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

[Hanjie Lin]

On 2018/12/4 20:00, Lorenzo Pieralisi wrote:
> On Tue, Dec 04, 2018 at 06:40:55PM +0800, Hanjie Lin wrote:
>>
>>
>> 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.
> 
> You can fix this too.
> 
>> 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
> 
> "There is a bug in the MESON AXG pcie controller whereby software cannot
> programme the PCI_CLASS_DEVICE register, so we must fabricate the return
> value in the config accessors."
> 

Ok, it's more clear :)

>>         * 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;
> 
> You can further filter it with (where & 0x1) == PCI_CLASS_DEVICE
> and handle the size accordingly, so that even a byte access would
> work, for completeness.
> 
> Lorenzo
>

Of course, I will add process to handle one-byte access from 0xa/0xb register 
in next-version.

thanks.

hanjie
 
>>         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
> 
> .
>