2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 1996, by Steve Passe
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. The name of the developer may NOT be used to endorse or promote products
13 * derived from this software without specific prior written permission.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28 #include <sys/cdefs.h>
29 __FBSDID("$FreeBSD$");
34 #include "opt_kstack_pages.h"
36 #include "opt_sched.h"
41 #error How did you get here?
45 #error The apic device is required for SMP, add "device apic" to your config file.
49 #include <sys/param.h>
50 #include <sys/systm.h>
52 #include <sys/cons.h> /* cngetc() */
53 #include <sys/cpuset.h>
58 #include <sys/kernel.h>
61 #include <sys/malloc.h>
62 #include <sys/memrange.h>
63 #include <sys/mutex.h>
66 #include <sys/sched.h>
68 #include <sys/sysctl.h>
71 #include <vm/vm_param.h>
73 #include <vm/vm_kern.h>
74 #include <vm/vm_extern.h>
76 #include <x86/apicreg.h>
77 #include <machine/clock.h>
78 #include <machine/cpu.h>
79 #include <machine/cputypes.h>
81 #include <machine/md_var.h>
82 #include <machine/pcb.h>
83 #include <machine/psl.h>
84 #include <machine/smp.h>
85 #include <machine/specialreg.h>
86 #include <x86/ucode.h>
89 #include <contrib/dev/acpica/include/acpi.h>
90 #include <dev/acpica/acpivar.h>
93 #define WARMBOOT_TARGET 0
94 #define WARMBOOT_OFF (PMAP_MAP_LOW + 0x0467)
95 #define WARMBOOT_SEG (PMAP_MAP_LOW + 0x0469)
97 #define CMOS_REG (0x70)
98 #define CMOS_DATA (0x71)
99 #define BIOS_RESET (0x0f)
100 #define BIOS_WARM (0x0a)
103 * this code MUST be enabled here and in mpboot.s.
104 * it follows the very early stages of AP boot by placing values in CMOS ram.
105 * it NORMALLY will never be needed and thus the primitive method for enabling.
110 #if defined(CHECK_POINTS)
111 #define CHECK_READ(A) (outb(CMOS_REG, (A)), inb(CMOS_DATA))
112 #define CHECK_WRITE(A,D) (outb(CMOS_REG, (A)), outb(CMOS_DATA, (D)))
114 #define CHECK_INIT(D); \
115 CHECK_WRITE(0x34, (D)); \
116 CHECK_WRITE(0x35, (D)); \
117 CHECK_WRITE(0x36, (D)); \
118 CHECK_WRITE(0x37, (D)); \
119 CHECK_WRITE(0x38, (D)); \
120 CHECK_WRITE(0x39, (D));
122 #define CHECK_PRINT(S); \
123 printf("%s: %d, %d, %d, %d, %d, %d\n", \
132 #else /* CHECK_POINTS */
134 #define CHECK_INIT(D)
135 #define CHECK_PRINT(S)
136 #define CHECK_WRITE(A, D)
138 #endif /* CHECK_POINTS */
141 * Local data and functions.
144 static void install_ap_tramp(void);
145 static int start_all_aps(void);
146 static int start_ap(int apic_id);
148 static char *ap_copyout_buf;
149 static char *ap_tramp_stack_base;
151 unsigned int boot_address;
153 #define MiB(v) (v ## ULL << 20)
155 /* Allocate memory for the AP trampoline. */
157 alloc_ap_trampoline(vm_paddr_t *physmap, unsigned int *physmap_idx)
163 for (i = *physmap_idx; i <= *physmap_idx; i -= 2) {
165 * Find a memory region big enough and below the 1MB boundary
166 * for the trampoline code.
167 * NB: needs to be page aligned.
169 if (physmap[i] >= MiB(1) ||
170 (trunc_page(physmap[i + 1]) - round_page(physmap[i])) <
171 round_page(bootMP_size))
176 * Try to steal from the end of the region to mimic previous
177 * behaviour, else fallback to steal from the start.
179 if (physmap[i + 1] < MiB(1)) {
180 boot_address = trunc_page(physmap[i + 1]);
181 if ((physmap[i + 1] - boot_address) < bootMP_size)
182 boot_address -= round_page(bootMP_size);
183 physmap[i + 1] = boot_address;
185 boot_address = round_page(physmap[i]);
186 physmap[i] = boot_address + round_page(bootMP_size);
188 if (physmap[i] == physmap[i + 1] && *physmap_idx != 0) {
189 memmove(&physmap[i], &physmap[i + 2],
190 sizeof(*physmap) * (*physmap_idx - i + 2));
197 boot_address = basemem * 1024 - bootMP_size;
200 "Cannot find enough space for the boot trampoline, placing it at %#x",
206 * Initialize the IPI handlers and start up the AP's.
213 /* Initialize the logical ID to APIC ID table. */
214 for (i = 0; i < MAXCPU; i++) {
215 cpu_apic_ids[i] = -1;
218 /* Install an inter-CPU IPI for TLB invalidation */
219 setidt(IPI_INVLTLB, IDTVEC(invltlb),
220 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
221 setidt(IPI_INVLPG, IDTVEC(invlpg),
222 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
223 setidt(IPI_INVLRNG, IDTVEC(invlrng),
224 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
226 /* Install an inter-CPU IPI for cache invalidation. */
227 setidt(IPI_INVLCACHE, IDTVEC(invlcache),
228 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
230 /* Install an inter-CPU IPI for all-CPU rendezvous */
231 setidt(IPI_RENDEZVOUS, IDTVEC(rendezvous),
232 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
234 /* Install generic inter-CPU IPI handler */
235 setidt(IPI_BITMAP_VECTOR, IDTVEC(ipi_intr_bitmap_handler),
236 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
238 /* Install an inter-CPU IPI for CPU stop/restart */
239 setidt(IPI_STOP, IDTVEC(cpustop),
240 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
242 /* Install an inter-CPU IPI for CPU suspend/resume */
243 setidt(IPI_SUSPEND, IDTVEC(cpususpend),
244 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
246 /* Install an IPI for calling delayed SWI */
247 setidt(IPI_SWI, IDTVEC(ipi_swi),
248 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
250 /* Set boot_cpu_id if needed. */
251 if (boot_cpu_id == -1) {
252 boot_cpu_id = PCPU_GET(apic_id);
253 cpu_info[boot_cpu_id].cpu_bsp = 1;
255 KASSERT(boot_cpu_id == PCPU_GET(apic_id),
256 ("BSP's APIC ID doesn't match boot_cpu_id"));
258 /* Probe logical/physical core configuration. */
263 /* Start each Application Processor */
266 set_interrupt_apic_ids();
268 #if defined(DEV_ACPI) && MAXMEMDOM > 1
269 acpi_pxm_set_cpu_locality();
274 * AP CPU's call this to initialize themselves.
280 struct i386tss *common_tssp;
281 struct region_descriptor r_gdt, r_idt;
282 int gsel_tss, myid, x;
285 /* bootAP is set in start_ap() to our ID. */
288 /* Update microcode before doing anything else. */
291 /* Get per-cpu data */
294 /* prime data page for it to use */
295 pcpu_init(pc, myid, sizeof(struct pcpu));
296 dpcpu_init(dpcpu, myid);
297 pc->pc_apic_id = cpu_apic_ids[myid];
298 pc->pc_prvspace = pc;
299 pc->pc_curthread = 0;
300 pc->pc_common_tssp = common_tssp = &(__pcpu[0].pc_common_tssp)[myid];
304 gdt_segs[GPRIV_SEL].ssd_base = (int)pc;
305 gdt_segs[GPROC0_SEL].ssd_base = (int)common_tssp;
306 gdt_segs[GLDT_SEL].ssd_base = (int)ldt;
308 for (x = 0; x < NGDT; x++) {
309 ssdtosd(&gdt_segs[x], &gdt[myid * NGDT + x].sd);
312 r_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
313 r_gdt.rd_base = (int) &gdt[myid * NGDT];
314 lgdt(&r_gdt); /* does magic intra-segment return */
316 r_idt.rd_limit = sizeof(struct gate_descriptor) * NIDT - 1;
317 r_idt.rd_base = (int)idt;
321 PCPU_SET(currentldt, _default_ldt);
323 PCPU_SET(trampstk, (uintptr_t)ap_tramp_stack_base + TRAMP_STACK_SZ -
326 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
327 gdt[myid * NGDT + GPROC0_SEL].sd.sd_type = SDT_SYS386TSS;
328 common_tssp->tss_esp0 = PCPU_GET(trampstk);
329 common_tssp->tss_ss0 = GSEL(GDATA_SEL, SEL_KPL);
330 common_tssp->tss_ioopt = sizeof(struct i386tss) << 16;
331 PCPU_SET(tss_gdt, &gdt[myid * NGDT + GPROC0_SEL].sd);
332 PCPU_SET(common_tssd, *PCPU_GET(tss_gdt));
335 PCPU_SET(fsgs_gdt, &gdt[myid * NGDT + GUFS_SEL].sd);
336 PCPU_SET(copyout_buf, ap_copyout_buf);
339 * Set to a known state:
340 * Set by mpboot.s: CR0_PG, CR0_PE
341 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
344 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
346 CHECK_WRITE(0x38, 5);
348 /* signal our startup to the BSP. */
350 CHECK_WRITE(0x39, 6);
352 /* Spin until the BSP releases the AP's. */
353 while (atomic_load_acq_int(&aps_ready) == 0)
356 /* BSP may have changed PTD while we were waiting */
359 #if defined(I586_CPU) && !defined(NO_F00F_HACK)
363 init_secondary_tail();
367 * start each AP in our list
369 #define TMPMAP_START 1
374 u_int32_t mpbioswarmvec;
377 mtx_init(&ap_boot_mtx, "ap boot", NULL, MTX_SPIN);
379 pmap_remap_lower(true);
381 /* install the AP 1st level boot code */
384 /* save the current value of the warm-start vector */
385 mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF);
386 outb(CMOS_REG, BIOS_RESET);
387 mpbiosreason = inb(CMOS_DATA);
389 /* take advantage of the P==V mapping for PTD[0] for AP boot */
392 for (cpu = 1; cpu < mp_ncpus; cpu++) {
393 apic_id = cpu_apic_ids[cpu];
395 /* allocate and set up a boot stack data page */
396 bootstacks[cpu] = (char *)kmem_malloc(kstack_pages * PAGE_SIZE,
398 dpcpu = (void *)kmem_malloc(DPCPU_SIZE, M_WAITOK | M_ZERO);
399 /* setup a vector to our boot code */
400 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
401 *((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4);
402 outb(CMOS_REG, BIOS_RESET);
403 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
405 bootSTK = (char *)bootstacks[cpu] + kstack_pages *
409 ap_tramp_stack_base = pmap_trm_alloc(TRAMP_STACK_SZ, M_NOWAIT);
410 ap_copyout_buf = pmap_trm_alloc(TRAMP_COPYOUT_SZ, M_NOWAIT);
412 /* attempt to start the Application Processor */
413 CHECK_INIT(99); /* setup checkpoints */
414 if (!start_ap(apic_id)) {
415 printf("AP #%d (PHY# %d) failed!\n", cpu, apic_id);
416 CHECK_PRINT("trace"); /* show checkpoints */
417 /* better panic as the AP may be running loose */
418 printf("panic y/n? [y] ");
422 CHECK_PRINT("trace"); /* show checkpoints */
424 CPU_SET(cpu, &all_cpus); /* record AP in CPU map */
427 pmap_remap_lower(false);
429 /* restore the warmstart vector */
430 *(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec;
432 outb(CMOS_REG, BIOS_RESET);
433 outb(CMOS_DATA, mpbiosreason);
435 /* number of APs actually started */
440 * load the 1st level AP boot code into base memory.
443 /* targets for relocation */
444 extern void bigJump(void);
445 extern void bootCodeSeg(void);
446 extern void bootDataSeg(void);
447 extern void MPentry(void);
449 extern u_int mp_gdtbase;
452 install_ap_tramp(void)
455 int size = *(int *) ((u_long) & bootMP_size);
456 vm_offset_t va = boot_address;
457 u_char *src = (u_char *) ((u_long) bootMP);
458 u_char *dst = (u_char *) va;
459 u_int boot_base = (u_int) bootMP;
464 KASSERT (size <= PAGE_SIZE,
465 ("'size' do not fit into PAGE_SIZE, as expected."));
466 pmap_kenter(va, boot_address);
467 pmap_invalidate_page (kernel_pmap, va);
468 for (x = 0; x < size; ++x)
472 * modify addresses in code we just moved to basemem. unfortunately we
473 * need fairly detailed info about mpboot.s for this to work. changes
474 * to mpboot.s might require changes here.
477 /* boot code is located in KERNEL space */
480 /* modify the lgdt arg */
481 dst32 = (u_int32_t *) (dst + ((u_int) & mp_gdtbase - boot_base));
482 *dst32 = boot_address + ((u_int) & MP_GDT - boot_base);
484 /* modify the ljmp target for MPentry() */
485 dst32 = (u_int32_t *) (dst + ((u_int) bigJump - boot_base) + 1);
486 *dst32 = (u_int)MPentry;
488 /* modify the target for boot code segment */
489 dst16 = (u_int16_t *) (dst + ((u_int) bootCodeSeg - boot_base));
490 dst8 = (u_int8_t *) (dst16 + 1);
491 *dst16 = (u_int) boot_address & 0xffff;
492 *dst8 = ((u_int) boot_address >> 16) & 0xff;
494 /* modify the target for boot data segment */
495 dst16 = (u_int16_t *) (dst + ((u_int) bootDataSeg - boot_base));
496 dst8 = (u_int8_t *) (dst16 + 1);
497 *dst16 = (u_int) boot_address & 0xffff;
498 *dst8 = ((u_int) boot_address >> 16) & 0xff;
502 * This function starts the AP (application processor) identified
503 * by the APIC ID 'physicalCpu'. It does quite a "song and dance"
504 * to accomplish this. This is necessary because of the nuances
505 * of the different hardware we might encounter. It isn't pretty,
506 * but it seems to work.
509 start_ap(int apic_id)
514 /* calculate the vector */
515 vector = (boot_address >> 12) & 0xff;
517 /* used as a watchpoint to signal AP startup */
520 ipi_startup(apic_id, vector);
522 /* Wait up to 5 seconds for it to start. */
523 for (ms = 0; ms < 5000; ms++) {
525 return 1; /* return SUCCESS */
528 return 0; /* return FAILURE */
532 * Flush the TLB on other CPU's
535 /* Variables needed for SMP tlb shootdown. */
536 vm_offset_t smp_tlb_addr1, smp_tlb_addr2;
538 volatile uint32_t smp_tlb_generation;
541 * Used by pmap to request cache or TLB invalidation on local and
542 * remote processors. Mask provides the set of remote CPUs which are
543 * to be signalled with the invalidation IPI. Vector specifies which
544 * invalidation IPI is used. As an optimization, the curcpu_cb
545 * callback is invoked on the calling CPU while waiting for remote
546 * CPUs to complete the operation.
548 * The callback function is called unconditionally on the caller's
549 * underlying processor, even when this processor is not set in the
550 * mask. So, the callback function must be prepared to handle such
551 * spurious invocations.
554 smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, pmap_t pmap,
555 vm_offset_t addr1, vm_offset_t addr2, smp_invl_cb_t curcpu_cb)
558 volatile uint32_t *p_cpudone;
563 * It is not necessary to signal other CPUs while booting or
564 * when in the debugger.
566 if (kdb_active || KERNEL_PANICKED() || !smp_started) {
567 curcpu_cb(pmap, addr1, addr2);
574 * Check for other cpus. Return if none.
576 if (CPU_ISFULLSET(&mask)) {
580 CPU_CLR(PCPU_GET(cpuid), &mask);
581 if (CPU_EMPTY(&mask))
585 KASSERT((read_eflags() & PSL_I) != 0,
586 ("smp_targeted_tlb_shootdown: interrupts disabled"));
587 mtx_lock_spin(&smp_ipi_mtx);
588 smp_tlb_addr1 = addr1;
589 smp_tlb_addr2 = addr2;
591 generation = ++smp_tlb_generation;
592 if (CPU_ISFULLSET(&mask)) {
593 ipi_all_but_self(vector);
594 other_cpus = all_cpus;
595 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
598 ipi_selected(mask, vector);
600 curcpu_cb(pmap, addr1, addr2);
601 CPU_FOREACH_ISSET(cpu, &other_cpus) {
602 p_cpudone = &cpuid_to_pcpu[cpu]->pc_smp_tlb_done;
603 while (*p_cpudone != generation)
606 mtx_unlock_spin(&smp_ipi_mtx);
611 curcpu_cb(pmap, addr1, addr2);
616 smp_masked_invltlb(cpuset_t mask, pmap_t pmap, smp_invl_cb_t curcpu_cb)
618 smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, pmap, 0, 0, curcpu_cb);
619 #ifdef COUNT_XINVLTLB_HITS
625 smp_masked_invlpg(cpuset_t mask, vm_offset_t addr, pmap_t pmap,
626 smp_invl_cb_t curcpu_cb)
628 smp_targeted_tlb_shootdown(mask, IPI_INVLPG, pmap, addr, 0, curcpu_cb);
629 #ifdef COUNT_XINVLTLB_HITS
635 smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2,
636 pmap_t pmap, smp_invl_cb_t curcpu_cb)
638 smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, pmap, addr1, addr2,
640 #ifdef COUNT_XINVLTLB_HITS
642 ipi_range_size += (addr2 - addr1) / PAGE_SIZE;
647 smp_cache_flush(smp_invl_cb_t curcpu_cb)
649 smp_targeted_tlb_shootdown(all_cpus, IPI_INVLCACHE, NULL, 0, 0,
654 * Handlers for TLB related IPIs
657 invltlb_handler(void)
662 #ifdef COUNT_XINVLTLB_HITS
663 xhits_gbl[PCPU_GET(cpuid)]++;
664 #endif /* COUNT_XINVLTLB_HITS */
666 (*ipi_invltlb_counts[PCPU_GET(cpuid)])++;
667 #endif /* COUNT_IPIS */
670 * Reading the generation here allows greater parallelism
671 * since invalidating the TLB is a serializing operation.
673 generation = smp_tlb_generation;
674 if (smp_tlb_pmap == kernel_pmap)
676 PCPU_SET(smp_tlb_done, generation);
685 #ifdef COUNT_XINVLTLB_HITS
686 xhits_pg[PCPU_GET(cpuid)]++;
687 #endif /* COUNT_XINVLTLB_HITS */
689 (*ipi_invlpg_counts[PCPU_GET(cpuid)])++;
690 #endif /* COUNT_IPIS */
692 generation = smp_tlb_generation; /* Overlap with serialization */
693 if (smp_tlb_pmap == kernel_pmap)
694 invlpg(smp_tlb_addr1);
695 PCPU_SET(smp_tlb_done, generation);
699 invlrng_handler(void)
701 vm_offset_t addr, addr2;
705 #ifdef COUNT_XINVLTLB_HITS
706 xhits_rng[PCPU_GET(cpuid)]++;
707 #endif /* COUNT_XINVLTLB_HITS */
709 (*ipi_invlrng_counts[PCPU_GET(cpuid)])++;
710 #endif /* COUNT_IPIS */
712 addr = smp_tlb_addr1;
713 addr2 = smp_tlb_addr2;
714 generation = smp_tlb_generation; /* Overlap with serialization */
715 if (smp_tlb_pmap == kernel_pmap) {
719 } while (addr < addr2);
722 PCPU_SET(smp_tlb_done, generation);
726 invlcache_handler(void)
732 (*ipi_invlcache_counts[PCPU_GET(cpuid)])++;
733 #endif /* COUNT_IPIS */
736 * Reading the generation here allows greater parallelism
737 * since wbinvd is a serializing instruction. Without the
738 * temporary, we'd wait for wbinvd to complete, then the read
739 * would execute, then the dependent write, which must then
740 * complete before return from interrupt.
742 generation = smp_tlb_generation;
744 PCPU_SET(smp_tlb_done, generation);