--- zzzz-none-000/linux-3.10.107/arch/x86/kernel/tsc.c 2017-06-27 09:49:32.000000000 +0000 +++ scorpion-7490-727/linux-3.10.107/arch/x86/kernel/tsc.c 2021-02-04 17:41:59.000000000 +0000 @@ -11,6 +11,7 @@ #include #include #include +#include #include #include @@ -38,13 +39,257 @@ erroneous rdtsc usage on !cpu_has_tsc processors */ static int __read_mostly tsc_disabled = -1; +static DEFINE_STATIC_KEY_FALSE(__use_tsc); + int tsc_clocksource_reliable; + +/* + * Use a ring-buffer like data structure, where a writer advances the head by + * writing a new data entry and a reader advances the tail when it observes a + * new entry. + * + * Writers are made to wait on readers until there's space to write a new + * entry. + * + * This means that we can always use an {offset, mul} pair to compute a ns + * value that is 'roughly' in the right direction, even if we're writing a new + * {offset, mul} pair during the clock read. + * + * The down-side is that we can no longer guarantee strict monotonicity anymore + * (assuming the TSC was that to begin with), because while we compute the + * intersection point of the two clock slopes and make sure the time is + * continuous at the point of switching; we can no longer guarantee a reader is + * strictly before or after the switch point. + * + * It does mean a reader no longer needs to disable IRQs in order to avoid + * CPU-Freq updates messing with his times, and similarly an NMI reader will + * no longer run the risk of hitting half-written state. + */ + +struct cyc2ns { + struct cyc2ns_data data[2]; /* 0 + 2*24 = 48 */ + struct cyc2ns_data *head; /* 48 + 8 = 56 */ + struct cyc2ns_data *tail; /* 56 + 8 = 64 */ +}; /* exactly fits one cacheline */ + +static DEFINE_PER_CPU_ALIGNED(struct cyc2ns, cyc2ns); + +struct cyc2ns_data *cyc2ns_read_begin(void) +{ + struct cyc2ns_data *head; + + preempt_disable(); + + head = this_cpu_read(cyc2ns.head); + /* + * Ensure we observe the entry when we observe the pointer to it. + * matches the wmb from cyc2ns_write_end(). + */ + smp_read_barrier_depends(); + head->__count++; + barrier(); + + return head; +} + +void cyc2ns_read_end(struct cyc2ns_data *head) +{ + barrier(); + /* + * If we're the outer most nested read; update the tail pointer + * when we're done. This notifies possible pending writers + * that we've observed the head pointer and that the other + * entry is now free. + */ + if (!--head->__count) { + /* + * x86-TSO does not reorder writes with older reads; + * therefore once this write becomes visible to another + * cpu, we must be finished reading the cyc2ns_data. + * + * matches with cyc2ns_write_begin(). + */ + this_cpu_write(cyc2ns.tail, head); + } + preempt_enable(); +} + +/* + * Begin writing a new @data entry for @cpu. + * + * Assumes some sort of write side lock; currently 'provided' by the assumption + * that cpufreq will call its notifiers sequentially. + */ +static struct cyc2ns_data *cyc2ns_write_begin(int cpu) +{ + struct cyc2ns *c2n = &per_cpu(cyc2ns, cpu); + struct cyc2ns_data *data = c2n->data; + + if (data == c2n->head) + data++; + + /* XXX send an IPI to @cpu in order to guarantee a read? */ + + /* + * When we observe the tail write from cyc2ns_read_end(), + * the cpu must be done with that entry and its safe + * to start writing to it. + */ + while (c2n->tail == data) + cpu_relax(); + + return data; +} + +static void cyc2ns_write_end(int cpu, struct cyc2ns_data *data) +{ + struct cyc2ns *c2n = &per_cpu(cyc2ns, cpu); + + /* + * Ensure the @data writes are visible before we publish the + * entry. Matches the data-depencency in cyc2ns_read_begin(). + */ + smp_wmb(); + + ACCESS_ONCE(c2n->head) = data; +} + +/* + * Accelerators for sched_clock() + * convert from cycles(64bits) => nanoseconds (64bits) + * basic equation: + * ns = cycles / (freq / ns_per_sec) + * ns = cycles * (ns_per_sec / freq) + * ns = cycles * (10^9 / (cpu_khz * 10^3)) + * ns = cycles * (10^6 / cpu_khz) + * + * Then we use scaling math (suggested by george@mvista.com) to get: + * ns = cycles * (10^6 * SC / cpu_khz) / SC + * ns = cycles * cyc2ns_scale / SC + * + * And since SC is a constant power of two, we can convert the div + * into a shift. The larger SC is, the more accurate the conversion, but + * cyc2ns_scale needs to be a 32-bit value so that 32-bit multiplication + * (64-bit result) can be used. + * + * We can use khz divisor instead of mhz to keep a better precision. + * (mathieu.desnoyers@polymtl.ca) + * + * -johnstul@us.ibm.com "math is hard, lets go shopping!" + */ + +static void cyc2ns_data_init(struct cyc2ns_data *data) +{ + data->cyc2ns_mul = 0; + data->cyc2ns_shift = 0; + data->cyc2ns_offset = 0; + data->__count = 0; +} + +static void cyc2ns_init(int cpu) +{ + struct cyc2ns *c2n = &per_cpu(cyc2ns, cpu); + + cyc2ns_data_init(&c2n->data[0]); + cyc2ns_data_init(&c2n->data[1]); + + c2n->head = c2n->data; + c2n->tail = c2n->data; +} + +static inline unsigned long long cycles_2_ns(unsigned long long cyc) +{ + struct cyc2ns_data *data, *tail; + unsigned long long ns; + + /* + * See cyc2ns_read_*() for details; replicated in order to avoid + * an extra few instructions that came with the abstraction. + * Notable, it allows us to only do the __count and tail update + * dance when its actually needed. + */ + + preempt_disable_notrace(); + data = this_cpu_read(cyc2ns.head); + tail = this_cpu_read(cyc2ns.tail); + + if (likely(data == tail)) { + ns = data->cyc2ns_offset; + ns += mul_u64_u32_shr(cyc, data->cyc2ns_mul, data->cyc2ns_shift); + } else { + data->__count++; + + barrier(); + + ns = data->cyc2ns_offset; + ns += mul_u64_u32_shr(cyc, data->cyc2ns_mul, data->cyc2ns_shift); + + barrier(); + + if (!--data->__count) + this_cpu_write(cyc2ns.tail, data); + } + preempt_enable_notrace(); + + return ns; +} + +static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu) +{ + unsigned long long tsc_now, ns_now; + struct cyc2ns_data *data; + unsigned long flags; + + local_irq_save(flags); + sched_clock_idle_sleep_event(); + + if (!cpu_khz) + goto done; + + data = cyc2ns_write_begin(cpu); + + tsc_now = rdtsc(); + ns_now = cycles_2_ns(tsc_now); + + /* + * Compute a new multiplier as per the above comment and ensure our + * time function is continuous; see the comment near struct + * cyc2ns_data. + */ + clocks_calc_mult_shift(&data->cyc2ns_mul, &data->cyc2ns_shift, cpu_khz, + NSEC_PER_MSEC, 0); + + /* + * cyc2ns_shift is exported via arch_perf_update_userpage() where it is + * not expected to be greater than 31 due to the original published + * conversion algorithm shifting a 32-bit value (now specifies a 64-bit + * value) - refer perf_event_mmap_page documentation in perf_event.h. + */ + if (data->cyc2ns_shift == 32) { + data->cyc2ns_shift = 31; + data->cyc2ns_mul >>= 1; + } + + data->cyc2ns_offset = ns_now - + mul_u64_u32_shr(tsc_now, data->cyc2ns_mul, data->cyc2ns_shift); + + cyc2ns_write_end(cpu, data); + +done: + sched_clock_idle_wakeup_event(0); + local_irq_restore(flags); +} /* * Scheduler clock - returns current time in nanosec units. */ u64 native_sched_clock(void) { - u64 this_offset; + if (static_branch_likely(&__use_tsc)) { + u64 tsc_now = rdtsc(); + + /* return the value in ns */ + return cycles_2_ns(tsc_now); + } /* * Fall back to jiffies if there's no TSC available: @@ -54,16 +299,17 @@ * very important for it to be as fast as the platform * can achieve it. ) */ - if (unlikely(tsc_disabled)) { - /* No locking but a rare wrong value is not a big deal: */ - return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ); - } - /* read the Time Stamp Counter: */ - rdtscll(this_offset); + /* No locking but a rare wrong value is not a big deal: */ + return (jiffies_64 - INITIAL_JIFFIES) * (1000000000 / HZ); +} - /* return the value in ns */ - return __cycles_2_ns(this_offset); +/* + * Generate a sched_clock if you already have a TSC value. + */ +u64 native_sched_clock_from_tsc(u64 tsc) +{ + return cycles_2_ns(tsc); } /* We need to define a real function for sched_clock, to override the @@ -78,18 +324,18 @@ sched_clock(void) __attribute__((alias("native_sched_clock"))); #endif -unsigned long long native_read_tsc(void) -{ - return __native_read_tsc(); -} -EXPORT_SYMBOL(native_read_tsc); - int check_tsc_unstable(void) { return tsc_unstable; } EXPORT_SYMBOL_GPL(check_tsc_unstable); +int check_tsc_disabled(void) +{ + return tsc_disabled; +} +EXPORT_SYMBOL_GPL(check_tsc_disabled); + #ifdef CONFIG_X86_TSC int __init notsc_setup(char *str) { @@ -362,10 +608,19 @@ if (!pit_expect_msb(0xff-i, &delta, &d2)) break; + delta -= tsc; + + /* + * Extrapolate the error and fail fast if the error will + * never be below 500 ppm. + */ + if (i == 1 && + d1 + d2 >= (delta * MAX_QUICK_PIT_ITERATIONS) >> 11) + return 0; + /* * Iterate until the error is less than 500 ppm */ - delta -= tsc; if (d1+d2 >= delta >> 11) continue; @@ -414,6 +669,13 @@ unsigned long flags, latch, ms, fast_calibrate; int hpet = is_hpet_enabled(), i, loopmin; + /* Calibrate TSC using MSR for Intel Atom SoCs */ + local_irq_save(flags); + fast_calibrate = try_msr_calibrate_tsc(); + local_irq_restore(flags); + if (fast_calibrate) + return fast_calibrate; + local_irq_save(flags); fast_calibrate = quick_pit_calibrate(); local_irq_restore(flags); @@ -584,61 +846,11 @@ EXPORT_SYMBOL(recalibrate_cpu_khz); -/* Accelerators for sched_clock() - * convert from cycles(64bits) => nanoseconds (64bits) - * basic equation: - * ns = cycles / (freq / ns_per_sec) - * ns = cycles * (ns_per_sec / freq) - * ns = cycles * (10^9 / (cpu_khz * 10^3)) - * ns = cycles * (10^6 / cpu_khz) - * - * Then we use scaling math (suggested by george@mvista.com) to get: - * ns = cycles * (10^6 * SC / cpu_khz) / SC - * ns = cycles * cyc2ns_scale / SC - * - * And since SC is a constant power of two, we can convert the div - * into a shift. - * - * We can use khz divisor instead of mhz to keep a better precision, since - * cyc2ns_scale is limited to 10^6 * 2^10, which fits in 32 bits. - * (mathieu.desnoyers@polymtl.ca) - * - * -johnstul@us.ibm.com "math is hard, lets go shopping!" - */ - -DEFINE_PER_CPU(unsigned long, cyc2ns); -DEFINE_PER_CPU(unsigned long long, cyc2ns_offset); - -static void set_cyc2ns_scale(unsigned long cpu_khz, int cpu) -{ - unsigned long long tsc_now, ns_now, *offset; - unsigned long flags, *scale; - - local_irq_save(flags); - sched_clock_idle_sleep_event(); - - scale = &per_cpu(cyc2ns, cpu); - offset = &per_cpu(cyc2ns_offset, cpu); - - rdtscll(tsc_now); - ns_now = __cycles_2_ns(tsc_now); - - if (cpu_khz) { - *scale = ((NSEC_PER_MSEC << CYC2NS_SCALE_FACTOR) + - cpu_khz / 2) / cpu_khz; - *offset = ns_now - mult_frac(tsc_now, *scale, - (1UL << CYC2NS_SCALE_FACTOR)); - } - - sched_clock_idle_wakeup_event(0); - local_irq_restore(flags); -} - static unsigned long long cyc2ns_suspend; void tsc_save_sched_clock_state(void) { - if (!sched_clock_stable) + if (!sched_clock_stable()) return; cyc2ns_suspend = sched_clock(); @@ -658,16 +870,26 @@ unsigned long flags; int cpu; - if (!sched_clock_stable) + if (!sched_clock_stable()) return; local_irq_save(flags); - __this_cpu_write(cyc2ns_offset, 0); + /* + * We're comming out of suspend, there's no concurrency yet; don't + * bother being nice about the RCU stuff, just write to both + * data fields. + */ + + this_cpu_write(cyc2ns.data[0].cyc2ns_offset, 0); + this_cpu_write(cyc2ns.data[1].cyc2ns_offset, 0); + offset = cyc2ns_suspend - sched_clock(); - for_each_possible_cpu(cpu) - per_cpu(cyc2ns_offset, cpu) = offset; + for_each_possible_cpu(cpu) { + per_cpu(cyc2ns.data[0].cyc2ns_offset, cpu) = offset; + per_cpu(cyc2ns.data[1].cyc2ns_offset, cpu) = offset; + } local_irq_restore(flags); } @@ -710,16 +932,15 @@ tsc_khz_ref = tsc_khz; } if ((val == CPUFREQ_PRECHANGE && freq->old < freq->new) || - (val == CPUFREQ_POSTCHANGE && freq->old > freq->new) || - (val == CPUFREQ_RESUMECHANGE)) { + (val == CPUFREQ_POSTCHANGE && freq->old > freq->new)) { *lpj = cpufreq_scale(loops_per_jiffy_ref, ref_freq, freq->new); tsc_khz = cpufreq_scale(tsc_khz_ref, ref_freq, freq->new); if (!(freq->flags & CPUFREQ_CONST_LOOPS)) mark_tsc_unstable("cpufreq changes"); - } - set_cyc2ns_scale(tsc_khz, freq->cpu); + set_cyc2ns_scale(tsc_khz, freq->cpu); + } return 0; } @@ -748,7 +969,7 @@ static struct clocksource clocksource_tsc; /* - * We compare the TSC to the cycle_last value in the clocksource + * We used to compare the TSC to the cycle_last value in the clocksource * structure to avoid a nasty time-warp. This can be observed in a * very small window right after one CPU updated cycle_last under * xtime/vsyscall_gtod lock and the other CPU reads a TSC value which @@ -758,39 +979,34 @@ * due to the unsigned delta calculation of the time keeping core * code, which is necessary to support wrapping clocksources like pm * timer. + * + * This sanity check is now done in the core timekeeping code. + * checking the result of read_tsc() - cycle_last for being negative. + * That works because CLOCKSOURCE_MASK(64) does not mask out any bit. */ static cycle_t read_tsc(struct clocksource *cs) { - cycle_t ret = (cycle_t)get_cycles(); - - return ret >= clocksource_tsc.cycle_last ? - ret : clocksource_tsc.cycle_last; -} - -static void resume_tsc(struct clocksource *cs) -{ - if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC_S3)) - clocksource_tsc.cycle_last = 0; + return (cycle_t)rdtsc_ordered(); } +/* + * .mask MUST be CLOCKSOURCE_MASK(64). See comment above read_tsc() + */ static struct clocksource clocksource_tsc = { .name = "tsc", .rating = 300, .read = read_tsc, - .resume = resume_tsc, .mask = CLOCKSOURCE_MASK(64), .flags = CLOCK_SOURCE_IS_CONTINUOUS | CLOCK_SOURCE_MUST_VERIFY, -#ifdef CONFIG_X86_64 .archdata = { .vclock_mode = VCLOCK_TSC }, -#endif }; void mark_tsc_unstable(char *reason) { if (!tsc_unstable) { tsc_unstable = 1; - sched_clock_stable = 0; + clear_sched_clock_stable(); disable_sched_clock_irqtime(); pr_info("Marking TSC unstable due to %s\n", reason); /* Change only the rating, when not registered */ @@ -827,7 +1043,7 @@ * Make an educated guess if the TSC is trustworthy and synchronized * over all CPUs. */ -__cpuinit int unsynchronized_tsc(void) +int unsynchronized_tsc(void) { if (!cpu_has_tsc || tsc_unstable) return 1; @@ -995,14 +1211,18 @@ * speed as the bootup CPU. (cpufreq notifiers will fix this * up if their speed diverges) */ - for_each_possible_cpu(cpu) + for_each_possible_cpu(cpu) { + cyc2ns_init(cpu); set_cyc2ns_scale(cpu_khz, cpu); + } if (tsc_disabled > 0) return; /* now allow native_sched_clock() to use rdtsc */ + tsc_disabled = 0; + static_branch_enable(&__use_tsc); if (!no_sched_irq_time) enable_sched_clock_irqtime(); @@ -1026,7 +1246,7 @@ * been calibrated. This assumes that CONSTANT_TSC applies to all * cpus in the socket - this should be a safe assumption. */ -unsigned long __cpuinit calibrate_delay_is_known(void) +unsigned long calibrate_delay_is_known(void) { int i, cpu = smp_processor_id();