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 <linux/hyperv.h>
25#include <clocksource/hyperv_timer.h>
26#include <asm/hyperv-tlfs.h>
27#include <asm/mshyperv.h>
28
29static struct clock_event_device __percpu *hv_clock_event;
30static u64 hv_sched_clock_offset __ro_after_init;
31
32/*
33 * If false, we're using the old mechanism for stimer0 interrupts
34 * where it sends a VMbus message when it expires. The old
35 * mechanism is used when running on older versions of Hyper-V
36 * that don't support Direct Mode. While Hyper-V provides
37 * four stimer's per CPU, Linux uses only stimer0.
38 *
39 * Because Direct Mode does not require processing a VMbus
40 * message, stimer interrupts can be enabled earlier in the
41 * process of booting a CPU, and consistent with when timer
42 * interrupts are enabled for other clocksource drivers.
43 * However, for legacy versions of Hyper-V when Direct Mode
44 * is not enabled, setting up stimer interrupts must be
45 * delayed until VMbus is initialized and can process the
46 * interrupt message.
47 */
48static bool direct_mode_enabled;
49
50static int stimer0_irq = -1;
51static int stimer0_message_sint;
52static __maybe_unused DEFINE_PER_CPU(long, stimer0_evt);
53
54/*
55 * Common code for stimer0 interrupts coming via Direct Mode or
56 * as a VMbus message.
57 */
58void hv_stimer0_isr(void)
59{
60 struct clock_event_device *ce;
61
62 ce = this_cpu_ptr(hv_clock_event);
63 ce->event_handler(ce);
64}
65EXPORT_SYMBOL_GPL(hv_stimer0_isr);
66
67/*
68 * stimer0 interrupt handler for architectures that support
69 * per-cpu interrupts, which also implies Direct Mode.
70 */
71static irqreturn_t __maybe_unused hv_stimer0_percpu_isr(int irq, void *dev_id)
72{
73 hv_stimer0_isr();
74 return IRQ_HANDLED;
75}
76
77static int hv_ce_set_next_event(unsigned long delta,
78 struct clock_event_device *evt)
79{
80 u64 current_tick;
81
82 current_tick = hv_read_reference_counter();
83 current_tick += delta;
84 hv_set_msr(HV_MSR_STIMER0_COUNT, value: current_tick);
85 return 0;
86}
87
88static int hv_ce_shutdown(struct clock_event_device *evt)
89{
90 hv_set_msr(HV_MSR_STIMER0_COUNT, value: 0);
91 hv_set_msr(HV_MSR_STIMER0_CONFIG, value: 0);
92 if (direct_mode_enabled && stimer0_irq >= 0)
93 disable_percpu_irq(irq: stimer0_irq);
94
95 return 0;
96}
97
98static int hv_ce_set_oneshot(struct clock_event_device *evt)
99{
100 union hv_stimer_config timer_cfg;
101
102 timer_cfg.as_uint64 = 0;
103 timer_cfg.enable = 1;
104 timer_cfg.auto_enable = 1;
105 if (direct_mode_enabled) {
106 /*
107 * When it expires, the timer will directly interrupt
108 * on the specified hardware vector/IRQ.
109 */
110 timer_cfg.direct_mode = 1;
111 timer_cfg.apic_vector = HYPERV_STIMER0_VECTOR;
112 if (stimer0_irq >= 0)
113 enable_percpu_irq(irq: stimer0_irq, type: IRQ_TYPE_NONE);
114 } else {
115 /*
116 * When it expires, the timer will generate a VMbus message,
117 * to be handled by the normal VMbus interrupt handler.
118 */
119 timer_cfg.direct_mode = 0;
120 timer_cfg.sintx = stimer0_message_sint;
121 }
122 hv_set_msr(HV_MSR_STIMER0_CONFIG, value: timer_cfg.as_uint64);
123 return 0;
124}
125
126/*
127 * hv_stimer_init - Per-cpu initialization of the clockevent
128 */
129static int hv_stimer_init(unsigned int cpu)
130{
131 struct clock_event_device *ce;
132
133 if (!hv_clock_event)
134 return 0;
135
136 ce = per_cpu_ptr(hv_clock_event, cpu);
137 ce->name = "Hyper-V clockevent";
138 ce->features = CLOCK_EVT_FEAT_ONESHOT;
139 ce->cpumask = cpumask_of(cpu);
140 ce->rating = 1000;
141 ce->set_state_shutdown = hv_ce_shutdown;
142 ce->set_state_oneshot = hv_ce_set_oneshot;
143 ce->set_next_event = hv_ce_set_next_event;
144
145 clockevents_config_and_register(dev: ce,
146 HV_CLOCK_HZ,
147 HV_MIN_DELTA_TICKS,
148 HV_MAX_MAX_DELTA_TICKS);
149 return 0;
150}
151
152/*
153 * hv_stimer_cleanup - Per-cpu cleanup of the clockevent
154 */
155int hv_stimer_cleanup(unsigned int cpu)
156{
157 struct clock_event_device *ce;
158
159 if (!hv_clock_event)
160 return 0;
161
162 /*
163 * In the legacy case where Direct Mode is not enabled
164 * (which can only be on x86/64), stimer cleanup happens
165 * relatively early in the CPU offlining process. We
166 * must unbind the stimer-based clockevent device so
167 * that the LAPIC timer can take over until clockevents
168 * are no longer needed in the offlining process. Note
169 * that clockevents_unbind_device() eventually calls
170 * hv_ce_shutdown().
171 *
172 * The unbind should not be done when Direct Mode is
173 * enabled because we may be on an architecture where
174 * there are no other clockevent devices to fallback to.
175 */
176 ce = per_cpu_ptr(hv_clock_event, cpu);
177 if (direct_mode_enabled)
178 hv_ce_shutdown(evt: ce);
179 else
180 clockevents_unbind_device(ced: ce, cpu);
181
182 return 0;
183}
184EXPORT_SYMBOL_GPL(hv_stimer_cleanup);
185
186/*
187 * These placeholders are overridden by arch specific code on
188 * architectures that need special setup of the stimer0 IRQ because
189 * they don't support per-cpu IRQs (such as x86/x64).
190 */
191void __weak hv_setup_stimer0_handler(void (*handler)(void))
192{
193};
194
195void __weak hv_remove_stimer0_handler(void)
196{
197};
198
199#ifdef CONFIG_ACPI
200/* Called only on architectures with per-cpu IRQs (i.e., not x86/x64) */
201static int hv_setup_stimer0_irq(void)
202{
203 int ret;
204
205 ret = acpi_register_gsi(NULL, HYPERV_STIMER0_VECTOR,
206 ACPI_EDGE_SENSITIVE, ACPI_ACTIVE_HIGH);
207 if (ret < 0) {
208 pr_err("Can't register Hyper-V stimer0 GSI. Error %d", ret);
209 return ret;
210 }
211 stimer0_irq = ret;
212
213 ret = request_percpu_irq(irq: stimer0_irq, handler: hv_stimer0_percpu_isr,
214 devname: "Hyper-V stimer0", percpu_dev_id: &stimer0_evt);
215 if (ret) {
216 pr_err("Can't request Hyper-V stimer0 IRQ %d. Error %d",
217 stimer0_irq, ret);
218 acpi_unregister_gsi(gsi: stimer0_irq);
219 stimer0_irq = -1;
220 }
221 return ret;
222}
223
224static void hv_remove_stimer0_irq(void)
225{
226 if (stimer0_irq == -1) {
227 hv_remove_stimer0_handler();
228 } else {
229 free_percpu_irq(stimer0_irq, &stimer0_evt);
230 acpi_unregister_gsi(gsi: stimer0_irq);
231 stimer0_irq = -1;
232 }
233}
234#else
235static int hv_setup_stimer0_irq(void)
236{
237 return 0;
238}
239
240static void hv_remove_stimer0_irq(void)
241{
242}
243#endif
244
245/* hv_stimer_alloc - Global initialization of the clockevent and stimer0 */
246int hv_stimer_alloc(bool have_percpu_irqs)
247{
248 int ret;
249
250 /*
251 * Synthetic timers are always available except on old versions of
252 * Hyper-V on x86. In that case, return as error as Linux will use a
253 * clockevent based on emulated LAPIC timer hardware.
254 */
255 if (!(ms_hyperv.features & HV_MSR_SYNTIMER_AVAILABLE))
256 return -EINVAL;
257
258 hv_clock_event = alloc_percpu(struct clock_event_device);
259 if (!hv_clock_event)
260 return -ENOMEM;
261
262 direct_mode_enabled = ms_hyperv.misc_features &
263 HV_STIMER_DIRECT_MODE_AVAILABLE;
264
265 /*
266 * If Direct Mode isn't enabled, the remainder of the initialization
267 * is done later by hv_stimer_legacy_init()
268 */
269 if (!direct_mode_enabled)
270 return 0;
271
272 if (have_percpu_irqs) {
273 ret = hv_setup_stimer0_irq();
274 if (ret)
275 goto free_clock_event;
276 } else {
277 hv_setup_stimer0_handler(handler: hv_stimer0_isr);
278 }
279
280 /*
281 * Since we are in Direct Mode, stimer initialization
282 * can be done now with a CPUHP value in the same range
283 * as other clockevent devices.
284 */
285 ret = cpuhp_setup_state(state: CPUHP_AP_HYPERV_TIMER_STARTING,
286 name: "clockevents/hyperv/stimer:starting",
287 startup: hv_stimer_init, teardown: hv_stimer_cleanup);
288 if (ret < 0) {
289 hv_remove_stimer0_irq();
290 goto free_clock_event;
291 }
292 return ret;
293
294free_clock_event:
295 free_percpu(pdata: hv_clock_event);
296 hv_clock_event = NULL;
297 return ret;
298}
299EXPORT_SYMBOL_GPL(hv_stimer_alloc);
300
301/*
302 * hv_stimer_legacy_init -- Called from the VMbus driver to handle
303 * the case when Direct Mode is not enabled, and the stimer
304 * must be initialized late in the CPU onlining process.
305 *
306 */
307void hv_stimer_legacy_init(unsigned int cpu, int sint)
308{
309 if (direct_mode_enabled)
310 return;
311
312 /*
313 * This function gets called by each vCPU, so setting the
314 * global stimer_message_sint value each time is conceptually
315 * not ideal, but the value passed in is always the same and
316 * it avoids introducing yet another interface into this
317 * clocksource driver just to set the sint in the legacy case.
318 */
319 stimer0_message_sint = sint;
320 (void)hv_stimer_init(cpu);
321}
322EXPORT_SYMBOL_GPL(hv_stimer_legacy_init);
323
324/*
325 * hv_stimer_legacy_cleanup -- Called from the VMbus driver to
326 * handle the case when Direct Mode is not enabled, and the
327 * stimer must be cleaned up early in the CPU offlining
328 * process.
329 */
330void hv_stimer_legacy_cleanup(unsigned int cpu)
331{
332 if (direct_mode_enabled)
333 return;
334 (void)hv_stimer_cleanup(cpu);
335}
336EXPORT_SYMBOL_GPL(hv_stimer_legacy_cleanup);
337
338/*
339 * Do a global cleanup of clockevents for the cases of kexec and
340 * vmbus exit
341 */
342void hv_stimer_global_cleanup(void)
343{
344 int cpu;
345
346 /*
347 * hv_stime_legacy_cleanup() will stop the stimer if Direct
348 * Mode is not enabled, and fallback to the LAPIC timer.
349 */
350 for_each_present_cpu(cpu) {
351 hv_stimer_legacy_cleanup(cpu);
352 }
353
354 if (!hv_clock_event)
355 return;
356
357 if (direct_mode_enabled) {
358 cpuhp_remove_state(state: CPUHP_AP_HYPERV_TIMER_STARTING);
359 hv_remove_stimer0_irq();
360 stimer0_irq = -1;
361 }
362 free_percpu(pdata: hv_clock_event);
363 hv_clock_event = NULL;
364
365}
366EXPORT_SYMBOL_GPL(hv_stimer_global_cleanup);
367
368static __always_inline u64 read_hv_clock_msr(void)
369{
370 /*
371 * Read the partition counter to get the current tick count. This count
372 * is set to 0 when the partition is created and is incremented in 100
373 * nanosecond units.
374 *
375 * Use hv_raw_get_msr() because this function is used from
376 * noinstr. Notable; while HV_MSR_TIME_REF_COUNT is a synthetic
377 * register it doesn't need the GHCB path.
378 */
379 return hv_raw_get_msr(HV_MSR_TIME_REF_COUNT);
380}
381
382/*
383 * Code and definitions for the Hyper-V clocksources. Two
384 * clocksources are defined: one that reads the Hyper-V defined MSR, and
385 * the other that uses the TSC reference page feature as defined in the
386 * TLFS. The MSR version is for compatibility with old versions of
387 * Hyper-V and 32-bit x86. The TSC reference page version is preferred.
388 */
389
390static union {
391 struct ms_hyperv_tsc_page page;
392 u8 reserved[PAGE_SIZE];
393} tsc_pg __bss_decrypted __aligned(PAGE_SIZE);
394
395static struct ms_hyperv_tsc_page *tsc_page = &tsc_pg.page;
396static unsigned long tsc_pfn;
397
398unsigned long hv_get_tsc_pfn(void)
399{
400 return tsc_pfn;
401}
402EXPORT_SYMBOL_GPL(hv_get_tsc_pfn);
403
404struct ms_hyperv_tsc_page *hv_get_tsc_page(void)
405{
406 return tsc_page;
407}
408EXPORT_SYMBOL_GPL(hv_get_tsc_page);
409
410static __always_inline u64 read_hv_clock_tsc(void)
411{
412 u64 cur_tsc, time;
413
414 /*
415 * The Hyper-V Top-Level Function Spec (TLFS), section Timers,
416 * subsection Refererence Counter, guarantees that the TSC and MSR
417 * times are in sync and monotonic. Therefore we can fall back
418 * to the MSR in case the TSC page indicates unavailability.
419 */
420 if (!hv_read_tsc_page_tsc(tsc_pg: tsc_page, cur_tsc: &cur_tsc, time: &time))
421 time = read_hv_clock_msr();
422
423 return time;
424}
425
426static u64 notrace read_hv_clock_tsc_cs(struct clocksource *arg)
427{
428 return read_hv_clock_tsc();
429}
430
431static u64 noinstr read_hv_sched_clock_tsc(void)
432{
433 return (read_hv_clock_tsc() - hv_sched_clock_offset) *
434 (NSEC_PER_SEC / HV_CLOCK_HZ);
435}
436
437static void suspend_hv_clock_tsc(struct clocksource *arg)
438{
439 union hv_reference_tsc_msr tsc_msr;
440
441 /* Disable the TSC page */
442 tsc_msr.as_uint64 = hv_get_msr(HV_MSR_REFERENCE_TSC);
443 tsc_msr.enable = 0;
444 hv_set_msr(HV_MSR_REFERENCE_TSC, value: tsc_msr.as_uint64);
445}
446
447
448static void resume_hv_clock_tsc(struct clocksource *arg)
449{
450 union hv_reference_tsc_msr tsc_msr;
451
452 /* Re-enable the TSC page */
453 tsc_msr.as_uint64 = hv_get_msr(HV_MSR_REFERENCE_TSC);
454 tsc_msr.enable = 1;
455 tsc_msr.pfn = tsc_pfn;
456 hv_set_msr(HV_MSR_REFERENCE_TSC, value: tsc_msr.as_uint64);
457}
458
459#ifdef HAVE_VDSO_CLOCKMODE_HVCLOCK
460static int hv_cs_enable(struct clocksource *cs)
461{
462 vclocks_set_used(which: VDSO_CLOCKMODE_HVCLOCK);
463 return 0;
464}
465#endif
466
467static struct clocksource hyperv_cs_tsc = {
468 .name = "hyperv_clocksource_tsc_page",
469 .rating = 500,
470 .read = read_hv_clock_tsc_cs,
471 .mask = CLOCKSOURCE_MASK(64),
472 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
473 .suspend= suspend_hv_clock_tsc,
474 .resume = resume_hv_clock_tsc,
475#ifdef HAVE_VDSO_CLOCKMODE_HVCLOCK
476 .enable = hv_cs_enable,
477 .vdso_clock_mode = VDSO_CLOCKMODE_HVCLOCK,
478#else
479 .vdso_clock_mode = VDSO_CLOCKMODE_NONE,
480#endif
481};
482
483static u64 notrace read_hv_clock_msr_cs(struct clocksource *arg)
484{
485 return read_hv_clock_msr();
486}
487
488static struct clocksource hyperv_cs_msr = {
489 .name = "hyperv_clocksource_msr",
490 .rating = 495,
491 .read = read_hv_clock_msr_cs,
492 .mask = CLOCKSOURCE_MASK(64),
493 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
494};
495
496/*
497 * Reference to pv_ops must be inline so objtool
498 * detection of noinstr violations can work correctly.
499 */
500#ifdef CONFIG_GENERIC_SCHED_CLOCK
501static __always_inline void hv_setup_sched_clock(void *sched_clock)
502{
503 /*
504 * We're on an architecture with generic sched clock (not x86/x64).
505 * The Hyper-V sched clock read function returns nanoseconds, not
506 * the normal 100ns units of the Hyper-V synthetic clock.
507 */
508 sched_clock_register(sched_clock, 64, NSEC_PER_SEC);
509}
510#elif defined CONFIG_PARAVIRT
511static __always_inline void hv_setup_sched_clock(void *sched_clock)
512{
513 /* We're on x86/x64 *and* using PV ops */
514 paravirt_set_sched_clock(func: sched_clock);
515}
516#else /* !CONFIG_GENERIC_SCHED_CLOCK && !CONFIG_PARAVIRT */
517static __always_inline void hv_setup_sched_clock(void *sched_clock) {}
518#endif /* CONFIG_GENERIC_SCHED_CLOCK */
519
520static void __init hv_init_tsc_clocksource(void)
521{
522 union hv_reference_tsc_msr tsc_msr;
523
524 /*
525 * If Hyper-V offers TSC_INVARIANT, then the virtualized TSC correctly
526 * handles frequency and offset changes due to live migration,
527 * pause/resume, and other VM management operations. So lower the
528 * Hyper-V Reference TSC rating, causing the generic TSC to be used.
529 * TSC_INVARIANT is not offered on ARM64, so the Hyper-V Reference
530 * TSC will be preferred over the virtualized ARM64 arch counter.
531 */
532 if (ms_hyperv.features & HV_ACCESS_TSC_INVARIANT) {
533 hyperv_cs_tsc.rating = 250;
534 hyperv_cs_msr.rating = 245;
535 }
536
537 if (!(ms_hyperv.features & HV_MSR_REFERENCE_TSC_AVAILABLE))
538 return;
539
540 hv_read_reference_counter = read_hv_clock_tsc;
541
542 /*
543 * TSC page mapping works differently in root compared to guest.
544 * - In guest partition the guest PFN has to be passed to the
545 * hypervisor.
546 * - In root partition it's other way around: it has to map the PFN
547 * provided by the hypervisor.
548 * But it can't be mapped right here as it's too early and MMU isn't
549 * ready yet. So, we only set the enable bit here and will remap the
550 * page later in hv_remap_tsc_clocksource().
551 *
552 * It worth mentioning, that TSC clocksource read function
553 * (read_hv_clock_tsc) has a MSR-based fallback mechanism, used when
554 * TSC page is zeroed (which is the case until the PFN is remapped) and
555 * thus TSC clocksource will work even without the real TSC page
556 * mapped.
557 */
558 tsc_msr.as_uint64 = hv_get_msr(HV_MSR_REFERENCE_TSC);
559 if (hv_root_partition)
560 tsc_pfn = tsc_msr.pfn;
561 else
562 tsc_pfn = HVPFN_DOWN(virt_to_phys(tsc_page));
563 tsc_msr.enable = 1;
564 tsc_msr.pfn = tsc_pfn;
565 hv_set_msr(HV_MSR_REFERENCE_TSC, value: tsc_msr.as_uint64);
566
567 clocksource_register_hz(cs: &hyperv_cs_tsc, NSEC_PER_SEC/100);
568
569 /*
570 * If TSC is invariant, then let it stay as the sched clock since it
571 * will be faster than reading the TSC page. But if not invariant, use
572 * the TSC page so that live migrations across hosts with different
573 * frequencies is handled correctly.
574 */
575 if (!(ms_hyperv.features & HV_ACCESS_TSC_INVARIANT)) {
576 hv_sched_clock_offset = hv_read_reference_counter();
577 hv_setup_sched_clock(sched_clock: read_hv_sched_clock_tsc);
578 }
579}
580
581void __init hv_init_clocksource(void)
582{
583 /*
584 * Try to set up the TSC page clocksource, then the MSR clocksource.
585 * At least one of these will always be available except on very old
586 * versions of Hyper-V on x86. In that case we won't have a Hyper-V
587 * clocksource, but Linux will still run with a clocksource based
588 * on the emulated PIT or LAPIC timer.
589 *
590 * Never use the MSR clocksource as sched clock. It's too slow.
591 * Better to use the native sched clock as the fallback.
592 */
593 hv_init_tsc_clocksource();
594
595 if (ms_hyperv.features & HV_MSR_TIME_REF_COUNT_AVAILABLE)
596 clocksource_register_hz(cs: &hyperv_cs_msr, NSEC_PER_SEC/100);
597}
598
599void __init hv_remap_tsc_clocksource(void)
600{
601 if (!(ms_hyperv.features & HV_MSR_REFERENCE_TSC_AVAILABLE))
602 return;
603
604 if (!hv_root_partition) {
605 WARN(1, "%s: attempt to remap TSC page in guest partition\n",
606 __func__);
607 return;
608 }
609
610 tsc_page = memremap(offset: tsc_pfn << HV_HYP_PAGE_SHIFT, size: sizeof(tsc_pg),
611 flags: MEMREMAP_WB);
612 if (!tsc_page)
613 pr_err("Failed to remap Hyper-V TSC page.\n");
614}
615

source code of linux/drivers/clocksource/hyperv_timer.c