Re: [PATCH v2] ARM: OMAP2+: Fix device node reference counts

[Guenter Roeck]

On Wed, Mar 01, 2017 at 10:04:39AM -0800, Tony Lindgren wrote:
> * Guenter Roeck <linux@xxxxxxxxxxxx> [170228 11:55]:
> > After commit 'of: fix of_node leak caused in of_find_node_opts_by_path',
> > the following error may be reported when running omap images.
> > 
> > OF: ERROR: Bad of_node_put() on /ocp@68000000
> > CPU: 0 PID: 0 Comm: swapper Not tainted 4.10.0-rc7-next-20170210 #1
> > Hardware name: Generic OMAP3-GP (Flattened Device Tree)
> > [<c0310604>] (unwind_backtrace) from [<c030bbf4>] (show_stack+0x10/0x14)
> > [<c030bbf4>] (show_stack) from [<c05add8c>] (dump_stack+0x98/0xac)
> > [<c05add8c>] (dump_stack) from [<c05af1b0>] (kobject_release+0x48/0x7c)
> > [<c05af1b0>] (kobject_release)
> > 	from [<c0ad1aa4>] (of_find_node_by_name+0x74/0x94)
> > [<c0ad1aa4>] (of_find_node_by_name)
> > 	from [<c1215bd4>] (omap3xxx_hwmod_is_hs_ip_block_usable+0x24/0x2c)
> > [<c1215bd4>] (omap3xxx_hwmod_is_hs_ip_block_usable) from
> > [<c1215d5c>] (omap3xxx_hwmod_init+0x180/0x274)
> > [<c1215d5c>] (omap3xxx_hwmod_init)
> > 	from [<c120faa8>] (omap3_init_early+0xa0/0x11c)
> > [<c120faa8>] (omap3_init_early)
> > 	from [<c120fb2c>] (omap3430_init_early+0x8/0x30)
> > [<c120fb2c>] (omap3430_init_early)
> > 	from [<c1204710>] (setup_arch+0xc04/0xc34)
> > [<c1204710>] (setup_arch) from [<c1200948>] (start_kernel+0x68/0x38c)
> > [<c1200948>] (start_kernel) from [<8020807c>] (0x8020807c)
> > 
> > of_find_node_by_name() drops the reference to the passed device node.
> > Use of_get_child_by_name() instead. Also, release references to
> > device nodes obtained with of_find_node_by_name() and
> > of_get_child_by_name() after they are no longer needed.
> > 
> > While at it, clean up the code and change the return type of
> > omap3xxx_hwmod_is_hs_ip_block_usable() to bool to match its use
> > and the return type of of_device_is_available().
> > 
> > Cc: Qi Hou <qi.hou@xxxxxxxxxxxxx>
> > Cc: Peter Rosin <peda@xxxxxxxxxx>
> > Cc: Rob Herring <robh@xxxxxxxxxx>
> > Signed-off-by: Guenter Roeck <linux@xxxxxxxxxxxx>
> > ---
> > v2: Change subject ('Grab reference to device nodes where needed'
> >     didn't really cover all the changes made)
> >     Use of_get_child_by_name() instead of of_find_node_by_name()
> >     Drop references to device nodes as needed
> >     Change return type of omap3xxx_hwmod_is_hs_ip_block_usable()
> >     to bool
> 
> OK so now we have a commit id for it and should have:
> 
> Fixes: 0549bde0fcb1 ("of: fix of_node leak caused in
> of_find_node_opts_by_path")
> 
Not really; 0549bde0fcb1 fixes a bug and hides _this_ bug.
More appropriate, if that would exist, would be something like

Exposed-by: 0549bde0fcb1 ("of: fix of_node leak caused in ...")

> What about other leaky users of of_find_node_by_name() and
> of_get_child_by_name()?
> 
> We have those in:
> 
> control.c
> display.c
> omap_device.c
> omap_hwmod.c
> 

I am sure there are many more. For example, the bug solved in my patch
"disappears" if one adds some delays (or just log messages) at strategic
areas in the code. Such delays result in node leaks elsewhere, which
again hides the bug I am trying to fix with my patch.

> So probably it really should be two fixes, one for the regression
> causing the error. Then another one for fixing the leaky np use.
> 
No problem; I 'll split into two patches.

Guenter

> Sorry for going back and forth here, I guess I did not full understand
> the second part of the change in your earlier version of the patch
> when you asked if it should be two patches.
> 

> Regards,
> 
> Tony