2 * Copyright (c) 1996, by Steve Passe
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. The name of the developer may NOT be used to endorse or promote products
11 * derived from this software without specific prior written permission.
13 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25 * $FreeBSD: src/sys/i386/i386/mpapic.c,v 1.37.2.7 2003/01/25 02:31:47 peter Exp $
28 #include <sys/param.h>
29 #include <sys/systm.h>
30 #include <sys/kernel.h>
33 #include <sys/machintr.h>
34 #include <machine/globaldata.h>
35 #include <machine/clock.h>
36 #include <machine/limits.h>
37 #include <machine/smp.h>
38 #include <machine/md_var.h>
39 #include <machine/pmap.h>
40 #include <machine/specialreg.h>
41 #include <machine_base/apic/lapic.h>
42 #include <machine_base/apic/ioapic.h>
43 #include <machine_base/apic/ioapic_abi.h>
44 #include <machine_base/apic/apicvar.h>
45 #include <machine_base/icu/icu_var.h>
46 #include <machine/segments.h>
47 #include <sys/thread2.h>
48 #include <sys/spinlock2.h>
50 #include <machine/cputypes.h>
51 #include <machine/intr_machdep.h>
53 #if !defined(KTR_LAPIC)
54 #define KTR_LAPIC KTR_ALL
56 KTR_INFO_MASTER(lapic);
57 KTR_INFO(KTR_LAPIC, lapic, mem_eoi, 0, "mem_eoi");
58 #define log_lapic(name) KTR_LOG(lapic_ ## name)
62 volatile lapic_t *lapic_mem;
64 static void lapic_timer_calibrate(void);
65 static void lapic_timer_set_divisor(int);
66 static void lapic_timer_fixup_handler(void *);
67 static void lapic_timer_restart_handler(void *);
70 static int lapic_timer_enable = 1;
71 TUNABLE_INT("hw.lapic_timer_enable", &lapic_timer_enable);
73 static int lapic_timer_tscdeadline = 1;
74 TUNABLE_INT("hw.lapic_timer_tscdeadline", &lapic_timer_tscdeadline);
76 static int lapic_calibrate_test = 0;
77 TUNABLE_INT("hw.lapic_calibrate_test", &lapic_calibrate_test);
79 static int lapic_calibrate_fast = 1;
80 TUNABLE_INT("hw.lapic_calibrate_fast", &lapic_calibrate_fast);
82 static void lapic_timer_tscdlt_reload(struct cputimer_intr *, sysclock_t);
83 static void lapic_mem_timer_intr_reload(struct cputimer_intr *, sysclock_t);
84 static void lapic_timer_intr_enable(struct cputimer_intr *);
85 static void lapic_timer_intr_restart(struct cputimer_intr *);
86 static void lapic_timer_intr_pmfixup(struct cputimer_intr *);
88 static struct cputimer_intr lapic_cputimer_intr = {
90 .reload = lapic_mem_timer_intr_reload,
91 .enable = lapic_timer_intr_enable,
92 .config = cputimer_intr_default_config,
93 .restart = lapic_timer_intr_restart,
94 .pmfixup = lapic_timer_intr_pmfixup,
95 .initclock = cputimer_intr_default_initclock,
97 .next = SLIST_ENTRY_INITIALIZER,
99 .type = CPUTIMER_INTR_LAPIC,
100 .prio = CPUTIMER_INTR_PRIO_LAPIC,
101 .caps = CPUTIMER_INTR_CAP_NONE,
105 static int lapic_timer_divisor_idx = -1;
106 static const uint32_t lapic_timer_divisors[] = {
107 APIC_TDCR_2, APIC_TDCR_4, APIC_TDCR_8, APIC_TDCR_16,
108 APIC_TDCR_32, APIC_TDCR_64, APIC_TDCR_128, APIC_TDCR_1
110 #define APIC_TIMER_NDIVISORS (int)(NELEM(lapic_timer_divisors))
112 static int lapic_use_tscdeadline = 0;
113 /* The raw TSC frequency might not fit into a sysclock_t value. */
114 static int lapic_timer_tscfreq_shift;
117 * APIC ID <-> CPU ID mapping structures.
119 int cpu_id_to_apic_id[NAPICID];
120 int apic_id_to_cpu_id[NAPICID];
121 int lapic_enable = 1;
122 int lapic_usable = 0;
124 /* Separate cachelines for each cpu's info. */
127 uint64_t downcount_time;
130 struct deadlines *tsc_deadlines = NULL;
132 static void lapic_mem_eoi(void);
133 static int lapic_mem_ipi(int dest_type, int vector, int delivery_mode);
134 static void lapic_mem_single_ipi(int cpu, int vector, int delivery_mode);
136 void (*lapic_eoi)(void);
137 int (*apic_ipi)(int dest_type, int vector, int delivery_mode);
138 void (*single_apic_ipi)(int cpu, int vector, int delivery_mode);
141 lapic_mem_icr_set(uint32_t apic_id, uint32_t icr_lo_val)
143 uint32_t icr_lo, icr_hi;
145 icr_hi = (LAPIC_MEM_READ(icr_hi) & ~APIC_ID_MASK) |
146 (apic_id << APIC_ID_SHIFT);
147 icr_lo = (LAPIC_MEM_READ(icr_lo) & APIC_ICRLO_RESV_MASK) | icr_lo_val;
149 LAPIC_MEM_WRITE(icr_hi, icr_hi);
150 LAPIC_MEM_WRITE(icr_lo, icr_lo);
154 * Enable LAPIC, configure interrupts.
157 lapic_init(boolean_t bsp)
163 /* Decide whether we want to use TSC Deadline mode. */
164 if (lapic_timer_tscdeadline != 0 &&
165 (cpu_feature2 & CPUID2_TSCDLT) &&
166 tsc_invariant && tsc_frequency != 0) {
167 lapic_use_tscdeadline = 1;
168 tsc_deadlines = kmalloc_cachealign(
169 sizeof(struct deadlines) * (naps + 1),
170 M_DEVBUF, M_WAITOK | M_ZERO);
177 * Since IDT is shared between BSP and APs, these vectors
178 * only need to be installed once; we do it on BSP.
181 if (cpu_vendor_id == CPU_VENDOR_AMD &&
182 CPUID_TO_FAMILY(cpu_id) >= 0x0f &&
183 CPUID_TO_FAMILY(cpu_id) < 0x17) { /* XXX */
187 * Set the LINTEN bit in the HyperTransport
188 * Transaction Control Register.
190 * This will cause EXTINT and NMI interrupts
191 * routed over the hypertransport bus to be
192 * fed into the LAPIC LINT0/LINT1. If the bit
193 * isn't set, the interrupts will go to the
194 * general cpu INTR/NMI pins. On a dual-core
195 * cpu the interrupt winds up going to BOTH cpus.
196 * The first cpu that does the interrupt ack
197 * cycle will get the correct interrupt. The
198 * second cpu that does it will get a spurious
199 * interrupt vector (typically IRQ 7).
202 (1 << 31) | /* enable */
203 (0 << 16) | /* bus */
204 (0x18 << 11) | /* dev (cpu + 0x18) */
205 (0 << 8) | /* func */
209 if ((tcr & 0x00010000) == 0) {
210 kprintf("LAPIC: AMD LINTEN on\n");
211 outl(0xcfc, tcr|0x00010000);
216 /* Install a 'Spurious INTerrupt' vector */
217 setidt_global(XSPURIOUSINT_OFFSET, Xspuriousint,
218 SDT_SYSIGT, SEL_KPL, 0);
220 /* Install a timer vector */
221 setidt_global(XTIMER_OFFSET, Xtimer,
222 SDT_SYSIGT, SEL_KPL, 0);
224 /* Install an inter-CPU IPI for TLB invalidation */
225 setidt_global(XINVLTLB_OFFSET, Xinvltlb,
226 SDT_SYSIGT, SEL_KPL, 0);
228 /* Install an inter-CPU IPI for IPIQ messaging */
229 setidt_global(XIPIQ_OFFSET, Xipiq,
230 SDT_SYSIGT, SEL_KPL, 0);
232 /* Install an inter-CPU IPI for CPU stop/restart */
233 setidt_global(XCPUSTOP_OFFSET, Xcpustop,
234 SDT_SYSIGT, SEL_KPL, 0);
236 /* Install an inter-CPU IPI for TLB invalidation */
237 setidt_global(XSNIFF_OFFSET, Xsniff,
238 SDT_SYSIGT, SEL_KPL, 0);
242 * Setup LINT0 as ExtINT on the BSP. This is theoretically an
243 * aggregate interrupt input from the 8259. The INTA cycle
244 * will be routed to the external controller (the 8259) which
245 * is expected to supply the vector.
247 * Must be setup edge triggered, active high.
249 * Disable LINT0 on BSP, if I/O APIC is enabled.
251 * Disable LINT0 on the APs. It doesn't matter what delivery
252 * mode we use because we leave it masked.
254 temp = LAPIC_READ(lvt_lint0);
255 temp &= ~(APIC_LVT_MASKED | APIC_LVT_TRIG_MASK |
256 APIC_LVT_POLARITY_MASK | APIC_LVT_DM_MASK);
258 temp |= APIC_LVT_DM_EXTINT;
260 temp |= APIC_LVT_MASKED;
262 temp |= APIC_LVT_DM_FIXED | APIC_LVT_MASKED;
264 LAPIC_WRITE(lvt_lint0, temp);
267 * Setup LINT1 as NMI.
269 * Must be setup edge trigger, active high.
271 * Enable LINT1 on BSP, if I/O APIC is enabled.
273 * Disable LINT1 on the APs.
275 temp = LAPIC_READ(lvt_lint1);
276 temp &= ~(APIC_LVT_MASKED | APIC_LVT_TRIG_MASK |
277 APIC_LVT_POLARITY_MASK | APIC_LVT_DM_MASK);
278 temp |= APIC_LVT_MASKED | APIC_LVT_DM_NMI;
279 if (bsp && ioapic_enable)
280 temp &= ~APIC_LVT_MASKED;
281 LAPIC_WRITE(lvt_lint1, temp);
284 * Mask the LAPIC error interrupt, LAPIC performance counter
287 LAPIC_WRITE(lvt_error, LAPIC_READ(lvt_error) | APIC_LVT_MASKED);
288 LAPIC_WRITE(lvt_pcint, LAPIC_READ(lvt_pcint) | APIC_LVT_MASKED);
291 * Set LAPIC timer vector and mask the LAPIC timer interrupt.
293 timer = LAPIC_READ(lvt_timer);
294 timer &= ~APIC_LVTT_VECTOR;
295 timer |= XTIMER_OFFSET;
296 timer |= APIC_LVTT_MASKED;
297 LAPIC_WRITE(lvt_timer, timer);
300 * Set the Task Priority Register as needed. At the moment allow
301 * interrupts on all cpus (the APs will remain CLId until they are
304 temp = LAPIC_READ(tpr);
305 temp &= ~APIC_TPR_PRIO; /* clear priority field */
306 LAPIC_WRITE(tpr, temp);
311 if (cpu_vendor_id == CPU_VENDOR_AMD && lapic_mem != NULL &&
312 (LAPIC_MEM_READ(version) & APIC_VER_AMD_EXT_SPACE)) {
319 ext_feat = LAPIC_MEM_READ(ext_feat);
320 count = (ext_feat & APIC_EXTFEAT_MASK) >> APIC_EXTFEAT_SHIFT;
321 max_count = sizeof(lapic_mem->ext_lvt) /
322 sizeof(lapic_mem->ext_lvt[0]);
323 if (count > max_count)
325 for (i = 0; i < count; ++i) {
326 lvt = LAPIC_MEM_READ(ext_lvt[i].lvt);
328 lvt &= ~(APIC_LVT_POLARITY_MASK | APIC_LVT_TRIG_MASK |
329 APIC_LVT_DM_MASK | APIC_LVT_MASKED);
330 lvt |= APIC_LVT_MASKED | APIC_LVT_DM_FIXED;
333 case APIC_EXTLVT_IBS:
335 case APIC_EXTLVT_MCA:
337 case APIC_EXTLVT_DEI:
339 case APIC_EXTLVT_SBI:
345 kprintf(" LAPIC AMD elvt%d: 0x%08x",
346 i, LAPIC_MEM_READ(ext_lvt[i].lvt));
347 if (LAPIC_MEM_READ(ext_lvt[i].lvt) != lvt)
348 kprintf(" -> 0x%08x", lvt);
351 LAPIC_MEM_WRITE(ext_lvt[i].lvt, lvt);
358 temp = LAPIC_READ(svr);
359 temp |= APIC_SVR_ENABLE; /* enable the LAPIC */
360 temp &= ~APIC_SVR_FOCUS_DISABLE; /* enable lopri focus processor */
362 if (LAPIC_READ(version) & APIC_VER_EOI_SUPP) {
363 if (temp & APIC_SVR_EOI_SUPP) {
364 temp &= ~APIC_SVR_EOI_SUPP;
366 kprintf(" LAPIC disabling EOI supp\n");
371 * Set the spurious interrupt vector. The low 4 bits of the vector
374 if ((XSPURIOUSINT_OFFSET & 0x0F) != 0x0F)
375 panic("bad XSPURIOUSINT_OFFSET: 0x%08x", XSPURIOUSINT_OFFSET);
376 temp &= ~APIC_SVR_VECTOR;
377 temp |= XSPURIOUSINT_OFFSET;
379 LAPIC_WRITE(svr, temp);
382 * Pump out a few EOIs to clean out interrupts that got through
383 * before we were able to set the TPR.
390 lapic_timer_calibrate();
391 if (lapic_timer_enable) {
392 if (cpu_thermal_feature & CPUID_THERMAL_ARAT) {
394 * Local APIC timer will not stop
397 lapic_cputimer_intr.caps |=
398 CPUTIMER_INTR_CAP_PS;
400 if (lapic_use_tscdeadline) {
401 lapic_cputimer_intr.reload =
402 lapic_timer_tscdlt_reload;
404 cputimer_intr_register(&lapic_cputimer_intr);
405 cputimer_intr_select(&lapic_cputimer_intr, 0);
407 } else if (!lapic_use_tscdeadline) {
408 lapic_timer_set_divisor(lapic_timer_divisor_idx);
412 apic_dump("apic_initialize()");
416 lapic_timer_set_divisor(int divisor_idx)
418 KKASSERT(divisor_idx >= 0 && divisor_idx < APIC_TIMER_NDIVISORS);
419 LAPIC_WRITE(dcr_timer, lapic_timer_divisors[divisor_idx]);
423 lapic_timer_oneshot(u_int count)
427 value = LAPIC_READ(lvt_timer);
428 value &= ~(APIC_LVTT_PERIODIC | APIC_LVTT_TSCDLT);
429 LAPIC_WRITE(lvt_timer, value);
430 LAPIC_WRITE(icr_timer, count);
434 lapic_timer_oneshot_quick(u_int count)
436 LAPIC_WRITE(icr_timer, count);
440 lapic_timer_tscdeadline_quick(uint64_t diff)
442 uint64_t val = rdtsc() + diff;
444 wrmsr(MSR_TSC_DEADLINE, val);
445 tsc_deadlines[mycpuid].timestamp = val;
449 lapic_scale_to_tsc(unsigned value, unsigned scale)
454 val *= tsc_frequency;
460 #define MAX_MEASURE_RETRIES 100
463 do_tsc_calibration(u_int us, u_int64_t apic_delay_tsc)
465 u_int64_t old_tsc1, old_tsc2, new_tsc1, new_tsc2;
466 u_int64_t diff, count;
468 u_int32_t start, end;
469 int retries1 = 0, retries2 = 0;
472 lapic_timer_oneshot_quick(APIC_TIMER_MAX_COUNT);
473 old_tsc1 = rdtsc_ordered();
474 start = LAPIC_READ(ccr_timer);
475 old_tsc2 = rdtsc_ordered();
476 if (apic_delay_tsc > 0 && retries1 < MAX_MEASURE_RETRIES &&
477 old_tsc2 - old_tsc1 > 2 * apic_delay_tsc) {
483 new_tsc1 = rdtsc_ordered();
484 end = LAPIC_READ(ccr_timer);
485 new_tsc2 = rdtsc_ordered();
486 if (apic_delay_tsc > 0 && retries2 < MAX_MEASURE_RETRIES &&
487 new_tsc2 - new_tsc1 > 2 * apic_delay_tsc) {
496 /* Make sure the lapic can count for up to 2s */
497 a = (unsigned)APIC_TIMER_MAX_COUNT;
498 if (us < 2000000 && (u_int64_t)count * 2000000 >= a * us)
501 if (lapic_calibrate_test > 0 && (retries1 > 0 || retries2 > 0)) {
502 kprintf("%s: retries1=%d retries2=%d\n",
503 __func__, retries1, retries2);
506 diff = (new_tsc1 - old_tsc1) + (new_tsc2 - old_tsc2);
507 /* XXX First estimate if the total TSC diff value makes sense */
508 /* This will almost overflow, but only almost :) */
509 count = (2 * count * tsc_frequency) / diff;
515 do_cputimer_calibration(u_int us)
518 sysclock_t start, end, beginning, finish;
520 lapic_timer_oneshot(APIC_TIMER_MAX_COUNT);
521 beginning = LAPIC_READ(ccr_timer);
522 start = sys_cputimer->count();
524 end = sys_cputimer->count();
525 finish = LAPIC_READ(ccr_timer);
528 /* value is the LAPIC timer difference. */
529 value = beginning - finish;
530 /* end is the sys_cputimer difference. */
534 value = ((uint64_t)value * sys_cputimer->freq) / end;
539 lapic_timer_calibrate(void)
542 u_int64_t apic_delay_tsc = 0;
543 int use_tsc_calibration = 0;
545 /* No need to calibrate lapic_timer, if we will use TSC Deadline mode */
546 if (lapic_use_tscdeadline) {
547 lapic_timer_tscfreq_shift = 0;
548 while ((tsc_frequency >> lapic_timer_tscfreq_shift) > INT_MAX)
549 lapic_timer_tscfreq_shift++;
550 lapic_cputimer_intr.freq =
551 tsc_frequency >> lapic_timer_tscfreq_shift;
553 "lapic: TSC Deadline Mode: shift %d, frequency %u Hz\n",
554 lapic_timer_tscfreq_shift, lapic_cputimer_intr.freq);
559 * On real hardware, tsc_invariant == 0 wouldn't be an issue, but in
560 * a virtual machine the frequency may get changed by the host.
562 if (tsc_frequency != 0 && tsc_invariant && lapic_calibrate_fast)
563 use_tsc_calibration = 1;
565 if (use_tsc_calibration) {
566 u_int64_t min_apic_tsc = 0, max_apic_tsc = 0;
567 u_int64_t old_tsc, new_tsc;
572 lapic_timer_oneshot(APIC_TIMER_MAX_COUNT);
573 for (i = 0; i < 10; i++)
574 val = LAPIC_READ(ccr_timer);
576 for (i = 0; i < 100; i++) {
577 old_tsc = rdtsc_ordered();
578 val = LAPIC_READ(ccr_timer);
579 new_tsc = rdtsc_ordered();
581 apic_delay_tsc += new_tsc;
582 if (min_apic_tsc == 0 ||
583 min_apic_tsc > new_tsc) {
584 min_apic_tsc = new_tsc;
586 if (max_apic_tsc < new_tsc)
587 max_apic_tsc = new_tsc;
589 apic_delay_tsc /= 100;
591 "LAPIC latency (in TSC ticks): %lu min: %lu max: %lu\n",
592 apic_delay_tsc, min_apic_tsc, max_apic_tsc);
593 apic_delay_tsc = min_apic_tsc;
596 if (!use_tsc_calibration) {
600 * Do some exercising of the lapic timer access. This improves
601 * precision of the subsequent calibration run in at least some
602 * virtualization cases.
604 lapic_timer_set_divisor(0);
605 for (i = 0; i < 10; i++)
606 (void)do_cputimer_calibration(100);
608 /* Try to calibrate the local APIC timer. */
609 for (lapic_timer_divisor_idx = 0;
610 lapic_timer_divisor_idx < APIC_TIMER_NDIVISORS;
611 lapic_timer_divisor_idx++) {
612 lapic_timer_set_divisor(lapic_timer_divisor_idx);
613 if (use_tsc_calibration) {
614 value = do_tsc_calibration(200*1000, apic_delay_tsc);
616 value = do_cputimer_calibration(2*1000*1000);
621 if (lapic_timer_divisor_idx >= APIC_TIMER_NDIVISORS)
622 panic("lapic: no proper timer divisor?!");
623 lapic_cputimer_intr.freq = value;
625 kprintf("lapic: divisor index %d, frequency %u Hz\n",
626 lapic_timer_divisor_idx, lapic_cputimer_intr.freq);
628 if (lapic_calibrate_test > 0) {
632 for (i = 1; i <= 20; i++) {
633 if (use_tsc_calibration) {
634 freq = do_tsc_calibration(i*100*1000,
637 freq = do_cputimer_calibration(i*100*1000);
640 kprintf("%ums: %lu\n", i * 100, freq);
646 lapic_timer_tscdlt_reload(struct cputimer_intr *cti, sysclock_t reload)
648 struct globaldata *gd = mycpu;
649 uint64_t diff, now, val;
651 if (reload > 1000*1000*1000)
652 reload = 1000*1000*1000;
653 diff = (uint64_t)reload * tsc_frequency / sys_cputimer->freq;
656 if (cpu_vendor_id == CPU_VENDOR_INTEL)
662 if (gd->gd_timer_running) {
663 uint64_t deadline = tsc_deadlines[mycpuid].timestamp;
664 if (deadline == 0 || now > deadline || val < deadline) {
665 wrmsr(MSR_TSC_DEADLINE, val);
666 tsc_deadlines[mycpuid].timestamp = val;
669 gd->gd_timer_running = 1;
670 wrmsr(MSR_TSC_DEADLINE, val);
671 tsc_deadlines[mycpuid].timestamp = val;
676 lapic_mem_timer_intr_reload(struct cputimer_intr *cti, sysclock_t reload)
678 struct globaldata *gd = mycpu;
680 reload = (int64_t)reload * cti->freq / sys_cputimer->freq;
684 if (gd->gd_timer_running) {
685 if (reload < LAPIC_MEM_READ(ccr_timer))
686 LAPIC_MEM_WRITE(icr_timer, reload);
688 gd->gd_timer_running = 1;
689 LAPIC_MEM_WRITE(icr_timer, reload);
694 lapic_timer_intr_enable(struct cputimer_intr *cti __unused)
698 timer = LAPIC_READ(lvt_timer);
699 timer &= ~(APIC_LVTT_MASKED | APIC_LVTT_PERIODIC | APIC_LVTT_TSCDLT);
700 if (lapic_use_tscdeadline)
701 timer |= APIC_LVTT_TSCDLT;
702 LAPIC_WRITE(lvt_timer, timer);
703 if (lapic_use_tscdeadline)
706 lapic_timer_fixup_handler(NULL);
710 lapic_timer_fixup_handler(void *arg)
717 if (cpu_vendor_id == CPU_VENDOR_AMD) {
719 * Detect the presence of C1E capability mostly on latest
720 * dual-cores (or future) k8 family. This feature renders
721 * the local APIC timer dead, so we disable it by reading
722 * the Interrupt Pending Message register and clearing both
723 * C1eOnCmpHalt (bit 28) and SmiOnCmpHalt (bit 27).
726 * "BIOS and Kernel Developer's Guide for AMD NPT
727 * Family 0Fh Processors"
728 * #32559 revision 3.00
730 if ((cpu_id & 0x00000f00) == 0x00000f00 &&
731 (cpu_id & 0x0fff0000) >= 0x00040000) {
734 msr = rdmsr(0xc0010055);
735 if (msr & 0x18000000) {
736 struct globaldata *gd = mycpu;
738 kprintf("cpu%d: AMD C1E detected\n",
740 wrmsr(0xc0010055, msr & ~0x18000000ULL);
743 * We are kinda stalled;
746 gd->gd_timer_running = 1;
747 if (lapic_use_tscdeadline) {
748 /* Maybe reached in Virtual Machines? */
749 lapic_timer_tscdeadline_quick(5000);
751 lapic_timer_oneshot_quick(2);
762 lapic_timer_restart_handler(void *dummy __unused)
766 lapic_timer_fixup_handler(&started);
768 struct globaldata *gd = mycpu;
770 gd->gd_timer_running = 1;
771 if (lapic_use_tscdeadline) {
772 /* Maybe reached in Virtual Machines? */
773 lapic_timer_tscdeadline_quick(5000);
775 lapic_timer_oneshot_quick(2);
781 * This function is called only by ACPICA code currently:
782 * - AMD C1E fixup. AMD C1E only seems to happen after ACPI
783 * module controls PM. So once ACPICA is attached, we try
784 * to apply the fixup to prevent LAPIC timer from hanging.
787 lapic_timer_intr_pmfixup(struct cputimer_intr *cti __unused)
789 lwkt_send_ipiq_mask(smp_active_mask,
790 lapic_timer_fixup_handler, NULL);
794 lapic_timer_intr_restart(struct cputimer_intr *cti __unused)
796 lwkt_send_ipiq_mask(smp_active_mask, lapic_timer_restart_handler, NULL);
801 * dump contents of local APIC registers
806 kprintf("SMP: CPU%d %s:\n", mycpu->gd_cpuid, str);
807 kprintf(" lint0: 0x%08x lint1: 0x%08x TPR: 0x%08x SVR: 0x%08x\n",
808 LAPIC_READ(lvt_lint0), LAPIC_READ(lvt_lint1), LAPIC_READ(tpr),
813 * Inter Processor Interrupt functions.
817 lapic_mem_icr_unpend(const char *func)
819 if (LAPIC_MEM_READ(icr_lo) & APIC_DELSTAT_PEND) {
824 while (LAPIC_MEM_READ(icr_lo) & APIC_DELSTAT_PEND) {
826 if ((tsc_sclock_t)(rdtsc() -
827 (tsc + tsc_frequency)) > 0) {
830 panic("%s: cpu%d apic stalled",
833 kprintf("%s: cpu%d apic stalled\n",
842 * Send APIC IPI 'vector' to 'destType' via 'deliveryMode'.
844 * destType is 1 of: APIC_DEST_SELF, APIC_DEST_ALLISELF, APIC_DEST_ALLESELF
845 * vector is any valid SYSTEM INT vector
846 * delivery_mode is 1 of: APIC_DELMODE_FIXED, APIC_DELMODE_LOWPRIO
850 * We now implement a per-cpu interlock (gd->gd_npoll) to prevent more than
851 * one IPI from being sent to any given cpu at a time. Thus we no longer
852 * have to process incoming IPIs while waiting for the status to clear.
853 * No deadlock should be possible.
855 * We now physically disable interrupts for the lapic ICR operation. If
856 * we do not do this then it looks like an EOI sent to the lapic (which
857 * occurs even with a critical section) can interfere with the command
858 * register ready status and cause an IPI to be lost.
860 * e.g. an interrupt can occur, issue the EOI, IRET, and cause the command
861 * register to busy just before we write to icr_lo, resulting in a lost
862 * issuance. This only appears to occur on Intel cpus and is not
863 * documented. It could simply be that cpus are so fast these days that
864 * it was always an issue, but is only now rearing its ugly head. This
868 lapic_mem_ipi(int dest_type, int vector, int delivery_mode)
870 lapic_mem_icr_unpend(__func__);
872 dest_type | APIC_LEVEL_ASSERT | delivery_mode | vector);
877 * Interrupts must be hard-disabled by caller
880 lapic_mem_single_ipi(int cpu, int vector, int delivery_mode)
882 lapic_mem_icr_unpend(__func__);
883 lapic_mem_icr_set(CPUID_TO_APICID(cpu),
884 APIC_DEST_DESTFLD | APIC_LEVEL_ASSERT | delivery_mode | vector);
888 * Send APIC IPI 'vector' to 'target's via 'delivery_mode'.
890 * target is a bitmask of destination cpus. Vector is any
891 * valid system INT vector. Delivery mode may be either
892 * APIC_DELMODE_FIXED or APIC_DELMODE_LOWPRIO.
894 * Interrupts must be hard-disabled by caller
897 selected_apic_ipi(cpumask_t target, int vector, int delivery_mode)
899 while (CPUMASK_TESTNZERO(target)) {
900 int n = BSFCPUMASK(target);
901 CPUMASK_NANDBIT(target, n);
902 single_apic_ipi(n, vector, delivery_mode);
907 * Load a 'downcount time' in uSeconds.
910 set_apic_timer(int us)
914 if (lapic_use_tscdeadline) {
917 val = lapic_scale_to_tsc(us, 1000000);
919 /* No need to arm the lapic here, just track the timeout. */
920 tsc_deadlines[mycpuid].downcount_time = val;
925 * When we reach here, lapic timer's frequency
926 * must have been calculated as well as the
927 * divisor (lapic->dcr_timer is setup during the
928 * divisor calculation).
930 KKASSERT(lapic_cputimer_intr.freq != 0 &&
931 lapic_timer_divisor_idx >= 0);
933 count = ((us * (int64_t)lapic_cputimer_intr.freq) + 999999) / 1000000;
934 lapic_timer_oneshot(count);
939 * Read remaining time in timer, in microseconds (rounded up).
942 read_apic_timer(void)
946 if (lapic_use_tscdeadline) {
949 val = tsc_deadlines[mycpuid].downcount_time;
951 if (val == 0 || now > val) {
956 val += (tsc_frequency - 1);
957 val /= tsc_frequency;
964 val = LAPIC_READ(ccr_timer);
968 KKASSERT(lapic_cputimer_intr.freq > 0);
970 val += (lapic_cputimer_intr.freq - 1);
971 val /= lapic_cputimer_intr.freq;
979 * Spin-style delay, set delay time in uS, spin till it drains.
984 set_apic_timer(count);
985 while (read_apic_timer())
990 lapic_unused_apic_id(int start)
994 for (i = start; i < APICID_MAX; ++i) {
995 if (APICID_TO_CPUID(i) == -1)
1002 lapic_map(vm_paddr_t lapic_addr)
1004 lapic_mem = pmap_mapdev_uncacheable(lapic_addr, sizeof(struct LAPIC));
1007 static TAILQ_HEAD(, lapic_enumerator) lapic_enumerators =
1008 TAILQ_HEAD_INITIALIZER(lapic_enumerators);
1013 struct lapic_enumerator *e;
1014 int error, i, ap_max;
1016 KKASSERT(lapic_enable);
1018 for (i = 0; i < NAPICID; ++i)
1019 APICID_TO_CPUID(i) = -1;
1021 TAILQ_FOREACH(e, &lapic_enumerators, lapic_link) {
1022 error = e->lapic_probe(e);
1027 kprintf("LAPIC: Can't find LAPIC\n");
1031 error = e->lapic_enumerate(e);
1033 kprintf("LAPIC: enumeration failed\n");
1037 /* LAPIC is usable now. */
1040 ap_max = MAXCPU - 1;
1041 TUNABLE_INT_FETCH("hw.ap_max", &ap_max);
1042 if (ap_max > MAXCPU - 1)
1043 ap_max = MAXCPU - 1;
1045 if (naps > ap_max) {
1046 kprintf("LAPIC: Warning use only %d out of %d "
1056 lapic_enumerator_register(struct lapic_enumerator *ne)
1058 struct lapic_enumerator *e;
1060 TAILQ_FOREACH(e, &lapic_enumerators, lapic_link) {
1061 if (e->lapic_prio < ne->lapic_prio) {
1062 TAILQ_INSERT_BEFORE(e, ne, lapic_link);
1066 TAILQ_INSERT_TAIL(&lapic_enumerators, ne, lapic_link);
1070 lapic_set_cpuid(int cpu_id, int apic_id)
1072 CPUID_TO_APICID(cpu_id) = apic_id;
1073 APICID_TO_CPUID(apic_id) = cpu_id;
1077 lapic_fixup_noioapic(void)
1081 /* Only allowed on BSP */
1082 KKASSERT(mycpuid == 0);
1083 KKASSERT(!ioapic_enable);
1085 temp = LAPIC_READ(lvt_lint0);
1086 temp &= ~APIC_LVT_MASKED;
1087 LAPIC_WRITE(lvt_lint0, temp);
1089 temp = LAPIC_READ(lvt_lint1);
1090 temp |= APIC_LVT_MASKED;
1091 LAPIC_WRITE(lvt_lint1, temp);
1098 LAPIC_MEM_WRITE(eoi, 0);
1102 lapic_mem_seticr_sync(uint32_t apic_id, uint32_t icr_lo_val)
1104 lapic_mem_icr_set(apic_id, icr_lo_val);
1105 while (LAPIC_MEM_READ(icr_lo) & APIC_DELSTAT_PEND)
1110 lapic_seticr_sync(uint32_t apic_id, uint32_t icr_lo_val)
1113 lapic_mem_seticr_sync(apic_id, icr_lo_val);
1117 lapic_sysinit(void *dummy __unused)
1122 lapic_eoi = lapic_mem_eoi;
1123 apic_ipi = lapic_mem_ipi;
1124 single_apic_ipi = lapic_mem_single_ipi;
1126 error = lapic_config();
1132 /* Initialize BSP's local APIC */
1134 } else if (ioapic_enable) {
1136 icu_reinit_noioapic();
1139 SYSINIT(lapic, SI_BOOT2_LAPIC, SI_ORDER_FIRST, lapic_sysinit, NULL);