Merge remote-tracking branch 'remotes/vivier2/tags/trivial-branch-for-6.2-pull-reques...
[qemu.git] / softmmu / icount.c
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-common.h"
27 #include "qemu/cutils.h"
28 #include "migration/vmstate.h"
29 #include "qapi/error.h"
30 #include "qemu/error-report.h"
31 #include "exec/exec-all.h"
32 #include "sysemu/cpus.h"
33 #include "sysemu/qtest.h"
34 #include "qemu/main-loop.h"
35 #include "qemu/option.h"
36 #include "qemu/seqlock.h"
37 #include "sysemu/replay.h"
38 #include "sysemu/runstate.h"
39 #include "hw/core/cpu.h"
40 #include "sysemu/cpu-timers.h"
41 #include "sysemu/cpu-throttle.h"
42 #include "timers-state.h"
43
44 /*
45 * ICOUNT: Instruction Counter
46 *
47 * this module is split off from cpu-timers because the icount part
48 * is TCG-specific, and does not need to be built for other accels.
49 */
50 static bool icount_sleep = true;
51 /* Arbitrarily pick 1MIPS as the minimum allowable speed. */
52 #define MAX_ICOUNT_SHIFT 10
53
54 /*
55 * 0 = Do not count executed instructions.
56 * 1 = Fixed conversion of insn to ns via "shift" option
57 * 2 = Runtime adaptive algorithm to compute shift
58 */
59 int use_icount;
60
61 static void icount_enable_precise(void)
62 {
63 use_icount = 1;
64 }
65
66 static void icount_enable_adaptive(void)
67 {
68 use_icount = 2;
69 }
70
71 /*
72 * The current number of executed instructions is based on what we
73 * originally budgeted minus the current state of the decrementing
74 * icount counters in extra/u16.low.
75 */
76 static int64_t icount_get_executed(CPUState *cpu)
77 {
78 return (cpu->icount_budget -
79 (cpu_neg(cpu)->icount_decr.u16.low + cpu->icount_extra));
80 }
81
82 /*
83 * Update the global shared timer_state.qemu_icount to take into
84 * account executed instructions. This is done by the TCG vCPU
85 * thread so the main-loop can see time has moved forward.
86 */
87 static void icount_update_locked(CPUState *cpu)
88 {
89 int64_t executed = icount_get_executed(cpu);
90 cpu->icount_budget -= executed;
91
92 qatomic_set_i64(&timers_state.qemu_icount,
93 timers_state.qemu_icount + executed);
94 }
95
96 /*
97 * Update the global shared timer_state.qemu_icount to take into
98 * account executed instructions. This is done by the TCG vCPU
99 * thread so the main-loop can see time has moved forward.
100 */
101 void icount_update(CPUState *cpu)
102 {
103 seqlock_write_lock(&timers_state.vm_clock_seqlock,
104 &timers_state.vm_clock_lock);
105 icount_update_locked(cpu);
106 seqlock_write_unlock(&timers_state.vm_clock_seqlock,
107 &timers_state.vm_clock_lock);
108 }
109
110 static int64_t icount_get_raw_locked(void)
111 {
112 CPUState *cpu = current_cpu;
113
114 if (cpu && cpu->running) {
115 if (!cpu->can_do_io) {
116 error_report("Bad icount read");
117 exit(1);
118 }
119 /* Take into account what has run */
120 icount_update_locked(cpu);
121 }
122 /* The read is protected by the seqlock, but needs atomic64 to avoid UB */
123 return qatomic_read_i64(&timers_state.qemu_icount);
124 }
125
126 static int64_t icount_get_locked(void)
127 {
128 int64_t icount = icount_get_raw_locked();
129 return qatomic_read_i64(&timers_state.qemu_icount_bias) +
130 icount_to_ns(icount);
131 }
132
133 int64_t icount_get_raw(void)
134 {
135 int64_t icount;
136 unsigned start;
137
138 do {
139 start = seqlock_read_begin(&timers_state.vm_clock_seqlock);
140 icount = icount_get_raw_locked();
141 } while (seqlock_read_retry(&timers_state.vm_clock_seqlock, start));
142
143 return icount;
144 }
145
146 /* Return the virtual CPU time, based on the instruction counter. */
147 int64_t icount_get(void)
148 {
149 int64_t icount;
150 unsigned start;
151
152 do {
153 start = seqlock_read_begin(&timers_state.vm_clock_seqlock);
154 icount = icount_get_locked();
155 } while (seqlock_read_retry(&timers_state.vm_clock_seqlock, start));
156
157 return icount;
158 }
159
160 int64_t icount_to_ns(int64_t icount)
161 {
162 return icount << qatomic_read(&timers_state.icount_time_shift);
163 }
164
165 /*
166 * Correlation between real and virtual time is always going to be
167 * fairly approximate, so ignore small variation.
168 * When the guest is idle real and virtual time will be aligned in
169 * the IO wait loop.
170 */
171 #define ICOUNT_WOBBLE (NANOSECONDS_PER_SECOND / 10)
172
173 static void icount_adjust(void)
174 {
175 int64_t cur_time;
176 int64_t cur_icount;
177 int64_t delta;
178
179 /* If the VM is not running, then do nothing. */
180 if (!runstate_is_running()) {
181 return;
182 }
183
184 seqlock_write_lock(&timers_state.vm_clock_seqlock,
185 &timers_state.vm_clock_lock);
186 cur_time = REPLAY_CLOCK_LOCKED(REPLAY_CLOCK_VIRTUAL_RT,
187 cpu_get_clock_locked());
188 cur_icount = icount_get_locked();
189
190 delta = cur_icount - cur_time;
191 /* FIXME: This is a very crude algorithm, somewhat prone to oscillation. */
192 if (delta > 0
193 && timers_state.last_delta + ICOUNT_WOBBLE < delta * 2
194 && timers_state.icount_time_shift > 0) {
195 /* The guest is getting too far ahead. Slow time down. */
196 qatomic_set(&timers_state.icount_time_shift,
197 timers_state.icount_time_shift - 1);
198 }
199 if (delta < 0
200 && timers_state.last_delta - ICOUNT_WOBBLE > delta * 2
201 && timers_state.icount_time_shift < MAX_ICOUNT_SHIFT) {
202 /* The guest is getting too far behind. Speed time up. */
203 qatomic_set(&timers_state.icount_time_shift,
204 timers_state.icount_time_shift + 1);
205 }
206 timers_state.last_delta = delta;
207 qatomic_set_i64(&timers_state.qemu_icount_bias,
208 cur_icount - (timers_state.qemu_icount
209 << timers_state.icount_time_shift));
210 seqlock_write_unlock(&timers_state.vm_clock_seqlock,
211 &timers_state.vm_clock_lock);
212 }
213
214 static void icount_adjust_rt(void *opaque)
215 {
216 timer_mod(timers_state.icount_rt_timer,
217 qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL_RT) + 1000);
218 icount_adjust();
219 }
220
221 static void icount_adjust_vm(void *opaque)
222 {
223 timer_mod(timers_state.icount_vm_timer,
224 qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
225 NANOSECONDS_PER_SECOND / 10);
226 icount_adjust();
227 }
228
229 int64_t icount_round(int64_t count)
230 {
231 int shift = qatomic_read(&timers_state.icount_time_shift);
232 return (count + (1 << shift) - 1) >> shift;
233 }
234
235 static void icount_warp_rt(void)
236 {
237 unsigned seq;
238 int64_t warp_start;
239
240 /*
241 * The icount_warp_timer is rescheduled soon after vm_clock_warp_start
242 * changes from -1 to another value, so the race here is okay.
243 */
244 do {
245 seq = seqlock_read_begin(&timers_state.vm_clock_seqlock);
246 warp_start = timers_state.vm_clock_warp_start;
247 } while (seqlock_read_retry(&timers_state.vm_clock_seqlock, seq));
248
249 if (warp_start == -1) {
250 return;
251 }
252
253 seqlock_write_lock(&timers_state.vm_clock_seqlock,
254 &timers_state.vm_clock_lock);
255 if (runstate_is_running()) {
256 int64_t clock = REPLAY_CLOCK_LOCKED(REPLAY_CLOCK_VIRTUAL_RT,
257 cpu_get_clock_locked());
258 int64_t warp_delta;
259
260 warp_delta = clock - timers_state.vm_clock_warp_start;
261 if (icount_enabled() == 2) {
262 /*
263 * In adaptive mode, do not let QEMU_CLOCK_VIRTUAL run too
264 * far ahead of real time.
265 */
266 int64_t cur_icount = icount_get_locked();
267 int64_t delta = clock - cur_icount;
268 warp_delta = MIN(warp_delta, delta);
269 }
270 qatomic_set_i64(&timers_state.qemu_icount_bias,
271 timers_state.qemu_icount_bias + warp_delta);
272 }
273 timers_state.vm_clock_warp_start = -1;
274 seqlock_write_unlock(&timers_state.vm_clock_seqlock,
275 &timers_state.vm_clock_lock);
276
277 if (qemu_clock_expired(QEMU_CLOCK_VIRTUAL)) {
278 qemu_clock_notify(QEMU_CLOCK_VIRTUAL);
279 }
280 }
281
282 static void icount_timer_cb(void *opaque)
283 {
284 /*
285 * No need for a checkpoint because the timer already synchronizes
286 * with CHECKPOINT_CLOCK_VIRTUAL_RT.
287 */
288 icount_warp_rt();
289 }
290
291 void icount_start_warp_timer(void)
292 {
293 int64_t clock;
294 int64_t deadline;
295
296 assert(icount_enabled());
297
298 /*
299 * Nothing to do if the VM is stopped: QEMU_CLOCK_VIRTUAL timers
300 * do not fire, so computing the deadline does not make sense.
301 */
302 if (!runstate_is_running()) {
303 return;
304 }
305
306 if (replay_mode != REPLAY_MODE_PLAY) {
307 if (!all_cpu_threads_idle()) {
308 return;
309 }
310
311 if (qtest_enabled()) {
312 /* When testing, qtest commands advance icount. */
313 return;
314 }
315
316 replay_checkpoint(CHECKPOINT_CLOCK_WARP_START);
317 } else {
318 /* warp clock deterministically in record/replay mode */
319 if (!replay_checkpoint(CHECKPOINT_CLOCK_WARP_START)) {
320 /*
321 * vCPU is sleeping and warp can't be started.
322 * It is probably a race condition: notification sent
323 * to vCPU was processed in advance and vCPU went to sleep.
324 * Therefore we have to wake it up for doing someting.
325 */
326 if (replay_has_checkpoint()) {
327 qemu_clock_notify(QEMU_CLOCK_VIRTUAL);
328 }
329 return;
330 }
331 }
332
333 /* We want to use the earliest deadline from ALL vm_clocks */
334 clock = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL_RT);
335 deadline = qemu_clock_deadline_ns_all(QEMU_CLOCK_VIRTUAL,
336 ~QEMU_TIMER_ATTR_EXTERNAL);
337 if (deadline < 0) {
338 static bool notified;
339 if (!icount_sleep && !notified) {
340 warn_report("icount sleep disabled and no active timers");
341 notified = true;
342 }
343 return;
344 }
345
346 if (deadline > 0) {
347 /*
348 * Ensure QEMU_CLOCK_VIRTUAL proceeds even when the virtual CPU goes to
349 * sleep. Otherwise, the CPU might be waiting for a future timer
350 * interrupt to wake it up, but the interrupt never comes because
351 * the vCPU isn't running any insns and thus doesn't advance the
352 * QEMU_CLOCK_VIRTUAL.
353 */
354 if (!icount_sleep) {
355 /*
356 * We never let VCPUs sleep in no sleep icount mode.
357 * If there is a pending QEMU_CLOCK_VIRTUAL timer we just advance
358 * to the next QEMU_CLOCK_VIRTUAL event and notify it.
359 * It is useful when we want a deterministic execution time,
360 * isolated from host latencies.
361 */
362 seqlock_write_lock(&timers_state.vm_clock_seqlock,
363 &timers_state.vm_clock_lock);
364 qatomic_set_i64(&timers_state.qemu_icount_bias,
365 timers_state.qemu_icount_bias + deadline);
366 seqlock_write_unlock(&timers_state.vm_clock_seqlock,
367 &timers_state.vm_clock_lock);
368 qemu_clock_notify(QEMU_CLOCK_VIRTUAL);
369 } else {
370 /*
371 * We do stop VCPUs and only advance QEMU_CLOCK_VIRTUAL after some
372 * "real" time, (related to the time left until the next event) has
373 * passed. The QEMU_CLOCK_VIRTUAL_RT clock will do this.
374 * This avoids that the warps are visible externally; for example,
375 * you will not be sending network packets continuously instead of
376 * every 100ms.
377 */
378 seqlock_write_lock(&timers_state.vm_clock_seqlock,
379 &timers_state.vm_clock_lock);
380 if (timers_state.vm_clock_warp_start == -1
381 || timers_state.vm_clock_warp_start > clock) {
382 timers_state.vm_clock_warp_start = clock;
383 }
384 seqlock_write_unlock(&timers_state.vm_clock_seqlock,
385 &timers_state.vm_clock_lock);
386 timer_mod_anticipate(timers_state.icount_warp_timer,
387 clock + deadline);
388 }
389 } else if (deadline == 0) {
390 qemu_clock_notify(QEMU_CLOCK_VIRTUAL);
391 }
392 }
393
394 void icount_account_warp_timer(void)
395 {
396 if (!icount_sleep) {
397 return;
398 }
399
400 /*
401 * Nothing to do if the VM is stopped: QEMU_CLOCK_VIRTUAL timers
402 * do not fire, so computing the deadline does not make sense.
403 */
404 if (!runstate_is_running()) {
405 return;
406 }
407
408 /* warp clock deterministically in record/replay mode */
409 if (!replay_checkpoint(CHECKPOINT_CLOCK_WARP_ACCOUNT)) {
410 return;
411 }
412
413 timer_del(timers_state.icount_warp_timer);
414 icount_warp_rt();
415 }
416
417 void icount_configure(QemuOpts *opts, Error **errp)
418 {
419 const char *option = qemu_opt_get(opts, "shift");
420 bool sleep = qemu_opt_get_bool(opts, "sleep", true);
421 bool align = qemu_opt_get_bool(opts, "align", false);
422 long time_shift = -1;
423
424 if (!option) {
425 if (qemu_opt_get(opts, "align") != NULL) {
426 error_setg(errp, "Please specify shift option when using align");
427 }
428 return;
429 }
430
431 if (align && !sleep) {
432 error_setg(errp, "align=on and sleep=off are incompatible");
433 return;
434 }
435
436 if (strcmp(option, "auto") != 0) {
437 if (qemu_strtol(option, NULL, 0, &time_shift) < 0
438 || time_shift < 0 || time_shift > MAX_ICOUNT_SHIFT) {
439 error_setg(errp, "icount: Invalid shift value");
440 return;
441 }
442 } else if (icount_align_option) {
443 error_setg(errp, "shift=auto and align=on are incompatible");
444 return;
445 } else if (!icount_sleep) {
446 error_setg(errp, "shift=auto and sleep=off are incompatible");
447 return;
448 }
449
450 icount_sleep = sleep;
451 if (icount_sleep) {
452 timers_state.icount_warp_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL_RT,
453 icount_timer_cb, NULL);
454 }
455
456 icount_align_option = align;
457
458 if (time_shift >= 0) {
459 timers_state.icount_time_shift = time_shift;
460 icount_enable_precise();
461 return;
462 }
463
464 icount_enable_adaptive();
465
466 /*
467 * 125MIPS seems a reasonable initial guess at the guest speed.
468 * It will be corrected fairly quickly anyway.
469 */
470 timers_state.icount_time_shift = 3;
471
472 /*
473 * Have both realtime and virtual time triggers for speed adjustment.
474 * The realtime trigger catches emulated time passing too slowly,
475 * the virtual time trigger catches emulated time passing too fast.
476 * Realtime triggers occur even when idle, so use them less frequently
477 * than VM triggers.
478 */
479 timers_state.vm_clock_warp_start = -1;
480 timers_state.icount_rt_timer = timer_new_ms(QEMU_CLOCK_VIRTUAL_RT,
481 icount_adjust_rt, NULL);
482 timer_mod(timers_state.icount_rt_timer,
483 qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL_RT) + 1000);
484 timers_state.icount_vm_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL,
485 icount_adjust_vm, NULL);
486 timer_mod(timers_state.icount_vm_timer,
487 qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
488 NANOSECONDS_PER_SECOND / 10);
489 }