Print this page
9600 LDT still not happy under KPTI
| Split |
Close |
| Expand all |
| Collapse all |
--- old/usr/src/uts/i86pc/os/mp_startup.c
+++ new/usr/src/uts/i86pc/os/mp_startup.c
1 1 /*
2 2 * CDDL HEADER START
3 3 *
4 4 * The contents of this file are subject to the terms of the
5 5 * Common Development and Distribution License (the "License").
6 6 * You may not use this file except in compliance with the License.
7 7 *
8 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 9 * or http://www.opensolaris.org/os/licensing.
10 10 * See the License for the specific language governing permissions
11 11 * and limitations under the License.
12 12 *
13 13 * When distributing Covered Code, include this CDDL HEADER in each
14 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 15 * If applicable, add the following below this CDDL HEADER, with the
16 16 * fields enclosed by brackets "[]" replaced with your own identifying
17 17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 18 *
19 19 * CDDL HEADER END
20 20 */
21 21
22 22 /*
23 23 * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
24 24 */
25 25 /*
26 26 * Copyright (c) 2010, Intel Corporation.
27 27 * All rights reserved.
28 28 */
29 29 /*
30 30 * Copyright 2018 Joyent, Inc.
31 31 * Copyright 2013 Nexenta Systems, Inc. All rights reserved.
32 32 * Copyright 2018 OmniOS Community Edition (OmniOSce) Association.
33 33 */
34 34
35 35 #include <sys/types.h>
36 36 #include <sys/thread.h>
37 37 #include <sys/cpuvar.h>
38 38 #include <sys/cpu.h>
39 39 #include <sys/t_lock.h>
40 40 #include <sys/param.h>
41 41 #include <sys/proc.h>
42 42 #include <sys/disp.h>
43 43 #include <sys/class.h>
44 44 #include <sys/cmn_err.h>
45 45 #include <sys/debug.h>
46 46 #include <sys/note.h>
47 47 #include <sys/asm_linkage.h>
48 48 #include <sys/x_call.h>
49 49 #include <sys/systm.h>
50 50 #include <sys/var.h>
51 51 #include <sys/vtrace.h>
52 52 #include <vm/hat.h>
53 53 #include <vm/as.h>
54 54 #include <vm/seg_kmem.h>
55 55 #include <vm/seg_kp.h>
56 56 #include <sys/segments.h>
57 57 #include <sys/kmem.h>
58 58 #include <sys/stack.h>
59 59 #include <sys/smp_impldefs.h>
60 60 #include <sys/x86_archext.h>
61 61 #include <sys/machsystm.h>
62 62 #include <sys/traptrace.h>
63 63 #include <sys/clock.h>
64 64 #include <sys/cpc_impl.h>
65 65 #include <sys/pg.h>
66 66 #include <sys/cmt.h>
67 67 #include <sys/dtrace.h>
68 68 #include <sys/archsystm.h>
69 69 #include <sys/fp.h>
70 70 #include <sys/reboot.h>
71 71 #include <sys/kdi_machimpl.h>
72 72 #include <vm/hat_i86.h>
73 73 #include <vm/vm_dep.h>
74 74 #include <sys/memnode.h>
75 75 #include <sys/pci_cfgspace.h>
76 76 #include <sys/mach_mmu.h>
77 77 #include <sys/sysmacros.h>
78 78 #if defined(__xpv)
79 79 #include <sys/hypervisor.h>
80 80 #endif
81 81 #include <sys/cpu_module.h>
82 82 #include <sys/ontrap.h>
83 83
84 84 struct cpu cpus[1] __aligned(MMU_PAGESIZE);
85 85 struct cpu *cpu[NCPU] = {&cpus[0]};
86 86 struct cpu *cpu_free_list;
87 87 cpu_core_t cpu_core[NCPU];
88 88
89 89 #define cpu_next_free cpu_prev
90 90
91 91 /*
92 92 * Useful for disabling MP bring-up on a MP capable system.
93 93 */
94 94 int use_mp = 1;
95 95
96 96 /*
97 97 * to be set by a PSM to indicate what cpus
98 98 * are sitting around on the system.
99 99 */
100 100 cpuset_t mp_cpus;
101 101
102 102 /*
103 103 * This variable is used by the hat layer to decide whether or not
104 104 * critical sections are needed to prevent race conditions. For sun4m,
105 105 * this variable is set once enough MP initialization has been done in
106 106 * order to allow cross calls.
107 107 */
108 108 int flushes_require_xcalls;
109 109
110 110 cpuset_t cpu_ready_set; /* initialized in startup() */
111 111
112 112 static void mp_startup_boot(void);
113 113 static void mp_startup_hotplug(void);
114 114
115 115 static void cpu_sep_enable(void);
116 116 static void cpu_sep_disable(void);
117 117 static void cpu_asysc_enable(void);
118 118 static void cpu_asysc_disable(void);
119 119
120 120 /*
121 121 * Init CPU info - get CPU type info for processor_info system call.
122 122 */
123 123 void
124 124 init_cpu_info(struct cpu *cp)
125 125 {
126 126 processor_info_t *pi = &cp->cpu_type_info;
127 127
128 128 /*
129 129 * Get clock-frequency property for the CPU.
130 130 */
131 131 pi->pi_clock = cpu_freq;
132 132
133 133 /*
134 134 * Current frequency in Hz.
135 135 */
136 136 cp->cpu_curr_clock = cpu_freq_hz;
137 137
138 138 /*
139 139 * Supported frequencies.
140 140 */
141 141 if (cp->cpu_supp_freqs == NULL) {
142 142 cpu_set_supp_freqs(cp, NULL);
143 143 }
144 144
145 145 (void) strcpy(pi->pi_processor_type, "i386");
146 146 if (fpu_exists)
147 147 (void) strcpy(pi->pi_fputypes, "i387 compatible");
148 148
149 149 cp->cpu_idstr = kmem_zalloc(CPU_IDSTRLEN, KM_SLEEP);
150 150 cp->cpu_brandstr = kmem_zalloc(CPU_IDSTRLEN, KM_SLEEP);
151 151
152 152 /*
153 153 * If called for the BSP, cp is equal to current CPU.
154 154 * For non-BSPs, cpuid info of cp is not ready yet, so use cpuid info
155 155 * of current CPU as default values for cpu_idstr and cpu_brandstr.
156 156 * They will be corrected in mp_startup_common() after cpuid_pass1()
157 157 * has been invoked on target CPU.
158 158 */
159 159 (void) cpuid_getidstr(CPU, cp->cpu_idstr, CPU_IDSTRLEN);
160 160 (void) cpuid_getbrandstr(CPU, cp->cpu_brandstr, CPU_IDSTRLEN);
161 161 }
162 162
163 163 /*
164 164 * Configure syscall support on this CPU.
165 165 */
166 166 /*ARGSUSED*/
167 167 void
168 168 init_cpu_syscall(struct cpu *cp)
169 169 {
170 170 kpreempt_disable();
171 171
172 172 if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
173 173 is_x86_feature(x86_featureset, X86FSET_ASYSC)) {
174 174 uint64_t flags;
175 175
176 176 #if !defined(__xpv)
177 177 /*
178 178 * The syscall instruction imposes a certain ordering on
179 179 * segment selectors, so we double-check that ordering
180 180 * here.
181 181 */
182 182 CTASSERT(KDS_SEL == KCS_SEL + 8);
183 183 CTASSERT(UDS_SEL == U32CS_SEL + 8);
184 184 CTASSERT(UCS_SEL == U32CS_SEL + 16);
185 185 #endif
186 186
187 187 /*
188 188 * Turn syscall/sysret extensions on.
189 189 */
190 190 cpu_asysc_enable();
191 191
192 192 /*
193 193 * Program the magic registers ..
194 194 */
195 195 wrmsr(MSR_AMD_STAR,
196 196 ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32);
197 197 if (kpti_enable == 1) {
198 198 wrmsr(MSR_AMD_LSTAR,
199 199 (uint64_t)(uintptr_t)tr_sys_syscall);
200 200 wrmsr(MSR_AMD_CSTAR,
201 201 (uint64_t)(uintptr_t)tr_sys_syscall32);
202 202 } else {
203 203 wrmsr(MSR_AMD_LSTAR,
204 204 (uint64_t)(uintptr_t)sys_syscall);
205 205 wrmsr(MSR_AMD_CSTAR,
206 206 (uint64_t)(uintptr_t)sys_syscall32);
207 207 }
208 208
209 209 /*
210 210 * This list of flags is masked off the incoming
211 211 * %rfl when we enter the kernel.
212 212 */
213 213 flags = PS_IE | PS_T;
214 214 if (is_x86_feature(x86_featureset, X86FSET_SMAP) == B_TRUE)
215 215 flags |= PS_ACHK;
216 216 wrmsr(MSR_AMD_SFMASK, flags);
217 217 }
218 218
219 219 /*
220 220 * On 64-bit kernels on Nocona machines, the 32-bit syscall
221 221 * variant isn't available to 32-bit applications, but sysenter is.
222 222 */
223 223 if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
224 224 is_x86_feature(x86_featureset, X86FSET_SEP)) {
225 225
226 226 #if !defined(__xpv)
227 227 /*
228 228 * The sysenter instruction imposes a certain ordering on
229 229 * segment selectors, so we double-check that ordering
230 230 * here. See "sysenter" in Intel document 245471-012, "IA-32
231 231 * Intel Architecture Software Developer's Manual Volume 2:
232 232 * Instruction Set Reference"
233 233 */
234 234 CTASSERT(KDS_SEL == KCS_SEL + 8);
235 235
236 236 CTASSERT(U32CS_SEL == ((KCS_SEL + 16) | 3));
237 237 CTASSERT(UDS_SEL == U32CS_SEL + 8);
238 238 #endif
239 239
240 240 cpu_sep_enable();
241 241
242 242 /*
243 243 * resume() sets this value to the base of the threads stack
244 244 * via a context handler.
245 245 */
246 246 wrmsr(MSR_INTC_SEP_ESP, 0);
247 247
248 248 if (kpti_enable == 1) {
249 249 wrmsr(MSR_INTC_SEP_EIP,
250 250 (uint64_t)(uintptr_t)tr_sys_sysenter);
251 251 } else {
252 252 wrmsr(MSR_INTC_SEP_EIP,
253 253 (uint64_t)(uintptr_t)sys_sysenter);
254 254 }
255 255 }
256 256
257 257 kpreempt_enable();
258 258 }
259 259
260 260 #if !defined(__xpv)
261 261 /*
262 262 * Configure per-cpu ID GDT
263 263 */
264 264 static void
265 265 init_cpu_id_gdt(struct cpu *cp)
266 266 {
267 267 /* Write cpu_id into limit field of GDT for usermode retrieval */
268 268 #if defined(__amd64)
269 269 set_usegd(&cp->cpu_gdt[GDT_CPUID], SDP_SHORT, NULL, cp->cpu_id,
270 270 SDT_MEMRODA, SEL_UPL, SDP_BYTES, SDP_OP32);
271 271 #elif defined(__i386)
272 272 set_usegd(&cp->cpu_gdt[GDT_CPUID], NULL, cp->cpu_id, SDT_MEMRODA,
273 273 SEL_UPL, SDP_BYTES, SDP_OP32);
274 274 #endif
275 275 }
276 276 #endif /* !defined(__xpv) */
277 277
278 278 /*
279 279 * Multiprocessor initialization.
280 280 *
281 281 * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the
282 282 * startup and idle threads for the specified CPU.
283 283 * Parameter boot is true for boot time operations and is false for CPU
284 284 * DR operations.
285 285 */
286 286 static struct cpu *
287 287 mp_cpu_configure_common(int cpun, boolean_t boot)
288 288 {
289 289 struct cpu *cp;
290 290 kthread_id_t tp;
291 291 caddr_t sp;
292 292 proc_t *procp;
293 293 #if !defined(__xpv)
294 294 extern int idle_cpu_prefer_mwait;
295 295 extern void cpu_idle_mwait();
296 296 #endif
297 297 extern void idle();
298 298 extern void cpu_idle();
299 299
300 300 #ifdef TRAPTRACE
301 301 trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun];
302 302 #endif
303 303
304 304 ASSERT(MUTEX_HELD(&cpu_lock));
305 305 ASSERT(cpun < NCPU && cpu[cpun] == NULL);
306 306
307 307 if (cpu_free_list == NULL) {
308 308 cp = kmem_zalloc(sizeof (*cp), KM_SLEEP);
309 309 } else {
310 310 cp = cpu_free_list;
311 311 cpu_free_list = cp->cpu_next_free;
312 312 }
313 313
314 314 cp->cpu_m.mcpu_istamp = cpun << 16;
315 315
316 316 /* Create per CPU specific threads in the process p0. */
317 317 procp = &p0;
318 318
319 319 /*
320 320 * Initialize the dispatcher first.
321 321 */
322 322 disp_cpu_init(cp);
323 323
324 324 cpu_vm_data_init(cp);
325 325
326 326 /*
327 327 * Allocate and initialize the startup thread for this CPU.
328 328 * Interrupt and process switch stacks get allocated later
329 329 * when the CPU starts running.
330 330 */
331 331 tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
332 332 TS_STOPPED, maxclsyspri);
333 333
334 334 /*
335 335 * Set state to TS_ONPROC since this thread will start running
336 336 * as soon as the CPU comes online.
337 337 *
338 338 * All the other fields of the thread structure are setup by
339 339 * thread_create().
340 340 */
341 341 THREAD_ONPROC(tp, cp);
342 342 tp->t_preempt = 1;
343 343 tp->t_bound_cpu = cp;
344 344 tp->t_affinitycnt = 1;
345 345 tp->t_cpu = cp;
346 346 tp->t_disp_queue = cp->cpu_disp;
347 347
348 348 /*
349 349 * Setup thread to start in mp_startup_common.
350 350 */
351 351 sp = tp->t_stk;
352 352 tp->t_sp = (uintptr_t)(sp - MINFRAME);
353 353 #if defined(__amd64)
354 354 tp->t_sp -= STACK_ENTRY_ALIGN; /* fake a call */
355 355 #endif
356 356 /*
357 357 * Setup thread start entry point for boot or hotplug.
358 358 */
359 359 if (boot) {
360 360 tp->t_pc = (uintptr_t)mp_startup_boot;
361 361 } else {
362 362 tp->t_pc = (uintptr_t)mp_startup_hotplug;
363 363 }
364 364
365 365 cp->cpu_id = cpun;
366 366 cp->cpu_self = cp;
367 367 cp->cpu_thread = tp;
368 368 cp->cpu_lwp = NULL;
369 369 cp->cpu_dispthread = tp;
370 370 cp->cpu_dispatch_pri = DISP_PRIO(tp);
371 371
372 372 /*
373 373 * cpu_base_spl must be set explicitly here to prevent any blocking
374 374 * operations in mp_startup_common from causing the spl of the cpu
375 375 * to drop to 0 (allowing device interrupts before we're ready) in
376 376 * resume().
377 377 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY.
378 378 * As an extra bit of security on DEBUG kernels, this is enforced with
379 379 * an assertion in mp_startup_common() -- before cpu_base_spl is set
380 380 * to its proper value.
381 381 */
382 382 cp->cpu_base_spl = ipltospl(LOCK_LEVEL);
383 383
384 384 /*
385 385 * Now, initialize per-CPU idle thread for this CPU.
386 386 */
387 387 tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);
388 388
389 389 cp->cpu_idle_thread = tp;
390 390
391 391 tp->t_preempt = 1;
392 392 tp->t_bound_cpu = cp;
393 393 tp->t_affinitycnt = 1;
394 394 tp->t_cpu = cp;
395 395 tp->t_disp_queue = cp->cpu_disp;
396 396
397 397 /*
398 398 * Bootstrap the CPU's PG data
399 399 */
400 400 pg_cpu_bootstrap(cp);
401 401
402 402 /*
403 403 * Perform CPC initialization on the new CPU.
404 404 */
405 405 kcpc_hw_init(cp);
406 406
407 407 /*
408 408 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
409 409 * for each CPU.
410 410 */
411 411 setup_vaddr_for_ppcopy(cp);
412 412
413 413 /*
414 414 * Allocate page for new GDT and initialize from current GDT.
415 415 */
416 416 #if !defined(__lint)
417 417 ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE);
418 418 #endif
419 419 cp->cpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP);
420 420 bcopy(CPU->cpu_gdt, cp->cpu_gdt, (sizeof (*cp->cpu_gdt) * NGDT));
421 421
422 422 #if defined(__i386)
423 423 /*
424 424 * setup kernel %gs.
425 425 */
426 426 set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA,
427 427 SEL_KPL, 0, 1);
428 428 #endif
429 429
430 430 /*
431 431 * Allocate pages for the CPU LDT.
432 432 */
433 433 cp->cpu_m.mcpu_ldt = kmem_zalloc(LDT_CPU_SIZE, KM_SLEEP);
434 434 cp->cpu_m.mcpu_ldt_len = 0;
435 435
436 436 /*
437 437 * Allocate a per-CPU IDT and initialize the new IDT to the currently
438 438 * runing CPU.
439 439 */
440 440 #if !defined(__lint)
441 441 ASSERT((sizeof (*CPU->cpu_idt) * NIDT) <= PAGESIZE);
442 442 #endif
443 443 cp->cpu_idt = kmem_alloc(PAGESIZE, KM_SLEEP);
444 444 bcopy(CPU->cpu_idt, cp->cpu_idt, PAGESIZE);
445 445
446 446 /*
447 447 * alloc space for cpuid info
448 448 */
449 449 cpuid_alloc_space(cp);
450 450 #if !defined(__xpv)
451 451 if (is_x86_feature(x86_featureset, X86FSET_MWAIT) &&
452 452 idle_cpu_prefer_mwait) {
453 453 cp->cpu_m.mcpu_mwait = cpuid_mwait_alloc(cp);
454 454 cp->cpu_m.mcpu_idle_cpu = cpu_idle_mwait;
455 455 } else
456 456 #endif
457 457 cp->cpu_m.mcpu_idle_cpu = cpu_idle;
458 458
459 459 init_cpu_info(cp);
460 460
461 461 #if !defined(__xpv)
462 462 init_cpu_id_gdt(cp);
463 463 #endif
464 464
465 465 /*
466 466 * alloc space for ucode_info
467 467 */
468 468 ucode_alloc_space(cp);
469 469 xc_init_cpu(cp);
470 470 hat_cpu_online(cp);
471 471
472 472 #ifdef TRAPTRACE
473 473 /*
474 474 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers
475 475 */
476 476 ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
477 477 ttc->ttc_next = ttc->ttc_first;
478 478 ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
479 479 #endif
480 480
481 481 /*
482 482 * Record that we have another CPU.
483 483 */
484 484 /*
485 485 * Initialize the interrupt threads for this CPU
486 486 */
487 487 cpu_intr_alloc(cp, NINTR_THREADS);
488 488
489 489 cp->cpu_flags = CPU_OFFLINE | CPU_QUIESCED | CPU_POWEROFF;
490 490 cpu_set_state(cp);
491 491
492 492 /*
493 493 * Add CPU to list of available CPUs. It'll be on the active list
494 494 * after mp_startup_common().
495 495 */
496 496 cpu_add_unit(cp);
497 497
498 498 return (cp);
499 499 }
500 500
501 501 /*
502 502 * Undo what was done in mp_cpu_configure_common
503 503 */
504 504 static void
505 505 mp_cpu_unconfigure_common(struct cpu *cp, int error)
506 506 {
507 507 ASSERT(MUTEX_HELD(&cpu_lock));
508 508
509 509 /*
510 510 * Remove the CPU from the list of available CPUs.
511 511 */
512 512 cpu_del_unit(cp->cpu_id);
513 513
514 514 if (error == ETIMEDOUT) {
515 515 /*
516 516 * The cpu was started, but never *seemed* to run any
517 517 * code in the kernel; it's probably off spinning in its
518 518 * own private world, though with potential references to
519 519 * our kmem-allocated IDTs and GDTs (for example).
520 520 *
521 521 * Worse still, it may actually wake up some time later,
522 522 * so rather than guess what it might or might not do, we
523 523 * leave the fundamental data structures intact.
524 524 */
525 525 cp->cpu_flags = 0;
526 526 return;
527 527 }
528 528
529 529 /*
530 530 * At this point, the only threads bound to this CPU should
531 531 * special per-cpu threads: it's idle thread, it's pause threads,
532 532 * and it's interrupt threads. Clean these up.
533 533 */
534 534 cpu_destroy_bound_threads(cp);
535 535 cp->cpu_idle_thread = NULL;
536 536
537 537 /*
538 538 * Free the interrupt stack.
539 539 */
540 540 segkp_release(segkp,
541 541 cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME)));
542 542 cp->cpu_intr_stack = NULL;
543 543
544 544 #ifdef TRAPTRACE
545 545 /*
546 546 * Discard the trap trace buffer
547 547 */
548 548 {
549 549 trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id];
550 550
551 551 kmem_free((void *)ttc->ttc_first, trap_trace_bufsize);
552 552 ttc->ttc_first = NULL;
553 553 }
554 554 #endif
555 555
556 556 hat_cpu_offline(cp);
557 557
558 558 ucode_free_space(cp);
559 559
560 560 /* Free CPU ID string and brand string. */
561 561 if (cp->cpu_idstr) {
562 562 kmem_free(cp->cpu_idstr, CPU_IDSTRLEN);
563 563 cp->cpu_idstr = NULL;
564 564 }
565 565 if (cp->cpu_brandstr) {
566 566 kmem_free(cp->cpu_brandstr, CPU_IDSTRLEN);
567 567 cp->cpu_brandstr = NULL;
568 568 }
569 569
570 570 #if !defined(__xpv)
571 571 if (cp->cpu_m.mcpu_mwait != NULL) {
572 572 cpuid_mwait_free(cp);
573 573 cp->cpu_m.mcpu_mwait = NULL;
574 574 }
575 575 #endif
576 576 cpuid_free_space(cp);
577 577
578 578 if (cp->cpu_idt != CPU->cpu_idt)
579 579 kmem_free(cp->cpu_idt, PAGESIZE);
580 580 cp->cpu_idt = NULL;
581 581
582 582 kmem_free(cp->cpu_m.mcpu_ldt, LDT_CPU_SIZE);
583 583 cp->cpu_m.mcpu_ldt = NULL;
584 584 cp->cpu_m.mcpu_ldt_len = 0;
585 585
586 586 kmem_free(cp->cpu_gdt, PAGESIZE);
587 587 cp->cpu_gdt = NULL;
588 588
589 589 if (cp->cpu_supp_freqs != NULL) {
590 590 size_t len = strlen(cp->cpu_supp_freqs) + 1;
591 591 kmem_free(cp->cpu_supp_freqs, len);
592 592 cp->cpu_supp_freqs = NULL;
593 593 }
594 594
595 595 teardown_vaddr_for_ppcopy(cp);
596 596
597 597 kcpc_hw_fini(cp);
598 598
599 599 cp->cpu_dispthread = NULL;
600 600 cp->cpu_thread = NULL; /* discarded by cpu_destroy_bound_threads() */
601 601
602 602 cpu_vm_data_destroy(cp);
603 603
604 604 xc_fini_cpu(cp);
605 605 disp_cpu_fini(cp);
606 606
607 607 ASSERT(cp != CPU0);
608 608 bzero(cp, sizeof (*cp));
609 609 cp->cpu_next_free = cpu_free_list;
610 610 cpu_free_list = cp;
611 611 }
612 612
613 613 /*
614 614 * Apply workarounds for known errata, and warn about those that are absent.
615 615 *
616 616 * System vendors occasionally create configurations which contain different
617 617 * revisions of the CPUs that are almost but not exactly the same. At the
618 618 * time of writing, this meant that their clock rates were the same, their
619 619 * feature sets were the same, but the required workaround were -not-
620 620 * necessarily the same. So, this routine is invoked on -every- CPU soon
621 621 * after starting to make sure that the resulting system contains the most
622 622 * pessimal set of workarounds needed to cope with *any* of the CPUs in the
623 623 * system.
624 624 *
625 625 * workaround_errata is invoked early in mlsetup() for CPU 0, and in
626 626 * mp_startup_common() for all slave CPUs. Slaves process workaround_errata
627 627 * prior to acknowledging their readiness to the master, so this routine will
628 628 * never be executed by multiple CPUs in parallel, thus making updates to
629 629 * global data safe.
630 630 *
631 631 * These workarounds are based on Rev 3.57 of the Revision Guide for
632 632 * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005.
633 633 */
634 634
635 635 #if defined(OPTERON_ERRATUM_88)
636 636 int opteron_erratum_88; /* if non-zero -> at least one cpu has it */
637 637 #endif
638 638
639 639 #if defined(OPTERON_ERRATUM_91)
640 640 int opteron_erratum_91; /* if non-zero -> at least one cpu has it */
641 641 #endif
642 642
643 643 #if defined(OPTERON_ERRATUM_93)
644 644 int opteron_erratum_93; /* if non-zero -> at least one cpu has it */
645 645 #endif
646 646
647 647 #if defined(OPTERON_ERRATUM_95)
648 648 int opteron_erratum_95; /* if non-zero -> at least one cpu has it */
649 649 #endif
650 650
651 651 #if defined(OPTERON_ERRATUM_100)
652 652 int opteron_erratum_100; /* if non-zero -> at least one cpu has it */
653 653 #endif
654 654
655 655 #if defined(OPTERON_ERRATUM_108)
656 656 int opteron_erratum_108; /* if non-zero -> at least one cpu has it */
657 657 #endif
658 658
659 659 #if defined(OPTERON_ERRATUM_109)
660 660 int opteron_erratum_109; /* if non-zero -> at least one cpu has it */
661 661 #endif
662 662
663 663 #if defined(OPTERON_ERRATUM_121)
664 664 int opteron_erratum_121; /* if non-zero -> at least one cpu has it */
665 665 #endif
666 666
667 667 #if defined(OPTERON_ERRATUM_122)
668 668 int opteron_erratum_122; /* if non-zero -> at least one cpu has it */
669 669 #endif
670 670
671 671 #if defined(OPTERON_ERRATUM_123)
672 672 int opteron_erratum_123; /* if non-zero -> at least one cpu has it */
673 673 #endif
674 674
675 675 #if defined(OPTERON_ERRATUM_131)
676 676 int opteron_erratum_131; /* if non-zero -> at least one cpu has it */
677 677 #endif
678 678
679 679 #if defined(OPTERON_WORKAROUND_6336786)
680 680 int opteron_workaround_6336786; /* non-zero -> WA relevant and applied */
681 681 int opteron_workaround_6336786_UP = 0; /* Not needed for UP */
682 682 #endif
683 683
684 684 #if defined(OPTERON_WORKAROUND_6323525)
685 685 int opteron_workaround_6323525; /* if non-zero -> at least one cpu has it */
686 686 #endif
687 687
688 688 #if defined(OPTERON_ERRATUM_298)
689 689 int opteron_erratum_298;
690 690 #endif
691 691
692 692 #if defined(OPTERON_ERRATUM_721)
693 693 int opteron_erratum_721;
694 694 #endif
695 695
696 696 static void
697 697 workaround_warning(cpu_t *cp, uint_t erratum)
698 698 {
699 699 cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u",
700 700 cp->cpu_id, erratum);
701 701 }
702 702
703 703 static void
704 704 workaround_applied(uint_t erratum)
705 705 {
706 706 if (erratum > 1000000)
707 707 cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n",
708 708 erratum);
709 709 else
710 710 cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n",
711 711 erratum);
712 712 }
713 713
714 714 static void
715 715 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error)
716 716 {
717 717 cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d",
718 718 cp->cpu_id, rw, msr, error);
719 719 }
720 720
721 721 /*
722 722 * Determine the number of nodes in a Hammer / Greyhound / Griffin family
723 723 * system.
724 724 */
725 725 static uint_t
726 726 opteron_get_nnodes(void)
727 727 {
728 728 static uint_t nnodes = 0;
729 729
730 730 if (nnodes == 0) {
731 731 #ifdef DEBUG
732 732 uint_t family;
733 733
734 734 /*
735 735 * This routine uses a PCI config space based mechanism
736 736 * for retrieving the number of nodes in the system.
737 737 * Device 24, function 0, offset 0x60 as used here is not
738 738 * AMD processor architectural, and may not work on processor
739 739 * families other than those listed below.
740 740 *
741 741 * Callers of this routine must ensure that we're running on
742 742 * a processor which supports this mechanism.
743 743 * The assertion below is meant to catch calls on unsupported
744 744 * processors.
745 745 */
746 746 family = cpuid_getfamily(CPU);
747 747 ASSERT(family == 0xf || family == 0x10 || family == 0x11);
748 748 #endif /* DEBUG */
749 749
750 750 /*
751 751 * Obtain the number of nodes in the system from
752 752 * bits [6:4] of the Node ID register on node 0.
753 753 *
754 754 * The actual node count is NodeID[6:4] + 1
755 755 *
756 756 * The Node ID register is accessed via function 0,
757 757 * offset 0x60. Node 0 is device 24.
758 758 */
759 759 nnodes = ((pci_getl_func(0, 24, 0, 0x60) & 0x70) >> 4) + 1;
760 760 }
761 761 return (nnodes);
762 762 }
763 763
764 764 uint_t
765 765 do_erratum_298(struct cpu *cpu)
766 766 {
767 767 static int osvwrc = -3;
768 768 extern int osvw_opteron_erratum(cpu_t *, uint_t);
769 769
770 770 /*
771 771 * L2 Eviction May Occur During Processor Operation To Set
772 772 * Accessed or Dirty Bit.
773 773 */
774 774 if (osvwrc == -3) {
775 775 osvwrc = osvw_opteron_erratum(cpu, 298);
776 776 } else {
777 777 /* osvw return codes should be consistent for all cpus */
778 778 ASSERT(osvwrc == osvw_opteron_erratum(cpu, 298));
779 779 }
780 780
781 781 switch (osvwrc) {
782 782 case 0: /* erratum is not present: do nothing */
783 783 break;
784 784 case 1: /* erratum is present: BIOS workaround applied */
785 785 /*
786 786 * check if workaround is actually in place and issue warning
787 787 * if not.
788 788 */
789 789 if (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
790 790 ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0)) {
791 791 #if defined(OPTERON_ERRATUM_298)
792 792 opteron_erratum_298++;
793 793 #else
794 794 workaround_warning(cpu, 298);
795 795 return (1);
796 796 #endif
797 797 }
798 798 break;
799 799 case -1: /* cannot determine via osvw: check cpuid */
800 800 if ((cpuid_opteron_erratum(cpu, 298) > 0) &&
801 801 (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
802 802 ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0))) {
803 803 #if defined(OPTERON_ERRATUM_298)
804 804 opteron_erratum_298++;
805 805 #else
806 806 workaround_warning(cpu, 298);
807 807 return (1);
808 808 #endif
809 809 }
810 810 break;
811 811 }
812 812 return (0);
813 813 }
814 814
815 815 uint_t
816 816 workaround_errata(struct cpu *cpu)
817 817 {
818 818 uint_t missing = 0;
819 819
820 820 ASSERT(cpu == CPU);
821 821
822 822 /*LINTED*/
823 823 if (cpuid_opteron_erratum(cpu, 88) > 0) {
824 824 /*
825 825 * SWAPGS May Fail To Read Correct GS Base
826 826 */
827 827 #if defined(OPTERON_ERRATUM_88)
828 828 /*
829 829 * The workaround is an mfence in the relevant assembler code
830 830 */
831 831 opteron_erratum_88++;
832 832 #else
833 833 workaround_warning(cpu, 88);
834 834 missing++;
835 835 #endif
836 836 }
837 837
838 838 if (cpuid_opteron_erratum(cpu, 91) > 0) {
839 839 /*
840 840 * Software Prefetches May Report A Page Fault
841 841 */
842 842 #if defined(OPTERON_ERRATUM_91)
843 843 /*
844 844 * fix is in trap.c
845 845 */
846 846 opteron_erratum_91++;
847 847 #else
848 848 workaround_warning(cpu, 91);
849 849 missing++;
850 850 #endif
851 851 }
852 852
853 853 if (cpuid_opteron_erratum(cpu, 93) > 0) {
854 854 /*
855 855 * RSM Auto-Halt Restart Returns to Incorrect RIP
856 856 */
857 857 #if defined(OPTERON_ERRATUM_93)
858 858 /*
859 859 * fix is in trap.c
860 860 */
861 861 opteron_erratum_93++;
862 862 #else
863 863 workaround_warning(cpu, 93);
864 864 missing++;
865 865 #endif
866 866 }
867 867
868 868 /*LINTED*/
869 869 if (cpuid_opteron_erratum(cpu, 95) > 0) {
870 870 /*
871 871 * RET Instruction May Return to Incorrect EIP
872 872 */
873 873 #if defined(OPTERON_ERRATUM_95)
874 874 #if defined(_LP64)
875 875 /*
876 876 * Workaround this by ensuring that 32-bit user code and
877 877 * 64-bit kernel code never occupy the same address
878 878 * range mod 4G.
879 879 */
880 880 if (_userlimit32 > 0xc0000000ul)
881 881 *(uintptr_t *)&_userlimit32 = 0xc0000000ul;
882 882
883 883 /*LINTED*/
884 884 ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
885 885 opteron_erratum_95++;
886 886 #endif /* _LP64 */
887 887 #else
888 888 workaround_warning(cpu, 95);
889 889 missing++;
890 890 #endif
891 891 }
892 892
893 893 if (cpuid_opteron_erratum(cpu, 100) > 0) {
894 894 /*
895 895 * Compatibility Mode Branches Transfer to Illegal Address
896 896 */
897 897 #if defined(OPTERON_ERRATUM_100)
898 898 /*
899 899 * fix is in trap.c
900 900 */
901 901 opteron_erratum_100++;
902 902 #else
903 903 workaround_warning(cpu, 100);
904 904 missing++;
905 905 #endif
906 906 }
907 907
908 908 /*LINTED*/
909 909 if (cpuid_opteron_erratum(cpu, 108) > 0) {
910 910 /*
911 911 * CPUID Instruction May Return Incorrect Model Number In
912 912 * Some Processors
913 913 */
914 914 #if defined(OPTERON_ERRATUM_108)
915 915 /*
916 916 * (Our cpuid-handling code corrects the model number on
917 917 * those processors)
918 918 */
919 919 #else
920 920 workaround_warning(cpu, 108);
921 921 missing++;
922 922 #endif
923 923 }
924 924
925 925 /*LINTED*/
926 926 if (cpuid_opteron_erratum(cpu, 109) > 0) do {
927 927 /*
928 928 * Certain Reverse REP MOVS May Produce Unpredictable Behavior
929 929 */
930 930 #if defined(OPTERON_ERRATUM_109)
931 931 /*
932 932 * The "workaround" is to print a warning to upgrade the BIOS
933 933 */
934 934 uint64_t value;
935 935 const uint_t msr = MSR_AMD_PATCHLEVEL;
936 936 int err;
937 937
938 938 if ((err = checked_rdmsr(msr, &value)) != 0) {
939 939 msr_warning(cpu, "rd", msr, err);
940 940 workaround_warning(cpu, 109);
941 941 missing++;
942 942 }
943 943 if (value == 0)
944 944 opteron_erratum_109++;
945 945 #else
946 946 workaround_warning(cpu, 109);
947 947 missing++;
948 948 #endif
949 949 /*CONSTANTCONDITION*/
950 950 } while (0);
951 951
952 952 /*LINTED*/
953 953 if (cpuid_opteron_erratum(cpu, 121) > 0) {
954 954 /*
955 955 * Sequential Execution Across Non_Canonical Boundary Caused
956 956 * Processor Hang
957 957 */
958 958 #if defined(OPTERON_ERRATUM_121)
959 959 #if defined(_LP64)
960 960 /*
961 961 * Erratum 121 is only present in long (64 bit) mode.
962 962 * Workaround is to include the page immediately before the
963 963 * va hole to eliminate the possibility of system hangs due to
964 964 * sequential execution across the va hole boundary.
965 965 */
966 966 if (opteron_erratum_121)
967 967 opteron_erratum_121++;
968 968 else {
969 969 if (hole_start) {
970 970 hole_start -= PAGESIZE;
971 971 } else {
972 972 /*
973 973 * hole_start not yet initialized by
974 974 * mmu_init. Initialize hole_start
975 975 * with value to be subtracted.
976 976 */
977 977 hole_start = PAGESIZE;
978 978 }
979 979 opteron_erratum_121++;
980 980 }
981 981 #endif /* _LP64 */
982 982 #else
983 983 workaround_warning(cpu, 121);
984 984 missing++;
985 985 #endif
986 986 }
987 987
988 988 /*LINTED*/
989 989 if (cpuid_opteron_erratum(cpu, 122) > 0) do {
990 990 /*
991 991 * TLB Flush Filter May Cause Coherency Problem in
992 992 * Multiprocessor Systems
993 993 */
994 994 #if defined(OPTERON_ERRATUM_122)
995 995 uint64_t value;
996 996 const uint_t msr = MSR_AMD_HWCR;
997 997 int error;
998 998
999 999 /*
1000 1000 * Erratum 122 is only present in MP configurations (multi-core
1001 1001 * or multi-processor).
1002 1002 */
1003 1003 #if defined(__xpv)
1004 1004 if (!DOMAIN_IS_INITDOMAIN(xen_info))
1005 1005 break;
1006 1006 if (!opteron_erratum_122 && xpv_nr_phys_cpus() == 1)
1007 1007 break;
1008 1008 #else
1009 1009 if (!opteron_erratum_122 && opteron_get_nnodes() == 1 &&
1010 1010 cpuid_get_ncpu_per_chip(cpu) == 1)
1011 1011 break;
1012 1012 #endif
1013 1013 /* disable TLB Flush Filter */
1014 1014
1015 1015 if ((error = checked_rdmsr(msr, &value)) != 0) {
1016 1016 msr_warning(cpu, "rd", msr, error);
1017 1017 workaround_warning(cpu, 122);
1018 1018 missing++;
1019 1019 } else {
1020 1020 value |= (uint64_t)AMD_HWCR_FFDIS;
1021 1021 if ((error = checked_wrmsr(msr, value)) != 0) {
1022 1022 msr_warning(cpu, "wr", msr, error);
1023 1023 workaround_warning(cpu, 122);
1024 1024 missing++;
1025 1025 }
1026 1026 }
1027 1027 opteron_erratum_122++;
1028 1028 #else
1029 1029 workaround_warning(cpu, 122);
1030 1030 missing++;
1031 1031 #endif
1032 1032 /*CONSTANTCONDITION*/
1033 1033 } while (0);
1034 1034
1035 1035 /*LINTED*/
1036 1036 if (cpuid_opteron_erratum(cpu, 123) > 0) do {
1037 1037 /*
1038 1038 * Bypassed Reads May Cause Data Corruption of System Hang in
1039 1039 * Dual Core Processors
1040 1040 */
1041 1041 #if defined(OPTERON_ERRATUM_123)
1042 1042 uint64_t value;
1043 1043 const uint_t msr = MSR_AMD_PATCHLEVEL;
1044 1044 int err;
1045 1045
1046 1046 /*
1047 1047 * Erratum 123 applies only to multi-core cpus.
1048 1048 */
1049 1049 if (cpuid_get_ncpu_per_chip(cpu) < 2)
1050 1050 break;
1051 1051 #if defined(__xpv)
1052 1052 if (!DOMAIN_IS_INITDOMAIN(xen_info))
1053 1053 break;
1054 1054 #endif
1055 1055 /*
1056 1056 * The "workaround" is to print a warning to upgrade the BIOS
1057 1057 */
1058 1058 if ((err = checked_rdmsr(msr, &value)) != 0) {
1059 1059 msr_warning(cpu, "rd", msr, err);
1060 1060 workaround_warning(cpu, 123);
1061 1061 missing++;
1062 1062 }
1063 1063 if (value == 0)
1064 1064 opteron_erratum_123++;
1065 1065 #else
1066 1066 workaround_warning(cpu, 123);
1067 1067 missing++;
1068 1068
1069 1069 #endif
1070 1070 /*CONSTANTCONDITION*/
1071 1071 } while (0);
1072 1072
1073 1073 /*LINTED*/
1074 1074 if (cpuid_opteron_erratum(cpu, 131) > 0) do {
1075 1075 /*
1076 1076 * Multiprocessor Systems with Four or More Cores May Deadlock
1077 1077 * Waiting for a Probe Response
1078 1078 */
1079 1079 #if defined(OPTERON_ERRATUM_131)
1080 1080 uint64_t nbcfg;
1081 1081 const uint_t msr = MSR_AMD_NB_CFG;
1082 1082 const uint64_t wabits =
1083 1083 AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR;
1084 1084 int error;
1085 1085
1086 1086 /*
1087 1087 * Erratum 131 applies to any system with four or more cores.
1088 1088 */
1089 1089 if (opteron_erratum_131)
1090 1090 break;
1091 1091 #if defined(__xpv)
1092 1092 if (!DOMAIN_IS_INITDOMAIN(xen_info))
1093 1093 break;
1094 1094 if (xpv_nr_phys_cpus() < 4)
1095 1095 break;
1096 1096 #else
1097 1097 if (opteron_get_nnodes() * cpuid_get_ncpu_per_chip(cpu) < 4)
1098 1098 break;
1099 1099 #endif
1100 1100 /*
1101 1101 * Print a warning if neither of the workarounds for
1102 1102 * erratum 131 is present.
1103 1103 */
1104 1104 if ((error = checked_rdmsr(msr, &nbcfg)) != 0) {
1105 1105 msr_warning(cpu, "rd", msr, error);
1106 1106 workaround_warning(cpu, 131);
1107 1107 missing++;
1108 1108 } else if ((nbcfg & wabits) == 0) {
1109 1109 opteron_erratum_131++;
1110 1110 } else {
1111 1111 /* cannot have both workarounds set */
1112 1112 ASSERT((nbcfg & wabits) != wabits);
1113 1113 }
1114 1114 #else
1115 1115 workaround_warning(cpu, 131);
1116 1116 missing++;
1117 1117 #endif
1118 1118 /*CONSTANTCONDITION*/
1119 1119 } while (0);
1120 1120
1121 1121 /*
1122 1122 * This isn't really an erratum, but for convenience the
1123 1123 * detection/workaround code lives here and in cpuid_opteron_erratum.
1124 1124 */
1125 1125 if (cpuid_opteron_erratum(cpu, 6336786) > 0) {
1126 1126 #if defined(OPTERON_WORKAROUND_6336786)
1127 1127 /*
1128 1128 * Disable C1-Clock ramping on multi-core/multi-processor
1129 1129 * K8 platforms to guard against TSC drift.
1130 1130 */
1131 1131 if (opteron_workaround_6336786) {
1132 1132 opteron_workaround_6336786++;
1133 1133 #if defined(__xpv)
1134 1134 } else if ((DOMAIN_IS_INITDOMAIN(xen_info) &&
1135 1135 xpv_nr_phys_cpus() > 1) ||
1136 1136 opteron_workaround_6336786_UP) {
1137 1137 /*
1138 1138 * XXPV Hmm. We can't walk the Northbridges on
1139 1139 * the hypervisor; so just complain and drive
1140 1140 * on. This probably needs to be fixed in
1141 1141 * the hypervisor itself.
1142 1142 */
1143 1143 opteron_workaround_6336786++;
1144 1144 workaround_warning(cpu, 6336786);
1145 1145 #else /* __xpv */
1146 1146 } else if ((opteron_get_nnodes() *
1147 1147 cpuid_get_ncpu_per_chip(cpu) > 1) ||
1148 1148 opteron_workaround_6336786_UP) {
1149 1149
1150 1150 uint_t node, nnodes;
1151 1151 uint8_t data;
1152 1152
1153 1153 nnodes = opteron_get_nnodes();
1154 1154 for (node = 0; node < nnodes; node++) {
1155 1155 /*
1156 1156 * Clear PMM7[1:0] (function 3, offset 0x87)
1157 1157 * Northbridge device is the node id + 24.
1158 1158 */
1159 1159 data = pci_getb_func(0, node + 24, 3, 0x87);
1160 1160 data &= 0xFC;
1161 1161 pci_putb_func(0, node + 24, 3, 0x87, data);
1162 1162 }
1163 1163 opteron_workaround_6336786++;
1164 1164 #endif /* __xpv */
1165 1165 }
1166 1166 #else
1167 1167 workaround_warning(cpu, 6336786);
1168 1168 missing++;
1169 1169 #endif
1170 1170 }
1171 1171
1172 1172 /*LINTED*/
1173 1173 /*
1174 1174 * Mutex primitives don't work as expected.
1175 1175 */
1176 1176 if (cpuid_opteron_erratum(cpu, 6323525) > 0) {
1177 1177 #if defined(OPTERON_WORKAROUND_6323525)
1178 1178 /*
1179 1179 * This problem only occurs with 2 or more cores. If bit in
1180 1180 * MSR_AMD_BU_CFG set, then not applicable. The workaround
1181 1181 * is to patch the semaphone routines with the lfence
1182 1182 * instruction to provide necessary load memory barrier with
1183 1183 * possible subsequent read-modify-write ops.
1184 1184 *
1185 1185 * It is too early in boot to call the patch routine so
1186 1186 * set erratum variable to be done in startup_end().
1187 1187 */
1188 1188 if (opteron_workaround_6323525) {
1189 1189 opteron_workaround_6323525++;
1190 1190 #if defined(__xpv)
1191 1191 } else if (is_x86_feature(x86_featureset, X86FSET_SSE2)) {
1192 1192 if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1193 1193 /*
1194 1194 * XXPV Use dom0_msr here when extended
1195 1195 * operations are supported?
1196 1196 */
1197 1197 if (xpv_nr_phys_cpus() > 1)
1198 1198 opteron_workaround_6323525++;
1199 1199 } else {
1200 1200 /*
1201 1201 * We have no way to tell how many physical
1202 1202 * cpus there are, or even if this processor
1203 1203 * has the problem, so enable the workaround
1204 1204 * unconditionally (at some performance cost).
1205 1205 */
1206 1206 opteron_workaround_6323525++;
1207 1207 }
1208 1208 #else /* __xpv */
1209 1209 } else if (is_x86_feature(x86_featureset, X86FSET_SSE2) &&
1210 1210 ((opteron_get_nnodes() *
1211 1211 cpuid_get_ncpu_per_chip(cpu)) > 1)) {
1212 1212 if ((xrdmsr(MSR_AMD_BU_CFG) & (UINT64_C(1) << 33)) == 0)
1213 1213 opteron_workaround_6323525++;
1214 1214 #endif /* __xpv */
1215 1215 }
1216 1216 #else
1217 1217 workaround_warning(cpu, 6323525);
1218 1218 missing++;
1219 1219 #endif
1220 1220 }
1221 1221
1222 1222 missing += do_erratum_298(cpu);
1223 1223
1224 1224 if (cpuid_opteron_erratum(cpu, 721) > 0) {
1225 1225 #if defined(OPTERON_ERRATUM_721)
1226 1226 on_trap_data_t otd;
1227 1227
1228 1228 if (!on_trap(&otd, OT_DATA_ACCESS))
1229 1229 wrmsr(MSR_AMD_DE_CFG,
1230 1230 rdmsr(MSR_AMD_DE_CFG) | AMD_DE_CFG_E721);
1231 1231 no_trap();
1232 1232
1233 1233 opteron_erratum_721++;
1234 1234 #else
1235 1235 workaround_warning(cpu, 721);
1236 1236 missing++;
1237 1237 #endif
1238 1238 }
1239 1239
1240 1240 #ifdef __xpv
1241 1241 return (0);
1242 1242 #else
1243 1243 return (missing);
1244 1244 #endif
1245 1245 }
1246 1246
1247 1247 void
1248 1248 workaround_errata_end()
1249 1249 {
1250 1250 #if defined(OPTERON_ERRATUM_88)
1251 1251 if (opteron_erratum_88)
1252 1252 workaround_applied(88);
1253 1253 #endif
1254 1254 #if defined(OPTERON_ERRATUM_91)
1255 1255 if (opteron_erratum_91)
1256 1256 workaround_applied(91);
1257 1257 #endif
1258 1258 #if defined(OPTERON_ERRATUM_93)
1259 1259 if (opteron_erratum_93)
1260 1260 workaround_applied(93);
1261 1261 #endif
1262 1262 #if defined(OPTERON_ERRATUM_95)
1263 1263 if (opteron_erratum_95)
1264 1264 workaround_applied(95);
1265 1265 #endif
1266 1266 #if defined(OPTERON_ERRATUM_100)
1267 1267 if (opteron_erratum_100)
1268 1268 workaround_applied(100);
1269 1269 #endif
1270 1270 #if defined(OPTERON_ERRATUM_108)
1271 1271 if (opteron_erratum_108)
1272 1272 workaround_applied(108);
1273 1273 #endif
1274 1274 #if defined(OPTERON_ERRATUM_109)
1275 1275 if (opteron_erratum_109) {
1276 1276 cmn_err(CE_WARN,
1277 1277 "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1278 1278 " processor\nerratum 109 was not detected; updating your"
1279 1279 " system's BIOS to a version\ncontaining this"
1280 1280 " microcode patch is HIGHLY recommended or erroneous"
1281 1281 " system\noperation may occur.\n");
1282 1282 }
1283 1283 #endif
1284 1284 #if defined(OPTERON_ERRATUM_121)
1285 1285 if (opteron_erratum_121)
1286 1286 workaround_applied(121);
1287 1287 #endif
1288 1288 #if defined(OPTERON_ERRATUM_122)
1289 1289 if (opteron_erratum_122)
1290 1290 workaround_applied(122);
1291 1291 #endif
1292 1292 #if defined(OPTERON_ERRATUM_123)
1293 1293 if (opteron_erratum_123) {
1294 1294 cmn_err(CE_WARN,
1295 1295 "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1296 1296 " processor\nerratum 123 was not detected; updating your"
1297 1297 " system's BIOS to a version\ncontaining this"
1298 1298 " microcode patch is HIGHLY recommended or erroneous"
1299 1299 " system\noperation may occur.\n");
1300 1300 }
1301 1301 #endif
1302 1302 #if defined(OPTERON_ERRATUM_131)
1303 1303 if (opteron_erratum_131) {
1304 1304 cmn_err(CE_WARN,
1305 1305 "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1306 1306 " processor\nerratum 131 was not detected; updating your"
1307 1307 " system's BIOS to a version\ncontaining this"
1308 1308 " microcode patch is HIGHLY recommended or erroneous"
1309 1309 " system\noperation may occur.\n");
1310 1310 }
1311 1311 #endif
1312 1312 #if defined(OPTERON_WORKAROUND_6336786)
1313 1313 if (opteron_workaround_6336786)
1314 1314 workaround_applied(6336786);
1315 1315 #endif
1316 1316 #if defined(OPTERON_WORKAROUND_6323525)
1317 1317 if (opteron_workaround_6323525)
1318 1318 workaround_applied(6323525);
1319 1319 #endif
1320 1320 #if defined(OPTERON_ERRATUM_298)
1321 1321 if (opteron_erratum_298) {
1322 1322 cmn_err(CE_WARN,
1323 1323 "BIOS microcode patch for AMD 64/Opteron(tm)"
1324 1324 " processor\nerratum 298 was not detected; updating your"
1325 1325 " system's BIOS to a version\ncontaining this"
1326 1326 " microcode patch is HIGHLY recommended or erroneous"
1327 1327 " system\noperation may occur.\n");
1328 1328 }
1329 1329 #endif
1330 1330 #if defined(OPTERON_ERRATUM_721)
1331 1331 if (opteron_erratum_721)
1332 1332 workaround_applied(721);
1333 1333 #endif
1334 1334 }
1335 1335
1336 1336 /*
1337 1337 * The procset_slave and procset_master are used to synchronize
1338 1338 * between the control CPU and the target CPU when starting CPUs.
1339 1339 */
1340 1340 static cpuset_t procset_slave, procset_master;
1341 1341
1342 1342 static void
1343 1343 mp_startup_wait(cpuset_t *sp, processorid_t cpuid)
1344 1344 {
1345 1345 cpuset_t tempset;
1346 1346
1347 1347 for (tempset = *sp; !CPU_IN_SET(tempset, cpuid);
1348 1348 tempset = *(volatile cpuset_t *)sp) {
1349 1349 SMT_PAUSE();
1350 1350 }
1351 1351 CPUSET_ATOMIC_DEL(*(cpuset_t *)sp, cpuid);
1352 1352 }
1353 1353
1354 1354 static void
1355 1355 mp_startup_signal(cpuset_t *sp, processorid_t cpuid)
1356 1356 {
1357 1357 cpuset_t tempset;
1358 1358
1359 1359 CPUSET_ATOMIC_ADD(*(cpuset_t *)sp, cpuid);
1360 1360 for (tempset = *sp; CPU_IN_SET(tempset, cpuid);
1361 1361 tempset = *(volatile cpuset_t *)sp) {
1362 1362 SMT_PAUSE();
1363 1363 }
1364 1364 }
1365 1365
1366 1366 int
1367 1367 mp_start_cpu_common(cpu_t *cp, boolean_t boot)
1368 1368 {
1369 1369 _NOTE(ARGUNUSED(boot));
1370 1370
1371 1371 void *ctx;
1372 1372 int delays;
1373 1373 int error = 0;
1374 1374 cpuset_t tempset;
1375 1375 processorid_t cpuid;
1376 1376 #ifndef __xpv
1377 1377 extern void cpupm_init(cpu_t *);
1378 1378 #endif
1379 1379
1380 1380 ASSERT(cp != NULL);
1381 1381 cpuid = cp->cpu_id;
1382 1382 ctx = mach_cpucontext_alloc(cp);
1383 1383 if (ctx == NULL) {
1384 1384 cmn_err(CE_WARN,
1385 1385 "cpu%d: failed to allocate context", cp->cpu_id);
1386 1386 return (EAGAIN);
1387 1387 }
1388 1388 error = mach_cpu_start(cp, ctx);
1389 1389 if (error != 0) {
1390 1390 cmn_err(CE_WARN,
1391 1391 "cpu%d: failed to start, error %d", cp->cpu_id, error);
1392 1392 mach_cpucontext_free(cp, ctx, error);
1393 1393 return (error);
1394 1394 }
1395 1395
1396 1396 for (delays = 0, tempset = procset_slave; !CPU_IN_SET(tempset, cpuid);
1397 1397 delays++) {
1398 1398 if (delays == 500) {
1399 1399 /*
1400 1400 * After five seconds, things are probably looking
1401 1401 * a bit bleak - explain the hang.
1402 1402 */
1403 1403 cmn_err(CE_NOTE, "cpu%d: started, "
1404 1404 "but not running in the kernel yet", cpuid);
1405 1405 } else if (delays > 2000) {
1406 1406 /*
1407 1407 * We waited at least 20 seconds, bail ..
1408 1408 */
1409 1409 error = ETIMEDOUT;
1410 1410 cmn_err(CE_WARN, "cpu%d: timed out", cpuid);
1411 1411 mach_cpucontext_free(cp, ctx, error);
1412 1412 return (error);
1413 1413 }
1414 1414
1415 1415 /*
1416 1416 * wait at least 10ms, then check again..
1417 1417 */
1418 1418 delay(USEC_TO_TICK_ROUNDUP(10000));
1419 1419 tempset = *((volatile cpuset_t *)&procset_slave);
1420 1420 }
1421 1421 CPUSET_ATOMIC_DEL(procset_slave, cpuid);
1422 1422
1423 1423 mach_cpucontext_free(cp, ctx, 0);
1424 1424
1425 1425 #ifndef __xpv
1426 1426 if (tsc_gethrtime_enable)
1427 1427 tsc_sync_master(cpuid);
1428 1428 #endif
1429 1429
1430 1430 if (dtrace_cpu_init != NULL) {
1431 1431 (*dtrace_cpu_init)(cpuid);
1432 1432 }
1433 1433
1434 1434 /*
1435 1435 * During CPU DR operations, the cpu_lock is held by current
1436 1436 * (the control) thread. We can't release the cpu_lock here
1437 1437 * because that will break the CPU DR logic.
1438 1438 * On the other hand, CPUPM and processor group initialization
1439 1439 * routines need to access the cpu_lock. So we invoke those
1440 1440 * routines here on behalf of mp_startup_common().
1441 1441 *
1442 1442 * CPUPM and processor group initialization routines depend
1443 1443 * on the cpuid probing results. Wait for mp_startup_common()
1444 1444 * to signal that cpuid probing is done.
1445 1445 */
1446 1446 mp_startup_wait(&procset_slave, cpuid);
1447 1447 #ifndef __xpv
1448 1448 cpupm_init(cp);
1449 1449 #endif
1450 1450 (void) pg_cpu_init(cp, B_FALSE);
1451 1451 cpu_set_state(cp);
1452 1452 mp_startup_signal(&procset_master, cpuid);
1453 1453
1454 1454 return (0);
1455 1455 }
1456 1456
1457 1457 /*
1458 1458 * Start a single cpu, assuming that the kernel context is available
1459 1459 * to successfully start another cpu.
1460 1460 *
1461 1461 * (For example, real mode code is mapped into the right place
1462 1462 * in memory and is ready to be run.)
1463 1463 */
1464 1464 int
1465 1465 start_cpu(processorid_t who)
1466 1466 {
1467 1467 cpu_t *cp;
1468 1468 int error = 0;
1469 1469 cpuset_t tempset;
1470 1470
1471 1471 ASSERT(who != 0);
1472 1472
1473 1473 /*
1474 1474 * Check if there's at least a Mbyte of kmem available
1475 1475 * before attempting to start the cpu.
1476 1476 */
1477 1477 if (kmem_avail() < 1024 * 1024) {
1478 1478 /*
1479 1479 * Kick off a reap in case that helps us with
1480 1480 * later attempts ..
1481 1481 */
1482 1482 kmem_reap();
1483 1483 return (ENOMEM);
1484 1484 }
1485 1485
1486 1486 /*
1487 1487 * First configure cpu.
1488 1488 */
1489 1489 cp = mp_cpu_configure_common(who, B_TRUE);
1490 1490 ASSERT(cp != NULL);
1491 1491
1492 1492 /*
1493 1493 * Then start cpu.
1494 1494 */
1495 1495 error = mp_start_cpu_common(cp, B_TRUE);
1496 1496 if (error != 0) {
1497 1497 mp_cpu_unconfigure_common(cp, error);
1498 1498 return (error);
1499 1499 }
1500 1500
1501 1501 mutex_exit(&cpu_lock);
1502 1502 tempset = cpu_ready_set;
1503 1503 while (!CPU_IN_SET(tempset, who)) {
1504 1504 drv_usecwait(1);
1505 1505 tempset = *((volatile cpuset_t *)&cpu_ready_set);
1506 1506 }
1507 1507 mutex_enter(&cpu_lock);
1508 1508
1509 1509 return (0);
1510 1510 }
1511 1511
1512 1512 void
1513 1513 start_other_cpus(int cprboot)
1514 1514 {
1515 1515 _NOTE(ARGUNUSED(cprboot));
1516 1516
1517 1517 uint_t who;
1518 1518 uint_t bootcpuid = 0;
1519 1519
1520 1520 /*
1521 1521 * Initialize our own cpu_info.
1522 1522 */
1523 1523 init_cpu_info(CPU);
1524 1524
1525 1525 #if !defined(__xpv)
1526 1526 init_cpu_id_gdt(CPU);
1527 1527 #endif
1528 1528
1529 1529 cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_idstr);
1530 1530 cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_brandstr);
1531 1531
1532 1532 /*
1533 1533 * KPTI initialisation happens very early in boot, before logging is
1534 1534 * set up. Output a status message now as the boot CPU comes online.
1535 1535 */
1536 1536 cmn_err(CE_CONT, "?KPTI %s (PCID %s, INVPCID %s)\n",
1537 1537 kpti_enable ? "enabled" : "disabled",
1538 1538 x86_use_pcid == 1 ? "in use" :
1539 1539 (is_x86_feature(x86_featureset, X86FSET_PCID) ? "disabled" :
1540 1540 "not supported"),
1541 1541 x86_use_pcid == 1 && x86_use_invpcid == 1 ? "in use" :
1542 1542 (is_x86_feature(x86_featureset, X86FSET_INVPCID) ? "disabled" :
1543 1543 "not supported"));
1544 1544
1545 1545 /*
1546 1546 * Initialize our syscall handlers
1547 1547 */
1548 1548 init_cpu_syscall(CPU);
1549 1549
1550 1550 /*
1551 1551 * Take the boot cpu out of the mp_cpus set because we know
1552 1552 * it's already running. Add it to the cpu_ready_set for
1553 1553 * precisely the same reason.
1554 1554 */
1555 1555 CPUSET_DEL(mp_cpus, bootcpuid);
1556 1556 CPUSET_ADD(cpu_ready_set, bootcpuid);
1557 1557
1558 1558 /*
1559 1559 * skip the rest of this if
1560 1560 * . only 1 cpu dectected and system isn't hotplug-capable
1561 1561 * . not using MP
1562 1562 */
1563 1563 if ((CPUSET_ISNULL(mp_cpus) && plat_dr_support_cpu() == 0) ||
1564 1564 use_mp == 0) {
1565 1565 if (use_mp == 0)
1566 1566 cmn_err(CE_CONT, "?***** Not in MP mode\n");
1567 1567 goto done;
1568 1568 }
1569 1569
1570 1570 /*
1571 1571 * perform such initialization as is needed
1572 1572 * to be able to take CPUs on- and off-line.
1573 1573 */
1574 1574 cpu_pause_init();
1575 1575
1576 1576 xc_init_cpu(CPU); /* initialize processor crosscalls */
1577 1577
1578 1578 if (mach_cpucontext_init() != 0)
1579 1579 goto done;
1580 1580
1581 1581 flushes_require_xcalls = 1;
1582 1582
1583 1583 /*
1584 1584 * We lock our affinity to the master CPU to ensure that all slave CPUs
1585 1585 * do their TSC syncs with the same CPU.
1586 1586 */
1587 1587 affinity_set(CPU_CURRENT);
1588 1588
1589 1589 for (who = 0; who < NCPU; who++) {
1590 1590 if (!CPU_IN_SET(mp_cpus, who))
1591 1591 continue;
1592 1592 ASSERT(who != bootcpuid);
1593 1593
1594 1594 mutex_enter(&cpu_lock);
1595 1595 if (start_cpu(who) != 0)
1596 1596 CPUSET_DEL(mp_cpus, who);
1597 1597 cpu_state_change_notify(who, CPU_SETUP);
1598 1598 mutex_exit(&cpu_lock);
1599 1599 }
1600 1600
1601 1601 /* Free the space allocated to hold the microcode file */
1602 1602 ucode_cleanup();
1603 1603
1604 1604 affinity_clear();
1605 1605
1606 1606 mach_cpucontext_fini();
1607 1607
1608 1608 done:
1609 1609 if (get_hwenv() == HW_NATIVE)
1610 1610 workaround_errata_end();
1611 1611 cmi_post_mpstartup();
1612 1612
1613 1613 if (use_mp && ncpus != boot_max_ncpus) {
1614 1614 cmn_err(CE_NOTE,
1615 1615 "System detected %d cpus, but "
1616 1616 "only %d cpu(s) were enabled during boot.",
1617 1617 boot_max_ncpus, ncpus);
1618 1618 cmn_err(CE_NOTE,
1619 1619 "Use \"boot-ncpus\" parameter to enable more CPU(s). "
1620 1620 "See eeprom(1M).");
1621 1621 }
1622 1622 }
1623 1623
1624 1624 int
1625 1625 mp_cpu_configure(int cpuid)
1626 1626 {
1627 1627 cpu_t *cp;
1628 1628
1629 1629 if (use_mp == 0 || plat_dr_support_cpu() == 0) {
1630 1630 return (ENOTSUP);
1631 1631 }
1632 1632
1633 1633 cp = cpu_get(cpuid);
1634 1634 if (cp != NULL) {
1635 1635 return (EALREADY);
1636 1636 }
1637 1637
1638 1638 /*
1639 1639 * Check if there's at least a Mbyte of kmem available
1640 1640 * before attempting to start the cpu.
1641 1641 */
1642 1642 if (kmem_avail() < 1024 * 1024) {
1643 1643 /*
1644 1644 * Kick off a reap in case that helps us with
1645 1645 * later attempts ..
1646 1646 */
1647 1647 kmem_reap();
1648 1648 return (ENOMEM);
1649 1649 }
1650 1650
1651 1651 cp = mp_cpu_configure_common(cpuid, B_FALSE);
1652 1652 ASSERT(cp != NULL && cpu_get(cpuid) == cp);
1653 1653
1654 1654 return (cp != NULL ? 0 : EAGAIN);
1655 1655 }
1656 1656
1657 1657 int
1658 1658 mp_cpu_unconfigure(int cpuid)
1659 1659 {
1660 1660 cpu_t *cp;
1661 1661
1662 1662 if (use_mp == 0 || plat_dr_support_cpu() == 0) {
1663 1663 return (ENOTSUP);
1664 1664 } else if (cpuid < 0 || cpuid >= max_ncpus) {
1665 1665 return (EINVAL);
1666 1666 }
1667 1667
1668 1668 cp = cpu_get(cpuid);
1669 1669 if (cp == NULL) {
1670 1670 return (ENODEV);
1671 1671 }
1672 1672 mp_cpu_unconfigure_common(cp, 0);
1673 1673
1674 1674 return (0);
1675 1675 }
1676 1676
1677 1677 /*
1678 1678 * Startup function for 'other' CPUs (besides boot cpu).
1679 1679 * Called from real_mode_start.
1680 1680 *
1681 1681 * WARNING: until CPU_READY is set, mp_startup_common and routines called by
1682 1682 * mp_startup_common should not call routines (e.g. kmem_free) that could call
1683 1683 * hat_unload which requires CPU_READY to be set.
1684 1684 */
1685 1685 static void
1686 1686 mp_startup_common(boolean_t boot)
1687 1687 {
1688 1688 cpu_t *cp = CPU;
1689 1689 uchar_t new_x86_featureset[BT_SIZEOFMAP(NUM_X86_FEATURES)];
1690 1690 extern void cpu_event_init_cpu(cpu_t *);
1691 1691
1692 1692 /*
1693 1693 * We need to get TSC on this proc synced (i.e., any delta
1694 1694 * from cpu0 accounted for) as soon as we can, because many
1695 1695 * many things use gethrtime/pc_gethrestime, including
1696 1696 * interrupts, cmn_err, etc. Before we can do that, we want to
1697 1697 * clear TSC if we're on a buggy Sandy/Ivy Bridge CPU, so do that
1698 1698 * right away.
1699 1699 */
1700 1700 bzero(new_x86_featureset, BT_SIZEOFMAP(NUM_X86_FEATURES));
1701 1701 cpuid_pass1(cp, new_x86_featureset);
1702 1702
1703 1703 if (boot && get_hwenv() == HW_NATIVE &&
1704 1704 cpuid_getvendor(CPU) == X86_VENDOR_Intel &&
1705 1705 cpuid_getfamily(CPU) == 6 &&
1706 1706 (cpuid_getmodel(CPU) == 0x2d || cpuid_getmodel(CPU) == 0x3e) &&
1707 1707 is_x86_feature(new_x86_featureset, X86FSET_TSC)) {
1708 1708 (void) wrmsr(REG_TSC, 0UL);
1709 1709 }
1710 1710
1711 1711 /* Let the control CPU continue into tsc_sync_master() */
1712 1712 mp_startup_signal(&procset_slave, cp->cpu_id);
1713 1713
1714 1714 #ifndef __xpv
1715 1715 if (tsc_gethrtime_enable)
1716 1716 tsc_sync_slave();
1717 1717 #endif
1718 1718
1719 1719 /*
1720 1720 * Once this was done from assembly, but it's safer here; if
1721 1721 * it blocks, we need to be able to swtch() to and from, and
1722 1722 * since we get here by calling t_pc, we need to do that call
1723 1723 * before swtch() overwrites it.
1724 1724 */
1725 1725 (void) (*ap_mlsetup)();
1726 1726
1727 1727 #ifndef __xpv
1728 1728 /*
1729 1729 * Program this cpu's PAT
1730 1730 */
1731 1731 pat_sync();
1732 1732 #endif
1733 1733
1734 1734 /*
1735 1735 * Set up TSC_AUX to contain the cpuid for this processor
1736 1736 * for the rdtscp instruction.
1737 1737 */
1738 1738 if (is_x86_feature(x86_featureset, X86FSET_TSCP))
1739 1739 (void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id);
1740 1740
1741 1741 /*
1742 1742 * Initialize this CPU's syscall handlers
1743 1743 */
1744 1744 init_cpu_syscall(cp);
1745 1745
1746 1746 /*
1747 1747 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
1748 1748 * highest level at which a routine is permitted to block on
1749 1749 * an adaptive mutex (allows for cpu poke interrupt in case
1750 1750 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
1751 1751 * device interrupts that may end up in the hat layer issuing cross
1752 1752 * calls before CPU_READY is set.
1753 1753 */
1754 1754 splx(ipltospl(LOCK_LEVEL));
1755 1755 sti();
1756 1756
1757 1757 /*
1758 1758 * Do a sanity check to make sure this new CPU is a sane thing
1759 1759 * to add to the collection of processors running this system.
1760 1760 *
1761 1761 * XXX Clearly this needs to get more sophisticated, if x86
1762 1762 * systems start to get built out of heterogenous CPUs; as is
1763 1763 * likely to happen once the number of processors in a configuration
1764 1764 * gets large enough.
1765 1765 */
1766 1766 if (compare_x86_featureset(x86_featureset, new_x86_featureset) ==
1767 1767 B_FALSE) {
1768 1768 cmn_err(CE_CONT, "cpu%d: featureset\n", cp->cpu_id);
1769 1769 print_x86_featureset(new_x86_featureset);
1770 1770 cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
1771 1771 }
1772 1772
1773 1773 /*
1774 1774 * There exists a small subset of systems which expose differing
1775 1775 * MWAIT/MONITOR support between CPUs. If MWAIT support is absent from
1776 1776 * the boot CPU, but is found on a later CPU, the system continues to
1777 1777 * operate as if no MWAIT support is available.
1778 1778 *
1779 1779 * The reverse case, where MWAIT is available on the boot CPU but not
1780 1780 * on a subsequently initialized CPU, is not presently allowed and will
1781 1781 * result in a panic.
1782 1782 */
1783 1783 if (is_x86_feature(x86_featureset, X86FSET_MWAIT) !=
1784 1784 is_x86_feature(new_x86_featureset, X86FSET_MWAIT)) {
1785 1785 if (!is_x86_feature(x86_featureset, X86FSET_MWAIT)) {
1786 1786 remove_x86_feature(new_x86_featureset, X86FSET_MWAIT);
1787 1787 } else {
1788 1788 panic("unsupported mixed cpu mwait support detected");
1789 1789 }
1790 1790 }
1791 1791
1792 1792 /*
1793 1793 * We could be more sophisticated here, and just mark the CPU
1794 1794 * as "faulted" but at this point we'll opt for the easier
1795 1795 * answer of dying horribly. Provided the boot cpu is ok,
1796 1796 * the system can be recovered by booting with use_mp set to zero.
1797 1797 */
1798 1798 if (workaround_errata(cp) != 0)
1799 1799 panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
1800 1800
1801 1801 /*
1802 1802 * We can touch cpu_flags here without acquiring the cpu_lock here
1803 1803 * because the cpu_lock is held by the control CPU which is running
1804 1804 * mp_start_cpu_common().
1805 1805 * Need to clear CPU_QUIESCED flag before calling any function which
1806 1806 * may cause thread context switching, such as kmem_alloc() etc.
1807 1807 * The idle thread checks for CPU_QUIESCED flag and loops for ever if
1808 1808 * it's set. So the startup thread may have no chance to switch back
1809 1809 * again if it's switched away with CPU_QUIESCED set.
1810 1810 */
1811 1811 cp->cpu_flags &= ~(CPU_POWEROFF | CPU_QUIESCED);
1812 1812
1813 1813 enable_pcid();
1814 1814
1815 1815 /*
1816 1816 * Setup this processor for XSAVE.
1817 1817 */
1818 1818 if (fp_save_mech == FP_XSAVE) {
1819 1819 xsave_setup_msr(cp);
1820 1820 }
1821 1821
1822 1822 cpuid_pass2(cp);
1823 1823 cpuid_pass3(cp);
1824 1824 cpuid_pass4(cp, NULL);
1825 1825
1826 1826 /*
1827 1827 * Correct cpu_idstr and cpu_brandstr on target CPU after
1828 1828 * cpuid_pass1() is done.
1829 1829 */
1830 1830 (void) cpuid_getidstr(cp, cp->cpu_idstr, CPU_IDSTRLEN);
1831 1831 (void) cpuid_getbrandstr(cp, cp->cpu_brandstr, CPU_IDSTRLEN);
1832 1832
1833 1833 cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_EXISTS;
1834 1834
1835 1835 post_startup_cpu_fixups();
1836 1836
1837 1837 cpu_event_init_cpu(cp);
1838 1838
1839 1839 /*
1840 1840 * Enable preemption here so that contention for any locks acquired
1841 1841 * later in mp_startup_common may be preempted if the thread owning
1842 1842 * those locks is continuously executing on other CPUs (for example,
1843 1843 * this CPU must be preemptible to allow other CPUs to pause it during
1844 1844 * their startup phases). It's safe to enable preemption here because
1845 1845 * the CPU state is pretty-much fully constructed.
1846 1846 */
1847 1847 curthread->t_preempt = 0;
1848 1848
1849 1849 /* The base spl should still be at LOCK LEVEL here */
1850 1850 ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL));
1851 1851 set_base_spl(); /* Restore the spl to its proper value */
1852 1852
1853 1853 pghw_physid_create(cp);
1854 1854 /*
1855 1855 * Delegate initialization tasks, which need to access the cpu_lock,
1856 1856 * to mp_start_cpu_common() because we can't acquire the cpu_lock here
1857 1857 * during CPU DR operations.
1858 1858 */
1859 1859 mp_startup_signal(&procset_slave, cp->cpu_id);
1860 1860 mp_startup_wait(&procset_master, cp->cpu_id);
1861 1861 pg_cmt_cpu_startup(cp);
1862 1862
1863 1863 if (boot) {
1864 1864 mutex_enter(&cpu_lock);
1865 1865 cp->cpu_flags &= ~CPU_OFFLINE;
1866 1866 cpu_enable_intr(cp);
1867 1867 cpu_add_active(cp);
1868 1868 mutex_exit(&cpu_lock);
1869 1869 }
1870 1870
1871 1871 /* Enable interrupts */
1872 1872 (void) spl0();
1873 1873
1874 1874 /*
1875 1875 * Fill out cpu_ucode_info. Update microcode if necessary.
1876 1876 */
1877 1877 ucode_check(cp);
1878 1878
1879 1879 #ifndef __xpv
1880 1880 {
1881 1881 /*
1882 1882 * Set up the CPU module for this CPU. This can't be done
1883 1883 * before this CPU is made CPU_READY, because we may (in
1884 1884 * heterogeneous systems) need to go load another CPU module.
1885 1885 * The act of attempting to load a module may trigger a
1886 1886 * cross-call, which will ASSERT unless this cpu is CPU_READY.
1887 1887 */
1888 1888 cmi_hdl_t hdl;
1889 1889
1890 1890 if ((hdl = cmi_init(CMI_HDL_NATIVE, cmi_ntv_hwchipid(CPU),
1891 1891 cmi_ntv_hwcoreid(CPU), cmi_ntv_hwstrandid(CPU))) != NULL) {
1892 1892 if (is_x86_feature(x86_featureset, X86FSET_MCA))
1893 1893 cmi_mca_init(hdl);
1894 1894 cp->cpu_m.mcpu_cmi_hdl = hdl;
1895 1895 }
1896 1896 }
1897 1897 #endif /* __xpv */
1898 1898
1899 1899 if (boothowto & RB_DEBUG)
1900 1900 kdi_cpu_init();
1901 1901
1902 1902 /*
1903 1903 * Setting the bit in cpu_ready_set must be the last operation in
1904 1904 * processor initialization; the boot CPU will continue to boot once
1905 1905 * it sees this bit set for all active CPUs.
1906 1906 */
1907 1907 CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1908 1908
1909 1909 (void) mach_cpu_create_device_node(cp, NULL);
1910 1910
1911 1911 cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr);
1912 1912 cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr);
1913 1913 cmn_err(CE_CONT, "?cpu%d initialization complete - online\n",
1914 1914 cp->cpu_id);
1915 1915
1916 1916 /*
1917 1917 * Now we are done with the startup thread, so free it up.
1918 1918 */
1919 1919 thread_exit();
1920 1920 panic("mp_startup: cannot return");
1921 1921 /*NOTREACHED*/
1922 1922 }
1923 1923
1924 1924 /*
1925 1925 * Startup function for 'other' CPUs at boot time (besides boot cpu).
1926 1926 */
1927 1927 static void
1928 1928 mp_startup_boot(void)
1929 1929 {
1930 1930 mp_startup_common(B_TRUE);
1931 1931 }
1932 1932
1933 1933 /*
1934 1934 * Startup function for hotplug CPUs at runtime.
1935 1935 */
1936 1936 void
1937 1937 mp_startup_hotplug(void)
1938 1938 {
1939 1939 mp_startup_common(B_FALSE);
1940 1940 }
1941 1941
1942 1942 /*
1943 1943 * Start CPU on user request.
1944 1944 */
1945 1945 /* ARGSUSED */
1946 1946 int
1947 1947 mp_cpu_start(struct cpu *cp)
1948 1948 {
1949 1949 ASSERT(MUTEX_HELD(&cpu_lock));
1950 1950 return (0);
1951 1951 }
1952 1952
1953 1953 /*
1954 1954 * Stop CPU on user request.
1955 1955 */
1956 1956 int
1957 1957 mp_cpu_stop(struct cpu *cp)
1958 1958 {
1959 1959 extern int cbe_psm_timer_mode;
1960 1960 ASSERT(MUTEX_HELD(&cpu_lock));
1961 1961
1962 1962 #ifdef __xpv
1963 1963 /*
1964 1964 * We can't offline vcpu0.
1965 1965 */
1966 1966 if (cp->cpu_id == 0)
1967 1967 return (EBUSY);
1968 1968 #endif
1969 1969
1970 1970 /*
1971 1971 * If TIMER_PERIODIC mode is used, CPU0 is the one running it;
1972 1972 * can't stop it. (This is true only for machines with no TSC.)
1973 1973 */
1974 1974
1975 1975 if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0))
1976 1976 return (EBUSY);
1977 1977
1978 1978 return (0);
1979 1979 }
1980 1980
1981 1981 /*
1982 1982 * Take the specified CPU out of participation in interrupts.
1983 1983 */
1984 1984 int
1985 1985 cpu_disable_intr(struct cpu *cp)
1986 1986 {
1987 1987 if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1988 1988 return (EBUSY);
1989 1989
1990 1990 cp->cpu_flags &= ~CPU_ENABLE;
1991 1991 return (0);
1992 1992 }
1993 1993
1994 1994 /*
1995 1995 * Allow the specified CPU to participate in interrupts.
1996 1996 */
1997 1997 void
1998 1998 cpu_enable_intr(struct cpu *cp)
1999 1999 {
2000 2000 ASSERT(MUTEX_HELD(&cpu_lock));
2001 2001 cp->cpu_flags |= CPU_ENABLE;
2002 2002 psm_enable_intr(cp->cpu_id);
2003 2003 }
2004 2004
2005 2005 void
2006 2006 mp_cpu_faulted_enter(struct cpu *cp)
2007 2007 {
2008 2008 #ifdef __xpv
2009 2009 _NOTE(ARGUNUSED(cp));
2010 2010 #else
2011 2011 cmi_hdl_t hdl = cp->cpu_m.mcpu_cmi_hdl;
2012 2012
2013 2013 if (hdl != NULL) {
2014 2014 cmi_hdl_hold(hdl);
2015 2015 } else {
2016 2016 hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
2017 2017 cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
2018 2018 }
2019 2019 if (hdl != NULL) {
2020 2020 cmi_faulted_enter(hdl);
2021 2021 cmi_hdl_rele(hdl);
2022 2022 }
2023 2023 #endif
2024 2024 }
2025 2025
2026 2026 void
2027 2027 mp_cpu_faulted_exit(struct cpu *cp)
2028 2028 {
2029 2029 #ifdef __xpv
2030 2030 _NOTE(ARGUNUSED(cp));
2031 2031 #else
2032 2032 cmi_hdl_t hdl = cp->cpu_m.mcpu_cmi_hdl;
2033 2033
2034 2034 if (hdl != NULL) {
2035 2035 cmi_hdl_hold(hdl);
2036 2036 } else {
2037 2037 hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
2038 2038 cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
2039 2039 }
2040 2040 if (hdl != NULL) {
2041 2041 cmi_faulted_exit(hdl);
2042 2042 cmi_hdl_rele(hdl);
2043 2043 }
2044 2044 #endif
2045 2045 }
2046 2046
|
↓ open down ↓ |
2046 lines elided |
↑ open up ↑ |
2047 2047 /*
2048 2048 * The following two routines are used as context operators on threads belonging
2049 2049 * to processes with a private LDT (see sysi86). Due to the rarity of such
2050 2050 * processes, these routines are currently written for best code readability and
2051 2051 * organization rather than speed. We could avoid checking x86_featureset at
2052 2052 * every context switch by installing different context ops, depending on
2053 2053 * x86_featureset, at LDT creation time -- one for each combination of fast
2054 2054 * syscall features.
2055 2055 */
2056 2056
2057 -/*ARGSUSED*/
2058 2057 void
2059 -cpu_fast_syscall_disable(void *arg)
2058 +cpu_fast_syscall_disable(void)
2060 2059 {
2061 2060 if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2062 2061 is_x86_feature(x86_featureset, X86FSET_SEP))
2063 2062 cpu_sep_disable();
2064 2063 if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2065 2064 is_x86_feature(x86_featureset, X86FSET_ASYSC))
2066 2065 cpu_asysc_disable();
2067 2066 }
2068 2067
2069 -/*ARGSUSED*/
2070 2068 void
2071 -cpu_fast_syscall_enable(void *arg)
2069 +cpu_fast_syscall_enable(void)
2072 2070 {
2073 2071 if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2074 2072 is_x86_feature(x86_featureset, X86FSET_SEP))
2075 2073 cpu_sep_enable();
2076 2074 if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2077 2075 is_x86_feature(x86_featureset, X86FSET_ASYSC))
2078 2076 cpu_asysc_enable();
2079 2077 }
2080 2078
2081 2079 static void
2082 2080 cpu_sep_enable(void)
2083 2081 {
2084 2082 ASSERT(is_x86_feature(x86_featureset, X86FSET_SEP));
2085 2083 ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2086 2084
2087 2085 wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL);
2088 2086 }
2089 2087
2090 2088 static void
2091 2089 cpu_sep_disable(void)
2092 2090 {
2093 2091 ASSERT(is_x86_feature(x86_featureset, X86FSET_SEP));
2094 2092 ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2095 2093
2096 2094 /*
2097 2095 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
2098 2096 * the sysenter or sysexit instruction to trigger a #gp fault.
2099 2097 */
2100 2098 wrmsr(MSR_INTC_SEP_CS, 0);
2101 2099 }
2102 2100
2103 2101 static void
2104 2102 cpu_asysc_enable(void)
2105 2103 {
2106 2104 ASSERT(is_x86_feature(x86_featureset, X86FSET_ASYSC));
2107 2105 ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2108 2106
2109 2107 wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) |
2110 2108 (uint64_t)(uintptr_t)AMD_EFER_SCE);
2111 2109 }
2112 2110
2113 2111 static void
2114 2112 cpu_asysc_disable(void)
2115 2113 {
2116 2114 ASSERT(is_x86_feature(x86_featureset, X86FSET_ASYSC));
2117 2115 ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2118 2116
2119 2117 /*
2120 2118 * Turn off the SCE (syscall enable) bit in the EFER register. Software
2121 2119 * executing syscall or sysret with this bit off will incur a #ud trap.
2122 2120 */
2123 2121 wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) &
2124 2122 ~((uint64_t)(uintptr_t)AMD_EFER_SCE));
2125 2123 }
|
↓ open down ↓ |
44 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX