]> git.karo-electronics.de Git - mv-sheeva.git/blobdiff - arch/x86/kernel/cpu/perf_event_p4.c
Merge tag 'v2.6.38' of git://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux-2.6
[mv-sheeva.git] / arch / x86 / kernel / cpu / perf_event_p4.c
index 81400b93e69483e18bedf1c541753e2f915910cd..ff751a9f182b15bcb5ff94a00079c26a3d80229d 100644 (file)
@@ -682,7 +682,7 @@ static int p4_validate_raw_event(struct perf_event *event)
         * if an event is shared accross the logical threads
         * the user needs special permissions to be able to use it
         */
-       if (p4_event_bind_map[v].shared) {
+       if (p4_ht_active() && p4_event_bind_map[v].shared) {
                if (perf_paranoid_cpu() && !capable(CAP_SYS_ADMIN))
                        return -EACCES;
        }
@@ -727,7 +727,8 @@ static int p4_hw_config(struct perf_event *event)
                event->hw.config = p4_set_ht_bit(event->hw.config);
 
        if (event->attr.type == PERF_TYPE_RAW) {
-
+               struct p4_event_bind *bind;
+               unsigned int esel;
                /*
                 * Clear bits we reserve to be managed by kernel itself
                 * and never allowed from a user space
@@ -743,6 +744,13 @@ static int p4_hw_config(struct perf_event *event)
                 * bits since we keep additional info here (for cache events and etc)
                 */
                event->hw.config |= event->attr.config;
+               bind = p4_config_get_bind(event->attr.config);
+               if (!bind) {
+                       rc = -EINVAL;
+                       goto out;
+               }
+               esel = P4_OPCODE_ESEL(bind->opcode);
+               event->hw.config |= p4_config_pack_cccr(P4_CCCR_ESEL(esel));
        }
 
        rc = x86_setup_perfctr(event);
@@ -753,19 +761,26 @@ out:
 
 static inline int p4_pmu_clear_cccr_ovf(struct hw_perf_event *hwc)
 {
-       int overflow = 0;
-       u32 low, high;
-
-       rdmsr(hwc->config_base + hwc->idx, low, high);
+       u64 v;
 
-       /* we need to check high bit for unflagged overflows */
-       if ((low & P4_CCCR_OVF) || !(high & (1 << 31))) {
-               overflow = 1;
-               (void)checking_wrmsrl(hwc->config_base + hwc->idx,
-                       ((u64)low) & ~P4_CCCR_OVF);
+       /* an official way for overflow indication */
+       rdmsrl(hwc->config_base + hwc->idx, v);
+       if (v & P4_CCCR_OVF) {
+               wrmsrl(hwc->config_base + hwc->idx, v & ~P4_CCCR_OVF);
+               return 1;
        }
 
-       return overflow;
+       /*
+        * In some circumstances the overflow might issue an NMI but did
+        * not set P4_CCCR_OVF bit. Because a counter holds a negative value
+        * we simply check for high bit being set, if it's cleared it means
+        * the counter has reached zero value and continued counting before
+        * real NMI signal was received:
+        */
+       if (!(v & ARCH_P4_UNFLAGGED_BIT))
+               return 1;
+
+       return 0;
 }
 
 static void p4_pmu_disable_pebs(void)
@@ -1152,9 +1167,9 @@ static __initconst const struct x86_pmu p4_pmu = {
         */
        .num_counters           = ARCH_P4_MAX_CCCR,
        .apic                   = 1,
-       .cntval_bits            = 40,
-       .cntval_mask            = (1ULL << 40) - 1,
-       .max_period             = (1ULL << 39) - 1,
+       .cntval_bits            = ARCH_P4_CNTRVAL_BITS,
+       .cntval_mask            = ARCH_P4_CNTRVAL_MASK,
+       .max_period             = (1ULL << (ARCH_P4_CNTRVAL_BITS - 1)) - 1,
        .hw_config              = p4_hw_config,
        .schedule_events        = p4_pmu_schedule_events,
        /*