Re: [PATCH 1/3] PCI: iproc: fix out of bound array access

[Ray Jui]

On 3/26/2020 1:48 PM, Bjorn Helgaas wrote:
> On Thu, Mar 26, 2020 at 01:27:36PM -0700, Ray Jui wrote:
>> On 3/26/2020 12:48 PM, Bjorn Helgaas wrote:
>>> ...
>>> It's outside the scope of this patch, but I'm not really a fan of the
>>> pcie->reg_offsets[] scheme this driver uses to deal with these
>>> differences.  There usually seems to be *something* that keeps the
>>> driver from referencing registers that don't exist, but it doesn't
>>> seem like the mechanism is very consistent or robust:
>>>
>>>   - IPROC_PCIE_LINK_STATUS is implemented by PAXB but not PAXC.
>>>     iproc_pcie_check_link() avoids using it if "ep_is_internal", which
>>>     is set for PAXC and PAXC_V2.  Not an obvious connection.
>>>
>>>   - IPROC_PCIE_CLK_CTRL is implemented for PAXB and PAXC_V1, but not
>>>     PAXC_V2.  iproc_pcie_perst_ctrl() avoids using it ep_is_internal",
>>>     so it *doesn't* use it for PAXC_V1, which does implement it.
>>>     Maybe a bug, maybe intentional; I can't tell.
>>>
>>>   - IPROC_PCIE_INTX_EN is only implemented by PAXB (not PAXC), but
>>>     AFAICT, we always call iproc_pcie_enable() and rely on
>>>     iproc_pcie_write_reg() silently drop the write to it on PAXC.
>>>
>>>   - IPROC_PCIE_OARR0 is implemented by PAXB and PAXB_V2 and used by
>>>     iproc_pcie_map_ranges(), which is called if "need_ob_cfg", which
>>>     is set if there's a "brcm,pcie-ob" DT property.  No clear
>>>     connection to PAXB.
>>>
>>> I think it would be more readable if we used a single variant
>>> identifier consistently, e.g., the "pcie->type" already used in
>>> iproc_pcie_msi_steer(), or maybe a set of variant-specific function
>>> pointers as pcie-qcom.c does.
>>
>> It is not possible to use a single variant identifier consistently,
>> i.e., 'pcie->type'. Many of these features are controller revision
>> specific, and certain revisions of the controllers may all have a
>> certain feature, while other revisions of the controllers do not. In
>> addition, there are overlap in features across different controllers.
>>
>> IMO, it makes sense to have feature specific flags or booleans, and have
>> those features enabled or disabled based on 'pcie->type', which is what
>> the current driver does, but like you pointed out, what the driver
>> failed is to do this consistently.
> 
> There are several drivers that have the same problem of dealing with
> different revisions of hardware.  It would be nice to do it in a
> consistent style, whatever that is.
> 

Sure, agree with you that it should be handled in a consistent way
within this driver, and the current driver is not handling this
consistently.

>> The IPROC_PCIE_INTX_EN example you pointed out is a good example. I
>> agree with you that we shouldn't rely on iproc_pcie_write_reg to
>> silently drop the operation for PAXC. We should add code to make it
>> explictly obvious that legacy interrupt is not supported in all PAXC
>> controllers.
>>
>> pcie->pcie->reg_offsets[] scheme was not intended to be used to silently
>> drop register access that are activated based on features. It's a
>> mistake that should be fixed if some code in the driver is done that
>> way, as you pointed out.
> 
> That's actually why I dug into this a bit -- the
> iproc_pcie_reg_is_invalid() case is really a design-time error, so it
> seemed like there should be a WARN() there instead of silently
> returning 0 or ignoring a write.
> 

I think 'iproc_pcie_reg_is_invalid' is a fall back protection. We should
aim to prevent this from happening in the first place using whatever
means we determined appropriate, and do that consistently. In addition,
I also agree with you that there should be a WARN instead of silently
returning zero (for reads) and dropping the writes.

We'll be looking into improving this as you suggested when we have a
chance. In the mean time, I think both of us agree this is out of the
scope of the issue that this patch is trying to fix, which is actually a
pretty critical issue that can cause potential corruption of memory and
the fix should be picked up ASAP (and for older LTS kernels too).

Thanks,

Ray

> Bjorn
>