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_mp_watchdog.h"
33 #include "opt_sched.h"
35 #include <sys/param.h>
36 #include <sys/systm.h>
41 #include <sys/kernel.h>
44 #include <sys/malloc.h>
45 #include <sys/memrange.h>
46 #include <sys/mutex.h>
49 #include <sys/sched.h>
51 #include <sys/sysctl.h>
54 #include <vm/vm_param.h>
56 #include <vm/vm_kern.h>
57 #include <vm/vm_extern.h>
59 #include <machine/apicreg.h>
60 #include <machine/md_var.h>
61 #include <machine/mp_watchdog.h>
62 #include <machine/pcb.h>
63 #include <machine/psl.h>
64 #include <machine/smp.h>
65 #include <machine/specialreg.h>
66 #include <machine/tss.h>
68 #define WARMBOOT_TARGET 0
69 #define WARMBOOT_OFF (KERNBASE + 0x0467)
70 #define WARMBOOT_SEG (KERNBASE + 0x0469)
72 #define CMOS_REG (0x70)
73 #define CMOS_DATA (0x71)
74 #define BIOS_RESET (0x0f)
75 #define BIOS_WARM (0x0a)
77 /* lock region used by kernel profiling */
80 int mp_naps; /* # of Applications processors */
81 int boot_cpu_id = -1; /* designated BSP */
83 extern struct pcpu __pcpu[];
86 * CPU topology map datastructures for HTT.
88 static struct cpu_group mp_groups[MAXCPU];
89 static struct cpu_top mp_top;
91 /* AP uses this during bootstrap. Do not staticize. */
95 /* Free these after use */
96 void *bootstacks[MAXCPU];
98 /* Temporary holder for double fault stack */
99 char *doublefault_stack;
101 /* Hotwire a 0->4MB V==P mapping */
102 extern pt_entry_t *KPTphys;
104 /* SMP page table page */
105 extern pt_entry_t *SMPpt;
107 struct pcb stoppcbs[MAXCPU];
109 /* Variables needed for SMP tlb shootdown. */
110 vm_offset_t smp_tlb_addr1;
111 vm_offset_t smp_tlb_addr2;
112 volatile int smp_tlb_wait;
114 extern inthand_t IDTVEC(fast_syscall), IDTVEC(fast_syscall32);
117 volatile cpumask_t ipi_nmi_pending;
119 static void ipi_nmi_selected(u_int32_t cpus);
123 * Local data and functions.
128 * Provide an alternate method of stopping other CPUs. If another CPU has
129 * disabled interrupts the conventional STOP IPI will be blocked. This
130 * NMI-based stop should get through in that case.
132 static int stop_cpus_with_nmi = 1;
133 SYSCTL_INT(_debug, OID_AUTO, stop_cpus_with_nmi, CTLTYPE_INT | CTLFLAG_RW,
134 &stop_cpus_with_nmi, 0, "");
135 TUNABLE_INT("debug.stop_cpus_with_nmi", &stop_cpus_with_nmi);
137 #define stop_cpus_with_nmi 0
140 static u_int logical_cpus;
142 /* used to hold the AP's until we are ready to release them */
143 static struct mtx ap_boot_mtx;
145 /* Set to 1 once we're ready to let the APs out of the pen. */
146 static volatile int aps_ready = 0;
149 * Store data from cpu_add() until later in the boot when we actually setup
156 int cpu_hyperthread:1;
157 } static cpu_info[MAX_APIC_ID + 1];
158 int cpu_apic_ids[MAXCPU];
160 /* Holds pending bitmap based IPIs per CPU */
161 static volatile u_int cpu_ipi_pending[MAXCPU];
163 static u_int boot_address;
165 static void assign_cpu_ids(void);
166 static void set_interrupt_apic_ids(void);
167 static int start_all_aps(void);
168 static int start_ap(int apic_id);
169 static void release_aps(void *dummy);
171 static int hlt_logical_cpus;
172 static u_int hyperthreading_cpus;
173 static cpumask_t hyperthreading_cpus_mask;
174 static int hyperthreading_allowed = 1;
175 static struct sysctl_ctx_list logical_cpu_clist;
176 static u_int bootMP_size;
179 mem_range_AP_init(void)
181 if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP)
182 mem_range_softc.mr_op->initAP(&mem_range_softc);
188 struct cpu_group *group;
193 /* Build the smp_topology map. */
194 /* Nothing to do if there is no HTT support. */
195 if (hyperthreading_cpus <= 1)
197 group = &mp_groups[0];
199 for (cpu = 0, apic_id = 0; apic_id <= MAX_APIC_ID; apic_id++) {
200 if (!cpu_info[apic_id].cpu_present)
203 * If the current group has members and we're not a logical
204 * cpu, create a new group.
206 if (group->cg_count != 0 &&
207 (apic_id % hyperthreading_cpus) == 0) {
212 group->cg_mask |= 1 << cpu;
216 mp_top.ct_count = groups;
217 mp_top.ct_group = mp_groups;
218 smp_topology = &mp_top;
222 * Calculate usable address in base memory for AP trampoline code.
225 mp_bootaddress(u_int basemem)
228 bootMP_size = mptramp_end - mptramp_start;
229 boot_address = trunc_page(basemem * 1024); /* round down to 4k boundary */
230 if (((basemem * 1024) - boot_address) < bootMP_size)
231 boot_address -= PAGE_SIZE; /* not enough, lower by 4k */
232 /* 3 levels of page table pages */
233 mptramp_pagetables = boot_address - (PAGE_SIZE * 3);
235 return mptramp_pagetables;
239 cpu_add(u_int apic_id, char boot_cpu)
242 if (apic_id > MAX_APIC_ID) {
243 panic("SMP: APIC ID %d too high", apic_id);
246 KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %d added twice",
248 cpu_info[apic_id].cpu_present = 1;
250 KASSERT(boot_cpu_id == -1,
251 ("CPU %d claims to be BSP, but CPU %d already is", apic_id,
253 boot_cpu_id = apic_id;
254 cpu_info[apic_id].cpu_bsp = 1;
256 if (mp_ncpus < MAXCPU) {
258 mp_maxid = mp_ncpus -1;
261 printf("SMP: Added CPU %d (%s)\n", apic_id, boot_cpu ? "BSP" :
266 cpu_mp_setmaxid(void)
270 * mp_maxid should be already set by calls to cpu_add().
271 * Just sanity check its value here.
274 KASSERT(mp_maxid == 0,
275 ("%s: mp_ncpus is zero, but mp_maxid is not", __func__));
276 else if (mp_ncpus == 1)
279 KASSERT(mp_maxid >= mp_ncpus - 1,
280 ("%s: counters out of sync: max %d, count %d", __func__,
281 mp_maxid, mp_ncpus));
289 * Always record BSP in CPU map so that the mbuf init code works
295 * No CPUs were found, so this must be a UP system. Setup
296 * the variables to represent a system with a single CPU
303 /* At least one CPU was found. */
306 * One CPU was found, so this must be a UP system with
313 /* At least two CPUs were found. */
318 * Initialize the IPI handlers and start up the AP's.
324 u_int threads_per_cache, p[4];
326 /* Initialize the logical ID to APIC ID table. */
327 for (i = 0; i < MAXCPU; i++) {
328 cpu_apic_ids[i] = -1;
329 cpu_ipi_pending[i] = 0;
332 /* Install an inter-CPU IPI for TLB invalidation */
333 setidt(IPI_INVLTLB, IDTVEC(invltlb), SDT_SYSIGT, SEL_KPL, 0);
334 setidt(IPI_INVLPG, IDTVEC(invlpg), SDT_SYSIGT, SEL_KPL, 0);
335 setidt(IPI_INVLRNG, IDTVEC(invlrng), SDT_SYSIGT, SEL_KPL, 0);
337 /* Install an inter-CPU IPI for cache invalidation. */
338 setidt(IPI_INVLCACHE, IDTVEC(invlcache), SDT_SYSIGT, SEL_KPL, 0);
340 /* Install an inter-CPU IPI for all-CPU rendezvous */
341 setidt(IPI_RENDEZVOUS, IDTVEC(rendezvous), SDT_SYSIGT, SEL_KPL, 0);
343 /* Install generic inter-CPU IPI handler */
344 setidt(IPI_BITMAP_VECTOR, IDTVEC(ipi_intr_bitmap_handler),
345 SDT_SYSIGT, SEL_KPL, 0);
347 /* Install an inter-CPU IPI for CPU stop/restart */
348 setidt(IPI_STOP, IDTVEC(cpustop), SDT_SYSIGT, SEL_KPL, 0);
350 /* Set boot_cpu_id if needed. */
351 if (boot_cpu_id == -1) {
352 boot_cpu_id = PCPU_GET(apic_id);
353 cpu_info[boot_cpu_id].cpu_bsp = 1;
355 KASSERT(boot_cpu_id == PCPU_GET(apic_id),
356 ("BSP's APIC ID doesn't match boot_cpu_id"));
357 cpu_apic_ids[0] = boot_cpu_id;
359 /* Setup the initial logical CPUs info. */
360 logical_cpus = logical_cpus_mask = 0;
361 if (cpu_feature & CPUID_HTT)
362 logical_cpus = (cpu_procinfo & CPUID_HTT_CORES) >> 16;
365 * Work out if hyperthreading is *really* enabled. This
366 * is made really ugly by the fact that processors lie: Dual
367 * core processors claim to be hyperthreaded even when they're
368 * not, presumably because they want to be treated the same
369 * way as HTT with respect to per-cpu software licensing.
370 * At the time of writing (May 12, 2005) the only hyperthreaded
371 * cpus are from Intel, and Intel's dual-core processors can be
372 * identified via the "deterministic cache parameters" cpuid
376 * First determine if this is an Intel processor which claims
377 * to have hyperthreading support.
379 if ((cpu_feature & CPUID_HTT) &&
380 (strcmp(cpu_vendor, "GenuineIntel") == 0)) {
382 * If the "deterministic cache parameters" cpuid calls
383 * are available, use them.
386 /* Ask the processor about the L1 cache. */
387 for (i = 0; i < 1; i++) {
388 cpuid_count(4, i, p);
389 threads_per_cache = ((p[0] & 0x3ffc000) >> 14) + 1;
390 if (hyperthreading_cpus < threads_per_cache)
391 hyperthreading_cpus = threads_per_cache;
392 if ((p[0] & 0x1f) == 0)
398 * If the deterministic cache parameters are not
399 * available, or if no caches were reported to exist,
400 * just accept what the HTT flag indicated.
402 if (hyperthreading_cpus == 0)
403 hyperthreading_cpus = logical_cpus;
408 /* Start each Application Processor */
411 set_interrupt_apic_ids();
413 /* Last, setup the cpu topology now that we have probed CPUs */
419 * Print various information about the SMP system hardware and setup.
422 cpu_mp_announce(void)
425 const char *hyperthread;
428 printf(" cpu0 (BSP): APIC ID: %2d\n", boot_cpu_id);
429 for (i = 1, x = 0; x <= MAX_APIC_ID; x++) {
430 if (!cpu_info[x].cpu_present || cpu_info[x].cpu_bsp)
432 if (cpu_info[x].cpu_hyperthread) {
437 if (cpu_info[x].cpu_disabled)
438 printf(" cpu (AP%s): APIC ID: %2d (disabled)\n",
441 KASSERT(i < mp_ncpus,
442 ("mp_ncpus and actual cpus are out of whack"));
443 printf(" cpu%d (AP%s): APIC ID: %2d\n", i++,
450 * AP CPU's call this to initialize themselves.
457 int cpu, gsel_tss, x;
458 struct region_descriptor ap_gdt;
460 /* Set by the startup code for us to use */
464 common_tss[cpu] = common_tss[0];
465 common_tss[cpu].tss_rsp0 = 0; /* not used until after switch */
466 common_tss[cpu].tss_iobase = sizeof(struct amd64tss);
467 common_tss[cpu].tss_ist1 = (long)&doublefault_stack[PAGE_SIZE];
469 /* Prepare private GDT */
470 gdt_segs[GPROC0_SEL].ssd_base = (long) &common_tss[cpu];
471 ssdtosyssd(&gdt_segs[GPROC0_SEL],
472 (struct system_segment_descriptor *)&gdt[NGDT * cpu + GPROC0_SEL]);
473 for (x = 0; x < NGDT; x++) {
474 if (x != GPROC0_SEL && x != (GPROC0_SEL + 1))
475 ssdtosd(&gdt_segs[x], &gdt[NGDT * cpu + x]);
477 ap_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
478 ap_gdt.rd_base = (long) &gdt[NGDT * cpu];
479 lgdt(&ap_gdt); /* does magic intra-segment return */
481 /* Get per-cpu data */
484 /* prime data page for it to use */
485 pcpu_init(pc, cpu, sizeof(struct pcpu));
486 pc->pc_apic_id = cpu_apic_ids[cpu];
487 pc->pc_prvspace = pc;
488 pc->pc_curthread = 0;
489 pc->pc_tssp = &common_tss[cpu];
491 pc->pc_gs32p = &gdt[NGDT * cpu + GUGS32_SEL];
493 wrmsr(MSR_FSBASE, 0); /* User value */
494 wrmsr(MSR_GSBASE, (u_int64_t)pc);
495 wrmsr(MSR_KGSBASE, (u_int64_t)pc); /* XXX User value while we're in the kernel */
499 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
503 * Set to a known state:
504 * Set by mpboot.s: CR0_PG, CR0_PE
505 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
508 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
511 /* Set up the fast syscall stuff */
512 msr = rdmsr(MSR_EFER) | EFER_SCE;
513 wrmsr(MSR_EFER, msr);
514 wrmsr(MSR_LSTAR, (u_int64_t)IDTVEC(fast_syscall));
515 wrmsr(MSR_CSTAR, (u_int64_t)IDTVEC(fast_syscall32));
516 msr = ((u_int64_t)GSEL(GCODE_SEL, SEL_KPL) << 32) |
517 ((u_int64_t)GSEL(GUCODE32_SEL, SEL_UPL) << 48);
518 wrmsr(MSR_STAR, msr);
519 wrmsr(MSR_SF_MASK, PSL_NT|PSL_T|PSL_I|PSL_C|PSL_D);
521 /* Disable local APIC just to be sure. */
524 /* signal our startup to the BSP. */
527 /* Spin until the BSP releases the AP's. */
531 /* Initialize the PAT MSR. */
534 /* set up CPU registers and state */
537 /* set up SSE/NX registers */
540 /* set up FPU state on the AP */
543 /* A quick check from sanity claus */
544 if (PCPU_GET(apic_id) != lapic_id()) {
545 printf("SMP: cpuid = %d\n", PCPU_GET(cpuid));
546 printf("SMP: actual apic_id = %d\n", lapic_id());
547 printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id));
548 panic("cpuid mismatch! boom!!");
551 /* Initialize curthread. */
552 KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread"));
553 PCPU_SET(curthread, PCPU_GET(idlethread));
555 mtx_lock_spin(&ap_boot_mtx);
557 /* Init local apic for irq's */
560 /* Set memory range attributes for this CPU to match the BSP */
565 CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", PCPU_GET(cpuid));
566 printf("SMP: AP CPU #%d Launched!\n", PCPU_GET(cpuid));
568 /* Determine if we are a logical CPU. */
569 if (logical_cpus > 1 && PCPU_GET(apic_id) % logical_cpus != 0)
570 logical_cpus_mask |= PCPU_GET(cpumask);
572 /* Determine if we are a hyperthread. */
573 if (hyperthreading_cpus > 1 &&
574 PCPU_GET(apic_id) % hyperthreading_cpus != 0)
575 hyperthreading_cpus_mask |= PCPU_GET(cpumask);
577 /* Build our map of 'other' CPUs. */
578 PCPU_SET(other_cpus, all_cpus & ~PCPU_GET(cpumask));
583 if (smp_cpus == mp_ncpus) {
584 /* enable IPI's, tlb shootdown, freezes etc */
585 atomic_store_rel_int(&smp_started, 1);
586 smp_active = 1; /* historic */
590 * Enable global pages TLB extension
591 * This also implicitly flushes the TLB
594 load_cr4(rcr4() | CR4_PGE);
596 mtx_unlock_spin(&ap_boot_mtx);
598 /* wait until all the AP's are up */
599 while (smp_started == 0)
604 panic("scheduler returned us to %s", __func__);
608 /*******************************************************************
609 * local functions and data
613 * We tell the I/O APIC code about all the CPUs we want to receive
614 * interrupts. If we don't want certain CPUs to receive IRQs we
615 * can simply not tell the I/O APIC code about them in this function.
616 * We also do not tell it about the BSP since it tells itself about
617 * the BSP internally to work with UP kernels and on UP machines.
620 set_interrupt_apic_ids(void)
624 for (i = 0; i < MAXCPU; i++) {
625 apic_id = cpu_apic_ids[i];
628 if (cpu_info[apic_id].cpu_bsp)
630 if (cpu_info[apic_id].cpu_disabled)
633 /* Don't let hyperthreads service interrupts. */
634 if (hyperthreading_cpus > 1 &&
635 apic_id % hyperthreading_cpus != 0)
643 * Assign logical CPU IDs to local APICs.
650 TUNABLE_INT_FETCH("machdep.hyperthreading_allowed",
651 &hyperthreading_allowed);
653 /* Check for explicitly disabled CPUs. */
654 for (i = 0; i <= MAX_APIC_ID; i++) {
655 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp)
658 if (hyperthreading_cpus > 1 && i % hyperthreading_cpus != 0) {
659 cpu_info[i].cpu_hyperthread = 1;
660 #if defined(SCHED_ULE)
662 * Don't use HT CPU if it has been disabled by a
665 if (hyperthreading_allowed == 0) {
666 cpu_info[i].cpu_disabled = 1;
672 /* Don't use this CPU if it has been disabled by a tunable. */
673 if (resource_disabled("lapic", i)) {
674 cpu_info[i].cpu_disabled = 1;
680 * Assign CPU IDs to local APIC IDs and disable any CPUs
681 * beyond MAXCPU. CPU 0 has already been assigned to the BSP,
682 * so we only have to assign IDs for APs.
685 for (i = 0; i <= MAX_APIC_ID; i++) {
686 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp ||
687 cpu_info[i].cpu_disabled)
690 if (mp_ncpus < MAXCPU) {
691 cpu_apic_ids[mp_ncpus] = i;
694 cpu_info[i].cpu_disabled = 1;
696 KASSERT(mp_maxid >= mp_ncpus - 1,
697 ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid,
702 * start each AP in our list
707 vm_offset_t va = boot_address + KERNBASE;
708 u_int64_t *pt4, *pt3, *pt2;
709 u_int32_t mpbioswarmvec;
713 mtx_init(&ap_boot_mtx, "ap boot", NULL, MTX_SPIN);
715 /* install the AP 1st level boot code */
716 pmap_kenter(va, boot_address);
717 pmap_invalidate_page(kernel_pmap, va);
718 bcopy(mptramp_start, (void *)va, bootMP_size);
720 /* Locate the page tables, they'll be below the trampoline */
721 pt4 = (u_int64_t *)(uintptr_t)(mptramp_pagetables + KERNBASE);
722 pt3 = pt4 + (PAGE_SIZE) / sizeof(u_int64_t);
723 pt2 = pt3 + (PAGE_SIZE) / sizeof(u_int64_t);
725 /* Create the initial 1GB replicated page tables */
726 for (i = 0; i < 512; i++) {
727 /* Each slot of the level 4 pages points to the same level 3 page */
728 pt4[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + PAGE_SIZE);
729 pt4[i] |= PG_V | PG_RW | PG_U;
731 /* Each slot of the level 3 pages points to the same level 2 page */
732 pt3[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + (2 * PAGE_SIZE));
733 pt3[i] |= PG_V | PG_RW | PG_U;
735 /* The level 2 page slots are mapped with 2MB pages for 1GB. */
736 pt2[i] = i * (2 * 1024 * 1024);
737 pt2[i] |= PG_V | PG_RW | PG_PS | PG_U;
740 /* save the current value of the warm-start vector */
741 mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF);
742 outb(CMOS_REG, BIOS_RESET);
743 mpbiosreason = inb(CMOS_DATA);
745 /* setup a vector to our boot code */
746 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
747 *((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4);
748 outb(CMOS_REG, BIOS_RESET);
749 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
752 for (cpu = 1; cpu < mp_ncpus; cpu++) {
753 apic_id = cpu_apic_ids[cpu];
755 /* allocate and set up an idle stack data page */
756 bootstacks[cpu] = (void *)kmem_alloc(kernel_map, KSTACK_PAGES * PAGE_SIZE);
757 doublefault_stack = (char *)kmem_alloc(kernel_map, PAGE_SIZE);
759 bootSTK = (char *)bootstacks[cpu] + KSTACK_PAGES * PAGE_SIZE - 8;
762 /* attempt to start the Application Processor */
763 if (!start_ap(apic_id)) {
764 /* restore the warmstart vector */
765 *(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec;
766 panic("AP #%d (PHY# %d) failed!", cpu, apic_id);
769 all_cpus |= (1 << cpu); /* record AP in CPU map */
772 /* build our map of 'other' CPUs */
773 PCPU_SET(other_cpus, all_cpus & ~PCPU_GET(cpumask));
775 /* restore the warmstart vector */
776 *(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec;
778 outb(CMOS_REG, BIOS_RESET);
779 outb(CMOS_DATA, mpbiosreason);
781 /* number of APs actually started */
787 * This function starts the AP (application processor) identified
788 * by the APIC ID 'physicalCpu'. It does quite a "song and dance"
789 * to accomplish this. This is necessary because of the nuances
790 * of the different hardware we might encounter. It isn't pretty,
791 * but it seems to work.
794 start_ap(int apic_id)
799 /* calculate the vector */
800 vector = (boot_address >> 12) & 0xff;
802 /* used as a watchpoint to signal AP startup */
806 * first we do an INIT/RESET IPI this INIT IPI might be run, reseting
807 * and running the target CPU. OR this INIT IPI might be latched (P5
808 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
812 /* do an INIT IPI: assert RESET */
813 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
814 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id);
816 /* wait for pending status end */
819 /* do an INIT IPI: deassert RESET */
820 lapic_ipi_raw(APIC_DEST_ALLESELF | APIC_TRIGMOD_LEVEL |
821 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, 0);
823 /* wait for pending status end */
824 DELAY(10000); /* wait ~10mS */
828 * next we do a STARTUP IPI: the previous INIT IPI might still be
829 * latched, (P5 bug) this 1st STARTUP would then terminate
830 * immediately, and the previously started INIT IPI would continue. OR
831 * the previous INIT IPI has already run. and this STARTUP IPI will
832 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
836 /* do a STARTUP IPI */
837 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
838 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
841 DELAY(200); /* wait ~200uS */
844 * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
845 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
846 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
847 * recognized after hardware RESET or INIT IPI.
850 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
851 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
854 DELAY(200); /* wait ~200uS */
856 /* Wait up to 5 seconds for it to start. */
857 for (ms = 0; ms < 5000; ms++) {
859 return 1; /* return SUCCESS */
862 return 0; /* return FAILURE */
866 * Flush the TLB on all other CPU's
869 smp_tlb_shootdown(u_int vector, vm_offset_t addr1, vm_offset_t addr2)
873 ncpu = mp_ncpus - 1; /* does not shootdown self */
875 return; /* no other cpus */
876 if (!(read_rflags() & PSL_I))
877 panic("%s: interrupts disabled", __func__);
878 mtx_lock_spin(&smp_ipi_mtx);
879 smp_tlb_addr1 = addr1;
880 smp_tlb_addr2 = addr2;
881 atomic_store_rel_int(&smp_tlb_wait, 0);
882 ipi_all_but_self(vector);
883 while (smp_tlb_wait < ncpu)
885 mtx_unlock_spin(&smp_ipi_mtx);
889 smp_targeted_tlb_shootdown(u_int mask, u_int vector, vm_offset_t addr1, vm_offset_t addr2)
893 othercpus = mp_ncpus - 1;
894 if (mask == (u_int)-1) {
899 mask &= ~PCPU_GET(cpumask);
902 ncpu = bitcount32(mask);
903 if (ncpu > othercpus) {
904 /* XXX this should be a panic offence */
905 printf("SMP: tlb shootdown to %d other cpus (only have %d)\n",
909 /* XXX should be a panic, implied by mask == 0 above */
913 if (!(read_rflags() & PSL_I))
914 panic("%s: interrupts disabled", __func__);
915 mtx_lock_spin(&smp_ipi_mtx);
916 smp_tlb_addr1 = addr1;
917 smp_tlb_addr2 = addr2;
918 atomic_store_rel_int(&smp_tlb_wait, 0);
919 if (mask == (u_int)-1)
920 ipi_all_but_self(vector);
922 ipi_selected(mask, vector);
923 while (smp_tlb_wait < ncpu)
925 mtx_unlock_spin(&smp_ipi_mtx);
929 smp_cache_flush(void)
933 smp_tlb_shootdown(IPI_INVLCACHE, 0, 0);
941 smp_tlb_shootdown(IPI_INVLTLB, 0, 0);
946 smp_invlpg(vm_offset_t addr)
950 smp_tlb_shootdown(IPI_INVLPG, addr, 0);
954 smp_invlpg_range(vm_offset_t addr1, vm_offset_t addr2)
958 smp_tlb_shootdown(IPI_INVLRNG, addr1, addr2);
963 smp_masked_invltlb(u_int mask)
967 smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, 0, 0);
972 smp_masked_invlpg(u_int mask, vm_offset_t addr)
976 smp_targeted_tlb_shootdown(mask, IPI_INVLPG, addr, 0);
981 smp_masked_invlpg_range(u_int mask, vm_offset_t addr1, vm_offset_t addr2)
985 smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, addr1, addr2);
990 ipi_bitmap_handler(struct trapframe frame)
992 int cpu = PCPU_GET(cpuid);
995 ipi_bitmap = atomic_readandclear_int(&cpu_ipi_pending[cpu]);
997 if (ipi_bitmap & (1 << IPI_PREEMPT)) {
998 struct thread *running_thread = curthread;
999 thread_lock(running_thread);
1000 if (running_thread->td_critnest > 1)
1001 running_thread->td_owepreempt = 1;
1003 mi_switch(SW_INVOL | SW_PREEMPT, NULL);
1004 thread_unlock(running_thread);
1007 /* Nothing to do for AST */
1011 * send an IPI to a set of cpus.
1014 ipi_selected(u_int32_t cpus, u_int ipi)
1021 if (IPI_IS_BITMAPED(ipi)) {
1023 ipi = IPI_BITMAP_VECTOR;
1027 if (ipi == IPI_STOP && stop_cpus_with_nmi) {
1028 ipi_nmi_selected(cpus);
1032 CTR3(KTR_SMP, "%s: cpus: %x ipi: %x", __func__, cpus, ipi);
1033 while ((cpu = ffs(cpus)) != 0) {
1035 cpus &= ~(1 << cpu);
1037 KASSERT(cpu_apic_ids[cpu] != -1,
1038 ("IPI to non-existent CPU %d", cpu));
1042 old_pending = cpu_ipi_pending[cpu];
1043 new_pending = old_pending | bitmap;
1044 } while (!atomic_cmpset_int(&cpu_ipi_pending[cpu],old_pending, new_pending));
1050 lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]);
1056 * send an IPI INTerrupt containing 'vector' to all CPUs, including myself
1062 if (IPI_IS_BITMAPED(ipi) || (ipi == IPI_STOP && stop_cpus_with_nmi)) {
1063 ipi_selected(all_cpus, ipi);
1066 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1067 lapic_ipi_vectored(ipi, APIC_IPI_DEST_ALL);
1071 * send an IPI to all CPUs EXCEPT myself
1074 ipi_all_but_self(u_int ipi)
1077 if (IPI_IS_BITMAPED(ipi) || (ipi == IPI_STOP && stop_cpus_with_nmi)) {
1078 ipi_selected(PCPU_GET(other_cpus), ipi);
1081 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1082 lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS);
1086 * send an IPI to myself
1092 if (IPI_IS_BITMAPED(ipi) || (ipi == IPI_STOP && stop_cpus_with_nmi)) {
1093 ipi_selected(PCPU_GET(cpumask), ipi);
1096 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1097 lapic_ipi_vectored(ipi, APIC_IPI_DEST_SELF);
1102 * send NMI IPI to selected CPUs
1105 #define BEFORE_SPIN 1000000
1108 ipi_nmi_selected(u_int32_t cpus)
1113 icrlo = APIC_DELMODE_NMI | APIC_DESTMODE_PHY | APIC_LEVEL_ASSERT
1114 | APIC_TRIGMOD_EDGE;
1116 CTR2(KTR_SMP, "%s: cpus: %x nmi", __func__, cpus);
1118 atomic_set_int(&ipi_nmi_pending, cpus);
1120 while ((cpu = ffs(cpus)) != 0) {
1122 cpus &= ~(1 << cpu);
1124 KASSERT(cpu_apic_ids[cpu] != -1,
1125 ("IPI NMI to non-existent CPU %d", cpu));
1127 /* Wait for an earlier IPI to finish. */
1128 if (!lapic_ipi_wait(BEFORE_SPIN))
1129 panic("ipi_nmi_selected: previous IPI has not cleared");
1131 lapic_ipi_raw(icrlo, cpu_apic_ids[cpu]);
1136 ipi_nmi_handler(void)
1138 int cpumask = PCPU_GET(cpumask);
1140 if (!(ipi_nmi_pending & cpumask))
1143 atomic_clear_int(&ipi_nmi_pending, cpumask);
1148 #endif /* STOP_NMI */
1151 * Handle an IPI_STOP by saving our current context and spinning until we
1155 cpustop_handler(void)
1157 int cpu = PCPU_GET(cpuid);
1158 int cpumask = PCPU_GET(cpumask);
1160 savectx(&stoppcbs[cpu]);
1162 /* Indicate that we are stopped */
1163 atomic_set_int(&stopped_cpus, cpumask);
1165 /* Wait for restart */
1166 while (!(started_cpus & cpumask))
1169 atomic_clear_int(&started_cpus, cpumask);
1170 atomic_clear_int(&stopped_cpus, cpumask);
1172 if (cpu == 0 && cpustop_restartfunc != NULL) {
1173 cpustop_restartfunc();
1174 cpustop_restartfunc = NULL;
1179 * This is called once the rest of the system is up and running and we're
1180 * ready to let the AP's out of the pen.
1183 release_aps(void *dummy __unused)
1188 atomic_store_rel_int(&aps_ready, 1);
1189 while (smp_started == 0)
1192 SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL);
1195 sysctl_hlt_cpus(SYSCTL_HANDLER_ARGS)
1200 mask = hlt_cpus_mask;
1201 error = sysctl_handle_int(oidp, &mask, 0, req);
1202 if (error || !req->newptr)
1205 if (logical_cpus_mask != 0 &&
1206 (mask & logical_cpus_mask) == logical_cpus_mask)
1207 hlt_logical_cpus = 1;
1209 hlt_logical_cpus = 0;
1211 if (! hyperthreading_allowed)
1212 mask |= hyperthreading_cpus_mask;
1214 if ((mask & all_cpus) == all_cpus)
1216 hlt_cpus_mask = mask;
1219 SYSCTL_PROC(_machdep, OID_AUTO, hlt_cpus, CTLTYPE_INT|CTLFLAG_RW,
1220 0, 0, sysctl_hlt_cpus, "IU",
1221 "Bitmap of CPUs to halt. 101 (binary) will halt CPUs 0 and 2.");
1224 sysctl_hlt_logical_cpus(SYSCTL_HANDLER_ARGS)
1228 disable = hlt_logical_cpus;
1229 error = sysctl_handle_int(oidp, &disable, 0, req);
1230 if (error || !req->newptr)
1234 hlt_cpus_mask |= logical_cpus_mask;
1236 hlt_cpus_mask &= ~logical_cpus_mask;
1238 if (! hyperthreading_allowed)
1239 hlt_cpus_mask |= hyperthreading_cpus_mask;
1241 if ((hlt_cpus_mask & all_cpus) == all_cpus)
1242 hlt_cpus_mask &= ~(1<<0);
1244 hlt_logical_cpus = disable;
1249 sysctl_hyperthreading_allowed(SYSCTL_HANDLER_ARGS)
1253 allowed = hyperthreading_allowed;
1254 error = sysctl_handle_int(oidp, &allowed, 0, req);
1255 if (error || !req->newptr)
1260 * SCHED_ULE doesn't allow enabling/disabling HT cores at
1263 if (allowed != hyperthreading_allowed)
1269 hlt_cpus_mask &= ~hyperthreading_cpus_mask;
1271 hlt_cpus_mask |= hyperthreading_cpus_mask;
1273 if (logical_cpus_mask != 0 &&
1274 (hlt_cpus_mask & logical_cpus_mask) == logical_cpus_mask)
1275 hlt_logical_cpus = 1;
1277 hlt_logical_cpus = 0;
1279 if ((hlt_cpus_mask & all_cpus) == all_cpus)
1280 hlt_cpus_mask &= ~(1<<0);
1282 hyperthreading_allowed = allowed;
1287 cpu_hlt_setup(void *dummy __unused)
1290 if (logical_cpus_mask != 0) {
1291 TUNABLE_INT_FETCH("machdep.hlt_logical_cpus",
1293 sysctl_ctx_init(&logical_cpu_clist);
1294 SYSCTL_ADD_PROC(&logical_cpu_clist,
1295 SYSCTL_STATIC_CHILDREN(_machdep), OID_AUTO,
1296 "hlt_logical_cpus", CTLTYPE_INT|CTLFLAG_RW, 0, 0,
1297 sysctl_hlt_logical_cpus, "IU", "");
1298 SYSCTL_ADD_UINT(&logical_cpu_clist,
1299 SYSCTL_STATIC_CHILDREN(_machdep), OID_AUTO,
1300 "logical_cpus_mask", CTLTYPE_INT|CTLFLAG_RD,
1301 &logical_cpus_mask, 0, "");
1303 if (hlt_logical_cpus)
1304 hlt_cpus_mask |= logical_cpus_mask;
1307 * If necessary for security purposes, force
1308 * hyperthreading off, regardless of the value
1309 * of hlt_logical_cpus.
1311 if (hyperthreading_cpus_mask) {
1312 SYSCTL_ADD_PROC(&logical_cpu_clist,
1313 SYSCTL_STATIC_CHILDREN(_machdep), OID_AUTO,
1314 "hyperthreading_allowed", CTLTYPE_INT|CTLFLAG_RW,
1315 0, 0, sysctl_hyperthreading_allowed, "IU", "");
1316 if (! hyperthreading_allowed)
1317 hlt_cpus_mask |= hyperthreading_cpus_mask;
1321 SYSINIT(cpu_hlt, SI_SUB_SMP, SI_ORDER_ANY, cpu_hlt_setup, NULL);
1324 mp_grab_cpu_hlt(void)
1326 u_int mask = PCPU_GET(cpumask);
1328 u_int cpuid = PCPU_GET(cpuid);
1336 retval = mask & hlt_cpus_mask;
1337 while (mask & hlt_cpus_mask)
1338 __asm __volatile("sti; hlt" : : : "memory");