tree: git://git.cmpxchg.org/linux-mmotm.git master
head: fa5441daae8ad99af4e198bcd4d57cffdd582961
commit: 60ad478b408f38c98f4df24d1cd87f5b3f130831 [170/329] psi: pressure stall information for CPU, memory, and IO
reproduce:
# apt-get install sparse
git checkout 60ad478b408f38c98f4df24d1cd87f5b3f130831
make ARCH=x86_64 allmodconfig
make C=1 CF=-D__CHECK_ENDIAN__
sparse warnings: (new ones prefixed by >>)
>> kernel/sched/psi.c:153:18: sparse: incorrect type in initializer (different address spaces) @@ expected struct psi_group_cpu *cpus @@ got struct psi_group_struct psi_group_cpu *cpus @@
kernel/sched/psi.c:153:18: expected struct psi_group_cpu *cpus
kernel/sched/psi.c:153:18: got struct psi_group_cpu [noderef] <asn:3>*<noident>
>> kernel/sched/psi.c:219:48: sparse: incorrect type in initializer (different address spaces) @@ expected void const [noderef] <asn:3>*__vpp_verify @@ got nst [noderef] <asn:3>*__vpp_verify @@
kernel/sched/psi.c:219:48: expected void const [noderef] <asn:3>*__vpp_verify
kernel/sched/psi.c:219:48: got struct psi_group_cpu *<noident>
>> kernel/sched/psi.c:255:17: sparse: expression using sizeof(void)
kernel/sched/psi.c:256:17: sparse: expression using sizeof(void)
kernel/sched/psi.c:338:18: sparse: incorrect type in initializer (different address spaces) @@ expected void const [noderef] <asn:3>*__vpp_verify @@ got nst [noderef] <asn:3>*__vpp_verify @@
kernel/sched/psi.c:338:18: expected void const [noderef] <asn:3>*__vpp_verify
kernel/sched/psi.c:338:18: got struct psi_group_cpu *<noident>
vim +153 kernel/sched/psi.c
149
150 /* System-level pressure and stall tracking */
151 static DEFINE_PER_CPU(struct psi_group_cpu, system_group_cpus);
152 static struct psi_group psi_system = {
> 153 .cpus = &system_group_cpus,
154 };
155
156 static void psi_clock(struct work_struct *work);
157
158 static void psi_group_init(struct psi_group *group)
159 {
160 group->period_expires = jiffies + PSI_FREQ;
161 INIT_DELAYED_WORK(&group->clock_work, psi_clock);
162 mutex_init(&group->stat_lock);
163 }
164
165 void __init psi_init(void)
166 {
167 if (psi_disabled)
168 return;
169
170 psi_period = jiffies_to_nsecs(PSI_FREQ);
171 psi_group_init(&psi_system);
172 }
173
174 static void calc_avgs(unsigned long avg[3], u64 time, int missed_periods)
175 {
176 unsigned long pct;
177
178 /* Sample the most recent active period */
179 pct = time * 100 / psi_period;
180 pct *= FIXED_1;
181 avg[0] = calc_load(avg[0], EXP_10s, pct);
182 avg[1] = calc_load(avg[1], EXP_60s, pct);
183 avg[2] = calc_load(avg[2], EXP_300s, pct);
184
185 /* Fill in zeroes for periods of no activity */
186 if (missed_periods) {
187 avg[0] = calc_load_n(avg[0], EXP_10s, 0, missed_periods);
188 avg[1] = calc_load_n(avg[1], EXP_60s, 0, missed_periods);
189 avg[2] = calc_load_n(avg[2], EXP_300s, 0, missed_periods);
190 }
191 }
192
193 static bool psi_update_stats(struct psi_group *group)
194 {
195 u64 some[NR_PSI_RESOURCES] = { 0, };
196 u64 full[NR_PSI_RESOURCES] = { 0, };
197 unsigned long nonidle_total = 0;
198 unsigned long missed_periods;
199 unsigned long expires;
200 int cpu;
201 int r;
202
203 mutex_lock(&group->stat_lock);
204
205 /*
206 * Collect the per-cpu time buckets and average them into a
207 * single time sample that is normalized to wallclock time.
208 *
209 * For averaging, each CPU is weighted by its non-idle time in
210 * the sampling period. This eliminates artifacts from uneven
211 * loading, or even entirely idle CPUs.
212 *
213 * We could pin the online CPUs here, but the noise introduced
214 * by missing up to one sample period from CPUs that are going
215 * away shouldn't matter in practice - just like the noise of
216 * previously offlined CPUs returning with a non-zero sample.
217 */
218 for_each_online_cpu(cpu) {
> 219 struct psi_group_cpu *groupc = per_cpu_ptr(group->cpus, cpu);
220 unsigned long nonidle;
221
222 if (!groupc->nonidle_time)
223 continue;
224
225 nonidle = nsecs_to_jiffies(groupc->nonidle_time);
226 groupc->nonidle_time = 0;
227 nonidle_total += nonidle;
228
229 for (r = 0; r < NR_PSI_RESOURCES; r++) {
230 struct psi_resource *res = &groupc->res[r];
231
232 some[r] += (res->times[0] + res->times[1]) * nonidle;
233 full[r] += res->times[1] * nonidle;
234
235 /* It's racy, but we can tolerate some error */
236 res->times[0] = 0;
237 res->times[1] = 0;
238 }
239 }
240
241 /*
242 * Integrate the sample into the running statistics that are
243 * reported to userspace: the cumulative stall times and the
244 * decaying averages.
245 *
246 * Pressure percentages are sampled at PSI_FREQ. We might be
247 * called more often when the user polls more frequently than
248 * that; we might be called less often when there is no task
249 * activity, thus no data, and clock ticks are sporadic. The
250 * below handles both.
251 */
252
253 /* total= */
254 for (r = 0; r < NR_PSI_RESOURCES; r++) {
> 255 do_div(some[r], max(nonidle_total, 1UL));
256 do_div(full[r], max(nonidle_total, 1UL));
257
258 group->some[r] += some[r];
259 group->full[r] += full[r];
260 }
261
262 /* avgX= */
263 expires = group->period_expires;
264 if (time_before(jiffies, expires))
265 goto out;
266
267 missed_periods = (jiffies - expires) / PSI_FREQ;
268 group->period_expires = expires + ((1 + missed_periods) * PSI_FREQ);
269
270 for (r = 0; r < NR_PSI_RESOURCES; r++) {
271 u64 some, full;
272
273 some = group->some[r] - group->last_some[r];
274 full = group->full[r] - group->last_full[r];
275
276 calc_avgs(group->avg_some[r], some, missed_periods);
277 calc_avgs(group->avg_full[r], full, missed_periods);
278
279 group->last_some[r] = group->some[r];
280 group->last_full[r] = group->full[r];
281 }
282 out:
283 mutex_unlock(&group->stat_lock);
284 return nonidle_total;
285 }
286
---
0-DAY kernel test infrastructure Open Source Technology Center
https://lists.01.org/pipermail/kbuild-all Intel Corporation
