x86, perfcounters: add support for fixed-function pmcs
authorIngo Molnar <mingo@elte.hu>
Mon, 22 Dec 2008 10:10:42 +0000 (11:10 +0100)
committerIngo Molnar <mingo@elte.hu>
Tue, 23 Dec 2008 11:45:25 +0000 (12:45 +0100)
Impact: extend performance counter support on x86 Intel CPUs

Modern Intel CPUs have 3 "fixed-function" performance counters, which
count these hardware events:

    Instr_Retired.Any
    CPU_CLK_Unhalted.Core
    CPU_CLK_Unhalted.Ref

Add support for them to the performance counters subsystem.

Their use is transparent to user-space: the counter scheduler is
extended to automatically recognize the cases where a fixed-function
PMC can be utilized instead of a generic PMC. In such cases the
generic PMC is kept available for more counters.

The above fixed-function events map to these generic counter hw events:

        PERF_COUNT_INSTRUCTIONS
        PERF_COUNT_CPU_CYCLES
        PERF_COUNT_BUS_CYCLES

(The 'bus' cycles are in reality often CPU-ish cycles, just with a fixed
 frequency.)

Signed-off-by: Ingo Molnar <mingo@elte.hu>
arch/x86/include/asm/perf_counter.h
arch/x86/kernel/cpu/perf_counter.c

index 13745deb16c8bd733613a8572b9d2177f15e0890..2e08ed736647eb055613fa244d1ad62257211cff 100644 (file)
 #define ARCH_PERFMON_EVENTSEL_OS                         (1 << 17)
 #define ARCH_PERFMON_EVENTSEL_USR                        (1 << 16)
 
+/*
+ * Includes eventsel and unit mask as well:
+ */
+#define ARCH_PERFMON_EVENT_MASK                                    0xffff
+
 #define ARCH_PERFMON_UNHALTED_CORE_CYCLES_SEL                0x3c
 #define ARCH_PERFMON_UNHALTED_CORE_CYCLES_UMASK                (0x00 << 8)
 #define ARCH_PERFMON_UNHALTED_CORE_CYCLES_INDEX                 0
@@ -69,12 +74,15 @@ union cpuid10_edx {
 
 /* Instr_Retired.Any: */
 #define MSR_ARCH_PERFMON_FIXED_CTR0                    0x309
+#define X86_PMC_IDX_FIXED_INSTRUCTIONS                 (X86_PMC_IDX_FIXED + 0)
 
 /* CPU_CLK_Unhalted.Core: */
 #define MSR_ARCH_PERFMON_FIXED_CTR1                    0x30a
+#define X86_PMC_IDX_FIXED_CPU_CYCLES                   (X86_PMC_IDX_FIXED + 1)
 
 /* CPU_CLK_Unhalted.Ref: */
 #define MSR_ARCH_PERFMON_FIXED_CTR2                    0x30b
+#define X86_PMC_IDX_FIXED_BUS_CYCLES                   (X86_PMC_IDX_FIXED + 2)
 
 #ifdef CONFIG_PERF_COUNTERS
 extern void init_hw_perf_counters(void);
index 86b2fdd344a62a765b07cf3d23a01d0a47c7177d..da46eca12543284264661b89d22b77bdcfa07c36 100644 (file)
@@ -26,6 +26,7 @@ static bool perf_counters_initialized __read_mostly;
  */
 static int nr_counters_generic __read_mostly;
 static u64 perf_counter_mask __read_mostly;
+static u64 counter_value_mask __read_mostly;
 
 static int nr_counters_fixed __read_mostly;
 
@@ -120,9 +121,6 @@ static int __hw_perf_counter_init(struct perf_counter *counter)
                        hwc->nmi = 1;
        }
 
-       hwc->config_base        = MSR_ARCH_PERFMON_EVENTSEL0;
-       hwc->counter_base       = MSR_ARCH_PERFMON_PERFCTR0;
-
        hwc->irq_period         = hw_event->irq_period;
        /*
         * Intel PMCs cannot be accessed sanely above 32 bit width,
@@ -183,16 +181,34 @@ void hw_perf_restore(u64 ctrl)
 }
 EXPORT_SYMBOL_GPL(hw_perf_restore);
 
+static inline void
+__pmc_fixed_disable(struct perf_counter *counter,
+                   struct hw_perf_counter *hwc, unsigned int __idx)
+{
+       int idx = __idx - X86_PMC_IDX_FIXED;
+       u64 ctrl_val, mask;
+       int err;
+
+       mask = 0xfULL << (idx * 4);
+
+       rdmsrl(hwc->config_base, ctrl_val);
+       ctrl_val &= ~mask;
+       err = checking_wrmsrl(hwc->config_base, ctrl_val);
+}
+
 static inline void
 __pmc_generic_disable(struct perf_counter *counter,
                           struct hw_perf_counter *hwc, unsigned int idx)
 {
        int err;
 
+       if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL))
+               return __pmc_fixed_disable(counter, hwc, idx);
+
        err = wrmsr_safe(hwc->config_base + idx, hwc->config, 0);
 }
 
-static DEFINE_PER_CPU(u64, prev_left[X86_PMC_MAX_GENERIC]);
+static DEFINE_PER_CPU(u64, prev_left[X86_PMC_IDX_MAX]);
 
 /*
  * Set the next IRQ period, based on the hwc->period_left value.
@@ -202,8 +218,9 @@ static void
 __hw_perf_counter_set_period(struct perf_counter *counter,
                             struct hw_perf_counter *hwc, int idx)
 {
-       s32 left = atomic64_read(&hwc->period_left);
+       s64 left = atomic64_read(&hwc->period_left);
        s32 period = hwc->irq_period;
+       int err;
 
        /*
         * If we are way outside a reasoable range then just skip forward:
@@ -224,21 +241,64 @@ __hw_perf_counter_set_period(struct perf_counter *counter,
         * The hw counter starts counting from this counter offset,
         * mark it to be able to extra future deltas:
         */
-       atomic64_set(&hwc->prev_count, (u64)(s64)-left);
+       atomic64_set(&hwc->prev_count, (u64)-left);
 
-       wrmsr(hwc->counter_base + idx, -left, 0);
+       err = checking_wrmsrl(hwc->counter_base + idx,
+                            (u64)(-left) & counter_value_mask);
+}
+
+static inline void
+__pmc_fixed_enable(struct perf_counter *counter,
+                  struct hw_perf_counter *hwc, unsigned int __idx)
+{
+       int idx = __idx - X86_PMC_IDX_FIXED;
+       u64 ctrl_val, bits, mask;
+       int err;
+
+       /*
+        * Enable IRQ generation (0x8) and ring-3 counting (0x2),
+        * and enable ring-0 counting if allowed:
+        */
+       bits = 0x8ULL | 0x2ULL;
+       if (hwc->config & ARCH_PERFMON_EVENTSEL_OS)
+               bits |= 0x1;
+       bits <<= (idx * 4);
+       mask = 0xfULL << (idx * 4);
+
+       rdmsrl(hwc->config_base, ctrl_val);
+       ctrl_val &= ~mask;
+       ctrl_val |= bits;
+       err = checking_wrmsrl(hwc->config_base, ctrl_val);
 }
 
 static void
 __pmc_generic_enable(struct perf_counter *counter,
                          struct hw_perf_counter *hwc, int idx)
 {
+       if (unlikely(hwc->config_base == MSR_ARCH_PERFMON_FIXED_CTR_CTRL))
+               return __pmc_fixed_enable(counter, hwc, idx);
+
        wrmsr(hwc->config_base + idx,
              hwc->config | ARCH_PERFMON_EVENTSEL0_ENABLE, 0);
 }
 
-static int fixed_mode_idx(struct hw_perf_counter *hwc)
+static int
+fixed_mode_idx(struct perf_counter *counter, struct hw_perf_counter *hwc)
 {
+       unsigned int event;
+
+       if (unlikely(hwc->nmi))
+               return -1;
+
+       event = hwc->config & ARCH_PERFMON_EVENT_MASK;
+
+       if (unlikely(event == intel_perfmon_event_map[PERF_COUNT_INSTRUCTIONS]))
+               return X86_PMC_IDX_FIXED_INSTRUCTIONS;
+       if (unlikely(event == intel_perfmon_event_map[PERF_COUNT_CPU_CYCLES]))
+               return X86_PMC_IDX_FIXED_CPU_CYCLES;
+       if (unlikely(event == intel_perfmon_event_map[PERF_COUNT_BUS_CYCLES]))
+               return X86_PMC_IDX_FIXED_BUS_CYCLES;
+
        return -1;
 }
 
@@ -249,16 +309,39 @@ static int pmc_generic_enable(struct perf_counter *counter)
 {
        struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
        struct hw_perf_counter *hwc = &counter->hw;
-       int idx = hwc->idx;
+       int idx;
 
-       /* Try to get the previous counter again */
-       if (test_and_set_bit(idx, cpuc->used)) {
-               idx = find_first_zero_bit(cpuc->used, nr_counters_generic);
-               if (idx == nr_counters_generic)
-                       return -EAGAIN;
+       idx = fixed_mode_idx(counter, hwc);
+       if (idx >= 0) {
+               /*
+                * Try to get the fixed counter, if that is already taken
+                * then try to get a generic counter:
+                */
+               if (test_and_set_bit(idx, cpuc->used))
+                       goto try_generic;
 
-               set_bit(idx, cpuc->used);
+               hwc->config_base = MSR_ARCH_PERFMON_FIXED_CTR_CTRL;
+               /*
+                * We set it so that counter_base + idx in wrmsr/rdmsr maps to
+                * MSR_ARCH_PERFMON_FIXED_CTR0 ... CTR2:
+                */
+               hwc->counter_base =
+                       MSR_ARCH_PERFMON_FIXED_CTR0 - X86_PMC_IDX_FIXED;
                hwc->idx = idx;
+       } else {
+               idx = hwc->idx;
+               /* Try to get the previous generic counter again */
+               if (test_and_set_bit(idx, cpuc->used)) {
+try_generic:
+                       idx = find_first_zero_bit(cpuc->used, nr_counters_generic);
+                       if (idx == nr_counters_generic)
+                               return -EAGAIN;
+
+                       set_bit(idx, cpuc->used);
+                       hwc->idx = idx;
+               }
+               hwc->config_base  = MSR_ARCH_PERFMON_EVENTSEL0;
+               hwc->counter_base = MSR_ARCH_PERFMON_PERFCTR0;
        }
 
        perf_counters_lapic_init(hwc->nmi);
@@ -266,6 +349,10 @@ static int pmc_generic_enable(struct perf_counter *counter)
        __pmc_generic_disable(counter, hwc, idx);
 
        cpuc->counters[idx] = counter;
+       /*
+        * Make it visible before enabling the hw:
+        */
+       smp_wmb();
 
        __hw_perf_counter_set_period(counter, hwc, idx);
        __pmc_generic_enable(counter, hwc, idx);
@@ -275,7 +362,7 @@ static int pmc_generic_enable(struct perf_counter *counter)
 
 void perf_counter_print_debug(void)
 {
-       u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left;
+       u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
        struct cpu_hw_counters *cpuc;
        int cpu, idx;
 
@@ -290,11 +377,13 @@ void perf_counter_print_debug(void)
        rdmsrl(MSR_CORE_PERF_GLOBAL_CTRL, ctrl);
        rdmsrl(MSR_CORE_PERF_GLOBAL_STATUS, status);
        rdmsrl(MSR_CORE_PERF_GLOBAL_OVF_CTRL, overflow);
+       rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR_CTRL, fixed);
 
        printk(KERN_INFO "\n");
        printk(KERN_INFO "CPU#%d: ctrl:       %016llx\n", cpu, ctrl);
        printk(KERN_INFO "CPU#%d: status:     %016llx\n", cpu, status);
        printk(KERN_INFO "CPU#%d: overflow:   %016llx\n", cpu, overflow);
+       printk(KERN_INFO "CPU#%d: fixed:      %016llx\n", cpu, fixed);
        printk(KERN_INFO "CPU#%d: used:       %016llx\n", cpu, *(u64 *)cpuc->used);
 
        for (idx = 0; idx < nr_counters_generic; idx++) {
@@ -303,13 +392,19 @@ void perf_counter_print_debug(void)
 
                prev_left = per_cpu(prev_left[idx], cpu);
 
-               printk(KERN_INFO "CPU#%d: PMC%d ctrl:  %016llx\n",
+               printk(KERN_INFO "CPU#%d:   gen-PMC%d ctrl:  %016llx\n",
                        cpu, idx, pmc_ctrl);
-               printk(KERN_INFO "CPU#%d: PMC%d count: %016llx\n",
+               printk(KERN_INFO "CPU#%d:   gen-PMC%d count: %016llx\n",
                        cpu, idx, pmc_count);
-               printk(KERN_INFO "CPU#%d: PMC%d left:  %016llx\n",
+               printk(KERN_INFO "CPU#%d:   gen-PMC%d left:  %016llx\n",
                        cpu, idx, prev_left);
        }
+       for (idx = 0; idx < nr_counters_fixed; idx++) {
+               rdmsrl(MSR_ARCH_PERFMON_FIXED_CTR0 + idx, pmc_count);
+
+               printk(KERN_INFO "CPU#%d: fixed-PMC%d count: %016llx\n",
+                       cpu, idx, pmc_count);
+       }
        local_irq_enable();
 }
 
@@ -323,6 +418,11 @@ static void pmc_generic_disable(struct perf_counter *counter)
 
        clear_bit(idx, cpuc->used);
        cpuc->counters[idx] = NULL;
+       /*
+        * Make sure the cleared pointer becomes visible before we
+        * (potentially) free the counter:
+        */
+       smp_wmb();
 
        /*
         * Drain the remaining delta count out of a counter
@@ -353,14 +453,11 @@ static void perf_save_and_restart(struct perf_counter *counter)
 {
        struct hw_perf_counter *hwc = &counter->hw;
        int idx = hwc->idx;
-       u64 pmc_ctrl;
-
-       rdmsrl(MSR_ARCH_PERFMON_EVENTSEL0 + idx, pmc_ctrl);
 
        x86_perf_counter_update(counter, hwc, idx);
        __hw_perf_counter_set_period(counter, hwc, idx);
 
-       if (pmc_ctrl & ARCH_PERFMON_EVENTSEL0_ENABLE)
+       if (counter->state == PERF_COUNTER_STATE_ACTIVE)
                __pmc_generic_enable(counter, hwc, idx);
 }
 
@@ -373,6 +470,7 @@ perf_handle_group(struct perf_counter *sibling, u64 *status, u64 *overflown)
         * Store sibling timestamps (if any):
         */
        list_for_each_entry(counter, &group_leader->sibling_list, list_entry) {
+
                x86_perf_counter_update(counter, &counter->hw, counter->hw.idx);
                perf_store_irq_data(sibling, counter->hw_event.type);
                perf_store_irq_data(sibling, atomic64_read(&counter->count));
@@ -403,7 +501,7 @@ static void __smp_perf_counter_interrupt(struct pt_regs *regs, int nmi)
 
 again:
        ack = status;
-       for_each_bit(bit, (unsigned long *) &status, nr_counters_generic) {
+       for_each_bit(bit, (unsigned long *)&status, X86_PMC_IDX_MAX) {
                struct perf_counter *counter = cpuc->counters[bit];
 
                clear_bit(bit, (unsigned long *) &status);
@@ -561,6 +659,9 @@ void __init init_hw_perf_counters(void)
        perf_max_counters = nr_counters_generic;
 
        printk(KERN_INFO "... bit width:       %d\n", eax.split.bit_width);
+       counter_value_mask = (1ULL << eax.split.bit_width) - 1;
+       printk(KERN_INFO "... value mask:      %016Lx\n", counter_value_mask);
+
        printk(KERN_INFO "... mask length:     %d\n", eax.split.mask_length);
 
        nr_counters_fixed = edx.split.num_counters_fixed;