[PATCH v4 1/2] ARM: dma-mapping: Don't use outer_flush_range when the L2C is coherent

On Tue, Sep 29, 2015 at 07:30:31PM +0200, Thomas Petazzoni wrote:

Hello Russell,

On Tue, 29 Sep 2015 18:10:54 +0100, Russell King - ARM Linux wrote:

quoted

quoted

In the current kernel implementation, the outer cache flush range
operation is triggered by the dma_alloc function.
This operation can be take place during runtime and in some
circumstances may lead to the PCIe/PL310 deadlock on Armada 375/38x
SoCs.

I wonder if that's what's causing the sporadic lockups I'm seeing on
38x with a SATA PCIe card - it happens at a very specific point during
boot while initialising the SATA card, right down to the kernel message
character that it stops at.

It might very well be the case. If there's enough PCIe traffic and a
PL310 cache maintenance operation happening at the same time, the
system will lockup. I'm a bit surprised that just the initialization of
the PCIe card generates enough traffic to trigger the deadlock, but
maybe I'm underestimating the problem.

It isn't every boot - the board has booted around 240 kernels so far and
maybe 5% of them have needed the reset button pressed because of this.
It only happens when I have the SATA card connected.  I don't have any
drives on the SATA card at the moment though.

-- 
FTTC broadband for 0.8mile line: currently at 9.6Mbps down 400kbps up
according to speedtest.net.