Remove coherent memory from the BL memory maps

This patch extends the build option `USE_COHERENT_MEMORY` to
conditionally remove coherent memory from the memory maps of
all boot loader stages. The patch also adds necessary
documentation for coherent memory removal in firmware-design,
porting and user guides.

Fixes ARM-Software/tf-issues#106

Change-Id: I260e8768c6a5c2efc402f5804a80657d8ce38773

26 files changed, 463 insertions, 87 deletions

diff --git a/bl1/aarch64/bl1_entrypoint.S b/bl1/aarch64/bl1_entrypoint.S
index 82330c11..cfc62921 100644
--- a/bl1/aarch64/bl1_entrypoint.S
+++ b/bl1/aarch64/bl1_entrypoint.S
@@ -131,9 +131,11 @@ func bl1_entrypoint
 	ldr	x1, =__BSS_SIZE__
 	bl	zeromem16
 
+#if USE_COHERENT_MEM
 	ldr	x0, =__COHERENT_RAM_START__
 	ldr	x1, =__COHERENT_RAM_UNALIGNED_SIZE__
 	bl	zeromem16
+#endif
 
 	ldr	x0, =__DATA_RAM_START__
 	ldr	x1, =__DATA_ROM_START__
diff --git a/bl1/bl1.ld.S b/bl1/bl1.ld.S
index 007149b5..d682384a 100644
--- a/bl1/bl1.ld.S
+++ b/bl1/bl1.ld.S
@@ -107,6 +107,7 @@ SECTIONS
         *(xlat_table)
     } >RAM
 
+#if USE_COHERENT_MEM
     /*
      * The base address of the coherent memory section must be page-aligned (4K)
      * to guarantee that the coherent data are stored on their own pages and
@@ -125,6 +126,7 @@ SECTIONS
         . = NEXT(4096);
         __COHERENT_RAM_END__ = .;
     } >RAM
+#endif
 
     __BL1_RAM_START__ = ADDR(.data);
     __BL1_RAM_END__ = .;
@@ -140,8 +142,10 @@ SECTIONS
 
     __BSS_SIZE__ = SIZEOF(.bss);
 
+#if USE_COHERENT_MEM
     __COHERENT_RAM_UNALIGNED_SIZE__ =
         __COHERENT_RAM_END_UNALIGNED__ - __COHERENT_RAM_START__;
+#endif
 
     ASSERT(. <= BL1_RW_LIMIT, "BL1's RW section has exceeded its limit.")
 }
diff --git a/bl2/aarch64/bl2_entrypoint.S b/bl2/aarch64/bl2_entrypoint.S
index 2f058da9..499dc373 100644
--- a/bl2/aarch64/bl2_entrypoint.S
+++ b/bl2/aarch64/bl2_entrypoint.S
@@ -91,9 +91,11 @@ func bl2_entrypoint
 	ldr	x1, =__BSS_SIZE__
 	bl	zeromem16
 
+#if USE_COHERENT_MEM
 	ldr	x0, =__COHERENT_RAM_START__
 	ldr	x1, =__COHERENT_RAM_UNALIGNED_SIZE__
 	bl	zeromem16
+#endif
 
 	/* --------------------------------------------
 	 * Allocate a stack whose memory will be marked
diff --git a/bl2/bl2.ld.S b/bl2/bl2.ld.S
index 65304de7..99333391 100644
--- a/bl2/bl2.ld.S
+++ b/bl2/bl2.ld.S
@@ -93,6 +93,7 @@ SECTIONS
         *(xlat_table)
     } >RAM
 
+#if USE_COHERENT_MEM
     /*
      * The base address of the coherent memory section must be page-aligned (4K)
      * to guarantee that the coherent data are stored on their own pages and
@@ -111,12 +112,16 @@ SECTIONS
         . = NEXT(4096);
         __COHERENT_RAM_END__ = .;
     } >RAM
+#endif
 
     __BL2_END__ = .;
 
     __BSS_SIZE__ = SIZEOF(.bss);
+
+#if USE_COHERENT_MEM
     __COHERENT_RAM_UNALIGNED_SIZE__ =
         __COHERENT_RAM_END_UNALIGNED__ - __COHERENT_RAM_START__;
+#endif
 
     ASSERT(. <= BL2_LIMIT, "BL2 image has exceeded its limit.")
 }
diff --git a/bl31/aarch64/bl31_entrypoint.S b/bl31/aarch64/bl31_entrypoint.S
index 04063e1c..b786b29d 100644
--- a/bl31/aarch64/bl31_entrypoint.S
+++ b/bl31/aarch64/bl31_entrypoint.S
@@ -149,9 +149,11 @@ func bl31_entrypoint
 	ldr	x1, =__BSS_SIZE__
 	bl	zeromem16
 
+#if USE_COHERENT_MEM
 	ldr	x0, =__COHERENT_RAM_START__
 	ldr	x1, =__COHERENT_RAM_UNALIGNED_SIZE__
 	bl	zeromem16
+#endif
 
 	/* ---------------------------------------------
 	 * Initialize the cpu_ops pointer.
diff --git a/bl31/bl31.ld.S b/bl31/bl31.ld.S
index 124be85c..3327f316 100644
--- a/bl31/bl31.ld.S
+++ b/bl31/bl31.ld.S
@@ -117,6 +117,7 @@ SECTIONS
         *(xlat_table)
     } >RAM
 
+#if USE_COHERENT_MEM
     /*
      * The base address of the coherent memory section must be page-aligned (4K)
      * to guarantee that the coherent data are stored on their own pages and
@@ -135,12 +136,15 @@ SECTIONS
         . = NEXT(4096);
         __COHERENT_RAM_END__ = .;
     } >RAM
+#endif
 
     __BL31_END__ = .;
 
     __BSS_SIZE__ = SIZEOF(.bss);
+#if USE_COHERENT_MEM
     __COHERENT_RAM_UNALIGNED_SIZE__ =
         __COHERENT_RAM_END_UNALIGNED__ - __COHERENT_RAM_START__;
+#endif
 
     ASSERT(. <= BL31_LIMIT, "BL3-1 image has exceeded its limit.")
 }
diff --git a/bl32/tsp/aarch64/tsp_entrypoint.S b/bl32/tsp/aarch64/tsp_entrypoint.S
index 1cda1653..2714282b 100644
--- a/bl32/tsp/aarch64/tsp_entrypoint.S
+++ b/bl32/tsp/aarch64/tsp_entrypoint.S
@@ -108,9 +108,11 @@ func tsp_entrypoint
 	ldr	x1, =__BSS_SIZE__
 	bl	zeromem16
 
+#if USE_COHERENT_MEM
 	ldr	x0, =__COHERENT_RAM_START__
 	ldr	x1, =__COHERENT_RAM_UNALIGNED_SIZE__
 	bl	zeromem16
+#endif
 
 	/* --------------------------------------------
 	 * Allocate a stack whose memory will be marked
diff --git a/bl32/tsp/tsp.ld.S b/bl32/tsp/tsp.ld.S
index 5d7ffa11..d411ad02 100644
--- a/bl32/tsp/tsp.ld.S
+++ b/bl32/tsp/tsp.ld.S
@@ -98,6 +98,7 @@ SECTIONS
         *(xlat_table)
     } >RAM
 
+#if USE_COHERENT_MEM
     /*
      * The base address of the coherent memory section must be page-aligned (4K)
      * to guarantee that the coherent data are stored on their own pages and
@@ -116,12 +117,15 @@ SECTIONS
         . = NEXT(4096);
         __COHERENT_RAM_END__ = .;
     } >RAM
+#endif
 
     __BL32_END__ = .;
 
     __BSS_SIZE__ = SIZEOF(.bss);
+#if USE_COHERENT_MEM
     __COHERENT_RAM_UNALIGNED_SIZE__ =
         __COHERENT_RAM_END_UNALIGNED__ - __COHERENT_RAM_START__;
+#endif
 
     ASSERT(. <= BL32_LIMIT, "BL3-2 image has exceeded its limit.")
 }
diff --git a/bl32/tsp/tsp_main.c b/bl32/tsp/tsp_main.c
index 193ba29b..2eaca7c9 100644
--- a/bl32/tsp/tsp_main.c
+++ b/bl32/tsp/tsp_main.c
@@ -43,7 +43,7 @@
  * of trusted SRAM
  ******************************************************************************/
 extern unsigned long __RO_START__;
-extern unsigned long __COHERENT_RAM_END__;
+extern unsigned long __BL32_END__;
 
 /*******************************************************************************
  * Lock to control access to the console
@@ -63,11 +63,11 @@ work_statistics_t tsp_stats[PLATFORM_CORE_COUNT];
 
 /*******************************************************************************
  * The BL32 memory footprint starts with an RO sections and ends
- * with a section for coherent RAM. Use it to find the memory size
+ * with the linker symbol __BL32_END__. Use it to find the memory size
  ******************************************************************************/
 #define BL32_TOTAL_BASE (unsigned long)(&__RO_START__)
 
-#define BL32_TOTAL_LIMIT (unsigned long)(&__COHERENT_RAM_END__)
+#define BL32_TOTAL_LIMIT (unsigned long)(&__BL32_END__)
 
 static tsp_args_t *set_smc_args(uint64_t arg0,
 			     uint64_t arg1,
diff --git a/docs/firmware-design.md b/docs/firmware-design.md
index 41aaf7f2..774ea436 100644
--- a/docs/firmware-design.md
+++ b/docs/firmware-design.md
@@ -12,8 +12,9 @@ Contents :
 7.  [CPU specific operations framework](#7--cpu-specific-operations-framework)
 8.  [Memory layout of BL images](#8-memory-layout-of-bl-images)
 9.  [Firmware Image Package (FIP)](#9--firmware-image-package-fip)
-10. [Code Structure](#10--code-structure)
-11. [References](#11--references)
+10. [Use of coherent memory in Trusted Firmware](#10--use-of-coherent-memory-in-trusted-firmware)
+11. [Code Structure](#11--code-structure)
+12. [References](#12--references)
 
 
 1.  Introduction
@@ -368,10 +369,10 @@ level implementation of the generic timer through the memory mapped interface.
     `ON`; any other cluster is `OFF`. BL3-1 initializes the data structures that
     implement the state machine, including the locks that protect them. BL3-1
     accesses the state of a CPU or cluster immediately after reset and before
-    the MMU is enabled in the warm boot path. It is not currently possible to
-    use 'exclusive' based spinlocks, therefore BL3-1 uses locks based on
-    Lamport's Bakery algorithm instead. BL3-1 allocates these locks in device
-    memory. They are accessible irrespective of MMU state.
+    the data cache is enabled in the warm boot path. It is not currently
+    possible to use 'exclusive' based spinlocks, therefore BL3-1 uses locks
+    based on Lamport's Bakery algorithm instead. BL3-1 allocates these locks in
+    device memory by default.
 
 *   Runtime services initialization:
 
@@ -1127,9 +1128,10 @@ this purpose:
 * `__BSS_START__` This address must be aligned on a 16-byte boundary.
 * `__BSS_SIZE__`
 
-Similarly, the coherent memory section must be zero-initialised. Also, the MMU
-setup code needs to know the extents of this section to set the right memory
-attributes for it. The following linker symbols are defined for this purpose:
+Similarly, the coherent memory section (if enabled) must be zero-initialised.
+Also, the MMU setup code needs to know the extents of this section to set the
+right memory attributes for it. The following linker symbols are defined for
+this purpose:
 
 * `__COHERENT_RAM_START__` This address must be aligned on a page-size boundary.
 * `__COHERENT_RAM_END__` This address must be aligned on a page-size boundary.
@@ -1443,7 +1445,208 @@ Currently the FVP's policy only allows loading of a known set of images. The
 platform policy can be modified to allow additional images.
 
 
-10.  Code Structure
+10. Use of coherent memory in Trusted Firmware
+----------------------------------------------
+
+There might be loss of coherency when physical memory with mismatched
+shareability, cacheability and memory attributes is accessed by multiple CPUs
+(refer to section B2.9 of [ARM ARM] for more details). This possibility occurs
+in Trusted Firmware during power up/down sequences when coherency, MMU and
+caches are turned on/off incrementally.
+
+Trusted Firmware defines coherent memory as a region of memory with Device
+nGnRE attributes in the translation tables. The translation granule size in
+Trusted Firmware is 4KB. This is the smallest possible size of the coherent
+memory region.
+
+By default, all data structures which are susceptible to accesses with
+mismatched attributes from various CPUs are allocated in a coherent memory
+region (refer to section 2.1 of [Porting Guide]). The coherent memory region
+accesses are Outer Shareable, non-cacheable and they can be accessed
+with the Device nGnRE attributes when the MMU is turned on. Hence, at the
+expense of at least an extra page of memory, Trusted Firmware is able to work
+around coherency issues due to mismatched memory attributes.
+
+The alternative to the above approach is to allocate the susceptible data
+structures in Normal WriteBack WriteAllocate Inner shareable memory. This
+approach requires the data structures to be designed so that it is possible to
+work around the issue of mismatched memory attributes by performing software
+cache maintenance on them.
+
+### Disabling the use of coherent memory in Trusted Firmware
+
+It might be desirable to avoid the cost of allocating coherent memory on
+platforms which are memory constrained. Trusted Firmware enables inclusion of
+coherent memory in firmware images through the build flag `USE_COHERENT_MEM`.
+This flag is enabled by default. It can be disabled to choose the second
+approach described above.
+
+The below sections analyze the data structures allocated in the coherent memory
+region and the changes required to allocate them in normal memory.
+
+### PSCI Affinity map nodes
+
+The `psci_aff_map` data structure stores the hierarchial node information for
+each affinity level in the system including the PSCI states associated with them.
+By default, this data structure is allocated in the coherent memory region in
+the Trusted Firmware because it can be accessed by multiple CPUs, either with
+their caches enabled or disabled.
+
+	typedef struct aff_map_node {
+		unsigned long mpidr;
+		unsigned char ref_count;
+		unsigned char state;
+		unsigned char level;
+	#if USE_COHERENT_MEM
+		bakery_lock_t lock;
+	#else
+		unsigned char aff_map_index;
+	#endif
+	} aff_map_node_t;
+
+In order to move this data structure to normal memory, the use of each of its
+fields must be analyzed. Fields like `mpidr` and `level` are only written once
+during cold boot. Hence removing them from coherent memory involves only doing
+a clean and invalidate of the cache lines after these fields are written.
+
+The fields `state` and `ref_count` can be concurrently accessed by multiple
+CPUs in different cache states. A Lamport's Bakery lock is used to ensure mutual
+exlusion to these fields. As a result, it is possible to move these fields out
+of coherent memory by performing software cache maintenance on them. The field
+`lock` is the bakery lock data structure when `USE_COHERENT_MEM` is enabled.
+The `aff_map_index` is used to identify the bakery lock when `USE_COHERENT_MEM`
+is disabled.
+
+### Bakery lock data
+
+The bakery lock data structure `bakery_lock_t` is allocated in coherent memory
+and is accessed by multiple CPUs with mismatched attributes. `bakery_lock_t` is
+defined as follows:
+
+    typedef struct bakery_lock {
+        int owner;
+        volatile char entering[BAKERY_LOCK_MAX_CPUS];
+        volatile unsigned number[BAKERY_LOCK_MAX_CPUS];
+    } bakery_lock_t;
+
+It is a characteristic of Lamport's Bakery algorithm that the volatile per-CPU
+fields can be read by all CPUs but only written to by the owning CPU.
+
+Depending upon the data cache line size, the per-CPU fields of the
+`bakery_lock_t` structure for multiple CPUs may exist on a single cache line.
+These per-CPU fields can be read and written during lock contention by multiple
+CPUs with mismatched memory attributes. Since these fields are a part of the
+lock implementation, they do not have access to any other locking primitive to
+safeguard against the resulting coherency issues. As a result, simple software
+cache maintenance is not enough to allocate them in coherent memory. Consider
+the following example.
+
+CPU0 updates its per-CPU field with data cache enabled. This write updates a
+local cache line which contains a copy of the fields for other CPUs as well. Now
+CPU1 updates its per-CPU field of the `bakery_lock_t` structure with data cache
+disabled. CPU1 then issues a DCIVAC operation to invalidate any stale copies of
+its field in any other cache line in the system. This operation will invalidate
+the update made by CPU0 as well.
+
+To use bakery locks when `USE_COHERENT_MEM` is disabled, the lock data structure
+has been redesigned. The changes utilise the characteristic of Lamport's Bakery
+algorithm mentioned earlier. The per-CPU fields of the new lock structure are
+aligned such that they are allocated on separate cache lines. The per-CPU data
+framework in Trusted Firmware is used to achieve this. This enables software to
+perform software cache maintenance on the lock data structure without running
+into coherency issues associated with mismatched attributes.
+
+The per-CPU data framework enables consolidation of data structures on the
+fewest cache lines possible. This saves memory as compared to the scenario where
+each data structure is separately aligned to the cache line boundary to achieve
+the same effect.
+
+The bakery lock data structure `bakery_info_t` is defined for use when
+`USE_COHERENT_MEM` is disabled as follows:
+
+    typedef struct bakery_info {
+        /*
+         * The lock_data is a bit-field of 2 members:
+         * Bit[0]       : choosing. This field is set when the CPU is
+         *                choosing its bakery number.
+         * Bits[1 - 15] : number. This is the bakery number allocated.
+         */
+         volatile uint16_t lock_data;
+    } bakery_info_t;
+
+The `bakery_info_t` represents a single per-CPU field of one lock and
+the combination of corresponding `bakery_info_t` structures for all CPUs in the
+system represents the complete bakery lock. It is embedded in the per-CPU
+data framework `cpu_data` as shown below:
+
+      CPU0 cpu_data
+    ------------------
+    | ....           |
+    |----------------|
+    | `bakery_info_t`| <-- Lock_0 per-CPU field
+    |    Lock_0      |     for CPU0
+    |----------------|
+    | `bakery_info_t`| <-- Lock_1 per-CPU field
+    |    Lock_1      |     for CPU0
+    |----------------|
+    | ....           |
+    |----------------|
+    | `bakery_info_t`| <-- Lock_N per-CPU field
+    |    Lock_N      |     for CPU0
+    ------------------
+
+
+      CPU1 cpu_data
+    ------------------
+    | ....           |
+    |----------------|
+    | `bakery_info_t`| <-- Lock_0 per-CPU field
+    |    Lock_0      |     for CPU1
+    |----------------|
+    | `bakery_info_t`| <-- Lock_1 per-CPU field
+    |    Lock_1      |     for CPU1
+    |----------------|
+    | ....           |
+    |----------------|
+    | `bakery_info_t`| <-- Lock_N per-CPU field
+    |    Lock_N      |     for CPU1
+    ------------------
+
+Consider a system of 2 CPUs with 'N' bakery locks as shown above.  For an
+operation on Lock_N, the corresponding `bakery_info_t` in both CPU0 and CPU1
+`cpu_data` need to be fetched and appropriate cache operations need to be
+performed for each access.
+
+For multiple bakery locks, an array of `bakery_info_t` is declared in `cpu_data`
+and each lock is given an `id` to identify it in the array.
+
+### Non Functional Impact of removing coherent memory
+
+Removal of the coherent memory region leads to the additional software overhead
+of performing cache maintenance for the affected data structures. However, since
+the memory where the data structures are allocated is cacheable, the overhead is
+mostly mitigated by an increase in performance.
+
+There is however a performance impact for bakery locks, due to:
+*   Additional cache maintenance operations, and
+*   Multiple cache line reads for each lock operation, since the bakery locks
+    for each CPU are distributed across different cache lines.
+
+The implementation has been optimized to mimimize this additional overhead.
+Measurements indicate that when bakery locks are allocated in Normal memory, the
+minimum latency of acquiring a lock is on an average 3-4 micro seconds whereas
+in Device memory the same is 2 micro seconds. The measurements were done on the
+Juno ARM development platform.
+
+As mentioned earlier, almost a page of memory can be saved by disabling
+`USE_COHERENT_MEM`. Each platform needs to consider these trade-offs to decide
+whether coherent memory should be used. If a platform disables
+`USE_COHERENT_MEM` and needs to use bakery locks in the porting layer, it should
+reserve memory in `cpu_data` by defining the macro `PLAT_PCPU_DATA_SIZE` (see
+the [Porting Guide]). Refer to the reference platform code for examples.
+
+
+11.  Code Structure
 -------------------
 
 Trusted Firmware code is logically divided between the three boot loader
@@ -1488,7 +1691,7 @@ FDTs provide a description of the hardware platform and are used by the Linux
 kernel at boot time. These can be found in the `fdts` directory.
 
 
-11.  References
+12.  References
 ---------------
 
 1.  Trusted Board Boot Requirements CLIENT PDD (ARM DEN 0006B-5). Available
@@ -1504,7 +1707,7 @@ kernel at boot time. These can be found in the `fdts` directory.
 
 _Copyright (c) 2013-2014, ARM Limited and Contributors. All rights reserved._
 
-
+[ARM ARM]:          http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0487a.e/index.html "ARMv8-A Reference Manual (ARM DDI0487A.E)"
 [PSCI]:             http://infocenter.arm.com/help/topic/com.arm.doc.den0022b/index.html "Power State Coordination Interface PDD (ARM DEN 0022B.b)"
 [SMCCC]:            http://infocenter.arm.com/help/topic/com.arm.doc.den0028a/index.html "SMC Calling Convention PDD (ARM DEN 0028A)"
 [UUID]:             https://tools.ietf.org/rfc/rfc4122.txt "A Universally Unique IDentifier (UUID) URN Namespace"
diff --git a/docs/porting-guide.md b/docs/porting-guide.md
index 3855ca7b..3d5e66fb 100644
--- a/docs/porting-guide.md
+++ b/docs/porting-guide.md
@@ -63,11 +63,11 @@ mapped page tables, and enable both the instruction and data caches for each BL
 stage. In the ARM FVP port, each BL stage configures the MMU in its platform-
 specific architecture setup function, for example `blX_plat_arch_setup()`.
 
-Each platform must allocate a block of identity mapped secure memory with
-Device-nGnRE attributes aligned to page boundary (4K) for each BL stage. This
-memory is identified by the section name `tzfw_coherent_mem` so that its
-possible for the firmware to place variables in it using the following C code
-directive:
+If the build option `USE_COHERENT_MEM` is enabled, each platform must allocate a
+block of identity mapped secure memory with Device-nGnRE attributes aligned to
+page boundary (4K) for each BL stage. This memory is identified by the section
+name `tzfw_coherent_mem` so that its possible for the firmware to place
+variables in it using the following C code directive:
 
     __attribute__ ((section("tzfw_coherent_mem")))
 
@@ -246,6 +246,17 @@ must also be defined:
     entities than this value using `io_open()` will fail with
     IO_RESOURCES_EXHAUSTED.
 
+If the platform needs to allocate data within the per-cpu data framework in
+BL3-1, it should define the following macro. Currently this is only required if
+the platform decides not to use the coherent memory section by undefining the
+USE_COHERENT_MEM build flag. In this case, the framework allocates the required
+memory within the the per-cpu data to minimize wastage.
+
+*   **#define : PLAT_PCPU_DATA_SIZE**
+
+    Defines the memory (in bytes) to be reserved within the per-cpu data
+    structure for use by the platform layer.
+
 The following constants are optional. They should be defined when the platform
 memory layout implies some image overlaying like on FVP.
 
diff --git a/docs/user-guide.md b/docs/user-guide.md
index b33c4c0c..5ad44a89 100644
--- a/docs/user-guide.md
+++ b/docs/user-guide.md
@@ -245,6 +245,12 @@ performed.
     synchronous method) or 1 (BL3-2 is initialized using asynchronous method).
     Default is 0.
 
+*   `USE_COHERENT_MEM`: This flag determines whether to include the coherent
+    memory region in the BL memory map or not (see "Use of Coherent memory in
+    Trusted Firmware" section in [Firmware Design]). It can take the value 1
+    (Coherent memory region is included) or 0 (Coherent memory region is
+    excluded). Default is 1.
+
 #### FVP specific build options
 
 *   `FVP_TSP_RAM_LOCATION`: location of the TSP binary. Options:
diff --git a/plat/fvp/aarch64/fvp_common.c b/plat/fvp/aarch64/fvp_common.c
index 987f48f6..e20fe7d9 100644
--- a/plat/fvp/aarch64/fvp_common.c
+++ b/plat/fvp/aarch64/fvp_common.c
@@ -136,7 +136,8 @@ const unsigned int num_sec_irqs = sizeof(irq_sec_array) /
  * Macro generating the code for the function setting up the pagetables as per
  * the platform memory map & initialize the mmu, for the given exception level
  ******************************************************************************/
-#define DEFINE_CONFIGURE_MMU_EL(_el)					\
+#if USE_COHERENT_MEM
+#define DEFINE_CONFIGURE_MMU_EL(_el)				\
 	void fvp_configure_mmu_el##_el(unsigned long total_base,	\
 				   unsigned long total_size,		\
 				   unsigned long ro_start,		\
@@ -158,6 +159,25 @@ const unsigned int num_sec_irqs = sizeof(irq_sec_array) /
 									\
 		enable_mmu_el##_el(0);					\
 	}
+#else
+#define DEFINE_CONFIGURE_MMU_EL(_el)				\
+	void fvp_configure_mmu_el##_el(unsigned long total_base,	\
+				   unsigned long total_size,		\
+				   unsigned long ro_start,		\
+				   unsigned long ro_limit)		\
+	{								\
+		mmap_add_region(total_base, total_base,			\
+				total_size,				\
+				MT_MEMORY | MT_RW | MT_SECURE);		\
+		mmap_add_region(ro_start, ro_start,			\
+				ro_limit - ro_start,			\
+				MT_MEMORY | MT_RO | MT_SECURE);		\
+		mmap_add(fvp_mmap);					\
+		init_xlat_tables();					\
+									\
+		enable_mmu_el##_el(0);					\
+	}
+#endif
 
 /* Define EL1 and EL3 variants of the function initialising the MMU */
 DEFINE_CONFIGURE_MMU_EL(1)
diff --git a/plat/fvp/bl1_fvp_setup.c b/plat/fvp/bl1_fvp_setup.c
index b1205d43..4b421d71 100644
--- a/plat/fvp/bl1_fvp_setup.c
+++ b/plat/fvp/bl1_fvp_setup.c
@@ -40,6 +40,7 @@
 #include "fvp_def.h"
 #include "fvp_private.h"
 
+#if USE_COHERENT_MEM
 /*******************************************************************************
  * Declarations of linker defined symbols which will help us find the layout
  * of trusted SRAM
@@ -56,6 +57,7 @@ extern unsigned long __COHERENT_RAM_END__;
  */
 #define BL1_COHERENT_RAM_BASE (unsigned long)(&__COHERENT_RAM_START__)
 #define BL1_COHERENT_RAM_LIMIT (unsigned long)(&__COHERENT_RAM_END__)
+#endif
 
 /* Data structure which holds the extents of the trusted SRAM for BL1*/
 static meminfo_t bl1_tzram_layout;
@@ -116,9 +118,12 @@ void bl1_plat_arch_setup(void)
 	fvp_configure_mmu_el3(bl1_tzram_layout.total_base,
 			      bl1_tzram_layout.total_size,
 			      BL1_RO_BASE,
-			      BL1_RO_LIMIT,
-			      BL1_COHERENT_RAM_BASE,
-			      BL1_COHERENT_RAM_LIMIT);
+			      BL1_RO_LIMIT
+#if USE_COHERENT_MEM
+			      , BL1_COHERENT_RAM_BASE,
+			      BL1_COHERENT_RAM_LIMIT
+#endif
+			     );
 }
 
 
diff --git a/plat/fvp/bl2_fvp_setup.c b/plat/fvp/bl2_fvp_setup.c
index 67f89bc4..71bd8c2a 100644
--- a/plat/fvp/bl2_fvp_setup.c
+++ b/plat/fvp/bl2_fvp_setup.c
@@ -45,8 +45,10 @@
 extern unsigned long __RO_START__;
 extern unsigned long __RO_END__;
 
+#if USE_COHERENT_MEM
 extern unsigned long __COHERENT_RAM_START__;
 extern unsigned long __COHERENT_RAM_END__;
+#endif
 
 /*
  * The next 2 constants identify the extents of the code & RO data region.
@@ -57,6 +59,7 @@ extern unsigned long __COHERENT_RAM_END__;
 #define BL2_RO_BASE (unsigned long)(&__RO_START__)
 #define BL2_RO_LIMIT (unsigned long)(&__RO_END__)
 
+#if USE_COHERENT_MEM
 /*
  * The next 2 constants identify the extents of the coherent memory region.
  * These addresses are used by the MMU setup code and therefore they must be
@@ -66,11 +69,11 @@ extern unsigned long __COHERENT_RAM_END__;
  */
 #define BL2_COHERENT_RAM_BASE (unsigned long)(&__COHERENT_RAM_START__)
 #define BL2_COHERENT_RAM_LIMIT (unsigned long)(&__COHERENT_RAM_END__)
+#endif
 
 /* Data structure which holds the extents of the trusted SRAM for BL2 */
 static meminfo_t bl2_tzram_layout
-__attribute__ ((aligned(PLATFORM_CACHE_LINE_SIZE),
-		section("tzfw_coherent_mem")));
+__attribute__ ((aligned(PLATFORM_CACHE_LINE_SIZE)));
 
 /* Assert that BL3-1 parameters fit in shared memory */
 CASSERT((PARAMS_BASE + sizeof(bl2_to_bl31_params_mem_t)) <
@@ -209,9 +212,12 @@ void bl2_plat_arch_setup(void)
 	fvp_configure_mmu_el1(bl2_tzram_layout.total_base,
 			      bl2_tzram_layout.total_size,
 			      BL2_RO_BASE,
-			      BL2_RO_LIMIT,
-			      BL2_COHERENT_RAM_BASE,
-			      BL2_COHERENT_RAM_LIMIT);
+			      BL2_RO_LIMIT
+#if USE_COHERENT_MEM
+			      , BL2_COHERENT_RAM_BASE,
+			      BL2_COHERENT_RAM_LIMIT
+#endif
+			      );
 }
 
 /*******************************************************************************
diff --git a/plat/fvp/bl31_fvp_setup.c b/plat/fvp/bl31_fvp_setup.c
index 69efc9cf..3874413f 100644
--- a/plat/fvp/bl31_fvp_setup.c
+++ b/plat/fvp/bl31_fvp_setup.c
@@ -48,19 +48,25 @@
  ******************************************************************************/
 extern unsigned long __RO_START__;
 extern unsigned long __RO_END__;
+extern unsigned long __BL31_END__;
 
+#if USE_COHERENT_MEM
 extern unsigned long __COHERENT_RAM_START__;
 extern unsigned long __COHERENT_RAM_END__;
+#endif
 
 /*
- * The next 2 constants identify the extents of the code & RO data region.
- * These addresses are used by the MMU setup code and therefore they must be
- * page-aligned.  It is the responsibility of the linker script to ensure that
- * __RO_START__ and __RO_END__ linker symbols refer to page-aligned addresses.
+ * The next 3 constants identify the extents of the code, RO data region and the
+ * limit of the BL3-1 image.  These addresses are used by the MMU setup code and
+ * therefore they must be page-aligned.  It is the responsibility of the linker
+ * script to ensure that __RO_START__, __RO_END__ & __BL31_END__ linker symbols
+ * refer to page-aligned addresses.
  */
 #define BL31_RO_BASE (unsigned long)(&__RO_START__)
 #define BL31_RO_LIMIT (unsigned long)(&__RO_END__)
+#define BL31_END (unsigned long)(&__BL31_END__)
 
+#if USE_COHERENT_MEM
 /*
  * The next 2 constants identify the extents of the coherent memory region.
  * These addresses are used by the MMU setup code and therefore they must be
@@ -70,7 +76,7 @@ extern unsigned long __COHERENT_RAM_END__;
  */
 #define BL31_COHERENT_RAM_BASE (unsigned long)(&__COHERENT_RAM_START__)
 #define BL31_COHERENT_RAM_LIMIT (unsigned long)(&__COHERENT_RAM_END__)
-
+#endif
 
 #if RESET_TO_BL31
 static entry_point_info_t bl32_image_ep_info;
@@ -235,9 +241,12 @@ void bl31_plat_arch_setup(void)
 	fvp_cci_enable();
 #endif
 	fvp_configure_mmu_el3(BL31_RO_BASE,
-			      (BL31_COHERENT_RAM_LIMIT - BL31_RO_BASE),
+			      (BL31_END - BL31_RO_BASE),
 			      BL31_RO_BASE,
-			      BL31_RO_LIMIT,
-			      BL31_COHERENT_RAM_BASE,
-			      BL31_COHERENT_RAM_LIMIT);
+			      BL31_RO_LIMIT
+#if USE_COHERENT_MEM
+			      , BL31_COHERENT_RAM_BASE,
+			      BL31_COHERENT_RAM_LIMIT
+#endif
+			      );
 }
diff --git a/plat/fvp/fvp_private.h b/plat/fvp/fvp_private.h
index 6f1a637e..3949754b 100644
--- a/plat/fvp/fvp_private.h
+++ b/plat/fvp/fvp_private.h
@@ -118,15 +118,21 @@ CASSERT(PLAT_PCPU_DATA_SIZE == sizeof(fvp_cpu_data_t),	\
 void fvp_configure_mmu_el1(unsigned long total_base,
 			   unsigned long total_size,
 			   unsigned long,
-			   unsigned long,
-			   unsigned long,
-			   unsigned long);
+			   unsigned long
+#if USE_COHERENT_MEM
+			   , unsigned long,
+			   unsigned long
+#endif
+			   );
 void fvp_configure_mmu_el3(unsigned long total_base,
 			   unsigned long total_size,
 			   unsigned long,
-			   unsigned long,
-			   unsigned long,
-			   unsigned long);
+			   unsigned long
+#if USE_COHERENT_MEM
+			   , unsigned long,
+			   unsigned long
+#endif
+			   );
 
 int fvp_config_setup(void);
 
diff --git a/plat/fvp/tsp/tsp_fvp_setup.c b/plat/fvp/tsp/tsp_fvp_setup.c
index 301f6693..d8f46bd2 100644
--- a/plat/fvp/tsp/tsp_fvp_setup.c
+++ b/plat/fvp/tsp/tsp_fvp_setup.c
@@ -40,19 +40,25 @@
  ******************************************************************************/
 extern unsigned long __RO_START__;
 extern unsigned long __RO_END__;
+extern unsigned long __BL32_END__;
 
+#if USE_COHERENT_MEM
 extern unsigned long __COHERENT_RAM_START__;
 extern unsigned long __COHERENT_RAM_END__;
+#endif
 
 /*
- * The next 2 constants identify the extents of the code & RO data region.
- * These addresses are used by the MMU setup code and therefore they must be
- * page-aligned.  It is the responsibility of the linker script to ensure that
- * __RO_START__ and __RO_END__ linker symbols refer to page-aligned addresses.
+ * The next 3 constants identify the extents of the code & RO data region and
+ * the limit of the BL3-2 image. These addresses are used by the MMU setup code
+ * and therefore they must be page-aligned.  It is the responsibility of the
+ * linker script to ensure that __RO_START__, __RO_END__ & & __BL32_END__
+ * linker symbols refer to page-aligned addresses.
  */
 #define BL32_RO_BASE (unsigned long)(&__RO_START__)
 #define BL32_RO_LIMIT (unsigned long)(&__RO_END__)
+#define BL32_END (unsigned long)(&__BL32_END__)
 
+#if USE_COHERENT_MEM
 /*
  * The next 2 constants identify the extents of the coherent memory region.
  * These addresses are used by the MMU setup code and therefore they must be
@@ -62,6 +68,7 @@ extern unsigned long __COHERENT_RAM_END__;
  */
 #define BL32_COHERENT_RAM_BASE (unsigned long)(&__COHERENT_RAM_START__)
 #define BL32_COHERENT_RAM_LIMIT (unsigned long)(&__COHERENT_RAM_END__)
+#endif
 
 /*******************************************************************************
  * Initialize the UART
@@ -93,9 +100,12 @@ void tsp_platform_setup(void)
 void tsp_plat_arch_setup(void)
 {
 	fvp_configure_mmu_el1(BL32_RO_BASE,
-			      (BL32_COHERENT_RAM_LIMIT - BL32_RO_BASE),
+			      (BL32_END - BL32_RO_BASE),
 			      BL32_RO_BASE,
-			      BL32_RO_LIMIT,
-			      BL32_COHERENT_RAM_BASE,
-			      BL32_COHERENT_RAM_LIMIT);
+			      BL32_RO_LIMIT
+#if USE_COHERENT_MEM
+			      , BL32_COHERENT_RAM_BASE,
+			      BL32_COHERENT_RAM_LIMIT
+#endif
+			      );
 }
diff --git a/plat/juno/aarch64/juno_common.c b/plat/juno/aarch64/juno_common.c
index 8129b051..7ad40d0d 100644
--- a/plat/juno/aarch64/juno_common.c
+++ b/plat/juno/aarch64/juno_common.c
@@ -140,6 +140,7 @@ const unsigned int num_sec_irqs = sizeof(irq_sec_array) /
  * Macro generating the code for the function setting up the pagetables as per
  * the platform memory map & initialize the mmu, for the given exception level
  ******************************************************************************/
+#if USE_COHERENT_MEM
 #define DEFINE_CONFIGURE_MMU_EL(_el)				\
 	void configure_mmu_el##_el(unsigned long total_base,	\
 				  unsigned long total_size,	\
@@ -162,7 +163,25 @@ const unsigned int num_sec_irqs = sizeof(irq_sec_array) /
 								\
 	       enable_mmu_el##_el(0);				\
 	}
-
+#else
+#define DEFINE_CONFIGURE_MMU_EL(_el)				\
+	void configure_mmu_el##_el(unsigned long total_base,	\
+				  unsigned long total_size,	\
+				  unsigned long ro_start,	\
+				  unsigned long ro_limit)	\
+	{							\
+	       mmap_add_region(total_base, total_base,		\
+			       total_size,			\
+			       MT_MEMORY | MT_RW | MT_SECURE);	\
+	       mmap_add_region(ro_start, ro_start,		\
+			       ro_limit - ro_start,		\
+			       MT_MEMORY | MT_RO | MT_SECURE);	\
+	       mmap_add(juno_mmap);				\
+	       init_xlat_tables();				\
+								\
+	       enable_mmu_el##_el(0);				\
+	}
+#endif
 /* Define EL1 and EL3 variants of the function initialising the MMU */
 DEFINE_CONFIGURE_MMU_EL(1)
 DEFINE_CONFIGURE_MMU_EL(3)
diff --git a/plat/juno/bl1_plat_setup.c b/plat/juno/bl1_plat_setup.c
index e27e3948..23e8592b 100644
--- a/plat/juno/bl1_plat_setup.c
+++ b/plat/juno/bl1_plat_setup.c
@@ -41,6 +41,7 @@
 #include "juno_def.h"
 #include "juno_private.h"
 
+#if USE_COHERENT_MEM
 /*******************************************************************************
  * Declarations of linker defined symbols which will help us find the layout
  * of trusted RAM
@@ -57,6 +58,7 @@ extern unsigned long __COHERENT_RAM_END__;
  */
 #define BL1_COHERENT_RAM_BASE (unsigned long)(&__COHERENT_RAM_START__)
 #define BL1_COHERENT_RAM_LIMIT (unsigned long)(&__COHERENT_RAM_END__)
+#endif
 
 /* Data structure which holds the extents of the trusted RAM for BL1 */
 static meminfo_t bl1_tzram_layout;
@@ -189,9 +191,12 @@ void bl1_plat_arch_setup(void)
 	configure_mmu_el3(bl1_tzram_layout.total_base,
 			  bl1_tzram_layout.total_size,
 			  TZROM_BASE,
-			  TZROM_BASE + TZROM_SIZE,
-			  BL1_COHERENT_RAM_BASE,
-			  BL1_COHERENT_RAM_LIMIT);
+			  TZROM_BASE + TZROM_SIZE
+#if USE_COHERENT_MEM
+			  , BL1_COHERENT_RAM_BASE,
+			  BL1_COHERENT_RAM_LIMIT
+#endif
+			  );
 }
 
 /*******************************************************************************
diff --git a/plat/juno/bl2_plat_setup.c b/plat/juno/bl2_plat_setup.c
index 900a587f..8e7b2a0a 100644
--- a/plat/juno/bl2_plat_setup.c
+++ b/plat/juno/bl2_plat_setup.c
@@ -47,8 +47,10 @@
 extern unsigned long __RO_START__;
 extern unsigned long __RO_END__;
 
+#if USE_COHERENT_MEM
 extern unsigned long __COHERENT_RAM_START__;
 extern unsigned long __COHERENT_RAM_END__;
+#endif
 
 /*
  * The next 2 constants identify the extents of the code & RO data region.
@@ -59,6 +61,7 @@ extern unsigned long __COHERENT_RAM_END__;
 #define BL2_RO_BASE (unsigned long)(&__RO_START__)
 #define BL2_RO_LIMIT (unsigned long)(&__RO_END__)
 
+#if USE_COHERENT_MEM
 /*
  * The next 2 constants identify the extents of the coherent memory region.
  * These addresses are used by the MMU setup code and therefore they must be
@@ -68,11 +71,11 @@ extern unsigned long __COHERENT_RAM_END__;
  */
 #define BL2_COHERENT_RAM_BASE (unsigned long)(&__COHERENT_RAM_START__)
 #define BL2_COHERENT_RAM_LIMIT (unsigned long)(&__COHERENT_RAM_END__)
+#endif
 
 /* Data structure which holds the extents of the trusted RAM for BL2 */
 static meminfo_t bl2_tzram_layout
-__attribute__ ((aligned(PLATFORM_CACHE_LINE_SIZE),
-		section("tzfw_coherent_mem")));
+__attribute__ ((aligned(PLATFORM_CACHE_LINE_SIZE)));
 
 /*******************************************************************************
  * Structure which holds the arguments which need to be passed to BL3-1
@@ -194,9 +197,12 @@ void bl2_plat_arch_setup(void)
 	configure_mmu_el1(bl2_tzram_layout.total_base,
 			  bl2_tzram_layout.total_size,
 			  BL2_RO_BASE,
-			  BL2_RO_LIMIT,
-			  BL2_COHERENT_RAM_BASE,
-			  BL2_COHERENT_RAM_LIMIT);
+			  BL2_RO_LIMIT
+#if USE_COHERENT_MEM
+			  , BL2_COHERENT_RAM_BASE,
+			  BL2_COHERENT_RAM_LIMIT
+#endif
+			  );
 }
 
 /*******************************************************************************
diff --git a/plat/juno/bl31_plat_setup.c b/plat/juno/bl31_plat_setup.c
index c4504622..ad8ea435 100644
--- a/plat/juno/bl31_plat_setup.c
+++ b/plat/juno/bl31_plat_setup.c
@@ -48,19 +48,25 @@
  ******************************************************************************/
 extern unsigned long __RO_START__;
 extern unsigned long __RO_END__;
+extern unsigned long __BL31_END__;
 
+#if USE_COHERENT_MEM
 extern unsigned long __COHERENT_RAM_START__;
 extern unsigned long __COHERENT_RAM_END__;
+#endif
 
 /*
- * The next 2 constants identify the extents of the code & RO data region.
- * These addresses are used by the MMU setup code and therefore they must be
- * page-aligned.  It is the responsibility of the linker script to ensure that
- * __RO_START__ and __RO_END__ linker symbols refer to page-aligned addresses.
+ * The next 3 constants identify the extents of the code, RO data region and the
+ * limit of the BL3-1 image.  These addresses are used by the MMU setup code and
+ * therefore they must be page-aligned.  It is the responsibility of the linker
+ * script to ensure that __RO_START__, __RO_END__ & __BL31_END__ linker symbols
+ * refer to page-aligned addresses.
  */
 #define BL31_RO_BASE (unsigned long)(&__RO_START__)
 #define BL31_RO_LIMIT (unsigned long)(&__RO_END__)
+#define BL31_END (unsigned long)(&__BL31_END__)
 
+#if USE_COHERENT_MEM
 /*
  * The next 2 constants identify the extents of the coherent memory region.
  * These addresses are used by the MMU setup code and therefore they must be
@@ -70,6 +76,7 @@ extern unsigned long __COHERENT_RAM_END__;
  */
 #define BL31_COHERENT_RAM_BASE (unsigned long)(&__COHERENT_RAM_START__)
 #define BL31_COHERENT_RAM_LIMIT (unsigned long)(&__COHERENT_RAM_END__)
+#endif
 
 /******************************************************************************
  * Placeholder variables for copying the arguments that have been passed to
@@ -178,9 +185,13 @@ void bl31_platform_setup(void)
 void bl31_plat_arch_setup()
 {
 	configure_mmu_el3(BL31_RO_BASE,
-			  BL31_COHERENT_RAM_LIMIT - BL31_RO_BASE,
+			  (BL31_END - BL31_RO_BASE),
 			  BL31_RO_BASE,
-			  BL31_RO_LIMIT,
+			  BL31_RO_LIMIT
+#if USE_COHERENT_MEM
+			  ,
 			  BL31_COHERENT_RAM_BASE,
-			  BL31_COHERENT_RAM_LIMIT);
+			  BL31_COHERENT_RAM_LIMIT
+#endif
+			  );
 }
diff --git a/plat/juno/juno_private.h b/plat/juno/juno_private.h
index b7ef4488..70439e8b 100644
--- a/plat/juno/juno_private.h
+++ b/plat/juno/juno_private.h
@@ -134,15 +134,21 @@ unsigned int platform_get_core_pos(unsigned long mpidr);
 void configure_mmu_el1(unsigned long total_base,
 		       unsigned long total_size,
 		       unsigned long ro_start,
-		       unsigned long ro_limit,
-		       unsigned long coh_start,
-		       unsigned long coh_limit);
+		       unsigned long ro_limit
+#if USE_COHERENT_MEM
+		       , unsigned long coh_start,
+		       unsigned long coh_limit
+#endif
+		       );
 void configure_mmu_el3(unsigned long total_base,
 		       unsigned long total_size,
 		       unsigned long ro_start,
-		       unsigned long ro_limit,
-		       unsigned long coh_start,
-		       unsigned long coh_limit);
+		       unsigned long ro_limit
+#if USE_COHERENT_MEM
+		       , unsigned long coh_start,
+		       unsigned long coh_limit
+#endif
+		       );
 void plat_report_exception(unsigned long type);
 unsigned long plat_get_ns_image_entrypoint(void);
 unsigned long platform_get_stack(unsigned long mpidr);
diff --git a/plat/juno/tsp/tsp_plat_setup.c b/plat/juno/tsp/tsp_plat_setup.c
index 0a9d4cbe..8293a132 100644
--- a/plat/juno/tsp/tsp_plat_setup.c
+++ b/plat/juno/tsp/tsp_plat_setup.c
@@ -40,19 +40,25 @@
  ******************************************************************************/
 extern unsigned long __RO_START__;
 extern unsigned long __RO_END__;
+extern unsigned long __BL32_END__;
 
+#if USE_COHERENT_MEM
 extern unsigned long __COHERENT_RAM_START__;
 extern unsigned long __COHERENT_RAM_END__;
+#endif
 
 /*
- * The next 2 constants identify the extents of the code & RO data region.
- * These addresses are used by the MMU setup code and therefore they must be
- * page-aligned.  It is the responsibility of the linker script to ensure that
- * __RO_START__ and __RO_END__ linker symbols refer to page-aligned addresses.
+ * The next 3 constants identify the extents of the code, RO data region and the
+ * limit of the BL3-2 image.  These addresses are used by the MMU setup code and
+ * therefore they must be page-aligned.  It is the responsibility of the linker
+ * script to ensure that __RO_START__, __RO_END__ & __BL32_END__ linker symbols
+ * refer to page-aligned addresses.
  */
 #define BL32_RO_BASE (unsigned long)(&__RO_START__)
 #define BL32_RO_LIMIT (unsigned long)(&__RO_END__)
+#define BL32_END (unsigned long)(&__BL32_END__)
 
+#if USE_COHERENT_MEM
 /*
  * The next 2 constants identify the extents of the coherent memory region.
  * These addresses are used by the MMU setup code and therefore they must be
@@ -62,6 +68,7 @@ extern unsigned long __COHERENT_RAM_END__;
  */
 #define BL32_COHERENT_RAM_BASE (unsigned long)(&__COHERENT_RAM_START__)
 #define BL32_COHERENT_RAM_LIMIT (unsigned long)(&__COHERENT_RAM_END__)
+#endif
 
 /*******************************************************************************
  * Initialize the UART
@@ -90,9 +97,12 @@ void tsp_platform_setup(void)
 void tsp_plat_arch_setup(void)
 {
 	configure_mmu_el1(BL32_RO_BASE,
-			  BL32_COHERENT_RAM_LIMIT - BL32_RO_BASE,
+			  (BL32_END - BL32_RO_BASE),
 			  BL32_RO_BASE,
-			  BL32_RO_LIMIT,
-			  BL32_COHERENT_RAM_BASE,
-			  BL32_COHERENT_RAM_LIMIT);
+			  BL32_RO_LIMIT
+#if USE_COHERENT_MEM
+			  , BL32_COHERENT_RAM_BASE,
+			  BL32_COHERENT_RAM_LIMIT
+#endif
+			  );
 }
diff --git a/services/std_svc/psci/psci_common.c b/services/std_svc/psci/psci_common.c
index c984e9ec..0a1cdf9e 100644
--- a/services/std_svc/psci/psci_common.c
+++ b/services/std_svc/psci/psci_common.c
@@ -51,7 +51,10 @@ const spd_pm_ops_t *psci_spd_pm;
  * corresponds to an affinity instance e.g. cluster, cpu within an mpidr
  ******************************************************************************/
 aff_map_node_t psci_aff_map[PSCI_NUM_AFFS]
-__attribute__ ((section("tzfw_coherent_mem")));
+#if USE_COHERENT_MEM
+__attribute__ ((section("tzfw_coherent_mem")))
+#endif
+;
 
 /*******************************************************************************
  * Pointer to functions exported by the platform to complete power mgmt. ops
@@ -352,6 +355,10 @@ int psci_save_ns_entry(uint64_t mpidr,
  ******************************************************************************/
 unsigned short psci_get_state(aff_map_node_t *node)
 {
+#if !USE_COHERENT_MEM
+	flush_dcache_range((uint64_t) node, sizeof(*node));
+#endif
+
 	assert(node->level >= MPIDR_AFFLVL0 && node->level <= MPIDR_MAX_AFFLVL);
 
 	/* A cpu node just contains the state which can be directly returned */
@@ -409,6 +416,10 @@ void psci_set_state(aff_map_node_t *node, unsigned short state)
 		node->state &= ~(PSCI_STATE_MASK << PSCI_STATE_SHIFT);
 		node->state |= (state & PSCI_STATE_MASK) << PSCI_STATE_SHIFT;
 	}
+
+#if !USE_COHERENT_MEM
+	flush_dcache_range((uint64_t) node, sizeof(*node));
+#endif
 }
 
 /*******************************************************************************
diff --git a/services/std_svc/psci/psci_setup.c b/services/std_svc/psci/psci_setup.c
index a750256e..be504e81 100644
--- a/services/std_svc/psci/psci_setup.c
+++ b/services/std_svc/psci/psci_setup.c
@@ -331,13 +331,20 @@ int32_t psci_setup(void)
 					       afflvl);
 	}
 
+#if !USE_COHERENT_MEM
+	/*
+	 * The psci_aff_map only needs flushing when it's not allocated in
+	 * coherent memory.
+	 */
+	flush_dcache_range((uint64_t) &psci_aff_map, sizeof(psci_aff_map));
+#endif
+
 	/*
 	 * Set the bounds for the affinity counts of each level in the map. Also
 	 * flush out the entire array so that it's visible to subsequent power
-	 * management operations. The 'psci_aff_map' array is allocated in
-	 * coherent memory so does not need flushing. The 'psci_aff_limits'
-	 * array is allocated in normal memory. It will be accessed when the mmu
-	 * is off e.g. after reset. Hence it needs to be flushed.
+	 * management operations. The 'psci_aff_limits' array is allocated in
+	 * normal memory. It will be accessed when the mmu is off e.g. after
+	 * reset. Hence it needs to be flushed.
 	 */
 	for (afflvl = MPIDR_AFFLVL0; afflvl < max_afflvl; afflvl++) {
 		psci_aff_limits[afflvl].min =