2 * Copyright (c) 1996, by Steve Passe
3 * Copyright (c) 2003, by Peter Wemm
6 * Redistribution and use in source and binary forms, with or without
7 * modification, are permitted provided that the following conditions
9 * 1. Redistributions of source code must retain the above copyright
10 * notice, this list of conditions and the following disclaimer.
11 * 2. The name of the developer may NOT be used to endorse or promote products
12 * derived from this software without specific prior written permission.
14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27 #include <sys/cdefs.h>
28 __FBSDID("$FreeBSD$");
31 #include "opt_kstack_pages.h"
32 #include "opt_sched.h"
35 #include <sys/param.h>
36 #include <sys/systm.h>
38 #include <sys/cpuset.h>
42 #include <sys/kernel.h>
45 #include <sys/malloc.h>
46 #include <sys/memrange.h>
47 #include <sys/mutex.h>
50 #include <sys/sched.h>
52 #include <sys/sysctl.h>
55 #include <vm/vm_param.h>
57 #include <vm/vm_kern.h>
58 #include <vm/vm_extern.h>
60 #include <x86/apicreg.h>
61 #include <machine/clock.h>
62 #include <machine/cputypes.h>
63 #include <machine/cpufunc.h>
65 #include <machine/md_var.h>
66 #include <machine/pcb.h>
67 #include <machine/psl.h>
68 #include <machine/smp.h>
69 #include <machine/specialreg.h>
70 #include <machine/tss.h>
72 #define WARMBOOT_TARGET 0
73 #define WARMBOOT_OFF (KERNBASE + 0x0467)
74 #define WARMBOOT_SEG (KERNBASE + 0x0469)
76 #define CMOS_REG (0x70)
77 #define CMOS_DATA (0x71)
78 #define BIOS_RESET (0x0f)
79 #define BIOS_WARM (0x0a)
81 /* lock region used by kernel profiling */
84 int mp_naps; /* # of Applications processors */
85 int boot_cpu_id = -1; /* designated BSP */
87 extern struct pcpu __pcpu[];
89 /* AP uses this during bootstrap. Do not staticize. */
93 /* Free these after use */
94 void *bootstacks[MAXCPU];
96 /* Temporary variables for init_secondary() */
97 char *doublefault_stack;
101 struct pcb stoppcbs[MAXCPU];
102 struct pcb **susppcbs;
105 /* Variables needed for SMP tlb shootdown. */
106 vm_offset_t smp_tlb_addr1;
107 vm_offset_t smp_tlb_addr2;
108 volatile int smp_tlb_wait;
111 /* Interrupt counts. */
112 static u_long *ipi_preempt_counts[MAXCPU];
113 static u_long *ipi_ast_counts[MAXCPU];
114 u_long *ipi_invltlb_counts[MAXCPU];
115 u_long *ipi_invlrng_counts[MAXCPU];
116 u_long *ipi_invlpg_counts[MAXCPU];
117 u_long *ipi_invlcache_counts[MAXCPU];
118 u_long *ipi_rendezvous_counts[MAXCPU];
119 static u_long *ipi_hardclock_counts[MAXCPU];
122 extern inthand_t IDTVEC(fast_syscall), IDTVEC(fast_syscall32);
125 * Local data and functions.
128 static volatile cpuset_t ipi_nmi_pending;
130 /* used to hold the AP's until we are ready to release them */
131 static struct mtx ap_boot_mtx;
133 /* Set to 1 once we're ready to let the APs out of the pen. */
134 static volatile int aps_ready = 0;
137 * Store data from cpu_add() until later in the boot when we actually setup
144 int cpu_hyperthread:1;
145 } static cpu_info[MAX_APIC_ID + 1];
146 int cpu_apic_ids[MAXCPU];
147 int apic_cpuids[MAX_APIC_ID + 1];
149 /* Holds pending bitmap based IPIs per CPU */
150 static volatile u_int cpu_ipi_pending[MAXCPU];
152 static u_int boot_address;
153 static int cpu_logical; /* logical cpus per core */
154 static int cpu_cores; /* cores per package */
156 static void assign_cpu_ids(void);
157 static void set_interrupt_apic_ids(void);
158 static int start_all_aps(void);
159 static int start_ap(int apic_id);
160 static void release_aps(void *dummy);
162 static u_int hyperthreading_cpus; /* logical cpus sharing L1 cache */
163 static int hyperthreading_allowed = 1;
164 static u_int bootMP_size;
167 mem_range_AP_init(void)
169 if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP)
170 mem_range_softc.mr_op->initAP(&mem_range_softc);
179 /* AMD processors do not support HTT. */
182 if ((amd_feature2 & AMDID2_CMP) == 0) {
187 core_id_bits = (cpu_procinfo2 & AMDID_COREID_SIZE) >>
188 AMDID_COREID_SIZE_SHIFT;
189 if (core_id_bits == 0) {
190 cpu_cores = (cpu_procinfo2 & AMDID_CMP_CORES) + 1;
194 /* Fam 10h and newer should get here. */
195 for (id = 0; id <= MAX_APIC_ID; id++) {
196 /* Check logical CPU availability. */
197 if (!cpu_info[id].cpu_present || cpu_info[id].cpu_disabled)
199 /* Check if logical CPU has the same package ID. */
200 if ((id >> core_id_bits) != (boot_cpu_id >> core_id_bits))
207 * Round up to the next power of two, if necessary, and then
209 * Returns -1 if argument is zero.
215 return (fls(x << (1 - powerof2(x))) - 1);
228 /* Both zero and one here mean one logical processor per package. */
229 max_logical = (cpu_feature & CPUID_HTT) != 0 ?
230 (cpu_procinfo & CPUID_HTT_CORES) >> 16 : 1;
231 if (max_logical <= 1)
235 * Because of uniformity assumption we examine only
236 * those logical processors that belong to the same
237 * package as BSP. Further, we count number of
238 * logical processors that belong to the same core
239 * as BSP thus deducing number of threads per core.
241 if (cpu_high >= 0x4) {
242 cpuid_count(0x04, 0, p);
243 max_cores = ((p[0] >> 26) & 0x3f) + 1;
246 core_id_bits = mask_width(max_logical/max_cores);
247 if (core_id_bits < 0)
249 pkg_id_bits = core_id_bits + mask_width(max_cores);
251 for (id = 0; id <= MAX_APIC_ID; id++) {
252 /* Check logical CPU availability. */
253 if (!cpu_info[id].cpu_present || cpu_info[id].cpu_disabled)
255 /* Check if logical CPU has the same package ID. */
256 if ((id >> pkg_id_bits) != (boot_cpu_id >> pkg_id_bits))
259 /* Check if logical CPU has the same package and core IDs. */
260 if ((id >> core_id_bits) == (boot_cpu_id >> core_id_bits))
264 KASSERT(cpu_cores >= 1 && cpu_logical >= 1,
265 ("topo_probe_0x4 couldn't find BSP"));
267 cpu_cores /= cpu_logical;
268 hyperthreading_cpus = cpu_logical;
282 /* We only support three levels for now. */
283 for (i = 0; i < 3; i++) {
284 cpuid_count(0x0b, i, p);
286 /* Fall back if CPU leaf 11 doesn't really exist. */
287 if (i == 0 && p[1] == 0) {
293 logical = p[1] &= 0xffff;
294 type = (p[2] >> 8) & 0xff;
295 if (type == 0 || logical == 0)
298 * Because of uniformity assumption we examine only
299 * those logical processors that belong to the same
302 for (cnt = 0, x = 0; x <= MAX_APIC_ID; x++) {
303 if (!cpu_info[x].cpu_present ||
304 cpu_info[x].cpu_disabled)
306 if (x >> bits == boot_cpu_id >> bits)
309 if (type == CPUID_TYPE_SMT)
311 else if (type == CPUID_TYPE_CORE)
314 if (cpu_logical == 0)
316 cpu_cores /= cpu_logical;
320 * Both topology discovery code and code that consumes topology
321 * information assume top-down uniformity of the topology.
322 * That is, all physical packages must be identical and each
323 * core in a package must have the same number of threads.
324 * Topology information is queried only on BSP, on which this
325 * code runs and for which it can query CPUID information.
326 * Then topology is extrapolated on all packages using the
327 * uniformity assumption.
332 static int cpu_topo_probed = 0;
337 CPU_ZERO(&logical_cpus_mask);
339 cpu_cores = cpu_logical = 1;
340 else if (cpu_vendor_id == CPU_VENDOR_AMD)
342 else if (cpu_vendor_id == CPU_VENDOR_INTEL) {
344 * See Intel(R) 64 Architecture Processor
345 * Topology Enumeration article for details.
347 * Note that 0x1 <= cpu_high < 4 case should be
348 * compatible with topo_probe_0x4() logic when
349 * CPUID.1:EBX[23:16] > 0 (cpu_cores will be 1)
350 * or it should trigger the fallback otherwise.
354 else if (cpu_high >= 0x1)
359 * Fallback: assume each logical CPU is in separate
360 * physical package. That is, no multi-core, no SMT.
362 if (cpu_cores == 0 || cpu_logical == 0)
363 cpu_cores = cpu_logical = 1;
373 * Determine whether any threading flags are
377 if (cpu_logical > 1 && hyperthreading_cpus)
378 cg_flags = CG_FLAG_HTT;
379 else if (cpu_logical > 1)
380 cg_flags = CG_FLAG_SMT;
383 if (mp_ncpus % (cpu_cores * cpu_logical) != 0) {
384 printf("WARNING: Non-uniform processors.\n");
385 printf("WARNING: Using suboptimal topology.\n");
386 return (smp_topo_none());
389 * No multi-core or hyper-threaded.
391 if (cpu_logical * cpu_cores == 1)
392 return (smp_topo_none());
394 * Only HTT no multi-core.
396 if (cpu_logical > 1 && cpu_cores == 1)
397 return (smp_topo_1level(CG_SHARE_L1, cpu_logical, cg_flags));
399 * Only multi-core no HTT.
401 if (cpu_cores > 1 && cpu_logical == 1)
402 return (smp_topo_1level(CG_SHARE_L2, cpu_cores, cg_flags));
404 * Both HTT and multi-core.
406 return (smp_topo_2level(CG_SHARE_L2, cpu_cores,
407 CG_SHARE_L1, cpu_logical, cg_flags));
411 * Calculate usable address in base memory for AP trampoline code.
414 mp_bootaddress(u_int basemem)
417 bootMP_size = mptramp_end - mptramp_start;
418 boot_address = trunc_page(basemem * 1024); /* round down to 4k boundary */
419 if (((basemem * 1024) - boot_address) < bootMP_size)
420 boot_address -= PAGE_SIZE; /* not enough, lower by 4k */
421 /* 3 levels of page table pages */
422 mptramp_pagetables = boot_address - (PAGE_SIZE * 3);
424 return mptramp_pagetables;
428 cpu_add(u_int apic_id, char boot_cpu)
431 if (apic_id > MAX_APIC_ID) {
432 panic("SMP: APIC ID %d too high", apic_id);
435 KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %d added twice",
437 cpu_info[apic_id].cpu_present = 1;
439 KASSERT(boot_cpu_id == -1,
440 ("CPU %d claims to be BSP, but CPU %d already is", apic_id,
442 boot_cpu_id = apic_id;
443 cpu_info[apic_id].cpu_bsp = 1;
445 if (mp_ncpus < MAXCPU) {
447 mp_maxid = mp_ncpus - 1;
450 printf("SMP: Added CPU %d (%s)\n", apic_id, boot_cpu ? "BSP" :
455 cpu_mp_setmaxid(void)
459 * mp_maxid should be already set by calls to cpu_add().
460 * Just sanity check its value here.
463 KASSERT(mp_maxid == 0,
464 ("%s: mp_ncpus is zero, but mp_maxid is not", __func__));
465 else if (mp_ncpus == 1)
468 KASSERT(mp_maxid >= mp_ncpus - 1,
469 ("%s: counters out of sync: max %d, count %d", __func__,
470 mp_maxid, mp_ncpus));
478 * Always record BSP in CPU map so that the mbuf init code works
481 CPU_SETOF(0, &all_cpus);
484 * No CPUs were found, so this must be a UP system. Setup
485 * the variables to represent a system with a single CPU
492 /* At least one CPU was found. */
495 * One CPU was found, so this must be a UP system with
502 /* At least two CPUs were found. */
507 * Initialize the IPI handlers and start up the AP's.
514 /* Initialize the logical ID to APIC ID table. */
515 for (i = 0; i < MAXCPU; i++) {
516 cpu_apic_ids[i] = -1;
517 cpu_ipi_pending[i] = 0;
520 /* Install an inter-CPU IPI for TLB invalidation */
521 setidt(IPI_INVLTLB, IDTVEC(invltlb), SDT_SYSIGT, SEL_KPL, 0);
522 setidt(IPI_INVLPG, IDTVEC(invlpg), SDT_SYSIGT, SEL_KPL, 0);
523 setidt(IPI_INVLRNG, IDTVEC(invlrng), SDT_SYSIGT, SEL_KPL, 0);
525 /* Install an inter-CPU IPI for cache invalidation. */
526 setidt(IPI_INVLCACHE, IDTVEC(invlcache), SDT_SYSIGT, SEL_KPL, 0);
528 /* Install an inter-CPU IPI for all-CPU rendezvous */
529 setidt(IPI_RENDEZVOUS, IDTVEC(rendezvous), SDT_SYSIGT, SEL_KPL, 0);
531 /* Install generic inter-CPU IPI handler */
532 setidt(IPI_BITMAP_VECTOR, IDTVEC(ipi_intr_bitmap_handler),
533 SDT_SYSIGT, SEL_KPL, 0);
535 /* Install an inter-CPU IPI for CPU stop/restart */
536 setidt(IPI_STOP, IDTVEC(cpustop), SDT_SYSIGT, SEL_KPL, 0);
538 /* Install an inter-CPU IPI for CPU suspend/resume */
539 setidt(IPI_SUSPEND, IDTVEC(cpususpend), SDT_SYSIGT, SEL_KPL, 0);
541 /* Set boot_cpu_id if needed. */
542 if (boot_cpu_id == -1) {
543 boot_cpu_id = PCPU_GET(apic_id);
544 cpu_info[boot_cpu_id].cpu_bsp = 1;
546 KASSERT(boot_cpu_id == PCPU_GET(apic_id),
547 ("BSP's APIC ID doesn't match boot_cpu_id"));
549 /* Probe logical/physical core configuration. */
554 /* Start each Application Processor */
557 set_interrupt_apic_ids();
562 * Print various information about the SMP system hardware and setup.
565 cpu_mp_announce(void)
567 const char *hyperthread;
570 printf("FreeBSD/SMP: %d package(s) x %d core(s)",
571 mp_ncpus / (cpu_cores * cpu_logical), cpu_cores);
572 if (hyperthreading_cpus > 1)
573 printf(" x %d HTT threads", cpu_logical);
574 else if (cpu_logical > 1)
575 printf(" x %d SMT threads", cpu_logical);
578 /* List active CPUs first. */
579 printf(" cpu0 (BSP): APIC ID: %2d\n", boot_cpu_id);
580 for (i = 1; i < mp_ncpus; i++) {
581 if (cpu_info[cpu_apic_ids[i]].cpu_hyperthread)
585 printf(" cpu%d (AP%s): APIC ID: %2d\n", i, hyperthread,
589 /* List disabled CPUs last. */
590 for (i = 0; i <= MAX_APIC_ID; i++) {
591 if (!cpu_info[i].cpu_present || !cpu_info[i].cpu_disabled)
593 if (cpu_info[i].cpu_hyperthread)
597 printf(" cpu (AP%s): APIC ID: %2d (disabled)\n", hyperthread,
603 * AP CPU's call this to initialize themselves.
612 int cpu, gsel_tss, x;
613 struct region_descriptor ap_gdt;
615 /* Set by the startup code for us to use */
619 common_tss[cpu] = common_tss[0];
620 common_tss[cpu].tss_rsp0 = 0; /* not used until after switch */
621 common_tss[cpu].tss_iobase = sizeof(struct amd64tss) +
623 common_tss[cpu].tss_ist1 = (long)&doublefault_stack[PAGE_SIZE];
625 /* The NMI stack runs on IST2. */
626 np = ((struct nmi_pcpu *) &nmi_stack[PAGE_SIZE]) - 1;
627 common_tss[cpu].tss_ist2 = (long) np;
629 /* Prepare private GDT */
630 gdt_segs[GPROC0_SEL].ssd_base = (long) &common_tss[cpu];
631 for (x = 0; x < NGDT; x++) {
632 if (x != GPROC0_SEL && x != (GPROC0_SEL + 1) &&
633 x != GUSERLDT_SEL && x != (GUSERLDT_SEL + 1))
634 ssdtosd(&gdt_segs[x], &gdt[NGDT * cpu + x]);
636 ssdtosyssd(&gdt_segs[GPROC0_SEL],
637 (struct system_segment_descriptor *)&gdt[NGDT * cpu + GPROC0_SEL]);
638 ap_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
639 ap_gdt.rd_base = (long) &gdt[NGDT * cpu];
640 lgdt(&ap_gdt); /* does magic intra-segment return */
642 /* Get per-cpu data */
645 /* prime data page for it to use */
646 pcpu_init(pc, cpu, sizeof(struct pcpu));
647 dpcpu_init(dpcpu, cpu);
648 pc->pc_apic_id = cpu_apic_ids[cpu];
649 pc->pc_prvspace = pc;
650 pc->pc_curthread = 0;
651 pc->pc_tssp = &common_tss[cpu];
652 pc->pc_commontssp = &common_tss[cpu];
654 pc->pc_tss = (struct system_segment_descriptor *)&gdt[NGDT * cpu +
656 pc->pc_fs32p = &gdt[NGDT * cpu + GUFS32_SEL];
657 pc->pc_gs32p = &gdt[NGDT * cpu + GUGS32_SEL];
658 pc->pc_ldt = (struct system_segment_descriptor *)&gdt[NGDT * cpu +
661 /* Save the per-cpu pointer for use by the NMI handler. */
662 np->np_pcpu = (register_t) pc;
664 wrmsr(MSR_FSBASE, 0); /* User value */
665 wrmsr(MSR_GSBASE, (u_int64_t)pc);
666 wrmsr(MSR_KGSBASE, (u_int64_t)pc); /* XXX User value while we're in the kernel */
670 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
674 * Set to a known state:
675 * Set by mpboot.s: CR0_PG, CR0_PE
676 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
679 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
682 /* Set up the fast syscall stuff */
683 msr = rdmsr(MSR_EFER) | EFER_SCE;
684 wrmsr(MSR_EFER, msr);
685 wrmsr(MSR_LSTAR, (u_int64_t)IDTVEC(fast_syscall));
686 wrmsr(MSR_CSTAR, (u_int64_t)IDTVEC(fast_syscall32));
687 msr = ((u_int64_t)GSEL(GCODE_SEL, SEL_KPL) << 32) |
688 ((u_int64_t)GSEL(GUCODE32_SEL, SEL_UPL) << 48);
689 wrmsr(MSR_STAR, msr);
690 wrmsr(MSR_SF_MASK, PSL_NT|PSL_T|PSL_I|PSL_C|PSL_D);
692 /* Disable local APIC just to be sure. */
695 /* signal our startup to the BSP. */
698 /* Spin until the BSP releases the AP's. */
702 /* Initialize the PAT MSR. */
705 /* set up CPU registers and state */
708 /* set up SSE/NX registers */
711 /* set up FPU state on the AP */
714 /* A quick check from sanity claus */
715 cpuid = PCPU_GET(cpuid);
716 if (PCPU_GET(apic_id) != lapic_id()) {
717 printf("SMP: cpuid = %d\n", cpuid);
718 printf("SMP: actual apic_id = %d\n", lapic_id());
719 printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id));
720 panic("cpuid mismatch! boom!!");
723 /* Initialize curthread. */
724 KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread"));
725 PCPU_SET(curthread, PCPU_GET(idlethread));
729 mtx_lock_spin(&ap_boot_mtx);
731 /* Init local apic for irq's */
734 /* Set memory range attributes for this CPU to match the BSP */
739 CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid);
740 printf("SMP: AP CPU #%d Launched!\n", cpuid);
742 /* Determine if we are a logical CPU. */
743 /* XXX Calculation depends on cpu_logical being a power of 2, e.g. 2 */
744 if (cpu_logical > 1 && PCPU_GET(apic_id) % cpu_logical != 0)
745 CPU_SET(cpuid, &logical_cpus_mask);
750 if (smp_cpus == mp_ncpus) {
751 /* enable IPI's, tlb shootdown, freezes etc */
752 atomic_store_rel_int(&smp_started, 1);
753 smp_active = 1; /* historic */
757 * Enable global pages TLB extension
758 * This also implicitly flushes the TLB
761 load_cr4(rcr4() | CR4_PGE);
765 mtx_unlock_spin(&ap_boot_mtx);
767 /* Wait until all the AP's are up. */
768 while (smp_started == 0)
771 /* Start per-CPU event timers. */
776 panic("scheduler returned us to %s", __func__);
780 /*******************************************************************
781 * local functions and data
785 * We tell the I/O APIC code about all the CPUs we want to receive
786 * interrupts. If we don't want certain CPUs to receive IRQs we
787 * can simply not tell the I/O APIC code about them in this function.
790 set_interrupt_apic_ids(void)
794 for (i = 0; i < MAXCPU; i++) {
795 apic_id = cpu_apic_ids[i];
798 if (cpu_info[apic_id].cpu_disabled)
801 /* Don't let hyperthreads service interrupts. */
802 if (hyperthreading_cpus > 1 &&
803 apic_id % hyperthreading_cpus != 0)
811 * Assign logical CPU IDs to local APICs.
818 TUNABLE_INT_FETCH("machdep.hyperthreading_allowed",
819 &hyperthreading_allowed);
821 /* Check for explicitly disabled CPUs. */
822 for (i = 0; i <= MAX_APIC_ID; i++) {
823 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp)
826 if (hyperthreading_cpus > 1 && i % hyperthreading_cpus != 0) {
827 cpu_info[i].cpu_hyperthread = 1;
830 * Don't use HT CPU if it has been disabled by a
833 if (hyperthreading_allowed == 0) {
834 cpu_info[i].cpu_disabled = 1;
839 /* Don't use this CPU if it has been disabled by a tunable. */
840 if (resource_disabled("lapic", i)) {
841 cpu_info[i].cpu_disabled = 1;
846 if (hyperthreading_allowed == 0 && hyperthreading_cpus > 1) {
847 hyperthreading_cpus = 0;
852 * Assign CPU IDs to local APIC IDs and disable any CPUs
853 * beyond MAXCPU. CPU 0 is always assigned to the BSP.
855 * To minimize confusion for userland, we attempt to number
856 * CPUs such that all threads and cores in a package are
857 * grouped together. For now we assume that the BSP is always
858 * the first thread in a package and just start adding APs
859 * starting with the BSP's APIC ID.
862 cpu_apic_ids[0] = boot_cpu_id;
863 apic_cpuids[boot_cpu_id] = 0;
864 for (i = boot_cpu_id + 1; i != boot_cpu_id;
865 i == MAX_APIC_ID ? i = 0 : i++) {
866 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp ||
867 cpu_info[i].cpu_disabled)
870 if (mp_ncpus < MAXCPU) {
871 cpu_apic_ids[mp_ncpus] = i;
872 apic_cpuids[i] = mp_ncpus;
875 cpu_info[i].cpu_disabled = 1;
877 KASSERT(mp_maxid >= mp_ncpus - 1,
878 ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid,
883 * start each AP in our list
888 vm_offset_t va = boot_address + KERNBASE;
889 u_int64_t *pt4, *pt3, *pt2;
890 u_int32_t mpbioswarmvec;
894 mtx_init(&ap_boot_mtx, "ap boot", NULL, MTX_SPIN);
896 /* install the AP 1st level boot code */
897 pmap_kenter(va, boot_address);
898 pmap_invalidate_page(kernel_pmap, va);
899 bcopy(mptramp_start, (void *)va, bootMP_size);
901 /* Locate the page tables, they'll be below the trampoline */
902 pt4 = (u_int64_t *)(uintptr_t)(mptramp_pagetables + KERNBASE);
903 pt3 = pt4 + (PAGE_SIZE) / sizeof(u_int64_t);
904 pt2 = pt3 + (PAGE_SIZE) / sizeof(u_int64_t);
906 /* Create the initial 1GB replicated page tables */
907 for (i = 0; i < 512; i++) {
908 /* Each slot of the level 4 pages points to the same level 3 page */
909 pt4[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + PAGE_SIZE);
910 pt4[i] |= PG_V | PG_RW | PG_U;
912 /* Each slot of the level 3 pages points to the same level 2 page */
913 pt3[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + (2 * PAGE_SIZE));
914 pt3[i] |= PG_V | PG_RW | PG_U;
916 /* The level 2 page slots are mapped with 2MB pages for 1GB. */
917 pt2[i] = i * (2 * 1024 * 1024);
918 pt2[i] |= PG_V | PG_RW | PG_PS | PG_U;
921 /* save the current value of the warm-start vector */
922 mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF);
923 outb(CMOS_REG, BIOS_RESET);
924 mpbiosreason = inb(CMOS_DATA);
926 /* setup a vector to our boot code */
927 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
928 *((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4);
929 outb(CMOS_REG, BIOS_RESET);
930 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
933 for (cpu = 1; cpu < mp_ncpus; cpu++) {
934 apic_id = cpu_apic_ids[cpu];
936 /* allocate and set up an idle stack data page */
937 bootstacks[cpu] = (void *)kmem_alloc(kernel_map, KSTACK_PAGES * PAGE_SIZE);
938 doublefault_stack = (char *)kmem_alloc(kernel_map, PAGE_SIZE);
939 nmi_stack = (char *)kmem_alloc(kernel_map, PAGE_SIZE);
940 dpcpu = (void *)kmem_alloc(kernel_map, DPCPU_SIZE);
942 bootSTK = (char *)bootstacks[cpu] + KSTACK_PAGES * PAGE_SIZE - 8;
945 /* attempt to start the Application Processor */
946 if (!start_ap(apic_id)) {
947 /* restore the warmstart vector */
948 *(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec;
949 panic("AP #%d (PHY# %d) failed!", cpu, apic_id);
952 CPU_SET(cpu, &all_cpus); /* record AP in CPU map */
955 /* restore the warmstart vector */
956 *(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec;
958 outb(CMOS_REG, BIOS_RESET);
959 outb(CMOS_DATA, mpbiosreason);
961 /* number of APs actually started */
967 * This function starts the AP (application processor) identified
968 * by the APIC ID 'physicalCpu'. It does quite a "song and dance"
969 * to accomplish this. This is necessary because of the nuances
970 * of the different hardware we might encounter. It isn't pretty,
971 * but it seems to work.
974 start_ap(int apic_id)
979 /* calculate the vector */
980 vector = (boot_address >> 12) & 0xff;
982 /* used as a watchpoint to signal AP startup */
986 * first we do an INIT/RESET IPI this INIT IPI might be run, reseting
987 * and running the target CPU. OR this INIT IPI might be latched (P5
988 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
992 /* do an INIT IPI: assert RESET */
993 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
994 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id);
996 /* wait for pending status end */
999 /* do an INIT IPI: deassert RESET */
1000 lapic_ipi_raw(APIC_DEST_ALLESELF | APIC_TRIGMOD_LEVEL |
1001 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, 0);
1003 /* wait for pending status end */
1004 DELAY(10000); /* wait ~10mS */
1008 * next we do a STARTUP IPI: the previous INIT IPI might still be
1009 * latched, (P5 bug) this 1st STARTUP would then terminate
1010 * immediately, and the previously started INIT IPI would continue. OR
1011 * the previous INIT IPI has already run. and this STARTUP IPI will
1012 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
1016 /* do a STARTUP IPI */
1017 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1018 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1021 DELAY(200); /* wait ~200uS */
1024 * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
1025 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
1026 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
1027 * recognized after hardware RESET or INIT IPI.
1030 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1031 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1034 DELAY(200); /* wait ~200uS */
1036 /* Wait up to 5 seconds for it to start. */
1037 for (ms = 0; ms < 5000; ms++) {
1039 return 1; /* return SUCCESS */
1042 return 0; /* return FAILURE */
1045 #ifdef COUNT_XINVLTLB_HITS
1046 u_int xhits_gbl[MAXCPU];
1047 u_int xhits_pg[MAXCPU];
1048 u_int xhits_rng[MAXCPU];
1049 static SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW, 0, "");
1050 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl,
1051 sizeof(xhits_gbl), "IU", "");
1052 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg,
1053 sizeof(xhits_pg), "IU", "");
1054 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng,
1055 sizeof(xhits_rng), "IU", "");
1060 u_int ipi_range_size;
1061 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, "");
1062 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, "");
1063 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, "");
1064 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW,
1065 &ipi_range_size, 0, "");
1067 u_int ipi_masked_global;
1068 u_int ipi_masked_page;
1069 u_int ipi_masked_range;
1070 u_int ipi_masked_range_size;
1071 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_masked_global, CTLFLAG_RW,
1072 &ipi_masked_global, 0, "");
1073 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_masked_page, CTLFLAG_RW,
1074 &ipi_masked_page, 0, "");
1075 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_masked_range, CTLFLAG_RW,
1076 &ipi_masked_range, 0, "");
1077 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_masked_range_size, CTLFLAG_RW,
1078 &ipi_masked_range_size, 0, "");
1079 #endif /* COUNT_XINVLTLB_HITS */
1082 * Send an IPI to specified CPU handling the bitmap logic.
1085 ipi_send_cpu(int cpu, u_int ipi)
1087 u_int bitmap, old_pending, new_pending;
1089 KASSERT(cpu_apic_ids[cpu] != -1, ("IPI to non-existent CPU %d", cpu));
1091 if (IPI_IS_BITMAPED(ipi)) {
1093 ipi = IPI_BITMAP_VECTOR;
1095 old_pending = cpu_ipi_pending[cpu];
1096 new_pending = old_pending | bitmap;
1097 } while (!atomic_cmpset_int(&cpu_ipi_pending[cpu],
1098 old_pending, new_pending));
1102 lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]);
1106 * Flush the TLB on all other CPU's
1109 smp_tlb_shootdown(u_int vector, vm_offset_t addr1, vm_offset_t addr2)
1113 ncpu = mp_ncpus - 1; /* does not shootdown self */
1115 return; /* no other cpus */
1116 if (!(read_rflags() & PSL_I))
1117 panic("%s: interrupts disabled", __func__);
1118 mtx_lock_spin(&smp_ipi_mtx);
1119 smp_tlb_addr1 = addr1;
1120 smp_tlb_addr2 = addr2;
1121 atomic_store_rel_int(&smp_tlb_wait, 0);
1122 ipi_all_but_self(vector);
1123 while (smp_tlb_wait < ncpu)
1125 mtx_unlock_spin(&smp_ipi_mtx);
1129 smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, vm_offset_t addr1, vm_offset_t addr2)
1131 int cpu, ncpu, othercpus;
1133 othercpus = mp_ncpus - 1;
1134 if (CPU_ISFULLSET(&mask)) {
1138 CPU_CLR(PCPU_GET(cpuid), &mask);
1139 if (CPU_EMPTY(&mask))
1142 if (!(read_rflags() & PSL_I))
1143 panic("%s: interrupts disabled", __func__);
1144 mtx_lock_spin(&smp_ipi_mtx);
1145 smp_tlb_addr1 = addr1;
1146 smp_tlb_addr2 = addr2;
1147 atomic_store_rel_int(&smp_tlb_wait, 0);
1148 if (CPU_ISFULLSET(&mask)) {
1150 ipi_all_but_self(vector);
1153 while ((cpu = cpusetobj_ffs(&mask)) != 0) {
1155 CPU_CLR(cpu, &mask);
1156 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__,
1158 ipi_send_cpu(cpu, vector);
1162 while (smp_tlb_wait < ncpu)
1164 mtx_unlock_spin(&smp_ipi_mtx);
1168 smp_cache_flush(void)
1172 smp_tlb_shootdown(IPI_INVLCACHE, 0, 0);
1180 smp_tlb_shootdown(IPI_INVLTLB, 0, 0);
1181 #ifdef COUNT_XINVLTLB_HITS
1188 smp_invlpg(vm_offset_t addr)
1192 smp_tlb_shootdown(IPI_INVLPG, addr, 0);
1193 #ifdef COUNT_XINVLTLB_HITS
1200 smp_invlpg_range(vm_offset_t addr1, vm_offset_t addr2)
1204 smp_tlb_shootdown(IPI_INVLRNG, addr1, addr2);
1205 #ifdef COUNT_XINVLTLB_HITS
1207 ipi_range_size += (addr2 - addr1) / PAGE_SIZE;
1213 smp_masked_invltlb(cpuset_t mask)
1217 smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, 0, 0);
1218 #ifdef COUNT_XINVLTLB_HITS
1219 ipi_masked_global++;
1225 smp_masked_invlpg(cpuset_t mask, vm_offset_t addr)
1229 smp_targeted_tlb_shootdown(mask, IPI_INVLPG, addr, 0);
1230 #ifdef COUNT_XINVLTLB_HITS
1237 smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2)
1241 smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, addr1, addr2);
1242 #ifdef COUNT_XINVLTLB_HITS
1244 ipi_masked_range_size += (addr2 - addr1) / PAGE_SIZE;
1250 ipi_bitmap_handler(struct trapframe frame)
1252 struct trapframe *oldframe;
1254 int cpu = PCPU_GET(cpuid);
1259 td->td_intr_nesting_level++;
1260 oldframe = td->td_intr_frame;
1261 td->td_intr_frame = &frame;
1262 ipi_bitmap = atomic_readandclear_int(&cpu_ipi_pending[cpu]);
1263 if (ipi_bitmap & (1 << IPI_PREEMPT)) {
1265 (*ipi_preempt_counts[cpu])++;
1269 if (ipi_bitmap & (1 << IPI_AST)) {
1271 (*ipi_ast_counts[cpu])++;
1273 /* Nothing to do for AST */
1275 if (ipi_bitmap & (1 << IPI_HARDCLOCK)) {
1277 (*ipi_hardclock_counts[cpu])++;
1281 td->td_intr_frame = oldframe;
1282 td->td_intr_nesting_level--;
1287 * send an IPI to a set of cpus.
1290 ipi_selected(cpuset_t cpus, u_int ipi)
1295 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1296 * of help in order to understand what is the source.
1297 * Set the mask of receiving CPUs for this purpose.
1299 if (ipi == IPI_STOP_HARD)
1300 CPU_OR_ATOMIC(&ipi_nmi_pending, &cpus);
1302 while ((cpu = cpusetobj_ffs(&cpus)) != 0) {
1304 CPU_CLR(cpu, &cpus);
1305 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1306 ipi_send_cpu(cpu, ipi);
1311 * send an IPI to a specific CPU.
1314 ipi_cpu(int cpu, u_int ipi)
1318 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1319 * of help in order to understand what is the source.
1320 * Set the mask of receiving CPUs for this purpose.
1322 if (ipi == IPI_STOP_HARD)
1323 CPU_SET_ATOMIC(cpu, &ipi_nmi_pending);
1325 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1326 ipi_send_cpu(cpu, ipi);
1330 * send an IPI to all CPUs EXCEPT myself
1333 ipi_all_but_self(u_int ipi)
1335 cpuset_t other_cpus;
1337 other_cpus = all_cpus;
1338 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1340 if (IPI_IS_BITMAPED(ipi)) {
1341 ipi_selected(other_cpus, ipi);
1346 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1347 * of help in order to understand what is the source.
1348 * Set the mask of receiving CPUs for this purpose.
1350 if (ipi == IPI_STOP_HARD)
1351 CPU_OR_ATOMIC(&ipi_nmi_pending, &other_cpus);
1353 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1354 lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS);
1363 * As long as there is not a simple way to know about a NMI's
1364 * source, if the bitmask for the current CPU is present in
1365 * the global pending bitword an IPI_STOP_HARD has been issued
1366 * and should be handled.
1368 cpuid = PCPU_GET(cpuid);
1369 if (!CPU_ISSET(cpuid, &ipi_nmi_pending))
1372 CPU_CLR_ATOMIC(cpuid, &ipi_nmi_pending);
1378 * Handle an IPI_STOP by saving our current context and spinning until we
1382 cpustop_handler(void)
1386 cpu = PCPU_GET(cpuid);
1388 savectx(&stoppcbs[cpu]);
1390 /* Indicate that we are stopped */
1391 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1393 /* Wait for restart */
1394 while (!CPU_ISSET(cpu, &started_cpus))
1397 CPU_CLR_ATOMIC(cpu, &started_cpus);
1398 CPU_CLR_ATOMIC(cpu, &stopped_cpus);
1400 if (cpu == 0 && cpustop_restartfunc != NULL) {
1401 cpustop_restartfunc();
1402 cpustop_restartfunc = NULL;
1407 * Handle an IPI_SUSPEND by saving our current context and spinning until we
1411 cpususpend_handler(void)
1415 cpu = PCPU_GET(cpuid);
1417 if (savectx(susppcbs[cpu])) {
1418 ctx_fpusave(suspfpusave[cpu]);
1420 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1423 load_cr3(susppcbs[cpu]->pcb_cr3);
1425 PCPU_SET(switchtime, 0);
1426 PCPU_SET(switchticks, ticks);
1429 /* Wait for resume */
1430 while (!CPU_ISSET(cpu, &started_cpus))
1433 CPU_CLR_ATOMIC(cpu, &started_cpus);
1434 CPU_CLR_ATOMIC(cpu, &stopped_cpus);
1436 /* Resume MCA and local APIC */
1442 * This is called once the rest of the system is up and running and we're
1443 * ready to let the AP's out of the pen.
1446 release_aps(void *dummy __unused)
1451 atomic_store_rel_int(&aps_ready, 1);
1452 while (smp_started == 0)
1455 SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL);
1459 * Setup interrupt counters for IPI handlers.
1462 mp_ipi_intrcnt(void *dummy)
1468 snprintf(buf, sizeof(buf), "cpu%d:invltlb", i);
1469 intrcnt_add(buf, &ipi_invltlb_counts[i]);
1470 snprintf(buf, sizeof(buf), "cpu%d:invlrng", i);
1471 intrcnt_add(buf, &ipi_invlrng_counts[i]);
1472 snprintf(buf, sizeof(buf), "cpu%d:invlpg", i);
1473 intrcnt_add(buf, &ipi_invlpg_counts[i]);
1474 snprintf(buf, sizeof(buf), "cpu%d:invlcache", i);
1475 intrcnt_add(buf, &ipi_invlcache_counts[i]);
1476 snprintf(buf, sizeof(buf), "cpu%d:preempt", i);
1477 intrcnt_add(buf, &ipi_preempt_counts[i]);
1478 snprintf(buf, sizeof(buf), "cpu%d:ast", i);
1479 intrcnt_add(buf, &ipi_ast_counts[i]);
1480 snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i);
1481 intrcnt_add(buf, &ipi_rendezvous_counts[i]);
1482 snprintf(buf, sizeof(buf), "cpu%d:hardclock", i);
1483 intrcnt_add(buf, &ipi_hardclock_counts[i]);
1486 SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL);