Re: [Qemu-ppc] [Qemu-devel] [PATCHv2 01/12] cuda: do not use old_mmio accesses

[David Gibson]

On Fri, Feb 09, 2018 at 05:05:23PM -0300, Philippe Mathieu-Daudé wrote:
> Hi Mark,
> 
> On 02/09/2018 03:51 PM, Mark Cave-Ayland wrote:
> > Signed-off-by: Mark Cave-Ayland <address@hidden>
> > ---
> >  hw/misc/macio/cuda.c | 40 ++++++++--------------------------------
> >  1 file changed, 8 insertions(+), 32 deletions(-)
> > 
> > diff --git a/hw/misc/macio/cuda.c b/hw/misc/macio/cuda.c
> > index 008d8bd4d5..6631017ca2 100644
> > --- a/hw/misc/macio/cuda.c
> > +++ b/hw/misc/macio/cuda.c
> > @@ -275,7 +275,7 @@ static void cuda_delay_set_sr_int(CUDAState *s)
> >      timer_mod(s->sr_delay_timer, expire);
> >  }
> >  
> > -static uint32_t cuda_readb(void *opaque, hwaddr addr)
> > +static uint64_t cuda_read(void *opaque, hwaddr addr, unsigned size)
> >  {
> >      CUDAState *s = opaque;
> >      uint32_t val;
> > @@ -350,7 +350,7 @@ static uint32_t cuda_readb(void *opaque, hwaddr addr)
> >      return val;
> >  }
> >  
> > -static void cuda_writeb(void *opaque, hwaddr addr, uint32_t val)
> > +static void cuda_write(void *opaque, hwaddr addr, uint64_t val, unsigned 
> > size)
> >  {
> >      CUDAState *s = opaque;
> >  
> > @@ -780,38 +780,14 @@ static void cuda_receive_packet_from_host(CUDAState 
> > *s,
> >      }
> >  }
> >  
> > -static void cuda_writew (void *opaque, hwaddr addr, uint32_t value)
> > -{
> > -}
> > -
> > -static void cuda_writel (void *opaque, hwaddr addr, uint32_t value)
> > -{
> > -}
> > -
> > -static uint32_t cuda_readw (void *opaque, hwaddr addr)
> > -{
> > -    return 0;
> > -}
> > -
> > -static uint32_t cuda_readl (void *opaque, hwaddr addr)
> > -{
> > -    return 0;
> > -}
> > -
> >  static const MemoryRegionOps cuda_ops = {
> > -    .old_mmio = {
> > -        .write = {
> > -            cuda_writeb,
> > -            cuda_writew,
> > -            cuda_writel,
> > -        },
> > -        .read = {
> > -            cuda_readb,
> > -            cuda_readw,
> > -            cuda_readl,
> > -        },
> > +    .read = cuda_read,
> > +    .write = cuda_write,
> > +    .endianness = DEVICE_BIG_ENDIAN,
> > +    .valid = {
> 
> This change the device behavior, previously 16/32bit access were
> ignored, now they are illegal.

This change looks correct to me.  As you note >1 byte accesses were
previously ignored - not broken or handled in any other way.  That
means there's no possibility using >1byte accesses could ever
accomplish, so it reads to me as being invalid.  I'd say the fact that
it previously didn't generate some sort of error is the bug.

> Using ".impl" instead also change the behavior, since 16/32bits access
> would be processed (as 2x or 4x 8bit access).
> 
> A comment "this change is safe because ..." would reassure me ;)
> 
> > +        .min_access_size = 1,
> > +        .max_access_size = 1,
> >      },
> > -    .endianness = DEVICE_NATIVE_ENDIAN,
> >  };
> >  
> >  static bool cuda_timer_exist(void *opaque, int version_id)
> > 
> 

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

signature.asc
Description: PGP signature