1 /* 2 * Xen time implementation. 3 * 4 * This is implemented in terms of a clocksource driver which uses 5 * the hypervisor clock as a nanosecond timebase, and a clockevent 6 * driver which uses the hypervisor's timer mechanism. 7 * 8 * Jeremy Fitzhardinge <jeremy@xensource.com>, XenSource Inc, 2007 9 */ 10 #include <linux/kernel.h> 11 #include <linux/interrupt.h> 12 #include <linux/clocksource.h> 13 #include <linux/clockchips.h> 14 #include <linux/kernel_stat.h> 15 #include <linux/math64.h> 16 17 #include <asm/pvclock.h> 18 #include <asm/xen/hypervisor.h> 19 #include <asm/xen/hypercall.h> 20 21 #include <xen/events.h> 22 #include <xen/interface/xen.h> 23 #include <xen/interface/vcpu.h> 24 25 #include "xen-ops.h" 26 27 #define XEN_SHIFT 22 28 29 /* Xen may fire a timer up to this many ns early */ 30 #define TIMER_SLOP 100000 31 #define NS_PER_TICK (1000000000LL / HZ) 32 33 /* runstate info updated by Xen */ 34 static DEFINE_PER_CPU(struct vcpu_runstate_info, xen_runstate); 35 36 /* snapshots of runstate info */ 37 static DEFINE_PER_CPU(struct vcpu_runstate_info, xen_runstate_snapshot); 38 39 /* unused ns of stolen and blocked time */ 40 static DEFINE_PER_CPU(u64, xen_residual_stolen); 41 static DEFINE_PER_CPU(u64, xen_residual_blocked); 42 43 /* return an consistent snapshot of 64-bit time/counter value */ 44 static u64 get64(const u64 *p) 45 { 46 u64 ret; 47 48 if (BITS_PER_LONG < 64) { 49 u32 *p32 = (u32 *)p; 50 u32 h, l; 51 52 /* 53 * Read high then low, and then make sure high is 54 * still the same; this will only loop if low wraps 55 * and carries into high. 56 * XXX some clean way to make this endian-proof? 57 */ 58 do { 59 h = p32[1]; 60 barrier(); 61 l = p32[0]; 62 barrier(); 63 } while (p32[1] != h); 64 65 ret = (((u64)h) << 32) | l; 66 } else 67 ret = *p; 68 69 return ret; 70 } 71 72 /* 73 * Runstate accounting 74 */ 75 static void get_runstate_snapshot(struct vcpu_runstate_info *res) 76 { 77 u64 state_time; 78 struct vcpu_runstate_info *state; 79 80 BUG_ON(preemptible()); 81 82 state = &__get_cpu_var(xen_runstate); 83 84 /* 85 * The runstate info is always updated by the hypervisor on 86 * the current CPU, so there's no need to use anything 87 * stronger than a compiler barrier when fetching it. 88 */ 89 do { 90 state_time = get64(&state->state_entry_time); 91 barrier(); 92 *res = *state; 93 barrier(); 94 } while (get64(&state->state_entry_time) != state_time); 95 } 96 97 /* return true when a vcpu could run but has no real cpu to run on */ 98 bool xen_vcpu_stolen(int vcpu) 99 { 100 return per_cpu(xen_runstate, vcpu).state == RUNSTATE_runnable; 101 } 102 103 void xen_setup_runstate_info(int cpu) 104 { 105 struct vcpu_register_runstate_memory_area area; 106 107 area.addr.v = &per_cpu(xen_runstate, cpu); 108 109 if (HYPERVISOR_vcpu_op(VCPUOP_register_runstate_memory_area, 110 cpu, &area)) 111 BUG(); 112 } 113 114 static void do_stolen_accounting(void) 115 { 116 struct vcpu_runstate_info state; 117 struct vcpu_runstate_info *snap; 118 s64 blocked, runnable, offline, stolen; 119 cputime_t ticks; 120 121 get_runstate_snapshot(&state); 122 123 WARN_ON(state.state != RUNSTATE_running); 124 125 snap = &__get_cpu_var(xen_runstate_snapshot); 126 127 /* work out how much time the VCPU has not been runn*ing* */ 128 blocked = state.time[RUNSTATE_blocked] - snap->time[RUNSTATE_blocked]; 129 runnable = state.time[RUNSTATE_runnable] - snap->time[RUNSTATE_runnable]; 130 offline = state.time[RUNSTATE_offline] - snap->time[RUNSTATE_offline]; 131 132 *snap = state; 133 134 /* Add the appropriate number of ticks of stolen time, 135 including any left-overs from last time. */ 136 stolen = runnable + offline + __get_cpu_var(xen_residual_stolen); 137 138 if (stolen < 0) 139 stolen = 0; 140 141 ticks = iter_div_u64_rem(stolen, NS_PER_TICK, &stolen); 142 __get_cpu_var(xen_residual_stolen) = stolen; 143 account_steal_ticks(ticks); 144 145 /* Add the appropriate number of ticks of blocked time, 146 including any left-overs from last time. */ 147 blocked += __get_cpu_var(xen_residual_blocked); 148 149 if (blocked < 0) 150 blocked = 0; 151 152 ticks = iter_div_u64_rem(blocked, NS_PER_TICK, &blocked); 153 __get_cpu_var(xen_residual_blocked) = blocked; 154 account_idle_ticks(ticks); 155 } 156 157 /* 158 * Xen sched_clock implementation. Returns the number of unstolen 159 * nanoseconds, which is nanoseconds the VCPU spent in RUNNING+BLOCKED 160 * states. 161 */ 162 unsigned long long xen_sched_clock(void) 163 { 164 struct vcpu_runstate_info state; 165 cycle_t now; 166 u64 ret; 167 s64 offset; 168 169 /* 170 * Ideally sched_clock should be called on a per-cpu basis 171 * anyway, so preempt should already be disabled, but that's 172 * not current practice at the moment. 173 */ 174 preempt_disable(); 175 176 now = xen_clocksource_read(); 177 178 get_runstate_snapshot(&state); 179 180 WARN_ON(state.state != RUNSTATE_running); 181 182 offset = now - state.state_entry_time; 183 if (offset < 0) 184 offset = 0; 185 186 ret = state.time[RUNSTATE_blocked] + 187 state.time[RUNSTATE_running] + 188 offset; 189 190 preempt_enable(); 191 192 return ret; 193 } 194 195 196 /* Get the TSC speed from Xen */ 197 unsigned long xen_tsc_khz(void) 198 { 199 struct pvclock_vcpu_time_info *info = 200 &HYPERVISOR_shared_info->vcpu_info[0].time; 201 202 return pvclock_tsc_khz(info); 203 } 204 205 cycle_t xen_clocksource_read(void) 206 { 207 struct pvclock_vcpu_time_info *src; 208 cycle_t ret; 209 210 src = &get_cpu_var(xen_vcpu)->time; 211 ret = pvclock_clocksource_read(src); 212 put_cpu_var(xen_vcpu); 213 return ret; 214 } 215 216 static cycle_t xen_clocksource_get_cycles(struct clocksource *cs) 217 { 218 return xen_clocksource_read(); 219 } 220 221 static void xen_read_wallclock(struct timespec *ts) 222 { 223 struct shared_info *s = HYPERVISOR_shared_info; 224 struct pvclock_wall_clock *wall_clock = &(s->wc); 225 struct pvclock_vcpu_time_info *vcpu_time; 226 227 vcpu_time = &get_cpu_var(xen_vcpu)->time; 228 pvclock_read_wallclock(wall_clock, vcpu_time, ts); 229 put_cpu_var(xen_vcpu); 230 } 231 232 unsigned long xen_get_wallclock(void) 233 { 234 struct timespec ts; 235 236 xen_read_wallclock(&ts); 237 return ts.tv_sec; 238 } 239 240 int xen_set_wallclock(unsigned long now) 241 { 242 /* do nothing for domU */ 243 return -1; 244 } 245 246 static struct clocksource xen_clocksource __read_mostly = { 247 .name = "xen", 248 .rating = 400, 249 .read = xen_clocksource_get_cycles, 250 .mask = ~0, 251 .mult = 1<<XEN_SHIFT, /* time directly in nanoseconds */ 252 .shift = XEN_SHIFT, 253 .flags = CLOCK_SOURCE_IS_CONTINUOUS, 254 }; 255 256 /* 257 Xen clockevent implementation 258 259 Xen has two clockevent implementations: 260 261 The old timer_op one works with all released versions of Xen prior 262 to version 3.0.4. This version of the hypervisor provides a 263 single-shot timer with nanosecond resolution. However, sharing the 264 same event channel is a 100Hz tick which is delivered while the 265 vcpu is running. We don't care about or use this tick, but it will 266 cause the core time code to think the timer fired too soon, and 267 will end up resetting it each time. It could be filtered, but 268 doing so has complications when the ktime clocksource is not yet 269 the xen clocksource (ie, at boot time). 270 271 The new vcpu_op-based timer interface allows the tick timer period 272 to be changed or turned off. The tick timer is not useful as a 273 periodic timer because events are only delivered to running vcpus. 274 The one-shot timer can report when a timeout is in the past, so 275 set_next_event is capable of returning -ETIME when appropriate. 276 This interface is used when available. 277 */ 278 279 280 /* 281 Get a hypervisor absolute time. In theory we could maintain an 282 offset between the kernel's time and the hypervisor's time, and 283 apply that to a kernel's absolute timeout. Unfortunately the 284 hypervisor and kernel times can drift even if the kernel is using 285 the Xen clocksource, because ntp can warp the kernel's clocksource. 286 */ 287 static s64 get_abs_timeout(unsigned long delta) 288 { 289 return xen_clocksource_read() + delta; 290 } 291 292 static void xen_timerop_set_mode(enum clock_event_mode mode, 293 struct clock_event_device *evt) 294 { 295 switch (mode) { 296 case CLOCK_EVT_MODE_PERIODIC: 297 /* unsupported */ 298 WARN_ON(1); 299 break; 300 301 case CLOCK_EVT_MODE_ONESHOT: 302 case CLOCK_EVT_MODE_RESUME: 303 break; 304 305 case CLOCK_EVT_MODE_UNUSED: 306 case CLOCK_EVT_MODE_SHUTDOWN: 307 HYPERVISOR_set_timer_op(0); /* cancel timeout */ 308 break; 309 } 310 } 311 312 static int xen_timerop_set_next_event(unsigned long delta, 313 struct clock_event_device *evt) 314 { 315 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT); 316 317 if (HYPERVISOR_set_timer_op(get_abs_timeout(delta)) < 0) 318 BUG(); 319 320 /* We may have missed the deadline, but there's no real way of 321 knowing for sure. If the event was in the past, then we'll 322 get an immediate interrupt. */ 323 324 return 0; 325 } 326 327 static const struct clock_event_device xen_timerop_clockevent = { 328 .name = "xen", 329 .features = CLOCK_EVT_FEAT_ONESHOT, 330 331 .max_delta_ns = 0xffffffff, 332 .min_delta_ns = TIMER_SLOP, 333 334 .mult = 1, 335 .shift = 0, 336 .rating = 500, 337 338 .set_mode = xen_timerop_set_mode, 339 .set_next_event = xen_timerop_set_next_event, 340 }; 341 342 343 344 static void xen_vcpuop_set_mode(enum clock_event_mode mode, 345 struct clock_event_device *evt) 346 { 347 int cpu = smp_processor_id(); 348 349 switch (mode) { 350 case CLOCK_EVT_MODE_PERIODIC: 351 WARN_ON(1); /* unsupported */ 352 break; 353 354 case CLOCK_EVT_MODE_ONESHOT: 355 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL)) 356 BUG(); 357 break; 358 359 case CLOCK_EVT_MODE_UNUSED: 360 case CLOCK_EVT_MODE_SHUTDOWN: 361 if (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, cpu, NULL) || 362 HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL)) 363 BUG(); 364 break; 365 case CLOCK_EVT_MODE_RESUME: 366 break; 367 } 368 } 369 370 static int xen_vcpuop_set_next_event(unsigned long delta, 371 struct clock_event_device *evt) 372 { 373 int cpu = smp_processor_id(); 374 struct vcpu_set_singleshot_timer single; 375 int ret; 376 377 WARN_ON(evt->mode != CLOCK_EVT_MODE_ONESHOT); 378 379 single.timeout_abs_ns = get_abs_timeout(delta); 380 single.flags = VCPU_SSHOTTMR_future; 381 382 ret = HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, cpu, &single); 383 384 BUG_ON(ret != 0 && ret != -ETIME); 385 386 return ret; 387 } 388 389 static const struct clock_event_device xen_vcpuop_clockevent = { 390 .name = "xen", 391 .features = CLOCK_EVT_FEAT_ONESHOT, 392 393 .max_delta_ns = 0xffffffff, 394 .min_delta_ns = TIMER_SLOP, 395 396 .mult = 1, 397 .shift = 0, 398 .rating = 500, 399 400 .set_mode = xen_vcpuop_set_mode, 401 .set_next_event = xen_vcpuop_set_next_event, 402 }; 403 404 static const struct clock_event_device *xen_clockevent = 405 &xen_timerop_clockevent; 406 static DEFINE_PER_CPU(struct clock_event_device, xen_clock_events); 407 408 static irqreturn_t xen_timer_interrupt(int irq, void *dev_id) 409 { 410 struct clock_event_device *evt = &__get_cpu_var(xen_clock_events); 411 irqreturn_t ret; 412 413 ret = IRQ_NONE; 414 if (evt->event_handler) { 415 evt->event_handler(evt); 416 ret = IRQ_HANDLED; 417 } 418 419 do_stolen_accounting(); 420 421 return ret; 422 } 423 424 void xen_setup_timer(int cpu) 425 { 426 const char *name; 427 struct clock_event_device *evt; 428 int irq; 429 430 printk(KERN_INFO "installing Xen timer for CPU %d\n", cpu); 431 432 name = kasprintf(GFP_KERNEL, "timer%d", cpu); 433 if (!name) 434 name = "<timer kasprintf failed>"; 435 436 irq = bind_virq_to_irqhandler(VIRQ_TIMER, cpu, xen_timer_interrupt, 437 IRQF_DISABLED|IRQF_PERCPU|IRQF_NOBALANCING|IRQF_TIMER, 438 name, NULL); 439 440 evt = &per_cpu(xen_clock_events, cpu); 441 memcpy(evt, xen_clockevent, sizeof(*evt)); 442 443 evt->cpumask = cpumask_of(cpu); 444 evt->irq = irq; 445 } 446 447 void xen_teardown_timer(int cpu) 448 { 449 struct clock_event_device *evt; 450 BUG_ON(cpu == 0); 451 evt = &per_cpu(xen_clock_events, cpu); 452 unbind_from_irqhandler(evt->irq, NULL); 453 } 454 455 void xen_setup_cpu_clockevents(void) 456 { 457 BUG_ON(preemptible()); 458 459 clockevents_register_device(&__get_cpu_var(xen_clock_events)); 460 } 461 462 void xen_timer_resume(void) 463 { 464 int cpu; 465 466 if (xen_clockevent != &xen_vcpuop_clockevent) 467 return; 468 469 for_each_online_cpu(cpu) { 470 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL)) 471 BUG(); 472 } 473 } 474 475 __init void xen_time_init(void) 476 { 477 int cpu = smp_processor_id(); 478 479 clocksource_register(&xen_clocksource); 480 481 if (HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, cpu, NULL) == 0) { 482 /* Successfully turned off 100Hz tick, so we have the 483 vcpuop-based timer interface */ 484 printk(KERN_DEBUG "Xen: using vcpuop timer interface\n"); 485 xen_clockevent = &xen_vcpuop_clockevent; 486 } 487 488 /* Set initial system time with full resolution */ 489 xen_read_wallclock(&xtime); 490 set_normalized_timespec(&wall_to_monotonic, 491 -xtime.tv_sec, -xtime.tv_nsec); 492 493 setup_force_cpu_cap(X86_FEATURE_TSC); 494 495 xen_setup_runstate_info(cpu); 496 xen_setup_timer(cpu); 497 xen_setup_cpu_clockevents(); 498 } 499