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>
37 #include <linux/pci.h>
38 #include <linux/sfi.h>
39 #include <linux/interrupt.h>
40 #include <linux/cpu.h>
41 #include <linux/irq.h>
43 #include <asm/fixmap.h>
44 #include <asm/apb_timer.h>
46 #define APBT_MASK CLOCKSOURCE_MASK(32)
48 #define APBT_CLOCKEVENT_RATING 150
49 #define APBT_CLOCKSOURCE_RATING 250
50 #define APBT_MIN_DELTA_USEC 200
52 #define EVT_TO_APBT_DEV(evt) container_of(evt, struct apbt_dev, evt)
53 #define APBT_CLOCKEVENT0_NUM (0)
54 #define APBT_CLOCKEVENT1_NUM (1)
55 #define APBT_CLOCKSOURCE_NUM (2)
57 static unsigned long apbt_address;
58 static int apb_timer_block_enabled;
59 static void __iomem *apbt_virt_address;
60 static int phy_cs_timer_id;
63 * Common DW APB timer info
65 static uint64_t apbt_freq;
67 static void apbt_set_mode(enum clock_event_mode mode,
68 struct clock_event_device *evt);
69 static int apbt_next_event(unsigned long delta,
70 struct clock_event_device *evt);
71 static cycle_t apbt_read_clocksource(struct clocksource *cs);
72 static void apbt_restart_clocksource(void);
75 struct clock_event_device evt;
85 int disable_apbt_percpu __cpuinitdata;
88 static unsigned int apbt_num_timers_used;
89 static DEFINE_PER_CPU(struct apbt_dev, cpu_apbt_dev);
90 static struct apbt_dev *apbt_devs;
93 static inline unsigned long apbt_readl_reg(unsigned long a)
95 return readl(apbt_virt_address + a);
98 static inline void apbt_writel_reg(unsigned long d, unsigned long a)
100 writel(d, apbt_virt_address + a);
103 static inline unsigned long apbt_readl(int n, unsigned long a)
105 return readl(apbt_virt_address + a + n * APBTMRS_REG_SIZE);
108 static inline void apbt_writel(int n, unsigned long d, unsigned long a)
110 writel(d, apbt_virt_address + a + n * APBTMRS_REG_SIZE);
113 static inline void apbt_set_mapping(void)
115 struct sfi_timer_table_entry *mtmr;
117 if (apbt_virt_address) {
118 pr_debug("APBT base already mapped\n");
121 mtmr = sfi_get_mtmr(APBT_CLOCKEVENT0_NUM);
123 printk(KERN_ERR "Failed to get MTMR %d from SFI\n",
124 APBT_CLOCKEVENT0_NUM);
127 apbt_address = (unsigned long)mtmr->phys_addr;
129 printk(KERN_WARNING "No timer base from SFI, use default\n");
130 apbt_address = APBT_DEFAULT_BASE;
132 apbt_virt_address = ioremap_nocache(apbt_address, APBT_MMAP_SIZE);
133 if (apbt_virt_address) {
134 pr_debug("Mapped APBT physical addr %p at virtual addr %p\n",\
135 (void *)apbt_address, (void *)apbt_virt_address);
137 pr_debug("Failed mapping APBT phy address at %p\n",\
138 (void *)apbt_address);
141 apbt_freq = mtmr->freq_hz / USEC_PER_SEC;
144 /* Now figure out the physical timer id for clocksource device */
145 mtmr = sfi_get_mtmr(APBT_CLOCKSOURCE_NUM);
149 /* Now figure out the physical timer id */
150 phy_cs_timer_id = (unsigned int)(mtmr->phys_addr & 0xff)
152 pr_debug("Use timer %d for clocksource\n", phy_cs_timer_id);
156 panic("Failed to setup APB system timer\n");
160 static inline void apbt_clear_mapping(void)
162 iounmap(apbt_virt_address);
163 apbt_virt_address = NULL;
167 * APBT timer interrupt enable / disable
169 static inline int is_apbt_capable(void)
171 return apbt_virt_address ? 1 : 0;
174 static struct clocksource clocksource_apbt = {
176 .rating = APBT_CLOCKSOURCE_RATING,
177 .read = apbt_read_clocksource,
180 .flags = CLOCK_SOURCE_IS_CONTINUOUS,
181 .resume = apbt_restart_clocksource,
184 /* boot APB clock event device */
185 static struct clock_event_device apbt_clockevent = {
187 .features = CLOCK_EVT_FEAT_PERIODIC | CLOCK_EVT_FEAT_ONESHOT,
188 .set_mode = apbt_set_mode,
189 .set_next_event = apbt_next_event,
192 .rating = APBT_CLOCKEVENT_RATING,
196 * if user does not want to use per CPU apb timer, just give it a lower rating
197 * than local apic timer and skip the late per cpu timer init.
199 static inline int __init setup_x86_mrst_timer(char *arg)
204 if (strcmp("apbt_only", arg) == 0)
205 disable_apbt_percpu = 0;
206 else if (strcmp("lapic_and_apbt", arg) == 0)
207 disable_apbt_percpu = 1;
209 pr_warning("X86 MRST timer option %s not recognised"
210 " use x86_mrst_timer=apbt_only or lapic_and_apbt\n",
216 __setup("x86_mrst_timer=", setup_x86_mrst_timer);
219 * start count down from 0xffff_ffff. this is done by toggling the enable bit
220 * then load initial load count to ~0.
222 static void apbt_start_counter(int n)
224 unsigned long ctrl = apbt_readl(n, APBTMR_N_CONTROL);
226 ctrl &= ~APBTMR_CONTROL_ENABLE;
227 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
228 apbt_writel(n, ~0, APBTMR_N_LOAD_COUNT);
229 /* enable, mask interrupt */
230 ctrl &= ~APBTMR_CONTROL_MODE_PERIODIC;
231 ctrl |= (APBTMR_CONTROL_ENABLE | APBTMR_CONTROL_INT);
232 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
233 /* read it once to get cached counter value initialized */
234 apbt_read_clocksource(&clocksource_apbt);
237 static irqreturn_t apbt_interrupt_handler(int irq, void *data)
239 struct apbt_dev *dev = (struct apbt_dev *)data;
240 struct clock_event_device *aevt = &dev->evt;
242 if (!aevt->event_handler) {
243 printk(KERN_INFO "Spurious APBT timer interrupt on %d\n",
247 aevt->event_handler(aevt);
251 static void apbt_restart_clocksource(void)
253 apbt_start_counter(phy_cs_timer_id);
256 /* Setup IRQ routing via IOAPIC */
258 static void apbt_setup_irq(struct apbt_dev *adev)
260 struct irq_chip *chip;
261 struct irq_desc *desc;
263 /* timer0 irq has been setup early */
266 desc = irq_to_desc(adev->irq);
267 chip = get_irq_chip(adev->irq);
268 disable_irq(adev->irq);
269 desc->status |= IRQ_MOVE_PCNTXT;
270 irq_set_affinity(adev->irq, cpumask_of(adev->cpu));
271 /* APB timer irqs are set up as mp_irqs, timer is edge triggerred */
272 set_irq_chip_and_handler_name(adev->irq, chip, handle_edge_irq, "edge");
273 enable_irq(adev->irq);
274 if (system_state == SYSTEM_BOOTING)
275 if (request_irq(adev->irq, apbt_interrupt_handler,
276 IRQF_TIMER | IRQF_DISABLED | IRQF_NOBALANCING,
278 printk(KERN_ERR "Failed request IRQ for APBT%d\n",
284 static void apbt_enable_int(int n)
286 unsigned long ctrl = apbt_readl(n, APBTMR_N_CONTROL);
287 /* clear pending intr */
288 apbt_readl(n, APBTMR_N_EOI);
289 ctrl &= ~APBTMR_CONTROL_INT;
290 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
293 static void apbt_disable_int(int n)
295 unsigned long ctrl = apbt_readl(n, APBTMR_N_CONTROL);
297 ctrl |= APBTMR_CONTROL_INT;
298 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
302 static int __init apbt_clockevent_register(void)
304 struct sfi_timer_table_entry *mtmr;
306 mtmr = sfi_get_mtmr(APBT_CLOCKEVENT0_NUM);
308 printk(KERN_ERR "Failed to get MTMR %d from SFI\n",
309 APBT_CLOCKEVENT0_NUM);
314 * We need to calculate the scaled math multiplication factor for
315 * nanosecond to apbt tick conversion.
316 * mult = (nsec/cycle)*2^APBT_SHIFT
318 apbt_clockevent.mult = div_sc((unsigned long) mtmr->freq_hz
319 , NSEC_PER_SEC, APBT_SHIFT);
321 /* Calculate the min / max delta */
322 apbt_clockevent.max_delta_ns = clockevent_delta2ns(0x7FFFFFFF,
324 apbt_clockevent.min_delta_ns = clockevent_delta2ns(
325 APBT_MIN_DELTA_USEC*apbt_freq,
328 * Start apbt with the boot cpu mask and make it
329 * global if not used for per cpu timer.
331 apbt_clockevent.cpumask = cpumask_of(smp_processor_id());
333 if (disable_apbt_percpu) {
334 apbt_clockevent.rating = APBT_CLOCKEVENT_RATING - 100;
335 global_clock_event = &apbt_clockevent;
336 printk(KERN_DEBUG "%s clockevent registered as global\n",
337 global_clock_event->name);
340 if (request_irq(apbt_clockevent.irq, apbt_interrupt_handler,
341 IRQF_TIMER | IRQF_DISABLED | IRQF_NOBALANCING,
342 apbt_clockevent.name, &apbt_clockevent)) {
343 printk(KERN_ERR "Failed request IRQ for APBT%d\n",
344 apbt_clockevent.irq);
347 clockevents_register_device(&apbt_clockevent);
348 /* Start APBT 0 interrupts */
349 apbt_enable_int(APBT_CLOCKEVENT0_NUM);
356 /* Should be called with per cpu */
357 void apbt_setup_secondary_clock(void)
359 struct apbt_dev *adev;
360 struct clock_event_device *aevt;
363 /* Don't register boot CPU clockevent */
364 cpu = smp_processor_id();
365 if (cpu == boot_cpu_id)
368 * We need to calculate the scaled math multiplication factor for
369 * nanosecond to apbt tick conversion.
370 * mult = (nsec/cycle)*2^APBT_SHIFT
372 printk(KERN_INFO "Init per CPU clockevent %d\n", cpu);
373 adev = &per_cpu(cpu_apbt_dev, cpu);
376 memcpy(aevt, &apbt_clockevent, sizeof(*aevt));
377 aevt->cpumask = cpumask_of(cpu);
378 aevt->name = adev->name;
379 aevt->mode = CLOCK_EVT_MODE_UNUSED;
381 printk(KERN_INFO "Registering CPU %d clockevent device %s, mask %08x\n",
382 cpu, aevt->name, *(u32 *)aevt->cpumask);
384 apbt_setup_irq(adev);
386 clockevents_register_device(aevt);
388 apbt_enable_int(cpu);
394 * this notify handler process CPU hotplug events. in case of S0i3, nonboot
395 * cpus are disabled/enabled frequently, for performance reasons, we keep the
396 * per cpu timer irq registered so that we do need to do free_irq/request_irq.
398 * TODO: it might be more reliable to directly disable percpu clockevent device
399 * without the notifier chain. currently, cpu 0 may get interrupts from other
400 * cpu timers during the offline process due to the ordering of notification.
401 * the extra interrupt is harmless.
403 static int apbt_cpuhp_notify(struct notifier_block *n,
404 unsigned long action, void *hcpu)
406 unsigned long cpu = (unsigned long)hcpu;
407 struct apbt_dev *adev = &per_cpu(cpu_apbt_dev, cpu);
409 switch (action & 0xf) {
411 apbt_disable_int(cpu);
412 if (system_state == SYSTEM_RUNNING)
413 pr_debug("skipping APBT CPU %lu offline\n", cpu);
415 pr_debug("APBT clockevent for cpu %lu offline\n", cpu);
416 free_irq(adev->irq, adev);
420 pr_debug(KERN_INFO "APBT notified %lu, no action\n", action);
425 static __init int apbt_late_init(void)
427 if (disable_apbt_percpu)
429 /* This notifier should be called after workqueue is ready */
430 hotcpu_notifier(apbt_cpuhp_notify, -20);
433 fs_initcall(apbt_late_init);
436 void apbt_setup_secondary_clock(void) {}
438 #endif /* CONFIG_SMP */
440 static void apbt_set_mode(enum clock_event_mode mode,
441 struct clock_event_device *evt)
446 struct apbt_dev *adev = EVT_TO_APBT_DEV(evt);
448 timer_num = adev->num;
449 pr_debug("%s CPU %d timer %d mode=%d\n",
450 __func__, first_cpu(*evt->cpumask), timer_num, mode);
453 case CLOCK_EVT_MODE_PERIODIC:
454 delta = ((uint64_t)(NSEC_PER_SEC/HZ)) * apbt_clockevent.mult;
455 delta >>= apbt_clockevent.shift;
456 ctrl = apbt_readl(timer_num, APBTMR_N_CONTROL);
457 ctrl |= APBTMR_CONTROL_MODE_PERIODIC;
458 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
460 * DW APB p. 46, have to disable timer before load counter,
461 * may cause sync problem.
463 ctrl &= ~APBTMR_CONTROL_ENABLE;
464 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
466 pr_debug("Setting clock period %d for HZ %d\n", (int)delta, HZ);
467 apbt_writel(timer_num, delta, APBTMR_N_LOAD_COUNT);
468 ctrl |= APBTMR_CONTROL_ENABLE;
469 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
471 /* APB timer does not have one-shot mode, use free running mode */
472 case CLOCK_EVT_MODE_ONESHOT:
473 ctrl = apbt_readl(timer_num, APBTMR_N_CONTROL);
475 * set free running mode, this mode will let timer reload max
476 * timeout which will give time (3min on 25MHz clock) to rearm
477 * the next event, therefore emulate the one-shot mode.
479 ctrl &= ~APBTMR_CONTROL_ENABLE;
480 ctrl &= ~APBTMR_CONTROL_MODE_PERIODIC;
482 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
483 /* write again to set free running mode */
484 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
487 * DW APB p. 46, load counter with all 1s before starting free
490 apbt_writel(timer_num, ~0, APBTMR_N_LOAD_COUNT);
491 ctrl &= ~APBTMR_CONTROL_INT;
492 ctrl |= APBTMR_CONTROL_ENABLE;
493 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
496 case CLOCK_EVT_MODE_UNUSED:
497 case CLOCK_EVT_MODE_SHUTDOWN:
498 apbt_disable_int(timer_num);
499 ctrl = apbt_readl(timer_num, APBTMR_N_CONTROL);
500 ctrl &= ~APBTMR_CONTROL_ENABLE;
501 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
504 case CLOCK_EVT_MODE_RESUME:
505 apbt_enable_int(timer_num);
510 static int apbt_next_event(unsigned long delta,
511 struct clock_event_device *evt)
516 struct apbt_dev *adev = EVT_TO_APBT_DEV(evt);
518 timer_num = adev->num;
520 ctrl = apbt_readl(timer_num, APBTMR_N_CONTROL);
521 ctrl &= ~APBTMR_CONTROL_ENABLE;
522 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
523 /* write new count */
524 apbt_writel(timer_num, delta, APBTMR_N_LOAD_COUNT);
525 ctrl |= APBTMR_CONTROL_ENABLE;
526 apbt_writel(timer_num, ctrl, APBTMR_N_CONTROL);
531 * APB timer clock is not in sync with pclk on Langwell, which translates to
532 * unreliable read value caused by sampling error. the error does not add up
533 * overtime and only happens when sampling a 0 as a 1 by mistake. so the time
534 * would go backwards. the following code is trying to prevent time traveling
535 * backwards. little bit paranoid.
537 static cycle_t apbt_read_clocksource(struct clocksource *cs)
539 unsigned long t0, t1, t2;
540 static unsigned long last_read;
543 t1 = apbt_readl(phy_cs_timer_id,
544 APBTMR_N_CURRENT_VALUE);
545 t2 = apbt_readl(phy_cs_timer_id,
546 APBTMR_N_CURRENT_VALUE);
547 if (unlikely(t1 < t2)) {
548 pr_debug("APBT: read current count error %lx:%lx:%lx\n",
553 * check against cached last read, makes sure time does not go back.
554 * it could be a normal rollover but we will do tripple check anyway
556 if (unlikely(t2 > last_read)) {
557 /* check if we have a normal rollover */
558 unsigned long raw_intr_status =
559 apbt_readl_reg(APBTMRS_RAW_INT_STATUS);
561 * cs timer interrupt is masked but raw intr bit is set if
562 * rollover occurs. then we read EOI reg to clear it.
564 if (raw_intr_status & (1 << phy_cs_timer_id)) {
565 apbt_readl(phy_cs_timer_id, APBTMR_N_EOI);
568 pr_debug("APB CS going back %lx:%lx:%lx ",
569 t2, last_read, t2 - last_read);
571 pr_debug(KERN_INFO "tripple check enforced\n");
572 t0 = apbt_readl(phy_cs_timer_id,
573 APBTMR_N_CURRENT_VALUE);
575 t1 = apbt_readl(phy_cs_timer_id,
576 APBTMR_N_CURRENT_VALUE);
578 t2 = apbt_readl(phy_cs_timer_id,
579 APBTMR_N_CURRENT_VALUE);
580 if ((t2 > t1) || (t1 > t0)) {
581 printk(KERN_ERR "Error: APB CS tripple check failed\n");
590 static int apbt_clocksource_register(void)
595 /* Start the counter, use timer 2 as source, timer 0/1 for event */
596 apbt_start_counter(phy_cs_timer_id);
598 /* Verify whether apbt counter works */
599 t1 = apbt_read_clocksource(&clocksource_apbt);
603 * We don't know the TSC frequency yet, but waiting for
604 * 200000 TSC cycles is safe:
611 } while ((now - start) < 200000UL);
613 /* APBT is the only always on clocksource, it has to work! */
614 if (t1 == apbt_read_clocksource(&clocksource_apbt))
615 panic("APBT counter not counting. APBT disabled\n");
618 * initialize and register APBT clocksource
619 * convert that to ns/clock cycle
620 * mult = (ns/c) * 2^APBT_SHIFT
622 clocksource_apbt.mult = div_sc(MSEC_PER_SEC,
623 (unsigned long) apbt_freq, APBT_SHIFT);
624 clocksource_register(&clocksource_apbt);
630 * Early setup the APBT timer, only use timer 0 for booting then switch to
631 * per CPU timer if possible.
632 * returns 1 if per cpu apbt is setup
633 * returns 0 if no per cpu apbt is chosen
634 * panic if set up failed, this is the only platform timer on Moorestown.
636 void __init apbt_time_init(void)
640 struct sfi_timer_table_entry *p_mtmr;
641 unsigned int percpu_timer;
642 struct apbt_dev *adev;
645 if (apb_timer_block_enabled)
648 if (apbt_virt_address) {
649 pr_debug("Found APBT version 0x%lx\n",\
650 apbt_readl_reg(APBTMRS_COMP_VERSION));
654 * Read the frequency and check for a sane value, for ESL model
655 * we extend the possible clock range to allow time scaling.
658 if (apbt_freq < APBT_MIN_FREQ || apbt_freq > APBT_MAX_FREQ) {
659 pr_debug("APBT has invalid freq 0x%llx\n", apbt_freq);
662 if (apbt_clocksource_register()) {
663 pr_debug("APBT has failed to register clocksource\n");
666 if (!apbt_clockevent_register())
667 apb_timer_block_enabled = 1;
669 pr_debug("APBT has failed to register clockevent\n");
673 /* kernel cmdline disable apb timer, so we will use lapic timers */
674 if (disable_apbt_percpu) {
675 printk(KERN_INFO "apbt: disabled per cpu timer\n");
678 pr_debug("%s: %d CPUs online\n", __func__, num_online_cpus());
679 if (num_possible_cpus() <= sfi_mtimer_num) {
681 apbt_num_timers_used = num_possible_cpus();
684 apbt_num_timers_used = 1;
685 adev = &per_cpu(cpu_apbt_dev, 0);
686 adev->flags &= ~APBT_DEV_USED;
688 pr_debug("%s: %d APB timers used\n", __func__, apbt_num_timers_used);
690 /* here we set up per CPU timer data structure */
691 apbt_devs = kzalloc(sizeof(struct apbt_dev) * apbt_num_timers_used,
694 printk(KERN_ERR "Failed to allocate APB timer devices\n");
697 for (i = 0; i < apbt_num_timers_used; i++) {
698 adev = &per_cpu(cpu_apbt_dev, i);
701 p_mtmr = sfi_get_mtmr(i);
703 adev->tick = p_mtmr->freq_hz;
704 adev->irq = p_mtmr->irq;
706 printk(KERN_ERR "Failed to get timer for cpu %d\n", i);
708 sprintf(adev->name, "apbt%d", i);
715 apbt_clear_mapping();
716 apb_timer_block_enabled = 0;
717 panic("failed to enable APB timer\n");
720 static inline void apbt_disable(int n)
722 if (is_apbt_capable()) {
723 unsigned long ctrl = apbt_readl(n, APBTMR_N_CONTROL);
724 ctrl &= ~APBTMR_CONTROL_ENABLE;
725 apbt_writel(n, ctrl, APBTMR_N_CONTROL);
729 /* called before apb_timer_enable, use early map */
730 unsigned long apbt_quick_calibrate()
735 unsigned long khz = 0;
739 apbt_start_counter(phy_cs_timer_id);
741 /* check if the timer can count down, otherwise return */
742 old = apbt_read_clocksource(&clocksource_apbt);
745 if (old != apbt_read_clocksource(&clocksource_apbt))
752 loop = (apbt_freq * 1000) << 4;
754 /* restart the timer to ensure it won't get to 0 in the calibration */
755 apbt_start_counter(phy_cs_timer_id);
757 old = apbt_read_clocksource(&clocksource_apbt);
760 t1 = __native_read_tsc();
763 new = apbt_read_clocksource(&clocksource_apbt);
766 t2 = __native_read_tsc();
769 if (unlikely(loop >> shift == 0)) {
771 "APBT TSC calibration failed, not enough resolution\n");
774 scale = (int)div_u64((t2 - t1), loop >> shift);
775 khz = (scale * apbt_freq * 1000) >> shift;
776 printk(KERN_INFO "TSC freq calculated by APB timer is %lu khz\n", khz);