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 /* which cpus have been started */
177 static cpumask_t smp_startup_mask = CPUMASK_INITIALIZER_ONLYONE;
178 /* which cpus have lapic been inited */
179 static cpumask_t smp_lapic_mask = CPUMASK_INITIALIZER_ONLYONE;
180 /* which cpus are ready for IPIs etc? */
181 cpumask_t smp_active_mask = CPUMASK_INITIALIZER_ONLYONE;
182 SYSCTL_LONG(_machdep, OID_AUTO, smp_active, CTLFLAG_RD,
183 &smp_active_mask, 0, "");
185 /* Local data for detecting CPU TOPOLOGY */
186 static int core_bits = 0;
187 static int logical_CPU_bits = 0;
191 * Calculate usable address in base memory for AP trampoline code.
194 mp_bootaddress(u_int basemem)
196 POSTCODE(MP_BOOTADDRESS_POST);
198 boot_address = basemem & ~0xfff; /* round down to 4k boundary */
199 if ((basemem - boot_address) < bootMP_size)
200 boot_address -= 4096; /* not enough, lower by 4k */
206 * Print various information about the SMP system hardware and setup.
213 POSTCODE(MP_ANNOUNCE_POST);
215 kprintf("DragonFly/MP: Multiprocessor motherboard\n");
216 kprintf(" cpu0 (BSP): apic id: %2d\n", CPUID_TO_APICID(0));
217 for (x = 1; x <= naps; ++x)
218 kprintf(" cpu%d (AP): apic id: %2d\n", x, CPUID_TO_APICID(x));
221 kprintf(" Warning: APIC I/O disabled\n");
225 * AP cpu's call this to sync up protected mode.
227 * WARNING! We must ensure that the cpu is sufficiently initialized to
228 * be able to use to the FP for our optimized bzero/bcopy code before
229 * we enter more mainstream C code.
231 * WARNING! %fs is not set up on entry. This routine sets up %fs.
237 int x, myid = bootAP;
239 struct mdglobaldata *md;
240 struct privatespace *ps;
242 ps = &CPU_prvspace[myid];
244 gdt_segs[GPRIV_SEL].ssd_base = (int)ps;
245 gdt_segs[GPROC0_SEL].ssd_base =
246 (int) &ps->mdglobaldata.gd_common_tss;
247 ps->mdglobaldata.mi.gd_prvspace = ps;
249 for (x = 0; x < NGDT; x++) {
250 ssdtosd(&gdt_segs[x], &gdt[myid * NGDT + x].sd);
253 r_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
254 r_gdt.rd_base = (int) &gdt[myid * NGDT];
255 lgdt(&r_gdt); /* does magic intra-segment return */
260 mdcpu->gd_currentldt = _default_ldt;
262 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
263 gdt[myid * NGDT + GPROC0_SEL].sd.sd_type = SDT_SYS386TSS;
265 md = mdcpu; /* loaded through %fs:0 (mdglobaldata.mi.gd_prvspace)*/
267 md->gd_common_tss.tss_esp0 = 0; /* not used until after switch */
268 md->gd_common_tss.tss_ss0 = GSEL(GDATA_SEL, SEL_KPL);
269 md->gd_common_tss.tss_ioopt = (sizeof md->gd_common_tss) << 16;
270 md->gd_tss_gdt = &gdt[myid * NGDT + GPROC0_SEL].sd;
271 md->gd_common_tssd = *md->gd_tss_gdt;
275 * Set to a known state:
276 * Set by mpboot.s: CR0_PG, CR0_PE
277 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
280 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
282 pmap_set_opt(); /* PSE/4MB pages, etc */
284 pmap_init_pat(); /* Page Attribute Table */
286 /* set up CPU registers and state */
289 /* set up FPU state on the AP */
292 /* set up SSE registers */
296 /*******************************************************************
297 * local functions and data
301 * Start the SMP system
304 mp_start_aps(void *dummy __unused)
307 /* start each Application Processor */
308 start_all_aps(boot_address);
310 mp_bsp_simple_setup();
313 SYSINIT(startaps, SI_BOOT2_START_APS, SI_ORDER_FIRST, mp_start_aps, NULL)
316 * start each AP in our list
319 start_all_aps(u_int boot_addr)
327 u_long mpbioswarmvec;
328 struct mdglobaldata *gd;
329 struct privatespace *ps;
333 POSTCODE(START_ALL_APS_POST);
335 /* install the AP 1st level boot code */
336 install_ap_tramp(boot_addr);
339 /* save the current value of the warm-start vector */
340 mpbioswarmvec = *((u_long *) WARMBOOT_OFF);
341 outb(CMOS_REG, BIOS_RESET);
342 mpbiosreason = inb(CMOS_DATA);
344 /* setup a vector to our boot code */
345 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
346 *((volatile u_short *) WARMBOOT_SEG) = (boot_addr >> 4);
347 outb(CMOS_REG, BIOS_RESET);
348 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
351 * If we have a TSC we can figure out the SMI interrupt rate.
352 * The SMI does not necessarily use a constant rate. Spend
353 * up to 250ms trying to figure it out.
356 if (cpu_feature & CPUID_TSC) {
357 set_apic_timer(275000);
358 smilast = read_apic_timer();
359 for (x = 0; x < 20 && read_apic_timer(); ++x) {
360 smicount = smitest();
361 if (smibest == 0 || smilast - smicount < smibest)
362 smibest = smilast - smicount;
365 if (smibest > 250000)
368 smibest = smibest * (int64_t)1000000 /
369 get_apic_timer_frequency();
373 kprintf("SMI Frequency (worst case): %d Hz (%d us)\n",
374 1000000 / smibest, smibest);
377 /* set up temporary P==V mapping for AP boot */
378 /* XXX this is a hack, we should boot the AP on its own stack/PTD */
379 kptbase = (uintptr_t)(void *)KPTphys;
380 for (x = 0; x < NKPT; x++) {
381 PTD[x] = (pd_entry_t)(PG_V | PG_RW |
382 ((kptbase + x * PAGE_SIZE) & PG_FRAME));
387 for (x = 1; x <= naps; ++x) {
389 /* This is a bit verbose, it will go away soon. */
391 /* first page of AP's private space */
392 pg = x * i386_btop(sizeof(struct privatespace));
394 /* allocate new private data page(s) */
395 gd = (struct mdglobaldata *)kmem_alloc(&kernel_map,
396 MDGLOBALDATA_BASEALLOC_SIZE);
397 /* wire it into the private page table page */
398 for (i = 0; i < MDGLOBALDATA_BASEALLOC_SIZE; i += PAGE_SIZE) {
399 SMPpt[pg + i / PAGE_SIZE] = (pt_entry_t)
400 (PG_V | PG_RW | vtophys_pte((char *)gd + i));
402 pg += MDGLOBALDATA_BASEALLOC_PAGES;
404 SMPpt[pg + 0] = 0; /* *gd_CMAP1 */
405 SMPpt[pg + 1] = 0; /* *gd_CMAP2 */
406 SMPpt[pg + 2] = 0; /* *gd_CMAP3 */
407 SMPpt[pg + 3] = 0; /* *gd_PMAP1 */
409 /* allocate and set up an idle stack data page */
410 stack = (char *)kmem_alloc(&kernel_map, UPAGES*PAGE_SIZE);
411 for (i = 0; i < UPAGES; i++) {
412 SMPpt[pg + 4 + i] = (pt_entry_t)
413 (PG_V | PG_RW | vtophys_pte(PAGE_SIZE * i + stack));
416 gd = &CPU_prvspace[x].mdglobaldata; /* official location */
417 bzero(gd, sizeof(*gd));
418 gd->mi.gd_prvspace = ps = &CPU_prvspace[x];
420 /* prime data page for it to use */
421 mi_gdinit(&gd->mi, x);
423 gd->gd_CMAP1 = &SMPpt[pg + 0];
424 gd->gd_CMAP2 = &SMPpt[pg + 1];
425 gd->gd_CMAP3 = &SMPpt[pg + 2];
426 gd->gd_PMAP1 = &SMPpt[pg + 3];
427 gd->gd_CADDR1 = ps->CPAGE1;
428 gd->gd_CADDR2 = ps->CPAGE2;
429 gd->gd_CADDR3 = ps->CPAGE3;
430 gd->gd_PADDR1 = (unsigned *)ps->PPAGE1;
433 * Per-cpu pmap for get_ptbase().
435 gd->gd_GDADDR1= (unsigned *)
436 kmem_alloc_nofault(&kernel_map, SEG_SIZE, SEG_SIZE);
437 gd->gd_GDMAP1 = &PTD[(vm_offset_t)gd->gd_GDADDR1 >> PDRSHIFT];
439 gd->mi.gd_ipiq = (void *)kmem_alloc(&kernel_map, sizeof(lwkt_ipiq) * (naps + 1));
440 bzero(gd->mi.gd_ipiq, sizeof(lwkt_ipiq) * (naps + 1));
443 * Setup the AP boot stack
445 bootSTK = &ps->idlestack[UPAGES*PAGE_SIZE/2];
448 /* attempt to start the Application Processor */
449 CHECK_INIT(99); /* setup checkpoints */
450 if (!start_ap(gd, boot_addr, smibest)) {
451 kprintf("AP #%d (PHY# %d) failed!\n", x,
453 CHECK_PRINT("trace"); /* show checkpoints */
454 /* better panic as the AP may be running loose */
455 kprintf("panic y/n? [y] ");
459 CHECK_PRINT("trace"); /* show checkpoints */
462 /* set ncpus to 1 + highest logical cpu. Not all may have come up */
465 /* ncpus2 -- ncpus rounded down to the nearest power of 2 */
466 for (shift = 0; (1 << shift) <= ncpus; ++shift)
469 ncpus2_shift = shift;
471 ncpus2_mask = ncpus2 - 1;
473 /* ncpus_fit -- ncpus rounded up to the nearest power of 2 */
474 if ((1 << shift) < ncpus)
476 ncpus_fit = 1 << shift;
477 ncpus_fit_mask = ncpus_fit - 1;
479 /* build our map of 'other' CPUs */
480 mycpu->gd_other_cpus = smp_startup_mask;
481 CPUMASK_NANDBIT(mycpu->gd_other_cpus, mycpu->gd_cpuid);
482 mycpu->gd_ipiq = (void *)kmem_alloc(&kernel_map,
483 sizeof(lwkt_ipiq) * ncpus);
484 bzero(mycpu->gd_ipiq, sizeof(lwkt_ipiq) * ncpus);
486 /* restore the warmstart vector */
487 *(u_long *) WARMBOOT_OFF = mpbioswarmvec;
488 outb(CMOS_REG, BIOS_RESET);
489 outb(CMOS_DATA, mpbiosreason);
492 * NOTE! The idlestack for the BSP was setup by locore. Finish
493 * up, clean out the P==V mapping we did earlier.
495 for (x = 0; x < NKPT; x++)
500 * Wait all APs to finish initializing LAPIC
504 kprintf("SMP: Waiting APs LAPIC initialization\n");
505 if (cpu_feature & CPUID_TSC)
506 tsc0_offset = rdtsc();
509 while (CPUMASK_CMPMASKNEQ(smp_lapic_mask, smp_startup_mask)) {
511 if (cpu_feature & CPUID_TSC)
512 tsc0_offset = rdtsc();
514 while (try_mplock() == 0)
517 /* number of APs actually started */
522 * load the 1st level AP boot code into base memory.
525 /* targets for relocation */
526 extern void bigJump(void);
527 extern void bootCodeSeg(void);
528 extern void bootDataSeg(void);
529 extern void MPentry(void);
531 extern u_int mp_gdtbase;
534 install_ap_tramp(u_int boot_addr)
537 int size = *(int *) ((u_long) & bootMP_size);
538 u_char *src = (u_char *) ((u_long) bootMP);
539 u_char *dst = (u_char *) boot_addr + KERNBASE;
540 u_int boot_base = (u_int) bootMP;
545 POSTCODE(INSTALL_AP_TRAMP_POST);
547 for (x = 0; x < size; ++x)
551 * modify addresses in code we just moved to basemem. unfortunately we
552 * need fairly detailed info about mpboot.s for this to work. changes
553 * to mpboot.s might require changes here.
556 /* boot code is located in KERNEL space */
557 dst = (u_char *) boot_addr + KERNBASE;
559 /* modify the lgdt arg */
560 dst32 = (u_int32_t *) (dst + ((u_int) & mp_gdtbase - boot_base));
561 *dst32 = boot_addr + ((u_int) & MP_GDT - boot_base);
563 /* modify the ljmp target for MPentry() */
564 dst32 = (u_int32_t *) (dst + ((u_int) bigJump - boot_base) + 1);
565 *dst32 = ((u_int) MPentry - KERNBASE);
567 /* modify the target for boot code segment */
568 dst16 = (u_int16_t *) (dst + ((u_int) bootCodeSeg - boot_base));
569 dst8 = (u_int8_t *) (dst16 + 1);
570 *dst16 = boot_addr & 0xffff;
571 *dst8 = (boot_addr >> 16) & 0xff;
573 /* modify the target for boot data segment */
574 dst16 = (u_int16_t *) (dst + ((u_int) bootDataSeg - boot_base));
575 dst8 = (u_int8_t *) (dst16 + 1);
576 *dst16 = boot_addr & 0xffff;
577 *dst8 = (boot_addr >> 16) & 0xff;
582 * This function starts the AP (application processor) identified
583 * by the APIC ID 'physicalCpu'. It does quite a "song and dance"
584 * to accomplish this. This is necessary because of the nuances
585 * of the different hardware we might encounter. It ain't pretty,
586 * but it seems to work.
588 * NOTE: eventually an AP gets to ap_init(), which is called just
589 * before the AP goes into the LWKT scheduler's idle loop.
592 start_ap(struct mdglobaldata *gd, u_int boot_addr, int smibest)
596 u_long icr_lo, icr_hi;
598 POSTCODE(START_AP_POST);
600 /* get the PHYSICAL APIC ID# */
601 physical_cpu = CPUID_TO_APICID(gd->mi.gd_cpuid);
603 /* calculate the vector */
604 vector = (boot_addr >> 12) & 0xff;
606 /* We don't want anything interfering */
609 /* Make sure the target cpu sees everything */
613 * Try to detect when a SMI has occurred, wait up to 200ms.
615 * If a SMI occurs during an AP reset but before we issue
616 * the STARTUP command, the AP may brick. To work around
617 * this problem we hold off doing the AP startup until
618 * after we have detected the SMI. Hopefully another SMI
619 * will not occur before we finish the AP startup.
621 * Retries don't seem to help. SMIs have a window of opportunity
622 * and if USB->legacy keyboard emulation is enabled in the BIOS
623 * the interrupt rate can be quite high.
625 * NOTE: Don't worry about the L1 cache load, it might bloat
626 * ldelta a little but ndelta will be so huge when the SMI
627 * occurs the detection logic will still work fine.
630 set_apic_timer(200000);
635 * first we do an INIT/RESET IPI this INIT IPI might be run, reseting
636 * and running the target CPU. OR this INIT IPI might be latched (P5
637 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
640 * see apic/apicreg.h for icr bit definitions.
642 * TIME CRITICAL CODE, DO NOT DO ANY KPRINTFS IN THE HOT PATH.
646 * Setup the address for the target AP. We can setup
647 * icr_hi once and then just trigger operations with
650 icr_hi = lapic->icr_hi & ~APIC_ID_MASK;
651 icr_hi |= (physical_cpu << 24);
652 icr_lo = lapic->icr_lo & 0xfff00000;
653 lapic->icr_hi = icr_hi;
656 * Do an INIT IPI: assert RESET
658 * Use edge triggered mode to assert INIT
660 lapic->icr_lo = icr_lo | 0x0000c500;
661 while (lapic->icr_lo & APIC_DELSTAT_MASK)
665 * The spec calls for a 10ms delay but we may have to use a
666 * MUCH lower delay to avoid bricking an AP due to a fast SMI
667 * interrupt. We have other loops here too and dividing by 2
668 * doesn't seem to be enough even after subtracting 350us,
671 * Our minimum delay is 150uS, maximum is 10ms. If no SMI
672 * interrupt was detected we use the full 10ms.
676 else if (smibest < 150 * 4 + 350)
678 else if ((smibest - 350) / 4 < 10000)
679 u_sleep((smibest - 350) / 4);
684 * Do an INIT IPI: deassert RESET
686 * Use level triggered mode to deassert. It is unclear
687 * why we need to do this.
689 lapic->icr_lo = icr_lo | 0x00008500;
690 while (lapic->icr_lo & APIC_DELSTAT_MASK)
692 u_sleep(150); /* wait 150us */
695 * Next we do a STARTUP IPI: the previous INIT IPI might still be
696 * latched, (P5 bug) this 1st STARTUP would then terminate
697 * immediately, and the previously started INIT IPI would continue. OR
698 * the previous INIT IPI has already run. and this STARTUP IPI will
699 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
702 lapic->icr_lo = icr_lo | 0x00000600 | vector;
703 while (lapic->icr_lo & APIC_DELSTAT_MASK)
705 u_sleep(200); /* wait ~200uS */
708 * Finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
709 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
710 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
711 * recognized after hardware RESET or INIT IPI.
713 lapic->icr_lo = icr_lo | 0x00000600 | vector;
714 while (lapic->icr_lo & APIC_DELSTAT_MASK)
717 /* Resume normal operation */
720 /* wait for it to start, see ap_init() */
721 set_apic_timer(5000000);/* == 5 seconds */
722 while (read_apic_timer()) {
723 if (CPUMASK_TESTBIT(smp_startup_mask, gd->mi.gd_cpuid))
724 return 1; /* return SUCCESS */
727 return 0; /* return FAILURE */
742 while (read_apic_timer()) {
744 for (count = 0; count < 100; ++count)
745 ntsc = rdtsc(); /* force loop to occur */
747 ndelta = ntsc - ltsc;
750 if (ndelta > ldelta * 2)
753 ldelta = ntsc - ltsc;
756 return(read_apic_timer());
760 * Lazy flush the TLB on all other CPU's. DEPRECATED.
762 * If for some reason we were unable to start all cpus we cannot safely
763 * use broadcast IPIs.
766 static cpumask_t smp_invltlb_req;
767 #define SMP_INVLTLB_DEBUG
772 struct mdglobaldata *md = mdcpu;
773 #ifdef SMP_INVLTLB_DEBUG
780 crit_enter_gd(&md->mi);
781 md->gd_invltlb_ret = 0;
782 ++md->mi.gd_cnt.v_smpinvltlb;
783 ATOMIC_CPUMASK_ORMASK(smp_invltlb_req, md->mi.gd_cpumask);
784 #ifdef SMP_INVLTLB_DEBUG
787 if (CPUMASK_CMPMASKEQ(smp_startup_mask, smp_active_mask)) {
788 all_but_self_ipi(XINVLTLB_OFFSET);
790 tmpmask = smp_active_mask;
791 CPUMASK_NANDMASK(tmpmask, md->mi.gd_cpumask);
792 selected_apic_ipi(tmpmask, XINVLTLB_OFFSET, APIC_DELMODE_FIXED);
795 #ifdef SMP_INVLTLB_DEBUG
797 kprintf("smp_invltlb: ipi sent\n");
800 tmpmask = smp_active_mask;
802 CPUMASK_ANDMASK(tmpmask, md->gd_invltlb_ret);
803 CPUMASK_NANDMASK(tmpmask, md->mi.gd_cpumask);
804 CPUMASK_NANDMASK(tmpmask2, md->mi.gd_cpumask);
805 if (CPUMASK_CMPMASKEQ(tmpmask, tmpmask2))
809 #ifdef SMP_INVLTLB_DEBUG
811 if (++count == 400000000) {
813 kprintf("smp_invltlb: endless loop %08lx %08lx, "
814 "eflags %016lx retry",
815 (long)md->gd_invltlb_ret,
816 (long)smp_invltlb_req,
817 (long)read_eflags());
818 __asm __volatile ("sti");
826 tmpmask = smp_active_mask;
827 CPUMASK_NANDMASK(tmpmask, md->gd_invltlb_ret);
828 CPUMASK_NANDMASK(tmpmask, md->mi.gd_cpumask);
829 bcpu = BSFCPUMASK(tmpmask);
831 kprintf("bcpu %d\n", bcpu);
832 xgd = globaldata_find(bcpu);
833 kprintf("thread %p %s\n",
835 xgd->gd_curthread->td_comm);
844 ATOMIC_CPUMASK_NANDMASK(smp_invltlb_req, md->mi.gd_cpumask);
845 crit_exit_gd(&md->mi);
849 * Called from Xinvltlb assembly with interrupts disabled. We didn't
850 * bother to bump the critical section count or nested interrupt count
851 * so only do very low level operations here.
854 smp_invltlb_intr(void)
856 struct mdglobaldata *md = mdcpu;
857 struct mdglobaldata *omd;
861 mask = smp_invltlb_req;
864 while (CPUMASK_TESTNZERO(mask)) {
865 cpu = BSFCPUMASK(mask);
866 CPUMASK_NANDBIT(mask, cpu);
867 omd = (struct mdglobaldata *)globaldata_find(cpu);
868 ATOMIC_CPUMASK_ORMASK(omd->gd_invltlb_ret, md->mi.gd_cpumask);
873 cpu_wbinvd_on_all_cpus_callback(void *arg)
879 * When called the executing CPU will send an IPI to all other CPUs
880 * requesting that they halt execution.
882 * Usually (but not necessarily) called with 'other_cpus' as its arg.
884 * - Signals all CPUs in map to stop.
885 * - Waits for each to stop.
892 * XXX FIXME: this is not MP-safe, needs a lock to prevent multiple CPUs
893 * from executing at same time.
896 stop_cpus(cpumask_t map)
900 map &= smp_active_mask;
902 /* send the Xcpustop IPI to all CPUs in map */
903 selected_apic_ipi(map, XCPUSTOP_OFFSET, APIC_DELMODE_FIXED);
907 CPUMASK_ANDMASK(tmpmask, stopped_cpus);
909 } while (CPUMASK_CMPMASKNEQ(tmpmask, map));
916 * Called by a CPU to restart stopped CPUs.
918 * Usually (but not necessarily) called with 'stopped_cpus' as its arg.
920 * - Signals all CPUs in map to restart.
921 * - Waits for each to restart.
929 restart_cpus(cpumask_t map)
931 /* signal other cpus to restart */
932 started_cpus = map & smp_active_mask;
934 /* wait for each to clear its bit */
935 while (CPUMASK_TESTMASK(stopped_cpus, map) != 0)
942 * This is called once the mpboot code has gotten us properly relocated
943 * and the MMU turned on, etc. ap_init() is actually the idle thread,
944 * and when it returns the scheduler will call the real cpu_idle() main
945 * loop for the idlethread. Interrupts are disabled on entry and should
946 * remain disabled at return.
954 * Adjust smp_startup_mask to signal the BSP that we have started
955 * up successfully. Note that we do not yet hold the BGL. The BSP
956 * is waiting for our signal.
958 * We can't set our bit in smp_active_mask yet because we are holding
959 * interrupts physically disabled and remote cpus could deadlock
960 * trying to send us an IPI.
962 ATOMIC_CPUMASK_ORBIT(smp_startup_mask, mycpu->gd_cpuid);
966 * Interlock for LAPIC initialization. Wait until mp_finish_lapic is
967 * non-zero, then get the MP lock.
969 * Note: We are in a critical section.
971 * Note: we are the idle thread, we can only spin.
973 * Note: The load fence is memory volatile and prevents the compiler
974 * from improperly caching mp_finish_lapic, and the cpu from improperly
977 while (mp_finish_lapic == 0)
979 while (try_mplock() == 0)
982 if (cpu_feature & CPUID_TSC) {
984 * The BSP is constantly updating tsc0_offset, figure out
985 * the relative difference to synchronize ktrdump.
987 tsc_offsets[mycpu->gd_cpuid] = rdtsc() - tsc0_offset;
990 /* BSP may have changed PTD while we're waiting for the lock */
993 #if defined(I586_CPU) && !defined(NO_F00F_HACK)
997 /* Build our map of 'other' CPUs. */
998 mycpu->gd_other_cpus = smp_startup_mask;
999 CPUMASK_NANDBIT(mycpu->gd_other_cpus, mycpu->gd_cpuid);
1001 /* A quick check from sanity claus */
1002 cpu_id = APICID_TO_CPUID((lapic->id & 0xff000000) >> 24);
1003 if (mycpu->gd_cpuid != cpu_id) {
1004 kprintf("SMP: assigned cpuid = %d\n", mycpu->gd_cpuid);
1005 kprintf("SMP: actual cpuid = %d\n", cpu_id);
1006 kprintf("PTD[MPPTDI] = %p\n", (void *)PTD[MPPTDI]);
1007 panic("cpuid mismatch! boom!!");
1010 /* Initialize AP's local APIC for irq's */
1013 /* LAPIC initialization is done */
1014 ATOMIC_CPUMASK_ORBIT(smp_lapic_mask, mycpu->gd_cpuid);
1017 /* Let BSP move onto the next initialization stage */
1021 * Interlock for finalization. Wait until mp_finish is non-zero,
1022 * then get the MP lock.
1024 * Note: We are in a critical section.
1026 * Note: we are the idle thread, we can only spin.
1028 * Note: The load fence is memory volatile and prevents the compiler
1029 * from improperly caching mp_finish, and the cpu from improperly
1032 while (mp_finish == 0)
1034 while (try_mplock() == 0)
1037 /* BSP may have changed PTD while we're waiting for the lock */
1040 /* Set memory range attributes for this CPU to match the BSP */
1041 mem_range_AP_init();
1044 * Once we go active we must process any IPIQ messages that may
1045 * have been queued, because no actual IPI will occur until we
1046 * set our bit in the smp_active_mask. If we don't the IPI
1047 * message interlock could be left set which would also prevent
1050 * The idle loop doesn't expect the BGL to be held and while
1051 * lwkt_switch() normally cleans things up this is a special case
1052 * because we returning almost directly into the idle loop.
1054 * The idle thread is never placed on the runq, make sure
1055 * nothing we've done put it there.
1057 KKASSERT(get_mplock_count(curthread) == 1);
1058 ATOMIC_CPUMASK_ORBIT(smp_active_mask, mycpu->gd_cpuid);
1061 * Enable interrupts here. idle_restore will also do it, but
1062 * doing it here lets us clean up any strays that got posted to
1063 * the CPU during the AP boot while we are still in a critical
1066 __asm __volatile("sti; pause; pause"::);
1067 bzero(mdcpu->gd_ipending, sizeof(mdcpu->gd_ipending));
1069 initclocks_pcpu(); /* clock interrupts (via IPIs) */
1070 lwkt_process_ipiq();
1073 * Releasing the mp lock lets the BSP finish up the SMP init
1076 KKASSERT((curthread->td_flags & TDF_RUNQ) == 0);
1080 * Get SMP fully working before we start initializing devices.
1088 kprintf("Finish MP startup\n");
1090 while (smp_active_mask != smp_startup_mask)
1092 while (try_mplock() == 0)
1095 kprintf("Active CPU Mask: %08x\n", smp_active_mask);
1098 SYSINIT(finishsmp, SI_BOOT2_FINISH_SMP, SI_ORDER_FIRST, ap_finish, NULL)
1101 cpu_send_ipiq(int dcpu)
1103 if (CPUMASK_TESTBIT(smp_active_mask, dcpu))
1104 single_apic_ipi(dcpu, XIPIQ_OFFSET, APIC_DELMODE_FIXED);
1107 #if 0 /* single_apic_ipi_passive() not working yet */
1109 * Returns 0 on failure, 1 on success
1112 cpu_send_ipiq_passive(int dcpu)
1115 if (CPUMASK_TESTBIT(smp_active_mask, dcpu)) {
1116 r = single_apic_ipi_passive(dcpu, XIPIQ_OFFSET,
1117 APIC_DELMODE_FIXED);
1124 mp_bsp_simple_setup(void)
1126 /* build our map of 'other' CPUs */
1127 mycpu->gd_other_cpus = smp_startup_mask;
1128 CPUMASK_NANDBIT(mycpu->gd_other_cpus, mycpu->gd_cpuid);
1129 mycpu->gd_ipiq = (void *)kmem_alloc(&kernel_map,
1130 sizeof(lwkt_ipiq) * ncpus);
1131 bzero(mycpu->gd_ipiq, sizeof(lwkt_ipiq) * ncpus);
1135 if (cpu_feature & CPUID_TSC)
1136 tsc0_offset = rdtsc();
1141 * CPU TOPOLOGY DETECTION FUNCTIONS
1144 /* Detect intel topology using CPUID
1145 * Ref: http://www.intel.com/Assets/PDF/appnote/241618.pdf, pg 41
1148 detect_intel_topology(int count_htt_cores)
1152 int core_plus_logical_bits = 0;
1153 int cores_per_package;
1154 int logical_per_package;
1155 int logical_per_core;
1158 if (cpu_high >= 0xb) {
1161 } else if (cpu_high >= 0x4) {
1166 for (shift = 0; (1 << shift) < count_htt_cores; ++shift)
1168 logical_CPU_bits = 1 << shift;
1173 cpuid_count(0xb, FUNC_B_THREAD_LEVEL, p);
1175 /* if 0xb not supported - fallback to 0x4 */
1176 if (p[1] == 0 || (FUNC_B_TYPE(p[2]) != FUNC_B_THREAD_TYPE)) {
1180 logical_CPU_bits = FUNC_B_BITS_SHIFT_NEXT_LEVEL(p[0]);
1182 ecx_index = FUNC_B_THREAD_LEVEL + 1;
1184 cpuid_count(0xb, ecx_index, p);
1185 /* Check for the Core type in the implemented sub leaves. */
1186 if (FUNC_B_TYPE(p[2]) == FUNC_B_CORE_TYPE) {
1187 core_plus_logical_bits = FUNC_B_BITS_SHIFT_NEXT_LEVEL(p[0]);
1191 } while (FUNC_B_TYPE(p[2]) != FUNC_B_INVALID_TYPE);
1193 core_bits = core_plus_logical_bits - logical_CPU_bits;
1198 cpuid_count(0x4, 0, p);
1199 cores_per_package = FUNC_4_MAX_CORE_NO(p[0]) + 1;
1201 logical_per_package = count_htt_cores;
1202 logical_per_core = logical_per_package / cores_per_package;
1204 for (shift = 0; (1 << shift) < logical_per_core; ++shift)
1206 logical_CPU_bits = shift;
1208 for (shift = 0; (1 << shift) < cores_per_package; ++shift)
1215 /* Detect AMD topology using CPUID
1216 * Ref: http://support.amd.com/us/Embedded_TechDocs/25481.pdf, last page
1219 detect_amd_topology(int count_htt_cores)
1223 if ((cpu_feature & CPUID_HTT)
1224 && (amd_feature2 & AMDID2_CMP)) {
1226 if (cpu_procinfo2 & AMDID_COREID_SIZE) {
1227 core_bits = (cpu_procinfo2 & AMDID_COREID_SIZE)
1228 >> AMDID_COREID_SIZE_SHIFT;
1230 core_bits = (cpu_procinfo2 & AMDID_CMP_CORES) + 1;
1231 for (shift = 0; (1 << shift) < core_bits; ++shift);
1235 logical_CPU_bits = count_htt_cores >> core_bits;
1236 for (shift = 0; (1 << shift) < logical_CPU_bits; ++shift)
1238 logical_CPU_bits = shift;
1240 for (shift = 0; (1 << shift) < count_htt_cores; ++shift)
1243 logical_CPU_bits = 0;
1248 * - logical_CPU_bits
1250 * With the values above (for AMD or INTEL) we are able to generally
1251 * detect the CPU topology (number of cores for each level):
1252 * Ref: http://wiki.osdev.org/Detecting_CPU_Topology_(80x86)
1253 * Ref: http://www.multicoreinfo.com/research/papers/whitepapers/Intel-detect-topology.pdf
1256 detect_cpu_topology(void)
1258 static int topology_detected = 0;
1261 if (topology_detected) {
1265 if ((cpu_feature & CPUID_HTT) == 0) {
1267 logical_CPU_bits = 0;
1270 count = (cpu_procinfo & CPUID_HTT_CORES)
1271 >> CPUID_HTT_CORE_SHIFT;
1274 if (cpu_vendor_id == CPU_VENDOR_INTEL) {
1275 detect_intel_topology(count);
1276 } else if (cpu_vendor_id == CPU_VENDOR_AMD) {
1277 detect_amd_topology(count);
1282 kprintf("Bits within APICID: logical_CPU_bits: %d; core_bits: %d\n",
1283 logical_CPU_bits, core_bits);
1285 topology_detected = 1;
1288 /* Interface functions to calculate chip_ID,
1289 * core_number and logical_number
1290 * Ref: http://wiki.osdev.org/Detecting_CPU_Topology_(80x86)
1293 get_chip_ID(int cpuid)
1295 return get_apicid_from_cpuid(cpuid) >>
1296 (logical_CPU_bits + core_bits);
1300 get_core_number_within_chip(int cpuid)
1302 return (get_apicid_from_cpuid(cpuid) >> logical_CPU_bits) &
1303 ( (1 << core_bits) -1);
1307 get_logical_CPU_number_within_core(int cpuid)
1309 return get_apicid_from_cpuid(cpuid) &
1310 ( (1 << logical_CPU_bits) -1);