]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * QEMU System Emulator | |
3 | * | |
4 | * Copyright (c) 2003-2008 Fabrice Bellard | |
5 | * | |
6 | * Permission is hereby granted, free of charge, to any person obtaining a copy | |
7 | * of this software and associated documentation files (the "Software"), to deal | |
8 | * in the Software without restriction, including without limitation the rights | |
9 | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell | |
10 | * copies of the Software, and to permit persons to whom the Software is | |
11 | * furnished to do so, subject to the following conditions: | |
12 | * | |
13 | * The above copyright notice and this permission notice shall be included in | |
14 | * all copies or substantial portions of the Software. | |
15 | * | |
16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR | |
17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, | |
18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL | |
19 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER | |
20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, | |
21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN | |
22 | * THE SOFTWARE. | |
23 | */ | |
24 | ||
25 | #include "qemu/osdep.h" | |
26 | #include "qemu/cutils.h" | |
27 | #include "migration/vmstate.h" | |
28 | #include "qapi/error.h" | |
29 | #include "qemu/error-report.h" | |
30 | #include "exec/exec-all.h" | |
31 | #include "sysemu/cpus.h" | |
32 | #include "sysemu/qtest.h" | |
33 | #include "qemu/main-loop.h" | |
34 | #include "qemu/option.h" | |
35 | #include "qemu/seqlock.h" | |
36 | #include "sysemu/replay.h" | |
37 | #include "sysemu/runstate.h" | |
38 | #include "hw/core/cpu.h" | |
39 | #include "sysemu/cpu-timers.h" | |
40 | #include "sysemu/cpu-throttle.h" | |
41 | #include "timers-state.h" | |
42 | ||
43 | /* | |
44 | * ICOUNT: Instruction Counter | |
45 | * | |
46 | * this module is split off from cpu-timers because the icount part | |
47 | * is TCG-specific, and does not need to be built for other accels. | |
48 | */ | |
49 | static bool icount_sleep = true; | |
50 | /* Arbitrarily pick 1MIPS as the minimum allowable speed. */ | |
51 | #define MAX_ICOUNT_SHIFT 10 | |
52 | ||
53 | /* | |
54 | * 0 = Do not count executed instructions. | |
55 | * 1 = Fixed conversion of insn to ns via "shift" option | |
56 | * 2 = Runtime adaptive algorithm to compute shift | |
57 | */ | |
58 | int use_icount; | |
59 | ||
60 | static void icount_enable_precise(void) | |
61 | { | |
62 | use_icount = 1; | |
63 | } | |
64 | ||
65 | static void icount_enable_adaptive(void) | |
66 | { | |
67 | use_icount = 2; | |
68 | } | |
69 | ||
70 | /* | |
71 | * The current number of executed instructions is based on what we | |
72 | * originally budgeted minus the current state of the decrementing | |
73 | * icount counters in extra/u16.low. | |
74 | */ | |
75 | static int64_t icount_get_executed(CPUState *cpu) | |
76 | { | |
77 | return (cpu->icount_budget - | |
78 | (cpu_neg(cpu)->icount_decr.u16.low + cpu->icount_extra)); | |
79 | } | |
80 | ||
81 | /* | |
82 | * Update the global shared timer_state.qemu_icount to take into | |
83 | * account executed instructions. This is done by the TCG vCPU | |
84 | * thread so the main-loop can see time has moved forward. | |
85 | */ | |
86 | static void icount_update_locked(CPUState *cpu) | |
87 | { | |
88 | int64_t executed = icount_get_executed(cpu); | |
89 | cpu->icount_budget -= executed; | |
90 | ||
91 | qatomic_set_i64(&timers_state.qemu_icount, | |
92 | timers_state.qemu_icount + executed); | |
93 | } | |
94 | ||
95 | /* | |
96 | * Update the global shared timer_state.qemu_icount to take into | |
97 | * account executed instructions. This is done by the TCG vCPU | |
98 | * thread so the main-loop can see time has moved forward. | |
99 | */ | |
100 | void icount_update(CPUState *cpu) | |
101 | { | |
102 | seqlock_write_lock(&timers_state.vm_clock_seqlock, | |
103 | &timers_state.vm_clock_lock); | |
104 | icount_update_locked(cpu); | |
105 | seqlock_write_unlock(&timers_state.vm_clock_seqlock, | |
106 | &timers_state.vm_clock_lock); | |
107 | } | |
108 | ||
109 | static int64_t icount_get_raw_locked(void) | |
110 | { | |
111 | CPUState *cpu = current_cpu; | |
112 | ||
113 | if (cpu && cpu->running) { | |
114 | if (!cpu->can_do_io) { | |
115 | error_report("Bad icount read"); | |
116 | exit(1); | |
117 | } | |
118 | /* Take into account what has run */ | |
119 | icount_update_locked(cpu); | |
120 | } | |
121 | /* The read is protected by the seqlock, but needs atomic64 to avoid UB */ | |
122 | return qatomic_read_i64(&timers_state.qemu_icount); | |
123 | } | |
124 | ||
125 | static int64_t icount_get_locked(void) | |
126 | { | |
127 | int64_t icount = icount_get_raw_locked(); | |
128 | return qatomic_read_i64(&timers_state.qemu_icount_bias) + | |
129 | icount_to_ns(icount); | |
130 | } | |
131 | ||
132 | int64_t icount_get_raw(void) | |
133 | { | |
134 | int64_t icount; | |
135 | unsigned start; | |
136 | ||
137 | do { | |
138 | start = seqlock_read_begin(&timers_state.vm_clock_seqlock); | |
139 | icount = icount_get_raw_locked(); | |
140 | } while (seqlock_read_retry(&timers_state.vm_clock_seqlock, start)); | |
141 | ||
142 | return icount; | |
143 | } | |
144 | ||
145 | /* Return the virtual CPU time, based on the instruction counter. */ | |
146 | int64_t icount_get(void) | |
147 | { | |
148 | int64_t icount; | |
149 | unsigned start; | |
150 | ||
151 | do { | |
152 | start = seqlock_read_begin(&timers_state.vm_clock_seqlock); | |
153 | icount = icount_get_locked(); | |
154 | } while (seqlock_read_retry(&timers_state.vm_clock_seqlock, start)); | |
155 | ||
156 | return icount; | |
157 | } | |
158 | ||
159 | int64_t icount_to_ns(int64_t icount) | |
160 | { | |
161 | return icount << qatomic_read(&timers_state.icount_time_shift); | |
162 | } | |
163 | ||
164 | /* | |
165 | * Correlation between real and virtual time is always going to be | |
166 | * fairly approximate, so ignore small variation. | |
167 | * When the guest is idle real and virtual time will be aligned in | |
168 | * the IO wait loop. | |
169 | */ | |
170 | #define ICOUNT_WOBBLE (NANOSECONDS_PER_SECOND / 10) | |
171 | ||
172 | static void icount_adjust(void) | |
173 | { | |
174 | int64_t cur_time; | |
175 | int64_t cur_icount; | |
176 | int64_t delta; | |
177 | ||
178 | /* If the VM is not running, then do nothing. */ | |
179 | if (!runstate_is_running()) { | |
180 | return; | |
181 | } | |
182 | ||
183 | seqlock_write_lock(&timers_state.vm_clock_seqlock, | |
184 | &timers_state.vm_clock_lock); | |
185 | cur_time = REPLAY_CLOCK_LOCKED(REPLAY_CLOCK_VIRTUAL_RT, | |
186 | cpu_get_clock_locked()); | |
187 | cur_icount = icount_get_locked(); | |
188 | ||
189 | delta = cur_icount - cur_time; | |
190 | /* FIXME: This is a very crude algorithm, somewhat prone to oscillation. */ | |
191 | if (delta > 0 | |
192 | && timers_state.last_delta + ICOUNT_WOBBLE < delta * 2 | |
193 | && timers_state.icount_time_shift > 0) { | |
194 | /* The guest is getting too far ahead. Slow time down. */ | |
195 | qatomic_set(&timers_state.icount_time_shift, | |
196 | timers_state.icount_time_shift - 1); | |
197 | } | |
198 | if (delta < 0 | |
199 | && timers_state.last_delta - ICOUNT_WOBBLE > delta * 2 | |
200 | && timers_state.icount_time_shift < MAX_ICOUNT_SHIFT) { | |
201 | /* The guest is getting too far behind. Speed time up. */ | |
202 | qatomic_set(&timers_state.icount_time_shift, | |
203 | timers_state.icount_time_shift + 1); | |
204 | } | |
205 | timers_state.last_delta = delta; | |
206 | qatomic_set_i64(&timers_state.qemu_icount_bias, | |
207 | cur_icount - (timers_state.qemu_icount | |
208 | << timers_state.icount_time_shift)); | |
209 | seqlock_write_unlock(&timers_state.vm_clock_seqlock, | |
210 | &timers_state.vm_clock_lock); | |
211 | } | |
212 | ||
213 | static void icount_adjust_rt(void *opaque) | |
214 | { | |
215 | timer_mod(timers_state.icount_rt_timer, | |
216 | qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL_RT) + 1000); | |
217 | icount_adjust(); | |
218 | } | |
219 | ||
220 | static void icount_adjust_vm(void *opaque) | |
221 | { | |
222 | timer_mod(timers_state.icount_vm_timer, | |
223 | qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + | |
224 | NANOSECONDS_PER_SECOND / 10); | |
225 | icount_adjust(); | |
226 | } | |
227 | ||
228 | int64_t icount_round(int64_t count) | |
229 | { | |
230 | int shift = qatomic_read(&timers_state.icount_time_shift); | |
231 | return (count + (1 << shift) - 1) >> shift; | |
232 | } | |
233 | ||
234 | static void icount_warp_rt(void) | |
235 | { | |
236 | unsigned seq; | |
237 | int64_t warp_start; | |
238 | ||
239 | /* | |
240 | * The icount_warp_timer is rescheduled soon after vm_clock_warp_start | |
241 | * changes from -1 to another value, so the race here is okay. | |
242 | */ | |
243 | do { | |
244 | seq = seqlock_read_begin(&timers_state.vm_clock_seqlock); | |
245 | warp_start = timers_state.vm_clock_warp_start; | |
246 | } while (seqlock_read_retry(&timers_state.vm_clock_seqlock, seq)); | |
247 | ||
248 | if (warp_start == -1) { | |
249 | return; | |
250 | } | |
251 | ||
252 | seqlock_write_lock(&timers_state.vm_clock_seqlock, | |
253 | &timers_state.vm_clock_lock); | |
254 | if (runstate_is_running()) { | |
255 | int64_t clock = REPLAY_CLOCK_LOCKED(REPLAY_CLOCK_VIRTUAL_RT, | |
256 | cpu_get_clock_locked()); | |
257 | int64_t warp_delta; | |
258 | ||
259 | warp_delta = clock - timers_state.vm_clock_warp_start; | |
260 | if (icount_enabled() == 2) { | |
261 | /* | |
262 | * In adaptive mode, do not let QEMU_CLOCK_VIRTUAL run too | |
263 | * far ahead of real time. | |
264 | */ | |
265 | int64_t cur_icount = icount_get_locked(); | |
266 | int64_t delta = clock - cur_icount; | |
267 | warp_delta = MIN(warp_delta, delta); | |
268 | } | |
269 | qatomic_set_i64(&timers_state.qemu_icount_bias, | |
270 | timers_state.qemu_icount_bias + warp_delta); | |
271 | } | |
272 | timers_state.vm_clock_warp_start = -1; | |
273 | seqlock_write_unlock(&timers_state.vm_clock_seqlock, | |
274 | &timers_state.vm_clock_lock); | |
275 | ||
276 | if (qemu_clock_expired(QEMU_CLOCK_VIRTUAL)) { | |
277 | qemu_clock_notify(QEMU_CLOCK_VIRTUAL); | |
278 | } | |
279 | } | |
280 | ||
281 | static void icount_timer_cb(void *opaque) | |
282 | { | |
283 | /* | |
284 | * No need for a checkpoint because the timer already synchronizes | |
285 | * with CHECKPOINT_CLOCK_VIRTUAL_RT. | |
286 | */ | |
287 | icount_warp_rt(); | |
288 | } | |
289 | ||
290 | void icount_start_warp_timer(void) | |
291 | { | |
292 | int64_t clock; | |
293 | int64_t deadline; | |
294 | ||
295 | assert(icount_enabled()); | |
296 | ||
297 | /* | |
298 | * Nothing to do if the VM is stopped: QEMU_CLOCK_VIRTUAL timers | |
299 | * do not fire, so computing the deadline does not make sense. | |
300 | */ | |
301 | if (!runstate_is_running()) { | |
302 | return; | |
303 | } | |
304 | ||
305 | if (replay_mode != REPLAY_MODE_PLAY) { | |
306 | if (!all_cpu_threads_idle()) { | |
307 | return; | |
308 | } | |
309 | ||
310 | if (qtest_enabled()) { | |
311 | /* When testing, qtest commands advance icount. */ | |
312 | return; | |
313 | } | |
314 | ||
315 | replay_checkpoint(CHECKPOINT_CLOCK_WARP_START); | |
316 | } else { | |
317 | /* warp clock deterministically in record/replay mode */ | |
318 | if (!replay_checkpoint(CHECKPOINT_CLOCK_WARP_START)) { | |
319 | /* | |
320 | * vCPU is sleeping and warp can't be started. | |
321 | * It is probably a race condition: notification sent | |
322 | * to vCPU was processed in advance and vCPU went to sleep. | |
323 | * Therefore we have to wake it up for doing someting. | |
324 | */ | |
325 | if (replay_has_checkpoint()) { | |
326 | qemu_clock_notify(QEMU_CLOCK_VIRTUAL); | |
327 | } | |
328 | return; | |
329 | } | |
330 | } | |
331 | ||
332 | /* We want to use the earliest deadline from ALL vm_clocks */ | |
333 | clock = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL_RT); | |
334 | deadline = qemu_clock_deadline_ns_all(QEMU_CLOCK_VIRTUAL, | |
335 | ~QEMU_TIMER_ATTR_EXTERNAL); | |
336 | if (deadline < 0) { | |
337 | static bool notified; | |
338 | if (!icount_sleep && !notified) { | |
339 | warn_report("icount sleep disabled and no active timers"); | |
340 | notified = true; | |
341 | } | |
342 | return; | |
343 | } | |
344 | ||
345 | if (deadline > 0) { | |
346 | /* | |
347 | * Ensure QEMU_CLOCK_VIRTUAL proceeds even when the virtual CPU goes to | |
348 | * sleep. Otherwise, the CPU might be waiting for a future timer | |
349 | * interrupt to wake it up, but the interrupt never comes because | |
350 | * the vCPU isn't running any insns and thus doesn't advance the | |
351 | * QEMU_CLOCK_VIRTUAL. | |
352 | */ | |
353 | if (!icount_sleep) { | |
354 | /* | |
355 | * We never let VCPUs sleep in no sleep icount mode. | |
356 | * If there is a pending QEMU_CLOCK_VIRTUAL timer we just advance | |
357 | * to the next QEMU_CLOCK_VIRTUAL event and notify it. | |
358 | * It is useful when we want a deterministic execution time, | |
359 | * isolated from host latencies. | |
360 | */ | |
361 | seqlock_write_lock(&timers_state.vm_clock_seqlock, | |
362 | &timers_state.vm_clock_lock); | |
363 | qatomic_set_i64(&timers_state.qemu_icount_bias, | |
364 | timers_state.qemu_icount_bias + deadline); | |
365 | seqlock_write_unlock(&timers_state.vm_clock_seqlock, | |
366 | &timers_state.vm_clock_lock); | |
367 | qemu_clock_notify(QEMU_CLOCK_VIRTUAL); | |
368 | } else { | |
369 | /* | |
370 | * We do stop VCPUs and only advance QEMU_CLOCK_VIRTUAL after some | |
371 | * "real" time, (related to the time left until the next event) has | |
372 | * passed. The QEMU_CLOCK_VIRTUAL_RT clock will do this. | |
373 | * This avoids that the warps are visible externally; for example, | |
374 | * you will not be sending network packets continuously instead of | |
375 | * every 100ms. | |
376 | */ | |
377 | seqlock_write_lock(&timers_state.vm_clock_seqlock, | |
378 | &timers_state.vm_clock_lock); | |
379 | if (timers_state.vm_clock_warp_start == -1 | |
380 | || timers_state.vm_clock_warp_start > clock) { | |
381 | timers_state.vm_clock_warp_start = clock; | |
382 | } | |
383 | seqlock_write_unlock(&timers_state.vm_clock_seqlock, | |
384 | &timers_state.vm_clock_lock); | |
385 | timer_mod_anticipate(timers_state.icount_warp_timer, | |
386 | clock + deadline); | |
387 | } | |
388 | } else if (deadline == 0) { | |
389 | qemu_clock_notify(QEMU_CLOCK_VIRTUAL); | |
390 | } | |
391 | } | |
392 | ||
393 | void icount_account_warp_timer(void) | |
394 | { | |
395 | if (!icount_sleep) { | |
396 | return; | |
397 | } | |
398 | ||
399 | /* | |
400 | * Nothing to do if the VM is stopped: QEMU_CLOCK_VIRTUAL timers | |
401 | * do not fire, so computing the deadline does not make sense. | |
402 | */ | |
403 | if (!runstate_is_running()) { | |
404 | return; | |
405 | } | |
406 | ||
407 | /* warp clock deterministically in record/replay mode */ | |
408 | if (!replay_checkpoint(CHECKPOINT_CLOCK_WARP_ACCOUNT)) { | |
409 | return; | |
410 | } | |
411 | ||
412 | timer_del(timers_state.icount_warp_timer); | |
413 | icount_warp_rt(); | |
414 | } | |
415 | ||
416 | void icount_configure(QemuOpts *opts, Error **errp) | |
417 | { | |
418 | const char *option = qemu_opt_get(opts, "shift"); | |
419 | bool sleep = qemu_opt_get_bool(opts, "sleep", true); | |
420 | bool align = qemu_opt_get_bool(opts, "align", false); | |
421 | long time_shift = -1; | |
422 | ||
423 | if (!option) { | |
424 | if (qemu_opt_get(opts, "align") != NULL) { | |
425 | error_setg(errp, "Please specify shift option when using align"); | |
426 | } | |
427 | return; | |
428 | } | |
429 | ||
430 | if (align && !sleep) { | |
431 | error_setg(errp, "align=on and sleep=off are incompatible"); | |
432 | return; | |
433 | } | |
434 | ||
435 | if (strcmp(option, "auto") != 0) { | |
436 | if (qemu_strtol(option, NULL, 0, &time_shift) < 0 | |
437 | || time_shift < 0 || time_shift > MAX_ICOUNT_SHIFT) { | |
438 | error_setg(errp, "icount: Invalid shift value"); | |
439 | return; | |
440 | } | |
441 | } else if (icount_align_option) { | |
442 | error_setg(errp, "shift=auto and align=on are incompatible"); | |
443 | return; | |
444 | } else if (!icount_sleep) { | |
445 | error_setg(errp, "shift=auto and sleep=off are incompatible"); | |
446 | return; | |
447 | } | |
448 | ||
449 | icount_sleep = sleep; | |
450 | if (icount_sleep) { | |
451 | timers_state.icount_warp_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL_RT, | |
452 | icount_timer_cb, NULL); | |
453 | } | |
454 | ||
455 | icount_align_option = align; | |
456 | ||
457 | if (time_shift >= 0) { | |
458 | timers_state.icount_time_shift = time_shift; | |
459 | icount_enable_precise(); | |
460 | return; | |
461 | } | |
462 | ||
463 | icount_enable_adaptive(); | |
464 | ||
465 | /* | |
466 | * 125MIPS seems a reasonable initial guess at the guest speed. | |
467 | * It will be corrected fairly quickly anyway. | |
468 | */ | |
469 | timers_state.icount_time_shift = 3; | |
470 | ||
471 | /* | |
472 | * Have both realtime and virtual time triggers for speed adjustment. | |
473 | * The realtime trigger catches emulated time passing too slowly, | |
474 | * the virtual time trigger catches emulated time passing too fast. | |
475 | * Realtime triggers occur even when idle, so use them less frequently | |
476 | * than VM triggers. | |
477 | */ | |
478 | timers_state.vm_clock_warp_start = -1; | |
479 | timers_state.icount_rt_timer = timer_new_ms(QEMU_CLOCK_VIRTUAL_RT, | |
480 | icount_adjust_rt, NULL); | |
481 | timer_mod(timers_state.icount_rt_timer, | |
482 | qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL_RT) + 1000); | |
483 | timers_state.icount_vm_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, | |
484 | icount_adjust_vm, NULL); | |
485 | timer_mod(timers_state.icount_vm_timer, | |
486 | qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + | |
487 | NANOSECONDS_PER_SECOND / 10); | |
488 | } |