]>
Commit | Line | Data |
---|---|---|
45ceebf7 | 1 | /* |
3289bdb4 | 2 | * kernel/sched/loadavg.c |
45ceebf7 | 3 | * |
3289bdb4 PZ |
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. | |
45ceebf7 PG |
7 | */ |
8 | ||
9 | #include <linux/export.h> | |
4f17722c | 10 | #include <linux/sched/loadavg.h> |
45ceebf7 PG |
11 | |
12 | #include "sched.h" | |
13 | ||
45ceebf7 PG |
14 | /* |
15 | * Global load-average calculations | |
16 | * | |
17 | * We take a distributed and async approach to calculating the global load-avg | |
18 | * in order to minimize overhead. | |
19 | * | |
20 | * The global load average is an exponentially decaying average of nr_running + | |
21 | * nr_uninterruptible. | |
22 | * | |
23 | * Once every LOAD_FREQ: | |
24 | * | |
25 | * nr_active = 0; | |
26 | * for_each_possible_cpu(cpu) | |
27 | * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible; | |
28 | * | |
29 | * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n) | |
30 | * | |
31 | * Due to a number of reasons the above turns in the mess below: | |
32 | * | |
33 | * - for_each_possible_cpu() is prohibitively expensive on machines with | |
34 | * serious number of cpus, therefore we need to take a distributed approach | |
35 | * to calculating nr_active. | |
36 | * | |
37 | * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0 | |
38 | * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) } | |
39 | * | |
40 | * So assuming nr_active := 0 when we start out -- true per definition, we | |
41 | * can simply take per-cpu deltas and fold those into a global accumulate | |
42 | * to obtain the same result. See calc_load_fold_active(). | |
43 | * | |
44 | * Furthermore, in order to avoid synchronizing all per-cpu delta folding | |
45 | * across the machine, we assume 10 ticks is sufficient time for every | |
46 | * cpu to have completed this task. | |
47 | * | |
48 | * This places an upper-bound on the IRQ-off latency of the machine. Then | |
49 | * again, being late doesn't loose the delta, just wrecks the sample. | |
50 | * | |
51 | * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because | |
52 | * this would add another cross-cpu cacheline miss and atomic operation | |
53 | * to the wakeup path. Instead we increment on whatever cpu the task ran | |
54 | * when it went into uninterruptible state and decrement on whatever cpu | |
55 | * did the wakeup. This means that only the sum of nr_uninterruptible over | |
56 | * all cpus yields the correct result. | |
57 | * | |
58 | * This covers the NO_HZ=n code, for extra head-aches, see the comment below. | |
59 | */ | |
60 | ||
61 | /* Variables and functions for calc_load */ | |
62 | atomic_long_t calc_load_tasks; | |
63 | unsigned long calc_load_update; | |
64 | unsigned long avenrun[3]; | |
65 | EXPORT_SYMBOL(avenrun); /* should be removed */ | |
66 | ||
67 | /** | |
68 | * get_avenrun - get the load average array | |
69 | * @loads: pointer to dest load array | |
70 | * @offset: offset to add | |
71 | * @shift: shift count to shift the result left | |
72 | * | |
73 | * These values are estimates at best, so no need for locking. | |
74 | */ | |
75 | void get_avenrun(unsigned long *loads, unsigned long offset, int shift) | |
76 | { | |
77 | loads[0] = (avenrun[0] + offset) << shift; | |
78 | loads[1] = (avenrun[1] + offset) << shift; | |
79 | loads[2] = (avenrun[2] + offset) << shift; | |
80 | } | |
81 | ||
d60585c5 | 82 | long calc_load_fold_active(struct rq *this_rq, long adjust) |
45ceebf7 PG |
83 | { |
84 | long nr_active, delta = 0; | |
85 | ||
d60585c5 | 86 | nr_active = this_rq->nr_running - adjust; |
3289bdb4 | 87 | nr_active += (long)this_rq->nr_uninterruptible; |
45ceebf7 PG |
88 | |
89 | if (nr_active != this_rq->calc_load_active) { | |
90 | delta = nr_active - this_rq->calc_load_active; | |
91 | this_rq->calc_load_active = nr_active; | |
92 | } | |
93 | ||
94 | return delta; | |
95 | } | |
96 | ||
97 | /* | |
98 | * a1 = a0 * e + a * (1 - e) | |
99 | */ | |
100 | static unsigned long | |
101 | calc_load(unsigned long load, unsigned long exp, unsigned long active) | |
102 | { | |
20878232 VH |
103 | unsigned long newload; |
104 | ||
105 | newload = load * exp + active * (FIXED_1 - exp); | |
106 | if (active >= load) | |
107 | newload += FIXED_1-1; | |
108 | ||
109 | return newload / FIXED_1; | |
45ceebf7 PG |
110 | } |
111 | ||
112 | #ifdef CONFIG_NO_HZ_COMMON | |
113 | /* | |
114 | * Handle NO_HZ for the global load-average. | |
115 | * | |
116 | * Since the above described distributed algorithm to compute the global | |
117 | * load-average relies on per-cpu sampling from the tick, it is affected by | |
118 | * NO_HZ. | |
119 | * | |
3c85d6db | 120 | * The basic idea is to fold the nr_active delta into a global NO_HZ-delta upon |
45ceebf7 PG |
121 | * entering NO_HZ state such that we can include this as an 'extra' cpu delta |
122 | * when we read the global state. | |
123 | * | |
124 | * Obviously reality has to ruin such a delightfully simple scheme: | |
125 | * | |
126 | * - When we go NO_HZ idle during the window, we can negate our sample | |
127 | * contribution, causing under-accounting. | |
128 | * | |
3c85d6db | 129 | * We avoid this by keeping two NO_HZ-delta counters and flipping them |
45ceebf7 PG |
130 | * when the window starts, thus separating old and new NO_HZ load. |
131 | * | |
132 | * The only trick is the slight shift in index flip for read vs write. | |
133 | * | |
134 | * 0s 5s 10s 15s | |
135 | * +10 +10 +10 +10 | |
136 | * |-|-----------|-|-----------|-|-----------|-| | |
137 | * r:0 0 1 1 0 0 1 1 0 | |
138 | * w:0 1 1 0 0 1 1 0 0 | |
139 | * | |
3c85d6db | 140 | * This ensures we'll fold the old NO_HZ contribution in this window while |
45ceebf7 PG |
141 | * accumlating the new one. |
142 | * | |
3c85d6db | 143 | * - When we wake up from NO_HZ during the window, we push up our |
45ceebf7 PG |
144 | * contribution, since we effectively move our sample point to a known |
145 | * busy state. | |
146 | * | |
147 | * This is solved by pushing the window forward, and thus skipping the | |
3c85d6db | 148 | * sample, for this cpu (effectively using the NO_HZ-delta for this cpu which |
45ceebf7 | 149 | * was in effect at the time the window opened). This also solves the issue |
3c85d6db FW |
150 | * of having to deal with a cpu having been in NO_HZ for multiple LOAD_FREQ |
151 | * intervals. | |
45ceebf7 PG |
152 | * |
153 | * When making the ILB scale, we should try to pull this in as well. | |
154 | */ | |
3c85d6db | 155 | static atomic_long_t calc_load_nohz[2]; |
45ceebf7 PG |
156 | static int calc_load_idx; |
157 | ||
158 | static inline int calc_load_write_idx(void) | |
159 | { | |
160 | int idx = calc_load_idx; | |
161 | ||
162 | /* | |
163 | * See calc_global_nohz(), if we observe the new index, we also | |
164 | * need to observe the new update time. | |
165 | */ | |
166 | smp_rmb(); | |
167 | ||
168 | /* | |
169 | * If the folding window started, make sure we start writing in the | |
3c85d6db | 170 | * next NO_HZ-delta. |
45ceebf7 | 171 | */ |
caeb5882 | 172 | if (!time_before(jiffies, READ_ONCE(calc_load_update))) |
45ceebf7 PG |
173 | idx++; |
174 | ||
175 | return idx & 1; | |
176 | } | |
177 | ||
178 | static inline int calc_load_read_idx(void) | |
179 | { | |
180 | return calc_load_idx & 1; | |
181 | } | |
182 | ||
3c85d6db | 183 | void calc_load_nohz_start(void) |
45ceebf7 PG |
184 | { |
185 | struct rq *this_rq = this_rq(); | |
186 | long delta; | |
187 | ||
188 | /* | |
3c85d6db FW |
189 | * We're going into NO_HZ mode, if there's any pending delta, fold it |
190 | * into the pending NO_HZ delta. | |
45ceebf7 | 191 | */ |
d60585c5 | 192 | delta = calc_load_fold_active(this_rq, 0); |
45ceebf7 PG |
193 | if (delta) { |
194 | int idx = calc_load_write_idx(); | |
3289bdb4 | 195 | |
3c85d6db | 196 | atomic_long_add(delta, &calc_load_nohz[idx]); |
45ceebf7 PG |
197 | } |
198 | } | |
199 | ||
3c85d6db | 200 | void calc_load_nohz_stop(void) |
45ceebf7 PG |
201 | { |
202 | struct rq *this_rq = this_rq(); | |
203 | ||
204 | /* | |
6e5f32f7 | 205 | * If we're still before the pending sample window, we're done. |
45ceebf7 | 206 | */ |
caeb5882 | 207 | this_rq->calc_load_update = READ_ONCE(calc_load_update); |
45ceebf7 PG |
208 | if (time_before(jiffies, this_rq->calc_load_update)) |
209 | return; | |
210 | ||
211 | /* | |
212 | * We woke inside or after the sample window, this means we're already | |
213 | * accounted through the nohz accounting, so skip the entire deal and | |
214 | * sync up for the next window. | |
215 | */ | |
45ceebf7 PG |
216 | if (time_before(jiffies, this_rq->calc_load_update + 10)) |
217 | this_rq->calc_load_update += LOAD_FREQ; | |
218 | } | |
219 | ||
3c85d6db | 220 | static long calc_load_nohz_fold(void) |
45ceebf7 PG |
221 | { |
222 | int idx = calc_load_read_idx(); | |
223 | long delta = 0; | |
224 | ||
3c85d6db FW |
225 | if (atomic_long_read(&calc_load_nohz[idx])) |
226 | delta = atomic_long_xchg(&calc_load_nohz[idx], 0); | |
45ceebf7 PG |
227 | |
228 | return delta; | |
229 | } | |
230 | ||
231 | /** | |
232 | * fixed_power_int - compute: x^n, in O(log n) time | |
233 | * | |
234 | * @x: base of the power | |
235 | * @frac_bits: fractional bits of @x | |
236 | * @n: power to raise @x to. | |
237 | * | |
238 | * By exploiting the relation between the definition of the natural power | |
239 | * function: x^n := x*x*...*x (x multiplied by itself for n times), and | |
240 | * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, | |
241 | * (where: n_i \elem {0, 1}, the binary vector representing n), | |
242 | * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is | |
243 | * of course trivially computable in O(log_2 n), the length of our binary | |
244 | * vector. | |
245 | */ | |
246 | static unsigned long | |
247 | fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) | |
248 | { | |
249 | unsigned long result = 1UL << frac_bits; | |
250 | ||
3289bdb4 PZ |
251 | if (n) { |
252 | for (;;) { | |
253 | if (n & 1) { | |
254 | result *= x; | |
255 | result += 1UL << (frac_bits - 1); | |
256 | result >>= frac_bits; | |
257 | } | |
258 | n >>= 1; | |
259 | if (!n) | |
260 | break; | |
261 | x *= x; | |
262 | x += 1UL << (frac_bits - 1); | |
263 | x >>= frac_bits; | |
45ceebf7 | 264 | } |
45ceebf7 PG |
265 | } |
266 | ||
267 | return result; | |
268 | } | |
269 | ||
270 | /* | |
271 | * a1 = a0 * e + a * (1 - e) | |
272 | * | |
273 | * a2 = a1 * e + a * (1 - e) | |
274 | * = (a0 * e + a * (1 - e)) * e + a * (1 - e) | |
275 | * = a0 * e^2 + a * (1 - e) * (1 + e) | |
276 | * | |
277 | * a3 = a2 * e + a * (1 - e) | |
278 | * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) | |
279 | * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) | |
280 | * | |
281 | * ... | |
282 | * | |
283 | * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] | |
284 | * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) | |
285 | * = a0 * e^n + a * (1 - e^n) | |
286 | * | |
287 | * [1] application of the geometric series: | |
288 | * | |
289 | * n 1 - x^(n+1) | |
290 | * S_n := \Sum x^i = ------------- | |
291 | * i=0 1 - x | |
292 | */ | |
293 | static unsigned long | |
294 | calc_load_n(unsigned long load, unsigned long exp, | |
295 | unsigned long active, unsigned int n) | |
296 | { | |
45ceebf7 PG |
297 | return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); |
298 | } | |
299 | ||
300 | /* | |
301 | * NO_HZ can leave us missing all per-cpu ticks calling | |
3c85d6db FW |
302 | * calc_load_fold_active(), but since a NO_HZ CPU folds its delta into |
303 | * calc_load_nohz per calc_load_nohz_start(), all we need to do is fold | |
304 | * in the pending NO_HZ delta if our NO_HZ period crossed a load cycle boundary. | |
45ceebf7 PG |
305 | * |
306 | * Once we've updated the global active value, we need to apply the exponential | |
307 | * weights adjusted to the number of cycles missed. | |
308 | */ | |
309 | static void calc_global_nohz(void) | |
310 | { | |
caeb5882 | 311 | unsigned long sample_window; |
45ceebf7 PG |
312 | long delta, active, n; |
313 | ||
caeb5882 MF |
314 | sample_window = READ_ONCE(calc_load_update); |
315 | if (!time_before(jiffies, sample_window + 10)) { | |
45ceebf7 PG |
316 | /* |
317 | * Catch-up, fold however many we are behind still | |
318 | */ | |
caeb5882 | 319 | delta = jiffies - sample_window - 10; |
45ceebf7 PG |
320 | n = 1 + (delta / LOAD_FREQ); |
321 | ||
322 | active = atomic_long_read(&calc_load_tasks); | |
323 | active = active > 0 ? active * FIXED_1 : 0; | |
324 | ||
325 | avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); | |
326 | avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); | |
327 | avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); | |
328 | ||
caeb5882 | 329 | WRITE_ONCE(calc_load_update, sample_window + n * LOAD_FREQ); |
45ceebf7 PG |
330 | } |
331 | ||
332 | /* | |
3c85d6db | 333 | * Flip the NO_HZ index... |
45ceebf7 PG |
334 | * |
335 | * Make sure we first write the new time then flip the index, so that | |
336 | * calc_load_write_idx() will see the new time when it reads the new | |
337 | * index, this avoids a double flip messing things up. | |
338 | */ | |
339 | smp_wmb(); | |
340 | calc_load_idx++; | |
341 | } | |
342 | #else /* !CONFIG_NO_HZ_COMMON */ | |
343 | ||
3c85d6db | 344 | static inline long calc_load_nohz_fold(void) { return 0; } |
45ceebf7 PG |
345 | static inline void calc_global_nohz(void) { } |
346 | ||
347 | #endif /* CONFIG_NO_HZ_COMMON */ | |
348 | ||
349 | /* | |
350 | * calc_load - update the avenrun load estimates 10 ticks after the | |
351 | * CPUs have updated calc_load_tasks. | |
3289bdb4 PZ |
352 | * |
353 | * Called from the global timer code. | |
45ceebf7 PG |
354 | */ |
355 | void calc_global_load(unsigned long ticks) | |
356 | { | |
caeb5882 | 357 | unsigned long sample_window; |
45ceebf7 PG |
358 | long active, delta; |
359 | ||
caeb5882 MF |
360 | sample_window = READ_ONCE(calc_load_update); |
361 | if (time_before(jiffies, sample_window + 10)) | |
45ceebf7 PG |
362 | return; |
363 | ||
364 | /* | |
3c85d6db | 365 | * Fold the 'old' NO_HZ-delta to include all NO_HZ cpus. |
45ceebf7 | 366 | */ |
3c85d6db | 367 | delta = calc_load_nohz_fold(); |
45ceebf7 PG |
368 | if (delta) |
369 | atomic_long_add(delta, &calc_load_tasks); | |
370 | ||
371 | active = atomic_long_read(&calc_load_tasks); | |
372 | active = active > 0 ? active * FIXED_1 : 0; | |
373 | ||
374 | avenrun[0] = calc_load(avenrun[0], EXP_1, active); | |
375 | avenrun[1] = calc_load(avenrun[1], EXP_5, active); | |
376 | avenrun[2] = calc_load(avenrun[2], EXP_15, active); | |
377 | ||
caeb5882 | 378 | WRITE_ONCE(calc_load_update, sample_window + LOAD_FREQ); |
45ceebf7 PG |
379 | |
380 | /* | |
3c85d6db FW |
381 | * In case we went to NO_HZ for multiple LOAD_FREQ intervals |
382 | * catch up in bulk. | |
45ceebf7 PG |
383 | */ |
384 | calc_global_nohz(); | |
385 | } | |
386 | ||
387 | /* | |
3289bdb4 | 388 | * Called from scheduler_tick() to periodically update this CPU's |
45ceebf7 PG |
389 | * active count. |
390 | */ | |
3289bdb4 | 391 | void calc_global_load_tick(struct rq *this_rq) |
45ceebf7 PG |
392 | { |
393 | long delta; | |
394 | ||
395 | if (time_before(jiffies, this_rq->calc_load_update)) | |
396 | return; | |
397 | ||
d60585c5 | 398 | delta = calc_load_fold_active(this_rq, 0); |
45ceebf7 PG |
399 | if (delta) |
400 | atomic_long_add(delta, &calc_load_tasks); | |
401 | ||
402 | this_rq->calc_load_update += LOAD_FREQ; | |
403 | } |