[ARM] OMAP2/3 clock: convert remaining MPU barriers into OCP barriers
commitf11fda6a9173e8e6b152ba5cb26fa20095a4c60f
authorPaul Walmsley <paul@pwsan.com>
Wed, 28 Jan 2009 19:35:06 +0000 (28 12:35 -0700)
committerRussell King <rmk+kernel@arm.linux.org.uk>
Sun, 8 Feb 2009 17:50:45 +0000 (8 17:50 +0000)
tree32130aefa922d5ca9519f0a07cea238ef5000db5
parent439764cc18beb20ef409991e75e29b460db71d33
[ARM] OMAP2/3 clock: convert remaining MPU barriers into OCP barriers

Several parts of the OMAP2/3 clock code use wmb() to try to ensure
that the hardware write completes before continuing.  This approach is
problematic: wmb() only ensures that the write leaves the ARM.  It
does not ensure that the write actually reaches the endpoint device.
The endpoint device in this case - either the PRM, CM, or SCM - is
three interconnects away from the ARM - and the final interconnect is
low-speed.  And the OCP interconnects will post the write, and who
knows how long that will take to complete.  So the wmb() is not what
we want.  Worse, the wmb() is indiscriminate; it causes the ARM to
flush any other unrelated buffered writes and wait for the local
interconnect to acknowledge them - potentially very expensive.

Fix this by converting the wmb()s into readbacks of the same PRM/CM/SCM
register.  Since the PRM/CM/SCM devices use a single OCP thread, this
will cause the MPU to block while waiting for posted writes to that device
to complete.

linux-omap source commit is 260f5487848681b4d8ea7430a709a601bbcb21d1.

Signed-off-by: Paul Walmsley <paul@pwsan.com>
Signed-off-by: Tony Lindgren <tony@atomide.com>
Signed-off-by: Russell King <rmk+kernel@arm.linux.org.uk>
arch/arm/mach-omap2/clock.c