2 * Copyright (c) 1996, by Steve Passe
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. The name of the developer may NOT be used to endorse or promote products
11 * derived from this software without specific prior written permission.
13 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
14 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
15 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
16 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
17 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
18 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
19 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
20 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
21 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
22 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25 * $FreeBSD: src/sys/i386/i386/mp_machdep.c,v 1.115.2.15 2003/03/14 21:22:35 jhb Exp $
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 #include <sys/cpu_topology.h>
40 #include <sys/mplock2.h>
43 #include <vm/vm_param.h>
45 #include <vm/vm_kern.h>
46 #include <vm/vm_extern.h>
48 #include <vm/vm_map.h>
54 #include <machine/smp.h>
55 #include <machine_base/apic/apicreg.h>
56 #include <machine/atomic.h>
57 #include <machine/cpufunc.h>
58 #include <machine/cputypes.h>
59 #include <machine_base/apic/lapic.h>
60 #include <machine_base/apic/ioapic.h>
61 #include <machine_base/acpica/acpi_md_cpu.h>
62 #include <machine/psl.h>
63 #include <machine/segments.h>
64 #include <machine/tss.h>
65 #include <machine/specialreg.h>
66 #include <machine/globaldata.h>
67 #include <machine/pmap_inval.h>
69 #include <machine/md_var.h> /* setidt() */
70 #include <machine_base/icu/icu.h> /* IPIs */
71 #include <machine_base/icu/icu_var.h>
72 #include <machine_base/apic/ioapic_abi.h>
73 #include <machine/intr_machdep.h> /* IPIs */
75 #define WARMBOOT_TARGET 0
76 #define WARMBOOT_OFF (KERNBASE + 0x0467)
77 #define WARMBOOT_SEG (KERNBASE + 0x0469)
79 #define CMOS_REG (0x70)
80 #define CMOS_DATA (0x71)
81 #define BIOS_RESET (0x0f)
82 #define BIOS_WARM (0x0a)
85 * this code MUST be enabled here and in mpboot.s.
86 * it follows the very early stages of AP boot by placing values in CMOS ram.
87 * it NORMALLY will never be needed and thus the primitive method for enabling.
90 #if defined(CHECK_POINTS)
91 #define CHECK_READ(A) (outb(CMOS_REG, (A)), inb(CMOS_DATA))
92 #define CHECK_WRITE(A,D) (outb(CMOS_REG, (A)), outb(CMOS_DATA, (D)))
94 #define CHECK_INIT(D); \
95 CHECK_WRITE(0x34, (D)); \
96 CHECK_WRITE(0x35, (D)); \
97 CHECK_WRITE(0x36, (D)); \
98 CHECK_WRITE(0x37, (D)); \
99 CHECK_WRITE(0x38, (D)); \
100 CHECK_WRITE(0x39, (D));
102 #define CHECK_PRINT(S); \
103 kprintf("%s: %d, %d, %d, %d, %d, %d\n", \
112 #else /* CHECK_POINTS */
114 #define CHECK_INIT(D)
115 #define CHECK_PRINT(S)
117 #endif /* CHECK_POINTS */
120 * Values to send to the POST hardware.
122 #define MP_BOOTADDRESS_POST 0x10
123 #define MP_PROBE_POST 0x11
124 #define MPTABLE_PASS1_POST 0x12
126 #define MP_START_POST 0x13
127 #define MP_ENABLE_POST 0x14
128 #define MPTABLE_PASS2_POST 0x15
130 #define START_ALL_APS_POST 0x16
131 #define INSTALL_AP_TRAMP_POST 0x17
132 #define START_AP_POST 0x18
134 #define MP_ANNOUNCE_POST 0x19
136 /** XXX FIXME: where does this really belong, isa.h/isa.c perhaps? */
137 int current_postcode;
139 /** XXX FIXME: what system files declare these??? */
140 extern struct region_descriptor r_gdt;
146 extern int64_t tsc_offsets[];
148 /* AP uses this during bootstrap. Do not staticize. */
152 struct pcb stoppcbs[MAXCPU];
154 extern inthand_t IDTVEC(fast_syscall), IDTVEC(fast_syscall32);
157 * Local data and functions.
160 static u_int boot_address;
161 static int mp_finish;
162 static int mp_finish_lapic;
164 static int start_all_aps(u_int boot_addr);
166 static void install_ap_tramp(u_int boot_addr);
168 static int start_ap(struct mdglobaldata *gd, u_int boot_addr, int smibest);
169 static int smitest(void);
170 static void mp_bsp_simple_setup(void);
172 /* which cpus have been started */
173 static cpumask_t smp_startup_mask = CPUMASK_INITIALIZER_ONLYONE;
174 /* which cpus have lapic been inited */
175 static cpumask_t smp_lapic_mask = CPUMASK_INITIALIZER_ONLYONE;
176 /* which cpus are ready for IPIs etc? */
177 cpumask_t smp_active_mask = CPUMASK_INITIALIZER_ONLYONE;
178 cpumask_t smp_finalize_mask = CPUMASK_INITIALIZER_ONLYONE;
180 SYSCTL_INT(_machdep, OID_AUTO, smp_active, CTLFLAG_RD, &smp_active_mask, 0, "");
181 static u_int bootMP_size;
183 /* Local data for detecting CPU TOPOLOGY */
184 static int core_bits = 0;
185 static int logical_CPU_bits = 0;
189 * Calculate usable address in base memory for AP trampoline code.
192 mp_bootaddress(u_int basemem)
194 POSTCODE(MP_BOOTADDRESS_POST);
196 bootMP_size = mptramp_end - mptramp_start;
197 boot_address = trunc_page(basemem * 1024); /* round down to 4k boundary */
198 if (((basemem * 1024) - boot_address) < bootMP_size)
199 boot_address -= PAGE_SIZE; /* not enough, lower by 4k */
200 /* 3 levels of page table pages */
201 mptramp_pagetables = boot_address - (PAGE_SIZE * 3);
203 return mptramp_pagetables;
207 * Print various information about the SMP system hardware and setup.
214 POSTCODE(MP_ANNOUNCE_POST);
216 kprintf("DragonFly/MP: Multiprocessor motherboard\n");
217 kprintf(" cpu0 (BSP): apic id: %2d\n", CPUID_TO_APICID(0));
218 for (x = 1; x <= naps; ++x)
219 kprintf(" cpu%d (AP): apic id: %2d\n", x, CPUID_TO_APICID(x));
222 kprintf(" Warning: APIC I/O disabled\n");
226 * AP cpu's call this to sync up protected mode.
228 * WARNING! %gs is not set up on entry. This routine sets up %gs.
234 int x, myid = bootAP;
236 struct mdglobaldata *md;
237 struct privatespace *ps;
239 ps = CPU_prvspace[myid];
241 gdt_segs[GPROC0_SEL].ssd_base =
242 (long) &ps->mdglobaldata.gd_common_tss;
243 ps->mdglobaldata.mi.gd_prvspace = ps;
245 /* We fill the 32-bit segment descriptors */
246 for (x = 0; x < NGDT; x++) {
247 if (x != GPROC0_SEL && x != (GPROC0_SEL + 1))
248 ssdtosd(&gdt_segs[x], &gdt[myid * NGDT + x]);
250 /* And now a 64-bit one */
251 ssdtosyssd(&gdt_segs[GPROC0_SEL],
252 (struct system_segment_descriptor *)&gdt[myid * NGDT + GPROC0_SEL]);
254 r_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
255 r_gdt.rd_base = (long) &gdt[myid * NGDT];
256 lgdt(&r_gdt); /* does magic intra-segment return */
258 /* lgdt() destroys the GSBASE value, so we load GSBASE after lgdt() */
259 wrmsr(MSR_FSBASE, 0); /* User value */
260 wrmsr(MSR_GSBASE, (u_int64_t)ps);
261 wrmsr(MSR_KGSBASE, 0); /* XXX User value while we're in the kernel */
263 lidt(&r_idt_arr[mdcpu->mi.gd_cpuid]);
267 mdcpu->gd_currentldt = _default_ldt;
270 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
271 gdt[myid * NGDT + GPROC0_SEL].sd_type = SDT_SYSTSS;
273 md = mdcpu; /* loaded through %gs:0 (mdglobaldata.mi.gd_prvspace)*/
275 md->gd_common_tss.tss_rsp0 = 0; /* not used until after switch */
277 md->gd_common_tss.tss_ioopt = (sizeof md->gd_common_tss) << 16;
279 md->gd_tss_gdt = &gdt[myid * NGDT + GPROC0_SEL];
280 md->gd_common_tssd = *md->gd_tss_gdt;
282 /* double fault stack */
283 md->gd_common_tss.tss_ist1 =
284 (long)&md->mi.gd_prvspace->idlestack[
285 sizeof(md->mi.gd_prvspace->idlestack)];
290 * Set to a known state:
291 * Set by mpboot.s: CR0_PG, CR0_PE
292 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
295 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
298 /* Set up the fast syscall stuff */
299 msr = rdmsr(MSR_EFER) | EFER_SCE;
300 wrmsr(MSR_EFER, msr);
301 wrmsr(MSR_LSTAR, (u_int64_t)IDTVEC(fast_syscall));
302 wrmsr(MSR_CSTAR, (u_int64_t)IDTVEC(fast_syscall32));
303 msr = ((u_int64_t)GSEL(GCODE_SEL, SEL_KPL) << 32) |
304 ((u_int64_t)GSEL(GUCODE32_SEL, SEL_UPL) << 48);
305 wrmsr(MSR_STAR, msr);
306 wrmsr(MSR_SF_MASK, PSL_NT|PSL_T|PSL_I|PSL_C|PSL_D|PSL_IOPL);
308 pmap_set_opt(); /* PSE/4MB pages, etc */
309 pmap_init_pat(); /* Page Attribute Table */
311 /* set up CPU registers and state */
314 /* set up SSE/NX registers */
317 /* set up FPU state on the AP */
320 /* disable the APIC, just to be SURE */
321 lapic->svr &= ~APIC_SVR_ENABLE;
324 /*******************************************************************
325 * local functions and data
329 * Start the SMP system
332 mp_start_aps(void *dummy __unused)
335 /* start each Application Processor */
336 start_all_aps(boot_address);
338 mp_bsp_simple_setup();
341 SYSINIT(startaps, SI_BOOT2_START_APS, SI_ORDER_FIRST, mp_start_aps, NULL);
344 * start each AP in our list
347 start_all_aps(u_int boot_addr)
349 vm_offset_t va = boot_address + KERNBASE;
350 u_int64_t *pt4, *pt3, *pt2;
358 u_long mpbioswarmvec;
359 struct mdglobaldata *gd;
360 struct privatespace *ps;
363 POSTCODE(START_ALL_APS_POST);
365 /* install the AP 1st level boot code */
366 pmap_kenter(va, boot_address);
367 cpu_invlpg((void *)va); /* JG XXX */
368 bcopy(mptramp_start, (void *)va, bootMP_size);
370 /* Locate the page tables, they'll be below the trampoline */
371 pt4 = (u_int64_t *)(uintptr_t)(mptramp_pagetables + KERNBASE);
372 pt3 = pt4 + (PAGE_SIZE) / sizeof(u_int64_t);
373 pt2 = pt3 + (PAGE_SIZE) / sizeof(u_int64_t);
375 /* Create the initial 1GB replicated page tables */
376 for (i = 0; i < 512; i++) {
377 /* Each slot of the level 4 pages points to the same level 3 page */
378 pt4[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + PAGE_SIZE);
379 pt4[i] |= kernel_pmap.pmap_bits[PG_V_IDX] |
380 kernel_pmap.pmap_bits[PG_RW_IDX] |
381 kernel_pmap.pmap_bits[PG_U_IDX];
383 /* Each slot of the level 3 pages points to the same level 2 page */
384 pt3[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + (2 * PAGE_SIZE));
385 pt3[i] |= kernel_pmap.pmap_bits[PG_V_IDX] |
386 kernel_pmap.pmap_bits[PG_RW_IDX] |
387 kernel_pmap.pmap_bits[PG_U_IDX];
389 /* The level 2 page slots are mapped with 2MB pages for 1GB. */
390 pt2[i] = i * (2 * 1024 * 1024);
391 pt2[i] |= kernel_pmap.pmap_bits[PG_V_IDX] |
392 kernel_pmap.pmap_bits[PG_RW_IDX] |
393 kernel_pmap.pmap_bits[PG_PS_IDX] |
394 kernel_pmap.pmap_bits[PG_U_IDX];
397 /* save the current value of the warm-start vector */
398 mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF);
399 outb(CMOS_REG, BIOS_RESET);
400 mpbiosreason = inb(CMOS_DATA);
402 /* setup a vector to our boot code */
403 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
404 *((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4);
405 outb(CMOS_REG, BIOS_RESET);
406 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
409 * If we have a TSC we can figure out the SMI interrupt rate.
410 * The SMI does not necessarily use a constant rate. Spend
411 * up to 250ms trying to figure it out.
414 if (cpu_feature & CPUID_TSC) {
415 set_apic_timer(275000);
416 smilast = read_apic_timer();
417 for (x = 0; x < 20 && read_apic_timer(); ++x) {
418 smicount = smitest();
419 if (smibest == 0 || smilast - smicount < smibest)
420 smibest = smilast - smicount;
423 if (smibest > 250000)
426 smibest = smibest * (int64_t)1000000 /
427 get_apic_timer_frequency();
431 kprintf("SMI Frequency (worst case): %d Hz (%d us)\n",
432 1000000 / smibest, smibest);
435 for (x = 1; x <= naps; ++x) {
436 /* This is a bit verbose, it will go away soon. */
438 pssize = sizeof(struct privatespace);
439 ps = (void *)kmem_alloc(&kernel_map, pssize);
440 CPU_prvspace[x] = ps;
442 kprintf("ps %d %p %d\n", x, ps, pssize);
445 gd = &ps->mdglobaldata;
446 gd->mi.gd_prvspace = ps;
448 /* prime data page for it to use */
449 mi_gdinit(&gd->mi, x);
451 ipiq_size = sizeof(struct lwkt_ipiq) * (naps + 1);
452 gd->mi.gd_ipiq = (void *)kmem_alloc(&kernel_map, ipiq_size);
453 bzero(gd->mi.gd_ipiq, ipiq_size);
455 gd->gd_acpi_id = CPUID_TO_ACPIID(gd->mi.gd_cpuid);
457 /* setup a vector to our boot code */
458 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
459 *((volatile u_short *) WARMBOOT_SEG) = (boot_addr >> 4);
460 outb(CMOS_REG, BIOS_RESET);
461 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
464 * Setup the AP boot stack
466 bootSTK = &ps->idlestack[UPAGES * PAGE_SIZE - PAGE_SIZE];
469 /* attempt to start the Application Processor */
470 CHECK_INIT(99); /* setup checkpoints */
471 if (!start_ap(gd, boot_addr, smibest)) {
472 kprintf("\nAP #%d (PHY# %d) failed!\n",
473 x, CPUID_TO_APICID(x));
474 CHECK_PRINT("trace"); /* show checkpoints */
475 /* better panic as the AP may be running loose */
476 kprintf("panic y/n? [y] ");
480 CHECK_PRINT("trace"); /* show checkpoints */
483 /* set ncpus to 1 + highest logical cpu. Not all may have come up */
486 /* ncpus2 -- ncpus rounded down to the nearest power of 2 */
487 for (shift = 0; (1 << shift) <= ncpus; ++shift)
490 ncpus2_shift = shift;
492 ncpus2_mask = ncpus2 - 1;
494 /* ncpus_fit -- ncpus rounded up to the nearest power of 2 */
495 if ((1 << shift) < ncpus)
497 ncpus_fit = 1 << shift;
498 ncpus_fit_mask = ncpus_fit - 1;
500 /* build our map of 'other' CPUs */
501 mycpu->gd_other_cpus = smp_startup_mask;
502 CPUMASK_NANDBIT(mycpu->gd_other_cpus, mycpu->gd_cpuid);
504 gd = (struct mdglobaldata *)mycpu;
505 gd->gd_acpi_id = CPUID_TO_ACPIID(mycpu->gd_cpuid);
507 ipiq_size = sizeof(struct lwkt_ipiq) * ncpus;
508 mycpu->gd_ipiq = (void *)kmem_alloc(&kernel_map, ipiq_size);
509 bzero(mycpu->gd_ipiq, ipiq_size);
511 /* restore the warmstart vector */
512 *(u_long *) WARMBOOT_OFF = mpbioswarmvec;
513 outb(CMOS_REG, BIOS_RESET);
514 outb(CMOS_DATA, mpbiosreason);
517 * NOTE! The idlestack for the BSP was setup by locore. Finish
518 * up, clean out the P==V mapping we did earlier.
523 * Wait all APs to finish initializing LAPIC
526 kprintf("SMP: Waiting APs LAPIC initialization\n");
527 if (cpu_feature & CPUID_TSC)
528 tsc0_offset = rdtsc();
533 while (CPUMASK_CMPMASKNEQ(smp_lapic_mask, smp_startup_mask)) {
536 if (cpu_feature & CPUID_TSC)
537 tsc0_offset = rdtsc();
539 while (try_mplock() == 0) {
544 /* number of APs actually started */
550 * load the 1st level AP boot code into base memory.
553 /* targets for relocation */
554 extern void bigJump(void);
555 extern void bootCodeSeg(void);
556 extern void bootDataSeg(void);
557 extern void MPentry(void);
559 extern u_int mp_gdtbase;
564 install_ap_tramp(u_int boot_addr)
567 int size = *(int *) ((u_long) & bootMP_size);
568 u_char *src = (u_char *) ((u_long) bootMP);
569 u_char *dst = (u_char *) boot_addr + KERNBASE;
570 u_int boot_base = (u_int) bootMP;
575 POSTCODE(INSTALL_AP_TRAMP_POST);
577 for (x = 0; x < size; ++x)
581 * modify addresses in code we just moved to basemem. unfortunately we
582 * need fairly detailed info about mpboot.s for this to work. changes
583 * to mpboot.s might require changes here.
586 /* boot code is located in KERNEL space */
587 dst = (u_char *) boot_addr + KERNBASE;
589 /* modify the lgdt arg */
590 dst32 = (u_int32_t *) (dst + ((u_int) & mp_gdtbase - boot_base));
591 *dst32 = boot_addr + ((u_int) & MP_GDT - boot_base);
593 /* modify the ljmp target for MPentry() */
594 dst32 = (u_int32_t *) (dst + ((u_int) bigJump - boot_base) + 1);
595 *dst32 = ((u_int) MPentry - KERNBASE);
597 /* modify the target for boot code segment */
598 dst16 = (u_int16_t *) (dst + ((u_int) bootCodeSeg - boot_base));
599 dst8 = (u_int8_t *) (dst16 + 1);
600 *dst16 = (u_int) boot_addr & 0xffff;
601 *dst8 = ((u_int) boot_addr >> 16) & 0xff;
603 /* modify the target for boot data segment */
604 dst16 = (u_int16_t *) (dst + ((u_int) bootDataSeg - boot_base));
605 dst8 = (u_int8_t *) (dst16 + 1);
606 *dst16 = (u_int) boot_addr & 0xffff;
607 *dst8 = ((u_int) boot_addr >> 16) & 0xff;
613 * This function starts the AP (application processor) identified
614 * by the APIC ID 'physicalCpu'. It does quite a "song and dance"
615 * to accomplish this. This is necessary because of the nuances
616 * of the different hardware we might encounter. It ain't pretty,
617 * but it seems to work.
619 * NOTE: eventually an AP gets to ap_init(), which is called just
620 * before the AP goes into the LWKT scheduler's idle loop.
623 start_ap(struct mdglobaldata *gd, u_int boot_addr, int smibest)
627 u_long icr_lo, icr_hi;
629 POSTCODE(START_AP_POST);
631 /* get the PHYSICAL APIC ID# */
632 physical_cpu = CPUID_TO_APICID(gd->mi.gd_cpuid);
634 /* calculate the vector */
635 vector = (boot_addr >> 12) & 0xff;
637 /* We don't want anything interfering */
640 /* Make sure the target cpu sees everything */
644 * Try to detect when a SMI has occurred, wait up to 200ms.
646 * If a SMI occurs during an AP reset but before we issue
647 * the STARTUP command, the AP may brick. To work around
648 * this problem we hold off doing the AP startup until
649 * after we have detected the SMI. Hopefully another SMI
650 * will not occur before we finish the AP startup.
652 * Retries don't seem to help. SMIs have a window of opportunity
653 * and if USB->legacy keyboard emulation is enabled in the BIOS
654 * the interrupt rate can be quite high.
656 * NOTE: Don't worry about the L1 cache load, it might bloat
657 * ldelta a little but ndelta will be so huge when the SMI
658 * occurs the detection logic will still work fine.
661 set_apic_timer(200000);
666 * first we do an INIT/RESET IPI this INIT IPI might be run, reseting
667 * and running the target CPU. OR this INIT IPI might be latched (P5
668 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
671 * see apic/apicreg.h for icr bit definitions.
673 * TIME CRITICAL CODE, DO NOT DO ANY KPRINTFS IN THE HOT PATH.
677 * Setup the address for the target AP. We can setup
678 * icr_hi once and then just trigger operations with
681 icr_hi = lapic->icr_hi & ~APIC_ID_MASK;
682 icr_hi |= (physical_cpu << 24);
683 icr_lo = lapic->icr_lo & 0xfff00000;
684 lapic->icr_hi = icr_hi;
687 * Do an INIT IPI: assert RESET
689 * Use edge triggered mode to assert INIT
691 lapic->icr_lo = icr_lo | 0x00004500;
692 while (lapic->icr_lo & APIC_DELSTAT_MASK)
696 * The spec calls for a 10ms delay but we may have to use a
697 * MUCH lower delay to avoid bricking an AP due to a fast SMI
698 * interrupt. We have other loops here too and dividing by 2
699 * doesn't seem to be enough even after subtracting 350us,
702 * Our minimum delay is 150uS, maximum is 10ms. If no SMI
703 * interrupt was detected we use the full 10ms.
707 else if (smibest < 150 * 4 + 350)
709 else if ((smibest - 350) / 4 < 10000)
710 u_sleep((smibest - 350) / 4);
715 * Do an INIT IPI: deassert RESET
717 * Use level triggered mode to deassert. It is unclear
718 * why we need to do this.
720 lapic->icr_lo = icr_lo | 0x00008500;
721 while (lapic->icr_lo & APIC_DELSTAT_MASK)
723 u_sleep(150); /* wait 150us */
726 * Next we do a STARTUP IPI: the previous INIT IPI might still be
727 * latched, (P5 bug) this 1st STARTUP would then terminate
728 * immediately, and the previously started INIT IPI would continue. OR
729 * the previous INIT IPI has already run. and this STARTUP IPI will
730 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
733 lapic->icr_lo = icr_lo | 0x00000600 | vector;
734 while (lapic->icr_lo & APIC_DELSTAT_MASK)
736 u_sleep(200); /* wait ~200uS */
739 * Finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
740 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
741 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
742 * recognized after hardware RESET or INIT IPI.
744 lapic->icr_lo = icr_lo | 0x00000600 | vector;
745 while (lapic->icr_lo & APIC_DELSTAT_MASK)
748 /* Resume normal operation */
751 /* wait for it to start, see ap_init() */
752 set_apic_timer(5000000);/* == 5 seconds */
753 while (read_apic_timer()) {
754 if (CPUMASK_TESTBIT(smp_startup_mask, gd->mi.gd_cpuid))
755 return 1; /* return SUCCESS */
758 return 0; /* return FAILURE */
773 while (read_apic_timer()) {
775 for (count = 0; count < 100; ++count)
776 ntsc = rdtsc(); /* force loop to occur */
778 ndelta = ntsc - ltsc;
781 if (ndelta > ldelta * 2)
784 ldelta = ntsc - ltsc;
787 return(read_apic_timer());
791 * Synchronously flush the TLB on all other CPU's. The current cpu's
792 * TLB is not flushed. If the caller wishes to flush the current cpu's
793 * TLB the caller must call cpu_invltlb() in addition to smp_invltlb().
795 * NOTE: If for some reason we were unable to start all cpus we cannot
796 * safely use broadcast IPIs.
799 static cpumask_t smp_invltlb_req;
801 #define SMP_INVLTLB_DEBUG
806 struct mdglobaldata *md = mdcpu;
807 #ifdef SMP_INVLTLB_DEBUG
814 crit_enter_gd(&md->mi);
815 CPUMASK_ASSZERO(md->gd_invltlb_ret);
816 ++md->mi.gd_cnt.v_smpinvltlb;
817 ATOMIC_CPUMASK_ORBIT(smp_invltlb_req, md->mi.gd_cpuid);
818 #ifdef SMP_INVLTLB_DEBUG
821 if (CPUMASK_CMPMASKEQ(smp_startup_mask, smp_active_mask)) {
822 all_but_self_ipi(XINVLTLB_OFFSET);
824 tmpmask = smp_active_mask;
825 CPUMASK_NANDMASK(tmpmask, md->mi.gd_cpumask);
826 selected_apic_ipi(tmpmask, XINVLTLB_OFFSET, APIC_DELMODE_FIXED);
829 #ifdef SMP_INVLTLB_DEBUG
831 kprintf("smp_invltlb: ipi sent\n");
834 tmpmask = smp_active_mask;
836 CPUMASK_ANDMASK(tmpmask, md->gd_invltlb_ret);
837 CPUMASK_NANDMASK(tmpmask, md->mi.gd_cpumask);
838 CPUMASK_NANDMASK(tmpmask2, md->mi.gd_cpumask);
840 if (CPUMASK_CMPMASKEQ(tmpmask, tmpmask2))
844 #ifdef SMP_INVLTLB_DEBUG
846 if (++count == 400000000) {
848 kprintf("smp_invltlb: endless loop %08lx %08lx, "
849 "rflags %016jx retry",
850 (long)CPUMASK_LOWMASK(md->gd_invltlb_ret),
851 (long)CPUMASK_LOWMASK(smp_invltlb_req),
852 (intmax_t)read_rflags());
853 __asm __volatile ("sti");
861 tmpmask = smp_active_mask;
862 CPUMASK_NANDMASK(tmpmask, md->gd_invltlb_ret);
863 CPUMASK_NANDMASK(tmpmask, md->mi.gd_cpumask);
864 bcpu = BSFCPUMASK(tmpmask);
866 kprintf("bcpu %d\n", bcpu);
867 xgd = globaldata_find(bcpu);
868 kprintf("thread %p %s\n", xgd->gd_curthread, xgd->gd_curthread->td_comm);
871 Debugger("giving up");
877 ATOMIC_CPUMASK_NANDBIT(smp_invltlb_req, md->mi.gd_cpuid);
878 crit_exit_gd(&md->mi);
882 * Called from Xinvltlb assembly with interrupts disabled. We didn't
883 * bother to bump the critical section count or nested interrupt count
884 * so only do very low level operations here.
887 smp_invltlb_intr(void)
889 struct mdglobaldata *md = mdcpu;
890 struct mdglobaldata *omd;
895 mask = smp_invltlb_req;
897 while (CPUMASK_TESTNZERO(mask)) {
898 cpu = BSFCPUMASK(mask);
899 CPUMASK_NANDBIT(mask, cpu);
900 omd = (struct mdglobaldata *)globaldata_find(cpu);
901 ATOMIC_CPUMASK_ORBIT(omd->gd_invltlb_ret, md->mi.gd_cpuid);
906 cpu_wbinvd_on_all_cpus_callback(void *arg)
912 smp_invlpg_range_cpusync(void *arg)
914 vm_offset_t eva, sva, addr;
915 sva = ((struct smp_invlpg_range_cpusync_arg *)arg)->sva;
916 eva = ((struct smp_invlpg_range_cpusync_arg *)arg)->eva;
918 for (addr = sva; addr < eva; addr += PAGE_SIZE) {
919 cpu_invlpg((void *)addr);
924 * When called the executing CPU will send an IPI to all other CPUs
925 * requesting that they halt execution.
927 * Usually (but not necessarily) called with 'other_cpus' as its arg.
929 * - Signals all CPUs in map to stop.
930 * - Waits for each to stop.
937 * XXX FIXME: this is not MP-safe, needs a lock to prevent multiple CPUs
938 * from executing at same time.
941 stop_cpus(cpumask_t map)
945 CPUMASK_ANDMASK(map, smp_active_mask);
947 /* send the Xcpustop IPI to all CPUs in map */
948 selected_apic_ipi(map, XCPUSTOP_OFFSET, APIC_DELMODE_FIXED);
952 CPUMASK_ANDMASK(mask, map);
954 } while (CPUMASK_CMPMASKNEQ(mask, map));
961 * Called by a CPU to restart stopped CPUs.
963 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
965 * - Signals all CPUs in map to restart.
966 * - Waits for each to restart.
974 restart_cpus(cpumask_t map)
978 /* signal other cpus to restart */
980 CPUMASK_ANDMASK(mask, smp_active_mask);
985 /* wait for each to clear its bit */
986 while (CPUMASK_CMPMASKNEQ(stopped_cpus, map))
993 * This is called once the mpboot code has gotten us properly relocated
994 * and the MMU turned on, etc. ap_init() is actually the idle thread,
995 * and when it returns the scheduler will call the real cpu_idle() main
996 * loop for the idlethread. Interrupts are disabled on entry and should
997 * remain disabled at return.
1005 * Adjust smp_startup_mask to signal the BSP that we have started
1006 * up successfully. Note that we do not yet hold the BGL. The BSP
1007 * is waiting for our signal.
1009 * We can't set our bit in smp_active_mask yet because we are holding
1010 * interrupts physically disabled and remote cpus could deadlock
1011 * trying to send us an IPI.
1013 ATOMIC_CPUMASK_ORBIT(smp_startup_mask, mycpu->gd_cpuid);
1017 * Interlock for LAPIC initialization. Wait until mp_finish_lapic is
1018 * non-zero, then get the MP lock.
1020 * Note: We are in a critical section.
1022 * Note: we are the idle thread, we can only spin.
1024 * Note: The load fence is memory volatile and prevents the compiler
1025 * from improperly caching mp_finish_lapic, and the cpu from improperly
1028 while (mp_finish_lapic == 0) {
1033 while (try_mplock() == 0) {
1039 if (cpu_feature & CPUID_TSC) {
1041 * The BSP is constantly updating tsc0_offset, figure out
1042 * the relative difference to synchronize ktrdump.
1044 tsc_offsets[mycpu->gd_cpuid] = rdtsc() - tsc0_offset;
1047 /* BSP may have changed PTD while we're waiting for the lock */
1050 /* Build our map of 'other' CPUs. */
1051 mycpu->gd_other_cpus = smp_startup_mask;
1052 ATOMIC_CPUMASK_NANDBIT(mycpu->gd_other_cpus, mycpu->gd_cpuid);
1054 /* A quick check from sanity claus */
1055 cpu_id = APICID_TO_CPUID((lapic->id & 0xff000000) >> 24);
1056 if (mycpu->gd_cpuid != cpu_id) {
1057 kprintf("SMP: assigned cpuid = %d\n", mycpu->gd_cpuid);
1058 kprintf("SMP: actual cpuid = %d lapicid %d\n",
1059 cpu_id, (lapic->id & 0xff000000) >> 24);
1061 kprintf("PTD[MPPTDI] = %p\n", (void *)PTD[MPPTDI]);
1063 panic("cpuid mismatch! boom!!");
1066 /* Initialize AP's local APIC for irq's */
1069 /* LAPIC initialization is done */
1070 ATOMIC_CPUMASK_ORBIT(smp_lapic_mask, mycpu->gd_cpuid);
1074 /* Let BSP move onto the next initialization stage */
1079 * Interlock for finalization. Wait until mp_finish is non-zero,
1080 * then get the MP lock.
1082 * Note: We are in a critical section.
1084 * Note: we are the idle thread, we can only spin.
1086 * Note: The load fence is memory volatile and prevents the compiler
1087 * from improperly caching mp_finish, and the cpu from improperly
1090 while (mp_finish == 0) {
1095 /* BSP may have changed PTD while we're waiting for the lock */
1098 /* Set memory range attributes for this CPU to match the BSP */
1099 mem_range_AP_init();
1102 * Once we go active we must process any IPIQ messages that may
1103 * have been queued, because no actual IPI will occur until we
1104 * set our bit in the smp_active_mask. If we don't the IPI
1105 * message interlock could be left set which would also prevent
1108 * The idle loop doesn't expect the BGL to be held and while
1109 * lwkt_switch() normally cleans things up this is a special case
1110 * because we returning almost directly into the idle loop.
1112 * The idle thread is never placed on the runq, make sure
1113 * nothing we've done put it there.
1117 * Hold a critical section and allow real interrupts to occur. Zero
1118 * any spurious interrupts which have accumulated, then set our
1119 * smp_active_mask indicating that we are fully operational.
1122 __asm __volatile("sti; pause; pause"::);
1123 bzero(mdcpu->gd_ipending, sizeof(mdcpu->gd_ipending));
1124 ATOMIC_CPUMASK_ORBIT(smp_active_mask, mycpu->gd_cpuid);
1127 * Wait until all cpus have set their smp_active_mask and have fully
1128 * operational interrupts before proceeding.
1130 * We need a final cpu_invltlb() because we would not have received
1131 * any until we set our bit in smp_active_mask.
1133 while (mp_finish == 1) {
1140 * Initialize per-cpu clocks and do other per-cpu initialization.
1141 * At this point code is expected to be able to use the full kernel
1144 initclocks_pcpu(); /* clock interrupts (via IPIs) */
1147 * Since we may have cleaned up the interrupt triggers, manually
1148 * process any pending IPIs before exiting our critical section.
1149 * Once the critical section has exited, normal interrupt processing
1152 lwkt_process_ipiq();
1153 crit_exit_noyield(mycpu->gd_curthread);
1156 * Final final, allow the waiting BSP to resume the boot process,
1157 * return 'into' the idle thread bootstrap.
1159 ATOMIC_CPUMASK_ORBIT(smp_finalize_mask, mycpu->gd_cpuid);
1160 KKASSERT((curthread->td_flags & TDF_RUNQ) == 0);
1164 * Get SMP fully working before we start initializing devices.
1171 kprintf("Finish MP startup\n");
1175 * Wait for the active mask to complete, after which all cpus will
1176 * be accepting interrupts.
1179 while (CPUMASK_CMPMASKNEQ(smp_active_mask, smp_startup_mask)) {
1185 * Wait for the finalization mask to complete, after which all cpus
1186 * have completely finished initializing and are entering or are in
1187 * their idle thread.
1189 * BSP should have received all required invltlbs but do another
1194 while (CPUMASK_CMPMASKNEQ(smp_finalize_mask, smp_startup_mask)) {
1199 while (try_mplock() == 0) {
1205 kprintf("Active CPU Mask: %016jx\n",
1206 (uintmax_t)CPUMASK_LOWMASK(smp_active_mask));
1210 SYSINIT(finishsmp, SI_BOOT2_FINISH_SMP, SI_ORDER_FIRST, ap_finish, NULL);
1213 cpu_send_ipiq(int dcpu)
1215 if (CPUMASK_TESTBIT(smp_active_mask, dcpu))
1216 single_apic_ipi(dcpu, XIPIQ_OFFSET, APIC_DELMODE_FIXED);
1219 #if 0 /* single_apic_ipi_passive() not working yet */
1221 * Returns 0 on failure, 1 on success
1224 cpu_send_ipiq_passive(int dcpu)
1227 if (CPUMASK_TESTBIT(smp_active_mask, dcpu)) {
1228 r = single_apic_ipi_passive(dcpu, XIPIQ_OFFSET,
1229 APIC_DELMODE_FIXED);
1236 mp_bsp_simple_setup(void)
1238 struct mdglobaldata *gd;
1241 /* build our map of 'other' CPUs */
1242 mycpu->gd_other_cpus = smp_startup_mask;
1243 CPUMASK_NANDBIT(mycpu->gd_other_cpus, mycpu->gd_cpuid);
1245 gd = (struct mdglobaldata *)mycpu;
1246 gd->gd_acpi_id = CPUID_TO_ACPIID(mycpu->gd_cpuid);
1248 ipiq_size = sizeof(struct lwkt_ipiq) * ncpus;
1249 mycpu->gd_ipiq = (void *)kmem_alloc(&kernel_map, ipiq_size);
1250 bzero(mycpu->gd_ipiq, ipiq_size);
1254 if (cpu_feature & CPUID_TSC)
1255 tsc0_offset = rdtsc();
1260 * CPU TOPOLOGY DETECTION FUNCTIONS
1263 /* Detect intel topology using CPUID
1264 * Ref: http://www.intel.com/Assets/PDF/appnote/241618.pdf, pg 41
1267 detect_intel_topology(int count_htt_cores)
1271 int core_plus_logical_bits = 0;
1272 int cores_per_package;
1273 int logical_per_package;
1274 int logical_per_core;
1277 if (cpu_high >= 0xb) {
1280 } else if (cpu_high >= 0x4) {
1285 for (shift = 0; (1 << shift) < count_htt_cores; ++shift)
1287 logical_CPU_bits = 1 << shift;
1292 cpuid_count(0xb, FUNC_B_THREAD_LEVEL, p);
1294 /* if 0xb not supported - fallback to 0x4 */
1295 if (p[1] == 0 || (FUNC_B_TYPE(p[2]) != FUNC_B_THREAD_TYPE)) {
1299 logical_CPU_bits = FUNC_B_BITS_SHIFT_NEXT_LEVEL(p[0]);
1301 ecx_index = FUNC_B_THREAD_LEVEL + 1;
1303 cpuid_count(0xb, ecx_index, p);
1305 /* Check for the Core type in the implemented sub leaves. */
1306 if (FUNC_B_TYPE(p[2]) == FUNC_B_CORE_TYPE) {
1307 core_plus_logical_bits = FUNC_B_BITS_SHIFT_NEXT_LEVEL(p[0]);
1313 } while (FUNC_B_TYPE(p[2]) != FUNC_B_INVALID_TYPE);
1315 core_bits = core_plus_logical_bits - logical_CPU_bits;
1320 cpuid_count(0x4, 0, p);
1321 cores_per_package = FUNC_4_MAX_CORE_NO(p[0]) + 1;
1323 logical_per_package = count_htt_cores;
1324 logical_per_core = logical_per_package / cores_per_package;
1326 for (shift = 0; (1 << shift) < logical_per_core; ++shift)
1328 logical_CPU_bits = shift;
1330 for (shift = 0; (1 << shift) < cores_per_package; ++shift)
1337 /* Detect AMD topology using CPUID
1338 * Ref: http://support.amd.com/us/Embedded_TechDocs/25481.pdf, last page
1341 detect_amd_topology(int count_htt_cores)
1344 if ((cpu_feature & CPUID_HTT)
1345 && (amd_feature2 & AMDID2_CMP)) {
1347 if (cpu_procinfo2 & AMDID_COREID_SIZE) {
1348 core_bits = (cpu_procinfo2 & AMDID_COREID_SIZE)
1349 >> AMDID_COREID_SIZE_SHIFT;
1351 core_bits = (cpu_procinfo2 & AMDID_CMP_CORES) + 1;
1352 for (shift = 0; (1 << shift) < core_bits; ++shift)
1357 logical_CPU_bits = count_htt_cores >> core_bits;
1358 for (shift = 0; (1 << shift) < logical_CPU_bits; ++shift)
1360 logical_CPU_bits = shift;
1362 for (shift = 0; (1 << shift) < count_htt_cores; ++shift)
1365 logical_CPU_bits = 0;
1370 amd_get_compute_unit_id(void *arg)
1374 do_cpuid(0x8000001e, regs);
1375 cpu_node_t * mynode = get_cpu_node_by_cpuid(mycpuid);
1377 * AMD - CPUID Specification September 2010
1378 * page 34 - //ComputeUnitID = ebx[0:7]//
1380 mynode->compute_unit_id = regs[1] & 0xff;
1384 fix_amd_topology(void)
1388 if (cpu_vendor_id != CPU_VENDOR_AMD)
1390 if ((amd_feature2 & AMDID2_TOPOEXT) == 0)
1393 CPUMASK_ASSALLONES(mask);
1394 lwkt_cpusync_simple(mask, amd_get_compute_unit_id, NULL);
1396 kprintf("Compute unit iDS:\n");
1398 for (i = 0; i < ncpus; i++) {
1399 kprintf("%d-%d; \n",
1400 i, get_cpu_node_by_cpuid(i)->compute_unit_id);
1407 * - logical_CPU_bits
1409 * With the values above (for AMD or INTEL) we are able to generally
1410 * detect the CPU topology (number of cores for each level):
1411 * Ref: http://wiki.osdev.org/Detecting_CPU_Topology_(80x86)
1412 * Ref: http://www.multicoreinfo.com/research/papers/whitepapers/Intel-detect-topology.pdf
1415 detect_cpu_topology(void)
1417 static int topology_detected = 0;
1420 if (topology_detected) {
1424 if ((cpu_feature & CPUID_HTT) == 0) {
1426 logical_CPU_bits = 0;
1429 count = (cpu_procinfo & CPUID_HTT_CORES)
1430 >> CPUID_HTT_CORE_SHIFT;
1433 if (cpu_vendor_id == CPU_VENDOR_INTEL) {
1434 detect_intel_topology(count);
1435 } else if (cpu_vendor_id == CPU_VENDOR_AMD) {
1436 detect_amd_topology(count);
1441 kprintf("Bits within APICID: logical_CPU_bits: %d; core_bits: %d\n",
1442 logical_CPU_bits, core_bits);
1444 topology_detected = 1;
1447 /* Interface functions to calculate chip_ID,
1448 * core_number and logical_number
1449 * Ref: http://wiki.osdev.org/Detecting_CPU_Topology_(80x86)
1452 get_chip_ID(int cpuid)
1454 return get_apicid_from_cpuid(cpuid) >>
1455 (logical_CPU_bits + core_bits);
1459 get_core_number_within_chip(int cpuid)
1461 return (get_apicid_from_cpuid(cpuid) >> logical_CPU_bits) &
1462 ( (1 << core_bits) -1);
1466 get_logical_CPU_number_within_core(int cpuid)
1468 return get_apicid_from_cpuid(cpuid) &
1469 ( (1 << logical_CPU_bits) -1);