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
25 * $FreeBSD: src/sys/i386/i386/mp_machdep.c,v 1.115.2.15 2003/03/14 21:22:35 jhb Exp $
26 * $DragonFly: src/sys/platform/pc32/i386/mp_machdep.c,v 1.60 2008/06/07 12:03:52 mneumann Exp $
31 #include <sys/param.h>
32 #include <sys/systm.h>
33 #include <sys/kernel.h>
34 #include <sys/sysctl.h>
35 #include <sys/malloc.h>
36 #include <sys/memrange.h>
37 #include <sys/cons.h> /* cngetc() */
38 #include <sys/machintr.h>
40 #include <sys/mplock2.h>
43 #include <vm/vm_param.h>
45 #include <vm/vm_kern.h>
46 #include <vm/vm_extern.h>
48 #include <vm/vm_map.h>
54 #include <machine/smp.h>
55 #include <machine_base/apic/apicreg.h>
56 #include <machine/atomic.h>
57 #include <machine/cpufunc.h>
58 #include <machine_base/apic/mpapic.h>
59 #include <machine/psl.h>
60 #include <machine/segments.h>
61 #include <machine/tss.h>
62 #include <machine/specialreg.h>
63 #include <machine/globaldata.h>
65 #include <machine/md_var.h> /* setidt() */
66 #include <machine_base/icu/icu.h> /* IPIs */
67 #include <machine_base/isa/intr_machdep.h> /* IPIs */
69 #define FIXUP_EXTRA_APIC_INTS 8 /* additional entries we may create */
71 #define WARMBOOT_TARGET 0
72 #define WARMBOOT_OFF (KERNBASE + 0x0467)
73 #define WARMBOOT_SEG (KERNBASE + 0x0469)
75 #define BIOS_BASE (0xf0000)
76 #define BIOS_SIZE (0x10000)
77 #define BIOS_COUNT (BIOS_SIZE/4)
79 #define CMOS_REG (0x70)
80 #define CMOS_DATA (0x71)
81 #define BIOS_RESET (0x0f)
82 #define BIOS_WARM (0x0a)
84 #define PROCENTRY_FLAG_EN 0x01
85 #define PROCENTRY_FLAG_BP 0x02
86 #define IOAPICENTRY_FLAG_EN 0x01
89 /* MP Floating Pointer Structure */
90 typedef struct MPFPS {
103 /* MP Configuration Table Header */
104 typedef struct MPCTH {
106 u_short base_table_length;
110 u_char product_id[12];
111 u_int32_t oem_table_pointer;
112 u_short oem_table_size;
114 u_int32_t apic_address;
115 u_short extended_table_length;
116 u_char extended_table_checksum;
121 typedef struct PROCENTRY {
126 u_int32_t cpu_signature;
127 u_int32_t feature_flags;
132 typedef struct BUSENTRY {
138 typedef struct IOAPICENTRY {
143 u_int32_t apic_address;
144 } *io_apic_entry_ptr;
146 typedef struct INTENTRY {
156 /* descriptions of MP basetable entries */
157 typedef struct BASETABLE_ENTRY {
166 vm_size_t mp_cth_mapsz;
169 typedef int (*mptable_iter_func)(void *, const void *, int);
172 * this code MUST be enabled here and in mpboot.s.
173 * it follows the very early stages of AP boot by placing values in CMOS ram.
174 * it NORMALLY will never be needed and thus the primitive method for enabling.
177 #if defined(CHECK_POINTS)
178 #define CHECK_READ(A) (outb(CMOS_REG, (A)), inb(CMOS_DATA))
179 #define CHECK_WRITE(A,D) (outb(CMOS_REG, (A)), outb(CMOS_DATA, (D)))
181 #define CHECK_INIT(D); \
182 CHECK_WRITE(0x34, (D)); \
183 CHECK_WRITE(0x35, (D)); \
184 CHECK_WRITE(0x36, (D)); \
185 CHECK_WRITE(0x37, (D)); \
186 CHECK_WRITE(0x38, (D)); \
187 CHECK_WRITE(0x39, (D));
189 #define CHECK_PRINT(S); \
190 kprintf("%s: %d, %d, %d, %d, %d, %d\n", \
199 #else /* CHECK_POINTS */
201 #define CHECK_INIT(D)
202 #define CHECK_PRINT(S)
204 #endif /* CHECK_POINTS */
207 * Values to send to the POST hardware.
209 #define MP_BOOTADDRESS_POST 0x10
210 #define MP_PROBE_POST 0x11
211 #define MPTABLE_PASS1_POST 0x12
213 #define MP_START_POST 0x13
214 #define MP_ENABLE_POST 0x14
215 #define MPTABLE_PASS2_POST 0x15
217 #define START_ALL_APS_POST 0x16
218 #define INSTALL_AP_TRAMP_POST 0x17
219 #define START_AP_POST 0x18
221 #define MP_ANNOUNCE_POST 0x19
223 static int madt_probe_test;
224 TUNABLE_INT("hw.madt_probe_test", &madt_probe_test);
226 /** XXX FIXME: where does this really belong, isa.h/isa.c perhaps? */
227 int current_postcode;
229 /** XXX FIXME: what system files declare these??? */
230 extern struct region_descriptor r_gdt, r_idt;
232 int mp_naps; /* # of Applications processors */
234 static int mp_nbusses; /* # of busses */
235 int mp_napics; /* # of IO APICs */
237 static vm_offset_t cpu_apic_address;
239 vm_offset_t io_apic_address[NAPICID]; /* NAPICID is more than enough */
240 u_int32_t *io_apic_versions;
244 u_int32_t cpu_apic_versions[MAXCPU];
246 extern int64_t tsc_offsets[];
248 extern u_long ebda_addr;
251 struct apic_intmapinfo int_to_apicintpin[APIC_INTMAPSIZE];
255 * APIC ID logical/physical mapping structures.
256 * We oversize these to simplify boot-time config.
258 int cpu_num_to_apic_id[NAPICID];
260 int io_num_to_apic_id[NAPICID];
262 int apic_id_to_logical[NAPICID];
264 /* AP uses this during bootstrap. Do not staticize. */
269 * SMP page table page. Setup by locore to point to a page table
270 * page from which we allocate per-cpu privatespace areas io_apics,
274 #define IO_MAPPING_START_INDEX \
275 (SMP_MAXCPU * sizeof(struct privatespace) / PAGE_SIZE)
277 extern pt_entry_t *SMPpt;
279 struct pcb stoppcbs[MAXCPU];
281 extern inthand_t IDTVEC(fast_syscall), IDTVEC(fast_syscall32);
283 static basetable_entry basetable_entry_types[] =
285 {0, 20, "Processor"},
293 * Local data and functions.
296 static u_int boot_address;
297 static u_int base_memory;
298 static int mp_finish;
300 static void mp_enable(u_int boot_addr);
302 static int mptable_iterate_entries(const mpcth_t,
303 mptable_iter_func, void *);
304 static int mptable_probe(void);
305 static int mptable_check(vm_paddr_t);
306 static long mptable_search_sig(u_int32_t target, int count);
307 static int mptable_hyperthread_fixup(u_int, int);
308 static void mptable_pass1(struct mptable_pos *);
309 static void mptable_pass2(struct mptable_pos *);
310 static void mptable_default(int type);
311 static void mptable_fix(void);
312 static int mptable_map(struct mptable_pos *, vm_paddr_t);
313 static void mptable_unmap(struct mptable_pos *);
314 static void mptable_lapic_enumerate(struct mptable_pos *);
315 static void mptable_lapic_default(void);
318 static void setup_apic_irq_mapping(void);
319 static int apic_int_is_bus_type(int intr, int bus_type);
321 static int start_all_aps(u_int boot_addr);
323 static void install_ap_tramp(u_int boot_addr);
325 static int start_ap(struct mdglobaldata *gd, u_int boot_addr, int smibest);
326 static void lapic_init(vm_offset_t);
327 static int smitest(void);
329 static cpumask_t smp_startup_mask = 1; /* which cpus have been started */
330 cpumask_t smp_active_mask = 1; /* which cpus are ready for IPIs etc? */
331 SYSCTL_INT(_machdep, OID_AUTO, smp_active, CTLFLAG_RD, &smp_active_mask, 0, "");
332 static u_int bootMP_size;
335 * Calculate usable address in base memory for AP trampoline code.
338 mp_bootaddress(u_int basemem)
340 POSTCODE(MP_BOOTADDRESS_POST);
342 base_memory = basemem;
344 bootMP_size = mptramp_end - mptramp_start;
345 boot_address = trunc_page(basemem * 1024); /* round down to 4k boundary */
346 if (((basemem * 1024) - boot_address) < bootMP_size)
347 boot_address -= PAGE_SIZE; /* not enough, lower by 4k */
348 /* 3 levels of page table pages */
349 mptramp_pagetables = boot_address - (PAGE_SIZE * 3);
351 return mptramp_pagetables;
356 * Look for an Intel MP spec table (ie, SMP capable hardware).
365 * Make sure our SMPpt[] page table is big enough to hold all the
368 KKASSERT(IO_MAPPING_START_INDEX < NPTEPG - 2);
370 POSTCODE(MP_PROBE_POST);
372 /* see if EBDA exists */
373 if (ebda_addr != 0) {
374 /* search first 1K of EBDA */
375 target = (u_int32_t)ebda_addr;
376 if ((x = mptable_search_sig(target, 1024 / 4)) > 0)
379 /* last 1K of base memory, effective 'top of base' passed in */
380 target = (u_int32_t)(base_memory - 0x400);
381 if ((x = mptable_search_sig(target, 1024 / 4)) > 0)
385 /* search the BIOS */
386 target = (u_int32_t)BIOS_BASE;
387 if ((x = mptable_search_sig(target, BIOS_COUNT)) > 0)
394 struct mptable_check_cbarg {
400 mptable_check_callback(void *xarg, const void *pos, int type)
402 const struct PROCENTRY *ent;
403 struct mptable_check_cbarg *arg = xarg;
409 if ((ent->cpu_flags & PROCENTRY_FLAG_EN) == 0)
413 if (ent->cpu_flags & PROCENTRY_FLAG_BP) {
414 if (arg->found_bsp) {
415 kprintf("more than one BSP in base MP table\n");
424 mptable_check(vm_paddr_t mpfps_paddr)
426 struct mptable_pos mpt;
427 struct mptable_check_cbarg arg;
431 if (mpfps_paddr == 0)
434 error = mptable_map(&mpt, mpfps_paddr);
438 if (mpt.mp_fps->mpfb1 != 0)
446 if (cth->apic_address == 0)
449 bzero(&arg, sizeof(arg));
450 error = mptable_iterate_entries(cth, mptable_check_callback, &arg);
452 if (arg.cpu_count == 0) {
453 kprintf("MP table contains no processor entries\n");
455 } else if (!arg.found_bsp) {
456 kprintf("MP table does not contains BSP entry\n");
466 mptable_iterate_entries(const mpcth_t cth, mptable_iter_func func, void *arg)
468 int count, total_size;
469 const void *position;
471 KKASSERT(cth->base_table_length >= sizeof(struct MPCTH));
472 total_size = cth->base_table_length - sizeof(struct MPCTH);
473 position = (const uint8_t *)cth + sizeof(struct MPCTH);
474 count = cth->entry_count;
479 KKASSERT(total_size >= 0);
480 if (total_size == 0) {
481 kprintf("invalid base MP table, "
482 "entry count and length mismatch\n");
486 type = *(const uint8_t *)position;
488 case 0: /* processor_entry */
489 case 1: /* bus_entry */
490 case 2: /* io_apic_entry */
491 case 3: /* int_entry */
492 case 4: /* int_entry */
495 kprintf("unknown base MP table entry type %d\n", type);
499 if (total_size < basetable_entry_types[type].length) {
500 kprintf("invalid base MP table length, "
501 "does not contain all entries\n");
504 total_size -= basetable_entry_types[type].length;
506 error = func(arg, position, type);
510 position = (const uint8_t *)position +
511 basetable_entry_types[type].length;
518 * Startup the SMP processors.
523 POSTCODE(MP_START_POST);
524 mp_enable(boot_address);
529 * Print various information about the SMP system hardware and setup.
536 POSTCODE(MP_ANNOUNCE_POST);
538 kprintf("DragonFly/MP: Multiprocessor motherboard\n");
539 kprintf(" cpu0 (BSP): apic id: %2d", CPU_TO_ID(0));
540 kprintf(", version: 0x%08x", cpu_apic_versions[0]);
541 kprintf(", at 0x%08jx\n", (intmax_t)cpu_apic_address);
542 for (x = 1; x <= mp_naps; ++x) {
543 kprintf(" cpu%d (AP): apic id: %2d", x, CPU_TO_ID(x));
544 kprintf(", version: 0x%08x", cpu_apic_versions[x]);
545 kprintf(", at 0x%08jx\n", (intmax_t)cpu_apic_address);
549 for (x = 0; x < mp_napics; ++x) {
550 kprintf(" io%d (APIC): apic id: %2d", x, IO_TO_ID(x));
551 kprintf(", version: 0x%08x", io_apic_versions[x]);
552 kprintf(", at 0x%08lx\n", io_apic_address[x]);
555 kprintf(" Warning: APIC I/O disabled\n");
560 * AP cpu's call this to sync up protected mode.
562 * WARNING! %gs is not set up on entry. This routine sets up %gs.
568 int x, myid = bootAP;
570 struct mdglobaldata *md;
571 struct privatespace *ps;
573 ps = &CPU_prvspace[myid];
575 gdt_segs[GPROC0_SEL].ssd_base =
576 (long) &ps->mdglobaldata.gd_common_tss;
577 ps->mdglobaldata.mi.gd_prvspace = ps;
579 /* We fill the 32-bit segment descriptors */
580 for (x = 0; x < NGDT; x++) {
581 if (x != GPROC0_SEL && x != (GPROC0_SEL + 1))
582 ssdtosd(&gdt_segs[x], &gdt[myid * NGDT + x]);
584 /* And now a 64-bit one */
585 ssdtosyssd(&gdt_segs[GPROC0_SEL],
586 (struct system_segment_descriptor *)&gdt[myid * NGDT + GPROC0_SEL]);
588 r_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
589 r_gdt.rd_base = (long) &gdt[myid * NGDT];
590 lgdt(&r_gdt); /* does magic intra-segment return */
592 /* lgdt() destroys the GSBASE value, so we load GSBASE after lgdt() */
593 wrmsr(MSR_FSBASE, 0); /* User value */
594 wrmsr(MSR_GSBASE, (u_int64_t)ps);
595 wrmsr(MSR_KGSBASE, 0); /* XXX User value while we're in the kernel */
601 mdcpu->gd_currentldt = _default_ldt;
604 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
605 gdt[myid * NGDT + GPROC0_SEL].sd_type = SDT_SYSTSS;
607 md = mdcpu; /* loaded through %gs:0 (mdglobaldata.mi.gd_prvspace)*/
609 md->gd_common_tss.tss_rsp0 = 0; /* not used until after switch */
611 md->gd_common_tss.tss_ioopt = (sizeof md->gd_common_tss) << 16;
613 md->gd_tss_gdt = &gdt[myid * NGDT + GPROC0_SEL];
614 md->gd_common_tssd = *md->gd_tss_gdt;
616 md->gd_common_tss.tss_ist1 = (long)&doublefault_stack[PAGE_SIZE];
621 * Set to a known state:
622 * Set by mpboot.s: CR0_PG, CR0_PE
623 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
626 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
629 /* Set up the fast syscall stuff */
630 msr = rdmsr(MSR_EFER) | EFER_SCE;
631 wrmsr(MSR_EFER, msr);
632 wrmsr(MSR_LSTAR, (u_int64_t)IDTVEC(fast_syscall));
633 wrmsr(MSR_CSTAR, (u_int64_t)IDTVEC(fast_syscall32));
634 msr = ((u_int64_t)GSEL(GCODE_SEL, SEL_KPL) << 32) |
635 ((u_int64_t)GSEL(GUCODE32_SEL, SEL_UPL) << 48);
636 wrmsr(MSR_STAR, msr);
637 wrmsr(MSR_SF_MASK, PSL_NT|PSL_T|PSL_I|PSL_C|PSL_D);
639 pmap_set_opt(); /* PSE/4MB pages, etc */
641 /* Initialize the PAT MSR. */
645 /* set up CPU registers and state */
648 /* set up SSE/NX registers */
651 /* set up FPU state on the AP */
652 npxinit(__INITIAL_NPXCW__);
654 /* disable the APIC, just to be SURE */
655 lapic->svr &= ~APIC_SVR_ENABLE;
657 /* data returned to BSP */
658 cpu_apic_versions[0] = lapic->version;
661 /*******************************************************************
662 * local functions and data
666 * start the SMP system
669 mp_enable(u_int boot_addr)
675 vm_paddr_t mpfps_paddr;
677 POSTCODE(MP_ENABLE_POST);
679 if (madt_probe_test) {
682 mpfps_paddr = mptable_probe();
683 if (mptable_check(mpfps_paddr))
688 struct mptable_pos mpt;
690 mptable_map(&mpt, mpfps_paddr);
692 mptable_lapic_enumerate(&mpt);
696 * We can safely map physical memory into SMPpt after
697 * mptable_pass1() completes.
702 * Examine the MP table for needed info
708 /* post scan cleanup */
712 vm_paddr_t madt_paddr;
713 vm_offset_t lapic_addr;
716 madt_paddr = madt_probe();
718 panic("mp_enable: madt_probe failed\n");
720 lapic_addr = madt_pass1(madt_paddr);
722 panic("mp_enable: no local apic (madt)!\n");
724 lapic_init(lapic_addr);
726 bsp_apic_id = APIC_ID(lapic->id);
727 if (madt_pass2(madt_paddr, bsp_apic_id))
728 panic("mp_enable: madt_pass2 failed\n");
733 setup_apic_irq_mapping();
735 /* fill the LOGICAL io_apic_versions table */
736 for (apic = 0; apic < mp_napics; ++apic) {
737 ux = io_apic_read(apic, IOAPIC_VER);
738 io_apic_versions[apic] = ux;
739 io_apic_set_id(apic, IO_TO_ID(apic));
742 /* program each IO APIC in the system */
743 for (apic = 0; apic < mp_napics; ++apic)
744 if (io_apic_setup(apic) < 0)
745 panic("IO APIC setup failure");
750 * These are required for SMP operation
753 /* install a 'Spurious INTerrupt' vector */
754 setidt(XSPURIOUSINT_OFFSET, Xspuriousint,
755 SDT_SYSIGT, SEL_KPL, 0);
757 /* install an inter-CPU IPI for TLB invalidation */
758 setidt(XINVLTLB_OFFSET, Xinvltlb,
759 SDT_SYSIGT, SEL_KPL, 0);
761 /* install an inter-CPU IPI for IPIQ messaging */
762 setidt(XIPIQ_OFFSET, Xipiq,
763 SDT_SYSIGT, SEL_KPL, 0);
765 /* install a timer vector */
766 setidt(XTIMER_OFFSET, Xtimer,
767 SDT_SYSIGT, SEL_KPL, 0);
769 /* install an inter-CPU IPI for CPU stop/restart */
770 setidt(XCPUSTOP_OFFSET, Xcpustop,
771 SDT_SYSIGT, SEL_KPL, 0);
773 /* start each Application Processor */
774 start_all_aps(boot_addr);
779 * look for the MP spec signature
782 /* string defined by the Intel MP Spec as identifying the MP table */
783 #define MP_SIG 0x5f504d5f /* _MP_ */
784 #define NEXT(X) ((X) += 4)
786 mptable_search_sig(u_int32_t target, int count)
792 KKASSERT(target != 0);
794 map_size = count * sizeof(u_int32_t);
795 addr = pmap_mapdev((vm_paddr_t)target, map_size);
798 for (x = 0; x < count; NEXT(x)) {
799 if (addr[x] == MP_SIG) {
800 /* make array index a byte index */
801 ret = target + (x * sizeof(u_int32_t));
806 pmap_unmapdev((vm_offset_t)addr, map_size);
811 typedef struct BUSDATA {
813 enum busTypes bus_type;
816 typedef struct INTDATA {
826 typedef struct BUSTYPENAME {
833 static bus_type_name bus_type_table[] =
839 {UNKNOWN_BUSTYPE, "---"},
842 {UNKNOWN_BUSTYPE, "---"},
843 {UNKNOWN_BUSTYPE, "---"},
844 {UNKNOWN_BUSTYPE, "---"},
845 {UNKNOWN_BUSTYPE, "---"},
846 {UNKNOWN_BUSTYPE, "---"},
848 {UNKNOWN_BUSTYPE, "---"},
849 {UNKNOWN_BUSTYPE, "---"},
850 {UNKNOWN_BUSTYPE, "---"},
851 {UNKNOWN_BUSTYPE, "---"},
853 {UNKNOWN_BUSTYPE, "---"}
856 /* from MP spec v1.4, table 5-1 */
857 static int default_data[7][5] =
859 /* nbus, id0, type0, id1, type1 */
860 {1, 0, ISA, 255, 255},
861 {1, 0, EISA, 255, 255},
862 {1, 0, EISA, 255, 255},
863 {1, 0, MCA, 255, 255},
865 {2, 0, EISA, 1, PCI},
870 static bus_datum *bus_data;
872 /* the IO INT data, one entry per possible APIC INTerrupt */
873 static io_int *io_apic_ints;
878 static int processor_entry (const struct PROCENTRY *entry, int cpu);
880 static int bus_entry (bus_entry_ptr entry, int bus);
881 static int io_apic_entry (io_apic_entry_ptr entry, int apic);
882 static int int_entry (int_entry_ptr entry, int intr);
883 static int lookup_bus_type (char *name);
889 mptable_ioapic_pass1_callback(void *xarg, const void *pos, int type)
891 const struct IOAPICENTRY *ioapic_ent;
894 case 1: /* bus_entry */
898 case 2: /* io_apic_entry */
900 if (ioapic_ent->apic_flags & IOAPICENTRY_FLAG_EN) {
901 io_apic_address[mp_napics++] =
902 (vm_offset_t)ioapic_ent->apic_address;
906 case 3: /* int_entry */
916 * 1st pass on motherboard's Intel MP specification table.
925 mptable_pass1(struct mptable_pos *mpt)
931 POSTCODE(MPTABLE_PASS1_POST);
934 KKASSERT(fps != NULL);
936 /* clear various tables */
937 for (x = 0; x < NAPICID; ++x)
938 io_apic_address[x] = ~0; /* IO APIC address table */
944 /* check for use of 'default' configuration */
945 if (fps->mpfb1 != 0) {
946 io_apic_address[0] = DEFAULT_IO_APIC_BASE;
947 mp_nbusses = default_data[fps->mpfb1 - 1][0];
953 error = mptable_iterate_entries(mpt->mp_cth,
954 mptable_ioapic_pass1_callback, NULL);
956 panic("mptable_iterate_entries(ioapic_pass1) failed\n");
963 * 2nd pass on motherboard's Intel MP specification table.
966 * ID_TO_IO(N), phy APIC ID to log CPU/IO table
967 * IO_TO_ID(N), logical IO to APIC ID table
972 mptable_pass2(struct mptable_pos *mpt)
986 POSTCODE(MPTABLE_PASS2_POST);
989 KKASSERT(fps != NULL);
992 MALLOC(io_apic_versions, u_int32_t *, sizeof(u_int32_t) * mp_napics,
994 MALLOC(ioapic, volatile ioapic_t **, sizeof(ioapic_t *) * mp_napics,
995 M_DEVBUF, M_WAITOK | M_ZERO);
996 MALLOC(io_apic_ints, io_int *, sizeof(io_int) * (nintrs + FIXUP_EXTRA_APIC_INTS),
998 MALLOC(bus_data, bus_datum *, sizeof(bus_datum) * mp_nbusses,
1003 for (i = 0; i < mp_napics; i++) {
1004 ioapic[i] = permanent_io_mapping(io_apic_address[i]);
1008 /* clear various tables */
1009 for (x = 0; x < NAPICID; ++x) {
1011 ID_TO_IO(x) = -1; /* phy APIC ID to log CPU/IO table */
1012 IO_TO_ID(x) = -1; /* logical IO to APIC ID table */
1017 /* clear bus data table */
1018 for (x = 0; x < mp_nbusses; ++x)
1019 bus_data[x].bus_id = 0xff;
1021 /* clear IO APIC INT table */
1022 for (x = 0; x < (nintrs + 1); ++x) {
1023 io_apic_ints[x].int_type = 0xff;
1024 io_apic_ints[x].int_vector = 0xff;
1028 /* record whether PIC or virtual-wire mode */
1029 machintr_setvar_simple(MACHINTR_VAR_IMCR_PRESENT, fps->mpfb2 & 0x80);
1031 /* check for use of 'default' configuration */
1032 if (fps->mpfb1 != 0) {
1033 mptable_default(fps->mpfb1);
1038 KKASSERT(cth != NULL);
1040 /* walk the table, recording info of interest */
1041 totalSize = cth->base_table_length - sizeof(struct MPCTH);
1042 position = (u_char *) cth + sizeof(struct MPCTH);
1043 count = cth->entry_count;
1044 apic = bus = intr = 0;
1047 switch (type = *(u_char *) position) {
1052 if (bus_entry(position, bus))
1058 if (io_apic_entry(position, apic))
1064 if (int_entry(position, intr))
1069 /* int_entry(position); */
1072 panic("mpfps Base Table HOSED!");
1076 totalSize -= basetable_entry_types[type].length;
1077 position = (uint8_t *)position + basetable_entry_types[type].length;
1083 * Check if we should perform a hyperthreading "fix-up" to
1084 * enumerate any logical CPU's that aren't already listed
1087 * XXX: We assume that all of the physical CPUs in the
1088 * system have the same number of logical CPUs.
1090 * XXX: We assume that APIC ID's are allocated such that
1091 * the APIC ID's for a physical processor are aligned
1092 * with the number of logical CPU's in the processor.
1095 mptable_hyperthread_fixup(u_int id_mask, int cpu_count)
1097 int i, id, lcpus_max, logical_cpus;
1099 if ((cpu_feature & CPUID_HTT) == 0)
1102 lcpus_max = (cpu_procinfo & CPUID_HTT_CORES) >> 16;
1106 if (strcmp(cpu_vendor, "GenuineIntel") == 0) {
1108 * INSTRUCTION SET REFERENCE, A-M (#253666)
1109 * Page 3-181, Table 3-20
1110 * "The nearest power-of-2 integer that is not smaller
1111 * than EBX[23:16] is the number of unique initial APIC
1112 * IDs reserved for addressing different logical
1113 * processors in a physical package."
1115 for (i = 0; ; ++i) {
1116 if ((1 << i) >= lcpus_max) {
1123 KKASSERT(cpu_count != 0);
1124 if (cpu_count == lcpus_max) {
1125 /* We have nothing to fix */
1127 } else if (cpu_count == 1) {
1128 /* XXX this may be incorrect */
1129 logical_cpus = lcpus_max;
1131 int cur, prev, dist;
1134 * Calculate the distances between two nearest
1135 * APIC IDs. If all such distances are same,
1136 * then it is the number of missing cpus that
1137 * we are going to fill later.
1139 dist = cur = prev = -1;
1140 for (id = 0; id < MAXCPU; ++id) {
1141 if ((id_mask & 1 << id) == 0)
1146 int new_dist = cur - prev;
1152 * Make sure that all distances
1153 * between two nearest APIC IDs
1156 if (dist != new_dist)
1164 /* Must be power of 2 */
1165 if (dist & (dist - 1))
1168 /* Can't exceed CPU package capacity */
1169 if (dist > lcpus_max)
1170 logical_cpus = lcpus_max;
1172 logical_cpus = dist;
1176 * For each APIC ID of a CPU that is set in the mask,
1177 * scan the other candidate APIC ID's for this
1178 * physical processor. If any of those ID's are
1179 * already in the table, then kill the fixup.
1181 for (id = 0; id < MAXCPU; id++) {
1182 if ((id_mask & 1 << id) == 0)
1184 /* First, make sure we are on a logical_cpus boundary. */
1185 if (id % logical_cpus != 0)
1187 for (i = id + 1; i < id + logical_cpus; i++)
1188 if ((id_mask & 1 << i) != 0)
1191 return logical_cpus;
1195 mptable_map(struct mptable_pos *mpt, vm_paddr_t mpfps_paddr)
1199 vm_size_t cth_mapsz = 0;
1201 bzero(mpt, sizeof(*mpt));
1203 fps = pmap_mapdev(mpfps_paddr, sizeof(*fps));
1204 if (fps->pap != 0) {
1206 * Map configuration table header to get
1207 * the base table size
1209 cth = pmap_mapdev(fps->pap, sizeof(*cth));
1210 cth_mapsz = cth->base_table_length;
1211 pmap_unmapdev((vm_offset_t)cth, sizeof(*cth));
1213 if (cth_mapsz < sizeof(*cth)) {
1214 kprintf("invalid base MP table length %d\n",
1216 pmap_unmapdev((vm_offset_t)fps, sizeof(*fps));
1221 * Map the base table
1223 cth = pmap_mapdev(fps->pap, cth_mapsz);
1228 mpt->mp_cth_mapsz = cth_mapsz;
1234 mptable_unmap(struct mptable_pos *mpt)
1236 if (mpt->mp_cth != NULL) {
1237 pmap_unmapdev((vm_offset_t)mpt->mp_cth, mpt->mp_cth_mapsz);
1239 mpt->mp_cth_mapsz = 0;
1241 if (mpt->mp_fps != NULL) {
1242 pmap_unmapdev((vm_offset_t)mpt->mp_fps, sizeof(*mpt->mp_fps));
1250 assign_apic_irq(int apic, int intpin, int irq)
1254 if (int_to_apicintpin[irq].ioapic != -1)
1255 panic("assign_apic_irq: inconsistent table");
1257 int_to_apicintpin[irq].ioapic = apic;
1258 int_to_apicintpin[irq].int_pin = intpin;
1259 int_to_apicintpin[irq].apic_address = ioapic[apic];
1260 int_to_apicintpin[irq].redirindex = IOAPIC_REDTBL + 2 * intpin;
1262 for (x = 0; x < nintrs; x++) {
1263 if ((io_apic_ints[x].int_type == 0 ||
1264 io_apic_ints[x].int_type == 3) &&
1265 io_apic_ints[x].int_vector == 0xff &&
1266 io_apic_ints[x].dst_apic_id == IO_TO_ID(apic) &&
1267 io_apic_ints[x].dst_apic_int == intpin)
1268 io_apic_ints[x].int_vector = irq;
1273 revoke_apic_irq(int irq)
1279 if (int_to_apicintpin[irq].ioapic == -1)
1280 panic("revoke_apic_irq: inconsistent table");
1282 oldapic = int_to_apicintpin[irq].ioapic;
1283 oldintpin = int_to_apicintpin[irq].int_pin;
1285 int_to_apicintpin[irq].ioapic = -1;
1286 int_to_apicintpin[irq].int_pin = 0;
1287 int_to_apicintpin[irq].apic_address = NULL;
1288 int_to_apicintpin[irq].redirindex = 0;
1290 for (x = 0; x < nintrs; x++) {
1291 if ((io_apic_ints[x].int_type == 0 ||
1292 io_apic_ints[x].int_type == 3) &&
1293 io_apic_ints[x].int_vector != 0xff &&
1294 io_apic_ints[x].dst_apic_id == IO_TO_ID(oldapic) &&
1295 io_apic_ints[x].dst_apic_int == oldintpin)
1296 io_apic_ints[x].int_vector = 0xff;
1304 allocate_apic_irq(int intr)
1310 if (io_apic_ints[intr].int_vector != 0xff)
1311 return; /* Interrupt handler already assigned */
1313 if (io_apic_ints[intr].int_type != 0 &&
1314 (io_apic_ints[intr].int_type != 3 ||
1315 (io_apic_ints[intr].dst_apic_id == IO_TO_ID(0) &&
1316 io_apic_ints[intr].dst_apic_int == 0)))
1317 return; /* Not INT or ExtInt on != (0, 0) */
1320 while (irq < APIC_INTMAPSIZE &&
1321 int_to_apicintpin[irq].ioapic != -1)
1324 if (irq >= APIC_INTMAPSIZE)
1325 return; /* No free interrupt handlers */
1327 apic = ID_TO_IO(io_apic_ints[intr].dst_apic_id);
1328 intpin = io_apic_ints[intr].dst_apic_int;
1330 assign_apic_irq(apic, intpin, irq);
1335 swap_apic_id(int apic, int oldid, int newid)
1342 return; /* Nothing to do */
1344 kprintf("Changing APIC ID for IO APIC #%d from %d to %d in MP table\n",
1345 apic, oldid, newid);
1347 /* Swap physical APIC IDs in interrupt entries */
1348 for (x = 0; x < nintrs; x++) {
1349 if (io_apic_ints[x].dst_apic_id == oldid)
1350 io_apic_ints[x].dst_apic_id = newid;
1351 else if (io_apic_ints[x].dst_apic_id == newid)
1352 io_apic_ints[x].dst_apic_id = oldid;
1355 /* Swap physical APIC IDs in IO_TO_ID mappings */
1356 for (oapic = 0; oapic < mp_napics; oapic++)
1357 if (IO_TO_ID(oapic) == newid)
1360 if (oapic < mp_napics) {
1361 kprintf("Changing APIC ID for IO APIC #%d from "
1362 "%d to %d in MP table\n",
1363 oapic, newid, oldid);
1364 IO_TO_ID(oapic) = oldid;
1366 IO_TO_ID(apic) = newid;
1371 fix_id_to_io_mapping(void)
1375 for (x = 0; x < NAPICID; x++)
1378 for (x = 0; x <= mp_naps; x++)
1379 if (CPU_TO_ID(x) < NAPICID)
1380 ID_TO_IO(CPU_TO_ID(x)) = x;
1382 for (x = 0; x < mp_napics; x++)
1383 if (IO_TO_ID(x) < NAPICID)
1384 ID_TO_IO(IO_TO_ID(x)) = x;
1389 first_free_apic_id(void)
1393 for (freeid = 0; freeid < NAPICID; freeid++) {
1394 for (x = 0; x <= mp_naps; x++)
1395 if (CPU_TO_ID(x) == freeid)
1399 for (x = 0; x < mp_napics; x++)
1400 if (IO_TO_ID(x) == freeid)
1411 io_apic_id_acceptable(int apic, int id)
1413 int cpu; /* Logical CPU number */
1414 int oapic; /* Logical IO APIC number for other IO APIC */
1417 return 0; /* Out of range */
1419 for (cpu = 0; cpu <= mp_naps; cpu++)
1420 if (CPU_TO_ID(cpu) == id)
1421 return 0; /* Conflict with CPU */
1423 for (oapic = 0; oapic < mp_napics && oapic < apic; oapic++)
1424 if (IO_TO_ID(oapic) == id)
1425 return 0; /* Conflict with other APIC */
1427 return 1; /* ID is acceptable for IO APIC */
1432 io_apic_find_int_entry(int apic, int pin)
1436 /* search each of the possible INTerrupt sources */
1437 for (x = 0; x < nintrs; ++x) {
1438 if ((apic == ID_TO_IO(io_apic_ints[x].dst_apic_id)) &&
1439 (pin == io_apic_ints[x].dst_apic_int))
1440 return (&io_apic_ints[x]);
1448 * parse an Intel MP specification table
1456 int apic; /* IO APIC unit number */
1457 int freeid; /* Free physical APIC ID */
1458 int physid; /* Current physical IO APIC ID */
1460 int bus_0 = 0; /* Stop GCC warning */
1461 int bus_pci = 0; /* Stop GCC warning */
1465 * Fix mis-numbering of the PCI bus and its INT entries if the BIOS
1466 * did it wrong. The MP spec says that when more than 1 PCI bus
1467 * exists the BIOS must begin with bus entries for the PCI bus and use
1468 * actual PCI bus numbering. This implies that when only 1 PCI bus
1469 * exists the BIOS can choose to ignore this ordering, and indeed many
1470 * MP motherboards do ignore it. This causes a problem when the PCI
1471 * sub-system makes requests of the MP sub-system based on PCI bus
1472 * numbers. So here we look for the situation and renumber the
1473 * busses and associated INTs in an effort to "make it right".
1476 /* find bus 0, PCI bus, count the number of PCI busses */
1477 for (num_pci_bus = 0, x = 0; x < mp_nbusses; ++x) {
1478 if (bus_data[x].bus_id == 0) {
1481 if (bus_data[x].bus_type == PCI) {
1487 * bus_0 == slot of bus with ID of 0
1488 * bus_pci == slot of last PCI bus encountered
1491 /* check the 1 PCI bus case for sanity */
1492 /* if it is number 0 all is well */
1493 if (num_pci_bus == 1 &&
1494 bus_data[bus_pci].bus_id != 0) {
1496 /* mis-numbered, swap with whichever bus uses slot 0 */
1498 /* swap the bus entry types */
1499 bus_data[bus_pci].bus_type = bus_data[bus_0].bus_type;
1500 bus_data[bus_0].bus_type = PCI;
1502 /* swap each relavant INTerrupt entry */
1503 id = bus_data[bus_pci].bus_id;
1504 for (x = 0; x < nintrs; ++x) {
1505 if (io_apic_ints[x].src_bus_id == id) {
1506 io_apic_ints[x].src_bus_id = 0;
1508 else if (io_apic_ints[x].src_bus_id == 0) {
1509 io_apic_ints[x].src_bus_id = id;
1514 /* Assign IO APIC IDs.
1516 * First try the existing ID. If a conflict is detected, try
1517 * the ID in the MP table. If a conflict is still detected, find
1520 * We cannot use the ID_TO_IO table before all conflicts has been
1521 * resolved and the table has been corrected.
1523 for (apic = 0; apic < mp_napics; ++apic) { /* For all IO APICs */
1525 /* First try to use the value set by the BIOS */
1526 physid = io_apic_get_id(apic);
1527 if (io_apic_id_acceptable(apic, physid)) {
1528 if (IO_TO_ID(apic) != physid)
1529 swap_apic_id(apic, IO_TO_ID(apic), physid);
1533 /* Then check if the value in the MP table is acceptable */
1534 if (io_apic_id_acceptable(apic, IO_TO_ID(apic)))
1537 /* Last resort, find a free APIC ID and use it */
1538 freeid = first_free_apic_id();
1539 if (freeid >= NAPICID)
1540 panic("No free physical APIC IDs found");
1542 if (io_apic_id_acceptable(apic, freeid)) {
1543 swap_apic_id(apic, IO_TO_ID(apic), freeid);
1546 panic("Free physical APIC ID not usable");
1548 fix_id_to_io_mapping();
1550 /* detect and fix broken Compaq MP table */
1551 if (apic_int_type(0, 0) == -1) {
1552 kprintf("APIC_IO: MP table broken: 8259->APIC entry missing!\n");
1553 io_apic_ints[nintrs].int_type = 3; /* ExtInt */
1554 io_apic_ints[nintrs].int_vector = 0xff; /* Unassigned */
1555 /* XXX fixme, set src bus id etc, but it doesn't seem to hurt */
1556 io_apic_ints[nintrs].dst_apic_id = IO_TO_ID(0);
1557 io_apic_ints[nintrs].dst_apic_int = 0; /* Pin 0 */
1559 } else if (apic_int_type(0, 0) == 0) {
1560 kprintf("APIC_IO: MP table broken: ExtINT entry corrupt!\n");
1561 for (x = 0; x < nintrs; ++x)
1562 if ((0 == ID_TO_IO(io_apic_ints[x].dst_apic_id)) &&
1563 (0 == io_apic_ints[x].dst_apic_int)) {
1564 io_apic_ints[x].int_type = 3;
1565 io_apic_ints[x].int_vector = 0xff;
1571 * Fix missing IRQ 15 when IRQ 14 is an ISA interrupt. IDE
1572 * controllers universally come in pairs. If IRQ 14 is specified
1573 * as an ISA interrupt, then IRQ 15 had better be too.
1575 * [ Shuttle XPC / AMD Athlon X2 ]
1576 * The MPTable is missing an entry for IRQ 15. Note that the
1577 * ACPI table has an entry for both 14 and 15.
1579 if (apic_int_type(0, 14) == 0 && apic_int_type(0, 15) == -1) {
1580 kprintf("APIC_IO: MP table broken: IRQ 15 not ISA when IRQ 14 is!\n");
1581 io14 = io_apic_find_int_entry(0, 14);
1582 io_apic_ints[nintrs] = *io14;
1583 io_apic_ints[nintrs].src_bus_irq = 15;
1584 io_apic_ints[nintrs].dst_apic_int = 15;
1592 /* Assign low level interrupt handlers */
1594 setup_apic_irq_mapping(void)
1600 for (x = 0; x < APIC_INTMAPSIZE; x++) {
1601 int_to_apicintpin[x].ioapic = -1;
1602 int_to_apicintpin[x].int_pin = 0;
1603 int_to_apicintpin[x].apic_address = NULL;
1604 int_to_apicintpin[x].redirindex = 0;
1607 /* First assign ISA/EISA interrupts */
1608 for (x = 0; x < nintrs; x++) {
1609 int_vector = io_apic_ints[x].src_bus_irq;
1610 if (int_vector < APIC_INTMAPSIZE &&
1611 io_apic_ints[x].int_vector == 0xff &&
1612 int_to_apicintpin[int_vector].ioapic == -1 &&
1613 (apic_int_is_bus_type(x, ISA) ||
1614 apic_int_is_bus_type(x, EISA)) &&
1615 io_apic_ints[x].int_type == 0) {
1616 assign_apic_irq(ID_TO_IO(io_apic_ints[x].dst_apic_id),
1617 io_apic_ints[x].dst_apic_int,
1622 /* Assign ExtInt entry if no ISA/EISA interrupt 0 entry */
1623 for (x = 0; x < nintrs; x++) {
1624 if (io_apic_ints[x].dst_apic_int == 0 &&
1625 io_apic_ints[x].dst_apic_id == IO_TO_ID(0) &&
1626 io_apic_ints[x].int_vector == 0xff &&
1627 int_to_apicintpin[0].ioapic == -1 &&
1628 io_apic_ints[x].int_type == 3) {
1629 assign_apic_irq(0, 0, 0);
1634 /* Assign PCI interrupts */
1635 for (x = 0; x < nintrs; ++x) {
1636 if (io_apic_ints[x].int_type == 0 &&
1637 io_apic_ints[x].int_vector == 0xff &&
1638 apic_int_is_bus_type(x, PCI))
1639 allocate_apic_irq(x);
1646 mp_set_cpuids(int cpu_id, int apic_id)
1648 CPU_TO_ID(cpu_id) = apic_id;
1649 ID_TO_CPU(apic_id) = cpu_id;
1653 processor_entry(const struct PROCENTRY *entry, int cpu)
1657 /* check for usability */
1658 if (!(entry->cpu_flags & PROCENTRY_FLAG_EN))
1661 /* check for BSP flag */
1662 if (entry->cpu_flags & PROCENTRY_FLAG_BP) {
1663 mp_set_cpuids(0, entry->apic_id);
1664 return 0; /* its already been counted */
1667 /* add another AP to list, if less than max number of CPUs */
1668 else if (cpu < MAXCPU) {
1669 mp_set_cpuids(cpu, entry->apic_id);
1679 bus_entry(bus_entry_ptr entry, int bus)
1684 /* encode the name into an index */
1685 for (x = 0; x < 6; ++x) {
1686 if ((c = entry->bus_type[x]) == ' ')
1692 if ((x = lookup_bus_type(name)) == UNKNOWN_BUSTYPE)
1693 panic("unknown bus type: '%s'", name);
1695 bus_data[bus].bus_id = entry->bus_id;
1696 bus_data[bus].bus_type = x;
1702 io_apic_entry(io_apic_entry_ptr entry, int apic)
1704 if (!(entry->apic_flags & IOAPICENTRY_FLAG_EN))
1707 IO_TO_ID(apic) = entry->apic_id;
1708 ID_TO_IO(entry->apic_id) = apic;
1714 lookup_bus_type(char *name)
1718 for (x = 0; x < MAX_BUSTYPE; ++x)
1719 if (strcmp(bus_type_table[x].name, name) == 0)
1720 return bus_type_table[x].type;
1722 return UNKNOWN_BUSTYPE;
1726 int_entry(int_entry_ptr entry, int intr)
1730 io_apic_ints[intr].int_type = entry->int_type;
1731 io_apic_ints[intr].int_flags = entry->int_flags;
1732 io_apic_ints[intr].src_bus_id = entry->src_bus_id;
1733 io_apic_ints[intr].src_bus_irq = entry->src_bus_irq;
1734 if (entry->dst_apic_id == 255) {
1735 /* This signal goes to all IO APICS. Select an IO APIC
1736 with sufficient number of interrupt pins */
1737 for (apic = 0; apic < mp_napics; apic++)
1738 if (((io_apic_read(apic, IOAPIC_VER) &
1739 IOART_VER_MAXREDIR) >> MAXREDIRSHIFT) >=
1740 entry->dst_apic_int)
1742 if (apic < mp_napics)
1743 io_apic_ints[intr].dst_apic_id = IO_TO_ID(apic);
1745 io_apic_ints[intr].dst_apic_id = entry->dst_apic_id;
1747 io_apic_ints[intr].dst_apic_id = entry->dst_apic_id;
1748 io_apic_ints[intr].dst_apic_int = entry->dst_apic_int;
1754 apic_int_is_bus_type(int intr, int bus_type)
1758 for (bus = 0; bus < mp_nbusses; ++bus)
1759 if ((bus_data[bus].bus_id == io_apic_ints[intr].src_bus_id)
1760 && ((int) bus_data[bus].bus_type == bus_type))
1767 * Given a traditional ISA INT mask, return an APIC mask.
1770 isa_apic_mask(u_int isa_mask)
1775 #if defined(SKIP_IRQ15_REDIRECT)
1776 if (isa_mask == (1 << 15)) {
1777 kprintf("skipping ISA IRQ15 redirect\n");
1780 #endif /* SKIP_IRQ15_REDIRECT */
1782 isa_irq = ffs(isa_mask); /* find its bit position */
1783 if (isa_irq == 0) /* doesn't exist */
1785 --isa_irq; /* make it zero based */
1787 apic_pin = isa_apic_irq(isa_irq); /* look for APIC connection */
1791 return (1 << apic_pin); /* convert pin# to a mask */
1795 * Determine which APIC pin an ISA/EISA INT is attached to.
1797 #define INTTYPE(I) (io_apic_ints[(I)].int_type)
1798 #define INTPIN(I) (io_apic_ints[(I)].dst_apic_int)
1799 #define INTIRQ(I) (io_apic_ints[(I)].int_vector)
1800 #define INTAPIC(I) (ID_TO_IO(io_apic_ints[(I)].dst_apic_id))
1802 #define SRCBUSIRQ(I) (io_apic_ints[(I)].src_bus_irq)
1804 isa_apic_irq(int isa_irq)
1808 for (intr = 0; intr < nintrs; ++intr) { /* check each record */
1809 if (INTTYPE(intr) == 0) { /* standard INT */
1810 if (SRCBUSIRQ(intr) == isa_irq) {
1811 if (apic_int_is_bus_type(intr, ISA) ||
1812 apic_int_is_bus_type(intr, EISA)) {
1813 if (INTIRQ(intr) == 0xff)
1814 return -1; /* unassigned */
1815 return INTIRQ(intr); /* found */
1820 return -1; /* NOT found */
1825 * Determine which APIC pin a PCI INT is attached to.
1827 #define SRCBUSID(I) (io_apic_ints[(I)].src_bus_id)
1828 #define SRCBUSDEVICE(I) ((io_apic_ints[(I)].src_bus_irq >> 2) & 0x1f)
1829 #define SRCBUSLINE(I) (io_apic_ints[(I)].src_bus_irq & 0x03)
1831 pci_apic_irq(int pciBus, int pciDevice, int pciInt)
1835 --pciInt; /* zero based */
1837 for (intr = 0; intr < nintrs; ++intr) { /* check each record */
1838 if ((INTTYPE(intr) == 0) /* standard INT */
1839 && (SRCBUSID(intr) == pciBus)
1840 && (SRCBUSDEVICE(intr) == pciDevice)
1841 && (SRCBUSLINE(intr) == pciInt)) { /* a candidate IRQ */
1842 if (apic_int_is_bus_type(intr, PCI)) {
1843 if (INTIRQ(intr) == 0xff) {
1844 kprintf("IOAPIC: pci_apic_irq() "
1846 return -1; /* unassigned */
1848 return INTIRQ(intr); /* exact match */
1853 return -1; /* NOT found */
1857 next_apic_irq(int irq)
1864 for (intr = 0; intr < nintrs; intr++) {
1865 if (INTIRQ(intr) != irq || INTTYPE(intr) != 0)
1867 bus = SRCBUSID(intr);
1868 bustype = apic_bus_type(bus);
1869 if (bustype != ISA &&
1875 if (intr >= nintrs) {
1878 for (ointr = intr + 1; ointr < nintrs; ointr++) {
1879 if (INTTYPE(ointr) != 0)
1881 if (bus != SRCBUSID(ointr))
1883 if (bustype == PCI) {
1884 if (SRCBUSDEVICE(intr) != SRCBUSDEVICE(ointr))
1886 if (SRCBUSLINE(intr) != SRCBUSLINE(ointr))
1889 if (bustype == ISA || bustype == EISA) {
1890 if (SRCBUSIRQ(intr) != SRCBUSIRQ(ointr))
1893 if (INTPIN(intr) == INTPIN(ointr))
1897 if (ointr >= nintrs) {
1900 return INTIRQ(ointr);
1915 * Reprogram the MB chipset to NOT redirect an ISA INTerrupt.
1918 * Exactly what this means is unclear at this point. It is a solution
1919 * for motherboards that redirect the MBIRQ0 pin. Generically a motherboard
1920 * could route any of the ISA INTs to upper (>15) IRQ values. But most would
1921 * NOT be redirected via MBIRQ0, thus "undirect()ing" them would NOT be an
1925 undirect_isa_irq(int rirq)
1929 kprintf("Freeing redirected ISA irq %d.\n", rirq);
1930 /** FIXME: tickle the MB redirector chip */
1934 kprintf("Freeing (NOT implemented) redirected ISA irq %d.\n", rirq);
1941 * Reprogram the MB chipset to NOT redirect a PCI INTerrupt
1944 undirect_pci_irq(int rirq)
1948 kprintf("Freeing redirected PCI irq %d.\n", rirq);
1950 /** FIXME: tickle the MB redirector chip */
1954 kprintf("Freeing (NOT implemented) redirected PCI irq %d.\n",
1964 * given a bus ID, return:
1965 * the bus type if found
1969 apic_bus_type(int id)
1973 for (x = 0; x < mp_nbusses; ++x)
1974 if (bus_data[x].bus_id == id)
1975 return bus_data[x].bus_type;
1981 * given a LOGICAL APIC# and pin#, return:
1982 * the associated src bus ID if found
1986 apic_src_bus_id(int apic, int pin)
1990 /* search each of the possible INTerrupt sources */
1991 for (x = 0; x < nintrs; ++x)
1992 if ((apic == ID_TO_IO(io_apic_ints[x].dst_apic_id)) &&
1993 (pin == io_apic_ints[x].dst_apic_int))
1994 return (io_apic_ints[x].src_bus_id);
1996 return -1; /* NOT found */
2000 * given a LOGICAL APIC# and pin#, return:
2001 * the associated src bus IRQ if found
2005 apic_src_bus_irq(int apic, int pin)
2009 for (x = 0; x < nintrs; x++)
2010 if ((apic == ID_TO_IO(io_apic_ints[x].dst_apic_id)) &&
2011 (pin == io_apic_ints[x].dst_apic_int))
2012 return (io_apic_ints[x].src_bus_irq);
2014 return -1; /* NOT found */
2019 * given a LOGICAL APIC# and pin#, return:
2020 * the associated INTerrupt type if found
2024 apic_int_type(int apic, int pin)
2028 /* search each of the possible INTerrupt sources */
2029 for (x = 0; x < nintrs; ++x) {
2030 if ((apic == ID_TO_IO(io_apic_ints[x].dst_apic_id)) &&
2031 (pin == io_apic_ints[x].dst_apic_int))
2032 return (io_apic_ints[x].int_type);
2034 return -1; /* NOT found */
2038 * Return the IRQ associated with an APIC pin
2041 apic_irq(int apic, int pin)
2046 for (x = 0; x < nintrs; ++x) {
2047 if ((apic == ID_TO_IO(io_apic_ints[x].dst_apic_id)) &&
2048 (pin == io_apic_ints[x].dst_apic_int)) {
2049 res = io_apic_ints[x].int_vector;
2052 if (apic != int_to_apicintpin[res].ioapic)
2053 panic("apic_irq: inconsistent table %d/%d", apic, int_to_apicintpin[res].ioapic);
2054 if (pin != int_to_apicintpin[res].int_pin)
2055 panic("apic_irq inconsistent table (2)");
2064 * given a LOGICAL APIC# and pin#, return:
2065 * the associated trigger mode if found
2069 apic_trigger(int apic, int pin)
2073 /* search each of the possible INTerrupt sources */
2074 for (x = 0; x < nintrs; ++x)
2075 if ((apic == ID_TO_IO(io_apic_ints[x].dst_apic_id)) &&
2076 (pin == io_apic_ints[x].dst_apic_int))
2077 return ((io_apic_ints[x].int_flags >> 2) & 0x03);
2079 return -1; /* NOT found */
2084 * given a LOGICAL APIC# and pin#, return:
2085 * the associated 'active' level if found
2089 apic_polarity(int apic, int pin)
2093 /* search each of the possible INTerrupt sources */
2094 for (x = 0; x < nintrs; ++x)
2095 if ((apic == ID_TO_IO(io_apic_ints[x].dst_apic_id)) &&
2096 (pin == io_apic_ints[x].dst_apic_int))
2097 return (io_apic_ints[x].int_flags & 0x03);
2099 return -1; /* NOT found */
2105 * set data according to MP defaults
2106 * FIXME: probably not complete yet...
2109 mptable_default(int type)
2111 #if defined(APIC_IO)
2116 kprintf(" MP default config type: %d\n", type);
2119 kprintf(" bus: ISA, APIC: 82489DX\n");
2122 kprintf(" bus: EISA, APIC: 82489DX\n");
2125 kprintf(" bus: EISA, APIC: 82489DX\n");
2128 kprintf(" bus: MCA, APIC: 82489DX\n");
2131 kprintf(" bus: ISA+PCI, APIC: Integrated\n");
2134 kprintf(" bus: EISA+PCI, APIC: Integrated\n");
2137 kprintf(" bus: MCA+PCI, APIC: Integrated\n");
2140 kprintf(" future type\n");
2146 /* one and only IO APIC */
2147 io_apic_id = (io_apic_read(0, IOAPIC_ID) & APIC_ID_MASK) >> 24;
2150 * sanity check, refer to MP spec section 3.6.6, last paragraph
2151 * necessary as some hardware isn't properly setting up the IO APIC
2153 #if defined(REALLY_ANAL_IOAPICID_VALUE)
2154 if (io_apic_id != 2) {
2156 if ((io_apic_id == 0) || (io_apic_id == 1) || (io_apic_id == 15)) {
2157 #endif /* REALLY_ANAL_IOAPICID_VALUE */
2158 io_apic_set_id(0, 2);
2161 IO_TO_ID(0) = io_apic_id;
2162 ID_TO_IO(io_apic_id) = 0;
2164 /* fill out bus entries */
2173 bus_data[0].bus_id = default_data[type - 1][1];
2174 bus_data[0].bus_type = default_data[type - 1][2];
2175 bus_data[1].bus_id = default_data[type - 1][3];
2176 bus_data[1].bus_type = default_data[type - 1][4];
2179 /* case 4: case 7: MCA NOT supported */
2180 default: /* illegal/reserved */
2181 panic("BAD default MP config: %d", type);
2185 /* general cases from MP v1.4, table 5-2 */
2186 for (pin = 0; pin < 16; ++pin) {
2187 io_apic_ints[pin].int_type = 0;
2188 io_apic_ints[pin].int_flags = 0x05; /* edge/active-hi */
2189 io_apic_ints[pin].src_bus_id = 0;
2190 io_apic_ints[pin].src_bus_irq = pin; /* IRQ2 caught below */
2191 io_apic_ints[pin].dst_apic_id = io_apic_id;
2192 io_apic_ints[pin].dst_apic_int = pin; /* 1-to-1 */
2195 /* special cases from MP v1.4, table 5-2 */
2197 io_apic_ints[2].int_type = 0xff; /* N/C */
2198 io_apic_ints[13].int_type = 0xff; /* N/C */
2199 #if !defined(APIC_MIXED_MODE)
2201 panic("sorry, can't support type 2 default yet");
2202 #endif /* APIC_MIXED_MODE */
2205 io_apic_ints[2].src_bus_irq = 0; /* ISA IRQ0 is on APIC INT 2 */
2208 io_apic_ints[0].int_type = 0xff; /* N/C */
2210 io_apic_ints[0].int_type = 3; /* vectored 8259 */
2211 #endif /* APIC_IO */
2215 * Map a physical memory address representing I/O into KVA. The I/O
2216 * block is assumed not to cross a page boundary.
2219 permanent_io_mapping(vm_paddr_t pa)
2221 KKASSERT(pa < 0x100000000LL);
2223 return pmap_mapdev_uncacheable(pa, PAGE_SIZE);
2227 * start each AP in our list
2230 start_all_aps(u_int boot_addr)
2232 vm_offset_t va = boot_address + KERNBASE;
2233 u_int64_t *pt4, *pt3, *pt2;
2239 u_char mpbiosreason;
2240 u_long mpbioswarmvec;
2241 struct mdglobaldata *gd;
2242 struct privatespace *ps;
2244 POSTCODE(START_ALL_APS_POST);
2246 /* Initialize BSP's local APIC */
2247 apic_initialize(TRUE);
2249 /* install the AP 1st level boot code */
2250 pmap_kenter(va, boot_address);
2251 cpu_invlpg((void *)va); /* JG XXX */
2252 bcopy(mptramp_start, (void *)va, bootMP_size);
2254 /* Locate the page tables, they'll be below the trampoline */
2255 pt4 = (u_int64_t *)(uintptr_t)(mptramp_pagetables + KERNBASE);
2256 pt3 = pt4 + (PAGE_SIZE) / sizeof(u_int64_t);
2257 pt2 = pt3 + (PAGE_SIZE) / sizeof(u_int64_t);
2259 /* Create the initial 1GB replicated page tables */
2260 for (i = 0; i < 512; i++) {
2261 /* Each slot of the level 4 pages points to the same level 3 page */
2262 pt4[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + PAGE_SIZE);
2263 pt4[i] |= PG_V | PG_RW | PG_U;
2265 /* Each slot of the level 3 pages points to the same level 2 page */
2266 pt3[i] = (u_int64_t)(uintptr_t)(mptramp_pagetables + (2 * PAGE_SIZE));
2267 pt3[i] |= PG_V | PG_RW | PG_U;
2269 /* The level 2 page slots are mapped with 2MB pages for 1GB. */
2270 pt2[i] = i * (2 * 1024 * 1024);
2271 pt2[i] |= PG_V | PG_RW | PG_PS | PG_U;
2274 /* save the current value of the warm-start vector */
2275 mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF);
2276 outb(CMOS_REG, BIOS_RESET);
2277 mpbiosreason = inb(CMOS_DATA);
2279 /* setup a vector to our boot code */
2280 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
2281 *((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4);
2282 outb(CMOS_REG, BIOS_RESET);
2283 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
2286 * If we have a TSC we can figure out the SMI interrupt rate.
2287 * The SMI does not necessarily use a constant rate. Spend
2288 * up to 250ms trying to figure it out.
2291 if (cpu_feature & CPUID_TSC) {
2292 set_apic_timer(275000);
2293 smilast = read_apic_timer();
2294 for (x = 0; x < 20 && read_apic_timer(); ++x) {
2295 smicount = smitest();
2296 if (smibest == 0 || smilast - smicount < smibest)
2297 smibest = smilast - smicount;
2300 if (smibest > 250000)
2303 smibest = smibest * (int64_t)1000000 /
2304 get_apic_timer_frequency();
2308 kprintf("SMI Frequency (worst case): %d Hz (%d us)\n",
2309 1000000 / smibest, smibest);
2312 for (x = 1; x <= mp_naps; ++x) {
2314 /* This is a bit verbose, it will go away soon. */
2316 /* first page of AP's private space */
2317 pg = x * x86_64_btop(sizeof(struct privatespace));
2319 /* allocate new private data page(s) */
2320 gd = (struct mdglobaldata *)kmem_alloc(&kernel_map,
2321 MDGLOBALDATA_BASEALLOC_SIZE);
2323 gd = &CPU_prvspace[x].mdglobaldata; /* official location */
2324 bzero(gd, sizeof(*gd));
2325 gd->mi.gd_prvspace = ps = &CPU_prvspace[x];
2327 /* prime data page for it to use */
2328 mi_gdinit(&gd->mi, x);
2330 gd->gd_CMAP1 = &SMPpt[pg + 0];
2331 gd->gd_CMAP2 = &SMPpt[pg + 1];
2332 gd->gd_CMAP3 = &SMPpt[pg + 2];
2333 gd->gd_PMAP1 = &SMPpt[pg + 3];
2334 gd->gd_CADDR1 = ps->CPAGE1;
2335 gd->gd_CADDR2 = ps->CPAGE2;
2336 gd->gd_CADDR3 = ps->CPAGE3;
2337 gd->gd_PADDR1 = (pt_entry_t *)ps->PPAGE1;
2338 gd->mi.gd_ipiq = (void *)kmem_alloc(&kernel_map, sizeof(lwkt_ipiq) * (mp_naps + 1));
2339 bzero(gd->mi.gd_ipiq, sizeof(lwkt_ipiq) * (mp_naps + 1));
2341 /* setup a vector to our boot code */
2342 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
2343 *((volatile u_short *) WARMBOOT_SEG) = (boot_addr >> 4);
2344 outb(CMOS_REG, BIOS_RESET);
2345 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
2348 * Setup the AP boot stack
2350 bootSTK = &ps->idlestack[UPAGES*PAGE_SIZE/2];
2353 /* attempt to start the Application Processor */
2354 CHECK_INIT(99); /* setup checkpoints */
2355 if (!start_ap(gd, boot_addr, smibest)) {
2356 kprintf("AP #%d (PHY# %d) failed!\n", x, CPU_TO_ID(x));
2357 CHECK_PRINT("trace"); /* show checkpoints */
2358 /* better panic as the AP may be running loose */
2359 kprintf("panic y/n? [y] ");
2360 if (cngetc() != 'n')
2363 CHECK_PRINT("trace"); /* show checkpoints */
2365 /* record its version info */
2366 cpu_apic_versions[x] = cpu_apic_versions[0];
2369 /* set ncpus to 1 + highest logical cpu. Not all may have come up */
2372 /* ncpus2 -- ncpus rounded down to the nearest power of 2 */
2373 for (shift = 0; (1 << shift) <= ncpus; ++shift)
2376 ncpus2_shift = shift;
2377 ncpus2 = 1 << shift;
2378 ncpus2_mask = ncpus2 - 1;
2380 /* ncpus_fit -- ncpus rounded up to the nearest power of 2 */
2381 if ((1 << shift) < ncpus)
2383 ncpus_fit = 1 << shift;
2384 ncpus_fit_mask = ncpus_fit - 1;
2386 /* build our map of 'other' CPUs */
2387 mycpu->gd_other_cpus = smp_startup_mask & ~(1 << mycpu->gd_cpuid);
2388 mycpu->gd_ipiq = (void *)kmem_alloc(&kernel_map, sizeof(lwkt_ipiq) * ncpus);
2389 bzero(mycpu->gd_ipiq, sizeof(lwkt_ipiq) * ncpus);
2391 /* fill in our (BSP) APIC version */
2392 cpu_apic_versions[0] = lapic->version;
2394 /* restore the warmstart vector */
2395 *(u_long *) WARMBOOT_OFF = mpbioswarmvec;
2396 outb(CMOS_REG, BIOS_RESET);
2397 outb(CMOS_DATA, mpbiosreason);
2400 * NOTE! The idlestack for the BSP was setup by locore. Finish
2401 * up, clean out the P==V mapping we did earlier.
2405 /* number of APs actually started */
2411 * load the 1st level AP boot code into base memory.
2414 /* targets for relocation */
2415 extern void bigJump(void);
2416 extern void bootCodeSeg(void);
2417 extern void bootDataSeg(void);
2418 extern void MPentry(void);
2419 extern u_int MP_GDT;
2420 extern u_int mp_gdtbase;
2425 install_ap_tramp(u_int boot_addr)
2428 int size = *(int *) ((u_long) & bootMP_size);
2429 u_char *src = (u_char *) ((u_long) bootMP);
2430 u_char *dst = (u_char *) boot_addr + KERNBASE;
2431 u_int boot_base = (u_int) bootMP;
2436 POSTCODE(INSTALL_AP_TRAMP_POST);
2438 for (x = 0; x < size; ++x)
2442 * modify addresses in code we just moved to basemem. unfortunately we
2443 * need fairly detailed info about mpboot.s for this to work. changes
2444 * to mpboot.s might require changes here.
2447 /* boot code is located in KERNEL space */
2448 dst = (u_char *) boot_addr + KERNBASE;
2450 /* modify the lgdt arg */
2451 dst32 = (u_int32_t *) (dst + ((u_int) & mp_gdtbase - boot_base));
2452 *dst32 = boot_addr + ((u_int) & MP_GDT - boot_base);
2454 /* modify the ljmp target for MPentry() */
2455 dst32 = (u_int32_t *) (dst + ((u_int) bigJump - boot_base) + 1);
2456 *dst32 = ((u_int) MPentry - KERNBASE);
2458 /* modify the target for boot code segment */
2459 dst16 = (u_int16_t *) (dst + ((u_int) bootCodeSeg - boot_base));
2460 dst8 = (u_int8_t *) (dst16 + 1);
2461 *dst16 = (u_int) boot_addr & 0xffff;
2462 *dst8 = ((u_int) boot_addr >> 16) & 0xff;
2464 /* modify the target for boot data segment */
2465 dst16 = (u_int16_t *) (dst + ((u_int) bootDataSeg - boot_base));
2466 dst8 = (u_int8_t *) (dst16 + 1);
2467 *dst16 = (u_int) boot_addr & 0xffff;
2468 *dst8 = ((u_int) boot_addr >> 16) & 0xff;
2474 * This function starts the AP (application processor) identified
2475 * by the APIC ID 'physicalCpu'. It does quite a "song and dance"
2476 * to accomplish this. This is necessary because of the nuances
2477 * of the different hardware we might encounter. It ain't pretty,
2478 * but it seems to work.
2480 * NOTE: eventually an AP gets to ap_init(), which is called just
2481 * before the AP goes into the LWKT scheduler's idle loop.
2484 start_ap(struct mdglobaldata *gd, u_int boot_addr, int smibest)
2488 u_long icr_lo, icr_hi;
2490 POSTCODE(START_AP_POST);
2492 /* get the PHYSICAL APIC ID# */
2493 physical_cpu = CPU_TO_ID(gd->mi.gd_cpuid);
2495 /* calculate the vector */
2496 vector = (boot_addr >> 12) & 0xff;
2498 /* We don't want anything interfering */
2501 /* Make sure the target cpu sees everything */
2505 * Try to detect when a SMI has occurred, wait up to 200ms.
2507 * If a SMI occurs during an AP reset but before we issue
2508 * the STARTUP command, the AP may brick. To work around
2509 * this problem we hold off doing the AP startup until
2510 * after we have detected the SMI. Hopefully another SMI
2511 * will not occur before we finish the AP startup.
2513 * Retries don't seem to help. SMIs have a window of opportunity
2514 * and if USB->legacy keyboard emulation is enabled in the BIOS
2515 * the interrupt rate can be quite high.
2517 * NOTE: Don't worry about the L1 cache load, it might bloat
2518 * ldelta a little but ndelta will be so huge when the SMI
2519 * occurs the detection logic will still work fine.
2522 set_apic_timer(200000);
2527 * first we do an INIT/RESET IPI this INIT IPI might be run, reseting
2528 * and running the target CPU. OR this INIT IPI might be latched (P5
2529 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
2532 * see apic/apicreg.h for icr bit definitions.
2534 * TIME CRITICAL CODE, DO NOT DO ANY KPRINTFS IN THE HOT PATH.
2538 * Setup the address for the target AP. We can setup
2539 * icr_hi once and then just trigger operations with
2542 icr_hi = lapic->icr_hi & ~APIC_ID_MASK;
2543 icr_hi |= (physical_cpu << 24);
2544 icr_lo = lapic->icr_lo & 0xfff00000;
2545 lapic->icr_hi = icr_hi;
2548 * Do an INIT IPI: assert RESET
2550 * Use edge triggered mode to assert INIT
2552 lapic->icr_lo = icr_lo | 0x00004500;
2553 while (lapic->icr_lo & APIC_DELSTAT_MASK)
2557 * The spec calls for a 10ms delay but we may have to use a
2558 * MUCH lower delay to avoid bricking an AP due to a fast SMI
2559 * interrupt. We have other loops here too and dividing by 2
2560 * doesn't seem to be enough even after subtracting 350us,
2561 * so we divide by 4.
2563 * Our minimum delay is 150uS, maximum is 10ms. If no SMI
2564 * interrupt was detected we use the full 10ms.
2568 else if (smibest < 150 * 4 + 350)
2570 else if ((smibest - 350) / 4 < 10000)
2571 u_sleep((smibest - 350) / 4);
2576 * Do an INIT IPI: deassert RESET
2578 * Use level triggered mode to deassert. It is unclear
2579 * why we need to do this.
2581 lapic->icr_lo = icr_lo | 0x00008500;
2582 while (lapic->icr_lo & APIC_DELSTAT_MASK)
2584 u_sleep(150); /* wait 150us */
2587 * Next we do a STARTUP IPI: the previous INIT IPI might still be
2588 * latched, (P5 bug) this 1st STARTUP would then terminate
2589 * immediately, and the previously started INIT IPI would continue. OR
2590 * the previous INIT IPI has already run. and this STARTUP IPI will
2591 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
2594 lapic->icr_lo = icr_lo | 0x00000600 | vector;
2595 while (lapic->icr_lo & APIC_DELSTAT_MASK)
2597 u_sleep(200); /* wait ~200uS */
2600 * Finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
2601 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
2602 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
2603 * recognized after hardware RESET or INIT IPI.
2605 lapic->icr_lo = icr_lo | 0x00000600 | vector;
2606 while (lapic->icr_lo & APIC_DELSTAT_MASK)
2609 /* Resume normal operation */
2612 /* wait for it to start, see ap_init() */
2613 set_apic_timer(5000000);/* == 5 seconds */
2614 while (read_apic_timer()) {
2615 if (smp_startup_mask & (1 << gd->mi.gd_cpuid))
2616 return 1; /* return SUCCESS */
2619 return 0; /* return FAILURE */
2634 while (read_apic_timer()) {
2636 for (count = 0; count < 100; ++count)
2637 ntsc = rdtsc(); /* force loop to occur */
2639 ndelta = ntsc - ltsc;
2640 if (ldelta > ndelta)
2642 if (ndelta > ldelta * 2)
2645 ldelta = ntsc - ltsc;
2648 return(read_apic_timer());
2652 * Lazy flush the TLB on all other CPU's. DEPRECATED.
2654 * If for some reason we were unable to start all cpus we cannot safely
2655 * use broadcast IPIs.
2661 if (smp_startup_mask == smp_active_mask) {
2662 all_but_self_ipi(XINVLTLB_OFFSET);
2664 selected_apic_ipi(smp_active_mask, XINVLTLB_OFFSET,
2665 APIC_DELMODE_FIXED);
2671 * When called the executing CPU will send an IPI to all other CPUs
2672 * requesting that they halt execution.
2674 * Usually (but not necessarily) called with 'other_cpus' as its arg.
2676 * - Signals all CPUs in map to stop.
2677 * - Waits for each to stop.
2684 * XXX FIXME: this is not MP-safe, needs a lock to prevent multiple CPUs
2685 * from executing at same time.
2688 stop_cpus(u_int map)
2690 map &= smp_active_mask;
2692 /* send the Xcpustop IPI to all CPUs in map */
2693 selected_apic_ipi(map, XCPUSTOP_OFFSET, APIC_DELMODE_FIXED);
2695 while ((stopped_cpus & map) != map)
2703 * Called by a CPU to restart stopped CPUs.
2705 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
2707 * - Signals all CPUs in map to restart.
2708 * - Waits for each to restart.
2716 restart_cpus(u_int map)
2718 /* signal other cpus to restart */
2719 started_cpus = map & smp_active_mask;
2721 while ((stopped_cpus & map) != 0) /* wait for each to clear its bit */
2728 * This is called once the mpboot code has gotten us properly relocated
2729 * and the MMU turned on, etc. ap_init() is actually the idle thread,
2730 * and when it returns the scheduler will call the real cpu_idle() main
2731 * loop for the idlethread. Interrupts are disabled on entry and should
2732 * remain disabled at return.
2740 * Adjust smp_startup_mask to signal the BSP that we have started
2741 * up successfully. Note that we do not yet hold the BGL. The BSP
2742 * is waiting for our signal.
2744 * We can't set our bit in smp_active_mask yet because we are holding
2745 * interrupts physically disabled and remote cpus could deadlock
2746 * trying to send us an IPI.
2748 smp_startup_mask |= 1 << mycpu->gd_cpuid;
2752 * Interlock for finalization. Wait until mp_finish is non-zero,
2753 * then get the MP lock.
2755 * Note: We are in a critical section.
2757 * Note: We have to synchronize td_mpcount to our desired MP state
2758 * before calling cpu_try_mplock().
2760 * Note: we are the idle thread, we can only spin.
2762 * Note: The load fence is memory volatile and prevents the compiler
2763 * from improperly caching mp_finish, and the cpu from improperly
2766 while (mp_finish == 0)
2768 ++curthread->td_mpcount;
2769 while (cpu_try_mplock() == 0)
2772 if (cpu_feature & CPUID_TSC) {
2774 * The BSP is constantly updating tsc0_offset, figure out the
2775 * relative difference to synchronize ktrdump.
2777 tsc_offsets[mycpu->gd_cpuid] = rdtsc() - tsc0_offset;
2780 /* BSP may have changed PTD while we're waiting for the lock */
2783 #if defined(I586_CPU) && !defined(NO_F00F_HACK)
2787 /* Build our map of 'other' CPUs. */
2788 mycpu->gd_other_cpus = smp_startup_mask & ~(1 << mycpu->gd_cpuid);
2790 kprintf("SMP: AP CPU #%d Launched!\n", mycpu->gd_cpuid);
2792 /* A quick check from sanity claus */
2793 apic_id = (apic_id_to_logical[(lapic->id & 0x0f000000) >> 24]);
2794 if (mycpu->gd_cpuid != apic_id) {
2795 kprintf("SMP: cpuid = %d\n", mycpu->gd_cpuid);
2796 kprintf("SMP: apic_id = %d\n", apic_id);
2798 kprintf("PTD[MPPTDI] = %p\n", (void *)PTD[MPPTDI]);
2800 panic("cpuid mismatch! boom!!");
2803 /* Initialize AP's local APIC for irq's */
2804 apic_initialize(FALSE);
2806 /* Set memory range attributes for this CPU to match the BSP */
2807 mem_range_AP_init();
2810 * Once we go active we must process any IPIQ messages that may
2811 * have been queued, because no actual IPI will occur until we
2812 * set our bit in the smp_active_mask. If we don't the IPI
2813 * message interlock could be left set which would also prevent
2816 * The idle loop doesn't expect the BGL to be held and while
2817 * lwkt_switch() normally cleans things up this is a special case
2818 * because we returning almost directly into the idle loop.
2820 * The idle thread is never placed on the runq, make sure
2821 * nothing we've done put it there.
2823 KKASSERT(curthread->td_mpcount == 1);
2824 smp_active_mask |= 1 << mycpu->gd_cpuid;
2827 * Enable interrupts here. idle_restore will also do it, but
2828 * doing it here lets us clean up any strays that got posted to
2829 * the CPU during the AP boot while we are still in a critical
2832 __asm __volatile("sti; pause; pause"::);
2833 mdcpu->gd_fpending = 0;
2835 initclocks_pcpu(); /* clock interrupts (via IPIs) */
2836 lwkt_process_ipiq();
2839 * Releasing the mp lock lets the BSP finish up the SMP init
2842 KKASSERT((curthread->td_flags & TDF_RUNQ) == 0);
2846 * Get SMP fully working before we start initializing devices.
2854 kprintf("Finish MP startup\n");
2855 if (cpu_feature & CPUID_TSC)
2856 tsc0_offset = rdtsc();
2859 while (smp_active_mask != smp_startup_mask) {
2861 if (cpu_feature & CPUID_TSC)
2862 tsc0_offset = rdtsc();
2864 while (try_mplock() == 0)
2867 kprintf("Active CPU Mask: %08x\n", smp_active_mask);
2870 SYSINIT(finishsmp, SI_BOOT2_FINISH_SMP, SI_ORDER_FIRST, ap_finish, NULL)
2873 cpu_send_ipiq(int dcpu)
2875 if ((1 << dcpu) & smp_active_mask)
2876 single_apic_ipi(dcpu, XIPIQ_OFFSET, APIC_DELMODE_FIXED);
2879 #if 0 /* single_apic_ipi_passive() not working yet */
2881 * Returns 0 on failure, 1 on success
2884 cpu_send_ipiq_passive(int dcpu)
2887 if ((1 << dcpu) & smp_active_mask) {
2888 r = single_apic_ipi_passive(dcpu, XIPIQ_OFFSET,
2889 APIC_DELMODE_FIXED);
2895 struct mptable_lapic_cbarg1 {
2898 u_int ht_apicid_mask;
2902 mptable_lapic_pass1_callback(void *xarg, const void *pos, int type)
2904 const struct PROCENTRY *ent;
2905 struct mptable_lapic_cbarg1 *arg = xarg;
2911 if ((ent->cpu_flags & PROCENTRY_FLAG_EN) == 0)
2915 if (ent->apic_id < 32) {
2916 arg->ht_apicid_mask |= 1 << ent->apic_id;
2917 } else if (arg->ht_fixup) {
2918 kprintf("MPTABLE: lapic id > 32, disable HTT fixup\n");
2924 struct mptable_lapic_cbarg2 {
2931 mptable_lapic_pass2_callback(void *xarg, const void *pos, int type)
2933 const struct PROCENTRY *ent;
2934 struct mptable_lapic_cbarg2 *arg = xarg;
2940 if (ent->cpu_flags & PROCENTRY_FLAG_BP) {
2941 KKASSERT(!arg->found_bsp);
2945 if (processor_entry(ent, arg->cpu))
2948 if (arg->logical_cpus) {
2949 struct PROCENTRY proc;
2953 * Create fake mptable processor entries
2954 * and feed them to processor_entry() to
2955 * enumerate the logical CPUs.
2957 bzero(&proc, sizeof(proc));
2959 proc.cpu_flags = PROCENTRY_FLAG_EN;
2960 proc.apic_id = ent->apic_id;
2962 for (i = 1; i < arg->logical_cpus; i++) {
2964 processor_entry(&proc, arg->cpu);
2972 mptable_lapic_default(void)
2974 int ap_apicid, bsp_apicid;
2976 mp_naps = 1; /* exclude BSP */
2978 /* Map local apic before the id field is accessed */
2979 lapic_init(DEFAULT_APIC_BASE);
2981 bsp_apicid = APIC_ID(lapic->id);
2982 ap_apicid = (bsp_apicid == 0) ? 1 : 0;
2985 mp_set_cpuids(0, bsp_apicid);
2986 /* one and only AP */
2987 mp_set_cpuids(1, ap_apicid);
2992 * cpu_apic_address (common to all CPUs)
2994 * ID_TO_CPU(N), APIC ID to logical CPU table
2995 * CPU_TO_ID(N), logical CPU to APIC ID table
2998 mptable_lapic_enumerate(struct mptable_pos *mpt)
3000 struct mptable_lapic_cbarg1 arg1;
3001 struct mptable_lapic_cbarg2 arg2;
3003 int error, logical_cpus = 0;
3004 vm_offset_t lapic_addr;
3006 KKASSERT(mpt->mp_fps != NULL);
3009 * Check for use of 'default' configuration
3011 if (mpt->mp_fps->mpfb1 != 0) {
3012 mptable_lapic_default();
3017 KKASSERT(cth != NULL);
3019 /* Save local apic address */
3020 lapic_addr = (vm_offset_t)cth->apic_address;
3021 KKASSERT(lapic_addr != 0);
3024 * Find out how many CPUs do we have
3026 bzero(&arg1, sizeof(arg1));
3027 arg1.ht_fixup = 1; /* Apply ht fixup by default */
3029 error = mptable_iterate_entries(cth,
3030 mptable_lapic_pass1_callback, &arg1);
3032 panic("mptable_iterate_entries(lapic_pass1) failed\n");
3033 KKASSERT(arg1.cpu_count != 0);
3035 /* See if we need to fixup HT logical CPUs. */
3036 if (arg1.ht_fixup) {
3037 logical_cpus = mptable_hyperthread_fixup(arg1.ht_apicid_mask,
3039 if (logical_cpus != 0)
3040 arg1.cpu_count *= logical_cpus;
3042 mp_naps = arg1.cpu_count;
3044 /* Qualify the numbers again, after possible HT fixup */
3045 if (mp_naps > MAXCPU) {
3046 kprintf("Warning: only using %d of %d available CPUs!\n",
3051 --mp_naps; /* subtract the BSP */
3054 * Link logical CPU id to local apic id
3056 bzero(&arg2, sizeof(arg2));
3058 arg2.logical_cpus = logical_cpus;
3060 error = mptable_iterate_entries(cth,
3061 mptable_lapic_pass2_callback, &arg2);
3063 panic("mptable_iterate_entries(lapic_pass2) failed\n");
3064 KKASSERT(arg2.found_bsp);
3066 /* Map local apic */
3067 lapic_init(lapic_addr);
3071 lapic_init(vm_offset_t lapic_addr)
3074 * lapic not mapped yet (pmap_init is called too late)
3076 lapic = pmap_mapdev_uncacheable(lapic_addr, sizeof(struct LAPIC));
3079 /* Local apic is mapped on last page */
3080 SMPpt[NPTEPG - 1] = (pt_entry_t)(PG_V | PG_RW | PG_N |
3081 pmap_get_pgeflag() | (lapic_addr & PG_FRAME));
3084 /* Just for printing */
3085 cpu_apic_address = lapic_addr;