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 2009 Sun Microsystems, Inc. All rights reserved. 24 * Use is subject to license terms. 25 */ 26 27 #include <sys/types.h> 28 #include <sys/thread.h> 29 #include <sys/cpuvar.h> 30 #include <sys/t_lock.h> 31 #include <sys/param.h> 32 #include <sys/proc.h> 33 #include <sys/disp.h> 34 #include <sys/class.h> 35 #include <sys/cmn_err.h> 36 #include <sys/debug.h> 37 #include <sys/asm_linkage.h> 38 #include <sys/x_call.h> 39 #include <sys/systm.h> 40 #include <sys/var.h> 41 #include <sys/vtrace.h> 42 #include <vm/hat.h> 43 #include <vm/as.h> 44 #include <vm/seg_kmem.h> 45 #include <vm/seg_kp.h> 46 #include <sys/segments.h> 47 #include <sys/kmem.h> 48 #include <sys/stack.h> 49 #include <sys/smp_impldefs.h> 50 #include <sys/x86_archext.h> 51 #include <sys/machsystm.h> 52 #include <sys/traptrace.h> 53 #include <sys/clock.h> 54 #include <sys/cpc_impl.h> 55 #include <sys/pg.h> 56 #include <sys/cmt.h> 57 #include <sys/dtrace.h> 58 #include <sys/archsystm.h> 59 #include <sys/fp.h> 60 #include <sys/reboot.h> 61 #include <sys/kdi_machimpl.h> 62 #include <vm/hat_i86.h> 63 #include <sys/memnode.h> 64 #include <sys/pci_cfgspace.h> 65 #include <sys/mach_mmu.h> 66 #include <sys/sysmacros.h> 67 #if defined(__xpv) 68 #include <sys/hypervisor.h> 69 #endif 70 #include <sys/cpu_module.h> 71 72 struct cpu cpus[1]; /* CPU data */ 73 struct cpu *cpu[NCPU] = {&cpus[0]}; /* pointers to all CPUs */ 74 cpu_core_t cpu_core[NCPU]; /* cpu_core structures */ 75 76 /* 77 * Useful for disabling MP bring-up on a MP capable system. 78 */ 79 int use_mp = 1; 80 81 /* 82 * to be set by a PSM to indicate what cpus 83 * are sitting around on the system. 84 */ 85 cpuset_t mp_cpus; 86 87 /* 88 * This variable is used by the hat layer to decide whether or not 89 * critical sections are needed to prevent race conditions. For sun4m, 90 * this variable is set once enough MP initialization has been done in 91 * order to allow cross calls. 92 */ 93 int flushes_require_xcalls; 94 95 cpuset_t cpu_ready_set; /* initialized in startup() */ 96 97 static void mp_startup(void); 98 99 static void cpu_sep_enable(void); 100 static void cpu_sep_disable(void); 101 static void cpu_asysc_enable(void); 102 static void cpu_asysc_disable(void); 103 104 /* 105 * Init CPU info - get CPU type info for processor_info system call. 106 */ 107 void 108 init_cpu_info(struct cpu *cp) 109 { 110 processor_info_t *pi = &cp->cpu_type_info; 111 char buf[CPU_IDSTRLEN]; 112 113 /* 114 * Get clock-frequency property for the CPU. 115 */ 116 pi->pi_clock = cpu_freq; 117 118 /* 119 * Current frequency in Hz. 120 */ 121 cp->cpu_curr_clock = cpu_freq_hz; 122 123 /* 124 * Supported frequencies. 125 */ 126 if (cp->cpu_supp_freqs == NULL) { 127 cpu_set_supp_freqs(cp, NULL); 128 } 129 130 (void) strcpy(pi->pi_processor_type, "i386"); 131 if (fpu_exists) 132 (void) strcpy(pi->pi_fputypes, "i387 compatible"); 133 134 (void) cpuid_getidstr(cp, buf, sizeof (buf)); 135 136 cp->cpu_idstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP); 137 (void) strcpy(cp->cpu_idstr, buf); 138 139 (void) cpuid_getbrandstr(cp, buf, sizeof (buf)); 140 cp->cpu_brandstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP); 141 (void) strcpy(cp->cpu_brandstr, buf); 142 } 143 144 /* 145 * Configure syscall support on this CPU. 146 */ 147 /*ARGSUSED*/ 148 void 149 init_cpu_syscall(struct cpu *cp) 150 { 151 kpreempt_disable(); 152 153 #if defined(__amd64) 154 if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC)) { 155 156 #if !defined(__lint) 157 /* 158 * The syscall instruction imposes a certain ordering on 159 * segment selectors, so we double-check that ordering 160 * here. 161 */ 162 ASSERT(KDS_SEL == KCS_SEL + 8); 163 ASSERT(UDS_SEL == U32CS_SEL + 8); 164 ASSERT(UCS_SEL == U32CS_SEL + 16); 165 #endif 166 /* 167 * Turn syscall/sysret extensions on. 168 */ 169 cpu_asysc_enable(); 170 171 /* 172 * Program the magic registers .. 173 */ 174 wrmsr(MSR_AMD_STAR, 175 ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32); 176 wrmsr(MSR_AMD_LSTAR, (uint64_t)(uintptr_t)sys_syscall); 177 wrmsr(MSR_AMD_CSTAR, (uint64_t)(uintptr_t)sys_syscall32); 178 179 /* 180 * This list of flags is masked off the incoming 181 * %rfl when we enter the kernel. 182 */ 183 wrmsr(MSR_AMD_SFMASK, (uint64_t)(uintptr_t)(PS_IE | PS_T)); 184 } 185 #endif 186 187 /* 188 * On 32-bit kernels, we use sysenter/sysexit because it's too 189 * hard to use syscall/sysret, and it is more portable anyway. 190 * 191 * On 64-bit kernels on Nocona machines, the 32-bit syscall 192 * variant isn't available to 32-bit applications, but sysenter is. 193 */ 194 if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP)) { 195 196 #if !defined(__lint) 197 /* 198 * The sysenter instruction imposes a certain ordering on 199 * segment selectors, so we double-check that ordering 200 * here. See "sysenter" in Intel document 245471-012, "IA-32 201 * Intel Architecture Software Developer's Manual Volume 2: 202 * Instruction Set Reference" 203 */ 204 ASSERT(KDS_SEL == KCS_SEL + 8); 205 206 ASSERT32(UCS_SEL == ((KCS_SEL + 16) | 3)); 207 ASSERT32(UDS_SEL == UCS_SEL + 8); 208 209 ASSERT64(U32CS_SEL == ((KCS_SEL + 16) | 3)); 210 ASSERT64(UDS_SEL == U32CS_SEL + 8); 211 #endif 212 213 cpu_sep_enable(); 214 215 /* 216 * resume() sets this value to the base of the threads stack 217 * via a context handler. 218 */ 219 wrmsr(MSR_INTC_SEP_ESP, 0); 220 wrmsr(MSR_INTC_SEP_EIP, (uint64_t)(uintptr_t)sys_sysenter); 221 } 222 223 kpreempt_enable(); 224 } 225 226 /* 227 * Multiprocessor initialization. 228 * 229 * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the 230 * startup and idle threads for the specified CPU. 231 */ 232 struct cpu * 233 mp_startup_init(int cpun) 234 { 235 struct cpu *cp; 236 kthread_id_t tp; 237 caddr_t sp; 238 proc_t *procp; 239 #if !defined(__xpv) 240 extern int idle_cpu_prefer_mwait; 241 extern void cpu_idle_mwait(); 242 #endif 243 extern void idle(); 244 extern void cpu_idle(); 245 246 #ifdef TRAPTRACE 247 trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun]; 248 #endif 249 250 ASSERT(cpun < NCPU && cpu[cpun] == NULL); 251 252 cp = kmem_zalloc(sizeof (*cp), KM_SLEEP); 253 #if !defined(__xpv) 254 if ((x86_feature & X86_MWAIT) && idle_cpu_prefer_mwait) { 255 cp->cpu_m.mcpu_mwait = cpuid_mwait_alloc(CPU); 256 cp->cpu_m.mcpu_idle_cpu = cpu_idle_mwait; 257 } else 258 #endif 259 cp->cpu_m.mcpu_idle_cpu = cpu_idle; 260 261 procp = curthread->t_procp; 262 263 mutex_enter(&cpu_lock); 264 /* 265 * Initialize the dispatcher first. 266 */ 267 disp_cpu_init(cp); 268 mutex_exit(&cpu_lock); 269 270 cpu_vm_data_init(cp); 271 272 /* 273 * Allocate and initialize the startup thread for this CPU. 274 * Interrupt and process switch stacks get allocated later 275 * when the CPU starts running. 276 */ 277 tp = thread_create(NULL, 0, NULL, NULL, 0, procp, 278 TS_STOPPED, maxclsyspri); 279 280 /* 281 * Set state to TS_ONPROC since this thread will start running 282 * as soon as the CPU comes online. 283 * 284 * All the other fields of the thread structure are setup by 285 * thread_create(). 286 */ 287 THREAD_ONPROC(tp, cp); 288 tp->t_preempt = 1; 289 tp->t_bound_cpu = cp; 290 tp->t_affinitycnt = 1; 291 tp->t_cpu = cp; 292 tp->t_disp_queue = cp->cpu_disp; 293 294 /* 295 * Setup thread to start in mp_startup. 296 */ 297 sp = tp->t_stk; 298 tp->t_pc = (uintptr_t)mp_startup; 299 tp->t_sp = (uintptr_t)(sp - MINFRAME); 300 #if defined(__amd64) 301 tp->t_sp -= STACK_ENTRY_ALIGN; /* fake a call */ 302 #endif 303 304 cp->cpu_id = cpun; 305 cp->cpu_self = cp; 306 cp->cpu_thread = tp; 307 cp->cpu_lwp = NULL; 308 cp->cpu_dispthread = tp; 309 cp->cpu_dispatch_pri = DISP_PRIO(tp); 310 311 /* 312 * cpu_base_spl must be set explicitly here to prevent any blocking 313 * operations in mp_startup from causing the spl of the cpu to drop 314 * to 0 (allowing device interrupts before we're ready) in resume(). 315 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY. 316 * As an extra bit of security on DEBUG kernels, this is enforced with 317 * an assertion in mp_startup() -- before cpu_base_spl is set to its 318 * proper value. 319 */ 320 cp->cpu_base_spl = ipltospl(LOCK_LEVEL); 321 322 /* 323 * Now, initialize per-CPU idle thread for this CPU. 324 */ 325 tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1); 326 327 cp->cpu_idle_thread = tp; 328 329 tp->t_preempt = 1; 330 tp->t_bound_cpu = cp; 331 tp->t_affinitycnt = 1; 332 tp->t_cpu = cp; 333 tp->t_disp_queue = cp->cpu_disp; 334 335 /* 336 * Bootstrap the CPU's PG data 337 */ 338 pg_cpu_bootstrap(cp); 339 340 /* 341 * Perform CPC initialization on the new CPU. 342 */ 343 kcpc_hw_init(cp); 344 345 /* 346 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2 347 * for each CPU. 348 */ 349 setup_vaddr_for_ppcopy(cp); 350 351 /* 352 * Allocate page for new GDT and initialize from current GDT. 353 */ 354 #if !defined(__lint) 355 ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE); 356 #endif 357 cp->cpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP); 358 bcopy(CPU->cpu_gdt, cp->cpu_gdt, (sizeof (*cp->cpu_gdt) * NGDT)); 359 360 #if defined(__i386) 361 /* 362 * setup kernel %gs. 363 */ 364 set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA, 365 SEL_KPL, 0, 1); 366 #endif 367 368 /* 369 * If we have more than one node, each cpu gets a copy of IDT 370 * local to its node. If this is a Pentium box, we use cpu 0's 371 * IDT. cpu 0's IDT has been made read-only to workaround the 372 * cmpxchgl register bug 373 */ 374 if (system_hardware.hd_nodes && x86_type != X86_TYPE_P5) { 375 #if !defined(__lint) 376 ASSERT((sizeof (*CPU->cpu_idt) * NIDT) <= PAGESIZE); 377 #endif 378 cp->cpu_idt = kmem_zalloc(PAGESIZE, KM_SLEEP); 379 bcopy(CPU->cpu_idt, cp->cpu_idt, PAGESIZE); 380 } else { 381 cp->cpu_idt = CPU->cpu_idt; 382 } 383 384 /* 385 * Get interrupt priority data from cpu 0. 386 */ 387 cp->cpu_pri_data = CPU->cpu_pri_data; 388 389 /* 390 * alloc space for cpuid info 391 */ 392 cpuid_alloc_space(cp); 393 394 /* 395 * alloc space for ucode_info 396 */ 397 ucode_alloc_space(cp); 398 399 hat_cpu_online(cp); 400 401 #ifdef TRAPTRACE 402 /* 403 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers 404 */ 405 ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP); 406 ttc->ttc_next = ttc->ttc_first; 407 ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize; 408 #endif 409 /* 410 * Record that we have another CPU. 411 */ 412 mutex_enter(&cpu_lock); 413 /* 414 * Initialize the interrupt threads for this CPU 415 */ 416 cpu_intr_alloc(cp, NINTR_THREADS); 417 /* 418 * Add CPU to list of available CPUs. It'll be on the active list 419 * after mp_startup(). 420 */ 421 cpu_add_unit(cp); 422 mutex_exit(&cpu_lock); 423 424 return (cp); 425 } 426 427 /* 428 * Undo what was done in mp_startup_init 429 */ 430 static void 431 mp_startup_fini(struct cpu *cp, int error) 432 { 433 mutex_enter(&cpu_lock); 434 435 /* 436 * Remove the CPU from the list of available CPUs. 437 */ 438 cpu_del_unit(cp->cpu_id); 439 440 if (error == ETIMEDOUT) { 441 /* 442 * The cpu was started, but never *seemed* to run any 443 * code in the kernel; it's probably off spinning in its 444 * own private world, though with potential references to 445 * our kmem-allocated IDTs and GDTs (for example). 446 * 447 * Worse still, it may actually wake up some time later, 448 * so rather than guess what it might or might not do, we 449 * leave the fundamental data structures intact. 450 */ 451 cp->cpu_flags = 0; 452 mutex_exit(&cpu_lock); 453 return; 454 } 455 456 /* 457 * At this point, the only threads bound to this CPU should 458 * special per-cpu threads: it's idle thread, it's pause threads, 459 * and it's interrupt threads. Clean these up. 460 */ 461 cpu_destroy_bound_threads(cp); 462 cp->cpu_idle_thread = NULL; 463 464 /* 465 * Free the interrupt stack. 466 */ 467 segkp_release(segkp, 468 cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME))); 469 470 mutex_exit(&cpu_lock); 471 472 #ifdef TRAPTRACE 473 /* 474 * Discard the trap trace buffer 475 */ 476 { 477 trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id]; 478 479 kmem_free((void *)ttc->ttc_first, trap_trace_bufsize); 480 ttc->ttc_first = NULL; 481 } 482 #endif 483 484 hat_cpu_offline(cp); 485 486 cpuid_free_space(cp); 487 488 ucode_free_space(cp); 489 490 if (cp->cpu_idt != CPU->cpu_idt) 491 kmem_free(cp->cpu_idt, PAGESIZE); 492 cp->cpu_idt = NULL; 493 494 kmem_free(cp->cpu_gdt, PAGESIZE); 495 cp->cpu_gdt = NULL; 496 497 teardown_vaddr_for_ppcopy(cp); 498 499 kcpc_hw_fini(cp); 500 501 cp->cpu_dispthread = NULL; 502 cp->cpu_thread = NULL; /* discarded by cpu_destroy_bound_threads() */ 503 504 cpu_vm_data_destroy(cp); 505 506 mutex_enter(&cpu_lock); 507 disp_cpu_fini(cp); 508 mutex_exit(&cpu_lock); 509 510 #if !defined(__xpv) 511 if (cp->cpu_m.mcpu_mwait != NULL) 512 cpuid_mwait_free(cp); 513 #endif 514 kmem_free(cp, sizeof (*cp)); 515 } 516 517 /* 518 * Apply workarounds for known errata, and warn about those that are absent. 519 * 520 * System vendors occasionally create configurations which contain different 521 * revisions of the CPUs that are almost but not exactly the same. At the 522 * time of writing, this meant that their clock rates were the same, their 523 * feature sets were the same, but the required workaround were -not- 524 * necessarily the same. So, this routine is invoked on -every- CPU soon 525 * after starting to make sure that the resulting system contains the most 526 * pessimal set of workarounds needed to cope with *any* of the CPUs in the 527 * system. 528 * 529 * workaround_errata is invoked early in mlsetup() for CPU 0, and in 530 * mp_startup() for all slave CPUs. Slaves process workaround_errata prior 531 * to acknowledging their readiness to the master, so this routine will 532 * never be executed by multiple CPUs in parallel, thus making updates to 533 * global data safe. 534 * 535 * These workarounds are based on Rev 3.57 of the Revision Guide for 536 * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005. 537 */ 538 539 #if defined(OPTERON_ERRATUM_88) 540 int opteron_erratum_88; /* if non-zero -> at least one cpu has it */ 541 #endif 542 543 #if defined(OPTERON_ERRATUM_91) 544 int opteron_erratum_91; /* if non-zero -> at least one cpu has it */ 545 #endif 546 547 #if defined(OPTERON_ERRATUM_93) 548 int opteron_erratum_93; /* if non-zero -> at least one cpu has it */ 549 #endif 550 551 #if defined(OPTERON_ERRATUM_95) 552 int opteron_erratum_95; /* if non-zero -> at least one cpu has it */ 553 #endif 554 555 #if defined(OPTERON_ERRATUM_100) 556 int opteron_erratum_100; /* if non-zero -> at least one cpu has it */ 557 #endif 558 559 #if defined(OPTERON_ERRATUM_108) 560 int opteron_erratum_108; /* if non-zero -> at least one cpu has it */ 561 #endif 562 563 #if defined(OPTERON_ERRATUM_109) 564 int opteron_erratum_109; /* if non-zero -> at least one cpu has it */ 565 #endif 566 567 #if defined(OPTERON_ERRATUM_121) 568 int opteron_erratum_121; /* if non-zero -> at least one cpu has it */ 569 #endif 570 571 #if defined(OPTERON_ERRATUM_122) 572 int opteron_erratum_122; /* if non-zero -> at least one cpu has it */ 573 #endif 574 575 #if defined(OPTERON_ERRATUM_123) 576 int opteron_erratum_123; /* if non-zero -> at least one cpu has it */ 577 #endif 578 579 #if defined(OPTERON_ERRATUM_131) 580 int opteron_erratum_131; /* if non-zero -> at least one cpu has it */ 581 #endif 582 583 #if defined(OPTERON_WORKAROUND_6336786) 584 int opteron_workaround_6336786; /* non-zero -> WA relevant and applied */ 585 int opteron_workaround_6336786_UP = 0; /* Not needed for UP */ 586 #endif 587 588 #if defined(OPTERON_WORKAROUND_6323525) 589 int opteron_workaround_6323525; /* if non-zero -> at least one cpu has it */ 590 #endif 591 592 #if defined(OPTERON_ERRATUM_298) 593 int opteron_erratum_298; 594 #endif 595 596 static void 597 workaround_warning(cpu_t *cp, uint_t erratum) 598 { 599 cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u", 600 cp->cpu_id, erratum); 601 } 602 603 static void 604 workaround_applied(uint_t erratum) 605 { 606 if (erratum > 1000000) 607 cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n", 608 erratum); 609 else 610 cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n", 611 erratum); 612 } 613 614 static void 615 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error) 616 { 617 cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d", 618 cp->cpu_id, rw, msr, error); 619 } 620 621 /* 622 * Determine the number of nodes in a Hammer / Greyhound / Griffin family 623 * system. 624 */ 625 static uint_t 626 opteron_get_nnodes(void) 627 { 628 static uint_t nnodes = 0; 629 630 if (nnodes == 0) { 631 #ifdef DEBUG 632 uint_t family; 633 634 /* 635 * This routine uses a PCI config space based mechanism 636 * for retrieving the number of nodes in the system. 637 * Device 24, function 0, offset 0x60 as used here is not 638 * AMD processor architectural, and may not work on processor 639 * families other than those listed below. 640 * 641 * Callers of this routine must ensure that we're running on 642 * a processor which supports this mechanism. 643 * The assertion below is meant to catch calls on unsupported 644 * processors. 645 */ 646 family = cpuid_getfamily(CPU); 647 ASSERT(family == 0xf || family == 0x10 || family == 0x11); 648 #endif /* DEBUG */ 649 650 /* 651 * Obtain the number of nodes in the system from 652 * bits [6:4] of the Node ID register on node 0. 653 * 654 * The actual node count is NodeID[6:4] + 1 655 * 656 * The Node ID register is accessed via function 0, 657 * offset 0x60. Node 0 is device 24. 658 */ 659 nnodes = ((pci_getl_func(0, 24, 0, 0x60) & 0x70) >> 4) + 1; 660 } 661 return (nnodes); 662 } 663 664 #if defined(__xpv) 665 666 /* 667 * On dom0, we can determine the number of physical cpus on the machine. 668 * This number is important when figuring out what workarounds are 669 * appropriate, so compute it now. 670 */ 671 uint_t 672 xen_get_nphyscpus(void) 673 { 674 static uint_t nphyscpus = 0; 675 676 ASSERT(DOMAIN_IS_INITDOMAIN(xen_info)); 677 678 if (nphyscpus == 0) { 679 xen_sysctl_t op; 680 xen_sysctl_physinfo_t *pi = &op.u.physinfo; 681 682 op.cmd = XEN_SYSCTL_physinfo; 683 op.interface_version = XEN_SYSCTL_INTERFACE_VERSION; 684 if (HYPERVISOR_sysctl(&op) == 0) 685 nphyscpus = pi->threads_per_core * 686 pi->cores_per_socket * pi->sockets_per_node * 687 pi->nr_nodes; 688 } 689 return (nphyscpus); 690 } 691 #endif 692 693 uint_t 694 do_erratum_298(struct cpu *cpu) 695 { 696 static int osvwrc = -3; 697 extern int osvw_opteron_erratum(cpu_t *, uint_t); 698 699 /* 700 * L2 Eviction May Occur During Processor Operation To Set 701 * Accessed or Dirty Bit. 702 */ 703 if (osvwrc == -3) { 704 osvwrc = osvw_opteron_erratum(cpu, 298); 705 } else { 706 /* osvw return codes should be consistent for all cpus */ 707 ASSERT(osvwrc == osvw_opteron_erratum(cpu, 298)); 708 } 709 710 switch (osvwrc) { 711 case 0: /* erratum is not present: do nothing */ 712 break; 713 case 1: /* erratum is present: BIOS workaround applied */ 714 /* 715 * check if workaround is actually in place and issue warning 716 * if not. 717 */ 718 if (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) || 719 ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0)) { 720 #if defined(OPTERON_ERRATUM_298) 721 opteron_erratum_298++; 722 #else 723 workaround_warning(cpu, 298); 724 return (1); 725 #endif 726 } 727 break; 728 case -1: /* cannot determine via osvw: check cpuid */ 729 if ((cpuid_opteron_erratum(cpu, 298) > 0) && 730 (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) || 731 ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0))) { 732 #if defined(OPTERON_ERRATUM_298) 733 opteron_erratum_298++; 734 #else 735 workaround_warning(cpu, 298); 736 return (1); 737 #endif 738 } 739 break; 740 } 741 return (0); 742 } 743 744 uint_t 745 workaround_errata(struct cpu *cpu) 746 { 747 uint_t missing = 0; 748 749 ASSERT(cpu == CPU); 750 751 /*LINTED*/ 752 if (cpuid_opteron_erratum(cpu, 88) > 0) { 753 /* 754 * SWAPGS May Fail To Read Correct GS Base 755 */ 756 #if defined(OPTERON_ERRATUM_88) 757 /* 758 * The workaround is an mfence in the relevant assembler code 759 */ 760 opteron_erratum_88++; 761 #else 762 workaround_warning(cpu, 88); 763 missing++; 764 #endif 765 } 766 767 if (cpuid_opteron_erratum(cpu, 91) > 0) { 768 /* 769 * Software Prefetches May Report A Page Fault 770 */ 771 #if defined(OPTERON_ERRATUM_91) 772 /* 773 * fix is in trap.c 774 */ 775 opteron_erratum_91++; 776 #else 777 workaround_warning(cpu, 91); 778 missing++; 779 #endif 780 } 781 782 if (cpuid_opteron_erratum(cpu, 93) > 0) { 783 /* 784 * RSM Auto-Halt Restart Returns to Incorrect RIP 785 */ 786 #if defined(OPTERON_ERRATUM_93) 787 /* 788 * fix is in trap.c 789 */ 790 opteron_erratum_93++; 791 #else 792 workaround_warning(cpu, 93); 793 missing++; 794 #endif 795 } 796 797 /*LINTED*/ 798 if (cpuid_opteron_erratum(cpu, 95) > 0) { 799 /* 800 * RET Instruction May Return to Incorrect EIP 801 */ 802 #if defined(OPTERON_ERRATUM_95) 803 #if defined(_LP64) 804 /* 805 * Workaround this by ensuring that 32-bit user code and 806 * 64-bit kernel code never occupy the same address 807 * range mod 4G. 808 */ 809 if (_userlimit32 > 0xc0000000ul) 810 *(uintptr_t *)&_userlimit32 = 0xc0000000ul; 811 812 /*LINTED*/ 813 ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u); 814 opteron_erratum_95++; 815 #endif /* _LP64 */ 816 #else 817 workaround_warning(cpu, 95); 818 missing++; 819 #endif 820 } 821 822 if (cpuid_opteron_erratum(cpu, 100) > 0) { 823 /* 824 * Compatibility Mode Branches Transfer to Illegal Address 825 */ 826 #if defined(OPTERON_ERRATUM_100) 827 /* 828 * fix is in trap.c 829 */ 830 opteron_erratum_100++; 831 #else 832 workaround_warning(cpu, 100); 833 missing++; 834 #endif 835 } 836 837 /*LINTED*/ 838 if (cpuid_opteron_erratum(cpu, 108) > 0) { 839 /* 840 * CPUID Instruction May Return Incorrect Model Number In 841 * Some Processors 842 */ 843 #if defined(OPTERON_ERRATUM_108) 844 /* 845 * (Our cpuid-handling code corrects the model number on 846 * those processors) 847 */ 848 #else 849 workaround_warning(cpu, 108); 850 missing++; 851 #endif 852 } 853 854 /*LINTED*/ 855 if (cpuid_opteron_erratum(cpu, 109) > 0) do { 856 /* 857 * Certain Reverse REP MOVS May Produce Unpredictable Behaviour 858 */ 859 #if defined(OPTERON_ERRATUM_109) 860 /* 861 * The "workaround" is to print a warning to upgrade the BIOS 862 */ 863 uint64_t value; 864 const uint_t msr = MSR_AMD_PATCHLEVEL; 865 int err; 866 867 if ((err = checked_rdmsr(msr, &value)) != 0) { 868 msr_warning(cpu, "rd", msr, err); 869 workaround_warning(cpu, 109); 870 missing++; 871 } 872 if (value == 0) 873 opteron_erratum_109++; 874 #else 875 workaround_warning(cpu, 109); 876 missing++; 877 #endif 878 /*CONSTANTCONDITION*/ 879 } while (0); 880 881 /*LINTED*/ 882 if (cpuid_opteron_erratum(cpu, 121) > 0) { 883 /* 884 * Sequential Execution Across Non_Canonical Boundary Caused 885 * Processor Hang 886 */ 887 #if defined(OPTERON_ERRATUM_121) 888 #if defined(_LP64) 889 /* 890 * Erratum 121 is only present in long (64 bit) mode. 891 * Workaround is to include the page immediately before the 892 * va hole to eliminate the possibility of system hangs due to 893 * sequential execution across the va hole boundary. 894 */ 895 if (opteron_erratum_121) 896 opteron_erratum_121++; 897 else { 898 if (hole_start) { 899 hole_start -= PAGESIZE; 900 } else { 901 /* 902 * hole_start not yet initialized by 903 * mmu_init. Initialize hole_start 904 * with value to be subtracted. 905 */ 906 hole_start = PAGESIZE; 907 } 908 opteron_erratum_121++; 909 } 910 #endif /* _LP64 */ 911 #else 912 workaround_warning(cpu, 121); 913 missing++; 914 #endif 915 } 916 917 /*LINTED*/ 918 if (cpuid_opteron_erratum(cpu, 122) > 0) do { 919 /* 920 * TLB Flush Filter May Cause Coherency Problem in 921 * Multiprocessor Systems 922 */ 923 #if defined(OPTERON_ERRATUM_122) 924 uint64_t value; 925 const uint_t msr = MSR_AMD_HWCR; 926 int error; 927 928 /* 929 * Erratum 122 is only present in MP configurations (multi-core 930 * or multi-processor). 931 */ 932 #if defined(__xpv) 933 if (!DOMAIN_IS_INITDOMAIN(xen_info)) 934 break; 935 if (!opteron_erratum_122 && xen_get_nphyscpus() == 1) 936 break; 937 #else 938 if (!opteron_erratum_122 && opteron_get_nnodes() == 1 && 939 cpuid_get_ncpu_per_chip(cpu) == 1) 940 break; 941 #endif 942 /* disable TLB Flush Filter */ 943 944 if ((error = checked_rdmsr(msr, &value)) != 0) { 945 msr_warning(cpu, "rd", msr, error); 946 workaround_warning(cpu, 122); 947 missing++; 948 } else { 949 value |= (uint64_t)AMD_HWCR_FFDIS; 950 if ((error = checked_wrmsr(msr, value)) != 0) { 951 msr_warning(cpu, "wr", msr, error); 952 workaround_warning(cpu, 122); 953 missing++; 954 } 955 } 956 opteron_erratum_122++; 957 #else 958 workaround_warning(cpu, 122); 959 missing++; 960 #endif 961 /*CONSTANTCONDITION*/ 962 } while (0); 963 964 /*LINTED*/ 965 if (cpuid_opteron_erratum(cpu, 123) > 0) do { 966 /* 967 * Bypassed Reads May Cause Data Corruption of System Hang in 968 * Dual Core Processors 969 */ 970 #if defined(OPTERON_ERRATUM_123) 971 uint64_t value; 972 const uint_t msr = MSR_AMD_PATCHLEVEL; 973 int err; 974 975 /* 976 * Erratum 123 applies only to multi-core cpus. 977 */ 978 if (cpuid_get_ncpu_per_chip(cpu) < 2) 979 break; 980 #if defined(__xpv) 981 if (!DOMAIN_IS_INITDOMAIN(xen_info)) 982 break; 983 #endif 984 /* 985 * The "workaround" is to print a warning to upgrade the BIOS 986 */ 987 if ((err = checked_rdmsr(msr, &value)) != 0) { 988 msr_warning(cpu, "rd", msr, err); 989 workaround_warning(cpu, 123); 990 missing++; 991 } 992 if (value == 0) 993 opteron_erratum_123++; 994 #else 995 workaround_warning(cpu, 123); 996 missing++; 997 998 #endif 999 /*CONSTANTCONDITION*/ 1000 } while (0); 1001 1002 /*LINTED*/ 1003 if (cpuid_opteron_erratum(cpu, 131) > 0) do { 1004 /* 1005 * Multiprocessor Systems with Four or More Cores May Deadlock 1006 * Waiting for a Probe Response 1007 */ 1008 #if defined(OPTERON_ERRATUM_131) 1009 uint64_t nbcfg; 1010 const uint_t msr = MSR_AMD_NB_CFG; 1011 const uint64_t wabits = 1012 AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR; 1013 int error; 1014 1015 /* 1016 * Erratum 131 applies to any system with four or more cores. 1017 */ 1018 if (opteron_erratum_131) 1019 break; 1020 #if defined(__xpv) 1021 if (!DOMAIN_IS_INITDOMAIN(xen_info)) 1022 break; 1023 if (xen_get_nphyscpus() < 4) 1024 break; 1025 #else 1026 if (opteron_get_nnodes() * cpuid_get_ncpu_per_chip(cpu) < 4) 1027 break; 1028 #endif 1029 /* 1030 * Print a warning if neither of the workarounds for 1031 * erratum 131 is present. 1032 */ 1033 if ((error = checked_rdmsr(msr, &nbcfg)) != 0) { 1034 msr_warning(cpu, "rd", msr, error); 1035 workaround_warning(cpu, 131); 1036 missing++; 1037 } else if ((nbcfg & wabits) == 0) { 1038 opteron_erratum_131++; 1039 } else { 1040 /* cannot have both workarounds set */ 1041 ASSERT((nbcfg & wabits) != wabits); 1042 } 1043 #else 1044 workaround_warning(cpu, 131); 1045 missing++; 1046 #endif 1047 /*CONSTANTCONDITION*/ 1048 } while (0); 1049 1050 /* 1051 * This isn't really an erratum, but for convenience the 1052 * detection/workaround code lives here and in cpuid_opteron_erratum. 1053 */ 1054 if (cpuid_opteron_erratum(cpu, 6336786) > 0) { 1055 #if defined(OPTERON_WORKAROUND_6336786) 1056 /* 1057 * Disable C1-Clock ramping on multi-core/multi-processor 1058 * K8 platforms to guard against TSC drift. 1059 */ 1060 if (opteron_workaround_6336786) { 1061 opteron_workaround_6336786++; 1062 #if defined(__xpv) 1063 } else if ((DOMAIN_IS_INITDOMAIN(xen_info) && 1064 xen_get_nphyscpus() > 1) || 1065 opteron_workaround_6336786_UP) { 1066 /* 1067 * XXPV Hmm. We can't walk the Northbridges on 1068 * the hypervisor; so just complain and drive 1069 * on. This probably needs to be fixed in 1070 * the hypervisor itself. 1071 */ 1072 opteron_workaround_6336786++; 1073 workaround_warning(cpu, 6336786); 1074 #else /* __xpv */ 1075 } else if ((opteron_get_nnodes() * 1076 cpuid_get_ncpu_per_chip(cpu) > 1) || 1077 opteron_workaround_6336786_UP) { 1078 1079 uint_t node, nnodes; 1080 uint8_t data; 1081 1082 nnodes = opteron_get_nnodes(); 1083 for (node = 0; node < nnodes; node++) { 1084 /* 1085 * Clear PMM7[1:0] (function 3, offset 0x87) 1086 * Northbridge device is the node id + 24. 1087 */ 1088 data = pci_getb_func(0, node + 24, 3, 0x87); 1089 data &= 0xFC; 1090 pci_putb_func(0, node + 24, 3, 0x87, data); 1091 } 1092 opteron_workaround_6336786++; 1093 #endif /* __xpv */ 1094 } 1095 #else 1096 workaround_warning(cpu, 6336786); 1097 missing++; 1098 #endif 1099 } 1100 1101 /*LINTED*/ 1102 /* 1103 * Mutex primitives don't work as expected. 1104 */ 1105 if (cpuid_opteron_erratum(cpu, 6323525) > 0) { 1106 #if defined(OPTERON_WORKAROUND_6323525) 1107 /* 1108 * This problem only occurs with 2 or more cores. If bit in 1109 * MSR_AMD_BU_CFG set, then not applicable. The workaround 1110 * is to patch the semaphone routines with the lfence 1111 * instruction to provide necessary load memory barrier with 1112 * possible subsequent read-modify-write ops. 1113 * 1114 * It is too early in boot to call the patch routine so 1115 * set erratum variable to be done in startup_end(). 1116 */ 1117 if (opteron_workaround_6323525) { 1118 opteron_workaround_6323525++; 1119 #if defined(__xpv) 1120 } else if (x86_feature & X86_SSE2) { 1121 if (DOMAIN_IS_INITDOMAIN(xen_info)) { 1122 /* 1123 * XXPV Use dom0_msr here when extended 1124 * operations are supported? 1125 */ 1126 if (xen_get_nphyscpus() > 1) 1127 opteron_workaround_6323525++; 1128 } else { 1129 /* 1130 * We have no way to tell how many physical 1131 * cpus there are, or even if this processor 1132 * has the problem, so enable the workaround 1133 * unconditionally (at some performance cost). 1134 */ 1135 opteron_workaround_6323525++; 1136 } 1137 #else /* __xpv */ 1138 } else if ((x86_feature & X86_SSE2) && ((opteron_get_nnodes() * 1139 cpuid_get_ncpu_per_chip(cpu)) > 1)) { 1140 if ((xrdmsr(MSR_AMD_BU_CFG) & (UINT64_C(1) << 33)) == 0) 1141 opteron_workaround_6323525++; 1142 #endif /* __xpv */ 1143 } 1144 #else 1145 workaround_warning(cpu, 6323525); 1146 missing++; 1147 #endif 1148 } 1149 1150 missing += do_erratum_298(cpu); 1151 1152 #ifdef __xpv 1153 return (0); 1154 #else 1155 return (missing); 1156 #endif 1157 } 1158 1159 void 1160 workaround_errata_end() 1161 { 1162 #if defined(OPTERON_ERRATUM_88) 1163 if (opteron_erratum_88) 1164 workaround_applied(88); 1165 #endif 1166 #if defined(OPTERON_ERRATUM_91) 1167 if (opteron_erratum_91) 1168 workaround_applied(91); 1169 #endif 1170 #if defined(OPTERON_ERRATUM_93) 1171 if (opteron_erratum_93) 1172 workaround_applied(93); 1173 #endif 1174 #if defined(OPTERON_ERRATUM_95) 1175 if (opteron_erratum_95) 1176 workaround_applied(95); 1177 #endif 1178 #if defined(OPTERON_ERRATUM_100) 1179 if (opteron_erratum_100) 1180 workaround_applied(100); 1181 #endif 1182 #if defined(OPTERON_ERRATUM_108) 1183 if (opteron_erratum_108) 1184 workaround_applied(108); 1185 #endif 1186 #if defined(OPTERON_ERRATUM_109) 1187 if (opteron_erratum_109) { 1188 cmn_err(CE_WARN, 1189 "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)" 1190 " processor\nerratum 109 was not detected; updating your" 1191 " system's BIOS to a version\ncontaining this" 1192 " microcode patch is HIGHLY recommended or erroneous" 1193 " system\noperation may occur.\n"); 1194 } 1195 #endif 1196 #if defined(OPTERON_ERRATUM_121) 1197 if (opteron_erratum_121) 1198 workaround_applied(121); 1199 #endif 1200 #if defined(OPTERON_ERRATUM_122) 1201 if (opteron_erratum_122) 1202 workaround_applied(122); 1203 #endif 1204 #if defined(OPTERON_ERRATUM_123) 1205 if (opteron_erratum_123) { 1206 cmn_err(CE_WARN, 1207 "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)" 1208 " processor\nerratum 123 was not detected; updating your" 1209 " system's BIOS to a version\ncontaining this" 1210 " microcode patch is HIGHLY recommended or erroneous" 1211 " system\noperation may occur.\n"); 1212 } 1213 #endif 1214 #if defined(OPTERON_ERRATUM_131) 1215 if (opteron_erratum_131) { 1216 cmn_err(CE_WARN, 1217 "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)" 1218 " processor\nerratum 131 was not detected; updating your" 1219 " system's BIOS to a version\ncontaining this" 1220 " microcode patch is HIGHLY recommended or erroneous" 1221 " system\noperation may occur.\n"); 1222 } 1223 #endif 1224 #if defined(OPTERON_WORKAROUND_6336786) 1225 if (opteron_workaround_6336786) 1226 workaround_applied(6336786); 1227 #endif 1228 #if defined(OPTERON_WORKAROUND_6323525) 1229 if (opteron_workaround_6323525) 1230 workaround_applied(6323525); 1231 #endif 1232 #if defined(OPTERON_ERRATUM_298) 1233 if (opteron_erratum_298) { 1234 cmn_err(CE_WARN, 1235 "BIOS microcode patch for AMD 64/Opteron(tm)" 1236 " processor\nerratum 298 was not detected; updating your" 1237 " system's BIOS to a version\ncontaining this" 1238 " microcode patch is HIGHLY recommended or erroneous" 1239 " system\noperation may occur.\n"); 1240 } 1241 #endif 1242 } 1243 1244 static cpuset_t procset; 1245 1246 /* 1247 * Start a single cpu, assuming that the kernel context is available 1248 * to successfully start another cpu. 1249 * 1250 * (For example, real mode code is mapped into the right place 1251 * in memory and is ready to be run.) 1252 */ 1253 int 1254 start_cpu(processorid_t who) 1255 { 1256 void *ctx; 1257 cpu_t *cp; 1258 int delays; 1259 int error = 0; 1260 1261 ASSERT(who != 0); 1262 1263 /* 1264 * Check if there's at least a Mbyte of kmem available 1265 * before attempting to start the cpu. 1266 */ 1267 if (kmem_avail() < 1024 * 1024) { 1268 /* 1269 * Kick off a reap in case that helps us with 1270 * later attempts .. 1271 */ 1272 kmem_reap(); 1273 return (ENOMEM); 1274 } 1275 1276 cp = mp_startup_init(who); 1277 if ((ctx = mach_cpucontext_alloc(cp)) == NULL || 1278 (error = mach_cpu_start(cp, ctx)) != 0) { 1279 1280 /* 1281 * Something went wrong before we even started it 1282 */ 1283 if (ctx) 1284 cmn_err(CE_WARN, 1285 "cpu%d: failed to start error %d", 1286 cp->cpu_id, error); 1287 else 1288 cmn_err(CE_WARN, 1289 "cpu%d: failed to allocate context", cp->cpu_id); 1290 1291 if (ctx) 1292 mach_cpucontext_free(cp, ctx, error); 1293 else 1294 error = EAGAIN; /* hmm. */ 1295 mp_startup_fini(cp, error); 1296 return (error); 1297 } 1298 1299 for (delays = 0; !CPU_IN_SET(procset, who); delays++) { 1300 if (delays == 500) { 1301 /* 1302 * After five seconds, things are probably looking 1303 * a bit bleak - explain the hang. 1304 */ 1305 cmn_err(CE_NOTE, "cpu%d: started, " 1306 "but not running in the kernel yet", who); 1307 } else if (delays > 2000) { 1308 /* 1309 * We waited at least 20 seconds, bail .. 1310 */ 1311 error = ETIMEDOUT; 1312 cmn_err(CE_WARN, "cpu%d: timed out", who); 1313 mach_cpucontext_free(cp, ctx, error); 1314 mp_startup_fini(cp, error); 1315 return (error); 1316 } 1317 1318 /* 1319 * wait at least 10ms, then check again.. 1320 */ 1321 delay(USEC_TO_TICK_ROUNDUP(10000)); 1322 } 1323 1324 mach_cpucontext_free(cp, ctx, 0); 1325 1326 #ifndef __xpv 1327 if (tsc_gethrtime_enable) 1328 tsc_sync_master(who); 1329 #endif 1330 1331 if (dtrace_cpu_init != NULL) { 1332 /* 1333 * DTrace CPU initialization expects cpu_lock to be held. 1334 */ 1335 mutex_enter(&cpu_lock); 1336 (*dtrace_cpu_init)(who); 1337 mutex_exit(&cpu_lock); 1338 } 1339 1340 while (!CPU_IN_SET(cpu_ready_set, who)) 1341 delay(1); 1342 1343 return (0); 1344 } 1345 1346 1347 /*ARGSUSED*/ 1348 void 1349 start_other_cpus(int cprboot) 1350 { 1351 uint_t who; 1352 uint_t skipped = 0; 1353 uint_t bootcpuid = 0; 1354 1355 /* 1356 * Initialize our own cpu_info. 1357 */ 1358 init_cpu_info(CPU); 1359 1360 cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_idstr); 1361 cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_brandstr); 1362 1363 /* 1364 * Initialize our syscall handlers 1365 */ 1366 init_cpu_syscall(CPU); 1367 1368 /* 1369 * Take the boot cpu out of the mp_cpus set because we know 1370 * it's already running. Add it to the cpu_ready_set for 1371 * precisely the same reason. 1372 */ 1373 CPUSET_DEL(mp_cpus, bootcpuid); 1374 CPUSET_ADD(cpu_ready_set, bootcpuid); 1375 1376 /* 1377 * if only 1 cpu or not using MP, skip the rest of this 1378 */ 1379 if (CPUSET_ISNULL(mp_cpus) || use_mp == 0) { 1380 if (use_mp == 0) 1381 cmn_err(CE_CONT, "?***** Not in MP mode\n"); 1382 goto done; 1383 } 1384 1385 /* 1386 * perform such initialization as is needed 1387 * to be able to take CPUs on- and off-line. 1388 */ 1389 cpu_pause_init(); 1390 1391 xc_init(); /* initialize processor crosscalls */ 1392 1393 if (mach_cpucontext_init() != 0) 1394 goto done; 1395 1396 flushes_require_xcalls = 1; 1397 1398 /* 1399 * We lock our affinity to the master CPU to ensure that all slave CPUs 1400 * do their TSC syncs with the same CPU. 1401 */ 1402 affinity_set(CPU_CURRENT); 1403 1404 for (who = 0; who < NCPU; who++) { 1405 1406 if (!CPU_IN_SET(mp_cpus, who)) 1407 continue; 1408 ASSERT(who != bootcpuid); 1409 if (ncpus >= max_ncpus) { 1410 skipped = who; 1411 continue; 1412 } 1413 if (start_cpu(who) != 0) 1414 CPUSET_DEL(mp_cpus, who); 1415 1416 mutex_enter(&cpu_lock); 1417 cpu_state_change_notify(who, CPU_SETUP); 1418 mutex_exit(&cpu_lock); 1419 } 1420 1421 /* Free the space allocated to hold the microcode file */ 1422 ucode_cleanup(); 1423 1424 affinity_clear(); 1425 1426 if (skipped) { 1427 cmn_err(CE_NOTE, 1428 "System detected %d cpus, but " 1429 "only %d cpu(s) were enabled during boot.", 1430 skipped + 1, ncpus); 1431 cmn_err(CE_NOTE, 1432 "Use \"boot-ncpus\" parameter to enable more CPU(s). " 1433 "See eeprom(1M)."); 1434 } 1435 1436 done: 1437 workaround_errata_end(); 1438 mach_cpucontext_fini(); 1439 1440 cmi_post_mpstartup(); 1441 } 1442 1443 /* 1444 * Dummy functions - no i86pc platforms support dynamic cpu allocation. 1445 */ 1446 /*ARGSUSED*/ 1447 int 1448 mp_cpu_configure(int cpuid) 1449 { 1450 return (ENOTSUP); /* not supported */ 1451 } 1452 1453 /*ARGSUSED*/ 1454 int 1455 mp_cpu_unconfigure(int cpuid) 1456 { 1457 return (ENOTSUP); /* not supported */ 1458 } 1459 1460 /* 1461 * Startup function for 'other' CPUs (besides boot cpu). 1462 * Called from real_mode_start. 1463 * 1464 * WARNING: until CPU_READY is set, mp_startup and routines called by 1465 * mp_startup should not call routines (e.g. kmem_free) that could call 1466 * hat_unload which requires CPU_READY to be set. 1467 */ 1468 void 1469 mp_startup(void) 1470 { 1471 struct cpu *cp = CPU; 1472 uint_t new_x86_feature; 1473 #ifndef __xpv 1474 extern void cpupm_init(cpu_t *); 1475 #endif 1476 const char *fmt = "?cpu%d: %b\n"; 1477 1478 /* 1479 * We need to get TSC on this proc synced (i.e., any delta 1480 * from cpu0 accounted for) as soon as we can, because many 1481 * many things use gethrtime/pc_gethrestime, including 1482 * interrupts, cmn_err, etc. 1483 */ 1484 1485 /* Let cpu0 continue into tsc_sync_master() */ 1486 CPUSET_ATOMIC_ADD(procset, cp->cpu_id); 1487 1488 #ifndef __xpv 1489 if (tsc_gethrtime_enable) 1490 tsc_sync_slave(); 1491 #endif 1492 1493 /* 1494 * Once this was done from assembly, but it's safer here; if 1495 * it blocks, we need to be able to swtch() to and from, and 1496 * since we get here by calling t_pc, we need to do that call 1497 * before swtch() overwrites it. 1498 */ 1499 (void) (*ap_mlsetup)(); 1500 1501 new_x86_feature = cpuid_pass1(cp); 1502 1503 #ifndef __xpv 1504 /* 1505 * Program this cpu's PAT 1506 */ 1507 if (x86_feature & X86_PAT) 1508 pat_sync(); 1509 #endif 1510 1511 /* 1512 * Set up TSC_AUX to contain the cpuid for this processor 1513 * for the rdtscp instruction. 1514 */ 1515 if (x86_feature & X86_TSCP) 1516 (void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id); 1517 1518 /* 1519 * Initialize this CPU's syscall handlers 1520 */ 1521 init_cpu_syscall(cp); 1522 1523 /* 1524 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the 1525 * highest level at which a routine is permitted to block on 1526 * an adaptive mutex (allows for cpu poke interrupt in case 1527 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks 1528 * device interrupts that may end up in the hat layer issuing cross 1529 * calls before CPU_READY is set. 1530 */ 1531 splx(ipltospl(LOCK_LEVEL)); 1532 sti(); 1533 1534 /* 1535 * Do a sanity check to make sure this new CPU is a sane thing 1536 * to add to the collection of processors running this system. 1537 * 1538 * XXX Clearly this needs to get more sophisticated, if x86 1539 * systems start to get built out of heterogenous CPUs; as is 1540 * likely to happen once the number of processors in a configuration 1541 * gets large enough. 1542 */ 1543 if ((x86_feature & new_x86_feature) != x86_feature) { 1544 cmn_err(CE_CONT, fmt, cp->cpu_id, new_x86_feature, 1545 FMT_X86_FEATURE); 1546 cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id); 1547 } 1548 1549 /* 1550 * We do not support cpus with mixed monitor/mwait support if the 1551 * boot cpu supports monitor/mwait. 1552 */ 1553 if ((x86_feature & ~new_x86_feature) & X86_MWAIT) 1554 panic("unsupported mixed cpu monitor/mwait support detected"); 1555 1556 /* 1557 * We could be more sophisticated here, and just mark the CPU 1558 * as "faulted" but at this point we'll opt for the easier 1559 * answer of dieing horribly. Provided the boot cpu is ok, 1560 * the system can be recovered by booting with use_mp set to zero. 1561 */ 1562 if (workaround_errata(cp) != 0) 1563 panic("critical workaround(s) missing for cpu%d", cp->cpu_id); 1564 1565 cpuid_pass2(cp); 1566 cpuid_pass3(cp); 1567 (void) cpuid_pass4(cp); 1568 1569 init_cpu_info(cp); 1570 1571 mutex_enter(&cpu_lock); 1572 1573 cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_EXISTS; 1574 1575 cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr); 1576 cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr); 1577 1578 if (dtrace_cpu_init != NULL) { 1579 (*dtrace_cpu_init)(cp->cpu_id); 1580 } 1581 1582 /* 1583 * Fill out cpu_ucode_info. Update microcode if necessary. 1584 */ 1585 ucode_check(cp); 1586 1587 mutex_exit(&cpu_lock); 1588 1589 post_startup_cpu_fixups(); 1590 1591 /* 1592 * Enable preemption here so that contention for any locks acquired 1593 * later in mp_startup may be preempted if the thread owning those 1594 * locks is continously executing on other CPUs (for example, this 1595 * CPU must be preemptible to allow other CPUs to pause it during their 1596 * startup phases). It's safe to enable preemption here because the 1597 * CPU state is pretty-much fully constructed. 1598 */ 1599 curthread->t_preempt = 0; 1600 1601 /* The base spl should still be at LOCK LEVEL here */ 1602 ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL)); 1603 set_base_spl(); /* Restore the spl to its proper value */ 1604 1605 #ifndef __xpv 1606 cpupm_init(cp); 1607 #endif 1608 add_cpunode2devtree(cp->cpu_id, cp->cpu_m.mcpu_cpi); 1609 1610 /* 1611 * Processor group initialization for this CPU is dependent on the 1612 * cpuid probing, which must be done in the context of the current 1613 * CPU, as well as the CPU's device node initialization (for ACPI). 1614 */ 1615 mutex_enter(&cpu_lock); 1616 pghw_physid_create(cp); 1617 pg_cpu_init(cp); 1618 pg_cmt_cpu_startup(cp); 1619 mutex_exit(&cpu_lock); 1620 1621 /* Enable interrupts */ 1622 (void) spl0(); 1623 1624 mutex_enter(&cpu_lock); 1625 cpu_enable_intr(cp); 1626 cpu_add_active(cp); 1627 mutex_exit(&cpu_lock); 1628 1629 #ifndef __xpv 1630 { 1631 /* 1632 * Set up the CPU module for this CPU. This can't be done 1633 * before this CPU is made CPU_READY, because we may (in 1634 * heterogeneous systems) need to go load another CPU module. 1635 * The act of attempting to load a module may trigger a 1636 * cross-call, which will ASSERT unless this cpu is CPU_READY. 1637 */ 1638 cmi_hdl_t hdl; 1639 1640 if ((hdl = cmi_init(CMI_HDL_NATIVE, cmi_ntv_hwchipid(CPU), 1641 cmi_ntv_hwcoreid(CPU), cmi_ntv_hwstrandid(CPU))) != NULL) { 1642 if (x86_feature & X86_MCA) 1643 cmi_mca_init(hdl); 1644 } 1645 } 1646 #endif /* __xpv */ 1647 1648 if (boothowto & RB_DEBUG) 1649 kdi_cpu_init(); 1650 1651 /* 1652 * Setting the bit in cpu_ready_set must be the last operation in 1653 * processor initialization; the boot CPU will continue to boot once 1654 * it sees this bit set for all active CPUs. 1655 */ 1656 CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id); 1657 1658 /* 1659 * Because mp_startup() gets fired off after init() starts, we 1660 * can't use the '?' trick to do 'boot -v' printing - so we 1661 * always direct the 'cpu .. online' messages to the log. 1662 */ 1663 cmn_err(CE_CONT, "!cpu%d initialization complete - online\n", 1664 cp->cpu_id); 1665 1666 /* 1667 * Now we are done with the startup thread, so free it up. 1668 */ 1669 thread_exit(); 1670 panic("mp_startup: cannot return"); 1671 /*NOTREACHED*/ 1672 } 1673 1674 1675 /* 1676 * Start CPU on user request. 1677 */ 1678 /* ARGSUSED */ 1679 int 1680 mp_cpu_start(struct cpu *cp) 1681 { 1682 ASSERT(MUTEX_HELD(&cpu_lock)); 1683 return (0); 1684 } 1685 1686 /* 1687 * Stop CPU on user request. 1688 */ 1689 /* ARGSUSED */ 1690 int 1691 mp_cpu_stop(struct cpu *cp) 1692 { 1693 extern int cbe_psm_timer_mode; 1694 ASSERT(MUTEX_HELD(&cpu_lock)); 1695 1696 #ifdef __xpv 1697 /* 1698 * We can't offline vcpu0. 1699 */ 1700 if (cp->cpu_id == 0) 1701 return (EBUSY); 1702 #endif 1703 1704 /* 1705 * If TIMER_PERIODIC mode is used, CPU0 is the one running it; 1706 * can't stop it. (This is true only for machines with no TSC.) 1707 */ 1708 1709 if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0)) 1710 return (EBUSY); 1711 1712 return (0); 1713 } 1714 1715 /* 1716 * Take the specified CPU out of participation in interrupts. 1717 */ 1718 int 1719 cpu_disable_intr(struct cpu *cp) 1720 { 1721 if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS) 1722 return (EBUSY); 1723 1724 cp->cpu_flags &= ~CPU_ENABLE; 1725 return (0); 1726 } 1727 1728 /* 1729 * Allow the specified CPU to participate in interrupts. 1730 */ 1731 void 1732 cpu_enable_intr(struct cpu *cp) 1733 { 1734 ASSERT(MUTEX_HELD(&cpu_lock)); 1735 cp->cpu_flags |= CPU_ENABLE; 1736 psm_enable_intr(cp->cpu_id); 1737 } 1738 1739 1740 /*ARGSUSED*/ 1741 void 1742 mp_cpu_faulted_enter(struct cpu *cp) 1743 { 1744 #ifndef __xpv 1745 cmi_hdl_t hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp), 1746 cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp)); 1747 1748 if (hdl != NULL) { 1749 cmi_faulted_enter(hdl); 1750 cmi_hdl_rele(hdl); 1751 } 1752 #endif 1753 } 1754 1755 /*ARGSUSED*/ 1756 void 1757 mp_cpu_faulted_exit(struct cpu *cp) 1758 { 1759 #ifndef __xpv 1760 cmi_hdl_t hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp), 1761 cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp)); 1762 1763 if (hdl != NULL) { 1764 cmi_faulted_exit(hdl); 1765 cmi_hdl_rele(hdl); 1766 } 1767 #endif 1768 } 1769 1770 /* 1771 * The following two routines are used as context operators on threads belonging 1772 * to processes with a private LDT (see sysi86). Due to the rarity of such 1773 * processes, these routines are currently written for best code readability and 1774 * organization rather than speed. We could avoid checking x86_feature at every 1775 * context switch by installing different context ops, depending on the 1776 * x86_feature flags, at LDT creation time -- one for each combination of fast 1777 * syscall feature flags. 1778 */ 1779 1780 /*ARGSUSED*/ 1781 void 1782 cpu_fast_syscall_disable(void *arg) 1783 { 1784 if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP)) 1785 cpu_sep_disable(); 1786 if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC)) 1787 cpu_asysc_disable(); 1788 } 1789 1790 /*ARGSUSED*/ 1791 void 1792 cpu_fast_syscall_enable(void *arg) 1793 { 1794 if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP)) 1795 cpu_sep_enable(); 1796 if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC)) 1797 cpu_asysc_enable(); 1798 } 1799 1800 static void 1801 cpu_sep_enable(void) 1802 { 1803 ASSERT(x86_feature & X86_SEP); 1804 ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL); 1805 1806 wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL); 1807 } 1808 1809 static void 1810 cpu_sep_disable(void) 1811 { 1812 ASSERT(x86_feature & X86_SEP); 1813 ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL); 1814 1815 /* 1816 * Setting the SYSENTER_CS_MSR register to 0 causes software executing 1817 * the sysenter or sysexit instruction to trigger a #gp fault. 1818 */ 1819 wrmsr(MSR_INTC_SEP_CS, 0); 1820 } 1821 1822 static void 1823 cpu_asysc_enable(void) 1824 { 1825 ASSERT(x86_feature & X86_ASYSC); 1826 ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL); 1827 1828 wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) | 1829 (uint64_t)(uintptr_t)AMD_EFER_SCE); 1830 } 1831 1832 static void 1833 cpu_asysc_disable(void) 1834 { 1835 ASSERT(x86_feature & X86_ASYSC); 1836 ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL); 1837 1838 /* 1839 * Turn off the SCE (syscall enable) bit in the EFER register. Software 1840 * executing syscall or sysret with this bit off will incur a #ud trap. 1841 */ 1842 wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) & 1843 ~((uint64_t)(uintptr_t)AMD_EFER_SCE)); 1844 } 1845