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>
68 #include <machine/clock.h>
70 #include <machine/md_var.h> /* setidt() */
71 #include <machine_base/icu/icu.h> /* IPIs */
72 #include <machine_base/icu/icu_var.h>
73 #include <machine_base/apic/ioapic_abi.h>
74 #include <machine/intr_machdep.h> /* IPIs */
76 #define WARMBOOT_TARGET 0
77 #define WARMBOOT_OFF (KERNBASE + 0x0467)
78 #define WARMBOOT_SEG (KERNBASE + 0x0469)
80 #define CMOS_REG (0x70)
81 #define CMOS_DATA (0x71)
82 #define BIOS_RESET (0x0f)
83 #define BIOS_WARM (0x0a)
86 * this code MUST be enabled here and in mpboot.s.
87 * it follows the very early stages of AP boot by placing values in CMOS ram.
88 * it NORMALLY will never be needed and thus the primitive method for enabling.
91 #if defined(CHECK_POINTS)
92 #define CHECK_READ(A) (outb(CMOS_REG, (A)), inb(CMOS_DATA))
93 #define CHECK_WRITE(A,D) (outb(CMOS_REG, (A)), outb(CMOS_DATA, (D)))
95 #define CHECK_INIT(D); \
96 CHECK_WRITE(0x34, (D)); \
97 CHECK_WRITE(0x35, (D)); \
98 CHECK_WRITE(0x36, (D)); \
99 CHECK_WRITE(0x37, (D)); \
100 CHECK_WRITE(0x38, (D)); \
101 CHECK_WRITE(0x39, (D));
103 #define CHECK_PRINT(S); \
104 kprintf("%s: %d, %d, %d, %d, %d, %d\n", \
113 #else /* CHECK_POINTS */
115 #define CHECK_INIT(D)
116 #define CHECK_PRINT(S)
118 #endif /* CHECK_POINTS */
121 * Values to send to the POST hardware.
123 #define MP_BOOTADDRESS_POST 0x10
124 #define MP_PROBE_POST 0x11
125 #define MPTABLE_PASS1_POST 0x12
127 #define MP_START_POST 0x13
128 #define MP_ENABLE_POST 0x14
129 #define MPTABLE_PASS2_POST 0x15
131 #define START_ALL_APS_POST 0x16
132 #define INSTALL_AP_TRAMP_POST 0x17
133 #define START_AP_POST 0x18
135 #define MP_ANNOUNCE_POST 0x19
137 /** XXX FIXME: where does this really belong, isa.h/isa.c perhaps? */
138 int current_postcode;
140 /** XXX FIXME: what system files declare these??? */
141 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_OPAQUE(_machdep, OID_AUTO, smp_active, CTLFLAG_RD,
181 &smp_active_mask, sizeof(smp_active_mask), "LU", "");
182 static u_int bootMP_size;
183 static u_int report_invlpg_src;
184 SYSCTL_INT(_machdep, OID_AUTO, report_invlpg_src, CTLFLAG_RW,
185 &report_invlpg_src, 0, "");
186 static u_int report_invltlb_src;
187 SYSCTL_INT(_machdep, OID_AUTO, report_invltlb_src, CTLFLAG_RW,
188 &report_invltlb_src, 0, "");
189 static int optimized_invltlb;
190 SYSCTL_INT(_machdep, OID_AUTO, optimized_invltlb, CTLFLAG_RW,
191 &optimized_invltlb, 0, "");
192 static int all_but_self_ipi_enable = 1;
193 SYSCTL_INT(_machdep, OID_AUTO, all_but_self_ipi_enable, CTLFLAG_RW,
194 &all_but_self_ipi_enable, 0, "");
196 /* Local data for detecting CPU TOPOLOGY */
197 static int core_bits = 0;
198 static int logical_CPU_bits = 0;
202 * Calculate usable address in base memory for AP trampoline code.
205 mp_bootaddress(u_int basemem)
207 POSTCODE(MP_BOOTADDRESS_POST);
209 bootMP_size = mptramp_end - mptramp_start;
210 boot_address = trunc_page(basemem * 1024); /* round down to 4k boundary */
211 if (((basemem * 1024) - boot_address) < bootMP_size)
212 boot_address -= PAGE_SIZE; /* not enough, lower by 4k */
213 /* 3 levels of page table pages */
214 mptramp_pagetables = boot_address - (PAGE_SIZE * 3);
216 return mptramp_pagetables;
220 * Print various information about the SMP system hardware and setup.
227 POSTCODE(MP_ANNOUNCE_POST);
229 kprintf("DragonFly/MP: Multiprocessor motherboard\n");
230 kprintf(" cpu0 (BSP): apic id: %2d\n", CPUID_TO_APICID(0));
231 for (x = 1; x <= naps; ++x)
232 kprintf(" cpu%d (AP): apic id: %2d\n", x, CPUID_TO_APICID(x));
235 kprintf(" Warning: APIC I/O disabled\n");
239 * AP cpu's call this to sync up protected mode.
241 * WARNING! %gs is not set up on entry. This routine sets up %gs.
247 int x, myid = bootAP;
249 struct mdglobaldata *md;
250 struct privatespace *ps;
252 ps = CPU_prvspace[myid];
254 gdt_segs[GPROC0_SEL].ssd_base =
255 (long) &ps->mdglobaldata.gd_common_tss;
256 ps->mdglobaldata.mi.gd_prvspace = ps;
258 /* We fill the 32-bit segment descriptors */
259 for (x = 0; x < NGDT; x++) {
260 if (x != GPROC0_SEL && x != (GPROC0_SEL + 1))
261 ssdtosd(&gdt_segs[x], &gdt[myid * NGDT + x]);
263 /* And now a 64-bit one */
264 ssdtosyssd(&gdt_segs[GPROC0_SEL],
265 (struct system_segment_descriptor *)&gdt[myid * NGDT + GPROC0_SEL]);
267 r_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
268 r_gdt.rd_base = (long) &gdt[myid * NGDT];
269 lgdt(&r_gdt); /* does magic intra-segment return */
271 /* lgdt() destroys the GSBASE value, so we load GSBASE after lgdt() */
272 wrmsr(MSR_FSBASE, 0); /* User value */
273 wrmsr(MSR_GSBASE, (u_int64_t)ps);
274 wrmsr(MSR_KGSBASE, 0); /* XXX User value while we're in the kernel */
276 lidt(&r_idt_arr[mdcpu->mi.gd_cpuid]);
280 mdcpu->gd_currentldt = _default_ldt;
283 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
284 gdt[myid * NGDT + GPROC0_SEL].sd_type = SDT_SYSTSS;
286 md = mdcpu; /* loaded through %gs:0 (mdglobaldata.mi.gd_prvspace)*/
289 * TSS entry point for interrupts, traps, and exceptions
290 * (sans NMI). This will always go to near the top of the pcpu
291 * trampoline area. Hardware-pushed data will be copied into
292 * the trap-frame on entry, and (if necessary) returned to the
293 * trampoline on exit.
295 * We store some pcb data for the trampoline code above the
296 * stack the cpu hw pushes into, and arrange things so the
297 * address of tr_pcb_rsp is the same as the desired top of
300 md->gd_common_tss.tss_rsp0 =
301 (register_t)&((struct privatespace *)md)->trampoline.tr_pcb_rsp;
302 ((struct privatespace *)md)->trampoline.tr_pcb_rsp =
303 md->gd_common_tss.tss_rsp0;
306 md->gd_common_tss.tss_ioopt = (sizeof md->gd_common_tss) << 16;
308 md->gd_tss_gdt = &gdt[myid * NGDT + GPROC0_SEL];
309 md->gd_common_tssd = *md->gd_tss_gdt;
311 /* double fault stack */
312 md->gd_common_tss.tss_ist1 =
313 (long)&md->mi.gd_prvspace->idlestack[
314 sizeof(md->mi.gd_prvspace->idlestack)];
319 * Set to a known state:
320 * Set by mpboot.s: CR0_PG, CR0_PE
321 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
324 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
327 /* Set up the fast syscall stuff */
328 msr = rdmsr(MSR_EFER) | EFER_SCE;
329 wrmsr(MSR_EFER, msr);
330 wrmsr(MSR_LSTAR, (u_int64_t)IDTVEC(fast_syscall));
331 wrmsr(MSR_CSTAR, (u_int64_t)IDTVEC(fast_syscall32));
332 msr = ((u_int64_t)GSEL(GCODE_SEL, SEL_KPL) << 32) |
333 ((u_int64_t)GSEL(GUCODE32_SEL, SEL_UPL) << 48);
334 wrmsr(MSR_STAR, msr);
335 wrmsr(MSR_SF_MASK, PSL_NT|PSL_T|PSL_I|PSL_C|PSL_D|PSL_IOPL);
337 pmap_set_opt(); /* PSE/4MB pages, etc */
338 pmap_init_pat(); /* Page Attribute Table */
340 /* set up CPU registers and state */
343 /* set up SSE/NX registers */
346 /* set up FPU state on the AP */
349 /* disable the APIC, just to be SURE */
350 lapic->svr &= ~APIC_SVR_ENABLE;
353 /*******************************************************************
354 * local functions and data
358 * Start the SMP system
361 mp_start_aps(void *dummy __unused)
364 /* start each Application Processor */
365 start_all_aps(boot_address);
367 mp_bsp_simple_setup();
370 SYSINIT(startaps, SI_BOOT2_START_APS, SI_ORDER_FIRST, mp_start_aps, NULL);
373 * start each AP in our list
376 start_all_aps(u_int boot_addr)
378 vm_offset_t va = boot_address + KERNBASE;
379 u_int64_t *pt4, *pt3, *pt2;
387 u_long mpbioswarmvec;
388 struct mdglobaldata *gd;
389 struct privatespace *ps;
392 POSTCODE(START_ALL_APS_POST);
394 /* install the AP 1st level boot code */
395 pmap_kenter(va, boot_address);
396 cpu_invlpg((void *)va); /* JG XXX */
397 bcopy(mptramp_start, (void *)va, bootMP_size);
399 /* Locate the page tables, they'll be below the trampoline */
400 pt4 = (u_int64_t *)(uintptr_t)(mptramp_pagetables + KERNBASE);
401 pt3 = pt4 + (PAGE_SIZE) / sizeof(u_int64_t);
402 pt2 = pt3 + (PAGE_SIZE) / sizeof(u_int64_t);
404 /* Create the initial 1GB replicated page tables */
405 for (i = 0; i < 512; i++) {
406 /* Each slot of the level 4 pages points to the same level 3 page */
407 pt4[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + PAGE_SIZE);
408 pt4[i] |= kernel_pmap.pmap_bits[PG_V_IDX] |
409 kernel_pmap.pmap_bits[PG_RW_IDX] |
410 kernel_pmap.pmap_bits[PG_U_IDX];
412 /* Each slot of the level 3 pages points to the same level 2 page */
413 pt3[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + (2 * PAGE_SIZE));
414 pt3[i] |= kernel_pmap.pmap_bits[PG_V_IDX] |
415 kernel_pmap.pmap_bits[PG_RW_IDX] |
416 kernel_pmap.pmap_bits[PG_U_IDX];
418 /* The level 2 page slots are mapped with 2MB pages for 1GB. */
419 pt2[i] = i * (2 * 1024 * 1024);
420 pt2[i] |= kernel_pmap.pmap_bits[PG_V_IDX] |
421 kernel_pmap.pmap_bits[PG_RW_IDX] |
422 kernel_pmap.pmap_bits[PG_PS_IDX] |
423 kernel_pmap.pmap_bits[PG_U_IDX];
426 /* save the current value of the warm-start vector */
427 mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF);
428 outb(CMOS_REG, BIOS_RESET);
429 mpbiosreason = inb(CMOS_DATA);
431 /* setup a vector to our boot code */
432 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
433 *((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4);
434 outb(CMOS_REG, BIOS_RESET);
435 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
438 * If we have a TSC we can figure out the SMI interrupt rate.
439 * The SMI does not necessarily use a constant rate. Spend
440 * up to 250ms trying to figure it out.
443 if (cpu_feature & CPUID_TSC) {
444 set_apic_timer(275000);
445 smilast = read_apic_timer();
446 for (x = 0; x < 20 && read_apic_timer(); ++x) {
447 smicount = smitest();
448 if (smibest == 0 || smilast - smicount < smibest)
449 smibest = smilast - smicount;
452 if (smibest > 250000)
456 kprintf("SMI Frequency (worst case): %d Hz (%d us)\n",
457 1000000 / smibest, smibest);
460 for (x = 1; x <= naps; ++x) {
461 /* This is a bit verbose, it will go away soon. */
463 pssize = sizeof(struct privatespace);
464 ps = (void *)kmem_alloc3(&kernel_map, pssize, VM_SUBSYS_GD,
466 CPU_prvspace[x] = ps;
468 kprintf("ps %d %p %d\n", x, ps, pssize);
471 gd = &ps->mdglobaldata;
472 gd->mi.gd_prvspace = ps;
474 /* prime data page for it to use */
475 mi_gdinit(&gd->mi, x);
477 ipiq_size = sizeof(struct lwkt_ipiq) * (naps + 1);
478 gd->mi.gd_ipiq = (void *)kmem_alloc3(&kernel_map, ipiq_size,
479 VM_SUBSYS_IPIQ, KM_CPU(x));
480 bzero(gd->mi.gd_ipiq, ipiq_size);
482 gd->gd_acpi_id = CPUID_TO_ACPIID(gd->mi.gd_cpuid);
484 /* initialize arc4random. */
487 /* setup a vector to our boot code */
488 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
489 *((volatile u_short *) WARMBOOT_SEG) = (boot_addr >> 4);
490 outb(CMOS_REG, BIOS_RESET);
491 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
494 * Setup the AP boot stack
496 bootSTK = &ps->idlestack[UPAGES * PAGE_SIZE - PAGE_SIZE];
499 /* attempt to start the Application Processor */
500 CHECK_INIT(99); /* setup checkpoints */
501 if (!start_ap(gd, boot_addr, smibest)) {
502 kprintf("\nAP #%d (PHY# %d) failed!\n",
503 x, CPUID_TO_APICID(x));
504 CHECK_PRINT("trace"); /* show checkpoints */
505 /* better panic as the AP may be running loose */
506 kprintf("panic y/n? [y] ");
512 CHECK_PRINT("trace"); /* show checkpoints */
515 /* set ncpus to 1 + highest logical cpu. Not all may have come up */
518 for (shift = 0; (1 << shift) <= ncpus; ++shift)
522 /* ncpus_fit -- ncpus rounded up to the nearest power of 2 */
523 if ((1 << shift) < ncpus)
525 ncpus_fit = 1 << shift;
526 ncpus_fit_mask = ncpus_fit - 1;
528 /* build our map of 'other' CPUs */
529 mycpu->gd_other_cpus = smp_startup_mask;
530 CPUMASK_NANDBIT(mycpu->gd_other_cpus, mycpu->gd_cpuid);
532 gd = (struct mdglobaldata *)mycpu;
533 gd->gd_acpi_id = CPUID_TO_ACPIID(mycpu->gd_cpuid);
535 ipiq_size = sizeof(struct lwkt_ipiq) * ncpus;
536 mycpu->gd_ipiq = (void *)kmem_alloc3(&kernel_map, ipiq_size,
537 VM_SUBSYS_IPIQ, KM_CPU(0));
538 bzero(mycpu->gd_ipiq, ipiq_size);
540 /* initialize arc4random. */
543 /* restore the warmstart vector */
544 *(u_long *) WARMBOOT_OFF = mpbioswarmvec;
545 outb(CMOS_REG, BIOS_RESET);
546 outb(CMOS_DATA, mpbiosreason);
549 * NOTE! The idlestack for the BSP was setup by locore. Finish
550 * up, clean out the P==V mapping we did earlier.
555 * Wait all APs to finish initializing LAPIC
558 kprintf("SMP: Waiting APs LAPIC initialization\n");
559 if (cpu_feature & CPUID_TSC)
560 tsc0_offset = rdtsc();
565 while (CPUMASK_CMPMASKNEQ(smp_lapic_mask, smp_startup_mask)) {
568 if (cpu_feature & CPUID_TSC)
569 tsc0_offset = rdtsc();
571 while (try_mplock() == 0) {
576 /* number of APs actually started */
582 * load the 1st level AP boot code into base memory.
585 /* targets for relocation */
586 extern void bigJump(void);
587 extern void bootCodeSeg(void);
588 extern void bootDataSeg(void);
589 extern void MPentry(void);
591 extern u_int mp_gdtbase;
596 install_ap_tramp(u_int boot_addr)
599 int size = *(int *) ((u_long) & bootMP_size);
600 u_char *src = (u_char *) ((u_long) bootMP);
601 u_char *dst = (u_char *) boot_addr + KERNBASE;
602 u_int boot_base = (u_int) bootMP;
607 POSTCODE(INSTALL_AP_TRAMP_POST);
609 for (x = 0; x < size; ++x)
613 * modify addresses in code we just moved to basemem. unfortunately we
614 * need fairly detailed info about mpboot.s for this to work. changes
615 * to mpboot.s might require changes here.
618 /* boot code is located in KERNEL space */
619 dst = (u_char *) boot_addr + KERNBASE;
621 /* modify the lgdt arg */
622 dst32 = (u_int32_t *) (dst + ((u_int) & mp_gdtbase - boot_base));
623 *dst32 = boot_addr + ((u_int) & MP_GDT - boot_base);
625 /* modify the ljmp target for MPentry() */
626 dst32 = (u_int32_t *) (dst + ((u_int) bigJump - boot_base) + 1);
627 *dst32 = ((u_int) MPentry - KERNBASE);
629 /* modify the target for boot code segment */
630 dst16 = (u_int16_t *) (dst + ((u_int) bootCodeSeg - boot_base));
631 dst8 = (u_int8_t *) (dst16 + 1);
632 *dst16 = (u_int) boot_addr & 0xffff;
633 *dst8 = ((u_int) boot_addr >> 16) & 0xff;
635 /* modify the target for boot data segment */
636 dst16 = (u_int16_t *) (dst + ((u_int) bootDataSeg - boot_base));
637 dst8 = (u_int8_t *) (dst16 + 1);
638 *dst16 = (u_int) boot_addr & 0xffff;
639 *dst8 = ((u_int) boot_addr >> 16) & 0xff;
645 * This function starts the AP (application processor) identified
646 * by the APIC ID 'physicalCpu'. It does quite a "song and dance"
647 * to accomplish this. This is necessary because of the nuances
648 * of the different hardware we might encounter. It ain't pretty,
649 * but it seems to work.
651 * NOTE: eventually an AP gets to ap_init(), which is called just
652 * before the AP goes into the LWKT scheduler's idle loop.
655 start_ap(struct mdglobaldata *gd, u_int boot_addr, int smibest)
659 u_long icr_lo, icr_hi;
661 POSTCODE(START_AP_POST);
663 /* get the PHYSICAL APIC ID# */
664 physical_cpu = CPUID_TO_APICID(gd->mi.gd_cpuid);
666 /* calculate the vector */
667 vector = (boot_addr >> 12) & 0xff;
669 /* We don't want anything interfering */
672 /* Make sure the target cpu sees everything */
676 * Try to detect when a SMI has occurred, wait up to 200ms.
678 * If a SMI occurs during an AP reset but before we issue
679 * the STARTUP command, the AP may brick. To work around
680 * this problem we hold off doing the AP startup until
681 * after we have detected the SMI. Hopefully another SMI
682 * will not occur before we finish the AP startup.
684 * Retries don't seem to help. SMIs have a window of opportunity
685 * and if USB->legacy keyboard emulation is enabled in the BIOS
686 * the interrupt rate can be quite high.
688 * NOTE: Don't worry about the L1 cache load, it might bloat
689 * ldelta a little but ndelta will be so huge when the SMI
690 * occurs the detection logic will still work fine.
693 set_apic_timer(200000);
698 * first we do an INIT/RESET IPI this INIT IPI might be run, reseting
699 * and running the target CPU. OR this INIT IPI might be latched (P5
700 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
703 * see apic/apicreg.h for icr bit definitions.
705 * TIME CRITICAL CODE, DO NOT DO ANY KPRINTFS IN THE HOT PATH.
709 * Setup the address for the target AP. We can setup
710 * icr_hi once and then just trigger operations with
713 icr_hi = lapic->icr_hi & ~APIC_ID_MASK;
714 icr_hi |= (physical_cpu << 24);
715 icr_lo = lapic->icr_lo & 0xfff00000;
716 lapic->icr_hi = icr_hi;
719 * Do an INIT IPI: assert RESET
721 * Use edge triggered mode to assert INIT
723 lapic->icr_lo = icr_lo | 0x00004500;
724 while (lapic->icr_lo & APIC_DELSTAT_MASK)
728 * The spec calls for a 10ms delay but we may have to use a
729 * MUCH lower delay to avoid bricking an AP due to a fast SMI
730 * interrupt. We have other loops here too and dividing by 2
731 * doesn't seem to be enough even after subtracting 350us,
734 * Our minimum delay is 150uS, maximum is 10ms. If no SMI
735 * interrupt was detected we use the full 10ms.
739 else if (smibest < 150 * 4 + 350)
741 else if ((smibest - 350) / 4 < 10000)
742 u_sleep((smibest - 350) / 4);
747 * Do an INIT IPI: deassert RESET
749 * Use level triggered mode to deassert. It is unclear
750 * why we need to do this.
752 lapic->icr_lo = icr_lo | 0x00008500;
753 while (lapic->icr_lo & APIC_DELSTAT_MASK)
755 u_sleep(150); /* wait 150us */
758 * Next we do a STARTUP IPI: the previous INIT IPI might still be
759 * latched, (P5 bug) this 1st STARTUP would then terminate
760 * immediately, and the previously started INIT IPI would continue. OR
761 * the previous INIT IPI has already run. and this STARTUP IPI will
762 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
765 lapic->icr_lo = icr_lo | 0x00000600 | vector;
766 while (lapic->icr_lo & APIC_DELSTAT_MASK)
768 u_sleep(200); /* wait ~200uS */
771 * Finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
772 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
773 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
774 * recognized after hardware RESET or INIT IPI.
776 lapic->icr_lo = icr_lo | 0x00000600 | vector;
777 while (lapic->icr_lo & APIC_DELSTAT_MASK)
780 /* Resume normal operation */
783 /* wait for it to start, see ap_init() */
784 set_apic_timer(5000000);/* == 5 seconds */
785 while (read_apic_timer()) {
786 if (CPUMASK_TESTBIT(smp_startup_mask, gd->mi.gd_cpuid))
787 return 1; /* return SUCCESS */
790 return 0; /* return FAILURE */
805 while (read_apic_timer()) {
807 for (count = 0; count < 100; ++count)
808 ntsc = rdtsc(); /* force loop to occur */
810 ndelta = ntsc - ltsc;
813 if (ndelta > ldelta * 2)
816 ldelta = ntsc - ltsc;
819 return(read_apic_timer());
823 * Synchronously flush the TLB on all other CPU's. The current cpu's
824 * TLB is not flushed. If the caller wishes to flush the current cpu's
825 * TLB the caller must call cpu_invltlb() in addition to smp_invltlb().
827 * This routine may be called concurrently from multiple cpus. When this
828 * happens, smp_invltlb() can wind up sticking around in the confirmation
829 * while() loop at the end as additional cpus are added to the global
830 * cpumask, until they are acknowledged by another IPI.
832 * NOTE: If for some reason we were unable to start all cpus we cannot
833 * safely use broadcast IPIs.
836 cpumask_t smp_smurf_mask;
837 static cpumask_t smp_invltlb_mask;
841 cpumask_t smp_in_mask;
843 cpumask_t smp_invmask;
844 extern cpumask_t smp_idleinvl_mask;
845 extern cpumask_t smp_idleinvl_reqs;
848 * Atomically OR bits in *mask to smp_smurf_mask. Adjust *mask to remove
849 * bits that do not need to be IPId. These bits are still part of the command,
850 * but the target cpus have already been signalled and do not need to be
853 #include <sys/spinlock.h>
854 #include <sys/spinlock2.h>
858 smp_smurf_fetchset(cpumask_t *mask)
866 while (i < CPUMASK_ELEMENTS) {
867 obits = smp_smurf_mask.ary[i];
869 nbits = obits | mask->ary[i];
870 if (atomic_cmpset_long(&smp_smurf_mask.ary[i], obits, nbits)) {
871 omask.ary[i] = obits;
875 CPUMASK_NANDMASK(*mask, omask);
879 * This is a mechanism which guarantees that cpu_invltlb() will be executed
880 * on idle cpus without having to signal or wake them up. The invltlb will be
881 * executed when they wake up, prior to any scheduling or interrupt thread.
883 * (*mask) is modified to remove the cpus we successfully negotiate this
884 * function with. This function may only be used with semi-synchronous
885 * commands (typically invltlb's or semi-synchronous invalidations which
886 * are usually associated only with kernel memory).
889 smp_smurf_idleinvlclr(cpumask_t *mask)
891 if (optimized_invltlb) {
892 ATOMIC_CPUMASK_ORMASK(smp_idleinvl_reqs, *mask);
893 /* cpu_lfence() not needed */
894 CPUMASK_NANDMASK(*mask, smp_idleinvl_mask);
899 * Issue cpu_invltlb() across all cpus except the current cpu.
901 * This function will arrange to avoid idle cpus, but still gurantee that
902 * invltlb is run on them when they wake up prior to any scheduling or
908 struct mdglobaldata *md = mdcpu;
910 unsigned long rflags;
912 tsc_uclock_t tsc_base = rdtsc();
916 if (report_invltlb_src > 0) {
917 if (--report_invltlb_src <= 0)
922 * Disallow normal interrupts, set all active cpus except our own
923 * in the global smp_invltlb_mask.
925 ++md->mi.gd_cnt.v_smpinvltlb;
926 crit_enter_gd(&md->mi);
929 * Bits we want to set in smp_invltlb_mask. We do not want to signal
930 * our own cpu. Also try to remove bits associated with idle cpus
931 * that we can flag for auto-invltlb.
933 mask = smp_active_mask;
934 CPUMASK_NANDBIT(mask, md->mi.gd_cpuid);
935 smp_smurf_idleinvlclr(&mask);
937 rflags = read_rflags();
939 ATOMIC_CPUMASK_ORMASK(smp_invltlb_mask, mask);
942 * IPI non-idle cpus represented by mask. The omask calculation
943 * removes cpus from the mask which already have a Xinvltlb IPI
944 * pending (avoid double-queueing the IPI).
946 * We must disable real interrupts when setting the smurf flags or
947 * we might race a XINVLTLB before we manage to send the ipi's for
950 * NOTE: We are not signalling ourselves, mask already does NOT
951 * include our own cpu.
953 smp_smurf_fetchset(&mask);
956 * Issue the IPI. Note that the XINVLTLB IPI runs regardless of
957 * the critical section count on the target cpus.
959 CPUMASK_ORMASK(mask, md->mi.gd_cpumask);
960 if (all_but_self_ipi_enable &&
961 (all_but_self_ipi_enable >= 2 ||
962 CPUMASK_CMPMASKEQ(smp_startup_mask, mask))) {
963 all_but_self_ipi(XINVLTLB_OFFSET);
965 CPUMASK_NANDMASK(mask, md->mi.gd_cpumask);
966 selected_apic_ipi(mask, XINVLTLB_OFFSET, APIC_DELMODE_FIXED);
970 * Wait for acknowledgement by all cpus. smp_inval_intr() will
971 * temporarily enable interrupts to avoid deadlocking the lapic,
972 * and will also handle running cpu_invltlb() and remote invlpg
973 * command son our cpu if some other cpu requests it of us.
975 * WARNING! I originally tried to implement this as a hard loop
976 * checking only smp_invltlb_mask (and issuing a local
977 * cpu_invltlb() if requested), with interrupts enabled
978 * and without calling smp_inval_intr(). This DID NOT WORK.
979 * It resulted in weird races where smurf bits would get
980 * cleared without any action being taken.
983 CPUMASK_ASSZERO(mask);
984 while (CPUMASK_CMPMASKNEQ(smp_invltlb_mask, mask)) {
988 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
990 * cpuid - cpu doing the waiting
991 * invltlb_mask - IPI in progress
993 kprintf("smp_invltlb %d: waited too long inv=%08jx "
998 "idle=%08jx/%08jx\n",
1000 smp_invltlb_mask.ary[0],
1001 smp_smurf_mask.ary[0],
1005 smp_idleinvl_mask.ary[0],
1006 smp_idleinvl_reqs.ary[0]);
1007 mdcpu->gd_xinvaltlb = 0;
1008 ATOMIC_CPUMASK_NANDMASK(smp_smurf_mask,
1010 smp_invlpg(&smp_active_mask);
1012 if (++repeats > 10) {
1013 kprintf("smp_invltlb: giving up\n");
1014 CPUMASK_ASSZERO(smp_invltlb_mask);
1019 write_rflags(rflags);
1020 crit_exit_gd(&md->mi);
1024 * Called from a critical section with interrupts hard-disabled.
1025 * This function issues an XINVLTLB IPI and then executes any pending
1026 * command on the current cpu before returning.
1029 smp_invlpg(cpumask_t *cmdmask)
1031 struct mdglobaldata *md = mdcpu;
1034 if (report_invlpg_src > 0) {
1035 if (--report_invlpg_src <= 0)
1040 * Disallow normal interrupts, set all active cpus in the pmap,
1041 * plus our own for completion processing (it might or might not
1042 * be part of the set).
1044 mask = smp_active_mask;
1045 CPUMASK_ANDMASK(mask, *cmdmask);
1046 CPUMASK_ORMASK(mask, md->mi.gd_cpumask);
1049 * Avoid double-queuing IPIs, which can deadlock us. We must disable
1050 * real interrupts when setting the smurf flags or we might race a
1051 * XINVLTLB before we manage to send the ipi's for the bits we set.
1053 * NOTE: We might be including our own cpu in the smurf mask.
1055 smp_smurf_fetchset(&mask);
1058 * Issue the IPI. Note that the XINVLTLB IPI runs regardless of
1059 * the critical section count on the target cpus.
1061 * We do not include our own cpu when issuing the IPI.
1063 if (all_but_self_ipi_enable &&
1064 (all_but_self_ipi_enable >= 2 ||
1065 CPUMASK_CMPMASKEQ(smp_startup_mask, mask))) {
1066 all_but_self_ipi(XINVLTLB_OFFSET);
1068 CPUMASK_NANDMASK(mask, md->mi.gd_cpumask);
1069 selected_apic_ipi(mask, XINVLTLB_OFFSET, APIC_DELMODE_FIXED);
1073 * This will synchronously wait for our command to complete,
1074 * as well as process commands from other cpus. It also handles
1077 * (interrupts are disabled and we are in a critical section here)
1085 globaldata_t gd = mycpu;
1090 * Ignore all_but_self_ipi_enable here and just use it.
1092 rflags = read_rflags();
1094 all_but_self_ipi(XSNIFF_OFFSET);
1095 gd->gd_sample_pc = smp_sniff;
1096 gd->gd_sample_sp = &dummy;
1097 write_rflags(rflags);
1103 globaldata_t rgd = globaldata_find(dcpu);
1108 * Ignore all_but_self_ipi_enable here and just use it.
1110 rflags = read_rflags();
1112 single_apic_ipi(dcpu, XSNIFF_OFFSET, APIC_DELMODE_FIXED);
1113 rgd->gd_sample_pc = cpu_sniff;
1114 rgd->gd_sample_sp = &dummy;
1115 write_rflags(rflags);
1119 * Called from Xinvltlb assembly with interrupts hard-disabled and in a
1120 * critical section. gd_intr_nesting_level may or may not be bumped
1121 * depending on entry.
1123 * THIS CODE IS INTENDED TO EXPLICITLY IGNORE THE CRITICAL SECTION COUNT.
1124 * THAT IS, THE INTERRUPT IS INTENDED TO FUNCTION EVEN WHEN MAINLINE CODE
1125 * IS IN A CRITICAL SECTION.
1128 smp_inval_intr(void)
1130 struct mdglobaldata *md = mdcpu;
1133 tsc_uclock_t tsc_base = rdtsc();
1138 * The idle code is in a critical section, but that doesn't stop
1139 * Xinvltlb from executing, so deal with the race which can occur
1140 * in that situation. Otherwise r-m-w operations by pmap_inval_intr()
1141 * may have problems.
1143 if (ATOMIC_CPUMASK_TESTANDCLR(smp_idleinvl_reqs, md->mi.gd_cpuid)) {
1144 ATOMIC_CPUMASK_NANDBIT(smp_invltlb_mask, md->mi.gd_cpuid);
1151 * This is a real mess. I'd like to just leave interrupts disabled
1152 * but it can cause the lapic to deadlock if too many interrupts queue
1153 * to it, due to the idiotic design of the lapic. So instead we have
1154 * to enter a critical section so normal interrupts are made pending
1155 * and track whether this one was reentered.
1157 if (md->gd_xinvaltlb) { /* reentrant on cpu */
1158 md->gd_xinvaltlb = 2;
1161 md->gd_xinvaltlb = 1;
1164 * Check only those cpus with active Xinvl* commands pending.
1166 * We are going to enable interrupts so make sure we are in a
1167 * critical section. This is necessary to avoid deadlocking
1168 * the lapic and to ensure that we execute our commands prior to
1169 * any nominal interrupt or preemption.
1171 * WARNING! It is very important that we only clear out but in
1172 * smp_smurf_mask once for each interrupt we take. In
1173 * this case, we clear it on initial entry and only loop
1174 * on the reentrancy detect (caused by another interrupt).
1176 cpumask = smp_invmask;
1178 ATOMIC_CPUMASK_ORBIT(smp_in_mask, md->mi.gd_cpuid);
1182 ATOMIC_CPUMASK_NANDBIT(smp_smurf_mask, md->mi.gd_cpuid);
1185 * Specific page request(s), and we can't return until all bits
1192 * Also execute any pending full invalidation request in
1195 if (CPUMASK_TESTBIT(smp_invltlb_mask, md->mi.gd_cpuid)) {
1196 ATOMIC_CPUMASK_NANDBIT(smp_invltlb_mask,
1203 if (tsc_frequency && rdtsc() - tsc_base > tsc_frequency) {
1205 * cpuid - cpu doing the waiting
1206 * invmask - IPI in progress
1207 * invltlb_mask - which ones are TLB invalidations?
1209 kprintf("smp_inval_intr %d inv=%08jx tlbm=%08jx "
1214 "idle=%08jx/%08jx\n",
1217 smp_invltlb_mask.ary[0],
1218 smp_smurf_mask.ary[0],
1222 smp_idleinvl_mask.ary[0],
1223 smp_idleinvl_reqs.ary[0]);
1234 * We can only add bits to the cpumask to test during the
1235 * loop because the smp_invmask bit is cleared once the
1236 * originator completes the command (the targets may still
1237 * be cycling their own completions in this loop, afterwords).
1239 * lfence required prior to all tests as this Xinvltlb
1240 * interrupt could race the originator (already be in progress
1241 * wnen the originator decides to issue, due to an issue by
1245 CPUMASK_ORMASK(cpumask, smp_invmask);
1246 /*cpumask = smp_active_mask;*/ /* XXX */
1249 if (pmap_inval_intr(&cpumask, toolong) == 0) {
1251 * Clear our smurf mask to allow new IPIs, but deal
1252 * with potential races.
1258 * Test if someone sent us another invalidation IPI, break
1259 * out so we can take it to avoid deadlocking the lapic
1260 * interrupt queue (? stupid intel, amd).
1262 if (md->gd_xinvaltlb == 2)
1265 if (CPUMASK_TESTBIT(smp_smurf_mask, md->mi.gd_cpuid))
1271 * Full invalidation request
1273 if (CPUMASK_TESTBIT(smp_invltlb_mask, md->mi.gd_cpuid)) {
1274 ATOMIC_CPUMASK_NANDBIT(smp_invltlb_mask,
1281 * Check to see if another Xinvltlb interrupt occurred and loop up
1285 if (md->gd_xinvaltlb == 2) {
1286 md->gd_xinvaltlb = 1;
1290 ATOMIC_CPUMASK_NANDBIT(smp_in_mask, md->mi.gd_cpuid);
1292 md->gd_xinvaltlb = 0;
1296 cpu_wbinvd_on_all_cpus_callback(void *arg)
1302 * When called the executing CPU will send an IPI to all other CPUs
1303 * requesting that they halt execution.
1305 * Usually (but not necessarily) called with 'other_cpus' as its arg.
1307 * - Signals all CPUs in map to stop.
1308 * - Waits for each to stop.
1315 * XXX FIXME: this is not MP-safe, needs a lock to prevent multiple CPUs
1316 * from executing at same time.
1319 stop_cpus(cpumask_t map)
1323 CPUMASK_ANDMASK(map, smp_active_mask);
1325 /* send the Xcpustop IPI to all CPUs in map */
1326 selected_apic_ipi(map, XCPUSTOP_OFFSET, APIC_DELMODE_FIXED);
1329 mask = stopped_cpus;
1330 CPUMASK_ANDMASK(mask, map);
1332 } while (CPUMASK_CMPMASKNEQ(mask, map));
1339 * Called by a CPU to restart stopped CPUs.
1341 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
1343 * - Signals all CPUs in map to restart.
1344 * - Waits for each to restart.
1352 restart_cpus(cpumask_t map)
1356 /* signal other cpus to restart */
1358 CPUMASK_ANDMASK(mask, smp_active_mask);
1360 started_cpus = mask;
1363 /* wait for each to clear its bit */
1364 while (CPUMASK_CMPMASKNEQ(stopped_cpus, map))
1371 * This is called once the mpboot code has gotten us properly relocated
1372 * and the MMU turned on, etc. ap_init() is actually the idle thread,
1373 * and when it returns the scheduler will call the real cpu_idle() main
1374 * loop for the idlethread. Interrupts are disabled on entry and should
1375 * remain disabled at return.
1383 * Adjust smp_startup_mask to signal the BSP that we have started
1384 * up successfully. Note that we do not yet hold the BGL. The BSP
1385 * is waiting for our signal.
1387 * We can't set our bit in smp_active_mask yet because we are holding
1388 * interrupts physically disabled and remote cpus could deadlock
1389 * trying to send us an IPI.
1391 ATOMIC_CPUMASK_ORBIT(smp_startup_mask, mycpu->gd_cpuid);
1395 * Interlock for LAPIC initialization. Wait until mp_finish_lapic is
1396 * non-zero, then get the MP lock.
1398 * Note: We are in a critical section.
1400 * Note: we are the idle thread, we can only spin.
1402 * Note: The load fence is memory volatile and prevents the compiler
1403 * from improperly caching mp_finish_lapic, and the cpu from improperly
1406 while (mp_finish_lapic == 0) {
1411 while (try_mplock() == 0) {
1417 if (cpu_feature & CPUID_TSC) {
1419 * The BSP is constantly updating tsc0_offset, figure out
1420 * the relative difference to synchronize ktrdump.
1422 tsc_offsets[mycpu->gd_cpuid] = rdtsc() - tsc0_offset;
1425 /* BSP may have changed PTD while we're waiting for the lock */
1428 /* Build our map of 'other' CPUs. */
1429 mycpu->gd_other_cpus = smp_startup_mask;
1430 ATOMIC_CPUMASK_NANDBIT(mycpu->gd_other_cpus, mycpu->gd_cpuid);
1432 /* A quick check from sanity claus */
1433 cpu_id = APICID_TO_CPUID((lapic->id & 0xff000000) >> 24);
1434 if (mycpu->gd_cpuid != cpu_id) {
1435 kprintf("SMP: assigned cpuid = %d\n", mycpu->gd_cpuid);
1436 kprintf("SMP: actual cpuid = %d lapicid %d\n",
1437 cpu_id, (lapic->id & 0xff000000) >> 24);
1439 kprintf("PTD[MPPTDI] = %p\n", (void *)PTD[MPPTDI]);
1441 panic("cpuid mismatch! boom!!");
1444 /* Initialize AP's local APIC for irq's */
1447 /* LAPIC initialization is done */
1448 ATOMIC_CPUMASK_ORBIT(smp_lapic_mask, mycpu->gd_cpuid);
1452 /* Let BSP move onto the next initialization stage */
1457 * Interlock for finalization. Wait until mp_finish is non-zero,
1458 * then get the MP lock.
1460 * Note: We are in a critical section.
1462 * Note: we are the idle thread, we can only spin.
1464 * Note: The load fence is memory volatile and prevents the compiler
1465 * from improperly caching mp_finish, and the cpu from improperly
1468 while (mp_finish == 0) {
1473 /* BSP may have changed PTD while we're waiting for the lock */
1476 /* Set memory range attributes for this CPU to match the BSP */
1477 mem_range_AP_init();
1480 * Once we go active we must process any IPIQ messages that may
1481 * have been queued, because no actual IPI will occur until we
1482 * set our bit in the smp_active_mask. If we don't the IPI
1483 * message interlock could be left set which would also prevent
1486 * The idle loop doesn't expect the BGL to be held and while
1487 * lwkt_switch() normally cleans things up this is a special case
1488 * because we returning almost directly into the idle loop.
1490 * The idle thread is never placed on the runq, make sure
1491 * nothing we've done put it there.
1495 * Hold a critical section and allow real interrupts to occur. Zero
1496 * any spurious interrupts which have accumulated, then set our
1497 * smp_active_mask indicating that we are fully operational.
1500 __asm __volatile("sti; pause; pause"::);
1501 bzero(mdcpu->gd_ipending, sizeof(mdcpu->gd_ipending));
1502 ATOMIC_CPUMASK_ORBIT(smp_active_mask, mycpu->gd_cpuid);
1505 * Wait until all cpus have set their smp_active_mask and have fully
1506 * operational interrupts before proceeding.
1508 * We need a final cpu_invltlb() because we would not have received
1509 * any until we set our bit in smp_active_mask.
1511 while (mp_finish == 1) {
1518 * Initialize per-cpu clocks and do other per-cpu initialization.
1519 * At this point code is expected to be able to use the full kernel
1522 initclocks_pcpu(); /* clock interrupts (via IPIs) */
1525 * Since we may have cleaned up the interrupt triggers, manually
1526 * process any pending IPIs before exiting our critical section.
1527 * Once the critical section has exited, normal interrupt processing
1530 atomic_swap_int(&mycpu->gd_npoll, 0);
1531 lwkt_process_ipiq();
1535 * Final final, allow the waiting BSP to resume the boot process,
1536 * return 'into' the idle thread bootstrap.
1538 ATOMIC_CPUMASK_ORBIT(smp_finalize_mask, mycpu->gd_cpuid);
1539 KKASSERT((curthread->td_flags & TDF_RUNQ) == 0);
1543 * Get SMP fully working before we start initializing devices.
1550 kprintf("Finish MP startup\n");
1554 * Wait for the active mask to complete, after which all cpus will
1555 * be accepting interrupts.
1558 while (CPUMASK_CMPMASKNEQ(smp_active_mask, smp_startup_mask)) {
1564 * Wait for the finalization mask to complete, after which all cpus
1565 * have completely finished initializing and are entering or are in
1566 * their idle thread.
1568 * BSP should have received all required invltlbs but do another
1573 while (CPUMASK_CMPMASKNEQ(smp_finalize_mask, smp_startup_mask)) {
1578 while (try_mplock() == 0) {
1584 kprintf("Active CPU Mask: %016jx\n",
1585 (uintmax_t)CPUMASK_LOWMASK(smp_active_mask));
1589 SYSINIT(finishsmp, SI_BOOT2_FINISH_SMP, SI_ORDER_FIRST, ap_finish, NULL);
1592 * Interrupts must be hard-disabled by caller
1595 cpu_send_ipiq(int dcpu)
1597 if (CPUMASK_TESTBIT(smp_active_mask, dcpu))
1598 single_apic_ipi(dcpu, XIPIQ_OFFSET, APIC_DELMODE_FIXED);
1601 #if 0 /* single_apic_ipi_passive() not working yet */
1603 * Returns 0 on failure, 1 on success
1606 cpu_send_ipiq_passive(int dcpu)
1609 if (CPUMASK_TESTBIT(smp_active_mask, dcpu)) {
1610 r = single_apic_ipi_passive(dcpu, XIPIQ_OFFSET,
1611 APIC_DELMODE_FIXED);
1618 mp_bsp_simple_setup(void)
1620 struct mdglobaldata *gd;
1623 /* build our map of 'other' CPUs */
1624 mycpu->gd_other_cpus = smp_startup_mask;
1625 CPUMASK_NANDBIT(mycpu->gd_other_cpus, mycpu->gd_cpuid);
1627 gd = (struct mdglobaldata *)mycpu;
1628 gd->gd_acpi_id = CPUID_TO_ACPIID(mycpu->gd_cpuid);
1630 ipiq_size = sizeof(struct lwkt_ipiq) * ncpus;
1631 mycpu->gd_ipiq = (void *)kmem_alloc(&kernel_map, ipiq_size,
1633 bzero(mycpu->gd_ipiq, ipiq_size);
1635 /* initialize arc4random. */
1640 if (cpu_feature & CPUID_TSC)
1641 tsc0_offset = rdtsc();
1646 * CPU TOPOLOGY DETECTION FUNCTIONS
1649 /* Detect intel topology using CPUID
1650 * Ref: http://www.intel.com/Assets/PDF/appnote/241618.pdf, pg 41
1653 detect_intel_topology(int count_htt_cores)
1657 int core_plus_logical_bits = 0;
1658 int cores_per_package;
1659 int logical_per_package;
1660 int logical_per_core;
1663 if (cpu_high >= 0xb) {
1666 } else if (cpu_high >= 0x4) {
1671 for (shift = 0; (1 << shift) < count_htt_cores; ++shift)
1673 logical_CPU_bits = 1 << shift;
1678 cpuid_count(0xb, FUNC_B_THREAD_LEVEL, p);
1680 /* if 0xb not supported - fallback to 0x4 */
1681 if (p[1] == 0 || (FUNC_B_TYPE(p[2]) != FUNC_B_THREAD_TYPE)) {
1685 logical_CPU_bits = FUNC_B_BITS_SHIFT_NEXT_LEVEL(p[0]);
1687 ecx_index = FUNC_B_THREAD_LEVEL + 1;
1689 cpuid_count(0xb, ecx_index, p);
1691 /* Check for the Core type in the implemented sub leaves. */
1692 if (FUNC_B_TYPE(p[2]) == FUNC_B_CORE_TYPE) {
1693 core_plus_logical_bits = FUNC_B_BITS_SHIFT_NEXT_LEVEL(p[0]);
1699 } while (FUNC_B_TYPE(p[2]) != FUNC_B_INVALID_TYPE);
1701 core_bits = core_plus_logical_bits - logical_CPU_bits;
1706 cpuid_count(0x4, 0, p);
1707 cores_per_package = FUNC_4_MAX_CORE_NO(p[0]) + 1;
1709 logical_per_package = count_htt_cores;
1710 logical_per_core = logical_per_package / cores_per_package;
1712 for (shift = 0; (1 << shift) < logical_per_core; ++shift)
1714 logical_CPU_bits = shift;
1716 for (shift = 0; (1 << shift) < cores_per_package; ++shift)
1723 /* Detect AMD topology using CPUID
1724 * Ref: http://support.amd.com/us/Embedded_TechDocs/25481.pdf, last page
1727 detect_amd_topology(int count_htt_cores)
1730 if ((cpu_feature & CPUID_HTT) && (amd_feature2 & AMDID2_CMP)) {
1731 if (cpu_procinfo2 & AMDID_COREID_SIZE) {
1732 core_bits = (cpu_procinfo2 & AMDID_COREID_SIZE) >>
1733 AMDID_COREID_SIZE_SHIFT;
1735 core_bits = (cpu_procinfo2 & AMDID_CMP_CORES) + 1;
1736 for (shift = 0; (1 << shift) < core_bits; ++shift)
1741 if (amd_feature2 & AMDID2_TOPOEXT) {
1747 for (i = 0; i < 256; ++i) {
1748 cpuid_count(0x8000001d, i, p);
1750 level = (p[0] >> 5) & 0x7;
1751 share_count = 1 + ((p[0] >> 14) & 0xfff);
1756 kprintf("Topology probe i=%2d type=%d level=%d share_count=%d\n",
1757 i, type, level, share_count);
1758 if (type == 1 && share_count) { /* CPUID_TYPE_SMT */
1759 for (shift = 0; (1 << shift) < count_htt_cores / share_count; ++shift)
1767 logical_CPU_bits = count_htt_cores >> core_bits;
1768 for (shift = 0; (1 << shift) < logical_CPU_bits; ++shift)
1770 logical_CPU_bits = shift;
1772 for (shift = 0; (1 << shift) < count_htt_cores; ++shift)
1775 logical_CPU_bits = 0;
1780 amd_get_compute_unit_id(void *arg)
1784 do_cpuid(0x8000001e, regs);
1785 cpu_node_t * mynode = get_cpu_node_by_cpuid(mycpuid);
1788 * AMD - CPUID Specification September 2010
1789 * page 34 - //ComputeUnitID = ebx[0:7]//
1791 mynode->compute_unit_id = regs[1] & 0xff;
1795 fix_amd_topology(void)
1799 if (cpu_vendor_id != CPU_VENDOR_AMD)
1801 if ((amd_feature2 & AMDID2_TOPOEXT) == 0)
1804 CPUMASK_ASSALLONES(mask);
1805 lwkt_cpusync_simple(mask, amd_get_compute_unit_id, NULL);
1807 kprintf("Compute unit iDS:\n");
1809 for (i = 0; i < ncpus; i++) {
1810 kprintf("%d-%d; \n",
1811 i, get_cpu_node_by_cpuid(i)->compute_unit_id);
1818 * - logical_CPU_bits
1820 * With the values above (for AMD or INTEL) we are able to generally
1821 * detect the CPU topology (number of cores for each level):
1822 * Ref: http://wiki.osdev.org/Detecting_CPU_Topology_(80x86)
1823 * Ref: http://www.multicoreinfo.com/research/papers/whitepapers/Intel-detect-topology.pdf
1826 detect_cpu_topology(void)
1828 static int topology_detected = 0;
1831 if (topology_detected)
1833 if ((cpu_feature & CPUID_HTT) == 0) {
1835 logical_CPU_bits = 0;
1838 count = (cpu_procinfo & CPUID_HTT_CORES) >> CPUID_HTT_CORE_SHIFT;
1840 if (cpu_vendor_id == CPU_VENDOR_INTEL)
1841 detect_intel_topology(count);
1842 else if (cpu_vendor_id == CPU_VENDOR_AMD)
1843 detect_amd_topology(count);
1844 topology_detected = 1;
1848 kprintf("Bits within APICID: logical_CPU_bits: %d; "
1850 logical_CPU_bits, core_bits);
1855 * Interface functions to calculate chip_ID,
1856 * core_number and logical_number
1857 * Ref: http://wiki.osdev.org/Detecting_CPU_Topology_(80x86)
1860 get_chip_ID(int cpuid)
1862 return get_apicid_from_cpuid(cpuid) >>
1863 (logical_CPU_bits + core_bits);
1867 get_chip_ID_from_APICID(int apicid)
1869 return apicid >> (logical_CPU_bits + core_bits);
1873 get_core_number_within_chip(int cpuid)
1875 return ((get_apicid_from_cpuid(cpuid) >> logical_CPU_bits) &
1876 ((1 << core_bits) - 1));
1880 get_logical_CPU_number_within_core(int cpuid)
1882 return (get_apicid_from_cpuid(cpuid) &
1883 ((1 << logical_CPU_bits) - 1));