File: | arch/amd64/amd64/tsc.c |
Warning: | line 95, column 4 Value stored to 'count' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
1 | /* $OpenBSD: tsc.c,v 1.24 2021/08/31 15:11:54 kettenis Exp $ */ |
2 | /* |
3 | * Copyright (c) 2008 The NetBSD Foundation, Inc. |
4 | * Copyright (c) 2016,2017 Reyk Floeter <reyk@openbsd.org> |
5 | * Copyright (c) 2017 Adam Steen <adam@adamsteen.com.au> |
6 | * Copyright (c) 2017 Mike Belopuhov <mike@openbsd.org> |
7 | * Copyright (c) 2019 Paul Irofti <paul@irofti.net> |
8 | * |
9 | * Permission to use, copy, modify, and distribute this software for any |
10 | * purpose with or without fee is hereby granted, provided that the above |
11 | * copyright notice and this permission notice appear in all copies. |
12 | * |
13 | * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES |
14 | * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF |
15 | * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR |
16 | * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES |
17 | * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN |
18 | * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF |
19 | * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE. |
20 | */ |
21 | |
22 | #include <sys/param.h> |
23 | #include <sys/systm.h> |
24 | #include <sys/timetc.h> |
25 | #include <sys/atomic.h> |
26 | |
27 | #include <machine/cpu.h> |
28 | #include <machine/cpufunc.h> |
29 | |
30 | #define RECALIBRATE_MAX_RETRIES5 5 |
31 | #define RECALIBRATE_SMI_THRESHOLD50000 50000 |
32 | #define RECALIBRATE_DELAY_THRESHOLD50 50 |
33 | |
34 | int tsc_recalibrate; |
35 | |
36 | uint64_t tsc_frequency; |
37 | int tsc_is_invariant; |
38 | |
39 | #define TSC_DRIFT_MAX250 250 |
40 | #define TSC_SKEW_MAX100 100 |
41 | int64_t tsc_drift_observed; |
42 | |
43 | volatile int64_t tsc_sync_val; |
44 | volatile struct cpu_info *tsc_sync_cpu; |
45 | |
46 | u_int tsc_get_timecount(struct timecounter *tc); |
47 | void tsc_delay(int usecs); |
48 | |
49 | #include "lapic.h" |
50 | #if NLAPIC1 > 0 |
51 | extern u_int32_t lapic_per_second; |
52 | #endif |
53 | |
54 | struct timecounter tsc_timecounter = { |
55 | .tc_get_timecount = tsc_get_timecount, |
56 | .tc_poll_pps = NULL((void *)0), |
57 | .tc_counter_mask = ~0u, |
58 | .tc_frequency = 0, |
59 | .tc_name = "tsc", |
60 | .tc_quality = -1000, |
61 | .tc_priv = NULL((void *)0), |
62 | .tc_user = TC_TSC1, |
63 | }; |
64 | |
65 | uint64_t |
66 | tsc_freq_cpuid(struct cpu_info *ci) |
67 | { |
68 | uint64_t count; |
69 | uint32_t eax, ebx, khz, dummy; |
70 | |
71 | if (!strcmp(cpu_vendor, "GenuineIntel") && |
72 | cpuid_level >= 0x15) { |
73 | eax = ebx = khz = dummy = 0; |
74 | CPUID(0x15, eax, ebx, khz, dummy)__asm volatile("cpuid" : "=a" (eax), "=b" (ebx), "=c" (khz), "=d" (dummy) : "a" (0x15)); |
75 | khz /= 1000; |
76 | if (khz == 0) { |
77 | switch (ci->ci_model) { |
78 | case 0x4e: /* Skylake mobile */ |
79 | case 0x5e: /* Skylake desktop */ |
80 | case 0x8e: /* Kabylake mobile */ |
81 | case 0x9e: /* Kabylake desktop */ |
82 | case 0xa5: /* CML-H CML-S62 CML-S102 */ |
83 | case 0xa6: /* CML-U62 */ |
84 | khz = 24000; /* 24.0 MHz */ |
85 | break; |
86 | case 0x5f: /* Atom Denverton */ |
87 | khz = 25000; /* 25.0 MHz */ |
88 | break; |
89 | case 0x5c: /* Atom Goldmont */ |
90 | khz = 19200; /* 19.2 MHz */ |
91 | break; |
92 | } |
93 | } |
94 | if (ebx == 0 || eax == 0) |
95 | count = 0; |
Value stored to 'count' is never read | |
96 | else if ((count = (uint64_t)khz * (uint64_t)ebx / eax) != 0) { |
97 | #if NLAPIC1 > 0 |
98 | lapic_per_second = khz * 1000; |
99 | #endif |
100 | return (count * 1000); |
101 | } |
102 | } |
103 | |
104 | return (0); |
105 | } |
106 | |
107 | void |
108 | tsc_identify(struct cpu_info *ci) |
109 | { |
110 | if (!(ci->ci_flags & CPUF_PRIMARY0x0008) || |
111 | !(ci->ci_flags & CPUF_CONST_TSC0x0040) || |
112 | !(ci->ci_flags & CPUF_INVAR_TSC0x0100)) |
113 | return; |
114 | |
115 | tsc_is_invariant = 1; |
116 | |
117 | tsc_frequency = tsc_freq_cpuid(ci); |
118 | if (tsc_frequency > 0) |
119 | delay_func = tsc_delay; |
120 | } |
121 | |
122 | static inline int |
123 | get_tsc_and_timecount(struct timecounter *tc, uint64_t *tsc, uint64_t *count) |
124 | { |
125 | uint64_t n, tsc1, tsc2; |
126 | int i; |
127 | |
128 | for (i = 0; i < RECALIBRATE_MAX_RETRIES5; i++) { |
129 | tsc1 = rdtsc_lfence(); |
130 | n = (tc->tc_get_timecount(tc) & tc->tc_counter_mask); |
131 | tsc2 = rdtsc_lfence(); |
132 | |
133 | if ((tsc2 - tsc1) < RECALIBRATE_SMI_THRESHOLD50000) { |
134 | *count = n; |
135 | *tsc = tsc2; |
136 | return (0); |
137 | } |
138 | } |
139 | return (1); |
140 | } |
141 | |
142 | static inline uint64_t |
143 | calculate_tsc_freq(uint64_t tsc1, uint64_t tsc2, int usec) |
144 | { |
145 | uint64_t delta; |
146 | |
147 | delta = (tsc2 - tsc1); |
148 | return (delta * 1000000 / usec); |
149 | } |
150 | |
151 | static inline uint64_t |
152 | calculate_tc_delay(struct timecounter *tc, uint64_t count1, uint64_t count2) |
153 | { |
154 | uint64_t delta; |
155 | |
156 | if (count2 < count1) |
157 | count2 += tc->tc_counter_mask; |
158 | |
159 | delta = (count2 - count1); |
160 | return (delta * 1000000 / tc->tc_frequency); |
161 | } |
162 | |
163 | uint64_t |
164 | measure_tsc_freq(struct timecounter *tc) |
165 | { |
166 | uint64_t count1, count2, frequency, min_freq, tsc1, tsc2; |
167 | u_long s; |
168 | int delay_usec, i, err1, err2, usec, success = 0; |
169 | |
170 | /* warmup the timers */ |
171 | for (i = 0; i < 3; i++) { |
172 | (void)tc->tc_get_timecount(tc); |
173 | (void)rdtsc(); |
174 | } |
175 | |
176 | min_freq = ULLONG_MAX0xffffffffffffffffULL; |
177 | |
178 | delay_usec = 100000; |
179 | for (i = 0; i < 3; i++) { |
180 | s = intr_disable(); |
181 | |
182 | err1 = get_tsc_and_timecount(tc, &tsc1, &count1); |
183 | delay(delay_usec)(*delay_func)(delay_usec); |
184 | err2 = get_tsc_and_timecount(tc, &tsc2, &count2); |
185 | |
186 | intr_restore(s); |
187 | |
188 | if (err1 || err2) |
189 | continue; |
190 | |
191 | usec = calculate_tc_delay(tc, count1, count2); |
192 | |
193 | if ((usec < (delay_usec - RECALIBRATE_DELAY_THRESHOLD50)) || |
194 | (usec > (delay_usec + RECALIBRATE_DELAY_THRESHOLD50))) |
195 | continue; |
196 | |
197 | frequency = calculate_tsc_freq(tsc1, tsc2, usec); |
198 | |
199 | min_freq = MIN(min_freq, frequency)(((min_freq)<(frequency))?(min_freq):(frequency)); |
200 | success++; |
201 | } |
202 | |
203 | return (success > 1 ? min_freq : 0); |
204 | } |
205 | |
206 | void |
207 | calibrate_tsc_freq(void) |
208 | { |
209 | struct timecounter *reference = tsc_timecounter.tc_priv; |
210 | uint64_t freq; |
211 | |
212 | if (!reference || !tsc_recalibrate) |
213 | return; |
214 | |
215 | if ((freq = measure_tsc_freq(reference)) == 0) |
216 | return; |
217 | tsc_frequency = freq; |
218 | tsc_timecounter.tc_frequency = freq; |
219 | if (tsc_is_invariant) |
220 | tsc_timecounter.tc_quality = 2000; |
221 | } |
222 | |
223 | void |
224 | cpu_recalibrate_tsc(struct timecounter *tc) |
225 | { |
226 | struct timecounter *reference = tsc_timecounter.tc_priv; |
227 | |
228 | /* Prevent recalibration with a worse timecounter source */ |
229 | if (reference && reference->tc_quality > tc->tc_quality) |
230 | return; |
231 | |
232 | tsc_timecounter.tc_priv = tc; |
233 | calibrate_tsc_freq(); |
234 | } |
235 | |
236 | u_int |
237 | tsc_get_timecount(struct timecounter *tc) |
238 | { |
239 | return rdtsc_lfence() + curcpu()({struct cpu_info *__ci; asm volatile("movq %%gs:%P1,%0" : "=r" (__ci) :"n" (__builtin_offsetof(struct cpu_info, ci_self))); __ci;})->ci_tsc_skew; |
240 | } |
241 | |
242 | void |
243 | tsc_timecounter_init(struct cpu_info *ci, uint64_t cpufreq) |
244 | { |
245 | #ifdef TSC_DEBUG |
246 | printf("%s: TSC skew=%lld observed drift=%lld\n", ci->ci_dev->dv_xname, |
247 | (long long)ci->ci_tsc_skew, (long long)tsc_drift_observed); |
248 | #endif |
249 | if (ci->ci_tsc_skew < -TSC_SKEW_MAX100 || ci->ci_tsc_skew > TSC_SKEW_MAX100) { |
250 | printf("%s: disabling user TSC (skew=%lld)\n", |
251 | ci->ci_dev->dv_xname, (long long)ci->ci_tsc_skew); |
252 | tsc_timecounter.tc_user = 0; |
253 | } |
254 | |
255 | if (!(ci->ci_flags & CPUF_PRIMARY0x0008) || |
256 | !(ci->ci_flags & CPUF_CONST_TSC0x0040) || |
257 | !(ci->ci_flags & CPUF_INVAR_TSC0x0100)) |
258 | return; |
259 | |
260 | /* Newer CPUs don't require recalibration */ |
261 | if (tsc_frequency > 0) { |
262 | tsc_timecounter.tc_frequency = tsc_frequency; |
263 | tsc_timecounter.tc_quality = 2000; |
264 | } else { |
265 | tsc_recalibrate = 1; |
266 | tsc_frequency = cpufreq; |
267 | tsc_timecounter.tc_frequency = cpufreq; |
268 | calibrate_tsc_freq(); |
269 | } |
270 | |
271 | if (tsc_drift_observed > TSC_DRIFT_MAX250) { |
272 | printf("ERROR: %lld cycle TSC drift observed\n", |
273 | (long long)tsc_drift_observed); |
274 | tsc_timecounter.tc_quality = -1000; |
275 | tsc_timecounter.tc_user = 0; |
276 | tsc_is_invariant = 0; |
277 | } |
278 | |
279 | tc_init(&tsc_timecounter); |
280 | } |
281 | |
282 | /* |
283 | * Record drift (in clock cycles). Called during AP startup. |
284 | */ |
285 | void |
286 | tsc_sync_drift(int64_t drift) |
287 | { |
288 | if (drift < 0) |
289 | drift = -drift; |
290 | if (drift > tsc_drift_observed) |
291 | tsc_drift_observed = drift; |
292 | } |
293 | |
294 | /* |
295 | * Called during startup of APs, by the boot processor. Interrupts |
296 | * are disabled on entry. |
297 | */ |
298 | void |
299 | tsc_read_bp(struct cpu_info *ci, uint64_t *bptscp, uint64_t *aptscp) |
300 | { |
301 | uint64_t bptsc; |
302 | |
303 | if (atomic_swap_ptr(&tsc_sync_cpu, ci)_atomic_swap_ptr((&tsc_sync_cpu), (ci)) != NULL((void *)0)) |
304 | panic("tsc_sync_bp: 1"); |
305 | |
306 | /* Flag it and read our TSC. */ |
307 | atomic_setbits_intx86_atomic_setbits_u32(&ci->ci_flags, CPUF_SYNCTSC0x0800); |
308 | bptsc = (rdtsc_lfence() >> 1); |
309 | |
310 | /* Wait for remote to complete, and read ours again. */ |
311 | while ((ci->ci_flags & CPUF_SYNCTSC0x0800) != 0) |
312 | membar_consumer()do { __asm volatile("" ::: "memory"); } while (0); |
313 | bptsc += (rdtsc_lfence() >> 1); |
314 | |
315 | /* Wait for the results to come in. */ |
316 | while (tsc_sync_cpu == ci) |
317 | CPU_BUSY_CYCLE()__asm volatile("pause": : : "memory"); |
318 | if (tsc_sync_cpu != NULL((void *)0)) |
319 | panic("tsc_sync_bp: 2"); |
320 | |
321 | *bptscp = bptsc; |
322 | *aptscp = tsc_sync_val; |
323 | } |
324 | |
325 | void |
326 | tsc_sync_bp(struct cpu_info *ci) |
327 | { |
328 | uint64_t bptsc, aptsc; |
329 | |
330 | tsc_read_bp(ci, &bptsc, &aptsc); /* discarded - cache effects */ |
331 | tsc_read_bp(ci, &bptsc, &aptsc); |
332 | |
333 | /* Compute final value to adjust for skew. */ |
334 | ci->ci_tsc_skew = bptsc - aptsc; |
335 | } |
336 | |
337 | /* |
338 | * Called during startup of AP, by the AP itself. Interrupts are |
339 | * disabled on entry. |
340 | */ |
341 | void |
342 | tsc_post_ap(struct cpu_info *ci) |
343 | { |
344 | uint64_t tsc; |
345 | |
346 | /* Wait for go-ahead from primary. */ |
347 | while ((ci->ci_flags & CPUF_SYNCTSC0x0800) == 0) |
348 | membar_consumer()do { __asm volatile("" ::: "memory"); } while (0); |
349 | tsc = (rdtsc_lfence() >> 1); |
350 | |
351 | /* Instruct primary to read its counter. */ |
352 | atomic_clearbits_intx86_atomic_clearbits_u32(&ci->ci_flags, CPUF_SYNCTSC0x0800); |
353 | tsc += (rdtsc_lfence() >> 1); |
354 | |
355 | /* Post result. Ensure the whole value goes out atomically. */ |
356 | (void)atomic_swap_64(&tsc_sync_val, tsc)_atomic_swap_64((&tsc_sync_val), (tsc)); |
357 | |
358 | if (atomic_swap_ptr(&tsc_sync_cpu, NULL)_atomic_swap_ptr((&tsc_sync_cpu), (((void *)0))) != ci) |
359 | panic("tsc_sync_ap"); |
360 | } |
361 | |
362 | void |
363 | tsc_sync_ap(struct cpu_info *ci) |
364 | { |
365 | tsc_post_ap(ci); |
366 | tsc_post_ap(ci); |
367 | } |
368 | |
369 | void |
370 | tsc_delay(int usecs) |
371 | { |
372 | uint64_t interval, start; |
373 | |
374 | interval = (uint64_t)usecs * tsc_frequency / 1000000; |
375 | start = rdtsc_lfence(); |
376 | while (rdtsc_lfence() - start < interval) |
377 | CPU_BUSY_CYCLE()__asm volatile("pause": : : "memory"); |
378 | } |