xref: /illumos-gate/usr/src/uts/sun4v/os/mach_cpu_states.c (revision 71269a2275bf5a143dad6461eee2710a344e7261)
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  * Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #include <sys/types.h>
27 #include <sys/systm.h>
28 #include <sys/archsystm.h>
29 #include <sys/t_lock.h>
30 #include <sys/uadmin.h>
31 #include <sys/panic.h>
32 #include <sys/reboot.h>
33 #include <sys/autoconf.h>
34 #include <sys/machsystm.h>
35 #include <sys/promif.h>
36 #include <sys/membar.h>
37 #include <vm/hat_sfmmu.h>
38 #include <sys/cpu_module.h>
39 #include <sys/cpu_sgnblk_defs.h>
40 #include <sys/intreg.h>
41 #include <sys/consdev.h>
42 #include <sys/kdi_impl.h>
43 #include <sys/traptrace.h>
44 #include <sys/hypervisor_api.h>
45 #include <sys/vmsystm.h>
46 #include <sys/dtrace.h>
47 #include <sys/xc_impl.h>
48 #include <sys/callb.h>
49 #include <sys/mdesc.h>
50 #include <sys/mach_descrip.h>
51 #include <sys/wdt.h>
52 #include <sys/soft_state.h>
53 #include <sys/promimpl.h>
54 #include <sys/hsvc.h>
55 #include <sys/ldoms.h>
56 
57 /*
58  * hvdump_buf_va is a pointer to the currently-configured hvdump_buf.
59  * A value of NULL indicates that this area is not configured.
60  * hvdump_buf_sz is tunable but will be clamped to HVDUMP_SIZE_MAX.
61  */
62 
63 caddr_t hvdump_buf_va;
64 uint64_t hvdump_buf_sz = HVDUMP_SIZE_DEFAULT;
65 static uint64_t hvdump_buf_pa;
66 
67 u_longlong_t panic_tick;
68 
69 extern u_longlong_t gettick();
70 static void reboot_machine(char *);
71 static void update_hvdump_buffer(void);
72 
73 /*
74  * For xt_sync synchronization.
75  */
76 extern uint64_t xc_tick_limit;
77 extern uint64_t xc_tick_jump_limit;
78 extern uint64_t xc_sync_tick_limit;
79 
80 /*
81  * We keep our own copies, used for cache flushing, because we can be called
82  * before cpu_fiximpl().
83  */
84 static int kdi_dcache_size;
85 static int kdi_dcache_linesize;
86 static int kdi_icache_size;
87 static int kdi_icache_linesize;
88 
89 /*
90  * Assembly support for generic modules in sun4v/ml/mach_xc.s
91  */
92 extern void init_mondo_nocheck(xcfunc_t *func, uint64_t arg1, uint64_t arg2);
93 extern void kdi_flush_idcache(int, int, int, int);
94 extern uint64_t get_cpuaddr(uint64_t, uint64_t);
95 
96 
97 #define	BOOT_CMD_MAX_LEN	256
98 #define	BOOT_CMD_BASE		"boot "
99 
100 /*
101  * In an LDoms system we do not save the user's boot args in NVRAM
102  * as is done on legacy systems.  Instead, we format and send a
103  * 'reboot-command' variable to the variable service.  The contents
104  * of the variable are retrieved by OBP and used verbatim for
105  * the next boot.
106  */
107 static void
108 store_boot_cmd(char *args, boolean_t add_boot_str)
109 {
110 	static char	cmd_buf[BOOT_CMD_MAX_LEN];
111 	size_t		len = 1;
112 	pnode_t		node;
113 	size_t		base_len = 0;
114 	size_t		args_len;
115 	size_t		args_max;
116 
117 	if (add_boot_str) {
118 		(void) strcpy(cmd_buf, BOOT_CMD_BASE);
119 
120 		base_len = strlen(BOOT_CMD_BASE);
121 		len = base_len + 1;
122 	}
123 
124 	if (args != NULL) {
125 		args_len = strlen(args);
126 		args_max = BOOT_CMD_MAX_LEN - len;
127 
128 		if (args_len > args_max) {
129 			cmn_err(CE_WARN, "Reboot command too long (%ld), "
130 			    "truncating command arguments", len + args_len);
131 
132 			args_len = args_max;
133 		}
134 
135 		len += args_len;
136 		(void) strncpy(&cmd_buf[base_len], args, args_len);
137 	}
138 
139 	node = prom_optionsnode();
140 	if ((node == OBP_NONODE) || (node == OBP_BADNODE) ||
141 	    prom_setprop(node, "reboot-command", cmd_buf, len) == -1)
142 		cmn_err(CE_WARN, "Unable to store boot command for "
143 		    "use on reboot");
144 }
145 
146 
147 /*
148  * Machine dependent code to reboot.
149  *
150  * "bootstr", when non-null, points to a string to be used as the
151  * argument string when rebooting.
152  *
153  * "invoke_cb" is a boolean. It is set to true when mdboot() can safely
154  * invoke CB_CL_MDBOOT callbacks before shutting the system down, i.e. when
155  * we are in a normal shutdown sequence (interrupts are not blocked, the
156  * system is not panic'ing or being suspended).
157  */
158 /*ARGSUSED*/
159 void
160 mdboot(int cmd, int fcn, char *bootstr, boolean_t invoke_cb)
161 {
162 	extern void pm_cfb_check_and_powerup(void);
163 
164 	/*
165 	 * XXX - rconsvp is set to NULL to ensure that output messages
166 	 * are sent to the underlying "hardware" device using the
167 	 * monitor's printf routine since we are in the process of
168 	 * either rebooting or halting the machine.
169 	 */
170 	rconsvp = NULL;
171 
172 	switch (fcn) {
173 	case AD_HALT:
174 		/*
175 		 * LDoms: By storing a no-op command
176 		 * in the 'reboot-command' variable we cause OBP
177 		 * to ignore the setting of 'auto-boot?' after
178 		 * it completes the reset.  This causes the system
179 		 * to stop at the ok prompt.
180 		 */
181 		if (domaining_enabled() && invoke_cb)
182 			store_boot_cmd("noop", B_FALSE);
183 		break;
184 
185 	case AD_POWEROFF:
186 		break;
187 
188 	default:
189 		if (bootstr == NULL) {
190 			switch (fcn) {
191 
192 			case AD_BOOT:
193 				bootstr = "";
194 				break;
195 
196 			case AD_IBOOT:
197 				bootstr = "-a";
198 				break;
199 
200 			case AD_SBOOT:
201 				bootstr = "-s";
202 				break;
203 
204 			case AD_SIBOOT:
205 				bootstr = "-sa";
206 				break;
207 			default:
208 				cmn_err(CE_WARN,
209 				    "mdboot: invalid function %d", fcn);
210 				bootstr = "";
211 				break;
212 			}
213 		}
214 
215 		/*
216 		 * If LDoms is running, we must save the boot string
217 		 * before we enter restricted mode.  This is possible
218 		 * only if we are not being called from panic.
219 		 */
220 		if (domaining_enabled() && invoke_cb)
221 			store_boot_cmd(bootstr, B_TRUE);
222 	}
223 
224 	/*
225 	 * At a high interrupt level we can't:
226 	 *	1) bring up the console
227 	 * or
228 	 *	2) wait for pending interrupts prior to redistribution
229 	 *	   to the current CPU
230 	 *
231 	 * so we do them now.
232 	 */
233 	pm_cfb_check_and_powerup();
234 
235 	/* make sure there are no more changes to the device tree */
236 	devtree_freeze();
237 
238 	if (invoke_cb)
239 		(void) callb_execute_class(CB_CL_MDBOOT, NULL);
240 
241 	/*
242 	 * Clear any unresolved UEs from memory.
243 	 */
244 	page_retire_mdboot();
245 
246 	/*
247 	 * stop other cpus which also raise our priority. since there is only
248 	 * one active cpu after this, and our priority will be too high
249 	 * for us to be preempted, we're essentially single threaded
250 	 * from here on out.
251 	 */
252 	stop_other_cpus();
253 
254 	/*
255 	 * try and reset leaf devices.  reset_leaves() should only
256 	 * be called when there are no other threads that could be
257 	 * accessing devices
258 	 */
259 	reset_leaves();
260 
261 	watchdog_clear();
262 
263 	if (fcn == AD_HALT) {
264 		mach_set_soft_state(SIS_TRANSITION,
265 		    &SOLARIS_SOFT_STATE_HALT_MSG);
266 		halt((char *)NULL);
267 	} else if (fcn == AD_POWEROFF) {
268 		mach_set_soft_state(SIS_TRANSITION,
269 		    &SOLARIS_SOFT_STATE_POWER_MSG);
270 		power_down(NULL);
271 	} else {
272 		mach_set_soft_state(SIS_TRANSITION,
273 		    &SOLARIS_SOFT_STATE_REBOOT_MSG);
274 		reboot_machine(bootstr);
275 	}
276 	/* MAYBE REACHED */
277 }
278 
279 /* mdpreboot - may be called prior to mdboot while root fs still mounted */
280 /*ARGSUSED*/
281 void
282 mdpreboot(int cmd, int fcn, char *bootstr)
283 {
284 }
285 
286 /*
287  * Halt the machine and then reboot with the device
288  * and arguments specified in bootstr.
289  */
290 static void
291 reboot_machine(char *bootstr)
292 {
293 	flush_windows();
294 	stop_other_cpus();		/* send stop signal to other CPUs */
295 	prom_printf("rebooting...\n");
296 	/*
297 	 * For platforms that use CPU signatures, we
298 	 * need to set the signature block to OS and
299 	 * the state to exiting for all the processors.
300 	 */
301 	CPU_SIGNATURE(OS_SIG, SIGST_EXIT, SIGSUBST_REBOOT, -1);
302 	prom_reboot(bootstr);
303 	/*NOTREACHED*/
304 }
305 
306 /*
307  * We use the x-trap mechanism and idle_stop_xcall() to stop the other CPUs.
308  * Once in panic_idle() they raise spl, record their location, and spin.
309  */
310 static void
311 panic_idle(void)
312 {
313 	(void) spl7();
314 
315 	debug_flush_windows();
316 	(void) setjmp(&curthread->t_pcb);
317 
318 	CPU->cpu_m.in_prom = 1;
319 	membar_stld();
320 
321 	for (;;)
322 		;
323 }
324 
325 /*
326  * Force the other CPUs to trap into panic_idle(), and then remove them
327  * from the cpu_ready_set so they will no longer receive cross-calls.
328  */
329 /*ARGSUSED*/
330 void
331 panic_stopcpus(cpu_t *cp, kthread_t *t, int spl)
332 {
333 	cpuset_t cps;
334 	int i;
335 
336 	(void) splzs();
337 	CPUSET_ALL_BUT(cps, cp->cpu_id);
338 	xt_some(cps, (xcfunc_t *)idle_stop_xcall, (uint64_t)&panic_idle, NULL);
339 
340 	for (i = 0; i < NCPU; i++) {
341 		if (i != cp->cpu_id && CPU_XCALL_READY(i)) {
342 			int ntries = 0x10000;
343 
344 			while (!cpu[i]->cpu_m.in_prom && ntries) {
345 				DELAY(50);
346 				ntries--;
347 			}
348 
349 			if (!cpu[i]->cpu_m.in_prom)
350 				printf("panic: failed to stop cpu%d\n", i);
351 
352 			cpu[i]->cpu_flags &= ~CPU_READY;
353 			cpu[i]->cpu_flags |= CPU_QUIESCED;
354 			CPUSET_DEL(cpu_ready_set, cpu[i]->cpu_id);
355 		}
356 	}
357 }
358 
359 /*
360  * Platform callback following each entry to panicsys().  If we've panicked at
361  * level 14, we examine t_panic_trap to see if a fatal trap occurred.  If so,
362  * we disable further %tick_cmpr interrupts.  If not, an explicit call to panic
363  * was made and so we re-enqueue an interrupt request structure to allow
364  * further level 14 interrupts to be processed once we lower PIL.  This allows
365  * us to handle panics from the deadman() CY_HIGH_LEVEL cyclic.
366  */
367 void
368 panic_enter_hw(int spl)
369 {
370 	if (!panic_tick) {
371 		panic_tick = gettick();
372 		if (mach_htraptrace_enable) {
373 			uint64_t prev_freeze;
374 
375 			/*  there are no possible error codes for this hcall */
376 			(void) hv_ttrace_freeze((uint64_t)TRAP_TFREEZE_ALL,
377 			    &prev_freeze);
378 		}
379 #ifdef TRAPTRACE
380 		TRAPTRACE_FREEZE;
381 #endif
382 	}
383 
384 	mach_set_soft_state(SIS_TRANSITION, &SOLARIS_SOFT_STATE_PANIC_MSG);
385 
386 	if (spl == ipltospl(PIL_14)) {
387 		uint_t opstate = disable_vec_intr();
388 
389 		if (curthread->t_panic_trap != NULL) {
390 			tickcmpr_disable();
391 			intr_dequeue_req(PIL_14, cbe_level14_inum);
392 		} else {
393 			if (!tickcmpr_disabled())
394 				intr_enqueue_req(PIL_14, cbe_level14_inum);
395 			/*
396 			 * Clear SOFTINT<14>, SOFTINT<0> (TICK_INT)
397 			 * and SOFTINT<16> (STICK_INT) to indicate
398 			 * that the current level 14 has been serviced.
399 			 */
400 			wr_clr_softint((1 << PIL_14) |
401 			    TICK_INT_MASK | STICK_INT_MASK);
402 		}
403 
404 		enable_vec_intr(opstate);
405 	}
406 }
407 
408 /*
409  * Miscellaneous hardware-specific code to execute after panicstr is set
410  * by the panic code: we also print and record PTL1 panic information here.
411  */
412 /*ARGSUSED*/
413 void
414 panic_quiesce_hw(panic_data_t *pdp)
415 {
416 	extern uint_t getpstate(void);
417 	extern void setpstate(uint_t);
418 
419 	/*
420 	 * Turn off TRAPTRACE and save the current %tick value in panic_tick.
421 	 */
422 	if (!panic_tick) {
423 		panic_tick = gettick();
424 		if (mach_htraptrace_enable) {
425 			uint64_t prev_freeze;
426 
427 			/*  there are no possible error codes for this hcall */
428 			(void) hv_ttrace_freeze((uint64_t)TRAP_TFREEZE_ALL,
429 			    &prev_freeze);
430 		}
431 #ifdef TRAPTRACE
432 		TRAPTRACE_FREEZE;
433 #endif
434 	}
435 	/*
436 	 * For Platforms that use CPU signatures, we
437 	 * need to set the signature block to OS, the state to
438 	 * exiting, and the substate to panic for all the processors.
439 	 */
440 	CPU_SIGNATURE(OS_SIG, SIGST_EXIT, SIGSUBST_PANIC, -1);
441 
442 	update_hvdump_buffer();
443 
444 	/*
445 	 * Disable further ECC errors from the bus nexus.
446 	 */
447 	(void) bus_func_invoke(BF_TYPE_ERRDIS);
448 
449 	/*
450 	 * Redirect all interrupts to the current CPU.
451 	 */
452 	intr_redist_all_cpus_shutdown();
453 
454 	/*
455 	 * This call exists solely to support dumps to network
456 	 * devices after sync from OBP.
457 	 *
458 	 * If we came here via the sync callback, then on some
459 	 * platforms, interrupts may have arrived while we were
460 	 * stopped in OBP.  OBP will arrange for those interrupts to
461 	 * be redelivered if you say "go", but not if you invoke a
462 	 * client callback like 'sync'.	 For some dump devices
463 	 * (network swap devices), we need interrupts to be
464 	 * delivered in order to dump, so we have to call the bus
465 	 * nexus driver to reset the interrupt state machines.
466 	 */
467 	(void) bus_func_invoke(BF_TYPE_RESINTR);
468 
469 	setpstate(getpstate() | PSTATE_IE);
470 }
471 
472 /*
473  * Platforms that use CPU signatures need to set the signature block to OS and
474  * the state to exiting for all CPUs. PANIC_CONT indicates that we're about to
475  * write the crash dump, which tells the SSP/SMS to begin a timeout routine to
476  * reboot the machine if the dump never completes.
477  */
478 /*ARGSUSED*/
479 void
480 panic_dump_hw(int spl)
481 {
482 	CPU_SIGNATURE(OS_SIG, SIGST_EXIT, SIGSUBST_DUMP, -1);
483 }
484 
485 /*
486  * for ptl1_panic
487  */
488 void
489 ptl1_init_cpu(struct cpu *cpu)
490 {
491 	ptl1_state_t *pstate = &cpu->cpu_m.ptl1_state;
492 
493 	/*CONSTCOND*/
494 	if (sizeof (struct cpu) + PTL1_SSIZE > CPU_ALLOC_SIZE) {
495 		panic("ptl1_init_cpu: not enough space left for ptl1_panic "
496 		    "stack, sizeof (struct cpu) = %lu",
497 		    (unsigned long)sizeof (struct cpu));
498 	}
499 
500 	pstate->ptl1_stktop = (uintptr_t)cpu + CPU_ALLOC_SIZE;
501 	cpu_pa[cpu->cpu_id] = va_to_pa(cpu);
502 }
503 
504 void
505 ptl1_panic_handler(ptl1_state_t *pstate)
506 {
507 	static const char *ptl1_reasons[] = {
508 #ifdef	PTL1_PANIC_DEBUG
509 		"trap for debug purpose",	/* PTL1_BAD_DEBUG */
510 #else
511 		"unknown trap",			/* PTL1_BAD_DEBUG */
512 #endif
513 		"register window trap",		/* PTL1_BAD_WTRAP */
514 		"kernel MMU miss",		/* PTL1_BAD_KMISS */
515 		"kernel protection fault",	/* PTL1_BAD_KPROT_FAULT */
516 		"ISM MMU miss",			/* PTL1_BAD_ISM */
517 		"kernel MMU trap",		/* PTL1_BAD_MMUTRAP */
518 		"kernel trap handler state",	/* PTL1_BAD_TRAP */
519 		"floating point trap",		/* PTL1_BAD_FPTRAP */
520 #ifdef	DEBUG
521 		"pointer to intr_vec",		/* PTL1_BAD_INTR_VEC */
522 #else
523 		"unknown trap",			/* PTL1_BAD_INTR_VEC */
524 #endif
525 #ifdef	TRAPTRACE
526 		"TRACE_PTR state",		/* PTL1_BAD_TRACE_PTR */
527 #else
528 		"unknown trap",			/* PTL1_BAD_TRACE_PTR */
529 #endif
530 		"stack overflow",		/* PTL1_BAD_STACK */
531 		"DTrace flags",			/* PTL1_BAD_DTRACE_FLAGS */
532 		"attempt to steal locked ctx",  /* PTL1_BAD_CTX_STEAL */
533 		"CPU ECC error loop",		/* PTL1_BAD_ECC */
534 		"unexpected error from hypervisor call", /* PTL1_BAD_HCALL */
535 		"unexpected global level(%gl)", /* PTL1_BAD_GL */
536 		"Watchdog Reset", 		/* PTL1_BAD_WATCHDOG */
537 		"unexpected RED mode trap", 	/* PTL1_BAD_RED */
538 		"return value EINVAL from hcall: "\
539 		    "UNMAP_PERM_ADDR",	/* PTL1_BAD_HCALL_UNMAP_PERM_EINVAL */
540 		"return value ENOMAP from hcall: "\
541 		    "UNMAP_PERM_ADDR", /* PTL1_BAD_HCALL_UNMAP_PERM_ENOMAP */
542 		"error raising a TSB exception", /* PTL1_BAD_RAISE_TSBEXCP */
543 		"missing shared TSB"	/* PTL1_NO_SCDTSB8K */
544 	};
545 
546 	uint_t reason = pstate->ptl1_regs.ptl1_gregs[0].ptl1_g1;
547 	uint_t tl = pstate->ptl1_regs.ptl1_trap_regs[0].ptl1_tl;
548 	struct panic_trap_info ti = { 0 };
549 
550 	/*
551 	 * Use trap_info for a place holder to call panic_savetrap() and
552 	 * panic_showtrap() to save and print out ptl1_panic information.
553 	 */
554 	if (curthread->t_panic_trap == NULL)
555 		curthread->t_panic_trap = &ti;
556 
557 	if (reason < sizeof (ptl1_reasons) / sizeof (ptl1_reasons[0]))
558 		panic("bad %s at TL %u", ptl1_reasons[reason], tl);
559 	else
560 		panic("ptl1_panic reason 0x%x at TL %u", reason, tl);
561 }
562 
563 void
564 clear_watchdog_on_exit(void)
565 {
566 	if (watchdog_enabled && watchdog_activated) {
567 		prom_printf("Debugging requested; hardware watchdog "
568 		    "suspended.\n");
569 		(void) watchdog_suspend();
570 	}
571 }
572 
573 /*
574  * Restore the watchdog timer when returning from a debugger
575  * after a panic or L1-A and resume watchdog pat.
576  */
577 void
578 restore_watchdog_on_entry()
579 {
580 	watchdog_resume();
581 }
582 
583 int
584 kdi_watchdog_disable(void)
585 {
586 	watchdog_suspend();
587 
588 	return (0);
589 }
590 
591 void
592 kdi_watchdog_restore(void)
593 {
594 	watchdog_resume();
595 }
596 
597 void
598 mach_dump_buffer_init(void)
599 {
600 	uint64_t  ret, minsize = 0;
601 
602 	if (hvdump_buf_sz > HVDUMP_SIZE_MAX)
603 		hvdump_buf_sz = HVDUMP_SIZE_MAX;
604 
605 	hvdump_buf_va = contig_mem_alloc_align(hvdump_buf_sz, PAGESIZE);
606 	if (hvdump_buf_va == NULL)
607 		return;
608 
609 	hvdump_buf_pa = va_to_pa(hvdump_buf_va);
610 
611 	ret = hv_dump_buf_update(hvdump_buf_pa, hvdump_buf_sz,
612 	    &minsize);
613 
614 	if (ret != H_EOK) {
615 		contig_mem_free(hvdump_buf_va, hvdump_buf_sz);
616 		hvdump_buf_va = NULL;
617 		cmn_err(CE_NOTE, "!Error in setting up hvstate"
618 		    "dump buffer. Error = 0x%lx, size = 0x%lx,"
619 		    "buf_pa = 0x%lx", ret, hvdump_buf_sz,
620 		    hvdump_buf_pa);
621 
622 		if (ret == H_EINVAL) {
623 			cmn_err(CE_NOTE, "!Buffer size too small."
624 			    "Available buffer size = 0x%lx,"
625 			    "Minimum buffer size required = 0x%lx",
626 			    hvdump_buf_sz, minsize);
627 		}
628 	}
629 }
630 
631 
632 static void
633 update_hvdump_buffer(void)
634 {
635 	uint64_t ret, dummy_val;
636 
637 	if (hvdump_buf_va == NULL)
638 		return;
639 
640 	ret = hv_dump_buf_update(hvdump_buf_pa, hvdump_buf_sz,
641 	    &dummy_val);
642 	if (ret != H_EOK) {
643 		cmn_err(CE_NOTE, "!Cannot update hvstate dump"
644 		    "buffer. Error = 0x%lx", ret);
645 	}
646 }
647 
648 
649 static int
650 getintprop(pnode_t node, char *name, int deflt)
651 {
652 	int	value;
653 
654 	switch (prom_getproplen(node, name)) {
655 	case 0:
656 		value = 1;	/* boolean properties */
657 		break;
658 
659 	case sizeof (int):
660 		(void) prom_getprop(node, name, (caddr_t)&value);
661 		break;
662 
663 	default:
664 		value = deflt;
665 		break;
666 	}
667 
668 	return (value);
669 }
670 
671 /*
672  * Called by setcpudelay
673  */
674 void
675 cpu_init_tick_freq(void)
676 {
677 	md_t *mdp;
678 	mde_cookie_t rootnode;
679 	int		listsz;
680 	mde_cookie_t	*listp = NULL;
681 	int	num_nodes;
682 	uint64_t stick_prop;
683 
684 	if (broken_md_flag) {
685 		sys_tick_freq = cpunodes[CPU->cpu_id].clock_freq;
686 		return;
687 	}
688 
689 	if ((mdp = md_get_handle()) == NULL)
690 		panic("stick_frequency property not found in MD");
691 
692 	rootnode = md_root_node(mdp);
693 	ASSERT(rootnode != MDE_INVAL_ELEM_COOKIE);
694 
695 	num_nodes = md_node_count(mdp);
696 
697 	ASSERT(num_nodes > 0);
698 	listsz = num_nodes * sizeof (mde_cookie_t);
699 	listp = (mde_cookie_t *)prom_alloc((caddr_t)0, listsz, 0);
700 
701 	if (listp == NULL)
702 		panic("cannot allocate list for MD properties");
703 
704 	num_nodes = md_scan_dag(mdp, rootnode, md_find_name(mdp, "platform"),
705 	    md_find_name(mdp, "fwd"), listp);
706 
707 	ASSERT(num_nodes == 1);
708 
709 	if (md_get_prop_val(mdp, *listp, "stick-frequency", &stick_prop) != 0)
710 		panic("stick_frequency property not found in MD");
711 
712 	sys_tick_freq = stick_prop;
713 
714 	prom_free((caddr_t)listp, listsz);
715 	(void) md_fini_handle(mdp);
716 }
717 
718 int shipit(int n, uint64_t cpu_list_ra);
719 
720 #ifdef DEBUG
721 #define	SEND_MONDO_STATS	1
722 #endif
723 
724 #ifdef SEND_MONDO_STATS
725 uint32_t x_one_stimes[64];
726 uint32_t x_one_ltimes[16];
727 uint32_t x_set_stimes[64];
728 uint32_t x_set_ltimes[16];
729 uint32_t x_set_cpus[NCPU];
730 #endif
731 
732 void
733 send_one_mondo(int cpuid)
734 {
735 	int retries, stat;
736 	uint64_t starttick, endtick, tick, lasttick;
737 	struct machcpu	*mcpup = &(CPU->cpu_m);
738 
739 	CPU_STATS_ADDQ(CPU, sys, xcalls, 1);
740 	starttick = lasttick = gettick();
741 	mcpup->cpu_list[0] = (uint16_t)cpuid;
742 	stat = shipit(1, mcpup->cpu_list_ra);
743 	endtick = starttick + xc_tick_limit;
744 	retries = 0;
745 	while (stat != H_EOK) {
746 		if (stat != H_EWOULDBLOCK) {
747 			if (panic_quiesce)
748 				return;
749 			if (stat == H_ECPUERROR)
750 				cmn_err(CE_PANIC, "send_one_mondo: "
751 				    "cpuid: 0x%x has been marked in "
752 				    "error", cpuid);
753 			else
754 				cmn_err(CE_PANIC, "send_one_mondo: "
755 				    "unexpected hypervisor error 0x%x "
756 				    "while sending a mondo to cpuid: "
757 				    "0x%x", stat, cpuid);
758 		}
759 		tick = gettick();
760 		/*
761 		 * If there is a big jump between the current tick
762 		 * count and lasttick, we have probably hit a break
763 		 * point.  Adjust endtick accordingly to avoid panic.
764 		 */
765 		if (tick > (lasttick + xc_tick_jump_limit))
766 			endtick += (tick - lasttick);
767 		lasttick = tick;
768 		if (tick > endtick) {
769 			if (panic_quiesce)
770 				return;
771 			cmn_err(CE_PANIC, "send mondo timeout "
772 			    "(target 0x%x) [retries: 0x%x hvstat: 0x%x]",
773 			    cpuid, retries, stat);
774 		}
775 		drv_usecwait(1);
776 		stat = shipit(1, mcpup->cpu_list_ra);
777 		retries++;
778 	}
779 #ifdef SEND_MONDO_STATS
780 	{
781 		uint64_t n = gettick() - starttick;
782 		if (n < 8192)
783 			x_one_stimes[n >> 7]++;
784 		else if (n < 15*8192)
785 			x_one_ltimes[n >> 13]++;
786 		else
787 			x_one_ltimes[0xf]++;
788 	}
789 #endif
790 }
791 
792 void
793 send_mondo_set(cpuset_t set)
794 {
795 	uint64_t starttick, endtick, tick, lasttick;
796 	uint_t largestid, smallestid;
797 	int i, j;
798 	int ncpuids = 0;
799 	int shipped = 0;
800 	int retries = 0;
801 	struct machcpu	*mcpup = &(CPU->cpu_m);
802 
803 	ASSERT(!CPUSET_ISNULL(set));
804 	CPUSET_BOUNDS(set, smallestid, largestid);
805 	if (smallestid == CPUSET_NOTINSET) {
806 		return;
807 	}
808 
809 	starttick = lasttick = gettick();
810 	endtick = starttick + xc_tick_limit;
811 
812 	/*
813 	 * Assemble CPU list for HV argument. We already know
814 	 * smallestid and largestid are members of set.
815 	 */
816 	mcpup->cpu_list[ncpuids++] = (uint16_t)smallestid;
817 	if (largestid != smallestid) {
818 		for (i = smallestid+1; i <= largestid-1; i++) {
819 			if (CPU_IN_SET(set, i)) {
820 				mcpup->cpu_list[ncpuids++] = (uint16_t)i;
821 			}
822 		}
823 		mcpup->cpu_list[ncpuids++] = (uint16_t)largestid;
824 	}
825 
826 	do {
827 		int stat;
828 
829 		stat = shipit(ncpuids, mcpup->cpu_list_ra);
830 		if (stat == H_EOK) {
831 			shipped += ncpuids;
832 			break;
833 		}
834 
835 		/*
836 		 * Either not all CPU mondos were sent, or an
837 		 * error occurred. CPUs that were sent mondos
838 		 * have their CPU IDs overwritten in cpu_list.
839 		 * Reset cpu_list so that it only holds those
840 		 * CPU IDs that still need to be sent.
841 		 */
842 		for (i = 0, j = 0; i < ncpuids; i++) {
843 			if (mcpup->cpu_list[i] == HV_SEND_MONDO_ENTRYDONE) {
844 				shipped++;
845 			} else {
846 				mcpup->cpu_list[j++] = mcpup->cpu_list[i];
847 			}
848 		}
849 		ncpuids = j;
850 
851 		/*
852 		 * Now handle possible errors returned
853 		 * from hypervisor.
854 		 */
855 		if (stat == H_ECPUERROR) {
856 			int errorcpus;
857 
858 			if (!panic_quiesce)
859 				cmn_err(CE_CONT, "send_mondo_set: cpuid(s) ");
860 
861 			/*
862 			 * Remove any CPUs in the error state from
863 			 * cpu_list. At this point cpu_list only
864 			 * contains the CPU IDs for mondos not
865 			 * succesfully sent.
866 			 */
867 			for (i = 0, errorcpus = 0; i < ncpuids; i++) {
868 				uint64_t state = CPU_STATE_INVALID;
869 				uint16_t id = mcpup->cpu_list[i];
870 
871 				(void) hv_cpu_state(id, &state);
872 				if (state == CPU_STATE_ERROR) {
873 					if (!panic_quiesce)
874 						cmn_err(CE_CONT, "0x%x ", id);
875 					errorcpus++;
876 				} else if (errorcpus > 0) {
877 					mcpup->cpu_list[i - errorcpus] =
878 					    mcpup->cpu_list[i];
879 				}
880 			}
881 			ncpuids -= errorcpus;
882 
883 			if (!panic_quiesce) {
884 				if (errorcpus == 0) {
885 					cmn_err(CE_CONT, "<none> have been "
886 					    "marked in error\n");
887 					cmn_err(CE_PANIC, "send_mondo_set: "
888 					    "hypervisor returned "
889 					    "H_ECPUERROR but no CPU in "
890 					    "cpu_list in error state");
891 				} else {
892 					cmn_err(CE_CONT, "have been marked in "
893 					    "error\n");
894 					cmn_err(CE_PANIC, "send_mondo_set: "
895 					    "CPU(s) in error state");
896 				}
897 			}
898 		} else if (stat != H_EWOULDBLOCK) {
899 			if (panic_quiesce)
900 				return;
901 			/*
902 			 * For all other errors, panic.
903 			 */
904 			cmn_err(CE_CONT, "send_mondo_set: unexpected "
905 			    "hypervisor error 0x%x while sending a "
906 			    "mondo to cpuid(s):", stat);
907 			for (i = 0; i < ncpuids; i++) {
908 				cmn_err(CE_CONT, " 0x%x", mcpup->cpu_list[i]);
909 			}
910 			cmn_err(CE_CONT, "\n");
911 			cmn_err(CE_PANIC, "send_mondo_set: unexpected "
912 			    "hypervisor error");
913 		}
914 
915 		tick = gettick();
916 		/*
917 		 * If there is a big jump between the current tick
918 		 * count and lasttick, we have probably hit a break
919 		 * point.  Adjust endtick accordingly to avoid panic.
920 		 */
921 		if (tick > (lasttick + xc_tick_jump_limit))
922 			endtick += (tick - lasttick);
923 		lasttick = tick;
924 		if (tick > endtick) {
925 			if (panic_quiesce)
926 				return;
927 			cmn_err(CE_CONT, "send mondo timeout "
928 			    "[retries: 0x%x]  cpuids: ", retries);
929 			for (i = 0; i < ncpuids; i++)
930 				cmn_err(CE_CONT, " 0x%x", mcpup->cpu_list[i]);
931 			cmn_err(CE_CONT, "\n");
932 			cmn_err(CE_PANIC, "send_mondo_set: timeout");
933 		}
934 
935 		while (gettick() < (tick + sys_clock_mhz))
936 			;
937 		retries++;
938 	} while (ncpuids > 0);
939 
940 	CPU_STATS_ADDQ(CPU, sys, xcalls, shipped);
941 
942 #ifdef SEND_MONDO_STATS
943 	{
944 		uint64_t n = gettick() - starttick;
945 		if (n < 8192)
946 			x_set_stimes[n >> 7]++;
947 		else if (n < 15*8192)
948 			x_set_ltimes[n >> 13]++;
949 		else
950 			x_set_ltimes[0xf]++;
951 	}
952 	x_set_cpus[shipped]++;
953 #endif
954 }
955 
956 void
957 syncfpu(void)
958 {
959 }
960 
961 void
962 sticksync_slave(void)
963 {}
964 
965 void
966 sticksync_master(void)
967 {}
968 
969 void
970 cpu_init_cache_scrub(void)
971 {
972 	mach_set_soft_state(SIS_NORMAL, &SOLARIS_SOFT_STATE_RUN_MSG);
973 }
974 
975 int
976 dtrace_blksuword32_err(uintptr_t addr, uint32_t *data)
977 {
978 	int ret, watched;
979 
980 	watched = watch_disable_addr((void *)addr, 4, S_WRITE);
981 	ret = dtrace_blksuword32(addr, data, 0);
982 	if (watched)
983 		watch_enable_addr((void *)addr, 4, S_WRITE);
984 
985 	return (ret);
986 }
987 
988 int
989 dtrace_blksuword32(uintptr_t addr, uint32_t *data, int tryagain)
990 {
991 	if (suword32((void *)addr, *data) == -1)
992 		return (tryagain ? dtrace_blksuword32_err(addr, data) : -1);
993 	dtrace_flush_sec(addr);
994 
995 	return (0);
996 }
997 
998 /*ARGSUSED*/
999 void
1000 cpu_faulted_enter(struct cpu *cp)
1001 {
1002 }
1003 
1004 /*ARGSUSED*/
1005 void
1006 cpu_faulted_exit(struct cpu *cp)
1007 {
1008 }
1009 
1010 static int
1011 kdi_cpu_ready_iter(int (*cb)(int, void *), void *arg)
1012 {
1013 	int rc, i;
1014 
1015 	for (rc = 0, i = 0; i < NCPU; i++) {
1016 		if (CPU_IN_SET(cpu_ready_set, i))
1017 			rc += cb(i, arg);
1018 	}
1019 
1020 	return (rc);
1021 }
1022 
1023 /*
1024  * Sends a cross-call to a specified processor.  The caller assumes
1025  * responsibility for repetition of cross-calls, as appropriate (MARSA for
1026  * debugging).
1027  */
1028 static int
1029 kdi_xc_one(int cpuid, void (*func)(uintptr_t, uintptr_t), uintptr_t arg1,
1030     uintptr_t arg2)
1031 {
1032 	int stat;
1033 	struct machcpu	*mcpup;
1034 	uint64_t cpuaddr_reg = 0, cpuaddr_scr = 0;
1035 
1036 	mcpup = &(((cpu_t *)get_cpuaddr(cpuaddr_reg, cpuaddr_scr))->cpu_m);
1037 
1038 	/*
1039 	 * if (idsr_busy())
1040 	 *	return (KDI_XC_RES_ERR);
1041 	 */
1042 
1043 	init_mondo_nocheck((xcfunc_t *)func, arg1, arg2);
1044 
1045 	mcpup->cpu_list[0] = (uint16_t)cpuid;
1046 	stat = shipit(1, mcpup->cpu_list_ra);
1047 
1048 	if (stat == 0)
1049 		return (KDI_XC_RES_OK);
1050 	else
1051 		return (KDI_XC_RES_NACK);
1052 }
1053 
1054 static void
1055 kdi_tickwait(clock_t nticks)
1056 {
1057 	clock_t endtick = gettick() + nticks;
1058 
1059 	while (gettick() < endtick)
1060 		;
1061 }
1062 
1063 static void
1064 kdi_cpu_init(int dcache_size, int dcache_linesize, int icache_size,
1065     int icache_linesize)
1066 {
1067 	kdi_dcache_size = dcache_size;
1068 	kdi_dcache_linesize = dcache_linesize;
1069 	kdi_icache_size = icache_size;
1070 	kdi_icache_linesize = icache_linesize;
1071 }
1072 
1073 /* used directly by kdi_read/write_phys */
1074 void
1075 kdi_flush_caches(void)
1076 {
1077 	/*
1078 	 * May not be implemented by all sun4v architectures.
1079 	 *
1080 	 * Cannot use hsvc_version to see if the group is already
1081 	 * negotiated or not because, this function is called by
1082 	 * KMDB when it is at the console prompt which is running
1083 	 * at highest PIL. hsvc_version grabs an adaptive mutex and
1084 	 * this is a no-no at this PIL level.
1085 	 */
1086 	if (hsvc_kdi_mem_iflush_negotiated) {
1087 		uint64_t	status = hv_mem_iflush_all();
1088 		if (status != H_EOK)
1089 			cmn_err(CE_PANIC, "Flushing all I$ entries failed");
1090 	}
1091 }
1092 
1093 /*ARGSUSED*/
1094 int
1095 kdi_get_stick(uint64_t *stickp)
1096 {
1097 	return (-1);
1098 }
1099 
1100 void
1101 cpu_kdi_init(kdi_t *kdi)
1102 {
1103 	/*
1104 	 * Any API negotiation this early in the boot will be unsuccessful.
1105 	 * Therefore firmware for Sun4v platforms that have incoherent I$ are
1106 	 * assumed to support pre-negotiated MEM_IFLUSH APIs. Successful
1107 	 * invokation the MEM_IFLUSH_ALL is a test for is availability.
1108 	 * Set a flag if successful indicating its availabitlity.
1109 	 */
1110 	if (hv_mem_iflush_all() == 0)
1111 		hsvc_kdi_mem_iflush_negotiated = B_TRUE;
1112 
1113 	kdi->kdi_flush_caches = kdi_flush_caches;
1114 	kdi->mkdi_cpu_init = kdi_cpu_init;
1115 	kdi->mkdi_cpu_ready_iter = kdi_cpu_ready_iter;
1116 	kdi->mkdi_xc_one = kdi_xc_one;
1117 	kdi->mkdi_tickwait = kdi_tickwait;
1118 	kdi->mkdi_get_stick = kdi_get_stick;
1119 }
1120 
1121 uint64_t	soft_state_message_ra[SOLARIS_SOFT_STATE_MSG_CNT];
1122 static uint64_t	soft_state_saved_state = (uint64_t)-1;
1123 static int	soft_state_initialized = 0;
1124 static uint64_t soft_state_sup_minor;		/* Supported minor number */
1125 static hsvc_info_t soft_state_hsvc = {
1126 			HSVC_REV_1, NULL, HSVC_GROUP_SOFT_STATE, 1, 0, NULL };
1127 
1128 
1129 static void
1130 sun4v_system_claim(void)
1131 {
1132 	watchdog_suspend();
1133 	/*
1134 	 * For "mdb -K", set soft state to debugging
1135 	 */
1136 	if (soft_state_saved_state == -1) {
1137 		mach_get_soft_state(&soft_state_saved_state,
1138 		    &SOLARIS_SOFT_STATE_SAVED_MSG);
1139 	}
1140 	/*
1141 	 * check again as the read above may or may not have worked and if
1142 	 * it didn't then soft state will still be -1
1143 	 */
1144 	if (soft_state_saved_state != -1) {
1145 		mach_set_soft_state(SIS_TRANSITION,
1146 		    &SOLARIS_SOFT_STATE_DEBUG_MSG);
1147 	}
1148 }
1149 
1150 static void
1151 sun4v_system_release(void)
1152 {
1153 	watchdog_resume();
1154 	/*
1155 	 * For "mdb -K", set soft_state state back to original state on exit
1156 	 */
1157 	if (soft_state_saved_state != -1) {
1158 		mach_set_soft_state(soft_state_saved_state,
1159 		    &SOLARIS_SOFT_STATE_SAVED_MSG);
1160 		soft_state_saved_state = -1;
1161 	}
1162 }
1163 
1164 void
1165 plat_kdi_init(kdi_t *kdi)
1166 {
1167 	kdi->pkdi_system_claim = sun4v_system_claim;
1168 	kdi->pkdi_system_release = sun4v_system_release;
1169 }
1170 
1171 /*
1172  * Routine to return memory information associated
1173  * with a physical address and syndrome.
1174  */
1175 /* ARGSUSED */
1176 int
1177 cpu_get_mem_info(uint64_t synd, uint64_t afar,
1178     uint64_t *mem_sizep, uint64_t *seg_sizep, uint64_t *bank_sizep,
1179     int *segsp, int *banksp, int *mcidp)
1180 {
1181 	return (ENOTSUP);
1182 }
1183 
1184 /*
1185  * This routine returns the size of the kernel's FRU name buffer.
1186  */
1187 size_t
1188 cpu_get_name_bufsize()
1189 {
1190 	return (UNUM_NAMLEN);
1191 }
1192 
1193 /*
1194  * This routine is a more generic interface to cpu_get_mem_unum(),
1195  * that may be used by other modules (e.g. mm).
1196  */
1197 /* ARGSUSED */
1198 int
1199 cpu_get_mem_name(uint64_t synd, uint64_t *afsr, uint64_t afar,
1200     char *buf, int buflen, int *lenp)
1201 {
1202 	return (ENOTSUP);
1203 }
1204 
1205 /* ARGSUSED */
1206 int
1207 cpu_get_mem_sid(char *unum, char *buf, int buflen, int *lenp)
1208 {
1209 	return (ENOTSUP);
1210 }
1211 
1212 /* ARGSUSED */
1213 int
1214 cpu_get_mem_addr(char *unum, char *sid, uint64_t offset, uint64_t *addrp)
1215 {
1216 	return (ENOTSUP);
1217 }
1218 
1219 /*
1220  * xt_sync - wait for previous x-traps to finish
1221  */
1222 void
1223 xt_sync(cpuset_t cpuset)
1224 {
1225 	union {
1226 		uint8_t volatile byte[NCPU];
1227 		uint64_t volatile xword[NCPU / 8];
1228 	} cpu_sync;
1229 	uint64_t starttick, endtick, tick, lasttick, traptrace_id;
1230 	uint_t largestid, smallestid;
1231 	int i, j;
1232 
1233 	kpreempt_disable();
1234 	CPUSET_DEL(cpuset, CPU->cpu_id);
1235 	CPUSET_AND(cpuset, cpu_ready_set);
1236 
1237 	CPUSET_BOUNDS(cpuset, smallestid, largestid);
1238 	if (smallestid == CPUSET_NOTINSET)
1239 		goto out;
1240 
1241 	/*
1242 	 * Sun4v uses a queue for receiving mondos. Successful
1243 	 * transmission of a mondo only indicates that the mondo
1244 	 * has been written into the queue.
1245 	 *
1246 	 * We use an array of bytes to let each cpu to signal back
1247 	 * to the cross trap sender that the cross trap has been
1248 	 * executed. Set the byte to 1 before sending the cross trap
1249 	 * and wait until other cpus reset it to 0.
1250 	 */
1251 	bzero((void *)&cpu_sync, NCPU);
1252 	cpu_sync.byte[smallestid] = 1;
1253 	if (largestid != smallestid) {
1254 		for (i = (smallestid + 1); i <= (largestid - 1); i++)
1255 			if (CPU_IN_SET(cpuset, i))
1256 				cpu_sync.byte[i] = 1;
1257 		cpu_sync.byte[largestid] = 1;
1258 	}
1259 
1260 	/*
1261 	 * To help debug xt_sync panic, each mondo is uniquely identified
1262 	 * by passing the tick value, traptrace_id as the second mondo
1263 	 * argument to xt_some which is logged in CPU's mondo queue,
1264 	 * traptrace buffer and the panic message.
1265 	 */
1266 	traptrace_id = gettick();
1267 	xt_some(cpuset, (xcfunc_t *)xt_sync_tl1,
1268 	    (uint64_t)cpu_sync.byte, traptrace_id);
1269 
1270 	starttick = lasttick = gettick();
1271 	endtick = starttick + xc_sync_tick_limit;
1272 
1273 	for (i = (smallestid / 8); i <= (largestid / 8); i++) {
1274 		while (cpu_sync.xword[i] != 0) {
1275 			tick = gettick();
1276 			/*
1277 			 * If there is a big jump between the current tick
1278 			 * count and lasttick, we have probably hit a break
1279 			 * point. Adjust endtick accordingly to avoid panic.
1280 			 */
1281 			if (tick > (lasttick + xc_tick_jump_limit)) {
1282 				endtick += (tick - lasttick);
1283 			}
1284 			lasttick = tick;
1285 			if (tick > endtick) {
1286 				if (panic_quiesce)
1287 					goto out;
1288 				cmn_err(CE_CONT, "Cross trap sync timeout:  "
1289 				    "at cpu_sync.xword[%d]: 0x%lx "
1290 				    "cpu_sync.byte: 0x%lx "
1291 				    "starttick: 0x%lx endtick: 0x%lx "
1292 				    "traptrace_id = 0x%lx\n",
1293 				    i, cpu_sync.xword[i],
1294 				    (uint64_t)cpu_sync.byte,
1295 				    starttick, endtick, traptrace_id);
1296 				cmn_err(CE_CONT, "CPUIDs:");
1297 				for (j = (i * 8); j <= largestid; j++) {
1298 					if (cpu_sync.byte[j] != 0)
1299 						cmn_err(CE_CONT, " 0x%x", j);
1300 				}
1301 				cmn_err(CE_PANIC, "xt_sync: timeout");
1302 			}
1303 		}
1304 	}
1305 
1306 out:
1307 	kpreempt_enable();
1308 }
1309 
1310 #define	QFACTOR		200
1311 /*
1312  * Recalculate the values of the cross-call timeout variables based
1313  * on the value of the 'inter-cpu-latency' property of the platform node.
1314  * The property sets the number of nanosec to wait for a cross-call
1315  * to be acknowledged.  Other timeout variables are derived from it.
1316  *
1317  * N.B. This implementation is aware of the internals of xc_init()
1318  * and updates many of the same variables.
1319  */
1320 void
1321 recalc_xc_timeouts(void)
1322 {
1323 	typedef union {
1324 		uint64_t whole;
1325 		struct {
1326 			uint_t high;
1327 			uint_t low;
1328 		} half;
1329 	} u_number;
1330 
1331 	/* See x_call.c for descriptions of these extern variables. */
1332 	extern uint64_t xc_tick_limit_scale;
1333 	extern uint64_t xc_mondo_time_limit;
1334 	extern uint64_t xc_func_time_limit;
1335 	extern uint64_t xc_scale;
1336 	extern uint64_t xc_mondo_multiplier;
1337 	extern uint_t   nsec_shift;
1338 
1339 	/* Temp versions of the target variables */
1340 	uint64_t tick_limit;
1341 	uint64_t tick_jump_limit;
1342 	uint64_t mondo_time_limit;
1343 	uint64_t func_time_limit;
1344 	uint64_t scale;
1345 
1346 	uint64_t latency;	/* nanoseconds */
1347 	uint64_t maxfreq;
1348 	uint64_t tick_limit_save = xc_tick_limit;
1349 	uint64_t sync_tick_limit_save = xc_sync_tick_limit;
1350 	uint_t   tick_scale;
1351 	uint64_t top;
1352 	uint64_t bottom;
1353 	u_number tk;
1354 
1355 	md_t *mdp;
1356 	int nrnode;
1357 	mde_cookie_t *platlist;
1358 
1359 	/*
1360 	 * Look up the 'inter-cpu-latency' (optional) property in the
1361 	 * platform node of the MD.  The units are nanoseconds.
1362 	 */
1363 	if ((mdp = md_get_handle()) == NULL) {
1364 		cmn_err(CE_WARN, "recalc_xc_timeouts: "
1365 		    "Unable to initialize machine description");
1366 		return;
1367 	}
1368 
1369 	nrnode = md_alloc_scan_dag(mdp,
1370 	    md_root_node(mdp), "platform", "fwd", &platlist);
1371 
1372 	ASSERT(nrnode == 1);
1373 	if (nrnode < 1) {
1374 		cmn_err(CE_WARN, "recalc_xc_timeouts: platform node missing");
1375 		goto done;
1376 	}
1377 	if (md_get_prop_val(mdp, platlist[0],
1378 	    "inter-cpu-latency", &latency) == -1)
1379 		goto done;
1380 
1381 	/*
1382 	 * clock.h defines an assembly-language macro
1383 	 * (NATIVE_TIME_TO_NSEC_SCALE) to convert from %stick
1384 	 * units to nanoseconds.  Since the inter-cpu-latency
1385 	 * units are nanoseconds and the xc_* variables require
1386 	 * %stick units, we need the inverse of that function.
1387 	 * The trick is to perform the calculation without
1388 	 * floating point, but also without integer truncation
1389 	 * or overflow.  To understand the calculation below,
1390 	 * please read the discussion of the macro in clock.h.
1391 	 * Since this new code will be invoked infrequently,
1392 	 * we can afford to implement it in C.
1393 	 *
1394 	 * tick_scale is the reciprocal of nsec_scale which is
1395 	 * calculated at startup in setcpudelay().  The calc
1396 	 * of tick_limit parallels that of NATIVE_TIME_TO_NSEC_SCALE
1397 	 * except we use tick_scale instead of nsec_scale and
1398 	 * C instead of assembler.
1399 	 */
1400 	tick_scale = (uint_t)(((u_longlong_t)sys_tick_freq
1401 	    << (32 - nsec_shift)) / NANOSEC);
1402 
1403 	tk.whole = latency;
1404 	top = ((uint64_t)tk.half.high << 4) * tick_scale;
1405 	bottom = (((uint64_t)tk.half.low << 4) * (uint64_t)tick_scale) >> 32;
1406 	tick_limit = top + bottom;
1407 
1408 	/*
1409 	 * xc_init() calculated 'maxfreq' by looking at all the cpus,
1410 	 * and used it to derive some of the timeout variables that we
1411 	 * recalculate below.  We can back into the original value by
1412 	 * using the inverse of one of those calculations.
1413 	 */
1414 	maxfreq = xc_mondo_time_limit / xc_scale;
1415 
1416 	/*
1417 	 * Don't allow the new timeout (xc_tick_limit) to fall below
1418 	 * the system tick frequency (stick).  Allowing the timeout
1419 	 * to be set more tightly than this empirically determined
1420 	 * value may cause panics.
1421 	 */
1422 	tick_limit = tick_limit < sys_tick_freq ? sys_tick_freq : tick_limit;
1423 
1424 	tick_jump_limit = tick_limit / 32;
1425 	tick_limit *= xc_tick_limit_scale;
1426 
1427 	/*
1428 	 * Recalculate xc_scale since it is used in a callback function
1429 	 * (xc_func_timeout_adj) to adjust two of the timeouts dynamically.
1430 	 * Make the change in xc_scale proportional to the change in
1431 	 * xc_tick_limit.
1432 	 */
1433 	scale = (xc_scale * tick_limit + sys_tick_freq / 2) / tick_limit_save;
1434 	if (scale == 0)
1435 		scale = 1;
1436 
1437 	mondo_time_limit = maxfreq * scale;
1438 	func_time_limit = mondo_time_limit * xc_mondo_multiplier;
1439 
1440 	/*
1441 	 * Don't modify the timeouts if nothing has changed.  Else,
1442 	 * stuff the variables with the freshly calculated (temp)
1443 	 * variables.  This minimizes the window where the set of
1444 	 * values could be inconsistent.
1445 	 */
1446 	if (tick_limit != xc_tick_limit) {
1447 		xc_tick_limit = tick_limit;
1448 		xc_tick_jump_limit = tick_jump_limit;
1449 		xc_scale = scale;
1450 		xc_mondo_time_limit = mondo_time_limit;
1451 		xc_func_time_limit = func_time_limit;
1452 	}
1453 
1454 done:
1455 	/*
1456 	 * Increase the timeout limit for xt_sync() cross calls.
1457 	 */
1458 	xc_sync_tick_limit = xc_tick_limit * (cpu_q_entries / QFACTOR);
1459 	xc_sync_tick_limit = xc_sync_tick_limit < xc_tick_limit ?
1460 	    xc_tick_limit : xc_sync_tick_limit;
1461 
1462 	/*
1463 	 * Force the new values to be used for future cross calls.
1464 	 * This is necessary only when we increase the timeouts.
1465 	 */
1466 	if ((xc_tick_limit > tick_limit_save) || (xc_sync_tick_limit >
1467 	    sync_tick_limit_save)) {
1468 		cpuset_t cpuset = cpu_ready_set;
1469 		xt_sync(cpuset);
1470 	}
1471 
1472 	if (nrnode > 0)
1473 		md_free_scan_dag(mdp, &platlist);
1474 	(void) md_fini_handle(mdp);
1475 }
1476 
1477 void
1478 mach_soft_state_init(void)
1479 {
1480 	int		i;
1481 	uint64_t	ra;
1482 
1483 	/*
1484 	 * Try to register soft_state api. If it fails, soft_state api has not
1485 	 * been implemented in the firmware, so do not bother to setup
1486 	 * soft_state in the kernel.
1487 	 */
1488 	if ((i = hsvc_register(&soft_state_hsvc, &soft_state_sup_minor)) != 0) {
1489 		return;
1490 	}
1491 	for (i = 0; i < SOLARIS_SOFT_STATE_MSG_CNT; i++) {
1492 		ASSERT(strlen((const char *)(void *)
1493 		    soft_state_message_strings + i) < SSM_SIZE);
1494 		if ((ra = va_to_pa(
1495 		    (void *)(soft_state_message_strings + i))) == -1ll) {
1496 			return;
1497 		}
1498 		soft_state_message_ra[i] = ra;
1499 	}
1500 	/*
1501 	 * Tell OBP that we are supporting Guest State
1502 	 */
1503 	prom_sun4v_soft_state_supported();
1504 	soft_state_initialized = 1;
1505 }
1506 
1507 void
1508 mach_set_soft_state(uint64_t state, uint64_t *string_ra)
1509 {
1510 	uint64_t	rc;
1511 
1512 	if (soft_state_initialized && *string_ra) {
1513 		rc = hv_soft_state_set(state, *string_ra);
1514 		if (rc != H_EOK) {
1515 			cmn_err(CE_WARN,
1516 			    "hv_soft_state_set returned %ld\n", rc);
1517 		}
1518 	}
1519 }
1520 
1521 void
1522 mach_get_soft_state(uint64_t *state, uint64_t *string_ra)
1523 {
1524 	uint64_t	rc;
1525 
1526 	if (soft_state_initialized && *string_ra) {
1527 		rc = hv_soft_state_get(*string_ra, state);
1528 		if (rc != H_EOK) {
1529 			cmn_err(CE_WARN,
1530 			    "hv_soft_state_get returned %ld\n", rc);
1531 			*state = -1;
1532 		}
1533 	}
1534 }
1535