Re: [PATCH v6 2/2] PCI: amlogic: Add the Amlogic Meson PCIe controller driver

[Hanjie Lin <hanjie.lin@amlogic.com>]

On 2018/12/4 6:57, Bjorn Helgaas wrote:
> On Mon, Dec 03, 2018 at 04:41:50PM +0000, Lorenzo Pieralisi wrote:
>> On Thu, Nov 22, 2018 at 04:53:54PM +0800, Hanjie Lin wrote:
>>
>> [...]
>>
>>> +static int meson_pcie_rd_own_conf(struct pcie_port *pp, int where, int size,
>>> +				  u32 *val)
>>> +{
>>> +	struct dw_pcie *pci = to_dw_pcie_from_pp(pp);
>>> +
>>> +	/*
>>> +	 * there is a bug of MESON AXG pcie controller that software can not
>>> +	 * programe PCI_CLASS_DEVICE register, so we must return a fake right
>>> +	 * value to ensure driver could probe successfully.
>>> +	 */
>>> +	if (where == PCI_CLASS_REVISION) {
>>> +		*val = readl(pci->dbi_base + PCI_CLASS_REVISION);
>>> +		/* keep revision id */
>>> +		*val &= PCI_CLASS_REVISION_MASK;
>>> +		*val |= PCI_CLASS_BRIDGE_PCI << 16;
>>> +		return PCIBIOS_SUCCESSFUL;
>>> +	}
>>
>> As I said before, this looks broken. If this code (or other drivers with
>> the same broken assumptions, eg dwc/pcie-qcom.c) carries out a, say,
>> byte sized config access of eg PCI_CLASS_DEVICE you will get junk out of
>> it according to your comment above.
>>
>> I would like to pick Bjorn's brain on this to see what we can really do
>> to fix this (and other) drivers.
> 
>   - Check to see whether you're reading anything in the 32-bit dword at
>     offset 0x08.
> 
>   - Do the 32-bit readl().
> 
>   - Insert the correct Sub-Class and Base Class code (you also throw
>     away the Programming Interface; not sure why that is)
> 
>   - If you're reading something smaller than 32 bits, mask & shift as
>     needed.  pci_bridge_emul_conf_read() does something similar that
>     you might be able to copy.
> 
> Out of curiosity, what code depends on PCI_CLASS_BRIDGE_PCI?  There
> are several places in the kernel that currently depend on it, but I
> think several of them *should* be checking dev->hdr_type to identify a
> type 1 header instead.
> 
> Bjorn
> 
> .
> 

Yes, it would be broken in particular scenes(eg: read 1 or 2 bytes from 0xa/PCI_CLASS_DEVICE)
that I didn't considered.
 
As your suggestion, I consider some code below may help this issue:
1, First call dw_pcie_read() help to read 1/2/4 bytes from register,
   request all other *size* bytes will return error and dw_pcie_read()
   will also check register alignment.
 
2, If dw_pcie_read() return success and *where* is 0x8/PCI_CLASS_DEVICE or 0xa/PCI_CLASS_REVISION,
   we may need to correct class code.
   As PCI_CLASS_REVISION is two-bytes register, so only when read 4 bytes from 0x8/PCI_CLASS_DEVICE
   or read 2 bytes from 0xa/PCI_CLASS_REVISION we should correct the class code.
 
ps: read 1 byte from 0xa/PCI_CLASS_REVISION or 0xb will get incorrect value.

static int meson_pcie_rd_own_conf(struct pcie_port *pp, int where, int size,
                u32 *val)
{
        struct dw_pcie *pci = to_dw_pcie_from_pp(pp);
        int ret;

        ret = dw_pcie_read(pci->dbi_base + where, size, val);
        if (ret != PCIBIOS_SUCCESSFUL)
                return ret;

        /*
        * there is a bug of MESON AXG pcie controller that software can not
        * programe PCI_CLASS_DEVICE register, so we must return a fake right
        * value to ensure driver could probe successfully.
        */
        if (where == PCI_CLASS_REVISION && size == 4)
                *val = (PCI_CLASS_BRIDGE_PCI << 16) | (*val & 0xffff);
        else if (where == PCI_CLASS_DEVICE && size == 2)
                *val = PCI_CLASS_BRIDGE_PCI;

        return PCIBIOS_SUCCESSFUL;
}

3, We must ensure class is PCI_CLASS_BRIDGE_PCI except right hdr_type,
   or pci_setup_device() will get failed: 

   ...
   class = pci_class(dev);
   dev->revision = class & 0xff;
   dev->class = class >> 8; /* upper 3 bytes */ 
   ....
   switch (dev->hdr_type) { /* header type */
   ... 
   case PCI_HEADER_TYPE_BRIDGE: /* bridge header */
      if (class != PCI_CLASS_BRIDGE_PCI)  /* class must be PCI_CLASS_BRIDGE_PCI  */
         goto bad; 


thanks.

hanjie