[BUG] completely bonkers use of set_need_resched + VM_FAULT_NOPAGE

Daniel Vetter daniel.vetter at ffwll.ch
Fri Sep 13 01:32:48 PDT 2013

On Fri, Sep 13, 2013 at 10:23 AM, Thomas Hellstrom
<thellstrom at vmware.com> wrote:
> As previously mentioned, copy_from_user should return -EFAULT, since the
> VMAs are marked with VM_IO. It should not recurse into fault(), so evil
> user-space looses.

I haven't put a printk in the code to prove this, but gem mmap also
sets VM_IO in drm_gem_mmap_obj. And we can very much hit our own fault
handler and deadlock ....

On a _very_ quick reading (and definitely not enough coffee yet for
reading mm/* stuff) it looks like it's get_user_pages that will return
an -EFAULT when hitting upon a VM_IO mapping (which makes sense since
there's really no page backing it). Actually using get_user_pages was
the original slowpath we've had in a few places until we've noticed
that for pwrite that breaks legit userspace (the glBufferData(glMap))
use-case), so we've switched to lock dropping and proper slowpaths
using copy_*_user everywhere instead of trying to pin the userspace
storage with get_user_pages.
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch

More information about the dri-devel mailing list