Re: [Qemu-devel] Memory API: handling unassigned physical memory

[Blue Swirl]

On Mon, Apr 30, 2012 at 13:52, Mark Cave-Ayland
<address@hidden> wrote:
> On 30/04/12 14:27, Peter Maydell wrote:
>
> Hi Peter,
>
>
>>> IMO the best fix is to unsysbus the device and qomify it instead.  This
>>> way we're 100% flexible in how we can attach it.
>>
>>
>> You don't need to wait for QOM to grow enough features to
>> replace sysbus. If you don't like what sysbus_mmio_map() does, you
>
>
> Oh - so does this mean that QOM is not feature-complete?

At least 'Pin' class to replace GPIOs and IRQs is not in HEAD (but
Anthony already has a tree somewhere) and I can't remember what was
the plan for buses.

>
>
>> can always use sysbus_mmio_get_region() to get the MemoryRegion* and
>> then deal with it however you need to. This is the standard way
>> to deal with "I have a sysbus device which I want to map into my
>> custom container object".
>
>
> I already have code to do this for the sun4m-only hardware modelled on
> sysbus_mmio_map() like this:
>
>
> static void sbus_mmio_map(void *sbus, SysBusDevice *s, int n,
> target_phys_addr_t addr)
> {
>    MemoryRegion *sbus_mem, *mem;
>    target_phys_addr_t sbus_base;
>    SBusState *sbus_state = FROM_SYSBUS(SBusState, (SysBusDevice *)sbus);
>
>    sbus_mem = sysbus_mmio_get_region((SysBusDevice *)sbus, 0);
>    mem = sysbus_mmio_get_region(s, n);
>
>    /* SBus addresses are physical addresses, so subtract start of region */
>    sbus_base = sbus_state->base;
>    memory_region_add_subregion(sbus_mem, addr - sbus_base, mem);
> }
>
>
> The key problem is that this doesn't worked with shared peripherals, such as
> the ESP device which is also used on various PPC Mac models as well as
> SPARC. That's because its init function looks like this:
>
>
>
> void esp_init(target_phys_addr_t espaddr, int it_shift,
>              ESPDMAMemoryReadWriteFunc dma_memory_read,
>              ESPDMAMemoryReadWriteFunc dma_memory_write,
>              void *dma_opaque, qemu_irq irq, qemu_irq *reset,
>              qemu_irq *dma_enable)
> {
>    ...
>    ...
>    sysbus_mmio_map(s, 0, espaddr);
>    ...
> }
>
>
> Therefore I can't change it to my (modified) sbus_mmio_map() function
> because it would break other non-SPARC platforms, and AIUI there is nothing
> in the memory API that allows me to move a subregion to a different
> MemoryRegion parent, even if I can get a reference to it with
> sysbus_mmio_get_region() after the sysbus_mmio_map() call - or have I
> misunderstood something?

Sysbus is used as a generic class for motherboard devices, there is an
assumption that there is no higher level bus. What we need here is a
full blown bus. The translations and mappigs between bus addresses and
motherboard addresses should be done in a Sysbus to SBus bridge
device, just like PCI host bridges do.

>
>
> ATB,
>
> Mark.
>