Groups | Search | Server Info | Keyboard shortcuts | Login | Register [http] [https] [nntp] [nntps]

Groups > linux.kernel > #1612223

Re: [PATCH v2 7/8] bus: brcmstb_gisb: add notifier handling

Show all headers | View raw

On Wed, Mar 29, 2017 at 10:39:11AM -0700, Doug Berger wrote:
> On 03/29/2017 03:13 AM, Mark Rutland wrote:

> >> +static int dump_gisb_error(struct notifier_block *self, unsigned long v,
> >> +			   void *p)

> >> +	return 0;
> > 
> > I think this should be NOTIFY_OK.
> > 
> I used dump_mem_limit() as a template and didn't catch this (work to
> do...).  Upon review I think I would prefer NOTIFY_DONE since this call
> is opportunistic (i.e. it is taking the opportunity to check whether
> additional diagnostic data is available to display) and has no interest
> in affecting the overall handling of the event.

That's fine by me.

Does the distinction matter here?

Most notifer users treat NOTIFY_OK and NOTIFY_DONE as equivalent, and
notifier_call_chain only terminates when it sees NOTIFY_STOP_MASK.

[...]

> >> +	if (list_is_singular(&brcmstb_gisb_arb_device_list)) {
> >> +		register_die_notifier(&gisb_error_notifier);
> >> +		atomic_notifier_chain_register(&panic_notifier_list,
> >> +					       &gisb_error_notifier);
> > 
> > I don't think this is quite right. A notifier_block can only be
> > registered to one notifier chain at a time, and this has the potential
> > to corrupt both chains.
> > 
> A VERY good point thanks for pointing this out.
> 
> > I also think you only need to register the panic notifier. An SError
> > should always result in a panic.
> > 
> That was my initial thought as well.  However, testing revealed that the
> bad mode Oops actually exits the user space process and doesn't reach
> the panic so there was no helpful diagnostic message.  This may be in
> line with your comments about insufficient fatality of failures in PATCH
> v2 6/8, but it actually is more in line with our desired behavior for
> the aborted write.  Setting the notify on die gave us the result we are
> looking for, but as noted above I should have created a separate notifier.
> 
> I had hoped that the same approach (i.e. die notifier) would remove the
> need for PATCH v2 6/8 as well, but I found that the Unhandled fault
> error didn't actually die from user mode.

In my mind it's a bug that we don't treat those errors more fatally.

I'll try to dig into that.

Thanks,
Mark.

Back to linux.kernel | Previous | Next — Previous in thread | Next in thread | Find similar | Unroll thread

Thread

[PATCH v2 0/8] bus: brcmstb_gisb: add support for GISBv7 arbiter Doug Berger <opendmb@gmail.com> - 2017-03-28 23:40 +0200
  [PATCH v2 5/8] bus: brcmstb_gisb: correct support for 64-bit address output Doug Berger <opendmb@gmail.com> - 2017-03-28 23:40 +0200
  [PATCH v2 2/8] arm64: mm: mark fault_info __ro_after_init Doug Berger <opendmb@gmail.com> - 2017-03-28 23:40 +0200
    Re: [PATCH v2 2/8] arm64: mm: mark fault_info __ro_after_init Mark Rutland <mark.rutland@arm.com> - 2017-03-29 13:30 +0200
  [PATCH v2 4/8] bus: brcmstb_gisb: Correct hooking of ARM aborts Doug Berger <opendmb@gmail.com> - 2017-03-28 23:40 +0200
  [PATCH v2 7/8] bus: brcmstb_gisb: add notifier handling Doug Berger <opendmb@gmail.com> - 2017-03-28 23:40 +0200
    Re: [PATCH v2 7/8] bus: brcmstb_gisb: add notifier handling Mark Rutland <mark.rutland@arm.com> - 2017-03-29 12:20 +0200
      Re: [PATCH v2 7/8] bus: brcmstb_gisb: add notifier handling Doug Berger <opendmb@gmail.com> - 2017-03-29 19:40 +0200
        Re: [PATCH v2 7/8] bus: brcmstb_gisb: add notifier handling Mark Rutland <mark.rutland@arm.com> - 2017-03-29 20:20 +0200
  [PATCH v2 1/8] arm64: mm: Allow installation of memory abort handlers Doug Berger <opendmb@gmail.com> - 2017-03-28 23:40 +0200
    Re: [PATCH v2 1/8] arm64: mm: Allow installation of memory abort  handlers Mark Rutland <mark.rutland@arm.com> - 2017-03-29 13:40 +0200
  [PATCH v2 3/8] bus: brcmstb_gisb: Use register offsets with writes too Doug Berger <opendmb@gmail.com> - 2017-03-28 23:40 +0200

csiph-web