DOWN/
+        INBOUND_NOT_COMING_UP (corresponding to GOING_DOWN in the basic
+        model) to CLUSTER_DOWN/INBOUND_COMING_UP (corresponding to
+        COMING_UP in the basic model).  The second path avoids cluster
+        teardown completely.
+
+        CLUSTER_UP/INBOUND_COMING_UP is equivalent to UP in the basic
+        model.  The final transition to CLUSTER_UP/INBOUND_NOT_COMING_UP
+        is trivial and merely resets the state machine ready for the
+        next cycle.
+
+        Details of the allowable transitions follow.
+
+        The next state in each case is notated
+
+                <cluster state>/<inbound state> (<transitioner>)
+
+        where the <transitioner> is the side on which the transition
+        can occur; either the inbound or the outbound side.
+
+
+CLUSTER_DOWN/INBOUND_NOT_COMING_UP:
+
+        Next state:     CLUSTER_DOWN/INBOUND_COMING_UP (inbound)
+        Conditions:     none
+        Trigger events:
+
+                a) an explicit hardware power-up operation, resulting
+                   from a policy decision on another CPU;
+
+                b) a hardware event, such as an interrupt.
+
+
+CLUSTER_DOWN/INBOUND_COMING_UP:
+
+        In this state, an inbound CPU sets up the cluster, including
+        enabling of hardware coherency at the cluster level and any
+        other operations (such as cache invalidation) which are required
+        in order to achieve this.
+
+        The purpose of this state is to do sufficient cluster-level
+        setup to enable other CPUs in the cluster to enter coherency
+        safely.
+
+        Next state:     CLUSTER_UP/INBOUND_COMING_UP (inbound)
+        Conditions:     cluster-level setup and hardware coherency complete
+        Trigger events: (spontaneous)
+
+
+CLUSTER_UP/INBOUND_COMING_UP:
+
+        Cluster-level setup is complete and hardware coherency is
+        enabled for the cluster.  Other CPUs in the cluster can safely
+        enter coherency.
+
+        This is a transient state, leading immediately to
+        CLUSTER_UP/INBOUND_NOT_COMING_UP.  All other CPUs on the cluster
+        should consider treat these two states as equivalent.
+
+        Next state:     CLUSTER_UP/INBOUND_NOT_COMING_UP (inbound)
+        Conditions:     none
+        Trigger events: (spontaneous)
+
+
+CLUSTER_UP/INBOUND_NOT_COMING_UP:
+
+        Cluster-level setup is complete and hardware coherency is
+        enabled for the cluster.  Other CPUs in the cluster can safely
+        enter coherency.
+
+        The cluster will remain in this state until a policy decision is
+        made to power the cluster down.
+
+        Next state:     CLUSTER_GOING_DOWN/INBOUND_NOT_COMING_UP (outbound)
+        Conditions:     none
+        Trigger events: policy decision to power down the cluster
+
+
+CLUSTER_GOING_DOWN/INBOUND_NOT_COMING_UP:
+
+        An outbound CPU is tearing the cluster down.  The selected CPU
+        must wait in this state until all CPUs in the cluster are in the
+        CPU_DOWN state.
+
+        When all CPUs are in the CPU_DOWN state, the cluster can be torn
+        down, for example by cleaning data caches and exiting
+        cluster-level coherency.
+
+        To avoid wasteful unnecessary teardown operations, the outbound
+        should check the inbound cluster state for asynchronous
+        transitions to INBOUND_COMING_UP.  Alternatively, individual
+        CPUs can be checked for entry into CPU_COMING_UP or CPU_UP.
+
+
+        Next states:
+
+        CLUSTER_DOWN/INBOUND_NOT_COMING_UP (outbound)
+                Conditions:     cluster torn down and ready to power off
+                Trigger events: (spontaneous)
+
+        CLUSTER_GOING_DOWN/INBOUND_COMING_UP (inbound)
+                Conditions:     none
+                Trigger events:
+
+                        a) an explicit hardware power-up operation,
+                           resulting from a policy decision on another
+                           CPU;
+
+                        b) a hardware event, such as an interrupt.
+
+
+CLUSTER_GOING_DOWN/INBOUND_COMING_UP:
+
+        The cluster is (or was) being torn down, but another CPU has
+        come online in the meantime and is trying to set up the cluster
+        again.
+
+        If the outbound CPU observes this state, it has two choices:
+
+                a) back out of teardown, restoring the cluster to the
+                   CLUSTER_UP state;
+
+                b) finish tearing the cluster down and put the cluster
+                   in the CLUSTER_DOWN state; the inbound CPU will
+                   set up the cluster again from there.
+
+        Choice (a) permits the removal of some latency by avoiding
+        unnecessary teardown and setup operations in situations where
+        the cluster is not really going to be powered down.
+
+
+        Next states:
+
+        CLUSTER_UP/INBOUND_COMING_UP (outbound)
+                Conditions:     cluster-level setup and hardware
+                                coherency complete
+                Trigger events: (spontaneous)
+
+        CLUSTER_DOWN/INBOUND_COMING_UP (outbound)
+                Conditions:     cluster torn down and ready to power off
+                Trigger events: (spontaneous)
+
+
+Last man and First man selection
+--------------------------------
+
+The CPU which performs cluster tear-down operations on the outbound side
+is commonly referred to as the "last man".
+
+The CPU which performs cluster setup on the inbound side is commonly
+referred to as the "first man".
+
+The race avoidance algorithm documented above does not provide a
+mechanism to choose which CPUs should play these roles.
+
+
+Last man:
+
+When shutting down the cluster, all the CPUs involved are initially
+executing Linux and hence coherent.  Therefore, ordinary spinlocks can
+be used to select a last man safely, before the CPUs become
+non-coherent.
+
+
+First man:
+
+Because CPUs may power up asynchronously in response to external wake-up
+events, a dynamic mechanism is needed to make sure that only one CPU
+attempts to play the first man role and do the cluster-level
+initialisation: any other CPUs must wait for this to complete before
+proceeding.
+
+Cluster-level initialisation may involve actions such as configuring
+coherency controls in the bus fabric.
+
+The current implementation in mcpm_head.S uses a separate mutual exclusion
+mechanism to do this arbitration.  This mechanism is documented in
+detail in vlocks.txt.
+
+
+Features and Limitations
+------------------------
+
+Implementation:
+
+        The current ARM-based implementation is split between
+        arch/arm/common/mcpm_head.S (low-level inbound CPU operations) and
+        arch/arm/common/mcpm_entry.c (everything else):
+
+        __mcpm_cpu_going_down() signals the transition of a CPU to the
+                CPU_GOING_DOWN state.
+
+        __mcpm_cpu_down() signals the transition of a CPU to the CPU_DOWN
+                state.
+
+        A CPU transitions to CPU_COMING_UP and then to CPU_UP via the
+                low-level power-up code in mcpm_head.S.  This could
+                involve CPU-specific setup code, but in the current
+                implementation it does not.
+
+        __mcpm_outbound_enter_critical() and __mcpm_outbound_leave_critical()
+                handle transitions from CLUSTER_UP to CLUSTER_GOING_DOWN
+                and from there to CLUSTER_DOWN or back to CLUSTER_UP (in
+                the case of an aborted cluster power-down).
+
+                These functions are more complex than the __mcpm_cpu_*()
+                functions due to the extra inter-CPU coordination which
+                is needed for safe transitions at the cluster level.
+
+        A cluster transitions from CLUSTER_DOWN back to CLUSTER_UP via
+                the low-level power-up code in mcpm_head.S.  This
+                typically involves platform-specific setup code,
+                provided by the platform-specific power_up_setup
+                function registered via mcpm_sync_init.
+
+Deep topologies:
+
+        As currently described and implemented, the algorithm does not
+        support CPU topologies involving more than two levels (i.e.,
+        clusters of clusters are not supported).  The algorithm could be
+        extended by replicating the cluster-level states for the
+        additional topological levels, and modifying the transition
+        rules for the intermediate (non-outermost) cluster levels.
+
+
+Colophon
+--------
+
+Originally created and documented by Dave Martin for Linaro Limited, in
+collaboration with Nicolas Pitre and Achin Gupta.
+
+Copyright (C) 2012-2013  Linaro Limited
+Distributed under the terms of Version 2 of the GNU General Public
+License, as defined in linux/COPYING.
diff --git a/arch/arm/common/mcpm_entry.c b/arch/arm/common/mcpm_entry.c
index 5d72889a58a4..370236dd1a03 100644
--- a/arch/arm/common/mcpm_entry.c
+++ b/arch/arm/common/mcpm_entry.c
@@ -16,6 +16,7 @@
 #include <asm/mcpm.h>
 #include <asm/cacheflush.h>
 #include <asm/idmap.h>
+#include <asm/cputype.h>
 
 extern unsigned long mcpm_entry_vectors[MAX_NR_CLUSTERS][MAX_CPUS_PER_CLUSTER];
 
@@ -111,3 +112,152 @@ int mcpm_cpu_powered_up(void)
                 platform_ops->powered_up();
         return 0;
 }
+
+struct sync_struct mcpm_sync;
+
+/*
+ * __mcpm_cpu_going_down: Indicates that the cpu is being torn down.
+ *    This must be called at the point of committing to teardown of a CPU.
+ *    The CPU cache (SCTRL.C bit) is expected to still be active.
+ */
+void __mcpm_cpu_going_down(unsigned int cpu, unsigned int cluster)
+{
+        mcpm_sync.clusters[cluster].cpus[cpu].cpu = CPU_GOING_DOWN;
+        sync_cache_w(&mcpm_sync.clusters[cluster].cpus[cpu].cpu);
+}
+
+/*
+ * __mcpm_cpu_down: Indicates that cpu teardown is complete and that the
+ *    cluster can be torn down without disrupting this CPU.
+ *    To avoid deadlocks, this must be called before a CPU is powered down.
+ *    The CPU cache (SCTRL.C bit) is expected to be off.
+ *    However L2 cache might or might not be active.
+ */
+void __mcpm_cpu_down(unsigned int cpu, unsigned int cluster)
+{
+        dmb();
+        mcpm_sync.clusters[cluster].cpus[cpu].cpu = CPU_DOWN;
+        sync_cache_w(&mcpm_sync.clusters[cluster].cpus[cpu].cpu);
+        dsb_sev();
+}
+
+/*
+ * __mcpm_outbound_leave_critical: Leave the cluster teardown critical section.
+ * @state: the final state of the cluster:
+ *     CLUSTER_UP: no destructive teardown was done and the cluster has been
+ *         restored to the previous state (CPU cache still active); or
+ *     CLUSTER_DOWN: the cluster has been torn-down, ready for power-off
+ *         (CPU cache disabled, L2 cache either enabled or disabled).
+ */
+void __mcpm_outbound_leave_critical(unsigned int cluster, int state)
+{
+        dmb();
+        mcpm_sync.clusters[cluster].cluster = state;
+        sync_cache_w(&mcpm_sync.clusters[cluster].cluster);
+        dsb_sev();
+}
+
+/*
+ * __mcpm_outbound_enter_critical: Enter the cluster teardown critical section.
+ * This function should be called by the last man, after local CPU teardown
+ * is complete.  CPU cache expected to be active.
+ *
+ * Returns:
+ *     false: the critical section was not entered because an inbound CPU was
+ *         observed, or the cluster is already being set up;
+ *     true: the critical section was entered: it is now safe to tear down the
+ *         cluster.
+ */
+bool __mcpm_outbound_enter_critical(unsigned int cpu, unsigned int cluster)
+{
+        unsigned int i;
+        struct mcpm_sync_struct *c = &mcpm_sync.clusters[cluster];
+
+        /* Warn inbound CPUs that the cluster is being torn down: */
+        c->cluster = CLUSTER_GOING_DOWN;
+        sync_cache_w(&c->cluster);
+
+        /* Back out if the inbound cluster is already in the critical region: */
+        sync_cache_r(&c->inbound);
+        if (c->inbound == INBOUND_COMING_UP)
+                goto abort;
+
+        /*
+         * Wait for all CPUs to get out of the GOING_DOWN state, so that local
+         * teardown is complete on each CPU before tearing down the cluster.
+         *
+         * If any CPU has been woken up again from the DOWN state, then we
+         * shouldn't be taking the cluster down at all: abort in that case.
+         */
+        sync_cache_r(&c->cpus);
+        for (i = 0; i < MAX_CPUS_PER_CLUSTER; i++) {
+                int cpustate;
+
+                if (i == cpu)
+                        continue;
+
+                while (1) {
+                        cpustate = c->cpus[i].cpu;
+                        if (cpustate != CPU_GOING_DOWN)
+                                break;
+
+                        wfe();
+                        sync_cache_r(&c->cpus[i].cpu);
+                }
+
+                switch (cpustate) {
+                case CPU_DOWN:
+                        continue;
+
+                default:
+                        goto abort;
+                }
+        }
+
+        return true;
+
+abort:
+        __mcpm_outbound_leave_critical(cluster, CLUSTER_UP);
+        return false;
+}
+
+int __mcpm_cluster_state(unsigned int cluster)
+{
+        sync_cache_r(&mcpm_sync.clusters[cluster].cluster);
+        return mcpm_sync.clusters[cluster].cluster;
+}
+
+extern unsigned long mcpm_power_up_setup_phys;
+
+int __init mcpm_sync_init(
+        void (*power_up_setup)(unsigned int affinity_level))
+{
+        unsigned int i, j, mpidr, this_cluster;
+
+        BUILD_BUG_ON(MCPM_SYNC_CLUSTER_SIZE * MAX_NR_CLUSTERS != sizeof mcpm_sync);
+        BUG_ON((unsigned long)&mcpm_sync & (__CACHE_WRITEBACK_GRANULE - 1));
+
+        /*
+         * Set initial CPU and cluster states.
+         * Only one cluster is assumed to be active at this point.
+         */
+        for (i = 0; i < MAX_NR_CLUSTERS; i++) {
+                mcpm_sync.clusters[i].cluster = CLUSTER_DOWN;
+                mcpm_sync.clusters[i].inbound = INBOUND_NOT_COMING_UP;
+                for (j = 0; j < MAX_CPUS_PER_CLUSTER; j++)
+                        mcpm_sync.clusters[i].cpus[j].cpu = CPU_DOWN;
+        }
+        mpidr = read_cpuid_mpidr();
+        this_cluster = MPIDR_AFFINITY_LEVEL(mpidr, 1);
+        for_each_online_cpu(i)
+                mcpm_sync.clusters[this_cluster].cpus[i].cpu = CPU_UP;
+        mcpm_sync.clusters[this_cluster].cluster = CLUSTER_UP;
+        sync_cache_w(&mcpm_sync);
+
+        if (power_up_setup) {
+                mcpm_power_up_setup_phys = virt_to_phys(power_up_setup);
+                sync_cache_w(&mcpm_power_up_setup_phys);
+        }
+
+        return 0;
+}
diff --git a/arch/arm/common/mcpm_head.S b/arch/arm/common/mcpm_head.S
index 68c9903075a9..7d729bd72674 100644
--- a/arch/arm/common/mcpm_head.S
+++ b/arch/arm/common/mcpm_head.S
@@ -7,11 +7,19 @@
  * This program is free software; you can redistribute it and/or modify
  * it under the terms of the GNU General Public License version 2 as
  * published by the Free Software Foundation.
+ *
+ *
+ * Refer to Documentation/arm/cluster-pm-race-avoidance.txt
+ * for details of the synchronisation algorithms used here.
  */
 
 #include <linux/linkage.h>
 #include <asm/mcpm.h>
 
+.if MCPM_SYNC_CLUSTER_CPUS
+.error "cpus must be the first member of struct mcpm_sync_struct"
+.endif
+
         .macro  pr_dbg  string
 #if defined(CONFIG_DEBUG_LL) && defined(DEBUG)
         b       1901f
@@ -57,24 +65,114 @@ ENTRY(mcpm_entry_point)
 2:      pr_dbg  "kernel mcpm_entry_point\n"
 
         /*
-         * MMU is off so we need to get to mcpm_entry_vectors in a
+         * MMU is off so we need to get to various variables in a
          * position independent way.
          */
         adr     r5, 3f
-        ldr     r6, [r5]
+        ldmia   r5, {r6, r7, r8}
         add     r6, r5, r6                      @ r6 = mcpm_entry_vectors
+        ldr     r7, [r5, r7]                    @ r7 = mcpm_power_up_setup_phys
+        add     r8, r5, r8                      @ r8 = mcpm_sync
+
+        mov     r0, #MCPM_SYNC_CLUSTER_SIZE
+        mla     r8, r0, r10, r8                 @ r8 = sync cluster base
+
+        @ Signal that this CPU is coming UP:
+        mov     r0, #CPU_COMING_UP
+        mov     r5, #MCPM_SYNC_CPU_SIZE
+        mla     r5, r9, r5, r8                  @ r5 = sync cpu address
+        strb    r0, [r5]
+
+        @ At this point, the cluster cannot unexpectedly enter the GOING_DOWN
+        @ state, because there is at least one active CPU (this CPU).
+
+        @ Note: the following is racy as another CPU might be testing
+        @ the same flag at the same moment.  That'll be fixed later.
+        ldrb    r0, [r8, #MCPM_SYNC_CLUSTER_CLUSTER]
+        cmp     r0, #CLUSTER_UP                 @ cluster already up?
+        bne     mcpm_setup                      @ if not, set up the cluster
+
+        @ Otherwise, skip setup:
+        b       mcpm_setup_complete
+
+mcpm_setup:
+        @ Control dependency implies strb not observable before previous ldrb.
+
+        @ Signal that the cluster is being brought up:
+        mov     r0, #INBOUND_COMING_UP
+        strb    r0, [r8, #MCPM_SYNC_CLUSTER_INBOUND]
+        dmb
+
+        @ Any CPU trying to take the cluster into CLUSTER_GOING_DOWN from this
+        @ point onwards will observe INBOUND_COMING_UP and abort.
+
+        @ Wait for any previously-pending cluster teardown operations to abort
+        @ or complete:
+mcpm_teardown_wait:
+        ldrb    r0, [r8, #MCPM_SYNC_CLUSTER_CLUSTER]
+        cmp     r0, #CLUSTER_GOING_DOWN
+        bne     first_man_setup
+        wfe
+        b       mcpm_teardown_wait
+
+first_man_setup:
+        dmb
+
+        @ If the outbound gave up before teardown started, skip cluster setup:
+
+        cmp     r0, #CLUSTER_UP
+        beq     mcpm_setup_leave
+
+        @ power_up_setup is now responsible for setting up the cluster:
+
+        cmp     r7, #0
+        mov     r0, #1          @ second (cluster) affinity level
+        blxne   r7              @ Call power_up_setup if defined
+        dmb
+
+        mov     r0, #CLUSTER_UP
+        strb    r0, [r8, #MCPM_SYNC_CLUSTER_CLUSTER]
+        dmb
+
+mcpm_setup_leave:
+        @ Leave the cluster setup critical section:
+
+        mov     r0, #INBOUND_NOT_COMING_UP
+        strb    r0, [r8, #MCPM_SYNC_CLUSTER_INBOUND]
+        dsb
+        sev
+
+mcpm_setup_complete:
+        @ If a platform-specific CPU setup hook is needed, it is
+        @ called from here.
+
+        cmp     r7, #0
+        mov     r0, #0          @ first (CPU) affinity level
+        blxne   r7              @ Call power_up_setup if defined
+        dmb
+
+        @ Mark the CPU as up:
+
+        mov     r0, #CPU_UP
+        strb    r0, [r5]
+
+        @ Observability order of CPU_UP and opening of the gate does not matter.
 
 mcpm_entry_gated:
         ldr     r5, [r6, r4, lsl #2]            @ r5 = CPU entry vector
         cmp     r5, #0
         wfeeq
         beq     mcpm_entry_gated
+        dmb
+
         pr_dbg  "released\n"
         bx      r5
 
         .align  2
 
 3:      .word   mcpm_entry_vectors - .
+        .word   mcpm_power_up_setup_phys - 3b
+        .word   mcpm_sync - 3b
 
 ENDPROC(mcpm_entry_point)
 
@@ -84,3 +182,7 @@ ENDPROC(mcpm_entry_point)
         .type   mcpm_entry_vectors, #object
 ENTRY(mcpm_entry_vectors)
         .space  4 * MAX_NR_CLUSTERS * MAX_CPUS_PER_CLUSTER
+
+        .type   mcpm_power_up_setup_phys, #object
+ENTRY(mcpm_power_up_setup_phys)
+        .space  4               @ set by mcpm_sync_init()
diff --git a/arch/arm/include/asm/mcpm.h b/arch/arm/include/asm/mcpm.h
index 627761fce780..3046e90210cb 100644
--- a/arch/arm/include/asm/mcpm.h
+++ b/arch/arm/include/asm/mcpm.h
@@ -24,6 +24,9 @@
 
 #ifndef __ASSEMBLY__
 
+#include <linux/types.h>
+#include <asm/cacheflush.h>
+
 /*
  * Platform specific code should use this symbol to set up secondary
  * entry location for processors to use when released from reset.
@@ -130,5 +133,75 @@ struct mcpm_platform_ops {
  */
 int __init mcpm_platform_register(const struct mcpm_platform_ops *ops);
 
+/* Synchronisation structures for coordinating safe cluster setup/teardown: */
+
+/*
+ * When modifying this structure, make sure you update the MCPM_SYNC_ defines
+ * to match.
+ */
+struct mcpm_sync_struct {
+        /* individual CPU states */
+        struct {
+                s8 cpu __aligned(__CACHE_WRITEBACK_GRANULE);
+        } cpus[MAX_CPUS_PER_CLUSTER];
+
+        /* cluster state */
+        s8 cluster __aligned(__CACHE_WRITEBACK_GRANULE);
+
+        /* inbound-side state */
+        s8 inbound __aligned(__CACHE_WRITEBACK_GRANULE);
+};
+
+struct sync_struct {
+        struct mcpm_sync_struct clusters[MAX_NR_CLUSTERS];
+};
+
+extern unsigned long sync_phys; /* physical address of *mcpm_sync */
+
+void __mcpm_cpu_going_down(unsigned int cpu, unsigned int cluster);
+void __mcpm_cpu_down(unsigned int cpu, unsigned int cluster);
+void __mcpm_outbound_leave_critical(unsigned int cluster, int state);
+bool __mcpm_outbound_enter_critical(unsigned int this_cpu, unsigned int cluster);
+int __mcpm_cluster_state(unsigned int cluster);
+
+int __init mcpm_sync_init(
+        void (*power_up_setup)(unsigned int affinity_level));
+
+#else
+
+/* 
+ * asm-offsets.h causes trouble when included in .c files, and cacheflush.h
+ * cannot be included in asm files.  Let's work around the conflict like this.
+ */
+#include <asm/asm-offsets.h>
+#define __CACHE_WRITEBACK_GRANULE CACHE_WRITEBACK_GRANULE
+
 #endif /* ! __ASSEMBLY__ */
+
+/* Definitions for mcpm_sync_struct */
+#define CPU_DOWN                0x11
+#define CPU_COMING_UP           0x12
+#define CPU_UP                  0x13
+#define CPU_GOING_DOWN          0x14
+
+#define CLUSTER_DOWN            0x21
+#define CLUSTER_UP              0x22
+#define CLUSTER_GOING_DOWN      0x23
+
+#define INBOUND_NOT_COMING_UP   0x31
+#define INBOUND_COMING_UP       0x32
+
+/*
+ * Offsets for the mcpm_sync_struct members, for use in asm.
+ * We don't want to make them global to the kernel via asm-offsets.c.
+ */
+#define MCPM_SYNC_CLUSTER_CPUS  0
+#define MCPM_SYNC_CPU_SIZE      __CACHE_WRITEBACK_GRANULE
+#define MCPM_SYNC_CLUSTER_CLUSTER \
+        (MCPM_SYNC_CLUSTER_CPUS + MCPM_SYNC_CPU_SIZE * MAX_CPUS_PER_CLUSTER)
+#define MCPM_SYNC_CLUSTER_INBOUND \
+        (MCPM_SYNC_CLUSTER_CLUSTER + __CACHE_WRITEBACK_GRANULE)
+#define MCPM_SYNC_CLUSTER_SIZE \
+        (MCPM_SYNC_CLUSTER_INBOUND + __CACHE_WRITEBACK_GRANULE)
+
 #endif
diff --git a/arch/arm/kernel/asm-offsets.c b/arch/arm/kernel/asm-offsets.c
index 923eec7105cf..1bed82a0a9e0 100644
--- a/arch/arm/kernel/asm-offsets.c
+++ b/arch/arm/kernel/asm-offsets.c
@@ -149,6 +149,9 @@ int main(void)
   DEFINE(DMA_BIDIRECTIONAL,     DMA_BIDIRECTIONAL);
   DEFINE(DMA_TO_DEVICE,         DMA_TO_DEVICE);
   DEFINE(DMA_FROM_DEVICE,       DMA_FROM_DEVICE);
+  BLANK();
+  DEFINE(CACHE_WRITEBACK_GRANULE, __CACHE_WRITEBACK_GRANULE);
+  BLANK();
 #ifdef CONFIG_KVM_ARM_HOST
   DEFINE(VCPU_KVM,              offsetof(struct kvm_vcpu, kvm));
   DEFINE(VCPU_MIDR,             offsetof(struct kvm_vcpu, arch.midr));