2 * apb_timer.c: Driver for Langwell APB timers
4 * (C) Copyright 2009 Intel Corporation
5 * Author: Jacob Pan (jacob.jun.pan@intel.com)
7 * This program is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU General Public License
9 * as published by the Free Software Foundation; version 2
13 * Langwell is the south complex of Intel Moorestown MID platform. There are
14 * eight external timers in total that can be used by the operating system.
15 * The timer information, such as frequency and addresses, is provided to the
17 * Timer interrupts are routed via FW/HW emulated IOAPIC independently via
18 * individual redirection table entries (RTE).
19 * Unlike HPET, there is no master counter, therefore one of the timers are
20 * used as clocksource. The overall allocation looks like:
21 * - timer 0 - NR_CPUs for per cpu timer
22 * - one timer for clocksource
23 * - one timer for watchdog driver.
24 * It is also worth notice that APB timer does not support true one-shot mode,
25 * free-running mode will be used here to emulate one-shot mode.
26 * APB timer can also be used as broadcast timer along with per cpu local APIC
27 * timer, but by default APB timer has higher rating than local APIC timers.
30 #include <linux/clocksource.h>
31 #include <linux/clockchips.h>
32 #include <linux/delay.h>
33 #include <linux/errno.h>
34 #include <linux/init.h>
35 #include <linux/sysdev.h>
36 #include <linux/slab.h>
38 #include <linux/pci.h>
39 #include <linux/sfi.h>
40 #include <linux/interrupt.h>
41 #include <linux/cpu.h>
42 #include <linux/irq.h>
44 #include <asm/fixmap.h>
45 #include <asm/apb_timer.h>
47 #define APBT_MASK CLOCKSOURCE_MASK(32)
49 #define APBT_CLOCKEVENT_RATING 150
50 #define APBT_CLOCKSOURCE_RATING 250
51 #define APBT_MIN_DELTA_USEC 200
53 #define EVT_TO_APBT_DEV(evt) container_of(evt, struct apbt_dev, evt)
54 #define APBT_CLOCKEVENT0_NUM (0)
55 #define APBT_CLOCKEVENT1_NUM (1)
56 #define APBT_CLOCKSOURCE_NUM (2)
58 static unsigned long apbt_address;
59 static int apb_timer_block_enabled;
60 static void __iomem *apbt_virt_address;
61 static int phy_cs_timer_id;
64 * Common DW APB timer info
66 static uint64_t apbt_freq;
68 static void apbt_set_mode(enum clock_event_mode mode,
69 struct clock_event_device *evt);
70 static int apbt_next_event(unsigned long delta,
71 struct clock_event_device *evt);
72 static cycle_t apbt_read_clocksource(struct clocksource *cs);
73 static void apbt_restart_clocksource(struct clocksource *cs);
76 struct clock_event_device evt;
86 int disable_apbt_percpu __cpuinitdata;
88 static DEFINE_PER_CPU(struct apbt_dev, cpu_apbt_dev);
91 static unsigned int apbt_num_timers_used;
92 static struct apbt_dev *apbt_devs;
95 static inline unsigned long apbt_readl_reg(unsigned long a)
97 return readl(apbt_virt_address + a);
100 static inline void apbt_writel_reg(unsigned long d, unsigned long a)
102 writel(d, apbt_virt_address + a);
105 static inline unsigned long apbt_readl(int n, unsigned long a)
107 return readl(apbt_virt_address + a + n * APBTMRS_REG_SIZE);
110 static inline void apbt_writel(int n, unsigned long d, unsigned long a)
112 writel(d, apbt_virt_address + a + n * APBTMRS_REG_SIZE);
115 static inline void apbt_set_mapping(void)
117 struct sfi_timer_table_entry *mtmr;
119 if (apbt_virt_address) {
120 pr_debug("APBT base already mapped\n");
123 mtmr = sfi_get_mtmr(APBT_CLOCKEVENT0_NUM);
125 printk(KERN_ERR "Failed to get MTMR %d from SFI\n",
126 APBT_CLOCKEVENT0_NUM);
129 apbt_address = (unsigned long)mtmr->phys_addr;
131 printk(KERN_WARNING "No timer base from SFI, use default\n");
132 apbt_address = APBT_DEFAULT_BASE;
134 apbt_virt_address = ioremap_nocache(apbt_address, APBT_MMAP_SIZE);
135 if (apbt_virt_address) {
136 pr_debug("Mapped APBT physical addr %p at virtual addr %p\n",\
137 (void *)apbt_address, (void *)apbt_virt_address);
139 pr_debug("Failed mapping APBT phy address at %p\n",\
140 (void *)apbt_address);
143 apbt_freq = mtmr->freq_hz / USEC_PER_SEC;
146 /* Now figure out the physical timer id for clocksource device */
147 mtmr = sfi_get_mtmr(APBT_CLOCKSOURCE_NUM);
151 /* Now figure out the physical timer id */
152 phy_cs_timer_id = (unsigned int)(mtmr->phys_addr & 0xff)
154 pr_debug("Use timer %d for clocksource\n", phy_cs_timer_id);
158 panic("Failed to setup APB system timer\n");
162 static inline void apbt_clear_mapping(void)
164 iounmap(apbt_virt_address);
165 apbt_virt_address = NULL;
169 * APBT timer interrupt enable / disable
171 static inline int is_apbt_capable(void)
173 return apbt_virt_address ? 1 : 0;
176 static struct clocksource clocksource_apbt = {
178 .rating = APBT_CLOCKSOURCE_RATING,
179 .read = apbt_read_clocksource,
182 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
183 .resume = apbt_restart_clocksource,
186 /* boot APB clock event device */
187 static struct clock_event_device apbt_clockevent = {
189 .features = CLOCK_EVT_FEAT_PERIODIC | CLOCK_EVT_FEAT_ONESHOT,
190 .set_mode = apbt_set_mode,
191 .set_next_event = apbt_next_event,
194 .rating = APBT_CLOCKEVENT_RATING,
198 * if user does not want to use per CPU apb timer, just give it a lower rating
199 * than local apic timer and skip the late per cpu timer init.
201 static inline int __init setup_x86_mrst_timer(char *arg)
206 if (strcmp("apbt_only", arg) == 0)
207 disable_apbt_percpu = 0;
208 else if (strcmp("lapic_and_apbt", arg) == 0)
209 disable_apbt_percpu = 1;
211 pr_warning("X86 MRST timer option %s not recognised"
212 " use x86_mrst_timer=apbt_only or lapic_and_apbt\n",
218 __setup("x86_mrst_timer=", setup_x86_mrst_timer);
221 * start count down from 0xffff_ffff. this is done by toggling the enable bit
222 * then load initial load count to ~0.
224 static void apbt_start_counter(int n)
226 unsigned long ctrl = apbt_readl(n, APBTMR_N_CONTROL);
228 ctrl &= ~APBTMR_CONTROL_ENABLE;
229 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
230 apbt_writel(n, ~0, APBTMR_N_LOAD_COUNT);
231 /* enable, mask interrupt */
232 ctrl &= ~APBTMR_CONTROL_MODE_PERIODIC;
233 ctrl |= (APBTMR_CONTROL_ENABLE | APBTMR_CONTROL_INT);
234 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
235 /* read it once to get cached counter value initialized */
236 apbt_read_clocksource(&clocksource_apbt);
239 static irqreturn_t apbt_interrupt_handler(int irq, void *data)
241 struct apbt_dev *dev = (struct apbt_dev *)data;
242 struct clock_event_device *aevt = &dev->evt;
244 if (!aevt->event_handler) {
245 printk(KERN_INFO "Spurious APBT timer interrupt on %d\n",
249 aevt->event_handler(aevt);
253 static void apbt_restart_clocksource(struct clocksource *cs)
255 apbt_start_counter(phy_cs_timer_id);
258 /* Setup IRQ routing via IOAPIC */
260 static void apbt_setup_irq(struct apbt_dev *adev)
262 struct irq_chip *chip;
263 struct irq_desc *desc;
265 /* timer0 irq has been setup early */
268 desc = irq_to_desc(adev->irq);
269 chip = get_irq_chip(adev->irq);
270 disable_irq(adev->irq);
271 desc->status |= IRQ_MOVE_PCNTXT;
272 irq_set_affinity(adev->irq, cpumask_of(adev->cpu));
273 /* APB timer irqs are set up as mp_irqs, timer is edge triggerred */
274 set_irq_chip_and_handler_name(adev->irq, chip, handle_edge_irq, "edge");
275 enable_irq(adev->irq);
276 if (system_state == SYSTEM_BOOTING)
277 if (request_irq(adev->irq, apbt_interrupt_handler,
278 IRQF_TIMER | IRQF_DISABLED | IRQF_NOBALANCING,
280 printk(KERN_ERR "Failed request IRQ for APBT%d\n",
286 static void apbt_enable_int(int n)
288 unsigned long ctrl = apbt_readl(n, APBTMR_N_CONTROL);
289 /* clear pending intr */
290 apbt_readl(n, APBTMR_N_EOI);
291 ctrl &= ~APBTMR_CONTROL_INT;
292 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
295 static void apbt_disable_int(int n)
297 unsigned long ctrl = apbt_readl(n, APBTMR_N_CONTROL);
299 ctrl |= APBTMR_CONTROL_INT;
300 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
304 static int __init apbt_clockevent_register(void)
306 struct sfi_timer_table_entry *mtmr;
307 struct apbt_dev *adev = &__get_cpu_var(cpu_apbt_dev);
309 mtmr = sfi_get_mtmr(APBT_CLOCKEVENT0_NUM);
311 printk(KERN_ERR "Failed to get MTMR %d from SFI\n",
312 APBT_CLOCKEVENT0_NUM);
317 * We need to calculate the scaled math multiplication factor for
318 * nanosecond to apbt tick conversion.
319 * mult = (nsec/cycle)*2^APBT_SHIFT
321 apbt_clockevent.mult = div_sc((unsigned long) mtmr->freq_hz
322 , NSEC_PER_SEC, APBT_SHIFT);
324 /* Calculate the min / max delta */
325 apbt_clockevent.max_delta_ns = clockevent_delta2ns(0x7FFFFFFF,
327 apbt_clockevent.min_delta_ns = clockevent_delta2ns(
328 APBT_MIN_DELTA_USEC*apbt_freq,
331 * Start apbt with the boot cpu mask and make it
332 * global if not used for per cpu timer.
334 apbt_clockevent.cpumask = cpumask_of(smp_processor_id());
335 adev->num = smp_processor_id();
336 memcpy(&adev->evt, &apbt_clockevent, sizeof(struct clock_event_device));
338 if (disable_apbt_percpu) {
339 apbt_clockevent.rating = APBT_CLOCKEVENT_RATING - 100;
340 global_clock_event = &adev->evt;
341 printk(KERN_DEBUG "%s clockevent registered as global\n",
342 global_clock_event->name);
345 if (request_irq(apbt_clockevent.irq, apbt_interrupt_handler,
346 IRQF_TIMER | IRQF_DISABLED | IRQF_NOBALANCING,
347 apbt_clockevent.name, adev)) {
348 printk(KERN_ERR "Failed request IRQ for APBT%d\n",
349 apbt_clockevent.irq);
352 clockevents_register_device(&adev->evt);
353 /* Start APBT 0 interrupts */
354 apbt_enable_int(APBT_CLOCKEVENT0_NUM);
361 /* Should be called with per cpu */
362 void apbt_setup_secondary_clock(void)
364 struct apbt_dev *adev;
365 struct clock_event_device *aevt;
368 /* Don't register boot CPU clockevent */
369 cpu = smp_processor_id();
370 if (cpu == boot_cpu_id)
373 * We need to calculate the scaled math multiplication factor for
374 * nanosecond to apbt tick conversion.
375 * mult = (nsec/cycle)*2^APBT_SHIFT
377 printk(KERN_INFO "Init per CPU clockevent %d\n", cpu);
378 adev = &per_cpu(cpu_apbt_dev, cpu);
381 memcpy(aevt, &apbt_clockevent, sizeof(*aevt));
382 aevt->cpumask = cpumask_of(cpu);
383 aevt->name = adev->name;
384 aevt->mode = CLOCK_EVT_MODE_UNUSED;
386 printk(KERN_INFO "Registering CPU %d clockevent device %s, mask %08x\n",
387 cpu, aevt->name, *(u32 *)aevt->cpumask);
389 apbt_setup_irq(adev);
391 clockevents_register_device(aevt);
393 apbt_enable_int(cpu);
399 * this notify handler process CPU hotplug events. in case of S0i3, nonboot
400 * cpus are disabled/enabled frequently, for performance reasons, we keep the
401 * per cpu timer irq registered so that we do need to do free_irq/request_irq.
403 * TODO: it might be more reliable to directly disable percpu clockevent device
404 * without the notifier chain. currently, cpu 0 may get interrupts from other
405 * cpu timers during the offline process due to the ordering of notification.
406 * the extra interrupt is harmless.
408 static int apbt_cpuhp_notify(struct notifier_block *n,
409 unsigned long action, void *hcpu)
411 unsigned long cpu = (unsigned long)hcpu;
412 struct apbt_dev *adev = &per_cpu(cpu_apbt_dev, cpu);
414 switch (action & 0xf) {
416 apbt_disable_int(cpu);
417 if (system_state == SYSTEM_RUNNING)
418 pr_debug("skipping APBT CPU %lu offline\n", cpu);
420 pr_debug("APBT clockevent for cpu %lu offline\n", cpu);
421 free_irq(adev->irq, adev);
425 pr_debug(KERN_INFO "APBT notified %lu, no action\n", action);
430 static __init int apbt_late_init(void)
432 if (disable_apbt_percpu)
434 /* This notifier should be called after workqueue is ready */
435 hotcpu_notifier(apbt_cpuhp_notify, -20);
438 fs_initcall(apbt_late_init);
441 void apbt_setup_secondary_clock(void) {}
443 #endif /* CONFIG_SMP */
445 static void apbt_set_mode(enum clock_event_mode mode,
446 struct clock_event_device *evt)
451 struct apbt_dev *adev = EVT_TO_APBT_DEV(evt);
453 timer_num = adev->num;
454 pr_debug("%s CPU %d timer %d mode=%d\n",
455 __func__, first_cpu(*evt->cpumask), timer_num, mode);
458 case CLOCK_EVT_MODE_PERIODIC:
459 delta = ((uint64_t)(NSEC_PER_SEC/HZ)) * apbt_clockevent.mult;
460 delta >>= apbt_clockevent.shift;
461 ctrl = apbt_readl(timer_num, APBTMR_N_CONTROL);
462 ctrl |= APBTMR_CONTROL_MODE_PERIODIC;
463 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
465 * DW APB p. 46, have to disable timer before load counter,
466 * may cause sync problem.
468 ctrl &= ~APBTMR_CONTROL_ENABLE;
469 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
471 pr_debug("Setting clock period %d for HZ %d\n", (int)delta, HZ);
472 apbt_writel(timer_num, delta, APBTMR_N_LOAD_COUNT);
473 ctrl |= APBTMR_CONTROL_ENABLE;
474 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
476 /* APB timer does not have one-shot mode, use free running mode */
477 case CLOCK_EVT_MODE_ONESHOT:
478 ctrl = apbt_readl(timer_num, APBTMR_N_CONTROL);
480 * set free running mode, this mode will let timer reload max
481 * timeout which will give time (3min on 25MHz clock) to rearm
482 * the next event, therefore emulate the one-shot mode.
484 ctrl &= ~APBTMR_CONTROL_ENABLE;
485 ctrl &= ~APBTMR_CONTROL_MODE_PERIODIC;
487 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
488 /* write again to set free running mode */
489 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
492 * DW APB p. 46, load counter with all 1s before starting free
495 apbt_writel(timer_num, ~0, APBTMR_N_LOAD_COUNT);
496 ctrl &= ~APBTMR_CONTROL_INT;
497 ctrl |= APBTMR_CONTROL_ENABLE;
498 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
501 case CLOCK_EVT_MODE_UNUSED:
502 case CLOCK_EVT_MODE_SHUTDOWN:
503 apbt_disable_int(timer_num);
504 ctrl = apbt_readl(timer_num, APBTMR_N_CONTROL);
505 ctrl &= ~APBTMR_CONTROL_ENABLE;
506 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
509 case CLOCK_EVT_MODE_RESUME:
510 apbt_enable_int(timer_num);
515 static int apbt_next_event(unsigned long delta,
516 struct clock_event_device *evt)
521 struct apbt_dev *adev = EVT_TO_APBT_DEV(evt);
523 timer_num = adev->num;
525 ctrl = apbt_readl(timer_num, APBTMR_N_CONTROL);
526 ctrl &= ~APBTMR_CONTROL_ENABLE;
527 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
528 /* write new count */
529 apbt_writel(timer_num, delta, APBTMR_N_LOAD_COUNT);
530 ctrl |= APBTMR_CONTROL_ENABLE;
531 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
536 * APB timer clock is not in sync with pclk on Langwell, which translates to
537 * unreliable read value caused by sampling error. the error does not add up
538 * overtime and only happens when sampling a 0 as a 1 by mistake. so the time
539 * would go backwards. the following code is trying to prevent time traveling
540 * backwards. little bit paranoid.
542 static cycle_t apbt_read_clocksource(struct clocksource *cs)
544 unsigned long t0, t1, t2;
545 static unsigned long last_read;
548 t1 = apbt_readl(phy_cs_timer_id,
549 APBTMR_N_CURRENT_VALUE);
550 t2 = apbt_readl(phy_cs_timer_id,
551 APBTMR_N_CURRENT_VALUE);
552 if (unlikely(t1 < t2)) {
553 pr_debug("APBT: read current count error %lx:%lx:%lx\n",
558 * check against cached last read, makes sure time does not go back.
559 * it could be a normal rollover but we will do tripple check anyway
561 if (unlikely(t2 > last_read)) {
562 /* check if we have a normal rollover */
563 unsigned long raw_intr_status =
564 apbt_readl_reg(APBTMRS_RAW_INT_STATUS);
566 * cs timer interrupt is masked but raw intr bit is set if
567 * rollover occurs. then we read EOI reg to clear it.
569 if (raw_intr_status & (1 << phy_cs_timer_id)) {
570 apbt_readl(phy_cs_timer_id, APBTMR_N_EOI);
573 pr_debug("APB CS going back %lx:%lx:%lx ",
574 t2, last_read, t2 - last_read);
576 pr_debug(KERN_INFO "tripple check enforced\n");
577 t0 = apbt_readl(phy_cs_timer_id,
578 APBTMR_N_CURRENT_VALUE);
580 t1 = apbt_readl(phy_cs_timer_id,
581 APBTMR_N_CURRENT_VALUE);
583 t2 = apbt_readl(phy_cs_timer_id,
584 APBTMR_N_CURRENT_VALUE);
585 if ((t2 > t1) || (t1 > t0)) {
586 printk(KERN_ERR "Error: APB CS tripple check failed\n");
595 static int apbt_clocksource_register(void)
600 /* Start the counter, use timer 2 as source, timer 0/1 for event */
601 apbt_start_counter(phy_cs_timer_id);
603 /* Verify whether apbt counter works */
604 t1 = apbt_read_clocksource(&clocksource_apbt);
608 * We don't know the TSC frequency yet, but waiting for
609 * 200000 TSC cycles is safe:
616 } while ((now - start) < 200000UL);
618 /* APBT is the only always on clocksource, it has to work! */
619 if (t1 == apbt_read_clocksource(&clocksource_apbt))
620 panic("APBT counter not counting. APBT disabled\n");
623 * initialize and register APBT clocksource
624 * convert that to ns/clock cycle
625 * mult = (ns/c) * 2^APBT_SHIFT
627 clocksource_apbt.mult = div_sc(MSEC_PER_SEC,
628 (unsigned long) apbt_freq, APBT_SHIFT);
629 clocksource_register(&clocksource_apbt);
635 * Early setup the APBT timer, only use timer 0 for booting then switch to
636 * per CPU timer if possible.
637 * returns 1 if per cpu apbt is setup
638 * returns 0 if no per cpu apbt is chosen
639 * panic if set up failed, this is the only platform timer on Moorestown.
641 void __init apbt_time_init(void)
645 struct sfi_timer_table_entry *p_mtmr;
646 unsigned int percpu_timer;
647 struct apbt_dev *adev;
650 if (apb_timer_block_enabled)
653 if (apbt_virt_address) {
654 pr_debug("Found APBT version 0x%lx\n",\
655 apbt_readl_reg(APBTMRS_COMP_VERSION));
659 * Read the frequency and check for a sane value, for ESL model
660 * we extend the possible clock range to allow time scaling.
663 if (apbt_freq < APBT_MIN_FREQ || apbt_freq > APBT_MAX_FREQ) {
664 pr_debug("APBT has invalid freq 0x%llx\n", apbt_freq);
667 if (apbt_clocksource_register()) {
668 pr_debug("APBT has failed to register clocksource\n");
671 if (!apbt_clockevent_register())
672 apb_timer_block_enabled = 1;
674 pr_debug("APBT has failed to register clockevent\n");
678 /* kernel cmdline disable apb timer, so we will use lapic timers */
679 if (disable_apbt_percpu) {
680 printk(KERN_INFO "apbt: disabled per cpu timer\n");
683 pr_debug("%s: %d CPUs online\n", __func__, num_online_cpus());
684 if (num_possible_cpus() <= sfi_mtimer_num) {
686 apbt_num_timers_used = num_possible_cpus();
689 apbt_num_timers_used = 1;
690 adev = &per_cpu(cpu_apbt_dev, 0);
691 adev->flags &= ~APBT_DEV_USED;
693 pr_debug("%s: %d APB timers used\n", __func__, apbt_num_timers_used);
695 /* here we set up per CPU timer data structure */
696 apbt_devs = kzalloc(sizeof(struct apbt_dev) * apbt_num_timers_used,
699 printk(KERN_ERR "Failed to allocate APB timer devices\n");
702 for (i = 0; i < apbt_num_timers_used; i++) {
703 adev = &per_cpu(cpu_apbt_dev, i);
706 p_mtmr = sfi_get_mtmr(i);
708 adev->tick = p_mtmr->freq_hz;
709 adev->irq = p_mtmr->irq;
711 printk(KERN_ERR "Failed to get timer for cpu %d\n", i);
713 sprintf(adev->name, "apbt%d", i);
720 apbt_clear_mapping();
721 apb_timer_block_enabled = 0;
722 panic("failed to enable APB timer\n");
725 static inline void apbt_disable(int n)
727 if (is_apbt_capable()) {
728 unsigned long ctrl = apbt_readl(n, APBTMR_N_CONTROL);
729 ctrl &= ~APBTMR_CONTROL_ENABLE;
730 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
734 /* called before apb_timer_enable, use early map */
735 unsigned long apbt_quick_calibrate()
740 unsigned long khz = 0;
744 apbt_start_counter(phy_cs_timer_id);
746 /* check if the timer can count down, otherwise return */
747 old = apbt_read_clocksource(&clocksource_apbt);
750 if (old != apbt_read_clocksource(&clocksource_apbt))
757 loop = (apbt_freq * 1000) << 4;
759 /* restart the timer to ensure it won't get to 0 in the calibration */
760 apbt_start_counter(phy_cs_timer_id);
762 old = apbt_read_clocksource(&clocksource_apbt);
765 t1 = __native_read_tsc();
768 new = apbt_read_clocksource(&clocksource_apbt);
771 t2 = __native_read_tsc();
774 if (unlikely(loop >> shift == 0)) {
776 "APBT TSC calibration failed, not enough resolution\n");
779 scale = (int)div_u64((t2 - t1), loop >> shift);
780 khz = (scale * apbt_freq * 1000) >> shift;
781 printk(KERN_INFO "TSC freq calculated by APB timer is %lu khz\n", khz);