| 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 (LTTNG_KERNEL_RANGE(4,8,0, 4,8,1) \ |
| 56 | || LTTNG_KERNEL_RANGE(4,7,4, 4,7,7) \ |
| 57 | || LTTNG_KERNEL_RANGE(4,4,20, 4,4,24) \ |
| 58 | || LTTNG_KERNEL_RANGE(4,1,32, 4,1,34)) |
| 59 | #define LTTNG_CLOCK_NMI_SAFE_BROKEN |
| 60 | #endif |
| 61 | |
| 62 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0) \ |
| 63 | && !defined(LTTNG_CLOCK_NMI_SAFE_BROKEN)) |
| 64 | |
| 65 | DECLARE_PER_CPU(local_t, lttng_last_tsc); |
| 66 | |
| 67 | #if (BITS_PER_LONG == 32) |
| 68 | /* |
| 69 | * Fixup "src_now" using the 32 LSB from "last". We need to handle overflow and |
| 70 | * underflow of the 32nd bit. "last" can be above, below or equal to the 32 LSB |
| 71 | * of "src_now". |
| 72 | */ |
| 73 | static inline u64 trace_clock_fixup(u64 src_now, u32 last) |
| 74 | { |
| 75 | u64 now; |
| 76 | |
| 77 | now = src_now & 0xFFFFFFFF00000000ULL; |
| 78 | now |= (u64) last; |
| 79 | /* Detect overflow or underflow between now and last. */ |
| 80 | if ((src_now & 0x80000000U) && !(last & 0x80000000U)) { |
| 81 | /* |
| 82 | * If 32nd bit transitions from 1 to 0, and we move forward in |
| 83 | * time from "now" to "last", then we have an overflow. |
| 84 | */ |
| 85 | if (((s32) now - (s32) last) < 0) |
| 86 | now += 0x0000000100000000ULL; |
| 87 | } else if (!(src_now & 0x80000000U) && (last & 0x80000000U)) { |
| 88 | /* |
| 89 | * If 32nd bit transitions from 0 to 1, and we move backward in |
| 90 | * time from "now" to "last", then we have an underflow. |
| 91 | */ |
| 92 | if (((s32) now - (s32) last) > 0) |
| 93 | now -= 0x0000000100000000ULL; |
| 94 | } |
| 95 | return now; |
| 96 | } |
| 97 | #else /* #if (BITS_PER_LONG == 32) */ |
| 98 | /* |
| 99 | * The fixup is pretty easy on 64-bit architectures: "last" is a 64-bit |
| 100 | * value, so we can use last directly as current time. |
| 101 | */ |
| 102 | static inline u64 trace_clock_fixup(u64 src_now, u64 last) |
| 103 | { |
| 104 | return last; |
| 105 | } |
| 106 | #endif /* #else #if (BITS_PER_LONG == 32) */ |
| 107 | |
| 108 | /* |
| 109 | * Sometimes called with preemption enabled. Can be interrupted. |
| 110 | */ |
| 111 | static inline u64 trace_clock_monotonic_wrapper(void) |
| 112 | { |
| 113 | u64 now; |
| 114 | unsigned long last, result; |
| 115 | local_t *last_tsc; |
| 116 | |
| 117 | /* Use fast nmi-safe monotonic clock provided by the Linux kernel. */ |
| 118 | preempt_disable(); |
| 119 | last_tsc = lttng_this_cpu_ptr(<tng_last_tsc); |
| 120 | last = local_read(last_tsc); |
| 121 | /* |
| 122 | * Read "last" before "now". It is not strictly required, but it ensures |
| 123 | * that an interrupt coming in won't artificially trigger a case where |
| 124 | * "now" < "last". This kind of situation should only happen if the |
| 125 | * mono_fast time source goes slightly backwards. |
| 126 | */ |
| 127 | barrier(); |
| 128 | now = ktime_get_mono_fast_ns(); |
| 129 | if (((long) now - (long) last) < 0) |
| 130 | now = trace_clock_fixup(now, last); |
| 131 | result = local_cmpxchg(last_tsc, last, (unsigned long) now); |
| 132 | preempt_enable(); |
| 133 | if (result == last) { |
| 134 | /* Update done. */ |
| 135 | return now; |
| 136 | } else { |
| 137 | /* |
| 138 | * Update not done, due to concurrent update. We can use |
| 139 | * "result", since it has been sampled concurrently with our |
| 140 | * time read, so it should not be far from "now". |
| 141 | */ |
| 142 | return trace_clock_fixup(now, result); |
| 143 | } |
| 144 | } |
| 145 | |
| 146 | #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0)) */ |
| 147 | static inline u64 trace_clock_monotonic_wrapper(void) |
| 148 | { |
| 149 | ktime_t ktime; |
| 150 | |
| 151 | /* |
| 152 | * Refuse to trace from NMIs with this wrapper, because an NMI could |
| 153 | * nest over the xtime write seqlock and deadlock. |
| 154 | */ |
| 155 | if (in_nmi()) |
| 156 | return (u64) -EIO; |
| 157 | |
| 158 | ktime = ktime_get(); |
| 159 | return ktime_to_ns(ktime); |
| 160 | } |
| 161 | #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0)) */ |
| 162 | |
| 163 | static inline u64 trace_clock_read64_monotonic(void) |
| 164 | { |
| 165 | return (u64) trace_clock_monotonic_wrapper(); |
| 166 | } |
| 167 | |
| 168 | static inline u64 trace_clock_freq_monotonic(void) |
| 169 | { |
| 170 | return (u64) NSEC_PER_SEC; |
| 171 | } |
| 172 | |
| 173 | static inline int trace_clock_uuid_monotonic(char *uuid) |
| 174 | { |
| 175 | return wrapper_get_bootid(uuid); |
| 176 | } |
| 177 | |
| 178 | static inline const char *trace_clock_name_monotonic(void) |
| 179 | { |
| 180 | return "monotonic"; |
| 181 | } |
| 182 | |
| 183 | static inline const char *trace_clock_description_monotonic(void) |
| 184 | { |
| 185 | return "Monotonic Clock"; |
| 186 | } |
| 187 | |
| 188 | #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0)) |
| 189 | static inline int get_trace_clock(void) |
| 190 | { |
| 191 | printk_once(KERN_WARNING "LTTng: Using mainline kernel monotonic fast clock, which is NMI-safe.\n"); |
| 192 | return 0; |
| 193 | } |
| 194 | #else /* #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0)) */ |
| 195 | static inline int get_trace_clock(void) |
| 196 | { |
| 197 | printk_once(KERN_WARNING "LTTng: Using mainline kernel monotonic clock. NMIs will not be traced.\n"); |
| 198 | return 0; |
| 199 | } |
| 200 | #endif /* #else #if (LINUX_VERSION_CODE >= KERNEL_VERSION(3,17,0)) */ |
| 201 | |
| 202 | static inline void put_trace_clock(void) |
| 203 | { |
| 204 | } |
| 205 | |
| 206 | static inline u64 trace_clock_read64(void) |
| 207 | { |
| 208 | struct lttng_trace_clock *ltc = ACCESS_ONCE(lttng_trace_clock); |
| 209 | |
| 210 | if (likely(!ltc)) { |
| 211 | return trace_clock_read64_monotonic(); |
| 212 | } else { |
| 213 | read_barrier_depends(); /* load ltc before content */ |
| 214 | return ltc->read64(); |
| 215 | } |
| 216 | } |
| 217 | |
| 218 | static inline u64 trace_clock_freq(void) |
| 219 | { |
| 220 | struct lttng_trace_clock *ltc = ACCESS_ONCE(lttng_trace_clock); |
| 221 | |
| 222 | if (!ltc) { |
| 223 | return trace_clock_freq_monotonic(); |
| 224 | } else { |
| 225 | read_barrier_depends(); /* load ltc before content */ |
| 226 | return ltc->freq(); |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | static inline int trace_clock_uuid(char *uuid) |
| 231 | { |
| 232 | struct lttng_trace_clock *ltc = ACCESS_ONCE(lttng_trace_clock); |
| 233 | |
| 234 | read_barrier_depends(); /* load ltc before content */ |
| 235 | /* Use default UUID cb when NULL */ |
| 236 | if (!ltc || !ltc->uuid) { |
| 237 | return trace_clock_uuid_monotonic(uuid); |
| 238 | } else { |
| 239 | return ltc->uuid(uuid); |
| 240 | } |
| 241 | } |
| 242 | |
| 243 | static inline const char *trace_clock_name(void) |
| 244 | { |
| 245 | struct lttng_trace_clock *ltc = ACCESS_ONCE(lttng_trace_clock); |
| 246 | |
| 247 | if (!ltc) { |
| 248 | return trace_clock_name_monotonic(); |
| 249 | } else { |
| 250 | read_barrier_depends(); /* load ltc before content */ |
| 251 | return ltc->name(); |
| 252 | } |
| 253 | } |
| 254 | |
| 255 | static inline const char *trace_clock_description(void) |
| 256 | { |
| 257 | struct lttng_trace_clock *ltc = ACCESS_ONCE(lttng_trace_clock); |
| 258 | |
| 259 | if (!ltc) { |
| 260 | return trace_clock_description_monotonic(); |
| 261 | } else { |
| 262 | read_barrier_depends(); /* load ltc before content */ |
| 263 | return ltc->description(); |
| 264 | } |
| 265 | } |
| 266 | |
| 267 | #endif /* CONFIG_HAVE_TRACE_CLOCK */ |
| 268 | |
| 269 | #endif /* _LTTNG_TRACE_CLOCK_H */ |