Re: [PATCH v2 2/2] of/pci: Fix memory leak in of_pci_get_host_bridge_resources

[jeffy]

Hi Bjorn,

On 04/05/2017 03:18 AM, Bjorn Helgaas wrote:
> On Thu, Mar 23, 2017 at 5:58 PM, Dmitry Torokhov <dtor@xxxxxxxxxxxx> wrote:
> > On Thu, Mar 23, 2017 at 3:07 PM, Rob Herring <robh@xxxxxxxxxx> wrote:
> > > On Thu, Mar 23, 2017 at 3:12 AM, Jeffy Chen <jeffy.chen@xxxxxxxxxxxxxx> wrote:
> > > > Currently we only free the allocated resource struct when error.
> > > > This would cause memory leak after pci_free_resource_list.
> > > >
> > > > Signed-off-by: Jeffy Chen <jeffy.chen@xxxxxxxxxxxxxx>
> > > > ---
> > > >
> > > > Changes in v2:
> > > > Don't change the resource_list_create_entry's behavior.
> > > >
> > > >   drivers/of/of_pci.c | 57 +++++++++++++++++++++++------------------------------
> > > >   1 file changed, 25 insertions(+), 32 deletions(-)
> > > >
> > > > diff --git a/drivers/of/of_pci.c b/drivers/of/of_pci.c
> > > > index 0ee42c3..a0ec246 100644
> > > > --- a/drivers/of/of_pci.c
> > > > +++ b/drivers/of/of_pci.c
> > > > @@ -190,8 +190,7 @@ int of_pci_get_host_bridge_resources(struct device_node *dev,
> > > >                          struct list_head *resources, resource_size_t *io_base)
> > > >   {
> > > >          struct resource_entry *window;
> > > > -       struct resource *res;
> > > > -       struct resource *bus_range;
> > > > +       struct resource res;
> > > >          struct of_pci_range range;
> > > >          struct of_pci_range_parser parser;
> > > >          char range_type[4];
> > > > @@ -200,24 +199,24 @@ int of_pci_get_host_bridge_resources(struct device_node *dev,
> > > >          if (io_base)
> > > >                  *io_base = (resource_size_t)OF_BAD_ADDR;
> > > >
> > > > -       bus_range = kzalloc(sizeof(*bus_range), GFP_KERNEL);
> > > > -       if (!bus_range)
> > > > -               return -ENOMEM;
> > > > -
> > > >          pr_info("host bridge %s ranges:\n", dev->full_name);
> > > >
> > > > -       err = of_pci_parse_bus_range(dev, bus_range);
> > > > +       err = of_pci_parse_bus_range(dev, &res);
> > > >          if (err) {
> > > > -               bus_range->start = busno;
> > > > -               bus_range->end = bus_max;
> > > > -               bus_range->flags = IORESOURCE_BUS;
> > > > -               pr_info("  No bus range found for %s, using %pR\n",
> > > > -                       dev->full_name, bus_range);
> > > > +               res.start = busno;
> > > > +               res.end = bus_max;
> > > > +               res.flags = IORESOURCE_BUS;
> > > > +               pr_info("  No bus range found for %s\n", dev->full_name);
> > > >          } else {
> > > > -               if (bus_range->end > bus_range->start + bus_max)
> > > > -                       bus_range->end = bus_range->start + bus_max;
> > > > +               if (res.end > res.start + bus_max)
> > > > +                       res.end = res.start + bus_max;
> > > > +       }
> > > > +       window = pci_add_resource(resources, NULL);
> > > > +       if (!window) {
> > > > +               err = -ENOMEM;
> > > > +               goto parse_failed;
> > > >          }
> > > > -       pci_add_resource(resources, bus_range);
> > > > +       *window->res = res;
> > >
> > > Well, now this seems racy. You add a blank resource to the list first
> > > and then fill it in.
> >
> > Huh? There is absolutely no guarantees for concurrent access here.
> > pcI_add_resource_offset() first adds a resource and then modifies
> > offset. Here we add an empty resource and then fill it in.
>
> I don't really like this pattern either.  Even if there's no actual
> racy behavior, it takes more analysis than necessary to figure that
> out.
>
> pci_add_resource_offset() allocates a resource list entry, sets the
> offset, then adds it to the list.  It doesn't update a resource entry
> that might be visible to anybody else.  Here we do update a resource
> that is already visible to others because it's already on the list.
i was following ./drivers/pnp/resource.c, but i'm agree this is not a 
good way.

i'll upload a new version to fix this in another way. more ideas:
1/ pass a struct device to of_pci_get_host_bridge_resources and use 
devm_kzalloc
2/ add a new type of flags(or reuse IORESOURCE_AUTO) to tell 
pci_free_resource_list to kfree them)
3/ add new helpers of of_pci_add_resource[_offset] to alloc empty res, 
fill it, add to list.
> Bjorn
>
> BTW, please CC linux-pci on the entire series so it's easier to
> review.  I don't know where you envision having this applied, but I
> only apply things to the PCI tree after they appear on linux-pci.
oh, sorry, didn't notice that, will do in next version.
>