- Stephen Rothwell: APM updates
[davej-history.git] / arch / i386 / kernel / apic.c
blob4179c8f13b3674e96c7a2cf61e8b6e3225831735
1 /*
2 * Local APIC handling, local APIC timers
4 * (c) 1999, 2000 Ingo Molnar <mingo@redhat.com>
6 * Fixes
7 * Maciej W. Rozycki : Bits for genuine 82489DX APICs;
8 * thanks to Eric Gilmore
9 * and Rolf G. Tews
10 * for testing these extensively.
11 * Maciej W. Rozycki : Various updates and fixes.
14 #include <linux/config.h>
15 #include <linux/init.h>
17 #include <linux/mm.h>
18 #include <linux/irq.h>
19 #include <linux/delay.h>
20 #include <linux/bootmem.h>
21 #include <linux/smp_lock.h>
22 #include <linux/interrupt.h>
23 #include <linux/mc146818rtc.h>
24 #include <linux/kernel_stat.h>
26 #include <asm/smp.h>
27 #include <asm/mtrr.h>
28 #include <asm/mpspec.h>
29 #include <asm/pgalloc.h>
31 int prof_multiplier[NR_CPUS] = { 1, };
32 int prof_old_multiplier[NR_CPUS] = { 1, };
33 int prof_counter[NR_CPUS] = { 1, };
35 int get_maxlvt(void)
37 unsigned int v, ver, maxlvt;
39 v = apic_read(APIC_LVR);
40 ver = GET_APIC_VERSION(v);
41 /* 82489DXs do not report # of LVT entries. */
42 maxlvt = APIC_INTEGRATED(ver) ? GET_APIC_MAXLVT(v) : 2;
43 return maxlvt;
46 static void clear_local_APIC(void)
48 int maxlvt;
49 unsigned long v;
51 maxlvt = get_maxlvt();
54 * Careful: we have to set masks only first to deassert
55 * any level-triggered sources.
57 v = apic_read(APIC_LVTT);
58 apic_write_around(APIC_LVTT, v | APIC_LVT_MASKED);
59 v = apic_read(APIC_LVT0);
60 apic_write_around(APIC_LVT0, v | APIC_LVT_MASKED);
61 v = apic_read(APIC_LVT1);
62 apic_write_around(APIC_LVT1, v | APIC_LVT_MASKED);
63 if (maxlvt >= 3) {
64 v = apic_read(APIC_LVTERR);
65 apic_write_around(APIC_LVTERR, v | APIC_LVT_MASKED);
67 if (maxlvt >= 4) {
68 v = apic_read(APIC_LVTPC);
69 apic_write_around(APIC_LVTPC, v | APIC_LVT_MASKED);
73 * Clean APIC state for other OSs:
75 apic_write_around(APIC_LVTT, APIC_LVT_MASKED);
76 apic_write_around(APIC_LVT0, APIC_LVT_MASKED);
77 apic_write_around(APIC_LVT1, APIC_LVT_MASKED);
78 if (maxlvt >= 3)
79 apic_write_around(APIC_LVTERR, APIC_LVT_MASKED);
80 if (maxlvt >= 4)
81 apic_write_around(APIC_LVTPC, APIC_LVT_MASKED);
84 void __init connect_bsp_APIC(void)
86 if (pic_mode) {
88 * Do not trust the local APIC being empty at bootup.
90 clear_local_APIC();
92 * PIC mode, enable symmetric IO mode in the IMCR,
93 * i.e. connect BSP's local APIC to INT and NMI lines.
95 printk("leaving PIC mode, enabling symmetric IO mode.\n");
96 outb(0x70, 0x22);
97 outb(0x01, 0x23);
101 void disconnect_bsp_APIC(void)
103 if (pic_mode) {
105 * Put the board back into PIC mode (has an effect
106 * only on certain older boards). Note that APIC
107 * interrupts, including IPIs, won't work beyond
108 * this point! The only exception are INIT IPIs.
110 printk("disabling symmetric IO mode, entering PIC mode.\n");
111 outb(0x70, 0x22);
112 outb(0x00, 0x23);
116 void disable_local_APIC(void)
118 unsigned long value;
120 clear_local_APIC();
123 * Disable APIC (implies clearing of registers
124 * for 82489DX!).
126 value = apic_read(APIC_SPIV);
127 value &= ~(1<<8);
128 apic_write_around(APIC_SPIV, value);
132 * This is to verify that we're looking at a real local APIC.
133 * Check these against your board if the CPUs aren't getting
134 * started for no apparent reason.
136 int __init verify_local_APIC(void)
138 unsigned int reg0, reg1;
141 * The version register is read-only in a real APIC.
143 reg0 = apic_read(APIC_LVR);
144 Dprintk("Getting VERSION: %x\n", reg0);
145 apic_write(APIC_LVR, reg0 ^ APIC_LVR_MASK);
146 reg1 = apic_read(APIC_LVR);
147 Dprintk("Getting VERSION: %x\n", reg1);
150 * The two version reads above should print the same
151 * numbers. If the second one is different, then we
152 * poke at a non-APIC.
154 if (reg1 != reg0)
155 return 0;
158 * Check if the version looks reasonably.
160 reg1 = GET_APIC_VERSION(reg0);
161 if (reg1 == 0x00 || reg1 == 0xff)
162 return 0;
163 reg1 = get_maxlvt();
164 if (reg1 < 0x02 || reg1 == 0xff)
165 return 0;
168 * The ID register is read/write in a real APIC.
170 reg0 = apic_read(APIC_ID);
171 Dprintk("Getting ID: %x\n", reg0);
172 apic_write(APIC_ID, reg0 ^ APIC_ID_MASK);
173 reg1 = apic_read(APIC_ID);
174 Dprintk("Getting ID: %x\n", reg1);
175 apic_write(APIC_ID, reg0);
176 if (reg1 != (reg0 ^ APIC_ID_MASK))
177 return 0;
180 * The next two are just to see if we have sane values.
181 * They're only really relevant if we're in Virtual Wire
182 * compatibility mode, but most boxes are anymore.
184 reg0 = apic_read(APIC_LVT0);
185 Dprintk("Getting LVT0: %x\n", reg0);
186 reg1 = apic_read(APIC_LVT1);
187 Dprintk("Getting LVT1: %x\n", reg1);
189 return 1;
192 void __init sync_Arb_IDs(void)
195 * Wait for idle.
197 apic_wait_icr_idle();
199 Dprintk("Synchronizing Arb IDs.\n");
200 apic_write_around(APIC_ICR, APIC_DEST_ALLINC | APIC_INT_LEVELTRIG
201 | APIC_DM_INIT);
204 extern void __error_in_apic_c (void);
206 void __init setup_local_APIC (void)
208 unsigned long value, ver, maxlvt;
210 value = apic_read(APIC_LVR);
211 ver = GET_APIC_VERSION(value);
213 if ((SPURIOUS_APIC_VECTOR & 0x0f) != 0x0f)
214 __error_in_apic_c();
217 * Double-check wether this APIC is really registered.
219 if (!test_bit(GET_APIC_ID(apic_read(APIC_ID)), &phys_cpu_present_map))
220 BUG();
223 * Intel recommends to set DFR, LDR and TPR before enabling
224 * an APIC. See e.g. "AP-388 82489DX User's Manual" (Intel
225 * document number 292116). So here it goes...
229 * Put the APIC into flat delivery mode.
230 * Must be "all ones" explicitly for 82489DX.
232 apic_write_around(APIC_DFR, 0xffffffff);
235 * Set up the logical destination ID.
237 value = apic_read(APIC_LDR);
238 value &= ~APIC_LDR_MASK;
239 value |= (1<<(smp_processor_id()+24));
240 apic_write_around(APIC_LDR, value);
243 * Set Task Priority to 'accept all'. We never change this
244 * later on.
246 value = apic_read(APIC_TASKPRI);
247 value &= ~APIC_TPRI_MASK;
248 apic_write_around(APIC_TASKPRI, value);
251 * Now that we are all set up, enable the APIC
253 value = apic_read(APIC_SPIV);
254 value &= ~APIC_VECTOR_MASK;
256 * Enable APIC
258 value |= (1<<8);
261 * Some unknown Intel IO/APIC (or APIC) errata is biting us with
262 * certain networking cards. If high frequency interrupts are
263 * happening on a particular IOAPIC pin, plus the IOAPIC routing
264 * entry is masked/unmasked at a high rate as well then sooner or
265 * later IOAPIC line gets 'stuck', no more interrupts are received
266 * from the device. If focus CPU is disabled then the hang goes
267 * away, oh well :-(
269 * [ This bug can be reproduced easily with a level-triggered
270 * PCI Ne2000 networking cards and PII/PIII processors, dual
271 * BX chipset. ]
273 #if 0
274 /* Enable focus processor (bit==0) */
275 value &= ~(1<<9);
276 #else
277 /* Disable focus processor (bit==1) */
278 value |= (1<<9);
279 #endif
281 * Set spurious IRQ vector
283 value |= SPURIOUS_APIC_VECTOR;
284 apic_write_around(APIC_SPIV, value);
287 * Set up LVT0, LVT1:
289 * set up through-local-APIC on the BP's LINT0. This is not
290 * strictly necessery in pure symmetric-IO mode, but sometimes
291 * we delegate interrupts to the 8259A.
294 * TODO: set up through-local-APIC from through-I/O-APIC? --macro
296 value = apic_read(APIC_LVT0) & APIC_LVT_MASKED;
297 if (!smp_processor_id() && (pic_mode || !value)) {
298 value = APIC_DM_EXTINT;
299 printk("enabled ExtINT on CPU#%d\n", smp_processor_id());
300 } else {
301 value = APIC_DM_EXTINT | APIC_LVT_MASKED;
302 printk("masked ExtINT on CPU#%d\n", smp_processor_id());
304 apic_write_around(APIC_LVT0, value);
307 * only the BP should see the LINT1 NMI signal, obviously.
309 if (!smp_processor_id())
310 value = APIC_DM_NMI;
311 else
312 value = APIC_DM_NMI | APIC_LVT_MASKED;
313 if (!APIC_INTEGRATED(ver)) /* 82489DX */
314 value |= APIC_LVT_LEVEL_TRIGGER;
315 apic_write_around(APIC_LVT1, value);
317 if (APIC_INTEGRATED(ver)) { /* !82489DX */
318 maxlvt = get_maxlvt();
319 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
320 apic_write(APIC_ESR, 0);
321 value = apic_read(APIC_ESR);
322 printk("ESR value before enabling vector: %08lx\n", value);
324 value = ERROR_APIC_VECTOR; // enables sending errors
325 apic_write_around(APIC_LVTERR, value);
327 * spec says clear errors after enabling vector.
329 if (maxlvt > 3)
330 apic_write(APIC_ESR, 0);
331 value = apic_read(APIC_ESR);
332 printk("ESR value after enabling vector: %08lx\n", value);
333 } else
334 printk("No ESR for 82489DX.\n");
337 void __init init_apic_mappings(void)
339 unsigned long apic_phys;
341 if (smp_found_config) {
342 apic_phys = mp_lapic_addr;
343 } else {
345 * set up a fake all zeroes page to simulate the
346 * local APIC and another one for the IO-APIC. We
347 * could use the real zero-page, but it's safer
348 * this way if some buggy code writes to this page ...
350 apic_phys = (unsigned long) alloc_bootmem_pages(PAGE_SIZE);
351 apic_phys = __pa(apic_phys);
353 set_fixmap_nocache(FIX_APIC_BASE, apic_phys);
354 Dprintk("mapped APIC to %08lx (%08lx)\n", APIC_BASE, apic_phys);
357 * Fetch the APIC ID of the BSP in case we have a
358 * default configuration (or the MP table is broken).
360 if (boot_cpu_id == -1U)
361 boot_cpu_id = GET_APIC_ID(apic_read(APIC_ID));
363 #ifdef CONFIG_X86_IO_APIC
365 unsigned long ioapic_phys, idx = FIX_IO_APIC_BASE_0;
366 int i;
368 for (i = 0; i < nr_ioapics; i++) {
369 if (smp_found_config) {
370 ioapic_phys = mp_ioapics[i].mpc_apicaddr;
371 } else {
372 ioapic_phys = (unsigned long) alloc_bootmem_pages(PAGE_SIZE);
373 ioapic_phys = __pa(ioapic_phys);
375 set_fixmap_nocache(idx, ioapic_phys);
376 Dprintk("mapped IOAPIC to %08lx (%08lx)\n",
377 __fix_to_virt(idx), ioapic_phys);
378 idx++;
381 #endif
385 * This part sets up the APIC 32 bit clock in LVTT1, with HZ interrupts
386 * per second. We assume that the caller has already set up the local
387 * APIC.
389 * The APIC timer is not exactly sync with the external timer chip, it
390 * closely follows bus clocks.
394 * The timer chip is already set up at HZ interrupts per second here,
395 * but we do not accept timer interrupts yet. We only allow the BP
396 * to calibrate.
398 static unsigned int __init get_8254_timer_count(void)
400 extern spinlock_t i8253_lock;
401 unsigned long flags;
403 unsigned int count;
405 spin_lock_irqsave(&i8253_lock, flags);
407 outb_p(0x00, 0x43);
408 count = inb_p(0x40);
409 count |= inb_p(0x40) << 8;
411 spin_unlock_irqrestore(&i8253_lock, flags);
413 return count;
416 void __init wait_8254_wraparound(void)
418 unsigned int curr_count, prev_count=~0;
419 int delta;
421 curr_count = get_8254_timer_count();
423 do {
424 prev_count = curr_count;
425 curr_count = get_8254_timer_count();
426 delta = curr_count-prev_count;
429 * This limit for delta seems arbitrary, but it isn't, it's
430 * slightly above the level of error a buggy Mercury/Neptune
431 * chipset timer can cause.
434 } while (delta < 300);
438 * This function sets up the local APIC timer, with a timeout of
439 * 'clocks' APIC bus clock. During calibration we actually call
440 * this function twice on the boot CPU, once with a bogus timeout
441 * value, second time for real. The other (noncalibrating) CPUs
442 * call this function only once, with the real, calibrated value.
444 * We do reads before writes even if unnecessary, to get around the
445 * P5 APIC double write bug.
448 #define APIC_DIVISOR 16
450 void __setup_APIC_LVTT(unsigned int clocks)
452 unsigned int lvtt1_value, tmp_value;
454 lvtt1_value = SET_APIC_TIMER_BASE(APIC_TIMER_BASE_DIV) |
455 APIC_LVT_TIMER_PERIODIC | LOCAL_TIMER_VECTOR;
456 apic_write_around(APIC_LVTT, lvtt1_value);
459 * Divide PICLK by 16
461 tmp_value = apic_read(APIC_TDCR);
462 apic_write_around(APIC_TDCR, (tmp_value
463 & ~(APIC_TDR_DIV_1 | APIC_TDR_DIV_TMBASE))
464 | APIC_TDR_DIV_16);
466 apic_write_around(APIC_TMICT, clocks/APIC_DIVISOR);
469 void setup_APIC_timer(void * data)
471 unsigned int clocks = (unsigned int) data, slice, t0, t1;
472 unsigned long flags;
473 int delta;
475 __save_flags(flags);
476 __sti();
478 * ok, Intel has some smart code in their APIC that knows
479 * if a CPU was in 'hlt' lowpower mode, and this increases
480 * its APIC arbitration priority. To avoid the external timer
481 * IRQ APIC event being in synchron with the APIC clock we
482 * introduce an interrupt skew to spread out timer events.
484 * The number of slices within a 'big' timeslice is smp_num_cpus+1
487 slice = clocks / (smp_num_cpus+1);
488 printk("cpu: %d, clocks: %d, slice: %d\n",
489 smp_processor_id(), clocks, slice);
492 * Wait for IRQ0's slice:
494 wait_8254_wraparound();
496 __setup_APIC_LVTT(clocks);
498 t0 = apic_read(APIC_TMICT)*APIC_DIVISOR;
499 /* Wait till TMCCT gets reloaded from TMICT... */
500 do {
501 t1 = apic_read(APIC_TMCCT)*APIC_DIVISOR;
502 delta = (int)(t0 - t1 - slice*(smp_processor_id()+1));
503 } while (delta >= 0);
504 /* Now wait for our slice for real. */
505 do {
506 t1 = apic_read(APIC_TMCCT)*APIC_DIVISOR;
507 delta = (int)(t0 - t1 - slice*(smp_processor_id()+1));
508 } while (delta < 0);
510 __setup_APIC_LVTT(clocks);
512 printk("CPU%d<T0:%d,T1:%d,D:%d,S:%d,C:%d>\n",
513 smp_processor_id(), t0, t1, delta, slice, clocks);
515 __restore_flags(flags);
519 * In this function we calibrate APIC bus clocks to the external
520 * timer. Unfortunately we cannot use jiffies and the timer irq
521 * to calibrate, since some later bootup code depends on getting
522 * the first irq? Ugh.
524 * We want to do the calibration only once since we
525 * want to have local timer irqs syncron. CPUs connected
526 * by the same APIC bus have the very same bus frequency.
527 * And we want to have irqs off anyways, no accidental
528 * APIC irq that way.
531 int __init calibrate_APIC_clock(void)
533 unsigned long long t1 = 0, t2 = 0;
534 long tt1, tt2;
535 long result;
536 int i;
537 const int LOOPS = HZ/10;
539 printk("calibrating APIC timer ...\n");
542 * Put whatever arbitrary (but long enough) timeout
543 * value into the APIC clock, we just want to get the
544 * counter running for calibration.
546 __setup_APIC_LVTT(1000000000);
549 * The timer chip counts down to zero. Let's wait
550 * for a wraparound to start exact measurement:
551 * (the current tick might have been already half done)
554 wait_8254_wraparound();
557 * We wrapped around just now. Let's start:
559 if (cpu_has_tsc)
560 rdtscll(t1);
561 tt1 = apic_read(APIC_TMCCT);
564 * Let's wait LOOPS wraprounds:
566 for (i = 0; i < LOOPS; i++)
567 wait_8254_wraparound();
569 tt2 = apic_read(APIC_TMCCT);
570 if (cpu_has_tsc)
571 rdtscll(t2);
574 * The APIC bus clock counter is 32 bits only, it
575 * might have overflown, but note that we use signed
576 * longs, thus no extra care needed.
578 * underflown to be exact, as the timer counts down ;)
581 result = (tt1-tt2)*APIC_DIVISOR/LOOPS;
583 if (cpu_has_tsc)
584 printk("..... CPU clock speed is %ld.%04ld MHz.\n",
585 ((long)(t2-t1)/LOOPS)/(1000000/HZ),
586 ((long)(t2-t1)/LOOPS)%(1000000/HZ));
588 printk("..... host bus clock speed is %ld.%04ld MHz.\n",
589 result/(1000000/HZ),
590 result%(1000000/HZ));
592 return result;
595 static unsigned int calibration_result;
597 void __init setup_APIC_clocks (void)
599 __cli();
601 calibration_result = calibrate_APIC_clock();
603 * Now set up the timer for real.
605 setup_APIC_timer((void *)calibration_result);
607 __sti();
609 /* and update all other cpus */
610 smp_call_function(setup_APIC_timer, (void *)calibration_result, 1, 1);
614 * the frequency of the profiling timer can be changed
615 * by writing a multiplier value into /proc/profile.
617 int setup_profiling_timer(unsigned int multiplier)
619 int i;
622 * Sanity check. [at least 500 APIC cycles should be
623 * between APIC interrupts as a rule of thumb, to avoid
624 * irqs flooding us]
626 if ( (!multiplier) || (calibration_result/multiplier < 500))
627 return -EINVAL;
630 * Set the new multiplier for each CPU. CPUs don't start using the
631 * new values until the next timer interrupt in which they do process
632 * accounting. At that time they also adjust their APIC timers
633 * accordingly.
635 for (i = 0; i < NR_CPUS; ++i)
636 prof_multiplier[i] = multiplier;
638 return 0;
641 #undef APIC_DIVISOR
644 * Local timer interrupt handler. It does both profiling and
645 * process statistics/rescheduling.
647 * We do profiling in every local tick, statistics/rescheduling
648 * happen only every 'profiling multiplier' ticks. The default
649 * multiplier is 1 and it can be changed by writing the new multiplier
650 * value into /proc/profile.
653 inline void smp_local_timer_interrupt(struct pt_regs * regs)
655 int user = user_mode(regs);
656 int cpu = smp_processor_id();
659 * The profiling function is SMP safe. (nothing can mess
660 * around with "current", and the profiling counters are
661 * updated with atomic operations). This is especially
662 * useful with a profiling multiplier != 1
664 if (!user)
665 x86_do_profile(regs->eip);
667 if (--prof_counter[cpu] <= 0) {
669 * The multiplier may have changed since the last time we got
670 * to this point as a result of the user writing to
671 * /proc/profile. In this case we need to adjust the APIC
672 * timer accordingly.
674 * Interrupts are already masked off at this point.
676 prof_counter[cpu] = prof_multiplier[cpu];
677 if (prof_counter[cpu] != prof_old_multiplier[cpu]) {
678 __setup_APIC_LVTT(calibration_result/prof_counter[cpu]);
679 prof_old_multiplier[cpu] = prof_counter[cpu];
682 #ifdef CONFIG_SMP
683 update_process_times(user);
684 #endif
688 * We take the 'long' return path, and there every subsystem
689 * grabs the apropriate locks (kernel lock/ irq lock).
691 * we might want to decouple profiling from the 'long path',
692 * and do the profiling totally in assembly.
694 * Currently this isn't too much of an issue (performance wise),
695 * we can take more than 100K local irqs per second on a 100 MHz P5.
700 * Local APIC timer interrupt. This is the most natural way for doing
701 * local interrupts, but local timer interrupts can be emulated by
702 * broadcast interrupts too. [in case the hw doesnt support APIC timers]
704 * [ if a single-CPU system runs an SMP kernel then we call the local
705 * interrupt as well. Thus we cannot inline the local irq ... ]
707 unsigned int apic_timer_irqs [NR_CPUS];
709 void smp_apic_timer_interrupt(struct pt_regs * regs)
711 int cpu = smp_processor_id();
714 * the NMI deadlock-detector uses this.
716 apic_timer_irqs[cpu]++;
719 * NOTE! We'd better ACK the irq immediately,
720 * because timer handling can be slow.
722 ack_APIC_irq();
724 * update_process_times() expects us to have done irq_enter().
725 * Besides, if we don't timer interrupts ignore the global
726 * interrupt lock, which is the WrongThing (tm) to do.
728 irq_enter(cpu, 0);
729 smp_local_timer_interrupt(regs);
730 irq_exit(cpu, 0);
734 * This interrupt should _never_ happen with our APIC/SMP architecture
736 asmlinkage void smp_spurious_interrupt(void)
738 unsigned long v;
741 * Check if this really is a spurious interrupt and ACK it
742 * if it is a vectored one. Just in case...
743 * Spurious interrupts should not be ACKed.
745 v = apic_read(APIC_ISR + ((SPURIOUS_APIC_VECTOR & ~0x1f) >> 1));
746 if (v & (1 << (SPURIOUS_APIC_VECTOR & 0x1f)))
747 ack_APIC_irq();
749 /* see sw-dev-man vol 3, chapter 7.4.13.5 */
750 printk(KERN_INFO "spurious APIC interrupt on CPU#%d, should never happen.\n",
751 smp_processor_id());
755 * This interrupt should never happen with our APIC/SMP architecture
758 asmlinkage void smp_error_interrupt(void)
760 unsigned long v, v1;
762 /* First tickle the hardware, only then report what went on. -- REW */
763 v = apic_read(APIC_ESR);
764 apic_write(APIC_ESR, 0);
765 v1 = apic_read(APIC_ESR);
766 ack_APIC_irq();
767 irq_err_count++;
769 /* Here is what the APIC error bits mean:
770 0: Send CS error
771 1: Receive CS error
772 2: Send accept error
773 3: Receive accept error
774 4: Reserved
775 5: Send illegal vector
776 6: Received illegal vector
777 7: Illegal register address
779 printk (KERN_ERR "APIC error on CPU%d: %02lx(%02lx)\n",
780 smp_processor_id(), v , v1);