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
26 #include <sys/cdefs.h>
27 __FBSDID("$FreeBSD$");
31 #include "opt_kstack_pages.h"
33 #include "opt_sched.h"
38 #error How did you get here?
42 #error The apic device is required for SMP, add "device apic" to your config file.
44 #if defined(CPU_DISABLE_CMPXCHG) && !defined(COMPILING_LINT)
45 #error SMP not supported with CPU_DISABLE_CMPXCHG
49 #include <sys/param.h>
50 #include <sys/systm.h>
52 #include <sys/cons.h> /* cngetc() */
53 #include <sys/cpuset.h>
57 #include <sys/kernel.h>
60 #include <sys/malloc.h>
61 #include <sys/memrange.h>
62 #include <sys/mutex.h>
65 #include <sys/sched.h>
67 #include <sys/sysctl.h>
70 #include <vm/vm_param.h>
72 #include <vm/vm_kern.h>
73 #include <vm/vm_extern.h>
75 #include <x86/apicreg.h>
76 #include <machine/clock.h>
77 #include <machine/cputypes.h>
79 #include <machine/md_var.h>
80 #include <machine/pcb.h>
81 #include <machine/psl.h>
82 #include <machine/smp.h>
83 #include <machine/specialreg.h>
85 #define WARMBOOT_TARGET 0
86 #define WARMBOOT_OFF (KERNBASE + 0x0467)
87 #define WARMBOOT_SEG (KERNBASE + 0x0469)
89 #define CMOS_REG (0x70)
90 #define CMOS_DATA (0x71)
91 #define BIOS_RESET (0x0f)
92 #define BIOS_WARM (0x0a)
95 * this code MUST be enabled here and in mpboot.s.
96 * it follows the very early stages of AP boot by placing values in CMOS ram.
97 * it NORMALLY will never be needed and thus the primitive method for enabling.
102 #if defined(CHECK_POINTS) && !defined(PC98)
103 #define CHECK_READ(A) (outb(CMOS_REG, (A)), inb(CMOS_DATA))
104 #define CHECK_WRITE(A,D) (outb(CMOS_REG, (A)), outb(CMOS_DATA, (D)))
106 #define CHECK_INIT(D); \
107 CHECK_WRITE(0x34, (D)); \
108 CHECK_WRITE(0x35, (D)); \
109 CHECK_WRITE(0x36, (D)); \
110 CHECK_WRITE(0x37, (D)); \
111 CHECK_WRITE(0x38, (D)); \
112 CHECK_WRITE(0x39, (D));
114 #define CHECK_PRINT(S); \
115 printf("%s: %d, %d, %d, %d, %d, %d\n", \
124 #else /* CHECK_POINTS */
126 #define CHECK_INIT(D)
127 #define CHECK_PRINT(S)
128 #define CHECK_WRITE(A, D)
130 #endif /* CHECK_POINTS */
132 /* lock region used by kernel profiling */
135 int mp_naps; /* # of Applications processors */
136 int boot_cpu_id = -1; /* designated BSP */
138 extern struct pcpu __pcpu[];
140 /* AP uses this during bootstrap. Do not staticize. */
144 /* Free these after use */
145 void *bootstacks[MAXCPU];
148 /* Hotwire a 0->4MB V==P mapping */
149 extern pt_entry_t *KPTphys;
151 struct pcb stoppcbs[MAXCPU];
152 struct pcb **susppcbs = NULL;
154 /* Variables needed for SMP tlb shootdown. */
155 vm_offset_t smp_tlb_addr1;
156 vm_offset_t smp_tlb_addr2;
157 volatile int smp_tlb_wait;
160 /* Interrupt counts. */
161 static u_long *ipi_preempt_counts[MAXCPU];
162 static u_long *ipi_ast_counts[MAXCPU];
163 u_long *ipi_invltlb_counts[MAXCPU];
164 u_long *ipi_invlrng_counts[MAXCPU];
165 u_long *ipi_invlpg_counts[MAXCPU];
166 u_long *ipi_invlcache_counts[MAXCPU];
167 u_long *ipi_rendezvous_counts[MAXCPU];
168 u_long *ipi_lazypmap_counts[MAXCPU];
169 static u_long *ipi_hardclock_counts[MAXCPU];
173 * Local data and functions.
176 static volatile cpuset_t ipi_nmi_pending;
178 /* used to hold the AP's until we are ready to release them */
179 static struct mtx ap_boot_mtx;
181 /* Set to 1 once we're ready to let the APs out of the pen. */
182 static volatile int aps_ready = 0;
185 * Store data from cpu_add() until later in the boot when we actually setup
192 int cpu_hyperthread:1;
193 } static cpu_info[MAX_APIC_ID + 1];
194 int cpu_apic_ids[MAXCPU];
195 int apic_cpuids[MAX_APIC_ID + 1];
197 /* Holds pending bitmap based IPIs per CPU */
198 static volatile u_int cpu_ipi_pending[MAXCPU];
200 static u_int boot_address;
201 static int cpu_logical; /* logical cpus per core */
202 static int cpu_cores; /* cores per package */
204 static void assign_cpu_ids(void);
205 static void install_ap_tramp(void);
206 static void set_interrupt_apic_ids(void);
207 static int start_all_aps(void);
208 static int start_ap(int apic_id);
209 static void release_aps(void *dummy);
211 static u_int hyperthreading_cpus; /* logical cpus sharing L1 cache */
212 static int hyperthreading_allowed = 1;
215 mem_range_AP_init(void)
217 if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP)
218 mem_range_softc.mr_op->initAP(&mem_range_softc);
227 /* AMD processors do not support HTT. */
230 if ((amd_feature2 & AMDID2_CMP) == 0) {
235 core_id_bits = (cpu_procinfo2 & AMDID_COREID_SIZE) >>
236 AMDID_COREID_SIZE_SHIFT;
237 if (core_id_bits == 0) {
238 cpu_cores = (cpu_procinfo2 & AMDID_CMP_CORES) + 1;
242 /* Fam 10h and newer should get here. */
243 for (id = 0; id <= MAX_APIC_ID; id++) {
244 /* Check logical CPU availability. */
245 if (!cpu_info[id].cpu_present || cpu_info[id].cpu_disabled)
247 /* Check if logical CPU has the same package ID. */
248 if ((id >> core_id_bits) != (boot_cpu_id >> core_id_bits))
255 * Round up to the next power of two, if necessary, and then
257 * Returns -1 if argument is zero.
263 return (fls(x << (1 - powerof2(x))) - 1);
276 /* Both zero and one here mean one logical processor per package. */
277 max_logical = (cpu_feature & CPUID_HTT) != 0 ?
278 (cpu_procinfo & CPUID_HTT_CORES) >> 16 : 1;
279 if (max_logical <= 1)
283 * Because of uniformity assumption we examine only
284 * those logical processors that belong to the same
285 * package as BSP. Further, we count number of
286 * logical processors that belong to the same core
287 * as BSP thus deducing number of threads per core.
289 if (cpu_high >= 0x4) {
290 cpuid_count(0x04, 0, p);
291 max_cores = ((p[0] >> 26) & 0x3f) + 1;
294 core_id_bits = mask_width(max_logical/max_cores);
295 if (core_id_bits < 0)
297 pkg_id_bits = core_id_bits + mask_width(max_cores);
299 for (id = 0; id <= MAX_APIC_ID; id++) {
300 /* Check logical CPU availability. */
301 if (!cpu_info[id].cpu_present || cpu_info[id].cpu_disabled)
303 /* Check if logical CPU has the same package ID. */
304 if ((id >> pkg_id_bits) != (boot_cpu_id >> pkg_id_bits))
307 /* Check if logical CPU has the same package and core IDs. */
308 if ((id >> core_id_bits) == (boot_cpu_id >> core_id_bits))
312 KASSERT(cpu_cores >= 1 && cpu_logical >= 1,
313 ("topo_probe_0x4 couldn't find BSP"));
315 cpu_cores /= cpu_logical;
316 hyperthreading_cpus = cpu_logical;
330 /* We only support three levels for now. */
331 for (i = 0; i < 3; i++) {
332 cpuid_count(0x0b, i, p);
334 /* Fall back if CPU leaf 11 doesn't really exist. */
335 if (i == 0 && p[1] == 0) {
341 logical = p[1] &= 0xffff;
342 type = (p[2] >> 8) & 0xff;
343 if (type == 0 || logical == 0)
346 * Because of uniformity assumption we examine only
347 * those logical processors that belong to the same
350 for (cnt = 0, x = 0; x <= MAX_APIC_ID; x++) {
351 if (!cpu_info[x].cpu_present ||
352 cpu_info[x].cpu_disabled)
354 if (x >> bits == boot_cpu_id >> bits)
357 if (type == CPUID_TYPE_SMT)
359 else if (type == CPUID_TYPE_CORE)
362 if (cpu_logical == 0)
364 cpu_cores /= cpu_logical;
368 * Both topology discovery code and code that consumes topology
369 * information assume top-down uniformity of the topology.
370 * That is, all physical packages must be identical and each
371 * core in a package must have the same number of threads.
372 * Topology information is queried only on BSP, on which this
373 * code runs and for which it can query CPUID information.
374 * Then topology is extrapolated on all packages using the
375 * uniformity assumption.
380 static int cpu_topo_probed = 0;
385 CPU_ZERO(&logical_cpus_mask);
387 cpu_cores = cpu_logical = 1;
388 else if (cpu_vendor_id == CPU_VENDOR_AMD)
390 else if (cpu_vendor_id == CPU_VENDOR_INTEL) {
392 * See Intel(R) 64 Architecture Processor
393 * Topology Enumeration article for details.
395 * Note that 0x1 <= cpu_high < 4 case should be
396 * compatible with topo_probe_0x4() logic when
397 * CPUID.1:EBX[23:16] > 0 (cpu_cores will be 1)
398 * or it should trigger the fallback otherwise.
402 else if (cpu_high >= 0x1)
407 * Fallback: assume each logical CPU is in separate
408 * physical package. That is, no multi-core, no SMT.
410 if (cpu_cores == 0 || cpu_logical == 0)
411 cpu_cores = cpu_logical = 1;
421 * Determine whether any threading flags are
425 if (cpu_logical > 1 && hyperthreading_cpus)
426 cg_flags = CG_FLAG_HTT;
427 else if (cpu_logical > 1)
428 cg_flags = CG_FLAG_SMT;
431 if (mp_ncpus % (cpu_cores * cpu_logical) != 0) {
432 printf("WARNING: Non-uniform processors.\n");
433 printf("WARNING: Using suboptimal topology.\n");
434 return (smp_topo_none());
437 * No multi-core or hyper-threaded.
439 if (cpu_logical * cpu_cores == 1)
440 return (smp_topo_none());
442 * Only HTT no multi-core.
444 if (cpu_logical > 1 && cpu_cores == 1)
445 return (smp_topo_1level(CG_SHARE_L1, cpu_logical, cg_flags));
447 * Only multi-core no HTT.
449 if (cpu_cores > 1 && cpu_logical == 1)
450 return (smp_topo_1level(CG_SHARE_L2, cpu_cores, cg_flags));
452 * Both HTT and multi-core.
454 return (smp_topo_2level(CG_SHARE_L2, cpu_cores,
455 CG_SHARE_L1, cpu_logical, cg_flags));
460 * Calculate usable address in base memory for AP trampoline code.
463 mp_bootaddress(u_int basemem)
466 boot_address = trunc_page(basemem); /* round down to 4k boundary */
467 if ((basemem - boot_address) < bootMP_size)
468 boot_address -= PAGE_SIZE; /* not enough, lower by 4k */
474 cpu_add(u_int apic_id, char boot_cpu)
477 if (apic_id > MAX_APIC_ID) {
478 panic("SMP: APIC ID %d too high", apic_id);
481 KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %d added twice",
483 cpu_info[apic_id].cpu_present = 1;
485 KASSERT(boot_cpu_id == -1,
486 ("CPU %d claims to be BSP, but CPU %d already is", apic_id,
488 boot_cpu_id = apic_id;
489 cpu_info[apic_id].cpu_bsp = 1;
491 if (mp_ncpus < MAXCPU) {
493 mp_maxid = mp_ncpus - 1;
496 printf("SMP: Added CPU %d (%s)\n", apic_id, boot_cpu ? "BSP" :
501 cpu_mp_setmaxid(void)
505 * mp_maxid should be already set by calls to cpu_add().
506 * Just sanity check its value here.
509 KASSERT(mp_maxid == 0,
510 ("%s: mp_ncpus is zero, but mp_maxid is not", __func__));
511 else if (mp_ncpus == 1)
514 KASSERT(mp_maxid >= mp_ncpus - 1,
515 ("%s: counters out of sync: max %d, count %d", __func__,
516 mp_maxid, mp_ncpus));
524 * Always record BSP in CPU map so that the mbuf init code works
527 CPU_SETOF(0, &all_cpus);
530 * No CPUs were found, so this must be a UP system. Setup
531 * the variables to represent a system with a single CPU
538 /* At least one CPU was found. */
541 * One CPU was found, so this must be a UP system with
548 /* At least two CPUs were found. */
553 * Initialize the IPI handlers and start up the AP's.
560 /* Initialize the logical ID to APIC ID table. */
561 for (i = 0; i < MAXCPU; i++) {
562 cpu_apic_ids[i] = -1;
563 cpu_ipi_pending[i] = 0;
566 /* Install an inter-CPU IPI for TLB invalidation */
567 setidt(IPI_INVLTLB, IDTVEC(invltlb),
568 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
569 setidt(IPI_INVLPG, IDTVEC(invlpg),
570 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
571 setidt(IPI_INVLRNG, IDTVEC(invlrng),
572 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
574 /* Install an inter-CPU IPI for cache invalidation. */
575 setidt(IPI_INVLCACHE, IDTVEC(invlcache),
576 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
578 /* Install an inter-CPU IPI for lazy pmap release */
579 setidt(IPI_LAZYPMAP, IDTVEC(lazypmap),
580 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
582 /* Install an inter-CPU IPI for all-CPU rendezvous */
583 setidt(IPI_RENDEZVOUS, IDTVEC(rendezvous),
584 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
586 /* Install generic inter-CPU IPI handler */
587 setidt(IPI_BITMAP_VECTOR, IDTVEC(ipi_intr_bitmap_handler),
588 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
590 /* Install an inter-CPU IPI for CPU stop/restart */
591 setidt(IPI_STOP, IDTVEC(cpustop),
592 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
594 /* Install an inter-CPU IPI for CPU suspend/resume */
595 setidt(IPI_SUSPEND, IDTVEC(cpususpend),
596 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
598 /* Set boot_cpu_id if needed. */
599 if (boot_cpu_id == -1) {
600 boot_cpu_id = PCPU_GET(apic_id);
601 cpu_info[boot_cpu_id].cpu_bsp = 1;
603 KASSERT(boot_cpu_id == PCPU_GET(apic_id),
604 ("BSP's APIC ID doesn't match boot_cpu_id"));
606 /* Probe logical/physical core configuration. */
611 /* Start each Application Processor */
614 set_interrupt_apic_ids();
619 * Print various information about the SMP system hardware and setup.
622 cpu_mp_announce(void)
624 const char *hyperthread;
627 printf("FreeBSD/SMP: %d package(s) x %d core(s)",
628 mp_ncpus / (cpu_cores * cpu_logical), cpu_cores);
629 if (hyperthreading_cpus > 1)
630 printf(" x %d HTT threads", cpu_logical);
631 else if (cpu_logical > 1)
632 printf(" x %d SMT threads", cpu_logical);
635 /* List active CPUs first. */
636 printf(" cpu0 (BSP): APIC ID: %2d\n", boot_cpu_id);
637 for (i = 1; i < mp_ncpus; i++) {
638 if (cpu_info[cpu_apic_ids[i]].cpu_hyperthread)
642 printf(" cpu%d (AP%s): APIC ID: %2d\n", i, hyperthread,
646 /* List disabled CPUs last. */
647 for (i = 0; i <= MAX_APIC_ID; i++) {
648 if (!cpu_info[i].cpu_present || !cpu_info[i].cpu_disabled)
650 if (cpu_info[i].cpu_hyperthread)
654 printf(" cpu (AP%s): APIC ID: %2d (disabled)\n", hyperthread,
660 * AP CPU's call this to initialize themselves.
671 /* bootAP is set in start_ap() to our ID. */
674 /* Get per-cpu data */
677 /* prime data page for it to use */
678 pcpu_init(pc, myid, sizeof(struct pcpu));
679 dpcpu_init(dpcpu, myid);
680 pc->pc_apic_id = cpu_apic_ids[myid];
681 pc->pc_prvspace = pc;
682 pc->pc_curthread = 0;
684 gdt_segs[GPRIV_SEL].ssd_base = (int) pc;
685 gdt_segs[GPROC0_SEL].ssd_base = (int) &pc->pc_common_tss;
687 for (x = 0; x < NGDT; x++) {
688 ssdtosd(&gdt_segs[x], &gdt[myid * NGDT + x].sd);
691 r_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
692 r_gdt.rd_base = (int) &gdt[myid * NGDT];
693 lgdt(&r_gdt); /* does magic intra-segment return */
698 PCPU_SET(currentldt, _default_ldt);
700 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
701 gdt[myid * NGDT + GPROC0_SEL].sd.sd_type = SDT_SYS386TSS;
702 PCPU_SET(common_tss.tss_esp0, 0); /* not used until after switch */
703 PCPU_SET(common_tss.tss_ss0, GSEL(GDATA_SEL, SEL_KPL));
704 PCPU_SET(common_tss.tss_ioopt, (sizeof (struct i386tss)) << 16);
705 PCPU_SET(tss_gdt, &gdt[myid * NGDT + GPROC0_SEL].sd);
706 PCPU_SET(common_tssd, *PCPU_GET(tss_gdt));
709 PCPU_SET(fsgs_gdt, &gdt[myid * NGDT + GUFS_SEL].sd);
712 * Set to a known state:
713 * Set by mpboot.s: CR0_PG, CR0_PE
714 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
717 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
719 CHECK_WRITE(0x38, 5);
721 /* Disable local APIC just to be sure. */
724 /* signal our startup to the BSP. */
726 CHECK_WRITE(0x39, 6);
728 /* Spin until the BSP releases the AP's. */
732 /* BSP may have changed PTD while we were waiting */
734 for (addr = 0; addr < NKPT * NBPDR - 1; addr += PAGE_SIZE)
737 #if defined(I586_CPU) && !defined(NO_F00F_HACK)
741 /* Initialize the PAT MSR if present. */
744 /* set up CPU registers and state */
747 /* set up FPU state on the AP */
750 /* set up SSE registers */
754 /* Enable the PTE no-execute bit. */
755 if ((amd_feature & AMDID_NX) != 0) {
758 msr = rdmsr(MSR_EFER) | EFER_NXE;
759 wrmsr(MSR_EFER, msr);
763 /* A quick check from sanity claus */
764 cpuid = PCPU_GET(cpuid);
765 if (PCPU_GET(apic_id) != lapic_id()) {
766 printf("SMP: cpuid = %d\n", cpuid);
767 printf("SMP: actual apic_id = %d\n", lapic_id());
768 printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id));
769 panic("cpuid mismatch! boom!!");
772 /* Initialize curthread. */
773 KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread"));
774 PCPU_SET(curthread, PCPU_GET(idlethread));
778 mtx_lock_spin(&ap_boot_mtx);
780 /* Init local apic for irq's */
783 /* Set memory range attributes for this CPU to match the BSP */
788 CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid);
789 printf("SMP: AP CPU #%d Launched!\n", cpuid);
791 /* Determine if we are a logical CPU. */
792 /* XXX Calculation depends on cpu_logical being a power of 2, e.g. 2 */
793 if (cpu_logical > 1 && PCPU_GET(apic_id) % cpu_logical != 0)
794 CPU_SET(cpuid, &logical_cpus_mask);
799 if (smp_cpus == mp_ncpus) {
800 /* enable IPI's, tlb shootdown, freezes etc */
801 atomic_store_rel_int(&smp_started, 1);
802 smp_active = 1; /* historic */
805 mtx_unlock_spin(&ap_boot_mtx);
807 /* Wait until all the AP's are up. */
808 while (smp_started == 0)
811 /* Start per-CPU event timers. */
814 /* Enter the scheduler. */
817 panic("scheduler returned us to %s", __func__);
821 /*******************************************************************
822 * local functions and data
826 * We tell the I/O APIC code about all the CPUs we want to receive
827 * interrupts. If we don't want certain CPUs to receive IRQs we
828 * can simply not tell the I/O APIC code about them in this function.
831 set_interrupt_apic_ids(void)
835 for (i = 0; i < MAXCPU; i++) {
836 apic_id = cpu_apic_ids[i];
839 if (cpu_info[apic_id].cpu_disabled)
842 /* Don't let hyperthreads service interrupts. */
843 if (hyperthreading_cpus > 1 &&
844 apic_id % hyperthreading_cpus != 0)
852 * Assign logical CPU IDs to local APICs.
859 TUNABLE_INT_FETCH("machdep.hyperthreading_allowed",
860 &hyperthreading_allowed);
862 /* Check for explicitly disabled CPUs. */
863 for (i = 0; i <= MAX_APIC_ID; i++) {
864 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp)
867 if (hyperthreading_cpus > 1 && i % hyperthreading_cpus != 0) {
868 cpu_info[i].cpu_hyperthread = 1;
871 * Don't use HT CPU if it has been disabled by a
874 if (hyperthreading_allowed == 0) {
875 cpu_info[i].cpu_disabled = 1;
880 /* Don't use this CPU if it has been disabled by a tunable. */
881 if (resource_disabled("lapic", i)) {
882 cpu_info[i].cpu_disabled = 1;
887 if (hyperthreading_allowed == 0 && hyperthreading_cpus > 1) {
888 hyperthreading_cpus = 0;
893 * Assign CPU IDs to local APIC IDs and disable any CPUs
894 * beyond MAXCPU. CPU 0 is always assigned to the BSP.
896 * To minimize confusion for userland, we attempt to number
897 * CPUs such that all threads and cores in a package are
898 * grouped together. For now we assume that the BSP is always
899 * the first thread in a package and just start adding APs
900 * starting with the BSP's APIC ID.
903 cpu_apic_ids[0] = boot_cpu_id;
904 apic_cpuids[boot_cpu_id] = 0;
905 for (i = boot_cpu_id + 1; i != boot_cpu_id;
906 i == MAX_APIC_ID ? i = 0 : i++) {
907 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp ||
908 cpu_info[i].cpu_disabled)
911 if (mp_ncpus < MAXCPU) {
912 cpu_apic_ids[mp_ncpus] = i;
913 apic_cpuids[i] = mp_ncpus;
916 cpu_info[i].cpu_disabled = 1;
918 KASSERT(mp_maxid >= mp_ncpus - 1,
919 ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid,
924 * start each AP in our list
926 /* Lowest 1MB is already mapped: don't touch*/
927 #define TMPMAP_START 1
935 u_int32_t mpbioswarmvec;
938 mtx_init(&ap_boot_mtx, "ap boot", NULL, MTX_SPIN);
940 /* install the AP 1st level boot code */
943 /* save the current value of the warm-start vector */
944 mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF);
946 outb(CMOS_REG, BIOS_RESET);
947 mpbiosreason = inb(CMOS_DATA);
950 /* set up temporary P==V mapping for AP boot */
951 /* XXX this is a hack, we should boot the AP on its own stack/PTD */
953 kptbase = (uintptr_t)(void *)KPTphys;
954 for (i = TMPMAP_START; i < NKPT; i++)
955 PTD[i] = (pd_entry_t)(PG_V | PG_RW |
956 ((kptbase + i * PAGE_SIZE) & PG_FRAME));
960 for (cpu = 1; cpu < mp_ncpus; cpu++) {
961 apic_id = cpu_apic_ids[cpu];
963 /* allocate and set up a boot stack data page */
965 (char *)kmem_alloc(kernel_map, KSTACK_PAGES * PAGE_SIZE);
966 dpcpu = (void *)kmem_alloc(kernel_map, DPCPU_SIZE);
967 /* setup a vector to our boot code */
968 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
969 *((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4);
971 outb(CMOS_REG, BIOS_RESET);
972 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
975 bootSTK = (char *)bootstacks[cpu] + KSTACK_PAGES * PAGE_SIZE - 4;
978 /* attempt to start the Application Processor */
979 CHECK_INIT(99); /* setup checkpoints */
980 if (!start_ap(apic_id)) {
981 printf("AP #%d (PHY# %d) failed!\n", cpu, apic_id);
982 CHECK_PRINT("trace"); /* show checkpoints */
983 /* better panic as the AP may be running loose */
984 printf("panic y/n? [y] ");
988 CHECK_PRINT("trace"); /* show checkpoints */
990 CPU_SET(cpu, &all_cpus); /* record AP in CPU map */
993 /* restore the warmstart vector */
994 *(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec;
997 outb(CMOS_REG, BIOS_RESET);
998 outb(CMOS_DATA, mpbiosreason);
1001 /* Undo V==P hack from above */
1002 for (i = TMPMAP_START; i < NKPT; i++)
1004 pmap_invalidate_range(kernel_pmap, 0, NKPT * NBPDR - 1);
1006 /* number of APs actually started */
1011 * load the 1st level AP boot code into base memory.
1014 /* targets for relocation */
1015 extern void bigJump(void);
1016 extern void bootCodeSeg(void);
1017 extern void bootDataSeg(void);
1018 extern void MPentry(void);
1019 extern u_int MP_GDT;
1020 extern u_int mp_gdtbase;
1023 install_ap_tramp(void)
1026 int size = *(int *) ((u_long) & bootMP_size);
1027 vm_offset_t va = boot_address + KERNBASE;
1028 u_char *src = (u_char *) ((u_long) bootMP);
1029 u_char *dst = (u_char *) va;
1030 u_int boot_base = (u_int) bootMP;
1035 KASSERT (size <= PAGE_SIZE,
1036 ("'size' do not fit into PAGE_SIZE, as expected."));
1037 pmap_kenter(va, boot_address);
1038 pmap_invalidate_page (kernel_pmap, va);
1039 for (x = 0; x < size; ++x)
1043 * modify addresses in code we just moved to basemem. unfortunately we
1044 * need fairly detailed info about mpboot.s for this to work. changes
1045 * to mpboot.s might require changes here.
1048 /* boot code is located in KERNEL space */
1049 dst = (u_char *) va;
1051 /* modify the lgdt arg */
1052 dst32 = (u_int32_t *) (dst + ((u_int) & mp_gdtbase - boot_base));
1053 *dst32 = boot_address + ((u_int) & MP_GDT - boot_base);
1055 /* modify the ljmp target for MPentry() */
1056 dst32 = (u_int32_t *) (dst + ((u_int) bigJump - boot_base) + 1);
1057 *dst32 = ((u_int) MPentry - KERNBASE);
1059 /* modify the target for boot code segment */
1060 dst16 = (u_int16_t *) (dst + ((u_int) bootCodeSeg - boot_base));
1061 dst8 = (u_int8_t *) (dst16 + 1);
1062 *dst16 = (u_int) boot_address & 0xffff;
1063 *dst8 = ((u_int) boot_address >> 16) & 0xff;
1065 /* modify the target for boot data segment */
1066 dst16 = (u_int16_t *) (dst + ((u_int) bootDataSeg - boot_base));
1067 dst8 = (u_int8_t *) (dst16 + 1);
1068 *dst16 = (u_int) boot_address & 0xffff;
1069 *dst8 = ((u_int) boot_address >> 16) & 0xff;
1073 * This function starts the AP (application processor) identified
1074 * by the APIC ID 'physicalCpu'. It does quite a "song and dance"
1075 * to accomplish this. This is necessary because of the nuances
1076 * of the different hardware we might encounter. It isn't pretty,
1077 * but it seems to work.
1080 start_ap(int apic_id)
1085 /* calculate the vector */
1086 vector = (boot_address >> 12) & 0xff;
1088 /* used as a watchpoint to signal AP startup */
1092 * first we do an INIT/RESET IPI this INIT IPI might be run, reseting
1093 * and running the target CPU. OR this INIT IPI might be latched (P5
1094 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
1098 /* do an INIT IPI: assert RESET */
1099 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1100 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id);
1102 /* wait for pending status end */
1105 /* do an INIT IPI: deassert RESET */
1106 lapic_ipi_raw(APIC_DEST_ALLESELF | APIC_TRIGMOD_LEVEL |
1107 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, 0);
1109 /* wait for pending status end */
1110 DELAY(10000); /* wait ~10mS */
1114 * next we do a STARTUP IPI: the previous INIT IPI might still be
1115 * latched, (P5 bug) this 1st STARTUP would then terminate
1116 * immediately, and the previously started INIT IPI would continue. OR
1117 * the previous INIT IPI has already run. and this STARTUP IPI will
1118 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
1122 /* do a STARTUP IPI */
1123 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1124 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1127 DELAY(200); /* wait ~200uS */
1130 * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
1131 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
1132 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
1133 * recognized after hardware RESET or INIT IPI.
1136 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1137 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1140 DELAY(200); /* wait ~200uS */
1142 /* Wait up to 5 seconds for it to start. */
1143 for (ms = 0; ms < 5000; ms++) {
1145 return 1; /* return SUCCESS */
1148 return 0; /* return FAILURE */
1151 #ifdef COUNT_XINVLTLB_HITS
1152 u_int xhits_gbl[MAXCPU];
1153 u_int xhits_pg[MAXCPU];
1154 u_int xhits_rng[MAXCPU];
1155 static SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW, 0, "");
1156 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl,
1157 sizeof(xhits_gbl), "IU", "");
1158 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg,
1159 sizeof(xhits_pg), "IU", "");
1160 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng,
1161 sizeof(xhits_rng), "IU", "");
1166 u_int ipi_range_size;
1167 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, "");
1168 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, "");
1169 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, "");
1170 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW, &ipi_range_size,
1173 u_int ipi_masked_global;
1174 u_int ipi_masked_page;
1175 u_int ipi_masked_range;
1176 u_int ipi_masked_range_size;
1177 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_masked_global, CTLFLAG_RW,
1178 &ipi_masked_global, 0, "");
1179 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_masked_page, CTLFLAG_RW,
1180 &ipi_masked_page, 0, "");
1181 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_masked_range, CTLFLAG_RW,
1182 &ipi_masked_range, 0, "");
1183 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_masked_range_size, CTLFLAG_RW,
1184 &ipi_masked_range_size, 0, "");
1185 #endif /* COUNT_XINVLTLB_HITS */
1188 * Send an IPI to specified CPU handling the bitmap logic.
1191 ipi_send_cpu(int cpu, u_int ipi)
1193 u_int bitmap, old_pending, new_pending;
1195 KASSERT(cpu_apic_ids[cpu] != -1, ("IPI to non-existent CPU %d", cpu));
1197 if (IPI_IS_BITMAPED(ipi)) {
1199 ipi = IPI_BITMAP_VECTOR;
1201 old_pending = cpu_ipi_pending[cpu];
1202 new_pending = old_pending | bitmap;
1203 } while (!atomic_cmpset_int(&cpu_ipi_pending[cpu],
1204 old_pending, new_pending));
1208 lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]);
1212 * Flush the TLB on all other CPU's
1215 smp_tlb_shootdown(u_int vector, vm_offset_t addr1, vm_offset_t addr2)
1219 ncpu = mp_ncpus - 1; /* does not shootdown self */
1221 return; /* no other cpus */
1222 if (!(read_eflags() & PSL_I))
1223 panic("%s: interrupts disabled", __func__);
1224 mtx_lock_spin(&smp_ipi_mtx);
1225 smp_tlb_addr1 = addr1;
1226 smp_tlb_addr2 = addr2;
1227 atomic_store_rel_int(&smp_tlb_wait, 0);
1228 ipi_all_but_self(vector);
1229 while (smp_tlb_wait < ncpu)
1231 mtx_unlock_spin(&smp_ipi_mtx);
1235 smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, vm_offset_t addr1, vm_offset_t addr2)
1237 int cpu, ncpu, othercpus;
1239 othercpus = mp_ncpus - 1;
1240 if (CPU_ISFULLSET(&mask)) {
1244 CPU_CLR(PCPU_GET(cpuid), &mask);
1245 if (CPU_EMPTY(&mask))
1248 if (!(read_eflags() & PSL_I))
1249 panic("%s: interrupts disabled", __func__);
1250 mtx_lock_spin(&smp_ipi_mtx);
1251 smp_tlb_addr1 = addr1;
1252 smp_tlb_addr2 = addr2;
1253 atomic_store_rel_int(&smp_tlb_wait, 0);
1254 if (CPU_ISFULLSET(&mask)) {
1256 ipi_all_but_self(vector);
1259 while ((cpu = CPU_FFS(&mask)) != 0) {
1261 CPU_CLR(cpu, &mask);
1262 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu,
1264 ipi_send_cpu(cpu, vector);
1268 while (smp_tlb_wait < ncpu)
1270 mtx_unlock_spin(&smp_ipi_mtx);
1274 smp_cache_flush(void)
1278 smp_tlb_shootdown(IPI_INVLCACHE, 0, 0);
1286 smp_tlb_shootdown(IPI_INVLTLB, 0, 0);
1287 #ifdef COUNT_XINVLTLB_HITS
1294 smp_invlpg(vm_offset_t addr)
1298 smp_tlb_shootdown(IPI_INVLPG, addr, 0);
1299 #ifdef COUNT_XINVLTLB_HITS
1306 smp_invlpg_range(vm_offset_t addr1, vm_offset_t addr2)
1310 smp_tlb_shootdown(IPI_INVLRNG, addr1, addr2);
1311 #ifdef COUNT_XINVLTLB_HITS
1313 ipi_range_size += (addr2 - addr1) / PAGE_SIZE;
1319 smp_masked_invltlb(cpuset_t mask)
1323 smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, 0, 0);
1324 #ifdef COUNT_XINVLTLB_HITS
1325 ipi_masked_global++;
1331 smp_masked_invlpg(cpuset_t mask, vm_offset_t addr)
1335 smp_targeted_tlb_shootdown(mask, IPI_INVLPG, addr, 0);
1336 #ifdef COUNT_XINVLTLB_HITS
1343 smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2)
1347 smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, addr1, addr2);
1348 #ifdef COUNT_XINVLTLB_HITS
1350 ipi_masked_range_size += (addr2 - addr1) / PAGE_SIZE;
1356 ipi_bitmap_handler(struct trapframe frame)
1358 struct trapframe *oldframe;
1360 int cpu = PCPU_GET(cpuid);
1365 td->td_intr_nesting_level++;
1366 oldframe = td->td_intr_frame;
1367 td->td_intr_frame = &frame;
1368 ipi_bitmap = atomic_readandclear_int(&cpu_ipi_pending[cpu]);
1369 if (ipi_bitmap & (1 << IPI_PREEMPT)) {
1371 (*ipi_preempt_counts[cpu])++;
1375 if (ipi_bitmap & (1 << IPI_AST)) {
1377 (*ipi_ast_counts[cpu])++;
1379 /* Nothing to do for AST */
1381 if (ipi_bitmap & (1 << IPI_HARDCLOCK)) {
1383 (*ipi_hardclock_counts[cpu])++;
1387 td->td_intr_frame = oldframe;
1388 td->td_intr_nesting_level--;
1393 * send an IPI to a set of cpus.
1396 ipi_selected(cpuset_t cpus, u_int ipi)
1401 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1402 * of help in order to understand what is the source.
1403 * Set the mask of receiving CPUs for this purpose.
1405 if (ipi == IPI_STOP_HARD)
1406 CPU_OR_ATOMIC(&ipi_nmi_pending, &cpus);
1408 while ((cpu = CPU_FFS(&cpus)) != 0) {
1410 CPU_CLR(cpu, &cpus);
1411 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1412 ipi_send_cpu(cpu, ipi);
1417 * send an IPI to a specific CPU.
1420 ipi_cpu(int cpu, u_int ipi)
1424 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1425 * of help in order to understand what is the source.
1426 * Set the mask of receiving CPUs for this purpose.
1428 if (ipi == IPI_STOP_HARD)
1429 CPU_SET_ATOMIC(cpu, &ipi_nmi_pending);
1431 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1432 ipi_send_cpu(cpu, ipi);
1436 * send an IPI to all CPUs EXCEPT myself
1439 ipi_all_but_self(u_int ipi)
1441 cpuset_t other_cpus;
1443 other_cpus = all_cpus;
1444 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1445 if (IPI_IS_BITMAPED(ipi)) {
1446 ipi_selected(other_cpus, ipi);
1451 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1452 * of help in order to understand what is the source.
1453 * Set the mask of receiving CPUs for this purpose.
1455 if (ipi == IPI_STOP_HARD)
1456 CPU_OR_ATOMIC(&ipi_nmi_pending, &other_cpus);
1458 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1459 lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS);
1468 * As long as there is not a simple way to know about a NMI's
1469 * source, if the bitmask for the current CPU is present in
1470 * the global pending bitword an IPI_STOP_HARD has been issued
1471 * and should be handled.
1473 cpuid = PCPU_GET(cpuid);
1474 if (!CPU_ISSET(cpuid, &ipi_nmi_pending))
1477 CPU_CLR_ATOMIC(cpuid, &ipi_nmi_pending);
1483 * Handle an IPI_STOP by saving our current context and spinning until we
1487 cpustop_handler(void)
1491 cpu = PCPU_GET(cpuid);
1493 savectx(&stoppcbs[cpu]);
1495 /* Indicate that we are stopped */
1496 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1498 /* Wait for restart */
1499 while (!CPU_ISSET(cpu, &started_cpus))
1502 CPU_CLR_ATOMIC(cpu, &started_cpus);
1503 CPU_CLR_ATOMIC(cpu, &stopped_cpus);
1505 if (cpu == 0 && cpustop_restartfunc != NULL) {
1506 cpustop_restartfunc();
1507 cpustop_restartfunc = NULL;
1512 * Handle an IPI_SUSPEND by saving our current context and spinning until we
1516 cpususpend_handler(void)
1520 cpu = PCPU_GET(cpuid);
1522 if (suspendctx(susppcbs[cpu])) {
1524 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1527 PCPU_SET(switchtime, 0);
1528 PCPU_SET(switchticks, ticks);
1529 susppcbs[cpu]->pcb_eip = 0;
1532 /* Wait for resume */
1533 while (!CPU_ISSET(cpu, &started_cpus))
1536 /* Resume MCA and local APIC */
1540 CPU_CLR_ATOMIC(cpu, &started_cpus);
1541 CPU_CLR_ATOMIC(cpu, &stopped_cpus);
1544 * This is called once the rest of the system is up and running and we're
1545 * ready to let the AP's out of the pen.
1548 release_aps(void *dummy __unused)
1553 atomic_store_rel_int(&aps_ready, 1);
1554 while (smp_started == 0)
1557 SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL);
1561 * Setup interrupt counters for IPI handlers.
1564 mp_ipi_intrcnt(void *dummy)
1570 snprintf(buf, sizeof(buf), "cpu%d:invltlb", i);
1571 intrcnt_add(buf, &ipi_invltlb_counts[i]);
1572 snprintf(buf, sizeof(buf), "cpu%d:invlrng", i);
1573 intrcnt_add(buf, &ipi_invlrng_counts[i]);
1574 snprintf(buf, sizeof(buf), "cpu%d:invlpg", i);
1575 intrcnt_add(buf, &ipi_invlpg_counts[i]);
1576 snprintf(buf, sizeof(buf), "cpu%d:invlcache", i);
1577 intrcnt_add(buf, &ipi_invlcache_counts[i]);
1578 snprintf(buf, sizeof(buf), "cpu%d:preempt", i);
1579 intrcnt_add(buf, &ipi_preempt_counts[i]);
1580 snprintf(buf, sizeof(buf), "cpu%d:ast", i);
1581 intrcnt_add(buf, &ipi_ast_counts[i]);
1582 snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i);
1583 intrcnt_add(buf, &ipi_rendezvous_counts[i]);
1584 snprintf(buf, sizeof(buf), "cpu%d:lazypmap", i);
1585 intrcnt_add(buf, &ipi_lazypmap_counts[i]);
1586 snprintf(buf, sizeof(buf), "cpu%d:hardclock", i);
1587 intrcnt_add(buf, &ipi_hardclock_counts[i]);
1590 SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL);