ice: Fix removing driver while bare-metal VFs pass traffic
authorBrett Creeley <brett.creeley@intel.com>
Thu, 27 Feb 2020 18:14:55 +0000 (10:14 -0800)
committerJeff Kirsher <jeffrey.t.kirsher@intel.com>
Tue, 10 Mar 2020 20:10:02 +0000 (13:10 -0700)
Currently, if there are bare-metal VFs passing traffic and the ice
driver is removed, there is a possibility of VFs triggering a Tx timeout
right before iavf_remove(). This is causing iavf_close() to not be
called because there is a check in the beginning of iavf_remove() that
bails out early if (adapter->state < IAVF_DOWN_PENDING). This makes it
so some resources do not get cleaned up. Specifically, free_irq()
is never called for data interrupts, which results in the following line
of code to trigger:

pci_disable_msix()
free_msi_irqs()
...
BUG_ON(irq_has_action(entry->irq + i));
...

To prevent the Tx timeout from occurring on the VF during driver unload
for ice and the iavf there are a few changes that are needed.

[1] Don't disable all active VF Tx/Rx queues prior to calling
pci_disable_sriov.

[2] Call ice_free_vfs() before disabling the service task.

[3] Disable VF resets when the ice driver is being unloaded by setting
the pf->state flag __ICE_VF_RESETS_DISABLED.

Changing [1] and [2] allow each VF driver's remove flow to successfully
send VIRTCHNL requests, which includes queue disable. This prevents
unexpected Tx timeouts because the PF driver is no longer forcefully
disabling queues.

Due to [1] and [2] there is a possibility that the PF driver will get a
VFLR or reset request over VIRTCHNL from a VF during PF driver unload.
Prevent that by doing [3].

Signed-off-by: Brett Creeley <brett.creeley@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
drivers/net/ethernet/intel/ice/ice.h
drivers/net/ethernet/intel/ice/ice_main.c
drivers/net/ethernet/intel/ice/ice_virtchnl_pf.c

index aed3ff31e064d477ee53dec6ae8616496c752730..4d5b1fdb0688e1614f68f7abfa02c7b87d504184 100644 (file)
@@ -212,6 +212,7 @@ enum ice_state {
        __ICE_SERVICE_DIS,
        __ICE_OICR_INTR_DIS,            /* Global OICR interrupt disabled */
        __ICE_MDD_VF_PRINT_PENDING,     /* set when MDD event handle */
+       __ICE_VF_RESETS_DISABLED,       /* disable resets during ice_remove */
        __ICE_STATE_NBITS               /* must be last */
 };
 
index b94a668b5c282e852930f5878132e2840e6438cb..19290cc0b83c320ecf3f3d37e11d2ea5cf7a933e 100644 (file)
@@ -2054,8 +2054,16 @@ static irqreturn_t ice_misc_intr(int __always_unused irq, void *data)
                set_bit(__ICE_MDD_EVENT_PENDING, pf->state);
        }
        if (oicr & PFINT_OICR_VFLR_M) {
-               ena_mask &= ~PFINT_OICR_VFLR_M;
-               set_bit(__ICE_VFLR_EVENT_PENDING, pf->state);
+               /* disable any further VFLR event notifications */
+               if (test_bit(__ICE_VF_RESETS_DISABLED, pf->state)) {
+                       u32 reg = rd32(hw, PFINT_OICR_ENA);
+
+                       reg &= ~PFINT_OICR_VFLR_M;
+                       wr32(hw, PFINT_OICR_ENA, reg);
+               } else {
+                       ena_mask &= ~PFINT_OICR_VFLR_M;
+                       set_bit(__ICE_VFLR_EVENT_PENDING, pf->state);
+               }
        }
 
        if (oicr & PFINT_OICR_GRST_M) {
@@ -3380,11 +3388,14 @@ static void ice_remove(struct pci_dev *pdev)
                msleep(100);
        }
 
+       if (test_bit(ICE_FLAG_SRIOV_ENA, pf->flags)) {
+               set_bit(__ICE_VF_RESETS_DISABLED, pf->state);
+               ice_free_vfs(pf);
+       }
+
        set_bit(__ICE_DOWN, pf->state);
        ice_service_task_stop(pf);
 
-       if (test_bit(ICE_FLAG_SRIOV_ENA, pf->flags))
-               ice_free_vfs(pf);
        ice_vsi_release_all(pf);
        ice_free_irq_msix_misc(pf);
        ice_for_each_vsi(pf, i) {
index e0277b49439f122b1aba0a765485742450b5bf96..6ee7f8c9449a97779034bb0b72ee6fdb1342b8c7 100644 (file)
@@ -300,11 +300,6 @@ void ice_free_vfs(struct ice_pf *pf)
        while (test_and_set_bit(__ICE_VF_DIS, pf->state))
                usleep_range(1000, 2000);
 
-       /* Avoid wait time by stopping all VFs at the same time */
-       ice_for_each_vf(pf, i)
-               if (test_bit(ICE_VF_STATE_QS_ENA, pf->vf[i].vf_states))
-                       ice_dis_vf_qs(&pf->vf[i]);
-
        /* Disable IOV before freeing resources. This lets any VF drivers
         * running in the host get themselves cleaned up before we yank
         * the carpet out from underneath their feet.
@@ -314,6 +309,11 @@ void ice_free_vfs(struct ice_pf *pf)
        else
                dev_warn(dev, "VFs are assigned - not disabling SR-IOV\n");
 
+       /* Avoid wait time by stopping all VFs at the same time */
+       ice_for_each_vf(pf, i)
+               if (test_bit(ICE_VF_STATE_QS_ENA, pf->vf[i].vf_states))
+                       ice_dis_vf_qs(&pf->vf[i]);
+
        tmp = pf->num_alloc_vfs;
        pf->num_qps_per_vf = 0;
        pf->num_alloc_vfs = 0;
@@ -1155,7 +1155,8 @@ static bool ice_is_vf_disabled(struct ice_vf *vf)
  * @vf: pointer to the VF structure
  * @is_vflr: true if VFLR was issued, false if not
  *
- * Returns true if the VF is reset, false otherwise.
+ * Returns true if the VF is currently in reset, resets successfully, or resets
+ * are disabled and false otherwise.
  */
 bool ice_reset_vf(struct ice_vf *vf, bool is_vflr)
 {
@@ -1170,6 +1171,12 @@ bool ice_reset_vf(struct ice_vf *vf, bool is_vflr)
 
        dev = ice_pf_to_dev(pf);
 
+       if (test_bit(__ICE_VF_RESETS_DISABLED, pf->state)) {
+               dev_dbg(dev, "Trying to reset VF %d, but all VF resets are disabled\n",
+                       vf->vf_id);
+               return true;
+       }
+
        if (ice_is_vf_disabled(vf)) {
                dev_dbg(dev, "VF is already disabled, there is no need for resetting it, telling VM, all is fine %d\n",
                        vf->vf_id);