Re: Validating dma_mmap_coherent() parameters before calling (was Re: WARNING in memtype_reserve)

[Jeremy Linton <jeremy.linton@arm.com>]

Hi,

On 5/14/20 1:31 AM, Christoph Hellwig wrote:
> On Thu, May 14, 2020 at 08:27:50AM +0200, Greg KH wrote:
>> On Thu, May 14, 2020 at 08:14:17AM +0200, Christoph Hellwig wrote:
>>> Guys, can you please start formal thread on this?  I have no
>>> idea where this came from and what the rationale is.  Btw, if the
>>> pfn is crap in dma_direct_mmap then the dma_addr_t passed in is
>>> crap, as it is derived from that.  What is the caller, and how is
>>> this triggered?
>>
>>
>> Ok, to summarize, commit 2bef9aed6f0e ("usb: usbfs: correct kernel->user
>> page attribute mismatch") changed a call from remap_pfn_range() to
>> dma_mmap_coherent() for usb data buffers being sent from userspace.
> 
> I only need to look at the commit for 3 seconds to tell you that it is
> completely buggy.  While using dma_mmap_coherent is fundamentally the
> right thing and absolutely required for dma_alloc_* allocations, USB
> also uses it's own local gen pool allocator or plain kmalloc for not
> DMA capable controller.  This need to use remap_pfn_range.  I'm pretty
> sure you hit one of those cases.

? The code path in question is usbdev_mmap() and the allocation is done 
~13 lines lines before as a usb_alloc_coherent().


> 
> The logic should be something like:
> 
> 	if (hcd->localmem_pool || !hcd_uses_dma(hcd))
> 		remap_pfn_range()
> 	else
> 		dma_mmap_coherent()
> 

That sort of makes sense, except for the above, and the fact that I 
would imagine the dma_mmap_coherent should be dealing with that case. 
I'm not really clear about the details of the GCE usb device here, but 
my first guess at this was the dma_pgprot() in dma_direct_mmap() is 
incorrectly picking a pgprot...