On 7/10/2015 9:50 PM, scottwood at freescale.com (Scott Wood) wrote:

On Fri, 2015-07-10 at 13:29 -0500, Pledge Roy-R01356 wrote:

quoted

quoted

  return in_be32((void *)bm + offset);

quoted

         ^
[...]/drivers/soc/fsl/qbman/bman.c: In function ‘__bm_out’:
[...]/drivers/soc/fsl/qbman/bman.c:172:2: error: implicit declaration
of function ‘out_be32’ [-Werror=implicit-function-declaration]
  out_be32((void *)bm + offset, val);

These PPCisms will need to be fixed.  LS1043A is an ARM chip with DPAA 
1.0.

LS1043 (ARM, Little Endian) support is still work in progress.  The patches 
for that are being tested now but are based on the SDK version of the 
driver and will need to be massaged in order to get them applied here.  Our 
plan of record is to add upstream support for this at a later time.

If you're already reworking this for upstream acceptance, why not fix the 
more obvious PPCisms now?

Other drivers are facing similar problems wrt. I/O accessors.
How should the problem be solved?

For caam driver (drivers/crypto/caam) we have the following combinations:
ARCH	CORE	CAAM		SoC, options
PPC	BIG	BIG		P4080 etc.
PPC	BIG	LITTLE		N/A
PPC	LITTLE	BIG		P4080 + CONFIG_CPU_LITTLE_ENDIAN
PPC	LITTLE	LITTLE		N/A
ARM	LITTLE	LITTLE		LS1021A, LS2085A etc.
ARM	LITTLE	BIG		LS1043A
ARM	BIG 	LITTLE		LS1021A + CONFIG_CPU_BIG_ENDIAN
ARM	BIG	BIG		LS1043A + CONFIG_CPU_BIG_ENDIAN

Not all need to be supported, but for e.g. we are being requested to run
on ARM in BE mode.

Existing MMIO accessors have one of the following drawbacks:
-readl/writel: on ARM do swapping (cpu <--> le), while on PPC are
bus-specific (for PCI)
-have weak semantics (no barriers / sequence points etc.), like __raw_*

Thus, it looks like drivers have to define their own MMIO accessors,
taking into account:
-core endianness (which is independent of arch)
-device endianness
-already-defined MMIO accessors availability, depending on arch
-semantics will have to match

Unfortunately, the code will look terrible. In each driver.

Thanks,
Horia

P.S. Sorry if the To/Cc list is screwed up, I've taken the message from
the archive.