1 /*
2  *	x86 SMP booting functions
3  *
4  *	(c) 1995 Alan Cox, Building #3 <alan@redhat.com>
5  *	(c) 1998, 1999, 2000 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 BogoMIP 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  */
35 
36 #include <linux/config.h>
37 #include <linux/init.h>
38 
39 #include <linux/mm.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/smp_lock.h>
42 #include <linux/irq.h>
43 #include <linux/bootmem.h>
44 
45 #include <linux/delay.h>
46 #include <linux/mc146818rtc.h>
47 #include <asm/mtrr.h>
48 #include <asm/pgalloc.h>
49 #include <asm/desc.h>
50 #include <asm/kdebug.h>
51 #include <asm/timex.h>
52 #include <asm/proto.h>
53 #include <asm/acpi.h>
54 
55 /* Setup configured maximum number of CPUs to activate */
56 unsigned int max_cpus = NR_CPUS;
57 
58 static int cpu_mask = -1;
59 
60 /* Total count of live CPUs */
61 int smp_num_cpus = 1;
62 
63 /* Number of siblings per CPU package */
64 int smp_num_siblings = 1;
65 int __initdata phys_proc_id[NR_CPUS]; /* Package ID of each logical CPU */
66 int cpu_sibling_map[NR_CPUS] __cacheline_aligned;
67 
68 /* Bitmask of currently online CPUs */
69 unsigned long cpu_online_map;
70 
71 /* which CPU (physical APIC ID) maps to which logical CPU number */
72 volatile int x86_apicid_to_cpu[NR_CPUS];
73 /* which logical CPU number maps to which CPU (physical APIC ID) */
74 volatile int x86_cpu_to_apicid[NR_CPUS];
75 
76 static volatile unsigned long cpu_callin_map;
77 static volatile unsigned long cpu_callout_map;
78 
79 /* Per CPU bogomips and other parameters */
80 struct cpuinfo_x86 cpu_data[NR_CPUS] __cacheline_aligned;
81 
82 /* Set when the idlers are all forked */
83 int smp_threads_ready;
84 
85 extern void time_init_smp(void);
86 
87 /*
88  * Setup routine for controlling SMP activation
89  *
90  * Command-line option of "nosmp" or "maxcpus=0" will disable SMP
91  * activation entirely (the MPS table probe still happens, though).
92  */
93 
nosmp(char * str)94 static int __init nosmp(char *str)
95 {
96 	max_cpus = 0;
97 	return 1;
98 }
99 
100 __setup("nosmp", nosmp);
101 
cpumask(char * str)102 static int __init cpumask(char *str)
103 {
104 	get_option(&str, &cpu_mask);
105 	return 1;
106 }
107 
108 __setup("cpumask=", cpumask);
109 
110 /*
111  * Trampoline 80x86 program as an array.
112  */
113 
114 extern unsigned char trampoline_data [];
115 extern unsigned char trampoline_end  [];
116 static unsigned char *trampoline_base;
117 
118 /*
119  * Currently trivial. Write the real->protected mode
120  * bootstrap into the page concerned. The caller
121  * has made sure it's suitably aligned.
122  */
123 
setup_trampoline(void)124 static unsigned long __init setup_trampoline(void)
125 {
126 	extern volatile __u32 tramp_gdt_ptr;
127 	tramp_gdt_ptr = __pa_symbol(&gdt_table);
128 	memcpy(trampoline_base, trampoline_data, trampoline_end - trampoline_data);
129 	return virt_to_phys(trampoline_base);
130 }
131 
132 /*
133  * We are called very early to get the low memory for the
134  * SMP bootup trampoline page.
135  */
smp_alloc_memory(void)136 void __init smp_alloc_memory(void)
137 {
138 	trampoline_base = __va(0x6000); /* reserved in setup.c */
139 }
140 
141 /*
142  * The bootstrap kernel entry code has set these up. Save them for
143  * a given CPU
144  */
145 
smp_store_cpu_info(int id)146 void __init smp_store_cpu_info(int id)
147 {
148 	struct cpuinfo_x86 *c = cpu_data + id;
149 
150 	*c = boot_cpu_data;
151 	identify_cpu(c);
152 }
153 
154 /*
155  * Architecture specific routine called by the kernel just before init is
156  * fired off. This allows the BP to have everything in order [we hope].
157  * At the end of this all the APs will hit the system scheduling and off
158  * we go. Each AP will load the system gdt's and jump through the kernel
159  * init into idle(). At this point the scheduler will one day take over
160  * and give them jobs to do. smp_callin is a standard routine
161  * we use to track CPUs as they power up.
162  */
163 
164 static atomic_t smp_commenced = ATOMIC_INIT(0);
165 
smp_commence(void)166 void __init smp_commence(void)
167 {
168 	/*
169 	 * Lets the callins below out of their loop.
170 	 */
171 	Dprintk("Setting commenced=1, go go go\n");
172 
173 	wmb();
174 	atomic_set(&smp_commenced,1);
175 }
176 
177 /*
178  * TSC synchronization.
179  *
180  * We first check wether all CPUs have their TSC's synchronized,
181  * then we print a warning if not, and always resync.
182  */
183 
184 static atomic_t tsc_start_flag = ATOMIC_INIT(0);
185 static atomic_t tsc_count_start = ATOMIC_INIT(0);
186 static atomic_t tsc_count_stop = ATOMIC_INIT(0);
187 static unsigned long long tsc_values[NR_CPUS];
188 
189 #define NR_LOOPS 5
190 
div64(unsigned long long a,unsigned long b)191 static inline unsigned long long div64 (unsigned long long a, unsigned long b)
192 {
193 	return a/b;
194 }
195 
synchronize_tsc_bp(void)196 static void __init synchronize_tsc_bp (void)
197 {
198 	int i;
199 	unsigned long long t0;
200 	unsigned long long sum, avg;
201 	long long delta;
202 	unsigned long one_usec;
203 	int buggy = 0;
204 
205 	printk("checking TSC synchronization across CPUs: ");
206 
207 	one_usec = cpu_khz / 1000;
208 
209 	atomic_set(&tsc_start_flag, 1);
210 	wmb();
211 
212 	/*
213 	 * We loop a few times to get a primed instruction cache,
214 	 * then the last pass is more or less synchronized and
215 	 * the BP and APs set their cycle counters to zero all at
216 	 * once. This reduces the chance of having random offsets
217 	 * between the processors, and guarantees that the maximum
218 	 * delay between the cycle counters is never bigger than
219 	 * the latency of information-passing (cachelines) between
220 	 * two CPUs.
221 	 */
222 	for (i = 0; i < NR_LOOPS; i++) {
223 		/*
224 		 * all APs synchronize but they loop on '== num_cpus'
225 		 */
226 		while (atomic_read(&tsc_count_start) != smp_num_cpus-1) mb();
227 		atomic_set(&tsc_count_stop, 0);
228 		wmb();
229 		/*
230 		 * this lets the APs save their current TSC:
231 		 */
232 		atomic_inc(&tsc_count_start);
233 
234 		sync_core();
235 		rdtscll(tsc_values[smp_processor_id()]);
236 
237 		/*
238 		 * We clear the TSC in the last loop:
239 		 */
240 
241 		if (i == NR_LOOPS-1) {
242 			write_tsc(0, 0);
243 		}
244 
245 		/*
246 		 * Wait for all APs to leave the synchronization point:
247 		 */
248 		while (atomic_read(&tsc_count_stop) != smp_num_cpus-1) mb();
249 		atomic_set(&tsc_count_start, 0);
250 		wmb();
251 		atomic_inc(&tsc_count_stop);
252 	}
253 
254 	sum = 0;
255 	for (i = 0; i < smp_num_cpus; i++) {
256 		t0 = tsc_values[i];
257 		sum += t0;
258 	}
259 	avg = div64(sum, smp_num_cpus);
260 
261 	sum = 0;
262 	for (i = 0; i < smp_num_cpus; i++) {
263 		delta = tsc_values[i] - avg;
264 		if (delta < 0)
265 			delta = -delta;
266 		/*
267 		 * We report bigger than 2 microseconds clock differences.
268 		 */
269 		if (delta > 2*one_usec) {
270 			long realdelta;
271 			if (!buggy) {
272 				buggy = 1;
273 				printk("\n");
274 			}
275 			realdelta = div64(delta, one_usec);
276 			if (tsc_values[i] < avg)
277 				realdelta = -realdelta;
278 
279 			printk("BIOS BUG: CPU#%d improperly initialized, has %ld usecs TSC skew! FIXED.\n",
280 				i, realdelta);
281 		}
282 
283 		sum += delta;
284 	}
285 	if (!buggy)
286 		printk("passed.\n");
287 }
288 
synchronize_tsc_ap(void)289 static void __init synchronize_tsc_ap (void)
290 {
291 	int i;
292 
293 	/*
294 	 * smp_num_cpus is not necessarily known at the time
295 	 * this gets called, so we first wait for the BP to
296 	 * finish SMP initialization:
297 	 */
298 	while (!atomic_read(&tsc_start_flag)) mb();
299 
300 	for (i = 0; i < NR_LOOPS; i++) {
301 		atomic_inc(&tsc_count_start);
302 		while (atomic_read(&tsc_count_start) != smp_num_cpus) mb();
303 
304 		sync_core();
305 		rdtscll(tsc_values[smp_processor_id()]);
306 		if (i == NR_LOOPS-1)
307 			write_tsc(0, 0);
308 
309 		atomic_inc(&tsc_count_stop);
310 		while (atomic_read(&tsc_count_stop) != smp_num_cpus) mb();
311 	}
312 }
313 #undef NR_LOOPS
314 
315 extern void calibrate_delay(void);
316 
317 static atomic_t init_deasserted;
318 
smp_callin(void)319 void __init smp_callin(void)
320 {
321 	int cpuid, phys_id;
322 	unsigned long timeout;
323 
324 	/*
325 	 * If waken up by an INIT in an 82489DX configuration
326 	 * we may get here before an INIT-deassert IPI reaches
327 	 * our local APIC.  We have to wait for the IPI or we'll
328 	 * lock up on an APIC access.
329 	 */
330 	while (!atomic_read(&init_deasserted));
331 
332 	/*
333 	 * (This works even if the APIC is not enabled.)
334 	 */
335 	phys_id = GET_APIC_ID(apic_read(APIC_ID));
336 	cpuid = current->processor;
337 	if (test_and_set_bit(cpuid, &cpu_online_map)) {
338 		printk("huh, phys CPU#%d, CPU#%d already present??\n",
339 					phys_id, cpuid);
340 		BUG();
341 	}
342 	Dprintk("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid, phys_id);
343 
344 	/*
345 	 * STARTUP IPIs are fragile beasts as they might sometimes
346 	 * trigger some glue motherboard logic. Complete APIC bus
347 	 * silence for 1 second, this overestimates the time the
348 	 * boot CPU is spending to send the up to 2 STARTUP IPIs
349 	 * by a factor of two. This should be enough.
350 	 */
351 
352 	/*
353 	 * Waiting 2s total for startup (udelay is not yet working)
354 	 */
355 	timeout = jiffies + 2*HZ;
356 	while (time_before(jiffies, timeout)) {
357 		/*
358 		 * Has the boot CPU finished it's STARTUP sequence?
359 		 */
360 		if (test_bit(cpuid, &cpu_callout_map))
361 			break;
362 		rep_nop();
363 	}
364 
365 	if (!time_before(jiffies, timeout)) {
366 		printk("BUG: CPU%d started up but did not get a callout!\n",
367 			cpuid);
368 		BUG();
369 	}
370 
371 	/*
372 	 * the boot CPU has finished the init stage and is spinning
373 	 * on callin_map until we finish. We are free to set up this
374 	 * CPU, first the APIC. (this is probably redundant on most
375 	 * boards)
376 	 */
377 
378 	Dprintk("CALLIN, before setup_local_APIC().\n");
379 	setup_local_APIC();
380 
381 	if (nmi_watchdog == NMI_IO_APIC) {
382 		disable_8259A_irq(0);
383 		enable_NMI_through_LVT0(NULL);
384 		enable_8259A_irq(0);
385 	}
386 
387 	sti();
388 
389 #ifdef CONFIG_MTRR
390 	/*
391 	 * Must be done before calibration delay is computed
392 	 */
393 	mtrr_init_secondary_cpu ();
394 #endif
395 	/*
396 	 * Get our bogomips.
397 	 */
398 	calibrate_delay();
399 	Dprintk("Stack at about %p\n",&cpuid);
400 
401 	/*
402 	 * Save our processor parameters
403 	 */
404  	smp_store_cpu_info(cpuid);
405 
406 	/*
407 	 * Allow the master to continue.
408 	 */
409 	set_bit(cpuid, &cpu_callin_map);
410 
411 	/*
412 	 *      Synchronize the TSC with the BP
413 	 */
414 	if (cpu_has_tsc)
415 		synchronize_tsc_ap();
416 }
417 
418 int cpucount;
419 
420 /*
421  * Activate a secondary processor.
422  */
start_secondary(void * unused)423 int __init start_secondary(void *unused)
424 {
425 	/*
426 	 * Dont put anything before smp_callin(), SMP
427 	 * booting is too fragile that we want to limit the
428 	 * things done here to the most necessary things.
429 	 */
430 	cpu_init();
431 	smp_callin();
432 	while (!atomic_read(&smp_commenced))
433 		rep_nop();
434 	/*
435 	 * low-memory mappings have been cleared, flush them from
436 	 * the local TLBs too.
437 	 */
438 	local_flush_tlb();
439 
440 	cpu_idle();
441 	return 0;
442 }
443 
444 /*
445  * Everything has been set up for the secondary
446  * CPUs - they just need to reload everything
447  * from the task structure
448  * This function must not return.
449  */
initialize_secondary(void)450 void __init initialize_secondary(void)
451 {
452 	struct task_struct *me = stack_current();
453 
454 	/*
455 	 * We don't actually need to load the full TSS,
456 	 * basically just the stack pointer and the eip.
457 	 */
458 
459 	asm volatile(
460 		"movq %0,%%rsp\n\t"
461 		"jmp *%1"
462 		:
463 		:"r" (me->thread.rsp),"r" (me->thread.rip));
464 }
465 
466 extern volatile void *init_rsp;
467 extern void (*initial_code)(void);
468 
fork_by_hand(void)469 static int __init fork_by_hand(void)
470 {
471 	struct pt_regs regs;
472 	/*
473 	 * don't care about the eip and regs settings since
474 	 * we'll never reschedule the forked task.
475 	 */
476 	return do_fork(CLONE_VM|CLONE_PID, 0, &regs, 0);
477 }
478 
479 #if APIC_DEBUG
inquire_remote_apic(int apicid)480 static inline void inquire_remote_apic(int apicid)
481 {
482 	int i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
483 	char *names[] = { "ID", "VERSION", "SPIV" };
484 	int timeout, status;
485 
486 	printk("Inquiring remote APIC #%d...\n", apicid);
487 
488 	for (i = 0; i < sizeof(regs) / sizeof(*regs); i++) {
489 		printk("... APIC #%d %s: ", apicid, names[i]);
490 
491 		/*
492 		 * Wait for idle.
493 		 */
494 		apic_wait_icr_idle();
495 
496 		apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(apicid));
497 		apic_write_around(APIC_ICR, APIC_DM_REMRD | regs[i]);
498 
499 		timeout = 0;
500 		do {
501 			udelay(100);
502 			status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
503 		} while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
504 
505 		switch (status) {
506 		case APIC_ICR_RR_VALID:
507 			status = apic_read(APIC_RRR);
508 			printk("%08x\n", status);
509 			break;
510 		default:
511 			printk("failed\n");
512 		}
513 	}
514 }
515 #endif
516 
do_boot_cpu(int apicid)517 static int __init do_boot_cpu (int apicid)
518 {
519 	struct task_struct *idle;
520 	unsigned long send_status, accept_status, boot_status, maxlvt;
521 	int timeout, num_starts, j, cpu;
522 	unsigned long start_eip;
523 
524 	cpu = ++cpucount;
525 
526 	/*
527 	 * We can't use kernel_thread since we must avoid to
528 	 * reschedule the child.
529 	 */
530 	if (fork_by_hand() < 0)
531 		panic("failed fork for CPU %d", cpu);
532 
533 	/*
534 	 * We remove it from the pidhash and the runqueue
535 	 * once we got the process:
536 	 */
537 	idle = init_task.prev_task;
538 	if (!idle)
539 		panic("No idle process for CPU %d", cpu);
540 
541 	idle->processor = cpu;
542 	x86_cpu_to_apicid[cpu] = apicid;
543 	x86_apicid_to_cpu[apicid] = cpu;
544 	idle->cpus_runnable = 1<<cpu;
545 	idle->cpus_allowed = 1<<cpu;
546 	idle->thread.rip = (unsigned long)start_secondary;
547 	idle->thread.rsp = (unsigned long)idle + THREAD_SIZE - 8;
548 
549 	del_from_runqueue(idle);
550 	unhash_process(idle);
551 	cpu_pda[cpu].pcurrent = init_tasks[cpu] = idle;
552 
553 	/* start_eip had better be page-aligned! */
554 	start_eip = setup_trampoline();
555 
556 	/* So we see what's up   */
557 	printk("Booting processor %d/%d rip %lx page %p\n", cpu, apicid, start_eip, idle);
558 	init_rsp = (void *) (THREAD_SIZE + (char *)idle - 16);
559 	initial_code = initialize_secondary;
560 
561 	/*
562 	 * This grunge runs the startup process for
563 	 * the targeted processor.
564 	 */
565 
566 	atomic_set(&init_deasserted, 0);
567 
568 	Dprintk("Setting warm reset code and vector.\n");
569 
570 	CMOS_WRITE(0xa, 0xf);
571 	local_flush_tlb();
572 	Dprintk("1.\n");
573 	*((volatile unsigned short *) phys_to_virt(0x469)) = start_eip >> 4;
574 	Dprintk("2.\n");
575 	*((volatile unsigned short *) phys_to_virt(0x467)) = start_eip & 0xf;
576 	Dprintk("3.\n");
577 
578 	/*
579 	 * Be paranoid about clearing APIC errors.
580 	 */
581 	if (APIC_INTEGRATED(apic_version[apicid])) {
582 		apic_read_around(APIC_SPIV);
583 		apic_write(APIC_ESR, 0);
584 		apic_read(APIC_ESR);
585 	}
586 
587 	/*
588 	 * Status is now clean
589 	 */
590 	send_status = 0;
591 	accept_status = 0;
592 	boot_status = 0;
593 
594 	/*
595 	 * Starting actual IPI sequence...
596 	 */
597 
598 	Dprintk("Asserting INIT.\n");
599 
600 	/*
601 	 * Turn INIT on target chip
602 	 */
603 	apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(apicid));
604 
605 	/*
606 	 * Send IPI
607 	 */
608 	apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_INT_ASSERT
609 				| APIC_DM_INIT);
610 
611 	Dprintk("Waiting for send to finish...\n");
612 	timeout = 0;
613 	do {
614 		Dprintk("+");
615 		udelay(100);
616 		send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
617 	} while (send_status && (timeout++ < 1000));
618 
619 	mdelay(10);
620 
621 	Dprintk("Deasserting INIT.\n");
622 
623 	/* Target chip */
624 	apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(apicid));
625 
626 	/* Send IPI */
627 	apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_DM_INIT);
628 
629 	Dprintk("Waiting for send to finish...\n");
630 	timeout = 0;
631 	do {
632 		Dprintk("+");
633 		udelay(100);
634 		send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
635 	} while (send_status && (timeout++ < 1000));
636 
637 	atomic_set(&init_deasserted, 1);
638 
639 	/*
640 	 * Should we send STARTUP IPIs ?
641 	 *
642 	 * Determine this based on the APIC version.
643 	 * If we don't have an integrated APIC, don't
644 	 * send the STARTUP IPIs.
645 	 */
646 	if (APIC_INTEGRATED(apic_version[apicid]))
647 		num_starts = 2;
648 	else
649 		num_starts = 0;
650 
651 	/*
652 	 * Run STARTUP IPI loop.
653 	 */
654 	Dprintk("#startup loops: %d.\n", num_starts);
655 
656 	maxlvt = get_maxlvt();
657 
658 	for (j = 1; j <= num_starts; j++) {
659 		Dprintk("Sending STARTUP #%d.\n",j);
660 		apic_read_around(APIC_SPIV);
661 		apic_write(APIC_ESR, 0);
662 		apic_read(APIC_ESR);
663 		Dprintk("After apic_write.\n");
664 
665 		/*
666 		 * STARTUP IPI
667 		 */
668 
669 		/* Target chip */
670 		Dprintk("target apic %x\n", SET_APIC_DEST_FIELD(apicid));
671 		apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(apicid));
672 
673 		Dprintk("after target chip\n");
674 
675 		/* Boot on the stack */
676 		/* Kick the second */
677 		apic_write_around(APIC_ICR, APIC_DM_STARTUP
678 					| (start_eip >> 12));
679 
680 		Dprintk("after eip write\n");
681 
682 		/*
683 		 * Give the other CPU some time to accept the IPI.
684 		 */
685 		udelay(300);
686 
687 		Dprintk("Startup point 1.\n");
688 
689 		Dprintk("Waiting for send to finish...\n");
690 		timeout = 0;
691 		do {
692 			Dprintk("+");
693 			udelay(100);
694 			send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
695 		} while (send_status && (timeout++ < 1000));
696 
697 		/*
698 		 * Give the other CPU some time to accept the IPI.
699 		 */
700 		udelay(200);
701 		/*
702 		 * Due to the Pentium erratum 3AP.
703 		 */
704 		if (maxlvt > 3) {
705 			apic_read_around(APIC_SPIV);
706 			apic_write(APIC_ESR, 0);
707 		}
708 		accept_status = (apic_read(APIC_ESR) & 0xEF);
709 		if (send_status || accept_status)
710 			break;
711 	}
712 	Dprintk("After Startup.\n");
713 
714 	if (send_status)
715 		printk("APIC never delivered???\n");
716 	if (accept_status)
717 		printk("APIC delivery error (%lx).\n", accept_status);
718 
719 	if (!send_status && !accept_status) {
720 		/*
721 		 * allow APs to start initializing.
722 		 */
723 		Dprintk("Before Callout %d.\n", cpu);
724 		set_bit(cpu, &cpu_callout_map);
725 		Dprintk("After Callout %d.\n", cpu);
726 
727 		/*
728 		 * Wait 5s total for a response
729 		 */
730 		for (timeout = 0; timeout < 50000; timeout++) {
731 			if (test_bit(cpu, &cpu_callin_map))
732 				break;	/* It has booted */
733 			udelay(100);
734 		}
735 
736 		if (test_bit(cpu, &cpu_callin_map)) {
737 			/* number CPUs logically, starting from 1 (BSP is 0) */
738 			Dprintk("OK.\n");
739 			printk("CPU%d: ", cpu);
740 			print_cpu_info(&cpu_data[cpu]);
741 			Dprintk("CPU has booted.\n");
742 		} else {
743 			boot_status = 1;
744 			if (*((volatile unsigned char *)phys_to_virt(8192))
745 					== 0xA5)
746 				/* trampoline started but...? */
747 				printk("Stuck ??\n");
748 			else
749 				/* trampoline code not run */
750 				printk("Not responding.\n");
751 #if APIC_DEBUG
752 			inquire_remote_apic(apicid);
753 #endif
754 		}
755 	}
756 	if (send_status || accept_status || boot_status) {
757 		x86_cpu_to_apicid[cpu] = -1;
758 		x86_apicid_to_cpu[apicid] = -1;
759 		cpucount--;
760 	}
761 
762 	/* mark "stuck" area as not stuck */
763 	*((volatile unsigned int *)phys_to_virt(8192)) = 0;
764 
765 	return cpu;
766 }
767 
768 cycles_t cacheflush_time;
769 
smp_tune_scheduling(void)770 static __init void smp_tune_scheduling (void)
771 {
772 	unsigned long cachesize;       /* kB   */
773 	unsigned long bandwidth = 2000; /* MB/s */
774 	/*
775 	 * Rough estimation for SMP scheduling, this is the number of
776 	 * cycles it takes for a fully memory-limited process to flush
777 	 * the SMP-local cache.
778 	 *
779 	 * (For a P5 this pretty much means we will choose another idle
780 	 *  CPU almost always at wakeup time (this is due to the small
781 	 *  L1 cache), on PIIs it's around 50-100 usecs, depending on
782 	 *  the cache size)
783 	 */
784 
785 	if (!cpu_khz) {
786 		/*
787 		 * this basically disables processor-affinity
788 		 * scheduling on SMP without a TSC.
789 		 */
790 		cacheflush_time = 0;
791 		return;
792 	} else {
793 		cachesize = boot_cpu_data.x86_cache_size;
794 		if (cachesize == -1) {
795 			cachesize = 16; /* Pentiums, 2x8kB cache */
796 			bandwidth = 100;
797 		}
798 
799 		cacheflush_time = (cpu_khz>>10) * (cachesize<<10) / bandwidth;
800 	}
801 
802 	cacheflush_time *= 10;  /* Add an NUMA factor */
803 
804 	printk("per-CPU timeslice cutoff: %ld.%02ld usecs.\n",
805 		(long)cacheflush_time/(cpu_khz/1000),
806 		((long)cacheflush_time*100/(cpu_khz/1000)) % 100);
807 }
808 
809 /*
810  * Cycle through the processors sending APIC IPIs to boot each.
811  */
812 
813 extern int prof_multiplier[NR_CPUS];
814 extern int prof_old_multiplier[NR_CPUS];
815 extern int prof_counter[NR_CPUS];
816 
smp_boot_cpus(void)817 void __init smp_boot_cpus(void)
818 {
819 	int apicid, cpu, maxcpu;
820 
821 #ifdef CONFIG_MTRR
822 	/*  Must be done before other processors booted  */
823 	mtrr_init_boot_cpu ();
824 #endif
825 	/*
826 	 * Initialize the logical to physical CPU number mapping
827 	 * and the per-CPU profiling counter/multiplier
828 	 */
829 
830 	for (apicid = 0; apicid < NR_CPUS; apicid++) {
831 		x86_apicid_to_cpu[apicid] = -1;
832 		prof_counter[apicid] = 1;
833 		prof_old_multiplier[apicid] = 1;
834 		prof_multiplier[apicid] = 1;
835 	}
836 
837 	/*
838 	 * Setup boot CPU information
839 	 */
840 	smp_store_cpu_info(0); /* Final full version of the data */
841 	printk("CPU%d: ", 0);
842 	print_cpu_info(&cpu_data[0]);
843 
844 	/*
845 	 * We have the boot CPU online for sure.
846 	 */
847 	set_bit(0, &cpu_online_map);
848 	x86_apicid_to_cpu[boot_cpu_id] = 0;
849 	x86_cpu_to_apicid[0] = boot_cpu_id;
850 	global_irq_holder = 0;
851 	current->processor = 0;
852 	init_idle();
853 	smp_tune_scheduling();
854 
855 	/*
856 	 * If we couldnt find an SMP configuration at boot time,
857 	 * get out of here now!
858 	 */
859 	if (!smp_found_config && !acpi_lapic) {
860 		printk(KERN_NOTICE "SMP motherboard not detected.\n");
861 		io_apic_irqs = 0;
862 		cpu_online_map = phys_cpu_present_map = 1;
863 		smp_num_cpus = 1;
864 		if (APIC_init_uniprocessor())
865 			printk(KERN_NOTICE "Local APIC not detected."
866 					   " Using dummy APIC emulation.\n");
867 		goto smp_done;
868 	}
869 
870 	/*
871 	 * Should not be necessary because the MP table should list the boot
872 	 * CPU too, but we do it for the sake of robustness anyway.
873 	 */
874 	if (!test_bit(boot_cpu_id, &phys_cpu_present_map)) {
875 		printk("weird, boot CPU (#%d) not listed by the BIOS.\n",
876 								 boot_cpu_id);
877 		phys_cpu_present_map |= (1 << hard_smp_processor_id());
878 	}
879 
880 	/*
881 	 * If we couldn't find a local APIC, then get out of here now!
882 	 */
883 	if (APIC_INTEGRATED(apic_version[boot_cpu_id]) &&
884 	    !test_bit(X86_FEATURE_APIC, boot_cpu_data.x86_capability)) {
885 		printk(KERN_ERR "BIOS bug, local APIC #%d not detected!...\n",
886 			boot_cpu_id);
887 		printk(KERN_ERR "... forcing use of dummy APIC emulation. (tell your hw vendor)\n");
888 		io_apic_irqs = 0;
889 		cpu_online_map = phys_cpu_present_map = 1;
890 		smp_num_cpus = 1;
891 		apic_disabled = 1;
892 		goto smp_done;
893 	}
894 
895 	verify_local_APIC();
896 
897 	/*
898 	 * If SMP should be disabled, then really disable it!
899 	 */
900 	if (!max_cpus) {
901 		smp_found_config = 0;
902 		printk(KERN_INFO "SMP mode deactivated, forcing use of dummy APIC emulation.\n");
903 		cpu_online_map = phys_cpu_present_map = 1;
904 		smp_num_cpus = 1;
905 		goto smp_done;
906 	}
907 
908 	connect_bsp_APIC();
909 	setup_local_APIC();
910 
911 	if (GET_APIC_ID(apic_read(APIC_ID)) != boot_cpu_id)
912 		BUG();
913 
914 	/*
915 	 * Now scan the CPU present map and fire up the other CPUs.
916 	 */
917 	Dprintk("CPU present map: %lx\n", phys_cpu_present_map);
918 
919 	maxcpu = 0;
920 	for (apicid = 0; apicid < NR_CPUS; apicid++) {
921 		/*
922 		 * Don't even attempt to start the boot CPU!
923 		 */
924 		if (apicid == boot_cpu_id)
925 			continue;
926 
927 		if (!(phys_cpu_present_map & (1 << apicid)))
928 			continue;
929 		if (((1<<apicid) & cpu_mask) == 0)
930 			continue;
931 
932 		cpu = do_boot_cpu(apicid);
933 
934 		/*
935 		 * Make sure we unmap all failed CPUs
936 		 */
937 		if ((x86_apicid_to_cpu[apicid] == -1) &&
938 				(phys_cpu_present_map & (1 << apicid)))
939 			printk("phys CPU #%d not responding - cannot use it.\n",apicid);
940 		else if (cpu > maxcpu)
941 			maxcpu = cpu;
942 	}
943 
944 	/*
945 	 * Cleanup possible dangling ends...
946 	 */
947 	{
948 		/*
949 		 * Install writable page 0 entry to set BIOS data area.
950 		 */
951 		local_flush_tlb();
952 
953 		/*
954 		 * Paranoid:  Set warm reset code and vector here back
955 		 * to default values.
956 		 */
957 		CMOS_WRITE(0, 0xf);
958 
959 		*((volatile int *) phys_to_virt(0x467)) = 0;
960 	}
961 
962 	/*
963 	 * Allow the user to impress friends.
964 	 */
965 
966 	Dprintk("Before bogomips.\n");
967 	{
968 		unsigned long bogosum = 0;
969 		for (cpu = 0; cpu < NR_CPUS; cpu++)
970 			if (cpu_online_map & (1<<cpu))
971 				bogosum += cpu_data[cpu].loops_per_jiffy;
972 		printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
973 			cpucount+1,
974 			bogosum/(500000/HZ),
975 			(bogosum/(5000/HZ))%100);
976 		Dprintk("Before bogocount - setting activated=1.\n");
977 	}
978 	smp_num_cpus = maxcpu + 1;
979 
980 	Dprintk("Boot done.\n");
981 
982 	/*
983 	 * If Hyper-Threading is avaialble, construct cpu_sibling_map[], so
984 	 * that we can tell the sibling CPU efficiently.
985 	 */
986 	if (test_bit(X86_FEATURE_HT, boot_cpu_data.x86_capability)
987 	    && smp_num_siblings > 1) {
988 		for (cpu = 0; cpu < NR_CPUS; cpu++)
989 			cpu_sibling_map[cpu] = NO_PROC_ID;
990 
991 		for (cpu = 0; cpu < smp_num_cpus; cpu++) {
992 			int 	i;
993 
994 			for (i = 0; i < smp_num_cpus; i++) {
995 				if (i == cpu)
996 					continue;
997 				if (phys_proc_id[cpu] == phys_proc_id[i]) {
998 					cpu_sibling_map[cpu] = i;
999 					printk("cpu_sibling_map[%d] = %d\n", cpu, cpu_sibling_map[cpu]);
1000 					break;
1001 				}
1002 			}
1003 			if (cpu_sibling_map[cpu] == NO_PROC_ID) {
1004 				smp_num_siblings = 1;
1005 				printk(KERN_WARNING "WARNING: No sibling found for CPU %d.\n", cpu);
1006 			}
1007 		}
1008 	}
1009 
1010 	/*
1011 	 * Here we can be sure that there is an IO-APIC in the system. Let's
1012 	 * go and set it up:
1013 	 */
1014 	if (!skip_ioapic_setup && nr_ioapics)
1015 		setup_IO_APIC();
1016 	else
1017 		nr_ioapics = 0;
1018 
1019 	/*
1020 	 * Set up all local APIC timers in the system:
1021 	 */
1022 	setup_APIC_clocks();
1023 
1024 	/*
1025 	 * Synchronize the TSC with the AP
1026 	 */
1027 	if (cpu_has_tsc && cpucount)
1028 		synchronize_tsc_bp();
1029 
1030 	if (nmi_watchdog != 0)
1031 		check_nmi_watchdog();
1032 
1033 smp_done:
1034 	zap_low_mappings();
1035 	time_init_smp();
1036 }
1037