Re: [RFC] [PATCH 2.6.37-rc5-tip 5/20] 5: Uprobes:register/unregister probes.

[Srikar Dronamraju]

* Peter Zijlstra <peterz@xxxxxxxxxxxxx> [2011-01-26 16:45:56]:

> On Wed, 2011-01-26 at 21:00 +0530, Srikar Dronamraju wrote:
> > * Peter Zijlstra <peterz@xxxxxxxxxxxxx> [2011-01-26 11:11:48]:
> > 
> > > On Wed, 2011-01-26 at 13:25 +0530, Srikar Dronamraju wrote:
> > > > 
> > > > > > +
> > > > > > +               list_add(&mm->uprobes_list, &tmp_list);
> > > > > > +               mm->uprobes_vaddr = vma->vm_start + offset;
> > > > > > +       }
> > > > > > +       spin_unlock(&mapping->i_mmap_lock);
> > > > > 
> > > > > Both this and unregister are racy, what is to say:
> > > > >  - the vma didn't get removed from the mm
> > > > >  - no new matching vma got added
> > > > > 
> > > > 
> > > > register_uprobe, unregister_uprobe, uprobe_mmap are all synchronized by
> > > > uprobes_mutex. So I dont see one unregister_uprobe getting thro when
> > > > another register_uprobe is working with a vma.
> > > > 
> > > > If I am missing something elementary, please explain a bit more.
> > > 
> > > afaict you're not holding the mmap_sem, so userspace can simply unmap
> > > the vma.
> > 
> > When we do the actual insert/remove of the breakpoint we hold the
> > mmap_sem. During the actual insertion/removal, if the vma for the
> > specific inode is not found, we just come out without doing the
> > actual insertion/deletion.
> 
> Right, but then install_uprobe() should:
> 
>  - lookup the vma relating to the address you stored,

We already do this thro get_user_pages in write_opcode().

>  - validate that the vma is indeed a map of the right inode

We can add a check in write_opcode( we need to pass the inode to
write_opcode).

>  - validate that the offset of the probe corresponds with the stored
> address

I am not clear on this. We would have derived the address from the
offset. So is that we check for
 (vaddr == vma->vm_start + uprobe->offset)

> 
> Otherwise you can race with unmap/map and end up installing the probe in
> a random location.
> 
> Also, I think the whole thing goes funny if someone maps the same text
> twice ;-)

I am not sure if we can map the same text twice. If something like
this is possible then we would have 2 addresses for each function.
So how does the linker know which address to jump to out of the 2 or
multiple matching addresses. What would be the usecases for same
text being mapped multiple times and both being executable?

-- 
Thanks and Regards
Srikar
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/