]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
45ceebf7 | 2 | /* |
3289bdb4 | 3 | * kernel/sched/loadavg.c |
45ceebf7 | 4 | * |
3289bdb4 PZ |
5 | * This file contains the magic bits required to compute the global loadavg |
6 | * figure. Its a silly number but people think its important. We go through | |
7 | * great pains to make it work on big machines and tickless kernels. | |
45ceebf7 PG |
8 | */ |
9 | ||
10 | #include <linux/export.h> | |
4f17722c | 11 | #include <linux/sched/loadavg.h> |
45ceebf7 PG |
12 | |
13 | #include "sched.h" | |
14 | ||
45ceebf7 PG |
15 | /* |
16 | * Global load-average calculations | |
17 | * | |
18 | * We take a distributed and async approach to calculating the global load-avg | |
19 | * in order to minimize overhead. | |
20 | * | |
21 | * The global load average is an exponentially decaying average of nr_running + | |
22 | * nr_uninterruptible. | |
23 | * | |
24 | * Once every LOAD_FREQ: | |
25 | * | |
26 | * nr_active = 0; | |
27 | * for_each_possible_cpu(cpu) | |
28 | * nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible; | |
29 | * | |
30 | * avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n) | |
31 | * | |
32 | * Due to a number of reasons the above turns in the mess below: | |
33 | * | |
34 | * - for_each_possible_cpu() is prohibitively expensive on machines with | |
35 | * serious number of cpus, therefore we need to take a distributed approach | |
36 | * to calculating nr_active. | |
37 | * | |
38 | * \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0 | |
39 | * = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) } | |
40 | * | |
41 | * So assuming nr_active := 0 when we start out -- true per definition, we | |
42 | * can simply take per-cpu deltas and fold those into a global accumulate | |
43 | * to obtain the same result. See calc_load_fold_active(). | |
44 | * | |
45 | * Furthermore, in order to avoid synchronizing all per-cpu delta folding | |
46 | * across the machine, we assume 10 ticks is sufficient time for every | |
47 | * cpu to have completed this task. | |
48 | * | |
49 | * This places an upper-bound on the IRQ-off latency of the machine. Then | |
50 | * again, being late doesn't loose the delta, just wrecks the sample. | |
51 | * | |
52 | * - cpu_rq()->nr_uninterruptible isn't accurately tracked per-cpu because | |
53 | * this would add another cross-cpu cacheline miss and atomic operation | |
54 | * to the wakeup path. Instead we increment on whatever cpu the task ran | |
55 | * when it went into uninterruptible state and decrement on whatever cpu | |
56 | * did the wakeup. This means that only the sum of nr_uninterruptible over | |
57 | * all cpus yields the correct result. | |
58 | * | |
59 | * This covers the NO_HZ=n code, for extra head-aches, see the comment below. | |
60 | */ | |
61 | ||
62 | /* Variables and functions for calc_load */ | |
63 | atomic_long_t calc_load_tasks; | |
64 | unsigned long calc_load_update; | |
65 | unsigned long avenrun[3]; | |
66 | EXPORT_SYMBOL(avenrun); /* should be removed */ | |
67 | ||
68 | /** | |
69 | * get_avenrun - get the load average array | |
70 | * @loads: pointer to dest load array | |
71 | * @offset: offset to add | |
72 | * @shift: shift count to shift the result left | |
73 | * | |
74 | * These values are estimates at best, so no need for locking. | |
75 | */ | |
76 | void get_avenrun(unsigned long *loads, unsigned long offset, int shift) | |
77 | { | |
78 | loads[0] = (avenrun[0] + offset) << shift; | |
79 | loads[1] = (avenrun[1] + offset) << shift; | |
80 | loads[2] = (avenrun[2] + offset) << shift; | |
81 | } | |
82 | ||
d60585c5 | 83 | long calc_load_fold_active(struct rq *this_rq, long adjust) |
45ceebf7 PG |
84 | { |
85 | long nr_active, delta = 0; | |
86 | ||
d60585c5 | 87 | nr_active = this_rq->nr_running - adjust; |
3289bdb4 | 88 | nr_active += (long)this_rq->nr_uninterruptible; |
45ceebf7 PG |
89 | |
90 | if (nr_active != this_rq->calc_load_active) { | |
91 | delta = nr_active - this_rq->calc_load_active; | |
92 | this_rq->calc_load_active = nr_active; | |
93 | } | |
94 | ||
95 | return delta; | |
96 | } | |
97 | ||
98 | /* | |
99 | * a1 = a0 * e + a * (1 - e) | |
100 | */ | |
101 | static unsigned long | |
102 | calc_load(unsigned long load, unsigned long exp, unsigned long active) | |
103 | { | |
20878232 VH |
104 | unsigned long newload; |
105 | ||
106 | newload = load * exp + active * (FIXED_1 - exp); | |
107 | if (active >= load) | |
108 | newload += FIXED_1-1; | |
109 | ||
110 | return newload / FIXED_1; | |
45ceebf7 PG |
111 | } |
112 | ||
113 | #ifdef CONFIG_NO_HZ_COMMON | |
114 | /* | |
115 | * Handle NO_HZ for the global load-average. | |
116 | * | |
117 | * Since the above described distributed algorithm to compute the global | |
118 | * load-average relies on per-cpu sampling from the tick, it is affected by | |
119 | * NO_HZ. | |
120 | * | |
3c85d6db | 121 | * The basic idea is to fold the nr_active delta into a global NO_HZ-delta upon |
45ceebf7 PG |
122 | * entering NO_HZ state such that we can include this as an 'extra' cpu delta |
123 | * when we read the global state. | |
124 | * | |
125 | * Obviously reality has to ruin such a delightfully simple scheme: | |
126 | * | |
127 | * - When we go NO_HZ idle during the window, we can negate our sample | |
128 | * contribution, causing under-accounting. | |
129 | * | |
3c85d6db | 130 | * We avoid this by keeping two NO_HZ-delta counters and flipping them |
45ceebf7 PG |
131 | * when the window starts, thus separating old and new NO_HZ load. |
132 | * | |
133 | * The only trick is the slight shift in index flip for read vs write. | |
134 | * | |
135 | * 0s 5s 10s 15s | |
136 | * +10 +10 +10 +10 | |
137 | * |-|-----------|-|-----------|-|-----------|-| | |
138 | * r:0 0 1 1 0 0 1 1 0 | |
139 | * w:0 1 1 0 0 1 1 0 0 | |
140 | * | |
3c85d6db | 141 | * This ensures we'll fold the old NO_HZ contribution in this window while |
45ceebf7 PG |
142 | * accumlating the new one. |
143 | * | |
3c85d6db | 144 | * - When we wake up from NO_HZ during the window, we push up our |
45ceebf7 PG |
145 | * contribution, since we effectively move our sample point to a known |
146 | * busy state. | |
147 | * | |
148 | * This is solved by pushing the window forward, and thus skipping the | |
3c85d6db | 149 | * sample, for this cpu (effectively using the NO_HZ-delta for this cpu which |
45ceebf7 | 150 | * was in effect at the time the window opened). This also solves the issue |
3c85d6db FW |
151 | * of having to deal with a cpu having been in NO_HZ for multiple LOAD_FREQ |
152 | * intervals. | |
45ceebf7 PG |
153 | * |
154 | * When making the ILB scale, we should try to pull this in as well. | |
155 | */ | |
3c85d6db | 156 | static atomic_long_t calc_load_nohz[2]; |
45ceebf7 PG |
157 | static int calc_load_idx; |
158 | ||
159 | static inline int calc_load_write_idx(void) | |
160 | { | |
161 | int idx = calc_load_idx; | |
162 | ||
163 | /* | |
164 | * See calc_global_nohz(), if we observe the new index, we also | |
165 | * need to observe the new update time. | |
166 | */ | |
167 | smp_rmb(); | |
168 | ||
169 | /* | |
170 | * If the folding window started, make sure we start writing in the | |
3c85d6db | 171 | * next NO_HZ-delta. |
45ceebf7 | 172 | */ |
caeb5882 | 173 | if (!time_before(jiffies, READ_ONCE(calc_load_update))) |
45ceebf7 PG |
174 | idx++; |
175 | ||
176 | return idx & 1; | |
177 | } | |
178 | ||
179 | static inline int calc_load_read_idx(void) | |
180 | { | |
181 | return calc_load_idx & 1; | |
182 | } | |
183 | ||
3c85d6db | 184 | void calc_load_nohz_start(void) |
45ceebf7 PG |
185 | { |
186 | struct rq *this_rq = this_rq(); | |
187 | long delta; | |
188 | ||
189 | /* | |
3c85d6db FW |
190 | * We're going into NO_HZ mode, if there's any pending delta, fold it |
191 | * into the pending NO_HZ delta. | |
45ceebf7 | 192 | */ |
d60585c5 | 193 | delta = calc_load_fold_active(this_rq, 0); |
45ceebf7 PG |
194 | if (delta) { |
195 | int idx = calc_load_write_idx(); | |
3289bdb4 | 196 | |
3c85d6db | 197 | atomic_long_add(delta, &calc_load_nohz[idx]); |
45ceebf7 PG |
198 | } |
199 | } | |
200 | ||
3c85d6db | 201 | void calc_load_nohz_stop(void) |
45ceebf7 PG |
202 | { |
203 | struct rq *this_rq = this_rq(); | |
204 | ||
205 | /* | |
6e5f32f7 | 206 | * If we're still before the pending sample window, we're done. |
45ceebf7 | 207 | */ |
caeb5882 | 208 | this_rq->calc_load_update = READ_ONCE(calc_load_update); |
45ceebf7 PG |
209 | if (time_before(jiffies, this_rq->calc_load_update)) |
210 | return; | |
211 | ||
212 | /* | |
213 | * We woke inside or after the sample window, this means we're already | |
214 | * accounted through the nohz accounting, so skip the entire deal and | |
215 | * sync up for the next window. | |
216 | */ | |
45ceebf7 PG |
217 | if (time_before(jiffies, this_rq->calc_load_update + 10)) |
218 | this_rq->calc_load_update += LOAD_FREQ; | |
219 | } | |
220 | ||
3c85d6db | 221 | static long calc_load_nohz_fold(void) |
45ceebf7 PG |
222 | { |
223 | int idx = calc_load_read_idx(); | |
224 | long delta = 0; | |
225 | ||
3c85d6db FW |
226 | if (atomic_long_read(&calc_load_nohz[idx])) |
227 | delta = atomic_long_xchg(&calc_load_nohz[idx], 0); | |
45ceebf7 PG |
228 | |
229 | return delta; | |
230 | } | |
231 | ||
232 | /** | |
233 | * fixed_power_int - compute: x^n, in O(log n) time | |
234 | * | |
235 | * @x: base of the power | |
236 | * @frac_bits: fractional bits of @x | |
237 | * @n: power to raise @x to. | |
238 | * | |
239 | * By exploiting the relation between the definition of the natural power | |
240 | * function: x^n := x*x*...*x (x multiplied by itself for n times), and | |
241 | * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, | |
242 | * (where: n_i \elem {0, 1}, the binary vector representing n), | |
243 | * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is | |
244 | * of course trivially computable in O(log_2 n), the length of our binary | |
245 | * vector. | |
246 | */ | |
247 | static unsigned long | |
248 | fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) | |
249 | { | |
250 | unsigned long result = 1UL << frac_bits; | |
251 | ||
3289bdb4 PZ |
252 | if (n) { |
253 | for (;;) { | |
254 | if (n & 1) { | |
255 | result *= x; | |
256 | result += 1UL << (frac_bits - 1); | |
257 | result >>= frac_bits; | |
258 | } | |
259 | n >>= 1; | |
260 | if (!n) | |
261 | break; | |
262 | x *= x; | |
263 | x += 1UL << (frac_bits - 1); | |
264 | x >>= frac_bits; | |
45ceebf7 | 265 | } |
45ceebf7 PG |
266 | } |
267 | ||
268 | return result; | |
269 | } | |
270 | ||
271 | /* | |
272 | * a1 = a0 * e + a * (1 - e) | |
273 | * | |
274 | * a2 = a1 * e + a * (1 - e) | |
275 | * = (a0 * e + a * (1 - e)) * e + a * (1 - e) | |
276 | * = a0 * e^2 + a * (1 - e) * (1 + e) | |
277 | * | |
278 | * a3 = a2 * e + a * (1 - e) | |
279 | * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) | |
280 | * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) | |
281 | * | |
282 | * ... | |
283 | * | |
284 | * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] | |
285 | * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) | |
286 | * = a0 * e^n + a * (1 - e^n) | |
287 | * | |
288 | * [1] application of the geometric series: | |
289 | * | |
290 | * n 1 - x^(n+1) | |
291 | * S_n := \Sum x^i = ------------- | |
292 | * i=0 1 - x | |
293 | */ | |
294 | static unsigned long | |
295 | calc_load_n(unsigned long load, unsigned long exp, | |
296 | unsigned long active, unsigned int n) | |
297 | { | |
45ceebf7 PG |
298 | return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); |
299 | } | |
300 | ||
301 | /* | |
302 | * NO_HZ can leave us missing all per-cpu ticks calling | |
3c85d6db FW |
303 | * calc_load_fold_active(), but since a NO_HZ CPU folds its delta into |
304 | * calc_load_nohz per calc_load_nohz_start(), all we need to do is fold | |
305 | * in the pending NO_HZ delta if our NO_HZ period crossed a load cycle boundary. | |
45ceebf7 PG |
306 | * |
307 | * Once we've updated the global active value, we need to apply the exponential | |
308 | * weights adjusted to the number of cycles missed. | |
309 | */ | |
310 | static void calc_global_nohz(void) | |
311 | { | |
caeb5882 | 312 | unsigned long sample_window; |
45ceebf7 PG |
313 | long delta, active, n; |
314 | ||
caeb5882 MF |
315 | sample_window = READ_ONCE(calc_load_update); |
316 | if (!time_before(jiffies, sample_window + 10)) { | |
45ceebf7 PG |
317 | /* |
318 | * Catch-up, fold however many we are behind still | |
319 | */ | |
caeb5882 | 320 | delta = jiffies - sample_window - 10; |
45ceebf7 PG |
321 | n = 1 + (delta / LOAD_FREQ); |
322 | ||
323 | active = atomic_long_read(&calc_load_tasks); | |
324 | active = active > 0 ? active * FIXED_1 : 0; | |
325 | ||
326 | avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); | |
327 | avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); | |
328 | avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); | |
329 | ||
caeb5882 | 330 | WRITE_ONCE(calc_load_update, sample_window + n * LOAD_FREQ); |
45ceebf7 PG |
331 | } |
332 | ||
333 | /* | |
3c85d6db | 334 | * Flip the NO_HZ index... |
45ceebf7 PG |
335 | * |
336 | * Make sure we first write the new time then flip the index, so that | |
337 | * calc_load_write_idx() will see the new time when it reads the new | |
338 | * index, this avoids a double flip messing things up. | |
339 | */ | |
340 | smp_wmb(); | |
341 | calc_load_idx++; | |
342 | } | |
343 | #else /* !CONFIG_NO_HZ_COMMON */ | |
344 | ||
3c85d6db | 345 | static inline long calc_load_nohz_fold(void) { return 0; } |
45ceebf7 PG |
346 | static inline void calc_global_nohz(void) { } |
347 | ||
348 | #endif /* CONFIG_NO_HZ_COMMON */ | |
349 | ||
350 | /* | |
351 | * calc_load - update the avenrun load estimates 10 ticks after the | |
352 | * CPUs have updated calc_load_tasks. | |
3289bdb4 PZ |
353 | * |
354 | * Called from the global timer code. | |
45ceebf7 PG |
355 | */ |
356 | void calc_global_load(unsigned long ticks) | |
357 | { | |
caeb5882 | 358 | unsigned long sample_window; |
45ceebf7 PG |
359 | long active, delta; |
360 | ||
caeb5882 MF |
361 | sample_window = READ_ONCE(calc_load_update); |
362 | if (time_before(jiffies, sample_window + 10)) | |
45ceebf7 PG |
363 | return; |
364 | ||
365 | /* | |
3c85d6db | 366 | * Fold the 'old' NO_HZ-delta to include all NO_HZ cpus. |
45ceebf7 | 367 | */ |
3c85d6db | 368 | delta = calc_load_nohz_fold(); |
45ceebf7 PG |
369 | if (delta) |
370 | atomic_long_add(delta, &calc_load_tasks); | |
371 | ||
372 | active = atomic_long_read(&calc_load_tasks); | |
373 | active = active > 0 ? active * FIXED_1 : 0; | |
374 | ||
375 | avenrun[0] = calc_load(avenrun[0], EXP_1, active); | |
376 | avenrun[1] = calc_load(avenrun[1], EXP_5, active); | |
377 | avenrun[2] = calc_load(avenrun[2], EXP_15, active); | |
378 | ||
caeb5882 | 379 | WRITE_ONCE(calc_load_update, sample_window + LOAD_FREQ); |
45ceebf7 PG |
380 | |
381 | /* | |
3c85d6db FW |
382 | * In case we went to NO_HZ for multiple LOAD_FREQ intervals |
383 | * catch up in bulk. | |
45ceebf7 PG |
384 | */ |
385 | calc_global_nohz(); | |
386 | } | |
387 | ||
388 | /* | |
3289bdb4 | 389 | * Called from scheduler_tick() to periodically update this CPU's |
45ceebf7 PG |
390 | * active count. |
391 | */ | |
3289bdb4 | 392 | void calc_global_load_tick(struct rq *this_rq) |
45ceebf7 PG |
393 | { |
394 | long delta; | |
395 | ||
396 | if (time_before(jiffies, this_rq->calc_load_update)) | |
397 | return; | |
398 | ||
d60585c5 | 399 | delta = calc_load_fold_active(this_rq, 0); |
45ceebf7 PG |
400 | if (delta) |
401 | atomic_long_add(delta, &calc_load_tasks); | |
402 | ||
403 | this_rq->calc_load_update += LOAD_FREQ; | |
404 | } |