Re: arch/s390/pci/pci_event.c:101 __zpci_event_availability() error: we previously assumed 'zdev->zbus' could be null (see line 83)

[Niklas Schnelle <schnelle@linux.ibm.com>]

On 12/3/20 11:27 AM, Dan Carpenter wrote:
> tree:   https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
> head:   3bb61aa61828499a7d0f5e560051625fd02ae7e4
> commit: 3047766bc6ec9c6bc9ece85b45a41ff401e8d988 s390/pci: fix enabling a reserved PCI function
> 
> If you fix the issue, kindly add following tag as appropriate
> Reported-by: kernel test robot <lkp@intel.com>
> Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
> 
> smatch warnings:
> arch/s390/pci/pci_event.c:101 __zpci_event_availability() error: we previously assumed 'zdev->zbus' could be null (see line 83)
> 
> vim +101 arch/s390/pci/pci_event.c
> 
> aa3b7c296732b43 Sebastian Ott   2013-12-12   76  static void __zpci_event_availability(struct zpci_ccdf_avail *ccdf)
> cbc0dd1f856b52b Jan Glauber     2012-11-29   77  {
> cbc0dd1f856b52b Jan Glauber     2012-11-29   78  	struct zpci_dev *zdev = get_zdev_by_fid(ccdf->fid);
> 9a99649f2a89fdf Sebastian Ott   2016-01-29   79  	struct pci_dev *pdev = NULL;
> 623bd44d3f277b7 Sebastian Ott   2017-05-09   80  	enum zpci_state state;
> d795ddad36cbc82 Sebastian Ott   2013-11-15   81  	int ret;
> cbc0dd1f856b52b Jan Glauber     2012-11-29   82  
> 05bc1be6db4b268 Pierre Morel    2020-03-23  @83  	if (zdev && zdev->zbus && zdev->zbus->bus)
>                                                                      ^^^^^^^^^
> Check for NULL
> 
> 44510d6fa0c00aa Pierre Morel    2020-04-22   84  		pdev = pci_get_slot(zdev->zbus->bus, zdev->devfn);
> 9a99649f2a89fdf Sebastian Ott   2016-01-29   85  
> 1f1dcbd4f23bd1f Sebastian Ott   2013-10-22   86  	zpci_err("avail CCDF:\n");
> 1f1dcbd4f23bd1f Sebastian Ott   2013-10-22   87  	zpci_err_hex(ccdf, sizeof(*ccdf));
> cbc0dd1f856b52b Jan Glauber     2012-11-29   88  
> cbc0dd1f856b52b Jan Glauber     2012-11-29   89  	switch (ccdf->pec) {
> 7fc611ff3ff1a0b Sebastian Ott   2015-06-16   90  	case 0x0301: /* Reserved|Standby -> Configured */
> 7fc611ff3ff1a0b Sebastian Ott   2015-06-16   91  		if (!zdev) {
> f606b3ef47c9f87 Pierre Morel    2020-03-25   92  			ret = clp_add_pci_device(ccdf->fid, ccdf->fh, 1);
> 7fc611ff3ff1a0b Sebastian Ott   2015-06-16   93  			break;
> 7fc611ff3ff1a0b Sebastian Ott   2015-06-16   94  		}
> fcf2f402937a669 Sebastian Ott   2013-12-18   95  		zdev->fh = ccdf->fh;
> f606b3ef47c9f87 Pierre Morel    2020-03-25   96  		zdev->state = ZPCI_FN_STATE_CONFIGURED;
> 3047766bc6ec9c6 Niklas Schnelle 2020-06-18   97  		ret = zpci_enable_device(zdev);
> 3047766bc6ec9c6 Niklas Schnelle 2020-06-18   98  		if (ret)
> 3047766bc6ec9c6 Niklas Schnelle 2020-06-18   99  			break;
> 3047766bc6ec9c6 Niklas Schnelle 2020-06-18  100  
> 3047766bc6ec9c6 Niklas Schnelle 2020-06-18 @101  		pdev = pci_scan_single_device(zdev->zbus->bus, zdev->devfn);
>                                                                                               ^^^^^^^^^^^^^^^^
> Unchecked dereference

First, thanks for reporting this is definitely appreciated!
We have also seen the same smatch report internally 
and I determined that this is a false positive.

This is because the existing zdev->zbus NULL check could already never
trigger. Sadly this is all a bit hard to see but let me try to share
my reasoning here. That said, we do plan a couple of cleanups/rework here 
that will also see the removal of the zdev->zbus NULL check at the start
of the function.

The existing check is here:

76 static void __zpci_event_availability(struct zpci_ccdf_avail *ccdf)
77 {
78 	struct zpci_dev *zdev = get_zdev_by_fid(ccdf->fid);
79 	struct pci_dev *pdev = NULL;
80 	enum zpci_state state;
81 	int ret;
82 
83 	if (zdev && zdev->zbus && zdev->zbus->bus)
84 		pdev = pci_get_slot(zdev->zbus->bus, zdev->devfn);

Now for both the new zdev->zbus->bus dereference and the check in line 83 we
check zdev != NULL beforehand. So that's a given.
If we have the zdev it came out of get_zdev_by_fid() (line 78) which found it in
the zpci_list.

It only ever gets into that list in zpci_create_device() and
gets removed if anything in there fails, including zpci_bus_register_device()
which when successful has definitely set zdev->zbus != NULL.
Furthermore zpci_create_device() is only called in clp_add_pci_device()
which is either called during init of zPCI (in the initial CLP List PCI)
or from the kthread that handles events via zpci_event_availability()
which checks the the initialization is complete.
Therefore there is no race for the window in which
zpci_create_device() has the zdev in the list but no zbus set.
Either way, this is definitely too hard to reason about e.g.
we should only add devices to the zpci_list once they are
fully initialized and we're working on a redesign of this.

If you prefer a short term fix so this doesn't trigger
again we could simply remove the zdev->zbus NULL check.

> 
... snip ...