xref: /linux/kernel/sched/loadavg.c (revision 0d3b051adbb72ed81956447d0d1e54d5943ee6f5)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * kernel/sched/loadavg.c
4  *
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.
8  */
9 #include "sched.h"
10 
11 /*
12  * Global load-average calculations
13  *
14  * We take a distributed and async approach to calculating the global load-avg
15  * in order to minimize overhead.
16  *
17  * The global load average is an exponentially decaying average of nr_running +
18  * nr_uninterruptible.
19  *
20  * Once every LOAD_FREQ:
21  *
22  *   nr_active = 0;
23  *   for_each_possible_cpu(cpu)
24  *	nr_active += cpu_of(cpu)->nr_running + cpu_of(cpu)->nr_uninterruptible;
25  *
26  *   avenrun[n] = avenrun[0] * exp_n + nr_active * (1 - exp_n)
27  *
28  * Due to a number of reasons the above turns in the mess below:
29  *
30  *  - for_each_possible_cpu() is prohibitively expensive on machines with
31  *    serious number of CPUs, therefore we need to take a distributed approach
32  *    to calculating nr_active.
33  *
34  *        \Sum_i x_i(t) = \Sum_i x_i(t) - x_i(t_0) | x_i(t_0) := 0
35  *                      = \Sum_i { \Sum_j=1 x_i(t_j) - x_i(t_j-1) }
36  *
37  *    So assuming nr_active := 0 when we start out -- true per definition, we
38  *    can simply take per-CPU deltas and fold those into a global accumulate
39  *    to obtain the same result. See calc_load_fold_active().
40  *
41  *    Furthermore, in order to avoid synchronizing all per-CPU delta folding
42  *    across the machine, we assume 10 ticks is sufficient time for every
43  *    CPU to have completed this task.
44  *
45  *    This places an upper-bound on the IRQ-off latency of the machine. Then
46  *    again, being late doesn't loose the delta, just wrecks the sample.
47  *
48  *  - cpu_rq()->nr_uninterruptible isn't accurately tracked per-CPU because
49  *    this would add another cross-CPU cacheline miss and atomic operation
50  *    to the wakeup path. Instead we increment on whatever CPU the task ran
51  *    when it went into uninterruptible state and decrement on whatever CPU
52  *    did the wakeup. This means that only the sum of nr_uninterruptible over
53  *    all CPUs yields the correct result.
54  *
55  *  This covers the NO_HZ=n code, for extra head-aches, see the comment below.
56  */
57 
58 /* Variables and functions for calc_load */
59 atomic_long_t calc_load_tasks;
60 unsigned long calc_load_update;
61 unsigned long avenrun[3];
62 EXPORT_SYMBOL(avenrun); /* should be removed */
63 
64 /**
65  * get_avenrun - get the load average array
66  * @loads:	pointer to dest load array
67  * @offset:	offset to add
68  * @shift:	shift count to shift the result left
69  *
70  * These values are estimates at best, so no need for locking.
71  */
72 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
73 {
74 	loads[0] = (avenrun[0] + offset) << shift;
75 	loads[1] = (avenrun[1] + offset) << shift;
76 	loads[2] = (avenrun[2] + offset) << shift;
77 }
78 
79 long calc_load_fold_active(struct rq *this_rq, long adjust)
80 {
81 	long nr_active, delta = 0;
82 
83 	nr_active = this_rq->nr_running - adjust;
84 	nr_active += (long)this_rq->nr_uninterruptible;
85 
86 	if (nr_active != this_rq->calc_load_active) {
87 		delta = nr_active - this_rq->calc_load_active;
88 		this_rq->calc_load_active = nr_active;
89 	}
90 
91 	return delta;
92 }
93 
94 /**
95  * fixed_power_int - compute: x^n, in O(log n) time
96  *
97  * @x:         base of the power
98  * @frac_bits: fractional bits of @x
99  * @n:         power to raise @x to.
100  *
101  * By exploiting the relation between the definition of the natural power
102  * function: x^n := x*x*...*x (x multiplied by itself for n times), and
103  * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
104  * (where: n_i \elem {0, 1}, the binary vector representing n),
105  * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
106  * of course trivially computable in O(log_2 n), the length of our binary
107  * vector.
108  */
109 static unsigned long
110 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
111 {
112 	unsigned long result = 1UL << frac_bits;
113 
114 	if (n) {
115 		for (;;) {
116 			if (n & 1) {
117 				result *= x;
118 				result += 1UL << (frac_bits - 1);
119 				result >>= frac_bits;
120 			}
121 			n >>= 1;
122 			if (!n)
123 				break;
124 			x *= x;
125 			x += 1UL << (frac_bits - 1);
126 			x >>= frac_bits;
127 		}
128 	}
129 
130 	return result;
131 }
132 
133 /*
134  * a1 = a0 * e + a * (1 - e)
135  *
136  * a2 = a1 * e + a * (1 - e)
137  *    = (a0 * e + a * (1 - e)) * e + a * (1 - e)
138  *    = a0 * e^2 + a * (1 - e) * (1 + e)
139  *
140  * a3 = a2 * e + a * (1 - e)
141  *    = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
142  *    = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
143  *
144  *  ...
145  *
146  * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
147  *    = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
148  *    = a0 * e^n + a * (1 - e^n)
149  *
150  * [1] application of the geometric series:
151  *
152  *              n         1 - x^(n+1)
153  *     S_n := \Sum x^i = -------------
154  *             i=0          1 - x
155  */
156 unsigned long
157 calc_load_n(unsigned long load, unsigned long exp,
158 	    unsigned long active, unsigned int n)
159 {
160 	return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
161 }
162 
163 #ifdef CONFIG_NO_HZ_COMMON
164 /*
165  * Handle NO_HZ for the global load-average.
166  *
167  * Since the above described distributed algorithm to compute the global
168  * load-average relies on per-CPU sampling from the tick, it is affected by
169  * NO_HZ.
170  *
171  * The basic idea is to fold the nr_active delta into a global NO_HZ-delta upon
172  * entering NO_HZ state such that we can include this as an 'extra' CPU delta
173  * when we read the global state.
174  *
175  * Obviously reality has to ruin such a delightfully simple scheme:
176  *
177  *  - When we go NO_HZ idle during the window, we can negate our sample
178  *    contribution, causing under-accounting.
179  *
180  *    We avoid this by keeping two NO_HZ-delta counters and flipping them
181  *    when the window starts, thus separating old and new NO_HZ load.
182  *
183  *    The only trick is the slight shift in index flip for read vs write.
184  *
185  *        0s            5s            10s           15s
186  *          +10           +10           +10           +10
187  *        |-|-----------|-|-----------|-|-----------|-|
188  *    r:0 0 1           1 0           0 1           1 0
189  *    w:0 1 1           0 0           1 1           0 0
190  *
191  *    This ensures we'll fold the old NO_HZ contribution in this window while
192  *    accumlating the new one.
193  *
194  *  - When we wake up from NO_HZ during the window, we push up our
195  *    contribution, since we effectively move our sample point to a known
196  *    busy state.
197  *
198  *    This is solved by pushing the window forward, and thus skipping the
199  *    sample, for this CPU (effectively using the NO_HZ-delta for this CPU which
200  *    was in effect at the time the window opened). This also solves the issue
201  *    of having to deal with a CPU having been in NO_HZ for multiple LOAD_FREQ
202  *    intervals.
203  *
204  * When making the ILB scale, we should try to pull this in as well.
205  */
206 static atomic_long_t calc_load_nohz[2];
207 static int calc_load_idx;
208 
209 static inline int calc_load_write_idx(void)
210 {
211 	int idx = calc_load_idx;
212 
213 	/*
214 	 * See calc_global_nohz(), if we observe the new index, we also
215 	 * need to observe the new update time.
216 	 */
217 	smp_rmb();
218 
219 	/*
220 	 * If the folding window started, make sure we start writing in the
221 	 * next NO_HZ-delta.
222 	 */
223 	if (!time_before(jiffies, READ_ONCE(calc_load_update)))
224 		idx++;
225 
226 	return idx & 1;
227 }
228 
229 static inline int calc_load_read_idx(void)
230 {
231 	return calc_load_idx & 1;
232 }
233 
234 static void calc_load_nohz_fold(struct rq *rq)
235 {
236 	long delta;
237 
238 	delta = calc_load_fold_active(rq, 0);
239 	if (delta) {
240 		int idx = calc_load_write_idx();
241 
242 		atomic_long_add(delta, &calc_load_nohz[idx]);
243 	}
244 }
245 
246 void calc_load_nohz_start(void)
247 {
248 	/*
249 	 * We're going into NO_HZ mode, if there's any pending delta, fold it
250 	 * into the pending NO_HZ delta.
251 	 */
252 	calc_load_nohz_fold(this_rq());
253 }
254 
255 /*
256  * Keep track of the load for NOHZ_FULL, must be called between
257  * calc_load_nohz_{start,stop}().
258  */
259 void calc_load_nohz_remote(struct rq *rq)
260 {
261 	calc_load_nohz_fold(rq);
262 }
263 
264 void calc_load_nohz_stop(void)
265 {
266 	struct rq *this_rq = this_rq();
267 
268 	/*
269 	 * If we're still before the pending sample window, we're done.
270 	 */
271 	this_rq->calc_load_update = READ_ONCE(calc_load_update);
272 	if (time_before(jiffies, this_rq->calc_load_update))
273 		return;
274 
275 	/*
276 	 * We woke inside or after the sample window, this means we're already
277 	 * accounted through the nohz accounting, so skip the entire deal and
278 	 * sync up for the next window.
279 	 */
280 	if (time_before(jiffies, this_rq->calc_load_update + 10))
281 		this_rq->calc_load_update += LOAD_FREQ;
282 }
283 
284 static long calc_load_nohz_read(void)
285 {
286 	int idx = calc_load_read_idx();
287 	long delta = 0;
288 
289 	if (atomic_long_read(&calc_load_nohz[idx]))
290 		delta = atomic_long_xchg(&calc_load_nohz[idx], 0);
291 
292 	return delta;
293 }
294 
295 /*
296  * NO_HZ can leave us missing all per-CPU ticks calling
297  * calc_load_fold_active(), but since a NO_HZ CPU folds its delta into
298  * calc_load_nohz per calc_load_nohz_start(), all we need to do is fold
299  * in the pending NO_HZ delta if our NO_HZ period crossed a load cycle boundary.
300  *
301  * Once we've updated the global active value, we need to apply the exponential
302  * weights adjusted to the number of cycles missed.
303  */
304 static void calc_global_nohz(void)
305 {
306 	unsigned long sample_window;
307 	long delta, active, n;
308 
309 	sample_window = READ_ONCE(calc_load_update);
310 	if (!time_before(jiffies, sample_window + 10)) {
311 		/*
312 		 * Catch-up, fold however many we are behind still
313 		 */
314 		delta = jiffies - sample_window - 10;
315 		n = 1 + (delta / LOAD_FREQ);
316 
317 		active = atomic_long_read(&calc_load_tasks);
318 		active = active > 0 ? active * FIXED_1 : 0;
319 
320 		avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
321 		avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
322 		avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
323 
324 		WRITE_ONCE(calc_load_update, sample_window + n * LOAD_FREQ);
325 	}
326 
327 	/*
328 	 * Flip the NO_HZ index...
329 	 *
330 	 * Make sure we first write the new time then flip the index, so that
331 	 * calc_load_write_idx() will see the new time when it reads the new
332 	 * index, this avoids a double flip messing things up.
333 	 */
334 	smp_wmb();
335 	calc_load_idx++;
336 }
337 #else /* !CONFIG_NO_HZ_COMMON */
338 
339 static inline long calc_load_nohz_read(void) { return 0; }
340 static inline void calc_global_nohz(void) { }
341 
342 #endif /* CONFIG_NO_HZ_COMMON */
343 
344 /*
345  * calc_load - update the avenrun load estimates 10 ticks after the
346  * CPUs have updated calc_load_tasks.
347  *
348  * Called from the global timer code.
349  */
350 void calc_global_load(void)
351 {
352 	unsigned long sample_window;
353 	long active, delta;
354 
355 	sample_window = READ_ONCE(calc_load_update);
356 	if (time_before(jiffies, sample_window + 10))
357 		return;
358 
359 	/*
360 	 * Fold the 'old' NO_HZ-delta to include all NO_HZ CPUs.
361 	 */
362 	delta = calc_load_nohz_read();
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 	WRITE_ONCE(calc_load_update, sample_window + LOAD_FREQ);
374 
375 	/*
376 	 * In case we went to NO_HZ for multiple LOAD_FREQ intervals
377 	 * catch up in bulk.
378 	 */
379 	calc_global_nohz();
380 }
381 
382 /*
383  * Called from scheduler_tick() to periodically update this CPU's
384  * active count.
385  */
386 void calc_global_load_tick(struct rq *this_rq)
387 {
388 	long delta;
389 
390 	if (time_before(jiffies, this_rq->calc_load_update))
391 		return;
392 
393 	delta  = calc_load_fold_active(this_rq, 0);
394 	if (delta)
395 		atomic_long_add(delta, &calc_load_tasks);
396 
397 	this_rq->calc_load_update += LOAD_FREQ;
398 }
399