xref: /titanic_44/usr/src/uts/i86pc/os/mp_startup.c (revision 8eea8e29cc4374d1ee24c25a07f45af132db3499)
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, Version 1.0 only
6  * (the "License").  You may not use this file except in compliance
7  * with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/types.h>
30 #include <sys/thread.h>
31 #include <sys/cpuvar.h>
32 #include <sys/t_lock.h>
33 #include <sys/param.h>
34 #include <sys/proc.h>
35 #include <sys/disp.h>
36 #include <sys/mmu.h>
37 #include <sys/class.h>
38 #include <sys/cmn_err.h>
39 #include <sys/debug.h>
40 #include <sys/asm_linkage.h>
41 #include <sys/x_call.h>
42 #include <sys/systm.h>
43 #include <sys/var.h>
44 #include <sys/vtrace.h>
45 #include <vm/hat.h>
46 #include <sys/mmu.h>
47 #include <vm/as.h>
48 #include <vm/seg_kmem.h>
49 #include <sys/segments.h>
50 #include <sys/kmem.h>
51 #include <sys/stack.h>
52 #include <sys/smp_impldefs.h>
53 #include <sys/x86_archext.h>
54 #include <sys/machsystm.h>
55 #include <sys/traptrace.h>
56 #include <sys/clock.h>
57 #include <sys/cpc_impl.h>
58 #include <sys/chip.h>
59 #include <sys/dtrace.h>
60 #include <sys/archsystm.h>
61 #include <sys/fp.h>
62 #include <sys/reboot.h>
63 #include <sys/kdi.h>
64 #include <vm/hat_i86.h>
65 #include <sys/memnode.h>
66 
67 struct cpu	cpus[1];			/* CPU data */
68 struct cpu	*cpu[NCPU] = {&cpus[0]};	/* pointers to all CPUs */
69 cpu_core_t	cpu_core[NCPU];			/* cpu_core structures */
70 
71 /*
72  * Useful for disabling MP bring-up for an MP capable kernel
73  * (a kernel that was built with MP defined)
74  */
75 int use_mp = 1;
76 
77 int mp_cpus = 0x1;	/* to be set by platform specific module	*/
78 
79 /*
80  * This variable is used by the hat layer to decide whether or not
81  * critical sections are needed to prevent race conditions.  For sun4m,
82  * this variable is set once enough MP initialization has been done in
83  * order to allow cross calls.
84  */
85 int flushes_require_xcalls = 0;
86 ulong_t	cpu_ready_set = 1;
87 
88 extern	void	real_mode_start(void);
89 extern	void	real_mode_end(void);
90 static 	void	mp_startup(void);
91 
92 static void cpu_sep_enable(void);
93 static void cpu_sep_disable(void);
94 static void cpu_asysc_enable(void);
95 static void cpu_asysc_disable(void);
96 
97 extern int tsc_gethrtime_enable;
98 
99 /*
100  * Init CPU info - get CPU type info for processor_info system call.
101  */
102 void
103 init_cpu_info(struct cpu *cp)
104 {
105 	processor_info_t *pi = &cp->cpu_type_info;
106 	char buf[CPU_IDSTRLEN];
107 
108 	/*
109 	 * Get clock-frequency property for the CPU.
110 	 */
111 	pi->pi_clock = cpu_freq;
112 
113 	(void) strcpy(pi->pi_processor_type, "i386");
114 	if (fpu_exists)
115 		(void) strcpy(pi->pi_fputypes, "i387 compatible");
116 
117 	(void) cpuid_getidstr(cp, buf, sizeof (buf));
118 
119 	cp->cpu_idstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
120 	(void) strcpy(cp->cpu_idstr, buf);
121 
122 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr);
123 
124 	(void) cpuid_getbrandstr(cp, buf, sizeof (buf));
125 	cp->cpu_brandstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
126 	(void) strcpy(cp->cpu_brandstr, buf);
127 
128 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr);
129 }
130 
131 /*
132  * Configure syscall support on this CPU.
133  */
134 /*ARGSUSED*/
135 static void
136 init_cpu_syscall(struct cpu *cp)
137 {
138 	uint64_t value;
139 
140 	kpreempt_disable();
141 
142 #if defined(__amd64)
143 	if (x86_feature & X86_ASYSC) {
144 
145 #if !defined(__lint)
146 		/*
147 		 * The syscall instruction imposes a certain ordering on
148 		 * segment selectors, so we double-check that ordering
149 		 * here.
150 		 */
151 		ASSERT(KDS_SEL == KCS_SEL + 8);
152 		ASSERT(UDS_SEL == U32CS_SEL + 8);
153 		ASSERT(UCS_SEL == U32CS_SEL + 16);
154 #endif
155 		/*
156 		 * Turn syscall/sysret extensions on.
157 		 */
158 		cpu_asysc_enable();
159 
160 		/*
161 		 * Program the magic registers ..
162 		 */
163 		value = ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32;
164 		wrmsr(MSR_AMD_STAR, &value);
165 		value = (uintptr_t)sys_syscall;
166 		wrmsr(MSR_AMD_LSTAR, &value);
167 		value = (uintptr_t)sys_syscall32;
168 		wrmsr(MSR_AMD_CSTAR, &value);
169 
170 		/*
171 		 * This list of flags is masked off the incoming
172 		 * %rfl when we enter the kernel.
173 		 */
174 		value = PS_IE | PS_T;
175 		wrmsr(MSR_AMD_SFMASK, &value);
176 	}
177 #endif
178 
179 	/*
180 	 * On 32-bit kernels, we use sysenter/sysexit because it's too
181 	 * hard to use syscall/sysret, and it is more portable anyway.
182 	 *
183 	 * On 64-bit kernels on Nocona machines, the 32-bit syscall
184 	 * variant isn't available to 32-bit applications, but sysenter is.
185 	 */
186 	if (x86_feature & X86_SEP) {
187 
188 #if !defined(__lint)
189 		/*
190 		 * The sysenter instruction imposes a certain ordering on
191 		 * segment selectors, so we double-check that ordering
192 		 * here. See "sysenter" in Intel document 245471-012, "IA-32
193 		 * Intel Architecture Software Developer's Manual Volume 2:
194 		 * Instruction Set Reference"
195 		 */
196 		ASSERT(KDS_SEL == KCS_SEL + 8);
197 
198 		ASSERT32(UCS_SEL == ((KCS_SEL + 16) | 3));
199 		ASSERT32(UDS_SEL == UCS_SEL + 8);
200 
201 		ASSERT64(U32CS_SEL == ((KCS_SEL + 16) | 3));
202 		ASSERT64(UDS_SEL == U32CS_SEL + 8);
203 #endif
204 
205 		cpu_sep_enable();
206 
207 		/*
208 		 * resume() sets this value to the base of the threads stack
209 		 * via a context handler.
210 		 */
211 		value = 0;
212 		wrmsr(MSR_INTC_SEP_ESP, &value);
213 
214 		value = (uintptr_t)sys_sysenter;
215 		wrmsr(MSR_INTC_SEP_EIP, &value);
216 	}
217 
218 	kpreempt_enable();
219 }
220 
221 /*
222  * Multiprocessor initialization.
223  *
224  * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the
225  * startup and idle threads for the specified CPU.
226  */
227 static void
228 mp_startup_init(int cpun)
229 {
230 #if defined(__amd64)
231 extern void *long_mode_64(void);
232 #endif	/* __amd64 */
233 
234 	struct cpu *cp;
235 	struct tss *ntss;
236 	kthread_id_t tp;
237 	caddr_t	sp;
238 	int size;
239 	proc_t *procp;
240 	extern void idle();
241 	extern void init_intr_threads(struct cpu *);
242 
243 	struct cpu_tables *tablesp;
244 	extern chip_t cpu0_chip;
245 	rm_platter_t *real_mode_platter = (rm_platter_t *)rm_platter_va;
246 
247 #ifdef TRAPTRACE
248 	trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun];
249 #endif
250 
251 	ASSERT(cpun < NCPU && cpu[cpun] == NULL);
252 
253 	if ((cp = kmem_zalloc(sizeof (*cp), KM_NOSLEEP)) == NULL) {
254 		panic("mp_startup_init: cpu%d: "
255 		    "no memory for cpu structure", cpun);
256 		/*NOTREACHED*/
257 	}
258 	procp = curthread->t_procp;
259 
260 	mutex_enter(&cpu_lock);
261 	/*
262 	 * Initialize the dispatcher first.
263 	 */
264 	disp_cpu_init(cp);
265 	mutex_exit(&cpu_lock);
266 
267 	/*
268 	 * Allocate and initialize the startup thread for this CPU.
269 	 * Interrupt and process switch stacks get allocated later
270 	 * when the CPU starts running.
271 	 */
272 	tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
273 	    TS_STOPPED, maxclsyspri);
274 
275 	/*
276 	 * Set state to TS_ONPROC since this thread will start running
277 	 * as soon as the CPU comes online.
278 	 *
279 	 * All the other fields of the thread structure are setup by
280 	 * thread_create().
281 	 */
282 	THREAD_ONPROC(tp, cp);
283 	tp->t_preempt = 1;
284 	tp->t_bound_cpu = cp;
285 	tp->t_affinitycnt = 1;
286 	tp->t_cpu = cp;
287 	tp->t_disp_queue = cp->cpu_disp;
288 
289 	/*
290 	 * Setup thread to start in mp_startup.
291 	 */
292 	sp = tp->t_stk;
293 	tp->t_pc = (uintptr_t)mp_startup;
294 	tp->t_sp = (uintptr_t)(sp - MINFRAME);
295 
296 	cp->cpu_id = cpun;
297 	cp->cpu_self = cp;
298 	cp->cpu_mask = 1 << cpun;
299 	cp->cpu_thread = tp;
300 	cp->cpu_lwp = NULL;
301 	cp->cpu_dispthread = tp;
302 	cp->cpu_dispatch_pri = DISP_PRIO(tp);
303 
304 	/*
305 	 * Bootstrap cpu_chip in case mp_startup blocks
306 	 */
307 	cp->cpu_chip = &cpu0_chip;
308 
309 	/*
310 	 * Now, initialize per-CPU idle thread for this CPU.
311 	 */
312 	tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);
313 
314 	cp->cpu_idle_thread = tp;
315 
316 	tp->t_preempt = 1;
317 	tp->t_bound_cpu = cp;
318 	tp->t_affinitycnt = 1;
319 	tp->t_cpu = cp;
320 	tp->t_disp_queue = cp->cpu_disp;
321 
322 	/*
323 	 * Perform CPC intialization on the new CPU.
324 	 */
325 	kcpc_hw_init(cp);
326 
327 	/*
328 	 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
329 	 * for each CPU.
330 	 */
331 
332 	setup_vaddr_for_ppcopy(cp);
333 
334 	/*
335 	 * Allocate space for page directory, stack, tss, gdt and idt.
336 	 * This assumes that kmem_alloc will return memory which is aligned
337 	 * to the next higher power of 2 or a page(if size > MAXABIG)
338 	 * If this assumption goes wrong at any time due to change in
339 	 * kmem alloc, things may not work as the page directory has to be
340 	 * page aligned
341 	 */
342 	if ((tablesp = kmem_zalloc(sizeof (*tablesp), KM_NOSLEEP)) == NULL)
343 		panic("mp_startup_init: cpu%d cannot allocate tables", cpun);
344 
345 	if ((uintptr_t)tablesp & ~MMU_STD_PAGEMASK) {
346 		kmem_free(tablesp, sizeof (struct cpu_tables));
347 		size = sizeof (struct cpu_tables) + MMU_STD_PAGESIZE;
348 		tablesp = kmem_zalloc(size, KM_NOSLEEP);
349 		tablesp = (struct cpu_tables *)
350 		    (((uintptr_t)tablesp + MMU_STD_PAGESIZE) &
351 		    MMU_STD_PAGEMASK);
352 	}
353 
354 	ntss = cp->cpu_tss = &tablesp->ct_tss;
355 	cp->cpu_gdt = tablesp->ct_gdt;
356 	bcopy(CPU->cpu_gdt, cp->cpu_gdt, NGDT * (sizeof (user_desc_t)));
357 
358 #if defined(__amd64)
359 
360 	/*
361 	 * #DF (double fault).
362 	 */
363 	ntss->tss_ist1 =
364 	    (uint64_t)&tablesp->ct_stack[sizeof (tablesp->ct_stack)];
365 
366 #elif defined(__i386)
367 
368 	ntss->tss_esp0 = ntss->tss_esp1 = ntss->tss_esp2 = ntss->tss_esp =
369 	    (uint32_t)&tablesp->ct_stack[sizeof (tablesp->ct_stack)];
370 
371 	ntss->tss_ss0 = ntss->tss_ss1 = ntss->tss_ss2 = ntss->tss_ss = KDS_SEL;
372 
373 	ntss->tss_eip = (uint32_t)mp_startup;
374 
375 	ntss->tss_cs = KCS_SEL;
376 	ntss->tss_fs = KFS_SEL;
377 	ntss->tss_gs = KGS_SEL;
378 
379 	/*
380 	 * setup kernel %gs.
381 	 */
382 	set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA,
383 	    SEL_KPL, 0, 1);
384 
385 #endif	/* __i386 */
386 
387 	/*
388 	 * Set I/O bit map offset equal to size of TSS segment limit
389 	 * for no I/O permission map. This will cause all user I/O
390 	 * instructions to generate #gp fault.
391 	 */
392 	ntss->tss_bitmapbase = sizeof (*ntss);
393 
394 	/*
395 	 * setup kernel tss.
396 	 */
397 	set_syssegd((system_desc_t *)&cp->cpu_gdt[GDT_KTSS], cp->cpu_tss,
398 	    sizeof (*cp->cpu_tss) -1, SDT_SYSTSS, SEL_KPL);
399 
400 	/*
401 	 * If we have more than one node, each cpu gets a copy of IDT
402 	 * local to its node. If this is a Pentium box, we use cpu 0's
403 	 * IDT. cpu 0's IDT has been made read-only to workaround the
404 	 * cmpxchgl register bug
405 	 */
406 	cp->cpu_idt = CPU->cpu_idt;
407 	if (system_hardware.hd_nodes && x86_type != X86_TYPE_P5) {
408 		cp->cpu_idt = kmem_alloc(sizeof (idt0), KM_SLEEP);
409 		bcopy(idt0, cp->cpu_idt, sizeof (idt0));
410 	}
411 
412 	/*
413 	 * Get interrupt priority data from cpu 0
414 	 */
415 	cp->cpu_pri_data = CPU->cpu_pri_data;
416 
417 	hat_cpu_online(cp);
418 
419 	/* Should remove all entries for the current process/thread here */
420 
421 	/*
422 	 * Fill up the real mode platter to make it easy for real mode code to
423 	 * kick it off. This area should really be one passed by boot to kernel
424 	 * and guaranteed to be below 1MB and aligned to 16 bytes. Should also
425 	 * have identical physical and virtual address in paged mode.
426 	 */
427 	real_mode_platter->rm_idt_base = cp->cpu_idt;
428 	real_mode_platter->rm_idt_lim = sizeof (idt0) - 1;
429 	real_mode_platter->rm_gdt_base = cp->cpu_gdt;
430 	real_mode_platter->rm_gdt_lim = sizeof (gdt0) -1;
431 	real_mode_platter->rm_pdbr = getcr3();
432 	real_mode_platter->rm_cpu = cpun;
433 	real_mode_platter->rm_x86feature = x86_feature;
434 	real_mode_platter->rm_cr4 = cr4_value;
435 
436 #if defined(__amd64)
437 	if (getcr3() > 0xffffffffUL)
438 		panic("Cannot initialize CPUs; kernel's 64-bit page tables\n"
439 			"located above 4G in physical memory (@ 0x%llx).",
440 			(unsigned long long)getcr3());
441 
442 	/*
443 	 * Setup pseudo-descriptors for temporary GDT and IDT for use ONLY
444 	 * by code in real_mode_start():
445 	 *
446 	 * GDT[0]:  NULL selector
447 	 * GDT[1]:  64-bit CS: Long = 1, Present = 1, bits 12, 11 = 1
448 	 *
449 	 * Clear the IDT as interrupts will be off and a limit of 0 will cause
450 	 * the CPU to triple fault and reset on an NMI, seemingly as reasonable
451 	 * a course of action as any other, though it may cause the entire
452 	 * platform to reset in some cases...
453 	 */
454 	real_mode_platter->rm_temp_gdt[0] = 0ULL;
455 	real_mode_platter->rm_temp_gdt[TEMPGDT_KCODE64] = 0x20980000000000ULL;
456 
457 	real_mode_platter->rm_temp_gdt_lim = (ushort_t)
458 	    (sizeof (real_mode_platter->rm_temp_gdt) - 1);
459 	real_mode_platter->rm_temp_gdt_base = rm_platter_pa +
460 	    (uint32_t)(&((rm_platter_t *)0)->rm_temp_gdt);
461 
462 	real_mode_platter->rm_temp_idt_lim = 0;
463 	real_mode_platter->rm_temp_idt_base = 0;
464 
465 	/*
466 	 * Since the CPU needs to jump to protected mode using an identity
467 	 * mapped address, we need to calculate it here.
468 	 */
469 	real_mode_platter->rm_longmode64_addr = rm_platter_pa +
470 	    ((uint32_t)long_mode_64 - (uint32_t)real_mode_start);
471 #endif	/* __amd64 */
472 
473 #ifdef TRAPTRACE
474 	/*
475 	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers for this
476 	 * CPU.
477 	 */
478 	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
479 	ttc->ttc_next = ttc->ttc_first;
480 	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
481 #endif
482 
483 	/*
484 	 * Record that we have another CPU.
485 	 */
486 	mutex_enter(&cpu_lock);
487 	/*
488 	 * Initialize the interrupt threads for this CPU
489 	 */
490 	init_intr_threads(cp);
491 	/*
492 	 * Add CPU to list of available CPUs.  It'll be on the active list
493 	 * after mp_startup().
494 	 */
495 	cpu_add_unit(cp);
496 	mutex_exit(&cpu_lock);
497 }
498 
499 /*
500  * Apply workarounds for known errata, and warn about those that are absent.
501  *
502  * System vendors occasionally create configurations which contain different
503  * revisions of the CPUs that are almost but not exactly the same.  At the
504  * time of writing, this meant that their clock rates were the same, their
505  * feature sets were the same, but the required workaround were -not-
506  * necessarily the same.  So, this routine is invoked on -every- CPU soon
507  * after starting to make sure that the resulting system contains the most
508  * pessimal set of workarounds needed to cope with *any* of the CPUs in the
509  * system.
510  *
511  * These workarounds are based on Rev 3.50 of the Revision Guide for
512  * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, May 2005.
513  */
514 
515 #if defined(OPTERON_ERRATUM_91)
516 int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
517 #endif
518 
519 #if defined(OPTERON_ERRATUM_93)
520 int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
521 #endif
522 
523 #if defined(OPTERON_ERRATUM_100)
524 int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
525 #endif
526 
527 #if defined(OPTERON_ERRATUM_109)
528 int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
529 #endif
530 
531 #if defined(OPTERON_ERRATUM_121)
532 int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
533 #endif
534 
535 #if defined(OPTERON_ERRATUM_122)
536 int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
537 #endif
538 
539 #if defined(OPTERON_ERRATUM_123)
540 int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
541 #endif
542 
543 
544 #define	WARNING(cpu, n)						\
545 	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %d",	\
546 	    (cpu)->cpu_id, (n))
547 
548 uint_t
549 workaround_errata(struct cpu *cpu)
550 {
551 	uint_t missing = 0;
552 
553 	ASSERT(cpu == CPU);
554 
555 	/*LINTED*/
556 	if (cpuid_opteron_erratum(cpu, 88) > 0) {
557 		/*
558 		 * SWAPGS May Fail To Read Correct GS Base
559 		 */
560 #if defined(OPTERON_ERRATUM_88)
561 		/*
562 		 * The workaround is an mfence in the relevant assembler code
563 		 */
564 #else
565 		WARNING(cpu, 88);
566 		missing++;
567 #endif
568 	}
569 
570 	if (cpuid_opteron_erratum(cpu, 91) > 0) {
571 		/*
572 		 * Software Prefetches May Report A Page Fault
573 		 */
574 #if defined(OPTERON_ERRATUM_91)
575 		/*
576 		 * fix is in trap.c
577 		 */
578 		opteron_erratum_91++;
579 #else
580 		WARNING(cpu, 91);
581 		missing++;
582 #endif
583 	}
584 
585 	if (cpuid_opteron_erratum(cpu, 93) > 0) {
586 		/*
587 		 * RSM Auto-Halt Restart Returns to Incorrect RIP
588 		 */
589 #if defined(OPTERON_ERRATUM_93)
590 		/*
591 		 * fix is in trap.c
592 		 */
593 		opteron_erratum_93++;
594 #else
595 		WARNING(cpu, 93);
596 		missing++;
597 #endif
598 	}
599 
600 	/*LINTED*/
601 	if (cpuid_opteron_erratum(cpu, 95) > 0) {
602 		/*
603 		 * RET Instruction May Return to Incorrect EIP
604 		 */
605 #if defined(OPTERON_ERRATUM_95)
606 #if defined(_LP64)
607 		/*
608 		 * Workaround this by ensuring that 32-bit user code and
609 		 * 64-bit kernel code never occupy the same address
610 		 * range mod 4G.
611 		 */
612 		if (_userlimit32 > 0xc0000000ul)
613 			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;
614 
615 		/*LINTED*/
616 		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
617 #endif	/* _LP64 */
618 #else
619 		WARNING(cpu, 95);
620 		missing++;
621 #endif	/* OPTERON_ERRATUM_95 */
622 	}
623 
624 	if (cpuid_opteron_erratum(cpu, 100) > 0) {
625 		/*
626 		 * Compatibility Mode Branches Transfer to Illegal Address
627 		 */
628 #if defined(OPTERON_ERRATUM_100)
629 		/*
630 		 * fix is in trap.c
631 		 */
632 		opteron_erratum_100++;
633 #else
634 		WARNING(cpu, 100);
635 		missing++;
636 #endif
637 	}
638 
639 	/*LINTED*/
640 	if (cpuid_opteron_erratum(cpu, 108) > 0) {
641 		/*
642 		 * CPUID Instruction May Return Incorrect Model Number In
643 		 * Some Processors
644 		 */
645 #if defined(OPTERON_ERRATUM_108)
646 		/*
647 		 * (Our cpuid-handling code corrects the model number on
648 		 * those processors)
649 		 */
650 #else
651 		WARNING(cpu, 108);
652 		missing++;
653 #endif
654 	}
655 
656 	/*LINTED*/
657 	if (cpuid_opteron_erratum(cpu, 109) > 0) {
658 		/*
659 		 * Certain Reverse REP MOVS May Produce Unpredictable Behaviour
660 		 */
661 #if defined(OPTERON_ERRATUM_109)
662 		uint64_t	patchlevel;
663 
664 		(void) rdmsr(MSR_AMD_PATCHLEVEL, &patchlevel);
665 		/* workaround is to print a warning to upgrade BIOS */
666 		if (patchlevel == 0)
667 			opteron_erratum_109++;
668 #else
669 		WARNING(cpu, 109);
670 		missing++;
671 #endif
672 	}
673 	/*LINTED*/
674 	if (cpuid_opteron_erratum(cpu, 121) > 0) {
675 		/*
676 		 * Sequential Execution Across Non_Canonical Boundary Caused
677 		 * Processor Hang
678 		 */
679 #if defined(OPTERON_ERRATUM_121)
680 		static int	lma;
681 
682 		if (opteron_erratum_121)
683 			opteron_erratum_121++;
684 
685 		/*
686 		 * Erratum 121 is only present in long (64 bit) mode.
687 		 * Workaround is to include the page immediately before the
688 		 * va hole to eliminate the possibility of system hangs due to
689 		 * sequential execution across the va hole boundary.
690 		 */
691 		if (lma == 0) {
692 			uint64_t	efer;
693 
694 			/*
695 			 * check LMA once: assume all cpus are in long mode
696 			 * or not.
697 			 */
698 			lma = 1;
699 
700 			(void) rdmsr(MSR_AMD_EFER, &efer);
701 			if (efer & AMD_EFER_LMA) {
702 				if (hole_start) {
703 					hole_start -= PAGESIZE;
704 				} else {
705 					/*
706 					 * hole_start not yet initialized by
707 					 * mmu_init. Initialize hole_start
708 					 * with value to be subtracted.
709 					 */
710 					hole_start = PAGESIZE;
711 				}
712 				opteron_erratum_121++;
713 			}
714 		}
715 #else
716 		WARNING(cpu, 121);
717 		missing++;
718 #endif
719 	}
720 
721 	/*LINTED*/
722 	if (cpuid_opteron_erratum(cpu, 122) > 0) {
723 		/*
724 		 * TLB Flush Filter May Cause Cohenrency Problem in
725 		 * Multiprocessor Systems
726 		 */
727 #if defined(OPTERON_ERRATUM_122)
728 		/*
729 		 * Erratum 122 is only present in MP configurations (multi-core
730 		 * or multi-processor).
731 		 */
732 
733 		if (opteron_erratum_122 || lgrp_plat_node_cnt > 1 ||
734 		    cpuid_get_ncpu_per_chip(cpu) > 1) {
735 			uint64_t	hwcrval;
736 
737 			/* disable TLB Flush Filter */
738 			(void) rdmsr(MSR_AMD_HWCR, &hwcrval);
739 			hwcrval |= AMD_HWCR_FFDIS;
740 			wrmsr(MSR_AMD_HWCR, &hwcrval);
741 			opteron_erratum_122++;
742 		}
743 
744 #else
745 		WARNING(cpu, 122);
746 		missing++;
747 #endif
748 	}
749 	/*LINTED*/
750 	if (cpuid_opteron_erratum(cpu, 123) > 0) {
751 		/*
752 		 * Bypassed Reads May Cause Data Corruption of System Hang in
753 		 * Dual Core Processors
754 		 */
755 #if defined(OPTERON_ERRATUM_123)
756 		/*
757 		 * Erratum 123 applies only to multi-core cpus.
758 		 */
759 
760 		if (cpuid_get_ncpu_per_chip(cpu) > 1) {
761 			uint64_t	patchlevel;
762 
763 			(void) rdmsr(MSR_AMD_PATCHLEVEL, &patchlevel);
764 			/* workaround is to print a warning to upgrade BIOS */
765 			if (patchlevel == 0)
766 				opteron_erratum_123++;
767 		}
768 #else
769 		WARNING(cpu, 123);
770 		missing++;
771 #endif
772 	}
773 	return (missing);
774 }
775 
776 void
777 workaround_errata_end()
778 {
779 #if defined(OPTERON_ERRATUM_109)
780 	if (opteron_erratum_109) {
781 		cmn_err(CE_WARN, "!BIOS microcode patch for AMD Processor"
782 		    " Erratum 109 was not detected. Updating BIOS with the"
783 		    " microcode patch is highly recommended.");
784 	}
785 #endif
786 #if defined(OPTERON_ERRATUM_123)
787 	if (opteron_erratum_123) {
788 		cmn_err(CE_WARN, "!BIOS microcode patch for AMD Processor"
789 		    " Erratum 123 was not detected. Updating BIOS with the"
790 		    " microcode patch is highly recommended.");
791 	}
792 #endif
793 }
794 
795 static ushort_t *mp_map_warm_reset_vector();
796 static void mp_unmap_warm_reset_vector(ushort_t *warm_reset_vector);
797 
798 /*ARGSUSED*/
799 void
800 start_other_cpus(int cprboot)
801 {
802 	unsigned who;
803 	int cpuid = getbootcpuid();
804 	int delays = 0;
805 	int started_cpu;
806 	ushort_t *warm_reset_vector = NULL;
807 	extern int procset;
808 
809 	/*
810 	 * Initialize our own cpu_info.
811 	 */
812 	init_cpu_info(CPU);
813 
814 	/*
815 	 * Initialize our syscall handlers
816 	 */
817 	init_cpu_syscall(CPU);
818 
819 	/*
820 	 * if only 1 cpu or not using MP, skip the rest of this
821 	 */
822 	if (!(mp_cpus & ~(1 << cpuid)) || use_mp == 0) {
823 		if (use_mp == 0)
824 			cmn_err(CE_CONT, "?***** Not in MP mode\n");
825 		goto done;
826 	}
827 
828 	/*
829 	 * perform such initialization as is needed
830 	 * to be able to take CPUs on- and off-line.
831 	 */
832 	cpu_pause_init();
833 
834 	xc_init();		/* initialize processor crosscalls */
835 
836 	/*
837 	 * Copy the real mode code at "real_mode_start" to the
838 	 * page at rm_platter_va.
839 	 */
840 	warm_reset_vector = mp_map_warm_reset_vector();
841 	if (warm_reset_vector == NULL)
842 		goto done;
843 
844 	bcopy((caddr_t)real_mode_start,
845 	    (caddr_t)((rm_platter_t *)rm_platter_va)->rm_code,
846 	    (size_t)real_mode_end - (size_t)real_mode_start);
847 
848 	flushes_require_xcalls = 1;
849 
850 	affinity_set(CPU_CURRENT);
851 
852 	for (who = 0; who < NCPU; who++) {
853 		if (who == cpuid)
854 			continue;
855 
856 		if ((mp_cpus & (1 << who)) == 0)
857 			continue;
858 
859 		mp_startup_init(who);
860 		started_cpu = 1;
861 		(*cpu_startf)(who, rm_platter_pa);
862 
863 		while ((procset & (1 << who)) == 0) {
864 
865 			delay(1);
866 			if (++delays > (20 * hz)) {
867 
868 				cmn_err(CE_WARN,
869 				    "cpu%d failed to start", who);
870 
871 				mutex_enter(&cpu_lock);
872 				cpu[who]->cpu_flags = 0;
873 				cpu_del_unit(who);
874 				mutex_exit(&cpu_lock);
875 
876 				started_cpu = 0;
877 				break;
878 			}
879 		}
880 		if (!started_cpu)
881 			continue;
882 		if (tsc_gethrtime_enable)
883 			tsc_sync_master(who);
884 
885 
886 		if (dtrace_cpu_init != NULL) {
887 			/*
888 			 * DTrace CPU initialization expects cpu_lock
889 			 * to be held.
890 			 */
891 			mutex_enter(&cpu_lock);
892 			(*dtrace_cpu_init)(who);
893 			mutex_exit(&cpu_lock);
894 		}
895 	}
896 
897 	affinity_clear();
898 
899 	for (who = 0; who < NCPU; who++) {
900 		if (who == cpuid)
901 			continue;
902 
903 		if (!(procset & (1 << who)))
904 			continue;
905 
906 		while (!(cpu_ready_set & (1 << who)))
907 			delay(1);
908 	}
909 
910 done:
911 	workaround_errata_end();
912 
913 	if (warm_reset_vector != NULL)
914 		mp_unmap_warm_reset_vector(warm_reset_vector);
915 	hat_unload(kas.a_hat, (caddr_t)(uintptr_t)rm_platter_pa, MMU_PAGESIZE,
916 	    HAT_UNLOAD);
917 }
918 
919 /*
920  * Dummy functions - no i86pc platforms support dynamic cpu allocation.
921  */
922 /*ARGSUSED*/
923 int
924 mp_cpu_configure(int cpuid)
925 {
926 	return (ENOTSUP);		/* not supported */
927 }
928 
929 /*ARGSUSED*/
930 int
931 mp_cpu_unconfigure(int cpuid)
932 {
933 	return (ENOTSUP);		/* not supported */
934 }
935 
936 /*
937  * Startup function for 'other' CPUs (besides boot cpu).
938  * Resumed from cpu_startup.
939  */
940 void
941 mp_startup(void)
942 {
943 	struct cpu *cp = CPU;
944 	extern int procset;
945 	uint_t new_x86_feature;
946 
947 	new_x86_feature = cpuid_pass1(cp);
948 
949 	/*
950 	 * We need to Sync MTRR with cpu0's MTRR. We have to do
951 	 * this with interrupts disabled.
952 	 */
953 	if (x86_feature & X86_MTRR)
954 		mtrr_sync();
955 	/*
956 	 * Enable machine check architecture
957 	 */
958 	if (x86_feature & X86_MCA)
959 		setup_mca();
960 
961 	/*
962 	 * Initialize this CPU's syscall handlers
963 	 */
964 	init_cpu_syscall(cp);
965 
966 	/*
967 	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
968 	 * highest level at which a routine is permitted to block on
969 	 * an adaptive mutex (allows for cpu poke interrupt in case
970 	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
971 	 * device interrupts that may end up in the hat layer issuing cross
972 	 * calls before CPU_READY is set.
973 	 */
974 	(void) splx(ipltospl(LOCK_LEVEL));
975 
976 	/*
977 	 * Do a sanity check to make sure this new CPU is a sane thing
978 	 * to add to the collection of processors running this system.
979 	 *
980 	 * XXX	Clearly this needs to get more sophisticated, if x86
981 	 * systems start to get built out of heterogenous CPUs; as is
982 	 * likely to happen once the number of processors in a configuration
983 	 * gets large enough.
984 	 */
985 	if ((x86_feature & new_x86_feature) != x86_feature) {
986 		cmn_err(CE_CONT, "?cpu%d: %b\n",
987 		    cp->cpu_id, new_x86_feature, FMT_X86_FEATURE);
988 		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
989 	}
990 
991 	/*
992 	 * We could be more sophisticated here, and just mark the CPU
993 	 * as "faulted" but at this point we'll opt for the easier
994 	 * answer of dieing horribly.  Provided the boot cpu is ok,
995 	 * the system can be recovered by booting with use_mp set to zero.
996 	 */
997 	if (workaround_errata(cp) != 0)
998 		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
999 
1000 	cpuid_pass2(cp);
1001 	cpuid_pass3(cp);
1002 	(void) cpuid_pass4(cp);
1003 
1004 	init_cpu_info(cp);
1005 
1006 	add_cpunode2devtree(cp->cpu_id, cp->cpu_m.mcpu_cpi);
1007 
1008 	mutex_enter(&cpu_lock);
1009 	procset |= 1 << cp->cpu_id;
1010 	mutex_exit(&cpu_lock);
1011 
1012 	if (tsc_gethrtime_enable)
1013 		tsc_sync_slave();
1014 
1015 	mutex_enter(&cpu_lock);
1016 	/*
1017 	 * It's unfortunate that chip_cpu_init() has to be called here.
1018 	 * It really belongs in cpu_add_unit(), but unfortunately it is
1019 	 * dependent on the cpuid probing, which must be done in the
1020 	 * context of the current CPU. Care must be taken on x86 to ensure
1021 	 * that mp_startup can safely block even though chip_cpu_init() and
1022 	 * cpu_add_active() have not yet been called.
1023 	 */
1024 	chip_cpu_init(cp);
1025 	chip_cpu_startup(cp);
1026 
1027 	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_ENABLE | CPU_EXISTS;
1028 	cpu_add_active(cp);
1029 	mutex_exit(&cpu_lock);
1030 
1031 	(void) spl0();				/* enable interrupts */
1032 
1033 	if (boothowto & RB_DEBUG)
1034 		kdi_dvec_cpu_init(cp);
1035 
1036 	/*
1037 	 * Setting the bit in cpu_ready_set must be the last operation in
1038 	 * processor initialization; the boot CPU will continue to boot once
1039 	 * it sees this bit set for all active CPUs.
1040 	 */
1041 	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1042 
1043 	/*
1044 	 * Because mp_startup() gets fired off after init() starts, we
1045 	 * can't use the '?' trick to do 'boot -v' printing - so we
1046 	 * always direct the 'cpu .. online' messages to the log.
1047 	 */
1048 	cmn_err(CE_CONT, "!cpu%d initialization complete - online\n",
1049 	    cp->cpu_id);
1050 
1051 	/*
1052 	 * Now we are done with the startup thread, so free it up.
1053 	 */
1054 	thread_exit();
1055 	panic("mp_startup: cannot return");
1056 	/*NOTREACHED*/
1057 }
1058 
1059 
1060 /*
1061  * Start CPU on user request.
1062  */
1063 /* ARGSUSED */
1064 int
1065 mp_cpu_start(struct cpu *cp)
1066 {
1067 	ASSERT(MUTEX_HELD(&cpu_lock));
1068 	if (cp->cpu_id == getbootcpuid())
1069 		return (EBUSY); 	/* Cannot start boot CPU */
1070 	return (0);
1071 }
1072 
1073 /*
1074  * Stop CPU on user request.
1075  */
1076 /* ARGSUSED */
1077 int
1078 mp_cpu_stop(struct cpu *cp)
1079 {
1080 	ASSERT(MUTEX_HELD(&cpu_lock));
1081 	if (cp->cpu_id == getbootcpuid())
1082 		return (EBUSY); 	/* Cannot stop boot CPU */
1083 
1084 	return (0);
1085 }
1086 
1087 /*
1088  * Power on CPU.
1089  */
1090 /* ARGSUSED */
1091 int
1092 mp_cpu_poweron(struct cpu *cp)
1093 {
1094 	ASSERT(MUTEX_HELD(&cpu_lock));
1095 	return (ENOTSUP);		/* not supported */
1096 }
1097 
1098 /*
1099  * Power off CPU.
1100  */
1101 /* ARGSUSED */
1102 int
1103 mp_cpu_poweroff(struct cpu *cp)
1104 {
1105 	ASSERT(MUTEX_HELD(&cpu_lock));
1106 	return (ENOTSUP);		/* not supported */
1107 }
1108 
1109 
1110 /*
1111  * Take the specified CPU out of participation in interrupts.
1112  */
1113 int
1114 cpu_disable_intr(struct cpu *cp)
1115 {
1116 	/*
1117 	 * cannot disable interrupts on boot cpu
1118 	 */
1119 	if (cp == cpu[getbootcpuid()])
1120 		return (EBUSY);
1121 
1122 	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1123 		return (EBUSY);
1124 
1125 	cp->cpu_flags &= ~CPU_ENABLE;
1126 	return (0);
1127 }
1128 
1129 /*
1130  * Allow the specified CPU to participate in interrupts.
1131  */
1132 void
1133 cpu_enable_intr(struct cpu *cp)
1134 {
1135 	ASSERT(MUTEX_HELD(&cpu_lock));
1136 	if (cp == cpu[getbootcpuid()])
1137 		return;
1138 
1139 	cp->cpu_flags |= CPU_ENABLE;
1140 	psm_enable_intr(cp->cpu_id);
1141 }
1142 
1143 
1144 /*
1145  * return the cpu id of the initial startup cpu
1146  */
1147 processorid_t
1148 getbootcpuid(void)
1149 {
1150 	return (0);
1151 }
1152 
1153 static ushort_t *
1154 mp_map_warm_reset_vector()
1155 {
1156 	ushort_t *warm_reset_vector;
1157 
1158 	if (!(warm_reset_vector = (ushort_t *)psm_map_phys(WARM_RESET_VECTOR,
1159 	    sizeof (ushort_t *), PROT_READ|PROT_WRITE)))
1160 		return (NULL);
1161 
1162 	/*
1163 	 * setup secondary cpu bios boot up vector
1164 	 */
1165 	*warm_reset_vector = (ushort_t)((caddr_t)
1166 		((struct rm_platter *)rm_platter_va)->rm_code - rm_platter_va
1167 		+ ((ulong_t)rm_platter_va & 0xf));
1168 	warm_reset_vector++;
1169 	*warm_reset_vector = (ushort_t)(rm_platter_pa >> 4);
1170 
1171 	--warm_reset_vector;
1172 	return (warm_reset_vector);
1173 }
1174 
1175 static void
1176 mp_unmap_warm_reset_vector(ushort_t *warm_reset_vector)
1177 {
1178 	psm_unmap_phys((caddr_t)warm_reset_vector, sizeof (ushort_t *));
1179 }
1180 
1181 /*ARGSUSED*/
1182 void
1183 mp_cpu_faulted_enter(struct cpu *cp)
1184 {}
1185 
1186 /*ARGSUSED*/
1187 void
1188 mp_cpu_faulted_exit(struct cpu *cp)
1189 {}
1190 
1191 /*
1192  * The following two routines are used as context operators on threads belonging
1193  * to processes with a private LDT (see sysi86).  Due to the rarity of such
1194  * processes, these routines are currently written for best code readability and
1195  * organization rather than speed.  We could avoid checking x86_feature at every
1196  * context switch by installing different context ops, depending on the
1197  * x86_feature flags, at LDT creation time -- one for each combination of fast
1198  * syscall feature flags.
1199  */
1200 
1201 /*ARGSUSED*/
1202 void
1203 cpu_fast_syscall_disable(void *arg)
1204 {
1205 	if (x86_feature & X86_SEP)
1206 		cpu_sep_disable();
1207 	if (x86_feature & X86_ASYSC)
1208 		cpu_asysc_disable();
1209 }
1210 
1211 /*ARGSUSED*/
1212 void
1213 cpu_fast_syscall_enable(void *arg)
1214 {
1215 	if (x86_feature & X86_SEP)
1216 		cpu_sep_enable();
1217 	if (x86_feature & X86_ASYSC)
1218 		cpu_asysc_enable();
1219 }
1220 
1221 static void
1222 cpu_sep_enable(void)
1223 {
1224 	uint64_t value;
1225 
1226 	ASSERT(x86_feature & X86_SEP);
1227 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1228 
1229 	value = KCS_SEL;
1230 	wrmsr(MSR_INTC_SEP_CS, &value);
1231 }
1232 
1233 static void
1234 cpu_sep_disable(void)
1235 {
1236 	uint64_t value;
1237 
1238 	ASSERT(x86_feature & X86_SEP);
1239 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1240 
1241 	/*
1242 	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
1243 	 * the sysenter or sysexit instruction to trigger a #gp fault.
1244 	 */
1245 	value = 0;
1246 	wrmsr(MSR_INTC_SEP_CS, &value);
1247 }
1248 
1249 static void
1250 cpu_asysc_enable(void)
1251 {
1252 	uint64_t value;
1253 
1254 	ASSERT(x86_feature & X86_ASYSC);
1255 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1256 
1257 	(void) rdmsr(MSR_AMD_EFER, &value);
1258 	value |= AMD_EFER_SCE;
1259 	wrmsr(MSR_AMD_EFER, &value);
1260 }
1261 
1262 static void
1263 cpu_asysc_disable(void)
1264 {
1265 	uint64_t value;
1266 
1267 	ASSERT(x86_feature & X86_ASYSC);
1268 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1269 
1270 	/*
1271 	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
1272 	 * executing syscall or sysret with this bit off will incur a #ud trap.
1273 	 */
1274 	(void) rdmsr(MSR_AMD_EFER, &value);
1275 	value &= ~AMD_EFER_SCE;
1276 	wrmsr(MSR_AMD_EFER, &value);
1277 }
1278