Re: [Qemu-devel] [PATCH v6 03/13] register: Add Memory API glue

[Alistair Francis]

On Thu, Jun 9, 2016 at 12:03 PM, Peter Maydell <address@hidden> wrote:
> On 12 May 2016 at 23:45, Alistair Francis <address@hidden> wrote:
>> Add memory io handlers that glue the register API to the memory API.
>> Just translation functions at this stage. Although it does allow for
>> devices to be created without all-in-one mmio r/w handlers.
>>
>> This patch also adds the RegisterInfoArray struct, which allows all of
>> the individual RegisterInfo structs to be grouped into a single memory
>> region.
>>
>> Signed-off-by: Peter Crosthwaite <address@hidden>
>> Signed-off-by: Alistair Francis <address@hidden>
>> ---
>> V6:
>>  - Add the memory region later
>> V5:
>>  - Convert to using only one memory region
>>
>>  hw/core/register.c    | 72 
>> +++++++++++++++++++++++++++++++++++++++++++++++++++
>>  include/hw/register.h | 50 +++++++++++++++++++++++++++++++++++
>>  2 files changed, 122 insertions(+)
>>
>> diff --git a/hw/core/register.c b/hw/core/register.c
>> index 5e6f621..25196e6 100644
>> --- a/hw/core/register.c
>> +++ b/hw/core/register.c
>> @@ -147,3 +147,75 @@ void register_reset(RegisterInfo *reg)
>>
>>      register_write_val(reg, reg->access->reset);
>>  }
>> +
>> +static inline void register_write_memory(void *opaque, hwaddr addr,
>> +                                         uint64_t value, unsigned size, 
>> bool be)
>> +{
>> +    RegisterInfoArray *reg_array = opaque;
>> +    RegisterInfo *reg = NULL;
>> +    uint64_t we = ~0;
>> +    int i, shift = 0;
>> +
>> +    for (i = 0; i < reg_array->num_elements; i++) {
>> +        if (reg_array->r[i]->access->decode.addr == addr) {
>> +            reg = reg_array->r[i];
>> +            break;
>> +        }
>> +    }
>> +    assert(reg);
>
> I'm surprised we don't support having the register array have
> gaps for unimplemented/undefined registers. Presumably users
> have to specify a lot of unimplemented entries ?
>
> If you're going to assert() on undecoded addresses it would be
> better to do a scan through at device init to sanity check
> the register array, so missing elements are an obvious failure
> rather than only showing up if the guest happens to access them.

You're right, this is a little harsh. I'm thinking I'll change it to a
qemu_log() (unimplemented, although it also could be a guest error)
and remove the assert().

>
>> +
>> +    /* Generate appropriate write enable mask and shift values */
>> +    if (reg->data_size < size) {
>> +        we = MAKE_64BIT_MASK(0, reg->data_size * 8);
>> +        shift = 8 * (be ? reg->data_size - size : 0);
>> +    } else if (reg->data_size >= size) {
>> +        we = MAKE_64BIT_MASK(0, size * 8);
>> +    }
>> +
>> +    register_write(reg, value << shift, we << shift, reg_array->prefix,
>> +                   reg_array->debug);
>> +}
>> +
>> +void register_write_memory_be(void *opaque, hwaddr addr, uint64_t value,
>> +                              unsigned size)
>> +{
>> +    register_write_memory(opaque, addr, value, size, true);
>> +}
>> +
>> +
>> +void register_write_memory_le(void *opaque, hwaddr addr, uint64_t value,
>> +                              unsigned size)
>> +{
>> +    register_write_memory(opaque, addr, value, size, false);
>> +}
>> +
>> +static inline uint64_t register_read_memory(void *opaque, hwaddr addr,
>> +                                            unsigned size, bool be)
>> +{
>> +    RegisterInfoArray *reg_array = opaque;
>> +    RegisterInfo *reg = NULL;
>> +    int i, shift;
>> +
>> +    for (i = 0; i < reg_array->num_elements; i++) {
>> +        if (reg_array->r[i]->access->decode.addr == addr) {
>> +            reg = reg_array->r[i];
>> +            break;
>> +        }
>> +    }
>> +    assert(reg);
>> +
>> +    shift = 8 * (be ? reg->data_size - size : 0);
>> +
>> +    return (register_read(reg, reg_array->prefix, reg_array->debug) >> 
>> shift) &
>> +           MAKE_64BIT_MASK(0, size * 8);
>
> This kind of thing is reimplementing extract64().

Ok, I'll update it to use extract64()

>
>> +}
>> +
>> +uint64_t register_read_memory_be(void *opaque, hwaddr addr, unsigned size)
>> +{
>> +    return register_read_memory(opaque, addr, size, true);
>> +}
>> +
>> +uint64_t register_read_memory_le(void *opaque, hwaddr addr, unsigned size)
>> +{
>> +    return register_read_memory(opaque, addr, size, false);
>> +}
>
> Why do we need to handle big vs little endian separately rather
> than just having the memory region say which it is and letting
> the core memory system handle things appropriately ?

I didn't realise that is an option. So I can remove all the endianess
handling from here and the core will handle it?

Thanks,

Alistair

>
> thanks
> -- PMM
>