Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* calibrate.c: default delay calibration |
2 | * | |
3 | * Excised from init/main.c | |
4 | * Copyright (C) 1991, 1992 Linus Torvalds | |
5 | */ | |
6 | ||
cd354f1a | 7 | #include <linux/jiffies.h> |
1da177e4 LT |
8 | #include <linux/delay.h> |
9 | #include <linux/init.h> | |
941e492b | 10 | #include <linux/timex.h> |
3da757da | 11 | #include <linux/smp.h> |
8a9e1b0f | 12 | |
f3f3149f | 13 | unsigned long lpj_fine; |
bfe8df3d | 14 | unsigned long preset_lpj; |
1da177e4 LT |
15 | static int __init lpj_setup(char *str) |
16 | { | |
17 | preset_lpj = simple_strtoul(str,NULL,0); | |
18 | return 1; | |
19 | } | |
20 | ||
21 | __setup("lpj=", lpj_setup); | |
22 | ||
8a9e1b0f VP |
23 | #ifdef ARCH_HAS_READ_CURRENT_TIMER |
24 | ||
25 | /* This routine uses the read_current_timer() routine and gets the | |
26 | * loops per jiffy directly, instead of guessing it using delay(). | |
27 | * Also, this code tries to handle non-maskable asynchronous events | |
28 | * (like SMIs) | |
29 | */ | |
30 | #define DELAY_CALIBRATION_TICKS ((HZ < 100) ? 1 : (HZ/100)) | |
31 | #define MAX_DIRECT_CALIBRATION_RETRIES 5 | |
32 | ||
6c81c32f | 33 | static unsigned long __cpuinit calibrate_delay_direct(void) |
8a9e1b0f VP |
34 | { |
35 | unsigned long pre_start, start, post_start; | |
36 | unsigned long pre_end, end, post_end; | |
37 | unsigned long start_jiffies; | |
f3f3149f AK |
38 | unsigned long timer_rate_min, timer_rate_max; |
39 | unsigned long good_timer_sum = 0; | |
40 | unsigned long good_timer_count = 0; | |
d2b46313 AW |
41 | unsigned long measured_times[MAX_DIRECT_CALIBRATION_RETRIES]; |
42 | int max = -1; /* index of measured_times with max/min values or not set */ | |
43 | int min = -1; | |
8a9e1b0f VP |
44 | int i; |
45 | ||
46 | if (read_current_timer(&pre_start) < 0 ) | |
47 | return 0; | |
48 | ||
49 | /* | |
50 | * A simple loop like | |
51 | * while ( jiffies < start_jiffies+1) | |
52 | * start = read_current_timer(); | |
53 | * will not do. As we don't really know whether jiffy switch | |
54 | * happened first or timer_value was read first. And some asynchronous | |
55 | * event can happen between these two events introducing errors in lpj. | |
56 | * | |
57 | * So, we do | |
58 | * 1. pre_start <- When we are sure that jiffy switch hasn't happened | |
59 | * 2. check jiffy switch | |
60 | * 3. start <- timer value before or after jiffy switch | |
61 | * 4. post_start <- When we are sure that jiffy switch has happened | |
62 | * | |
63 | * Note, we don't know anything about order of 2 and 3. | |
64 | * Now, by looking at post_start and pre_start difference, we can | |
65 | * check whether any asynchronous event happened or not | |
66 | */ | |
67 | ||
68 | for (i = 0; i < MAX_DIRECT_CALIBRATION_RETRIES; i++) { | |
69 | pre_start = 0; | |
70 | read_current_timer(&start); | |
71 | start_jiffies = jiffies; | |
70a06228 | 72 | while (time_before_eq(jiffies, start_jiffies + 1)) { |
8a9e1b0f VP |
73 | pre_start = start; |
74 | read_current_timer(&start); | |
75 | } | |
76 | read_current_timer(&post_start); | |
77 | ||
78 | pre_end = 0; | |
79 | end = post_start; | |
70a06228 TD |
80 | while (time_before_eq(jiffies, start_jiffies + 1 + |
81 | DELAY_CALIBRATION_TICKS)) { | |
8a9e1b0f VP |
82 | pre_end = end; |
83 | read_current_timer(&end); | |
84 | } | |
85 | read_current_timer(&post_end); | |
86 | ||
f3f3149f AK |
87 | timer_rate_max = (post_end - pre_start) / |
88 | DELAY_CALIBRATION_TICKS; | |
89 | timer_rate_min = (pre_end - post_start) / | |
90 | DELAY_CALIBRATION_TICKS; | |
8a9e1b0f VP |
91 | |
92 | /* | |
f3f3149f | 93 | * If the upper limit and lower limit of the timer_rate is |
8a9e1b0f VP |
94 | * >= 12.5% apart, redo calibration. |
95 | */ | |
d2b46313 AW |
96 | printk(KERN_DEBUG "calibrate_delay_direct() timer_rate_max=%lu " |
97 | "timer_rate_min=%lu pre_start=%lu pre_end=%lu\n", | |
98 | timer_rate_max, timer_rate_min, pre_start, pre_end); | |
99 | if (start >= post_end) | |
100 | printk(KERN_NOTICE "calibrate_delay_direct() ignoring " | |
101 | "timer_rate as we had a TSC wrap around" | |
102 | " start=%lu >=post_end=%lu\n", | |
103 | start, post_end); | |
104 | if (start < post_end && pre_start != 0 && pre_end != 0 && | |
f3f3149f AK |
105 | (timer_rate_max - timer_rate_min) < (timer_rate_max >> 3)) { |
106 | good_timer_count++; | |
107 | good_timer_sum += timer_rate_max; | |
d2b46313 AW |
108 | measured_times[i] = timer_rate_max; |
109 | if (max < 0 || timer_rate_max > measured_times[max]) | |
110 | max = i; | |
111 | if (min < 0 || timer_rate_max < measured_times[min]) | |
112 | min = i; | |
113 | } else | |
114 | measured_times[i] = 0; | |
115 | ||
8a9e1b0f VP |
116 | } |
117 | ||
d2b46313 AW |
118 | /* |
119 | * Find the maximum & minimum - if they differ too much throw out the | |
120 | * one with the largest difference from the mean and try again... | |
121 | */ | |
122 | while (good_timer_count > 1) { | |
123 | unsigned long estimate; | |
124 | unsigned long maxdiff; | |
125 | ||
126 | /* compute the estimate */ | |
127 | estimate = (good_timer_sum/good_timer_count); | |
128 | maxdiff = estimate >> 3; | |
129 | ||
130 | /* if range is within 12% let's take it */ | |
131 | if ((measured_times[max] - measured_times[min]) < maxdiff) | |
132 | return estimate; | |
133 | ||
134 | /* ok - drop the worse value and try again... */ | |
135 | good_timer_sum = 0; | |
136 | good_timer_count = 0; | |
137 | if ((measured_times[max] - estimate) < | |
138 | (estimate - measured_times[min])) { | |
139 | printk(KERN_NOTICE "calibrate_delay_direct() dropping " | |
140 | "min bogoMips estimate %d = %lu\n", | |
141 | min, measured_times[min]); | |
142 | measured_times[min] = 0; | |
143 | min = max; | |
144 | } else { | |
145 | printk(KERN_NOTICE "calibrate_delay_direct() dropping " | |
146 | "max bogoMips estimate %d = %lu\n", | |
147 | max, measured_times[max]); | |
148 | measured_times[max] = 0; | |
149 | max = min; | |
150 | } | |
151 | ||
152 | for (i = 0; i < MAX_DIRECT_CALIBRATION_RETRIES; i++) { | |
153 | if (measured_times[i] == 0) | |
154 | continue; | |
155 | good_timer_count++; | |
156 | good_timer_sum += measured_times[i]; | |
157 | if (measured_times[i] < measured_times[min]) | |
158 | min = i; | |
159 | if (measured_times[i] > measured_times[max]) | |
160 | max = i; | |
161 | } | |
162 | ||
163 | } | |
8a9e1b0f | 164 | |
d2b46313 AW |
165 | printk(KERN_NOTICE "calibrate_delay_direct() failed to get a good " |
166 | "estimate for loops_per_jiffy.\nProbably due to long platform " | |
167 | "interrupts. Consider using \"lpj=\" boot option.\n"); | |
8a9e1b0f VP |
168 | return 0; |
169 | } | |
170 | #else | |
6c81c32f | 171 | static unsigned long __cpuinit calibrate_delay_direct(void) {return 0;} |
8a9e1b0f VP |
172 | #endif |
173 | ||
1da177e4 LT |
174 | /* |
175 | * This is the number of bits of precision for the loops_per_jiffy. Each | |
191e5688 PC |
176 | * time we refine our estimate after the first takes 1.5/HZ seconds, so try |
177 | * to start with a good estimate. | |
3da757da | 178 | * For the boot cpu we can skip the delay calibration and assign it a value |
f3f3149f AK |
179 | * calculated based on the timer frequency. |
180 | * For the rest of the CPUs we cannot assume that the timer frequency is same as | |
3da757da | 181 | * the cpu frequency, hence do the calibration for those. |
1da177e4 LT |
182 | */ |
183 | #define LPS_PREC 8 | |
184 | ||
71c696b1 | 185 | static unsigned long __cpuinit calibrate_delay_converge(void) |
1da177e4 | 186 | { |
191e5688 | 187 | /* First stage - slowly accelerate to find initial bounds */ |
b1b5f65e | 188 | unsigned long lpj, lpj_base, ticks, loopadd, loopadd_base, chop_limit; |
191e5688 | 189 | int trials = 0, band = 0, trial_in_band = 0; |
71c696b1 PC |
190 | |
191 | lpj = (1<<12); | |
191e5688 PC |
192 | |
193 | /* wait for "start of" clock tick */ | |
194 | ticks = jiffies; | |
195 | while (ticks == jiffies) | |
196 | ; /* nothing */ | |
197 | /* Go .. */ | |
198 | ticks = jiffies; | |
199 | do { | |
200 | if (++trial_in_band == (1<<band)) { | |
201 | ++band; | |
202 | trial_in_band = 0; | |
203 | } | |
204 | __delay(lpj * band); | |
205 | trials += band; | |
206 | } while (ticks == jiffies); | |
207 | /* | |
208 | * We overshot, so retreat to a clear underestimate. Then estimate | |
209 | * the largest likely undershoot. This defines our chop bounds. | |
210 | */ | |
211 | trials -= band; | |
b1b5f65e PC |
212 | loopadd_base = lpj * band; |
213 | lpj_base = lpj * trials; | |
214 | ||
215 | recalibrate: | |
216 | lpj = lpj_base; | |
217 | loopadd = loopadd_base; | |
71c696b1 PC |
218 | |
219 | /* | |
220 | * Do a binary approximation to get lpj set to | |
191e5688 | 221 | * equal one clock (up to LPS_PREC bits) |
71c696b1 | 222 | */ |
b1b5f65e | 223 | chop_limit = lpj >> LPS_PREC; |
191e5688 PC |
224 | while (loopadd > chop_limit) { |
225 | lpj += loopadd; | |
71c696b1 PC |
226 | ticks = jiffies; |
227 | while (ticks == jiffies) | |
191e5688 | 228 | ; /* nothing */ |
71c696b1 PC |
229 | ticks = jiffies; |
230 | __delay(lpj); | |
231 | if (jiffies != ticks) /* longer than 1 tick */ | |
191e5688 PC |
232 | lpj -= loopadd; |
233 | loopadd >>= 1; | |
71c696b1 | 234 | } |
b1b5f65e PC |
235 | /* |
236 | * If we incremented every single time possible, presume we've | |
237 | * massively underestimated initially, and retry with a higher | |
238 | * start, and larger range. (Only seen on x86_64, due to SMIs) | |
239 | */ | |
240 | if (lpj + loopadd * 2 == lpj_base + loopadd_base * 2) { | |
241 | lpj_base = lpj; | |
242 | loopadd_base <<= 2; | |
243 | goto recalibrate; | |
244 | } | |
71c696b1 PC |
245 | |
246 | return lpj; | |
247 | } | |
248 | ||
249 | void __cpuinit calibrate_delay(void) | |
250 | { | |
feae3203 | 251 | static bool printed; |
1da177e4 LT |
252 | |
253 | if (preset_lpj) { | |
254 | loops_per_jiffy = preset_lpj; | |
feae3203 MT |
255 | if (!printed) |
256 | pr_info("Calibrating delay loop (skipped) " | |
257 | "preset value.. "); | |
258 | } else if ((!printed) && lpj_fine) { | |
f3f3149f | 259 | loops_per_jiffy = lpj_fine; |
feae3203 | 260 | pr_info("Calibrating delay loop (skipped), " |
f3f3149f | 261 | "value calculated using timer frequency.. "); |
8a9e1b0f | 262 | } else if ((loops_per_jiffy = calibrate_delay_direct()) != 0) { |
feae3203 MT |
263 | if (!printed) |
264 | pr_info("Calibrating delay using timer " | |
265 | "specific routine.. "); | |
1da177e4 | 266 | } else { |
feae3203 MT |
267 | if (!printed) |
268 | pr_info("Calibrating delay loop... "); | |
71c696b1 | 269 | loops_per_jiffy = calibrate_delay_converge(); |
1da177e4 | 270 | } |
feae3203 MT |
271 | if (!printed) |
272 | pr_cont("%lu.%02lu BogoMIPS (lpj=%lu)\n", | |
3da757da AK |
273 | loops_per_jiffy/(500000/HZ), |
274 | (loops_per_jiffy/(5000/HZ)) % 100, loops_per_jiffy); | |
feae3203 MT |
275 | |
276 | printed = true; | |
1da177e4 | 277 | } |