1 /* 2 * CDDL HEADER START 3 * 4 * The contents of this file are subject to the terms of the 5 * Common Development and Distribution License (the "License"). 6 * You may not use this file except in compliance with the License. 7 * 8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 9 * or http://www.opensolaris.org/os/licensing. 10 * See the License for the specific language governing permissions 11 * and limitations under the License. 12 * 13 * When distributing Covered Code, include this CDDL HEADER in each 14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE. 15 * If applicable, add the following below this CDDL HEADER, with the 16 * fields enclosed by brackets "[]" replaced with your own identifying 17 * information: Portions Copyright [yyyy] [name of copyright owner] 18 * 19 * CDDL HEADER END 20 */ 21 22 /* 23 * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved. 24 */ 25 /* 26 * Copyright (c) 2010, Intel Corporation. 27 * All rights reserved. 28 */ 29 /* 30 * Copyright (c) 2012, Joyent, Inc. All rights reserved. 31 * Copyright 2013 Nexenta Systems, Inc. All rights reserved. 32 */ 33 34 #include <sys/types.h> 35 #include <sys/thread.h> 36 #include <sys/cpuvar.h> 37 #include <sys/cpu.h> 38 #include <sys/t_lock.h> 39 #include <sys/param.h> 40 #include <sys/proc.h> 41 #include <sys/disp.h> 42 #include <sys/class.h> 43 #include <sys/cmn_err.h> 44 #include <sys/debug.h> 45 #include <sys/note.h> 46 #include <sys/asm_linkage.h> 47 #include <sys/x_call.h> 48 #include <sys/systm.h> 49 #include <sys/var.h> 50 #include <sys/vtrace.h> 51 #include <vm/hat.h> 52 #include <vm/as.h> 53 #include <vm/seg_kmem.h> 54 #include <vm/seg_kp.h> 55 #include <sys/segments.h> 56 #include <sys/kmem.h> 57 #include <sys/stack.h> 58 #include <sys/smp_impldefs.h> 59 #include <sys/x86_archext.h> 60 #include <sys/machsystm.h> 61 #include <sys/traptrace.h> 62 #include <sys/clock.h> 63 #include <sys/cpc_impl.h> 64 #include <sys/pg.h> 65 #include <sys/cmt.h> 66 #include <sys/dtrace.h> 67 #include <sys/archsystm.h> 68 #include <sys/fp.h> 69 #include <sys/reboot.h> 70 #include <sys/kdi_machimpl.h> 71 #include <vm/hat_i86.h> 72 #include <vm/vm_dep.h> 73 #include <sys/memnode.h> 74 #include <sys/pci_cfgspace.h> 75 #include <sys/mach_mmu.h> 76 #include <sys/sysmacros.h> 77 #if defined(__xpv) 78 #include <sys/hypervisor.h> 79 #endif 80 #include <sys/cpu_module.h> 81 #include <sys/ontrap.h> 82 83 struct cpu cpus[1]; /* CPU data */ 84 struct cpu *cpu[NCPU] = {&cpus[0]}; /* pointers to all CPUs */ 85 struct cpu *cpu_free_list; /* list for released CPUs */ 86 cpu_core_t cpu_core[NCPU]; /* cpu_core structures */ 87 88 #define cpu_next_free cpu_prev 89 90 /* 91 * Useful for disabling MP bring-up on a MP capable system. 92 */ 93 int use_mp = 1; 94 95 /* 96 * to be set by a PSM to indicate what cpus 97 * are sitting around on the system. 98 */ 99 cpuset_t mp_cpus; 100 101 /* 102 * This variable is used by the hat layer to decide whether or not 103 * critical sections are needed to prevent race conditions. For sun4m, 104 * this variable is set once enough MP initialization has been done in 105 * order to allow cross calls. 106 */ 107 int flushes_require_xcalls; 108 109 cpuset_t cpu_ready_set; /* initialized in startup() */ 110 111 static void mp_startup_boot(void); 112 static void mp_startup_hotplug(void); 113 114 static void cpu_sep_enable(void); 115 static void cpu_sep_disable(void); 116 static void cpu_asysc_enable(void); 117 static void cpu_asysc_disable(void); 118 119 /* 120 * Init CPU info - get CPU type info for processor_info system call. 121 */ 122 void 123 init_cpu_info(struct cpu *cp) 124 { 125 processor_info_t *pi = &cp->cpu_type_info; 126 127 /* 128 * Get clock-frequency property for the CPU. 129 */ 130 pi->pi_clock = cpu_freq; 131 132 /* 133 * Current frequency in Hz. 134 */ 135 cp->cpu_curr_clock = cpu_freq_hz; 136 137 /* 138 * Supported frequencies. 139 */ 140 if (cp->cpu_supp_freqs == NULL) { 141 cpu_set_supp_freqs(cp, NULL); 142 } 143 144 (void) strcpy(pi->pi_processor_type, "i386"); 145 if (fpu_exists) 146 (void) strcpy(pi->pi_fputypes, "i387 compatible"); 147 148 cp->cpu_idstr = kmem_zalloc(CPU_IDSTRLEN, KM_SLEEP); 149 cp->cpu_brandstr = kmem_zalloc(CPU_IDSTRLEN, KM_SLEEP); 150 151 /* 152 * If called for the BSP, cp is equal to current CPU. 153 * For non-BSPs, cpuid info of cp is not ready yet, so use cpuid info 154 * of current CPU as default values for cpu_idstr and cpu_brandstr. 155 * They will be corrected in mp_startup_common() after cpuid_pass1() 156 * has been invoked on target CPU. 157 */ 158 (void) cpuid_getidstr(CPU, cp->cpu_idstr, CPU_IDSTRLEN); 159 (void) cpuid_getbrandstr(CPU, cp->cpu_brandstr, CPU_IDSTRLEN); 160 } 161 162 /* 163 * Configure syscall support on this CPU. 164 */ 165 /*ARGSUSED*/ 166 void 167 init_cpu_syscall(struct cpu *cp) 168 { 169 kpreempt_disable(); 170 171 #if defined(__amd64) 172 if (is_x86_feature(x86_featureset, X86FSET_MSR) && 173 is_x86_feature(x86_featureset, X86FSET_ASYSC)) { 174 175 #if !defined(__lint) 176 /* 177 * The syscall instruction imposes a certain ordering on 178 * segment selectors, so we double-check that ordering 179 * here. 180 */ 181 ASSERT(KDS_SEL == KCS_SEL + 8); 182 ASSERT(UDS_SEL == U32CS_SEL + 8); 183 ASSERT(UCS_SEL == U32CS_SEL + 16); 184 #endif 185 /* 186 * Turn syscall/sysret extensions on. 187 */ 188 cpu_asysc_enable(); 189 190 /* 191 * Program the magic registers .. 192 */ 193 wrmsr(MSR_AMD_STAR, 194 ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32); 195 wrmsr(MSR_AMD_LSTAR, (uint64_t)(uintptr_t)sys_syscall); 196 wrmsr(MSR_AMD_CSTAR, (uint64_t)(uintptr_t)sys_syscall32); 197 198 /* 199 * This list of flags is masked off the incoming 200 * %rfl when we enter the kernel. 201 */ 202 wrmsr(MSR_AMD_SFMASK, (uint64_t)(uintptr_t)(PS_IE | PS_T)); 203 } 204 #endif 205 206 /* 207 * On 32-bit kernels, we use sysenter/sysexit because it's too 208 * hard to use syscall/sysret, and it is more portable anyway. 209 * 210 * On 64-bit kernels on Nocona machines, the 32-bit syscall 211 * variant isn't available to 32-bit applications, but sysenter is. 212 */ 213 if (is_x86_feature(x86_featureset, X86FSET_MSR) && 214 is_x86_feature(x86_featureset, X86FSET_SEP)) { 215 216 #if !defined(__lint) 217 /* 218 * The sysenter instruction imposes a certain ordering on 219 * segment selectors, so we double-check that ordering 220 * here. See "sysenter" in Intel document 245471-012, "IA-32 221 * Intel Architecture Software Developer's Manual Volume 2: 222 * Instruction Set Reference" 223 */ 224 ASSERT(KDS_SEL == KCS_SEL + 8); 225 226 ASSERT32(UCS_SEL == ((KCS_SEL + 16) | 3)); 227 ASSERT32(UDS_SEL == UCS_SEL + 8); 228 229 ASSERT64(U32CS_SEL == ((KCS_SEL + 16) | 3)); 230 ASSERT64(UDS_SEL == U32CS_SEL + 8); 231 #endif 232 233 cpu_sep_enable(); 234 235 /* 236 * resume() sets this value to the base of the threads stack 237 * via a context handler. 238 */ 239 wrmsr(MSR_INTC_SEP_ESP, 0); 240 wrmsr(MSR_INTC_SEP_EIP, (uint64_t)(uintptr_t)sys_sysenter); 241 } 242 243 kpreempt_enable(); 244 } 245 246 /* 247 * Multiprocessor initialization. 248 * 249 * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the 250 * startup and idle threads for the specified CPU. 251 * Parameter boot is true for boot time operations and is false for CPU 252 * DR operations. 253 */ 254 static struct cpu * 255 mp_cpu_configure_common(int cpun, boolean_t boot) 256 { 257 struct cpu *cp; 258 kthread_id_t tp; 259 caddr_t sp; 260 proc_t *procp; 261 #if !defined(__xpv) 262 extern int idle_cpu_prefer_mwait; 263 extern void cpu_idle_mwait(); 264 #endif 265 extern void idle(); 266 extern void cpu_idle(); 267 268 #ifdef TRAPTRACE 269 trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun]; 270 #endif 271 272 ASSERT(MUTEX_HELD(&cpu_lock)); 273 ASSERT(cpun < NCPU && cpu[cpun] == NULL); 274 275 if (cpu_free_list == NULL) { 276 cp = kmem_zalloc(sizeof (*cp), KM_SLEEP); 277 } else { 278 cp = cpu_free_list; 279 cpu_free_list = cp->cpu_next_free; 280 } 281 282 cp->cpu_m.mcpu_istamp = cpun << 16; 283 284 /* Create per CPU specific threads in the process p0. */ 285 procp = &p0; 286 287 /* 288 * Initialize the dispatcher first. 289 */ 290 disp_cpu_init(cp); 291 292 cpu_vm_data_init(cp); 293 294 /* 295 * Allocate and initialize the startup thread for this CPU. 296 * Interrupt and process switch stacks get allocated later 297 * when the CPU starts running. 298 */ 299 tp = thread_create(NULL, 0, NULL, NULL, 0, procp, 300 TS_STOPPED, maxclsyspri); 301 302 /* 303 * Set state to TS_ONPROC since this thread will start running 304 * as soon as the CPU comes online. 305 * 306 * All the other fields of the thread structure are setup by 307 * thread_create(). 308 */ 309 THREAD_ONPROC(tp, cp); 310 tp->t_preempt = 1; 311 tp->t_bound_cpu = cp; 312 tp->t_affinitycnt = 1; 313 tp->t_cpu = cp; 314 tp->t_disp_queue = cp->cpu_disp; 315 316 /* 317 * Setup thread to start in mp_startup_common. 318 */ 319 sp = tp->t_stk; 320 tp->t_sp = (uintptr_t)(sp - MINFRAME); 321 #if defined(__amd64) 322 tp->t_sp -= STACK_ENTRY_ALIGN; /* fake a call */ 323 #endif 324 /* 325 * Setup thread start entry point for boot or hotplug. 326 */ 327 if (boot) { 328 tp->t_pc = (uintptr_t)mp_startup_boot; 329 } else { 330 tp->t_pc = (uintptr_t)mp_startup_hotplug; 331 } 332 333 cp->cpu_id = cpun; 334 cp->cpu_self = cp; 335 cp->cpu_thread = tp; 336 cp->cpu_lwp = NULL; 337 cp->cpu_dispthread = tp; 338 cp->cpu_dispatch_pri = DISP_PRIO(tp); 339 340 /* 341 * cpu_base_spl must be set explicitly here to prevent any blocking 342 * operations in mp_startup_common from causing the spl of the cpu 343 * to drop to 0 (allowing device interrupts before we're ready) in 344 * resume(). 345 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY. 346 * As an extra bit of security on DEBUG kernels, this is enforced with 347 * an assertion in mp_startup_common() -- before cpu_base_spl is set 348 * to its proper value. 349 */ 350 cp->cpu_base_spl = ipltospl(LOCK_LEVEL); 351 352 /* 353 * Now, initialize per-CPU idle thread for this CPU. 354 */ 355 tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1); 356 357 cp->cpu_idle_thread = tp; 358 359 tp->t_preempt = 1; 360 tp->t_bound_cpu = cp; 361 tp->t_affinitycnt = 1; 362 tp->t_cpu = cp; 363 tp->t_disp_queue = cp->cpu_disp; 364 365 /* 366 * Bootstrap the CPU's PG data 367 */ 368 pg_cpu_bootstrap(cp); 369 370 /* 371 * Perform CPC initialization on the new CPU. 372 */ 373 kcpc_hw_init(cp); 374 375 /* 376 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2 377 * for each CPU. 378 */ 379 setup_vaddr_for_ppcopy(cp); 380 381 /* 382 * Allocate page for new GDT and initialize from current GDT. 383 */ 384 #if !defined(__lint) 385 ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE); 386 #endif 387 cp->cpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP); 388 bcopy(CPU->cpu_gdt, cp->cpu_gdt, (sizeof (*cp->cpu_gdt) * NGDT)); 389 390 #if defined(__i386) 391 /* 392 * setup kernel %gs. 393 */ 394 set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA, 395 SEL_KPL, 0, 1); 396 #endif 397 398 /* 399 * If we have more than one node, each cpu gets a copy of IDT 400 * local to its node. If this is a Pentium box, we use cpu 0's 401 * IDT. cpu 0's IDT has been made read-only to workaround the 402 * cmpxchgl register bug 403 */ 404 if (system_hardware.hd_nodes && x86_type != X86_TYPE_P5) { 405 #if !defined(__lint) 406 ASSERT((sizeof (*CPU->cpu_idt) * NIDT) <= PAGESIZE); 407 #endif 408 cp->cpu_idt = kmem_zalloc(PAGESIZE, KM_SLEEP); 409 bcopy(CPU->cpu_idt, cp->cpu_idt, PAGESIZE); 410 } else { 411 cp->cpu_idt = CPU->cpu_idt; 412 } 413 414 /* 415 * alloc space for cpuid info 416 */ 417 cpuid_alloc_space(cp); 418 #if !defined(__xpv) 419 if (is_x86_feature(x86_featureset, X86FSET_MWAIT) && 420 idle_cpu_prefer_mwait) { 421 cp->cpu_m.mcpu_mwait = cpuid_mwait_alloc(cp); 422 cp->cpu_m.mcpu_idle_cpu = cpu_idle_mwait; 423 } else 424 #endif 425 cp->cpu_m.mcpu_idle_cpu = cpu_idle; 426 427 init_cpu_info(cp); 428 429 /* 430 * alloc space for ucode_info 431 */ 432 ucode_alloc_space(cp); 433 xc_init_cpu(cp); 434 hat_cpu_online(cp); 435 436 #ifdef TRAPTRACE 437 /* 438 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers 439 */ 440 ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP); 441 ttc->ttc_next = ttc->ttc_first; 442 ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize; 443 #endif 444 445 /* 446 * Record that we have another CPU. 447 */ 448 /* 449 * Initialize the interrupt threads for this CPU 450 */ 451 cpu_intr_alloc(cp, NINTR_THREADS); 452 453 cp->cpu_flags = CPU_OFFLINE | CPU_QUIESCED | CPU_POWEROFF; 454 cpu_set_state(cp); 455 456 /* 457 * Add CPU to list of available CPUs. It'll be on the active list 458 * after mp_startup_common(). 459 */ 460 cpu_add_unit(cp); 461 462 return (cp); 463 } 464 465 /* 466 * Undo what was done in mp_cpu_configure_common 467 */ 468 static void 469 mp_cpu_unconfigure_common(struct cpu *cp, int error) 470 { 471 ASSERT(MUTEX_HELD(&cpu_lock)); 472 473 /* 474 * Remove the CPU from the list of available CPUs. 475 */ 476 cpu_del_unit(cp->cpu_id); 477 478 if (error == ETIMEDOUT) { 479 /* 480 * The cpu was started, but never *seemed* to run any 481 * code in the kernel; it's probably off spinning in its 482 * own private world, though with potential references to 483 * our kmem-allocated IDTs and GDTs (for example). 484 * 485 * Worse still, it may actually wake up some time later, 486 * so rather than guess what it might or might not do, we 487 * leave the fundamental data structures intact. 488 */ 489 cp->cpu_flags = 0; 490 return; 491 } 492 493 /* 494 * At this point, the only threads bound to this CPU should 495 * special per-cpu threads: it's idle thread, it's pause threads, 496 * and it's interrupt threads. Clean these up. 497 */ 498 cpu_destroy_bound_threads(cp); 499 cp->cpu_idle_thread = NULL; 500 501 /* 502 * Free the interrupt stack. 503 */ 504 segkp_release(segkp, 505 cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME))); 506 cp->cpu_intr_stack = NULL; 507 508 #ifdef TRAPTRACE 509 /* 510 * Discard the trap trace buffer 511 */ 512 { 513 trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id]; 514 515 kmem_free((void *)ttc->ttc_first, trap_trace_bufsize); 516 ttc->ttc_first = NULL; 517 } 518 #endif 519 520 hat_cpu_offline(cp); 521 522 ucode_free_space(cp); 523 524 /* Free CPU ID string and brand string. */ 525 if (cp->cpu_idstr) { 526 kmem_free(cp->cpu_idstr, CPU_IDSTRLEN); 527 cp->cpu_idstr = NULL; 528 } 529 if (cp->cpu_brandstr) { 530 kmem_free(cp->cpu_brandstr, CPU_IDSTRLEN); 531 cp->cpu_brandstr = NULL; 532 } 533 534 #if !defined(__xpv) 535 if (cp->cpu_m.mcpu_mwait != NULL) { 536 cpuid_mwait_free(cp); 537 cp->cpu_m.mcpu_mwait = NULL; 538 } 539 #endif 540 cpuid_free_space(cp); 541 542 if (cp->cpu_idt != CPU->cpu_idt) 543 kmem_free(cp->cpu_idt, PAGESIZE); 544 cp->cpu_idt = NULL; 545 546 kmem_free(cp->cpu_gdt, PAGESIZE); 547 cp->cpu_gdt = NULL; 548 549 if (cp->cpu_supp_freqs != NULL) { 550 size_t len = strlen(cp->cpu_supp_freqs) + 1; 551 kmem_free(cp->cpu_supp_freqs, len); 552 cp->cpu_supp_freqs = NULL; 553 } 554 555 teardown_vaddr_for_ppcopy(cp); 556 557 kcpc_hw_fini(cp); 558 559 cp->cpu_dispthread = NULL; 560 cp->cpu_thread = NULL; /* discarded by cpu_destroy_bound_threads() */ 561 562 cpu_vm_data_destroy(cp); 563 564 xc_fini_cpu(cp); 565 disp_cpu_fini(cp); 566 567 ASSERT(cp != CPU0); 568 bzero(cp, sizeof (*cp)); 569 cp->cpu_next_free = cpu_free_list; 570 cpu_free_list = cp; 571 } 572 573 /* 574 * Apply workarounds for known errata, and warn about those that are absent. 575 * 576 * System vendors occasionally create configurations which contain different 577 * revisions of the CPUs that are almost but not exactly the same. At the 578 * time of writing, this meant that their clock rates were the same, their 579 * feature sets were the same, but the required workaround were -not- 580 * necessarily the same. So, this routine is invoked on -every- CPU soon 581 * after starting to make sure that the resulting system contains the most 582 * pessimal set of workarounds needed to cope with *any* of the CPUs in the 583 * system. 584 * 585 * workaround_errata is invoked early in mlsetup() for CPU 0, and in 586 * mp_startup_common() for all slave CPUs. Slaves process workaround_errata 587 * prior to acknowledging their readiness to the master, so this routine will 588 * never be executed by multiple CPUs in parallel, thus making updates to 589 * global data safe. 590 * 591 * These workarounds are based on Rev 3.57 of the Revision Guide for 592 * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005. 593 */ 594 595 #if defined(OPTERON_ERRATUM_88) 596 int opteron_erratum_88; /* if non-zero -> at least one cpu has it */ 597 #endif 598 599 #if defined(OPTERON_ERRATUM_91) 600 int opteron_erratum_91; /* if non-zero -> at least one cpu has it */ 601 #endif 602 603 #if defined(OPTERON_ERRATUM_93) 604 int opteron_erratum_93; /* if non-zero -> at least one cpu has it */ 605 #endif 606 607 #if defined(OPTERON_ERRATUM_95) 608 int opteron_erratum_95; /* if non-zero -> at least one cpu has it */ 609 #endif 610 611 #if defined(OPTERON_ERRATUM_100) 612 int opteron_erratum_100; /* if non-zero -> at least one cpu has it */ 613 #endif 614 615 #if defined(OPTERON_ERRATUM_108) 616 int opteron_erratum_108; /* if non-zero -> at least one cpu has it */ 617 #endif 618 619 #if defined(OPTERON_ERRATUM_109) 620 int opteron_erratum_109; /* if non-zero -> at least one cpu has it */ 621 #endif 622 623 #if defined(OPTERON_ERRATUM_121) 624 int opteron_erratum_121; /* if non-zero -> at least one cpu has it */ 625 #endif 626 627 #if defined(OPTERON_ERRATUM_122) 628 int opteron_erratum_122; /* if non-zero -> at least one cpu has it */ 629 #endif 630 631 #if defined(OPTERON_ERRATUM_123) 632 int opteron_erratum_123; /* if non-zero -> at least one cpu has it */ 633 #endif 634 635 #if defined(OPTERON_ERRATUM_131) 636 int opteron_erratum_131; /* if non-zero -> at least one cpu has it */ 637 #endif 638 639 #if defined(OPTERON_WORKAROUND_6336786) 640 int opteron_workaround_6336786; /* non-zero -> WA relevant and applied */ 641 int opteron_workaround_6336786_UP = 0; /* Not needed for UP */ 642 #endif 643 644 #if defined(OPTERON_WORKAROUND_6323525) 645 int opteron_workaround_6323525; /* if non-zero -> at least one cpu has it */ 646 #endif 647 648 #if defined(OPTERON_ERRATUM_298) 649 int opteron_erratum_298; 650 #endif 651 652 #if defined(OPTERON_ERRATUM_721) 653 int opteron_erratum_721; 654 #endif 655 656 static void 657 workaround_warning(cpu_t *cp, uint_t erratum) 658 { 659 cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u", 660 cp->cpu_id, erratum); 661 } 662 663 static void 664 workaround_applied(uint_t erratum) 665 { 666 if (erratum > 1000000) 667 cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n", 668 erratum); 669 else 670 cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n", 671 erratum); 672 } 673 674 static void 675 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error) 676 { 677 cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d", 678 cp->cpu_id, rw, msr, error); 679 } 680 681 /* 682 * Determine the number of nodes in a Hammer / Greyhound / Griffin family 683 * system. 684 */ 685 static uint_t 686 opteron_get_nnodes(void) 687 { 688 static uint_t nnodes = 0; 689 690 if (nnodes == 0) { 691 #ifdef DEBUG 692 uint_t family; 693 694 /* 695 * This routine uses a PCI config space based mechanism 696 * for retrieving the number of nodes in the system. 697 * Device 24, function 0, offset 0x60 as used here is not 698 * AMD processor architectural, and may not work on processor 699 * families other than those listed below. 700 * 701 * Callers of this routine must ensure that we're running on 702 * a processor which supports this mechanism. 703 * The assertion below is meant to catch calls on unsupported 704 * processors. 705 */ 706 family = cpuid_getfamily(CPU); 707 ASSERT(family == 0xf || family == 0x10 || family == 0x11); 708 #endif /* DEBUG */ 709 710 /* 711 * Obtain the number of nodes in the system from 712 * bits [6:4] of the Node ID register on node 0. 713 * 714 * The actual node count is NodeID[6:4] + 1 715 * 716 * The Node ID register is accessed via function 0, 717 * offset 0x60. Node 0 is device 24. 718 */ 719 nnodes = ((pci_getl_func(0, 24, 0, 0x60) & 0x70) >> 4) + 1; 720 } 721 return (nnodes); 722 } 723 724 uint_t 725 do_erratum_298(struct cpu *cpu) 726 { 727 static int osvwrc = -3; 728 extern int osvw_opteron_erratum(cpu_t *, uint_t); 729 730 /* 731 * L2 Eviction May Occur During Processor Operation To Set 732 * Accessed or Dirty Bit. 733 */ 734 if (osvwrc == -3) { 735 osvwrc = osvw_opteron_erratum(cpu, 298); 736 } else { 737 /* osvw return codes should be consistent for all cpus */ 738 ASSERT(osvwrc == osvw_opteron_erratum(cpu, 298)); 739 } 740 741 switch (osvwrc) { 742 case 0: /* erratum is not present: do nothing */ 743 break; 744 case 1: /* erratum is present: BIOS workaround applied */ 745 /* 746 * check if workaround is actually in place and issue warning 747 * if not. 748 */ 749 if (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) || 750 ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0)) { 751 #if defined(OPTERON_ERRATUM_298) 752 opteron_erratum_298++; 753 #else 754 workaround_warning(cpu, 298); 755 return (1); 756 #endif 757 } 758 break; 759 case -1: /* cannot determine via osvw: check cpuid */ 760 if ((cpuid_opteron_erratum(cpu, 298) > 0) && 761 (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) || 762 ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0))) { 763 #if defined(OPTERON_ERRATUM_298) 764 opteron_erratum_298++; 765 #else 766 workaround_warning(cpu, 298); 767 return (1); 768 #endif 769 } 770 break; 771 } 772 return (0); 773 } 774 775 uint_t 776 workaround_errata(struct cpu *cpu) 777 { 778 uint_t missing = 0; 779 780 ASSERT(cpu == CPU); 781 782 /*LINTED*/ 783 if (cpuid_opteron_erratum(cpu, 88) > 0) { 784 /* 785 * SWAPGS May Fail To Read Correct GS Base 786 */ 787 #if defined(OPTERON_ERRATUM_88) 788 /* 789 * The workaround is an mfence in the relevant assembler code 790 */ 791 opteron_erratum_88++; 792 #else 793 workaround_warning(cpu, 88); 794 missing++; 795 #endif 796 } 797 798 if (cpuid_opteron_erratum(cpu, 91) > 0) { 799 /* 800 * Software Prefetches May Report A Page Fault 801 */ 802 #if defined(OPTERON_ERRATUM_91) 803 /* 804 * fix is in trap.c 805 */ 806 opteron_erratum_91++; 807 #else 808 workaround_warning(cpu, 91); 809 missing++; 810 #endif 811 } 812 813 if (cpuid_opteron_erratum(cpu, 93) > 0) { 814 /* 815 * RSM Auto-Halt Restart Returns to Incorrect RIP 816 */ 817 #if defined(OPTERON_ERRATUM_93) 818 /* 819 * fix is in trap.c 820 */ 821 opteron_erratum_93++; 822 #else 823 workaround_warning(cpu, 93); 824 missing++; 825 #endif 826 } 827 828 /*LINTED*/ 829 if (cpuid_opteron_erratum(cpu, 95) > 0) { 830 /* 831 * RET Instruction May Return to Incorrect EIP 832 */ 833 #if defined(OPTERON_ERRATUM_95) 834 #if defined(_LP64) 835 /* 836 * Workaround this by ensuring that 32-bit user code and 837 * 64-bit kernel code never occupy the same address 838 * range mod 4G. 839 */ 840 if (_userlimit32 > 0xc0000000ul) 841 *(uintptr_t *)&_userlimit32 = 0xc0000000ul; 842 843 /*LINTED*/ 844 ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u); 845 opteron_erratum_95++; 846 #endif /* _LP64 */ 847 #else 848 workaround_warning(cpu, 95); 849 missing++; 850 #endif 851 } 852 853 if (cpuid_opteron_erratum(cpu, 100) > 0) { 854 /* 855 * Compatibility Mode Branches Transfer to Illegal Address 856 */ 857 #if defined(OPTERON_ERRATUM_100) 858 /* 859 * fix is in trap.c 860 */ 861 opteron_erratum_100++; 862 #else 863 workaround_warning(cpu, 100); 864 missing++; 865 #endif 866 } 867 868 /*LINTED*/ 869 if (cpuid_opteron_erratum(cpu, 108) > 0) { 870 /* 871 * CPUID Instruction May Return Incorrect Model Number In 872 * Some Processors 873 */ 874 #if defined(OPTERON_ERRATUM_108) 875 /* 876 * (Our cpuid-handling code corrects the model number on 877 * those processors) 878 */ 879 #else 880 workaround_warning(cpu, 108); 881 missing++; 882 #endif 883 } 884 885 /*LINTED*/ 886 if (cpuid_opteron_erratum(cpu, 109) > 0) do { 887 /* 888 * Certain Reverse REP MOVS May Produce Unpredictable Behavior 889 */ 890 #if defined(OPTERON_ERRATUM_109) 891 /* 892 * The "workaround" is to print a warning to upgrade the BIOS 893 */ 894 uint64_t value; 895 const uint_t msr = MSR_AMD_PATCHLEVEL; 896 int err; 897 898 if ((err = checked_rdmsr(msr, &value)) != 0) { 899 msr_warning(cpu, "rd", msr, err); 900 workaround_warning(cpu, 109); 901 missing++; 902 } 903 if (value == 0) 904 opteron_erratum_109++; 905 #else 906 workaround_warning(cpu, 109); 907 missing++; 908 #endif 909 /*CONSTANTCONDITION*/ 910 } while (0); 911 912 /*LINTED*/ 913 if (cpuid_opteron_erratum(cpu, 121) > 0) { 914 /* 915 * Sequential Execution Across Non_Canonical Boundary Caused 916 * Processor Hang 917 */ 918 #if defined(OPTERON_ERRATUM_121) 919 #if defined(_LP64) 920 /* 921 * Erratum 121 is only present in long (64 bit) mode. 922 * Workaround is to include the page immediately before the 923 * va hole to eliminate the possibility of system hangs due to 924 * sequential execution across the va hole boundary. 925 */ 926 if (opteron_erratum_121) 927 opteron_erratum_121++; 928 else { 929 if (hole_start) { 930 hole_start -= PAGESIZE; 931 } else { 932 /* 933 * hole_start not yet initialized by 934 * mmu_init. Initialize hole_start 935 * with value to be subtracted. 936 */ 937 hole_start = PAGESIZE; 938 } 939 opteron_erratum_121++; 940 } 941 #endif /* _LP64 */ 942 #else 943 workaround_warning(cpu, 121); 944 missing++; 945 #endif 946 } 947 948 /*LINTED*/ 949 if (cpuid_opteron_erratum(cpu, 122) > 0) do { 950 /* 951 * TLB Flush Filter May Cause Coherency Problem in 952 * Multiprocessor Systems 953 */ 954 #if defined(OPTERON_ERRATUM_122) 955 uint64_t value; 956 const uint_t msr = MSR_AMD_HWCR; 957 int error; 958 959 /* 960 * Erratum 122 is only present in MP configurations (multi-core 961 * or multi-processor). 962 */ 963 #if defined(__xpv) 964 if (!DOMAIN_IS_INITDOMAIN(xen_info)) 965 break; 966 if (!opteron_erratum_122 && xpv_nr_phys_cpus() == 1) 967 break; 968 #else 969 if (!opteron_erratum_122 && opteron_get_nnodes() == 1 && 970 cpuid_get_ncpu_per_chip(cpu) == 1) 971 break; 972 #endif 973 /* disable TLB Flush Filter */ 974 975 if ((error = checked_rdmsr(msr, &value)) != 0) { 976 msr_warning(cpu, "rd", msr, error); 977 workaround_warning(cpu, 122); 978 missing++; 979 } else { 980 value |= (uint64_t)AMD_HWCR_FFDIS; 981 if ((error = checked_wrmsr(msr, value)) != 0) { 982 msr_warning(cpu, "wr", msr, error); 983 workaround_warning(cpu, 122); 984 missing++; 985 } 986 } 987 opteron_erratum_122++; 988 #else 989 workaround_warning(cpu, 122); 990 missing++; 991 #endif 992 /*CONSTANTCONDITION*/ 993 } while (0); 994 995 /*LINTED*/ 996 if (cpuid_opteron_erratum(cpu, 123) > 0) do { 997 /* 998 * Bypassed Reads May Cause Data Corruption of System Hang in 999 * Dual Core Processors 1000 */ 1001 #if defined(OPTERON_ERRATUM_123) 1002 uint64_t value; 1003 const uint_t msr = MSR_AMD_PATCHLEVEL; 1004 int err; 1005 1006 /* 1007 * Erratum 123 applies only to multi-core cpus. 1008 */ 1009 if (cpuid_get_ncpu_per_chip(cpu) < 2) 1010 break; 1011 #if defined(__xpv) 1012 if (!DOMAIN_IS_INITDOMAIN(xen_info)) 1013 break; 1014 #endif 1015 /* 1016 * The "workaround" is to print a warning to upgrade the BIOS 1017 */ 1018 if ((err = checked_rdmsr(msr, &value)) != 0) { 1019 msr_warning(cpu, "rd", msr, err); 1020 workaround_warning(cpu, 123); 1021 missing++; 1022 } 1023 if (value == 0) 1024 opteron_erratum_123++; 1025 #else 1026 workaround_warning(cpu, 123); 1027 missing++; 1028 1029 #endif 1030 /*CONSTANTCONDITION*/ 1031 } while (0); 1032 1033 /*LINTED*/ 1034 if (cpuid_opteron_erratum(cpu, 131) > 0) do { 1035 /* 1036 * Multiprocessor Systems with Four or More Cores May Deadlock 1037 * Waiting for a Probe Response 1038 */ 1039 #if defined(OPTERON_ERRATUM_131) 1040 uint64_t nbcfg; 1041 const uint_t msr = MSR_AMD_NB_CFG; 1042 const uint64_t wabits = 1043 AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR; 1044 int error; 1045 1046 /* 1047 * Erratum 131 applies to any system with four or more cores. 1048 */ 1049 if (opteron_erratum_131) 1050 break; 1051 #if defined(__xpv) 1052 if (!DOMAIN_IS_INITDOMAIN(xen_info)) 1053 break; 1054 if (xpv_nr_phys_cpus() < 4) 1055 break; 1056 #else 1057 if (opteron_get_nnodes() * cpuid_get_ncpu_per_chip(cpu) < 4) 1058 break; 1059 #endif 1060 /* 1061 * Print a warning if neither of the workarounds for 1062 * erratum 131 is present. 1063 */ 1064 if ((error = checked_rdmsr(msr, &nbcfg)) != 0) { 1065 msr_warning(cpu, "rd", msr, error); 1066 workaround_warning(cpu, 131); 1067 missing++; 1068 } else if ((nbcfg & wabits) == 0) { 1069 opteron_erratum_131++; 1070 } else { 1071 /* cannot have both workarounds set */ 1072 ASSERT((nbcfg & wabits) != wabits); 1073 } 1074 #else 1075 workaround_warning(cpu, 131); 1076 missing++; 1077 #endif 1078 /*CONSTANTCONDITION*/ 1079 } while (0); 1080 1081 /* 1082 * This isn't really an erratum, but for convenience the 1083 * detection/workaround code lives here and in cpuid_opteron_erratum. 1084 */ 1085 if (cpuid_opteron_erratum(cpu, 6336786) > 0) { 1086 #if defined(OPTERON_WORKAROUND_6336786) 1087 /* 1088 * Disable C1-Clock ramping on multi-core/multi-processor 1089 * K8 platforms to guard against TSC drift. 1090 */ 1091 if (opteron_workaround_6336786) { 1092 opteron_workaround_6336786++; 1093 #if defined(__xpv) 1094 } else if ((DOMAIN_IS_INITDOMAIN(xen_info) && 1095 xpv_nr_phys_cpus() > 1) || 1096 opteron_workaround_6336786_UP) { 1097 /* 1098 * XXPV Hmm. We can't walk the Northbridges on 1099 * the hypervisor; so just complain and drive 1100 * on. This probably needs to be fixed in 1101 * the hypervisor itself. 1102 */ 1103 opteron_workaround_6336786++; 1104 workaround_warning(cpu, 6336786); 1105 #else /* __xpv */ 1106 } else if ((opteron_get_nnodes() * 1107 cpuid_get_ncpu_per_chip(cpu) > 1) || 1108 opteron_workaround_6336786_UP) { 1109 1110 uint_t node, nnodes; 1111 uint8_t data; 1112 1113 nnodes = opteron_get_nnodes(); 1114 for (node = 0; node < nnodes; node++) { 1115 /* 1116 * Clear PMM7[1:0] (function 3, offset 0x87) 1117 * Northbridge device is the node id + 24. 1118 */ 1119 data = pci_getb_func(0, node + 24, 3, 0x87); 1120 data &= 0xFC; 1121 pci_putb_func(0, node + 24, 3, 0x87, data); 1122 } 1123 opteron_workaround_6336786++; 1124 #endif /* __xpv */ 1125 } 1126 #else 1127 workaround_warning(cpu, 6336786); 1128 missing++; 1129 #endif 1130 } 1131 1132 /*LINTED*/ 1133 /* 1134 * Mutex primitives don't work as expected. 1135 */ 1136 if (cpuid_opteron_erratum(cpu, 6323525) > 0) { 1137 #if defined(OPTERON_WORKAROUND_6323525) 1138 /* 1139 * This problem only occurs with 2 or more cores. If bit in 1140 * MSR_AMD_BU_CFG set, then not applicable. The workaround 1141 * is to patch the semaphone routines with the lfence 1142 * instruction to provide necessary load memory barrier with 1143 * possible subsequent read-modify-write ops. 1144 * 1145 * It is too early in boot to call the patch routine so 1146 * set erratum variable to be done in startup_end(). 1147 */ 1148 if (opteron_workaround_6323525) { 1149 opteron_workaround_6323525++; 1150 #if defined(__xpv) 1151 } else if (is_x86_feature(x86_featureset, X86FSET_SSE2)) { 1152 if (DOMAIN_IS_INITDOMAIN(xen_info)) { 1153 /* 1154 * XXPV Use dom0_msr here when extended 1155 * operations are supported? 1156 */ 1157 if (xpv_nr_phys_cpus() > 1) 1158 opteron_workaround_6323525++; 1159 } else { 1160 /* 1161 * We have no way to tell how many physical 1162 * cpus there are, or even if this processor 1163 * has the problem, so enable the workaround 1164 * unconditionally (at some performance cost). 1165 */ 1166 opteron_workaround_6323525++; 1167 } 1168 #else /* __xpv */ 1169 } else if (is_x86_feature(x86_featureset, X86FSET_SSE2) && 1170 ((opteron_get_nnodes() * 1171 cpuid_get_ncpu_per_chip(cpu)) > 1)) { 1172 if ((xrdmsr(MSR_AMD_BU_CFG) & (UINT64_C(1) << 33)) == 0) 1173 opteron_workaround_6323525++; 1174 #endif /* __xpv */ 1175 } 1176 #else 1177 workaround_warning(cpu, 6323525); 1178 missing++; 1179 #endif 1180 } 1181 1182 missing += do_erratum_298(cpu); 1183 1184 if (cpuid_opteron_erratum(cpu, 721) > 0) { 1185 #if defined(OPTERON_ERRATUM_721) 1186 on_trap_data_t otd; 1187 1188 if (!on_trap(&otd, OT_DATA_ACCESS)) 1189 wrmsr(MSR_AMD_DE_CFG, 1190 rdmsr(MSR_AMD_DE_CFG) | AMD_DE_CFG_E721); 1191 no_trap(); 1192 1193 opteron_erratum_721++; 1194 #else 1195 workaround_warning(cpu, 721); 1196 missing++; 1197 #endif 1198 } 1199 1200 #ifdef __xpv 1201 return (0); 1202 #else 1203 return (missing); 1204 #endif 1205 } 1206 1207 void 1208 workaround_errata_end() 1209 { 1210 #if defined(OPTERON_ERRATUM_88) 1211 if (opteron_erratum_88) 1212 workaround_applied(88); 1213 #endif 1214 #if defined(OPTERON_ERRATUM_91) 1215 if (opteron_erratum_91) 1216 workaround_applied(91); 1217 #endif 1218 #if defined(OPTERON_ERRATUM_93) 1219 if (opteron_erratum_93) 1220 workaround_applied(93); 1221 #endif 1222 #if defined(OPTERON_ERRATUM_95) 1223 if (opteron_erratum_95) 1224 workaround_applied(95); 1225 #endif 1226 #if defined(OPTERON_ERRATUM_100) 1227 if (opteron_erratum_100) 1228 workaround_applied(100); 1229 #endif 1230 #if defined(OPTERON_ERRATUM_108) 1231 if (opteron_erratum_108) 1232 workaround_applied(108); 1233 #endif 1234 #if defined(OPTERON_ERRATUM_109) 1235 if (opteron_erratum_109) { 1236 cmn_err(CE_WARN, 1237 "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)" 1238 " processor\nerratum 109 was not detected; updating your" 1239 " system's BIOS to a version\ncontaining this" 1240 " microcode patch is HIGHLY recommended or erroneous" 1241 " system\noperation may occur.\n"); 1242 } 1243 #endif 1244 #if defined(OPTERON_ERRATUM_121) 1245 if (opteron_erratum_121) 1246 workaround_applied(121); 1247 #endif 1248 #if defined(OPTERON_ERRATUM_122) 1249 if (opteron_erratum_122) 1250 workaround_applied(122); 1251 #endif 1252 #if defined(OPTERON_ERRATUM_123) 1253 if (opteron_erratum_123) { 1254 cmn_err(CE_WARN, 1255 "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)" 1256 " processor\nerratum 123 was not detected; updating your" 1257 " system's BIOS to a version\ncontaining this" 1258 " microcode patch is HIGHLY recommended or erroneous" 1259 " system\noperation may occur.\n"); 1260 } 1261 #endif 1262 #if defined(OPTERON_ERRATUM_131) 1263 if (opteron_erratum_131) { 1264 cmn_err(CE_WARN, 1265 "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)" 1266 " processor\nerratum 131 was not detected; updating your" 1267 " system's BIOS to a version\ncontaining this" 1268 " microcode patch is HIGHLY recommended or erroneous" 1269 " system\noperation may occur.\n"); 1270 } 1271 #endif 1272 #if defined(OPTERON_WORKAROUND_6336786) 1273 if (opteron_workaround_6336786) 1274 workaround_applied(6336786); 1275 #endif 1276 #if defined(OPTERON_WORKAROUND_6323525) 1277 if (opteron_workaround_6323525) 1278 workaround_applied(6323525); 1279 #endif 1280 #if defined(OPTERON_ERRATUM_298) 1281 if (opteron_erratum_298) { 1282 cmn_err(CE_WARN, 1283 "BIOS microcode patch for AMD 64/Opteron(tm)" 1284 " processor\nerratum 298 was not detected; updating your" 1285 " system's BIOS to a version\ncontaining this" 1286 " microcode patch is HIGHLY recommended or erroneous" 1287 " system\noperation may occur.\n"); 1288 } 1289 #endif 1290 #if defined(OPTERON_ERRATUM_721) 1291 if (opteron_erratum_721) 1292 workaround_applied(721); 1293 #endif 1294 } 1295 1296 /* 1297 * The procset_slave and procset_master are used to synchronize 1298 * between the control CPU and the target CPU when starting CPUs. 1299 */ 1300 static cpuset_t procset_slave, procset_master; 1301 1302 static void 1303 mp_startup_wait(cpuset_t *sp, processorid_t cpuid) 1304 { 1305 cpuset_t tempset; 1306 1307 for (tempset = *sp; !CPU_IN_SET(tempset, cpuid); 1308 tempset = *(volatile cpuset_t *)sp) { 1309 SMT_PAUSE(); 1310 } 1311 CPUSET_ATOMIC_DEL(*(cpuset_t *)sp, cpuid); 1312 } 1313 1314 static void 1315 mp_startup_signal(cpuset_t *sp, processorid_t cpuid) 1316 { 1317 cpuset_t tempset; 1318 1319 CPUSET_ATOMIC_ADD(*(cpuset_t *)sp, cpuid); 1320 for (tempset = *sp; CPU_IN_SET(tempset, cpuid); 1321 tempset = *(volatile cpuset_t *)sp) { 1322 SMT_PAUSE(); 1323 } 1324 } 1325 1326 int 1327 mp_start_cpu_common(cpu_t *cp, boolean_t boot) 1328 { 1329 _NOTE(ARGUNUSED(boot)); 1330 1331 void *ctx; 1332 int delays; 1333 int error = 0; 1334 cpuset_t tempset; 1335 processorid_t cpuid; 1336 #ifndef __xpv 1337 extern void cpupm_init(cpu_t *); 1338 #endif 1339 1340 ASSERT(cp != NULL); 1341 cpuid = cp->cpu_id; 1342 ctx = mach_cpucontext_alloc(cp); 1343 if (ctx == NULL) { 1344 cmn_err(CE_WARN, 1345 "cpu%d: failed to allocate context", cp->cpu_id); 1346 return (EAGAIN); 1347 } 1348 error = mach_cpu_start(cp, ctx); 1349 if (error != 0) { 1350 cmn_err(CE_WARN, 1351 "cpu%d: failed to start, error %d", cp->cpu_id, error); 1352 mach_cpucontext_free(cp, ctx, error); 1353 return (error); 1354 } 1355 1356 for (delays = 0, tempset = procset_slave; !CPU_IN_SET(tempset, cpuid); 1357 delays++) { 1358 if (delays == 500) { 1359 /* 1360 * After five seconds, things are probably looking 1361 * a bit bleak - explain the hang. 1362 */ 1363 cmn_err(CE_NOTE, "cpu%d: started, " 1364 "but not running in the kernel yet", cpuid); 1365 } else if (delays > 2000) { 1366 /* 1367 * We waited at least 20 seconds, bail .. 1368 */ 1369 error = ETIMEDOUT; 1370 cmn_err(CE_WARN, "cpu%d: timed out", cpuid); 1371 mach_cpucontext_free(cp, ctx, error); 1372 return (error); 1373 } 1374 1375 /* 1376 * wait at least 10ms, then check again.. 1377 */ 1378 delay(USEC_TO_TICK_ROUNDUP(10000)); 1379 tempset = *((volatile cpuset_t *)&procset_slave); 1380 } 1381 CPUSET_ATOMIC_DEL(procset_slave, cpuid); 1382 1383 mach_cpucontext_free(cp, ctx, 0); 1384 1385 #ifndef __xpv 1386 if (tsc_gethrtime_enable) 1387 tsc_sync_master(cpuid); 1388 #endif 1389 1390 if (dtrace_cpu_init != NULL) { 1391 (*dtrace_cpu_init)(cpuid); 1392 } 1393 1394 /* 1395 * During CPU DR operations, the cpu_lock is held by current 1396 * (the control) thread. We can't release the cpu_lock here 1397 * because that will break the CPU DR logic. 1398 * On the other hand, CPUPM and processor group initialization 1399 * routines need to access the cpu_lock. So we invoke those 1400 * routines here on behalf of mp_startup_common(). 1401 * 1402 * CPUPM and processor group initialization routines depend 1403 * on the cpuid probing results. Wait for mp_startup_common() 1404 * to signal that cpuid probing is done. 1405 */ 1406 mp_startup_wait(&procset_slave, cpuid); 1407 #ifndef __xpv 1408 cpupm_init(cp); 1409 #endif 1410 (void) pg_cpu_init(cp, B_FALSE); 1411 cpu_set_state(cp); 1412 mp_startup_signal(&procset_master, cpuid); 1413 1414 return (0); 1415 } 1416 1417 /* 1418 * Start a single cpu, assuming that the kernel context is available 1419 * to successfully start another cpu. 1420 * 1421 * (For example, real mode code is mapped into the right place 1422 * in memory and is ready to be run.) 1423 */ 1424 int 1425 start_cpu(processorid_t who) 1426 { 1427 cpu_t *cp; 1428 int error = 0; 1429 cpuset_t tempset; 1430 1431 ASSERT(who != 0); 1432 1433 /* 1434 * Check if there's at least a Mbyte of kmem available 1435 * before attempting to start the cpu. 1436 */ 1437 if (kmem_avail() < 1024 * 1024) { 1438 /* 1439 * Kick off a reap in case that helps us with 1440 * later attempts .. 1441 */ 1442 kmem_reap(); 1443 return (ENOMEM); 1444 } 1445 1446 /* 1447 * First configure cpu. 1448 */ 1449 cp = mp_cpu_configure_common(who, B_TRUE); 1450 ASSERT(cp != NULL); 1451 1452 /* 1453 * Then start cpu. 1454 */ 1455 error = mp_start_cpu_common(cp, B_TRUE); 1456 if (error != 0) { 1457 mp_cpu_unconfigure_common(cp, error); 1458 return (error); 1459 } 1460 1461 mutex_exit(&cpu_lock); 1462 tempset = cpu_ready_set; 1463 while (!CPU_IN_SET(tempset, who)) { 1464 drv_usecwait(1); 1465 tempset = *((volatile cpuset_t *)&cpu_ready_set); 1466 } 1467 mutex_enter(&cpu_lock); 1468 1469 return (0); 1470 } 1471 1472 void 1473 start_other_cpus(int cprboot) 1474 { 1475 _NOTE(ARGUNUSED(cprboot)); 1476 1477 uint_t who; 1478 uint_t bootcpuid = 0; 1479 1480 /* 1481 * Initialize our own cpu_info. 1482 */ 1483 init_cpu_info(CPU); 1484 1485 cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_idstr); 1486 cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_brandstr); 1487 1488 /* 1489 * Initialize our syscall handlers 1490 */ 1491 init_cpu_syscall(CPU); 1492 1493 /* 1494 * Take the boot cpu out of the mp_cpus set because we know 1495 * it's already running. Add it to the cpu_ready_set for 1496 * precisely the same reason. 1497 */ 1498 CPUSET_DEL(mp_cpus, bootcpuid); 1499 CPUSET_ADD(cpu_ready_set, bootcpuid); 1500 1501 /* 1502 * skip the rest of this if 1503 * . only 1 cpu dectected and system isn't hotplug-capable 1504 * . not using MP 1505 */ 1506 if ((CPUSET_ISNULL(mp_cpus) && plat_dr_support_cpu() == 0) || 1507 use_mp == 0) { 1508 if (use_mp == 0) 1509 cmn_err(CE_CONT, "?***** Not in MP mode\n"); 1510 goto done; 1511 } 1512 1513 /* 1514 * perform such initialization as is needed 1515 * to be able to take CPUs on- and off-line. 1516 */ 1517 cpu_pause_init(); 1518 1519 xc_init_cpu(CPU); /* initialize processor crosscalls */ 1520 1521 if (mach_cpucontext_init() != 0) 1522 goto done; 1523 1524 flushes_require_xcalls = 1; 1525 1526 /* 1527 * We lock our affinity to the master CPU to ensure that all slave CPUs 1528 * do their TSC syncs with the same CPU. 1529 */ 1530 affinity_set(CPU_CURRENT); 1531 1532 for (who = 0; who < NCPU; who++) { 1533 if (!CPU_IN_SET(mp_cpus, who)) 1534 continue; 1535 ASSERT(who != bootcpuid); 1536 1537 mutex_enter(&cpu_lock); 1538 if (start_cpu(who) != 0) 1539 CPUSET_DEL(mp_cpus, who); 1540 cpu_state_change_notify(who, CPU_SETUP); 1541 mutex_exit(&cpu_lock); 1542 } 1543 1544 /* Free the space allocated to hold the microcode file */ 1545 ucode_cleanup(); 1546 1547 affinity_clear(); 1548 1549 mach_cpucontext_fini(); 1550 1551 done: 1552 if (get_hwenv() == HW_NATIVE) 1553 workaround_errata_end(); 1554 cmi_post_mpstartup(); 1555 1556 if (use_mp && ncpus != boot_max_ncpus) { 1557 cmn_err(CE_NOTE, 1558 "System detected %d cpus, but " 1559 "only %d cpu(s) were enabled during boot.", 1560 boot_max_ncpus, ncpus); 1561 cmn_err(CE_NOTE, 1562 "Use \"boot-ncpus\" parameter to enable more CPU(s). " 1563 "See eeprom(1M)."); 1564 } 1565 } 1566 1567 int 1568 mp_cpu_configure(int cpuid) 1569 { 1570 cpu_t *cp; 1571 1572 if (use_mp == 0 || plat_dr_support_cpu() == 0) { 1573 return (ENOTSUP); 1574 } 1575 1576 cp = cpu_get(cpuid); 1577 if (cp != NULL) { 1578 return (EALREADY); 1579 } 1580 1581 /* 1582 * Check if there's at least a Mbyte of kmem available 1583 * before attempting to start the cpu. 1584 */ 1585 if (kmem_avail() < 1024 * 1024) { 1586 /* 1587 * Kick off a reap in case that helps us with 1588 * later attempts .. 1589 */ 1590 kmem_reap(); 1591 return (ENOMEM); 1592 } 1593 1594 cp = mp_cpu_configure_common(cpuid, B_FALSE); 1595 ASSERT(cp != NULL && cpu_get(cpuid) == cp); 1596 1597 return (cp != NULL ? 0 : EAGAIN); 1598 } 1599 1600 int 1601 mp_cpu_unconfigure(int cpuid) 1602 { 1603 cpu_t *cp; 1604 1605 if (use_mp == 0 || plat_dr_support_cpu() == 0) { 1606 return (ENOTSUP); 1607 } else if (cpuid < 0 || cpuid >= max_ncpus) { 1608 return (EINVAL); 1609 } 1610 1611 cp = cpu_get(cpuid); 1612 if (cp == NULL) { 1613 return (ENODEV); 1614 } 1615 mp_cpu_unconfigure_common(cp, 0); 1616 1617 return (0); 1618 } 1619 1620 /* 1621 * Startup function for 'other' CPUs (besides boot cpu). 1622 * Called from real_mode_start. 1623 * 1624 * WARNING: until CPU_READY is set, mp_startup_common and routines called by 1625 * mp_startup_common should not call routines (e.g. kmem_free) that could call 1626 * hat_unload which requires CPU_READY to be set. 1627 */ 1628 static void 1629 mp_startup_common(boolean_t boot) 1630 { 1631 cpu_t *cp = CPU; 1632 uchar_t new_x86_featureset[BT_SIZEOFMAP(NUM_X86_FEATURES)]; 1633 extern void cpu_event_init_cpu(cpu_t *); 1634 1635 /* 1636 * We need to get TSC on this proc synced (i.e., any delta 1637 * from cpu0 accounted for) as soon as we can, because many 1638 * many things use gethrtime/pc_gethrestime, including 1639 * interrupts, cmn_err, etc. Before we can do that, we want to 1640 * clear TSC if we're on a buggy Sandy/Ivy Bridge CPU, so do that 1641 * right away. 1642 */ 1643 bzero(new_x86_featureset, BT_SIZEOFMAP(NUM_X86_FEATURES)); 1644 cpuid_pass1(cp, new_x86_featureset); 1645 1646 if (boot && get_hwenv() == HW_NATIVE && 1647 cpuid_getvendor(CPU) == X86_VENDOR_Intel && 1648 cpuid_getfamily(CPU) == 6 && 1649 (cpuid_getmodel(CPU) == 0x2d || cpuid_getmodel(CPU) == 0x3e) && 1650 is_x86_feature(new_x86_featureset, X86FSET_TSC)) { 1651 (void) wrmsr(REG_TSC, 0UL); 1652 } 1653 1654 /* Let the control CPU continue into tsc_sync_master() */ 1655 mp_startup_signal(&procset_slave, cp->cpu_id); 1656 1657 #ifndef __xpv 1658 if (tsc_gethrtime_enable) 1659 tsc_sync_slave(); 1660 #endif 1661 1662 /* 1663 * Once this was done from assembly, but it's safer here; if 1664 * it blocks, we need to be able to swtch() to and from, and 1665 * since we get here by calling t_pc, we need to do that call 1666 * before swtch() overwrites it. 1667 */ 1668 (void) (*ap_mlsetup)(); 1669 1670 #ifndef __xpv 1671 /* 1672 * Program this cpu's PAT 1673 */ 1674 pat_sync(); 1675 #endif 1676 1677 /* 1678 * Set up TSC_AUX to contain the cpuid for this processor 1679 * for the rdtscp instruction. 1680 */ 1681 if (is_x86_feature(x86_featureset, X86FSET_TSCP)) 1682 (void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id); 1683 1684 /* 1685 * Initialize this CPU's syscall handlers 1686 */ 1687 init_cpu_syscall(cp); 1688 1689 /* 1690 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the 1691 * highest level at which a routine is permitted to block on 1692 * an adaptive mutex (allows for cpu poke interrupt in case 1693 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks 1694 * device interrupts that may end up in the hat layer issuing cross 1695 * calls before CPU_READY is set. 1696 */ 1697 splx(ipltospl(LOCK_LEVEL)); 1698 sti(); 1699 1700 /* 1701 * Do a sanity check to make sure this new CPU is a sane thing 1702 * to add to the collection of processors running this system. 1703 * 1704 * XXX Clearly this needs to get more sophisticated, if x86 1705 * systems start to get built out of heterogenous CPUs; as is 1706 * likely to happen once the number of processors in a configuration 1707 * gets large enough. 1708 */ 1709 if (compare_x86_featureset(x86_featureset, new_x86_featureset) == 1710 B_FALSE) { 1711 cmn_err(CE_CONT, "cpu%d: featureset\n", cp->cpu_id); 1712 print_x86_featureset(new_x86_featureset); 1713 cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id); 1714 } 1715 1716 /* 1717 * We do not support cpus with mixed monitor/mwait support if the 1718 * boot cpu supports monitor/mwait. 1719 */ 1720 if (is_x86_feature(x86_featureset, X86FSET_MWAIT) != 1721 is_x86_feature(new_x86_featureset, X86FSET_MWAIT)) 1722 panic("unsupported mixed cpu monitor/mwait support detected"); 1723 1724 /* 1725 * We could be more sophisticated here, and just mark the CPU 1726 * as "faulted" but at this point we'll opt for the easier 1727 * answer of dying horribly. Provided the boot cpu is ok, 1728 * the system can be recovered by booting with use_mp set to zero. 1729 */ 1730 if (workaround_errata(cp) != 0) 1731 panic("critical workaround(s) missing for cpu%d", cp->cpu_id); 1732 1733 /* 1734 * We can touch cpu_flags here without acquiring the cpu_lock here 1735 * because the cpu_lock is held by the control CPU which is running 1736 * mp_start_cpu_common(). 1737 * Need to clear CPU_QUIESCED flag before calling any function which 1738 * may cause thread context switching, such as kmem_alloc() etc. 1739 * The idle thread checks for CPU_QUIESCED flag and loops for ever if 1740 * it's set. So the startup thread may have no chance to switch back 1741 * again if it's switched away with CPU_QUIESCED set. 1742 */ 1743 cp->cpu_flags &= ~(CPU_POWEROFF | CPU_QUIESCED); 1744 1745 /* 1746 * Setup this processor for XSAVE. 1747 */ 1748 if (fp_save_mech == FP_XSAVE) { 1749 xsave_setup_msr(cp); 1750 } 1751 1752 cpuid_pass2(cp); 1753 cpuid_pass3(cp); 1754 cpuid_pass4(cp, NULL); 1755 1756 /* 1757 * Correct cpu_idstr and cpu_brandstr on target CPU after 1758 * cpuid_pass1() is done. 1759 */ 1760 (void) cpuid_getidstr(cp, cp->cpu_idstr, CPU_IDSTRLEN); 1761 (void) cpuid_getbrandstr(cp, cp->cpu_brandstr, CPU_IDSTRLEN); 1762 1763 cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_EXISTS; 1764 1765 post_startup_cpu_fixups(); 1766 1767 cpu_event_init_cpu(cp); 1768 1769 /* 1770 * Enable preemption here so that contention for any locks acquired 1771 * later in mp_startup_common may be preempted if the thread owning 1772 * those locks is continuously executing on other CPUs (for example, 1773 * this CPU must be preemptible to allow other CPUs to pause it during 1774 * their startup phases). It's safe to enable preemption here because 1775 * the CPU state is pretty-much fully constructed. 1776 */ 1777 curthread->t_preempt = 0; 1778 1779 /* The base spl should still be at LOCK LEVEL here */ 1780 ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL)); 1781 set_base_spl(); /* Restore the spl to its proper value */ 1782 1783 pghw_physid_create(cp); 1784 /* 1785 * Delegate initialization tasks, which need to access the cpu_lock, 1786 * to mp_start_cpu_common() because we can't acquire the cpu_lock here 1787 * during CPU DR operations. 1788 */ 1789 mp_startup_signal(&procset_slave, cp->cpu_id); 1790 mp_startup_wait(&procset_master, cp->cpu_id); 1791 pg_cmt_cpu_startup(cp); 1792 1793 if (boot) { 1794 mutex_enter(&cpu_lock); 1795 cp->cpu_flags &= ~CPU_OFFLINE; 1796 cpu_enable_intr(cp); 1797 cpu_add_active(cp); 1798 mutex_exit(&cpu_lock); 1799 } 1800 1801 /* Enable interrupts */ 1802 (void) spl0(); 1803 1804 /* 1805 * Fill out cpu_ucode_info. Update microcode if necessary. 1806 */ 1807 ucode_check(cp); 1808 1809 #ifndef __xpv 1810 { 1811 /* 1812 * Set up the CPU module for this CPU. This can't be done 1813 * before this CPU is made CPU_READY, because we may (in 1814 * heterogeneous systems) need to go load another CPU module. 1815 * The act of attempting to load a module may trigger a 1816 * cross-call, which will ASSERT unless this cpu is CPU_READY. 1817 */ 1818 cmi_hdl_t hdl; 1819 1820 if ((hdl = cmi_init(CMI_HDL_NATIVE, cmi_ntv_hwchipid(CPU), 1821 cmi_ntv_hwcoreid(CPU), cmi_ntv_hwstrandid(CPU))) != NULL) { 1822 if (is_x86_feature(x86_featureset, X86FSET_MCA)) 1823 cmi_mca_init(hdl); 1824 cp->cpu_m.mcpu_cmi_hdl = hdl; 1825 } 1826 } 1827 #endif /* __xpv */ 1828 1829 if (boothowto & RB_DEBUG) 1830 kdi_cpu_init(); 1831 1832 /* 1833 * Setting the bit in cpu_ready_set must be the last operation in 1834 * processor initialization; the boot CPU will continue to boot once 1835 * it sees this bit set for all active CPUs. 1836 */ 1837 CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id); 1838 1839 (void) mach_cpu_create_device_node(cp, NULL); 1840 1841 cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr); 1842 cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr); 1843 cmn_err(CE_CONT, "?cpu%d initialization complete - online\n", 1844 cp->cpu_id); 1845 1846 /* 1847 * Now we are done with the startup thread, so free it up. 1848 */ 1849 thread_exit(); 1850 panic("mp_startup: cannot return"); 1851 /*NOTREACHED*/ 1852 } 1853 1854 /* 1855 * Startup function for 'other' CPUs at boot time (besides boot cpu). 1856 */ 1857 static void 1858 mp_startup_boot(void) 1859 { 1860 mp_startup_common(B_TRUE); 1861 } 1862 1863 /* 1864 * Startup function for hotplug CPUs at runtime. 1865 */ 1866 void 1867 mp_startup_hotplug(void) 1868 { 1869 mp_startup_common(B_FALSE); 1870 } 1871 1872 /* 1873 * Start CPU on user request. 1874 */ 1875 /* ARGSUSED */ 1876 int 1877 mp_cpu_start(struct cpu *cp) 1878 { 1879 ASSERT(MUTEX_HELD(&cpu_lock)); 1880 return (0); 1881 } 1882 1883 /* 1884 * Stop CPU on user request. 1885 */ 1886 int 1887 mp_cpu_stop(struct cpu *cp) 1888 { 1889 extern int cbe_psm_timer_mode; 1890 ASSERT(MUTEX_HELD(&cpu_lock)); 1891 1892 #ifdef __xpv 1893 /* 1894 * We can't offline vcpu0. 1895 */ 1896 if (cp->cpu_id == 0) 1897 return (EBUSY); 1898 #endif 1899 1900 /* 1901 * If TIMER_PERIODIC mode is used, CPU0 is the one running it; 1902 * can't stop it. (This is true only for machines with no TSC.) 1903 */ 1904 1905 if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0)) 1906 return (EBUSY); 1907 1908 return (0); 1909 } 1910 1911 /* 1912 * Take the specified CPU out of participation in interrupts. 1913 */ 1914 int 1915 cpu_disable_intr(struct cpu *cp) 1916 { 1917 if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS) 1918 return (EBUSY); 1919 1920 cp->cpu_flags &= ~CPU_ENABLE; 1921 return (0); 1922 } 1923 1924 /* 1925 * Allow the specified CPU to participate in interrupts. 1926 */ 1927 void 1928 cpu_enable_intr(struct cpu *cp) 1929 { 1930 ASSERT(MUTEX_HELD(&cpu_lock)); 1931 cp->cpu_flags |= CPU_ENABLE; 1932 psm_enable_intr(cp->cpu_id); 1933 } 1934 1935 void 1936 mp_cpu_faulted_enter(struct cpu *cp) 1937 { 1938 #ifdef __xpv 1939 _NOTE(ARGUNUSED(cp)); 1940 #else 1941 cmi_hdl_t hdl = cp->cpu_m.mcpu_cmi_hdl; 1942 1943 if (hdl != NULL) { 1944 cmi_hdl_hold(hdl); 1945 } else { 1946 hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp), 1947 cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp)); 1948 } 1949 if (hdl != NULL) { 1950 cmi_faulted_enter(hdl); 1951 cmi_hdl_rele(hdl); 1952 } 1953 #endif 1954 } 1955 1956 void 1957 mp_cpu_faulted_exit(struct cpu *cp) 1958 { 1959 #ifdef __xpv 1960 _NOTE(ARGUNUSED(cp)); 1961 #else 1962 cmi_hdl_t hdl = cp->cpu_m.mcpu_cmi_hdl; 1963 1964 if (hdl != NULL) { 1965 cmi_hdl_hold(hdl); 1966 } else { 1967 hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp), 1968 cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp)); 1969 } 1970 if (hdl != NULL) { 1971 cmi_faulted_exit(hdl); 1972 cmi_hdl_rele(hdl); 1973 } 1974 #endif 1975 } 1976 1977 /* 1978 * The following two routines are used as context operators on threads belonging 1979 * to processes with a private LDT (see sysi86). Due to the rarity of such 1980 * processes, these routines are currently written for best code readability and 1981 * organization rather than speed. We could avoid checking x86_featureset at 1982 * every context switch by installing different context ops, depending on 1983 * x86_featureset, at LDT creation time -- one for each combination of fast 1984 * syscall features. 1985 */ 1986 1987 /*ARGSUSED*/ 1988 void 1989 cpu_fast_syscall_disable(void *arg) 1990 { 1991 if (is_x86_feature(x86_featureset, X86FSET_MSR) && 1992 is_x86_feature(x86_featureset, X86FSET_SEP)) 1993 cpu_sep_disable(); 1994 if (is_x86_feature(x86_featureset, X86FSET_MSR) && 1995 is_x86_feature(x86_featureset, X86FSET_ASYSC)) 1996 cpu_asysc_disable(); 1997 } 1998 1999 /*ARGSUSED*/ 2000 void 2001 cpu_fast_syscall_enable(void *arg) 2002 { 2003 if (is_x86_feature(x86_featureset, X86FSET_MSR) && 2004 is_x86_feature(x86_featureset, X86FSET_SEP)) 2005 cpu_sep_enable(); 2006 if (is_x86_feature(x86_featureset, X86FSET_MSR) && 2007 is_x86_feature(x86_featureset, X86FSET_ASYSC)) 2008 cpu_asysc_enable(); 2009 } 2010 2011 static void 2012 cpu_sep_enable(void) 2013 { 2014 ASSERT(is_x86_feature(x86_featureset, X86FSET_SEP)); 2015 ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL); 2016 2017 wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL); 2018 } 2019 2020 static void 2021 cpu_sep_disable(void) 2022 { 2023 ASSERT(is_x86_feature(x86_featureset, X86FSET_SEP)); 2024 ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL); 2025 2026 /* 2027 * Setting the SYSENTER_CS_MSR register to 0 causes software executing 2028 * the sysenter or sysexit instruction to trigger a #gp fault. 2029 */ 2030 wrmsr(MSR_INTC_SEP_CS, 0); 2031 } 2032 2033 static void 2034 cpu_asysc_enable(void) 2035 { 2036 ASSERT(is_x86_feature(x86_featureset, X86FSET_ASYSC)); 2037 ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL); 2038 2039 wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) | 2040 (uint64_t)(uintptr_t)AMD_EFER_SCE); 2041 } 2042 2043 static void 2044 cpu_asysc_disable(void) 2045 { 2046 ASSERT(is_x86_feature(x86_featureset, X86FSET_ASYSC)); 2047 ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL); 2048 2049 /* 2050 * Turn off the SCE (syscall enable) bit in the EFER register. Software 2051 * executing syscall or sysret with this bit off will incur a #ud trap. 2052 */ 2053 wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) & 2054 ~((uint64_t)(uintptr_t)AMD_EFER_SCE)); 2055 } 2056