Re: Gem GTT mmaps..

[Jesse Barnes <jbarnes@virtuousgeek.org>]

On Wednesday, February 4, 2009 2:32 pm Thomas Hellström wrote:
> Jesse,
>
> I have some concerns about the GEM GTT mmap functionality.

Thanks for looking it over again; you would know since some of this code came 
from you in the first place. :)

> First, a gem object pointer is copied to map->offset and then to the
> vma->vm_private_data without proper reference counting. This pointer is
> used in i915_gem_fault() to access the gem object. However if the gem
> object is destroyed and a process then tries to access data in a vma
> mapping the (now destroyed) object, it would dereference a stale pointer
> into kernel space? Shouldn't those pointers be reference counted, and to
> account for fork(), a vm open and close would be needed to  reference
> count corresponding pointers of newly created and destroyed vmas?

Yeah looks like we don't protect against vm_private_data pointing at a freed 
or other object.  But rather than refcounting the pointers I wonder if we 
could make the private data use the GEM object name instead, then do the 
lookup in the fault handler?

> Second, the i915_gem_fault method  returns VM_FAULT_SIGBUS if
> vm_insert_pfn() fails with an -EBUSY. I think that's an error, since
> that would indicate that the pte was already populated by a racing thread.

Ah ok that's easy enough to fix up; I didn't see that EBUSY meant "pte already 
valid".

-- 
Jesse Barnes, Intel Open Source Technology Center