Re: [PATCH v7 06/16] lockdep: Detect and handle hist_lock ring buffer overwrite

From: Peter Zijlstra
Date: Thu Jul 13 2017 - 05:51:17 EST


On Thu, Jul 13, 2017 at 05:57:46PM +0900, Byungchul Park wrote:
> On Thu, Jul 13, 2017 at 10:14:42AM +0200, Peter Zijlstra wrote:
> > On Thu, Jul 13, 2017 at 11:07:45AM +0900, Byungchul Park wrote:
> > > Does my approach have problems, rewinding to 'original idx' on exit and
> > > deciding whether overwrite or not? I think, this way, no need to do the
> > > drastic work. Or.. does my one get more overhead in usual case?
> >
> > So I think that invalidating just the one entry doesn't work; the moment
>
> I think invalidating just the one is enough. After rewinding, the entry
> will be invalidated and the ring buffer starts to be filled forward from
> the point with valid ones. When commit, it will proceed backward with
> valid ones until meeting the invalidated entry and stop.
>
> IOW, in case of (overwritten)
>
> rewind to here
> |
> ppppppppppiiiiiiiiiiiiiiii
> iiiiiiiiiiiiiii
>
> invalidate it on exit_irq
> and start to fill from here again
> |
> pppppppppxiiiiiiiiiiiiiiii
> iiiiiiiiiiiiiii
>
> when commit occurs here
> |
> pppppppppxpppppppppppiiiii
>
> do commit within this range
> |<---------|
> pppppppppxpppppppppppiiiii
>
> So I think this works and is much simple. Anything I missed?


wait_for_completion(&C);
atomic_inc_return();

mutex_lock(A1);
mutex_unlock(A1);


<IRQ>
spin_lock(B1);
spin_unlock(B1);

...

spin_lock(B64);
spin_unlock(B64);
</IRQ>


mutex_lock(A2);
mutex_unlock(A2);

complete(&C);


That gives:

xhist[ 0] = A1
xhist[ 1] = B1
...
xhist[63] = B63

then we wrap and have:

xhist[0] = B64

then we rewind to 1 and invalidate to arrive at:

xhist[ 0] = B64
xhist[ 1] = NULL <-- idx
xhist[ 2] = B2
...
xhist[63] = B63


Then we do A2 and get

xhist[ 0] = B64
xhist[ 1] = A2 <-- idx
xhist[ 2] = B2
...
xhist[63] = B63

and the commit_xhlocks() will happily create links between C and A2,
B2..B64.

The C<->A2 link is desired, the C<->B* are not.