]>
Commit | Line | Data |
---|---|---|
fb1afad0 GKH |
1 | From 7d5905dc14a87805a59f3c5bf70173aac2bb18f8 Mon Sep 17 00:00:00 2001 |
2 | From: "Rafael J. Wysocki" <rafael.j.wysocki@intel.com> | |
3 | Date: Wed, 15 Nov 2017 02:13:40 +0100 | |
4 | Subject: x86 / CPU: Always show current CPU frequency in /proc/cpuinfo | |
5 | ||
6 | From: Rafael J. Wysocki <rafael.j.wysocki@intel.com> | |
7 | ||
8 | commit 7d5905dc14a87805a59f3c5bf70173aac2bb18f8 upstream. | |
9 | ||
10 | After commit 890da9cf0983 (Revert "x86: do not use cpufreq_quick_get() | |
11 | for /proc/cpuinfo "cpu MHz"") the "cpu MHz" number in /proc/cpuinfo | |
12 | on x86 can be either the nominal CPU frequency (which is constant) | |
13 | or the frequency most recently requested by a scaling governor in | |
14 | cpufreq, depending on the cpufreq configuration. That is somewhat | |
15 | inconsistent and is different from what it was before 4.13, so in | |
16 | order to restore the previous behavior, make it report the current | |
17 | CPU frequency like the scaling_cur_freq sysfs file in cpufreq. | |
18 | ||
19 | To that end, modify the /proc/cpuinfo implementation on x86 to use | |
20 | aperfmperf_snapshot_khz() to snapshot the APERF and MPERF feedback | |
21 | registers, if available, and use their values to compute the CPU | |
22 | frequency to be reported as "cpu MHz". | |
23 | ||
24 | However, do that carefully enough to avoid accumulating delays that | |
25 | lead to unacceptable access times for /proc/cpuinfo on systems with | |
26 | many CPUs. Run aperfmperf_snapshot_khz() once on all CPUs | |
27 | asynchronously at the /proc/cpuinfo open time, add a single delay | |
28 | upfront (if necessary) at that point and simply compute the current | |
29 | frequency while running show_cpuinfo() for each individual CPU. | |
30 | ||
31 | Also, to avoid slowing down /proc/cpuinfo accesses too much, reduce | |
32 | the default delay between consecutive APERF and MPERF reads to 10 ms, | |
33 | which should be sufficient to get large enough numbers for the | |
34 | frequency computation in all cases. | |
35 | ||
36 | Fixes: 890da9cf0983 (Revert "x86: do not use cpufreq_quick_get() for /proc/cpuinfo "cpu MHz"") | |
37 | Signed-off-by: Rafael J. Wysocki <rafael.j.wysocki@intel.com> | |
38 | Acked-by: Thomas Gleixner <tglx@linutronix.de> | |
39 | Tested-by: Thomas Gleixner <tglx@linutronix.de> | |
40 | Acked-by: Ingo Molnar <mingo@kernel.org> | |
41 | Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org> | |
42 | ||
43 | --- | |
44 | arch/x86/kernel/cpu/Makefile | 2 - | |
45 | arch/x86/kernel/cpu/aperfmperf.c | 74 +++++++++++++++++++++++++++------------ | |
46 | arch/x86/kernel/cpu/cpu.h | 3 + | |
47 | arch/x86/kernel/cpu/proc.c | 6 ++- | |
48 | fs/proc/cpuinfo.c | 6 +++ | |
49 | include/linux/cpufreq.h | 1 | |
50 | 6 files changed, 68 insertions(+), 24 deletions(-) | |
51 | ||
52 | --- a/arch/x86/kernel/cpu/Makefile | |
53 | +++ b/arch/x86/kernel/cpu/Makefile | |
54 | @@ -22,7 +22,7 @@ obj-y += common.o | |
55 | obj-y += rdrand.o | |
56 | obj-y += match.o | |
57 | obj-y += bugs.o | |
58 | -obj-$(CONFIG_CPU_FREQ) += aperfmperf.o | |
59 | +obj-y += aperfmperf.o | |
60 | obj-y += cpuid-deps.o | |
61 | ||
62 | obj-$(CONFIG_PROC_FS) += proc.o | |
63 | --- a/arch/x86/kernel/cpu/aperfmperf.c | |
64 | +++ b/arch/x86/kernel/cpu/aperfmperf.c | |
65 | @@ -14,6 +14,8 @@ | |
66 | #include <linux/percpu.h> | |
67 | #include <linux/smp.h> | |
68 | ||
69 | +#include "cpu.h" | |
70 | + | |
71 | struct aperfmperf_sample { | |
72 | unsigned int khz; | |
73 | ktime_t time; | |
74 | @@ -24,7 +26,7 @@ struct aperfmperf_sample { | |
75 | static DEFINE_PER_CPU(struct aperfmperf_sample, samples); | |
76 | ||
77 | #define APERFMPERF_CACHE_THRESHOLD_MS 10 | |
78 | -#define APERFMPERF_REFRESH_DELAY_MS 20 | |
79 | +#define APERFMPERF_REFRESH_DELAY_MS 10 | |
80 | #define APERFMPERF_STALE_THRESHOLD_MS 1000 | |
81 | ||
82 | /* | |
83 | @@ -38,8 +40,6 @@ static void aperfmperf_snapshot_khz(void | |
84 | u64 aperf, aperf_delta; | |
85 | u64 mperf, mperf_delta; | |
86 | struct aperfmperf_sample *s = this_cpu_ptr(&samples); | |
87 | - ktime_t now = ktime_get(); | |
88 | - s64 time_delta = ktime_ms_delta(now, s->time); | |
89 | unsigned long flags; | |
90 | ||
91 | local_irq_save(flags); | |
92 | @@ -57,38 +57,68 @@ static void aperfmperf_snapshot_khz(void | |
93 | if (mperf_delta == 0) | |
94 | return; | |
95 | ||
96 | - s->time = now; | |
97 | + s->time = ktime_get(); | |
98 | s->aperf = aperf; | |
99 | s->mperf = mperf; | |
100 | - | |
101 | - /* If the previous iteration was too long ago, discard it. */ | |
102 | - if (time_delta > APERFMPERF_STALE_THRESHOLD_MS) | |
103 | - s->khz = 0; | |
104 | - else | |
105 | - s->khz = div64_u64((cpu_khz * aperf_delta), mperf_delta); | |
106 | + s->khz = div64_u64((cpu_khz * aperf_delta), mperf_delta); | |
107 | } | |
108 | ||
109 | -unsigned int arch_freq_get_on_cpu(int cpu) | |
110 | +static bool aperfmperf_snapshot_cpu(int cpu, ktime_t now, bool wait) | |
111 | { | |
112 | - s64 time_delta; | |
113 | - unsigned int khz; | |
114 | + s64 time_delta = ktime_ms_delta(now, per_cpu(samples.time, cpu)); | |
115 | + | |
116 | + /* Don't bother re-computing within the cache threshold time. */ | |
117 | + if (time_delta < APERFMPERF_CACHE_THRESHOLD_MS) | |
118 | + return true; | |
119 | + | |
120 | + smp_call_function_single(cpu, aperfmperf_snapshot_khz, NULL, wait); | |
121 | + | |
122 | + /* Return false if the previous iteration was too long ago. */ | |
123 | + return time_delta <= APERFMPERF_STALE_THRESHOLD_MS; | |
124 | +} | |
125 | ||
126 | +unsigned int aperfmperf_get_khz(int cpu) | |
127 | +{ | |
128 | if (!cpu_khz) | |
129 | return 0; | |
130 | ||
131 | if (!static_cpu_has(X86_FEATURE_APERFMPERF)) | |
132 | return 0; | |
133 | ||
134 | - /* Don't bother re-computing within the cache threshold time. */ | |
135 | - time_delta = ktime_ms_delta(ktime_get(), per_cpu(samples.time, cpu)); | |
136 | - khz = per_cpu(samples.khz, cpu); | |
137 | - if (khz && time_delta < APERFMPERF_CACHE_THRESHOLD_MS) | |
138 | - return khz; | |
139 | + aperfmperf_snapshot_cpu(cpu, ktime_get(), true); | |
140 | + return per_cpu(samples.khz, cpu); | |
141 | +} | |
142 | ||
143 | - smp_call_function_single(cpu, aperfmperf_snapshot_khz, NULL, 1); | |
144 | - khz = per_cpu(samples.khz, cpu); | |
145 | - if (khz) | |
146 | - return khz; | |
147 | +void arch_freq_prepare_all(void) | |
148 | +{ | |
149 | + ktime_t now = ktime_get(); | |
150 | + bool wait = false; | |
151 | + int cpu; | |
152 | + | |
153 | + if (!cpu_khz) | |
154 | + return; | |
155 | + | |
156 | + if (!static_cpu_has(X86_FEATURE_APERFMPERF)) | |
157 | + return; | |
158 | + | |
159 | + for_each_online_cpu(cpu) | |
160 | + if (!aperfmperf_snapshot_cpu(cpu, now, false)) | |
161 | + wait = true; | |
162 | + | |
163 | + if (wait) | |
164 | + msleep(APERFMPERF_REFRESH_DELAY_MS); | |
165 | +} | |
166 | + | |
167 | +unsigned int arch_freq_get_on_cpu(int cpu) | |
168 | +{ | |
169 | + if (!cpu_khz) | |
170 | + return 0; | |
171 | + | |
172 | + if (!static_cpu_has(X86_FEATURE_APERFMPERF)) | |
173 | + return 0; | |
174 | + | |
175 | + if (aperfmperf_snapshot_cpu(cpu, ktime_get(), true)) | |
176 | + return per_cpu(samples.khz, cpu); | |
177 | ||
178 | msleep(APERFMPERF_REFRESH_DELAY_MS); | |
179 | smp_call_function_single(cpu, aperfmperf_snapshot_khz, NULL, 1); | |
180 | --- a/arch/x86/kernel/cpu/cpu.h | |
181 | +++ b/arch/x86/kernel/cpu/cpu.h | |
182 | @@ -47,4 +47,7 @@ extern const struct cpu_dev *const __x86 | |
183 | ||
184 | extern void get_cpu_cap(struct cpuinfo_x86 *c); | |
185 | extern void cpu_detect_cache_sizes(struct cpuinfo_x86 *c); | |
186 | + | |
187 | +unsigned int aperfmperf_get_khz(int cpu); | |
188 | + | |
189 | #endif /* ARCH_X86_CPU_H */ | |
190 | --- a/arch/x86/kernel/cpu/proc.c | |
191 | +++ b/arch/x86/kernel/cpu/proc.c | |
192 | @@ -5,6 +5,8 @@ | |
193 | #include <linux/seq_file.h> | |
194 | #include <linux/cpufreq.h> | |
195 | ||
196 | +#include "cpu.h" | |
197 | + | |
198 | /* | |
199 | * Get CPU information for use by the procfs. | |
200 | */ | |
201 | @@ -78,9 +80,11 @@ static int show_cpuinfo(struct seq_file | |
202 | seq_printf(m, "microcode\t: 0x%x\n", c->microcode); | |
203 | ||
204 | if (cpu_has(c, X86_FEATURE_TSC)) { | |
205 | - unsigned int freq = cpufreq_quick_get(cpu); | |
206 | + unsigned int freq = aperfmperf_get_khz(cpu); | |
207 | ||
208 | if (!freq) | |
209 | + freq = cpufreq_quick_get(cpu); | |
210 | + if (!freq) | |
211 | freq = cpu_khz; | |
212 | seq_printf(m, "cpu MHz\t\t: %u.%03u\n", | |
213 | freq / 1000, (freq % 1000)); | |
214 | --- a/fs/proc/cpuinfo.c | |
215 | +++ b/fs/proc/cpuinfo.c | |
216 | @@ -1,12 +1,18 @@ | |
217 | // SPDX-License-Identifier: GPL-2.0 | |
218 | +#include <linux/cpufreq.h> | |
219 | #include <linux/fs.h> | |
220 | #include <linux/init.h> | |
221 | #include <linux/proc_fs.h> | |
222 | #include <linux/seq_file.h> | |
223 | ||
224 | +__weak void arch_freq_prepare_all(void) | |
225 | +{ | |
226 | +} | |
227 | + | |
228 | extern const struct seq_operations cpuinfo_op; | |
229 | static int cpuinfo_open(struct inode *inode, struct file *file) | |
230 | { | |
231 | + arch_freq_prepare_all(); | |
232 | return seq_open(file, &cpuinfo_op); | |
233 | } | |
234 | ||
235 | --- a/include/linux/cpufreq.h | |
236 | +++ b/include/linux/cpufreq.h | |
237 | @@ -917,6 +917,7 @@ static inline bool policy_has_boost_freq | |
238 | } | |
239 | #endif | |
240 | ||
241 | +extern void arch_freq_prepare_all(void); | |
242 | extern unsigned int arch_freq_get_on_cpu(int cpu); | |
243 | ||
244 | /* the following are really really optional */ |