1 /*
2  * Copyright (c) 1996, by Steve Passe
3  * All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. The name of the developer may NOT be used to endorse or promote products
11  *    derived from this software without specific prior written permission.
12  *
13  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
17  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
23  * SUCH DAMAGE.
24  *
25  * $FreeBSD: src/sys/i386/i386/mp_machdep.c,v 1.115.2.15 2003/03/14 21:22:35 jhb Exp $
26  */
27 
28 #include "opt_cpu.h"
29 
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/kernel.h>
33 #include <sys/sysctl.h>
34 #include <sys/malloc.h>
35 #include <sys/memrange.h>
36 #include <sys/cons.h>	/* cngetc() */
37 #include <sys/machintr.h>
38 
39 #include <sys/mplock2.h>
40 
41 #include <vm/vm.h>
42 #include <vm/vm_param.h>
43 #include <vm/pmap.h>
44 #include <vm/vm_kern.h>
45 #include <vm/vm_extern.h>
46 #include <sys/lock.h>
47 #include <vm/vm_map.h>
48 #include <sys/user.h>
49 #ifdef GPROF
50 #include <sys/gmon.h>
51 #endif
52 
53 #include <machine/smp.h>
54 #include <machine_base/apic/apicreg.h>
55 #include <machine/atomic.h>
56 #include <machine/cpufunc.h>
57 #include <machine/cputypes.h>
58 #include <machine_base/apic/lapic.h>
59 #include <machine_base/apic/ioapic.h>
60 #include <machine/psl.h>
61 #include <machine/segments.h>
62 #include <machine/tss.h>
63 #include <machine/specialreg.h>
64 #include <machine/globaldata.h>
65 #include <machine/pmap_inval.h>
66 
67 #include <machine/md_var.h>		/* setidt() */
68 #include <machine_base/icu/icu.h>	/* IPIs */
69 #include <machine_base/icu/icu_var.h>
70 #include <machine_base/apic/ioapic_abi.h>
71 #include <machine/intr_machdep.h>	/* IPIs */
72 
73 #define WARMBOOT_TARGET		0
74 #define WARMBOOT_OFF		(KERNBASE + 0x0467)
75 #define WARMBOOT_SEG		(KERNBASE + 0x0469)
76 
77 #define CMOS_REG		(0x70)
78 #define CMOS_DATA		(0x71)
79 #define BIOS_RESET		(0x0f)
80 #define BIOS_WARM		(0x0a)
81 
82 /*
83  * this code MUST be enabled here and in mpboot.s.
84  * it follows the very early stages of AP boot by placing values in CMOS ram.
85  * it NORMALLY will never be needed and thus the primitive method for enabling.
86  *
87  */
88 #if defined(CHECK_POINTS)
89 #define CHECK_READ(A)	 (outb(CMOS_REG, (A)), inb(CMOS_DATA))
90 #define CHECK_WRITE(A,D) (outb(CMOS_REG, (A)), outb(CMOS_DATA, (D)))
91 
92 #define CHECK_INIT(D);				\
93 	CHECK_WRITE(0x34, (D));			\
94 	CHECK_WRITE(0x35, (D));			\
95 	CHECK_WRITE(0x36, (D));			\
96 	CHECK_WRITE(0x37, (D));			\
97 	CHECK_WRITE(0x38, (D));			\
98 	CHECK_WRITE(0x39, (D));
99 
100 #define CHECK_PRINT(S);				\
101 	kprintf("%s: %d, %d, %d, %d, %d, %d\n",	\
102 	   (S),					\
103 	   CHECK_READ(0x34),			\
104 	   CHECK_READ(0x35),			\
105 	   CHECK_READ(0x36),			\
106 	   CHECK_READ(0x37),			\
107 	   CHECK_READ(0x38),			\
108 	   CHECK_READ(0x39));
109 
110 #else				/* CHECK_POINTS */
111 
112 #define CHECK_INIT(D)
113 #define CHECK_PRINT(S)
114 
115 #endif				/* CHECK_POINTS */
116 
117 /*
118  * Values to send to the POST hardware.
119  */
120 #define MP_BOOTADDRESS_POST	0x10
121 #define MP_PROBE_POST		0x11
122 #define MPTABLE_PASS1_POST	0x12
123 
124 #define MP_START_POST		0x13
125 #define MP_ENABLE_POST		0x14
126 #define MPTABLE_PASS2_POST	0x15
127 
128 #define START_ALL_APS_POST	0x16
129 #define INSTALL_AP_TRAMP_POST	0x17
130 #define START_AP_POST		0x18
131 
132 #define MP_ANNOUNCE_POST	0x19
133 
134 /** XXX FIXME: where does this really belong, isa.h/isa.c perhaps? */
135 int	current_postcode;
136 
137 /** XXX FIXME: what system files declare these??? */
138 extern struct region_descriptor r_gdt;
139 
140 extern int nkpt;
141 extern int naps;
142 
143 int64_t tsc0_offset;
144 extern int64_t tsc_offsets[];
145 
146 /* AP uses this during bootstrap.  Do not staticize.  */
147 char *bootSTK;
148 static int bootAP;
149 
150 struct pcb stoppcbs[MAXCPU];
151 
152 extern inthand_t IDTVEC(fast_syscall), IDTVEC(fast_syscall32);
153 
154 /*
155  * Local data and functions.
156  */
157 
158 static u_int	boot_address;
159 static int	mp_finish;
160 static int	mp_finish_lapic;
161 
162 static int	start_all_aps(u_int boot_addr);
163 #if 0
164 static void	install_ap_tramp(u_int boot_addr);
165 #endif
166 static int	start_ap(struct mdglobaldata *gd, u_int boot_addr, int smibest);
167 static int	smitest(void);
168 static void	mp_bsp_simple_setup(void);
169 
170 static cpumask_t smp_startup_mask = 1;	/* which cpus have been started */
171 static cpumask_t smp_lapic_mask = 1;	/* which cpus have lapic been inited */
172 cpumask_t smp_active_mask = 1;	/* which cpus are ready for IPIs etc? */
173 SYSCTL_INT(_machdep, OID_AUTO, smp_active, CTLFLAG_RD, &smp_active_mask, 0, "");
174 static u_int	bootMP_size;
175 
176 /* Local data for detecting CPU TOPOLOGY */
177 static int core_bits = 0;
178 static int logical_CPU_bits = 0;
179 
180 
181 /*
182  * Calculate usable address in base memory for AP trampoline code.
183  */
184 u_int
185 mp_bootaddress(u_int basemem)
186 {
187 	POSTCODE(MP_BOOTADDRESS_POST);
188 
189 	bootMP_size = mptramp_end - mptramp_start;
190 	boot_address = trunc_page(basemem * 1024); /* round down to 4k boundary */
191 	if (((basemem * 1024) - boot_address) < bootMP_size)
192 		boot_address -= PAGE_SIZE;	/* not enough, lower by 4k */
193 	/* 3 levels of page table pages */
194 	mptramp_pagetables = boot_address - (PAGE_SIZE * 3);
195 
196 	return mptramp_pagetables;
197 }
198 
199 /*
200  * Print various information about the SMP system hardware and setup.
201  */
202 void
203 mp_announce(void)
204 {
205 	int     x;
206 
207 	POSTCODE(MP_ANNOUNCE_POST);
208 
209 	kprintf("DragonFly/MP: Multiprocessor motherboard\n");
210 	kprintf(" cpu0 (BSP): apic id: %2d\n", CPUID_TO_APICID(0));
211 	for (x = 1; x <= naps; ++x)
212 		kprintf(" cpu%d (AP):  apic id: %2d\n", x, CPUID_TO_APICID(x));
213 
214 	if (!ioapic_enable)
215 		kprintf(" Warning: APIC I/O disabled\n");
216 }
217 
218 /*
219  * AP cpu's call this to sync up protected mode.
220  *
221  * WARNING! %gs is not set up on entry.  This routine sets up %gs.
222  */
223 void
224 init_secondary(void)
225 {
226 	int	gsel_tss;
227 	int	x, myid = bootAP;
228 	u_int64_t msr, cr0;
229 	struct mdglobaldata *md;
230 	struct privatespace *ps;
231 
232 	ps = &CPU_prvspace[myid];
233 
234 	gdt_segs[GPROC0_SEL].ssd_base =
235 		(long) &ps->mdglobaldata.gd_common_tss;
236 	ps->mdglobaldata.mi.gd_prvspace = ps;
237 
238 	/* We fill the 32-bit segment descriptors */
239 	for (x = 0; x < NGDT; x++) {
240 		if (x != GPROC0_SEL && x != (GPROC0_SEL + 1))
241 			ssdtosd(&gdt_segs[x], &gdt[myid * NGDT + x]);
242 	}
243 	/* And now a 64-bit one */
244 	ssdtosyssd(&gdt_segs[GPROC0_SEL],
245 	    (struct system_segment_descriptor *)&gdt[myid * NGDT + GPROC0_SEL]);
246 
247 	r_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
248 	r_gdt.rd_base = (long) &gdt[myid * NGDT];
249 	lgdt(&r_gdt);			/* does magic intra-segment return */
250 
251 	/* lgdt() destroys the GSBASE value, so we load GSBASE after lgdt() */
252 	wrmsr(MSR_FSBASE, 0);		/* User value */
253 	wrmsr(MSR_GSBASE, (u_int64_t)ps);
254 	wrmsr(MSR_KGSBASE, 0);		/* XXX User value while we're in the kernel */
255 
256 	lidt(&r_idt_arr[mdcpu->mi.gd_cpuid]);
257 
258 #if 0
259 	lldt(_default_ldt);
260 	mdcpu->gd_currentldt = _default_ldt;
261 #endif
262 
263 	gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
264 	gdt[myid * NGDT + GPROC0_SEL].sd_type = SDT_SYSTSS;
265 
266 	md = mdcpu;	/* loaded through %gs:0 (mdglobaldata.mi.gd_prvspace)*/
267 
268 	md->gd_common_tss.tss_rsp0 = 0;	/* not used until after switch */
269 #if 0 /* JG XXX */
270 	md->gd_common_tss.tss_ioopt = (sizeof md->gd_common_tss) << 16;
271 #endif
272 	md->gd_tss_gdt = &gdt[myid * NGDT + GPROC0_SEL];
273 	md->gd_common_tssd = *md->gd_tss_gdt;
274 
275 	/* double fault stack */
276 	md->gd_common_tss.tss_ist1 =
277 		(long)&md->mi.gd_prvspace->idlestack[
278 			sizeof(md->mi.gd_prvspace->idlestack)];
279 
280 	ltr(gsel_tss);
281 
282 	/*
283 	 * Set to a known state:
284 	 * Set by mpboot.s: CR0_PG, CR0_PE
285 	 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
286 	 */
287 	cr0 = rcr0();
288 	cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
289 	load_cr0(cr0);
290 
291 	/* Set up the fast syscall stuff */
292 	msr = rdmsr(MSR_EFER) | EFER_SCE;
293 	wrmsr(MSR_EFER, msr);
294 	wrmsr(MSR_LSTAR, (u_int64_t)IDTVEC(fast_syscall));
295 	wrmsr(MSR_CSTAR, (u_int64_t)IDTVEC(fast_syscall32));
296 	msr = ((u_int64_t)GSEL(GCODE_SEL, SEL_KPL) << 32) |
297 	      ((u_int64_t)GSEL(GUCODE32_SEL, SEL_UPL) << 48);
298 	wrmsr(MSR_STAR, msr);
299 	wrmsr(MSR_SF_MASK, PSL_NT|PSL_T|PSL_I|PSL_C|PSL_D|PSL_IOPL);
300 
301 	pmap_set_opt();		/* PSE/4MB pages, etc */
302 	pmap_init_pat();	/* Page Attribute Table */
303 
304 	/* set up CPU registers and state */
305 	cpu_setregs();
306 
307 	/* set up SSE/NX registers */
308 	initializecpu();
309 
310 	/* set up FPU state on the AP */
311 	npxinit(__INITIAL_FPUCW__);
312 
313 	/* disable the APIC, just to be SURE */
314 	lapic->svr &= ~APIC_SVR_ENABLE;
315 }
316 
317 /*******************************************************************
318  * local functions and data
319  */
320 
321 /*
322  * Start the SMP system
323  */
324 static void
325 mp_start_aps(void *dummy __unused)
326 {
327 	if (lapic_enable) {
328 		/* start each Application Processor */
329 		start_all_aps(boot_address);
330 	} else {
331 		mp_bsp_simple_setup();
332 	}
333 }
334 SYSINIT(startaps, SI_BOOT2_START_APS, SI_ORDER_FIRST, mp_start_aps, NULL)
335 
336 /*
337  * start each AP in our list
338  */
339 static int
340 start_all_aps(u_int boot_addr)
341 {
342 	vm_offset_t va = boot_address + KERNBASE;
343 	u_int64_t *pt4, *pt3, *pt2;
344 	int     x, i;
345 	int	shift;
346 	int	smicount;
347 	int	smibest;
348 	int	smilast;
349 	u_char  mpbiosreason;
350 	u_long  mpbioswarmvec;
351 	struct mdglobaldata *gd;
352 	struct privatespace *ps;
353 
354 	POSTCODE(START_ALL_APS_POST);
355 
356 	/* install the AP 1st level boot code */
357 	pmap_kenter(va, boot_address);
358 	cpu_invlpg((void *)va);		/* JG XXX */
359 	bcopy(mptramp_start, (void *)va, bootMP_size);
360 
361 	/* Locate the page tables, they'll be below the trampoline */
362 	pt4 = (u_int64_t *)(uintptr_t)(mptramp_pagetables + KERNBASE);
363 	pt3 = pt4 + (PAGE_SIZE) / sizeof(u_int64_t);
364 	pt2 = pt3 + (PAGE_SIZE) / sizeof(u_int64_t);
365 
366 	/* Create the initial 1GB replicated page tables */
367 	for (i = 0; i < 512; i++) {
368 		/* Each slot of the level 4 pages points to the same level 3 page */
369 		pt4[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + PAGE_SIZE);
370 		pt4[i] |= PG_V | PG_RW | PG_U;
371 
372 		/* Each slot of the level 3 pages points to the same level 2 page */
373 		pt3[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + (2 * PAGE_SIZE));
374 		pt3[i] |= PG_V | PG_RW | PG_U;
375 
376 		/* The level 2 page slots are mapped with 2MB pages for 1GB. */
377 		pt2[i] = i * (2 * 1024 * 1024);
378 		pt2[i] |= PG_V | PG_RW | PG_PS | PG_U;
379 	}
380 
381 	/* save the current value of the warm-start vector */
382 	mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF);
383 	outb(CMOS_REG, BIOS_RESET);
384 	mpbiosreason = inb(CMOS_DATA);
385 
386 	/* setup a vector to our boot code */
387 	*((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
388 	*((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4);
389 	outb(CMOS_REG, BIOS_RESET);
390 	outb(CMOS_DATA, BIOS_WARM);	/* 'warm-start' */
391 
392 	/*
393 	 * If we have a TSC we can figure out the SMI interrupt rate.
394 	 * The SMI does not necessarily use a constant rate.  Spend
395 	 * up to 250ms trying to figure it out.
396 	 */
397 	smibest = 0;
398 	if (cpu_feature & CPUID_TSC) {
399 		set_apic_timer(275000);
400 		smilast = read_apic_timer();
401 		for (x = 0; x < 20 && read_apic_timer(); ++x) {
402 			smicount = smitest();
403 			if (smibest == 0 || smilast - smicount < smibest)
404 				smibest = smilast - smicount;
405 			smilast = smicount;
406 		}
407 		if (smibest > 250000)
408 			smibest = 0;
409 		if (smibest) {
410 			smibest = smibest * (int64_t)1000000 /
411 				  get_apic_timer_frequency();
412 		}
413 	}
414 	if (smibest)
415 		kprintf("SMI Frequency (worst case): %d Hz (%d us)\n",
416 			1000000 / smibest, smibest);
417 
418 	/* start each AP */
419 	for (x = 1; x <= naps; ++x) {
420 
421 		/* This is a bit verbose, it will go away soon.  */
422 
423 		/* allocate new private data page(s) */
424 		gd = (struct mdglobaldata *)kmem_alloc(&kernel_map,
425 				MDGLOBALDATA_BASEALLOC_SIZE);
426 
427 		gd = &CPU_prvspace[x].mdglobaldata;	/* official location */
428 		bzero(gd, sizeof(*gd));
429 		gd->mi.gd_prvspace = ps = &CPU_prvspace[x];
430 
431 		/* prime data page for it to use */
432 		mi_gdinit(&gd->mi, x);
433 		cpu_gdinit(gd, x);
434 		gd->mi.gd_ipiq = (void *)kmem_alloc(&kernel_map, sizeof(lwkt_ipiq) * (naps + 1));
435 		bzero(gd->mi.gd_ipiq, sizeof(lwkt_ipiq) * (naps + 1));
436 
437 		/* setup a vector to our boot code */
438 		*((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
439 		*((volatile u_short *) WARMBOOT_SEG) = (boot_addr >> 4);
440 		outb(CMOS_REG, BIOS_RESET);
441 		outb(CMOS_DATA, BIOS_WARM);	/* 'warm-start' */
442 
443 		/*
444 		 * Setup the AP boot stack
445 		 */
446 		bootSTK = &ps->idlestack[UPAGES*PAGE_SIZE/2];
447 		bootAP = x;
448 
449 		/* attempt to start the Application Processor */
450 		CHECK_INIT(99);	/* setup checkpoints */
451 		if (!start_ap(gd, boot_addr, smibest)) {
452 			kprintf("\nAP #%d (PHY# %d) failed!\n",
453 				x, CPUID_TO_APICID(x));
454 			CHECK_PRINT("trace");	/* show checkpoints */
455 			/* better panic as the AP may be running loose */
456 			kprintf("panic y/n? [y] ");
457 			if (cngetc() != 'n')
458 				panic("bye-bye");
459 		}
460 		CHECK_PRINT("trace");		/* show checkpoints */
461 	}
462 
463 	/* set ncpus to 1 + highest logical cpu.  Not all may have come up */
464 	ncpus = x;
465 
466 	/* ncpus2 -- ncpus rounded down to the nearest power of 2 */
467 	for (shift = 0; (1 << shift) <= ncpus; ++shift)
468 		;
469 	--shift;
470 	ncpus2_shift = shift;
471 	ncpus2 = 1 << shift;
472 	ncpus2_mask = ncpus2 - 1;
473 
474 	/* ncpus_fit -- ncpus rounded up to the nearest power of 2 */
475 	if ((1 << shift) < ncpus)
476 		++shift;
477 	ncpus_fit = 1 << shift;
478 	ncpus_fit_mask = ncpus_fit - 1;
479 
480 	/* build our map of 'other' CPUs */
481 	mycpu->gd_other_cpus = smp_startup_mask & ~CPUMASK(mycpu->gd_cpuid);
482 	mycpu->gd_ipiq = (void *)kmem_alloc(&kernel_map, sizeof(lwkt_ipiq) * ncpus);
483 	bzero(mycpu->gd_ipiq, sizeof(lwkt_ipiq) * ncpus);
484 
485 	/* restore the warmstart vector */
486 	*(u_long *) WARMBOOT_OFF = mpbioswarmvec;
487 	outb(CMOS_REG, BIOS_RESET);
488 	outb(CMOS_DATA, mpbiosreason);
489 
490 	/*
491 	 * NOTE!  The idlestack for the BSP was setup by locore.  Finish
492 	 * up, clean out the P==V mapping we did earlier.
493 	 */
494 	pmap_set_opt();
495 
496 	/*
497 	 * Wait all APs to finish initializing LAPIC
498 	 */
499 	mp_finish_lapic = 1;
500 	if (bootverbose)
501 		kprintf("SMP: Waiting APs LAPIC initialization\n");
502 	if (cpu_feature & CPUID_TSC)
503 		tsc0_offset = rdtsc();
504 	tsc_offsets[0] = 0;
505 	rel_mplock();
506 	while (smp_lapic_mask != smp_startup_mask) {
507 		cpu_lfence();
508 		if (cpu_feature & CPUID_TSC)
509 			tsc0_offset = rdtsc();
510 	}
511 	while (try_mplock() == 0)
512 		;
513 
514 	/* number of APs actually started */
515 	return ncpus - 1;
516 }
517 
518 
519 /*
520  * load the 1st level AP boot code into base memory.
521  */
522 
523 /* targets for relocation */
524 extern void bigJump(void);
525 extern void bootCodeSeg(void);
526 extern void bootDataSeg(void);
527 extern void MPentry(void);
528 extern u_int MP_GDT;
529 extern u_int mp_gdtbase;
530 
531 #if 0
532 
533 static void
534 install_ap_tramp(u_int boot_addr)
535 {
536 	int     x;
537 	int     size = *(int *) ((u_long) & bootMP_size);
538 	u_char *src = (u_char *) ((u_long) bootMP);
539 	u_char *dst = (u_char *) boot_addr + KERNBASE;
540 	u_int   boot_base = (u_int) bootMP;
541 	u_int8_t *dst8;
542 	u_int16_t *dst16;
543 	u_int32_t *dst32;
544 
545 	POSTCODE(INSTALL_AP_TRAMP_POST);
546 
547 	for (x = 0; x < size; ++x)
548 		*dst++ = *src++;
549 
550 	/*
551 	 * modify addresses in code we just moved to basemem. unfortunately we
552 	 * need fairly detailed info about mpboot.s for this to work.  changes
553 	 * to mpboot.s might require changes here.
554 	 */
555 
556 	/* boot code is located in KERNEL space */
557 	dst = (u_char *) boot_addr + KERNBASE;
558 
559 	/* modify the lgdt arg */
560 	dst32 = (u_int32_t *) (dst + ((u_int) & mp_gdtbase - boot_base));
561 	*dst32 = boot_addr + ((u_int) & MP_GDT - boot_base);
562 
563 	/* modify the ljmp target for MPentry() */
564 	dst32 = (u_int32_t *) (dst + ((u_int) bigJump - boot_base) + 1);
565 	*dst32 = ((u_int) MPentry - KERNBASE);
566 
567 	/* modify the target for boot code segment */
568 	dst16 = (u_int16_t *) (dst + ((u_int) bootCodeSeg - boot_base));
569 	dst8 = (u_int8_t *) (dst16 + 1);
570 	*dst16 = (u_int) boot_addr & 0xffff;
571 	*dst8 = ((u_int) boot_addr >> 16) & 0xff;
572 
573 	/* modify the target for boot data segment */
574 	dst16 = (u_int16_t *) (dst + ((u_int) bootDataSeg - boot_base));
575 	dst8 = (u_int8_t *) (dst16 + 1);
576 	*dst16 = (u_int) boot_addr & 0xffff;
577 	*dst8 = ((u_int) boot_addr >> 16) & 0xff;
578 }
579 
580 #endif
581 
582 /*
583  * This function starts the AP (application processor) identified
584  * by the APIC ID 'physicalCpu'.  It does quite a "song and dance"
585  * to accomplish this.  This is necessary because of the nuances
586  * of the different hardware we might encounter.  It ain't pretty,
587  * but it seems to work.
588  *
589  * NOTE: eventually an AP gets to ap_init(), which is called just
590  * before the AP goes into the LWKT scheduler's idle loop.
591  */
592 static int
593 start_ap(struct mdglobaldata *gd, u_int boot_addr, int smibest)
594 {
595 	int     physical_cpu;
596 	int     vector;
597 	u_long  icr_lo, icr_hi;
598 
599 	POSTCODE(START_AP_POST);
600 
601 	/* get the PHYSICAL APIC ID# */
602 	physical_cpu = CPUID_TO_APICID(gd->mi.gd_cpuid);
603 
604 	/* calculate the vector */
605 	vector = (boot_addr >> 12) & 0xff;
606 
607 	/* We don't want anything interfering */
608 	cpu_disable_intr();
609 
610 	/* Make sure the target cpu sees everything */
611 	wbinvd();
612 
613 	/*
614 	 * Try to detect when a SMI has occurred, wait up to 200ms.
615 	 *
616 	 * If a SMI occurs during an AP reset but before we issue
617 	 * the STARTUP command, the AP may brick.  To work around
618 	 * this problem we hold off doing the AP startup until
619 	 * after we have detected the SMI.  Hopefully another SMI
620 	 * will not occur before we finish the AP startup.
621 	 *
622 	 * Retries don't seem to help.  SMIs have a window of opportunity
623 	 * and if USB->legacy keyboard emulation is enabled in the BIOS
624 	 * the interrupt rate can be quite high.
625 	 *
626 	 * NOTE: Don't worry about the L1 cache load, it might bloat
627 	 *	 ldelta a little but ndelta will be so huge when the SMI
628 	 *	 occurs the detection logic will still work fine.
629 	 */
630 	if (smibest) {
631 		set_apic_timer(200000);
632 		smitest();
633 	}
634 
635 	/*
636 	 * first we do an INIT/RESET IPI this INIT IPI might be run, reseting
637 	 * and running the target CPU. OR this INIT IPI might be latched (P5
638 	 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
639 	 * ignored.
640 	 *
641 	 * see apic/apicreg.h for icr bit definitions.
642 	 *
643 	 * TIME CRITICAL CODE, DO NOT DO ANY KPRINTFS IN THE HOT PATH.
644 	 */
645 
646 	/*
647 	 * Setup the address for the target AP.  We can setup
648 	 * icr_hi once and then just trigger operations with
649 	 * icr_lo.
650 	 */
651 	icr_hi = lapic->icr_hi & ~APIC_ID_MASK;
652 	icr_hi |= (physical_cpu << 24);
653 	icr_lo = lapic->icr_lo & 0xfff00000;
654 	lapic->icr_hi = icr_hi;
655 
656 	/*
657 	 * Do an INIT IPI: assert RESET
658 	 *
659 	 * Use edge triggered mode to assert INIT
660 	 */
661 	lapic->icr_lo = icr_lo | 0x00004500;
662 	while (lapic->icr_lo & APIC_DELSTAT_MASK)
663 		 /* spin */ ;
664 
665 	/*
666 	 * The spec calls for a 10ms delay but we may have to use a
667 	 * MUCH lower delay to avoid bricking an AP due to a fast SMI
668 	 * interrupt.  We have other loops here too and dividing by 2
669 	 * doesn't seem to be enough even after subtracting 350us,
670 	 * so we divide by 4.
671 	 *
672 	 * Our minimum delay is 150uS, maximum is 10ms.  If no SMI
673 	 * interrupt was detected we use the full 10ms.
674 	 */
675 	if (smibest == 0)
676 		u_sleep(10000);
677 	else if (smibest < 150 * 4 + 350)
678 		u_sleep(150);
679 	else if ((smibest - 350) / 4 < 10000)
680 		u_sleep((smibest - 350) / 4);
681 	else
682 		u_sleep(10000);
683 
684 	/*
685 	 * Do an INIT IPI: deassert RESET
686 	 *
687 	 * Use level triggered mode to deassert.  It is unclear
688 	 * why we need to do this.
689 	 */
690 	lapic->icr_lo = icr_lo | 0x00008500;
691 	while (lapic->icr_lo & APIC_DELSTAT_MASK)
692 		 /* spin */ ;
693 	u_sleep(150);				/* wait 150us */
694 
695 	/*
696 	 * Next we do a STARTUP IPI: the previous INIT IPI might still be
697 	 * latched, (P5 bug) this 1st STARTUP would then terminate
698 	 * immediately, and the previously started INIT IPI would continue. OR
699 	 * the previous INIT IPI has already run. and this STARTUP IPI will
700 	 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
701 	 * will run.
702 	 */
703 	lapic->icr_lo = icr_lo | 0x00000600 | vector;
704 	while (lapic->icr_lo & APIC_DELSTAT_MASK)
705 		 /* spin */ ;
706 	u_sleep(200);		/* wait ~200uS */
707 
708 	/*
709 	 * Finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
710 	 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
711 	 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
712 	 * recognized after hardware RESET or INIT IPI.
713 	 */
714 	lapic->icr_lo = icr_lo | 0x00000600 | vector;
715 	while (lapic->icr_lo & APIC_DELSTAT_MASK)
716 		 /* spin */ ;
717 
718 	/* Resume normal operation */
719 	cpu_enable_intr();
720 
721 	/* wait for it to start, see ap_init() */
722 	set_apic_timer(5000000);/* == 5 seconds */
723 	while (read_apic_timer()) {
724 		if (smp_startup_mask & CPUMASK(gd->mi.gd_cpuid))
725 			return 1;	/* return SUCCESS */
726 	}
727 
728 	return 0;		/* return FAILURE */
729 }
730 
731 static
732 int
733 smitest(void)
734 {
735 	int64_t	ltsc;
736 	int64_t	ntsc;
737 	int64_t	ldelta;
738 	int64_t	ndelta;
739 	int count;
740 
741 	ldelta = 0;
742 	ndelta = 0;
743 	while (read_apic_timer()) {
744 		ltsc = rdtsc();
745 		for (count = 0; count < 100; ++count)
746 			ntsc = rdtsc();	/* force loop to occur */
747 		if (ldelta) {
748 			ndelta = ntsc - ltsc;
749 			if (ldelta > ndelta)
750 				ldelta = ndelta;
751 			if (ndelta > ldelta * 2)
752 				break;
753 		} else {
754 			ldelta = ntsc - ltsc;
755 		}
756 	}
757 	return(read_apic_timer());
758 }
759 
760 /*
761  * Synchronously flush the TLB on all other CPU's.  The current cpu's
762  * TLB is not flushed.  If the caller wishes to flush the current cpu's
763  * TLB the caller must call cpu_invltlb() in addition to smp_invltlb().
764  *
765  * NOTE: If for some reason we were unable to start all cpus we cannot
766  *	 safely use broadcast IPIs.
767  */
768 
769 static cpumask_t smp_invltlb_req;
770 
771 #define SMP_INVLTLB_DEBUG
772 
773 void
774 smp_invltlb(void)
775 {
776 	struct mdglobaldata *md = mdcpu;
777 #ifdef SMP_INVLTLB_DEBUG
778 	long count = 0;
779 	long xcount = 0;
780 #endif
781 
782 	crit_enter_gd(&md->mi);
783 	md->gd_invltlb_ret = 0;
784 	++md->mi.gd_cnt.v_smpinvltlb;
785 	atomic_set_cpumask(&smp_invltlb_req, md->mi.gd_cpumask);
786 #ifdef SMP_INVLTLB_DEBUG
787 again:
788 #endif
789 	if (smp_startup_mask == smp_active_mask) {
790 		all_but_self_ipi(XINVLTLB_OFFSET);
791 	} else {
792 		selected_apic_ipi(smp_active_mask & ~md->mi.gd_cpumask,
793 				  XINVLTLB_OFFSET, APIC_DELMODE_FIXED);
794 	}
795 
796 #ifdef SMP_INVLTLB_DEBUG
797 	if (xcount)
798 		kprintf("smp_invltlb: ipi sent\n");
799 #endif
800 	while ((md->gd_invltlb_ret & smp_active_mask & ~md->mi.gd_cpumask) !=
801 	       (smp_active_mask & ~md->mi.gd_cpumask)) {
802 		cpu_mfence();
803 		cpu_pause();
804 #ifdef SMP_INVLTLB_DEBUG
805 		/* DEBUGGING */
806 		if (++count == 400000000) {
807 			print_backtrace(-1);
808 			kprintf("smp_invltlb: endless loop %08lx %08lx, "
809 				"rflags %016jx retry",
810 			      (long)md->gd_invltlb_ret,
811 			      (long)smp_invltlb_req,
812 			      (intmax_t)read_rflags());
813 			__asm __volatile ("sti");
814 			++xcount;
815 			if (xcount > 2)
816 				lwkt_process_ipiq();
817 			if (xcount > 3) {
818 				int bcpu = BSFCPUMASK(~md->gd_invltlb_ret &
819 						      ~md->mi.gd_cpumask &
820 						      smp_active_mask);
821 				globaldata_t xgd;
822 
823 				kprintf("bcpu %d\n", bcpu);
824 				xgd = globaldata_find(bcpu);
825 				kprintf("thread %p %s\n", xgd->gd_curthread, xgd->gd_curthread->td_comm);
826 			}
827 			if (xcount > 5)
828 				Debugger("giving up");
829 			count = 0;
830 			goto again;
831 		}
832 #endif
833 	}
834 	atomic_clear_cpumask(&smp_invltlb_req, md->mi.gd_cpumask);
835 	crit_exit_gd(&md->mi);
836 }
837 
838 /*
839  * Called from Xinvltlb assembly with interrupts disabled.  We didn't
840  * bother to bump the critical section count or nested interrupt count
841  * so only do very low level operations here.
842  */
843 void
844 smp_invltlb_intr(void)
845 {
846 	struct mdglobaldata *md = mdcpu;
847 	struct mdglobaldata *omd;
848 	cpumask_t mask;
849 	int cpu;
850 
851 	cpu_mfence();
852 	mask = smp_invltlb_req;
853 	cpu_invltlb();
854 	while (mask) {
855 		cpu = BSFCPUMASK(mask);
856 		mask &= ~CPUMASK(cpu);
857 		omd = (struct mdglobaldata *)globaldata_find(cpu);
858 		atomic_set_cpumask(&omd->gd_invltlb_ret, md->mi.gd_cpumask);
859 	}
860 }
861 
862 void
863 cpu_wbinvd_on_all_cpus_callback(void *arg)
864 {
865     wbinvd();
866 }
867 
868 void
869 smp_invlpg_range_cpusync(void *arg)
870 {
871 	vm_offset_t eva, sva, addr;
872 	sva = ((struct smp_invlpg_range_cpusync_arg *)arg)->sva;
873 	eva = ((struct smp_invlpg_range_cpusync_arg *)arg)->eva;
874 
875 	for (addr = sva; addr < eva; addr += PAGE_SIZE) {
876 		cpu_invlpg((void *)addr);
877 	}
878 }
879 
880 /*
881  * When called the executing CPU will send an IPI to all other CPUs
882  *  requesting that they halt execution.
883  *
884  * Usually (but not necessarily) called with 'other_cpus' as its arg.
885  *
886  *  - Signals all CPUs in map to stop.
887  *  - Waits for each to stop.
888  *
889  * Returns:
890  *  -1: error
891  *   0: NA
892  *   1: ok
893  *
894  * XXX FIXME: this is not MP-safe, needs a lock to prevent multiple CPUs
895  *            from executing at same time.
896  */
897 int
898 stop_cpus(cpumask_t map)
899 {
900 	map &= smp_active_mask;
901 
902 	/* send the Xcpustop IPI to all CPUs in map */
903 	selected_apic_ipi(map, XCPUSTOP_OFFSET, APIC_DELMODE_FIXED);
904 
905 	while ((stopped_cpus & map) != map)
906 		/* spin */ ;
907 
908 	return 1;
909 }
910 
911 
912 /*
913  * Called by a CPU to restart stopped CPUs.
914  *
915  * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
916  *
917  *  - Signals all CPUs in map to restart.
918  *  - Waits for each to restart.
919  *
920  * Returns:
921  *  -1: error
922  *   0: NA
923  *   1: ok
924  */
925 int
926 restart_cpus(cpumask_t map)
927 {
928 	/* signal other cpus to restart */
929 	started_cpus = map & smp_active_mask;
930 
931 	while ((stopped_cpus & map) != 0) /* wait for each to clear its bit */
932 		/* spin */ ;
933 
934 	return 1;
935 }
936 
937 /*
938  * This is called once the mpboot code has gotten us properly relocated
939  * and the MMU turned on, etc.   ap_init() is actually the idle thread,
940  * and when it returns the scheduler will call the real cpu_idle() main
941  * loop for the idlethread.  Interrupts are disabled on entry and should
942  * remain disabled at return.
943  */
944 void
945 ap_init(void)
946 {
947 	int	cpu_id;
948 
949 	/*
950 	 * Adjust smp_startup_mask to signal the BSP that we have started
951 	 * up successfully.  Note that we do not yet hold the BGL.  The BSP
952 	 * is waiting for our signal.
953 	 *
954 	 * We can't set our bit in smp_active_mask yet because we are holding
955 	 * interrupts physically disabled and remote cpus could deadlock
956 	 * trying to send us an IPI.
957 	 */
958 	smp_startup_mask |= CPUMASK(mycpu->gd_cpuid);
959 	cpu_mfence();
960 
961 	/*
962 	 * Interlock for LAPIC initialization.  Wait until mp_finish_lapic is
963 	 * non-zero, then get the MP lock.
964 	 *
965 	 * Note: We are in a critical section.
966 	 *
967 	 * Note: we are the idle thread, we can only spin.
968 	 *
969 	 * Note: The load fence is memory volatile and prevents the compiler
970 	 * from improperly caching mp_finish_lapic, and the cpu from improperly
971 	 * caching it.
972 	 */
973 	while (mp_finish_lapic == 0)
974 		cpu_lfence();
975 	while (try_mplock() == 0)
976 		;
977 
978 	if (cpu_feature & CPUID_TSC) {
979 		/*
980 		 * The BSP is constantly updating tsc0_offset, figure out
981 		 * the relative difference to synchronize ktrdump.
982 		 */
983 		tsc_offsets[mycpu->gd_cpuid] = rdtsc() - tsc0_offset;
984 	}
985 
986 	/* BSP may have changed PTD while we're waiting for the lock */
987 	cpu_invltlb();
988 
989 	/* Build our map of 'other' CPUs. */
990 	mycpu->gd_other_cpus = smp_startup_mask & ~CPUMASK(mycpu->gd_cpuid);
991 
992 	/* A quick check from sanity claus */
993 	cpu_id = APICID_TO_CPUID((lapic->id & 0xff000000) >> 24);
994 	if (mycpu->gd_cpuid != cpu_id) {
995 		kprintf("SMP: assigned cpuid = %d\n", mycpu->gd_cpuid);
996 		kprintf("SMP: actual cpuid = %d lapicid %d\n",
997 			cpu_id, (lapic->id & 0xff000000) >> 24);
998 #if JGXXX
999 		kprintf("PTD[MPPTDI] = %p\n", (void *)PTD[MPPTDI]);
1000 #endif
1001 		panic("cpuid mismatch! boom!!");
1002 	}
1003 
1004 	/* Initialize AP's local APIC for irq's */
1005 	lapic_init(FALSE);
1006 
1007 	/* LAPIC initialization is done */
1008 	smp_lapic_mask |= CPUMASK(mycpu->gd_cpuid);
1009 	cpu_mfence();
1010 
1011 	/* Let BSP move onto the next initialization stage */
1012 	rel_mplock();
1013 
1014 	/*
1015 	 * Interlock for finalization.  Wait until mp_finish is non-zero,
1016 	 * then get the MP lock.
1017 	 *
1018 	 * Note: We are in a critical section.
1019 	 *
1020 	 * Note: we are the idle thread, we can only spin.
1021 	 *
1022 	 * Note: The load fence is memory volatile and prevents the compiler
1023 	 * from improperly caching mp_finish, and the cpu from improperly
1024 	 * caching it.
1025 	 */
1026 	while (mp_finish == 0)
1027 		cpu_lfence();
1028 	while (try_mplock() == 0)
1029 		;
1030 
1031 	/* BSP may have changed PTD while we're waiting for the lock */
1032 	cpu_invltlb();
1033 
1034 	/* Set memory range attributes for this CPU to match the BSP */
1035 	mem_range_AP_init();
1036 
1037 	/*
1038 	 * Once we go active we must process any IPIQ messages that may
1039 	 * have been queued, because no actual IPI will occur until we
1040 	 * set our bit in the smp_active_mask.  If we don't the IPI
1041 	 * message interlock could be left set which would also prevent
1042 	 * further IPIs.
1043 	 *
1044 	 * The idle loop doesn't expect the BGL to be held and while
1045 	 * lwkt_switch() normally cleans things up this is a special case
1046 	 * because we returning almost directly into the idle loop.
1047 	 *
1048 	 * The idle thread is never placed on the runq, make sure
1049 	 * nothing we've done put it there.
1050 	 */
1051 	KKASSERT(get_mplock_count(curthread) == 1);
1052 	smp_active_mask |= CPUMASK(mycpu->gd_cpuid);
1053 
1054 	/*
1055 	 * Enable interrupts here.  idle_restore will also do it, but
1056 	 * doing it here lets us clean up any strays that got posted to
1057 	 * the CPU during the AP boot while we are still in a critical
1058 	 * section.
1059 	 */
1060 	__asm __volatile("sti; pause; pause"::);
1061 	bzero(mdcpu->gd_ipending, sizeof(mdcpu->gd_ipending));
1062 
1063 	initclocks_pcpu();	/* clock interrupts (via IPIs) */
1064 	lwkt_process_ipiq();
1065 
1066 	/*
1067 	 * Releasing the mp lock lets the BSP finish up the SMP init
1068 	 */
1069 	rel_mplock();
1070 	KKASSERT((curthread->td_flags & TDF_RUNQ) == 0);
1071 }
1072 
1073 /*
1074  * Get SMP fully working before we start initializing devices.
1075  */
1076 static
1077 void
1078 ap_finish(void)
1079 {
1080 	mp_finish = 1;
1081 	if (bootverbose)
1082 		kprintf("Finish MP startup\n");
1083 	rel_mplock();
1084 	while (smp_active_mask != smp_startup_mask)
1085 		cpu_lfence();
1086 	while (try_mplock() == 0)
1087 		;
1088 	if (bootverbose) {
1089 		kprintf("Active CPU Mask: %016jx\n",
1090 			(uintmax_t)smp_active_mask);
1091 	}
1092 }
1093 
1094 SYSINIT(finishsmp, SI_BOOT2_FINISH_SMP, SI_ORDER_FIRST, ap_finish, NULL)
1095 
1096 void
1097 cpu_send_ipiq(int dcpu)
1098 {
1099         if (CPUMASK(dcpu) & smp_active_mask)
1100                 single_apic_ipi(dcpu, XIPIQ_OFFSET, APIC_DELMODE_FIXED);
1101 }
1102 
1103 #if 0	/* single_apic_ipi_passive() not working yet */
1104 /*
1105  * Returns 0 on failure, 1 on success
1106  */
1107 int
1108 cpu_send_ipiq_passive(int dcpu)
1109 {
1110         int r = 0;
1111         if (CPUMASK(dcpu) & smp_active_mask) {
1112                 r = single_apic_ipi_passive(dcpu, XIPIQ_OFFSET,
1113                                         APIC_DELMODE_FIXED);
1114         }
1115 	return(r);
1116 }
1117 #endif
1118 
1119 static void
1120 mp_bsp_simple_setup(void)
1121 {
1122 	/* build our map of 'other' CPUs */
1123 	mycpu->gd_other_cpus = smp_startup_mask & ~CPUMASK(mycpu->gd_cpuid);
1124 	mycpu->gd_ipiq = (void *)kmem_alloc(&kernel_map, sizeof(lwkt_ipiq) * ncpus);
1125 	bzero(mycpu->gd_ipiq, sizeof(lwkt_ipiq) * ncpus);
1126 
1127 	pmap_set_opt();
1128 
1129 	if (cpu_feature & CPUID_TSC)
1130 		tsc0_offset = rdtsc();
1131 }
1132 
1133 
1134 /*
1135  * CPU TOPOLOGY DETECTION FUNCTIONS
1136  */
1137 
1138 /* Detect intel topology using CPUID
1139  * Ref: http://www.intel.com/Assets/PDF/appnote/241618.pdf, pg 41
1140  */
1141 static void
1142 detect_intel_topology(int count_htt_cores)
1143 {
1144 	int shift = 0;
1145 	int ecx_index = 0;
1146 	int core_plus_logical_bits = 0;
1147 	int cores_per_package;
1148 	int logical_per_package;
1149 	int logical_per_core;
1150 	unsigned int p[4];
1151 
1152 	if (cpu_high >= 0xb) {
1153 		goto FUNC_B;
1154 
1155 	} else if (cpu_high >= 0x4) {
1156 		goto FUNC_4;
1157 
1158 	} else {
1159 		core_bits = 0;
1160 		for (shift = 0; (1 << shift) < count_htt_cores; ++shift)
1161 			;
1162 		logical_CPU_bits = 1 << shift;
1163 		return;
1164 	}
1165 
1166 FUNC_B:
1167 	cpuid_count(0xb, FUNC_B_THREAD_LEVEL, p);
1168 
1169 	/* if 0xb not supported - fallback to 0x4 */
1170 	if (p[1] == 0 || (FUNC_B_TYPE(p[2]) != FUNC_B_THREAD_TYPE)) {
1171 		goto FUNC_4;
1172 	}
1173 
1174 	logical_CPU_bits = FUNC_B_BITS_SHIFT_NEXT_LEVEL(p[0]);
1175 
1176 	ecx_index = FUNC_B_THREAD_LEVEL + 1;
1177 	do {
1178 		cpuid_count(0xb, ecx_index, p);
1179 
1180 		/* Check for the Core type in the implemented sub leaves. */
1181 		if (FUNC_B_TYPE(p[2]) == FUNC_B_CORE_TYPE) {
1182 			core_plus_logical_bits = FUNC_B_BITS_SHIFT_NEXT_LEVEL(p[0]);
1183 			break;
1184 		}
1185 
1186 		ecx_index++;
1187 
1188 	} while (FUNC_B_TYPE(p[2]) != FUNC_B_INVALID_TYPE);
1189 
1190 	core_bits = core_plus_logical_bits - logical_CPU_bits;
1191 
1192 	return;
1193 
1194 FUNC_4:
1195 	cpuid_count(0x4, 0, p);
1196 	cores_per_package = FUNC_4_MAX_CORE_NO(p[0]) + 1;
1197 
1198 	logical_per_package = count_htt_cores;
1199 	logical_per_core = logical_per_package / cores_per_package;
1200 
1201 	for (shift = 0; (1 << shift) < logical_per_core; ++shift)
1202 		;
1203 	logical_CPU_bits = shift;
1204 
1205 	for (shift = 0; (1 << shift) < cores_per_package; ++shift)
1206 		;
1207 	core_bits = shift;
1208 
1209 	return;
1210 }
1211 
1212 /* Detect AMD topology using CPUID
1213  * Ref: http://support.amd.com/us/Embedded_TechDocs/25481.pdf, last page
1214  */
1215 static void
1216 detect_amd_topology(int count_htt_cores)
1217 {
1218 	int shift = 0;
1219 	if ((cpu_feature & CPUID_HTT)
1220 			&& (amd_feature2 & AMDID2_CMP)) {
1221 
1222 		if (cpu_procinfo2 & AMDID_COREID_SIZE) {
1223 			core_bits = (cpu_procinfo2 & AMDID_COREID_SIZE)
1224 			    >> AMDID_COREID_SIZE_SHIFT;
1225 		} else {
1226 			core_bits = (cpu_procinfo2 & AMDID_CMP_CORES) + 1;
1227 			for (shift = 0; (1 << shift) < core_bits; ++shift)
1228 				;
1229 			core_bits = shift;
1230 		}
1231 
1232 		logical_CPU_bits = count_htt_cores >> core_bits;
1233 		for (shift = 0; (1 << shift) < logical_CPU_bits; ++shift)
1234 			;
1235 		logical_CPU_bits = shift;
1236 	} else {
1237 		for (shift = 0; (1 << shift) < count_htt_cores; ++shift)
1238 			;
1239 		core_bits = shift;
1240 		logical_CPU_bits = 0;
1241 	}
1242 }
1243 
1244 /* Calculate
1245  * - logical_CPU_bits
1246  * - core_bits
1247  * With the values above (for AMD or INTEL) we are able to generally
1248  * detect the CPU topology (number of cores for each level):
1249  * Ref: http://wiki.osdev.org/Detecting_CPU_Topology_(80x86)
1250  * Ref: http://www.multicoreinfo.com/research/papers/whitepapers/Intel-detect-topology.pdf
1251  */
1252 void
1253 detect_cpu_topology(void)
1254 {
1255 	static int topology_detected = 0;
1256 	int count = 0;
1257 
1258 	if (topology_detected) {
1259 		goto OUT;
1260 	}
1261 
1262 	if ((cpu_feature & CPUID_HTT) == 0) {
1263 		core_bits = 0;
1264 		logical_CPU_bits = 0;
1265 		goto OUT;
1266 	} else {
1267 		count = (cpu_procinfo & CPUID_HTT_CORES)
1268 		    >> CPUID_HTT_CORE_SHIFT;
1269 	}
1270 
1271 	if (cpu_vendor_id == CPU_VENDOR_INTEL) {
1272 		detect_intel_topology(count);
1273 	} else if (cpu_vendor_id == CPU_VENDOR_AMD) {
1274 		detect_amd_topology(count);
1275 	}
1276 
1277 OUT:
1278 	if (bootverbose)
1279 		kprintf("Bits within APICID: logical_CPU_bits: %d; core_bits: %d\n",
1280 		    logical_CPU_bits, core_bits);
1281 
1282 	topology_detected = 1;
1283 }
1284 
1285 /* Interface functions to calculate chip_ID,
1286  * core_number and logical_number
1287  * Ref: http://wiki.osdev.org/Detecting_CPU_Topology_(80x86)
1288  */
1289 int
1290 get_chip_ID(int cpuid)
1291 {
1292 	return get_apicid_from_cpuid(cpuid) >>
1293 	    (logical_CPU_bits + core_bits);
1294 }
1295 
1296 int
1297 get_core_number_within_chip(int cpuid)
1298 {
1299 	return (get_apicid_from_cpuid(cpuid) >> logical_CPU_bits) &
1300 	    ( (1 << core_bits) -1);
1301 }
1302 
1303 int
1304 get_logical_CPU_number_within_core(int cpuid)
1305 {
1306 	return get_apicid_from_cpuid(cpuid) &
1307 	    ( (1 << logical_CPU_bits) -1);
1308 }
1309