Re: [PATCH] PCI: rework new_id interface for known vendor/device values

[Alex Williamson]

On Tue, 2014-04-01 at 14:32 -0400, Bandan Das wrote:
> Alex Williamson <alex.williamson@xxxxxxxxxx> writes:
> 
> > On Mon, 2014-03-31 at 00:28 -0400, Bandan Das wrote:
> >> While using the new_id interface, the user can unintentionally feed
> >> incorrect values if the driver static table has a matching entry.
> >> This is possible since only the device and vendor fields are
> >> mandatory and the rest are optional. As a result, store_new_id
> >> will fill in default values that are then passed on to the driver
> >> and can have unintended consequences.
> >> 
> >> As an example, consider the ixgbe driver and the 82599EB network card :
> >> echo "8086 10fb" > /sys/bus/pci/drivers/ixgbe/new_id
> >> 
> >> This will pass a driver_data value of 0 to the driver whereas
> >> the index 0 in ixgbe actually points to a different set of card
> >> operations.
> >> 
> >> This change automatically selects the matching static entry if there
> >> is one for the newly created dynid. However, if the user intentionally
> >> wants a different set of values, she must provide all the 7 fields
> >> and the static entry will be ignored.
> >> 
> >> In most cases, this use case seems unnecessary, however, this
> >> is a common libvirt/KVM/device assignment scenario where the
> >> user might want to bind a device back to the host driver.
> >> 
> >> Signed-off-by: Bandan Das <bsd@xxxxxxxxxx>
> >> ---
> >>  drivers/pci/pci-driver.c | 42 ++++++++++++++++++++++++++++++++++++++----
> >>  1 file changed, 38 insertions(+), 4 deletions(-)
> >> 
> >> diff --git a/drivers/pci/pci-driver.c b/drivers/pci/pci-driver.c
> >> index 25f0bc6..187e572 100644
> >> --- a/drivers/pci/pci-driver.c
> >> +++ b/drivers/pci/pci-driver.c
> >> @@ -90,6 +90,24 @@ static void pci_free_dynids(struct pci_driver *drv)
> >>  	spin_unlock(&drv->dynids.lock);
> >>  }
> >>  
> >> +static const struct
> >> +pci_device_id *match_id_table_entry(struct device_driver *driver,
> >> +				    __u32 vendor, __u32 device)
> >> +{
> >> +	struct pci_driver *pdrv = to_pci_driver(driver);
> >> +	const struct pci_device_id *ids = pdrv->id_table;
> >> +
> >> +	if (ids) {
> >> +		while (ids->vendor || ids->subvendor || ids->class_mask) {
> >> +			if ((ids->vendor == vendor) && (ids->device == device))
> >> +				return ids;
> >> +			ids++;
> >> +		}
> >> +	}
> >> +
> >> +	return NULL;
> >> +}
> >> +
> >>  /**
> >>   * store_new_id - sysfs frontend to pci_add_dynid()
> >>   * @driver: target device driver
> >> @@ -102,7 +120,8 @@ static ssize_t
> >>  store_new_id(struct device_driver *driver, const char *buf, size_t count)
> >>  {
> >>  	struct pci_driver *pdrv = to_pci_driver(driver);
> >> -	const struct pci_device_id *ids = pdrv->id_table;
> >> +	const struct pci_device_id *ids = pdrv->id_table,
> >> +		*tids = NULL;
> >>  	__u32 vendor, device, subvendor=PCI_ANY_ID,
> >>  		subdevice=PCI_ANY_ID, class=0, class_mask=0;
> >>  	unsigned long driver_data=0;
> >> @@ -115,9 +134,24 @@ store_new_id(struct device_driver *driver, const char *buf, size_t count)
> >>  	if (fields < 2)
> >>  		return -EINVAL;
> >>  
> >> -	/* Only accept driver_data values that match an existing id_table
> >> -	   entry */
> >> -	if (ids) {
> >> +	tids = match_id_table_entry(driver, vendor, device);
> >> +
> >
> > Would it make more sense to construct a pci_dev, ex:
> >
> > if (fields != 7) {
> > 	struct pci_dev dev = { .subvendor = PCI_ANY_ID, .subdevice = PCI_ANY_ID };
> >
> > 	dev.vendor = vendor;
> > 	dev.device = device;
> > 	if (fields > 2)
> > 		dev.subvendor = subvendor;
> > 	if (fields > 3)
> > 		dev.subdevice = subdevice;
> > 	...
> >
> > 	if (pci_match_id(drv->id_table, &dev))
> > 		return -EEXIST;
> > }
> 
> I initially went ahead this way, but the compilation warns about frame size 
> being larger, possibly because of a kernel config option that's set in my config
> 
> drivers/pci/pci-driver.c:193:1: warning: the frame size of 
> 2264 bytes is larger than 2048 bytes [-Wframe-larger-than=]
> 
> Do you know if it is safe to ignore this ? This seems to be coming up 
> if I add the struct pdev.

struct pdev is pretty large, maybe just allocate one rather than putting
it on the stack.  The alternative might be the more intrusive path of
passing a struct pci_device_id rather than a struct pci_dev to the
existing match functions.

> >
> >> +	if (tids && (fields != 7)) {
> >> +
> >> +		subvendor = tids->subvendor;
> >> +		subdevice = tids->subdevice;
> >> +		class = tids->class;
> >> +		class_mask = tids->class_mask;
> >> +		driver_data = tids->driver_data;
> >
> > This doesn't look right.  First, we're potentially overwriting user
> > stored data for fields >2 but <7.  Second, we only matched on vendor &
> > device and could be filling the rest with data that isn't the best match
> > (and is guaranteed to just be a duplicate of a static table ID).
> >
> >> +
> >> +		pr_warn("pci: Using driver (%s) static DeviceID table entry for vendor 0x%04x and device 0x%04x",
> >> +			driver->name, vendor, device);
> >
> > I think we should be error'ing rather than inventing a duplicate ID to
> > insert.  How would a user ever know how to use remove_id to clean out
> > this new_id?  Thanks,
> 
> Ok, makes sense to just error out. Good point about remove_id, 
> didn't think about that.

Probably a good idea to check whether libvirt explodes from a write
failure or just ignores it.  Thanks,

Alex

> >> +
> >> +	} else if (ids) {
> >> +
> >> +		/* Only accept driver_data values that match an existing
> >> +		   id_table entry */
> >> +
> >>  		retval = -EINVAL;
> >>  		while (ids->vendor || ids->subvendor || ids->class_mask) {
> >>  			if (driver_data == ids->driver_data) {



--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/