Re: [PATCH RFC 1/1] x86: fix bad memory access in fb_is_primary_device()

[Alexander Popov]

On 16.02.2016 18:18, Peter Jones wrote:
> On Tue, Feb 16, 2016 at 01:49:18PM +0000, Matt Fleming wrote:
>> [ Including Peter, the efifb maintainer. Original email is here,
>>
>>     http://marc.info/?l=linux-kernel&m=145552936131335&w=2
>>
>>   I've snipped some of the quoted text ]
>>
>> On Tue, 16 Feb, at 08:55:22AM, Ingo Molnar wrote:
>>>
>>> (I've Cc:-ed the EFI-FB and FB gents. Mail quoted below.)
>>>
>>> * Alexander Popov <alpopov@xxxxxxxxxxxxxx> wrote:
>>>
>>>> Currently the code in fb_is_primary_device() contains to_pci_dev() macro
>>>> which is applied to dev from struct fb_info. In some cases this causes
>>>> bad memory access when fb_is_primary_device() handles fb_info of efifb.
>>>> The reason is that fb dev of efifb is embedded into struct platform_device
>>>> but not into struct pci_dev.
>>>>
>>>> We can fix this by checking fb dev bus name in fb_is_primary_device().
>>>>
>>>> It seems that this bug reveals some bigger problem with to_pci_dev(),
>>>> to_platform_device() and others, which just do container_of() and
>>>> don't check whether struct device is a part of the appropriate structure.
>>>> Should we do something more about it?
>>>>
>>>> KASan report:
>>
>> [...]
>>
>>>>
>>>> Signed-off-by: Alexander Popov <alpopov@xxxxxxxxxxxxxx>
>>>> ---
>>>>  arch/x86/video/fbdev.c | 9 +++++----
>>>>  1 file changed, 5 insertions(+), 4 deletions(-)
>>>>
>>>> diff --git a/arch/x86/video/fbdev.c b/arch/x86/video/fbdev.c
>>>> index d5644bb..4999f78 100644
>>>> --- a/arch/x86/video/fbdev.c
>>>> +++ b/arch/x86/video/fbdev.c
>>>> @@ -18,11 +18,12 @@ int fb_is_primary_device(struct fb_info *info)
>>>>  	struct pci_dev *default_device = vga_default_device();
>>>>  	struct resource *res = NULL;
>>>>  
>>>> -	if (device)
>>>> -		pci_dev = to_pci_dev(device);
>>>> -
>>>> -	if (!pci_dev)
>>>> +	if (!device || !device->bus ||
>>>> +		    !device->bus->name || strcmp(device->bus->name, "pci")) {
>>>>  		return 0;
>>>> +	}
>>>> +
>>>> +	pci_dev = to_pci_dev(device);
>>>>  
>>>>  	if (default_device) {
>>>>  		if (pci_dev == default_device)
>>>> -- 
>>>> 1.9.1
>>>>
>>
>> I wonder if this issue could explain some of the efifb issues we've
>> seen reported on bugzilla.kernel.org in the past where switching from
>> efifb to some other framebuffer device caused hangs during boot. I'm
>> struggling to find the relevant bugzilla entries now, though.
> 
> It's possible it could, but I don't have them handy either.  I've also
> wondered if some of them were due to bad data from the firmware - at
> plugfests we've seen some cases where the actual video mode as measured
> with a ruler is clearly not what the firmware claims it to be, so it's
> entirely possible we're occasionally told a memory region that is not
> what's actually mapped, or that's mapped but is only partially backed
> by the actual frame buffer memory.
> 
> But aside from that diversion, I think Alexander has a legitimate
> question about use of to_pci_dev().  If I ask the question: can we fix
> this in efifb by making it live on a pci_dev, I have a couple of
> fundamental problems:
> 
> 1) technically it doesn't have to be a pci_dev at all (but, practically,
>    so far it always is on PCI...)
> 2) From EFI, we can't necessarily pin it down to a single PCI device
>    even if it is PCI.  Before we do EFI's ExitBootServices() call, we
>    can try to find the PCI_IO handle our GOP instance is connected to,
>    but not all firmware GOP drivers use that, so it doesn't always work.
>    And even if it did, there can be more than one instance pointing to
>    the same memory with different PCI devices - lots of laptops have
>    this sort of thing.
> 3) Ignoring the EFI side and just focusing on PCI, if there's two
>    devices configured that could do scanout, it can be mapped to one
>    device's BAR but the other device be the actual device using it.  In
>    this case either choice is probably wrong for something, and the
>    things that have the information to resolve which one don't include
>    efifb - they're the drivers we'll likely hand off to later.
> 
> So it's most likely right for efifb to be embedded in a platform_device
> instead of a pci_dev.  Which leads back to Alexander's question - if it
> isn't in a pci_dev, that means fb_is_primary_device() needs to not
> assume it is.  So the patch appears correct, but so is the question -
> should to_pci_dev() be checking this and returning NULL here?

The discussion has suspended. May I activate it again?

So there are two ways to fix the bad memory access in fb_is_primary_device().

The first one is proposed in my patch. Checking the bus name string doesn't
look good but I didn't manage to come up with anything better.

The second way is changing to_pci_dev() similarly. It may return NULL or
call BUG() when struct device is a part of an inappropriate structure.

Which way is better? Do we need to do anything with other similar macros?

Thanks.
Best regards,
Alexander