On Thu, Aug 09, 2018 at 04:17:13PM +0200, Oleg Nesterov wrote:
> On 08/09, Jiri Olsa wrote:
> >
> > @@ -523,8 +523,10 @@ int modify_user_hw_breakpoint(struct perf_event *bp,
> > struct perf_event_attr *att
> > perf_event_disable(bp);
> >
> > err = modify_user_hw_breakpoint_check(bp, attr, false);
> > - if (err)
> > + if (err) {
> > + bp->attr.disabled = 1;
> > return err;
>
> Yes, but on the second thought... Can't we simply do
>
> int modify_user_hw_breakpoint(struct perf_event *bp, struct
> perf_event_attr *attr)
> {
> int err;
>
> /*
> * modify_user_hw_breakpoint can be invoked with IRQs disabled
> and hence it
> * will not be possible to raise IPIs that invoke
> __perf_event_disable.
> * So call the function directly after making sure we are
> targeting the
> * current task.
> */
> if (irqs_disabled() && bp->ctx && bp->ctx->task == current)
> perf_event_disable_local(bp);
> else
> perf_event_disable(bp);
>
> err = modify_user_hw_breakpoint_check(bp, attr, false);
>
> if (!bp.attr->disabled)
> perf_event_enable(bp);
>
> return err;
> }
>
> instead of this and the next patch?
>
> We can do this because (as you pointed out) validate_hw_breakpoint() has
> already
> gone in -tip tree, and (afaics) modify_user_hw_breakpoint_check() never
> changes
> perf_event's state on failure, so we can safely "restart" this bp if it was
> enabled
> before.
>
> 1. This is what we had before the recent
> f67b15037a7a50c57f72e69a6d59941ad90a0f0f
> "perf/hwbp: Simplify the perf-hwbp code, fix documentation".
>
> Note that this commit was actually the bug fix which introduced another
> problem
> fixed by your 2/5.
>
> But see above, perf_event_enable() is no longer unsafe after
> modify_user_hw_breakpoint_check(), we can restore the previous semantics.
>
> 2. This matches perf_event_modify_breakpoint(). Which btw can be simplified a
> bit,
> it too can simply do
>
> if (!bp->attr.disabled)
> _perf_event_enable(bp);
>
> return err;
yep, seems good.. will send new version
thanks,
jirka
