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>
41 #include <vm/vm_param.h>
43 #include <vm/vm_kern.h>
44 #include <vm/vm_extern.h>
46 #include <vm/vm_map.h>
52 #include <sys/mplock2.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/cputypes.h>
59 #include <machine_base/icu/icu_var.h>
60 #include <machine_base/apic/ioapic_abi.h>
61 #include <machine_base/apic/lapic.h>
62 #include <machine_base/apic/ioapic.h>
63 #include <machine/psl.h>
64 #include <machine/segments.h>
65 #include <machine/tss.h>
66 #include <machine/specialreg.h>
67 #include <machine/globaldata.h>
68 #include <machine/pmap_inval.h>
70 #include <machine/md_var.h> /* setidt() */
71 #include <machine_base/icu/icu.h> /* IPIs */
72 #include <machine/intr_machdep.h> /* IPIs */
74 #define WARMBOOT_TARGET 0
75 #define WARMBOOT_OFF (KERNBASE + 0x0467)
76 #define WARMBOOT_SEG (KERNBASE + 0x0469)
78 #define CMOS_REG (0x70)
79 #define CMOS_DATA (0x71)
80 #define BIOS_RESET (0x0f)
81 #define BIOS_WARM (0x0a)
84 * this code MUST be enabled here and in mpboot.s.
85 * it follows the very early stages of AP boot by placing values in CMOS ram.
86 * it NORMALLY will never be needed and thus the primitive method for enabling.
89 #if defined(CHECK_POINTS)
90 #define CHECK_READ(A) (outb(CMOS_REG, (A)), inb(CMOS_DATA))
91 #define CHECK_WRITE(A,D) (outb(CMOS_REG, (A)), outb(CMOS_DATA, (D)))
93 #define CHECK_INIT(D); \
94 CHECK_WRITE(0x34, (D)); \
95 CHECK_WRITE(0x35, (D)); \
96 CHECK_WRITE(0x36, (D)); \
97 CHECK_WRITE(0x37, (D)); \
98 CHECK_WRITE(0x38, (D)); \
99 CHECK_WRITE(0x39, (D));
101 #define CHECK_PRINT(S); \
102 kprintf("%s: %d, %d, %d, %d, %d, %d\n", \
111 #else /* CHECK_POINTS */
113 #define CHECK_INIT(D)
114 #define CHECK_PRINT(S)
116 #endif /* CHECK_POINTS */
119 * Values to send to the POST hardware.
121 #define MP_BOOTADDRESS_POST 0x10
122 #define MP_PROBE_POST 0x11
123 #define MPTABLE_PASS1_POST 0x12
125 #define MP_START_POST 0x13
126 #define MP_ENABLE_POST 0x14
127 #define MPTABLE_PASS2_POST 0x15
129 #define START_ALL_APS_POST 0x16
130 #define INSTALL_AP_TRAMP_POST 0x17
131 #define START_AP_POST 0x18
133 #define MP_ANNOUNCE_POST 0x19
135 /** XXX FIXME: where does this really belong, isa.h/isa.c perhaps? */
136 int current_postcode;
138 /** XXX FIXME: what system files declare these??? */
139 extern struct region_descriptor r_gdt, r_idt;
145 extern int64_t tsc_offsets[];
147 /* AP uses this during bootstrap. Do not staticize. */
151 /* Hotwire a 0->4MB V==P mapping */
152 extern pt_entry_t *KPTphys;
155 * SMP page table page. Setup by locore to point to a page table
156 * page from which we allocate per-cpu privatespace areas io_apics,
159 extern pt_entry_t *SMPpt;
161 struct pcb stoppcbs[MAXCPU];
164 * Local data and functions.
167 static u_int boot_address;
168 static int mp_finish;
169 static int mp_finish_lapic;
171 static void mp_enable(u_int boot_addr);
173 static int start_all_aps(u_int boot_addr);
174 static void install_ap_tramp(u_int boot_addr);
175 static int start_ap(struct mdglobaldata *gd, u_int boot_addr, int smibest);
176 static int smitest(void);
177 static void cpu_simple_setup(void);
179 static cpumask_t smp_startup_mask = 1; /* which cpus have been started */
180 static cpumask_t smp_lapic_mask = 1; /* which cpus have lapic been inited */
181 cpumask_t smp_active_mask = 1; /* which cpus are ready for IPIs etc? */
182 SYSCTL_INT(_machdep, OID_AUTO, smp_active, CTLFLAG_RD, &smp_active_mask, 0, "");
185 * Calculate usable address in base memory for AP trampoline code.
188 mp_bootaddress(u_int basemem)
190 POSTCODE(MP_BOOTADDRESS_POST);
192 boot_address = basemem & ~0xfff; /* round down to 4k boundary */
193 if ((basemem - boot_address) < bootMP_size)
194 boot_address -= 4096; /* not enough, lower by 4k */
200 * Startup the SMP processors.
205 POSTCODE(MP_START_POST);
206 mp_enable(boot_address);
211 * Print various information about the SMP system hardware and setup.
218 POSTCODE(MP_ANNOUNCE_POST);
220 kprintf("DragonFly/MP: Multiprocessor motherboard\n");
221 kprintf(" cpu0 (BSP): apic id: %2d\n", CPUID_TO_APICID(0));
222 for (x = 1; x <= naps; ++x)
223 kprintf(" cpu%d (AP): apic id: %2d\n", x, CPUID_TO_APICID(x));
226 kprintf(" Warning: APIC I/O disabled\n");
230 * AP cpu's call this to sync up protected mode.
232 * WARNING! We must ensure that the cpu is sufficiently initialized to
233 * be able to use to the FP for our optimized bzero/bcopy code before
234 * we enter more mainstream C code.
236 * WARNING! %fs is not set up on entry. This routine sets up %fs.
242 int x, myid = bootAP;
244 struct mdglobaldata *md;
245 struct privatespace *ps;
247 ps = &CPU_prvspace[myid];
249 gdt_segs[GPRIV_SEL].ssd_base = (int)ps;
250 gdt_segs[GPROC0_SEL].ssd_base =
251 (int) &ps->mdglobaldata.gd_common_tss;
252 ps->mdglobaldata.mi.gd_prvspace = ps;
254 for (x = 0; x < NGDT; x++) {
255 ssdtosd(&gdt_segs[x], &gdt[myid * NGDT + x].sd);
258 r_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
259 r_gdt.rd_base = (int) &gdt[myid * NGDT];
260 lgdt(&r_gdt); /* does magic intra-segment return */
265 mdcpu->gd_currentldt = _default_ldt;
267 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
268 gdt[myid * NGDT + GPROC0_SEL].sd.sd_type = SDT_SYS386TSS;
270 md = mdcpu; /* loaded through %fs:0 (mdglobaldata.mi.gd_prvspace)*/
272 md->gd_common_tss.tss_esp0 = 0; /* not used until after switch */
273 md->gd_common_tss.tss_ss0 = GSEL(GDATA_SEL, SEL_KPL);
274 md->gd_common_tss.tss_ioopt = (sizeof md->gd_common_tss) << 16;
275 md->gd_tss_gdt = &gdt[myid * NGDT + GPROC0_SEL].sd;
276 md->gd_common_tssd = *md->gd_tss_gdt;
280 * Set to a known state:
281 * Set by mpboot.s: CR0_PG, CR0_PE
282 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
285 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
287 pmap_set_opt(); /* PSE/4MB pages, etc */
289 /* set up CPU registers and state */
292 /* set up FPU state on the AP */
293 npxinit(__INITIAL_NPXCW__);
295 /* set up SSE registers */
299 /*******************************************************************
300 * local functions and data
304 * start the SMP system
307 mp_enable(u_int boot_addr)
311 POSTCODE(MP_ENABLE_POST);
314 error = lapic_config();
320 /* Initialize BSP's local APIC */
322 } else if (ioapic_enable) {
324 icu_reinit_noioapic();
328 /* start each Application Processor */
329 start_all_aps(boot_addr);
335 KASSERT(lapic_enable,
336 ("I/O APIC is enabled, but LAPIC is disabled\n"));
337 error = ioapic_config();
340 icu_reinit_noioapic();
341 lapic_fixup_noioapic();
347 * start each AP in our list
350 start_all_aps(u_int boot_addr)
358 u_long mpbioswarmvec;
359 struct mdglobaldata *gd;
360 struct privatespace *ps;
364 POSTCODE(START_ALL_APS_POST);
366 /* install the AP 1st level boot code */
367 install_ap_tramp(boot_addr);
370 /* save the current value of the warm-start vector */
371 mpbioswarmvec = *((u_long *) WARMBOOT_OFF);
372 outb(CMOS_REG, BIOS_RESET);
373 mpbiosreason = inb(CMOS_DATA);
375 /* setup a vector to our boot code */
376 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
377 *((volatile u_short *) WARMBOOT_SEG) = (boot_addr >> 4);
378 outb(CMOS_REG, BIOS_RESET);
379 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
382 * If we have a TSC we can figure out the SMI interrupt rate.
383 * The SMI does not necessarily use a constant rate. Spend
384 * up to 250ms trying to figure it out.
387 if (cpu_feature & CPUID_TSC) {
388 set_apic_timer(275000);
389 smilast = read_apic_timer();
390 for (x = 0; x < 20 && read_apic_timer(); ++x) {
391 smicount = smitest();
392 if (smibest == 0 || smilast - smicount < smibest)
393 smibest = smilast - smicount;
396 if (smibest > 250000)
399 smibest = smibest * (int64_t)1000000 /
400 get_apic_timer_frequency();
404 kprintf("SMI Frequency (worst case): %d Hz (%d us)\n",
405 1000000 / smibest, smibest);
408 /* set up temporary P==V mapping for AP boot */
409 /* XXX this is a hack, we should boot the AP on its own stack/PTD */
410 kptbase = (uintptr_t)(void *)KPTphys;
411 for (x = 0; x < NKPT; x++) {
412 PTD[x] = (pd_entry_t)(PG_V | PG_RW |
413 ((kptbase + x * PAGE_SIZE) & PG_FRAME));
418 for (x = 1; x <= naps; ++x) {
420 /* This is a bit verbose, it will go away soon. */
422 /* first page of AP's private space */
423 pg = x * i386_btop(sizeof(struct privatespace));
425 /* allocate new private data page(s) */
426 gd = (struct mdglobaldata *)kmem_alloc(&kernel_map,
427 MDGLOBALDATA_BASEALLOC_SIZE);
428 /* wire it into the private page table page */
429 for (i = 0; i < MDGLOBALDATA_BASEALLOC_SIZE; i += PAGE_SIZE) {
430 SMPpt[pg + i / PAGE_SIZE] = (pt_entry_t)
431 (PG_V | PG_RW | vtophys_pte((char *)gd + i));
433 pg += MDGLOBALDATA_BASEALLOC_PAGES;
435 SMPpt[pg + 0] = 0; /* *gd_CMAP1 */
436 SMPpt[pg + 1] = 0; /* *gd_CMAP2 */
437 SMPpt[pg + 2] = 0; /* *gd_CMAP3 */
438 SMPpt[pg + 3] = 0; /* *gd_PMAP1 */
440 /* allocate and set up an idle stack data page */
441 stack = (char *)kmem_alloc(&kernel_map, UPAGES*PAGE_SIZE);
442 for (i = 0; i < UPAGES; i++) {
443 SMPpt[pg + 4 + i] = (pt_entry_t)
444 (PG_V | PG_RW | vtophys_pte(PAGE_SIZE * i + stack));
447 gd = &CPU_prvspace[x].mdglobaldata; /* official location */
448 bzero(gd, sizeof(*gd));
449 gd->mi.gd_prvspace = ps = &CPU_prvspace[x];
451 /* prime data page for it to use */
452 mi_gdinit(&gd->mi, x);
454 gd->gd_CMAP1 = &SMPpt[pg + 0];
455 gd->gd_CMAP2 = &SMPpt[pg + 1];
456 gd->gd_CMAP3 = &SMPpt[pg + 2];
457 gd->gd_PMAP1 = &SMPpt[pg + 3];
458 gd->gd_CADDR1 = ps->CPAGE1;
459 gd->gd_CADDR2 = ps->CPAGE2;
460 gd->gd_CADDR3 = ps->CPAGE3;
461 gd->gd_PADDR1 = (unsigned *)ps->PPAGE1;
464 * Per-cpu pmap for get_ptbase().
466 gd->gd_GDADDR1= (unsigned *)
467 kmem_alloc_nofault(&kernel_map, SEG_SIZE, SEG_SIZE);
468 gd->gd_GDMAP1 = &PTD[(vm_offset_t)gd->gd_GDADDR1 >> PDRSHIFT];
470 gd->mi.gd_ipiq = (void *)kmem_alloc(&kernel_map, sizeof(lwkt_ipiq) * (naps + 1));
471 bzero(gd->mi.gd_ipiq, sizeof(lwkt_ipiq) * (naps + 1));
474 * Setup the AP boot stack
476 bootSTK = &ps->idlestack[UPAGES*PAGE_SIZE/2];
479 /* attempt to start the Application Processor */
480 CHECK_INIT(99); /* setup checkpoints */
481 if (!start_ap(gd, boot_addr, smibest)) {
482 kprintf("AP #%d (PHY# %d) failed!\n", x,
484 CHECK_PRINT("trace"); /* show checkpoints */
485 /* better panic as the AP may be running loose */
486 kprintf("panic y/n? [y] ");
490 CHECK_PRINT("trace"); /* show checkpoints */
493 /* set ncpus to 1 + highest logical cpu. Not all may have come up */
496 /* ncpus2 -- ncpus rounded down to the nearest power of 2 */
497 for (shift = 0; (1 << shift) <= ncpus; ++shift)
500 ncpus2_shift = shift;
502 ncpus2_mask = ncpus2 - 1;
504 /* ncpus_fit -- ncpus rounded up to the nearest power of 2 */
505 if ((1 << shift) < ncpus)
507 ncpus_fit = 1 << shift;
508 ncpus_fit_mask = ncpus_fit - 1;
510 /* build our map of 'other' CPUs */
511 mycpu->gd_other_cpus = smp_startup_mask & ~CPUMASK(mycpu->gd_cpuid);
512 mycpu->gd_ipiq = (void *)kmem_alloc(&kernel_map, sizeof(lwkt_ipiq) * ncpus);
513 bzero(mycpu->gd_ipiq, sizeof(lwkt_ipiq) * ncpus);
515 /* restore the warmstart vector */
516 *(u_long *) WARMBOOT_OFF = mpbioswarmvec;
517 outb(CMOS_REG, BIOS_RESET);
518 outb(CMOS_DATA, mpbiosreason);
521 * NOTE! The idlestack for the BSP was setup by locore. Finish
522 * up, clean out the P==V mapping we did earlier.
524 for (x = 0; x < NKPT; x++)
529 * Wait all APs to finish initializing LAPIC
533 kprintf("SMP: Waiting APs LAPIC initialization\n");
534 if (cpu_feature & CPUID_TSC)
535 tsc0_offset = rdtsc();
538 while (smp_lapic_mask != smp_startup_mask) {
540 if (cpu_feature & CPUID_TSC)
541 tsc0_offset = rdtsc();
543 while (try_mplock() == 0)
546 /* number of APs actually started */
551 * load the 1st level AP boot code into base memory.
554 /* targets for relocation */
555 extern void bigJump(void);
556 extern void bootCodeSeg(void);
557 extern void bootDataSeg(void);
558 extern void MPentry(void);
560 extern u_int mp_gdtbase;
563 install_ap_tramp(u_int boot_addr)
566 int size = *(int *) ((u_long) & bootMP_size);
567 u_char *src = (u_char *) ((u_long) bootMP);
568 u_char *dst = (u_char *) boot_addr + KERNBASE;
569 u_int boot_base = (u_int) bootMP;
574 POSTCODE(INSTALL_AP_TRAMP_POST);
576 for (x = 0; x < size; ++x)
580 * modify addresses in code we just moved to basemem. unfortunately we
581 * need fairly detailed info about mpboot.s for this to work. changes
582 * to mpboot.s might require changes here.
585 /* boot code is located in KERNEL space */
586 dst = (u_char *) boot_addr + KERNBASE;
588 /* modify the lgdt arg */
589 dst32 = (u_int32_t *) (dst + ((u_int) & mp_gdtbase - boot_base));
590 *dst32 = boot_addr + ((u_int) & MP_GDT - boot_base);
592 /* modify the ljmp target for MPentry() */
593 dst32 = (u_int32_t *) (dst + ((u_int) bigJump - boot_base) + 1);
594 *dst32 = ((u_int) MPentry - KERNBASE);
596 /* modify the target for boot code segment */
597 dst16 = (u_int16_t *) (dst + ((u_int) bootCodeSeg - boot_base));
598 dst8 = (u_int8_t *) (dst16 + 1);
599 *dst16 = (u_int) boot_addr & 0xffff;
600 *dst8 = ((u_int) boot_addr >> 16) & 0xff;
602 /* modify the target for boot data segment */
603 dst16 = (u_int16_t *) (dst + ((u_int) bootDataSeg - boot_base));
604 dst8 = (u_int8_t *) (dst16 + 1);
605 *dst16 = (u_int) boot_addr & 0xffff;
606 *dst8 = ((u_int) boot_addr >> 16) & 0xff;
611 * This function starts the AP (application processor) identified
612 * by the APIC ID 'physicalCpu'. It does quite a "song and dance"
613 * to accomplish this. This is necessary because of the nuances
614 * of the different hardware we might encounter. It ain't pretty,
615 * but it seems to work.
617 * NOTE: eventually an AP gets to ap_init(), which is called just
618 * before the AP goes into the LWKT scheduler's idle loop.
621 start_ap(struct mdglobaldata *gd, u_int boot_addr, int smibest)
625 u_long icr_lo, icr_hi;
627 POSTCODE(START_AP_POST);
629 /* get the PHYSICAL APIC ID# */
630 physical_cpu = CPUID_TO_APICID(gd->mi.gd_cpuid);
632 /* calculate the vector */
633 vector = (boot_addr >> 12) & 0xff;
635 /* We don't want anything interfering */
638 /* Make sure the target cpu sees everything */
642 * Try to detect when a SMI has occurred, wait up to 200ms.
644 * If a SMI occurs during an AP reset but before we issue
645 * the STARTUP command, the AP may brick. To work around
646 * this problem we hold off doing the AP startup until
647 * after we have detected the SMI. Hopefully another SMI
648 * will not occur before we finish the AP startup.
650 * Retries don't seem to help. SMIs have a window of opportunity
651 * and if USB->legacy keyboard emulation is enabled in the BIOS
652 * the interrupt rate can be quite high.
654 * NOTE: Don't worry about the L1 cache load, it might bloat
655 * ldelta a little but ndelta will be so huge when the SMI
656 * occurs the detection logic will still work fine.
659 set_apic_timer(200000);
664 * first we do an INIT/RESET IPI this INIT IPI might be run, reseting
665 * and running the target CPU. OR this INIT IPI might be latched (P5
666 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
669 * see apic/apicreg.h for icr bit definitions.
671 * TIME CRITICAL CODE, DO NOT DO ANY KPRINTFS IN THE HOT PATH.
675 * Setup the address for the target AP. We can setup
676 * icr_hi once and then just trigger operations with
679 icr_hi = lapic->icr_hi & ~APIC_ID_MASK;
680 icr_hi |= (physical_cpu << 24);
681 icr_lo = lapic->icr_lo & 0xfff00000;
682 lapic->icr_hi = icr_hi;
685 * Do an INIT IPI: assert RESET
687 * Use edge triggered mode to assert INIT
689 lapic->icr_lo = icr_lo | 0x0000c500;
690 while (lapic->icr_lo & APIC_DELSTAT_MASK)
694 * The spec calls for a 10ms delay but we may have to use a
695 * MUCH lower delay to avoid bricking an AP due to a fast SMI
696 * interrupt. We have other loops here too and dividing by 2
697 * doesn't seem to be enough even after subtracting 350us,
700 * Our minimum delay is 150uS, maximum is 10ms. If no SMI
701 * interrupt was detected we use the full 10ms.
705 else if (smibest < 150 * 4 + 350)
707 else if ((smibest - 350) / 4 < 10000)
708 u_sleep((smibest - 350) / 4);
713 * Do an INIT IPI: deassert RESET
715 * Use level triggered mode to deassert. It is unclear
716 * why we need to do this.
718 lapic->icr_lo = icr_lo | 0x00008500;
719 while (lapic->icr_lo & APIC_DELSTAT_MASK)
721 u_sleep(150); /* wait 150us */
724 * Next we do a STARTUP IPI: the previous INIT IPI might still be
725 * latched, (P5 bug) this 1st STARTUP would then terminate
726 * immediately, and the previously started INIT IPI would continue. OR
727 * the previous INIT IPI has already run. and this STARTUP IPI will
728 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
731 lapic->icr_lo = icr_lo | 0x00000600 | vector;
732 while (lapic->icr_lo & APIC_DELSTAT_MASK)
734 u_sleep(200); /* wait ~200uS */
737 * Finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
738 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
739 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
740 * recognized after hardware RESET or INIT IPI.
742 lapic->icr_lo = icr_lo | 0x00000600 | vector;
743 while (lapic->icr_lo & APIC_DELSTAT_MASK)
746 /* Resume normal operation */
749 /* wait for it to start, see ap_init() */
750 set_apic_timer(5000000);/* == 5 seconds */
751 while (read_apic_timer()) {
752 if (smp_startup_mask & CPUMASK(gd->mi.gd_cpuid))
753 return 1; /* return SUCCESS */
756 return 0; /* return FAILURE */
771 while (read_apic_timer()) {
773 for (count = 0; count < 100; ++count)
774 ntsc = rdtsc(); /* force loop to occur */
776 ndelta = ntsc - ltsc;
779 if (ndelta > ldelta * 2)
782 ldelta = ntsc - ltsc;
785 return(read_apic_timer());
789 * Lazy flush the TLB on all other CPU's. DEPRECATED.
791 * If for some reason we were unable to start all cpus we cannot safely
792 * use broadcast IPIs.
795 static cpumask_t smp_invltlb_req;
796 #define SMP_INVLTLB_DEBUG
802 struct mdglobaldata *md = mdcpu;
803 #ifdef SMP_INVLTLB_DEBUG
808 crit_enter_gd(&md->mi);
809 md->gd_invltlb_ret = 0;
810 ++md->mi.gd_cnt.v_smpinvltlb;
811 atomic_set_cpumask(&smp_invltlb_req, md->mi.gd_cpumask);
812 #ifdef SMP_INVLTLB_DEBUG
815 if (smp_startup_mask == smp_active_mask) {
816 all_but_self_ipi(XINVLTLB_OFFSET);
818 selected_apic_ipi(smp_active_mask & ~md->mi.gd_cpumask,
819 XINVLTLB_OFFSET, APIC_DELMODE_FIXED);
822 #ifdef SMP_INVLTLB_DEBUG
824 kprintf("smp_invltlb: ipi sent\n");
826 while ((md->gd_invltlb_ret & smp_active_mask & ~md->mi.gd_cpumask) !=
827 (smp_active_mask & ~md->mi.gd_cpumask)) {
830 #ifdef SMP_INVLTLB_DEBUG
832 if (++count == 400000000) {
834 kprintf("smp_invltlb: endless loop %08lx %08lx, "
835 "rflags %016lx retry",
836 (long)md->gd_invltlb_ret,
837 (long)smp_invltlb_req,
838 (long)read_eflags());
839 __asm __volatile ("sti");
844 int bcpu = BSFCPUMASK(~md->gd_invltlb_ret &
848 kprintf("bcpu %d\n", bcpu);
849 xgd = globaldata_find(bcpu);
850 kprintf("thread %p %s\n", xgd->gd_curthread, xgd->gd_curthread->td_comm);
859 atomic_clear_cpumask(&smp_invltlb_req, md->mi.gd_cpumask);
860 crit_exit_gd(&md->mi);
867 * Called from Xinvltlb assembly with interrupts disabled. We didn't
868 * bother to bump the critical section count or nested interrupt count
869 * so only do very low level operations here.
872 smp_invltlb_intr(void)
874 struct mdglobaldata *md = mdcpu;
875 struct mdglobaldata *omd;
879 mask = smp_invltlb_req;
883 cpu = BSFCPUMASK(mask);
884 mask &= ~CPUMASK(cpu);
885 omd = (struct mdglobaldata *)globaldata_find(cpu);
886 atomic_set_cpumask(&omd->gd_invltlb_ret, md->mi.gd_cpumask);
893 * When called the executing CPU will send an IPI to all other CPUs
894 * requesting that they halt execution.
896 * Usually (but not necessarily) called with 'other_cpus' as its arg.
898 * - Signals all CPUs in map to stop.
899 * - Waits for each to stop.
906 * XXX FIXME: this is not MP-safe, needs a lock to prevent multiple CPUs
907 * from executing at same time.
910 stop_cpus(cpumask_t map)
912 map &= smp_active_mask;
914 /* send the Xcpustop IPI to all CPUs in map */
915 selected_apic_ipi(map, XCPUSTOP_OFFSET, APIC_DELMODE_FIXED);
917 while ((stopped_cpus & map) != map)
925 * Called by a CPU to restart stopped CPUs.
927 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
929 * - Signals all CPUs in map to restart.
930 * - Waits for each to restart.
938 restart_cpus(cpumask_t map)
940 /* signal other cpus to restart */
941 started_cpus = map & smp_active_mask;
943 while ((stopped_cpus & map) != 0) /* wait for each to clear its bit */
950 * This is called once the mpboot code has gotten us properly relocated
951 * and the MMU turned on, etc. ap_init() is actually the idle thread,
952 * and when it returns the scheduler will call the real cpu_idle() main
953 * loop for the idlethread. Interrupts are disabled on entry and should
954 * remain disabled at return.
962 * Adjust smp_startup_mask to signal the BSP that we have started
963 * up successfully. Note that we do not yet hold the BGL. The BSP
964 * is waiting for our signal.
966 * We can't set our bit in smp_active_mask yet because we are holding
967 * interrupts physically disabled and remote cpus could deadlock
968 * trying to send us an IPI.
970 smp_startup_mask |= CPUMASK(mycpu->gd_cpuid);
974 * Interlock for LAPIC initialization. Wait until mp_finish_lapic is
975 * non-zero, then get the MP lock.
977 * Note: We are in a critical section.
979 * Note: we are the idle thread, we can only spin.
981 * Note: The load fence is memory volatile and prevents the compiler
982 * from improperly caching mp_finish_lapic, and the cpu from improperly
985 while (mp_finish_lapic == 0)
987 while (try_mplock() == 0)
990 if (cpu_feature & CPUID_TSC) {
992 * The BSP is constantly updating tsc0_offset, figure out
993 * the relative difference to synchronize ktrdump.
995 tsc_offsets[mycpu->gd_cpuid] = rdtsc() - tsc0_offset;
998 /* BSP may have changed PTD while we're waiting for the lock */
1001 #if defined(I586_CPU) && !defined(NO_F00F_HACK)
1005 /* Build our map of 'other' CPUs. */
1006 mycpu->gd_other_cpus = smp_startup_mask & ~CPUMASK(mycpu->gd_cpuid);
1008 /* A quick check from sanity claus */
1009 cpu_id = APICID_TO_CPUID((lapic->id & 0xff000000) >> 24);
1010 if (mycpu->gd_cpuid != cpu_id) {
1011 kprintf("SMP: assigned cpuid = %d\n", mycpu->gd_cpuid);
1012 kprintf("SMP: actual cpuid = %d\n", cpu_id);
1013 kprintf("PTD[MPPTDI] = %p\n", (void *)PTD[MPPTDI]);
1014 panic("cpuid mismatch! boom!!");
1017 /* Initialize AP's local APIC for irq's */
1020 /* LAPIC initialization is done */
1021 smp_lapic_mask |= CPUMASK(mycpu->gd_cpuid);
1024 /* Let BSP move onto the next initialization stage */
1028 * Interlock for finalization. Wait until mp_finish is non-zero,
1029 * then get the MP lock.
1031 * Note: We are in a critical section.
1033 * Note: we are the idle thread, we can only spin.
1035 * Note: The load fence is memory volatile and prevents the compiler
1036 * from improperly caching mp_finish, and the cpu from improperly
1039 while (mp_finish == 0)
1041 while (try_mplock() == 0)
1044 /* BSP may have changed PTD while we're waiting for the lock */
1047 /* Set memory range attributes for this CPU to match the BSP */
1048 mem_range_AP_init();
1051 * Once we go active we must process any IPIQ messages that may
1052 * have been queued, because no actual IPI will occur until we
1053 * set our bit in the smp_active_mask. If we don't the IPI
1054 * message interlock could be left set which would also prevent
1057 * The idle loop doesn't expect the BGL to be held and while
1058 * lwkt_switch() normally cleans things up this is a special case
1059 * because we returning almost directly into the idle loop.
1061 * The idle thread is never placed on the runq, make sure
1062 * nothing we've done put it there.
1064 KKASSERT(get_mplock_count(curthread) == 1);
1065 smp_active_mask |= CPUMASK(mycpu->gd_cpuid);
1068 * Enable interrupts here. idle_restore will also do it, but
1069 * doing it here lets us clean up any strays that got posted to
1070 * the CPU during the AP boot while we are still in a critical
1073 __asm __volatile("sti; pause; pause"::);
1074 bzero(mdcpu->gd_ipending, sizeof(mdcpu->gd_ipending));
1076 initclocks_pcpu(); /* clock interrupts (via IPIs) */
1077 lwkt_process_ipiq();
1080 * Releasing the mp lock lets the BSP finish up the SMP init
1083 KKASSERT((curthread->td_flags & TDF_RUNQ) == 0);
1087 * Get SMP fully working before we start initializing devices.
1095 kprintf("Finish MP startup\n");
1097 while (smp_active_mask != smp_startup_mask)
1099 while (try_mplock() == 0)
1102 kprintf("Active CPU Mask: %08x\n", smp_active_mask);
1105 SYSINIT(finishsmp, SI_BOOT2_FINISH_SMP, SI_ORDER_FIRST, ap_finish, NULL)
1108 cpu_send_ipiq(int dcpu)
1110 if (CPUMASK(dcpu) & smp_active_mask)
1111 single_apic_ipi(dcpu, XIPIQ_OFFSET, APIC_DELMODE_FIXED);
1114 #if 0 /* single_apic_ipi_passive() not working yet */
1116 * Returns 0 on failure, 1 on success
1119 cpu_send_ipiq_passive(int dcpu)
1122 if (CPUMASK(dcpu) & smp_active_mask) {
1123 r = single_apic_ipi_passive(dcpu, XIPIQ_OFFSET,
1124 APIC_DELMODE_FIXED);
1131 cpu_simple_setup(void)
1133 /* build our map of 'other' CPUs */
1134 mycpu->gd_other_cpus = smp_startup_mask & ~CPUMASK(mycpu->gd_cpuid);
1135 mycpu->gd_ipiq = (void *)kmem_alloc(&kernel_map, sizeof(lwkt_ipiq) * ncpus);
1136 bzero(mycpu->gd_ipiq, sizeof(lwkt_ipiq) * ncpus);
1140 if (cpu_feature & CPUID_TSC)
1141 tsc0_offset = rdtsc();