Skip to contentSkip to conversation

fbmem: VM_IO set, but not propagated.

PATCH Konrad Rzeszutek Wilk 3 participants 4 messages Jul 22, 2010 – Jul 27, 2010 Framebuffer layer

lore
Share & tools
lore
#0 lore raw

Hey,

This bug was found when Linux kernel was running under Xen. In that scenario, any page that has VM_IO flag to it, means that it MUST be a MMIO/VRAM backend memory , _not_ System RAM. That is what the fbmem.c does: sets VM_IO, ioremaps the region - everything is peachy.

Well, not exactly. The vm_page_prot does not get the relevant PTE flags set (_PAGE_IOMAP) which under Xen is a death-kneel to pages that are referencing real physical devices but don't have that flag set.

Here is the patch:

Author: Daniel De Graaf <[email protected]>
Date:   Wed Jul 21 16:52:46 2010 -0400

fb: propagate VM_IO to VMA.

When we setup up the VMA flags for the mmap flag and we end up using the fallback mmap functionality we set the vma->vm_flags |= VM_IO. However we neglect to propagate the flag to the vma->vm_page_prot.

This patch fixes this.

drivers/video/fbmem.c +1 -0
index 99bbd28..057433a 100644
@@ -1362,6 +1362,7 @@ fb_mmap(struct file *file, struct vm_area_struct * vma)
 	vma->vm_pgoff = off >> PAGE_SHIFT;
 	/* This is an IO map - tell maydump to skip this VMA */
 	vma->vm_flags |= VM_IO | VM_RESERVED;
+	vma->vm_page_prot = vm_get_page_prot(vma->vm_flags);
 	fb_pgprotect(file, vma, off);
 	if (io_remap_pfn_range(vma, vma->vm_start, off >> PAGE_SHIFT,
 			     vma->vm_end - vma->vm_start, vma->vm_page_prot))
#1 lore raw
Quoted 16 lines · On Thu, 22 Jul 2010 17:31:28 -0400 Konrad Rzeszutek Wilk…

This bug was found when Linux kernel was running under Xen. In that scenario, any page that has VM_IO flag to it, means that it MUST be a MMIO/VRAM backend memory , _not_ System RAM. That is what the fbmem.c does: sets VM_IO, ioremaps the region - everything is peachy.

Well, not exactly. The vm_page_prot does not get the relevant PTE flags set (_PAGE_IOMAP) which under Xen is a death-kneel to pages that are referencing real physical devices but don't have that flag set.

Here is the patch:

Author: Daniel De Graaf <[email protected]>
Date:   Wed Jul 21 16:52:46 2010 -0400

fb: propagate VM_IO to VMA.

Quoted 9 lines

When we setup up the VMA flags for the mmap flag and we end up using the fallback mmap functionality we set the vma->vm_flags |= VM_IO. However we neglect to propagate the flag to the vma->vm_page_prot.

This patch fixes this.

Confused. We have From:Konrad and Author:Daniel and no signoff from Daniel.

I've committed the patch assuming that Daniel was the author, but didn't sign off the patch. Your signoff is sufficient for merging purposes.

But maybe I was wrong.

I'm also assuming that we can merge this into 2.6.36 and not backport it into -stable. But maybe I'm wrong about that too! Talk to me.

#2 lore raw
In reply to Konrad Rzeszutek Wilk
Quoted 35 lines · Andrew Morton
Quoted 16 lines · On Thu, 22 Jul 2010 17:31:28 -0400 Konrad Rzeszutek Wilk…

This bug was found when Linux kernel was running under Xen. In that scenario, any page that has VM_IO flag to it, means that it MUST be a MMIO/VRAM backend memory , _not_ System RAM. That is what the fbmem.c does: sets VM_IO, ioremaps the region - everything is peachy.

Well, not exactly. The vm_page_prot does not get the relevant PTE flags set (_PAGE_IOMAP) which under Xen is a death-kneel to pages that are referencing real physical devices but don't have that flag set.

Here is the patch:

Author: Daniel De Graaf <[email protected]>
Date:   Wed Jul 21 16:52:46 2010 -0400

fb: propagate VM_IO to VMA.

Quoted 9 lines

When we setup up the VMA flags for the mmap flag and we end up using the fallback mmap functionality we set the vma->vm_flags |= VM_IO. However we neglect to propagate the flag to the vma->vm_page_prot.

This patch fixes this.

Confused. We have From:Konrad and Author:Daniel and no signoff from Daniel.

I've committed the patch assuming that Daniel was the author, but didn't sign off the patch. Your signoff is sufficient for merging purposes.

Quoted 6 lines

But maybe I was wrong.

I'm also assuming that we can merge this into 2.6.36 and not backport it into -stable. But maybe I'm wrong about that too! Talk to me.

If it's useful to have a signoff line from me, you can add it; I guess it wasn't clear in my previous reply that I was signing off on it.

#3 lore raw
In reply to Konrad Rzeszutek Wilk
Quoted 23 lines · Andrew Morton
Quoted 20 lines · On Thu, 22 Jul 2010 17:31:28 -0400 Konrad Rzeszutek Wilk…

This bug was found when Linux kernel was running under Xen. In that scenario, any page that has VM_IO flag to it, means that it MUST be a MMIO/VRAM backend memory , _not_ System RAM. That is what the fbmem.c does: sets VM_IO, ioremaps the region - everything is peachy.

Well, not exactly. The vm_page_prot does not get the relevant PTE flags set (_PAGE_IOMAP) which under Xen is a death-kneel to pages that are referencing real physical devices but don't have that flag set.

Here is the patch:

Author: Daniel De Graaf <[email protected]>
Date:   Wed Jul 21 16:52:46 2010 -0400

fb: propagate VM_IO to VMA.

When we setup up the VMA flags for the mmap flag and we end up using the fallback mmap functionality we set the vma->vm_flags |= VM_IO. However we neglect to propagate the flag to the vma->vm_page_prot.

Quoted 7 lines
Quoted 5 lines

This patch fixes this.

Confused. We have From:Konrad and Author:Daniel and no signoff from Daniel.

So Daniel came up with the original fix, I've came to him and suggested that perhaps we should use the vm_get_page_prot, and he agreed.

Since I've kept the patch in my branch of "Xen-KMS/DRM/TTM-fixes" I figured I should post the patch upstream and hence I signed off on it.

Quoted 3 lines

I've committed the patch assuming that Daniel was the author, but didn't sign off the patch. Your signoff is sufficient for merging

Yes. He is the author. Let me refer back to my notes but I figured if the patch has an Author: set there was no need for an extra S-o-b?

Quoted 7 lines

purposes.

But maybe I was wrong.

I'm also assuming that we can merge this into 2.6.36 and not backport it into -stable. But maybe I'm wrong about that too! Talk to me.

No need to backport it.