1 /* 2 * kernel/sched/loadavg.c 3 * 4 * This file contains the magic bits required to compute the global loadavg 5 * figure. Its a silly number but people think its important. We go through 6 * great pains to make it work on big machines and tickless kernels. 7 */ 8 9 #include <linux/export.h> 10 11 #include "sched.h" 12 13 /* 14 * Global load-average calculations 15 * 16 * We take a distributed and async approach to calculating the global load-avg 17 * in order to minimize overhead. 18 * 19 * The global load average is an exponentially decaying average of nr_running + 20 * nr_uninterruptible. 21 * 22 * Once every LOAD_FREQ: 23 * 24 * nr_active = 0; 25 * for_each_possible_cpu(cpu) 26 * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible; 27 * 28 * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n) 29 * 30 * Due to a number of reasons the above turns in the mess below: 31 * 32 * - for_each_possible_cpu() is prohibitively expensive on machines with 33 * serious number of cpus, therefore we need to take a distributed approach 34 * to calculating nr_active. 35 * 36 * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0 37 * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) } 38 * 39 * So assuming nr_active := 0 when we start out -- true per definition, we 40 * can simply take per-cpu deltas and fold those into a global accumulate 41 * to obtain the same result. See calc_load_fold_active(). 42 * 43 * Furthermore, in order to avoid synchronizing all per-cpu delta folding 44 * across the machine, we assume 10 ticks is sufficient time for every 45 * cpu to have completed this task. 46 * 47 * This places an upper-bound on the IRQ-off latency of the machine. Then 48 * again, being late doesn't loose the delta, just wrecks the sample. 49 * 50 * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because 51 * this would add another cross-cpu cacheline miss and atomic operation 52 * to the wakeup path. Instead we increment on whatever cpu the task ran 53 * when it went into uninterruptible state and decrement on whatever cpu 54 * did the wakeup. This means that only the sum of nr_uninterruptible over 55 * all cpus yields the correct result. 56 * 57 * This covers the NO_HZ=n code, for extra head-aches, see the comment below. 58 */ 59 60 /* Variables and functions for calc_load */ 61 atomic_long_t calc_load_tasks; 62 unsigned long calc_load_update; 63 unsigned long avenrun[3]; 64 EXPORT_SYMBOL(avenrun); /* should be removed */ 65 66 /** 67 * get_avenrun - get the load average array 68 * @loads: pointer to dest load array 69 * @offset: offset to add 70 * @shift: shift count to shift the result left 71 * 72 * These values are estimates at best, so no need for locking. 73 */ 74 void get_avenrun(unsigned long *loads, unsigned long offset, int shift) 75 { 76 loads[0] = (avenrun[0] + offset) << shift; 77 loads[1] = (avenrun[1] + offset) << shift; 78 loads[2] = (avenrun[2] + offset) << shift; 79 } 80 81 long calc_load_fold_active(struct rq *this_rq) 82 { 83 long nr_active, delta = 0; 84 85 nr_active = this_rq->nr_running; 86 nr_active += (long)this_rq->nr_uninterruptible; 87 88 if (nr_active != this_rq->calc_load_active) { 89 delta = nr_active - this_rq->calc_load_active; 90 this_rq->calc_load_active = nr_active; 91 } 92 93 return delta; 94 } 95 96 /* 97 * a1 = a0 * e + a * (1 - e) 98 */ 99 static unsigned long 100 calc_load(unsigned long load, unsigned long exp, unsigned long active) 101 { 102 load *= exp; 103 load += active * (FIXED_1 - exp); 104 load += 1UL << (FSHIFT - 1); 105 return load >> FSHIFT; 106 } 107 108 #ifdef CONFIG_NO_HZ_COMMON 109 /* 110 * Handle NO_HZ for the global load-average. 111 * 112 * Since the above described distributed algorithm to compute the global 113 * load-average relies on per-cpu sampling from the tick, it is affected by 114 * NO_HZ. 115 * 116 * The basic idea is to fold the nr_active delta into a global idle-delta upon 117 * entering NO_HZ state such that we can include this as an 'extra' cpu delta 118 * when we read the global state. 119 * 120 * Obviously reality has to ruin such a delightfully simple scheme: 121 * 122 * - When we go NO_HZ idle during the window, we can negate our sample 123 * contribution, causing under-accounting. 124 * 125 * We avoid this by keeping two idle-delta counters and flipping them 126 * when the window starts, thus separating old and new NO_HZ load. 127 * 128 * The only trick is the slight shift in index flip for read vs write. 129 * 130 * 0s 5s 10s 15s 131 * +10 +10 +10 +10 132 * |-|-----------|-|-----------|-|-----------|-| 133 * r:0 0 1 1 0 0 1 1 0 134 * w:0 1 1 0 0 1 1 0 0 135 * 136 * This ensures we'll fold the old idle contribution in this window while 137 * accumlating the new one. 138 * 139 * - When we wake up from NO_HZ idle during the window, we push up our 140 * contribution, since we effectively move our sample point to a known 141 * busy state. 142 * 143 * This is solved by pushing the window forward, and thus skipping the 144 * sample, for this cpu (effectively using the idle-delta for this cpu which 145 * was in effect at the time the window opened). This also solves the issue 146 * of having to deal with a cpu having been in NOHZ idle for multiple 147 * LOAD_FREQ intervals. 148 * 149 * When making the ILB scale, we should try to pull this in as well. 150 */ 151 static atomic_long_t calc_load_idle[2]; 152 static int calc_load_idx; 153 154 static inline int calc_load_write_idx(void) 155 { 156 int idx = calc_load_idx; 157 158 /* 159 * See calc_global_nohz(), if we observe the new index, we also 160 * need to observe the new update time. 161 */ 162 smp_rmb(); 163 164 /* 165 * If the folding window started, make sure we start writing in the 166 * next idle-delta. 167 */ 168 if (!time_before(jiffies, calc_load_update)) 169 idx++; 170 171 return idx & 1; 172 } 173 174 static inline int calc_load_read_idx(void) 175 { 176 return calc_load_idx & 1; 177 } 178 179 void calc_load_enter_idle(void) 180 { 181 struct rq *this_rq = this_rq(); 182 long delta; 183 184 /* 185 * We're going into NOHZ mode, if there's any pending delta, fold it 186 * into the pending idle delta. 187 */ 188 delta = calc_load_fold_active(this_rq); 189 if (delta) { 190 int idx = calc_load_write_idx(); 191 192 atomic_long_add(delta, &calc_load_idle[idx]); 193 } 194 } 195 196 void calc_load_exit_idle(void) 197 { 198 struct rq *this_rq = this_rq(); 199 200 /* 201 * If we're still before the sample window, we're done. 202 */ 203 if (time_before(jiffies, this_rq->calc_load_update)) 204 return; 205 206 /* 207 * We woke inside or after the sample window, this means we're already 208 * accounted through the nohz accounting, so skip the entire deal and 209 * sync up for the next window. 210 */ 211 this_rq->calc_load_update = calc_load_update; 212 if (time_before(jiffies, this_rq->calc_load_update + 10)) 213 this_rq->calc_load_update += LOAD_FREQ; 214 } 215 216 static long calc_load_fold_idle(void) 217 { 218 int idx = calc_load_read_idx(); 219 long delta = 0; 220 221 if (atomic_long_read(&calc_load_idle[idx])) 222 delta = atomic_long_xchg(&calc_load_idle[idx], 0); 223 224 return delta; 225 } 226 227 /** 228 * fixed_power_int - compute: x^n, in O(log n) time 229 * 230 * @x: base of the power 231 * @frac_bits: fractional bits of @x 232 * @n: power to raise @x to. 233 * 234 * By exploiting the relation between the definition of the natural power 235 * function: x^n := x*x*...*x (x multiplied by itself for n times), and 236 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, 237 * (where: n_i \elem {0, 1}, the binary vector representing n), 238 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is 239 * of course trivially computable in O(log_2 n), the length of our binary 240 * vector. 241 */ 242 static unsigned long 243 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) 244 { 245 unsigned long result = 1UL << frac_bits; 246 247 if (n) { 248 for (;;) { 249 if (n & 1) { 250 result *= x; 251 result += 1UL << (frac_bits - 1); 252 result >>= frac_bits; 253 } 254 n >>= 1; 255 if (!n) 256 break; 257 x *= x; 258 x += 1UL << (frac_bits - 1); 259 x >>= frac_bits; 260 } 261 } 262 263 return result; 264 } 265 266 /* 267 * a1 = a0 * e + a * (1 - e) 268 * 269 * a2 = a1 * e + a * (1 - e) 270 * = (a0 * e + a * (1 - e)) * e + a * (1 - e) 271 * = a0 * e^2 + a * (1 - e) * (1 + e) 272 * 273 * a3 = a2 * e + a * (1 - e) 274 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) 275 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) 276 * 277 * ... 278 * 279 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] 280 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) 281 * = a0 * e^n + a * (1 - e^n) 282 * 283 * [1] application of the geometric series: 284 * 285 * n 1 - x^(n+1) 286 * S_n := \Sum x^i = ------------- 287 * i=0 1 - x 288 */ 289 static unsigned long 290 calc_load_n(unsigned long load, unsigned long exp, 291 unsigned long active, unsigned int n) 292 { 293 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); 294 } 295 296 /* 297 * NO_HZ can leave us missing all per-cpu ticks calling 298 * calc_load_account_active(), but since an idle CPU folds its delta into 299 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold 300 * in the pending idle delta if our idle period crossed a load cycle boundary. 301 * 302 * Once we've updated the global active value, we need to apply the exponential 303 * weights adjusted to the number of cycles missed. 304 */ 305 static void calc_global_nohz(void) 306 { 307 long delta, active, n; 308 309 if (!time_before(jiffies, calc_load_update + 10)) { 310 /* 311 * Catch-up, fold however many we are behind still 312 */ 313 delta = jiffies - calc_load_update - 10; 314 n = 1 + (delta / LOAD_FREQ); 315 316 active = atomic_long_read(&calc_load_tasks); 317 active = active > 0 ? active * FIXED_1 : 0; 318 319 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); 320 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); 321 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); 322 323 calc_load_update += n * LOAD_FREQ; 324 } 325 326 /* 327 * Flip the idle index... 328 * 329 * Make sure we first write the new time then flip the index, so that 330 * calc_load_write_idx() will see the new time when it reads the new 331 * index, this avoids a double flip messing things up. 332 */ 333 smp_wmb(); 334 calc_load_idx++; 335 } 336 #else /* !CONFIG_NO_HZ_COMMON */ 337 338 static inline long calc_load_fold_idle(void) { return 0; } 339 static inline void calc_global_nohz(void) { } 340 341 #endif /* CONFIG_NO_HZ_COMMON */ 342 343 /* 344 * calc_load - update the avenrun load estimates 10 ticks after the 345 * CPUs have updated calc_load_tasks. 346 * 347 * Called from the global timer code. 348 */ 349 void calc_global_load(unsigned long ticks) 350 { 351 long active, delta; 352 353 if (time_before(jiffies, calc_load_update + 10)) 354 return; 355 356 /* 357 * Fold the 'old' idle-delta to include all NO_HZ cpus. 358 */ 359 delta = calc_load_fold_idle(); 360 if (delta) 361 atomic_long_add(delta, &calc_load_tasks); 362 363 active = atomic_long_read(&calc_load_tasks); 364 active = active > 0 ? active * FIXED_1 : 0; 365 366 avenrun[0] = calc_load(avenrun[0], EXP_1, active); 367 avenrun[1] = calc_load(avenrun[1], EXP_5, active); 368 avenrun[2] = calc_load(avenrun[2], EXP_15, active); 369 370 calc_load_update += LOAD_FREQ; 371 372 /* 373 * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk. 374 */ 375 calc_global_nohz(); 376 } 377 378 /* 379 * Called from scheduler_tick() to periodically update this CPU's 380 * active count. 381 */ 382 void calc_global_load_tick(struct rq *this_rq) 383 { 384 long delta; 385 386 if (time_before(jiffies, this_rq->calc_load_update)) 387 return; 388 389 delta = calc_load_fold_active(this_rq); 390 if (delta) 391 atomic_long_add(delta, &calc_load_tasks); 392 393 this_rq->calc_load_update += LOAD_FREQ; 394 } 395