]>
Commit | Line | Data |
---|---|---|
eb414681 JW |
1 | /* |
2 | * Pressure stall information for CPU, memory and IO | |
3 | * | |
4 | * Copyright (c) 2018 Facebook, Inc. | |
5 | * Author: Johannes Weiner <hannes@cmpxchg.org> | |
6 | * | |
0e94682b SB |
7 | * Polling support by Suren Baghdasaryan <surenb@google.com> |
8 | * Copyright (c) 2018 Google, Inc. | |
9 | * | |
eb414681 JW |
10 | * When CPU, memory and IO are contended, tasks experience delays that |
11 | * reduce throughput and introduce latencies into the workload. Memory | |
12 | * and IO contention, in addition, can cause a full loss of forward | |
13 | * progress in which the CPU goes idle. | |
14 | * | |
15 | * This code aggregates individual task delays into resource pressure | |
16 | * metrics that indicate problems with both workload health and | |
17 | * resource utilization. | |
18 | * | |
19 | * Model | |
20 | * | |
21 | * The time in which a task can execute on a CPU is our baseline for | |
22 | * productivity. Pressure expresses the amount of time in which this | |
23 | * potential cannot be realized due to resource contention. | |
24 | * | |
25 | * This concept of productivity has two components: the workload and | |
26 | * the CPU. To measure the impact of pressure on both, we define two | |
27 | * contention states for a resource: SOME and FULL. | |
28 | * | |
29 | * In the SOME state of a given resource, one or more tasks are | |
30 | * delayed on that resource. This affects the workload's ability to | |
31 | * perform work, but the CPU may still be executing other tasks. | |
32 | * | |
33 | * In the FULL state of a given resource, all non-idle tasks are | |
34 | * delayed on that resource such that nobody is advancing and the CPU | |
35 | * goes idle. This leaves both workload and CPU unproductive. | |
36 | * | |
eb414681 | 37 | * SOME = nr_delayed_tasks != 0 |
221276cb BC |
38 | * FULL = nr_delayed_tasks != 0 && nr_productive_tasks == 0 |
39 | * | |
40 | * What it means for a task to be productive is defined differently | |
41 | * for each resource. For IO, productive means a running task. For | |
42 | * memory, productive means a running task that isn't a reclaimer. For | |
43 | * CPU, productive means an oncpu task. | |
44 | * | |
45 | * Naturally, the FULL state doesn't exist for the CPU resource at the | |
46 | * system level, but exist at the cgroup level. At the cgroup level, | |
47 | * FULL means all non-idle tasks in the cgroup are delayed on the CPU | |
48 | * resource which is being used by others outside of the cgroup or | |
49 | * throttled by the cgroup cpu.max configuration. | |
eb414681 JW |
50 | * |
51 | * The percentage of wallclock time spent in those compound stall | |
52 | * states gives pressure numbers between 0 and 100 for each resource, | |
53 | * where the SOME percentage indicates workload slowdowns and the FULL | |
54 | * percentage indicates reduced CPU utilization: | |
55 | * | |
56 | * %SOME = time(SOME) / period | |
57 | * %FULL = time(FULL) / period | |
58 | * | |
59 | * Multiple CPUs | |
60 | * | |
61 | * The more tasks and available CPUs there are, the more work can be | |
62 | * performed concurrently. This means that the potential that can go | |
63 | * unrealized due to resource contention *also* scales with non-idle | |
64 | * tasks and CPUs. | |
65 | * | |
66 | * Consider a scenario where 257 number crunching tasks are trying to | |
67 | * run concurrently on 256 CPUs. If we simply aggregated the task | |
68 | * states, we would have to conclude a CPU SOME pressure number of | |
69 | * 100%, since *somebody* is waiting on a runqueue at all | |
70 | * times. However, that is clearly not the amount of contention the | |
3b03706f | 71 | * workload is experiencing: only one out of 256 possible execution |
eb414681 JW |
72 | * threads will be contended at any given time, or about 0.4%. |
73 | * | |
74 | * Conversely, consider a scenario of 4 tasks and 4 CPUs where at any | |
75 | * given time *one* of the tasks is delayed due to a lack of memory. | |
76 | * Again, looking purely at the task state would yield a memory FULL | |
77 | * pressure number of 0%, since *somebody* is always making forward | |
78 | * progress. But again this wouldn't capture the amount of execution | |
79 | * potential lost, which is 1 out of 4 CPUs, or 25%. | |
80 | * | |
81 | * To calculate wasted potential (pressure) with multiple processors, | |
82 | * we have to base our calculation on the number of non-idle tasks in | |
83 | * conjunction with the number of available CPUs, which is the number | |
84 | * of potential execution threads. SOME becomes then the proportion of | |
3b03706f | 85 | * delayed tasks to possible threads, and FULL is the share of possible |
eb414681 JW |
86 | * threads that are unproductive due to delays: |
87 | * | |
88 | * threads = min(nr_nonidle_tasks, nr_cpus) | |
89 | * SOME = min(nr_delayed_tasks / threads, 1) | |
221276cb | 90 | * FULL = (threads - min(nr_productive_tasks, threads)) / threads |
eb414681 JW |
91 | * |
92 | * For the 257 number crunchers on 256 CPUs, this yields: | |
93 | * | |
94 | * threads = min(257, 256) | |
95 | * SOME = min(1 / 256, 1) = 0.4% | |
221276cb | 96 | * FULL = (256 - min(256, 256)) / 256 = 0% |
eb414681 JW |
97 | * |
98 | * For the 1 out of 4 memory-delayed tasks, this yields: | |
99 | * | |
100 | * threads = min(4, 4) | |
101 | * SOME = min(1 / 4, 1) = 25% | |
102 | * FULL = (4 - min(3, 4)) / 4 = 25% | |
103 | * | |
104 | * [ Substitute nr_cpus with 1, and you can see that it's a natural | |
105 | * extension of the single-CPU model. ] | |
106 | * | |
107 | * Implementation | |
108 | * | |
109 | * To assess the precise time spent in each such state, we would have | |
110 | * to freeze the system on task changes and start/stop the state | |
111 | * clocks accordingly. Obviously that doesn't scale in practice. | |
112 | * | |
113 | * Because the scheduler aims to distribute the compute load evenly | |
114 | * among the available CPUs, we can track task state locally to each | |
115 | * CPU and, at much lower frequency, extrapolate the global state for | |
116 | * the cumulative stall times and the running averages. | |
117 | * | |
118 | * For each runqueue, we track: | |
119 | * | |
120 | * tSOME[cpu] = time(nr_delayed_tasks[cpu] != 0) | |
221276cb | 121 | * tFULL[cpu] = time(nr_delayed_tasks[cpu] && !nr_productive_tasks[cpu]) |
eb414681 JW |
122 | * tNONIDLE[cpu] = time(nr_nonidle_tasks[cpu] != 0) |
123 | * | |
124 | * and then periodically aggregate: | |
125 | * | |
126 | * tNONIDLE = sum(tNONIDLE[i]) | |
127 | * | |
128 | * tSOME = sum(tSOME[i] * tNONIDLE[i]) / tNONIDLE | |
129 | * tFULL = sum(tFULL[i] * tNONIDLE[i]) / tNONIDLE | |
130 | * | |
131 | * %SOME = tSOME / period | |
132 | * %FULL = tFULL / period | |
133 | * | |
134 | * This gives us an approximation of pressure that is practical | |
135 | * cost-wise, yet way more sensitive and accurate than periodic | |
136 | * sampling of the aggregate task states would be. | |
137 | */ | |
138 | ||
1b69ac6b | 139 | #include "../workqueue_internal.h" |
eb414681 JW |
140 | #include <linux/sched/loadavg.h> |
141 | #include <linux/seq_file.h> | |
142 | #include <linux/proc_fs.h> | |
143 | #include <linux/seqlock.h> | |
0e94682b | 144 | #include <linux/uaccess.h> |
eb414681 JW |
145 | #include <linux/cgroup.h> |
146 | #include <linux/module.h> | |
147 | #include <linux/sched.h> | |
0e94682b SB |
148 | #include <linux/ctype.h> |
149 | #include <linux/file.h> | |
150 | #include <linux/poll.h> | |
eb414681 JW |
151 | #include <linux/psi.h> |
152 | #include "sched.h" | |
153 | ||
154 | static int psi_bug __read_mostly; | |
155 | ||
e0c27447 | 156 | DEFINE_STATIC_KEY_FALSE(psi_disabled); |
3958e2d0 | 157 | DEFINE_STATIC_KEY_TRUE(psi_cgroups_enabled); |
e0c27447 JW |
158 | |
159 | #ifdef CONFIG_PSI_DEFAULT_DISABLED | |
9289c5e6 | 160 | static bool psi_enable; |
e0c27447 | 161 | #else |
9289c5e6 | 162 | static bool psi_enable = true; |
e0c27447 JW |
163 | #endif |
164 | static int __init setup_psi(char *str) | |
165 | { | |
166 | return kstrtobool(str, &psi_enable) == 0; | |
167 | } | |
168 | __setup("psi=", setup_psi); | |
eb414681 JW |
169 | |
170 | /* Running averages - we need to be higher-res than loadavg */ | |
171 | #define PSI_FREQ (2*HZ+1) /* 2 sec intervals */ | |
172 | #define EXP_10s 1677 /* 1/exp(2s/10s) as fixed-point */ | |
173 | #define EXP_60s 1981 /* 1/exp(2s/60s) */ | |
174 | #define EXP_300s 2034 /* 1/exp(2s/300s) */ | |
175 | ||
0e94682b SB |
176 | /* PSI trigger definitions */ |
177 | #define WINDOW_MIN_US 500000 /* Min window size is 500ms */ | |
178 | #define WINDOW_MAX_US 10000000 /* Max window size is 10s */ | |
179 | #define UPDATES_PER_WINDOW 10 /* 10 updates per window */ | |
180 | ||
eb414681 JW |
181 | /* Sampling frequency in nanoseconds */ |
182 | static u64 psi_period __read_mostly; | |
183 | ||
184 | /* System-level pressure and stall tracking */ | |
185 | static DEFINE_PER_CPU(struct psi_group_cpu, system_group_pcpu); | |
df5ba5be | 186 | struct psi_group psi_system = { |
eb414681 JW |
187 | .pcpu = &system_group_pcpu, |
188 | }; | |
189 | ||
bcc78db6 | 190 | static void psi_avgs_work(struct work_struct *work); |
eb414681 | 191 | |
8f91efd8 ZH |
192 | static void poll_timer_fn(struct timer_list *t); |
193 | ||
eb414681 JW |
194 | static void group_init(struct psi_group *group) |
195 | { | |
196 | int cpu; | |
197 | ||
198 | for_each_possible_cpu(cpu) | |
199 | seqcount_init(&per_cpu_ptr(group->pcpu, cpu)->seq); | |
3dfbe25c JW |
200 | group->avg_last_update = sched_clock(); |
201 | group->avg_next_update = group->avg_last_update + psi_period; | |
bcc78db6 SB |
202 | INIT_DELAYED_WORK(&group->avgs_work, psi_avgs_work); |
203 | mutex_init(&group->avgs_lock); | |
0e94682b | 204 | /* Init trigger-related members */ |
0e94682b SB |
205 | mutex_init(&group->trigger_lock); |
206 | INIT_LIST_HEAD(&group->triggers); | |
207 | memset(group->nr_triggers, 0, sizeof(group->nr_triggers)); | |
208 | group->poll_states = 0; | |
209 | group->poll_min_period = U32_MAX; | |
210 | memset(group->polling_total, 0, sizeof(group->polling_total)); | |
211 | group->polling_next_update = ULLONG_MAX; | |
212 | group->polling_until = 0; | |
8f91efd8 ZH |
213 | init_waitqueue_head(&group->poll_wait); |
214 | timer_setup(&group->poll_timer, poll_timer_fn, 0); | |
461daba0 | 215 | rcu_assign_pointer(group->poll_task, NULL); |
eb414681 JW |
216 | } |
217 | ||
218 | void __init psi_init(void) | |
219 | { | |
e0c27447 JW |
220 | if (!psi_enable) { |
221 | static_branch_enable(&psi_disabled); | |
eb414681 | 222 | return; |
e0c27447 | 223 | } |
eb414681 | 224 | |
3958e2d0 SB |
225 | if (!cgroup_psi_enabled()) |
226 | static_branch_disable(&psi_cgroups_enabled); | |
227 | ||
eb414681 JW |
228 | psi_period = jiffies_to_nsecs(PSI_FREQ); |
229 | group_init(&psi_system); | |
230 | } | |
231 | ||
232 | static bool test_state(unsigned int *tasks, enum psi_states state) | |
233 | { | |
234 | switch (state) { | |
235 | case PSI_IO_SOME: | |
fddc8bab | 236 | return unlikely(tasks[NR_IOWAIT]); |
eb414681 | 237 | case PSI_IO_FULL: |
fddc8bab | 238 | return unlikely(tasks[NR_IOWAIT] && !tasks[NR_RUNNING]); |
eb414681 | 239 | case PSI_MEM_SOME: |
fddc8bab | 240 | return unlikely(tasks[NR_MEMSTALL]); |
eb414681 | 241 | case PSI_MEM_FULL: |
221276cb BC |
242 | return unlikely(tasks[NR_MEMSTALL] && |
243 | tasks[NR_RUNNING] == tasks[NR_MEMSTALL_RUNNING]); | |
eb414681 | 244 | case PSI_CPU_SOME: |
fddc8bab | 245 | return unlikely(tasks[NR_RUNNING] > tasks[NR_ONCPU]); |
e7fcd762 | 246 | case PSI_CPU_FULL: |
fddc8bab | 247 | return unlikely(tasks[NR_RUNNING] && !tasks[NR_ONCPU]); |
eb414681 JW |
248 | case PSI_NONIDLE: |
249 | return tasks[NR_IOWAIT] || tasks[NR_MEMSTALL] || | |
250 | tasks[NR_RUNNING]; | |
251 | default: | |
252 | return false; | |
253 | } | |
254 | } | |
255 | ||
0e94682b SB |
256 | static void get_recent_times(struct psi_group *group, int cpu, |
257 | enum psi_aggregators aggregator, u32 *times, | |
333f3017 | 258 | u32 *pchanged_states) |
eb414681 JW |
259 | { |
260 | struct psi_group_cpu *groupc = per_cpu_ptr(group->pcpu, cpu); | |
eb414681 | 261 | u64 now, state_start; |
33b2d630 | 262 | enum psi_states s; |
eb414681 | 263 | unsigned int seq; |
33b2d630 | 264 | u32 state_mask; |
eb414681 | 265 | |
333f3017 SB |
266 | *pchanged_states = 0; |
267 | ||
eb414681 JW |
268 | /* Snapshot a coherent view of the CPU state */ |
269 | do { | |
270 | seq = read_seqcount_begin(&groupc->seq); | |
271 | now = cpu_clock(cpu); | |
272 | memcpy(times, groupc->times, sizeof(groupc->times)); | |
33b2d630 | 273 | state_mask = groupc->state_mask; |
eb414681 JW |
274 | state_start = groupc->state_start; |
275 | } while (read_seqcount_retry(&groupc->seq, seq)); | |
276 | ||
277 | /* Calculate state time deltas against the previous snapshot */ | |
278 | for (s = 0; s < NR_PSI_STATES; s++) { | |
279 | u32 delta; | |
280 | /* | |
281 | * In addition to already concluded states, we also | |
282 | * incorporate currently active states on the CPU, | |
283 | * since states may last for many sampling periods. | |
284 | * | |
285 | * This way we keep our delta sampling buckets small | |
286 | * (u32) and our reported pressure close to what's | |
287 | * actually happening. | |
288 | */ | |
33b2d630 | 289 | if (state_mask & (1 << s)) |
eb414681 JW |
290 | times[s] += now - state_start; |
291 | ||
0e94682b SB |
292 | delta = times[s] - groupc->times_prev[aggregator][s]; |
293 | groupc->times_prev[aggregator][s] = times[s]; | |
eb414681 JW |
294 | |
295 | times[s] = delta; | |
333f3017 SB |
296 | if (delta) |
297 | *pchanged_states |= (1 << s); | |
eb414681 JW |
298 | } |
299 | } | |
300 | ||
301 | static void calc_avgs(unsigned long avg[3], int missed_periods, | |
302 | u64 time, u64 period) | |
303 | { | |
304 | unsigned long pct; | |
305 | ||
306 | /* Fill in zeroes for periods of no activity */ | |
307 | if (missed_periods) { | |
308 | avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods); | |
309 | avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods); | |
310 | avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods); | |
311 | } | |
312 | ||
313 | /* Sample the most recent active period */ | |
314 | pct = div_u64(time * 100, period); | |
315 | pct *= FIXED_1; | |
316 | avg[0] = calc_load(avg[0], EXP_10s, pct); | |
317 | avg[1] = calc_load(avg[1], EXP_60s, pct); | |
318 | avg[2] = calc_load(avg[2], EXP_300s, pct); | |
319 | } | |
320 | ||
0e94682b SB |
321 | static void collect_percpu_times(struct psi_group *group, |
322 | enum psi_aggregators aggregator, | |
323 | u32 *pchanged_states) | |
eb414681 JW |
324 | { |
325 | u64 deltas[NR_PSI_STATES - 1] = { 0, }; | |
eb414681 | 326 | unsigned long nonidle_total = 0; |
333f3017 | 327 | u32 changed_states = 0; |
eb414681 JW |
328 | int cpu; |
329 | int s; | |
330 | ||
eb414681 JW |
331 | /* |
332 | * Collect the per-cpu time buckets and average them into a | |
333 | * single time sample that is normalized to wallclock time. | |
334 | * | |
335 | * For averaging, each CPU is weighted by its non-idle time in | |
336 | * the sampling period. This eliminates artifacts from uneven | |
337 | * loading, or even entirely idle CPUs. | |
338 | */ | |
339 | for_each_possible_cpu(cpu) { | |
340 | u32 times[NR_PSI_STATES]; | |
341 | u32 nonidle; | |
333f3017 | 342 | u32 cpu_changed_states; |
eb414681 | 343 | |
0e94682b | 344 | get_recent_times(group, cpu, aggregator, times, |
333f3017 SB |
345 | &cpu_changed_states); |
346 | changed_states |= cpu_changed_states; | |
eb414681 JW |
347 | |
348 | nonidle = nsecs_to_jiffies(times[PSI_NONIDLE]); | |
349 | nonidle_total += nonidle; | |
350 | ||
351 | for (s = 0; s < PSI_NONIDLE; s++) | |
352 | deltas[s] += (u64)times[s] * nonidle; | |
353 | } | |
354 | ||
355 | /* | |
356 | * Integrate the sample into the running statistics that are | |
357 | * reported to userspace: the cumulative stall times and the | |
358 | * decaying averages. | |
359 | * | |
360 | * Pressure percentages are sampled at PSI_FREQ. We might be | |
361 | * called more often when the user polls more frequently than | |
362 | * that; we might be called less often when there is no task | |
363 | * activity, thus no data, and clock ticks are sporadic. The | |
364 | * below handles both. | |
365 | */ | |
366 | ||
367 | /* total= */ | |
368 | for (s = 0; s < NR_PSI_STATES - 1; s++) | |
0e94682b SB |
369 | group->total[aggregator][s] += |
370 | div_u64(deltas[s], max(nonidle_total, 1UL)); | |
eb414681 | 371 | |
333f3017 SB |
372 | if (pchanged_states) |
373 | *pchanged_states = changed_states; | |
7fc70a39 SB |
374 | } |
375 | ||
376 | static u64 update_averages(struct psi_group *group, u64 now) | |
377 | { | |
378 | unsigned long missed_periods = 0; | |
379 | u64 expires, period; | |
380 | u64 avg_next_update; | |
381 | int s; | |
382 | ||
eb414681 | 383 | /* avgX= */ |
bcc78db6 | 384 | expires = group->avg_next_update; |
4e37504d | 385 | if (now - expires >= psi_period) |
eb414681 JW |
386 | missed_periods = div_u64(now - expires, psi_period); |
387 | ||
388 | /* | |
389 | * The periodic clock tick can get delayed for various | |
390 | * reasons, especially on loaded systems. To avoid clock | |
391 | * drift, we schedule the clock in fixed psi_period intervals. | |
392 | * But the deltas we sample out of the per-cpu buckets above | |
393 | * are based on the actual time elapsing between clock ticks. | |
394 | */ | |
7fc70a39 | 395 | avg_next_update = expires + ((1 + missed_periods) * psi_period); |
bcc78db6 SB |
396 | period = now - (group->avg_last_update + (missed_periods * psi_period)); |
397 | group->avg_last_update = now; | |
eb414681 JW |
398 | |
399 | for (s = 0; s < NR_PSI_STATES - 1; s++) { | |
400 | u32 sample; | |
401 | ||
0e94682b | 402 | sample = group->total[PSI_AVGS][s] - group->avg_total[s]; |
eb414681 JW |
403 | /* |
404 | * Due to the lockless sampling of the time buckets, | |
405 | * recorded time deltas can slip into the next period, | |
406 | * which under full pressure can result in samples in | |
407 | * excess of the period length. | |
408 | * | |
409 | * We don't want to report non-sensical pressures in | |
410 | * excess of 100%, nor do we want to drop such events | |
411 | * on the floor. Instead we punt any overage into the | |
412 | * future until pressure subsides. By doing this we | |
413 | * don't underreport the occurring pressure curve, we | |
414 | * just report it delayed by one period length. | |
415 | * | |
416 | * The error isn't cumulative. As soon as another | |
417 | * delta slips from a period P to P+1, by definition | |
418 | * it frees up its time T in P. | |
419 | */ | |
420 | if (sample > period) | |
421 | sample = period; | |
bcc78db6 | 422 | group->avg_total[s] += sample; |
eb414681 JW |
423 | calc_avgs(group->avg[s], missed_periods, sample, period); |
424 | } | |
7fc70a39 SB |
425 | |
426 | return avg_next_update; | |
eb414681 JW |
427 | } |
428 | ||
bcc78db6 | 429 | static void psi_avgs_work(struct work_struct *work) |
eb414681 JW |
430 | { |
431 | struct delayed_work *dwork; | |
432 | struct psi_group *group; | |
333f3017 | 433 | u32 changed_states; |
eb414681 | 434 | bool nonidle; |
7fc70a39 | 435 | u64 now; |
eb414681 JW |
436 | |
437 | dwork = to_delayed_work(work); | |
bcc78db6 | 438 | group = container_of(dwork, struct psi_group, avgs_work); |
eb414681 | 439 | |
7fc70a39 SB |
440 | mutex_lock(&group->avgs_lock); |
441 | ||
442 | now = sched_clock(); | |
443 | ||
0e94682b | 444 | collect_percpu_times(group, PSI_AVGS, &changed_states); |
333f3017 | 445 | nonidle = changed_states & (1 << PSI_NONIDLE); |
eb414681 JW |
446 | /* |
447 | * If there is task activity, periodically fold the per-cpu | |
448 | * times and feed samples into the running averages. If things | |
449 | * are idle and there is no data to process, stop the clock. | |
450 | * Once restarted, we'll catch up the running averages in one | |
451 | * go - see calc_avgs() and missed_periods. | |
452 | */ | |
7fc70a39 SB |
453 | if (now >= group->avg_next_update) |
454 | group->avg_next_update = update_averages(group, now); | |
eb414681 JW |
455 | |
456 | if (nonidle) { | |
7fc70a39 SB |
457 | schedule_delayed_work(dwork, nsecs_to_jiffies( |
458 | group->avg_next_update - now) + 1); | |
eb414681 | 459 | } |
7fc70a39 SB |
460 | |
461 | mutex_unlock(&group->avgs_lock); | |
eb414681 JW |
462 | } |
463 | ||
3b03706f | 464 | /* Trigger tracking window manipulations */ |
0e94682b SB |
465 | static void window_reset(struct psi_window *win, u64 now, u64 value, |
466 | u64 prev_growth) | |
467 | { | |
468 | win->start_time = now; | |
469 | win->start_value = value; | |
470 | win->prev_growth = prev_growth; | |
471 | } | |
472 | ||
473 | /* | |
474 | * PSI growth tracking window update and growth calculation routine. | |
475 | * | |
476 | * This approximates a sliding tracking window by interpolating | |
477 | * partially elapsed windows using historical growth data from the | |
478 | * previous intervals. This minimizes memory requirements (by not storing | |
479 | * all the intermediate values in the previous window) and simplifies | |
480 | * the calculations. It works well because PSI signal changes only in | |
481 | * positive direction and over relatively small window sizes the growth | |
482 | * is close to linear. | |
483 | */ | |
484 | static u64 window_update(struct psi_window *win, u64 now, u64 value) | |
485 | { | |
486 | u64 elapsed; | |
487 | u64 growth; | |
488 | ||
489 | elapsed = now - win->start_time; | |
490 | growth = value - win->start_value; | |
491 | /* | |
492 | * After each tracking window passes win->start_value and | |
493 | * win->start_time get reset and win->prev_growth stores | |
494 | * the average per-window growth of the previous window. | |
495 | * win->prev_growth is then used to interpolate additional | |
496 | * growth from the previous window assuming it was linear. | |
497 | */ | |
498 | if (elapsed > win->size) | |
499 | window_reset(win, now, value, growth); | |
500 | else { | |
501 | u32 remaining; | |
502 | ||
503 | remaining = win->size - elapsed; | |
c3466952 | 504 | growth += div64_u64(win->prev_growth * remaining, win->size); |
0e94682b SB |
505 | } |
506 | ||
507 | return growth; | |
508 | } | |
509 | ||
510 | static void init_triggers(struct psi_group *group, u64 now) | |
511 | { | |
512 | struct psi_trigger *t; | |
513 | ||
514 | list_for_each_entry(t, &group->triggers, node) | |
515 | window_reset(&t->win, now, | |
516 | group->total[PSI_POLL][t->state], 0); | |
517 | memcpy(group->polling_total, group->total[PSI_POLL], | |
518 | sizeof(group->polling_total)); | |
519 | group->polling_next_update = now + group->poll_min_period; | |
520 | } | |
521 | ||
522 | static u64 update_triggers(struct psi_group *group, u64 now) | |
523 | { | |
524 | struct psi_trigger *t; | |
525 | bool new_stall = false; | |
526 | u64 *total = group->total[PSI_POLL]; | |
527 | ||
528 | /* | |
529 | * On subsequent updates, calculate growth deltas and let | |
530 | * watchers know when their specified thresholds are exceeded. | |
531 | */ | |
532 | list_for_each_entry(t, &group->triggers, node) { | |
533 | u64 growth; | |
534 | ||
535 | /* Check for stall activity */ | |
536 | if (group->polling_total[t->state] == total[t->state]) | |
537 | continue; | |
538 | ||
539 | /* | |
540 | * Multiple triggers might be looking at the same state, | |
541 | * remember to update group->polling_total[] once we've | |
542 | * been through all of them. Also remember to extend the | |
543 | * polling time if we see new stall activity. | |
544 | */ | |
545 | new_stall = true; | |
546 | ||
547 | /* Calculate growth since last update */ | |
548 | growth = window_update(&t->win, now, total[t->state]); | |
549 | if (growth < t->threshold) | |
550 | continue; | |
551 | ||
552 | /* Limit event signaling to once per window */ | |
553 | if (now < t->last_event_time + t->win.size) | |
554 | continue; | |
555 | ||
556 | /* Generate an event */ | |
557 | if (cmpxchg(&t->event, 0, 1) == 0) | |
558 | wake_up_interruptible(&t->event_wait); | |
559 | t->last_event_time = now; | |
560 | } | |
561 | ||
562 | if (new_stall) | |
563 | memcpy(group->polling_total, total, | |
564 | sizeof(group->polling_total)); | |
565 | ||
566 | return now + group->poll_min_period; | |
567 | } | |
568 | ||
461daba0 | 569 | /* Schedule polling if it's not already scheduled. */ |
0e94682b SB |
570 | static void psi_schedule_poll_work(struct psi_group *group, unsigned long delay) |
571 | { | |
461daba0 | 572 | struct task_struct *task; |
0e94682b | 573 | |
461daba0 SB |
574 | /* |
575 | * Do not reschedule if already scheduled. | |
576 | * Possible race with a timer scheduled after this check but before | |
577 | * mod_timer below can be tolerated because group->polling_next_update | |
578 | * will keep updates on schedule. | |
579 | */ | |
580 | if (timer_pending(&group->poll_timer)) | |
0e94682b SB |
581 | return; |
582 | ||
583 | rcu_read_lock(); | |
584 | ||
461daba0 | 585 | task = rcu_dereference(group->poll_task); |
0e94682b SB |
586 | /* |
587 | * kworker might be NULL in case psi_trigger_destroy races with | |
588 | * psi_task_change (hotpath) which can't use locks | |
589 | */ | |
461daba0 SB |
590 | if (likely(task)) |
591 | mod_timer(&group->poll_timer, jiffies + delay); | |
0e94682b SB |
592 | |
593 | rcu_read_unlock(); | |
594 | } | |
595 | ||
461daba0 | 596 | static void psi_poll_work(struct psi_group *group) |
0e94682b | 597 | { |
0e94682b SB |
598 | u32 changed_states; |
599 | u64 now; | |
600 | ||
0e94682b SB |
601 | mutex_lock(&group->trigger_lock); |
602 | ||
603 | now = sched_clock(); | |
604 | ||
605 | collect_percpu_times(group, PSI_POLL, &changed_states); | |
606 | ||
607 | if (changed_states & group->poll_states) { | |
608 | /* Initialize trigger windows when entering polling mode */ | |
609 | if (now > group->polling_until) | |
610 | init_triggers(group, now); | |
611 | ||
612 | /* | |
613 | * Keep the monitor active for at least the duration of the | |
614 | * minimum tracking window as long as monitor states are | |
615 | * changing. | |
616 | */ | |
617 | group->polling_until = now + | |
618 | group->poll_min_period * UPDATES_PER_WINDOW; | |
619 | } | |
620 | ||
621 | if (now > group->polling_until) { | |
622 | group->polling_next_update = ULLONG_MAX; | |
623 | goto out; | |
624 | } | |
625 | ||
626 | if (now >= group->polling_next_update) | |
627 | group->polling_next_update = update_triggers(group, now); | |
628 | ||
629 | psi_schedule_poll_work(group, | |
630 | nsecs_to_jiffies(group->polling_next_update - now) + 1); | |
631 | ||
632 | out: | |
633 | mutex_unlock(&group->trigger_lock); | |
634 | } | |
635 | ||
461daba0 SB |
636 | static int psi_poll_worker(void *data) |
637 | { | |
638 | struct psi_group *group = (struct psi_group *)data; | |
461daba0 | 639 | |
2cca5426 | 640 | sched_set_fifo_low(current); |
461daba0 SB |
641 | |
642 | while (true) { | |
643 | wait_event_interruptible(group->poll_wait, | |
644 | atomic_cmpxchg(&group->poll_wakeup, 1, 0) || | |
645 | kthread_should_stop()); | |
646 | if (kthread_should_stop()) | |
647 | break; | |
648 | ||
649 | psi_poll_work(group); | |
650 | } | |
651 | return 0; | |
652 | } | |
653 | ||
654 | static void poll_timer_fn(struct timer_list *t) | |
655 | { | |
656 | struct psi_group *group = from_timer(group, t, poll_timer); | |
657 | ||
658 | atomic_set(&group->poll_wakeup, 1); | |
659 | wake_up_interruptible(&group->poll_wait); | |
660 | } | |
661 | ||
df774306 | 662 | static void record_times(struct psi_group_cpu *groupc, u64 now) |
eb414681 JW |
663 | { |
664 | u32 delta; | |
eb414681 | 665 | |
eb414681 JW |
666 | delta = now - groupc->state_start; |
667 | groupc->state_start = now; | |
668 | ||
33b2d630 | 669 | if (groupc->state_mask & (1 << PSI_IO_SOME)) { |
eb414681 | 670 | groupc->times[PSI_IO_SOME] += delta; |
33b2d630 | 671 | if (groupc->state_mask & (1 << PSI_IO_FULL)) |
eb414681 JW |
672 | groupc->times[PSI_IO_FULL] += delta; |
673 | } | |
674 | ||
33b2d630 | 675 | if (groupc->state_mask & (1 << PSI_MEM_SOME)) { |
eb414681 | 676 | groupc->times[PSI_MEM_SOME] += delta; |
33b2d630 | 677 | if (groupc->state_mask & (1 << PSI_MEM_FULL)) |
eb414681 | 678 | groupc->times[PSI_MEM_FULL] += delta; |
eb414681 JW |
679 | } |
680 | ||
e7fcd762 | 681 | if (groupc->state_mask & (1 << PSI_CPU_SOME)) { |
eb414681 | 682 | groupc->times[PSI_CPU_SOME] += delta; |
e7fcd762 CZ |
683 | if (groupc->state_mask & (1 << PSI_CPU_FULL)) |
684 | groupc->times[PSI_CPU_FULL] += delta; | |
685 | } | |
eb414681 | 686 | |
33b2d630 | 687 | if (groupc->state_mask & (1 << PSI_NONIDLE)) |
eb414681 JW |
688 | groupc->times[PSI_NONIDLE] += delta; |
689 | } | |
690 | ||
36b238d5 | 691 | static void psi_group_change(struct psi_group *group, int cpu, |
df774306 | 692 | unsigned int clear, unsigned int set, u64 now, |
36b238d5 | 693 | bool wake_clock) |
eb414681 JW |
694 | { |
695 | struct psi_group_cpu *groupc; | |
36b238d5 | 696 | u32 state_mask = 0; |
eb414681 | 697 | unsigned int t, m; |
33b2d630 | 698 | enum psi_states s; |
eb414681 JW |
699 | |
700 | groupc = per_cpu_ptr(group->pcpu, cpu); | |
701 | ||
702 | /* | |
703 | * First we assess the aggregate resource states this CPU's | |
704 | * tasks have been in since the last change, and account any | |
705 | * SOME and FULL time these may have resulted in. | |
706 | * | |
707 | * Then we update the task counts according to the state | |
708 | * change requested through the @clear and @set bits. | |
709 | */ | |
710 | write_seqcount_begin(&groupc->seq); | |
711 | ||
df774306 | 712 | record_times(groupc, now); |
eb414681 JW |
713 | |
714 | for (t = 0, m = clear; m; m &= ~(1 << t), t++) { | |
715 | if (!(m & (1 << t))) | |
716 | continue; | |
9d10a13d CTR |
717 | if (groupc->tasks[t]) { |
718 | groupc->tasks[t]--; | |
719 | } else if (!psi_bug) { | |
221276cb | 720 | printk_deferred(KERN_ERR "psi: task underflow! cpu=%d t=%d tasks=[%u %u %u %u %u] clear=%x set=%x\n", |
eb414681 JW |
721 | cpu, t, groupc->tasks[0], |
722 | groupc->tasks[1], groupc->tasks[2], | |
221276cb BC |
723 | groupc->tasks[3], groupc->tasks[4], |
724 | clear, set); | |
eb414681 JW |
725 | psi_bug = 1; |
726 | } | |
eb414681 JW |
727 | } |
728 | ||
729 | for (t = 0; set; set &= ~(1 << t), t++) | |
730 | if (set & (1 << t)) | |
731 | groupc->tasks[t]++; | |
732 | ||
33b2d630 SB |
733 | /* Calculate state mask representing active states */ |
734 | for (s = 0; s < NR_PSI_STATES; s++) { | |
735 | if (test_state(groupc->tasks, s)) | |
736 | state_mask |= (1 << s); | |
737 | } | |
7fae6c81 CZ |
738 | |
739 | /* | |
740 | * Since we care about lost potential, a memstall is FULL | |
741 | * when there are no other working tasks, but also when | |
742 | * the CPU is actively reclaiming and nothing productive | |
743 | * could run even if it were runnable. So when the current | |
744 | * task in a cgroup is in_memstall, the corresponding groupc | |
745 | * on that cpu is in PSI_MEM_FULL state. | |
746 | */ | |
fddc8bab | 747 | if (unlikely(groupc->tasks[NR_ONCPU] && cpu_curr(cpu)->in_memstall)) |
7fae6c81 CZ |
748 | state_mask |= (1 << PSI_MEM_FULL); |
749 | ||
33b2d630 SB |
750 | groupc->state_mask = state_mask; |
751 | ||
eb414681 | 752 | write_seqcount_end(&groupc->seq); |
0e94682b | 753 | |
36b238d5 JW |
754 | if (state_mask & group->poll_states) |
755 | psi_schedule_poll_work(group, 1); | |
756 | ||
757 | if (wake_clock && !delayed_work_pending(&group->avgs_work)) | |
758 | schedule_delayed_work(&group->avgs_work, PSI_FREQ); | |
eb414681 JW |
759 | } |
760 | ||
2ce7135a JW |
761 | static struct psi_group *iterate_groups(struct task_struct *task, void **iter) |
762 | { | |
3958e2d0 SB |
763 | if (*iter == &psi_system) |
764 | return NULL; | |
765 | ||
2ce7135a | 766 | #ifdef CONFIG_CGROUPS |
3958e2d0 SB |
767 | if (static_branch_likely(&psi_cgroups_enabled)) { |
768 | struct cgroup *cgroup = NULL; | |
2ce7135a | 769 | |
3958e2d0 SB |
770 | if (!*iter) |
771 | cgroup = task->cgroups->dfl_cgrp; | |
772 | else | |
773 | cgroup = cgroup_parent(*iter); | |
2ce7135a | 774 | |
3958e2d0 SB |
775 | if (cgroup && cgroup_parent(cgroup)) { |
776 | *iter = cgroup; | |
777 | return cgroup_psi(cgroup); | |
778 | } | |
2ce7135a | 779 | } |
2ce7135a JW |
780 | #endif |
781 | *iter = &psi_system; | |
782 | return &psi_system; | |
783 | } | |
784 | ||
36b238d5 | 785 | static void psi_flags_change(struct task_struct *task, int clear, int set) |
eb414681 | 786 | { |
eb414681 JW |
787 | if (((task->psi_flags & set) || |
788 | (task->psi_flags & clear) != clear) && | |
789 | !psi_bug) { | |
790 | printk_deferred(KERN_ERR "psi: inconsistent task state! task=%d:%s cpu=%d psi_flags=%x clear=%x set=%x\n", | |
36b238d5 | 791 | task->pid, task->comm, task_cpu(task), |
eb414681 JW |
792 | task->psi_flags, clear, set); |
793 | psi_bug = 1; | |
794 | } | |
795 | ||
796 | task->psi_flags &= ~clear; | |
797 | task->psi_flags |= set; | |
36b238d5 JW |
798 | } |
799 | ||
800 | void psi_task_change(struct task_struct *task, int clear, int set) | |
801 | { | |
802 | int cpu = task_cpu(task); | |
803 | struct psi_group *group; | |
804 | bool wake_clock = true; | |
805 | void *iter = NULL; | |
df774306 | 806 | u64 now; |
36b238d5 JW |
807 | |
808 | if (!task->pid) | |
809 | return; | |
810 | ||
811 | psi_flags_change(task, clear, set); | |
eb414681 | 812 | |
df774306 | 813 | now = cpu_clock(cpu); |
1b69ac6b JW |
814 | /* |
815 | * Periodic aggregation shuts off if there is a period of no | |
816 | * task changes, so we wake it back up if necessary. However, | |
817 | * don't do this if the task change is the aggregation worker | |
818 | * itself going to sleep, or we'll ping-pong forever. | |
819 | */ | |
820 | if (unlikely((clear & TSK_RUNNING) && | |
821 | (task->flags & PF_WQ_WORKER) && | |
bcc78db6 | 822 | wq_worker_last_func(task) == psi_avgs_work)) |
1b69ac6b JW |
823 | wake_clock = false; |
824 | ||
36b238d5 | 825 | while ((group = iterate_groups(task, &iter))) |
df774306 | 826 | psi_group_change(group, cpu, clear, set, now, wake_clock); |
36b238d5 JW |
827 | } |
828 | ||
829 | void psi_task_switch(struct task_struct *prev, struct task_struct *next, | |
830 | bool sleep) | |
831 | { | |
832 | struct psi_group *group, *common = NULL; | |
833 | int cpu = task_cpu(prev); | |
834 | void *iter; | |
df774306 | 835 | u64 now = cpu_clock(cpu); |
36b238d5 JW |
836 | |
837 | if (next->pid) { | |
7fae6c81 CZ |
838 | bool identical_state; |
839 | ||
36b238d5 JW |
840 | psi_flags_change(next, 0, TSK_ONCPU); |
841 | /* | |
7fae6c81 CZ |
842 | * When switching between tasks that have an identical |
843 | * runtime state, the cgroup that contains both tasks | |
844 | * runtime state, the cgroup that contains both tasks | |
845 | * we reach the first common ancestor. Iterate @next's | |
846 | * ancestors only until we encounter @prev's ONCPU. | |
36b238d5 | 847 | */ |
7fae6c81 | 848 | identical_state = prev->psi_flags == next->psi_flags; |
36b238d5 JW |
849 | iter = NULL; |
850 | while ((group = iterate_groups(next, &iter))) { | |
7fae6c81 CZ |
851 | if (identical_state && |
852 | per_cpu_ptr(group->pcpu, cpu)->tasks[NR_ONCPU]) { | |
36b238d5 JW |
853 | common = group; |
854 | break; | |
855 | } | |
856 | ||
df774306 | 857 | psi_group_change(group, cpu, 0, TSK_ONCPU, now, true); |
36b238d5 JW |
858 | } |
859 | } | |
860 | ||
36b238d5 | 861 | if (prev->pid) { |
4117cebf CZ |
862 | int clear = TSK_ONCPU, set = 0; |
863 | ||
864 | /* | |
221276cb BC |
865 | * When we're going to sleep, psi_dequeue() lets us |
866 | * handle TSK_RUNNING, TSK_MEMSTALL_RUNNING and | |
867 | * TSK_IOWAIT here, where we can combine it with | |
868 | * TSK_ONCPU and save walking common ancestors twice. | |
4117cebf CZ |
869 | */ |
870 | if (sleep) { | |
871 | clear |= TSK_RUNNING; | |
221276cb BC |
872 | if (prev->in_memstall) |
873 | clear |= TSK_MEMSTALL_RUNNING; | |
4117cebf CZ |
874 | if (prev->in_iowait) |
875 | set |= TSK_IOWAIT; | |
876 | } | |
877 | ||
878 | psi_flags_change(prev, clear, set); | |
0e94682b | 879 | |
36b238d5 JW |
880 | iter = NULL; |
881 | while ((group = iterate_groups(prev, &iter)) && group != common) | |
df774306 | 882 | psi_group_change(group, cpu, clear, set, now, true); |
4117cebf CZ |
883 | |
884 | /* | |
885 | * TSK_ONCPU is handled up to the common ancestor. If we're tasked | |
886 | * with dequeuing too, finish that for the rest of the hierarchy. | |
887 | */ | |
888 | if (sleep) { | |
889 | clear &= ~TSK_ONCPU; | |
890 | for (; group; group = iterate_groups(prev, &iter)) | |
df774306 | 891 | psi_group_change(group, cpu, clear, set, now, true); |
4117cebf | 892 | } |
1b69ac6b | 893 | } |
eb414681 JW |
894 | } |
895 | ||
eb414681 JW |
896 | /** |
897 | * psi_memstall_enter - mark the beginning of a memory stall section | |
898 | * @flags: flags to handle nested sections | |
899 | * | |
900 | * Marks the calling task as being stalled due to a lack of memory, | |
901 | * such as waiting for a refault or performing reclaim. | |
902 | */ | |
903 | void psi_memstall_enter(unsigned long *flags) | |
904 | { | |
905 | struct rq_flags rf; | |
906 | struct rq *rq; | |
907 | ||
e0c27447 | 908 | if (static_branch_likely(&psi_disabled)) |
eb414681 JW |
909 | return; |
910 | ||
1066d1b6 | 911 | *flags = current->in_memstall; |
eb414681 JW |
912 | if (*flags) |
913 | return; | |
914 | /* | |
1066d1b6 | 915 | * in_memstall setting & accounting needs to be atomic wrt |
eb414681 JW |
916 | * changes to the task's scheduling state, otherwise we can |
917 | * race with CPU migration. | |
918 | */ | |
919 | rq = this_rq_lock_irq(&rf); | |
920 | ||
1066d1b6 | 921 | current->in_memstall = 1; |
221276cb | 922 | psi_task_change(current, 0, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING); |
eb414681 JW |
923 | |
924 | rq_unlock_irq(rq, &rf); | |
925 | } | |
926 | ||
927 | /** | |
928 | * psi_memstall_leave - mark the end of an memory stall section | |
929 | * @flags: flags to handle nested memdelay sections | |
930 | * | |
931 | * Marks the calling task as no longer stalled due to lack of memory. | |
932 | */ | |
933 | void psi_memstall_leave(unsigned long *flags) | |
934 | { | |
935 | struct rq_flags rf; | |
936 | struct rq *rq; | |
937 | ||
e0c27447 | 938 | if (static_branch_likely(&psi_disabled)) |
eb414681 JW |
939 | return; |
940 | ||
941 | if (*flags) | |
942 | return; | |
943 | /* | |
1066d1b6 | 944 | * in_memstall clearing & accounting needs to be atomic wrt |
eb414681 JW |
945 | * changes to the task's scheduling state, otherwise we could |
946 | * race with CPU migration. | |
947 | */ | |
948 | rq = this_rq_lock_irq(&rf); | |
949 | ||
1066d1b6 | 950 | current->in_memstall = 0; |
221276cb | 951 | psi_task_change(current, TSK_MEMSTALL | TSK_MEMSTALL_RUNNING, 0); |
eb414681 JW |
952 | |
953 | rq_unlock_irq(rq, &rf); | |
954 | } | |
955 | ||
2ce7135a JW |
956 | #ifdef CONFIG_CGROUPS |
957 | int psi_cgroup_alloc(struct cgroup *cgroup) | |
958 | { | |
e0c27447 | 959 | if (static_branch_likely(&psi_disabled)) |
2ce7135a JW |
960 | return 0; |
961 | ||
962 | cgroup->psi.pcpu = alloc_percpu(struct psi_group_cpu); | |
963 | if (!cgroup->psi.pcpu) | |
964 | return -ENOMEM; | |
965 | group_init(&cgroup->psi); | |
966 | return 0; | |
967 | } | |
968 | ||
969 | void psi_cgroup_free(struct cgroup *cgroup) | |
970 | { | |
e0c27447 | 971 | if (static_branch_likely(&psi_disabled)) |
2ce7135a JW |
972 | return; |
973 | ||
bcc78db6 | 974 | cancel_delayed_work_sync(&cgroup->psi.avgs_work); |
2ce7135a | 975 | free_percpu(cgroup->psi.pcpu); |
0e94682b SB |
976 | /* All triggers must be removed by now */ |
977 | WARN_ONCE(cgroup->psi.poll_states, "psi: trigger leak\n"); | |
2ce7135a JW |
978 | } |
979 | ||
980 | /** | |
981 | * cgroup_move_task - move task to a different cgroup | |
982 | * @task: the task | |
983 | * @to: the target css_set | |
984 | * | |
985 | * Move task to a new cgroup and safely migrate its associated stall | |
986 | * state between the different groups. | |
987 | * | |
988 | * This function acquires the task's rq lock to lock out concurrent | |
989 | * changes to the task's scheduling state and - in case the task is | |
990 | * running - concurrent changes to its stall state. | |
991 | */ | |
992 | void cgroup_move_task(struct task_struct *task, struct css_set *to) | |
993 | { | |
d583d360 | 994 | unsigned int task_flags; |
2ce7135a JW |
995 | struct rq_flags rf; |
996 | struct rq *rq; | |
997 | ||
e0c27447 | 998 | if (static_branch_likely(&psi_disabled)) { |
8fcb2312 OJ |
999 | /* |
1000 | * Lame to do this here, but the scheduler cannot be locked | |
1001 | * from the outside, so we move cgroups from inside sched/. | |
1002 | */ | |
1003 | rcu_assign_pointer(task->cgroups, to); | |
1004 | return; | |
1005 | } | |
2ce7135a | 1006 | |
8fcb2312 | 1007 | rq = task_rq_lock(task, &rf); |
2ce7135a | 1008 | |
d583d360 JW |
1009 | /* |
1010 | * We may race with schedule() dropping the rq lock between | |
1011 | * deactivating prev and switching to next. Because the psi | |
1012 | * updates from the deactivation are deferred to the switch | |
1013 | * callback to save cgroup tree updates, the task's scheduling | |
1014 | * state here is not coherent with its psi state: | |
1015 | * | |
1016 | * schedule() cgroup_move_task() | |
1017 | * rq_lock() | |
1018 | * deactivate_task() | |
1019 | * p->on_rq = 0 | |
1020 | * psi_dequeue() // defers TSK_RUNNING & TSK_IOWAIT updates | |
1021 | * pick_next_task() | |
1022 | * rq_unlock() | |
1023 | * rq_lock() | |
1024 | * psi_task_change() // old cgroup | |
1025 | * task->cgroups = to | |
1026 | * psi_task_change() // new cgroup | |
1027 | * rq_unlock() | |
1028 | * rq_lock() | |
1029 | * psi_sched_switch() // does deferred updates in new cgroup | |
1030 | * | |
1031 | * Don't rely on the scheduling state. Use psi_flags instead. | |
1032 | */ | |
1033 | task_flags = task->psi_flags; | |
2ce7135a | 1034 | |
8fcb2312 OJ |
1035 | if (task_flags) |
1036 | psi_task_change(task, task_flags, 0); | |
1037 | ||
1038 | /* See comment above */ | |
2ce7135a JW |
1039 | rcu_assign_pointer(task->cgroups, to); |
1040 | ||
8fcb2312 OJ |
1041 | if (task_flags) |
1042 | psi_task_change(task, 0, task_flags); | |
2ce7135a | 1043 | |
8fcb2312 | 1044 | task_rq_unlock(rq, task, &rf); |
2ce7135a JW |
1045 | } |
1046 | #endif /* CONFIG_CGROUPS */ | |
1047 | ||
1048 | int psi_show(struct seq_file *m, struct psi_group *group, enum psi_res res) | |
eb414681 JW |
1049 | { |
1050 | int full; | |
7fc70a39 | 1051 | u64 now; |
eb414681 | 1052 | |
e0c27447 | 1053 | if (static_branch_likely(&psi_disabled)) |
eb414681 JW |
1054 | return -EOPNOTSUPP; |
1055 | ||
7fc70a39 SB |
1056 | /* Update averages before reporting them */ |
1057 | mutex_lock(&group->avgs_lock); | |
1058 | now = sched_clock(); | |
0e94682b | 1059 | collect_percpu_times(group, PSI_AVGS, NULL); |
7fc70a39 SB |
1060 | if (now >= group->avg_next_update) |
1061 | group->avg_next_update = update_averages(group, now); | |
1062 | mutex_unlock(&group->avgs_lock); | |
eb414681 | 1063 | |
e7fcd762 | 1064 | for (full = 0; full < 2; full++) { |
eb414681 JW |
1065 | unsigned long avg[3]; |
1066 | u64 total; | |
1067 | int w; | |
1068 | ||
1069 | for (w = 0; w < 3; w++) | |
1070 | avg[w] = group->avg[res * 2 + full][w]; | |
0e94682b SB |
1071 | total = div_u64(group->total[PSI_AVGS][res * 2 + full], |
1072 | NSEC_PER_USEC); | |
eb414681 JW |
1073 | |
1074 | seq_printf(m, "%s avg10=%lu.%02lu avg60=%lu.%02lu avg300=%lu.%02lu total=%llu\n", | |
1075 | full ? "full" : "some", | |
1076 | LOAD_INT(avg[0]), LOAD_FRAC(avg[0]), | |
1077 | LOAD_INT(avg[1]), LOAD_FRAC(avg[1]), | |
1078 | LOAD_INT(avg[2]), LOAD_FRAC(avg[2]), | |
1079 | total); | |
1080 | } | |
1081 | ||
1082 | return 0; | |
1083 | } | |
1084 | ||
0e94682b SB |
1085 | struct psi_trigger *psi_trigger_create(struct psi_group *group, |
1086 | char *buf, size_t nbytes, enum psi_res res) | |
1087 | { | |
1088 | struct psi_trigger *t; | |
1089 | enum psi_states state; | |
1090 | u32 threshold_us; | |
1091 | u32 window_us; | |
1092 | ||
1093 | if (static_branch_likely(&psi_disabled)) | |
1094 | return ERR_PTR(-EOPNOTSUPP); | |
1095 | ||
1096 | if (sscanf(buf, "some %u %u", &threshold_us, &window_us) == 2) | |
1097 | state = PSI_IO_SOME + res * 2; | |
1098 | else if (sscanf(buf, "full %u %u", &threshold_us, &window_us) == 2) | |
1099 | state = PSI_IO_FULL + res * 2; | |
1100 | else | |
1101 | return ERR_PTR(-EINVAL); | |
1102 | ||
1103 | if (state >= PSI_NONIDLE) | |
1104 | return ERR_PTR(-EINVAL); | |
1105 | ||
1106 | if (window_us < WINDOW_MIN_US || | |
1107 | window_us > WINDOW_MAX_US) | |
1108 | return ERR_PTR(-EINVAL); | |
1109 | ||
1110 | /* Check threshold */ | |
1111 | if (threshold_us == 0 || threshold_us > window_us) | |
1112 | return ERR_PTR(-EINVAL); | |
1113 | ||
1114 | t = kmalloc(sizeof(*t), GFP_KERNEL); | |
1115 | if (!t) | |
1116 | return ERR_PTR(-ENOMEM); | |
1117 | ||
1118 | t->group = group; | |
1119 | t->state = state; | |
1120 | t->threshold = threshold_us * NSEC_PER_USEC; | |
1121 | t->win.size = window_us * NSEC_PER_USEC; | |
1122 | window_reset(&t->win, 0, 0, 0); | |
1123 | ||
1124 | t->event = 0; | |
1125 | t->last_event_time = 0; | |
1126 | init_waitqueue_head(&t->event_wait); | |
0e94682b SB |
1127 | |
1128 | mutex_lock(&group->trigger_lock); | |
1129 | ||
461daba0 SB |
1130 | if (!rcu_access_pointer(group->poll_task)) { |
1131 | struct task_struct *task; | |
0e94682b | 1132 | |
461daba0 SB |
1133 | task = kthread_create(psi_poll_worker, group, "psimon"); |
1134 | if (IS_ERR(task)) { | |
0e94682b SB |
1135 | kfree(t); |
1136 | mutex_unlock(&group->trigger_lock); | |
461daba0 | 1137 | return ERR_CAST(task); |
0e94682b | 1138 | } |
461daba0 | 1139 | atomic_set(&group->poll_wakeup, 0); |
461daba0 | 1140 | wake_up_process(task); |
461daba0 | 1141 | rcu_assign_pointer(group->poll_task, task); |
0e94682b SB |
1142 | } |
1143 | ||
1144 | list_add(&t->node, &group->triggers); | |
1145 | group->poll_min_period = min(group->poll_min_period, | |
1146 | div_u64(t->win.size, UPDATES_PER_WINDOW)); | |
1147 | group->nr_triggers[t->state]++; | |
1148 | group->poll_states |= (1 << t->state); | |
1149 | ||
1150 | mutex_unlock(&group->trigger_lock); | |
1151 | ||
1152 | return t; | |
1153 | } | |
1154 | ||
cccc8aa4 | 1155 | void psi_trigger_destroy(struct psi_trigger *t) |
0e94682b | 1156 | { |
cccc8aa4 | 1157 | struct psi_group *group; |
461daba0 | 1158 | struct task_struct *task_to_destroy = NULL; |
0e94682b | 1159 | |
cccc8aa4 SB |
1160 | /* |
1161 | * We do not check psi_disabled since it might have been disabled after | |
1162 | * the trigger got created. | |
1163 | */ | |
1164 | if (!t) | |
0e94682b SB |
1165 | return; |
1166 | ||
cccc8aa4 | 1167 | group = t->group; |
0e94682b SB |
1168 | /* |
1169 | * Wakeup waiters to stop polling. Can happen if cgroup is deleted | |
1170 | * from under a polling process. | |
1171 | */ | |
1172 | wake_up_interruptible(&t->event_wait); | |
1173 | ||
1174 | mutex_lock(&group->trigger_lock); | |
1175 | ||
1176 | if (!list_empty(&t->node)) { | |
1177 | struct psi_trigger *tmp; | |
1178 | u64 period = ULLONG_MAX; | |
1179 | ||
1180 | list_del(&t->node); | |
1181 | group->nr_triggers[t->state]--; | |
1182 | if (!group->nr_triggers[t->state]) | |
1183 | group->poll_states &= ~(1 << t->state); | |
1184 | /* reset min update period for the remaining triggers */ | |
1185 | list_for_each_entry(tmp, &group->triggers, node) | |
1186 | period = min(period, div_u64(tmp->win.size, | |
1187 | UPDATES_PER_WINDOW)); | |
1188 | group->poll_min_period = period; | |
461daba0 | 1189 | /* Destroy poll_task when the last trigger is destroyed */ |
0e94682b SB |
1190 | if (group->poll_states == 0) { |
1191 | group->polling_until = 0; | |
461daba0 SB |
1192 | task_to_destroy = rcu_dereference_protected( |
1193 | group->poll_task, | |
0e94682b | 1194 | lockdep_is_held(&group->trigger_lock)); |
461daba0 | 1195 | rcu_assign_pointer(group->poll_task, NULL); |
8f91efd8 | 1196 | del_timer(&group->poll_timer); |
0e94682b SB |
1197 | } |
1198 | } | |
1199 | ||
1200 | mutex_unlock(&group->trigger_lock); | |
1201 | ||
1202 | /* | |
cccc8aa4 SB |
1203 | * Wait for psi_schedule_poll_work RCU to complete its read-side |
1204 | * critical section before destroying the trigger and optionally the | |
1205 | * poll_task. | |
0e94682b SB |
1206 | */ |
1207 | synchronize_rcu(); | |
1208 | /* | |
8f91efd8 | 1209 | * Stop kthread 'psimon' after releasing trigger_lock to prevent a |
0e94682b SB |
1210 | * deadlock while waiting for psi_poll_work to acquire trigger_lock |
1211 | */ | |
461daba0 | 1212 | if (task_to_destroy) { |
7b2b55da JX |
1213 | /* |
1214 | * After the RCU grace period has expired, the worker | |
461daba0 | 1215 | * can no longer be found through group->poll_task. |
7b2b55da | 1216 | */ |
461daba0 | 1217 | kthread_stop(task_to_destroy); |
0e94682b SB |
1218 | } |
1219 | kfree(t); | |
1220 | } | |
1221 | ||
0e94682b SB |
1222 | __poll_t psi_trigger_poll(void **trigger_ptr, |
1223 | struct file *file, poll_table *wait) | |
1224 | { | |
1225 | __poll_t ret = DEFAULT_POLLMASK; | |
1226 | struct psi_trigger *t; | |
1227 | ||
1228 | if (static_branch_likely(&psi_disabled)) | |
1229 | return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI; | |
1230 | ||
cccc8aa4 SB |
1231 | t = smp_load_acquire(trigger_ptr); |
1232 | if (!t) | |
0e94682b | 1233 | return DEFAULT_POLLMASK | EPOLLERR | EPOLLPRI; |
0e94682b SB |
1234 | |
1235 | poll_wait(file, &t->event_wait, wait); | |
1236 | ||
1237 | if (cmpxchg(&t->event, 1, 0) == 1) | |
1238 | ret |= EPOLLPRI; | |
1239 | ||
0e94682b SB |
1240 | return ret; |
1241 | } | |
1242 | ||
f1c2eef5 SB |
1243 | #ifdef CONFIG_PROC_FS |
1244 | static int psi_io_show(struct seq_file *m, void *v) | |
1245 | { | |
1246 | return psi_show(m, &psi_system, PSI_IO); | |
1247 | } | |
1248 | ||
1249 | static int psi_memory_show(struct seq_file *m, void *v) | |
1250 | { | |
1251 | return psi_show(m, &psi_system, PSI_MEM); | |
1252 | } | |
1253 | ||
1254 | static int psi_cpu_show(struct seq_file *m, void *v) | |
1255 | { | |
1256 | return psi_show(m, &psi_system, PSI_CPU); | |
1257 | } | |
1258 | ||
1259 | static int psi_open(struct file *file, int (*psi_show)(struct seq_file *, void *)) | |
1260 | { | |
1261 | if (file->f_mode & FMODE_WRITE && !capable(CAP_SYS_RESOURCE)) | |
1262 | return -EPERM; | |
1263 | ||
1264 | return single_open(file, psi_show, NULL); | |
1265 | } | |
1266 | ||
1267 | static int psi_io_open(struct inode *inode, struct file *file) | |
1268 | { | |
1269 | return psi_open(file, psi_io_show); | |
1270 | } | |
1271 | ||
1272 | static int psi_memory_open(struct inode *inode, struct file *file) | |
1273 | { | |
1274 | return psi_open(file, psi_memory_show); | |
1275 | } | |
1276 | ||
1277 | static int psi_cpu_open(struct inode *inode, struct file *file) | |
1278 | { | |
1279 | return psi_open(file, psi_cpu_show); | |
1280 | } | |
1281 | ||
0e94682b SB |
1282 | static ssize_t psi_write(struct file *file, const char __user *user_buf, |
1283 | size_t nbytes, enum psi_res res) | |
1284 | { | |
1285 | char buf[32]; | |
1286 | size_t buf_size; | |
1287 | struct seq_file *seq; | |
1288 | struct psi_trigger *new; | |
1289 | ||
1290 | if (static_branch_likely(&psi_disabled)) | |
1291 | return -EOPNOTSUPP; | |
1292 | ||
6fcca0fa SB |
1293 | if (!nbytes) |
1294 | return -EINVAL; | |
1295 | ||
4adcdcea | 1296 | buf_size = min(nbytes, sizeof(buf)); |
0e94682b SB |
1297 | if (copy_from_user(buf, user_buf, buf_size)) |
1298 | return -EFAULT; | |
1299 | ||
1300 | buf[buf_size - 1] = '\0'; | |
1301 | ||
0e94682b | 1302 | seq = file->private_data; |
cccc8aa4 | 1303 | |
0e94682b SB |
1304 | /* Take seq->lock to protect seq->private from concurrent writes */ |
1305 | mutex_lock(&seq->lock); | |
cccc8aa4 SB |
1306 | |
1307 | /* Allow only one trigger per file descriptor */ | |
1308 | if (seq->private) { | |
1309 | mutex_unlock(&seq->lock); | |
1310 | return -EBUSY; | |
1311 | } | |
1312 | ||
1313 | new = psi_trigger_create(&psi_system, buf, nbytes, res); | |
1314 | if (IS_ERR(new)) { | |
1315 | mutex_unlock(&seq->lock); | |
1316 | return PTR_ERR(new); | |
1317 | } | |
1318 | ||
1319 | smp_store_release(&seq->private, new); | |
0e94682b SB |
1320 | mutex_unlock(&seq->lock); |
1321 | ||
1322 | return nbytes; | |
1323 | } | |
1324 | ||
1325 | static ssize_t psi_io_write(struct file *file, const char __user *user_buf, | |
1326 | size_t nbytes, loff_t *ppos) | |
1327 | { | |
1328 | return psi_write(file, user_buf, nbytes, PSI_IO); | |
1329 | } | |
1330 | ||
1331 | static ssize_t psi_memory_write(struct file *file, const char __user *user_buf, | |
1332 | size_t nbytes, loff_t *ppos) | |
1333 | { | |
1334 | return psi_write(file, user_buf, nbytes, PSI_MEM); | |
1335 | } | |
1336 | ||
1337 | static ssize_t psi_cpu_write(struct file *file, const char __user *user_buf, | |
1338 | size_t nbytes, loff_t *ppos) | |
1339 | { | |
1340 | return psi_write(file, user_buf, nbytes, PSI_CPU); | |
1341 | } | |
1342 | ||
1343 | static __poll_t psi_fop_poll(struct file *file, poll_table *wait) | |
1344 | { | |
1345 | struct seq_file *seq = file->private_data; | |
1346 | ||
1347 | return psi_trigger_poll(&seq->private, file, wait); | |
1348 | } | |
1349 | ||
1350 | static int psi_fop_release(struct inode *inode, struct file *file) | |
1351 | { | |
1352 | struct seq_file *seq = file->private_data; | |
1353 | ||
cccc8aa4 | 1354 | psi_trigger_destroy(seq->private); |
0e94682b SB |
1355 | return single_release(inode, file); |
1356 | } | |
1357 | ||
97a32539 AD |
1358 | static const struct proc_ops psi_io_proc_ops = { |
1359 | .proc_open = psi_io_open, | |
1360 | .proc_read = seq_read, | |
1361 | .proc_lseek = seq_lseek, | |
1362 | .proc_write = psi_io_write, | |
1363 | .proc_poll = psi_fop_poll, | |
1364 | .proc_release = psi_fop_release, | |
eb414681 JW |
1365 | }; |
1366 | ||
97a32539 AD |
1367 | static const struct proc_ops psi_memory_proc_ops = { |
1368 | .proc_open = psi_memory_open, | |
1369 | .proc_read = seq_read, | |
1370 | .proc_lseek = seq_lseek, | |
1371 | .proc_write = psi_memory_write, | |
1372 | .proc_poll = psi_fop_poll, | |
1373 | .proc_release = psi_fop_release, | |
eb414681 JW |
1374 | }; |
1375 | ||
97a32539 AD |
1376 | static const struct proc_ops psi_cpu_proc_ops = { |
1377 | .proc_open = psi_cpu_open, | |
1378 | .proc_read = seq_read, | |
1379 | .proc_lseek = seq_lseek, | |
1380 | .proc_write = psi_cpu_write, | |
1381 | .proc_poll = psi_fop_poll, | |
1382 | .proc_release = psi_fop_release, | |
eb414681 JW |
1383 | }; |
1384 | ||
1385 | static int __init psi_proc_init(void) | |
1386 | { | |
3d817689 WL |
1387 | if (psi_enable) { |
1388 | proc_mkdir("pressure", NULL); | |
6db12ee0 JH |
1389 | proc_create("pressure/io", 0666, NULL, &psi_io_proc_ops); |
1390 | proc_create("pressure/memory", 0666, NULL, &psi_memory_proc_ops); | |
1391 | proc_create("pressure/cpu", 0666, NULL, &psi_cpu_proc_ops); | |
3d817689 | 1392 | } |
eb414681 JW |
1393 | return 0; |
1394 | } | |
1395 | module_init(psi_proc_init); | |
f1c2eef5 SB |
1396 | |
1397 | #endif /* CONFIG_PROC_FS */ |