Re: 2.6.23 regression: accessing invalid mmap'ed memory from gdb causes unkillable spinning

[Nick Piggin <npiggin@suse.de>]

On Thu, Nov 01, 2007 at 09:08:45AM -0700, Linus Torvalds wrote:
> 
> 
> On Thu, 1 Nov 2007, Nick Piggin wrote:
> > 
> > Untested patch follows
> 
> Ok, this looks ok.
> 
> Except I would remove the VM_MAYSHARE bit from the test.

But we do want to allow forced COW faults for MAP_PRIVATE mappings. gdb
uses this for inserting breakpoints (but fortunately, a COW page in a
MAP_PRIVATE mapping is a much more natural thing for the VM).


> That whole bit should go, in fact.
> 
> We used to make it something different: iirc, a read-only SHARED mapping 
> was downgraded to a non-shared mapping, because we wanted to avoid some of 
> the costs we used to have with the VM implementation (actually, I think it 
> was various filesystems that don't like shared mappings because they don't 
> have a per-page writeback). But we left the VM_MAYSHARE bit on, to get 
> /proc/<pid>/mmap things right.
> 
> Or something like that. I forget the details. But I *think* we don't 
> actually need this any more.
> 
> But basically, the "right" way to test for shared mappings is historically 
> to just test the VM_MAYSHARE bit - but not *both* bits. Because VM_SHARE 
> may have been artificially cleared.
 
I think you're right -- VM_MAYSHARE is basically testing for MAP_SHARED.
I just don't know exactly what you're proposing here.

Thanks,
Nick