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> | |
10 | ||
11 | #include "sched.h" | |
12 | ||
45ceebf7 PG |
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 | ||
d60585c5 | 81 | long calc_load_fold_active(struct rq *this_rq, long adjust) |
45ceebf7 PG |
82 | { |
83 | long nr_active, delta = 0; | |
84 | ||
d60585c5 | 85 | nr_active = this_rq->nr_running - adjust; |
3289bdb4 | 86 | nr_active += (long)this_rq->nr_uninterruptible; |
45ceebf7 PG |
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 | { | |
20878232 VH |
102 | unsigned long newload; |
103 | ||
104 | newload = load * exp + active * (FIXED_1 - exp); | |
105 | if (active >= load) | |
106 | newload += FIXED_1-1; | |
107 | ||
108 | return newload / FIXED_1; | |
45ceebf7 PG |
109 | } |
110 | ||
111 | #ifdef CONFIG_NO_HZ_COMMON | |
112 | /* | |
113 | * Handle NO_HZ for the global load-average. | |
114 | * | |
115 | * Since the above described distributed algorithm to compute the global | |
116 | * load-average relies on per-cpu sampling from the tick, it is affected by | |
117 | * NO_HZ. | |
118 | * | |
119 | * The basic idea is to fold the nr_active delta into a global idle-delta upon | |
120 | * entering NO_HZ state such that we can include this as an 'extra' cpu delta | |
121 | * when we read the global state. | |
122 | * | |
123 | * Obviously reality has to ruin such a delightfully simple scheme: | |
124 | * | |
125 | * - When we go NO_HZ idle during the window, we can negate our sample | |
126 | * contribution, causing under-accounting. | |
127 | * | |
128 | * We avoid this by keeping two idle-delta counters and flipping them | |
129 | * when the window starts, thus separating old and new NO_HZ load. | |
130 | * | |
131 | * The only trick is the slight shift in index flip for read vs write. | |
132 | * | |
133 | * 0s 5s 10s 15s | |
134 | * +10 +10 +10 +10 | |
135 | * |-|-----------|-|-----------|-|-----------|-| | |
136 | * r:0 0 1 1 0 0 1 1 0 | |
137 | * w:0 1 1 0 0 1 1 0 0 | |
138 | * | |
139 | * This ensures we'll fold the old idle contribution in this window while | |
140 | * accumlating the new one. | |
141 | * | |
142 | * - When we wake up from NO_HZ idle during the window, we push up our | |
143 | * contribution, since we effectively move our sample point to a known | |
144 | * busy state. | |
145 | * | |
146 | * This is solved by pushing the window forward, and thus skipping the | |
147 | * sample, for this cpu (effectively using the idle-delta for this cpu which | |
148 | * was in effect at the time the window opened). This also solves the issue | |
149 | * of having to deal with a cpu having been in NOHZ idle for multiple | |
150 | * LOAD_FREQ intervals. | |
151 | * | |
152 | * When making the ILB scale, we should try to pull this in as well. | |
153 | */ | |
154 | static atomic_long_t calc_load_idle[2]; | |
155 | static int calc_load_idx; | |
156 | ||
157 | static inline int calc_load_write_idx(void) | |
158 | { | |
159 | int idx = calc_load_idx; | |
160 | ||
161 | /* | |
162 | * See calc_global_nohz(), if we observe the new index, we also | |
163 | * need to observe the new update time. | |
164 | */ | |
165 | smp_rmb(); | |
166 | ||
167 | /* | |
168 | * If the folding window started, make sure we start writing in the | |
169 | * next idle-delta. | |
170 | */ | |
171 | if (!time_before(jiffies, calc_load_update)) | |
172 | idx++; | |
173 | ||
174 | return idx & 1; | |
175 | } | |
176 | ||
177 | static inline int calc_load_read_idx(void) | |
178 | { | |
179 | return calc_load_idx & 1; | |
180 | } | |
181 | ||
182 | void calc_load_enter_idle(void) | |
183 | { | |
184 | struct rq *this_rq = this_rq(); | |
185 | long delta; | |
186 | ||
187 | /* | |
188 | * We're going into NOHZ mode, if there's any pending delta, fold it | |
189 | * into the pending idle delta. | |
190 | */ | |
d60585c5 | 191 | delta = calc_load_fold_active(this_rq, 0); |
45ceebf7 PG |
192 | if (delta) { |
193 | int idx = calc_load_write_idx(); | |
3289bdb4 | 194 | |
45ceebf7 PG |
195 | atomic_long_add(delta, &calc_load_idle[idx]); |
196 | } | |
197 | } | |
198 | ||
199 | void calc_load_exit_idle(void) | |
200 | { | |
201 | struct rq *this_rq = this_rq(); | |
202 | ||
203 | /* | |
204 | * If we're still before the sample window, we're done. | |
205 | */ | |
206 | if (time_before(jiffies, this_rq->calc_load_update)) | |
207 | return; | |
208 | ||
209 | /* | |
210 | * We woke inside or after the sample window, this means we're already | |
211 | * accounted through the nohz accounting, so skip the entire deal and | |
212 | * sync up for the next window. | |
213 | */ | |
214 | this_rq->calc_load_update = calc_load_update; | |
215 | if (time_before(jiffies, this_rq->calc_load_update + 10)) | |
216 | this_rq->calc_load_update += LOAD_FREQ; | |
217 | } | |
218 | ||
219 | static long calc_load_fold_idle(void) | |
220 | { | |
221 | int idx = calc_load_read_idx(); | |
222 | long delta = 0; | |
223 | ||
224 | if (atomic_long_read(&calc_load_idle[idx])) | |
225 | delta = atomic_long_xchg(&calc_load_idle[idx], 0); | |
226 | ||
227 | return delta; | |
228 | } | |
229 | ||
230 | /** | |
231 | * fixed_power_int - compute: x^n, in O(log n) time | |
232 | * | |
233 | * @x: base of the power | |
234 | * @frac_bits: fractional bits of @x | |
235 | * @n: power to raise @x to. | |
236 | * | |
237 | * By exploiting the relation between the definition of the natural power | |
238 | * function: x^n := x*x*...*x (x multiplied by itself for n times), and | |
239 | * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, | |
240 | * (where: n_i \elem {0, 1}, the binary vector representing n), | |
241 | * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is | |
242 | * of course trivially computable in O(log_2 n), the length of our binary | |
243 | * vector. | |
244 | */ | |
245 | static unsigned long | |
246 | fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) | |
247 | { | |
248 | unsigned long result = 1UL << frac_bits; | |
249 | ||
3289bdb4 PZ |
250 | if (n) { |
251 | for (;;) { | |
252 | if (n & 1) { | |
253 | result *= x; | |
254 | result += 1UL << (frac_bits - 1); | |
255 | result >>= frac_bits; | |
256 | } | |
257 | n >>= 1; | |
258 | if (!n) | |
259 | break; | |
260 | x *= x; | |
261 | x += 1UL << (frac_bits - 1); | |
262 | x >>= frac_bits; | |
45ceebf7 | 263 | } |
45ceebf7 PG |
264 | } |
265 | ||
266 | return result; | |
267 | } | |
268 | ||
269 | /* | |
270 | * a1 = a0 * e + a * (1 - e) | |
271 | * | |
272 | * a2 = a1 * e + a * (1 - e) | |
273 | * = (a0 * e + a * (1 - e)) * e + a * (1 - e) | |
274 | * = a0 * e^2 + a * (1 - e) * (1 + e) | |
275 | * | |
276 | * a3 = a2 * e + a * (1 - e) | |
277 | * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) | |
278 | * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) | |
279 | * | |
280 | * ... | |
281 | * | |
282 | * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] | |
283 | * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) | |
284 | * = a0 * e^n + a * (1 - e^n) | |
285 | * | |
286 | * [1] application of the geometric series: | |
287 | * | |
288 | * n 1 - x^(n+1) | |
289 | * S_n := \Sum x^i = ------------- | |
290 | * i=0 1 - x | |
291 | */ | |
292 | static unsigned long | |
293 | calc_load_n(unsigned long load, unsigned long exp, | |
294 | unsigned long active, unsigned int n) | |
295 | { | |
45ceebf7 PG |
296 | return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); |
297 | } | |
298 | ||
299 | /* | |
300 | * NO_HZ can leave us missing all per-cpu ticks calling | |
301 | * calc_load_account_active(), but since an idle CPU folds its delta into | |
302 | * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold | |
303 | * in the pending idle delta if our idle period crossed a load cycle boundary. | |
304 | * | |
305 | * Once we've updated the global active value, we need to apply the exponential | |
306 | * weights adjusted to the number of cycles missed. | |
307 | */ | |
308 | static void calc_global_nohz(void) | |
309 | { | |
310 | long delta, active, n; | |
311 | ||
312 | if (!time_before(jiffies, calc_load_update + 10)) { | |
313 | /* | |
314 | * Catch-up, fold however many we are behind still | |
315 | */ | |
316 | delta = jiffies - calc_load_update - 10; | |
317 | n = 1 + (delta / LOAD_FREQ); | |
318 | ||
319 | active = atomic_long_read(&calc_load_tasks); | |
320 | active = active > 0 ? active * FIXED_1 : 0; | |
321 | ||
322 | avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); | |
323 | avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); | |
324 | avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); | |
325 | ||
326 | calc_load_update += n * LOAD_FREQ; | |
327 | } | |
328 | ||
329 | /* | |
330 | * Flip the idle index... | |
331 | * | |
332 | * Make sure we first write the new time then flip the index, so that | |
333 | * calc_load_write_idx() will see the new time when it reads the new | |
334 | * index, this avoids a double flip messing things up. | |
335 | */ | |
336 | smp_wmb(); | |
337 | calc_load_idx++; | |
338 | } | |
339 | #else /* !CONFIG_NO_HZ_COMMON */ | |
340 | ||
341 | static inline long calc_load_fold_idle(void) { return 0; } | |
342 | static inline void calc_global_nohz(void) { } | |
343 | ||
344 | #endif /* CONFIG_NO_HZ_COMMON */ | |
345 | ||
346 | /* | |
347 | * calc_load - update the avenrun load estimates 10 ticks after the | |
348 | * CPUs have updated calc_load_tasks. | |
3289bdb4 PZ |
349 | * |
350 | * Called from the global timer code. | |
45ceebf7 PG |
351 | */ |
352 | void calc_global_load(unsigned long ticks) | |
353 | { | |
354 | long active, delta; | |
355 | ||
356 | if (time_before(jiffies, calc_load_update + 10)) | |
357 | return; | |
358 | ||
359 | /* | |
360 | * Fold the 'old' idle-delta to include all NO_HZ cpus. | |
361 | */ | |
362 | delta = calc_load_fold_idle(); | |
363 | if (delta) | |
364 | atomic_long_add(delta, &calc_load_tasks); | |
365 | ||
366 | active = atomic_long_read(&calc_load_tasks); | |
367 | active = active > 0 ? active * FIXED_1 : 0; | |
368 | ||
369 | avenrun[0] = calc_load(avenrun[0], EXP_1, active); | |
370 | avenrun[1] = calc_load(avenrun[1], EXP_5, active); | |
371 | avenrun[2] = calc_load(avenrun[2], EXP_15, active); | |
372 | ||
373 | calc_load_update += LOAD_FREQ; | |
374 | ||
375 | /* | |
376 | * In case we idled for multiple LOAD_FREQ intervals, catch up in bulk. | |
377 | */ | |
378 | calc_global_nohz(); | |
379 | } | |
380 | ||
381 | /* | |
3289bdb4 | 382 | * Called from scheduler_tick() to periodically update this CPU's |
45ceebf7 PG |
383 | * active count. |
384 | */ | |
3289bdb4 | 385 | void calc_global_load_tick(struct rq *this_rq) |
45ceebf7 PG |
386 | { |
387 | long delta; | |
388 | ||
389 | if (time_before(jiffies, this_rq->calc_load_update)) | |
390 | return; | |
391 | ||
d60585c5 | 392 | delta = calc_load_fold_active(this_rq, 0); |
45ceebf7 PG |
393 | if (delta) |
394 | atomic_long_add(delta, &calc_load_tasks); | |
395 | ||
396 | this_rq->calc_load_update += LOAD_FREQ; | |
397 | } |