projects/linux·blob master · refs

558 lines · 14.6 KiB

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 <[email protected]>
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 <linux/export.h>
26#include <clocksource/hyperv_timer.h>
27#include <hyperv/hvhdk.h>
28#include <asm/mshyperv.h>
29
30static struct clock_event_device __percpu *hv_clock_event;
31/* Note: offset can hold negative values after hibernation. */
32static u64 hv_sched_clock_offset __read_mostly;
33
34static int stimer0_irq = -1;
35static __maybe_unused DEFINE_PER_CPU(long, stimer0_evt);
36
37static void hv_stimer0_isr(void)
38{
39 struct clock_event_device *ce;
40
41 ce = this_cpu_ptr(hv_clock_event);
42 ce->event_handler(ce);
43}
44
45/*
46 * stimer0 interrupt handler for architectures that support
47 * per-cpu interrupts
48 */
49static irqreturn_t __maybe_unused hv_stimer0_percpu_isr(int irq, void *dev_id)
50{
51 hv_stimer0_isr();
52 return IRQ_HANDLED;
53}
54
55static int hv_ce_set_next_event(unsigned long delta,
56 struct clock_event_device *evt)
57{
58 u64 current_tick;
59
60 current_tick = hv_read_reference_counter();
61 current_tick += delta;
62 hv_set_msr(HV_MSR_STIMER0_COUNT, current_tick);
63 return 0;
64}
65
66static int hv_ce_shutdown(struct clock_event_device *evt)
67{
68 hv_set_msr(HV_MSR_STIMER0_COUNT, 0);
69 hv_set_msr(HV_MSR_STIMER0_CONFIG, 0);
70 if (stimer0_irq >= 0)
71 disable_percpu_irq(stimer0_irq);
72
73 return 0;
74}
75
76static int hv_ce_set_oneshot(struct clock_event_device *evt)
77{
78 union hv_stimer_config timer_cfg;
79
80 timer_cfg.as_uint64 = 0;
81 timer_cfg.enable = 1;
82 timer_cfg.auto_enable = 1;
83
84 /*
85 * When it expires, the timer will directly interrupt
86 * on the specified hardware vector/IRQ.
87 */
88 timer_cfg.direct_mode = 1;
89 timer_cfg.apic_vector = HYPERV_STIMER0_VECTOR;
90 if (stimer0_irq >= 0)
91 enable_percpu_irq(stimer0_irq, IRQ_TYPE_NONE);
92
93 hv_set_msr(HV_MSR_STIMER0_CONFIG, timer_cfg.as_uint64);
94 return 0;
95}
96
97/*
98 * hv_stimer_init - Per-cpu initialization of the clockevent
99 */
100static int hv_stimer_init(unsigned int cpu)
101{
102 struct clock_event_device *ce;
103
104 if (!hv_clock_event)
105 return 0;
106
107 ce = per_cpu_ptr(hv_clock_event, cpu);
108 ce->name = "Hyper-V clockevent";
109 ce->features = CLOCK_EVT_FEAT_ONESHOT;
110 ce->cpumask = cpumask_of(cpu);
111
112 /*
113 * Lower the rating of the Hyper-V timer in a TDX VM without paravisor,
114 * so the local APIC timer (lapic_clockevent) is the default timer in
115 * such a VM. The Hyper-V timer is not preferred in such a VM because
116 * it depends on the slow VM Reference Counter MSR (the Hyper-V TSC
117 * page is not enbled in such a VM because the VM uses Invariant TSC
118 * as a better clocksource and it's challenging to mark the Hyper-V
119 * TSC page shared in very early boot).
120 */
121 if (!ms_hyperv.paravisor_present && hv_isolation_type_tdx())
122 ce->rating = 90;
123 else
124 ce->rating = 1000;
125
126 ce->set_state_shutdown = hv_ce_shutdown;
127 ce->set_state_oneshot = hv_ce_set_oneshot;
128 ce->set_next_event = hv_ce_set_next_event;
129
130 clockevents_config_and_register(ce,
131 HV_CLOCK_HZ,
132 HV_MIN_DELTA_TICKS,
133 HV_MAX_MAX_DELTA_TICKS);
134 return 0;
135}
136
137/*
138 * hv_stimer_cleanup - Per-cpu cleanup of the clockevent
139 */
140int hv_stimer_cleanup(unsigned int cpu)
141{
142 struct clock_event_device *ce;
143
144 if (!hv_clock_event)
145 return 0;
146
147 ce = per_cpu_ptr(hv_clock_event, cpu);
148 hv_ce_shutdown(ce);
149
150 return 0;
151}
152EXPORT_SYMBOL_GPL(hv_stimer_cleanup);
153
154/*
155 * These placeholders are overridden by arch specific code on
156 * architectures that need special setup of the stimer0 IRQ because
157 * they don't support per-cpu IRQs (such as x86/x64).
158 */
159void __weak hv_setup_stimer0_handler(void (*handler)(void))
160{
161};
162
163void __weak hv_remove_stimer0_handler(void)
164{
165};
166
167#ifdef CONFIG_ACPI
168/* Called only on architectures with per-cpu IRQs (i.e., not x86/x64) */
169static int hv_setup_stimer0_irq(void)
170{
171 int ret;
172
173 ret = acpi_register_gsi(NULL, HYPERV_STIMER0_VECTOR,
174 ACPI_EDGE_SENSITIVE, ACPI_ACTIVE_HIGH);
175 if (ret < 0) {
176 pr_err("Can't register Hyper-V stimer0 GSI. Error %d", ret);
177 return ret;
178 }
179 stimer0_irq = ret;
180
181 ret = request_percpu_irq(stimer0_irq, hv_stimer0_percpu_isr,
182 "Hyper-V stimer0", &stimer0_evt);
183 if (ret) {
184 pr_err("Can't request Hyper-V stimer0 IRQ %d. Error %d",
185 stimer0_irq, ret);
186 acpi_unregister_gsi(stimer0_irq);
187 stimer0_irq = -1;
188 }
189 return ret;
190}
191
192static void hv_remove_stimer0_irq(void)
193{
194 if (stimer0_irq == -1) {
195 hv_remove_stimer0_handler();
196 } else {
197 free_percpu_irq(stimer0_irq, &stimer0_evt);
198 acpi_unregister_gsi(stimer0_irq);
199 stimer0_irq = -1;
200 }
201}
202#else
203static int hv_setup_stimer0_irq(void)
204{
205 return 0;
206}
207
208static void hv_remove_stimer0_irq(void)
209{
210}
211#endif
212
213/* hv_stimer_alloc - Global initialization of the clockevent and stimer0 */
214int hv_stimer_alloc(bool have_percpu_irqs)
215{
216 int ret;
217
218 /*
219 * Synthetic timers are always available except on old versions of
220 * Hyper-V on x86. In that case, return as error as Linux will use a
221 * clockevent based on emulated LAPIC timer hardware.
222 */
223 if (!(ms_hyperv.features & HV_MSR_SYNTIMER_AVAILABLE) ||
224 !(ms_hyperv.misc_features & HV_STIMER_DIRECT_MODE_AVAILABLE))
225 return -EINVAL;
226
227 hv_clock_event = alloc_percpu(struct clock_event_device);
228 if (!hv_clock_event)
229 return -ENOMEM;
230
231 if (have_percpu_irqs) {
232 ret = hv_setup_stimer0_irq();
233 if (ret)
234 goto free_clock_event;
235 } else {
236 hv_setup_stimer0_handler(hv_stimer0_isr);
237 }
238
239 ret = cpuhp_setup_state(CPUHP_AP_HYPERV_TIMER_STARTING,
240 "clockevents/hyperv/stimer:starting",
241 hv_stimer_init, hv_stimer_cleanup);
242 if (ret < 0) {
243 hv_remove_stimer0_irq();
244 goto free_clock_event;
245 }
246 return ret;
247
248free_clock_event:
249 free_percpu(hv_clock_event);
250 hv_clock_event = NULL;
251 return ret;
252}
253EXPORT_SYMBOL_GPL(hv_stimer_alloc);
254
255/*
256 * Do a global cleanup of clockevents for the cases of kexec and
257 * vmbus exit
258 */
259void hv_stimer_global_cleanup(void)
260{
261 if (!hv_clock_event)
262 return;
263
264 cpuhp_remove_state(CPUHP_AP_HYPERV_TIMER_STARTING);
265 hv_remove_stimer0_irq();
266 stimer0_irq = -1;
267
268 free_percpu(hv_clock_event);
269 hv_clock_event = NULL;
270
271}
272EXPORT_SYMBOL_GPL(hv_stimer_global_cleanup);
273
274static __always_inline u64 read_hv_clock_msr(void)
275{
276 /*
277 * Read the partition counter to get the current tick count. This count
278 * is set to 0 when the partition is created and is incremented in 100
279 * nanosecond units.
280 *
281 * Use hv_raw_get_msr() because this function is used from
282 * noinstr. Notable; while HV_MSR_TIME_REF_COUNT is a synthetic
283 * register it doesn't need the GHCB path.
284 */
285 return hv_raw_get_msr(HV_MSR_TIME_REF_COUNT);
286}
287
288/*
289 * Code and definitions for the Hyper-V clocksources. Two
290 * clocksources are defined: one that reads the Hyper-V defined MSR, and
291 * the other that uses the TSC reference page feature as defined in the
292 * TLFS. The MSR version is for compatibility with old versions of
293 * Hyper-V and 32-bit x86. The TSC reference page version is preferred.
294 */
295
296static union {
297 struct ms_hyperv_tsc_page page;
298 u8 reserved[PAGE_SIZE];
299} tsc_pg __bss_decrypted __aligned(PAGE_SIZE);
300
301static struct ms_hyperv_tsc_page *tsc_page = &tsc_pg.page;
302static unsigned long tsc_pfn;
303
304unsigned long hv_get_tsc_pfn(void)
305{
306 return tsc_pfn;
307}
308EXPORT_SYMBOL_GPL(hv_get_tsc_pfn);
309
310struct ms_hyperv_tsc_page *hv_get_tsc_page(void)
311{
312 return tsc_page;
313}
314EXPORT_SYMBOL_GPL(hv_get_tsc_page);
315
316static __always_inline u64 read_hv_clock_tsc(void)
317{
318 u64 cur_tsc, time;
319
320 /*
321 * The Hyper-V Top-Level Function Spec (TLFS), section Timers,
322 * subsection Refererence Counter, guarantees that the TSC and MSR
323 * times are in sync and monotonic. Therefore we can fall back
324 * to the MSR in case the TSC page indicates unavailability.
325 */
326 if (!hv_read_tsc_page_tsc(tsc_page, &cur_tsc, &time))
327 time = read_hv_clock_msr();
328
329 return time;
330}
331
332static u64 notrace read_hv_clock_tsc_cs(struct clocksource *arg)
333{
334 return read_hv_clock_tsc();
335}
336
337static u64 notrace read_hv_clock_tsc_cs_snapshot(struct clocksource *arg,
338 struct clocksource_hw_snapshot *chs)
339{
340 u64 time;
341
342 if (hv_read_tsc_page_tsc(tsc_page, &chs->hw_cycles, &time)) {
343 chs->hw_csid = CSID_X86_TSC;
344 } else {
345 chs->hw_cycles = 0;
346 chs->hw_csid = CSID_GENERIC;
347 time = read_hv_clock_msr();
348 }
349
350 return time;
351}
352
353static u64 noinstr read_hv_sched_clock_tsc(void)
354{
355 return (read_hv_clock_tsc() - hv_sched_clock_offset) *
356 (NSEC_PER_SEC / HV_CLOCK_HZ);
357}
358
359static void suspend_hv_clock_tsc(struct clocksource *arg)
360{
361 union hv_reference_tsc_msr tsc_msr;
362
363 /* Disable the TSC page */
364 tsc_msr.as_uint64 = hv_get_msr(HV_MSR_REFERENCE_TSC);
365 tsc_msr.enable = 0;
366 hv_set_msr(HV_MSR_REFERENCE_TSC, tsc_msr.as_uint64);
367}
368
369
370static void resume_hv_clock_tsc(struct clocksource *arg)
371{
372 union hv_reference_tsc_msr tsc_msr;
373
374 /* Re-enable the TSC page */
375 tsc_msr.as_uint64 = hv_get_msr(HV_MSR_REFERENCE_TSC);
376 tsc_msr.enable = 1;
377 tsc_msr.pfn = tsc_pfn;
378 hv_set_msr(HV_MSR_REFERENCE_TSC, tsc_msr.as_uint64);
379}
380
381/*
382 * Called during resume from hibernation, from overridden
383 * x86_platform.restore_sched_clock_state routine. This is to adjust offsets
384 * used to calculate time for hv tsc page based sched_clock, to account for
385 * time spent before hibernation.
386 */
387void hv_adj_sched_clock_offset(u64 offset)
388{
389 hv_sched_clock_offset -= offset;
390}
391
392#ifdef HAVE_VDSO_CLOCKMODE_HVCLOCK
393static int hv_cs_enable(struct clocksource *cs)
394{
395 vclocks_set_used(VDSO_CLOCKMODE_HVCLOCK);
396 return 0;
397}
398#endif
399
400static struct clocksource hyperv_cs_tsc = {
401 .name = "hyperv_clocksource_tsc_page",
402 .rating = 500,
403 .read = read_hv_clock_tsc_cs,
404 .read_snapshot = read_hv_clock_tsc_cs_snapshot,
405 .mask = CLOCKSOURCE_MASK(64),
406 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
407 .suspend = suspend_hv_clock_tsc,
408 .resume = resume_hv_clock_tsc,
409#ifdef HAVE_VDSO_CLOCKMODE_HVCLOCK
410 .enable = hv_cs_enable,
411 .vdso_clock_mode = VDSO_CLOCKMODE_HVCLOCK,
412#else
413 .vdso_clock_mode = VDSO_CLOCKMODE_NONE,
414#endif
415};
416
417static u64 notrace read_hv_clock_msr_cs(struct clocksource *arg)
418{
419 return read_hv_clock_msr();
420}
421
422static struct clocksource hyperv_cs_msr = {
423 .name = "hyperv_clocksource_msr",
424 .rating = 495,
425 .read = read_hv_clock_msr_cs,
426 .mask = CLOCKSOURCE_MASK(64),
427 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
428};
429
430/*
431 * Reference to pv_ops must be inline so objtool
432 * detection of noinstr violations can work correctly.
433 */
434#ifdef CONFIG_GENERIC_SCHED_CLOCK
435static __always_inline void hv_setup_sched_clock(void *sched_clock)
436{
437 /*
438 * We're on an architecture with generic sched clock (not x86/x64).
439 * The Hyper-V sched clock read function returns nanoseconds, not
440 * the normal 100ns units of the Hyper-V synthetic clock.
441 */
442 sched_clock_register(sched_clock, 64, NSEC_PER_SEC);
443}
444#elif defined CONFIG_PARAVIRT
445#include <asm/timer.h>
446
447static __always_inline void hv_setup_sched_clock(void *sched_clock)
448{
449 /* We're on x86/x64 *and* using PV ops */
450 paravirt_set_sched_clock(sched_clock);
451}
452#else /* !CONFIG_GENERIC_SCHED_CLOCK && !CONFIG_PARAVIRT */
453static __always_inline void hv_setup_sched_clock(void *sched_clock) {}
454#endif /* CONFIG_GENERIC_SCHED_CLOCK */
455
456static void __init hv_init_tsc_clocksource(void)
457{
458 union hv_reference_tsc_msr tsc_msr;
459
460 /*
461 * When running as a guest partition:
462 *
463 * If Hyper-V offers TSC_INVARIANT, then the virtualized TSC correctly
464 * handles frequency and offset changes due to live migration,
465 * pause/resume, and other VM management operations. So lower the
466 * Hyper-V Reference TSC rating, causing the generic TSC to be used.
467 * TSC_INVARIANT is not offered on ARM64, so the Hyper-V Reference
468 * TSC will be preferred over the virtualized ARM64 arch counter.
469 *
470 * When running as the root partition:
471 *
472 * There is no HV_ACCESS_TSC_INVARIANT feature. Always lower the rating
473 * of the Hyper-V Reference TSC.
474 */
475 if ((ms_hyperv.features & HV_ACCESS_TSC_INVARIANT) ||
476 hv_root_partition()) {
477 hyperv_cs_tsc.rating = 250;
478 hyperv_cs_msr.rating = 245;
479 }
480
481 if (!(ms_hyperv.features & HV_MSR_REFERENCE_TSC_AVAILABLE))
482 return;
483
484 hv_read_reference_counter = read_hv_clock_tsc;
485
486 /*
487 * TSC page mapping works differently in root compared to guest.
488 * - In guest partition the guest PFN has to be passed to the
489 * hypervisor.
490 * - In root partition it's other way around: it has to map the PFN
491 * provided by the hypervisor.
492 * But it can't be mapped right here as it's too early and MMU isn't
493 * ready yet. So, we only set the enable bit here and will remap the
494 * page later in hv_remap_tsc_clocksource().
495 *
496 * It worth mentioning, that TSC clocksource read function
497 * (read_hv_clock_tsc) has a MSR-based fallback mechanism, used when
498 * TSC page is zeroed (which is the case until the PFN is remapped) and
499 * thus TSC clocksource will work even without the real TSC page
500 * mapped.
501 */
502 tsc_msr.as_uint64 = hv_get_msr(HV_MSR_REFERENCE_TSC);
503 if (hv_root_partition())
504 tsc_pfn = tsc_msr.pfn;
505 else
506 tsc_pfn = HVPFN_DOWN(virt_to_phys(tsc_page));
507 tsc_msr.enable = 1;
508 tsc_msr.pfn = tsc_pfn;
509 hv_set_msr(HV_MSR_REFERENCE_TSC, tsc_msr.as_uint64);
510
511 clocksource_register_hz(&hyperv_cs_tsc, NSEC_PER_SEC/100);
512
513 /*
514 * If TSC is invariant, then let it stay as the sched clock since it
515 * will be faster than reading the TSC page. But if not invariant, use
516 * the TSC page so that live migrations across hosts with different
517 * frequencies is handled correctly.
518 */
519 if (!(ms_hyperv.features & HV_ACCESS_TSC_INVARIANT)) {
520 hv_sched_clock_offset = hv_read_reference_counter();
521 hv_setup_sched_clock(read_hv_sched_clock_tsc);
522 }
523}
524
525void __init hv_init_clocksource(void)
526{
527 /*
528 * Try to set up the TSC page clocksource, then the MSR clocksource.
529 * At least one of these will always be available except on very old
530 * versions of Hyper-V on x86. In that case we won't have a Hyper-V
531 * clocksource, but Linux will still run with a clocksource based
532 * on the emulated PIT or LAPIC timer.
533 *
534 * Never use the MSR clocksource as sched clock. It's too slow.
535 * Better to use the native sched clock as the fallback.
536 */
537 hv_init_tsc_clocksource();
538
539 if (ms_hyperv.features & HV_MSR_TIME_REF_COUNT_AVAILABLE)
540 clocksource_register_hz(&hyperv_cs_msr, NSEC_PER_SEC/100);
541}
542
543void __init hv_remap_tsc_clocksource(void)
544{
545 if (!(ms_hyperv.features & HV_MSR_REFERENCE_TSC_AVAILABLE))
546 return;
547
548 if (!hv_root_partition()) {
549 WARN(1, "%s: attempt to remap TSC page in guest partition\n",
550 __func__);
551 return;
552 }
553
554 tsc_page = memremap(tsc_pfn << HV_HYP_PAGE_SHIFT, sizeof(tsc_pg),
555 MEMREMAP_WB);
556 if (!tsc_page)
557 pr_err("Failed to remap Hyper-V TSC page.\n");
558}