xref: /linux/drivers/clocksource/arm_arch_timer.c (revision 981368e1440b76f68b1ac8f5fb14e739f80ecc4e)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  *  linux/drivers/clocksource/arm_arch_timer.c
4  *
5  *  Copyright (C) 2011 ARM Ltd.
6  *  All Rights Reserved
7  */
8 
9 #define pr_fmt(fmt) 	"arch_timer: " fmt
10 
11 #include <linux/init.h>
12 #include <linux/kernel.h>
13 #include <linux/device.h>
14 #include <linux/smp.h>
15 #include <linux/cpu.h>
16 #include <linux/cpu_pm.h>
17 #include <linux/clockchips.h>
18 #include <linux/clocksource.h>
19 #include <linux/clocksource_ids.h>
20 #include <linux/interrupt.h>
21 #include <linux/kstrtox.h>
22 #include <linux/of_irq.h>
23 #include <linux/of_address.h>
24 #include <linux/io.h>
25 #include <linux/slab.h>
26 #include <linux/sched/clock.h>
27 #include <linux/sched_clock.h>
28 #include <linux/acpi.h>
29 #include <linux/arm-smccc.h>
30 #include <linux/ptp_kvm.h>
31 
32 #include <asm/arch_timer.h>
33 #include <asm/virt.h>
34 
35 #include <clocksource/arm_arch_timer.h>
36 
37 #define CNTTIDR		0x08
38 #define CNTTIDR_VIRT(n)	(BIT(1) << ((n) * 4))
39 
40 #define CNTACR(n)	(0x40 + ((n) * 4))
41 #define CNTACR_RPCT	BIT(0)
42 #define CNTACR_RVCT	BIT(1)
43 #define CNTACR_RFRQ	BIT(2)
44 #define CNTACR_RVOFF	BIT(3)
45 #define CNTACR_RWVT	BIT(4)
46 #define CNTACR_RWPT	BIT(5)
47 
48 #define CNTPCT_LO	0x00
49 #define CNTVCT_LO	0x08
50 #define CNTFRQ		0x10
51 #define CNTP_CVAL_LO	0x20
52 #define CNTP_CTL	0x2c
53 #define CNTV_CVAL_LO	0x30
54 #define CNTV_CTL	0x3c
55 
56 /*
57  * The minimum amount of time a generic counter is guaranteed to not roll over
58  * (40 years)
59  */
60 #define MIN_ROLLOVER_SECS	(40ULL * 365 * 24 * 3600)
61 
62 static unsigned arch_timers_present __initdata;
63 
64 struct arch_timer {
65 	void __iomem *base;
66 	struct clock_event_device evt;
67 };
68 
69 static struct arch_timer *arch_timer_mem __ro_after_init;
70 
71 #define to_arch_timer(e) container_of(e, struct arch_timer, evt)
72 
73 static u32 arch_timer_rate __ro_after_init;
74 static int arch_timer_ppi[ARCH_TIMER_MAX_TIMER_PPI] __ro_after_init;
75 
76 static const char *arch_timer_ppi_names[ARCH_TIMER_MAX_TIMER_PPI] = {
77 	[ARCH_TIMER_PHYS_SECURE_PPI]	= "sec-phys",
78 	[ARCH_TIMER_PHYS_NONSECURE_PPI]	= "phys",
79 	[ARCH_TIMER_VIRT_PPI]		= "virt",
80 	[ARCH_TIMER_HYP_PPI]		= "hyp-phys",
81 	[ARCH_TIMER_HYP_VIRT_PPI]	= "hyp-virt",
82 };
83 
84 static struct clock_event_device __percpu *arch_timer_evt;
85 
86 static enum arch_timer_ppi_nr arch_timer_uses_ppi __ro_after_init = ARCH_TIMER_VIRT_PPI;
87 static bool arch_timer_c3stop __ro_after_init;
88 static bool arch_timer_mem_use_virtual __ro_after_init;
89 static bool arch_counter_suspend_stop __ro_after_init;
90 #ifdef CONFIG_GENERIC_GETTIMEOFDAY
91 static enum vdso_clock_mode vdso_default = VDSO_CLOCKMODE_ARCHTIMER;
92 #else
93 static enum vdso_clock_mode vdso_default = VDSO_CLOCKMODE_NONE;
94 #endif /* CONFIG_GENERIC_GETTIMEOFDAY */
95 
96 static cpumask_t evtstrm_available = CPU_MASK_NONE;
97 static bool evtstrm_enable __ro_after_init = IS_ENABLED(CONFIG_ARM_ARCH_TIMER_EVTSTREAM);
98 
99 static int __init early_evtstrm_cfg(char *buf)
100 {
101 	return kstrtobool(buf, &evtstrm_enable);
102 }
103 early_param("clocksource.arm_arch_timer.evtstrm", early_evtstrm_cfg);
104 
105 /*
106  * Makes an educated guess at a valid counter width based on the Generic Timer
107  * specification. Of note:
108  *   1) the system counter is at least 56 bits wide
109  *   2) a roll-over time of not less than 40 years
110  *
111  * See 'ARM DDI 0487G.a D11.1.2 ("The system counter")' for more details.
112  */
113 static int arch_counter_get_width(void)
114 {
115 	u64 min_cycles = MIN_ROLLOVER_SECS * arch_timer_rate;
116 
117 	/* guarantee the returned width is within the valid range */
118 	return clamp_val(ilog2(min_cycles - 1) + 1, 56, 64);
119 }
120 
121 /*
122  * Architected system timer support.
123  */
124 
125 static __always_inline
126 void arch_timer_reg_write(int access, enum arch_timer_reg reg, u64 val,
127 			  struct clock_event_device *clk)
128 {
129 	if (access == ARCH_TIMER_MEM_PHYS_ACCESS) {
130 		struct arch_timer *timer = to_arch_timer(clk);
131 		switch (reg) {
132 		case ARCH_TIMER_REG_CTRL:
133 			writel_relaxed((u32)val, timer->base + CNTP_CTL);
134 			break;
135 		case ARCH_TIMER_REG_CVAL:
136 			/*
137 			 * Not guaranteed to be atomic, so the timer
138 			 * must be disabled at this point.
139 			 */
140 			writeq_relaxed(val, timer->base + CNTP_CVAL_LO);
141 			break;
142 		default:
143 			BUILD_BUG();
144 		}
145 	} else if (access == ARCH_TIMER_MEM_VIRT_ACCESS) {
146 		struct arch_timer *timer = to_arch_timer(clk);
147 		switch (reg) {
148 		case ARCH_TIMER_REG_CTRL:
149 			writel_relaxed((u32)val, timer->base + CNTV_CTL);
150 			break;
151 		case ARCH_TIMER_REG_CVAL:
152 			/* Same restriction as above */
153 			writeq_relaxed(val, timer->base + CNTV_CVAL_LO);
154 			break;
155 		default:
156 			BUILD_BUG();
157 		}
158 	} else {
159 		arch_timer_reg_write_cp15(access, reg, val);
160 	}
161 }
162 
163 static __always_inline
164 u32 arch_timer_reg_read(int access, enum arch_timer_reg reg,
165 			struct clock_event_device *clk)
166 {
167 	u32 val;
168 
169 	if (access == ARCH_TIMER_MEM_PHYS_ACCESS) {
170 		struct arch_timer *timer = to_arch_timer(clk);
171 		switch (reg) {
172 		case ARCH_TIMER_REG_CTRL:
173 			val = readl_relaxed(timer->base + CNTP_CTL);
174 			break;
175 		default:
176 			BUILD_BUG();
177 		}
178 	} else if (access == ARCH_TIMER_MEM_VIRT_ACCESS) {
179 		struct arch_timer *timer = to_arch_timer(clk);
180 		switch (reg) {
181 		case ARCH_TIMER_REG_CTRL:
182 			val = readl_relaxed(timer->base + CNTV_CTL);
183 			break;
184 		default:
185 			BUILD_BUG();
186 		}
187 	} else {
188 		val = arch_timer_reg_read_cp15(access, reg);
189 	}
190 
191 	return val;
192 }
193 
194 static noinstr u64 raw_counter_get_cntpct_stable(void)
195 {
196 	return __arch_counter_get_cntpct_stable();
197 }
198 
199 static notrace u64 arch_counter_get_cntpct_stable(void)
200 {
201 	u64 val;
202 	preempt_disable_notrace();
203 	val = __arch_counter_get_cntpct_stable();
204 	preempt_enable_notrace();
205 	return val;
206 }
207 
208 static noinstr u64 arch_counter_get_cntpct(void)
209 {
210 	return __arch_counter_get_cntpct();
211 }
212 
213 static noinstr u64 raw_counter_get_cntvct_stable(void)
214 {
215 	return __arch_counter_get_cntvct_stable();
216 }
217 
218 static notrace u64 arch_counter_get_cntvct_stable(void)
219 {
220 	u64 val;
221 	preempt_disable_notrace();
222 	val = __arch_counter_get_cntvct_stable();
223 	preempt_enable_notrace();
224 	return val;
225 }
226 
227 static noinstr u64 arch_counter_get_cntvct(void)
228 {
229 	return __arch_counter_get_cntvct();
230 }
231 
232 /*
233  * Default to cp15 based access because arm64 uses this function for
234  * sched_clock() before DT is probed and the cp15 method is guaranteed
235  * to exist on arm64. arm doesn't use this before DT is probed so even
236  * if we don't have the cp15 accessors we won't have a problem.
237  */
238 u64 (*arch_timer_read_counter)(void) __ro_after_init = arch_counter_get_cntvct;
239 EXPORT_SYMBOL_GPL(arch_timer_read_counter);
240 
241 static u64 arch_counter_read(struct clocksource *cs)
242 {
243 	return arch_timer_read_counter();
244 }
245 
246 static u64 arch_counter_read_cc(const struct cyclecounter *cc)
247 {
248 	return arch_timer_read_counter();
249 }
250 
251 static struct clocksource clocksource_counter = {
252 	.name	= "arch_sys_counter",
253 	.id	= CSID_ARM_ARCH_COUNTER,
254 	.rating	= 400,
255 	.read	= arch_counter_read,
256 	.flags	= CLOCK_SOURCE_IS_CONTINUOUS,
257 };
258 
259 static struct cyclecounter cyclecounter __ro_after_init = {
260 	.read	= arch_counter_read_cc,
261 };
262 
263 struct ate_acpi_oem_info {
264 	char oem_id[ACPI_OEM_ID_SIZE + 1];
265 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE + 1];
266 	u32 oem_revision;
267 };
268 
269 #ifdef CONFIG_FSL_ERRATUM_A008585
270 /*
271  * The number of retries is an arbitrary value well beyond the highest number
272  * of iterations the loop has been observed to take.
273  */
274 #define __fsl_a008585_read_reg(reg) ({			\
275 	u64 _old, _new;					\
276 	int _retries = 200;				\
277 							\
278 	do {						\
279 		_old = read_sysreg(reg);		\
280 		_new = read_sysreg(reg);		\
281 		_retries--;				\
282 	} while (unlikely(_old != _new) && _retries);	\
283 							\
284 	WARN_ON_ONCE(!_retries);			\
285 	_new;						\
286 })
287 
288 static u64 notrace fsl_a008585_read_cntpct_el0(void)
289 {
290 	return __fsl_a008585_read_reg(cntpct_el0);
291 }
292 
293 static u64 notrace fsl_a008585_read_cntvct_el0(void)
294 {
295 	return __fsl_a008585_read_reg(cntvct_el0);
296 }
297 #endif
298 
299 #ifdef CONFIG_HISILICON_ERRATUM_161010101
300 /*
301  * Verify whether the value of the second read is larger than the first by
302  * less than 32 is the only way to confirm the value is correct, so clear the
303  * lower 5 bits to check whether the difference is greater than 32 or not.
304  * Theoretically the erratum should not occur more than twice in succession
305  * when reading the system counter, but it is possible that some interrupts
306  * may lead to more than twice read errors, triggering the warning, so setting
307  * the number of retries far beyond the number of iterations the loop has been
308  * observed to take.
309  */
310 #define __hisi_161010101_read_reg(reg) ({				\
311 	u64 _old, _new;						\
312 	int _retries = 50;					\
313 								\
314 	do {							\
315 		_old = read_sysreg(reg);			\
316 		_new = read_sysreg(reg);			\
317 		_retries--;					\
318 	} while (unlikely((_new - _old) >> 5) && _retries);	\
319 								\
320 	WARN_ON_ONCE(!_retries);				\
321 	_new;							\
322 })
323 
324 static u64 notrace hisi_161010101_read_cntpct_el0(void)
325 {
326 	return __hisi_161010101_read_reg(cntpct_el0);
327 }
328 
329 static u64 notrace hisi_161010101_read_cntvct_el0(void)
330 {
331 	return __hisi_161010101_read_reg(cntvct_el0);
332 }
333 
334 static struct ate_acpi_oem_info hisi_161010101_oem_info[] = {
335 	/*
336 	 * Note that trailing spaces are required to properly match
337 	 * the OEM table information.
338 	 */
339 	{
340 		.oem_id		= "HISI  ",
341 		.oem_table_id	= "HIP05   ",
342 		.oem_revision	= 0,
343 	},
344 	{
345 		.oem_id		= "HISI  ",
346 		.oem_table_id	= "HIP06   ",
347 		.oem_revision	= 0,
348 	},
349 	{
350 		.oem_id		= "HISI  ",
351 		.oem_table_id	= "HIP07   ",
352 		.oem_revision	= 0,
353 	},
354 	{ /* Sentinel indicating the end of the OEM array */ },
355 };
356 #endif
357 
358 #ifdef CONFIG_ARM64_ERRATUM_858921
359 static u64 notrace arm64_858921_read_cntpct_el0(void)
360 {
361 	u64 old, new;
362 
363 	old = read_sysreg(cntpct_el0);
364 	new = read_sysreg(cntpct_el0);
365 	return (((old ^ new) >> 32) & 1) ? old : new;
366 }
367 
368 static u64 notrace arm64_858921_read_cntvct_el0(void)
369 {
370 	u64 old, new;
371 
372 	old = read_sysreg(cntvct_el0);
373 	new = read_sysreg(cntvct_el0);
374 	return (((old ^ new) >> 32) & 1) ? old : new;
375 }
376 #endif
377 
378 #ifdef CONFIG_SUN50I_ERRATUM_UNKNOWN1
379 /*
380  * The low bits of the counter registers are indeterminate while bit 10 or
381  * greater is rolling over. Since the counter value can jump both backward
382  * (7ff -> 000 -> 800) and forward (7ff -> fff -> 800), ignore register values
383  * with all ones or all zeros in the low bits. Bound the loop by the maximum
384  * number of CPU cycles in 3 consecutive 24 MHz counter periods.
385  */
386 #define __sun50i_a64_read_reg(reg) ({					\
387 	u64 _val;							\
388 	int _retries = 150;						\
389 									\
390 	do {								\
391 		_val = read_sysreg(reg);				\
392 		_retries--;						\
393 	} while (((_val + 1) & GENMASK(8, 0)) <= 1 && _retries);	\
394 									\
395 	WARN_ON_ONCE(!_retries);					\
396 	_val;								\
397 })
398 
399 static u64 notrace sun50i_a64_read_cntpct_el0(void)
400 {
401 	return __sun50i_a64_read_reg(cntpct_el0);
402 }
403 
404 static u64 notrace sun50i_a64_read_cntvct_el0(void)
405 {
406 	return __sun50i_a64_read_reg(cntvct_el0);
407 }
408 #endif
409 
410 #ifdef CONFIG_ARM_ARCH_TIMER_OOL_WORKAROUND
411 DEFINE_PER_CPU(const struct arch_timer_erratum_workaround *, timer_unstable_counter_workaround);
412 EXPORT_SYMBOL_GPL(timer_unstable_counter_workaround);
413 
414 static atomic_t timer_unstable_counter_workaround_in_use = ATOMIC_INIT(0);
415 
416 /*
417  * Force the inlining of this function so that the register accesses
418  * can be themselves correctly inlined.
419  */
420 static __always_inline
421 void erratum_set_next_event_generic(const int access, unsigned long evt,
422 				    struct clock_event_device *clk)
423 {
424 	unsigned long ctrl;
425 	u64 cval;
426 
427 	ctrl = arch_timer_reg_read(access, ARCH_TIMER_REG_CTRL, clk);
428 	ctrl |= ARCH_TIMER_CTRL_ENABLE;
429 	ctrl &= ~ARCH_TIMER_CTRL_IT_MASK;
430 
431 	if (access == ARCH_TIMER_PHYS_ACCESS) {
432 		cval = evt + arch_counter_get_cntpct_stable();
433 		write_sysreg(cval, cntp_cval_el0);
434 	} else {
435 		cval = evt + arch_counter_get_cntvct_stable();
436 		write_sysreg(cval, cntv_cval_el0);
437 	}
438 
439 	arch_timer_reg_write(access, ARCH_TIMER_REG_CTRL, ctrl, clk);
440 }
441 
442 static __maybe_unused int erratum_set_next_event_virt(unsigned long evt,
443 					    struct clock_event_device *clk)
444 {
445 	erratum_set_next_event_generic(ARCH_TIMER_VIRT_ACCESS, evt, clk);
446 	return 0;
447 }
448 
449 static __maybe_unused int erratum_set_next_event_phys(unsigned long evt,
450 					    struct clock_event_device *clk)
451 {
452 	erratum_set_next_event_generic(ARCH_TIMER_PHYS_ACCESS, evt, clk);
453 	return 0;
454 }
455 
456 static const struct arch_timer_erratum_workaround ool_workarounds[] = {
457 #ifdef CONFIG_FSL_ERRATUM_A008585
458 	{
459 		.match_type = ate_match_dt,
460 		.id = "fsl,erratum-a008585",
461 		.desc = "Freescale erratum a005858",
462 		.read_cntpct_el0 = fsl_a008585_read_cntpct_el0,
463 		.read_cntvct_el0 = fsl_a008585_read_cntvct_el0,
464 		.set_next_event_phys = erratum_set_next_event_phys,
465 		.set_next_event_virt = erratum_set_next_event_virt,
466 	},
467 #endif
468 #ifdef CONFIG_HISILICON_ERRATUM_161010101
469 	{
470 		.match_type = ate_match_dt,
471 		.id = "hisilicon,erratum-161010101",
472 		.desc = "HiSilicon erratum 161010101",
473 		.read_cntpct_el0 = hisi_161010101_read_cntpct_el0,
474 		.read_cntvct_el0 = hisi_161010101_read_cntvct_el0,
475 		.set_next_event_phys = erratum_set_next_event_phys,
476 		.set_next_event_virt = erratum_set_next_event_virt,
477 	},
478 	{
479 		.match_type = ate_match_acpi_oem_info,
480 		.id = hisi_161010101_oem_info,
481 		.desc = "HiSilicon erratum 161010101",
482 		.read_cntpct_el0 = hisi_161010101_read_cntpct_el0,
483 		.read_cntvct_el0 = hisi_161010101_read_cntvct_el0,
484 		.set_next_event_phys = erratum_set_next_event_phys,
485 		.set_next_event_virt = erratum_set_next_event_virt,
486 	},
487 #endif
488 #ifdef CONFIG_ARM64_ERRATUM_858921
489 	{
490 		.match_type = ate_match_local_cap_id,
491 		.id = (void *)ARM64_WORKAROUND_858921,
492 		.desc = "ARM erratum 858921",
493 		.read_cntpct_el0 = arm64_858921_read_cntpct_el0,
494 		.read_cntvct_el0 = arm64_858921_read_cntvct_el0,
495 		.set_next_event_phys = erratum_set_next_event_phys,
496 		.set_next_event_virt = erratum_set_next_event_virt,
497 	},
498 #endif
499 #ifdef CONFIG_SUN50I_ERRATUM_UNKNOWN1
500 	{
501 		.match_type = ate_match_dt,
502 		.id = "allwinner,erratum-unknown1",
503 		.desc = "Allwinner erratum UNKNOWN1",
504 		.read_cntpct_el0 = sun50i_a64_read_cntpct_el0,
505 		.read_cntvct_el0 = sun50i_a64_read_cntvct_el0,
506 		.set_next_event_phys = erratum_set_next_event_phys,
507 		.set_next_event_virt = erratum_set_next_event_virt,
508 	},
509 #endif
510 #ifdef CONFIG_ARM64_ERRATUM_1418040
511 	{
512 		.match_type = ate_match_local_cap_id,
513 		.id = (void *)ARM64_WORKAROUND_1418040,
514 		.desc = "ARM erratum 1418040",
515 		.disable_compat_vdso = true,
516 	},
517 #endif
518 };
519 
520 typedef bool (*ate_match_fn_t)(const struct arch_timer_erratum_workaround *,
521 			       const void *);
522 
523 static
524 bool arch_timer_check_dt_erratum(const struct arch_timer_erratum_workaround *wa,
525 				 const void *arg)
526 {
527 	const struct device_node *np = arg;
528 
529 	return of_property_read_bool(np, wa->id);
530 }
531 
532 static
533 bool arch_timer_check_local_cap_erratum(const struct arch_timer_erratum_workaround *wa,
534 					const void *arg)
535 {
536 	return this_cpu_has_cap((uintptr_t)wa->id);
537 }
538 
539 
540 static
541 bool arch_timer_check_acpi_oem_erratum(const struct arch_timer_erratum_workaround *wa,
542 				       const void *arg)
543 {
544 	static const struct ate_acpi_oem_info empty_oem_info = {};
545 	const struct ate_acpi_oem_info *info = wa->id;
546 	const struct acpi_table_header *table = arg;
547 
548 	/* Iterate over the ACPI OEM info array, looking for a match */
549 	while (memcmp(info, &empty_oem_info, sizeof(*info))) {
550 		if (!memcmp(info->oem_id, table->oem_id, ACPI_OEM_ID_SIZE) &&
551 		    !memcmp(info->oem_table_id, table->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) &&
552 		    info->oem_revision == table->oem_revision)
553 			return true;
554 
555 		info++;
556 	}
557 
558 	return false;
559 }
560 
561 static const struct arch_timer_erratum_workaround *
562 arch_timer_iterate_errata(enum arch_timer_erratum_match_type type,
563 			  ate_match_fn_t match_fn,
564 			  void *arg)
565 {
566 	int i;
567 
568 	for (i = 0; i < ARRAY_SIZE(ool_workarounds); i++) {
569 		if (ool_workarounds[i].match_type != type)
570 			continue;
571 
572 		if (match_fn(&ool_workarounds[i], arg))
573 			return &ool_workarounds[i];
574 	}
575 
576 	return NULL;
577 }
578 
579 static
580 void arch_timer_enable_workaround(const struct arch_timer_erratum_workaround *wa,
581 				  bool local)
582 {
583 	int i;
584 
585 	if (local) {
586 		__this_cpu_write(timer_unstable_counter_workaround, wa);
587 	} else {
588 		for_each_possible_cpu(i)
589 			per_cpu(timer_unstable_counter_workaround, i) = wa;
590 	}
591 
592 	if (wa->read_cntvct_el0 || wa->read_cntpct_el0)
593 		atomic_set(&timer_unstable_counter_workaround_in_use, 1);
594 
595 	/*
596 	 * Don't use the vdso fastpath if errata require using the
597 	 * out-of-line counter accessor. We may change our mind pretty
598 	 * late in the game (with a per-CPU erratum, for example), so
599 	 * change both the default value and the vdso itself.
600 	 */
601 	if (wa->read_cntvct_el0) {
602 		clocksource_counter.vdso_clock_mode = VDSO_CLOCKMODE_NONE;
603 		vdso_default = VDSO_CLOCKMODE_NONE;
604 	} else if (wa->disable_compat_vdso && vdso_default != VDSO_CLOCKMODE_NONE) {
605 		vdso_default = VDSO_CLOCKMODE_ARCHTIMER_NOCOMPAT;
606 		clocksource_counter.vdso_clock_mode = vdso_default;
607 	}
608 }
609 
610 static void arch_timer_check_ool_workaround(enum arch_timer_erratum_match_type type,
611 					    void *arg)
612 {
613 	const struct arch_timer_erratum_workaround *wa, *__wa;
614 	ate_match_fn_t match_fn = NULL;
615 	bool local = false;
616 
617 	switch (type) {
618 	case ate_match_dt:
619 		match_fn = arch_timer_check_dt_erratum;
620 		break;
621 	case ate_match_local_cap_id:
622 		match_fn = arch_timer_check_local_cap_erratum;
623 		local = true;
624 		break;
625 	case ate_match_acpi_oem_info:
626 		match_fn = arch_timer_check_acpi_oem_erratum;
627 		break;
628 	default:
629 		WARN_ON(1);
630 		return;
631 	}
632 
633 	wa = arch_timer_iterate_errata(type, match_fn, arg);
634 	if (!wa)
635 		return;
636 
637 	__wa = __this_cpu_read(timer_unstable_counter_workaround);
638 	if (__wa && wa != __wa)
639 		pr_warn("Can't enable workaround for %s (clashes with %s\n)",
640 			wa->desc, __wa->desc);
641 
642 	if (__wa)
643 		return;
644 
645 	arch_timer_enable_workaround(wa, local);
646 	pr_info("Enabling %s workaround for %s\n",
647 		local ? "local" : "global", wa->desc);
648 }
649 
650 static bool arch_timer_this_cpu_has_cntvct_wa(void)
651 {
652 	return has_erratum_handler(read_cntvct_el0);
653 }
654 
655 static bool arch_timer_counter_has_wa(void)
656 {
657 	return atomic_read(&timer_unstable_counter_workaround_in_use);
658 }
659 #else
660 #define arch_timer_check_ool_workaround(t,a)		do { } while(0)
661 #define arch_timer_this_cpu_has_cntvct_wa()		({false;})
662 #define arch_timer_counter_has_wa()			({false;})
663 #endif /* CONFIG_ARM_ARCH_TIMER_OOL_WORKAROUND */
664 
665 static __always_inline irqreturn_t timer_handler(const int access,
666 					struct clock_event_device *evt)
667 {
668 	unsigned long ctrl;
669 
670 	ctrl = arch_timer_reg_read(access, ARCH_TIMER_REG_CTRL, evt);
671 	if (ctrl & ARCH_TIMER_CTRL_IT_STAT) {
672 		ctrl |= ARCH_TIMER_CTRL_IT_MASK;
673 		arch_timer_reg_write(access, ARCH_TIMER_REG_CTRL, ctrl, evt);
674 		evt->event_handler(evt);
675 		return IRQ_HANDLED;
676 	}
677 
678 	return IRQ_NONE;
679 }
680 
681 static irqreturn_t arch_timer_handler_virt(int irq, void *dev_id)
682 {
683 	struct clock_event_device *evt = dev_id;
684 
685 	return timer_handler(ARCH_TIMER_VIRT_ACCESS, evt);
686 }
687 
688 static irqreturn_t arch_timer_handler_phys(int irq, void *dev_id)
689 {
690 	struct clock_event_device *evt = dev_id;
691 
692 	return timer_handler(ARCH_TIMER_PHYS_ACCESS, evt);
693 }
694 
695 static irqreturn_t arch_timer_handler_phys_mem(int irq, void *dev_id)
696 {
697 	struct clock_event_device *evt = dev_id;
698 
699 	return timer_handler(ARCH_TIMER_MEM_PHYS_ACCESS, evt);
700 }
701 
702 static irqreturn_t arch_timer_handler_virt_mem(int irq, void *dev_id)
703 {
704 	struct clock_event_device *evt = dev_id;
705 
706 	return timer_handler(ARCH_TIMER_MEM_VIRT_ACCESS, evt);
707 }
708 
709 static __always_inline int arch_timer_shutdown(const int access,
710 					       struct clock_event_device *clk)
711 {
712 	unsigned long ctrl;
713 
714 	ctrl = arch_timer_reg_read(access, ARCH_TIMER_REG_CTRL, clk);
715 	ctrl &= ~ARCH_TIMER_CTRL_ENABLE;
716 	arch_timer_reg_write(access, ARCH_TIMER_REG_CTRL, ctrl, clk);
717 
718 	return 0;
719 }
720 
721 static int arch_timer_shutdown_virt(struct clock_event_device *clk)
722 {
723 	return arch_timer_shutdown(ARCH_TIMER_VIRT_ACCESS, clk);
724 }
725 
726 static int arch_timer_shutdown_phys(struct clock_event_device *clk)
727 {
728 	return arch_timer_shutdown(ARCH_TIMER_PHYS_ACCESS, clk);
729 }
730 
731 static int arch_timer_shutdown_virt_mem(struct clock_event_device *clk)
732 {
733 	return arch_timer_shutdown(ARCH_TIMER_MEM_VIRT_ACCESS, clk);
734 }
735 
736 static int arch_timer_shutdown_phys_mem(struct clock_event_device *clk)
737 {
738 	return arch_timer_shutdown(ARCH_TIMER_MEM_PHYS_ACCESS, clk);
739 }
740 
741 static __always_inline void set_next_event(const int access, unsigned long evt,
742 					   struct clock_event_device *clk)
743 {
744 	unsigned long ctrl;
745 	u64 cnt;
746 
747 	ctrl = arch_timer_reg_read(access, ARCH_TIMER_REG_CTRL, clk);
748 	ctrl |= ARCH_TIMER_CTRL_ENABLE;
749 	ctrl &= ~ARCH_TIMER_CTRL_IT_MASK;
750 
751 	if (access == ARCH_TIMER_PHYS_ACCESS)
752 		cnt = __arch_counter_get_cntpct();
753 	else
754 		cnt = __arch_counter_get_cntvct();
755 
756 	arch_timer_reg_write(access, ARCH_TIMER_REG_CVAL, evt + cnt, clk);
757 	arch_timer_reg_write(access, ARCH_TIMER_REG_CTRL, ctrl, clk);
758 }
759 
760 static int arch_timer_set_next_event_virt(unsigned long evt,
761 					  struct clock_event_device *clk)
762 {
763 	set_next_event(ARCH_TIMER_VIRT_ACCESS, evt, clk);
764 	return 0;
765 }
766 
767 static int arch_timer_set_next_event_phys(unsigned long evt,
768 					  struct clock_event_device *clk)
769 {
770 	set_next_event(ARCH_TIMER_PHYS_ACCESS, evt, clk);
771 	return 0;
772 }
773 
774 static noinstr u64 arch_counter_get_cnt_mem(struct arch_timer *t, int offset_lo)
775 {
776 	u32 cnt_lo, cnt_hi, tmp_hi;
777 
778 	do {
779 		cnt_hi = __le32_to_cpu((__le32 __force)__raw_readl(t->base + offset_lo + 4));
780 		cnt_lo = __le32_to_cpu((__le32 __force)__raw_readl(t->base + offset_lo));
781 		tmp_hi = __le32_to_cpu((__le32 __force)__raw_readl(t->base + offset_lo + 4));
782 	} while (cnt_hi != tmp_hi);
783 
784 	return ((u64) cnt_hi << 32) | cnt_lo;
785 }
786 
787 static __always_inline void set_next_event_mem(const int access, unsigned long evt,
788 					   struct clock_event_device *clk)
789 {
790 	struct arch_timer *timer = to_arch_timer(clk);
791 	unsigned long ctrl;
792 	u64 cnt;
793 
794 	ctrl = arch_timer_reg_read(access, ARCH_TIMER_REG_CTRL, clk);
795 
796 	/* Timer must be disabled before programming CVAL */
797 	if (ctrl & ARCH_TIMER_CTRL_ENABLE) {
798 		ctrl &= ~ARCH_TIMER_CTRL_ENABLE;
799 		arch_timer_reg_write(access, ARCH_TIMER_REG_CTRL, ctrl, clk);
800 	}
801 
802 	ctrl |= ARCH_TIMER_CTRL_ENABLE;
803 	ctrl &= ~ARCH_TIMER_CTRL_IT_MASK;
804 
805 	if (access ==  ARCH_TIMER_MEM_VIRT_ACCESS)
806 		cnt = arch_counter_get_cnt_mem(timer, CNTVCT_LO);
807 	else
808 		cnt = arch_counter_get_cnt_mem(timer, CNTPCT_LO);
809 
810 	arch_timer_reg_write(access, ARCH_TIMER_REG_CVAL, evt + cnt, clk);
811 	arch_timer_reg_write(access, ARCH_TIMER_REG_CTRL, ctrl, clk);
812 }
813 
814 static int arch_timer_set_next_event_virt_mem(unsigned long evt,
815 					      struct clock_event_device *clk)
816 {
817 	set_next_event_mem(ARCH_TIMER_MEM_VIRT_ACCESS, evt, clk);
818 	return 0;
819 }
820 
821 static int arch_timer_set_next_event_phys_mem(unsigned long evt,
822 					      struct clock_event_device *clk)
823 {
824 	set_next_event_mem(ARCH_TIMER_MEM_PHYS_ACCESS, evt, clk);
825 	return 0;
826 }
827 
828 static u64 __arch_timer_check_delta(void)
829 {
830 #ifdef CONFIG_ARM64
831 	const struct midr_range broken_cval_midrs[] = {
832 		/*
833 		 * XGene-1 implements CVAL in terms of TVAL, meaning
834 		 * that the maximum timer range is 32bit. Shame on them.
835 		 *
836 		 * Note that TVAL is signed, thus has only 31 of its
837 		 * 32 bits to express magnitude.
838 		 */
839 		MIDR_ALL_VERSIONS(MIDR_CPU_MODEL(ARM_CPU_IMP_APM,
840 						 APM_CPU_PART_POTENZA)),
841 		{},
842 	};
843 
844 	if (is_midr_in_range_list(read_cpuid_id(), broken_cval_midrs)) {
845 		pr_warn_once("Broken CNTx_CVAL_EL1, using 31 bit TVAL instead.\n");
846 		return CLOCKSOURCE_MASK(31);
847 	}
848 #endif
849 	return CLOCKSOURCE_MASK(arch_counter_get_width());
850 }
851 
852 static void __arch_timer_setup(unsigned type,
853 			       struct clock_event_device *clk)
854 {
855 	u64 max_delta;
856 
857 	clk->features = CLOCK_EVT_FEAT_ONESHOT;
858 
859 	if (type == ARCH_TIMER_TYPE_CP15) {
860 		typeof(clk->set_next_event) sne;
861 
862 		arch_timer_check_ool_workaround(ate_match_local_cap_id, NULL);
863 
864 		if (arch_timer_c3stop)
865 			clk->features |= CLOCK_EVT_FEAT_C3STOP;
866 		clk->name = "arch_sys_timer";
867 		clk->rating = 450;
868 		clk->cpumask = cpumask_of(smp_processor_id());
869 		clk->irq = arch_timer_ppi[arch_timer_uses_ppi];
870 		switch (arch_timer_uses_ppi) {
871 		case ARCH_TIMER_VIRT_PPI:
872 			clk->set_state_shutdown = arch_timer_shutdown_virt;
873 			clk->set_state_oneshot_stopped = arch_timer_shutdown_virt;
874 			sne = erratum_handler(set_next_event_virt);
875 			break;
876 		case ARCH_TIMER_PHYS_SECURE_PPI:
877 		case ARCH_TIMER_PHYS_NONSECURE_PPI:
878 		case ARCH_TIMER_HYP_PPI:
879 			clk->set_state_shutdown = arch_timer_shutdown_phys;
880 			clk->set_state_oneshot_stopped = arch_timer_shutdown_phys;
881 			sne = erratum_handler(set_next_event_phys);
882 			break;
883 		default:
884 			BUG();
885 		}
886 
887 		clk->set_next_event = sne;
888 		max_delta = __arch_timer_check_delta();
889 	} else {
890 		clk->features |= CLOCK_EVT_FEAT_DYNIRQ;
891 		clk->name = "arch_mem_timer";
892 		clk->rating = 400;
893 		clk->cpumask = cpu_possible_mask;
894 		if (arch_timer_mem_use_virtual) {
895 			clk->set_state_shutdown = arch_timer_shutdown_virt_mem;
896 			clk->set_state_oneshot_stopped = arch_timer_shutdown_virt_mem;
897 			clk->set_next_event =
898 				arch_timer_set_next_event_virt_mem;
899 		} else {
900 			clk->set_state_shutdown = arch_timer_shutdown_phys_mem;
901 			clk->set_state_oneshot_stopped = arch_timer_shutdown_phys_mem;
902 			clk->set_next_event =
903 				arch_timer_set_next_event_phys_mem;
904 		}
905 
906 		max_delta = CLOCKSOURCE_MASK(56);
907 	}
908 
909 	clk->set_state_shutdown(clk);
910 
911 	clockevents_config_and_register(clk, arch_timer_rate, 0xf, max_delta);
912 }
913 
914 static void arch_timer_evtstrm_enable(unsigned int divider)
915 {
916 	u32 cntkctl = arch_timer_get_cntkctl();
917 
918 #ifdef CONFIG_ARM64
919 	/* ECV is likely to require a large divider. Use the EVNTIS flag. */
920 	if (cpus_have_const_cap(ARM64_HAS_ECV) && divider > 15) {
921 		cntkctl |= ARCH_TIMER_EVT_INTERVAL_SCALE;
922 		divider -= 8;
923 	}
924 #endif
925 
926 	divider = min(divider, 15U);
927 	cntkctl &= ~ARCH_TIMER_EVT_TRIGGER_MASK;
928 	/* Set the divider and enable virtual event stream */
929 	cntkctl |= (divider << ARCH_TIMER_EVT_TRIGGER_SHIFT)
930 			| ARCH_TIMER_VIRT_EVT_EN;
931 	arch_timer_set_cntkctl(cntkctl);
932 	arch_timer_set_evtstrm_feature();
933 	cpumask_set_cpu(smp_processor_id(), &evtstrm_available);
934 }
935 
936 static void arch_timer_configure_evtstream(void)
937 {
938 	int evt_stream_div, lsb;
939 
940 	/*
941 	 * As the event stream can at most be generated at half the frequency
942 	 * of the counter, use half the frequency when computing the divider.
943 	 */
944 	evt_stream_div = arch_timer_rate / ARCH_TIMER_EVT_STREAM_FREQ / 2;
945 
946 	/*
947 	 * Find the closest power of two to the divisor. If the adjacent bit
948 	 * of lsb (last set bit, starts from 0) is set, then we use (lsb + 1).
949 	 */
950 	lsb = fls(evt_stream_div) - 1;
951 	if (lsb > 0 && (evt_stream_div & BIT(lsb - 1)))
952 		lsb++;
953 
954 	/* enable event stream */
955 	arch_timer_evtstrm_enable(max(0, lsb));
956 }
957 
958 static void arch_counter_set_user_access(void)
959 {
960 	u32 cntkctl = arch_timer_get_cntkctl();
961 
962 	/* Disable user access to the timers and both counters */
963 	/* Also disable virtual event stream */
964 	cntkctl &= ~(ARCH_TIMER_USR_PT_ACCESS_EN
965 			| ARCH_TIMER_USR_VT_ACCESS_EN
966 		        | ARCH_TIMER_USR_VCT_ACCESS_EN
967 			| ARCH_TIMER_VIRT_EVT_EN
968 			| ARCH_TIMER_USR_PCT_ACCESS_EN);
969 
970 	/*
971 	 * Enable user access to the virtual counter if it doesn't
972 	 * need to be workaround. The vdso may have been already
973 	 * disabled though.
974 	 */
975 	if (arch_timer_this_cpu_has_cntvct_wa())
976 		pr_info("CPU%d: Trapping CNTVCT access\n", smp_processor_id());
977 	else
978 		cntkctl |= ARCH_TIMER_USR_VCT_ACCESS_EN;
979 
980 	arch_timer_set_cntkctl(cntkctl);
981 }
982 
983 static bool arch_timer_has_nonsecure_ppi(void)
984 {
985 	return (arch_timer_uses_ppi == ARCH_TIMER_PHYS_SECURE_PPI &&
986 		arch_timer_ppi[ARCH_TIMER_PHYS_NONSECURE_PPI]);
987 }
988 
989 static u32 check_ppi_trigger(int irq)
990 {
991 	u32 flags = irq_get_trigger_type(irq);
992 
993 	if (flags != IRQF_TRIGGER_HIGH && flags != IRQF_TRIGGER_LOW) {
994 		pr_warn("WARNING: Invalid trigger for IRQ%d, assuming level low\n", irq);
995 		pr_warn("WARNING: Please fix your firmware\n");
996 		flags = IRQF_TRIGGER_LOW;
997 	}
998 
999 	return flags;
1000 }
1001 
1002 static int arch_timer_starting_cpu(unsigned int cpu)
1003 {
1004 	struct clock_event_device *clk = this_cpu_ptr(arch_timer_evt);
1005 	u32 flags;
1006 
1007 	__arch_timer_setup(ARCH_TIMER_TYPE_CP15, clk);
1008 
1009 	flags = check_ppi_trigger(arch_timer_ppi[arch_timer_uses_ppi]);
1010 	enable_percpu_irq(arch_timer_ppi[arch_timer_uses_ppi], flags);
1011 
1012 	if (arch_timer_has_nonsecure_ppi()) {
1013 		flags = check_ppi_trigger(arch_timer_ppi[ARCH_TIMER_PHYS_NONSECURE_PPI]);
1014 		enable_percpu_irq(arch_timer_ppi[ARCH_TIMER_PHYS_NONSECURE_PPI],
1015 				  flags);
1016 	}
1017 
1018 	arch_counter_set_user_access();
1019 	if (evtstrm_enable)
1020 		arch_timer_configure_evtstream();
1021 
1022 	return 0;
1023 }
1024 
1025 static int validate_timer_rate(void)
1026 {
1027 	if (!arch_timer_rate)
1028 		return -EINVAL;
1029 
1030 	/* Arch timer frequency < 1MHz can cause trouble */
1031 	WARN_ON(arch_timer_rate < 1000000);
1032 
1033 	return 0;
1034 }
1035 
1036 /*
1037  * For historical reasons, when probing with DT we use whichever (non-zero)
1038  * rate was probed first, and don't verify that others match. If the first node
1039  * probed has a clock-frequency property, this overrides the HW register.
1040  */
1041 static void __init arch_timer_of_configure_rate(u32 rate, struct device_node *np)
1042 {
1043 	/* Who has more than one independent system counter? */
1044 	if (arch_timer_rate)
1045 		return;
1046 
1047 	if (of_property_read_u32(np, "clock-frequency", &arch_timer_rate))
1048 		arch_timer_rate = rate;
1049 
1050 	/* Check the timer frequency. */
1051 	if (validate_timer_rate())
1052 		pr_warn("frequency not available\n");
1053 }
1054 
1055 static void __init arch_timer_banner(unsigned type)
1056 {
1057 	pr_info("%s%s%s timer(s) running at %lu.%02luMHz (%s%s%s).\n",
1058 		type & ARCH_TIMER_TYPE_CP15 ? "cp15" : "",
1059 		type == (ARCH_TIMER_TYPE_CP15 | ARCH_TIMER_TYPE_MEM) ?
1060 			" and " : "",
1061 		type & ARCH_TIMER_TYPE_MEM ? "mmio" : "",
1062 		(unsigned long)arch_timer_rate / 1000000,
1063 		(unsigned long)(arch_timer_rate / 10000) % 100,
1064 		type & ARCH_TIMER_TYPE_CP15 ?
1065 			(arch_timer_uses_ppi == ARCH_TIMER_VIRT_PPI) ? "virt" : "phys" :
1066 			"",
1067 		type == (ARCH_TIMER_TYPE_CP15 | ARCH_TIMER_TYPE_MEM) ? "/" : "",
1068 		type & ARCH_TIMER_TYPE_MEM ?
1069 			arch_timer_mem_use_virtual ? "virt" : "phys" :
1070 			"");
1071 }
1072 
1073 u32 arch_timer_get_rate(void)
1074 {
1075 	return arch_timer_rate;
1076 }
1077 
1078 bool arch_timer_evtstrm_available(void)
1079 {
1080 	/*
1081 	 * We might get called from a preemptible context. This is fine
1082 	 * because availability of the event stream should be always the same
1083 	 * for a preemptible context and context where we might resume a task.
1084 	 */
1085 	return cpumask_test_cpu(raw_smp_processor_id(), &evtstrm_available);
1086 }
1087 
1088 static noinstr u64 arch_counter_get_cntvct_mem(void)
1089 {
1090 	return arch_counter_get_cnt_mem(arch_timer_mem, CNTVCT_LO);
1091 }
1092 
1093 static struct arch_timer_kvm_info arch_timer_kvm_info;
1094 
1095 struct arch_timer_kvm_info *arch_timer_get_kvm_info(void)
1096 {
1097 	return &arch_timer_kvm_info;
1098 }
1099 
1100 static void __init arch_counter_register(unsigned type)
1101 {
1102 	u64 (*scr)(void);
1103 	u64 start_count;
1104 	int width;
1105 
1106 	/* Register the CP15 based counter if we have one */
1107 	if (type & ARCH_TIMER_TYPE_CP15) {
1108 		u64 (*rd)(void);
1109 
1110 		if ((IS_ENABLED(CONFIG_ARM64) && !is_hyp_mode_available()) ||
1111 		    arch_timer_uses_ppi == ARCH_TIMER_VIRT_PPI) {
1112 			if (arch_timer_counter_has_wa()) {
1113 				rd = arch_counter_get_cntvct_stable;
1114 				scr = raw_counter_get_cntvct_stable;
1115 			} else {
1116 				rd = arch_counter_get_cntvct;
1117 				scr = arch_counter_get_cntvct;
1118 			}
1119 		} else {
1120 			if (arch_timer_counter_has_wa()) {
1121 				rd = arch_counter_get_cntpct_stable;
1122 				scr = raw_counter_get_cntpct_stable;
1123 			} else {
1124 				rd = arch_counter_get_cntpct;
1125 				scr = arch_counter_get_cntpct;
1126 			}
1127 		}
1128 
1129 		arch_timer_read_counter = rd;
1130 		clocksource_counter.vdso_clock_mode = vdso_default;
1131 	} else {
1132 		arch_timer_read_counter = arch_counter_get_cntvct_mem;
1133 		scr = arch_counter_get_cntvct_mem;
1134 	}
1135 
1136 	width = arch_counter_get_width();
1137 	clocksource_counter.mask = CLOCKSOURCE_MASK(width);
1138 	cyclecounter.mask = CLOCKSOURCE_MASK(width);
1139 
1140 	if (!arch_counter_suspend_stop)
1141 		clocksource_counter.flags |= CLOCK_SOURCE_SUSPEND_NONSTOP;
1142 	start_count = arch_timer_read_counter();
1143 	clocksource_register_hz(&clocksource_counter, arch_timer_rate);
1144 	cyclecounter.mult = clocksource_counter.mult;
1145 	cyclecounter.shift = clocksource_counter.shift;
1146 	timecounter_init(&arch_timer_kvm_info.timecounter,
1147 			 &cyclecounter, start_count);
1148 
1149 	sched_clock_register(scr, width, arch_timer_rate);
1150 }
1151 
1152 static void arch_timer_stop(struct clock_event_device *clk)
1153 {
1154 	pr_debug("disable IRQ%d cpu #%d\n", clk->irq, smp_processor_id());
1155 
1156 	disable_percpu_irq(arch_timer_ppi[arch_timer_uses_ppi]);
1157 	if (arch_timer_has_nonsecure_ppi())
1158 		disable_percpu_irq(arch_timer_ppi[ARCH_TIMER_PHYS_NONSECURE_PPI]);
1159 
1160 	clk->set_state_shutdown(clk);
1161 }
1162 
1163 static int arch_timer_dying_cpu(unsigned int cpu)
1164 {
1165 	struct clock_event_device *clk = this_cpu_ptr(arch_timer_evt);
1166 
1167 	cpumask_clear_cpu(smp_processor_id(), &evtstrm_available);
1168 
1169 	arch_timer_stop(clk);
1170 	return 0;
1171 }
1172 
1173 #ifdef CONFIG_CPU_PM
1174 static DEFINE_PER_CPU(unsigned long, saved_cntkctl);
1175 static int arch_timer_cpu_pm_notify(struct notifier_block *self,
1176 				    unsigned long action, void *hcpu)
1177 {
1178 	if (action == CPU_PM_ENTER) {
1179 		__this_cpu_write(saved_cntkctl, arch_timer_get_cntkctl());
1180 
1181 		cpumask_clear_cpu(smp_processor_id(), &evtstrm_available);
1182 	} else if (action == CPU_PM_ENTER_FAILED || action == CPU_PM_EXIT) {
1183 		arch_timer_set_cntkctl(__this_cpu_read(saved_cntkctl));
1184 
1185 		if (arch_timer_have_evtstrm_feature())
1186 			cpumask_set_cpu(smp_processor_id(), &evtstrm_available);
1187 	}
1188 	return NOTIFY_OK;
1189 }
1190 
1191 static struct notifier_block arch_timer_cpu_pm_notifier = {
1192 	.notifier_call = arch_timer_cpu_pm_notify,
1193 };
1194 
1195 static int __init arch_timer_cpu_pm_init(void)
1196 {
1197 	return cpu_pm_register_notifier(&arch_timer_cpu_pm_notifier);
1198 }
1199 
1200 static void __init arch_timer_cpu_pm_deinit(void)
1201 {
1202 	WARN_ON(cpu_pm_unregister_notifier(&arch_timer_cpu_pm_notifier));
1203 }
1204 
1205 #else
1206 static int __init arch_timer_cpu_pm_init(void)
1207 {
1208 	return 0;
1209 }
1210 
1211 static void __init arch_timer_cpu_pm_deinit(void)
1212 {
1213 }
1214 #endif
1215 
1216 static int __init arch_timer_register(void)
1217 {
1218 	int err;
1219 	int ppi;
1220 
1221 	arch_timer_evt = alloc_percpu(struct clock_event_device);
1222 	if (!arch_timer_evt) {
1223 		err = -ENOMEM;
1224 		goto out;
1225 	}
1226 
1227 	ppi = arch_timer_ppi[arch_timer_uses_ppi];
1228 	switch (arch_timer_uses_ppi) {
1229 	case ARCH_TIMER_VIRT_PPI:
1230 		err = request_percpu_irq(ppi, arch_timer_handler_virt,
1231 					 "arch_timer", arch_timer_evt);
1232 		break;
1233 	case ARCH_TIMER_PHYS_SECURE_PPI:
1234 	case ARCH_TIMER_PHYS_NONSECURE_PPI:
1235 		err = request_percpu_irq(ppi, arch_timer_handler_phys,
1236 					 "arch_timer", arch_timer_evt);
1237 		if (!err && arch_timer_has_nonsecure_ppi()) {
1238 			ppi = arch_timer_ppi[ARCH_TIMER_PHYS_NONSECURE_PPI];
1239 			err = request_percpu_irq(ppi, arch_timer_handler_phys,
1240 						 "arch_timer", arch_timer_evt);
1241 			if (err)
1242 				free_percpu_irq(arch_timer_ppi[ARCH_TIMER_PHYS_SECURE_PPI],
1243 						arch_timer_evt);
1244 		}
1245 		break;
1246 	case ARCH_TIMER_HYP_PPI:
1247 		err = request_percpu_irq(ppi, arch_timer_handler_phys,
1248 					 "arch_timer", arch_timer_evt);
1249 		break;
1250 	default:
1251 		BUG();
1252 	}
1253 
1254 	if (err) {
1255 		pr_err("can't register interrupt %d (%d)\n", ppi, err);
1256 		goto out_free;
1257 	}
1258 
1259 	err = arch_timer_cpu_pm_init();
1260 	if (err)
1261 		goto out_unreg_notify;
1262 
1263 	/* Register and immediately configure the timer on the boot CPU */
1264 	err = cpuhp_setup_state(CPUHP_AP_ARM_ARCH_TIMER_STARTING,
1265 				"clockevents/arm/arch_timer:starting",
1266 				arch_timer_starting_cpu, arch_timer_dying_cpu);
1267 	if (err)
1268 		goto out_unreg_cpupm;
1269 	return 0;
1270 
1271 out_unreg_cpupm:
1272 	arch_timer_cpu_pm_deinit();
1273 
1274 out_unreg_notify:
1275 	free_percpu_irq(arch_timer_ppi[arch_timer_uses_ppi], arch_timer_evt);
1276 	if (arch_timer_has_nonsecure_ppi())
1277 		free_percpu_irq(arch_timer_ppi[ARCH_TIMER_PHYS_NONSECURE_PPI],
1278 				arch_timer_evt);
1279 
1280 out_free:
1281 	free_percpu(arch_timer_evt);
1282 out:
1283 	return err;
1284 }
1285 
1286 static int __init arch_timer_mem_register(void __iomem *base, unsigned int irq)
1287 {
1288 	int ret;
1289 	irq_handler_t func;
1290 
1291 	arch_timer_mem = kzalloc(sizeof(*arch_timer_mem), GFP_KERNEL);
1292 	if (!arch_timer_mem)
1293 		return -ENOMEM;
1294 
1295 	arch_timer_mem->base = base;
1296 	arch_timer_mem->evt.irq = irq;
1297 	__arch_timer_setup(ARCH_TIMER_TYPE_MEM, &arch_timer_mem->evt);
1298 
1299 	if (arch_timer_mem_use_virtual)
1300 		func = arch_timer_handler_virt_mem;
1301 	else
1302 		func = arch_timer_handler_phys_mem;
1303 
1304 	ret = request_irq(irq, func, IRQF_TIMER, "arch_mem_timer", &arch_timer_mem->evt);
1305 	if (ret) {
1306 		pr_err("Failed to request mem timer irq\n");
1307 		kfree(arch_timer_mem);
1308 		arch_timer_mem = NULL;
1309 	}
1310 
1311 	return ret;
1312 }
1313 
1314 static const struct of_device_id arch_timer_of_match[] __initconst = {
1315 	{ .compatible   = "arm,armv7-timer",    },
1316 	{ .compatible   = "arm,armv8-timer",    },
1317 	{},
1318 };
1319 
1320 static const struct of_device_id arch_timer_mem_of_match[] __initconst = {
1321 	{ .compatible   = "arm,armv7-timer-mem", },
1322 	{},
1323 };
1324 
1325 static bool __init arch_timer_needs_of_probing(void)
1326 {
1327 	struct device_node *dn;
1328 	bool needs_probing = false;
1329 	unsigned int mask = ARCH_TIMER_TYPE_CP15 | ARCH_TIMER_TYPE_MEM;
1330 
1331 	/* We have two timers, and both device-tree nodes are probed. */
1332 	if ((arch_timers_present & mask) == mask)
1333 		return false;
1334 
1335 	/*
1336 	 * Only one type of timer is probed,
1337 	 * check if we have another type of timer node in device-tree.
1338 	 */
1339 	if (arch_timers_present & ARCH_TIMER_TYPE_CP15)
1340 		dn = of_find_matching_node(NULL, arch_timer_mem_of_match);
1341 	else
1342 		dn = of_find_matching_node(NULL, arch_timer_of_match);
1343 
1344 	if (dn && of_device_is_available(dn))
1345 		needs_probing = true;
1346 
1347 	of_node_put(dn);
1348 
1349 	return needs_probing;
1350 }
1351 
1352 static int __init arch_timer_common_init(void)
1353 {
1354 	arch_timer_banner(arch_timers_present);
1355 	arch_counter_register(arch_timers_present);
1356 	return arch_timer_arch_init();
1357 }
1358 
1359 /**
1360  * arch_timer_select_ppi() - Select suitable PPI for the current system.
1361  *
1362  * If HYP mode is available, we know that the physical timer
1363  * has been configured to be accessible from PL1. Use it, so
1364  * that a guest can use the virtual timer instead.
1365  *
1366  * On ARMv8.1 with VH extensions, the kernel runs in HYP. VHE
1367  * accesses to CNTP_*_EL1 registers are silently redirected to
1368  * their CNTHP_*_EL2 counterparts, and use a different PPI
1369  * number.
1370  *
1371  * If no interrupt provided for virtual timer, we'll have to
1372  * stick to the physical timer. It'd better be accessible...
1373  * For arm64 we never use the secure interrupt.
1374  *
1375  * Return: a suitable PPI type for the current system.
1376  */
1377 static enum arch_timer_ppi_nr __init arch_timer_select_ppi(void)
1378 {
1379 	if (is_kernel_in_hyp_mode())
1380 		return ARCH_TIMER_HYP_PPI;
1381 
1382 	if (!is_hyp_mode_available() && arch_timer_ppi[ARCH_TIMER_VIRT_PPI])
1383 		return ARCH_TIMER_VIRT_PPI;
1384 
1385 	if (IS_ENABLED(CONFIG_ARM64))
1386 		return ARCH_TIMER_PHYS_NONSECURE_PPI;
1387 
1388 	return ARCH_TIMER_PHYS_SECURE_PPI;
1389 }
1390 
1391 static void __init arch_timer_populate_kvm_info(void)
1392 {
1393 	arch_timer_kvm_info.virtual_irq = arch_timer_ppi[ARCH_TIMER_VIRT_PPI];
1394 	if (is_kernel_in_hyp_mode())
1395 		arch_timer_kvm_info.physical_irq = arch_timer_ppi[ARCH_TIMER_PHYS_NONSECURE_PPI];
1396 }
1397 
1398 static int __init arch_timer_of_init(struct device_node *np)
1399 {
1400 	int i, irq, ret;
1401 	u32 rate;
1402 	bool has_names;
1403 
1404 	if (arch_timers_present & ARCH_TIMER_TYPE_CP15) {
1405 		pr_warn("multiple nodes in dt, skipping\n");
1406 		return 0;
1407 	}
1408 
1409 	arch_timers_present |= ARCH_TIMER_TYPE_CP15;
1410 
1411 	has_names = of_property_read_bool(np, "interrupt-names");
1412 
1413 	for (i = ARCH_TIMER_PHYS_SECURE_PPI; i < ARCH_TIMER_MAX_TIMER_PPI; i++) {
1414 		if (has_names)
1415 			irq = of_irq_get_byname(np, arch_timer_ppi_names[i]);
1416 		else
1417 			irq = of_irq_get(np, i);
1418 		if (irq > 0)
1419 			arch_timer_ppi[i] = irq;
1420 	}
1421 
1422 	arch_timer_populate_kvm_info();
1423 
1424 	rate = arch_timer_get_cntfrq();
1425 	arch_timer_of_configure_rate(rate, np);
1426 
1427 	arch_timer_c3stop = !of_property_read_bool(np, "always-on");
1428 
1429 	/* Check for globally applicable workarounds */
1430 	arch_timer_check_ool_workaround(ate_match_dt, np);
1431 
1432 	/*
1433 	 * If we cannot rely on firmware initializing the timer registers then
1434 	 * we should use the physical timers instead.
1435 	 */
1436 	if (IS_ENABLED(CONFIG_ARM) &&
1437 	    of_property_read_bool(np, "arm,cpu-registers-not-fw-configured"))
1438 		arch_timer_uses_ppi = ARCH_TIMER_PHYS_SECURE_PPI;
1439 	else
1440 		arch_timer_uses_ppi = arch_timer_select_ppi();
1441 
1442 	if (!arch_timer_ppi[arch_timer_uses_ppi]) {
1443 		pr_err("No interrupt available, giving up\n");
1444 		return -EINVAL;
1445 	}
1446 
1447 	/* On some systems, the counter stops ticking when in suspend. */
1448 	arch_counter_suspend_stop = of_property_read_bool(np,
1449 							 "arm,no-tick-in-suspend");
1450 
1451 	ret = arch_timer_register();
1452 	if (ret)
1453 		return ret;
1454 
1455 	if (arch_timer_needs_of_probing())
1456 		return 0;
1457 
1458 	return arch_timer_common_init();
1459 }
1460 TIMER_OF_DECLARE(armv7_arch_timer, "arm,armv7-timer", arch_timer_of_init);
1461 TIMER_OF_DECLARE(armv8_arch_timer, "arm,armv8-timer", arch_timer_of_init);
1462 
1463 static u32 __init
1464 arch_timer_mem_frame_get_cntfrq(struct arch_timer_mem_frame *frame)
1465 {
1466 	void __iomem *base;
1467 	u32 rate;
1468 
1469 	base = ioremap(frame->cntbase, frame->size);
1470 	if (!base) {
1471 		pr_err("Unable to map frame @ %pa\n", &frame->cntbase);
1472 		return 0;
1473 	}
1474 
1475 	rate = readl_relaxed(base + CNTFRQ);
1476 
1477 	iounmap(base);
1478 
1479 	return rate;
1480 }
1481 
1482 static struct arch_timer_mem_frame * __init
1483 arch_timer_mem_find_best_frame(struct arch_timer_mem *timer_mem)
1484 {
1485 	struct arch_timer_mem_frame *frame, *best_frame = NULL;
1486 	void __iomem *cntctlbase;
1487 	u32 cnttidr;
1488 	int i;
1489 
1490 	cntctlbase = ioremap(timer_mem->cntctlbase, timer_mem->size);
1491 	if (!cntctlbase) {
1492 		pr_err("Can't map CNTCTLBase @ %pa\n",
1493 			&timer_mem->cntctlbase);
1494 		return NULL;
1495 	}
1496 
1497 	cnttidr = readl_relaxed(cntctlbase + CNTTIDR);
1498 
1499 	/*
1500 	 * Try to find a virtual capable frame. Otherwise fall back to a
1501 	 * physical capable frame.
1502 	 */
1503 	for (i = 0; i < ARCH_TIMER_MEM_MAX_FRAMES; i++) {
1504 		u32 cntacr = CNTACR_RFRQ | CNTACR_RWPT | CNTACR_RPCT |
1505 			     CNTACR_RWVT | CNTACR_RVOFF | CNTACR_RVCT;
1506 
1507 		frame = &timer_mem->frame[i];
1508 		if (!frame->valid)
1509 			continue;
1510 
1511 		/* Try enabling everything, and see what sticks */
1512 		writel_relaxed(cntacr, cntctlbase + CNTACR(i));
1513 		cntacr = readl_relaxed(cntctlbase + CNTACR(i));
1514 
1515 		if ((cnttidr & CNTTIDR_VIRT(i)) &&
1516 		    !(~cntacr & (CNTACR_RWVT | CNTACR_RVCT))) {
1517 			best_frame = frame;
1518 			arch_timer_mem_use_virtual = true;
1519 			break;
1520 		}
1521 
1522 		if (~cntacr & (CNTACR_RWPT | CNTACR_RPCT))
1523 			continue;
1524 
1525 		best_frame = frame;
1526 	}
1527 
1528 	iounmap(cntctlbase);
1529 
1530 	return best_frame;
1531 }
1532 
1533 static int __init
1534 arch_timer_mem_frame_register(struct arch_timer_mem_frame *frame)
1535 {
1536 	void __iomem *base;
1537 	int ret, irq = 0;
1538 
1539 	if (arch_timer_mem_use_virtual)
1540 		irq = frame->virt_irq;
1541 	else
1542 		irq = frame->phys_irq;
1543 
1544 	if (!irq) {
1545 		pr_err("Frame missing %s irq.\n",
1546 		       arch_timer_mem_use_virtual ? "virt" : "phys");
1547 		return -EINVAL;
1548 	}
1549 
1550 	if (!request_mem_region(frame->cntbase, frame->size,
1551 				"arch_mem_timer"))
1552 		return -EBUSY;
1553 
1554 	base = ioremap(frame->cntbase, frame->size);
1555 	if (!base) {
1556 		pr_err("Can't map frame's registers\n");
1557 		return -ENXIO;
1558 	}
1559 
1560 	ret = arch_timer_mem_register(base, irq);
1561 	if (ret) {
1562 		iounmap(base);
1563 		return ret;
1564 	}
1565 
1566 	arch_timers_present |= ARCH_TIMER_TYPE_MEM;
1567 
1568 	return 0;
1569 }
1570 
1571 static int __init arch_timer_mem_of_init(struct device_node *np)
1572 {
1573 	struct arch_timer_mem *timer_mem;
1574 	struct arch_timer_mem_frame *frame;
1575 	struct device_node *frame_node;
1576 	struct resource res;
1577 	int ret = -EINVAL;
1578 	u32 rate;
1579 
1580 	timer_mem = kzalloc(sizeof(*timer_mem), GFP_KERNEL);
1581 	if (!timer_mem)
1582 		return -ENOMEM;
1583 
1584 	if (of_address_to_resource(np, 0, &res))
1585 		goto out;
1586 	timer_mem->cntctlbase = res.start;
1587 	timer_mem->size = resource_size(&res);
1588 
1589 	for_each_available_child_of_node(np, frame_node) {
1590 		u32 n;
1591 		struct arch_timer_mem_frame *frame;
1592 
1593 		if (of_property_read_u32(frame_node, "frame-number", &n)) {
1594 			pr_err(FW_BUG "Missing frame-number.\n");
1595 			of_node_put(frame_node);
1596 			goto out;
1597 		}
1598 		if (n >= ARCH_TIMER_MEM_MAX_FRAMES) {
1599 			pr_err(FW_BUG "Wrong frame-number, only 0-%u are permitted.\n",
1600 			       ARCH_TIMER_MEM_MAX_FRAMES - 1);
1601 			of_node_put(frame_node);
1602 			goto out;
1603 		}
1604 		frame = &timer_mem->frame[n];
1605 
1606 		if (frame->valid) {
1607 			pr_err(FW_BUG "Duplicated frame-number.\n");
1608 			of_node_put(frame_node);
1609 			goto out;
1610 		}
1611 
1612 		if (of_address_to_resource(frame_node, 0, &res)) {
1613 			of_node_put(frame_node);
1614 			goto out;
1615 		}
1616 		frame->cntbase = res.start;
1617 		frame->size = resource_size(&res);
1618 
1619 		frame->virt_irq = irq_of_parse_and_map(frame_node,
1620 						       ARCH_TIMER_VIRT_SPI);
1621 		frame->phys_irq = irq_of_parse_and_map(frame_node,
1622 						       ARCH_TIMER_PHYS_SPI);
1623 
1624 		frame->valid = true;
1625 	}
1626 
1627 	frame = arch_timer_mem_find_best_frame(timer_mem);
1628 	if (!frame) {
1629 		pr_err("Unable to find a suitable frame in timer @ %pa\n",
1630 			&timer_mem->cntctlbase);
1631 		ret = -EINVAL;
1632 		goto out;
1633 	}
1634 
1635 	rate = arch_timer_mem_frame_get_cntfrq(frame);
1636 	arch_timer_of_configure_rate(rate, np);
1637 
1638 	ret = arch_timer_mem_frame_register(frame);
1639 	if (!ret && !arch_timer_needs_of_probing())
1640 		ret = arch_timer_common_init();
1641 out:
1642 	kfree(timer_mem);
1643 	return ret;
1644 }
1645 TIMER_OF_DECLARE(armv7_arch_timer_mem, "arm,armv7-timer-mem",
1646 		       arch_timer_mem_of_init);
1647 
1648 #ifdef CONFIG_ACPI_GTDT
1649 static int __init
1650 arch_timer_mem_verify_cntfrq(struct arch_timer_mem *timer_mem)
1651 {
1652 	struct arch_timer_mem_frame *frame;
1653 	u32 rate;
1654 	int i;
1655 
1656 	for (i = 0; i < ARCH_TIMER_MEM_MAX_FRAMES; i++) {
1657 		frame = &timer_mem->frame[i];
1658 
1659 		if (!frame->valid)
1660 			continue;
1661 
1662 		rate = arch_timer_mem_frame_get_cntfrq(frame);
1663 		if (rate == arch_timer_rate)
1664 			continue;
1665 
1666 		pr_err(FW_BUG "CNTFRQ mismatch: frame @ %pa: (0x%08lx), CPU: (0x%08lx)\n",
1667 			&frame->cntbase,
1668 			(unsigned long)rate, (unsigned long)arch_timer_rate);
1669 
1670 		return -EINVAL;
1671 	}
1672 
1673 	return 0;
1674 }
1675 
1676 static int __init arch_timer_mem_acpi_init(int platform_timer_count)
1677 {
1678 	struct arch_timer_mem *timers, *timer;
1679 	struct arch_timer_mem_frame *frame, *best_frame = NULL;
1680 	int timer_count, i, ret = 0;
1681 
1682 	timers = kcalloc(platform_timer_count, sizeof(*timers),
1683 			    GFP_KERNEL);
1684 	if (!timers)
1685 		return -ENOMEM;
1686 
1687 	ret = acpi_arch_timer_mem_init(timers, &timer_count);
1688 	if (ret || !timer_count)
1689 		goto out;
1690 
1691 	/*
1692 	 * While unlikely, it's theoretically possible that none of the frames
1693 	 * in a timer expose the combination of feature we want.
1694 	 */
1695 	for (i = 0; i < timer_count; i++) {
1696 		timer = &timers[i];
1697 
1698 		frame = arch_timer_mem_find_best_frame(timer);
1699 		if (!best_frame)
1700 			best_frame = frame;
1701 
1702 		ret = arch_timer_mem_verify_cntfrq(timer);
1703 		if (ret) {
1704 			pr_err("Disabling MMIO timers due to CNTFRQ mismatch\n");
1705 			goto out;
1706 		}
1707 
1708 		if (!best_frame) /* implies !frame */
1709 			/*
1710 			 * Only complain about missing suitable frames if we
1711 			 * haven't already found one in a previous iteration.
1712 			 */
1713 			pr_err("Unable to find a suitable frame in timer @ %pa\n",
1714 				&timer->cntctlbase);
1715 	}
1716 
1717 	if (best_frame)
1718 		ret = arch_timer_mem_frame_register(best_frame);
1719 out:
1720 	kfree(timers);
1721 	return ret;
1722 }
1723 
1724 /* Initialize per-processor generic timer and memory-mapped timer(if present) */
1725 static int __init arch_timer_acpi_init(struct acpi_table_header *table)
1726 {
1727 	int ret, platform_timer_count;
1728 
1729 	if (arch_timers_present & ARCH_TIMER_TYPE_CP15) {
1730 		pr_warn("already initialized, skipping\n");
1731 		return -EINVAL;
1732 	}
1733 
1734 	arch_timers_present |= ARCH_TIMER_TYPE_CP15;
1735 
1736 	ret = acpi_gtdt_init(table, &platform_timer_count);
1737 	if (ret)
1738 		return ret;
1739 
1740 	arch_timer_ppi[ARCH_TIMER_PHYS_NONSECURE_PPI] =
1741 		acpi_gtdt_map_ppi(ARCH_TIMER_PHYS_NONSECURE_PPI);
1742 
1743 	arch_timer_ppi[ARCH_TIMER_VIRT_PPI] =
1744 		acpi_gtdt_map_ppi(ARCH_TIMER_VIRT_PPI);
1745 
1746 	arch_timer_ppi[ARCH_TIMER_HYP_PPI] =
1747 		acpi_gtdt_map_ppi(ARCH_TIMER_HYP_PPI);
1748 
1749 	arch_timer_populate_kvm_info();
1750 
1751 	/*
1752 	 * When probing via ACPI, we have no mechanism to override the sysreg
1753 	 * CNTFRQ value. This *must* be correct.
1754 	 */
1755 	arch_timer_rate = arch_timer_get_cntfrq();
1756 	ret = validate_timer_rate();
1757 	if (ret) {
1758 		pr_err(FW_BUG "frequency not available.\n");
1759 		return ret;
1760 	}
1761 
1762 	arch_timer_uses_ppi = arch_timer_select_ppi();
1763 	if (!arch_timer_ppi[arch_timer_uses_ppi]) {
1764 		pr_err("No interrupt available, giving up\n");
1765 		return -EINVAL;
1766 	}
1767 
1768 	/* Always-on capability */
1769 	arch_timer_c3stop = acpi_gtdt_c3stop(arch_timer_uses_ppi);
1770 
1771 	/* Check for globally applicable workarounds */
1772 	arch_timer_check_ool_workaround(ate_match_acpi_oem_info, table);
1773 
1774 	ret = arch_timer_register();
1775 	if (ret)
1776 		return ret;
1777 
1778 	if (platform_timer_count &&
1779 	    arch_timer_mem_acpi_init(platform_timer_count))
1780 		pr_err("Failed to initialize memory-mapped timer.\n");
1781 
1782 	return arch_timer_common_init();
1783 }
1784 TIMER_ACPI_DECLARE(arch_timer, ACPI_SIG_GTDT, arch_timer_acpi_init);
1785 #endif
1786 
1787 int kvm_arch_ptp_get_crosststamp(u64 *cycle, struct timespec64 *ts,
1788 				 struct clocksource **cs)
1789 {
1790 	struct arm_smccc_res hvc_res;
1791 	u32 ptp_counter;
1792 	ktime_t ktime;
1793 
1794 	if (!IS_ENABLED(CONFIG_HAVE_ARM_SMCCC_DISCOVERY))
1795 		return -EOPNOTSUPP;
1796 
1797 	if (arch_timer_uses_ppi == ARCH_TIMER_VIRT_PPI)
1798 		ptp_counter = KVM_PTP_VIRT_COUNTER;
1799 	else
1800 		ptp_counter = KVM_PTP_PHYS_COUNTER;
1801 
1802 	arm_smccc_1_1_invoke(ARM_SMCCC_VENDOR_HYP_KVM_PTP_FUNC_ID,
1803 			     ptp_counter, &hvc_res);
1804 
1805 	if ((int)(hvc_res.a0) < 0)
1806 		return -EOPNOTSUPP;
1807 
1808 	ktime = (u64)hvc_res.a0 << 32 | hvc_res.a1;
1809 	*ts = ktime_to_timespec64(ktime);
1810 	if (cycle)
1811 		*cycle = (u64)hvc_res.a2 << 32 | hvc_res.a3;
1812 	if (cs)
1813 		*cs = &clocksource_counter;
1814 
1815 	return 0;
1816 }
1817 EXPORT_SYMBOL_GPL(kvm_arch_ptp_get_crosststamp);
1818