Message ID | 1477944749-22822-1-git-send-email-gpiccoli@linux.vnet.ibm.com |
---|---|
State | Changes Requested |
Delegated to: | Jeff Kirsher |
Headers | show |
On Mon, Oct 31, 2016 at 1:12 PM, Guilherme G. Piccoli <gpiccoli@linux.vnet.ibm.com> wrote: > Whenever the igb driver detects the result of a read operation returns > a value composed only by F's (like 0xFFFFFFFF), it will detach the > net_device, clear the hw_addr pointer and warn to the user that adapter's > link is lost - those steps happen on igb_rd32(). > > In case a PCI error happens on Power architecture, there's a recovery > mechanism called EEH, that will reset the PCI slot and call driver's > handlers to reset the adapter and network functionality as well. > > We observed that once hw_addr is NULL after the error is detected on > igb_rd32(), it's never assigned back, so in the process of resetting > the network functionality we got a NULL pointer dereference in both > igb_configure_tx_ring() and igb_configure_rx_ring(). In order to avoid > such bug, we re-assign the hw_addr value in the beginning of the > function igb_reset(), in case the hw_addr is NULL when we reach that > path. > > Reported-by: Anthony H. Thai <ahthai@us.ibm.com> > Reported-by: Harsha Thyagaraja <hathyaga@in.ibm.com> > Signed-off-by: Guilherme G. Piccoli <gpiccoli@linux.vnet.ibm.com> > --- > drivers/net/ethernet/intel/igb/igb_main.c | 7 +++++++ > 1 file changed, 7 insertions(+) > > diff --git a/drivers/net/ethernet/intel/igb/igb_main.c b/drivers/net/ethernet/intel/igb/igb_main.c > index edc9a6a..c19119c 100644 > --- a/drivers/net/ethernet/intel/igb/igb_main.c > +++ b/drivers/net/ethernet/intel/igb/igb_main.c > @@ -1873,6 +1873,13 @@ void igb_reset(struct igb_adapter *adapter) > struct e1000_fc_info *fc = &hw->fc; > u32 pba, hwm; > > + /* In case of PCI error, adapter might have lost its HW > + * address; if we reached this point after an error scenario, > + * we should re-assign the hw_addr based on the saved io_addr. > + */ > + if (!hw->hw_addr) > + hw->hw_addr = adapter->io_addr; > + > /* Repartition Pba for greater than 9k mtu > * To take effect CTRL.RST is required. > */ It seems like this would have the potential to get noisy pretty quickly since every reset would retrigger this. It might make more sense to move this line into igb_io_slot_reset and igb_resume where the device would have gone through a reset of some sort and then had the state of the device restored and the device memory access was re-enabled via pci_enable_device_mem. Also there is no point in really doing the "if" since you should always be okay to overwrite hw->hw_addr with adapter->io_addr. Thanks. - Alex
On 11/09/2016 03:12 PM, Alexander Duyck wrote: > On Mon, Oct 31, 2016 at 1:12 PM, Guilherme G. Piccoli > <gpiccoli@linux.vnet.ibm.com> wrote: >> Whenever the igb driver detects the result of a read operation returns >> a value composed only by F's (like 0xFFFFFFFF), it will detach the >> net_device, clear the hw_addr pointer and warn to the user that adapter's >> link is lost - those steps happen on igb_rd32(). >> >> In case a PCI error happens on Power architecture, there's a recovery >> mechanism called EEH, that will reset the PCI slot and call driver's >> handlers to reset the adapter and network functionality as well. >> >> We observed that once hw_addr is NULL after the error is detected on >> igb_rd32(), it's never assigned back, so in the process of resetting >> the network functionality we got a NULL pointer dereference in both >> igb_configure_tx_ring() and igb_configure_rx_ring(). In order to avoid >> such bug, we re-assign the hw_addr value in the beginning of the >> function igb_reset(), in case the hw_addr is NULL when we reach that >> path. >> >> Reported-by: Anthony H. Thai <ahthai@us.ibm.com> >> Reported-by: Harsha Thyagaraja <hathyaga@in.ibm.com> >> Signed-off-by: Guilherme G. Piccoli <gpiccoli@linux.vnet.ibm.com> >> --- >> drivers/net/ethernet/intel/igb/igb_main.c | 7 +++++++ >> 1 file changed, 7 insertions(+) >> >> diff --git a/drivers/net/ethernet/intel/igb/igb_main.c b/drivers/net/ethernet/intel/igb/igb_main.c >> index edc9a6a..c19119c 100644 >> --- a/drivers/net/ethernet/intel/igb/igb_main.c >> +++ b/drivers/net/ethernet/intel/igb/igb_main.c >> @@ -1873,6 +1873,13 @@ void igb_reset(struct igb_adapter *adapter) >> struct e1000_fc_info *fc = &hw->fc; >> u32 pba, hwm; >> >> + /* In case of PCI error, adapter might have lost its HW >> + * address; if we reached this point after an error scenario, >> + * we should re-assign the hw_addr based on the saved io_addr. >> + */ >> + if (!hw->hw_addr) >> + hw->hw_addr = adapter->io_addr; >> + >> /* Repartition Pba for greater than 9k mtu >> * To take effect CTRL.RST is required. >> */ > > It seems like this would have the potential to get noisy pretty > quickly since every reset would retrigger this. > > It might make more sense to move this line into igb_io_slot_reset and > igb_resume where the device would have gone through a reset of some > sort and then had the state of the device restored and the device > memory access was re-enabled via pci_enable_device_mem. Also there is > no point in really doing the "if" since you should always be okay to > overwrite hw->hw_addr with adapter->io_addr. Thanks for the good suggestion Alexander, will send a V2. Cheers, Guilherme > > Thanks. > > - Alex >
diff --git a/drivers/net/ethernet/intel/igb/igb_main.c b/drivers/net/ethernet/intel/igb/igb_main.c index edc9a6a..c19119c 100644 --- a/drivers/net/ethernet/intel/igb/igb_main.c +++ b/drivers/net/ethernet/intel/igb/igb_main.c @@ -1873,6 +1873,13 @@ void igb_reset(struct igb_adapter *adapter) struct e1000_fc_info *fc = &hw->fc; u32 pba, hwm; + /* In case of PCI error, adapter might have lost its HW + * address; if we reached this point after an error scenario, + * we should re-assign the hw_addr based on the saved io_addr. + */ + if (!hw->hw_addr) + hw->hw_addr = adapter->io_addr; + /* Repartition Pba for greater than 9k mtu * To take effect CTRL.RST is required. */
Whenever the igb driver detects the result of a read operation returns a value composed only by F's (like 0xFFFFFFFF), it will detach the net_device, clear the hw_addr pointer and warn to the user that adapter's link is lost - those steps happen on igb_rd32(). In case a PCI error happens on Power architecture, there's a recovery mechanism called EEH, that will reset the PCI slot and call driver's handlers to reset the adapter and network functionality as well. We observed that once hw_addr is NULL after the error is detected on igb_rd32(), it's never assigned back, so in the process of resetting the network functionality we got a NULL pointer dereference in both igb_configure_tx_ring() and igb_configure_rx_ring(). In order to avoid such bug, we re-assign the hw_addr value in the beginning of the function igb_reset(), in case the hw_addr is NULL when we reach that path. Reported-by: Anthony H. Thai <ahthai@us.ibm.com> Reported-by: Harsha Thyagaraja <hathyaga@in.ibm.com> Signed-off-by: Guilherme G. Piccoli <gpiccoli@linux.vnet.ibm.com> --- drivers/net/ethernet/intel/igb/igb_main.c | 7 +++++++ 1 file changed, 7 insertions(+)