10.12.2021 22:14, Rafael J. Wysocki пишет:

On Fri, Dec 10, 2021 at 8:04 PM Dmitry Osipenko [off-list ref] wrote:

quoted

10.12.2021 21:27, Rafael J. Wysocki пишет:

quoted

On Mon, Nov 29, 2021 at 12:34 PM Dmitry Osipenko [off-list ref] wrote:

quoted

29.11.2021 03:26, Michał Mirosław пишет:

quoted

On Mon, Nov 29, 2021 at 12:06:19AM +0300, Dmitry Osipenko wrote:

quoted

28.11.2021 03:28, Michał Mirosław пишет:

quoted

On Fri, Nov 26, 2021 at 09:00:41PM +0300, Dmitry Osipenko wrote:

quoted

Add sanity check which ensures that there are no two restart handlers
registered with the same priority. Normally it's a direct sign of a
problem if two handlers use the same priority.

The patch doesn't ensure the property that there are no duplicated-priority
entries on the chain.

It's not the exact point of this patch.

quoted

I'd rather see a atomic_notifier_chain_register_unique() that returns
-EBUSY or something istead of adding an entry with duplicate priority.
That way it would need only one list traversal unless you want to
register the duplicate anyway (then you would call the older
atomic_notifier_chain_register() after reporting the error).

The point of this patch is to warn developers about the problem that
needs to be fixed. We already have such troubling drivers in mainline.

It's not critical to register different handlers with a duplicated
priorities, but such cases really need to be corrected. We shouldn't
break users' machines during transition to the new API, meanwhile
developers should take action of fixing theirs drivers.

quoted

(Or you could return > 0 when a duplicate is registered in
atomic_notifier_chain_register() if the callers are prepared
for that. I don't really like this way, though.)

I had a similar thought at some point before and decided that I'm not in
favor of this approach. It's nicer to have a dedicated function that
verifies the uniqueness, IMO.

I don't like the part that it traverses the list second time to check
the uniqueness. But actually you could avoid that if
notifier_chain_register() would always add equal-priority entries in
reverse order:

 static int notifier_chain_register(struct notifier_block **nl,
              struct notifier_block *n)
 {
      while ((*nl) != NULL) {
              if (unlikely((*nl) == n)) {
                      WARN(1, "double register detected");
                      return 0;
              }
-             if (n->priority > (*nl)->priority)
+             if (n->priority >= (*nl)->priority)
                      break;
              nl = &((*nl)->next);
      }
      n->next = *nl;
      rcu_assign_pointer(*nl, n);
      return 0;
 }

Then the check for uniqueness after adding would be:

 WARN(nb->next && nb->priority == nb->next->priority);

We can't just change the registration order because invocation order of
the call chain depends on the registration order

It doesn't if unique priorities are required and isn't that what you want?

quoted

and some of current
users may rely on that order. I'm pretty sure that changing the order
will have unfortunate consequences.

Well, the WARN() doesn't help much then.

Either you can make all of the users register with unique priorities,
and then you can make the registration reject non-unique ones, or you
cannot assume them to be unique.

There is no strong requirement for priorities to be unique, the reboot.c
code will work properly.

In which case adding the WARN() is not appropriate IMV.

Also I've looked at the existing code and at least in some cases the
order in which the notifiers run doesn't matter.  I'm not sure what
the purpose of this patch is TBH.

The purpose is to let developer know that driver needs to be corrected.

quoted

The potential problem is on the user's side and the warning is intended
to aid the user.

Unless somebody has the panic_on_warn mentioned previously set and
really the user need not understand what the WARN() is about.  IOW,
WARN() helps developers, not users.

quoted

We can make it a strong requirement, but only after converting and
testing all kernel drivers.

Right.

quoted

I'll consider to add patches for that.

But can you avoid adding more patches to this series?

I won't add more patches since such patches can be added only after
completion of transition to the new API of the whole kernel.