Re: [PATCH 0/3] Get device's parent from parent field, fix sleeping IRQs disabled

[Andy Shevchenko <andriy.shevchenko@intel.com>]

On Wed, Nov 03, 2021 at 04:47:44PM +0100, Petr Mladek wrote:
> On Wed 2021-11-03 15:50:04, John Ogness wrote:
> > On 2021-11-03, Sakari Ailus <sakari.ailus@linux.intel.com> wrote:
> > > This set changes getting fwnode's parent on ACPI fwnode so it no longer
> > > needs a semaphore, using struct acpi_device->parent field instead of
> > > calling acpi_get_parent(). The semaphore is being acquired when the
> > > device's full path is printed which now takes place local IRQs disabled:
> > >
> > > --------8<------------------------
> > > BUG: sleeping function called from invalid context at kernel/locking/semaphore.c:163
> > >
> > > ...
> > >
> > > Call Trace:
> > >  <TASK>
> > >  dump_stack_lvl+0x57/0x7d
> > >  __might_resched.cold+0xf4/0x12f
> > >  down_timeout+0x21/0x70
> > >  acpi_os_wait_semaphore+0x63/0x180
> > >  acpi_ut_acquire_mutex+0x123/0x1ba
> > >  acpi_get_parent+0x30/0x71
> > >  acpi_node_get_parent+0x64/0x90
> > >  ? lock_acquire+0x1a0/0x300
> > >  fwnode_count_parents+0x6d/0xb0
> > >  fwnode_full_name_string+0x18/0x90
> > >  fwnode_string+0xd7/0x140
> > >  vsnprintf+0x1ec/0x4f0
> > >  va_format.constprop.0+0x6a/0x130
> > >  vsnprintf+0x1ec/0x4f0
> > >  vprintk_store+0x271/0x5a0
> > >  ? rcu_read_lock_sched_held+0x12/0x70
> > >  ? lock_release+0x228/0x310
> > >  ? acpi_initialize_hp_context+0x50/0x50
> > >  vprintk_emit+0xd5/0x340
> > >  _printk+0x58/0x6f
> > ...
> > > --------8<------------------------
> > >
> > > I guess one could argue it wasn't great to begin with that getting
> > > fwnode's parent required a semaphore to begin with, nevertheless John's
> > > patch made it a concrete problem. Added Cc: stable, too.
> 
> It looks like a generic problem.
> 
> If I get it properly, we should make sure that any struct
> fwnode_operations implementation will _not_ use sleeping locks in
> the .get_parent() callback. Or anything that is called indirectly
> from from vsprintf.

Not sure how I may help here, but if there is a requirement to all *printf()
to be non-sleepable, then yes, this should be done in that way. I'm not sure
if every single %p extension does this...

It may require to update fwnode (source code) documentation to reflect the
requirement. Also might_sleep() / migth_sleep_if() in the API calls may
shed a light on (potential) issues.

-- 
With Best Regards,
Andy Shevchenko