Commit | Line | Data |
---|---|---|
2fb75e1b | 1 | // SPDX-License-Identifier: GPL-2.0 |
eb414681 JW |
2 | /* |
3 | * Pressure stall information for CPU, memory and IO | |
4 | * | |
5 | * Copyright (c) 2018 Facebook, Inc. | |
6 | * Author: Johannes Weiner <hannes@cmpxchg.org> | |
7 | * | |
0e94682b SB |
8 | * Polling support by Suren Baghdasaryan <surenb@google.com> |
9 | * Copyright (c) 2018 Google, Inc. | |
10 | * | |
eb414681 JW |
11 | * When CPU, memory and IO are contended, tasks experience delays that |
12 | * reduce throughput and introduce latencies into the workload. Memory | |
13 | * and IO contention, in addition, can cause a full loss of forward | |
14 | * progress in which the CPU goes idle. | |
15 | * | |
16 | * This code aggregates individual task delays into resource pressure | |
17 | * metrics that indicate problems with both workload health and | |
18 | * resource utilization. | |
19 | * | |
20 | * Model | |
21 | * | |
22 | * The time in which a task can execute on a CPU is our baseline for | |
23 | * productivity. Pressure expresses the amount of time in which this | |
24 | * potential cannot be realized due to resource contention. | |
25 | * | |
26 | * This concept of productivity has two components: the workload and | |
27 | * the CPU. To measure the impact of pressure on both, we define two | |
28 | * contention states for a resource: SOME and FULL. | |
29 | * | |
30 | * In the SOME state of a given resource, one or more tasks are | |
31 | * delayed on that resource. This affects the workload's ability to | |
32 | * perform work, but the CPU may still be executing other tasks. | |
33 | * | |
34 | * In the FULL state of a given resource, all non-idle tasks are | |
35 | * delayed on that resource such that nobody is advancing and the CPU | |
36 | * goes idle. This leaves both workload and CPU unproductive. | |
37 | * | |
e7fcd762 CZ |
38 | * Naturally, the FULL state doesn't exist for the CPU resource at the |
39 | * system level, but exist at the cgroup level, means all non-idle tasks | |
40 | * in a cgroup are delayed on the CPU resource which used by others outside | |
41 | * of the cgroup or throttled by the cgroup cpu.max configuration. | |
eb414681 JW |
42 | * |
43 | * SOME = nr_delayed_tasks != 0 | |
44 | * FULL = nr_delayed_tasks != 0 && nr_running_tasks == 0 | |
45 | * | |
46 | * The percentage of wallclock time spent in those compound stall | |
47 | * states gives pressure numbers between 0 and 100 for each resource, | |
48 | * where the SOME percentage indicates workload slowdowns and the FULL | |
49 | * percentage indicates reduced CPU utilization: | |
50 | * | |
51 | * %SOME = time(SOME) / period | |
52 | * %FULL = time(FULL) / period | |
53 | * | |
54 | * Multiple CPUs | |
55 | * | |
56 | * The more tasks and available CPUs there are, the more work can be | |
57 | * performed concurrently. This means that the potential that can go | |
58 | * unrealized due to resource contention *also* scales with non-idle | |
59 | * tasks and CPUs. | |
60 | * | |
61 | * Consider a scenario where 257 number crunching tasks are trying to | |
62 | * run concurrently on 256 CPUs. If we simply aggregated the task | |
63 | * states, we would have to conclude a CPU SOME pressure number of | |
64 | * 100%, since *somebody* is waiting on a runqueue at all | |
65 | * times. However, that is clearly not the amount of contention the | |
3b03706f | 66 | * workload is experiencing: only one out of 256 possible execution |
eb414681 JW |
67 | * threads will be contended at any given time, or about 0.4%. |
68 | * | |
69 | * Conversely, consider a scenario of 4 tasks and 4 CPUs where at any | |
70 | * given time *one* of the tasks is delayed due to a lack of memory. | |
71 | * Again, looking purely at the task state would yield a memory FULL | |
72 | * pressure number of 0%, since *somebody* is always making forward | |
73 | * progress. But again this wouldn't capture the amount of execution | |
74 | * potential lost, which is 1 out of 4 CPUs, or 25%. | |
75 | * | |
76 | * To calculate wasted potential (pressure) with multiple processors, | |
77 | * we have to base our calculation on the number of non-idle tasks in | |
78 | * conjunction with the number of available CPUs, which is the number | |
79 | * of potential execution threads. SOME becomes then the proportion of | |
3b03706f | 80 | * delayed tasks to possible threads, and FULL is the share of possible |
eb414681 JW |
81 | * threads that are unproductive due to delays: |
82 | * | |
83 | * threads = min(nr_nonidle_tasks, nr_cpus) | |
84 | * SOME = min(nr_delayed_tasks / threads, 1) | |
85 | * FULL = (threads - min(nr_running_tasks, threads)) / threads | |
86 | * | |
87 | * For the 257 number crunchers on 256 CPUs, this yields: | |
88 | * | |
89 | * threads = min(257, 256) | |
90 | * SOME = min(1 / 256, 1) = 0.4% | |
91 | * FULL = (256 - min(257, 256)) / 256 = 0% | |
92 | * | |
93 | * For the 1 out of 4 memory-delayed tasks, this yields: | |
94 | * | |
95 | * threads = min(4, 4) | |
96 | * SOME = min(1 / 4, 1) = 25% | |
97 | * FULL = (4 - min(3, 4)) / 4 = 25% | |
98 | * | |
99 | * [ Substitute nr_cpus with 1, and you can see that it's a natural | |
100 | * extension of the single-CPU model. ] | |
101 | * | |
102 | * Implementation | |
103 | * | |
104 | * To assess the precise time spent in each such state, we would have | |
105 | * to freeze the system on task changes and start/stop the state | |
106 | * clocks accordingly. Obviously that doesn't scale in practice. | |
107 | * | |
108 | * Because the scheduler aims to distribute the compute load evenly | |
109 | * among the available CPUs, we can track task state locally to each | |
110 | * CPU and, at much lower frequency, extrapolate the global state for | |
111 | * the cumulative stall times and the running averages. | |
112 | * | |
113 | * For each runqueue, we track: | |
114 | * | |
115 | * tSOME[cpu] = time(nr_delayed_tasks[cpu] != 0) | |
116 | * tFULL[cpu] = time(nr_delayed_tasks[cpu] && !nr_running_tasks[cpu]) | |
117 | * tNONIDLE[cpu] = time(nr_nonidle_tasks[cpu] != 0) | |
118 | * | |
119 | * and then periodically aggregate: | |
120 | * | |
121 | * tNONIDLE = sum(tNONIDLE[i]) | |
122 | * | |
123 | * tSOME = sum(tSOME[i] * tNONIDLE[i]) / tNONIDLE | |
124 | * tFULL = sum(tFULL[i] * tNONIDLE[i]) / tNONIDLE | |
125 | * | |
126 | * %SOME = tSOME / period | |
127 | * %FULL = tFULL / period | |
128 | * | |
129 | * This gives us an approximation of pressure that is practical | |
130 | * cost-wise, yet way more sensitive and accurate than periodic | |
131 | * sampling of the aggregate task states would be. | |
132 | */ | |
133 | ||
1b69ac6b | 134 | #include "../workqueue_internal.h" |
eb414681 JW |
135 | #include <linux/sched/loadavg.h> |
136 | #include <linux/seq_file.h> | |
137 | #include <linux/proc_fs.h> | |
138 | #include <linux/seqlock.h> | |
0e94682b | 139 | #include <linux/uaccess.h> |
eb414681 JW |
140 | #include <linux/cgroup.h> |
141 | #include <linux/module.h> | |
142 | #include <linux/sched.h> | |
0e94682b SB |
143 | #include <linux/ctype.h> |
144 | #include <linux/file.h> | |
145 | #include <linux/poll.h> | |
eb414681 JW |
146 | #include <linux/psi.h> |
147 | #include "sched.h" | |
148 | ||
149 | static int psi_bug __read_mostly; | |
150 | ||
e0c27447 | 151 | DEFINE_STATIC_KEY_FALSE(psi_disabled); |
3958e2d0 | 152 | DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled); |
e0c27447 JW |
153 | |
154 | #ifdef CONFIG_PSI_DEFAULT_DISABLED | |
9289c5e6 | 155 | static bool psi_enable; |
e0c27447 | 156 | #else |
9289c5e6 | 157 | static bool psi_enable = true; |
e0c27447 JW |
158 | #endif |
159 | static int __init setup_psi(char *str) | |
160 | { | |
161 | return kstrtobool(str, &psi_enable) == 0; | |
162 | } | |
163 | __setup("psi=", setup_psi); | |
eb414681 JW |
164 | |
165 | /* Running averages - we need to be higher-res than loadavg */ | |
166 | #define PSI_FREQ (2*HZ+1) /* 2 sec intervals */ | |
167 | #define EXP_10s 1677 /* 1/exp(2s/10s) as fixed-point */ | |
168 | #define EXP_60s 1981 /* 1/exp(2s/60s) */ | |
169 | #define EXP_300s 2034 /* 1/exp(2s/300s) */ | |
170 | ||
0e94682b SB |
171 | /* PSI trigger definitions */ |
172 | #define WINDOW_MIN_US 500000 /* Min window size is 500ms */ | |
173 | #define WINDOW_MAX_US 10000000 /* Max window size is 10s */ | |
174 | #define UPDATES_PER_WINDOW 10 /* 10 updates per window */ | |
175 | ||
eb414681 JW |
176 | /* Sampling frequency in nanoseconds */ |
177 | static u64 psi_period __read_mostly; | |
178 | ||
179 | /* System-level pressure and stall tracking */ | |
180 | static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu); | |
df5ba5be | 181 | struct psi_group psi_system = { |
eb414681 JW |
182 | .pcpu = &system_group_pcpu, |
183 | }; | |
184 | ||
bcc78db6 | 185 | static void psi_avgs_work(struct work_struct *work); |
eb414681 | 186 | |
8f91efd8 ZH |
187 | static void poll_timer_fn(struct timer_list *t); |
188 | ||
eb414681 JW |
189 | static void group_init(struct psi_group *group) |
190 | { | |
191 | int cpu; | |
192 | ||
193 | for_each_possible_cpu(cpu) | |
194 | seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq); | |
3dfbe25c JW |
195 | group->avg_last_update = sched_clock(); |
196 | group->avg_next_update = group->avg_last_update + psi_period; | |
bcc78db6 SB |
197 | INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work); |
198 | mutex_init(&group->avgs_lock); | |
0e94682b | 199 | /* Init trigger-related members */ |
0e94682b SB |
200 | mutex_init(&group->trigger_lock); |
201 | INIT_LIST_HEAD(&group->triggers); | |
202 | memset(group->nr_triggers, 0, sizeof(group->nr_triggers)); | |
203 | group->poll_states = 0; | |
204 | group->poll_min_period = U32_MAX; | |
205 | memset(group->polling_total, 0, sizeof(group->polling_total)); | |
206 | group->polling_next_update = ULLONG_MAX; | |
207 | group->polling_until = 0; | |
8f91efd8 ZH |
208 | init_waitqueue_head(&group->poll_wait); |
209 | timer_setup(&group->poll_timer, poll_timer_fn, 0); | |
461daba0 | 210 | rcu_assign_pointer(group->poll_task, NULL); |
eb414681 JW |
211 | } |
212 | ||
213 | void __init psi_init(void) | |
214 | { | |
e0c27447 JW |
215 | if (!psi_enable) { |
216 | static_branch_enable(&psi_disabled); | |
eb414681 | 217 | return; |
e0c27447 | 218 | } |
eb414681 | 219 | |
3958e2d0 SB |
220 | if (!cgroup_psi_enabled()) |
221 | static_branch_disable(&psi_cgroups_enabled); | |
222 | ||
eb414681 JW |
223 | psi_period = jiffies_to_nsecs(PSI_FREQ); |
224 | group_init(&psi_system); | |
225 | } | |
226 | ||
227 | static bool test_state(unsigned int *tasks, enum psi_states state) | |
228 | { | |
229 | switch (state) { | |
230 | case PSI_IO_SOME: | |
fddc8bab | 231 | return unlikely(tasks[NR_IOWAIT]); |
eb414681 | 232 | case PSI_IO_FULL: |
fddc8bab | 233 | return unlikely(tasks[NR_IOWAIT] && !tasks[NR_RUNNING]); |
eb414681 | 234 | case PSI_MEM_SOME: |
fddc8bab | 235 | return unlikely(tasks[NR_MEMSTALL]); |
eb414681 | 236 | case PSI_MEM_FULL: |
fddc8bab | 237 | return unlikely(tasks[NR_MEMSTALL] && !tasks[NR_RUNNING]); |
eb414681 | 238 | case PSI_CPU_SOME: |
fddc8bab | 239 | return unlikely(tasks[NR_RUNNING] > tasks[NR_ONCPU]); |
e7fcd762 | 240 | case PSI_CPU_FULL: |
fddc8bab | 241 | return unlikely(tasks[NR_RUNNING] && !tasks[NR_ONCPU]); |
eb414681 JW |
242 | case PSI_NONIDLE: |
243 | return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] || | |
244 | tasks[NR_RUNNING]; | |
245 | default: | |
246 | return false; | |
247 | } | |
248 | } | |
249 | ||
0e94682b SB |
250 | static void get_recent_times(struct psi_group *group, int cpu, |
251 | enum psi_aggregators aggregator, u32 *times, | |
333f3017 | 252 | u32 *pchanged_states) |
eb414681 JW |
253 | { |
254 | struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu); | |
eb414681 | 255 | u64 now, state_start; |
33b2d630 | 256 | enum psi_states s; |
eb414681 | 257 | unsigned int seq; |
33b2d630 | 258 | u32 state_mask; |
eb414681 | 259 | |
333f3017 SB |
260 | *pchanged_states = 0; |
261 | ||
eb414681 JW |
262 | /* Snapshot a coherent view of the CPU state */ |
263 | do { | |
264 | seq = read_seqcount_begin(&groupc->seq); | |
265 | now = cpu_clock(cpu); | |
266 | memcpy(times, groupc->times, sizeof(groupc->times)); | |
33b2d630 | 267 | state_mask = groupc->state_mask; |
eb414681 JW |
268 | state_start = groupc->state_start; |
269 | } while (read_seqcount_retry(&groupc->seq, seq)); | |
270 | ||
271 | /* Calculate state time deltas against the previous snapshot */ | |
272 | for (s = 0; s < NR_PSI_STATES; s++) { | |
273 | u32 delta; | |
274 | /* | |
275 | * In addition to already concluded states, we also | |
276 | * incorporate currently active states on the CPU, | |
277 | * since states may last for many sampling periods. | |
278 | * | |
279 | * This way we keep our delta sampling buckets small | |
280 | * (u32) and our reported pressure close to what's | |
281 | * actually happening. | |
282 | */ | |
33b2d630 | 283 | if (state_mask & (1 << s)) |
eb414681 JW |
284 | times[s] += now - state_start; |
285 | ||
0e94682b SB |
286 | delta = times[s] - groupc->times_prev[aggregator][s]; |
287 | groupc->times_prev[aggregator][s] = times[s]; | |
eb414681 JW |
288 | |
289 | times[s] = delta; | |
333f3017 SB |
290 | if (delta) |
291 | *pchanged_states |= (1 << s); | |
eb414681 JW |
292 | } |
293 | } | |
294 | ||
295 | static void calc_avgs(unsigned long avg[3], int missed_periods, | |
296 | u64 time, u64 period) | |
297 | { | |
298 | unsigned long pct; | |
299 | ||
300 | /* Fill in zeroes for periods of no activity */ | |
301 | if (missed_periods) { | |
302 | avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods); | |
303 | avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods); | |
304 | avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods); | |
305 | } | |
306 | ||
307 | /* Sample the most recent active period */ | |
308 | pct = div_u64(time * 100, period); | |
309 | pct *= FIXED_1; | |
310 | avg[0] = calc_load(avg[0], EXP_10s, pct); | |
311 | avg[1] = calc_load(avg[1], EXP_60s, pct); | |
312 | avg[2] = calc_load(avg[2], EXP_300s, pct); | |
313 | } | |
314 | ||
0e94682b SB |
315 | static void collect_percpu_times(struct psi_group *group, |
316 | enum psi_aggregators aggregator, | |
317 | u32 *pchanged_states) | |
eb414681 JW |
318 | { |
319 | u64 deltas[NR_PSI_STATES - 1] = { 0, }; | |
eb414681 | 320 | unsigned long nonidle_total = 0; |
333f3017 | 321 | u32 changed_states = 0; |
eb414681 JW |
322 | int cpu; |
323 | int s; | |
324 | ||
eb414681 JW |
325 | /* |
326 | * Collect the per-cpu time buckets and average them into a | |
327 | * single time sample that is normalized to wallclock time. | |
328 | * | |
329 | * For averaging, each CPU is weighted by its non-idle time in | |
330 | * the sampling period. This eliminates artifacts from uneven | |
331 | * loading, or even entirely idle CPUs. | |
332 | */ | |
333 | for_each_possible_cpu(cpu) { | |
334 | u32 times[NR_PSI_STATES]; | |
335 | u32 nonidle; | |
333f3017 | 336 | u32 cpu_changed_states; |
eb414681 | 337 | |
0e94682b | 338 | get_recent_times(group, cpu, aggregator, times, |
333f3017 SB |
339 | &cpu_changed_states); |
340 | changed_states |= cpu_changed_states; | |
eb414681 JW |
341 | |
342 | nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]); | |
343 | nonidle_total += nonidle; | |
344 | ||
345 | for (s = 0; s < PSI_NONIDLE; s++) | |
346 | deltas[s] += (u64)times[s] * nonidle; | |
347 | } | |
348 | ||
349 | /* | |
350 | * Integrate the sample into the running statistics that are | |
351 | * reported to userspace: the cumulative stall times and the | |
352 | * decaying averages. | |
353 | * | |
354 | * Pressure percentages are sampled at PSI_FREQ. We might be | |
355 | * called more often when the user polls more frequently than | |
356 | * that; we might be called less often when there is no task | |
357 | * activity, thus no data, and clock ticks are sporadic. The | |
358 | * below handles both. | |
359 | */ | |
360 | ||
361 | /* total= */ | |
362 | for (s = 0; s < NR_PSI_STATES - 1; s++) | |
0e94682b SB |
363 | group->total[aggregator][s] += |
364 | div_u64(deltas[s], max(nonidle_total, 1UL)); | |
eb414681 | 365 | |
333f3017 SB |
366 | if (pchanged_states) |
367 | *pchanged_states = changed_states; | |
7fc70a39 SB |
368 | } |
369 | ||
370 | static u64 update_averages(struct psi_group *group, u64 now) | |
371 | { | |
372 | unsigned long missed_periods = 0; | |
373 | u64 expires, period; | |
374 | u64 avg_next_update; | |
375 | int s; | |
376 | ||
eb414681 | 377 | /* avgX= */ |
bcc78db6 | 378 | expires = group->avg_next_update; |
4e37504d | 379 | if (now - expires >= psi_period) |
eb414681 JW |
380 | missed_periods = div_u64(now - expires, psi_period); |
381 | ||
382 | /* | |
383 | * The periodic clock tick can get delayed for various | |
384 | * reasons, especially on loaded systems. To avoid clock | |
385 | * drift, we schedule the clock in fixed psi_period intervals. | |
386 | * But the deltas we sample out of the per-cpu buckets above | |
387 | * are based on the actual time elapsing between clock ticks. | |
388 | */ | |
7fc70a39 | 389 | avg_next_update = expires + ((1 + missed_periods) * psi_period); |
bcc78db6 SB |
390 | period = now - (group->avg_last_update + (missed_periods * psi_period)); |
391 | group->avg_last_update = now; | |
eb414681 JW |
392 | |
393 | for (s = 0; s < NR_PSI_STATES - 1; s++) { | |
394 | u32 sample; | |
395 | ||
0e94682b | 396 | sample = group->total[PSI_AVGS][s] - group->avg_total[s]; |
eb414681 JW |
397 | /* |
398 | * Due to the lockless sampling of the time buckets, | |
399 | * recorded time deltas can slip into the next period, | |
400 | * which under full pressure can result in samples in | |
401 | * excess of the period length. | |
402 | * | |
403 | * We don't want to report non-sensical pressures in | |
404 | * excess of 100%, nor do we want to drop such events | |
405 | * on the floor. Instead we punt any overage into the | |
406 | * future until pressure subsides. By doing this we | |
407 | * don't underreport the occurring pressure curve, we | |
408 | * just report it delayed by one period length. | |
409 | * | |
410 | * The error isn't cumulative. As soon as another | |
411 | * delta slips from a period P to P+1, by definition | |
412 | * it frees up its time T in P. | |
413 | */ | |
414 | if (sample > period) | |
415 | sample = period; | |
bcc78db6 | 416 | group->avg_total[s] += sample; |
eb414681 JW |
417 | calc_avgs(group->avg[s], missed_periods, sample, period); |
418 | } | |
7fc70a39 SB |
419 | |
420 | return avg_next_update; | |
eb414681 JW |
421 | } |
422 | ||
bcc78db6 | 423 | static void psi_avgs_work(struct work_struct *work) |
eb414681 JW |
424 | { |
425 | struct delayed_work *dwork; | |
426 | struct psi_group *group; | |
333f3017 | 427 | u32 changed_states; |
eb414681 | 428 | bool nonidle; |
7fc70a39 | 429 | u64 now; |
eb414681 JW |
430 | |
431 | dwork = to_delayed_work(work); | |
bcc78db6 | 432 | group = container_of(dwork, struct psi_group, avgs_work); |
eb414681 | 433 | |
7fc70a39 SB |
434 | mutex_lock(&group->avgs_lock); |
435 | ||
436 | now = sched_clock(); | |
437 | ||
0e94682b | 438 | collect_percpu_times(group, PSI_AVGS, &changed_states); |
333f3017 | 439 | nonidle = changed_states & (1 << PSI_NONIDLE); |
eb414681 JW |
440 | /* |
441 | * If there is task activity, periodically fold the per-cpu | |
442 | * times and feed samples into the running averages. If things | |
443 | * are idle and there is no data to process, stop the clock. | |
444 | * Once restarted, we'll catch up the running averages in one | |
445 | * go - see calc_avgs() and missed_periods. | |
446 | */ | |
7fc70a39 SB |
447 | if (now >= group->avg_next_update) |
448 | group->avg_next_update = update_averages(group, now); | |
eb414681 JW |
449 | |
450 | if (nonidle) { | |
7fc70a39 SB |
451 | schedule_delayed_work(dwork, nsecs_to_jiffies( |
452 | group->avg_next_update - now) + 1); | |
eb414681 | 453 | } |
7fc70a39 SB |
454 | |
455 | mutex_unlock(&group->avgs_lock); | |
eb414681 JW |
456 | } |
457 | ||
3b03706f | 458 | /* Trigger tracking window manipulations */ |
0e94682b SB |
459 | static void window_reset(struct psi_window *win, u64 now, u64 value, |
460 | u64 prev_growth) | |
461 | { | |
462 | win->start_time = now; | |
463 | win->start_value = value; | |
464 | win->prev_growth = prev_growth; | |
465 | } | |
466 | ||
467 | /* | |
468 | * PSI growth tracking window update and growth calculation routine. | |
469 | * | |
470 | * This approximates a sliding tracking window by interpolating | |
471 | * partially elapsed windows using historical growth data from the | |
472 | * previous intervals. This minimizes memory requirements (by not storing | |
473 | * all the intermediate values in the previous window) and simplifies | |
474 | * the calculations. It works well because PSI signal changes only in | |
475 | * positive direction and over relatively small window sizes the growth | |
476 | * is close to linear. | |
477 | */ | |
478 | static u64 window_update(struct psi_window *win, u64 now, u64 value) | |
479 | { | |
480 | u64 elapsed; | |
481 | u64 growth; | |
482 | ||
483 | elapsed = now - win->start_time; | |
484 | growth = value - win->start_value; | |
485 | /* | |
486 | * After each tracking window passes win->start_value and | |
487 | * win->start_time get reset and win->prev_growth stores | |
488 | * the average per-window growth of the previous window. | |
489 | * win->prev_growth is then used to interpolate additional | |
490 | * growth from the previous window assuming it was linear. | |
491 | */ | |
492 | if (elapsed > win->size) | |
493 | window_reset(win, now, value, growth); | |
494 | else { | |
495 | u32 remaining; | |
496 | ||
497 | remaining = win->size - elapsed; | |
c3466952 | 498 | growth += div64_u64(win->prev_growth * remaining, win->size); |
0e94682b SB |
499 | } |
500 | ||
501 | return growth; | |
502 | } | |
503 | ||
504 | static void init_triggers(struct psi_group *group, u64 now) | |
505 | { | |
506 | struct psi_trigger *t; | |
507 | ||
508 | list_for_each_entry(t, &group->triggers, node) | |
509 | window_reset(&t->win, now, | |
510 | group->total[PSI_POLL][t->state], 0); | |
511 | memcpy(group->polling_total, group->total[PSI_POLL], | |
512 | sizeof(group->polling_total)); | |
513 | group->polling_next_update = now + group->poll_min_period; | |
514 | } | |
515 | ||
516 | static u64 update_triggers(struct psi_group *group, u64 now) | |
517 | { | |
518 | struct psi_trigger *t; | |
519 | bool new_stall = false; | |
520 | u64 *total = group->total[PSI_POLL]; | |
521 | ||
522 | /* | |
523 | * On subsequent updates, calculate growth deltas and let | |
524 | * watchers know when their specified thresholds are exceeded. | |
525 | */ | |
526 | list_for_each_entry(t, &group->triggers, node) { | |
527 | u64 growth; | |
528 | ||
529 | /* Check for stall activity */ | |
530 | if (group->polling_total[t->state] == total[t->state]) | |
531 | continue; | |
532 | ||
533 | /* | |
534 | * Multiple triggers might be looking at the same state, | |
535 | * remember to update group->polling_total[] once we've | |
536 | * been through all of them. Also remember to extend the | |
537 | * polling time if we see new stall activity. | |
538 | */ | |
539 | new_stall = true; | |
540 | ||
541 | /* Calculate growth since last update */ | |
542 | growth = window_update(&t->win, now, total[t->state]); | |
543 | if (growth < t->threshold) | |
544 | continue; | |
545 | ||
546 | /* Limit event signaling to once per window */ | |
547 | if (now < t->last_event_time + t->win.size) | |
548 | continue; | |
549 | ||
550 | /* Generate an event */ | |
551 | if (cmpxchg(&t->event, 0, 1) == 0) | |
552 | wake_up_interruptible(&t->event_wait); | |
553 | t->last_event_time = now; | |
554 | } | |
555 | ||
556 | if (new_stall) | |
557 | memcpy(group->polling_total, total, | |
558 | sizeof(group->polling_total)); | |
559 | ||
560 | return now + group->poll_min_period; | |
561 | } | |
562 | ||
461daba0 | 563 | /* Schedule polling if it's not already scheduled. */ |
0e94682b SB |
564 | static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay) |
565 | { | |
461daba0 | 566 | struct task_struct *task; |
0e94682b | 567 | |
461daba0 SB |
568 | /* |
569 | * Do not reschedule if already scheduled. | |
570 | * Possible race with a timer scheduled after this check but before | |
571 | * mod_timer below can be tolerated because group->polling_next_update | |
572 | * will keep updates on schedule. | |
573 | */ | |
574 | if (timer_pending(&group->poll_timer)) | |
0e94682b SB |
575 | return; |
576 | ||
577 | rcu_read_lock(); | |
578 | ||
461daba0 | 579 | task = rcu_dereference(group->poll_task); |
0e94682b SB |
580 | /* |
581 | * kworker might be NULL in case psi_trigger_destroy races with | |
582 | * psi_task_change (hotpath) which can't use locks | |
583 | */ | |
461daba0 SB |
584 | if (likely(task)) |
585 | mod_timer(&group->poll_timer, jiffies + delay); | |
0e94682b SB |
586 | |
587 | rcu_read_unlock(); | |
588 | } | |
589 | ||
461daba0 | 590 | static void psi_poll_work(struct psi_group *group) |
0e94682b | 591 | { |
0e94682b SB |
592 | u32 changed_states; |
593 | u64 now; | |
594 | ||
0e94682b SB |
595 | mutex_lock(&group->trigger_lock); |
596 | ||
597 | now = sched_clock(); | |
598 | ||
599 | collect_percpu_times(group, PSI_POLL, &changed_states); | |
600 | ||
601 | if (changed_states & group->poll_states) { | |
602 | /* Initialize trigger windows when entering polling mode */ | |
603 | if (now > group->polling_until) | |
604 | init_triggers(group, now); | |
605 | ||
606 | /* | |
607 | * Keep the monitor active for at least the duration of the | |
608 | * minimum tracking window as long as monitor states are | |
609 | * changing. | |
610 | */ | |
611 | group->polling_until = now + | |
612 | group->poll_min_period * UPDATES_PER_WINDOW; | |
613 | } | |
614 | ||
615 | if (now > group->polling_until) { | |
616 | group->polling_next_update = ULLONG_MAX; | |
617 | goto out; | |
618 | } | |
619 | ||
620 | if (now >= group->polling_next_update) | |
621 | group->polling_next_update = update_triggers(group, now); | |
622 | ||
623 | psi_schedule_poll_work(group, | |
624 | nsecs_to_jiffies(group->polling_next_update - now) + 1); | |
625 | ||
626 | out: | |
627 | mutex_unlock(&group->trigger_lock); | |
628 | } | |
629 | ||
461daba0 SB |
630 | static int psi_poll_worker(void *data) |
631 | { | |
632 | struct psi_group *group = (struct psi_group *)data; | |
461daba0 | 633 | |
2cca5426 | 634 | sched_set_fifo_low(current); |
461daba0 SB |
635 | |
636 | while (true) { | |
637 | wait_event_interruptible(group->poll_wait, | |
638 | atomic_cmpxchg(&group->poll_wakeup, 1, 0) || | |
639 | kthread_should_stop()); | |
640 | if (kthread_should_stop()) | |
641 | break; | |
642 | ||
643 | psi_poll_work(group); | |
644 | } | |
645 | return 0; | |
646 | } | |
647 | ||
648 | static void poll_timer_fn(struct timer_list *t) | |
649 | { | |
650 | struct psi_group *group = from_timer(group, t, poll_timer); | |
651 | ||
652 | atomic_set(&group->poll_wakeup, 1); | |
653 | wake_up_interruptible(&group->poll_wait); | |
654 | } | |
655 | ||
df774306 | 656 | static void record_times(struct psi_group_cpu *groupc, u64 now) |
eb414681 JW |
657 | { |
658 | u32 delta; | |
eb414681 | 659 | |
eb414681 JW |
660 | delta = now - groupc->state_start; |
661 | groupc->state_start = now; | |
662 | ||
33b2d630 | 663 | if (groupc->state_mask & (1 << PSI_IO_SOME)) { |
eb414681 | 664 | groupc->times[PSI_IO_SOME] += delta; |
33b2d630 | 665 | if (groupc->state_mask & (1 << PSI_IO_FULL)) |
eb414681 JW |
666 | groupc->times[PSI_IO_FULL] += delta; |
667 | } | |
668 | ||
33b2d630 | 669 | if (groupc->state_mask & (1 << PSI_MEM_SOME)) { |
eb414681 | 670 | groupc->times[PSI_MEM_SOME] += delta; |
33b2d630 | 671 | if (groupc->state_mask & (1 << PSI_MEM_FULL)) |
eb414681 | 672 | groupc->times[PSI_MEM_FULL] += delta; |
eb414681 JW |
673 | } |
674 | ||
e7fcd762 | 675 | if (groupc->state_mask & (1 << PSI_CPU_SOME)) { |
eb414681 | 676 | groupc->times[PSI_CPU_SOME] += delta; |
e7fcd762 CZ |
677 | if (groupc->state_mask & (1 << PSI_CPU_FULL)) |
678 | groupc->times[PSI_CPU_FULL] += delta; | |
679 | } | |
eb414681 | 680 | |
33b2d630 | 681 | if (groupc->state_mask & (1 << PSI_NONIDLE)) |
eb414681 JW |
682 | groupc->times[PSI_NONIDLE] += delta; |
683 | } | |
684 | ||
36b238d5 | 685 | static void psi_group_change(struct psi_group *group, int cpu, |
df774306 | 686 | unsigned int clear, unsigned int set, u64 now, |
36b238d5 | 687 | bool wake_clock) |
eb414681 JW |
688 | { |
689 | struct psi_group_cpu *groupc; | |
36b238d5 | 690 | u32 state_mask = 0; |
eb414681 | 691 | unsigned int t, m; |
33b2d630 | 692 | enum psi_states s; |
eb414681 JW |
693 | |
694 | groupc = per_cpu_ptr(group->pcpu, cpu); | |
695 | ||
696 | /* | |
697 | * First we assess the aggregate resource states this CPU's | |
698 | * tasks have been in since the last change, and account any | |
699 | * SOME and FULL time these may have resulted in. | |
700 | * | |
701 | * Then we update the task counts according to the state | |
702 | * change requested through the @clear and @set bits. | |
703 | */ | |
704 | write_seqcount_begin(&groupc->seq); | |
705 | ||
df774306 | 706 | record_times(groupc, now); |
eb414681 JW |
707 | |
708 | for (t = 0, m = clear; m; m &= ~(1 << t), t++) { | |
709 | if (!(m & (1 << t))) | |
710 | continue; | |
9d10a13d CTR |
711 | if (groupc->tasks[t]) { |
712 | groupc->tasks[t]--; | |
713 | } else if (!psi_bug) { | |
b05e75d6 | 714 | printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u] clear=%x set=%x\n", |
eb414681 JW |
715 | cpu, t, groupc->tasks[0], |
716 | groupc->tasks[1], groupc->tasks[2], | |
b05e75d6 | 717 | groupc->tasks[3], clear, set); |
eb414681 JW |
718 | psi_bug = 1; |
719 | } | |
eb414681 JW |
720 | } |
721 | ||
722 | for (t = 0; set; set &= ~(1 << t), t++) | |
723 | if (set & (1 << t)) | |
724 | groupc->tasks[t]++; | |
725 | ||
33b2d630 SB |
726 | /* Calculate state mask representing active states */ |
727 | for (s = 0; s < NR_PSI_STATES; s++) { | |
728 | if (test_state(groupc->tasks, s)) | |
729 | state_mask |= (1 << s); | |
730 | } | |
7fae6c81 CZ |
731 | |
732 | /* | |
733 | * Since we care about lost potential, a memstall is FULL | |
734 | * when there are no other working tasks, but also when | |
735 | * the CPU is actively reclaiming and nothing productive | |
736 | * could run even if it were runnable. So when the current | |
737 | * task in a cgroup is in_memstall, the corresponding groupc | |
738 | * on that cpu is in PSI_MEM_FULL state. | |
739 | */ | |
fddc8bab | 740 | if (unlikely(groupc->tasks[NR_ONCPU] && cpu_curr(cpu)->in_memstall)) |
7fae6c81 CZ |
741 | state_mask |= (1 << PSI_MEM_FULL); |
742 | ||
33b2d630 SB |
743 | groupc->state_mask = state_mask; |
744 | ||
eb414681 | 745 | write_seqcount_end(&groupc->seq); |
0e94682b | 746 | |
36b238d5 JW |
747 | if (state_mask & group->poll_states) |
748 | psi_schedule_poll_work(group, 1); | |
749 | ||
750 | if (wake_clock && !delayed_work_pending(&group->avgs_work)) | |
751 | schedule_delayed_work(&group->avgs_work, PSI_FREQ); | |
eb414681 JW |
752 | } |
753 | ||
2ce7135a JW |
754 | static struct psi_group *iterate_groups(struct task_struct *task, void **iter) |
755 | { | |
3958e2d0 SB |
756 | if (*iter == &psi_system) |
757 | return NULL; | |
758 | ||
2ce7135a | 759 | #ifdef CONFIG_CGROUPS |
3958e2d0 SB |
760 | if (static_branch_likely(&psi_cgroups_enabled)) { |
761 | struct cgroup *cgroup = NULL; | |
2ce7135a | 762 | |
3958e2d0 SB |
763 | if (!*iter) |
764 | cgroup = task->cgroups->dfl_cgrp; | |
765 | else | |
766 | cgroup = cgroup_parent(*iter); | |
2ce7135a | 767 | |
3958e2d0 SB |
768 | if (cgroup && cgroup_parent(cgroup)) { |
769 | *iter = cgroup; | |
770 | return cgroup_psi(cgroup); | |
771 | } | |
2ce7135a | 772 | } |
2ce7135a JW |
773 | #endif |
774 | *iter = &psi_system; | |
775 | return &psi_system; | |
776 | } | |
777 | ||
36b238d5 | 778 | static void psi_flags_change(struct task_struct *task, int clear, int set) |
eb414681 | 779 | { |
eb414681 JW |
780 | if (((task->psi_flags & set) || |
781 | (task->psi_flags & clear) != clear) && | |
782 | !psi_bug) { | |
783 | printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n", | |
36b238d5 | 784 | task->pid, task->comm, task_cpu(task), |
eb414681 JW |
785 | task->psi_flags, clear, set); |
786 | psi_bug = 1; | |
787 | } | |
788 | ||
789 | task->psi_flags &= ~clear; | |
790 | task->psi_flags |= set; | |
36b238d5 JW |
791 | } |
792 | ||
793 | void psi_task_change(struct task_struct *task, int clear, int set) | |
794 | { | |
795 | int cpu = task_cpu(task); | |
796 | struct psi_group *group; | |
797 | bool wake_clock = true; | |
798 | void *iter = NULL; | |
df774306 | 799 | u64 now; |
36b238d5 JW |
800 | |
801 | if (!task->pid) | |
802 | return; | |
803 | ||
804 | psi_flags_change(task, clear, set); | |
eb414681 | 805 | |
df774306 | 806 | now = cpu_clock(cpu); |
1b69ac6b JW |
807 | /* |
808 | * Periodic aggregation shuts off if there is a period of no | |
809 | * task changes, so we wake it back up if necessary. However, | |
810 | * don't do this if the task change is the aggregation worker | |
811 | * itself going to sleep, or we'll ping-pong forever. | |
812 | */ | |
813 | if (unlikely((clear & TSK_RUNNING) && | |
814 | (task->flags & PF_WQ_WORKER) && | |
bcc78db6 | 815 | wq_worker_last_func(task) == psi_avgs_work)) |
1b69ac6b JW |
816 | wake_clock = false; |
817 | ||
36b238d5 | 818 | while ((group = iterate_groups(task, &iter))) |
df774306 | 819 | psi_group_change(group, cpu, clear, set, now, wake_clock); |
36b238d5 JW |
820 | } |
821 | ||
822 | void psi_task_switch(struct task_struct *prev, struct task_struct *next, | |
823 | bool sleep) | |
824 | { | |
825 | struct psi_group *group, *common = NULL; | |
826 | int cpu = task_cpu(prev); | |
827 | void *iter; | |
df774306 | 828 | u64 now = cpu_clock(cpu); |
36b238d5 JW |
829 | |
830 | if (next->pid) { | |
7fae6c81 CZ |
831 | bool identical_state; |
832 | ||
36b238d5 JW |
833 | psi_flags_change(next, 0, TSK_ONCPU); |
834 | /* | |
7fae6c81 CZ |
835 | * When switching between tasks that have an identical |
836 | * runtime state, the cgroup that contains both tasks | |
7fae6c81 CZ |
837 | * we reach the first common ancestor. Iterate @next's |
838 | * ancestors only until we encounter @prev's ONCPU. | |
36b238d5 | 839 | */ |
7fae6c81 | 840 | identical_state = prev->psi_flags == next->psi_flags; |
36b238d5 JW |
841 | iter = NULL; |
842 | while ((group = iterate_groups(next, &iter))) { | |
7fae6c81 CZ |
843 | if (identical_state && |
844 | per_cpu_ptr(group->pcpu, cpu)->tasks[NR_ONCPU]) { | |
36b238d5 JW |
845 | common = group; |
846 | break; | |
847 | } | |
848 | ||
df774306 | 849 | psi_group_change(group, cpu, 0, TSK_ONCPU, now, true); |
36b238d5 JW |
850 | } |
851 | } | |
852 | ||
36b238d5 | 853 | if (prev->pid) { |
4117cebf CZ |
854 | int clear = TSK_ONCPU, set = 0; |
855 | ||
856 | /* | |
857 | * When we're going to sleep, psi_dequeue() lets us handle | |
858 | * TSK_RUNNING and TSK_IOWAIT here, where we can combine it | |
859 | * with TSK_ONCPU and save walking common ancestors twice. | |
860 | */ | |
861 | if (sleep) { | |
862 | clear |= TSK_RUNNING; | |
863 | if (prev->in_iowait) | |
864 | set |= TSK_IOWAIT; | |
865 | } | |
866 | ||
867 | psi_flags_change(prev, clear, set); | |
0e94682b | 868 | |
36b238d5 JW |
869 | iter = NULL; |
870 | while ((group = iterate_groups(prev, &iter)) && group != common) | |
df774306 | 871 | psi_group_change(group, cpu, clear, set, now, true); |
4117cebf CZ |
872 | |
873 | /* | |
874 | * TSK_ONCPU is handled up to the common ancestor. If we're tasked | |
875 | * with dequeuing too, finish that for the rest of the hierarchy. | |
876 | */ | |
877 | if (sleep) { | |
878 | clear &= ~TSK_ONCPU; | |
879 | for (; group; group = iterate_groups(prev, &iter)) | |
df774306 | 880 | psi_group_change(group, cpu, clear, set, now, true); |
4117cebf | 881 | } |
1b69ac6b | 882 | } |
eb414681 JW |
883 | } |
884 | ||
eb414681 JW |
885 | /** |
886 | * psi_memstall_enter - mark the beginning of a memory stall section | |
887 | * @flags: flags to handle nested sections | |
888 | * | |
889 | * Marks the calling task as being stalled due to a lack of memory, | |
890 | * such as waiting for a refault or performing reclaim. | |
891 | */ | |
892 | void psi_memstall_enter(unsigned long *flags) | |
893 | { | |
894 | struct rq_flags rf; | |
895 | struct rq *rq; | |
896 | ||
e0c27447 | 897 | if (static_branch_likely(&psi_disabled)) |
eb414681 JW |
898 | return; |
899 | ||
1066d1b6 | 900 | *flags = current->in_memstall; |
eb414681 JW |
901 | if (*flags) |
902 | return; | |
903 | /* | |
1066d1b6 | 904 | * in_memstall setting & accounting needs to be atomic wrt |
eb414681 JW |
905 | * changes to the task's scheduling state, otherwise we can |
906 | * race with CPU migration. | |
907 | */ | |
908 | rq = this_rq_lock_irq(&rf); | |
909 | ||
1066d1b6 | 910 | current->in_memstall = 1; |
eb414681 JW |
911 | psi_task_change(current, 0, TSK_MEMSTALL); |
912 | ||
913 | rq_unlock_irq(rq, &rf); | |
914 | } | |
915 | ||
916 | /** | |
917 | * psi_memstall_leave - mark the end of an memory stall section | |
918 | * @flags: flags to handle nested memdelay sections | |
919 | * | |
920 | * Marks the calling task as no longer stalled due to lack of memory. | |
921 | */ | |
922 | void psi_memstall_leave(unsigned long *flags) | |
923 | { | |
924 | struct rq_flags rf; | |
925 | struct rq *rq; | |
926 | ||
e0c27447 | 927 | if (static_branch_likely(&psi_disabled)) |
eb414681 JW |
928 | return; |
929 | ||
930 | if (*flags) | |
931 | return; | |
932 | /* | |
1066d1b6 | 933 | * in_memstall clearing & accounting needs to be atomic wrt |
eb414681 JW |
934 | * changes to the task's scheduling state, otherwise we could |
935 | * race with CPU migration. | |
936 | */ | |
937 | rq = this_rq_lock_irq(&rf); | |
938 | ||
1066d1b6 | 939 | current->in_memstall = 0; |
eb414681 JW |
940 | psi_task_change(current, TSK_MEMSTALL, 0); |
941 | ||
942 | rq_unlock_irq(rq, &rf); | |
943 | } | |
944 | ||
2ce7135a JW |
945 | #ifdef CONFIG_CGROUPS |
946 | int psi_cgroup_alloc(struct cgroup *cgroup) | |
947 | { | |
e0c27447 | 948 | if (static_branch_likely(&psi_disabled)) |
2ce7135a JW |
949 | return 0; |
950 | ||
951 | cgroup->psi.pcpu = alloc_percpu(struct psi_group_cpu); | |
952 | if (!cgroup->psi.pcpu) | |
953 | return -ENOMEM; | |
954 | group_init(&cgroup->psi); | |
955 | return 0; | |
956 | } | |
957 | ||
958 | void psi_cgroup_free(struct cgroup *cgroup) | |
959 | { | |
e0c27447 | 960 | if (static_branch_likely(&psi_disabled)) |
2ce7135a JW |
961 | return; |
962 | ||
bcc78db6 | 963 | cancel_delayed_work_sync(&cgroup->psi.avgs_work); |
2ce7135a | 964 | free_percpu(cgroup->psi.pcpu); |
0e94682b SB |
965 | /* All triggers must be removed by now */ |
966 | WARN_ONCE(cgroup->psi.poll_states, "psi: trigger leak\n"); | |
2ce7135a JW |
967 | } |
968 | ||
969 | /** | |
970 | * cgroup_move_task - move task to a different cgroup | |
971 | * @task: the task | |
972 | * @to: the target css_set | |
973 | * | |
974 | * Move task to a new cgroup and safely migrate its associated stall | |
975 | * state between the different groups. | |
976 | * | |
977 | * This function acquires the task's rq lock to lock out concurrent | |
978 | * changes to the task's scheduling state and - in case the task is | |
979 | * running - concurrent changes to its stall state. | |
980 | */ | |
981 | void cgroup_move_task(struct task_struct *task, struct css_set *to) | |
982 | { | |
d583d360 | 983 | unsigned int task_flags; |
2ce7135a JW |
984 | struct rq_flags rf; |
985 | struct rq *rq; | |
986 | ||
e0c27447 | 987 | if (static_branch_likely(&psi_disabled)) { |
8fcb2312 OJ |
988 | /* |
989 | * Lame to do this here, but the scheduler cannot be locked | |
990 | * from the outside, so we move cgroups from inside sched/. | |
991 | */ | |
992 | rcu_assign_pointer(task->cgroups, to); | |
993 | return; | |
994 | } | |
2ce7135a | 995 | |
8fcb2312 | 996 | rq = task_rq_lock(task, &rf); |
2ce7135a | 997 | |
d583d360 JW |
998 | /* |
999 | * We may race with schedule() dropping the rq lock between | |
1000 | * deactivating prev and switching to next. Because the psi | |
1001 | * updates from the deactivation are deferred to the switch | |
1002 | * callback to save cgroup tree updates, the task's scheduling | |
1003 | * state here is not coherent with its psi state: | |
1004 | * | |
1005 | * schedule() cgroup_move_task() | |
1006 | * rq_lock() | |
1007 | * deactivate_task() | |
1008 | * p->on_rq = 0 | |
1009 | * psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates | |
1010 | * pick_next_task() | |
1011 | * rq_unlock() | |
1012 | * rq_lock() | |
1013 | * psi_task_change() // old cgroup | |
1014 | * task->cgroups = to | |
1015 | * psi_task_change() // new cgroup | |
1016 | * rq_unlock() | |
1017 | * rq_lock() | |
1018 | * psi_sched_switch() // does deferred updates in new cgroup | |
1019 | * | |
1020 | * Don't rely on the scheduling state. Use psi_flags instead. | |
1021 | */ | |
1022 | task_flags = task->psi_flags; | |
2ce7135a | 1023 | |
8fcb2312 OJ |
1024 | if (task_flags) |
1025 | psi_task_change(task, task_flags, 0); | |
1026 | ||
1027 | /* See comment above */ | |
2ce7135a JW |
1028 | rcu_assign_pointer(task->cgroups, to); |
1029 | ||
8fcb2312 OJ |
1030 | if (task_flags) |
1031 | psi_task_change(task, 0, task_flags); | |
2ce7135a | 1032 | |
8fcb2312 | 1033 | task_rq_unlock(rq, task, &rf); |
2ce7135a JW |
1034 | } |
1035 | #endif /* CONFIG_CGROUPS */ | |
1036 | ||
1037 | int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res) | |
eb414681 JW |
1038 | { |
1039 | int full; | |
7fc70a39 | 1040 | u64 now; |
eb414681 | 1041 | |
e0c27447 | 1042 | if (static_branch_likely(&psi_disabled)) |
eb414681 JW |
1043 | return -EOPNOTSUPP; |
1044 | ||
7fc70a39 SB |
1045 | /* Update averages before reporting them */ |
1046 | mutex_lock(&group->avgs_lock); | |
1047 | now = sched_clock(); | |
0e94682b | 1048 | collect_percpu_times(group, PSI_AVGS, NULL); |
7fc70a39 SB |
1049 | if (now >= group->avg_next_update) |
1050 | group->avg_next_update = update_averages(group, now); | |
1051 | mutex_unlock(&group->avgs_lock); | |
eb414681 | 1052 | |
e7fcd762 | 1053 | for (full = 0; full < 2; full++) { |
eb414681 JW |
1054 | unsigned long avg[3]; |
1055 | u64 total; | |
1056 | int w; | |
1057 | ||
1058 | for (w = 0; w < 3; w++) | |
1059 | avg[w] = group->avg[res * 2 + full][w]; | |
0e94682b SB |
1060 | total = div_u64(group->total[PSI_AVGS][res * 2 + full], |
1061 | NSEC_PER_USEC); | |
eb414681 JW |
1062 | |
1063 | seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n", | |
1064 | full ? "full" : "some", | |
1065 | LOAD_INT(avg[0]), LOAD_FRAC(avg[0]), | |
1066 | LOAD_INT(avg[1]), LOAD_FRAC(avg[1]), | |
1067 | LOAD_INT(avg[2]), LOAD_FRAC(avg[2]), | |
1068 | total); | |
1069 | } | |
1070 | ||
1071 | return 0; | |
1072 | } | |
1073 | ||
1074 | static int psi_io_show(struct seq_file *m, void *v) | |
1075 | { | |
1076 | return psi_show(m, &psi_system, PSI_IO); | |
1077 | } | |
1078 | ||
1079 | static int psi_memory_show(struct seq_file *m, void *v) | |
1080 | { | |
1081 | return psi_show(m, &psi_system, PSI_MEM); | |
1082 | } | |
1083 | ||
1084 | static int psi_cpu_show(struct seq_file *m, void *v) | |
1085 | { | |
1086 | return psi_show(m, &psi_system, PSI_CPU); | |
1087 | } | |
1088 | ||
6db12ee0 JH |
1089 | static int psi_open(struct file *file, int (*psi_show)(struct seq_file *, void *)) |
1090 | { | |
1091 | if (file->f_mode & FMODE_WRITE && !capable(CAP_SYS_RESOURCE)) | |
1092 | return -EPERM; | |
1093 | ||
1094 | return single_open(file, psi_show, NULL); | |
1095 | } | |
1096 | ||
eb414681 JW |
1097 | static int psi_io_open(struct inode *inode, struct file *file) |
1098 | { | |
6db12ee0 | 1099 | return psi_open(file, psi_io_show); |
eb414681 JW |
1100 | } |
1101 | ||
1102 | static int psi_memory_open(struct inode *inode, struct file *file) | |
1103 | { | |
6db12ee0 | 1104 | return psi_open(file, psi_memory_show); |
eb414681 JW |
1105 | } |
1106 | ||
1107 | static int psi_cpu_open(struct inode *inode, struct file *file) | |
1108 | { | |
6db12ee0 | 1109 | return psi_open(file, psi_cpu_show); |
eb414681 JW |
1110 | } |
1111 | ||
0e94682b SB |
1112 | struct psi_trigger *psi_trigger_create(struct psi_group *group, |
1113 | char *buf, size_t nbytes, enum psi_res res) | |
1114 | { | |
1115 | struct psi_trigger *t; | |
1116 | enum psi_states state; | |
1117 | u32 threshold_us; | |
1118 | u32 window_us; | |
1119 | ||
1120 | if (static_branch_likely(&psi_disabled)) | |
1121 | return ERR_PTR(-EOPNOTSUPP); | |
1122 | ||
1123 | if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2) | |
1124 | state = PSI_IO_SOME + res * 2; | |
1125 | else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2) | |
1126 | state = PSI_IO_FULL + res * 2; | |
1127 | else | |
1128 | return ERR_PTR(-EINVAL); | |
1129 | ||
1130 | if (state >= PSI_NONIDLE) | |
1131 | return ERR_PTR(-EINVAL); | |
1132 | ||
1133 | if (window_us < WINDOW_MIN_US || | |
1134 | window_us > WINDOW_MAX_US) | |
1135 | return ERR_PTR(-EINVAL); | |
1136 | ||
1137 | /* Check threshold */ | |
1138 | if (threshold_us == 0 || threshold_us > window_us) | |
1139 | return ERR_PTR(-EINVAL); | |
1140 | ||
1141 | t = kmalloc(sizeof(*t), GFP_KERNEL); | |
1142 | if (!t) | |
1143 | return ERR_PTR(-ENOMEM); | |
1144 | ||
1145 | t->group = group; | |
1146 | t->state = state; | |
1147 | t->threshold = threshold_us * NSEC_PER_USEC; | |
1148 | t->win.size = window_us * NSEC_PER_USEC; | |
1149 | window_reset(&t->win, 0, 0, 0); | |
1150 | ||
1151 | t->event = 0; | |
1152 | t->last_event_time = 0; | |
1153 | init_waitqueue_head(&t->event_wait); | |
1154 | kref_init(&t->refcount); | |
1155 | ||
1156 | mutex_lock(&group->trigger_lock); | |
1157 | ||
461daba0 SB |
1158 | if (!rcu_access_pointer(group->poll_task)) { |
1159 | struct task_struct *task; | |
0e94682b | 1160 | |
461daba0 SB |
1161 | task = kthread_create(psi_poll_worker, group, "psimon"); |
1162 | if (IS_ERR(task)) { | |
0e94682b SB |
1163 | kfree(t); |
1164 | mutex_unlock(&group->trigger_lock); | |
461daba0 | 1165 | return ERR_CAST(task); |
0e94682b | 1166 | } |
461daba0 | 1167 | atomic_set(&group->poll_wakeup, 0); |
461daba0 | 1168 | wake_up_process(task); |
461daba0 | 1169 | rcu_assign_pointer(group->poll_task, task); |
0e94682b SB |
1170 | } |
1171 | ||
1172 | list_add(&t->node, &group->triggers); | |
1173 | group->poll_min_period = min(group->poll_min_period, | |
1174 | div_u64(t->win.size, UPDATES_PER_WINDOW)); | |
1175 | group->nr_triggers[t->state]++; | |
1176 | group->poll_states |= (1 << t->state); | |
1177 | ||
1178 | mutex_unlock(&group->trigger_lock); | |
1179 | ||
1180 | return t; | |
1181 | } | |
1182 | ||
1183 | static void psi_trigger_destroy(struct kref *ref) | |
1184 | { | |
1185 | struct psi_trigger *t = container_of(ref, struct psi_trigger, refcount); | |
1186 | struct psi_group *group = t->group; | |
461daba0 | 1187 | struct task_struct *task_to_destroy = NULL; |
0e94682b SB |
1188 | |
1189 | if (static_branch_likely(&psi_disabled)) | |
1190 | return; | |
1191 | ||
1192 | /* | |
1193 | * Wakeup waiters to stop polling. Can happen if cgroup is deleted | |
1194 | * from under a polling process. | |
1195 | */ | |
1196 | wake_up_interruptible(&t->event_wait); | |
1197 | ||
1198 | mutex_lock(&group->trigger_lock); | |
1199 | ||
1200 | if (!list_empty(&t->node)) { | |
1201 | struct psi_trigger *tmp; | |
1202 | u64 period = ULLONG_MAX; | |
1203 | ||
1204 | list_del(&t->node); | |
1205 | group->nr_triggers[t->state]--; | |
1206 | if (!group->nr_triggers[t->state]) | |
1207 | group->poll_states &= ~(1 << t->state); | |
1208 | /* reset min update period for the remaining triggers */ | |
1209 | list_for_each_entry(tmp, &group->triggers, node) | |
1210 | period = min(period, div_u64(tmp->win.size, | |
1211 | UPDATES_PER_WINDOW)); | |
1212 | group->poll_min_period = period; | |
461daba0 | 1213 | /* Destroy poll_task when the last trigger is destroyed */ |
0e94682b SB |
1214 | if (group->poll_states == 0) { |
1215 | group->polling_until = 0; | |
461daba0 SB |
1216 | task_to_destroy = rcu_dereference_protected( |
1217 | group->poll_task, | |
0e94682b | 1218 | lockdep_is_held(&group->trigger_lock)); |
461daba0 | 1219 | rcu_assign_pointer(group->poll_task, NULL); |
8f91efd8 | 1220 | del_timer(&group->poll_timer); |
0e94682b SB |
1221 | } |
1222 | } | |
1223 | ||
1224 | mutex_unlock(&group->trigger_lock); | |
1225 | ||
1226 | /* | |
1227 | * Wait for both *trigger_ptr from psi_trigger_replace and | |
461daba0 SB |
1228 | * poll_task RCUs to complete their read-side critical sections |
1229 | * before destroying the trigger and optionally the poll_task | |
0e94682b SB |
1230 | */ |
1231 | synchronize_rcu(); | |
1232 | /* | |
8f91efd8 | 1233 | * Stop kthread 'psimon' after releasing trigger_lock to prevent a |
0e94682b SB |
1234 | * deadlock while waiting for psi_poll_work to acquire trigger_lock |
1235 | */ | |
461daba0 | 1236 | if (task_to_destroy) { |
7b2b55da JX |
1237 | /* |
1238 | * After the RCU grace period has expired, the worker | |
461daba0 | 1239 | * can no longer be found through group->poll_task. |
7b2b55da | 1240 | */ |
461daba0 | 1241 | kthread_stop(task_to_destroy); |
0e94682b SB |
1242 | } |
1243 | kfree(t); | |
1244 | } | |
1245 | ||
1246 | void psi_trigger_replace(void **trigger_ptr, struct psi_trigger *new) | |
1247 | { | |
1248 | struct psi_trigger *old = *trigger_ptr; | |
1249 | ||
1250 | if (static_branch_likely(&psi_disabled)) | |
1251 | return; | |
1252 | ||
1253 | rcu_assign_pointer(*trigger_ptr, new); | |
1254 | if (old) | |
1255 | kref_put(&old->refcount, psi_trigger_destroy); | |
1256 | } | |
1257 | ||
1258 | __poll_t psi_trigger_poll(void **trigger_ptr, | |
1259 | struct file *file, poll_table *wait) | |
1260 | { | |
1261 | __poll_t ret = DEFAULT_POLLMASK; | |
1262 | struct psi_trigger *t; | |
1263 | ||
1264 | if (static_branch_likely(&psi_disabled)) | |
1265 | return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI; | |
1266 | ||
1267 | rcu_read_lock(); | |
1268 | ||
1269 | t = rcu_dereference(*(void __rcu __force **)trigger_ptr); | |
1270 | if (!t) { | |
1271 | rcu_read_unlock(); | |
1272 | return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI; | |
1273 | } | |
1274 | kref_get(&t->refcount); | |
1275 | ||
1276 | rcu_read_unlock(); | |
1277 | ||
1278 | poll_wait(file, &t->event_wait, wait); | |
1279 | ||
1280 | if (cmpxchg(&t->event, 1, 0) == 1) | |
1281 | ret |= EPOLLPRI; | |
1282 | ||
1283 | kref_put(&t->refcount, psi_trigger_destroy); | |
1284 | ||
1285 | return ret; | |
1286 | } | |
1287 | ||
1288 | static ssize_t psi_write(struct file *file, const char __user *user_buf, | |
1289 | size_t nbytes, enum psi_res res) | |
1290 | { | |
1291 | char buf[32]; | |
1292 | size_t buf_size; | |
1293 | struct seq_file *seq; | |
1294 | struct psi_trigger *new; | |
1295 | ||
1296 | if (static_branch_likely(&psi_disabled)) | |
1297 | return -EOPNOTSUPP; | |
1298 | ||
6fcca0fa SB |
1299 | if (!nbytes) |
1300 | return -EINVAL; | |
1301 | ||
4adcdcea | 1302 | buf_size = min(nbytes, sizeof(buf)); |
0e94682b SB |
1303 | if (copy_from_user(buf, user_buf, buf_size)) |
1304 | return -EFAULT; | |
1305 | ||
1306 | buf[buf_size - 1] = '\0'; | |
1307 | ||
1308 | new = psi_trigger_create(&psi_system, buf, nbytes, res); | |
1309 | if (IS_ERR(new)) | |
1310 | return PTR_ERR(new); | |
1311 | ||
1312 | seq = file->private_data; | |
1313 | /* Take seq->lock to protect seq->private from concurrent writes */ | |
1314 | mutex_lock(&seq->lock); | |
1315 | psi_trigger_replace(&seq->private, new); | |
1316 | mutex_unlock(&seq->lock); | |
1317 | ||
1318 | return nbytes; | |
1319 | } | |
1320 | ||
1321 | static ssize_t psi_io_write(struct file *file, const char __user *user_buf, | |
1322 | size_t nbytes, loff_t *ppos) | |
1323 | { | |
1324 | return psi_write(file, user_buf, nbytes, PSI_IO); | |
1325 | } | |
1326 | ||
1327 | static ssize_t psi_memory_write(struct file *file, const char __user *user_buf, | |
1328 | size_t nbytes, loff_t *ppos) | |
1329 | { | |
1330 | return psi_write(file, user_buf, nbytes, PSI_MEM); | |
1331 | } | |
1332 | ||
1333 | static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf, | |
1334 | size_t nbytes, loff_t *ppos) | |
1335 | { | |
1336 | return psi_write(file, user_buf, nbytes, PSI_CPU); | |
1337 | } | |
1338 | ||
1339 | static __poll_t psi_fop_poll(struct file *file, poll_table *wait) | |
1340 | { | |
1341 | struct seq_file *seq = file->private_data; | |
1342 | ||
1343 | return psi_trigger_poll(&seq->private, file, wait); | |
1344 | } | |
1345 | ||
1346 | static int psi_fop_release(struct inode *inode, struct file *file) | |
1347 | { | |
1348 | struct seq_file *seq = file->private_data; | |
1349 | ||
1350 | psi_trigger_replace(&seq->private, NULL); | |
1351 | return single_release(inode, file); | |
1352 | } | |
1353 | ||
97a32539 AD |
1354 | static const struct proc_ops psi_io_proc_ops = { |
1355 | .proc_open = psi_io_open, | |
1356 | .proc_read = seq_read, | |
1357 | .proc_lseek = seq_lseek, | |
1358 | .proc_write = psi_io_write, | |
1359 | .proc_poll = psi_fop_poll, | |
1360 | .proc_release = psi_fop_release, | |
eb414681 JW |
1361 | }; |
1362 | ||
97a32539 AD |
1363 | static const struct proc_ops psi_memory_proc_ops = { |
1364 | .proc_open = psi_memory_open, | |
1365 | .proc_read = seq_read, | |
1366 | .proc_lseek = seq_lseek, | |
1367 | .proc_write = psi_memory_write, | |
1368 | .proc_poll = psi_fop_poll, | |
1369 | .proc_release = psi_fop_release, | |
eb414681 JW |
1370 | }; |
1371 | ||
97a32539 AD |
1372 | static const struct proc_ops psi_cpu_proc_ops = { |
1373 | .proc_open = psi_cpu_open, | |
1374 | .proc_read = seq_read, | |
1375 | .proc_lseek = seq_lseek, | |
1376 | .proc_write = psi_cpu_write, | |
1377 | .proc_poll = psi_fop_poll, | |
1378 | .proc_release = psi_fop_release, | |
eb414681 JW |
1379 | }; |
1380 | ||
1381 | static int __init psi_proc_init(void) | |
1382 | { | |
3d817689 WL |
1383 | if (psi_enable) { |
1384 | proc_mkdir("pressure", NULL); | |
6db12ee0 JH |
1385 | proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops); |
1386 | proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops); | |
1387 | proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops); | |
3d817689 | 1388 | } |
eb414681 JW |
1389 | return 0; |
1390 | } | |
1391 | module_init(psi_proc_init); |