Loading...
1/*
2 * x86 SMP booting functions
3 *
4 * (c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
5 * (c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com>
6 * Copyright 2001 Andi Kleen, SuSE Labs.
7 *
8 * Much of the core SMP work is based on previous work by Thomas Radke, to
9 * whom a great many thanks are extended.
10 *
11 * Thanks to Intel for making available several different Pentium,
12 * Pentium Pro and Pentium-II/Xeon MP machines.
13 * Original development of Linux SMP code supported by Caldera.
14 *
15 * This code is released under the GNU General Public License version 2 or
16 * later.
17 *
18 * Fixes
19 * Felix Koop : NR_CPUS used properly
20 * Jose Renau : Handle single CPU case.
21 * Alan Cox : By repeated request 8) - Total BogoMIPS report.
22 * Greg Wright : Fix for kernel stacks panic.
23 * Erich Boleyn : MP v1.4 and additional changes.
24 * Matthias Sattler : Changes for 2.1 kernel map.
25 * Michel Lespinasse : Changes for 2.1 kernel map.
26 * Michael Chastain : Change trampoline.S to gnu as.
27 * Alan Cox : Dumb bug: 'B' step PPro's are fine
28 * Ingo Molnar : Added APIC timers, based on code
29 * from Jose Renau
30 * Ingo Molnar : various cleanups and rewrites
31 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug.
32 * Maciej W. Rozycki : Bits for genuine 82489DX APICs
33 * Andi Kleen : Changed for SMP boot into long mode.
34 * Martin J. Bligh : Added support for multi-quad systems
35 * Dave Jones : Report invalid combinations of Athlon CPUs.
36 * Rusty Russell : Hacked into shape for new "hotplug" boot process.
37 * Andi Kleen : Converted to new state machine.
38 * Ashok Raj : CPU hotplug support
39 * Glauber Costa : i386 and x86_64 integration
40 */
41
42#include <linux/init.h>
43#include <linux/smp.h>
44#include <linux/module.h>
45#include <linux/sched.h>
46#include <linux/percpu.h>
47#include <linux/bootmem.h>
48#include <linux/err.h>
49#include <linux/nmi.h>
50#include <linux/tboot.h>
51#include <linux/stackprotector.h>
52#include <linux/gfp.h>
53#include <linux/cpuidle.h>
54
55#include <asm/acpi.h>
56#include <asm/desc.h>
57#include <asm/nmi.h>
58#include <asm/irq.h>
59#include <asm/idle.h>
60#include <asm/realmode.h>
61#include <asm/cpu.h>
62#include <asm/numa.h>
63#include <asm/pgtable.h>
64#include <asm/tlbflush.h>
65#include <asm/mtrr.h>
66#include <asm/mwait.h>
67#include <asm/apic.h>
68#include <asm/io_apic.h>
69#include <asm/setup.h>
70#include <asm/uv/uv.h>
71#include <linux/mc146818rtc.h>
72
73#include <asm/smpboot_hooks.h>
74#include <asm/i8259.h>
75
76#include <asm/realmode.h>
77
78/* State of each CPU */
79DEFINE_PER_CPU(int, cpu_state) = { 0 };
80
81#ifdef CONFIG_HOTPLUG_CPU
82/*
83 * We need this for trampoline_base protection from concurrent accesses when
84 * off- and onlining cores wildly.
85 */
86static DEFINE_MUTEX(x86_cpu_hotplug_driver_mutex);
87
88void cpu_hotplug_driver_lock(void)
89{
90 mutex_lock(&x86_cpu_hotplug_driver_mutex);
91}
92
93void cpu_hotplug_driver_unlock(void)
94{
95 mutex_unlock(&x86_cpu_hotplug_driver_mutex);
96}
97
98ssize_t arch_cpu_probe(const char *buf, size_t count) { return -1; }
99ssize_t arch_cpu_release(const char *buf, size_t count) { return -1; }
100#endif
101
102/* Number of siblings per CPU package */
103int smp_num_siblings = 1;
104EXPORT_SYMBOL(smp_num_siblings);
105
106/* Last level cache ID of each logical CPU */
107DEFINE_PER_CPU(u16, cpu_llc_id) = BAD_APICID;
108
109/* representing HT siblings of each logical CPU */
110DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
111EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
112
113/* representing HT and core siblings of each logical CPU */
114DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
115EXPORT_PER_CPU_SYMBOL(cpu_core_map);
116
117DEFINE_PER_CPU(cpumask_var_t, cpu_llc_shared_map);
118
119/* Per CPU bogomips and other parameters */
120DEFINE_PER_CPU_SHARED_ALIGNED(struct cpuinfo_x86, cpu_info);
121EXPORT_PER_CPU_SYMBOL(cpu_info);
122
123atomic_t init_deasserted;
124
125/*
126 * Report back to the Boot Processor.
127 * Running on AP.
128 */
129static void __cpuinit smp_callin(void)
130{
131 int cpuid, phys_id;
132 unsigned long timeout;
133
134 /*
135 * If waken up by an INIT in an 82489DX configuration
136 * we may get here before an INIT-deassert IPI reaches
137 * our local APIC. We have to wait for the IPI or we'll
138 * lock up on an APIC access.
139 */
140 if (apic->wait_for_init_deassert)
141 apic->wait_for_init_deassert(&init_deasserted);
142
143 /*
144 * (This works even if the APIC is not enabled.)
145 */
146 phys_id = read_apic_id();
147 cpuid = smp_processor_id();
148 if (cpumask_test_cpu(cpuid, cpu_callin_mask)) {
149 panic("%s: phys CPU#%d, CPU#%d already present??\n", __func__,
150 phys_id, cpuid);
151 }
152 pr_debug("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid, phys_id);
153
154 /*
155 * STARTUP IPIs are fragile beasts as they might sometimes
156 * trigger some glue motherboard logic. Complete APIC bus
157 * silence for 1 second, this overestimates the time the
158 * boot CPU is spending to send the up to 2 STARTUP IPIs
159 * by a factor of two. This should be enough.
160 */
161
162 /*
163 * Waiting 2s total for startup (udelay is not yet working)
164 */
165 timeout = jiffies + 2*HZ;
166 while (time_before(jiffies, timeout)) {
167 /*
168 * Has the boot CPU finished it's STARTUP sequence?
169 */
170 if (cpumask_test_cpu(cpuid, cpu_callout_mask))
171 break;
172 cpu_relax();
173 }
174
175 if (!time_before(jiffies, timeout)) {
176 panic("%s: CPU%d started up but did not get a callout!\n",
177 __func__, cpuid);
178 }
179
180 /*
181 * the boot CPU has finished the init stage and is spinning
182 * on callin_map until we finish. We are free to set up this
183 * CPU, first the APIC. (this is probably redundant on most
184 * boards)
185 */
186
187 pr_debug("CALLIN, before setup_local_APIC().\n");
188 if (apic->smp_callin_clear_local_apic)
189 apic->smp_callin_clear_local_apic();
190 setup_local_APIC();
191 end_local_APIC_setup();
192
193 /*
194 * Need to setup vector mappings before we enable interrupts.
195 */
196 setup_vector_irq(smp_processor_id());
197
198 /*
199 * Save our processor parameters. Note: this information
200 * is needed for clock calibration.
201 */
202 smp_store_cpu_info(cpuid);
203
204 /*
205 * Get our bogomips.
206 * Update loops_per_jiffy in cpu_data. Previous call to
207 * smp_store_cpu_info() stored a value that is close but not as
208 * accurate as the value just calculated.
209 */
210 calibrate_delay();
211 cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
212 pr_debug("Stack at about %p\n", &cpuid);
213
214 /*
215 * This must be done before setting cpu_online_mask
216 * or calling notify_cpu_starting.
217 */
218 set_cpu_sibling_map(raw_smp_processor_id());
219 wmb();
220
221 notify_cpu_starting(cpuid);
222
223 /*
224 * Allow the master to continue.
225 */
226 cpumask_set_cpu(cpuid, cpu_callin_mask);
227}
228
229/*
230 * Activate a secondary processor.
231 */
232notrace static void __cpuinit start_secondary(void *unused)
233{
234 /*
235 * Don't put *anything* before cpu_init(), SMP booting is too
236 * fragile that we want to limit the things done here to the
237 * most necessary things.
238 */
239 cpu_init();
240 x86_cpuinit.early_percpu_clock_init();
241 preempt_disable();
242 smp_callin();
243
244#ifdef CONFIG_X86_32
245 /* switch away from the initial page table */
246 load_cr3(swapper_pg_dir);
247 __flush_tlb_all();
248#endif
249
250 /* otherwise gcc will move up smp_processor_id before the cpu_init */
251 barrier();
252 /*
253 * Check TSC synchronization with the BP:
254 */
255 check_tsc_sync_target();
256
257 /*
258 * We need to hold call_lock, so there is no inconsistency
259 * between the time smp_call_function() determines number of
260 * IPI recipients, and the time when the determination is made
261 * for which cpus receive the IPI. Holding this
262 * lock helps us to not include this cpu in a currently in progress
263 * smp_call_function().
264 *
265 * We need to hold vector_lock so there the set of online cpus
266 * does not change while we are assigning vectors to cpus. Holding
267 * this lock ensures we don't half assign or remove an irq from a cpu.
268 */
269 ipi_call_lock();
270 lock_vector_lock();
271 set_cpu_online(smp_processor_id(), true);
272 unlock_vector_lock();
273 ipi_call_unlock();
274 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
275 x86_platform.nmi_init();
276
277 /* enable local interrupts */
278 local_irq_enable();
279
280 /* to prevent fake stack check failure in clock setup */
281 boot_init_stack_canary();
282
283 x86_cpuinit.setup_percpu_clockev();
284
285 wmb();
286 cpu_idle();
287}
288
289/*
290 * The bootstrap kernel entry code has set these up. Save them for
291 * a given CPU
292 */
293
294void __cpuinit smp_store_cpu_info(int id)
295{
296 struct cpuinfo_x86 *c = &cpu_data(id);
297
298 *c = boot_cpu_data;
299 c->cpu_index = id;
300 if (id != 0)
301 identify_secondary_cpu(c);
302}
303
304static bool __cpuinit
305topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
306{
307 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
308
309 return !WARN_ONCE(cpu_to_node(cpu1) != cpu_to_node(cpu2),
310 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
311 "[node: %d != %d]. Ignoring dependency.\n",
312 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
313}
314
315#define link_mask(_m, c1, c2) \
316do { \
317 cpumask_set_cpu((c1), cpu_##_m##_mask(c2)); \
318 cpumask_set_cpu((c2), cpu_##_m##_mask(c1)); \
319} while (0)
320
321static bool __cpuinit match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
322{
323 if (cpu_has(c, X86_FEATURE_TOPOEXT)) {
324 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
325
326 if (c->phys_proc_id == o->phys_proc_id &&
327 per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2) &&
328 c->compute_unit_id == o->compute_unit_id)
329 return topology_sane(c, o, "smt");
330
331 } else if (c->phys_proc_id == o->phys_proc_id &&
332 c->cpu_core_id == o->cpu_core_id) {
333 return topology_sane(c, o, "smt");
334 }
335
336 return false;
337}
338
339static bool __cpuinit match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
340{
341 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
342
343 if (per_cpu(cpu_llc_id, cpu1) != BAD_APICID &&
344 per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2))
345 return topology_sane(c, o, "llc");
346
347 return false;
348}
349
350static bool __cpuinit match_mc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
351{
352 if (c->phys_proc_id == o->phys_proc_id) {
353 if (cpu_has(c, X86_FEATURE_AMD_DCM))
354 return true;
355
356 return topology_sane(c, o, "mc");
357 }
358 return false;
359}
360
361void __cpuinit set_cpu_sibling_map(int cpu)
362{
363 bool has_mc = boot_cpu_data.x86_max_cores > 1;
364 bool has_smt = smp_num_siblings > 1;
365 struct cpuinfo_x86 *c = &cpu_data(cpu);
366 struct cpuinfo_x86 *o;
367 int i;
368
369 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
370
371 if (!has_smt && !has_mc) {
372 cpumask_set_cpu(cpu, cpu_sibling_mask(cpu));
373 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
374 cpumask_set_cpu(cpu, cpu_core_mask(cpu));
375 c->booted_cores = 1;
376 return;
377 }
378
379 for_each_cpu(i, cpu_sibling_setup_mask) {
380 o = &cpu_data(i);
381
382 if ((i == cpu) || (has_smt && match_smt(c, o)))
383 link_mask(sibling, cpu, i);
384
385 if ((i == cpu) || (has_mc && match_llc(c, o)))
386 link_mask(llc_shared, cpu, i);
387
388 }
389
390 /*
391 * This needs a separate iteration over the cpus because we rely on all
392 * cpu_sibling_mask links to be set-up.
393 */
394 for_each_cpu(i, cpu_sibling_setup_mask) {
395 o = &cpu_data(i);
396
397 if ((i == cpu) || (has_mc && match_mc(c, o))) {
398 link_mask(core, cpu, i);
399
400 /*
401 * Does this new cpu bringup a new core?
402 */
403 if (cpumask_weight(cpu_sibling_mask(cpu)) == 1) {
404 /*
405 * for each core in package, increment
406 * the booted_cores for this new cpu
407 */
408 if (cpumask_first(cpu_sibling_mask(i)) == i)
409 c->booted_cores++;
410 /*
411 * increment the core count for all
412 * the other cpus in this package
413 */
414 if (i != cpu)
415 cpu_data(i).booted_cores++;
416 } else if (i != cpu && !c->booted_cores)
417 c->booted_cores = cpu_data(i).booted_cores;
418 }
419 }
420}
421
422/* maps the cpu to the sched domain representing multi-core */
423const struct cpumask *cpu_coregroup_mask(int cpu)
424{
425 return cpu_llc_shared_mask(cpu);
426}
427
428static void impress_friends(void)
429{
430 int cpu;
431 unsigned long bogosum = 0;
432 /*
433 * Allow the user to impress friends.
434 */
435 pr_debug("Before bogomips.\n");
436 for_each_possible_cpu(cpu)
437 if (cpumask_test_cpu(cpu, cpu_callout_mask))
438 bogosum += cpu_data(cpu).loops_per_jiffy;
439 printk(KERN_INFO
440 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
441 num_online_cpus(),
442 bogosum/(500000/HZ),
443 (bogosum/(5000/HZ))%100);
444
445 pr_debug("Before bogocount - setting activated=1.\n");
446}
447
448void __inquire_remote_apic(int apicid)
449{
450 unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
451 const char * const names[] = { "ID", "VERSION", "SPIV" };
452 int timeout;
453 u32 status;
454
455 printk(KERN_INFO "Inquiring remote APIC 0x%x...\n", apicid);
456
457 for (i = 0; i < ARRAY_SIZE(regs); i++) {
458 printk(KERN_INFO "... APIC 0x%x %s: ", apicid, names[i]);
459
460 /*
461 * Wait for idle.
462 */
463 status = safe_apic_wait_icr_idle();
464 if (status)
465 printk(KERN_CONT
466 "a previous APIC delivery may have failed\n");
467
468 apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
469
470 timeout = 0;
471 do {
472 udelay(100);
473 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
474 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
475
476 switch (status) {
477 case APIC_ICR_RR_VALID:
478 status = apic_read(APIC_RRR);
479 printk(KERN_CONT "%08x\n", status);
480 break;
481 default:
482 printk(KERN_CONT "failed\n");
483 }
484 }
485}
486
487/*
488 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
489 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
490 * won't ... remember to clear down the APIC, etc later.
491 */
492int __cpuinit
493wakeup_secondary_cpu_via_nmi(int logical_apicid, unsigned long start_eip)
494{
495 unsigned long send_status, accept_status = 0;
496 int maxlvt;
497
498 /* Target chip */
499 /* Boot on the stack */
500 /* Kick the second */
501 apic_icr_write(APIC_DM_NMI | apic->dest_logical, logical_apicid);
502
503 pr_debug("Waiting for send to finish...\n");
504 send_status = safe_apic_wait_icr_idle();
505
506 /*
507 * Give the other CPU some time to accept the IPI.
508 */
509 udelay(200);
510 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid])) {
511 maxlvt = lapic_get_maxlvt();
512 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
513 apic_write(APIC_ESR, 0);
514 accept_status = (apic_read(APIC_ESR) & 0xEF);
515 }
516 pr_debug("NMI sent.\n");
517
518 if (send_status)
519 printk(KERN_ERR "APIC never delivered???\n");
520 if (accept_status)
521 printk(KERN_ERR "APIC delivery error (%lx).\n", accept_status);
522
523 return (send_status | accept_status);
524}
525
526static int __cpuinit
527wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
528{
529 unsigned long send_status, accept_status = 0;
530 int maxlvt, num_starts, j;
531
532 maxlvt = lapic_get_maxlvt();
533
534 /*
535 * Be paranoid about clearing APIC errors.
536 */
537 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
538 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
539 apic_write(APIC_ESR, 0);
540 apic_read(APIC_ESR);
541 }
542
543 pr_debug("Asserting INIT.\n");
544
545 /*
546 * Turn INIT on target chip
547 */
548 /*
549 * Send IPI
550 */
551 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
552 phys_apicid);
553
554 pr_debug("Waiting for send to finish...\n");
555 send_status = safe_apic_wait_icr_idle();
556
557 mdelay(10);
558
559 pr_debug("Deasserting INIT.\n");
560
561 /* Target chip */
562 /* Send IPI */
563 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
564
565 pr_debug("Waiting for send to finish...\n");
566 send_status = safe_apic_wait_icr_idle();
567
568 mb();
569 atomic_set(&init_deasserted, 1);
570
571 /*
572 * Should we send STARTUP IPIs ?
573 *
574 * Determine this based on the APIC version.
575 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
576 */
577 if (APIC_INTEGRATED(apic_version[phys_apicid]))
578 num_starts = 2;
579 else
580 num_starts = 0;
581
582 /*
583 * Paravirt / VMI wants a startup IPI hook here to set up the
584 * target processor state.
585 */
586 startup_ipi_hook(phys_apicid, (unsigned long) start_secondary,
587 stack_start);
588
589 /*
590 * Run STARTUP IPI loop.
591 */
592 pr_debug("#startup loops: %d.\n", num_starts);
593
594 for (j = 1; j <= num_starts; j++) {
595 pr_debug("Sending STARTUP #%d.\n", j);
596 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
597 apic_write(APIC_ESR, 0);
598 apic_read(APIC_ESR);
599 pr_debug("After apic_write.\n");
600
601 /*
602 * STARTUP IPI
603 */
604
605 /* Target chip */
606 /* Boot on the stack */
607 /* Kick the second */
608 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
609 phys_apicid);
610
611 /*
612 * Give the other CPU some time to accept the IPI.
613 */
614 udelay(300);
615
616 pr_debug("Startup point 1.\n");
617
618 pr_debug("Waiting for send to finish...\n");
619 send_status = safe_apic_wait_icr_idle();
620
621 /*
622 * Give the other CPU some time to accept the IPI.
623 */
624 udelay(200);
625 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
626 apic_write(APIC_ESR, 0);
627 accept_status = (apic_read(APIC_ESR) & 0xEF);
628 if (send_status || accept_status)
629 break;
630 }
631 pr_debug("After Startup.\n");
632
633 if (send_status)
634 printk(KERN_ERR "APIC never delivered???\n");
635 if (accept_status)
636 printk(KERN_ERR "APIC delivery error (%lx).\n", accept_status);
637
638 return (send_status | accept_status);
639}
640
641/* reduce the number of lines printed when booting a large cpu count system */
642static void __cpuinit announce_cpu(int cpu, int apicid)
643{
644 static int current_node = -1;
645 int node = early_cpu_to_node(cpu);
646
647 if (system_state == SYSTEM_BOOTING) {
648 if (node != current_node) {
649 if (current_node > (-1))
650 pr_cont(" Ok.\n");
651 current_node = node;
652 pr_info("Booting Node %3d, Processors ", node);
653 }
654 pr_cont(" #%d%s", cpu, cpu == (nr_cpu_ids - 1) ? " Ok.\n" : "");
655 return;
656 } else
657 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
658 node, cpu, apicid);
659}
660
661/*
662 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
663 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
664 * Returns zero if CPU booted OK, else error code from
665 * ->wakeup_secondary_cpu.
666 */
667static int __cpuinit do_boot_cpu(int apicid, int cpu, struct task_struct *idle)
668{
669 volatile u32 *trampoline_status =
670 (volatile u32 *) __va(real_mode_header->trampoline_status);
671 /* start_ip had better be page-aligned! */
672 unsigned long start_ip = real_mode_header->trampoline_start;
673
674 unsigned long boot_error = 0;
675 int timeout;
676
677 alternatives_smp_switch(1);
678
679 idle->thread.sp = (unsigned long) (((struct pt_regs *)
680 (THREAD_SIZE + task_stack_page(idle))) - 1);
681 per_cpu(current_task, cpu) = idle;
682
683#ifdef CONFIG_X86_32
684 /* Stack for startup_32 can be just as for start_secondary onwards */
685 irq_ctx_init(cpu);
686#else
687 clear_tsk_thread_flag(idle, TIF_FORK);
688 initial_gs = per_cpu_offset(cpu);
689 per_cpu(kernel_stack, cpu) =
690 (unsigned long)task_stack_page(idle) -
691 KERNEL_STACK_OFFSET + THREAD_SIZE;
692#endif
693 early_gdt_descr.address = (unsigned long)get_cpu_gdt_table(cpu);
694 initial_code = (unsigned long)start_secondary;
695 stack_start = idle->thread.sp;
696
697 /* So we see what's up */
698 announce_cpu(cpu, apicid);
699
700 /*
701 * This grunge runs the startup process for
702 * the targeted processor.
703 */
704
705 atomic_set(&init_deasserted, 0);
706
707 if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
708
709 pr_debug("Setting warm reset code and vector.\n");
710
711 smpboot_setup_warm_reset_vector(start_ip);
712 /*
713 * Be paranoid about clearing APIC errors.
714 */
715 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid])) {
716 apic_write(APIC_ESR, 0);
717 apic_read(APIC_ESR);
718 }
719 }
720
721 /*
722 * Kick the secondary CPU. Use the method in the APIC driver
723 * if it's defined - or use an INIT boot APIC message otherwise:
724 */
725 if (apic->wakeup_secondary_cpu)
726 boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
727 else
728 boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
729
730 if (!boot_error) {
731 /*
732 * allow APs to start initializing.
733 */
734 pr_debug("Before Callout %d.\n", cpu);
735 cpumask_set_cpu(cpu, cpu_callout_mask);
736 pr_debug("After Callout %d.\n", cpu);
737
738 /*
739 * Wait 5s total for a response
740 */
741 for (timeout = 0; timeout < 50000; timeout++) {
742 if (cpumask_test_cpu(cpu, cpu_callin_mask))
743 break; /* It has booted */
744 udelay(100);
745 /*
746 * Allow other tasks to run while we wait for the
747 * AP to come online. This also gives a chance
748 * for the MTRR work(triggered by the AP coming online)
749 * to be completed in the stop machine context.
750 */
751 schedule();
752 }
753
754 if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
755 print_cpu_msr(&cpu_data(cpu));
756 pr_debug("CPU%d: has booted.\n", cpu);
757 } else {
758 boot_error = 1;
759 if (*trampoline_status == 0xA5A5A5A5)
760 /* trampoline started but...? */
761 pr_err("CPU%d: Stuck ??\n", cpu);
762 else
763 /* trampoline code not run */
764 pr_err("CPU%d: Not responding.\n", cpu);
765 if (apic->inquire_remote_apic)
766 apic->inquire_remote_apic(apicid);
767 }
768 }
769
770 if (boot_error) {
771 /* Try to put things back the way they were before ... */
772 numa_remove_cpu(cpu); /* was set by numa_add_cpu */
773
774 /* was set by do_boot_cpu() */
775 cpumask_clear_cpu(cpu, cpu_callout_mask);
776
777 /* was set by cpu_init() */
778 cpumask_clear_cpu(cpu, cpu_initialized_mask);
779
780 set_cpu_present(cpu, false);
781 per_cpu(x86_cpu_to_apicid, cpu) = BAD_APICID;
782 }
783
784 /* mark "stuck" area as not stuck */
785 *trampoline_status = 0;
786
787 if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
788 /*
789 * Cleanup possible dangling ends...
790 */
791 smpboot_restore_warm_reset_vector();
792 }
793 return boot_error;
794}
795
796int __cpuinit native_cpu_up(unsigned int cpu, struct task_struct *tidle)
797{
798 int apicid = apic->cpu_present_to_apicid(cpu);
799 unsigned long flags;
800 int err;
801
802 WARN_ON(irqs_disabled());
803
804 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
805
806 if (apicid == BAD_APICID || apicid == boot_cpu_physical_apicid ||
807 !physid_isset(apicid, phys_cpu_present_map) ||
808 !apic->apic_id_valid(apicid)) {
809 printk(KERN_ERR "%s: bad cpu %d\n", __func__, cpu);
810 return -EINVAL;
811 }
812
813 /*
814 * Already booted CPU?
815 */
816 if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
817 pr_debug("do_boot_cpu %d Already started\n", cpu);
818 return -ENOSYS;
819 }
820
821 /*
822 * Save current MTRR state in case it was changed since early boot
823 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
824 */
825 mtrr_save_state();
826
827 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
828
829 err = do_boot_cpu(apicid, cpu, tidle);
830 if (err) {
831 pr_debug("do_boot_cpu failed %d\n", err);
832 return -EIO;
833 }
834
835 /*
836 * Check TSC synchronization with the AP (keep irqs disabled
837 * while doing so):
838 */
839 local_irq_save(flags);
840 check_tsc_sync_source(cpu);
841 local_irq_restore(flags);
842
843 while (!cpu_online(cpu)) {
844 cpu_relax();
845 touch_nmi_watchdog();
846 }
847
848 return 0;
849}
850
851/**
852 * arch_disable_smp_support() - disables SMP support for x86 at runtime
853 */
854void arch_disable_smp_support(void)
855{
856 disable_ioapic_support();
857}
858
859/*
860 * Fall back to non SMP mode after errors.
861 *
862 * RED-PEN audit/test this more. I bet there is more state messed up here.
863 */
864static __init void disable_smp(void)
865{
866 init_cpu_present(cpumask_of(0));
867 init_cpu_possible(cpumask_of(0));
868 smpboot_clear_io_apic_irqs();
869
870 if (smp_found_config)
871 physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
872 else
873 physid_set_mask_of_physid(0, &phys_cpu_present_map);
874 cpumask_set_cpu(0, cpu_sibling_mask(0));
875 cpumask_set_cpu(0, cpu_core_mask(0));
876}
877
878/*
879 * Various sanity checks.
880 */
881static int __init smp_sanity_check(unsigned max_cpus)
882{
883 preempt_disable();
884
885#if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
886 if (def_to_bigsmp && nr_cpu_ids > 8) {
887 unsigned int cpu;
888 unsigned nr;
889
890 printk(KERN_WARNING
891 "More than 8 CPUs detected - skipping them.\n"
892 "Use CONFIG_X86_BIGSMP.\n");
893
894 nr = 0;
895 for_each_present_cpu(cpu) {
896 if (nr >= 8)
897 set_cpu_present(cpu, false);
898 nr++;
899 }
900
901 nr = 0;
902 for_each_possible_cpu(cpu) {
903 if (nr >= 8)
904 set_cpu_possible(cpu, false);
905 nr++;
906 }
907
908 nr_cpu_ids = 8;
909 }
910#endif
911
912 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
913 printk(KERN_WARNING
914 "weird, boot CPU (#%d) not listed by the BIOS.\n",
915 hard_smp_processor_id());
916
917 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
918 }
919
920 /*
921 * If we couldn't find an SMP configuration at boot time,
922 * get out of here now!
923 */
924 if (!smp_found_config && !acpi_lapic) {
925 preempt_enable();
926 printk(KERN_NOTICE "SMP motherboard not detected.\n");
927 disable_smp();
928 if (APIC_init_uniprocessor())
929 printk(KERN_NOTICE "Local APIC not detected."
930 " Using dummy APIC emulation.\n");
931 return -1;
932 }
933
934 /*
935 * Should not be necessary because the MP table should list the boot
936 * CPU too, but we do it for the sake of robustness anyway.
937 */
938 if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
939 printk(KERN_NOTICE
940 "weird, boot CPU (#%d) not listed by the BIOS.\n",
941 boot_cpu_physical_apicid);
942 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
943 }
944 preempt_enable();
945
946 /*
947 * If we couldn't find a local APIC, then get out of here now!
948 */
949 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) &&
950 !cpu_has_apic) {
951 if (!disable_apic) {
952 pr_err("BIOS bug, local APIC #%d not detected!...\n",
953 boot_cpu_physical_apicid);
954 pr_err("... forcing use of dummy APIC emulation."
955 "(tell your hw vendor)\n");
956 }
957 smpboot_clear_io_apic();
958 disable_ioapic_support();
959 return -1;
960 }
961
962 verify_local_APIC();
963
964 /*
965 * If SMP should be disabled, then really disable it!
966 */
967 if (!max_cpus) {
968 printk(KERN_INFO "SMP mode deactivated.\n");
969 smpboot_clear_io_apic();
970
971 connect_bsp_APIC();
972 setup_local_APIC();
973 bsp_end_local_APIC_setup();
974 return -1;
975 }
976
977 return 0;
978}
979
980static void __init smp_cpu_index_default(void)
981{
982 int i;
983 struct cpuinfo_x86 *c;
984
985 for_each_possible_cpu(i) {
986 c = &cpu_data(i);
987 /* mark all to hotplug */
988 c->cpu_index = nr_cpu_ids;
989 }
990}
991
992/*
993 * Prepare for SMP bootup. The MP table or ACPI has been read
994 * earlier. Just do some sanity checking here and enable APIC mode.
995 */
996void __init native_smp_prepare_cpus(unsigned int max_cpus)
997{
998 unsigned int i;
999
1000 preempt_disable();
1001 smp_cpu_index_default();
1002
1003 /*
1004 * Setup boot CPU information
1005 */
1006 smp_store_cpu_info(0); /* Final full version of the data */
1007 cpumask_copy(cpu_callin_mask, cpumask_of(0));
1008 mb();
1009
1010 current_thread_info()->cpu = 0; /* needed? */
1011 for_each_possible_cpu(i) {
1012 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1013 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1014 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1015 }
1016 set_cpu_sibling_map(0);
1017
1018
1019 if (smp_sanity_check(max_cpus) < 0) {
1020 printk(KERN_INFO "SMP disabled\n");
1021 disable_smp();
1022 goto out;
1023 }
1024
1025 default_setup_apic_routing();
1026
1027 preempt_disable();
1028 if (read_apic_id() != boot_cpu_physical_apicid) {
1029 panic("Boot APIC ID in local APIC unexpected (%d vs %d)",
1030 read_apic_id(), boot_cpu_physical_apicid);
1031 /* Or can we switch back to PIC here? */
1032 }
1033 preempt_enable();
1034
1035 connect_bsp_APIC();
1036
1037 /*
1038 * Switch from PIC to APIC mode.
1039 */
1040 setup_local_APIC();
1041
1042 /*
1043 * Enable IO APIC before setting up error vector
1044 */
1045 if (!skip_ioapic_setup && nr_ioapics)
1046 enable_IO_APIC();
1047
1048 bsp_end_local_APIC_setup();
1049
1050 if (apic->setup_portio_remap)
1051 apic->setup_portio_remap();
1052
1053 smpboot_setup_io_apic();
1054 /*
1055 * Set up local APIC timer on boot CPU.
1056 */
1057
1058 printk(KERN_INFO "CPU%d: ", 0);
1059 print_cpu_info(&cpu_data(0));
1060 x86_init.timers.setup_percpu_clockev();
1061
1062 if (is_uv_system())
1063 uv_system_init();
1064
1065 set_mtrr_aps_delayed_init();
1066out:
1067 preempt_enable();
1068}
1069
1070void arch_disable_nonboot_cpus_begin(void)
1071{
1072 /*
1073 * Avoid the smp alternatives switch during the disable_nonboot_cpus().
1074 * In the suspend path, we will be back in the SMP mode shortly anyways.
1075 */
1076 skip_smp_alternatives = true;
1077}
1078
1079void arch_disable_nonboot_cpus_end(void)
1080{
1081 skip_smp_alternatives = false;
1082}
1083
1084void arch_enable_nonboot_cpus_begin(void)
1085{
1086 set_mtrr_aps_delayed_init();
1087}
1088
1089void arch_enable_nonboot_cpus_end(void)
1090{
1091 mtrr_aps_init();
1092}
1093
1094/*
1095 * Early setup to make printk work.
1096 */
1097void __init native_smp_prepare_boot_cpu(void)
1098{
1099 int me = smp_processor_id();
1100 switch_to_new_gdt(me);
1101 /* already set me in cpu_online_mask in boot_cpu_init() */
1102 cpumask_set_cpu(me, cpu_callout_mask);
1103 per_cpu(cpu_state, me) = CPU_ONLINE;
1104}
1105
1106void __init native_smp_cpus_done(unsigned int max_cpus)
1107{
1108 pr_debug("Boot done.\n");
1109
1110 nmi_selftest();
1111 impress_friends();
1112#ifdef CONFIG_X86_IO_APIC
1113 setup_ioapic_dest();
1114#endif
1115 mtrr_aps_init();
1116}
1117
1118static int __initdata setup_possible_cpus = -1;
1119static int __init _setup_possible_cpus(char *str)
1120{
1121 get_option(&str, &setup_possible_cpus);
1122 return 0;
1123}
1124early_param("possible_cpus", _setup_possible_cpus);
1125
1126
1127/*
1128 * cpu_possible_mask should be static, it cannot change as cpu's
1129 * are onlined, or offlined. The reason is per-cpu data-structures
1130 * are allocated by some modules at init time, and dont expect to
1131 * do this dynamically on cpu arrival/departure.
1132 * cpu_present_mask on the other hand can change dynamically.
1133 * In case when cpu_hotplug is not compiled, then we resort to current
1134 * behaviour, which is cpu_possible == cpu_present.
1135 * - Ashok Raj
1136 *
1137 * Three ways to find out the number of additional hotplug CPUs:
1138 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1139 * - The user can overwrite it with possible_cpus=NUM
1140 * - Otherwise don't reserve additional CPUs.
1141 * We do this because additional CPUs waste a lot of memory.
1142 * -AK
1143 */
1144__init void prefill_possible_map(void)
1145{
1146 int i, possible;
1147
1148 /* no processor from mptable or madt */
1149 if (!num_processors)
1150 num_processors = 1;
1151
1152 i = setup_max_cpus ?: 1;
1153 if (setup_possible_cpus == -1) {
1154 possible = num_processors;
1155#ifdef CONFIG_HOTPLUG_CPU
1156 if (setup_max_cpus)
1157 possible += disabled_cpus;
1158#else
1159 if (possible > i)
1160 possible = i;
1161#endif
1162 } else
1163 possible = setup_possible_cpus;
1164
1165 total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1166
1167 /* nr_cpu_ids could be reduced via nr_cpus= */
1168 if (possible > nr_cpu_ids) {
1169 printk(KERN_WARNING
1170 "%d Processors exceeds NR_CPUS limit of %d\n",
1171 possible, nr_cpu_ids);
1172 possible = nr_cpu_ids;
1173 }
1174
1175#ifdef CONFIG_HOTPLUG_CPU
1176 if (!setup_max_cpus)
1177#endif
1178 if (possible > i) {
1179 printk(KERN_WARNING
1180 "%d Processors exceeds max_cpus limit of %u\n",
1181 possible, setup_max_cpus);
1182 possible = i;
1183 }
1184
1185 printk(KERN_INFO "SMP: Allowing %d CPUs, %d hotplug CPUs\n",
1186 possible, max_t(int, possible - num_processors, 0));
1187
1188 for (i = 0; i < possible; i++)
1189 set_cpu_possible(i, true);
1190 for (; i < NR_CPUS; i++)
1191 set_cpu_possible(i, false);
1192
1193 nr_cpu_ids = possible;
1194}
1195
1196#ifdef CONFIG_HOTPLUG_CPU
1197
1198static void remove_siblinginfo(int cpu)
1199{
1200 int sibling;
1201 struct cpuinfo_x86 *c = &cpu_data(cpu);
1202
1203 for_each_cpu(sibling, cpu_core_mask(cpu)) {
1204 cpumask_clear_cpu(cpu, cpu_core_mask(sibling));
1205 /*/
1206 * last thread sibling in this cpu core going down
1207 */
1208 if (cpumask_weight(cpu_sibling_mask(cpu)) == 1)
1209 cpu_data(sibling).booted_cores--;
1210 }
1211
1212 for_each_cpu(sibling, cpu_sibling_mask(cpu))
1213 cpumask_clear_cpu(cpu, cpu_sibling_mask(sibling));
1214 cpumask_clear(cpu_sibling_mask(cpu));
1215 cpumask_clear(cpu_core_mask(cpu));
1216 c->phys_proc_id = 0;
1217 c->cpu_core_id = 0;
1218 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1219}
1220
1221static void __ref remove_cpu_from_maps(int cpu)
1222{
1223 set_cpu_online(cpu, false);
1224 cpumask_clear_cpu(cpu, cpu_callout_mask);
1225 cpumask_clear_cpu(cpu, cpu_callin_mask);
1226 /* was set by cpu_init() */
1227 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1228 numa_remove_cpu(cpu);
1229}
1230
1231void cpu_disable_common(void)
1232{
1233 int cpu = smp_processor_id();
1234
1235 remove_siblinginfo(cpu);
1236
1237 /* It's now safe to remove this processor from the online map */
1238 lock_vector_lock();
1239 remove_cpu_from_maps(cpu);
1240 unlock_vector_lock();
1241 fixup_irqs();
1242}
1243
1244int native_cpu_disable(void)
1245{
1246 int cpu = smp_processor_id();
1247
1248 /*
1249 * Perhaps use cpufreq to drop frequency, but that could go
1250 * into generic code.
1251 *
1252 * We won't take down the boot processor on i386 due to some
1253 * interrupts only being able to be serviced by the BSP.
1254 * Especially so if we're not using an IOAPIC -zwane
1255 */
1256 if (cpu == 0)
1257 return -EBUSY;
1258
1259 clear_local_APIC();
1260
1261 cpu_disable_common();
1262 return 0;
1263}
1264
1265void native_cpu_die(unsigned int cpu)
1266{
1267 /* We don't do anything here: idle task is faking death itself. */
1268 unsigned int i;
1269
1270 for (i = 0; i < 10; i++) {
1271 /* They ack this in play_dead by setting CPU_DEAD */
1272 if (per_cpu(cpu_state, cpu) == CPU_DEAD) {
1273 if (system_state == SYSTEM_RUNNING)
1274 pr_info("CPU %u is now offline\n", cpu);
1275
1276 if (1 == num_online_cpus())
1277 alternatives_smp_switch(0);
1278 return;
1279 }
1280 msleep(100);
1281 }
1282 pr_err("CPU %u didn't die...\n", cpu);
1283}
1284
1285void play_dead_common(void)
1286{
1287 idle_task_exit();
1288 reset_lazy_tlbstate();
1289 amd_e400_remove_cpu(raw_smp_processor_id());
1290
1291 mb();
1292 /* Ack it */
1293 __this_cpu_write(cpu_state, CPU_DEAD);
1294
1295 /*
1296 * With physical CPU hotplug, we should halt the cpu
1297 */
1298 local_irq_disable();
1299}
1300
1301/*
1302 * We need to flush the caches before going to sleep, lest we have
1303 * dirty data in our caches when we come back up.
1304 */
1305static inline void mwait_play_dead(void)
1306{
1307 unsigned int eax, ebx, ecx, edx;
1308 unsigned int highest_cstate = 0;
1309 unsigned int highest_subcstate = 0;
1310 int i;
1311 void *mwait_ptr;
1312 struct cpuinfo_x86 *c = __this_cpu_ptr(&cpu_info);
1313
1314 if (!(this_cpu_has(X86_FEATURE_MWAIT) && mwait_usable(c)))
1315 return;
1316 if (!this_cpu_has(X86_FEATURE_CLFLSH))
1317 return;
1318 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1319 return;
1320
1321 eax = CPUID_MWAIT_LEAF;
1322 ecx = 0;
1323 native_cpuid(&eax, &ebx, &ecx, &edx);
1324
1325 /*
1326 * eax will be 0 if EDX enumeration is not valid.
1327 * Initialized below to cstate, sub_cstate value when EDX is valid.
1328 */
1329 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1330 eax = 0;
1331 } else {
1332 edx >>= MWAIT_SUBSTATE_SIZE;
1333 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1334 if (edx & MWAIT_SUBSTATE_MASK) {
1335 highest_cstate = i;
1336 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1337 }
1338 }
1339 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1340 (highest_subcstate - 1);
1341 }
1342
1343 /*
1344 * This should be a memory location in a cache line which is
1345 * unlikely to be touched by other processors. The actual
1346 * content is immaterial as it is not actually modified in any way.
1347 */
1348 mwait_ptr = ¤t_thread_info()->flags;
1349
1350 wbinvd();
1351
1352 while (1) {
1353 /*
1354 * The CLFLUSH is a workaround for erratum AAI65 for
1355 * the Xeon 7400 series. It's not clear it is actually
1356 * needed, but it should be harmless in either case.
1357 * The WBINVD is insufficient due to the spurious-wakeup
1358 * case where we return around the loop.
1359 */
1360 clflush(mwait_ptr);
1361 __monitor(mwait_ptr, 0, 0);
1362 mb();
1363 __mwait(eax, 0);
1364 }
1365}
1366
1367static inline void hlt_play_dead(void)
1368{
1369 if (__this_cpu_read(cpu_info.x86) >= 4)
1370 wbinvd();
1371
1372 while (1) {
1373 native_halt();
1374 }
1375}
1376
1377void native_play_dead(void)
1378{
1379 play_dead_common();
1380 tboot_shutdown(TB_SHUTDOWN_WFS);
1381
1382 mwait_play_dead(); /* Only returns on failure */
1383 if (cpuidle_play_dead())
1384 hlt_play_dead();
1385}
1386
1387#else /* ... !CONFIG_HOTPLUG_CPU */
1388int native_cpu_disable(void)
1389{
1390 return -ENOSYS;
1391}
1392
1393void native_cpu_die(unsigned int cpu)
1394{
1395 /* We said "no" in __cpu_disable */
1396 BUG();
1397}
1398
1399void native_play_dead(void)
1400{
1401 BUG();
1402}
1403
1404#endif
1/*
2 * x86 SMP booting functions
3 *
4 * (c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
5 * (c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com>
6 * Copyright 2001 Andi Kleen, SuSE Labs.
7 *
8 * Much of the core SMP work is based on previous work by Thomas Radke, to
9 * whom a great many thanks are extended.
10 *
11 * Thanks to Intel for making available several different Pentium,
12 * Pentium Pro and Pentium-II/Xeon MP machines.
13 * Original development of Linux SMP code supported by Caldera.
14 *
15 * This code is released under the GNU General Public License version 2 or
16 * later.
17 *
18 * Fixes
19 * Felix Koop : NR_CPUS used properly
20 * Jose Renau : Handle single CPU case.
21 * Alan Cox : By repeated request 8) - Total BogoMIPS report.
22 * Greg Wright : Fix for kernel stacks panic.
23 * Erich Boleyn : MP v1.4 and additional changes.
24 * Matthias Sattler : Changes for 2.1 kernel map.
25 * Michel Lespinasse : Changes for 2.1 kernel map.
26 * Michael Chastain : Change trampoline.S to gnu as.
27 * Alan Cox : Dumb bug: 'B' step PPro's are fine
28 * Ingo Molnar : Added APIC timers, based on code
29 * from Jose Renau
30 * Ingo Molnar : various cleanups and rewrites
31 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug.
32 * Maciej W. Rozycki : Bits for genuine 82489DX APICs
33 * Andi Kleen : Changed for SMP boot into long mode.
34 * Martin J. Bligh : Added support for multi-quad systems
35 * Dave Jones : Report invalid combinations of Athlon CPUs.
36 * Rusty Russell : Hacked into shape for new "hotplug" boot process.
37 * Andi Kleen : Converted to new state machine.
38 * Ashok Raj : CPU hotplug support
39 * Glauber Costa : i386 and x86_64 integration
40 */
41
42#include <linux/init.h>
43#include <linux/smp.h>
44#include <linux/module.h>
45#include <linux/sched.h>
46#include <linux/percpu.h>
47#include <linux/bootmem.h>
48#include <linux/err.h>
49#include <linux/nmi.h>
50#include <linux/tboot.h>
51#include <linux/stackprotector.h>
52#include <linux/gfp.h>
53
54#include <asm/acpi.h>
55#include <asm/desc.h>
56#include <asm/nmi.h>
57#include <asm/irq.h>
58#include <asm/idle.h>
59#include <asm/trampoline.h>
60#include <asm/cpu.h>
61#include <asm/numa.h>
62#include <asm/pgtable.h>
63#include <asm/tlbflush.h>
64#include <asm/mtrr.h>
65#include <asm/mwait.h>
66#include <asm/apic.h>
67#include <asm/io_apic.h>
68#include <asm/setup.h>
69#include <asm/uv/uv.h>
70#include <linux/mc146818rtc.h>
71
72#include <asm/smpboot_hooks.h>
73#include <asm/i8259.h>
74
75/* State of each CPU */
76DEFINE_PER_CPU(int, cpu_state) = { 0 };
77
78/* Store all idle threads, this can be reused instead of creating
79* a new thread. Also avoids complicated thread destroy functionality
80* for idle threads.
81*/
82#ifdef CONFIG_HOTPLUG_CPU
83/*
84 * Needed only for CONFIG_HOTPLUG_CPU because __cpuinitdata is
85 * removed after init for !CONFIG_HOTPLUG_CPU.
86 */
87static DEFINE_PER_CPU(struct task_struct *, idle_thread_array);
88#define get_idle_for_cpu(x) (per_cpu(idle_thread_array, x))
89#define set_idle_for_cpu(x, p) (per_cpu(idle_thread_array, x) = (p))
90
91/*
92 * We need this for trampoline_base protection from concurrent accesses when
93 * off- and onlining cores wildly.
94 */
95static DEFINE_MUTEX(x86_cpu_hotplug_driver_mutex);
96
97void cpu_hotplug_driver_lock(void)
98{
99 mutex_lock(&x86_cpu_hotplug_driver_mutex);
100}
101
102void cpu_hotplug_driver_unlock(void)
103{
104 mutex_unlock(&x86_cpu_hotplug_driver_mutex);
105}
106
107ssize_t arch_cpu_probe(const char *buf, size_t count) { return -1; }
108ssize_t arch_cpu_release(const char *buf, size_t count) { return -1; }
109#else
110static struct task_struct *idle_thread_array[NR_CPUS] __cpuinitdata ;
111#define get_idle_for_cpu(x) (idle_thread_array[(x)])
112#define set_idle_for_cpu(x, p) (idle_thread_array[(x)] = (p))
113#endif
114
115/* Number of siblings per CPU package */
116int smp_num_siblings = 1;
117EXPORT_SYMBOL(smp_num_siblings);
118
119/* Last level cache ID of each logical CPU */
120DEFINE_PER_CPU(u16, cpu_llc_id) = BAD_APICID;
121
122/* representing HT siblings of each logical CPU */
123DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
124EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
125
126/* representing HT and core siblings of each logical CPU */
127DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
128EXPORT_PER_CPU_SYMBOL(cpu_core_map);
129
130DEFINE_PER_CPU(cpumask_var_t, cpu_llc_shared_map);
131
132/* Per CPU bogomips and other parameters */
133DEFINE_PER_CPU_SHARED_ALIGNED(struct cpuinfo_x86, cpu_info);
134EXPORT_PER_CPU_SYMBOL(cpu_info);
135
136atomic_t init_deasserted;
137
138/*
139 * Report back to the Boot Processor.
140 * Running on AP.
141 */
142static void __cpuinit smp_callin(void)
143{
144 int cpuid, phys_id;
145 unsigned long timeout;
146
147 /*
148 * If waken up by an INIT in an 82489DX configuration
149 * we may get here before an INIT-deassert IPI reaches
150 * our local APIC. We have to wait for the IPI or we'll
151 * lock up on an APIC access.
152 */
153 if (apic->wait_for_init_deassert)
154 apic->wait_for_init_deassert(&init_deasserted);
155
156 /*
157 * (This works even if the APIC is not enabled.)
158 */
159 phys_id = read_apic_id();
160 cpuid = smp_processor_id();
161 if (cpumask_test_cpu(cpuid, cpu_callin_mask)) {
162 panic("%s: phys CPU#%d, CPU#%d already present??\n", __func__,
163 phys_id, cpuid);
164 }
165 pr_debug("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid, phys_id);
166
167 /*
168 * STARTUP IPIs are fragile beasts as they might sometimes
169 * trigger some glue motherboard logic. Complete APIC bus
170 * silence for 1 second, this overestimates the time the
171 * boot CPU is spending to send the up to 2 STARTUP IPIs
172 * by a factor of two. This should be enough.
173 */
174
175 /*
176 * Waiting 2s total for startup (udelay is not yet working)
177 */
178 timeout = jiffies + 2*HZ;
179 while (time_before(jiffies, timeout)) {
180 /*
181 * Has the boot CPU finished it's STARTUP sequence?
182 */
183 if (cpumask_test_cpu(cpuid, cpu_callout_mask))
184 break;
185 cpu_relax();
186 }
187
188 if (!time_before(jiffies, timeout)) {
189 panic("%s: CPU%d started up but did not get a callout!\n",
190 __func__, cpuid);
191 }
192
193 /*
194 * the boot CPU has finished the init stage and is spinning
195 * on callin_map until we finish. We are free to set up this
196 * CPU, first the APIC. (this is probably redundant on most
197 * boards)
198 */
199
200 pr_debug("CALLIN, before setup_local_APIC().\n");
201 if (apic->smp_callin_clear_local_apic)
202 apic->smp_callin_clear_local_apic();
203 setup_local_APIC();
204 end_local_APIC_setup();
205
206 /*
207 * Need to setup vector mappings before we enable interrupts.
208 */
209 setup_vector_irq(smp_processor_id());
210 /*
211 * Get our bogomips.
212 *
213 * Need to enable IRQs because it can take longer and then
214 * the NMI watchdog might kill us.
215 */
216 local_irq_enable();
217 calibrate_delay();
218 local_irq_disable();
219 pr_debug("Stack at about %p\n", &cpuid);
220
221 /*
222 * Save our processor parameters
223 */
224 smp_store_cpu_info(cpuid);
225
226 /*
227 * This must be done before setting cpu_online_mask
228 * or calling notify_cpu_starting.
229 */
230 set_cpu_sibling_map(raw_smp_processor_id());
231 wmb();
232
233 notify_cpu_starting(cpuid);
234
235 /*
236 * Allow the master to continue.
237 */
238 cpumask_set_cpu(cpuid, cpu_callin_mask);
239}
240
241/*
242 * Activate a secondary processor.
243 */
244notrace static void __cpuinit start_secondary(void *unused)
245{
246 /*
247 * Don't put *anything* before cpu_init(), SMP booting is too
248 * fragile that we want to limit the things done here to the
249 * most necessary things.
250 */
251 cpu_init();
252 preempt_disable();
253 smp_callin();
254
255#ifdef CONFIG_X86_32
256 /* switch away from the initial page table */
257 load_cr3(swapper_pg_dir);
258 __flush_tlb_all();
259#endif
260
261 /* otherwise gcc will move up smp_processor_id before the cpu_init */
262 barrier();
263 /*
264 * Check TSC synchronization with the BP:
265 */
266 check_tsc_sync_target();
267
268 /*
269 * We need to hold call_lock, so there is no inconsistency
270 * between the time smp_call_function() determines number of
271 * IPI recipients, and the time when the determination is made
272 * for which cpus receive the IPI. Holding this
273 * lock helps us to not include this cpu in a currently in progress
274 * smp_call_function().
275 *
276 * We need to hold vector_lock so there the set of online cpus
277 * does not change while we are assigning vectors to cpus. Holding
278 * this lock ensures we don't half assign or remove an irq from a cpu.
279 */
280 ipi_call_lock();
281 lock_vector_lock();
282 set_cpu_online(smp_processor_id(), true);
283 unlock_vector_lock();
284 ipi_call_unlock();
285 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
286 x86_platform.nmi_init();
287
288 /*
289 * Wait until the cpu which brought this one up marked it
290 * online before enabling interrupts. If we don't do that then
291 * we can end up waking up the softirq thread before this cpu
292 * reached the active state, which makes the scheduler unhappy
293 * and schedule the softirq thread on the wrong cpu. This is
294 * only observable with forced threaded interrupts, but in
295 * theory it could also happen w/o them. It's just way harder
296 * to achieve.
297 */
298 while (!cpumask_test_cpu(smp_processor_id(), cpu_active_mask))
299 cpu_relax();
300
301 /* enable local interrupts */
302 local_irq_enable();
303
304 /* to prevent fake stack check failure in clock setup */
305 boot_init_stack_canary();
306
307 x86_cpuinit.setup_percpu_clockev();
308
309 wmb();
310 cpu_idle();
311}
312
313/*
314 * The bootstrap kernel entry code has set these up. Save them for
315 * a given CPU
316 */
317
318void __cpuinit smp_store_cpu_info(int id)
319{
320 struct cpuinfo_x86 *c = &cpu_data(id);
321
322 *c = boot_cpu_data;
323 c->cpu_index = id;
324 if (id != 0)
325 identify_secondary_cpu(c);
326}
327
328static void __cpuinit link_thread_siblings(int cpu1, int cpu2)
329{
330 cpumask_set_cpu(cpu1, cpu_sibling_mask(cpu2));
331 cpumask_set_cpu(cpu2, cpu_sibling_mask(cpu1));
332 cpumask_set_cpu(cpu1, cpu_core_mask(cpu2));
333 cpumask_set_cpu(cpu2, cpu_core_mask(cpu1));
334 cpumask_set_cpu(cpu1, cpu_llc_shared_mask(cpu2));
335 cpumask_set_cpu(cpu2, cpu_llc_shared_mask(cpu1));
336}
337
338
339void __cpuinit set_cpu_sibling_map(int cpu)
340{
341 int i;
342 struct cpuinfo_x86 *c = &cpu_data(cpu);
343
344 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
345
346 if (smp_num_siblings > 1) {
347 for_each_cpu(i, cpu_sibling_setup_mask) {
348 struct cpuinfo_x86 *o = &cpu_data(i);
349
350 if (cpu_has(c, X86_FEATURE_TOPOEXT)) {
351 if (c->phys_proc_id == o->phys_proc_id &&
352 per_cpu(cpu_llc_id, cpu) == per_cpu(cpu_llc_id, i) &&
353 c->compute_unit_id == o->compute_unit_id)
354 link_thread_siblings(cpu, i);
355 } else if (c->phys_proc_id == o->phys_proc_id &&
356 c->cpu_core_id == o->cpu_core_id) {
357 link_thread_siblings(cpu, i);
358 }
359 }
360 } else {
361 cpumask_set_cpu(cpu, cpu_sibling_mask(cpu));
362 }
363
364 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
365
366 if (__this_cpu_read(cpu_info.x86_max_cores) == 1) {
367 cpumask_copy(cpu_core_mask(cpu), cpu_sibling_mask(cpu));
368 c->booted_cores = 1;
369 return;
370 }
371
372 for_each_cpu(i, cpu_sibling_setup_mask) {
373 if (per_cpu(cpu_llc_id, cpu) != BAD_APICID &&
374 per_cpu(cpu_llc_id, cpu) == per_cpu(cpu_llc_id, i)) {
375 cpumask_set_cpu(i, cpu_llc_shared_mask(cpu));
376 cpumask_set_cpu(cpu, cpu_llc_shared_mask(i));
377 }
378 if (c->phys_proc_id == cpu_data(i).phys_proc_id) {
379 cpumask_set_cpu(i, cpu_core_mask(cpu));
380 cpumask_set_cpu(cpu, cpu_core_mask(i));
381 /*
382 * Does this new cpu bringup a new core?
383 */
384 if (cpumask_weight(cpu_sibling_mask(cpu)) == 1) {
385 /*
386 * for each core in package, increment
387 * the booted_cores for this new cpu
388 */
389 if (cpumask_first(cpu_sibling_mask(i)) == i)
390 c->booted_cores++;
391 /*
392 * increment the core count for all
393 * the other cpus in this package
394 */
395 if (i != cpu)
396 cpu_data(i).booted_cores++;
397 } else if (i != cpu && !c->booted_cores)
398 c->booted_cores = cpu_data(i).booted_cores;
399 }
400 }
401}
402
403/* maps the cpu to the sched domain representing multi-core */
404const struct cpumask *cpu_coregroup_mask(int cpu)
405{
406 struct cpuinfo_x86 *c = &cpu_data(cpu);
407 /*
408 * For perf, we return last level cache shared map.
409 * And for power savings, we return cpu_core_map
410 */
411 if ((sched_mc_power_savings || sched_smt_power_savings) &&
412 !(cpu_has(c, X86_FEATURE_AMD_DCM)))
413 return cpu_core_mask(cpu);
414 else
415 return cpu_llc_shared_mask(cpu);
416}
417
418static void impress_friends(void)
419{
420 int cpu;
421 unsigned long bogosum = 0;
422 /*
423 * Allow the user to impress friends.
424 */
425 pr_debug("Before bogomips.\n");
426 for_each_possible_cpu(cpu)
427 if (cpumask_test_cpu(cpu, cpu_callout_mask))
428 bogosum += cpu_data(cpu).loops_per_jiffy;
429 printk(KERN_INFO
430 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
431 num_online_cpus(),
432 bogosum/(500000/HZ),
433 (bogosum/(5000/HZ))%100);
434
435 pr_debug("Before bogocount - setting activated=1.\n");
436}
437
438void __inquire_remote_apic(int apicid)
439{
440 unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
441 const char * const names[] = { "ID", "VERSION", "SPIV" };
442 int timeout;
443 u32 status;
444
445 printk(KERN_INFO "Inquiring remote APIC 0x%x...\n", apicid);
446
447 for (i = 0; i < ARRAY_SIZE(regs); i++) {
448 printk(KERN_INFO "... APIC 0x%x %s: ", apicid, names[i]);
449
450 /*
451 * Wait for idle.
452 */
453 status = safe_apic_wait_icr_idle();
454 if (status)
455 printk(KERN_CONT
456 "a previous APIC delivery may have failed\n");
457
458 apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
459
460 timeout = 0;
461 do {
462 udelay(100);
463 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
464 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
465
466 switch (status) {
467 case APIC_ICR_RR_VALID:
468 status = apic_read(APIC_RRR);
469 printk(KERN_CONT "%08x\n", status);
470 break;
471 default:
472 printk(KERN_CONT "failed\n");
473 }
474 }
475}
476
477/*
478 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
479 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
480 * won't ... remember to clear down the APIC, etc later.
481 */
482int __cpuinit
483wakeup_secondary_cpu_via_nmi(int logical_apicid, unsigned long start_eip)
484{
485 unsigned long send_status, accept_status = 0;
486 int maxlvt;
487
488 /* Target chip */
489 /* Boot on the stack */
490 /* Kick the second */
491 apic_icr_write(APIC_DM_NMI | apic->dest_logical, logical_apicid);
492
493 pr_debug("Waiting for send to finish...\n");
494 send_status = safe_apic_wait_icr_idle();
495
496 /*
497 * Give the other CPU some time to accept the IPI.
498 */
499 udelay(200);
500 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid])) {
501 maxlvt = lapic_get_maxlvt();
502 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
503 apic_write(APIC_ESR, 0);
504 accept_status = (apic_read(APIC_ESR) & 0xEF);
505 }
506 pr_debug("NMI sent.\n");
507
508 if (send_status)
509 printk(KERN_ERR "APIC never delivered???\n");
510 if (accept_status)
511 printk(KERN_ERR "APIC delivery error (%lx).\n", accept_status);
512
513 return (send_status | accept_status);
514}
515
516static int __cpuinit
517wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
518{
519 unsigned long send_status, accept_status = 0;
520 int maxlvt, num_starts, j;
521
522 maxlvt = lapic_get_maxlvt();
523
524 /*
525 * Be paranoid about clearing APIC errors.
526 */
527 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
528 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
529 apic_write(APIC_ESR, 0);
530 apic_read(APIC_ESR);
531 }
532
533 pr_debug("Asserting INIT.\n");
534
535 /*
536 * Turn INIT on target chip
537 */
538 /*
539 * Send IPI
540 */
541 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
542 phys_apicid);
543
544 pr_debug("Waiting for send to finish...\n");
545 send_status = safe_apic_wait_icr_idle();
546
547 mdelay(10);
548
549 pr_debug("Deasserting INIT.\n");
550
551 /* Target chip */
552 /* Send IPI */
553 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
554
555 pr_debug("Waiting for send to finish...\n");
556 send_status = safe_apic_wait_icr_idle();
557
558 mb();
559 atomic_set(&init_deasserted, 1);
560
561 /*
562 * Should we send STARTUP IPIs ?
563 *
564 * Determine this based on the APIC version.
565 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
566 */
567 if (APIC_INTEGRATED(apic_version[phys_apicid]))
568 num_starts = 2;
569 else
570 num_starts = 0;
571
572 /*
573 * Paravirt / VMI wants a startup IPI hook here to set up the
574 * target processor state.
575 */
576 startup_ipi_hook(phys_apicid, (unsigned long) start_secondary,
577 stack_start);
578
579 /*
580 * Run STARTUP IPI loop.
581 */
582 pr_debug("#startup loops: %d.\n", num_starts);
583
584 for (j = 1; j <= num_starts; j++) {
585 pr_debug("Sending STARTUP #%d.\n", j);
586 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
587 apic_write(APIC_ESR, 0);
588 apic_read(APIC_ESR);
589 pr_debug("After apic_write.\n");
590
591 /*
592 * STARTUP IPI
593 */
594
595 /* Target chip */
596 /* Boot on the stack */
597 /* Kick the second */
598 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
599 phys_apicid);
600
601 /*
602 * Give the other CPU some time to accept the IPI.
603 */
604 udelay(300);
605
606 pr_debug("Startup point 1.\n");
607
608 pr_debug("Waiting for send to finish...\n");
609 send_status = safe_apic_wait_icr_idle();
610
611 /*
612 * Give the other CPU some time to accept the IPI.
613 */
614 udelay(200);
615 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
616 apic_write(APIC_ESR, 0);
617 accept_status = (apic_read(APIC_ESR) & 0xEF);
618 if (send_status || accept_status)
619 break;
620 }
621 pr_debug("After Startup.\n");
622
623 if (send_status)
624 printk(KERN_ERR "APIC never delivered???\n");
625 if (accept_status)
626 printk(KERN_ERR "APIC delivery error (%lx).\n", accept_status);
627
628 return (send_status | accept_status);
629}
630
631struct create_idle {
632 struct work_struct work;
633 struct task_struct *idle;
634 struct completion done;
635 int cpu;
636};
637
638static void __cpuinit do_fork_idle(struct work_struct *work)
639{
640 struct create_idle *c_idle =
641 container_of(work, struct create_idle, work);
642
643 c_idle->idle = fork_idle(c_idle->cpu);
644 complete(&c_idle->done);
645}
646
647/* reduce the number of lines printed when booting a large cpu count system */
648static void __cpuinit announce_cpu(int cpu, int apicid)
649{
650 static int current_node = -1;
651 int node = early_cpu_to_node(cpu);
652
653 if (system_state == SYSTEM_BOOTING) {
654 if (node != current_node) {
655 if (current_node > (-1))
656 pr_cont(" Ok.\n");
657 current_node = node;
658 pr_info("Booting Node %3d, Processors ", node);
659 }
660 pr_cont(" #%d%s", cpu, cpu == (nr_cpu_ids - 1) ? " Ok.\n" : "");
661 return;
662 } else
663 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
664 node, cpu, apicid);
665}
666
667/*
668 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
669 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
670 * Returns zero if CPU booted OK, else error code from
671 * ->wakeup_secondary_cpu.
672 */
673static int __cpuinit do_boot_cpu(int apicid, int cpu)
674{
675 unsigned long boot_error = 0;
676 unsigned long start_ip;
677 int timeout;
678 struct create_idle c_idle = {
679 .cpu = cpu,
680 .done = COMPLETION_INITIALIZER_ONSTACK(c_idle.done),
681 };
682
683 INIT_WORK_ONSTACK(&c_idle.work, do_fork_idle);
684
685 alternatives_smp_switch(1);
686
687 c_idle.idle = get_idle_for_cpu(cpu);
688
689 /*
690 * We can't use kernel_thread since we must avoid to
691 * reschedule the child.
692 */
693 if (c_idle.idle) {
694 c_idle.idle->thread.sp = (unsigned long) (((struct pt_regs *)
695 (THREAD_SIZE + task_stack_page(c_idle.idle))) - 1);
696 init_idle(c_idle.idle, cpu);
697 goto do_rest;
698 }
699
700 schedule_work(&c_idle.work);
701 wait_for_completion(&c_idle.done);
702
703 if (IS_ERR(c_idle.idle)) {
704 printk("failed fork for CPU %d\n", cpu);
705 destroy_work_on_stack(&c_idle.work);
706 return PTR_ERR(c_idle.idle);
707 }
708
709 set_idle_for_cpu(cpu, c_idle.idle);
710do_rest:
711 per_cpu(current_task, cpu) = c_idle.idle;
712#ifdef CONFIG_X86_32
713 /* Stack for startup_32 can be just as for start_secondary onwards */
714 irq_ctx_init(cpu);
715#else
716 clear_tsk_thread_flag(c_idle.idle, TIF_FORK);
717 initial_gs = per_cpu_offset(cpu);
718 per_cpu(kernel_stack, cpu) =
719 (unsigned long)task_stack_page(c_idle.idle) -
720 KERNEL_STACK_OFFSET + THREAD_SIZE;
721#endif
722 early_gdt_descr.address = (unsigned long)get_cpu_gdt_table(cpu);
723 initial_code = (unsigned long)start_secondary;
724 stack_start = c_idle.idle->thread.sp;
725
726 /* start_ip had better be page-aligned! */
727 start_ip = trampoline_address();
728
729 /* So we see what's up */
730 announce_cpu(cpu, apicid);
731
732 /*
733 * This grunge runs the startup process for
734 * the targeted processor.
735 */
736
737 printk(KERN_DEBUG "smpboot cpu %d: start_ip = %lx\n", cpu, start_ip);
738
739 atomic_set(&init_deasserted, 0);
740
741 if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
742
743 pr_debug("Setting warm reset code and vector.\n");
744
745 smpboot_setup_warm_reset_vector(start_ip);
746 /*
747 * Be paranoid about clearing APIC errors.
748 */
749 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid])) {
750 apic_write(APIC_ESR, 0);
751 apic_read(APIC_ESR);
752 }
753 }
754
755 /*
756 * Kick the secondary CPU. Use the method in the APIC driver
757 * if it's defined - or use an INIT boot APIC message otherwise:
758 */
759 if (apic->wakeup_secondary_cpu)
760 boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
761 else
762 boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
763
764 if (!boot_error) {
765 /*
766 * allow APs to start initializing.
767 */
768 pr_debug("Before Callout %d.\n", cpu);
769 cpumask_set_cpu(cpu, cpu_callout_mask);
770 pr_debug("After Callout %d.\n", cpu);
771
772 /*
773 * Wait 5s total for a response
774 */
775 for (timeout = 0; timeout < 50000; timeout++) {
776 if (cpumask_test_cpu(cpu, cpu_callin_mask))
777 break; /* It has booted */
778 udelay(100);
779 /*
780 * Allow other tasks to run while we wait for the
781 * AP to come online. This also gives a chance
782 * for the MTRR work(triggered by the AP coming online)
783 * to be completed in the stop machine context.
784 */
785 schedule();
786 }
787
788 if (cpumask_test_cpu(cpu, cpu_callin_mask))
789 pr_debug("CPU%d: has booted.\n", cpu);
790 else {
791 boot_error = 1;
792 if (*(volatile u32 *)TRAMPOLINE_SYM(trampoline_status)
793 == 0xA5A5A5A5)
794 /* trampoline started but...? */
795 pr_err("CPU%d: Stuck ??\n", cpu);
796 else
797 /* trampoline code not run */
798 pr_err("CPU%d: Not responding.\n", cpu);
799 if (apic->inquire_remote_apic)
800 apic->inquire_remote_apic(apicid);
801 }
802 }
803
804 if (boot_error) {
805 /* Try to put things back the way they were before ... */
806 numa_remove_cpu(cpu); /* was set by numa_add_cpu */
807
808 /* was set by do_boot_cpu() */
809 cpumask_clear_cpu(cpu, cpu_callout_mask);
810
811 /* was set by cpu_init() */
812 cpumask_clear_cpu(cpu, cpu_initialized_mask);
813
814 set_cpu_present(cpu, false);
815 per_cpu(x86_cpu_to_apicid, cpu) = BAD_APICID;
816 }
817
818 /* mark "stuck" area as not stuck */
819 *(volatile u32 *)TRAMPOLINE_SYM(trampoline_status) = 0;
820
821 if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
822 /*
823 * Cleanup possible dangling ends...
824 */
825 smpboot_restore_warm_reset_vector();
826 }
827
828 destroy_work_on_stack(&c_idle.work);
829 return boot_error;
830}
831
832int __cpuinit native_cpu_up(unsigned int cpu)
833{
834 int apicid = apic->cpu_present_to_apicid(cpu);
835 unsigned long flags;
836 int err;
837
838 WARN_ON(irqs_disabled());
839
840 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
841
842 if (apicid == BAD_APICID || apicid == boot_cpu_physical_apicid ||
843 !physid_isset(apicid, phys_cpu_present_map)) {
844 printk(KERN_ERR "%s: bad cpu %d\n", __func__, cpu);
845 return -EINVAL;
846 }
847
848 /*
849 * Already booted CPU?
850 */
851 if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
852 pr_debug("do_boot_cpu %d Already started\n", cpu);
853 return -ENOSYS;
854 }
855
856 /*
857 * Save current MTRR state in case it was changed since early boot
858 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
859 */
860 mtrr_save_state();
861
862 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
863
864 err = do_boot_cpu(apicid, cpu);
865 if (err) {
866 pr_debug("do_boot_cpu failed %d\n", err);
867 return -EIO;
868 }
869
870 /*
871 * Check TSC synchronization with the AP (keep irqs disabled
872 * while doing so):
873 */
874 local_irq_save(flags);
875 check_tsc_sync_source(cpu);
876 local_irq_restore(flags);
877
878 while (!cpu_online(cpu)) {
879 cpu_relax();
880 touch_nmi_watchdog();
881 }
882
883 return 0;
884}
885
886/**
887 * arch_disable_smp_support() - disables SMP support for x86 at runtime
888 */
889void arch_disable_smp_support(void)
890{
891 disable_ioapic_support();
892}
893
894/*
895 * Fall back to non SMP mode after errors.
896 *
897 * RED-PEN audit/test this more. I bet there is more state messed up here.
898 */
899static __init void disable_smp(void)
900{
901 init_cpu_present(cpumask_of(0));
902 init_cpu_possible(cpumask_of(0));
903 smpboot_clear_io_apic_irqs();
904
905 if (smp_found_config)
906 physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
907 else
908 physid_set_mask_of_physid(0, &phys_cpu_present_map);
909 cpumask_set_cpu(0, cpu_sibling_mask(0));
910 cpumask_set_cpu(0, cpu_core_mask(0));
911}
912
913/*
914 * Various sanity checks.
915 */
916static int __init smp_sanity_check(unsigned max_cpus)
917{
918 preempt_disable();
919
920#if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
921 if (def_to_bigsmp && nr_cpu_ids > 8) {
922 unsigned int cpu;
923 unsigned nr;
924
925 printk(KERN_WARNING
926 "More than 8 CPUs detected - skipping them.\n"
927 "Use CONFIG_X86_BIGSMP.\n");
928
929 nr = 0;
930 for_each_present_cpu(cpu) {
931 if (nr >= 8)
932 set_cpu_present(cpu, false);
933 nr++;
934 }
935
936 nr = 0;
937 for_each_possible_cpu(cpu) {
938 if (nr >= 8)
939 set_cpu_possible(cpu, false);
940 nr++;
941 }
942
943 nr_cpu_ids = 8;
944 }
945#endif
946
947 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
948 printk(KERN_WARNING
949 "weird, boot CPU (#%d) not listed by the BIOS.\n",
950 hard_smp_processor_id());
951
952 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
953 }
954
955 /*
956 * If we couldn't find an SMP configuration at boot time,
957 * get out of here now!
958 */
959 if (!smp_found_config && !acpi_lapic) {
960 preempt_enable();
961 printk(KERN_NOTICE "SMP motherboard not detected.\n");
962 disable_smp();
963 if (APIC_init_uniprocessor())
964 printk(KERN_NOTICE "Local APIC not detected."
965 " Using dummy APIC emulation.\n");
966 return -1;
967 }
968
969 /*
970 * Should not be necessary because the MP table should list the boot
971 * CPU too, but we do it for the sake of robustness anyway.
972 */
973 if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
974 printk(KERN_NOTICE
975 "weird, boot CPU (#%d) not listed by the BIOS.\n",
976 boot_cpu_physical_apicid);
977 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
978 }
979 preempt_enable();
980
981 /*
982 * If we couldn't find a local APIC, then get out of here now!
983 */
984 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) &&
985 !cpu_has_apic) {
986 if (!disable_apic) {
987 pr_err("BIOS bug, local APIC #%d not detected!...\n",
988 boot_cpu_physical_apicid);
989 pr_err("... forcing use of dummy APIC emulation."
990 "(tell your hw vendor)\n");
991 }
992 smpboot_clear_io_apic();
993 disable_ioapic_support();
994 return -1;
995 }
996
997 verify_local_APIC();
998
999 /*
1000 * If SMP should be disabled, then really disable it!
1001 */
1002 if (!max_cpus) {
1003 printk(KERN_INFO "SMP mode deactivated.\n");
1004 smpboot_clear_io_apic();
1005
1006 connect_bsp_APIC();
1007 setup_local_APIC();
1008 bsp_end_local_APIC_setup();
1009 return -1;
1010 }
1011
1012 return 0;
1013}
1014
1015static void __init smp_cpu_index_default(void)
1016{
1017 int i;
1018 struct cpuinfo_x86 *c;
1019
1020 for_each_possible_cpu(i) {
1021 c = &cpu_data(i);
1022 /* mark all to hotplug */
1023 c->cpu_index = nr_cpu_ids;
1024 }
1025}
1026
1027/*
1028 * Prepare for SMP bootup. The MP table or ACPI has been read
1029 * earlier. Just do some sanity checking here and enable APIC mode.
1030 */
1031void __init native_smp_prepare_cpus(unsigned int max_cpus)
1032{
1033 unsigned int i;
1034
1035 preempt_disable();
1036 smp_cpu_index_default();
1037
1038 /*
1039 * Setup boot CPU information
1040 */
1041 smp_store_cpu_info(0); /* Final full version of the data */
1042 cpumask_copy(cpu_callin_mask, cpumask_of(0));
1043 mb();
1044
1045 current_thread_info()->cpu = 0; /* needed? */
1046 for_each_possible_cpu(i) {
1047 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1048 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1049 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1050 }
1051 set_cpu_sibling_map(0);
1052
1053
1054 if (smp_sanity_check(max_cpus) < 0) {
1055 printk(KERN_INFO "SMP disabled\n");
1056 disable_smp();
1057 goto out;
1058 }
1059
1060 default_setup_apic_routing();
1061
1062 preempt_disable();
1063 if (read_apic_id() != boot_cpu_physical_apicid) {
1064 panic("Boot APIC ID in local APIC unexpected (%d vs %d)",
1065 read_apic_id(), boot_cpu_physical_apicid);
1066 /* Or can we switch back to PIC here? */
1067 }
1068 preempt_enable();
1069
1070 connect_bsp_APIC();
1071
1072 /*
1073 * Switch from PIC to APIC mode.
1074 */
1075 setup_local_APIC();
1076
1077 /*
1078 * Enable IO APIC before setting up error vector
1079 */
1080 if (!skip_ioapic_setup && nr_ioapics)
1081 enable_IO_APIC();
1082
1083 bsp_end_local_APIC_setup();
1084
1085 if (apic->setup_portio_remap)
1086 apic->setup_portio_remap();
1087
1088 smpboot_setup_io_apic();
1089 /*
1090 * Set up local APIC timer on boot CPU.
1091 */
1092
1093 printk(KERN_INFO "CPU%d: ", 0);
1094 print_cpu_info(&cpu_data(0));
1095 x86_init.timers.setup_percpu_clockev();
1096
1097 if (is_uv_system())
1098 uv_system_init();
1099
1100 set_mtrr_aps_delayed_init();
1101out:
1102 preempt_enable();
1103}
1104
1105void arch_disable_nonboot_cpus_begin(void)
1106{
1107 /*
1108 * Avoid the smp alternatives switch during the disable_nonboot_cpus().
1109 * In the suspend path, we will be back in the SMP mode shortly anyways.
1110 */
1111 skip_smp_alternatives = true;
1112}
1113
1114void arch_disable_nonboot_cpus_end(void)
1115{
1116 skip_smp_alternatives = false;
1117}
1118
1119void arch_enable_nonboot_cpus_begin(void)
1120{
1121 set_mtrr_aps_delayed_init();
1122}
1123
1124void arch_enable_nonboot_cpus_end(void)
1125{
1126 mtrr_aps_init();
1127}
1128
1129/*
1130 * Early setup to make printk work.
1131 */
1132void __init native_smp_prepare_boot_cpu(void)
1133{
1134 int me = smp_processor_id();
1135 switch_to_new_gdt(me);
1136 /* already set me in cpu_online_mask in boot_cpu_init() */
1137 cpumask_set_cpu(me, cpu_callout_mask);
1138 per_cpu(cpu_state, me) = CPU_ONLINE;
1139}
1140
1141void __init native_smp_cpus_done(unsigned int max_cpus)
1142{
1143 pr_debug("Boot done.\n");
1144
1145 impress_friends();
1146#ifdef CONFIG_X86_IO_APIC
1147 setup_ioapic_dest();
1148#endif
1149 mtrr_aps_init();
1150}
1151
1152static int __initdata setup_possible_cpus = -1;
1153static int __init _setup_possible_cpus(char *str)
1154{
1155 get_option(&str, &setup_possible_cpus);
1156 return 0;
1157}
1158early_param("possible_cpus", _setup_possible_cpus);
1159
1160
1161/*
1162 * cpu_possible_mask should be static, it cannot change as cpu's
1163 * are onlined, or offlined. The reason is per-cpu data-structures
1164 * are allocated by some modules at init time, and dont expect to
1165 * do this dynamically on cpu arrival/departure.
1166 * cpu_present_mask on the other hand can change dynamically.
1167 * In case when cpu_hotplug is not compiled, then we resort to current
1168 * behaviour, which is cpu_possible == cpu_present.
1169 * - Ashok Raj
1170 *
1171 * Three ways to find out the number of additional hotplug CPUs:
1172 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1173 * - The user can overwrite it with possible_cpus=NUM
1174 * - Otherwise don't reserve additional CPUs.
1175 * We do this because additional CPUs waste a lot of memory.
1176 * -AK
1177 */
1178__init void prefill_possible_map(void)
1179{
1180 int i, possible;
1181
1182 /* no processor from mptable or madt */
1183 if (!num_processors)
1184 num_processors = 1;
1185
1186 i = setup_max_cpus ?: 1;
1187 if (setup_possible_cpus == -1) {
1188 possible = num_processors;
1189#ifdef CONFIG_HOTPLUG_CPU
1190 if (setup_max_cpus)
1191 possible += disabled_cpus;
1192#else
1193 if (possible > i)
1194 possible = i;
1195#endif
1196 } else
1197 possible = setup_possible_cpus;
1198
1199 total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1200
1201 /* nr_cpu_ids could be reduced via nr_cpus= */
1202 if (possible > nr_cpu_ids) {
1203 printk(KERN_WARNING
1204 "%d Processors exceeds NR_CPUS limit of %d\n",
1205 possible, nr_cpu_ids);
1206 possible = nr_cpu_ids;
1207 }
1208
1209#ifdef CONFIG_HOTPLUG_CPU
1210 if (!setup_max_cpus)
1211#endif
1212 if (possible > i) {
1213 printk(KERN_WARNING
1214 "%d Processors exceeds max_cpus limit of %u\n",
1215 possible, setup_max_cpus);
1216 possible = i;
1217 }
1218
1219 printk(KERN_INFO "SMP: Allowing %d CPUs, %d hotplug CPUs\n",
1220 possible, max_t(int, possible - num_processors, 0));
1221
1222 for (i = 0; i < possible; i++)
1223 set_cpu_possible(i, true);
1224 for (; i < NR_CPUS; i++)
1225 set_cpu_possible(i, false);
1226
1227 nr_cpu_ids = possible;
1228}
1229
1230#ifdef CONFIG_HOTPLUG_CPU
1231
1232static void remove_siblinginfo(int cpu)
1233{
1234 int sibling;
1235 struct cpuinfo_x86 *c = &cpu_data(cpu);
1236
1237 for_each_cpu(sibling, cpu_core_mask(cpu)) {
1238 cpumask_clear_cpu(cpu, cpu_core_mask(sibling));
1239 /*/
1240 * last thread sibling in this cpu core going down
1241 */
1242 if (cpumask_weight(cpu_sibling_mask(cpu)) == 1)
1243 cpu_data(sibling).booted_cores--;
1244 }
1245
1246 for_each_cpu(sibling, cpu_sibling_mask(cpu))
1247 cpumask_clear_cpu(cpu, cpu_sibling_mask(sibling));
1248 cpumask_clear(cpu_sibling_mask(cpu));
1249 cpumask_clear(cpu_core_mask(cpu));
1250 c->phys_proc_id = 0;
1251 c->cpu_core_id = 0;
1252 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1253}
1254
1255static void __ref remove_cpu_from_maps(int cpu)
1256{
1257 set_cpu_online(cpu, false);
1258 cpumask_clear_cpu(cpu, cpu_callout_mask);
1259 cpumask_clear_cpu(cpu, cpu_callin_mask);
1260 /* was set by cpu_init() */
1261 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1262 numa_remove_cpu(cpu);
1263}
1264
1265void cpu_disable_common(void)
1266{
1267 int cpu = smp_processor_id();
1268
1269 remove_siblinginfo(cpu);
1270
1271 /* It's now safe to remove this processor from the online map */
1272 lock_vector_lock();
1273 remove_cpu_from_maps(cpu);
1274 unlock_vector_lock();
1275 fixup_irqs();
1276}
1277
1278int native_cpu_disable(void)
1279{
1280 int cpu = smp_processor_id();
1281
1282 /*
1283 * Perhaps use cpufreq to drop frequency, but that could go
1284 * into generic code.
1285 *
1286 * We won't take down the boot processor on i386 due to some
1287 * interrupts only being able to be serviced by the BSP.
1288 * Especially so if we're not using an IOAPIC -zwane
1289 */
1290 if (cpu == 0)
1291 return -EBUSY;
1292
1293 clear_local_APIC();
1294
1295 cpu_disable_common();
1296 return 0;
1297}
1298
1299void native_cpu_die(unsigned int cpu)
1300{
1301 /* We don't do anything here: idle task is faking death itself. */
1302 unsigned int i;
1303
1304 for (i = 0; i < 10; i++) {
1305 /* They ack this in play_dead by setting CPU_DEAD */
1306 if (per_cpu(cpu_state, cpu) == CPU_DEAD) {
1307 if (system_state == SYSTEM_RUNNING)
1308 pr_info("CPU %u is now offline\n", cpu);
1309
1310 if (1 == num_online_cpus())
1311 alternatives_smp_switch(0);
1312 return;
1313 }
1314 msleep(100);
1315 }
1316 pr_err("CPU %u didn't die...\n", cpu);
1317}
1318
1319void play_dead_common(void)
1320{
1321 idle_task_exit();
1322 reset_lazy_tlbstate();
1323 amd_e400_remove_cpu(raw_smp_processor_id());
1324
1325 mb();
1326 /* Ack it */
1327 __this_cpu_write(cpu_state, CPU_DEAD);
1328
1329 /*
1330 * With physical CPU hotplug, we should halt the cpu
1331 */
1332 local_irq_disable();
1333}
1334
1335/*
1336 * We need to flush the caches before going to sleep, lest we have
1337 * dirty data in our caches when we come back up.
1338 */
1339static inline void mwait_play_dead(void)
1340{
1341 unsigned int eax, ebx, ecx, edx;
1342 unsigned int highest_cstate = 0;
1343 unsigned int highest_subcstate = 0;
1344 int i;
1345 void *mwait_ptr;
1346 struct cpuinfo_x86 *c = __this_cpu_ptr(&cpu_info);
1347
1348 if (!(this_cpu_has(X86_FEATURE_MWAIT) && mwait_usable(c)))
1349 return;
1350 if (!this_cpu_has(X86_FEATURE_CLFLSH))
1351 return;
1352 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1353 return;
1354
1355 eax = CPUID_MWAIT_LEAF;
1356 ecx = 0;
1357 native_cpuid(&eax, &ebx, &ecx, &edx);
1358
1359 /*
1360 * eax will be 0 if EDX enumeration is not valid.
1361 * Initialized below to cstate, sub_cstate value when EDX is valid.
1362 */
1363 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1364 eax = 0;
1365 } else {
1366 edx >>= MWAIT_SUBSTATE_SIZE;
1367 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1368 if (edx & MWAIT_SUBSTATE_MASK) {
1369 highest_cstate = i;
1370 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1371 }
1372 }
1373 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1374 (highest_subcstate - 1);
1375 }
1376
1377 /*
1378 * This should be a memory location in a cache line which is
1379 * unlikely to be touched by other processors. The actual
1380 * content is immaterial as it is not actually modified in any way.
1381 */
1382 mwait_ptr = ¤t_thread_info()->flags;
1383
1384 wbinvd();
1385
1386 while (1) {
1387 /*
1388 * The CLFLUSH is a workaround for erratum AAI65 for
1389 * the Xeon 7400 series. It's not clear it is actually
1390 * needed, but it should be harmless in either case.
1391 * The WBINVD is insufficient due to the spurious-wakeup
1392 * case where we return around the loop.
1393 */
1394 clflush(mwait_ptr);
1395 __monitor(mwait_ptr, 0, 0);
1396 mb();
1397 __mwait(eax, 0);
1398 }
1399}
1400
1401static inline void hlt_play_dead(void)
1402{
1403 if (__this_cpu_read(cpu_info.x86) >= 4)
1404 wbinvd();
1405
1406 while (1) {
1407 native_halt();
1408 }
1409}
1410
1411void native_play_dead(void)
1412{
1413 play_dead_common();
1414 tboot_shutdown(TB_SHUTDOWN_WFS);
1415
1416 mwait_play_dead(); /* Only returns on failure */
1417 hlt_play_dead();
1418}
1419
1420#else /* ... !CONFIG_HOTPLUG_CPU */
1421int native_cpu_disable(void)
1422{
1423 return -ENOSYS;
1424}
1425
1426void native_cpu_die(unsigned int cpu)
1427{
1428 /* We said "no" in __cpu_disable */
1429 BUG();
1430}
1431
1432void native_play_dead(void)
1433{
1434 BUG();
1435}
1436
1437#endif