1 // SPDX-License-Identifier: GPL-2.0
2 
3 /*
4  * Clocksource driver for the synthetic counter and timers
5  * provided by the Hyper-V hypervisor to guest VMs, as described
6  * in the Hyper-V Top Level Functional Spec (TLFS). This driver
7  * is instruction set architecture independent.
8  *
9  * Copyright (C) 2019, Microsoft, Inc.
10  *
11  * Author:  Michael Kelley <mikelley@microsoft.com>
12  */
13 
14 #include <linux/percpu.h>
15 #include <linux/cpumask.h>
16 #include <linux/clockchips.h>
17 #include <linux/clocksource.h>
18 #include <linux/sched_clock.h>
19 #include <linux/mm.h>
20 #include <linux/cpuhotplug.h>
21 #include <linux/interrupt.h>
22 #include <linux/irq.h>
23 #include <linux/acpi.h>
24 #include <clocksource/hyperv_timer.h>
25 #include <asm/hyperv-tlfs.h>
26 #include <asm/mshyperv.h>
27 
28 static struct clock_event_device __percpu *hv_clock_event;
29 static u64 hv_sched_clock_offset __ro_after_init;
30 
31 /*
32  * If false, we're using the old mechanism for stimer0 interrupts
33  * where it sends a VMbus message when it expires. The old
34  * mechanism is used when running on older versions of Hyper-V
35  * that don't support Direct Mode. While Hyper-V provides
36  * four stimer's per CPU, Linux uses only stimer0.
37  *
38  * Because Direct Mode does not require processing a VMbus
39  * message, stimer interrupts can be enabled earlier in the
40  * process of booting a CPU, and consistent with when timer
41  * interrupts are enabled for other clocksource drivers.
42  * However, for legacy versions of Hyper-V when Direct Mode
43  * is not enabled, setting up stimer interrupts must be
44  * delayed until VMbus is initialized and can process the
45  * interrupt message.
46  */
47 static bool direct_mode_enabled;
48 
49 static int stimer0_irq = -1;
50 static int stimer0_message_sint;
51 static DEFINE_PER_CPU(long, stimer0_evt);
52 
53 /*
54  * Common code for stimer0 interrupts coming via Direct Mode or
55  * as a VMbus message.
56  */
57 void hv_stimer0_isr(void)
58 {
59 	struct clock_event_device *ce;
60 
61 	ce = this_cpu_ptr(hv_clock_event);
62 	ce->event_handler(ce);
63 }
64 EXPORT_SYMBOL_GPL(hv_stimer0_isr);
65 
66 /*
67  * stimer0 interrupt handler for architectures that support
68  * per-cpu interrupts, which also implies Direct Mode.
69  */
70 static irqreturn_t hv_stimer0_percpu_isr(int irq, void *dev_id)
71 {
72 	hv_stimer0_isr();
73 	return IRQ_HANDLED;
74 }
75 
76 static int hv_ce_set_next_event(unsigned long delta,
77 				struct clock_event_device *evt)
78 {
79 	u64 current_tick;
80 
81 	current_tick = hv_read_reference_counter();
82 	current_tick += delta;
83 	hv_set_register(HV_REGISTER_STIMER0_COUNT, current_tick);
84 	return 0;
85 }
86 
87 static int hv_ce_shutdown(struct clock_event_device *evt)
88 {
89 	hv_set_register(HV_REGISTER_STIMER0_COUNT, 0);
90 	hv_set_register(HV_REGISTER_STIMER0_CONFIG, 0);
91 	if (direct_mode_enabled && stimer0_irq >= 0)
92 		disable_percpu_irq(stimer0_irq);
93 
94 	return 0;
95 }
96 
97 static int hv_ce_set_oneshot(struct clock_event_device *evt)
98 {
99 	union hv_stimer_config timer_cfg;
100 
101 	timer_cfg.as_uint64 = 0;
102 	timer_cfg.enable = 1;
103 	timer_cfg.auto_enable = 1;
104 	if (direct_mode_enabled) {
105 		/*
106 		 * When it expires, the timer will directly interrupt
107 		 * on the specified hardware vector/IRQ.
108 		 */
109 		timer_cfg.direct_mode = 1;
110 		timer_cfg.apic_vector = HYPERV_STIMER0_VECTOR;
111 		if (stimer0_irq >= 0)
112 			enable_percpu_irq(stimer0_irq, IRQ_TYPE_NONE);
113 	} else {
114 		/*
115 		 * When it expires, the timer will generate a VMbus message,
116 		 * to be handled by the normal VMbus interrupt handler.
117 		 */
118 		timer_cfg.direct_mode = 0;
119 		timer_cfg.sintx = stimer0_message_sint;
120 	}
121 	hv_set_register(HV_REGISTER_STIMER0_CONFIG, timer_cfg.as_uint64);
122 	return 0;
123 }
124 
125 /*
126  * hv_stimer_init - Per-cpu initialization of the clockevent
127  */
128 static int hv_stimer_init(unsigned int cpu)
129 {
130 	struct clock_event_device *ce;
131 
132 	if (!hv_clock_event)
133 		return 0;
134 
135 	ce = per_cpu_ptr(hv_clock_event, cpu);
136 	ce->name = "Hyper-V clockevent";
137 	ce->features = CLOCK_EVT_FEAT_ONESHOT;
138 	ce->cpumask = cpumask_of(cpu);
139 	ce->rating = 1000;
140 	ce->set_state_shutdown = hv_ce_shutdown;
141 	ce->set_state_oneshot = hv_ce_set_oneshot;
142 	ce->set_next_event = hv_ce_set_next_event;
143 
144 	clockevents_config_and_register(ce,
145 					HV_CLOCK_HZ,
146 					HV_MIN_DELTA_TICKS,
147 					HV_MAX_MAX_DELTA_TICKS);
148 	return 0;
149 }
150 
151 /*
152  * hv_stimer_cleanup - Per-cpu cleanup of the clockevent
153  */
154 int hv_stimer_cleanup(unsigned int cpu)
155 {
156 	struct clock_event_device *ce;
157 
158 	if (!hv_clock_event)
159 		return 0;
160 
161 	/*
162 	 * In the legacy case where Direct Mode is not enabled
163 	 * (which can only be on x86/64), stimer cleanup happens
164 	 * relatively early in the CPU offlining process. We
165 	 * must unbind the stimer-based clockevent device so
166 	 * that the LAPIC timer can take over until clockevents
167 	 * are no longer needed in the offlining process. Note
168 	 * that clockevents_unbind_device() eventually calls
169 	 * hv_ce_shutdown().
170 	 *
171 	 * The unbind should not be done when Direct Mode is
172 	 * enabled because we may be on an architecture where
173 	 * there are no other clockevent devices to fallback to.
174 	 */
175 	ce = per_cpu_ptr(hv_clock_event, cpu);
176 	if (direct_mode_enabled)
177 		hv_ce_shutdown(ce);
178 	else
179 		clockevents_unbind_device(ce, cpu);
180 
181 	return 0;
182 }
183 EXPORT_SYMBOL_GPL(hv_stimer_cleanup);
184 
185 /*
186  * These placeholders are overridden by arch specific code on
187  * architectures that need special setup of the stimer0 IRQ because
188  * they don't support per-cpu IRQs (such as x86/x64).
189  */
190 void __weak hv_setup_stimer0_handler(void (*handler)(void))
191 {
192 };
193 
194 void __weak hv_remove_stimer0_handler(void)
195 {
196 };
197 
198 /* Called only on architectures with per-cpu IRQs (i.e., not x86/x64) */
199 static int hv_setup_stimer0_irq(void)
200 {
201 	int ret;
202 
203 	ret = acpi_register_gsi(NULL, HYPERV_STIMER0_VECTOR,
204 			ACPI_EDGE_SENSITIVE, ACPI_ACTIVE_HIGH);
205 	if (ret < 0) {
206 		pr_err("Can't register Hyper-V stimer0 GSI. Error %d", ret);
207 		return ret;
208 	}
209 	stimer0_irq = ret;
210 
211 	ret = request_percpu_irq(stimer0_irq, hv_stimer0_percpu_isr,
212 		"Hyper-V stimer0", &stimer0_evt);
213 	if (ret) {
214 		pr_err("Can't request Hyper-V stimer0 IRQ %d. Error %d",
215 			stimer0_irq, ret);
216 		acpi_unregister_gsi(stimer0_irq);
217 		stimer0_irq = -1;
218 	}
219 	return ret;
220 }
221 
222 static void hv_remove_stimer0_irq(void)
223 {
224 	if (stimer0_irq == -1) {
225 		hv_remove_stimer0_handler();
226 	} else {
227 		free_percpu_irq(stimer0_irq, &stimer0_evt);
228 		acpi_unregister_gsi(stimer0_irq);
229 		stimer0_irq = -1;
230 	}
231 }
232 
233 /* hv_stimer_alloc - Global initialization of the clockevent and stimer0 */
234 int hv_stimer_alloc(bool have_percpu_irqs)
235 {
236 	int ret;
237 
238 	/*
239 	 * Synthetic timers are always available except on old versions of
240 	 * Hyper-V on x86.  In that case, return as error as Linux will use a
241 	 * clockevent based on emulated LAPIC timer hardware.
242 	 */
243 	if (!(ms_hyperv.features & HV_MSR_SYNTIMER_AVAILABLE))
244 		return -EINVAL;
245 
246 	hv_clock_event = alloc_percpu(struct clock_event_device);
247 	if (!hv_clock_event)
248 		return -ENOMEM;
249 
250 	direct_mode_enabled = ms_hyperv.misc_features &
251 			HV_STIMER_DIRECT_MODE_AVAILABLE;
252 
253 	/*
254 	 * If Direct Mode isn't enabled, the remainder of the initialization
255 	 * is done later by hv_stimer_legacy_init()
256 	 */
257 	if (!direct_mode_enabled)
258 		return 0;
259 
260 	if (have_percpu_irqs) {
261 		ret = hv_setup_stimer0_irq();
262 		if (ret)
263 			goto free_clock_event;
264 	} else {
265 		hv_setup_stimer0_handler(hv_stimer0_isr);
266 	}
267 
268 	/*
269 	 * Since we are in Direct Mode, stimer initialization
270 	 * can be done now with a CPUHP value in the same range
271 	 * as other clockevent devices.
272 	 */
273 	ret = cpuhp_setup_state(CPUHP_AP_HYPERV_TIMER_STARTING,
274 			"clockevents/hyperv/stimer:starting",
275 			hv_stimer_init, hv_stimer_cleanup);
276 	if (ret < 0) {
277 		hv_remove_stimer0_irq();
278 		goto free_clock_event;
279 	}
280 	return ret;
281 
282 free_clock_event:
283 	free_percpu(hv_clock_event);
284 	hv_clock_event = NULL;
285 	return ret;
286 }
287 EXPORT_SYMBOL_GPL(hv_stimer_alloc);
288 
289 /*
290  * hv_stimer_legacy_init -- Called from the VMbus driver to handle
291  * the case when Direct Mode is not enabled, and the stimer
292  * must be initialized late in the CPU onlining process.
293  *
294  */
295 void hv_stimer_legacy_init(unsigned int cpu, int sint)
296 {
297 	if (direct_mode_enabled)
298 		return;
299 
300 	/*
301 	 * This function gets called by each vCPU, so setting the
302 	 * global stimer_message_sint value each time is conceptually
303 	 * not ideal, but the value passed in is always the same and
304 	 * it avoids introducing yet another interface into this
305 	 * clocksource driver just to set the sint in the legacy case.
306 	 */
307 	stimer0_message_sint = sint;
308 	(void)hv_stimer_init(cpu);
309 }
310 EXPORT_SYMBOL_GPL(hv_stimer_legacy_init);
311 
312 /*
313  * hv_stimer_legacy_cleanup -- Called from the VMbus driver to
314  * handle the case when Direct Mode is not enabled, and the
315  * stimer must be cleaned up early in the CPU offlining
316  * process.
317  */
318 void hv_stimer_legacy_cleanup(unsigned int cpu)
319 {
320 	if (direct_mode_enabled)
321 		return;
322 	(void)hv_stimer_cleanup(cpu);
323 }
324 EXPORT_SYMBOL_GPL(hv_stimer_legacy_cleanup);
325 
326 /*
327  * Do a global cleanup of clockevents for the cases of kexec and
328  * vmbus exit
329  */
330 void hv_stimer_global_cleanup(void)
331 {
332 	int	cpu;
333 
334 	/*
335 	 * hv_stime_legacy_cleanup() will stop the stimer if Direct
336 	 * Mode is not enabled, and fallback to the LAPIC timer.
337 	 */
338 	for_each_present_cpu(cpu) {
339 		hv_stimer_legacy_cleanup(cpu);
340 	}
341 
342 	if (!hv_clock_event)
343 		return;
344 
345 	if (direct_mode_enabled) {
346 		cpuhp_remove_state(CPUHP_AP_HYPERV_TIMER_STARTING);
347 		hv_remove_stimer0_irq();
348 		stimer0_irq = -1;
349 	}
350 	free_percpu(hv_clock_event);
351 	hv_clock_event = NULL;
352 
353 }
354 EXPORT_SYMBOL_GPL(hv_stimer_global_cleanup);
355 
356 /*
357  * Code and definitions for the Hyper-V clocksources.  Two
358  * clocksources are defined: one that reads the Hyper-V defined MSR, and
359  * the other that uses the TSC reference page feature as defined in the
360  * TLFS.  The MSR version is for compatibility with old versions of
361  * Hyper-V and 32-bit x86.  The TSC reference page version is preferred.
362  */
363 
364 u64 (*hv_read_reference_counter)(void);
365 EXPORT_SYMBOL_GPL(hv_read_reference_counter);
366 
367 static union {
368 	struct ms_hyperv_tsc_page page;
369 	u8 reserved[PAGE_SIZE];
370 } tsc_pg __aligned(PAGE_SIZE);
371 
372 struct ms_hyperv_tsc_page *hv_get_tsc_page(void)
373 {
374 	return &tsc_pg.page;
375 }
376 EXPORT_SYMBOL_GPL(hv_get_tsc_page);
377 
378 static u64 notrace read_hv_clock_tsc(void)
379 {
380 	u64 current_tick = hv_read_tsc_page(hv_get_tsc_page());
381 
382 	if (current_tick == U64_MAX)
383 		current_tick = hv_get_register(HV_REGISTER_TIME_REF_COUNT);
384 
385 	return current_tick;
386 }
387 
388 static u64 notrace read_hv_clock_tsc_cs(struct clocksource *arg)
389 {
390 	return read_hv_clock_tsc();
391 }
392 
393 static u64 notrace read_hv_sched_clock_tsc(void)
394 {
395 	return (read_hv_clock_tsc() - hv_sched_clock_offset) *
396 		(NSEC_PER_SEC / HV_CLOCK_HZ);
397 }
398 
399 static void suspend_hv_clock_tsc(struct clocksource *arg)
400 {
401 	u64 tsc_msr;
402 
403 	/* Disable the TSC page */
404 	tsc_msr = hv_get_register(HV_REGISTER_REFERENCE_TSC);
405 	tsc_msr &= ~BIT_ULL(0);
406 	hv_set_register(HV_REGISTER_REFERENCE_TSC, tsc_msr);
407 }
408 
409 
410 static void resume_hv_clock_tsc(struct clocksource *arg)
411 {
412 	phys_addr_t phys_addr = virt_to_phys(&tsc_pg);
413 	u64 tsc_msr;
414 
415 	/* Re-enable the TSC page */
416 	tsc_msr = hv_get_register(HV_REGISTER_REFERENCE_TSC);
417 	tsc_msr &= GENMASK_ULL(11, 0);
418 	tsc_msr |= BIT_ULL(0) | (u64)phys_addr;
419 	hv_set_register(HV_REGISTER_REFERENCE_TSC, tsc_msr);
420 }
421 
422 #ifdef HAVE_VDSO_CLOCKMODE_HVCLOCK
423 static int hv_cs_enable(struct clocksource *cs)
424 {
425 	vclocks_set_used(VDSO_CLOCKMODE_HVCLOCK);
426 	return 0;
427 }
428 #endif
429 
430 static struct clocksource hyperv_cs_tsc = {
431 	.name	= "hyperv_clocksource_tsc_page",
432 	.rating	= 500,
433 	.read	= read_hv_clock_tsc_cs,
434 	.mask	= CLOCKSOURCE_MASK(64),
435 	.flags	= CLOCK_SOURCE_IS_CONTINUOUS,
436 	.suspend= suspend_hv_clock_tsc,
437 	.resume	= resume_hv_clock_tsc,
438 #ifdef HAVE_VDSO_CLOCKMODE_HVCLOCK
439 	.enable = hv_cs_enable,
440 	.vdso_clock_mode = VDSO_CLOCKMODE_HVCLOCK,
441 #else
442 	.vdso_clock_mode = VDSO_CLOCKMODE_NONE,
443 #endif
444 };
445 
446 static u64 notrace read_hv_clock_msr(void)
447 {
448 	/*
449 	 * Read the partition counter to get the current tick count. This count
450 	 * is set to 0 when the partition is created and is incremented in
451 	 * 100 nanosecond units.
452 	 */
453 	return hv_get_register(HV_REGISTER_TIME_REF_COUNT);
454 }
455 
456 static u64 notrace read_hv_clock_msr_cs(struct clocksource *arg)
457 {
458 	return read_hv_clock_msr();
459 }
460 
461 static u64 notrace read_hv_sched_clock_msr(void)
462 {
463 	return (read_hv_clock_msr() - hv_sched_clock_offset) *
464 		(NSEC_PER_SEC / HV_CLOCK_HZ);
465 }
466 
467 static struct clocksource hyperv_cs_msr = {
468 	.name	= "hyperv_clocksource_msr",
469 	.rating	= 500,
470 	.read	= read_hv_clock_msr_cs,
471 	.mask	= CLOCKSOURCE_MASK(64),
472 	.flags	= CLOCK_SOURCE_IS_CONTINUOUS,
473 };
474 
475 /*
476  * Reference to pv_ops must be inline so objtool
477  * detection of noinstr violations can work correctly.
478  */
479 #ifdef CONFIG_GENERIC_SCHED_CLOCK
480 static __always_inline void hv_setup_sched_clock(void *sched_clock)
481 {
482 	/*
483 	 * We're on an architecture with generic sched clock (not x86/x64).
484 	 * The Hyper-V sched clock read function returns nanoseconds, not
485 	 * the normal 100ns units of the Hyper-V synthetic clock.
486 	 */
487 	sched_clock_register(sched_clock, 64, NSEC_PER_SEC);
488 }
489 #elif defined CONFIG_PARAVIRT
490 static __always_inline void hv_setup_sched_clock(void *sched_clock)
491 {
492 	/* We're on x86/x64 *and* using PV ops */
493 	paravirt_set_sched_clock(sched_clock);
494 }
495 #else /* !CONFIG_GENERIC_SCHED_CLOCK && !CONFIG_PARAVIRT */
496 static __always_inline void hv_setup_sched_clock(void *sched_clock) {}
497 #endif /* CONFIG_GENERIC_SCHED_CLOCK */
498 
499 static bool __init hv_init_tsc_clocksource(void)
500 {
501 	u64		tsc_msr;
502 	phys_addr_t	phys_addr;
503 
504 	if (!(ms_hyperv.features & HV_MSR_REFERENCE_TSC_AVAILABLE))
505 		return false;
506 
507 	if (hv_root_partition)
508 		return false;
509 
510 	/*
511 	 * If Hyper-V offers TSC_INVARIANT, then the virtualized TSC correctly
512 	 * handles frequency and offset changes due to live migration,
513 	 * pause/resume, and other VM management operations.  So lower the
514 	 * Hyper-V Reference TSC rating, causing the generic TSC to be used.
515 	 * TSC_INVARIANT is not offered on ARM64, so the Hyper-V Reference
516 	 * TSC will be preferred over the virtualized ARM64 arch counter.
517 	 * While the Hyper-V MSR clocksource won't be used since the
518 	 * Reference TSC clocksource is present, change its rating as
519 	 * well for consistency.
520 	 */
521 	if (ms_hyperv.features & HV_ACCESS_TSC_INVARIANT) {
522 		hyperv_cs_tsc.rating = 250;
523 		hyperv_cs_msr.rating = 250;
524 	}
525 
526 	hv_read_reference_counter = read_hv_clock_tsc;
527 	phys_addr = virt_to_phys(hv_get_tsc_page());
528 
529 	/*
530 	 * The Hyper-V TLFS specifies to preserve the value of reserved
531 	 * bits in registers. So read the existing value, preserve the
532 	 * low order 12 bits, and add in the guest physical address
533 	 * (which already has at least the low 12 bits set to zero since
534 	 * it is page aligned). Also set the "enable" bit, which is bit 0.
535 	 */
536 	tsc_msr = hv_get_register(HV_REGISTER_REFERENCE_TSC);
537 	tsc_msr &= GENMASK_ULL(11, 0);
538 	tsc_msr = tsc_msr | 0x1 | (u64)phys_addr;
539 	hv_set_register(HV_REGISTER_REFERENCE_TSC, tsc_msr);
540 
541 	clocksource_register_hz(&hyperv_cs_tsc, NSEC_PER_SEC/100);
542 
543 	hv_sched_clock_offset = hv_read_reference_counter();
544 	hv_setup_sched_clock(read_hv_sched_clock_tsc);
545 
546 	return true;
547 }
548 
549 void __init hv_init_clocksource(void)
550 {
551 	/*
552 	 * Try to set up the TSC page clocksource. If it succeeds, we're
553 	 * done. Otherwise, set up the MSR clocksource.  At least one of
554 	 * these will always be available except on very old versions of
555 	 * Hyper-V on x86.  In that case we won't have a Hyper-V
556 	 * clocksource, but Linux will still run with a clocksource based
557 	 * on the emulated PIT or LAPIC timer.
558 	 */
559 	if (hv_init_tsc_clocksource())
560 		return;
561 
562 	if (!(ms_hyperv.features & HV_MSR_TIME_REF_COUNT_AVAILABLE))
563 		return;
564 
565 	hv_read_reference_counter = read_hv_clock_msr;
566 	clocksource_register_hz(&hyperv_cs_msr, NSEC_PER_SEC/100);
567 
568 	hv_sched_clock_offset = hv_read_reference_counter();
569 	hv_setup_sched_clock(read_hv_sched_clock_msr);
570 }
571 EXPORT_SYMBOL_GPL(hv_init_clocksource);
572