igb: release software locked semaphores on initialization
authorDidier Pallard <didier.pallard@6wind.com>
Tue, 8 Apr 2014 13:29:58 +0000 (15:29 +0200)
committerThomas Monjalon <thomas.monjalon@6wind.com>
Wed, 9 Apr 2014 16:30:02 +0000 (18:30 +0200)
It may happen that DPDK application gets killed while having
acquired locks on the ethernet hardware, causing these locks to
be never released. On next restart of the application, DPDK
skip those ports because it can not acquire the lock,
this may cause some ports (or even complete board if SMBI is locked)
to be inaccessible from DPDK application until reboot of the
hardware.

This patch release locks that are supposed to be locked due to
an improper exit of the application.

Signed-off-by: Didier Pallard <didier.pallard@6wind.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
lib/librte_pmd_e1000/igb_ethdev.c

index 184e7d6..673b4de 100644 (file)
@@ -319,6 +319,61 @@ igb_identify_hardware(struct rte_eth_dev *dev)
        /* need to check if it is a vf device below */
 }
 
+static int
+igb_reset_swfw_lock(struct e1000_hw *hw)
+{
+       int ret_val;
+
+       /*
+        * Do mac ops initialization manually here, since we will need
+        * some function pointers set by this call.
+        */
+       ret_val = e1000_init_mac_params(hw);
+       if (ret_val)
+               return ret_val;
+
+       /*
+        * SMBI lock should not fail in this early stage. If this is the case,
+        * it is due to an improper exit of the application.
+        * So force the release of the faulty lock.
+        */
+       if (e1000_get_hw_semaphore_generic(hw) < 0) {
+               DEBUGOUT("SMBI lock released");
+       }
+       e1000_put_hw_semaphore_generic(hw);
+
+       if (hw->mac.ops.acquire_swfw_sync != NULL) {
+               uint16_t mask;
+
+               /*
+                * Phy lock should not fail in this early stage. If this is the case,
+                * it is due to an improper exit of the application.
+                * So force the release of the faulty lock.
+                */
+               mask = E1000_SWFW_PHY0_SM << hw->bus.func;
+               if (hw->bus.func > E1000_FUNC_1)
+                       mask <<= 2;
+               if (hw->mac.ops.acquire_swfw_sync(hw, mask) < 0) {
+                       DEBUGOUT1("SWFW phy%d lock released", hw->bus.func);
+               }
+               hw->mac.ops.release_swfw_sync(hw, mask);
+
+               /*
+                * This one is more tricky since it is common to all ports; but
+                * swfw_sync retries last long enough (1s) to be almost sure that if
+                * lock can not be taken it is due to an improper lock of the
+                * semaphore.
+                */
+               mask = E1000_SWFW_EEP_SM;
+               if (hw->mac.ops.acquire_swfw_sync(hw, mask) < 0) {
+                       DEBUGOUT("SWFW common locks released");
+               }
+               hw->mac.ops.release_swfw_sync(hw, mask);
+       }
+
+       return E1000_SUCCESS;
+}
+
 static int
 eth_igb_dev_init(__attribute__((unused)) struct eth_driver *eth_drv,
                   struct rte_eth_dev *eth_dev)
@@ -348,13 +403,25 @@ eth_igb_dev_init(__attribute__((unused)) struct eth_driver *eth_drv,
        hw->hw_addr= (void *)pci_dev->mem_resource[0].addr;
 
        igb_identify_hardware(eth_dev);
-       if (e1000_setup_init_funcs(hw, TRUE) != E1000_SUCCESS) {
+       if (e1000_setup_init_funcs(hw, FALSE) != E1000_SUCCESS) {
                error = -EIO;
                goto err_late;
        }
 
        e1000_get_bus_info(hw);
 
+       /* Reset any pending lock */
+       if (igb_reset_swfw_lock(hw) != E1000_SUCCESS) {
+               error = -EIO;
+               goto err_late;
+       }
+
+       /* Finish initialization */
+       if (e1000_setup_init_funcs(hw, TRUE) != E1000_SUCCESS) {
+               error = -EIO;
+               goto err_late;
+       }
+
        hw->mac.autoneg = 1;
        hw->phy.autoneg_wait_to_complete = 0;
        hw->phy.autoneg_advertised = E1000_ALL_SPEED_DUPLEX;