wrapper/trace-clock.h

   1 #ifndef _LTTNG_TRACE_CLOCK_H
   2 #define _LTTNG_TRACE_CLOCK_H
   3
   4 /*
   5  * wrapper/trace-clock.h
   6  *
   7  * Contains LTTng trace clock mapping to LTTng trace clock or mainline monotonic
   8  * clock. This wrapper depends on CONFIG_HIGH_RES_TIMERS=y.
   9  *
  10  * Copyright (C) 2011-2012 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
  11  *
  12  * This library is free software; you can redistribute it and/or
  13  * modify it under the terms of the GNU Lesser General Public
  14  * License as published by the Free Software Foundation; only
  15  * version 2.1 of the License.
  16  *
  17  * This library is distributed in the hope that it will be useful,
  18  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  19  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  20  * Lesser General Public License for more details.
  21  *
  22  * You should have received a copy of the GNU Lesser General Public
  23  * License along with this library; if not, write to the Free Software
  24  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  25  */
  26
  27 #ifdef CONFIG_HAVE_TRACE_CLOCK
  28 #include <linux/trace-clock.h>
  29 #else /* CONFIG_HAVE_TRACE_CLOCK */
  30
  31 #include <linux/hardirq.h>
  32 #include <linux/ktime.h>
  33 #include <linux/time.h>
  34 #include <linux/hrtimer.h>
  35 #include <linux/percpu.h>
  36 #include <linux/version.h>
  37 #include <asm/local.h>
  38 #include <lttng-kernel-version.h>
  39 #include <lttng-clock.h>
  40 #include <wrapper/percpu-defs.h>
  41 #include <wrapper/random.h>
  42
  43 #if ((LTTNG_KERNEL_RANGE(3,10,0, 3,10,14) && !LTTNG_RHEL_KERNEL_RANGE(3,10,0,123,0,0, 3,10,14,0,0,0)) \
  44         || LTTNG_KERNEL_RANGE(3,11,0, 3,11,3))
  45 #error "Linux kernels 3.10 and 3.11 introduce a deadlock in the timekeeping subsystem. Fixed by commit 7bd36014460f793c19e7d6c94dab67b0afcfcb7f \"timekeeping: Fix HRTICK related deadlock from ntp lock changes\" in Linux."
  46 #endif
  47
  48 extern struct lttng_trace_clock *lttng_trace_clock;
  49
  50 /*
  51  * Upstream Linux commit 27727df240c7 ("Avoid taking lock in NMI path with
  52  * CONFIG_DEBUG_TIMEKEEPING") introduces a buggy ktime_get_mono_fast_ns().
  53  * This is fixed by patch "timekeeping: Fix __ktime_get_fast_ns() regression".
  54  */
  55 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) \
  56         && !LTTNG_KERNEL_RANGE(4,8,0, 4,8,1) \
  57         && !LTTNG_KERNEL_RANGE(4,7,4, 4,7,7) \
  58         && !LTTNG_KERNEL_RANGE(4,4,20, 4,4,24) \
  59         && !LTTNG_KERNEL_RANGE(4,1,32, 4,1,34))
  60
  61 DECLARE_PER_CPU(local_t, lttng_last_tsc);
  62
  63 #if (BITS_PER_LONG == 32)
  64 /*
  65  * Fixup "src_now" using the 32 LSB from "last". We need to handle overflow and
  66  * underflow of the 32nd bit. "last" can be above, below or equal to the 32 LSB
  67  * of "src_now".
  68  */
  69 static inline u64 trace_clock_fixup(u64 src_now, u32 last)
  70 {
  71         u64 now;
  72
  73         now = src_now & 0xFFFFFFFF00000000ULL;
  74         now |= (u64) last;
  75         /* Detect overflow or underflow between now and last. */
  76         if ((src_now & 0x80000000U) && !(last & 0x80000000U)) {
  77                 /*
  78                  * If 32nd bit transitions from 1 to 0, and we move forward in
  79                  * time from "now" to "last", then we have an overflow.
  80                  */
  81                 if (((s32) now - (s32) last) < 0)
  82                         now += 0x0000000100000000ULL;
  83         } else if (!(src_now & 0x80000000U) && (last & 0x80000000U)) {
  84                 /*
  85                  * If 32nd bit transitions from 0 to 1, and we move backward in
  86                  * time from "now" to "last", then we have an underflow.
  87                  */
  88                 if (((s32) now - (s32) last) > 0)
  89                         now -= 0x0000000100000000ULL;
  90         }
  91         return now;
  92 }
  93 #else /* #if (BITS_PER_LONG == 32) */
  94 /*
  95  * The fixup is pretty easy on 64-bit architectures: "last" is a 64-bit
  96  * value, so we can use last directly as current time.
  97  */
  98 static inline u64 trace_clock_fixup(u64 src_now, u64 last)
  99 {
 100         return last;
 101 }
 102 #endif /* #else #if (BITS_PER_LONG == 32) */
 103
 104 /*
 105  * Sometimes called with preemption enabled. Can be interrupted.
 106  */
 107 static inline u64 trace_clock_monotonic_wrapper(void)
 108 {
 109         u64 now;
 110         unsigned long last, result;
 111         local_t *last_tsc;
 112
 113         /* Use fast nmi-safe monotonic clock provided by the Linux kernel. */
 114         preempt_disable();
 115         last_tsc = lttng_this_cpu_ptr(&lttng_last_tsc);
 116         last = local_read(last_tsc);
 117         /*
 118          * Read "last" before "now". It is not strictly required, but it ensures
 119          * that an interrupt coming in won't artificially trigger a case where
 120          * "now" < "last". This kind of situation should only happen if the
 121          * mono_fast time source goes slightly backwards.
 122          */
 123         barrier();
 124         now = ktime_get_mono_fast_ns();
 125         if (((long) now - (long) last) < 0)
 126                 now = trace_clock_fixup(now, last);
 127         result = local_cmpxchg(last_tsc, last, (unsigned long) now);
 128         preempt_enable();
 129         if (result == last) {
 130                 /* Update done. */
 131                 return now;
 132         } else {
 133                 /*
 134                  * Update not done, due to concurrent update. We can use
 135                  * "result", since it has been sampled concurrently with our
 136                  * time read, so it should not be far from "now".
 137                  */
 138                 return trace_clock_fixup(now, result);
 139         }
 140 }
 141
 142 #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0)) */
 143 static inline u64 trace_clock_monotonic_wrapper(void)
 144 {
 145         ktime_t ktime;
 146
 147         /*
 148          * Refuse to trace from NMIs with this wrapper, because an NMI could
 149          * nest over the xtime write seqlock and deadlock.
 150          */
 151         if (in_nmi())
 152                 return (u64) -EIO;
 153
 154         ktime = ktime_get();
 155         return ktime_to_ns(ktime);
 156 }
 157 #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0)) */
 158
 159 static inline u64 trace_clock_read64_monotonic(void)
 160 {
 161         return (u64) trace_clock_monotonic_wrapper();
 162 }
 163
 164 static inline u64 trace_clock_freq_monotonic(void)
 165 {
 166         return (u64) NSEC_PER_SEC;
 167 }
 168
 169 static inline int trace_clock_uuid_monotonic(char *uuid)
 170 {
 171         return wrapper_get_bootid(uuid);
 172 }
 173
 174 static inline const char *trace_clock_name_monotonic(void)
 175 {
 176         return "monotonic";
 177 }
 178
 179 static inline const char *trace_clock_description_monotonic(void)
 180 {
 181         return "Monotonic Clock";
 182 }
 183
 184 #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0))
 185 static inline int get_trace_clock(void)
 186 {
 187         printk_once(KERN_WARNING "LTTng: Using mainline kernel monotonic fast clock, which is NMI-safe.\n");
 188         return 0;
 189 }
 190 #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0)) */
 191 static inline int get_trace_clock(void)
 192 {
 193         printk_once(KERN_WARNING "LTTng: Using mainline kernel monotonic clock. NMIs will not be traced.\n");
 194         return 0;
 195 }
 196 #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0)) */
 197
 198 static inline void put_trace_clock(void)
 199 {
 200 }
 201
 202 static inline u64 trace_clock_read64(void)
 203 {
 204         struct lttng_trace_clock *ltc = ACCESS_ONCE(lttng_trace_clock);
 205
 206         if (likely(!ltc)) {
 207                 return trace_clock_read64_monotonic();
 208         } else {
 209                 read_barrier_depends(); /* load ltc before content */
 210                 return ltc->read64();
 211         }
 212 }
 213
 214 static inline u64 trace_clock_freq(void)
 215 {
 216         struct lttng_trace_clock *ltc = ACCESS_ONCE(lttng_trace_clock);
 217
 218         if (!ltc) {
 219                 return trace_clock_freq_monotonic();
 220         } else {
 221                 read_barrier_depends(); /* load ltc before content */
 222                 return ltc->freq();
 223         }
 224 }
 225
 226 static inline int trace_clock_uuid(char *uuid)
 227 {
 228         struct lttng_trace_clock *ltc = ACCESS_ONCE(lttng_trace_clock);
 229
 230         read_barrier_depends(); /* load ltc before content */
 231         /* Use default UUID cb when NULL */
 232         if (!ltc || !ltc->uuid) {
 233                 return trace_clock_uuid_monotonic(uuid);
 234         } else {
 235                 return ltc->uuid(uuid);
 236         }
 237 }
 238
 239 static inline const char *trace_clock_name(void)
 240 {
 241         struct lttng_trace_clock *ltc = ACCESS_ONCE(lttng_trace_clock);
 242
 243         if (!ltc) {
 244                 return trace_clock_name_monotonic();
 245         } else {
 246                 read_barrier_depends(); /* load ltc before content */
 247                 return ltc->name();
 248         }
 249 }
 250
 251 static inline const char *trace_clock_description(void)
 252 {
 253         struct lttng_trace_clock *ltc = ACCESS_ONCE(lttng_trace_clock);
 254
 255         if (!ltc) {
 256                 return trace_clock_description_monotonic();
 257         } else {
 258                 read_barrier_depends(); /* load ltc before content */
 259                 return ltc->description();
 260         }
 261 }
 262
 263 #endif /* CONFIG_HAVE_TRACE_CLOCK */
 264
 265 #endif /* _LTTNG_TRACE_CLOCK_H */