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>
48 #define APBT_MASK CLOCKSOURCE_MASK(32)
50 #define APBT_CLOCKEVENT_RATING 110
51 #define APBT_CLOCKSOURCE_RATING 250
52 #define APBT_MIN_DELTA_USEC 200
54 #define EVT_TO_APBT_DEV(evt) container_of(evt, struct apbt_dev, evt)
55 #define APBT_CLOCKEVENT0_NUM (0)
56 #define APBT_CLOCKEVENT1_NUM (1)
57 #define APBT_CLOCKSOURCE_NUM (2)
59 static unsigned long apbt_address;
60 static int apb_timer_block_enabled;
61 static void __iomem *apbt_virt_address;
62 static int phy_cs_timer_id;
65 * Common DW APB timer info
67 static uint64_t apbt_freq;
69 static void apbt_set_mode(enum clock_event_mode mode,
70 struct clock_event_device *evt);
71 static int apbt_next_event(unsigned long delta,
72 struct clock_event_device *evt);
73 static cycle_t apbt_read_clocksource(struct clocksource *cs);
74 static void apbt_restart_clocksource(struct clocksource *cs);
77 struct clock_event_device evt;
87 static DEFINE_PER_CPU(struct apbt_dev, cpu_apbt_dev);
90 static unsigned int apbt_num_timers_used;
91 static struct apbt_dev *apbt_devs;
94 static inline unsigned long apbt_readl_reg(unsigned long a)
96 return readl(apbt_virt_address + a);
99 static inline void apbt_writel_reg(unsigned long d, unsigned long a)
101 writel(d, apbt_virt_address + a);
104 static inline unsigned long apbt_readl(int n, unsigned long a)
106 return readl(apbt_virt_address + a + n * APBTMRS_REG_SIZE);
109 static inline void apbt_writel(int n, unsigned long d, unsigned long a)
111 writel(d, apbt_virt_address + a + n * APBTMRS_REG_SIZE);
114 static inline void apbt_set_mapping(void)
116 struct sfi_timer_table_entry *mtmr;
118 if (apbt_virt_address) {
119 pr_debug("APBT base already mapped\n");
122 mtmr = sfi_get_mtmr(APBT_CLOCKEVENT0_NUM);
124 printk(KERN_ERR "Failed to get MTMR %d from SFI\n",
125 APBT_CLOCKEVENT0_NUM);
128 apbt_address = (unsigned long)mtmr->phys_addr;
130 printk(KERN_WARNING "No timer base from SFI, use default\n");
131 apbt_address = APBT_DEFAULT_BASE;
133 apbt_virt_address = ioremap_nocache(apbt_address, APBT_MMAP_SIZE);
134 if (apbt_virt_address) {
135 pr_debug("Mapped APBT physical addr %p at virtual addr %p\n",\
136 (void *)apbt_address, (void *)apbt_virt_address);
138 pr_debug("Failed mapping APBT phy address at %p\n",\
139 (void *)apbt_address);
142 apbt_freq = mtmr->freq_hz / USEC_PER_SEC;
145 /* Now figure out the physical timer id for clocksource device */
146 mtmr = sfi_get_mtmr(APBT_CLOCKSOURCE_NUM);
150 /* Now figure out the physical timer id */
151 phy_cs_timer_id = (unsigned int)(mtmr->phys_addr & 0xff)
153 pr_debug("Use timer %d for clocksource\n", phy_cs_timer_id);
157 panic("Failed to setup APB system timer\n");
161 static inline void apbt_clear_mapping(void)
163 iounmap(apbt_virt_address);
164 apbt_virt_address = NULL;
168 * APBT timer interrupt enable / disable
170 static inline int is_apbt_capable(void)
172 return apbt_virt_address ? 1 : 0;
175 static struct clocksource clocksource_apbt = {
177 .rating = APBT_CLOCKSOURCE_RATING,
178 .read = apbt_read_clocksource,
181 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
182 .resume = apbt_restart_clocksource,
185 /* boot APB clock event device */
186 static struct clock_event_device apbt_clockevent = {
188 .features = CLOCK_EVT_FEAT_PERIODIC | CLOCK_EVT_FEAT_ONESHOT,
189 .set_mode = apbt_set_mode,
190 .set_next_event = apbt_next_event,
193 .rating = APBT_CLOCKEVENT_RATING,
197 * start count down from 0xffff_ffff. this is done by toggling the enable bit
198 * then load initial load count to ~0.
200 static void apbt_start_counter(int n)
202 unsigned long ctrl = apbt_readl(n, APBTMR_N_CONTROL);
204 ctrl &= ~APBTMR_CONTROL_ENABLE;
205 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
206 apbt_writel(n, ~0, APBTMR_N_LOAD_COUNT);
207 /* enable, mask interrupt */
208 ctrl &= ~APBTMR_CONTROL_MODE_PERIODIC;
209 ctrl |= (APBTMR_CONTROL_ENABLE | APBTMR_CONTROL_INT);
210 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
211 /* read it once to get cached counter value initialized */
212 apbt_read_clocksource(&clocksource_apbt);
215 static irqreturn_t apbt_interrupt_handler(int irq, void *data)
217 struct apbt_dev *dev = (struct apbt_dev *)data;
218 struct clock_event_device *aevt = &dev->evt;
220 if (!aevt->event_handler) {
221 printk(KERN_INFO "Spurious APBT timer interrupt on %d\n",
225 aevt->event_handler(aevt);
229 static void apbt_restart_clocksource(struct clocksource *cs)
231 apbt_start_counter(phy_cs_timer_id);
234 /* Setup IRQ routing via IOAPIC */
236 static void apbt_setup_irq(struct apbt_dev *adev)
238 struct irq_chip *chip;
239 struct irq_desc *desc;
241 /* timer0 irq has been setup early */
244 desc = irq_to_desc(adev->irq);
245 chip = get_irq_chip(adev->irq);
246 disable_irq(adev->irq);
247 desc->status |= IRQ_MOVE_PCNTXT;
248 irq_set_affinity(adev->irq, cpumask_of(adev->cpu));
249 /* APB timer irqs are set up as mp_irqs, timer is edge triggerred */
250 set_irq_chip_and_handler_name(adev->irq, chip, handle_edge_irq, "edge");
251 enable_irq(adev->irq);
252 if (system_state == SYSTEM_BOOTING)
253 if (request_irq(adev->irq, apbt_interrupt_handler,
254 IRQF_TIMER | IRQF_DISABLED | IRQF_NOBALANCING,
256 printk(KERN_ERR "Failed request IRQ for APBT%d\n",
262 static void apbt_enable_int(int n)
264 unsigned long ctrl = apbt_readl(n, APBTMR_N_CONTROL);
265 /* clear pending intr */
266 apbt_readl(n, APBTMR_N_EOI);
267 ctrl &= ~APBTMR_CONTROL_INT;
268 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
271 static void apbt_disable_int(int n)
273 unsigned long ctrl = apbt_readl(n, APBTMR_N_CONTROL);
275 ctrl |= APBTMR_CONTROL_INT;
276 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
280 static int __init apbt_clockevent_register(void)
282 struct sfi_timer_table_entry *mtmr;
283 struct apbt_dev *adev = &__get_cpu_var(cpu_apbt_dev);
285 mtmr = sfi_get_mtmr(APBT_CLOCKEVENT0_NUM);
287 printk(KERN_ERR "Failed to get MTMR %d from SFI\n",
288 APBT_CLOCKEVENT0_NUM);
293 * We need to calculate the scaled math multiplication factor for
294 * nanosecond to apbt tick conversion.
295 * mult = (nsec/cycle)*2^APBT_SHIFT
297 apbt_clockevent.mult = div_sc((unsigned long) mtmr->freq_hz
298 , NSEC_PER_SEC, APBT_SHIFT);
300 /* Calculate the min / max delta */
301 apbt_clockevent.max_delta_ns = clockevent_delta2ns(0x7FFFFFFF,
303 apbt_clockevent.min_delta_ns = clockevent_delta2ns(
304 APBT_MIN_DELTA_USEC*apbt_freq,
307 * Start apbt with the boot cpu mask and make it
308 * global if not used for per cpu timer.
310 apbt_clockevent.cpumask = cpumask_of(smp_processor_id());
311 adev->num = smp_processor_id();
312 memcpy(&adev->evt, &apbt_clockevent, sizeof(struct clock_event_device));
314 if (mrst_timer_options == MRST_TIMER_LAPIC_APBT) {
315 apbt_clockevent.rating = APBT_CLOCKEVENT_RATING - 100;
316 global_clock_event = &adev->evt;
317 printk(KERN_DEBUG "%s clockevent registered as global\n",
318 global_clock_event->name);
321 if (request_irq(apbt_clockevent.irq, apbt_interrupt_handler,
322 IRQF_TIMER | IRQF_DISABLED | IRQF_NOBALANCING,
323 apbt_clockevent.name, adev)) {
324 printk(KERN_ERR "Failed request IRQ for APBT%d\n",
325 apbt_clockevent.irq);
328 clockevents_register_device(&adev->evt);
329 /* Start APBT 0 interrupts */
330 apbt_enable_int(APBT_CLOCKEVENT0_NUM);
337 /* Should be called with per cpu */
338 void apbt_setup_secondary_clock(void)
340 struct apbt_dev *adev;
341 struct clock_event_device *aevt;
344 /* Don't register boot CPU clockevent */
345 cpu = smp_processor_id();
346 if (cpu == boot_cpu_id)
349 * We need to calculate the scaled math multiplication factor for
350 * nanosecond to apbt tick conversion.
351 * mult = (nsec/cycle)*2^APBT_SHIFT
353 printk(KERN_INFO "Init per CPU clockevent %d\n", cpu);
354 adev = &per_cpu(cpu_apbt_dev, cpu);
357 memcpy(aevt, &apbt_clockevent, sizeof(*aevt));
358 aevt->cpumask = cpumask_of(cpu);
359 aevt->name = adev->name;
360 aevt->mode = CLOCK_EVT_MODE_UNUSED;
362 printk(KERN_INFO "Registering CPU %d clockevent device %s, mask %08x\n",
363 cpu, aevt->name, *(u32 *)aevt->cpumask);
365 apbt_setup_irq(adev);
367 clockevents_register_device(aevt);
369 apbt_enable_int(cpu);
375 * this notify handler process CPU hotplug events. in case of S0i3, nonboot
376 * cpus are disabled/enabled frequently, for performance reasons, we keep the
377 * per cpu timer irq registered so that we do need to do free_irq/request_irq.
379 * TODO: it might be more reliable to directly disable percpu clockevent device
380 * without the notifier chain. currently, cpu 0 may get interrupts from other
381 * cpu timers during the offline process due to the ordering of notification.
382 * the extra interrupt is harmless.
384 static int apbt_cpuhp_notify(struct notifier_block *n,
385 unsigned long action, void *hcpu)
387 unsigned long cpu = (unsigned long)hcpu;
388 struct apbt_dev *adev = &per_cpu(cpu_apbt_dev, cpu);
390 switch (action & 0xf) {
392 apbt_disable_int(cpu);
393 if (system_state == SYSTEM_RUNNING)
394 pr_debug("skipping APBT CPU %lu offline\n", cpu);
396 pr_debug("APBT clockevent for cpu %lu offline\n", cpu);
397 free_irq(adev->irq, adev);
401 pr_debug(KERN_INFO "APBT notified %lu, no action\n", action);
406 static __init int apbt_late_init(void)
408 if (mrst_timer_options == MRST_TIMER_LAPIC_APBT ||
409 !apb_timer_block_enabled)
411 /* This notifier should be called after workqueue is ready */
412 hotcpu_notifier(apbt_cpuhp_notify, -20);
415 fs_initcall(apbt_late_init);
418 void apbt_setup_secondary_clock(void) {}
420 #endif /* CONFIG_SMP */
422 static void apbt_set_mode(enum clock_event_mode mode,
423 struct clock_event_device *evt)
428 struct apbt_dev *adev = EVT_TO_APBT_DEV(evt);
430 BUG_ON(!apbt_virt_address);
432 timer_num = adev->num;
433 pr_debug("%s CPU %d timer %d mode=%d\n",
434 __func__, first_cpu(*evt->cpumask), timer_num, mode);
437 case CLOCK_EVT_MODE_PERIODIC:
438 delta = ((uint64_t)(NSEC_PER_SEC/HZ)) * apbt_clockevent.mult;
439 delta >>= apbt_clockevent.shift;
440 ctrl = apbt_readl(timer_num, APBTMR_N_CONTROL);
441 ctrl |= APBTMR_CONTROL_MODE_PERIODIC;
442 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
444 * DW APB p. 46, have to disable timer before load counter,
445 * may cause sync problem.
447 ctrl &= ~APBTMR_CONTROL_ENABLE;
448 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
450 pr_debug("Setting clock period %d for HZ %d\n", (int)delta, HZ);
451 apbt_writel(timer_num, delta, APBTMR_N_LOAD_COUNT);
452 ctrl |= APBTMR_CONTROL_ENABLE;
453 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
455 /* APB timer does not have one-shot mode, use free running mode */
456 case CLOCK_EVT_MODE_ONESHOT:
457 ctrl = apbt_readl(timer_num, APBTMR_N_CONTROL);
459 * set free running mode, this mode will let timer reload max
460 * timeout which will give time (3min on 25MHz clock) to rearm
461 * the next event, therefore emulate the one-shot mode.
463 ctrl &= ~APBTMR_CONTROL_ENABLE;
464 ctrl &= ~APBTMR_CONTROL_MODE_PERIODIC;
466 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
467 /* write again to set free running mode */
468 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
471 * DW APB p. 46, load counter with all 1s before starting free
474 apbt_writel(timer_num, ~0, APBTMR_N_LOAD_COUNT);
475 ctrl &= ~APBTMR_CONTROL_INT;
476 ctrl |= APBTMR_CONTROL_ENABLE;
477 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
480 case CLOCK_EVT_MODE_UNUSED:
481 case CLOCK_EVT_MODE_SHUTDOWN:
482 apbt_disable_int(timer_num);
483 ctrl = apbt_readl(timer_num, APBTMR_N_CONTROL);
484 ctrl &= ~APBTMR_CONTROL_ENABLE;
485 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
488 case CLOCK_EVT_MODE_RESUME:
489 apbt_enable_int(timer_num);
494 static int apbt_next_event(unsigned long delta,
495 struct clock_event_device *evt)
500 struct apbt_dev *adev = EVT_TO_APBT_DEV(evt);
502 timer_num = adev->num;
504 ctrl = apbt_readl(timer_num, APBTMR_N_CONTROL);
505 ctrl &= ~APBTMR_CONTROL_ENABLE;
506 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
507 /* write new count */
508 apbt_writel(timer_num, delta, APBTMR_N_LOAD_COUNT);
509 ctrl |= APBTMR_CONTROL_ENABLE;
510 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
515 * APB timer clock is not in sync with pclk on Langwell, which translates to
516 * unreliable read value caused by sampling error. the error does not add up
517 * overtime and only happens when sampling a 0 as a 1 by mistake. so the time
518 * would go backwards. the following code is trying to prevent time traveling
519 * backwards. little bit paranoid.
521 static cycle_t apbt_read_clocksource(struct clocksource *cs)
523 unsigned long t0, t1, t2;
524 static unsigned long last_read;
527 t1 = apbt_readl(phy_cs_timer_id,
528 APBTMR_N_CURRENT_VALUE);
529 t2 = apbt_readl(phy_cs_timer_id,
530 APBTMR_N_CURRENT_VALUE);
531 if (unlikely(t1 < t2)) {
532 pr_debug("APBT: read current count error %lx:%lx:%lx\n",
537 * check against cached last read, makes sure time does not go back.
538 * it could be a normal rollover but we will do tripple check anyway
540 if (unlikely(t2 > last_read)) {
541 /* check if we have a normal rollover */
542 unsigned long raw_intr_status =
543 apbt_readl_reg(APBTMRS_RAW_INT_STATUS);
545 * cs timer interrupt is masked but raw intr bit is set if
546 * rollover occurs. then we read EOI reg to clear it.
548 if (raw_intr_status & (1 << phy_cs_timer_id)) {
549 apbt_readl(phy_cs_timer_id, APBTMR_N_EOI);
552 pr_debug("APB CS going back %lx:%lx:%lx ",
553 t2, last_read, t2 - last_read);
555 pr_debug(KERN_INFO "tripple check enforced\n");
556 t0 = apbt_readl(phy_cs_timer_id,
557 APBTMR_N_CURRENT_VALUE);
559 t1 = apbt_readl(phy_cs_timer_id,
560 APBTMR_N_CURRENT_VALUE);
562 t2 = apbt_readl(phy_cs_timer_id,
563 APBTMR_N_CURRENT_VALUE);
564 if ((t2 > t1) || (t1 > t0)) {
565 printk(KERN_ERR "Error: APB CS tripple check failed\n");
574 static int apbt_clocksource_register(void)
579 /* Start the counter, use timer 2 as source, timer 0/1 for event */
580 apbt_start_counter(phy_cs_timer_id);
582 /* Verify whether apbt counter works */
583 t1 = apbt_read_clocksource(&clocksource_apbt);
587 * We don't know the TSC frequency yet, but waiting for
588 * 200000 TSC cycles is safe:
595 } while ((now - start) < 200000UL);
597 /* APBT is the only always on clocksource, it has to work! */
598 if (t1 == apbt_read_clocksource(&clocksource_apbt))
599 panic("APBT counter not counting. APBT disabled\n");
602 * initialize and register APBT clocksource
603 * convert that to ns/clock cycle
604 * mult = (ns/c) * 2^APBT_SHIFT
606 clocksource_apbt.mult = div_sc(MSEC_PER_SEC,
607 (unsigned long) apbt_freq, APBT_SHIFT);
608 clocksource_register(&clocksource_apbt);
614 * Early setup the APBT timer, only use timer 0 for booting then switch to
615 * per CPU timer if possible.
616 * returns 1 if per cpu apbt is setup
617 * returns 0 if no per cpu apbt is chosen
618 * panic if set up failed, this is the only platform timer on Moorestown.
620 void __init apbt_time_init(void)
624 struct sfi_timer_table_entry *p_mtmr;
625 unsigned int percpu_timer;
626 struct apbt_dev *adev;
629 if (apb_timer_block_enabled)
632 if (apbt_virt_address) {
633 pr_debug("Found APBT version 0x%lx\n",\
634 apbt_readl_reg(APBTMRS_COMP_VERSION));
638 * Read the frequency and check for a sane value, for ESL model
639 * we extend the possible clock range to allow time scaling.
642 if (apbt_freq < APBT_MIN_FREQ || apbt_freq > APBT_MAX_FREQ) {
643 pr_debug("APBT has invalid freq 0x%llx\n", apbt_freq);
646 if (apbt_clocksource_register()) {
647 pr_debug("APBT has failed to register clocksource\n");
650 if (!apbt_clockevent_register())
651 apb_timer_block_enabled = 1;
653 pr_debug("APBT has failed to register clockevent\n");
657 /* kernel cmdline disable apb timer, so we will use lapic timers */
658 if (mrst_timer_options == MRST_TIMER_LAPIC_APBT) {
659 printk(KERN_INFO "apbt: disabled per cpu timer\n");
662 pr_debug("%s: %d CPUs online\n", __func__, num_online_cpus());
663 if (num_possible_cpus() <= sfi_mtimer_num) {
665 apbt_num_timers_used = num_possible_cpus();
668 apbt_num_timers_used = 1;
669 adev = &per_cpu(cpu_apbt_dev, 0);
670 adev->flags &= ~APBT_DEV_USED;
672 pr_debug("%s: %d APB timers used\n", __func__, apbt_num_timers_used);
674 /* here we set up per CPU timer data structure */
675 apbt_devs = kzalloc(sizeof(struct apbt_dev) * apbt_num_timers_used,
678 printk(KERN_ERR "Failed to allocate APB timer devices\n");
681 for (i = 0; i < apbt_num_timers_used; i++) {
682 adev = &per_cpu(cpu_apbt_dev, i);
685 p_mtmr = sfi_get_mtmr(i);
687 adev->tick = p_mtmr->freq_hz;
688 adev->irq = p_mtmr->irq;
690 printk(KERN_ERR "Failed to get timer for cpu %d\n", i);
692 sprintf(adev->name, "apbt%d", i);
699 apbt_clear_mapping();
700 apb_timer_block_enabled = 0;
701 panic("failed to enable APB timer\n");
704 static inline void apbt_disable(int n)
706 if (is_apbt_capable()) {
707 unsigned long ctrl = apbt_readl(n, APBTMR_N_CONTROL);
708 ctrl &= ~APBTMR_CONTROL_ENABLE;
709 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
713 /* called before apb_timer_enable, use early map */
714 unsigned long apbt_quick_calibrate()
719 unsigned long khz = 0;
723 apbt_start_counter(phy_cs_timer_id);
725 /* check if the timer can count down, otherwise return */
726 old = apbt_read_clocksource(&clocksource_apbt);
729 if (old != apbt_read_clocksource(&clocksource_apbt))
736 loop = (apbt_freq * 1000) << 4;
738 /* restart the timer to ensure it won't get to 0 in the calibration */
739 apbt_start_counter(phy_cs_timer_id);
741 old = apbt_read_clocksource(&clocksource_apbt);
744 t1 = __native_read_tsc();
747 new = apbt_read_clocksource(&clocksource_apbt);
750 t2 = __native_read_tsc();
753 if (unlikely(loop >> shift == 0)) {
755 "APBT TSC calibration failed, not enough resolution\n");
758 scale = (int)div_u64((t2 - t1), loop >> shift);
759 khz = (scale * apbt_freq * 1000) >> shift;
760 printk(KERN_INFO "TSC freq calculated by APB timer is %lu khz\n", khz);