1 // SPDX-License-Identifier: GPL-2.0 <<
2 /* calibrate.c: default delay calibration 1 /* calibrate.c: default delay calibration
3 * 2 *
4 * Excised from init/main.c 3 * Excised from init/main.c
5 * Copyright (C) 1991, 1992 Linus Torvalds 4 * Copyright (C) 1991, 1992 Linus Torvalds
6 */ 5 */
7 6
8 #include <linux/jiffies.h> 7 #include <linux/jiffies.h>
9 #include <linux/delay.h> 8 #include <linux/delay.h>
10 #include <linux/init.h> 9 #include <linux/init.h>
11 #include <linux/timex.h> 10 #include <linux/timex.h>
12 #include <linux/smp.h> 11 #include <linux/smp.h>
13 #include <linux/percpu.h> 12 #include <linux/percpu.h>
14 13
15 unsigned long lpj_fine; 14 unsigned long lpj_fine;
16 unsigned long preset_lpj; 15 unsigned long preset_lpj;
17 static int __init lpj_setup(char *str) 16 static int __init lpj_setup(char *str)
18 { 17 {
19 preset_lpj = simple_strtoul(str,NULL,0 18 preset_lpj = simple_strtoul(str,NULL,0);
20 return 1; 19 return 1;
21 } 20 }
22 21
23 __setup("lpj=", lpj_setup); 22 __setup("lpj=", lpj_setup);
24 23
25 #ifdef ARCH_HAS_READ_CURRENT_TIMER 24 #ifdef ARCH_HAS_READ_CURRENT_TIMER
26 25
27 /* This routine uses the read_current_timer() 26 /* This routine uses the read_current_timer() routine and gets the
28 * loops per jiffy directly, instead of guessi 27 * loops per jiffy directly, instead of guessing it using delay().
29 * Also, this code tries to handle non-maskabl 28 * Also, this code tries to handle non-maskable asynchronous events
30 * (like SMIs) 29 * (like SMIs)
31 */ 30 */
32 #define DELAY_CALIBRATION_TICKS 31 #define DELAY_CALIBRATION_TICKS ((HZ < 100) ? 1 : (HZ/100))
33 #define MAX_DIRECT_CALIBRATION_RETRIES 32 #define MAX_DIRECT_CALIBRATION_RETRIES 5
34 33
35 static unsigned long calibrate_delay_direct(vo 34 static unsigned long calibrate_delay_direct(void)
36 { 35 {
37 unsigned long pre_start, start, post_s 36 unsigned long pre_start, start, post_start;
38 unsigned long pre_end, end, post_end; 37 unsigned long pre_end, end, post_end;
39 unsigned long start_jiffies; 38 unsigned long start_jiffies;
40 unsigned long timer_rate_min, timer_ra 39 unsigned long timer_rate_min, timer_rate_max;
41 unsigned long good_timer_sum = 0; 40 unsigned long good_timer_sum = 0;
42 unsigned long good_timer_count = 0; 41 unsigned long good_timer_count = 0;
43 unsigned long measured_times[MAX_DIREC 42 unsigned long measured_times[MAX_DIRECT_CALIBRATION_RETRIES];
44 int max = -1; /* index of measured_tim 43 int max = -1; /* index of measured_times with max/min values or not set */
45 int min = -1; 44 int min = -1;
46 int i; 45 int i;
47 46
48 if (read_current_timer(&pre_start) < 0 47 if (read_current_timer(&pre_start) < 0 )
49 return 0; 48 return 0;
50 49
51 /* 50 /*
52 * A simple loop like 51 * A simple loop like
53 * while ( jiffies < start_jiffie 52 * while ( jiffies < start_jiffies+1)
54 * start = read_current_t 53 * start = read_current_timer();
55 * will not do. As we don't really kno 54 * will not do. As we don't really know whether jiffy switch
56 * happened first or timer_value was r 55 * happened first or timer_value was read first. And some asynchronous
57 * event can happen between these two 56 * event can happen between these two events introducing errors in lpj.
58 * 57 *
59 * So, we do 58 * So, we do
60 * 1. pre_start <- When we are sure th 59 * 1. pre_start <- When we are sure that jiffy switch hasn't happened
61 * 2. check jiffy switch 60 * 2. check jiffy switch
62 * 3. start <- timer value before or a 61 * 3. start <- timer value before or after jiffy switch
63 * 4. post_start <- When we are sure t 62 * 4. post_start <- When we are sure that jiffy switch has happened
64 * 63 *
65 * Note, we don't know anything about 64 * Note, we don't know anything about order of 2 and 3.
66 * Now, by looking at post_start and p 65 * Now, by looking at post_start and pre_start difference, we can
67 * check whether any asynchronous even 66 * check whether any asynchronous event happened or not
68 */ 67 */
69 68
70 for (i = 0; i < MAX_DIRECT_CALIBRATION 69 for (i = 0; i < MAX_DIRECT_CALIBRATION_RETRIES; i++) {
71 pre_start = 0; 70 pre_start = 0;
72 read_current_timer(&start); 71 read_current_timer(&start);
73 start_jiffies = jiffies; 72 start_jiffies = jiffies;
74 while (time_before_eq(jiffies, 73 while (time_before_eq(jiffies, start_jiffies + 1)) {
75 pre_start = start; 74 pre_start = start;
76 read_current_timer(&st 75 read_current_timer(&start);
77 } 76 }
78 read_current_timer(&post_start 77 read_current_timer(&post_start);
79 78
80 pre_end = 0; 79 pre_end = 0;
81 end = post_start; 80 end = post_start;
82 while (time_before_eq(jiffies, 81 while (time_before_eq(jiffies, start_jiffies + 1 +
83 82 DELAY_CALIBRATION_TICKS)) {
84 pre_end = end; 83 pre_end = end;
85 read_current_timer(&en 84 read_current_timer(&end);
86 } 85 }
87 read_current_timer(&post_end); 86 read_current_timer(&post_end);
88 87
89 timer_rate_max = (post_end - p 88 timer_rate_max = (post_end - pre_start) /
90 DELAY_ 89 DELAY_CALIBRATION_TICKS;
91 timer_rate_min = (pre_end - po 90 timer_rate_min = (pre_end - post_start) /
92 DELAY_ 91 DELAY_CALIBRATION_TICKS;
93 92
94 /* 93 /*
95 * If the upper limit and lowe 94 * If the upper limit and lower limit of the timer_rate is
96 * >= 12.5% apart, redo calibr 95 * >= 12.5% apart, redo calibration.
97 */ 96 */
98 if (start >= post_end) 97 if (start >= post_end)
99 printk(KERN_NOTICE "ca 98 printk(KERN_NOTICE "calibrate_delay_direct() ignoring "
100 "timer 99 "timer_rate as we had a TSC wrap around"
101 " star 100 " start=%lu >=post_end=%lu\n",
102 start, post_en 101 start, post_end);
103 if (start < post_end && pre_st 102 if (start < post_end && pre_start != 0 && pre_end != 0 &&
104 (timer_rate_max - timer_ra 103 (timer_rate_max - timer_rate_min) < (timer_rate_max >> 3)) {
105 good_timer_count++; 104 good_timer_count++;
106 good_timer_sum += time 105 good_timer_sum += timer_rate_max;
107 measured_times[i] = ti 106 measured_times[i] = timer_rate_max;
108 if (max < 0 || timer_r 107 if (max < 0 || timer_rate_max > measured_times[max])
109 max = i; 108 max = i;
110 if (min < 0 || timer_r 109 if (min < 0 || timer_rate_max < measured_times[min])
111 min = i; 110 min = i;
112 } else 111 } else
113 measured_times[i] = 0; 112 measured_times[i] = 0;
114 113
115 } 114 }
116 115
117 /* 116 /*
118 * Find the maximum & minimum - if the 117 * Find the maximum & minimum - if they differ too much throw out the
119 * one with the largest difference fro 118 * one with the largest difference from the mean and try again...
120 */ 119 */
121 while (good_timer_count > 1) { 120 while (good_timer_count > 1) {
122 unsigned long estimate; 121 unsigned long estimate;
123 unsigned long maxdiff; 122 unsigned long maxdiff;
124 123
125 /* compute the estimate */ 124 /* compute the estimate */
126 estimate = (good_timer_sum/goo 125 estimate = (good_timer_sum/good_timer_count);
127 maxdiff = estimate >> 3; 126 maxdiff = estimate >> 3;
128 127
129 /* if range is within 12% let' 128 /* if range is within 12% let's take it */
130 if ((measured_times[max] - mea 129 if ((measured_times[max] - measured_times[min]) < maxdiff)
131 return estimate; 130 return estimate;
132 131
133 /* ok - drop the worse value a 132 /* ok - drop the worse value and try again... */
134 good_timer_sum = 0; 133 good_timer_sum = 0;
135 good_timer_count = 0; 134 good_timer_count = 0;
136 if ((measured_times[max] - est 135 if ((measured_times[max] - estimate) <
137 (estimate - me 136 (estimate - measured_times[min])) {
138 printk(KERN_NOTICE "ca 137 printk(KERN_NOTICE "calibrate_delay_direct() dropping "
139 "min b 138 "min bogoMips estimate %d = %lu\n",
140 min, measured_ 139 min, measured_times[min]);
141 measured_times[min] = 140 measured_times[min] = 0;
142 min = max; 141 min = max;
143 } else { 142 } else {
144 printk(KERN_NOTICE "ca 143 printk(KERN_NOTICE "calibrate_delay_direct() dropping "
145 "max b 144 "max bogoMips estimate %d = %lu\n",
146 max, measured_ 145 max, measured_times[max]);
147 measured_times[max] = 146 measured_times[max] = 0;
148 max = min; 147 max = min;
149 } 148 }
150 149
151 for (i = 0; i < MAX_DIRECT_CAL 150 for (i = 0; i < MAX_DIRECT_CALIBRATION_RETRIES; i++) {
152 if (measured_times[i] 151 if (measured_times[i] == 0)
153 continue; 152 continue;
154 good_timer_count++; 153 good_timer_count++;
155 good_timer_sum += meas 154 good_timer_sum += measured_times[i];
156 if (measured_times[i] 155 if (measured_times[i] < measured_times[min])
157 min = i; 156 min = i;
158 if (measured_times[i] 157 if (measured_times[i] > measured_times[max])
159 max = i; 158 max = i;
160 } 159 }
161 160
162 } 161 }
163 162
164 printk(KERN_NOTICE "calibrate_delay_di 163 printk(KERN_NOTICE "calibrate_delay_direct() failed to get a good "
165 "estimate for loops_per_jiffy.\ 164 "estimate for loops_per_jiffy.\nProbably due to long platform "
166 "interrupts. Consider using \" 165 "interrupts. Consider using \"lpj=\" boot option.\n");
167 return 0; 166 return 0;
168 } 167 }
169 #else 168 #else
170 static unsigned long calibrate_delay_direct(vo 169 static unsigned long calibrate_delay_direct(void)
171 { 170 {
172 return 0; 171 return 0;
173 } 172 }
174 #endif 173 #endif
175 174
176 /* 175 /*
177 * This is the number of bits of precision for 176 * This is the number of bits of precision for the loops_per_jiffy. Each
178 * time we refine our estimate after the first 177 * time we refine our estimate after the first takes 1.5/HZ seconds, so try
179 * to start with a good estimate. 178 * to start with a good estimate.
180 * For the boot cpu we can skip the delay cali 179 * For the boot cpu we can skip the delay calibration and assign it a value
181 * calculated based on the timer frequency. 180 * calculated based on the timer frequency.
182 * For the rest of the CPUs we cannot assume t 181 * For the rest of the CPUs we cannot assume that the timer frequency is same as
183 * the cpu frequency, hence do the calibration 182 * the cpu frequency, hence do the calibration for those.
184 */ 183 */
185 #define LPS_PREC 8 184 #define LPS_PREC 8
186 185
187 static unsigned long calibrate_delay_converge( 186 static unsigned long calibrate_delay_converge(void)
188 { 187 {
189 /* First stage - slowly accelerate to 188 /* First stage - slowly accelerate to find initial bounds */
190 unsigned long lpj, lpj_base, ticks, lo 189 unsigned long lpj, lpj_base, ticks, loopadd, loopadd_base, chop_limit;
191 int trials = 0, band = 0, trial_in_ban 190 int trials = 0, band = 0, trial_in_band = 0;
192 191
193 lpj = (1<<12); 192 lpj = (1<<12);
194 193
195 /* wait for "start of" clock tick */ 194 /* wait for "start of" clock tick */
196 ticks = jiffies; 195 ticks = jiffies;
197 while (ticks == jiffies) 196 while (ticks == jiffies)
198 ; /* nothing */ 197 ; /* nothing */
199 /* Go .. */ 198 /* Go .. */
200 ticks = jiffies; 199 ticks = jiffies;
201 do { 200 do {
202 if (++trial_in_band == (1<<ban 201 if (++trial_in_band == (1<<band)) {
203 ++band; 202 ++band;
204 trial_in_band = 0; 203 trial_in_band = 0;
205 } 204 }
206 __delay(lpj * band); 205 __delay(lpj * band);
207 trials += band; 206 trials += band;
208 } while (ticks == jiffies); 207 } while (ticks == jiffies);
209 /* 208 /*
210 * We overshot, so retreat to a clear 209 * We overshot, so retreat to a clear underestimate. Then estimate
211 * the largest likely undershoot. This 210 * the largest likely undershoot. This defines our chop bounds.
212 */ 211 */
213 trials -= band; 212 trials -= band;
214 loopadd_base = lpj * band; 213 loopadd_base = lpj * band;
215 lpj_base = lpj * trials; 214 lpj_base = lpj * trials;
216 215
217 recalibrate: 216 recalibrate:
218 lpj = lpj_base; 217 lpj = lpj_base;
219 loopadd = loopadd_base; 218 loopadd = loopadd_base;
220 219
221 /* 220 /*
222 * Do a binary approximation to get lp 221 * Do a binary approximation to get lpj set to
223 * equal one clock (up to LPS_PREC bit 222 * equal one clock (up to LPS_PREC bits)
224 */ 223 */
225 chop_limit = lpj >> LPS_PREC; 224 chop_limit = lpj >> LPS_PREC;
226 while (loopadd > chop_limit) { 225 while (loopadd > chop_limit) {
227 lpj += loopadd; 226 lpj += loopadd;
228 ticks = jiffies; 227 ticks = jiffies;
229 while (ticks == jiffies) 228 while (ticks == jiffies)
230 ; /* nothing */ 229 ; /* nothing */
231 ticks = jiffies; 230 ticks = jiffies;
232 __delay(lpj); 231 __delay(lpj);
233 if (jiffies != ticks) /* lon 232 if (jiffies != ticks) /* longer than 1 tick */
234 lpj -= loopadd; 233 lpj -= loopadd;
235 loopadd >>= 1; 234 loopadd >>= 1;
236 } 235 }
237 /* 236 /*
238 * If we incremented every single time 237 * If we incremented every single time possible, presume we've
239 * massively underestimated initially, 238 * massively underestimated initially, and retry with a higher
240 * start, and larger range. (Only seen 239 * start, and larger range. (Only seen on x86_64, due to SMIs)
241 */ 240 */
242 if (lpj + loopadd * 2 == lpj_base + lo 241 if (lpj + loopadd * 2 == lpj_base + loopadd_base * 2) {
243 lpj_base = lpj; 242 lpj_base = lpj;
244 loopadd_base <<= 2; 243 loopadd_base <<= 2;
245 goto recalibrate; 244 goto recalibrate;
246 } 245 }
247 246
248 return lpj; 247 return lpj;
249 } 248 }
250 249
251 static DEFINE_PER_CPU(unsigned long, cpu_loops 250 static DEFINE_PER_CPU(unsigned long, cpu_loops_per_jiffy) = { 0 };
252 251
253 /* 252 /*
254 * Check if cpu calibration delay is already k 253 * Check if cpu calibration delay is already known. For example,
255 * some processors with multi-core sockets may 254 * some processors with multi-core sockets may have all cores
256 * with the same calibration delay. 255 * with the same calibration delay.
257 * 256 *
258 * Architectures should override this function 257 * Architectures should override this function if a faster calibration
259 * method is available. 258 * method is available.
260 */ 259 */
261 unsigned long __attribute__((weak)) calibrate_ 260 unsigned long __attribute__((weak)) calibrate_delay_is_known(void)
262 { 261 {
263 return 0; 262 return 0;
264 } 263 }
265 264
266 /* 265 /*
267 * Indicate the cpu delay calibration is done. 266 * Indicate the cpu delay calibration is done. This can be used by
268 * architectures to stop accepting delay timer 267 * architectures to stop accepting delay timer registrations after this point.
269 */ 268 */
270 269
271 void __attribute__((weak)) calibration_delay_d 270 void __attribute__((weak)) calibration_delay_done(void)
272 { 271 {
273 } 272 }
274 273
275 void calibrate_delay(void) 274 void calibrate_delay(void)
276 { 275 {
277 unsigned long lpj; 276 unsigned long lpj;
278 static bool printed; 277 static bool printed;
279 int this_cpu = smp_processor_id(); 278 int this_cpu = smp_processor_id();
280 279
281 if (per_cpu(cpu_loops_per_jiffy, this_ 280 if (per_cpu(cpu_loops_per_jiffy, this_cpu)) {
282 lpj = per_cpu(cpu_loops_per_ji 281 lpj = per_cpu(cpu_loops_per_jiffy, this_cpu);
283 if (!printed) 282 if (!printed)
284 pr_info("Calibrating d 283 pr_info("Calibrating delay loop (skipped) "
285 "already calib 284 "already calibrated this CPU");
286 } else if (preset_lpj) { 285 } else if (preset_lpj) {
287 lpj = preset_lpj; 286 lpj = preset_lpj;
288 if (!printed) 287 if (!printed)
289 pr_info("Calibrating d 288 pr_info("Calibrating delay loop (skipped) "
290 "preset value. 289 "preset value.. ");
291 } else if ((!printed) && lpj_fine) { 290 } else if ((!printed) && lpj_fine) {
292 lpj = lpj_fine; 291 lpj = lpj_fine;
293 pr_info("Calibrating delay loo 292 pr_info("Calibrating delay loop (skipped), "
294 "value calculated usin 293 "value calculated using timer frequency.. ");
295 } else if ((lpj = calibrate_delay_is_k 294 } else if ((lpj = calibrate_delay_is_known())) {
296 ; 295 ;
297 } else if ((lpj = calibrate_delay_dire 296 } else if ((lpj = calibrate_delay_direct()) != 0) {
298 if (!printed) 297 if (!printed)
299 pr_info("Calibrating d 298 pr_info("Calibrating delay using timer "
300 "specific rout 299 "specific routine.. ");
301 } else { 300 } else {
302 if (!printed) 301 if (!printed)
303 pr_info("Calibrating d 302 pr_info("Calibrating delay loop... ");
304 lpj = calibrate_delay_converge 303 lpj = calibrate_delay_converge();
305 } 304 }
306 per_cpu(cpu_loops_per_jiffy, this_cpu) 305 per_cpu(cpu_loops_per_jiffy, this_cpu) = lpj;
307 if (!printed) 306 if (!printed)
308 pr_cont("%lu.%02lu BogoMIPS (l 307 pr_cont("%lu.%02lu BogoMIPS (lpj=%lu)\n",
309 lpj/(500000/HZ), 308 lpj/(500000/HZ),
310 (lpj/(5000/HZ)) % 100, 309 (lpj/(5000/HZ)) % 100, lpj);
311 310
312 loops_per_jiffy = lpj; 311 loops_per_jiffy = lpj;
313 printed = true; 312 printed = true;
314 313
315 calibration_delay_done(); 314 calibration_delay_done();
316 } 315 }
317 316
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.