Re: [PATCH v2] powerpc/mpic: Fix memory leak of mpic on error paths

[Amit Machhiwal <amachhiw@linux.ibm.com>]

Hi Jyoti,

Please find my responses inline below.

On 2026/10/01 05:29 PM, Jyoti Gupta wrote:
> On 2026-10-01 00:20, Amit Machhiwal wrote:
> > On 2026/09/30 11:39 PM, Jyoti Gupta wrote:
> > > mpic_alloc() allocates struct mpic (and optionally the protected
> > > sources bitmap) but does not free them when mpic_setup_error_int() or
> > > irq_domain_create_linear() fails, leaking the memory.
> > > 
> > > Free the bitmap and the mpic in the common error path before
> > > dropping the
> > > device node reference.
> > > 
> > > Fixes: 996983b75ceb ("powerpc/mpic: Search for open-pic device-tree
> > > node if NULL")
> > > Signed-off-by: Jyoti Gupta <jyoti@linux.ibm.com>
> > > ---
> > >  arch/powerpc/sysdev/mpic.c | 11 +++++------
> > >  1 file changed, 5 insertions(+), 6 deletions(-)
> > 
> > Thanks for addressing the comments from v1.
> > 
> > A couple of minor suggestions:
> > 
> > 1. Changelog & Links: It is always a good practice to include a
> > changelog below
> >    the "---" line detailing the changes made in this version (e.g. "v1
> > -> v2:
> >    ...") along with a link to the previous version(s) on
> > lore.kernel.org. This
> >    helps reviewers and maintainers quickly track the evolution of the
> > patch.
> > 
> 
> Got it, thank you for the suggestion. I'll keep it in mind.
> 
> > 2. Minor nit in cleanup logic below:
> > 
> > > 
> > > diff --git a/arch/powerpc/sysdev/mpic.c b/arch/powerpc/sysdev/mpic.c
> > > index 257886ec3f70..b21a880777cc 100644
> > > --- a/arch/powerpc/sysdev/mpic.c
> > > +++ b/arch/powerpc/sysdev/mpic.c
> > > @@ -1389,7 +1389,7 @@ struct mpic * __init mpic_alloc(struct
> > > device_node *node,
> > >  		if (fsl_version >= 0x401) {
> > >  			ret = mpic_setup_error_int(mpic, intvec_top - 13);
> > >  			if (ret)
> > > -				return NULL;
> > > +				goto err_of_node_put;
> > >  		}
> > > 
> > >  	}
> > > @@ -1488,12 +1488,8 @@ struct mpic * __init mpic_alloc(struct
> > > device_node *node,
> > >  						 intvec_top,
> > >  						 &mpic_host_ops, mpic);
> > > 
> > > -	/*
> > > -	 * FIXME: The code leaks the MPIC object and mappings here; this
> > > -	 * is very unlikely to fail but it ought to be fixed anyways.
> > > -	 */
> > >  	if (mpic->irqhost == NULL)
> > > -		return NULL;
> > > +		goto err_of_node_put;
> > > 
> > >  	/* Display version */
> > >  	switch (greg_feature & MPIC_GREG_FEATURE_VERSION_MASK) {
> > > @@ -1527,6 +1523,9 @@ struct mpic * __init mpic_alloc(struct
> > > device_node *node,
> > >  	return mpic;
> > > 
> > >  err_of_node_put:
> > > +	if (mpic)
> > > +		bitmap_free(mpic->protected);
> > > +	kfree(mpic);
> > 
> > Both bitmap_free() and kfree() safely handle NULL pointers (and mpic is
> > guaranteed to be non-NULL at this label anyway), so the 'if (mpic)'
> > check is
> > redundant.
> > 
> 
> Thanks for the review.
> 
> Regarding the 'if (mpic)' check, I think we may need to keep it. While
> bitmap_free() and kfree() both safely handle NULL, the concern here is
> dereferencing mpic to obtain mpic->protected. There are two paths that
> reach err_of_node_put before mpic is allocated (line 1276):
> of_address_to_resource() failure at line 1258 and kzalloc_obj() returning
> NULL at line 1278. In both cases, mpic->protected would dereference a
> NULL/garbage pointer and crash.

Good catch on the early error paths before mpic allocation!

However, in v2, 'struct mpic *mpic' is left uninitialized at function entry
(since the '= NULL' initialization was dropped).

Looking more closely, if of_address_to_resource() fails at line 1258, control
jumps to err_of_node_put where 'mpic' contains uninitialized stack garbage.  In
that case, 'if (mpic)' evaluates to true, leading to an invalid pointer
dereference when evaluating mpic->protected and a crash on kfree(mpic).

Rather than initializing mpic to NULL and checking 'if (mpic)' on error paths,
the preferred approach in Linux is to use layered error labels.  In fact,
Documentation/process/coding-style.rst (Section 7, "Centralized exiting of
functions") specifically describes this pattern when discussing "one err bugs":


  A common type of bug to be aware of is ``one err bugs`` which look like this:

  .. code-block:: c

  	err:
  		kfree(foo->bar);
  		kfree(foo);
  		return ret;

  The bug in this code is that on some exit paths ``foo`` is NULL.  Normally the
  fix for this is to split it up into two error labels ``err_free_bar:`` and
  ``err_free_foo:``:

  .. code-block:: c

  	err_free_bar:
  		kfree(foo->bar);
  	err_free_foo:
  		kfree(foo);
  		return ret;

Applying that pattern here gives:

diff --git a/arch/powerpc/sysdev/mpic.c b/arch/powerpc/sysdev/mpic.c
index 257886ec3f70..d799cf4da7bd 100644
--- a/arch/powerpc/sysdev/mpic.c
+++ b/arch/powerpc/sysdev/mpic.c
@@ -1389,7 +1389,7 @@ struct mpic * __init mpic_alloc(struct device_node *node,
                if (fsl_version >= 0x401) {
                        ret = mpic_setup_error_int(mpic, intvec_top - 13);
                        if (ret)
-                               return NULL;
+                               goto err_free_mpic;
                }
 
        }
@@ -1488,12 +1488,8 @@ struct mpic * __init mpic_alloc(struct device_node *node,
                                                 intvec_top,
                                                 &mpic_host_ops, mpic);
 
-       /*
-        * FIXME: The code leaks the MPIC object and mappings here; this
-        * is very unlikely to fail but it ought to be fixed anyways.
-        */
        if (mpic->irqhost == NULL)
-               return NULL;
+               goto err_free_mpic;
 
        /* Display version */
        switch (greg_feature & MPIC_GREG_FEATURE_VERSION_MASK) {
@@ -1526,6 +1522,9 @@ struct mpic * __init mpic_alloc(struct device_node *node,
 
        return mpic;
 
+err_free_mpic:
+       bitmap_free(mpic->protected);
+       kfree(mpic);
 err_of_node_put:
        of_node_put(node);
        return NULL;

This ensures:

1. Early error paths (of_address_to_resource and kzalloc_obj failures) jump
   directly to err_of_node_put without touching mpic.
2. Later error paths jump to err_free_mpic where mpic is guaranteed to be valid.
3. No 'if (mpic)' check or 'struct mpic *mpic = NULL;' is needed.

What do you think?

Thanks,
Amit