OSDN Git Service

powerpc/eeh: Avoid event on passed PE
authorGavin Shan <gwshan@linux.vnet.ibm.com>
Tue, 10 Jun 2014 01:41:55 +0000 (11:41 +1000)
committerBenjamin Herrenschmidt <benh@kernel.crashing.org>
Tue, 5 Aug 2014 05:28:47 +0000 (15:28 +1000)
We must not handle EEH error on devices which are passed to somebody
else. Instead, we expect that the frozen device owner detects an EEH
error and recovers from it.

This avoids EEH error handling on passed through devices so the device
owner gets a chance to handle them.

Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Acked-by: Alexander Graf <agraf@suse.de>
Signed-off-by: Benjamin Herrenschmidt <benh@kernel.crashing.org>
arch/powerpc/include/asm/eeh.h
arch/powerpc/kernel/eeh.c
arch/powerpc/platforms/powernv/eeh-ioda.c

index fab7743..9537d83 100644 (file)
@@ -25,6 +25,7 @@
 #include <linux/list.h>
 #include <linux/string.h>
 #include <linux/time.h>
+#include <linux/atomic.h>
 
 struct pci_dev;
 struct pci_bus;
@@ -84,6 +85,7 @@ struct eeh_pe {
        int freeze_count;               /* Times of froze up            */
        struct timeval tstamp;          /* Time on first-time freeze    */
        int false_positives;            /* Times of reported #ff's      */
+       atomic_t pass_dev_cnt;          /* Count of passed through devs */
        struct eeh_pe *parent;          /* Parent PE                    */
        struct list_head child_list;    /* Link PE to the child list    */
        struct list_head edevs;         /* Link list of EEH devices     */
@@ -93,6 +95,11 @@ struct eeh_pe {
 #define eeh_pe_for_each_dev(pe, edev, tmp) \
                list_for_each_entry_safe(edev, tmp, &pe->edevs, list)
 
+static inline bool eeh_pe_passed(struct eeh_pe *pe)
+{
+       return pe ? !!atomic_read(&pe->pass_dev_cnt) : false;
+}
+
 /*
  * The struct is used to trace EEH state for the associated
  * PCI device node or PCI device. In future, it might
index 86e2570..c8f1a9d 100644 (file)
@@ -400,6 +400,14 @@ int eeh_dev_check_failure(struct eeh_dev *edev)
        if (ret > 0)
                return ret;
 
+       /*
+        * If the PE isn't owned by us, we shouldn't check the
+        * state. Instead, let the owner handle it if the PE has
+        * been frozen.
+        */
+       if (eeh_pe_passed(pe))
+               return 0;
+
        /* If we already have a pending isolation event for this
         * slot, we know it's bad already, we don't need to check.
         * Do this checking under a lock; as multiple PCI devices
index 8ad0c5b..f6abdb1 100644 (file)
@@ -812,7 +812,8 @@ static int ioda_eeh_next_error(struct eeh_pe **pe)
                                opal_pci_eeh_freeze_clear(phb->opal_id, frozen_pe_no,
                                        OPAL_EEH_ACTION_CLEAR_FREEZE_ALL);
                                ret = EEH_NEXT_ERR_NONE;
-                       } else if ((*pe)->state & EEH_PE_ISOLATED) {
+                       } else if ((*pe)->state & EEH_PE_ISOLATED ||
+                                  eeh_pe_passed(*pe)) {
                                ret = EEH_NEXT_ERR_NONE;
                        } else {
                                pr_err("EEH: Frozen PE#%x on PHB#%x detected\n",