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/9.0/sys/i386/i386/mp_machdep.c 223758 2011-07-04 12:04:52Z attilio $");
28
29 #include "opt_apic.h"
30 #include "opt_cpu.h"
31 #include "opt_kstack_pages.h"
32 #include "opt_pmap.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 #include <sys/cpuset.h>
54 #ifdef GPROF
55 #include <sys/gmon.h>
56 #endif
57 #include <sys/kernel.h>
58 #include <sys/ktr.h>
59 #include <sys/lock.h>
60 #include <sys/malloc.h>
61 #include <sys/memrange.h>
62 #include <sys/mutex.h>
63 #include <sys/pcpu.h>
64 #include <sys/proc.h>
65 #include <sys/sched.h>
66 #include <sys/smp.h>
67 #include <sys/sysctl.h>
68
69 #include <vm/vm.h>
70 #include <vm/vm_param.h>
71 #include <vm/pmap.h>
72 #include <vm/vm_kern.h>
73 #include <vm/vm_extern.h>
74
75 #include <x86/apicreg.h>
76 #include <machine/clock.h>
77 #include <machine/cputypes.h>
78 #include <x86/mca.h>
79 #include <machine/md_var.h>
80 #include <machine/pcb.h>
81 #include <machine/psl.h>
82 #include <machine/smp.h>
83 #include <machine/specialreg.h>
84
85 #define WARMBOOT_TARGET 0
86 #define WARMBOOT_OFF (KERNBASE + 0x0467)
87 #define WARMBOOT_SEG (KERNBASE + 0x0469)
88
89 #define CMOS_REG (0x70)
90 #define CMOS_DATA (0x71)
91 #define BIOS_RESET (0x0f)
92 #define BIOS_WARM (0x0a)
93
94 /*
95 * this code MUST be enabled here and in mpboot.s.
96 * it follows the very early stages of AP boot by placing values in CMOS ram.
97 * it NORMALLY will never be needed and thus the primitive method for enabling.
98 *
99 #define CHECK_POINTS
100 */
101
102 #if defined(CHECK_POINTS) && !defined(PC98)
103 #define CHECK_READ(A) (outb(CMOS_REG, (A)), inb(CMOS_DATA))
104 #define CHECK_WRITE(A,D) (outb(CMOS_REG, (A)), outb(CMOS_DATA, (D)))
105
106 #define CHECK_INIT(D); \
107 CHECK_WRITE(0x34, (D)); \
108 CHECK_WRITE(0x35, (D)); \
109 CHECK_WRITE(0x36, (D)); \
110 CHECK_WRITE(0x37, (D)); \
111 CHECK_WRITE(0x38, (D)); \
112 CHECK_WRITE(0x39, (D));
113
114 #define CHECK_PRINT(S); \
115 printf("%s: %d, %d, %d, %d, %d, %d\n", \
116 (S), \
117 CHECK_READ(0x34), \
118 CHECK_READ(0x35), \
119 CHECK_READ(0x36), \
120 CHECK_READ(0x37), \
121 CHECK_READ(0x38), \
122 CHECK_READ(0x39));
123
124 #else /* CHECK_POINTS */
125
126 #define CHECK_INIT(D)
127 #define CHECK_PRINT(S)
128 #define CHECK_WRITE(A, D)
129
130 #endif /* CHECK_POINTS */
131
132 /* lock region used by kernel profiling */
133 int mcount_lock;
134
135 int mp_naps; /* # of Applications processors */
136 int boot_cpu_id = -1; /* designated BSP */
137
138 extern struct pcpu __pcpu[];
139
140 /* AP uses this during bootstrap. Do not staticize. */
141 char *bootSTK;
142 static int bootAP;
143
144 /* Free these after use */
145 void *bootstacks[MAXCPU];
146 static void *dpcpu;
147
148 /* Hotwire a 0->4MB V==P mapping */
149 extern pt_entry_t *KPTphys;
150
151 struct pcb stoppcbs[MAXCPU];
152
153 /* Variables needed for SMP tlb shootdown. */
154 vm_offset_t smp_tlb_addr1;
155 vm_offset_t smp_tlb_addr2;
156 volatile int smp_tlb_wait;
157
158 #ifdef COUNT_IPIS
159 /* Interrupt counts. */
160 static u_long *ipi_preempt_counts[MAXCPU];
161 static u_long *ipi_ast_counts[MAXCPU];
162 u_long *ipi_invltlb_counts[MAXCPU];
163 u_long *ipi_invlrng_counts[MAXCPU];
164 u_long *ipi_invlpg_counts[MAXCPU];
165 u_long *ipi_invlcache_counts[MAXCPU];
166 u_long *ipi_rendezvous_counts[MAXCPU];
167 u_long *ipi_lazypmap_counts[MAXCPU];
168 static u_long *ipi_hardclock_counts[MAXCPU];
169 #endif
170
171 /*
172 * Local data and functions.
173 */
174
175 static volatile cpuset_t ipi_nmi_pending;
176
177 /* used to hold the AP's until we are ready to release them */
178 static struct mtx ap_boot_mtx;
179
180 /* Set to 1 once we're ready to let the APs out of the pen. */
181 static volatile int aps_ready = 0;
182
183 /*
184 * Store data from cpu_add() until later in the boot when we actually setup
185 * the APs.
186 */
187 struct cpu_info {
188 int cpu_present:1;
189 int cpu_bsp:1;
190 int cpu_disabled:1;
191 int cpu_hyperthread:1;
192 } static cpu_info[MAX_APIC_ID + 1];
193 int cpu_apic_ids[MAXCPU];
194 int apic_cpuids[MAX_APIC_ID + 1];
195
196 /* Holds pending bitmap based IPIs per CPU */
197 static volatile u_int cpu_ipi_pending[MAXCPU];
198
199 static u_int boot_address;
200 static int cpu_logical; /* logical cpus per core */
201 static int cpu_cores; /* cores per package */
202
203 static void assign_cpu_ids(void);
204 static void install_ap_tramp(void);
205 static void set_interrupt_apic_ids(void);
206 static int start_all_aps(void);
207 static int start_ap(int apic_id);
208 static void release_aps(void *dummy);
209
210 static u_int hyperthreading_cpus; /* logical cpus sharing L1 cache */
211 static int hyperthreading_allowed = 1;
212
213 static void
214 mem_range_AP_init(void)
215 {
216 if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP)
217 mem_range_softc.mr_op->initAP(&mem_range_softc);
218 }
219
220 static void
221 topo_probe_amd(void)
222 {
223 int core_id_bits;
224 int id;
225
226 /* AMD processors do not support HTT. */
227 cpu_logical = 1;
228
229 if ((amd_feature2 & AMDID2_CMP) == 0) {
230 cpu_cores = 1;
231 return;
232 }
233
234 core_id_bits = (cpu_procinfo2 & AMDID_COREID_SIZE) >>
235 AMDID_COREID_SIZE_SHIFT;
236 if (core_id_bits == 0) {
237 cpu_cores = (cpu_procinfo2 & AMDID_CMP_CORES) + 1;
238 return;
239 }
240
241 /* Fam 10h and newer should get here. */
242 for (id = 0; id <= MAX_APIC_ID; id++) {
243 /* Check logical CPU availability. */
244 if (!cpu_info[id].cpu_present || cpu_info[id].cpu_disabled)
245 continue;
246 /* Check if logical CPU has the same package ID. */
247 if ((id >> core_id_bits) != (boot_cpu_id >> core_id_bits))
248 continue;
249 cpu_cores++;
250 }
251 }
252
253 /*
254 * Round up to the next power of two, if necessary, and then
255 * take log2.
256 * Returns -1 if argument is zero.
257 */
258 static __inline int
259 mask_width(u_int x)
260 {
261
262 return (fls(x << (1 - powerof2(x))) - 1);
263 }
264
265 static void
266 topo_probe_0x4(void)
267 {
268 u_int p[4];
269 int pkg_id_bits;
270 int core_id_bits;
271 int max_cores;
272 int max_logical;
273 int id;
274
275 /* Both zero and one here mean one logical processor per package. */
276 max_logical = (cpu_feature & CPUID_HTT) != 0 ?
277 (cpu_procinfo & CPUID_HTT_CORES) >> 16 : 1;
278 if (max_logical <= 1)
279 return;
280
281 /*
282 * Because of uniformity assumption we examine only
283 * those logical processors that belong to the same
284 * package as BSP. Further, we count number of
285 * logical processors that belong to the same core
286 * as BSP thus deducing number of threads per core.
287 */
288 if (cpu_high >= 0x4) {
289 cpuid_count(0x04, 0, p);
290 max_cores = ((p[0] >> 26) & 0x3f) + 1;
291 } else
292 max_cores = 1;
293 core_id_bits = mask_width(max_logical/max_cores);
294 if (core_id_bits < 0)
295 return;
296 pkg_id_bits = core_id_bits + mask_width(max_cores);
297
298 for (id = 0; id <= MAX_APIC_ID; id++) {
299 /* Check logical CPU availability. */
300 if (!cpu_info[id].cpu_present || cpu_info[id].cpu_disabled)
301 continue;
302 /* Check if logical CPU has the same package ID. */
303 if ((id >> pkg_id_bits) != (boot_cpu_id >> pkg_id_bits))
304 continue;
305 cpu_cores++;
306 /* Check if logical CPU has the same package and core IDs. */
307 if ((id >> core_id_bits) == (boot_cpu_id >> core_id_bits))
308 cpu_logical++;
309 }
310
311 KASSERT(cpu_cores >= 1 && cpu_logical >= 1,
312 ("topo_probe_0x4 couldn't find BSP"));
313
314 cpu_cores /= cpu_logical;
315 hyperthreading_cpus = cpu_logical;
316 }
317
318 static void
319 topo_probe_0xb(void)
320 {
321 u_int p[4];
322 int bits;
323 int cnt;
324 int i;
325 int logical;
326 int type;
327 int x;
328
329 /* We only support three levels for now. */
330 for (i = 0; i < 3; i++) {
331 cpuid_count(0x0b, i, p);
332
333 /* Fall back if CPU leaf 11 doesn't really exist. */
334 if (i == 0 && p[1] == 0) {
335 topo_probe_0x4();
336 return;
337 }
338
339 bits = p[0] & 0x1f;
340 logical = p[1] &= 0xffff;
341 type = (p[2] >> 8) & 0xff;
342 if (type == 0 || logical == 0)
343 break;
344 /*
345 * Because of uniformity assumption we examine only
346 * those logical processors that belong to the same
347 * package as BSP.
348 */
349 for (cnt = 0, x = 0; x <= MAX_APIC_ID; x++) {
350 if (!cpu_info[x].cpu_present ||
351 cpu_info[x].cpu_disabled)
352 continue;
353 if (x >> bits == boot_cpu_id >> bits)
354 cnt++;
355 }
356 if (type == CPUID_TYPE_SMT)
357 cpu_logical = cnt;
358 else if (type == CPUID_TYPE_CORE)
359 cpu_cores = cnt;
360 }
361 if (cpu_logical == 0)
362 cpu_logical = 1;
363 cpu_cores /= cpu_logical;
364 }
365
366 /*
367 * Both topology discovery code and code that consumes topology
368 * information assume top-down uniformity of the topology.
369 * That is, all physical packages must be identical and each
370 * core in a package must have the same number of threads.
371 * Topology information is queried only on BSP, on which this
372 * code runs and for which it can query CPUID information.
373 * Then topology is extrapolated on all packages using the
374 * uniformity assumption.
375 */
376 static void
377 topo_probe(void)
378 {
379 static int cpu_topo_probed = 0;
380
381 if (cpu_topo_probed)
382 return;
383
384 CPU_ZERO(&logical_cpus_mask);
385 if (mp_ncpus <= 1)
386 cpu_cores = cpu_logical = 1;
387 else if (cpu_vendor_id == CPU_VENDOR_AMD)
388 topo_probe_amd();
389 else if (cpu_vendor_id == CPU_VENDOR_INTEL) {
390 /*
391 * See Intel(R) 64 Architecture Processor
392 * Topology Enumeration article for details.
393 *
394 * Note that 0x1 <= cpu_high < 4 case should be
395 * compatible with topo_probe_0x4() logic when
396 * CPUID.1:EBX[23:16] > 0 (cpu_cores will be 1)
397 * or it should trigger the fallback otherwise.
398 */
399 if (cpu_high >= 0xb)
400 topo_probe_0xb();
401 else if (cpu_high >= 0x1)
402 topo_probe_0x4();
403 }
404
405 /*
406 * Fallback: assume each logical CPU is in separate
407 * physical package. That is, no multi-core, no SMT.
408 */
409 if (cpu_cores == 0 || cpu_logical == 0)
410 cpu_cores = cpu_logical = 1;
411 cpu_topo_probed = 1;
412 }
413
414 struct cpu_group *
415 cpu_topo(void)
416 {
417 int cg_flags;
418
419 /*
420 * Determine whether any threading flags are
421 * necessry.
422 */
423 topo_probe();
424 if (cpu_logical > 1 && hyperthreading_cpus)
425 cg_flags = CG_FLAG_HTT;
426 else if (cpu_logical > 1)
427 cg_flags = CG_FLAG_SMT;
428 else
429 cg_flags = 0;
430 if (mp_ncpus % (cpu_cores * cpu_logical) != 0) {
431 printf("WARNING: Non-uniform processors.\n");
432 printf("WARNING: Using suboptimal topology.\n");
433 return (smp_topo_none());
434 }
435 /*
436 * No multi-core or hyper-threaded.
437 */
438 if (cpu_logical * cpu_cores == 1)
439 return (smp_topo_none());
440 /*
441 * Only HTT no multi-core.
442 */
443 if (cpu_logical > 1 && cpu_cores == 1)
444 return (smp_topo_1level(CG_SHARE_L1, cpu_logical, cg_flags));
445 /*
446 * Only multi-core no HTT.
447 */
448 if (cpu_cores > 1 && cpu_logical == 1)
449 return (smp_topo_1level(CG_SHARE_L2, cpu_cores, cg_flags));
450 /*
451 * Both HTT and multi-core.
452 */
453 return (smp_topo_2level(CG_SHARE_L2, cpu_cores,
454 CG_SHARE_L1, cpu_logical, cg_flags));
455 }
456
457
458 /*
459 * Calculate usable address in base memory for AP trampoline code.
460 */
461 u_int
462 mp_bootaddress(u_int basemem)
463 {
464
465 boot_address = trunc_page(basemem); /* round down to 4k boundary */
466 if ((basemem - boot_address) < bootMP_size)
467 boot_address -= PAGE_SIZE; /* not enough, lower by 4k */
468
469 return boot_address;
470 }
471
472 void
473 cpu_add(u_int apic_id, char boot_cpu)
474 {
475
476 if (apic_id > MAX_APIC_ID) {
477 panic("SMP: APIC ID %d too high", apic_id);
478 return;
479 }
480 KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %d added twice",
481 apic_id));
482 cpu_info[apic_id].cpu_present = 1;
483 if (boot_cpu) {
484 KASSERT(boot_cpu_id == -1,
485 ("CPU %d claims to be BSP, but CPU %d already is", apic_id,
486 boot_cpu_id));
487 boot_cpu_id = apic_id;
488 cpu_info[apic_id].cpu_bsp = 1;
489 }
490 if (mp_ncpus < MAXCPU) {
491 mp_ncpus++;
492 mp_maxid = mp_ncpus - 1;
493 }
494 if (bootverbose)
495 printf("SMP: Added CPU %d (%s)\n", apic_id, boot_cpu ? "BSP" :
496 "AP");
497 }
498
499 void
500 cpu_mp_setmaxid(void)
501 {
502
503 /*
504 * mp_maxid should be already set by calls to cpu_add().
505 * Just sanity check its value here.
506 */
507 if (mp_ncpus == 0)
508 KASSERT(mp_maxid == 0,
509 ("%s: mp_ncpus is zero, but mp_maxid is not", __func__));
510 else if (mp_ncpus == 1)
511 mp_maxid = 0;
512 else
513 KASSERT(mp_maxid >= mp_ncpus - 1,
514 ("%s: counters out of sync: max %d, count %d", __func__,
515 mp_maxid, mp_ncpus));
516 }
517
518 int
519 cpu_mp_probe(void)
520 {
521
522 /*
523 * Always record BSP in CPU map so that the mbuf init code works
524 * correctly.
525 */
526 CPU_SETOF(0, &all_cpus);
527 if (mp_ncpus == 0) {
528 /*
529 * No CPUs were found, so this must be a UP system. Setup
530 * the variables to represent a system with a single CPU
531 * with an id of 0.
532 */
533 mp_ncpus = 1;
534 return (0);
535 }
536
537 /* At least one CPU was found. */
538 if (mp_ncpus == 1) {
539 /*
540 * One CPU was found, so this must be a UP system with
541 * an I/O APIC.
542 */
543 mp_maxid = 0;
544 return (0);
545 }
546
547 /* At least two CPUs were found. */
548 return (1);
549 }
550
551 /*
552 * Initialize the IPI handlers and start up the AP's.
553 */
554 void
555 cpu_mp_start(void)
556 {
557 int i;
558
559 /* Initialize the logical ID to APIC ID table. */
560 for (i = 0; i < MAXCPU; i++) {
561 cpu_apic_ids[i] = -1;
562 cpu_ipi_pending[i] = 0;
563 }
564
565 /* Install an inter-CPU IPI for TLB invalidation */
566 setidt(IPI_INVLTLB, IDTVEC(invltlb),
567 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
568 setidt(IPI_INVLPG, IDTVEC(invlpg),
569 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
570 setidt(IPI_INVLRNG, IDTVEC(invlrng),
571 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
572
573 /* Install an inter-CPU IPI for cache invalidation. */
574 setidt(IPI_INVLCACHE, IDTVEC(invlcache),
575 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
576
577 /* Install an inter-CPU IPI for lazy pmap release */
578 setidt(IPI_LAZYPMAP, IDTVEC(lazypmap),
579 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
580
581 /* Install an inter-CPU IPI for all-CPU rendezvous */
582 setidt(IPI_RENDEZVOUS, IDTVEC(rendezvous),
583 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
584
585 /* Install generic inter-CPU IPI handler */
586 setidt(IPI_BITMAP_VECTOR, IDTVEC(ipi_intr_bitmap_handler),
587 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
588
589 /* Install an inter-CPU IPI for CPU stop/restart */
590 setidt(IPI_STOP, IDTVEC(cpustop),
591 SDT_SYS386IGT, SEL_KPL, GSEL(GCODE_SEL, SEL_KPL));
592
593
594 /* Set boot_cpu_id if needed. */
595 if (boot_cpu_id == -1) {
596 boot_cpu_id = PCPU_GET(apic_id);
597 cpu_info[boot_cpu_id].cpu_bsp = 1;
598 } else
599 KASSERT(boot_cpu_id == PCPU_GET(apic_id),
600 ("BSP's APIC ID doesn't match boot_cpu_id"));
601
602 /* Probe logical/physical core configuration. */
603 topo_probe();
604
605 assign_cpu_ids();
606
607 /* Start each Application Processor */
608 start_all_aps();
609
610 set_interrupt_apic_ids();
611 }
612
613
614 /*
615 * Print various information about the SMP system hardware and setup.
616 */
617 void
618 cpu_mp_announce(void)
619 {
620 const char *hyperthread;
621 int i;
622
623 printf("FreeBSD/SMP: %d package(s) x %d core(s)",
624 mp_ncpus / (cpu_cores * cpu_logical), cpu_cores);
625 if (hyperthreading_cpus > 1)
626 printf(" x %d HTT threads", cpu_logical);
627 else if (cpu_logical > 1)
628 printf(" x %d SMT threads", cpu_logical);
629 printf("\n");
630
631 /* List active CPUs first. */
632 printf(" cpu0 (BSP): APIC ID: %2d\n", boot_cpu_id);
633 for (i = 1; i < mp_ncpus; i++) {
634 if (cpu_info[cpu_apic_ids[i]].cpu_hyperthread)
635 hyperthread = "/HT";
636 else
637 hyperthread = "";
638 printf(" cpu%d (AP%s): APIC ID: %2d\n", i, hyperthread,
639 cpu_apic_ids[i]);
640 }
641
642 /* List disabled CPUs last. */
643 for (i = 0; i <= MAX_APIC_ID; i++) {
644 if (!cpu_info[i].cpu_present || !cpu_info[i].cpu_disabled)
645 continue;
646 if (cpu_info[i].cpu_hyperthread)
647 hyperthread = "/HT";
648 else
649 hyperthread = "";
650 printf(" cpu (AP%s): APIC ID: %2d (disabled)\n", hyperthread,
651 i);
652 }
653 }
654
655 /*
656 * AP CPU's call this to initialize themselves.
657 */
658 void
659 init_secondary(void)
660 {
661 struct pcpu *pc;
662 vm_offset_t addr;
663 int gsel_tss;
664 int x, myid;
665 u_int cpuid, cr0;
666
667 /* bootAP is set in start_ap() to our ID. */
668 myid = bootAP;
669
670 /* Get per-cpu data */
671 pc = &__pcpu[myid];
672
673 /* prime data page for it to use */
674 pcpu_init(pc, myid, sizeof(struct pcpu));
675 dpcpu_init(dpcpu, myid);
676 pc->pc_apic_id = cpu_apic_ids[myid];
677 pc->pc_prvspace = pc;
678 pc->pc_curthread = 0;
679
680 gdt_segs[GPRIV_SEL].ssd_base = (int) pc;
681 gdt_segs[GPROC0_SEL].ssd_base = (int) &pc->pc_common_tss;
682
683 for (x = 0; x < NGDT; x++) {
684 ssdtosd(&gdt_segs[x], &gdt[myid * NGDT + x].sd);
685 }
686
687 r_gdt.rd_limit = NGDT * sizeof(gdt[0]) - 1;
688 r_gdt.rd_base = (int) &gdt[myid * NGDT];
689 lgdt(&r_gdt); /* does magic intra-segment return */
690
691 lidt(&r_idt);
692
693 lldt(_default_ldt);
694 PCPU_SET(currentldt, _default_ldt);
695
696 gsel_tss = GSEL(GPROC0_SEL, SEL_KPL);
697 gdt[myid * NGDT + GPROC0_SEL].sd.sd_type = SDT_SYS386TSS;
698 PCPU_SET(common_tss.tss_esp0, 0); /* not used until after switch */
699 PCPU_SET(common_tss.tss_ss0, GSEL(GDATA_SEL, SEL_KPL));
700 PCPU_SET(common_tss.tss_ioopt, (sizeof (struct i386tss)) << 16);
701 PCPU_SET(tss_gdt, &gdt[myid * NGDT + GPROC0_SEL].sd);
702 PCPU_SET(common_tssd, *PCPU_GET(tss_gdt));
703 ltr(gsel_tss);
704
705 PCPU_SET(fsgs_gdt, &gdt[myid * NGDT + GUFS_SEL].sd);
706
707 /*
708 * Set to a known state:
709 * Set by mpboot.s: CR0_PG, CR0_PE
710 * Set by cpu_setregs: CR0_NE, CR0_MP, CR0_TS, CR0_WP, CR0_AM
711 */
712 cr0 = rcr0();
713 cr0 &= ~(CR0_CD | CR0_NW | CR0_EM);
714 load_cr0(cr0);
715 CHECK_WRITE(0x38, 5);
716
717 /* Disable local APIC just to be sure. */
718 lapic_disable();
719
720 /* signal our startup to the BSP. */
721 mp_naps++;
722 CHECK_WRITE(0x39, 6);
723
724 /* Spin until the BSP releases the AP's. */
725 while (!aps_ready)
726 ia32_pause();
727
728 /* BSP may have changed PTD while we were waiting */
729 invltlb();
730 for (addr = 0; addr < NKPT * NBPDR - 1; addr += PAGE_SIZE)
731 invlpg(addr);
732
733 #if defined(I586_CPU) && !defined(NO_F00F_HACK)
734 lidt(&r_idt);
735 #endif
736
737 /* Initialize the PAT MSR if present. */
738 pmap_init_pat();
739
740 /* set up CPU registers and state */
741 cpu_setregs();
742
743 /* set up FPU state on the AP */
744 npxinit();
745
746 /* set up SSE registers */
747 enable_sse();
748
749 #ifdef PAE
750 /* Enable the PTE no-execute bit. */
751 if ((amd_feature & AMDID_NX) != 0) {
752 uint64_t msr;
753
754 msr = rdmsr(MSR_EFER) | EFER_NXE;
755 wrmsr(MSR_EFER, msr);
756 }
757 #endif
758
759 /* A quick check from sanity claus */
760 cpuid = PCPU_GET(cpuid);
761 if (PCPU_GET(apic_id) != lapic_id()) {
762 printf("SMP: cpuid = %d\n", cpuid);
763 printf("SMP: actual apic_id = %d\n", lapic_id());
764 printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id));
765 panic("cpuid mismatch! boom!!");
766 }
767
768 /* Initialize curthread. */
769 KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread"));
770 PCPU_SET(curthread, PCPU_GET(idlethread));
771
772 mca_init();
773
774 mtx_lock_spin(&ap_boot_mtx);
775
776 /* Init local apic for irq's */
777 lapic_setup(1);
778
779 /* Set memory range attributes for this CPU to match the BSP */
780 mem_range_AP_init();
781
782 smp_cpus++;
783
784 CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid);
785 printf("SMP: AP CPU #%d Launched!\n", cpuid);
786
787 /* Determine if we are a logical CPU. */
788 /* XXX Calculation depends on cpu_logical being a power of 2, e.g. 2 */
789 if (cpu_logical > 1 && PCPU_GET(apic_id) % cpu_logical != 0)
790 CPU_SET(cpuid, &logical_cpus_mask);
791
792 if (bootverbose)
793 lapic_dump("AP");
794
795 if (smp_cpus == mp_ncpus) {
796 /* enable IPI's, tlb shootdown, freezes etc */
797 atomic_store_rel_int(&smp_started, 1);
798 smp_active = 1; /* historic */
799 }
800
801 mtx_unlock_spin(&ap_boot_mtx);
802
803 /* Wait until all the AP's are up. */
804 while (smp_started == 0)
805 ia32_pause();
806
807 /* Start per-CPU event timers. */
808 cpu_initclocks_ap();
809
810 /* Enter the scheduler. */
811 sched_throw(NULL);
812
813 panic("scheduler returned us to %s", __func__);
814 /* NOTREACHED */
815 }
816
817 /*******************************************************************
818 * local functions and data
819 */
820
821 /*
822 * We tell the I/O APIC code about all the CPUs we want to receive
823 * interrupts. If we don't want certain CPUs to receive IRQs we
824 * can simply not tell the I/O APIC code about them in this function.
825 * We also do not tell it about the BSP since it tells itself about
826 * the BSP internally to work with UP kernels and on UP machines.
827 */
828 static void
829 set_interrupt_apic_ids(void)
830 {
831 u_int i, apic_id;
832
833 for (i = 0; i < MAXCPU; i++) {
834 apic_id = cpu_apic_ids[i];
835 if (apic_id == -1)
836 continue;
837 if (cpu_info[apic_id].cpu_bsp)
838 continue;
839 if (cpu_info[apic_id].cpu_disabled)
840 continue;
841
842 /* Don't let hyperthreads service interrupts. */
843 if (hyperthreading_cpus > 1 &&
844 apic_id % hyperthreading_cpus != 0)
845 continue;
846
847 intr_add_cpu(i);
848 }
849 }
850
851 /*
852 * Assign logical CPU IDs to local APICs.
853 */
854 static void
855 assign_cpu_ids(void)
856 {
857 u_int i;
858
859 TUNABLE_INT_FETCH("machdep.hyperthreading_allowed",
860 &hyperthreading_allowed);
861
862 /* Check for explicitly disabled CPUs. */
863 for (i = 0; i <= MAX_APIC_ID; i++) {
864 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp)
865 continue;
866
867 if (hyperthreading_cpus > 1 && i % hyperthreading_cpus != 0) {
868 cpu_info[i].cpu_hyperthread = 1;
869
870 /*
871 * Don't use HT CPU if it has been disabled by a
872 * tunable.
873 */
874 if (hyperthreading_allowed == 0) {
875 cpu_info[i].cpu_disabled = 1;
876 continue;
877 }
878 }
879
880 /* Don't use this CPU if it has been disabled by a tunable. */
881 if (resource_disabled("lapic", i)) {
882 cpu_info[i].cpu_disabled = 1;
883 continue;
884 }
885 }
886
887 if (hyperthreading_allowed == 0 && hyperthreading_cpus > 1) {
888 hyperthreading_cpus = 0;
889 cpu_logical = 1;
890 }
891
892 /*
893 * Assign CPU IDs to local APIC IDs and disable any CPUs
894 * beyond MAXCPU. CPU 0 is always assigned to the BSP.
895 *
896 * To minimize confusion for userland, we attempt to number
897 * CPUs such that all threads and cores in a package are
898 * grouped together. For now we assume that the BSP is always
899 * the first thread in a package and just start adding APs
900 * starting with the BSP's APIC ID.
901 */
902 mp_ncpus = 1;
903 cpu_apic_ids[0] = boot_cpu_id;
904 apic_cpuids[boot_cpu_id] = 0;
905 for (i = boot_cpu_id + 1; i != boot_cpu_id;
906 i == MAX_APIC_ID ? i = 0 : i++) {
907 if (!cpu_info[i].cpu_present || cpu_info[i].cpu_bsp ||
908 cpu_info[i].cpu_disabled)
909 continue;
910
911 if (mp_ncpus < MAXCPU) {
912 cpu_apic_ids[mp_ncpus] = i;
913 apic_cpuids[i] = mp_ncpus;
914 mp_ncpus++;
915 } else
916 cpu_info[i].cpu_disabled = 1;
917 }
918 KASSERT(mp_maxid >= mp_ncpus - 1,
919 ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid,
920 mp_ncpus));
921 }
922
923 /*
924 * start each AP in our list
925 */
926 /* Lowest 1MB is already mapped: don't touch*/
927 #define TMPMAP_START 1
928 static int
929 start_all_aps(void)
930 {
931 #ifndef PC98
932 u_char mpbiosreason;
933 #endif
934 uintptr_t kptbase;
935 u_int32_t mpbioswarmvec;
936 int apic_id, cpu, i;
937
938 mtx_init(&ap_boot_mtx, "ap boot", NULL, MTX_SPIN);
939
940 /* install the AP 1st level boot code */
941 install_ap_tramp();
942
943 /* save the current value of the warm-start vector */
944 mpbioswarmvec = *((u_int32_t *) WARMBOOT_OFF);
945 #ifndef PC98
946 outb(CMOS_REG, BIOS_RESET);
947 mpbiosreason = inb(CMOS_DATA);
948 #endif
949
950 /* set up temporary P==V mapping for AP boot */
951 /* XXX this is a hack, we should boot the AP on its own stack/PTD */
952
953 kptbase = (uintptr_t)(void *)KPTphys;
954 for (i = TMPMAP_START; i < NKPT; i++)
955 PTD[i] = (pd_entry_t)(PG_V | PG_RW |
956 ((kptbase + i * PAGE_SIZE) & PG_FRAME));
957 invltlb();
958
959 /* start each AP */
960 for (cpu = 1; cpu < mp_ncpus; cpu++) {
961 apic_id = cpu_apic_ids[cpu];
962
963 /* allocate and set up a boot stack data page */
964 bootstacks[cpu] =
965 (char *)kmem_alloc(kernel_map, KSTACK_PAGES * PAGE_SIZE);
966 dpcpu = (void *)kmem_alloc(kernel_map, DPCPU_SIZE);
967 /* setup a vector to our boot code */
968 *((volatile u_short *) WARMBOOT_OFF) = WARMBOOT_TARGET;
969 *((volatile u_short *) WARMBOOT_SEG) = (boot_address >> 4);
970 #ifndef PC98
971 outb(CMOS_REG, BIOS_RESET);
972 outb(CMOS_DATA, BIOS_WARM); /* 'warm-start' */
973 #endif
974
975 bootSTK = (char *)bootstacks[cpu] + KSTACK_PAGES * PAGE_SIZE - 4;
976 bootAP = cpu;
977
978 /* attempt to start the Application Processor */
979 CHECK_INIT(99); /* setup checkpoints */
980 if (!start_ap(apic_id)) {
981 printf("AP #%d (PHY# %d) failed!\n", cpu, apic_id);
982 CHECK_PRINT("trace"); /* show checkpoints */
983 /* better panic as the AP may be running loose */
984 printf("panic y/n? [y] ");
985 if (cngetc() != 'n')
986 panic("bye-bye");
987 }
988 CHECK_PRINT("trace"); /* show checkpoints */
989
990 CPU_SET(cpu, &all_cpus); /* record AP in CPU map */
991 }
992
993 /* restore the warmstart vector */
994 *(u_int32_t *) WARMBOOT_OFF = mpbioswarmvec;
995
996 #ifndef PC98
997 outb(CMOS_REG, BIOS_RESET);
998 outb(CMOS_DATA, mpbiosreason);
999 #endif
1000
1001 /* Undo V==P hack from above */
1002 for (i = TMPMAP_START; i < NKPT; i++)
1003 PTD[i] = 0;
1004 pmap_invalidate_range(kernel_pmap, 0, NKPT * NBPDR - 1);
1005
1006 /* number of APs actually started */
1007 return mp_naps;
1008 }
1009
1010 /*
1011 * load the 1st level AP boot code into base memory.
1012 */
1013
1014 /* targets for relocation */
1015 extern void bigJump(void);
1016 extern void bootCodeSeg(void);
1017 extern void bootDataSeg(void);
1018 extern void MPentry(void);
1019 extern u_int MP_GDT;
1020 extern u_int mp_gdtbase;
1021
1022 static void
1023 install_ap_tramp(void)
1024 {
1025 int x;
1026 int size = *(int *) ((u_long) & bootMP_size);
1027 vm_offset_t va = boot_address + KERNBASE;
1028 u_char *src = (u_char *) ((u_long) bootMP);
1029 u_char *dst = (u_char *) va;
1030 u_int boot_base = (u_int) bootMP;
1031 u_int8_t *dst8;
1032 u_int16_t *dst16;
1033 u_int32_t *dst32;
1034
1035 KASSERT (size <= PAGE_SIZE,
1036 ("'size' do not fit into PAGE_SIZE, as expected."));
1037 pmap_kenter(va, boot_address);
1038 pmap_invalidate_page (kernel_pmap, va);
1039 for (x = 0; x < size; ++x)
1040 *dst++ = *src++;
1041
1042 /*
1043 * modify addresses in code we just moved to basemem. unfortunately we
1044 * need fairly detailed info about mpboot.s for this to work. changes
1045 * to mpboot.s might require changes here.
1046 */
1047
1048 /* boot code is located in KERNEL space */
1049 dst = (u_char *) va;
1050
1051 /* modify the lgdt arg */
1052 dst32 = (u_int32_t *) (dst + ((u_int) & mp_gdtbase - boot_base));
1053 *dst32 = boot_address + ((u_int) & MP_GDT - boot_base);
1054
1055 /* modify the ljmp target for MPentry() */
1056 dst32 = (u_int32_t *) (dst + ((u_int) bigJump - boot_base) + 1);
1057 *dst32 = ((u_int) MPentry - KERNBASE);
1058
1059 /* modify the target for boot code segment */
1060 dst16 = (u_int16_t *) (dst + ((u_int) bootCodeSeg - boot_base));
1061 dst8 = (u_int8_t *) (dst16 + 1);
1062 *dst16 = (u_int) boot_address & 0xffff;
1063 *dst8 = ((u_int) boot_address >> 16) & 0xff;
1064
1065 /* modify the target for boot data segment */
1066 dst16 = (u_int16_t *) (dst + ((u_int) bootDataSeg - boot_base));
1067 dst8 = (u_int8_t *) (dst16 + 1);
1068 *dst16 = (u_int) boot_address & 0xffff;
1069 *dst8 = ((u_int) boot_address >> 16) & 0xff;
1070 }
1071
1072 /*
1073 * This function starts the AP (application processor) identified
1074 * by the APIC ID 'physicalCpu'. It does quite a "song and dance"
1075 * to accomplish this. This is necessary because of the nuances
1076 * of the different hardware we might encounter. It isn't pretty,
1077 * but it seems to work.
1078 */
1079 static int
1080 start_ap(int apic_id)
1081 {
1082 int vector, ms;
1083 int cpus;
1084
1085 /* calculate the vector */
1086 vector = (boot_address >> 12) & 0xff;
1087
1088 /* used as a watchpoint to signal AP startup */
1089 cpus = mp_naps;
1090
1091 /*
1092 * first we do an INIT/RESET IPI this INIT IPI might be run, reseting
1093 * and running the target CPU. OR this INIT IPI might be latched (P5
1094 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be
1095 * ignored.
1096 */
1097
1098 /* do an INIT IPI: assert RESET */
1099 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1100 APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id);
1101
1102 /* wait for pending status end */
1103 lapic_ipi_wait(-1);
1104
1105 /* do an INIT IPI: deassert RESET */
1106 lapic_ipi_raw(APIC_DEST_ALLESELF | APIC_TRIGMOD_LEVEL |
1107 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, 0);
1108
1109 /* wait for pending status end */
1110 DELAY(10000); /* wait ~10mS */
1111 lapic_ipi_wait(-1);
1112
1113 /*
1114 * next we do a STARTUP IPI: the previous INIT IPI might still be
1115 * latched, (P5 bug) this 1st STARTUP would then terminate
1116 * immediately, and the previously started INIT IPI would continue. OR
1117 * the previous INIT IPI has already run. and this STARTUP IPI will
1118 * run. OR the previous INIT IPI was ignored. and this STARTUP IPI
1119 * will run.
1120 */
1121
1122 /* do a STARTUP IPI */
1123 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1124 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1125 vector, apic_id);
1126 lapic_ipi_wait(-1);
1127 DELAY(200); /* wait ~200uS */
1128
1129 /*
1130 * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF
1131 * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR
1132 * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is
1133 * recognized after hardware RESET or INIT IPI.
1134 */
1135
1136 lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE |
1137 APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP |
1138 vector, apic_id);
1139 lapic_ipi_wait(-1);
1140 DELAY(200); /* wait ~200uS */
1141
1142 /* Wait up to 5 seconds for it to start. */
1143 for (ms = 0; ms < 5000; ms++) {
1144 if (mp_naps > cpus)
1145 return 1; /* return SUCCESS */
1146 DELAY(1000);
1147 }
1148 return 0; /* return FAILURE */
1149 }
1150
1151 #ifdef COUNT_XINVLTLB_HITS
1152 u_int xhits_gbl[MAXCPU];
1153 u_int xhits_pg[MAXCPU];
1154 u_int xhits_rng[MAXCPU];
1155 SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW, 0, "");
1156 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl,
1157 sizeof(xhits_gbl), "IU", "");
1158 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg,
1159 sizeof(xhits_pg), "IU", "");
1160 SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng,
1161 sizeof(xhits_rng), "IU", "");
1162
1163 u_int ipi_global;
1164 u_int ipi_page;
1165 u_int ipi_range;
1166 u_int ipi_range_size;
1167 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, "");
1168 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, "");
1169 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, "");
1170 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW, &ipi_range_size,
1171 0, "");
1172
1173 u_int ipi_masked_global;
1174 u_int ipi_masked_page;
1175 u_int ipi_masked_range;
1176 u_int ipi_masked_range_size;
1177 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_masked_global, CTLFLAG_RW,
1178 &ipi_masked_global, 0, "");
1179 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_masked_page, CTLFLAG_RW,
1180 &ipi_masked_page, 0, "");
1181 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_masked_range, CTLFLAG_RW,
1182 &ipi_masked_range, 0, "");
1183 SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_masked_range_size, CTLFLAG_RW,
1184 &ipi_masked_range_size, 0, "");
1185 #endif /* COUNT_XINVLTLB_HITS */
1186
1187 /*
1188 * Send an IPI to specified CPU handling the bitmap logic.
1189 */
1190 static void
1191 ipi_send_cpu(int cpu, u_int ipi)
1192 {
1193 u_int bitmap, old_pending, new_pending;
1194
1195 KASSERT(cpu_apic_ids[cpu] != -1, ("IPI to non-existent CPU %d", cpu));
1196
1197 if (IPI_IS_BITMAPED(ipi)) {
1198 bitmap = 1 << ipi;
1199 ipi = IPI_BITMAP_VECTOR;
1200 do {
1201 old_pending = cpu_ipi_pending[cpu];
1202 new_pending = old_pending | bitmap;
1203 } while (!atomic_cmpset_int(&cpu_ipi_pending[cpu],
1204 old_pending, new_pending));
1205 if (old_pending)
1206 return;
1207 }
1208 lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]);
1209 }
1210
1211 /*
1212 * Flush the TLB on all other CPU's
1213 */
1214 static void
1215 smp_tlb_shootdown(u_int vector, vm_offset_t addr1, vm_offset_t addr2)
1216 {
1217 u_int ncpu;
1218
1219 ncpu = mp_ncpus - 1; /* does not shootdown self */
1220 if (ncpu < 1)
1221 return; /* no other cpus */
1222 if (!(read_eflags() & PSL_I))
1223 panic("%s: interrupts disabled", __func__);
1224 mtx_lock_spin(&smp_ipi_mtx);
1225 smp_tlb_addr1 = addr1;
1226 smp_tlb_addr2 = addr2;
1227 atomic_store_rel_int(&smp_tlb_wait, 0);
1228 ipi_all_but_self(vector);
1229 while (smp_tlb_wait < ncpu)
1230 ia32_pause();
1231 mtx_unlock_spin(&smp_ipi_mtx);
1232 }
1233
1234 static void
1235 smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, vm_offset_t addr1, vm_offset_t addr2)
1236 {
1237 int cpu, ncpu, othercpus;
1238
1239 othercpus = mp_ncpus - 1;
1240 if (CPU_ISFULLSET(&mask)) {
1241 if (othercpus < 1)
1242 return;
1243 } else {
1244 CPU_CLR(PCPU_GET(cpuid), &mask);
1245 if (CPU_EMPTY(&mask))
1246 return;
1247 }
1248 if (!(read_eflags() & PSL_I))
1249 panic("%s: interrupts disabled", __func__);
1250 mtx_lock_spin(&smp_ipi_mtx);
1251 smp_tlb_addr1 = addr1;
1252 smp_tlb_addr2 = addr2;
1253 atomic_store_rel_int(&smp_tlb_wait, 0);
1254 if (CPU_ISFULLSET(&mask)) {
1255 ncpu = othercpus;
1256 ipi_all_but_self(vector);
1257 } else {
1258 ncpu = 0;
1259 while ((cpu = cpusetobj_ffs(&mask)) != 0) {
1260 cpu--;
1261 CPU_CLR(cpu, &mask);
1262 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu,
1263 vector);
1264 ipi_send_cpu(cpu, vector);
1265 ncpu++;
1266 }
1267 }
1268 while (smp_tlb_wait < ncpu)
1269 ia32_pause();
1270 mtx_unlock_spin(&smp_ipi_mtx);
1271 }
1272
1273 void
1274 smp_cache_flush(void)
1275 {
1276
1277 if (smp_started)
1278 smp_tlb_shootdown(IPI_INVLCACHE, 0, 0);
1279 }
1280
1281 void
1282 smp_invltlb(void)
1283 {
1284
1285 if (smp_started) {
1286 smp_tlb_shootdown(IPI_INVLTLB, 0, 0);
1287 #ifdef COUNT_XINVLTLB_HITS
1288 ipi_global++;
1289 #endif
1290 }
1291 }
1292
1293 void
1294 smp_invlpg(vm_offset_t addr)
1295 {
1296
1297 if (smp_started) {
1298 smp_tlb_shootdown(IPI_INVLPG, addr, 0);
1299 #ifdef COUNT_XINVLTLB_HITS
1300 ipi_page++;
1301 #endif
1302 }
1303 }
1304
1305 void
1306 smp_invlpg_range(vm_offset_t addr1, vm_offset_t addr2)
1307 {
1308
1309 if (smp_started) {
1310 smp_tlb_shootdown(IPI_INVLRNG, addr1, addr2);
1311 #ifdef COUNT_XINVLTLB_HITS
1312 ipi_range++;
1313 ipi_range_size += (addr2 - addr1) / PAGE_SIZE;
1314 #endif
1315 }
1316 }
1317
1318 void
1319 smp_masked_invltlb(cpuset_t mask)
1320 {
1321
1322 if (smp_started) {
1323 smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, 0, 0);
1324 #ifdef COUNT_XINVLTLB_HITS
1325 ipi_masked_global++;
1326 #endif
1327 }
1328 }
1329
1330 void
1331 smp_masked_invlpg(cpuset_t mask, vm_offset_t addr)
1332 {
1333
1334 if (smp_started) {
1335 smp_targeted_tlb_shootdown(mask, IPI_INVLPG, addr, 0);
1336 #ifdef COUNT_XINVLTLB_HITS
1337 ipi_masked_page++;
1338 #endif
1339 }
1340 }
1341
1342 void
1343 smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2)
1344 {
1345
1346 if (smp_started) {
1347 smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, addr1, addr2);
1348 #ifdef COUNT_XINVLTLB_HITS
1349 ipi_masked_range++;
1350 ipi_masked_range_size += (addr2 - addr1) / PAGE_SIZE;
1351 #endif
1352 }
1353 }
1354
1355 void
1356 ipi_bitmap_handler(struct trapframe frame)
1357 {
1358 struct trapframe *oldframe;
1359 struct thread *td;
1360 int cpu = PCPU_GET(cpuid);
1361 u_int ipi_bitmap;
1362
1363 critical_enter();
1364 td = curthread;
1365 td->td_intr_nesting_level++;
1366 oldframe = td->td_intr_frame;
1367 td->td_intr_frame = &frame;
1368 ipi_bitmap = atomic_readandclear_int(&cpu_ipi_pending[cpu]);
1369 if (ipi_bitmap & (1 << IPI_PREEMPT)) {
1370 #ifdef COUNT_IPIS
1371 (*ipi_preempt_counts[cpu])++;
1372 #endif
1373 sched_preempt(td);
1374 }
1375 if (ipi_bitmap & (1 << IPI_AST)) {
1376 #ifdef COUNT_IPIS
1377 (*ipi_ast_counts[cpu])++;
1378 #endif
1379 /* Nothing to do for AST */
1380 }
1381 if (ipi_bitmap & (1 << IPI_HARDCLOCK)) {
1382 #ifdef COUNT_IPIS
1383 (*ipi_hardclock_counts[cpu])++;
1384 #endif
1385 hardclockintr();
1386 }
1387 td->td_intr_frame = oldframe;
1388 td->td_intr_nesting_level--;
1389 critical_exit();
1390 }
1391
1392 /*
1393 * send an IPI to a set of cpus.
1394 */
1395 void
1396 ipi_selected(cpuset_t cpus, u_int ipi)
1397 {
1398 int cpu;
1399
1400 /*
1401 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1402 * of help in order to understand what is the source.
1403 * Set the mask of receiving CPUs for this purpose.
1404 */
1405 if (ipi == IPI_STOP_HARD)
1406 CPU_OR_ATOMIC(&ipi_nmi_pending, &cpus);
1407
1408 while ((cpu = cpusetobj_ffs(&cpus)) != 0) {
1409 cpu--;
1410 CPU_CLR(cpu, &cpus);
1411 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1412 ipi_send_cpu(cpu, ipi);
1413 }
1414 }
1415
1416 /*
1417 * send an IPI to a specific CPU.
1418 */
1419 void
1420 ipi_cpu(int cpu, u_int ipi)
1421 {
1422
1423 /*
1424 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1425 * of help in order to understand what is the source.
1426 * Set the mask of receiving CPUs for this purpose.
1427 */
1428 if (ipi == IPI_STOP_HARD)
1429 CPU_SET_ATOMIC(cpu, &ipi_nmi_pending);
1430
1431 CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi);
1432 ipi_send_cpu(cpu, ipi);
1433 }
1434
1435 /*
1436 * send an IPI to all CPUs EXCEPT myself
1437 */
1438 void
1439 ipi_all_but_self(u_int ipi)
1440 {
1441 cpuset_t other_cpus;
1442
1443 other_cpus = all_cpus;
1444 CPU_CLR(PCPU_GET(cpuid), &other_cpus);
1445 if (IPI_IS_BITMAPED(ipi)) {
1446 ipi_selected(other_cpus, ipi);
1447 return;
1448 }
1449
1450 /*
1451 * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit
1452 * of help in order to understand what is the source.
1453 * Set the mask of receiving CPUs for this purpose.
1454 */
1455 if (ipi == IPI_STOP_HARD)
1456 CPU_OR_ATOMIC(&ipi_nmi_pending, &other_cpus);
1457
1458 CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi);
1459 lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS);
1460 }
1461
1462 int
1463 ipi_nmi_handler()
1464 {
1465 u_int cpuid;
1466
1467 /*
1468 * As long as there is not a simple way to know about a NMI's
1469 * source, if the bitmask for the current CPU is present in
1470 * the global pending bitword an IPI_STOP_HARD has been issued
1471 * and should be handled.
1472 */
1473 cpuid = PCPU_GET(cpuid);
1474 if (!CPU_ISSET(cpuid, &ipi_nmi_pending))
1475 return (1);
1476
1477 CPU_CLR_ATOMIC(cpuid, &ipi_nmi_pending);
1478 cpustop_handler();
1479 return (0);
1480 }
1481
1482 /*
1483 * Handle an IPI_STOP by saving our current context and spinning until we
1484 * are resumed.
1485 */
1486 void
1487 cpustop_handler(void)
1488 {
1489 u_int cpu;
1490
1491 cpu = PCPU_GET(cpuid);
1492
1493 savectx(&stoppcbs[cpu]);
1494
1495 /* Indicate that we are stopped */
1496 CPU_SET_ATOMIC(cpu, &stopped_cpus);
1497
1498 /* Wait for restart */
1499 while (!CPU_ISSET(cpu, &started_cpus))
1500 ia32_pause();
1501
1502 CPU_CLR_ATOMIC(cpu, &started_cpus);
1503 CPU_CLR_ATOMIC(cpu, &stopped_cpus);
1504
1505 if (cpu == 0 && cpustop_restartfunc != NULL) {
1506 cpustop_restartfunc();
1507 cpustop_restartfunc = NULL;
1508 }
1509 }
1510
1511 /*
1512 * This is called once the rest of the system is up and running and we're
1513 * ready to let the AP's out of the pen.
1514 */
1515 static void
1516 release_aps(void *dummy __unused)
1517 {
1518
1519 if (mp_ncpus == 1)
1520 return;
1521 atomic_store_rel_int(&aps_ready, 1);
1522 while (smp_started == 0)
1523 ia32_pause();
1524 }
1525 SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL);
1526
1527 #ifdef COUNT_IPIS
1528 /*
1529 * Setup interrupt counters for IPI handlers.
1530 */
1531 static void
1532 mp_ipi_intrcnt(void *dummy)
1533 {
1534 char buf[64];
1535 int i;
1536
1537 CPU_FOREACH(i) {
1538 snprintf(buf, sizeof(buf), "cpu%d:invltlb", i);
1539 intrcnt_add(buf, &ipi_invltlb_counts[i]);
1540 snprintf(buf, sizeof(buf), "cpu%d:invlrng", i);
1541 intrcnt_add(buf, &ipi_invlrng_counts[i]);
1542 snprintf(buf, sizeof(buf), "cpu%d:invlpg", i);
1543 intrcnt_add(buf, &ipi_invlpg_counts[i]);
1544 snprintf(buf, sizeof(buf), "cpu%d:preempt", i);
1545 intrcnt_add(buf, &ipi_preempt_counts[i]);
1546 snprintf(buf, sizeof(buf), "cpu%d:ast", i);
1547 intrcnt_add(buf, &ipi_ast_counts[i]);
1548 snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i);
1549 intrcnt_add(buf, &ipi_rendezvous_counts[i]);
1550 snprintf(buf, sizeof(buf), "cpu%d:lazypmap", i);
1551 intrcnt_add(buf, &ipi_lazypmap_counts[i]);
1552 snprintf(buf, sizeof(buf), "cpu%d:hardclock", i);
1553 intrcnt_add(buf, &ipi_hardclock_counts[i]);
1554 }
1555 }
1556 SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL);
1557 #endif
Cache object: 69face5fd3b7a931bb584054ffa019f1
|