Regression (ARM) arch/arm/mm/init.c doesn't build without CONFIG_ARM_PATCH_PHYS_VIRT.

On Monday 06 January 2014 05:39 PM, Russell King - ARM Linux wrote:

On Mon, Jan 06, 2014 at 08:28:20PM +0100, Krzysztof Ha?asa wrote:

quoted

Russell, Santosh,

the unneeded commit causing regression is still in place. Please try to
compile an ARM kernel without CONFIG_ARM_PATCH_PHYS_VIRT and with
CONFIG_ZONE_DMA and see for yourself, if you don't believe me.

Please be aware that this commit fixes nothing, its only function is
causing the regression - so we don't lose anything by reverting it.

If the attached wasn't clear, what the defective commit presently does
is changing a perfectly valid code into a code referencing a variable
which (without CONFIG_ARM_PATCH_PHYS_VIRT set) doesn't at all exist.

With CONFIG_ARM_PATCH_PHYS_VIRT set, this commit does precisely nothing.

Right, so, with Assabet, which has CONFIG_DMA_ZONE=y and
CONFIG_ARM_PATCH_PHYS_VIRT=y:

$ make O=../build/assabet arch/arm/mm/init.i

gives:
  arm_dma_limit = __pv_phys_offset + arm_dma_zone_size - 1;

with or without Santosh's patch.  So, with V2P patching in place, there's
absolutely no functional difference.

With CONFIG_ARM_PATCH_PHYS_VIRT=n, before Santosh's patch:

  arm_dma_limit = (0xc0000000UL) + arm_dma_zone_size - 1;

After:

  arm_dma_limit = __pv_phys_offset + arm_dma_zone_size - 1;

and this breaks the build because there is no __pv_phys_offset symbol.

Now, the case which matters for Santosh is the first case - the one
where CONFIG_ARM_PATCH_PHYS_VIRT=y.  We can clearly see that after
preprocessing, the results are 100% identical.

Therefore, I find myself agreeing with Krzysztof that the commit is
bad, has no functional change for the case it was proposed to solve,
and needs to be reverted.

May be I missed your point but I ended up creating the patch because
the CMA reservation was failing on Keystone since the arm_dma_limit
wasn't right since it wasn't considering the actual __pv_phys_offset.

Isn't that an issue ?

Regards,
Santosh