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 $
30 #include <sys/param.h>
31 #include <sys/systm.h>
32 #include <sys/kernel.h>
33 #include <sys/sysctl.h>
34 #include <sys/malloc.h>
35 #include <sys/memrange.h>
36 #include <sys/cons.h> /* cngetc() */
37 #include <sys/machintr.h>
40 #include <vm/vm_param.h>
42 #include <vm/vm_kern.h>
43 #include <vm/vm_extern.h>
45 #include <vm/vm_map.h>
51 #include <sys/mplock2.h>
53 #include <machine/smp.h>
54 #include <machine_base/apic/apicreg.h>
55 #include <machine/atomic.h>
56 #include <machine/cpufunc.h>
57 #include <machine/cputypes.h>
58 #include <machine_base/icu/icu_var.h>
59 #include <machine_base/apic/ioapic_abi.h>
60 #include <machine_base/apic/lapic.h>
61 #include <machine_base/apic/ioapic.h>
62 #include <machine/psl.h>
63 #include <machine/segments.h>
64 #include <machine/tss.h>
65 #include <machine/specialreg.h>
66 #include <machine/globaldata.h>
67 #include <machine/pmap_inval.h>
69 #include <machine/md_var.h> /* setidt() */
70 #include <machine_base/icu/icu.h> /* IPIs */
71 #include <machine/intr_machdep.h> /* IPIs */
73 #define WARMBOOT_TARGET 0
74 #define WARMBOOT_OFF (KERNBASE + 0x0467)
75 #define WARMBOOT_SEG (KERNBASE + 0x0469)
77 #define CMOS_REG (0x70)
78 #define CMOS_DATA (0x71)
79 #define BIOS_RESET (0x0f)
80 #define BIOS_WARM (0x0a)
83 * this code MUST be enabled here and in mpboot.s.
84 * it follows the very early stages of AP boot by placing values in CMOS ram.
85 * it NORMALLY will never be needed and thus the primitive method for enabling.
88 #if defined(CHECK_POINTS)
89 #define CHECK_READ(A) (outb(CMOS_REG, (A)), inb(CMOS_DATA))
90 #define CHECK_WRITE(A,D) (outb(CMOS_REG, (A)), outb(CMOS_DATA, (D)))
92 #define CHECK_INIT(D); \
93 CHECK_WRITE(0x34, (D)); \
94 CHECK_WRITE(0x35, (D)); \
95 CHECK_WRITE(0x36, (D)); \
96 CHECK_WRITE(0x37, (D)); \
97 CHECK_WRITE(0x38, (D)); \
98 CHECK_WRITE(0x39, (D));
100 #define CHECK_PRINT(S); \
101 kprintf("%s: %d, %d, %d, %d, %d, %d\n", \
110 #else /* CHECK_POINTS */
112 #define CHECK_INIT(D)
113 #define CHECK_PRINT(S)
115 #endif /* CHECK_POINTS */
118 * Values to send to the POST hardware.
120 #define MP_BOOTADDRESS_POST 0x10
121 #define MP_PROBE_POST 0x11
122 #define MPTABLE_PASS1_POST 0x12
124 #define MP_START_POST 0x13
125 #define MP_ENABLE_POST 0x14
126 #define MPTABLE_PASS2_POST 0x15
128 #define START_ALL_APS_POST 0x16
129 #define INSTALL_AP_TRAMP_POST 0x17
130 #define START_AP_POST 0x18
132 #define MP_ANNOUNCE_POST 0x19
134 /** XXX FIXME: where does this really belong, isa.h/isa.c perhaps? */
135 int current_postcode;
137 /** XXX FIXME: what system files declare these??? */
138 extern struct region_descriptor r_gdt, r_idt;
144 extern int64_t tsc_offsets[];
146 /* AP uses this during bootstrap. Do not staticize. */
150 /* Hotwire a 0->4MB V==P mapping */
151 extern pt_entry_t *KPTphys;
154 * SMP page table page. Setup by locore to point to a page table
155 * page from which we allocate per-cpu privatespace areas io_apics,
158 extern pt_entry_t *SMPpt;
160 struct pcb stoppcbs[MAXCPU];
163 * Local data and functions.
166 static u_int boot_address;
167 static int mp_finish;
168 static int mp_finish_lapic;
170 static int start_all_aps(u_int boot_addr);
171 static void install_ap_tramp(u_int boot_addr);
172 static int start_ap(struct mdglobaldata *gd, u_int boot_addr, int smibest);
173 static int smitest(void);
174 static void mp_bsp_simple_setup(void);
176 static cpumask_t smp_startup_mask = 1; /* which cpus have been started */
177 static cpumask_t smp_lapic_mask = 1; /* which cpus have lapic been inited */
178 cpumask_t smp_active_mask = 1; /* which cpus are ready for IPIs etc? */
179 SYSCTL_INT(_machdep, OID_AUTO, smp_active, CTLFLAG_RD, &smp_active_mask, 0, "");
181 /* Local data for detecting CPU TOPOLOGY */
182 static int core_bits = 0;
183 static int logical_CPU_bits = 0;
187 * Calculate usable address in base memory for AP trampoline code.
190 mp_bootaddress(u_int basemem)
192 POSTCODE(MP_BOOTADDRESS_POST);
194 boot_address = basemem & ~0xfff; /* round down to 4k boundary */
195 if ((basemem - boot_address) < bootMP_size)
196 boot_address -= 4096; /* not enough, lower by 4k */
202 * Print various information about the SMP system hardware and setup.
209 POSTCODE(MP_ANNOUNCE_POST);
211 kprintf("DragonFly/MP: Multiprocessor motherboard\n");
212 kprintf(" cpu0 (BSP): apic id: %2d\n", CPUID_TO_APICID(0));
213 for (x = 1; x <= naps; ++x)
214 kprintf(" cpu%d (AP): apic id: %2d\n", x, CPUID_TO_APICID(x));
217 kprintf(" Warning: APIC I/O disabled\n");
221 * AP cpu's call this to sync up protected mode.
223 * WARNING! We must ensure that the cpu is sufficiently initialized to
224 * be able to use to the FP for our optimized bzero/bcopy code before
225 * we enter more mainstream C code.
227 * WARNING! %fs is not set up on entry. This routine sets up %fs.
233 int x, myid = bootAP;
235 struct mdglobaldata *md;
236 struct privatespace *ps;
238 ps = &CPU_prvspace[myid];
240 gdt_segs[GPRIV_SEL].ssd_base = (int)ps;
241 gdt_segs[GPROC0_SEL].ssd_base =
242 (int) &ps->mdglobaldata.gd_common_tss;
243 ps->mdglobaldata.mi.gd_prvspace = ps;
245 for (x = 0; x < NGDT; x++) {
246 ssdtosd(&gdt_segs[x], &gdt[myid * NGDT + x].sd);
249 r_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
250 r_gdt.rd_base = (int) &gdt[myid * NGDT];
251 lgdt(&r_gdt); /* does magic intra-segment return */
256 mdcpu->gd_currentldt = _default_ldt;
258 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
259 gdt[myid * NGDT + GPROC0_SEL].sd.sd_type = SDT_SYS386TSS;
261 md = mdcpu; /* loaded through %fs:0 (mdglobaldata.mi.gd_prvspace)*/
263 md->gd_common_tss.tss_esp0 = 0; /* not used until after switch */
264 md->gd_common_tss.tss_ss0 = GSEL(GDATA_SEL, SEL_KPL);
265 md->gd_common_tss.tss_ioopt = (sizeof md->gd_common_tss) << 16;
266 md->gd_tss_gdt = &gdt[myid * NGDT + GPROC0_SEL].sd;
267 md->gd_common_tssd = *md->gd_tss_gdt;
271 * Set to a known state:
272 * Set by mpboot.s: CR0_PG, CR0_PE
273 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
276 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
278 pmap_set_opt(); /* PSE/4MB pages, etc */
280 pmap_init_pat(); /* Page Attribute Table */
282 /* set up CPU registers and state */
285 /* set up FPU state on the AP */
286 npxinit(__INITIAL_NPXCW__);
288 /* set up SSE registers */
292 /*******************************************************************
293 * local functions and data
297 * Start the SMP system
300 mp_start_aps(void *dummy __unused)
303 /* start each Application Processor */
304 start_all_aps(boot_address);
306 mp_bsp_simple_setup();
309 SYSINIT(startaps, SI_BOOT2_START_APS, SI_ORDER_FIRST, mp_start_aps, NULL)
312 * start each AP in our list
315 start_all_aps(u_int boot_addr)
323 u_long mpbioswarmvec;
324 struct mdglobaldata *gd;
325 struct privatespace *ps;
329 POSTCODE(START_ALL_APS_POST);
331 /* install the AP 1st level boot code */
332 install_ap_tramp(boot_addr);
335 /* save the current value of the warm-start vector */
336 mpbioswarmvec = *((u_long *) WARMBOOT_OFF);
337 outb(CMOS_REG, BIOS_RESET);
338 mpbiosreason = inb(CMOS_DATA);
340 /* setup a vector to our boot code */
341 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
342 *((volatile u_short *) WARMBOOT_SEG) = (boot_addr >> 4);
343 outb(CMOS_REG, BIOS_RESET);
344 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
347 * If we have a TSC we can figure out the SMI interrupt rate.
348 * The SMI does not necessarily use a constant rate. Spend
349 * up to 250ms trying to figure it out.
352 if (cpu_feature & CPUID_TSC) {
353 set_apic_timer(275000);
354 smilast = read_apic_timer();
355 for (x = 0; x < 20 && read_apic_timer(); ++x) {
356 smicount = smitest();
357 if (smibest == 0 || smilast - smicount < smibest)
358 smibest = smilast - smicount;
361 if (smibest > 250000)
364 smibest = smibest * (int64_t)1000000 /
365 get_apic_timer_frequency();
369 kprintf("SMI Frequency (worst case): %d Hz (%d us)\n",
370 1000000 / smibest, smibest);
373 /* set up temporary P==V mapping for AP boot */
374 /* XXX this is a hack, we should boot the AP on its own stack/PTD */
375 kptbase = (uintptr_t)(void *)KPTphys;
376 for (x = 0; x < NKPT; x++) {
377 PTD[x] = (pd_entry_t)(PG_V | PG_RW |
378 ((kptbase + x * PAGE_SIZE) & PG_FRAME));
383 for (x = 1; x <= naps; ++x) {
385 /* This is a bit verbose, it will go away soon. */
387 /* first page of AP's private space */
388 pg = x * i386_btop(sizeof(struct privatespace));
390 /* allocate new private data page(s) */
391 gd = (struct mdglobaldata *)kmem_alloc(&kernel_map,
392 MDGLOBALDATA_BASEALLOC_SIZE);
393 /* wire it into the private page table page */
394 for (i = 0; i < MDGLOBALDATA_BASEALLOC_SIZE; i += PAGE_SIZE) {
395 SMPpt[pg + i / PAGE_SIZE] = (pt_entry_t)
396 (PG_V | PG_RW | vtophys_pte((char *)gd + i));
398 pg += MDGLOBALDATA_BASEALLOC_PAGES;
400 SMPpt[pg + 0] = 0; /* *gd_CMAP1 */
401 SMPpt[pg + 1] = 0; /* *gd_CMAP2 */
402 SMPpt[pg + 2] = 0; /* *gd_CMAP3 */
403 SMPpt[pg + 3] = 0; /* *gd_PMAP1 */
405 /* allocate and set up an idle stack data page */
406 stack = (char *)kmem_alloc(&kernel_map, UPAGES*PAGE_SIZE);
407 for (i = 0; i < UPAGES; i++) {
408 SMPpt[pg + 4 + i] = (pt_entry_t)
409 (PG_V | PG_RW | vtophys_pte(PAGE_SIZE * i + stack));
412 gd = &CPU_prvspace[x].mdglobaldata; /* official location */
413 bzero(gd, sizeof(*gd));
414 gd->mi.gd_prvspace = ps = &CPU_prvspace[x];
416 /* prime data page for it to use */
417 mi_gdinit(&gd->mi, x);
419 gd->gd_CMAP1 = &SMPpt[pg + 0];
420 gd->gd_CMAP2 = &SMPpt[pg + 1];
421 gd->gd_CMAP3 = &SMPpt[pg + 2];
422 gd->gd_PMAP1 = &SMPpt[pg + 3];
423 gd->gd_CADDR1 = ps->CPAGE1;
424 gd->gd_CADDR2 = ps->CPAGE2;
425 gd->gd_CADDR3 = ps->CPAGE3;
426 gd->gd_PADDR1 = (unsigned *)ps->PPAGE1;
429 * Per-cpu pmap for get_ptbase().
431 gd->gd_GDADDR1= (unsigned *)
432 kmem_alloc_nofault(&kernel_map, SEG_SIZE, SEG_SIZE);
433 gd->gd_GDMAP1 = &PTD[(vm_offset_t)gd->gd_GDADDR1 >> PDRSHIFT];
435 gd->mi.gd_ipiq = (void *)kmem_alloc(&kernel_map, sizeof(lwkt_ipiq) * (naps + 1));
436 bzero(gd->mi.gd_ipiq, sizeof(lwkt_ipiq) * (naps + 1));
439 * Setup the AP boot stack
441 bootSTK = &ps->idlestack[UPAGES*PAGE_SIZE/2];
444 /* attempt to start the Application Processor */
445 CHECK_INIT(99); /* setup checkpoints */
446 if (!start_ap(gd, boot_addr, smibest)) {
447 kprintf("AP #%d (PHY# %d) failed!\n", x,
449 CHECK_PRINT("trace"); /* show checkpoints */
450 /* better panic as the AP may be running loose */
451 kprintf("panic y/n? [y] ");
455 CHECK_PRINT("trace"); /* show checkpoints */
458 /* set ncpus to 1 + highest logical cpu. Not all may have come up */
461 /* ncpus2 -- ncpus rounded down to the nearest power of 2 */
462 for (shift = 0; (1 << shift) <= ncpus; ++shift)
465 ncpus2_shift = shift;
467 ncpus2_mask = ncpus2 - 1;
469 /* ncpus_fit -- ncpus rounded up to the nearest power of 2 */
470 if ((1 << shift) < ncpus)
472 ncpus_fit = 1 << shift;
473 ncpus_fit_mask = ncpus_fit - 1;
475 /* build our map of 'other' CPUs */
476 mycpu->gd_other_cpus = smp_startup_mask & ~CPUMASK(mycpu->gd_cpuid);
477 mycpu->gd_ipiq = (void *)kmem_alloc(&kernel_map, sizeof(lwkt_ipiq) * ncpus);
478 bzero(mycpu->gd_ipiq, sizeof(lwkt_ipiq) * ncpus);
480 /* restore the warmstart vector */
481 *(u_long *) WARMBOOT_OFF = mpbioswarmvec;
482 outb(CMOS_REG, BIOS_RESET);
483 outb(CMOS_DATA, mpbiosreason);
486 * NOTE! The idlestack for the BSP was setup by locore. Finish
487 * up, clean out the P==V mapping we did earlier.
489 for (x = 0; x < NKPT; x++)
494 * Wait all APs to finish initializing LAPIC
498 kprintf("SMP: Waiting APs LAPIC initialization\n");
499 if (cpu_feature & CPUID_TSC)
500 tsc0_offset = rdtsc();
503 while (smp_lapic_mask != smp_startup_mask) {
505 if (cpu_feature & CPUID_TSC)
506 tsc0_offset = rdtsc();
508 while (try_mplock() == 0)
511 /* number of APs actually started */
516 * load the 1st level AP boot code into base memory.
519 /* targets for relocation */
520 extern void bigJump(void);
521 extern void bootCodeSeg(void);
522 extern void bootDataSeg(void);
523 extern void MPentry(void);
525 extern u_int mp_gdtbase;
528 install_ap_tramp(u_int boot_addr)
531 int size = *(int *) ((u_long) & bootMP_size);
532 u_char *src = (u_char *) ((u_long) bootMP);
533 u_char *dst = (u_char *) boot_addr + KERNBASE;
534 u_int boot_base = (u_int) bootMP;
539 POSTCODE(INSTALL_AP_TRAMP_POST);
541 for (x = 0; x < size; ++x)
545 * modify addresses in code we just moved to basemem. unfortunately we
546 * need fairly detailed info about mpboot.s for this to work. changes
547 * to mpboot.s might require changes here.
550 /* boot code is located in KERNEL space */
551 dst = (u_char *) boot_addr + KERNBASE;
553 /* modify the lgdt arg */
554 dst32 = (u_int32_t *) (dst + ((u_int) & mp_gdtbase - boot_base));
555 *dst32 = boot_addr + ((u_int) & MP_GDT - boot_base);
557 /* modify the ljmp target for MPentry() */
558 dst32 = (u_int32_t *) (dst + ((u_int) bigJump - boot_base) + 1);
559 *dst32 = ((u_int) MPentry - KERNBASE);
561 /* modify the target for boot code segment */
562 dst16 = (u_int16_t *) (dst + ((u_int) bootCodeSeg - boot_base));
563 dst8 = (u_int8_t *) (dst16 + 1);
564 *dst16 = boot_addr & 0xffff;
565 *dst8 = (boot_addr >> 16) & 0xff;
567 /* modify the target for boot data segment */
568 dst16 = (u_int16_t *) (dst + ((u_int) bootDataSeg - boot_base));
569 dst8 = (u_int8_t *) (dst16 + 1);
570 *dst16 = boot_addr & 0xffff;
571 *dst8 = (boot_addr >> 16) & 0xff;
576 * This function starts the AP (application processor) identified
577 * by the APIC ID 'physicalCpu'. It does quite a "song and dance"
578 * to accomplish this. This is necessary because of the nuances
579 * of the different hardware we might encounter. It ain't pretty,
580 * but it seems to work.
582 * NOTE: eventually an AP gets to ap_init(), which is called just
583 * before the AP goes into the LWKT scheduler's idle loop.
586 start_ap(struct mdglobaldata *gd, u_int boot_addr, int smibest)
590 u_long icr_lo, icr_hi;
592 POSTCODE(START_AP_POST);
594 /* get the PHYSICAL APIC ID# */
595 physical_cpu = CPUID_TO_APICID(gd->mi.gd_cpuid);
597 /* calculate the vector */
598 vector = (boot_addr >> 12) & 0xff;
600 /* We don't want anything interfering */
603 /* Make sure the target cpu sees everything */
607 * Try to detect when a SMI has occurred, wait up to 200ms.
609 * If a SMI occurs during an AP reset but before we issue
610 * the STARTUP command, the AP may brick. To work around
611 * this problem we hold off doing the AP startup until
612 * after we have detected the SMI. Hopefully another SMI
613 * will not occur before we finish the AP startup.
615 * Retries don't seem to help. SMIs have a window of opportunity
616 * and if USB->legacy keyboard emulation is enabled in the BIOS
617 * the interrupt rate can be quite high.
619 * NOTE: Don't worry about the L1 cache load, it might bloat
620 * ldelta a little but ndelta will be so huge when the SMI
621 * occurs the detection logic will still work fine.
624 set_apic_timer(200000);
629 * first we do an INIT/RESET IPI this INIT IPI might be run, reseting
630 * and running the target CPU. OR this INIT IPI might be latched (P5
631 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
634 * see apic/apicreg.h for icr bit definitions.
636 * TIME CRITICAL CODE, DO NOT DO ANY KPRINTFS IN THE HOT PATH.
640 * Setup the address for the target AP. We can setup
641 * icr_hi once and then just trigger operations with
644 icr_hi = lapic->icr_hi & ~APIC_ID_MASK;
645 icr_hi |= (physical_cpu << 24);
646 icr_lo = lapic->icr_lo & 0xfff00000;
647 lapic->icr_hi = icr_hi;
650 * Do an INIT IPI: assert RESET
652 * Use edge triggered mode to assert INIT
654 lapic->icr_lo = icr_lo | 0x0000c500;
655 while (lapic->icr_lo & APIC_DELSTAT_MASK)
659 * The spec calls for a 10ms delay but we may have to use a
660 * MUCH lower delay to avoid bricking an AP due to a fast SMI
661 * interrupt. We have other loops here too and dividing by 2
662 * doesn't seem to be enough even after subtracting 350us,
665 * Our minimum delay is 150uS, maximum is 10ms. If no SMI
666 * interrupt was detected we use the full 10ms.
670 else if (smibest < 150 * 4 + 350)
672 else if ((smibest - 350) / 4 < 10000)
673 u_sleep((smibest - 350) / 4);
678 * Do an INIT IPI: deassert RESET
680 * Use level triggered mode to deassert. It is unclear
681 * why we need to do this.
683 lapic->icr_lo = icr_lo | 0x00008500;
684 while (lapic->icr_lo & APIC_DELSTAT_MASK)
686 u_sleep(150); /* wait 150us */
689 * Next we do a STARTUP IPI: the previous INIT IPI might still be
690 * latched, (P5 bug) this 1st STARTUP would then terminate
691 * immediately, and the previously started INIT IPI would continue. OR
692 * the previous INIT IPI has already run. and this STARTUP IPI will
693 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
696 lapic->icr_lo = icr_lo | 0x00000600 | vector;
697 while (lapic->icr_lo & APIC_DELSTAT_MASK)
699 u_sleep(200); /* wait ~200uS */
702 * Finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
703 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
704 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
705 * recognized after hardware RESET or INIT IPI.
707 lapic->icr_lo = icr_lo | 0x00000600 | vector;
708 while (lapic->icr_lo & APIC_DELSTAT_MASK)
711 /* Resume normal operation */
714 /* wait for it to start, see ap_init() */
715 set_apic_timer(5000000);/* == 5 seconds */
716 while (read_apic_timer()) {
717 if (smp_startup_mask & CPUMASK(gd->mi.gd_cpuid))
718 return 1; /* return SUCCESS */
721 return 0; /* return FAILURE */
736 while (read_apic_timer()) {
738 for (count = 0; count < 100; ++count)
739 ntsc = rdtsc(); /* force loop to occur */
741 ndelta = ntsc - ltsc;
744 if (ndelta > ldelta * 2)
747 ldelta = ntsc - ltsc;
750 return(read_apic_timer());
754 * Lazy flush the TLB on all other CPU's. DEPRECATED.
756 * If for some reason we were unable to start all cpus we cannot safely
757 * use broadcast IPIs.
760 static cpumask_t smp_invltlb_req;
761 #define SMP_INVLTLB_DEBUG
766 struct mdglobaldata *md = mdcpu;
767 #ifdef SMP_INVLTLB_DEBUG
772 crit_enter_gd(&md->mi);
773 md->gd_invltlb_ret = 0;
774 ++md->mi.gd_cnt.v_smpinvltlb;
775 atomic_set_cpumask(&smp_invltlb_req, md->mi.gd_cpumask);
776 #ifdef SMP_INVLTLB_DEBUG
779 if (smp_startup_mask == smp_active_mask) {
780 all_but_self_ipi(XINVLTLB_OFFSET);
782 selected_apic_ipi(smp_active_mask & ~md->mi.gd_cpumask,
783 XINVLTLB_OFFSET, APIC_DELMODE_FIXED);
786 #ifdef SMP_INVLTLB_DEBUG
788 kprintf("smp_invltlb: ipi sent\n");
790 while ((md->gd_invltlb_ret & smp_active_mask & ~md->mi.gd_cpumask) !=
791 (smp_active_mask & ~md->mi.gd_cpumask)) {
794 #ifdef SMP_INVLTLB_DEBUG
796 if (++count == 400000000) {
798 kprintf("smp_invltlb: endless loop %08lx %08lx, "
799 "eflags %016lx retry",
800 (long)md->gd_invltlb_ret,
801 (long)smp_invltlb_req,
802 (long)read_eflags());
803 __asm __volatile ("sti");
808 int bcpu = BSFCPUMASK(~md->gd_invltlb_ret &
812 kprintf("bcpu %d\n", bcpu);
813 xgd = globaldata_find(bcpu);
814 kprintf("thread %p %s\n", xgd->gd_curthread, xgd->gd_curthread->td_comm);
823 atomic_clear_cpumask(&smp_invltlb_req, md->mi.gd_cpumask);
824 crit_exit_gd(&md->mi);
828 * Called from Xinvltlb assembly with interrupts disabled. We didn't
829 * bother to bump the critical section count or nested interrupt count
830 * so only do very low level operations here.
833 smp_invltlb_intr(void)
835 struct mdglobaldata *md = mdcpu;
836 struct mdglobaldata *omd;
840 mask = smp_invltlb_req;
844 cpu = BSFCPUMASK(mask);
845 mask &= ~CPUMASK(cpu);
846 omd = (struct mdglobaldata *)globaldata_find(cpu);
847 atomic_set_cpumask(&omd->gd_invltlb_ret, md->mi.gd_cpumask);
852 cpu_wbinvd_on_all_cpus_callback(void *arg)
858 * When called the executing CPU will send an IPI to all other CPUs
859 * requesting that they halt execution.
861 * Usually (but not necessarily) called with 'other_cpus' as its arg.
863 * - Signals all CPUs in map to stop.
864 * - Waits for each to stop.
871 * XXX FIXME: this is not MP-safe, needs a lock to prevent multiple CPUs
872 * from executing at same time.
875 stop_cpus(cpumask_t map)
877 map &= smp_active_mask;
879 /* send the Xcpustop IPI to all CPUs in map */
880 selected_apic_ipi(map, XCPUSTOP_OFFSET, APIC_DELMODE_FIXED);
882 while ((stopped_cpus & map) != map)
890 * Called by a CPU to restart stopped CPUs.
892 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
894 * - Signals all CPUs in map to restart.
895 * - Waits for each to restart.
903 restart_cpus(cpumask_t map)
905 /* signal other cpus to restart */
906 started_cpus = map & smp_active_mask;
908 while ((stopped_cpus & map) != 0) /* wait for each to clear its bit */
915 * This is called once the mpboot code has gotten us properly relocated
916 * and the MMU turned on, etc. ap_init() is actually the idle thread,
917 * and when it returns the scheduler will call the real cpu_idle() main
918 * loop for the idlethread. Interrupts are disabled on entry and should
919 * remain disabled at return.
927 * Adjust smp_startup_mask to signal the BSP that we have started
928 * up successfully. Note that we do not yet hold the BGL. The BSP
929 * is waiting for our signal.
931 * We can't set our bit in smp_active_mask yet because we are holding
932 * interrupts physically disabled and remote cpus could deadlock
933 * trying to send us an IPI.
935 smp_startup_mask |= CPUMASK(mycpu->gd_cpuid);
939 * Interlock for LAPIC initialization. Wait until mp_finish_lapic is
940 * non-zero, then get the MP lock.
942 * Note: We are in a critical section.
944 * Note: we are the idle thread, we can only spin.
946 * Note: The load fence is memory volatile and prevents the compiler
947 * from improperly caching mp_finish_lapic, and the cpu from improperly
950 while (mp_finish_lapic == 0)
952 while (try_mplock() == 0)
955 if (cpu_feature & CPUID_TSC) {
957 * The BSP is constantly updating tsc0_offset, figure out
958 * the relative difference to synchronize ktrdump.
960 tsc_offsets[mycpu->gd_cpuid] = rdtsc() - tsc0_offset;
963 /* BSP may have changed PTD while we're waiting for the lock */
966 #if defined(I586_CPU) && !defined(NO_F00F_HACK)
970 /* Build our map of 'other' CPUs. */
971 mycpu->gd_other_cpus = smp_startup_mask & ~CPUMASK(mycpu->gd_cpuid);
973 /* A quick check from sanity claus */
974 cpu_id = APICID_TO_CPUID((lapic->id & 0xff000000) >> 24);
975 if (mycpu->gd_cpuid != cpu_id) {
976 kprintf("SMP: assigned cpuid = %d\n", mycpu->gd_cpuid);
977 kprintf("SMP: actual cpuid = %d\n", cpu_id);
978 kprintf("PTD[MPPTDI] = %p\n", (void *)PTD[MPPTDI]);
979 panic("cpuid mismatch! boom!!");
982 /* Initialize AP's local APIC for irq's */
985 /* LAPIC initialization is done */
986 smp_lapic_mask |= CPUMASK(mycpu->gd_cpuid);
989 /* Let BSP move onto the next initialization stage */
993 * Interlock for finalization. Wait until mp_finish is non-zero,
994 * then get the MP lock.
996 * Note: We are in a critical section.
998 * Note: we are the idle thread, we can only spin.
1000 * Note: The load fence is memory volatile and prevents the compiler
1001 * from improperly caching mp_finish, and the cpu from improperly
1004 while (mp_finish == 0)
1006 while (try_mplock() == 0)
1009 /* BSP may have changed PTD while we're waiting for the lock */
1012 /* Set memory range attributes for this CPU to match the BSP */
1013 mem_range_AP_init();
1016 * Once we go active we must process any IPIQ messages that may
1017 * have been queued, because no actual IPI will occur until we
1018 * set our bit in the smp_active_mask. If we don't the IPI
1019 * message interlock could be left set which would also prevent
1022 * The idle loop doesn't expect the BGL to be held and while
1023 * lwkt_switch() normally cleans things up this is a special case
1024 * because we returning almost directly into the idle loop.
1026 * The idle thread is never placed on the runq, make sure
1027 * nothing we've done put it there.
1029 KKASSERT(get_mplock_count(curthread) == 1);
1030 smp_active_mask |= CPUMASK(mycpu->gd_cpuid);
1033 * Enable interrupts here. idle_restore will also do it, but
1034 * doing it here lets us clean up any strays that got posted to
1035 * the CPU during the AP boot while we are still in a critical
1038 __asm __volatile("sti; pause; pause"::);
1039 bzero(mdcpu->gd_ipending, sizeof(mdcpu->gd_ipending));
1041 initclocks_pcpu(); /* clock interrupts (via IPIs) */
1042 lwkt_process_ipiq();
1045 * Releasing the mp lock lets the BSP finish up the SMP init
1048 KKASSERT((curthread->td_flags & TDF_RUNQ) == 0);
1052 * Get SMP fully working before we start initializing devices.
1060 kprintf("Finish MP startup\n");
1062 while (smp_active_mask != smp_startup_mask)
1064 while (try_mplock() == 0)
1067 kprintf("Active CPU Mask: %08x\n", smp_active_mask);
1070 SYSINIT(finishsmp, SI_BOOT2_FINISH_SMP, SI_ORDER_FIRST, ap_finish, NULL)
1073 cpu_send_ipiq(int dcpu)
1075 if (CPUMASK(dcpu) & smp_active_mask)
1076 single_apic_ipi(dcpu, XIPIQ_OFFSET, APIC_DELMODE_FIXED);
1079 #if 0 /* single_apic_ipi_passive() not working yet */
1081 * Returns 0 on failure, 1 on success
1084 cpu_send_ipiq_passive(int dcpu)
1087 if (CPUMASK(dcpu) & smp_active_mask) {
1088 r = single_apic_ipi_passive(dcpu, XIPIQ_OFFSET,
1089 APIC_DELMODE_FIXED);
1096 mp_bsp_simple_setup(void)
1098 /* build our map of 'other' CPUs */
1099 mycpu->gd_other_cpus = smp_startup_mask & ~CPUMASK(mycpu->gd_cpuid);
1100 mycpu->gd_ipiq = (void *)kmem_alloc(&kernel_map, sizeof(lwkt_ipiq) * ncpus);
1101 bzero(mycpu->gd_ipiq, sizeof(lwkt_ipiq) * ncpus);
1105 if (cpu_feature & CPUID_TSC)
1106 tsc0_offset = rdtsc();
1111 * CPU TOPOLOGY DETECTION FUNCTIONS
1114 /* Detect intel topology using CPUID
1115 * Ref: http://www.intel.com/Assets/PDF/appnote/241618.pdf, pg 41
1118 detect_intel_topology(int count_htt_cores)
1122 int core_plus_logical_bits = 0;
1123 int cores_per_package;
1124 int logical_per_package;
1125 int logical_per_core;
1128 if (cpu_high >= 0xb) {
1131 } else if (cpu_high >= 0x4) {
1136 for (shift = 0; (1 << shift) < count_htt_cores; ++shift)
1138 logical_CPU_bits = 1 << shift;
1143 cpuid_count(0xb, FUNC_B_THREAD_LEVEL, p);
1145 /* if 0xb not supported - fallback to 0x4 */
1146 if (p[1] == 0 || (FUNC_B_TYPE(p[2]) != FUNC_B_THREAD_TYPE)) {
1150 logical_CPU_bits = FUNC_B_BITS_SHIFT_NEXT_LEVEL(p[0]);
1152 ecx_index = FUNC_B_THREAD_LEVEL + 1;
1154 cpuid_count(0xb, ecx_index, p);
1155 /* Check for the Core type in the implemented sub leaves. */
1156 if (FUNC_B_TYPE(p[2]) == FUNC_B_CORE_TYPE) {
1157 core_plus_logical_bits = FUNC_B_BITS_SHIFT_NEXT_LEVEL(p[0]);
1161 } while (FUNC_B_TYPE(p[2]) != FUNC_B_INVALID_TYPE);
1163 core_bits = core_plus_logical_bits - logical_CPU_bits;
1168 cpuid_count(0x4, 0, p);
1169 cores_per_package = FUNC_4_MAX_CORE_NO(p[0]) + 1;
1171 logical_per_package = count_htt_cores;
1172 logical_per_core = logical_per_package / cores_per_package;
1174 for (shift = 0; (1 << shift) < logical_per_core; ++shift)
1176 logical_CPU_bits = shift;
1178 for (shift = 0; (1 << shift) < cores_per_package; ++shift)
1185 /* Detect AMD topology using CPUID
1186 * Ref: http://support.amd.com/us/Embedded_TechDocs/25481.pdf, last page
1189 detect_amd_topology(int count_htt_cores)
1193 if ((cpu_feature & CPUID_HTT)
1194 && (amd_feature2 & AMDID2_CMP)) {
1196 if (cpu_procinfo2 & AMDID_COREID_SIZE) {
1197 core_bits = (cpu_procinfo2 & AMDID_COREID_SIZE)
1198 >> AMDID_COREID_SIZE_SHIFT;
1200 core_bits = (cpu_procinfo2 & AMDID_CMP_CORES) + 1;
1201 for (shift = 0; (1 << shift) < core_bits; ++shift);
1205 logical_CPU_bits = count_htt_cores >> core_bits;
1206 for (shift = 0; (1 << shift) < logical_CPU_bits; ++shift)
1208 logical_CPU_bits = shift;
1210 for (shift = 0; (1 << shift) < count_htt_cores; ++shift)
1213 logical_CPU_bits = 0;
1218 * - logical_CPU_bits
1220 * With the values above (for AMD or INTEL) we are able to generally
1221 * detect the CPU topology (number of cores for each level):
1222 * Ref: http://wiki.osdev.org/Detecting_CPU_Topology_(80x86)
1223 * Ref: http://www.multicoreinfo.com/research/papers/whitepapers/Intel-detect-topology.pdf
1226 detect_cpu_topology(void)
1228 static int topology_detected = 0;
1231 if (topology_detected) {
1235 if ((cpu_feature & CPUID_HTT) == 0) {
1237 logical_CPU_bits = 0;
1240 count = (cpu_procinfo & CPUID_HTT_CORES)
1241 >> CPUID_HTT_CORE_SHIFT;
1244 if (cpu_vendor_id == CPU_VENDOR_INTEL) {
1245 detect_intel_topology(count);
1246 } else if (cpu_vendor_id == CPU_VENDOR_AMD) {
1247 detect_amd_topology(count);
1252 kprintf("Bits within APICID: logical_CPU_bits: %d; core_bits: %d\n",
1253 logical_CPU_bits, core_bits);
1255 topology_detected = 1;
1258 /* Interface functions to calculate chip_ID,
1259 * core_number and logical_number
1260 * Ref: http://wiki.osdev.org/Detecting_CPU_Topology_(80x86)
1263 get_chip_ID(int cpuid)
1265 return get_apicid_from_cpuid(cpuid) >>
1266 (logical_CPU_bits + core_bits);
1270 get_core_number_within_chip(int cpuid)
1272 return (get_apicid_from_cpuid(cpuid) >> logical_CPU_bits) &
1273 ( (1 << core_bits) -1);
1277 get_logical_CPU_number_within_core(int cpuid)
1279 return get_apicid_from_cpuid(cpuid) &
1280 ( (1 << logical_CPU_bits) -1);