[PATCH] clocksource/drivers/pistachio: Fix wrong calculated clocksource read value

[jszhang@marvell.com (Jisheng Zhang)]

On Wed, 16 Dec 2015 10:38:03 +0000 Russell King - ARM Linux wrote:

> On Wed, Dec 16, 2015 at 11:32:17AM +0100, Daniel Lezcano wrote:
> > On 12/16/2015 10:33 AM, Russell King - ARM Linux wrote:  
> > >On Wed, Dec 16, 2015 at 10:21:55AM +0100, Daniel Lezcano wrote:  
> > >>On 12/16/2015 08:36 AM, Jisheng Zhang wrote:  
> > >>>And in fact, clocksource_mmio_readw_down() also has similar issue, but it masks
> > >>>with c->mask before return, the c->mask is less than 32 bit (because the
> > >>>clocksource_mmio_init think number of valid bits > 32 or < 16 is invalid.)
> > >>>the higher 32 bits are masked off, so we never saw such issue. But we'd better
> > >>>to fix that, what's your opinion?  
> > >>
> > >>I think we should have a look to this portion closely.  
> > >
> > >There is no need to return more bits than are specified.  If you have
> > >a N-bit counter, then the high (64-N)-bits can be any value, because:
> > >
> > >static inline cycle_t clocksource_delta(cycle_t now, cycle_t last, cycle_t mask)
> > >{
> > >         return (now - last) & mask;
> > >}

So the "& c->mask" in "~(cycle_t)readl_relaxed(to_mmio_clksrc(c)->reg) & c->mask;"

isn't needed, I'm not sure I understand this correctly.


> > >
> > >where 'now' is the current value returned from the clock source read
> > >function, 'last' is a previously returned value, and 'mask' is the
> > >bit mask.  This has the effect of ignoring the high order bits.  
> > 
> > I think this approach is perfectly sane. When I said we should look at this
> > portion closely, I meant we should double check the bitwise-nor order
> > regarding the explicit cast. The clocksource's mask makes sense and must
> > stay untouched.  
> 
> That's not my point.  Whether you do:
> 
> 	~(cycle_t)readl(...)
> 
> or
> 
> 	(cycle_t)~readl(...)
> 
> is irrelevant - the result is the same as far as the core code is
> concerned as it doesn't care about the higher order bits.
> 
> The only thing about which should be done is really which is faster
> in the general case, since this is a fast path in the time keeping
> code.
> 

Got it.

If there's no "& c->mask", just as the pistachio does, 

return (cycle_t)~readl_relaxed(to_mmio_clksrc(c)->reg)
  1c:   e1a0c00d        mov     ip, sp
  20:   e92dd800        push    {fp, ip, lr, pc}
  24:   e24cb004        sub     fp, ip, #4
  28:   e5103040        ldr     r3, [r0, #-64]  ; 0x40
  2c:   e5930000        ldr     r0, [r3]
  30:   e3a01000        mov     r1, #0
  34:   e1e00000        mvn     r0, r0
  38:   e89da800        ldm     sp, {fp, sp, pc}


is better than

return ~(cycle_t)readl_relaxed(to_mmio_clksrc(c)->reg);

  1c:   e1a0c00d        mov     ip, sp
  20:   e92dd800        push    {fp, ip, lr, pc}
  24:   e24cb004        sub     fp, ip, #4
  28:   e5103040        ldr     r3, [r0, #-64]  ; 0x40
  2c:   e5932000        ldr     r2, [r3]
  30:   e3a01000        mov     r1, #0
  34:   e1e00002        mvn     r0, r2
  38:   e1e01001        mvn     r1, r1
  3c:   e89da800        ldm     sp, {fp, sp, pc}

Thanks,
Jisheng