1 /*- 2 * Copyright (c) 2009 Adrian Chadd 3 * Copyright (c) 2012 Spectra Logic Corporation 4 * All rights reserved. 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 * 27 */ 28 29 /** 30 * \file dev/xen/timer/timer.c 31 * \brief A timer driver for the Xen hypervisor's PV clock. 32 */ 33 34 #include <sys/cdefs.h> 35 __FBSDID("$FreeBSD$"); 36 37 #include <sys/param.h> 38 #include <sys/systm.h> 39 #include <sys/bus.h> 40 #include <sys/kernel.h> 41 #include <sys/module.h> 42 #include <sys/time.h> 43 #include <sys/timetc.h> 44 #include <sys/timeet.h> 45 #include <sys/smp.h> 46 #include <sys/limits.h> 47 #include <sys/clock.h> 48 #include <sys/proc.h> 49 50 #include <xen/xen-os.h> 51 #include <xen/features.h> 52 #include <xen/xen_intr.h> 53 #include <xen/hypervisor.h> 54 #include <xen/interface/io/xenbus.h> 55 #include <xen/interface/vcpu.h> 56 #include <xen/error.h> 57 58 #include <machine/cpu.h> 59 #include <machine/cpufunc.h> 60 #include <machine/clock.h> 61 #include <machine/_inttypes.h> 62 #include <machine/smp.h> 63 #include <machine/pvclock.h> 64 65 #include <dev/xen/timer/timer.h> 66 67 #include <isa/rtc.h> 68 69 #include "clock_if.h" 70 71 static devclass_t xentimer_devclass; 72 73 #define NSEC_IN_SEC 1000000000ULL 74 #define NSEC_IN_USEC 1000ULL 75 /* 18446744073 = int(2^64 / NSEC_IN_SC) = 1 ns in 64-bit fractions */ 76 #define FRAC_IN_NSEC 18446744073LL 77 78 /* Xen timers may fire up to 100us off */ 79 #define XENTIMER_MIN_PERIOD_IN_NSEC 100*NSEC_IN_USEC 80 81 /* 82 * The real resolution of the PV clock is 1ns, but the highest 83 * resolution that FreeBSD supports is 1us, so just use that. 84 */ 85 #define XENCLOCK_RESOLUTION 1 86 87 #define XENTIMER_QUALITY 950 88 89 struct xentimer_pcpu_data { 90 uint64_t timer; 91 uint64_t last_processed; 92 void *irq_handle; 93 }; 94 95 DPCPU_DEFINE(struct xentimer_pcpu_data, xentimer_pcpu); 96 97 DPCPU_DECLARE(struct vcpu_info *, vcpu_info); 98 99 struct xentimer_softc { 100 device_t dev; 101 struct timecounter tc; 102 struct eventtimer et; 103 }; 104 105 static void 106 xentimer_identify(driver_t *driver, device_t parent) 107 { 108 if (!xen_domain()) 109 return; 110 111 /* Handle all Xen PV timers in one device instance. */ 112 if (devclass_get_device(xentimer_devclass, 0)) 113 return; 114 115 BUS_ADD_CHILD(parent, 0, "xen_et", 0); 116 } 117 118 static int 119 xentimer_probe(device_t dev) 120 { 121 KASSERT((xen_domain()), ("Trying to use Xen timer on bare metal")); 122 /* 123 * In order to attach, this driver requires the following: 124 * - Vector callback support by the hypervisor, in order to deliver 125 * timer interrupts to the correct CPU for CPUs other than 0. 126 * - Access to the hypervisor shared info page, in order to look up 127 * each VCPU's timer information and the Xen wallclock time. 128 * - The hypervisor must say its PV clock is "safe" to use. 129 * - The hypervisor must support VCPUOP hypercalls. 130 * - The maximum number of CPUs supported by FreeBSD must not exceed 131 * the number of VCPUs supported by the hypervisor. 132 */ 133 #define XTREQUIRES(condition, reason...) \ 134 if (!(condition)) { \ 135 device_printf(dev, ## reason); \ 136 device_detach(dev); \ 137 return (ENXIO); \ 138 } 139 140 if (xen_hvm_domain()) { 141 XTREQUIRES(xen_vector_callback_enabled, 142 "vector callbacks unavailable\n"); 143 XTREQUIRES(xen_feature(XENFEAT_hvm_safe_pvclock), 144 "HVM safe pvclock unavailable\n"); 145 } 146 XTREQUIRES(HYPERVISOR_shared_info != NULL, 147 "shared info page unavailable\n"); 148 XTREQUIRES(HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, 0, NULL) == 0, 149 "VCPUOPs interface unavailable\n"); 150 #undef XTREQUIRES 151 device_set_desc(dev, "Xen PV Clock"); 152 return (BUS_PROBE_NOWILDCARD); 153 } 154 155 /** 156 * \brief Get the current time, in nanoseconds, since the hypervisor booted. 157 * 158 * \param vcpu vcpu_info structure to fetch the time from. 159 * 160 */ 161 static uint64_t 162 xen_fetch_vcpu_time(struct vcpu_info *vcpu) 163 { 164 struct pvclock_vcpu_time_info *time; 165 166 time = (struct pvclock_vcpu_time_info *) &vcpu->time; 167 168 return (pvclock_get_timecount(time)); 169 } 170 171 static uint32_t 172 xentimer_get_timecount(struct timecounter *tc) 173 { 174 uint64_t vcpu_time; 175 176 /* 177 * We don't disable preemption here because the worst that can 178 * happen is reading the vcpu_info area of a different CPU than 179 * the one we are currently running on, but that would also 180 * return a valid tc (and we avoid the overhead of 181 * critical_{enter/exit} calls). 182 */ 183 vcpu_time = xen_fetch_vcpu_time(DPCPU_GET(vcpu_info)); 184 185 return (vcpu_time & UINT32_MAX); 186 } 187 188 /** 189 * \brief Fetch the hypervisor boot time, known as the "Xen wallclock". 190 * 191 * \param ts Timespec to store the current stable value. 192 * \param version Pointer to store the corresponding wallclock version. 193 * 194 * \note This value is updated when Domain-0 shifts its clock to follow 195 * clock drift, e.g. as detected by NTP. 196 */ 197 static void 198 xen_fetch_wallclock(struct timespec *ts) 199 { 200 shared_info_t *src = HYPERVISOR_shared_info; 201 struct pvclock_wall_clock *wc; 202 203 wc = (struct pvclock_wall_clock *) &src->wc_version; 204 205 pvclock_get_wallclock(wc, ts); 206 } 207 208 static void 209 xen_fetch_uptime(struct timespec *ts) 210 { 211 uint64_t uptime; 212 213 uptime = xen_fetch_vcpu_time(DPCPU_GET(vcpu_info)); 214 215 ts->tv_sec = uptime / NSEC_IN_SEC; 216 ts->tv_nsec = uptime % NSEC_IN_SEC; 217 } 218 219 static int 220 xentimer_settime(device_t dev __unused, struct timespec *ts) 221 { 222 struct xen_platform_op settime; 223 int ret; 224 225 /* 226 * Don't return EINVAL here; just silently fail if the domain isn't 227 * privileged enough to set the TOD. 228 */ 229 if (!xen_initial_domain()) 230 return (0); 231 232 /* Set the native RTC. */ 233 atrtc_set(ts); 234 235 settime.cmd = XENPF_settime64; 236 settime.u.settime64.mbz = 0; 237 settime.u.settime64.secs = ts->tv_sec; 238 settime.u.settime64.nsecs = ts->tv_nsec; 239 settime.u.settime64.system_time = 240 xen_fetch_vcpu_time(DPCPU_GET(vcpu_info)); 241 242 ret = HYPERVISOR_platform_op(&settime); 243 ret = ret != 0 ? xen_translate_error(ret) : 0; 244 if (ret != 0 && bootverbose) 245 device_printf(dev, "failed to set Xen PV clock: %d\n", ret); 246 247 return (ret); 248 } 249 250 /** 251 * \brief Return current time according to the Xen Hypervisor wallclock. 252 * 253 * \param dev Xentimer device. 254 * \param ts Pointer to store the wallclock time. 255 * 256 * \note The Xen time structures document the hypervisor start time and the 257 * uptime-since-hypervisor-start (in nsec.) They need to be combined 258 * in order to calculate a TOD clock. 259 */ 260 static int 261 xentimer_gettime(device_t dev, struct timespec *ts) 262 { 263 struct timespec u_ts; 264 265 timespecclear(ts); 266 xen_fetch_wallclock(ts); 267 xen_fetch_uptime(&u_ts); 268 timespecadd(ts, &u_ts); 269 270 return (0); 271 } 272 273 /** 274 * \brief Handle a timer interrupt for the Xen PV timer driver. 275 * 276 * \param arg Xen timer driver softc that is expecting the interrupt. 277 */ 278 static int 279 xentimer_intr(void *arg) 280 { 281 struct xentimer_softc *sc = (struct xentimer_softc *)arg; 282 struct xentimer_pcpu_data *pcpu = DPCPU_PTR(xentimer_pcpu); 283 284 pcpu->last_processed = xen_fetch_vcpu_time(DPCPU_GET(vcpu_info)); 285 if (pcpu->timer != 0 && sc->et.et_active) 286 sc->et.et_event_cb(&sc->et, sc->et.et_arg); 287 288 return (FILTER_HANDLED); 289 } 290 291 static int 292 xentimer_vcpu_start_timer(int vcpu, uint64_t next_time) 293 { 294 struct vcpu_set_singleshot_timer single; 295 296 single.timeout_abs_ns = next_time; 297 /* Get an event anyway, even if the timeout is already expired */ 298 single.flags = 0; 299 return (HYPERVISOR_vcpu_op(VCPUOP_set_singleshot_timer, vcpu, &single)); 300 } 301 302 static int 303 xentimer_vcpu_stop_timer(int vcpu) 304 { 305 306 return (HYPERVISOR_vcpu_op(VCPUOP_stop_singleshot_timer, vcpu, NULL)); 307 } 308 309 /** 310 * \brief Set the next oneshot time for the current CPU. 311 * 312 * \param et Xen timer driver event timer to schedule on. 313 * \param first Delta to the next time to schedule the interrupt for. 314 * \param period Not used. 315 * 316 * \note See eventtimers(9) for more information. 317 * \note 318 * 319 * \returns 0 320 */ 321 static int 322 xentimer_et_start(struct eventtimer *et, 323 sbintime_t first, sbintime_t period) 324 { 325 int error; 326 struct xentimer_softc *sc = et->et_priv; 327 int cpu = PCPU_GET(vcpu_id); 328 struct xentimer_pcpu_data *pcpu = DPCPU_PTR(xentimer_pcpu); 329 struct vcpu_info *vcpu = DPCPU_GET(vcpu_info); 330 uint64_t first_in_ns, next_time; 331 #ifdef INVARIANTS 332 struct thread *td = curthread; 333 #endif 334 335 KASSERT(td->td_critnest != 0, 336 ("xentimer_et_start called without preemption disabled")); 337 338 /* See sbttots() for this formula. */ 339 first_in_ns = (((first >> 32) * NSEC_IN_SEC) + 340 (((uint64_t)NSEC_IN_SEC * (uint32_t)first) >> 32)); 341 342 next_time = xen_fetch_vcpu_time(vcpu) + first_in_ns; 343 error = xentimer_vcpu_start_timer(cpu, next_time); 344 if (error) 345 panic("%s: Error %d setting singleshot timer to %"PRIu64"\n", 346 device_get_nameunit(sc->dev), error, next_time); 347 348 pcpu->timer = next_time; 349 return (error); 350 } 351 352 /** 353 * \brief Cancel the event timer's currently running timer, if any. 354 */ 355 static int 356 xentimer_et_stop(struct eventtimer *et) 357 { 358 int cpu = PCPU_GET(vcpu_id); 359 struct xentimer_pcpu_data *pcpu = DPCPU_PTR(xentimer_pcpu); 360 361 pcpu->timer = 0; 362 return (xentimer_vcpu_stop_timer(cpu)); 363 } 364 365 /** 366 * \brief Attach a Xen PV timer driver instance. 367 * 368 * \param dev Bus device object to attach. 369 * 370 * \note 371 * \returns EINVAL 372 */ 373 static int 374 xentimer_attach(device_t dev) 375 { 376 struct xentimer_softc *sc = device_get_softc(dev); 377 int error, i; 378 379 sc->dev = dev; 380 381 /* Bind an event channel to a VIRQ on each VCPU. */ 382 CPU_FOREACH(i) { 383 struct xentimer_pcpu_data *pcpu; 384 385 pcpu = DPCPU_ID_PTR(i, xentimer_pcpu); 386 error = HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, i, NULL); 387 if (error) { 388 device_printf(dev, "Error disabling Xen periodic timer " 389 "on CPU %d\n", i); 390 return (error); 391 } 392 393 error = xen_intr_bind_virq(dev, VIRQ_TIMER, i, xentimer_intr, 394 NULL, sc, INTR_TYPE_CLK, &pcpu->irq_handle); 395 if (error) { 396 device_printf(dev, "Error %d binding VIRQ_TIMER " 397 "to VCPU %d\n", error, i); 398 return (error); 399 } 400 xen_intr_describe(pcpu->irq_handle, "c%d", i); 401 } 402 403 /* Register the event timer. */ 404 sc->et.et_name = "XENTIMER"; 405 sc->et.et_quality = XENTIMER_QUALITY; 406 sc->et.et_flags = ET_FLAGS_ONESHOT | ET_FLAGS_PERCPU; 407 sc->et.et_frequency = NSEC_IN_SEC; 408 /* See tstosbt() for this formula */ 409 sc->et.et_min_period = (XENTIMER_MIN_PERIOD_IN_NSEC * 410 (((uint64_t)1 << 63) / 500000000) >> 32); 411 sc->et.et_max_period = ((sbintime_t)4 << 32); 412 sc->et.et_start = xentimer_et_start; 413 sc->et.et_stop = xentimer_et_stop; 414 sc->et.et_priv = sc; 415 et_register(&sc->et); 416 417 /* Register the timecounter. */ 418 sc->tc.tc_name = "XENTIMER"; 419 sc->tc.tc_quality = XENTIMER_QUALITY; 420 /* 421 * FIXME: due to the lack of ordering during resume, FreeBSD cannot 422 * guarantee that the Xen PV timer is resumed before any other device 423 * attempts to make use of it, so mark it as not safe for suspension 424 * (ie: remove the TC_FLAGS_SUSPEND_SAFE flag). 425 * 426 * NB: This was not a problem in previous FreeBSD versions because the 427 * timer was directly attached to the nexus, but it is an issue now 428 * that the timer is attached to the xenpv bus, and thus resumed 429 * later. 430 * 431 * sc->tc.tc_flags = TC_FLAGS_SUSPEND_SAFE; 432 */ 433 /* 434 * The underlying resolution is in nanoseconds, since the timer info 435 * scales TSC frequencies using a fraction that represents time in 436 * terms of nanoseconds. 437 */ 438 sc->tc.tc_frequency = NSEC_IN_SEC; 439 sc->tc.tc_counter_mask = ~0u; 440 sc->tc.tc_get_timecount = xentimer_get_timecount; 441 sc->tc.tc_priv = sc; 442 tc_init(&sc->tc); 443 444 /* Register the Hypervisor wall clock */ 445 clock_register(dev, XENCLOCK_RESOLUTION); 446 447 return (0); 448 } 449 450 static int 451 xentimer_detach(device_t dev) 452 { 453 454 /* Implement Xen PV clock teardown - XXX see hpet_detach ? */ 455 /* If possible: 456 * 1. need to deregister timecounter 457 * 2. need to deregister event timer 458 * 3. need to deregister virtual IRQ event channels 459 */ 460 return (EBUSY); 461 } 462 463 static void 464 xentimer_percpu_resume(void *arg) 465 { 466 device_t dev = (device_t) arg; 467 struct xentimer_softc *sc = device_get_softc(dev); 468 469 xentimer_et_start(&sc->et, sc->et.et_min_period, 0); 470 } 471 472 static int 473 xentimer_resume(device_t dev) 474 { 475 int error; 476 int i; 477 478 /* Disable the periodic timer */ 479 CPU_FOREACH(i) { 480 error = HYPERVISOR_vcpu_op(VCPUOP_stop_periodic_timer, i, NULL); 481 if (error != 0) { 482 device_printf(dev, 483 "Error disabling Xen periodic timer on CPU %d\n", 484 i); 485 return (error); 486 } 487 } 488 489 /* Reset the last uptime value */ 490 pvclock_resume(); 491 492 /* Reset the RTC clock */ 493 inittodr(time_second); 494 495 /* Kick the timers on all CPUs */ 496 smp_rendezvous(NULL, xentimer_percpu_resume, NULL, dev); 497 498 if (bootverbose) 499 device_printf(dev, "resumed operation after suspension\n"); 500 501 return (0); 502 } 503 504 static int 505 xentimer_suspend(device_t dev) 506 { 507 return (0); 508 } 509 510 /* 511 * Xen early clock init 512 */ 513 void 514 xen_clock_init(void) 515 { 516 } 517 518 /* 519 * Xen PV DELAY function 520 * 521 * When running on PVH mode we don't have an emulated i8524, so 522 * make use of the Xen time info in order to code a simple DELAY 523 * function that can be used during early boot. 524 */ 525 void 526 xen_delay(int n) 527 { 528 struct vcpu_info *vcpu = &HYPERVISOR_shared_info->vcpu_info[0]; 529 uint64_t end_ns; 530 uint64_t current; 531 532 end_ns = xen_fetch_vcpu_time(vcpu); 533 end_ns += n * NSEC_IN_USEC; 534 535 for (;;) { 536 current = xen_fetch_vcpu_time(vcpu); 537 if (current >= end_ns) 538 break; 539 } 540 } 541 542 static device_method_t xentimer_methods[] = { 543 DEVMETHOD(device_identify, xentimer_identify), 544 DEVMETHOD(device_probe, xentimer_probe), 545 DEVMETHOD(device_attach, xentimer_attach), 546 DEVMETHOD(device_detach, xentimer_detach), 547 DEVMETHOD(device_suspend, xentimer_suspend), 548 DEVMETHOD(device_resume, xentimer_resume), 549 /* clock interface */ 550 DEVMETHOD(clock_gettime, xentimer_gettime), 551 DEVMETHOD(clock_settime, xentimer_settime), 552 DEVMETHOD_END 553 }; 554 555 static driver_t xentimer_driver = { 556 "xen_et", 557 xentimer_methods, 558 sizeof(struct xentimer_softc), 559 }; 560 561 DRIVER_MODULE(xentimer, xenpv, xentimer_driver, xentimer_devclass, 0, 0); 562 MODULE_DEPEND(xentimer, xenpv, 1, 1, 1); 563