Hi Vineet,
On Thu, 2016-03-10 at 05:05 +0000, Vineet Gupta wrote:
> +CC Noam
>
> On Wednesday 09 March 2016 10:51 PM, Lada Trimasova wrote:
> >
> > Memory access primitives should use cpu_to_le16, cpu_to_le32, le16_to_cpu
> > and le32_to_cpu because it is not really guaranteed that drivers handles
> > any ordering themselves.
> That is the driver issue. readxx as API simply returns data in native
> endianness.
> We've had EZChip running big endian and so far and they didn't need this
> change.
Let me disagree with you here.
See what is said in "include/asm-generic/io.h":
---------------------->8---------------------
/*
* __raw_{read,write}{b,w,l,q}() access memory in native endianness.
*
* On some architectures memory mapped IO needs to be accessed differently.
* On the simple architectures, we just read/write the memory location
* directly.
*/
...
/*
* {read,write}{b,w,l,q}() access little endian memory and return result in
* native endianness.
*/
---------------------->8---------------------
And that's an implementation we have for ARC:
---------------------->8---------------------
#define readb(c) ({ u8 __v = readb_relaxed(c); __iormb(); __v;
})
#define readw(c) ({ u16 __v = readw_relaxed(c); __iormb(); __v;
})
#define readl(c) ({ u32 __v = readl_relaxed(c); __iormb(); __v;
})
#define writeb(v,c) ({ __iowmb(); writeb_relaxed(v,c); })
#define writew(v,c) ({ __iowmb(); writew_relaxed(v,c); })
#define writel(v,c) ({ __iowmb(); writel_relaxed(v,c); })
/*
* Relaxed API for drivers which can handle any ordering themselves
*/
#define readb_relaxed(c) __raw_readb(c)
#define readw_relaxed(c) __raw_readw(c)
#define readl_relaxed(c) __raw_readl(c)
#define writeb_relaxed(v,c) __raw_writeb(v,c)
#define writew_relaxed(v,c) __raw_writew(v,c)
#define writel_relaxed(v,c) __raw_writel(v,c)
---------------------->8---------------------
Which is effectively (related to endianess discussion):
---------------------->8---------------------
#define readX(c) __raw_readX(c)
#define writeX(v,c) __raw_writeX(v,c)
---------------------->8---------------------
That looks IMHO incorrect if we read API description in
"include/asm-generic/io.h".
BTW description of {read,write}{b,w,l,q}() is a bit misleading in part saying
"... and return result in __native_endianness__".
But real implementation of {read,write}{b,w,l,q}() in "include/asm-generic/io.h"
really shows what was meant - note __leXX_to_cpu() and cpu_to_leXX are used.
> >
> > For example, serial port driver doesn't work when kernel is build for
> > arc big endian architecture.
> Last I tested Big Endian on SDP with 8250 part + 8250 driver it was working
> fine.
> I presume this is the systemC model for device and standard 8250 driver and
> very
> likely the model is not fixed endian or something.
Model is indeed little-endian. We build it only once and than changing
only "nsim_isa_big_endian" property (which changes only CPU core endianess) may
use
it equally well with little- and big-endian builds of Linux kernel.
-Alexey
