Re: [Qemu-devel] [PATCH v6 1/3] pci: Add support for Designware IP block

[Andrey Smirnov]

On Tue, Feb 13, 2018 at 2:15 PM, Michael S. Tsirkin <address@hidden> wrote:
> On Tue, Feb 13, 2018 at 12:24:40PM -0800, Andrey Smirnov wrote:
>> On Tue, Feb 13, 2018 at 10:13 AM, Michael S. Tsirkin <address@hidden> wrote:
>> > On Tue, Feb 13, 2018 at 09:07:10AM -0800, Andrey Smirnov wrote:
>> >> +static void designware_pcie_root_class_init(ObjectClass *klass, void 
>> >> *data)
>> >> +{
>> >> +    PCIDeviceClass *k = PCI_DEVICE_CLASS(klass);
>> >> +    DeviceClass *dc = DEVICE_CLASS(klass);
>> >> +
>> >> +    set_bit(DEVICE_CATEGORY_BRIDGE, dc->categories);
>> >> +
>> >> +    k->vendor_id = PCI_VENDOR_ID_SYNOPSYS;
>> >> +    k->device_id = 0xABCD;
>> >> +    k->revision = 0;
>> >> +    k->class_id = PCI_CLASS_BRIDGE_PCI;
>> >> +    k->is_express = true;
>> >> +    k->is_bridge = true;
>> >> +    k->exit = pci_bridge_exitfn;
>> >> +    k->realize = designware_pcie_root_realize;
>> >> +    k->config_read = designware_pcie_root_config_read;
>> >> +    k->config_write = designware_pcie_root_config_write;
>> >> +
>> >> +    dc->reset = pci_bridge_reset;
>> >> +    /*
>> >> +     * PCI-facing part of the host bridge, not usable without the
>> >> +     * host-facing part, which can't be device_add'ed, yet.
>> >> +     */
>> >> +    dc->user_creatable = false;
>> >> +    dc->vmsd = &vmstate_designware_pcie_root;
>> >> +}
>> >> +
>> >> +static uint64_t designware_pcie_host_mmio_read(void *opaque, hwaddr addr,
>> >> +                                               unsigned int size)
>> >> +{
>> >> +    PCIHostState *pci = PCI_HOST_BRIDGE(opaque);
>> >> +    PCIDevice *device = pci_find_device(pci->bus, 0, 0);
>> >> +
>> >> +    return pci_host_config_read_common(device,
>> >> +                                       addr,
>> >> +                                       pci_config_size(device),
>> >> +                                       size);
>> >> +}
>> >> +
>> >> +static void designware_pcie_host_mmio_write(void *opaque, hwaddr addr,
>> >> +                                            uint64_t val, unsigned int 
>> >> size)
>> >> +{
>> >> +    PCIHostState *pci = PCI_HOST_BRIDGE(opaque);
>> >> +    PCIDevice *device = pci_find_device(pci->bus, 0, 0);
>> >> +
>> >> +    return pci_host_config_write_common(device,
>> >> +                                        addr,
>> >> +                                        pci_config_size(device),
>> >> +                                        val, size);
>> >> +}
>> >> +
>> >> +static const MemoryRegionOps designware_pci_mmio_ops = {
>> >> +    .read       = designware_pcie_host_mmio_read,
>> >> +    .write      = designware_pcie_host_mmio_write,
>> >> +    .endianness = DEVICE_NATIVE_ENDIAN,
>> >> +    .impl = {
>> >> +        /*
>> >> +         * Our device would not work correctly if the guest was doing
>> >> +         * unaligned access. This might not be a limitation on the real
>> >> +         * device but in practice there is no reason for a guest to 
>> >> access
>> >> +         * this device unaligned.
>> >> +         */
>> >> +        .min_access_size = 4,
>> >> +        .max_access_size = 4,
>> >> +        .unaligned = false,
>> >> +    },
>> >> +};
>> >
>> > Could you pls add some comments explaining why is DEVICE_NATIVE_ENDIAN
>> > appropriate here?  Most of these cases are plain "we never bothered
>> > about cross-endian setups". Some are "there's a mix of different
>> > endian-ness values, need to handle in a special way".
>> >
>> > I suspect you really need DEVICE_LITTLE_ENDIAN.
>> >
>>
>> That MemoryRegion corresponds to a register file permanently mapped
>> into CPU's address space, so my assumption is that SoC designers will
>> wire it according to CPUs endianness be it big or little. I am not
>> aware of any big-endian CPU based SoC on the market using Designware's
>> IP block, so I don't think there are any precedent confirming or
>> denying correctness of my assumption. IMHO, this is also the reason
>> why all of Linux driver code for that IP assumes little endianness.
>
> IMHO if Linux driver code does cpu_to_le then it seems best to be
> consistent with that.
>

Well, all of the DW code does so implicitly by using readl()/writel()
helpers which will perform cpu_to_le/le_to_cpu under the hood. But is
seems to me that it could be either because the access does have to be
LE always or simply because readl()/writel() are goto memory helpers
on ARM/LE-platforms.

FWIW: Somewhat similar precedent of MIPS/Boston machine can serve as
counter-example to my assumption, since Xilinx PCIE IP there seem to
be wired to be LE despite being attached to BE CPU.

Thanks,
Andrey Smirnov