Skip to content

Commit

Permalink
[PARISC] fix virtual aliasing issue in get_shared_area()
Browse files Browse the repository at this point in the history
On Thu, 2012-11-01 at 16:45 -0700, Michel Lespinasse wrote:
> Looking at the arch/parisc/kernel/sys_parisc.c implementation of
> get_shared_area(), I do have a concern though. The function basically
> ignores the pgoff argument, so that if one creates a shared mapping of
> pages 0-N of a file, and then a separate shared mapping of pages 1-N
> of that same file, both will have the same cache offset for their
> starting address.
>
> This looks like this would create obvious aliasing issues. Am I
> misreading this ? I can't understand how this could work good enough
> to be undetected, so there must be something I'm missing here ???

This turns out to be correct and we need to pay attention to the pgoff as
well as the address when creating the virtual address for the area.
Fortunately, the bug is rarely triggered as most applications which use pgoff
tend to use large values (git being the primary one, and it uses pgoff in
multiples of 16MB) which are larger than our cache coherency modulus, so the
problem isn't often seen in practise.

Reported-by: Michel Lespinasse <walken@google.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: James Bottomley <JBottomley@Parallels.com>
  • Loading branch information
James Bottomley authored and James Bottomley committed Nov 15, 2012
1 parent 77b6706 commit 949a05d
Showing 1 changed file with 2 additions and 0 deletions.
2 changes: 2 additions & 0 deletions arch/parisc/kernel/sys_parisc.c
Original file line number Diff line number Diff line change
Expand Up @@ -73,6 +73,8 @@ static unsigned long get_shared_area(struct address_space *mapping,
struct vm_area_struct *vma;
int offset = mapping ? get_offset(mapping) : 0;

offset = (offset + (pgoff << PAGE_SHIFT)) & 0x3FF000;

addr = DCACHE_ALIGN(addr - offset) + offset;

for (vma = find_vma(current->mm, addr); ; vma = vma->vm_next) {
Expand Down

0 comments on commit 949a05d

Please sign in to comment.