lkml.org 
[lkml]   [2011]   [Jun]   [16]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH v4 3.0-rc2-tip 7/22] 7: uprobes: mmap and fork hooks.
* Peter Zijlstra <peterz@infradead.org> [2011-06-16 14:00:26]:

> On Thu, 2011-06-16 at 08:56 +0530, Srikar Dronamraju wrote:
> > * Peter Zijlstra <peterz@infradead.org> [2011-06-15 20:11:26]:
> >
> > > On Tue, 2011-06-07 at 18:29 +0530, Srikar Dronamraju wrote:
> > > > + up_write(&mm->mmap_sem);
> > > > + mutex_lock(&uprobes_mutex);
> > > > + down_read(&mm->mmap_sem);
> > >
> > > egads, and all that without a comment explaining why you think that is
> > > even remotely sane.
> > >
> > > I'm not at all convinced, it would expose the mmap() even though you
> > > could still decide to tear it down if this function were to fail, I bet
> > > there's some funnies there.
> >
> > The problem is with lock ordering. register/unregister operations
> > acquire uprobes_mutex (which serializes register unregister and the
> > mmap_hook) and then holds mmap_sem for read before they insert a
> > breakpoint.
> >
> > But the mmap hook would be called with mmap_sem held for write. So
> > acquiring uprobes_mutex can result in deadlock. Hence we release the
> > mmap_sem, take the uprobes_mutex and then again hold the mmap_sem.
>
> Sure, I saw why you wanted to do it, I'm just not quite convinced its
> safe to do and something like this definitely wants a comment explaining
> why its safe to drop mmap_sem.
>
> > After we re-acquire the mmap_sem, we do check if the vma is valid.
>
> But you don't on the return path, and if !ret
> mmap_region():unmap_and_free_vma will be touching vma again to remove
> it.
>

Agree.

> > Do we have better solutions?
>
> /me kicks the brain into gear and walks off to get a fresh cup of tea.
>
> So the reason we take uprobes_mutex there is to avoid probes from going
> away while you're installing them, right?

It serializes register/unregister/mmap operations.

>
> So we start by doing this add_to_temp_list() thing (horrid name), which
> iterates the probes on this inode under uprobes_treelock and adds them
> to a list.
>
> Then we iterate the list, installing the probles.
>
> How about we make the initial pass under uprobes_treelock take a
> references on the probe, and then after install_breakpoint() succeeds we
> again take uprobes_treelock and validate the uprobe still exists in the
> tree and drop the extra reference, if not we simply remove the
> breakpoint again and continue like it never existed.
>
> That should avoid the need to take uprobes_mutex and not require
> dropping mmap_sem, right?

Now since a register and mmap operations can run in parallel, we could
have subtle race conditions like this:

1. register_uprobe inserts the uprobe in RB tree.
2. register_uprobe loops thro vmas and inserts breakpoints.

3. mmap is called for same inode, mmap_uprobe() takes reference;
4. mmap completes insertion and releases reference.
5. register uprobe tries to install breakpoint on one vma fails and not
due to -ESRCH or -EEXIST.
6. register_uprobe rolls back all install breakpoints except the one
inserted by mmap.

We end up with breakpoints that we have inserted by havent cleared.

Similarly unregister_uprobe might be looping to remove the breakpoints
when mmap comes in installs the breakpoint and returns.
unregister_uprobe might erase the uprobe from rbtree after mmap is done.

--
Thanks and Regards
Srikar


\
 
 \ /
  Last update: 2011-06-16 15:15    [from the cache]
©2003-2011 Jasper Spaans