]> git.openfabrics.org - ~shefty/rdma-dev.git/commitdiff
ixgbe: Add protection from VF invalid target DMA
authorGreg Rose <gregory.v.rose@intel.com>
Wed, 7 Sep 2011 05:59:35 +0000 (05:59 +0000)
committerJeff Kirsher <jeffrey.t.kirsher@intel.com>
Thu, 13 Oct 2011 05:45:24 +0000 (22:45 -0700)
It is possible for a VF to set an invalid target DMA address in its
Tx/Rx descriptor buffer pointers.  The workarounds in this patch
will guard against such an event and issue a VFLR to the VF in response.
The VFLR will shut down the VF until an administrator can take action
to investigate the event and correct the problem.

Signed-off-by: Greg Rose <gregory.v.rose@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
drivers/net/ethernet/intel/ixgbe/ixgbe.h
drivers/net/ethernet/intel/ixgbe/ixgbe_main.c

index 38940d72991ddeb3259595d65d2a85df8eaf39a6..c1f76aaf87744ea444a2eb4554f0cf1b431f537b 100644 (file)
 #define MAX_EMULATION_MAC_ADDRS         16
 #define IXGBE_MAX_PF_MACVLANS           15
 #define VMDQ_P(p)   ((p) + adapter->num_vfs)
+#define IXGBE_82599_VF_DEVICE_ID        0x10ED
+#define IXGBE_X540_VF_DEVICE_ID         0x1515
 
 struct vf_data_storage {
        unsigned char vf_mac_addresses[ETH_ALEN];
@@ -512,6 +514,8 @@ struct ixgbe_adapter {
        struct hlist_head fdir_filter_list;
        union ixgbe_atr_input fdir_mask;
        int fdir_filter_count;
+       u32 timer_event_accumulator;
+       u32 vferr_refcount;
 };
 
 struct ixgbe_fdir_filter {
index 1519a23421af18883770fe68779a740cc29d9074..b95c6e979832927cb3f25250c6da7f85be8ab199 100644 (file)
@@ -6112,6 +6112,51 @@ static void ixgbe_sfp_link_config_subtask(struct ixgbe_adapter *adapter)
        clear_bit(__IXGBE_IN_SFP_INIT, &adapter->state);
 }
 
+#ifdef CONFIG_PCI_IOV
+static void ixgbe_check_for_bad_vf(struct ixgbe_adapter *adapter)
+{
+       int vf;
+       struct ixgbe_hw *hw = &adapter->hw;
+       struct net_device *netdev = adapter->netdev;
+       u32 gpc;
+       u32 ciaa, ciad;
+
+       gpc = IXGBE_READ_REG(hw, IXGBE_TXDGPC);
+       if (gpc) /* If incrementing then no need for the check below */
+               return;
+       /*
+        * Check to see if a bad DMA write target from an errant or
+        * malicious VF has caused a PCIe error.  If so then we can
+        * issue a VFLR to the offending VF(s) and then resume without
+        * requesting a full slot reset.
+        */
+
+       for (vf = 0; vf < adapter->num_vfs; vf++) {
+               ciaa = (vf << 16) | 0x80000000;
+               /* 32 bit read so align, we really want status at offset 6 */
+               ciaa |= PCI_COMMAND;
+               IXGBE_WRITE_REG(hw, IXGBE_CIAA_82599, ciaa);
+               ciad = IXGBE_READ_REG(hw, IXGBE_CIAD_82599);
+               ciaa &= 0x7FFFFFFF;
+               /* disable debug mode asap after reading data */
+               IXGBE_WRITE_REG(hw, IXGBE_CIAA_82599, ciaa);
+               /* Get the upper 16 bits which will be the PCI status reg */
+               ciad >>= 16;
+               if (ciad & PCI_STATUS_REC_MASTER_ABORT) {
+                       netdev_err(netdev, "VF %d Hung DMA\n", vf);
+                       /* Issue VFLR */
+                       ciaa = (vf << 16) | 0x80000000;
+                       ciaa |= 0xA8;
+                       IXGBE_WRITE_REG(hw, IXGBE_CIAA_82599, ciaa);
+                       ciad = 0x00008000;  /* VFLR */
+                       IXGBE_WRITE_REG(hw, IXGBE_CIAD_82599, ciad);
+                       ciaa &= 0x7FFFFFFF;
+                       IXGBE_WRITE_REG(hw, IXGBE_CIAA_82599, ciaa);
+               }
+       }
+}
+
+#endif
 /**
  * ixgbe_service_timer - Timer Call-back
  * @data: pointer to adapter cast into an unsigned long
@@ -6120,17 +6165,49 @@ static void ixgbe_service_timer(unsigned long data)
 {
        struct ixgbe_adapter *adapter = (struct ixgbe_adapter *)data;
        unsigned long next_event_offset;
+       bool ready = true;
 
+#ifdef CONFIG_PCI_IOV
+       ready = false;
+
+       /*
+        * don't bother with SR-IOV VF DMA hang check if there are
+        * no VFs or the link is down
+        */
+       if (!adapter->num_vfs ||
+           (adapter->flags & IXGBE_FLAG_NEED_LINK_UPDATE)) {
+               ready = true;
+               goto normal_timer_service;
+       }
+
+       /* If we have VFs allocated then we must check for DMA hangs */
+       ixgbe_check_for_bad_vf(adapter);
+       next_event_offset = HZ / 50;
+       adapter->timer_event_accumulator++;
+
+       if (adapter->timer_event_accumulator >= 100) {
+               ready = true;
+               adapter->timer_event_accumulator = 0;
+       }
+
+       goto schedule_event;
+
+normal_timer_service:
+#endif
        /* poll faster when waiting for link */
        if (adapter->flags & IXGBE_FLAG_NEED_LINK_UPDATE)
                next_event_offset = HZ / 10;
        else
                next_event_offset = HZ * 2;
 
+#ifdef CONFIG_PCI_IOV
+schedule_event:
+#endif
        /* Reset the timer */
        mod_timer(&adapter->service_timer, next_event_offset + jiffies);
 
-       ixgbe_service_event_schedule(adapter);
+       if (ready)
+               ixgbe_service_event_schedule(adapter);
 }
 
 static void ixgbe_reset_subtask(struct ixgbe_adapter *adapter)
@@ -7717,6 +7794,91 @@ static pci_ers_result_t ixgbe_io_error_detected(struct pci_dev *pdev,
        struct ixgbe_adapter *adapter = pci_get_drvdata(pdev);
        struct net_device *netdev = adapter->netdev;
 
+#ifdef CONFIG_PCI_IOV
+       struct pci_dev *bdev, *vfdev;
+       u32 dw0, dw1, dw2, dw3;
+       int vf, pos;
+       u16 req_id, pf_func;
+
+       if (adapter->hw.mac.type == ixgbe_mac_82598EB ||
+           adapter->num_vfs == 0)
+               goto skip_bad_vf_detection;
+
+       bdev = pdev->bus->self;
+       while (bdev && (bdev->pcie_type != PCI_EXP_TYPE_ROOT_PORT))
+               bdev = bdev->bus->self;
+
+       if (!bdev)
+               goto skip_bad_vf_detection;
+
+       pos = pci_find_ext_capability(bdev, PCI_EXT_CAP_ID_ERR);
+       if (!pos)
+               goto skip_bad_vf_detection;
+
+       pci_read_config_dword(bdev, pos + PCI_ERR_HEADER_LOG, &dw0);
+       pci_read_config_dword(bdev, pos + PCI_ERR_HEADER_LOG + 4, &dw1);
+       pci_read_config_dword(bdev, pos + PCI_ERR_HEADER_LOG + 8, &dw2);
+       pci_read_config_dword(bdev, pos + PCI_ERR_HEADER_LOG + 12, &dw3);
+
+       req_id = dw1 >> 16;
+       /* On the 82599 if bit 7 of the requestor ID is set then it's a VF */
+       if (!(req_id & 0x0080))
+               goto skip_bad_vf_detection;
+
+       pf_func = req_id & 0x01;
+       if ((pf_func & 1) == (pdev->devfn & 1)) {
+               unsigned int device_id;
+
+               vf = (req_id & 0x7F) >> 1;
+               e_dev_err("VF %d has caused a PCIe error\n", vf);
+               e_dev_err("TLP: dw0: %8.8x\tdw1: %8.8x\tdw2: "
+                               "%8.8x\tdw3: %8.8x\n",
+               dw0, dw1, dw2, dw3);
+               switch (adapter->hw.mac.type) {
+               case ixgbe_mac_82599EB:
+                       device_id = IXGBE_82599_VF_DEVICE_ID;
+                       break;
+               case ixgbe_mac_X540:
+                       device_id = IXGBE_X540_VF_DEVICE_ID;
+                       break;
+               default:
+                       device_id = 0;
+                       break;
+               }
+
+               /* Find the pci device of the offending VF */
+               vfdev = pci_get_device(IXGBE_INTEL_VENDOR_ID, device_id, NULL);
+               while (vfdev) {
+                       if (vfdev->devfn == (req_id & 0xFF))
+                               break;
+                       vfdev = pci_get_device(IXGBE_INTEL_VENDOR_ID,
+                                              device_id, vfdev);
+               }
+               /*
+                * There's a slim chance the VF could have been hot plugged,
+                * so if it is no longer present we don't need to issue the
+                * VFLR.  Just clean up the AER in that case.
+                */
+               if (vfdev) {
+                       e_dev_err("Issuing VFLR to VF %d\n", vf);
+                       pci_write_config_dword(vfdev, 0xA8, 0x00008000);
+               }
+
+               pci_cleanup_aer_uncorrect_error_status(pdev);
+       }
+
+       /*
+        * Even though the error may have occurred on the other port
+        * we still need to increment the vf error reference count for
+        * both ports because the I/O resume function will be called
+        * for both of them.
+        */
+       adapter->vferr_refcount++;
+
+       return PCI_ERS_RESULT_RECOVERED;
+
+skip_bad_vf_detection:
+#endif /* CONFIG_PCI_IOV */
        netif_device_detach(netdev);
 
        if (state == pci_channel_io_perm_failure)
@@ -7779,6 +7941,14 @@ static void ixgbe_io_resume(struct pci_dev *pdev)
        struct ixgbe_adapter *adapter = pci_get_drvdata(pdev);
        struct net_device *netdev = adapter->netdev;
 
+#ifdef CONFIG_PCI_IOV
+       if (adapter->vferr_refcount) {
+               e_info(drv, "Resuming after VF err\n");
+               adapter->vferr_refcount--;
+               return;
+       }
+
+#endif
        if (netif_running(netdev))
                ixgbe_up(adapter);