Comment 51 for bug 709245

Finally, Will Deacon figured out a workaround for the issue on arm code, see below upstream commit:
(CONFIG_PL310_ERRATA_769419 is needed for the fix)

commit 11ed0ba1754841316d4095478944300acf19acc3
Author: Will Deacon <email address hidden>
Date: Mon Nov 14 17:24:58 2011 +0100

    ARM: 7161/1: errata: no automatic store buffer drain

    This patch implements a workaround for PL310 erratum 769419. On
    revisions of the PL310 prior to r3p2, the Store Buffer does not
    automatically drain. This can cause normal, non-cacheable writes to be
    retained when the memory system is idle, leading to suboptimal I/O
    performance for drivers using coherent DMA.

    This patch adds an optional wmb() call to the cpu_idle loop. On systems
    with an outer cache, this causes an explicit flush of the store buffer.

    Cc: <email address hidden>
    Acked-by: Catalin Marinas <email address hidden>
    Tested-by: Marc Zyngier <email address hidden>
    Signed-off-by: Will Deacon <email address hidden>
    Signed-off-by: Russell King <email address hidden>

diff --git a/arch/arm/Kconfig b/arch/arm/Kconfig
index 44789ef..83aa746 100644
--- a/arch/arm/Kconfig
+++ b/arch/arm/Kconfig
@@ -1352,6 +1352,18 @@ config ARM_ERRATA_764369
    relevant cache maintenance functions and sets a specific bit
    in the diagnostic control register of the SCU.

+config PL310_ERRATA_769419
+ bool "PL310 errata: no automatic Store Buffer drain"
+ depends on CACHE_L2X0
+ help
+ On revisions of the PL310 prior to r3p2, the Store Buffer does
+ not automatically drain. This can cause normal, non-cacheable
+ writes to be retained when the memory system is idle, leading
+ to suboptimal I/O performance for drivers using coherent DMA.
+ This option adds a write barrier to the cpu_idle loop so that,
+ on systems with an outer cache, the store buffer is drained
+ explicitly.
+
 endmenu

 source "arch/arm/common/Kconfig"
diff --git a/arch/arm/kernel/process.c b/arch/arm/kernel/process.c
index 75316f0..3d0c6fb 100644
--- a/arch/arm/kernel/process.c
+++ b/arch/arm/kernel/process.c
@@ -192,6 +192,9 @@ void cpu_idle(void)
 #endif

    local_irq_disable();
+#ifdef CONFIG_PL310_ERRATA_769419
+ wmb();
+#endif
    if (hlt_counter) {
     local_irq_enable();
     cpu_relax();