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("$FreeBSD$");
32 #include "opt_kstack_pages.h"
33 #include "opt_sched.h"
36 #include <sys/param.h>
37 #include <sys/systm.h>
39 #include <sys/cpuset.h>
43 #include <sys/kernel.h>
46 #include <sys/malloc.h>
47 #include <sys/memrange.h>
48 #include <sys/mutex.h>
51 #include <sys/sched.h>
53 #include <sys/sysctl.h>
56 #include <vm/vm_param.h>
58 #include <vm/vm_kern.h>
59 #include <vm/vm_extern.h>
61 #include <x86/apicreg.h>
62 #include <machine/clock.h>
63 #include <machine/cputypes.h>
64 #include <machine/cpufunc.h>
66 #include <machine/md_var.h>
67 #include <machine/pcb.h>
68 #include <machine/psl.h>
69 #include <machine/smp.h>
70 #include <machine/specialreg.h>
71 #include <machine/tss.h>
73 #define WARMBOOT_TARGET 0
74 #define WARMBOOT_OFF (KERNBASE + 0x0467)
75 #define WARMBOOT_SEG (KERNBASE + 0x0469)
77 #define CMOS_REG (0x70)
78 #define CMOS_DATA (0x71)
79 #define BIOS_RESET (0x0f)
80 #define BIOS_WARM (0x0a)
82 /* lock region used by kernel profiling */
85 int mp_naps; /* # of Applications processors */
86 int boot_cpu_id = -1; /* designated BSP */
88 extern struct pcpu __pcpu[];
90 /* AP uses this during bootstrap. Do not staticize. */
94 /* Free these after use */
95 void *bootstacks[MAXCPU];
97 /* Temporary variables for init_secondary() */
98 char *doublefault_stack;
102 struct pcb stoppcbs[MAXCPU];
103 struct pcb **susppcbs;
106 /* Variables needed for SMP tlb shootdown. */
107 vm_offset_t smp_tlb_addr1;
108 vm_offset_t smp_tlb_addr2;
109 volatile int smp_tlb_wait;
112 /* Interrupt counts. */
113 static u_long *ipi_preempt_counts[MAXCPU];
114 static u_long *ipi_ast_counts[MAXCPU];
115 u_long *ipi_invltlb_counts[MAXCPU];
116 u_long *ipi_invlrng_counts[MAXCPU];
117 u_long *ipi_invlpg_counts[MAXCPU];
118 u_long *ipi_invlcache_counts[MAXCPU];
119 u_long *ipi_rendezvous_counts[MAXCPU];
120 static u_long *ipi_hardclock_counts[MAXCPU];
123 extern inthand_t IDTVEC(fast_syscall), IDTVEC(fast_syscall32);
126 * Local data and functions.
129 static volatile cpuset_t ipi_nmi_pending;
131 /* used to hold the AP's until we are ready to release them */
132 static struct mtx ap_boot_mtx;
134 /* Set to 1 once we're ready to let the APs out of the pen. */
135 static volatile int aps_ready = 0;
138 * Store data from cpu_add() until later in the boot when we actually setup
145 int cpu_hyperthread:1;
146 } static cpu_info[MAX_APIC_ID + 1];
147 int cpu_apic_ids[MAXCPU];
148 int apic_cpuids[MAX_APIC_ID + 1];
150 /* Holds pending bitmap based IPIs per CPU */
151 static volatile u_int cpu_ipi_pending[MAXCPU];
153 static u_int boot_address;
154 static int cpu_logical; /* logical cpus per core */
155 static int cpu_cores; /* cores per package */
157 static void assign_cpu_ids(void);
158 static void set_interrupt_apic_ids(void);
159 static int start_all_aps(void);
160 static int start_ap(int apic_id);
161 static void release_aps(void *dummy);
163 static u_int hyperthreading_cpus; /* logical cpus sharing L1 cache */
164 static int hyperthreading_allowed = 1;
165 static u_int bootMP_size;
168 mem_range_AP_init(void)
170 if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP)
171 mem_range_softc.mr_op->initAP(&mem_range_softc);
180 /* AMD processors do not support HTT. */
183 if ((amd_feature2 & AMDID2_CMP) == 0) {
188 core_id_bits = (cpu_procinfo2 & AMDID_COREID_SIZE) >>
189 AMDID_COREID_SIZE_SHIFT;
190 if (core_id_bits == 0) {
191 cpu_cores = (cpu_procinfo2 & AMDID_CMP_CORES) + 1;
195 /* Fam 10h and newer should get here. */
196 for (id = 0; id <= MAX_APIC_ID; id++) {
197 /* Check logical CPU availability. */
198 if (!cpu_info[id].cpu_present || cpu_info[id].cpu_disabled)
200 /* Check if logical CPU has the same package ID. */
201 if ((id >> core_id_bits) != (boot_cpu_id >> core_id_bits))
208 * Round up to the next power of two, if necessary, and then
210 * Returns -1 if argument is zero.
216 return (fls(x << (1 - powerof2(x))) - 1);
229 /* Both zero and one here mean one logical processor per package. */
230 max_logical = (cpu_feature & CPUID_HTT) != 0 ?
231 (cpu_procinfo & CPUID_HTT_CORES) >> 16 : 1;
232 if (max_logical <= 1)
236 * Because of uniformity assumption we examine only
237 * those logical processors that belong to the same
238 * package as BSP. Further, we count number of
239 * logical processors that belong to the same core
240 * as BSP thus deducing number of threads per core.
242 if (cpu_high >= 0x4) {
243 cpuid_count(0x04, 0, p);
244 max_cores = ((p[0] >> 26) & 0x3f) + 1;
247 core_id_bits = mask_width(max_logical/max_cores);
248 if (core_id_bits < 0)
250 pkg_id_bits = core_id_bits + mask_width(max_cores);
252 for (id = 0; id <= MAX_APIC_ID; id++) {
253 /* Check logical CPU availability. */
254 if (!cpu_info[id].cpu_present || cpu_info[id].cpu_disabled)
256 /* Check if logical CPU has the same package ID. */
257 if ((id >> pkg_id_bits) != (boot_cpu_id >> pkg_id_bits))
260 /* Check if logical CPU has the same package and core IDs. */
261 if ((id >> core_id_bits) == (boot_cpu_id >> core_id_bits))
265 KASSERT(cpu_cores >= 1 && cpu_logical >= 1,
266 ("topo_probe_0x4 couldn't find BSP"));
268 cpu_cores /= cpu_logical;
269 hyperthreading_cpus = cpu_logical;
283 /* We only support three levels for now. */
284 for (i = 0; i < 3; i++) {
285 cpuid_count(0x0b, i, p);
287 /* Fall back if CPU leaf 11 doesn't really exist. */
288 if (i == 0 && p[1] == 0) {
294 logical = p[1] &= 0xffff;
295 type = (p[2] >> 8) & 0xff;
296 if (type == 0 || logical == 0)
299 * Because of uniformity assumption we examine only
300 * those logical processors that belong to the same
303 for (cnt = 0, x = 0; x <= MAX_APIC_ID; x++) {
304 if (!cpu_info[x].cpu_present ||
305 cpu_info[x].cpu_disabled)
307 if (x >> bits == boot_cpu_id >> bits)
310 if (type == CPUID_TYPE_SMT)
312 else if (type == CPUID_TYPE_CORE)
315 if (cpu_logical == 0)
317 cpu_cores /= cpu_logical;
321 * Both topology discovery code and code that consumes topology
322 * information assume top-down uniformity of the topology.
323 * That is, all physical packages must be identical and each
324 * core in a package must have the same number of threads.
325 * Topology information is queried only on BSP, on which this
326 * code runs and for which it can query CPUID information.
327 * Then topology is extrapolated on all packages using the
328 * uniformity assumption.
333 static int cpu_topo_probed = 0;
338 CPU_ZERO(&logical_cpus_mask);
340 cpu_cores = cpu_logical = 1;
341 else if (cpu_vendor_id == CPU_VENDOR_AMD)
343 else if (cpu_vendor_id == CPU_VENDOR_INTEL) {
345 * See Intel(R) 64 Architecture Processor
346 * Topology Enumeration article for details.
348 * Note that 0x1 <= cpu_high < 4 case should be
349 * compatible with topo_probe_0x4() logic when
350 * CPUID.1:EBX[23:16] > 0 (cpu_cores will be 1)
351 * or it should trigger the fallback otherwise.
355 else if (cpu_high >= 0x1)
360 * Fallback: assume each logical CPU is in separate
361 * physical package. That is, no multi-core, no SMT.
363 if (cpu_cores == 0 || cpu_logical == 0)
364 cpu_cores = cpu_logical = 1;
374 * Determine whether any threading flags are
378 if (cpu_logical > 1 && hyperthreading_cpus)
379 cg_flags = CG_FLAG_HTT;
380 else if (cpu_logical > 1)
381 cg_flags = CG_FLAG_SMT;
384 if (mp_ncpus % (cpu_cores * cpu_logical) != 0) {
385 printf("WARNING: Non-uniform processors.\n");
386 printf("WARNING: Using suboptimal topology.\n");
387 return (smp_topo_none());
390 * No multi-core or hyper-threaded.
392 if (cpu_logical * cpu_cores == 1)
393 return (smp_topo_none());
395 * Only HTT no multi-core.
397 if (cpu_logical > 1 && cpu_cores == 1)
398 return (smp_topo_1level(CG_SHARE_L1, cpu_logical, cg_flags));
400 * Only multi-core no HTT.
402 if (cpu_cores > 1 && cpu_logical == 1)
403 return (smp_topo_1level(CG_SHARE_L2, cpu_cores, cg_flags));
405 * Both HTT and multi-core.
407 return (smp_topo_2level(CG_SHARE_L2, cpu_cores,
408 CG_SHARE_L1, cpu_logical, cg_flags));
412 * Calculate usable address in base memory for AP trampoline code.
415 mp_bootaddress(u_int basemem)
418 bootMP_size = mptramp_end - mptramp_start;
419 boot_address = trunc_page(basemem * 1024); /* round down to 4k boundary */
420 if (((basemem * 1024) - boot_address) < bootMP_size)
421 boot_address -= PAGE_SIZE; /* not enough, lower by 4k */
422 /* 3 levels of page table pages */
423 mptramp_pagetables = boot_address - (PAGE_SIZE * 3);
425 return mptramp_pagetables;
429 cpu_add(u_int apic_id, char boot_cpu)
432 if (apic_id > MAX_APIC_ID) {
433 panic("SMP: APIC ID %d too high", apic_id);
436 KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %d added twice",
438 cpu_info[apic_id].cpu_present = 1;
440 KASSERT(boot_cpu_id == -1,
441 ("CPU %d claims to be BSP, but CPU %d already is", apic_id,
443 boot_cpu_id = apic_id;
444 cpu_info[apic_id].cpu_bsp = 1;
446 if (mp_ncpus < MAXCPU) {
448 mp_maxid = mp_ncpus - 1;
451 printf("SMP: Added CPU %d (%s)\n", apic_id, boot_cpu ? "BSP" :
456 cpu_mp_setmaxid(void)
460 * mp_maxid should be already set by calls to cpu_add().
461 * Just sanity check its value here.
464 KASSERT(mp_maxid == 0,
465 ("%s: mp_ncpus is zero, but mp_maxid is not", __func__));
466 else if (mp_ncpus == 1)
469 KASSERT(mp_maxid >= mp_ncpus - 1,
470 ("%s: counters out of sync: max %d, count %d", __func__,
471 mp_maxid, mp_ncpus));
479 * Always record BSP in CPU map so that the mbuf init code works
482 CPU_SETOF(0, &all_cpus);
485 * No CPUs were found, so this must be a UP system. Setup
486 * the variables to represent a system with a single CPU
493 /* At least one CPU was found. */
496 * One CPU was found, so this must be a UP system with
503 /* At least two CPUs were found. */
508 * Initialize the IPI handlers and start up the AP's.
515 /* Initialize the logical ID to APIC ID table. */
516 for (i = 0; i < MAXCPU; i++) {
517 cpu_apic_ids[i] = -1;
518 cpu_ipi_pending[i] = 0;
521 /* Install an inter-CPU IPI for TLB invalidation */
522 setidt(IPI_INVLTLB, IDTVEC(invltlb), SDT_SYSIGT, SEL_KPL, 0);
523 setidt(IPI_INVLPG, IDTVEC(invlpg), SDT_SYSIGT, SEL_KPL, 0);
524 setidt(IPI_INVLRNG, IDTVEC(invlrng), SDT_SYSIGT, SEL_KPL, 0);
526 /* Install an inter-CPU IPI for cache invalidation. */
527 setidt(IPI_INVLCACHE, IDTVEC(invlcache), SDT_SYSIGT, SEL_KPL, 0);
529 /* Install an inter-CPU IPI for all-CPU rendezvous */
530 setidt(IPI_RENDEZVOUS, IDTVEC(rendezvous), SDT_SYSIGT, SEL_KPL, 0);
532 /* Install generic inter-CPU IPI handler */
533 setidt(IPI_BITMAP_VECTOR, IDTVEC(ipi_intr_bitmap_handler),
534 SDT_SYSIGT, SEL_KPL, 0);
536 /* Install an inter-CPU IPI for CPU stop/restart */
537 setidt(IPI_STOP, IDTVEC(cpustop), SDT_SYSIGT, SEL_KPL, 0);
539 /* Install an inter-CPU IPI for CPU suspend/resume */
540 setidt(IPI_SUSPEND, IDTVEC(cpususpend), SDT_SYSIGT, SEL_KPL, 0);
542 /* Set boot_cpu_id if needed. */
543 if (boot_cpu_id == -1) {
544 boot_cpu_id = PCPU_GET(apic_id);
545 cpu_info[boot_cpu_id].cpu_bsp = 1;
547 KASSERT(boot_cpu_id == PCPU_GET(apic_id),
548 ("BSP's APIC ID doesn't match boot_cpu_id"));
550 /* Probe logical/physical core configuration. */
555 /* Start each Application Processor */
558 set_interrupt_apic_ids();
563 * Print various information about the SMP system hardware and setup.
566 cpu_mp_announce(void)
568 const char *hyperthread;
571 printf("FreeBSD/SMP: %d package(s) x %d core(s)",
572 mp_ncpus / (cpu_cores * cpu_logical), cpu_cores);
573 if (hyperthreading_cpus > 1)
574 printf(" x %d HTT threads", cpu_logical);
575 else if (cpu_logical > 1)
576 printf(" x %d SMT threads", cpu_logical);
579 /* List active CPUs first. */
580 printf(" cpu0 (BSP): APIC ID: %2d\n", boot_cpu_id);
581 for (i = 1; i < mp_ncpus; i++) {
582 if (cpu_info[cpu_apic_ids[i]].cpu_hyperthread)
586 printf(" cpu%d (AP%s): APIC ID: %2d\n", i, hyperthread,
590 /* List disabled CPUs last. */
591 for (i = 0; i <= MAX_APIC_ID; i++) {
592 if (!cpu_info[i].cpu_present || !cpu_info[i].cpu_disabled)
594 if (cpu_info[i].cpu_hyperthread)
598 printf(" cpu (AP%s): APIC ID: %2d (disabled)\n", hyperthread,
604 * AP CPU's call this to initialize themselves.
613 int cpu, gsel_tss, x;
614 struct region_descriptor ap_gdt;
616 /* Set by the startup code for us to use */
620 common_tss[cpu] = common_tss[0];
621 common_tss[cpu].tss_rsp0 = 0; /* not used until after switch */
622 common_tss[cpu].tss_iobase = sizeof(struct amd64tss) +
624 common_tss[cpu].tss_ist1 = (long)&doublefault_stack[PAGE_SIZE];
626 /* The NMI stack runs on IST2. */
627 np = ((struct nmi_pcpu *) &nmi_stack[PAGE_SIZE]) - 1;
628 common_tss[cpu].tss_ist2 = (long) np;
630 /* Prepare private GDT */
631 gdt_segs[GPROC0_SEL].ssd_base = (long) &common_tss[cpu];
632 for (x = 0; x < NGDT; x++) {
633 if (x != GPROC0_SEL && x != (GPROC0_SEL + 1) &&
634 x != GUSERLDT_SEL && x != (GUSERLDT_SEL + 1))
635 ssdtosd(&gdt_segs[x], &gdt[NGDT * cpu + x]);
637 ssdtosyssd(&gdt_segs[GPROC0_SEL],
638 (struct system_segment_descriptor *)&gdt[NGDT * cpu + GPROC0_SEL]);
639 ap_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
640 ap_gdt.rd_base = (long) &gdt[NGDT * cpu];
641 lgdt(&ap_gdt); /* does magic intra-segment return */
643 /* Get per-cpu data */
646 /* prime data page for it to use */
647 pcpu_init(pc, cpu, sizeof(struct pcpu));
648 dpcpu_init(dpcpu, cpu);
649 pc->pc_apic_id = cpu_apic_ids[cpu];
650 pc->pc_prvspace = pc;
651 pc->pc_curthread = 0;
652 pc->pc_tssp = &common_tss[cpu];
653 pc->pc_commontssp = &common_tss[cpu];
655 pc->pc_tss = (struct system_segment_descriptor *)&gdt[NGDT * cpu +
657 pc->pc_fs32p = &gdt[NGDT * cpu + GUFS32_SEL];
658 pc->pc_gs32p = &gdt[NGDT * cpu + GUGS32_SEL];
659 pc->pc_ldt = (struct system_segment_descriptor *)&gdt[NGDT * cpu +
662 /* Save the per-cpu pointer for use by the NMI handler. */
663 np->np_pcpu = (register_t) pc;
665 wrmsr(MSR_FSBASE, 0); /* User value */
666 wrmsr(MSR_GSBASE, (u_int64_t)pc);
667 wrmsr(MSR_KGSBASE, (u_int64_t)pc); /* XXX User value while we're in the kernel */
671 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
675 * Set to a known state:
676 * Set by mpboot.s: CR0_PG, CR0_PE
677 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
680 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
683 /* Set up the fast syscall stuff */
684 msr = rdmsr(MSR_EFER) | EFER_SCE;
685 wrmsr(MSR_EFER, msr);
686 wrmsr(MSR_LSTAR, (u_int64_t)IDTVEC(fast_syscall));
687 wrmsr(MSR_CSTAR, (u_int64_t)IDTVEC(fast_syscall32));
688 msr = ((u_int64_t)GSEL(GCODE_SEL, SEL_KPL) << 32) |
689 ((u_int64_t)GSEL(GUCODE32_SEL, SEL_UPL) << 48);
690 wrmsr(MSR_STAR, msr);
691 wrmsr(MSR_SF_MASK, PSL_NT|PSL_T|PSL_I|PSL_C|PSL_D);
693 /* Disable local APIC just to be sure. */
696 /* signal our startup to the BSP. */
699 /* Spin until the BSP releases the AP's. */
703 /* Initialize the PAT MSR. */
706 /* set up CPU registers and state */
709 /* set up SSE/NX registers */
712 /* set up FPU state on the AP */
715 /* A quick check from sanity claus */
716 cpuid = PCPU_GET(cpuid);
717 if (PCPU_GET(apic_id) != lapic_id()) {
718 printf("SMP: cpuid = %d\n", cpuid);
719 printf("SMP: actual apic_id = %d\n", lapic_id());
720 printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id));
721 panic("cpuid mismatch! boom!!");
724 /* Initialize curthread. */
725 KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread"));
726 PCPU_SET(curthread, PCPU_GET(idlethread));
730 mtx_lock_spin(&ap_boot_mtx);
732 /* Init local apic for irq's */
735 /* Set memory range attributes for this CPU to match the BSP */
740 CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid);
741 printf("SMP: AP CPU #%d Launched!\n", cpuid);
743 /* Determine if we are a logical CPU. */
744 /* XXX Calculation depends on cpu_logical being a power of 2, e.g. 2 */
745 if (cpu_logical > 1 && PCPU_GET(apic_id) % cpu_logical != 0)
746 CPU_SET(cpuid, &logical_cpus_mask);
751 if (smp_cpus == mp_ncpus) {
752 /* enable IPI's, tlb shootdown, freezes etc */
753 atomic_store_rel_int(&smp_started, 1);
754 smp_active = 1; /* historic */
758 * Enable global pages TLB extension
759 * This also implicitly flushes the TLB
762 load_cr4(rcr4() | CR4_PGE);
766 mtx_unlock_spin(&ap_boot_mtx);
768 /* Wait until all the AP's are up. */
769 while (smp_started == 0)
772 /* Start per-CPU event timers. */
777 panic("scheduler returned us to %s", __func__);
781 /*******************************************************************
782 * local functions and data
786 * We tell the I/O APIC code about all the CPUs we want to receive
787 * interrupts. If we don't want certain CPUs to receive IRQs we
788 * can simply not tell the I/O APIC code about them in this function.
791 set_interrupt_apic_ids(void)
795 for (i = 0; i < MAXCPU; i++) {
796 apic_id = cpu_apic_ids[i];
799 if (cpu_info[apic_id].cpu_disabled)
802 /* Don't let hyperthreads service interrupts. */
803 if (hyperthreading_cpus > 1 &&
804 apic_id % hyperthreading_cpus != 0)
812 * Assign logical CPU IDs to local APICs.
819 TUNABLE_INT_FETCH("machdep.hyperthreading_allowed",
820 &hyperthreading_allowed);
822 /* Check for explicitly disabled CPUs. */
823 for (i = 0; i <= MAX_APIC_ID; i++) {
824 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp)
827 if (hyperthreading_cpus > 1 && i % hyperthreading_cpus != 0) {
828 cpu_info[i].cpu_hyperthread = 1;
831 * Don't use HT CPU if it has been disabled by a
834 if (hyperthreading_allowed == 0) {
835 cpu_info[i].cpu_disabled = 1;
840 /* Don't use this CPU if it has been disabled by a tunable. */
841 if (resource_disabled("lapic", i)) {
842 cpu_info[i].cpu_disabled = 1;
847 if (hyperthreading_allowed == 0 && hyperthreading_cpus > 1) {
848 hyperthreading_cpus = 0;
853 * Assign CPU IDs to local APIC IDs and disable any CPUs
854 * beyond MAXCPU. CPU 0 is always assigned to the BSP.
856 * To minimize confusion for userland, we attempt to number
857 * CPUs such that all threads and cores in a package are
858 * grouped together. For now we assume that the BSP is always
859 * the first thread in a package and just start adding APs
860 * starting with the BSP's APIC ID.
863 cpu_apic_ids[0] = boot_cpu_id;
864 apic_cpuids[boot_cpu_id] = 0;
865 for (i = boot_cpu_id + 1; i != boot_cpu_id;
866 i == MAX_APIC_ID ? i = 0 : i++) {
867 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp ||
868 cpu_info[i].cpu_disabled)
871 if (mp_ncpus < MAXCPU) {
872 cpu_apic_ids[mp_ncpus] = i;
873 apic_cpuids[i] = mp_ncpus;
876 cpu_info[i].cpu_disabled = 1;
878 KASSERT(mp_maxid >= mp_ncpus - 1,
879 ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid,
884 * start each AP in our list
889 vm_offset_t va = boot_address + KERNBASE;
890 u_int64_t *pt4, *pt3, *pt2;
891 u_int32_t mpbioswarmvec;
895 mtx_init(&ap_boot_mtx, "ap boot", NULL, MTX_SPIN);
897 /* install the AP 1st level boot code */
898 pmap_kenter(va, boot_address);
899 pmap_invalidate_page(kernel_pmap, va);
900 bcopy(mptramp_start, (void *)va, bootMP_size);
902 /* Locate the page tables, they'll be below the trampoline */
903 pt4 = (u_int64_t *)(uintptr_t)(mptramp_pagetables + KERNBASE);
904 pt3 = pt4 + (PAGE_SIZE) / sizeof(u_int64_t);
905 pt2 = pt3 + (PAGE_SIZE) / sizeof(u_int64_t);
907 /* Create the initial 1GB replicated page tables */
908 for (i = 0; i < 512; i++) {
909 /* Each slot of the level 4 pages points to the same level 3 page */
910 pt4[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + PAGE_SIZE);
911 pt4[i] |= PG_V | PG_RW | PG_U;
913 /* Each slot of the level 3 pages points to the same level 2 page */
914 pt3[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + (2 * PAGE_SIZE));
915 pt3[i] |= PG_V | PG_RW | PG_U;
917 /* The level 2 page slots are mapped with 2MB pages for 1GB. */
918 pt2[i] = i * (2 * 1024 * 1024);
919 pt2[i] |= PG_V | PG_RW | PG_PS | PG_U;
922 /* save the current value of the warm-start vector */
923 mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF);
924 outb(CMOS_REG, BIOS_RESET);
925 mpbiosreason = inb(CMOS_DATA);
927 /* setup a vector to our boot code */
928 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
929 *((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4);
930 outb(CMOS_REG, BIOS_RESET);
931 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
934 for (cpu = 1; cpu < mp_ncpus; cpu++) {
935 apic_id = cpu_apic_ids[cpu];
937 /* allocate and set up an idle stack data page */
938 bootstacks[cpu] = (void *)kmem_alloc(kernel_map, KSTACK_PAGES * PAGE_SIZE);
939 doublefault_stack = (char *)kmem_alloc(kernel_map, PAGE_SIZE);
940 nmi_stack = (char *)kmem_alloc(kernel_map, PAGE_SIZE);
941 dpcpu = (void *)kmem_alloc(kernel_map, DPCPU_SIZE);
943 bootSTK = (char *)bootstacks[cpu] + KSTACK_PAGES * PAGE_SIZE - 8;
946 /* attempt to start the Application Processor */
947 if (!start_ap(apic_id)) {
948 /* restore the warmstart vector */
949 *(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec;
950 panic("AP #%d (PHY# %d) failed!", cpu, apic_id);
953 CPU_SET(cpu, &all_cpus); /* record AP in CPU map */
956 /* restore the warmstart vector */
957 *(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec;
959 outb(CMOS_REG, BIOS_RESET);
960 outb(CMOS_DATA, mpbiosreason);
962 /* number of APs actually started */
968 * This function starts the AP (application processor) identified
969 * by the APIC ID 'physicalCpu'. It does quite a "song and dance"
970 * to accomplish this. This is necessary because of the nuances
971 * of the different hardware we might encounter. It isn't pretty,
972 * but it seems to work.
975 start_ap(int apic_id)
980 /* calculate the vector */
981 vector = (boot_address >> 12) & 0xff;
983 /* used as a watchpoint to signal AP startup */
987 * first we do an INIT/RESET IPI this INIT IPI might be run, reseting
988 * and running the target CPU. OR this INIT IPI might be latched (P5
989 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
993 /* do an INIT IPI: assert RESET */
994 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
995 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id);
997 /* wait for pending status end */
1000 /* do an INIT IPI: deassert RESET */
1001 lapic_ipi_raw(APIC_DEST_ALLESELF | APIC_TRIGMOD_LEVEL |
1002 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, 0);
1004 /* wait for pending status end */
1005 DELAY(10000); /* wait ~10mS */
1009 * next we do a STARTUP IPI: the previous INIT IPI might still be
1010 * latched, (P5 bug) this 1st STARTUP would then terminate
1011 * immediately, and the previously started INIT IPI would continue. OR
1012 * the previous INIT IPI has already run. and this STARTUP IPI will
1013 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
1017 /* do a STARTUP IPI */
1018 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1019 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1022 DELAY(200); /* wait ~200uS */
1025 * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
1026 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
1027 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
1028 * recognized after hardware RESET or INIT IPI.
1031 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1032 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1035 DELAY(200); /* wait ~200uS */
1037 /* Wait up to 5 seconds for it to start. */
1038 for (ms = 0; ms < 5000; ms++) {
1040 return 1; /* return SUCCESS */
1043 return 0; /* return FAILURE */
1046 #ifdef COUNT_XINVLTLB_HITS
1047 u_int xhits_gbl[MAXCPU];
1048 u_int xhits_pg[MAXCPU];
1049 u_int xhits_rng[MAXCPU];
1050 static SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW, 0, "");
1051 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl,
1052 sizeof(xhits_gbl), "IU", "");
1053 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg,
1054 sizeof(xhits_pg), "IU", "");
1055 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng,
1056 sizeof(xhits_rng), "IU", "");
1061 u_int ipi_range_size;
1062 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, "");
1063 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, "");
1064 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, "");
1065 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW,
1066 &ipi_range_size, 0, "");
1068 u_int ipi_masked_global;
1069 u_int ipi_masked_page;
1070 u_int ipi_masked_range;
1071 u_int ipi_masked_range_size;
1072 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_masked_global, CTLFLAG_RW,
1073 &ipi_masked_global, 0, "");
1074 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_masked_page, CTLFLAG_RW,
1075 &ipi_masked_page, 0, "");
1076 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_masked_range, CTLFLAG_RW,
1077 &ipi_masked_range, 0, "");
1078 SYSCTL_UINT(_debug_xhits, OID_AUTO, ipi_masked_range_size, CTLFLAG_RW,
1079 &ipi_masked_range_size, 0, "");
1080 #endif /* COUNT_XINVLTLB_HITS */
1083 * Send an IPI to specified CPU handling the bitmap logic.
1086 ipi_send_cpu(int cpu, u_int ipi)
1088 u_int bitmap, old_pending, new_pending;
1090 KASSERT(cpu_apic_ids[cpu] != -1, ("IPI to non-existent CPU %d", cpu));
1092 if (IPI_IS_BITMAPED(ipi)) {
1094 ipi = IPI_BITMAP_VECTOR;
1096 old_pending = cpu_ipi_pending[cpu];
1097 new_pending = old_pending | bitmap;
1098 } while (!atomic_cmpset_int(&cpu_ipi_pending[cpu],
1099 old_pending, new_pending));
1103 lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]);
1107 * Flush the TLB on all other CPU's
1110 smp_tlb_shootdown(u_int vector, vm_offset_t addr1, vm_offset_t addr2)
1114 ncpu = mp_ncpus - 1; /* does not shootdown self */
1116 return; /* no other cpus */
1117 if (!(read_rflags() & PSL_I))
1118 panic("%s: interrupts disabled", __func__);
1119 mtx_lock_spin(&smp_ipi_mtx);
1120 smp_tlb_addr1 = addr1;
1121 smp_tlb_addr2 = addr2;
1122 atomic_store_rel_int(&smp_tlb_wait, 0);
1123 ipi_all_but_self(vector);
1124 while (smp_tlb_wait < ncpu)
1126 mtx_unlock_spin(&smp_ipi_mtx);
1130 smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, vm_offset_t addr1, vm_offset_t addr2)
1132 int cpu, ncpu, othercpus;
1134 othercpus = mp_ncpus - 1;
1135 if (CPU_ISFULLSET(&mask)) {
1139 CPU_CLR(PCPU_GET(cpuid), &mask);
1140 if (CPU_EMPTY(&mask))
1143 if (!(read_rflags() & PSL_I))
1144 panic("%s: interrupts disabled", __func__);
1145 mtx_lock_spin(&smp_ipi_mtx);
1146 smp_tlb_addr1 = addr1;
1147 smp_tlb_addr2 = addr2;
1148 atomic_store_rel_int(&smp_tlb_wait, 0);
1149 if (CPU_ISFULLSET(&mask)) {
1151 ipi_all_but_self(vector);
1154 while ((cpu = cpusetobj_ffs(&mask)) != 0) {
1156 CPU_CLR(cpu, &mask);
1157 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__,
1159 ipi_send_cpu(cpu, vector);
1163 while (smp_tlb_wait < ncpu)
1165 mtx_unlock_spin(&smp_ipi_mtx);
1169 smp_cache_flush(void)
1173 smp_tlb_shootdown(IPI_INVLCACHE, 0, 0);
1181 smp_tlb_shootdown(IPI_INVLTLB, 0, 0);
1182 #ifdef COUNT_XINVLTLB_HITS
1189 smp_invlpg(vm_offset_t addr)
1193 smp_tlb_shootdown(IPI_INVLPG, addr, 0);
1194 #ifdef COUNT_XINVLTLB_HITS
1201 smp_invlpg_range(vm_offset_t addr1, vm_offset_t addr2)
1205 smp_tlb_shootdown(IPI_INVLRNG, addr1, addr2);
1206 #ifdef COUNT_XINVLTLB_HITS
1208 ipi_range_size += (addr2 - addr1) / PAGE_SIZE;
1214 smp_masked_invltlb(cpuset_t mask)
1218 smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, 0, 0);
1219 #ifdef COUNT_XINVLTLB_HITS
1220 ipi_masked_global++;
1226 smp_masked_invlpg(cpuset_t mask, vm_offset_t addr)
1230 smp_targeted_tlb_shootdown(mask, IPI_INVLPG, addr, 0);
1231 #ifdef COUNT_XINVLTLB_HITS
1238 smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2)
1242 smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, addr1, addr2);
1243 #ifdef COUNT_XINVLTLB_HITS
1245 ipi_masked_range_size += (addr2 - addr1) / PAGE_SIZE;
1251 ipi_bitmap_handler(struct trapframe frame)
1253 struct trapframe *oldframe;
1255 int cpu = PCPU_GET(cpuid);
1260 td->td_intr_nesting_level++;
1261 oldframe = td->td_intr_frame;
1262 td->td_intr_frame = &frame;
1263 ipi_bitmap = atomic_readandclear_int(&cpu_ipi_pending[cpu]);
1264 if (ipi_bitmap & (1 << IPI_PREEMPT)) {
1266 (*ipi_preempt_counts[cpu])++;
1270 if (ipi_bitmap & (1 << IPI_AST)) {
1272 (*ipi_ast_counts[cpu])++;
1274 /* Nothing to do for AST */
1276 if (ipi_bitmap & (1 << IPI_HARDCLOCK)) {
1278 (*ipi_hardclock_counts[cpu])++;
1282 td->td_intr_frame = oldframe;
1283 td->td_intr_nesting_level--;
1288 * send an IPI to a set of cpus.
1291 ipi_selected(cpuset_t cpus, u_int ipi)
1296 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1297 * of help in order to understand what is the source.
1298 * Set the mask of receiving CPUs for this purpose.
1300 if (ipi == IPI_STOP_HARD)
1301 CPU_OR_ATOMIC(&ipi_nmi_pending, &cpus);
1303 while ((cpu = cpusetobj_ffs(&cpus)) != 0) {
1305 CPU_CLR(cpu, &cpus);
1306 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1307 ipi_send_cpu(cpu, ipi);
1312 * send an IPI to a specific CPU.
1315 ipi_cpu(int cpu, u_int ipi)
1319 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1320 * of help in order to understand what is the source.
1321 * Set the mask of receiving CPUs for this purpose.
1323 if (ipi == IPI_STOP_HARD)
1324 CPU_SET_ATOMIC(cpu, &ipi_nmi_pending);
1326 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1327 ipi_send_cpu(cpu, ipi);
1331 * send an IPI to all CPUs EXCEPT myself
1334 ipi_all_but_self(u_int ipi)
1336 cpuset_t other_cpus;
1338 other_cpus = all_cpus;
1339 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1341 if (IPI_IS_BITMAPED(ipi)) {
1342 ipi_selected(other_cpus, ipi);
1347 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1348 * of help in order to understand what is the source.
1349 * Set the mask of receiving CPUs for this purpose.
1351 if (ipi == IPI_STOP_HARD)
1352 CPU_OR_ATOMIC(&ipi_nmi_pending, &other_cpus);
1354 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1355 lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS);
1364 * As long as there is not a simple way to know about a NMI's
1365 * source, if the bitmask for the current CPU is present in
1366 * the global pending bitword an IPI_STOP_HARD has been issued
1367 * and should be handled.
1369 cpuid = PCPU_GET(cpuid);
1370 if (!CPU_ISSET(cpuid, &ipi_nmi_pending))
1373 CPU_CLR_ATOMIC(cpuid, &ipi_nmi_pending);
1379 * Handle an IPI_STOP by saving our current context and spinning until we
1383 cpustop_handler(void)
1387 cpu = PCPU_GET(cpuid);
1389 savectx(&stoppcbs[cpu]);
1391 /* Indicate that we are stopped */
1392 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1394 /* Wait for restart */
1395 while (!CPU_ISSET(cpu, &started_cpus))
1398 CPU_CLR_ATOMIC(cpu, &started_cpus);
1399 CPU_CLR_ATOMIC(cpu, &stopped_cpus);
1402 amd64_db_resume_dbreg();
1405 if (cpu == 0 && cpustop_restartfunc != NULL) {
1406 cpustop_restartfunc();
1407 cpustop_restartfunc = NULL;
1412 * Handle an IPI_SUSPEND by saving our current context and spinning until we
1416 cpususpend_handler(void)
1420 cpu = PCPU_GET(cpuid);
1422 if (savectx(susppcbs[cpu])) {
1423 ctx_fpusave(suspfpusave[cpu]);
1425 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1428 load_cr3(susppcbs[cpu]->pcb_cr3);
1430 PCPU_SET(switchtime, 0);
1431 PCPU_SET(switchticks, ticks);
1434 /* Wait for resume */
1435 while (!CPU_ISSET(cpu, &started_cpus))
1438 /* Resume MCA and local APIC */
1442 CPU_CLR_ATOMIC(cpu, &started_cpus);
1443 CPU_CLR_ATOMIC(cpu, &stopped_cpus);
1447 * This is called once the rest of the system is up and running and we're
1448 * ready to let the AP's out of the pen.
1451 release_aps(void *dummy __unused)
1456 atomic_store_rel_int(&aps_ready, 1);
1457 while (smp_started == 0)
1460 SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL);
1464 * Setup interrupt counters for IPI handlers.
1467 mp_ipi_intrcnt(void *dummy)
1473 snprintf(buf, sizeof(buf), "cpu%d:invltlb", i);
1474 intrcnt_add(buf, &ipi_invltlb_counts[i]);
1475 snprintf(buf, sizeof(buf), "cpu%d:invlrng", i);
1476 intrcnt_add(buf, &ipi_invlrng_counts[i]);
1477 snprintf(buf, sizeof(buf), "cpu%d:invlpg", i);
1478 intrcnt_add(buf, &ipi_invlpg_counts[i]);
1479 snprintf(buf, sizeof(buf), "cpu%d:invlcache", i);
1480 intrcnt_add(buf, &ipi_invlcache_counts[i]);
1481 snprintf(buf, sizeof(buf), "cpu%d:preempt", i);
1482 intrcnt_add(buf, &ipi_preempt_counts[i]);
1483 snprintf(buf, sizeof(buf), "cpu%d:ast", i);
1484 intrcnt_add(buf, &ipi_ast_counts[i]);
1485 snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i);
1486 intrcnt_add(buf, &ipi_rendezvous_counts[i]);
1487 snprintf(buf, sizeof(buf), "cpu%d:hardclock", i);
1488 intrcnt_add(buf, &ipi_hardclock_counts[i]);
1491 SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL);