Re: [PATCH] MIPS: Calculate proper ebase value for 64-bit kernels

[Wu Zhangjin <wuzhangjin@gmail.com>]

On Tue, 2010-04-13 at 18:16 +0100, Ralf Baechle wrote:
> On Tue, Apr 13, 2010 at 09:34:38AM +0200, Thomas Bogendoerfer wrote:
> 
> > On Tue, Apr 13, 2010 at 01:03:54PM +0800, Wu Zhangjin wrote:
> > > This patch have broken the support to the MIPS variants whose
> > > cpu_has_mips_r2 is 0 for the CAC_BASE and CKSEG0 is completely different
> > > in these MIPSs.
> > 
> > I've checked R4k and R10k manulas and the exception base is at CKSEG0, so
> > about CPU we are talking ? And wouldn't it make for senso to have
> > an extra define for the exception base then ?
> 
> C0_ebase's design was a short-sigthed only considering 32-bit processors.
> So the exception base is in CKSEG0 on every 64-bit processor, be it R2 or
> older.  So yes, there is a bug as I've verified by testing but the patch
> is unfortunately incorrect.

Just debugged it via PMON:

loaded the kernel and used "g console=tty root=/dev/hda5 init=/bin/bash"
to start the kernel, there was a bad address exception.

the kernel stopped at:

Exception Cause=address error on store, SR=0x24000002, PC=0x8020526c
...
BADVADDR=0x97ffffff80000100, ENTHI=0xfffffe000
...
...
__copy_user+0x48  ... sd  t0,0(a0)  # addr = 0x80000100 rt=0x401a8000

Seems the a0 argument of __copy_user is _bad_.

And tried to set a break pointer to trap_init() and per_cpu_trap_init(),
and then cpu_cache_init() ... r4k_cache_init() and at last found that
set_uncached_handler(0x100, &except_vec2_generic, 0x80);

/*
 * Install uncached CPU exception handler.
 * This is suitable only for the cache error exception which is the only
 * exception handler that is being run uncached.
 */
void __cpuinit set_uncached_handler(unsigned long offset, void *addr,
        unsigned long size)
{
#ifdef CONFIG_32BIT
        unsigned long uncached_ebase = KSEG1ADDR(ebase);
#endif       
#ifdef CONFIG_64BIT
        unsigned long uncached_ebase = TO_UNCAC(ebase);
#endif       

        if (!addr)
                panic(panic_null_cerr);

        memcpy((void *)(uncached_ebase + offset), addr, size);
}

memcpy() called __copy_user... and the a0 is uncached_ebase + offset,
and uncached_ebase is defined by TO_UNCAC:

#define TO_UNCAC(x)             (UNCAC_BASE | ((x) & TO_PHYS_MASK))
#define TO_PHYS_MASK _CONST64_(0x07ffffffffffffff)
#define UNCAC_BASE _AC(0x9000000000000000, UL)

If using CKSEG0 as the ebase, CKSEG0 is defined as 0xffffffff80000000,
then we get the address: 0x97ffffff80000100, is this address ok?

And before, we have used the CAC_BASE as the ebase, the CAC_BASE is
defined as following:

#ifndef CAC_BASE
#ifdef CONFIG_DMA_NONCOHERENT
#define CAC_BASE                _AC(0x9800000000000000, UL)
#else
#define CAC_BASE                _AC(0xa800000000000000, UL)
#endif
#endif

So, before, the uncached_base is 0x9000000000000000.

Regards,
	Wu Zhangjin