1 /*-
2 * Copyright (c) 1996, by Steve Passe
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
7 * are met:
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.
12 *
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
23 * SUCH DAMAGE.
24 */
25
26 #include <sys/cdefs.h>
27 __FBSDID("$FreeBSD: releng/7.3/sys/i386/i386/mp_machdep.c 198589 2009-10-29 14:34:02Z jhb $");
28
29 #include "opt_apic.h"
30 #include "opt_cpu.h"
31 #include "opt_kstack_pages.h"
32 #include "opt_mp_watchdog.h"
33 #include "opt_sched.h"
34 #include "opt_smp.h"
35
36 #if !defined(lint)
37 #if !defined(SMP)
38 #error How did you get here?
39 #endif
40
41 #ifndef DEV_APIC
42 #error The apic device is required for SMP, add "device apic" to your config file.
43 #endif
44 #if defined(CPU_DISABLE_CMPXCHG) && !defined(COMPILING_LINT)
45 #error SMP not supported with CPU_DISABLE_CMPXCHG
46 #endif
47 #endif /* not lint */
48
49 #include <sys/param.h>
50 #include <sys/systm.h>
51 #include <sys/bus.h>
52 #include <sys/cons.h> /* cngetc() */
53 #ifdef GPROF
54 #include <sys/gmon.h>
55 #endif
56 #include <sys/kernel.h>
57 #include <sys/ktr.h>
58 #include <sys/lock.h>
59 #include <sys/malloc.h>
60 #include <sys/memrange.h>
61 #include <sys/mutex.h>
62 #include <sys/pcpu.h>
63 #include <sys/proc.h>
64 #include <sys/sched.h>
65 #include <sys/smp.h>
66 #include <sys/sysctl.h>
67
68 #include <vm/vm.h>
69 #include <vm/vm_param.h>
70 #include <vm/pmap.h>
71 #include <vm/vm_kern.h>
72 #include <vm/vm_extern.h>
73
74 #include <machine/apicreg.h>
75 #include <machine/cputypes.h>
76 #include <machine/mca.h>
77 #include <machine/md_var.h>
78 #include <machine/mp_watchdog.h>
79 #include <machine/pcb.h>
80 #include <machine/psl.h>
81 #include <machine/smp.h>
82 #include <machine/specialreg.h>
83
84 #define WARMBOOT_TARGET 0
85 #define WARMBOOT_OFF (KERNBASE + 0x0467)
86 #define WARMBOOT_SEG (KERNBASE + 0x0469)
87
88 #define CMOS_REG (0x70)
89 #define CMOS_DATA (0x71)
90 #define BIOS_RESET (0x0f)
91 #define BIOS_WARM (0x0a)
92
93 /*
94 * this code MUST be enabled here and in mpboot.s.
95 * it follows the very early stages of AP boot by placing values in CMOS ram.
96 * it NORMALLY will never be needed and thus the primitive method for enabling.
97 *
98 #define CHECK_POINTS
99 */
100
101 #if defined(CHECK_POINTS) && !defined(PC98)
102 #define CHECK_READ(A) (outb(CMOS_REG, (A)), inb(CMOS_DATA))
103 #define CHECK_WRITE(A,D) (outb(CMOS_REG, (A)), outb(CMOS_DATA, (D)))
104
105 #define CHECK_INIT(D); \
106 CHECK_WRITE(0x34, (D)); \
107 CHECK_WRITE(0x35, (D)); \
108 CHECK_WRITE(0x36, (D)); \
109 CHECK_WRITE(0x37, (D)); \
110 CHECK_WRITE(0x38, (D)); \
111 CHECK_WRITE(0x39, (D));
112
113 #define CHECK_PRINT(S); \
114 printf("%s: %d, %d, %d, %d, %d, %d\n", \
115 (S), \
116 CHECK_READ(0x34), \
117 CHECK_READ(0x35), \
118 CHECK_READ(0x36), \
119 CHECK_READ(0x37), \
120 CHECK_READ(0x38), \
121 CHECK_READ(0x39));
122
123 #else /* CHECK_POINTS */
124
125 #define CHECK_INIT(D)
126 #define CHECK_PRINT(S)
127 #define CHECK_WRITE(A, D)
128
129 #endif /* CHECK_POINTS */
130
131 /* lock region used by kernel profiling */
132 int mcount_lock;
133
134 int mp_naps; /* # of Applications processors */
135 int boot_cpu_id = -1; /* designated BSP */
136
137 extern struct pcpu __pcpu[];
138
139 /*
140 * CPU topology map datastructures for HTT.
141 */
142 static struct cpu_group mp_groups[MAXCPU];
143 static struct cpu_top mp_top;
144
145 /* AP uses this during bootstrap. Do not staticize. */
146 char *bootSTK;
147 static int bootAP;
148
149 /* Free these after use */
150 void *bootstacks[MAXCPU];
151
152 /* Hotwire a 0->4MB V==P mapping */
153 extern pt_entry_t *KPTphys;
154
155 struct pcb stoppcbs[MAXCPU];
156
157 /* Variables needed for SMP tlb shootdown. */
158 vm_offset_t smp_tlb_addr1;
159 vm_offset_t smp_tlb_addr2;
160 volatile int smp_tlb_wait;
161
162 #ifdef STOP_NMI
163 volatile cpumask_t ipi_nmi_pending;
164
165 static void ipi_nmi_selected(u_int32_t cpus);
166 #endif
167
168 #ifdef COUNT_IPIS
169 /* Interrupt counts. */
170 static u_long *ipi_preempt_counts[MAXCPU];
171 static u_long *ipi_ast_counts[MAXCPU];
172 u_long *ipi_invltlb_counts[MAXCPU];
173 u_long *ipi_invlrng_counts[MAXCPU];
174 u_long *ipi_invlpg_counts[MAXCPU];
175 u_long *ipi_invlcache_counts[MAXCPU];
176 u_long *ipi_rendezvous_counts[MAXCPU];
177 u_long *ipi_lazypmap_counts[MAXCPU];
178 #endif
179
180 /*
181 * Local data and functions.
182 */
183
184 #ifdef STOP_NMI
185 /*
186 * Provide an alternate method of stopping other CPUs. If another CPU has
187 * disabled interrupts the conventional STOP IPI will be blocked. This
188 * NMI-based stop should get through in that case.
189 */
190 static int stop_cpus_with_nmi = 1;
191 SYSCTL_INT(_debug, OID_AUTO, stop_cpus_with_nmi, CTLTYPE_INT | CTLFLAG_RW,
192 &stop_cpus_with_nmi, 0, "");
193 TUNABLE_INT("debug.stop_cpus_with_nmi", &stop_cpus_with_nmi);
194 #else
195 #define stop_cpus_with_nmi 0
196 #endif
197
198 static u_int logical_cpus;
199
200 /* used to hold the AP's until we are ready to release them */
201 static struct mtx ap_boot_mtx;
202
203 /* Set to 1 once we're ready to let the APs out of the pen. */
204 static volatile int aps_ready = 0;
205
206 /*
207 * Store data from cpu_add() until later in the boot when we actually setup
208 * the APs.
209 */
210 struct cpu_info {
211 int cpu_present:1;
212 int cpu_bsp:1;
213 int cpu_disabled:1;
214 int cpu_hyperthread:1;
215 } static cpu_info[MAX_APIC_ID + 1];
216 int cpu_apic_ids[MAXCPU];
217
218 /* Holds pending bitmap based IPIs per CPU */
219 static volatile u_int cpu_ipi_pending[MAXCPU];
220
221 static u_int boot_address;
222
223 static void assign_cpu_ids(void);
224 static void install_ap_tramp(void);
225 static void set_interrupt_apic_ids(void);
226 static int start_all_aps(void);
227 static int start_ap(int apic_id);
228 static void release_aps(void *dummy);
229
230 static int hlt_logical_cpus;
231 static u_int hyperthreading_cpus;
232 static cpumask_t hyperthreading_cpus_mask;
233 static int hyperthreading_allowed = 1;
234 static struct sysctl_ctx_list logical_cpu_clist;
235
236 static void
237 mem_range_AP_init(void)
238 {
239 if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP)
240 mem_range_softc.mr_op->initAP(&mem_range_softc);
241 }
242
243 void
244 mp_topology(void)
245 {
246 struct cpu_group *group;
247 int apic_id;
248 int groups;
249 int cpu;
250
251 /* Build the smp_topology map. */
252 /* Nothing to do if there is no HTT support. */
253 if (hyperthreading_cpus <= 1)
254 return;
255 group = &mp_groups[0];
256 groups = 1;
257 for (cpu = 0, apic_id = 0; apic_id <= MAX_APIC_ID; apic_id++) {
258 if (!cpu_info[apic_id].cpu_present)
259 continue;
260 /*
261 * If the current group has members and we're not a logical
262 * cpu, create a new group.
263 */
264 if (group->cg_count != 0 &&
265 (apic_id % hyperthreading_cpus) == 0) {
266 group++;
267 groups++;
268 }
269 group->cg_count++;
270 group->cg_mask |= 1 << cpu;
271 cpu++;
272 }
273
274 mp_top.ct_count = groups;
275 mp_top.ct_group = mp_groups;
276 smp_topology = &mp_top;
277 }
278
279
280 /*
281 * Calculate usable address in base memory for AP trampoline code.
282 */
283 u_int
284 mp_bootaddress(u_int basemem)
285 {
286
287 boot_address = trunc_page(basemem); /* round down to 4k boundary */
288 if ((basemem - boot_address) < bootMP_size)
289 boot_address -= PAGE_SIZE; /* not enough, lower by 4k */
290
291 return boot_address;
292 }
293
294 void
295 cpu_add(u_int apic_id, char boot_cpu)
296 {
297
298 if (apic_id > MAX_APIC_ID) {
299 panic("SMP: APIC ID %d too high", apic_id);
300 return;
301 }
302 KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %d added twice",
303 apic_id));
304 cpu_info[apic_id].cpu_present = 1;
305 if (boot_cpu) {
306 KASSERT(boot_cpu_id == -1,
307 ("CPU %d claims to be BSP, but CPU %d already is", apic_id,
308 boot_cpu_id));
309 boot_cpu_id = apic_id;
310 cpu_info[apic_id].cpu_bsp = 1;
311 }
312 if (mp_ncpus < MAXCPU)
313 mp_ncpus++;
314 if (bootverbose)
315 printf("SMP: Added CPU %d (%s)\n", apic_id, boot_cpu ? "BSP" :
316 "AP");
317 }
318
319 void
320 cpu_mp_setmaxid(void)
321 {
322
323 mp_maxid = MAXCPU - 1;
324 }
325
326 int
327 cpu_mp_probe(void)
328 {
329
330 /*
331 * Always record BSP in CPU map so that the mbuf init code works
332 * correctly.
333 */
334 all_cpus = 1;
335 if (mp_ncpus == 0) {
336 /*
337 * No CPUs were found, so this must be a UP system. Setup
338 * the variables to represent a system with a single CPU
339 * with an id of 0.
340 */
341 mp_ncpus = 1;
342 return (0);
343 }
344
345 /* At least one CPU was found. */
346 if (mp_ncpus == 1) {
347 /*
348 * One CPU was found, so this must be a UP system with
349 * an I/O APIC.
350 */
351 return (0);
352 }
353
354 /* At least two CPUs were found. */
355 return (1);
356 }
357
358 /*
359 * Initialize the IPI handlers and start up the AP's.
360 */
361 void
362 cpu_mp_start(void)
363 {
364 int i;
365 u_int threads_per_cache, p[4];
366
367 /* Initialize the logical ID to APIC ID table. */
368 for (i = 0; i < MAXCPU; i++) {
369 cpu_apic_ids[i] = -1;
370 cpu_ipi_pending[i] = 0;
371 }
372
373 /* Install an inter-CPU IPI for TLB invalidation */
374 setidt(IPI_INVLTLB, IDTVEC(invltlb),
375 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
376 setidt(IPI_INVLPG, IDTVEC(invlpg),
377 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
378 setidt(IPI_INVLRNG, IDTVEC(invlrng),
379 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
380
381 /* Install an inter-CPU IPI for cache invalidation. */
382 setidt(IPI_INVLCACHE, IDTVEC(invlcache),
383 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
384
385 /* Install an inter-CPU IPI for lazy pmap release */
386 setidt(IPI_LAZYPMAP, IDTVEC(lazypmap),
387 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
388
389 /* Install an inter-CPU IPI for all-CPU rendezvous */
390 setidt(IPI_RENDEZVOUS, IDTVEC(rendezvous),
391 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
392
393 /* Install generic inter-CPU IPI handler */
394 setidt(IPI_BITMAP_VECTOR, IDTVEC(ipi_intr_bitmap_handler),
395 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
396
397 /* Install an inter-CPU IPI for CPU stop/restart */
398 setidt(IPI_STOP, IDTVEC(cpustop),
399 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
400
401
402 /* Set boot_cpu_id if needed. */
403 if (boot_cpu_id == -1) {
404 boot_cpu_id = PCPU_GET(apic_id);
405 cpu_info[boot_cpu_id].cpu_bsp = 1;
406 } else
407 KASSERT(boot_cpu_id == PCPU_GET(apic_id),
408 ("BSP's APIC ID doesn't match boot_cpu_id"));
409
410 /* Setup the initial logical CPUs info. */
411 logical_cpus = logical_cpus_mask = 0;
412 if (cpu_feature & CPUID_HTT)
413 logical_cpus = (cpu_procinfo & CPUID_HTT_CORES) >> 16;
414
415 /*
416 * Work out if hyperthreading is *really* enabled. This
417 * is made really ugly by the fact that processors lie: Dual
418 * core processors claim to be hyperthreaded even when they're
419 * not, presumably because they want to be treated the same
420 * way as HTT with respect to per-cpu software licensing.
421 * At the time of writing (May 12, 2005) the only hyperthreaded
422 * cpus are from Intel, and Intel's dual-core processors can be
423 * identified via the "deterministic cache parameters" cpuid
424 * calls.
425 */
426 /*
427 * First determine if this is an Intel processor which claims
428 * to have hyperthreading support.
429 */
430 if ((cpu_feature & CPUID_HTT) && cpu_vendor_id == CPU_VENDOR_INTEL) {
431 /*
432 * If the "deterministic cache parameters" cpuid calls
433 * are available, use them.
434 */
435 if (cpu_high >= 4) {
436 /* Ask the processor about the L1 cache. */
437 for (i = 0; i < 1; i++) {
438 cpuid_count(4, i, p);
439 threads_per_cache = ((p[0] & 0x3ffc000) >> 14) + 1;
440 if (hyperthreading_cpus < threads_per_cache)
441 hyperthreading_cpus = threads_per_cache;
442 if ((p[0] & 0x1f) == 0)
443 break;
444 }
445 }
446
447 /*
448 * If the deterministic cache parameters are not
449 * available, or if no caches were reported to exist,
450 * just accept what the HTT flag indicated.
451 */
452 if (hyperthreading_cpus == 0)
453 hyperthreading_cpus = logical_cpus;
454 }
455
456 assign_cpu_ids();
457
458 /* Start each Application Processor */
459 start_all_aps();
460
461 set_interrupt_apic_ids();
462
463 /* Last, setup the cpu topology now that we have probed CPUs */
464 mp_topology();
465 }
466
467
468 /*
469 * Print various information about the SMP system hardware and setup.
470 */
471 void
472 cpu_mp_announce(void)
473 {
474 const char *hyperthread;
475 int i;
476
477 /* List active CPUs first. */
478 printf(" cpu0 (BSP): APIC ID: %2d\n", boot_cpu_id);
479 for (i = 1; i < mp_ncpus; i++) {
480 if (cpu_info[cpu_apic_ids[i]].cpu_hyperthread)
481 hyperthread = "/HT";
482 else
483 hyperthread = "";
484 printf(" cpu%d (AP%s): APIC ID: %2d\n", i, hyperthread,
485 cpu_apic_ids[i]);
486 }
487
488 /* List disabled CPUs last. */
489 for (i = 0; i <= MAX_APIC_ID; i++) {
490 if (!cpu_info[i].cpu_present || !cpu_info[i].cpu_disabled)
491 continue;
492 if (cpu_info[i].cpu_hyperthread)
493 hyperthread = "/HT";
494 else
495 hyperthread = "";
496 printf(" cpu (AP%s): APIC ID: %2d (disabled)\n", hyperthread,
497 i);
498 }
499 }
500
501 /*
502 * AP CPU's call this to initialize themselves.
503 */
504 void
505 init_secondary(void)
506 {
507 struct pcpu *pc;
508 vm_offset_t addr;
509 int gsel_tss;
510 int x, myid;
511 u_int cr0;
512
513 /* bootAP is set in start_ap() to our ID. */
514 myid = bootAP;
515
516 /* Get per-cpu data */
517 pc = &__pcpu[myid];
518
519 /* prime data page for it to use */
520 pcpu_init(pc, myid, sizeof(struct pcpu));
521 pc->pc_apic_id = cpu_apic_ids[myid];
522 pc->pc_prvspace = pc;
523 pc->pc_curthread = 0;
524
525 gdt_segs[GPRIV_SEL].ssd_base = (int) pc;
526 gdt_segs[GPROC0_SEL].ssd_base = (int) &pc->pc_common_tss;
527
528 for (x = 0; x < NGDT; x++) {
529 ssdtosd(&gdt_segs[x], &gdt[myid * NGDT + x].sd);
530 }
531
532 r_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
533 r_gdt.rd_base = (int) &gdt[myid * NGDT];
534 lgdt(&r_gdt); /* does magic intra-segment return */
535
536 lidt(&r_idt);
537
538 lldt(_default_ldt);
539 PCPU_SET(currentldt, _default_ldt);
540
541 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
542 gdt[myid * NGDT + GPROC0_SEL].sd.sd_type = SDT_SYS386TSS;
543 PCPU_SET(common_tss.tss_esp0, 0); /* not used until after switch */
544 PCPU_SET(common_tss.tss_ss0, GSEL(GDATA_SEL, SEL_KPL));
545 PCPU_SET(common_tss.tss_ioopt, (sizeof (struct i386tss)) << 16);
546 PCPU_SET(tss_gdt, &gdt[myid * NGDT + GPROC0_SEL].sd);
547 PCPU_SET(common_tssd, *PCPU_GET(tss_gdt));
548 ltr(gsel_tss);
549
550 PCPU_SET(fsgs_gdt, &gdt[myid * NGDT + GUFS_SEL].sd);
551
552 /*
553 * Set to a known state:
554 * Set by mpboot.s: CR0_PG, CR0_PE
555 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
556 */
557 cr0 = rcr0();
558 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
559 load_cr0(cr0);
560 CHECK_WRITE(0x38, 5);
561
562 /* Disable local APIC just to be sure. */
563 lapic_disable();
564
565 /* signal our startup to the BSP. */
566 mp_naps++;
567 CHECK_WRITE(0x39, 6);
568
569 /* Spin until the BSP releases the AP's. */
570 while (!aps_ready)
571 ia32_pause();
572
573 /* BSP may have changed PTD while we were waiting */
574 invltlb();
575 for (addr = 0; addr < NKPT * NBPDR - 1; addr += PAGE_SIZE)
576 invlpg(addr);
577
578 #if defined(I586_CPU) && !defined(NO_F00F_HACK)
579 lidt(&r_idt);
580 #endif
581
582 /* Initialize the PAT MSR if present. */
583 pmap_init_pat();
584
585 /* set up CPU registers and state */
586 cpu_setregs();
587
588 /* set up FPU state on the AP */
589 npxinit();
590
591 /* set up SSE registers */
592 enable_sse();
593
594 #ifdef PAE
595 /* Enable the PTE no-execute bit. */
596 if ((amd_feature & AMDID_NX) != 0) {
597 uint64_t msr;
598
599 msr = rdmsr(MSR_EFER) | EFER_NXE;
600 wrmsr(MSR_EFER, msr);
601 }
602 #endif
603
604 /* A quick check from sanity claus */
605 if (PCPU_GET(apic_id) != lapic_id()) {
606 printf("SMP: cpuid = %d\n", PCPU_GET(cpuid));
607 printf("SMP: actual apic_id = %d\n", lapic_id());
608 printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id));
609 panic("cpuid mismatch! boom!!");
610 }
611
612 /* Initialize curthread. */
613 KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread"));
614 PCPU_SET(curthread, PCPU_GET(idlethread));
615
616 mca_init();
617
618 mtx_lock_spin(&ap_boot_mtx);
619
620 /* Init local apic for irq's */
621 lapic_setup(1);
622
623 /* Set memory range attributes for this CPU to match the BSP */
624 mem_range_AP_init();
625
626 smp_cpus++;
627
628 CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", PCPU_GET(cpuid));
629 printf("SMP: AP CPU #%d Launched!\n", PCPU_GET(cpuid));
630
631 /* Determine if we are a logical CPU. */
632 if (logical_cpus > 1 && PCPU_GET(apic_id) % logical_cpus != 0)
633 logical_cpus_mask |= PCPU_GET(cpumask);
634
635 /* Determine if we are a hyperthread. */
636 if (hyperthreading_cpus > 1 &&
637 PCPU_GET(apic_id) % hyperthreading_cpus != 0)
638 hyperthreading_cpus_mask |= PCPU_GET(cpumask);
639
640 /* Build our map of 'other' CPUs. */
641 PCPU_SET(other_cpus, all_cpus & ~PCPU_GET(cpumask));
642
643 if (bootverbose)
644 lapic_dump("AP");
645
646 if (smp_cpus == mp_ncpus) {
647 /* enable IPI's, tlb shootdown, freezes etc */
648 atomic_store_rel_int(&smp_started, 1);
649 smp_active = 1; /* historic */
650 }
651
652 mtx_unlock_spin(&ap_boot_mtx);
653
654 /* wait until all the AP's are up */
655 while (smp_started == 0)
656 ia32_pause();
657
658 /* enter the scheduler */
659 sched_throw(NULL);
660
661 panic("scheduler returned us to %s", __func__);
662 /* NOTREACHED */
663 }
664
665 /*******************************************************************
666 * local functions and data
667 */
668
669 /*
670 * We tell the I/O APIC code about all the CPUs we want to receive
671 * interrupts. If we don't want certain CPUs to receive IRQs we
672 * can simply not tell the I/O APIC code about them in this function.
673 * We also do not tell it about the BSP since it tells itself about
674 * the BSP internally to work with UP kernels and on UP machines.
675 */
676 static void
677 set_interrupt_apic_ids(void)
678 {
679 u_int i, apic_id;
680
681 for (i = 0; i < MAXCPU; i++) {
682 apic_id = cpu_apic_ids[i];
683 if (apic_id == -1)
684 continue;
685 if (cpu_info[apic_id].cpu_bsp)
686 continue;
687 if (cpu_info[apic_id].cpu_disabled)
688 continue;
689
690 /* Don't let hyperthreads service interrupts. */
691 if (hyperthreading_cpus > 1 &&
692 apic_id % hyperthreading_cpus != 0)
693 continue;
694
695 intr_add_cpu(i);
696 }
697 }
698
699 /*
700 * Assign logical CPU IDs to local APICs.
701 */
702 static void
703 assign_cpu_ids(void)
704 {
705 u_int i;
706
707 TUNABLE_INT_FETCH("machdep.hyperthreading_allowed",
708 &hyperthreading_allowed);
709
710 /* Check for explicitly disabled CPUs. */
711 for (i = 0; i <= MAX_APIC_ID; i++) {
712 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp)
713 continue;
714
715 if (hyperthreading_cpus > 1 && i % hyperthreading_cpus != 0) {
716 cpu_info[i].cpu_hyperthread = 1;
717 #if defined(SCHED_ULE)
718 /*
719 * Don't use HT CPU if it has been disabled by a
720 * tunable.
721 */
722 if (hyperthreading_allowed == 0) {
723 cpu_info[i].cpu_disabled = 1;
724 continue;
725 }
726 #endif
727 }
728
729 /* Don't use this CPU if it has been disabled by a tunable. */
730 if (resource_disabled("lapic", i)) {
731 cpu_info[i].cpu_disabled = 1;
732 continue;
733 }
734 }
735
736 /*
737 * Assign CPU IDs to local APIC IDs and disable any CPUs
738 * beyond MAXCPU. CPU 0 is always assigned to the BSP.
739 *
740 * To minimize confusion for userland, we attempt to number
741 * CPUs such that all threads and cores in a package are
742 * grouped together. For now we assume that the BSP is always
743 * the first thread in a package and just start adding APs
744 * starting with the BSP's APIC ID.
745 */
746 mp_ncpus = 1;
747 cpu_apic_ids[0] = boot_cpu_id;
748 for (i = boot_cpu_id + 1; i != boot_cpu_id;
749 i == MAX_APIC_ID ? i = 0 : i++) {
750 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp ||
751 cpu_info[i].cpu_disabled)
752 continue;
753
754 if (mp_ncpus < MAXCPU) {
755 cpu_apic_ids[mp_ncpus] = i;
756 mp_ncpus++;
757 } else
758 cpu_info[i].cpu_disabled = 1;
759 }
760 KASSERT(mp_maxid >= mp_ncpus - 1,
761 ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid,
762 mp_ncpus));
763 }
764
765 /*
766 * start each AP in our list
767 */
768 static int
769 start_all_aps(void)
770 {
771 #ifndef PC98
772 u_char mpbiosreason;
773 #endif
774 uintptr_t kptbase;
775 u_int32_t mpbioswarmvec;
776 int apic_id, cpu, i;
777
778 mtx_init(&ap_boot_mtx, "ap boot", NULL, MTX_SPIN);
779
780 /* install the AP 1st level boot code */
781 install_ap_tramp();
782
783 /* save the current value of the warm-start vector */
784 mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF);
785 #ifndef PC98
786 outb(CMOS_REG, BIOS_RESET);
787 mpbiosreason = inb(CMOS_DATA);
788 #endif
789
790 /* set up temporary P==V mapping for AP boot */
791 /* XXX this is a hack, we should boot the AP on its own stack/PTD */
792 kptbase = (uintptr_t)(void *)KPTphys;
793 for (i = 0; i < NKPT; i++)
794 PTD[i] = (pd_entry_t)(PG_V | PG_RW |
795 ((kptbase + i * PAGE_SIZE) & PG_FRAME));
796 invltlb();
797
798 /* start each AP */
799 for (cpu = 1; cpu < mp_ncpus; cpu++) {
800 apic_id = cpu_apic_ids[cpu];
801
802 /* allocate and set up a boot stack data page */
803 bootstacks[cpu] = (char *)kmem_alloc(kernel_map, KSTACK_PAGES * PAGE_SIZE);
804
805 /* setup a vector to our boot code */
806 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
807 *((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4);
808 #ifndef PC98
809 outb(CMOS_REG, BIOS_RESET);
810 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
811 #endif
812
813 bootSTK = (char *)bootstacks[cpu] + KSTACK_PAGES * PAGE_SIZE - 4;
814 bootAP = cpu;
815
816 /* attempt to start the Application Processor */
817 CHECK_INIT(99); /* setup checkpoints */
818 if (!start_ap(apic_id)) {
819 printf("AP #%d (PHY# %d) failed!\n", cpu, apic_id);
820 CHECK_PRINT("trace"); /* show checkpoints */
821 /* better panic as the AP may be running loose */
822 printf("panic y/n? [y] ");
823 if (cngetc() != 'n')
824 panic("bye-bye");
825 }
826 CHECK_PRINT("trace"); /* show checkpoints */
827
828 all_cpus |= (1 << cpu); /* record AP in CPU map */
829 }
830
831 /* build our map of 'other' CPUs */
832 PCPU_SET(other_cpus, all_cpus & ~PCPU_GET(cpumask));
833
834 /* restore the warmstart vector */
835 *(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec;
836
837 #ifndef PC98
838 outb(CMOS_REG, BIOS_RESET);
839 outb(CMOS_DATA, mpbiosreason);
840 #endif
841
842 /* Undo V==P hack from above */
843 for (i = 0; i < NKPT; i++)
844 PTD[i] = 0;
845 pmap_invalidate_range(kernel_pmap, 0, NKPT * NBPDR - 1);
846
847 /* number of APs actually started */
848 return mp_naps;
849 }
850
851 /*
852 * load the 1st level AP boot code into base memory.
853 */
854
855 /* targets for relocation */
856 extern void bigJump(void);
857 extern void bootCodeSeg(void);
858 extern void bootDataSeg(void);
859 extern void MPentry(void);
860 extern u_int MP_GDT;
861 extern u_int mp_gdtbase;
862
863 static void
864 install_ap_tramp(void)
865 {
866 int x;
867 int size = *(int *) ((u_long) & bootMP_size);
868 vm_offset_t va = boot_address + KERNBASE;
869 u_char *src = (u_char *) ((u_long) bootMP);
870 u_char *dst = (u_char *) va;
871 u_int boot_base = (u_int) bootMP;
872 u_int8_t *dst8;
873 u_int16_t *dst16;
874 u_int32_t *dst32;
875
876 KASSERT (size <= PAGE_SIZE,
877 ("'size' do not fit into PAGE_SIZE, as expected."));
878 pmap_kenter(va, boot_address);
879 pmap_invalidate_page (kernel_pmap, va);
880 for (x = 0; x < size; ++x)
881 *dst++ = *src++;
882
883 /*
884 * modify addresses in code we just moved to basemem. unfortunately we
885 * need fairly detailed info about mpboot.s for this to work. changes
886 * to mpboot.s might require changes here.
887 */
888
889 /* boot code is located in KERNEL space */
890 dst = (u_char *) va;
891
892 /* modify the lgdt arg */
893 dst32 = (u_int32_t *) (dst + ((u_int) & mp_gdtbase - boot_base));
894 *dst32 = boot_address + ((u_int) & MP_GDT - boot_base);
895
896 /* modify the ljmp target for MPentry() */
897 dst32 = (u_int32_t *) (dst + ((u_int) bigJump - boot_base) + 1);
898 *dst32 = ((u_int) MPentry - KERNBASE);
899
900 /* modify the target for boot code segment */
901 dst16 = (u_int16_t *) (dst + ((u_int) bootCodeSeg - boot_base));
902 dst8 = (u_int8_t *) (dst16 + 1);
903 *dst16 = (u_int) boot_address & 0xffff;
904 *dst8 = ((u_int) boot_address >> 16) & 0xff;
905
906 /* modify the target for boot data segment */
907 dst16 = (u_int16_t *) (dst + ((u_int) bootDataSeg - boot_base));
908 dst8 = (u_int8_t *) (dst16 + 1);
909 *dst16 = (u_int) boot_address & 0xffff;
910 *dst8 = ((u_int) boot_address >> 16) & 0xff;
911 }
912
913 /*
914 * This function starts the AP (application processor) identified
915 * by the APIC ID 'physicalCpu'. It does quite a "song and dance"
916 * to accomplish this. This is necessary because of the nuances
917 * of the different hardware we might encounter. It isn't pretty,
918 * but it seems to work.
919 */
920 static int
921 start_ap(int apic_id)
922 {
923 int vector, ms;
924 int cpus;
925
926 /* calculate the vector */
927 vector = (boot_address >> 12) & 0xff;
928
929 /* used as a watchpoint to signal AP startup */
930 cpus = mp_naps;
931
932 /*
933 * first we do an INIT/RESET IPI this INIT IPI might be run, reseting
934 * and running the target CPU. OR this INIT IPI might be latched (P5
935 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
936 * ignored.
937 */
938
939 /* do an INIT IPI: assert RESET */
940 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
941 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id);
942
943 /* wait for pending status end */
944 lapic_ipi_wait(-1);
945
946 /* do an INIT IPI: deassert RESET */
947 lapic_ipi_raw(APIC_DEST_ALLESELF | APIC_TRIGMOD_LEVEL |
948 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, 0);
949
950 /* wait for pending status end */
951 DELAY(10000); /* wait ~10mS */
952 lapic_ipi_wait(-1);
953
954 /*
955 * next we do a STARTUP IPI: the previous INIT IPI might still be
956 * latched, (P5 bug) this 1st STARTUP would then terminate
957 * immediately, and the previously started INIT IPI would continue. OR
958 * the previous INIT IPI has already run. and this STARTUP IPI will
959 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
960 * will run.
961 */
962
963 /* do a STARTUP IPI */
964 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
965 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
966 vector, apic_id);
967 lapic_ipi_wait(-1);
968 DELAY(200); /* wait ~200uS */
969
970 /*
971 * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
972 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
973 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
974 * recognized after hardware RESET or INIT IPI.
975 */
976
977 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
978 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
979 vector, apic_id);
980 lapic_ipi_wait(-1);
981 DELAY(200); /* wait ~200uS */
982
983 /* Wait up to 5 seconds for it to start. */
984 for (ms = 0; ms < 5000; ms++) {
985 if (mp_naps > cpus)
986 return 1; /* return SUCCESS */
987 DELAY(1000);
988 }
989 return 0; /* return FAILURE */
990 }
991
992 #ifdef COUNT_XINVLTLB_HITS
993 u_int xhits_gbl[MAXCPU];
994 u_int xhits_pg[MAXCPU];
995 u_int xhits_rng[MAXCPU];
996 SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW, 0, "");
997 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl,
998 sizeof(xhits_gbl), "IU", "");
999 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg,
1000 sizeof(xhits_pg), "IU", "");
1001 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng,
1002 sizeof(xhits_rng), "IU", "");
1003
1004 u_int ipi_global;
1005 u_int ipi_page;
1006 u_int ipi_range;
1007 u_int ipi_range_size;
1008 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, "");
1009 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, "");
1010 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, "");
1011 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW, &ipi_range_size,
1012 0, "");
1013
1014 u_int ipi_masked_global;
1015 u_int ipi_masked_page;
1016 u_int ipi_masked_range;
1017 u_int ipi_masked_range_size;
1018 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_masked_global, CTLFLAG_RW,
1019 &ipi_masked_global, 0, "");
1020 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_masked_page, CTLFLAG_RW,
1021 &ipi_masked_page, 0, "");
1022 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_masked_range, CTLFLAG_RW,
1023 &ipi_masked_range, 0, "");
1024 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_masked_range_size, CTLFLAG_RW,
1025 &ipi_masked_range_size, 0, "");
1026 #endif /* COUNT_XINVLTLB_HITS */
1027
1028 /*
1029 * Flush the TLB on all other CPU's
1030 */
1031 static void
1032 smp_tlb_shootdown(u_int vector, vm_offset_t addr1, vm_offset_t addr2)
1033 {
1034 u_int ncpu;
1035
1036 ncpu = mp_ncpus - 1; /* does not shootdown self */
1037 if (ncpu < 1)
1038 return; /* no other cpus */
1039 if (!(read_eflags() & PSL_I))
1040 panic("%s: interrupts disabled", __func__);
1041 mtx_lock_spin(&smp_ipi_mtx);
1042 smp_tlb_addr1 = addr1;
1043 smp_tlb_addr2 = addr2;
1044 atomic_store_rel_int(&smp_tlb_wait, 0);
1045 ipi_all_but_self(vector);
1046 while (smp_tlb_wait < ncpu)
1047 ia32_pause();
1048 mtx_unlock_spin(&smp_ipi_mtx);
1049 }
1050
1051 static void
1052 smp_targeted_tlb_shootdown(u_int mask, u_int vector, vm_offset_t addr1, vm_offset_t addr2)
1053 {
1054 int ncpu, othercpus;
1055
1056 othercpus = mp_ncpus - 1;
1057 if (mask == (u_int)-1) {
1058 ncpu = othercpus;
1059 if (ncpu < 1)
1060 return;
1061 } else {
1062 mask &= ~PCPU_GET(cpumask);
1063 if (mask == 0)
1064 return;
1065 ncpu = bitcount32(mask);
1066 if (ncpu > othercpus) {
1067 /* XXX this should be a panic offence */
1068 printf("SMP: tlb shootdown to %d other cpus (only have %d)\n",
1069 ncpu, othercpus);
1070 ncpu = othercpus;
1071 }
1072 /* XXX should be a panic, implied by mask == 0 above */
1073 if (ncpu < 1)
1074 return;
1075 }
1076 if (!(read_eflags() & PSL_I))
1077 panic("%s: interrupts disabled", __func__);
1078 mtx_lock_spin(&smp_ipi_mtx);
1079 smp_tlb_addr1 = addr1;
1080 smp_tlb_addr2 = addr2;
1081 atomic_store_rel_int(&smp_tlb_wait, 0);
1082 if (mask == (u_int)-1)
1083 ipi_all_but_self(vector);
1084 else
1085 ipi_selected(mask, vector);
1086 while (smp_tlb_wait < ncpu)
1087 ia32_pause();
1088 mtx_unlock_spin(&smp_ipi_mtx);
1089 }
1090
1091 void
1092 smp_cache_flush(void)
1093 {
1094
1095 if (smp_started)
1096 smp_tlb_shootdown(IPI_INVLCACHE, 0, 0);
1097 }
1098
1099 void
1100 smp_invltlb(void)
1101 {
1102
1103 if (smp_started) {
1104 smp_tlb_shootdown(IPI_INVLTLB, 0, 0);
1105 #ifdef COUNT_XINVLTLB_HITS
1106 ipi_global++;
1107 #endif
1108 }
1109 }
1110
1111 void
1112 smp_invlpg(vm_offset_t addr)
1113 {
1114
1115 if (smp_started) {
1116 smp_tlb_shootdown(IPI_INVLPG, addr, 0);
1117 #ifdef COUNT_XINVLTLB_HITS
1118 ipi_page++;
1119 #endif
1120 }
1121 }
1122
1123 void
1124 smp_invlpg_range(vm_offset_t addr1, vm_offset_t addr2)
1125 {
1126
1127 if (smp_started) {
1128 smp_tlb_shootdown(IPI_INVLRNG, addr1, addr2);
1129 #ifdef COUNT_XINVLTLB_HITS
1130 ipi_range++;
1131 ipi_range_size += (addr2 - addr1) / PAGE_SIZE;
1132 #endif
1133 }
1134 }
1135
1136 void
1137 smp_masked_invltlb(u_int mask)
1138 {
1139
1140 if (smp_started) {
1141 smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, 0, 0);
1142 #ifdef COUNT_XINVLTLB_HITS
1143 ipi_masked_global++;
1144 #endif
1145 }
1146 }
1147
1148 void
1149 smp_masked_invlpg(u_int mask, vm_offset_t addr)
1150 {
1151
1152 if (smp_started) {
1153 smp_targeted_tlb_shootdown(mask, IPI_INVLPG, addr, 0);
1154 #ifdef COUNT_XINVLTLB_HITS
1155 ipi_masked_page++;
1156 #endif
1157 }
1158 }
1159
1160 void
1161 smp_masked_invlpg_range(u_int mask, vm_offset_t addr1, vm_offset_t addr2)
1162 {
1163
1164 if (smp_started) {
1165 smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, addr1, addr2);
1166 #ifdef COUNT_XINVLTLB_HITS
1167 ipi_masked_range++;
1168 ipi_masked_range_size += (addr2 - addr1) / PAGE_SIZE;
1169 #endif
1170 }
1171 }
1172
1173 void
1174 ipi_bitmap_handler(struct trapframe frame)
1175 {
1176 int cpu = PCPU_GET(cpuid);
1177 u_int ipi_bitmap;
1178
1179 ipi_bitmap = atomic_readandclear_int(&cpu_ipi_pending[cpu]);
1180
1181 if (ipi_bitmap & (1 << IPI_PREEMPT)) {
1182 struct thread *running_thread = curthread;
1183 #ifdef COUNT_IPIS
1184 (*ipi_preempt_counts[cpu])++;
1185 #endif
1186 thread_lock(running_thread);
1187 if (running_thread->td_critnest > 1)
1188 running_thread->td_owepreempt = 1;
1189 else
1190 mi_switch(SW_INVOL | SW_PREEMPT, NULL);
1191 thread_unlock(running_thread);
1192 }
1193
1194 if (ipi_bitmap & (1 << IPI_AST)) {
1195 #ifdef COUNT_IPIS
1196 (*ipi_ast_counts[cpu])++;
1197 #endif
1198 /* Nothing to do for AST */
1199 }
1200 }
1201
1202 /*
1203 * send an IPI to a set of cpus.
1204 */
1205 void
1206 ipi_selected(u_int32_t cpus, u_int ipi)
1207 {
1208 int cpu;
1209 u_int bitmap = 0;
1210 u_int old_pending;
1211 u_int new_pending;
1212
1213 if (IPI_IS_BITMAPED(ipi)) {
1214 bitmap = 1 << ipi;
1215 ipi = IPI_BITMAP_VECTOR;
1216 }
1217
1218 #ifdef STOP_NMI
1219 if (ipi == IPI_STOP && stop_cpus_with_nmi) {
1220 ipi_nmi_selected(cpus);
1221 return;
1222 }
1223 #endif
1224 CTR3(KTR_SMP, "%s: cpus: %x ipi: %x", __func__, cpus, ipi);
1225 while ((cpu = ffs(cpus)) != 0) {
1226 cpu--;
1227 cpus &= ~(1 << cpu);
1228
1229 KASSERT(cpu_apic_ids[cpu] != -1,
1230 ("IPI to non-existent CPU %d", cpu));
1231
1232 if (bitmap) {
1233 do {
1234 old_pending = cpu_ipi_pending[cpu];
1235 new_pending = old_pending | bitmap;
1236 } while (!atomic_cmpset_int(&cpu_ipi_pending[cpu],old_pending, new_pending));
1237
1238 if (old_pending)
1239 continue;
1240 }
1241
1242 lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]);
1243 }
1244
1245 }
1246
1247 /*
1248 * send an IPI INTerrupt containing 'vector' to all CPUs, including myself
1249 */
1250 void
1251 ipi_all(u_int ipi)
1252 {
1253
1254 if (IPI_IS_BITMAPED(ipi) || (ipi == IPI_STOP && stop_cpus_with_nmi)) {
1255 ipi_selected(all_cpus, ipi);
1256 return;
1257 }
1258 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1259 lapic_ipi_vectored(ipi, APIC_IPI_DEST_ALL);
1260 }
1261
1262 /*
1263 * send an IPI to all CPUs EXCEPT myself
1264 */
1265 void
1266 ipi_all_but_self(u_int ipi)
1267 {
1268
1269 if (IPI_IS_BITMAPED(ipi) || (ipi == IPI_STOP && stop_cpus_with_nmi)) {
1270 ipi_selected(PCPU_GET(other_cpus), ipi);
1271 return;
1272 }
1273 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1274 lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS);
1275 }
1276
1277 /*
1278 * send an IPI to myself
1279 */
1280 void
1281 ipi_self(u_int ipi)
1282 {
1283
1284 if (IPI_IS_BITMAPED(ipi) || (ipi == IPI_STOP && stop_cpus_with_nmi)) {
1285 ipi_selected(PCPU_GET(cpumask), ipi);
1286 return;
1287 }
1288 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1289 lapic_ipi_vectored(ipi, APIC_IPI_DEST_SELF);
1290 }
1291
1292 #ifdef STOP_NMI
1293 /*
1294 * send NMI IPI to selected CPUs
1295 */
1296
1297 #define BEFORE_SPIN 1000000
1298
1299 void
1300 ipi_nmi_selected(u_int32_t cpus)
1301 {
1302 int cpu;
1303 register_t icrlo;
1304
1305 icrlo = APIC_DELMODE_NMI | APIC_DESTMODE_PHY | APIC_LEVEL_ASSERT
1306 | APIC_TRIGMOD_EDGE;
1307
1308 CTR2(KTR_SMP, "%s: cpus: %x nmi", __func__, cpus);
1309
1310 atomic_set_int(&ipi_nmi_pending, cpus);
1311
1312 while ((cpu = ffs(cpus)) != 0) {
1313 cpu--;
1314 cpus &= ~(1 << cpu);
1315
1316 KASSERT(cpu_apic_ids[cpu] != -1,
1317 ("IPI NMI to non-existent CPU %d", cpu));
1318
1319 /* Wait for an earlier IPI to finish. */
1320 if (!lapic_ipi_wait(BEFORE_SPIN))
1321 panic("ipi_nmi_selected: previous IPI has not cleared");
1322
1323 lapic_ipi_raw(icrlo, cpu_apic_ids[cpu]);
1324 }
1325 }
1326
1327 int
1328 ipi_nmi_handler(void)
1329 {
1330 int cpumask = PCPU_GET(cpumask);
1331
1332 if (!(ipi_nmi_pending & cpumask))
1333 return 1;
1334
1335 atomic_clear_int(&ipi_nmi_pending, cpumask);
1336 cpustop_handler();
1337 return 0;
1338 }
1339
1340 #endif /* STOP_NMI */
1341
1342 /*
1343 * Handle an IPI_STOP by saving our current context and spinning until we
1344 * are resumed.
1345 */
1346 void
1347 cpustop_handler(void)
1348 {
1349 int cpu = PCPU_GET(cpuid);
1350 int cpumask = PCPU_GET(cpumask);
1351
1352 savectx(&stoppcbs[cpu]);
1353
1354 /* Indicate that we are stopped */
1355 atomic_set_int(&stopped_cpus, cpumask);
1356
1357 /* Wait for restart */
1358 while (!(started_cpus & cpumask))
1359 ia32_pause();
1360
1361 atomic_clear_int(&started_cpus, cpumask);
1362 atomic_clear_int(&stopped_cpus, cpumask);
1363
1364 if (cpu == 0 && cpustop_restartfunc != NULL) {
1365 cpustop_restartfunc();
1366 cpustop_restartfunc = NULL;
1367 }
1368 }
1369
1370 /*
1371 * This is called once the rest of the system is up and running and we're
1372 * ready to let the AP's out of the pen.
1373 */
1374 static void
1375 release_aps(void *dummy __unused)
1376 {
1377
1378 if (mp_ncpus == 1)
1379 return;
1380 atomic_store_rel_int(&aps_ready, 1);
1381 while (smp_started == 0)
1382 ia32_pause();
1383 }
1384 SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL);
1385
1386 static int
1387 sysctl_hlt_cpus(SYSCTL_HANDLER_ARGS)
1388 {
1389 u_int mask;
1390 int error;
1391
1392 mask = hlt_cpus_mask;
1393 error = sysctl_handle_int(oidp, &mask, 0, req);
1394 if (error || !req->newptr)
1395 return (error);
1396
1397 if (logical_cpus_mask != 0 &&
1398 (mask & logical_cpus_mask) == logical_cpus_mask)
1399 hlt_logical_cpus = 1;
1400 else
1401 hlt_logical_cpus = 0;
1402
1403 if (! hyperthreading_allowed)
1404 mask |= hyperthreading_cpus_mask;
1405
1406 if ((mask & all_cpus) == all_cpus)
1407 mask &= ~(1<<0);
1408 hlt_cpus_mask = mask;
1409 return (error);
1410 }
1411 SYSCTL_PROC(_machdep, OID_AUTO, hlt_cpus, CTLTYPE_INT|CTLFLAG_RW,
1412 0, 0, sysctl_hlt_cpus, "IU",
1413 "Bitmap of CPUs to halt. 101 (binary) will halt CPUs 0 and 2.");
1414
1415 static int
1416 sysctl_hlt_logical_cpus(SYSCTL_HANDLER_ARGS)
1417 {
1418 int disable, error;
1419
1420 disable = hlt_logical_cpus;
1421 error = sysctl_handle_int(oidp, &disable, 0, req);
1422 if (error || !req->newptr)
1423 return (error);
1424
1425 if (disable)
1426 hlt_cpus_mask |= logical_cpus_mask;
1427 else
1428 hlt_cpus_mask &= ~logical_cpus_mask;
1429
1430 if (! hyperthreading_allowed)
1431 hlt_cpus_mask |= hyperthreading_cpus_mask;
1432
1433 if ((hlt_cpus_mask & all_cpus) == all_cpus)
1434 hlt_cpus_mask &= ~(1<<0);
1435
1436 hlt_logical_cpus = disable;
1437 return (error);
1438 }
1439
1440 static int
1441 sysctl_hyperthreading_allowed(SYSCTL_HANDLER_ARGS)
1442 {
1443 int allowed, error;
1444
1445 allowed = hyperthreading_allowed;
1446 error = sysctl_handle_int(oidp, &allowed, 0, req);
1447 if (error || !req->newptr)
1448 return (error);
1449
1450 #ifdef SCHED_ULE
1451 /*
1452 * SCHED_ULE doesn't allow enabling/disabling HT cores at
1453 * run-time.
1454 */
1455 if (allowed != hyperthreading_allowed)
1456 return (ENOTSUP);
1457 return (error);
1458 #endif
1459
1460 if (allowed)
1461 hlt_cpus_mask &= ~hyperthreading_cpus_mask;
1462 else
1463 hlt_cpus_mask |= hyperthreading_cpus_mask;
1464
1465 if (logical_cpus_mask != 0 &&
1466 (hlt_cpus_mask & logical_cpus_mask) == logical_cpus_mask)
1467 hlt_logical_cpus = 1;
1468 else
1469 hlt_logical_cpus = 0;
1470
1471 if ((hlt_cpus_mask & all_cpus) == all_cpus)
1472 hlt_cpus_mask &= ~(1<<0);
1473
1474 hyperthreading_allowed = allowed;
1475 return (error);
1476 }
1477
1478 static void
1479 cpu_hlt_setup(void *dummy __unused)
1480 {
1481
1482 if (logical_cpus_mask != 0) {
1483 TUNABLE_INT_FETCH("machdep.hlt_logical_cpus",
1484 &hlt_logical_cpus);
1485 sysctl_ctx_init(&logical_cpu_clist);
1486 SYSCTL_ADD_PROC(&logical_cpu_clist,
1487 SYSCTL_STATIC_CHILDREN(_machdep), OID_AUTO,
1488 "hlt_logical_cpus", CTLTYPE_INT|CTLFLAG_RW, 0, 0,
1489 sysctl_hlt_logical_cpus, "IU", "");
1490 SYSCTL_ADD_UINT(&logical_cpu_clist,
1491 SYSCTL_STATIC_CHILDREN(_machdep), OID_AUTO,
1492 "logical_cpus_mask", CTLTYPE_INT|CTLFLAG_RD,
1493 &logical_cpus_mask, 0, "");
1494
1495 if (hlt_logical_cpus)
1496 hlt_cpus_mask |= logical_cpus_mask;
1497
1498 /*
1499 * If necessary for security purposes, force
1500 * hyperthreading off, regardless of the value
1501 * of hlt_logical_cpus.
1502 */
1503 if (hyperthreading_cpus_mask) {
1504 SYSCTL_ADD_PROC(&logical_cpu_clist,
1505 SYSCTL_STATIC_CHILDREN(_machdep), OID_AUTO,
1506 "hyperthreading_allowed", CTLTYPE_INT|CTLFLAG_RW,
1507 0, 0, sysctl_hyperthreading_allowed, "IU", "");
1508 if (! hyperthreading_allowed)
1509 hlt_cpus_mask |= hyperthreading_cpus_mask;
1510 }
1511 }
1512 }
1513 SYSINIT(cpu_hlt, SI_SUB_SMP, SI_ORDER_ANY, cpu_hlt_setup, NULL);
1514
1515 int
1516 mp_grab_cpu_hlt(void)
1517 {
1518 u_int mask = PCPU_GET(cpumask);
1519 #ifdef MP_WATCHDOG
1520 u_int cpuid = PCPU_GET(cpuid);
1521 #endif
1522 int retval;
1523
1524 #ifdef MP_WATCHDOG
1525 ap_watchdog(cpuid);
1526 #endif
1527
1528 retval = mask & hlt_cpus_mask;
1529 while (mask & hlt_cpus_mask)
1530 __asm __volatile("sti; hlt" : : : "memory");
1531 return (retval);
1532 }
1533
1534 #ifdef COUNT_IPIS
1535 /*
1536 * Setup interrupt counters for IPI handlers.
1537 */
1538 static void
1539 mp_ipi_intrcnt(void *dummy)
1540 {
1541 char buf[64];
1542 int i;
1543
1544 for (i = 0; i < mp_maxid; i++) {
1545 if (CPU_ABSENT(i))
1546 continue;
1547 snprintf(buf, sizeof(buf), "cpu%d: invltlb", i);
1548 intrcnt_add(buf, &ipi_invltlb_counts[i]);
1549 snprintf(buf, sizeof(buf), "cpu%d: invlrng", i);
1550 intrcnt_add(buf, &ipi_invlrng_counts[i]);
1551 snprintf(buf, sizeof(buf), "cpu%d: invlpg", i);
1552 intrcnt_add(buf, &ipi_invlpg_counts[i]);
1553 snprintf(buf, sizeof(buf), "cpu%d: preempt", i);
1554 intrcnt_add(buf, &ipi_preempt_counts[i]);
1555 snprintf(buf, sizeof(buf), "cpu%d: ast", i);
1556 intrcnt_add(buf, &ipi_ast_counts[i]);
1557 snprintf(buf, sizeof(buf), "cpu%d: rendezvous", i);
1558 intrcnt_add(buf, &ipi_rendezvous_counts[i]);
1559 snprintf(buf, sizeof(buf), "cpu%d: lazypmap", i);
1560 intrcnt_add(buf, &ipi_lazypmap_counts[i]);
1561 }
1562 }
1563 SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL);
1564 #endif
Cache object: 10310e95488be3e3d0030f320ae2da73
|