xref: /titanic_41/usr/src/uts/i86pc/os/machdep.c (revision 9e39c5ba00a55fa05777cc94b148296af305e135)
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/t_lock.h>
29 #include <sys/param.h>
30 #include <sys/segments.h>
31 #include <sys/sysmacros.h>
32 #include <sys/signal.h>
33 #include <sys/systm.h>
34 #include <sys/user.h>
35 #include <sys/mman.h>
36 #include <sys/vm.h>
37 
38 #include <sys/disp.h>
39 #include <sys/class.h>
40 
41 #include <sys/proc.h>
42 #include <sys/buf.h>
43 #include <sys/kmem.h>
44 
45 #include <sys/reboot.h>
46 #include <sys/uadmin.h>
47 #include <sys/callb.h>
48 
49 #include <sys/cred.h>
50 #include <sys/vnode.h>
51 #include <sys/file.h>
52 
53 #include <sys/procfs.h>
54 #include <sys/acct.h>
55 
56 #include <sys/vfs.h>
57 #include <sys/dnlc.h>
58 #include <sys/var.h>
59 #include <sys/cmn_err.h>
60 #include <sys/utsname.h>
61 #include <sys/debug.h>
62 
63 #include <sys/dumphdr.h>
64 #include <sys/bootconf.h>
65 #include <sys/varargs.h>
66 #include <sys/promif.h>
67 #include <sys/modctl.h>
68 
69 #include <sys/consdev.h>
70 #include <sys/frame.h>
71 
72 #include <sys/sunddi.h>
73 #include <sys/ddidmareq.h>
74 #include <sys/psw.h>
75 #include <sys/regset.h>
76 #include <sys/privregs.h>
77 #include <sys/clock.h>
78 #include <sys/tss.h>
79 #include <sys/cpu.h>
80 #include <sys/stack.h>
81 #include <sys/trap.h>
82 #include <sys/pic.h>
83 #include <vm/hat.h>
84 #include <vm/anon.h>
85 #include <vm/as.h>
86 #include <vm/page.h>
87 #include <vm/seg.h>
88 #include <vm/seg_kmem.h>
89 #include <vm/seg_map.h>
90 #include <vm/seg_vn.h>
91 #include <vm/seg_kp.h>
92 #include <vm/hat_i86.h>
93 #include <sys/swap.h>
94 #include <sys/thread.h>
95 #include <sys/sysconf.h>
96 #include <sys/vm_machparam.h>
97 #include <sys/archsystm.h>
98 #include <sys/machsystm.h>
99 #include <sys/machlock.h>
100 #include <sys/x_call.h>
101 #include <sys/instance.h>
102 
103 #include <sys/time.h>
104 #include <sys/smp_impldefs.h>
105 #include <sys/psm_types.h>
106 #include <sys/atomic.h>
107 #include <sys/panic.h>
108 #include <sys/cpuvar.h>
109 #include <sys/dtrace.h>
110 #include <sys/bl.h>
111 #include <sys/nvpair.h>
112 #include <sys/x86_archext.h>
113 #include <sys/pool_pset.h>
114 #include <sys/autoconf.h>
115 #include <sys/mem.h>
116 #include <sys/dumphdr.h>
117 #include <sys/compress.h>
118 #include <sys/cpu_module.h>
119 #if defined(__xpv)
120 #include <sys/hypervisor.h>
121 #include <sys/xpv_panic.h>
122 #endif
123 
124 #include <sys/fastboot.h>
125 #include <sys/machelf.h>
126 #include <sys/kobj.h>
127 #include <sys/multiboot.h>
128 
129 #ifdef	TRAPTRACE
130 #include <sys/traptrace.h>
131 #endif	/* TRAPTRACE */
132 
133 extern void audit_enterprom(int);
134 extern void audit_exitprom(int);
135 
136 /*
137  * Occassionally the kernel knows better whether to power-off or reboot.
138  */
139 int force_shutdown_method = AD_UNKNOWN;
140 
141 /*
142  * The panicbuf array is used to record messages and state:
143  */
144 char panicbuf[PANICBUFSIZE];
145 
146 /*
147  * maxphys - used during physio
148  * klustsize - used for klustering by swapfs and specfs
149  */
150 int maxphys = 56 * 1024;    /* XXX See vm_subr.c - max b_count in physio */
151 int klustsize = 56 * 1024;
152 
153 caddr_t	p0_va;		/* Virtual address for accessing physical page 0 */
154 
155 /*
156  * defined here, though unused on x86,
157  * to make kstat_fr.c happy.
158  */
159 int vac;
160 
161 void debug_enter(char *);
162 
163 extern void pm_cfb_check_and_powerup(void);
164 extern void pm_cfb_rele(void);
165 
166 extern fastboot_info_t newkernel;
167 
168 /*
169  * Machine dependent code to reboot.
170  * "mdep" is interpreted as a character pointer; if non-null, it is a pointer
171  * to a string to be used as the argument string when rebooting.
172  *
173  * "invoke_cb" is a boolean. It is set to true when mdboot() can safely
174  * invoke CB_CL_MDBOOT callbacks before shutting the system down, i.e. when
175  * we are in a normal shutdown sequence (interrupts are not blocked, the
176  * system is not panic'ing or being suspended).
177  */
178 /*ARGSUSED*/
179 void
180 mdboot(int cmd, int fcn, char *mdep, boolean_t invoke_cb)
181 {
182 	processorid_t bootcpuid = 0;
183 	static int is_first_quiesce = 1;
184 	static int is_first_reset = 1;
185 	int reset_status = 0;
186 	static char fallback_str[] = "Falling back to regular reboot.\n";
187 
188 	if (fcn == AD_FASTREBOOT && !newkernel.fi_valid)
189 		fcn = AD_BOOT;
190 
191 	if (!panicstr) {
192 		kpreempt_disable();
193 		if (fcn == AD_FASTREBOOT) {
194 			mutex_enter(&cpu_lock);
195 			if (CPU_ACTIVE(cpu_get(bootcpuid))) {
196 				affinity_set(bootcpuid);
197 			}
198 			mutex_exit(&cpu_lock);
199 		} else {
200 			affinity_set(CPU_CURRENT);
201 		}
202 	}
203 
204 	if (force_shutdown_method != AD_UNKNOWN)
205 		fcn = force_shutdown_method;
206 
207 	/*
208 	 * XXX - rconsvp is set to NULL to ensure that output messages
209 	 * are sent to the underlying "hardware" device using the
210 	 * monitor's printf routine since we are in the process of
211 	 * either rebooting or halting the machine.
212 	 */
213 	rconsvp = NULL;
214 
215 	/*
216 	 * Print the reboot message now, before pausing other cpus.
217 	 * There is a race condition in the printing support that
218 	 * can deadlock multiprocessor machines.
219 	 */
220 	if (!(fcn == AD_HALT || fcn == AD_POWEROFF))
221 		prom_printf("rebooting...\n");
222 
223 	if (IN_XPV_PANIC())
224 		reset();
225 
226 	/*
227 	 * We can't bring up the console from above lock level, so do it now
228 	 */
229 	pm_cfb_check_and_powerup();
230 
231 	/* make sure there are no more changes to the device tree */
232 	devtree_freeze();
233 
234 	if (invoke_cb)
235 		(void) callb_execute_class(CB_CL_MDBOOT, NULL);
236 
237 	/*
238 	 * Clear any unresolved UEs from memory.
239 	 */
240 	page_retire_mdboot();
241 
242 #if defined(__xpv)
243 	/*
244 	 * XXPV	Should probably think some more about how we deal
245 	 *	with panicing before it's really safe to panic.
246 	 *	On hypervisors, we reboot very quickly..  Perhaps panic
247 	 *	should only attempt to recover by rebooting if,
248 	 *	say, we were able to mount the root filesystem,
249 	 *	or if we successfully launched init(1m).
250 	 */
251 	if (panicstr && proc_init == NULL)
252 		(void) HYPERVISOR_shutdown(SHUTDOWN_poweroff);
253 #endif
254 	/*
255 	 * stop other cpus and raise our priority.  since there is only
256 	 * one active cpu after this, and our priority will be too high
257 	 * for us to be preempted, we're essentially single threaded
258 	 * from here on out.
259 	 */
260 	(void) spl6();
261 	if (!panicstr) {
262 		mutex_enter(&cpu_lock);
263 		pause_cpus(NULL);
264 		mutex_exit(&cpu_lock);
265 	}
266 
267 	/*
268 	 * If the system is panicking, the preloaded kernel is valid,
269 	 * and fastreboot_onpanic has been set, choose Fast Reboot.
270 	 */
271 	if (fcn == AD_BOOT && panicstr && newkernel.fi_valid &&
272 	    fastreboot_onpanic)
273 		fcn = AD_FASTREBOOT;
274 
275 	/*
276 	 * Try to quiesce devices.
277 	 */
278 	if (is_first_quiesce) {
279 		/*
280 		 * Clear is_first_quiesce before calling quiesce_devices()
281 		 * so that if quiesce_devices() causes panics, it will not
282 		 * be invoked again.
283 		 */
284 		is_first_quiesce = 0;
285 
286 		quiesce_active = 1;
287 		quiesce_devices(ddi_root_node(), &reset_status);
288 		if (reset_status == -1) {
289 			if (fcn == AD_FASTREBOOT && !force_fastreboot) {
290 				prom_printf("Driver(s) not capable of fast "
291 				    "reboot.\n");
292 				prom_printf(fallback_str);
293 				fastreboot_capable = 0;
294 				fcn = AD_BOOT;
295 			} else if (fcn != AD_FASTREBOOT)
296 				fastreboot_capable = 0;
297 		}
298 		quiesce_active = 0;
299 	}
300 
301 	/*
302 	 * Try to reset devices. reset_leaves() should only be called
303 	 * a) when there are no other threads that could be accessing devices,
304 	 *    and
305 	 * b) on a system that's not capable of fast reboot (fastreboot_capable
306 	 *    being 0), or on a system where quiesce_devices() failed to
307 	 *    complete (quiesce_active being 1).
308 	 */
309 	if (is_first_reset && (!fastreboot_capable || quiesce_active)) {
310 		/*
311 		 * Clear is_first_reset before calling reset_devices()
312 		 * so that if reset_devices() causes panics, it will not
313 		 * be invoked again.
314 		 */
315 		is_first_reset = 0;
316 		reset_leaves();
317 	}
318 
319 	/* Verify newkernel checksum */
320 	if (fastreboot_capable && fcn == AD_FASTREBOOT &&
321 	    fastboot_cksum_verify(&newkernel) != 0) {
322 		fastreboot_capable = 0;
323 		prom_printf("Fast reboot: checksum failed for the new "
324 		    "kernel.\n");
325 		prom_printf(fallback_str);
326 	}
327 
328 	(void) spl8();
329 
330 	if (fastreboot_capable && fcn == AD_FASTREBOOT) {
331 		/*
332 		 * psm_shutdown is called within fast_reboot()
333 		 */
334 		fast_reboot();
335 	} else {
336 		(*psm_shutdownf)(cmd, fcn);
337 
338 		if (fcn == AD_HALT || fcn == AD_POWEROFF)
339 			halt((char *)NULL);
340 		else
341 			prom_reboot("");
342 	}
343 	/*NOTREACHED*/
344 }
345 
346 /* mdpreboot - may be called prior to mdboot while root fs still mounted */
347 /*ARGSUSED*/
348 void
349 mdpreboot(int cmd, int fcn, char *mdep)
350 {
351 	if (fcn == AD_FASTREBOOT && !fastreboot_capable) {
352 		fcn = AD_BOOT;
353 #ifdef	__xpv
354 		cmn_err(CE_WARN, "Fast reboot is not supported on xVM");
355 #else
356 		cmn_err(CE_WARN,
357 		    "Fast reboot is not supported on this platform");
358 #endif
359 	}
360 
361 	if (fcn == AD_FASTREBOOT) {
362 		fastboot_load_kernel(mdep);
363 		if (!newkernel.fi_valid)
364 			fcn = AD_BOOT;
365 	}
366 
367 	(*psm_preshutdownf)(cmd, fcn);
368 }
369 
370 static void
371 stop_other_cpus(void)
372 {
373 	ulong_t s = clear_int_flag(); /* fast way to keep CPU from changing */
374 	cpuset_t xcset;
375 
376 	CPUSET_ALL_BUT(xcset, CPU->cpu_id);
377 	xc_priority(0, 0, 0, CPUSET2BV(xcset), (xc_func_t)mach_cpu_halt);
378 	restore_int_flag(s);
379 }
380 
381 /*
382  *	Machine dependent abort sequence handling
383  */
384 void
385 abort_sequence_enter(char *msg)
386 {
387 	if (abort_enable == 0) {
388 		if (audit_active)
389 			audit_enterprom(0);
390 		return;
391 	}
392 	if (audit_active)
393 		audit_enterprom(1);
394 	debug_enter(msg);
395 	if (audit_active)
396 		audit_exitprom(1);
397 }
398 
399 /*
400  * Enter debugger.  Called when the user types ctrl-alt-d or whenever
401  * code wants to enter the debugger and possibly resume later.
402  */
403 void
404 debug_enter(
405 	char	*msg)		/* message to print, possibly NULL */
406 {
407 	if (dtrace_debugger_init != NULL)
408 		(*dtrace_debugger_init)();
409 
410 	if (msg)
411 		prom_printf("%s\n", msg);
412 
413 	if (boothowto & RB_DEBUG)
414 		kmdb_enter();
415 
416 	if (dtrace_debugger_fini != NULL)
417 		(*dtrace_debugger_fini)();
418 }
419 
420 void
421 reset(void)
422 {
423 #if !defined(__xpv)
424 	ushort_t *bios_memchk;
425 
426 	/*
427 	 * Can't use psm_map_phys before the hat is initialized.
428 	 */
429 	if (khat_running) {
430 		bios_memchk = (ushort_t *)psm_map_phys(0x472,
431 		    sizeof (ushort_t), PROT_READ | PROT_WRITE);
432 		if (bios_memchk)
433 			*bios_memchk = 0x1234;	/* bios memory check disable */
434 	}
435 
436 	if (ddi_prop_exists(DDI_DEV_T_ANY, ddi_root_node(), 0, "efi-systab"))
437 		efi_reset();
438 	pc_reset();
439 #else
440 	if (IN_XPV_PANIC())
441 		pc_reset();
442 	(void) HYPERVISOR_shutdown(SHUTDOWN_reboot);
443 	panic("HYPERVISOR_shutdown() failed");
444 #endif
445 	/*NOTREACHED*/
446 }
447 
448 /*
449  * Halt the machine and return to the monitor
450  */
451 void
452 halt(char *s)
453 {
454 	stop_other_cpus();	/* send stop signal to other CPUs */
455 	if (s)
456 		prom_printf("(%s) \n", s);
457 	prom_exit_to_mon();
458 	/*NOTREACHED*/
459 }
460 
461 /*
462  * Initiate interrupt redistribution.
463  */
464 void
465 i_ddi_intr_redist_all_cpus()
466 {
467 }
468 
469 /*
470  * XXX These probably ought to live somewhere else
471  * XXX They are called from mem.c
472  */
473 
474 /*
475  * Convert page frame number to an OBMEM page frame number
476  * (i.e. put in the type bits -- zero for this implementation)
477  */
478 pfn_t
479 impl_obmem_pfnum(pfn_t pf)
480 {
481 	return (pf);
482 }
483 
484 #ifdef	NM_DEBUG
485 int nmi_test = 0;	/* checked in intentry.s during clock int */
486 int nmtest = -1;
487 nmfunc1(arg, rp)
488 int	arg;
489 struct regs *rp;
490 {
491 	printf("nmi called with arg = %x, regs = %x\n", arg, rp);
492 	nmtest += 50;
493 	if (arg == nmtest) {
494 		printf("ip = %x\n", rp->r_pc);
495 		return (1);
496 	}
497 	return (0);
498 }
499 
500 #endif
501 
502 #include <sys/bootsvcs.h>
503 
504 /* Hacked up initialization for initial kernel check out is HERE. */
505 /* The basic steps are: */
506 /*	kernel bootfuncs definition/initialization for KADB */
507 /*	kadb bootfuncs pointer initialization */
508 /*	putchar/getchar (interrupts disabled) */
509 
510 /* kadb bootfuncs pointer initialization */
511 
512 int
513 sysp_getchar()
514 {
515 	int i;
516 	ulong_t s;
517 
518 	if (cons_polledio == NULL) {
519 		/* Uh oh */
520 		prom_printf("getchar called with no console\n");
521 		for (;;)
522 			/* LOOP FOREVER */;
523 	}
524 
525 	s = clear_int_flag();
526 	i = cons_polledio->cons_polledio_getchar(
527 	    cons_polledio->cons_polledio_argument);
528 	restore_int_flag(s);
529 	return (i);
530 }
531 
532 void
533 sysp_putchar(int c)
534 {
535 	ulong_t s;
536 
537 	/*
538 	 * We have no alternative but to drop the output on the floor.
539 	 */
540 	if (cons_polledio == NULL ||
541 	    cons_polledio->cons_polledio_putchar == NULL)
542 		return;
543 
544 	s = clear_int_flag();
545 	cons_polledio->cons_polledio_putchar(
546 	    cons_polledio->cons_polledio_argument, c);
547 	restore_int_flag(s);
548 }
549 
550 int
551 sysp_ischar()
552 {
553 	int i;
554 	ulong_t s;
555 
556 	if (cons_polledio == NULL ||
557 	    cons_polledio->cons_polledio_ischar == NULL)
558 		return (0);
559 
560 	s = clear_int_flag();
561 	i = cons_polledio->cons_polledio_ischar(
562 	    cons_polledio->cons_polledio_argument);
563 	restore_int_flag(s);
564 	return (i);
565 }
566 
567 int
568 goany(void)
569 {
570 	prom_printf("Type any key to continue ");
571 	(void) prom_getchar();
572 	prom_printf("\n");
573 	return (1);
574 }
575 
576 static struct boot_syscalls kern_sysp = {
577 	sysp_getchar,	/*	unchar	(*getchar)();	7  */
578 	sysp_putchar,	/*	int	(*putchar)();	8  */
579 	sysp_ischar,	/*	int	(*ischar)();	9  */
580 };
581 
582 #if defined(__xpv)
583 int using_kern_polledio;
584 #endif
585 
586 void
587 kadb_uses_kernel()
588 {
589 	/*
590 	 * This routine is now totally misnamed, since it does not in fact
591 	 * control kadb's I/O; it only controls the kernel's prom_* I/O.
592 	 */
593 	sysp = &kern_sysp;
594 #if defined(__xpv)
595 	using_kern_polledio = 1;
596 #endif
597 }
598 
599 /*
600  *	the interface to the outside world
601  */
602 
603 /*
604  * poll_port -- wait for a register to achieve a
605  *		specific state.  Arguments are a mask of bits we care about,
606  *		and two sub-masks.  To return normally, all the bits in the
607  *		first sub-mask must be ON, all the bits in the second sub-
608  *		mask must be OFF.  If about seconds pass without the register
609  *		achieving the desired bit configuration, we return 1, else
610  *		0.
611  */
612 int
613 poll_port(ushort_t port, ushort_t mask, ushort_t onbits, ushort_t offbits)
614 {
615 	int i;
616 	ushort_t maskval;
617 
618 	for (i = 500000; i; i--) {
619 		maskval = inb(port) & mask;
620 		if (((maskval & onbits) == onbits) &&
621 		    ((maskval & offbits) == 0))
622 			return (0);
623 		drv_usecwait(10);
624 	}
625 	return (1);
626 }
627 
628 /*
629  * set_idle_cpu is called from idle() when a CPU becomes idle.
630  */
631 /*LINTED: static unused */
632 static uint_t last_idle_cpu;
633 
634 /*ARGSUSED*/
635 void
636 set_idle_cpu(int cpun)
637 {
638 	last_idle_cpu = cpun;
639 	(*psm_set_idle_cpuf)(cpun);
640 }
641 
642 /*
643  * unset_idle_cpu is called from idle() when a CPU is no longer idle.
644  */
645 /*ARGSUSED*/
646 void
647 unset_idle_cpu(int cpun)
648 {
649 	(*psm_unset_idle_cpuf)(cpun);
650 }
651 
652 /*
653  * This routine is almost correct now, but not quite.  It still needs the
654  * equivalent concept of "hres_last_tick", just like on the sparc side.
655  * The idea is to take a snapshot of the hi-res timer while doing the
656  * hrestime_adj updates under hres_lock in locore, so that the small
657  * interval between interrupt assertion and interrupt processing is
658  * accounted for correctly.  Once we have this, the code below should
659  * be modified to subtract off hres_last_tick rather than hrtime_base.
660  *
661  * I'd have done this myself, but I don't have source to all of the
662  * vendor-specific hi-res timer routines (grrr...).  The generic hook I
663  * need is something like "gethrtime_unlocked()", which would be just like
664  * gethrtime() but would assume that you're already holding CLOCK_LOCK().
665  * This is what the GET_HRTIME() macro is for on sparc (although it also
666  * serves the function of making time available without a function call
667  * so you don't take a register window overflow while traps are disabled).
668  */
669 void
670 pc_gethrestime(timestruc_t *tp)
671 {
672 	int lock_prev;
673 	timestruc_t now;
674 	int nslt;		/* nsec since last tick */
675 	int adj;		/* amount of adjustment to apply */
676 
677 loop:
678 	lock_prev = hres_lock;
679 	now = hrestime;
680 	nslt = (int)(gethrtime() - hres_last_tick);
681 	if (nslt < 0) {
682 		/*
683 		 * nslt < 0 means a tick came between sampling
684 		 * gethrtime() and hres_last_tick; restart the loop
685 		 */
686 
687 		goto loop;
688 	}
689 	now.tv_nsec += nslt;
690 	if (hrestime_adj != 0) {
691 		if (hrestime_adj > 0) {
692 			adj = (nslt >> ADJ_SHIFT);
693 			if (adj > hrestime_adj)
694 				adj = (int)hrestime_adj;
695 		} else {
696 			adj = -(nslt >> ADJ_SHIFT);
697 			if (adj < hrestime_adj)
698 				adj = (int)hrestime_adj;
699 		}
700 		now.tv_nsec += adj;
701 	}
702 	while ((unsigned long)now.tv_nsec >= NANOSEC) {
703 
704 		/*
705 		 * We might have a large adjustment or have been in the
706 		 * debugger for a long time; take care of (at most) four
707 		 * of those missed seconds (tv_nsec is 32 bits, so
708 		 * anything >4s will be wrapping around).  However,
709 		 * anything more than 2 seconds out of sync will trigger
710 		 * timedelta from clock() to go correct the time anyway,
711 		 * so do what we can, and let the big crowbar do the
712 		 * rest.  A similar correction while loop exists inside
713 		 * hres_tick(); in all cases we'd like tv_nsec to
714 		 * satisfy 0 <= tv_nsec < NANOSEC to avoid confusing
715 		 * user processes, but if tv_sec's a little behind for a
716 		 * little while, that's OK; time still monotonically
717 		 * increases.
718 		 */
719 
720 		now.tv_nsec -= NANOSEC;
721 		now.tv_sec++;
722 	}
723 	if ((hres_lock & ~1) != lock_prev)
724 		goto loop;
725 
726 	*tp = now;
727 }
728 
729 void
730 gethrestime_lasttick(timespec_t *tp)
731 {
732 	int s;
733 
734 	s = hr_clock_lock();
735 	*tp = hrestime;
736 	hr_clock_unlock(s);
737 }
738 
739 time_t
740 gethrestime_sec(void)
741 {
742 	timestruc_t now;
743 
744 	gethrestime(&now);
745 	return (now.tv_sec);
746 }
747 
748 /*
749  * Initialize a kernel thread's stack
750  */
751 
752 caddr_t
753 thread_stk_init(caddr_t stk)
754 {
755 	ASSERT(((uintptr_t)stk & (STACK_ALIGN - 1)) == 0);
756 	return (stk - SA(MINFRAME));
757 }
758 
759 /*
760  * Initialize lwp's kernel stack.
761  */
762 
763 #ifdef TRAPTRACE
764 /*
765  * There's a tricky interdependency here between use of sysenter and
766  * TRAPTRACE which needs recording to avoid future confusion (this is
767  * about the third time I've re-figured this out ..)
768  *
769  * Here's how debugging lcall works with TRAPTRACE.
770  *
771  * 1 We're in userland with a breakpoint on the lcall instruction.
772  * 2 We execute the instruction - the instruction pushes the userland
773  *   %ss, %esp, %efl, %cs, %eip on the stack and zips into the kernel
774  *   via the call gate.
775  * 3 The hardware raises a debug trap in kernel mode, the hardware
776  *   pushes %efl, %cs, %eip and gets to dbgtrap via the idt.
777  * 4 dbgtrap pushes the error code and trapno and calls cmntrap
778  * 5 cmntrap finishes building a trap frame
779  * 6 The TRACE_REGS macros in cmntrap copy a REGSIZE worth chunk
780  *   off the stack into the traptrace buffer.
781  *
782  * This means that the traptrace buffer contains the wrong values in
783  * %esp and %ss, but everything else in there is correct.
784  *
785  * Here's how debugging sysenter works with TRAPTRACE.
786  *
787  * a We're in userland with a breakpoint on the sysenter instruction.
788  * b We execute the instruction - the instruction pushes -nothing-
789  *   on the stack, but sets %cs, %eip, %ss, %esp to prearranged
790  *   values to take us to sys_sysenter, at the top of the lwp's
791  *   stack.
792  * c goto 3
793  *
794  * At this point, because we got into the kernel without the requisite
795  * five pushes on the stack, if we didn't make extra room, we'd
796  * end up with the TRACE_REGS macro fetching the saved %ss and %esp
797  * values from negative (unmapped) stack addresses -- which really bites.
798  * That's why we do the '-= 8' below.
799  *
800  * XXX	Note that reading "up" lwp0's stack works because t0 is declared
801  *	right next to t0stack in locore.s
802  */
803 #endif
804 
805 caddr_t
806 lwp_stk_init(klwp_t *lwp, caddr_t stk)
807 {
808 	caddr_t oldstk;
809 	struct pcb *pcb = &lwp->lwp_pcb;
810 
811 	oldstk = stk;
812 	stk -= SA(sizeof (struct regs) + SA(MINFRAME));
813 #ifdef TRAPTRACE
814 	stk -= 2 * sizeof (greg_t); /* space for phony %ss:%sp (see above) */
815 #endif
816 	stk = (caddr_t)((uintptr_t)stk & ~(STACK_ALIGN - 1ul));
817 	bzero(stk, oldstk - stk);
818 	lwp->lwp_regs = (void *)(stk + SA(MINFRAME));
819 
820 	/*
821 	 * Arrange that the virtualized %fs and %gs GDT descriptors
822 	 * have a well-defined initial state (present, ring 3
823 	 * and of type data).
824 	 */
825 #if defined(__amd64)
826 	if (lwp_getdatamodel(lwp) == DATAMODEL_NATIVE)
827 		pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_udesc;
828 	else
829 		pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_u32desc;
830 #elif defined(__i386)
831 	pcb->pcb_fsdesc = pcb->pcb_gsdesc = zero_udesc;
832 #endif	/* __i386 */
833 	lwp_installctx(lwp);
834 	return (stk);
835 }
836 
837 /*ARGSUSED*/
838 void
839 lwp_stk_fini(klwp_t *lwp)
840 {}
841 
842 /*
843  * If we're not the panic CPU, we wait in panic_idle for reboot.
844  */
845 void
846 panic_idle(void)
847 {
848 	splx(ipltospl(CLOCK_LEVEL));
849 	(void) setjmp(&curthread->t_pcb);
850 
851 #ifndef __xpv
852 	for (;;)
853 		i86_halt();
854 #else
855 	for (;;)
856 		;
857 #endif
858 }
859 
860 /*
861  * Stop the other CPUs by cross-calling them and forcing them to enter
862  * the panic_idle() loop above.
863  */
864 /*ARGSUSED*/
865 void
866 panic_stopcpus(cpu_t *cp, kthread_t *t, int spl)
867 {
868 	processorid_t i;
869 	cpuset_t xcset;
870 
871 	/*
872 	 * In the case of a Xen panic, the hypervisor has already stopped
873 	 * all of the CPUs.
874 	 */
875 	if (!IN_XPV_PANIC()) {
876 		(void) splzs();
877 
878 		CPUSET_ALL_BUT(xcset, cp->cpu_id);
879 		xc_priority(0, 0, 0, CPUSET2BV(xcset), (xc_func_t)panic_idle);
880 	}
881 
882 	for (i = 0; i < NCPU; i++) {
883 		if (i != cp->cpu_id && cpu[i] != NULL &&
884 		    (cpu[i]->cpu_flags & CPU_EXISTS))
885 			cpu[i]->cpu_flags |= CPU_QUIESCED;
886 	}
887 }
888 
889 /*
890  * Platform callback following each entry to panicsys().
891  */
892 /*ARGSUSED*/
893 void
894 panic_enter_hw(int spl)
895 {
896 	/* Nothing to do here */
897 }
898 
899 /*
900  * Platform-specific code to execute after panicstr is set: we invoke
901  * the PSM entry point to indicate that a panic has occurred.
902  */
903 /*ARGSUSED*/
904 void
905 panic_quiesce_hw(panic_data_t *pdp)
906 {
907 	psm_notifyf(PSM_PANIC_ENTER);
908 
909 	cmi_panic_callback();
910 
911 #ifdef	TRAPTRACE
912 	/*
913 	 * Turn off TRAPTRACE
914 	 */
915 	TRAPTRACE_FREEZE;
916 #endif	/* TRAPTRACE */
917 }
918 
919 /*
920  * Platform callback prior to writing crash dump.
921  */
922 /*ARGSUSED*/
923 void
924 panic_dump_hw(int spl)
925 {
926 	/* Nothing to do here */
927 }
928 
929 void *
930 plat_traceback(void *fpreg)
931 {
932 #ifdef __xpv
933 	if (IN_XPV_PANIC())
934 		return (xpv_traceback(fpreg));
935 #endif
936 	return (fpreg);
937 }
938 
939 /*ARGSUSED*/
940 void
941 plat_tod_fault(enum tod_fault_type tod_bad)
942 {}
943 
944 /*ARGSUSED*/
945 int
946 blacklist(int cmd, const char *scheme, nvlist_t *fmri, const char *class)
947 {
948 	return (ENOTSUP);
949 }
950 
951 /*
952  * The underlying console output routines are protected by raising IPL in case
953  * we are still calling into the early boot services.  Once we start calling
954  * the kernel console emulator, it will disable interrupts completely during
955  * character rendering (see sysp_putchar, for example).  Refer to the comments
956  * and code in common/os/console.c for more information on these callbacks.
957  */
958 /*ARGSUSED*/
959 int
960 console_enter(int busy)
961 {
962 	return (splzs());
963 }
964 
965 /*ARGSUSED*/
966 void
967 console_exit(int busy, int spl)
968 {
969 	splx(spl);
970 }
971 
972 /*
973  * Allocate a region of virtual address space, unmapped.
974  * Stubbed out except on sparc, at least for now.
975  */
976 /*ARGSUSED*/
977 void *
978 boot_virt_alloc(void *addr, size_t size)
979 {
980 	return (addr);
981 }
982 
983 volatile unsigned long	tenmicrodata;
984 
985 void
986 tenmicrosec(void)
987 {
988 	extern int gethrtime_hires;
989 
990 	if (gethrtime_hires) {
991 		hrtime_t start, end;
992 		start = end =  gethrtime();
993 		while ((end - start) < (10 * (NANOSEC / MICROSEC))) {
994 			SMT_PAUSE();
995 			end = gethrtime();
996 		}
997 	} else {
998 #if defined(__xpv)
999 		hrtime_t newtime;
1000 
1001 		newtime = xpv_gethrtime() + 10000; /* now + 10 us */
1002 		while (xpv_gethrtime() < newtime)
1003 			SMT_PAUSE();
1004 #else	/* __xpv */
1005 		int i;
1006 
1007 		/*
1008 		 * Artificial loop to induce delay.
1009 		 */
1010 		for (i = 0; i < microdata; i++)
1011 			tenmicrodata = microdata;
1012 #endif	/* __xpv */
1013 	}
1014 }
1015 
1016 /*
1017  * get_cpu_mstate() is passed an array of timestamps, NCMSTATES
1018  * long, and it fills in the array with the time spent on cpu in
1019  * each of the mstates, where time is returned in nsec.
1020  *
1021  * No guarantee is made that the returned values in times[] will
1022  * monotonically increase on sequential calls, although this will
1023  * be true in the long run. Any such guarantee must be handled by
1024  * the caller, if needed. This can happen if we fail to account
1025  * for elapsed time due to a generation counter conflict, yet we
1026  * did account for it on a prior call (see below).
1027  *
1028  * The complication is that the cpu in question may be updating
1029  * its microstate at the same time that we are reading it.
1030  * Because the microstate is only updated when the CPU's state
1031  * changes, the values in cpu_intracct[] can be indefinitely out
1032  * of date. To determine true current values, it is necessary to
1033  * compare the current time with cpu_mstate_start, and add the
1034  * difference to times[cpu_mstate].
1035  *
1036  * This can be a problem if those values are changing out from
1037  * under us. Because the code path in new_cpu_mstate() is
1038  * performance critical, we have not added a lock to it. Instead,
1039  * we have added a generation counter. Before beginning
1040  * modifications, the counter is set to 0. After modifications,
1041  * it is set to the old value plus one.
1042  *
1043  * get_cpu_mstate() will not consider the values of cpu_mstate
1044  * and cpu_mstate_start to be usable unless the value of
1045  * cpu_mstate_gen is both non-zero and unchanged, both before and
1046  * after reading the mstate information. Note that we must
1047  * protect against out-of-order loads around accesses to the
1048  * generation counter. Also, this is a best effort approach in
1049  * that we do not retry should the counter be found to have
1050  * changed.
1051  *
1052  * cpu_intracct[] is used to identify time spent in each CPU
1053  * mstate while handling interrupts. Such time should be reported
1054  * against system time, and so is subtracted out from its
1055  * corresponding cpu_acct[] time and added to
1056  * cpu_acct[CMS_SYSTEM].
1057  */
1058 
1059 void
1060 get_cpu_mstate(cpu_t *cpu, hrtime_t *times)
1061 {
1062 	int i;
1063 	hrtime_t now, start;
1064 	uint16_t gen;
1065 	uint16_t state;
1066 	hrtime_t intracct[NCMSTATES];
1067 
1068 	/*
1069 	 * Load all volatile state under the protection of membar.
1070 	 * cpu_acct[cpu_mstate] must be loaded to avoid double counting
1071 	 * of (now - cpu_mstate_start) by a change in CPU mstate that
1072 	 * arrives after we make our last check of cpu_mstate_gen.
1073 	 */
1074 
1075 	now = gethrtime_unscaled();
1076 	gen = cpu->cpu_mstate_gen;
1077 
1078 	membar_consumer();	/* guarantee load ordering */
1079 	start = cpu->cpu_mstate_start;
1080 	state = cpu->cpu_mstate;
1081 	for (i = 0; i < NCMSTATES; i++) {
1082 		intracct[i] = cpu->cpu_intracct[i];
1083 		times[i] = cpu->cpu_acct[i];
1084 	}
1085 	membar_consumer();	/* guarantee load ordering */
1086 
1087 	if (gen != 0 && gen == cpu->cpu_mstate_gen && now > start)
1088 		times[state] += now - start;
1089 
1090 	for (i = 0; i < NCMSTATES; i++) {
1091 		if (i == CMS_SYSTEM)
1092 			continue;
1093 		times[i] -= intracct[i];
1094 		if (times[i] < 0) {
1095 			intracct[i] += times[i];
1096 			times[i] = 0;
1097 		}
1098 		times[CMS_SYSTEM] += intracct[i];
1099 		scalehrtime(&times[i]);
1100 	}
1101 	scalehrtime(&times[CMS_SYSTEM]);
1102 }
1103 
1104 /*
1105  * This is a version of the rdmsr instruction that allows
1106  * an error code to be returned in the case of failure.
1107  */
1108 int
1109 checked_rdmsr(uint_t msr, uint64_t *value)
1110 {
1111 	if ((x86_feature & X86_MSR) == 0)
1112 		return (ENOTSUP);
1113 	*value = rdmsr(msr);
1114 	return (0);
1115 }
1116 
1117 /*
1118  * This is a version of the wrmsr instruction that allows
1119  * an error code to be returned in the case of failure.
1120  */
1121 int
1122 checked_wrmsr(uint_t msr, uint64_t value)
1123 {
1124 	if ((x86_feature & X86_MSR) == 0)
1125 		return (ENOTSUP);
1126 	wrmsr(msr, value);
1127 	return (0);
1128 }
1129 
1130 /*
1131  * The mem driver's usual method of using hat_devload() to establish a
1132  * temporary mapping will not work for foreign pages mapped into this
1133  * domain or for the special hypervisor-provided pages.  For the foreign
1134  * pages, we often don't know which domain owns them, so we can't ask the
1135  * hypervisor to set up a new mapping.  For the other pages, we don't have
1136  * a pfn, so we can't create a new PTE.  For these special cases, we do a
1137  * direct uiomove() from the existing kernel virtual address.
1138  */
1139 /*ARGSUSED*/
1140 int
1141 plat_mem_do_mmio(struct uio *uio, enum uio_rw rw)
1142 {
1143 #if defined(__xpv)
1144 	void *va = (void *)(uintptr_t)uio->uio_loffset;
1145 	off_t pageoff = uio->uio_loffset & PAGEOFFSET;
1146 	size_t nbytes = MIN((size_t)(PAGESIZE - pageoff),
1147 	    (size_t)uio->uio_iov->iov_len);
1148 
1149 	if ((rw == UIO_READ &&
1150 	    (va == HYPERVISOR_shared_info || va == xen_info)) ||
1151 	    (pfn_is_foreign(hat_getpfnum(kas.a_hat, va))))
1152 		return (uiomove(va, nbytes, rw, uio));
1153 #endif
1154 	return (ENOTSUP);
1155 }
1156 
1157 pgcnt_t
1158 num_phys_pages()
1159 {
1160 	pgcnt_t npages = 0;
1161 	struct memlist *mp;
1162 
1163 #if defined(__xpv)
1164 	if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1165 		xen_sysctl_t op;
1166 
1167 		op.cmd = XEN_SYSCTL_physinfo;
1168 		op.interface_version = XEN_SYSCTL_INTERFACE_VERSION;
1169 		if (HYPERVISOR_sysctl(&op) != 0)
1170 			panic("physinfo op refused");
1171 
1172 		return ((pgcnt_t)op.u.physinfo.total_pages);
1173 	}
1174 #endif /* __xpv */
1175 
1176 	for (mp = phys_install; mp != NULL; mp = mp->next)
1177 		npages += mp->size >> PAGESHIFT;
1178 
1179 	return (npages);
1180 }
1181 
1182 int
1183 dump_plat_addr()
1184 {
1185 #ifdef __xpv
1186 	pfn_t pfn = mmu_btop(xen_info->shared_info) | PFN_IS_FOREIGN_MFN;
1187 	mem_vtop_t mem_vtop;
1188 	int cnt;
1189 
1190 	/*
1191 	 * On the hypervisor, we want to dump the page with shared_info on it.
1192 	 */
1193 	if (!IN_XPV_PANIC()) {
1194 		mem_vtop.m_as = &kas;
1195 		mem_vtop.m_va = HYPERVISOR_shared_info;
1196 		mem_vtop.m_pfn = pfn;
1197 		dumpvp_write(&mem_vtop, sizeof (mem_vtop_t));
1198 		cnt = 1;
1199 	} else {
1200 		cnt = dump_xpv_addr();
1201 	}
1202 	return (cnt);
1203 #else
1204 	return (0);
1205 #endif
1206 }
1207 
1208 void
1209 dump_plat_pfn()
1210 {
1211 #ifdef __xpv
1212 	pfn_t pfn = mmu_btop(xen_info->shared_info) | PFN_IS_FOREIGN_MFN;
1213 
1214 	if (!IN_XPV_PANIC())
1215 		dumpvp_write(&pfn, sizeof (pfn));
1216 	else
1217 		dump_xpv_pfn();
1218 #endif
1219 }
1220 
1221 /*ARGSUSED*/
1222 int
1223 dump_plat_data(void *dump_cbuf)
1224 {
1225 #ifdef __xpv
1226 	uint32_t csize;
1227 	int cnt;
1228 
1229 	if (!IN_XPV_PANIC()) {
1230 		csize = (uint32_t)compress(HYPERVISOR_shared_info, dump_cbuf,
1231 		    PAGESIZE);
1232 		dumpvp_write(&csize, sizeof (uint32_t));
1233 		dumpvp_write(dump_cbuf, csize);
1234 		cnt = 1;
1235 	} else {
1236 		cnt = dump_xpv_data(dump_cbuf);
1237 	}
1238 	return (cnt);
1239 #else
1240 	return (0);
1241 #endif
1242 }
1243 
1244 /*
1245  * Calculates a linear address, given the CS selector and PC values
1246  * by looking up the %cs selector process's LDT or the CPU's GDT.
1247  * proc->p_ldtlock must be held across this call.
1248  */
1249 int
1250 linear_pc(struct regs *rp, proc_t *p, caddr_t *linearp)
1251 {
1252 	user_desc_t	*descrp;
1253 	caddr_t		baseaddr;
1254 	uint16_t	idx = SELTOIDX(rp->r_cs);
1255 
1256 	ASSERT(rp->r_cs <= 0xFFFF);
1257 	ASSERT(MUTEX_HELD(&p->p_ldtlock));
1258 
1259 	if (SELISLDT(rp->r_cs)) {
1260 		/*
1261 		 * Currently 64 bit processes cannot have private LDTs.
1262 		 */
1263 		ASSERT(p->p_model != DATAMODEL_LP64);
1264 
1265 		if (p->p_ldt == NULL)
1266 			return (-1);
1267 
1268 		descrp = &p->p_ldt[idx];
1269 		baseaddr = (caddr_t)(uintptr_t)USEGD_GETBASE(descrp);
1270 
1271 		/*
1272 		 * Calculate the linear address (wraparound is not only ok,
1273 		 * it's expected behavior).  The cast to uint32_t is because
1274 		 * LDT selectors are only allowed in 32-bit processes.
1275 		 */
1276 		*linearp = (caddr_t)(uintptr_t)(uint32_t)((uintptr_t)baseaddr +
1277 		    rp->r_pc);
1278 	} else {
1279 #ifdef DEBUG
1280 		descrp = &CPU->cpu_gdt[idx];
1281 		baseaddr = (caddr_t)(uintptr_t)USEGD_GETBASE(descrp);
1282 		/* GDT-based descriptors' base addresses should always be 0 */
1283 		ASSERT(baseaddr == 0);
1284 #endif
1285 		*linearp = (caddr_t)(uintptr_t)rp->r_pc;
1286 	}
1287 
1288 	return (0);
1289 }
1290 
1291 /*
1292  * The implementation of dtrace_linear_pc is similar to the that of
1293  * linear_pc, above, but here we acquire p_ldtlock before accessing
1294  * p_ldt.  This implementation is used by the pid provider; we prefix
1295  * it with "dtrace_" to avoid inducing spurious tracing events.
1296  */
1297 int
1298 dtrace_linear_pc(struct regs *rp, proc_t *p, caddr_t *linearp)
1299 {
1300 	user_desc_t	*descrp;
1301 	caddr_t		baseaddr;
1302 	uint16_t	idx = SELTOIDX(rp->r_cs);
1303 
1304 	ASSERT(rp->r_cs <= 0xFFFF);
1305 
1306 	if (SELISLDT(rp->r_cs)) {
1307 		/*
1308 		 * Currently 64 bit processes cannot have private LDTs.
1309 		 */
1310 		ASSERT(p->p_model != DATAMODEL_LP64);
1311 
1312 		mutex_enter(&p->p_ldtlock);
1313 		if (p->p_ldt == NULL) {
1314 			mutex_exit(&p->p_ldtlock);
1315 			return (-1);
1316 		}
1317 		descrp = &p->p_ldt[idx];
1318 		baseaddr = (caddr_t)(uintptr_t)USEGD_GETBASE(descrp);
1319 		mutex_exit(&p->p_ldtlock);
1320 
1321 		/*
1322 		 * Calculate the linear address (wraparound is not only ok,
1323 		 * it's expected behavior).  The cast to uint32_t is because
1324 		 * LDT selectors are only allowed in 32-bit processes.
1325 		 */
1326 		*linearp = (caddr_t)(uintptr_t)(uint32_t)((uintptr_t)baseaddr +
1327 		    rp->r_pc);
1328 	} else {
1329 #ifdef DEBUG
1330 		descrp = &CPU->cpu_gdt[idx];
1331 		baseaddr = (caddr_t)(uintptr_t)USEGD_GETBASE(descrp);
1332 		/* GDT-based descriptors' base addresses should always be 0 */
1333 		ASSERT(baseaddr == 0);
1334 #endif
1335 		*linearp = (caddr_t)(uintptr_t)rp->r_pc;
1336 	}
1337 
1338 	return (0);
1339 }
1340