xref: /titanic_41/usr/src/uts/i86pc/os/mp_startup.c (revision d87bb7db56bb1b04f743195ba206a58728f6ffaf)
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 	rm_platter_t *real_mode_platter = (rm_platter_t *)rm_platter_va;
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 	if ((cp = kmem_zalloc(sizeof (*cp), KM_NOSLEEP)) == NULL) {
253 		panic("mp_startup_init: cpu%d: "
254 		    "no memory for cpu structure", cpun);
255 		/*NOTREACHED*/
256 	}
257 	procp = curthread->t_procp;
258 
259 	mutex_enter(&cpu_lock);
260 	/*
261 	 * Initialize the dispatcher first.
262 	 */
263 	disp_cpu_init(cp);
264 	mutex_exit(&cpu_lock);
265 
266 	/*
267 	 * Allocate and initialize the startup thread for this CPU.
268 	 * Interrupt and process switch stacks get allocated later
269 	 * when the CPU starts running.
270 	 */
271 	tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
272 	    TS_STOPPED, maxclsyspri);
273 
274 	/*
275 	 * Set state to TS_ONPROC since this thread will start running
276 	 * as soon as the CPU comes online.
277 	 *
278 	 * All the other fields of the thread structure are setup by
279 	 * thread_create().
280 	 */
281 	THREAD_ONPROC(tp, cp);
282 	tp->t_preempt = 1;
283 	tp->t_bound_cpu = cp;
284 	tp->t_affinitycnt = 1;
285 	tp->t_cpu = cp;
286 	tp->t_disp_queue = cp->cpu_disp;
287 
288 	/*
289 	 * Setup thread to start in mp_startup.
290 	 */
291 	sp = tp->t_stk;
292 	tp->t_pc = (uintptr_t)mp_startup;
293 	tp->t_sp = (uintptr_t)(sp - MINFRAME);
294 
295 	cp->cpu_id = cpun;
296 	cp->cpu_self = cp;
297 	cp->cpu_mask = 1 << cpun;
298 	cp->cpu_thread = tp;
299 	cp->cpu_lwp = NULL;
300 	cp->cpu_dispthread = tp;
301 	cp->cpu_dispatch_pri = DISP_PRIO(tp);
302 
303 	/*
304 	 * Now, initialize per-CPU idle thread for this CPU.
305 	 */
306 	tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);
307 
308 	cp->cpu_idle_thread = tp;
309 
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 	 * Bootstrap the CPU for CMT aware scheduling
318 	 * The rest of the initialization will happen from
319 	 * mp_startup()
320 	 */
321 	chip_bootstrap_cpu(cp);
322 
323 	/*
324 	 * Perform CPC intialization on the new CPU.
325 	 */
326 	kcpc_hw_init(cp);
327 
328 	/*
329 	 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
330 	 * for each CPU.
331 	 */
332 
333 	setup_vaddr_for_ppcopy(cp);
334 
335 	/*
336 	 * Allocate space for page directory, stack, tss, gdt and idt.
337 	 * This assumes that kmem_alloc will return memory which is aligned
338 	 * to the next higher power of 2 or a page(if size > MAXABIG)
339 	 * If this assumption goes wrong at any time due to change in
340 	 * kmem alloc, things may not work as the page directory has to be
341 	 * page aligned
342 	 */
343 	if ((tablesp = kmem_zalloc(sizeof (*tablesp), KM_NOSLEEP)) == NULL)
344 		panic("mp_startup_init: cpu%d cannot allocate tables", cpun);
345 
346 	if ((uintptr_t)tablesp & ~MMU_STD_PAGEMASK) {
347 		kmem_free(tablesp, sizeof (struct cpu_tables));
348 		size = sizeof (struct cpu_tables) + MMU_STD_PAGESIZE;
349 		tablesp = kmem_zalloc(size, KM_NOSLEEP);
350 		tablesp = (struct cpu_tables *)
351 		    (((uintptr_t)tablesp + MMU_STD_PAGESIZE) &
352 		    MMU_STD_PAGEMASK);
353 	}
354 
355 	ntss = cp->cpu_tss = &tablesp->ct_tss;
356 	cp->cpu_gdt = tablesp->ct_gdt;
357 	bcopy(CPU->cpu_gdt, cp->cpu_gdt, NGDT * (sizeof (user_desc_t)));
358 
359 #if defined(__amd64)
360 
361 	/*
362 	 * #DF (double fault).
363 	 */
364 	ntss->tss_ist1 =
365 	    (uint64_t)&tablesp->ct_stack[sizeof (tablesp->ct_stack)];
366 
367 #elif defined(__i386)
368 
369 	ntss->tss_esp0 = ntss->tss_esp1 = ntss->tss_esp2 = ntss->tss_esp =
370 	    (uint32_t)&tablesp->ct_stack[sizeof (tablesp->ct_stack)];
371 
372 	ntss->tss_ss0 = ntss->tss_ss1 = ntss->tss_ss2 = ntss->tss_ss = KDS_SEL;
373 
374 	ntss->tss_eip = (uint32_t)mp_startup;
375 
376 	ntss->tss_cs = KCS_SEL;
377 	ntss->tss_fs = KFS_SEL;
378 	ntss->tss_gs = KGS_SEL;
379 
380 	/*
381 	 * setup kernel %gs.
382 	 */
383 	set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA,
384 	    SEL_KPL, 0, 1);
385 
386 #endif	/* __i386 */
387 
388 	/*
389 	 * Set I/O bit map offset equal to size of TSS segment limit
390 	 * for no I/O permission map. This will cause all user I/O
391 	 * instructions to generate #gp fault.
392 	 */
393 	ntss->tss_bitmapbase = sizeof (*ntss);
394 
395 	/*
396 	 * setup kernel tss.
397 	 */
398 	set_syssegd((system_desc_t *)&cp->cpu_gdt[GDT_KTSS], cp->cpu_tss,
399 	    sizeof (*cp->cpu_tss) -1, SDT_SYSTSS, SEL_KPL);
400 
401 	/*
402 	 * If we have more than one node, each cpu gets a copy of IDT
403 	 * local to its node. If this is a Pentium box, we use cpu 0's
404 	 * IDT. cpu 0's IDT has been made read-only to workaround the
405 	 * cmpxchgl register bug
406 	 */
407 	cp->cpu_idt = CPU->cpu_idt;
408 	if (system_hardware.hd_nodes && x86_type != X86_TYPE_P5) {
409 		cp->cpu_idt = kmem_alloc(sizeof (idt0), KM_SLEEP);
410 		bcopy(idt0, cp->cpu_idt, sizeof (idt0));
411 	}
412 
413 	/*
414 	 * Get interrupt priority data from cpu 0
415 	 */
416 	cp->cpu_pri_data = CPU->cpu_pri_data;
417 
418 	hat_cpu_online(cp);
419 
420 	/* Should remove all entries for the current process/thread here */
421 
422 	/*
423 	 * Fill up the real mode platter to make it easy for real mode code to
424 	 * kick it off. This area should really be one passed by boot to kernel
425 	 * and guaranteed to be below 1MB and aligned to 16 bytes. Should also
426 	 * have identical physical and virtual address in paged mode.
427 	 */
428 	real_mode_platter->rm_idt_base = cp->cpu_idt;
429 	real_mode_platter->rm_idt_lim = sizeof (idt0) - 1;
430 	real_mode_platter->rm_gdt_base = cp->cpu_gdt;
431 	real_mode_platter->rm_gdt_lim = sizeof (gdt0) -1;
432 	real_mode_platter->rm_pdbr = getcr3();
433 	real_mode_platter->rm_cpu = cpun;
434 	real_mode_platter->rm_x86feature = x86_feature;
435 	real_mode_platter->rm_cr4 = cr4_value;
436 
437 #if defined(__amd64)
438 	if (getcr3() > 0xffffffffUL)
439 		panic("Cannot initialize CPUs; kernel's 64-bit page tables\n"
440 			"located above 4G in physical memory (@ 0x%llx).",
441 			(unsigned long long)getcr3());
442 
443 	/*
444 	 * Setup pseudo-descriptors for temporary GDT and IDT for use ONLY
445 	 * by code in real_mode_start():
446 	 *
447 	 * GDT[0]:  NULL selector
448 	 * GDT[1]:  64-bit CS: Long = 1, Present = 1, bits 12, 11 = 1
449 	 *
450 	 * Clear the IDT as interrupts will be off and a limit of 0 will cause
451 	 * the CPU to triple fault and reset on an NMI, seemingly as reasonable
452 	 * a course of action as any other, though it may cause the entire
453 	 * platform to reset in some cases...
454 	 */
455 	real_mode_platter->rm_temp_gdt[0] = 0ULL;
456 	real_mode_platter->rm_temp_gdt[TEMPGDT_KCODE64] = 0x20980000000000ULL;
457 
458 	real_mode_platter->rm_temp_gdt_lim = (ushort_t)
459 	    (sizeof (real_mode_platter->rm_temp_gdt) - 1);
460 	real_mode_platter->rm_temp_gdt_base = rm_platter_pa +
461 	    (uint32_t)(&((rm_platter_t *)0)->rm_temp_gdt);
462 
463 	real_mode_platter->rm_temp_idt_lim = 0;
464 	real_mode_platter->rm_temp_idt_base = 0;
465 
466 	/*
467 	 * Since the CPU needs to jump to protected mode using an identity
468 	 * mapped address, we need to calculate it here.
469 	 */
470 	real_mode_platter->rm_longmode64_addr = rm_platter_pa +
471 	    ((uint32_t)long_mode_64 - (uint32_t)real_mode_start);
472 #endif	/* __amd64 */
473 
474 #ifdef TRAPTRACE
475 	/*
476 	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers for this
477 	 * CPU.
478 	 */
479 	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
480 	ttc->ttc_next = ttc->ttc_first;
481 	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
482 #endif
483 
484 	/*
485 	 * Record that we have another CPU.
486 	 */
487 	mutex_enter(&cpu_lock);
488 	/*
489 	 * Initialize the interrupt threads for this CPU
490 	 */
491 	init_intr_threads(cp);
492 	/*
493 	 * Add CPU to list of available CPUs.  It'll be on the active list
494 	 * after mp_startup().
495 	 */
496 	cpu_add_unit(cp);
497 	mutex_exit(&cpu_lock);
498 }
499 
500 /*
501  * Apply workarounds for known errata, and warn about those that are absent.
502  *
503  * System vendors occasionally create configurations which contain different
504  * revisions of the CPUs that are almost but not exactly the same.  At the
505  * time of writing, this meant that their clock rates were the same, their
506  * feature sets were the same, but the required workaround were -not-
507  * necessarily the same.  So, this routine is invoked on -every- CPU soon
508  * after starting to make sure that the resulting system contains the most
509  * pessimal set of workarounds needed to cope with *any* of the CPUs in the
510  * system.
511  *
512  * These workarounds are based on Rev 3.50 of the Revision Guide for
513  * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, May 2005.
514  */
515 
516 #if defined(OPTERON_ERRATUM_91)
517 int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
518 #endif
519 
520 #if defined(OPTERON_ERRATUM_93)
521 int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
522 #endif
523 
524 #if defined(OPTERON_ERRATUM_100)
525 int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
526 #endif
527 
528 #if defined(OPTERON_ERRATUM_109)
529 int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
530 #endif
531 
532 #if defined(OPTERON_ERRATUM_121)
533 int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
534 #endif
535 
536 #if defined(OPTERON_ERRATUM_122)
537 int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
538 #endif
539 
540 #if defined(OPTERON_ERRATUM_123)
541 int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
542 #endif
543 
544 
545 #define	WARNING(cpu, n)						\
546 	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %d",	\
547 	    (cpu)->cpu_id, (n))
548 
549 uint_t
550 workaround_errata(struct cpu *cpu)
551 {
552 	uint_t missing = 0;
553 
554 	ASSERT(cpu == CPU);
555 
556 	/*LINTED*/
557 	if (cpuid_opteron_erratum(cpu, 88) > 0) {
558 		/*
559 		 * SWAPGS May Fail To Read Correct GS Base
560 		 */
561 #if defined(OPTERON_ERRATUM_88)
562 		/*
563 		 * The workaround is an mfence in the relevant assembler code
564 		 */
565 #else
566 		WARNING(cpu, 88);
567 		missing++;
568 #endif
569 	}
570 
571 	if (cpuid_opteron_erratum(cpu, 91) > 0) {
572 		/*
573 		 * Software Prefetches May Report A Page Fault
574 		 */
575 #if defined(OPTERON_ERRATUM_91)
576 		/*
577 		 * fix is in trap.c
578 		 */
579 		opteron_erratum_91++;
580 #else
581 		WARNING(cpu, 91);
582 		missing++;
583 #endif
584 	}
585 
586 	if (cpuid_opteron_erratum(cpu, 93) > 0) {
587 		/*
588 		 * RSM Auto-Halt Restart Returns to Incorrect RIP
589 		 */
590 #if defined(OPTERON_ERRATUM_93)
591 		/*
592 		 * fix is in trap.c
593 		 */
594 		opteron_erratum_93++;
595 #else
596 		WARNING(cpu, 93);
597 		missing++;
598 #endif
599 	}
600 
601 	/*LINTED*/
602 	if (cpuid_opteron_erratum(cpu, 95) > 0) {
603 		/*
604 		 * RET Instruction May Return to Incorrect EIP
605 		 */
606 #if defined(OPTERON_ERRATUM_95)
607 #if defined(_LP64)
608 		/*
609 		 * Workaround this by ensuring that 32-bit user code and
610 		 * 64-bit kernel code never occupy the same address
611 		 * range mod 4G.
612 		 */
613 		if (_userlimit32 > 0xc0000000ul)
614 			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;
615 
616 		/*LINTED*/
617 		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
618 #endif	/* _LP64 */
619 #else
620 		WARNING(cpu, 95);
621 		missing++;
622 #endif	/* OPTERON_ERRATUM_95 */
623 	}
624 
625 	if (cpuid_opteron_erratum(cpu, 100) > 0) {
626 		/*
627 		 * Compatibility Mode Branches Transfer to Illegal Address
628 		 */
629 #if defined(OPTERON_ERRATUM_100)
630 		/*
631 		 * fix is in trap.c
632 		 */
633 		opteron_erratum_100++;
634 #else
635 		WARNING(cpu, 100);
636 		missing++;
637 #endif
638 	}
639 
640 	/*LINTED*/
641 	if (cpuid_opteron_erratum(cpu, 108) > 0) {
642 		/*
643 		 * CPUID Instruction May Return Incorrect Model Number In
644 		 * Some Processors
645 		 */
646 #if defined(OPTERON_ERRATUM_108)
647 		/*
648 		 * (Our cpuid-handling code corrects the model number on
649 		 * those processors)
650 		 */
651 #else
652 		WARNING(cpu, 108);
653 		missing++;
654 #endif
655 	}
656 
657 	/*LINTED*/
658 	if (cpuid_opteron_erratum(cpu, 109) > 0) {
659 		/*
660 		 * Certain Reverse REP MOVS May Produce Unpredictable Behaviour
661 		 */
662 #if defined(OPTERON_ERRATUM_109)
663 		uint64_t	patchlevel;
664 
665 		(void) rdmsr(MSR_AMD_PATCHLEVEL, &patchlevel);
666 		/* workaround is to print a warning to upgrade BIOS */
667 		if (patchlevel == 0)
668 			opteron_erratum_109++;
669 #else
670 		WARNING(cpu, 109);
671 		missing++;
672 #endif
673 	}
674 	/*LINTED*/
675 	if (cpuid_opteron_erratum(cpu, 121) > 0) {
676 		/*
677 		 * Sequential Execution Across Non_Canonical Boundary Caused
678 		 * Processor Hang
679 		 */
680 #if defined(OPTERON_ERRATUM_121)
681 		static int	lma;
682 
683 		if (opteron_erratum_121)
684 			opteron_erratum_121++;
685 
686 		/*
687 		 * Erratum 121 is only present in long (64 bit) mode.
688 		 * Workaround is to include the page immediately before the
689 		 * va hole to eliminate the possibility of system hangs due to
690 		 * sequential execution across the va hole boundary.
691 		 */
692 		if (lma == 0) {
693 			uint64_t	efer;
694 
695 			/*
696 			 * check LMA once: assume all cpus are in long mode
697 			 * or not.
698 			 */
699 			lma = 1;
700 
701 			(void) rdmsr(MSR_AMD_EFER, &efer);
702 			if (efer & AMD_EFER_LMA) {
703 				if (hole_start) {
704 					hole_start -= PAGESIZE;
705 				} else {
706 					/*
707 					 * hole_start not yet initialized by
708 					 * mmu_init. Initialize hole_start
709 					 * with value to be subtracted.
710 					 */
711 					hole_start = PAGESIZE;
712 				}
713 				opteron_erratum_121++;
714 			}
715 		}
716 #else
717 		WARNING(cpu, 121);
718 		missing++;
719 #endif
720 	}
721 
722 	/*LINTED*/
723 	if (cpuid_opteron_erratum(cpu, 122) > 0) {
724 		/*
725 		 * TLB Flush Filter May Cause Cohenrency Problem in
726 		 * Multiprocessor Systems
727 		 */
728 #if defined(OPTERON_ERRATUM_122)
729 		/*
730 		 * Erratum 122 is only present in MP configurations (multi-core
731 		 * or multi-processor).
732 		 */
733 
734 		if (opteron_erratum_122 || lgrp_plat_node_cnt > 1 ||
735 		    cpuid_get_ncpu_per_chip(cpu) > 1) {
736 			uint64_t	hwcrval;
737 
738 			/* disable TLB Flush Filter */
739 			(void) rdmsr(MSR_AMD_HWCR, &hwcrval);
740 			hwcrval |= AMD_HWCR_FFDIS;
741 			wrmsr(MSR_AMD_HWCR, &hwcrval);
742 			opteron_erratum_122++;
743 		}
744 
745 #else
746 		WARNING(cpu, 122);
747 		missing++;
748 #endif
749 	}
750 	/*LINTED*/
751 	if (cpuid_opteron_erratum(cpu, 123) > 0) {
752 		/*
753 		 * Bypassed Reads May Cause Data Corruption of System Hang in
754 		 * Dual Core Processors
755 		 */
756 #if defined(OPTERON_ERRATUM_123)
757 		/*
758 		 * Erratum 123 applies only to multi-core cpus.
759 		 */
760 
761 		if (cpuid_get_ncpu_per_chip(cpu) > 1) {
762 			uint64_t	patchlevel;
763 
764 			(void) rdmsr(MSR_AMD_PATCHLEVEL, &patchlevel);
765 			/* workaround is to print a warning to upgrade BIOS */
766 			if (patchlevel == 0)
767 				opteron_erratum_123++;
768 		}
769 #else
770 		WARNING(cpu, 123);
771 		missing++;
772 #endif
773 	}
774 	return (missing);
775 }
776 
777 void
778 workaround_errata_end()
779 {
780 #if defined(OPTERON_ERRATUM_109)
781 	if (opteron_erratum_109) {
782 		cmn_err(CE_WARN, "!BIOS microcode patch for AMD Processor"
783 		    " Erratum 109 was not detected. Updating BIOS with the"
784 		    " microcode patch is highly recommended.");
785 	}
786 #endif
787 #if defined(OPTERON_ERRATUM_123)
788 	if (opteron_erratum_123) {
789 		cmn_err(CE_WARN, "!BIOS microcode patch for AMD Processor"
790 		    " Erratum 123 was not detected. Updating BIOS with the"
791 		    " microcode patch is highly recommended.");
792 	}
793 #endif
794 }
795 
796 static ushort_t *mp_map_warm_reset_vector();
797 static void mp_unmap_warm_reset_vector(ushort_t *warm_reset_vector);
798 
799 /*ARGSUSED*/
800 void
801 start_other_cpus(int cprboot)
802 {
803 	unsigned who;
804 	int cpuid = getbootcpuid();
805 	int delays = 0;
806 	int started_cpu;
807 	ushort_t *warm_reset_vector = NULL;
808 	extern int procset;
809 
810 	/*
811 	 * Initialize our own cpu_info.
812 	 */
813 	init_cpu_info(CPU);
814 
815 	/*
816 	 * Initialize our syscall handlers
817 	 */
818 	init_cpu_syscall(CPU);
819 
820 	/*
821 	 * if only 1 cpu or not using MP, skip the rest of this
822 	 */
823 	if (!(mp_cpus & ~(1 << cpuid)) || use_mp == 0) {
824 		if (use_mp == 0)
825 			cmn_err(CE_CONT, "?***** Not in MP mode\n");
826 		goto done;
827 	}
828 
829 	/*
830 	 * perform such initialization as is needed
831 	 * to be able to take CPUs on- and off-line.
832 	 */
833 	cpu_pause_init();
834 
835 	xc_init();		/* initialize processor crosscalls */
836 
837 	/*
838 	 * Copy the real mode code at "real_mode_start" to the
839 	 * page at rm_platter_va.
840 	 */
841 	warm_reset_vector = mp_map_warm_reset_vector();
842 	if (warm_reset_vector == NULL)
843 		goto done;
844 
845 	bcopy((caddr_t)real_mode_start,
846 	    (caddr_t)((rm_platter_t *)rm_platter_va)->rm_code,
847 	    (size_t)real_mode_end - (size_t)real_mode_start);
848 
849 	flushes_require_xcalls = 1;
850 
851 	affinity_set(CPU_CURRENT);
852 
853 	for (who = 0; who < NCPU; who++) {
854 		if (who == cpuid)
855 			continue;
856 
857 		if ((mp_cpus & (1 << who)) == 0)
858 			continue;
859 
860 		mp_startup_init(who);
861 		started_cpu = 1;
862 		(*cpu_startf)(who, rm_platter_pa);
863 
864 		while ((procset & (1 << who)) == 0) {
865 
866 			delay(1);
867 			if (++delays > (20 * hz)) {
868 
869 				cmn_err(CE_WARN,
870 				    "cpu%d failed to start", who);
871 
872 				mutex_enter(&cpu_lock);
873 				cpu[who]->cpu_flags = 0;
874 				cpu_del_unit(who);
875 				mutex_exit(&cpu_lock);
876 
877 				started_cpu = 0;
878 				break;
879 			}
880 		}
881 		if (!started_cpu)
882 			continue;
883 		if (tsc_gethrtime_enable)
884 			tsc_sync_master(who);
885 
886 
887 		if (dtrace_cpu_init != NULL) {
888 			/*
889 			 * DTrace CPU initialization expects cpu_lock
890 			 * to be held.
891 			 */
892 			mutex_enter(&cpu_lock);
893 			(*dtrace_cpu_init)(who);
894 			mutex_exit(&cpu_lock);
895 		}
896 	}
897 
898 	affinity_clear();
899 
900 	for (who = 0; who < NCPU; who++) {
901 		if (who == cpuid)
902 			continue;
903 
904 		if (!(procset & (1 << who)))
905 			continue;
906 
907 		while (!(cpu_ready_set & (1 << who)))
908 			delay(1);
909 	}
910 
911 done:
912 	workaround_errata_end();
913 
914 	if (warm_reset_vector != NULL)
915 		mp_unmap_warm_reset_vector(warm_reset_vector);
916 	hat_unload(kas.a_hat, (caddr_t)(uintptr_t)rm_platter_pa, MMU_PAGESIZE,
917 	    HAT_UNLOAD);
918 }
919 
920 /*
921  * Dummy functions - no i86pc platforms support dynamic cpu allocation.
922  */
923 /*ARGSUSED*/
924 int
925 mp_cpu_configure(int cpuid)
926 {
927 	return (ENOTSUP);		/* not supported */
928 }
929 
930 /*ARGSUSED*/
931 int
932 mp_cpu_unconfigure(int cpuid)
933 {
934 	return (ENOTSUP);		/* not supported */
935 }
936 
937 /*
938  * Startup function for 'other' CPUs (besides boot cpu).
939  * Resumed from cpu_startup.
940  */
941 void
942 mp_startup(void)
943 {
944 	struct cpu *cp = CPU;
945 	extern int procset;
946 	uint_t new_x86_feature;
947 
948 	new_x86_feature = cpuid_pass1(cp);
949 
950 	/*
951 	 * We need to Sync MTRR with cpu0's MTRR. We have to do
952 	 * this with interrupts disabled.
953 	 */
954 	if (x86_feature & X86_MTRR)
955 		mtrr_sync();
956 	/*
957 	 * Enable machine check architecture
958 	 */
959 	if (x86_feature & X86_MCA)
960 		setup_mca();
961 
962 	/*
963 	 * Initialize this CPU's syscall handlers
964 	 */
965 	init_cpu_syscall(cp);
966 
967 	/*
968 	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
969 	 * highest level at which a routine is permitted to block on
970 	 * an adaptive mutex (allows for cpu poke interrupt in case
971 	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
972 	 * device interrupts that may end up in the hat layer issuing cross
973 	 * calls before CPU_READY is set.
974 	 */
975 	(void) splx(ipltospl(LOCK_LEVEL));
976 
977 	/*
978 	 * Do a sanity check to make sure this new CPU is a sane thing
979 	 * to add to the collection of processors running this system.
980 	 *
981 	 * XXX	Clearly this needs to get more sophisticated, if x86
982 	 * systems start to get built out of heterogenous CPUs; as is
983 	 * likely to happen once the number of processors in a configuration
984 	 * gets large enough.
985 	 */
986 	if ((x86_feature & new_x86_feature) != x86_feature) {
987 		cmn_err(CE_CONT, "?cpu%d: %b\n",
988 		    cp->cpu_id, new_x86_feature, FMT_X86_FEATURE);
989 		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
990 	}
991 
992 	/*
993 	 * We could be more sophisticated here, and just mark the CPU
994 	 * as "faulted" but at this point we'll opt for the easier
995 	 * answer of dieing horribly.  Provided the boot cpu is ok,
996 	 * the system can be recovered by booting with use_mp set to zero.
997 	 */
998 	if (workaround_errata(cp) != 0)
999 		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
1000 
1001 	cpuid_pass2(cp);
1002 	cpuid_pass3(cp);
1003 	(void) cpuid_pass4(cp);
1004 
1005 	init_cpu_info(cp);
1006 
1007 	add_cpunode2devtree(cp->cpu_id, cp->cpu_m.mcpu_cpi);
1008 
1009 	mutex_enter(&cpu_lock);
1010 	procset |= 1 << cp->cpu_id;
1011 	mutex_exit(&cpu_lock);
1012 
1013 	if (tsc_gethrtime_enable)
1014 		tsc_sync_slave();
1015 
1016 	mutex_enter(&cpu_lock);
1017 	/*
1018 	 * It's unfortunate that chip_cpu_init() has to be called here.
1019 	 * It really belongs in cpu_add_unit(), but unfortunately it is
1020 	 * dependent on the cpuid probing, which must be done in the
1021 	 * context of the current CPU. Care must be taken on x86 to ensure
1022 	 * that mp_startup can safely block even though chip_cpu_init() and
1023 	 * cpu_add_active() have not yet been called.
1024 	 */
1025 	chip_cpu_init(cp);
1026 	chip_cpu_startup(cp);
1027 
1028 	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_ENABLE | CPU_EXISTS;
1029 	cpu_add_active(cp);
1030 	mutex_exit(&cpu_lock);
1031 
1032 	(void) spl0();				/* enable interrupts */
1033 
1034 	if (boothowto & RB_DEBUG)
1035 		kdi_dvec_cpu_init(cp);
1036 
1037 	/*
1038 	 * Setting the bit in cpu_ready_set must be the last operation in
1039 	 * processor initialization; the boot CPU will continue to boot once
1040 	 * it sees this bit set for all active CPUs.
1041 	 */
1042 	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1043 
1044 	/*
1045 	 * Because mp_startup() gets fired off after init() starts, we
1046 	 * can't use the '?' trick to do 'boot -v' printing - so we
1047 	 * always direct the 'cpu .. online' messages to the log.
1048 	 */
1049 	cmn_err(CE_CONT, "!cpu%d initialization complete - online\n",
1050 	    cp->cpu_id);
1051 
1052 	/*
1053 	 * Now we are done with the startup thread, so free it up.
1054 	 */
1055 	thread_exit();
1056 	panic("mp_startup: cannot return");
1057 	/*NOTREACHED*/
1058 }
1059 
1060 
1061 /*
1062  * Start CPU on user request.
1063  */
1064 /* ARGSUSED */
1065 int
1066 mp_cpu_start(struct cpu *cp)
1067 {
1068 	ASSERT(MUTEX_HELD(&cpu_lock));
1069 	if (cp->cpu_id == getbootcpuid())
1070 		return (EBUSY); 	/* Cannot start boot CPU */
1071 	return (0);
1072 }
1073 
1074 /*
1075  * Stop CPU on user request.
1076  */
1077 /* ARGSUSED */
1078 int
1079 mp_cpu_stop(struct cpu *cp)
1080 {
1081 	ASSERT(MUTEX_HELD(&cpu_lock));
1082 	if (cp->cpu_id == getbootcpuid())
1083 		return (EBUSY); 	/* Cannot stop boot CPU */
1084 
1085 	return (0);
1086 }
1087 
1088 /*
1089  * Power on CPU.
1090  */
1091 /* ARGSUSED */
1092 int
1093 mp_cpu_poweron(struct cpu *cp)
1094 {
1095 	ASSERT(MUTEX_HELD(&cpu_lock));
1096 	return (ENOTSUP);		/* not supported */
1097 }
1098 
1099 /*
1100  * Power off CPU.
1101  */
1102 /* ARGSUSED */
1103 int
1104 mp_cpu_poweroff(struct cpu *cp)
1105 {
1106 	ASSERT(MUTEX_HELD(&cpu_lock));
1107 	return (ENOTSUP);		/* not supported */
1108 }
1109 
1110 
1111 /*
1112  * Take the specified CPU out of participation in interrupts.
1113  */
1114 int
1115 cpu_disable_intr(struct cpu *cp)
1116 {
1117 	/*
1118 	 * cannot disable interrupts on boot cpu
1119 	 */
1120 	if (cp == cpu[getbootcpuid()])
1121 		return (EBUSY);
1122 
1123 	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1124 		return (EBUSY);
1125 
1126 	cp->cpu_flags &= ~CPU_ENABLE;
1127 	return (0);
1128 }
1129 
1130 /*
1131  * Allow the specified CPU to participate in interrupts.
1132  */
1133 void
1134 cpu_enable_intr(struct cpu *cp)
1135 {
1136 	ASSERT(MUTEX_HELD(&cpu_lock));
1137 	if (cp == cpu[getbootcpuid()])
1138 		return;
1139 
1140 	cp->cpu_flags |= CPU_ENABLE;
1141 	psm_enable_intr(cp->cpu_id);
1142 }
1143 
1144 
1145 /*
1146  * return the cpu id of the initial startup cpu
1147  */
1148 processorid_t
1149 getbootcpuid(void)
1150 {
1151 	return (0);
1152 }
1153 
1154 static ushort_t *
1155 mp_map_warm_reset_vector()
1156 {
1157 	ushort_t *warm_reset_vector;
1158 
1159 	if (!(warm_reset_vector = (ushort_t *)psm_map_phys(WARM_RESET_VECTOR,
1160 	    sizeof (ushort_t *), PROT_READ|PROT_WRITE)))
1161 		return (NULL);
1162 
1163 	/*
1164 	 * setup secondary cpu bios boot up vector
1165 	 */
1166 	*warm_reset_vector = (ushort_t)((caddr_t)
1167 		((struct rm_platter *)rm_platter_va)->rm_code - rm_platter_va
1168 		+ ((ulong_t)rm_platter_va & 0xf));
1169 	warm_reset_vector++;
1170 	*warm_reset_vector = (ushort_t)(rm_platter_pa >> 4);
1171 
1172 	--warm_reset_vector;
1173 	return (warm_reset_vector);
1174 }
1175 
1176 static void
1177 mp_unmap_warm_reset_vector(ushort_t *warm_reset_vector)
1178 {
1179 	psm_unmap_phys((caddr_t)warm_reset_vector, sizeof (ushort_t *));
1180 }
1181 
1182 /*ARGSUSED*/
1183 void
1184 mp_cpu_faulted_enter(struct cpu *cp)
1185 {}
1186 
1187 /*ARGSUSED*/
1188 void
1189 mp_cpu_faulted_exit(struct cpu *cp)
1190 {}
1191 
1192 /*
1193  * The following two routines are used as context operators on threads belonging
1194  * to processes with a private LDT (see sysi86).  Due to the rarity of such
1195  * processes, these routines are currently written for best code readability and
1196  * organization rather than speed.  We could avoid checking x86_feature at every
1197  * context switch by installing different context ops, depending on the
1198  * x86_feature flags, at LDT creation time -- one for each combination of fast
1199  * syscall feature flags.
1200  */
1201 
1202 /*ARGSUSED*/
1203 void
1204 cpu_fast_syscall_disable(void *arg)
1205 {
1206 	if (x86_feature & X86_SEP)
1207 		cpu_sep_disable();
1208 	if (x86_feature & X86_ASYSC)
1209 		cpu_asysc_disable();
1210 }
1211 
1212 /*ARGSUSED*/
1213 void
1214 cpu_fast_syscall_enable(void *arg)
1215 {
1216 	if (x86_feature & X86_SEP)
1217 		cpu_sep_enable();
1218 	if (x86_feature & X86_ASYSC)
1219 		cpu_asysc_enable();
1220 }
1221 
1222 static void
1223 cpu_sep_enable(void)
1224 {
1225 	uint64_t value;
1226 
1227 	ASSERT(x86_feature & X86_SEP);
1228 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1229 
1230 	value = KCS_SEL;
1231 	wrmsr(MSR_INTC_SEP_CS, &value);
1232 }
1233 
1234 static void
1235 cpu_sep_disable(void)
1236 {
1237 	uint64_t value;
1238 
1239 	ASSERT(x86_feature & X86_SEP);
1240 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1241 
1242 	/*
1243 	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
1244 	 * the sysenter or sysexit instruction to trigger a #gp fault.
1245 	 */
1246 	value = 0;
1247 	wrmsr(MSR_INTC_SEP_CS, &value);
1248 }
1249 
1250 static void
1251 cpu_asysc_enable(void)
1252 {
1253 	uint64_t value;
1254 
1255 	ASSERT(x86_feature & X86_ASYSC);
1256 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1257 
1258 	(void) rdmsr(MSR_AMD_EFER, &value);
1259 	value |= AMD_EFER_SCE;
1260 	wrmsr(MSR_AMD_EFER, &value);
1261 }
1262 
1263 static void
1264 cpu_asysc_disable(void)
1265 {
1266 	uint64_t value;
1267 
1268 	ASSERT(x86_feature & X86_ASYSC);
1269 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1270 
1271 	/*
1272 	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
1273 	 * executing syscall or sysret with this bit off will incur a #ud trap.
1274 	 */
1275 	(void) rdmsr(MSR_AMD_EFER, &value);
1276 	value &= ~AMD_EFER_SCE;
1277 	wrmsr(MSR_AMD_EFER, &value);
1278 }
1279