1 /* 2 * tracing clocks 3 * 4 * Copyright (C) 2009 Red Hat, Inc., Ingo Molnar <mingo@redhat.com> 5 * 6 * Implements 3 trace clock variants, with differing scalability/precision 7 * tradeoffs: 8 * 9 * - local: CPU-local trace clock 10 * - medium: scalable global clock with some jitter 11 * - global: globally monotonic, serialized clock 12 * 13 * Tracer plugins will chose a default from these clocks. 14 */ 15 #include <linux/spinlock.h> 16 #include <linux/irqflags.h> 17 #include <linux/hardirq.h> 18 #include <linux/module.h> 19 #include <linux/percpu.h> 20 #include <linux/sched.h> 21 #include <linux/ktime.h> 22 #include <linux/trace_clock.h> 23 24 /* 25 * trace_clock_local(): the simplest and least coherent tracing clock. 26 * 27 * Useful for tracing that does not cross to other CPUs nor 28 * does it go through idle events. 29 */ 30 u64 notrace trace_clock_local(void) 31 { 32 u64 clock; 33 34 /* 35 * sched_clock() is an architecture implemented, fast, scalable, 36 * lockless clock. It is not guaranteed to be coherent across 37 * CPUs, nor across CPU idle events. 38 */ 39 preempt_disable_notrace(); 40 clock = sched_clock(); 41 preempt_enable_notrace(); 42 43 return clock; 44 } 45 46 /* 47 * trace_clock(): 'between' trace clock. Not completely serialized, 48 * but not completely incorrect when crossing CPUs either. 49 * 50 * This is based on cpu_clock(), which will allow at most ~1 jiffy of 51 * jitter between CPUs. So it's a pretty scalable clock, but there 52 * can be offsets in the trace data. 53 */ 54 u64 notrace trace_clock(void) 55 { 56 return local_clock(); 57 } 58 59 60 /* 61 * trace_clock_global(): special globally coherent trace clock 62 * 63 * It has higher overhead than the other trace clocks but is still 64 * an order of magnitude faster than GTOD derived hardware clocks. 65 * 66 * Used by plugins that need globally coherent timestamps. 67 */ 68 69 /* keep prev_time and lock in the same cacheline. */ 70 static struct { 71 u64 prev_time; 72 arch_spinlock_t lock; 73 } trace_clock_struct ____cacheline_aligned_in_smp = 74 { 75 .lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED, 76 }; 77 78 u64 notrace trace_clock_global(void) 79 { 80 unsigned long flags; 81 int this_cpu; 82 u64 now; 83 84 local_irq_save(flags); 85 86 this_cpu = raw_smp_processor_id(); 87 now = cpu_clock(this_cpu); 88 /* 89 * If in an NMI context then dont risk lockups and return the 90 * cpu_clock() time: 91 */ 92 if (unlikely(in_nmi())) 93 goto out; 94 95 arch_spin_lock(&trace_clock_struct.lock); 96 97 /* 98 * TODO: if this happens often then maybe we should reset 99 * my_scd->clock to prev_time+1, to make sure 100 * we start ticking with the local clock from now on? 101 */ 102 if ((s64)(now - trace_clock_struct.prev_time) < 0) 103 now = trace_clock_struct.prev_time + 1; 104 105 trace_clock_struct.prev_time = now; 106 107 arch_spin_unlock(&trace_clock_struct.lock); 108 109 out: 110 local_irq_restore(flags); 111 112 return now; 113 } 114 115 static atomic64_t trace_counter; 116 117 /* 118 * trace_clock_counter(): simply an atomic counter. 119 * Use the trace_counter "counter" for cases where you do not care 120 * about timings, but are interested in strict ordering. 121 */ 122 u64 notrace trace_clock_counter(void) 123 { 124 return atomic64_add_return(1, &trace_counter); 125 } 126