Re: [PATCH 0/2] perf/hw_breakpoint: Fix breakpoint overcommit issue

[Peter Zijlstra]

On Wed, Jul 10, 2019 at 05:34:07PM +0200, Frederic Weisbecker wrote:
> On Wed, Jul 10, 2019 at 04:04:21PM +0200, Peter Zijlstra wrote:
> > On Tue, Jul 09, 2019 at 03:48:19PM +0200, Frederic Weisbecker wrote:
> > > 
> > > Syzbot has found a breakpoint overcommit issue:
> > > 
> > > https://lore.kernel.org/lkml/000000000000639f6a0584d11b82@xxxxxxxxxx/
> > > 
> > > It took me a long time to find out what the actual root problem was. Also
> > > its reproducer only worked on a few month old kernel but it didn't feel like
> > > the issue was actually solved.
> > > 
> > > I eventually cooked a reproducer that works with latest upstream, see in
> > > the end of this message.
> > > 
> > > The fix is just a few liner but implies to shut down the context swapping
> > > optimization for contexts containing breakpoints.
> > > 
> > > Also I feel like uprobes may be concerned as well as it seems to make use
> > > of event.hw->target after pmu::init().
> > 
> > Can't we simply swizzle event.hw->target along too?
> 
> You mean remove it? But it's still needed by breakpoint code during all the event
> lifecycle (init, destroy and anytime in-between).

No, I meant flip hw->target when we flip the context. It would mean
iterating the events, which I suppose would suck.

> I wish we could use event->ctx->task instead but on pmu::init() there
> is no ctx yet (we could pass the task in parameter though) 

Right, that should be fairly easy.

> and on event->destroy() it's TASK_TOMBSTONE and retrieving the task at
> that time would be non trivial.

Well, right, we can maybe make TOMBSTONE be the LSB instead of the whole
word, then we can recover the task pointer... *yuck* though.