xref: /titanic_44/usr/src/uts/sun4u/cpu/opl_olympus.c (revision 18c2aff776a775d34a4c9893a4c72e0434d68e36)
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 2006 Sun Microsystems, Inc.  All rights reserved.
23  * Use is subject to license terms.
24  */
25 
26 #pragma ident	"%Z%%M%	%I%	%E% SMI"
27 
28 #include <sys/types.h>
29 #include <sys/systm.h>
30 #include <sys/ddi.h>
31 #include <sys/sysmacros.h>
32 #include <sys/archsystm.h>
33 #include <sys/vmsystm.h>
34 #include <sys/machparam.h>
35 #include <sys/machsystm.h>
36 #include <sys/machthread.h>
37 #include <sys/cpu.h>
38 #include <sys/cmp.h>
39 #include <sys/elf_SPARC.h>
40 #include <vm/vm_dep.h>
41 #include <vm/hat_sfmmu.h>
42 #include <vm/seg_kpm.h>
43 #include <sys/cpuvar.h>
44 #include <sys/opl_olympus_regs.h>
45 #include <sys/opl_module.h>
46 #include <sys/async.h>
47 #include <sys/cmn_err.h>
48 #include <sys/debug.h>
49 #include <sys/dditypes.h>
50 #include <sys/cpu_module.h>
51 #include <sys/sysmacros.h>
52 #include <sys/intreg.h>
53 #include <sys/clock.h>
54 #include <sys/platform_module.h>
55 #include <sys/ontrap.h>
56 #include <sys/panic.h>
57 #include <sys/memlist.h>
58 #include <sys/ndifm.h>
59 #include <sys/ddifm.h>
60 #include <sys/fm/protocol.h>
61 #include <sys/fm/util.h>
62 #include <sys/fm/cpu/SPARC64-VI.h>
63 #include <sys/dtrace.h>
64 #include <sys/watchpoint.h>
65 #include <sys/promif.h>
66 
67 /*
68  * Internal functions.
69  */
70 static int cpu_sync_log_err(void *flt);
71 static void cpu_payload_add_aflt(struct async_flt *, nvlist_t *, nvlist_t *);
72 static void opl_cpu_sync_error(struct regs *, ulong_t, ulong_t, uint_t, uint_t);
73 static int  cpu_flt_in_memory(opl_async_flt_t *, uint64_t);
74 
75 /*
76  * Error counters resetting interval.
77  */
78 static int opl_async_check_interval = 60;		/* 1 min */
79 
80 uint_t cpu_impl_dual_pgsz = 1;
81 
82 /*
83  * PA[22:0] represent Displacement in Jupiter
84  * configuration space.
85  */
86 uint_t	root_phys_addr_lo_mask = 0x7fffffu;
87 
88 /*
89  * set in /etc/system to control logging of user BERR/TO's
90  */
91 int cpu_berr_to_verbose = 0;
92 
93 static int min_ecache_size;
94 static uint_t priv_hcl_1;
95 static uint_t priv_hcl_2;
96 static uint_t priv_hcl_4;
97 static uint_t priv_hcl_8;
98 
99 /*
100  * Olympus error log
101  */
102 static opl_errlog_t	*opl_err_log;
103 
104 /*
105  * UE is classified into four classes (MEM, CHANNEL, CPU, PATH).
106  * No any other ecc_type_info insertion is allowed in between the following
107  * four UE classess.
108  */
109 ecc_type_to_info_t ecc_type_to_info[] = {
110 	SFSR_UE,	"UE ",	(OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_UE,
111 	"Uncorrectable ECC",  FM_EREPORT_PAYLOAD_SYNC,
112 	FM_EREPORT_CPU_UE_MEM,
113 	SFSR_UE,	"UE ",	(OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_UE,
114 	"Uncorrectable ECC",  FM_EREPORT_PAYLOAD_SYNC,
115 	FM_EREPORT_CPU_UE_CHANNEL,
116 	SFSR_UE,	"UE ",	(OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_UE,
117 	"Uncorrectable ECC",  FM_EREPORT_PAYLOAD_SYNC,
118 	FM_EREPORT_CPU_UE_CPU,
119 	SFSR_UE,	"UE ",	(OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_UE,
120 	"Uncorrectable ECC",  FM_EREPORT_PAYLOAD_SYNC,
121 	FM_EREPORT_CPU_UE_PATH,
122 	SFSR_BERR, "BERR ", (OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_OTHERS,
123 	"Bus Error",  FM_EREPORT_PAYLOAD_SYNC,
124 	FM_EREPORT_CPU_BERR,
125 	SFSR_TO, "TO ", (OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_OTHERS,
126 	"Bus Timeout",  FM_EREPORT_PAYLOAD_SYNC,
127 	FM_EREPORT_CPU_BTO,
128 	SFSR_TLB_MUL, "TLB_MUL ", (OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_OTHERS,
129 	"TLB MultiHit",  FM_EREPORT_PAYLOAD_SYNC,
130 	FM_EREPORT_CPU_MTLB,
131 	SFSR_TLB_PRT, "TLB_PRT ", (OPL_ECC_SYNC_TRAP), OPL_CPU_SYNC_OTHERS,
132 	"TLB Parity",  FM_EREPORT_PAYLOAD_SYNC,
133 	FM_EREPORT_CPU_TLBP,
134 
135 	UGESR_IAUG_CRE, "IAUG_CRE", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
136 	"IAUG CRE",  FM_EREPORT_PAYLOAD_URGENT,
137 	FM_EREPORT_CPU_CRE,
138 	UGESR_IAUG_TSBCTXT, "IAUG_TSBCTXT",
139 	OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
140 	"IAUG TSBCTXT",  FM_EREPORT_PAYLOAD_URGENT,
141 	FM_EREPORT_CPU_TSBCTX,
142 	UGESR_IUG_TSBP, "IUG_TSBP", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
143 	"IUG TSBP",  FM_EREPORT_PAYLOAD_URGENT,
144 	FM_EREPORT_CPU_TSBP,
145 	UGESR_IUG_PSTATE, "IUG_PSTATE", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
146 	"IUG PSTATE",  FM_EREPORT_PAYLOAD_URGENT,
147 	FM_EREPORT_CPU_PSTATE,
148 	UGESR_IUG_TSTATE, "IUG_TSTATE", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
149 	"IUG TSTATE",  FM_EREPORT_PAYLOAD_URGENT,
150 	FM_EREPORT_CPU_TSTATE,
151 	UGESR_IUG_F, "IUG_F", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
152 	"IUG FREG",  FM_EREPORT_PAYLOAD_URGENT,
153 	FM_EREPORT_CPU_IUG_F,
154 	UGESR_IUG_R, "IUG_R", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
155 	"IUG RREG",  FM_EREPORT_PAYLOAD_URGENT,
156 	FM_EREPORT_CPU_IUG_R,
157 	UGESR_AUG_SDC, "AUG_SDC", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
158 	"AUG SDC",  FM_EREPORT_PAYLOAD_URGENT,
159 	FM_EREPORT_CPU_SDC,
160 	UGESR_IUG_WDT, "IUG_WDT", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
161 	"IUG WDT",  FM_EREPORT_PAYLOAD_URGENT,
162 	FM_EREPORT_CPU_WDT,
163 	UGESR_IUG_DTLB, "IUG_DTLB", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
164 	"IUG DTLB",  FM_EREPORT_PAYLOAD_URGENT,
165 	FM_EREPORT_CPU_DTLB,
166 	UGESR_IUG_ITLB, "IUG_ITLB", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
167 	"IUG ITLB",  FM_EREPORT_PAYLOAD_URGENT,
168 	FM_EREPORT_CPU_ITLB,
169 	UGESR_IUG_COREERR, "IUG_COREERR",
170 	OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
171 	"IUG COREERR",  FM_EREPORT_PAYLOAD_URGENT,
172 	FM_EREPORT_CPU_CORE,
173 	UGESR_MULTI_DAE, "MULTI_DAE", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
174 	"MULTI DAE",  FM_EREPORT_PAYLOAD_URGENT,
175 	FM_EREPORT_CPU_DAE,
176 	UGESR_MULTI_IAE, "MULTI_IAE", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
177 	"MULTI IAE",  FM_EREPORT_PAYLOAD_URGENT,
178 	FM_EREPORT_CPU_IAE,
179 	UGESR_MULTI_UGE, "MULTI_UGE", OPL_ECC_URGENT_TRAP, OPL_CPU_URGENT,
180 	"MULTI UGE",  FM_EREPORT_PAYLOAD_URGENT,
181 	FM_EREPORT_CPU_UGE,
182 	0,		NULL,		0,		0,
183 	NULL,  0,	   0,
184 };
185 
186 int (*p2get_mem_info)(int synd_code, uint64_t paddr,
187 		uint64_t *mem_sizep, uint64_t *seg_sizep, uint64_t *bank_sizep,
188 		int *segsp, int *banksp, int *mcidp);
189 
190 
191 /*
192  * Setup trap handlers for 0xA, 0x32, 0x40 trap types.
193  */
194 void
195 cpu_init_trap(void)
196 {
197 	OPL_SET_TRAP(tt0_iae, opl_serr_instr);
198 	OPL_SET_TRAP(tt1_iae, opl_serr_instr);
199 	OPL_SET_TRAP(tt0_dae, opl_serr_instr);
200 	OPL_SET_TRAP(tt1_dae, opl_serr_instr);
201 	OPL_SET_TRAP(tt0_asdat, opl_ugerr_instr);
202 	OPL_SET_TRAP(tt1_asdat, opl_ugerr_instr);
203 }
204 
205 static int
206 getintprop(pnode_t node, char *name, int deflt)
207 {
208 	int	value;
209 
210 	switch (prom_getproplen(node, name)) {
211 	case sizeof (int):
212 		(void) prom_getprop(node, name, (caddr_t)&value);
213 		break;
214 
215 	default:
216 		value = deflt;
217 		break;
218 	}
219 
220 	return (value);
221 }
222 
223 /*
224  * Set the magic constants of the implementation.
225  */
226 /*ARGSUSED*/
227 void
228 cpu_fiximp(pnode_t dnode)
229 {
230 	int i, a;
231 	extern int vac_size, vac_shift;
232 	extern uint_t vac_mask;
233 
234 	static struct {
235 		char	*name;
236 		int	*var;
237 		int	defval;
238 	} prop[] = {
239 		"l1-dcache-size", &dcache_size, OPL_DCACHE_SIZE,
240 		"l1-dcache-line-size", &dcache_linesize, OPL_DCACHE_LSIZE,
241 		"l1-icache-size", &icache_size, OPL_ICACHE_SIZE,
242 		"l1-icache-line-size", &icache_linesize, OPL_ICACHE_LSIZE,
243 		"l2-cache-size", &ecache_size, OPL_ECACHE_SIZE,
244 		"l2-cache-line-size", &ecache_alignsize, OPL_ECACHE_LSIZE,
245 		"l2-cache-associativity", &ecache_associativity, OPL_ECACHE_NWAY
246 	};
247 
248 	for (i = 0; i < sizeof (prop) / sizeof (prop[0]); i++)
249 		*prop[i].var = getintprop(dnode, prop[i].name, prop[i].defval);
250 
251 	ecache_setsize = ecache_size / ecache_associativity;
252 
253 	vac_size = OPL_VAC_SIZE;
254 	vac_mask = MMU_PAGEMASK & (vac_size - 1);
255 	i = 0; a = vac_size;
256 	while (a >>= 1)
257 		++i;
258 	vac_shift = i;
259 	shm_alignment = vac_size;
260 	vac = 1;
261 }
262 
263 #ifdef	OLYMPUS_C_REV_B_ERRATA_XCALL
264 /*
265  * Quick and dirty way to redefine locally in
266  * OPL the value of IDSR_BN_SETS to 31 instead
267  * of the standard 32 value. This is to workaround
268  * REV_B of Olympus_c processor's problem in handling
269  * more than 31 xcall broadcast.
270  */
271 #undef	IDSR_BN_SETS
272 #define	IDSR_BN_SETS    31
273 #endif	/* OLYMPUS_C_REV_B_ERRATA_XCALL */
274 
275 void
276 send_mondo_set(cpuset_t set)
277 {
278 	int lo, busy, nack, shipped = 0;
279 	uint16_t i, cpuids[IDSR_BN_SETS];
280 	uint64_t idsr, nackmask = 0, busymask, curnack, curbusy;
281 	uint64_t starttick, endtick, tick, lasttick;
282 #if (NCPU > IDSR_BN_SETS)
283 	int index = 0;
284 	int ncpuids = 0;
285 #endif
286 #ifdef	OLYMPUS_C_REV_A_ERRATA_XCALL
287 	int bn_sets = IDSR_BN_SETS;
288 	uint64_t ver;
289 
290 	ASSERT(NCPU > bn_sets);
291 #endif
292 
293 	ASSERT(!CPUSET_ISNULL(set));
294 	starttick = lasttick = gettick();
295 
296 #ifdef	OLYMPUS_C_REV_A_ERRATA_XCALL
297 	ver = ultra_getver();
298 	if (((ULTRA_VER_IMPL(ver)) == OLYMPUS_C_IMPL) &&
299 		((OLYMPUS_REV_MASK(ver)) == OLYMPUS_C_A))
300 		bn_sets = 1;
301 #endif
302 
303 #if (NCPU <= IDSR_BN_SETS)
304 	for (i = 0; i < NCPU; i++)
305 		if (CPU_IN_SET(set, i)) {
306 			shipit(i, shipped);
307 			nackmask |= IDSR_NACK_BIT(shipped);
308 			cpuids[shipped++] = i;
309 			CPUSET_DEL(set, i);
310 			if (CPUSET_ISNULL(set))
311 				break;
312 		}
313 	CPU_STATS_ADDQ(CPU, sys, xcalls, shipped);
314 #else
315 	for (i = 0; i < NCPU; i++)
316 		if (CPU_IN_SET(set, i)) {
317 			ncpuids++;
318 
319 			/*
320 			 * Ship only to the first (IDSR_BN_SETS) CPUs.  If we
321 			 * find we have shipped to more than (IDSR_BN_SETS)
322 			 * CPUs, set "index" to the highest numbered CPU in
323 			 * the set so we can ship to other CPUs a bit later on.
324 			 */
325 #ifdef	OLYMPUS_C_REV_A_ERRATA_XCALL
326 			if (shipped < bn_sets) {
327 #else
328 			if (shipped < IDSR_BN_SETS) {
329 #endif
330 				shipit(i, shipped);
331 				nackmask |= IDSR_NACK_BIT(shipped);
332 				cpuids[shipped++] = i;
333 				CPUSET_DEL(set, i);
334 				if (CPUSET_ISNULL(set))
335 					break;
336 			} else
337 				index = (int)i;
338 		}
339 
340 	CPU_STATS_ADDQ(CPU, sys, xcalls, ncpuids);
341 #endif
342 
343 	busymask = IDSR_NACK_TO_BUSY(nackmask);
344 	busy = nack = 0;
345 	endtick = starttick + xc_tick_limit;
346 	for (;;) {
347 		idsr = getidsr();
348 #if (NCPU <= IDSR_BN_SETS)
349 		if (idsr == 0)
350 			break;
351 #else
352 		if (idsr == 0 && shipped == ncpuids)
353 			break;
354 #endif
355 		tick = gettick();
356 		/*
357 		 * If there is a big jump between the current tick
358 		 * count and lasttick, we have probably hit a break
359 		 * point.  Adjust endtick accordingly to avoid panic.
360 		 */
361 		if (tick > (lasttick + xc_tick_jump_limit))
362 			endtick += (tick - lasttick);
363 		lasttick = tick;
364 		if (tick > endtick) {
365 			if (panic_quiesce)
366 				return;
367 			cmn_err(CE_CONT, "send mondo timeout "
368 				"[%d NACK %d BUSY]\nIDSR 0x%"
369 				"" PRIx64 "  cpuids:", nack, busy, idsr);
370 #ifdef	OLYMPUS_C_REV_A_ERRATA_XCALL
371 			for (i = 0; i < bn_sets; i++) {
372 #else
373 			for (i = 0; i < IDSR_BN_SETS; i++) {
374 #endif
375 				if (idsr & (IDSR_NACK_BIT(i) |
376 				    IDSR_BUSY_BIT(i))) {
377 					cmn_err(CE_CONT, " 0x%x",
378 						cpuids[i]);
379 				}
380 			}
381 			cmn_err(CE_CONT, "\n");
382 			cmn_err(CE_PANIC, "send_mondo_set: timeout");
383 		}
384 		curnack = idsr & nackmask;
385 		curbusy = idsr & busymask;
386 
387 #ifdef OLYMPUS_C_REV_B_ERRATA_XCALL
388 		/*
389 		 * Only proceed to send more xcalls if all the
390 		 * cpus in the previous IDSR_BN_SETS were completed.
391 		 */
392 		if (curbusy) {
393 			busy++;
394 			continue;
395 		}
396 #endif /* OLYMPUS_C_REV_B_ERRATA_XCALL */
397 
398 #if (NCPU > IDSR_BN_SETS)
399 		if (shipped < ncpuids) {
400 			uint64_t cpus_left;
401 			uint16_t next = (uint16_t)index;
402 
403 			cpus_left = ~(IDSR_NACK_TO_BUSY(curnack) | curbusy) &
404 			    busymask;
405 
406 			if (cpus_left) {
407 				do {
408 					/*
409 					 * Sequence through and ship to the
410 					 * remainder of the CPUs in the system
411 					 * (e.g. other than the first
412 					 * (IDSR_BN_SETS)) in reverse order.
413 					 */
414 					lo = lowbit(cpus_left) - 1;
415 					i = IDSR_BUSY_IDX(lo);
416 					shipit(next, i);
417 					shipped++;
418 					cpuids[i] = next;
419 
420 					/*
421 					 * If we've processed all the CPUs,
422 					 * exit the loop now and save
423 					 * instructions.
424 					 */
425 					if (shipped == ncpuids)
426 						break;
427 
428 					for ((index = ((int)next - 1));
429 						index >= 0; index--)
430 						if (CPU_IN_SET(set, index)) {
431 							next = (uint16_t)index;
432 							break;
433 						}
434 
435 					cpus_left &= ~(1ull << lo);
436 				} while (cpus_left);
437 				continue;
438 			}
439 		}
440 #endif
441 #ifndef	OLYMPUS_C_REV_B_ERRATA_XCALL
442 		if (curbusy) {
443 			busy++;
444 			continue;
445 		}
446 #endif	/* OLYMPUS_C_REV_B_ERRATA_XCALL */
447 #ifdef SEND_MONDO_STATS
448 		{
449 			int n = gettick() - starttick;
450 			if (n < 8192)
451 				x_nack_stimes[n >> 7]++;
452 		}
453 #endif
454 		while (gettick() < (tick + sys_clock_mhz))
455 			;
456 		do {
457 			lo = lowbit(curnack) - 1;
458 			i = IDSR_NACK_IDX(lo);
459 			shipit(cpuids[i], i);
460 			curnack &= ~(1ull << lo);
461 		} while (curnack);
462 		nack++;
463 		busy = 0;
464 	}
465 #ifdef SEND_MONDO_STATS
466 	{
467 		int n = gettick() - starttick;
468 		if (n < 8192)
469 			x_set_stimes[n >> 7]++;
470 		else
471 			x_set_ltimes[(n >> 13) & 0xf]++;
472 	}
473 	x_set_cpus[shipped]++;
474 #endif
475 }
476 
477 /*
478  * Cpu private initialization.
479  */
480 void
481 cpu_init_private(struct cpu *cp)
482 {
483 	if (!(IS_OLYMPUS_C(cpunodes[cp->cpu_id].implementation))) {
484 		cmn_err(CE_PANIC, "CPU%d Impl %d: Only SPARC64-VI is supported",
485 			cp->cpu_id, cpunodes[cp->cpu_id].implementation);
486 	}
487 
488 	adjust_hw_copy_limits(cpunodes[cp->cpu_id].ecache_size);
489 }
490 
491 void
492 cpu_setup(void)
493 {
494 	extern int at_flags;
495 	extern int disable_delay_tlb_flush, delay_tlb_flush;
496 	extern int cpc_has_overflow_intr;
497 	extern int disable_text_largepages;
498 	extern int use_text_pgsz4m;
499 	uint64_t cpu0_log;
500 	extern	 uint64_t opl_cpu0_err_log;
501 
502 	/*
503 	 * Initialize Error log Scratch register for error handling.
504 	 */
505 
506 	cpu0_log = va_to_pa(&opl_cpu0_err_log);
507 	opl_error_setup(cpu0_log);
508 
509 	/*
510 	 * Enable MMU translating multiple page sizes for
511 	 * sITLB and sDTLB.
512 	 */
513 	opl_mpg_enable();
514 
515 	/*
516 	 * Setup chip-specific trap handlers.
517 	 */
518 	cpu_init_trap();
519 
520 	cache |= (CACHE_VAC | CACHE_PTAG | CACHE_IOCOHERENT);
521 
522 	at_flags = EF_SPARC_32PLUS | EF_SPARC_SUN_US1 | EF_SPARC_SUN_US3;
523 
524 	/*
525 	 * Due to the number of entries in the fully-associative tlb
526 	 * this may have to be tuned lower than in spitfire.
527 	 */
528 	pp_slots = MIN(8, MAXPP_SLOTS);
529 
530 	/*
531 	 * Block stores do not invalidate all pages of the d$, pagecopy
532 	 * et. al. need virtual translations with virtual coloring taken
533 	 * into consideration.  prefetch/ldd will pollute the d$ on the
534 	 * load side.
535 	 */
536 	pp_consistent_coloring = PPAGE_STORE_VCOLORING | PPAGE_LOADS_POLLUTE;
537 
538 	if (use_page_coloring) {
539 		do_pg_coloring = 1;
540 		if (use_virtual_coloring)
541 			do_virtual_coloring = 1;
542 	}
543 
544 	isa_list =
545 	    "sparcv9+vis2 sparcv9+vis sparcv9 "
546 	    "sparcv8plus+vis2 sparcv8plus+vis sparcv8plus "
547 	    "sparcv8 sparcv8-fsmuld sparcv7 sparc";
548 
549 	cpu_hwcap_flags = AV_SPARC_VIS | AV_SPARC_VIS2 |
550 	    AV_SPARC_POPC | AV_SPARC_FMAF;
551 
552 	/*
553 	 * On SPARC64-VI, there's no hole in the virtual address space
554 	 */
555 	hole_start = hole_end = 0;
556 
557 	/*
558 	 * The kpm mapping window.
559 	 * kpm_size:
560 	 *	The size of a single kpm range.
561 	 *	The overall size will be: kpm_size * vac_colors.
562 	 * kpm_vbase:
563 	 *	The virtual start address of the kpm range within the kernel
564 	 *	virtual address space. kpm_vbase has to be kpm_size aligned.
565 	 */
566 	kpm_size = (size_t)(128ull * 1024 * 1024 * 1024 * 1024); /* 128TB */
567 	kpm_size_shift = 47;
568 	kpm_vbase = (caddr_t)0x8000000000000000ull; /* 8EB */
569 	kpm_smallpages = 1;
570 
571 	/*
572 	 * The traptrace code uses either %tick or %stick for
573 	 * timestamping.  We have %stick so we can use it.
574 	 */
575 	traptrace_use_stick = 1;
576 
577 	/*
578 	 * SPARC64-VI has a performance counter overflow interrupt
579 	 */
580 	cpc_has_overflow_intr = 1;
581 
582 	/*
583 	 * Use SPARC64-VI flush-all support
584 	 */
585 	if (!disable_delay_tlb_flush)
586 		delay_tlb_flush = 1;
587 
588 	/*
589 	 * Declare that this architecture/cpu combination does not support
590 	 * fpRAS.
591 	 */
592 	fpras_implemented = 0;
593 
594 	/*
595 	 * Enable 4M pages to be used for mapping user text by default.  Don't
596 	 * use large pages for initialized data segments since we may not know
597 	 * at exec() time what should be the preferred large page size for DTLB
598 	 * programming.
599 	 */
600 	use_text_pgsz4m = 1;
601 	disable_text_largepages = (1 << TTE64K) | (1 << TTE512K) |
602 	    (1 << TTE32M) | (1 << TTE256M);
603 }
604 
605 /*
606  * Called by setcpudelay
607  */
608 void
609 cpu_init_tick_freq(void)
610 {
611 	/*
612 	 * For SPARC64-VI we want to use the system clock rate as
613 	 * the basis for low level timing, due to support of mixed
614 	 * speed CPUs and power managment.
615 	 */
616 	if (system_clock_freq == 0)
617 		cmn_err(CE_PANIC, "setcpudelay: invalid system_clock_freq");
618 
619 	sys_tick_freq = system_clock_freq;
620 }
621 
622 #ifdef SEND_MONDO_STATS
623 uint32_t x_one_stimes[64];
624 uint32_t x_one_ltimes[16];
625 uint32_t x_set_stimes[64];
626 uint32_t x_set_ltimes[16];
627 uint32_t x_set_cpus[NCPU];
628 uint32_t x_nack_stimes[64];
629 #endif
630 
631 /*
632  * Note: A version of this function is used by the debugger via the KDI,
633  * and must be kept in sync with this version.  Any changes made to this
634  * function to support new chips or to accomodate errata must also be included
635  * in the KDI-specific version.  See us3_kdi.c.
636  */
637 void
638 send_one_mondo(int cpuid)
639 {
640 	int busy, nack;
641 	uint64_t idsr, starttick, endtick, tick, lasttick;
642 	uint64_t busymask;
643 
644 	CPU_STATS_ADDQ(CPU, sys, xcalls, 1);
645 	starttick = lasttick = gettick();
646 	shipit(cpuid, 0);
647 	endtick = starttick + xc_tick_limit;
648 	busy = nack = 0;
649 	busymask = IDSR_BUSY;
650 	for (;;) {
651 		idsr = getidsr();
652 		if (idsr == 0)
653 			break;
654 
655 		tick = gettick();
656 		/*
657 		 * If there is a big jump between the current tick
658 		 * count and lasttick, we have probably hit a break
659 		 * point.  Adjust endtick accordingly to avoid panic.
660 		 */
661 		if (tick > (lasttick + xc_tick_jump_limit))
662 			endtick += (tick - lasttick);
663 		lasttick = tick;
664 		if (tick > endtick) {
665 			if (panic_quiesce)
666 				return;
667 			cmn_err(CE_PANIC, "send mondo timeout "
668 				"(target 0x%x) [%d NACK %d BUSY]",
669 					cpuid, nack, busy);
670 		}
671 
672 		if (idsr & busymask) {
673 			busy++;
674 			continue;
675 		}
676 		drv_usecwait(1);
677 		shipit(cpuid, 0);
678 		nack++;
679 		busy = 0;
680 	}
681 #ifdef SEND_MONDO_STATS
682 	{
683 		int n = gettick() - starttick;
684 		if (n < 8192)
685 			x_one_stimes[n >> 7]++;
686 		else
687 			x_one_ltimes[(n >> 13) & 0xf]++;
688 	}
689 #endif
690 }
691 
692 /*
693  * init_mmu_page_sizes is set to one after the bootup time initialization
694  * via mmu_init_mmu_page_sizes, to indicate that mmu_page_sizes has a
695  * valid value.
696  *
697  * mmu_disable_ism_large_pages and mmu_disable_large_pages are the mmu-specific
698  * versions of disable_ism_large_pages and disable_large_pages, and feed back
699  * into those two hat variables at hat initialization time.
700  *
701  */
702 int init_mmu_page_sizes = 0;
703 static int mmu_disable_ism_large_pages = ((1 << TTE64K) |
704 	(1 << TTE512K) | (1 << TTE32M) | (1 << TTE256M));
705 static int mmu_disable_auto_large_pages = ((1 << TTE64K) |
706 	(1 << TTE512K) | (1 << TTE32M) | (1 << TTE256M));
707 static int mmu_disable_large_pages = 0;
708 
709 /*
710  * Re-initialize mmu_page_sizes and friends, for SPARC64-VI mmu support.
711  * Called during very early bootup from check_cpus_set().
712  * Can be called to verify that mmu_page_sizes are set up correctly.
713  *
714  * Set Olympus defaults. We do not use the function parameter.
715  */
716 /*ARGSUSED*/
717 int
718 mmu_init_mmu_page_sizes(int32_t not_used)
719 {
720 	if (!init_mmu_page_sizes) {
721 		mmu_page_sizes = MMU_PAGE_SIZES;
722 		mmu_hashcnt = MAX_HASHCNT;
723 		mmu_ism_pagesize = DEFAULT_ISM_PAGESIZE;
724 		auto_lpg_maxszc = TTE4M;
725 		mmu_exported_pagesize_mask = (1 << TTE8K) |
726 		    (1 << TTE64K) | (1 << TTE512K) | (1 << TTE4M) |
727 		    (1 << TTE32M) | (1 << TTE256M);
728 		init_mmu_page_sizes = 1;
729 		return (0);
730 	}
731 	return (1);
732 }
733 
734 /* SPARC64-VI worst case DTLB parameters */
735 #ifndef	LOCKED_DTLB_ENTRIES
736 #define	LOCKED_DTLB_ENTRIES	5	/* 2 user TSBs, 2 nucleus, + OBP */
737 #endif
738 #define	TOTAL_DTLB_ENTRIES	32
739 #define	AVAIL_32M_ENTRIES	0
740 #define	AVAIL_256M_ENTRIES	0
741 #define	AVAIL_DTLB_ENTRIES	(TOTAL_DTLB_ENTRIES - LOCKED_DTLB_ENTRIES)
742 static uint64_t ttecnt_threshold[MMU_PAGE_SIZES] = {
743 	AVAIL_DTLB_ENTRIES, AVAIL_DTLB_ENTRIES,
744 	AVAIL_DTLB_ENTRIES, AVAIL_DTLB_ENTRIES,
745 	AVAIL_DTLB_ENTRIES, AVAIL_DTLB_ENTRIES};
746 
747 /*
748  * The function returns the mmu-specific values for the
749  * hat's disable_large_pages, disable_ism_large_pages, and
750  * disable_auto_large_pages variables.
751  */
752 int
753 mmu_large_pages_disabled(uint_t flag)
754 {
755 	int pages_disable = 0;
756 
757 	if (flag == HAT_LOAD) {
758 		pages_disable =  mmu_disable_large_pages;
759 	} else if (flag == HAT_LOAD_SHARE) {
760 		pages_disable = mmu_disable_ism_large_pages;
761 	} else if (flag == HAT_LOAD_AUTOLPG) {
762 		pages_disable = mmu_disable_auto_large_pages;
763 	}
764 	return (pages_disable);
765 }
766 
767 /*
768  * mmu_init_large_pages is called with the desired ism_pagesize parameter.
769  * It may be called from set_platform_defaults, if some value other than 32M
770  * is desired.  mmu_ism_pagesize is the tunable.  If it has a bad value,
771  * then only warn, since it would be bad form to panic due to a user typo.
772  *
773  * The function re-initializes the mmu_disable_ism_large_pages variable.
774  */
775 void
776 mmu_init_large_pages(size_t ism_pagesize)
777 {
778 	switch (ism_pagesize) {
779 	case MMU_PAGESIZE4M:
780 		mmu_disable_ism_large_pages = ((1 << TTE64K) |
781 		    (1 << TTE512K) | (1 << TTE32M) | (1 << TTE256M));
782 		mmu_disable_auto_large_pages = ((1 << TTE64K) |
783 		    (1 << TTE512K) | (1 << TTE32M) | (1 << TTE256M));
784 		auto_lpg_maxszc = TTE4M;
785 		break;
786 	case MMU_PAGESIZE32M:
787 		mmu_disable_ism_large_pages = ((1 << TTE64K) |
788 		    (1 << TTE512K) | (1 << TTE256M));
789 		mmu_disable_auto_large_pages = ((1 << TTE64K) |
790 		    (1 << TTE512K) | (1 << TTE4M) | (1 << TTE256M));
791 		auto_lpg_maxszc = TTE32M;
792 		break;
793 	case MMU_PAGESIZE256M:
794 		mmu_disable_ism_large_pages = ((1 << TTE64K) |
795 		    (1 << TTE512K) | (1 << TTE32M));
796 		mmu_disable_auto_large_pages = ((1 << TTE64K) |
797 		    (1 << TTE512K) | (1 << TTE4M) | (1 << TTE32M));
798 		auto_lpg_maxszc = TTE256M;
799 		break;
800 	default:
801 		cmn_err(CE_WARN, "Unrecognized mmu_ism_pagesize value 0x%lx",
802 		    ism_pagesize);
803 		break;
804 	}
805 }
806 
807 /*
808  * Function to reprogram the TLBs when page sizes used
809  * by a process change significantly.
810  */
811 void
812 mmu_setup_page_sizes(struct hat *hat, uint64_t *ttecnt, uint8_t *tmp_pgsz)
813 {
814 	uint8_t pgsz0, pgsz1;
815 
816 	/*
817 	 * Don't program 2nd dtlb for kernel and ism hat
818 	 */
819 	ASSERT(hat->sfmmu_ismhat == NULL);
820 	ASSERT(hat != ksfmmup);
821 
822 	/*
823 	 * hat->sfmmu_pgsz[] is an array whose elements
824 	 * contain a sorted order of page sizes.  Element
825 	 * 0 is the most commonly used page size, followed
826 	 * by element 1, and so on.
827 	 *
828 	 * ttecnt[] is an array of per-page-size page counts
829 	 * mapped into the process.
830 	 *
831 	 * If the HAT's choice for page sizes is unsuitable,
832 	 * we can override it here.  The new values written
833 	 * to the array will be handed back to us later to
834 	 * do the actual programming of the TLB hardware.
835 	 *
836 	 */
837 	pgsz0 = (uint8_t)MIN(tmp_pgsz[0], tmp_pgsz[1]);
838 	pgsz1 = (uint8_t)MAX(tmp_pgsz[0], tmp_pgsz[1]);
839 
840 	/*
841 	 * This implements PAGESIZE programming of the sTLB
842 	 * if large TTE counts don't exceed the thresholds.
843 	 */
844 	if (ttecnt[pgsz0] < ttecnt_threshold[pgsz0])
845 		pgsz0 = page_szc(MMU_PAGESIZE);
846 	if (ttecnt[pgsz1] < ttecnt_threshold[pgsz1])
847 		pgsz1 = page_szc(MMU_PAGESIZE);
848 	tmp_pgsz[0] = pgsz0;
849 	tmp_pgsz[1] = pgsz1;
850 	/* otherwise, accept what the HAT chose for us */
851 }
852 
853 /*
854  * The HAT calls this function when an MMU context is allocated so that we
855  * can reprogram the large TLBs appropriately for the new process using
856  * the context.
857  *
858  * The caller must hold the HAT lock.
859  */
860 void
861 mmu_set_ctx_page_sizes(struct hat *hat)
862 {
863 	uint8_t pgsz0, pgsz1;
864 	uint8_t new_cext;
865 
866 	ASSERT(sfmmu_hat_lock_held(hat));
867 	/*
868 	 * Don't program 2nd dtlb for kernel and ism hat
869 	 */
870 	if (hat->sfmmu_ismhat || hat == ksfmmup)
871 		return;
872 
873 	/*
874 	 * If supported, reprogram the TLBs to a larger pagesize.
875 	 */
876 	pgsz0 = hat->sfmmu_pgsz[0];
877 	pgsz1 = hat->sfmmu_pgsz[1];
878 	ASSERT(pgsz0 < mmu_page_sizes);
879 	ASSERT(pgsz1 < mmu_page_sizes);
880 	new_cext = TAGACCEXT_MKSZPAIR(pgsz1, pgsz0);
881 	if (hat->sfmmu_cext != new_cext) {
882 #ifdef DEBUG
883 		int i;
884 		/*
885 		 * assert cnum should be invalid, this is because pagesize
886 		 * can only be changed after a proc's ctxs are invalidated.
887 		 */
888 		for (i = 0; i < max_mmu_ctxdoms; i++) {
889 			ASSERT(hat->sfmmu_ctxs[i].cnum == INVALID_CONTEXT);
890 		}
891 #endif /* DEBUG */
892 		hat->sfmmu_cext = new_cext;
893 	}
894 	/*
895 	 * sfmmu_setctx_sec() will take care of the
896 	 * rest of the dirty work for us.
897 	 */
898 }
899 
900 /*
901  * This function assumes that there are either four or six supported page
902  * sizes and at most two programmable TLBs, so we need to decide which
903  * page sizes are most important and then adjust the TLB page sizes
904  * accordingly (if supported).
905  *
906  * If these assumptions change, this function will need to be
907  * updated to support whatever the new limits are.
908  */
909 void
910 mmu_check_page_sizes(sfmmu_t *sfmmup, uint64_t *ttecnt)
911 {
912 	uint64_t sortcnt[MMU_PAGE_SIZES];
913 	uint8_t tmp_pgsz[MMU_PAGE_SIZES];
914 	uint8_t i, j, max;
915 	uint16_t oldval, newval;
916 
917 	/*
918 	 * We only consider reprogramming the TLBs if one or more of
919 	 * the two most used page sizes changes and we're using
920 	 * large pages in this process.
921 	 */
922 	if (sfmmup->sfmmu_flags & HAT_LGPG_FLAGS) {
923 		/* Sort page sizes. */
924 		for (i = 0; i < mmu_page_sizes; i++) {
925 			sortcnt[i] = ttecnt[i];
926 		}
927 		for (j = 0; j < mmu_page_sizes; j++) {
928 			for (i = mmu_page_sizes - 1, max = 0; i > 0; i--) {
929 				if (sortcnt[i] > sortcnt[max])
930 					max = i;
931 			}
932 			tmp_pgsz[j] = max;
933 			sortcnt[max] = 0;
934 		}
935 
936 		oldval = sfmmup->sfmmu_pgsz[0] << 8 | sfmmup->sfmmu_pgsz[1];
937 
938 		mmu_setup_page_sizes(sfmmup, ttecnt, tmp_pgsz);
939 
940 		/* Check 2 largest values after the sort. */
941 		newval = tmp_pgsz[0] << 8 | tmp_pgsz[1];
942 		if (newval != oldval) {
943 			sfmmu_reprog_pgsz_arr(sfmmup, tmp_pgsz);
944 		}
945 	}
946 }
947 
948 /*
949  * Return processor specific async error structure
950  * size used.
951  */
952 int
953 cpu_aflt_size(void)
954 {
955 	return (sizeof (opl_async_flt_t));
956 }
957 
958 /*
959  * The cpu_sync_log_err() function is called via the [uc]e_drain() function to
960  * post-process CPU events that are dequeued.  As such, it can be invoked
961  * from softint context, from AST processing in the trap() flow, or from the
962  * panic flow.  We decode the CPU-specific data, and take appropriate actions.
963  * Historically this entry point was used to log the actual cmn_err(9F) text;
964  * now with FMA it is used to prepare 'flt' to be converted into an ereport.
965  * With FMA this function now also returns a flag which indicates to the
966  * caller whether the ereport should be posted (1) or suppressed (0).
967  */
968 /*ARGSUSED*/
969 static int
970 cpu_sync_log_err(void *flt)
971 {
972 	opl_async_flt_t *opl_flt = (opl_async_flt_t *)flt;
973 	struct async_flt *aflt = (struct async_flt *)flt;
974 
975 	/*
976 	 * No extra processing of urgent error events.
977 	 * Always generate ereports for these events.
978 	 */
979 	if (aflt->flt_status == OPL_ECC_URGENT_TRAP)
980 		return (1);
981 
982 	/*
983 	 * Additional processing for synchronous errors.
984 	 */
985 	switch (opl_flt->flt_type) {
986 	case OPL_CPU_INV_SFSR:
987 		return (1);
988 
989 	case OPL_CPU_SYNC_UE:
990 		/*
991 		 * The validity: SFSR_MK_UE bit has been checked
992 		 * in opl_cpu_sync_error()
993 		 * No more check is required.
994 		 *
995 		 * opl_flt->flt_eid_mod and flt_eid_sid have been set by H/W,
996 		 * and they have been retrieved in cpu_queue_events()
997 		 */
998 
999 		if (opl_flt->flt_eid_mod == OPL_ERRID_MEM) {
1000 			ASSERT(aflt->flt_in_memory);
1001 			/*
1002 			 * We want to skip logging only if ALL the following
1003 			 * conditions are true:
1004 			 *
1005 			 *	1. We are not panicing already.
1006 			 *	2. The error is a memory error.
1007 			 *	3. There is only one error.
1008 			 *	4. The error is on a retired page.
1009 			 *	5. The error occurred under on_trap
1010 			 *	protection AFLT_PROT_EC
1011 			 */
1012 			if (!panicstr && aflt->flt_prot == AFLT_PROT_EC &&
1013 			    page_retire_check(aflt->flt_addr, NULL) == 0) {
1014 				/*
1015 				 * Do not log an error from
1016 				 * the retired page
1017 				 */
1018 				softcall(ecc_page_zero, (void *)aflt->flt_addr);
1019 				return (0);
1020 			}
1021 			if (!panicstr)
1022 				cpu_page_retire(opl_flt);
1023 		}
1024 		return (1);
1025 
1026 	case OPL_CPU_SYNC_OTHERS:
1027 		/*
1028 		 * For the following error cases, the processor HW does
1029 		 * not set the flt_eid_mod/flt_eid_sid. Instead, SW will attempt
1030 		 * to assign appropriate values here to reflect what we
1031 		 * think is the most likely cause of the problem w.r.t to
1032 		 * the particular error event.  For Buserr and timeout
1033 		 * error event, we will assign OPL_ERRID_CHANNEL as the
1034 		 * most likely reason.  For TLB parity or multiple hit
1035 		 * error events, we will assign the reason as
1036 		 * OPL_ERRID_CPU (cpu related problem) and set the
1037 		 * flt_eid_sid to point to the cpuid.
1038 		 */
1039 
1040 		if (opl_flt->flt_bit & (SFSR_BERR|SFSR_TO)) {
1041 			/*
1042 			 * flt_eid_sid will not be used for this case.
1043 			 */
1044 			opl_flt->flt_eid_mod = OPL_ERRID_CHANNEL;
1045 		}
1046 		if (opl_flt->flt_bit & (SFSR_TLB_MUL|SFSR_TLB_PRT)) {
1047 			    opl_flt->flt_eid_mod = OPL_ERRID_CPU;
1048 			    opl_flt->flt_eid_sid = aflt->flt_inst;
1049 		}
1050 
1051 		/*
1052 		 * In case of no effective error bit
1053 		 */
1054 		if ((opl_flt->flt_bit & SFSR_ERRS) == 0) {
1055 			    opl_flt->flt_eid_mod = OPL_ERRID_CPU;
1056 			    opl_flt->flt_eid_sid = aflt->flt_inst;
1057 		}
1058 		break;
1059 
1060 		default:
1061 			return (1);
1062 	}
1063 	return (1);
1064 }
1065 
1066 /*
1067  * Retire the bad page that may contain the flushed error.
1068  */
1069 void
1070 cpu_page_retire(opl_async_flt_t *opl_flt)
1071 {
1072 	struct async_flt *aflt = (struct async_flt *)opl_flt;
1073 	(void) page_retire(aflt->flt_addr, PR_UE);
1074 }
1075 
1076 /*
1077  * Invoked by error_init() early in startup and therefore before
1078  * startup_errorq() is called to drain any error Q -
1079  *
1080  * startup()
1081  *   startup_end()
1082  *     error_init()
1083  *       cpu_error_init()
1084  * errorq_init()
1085  *   errorq_drain()
1086  * start_other_cpus()
1087  *
1088  * The purpose of this routine is to create error-related taskqs.  Taskqs
1089  * are used for this purpose because cpu_lock can't be grabbed from interrupt
1090  * context.
1091  *
1092  */
1093 /*ARGSUSED*/
1094 void
1095 cpu_error_init(int items)
1096 {
1097 	opl_err_log = (opl_errlog_t *)
1098 	    kmem_alloc(ERRLOG_ALLOC_SZ, KM_SLEEP);
1099 	if ((uint64_t)opl_err_log & MMU_PAGEOFFSET)
1100 		cmn_err(CE_PANIC, "The base address of the error log "
1101 		    "is not page aligned");
1102 }
1103 
1104 /*
1105  * We route all errors through a single switch statement.
1106  */
1107 void
1108 cpu_ue_log_err(struct async_flt *aflt)
1109 {
1110 	switch (aflt->flt_class) {
1111 	case CPU_FAULT:
1112 		if (cpu_sync_log_err(aflt))
1113 			cpu_ereport_post(aflt);
1114 		break;
1115 
1116 	case BUS_FAULT:
1117 		bus_async_log_err(aflt);
1118 		break;
1119 
1120 	default:
1121 		cmn_err(CE_WARN, "discarding async error %p with invalid "
1122 		    "fault class (0x%x)", (void *)aflt, aflt->flt_class);
1123 		return;
1124 	}
1125 }
1126 
1127 /*
1128  * Routine for panic hook callback from panic_idle().
1129  *
1130  * Nothing to do here.
1131  */
1132 void
1133 cpu_async_panic_callb(void)
1134 {
1135 }
1136 
1137 /*
1138  * Routine to return a string identifying the physical name
1139  * associated with a memory/cache error.
1140  */
1141 /*ARGSUSED*/
1142 int
1143 cpu_get_mem_unum(int synd_status, ushort_t flt_synd, uint64_t flt_stat,
1144     uint64_t flt_addr, int flt_bus_id, int flt_in_memory,
1145     ushort_t flt_status, char *buf, int buflen, int *lenp)
1146 {
1147 	int synd_code;
1148 	int ret;
1149 
1150 	/*
1151 	 * An AFSR of -1 defaults to a memory syndrome.
1152 	 */
1153 	synd_code = (int)flt_synd;
1154 
1155 	if (&plat_get_mem_unum) {
1156 		if ((ret = plat_get_mem_unum(synd_code, flt_addr, flt_bus_id,
1157 			flt_in_memory, flt_status, buf, buflen, lenp)) != 0) {
1158 			buf[0] = '\0';
1159 			*lenp = 0;
1160 		}
1161 		return (ret);
1162 	}
1163 	buf[0] = '\0';
1164 	*lenp = 0;
1165 	return (ENOTSUP);
1166 }
1167 
1168 /*
1169  * Wrapper for cpu_get_mem_unum() routine that takes an
1170  * async_flt struct rather than explicit arguments.
1171  */
1172 int
1173 cpu_get_mem_unum_aflt(int synd_status, struct async_flt *aflt,
1174     char *buf, int buflen, int *lenp)
1175 {
1176 	/*
1177 	 * We always pass -1 so that cpu_get_mem_unum will interpret this as a
1178 	 * memory error.
1179 	 */
1180 	return (cpu_get_mem_unum(synd_status, aflt->flt_synd,
1181 	    (uint64_t)-1,
1182 	    aflt->flt_addr, aflt->flt_bus_id, aflt->flt_in_memory,
1183 	    aflt->flt_status, buf, buflen, lenp));
1184 }
1185 
1186 /*
1187  * This routine is a more generic interface to cpu_get_mem_unum()
1188  * that may be used by other modules (e.g. mm).
1189  */
1190 /*ARGSUSED*/
1191 int
1192 cpu_get_mem_name(uint64_t synd, uint64_t *afsr, uint64_t afar,
1193     char *buf, int buflen, int *lenp)
1194 {
1195 	int synd_status, flt_in_memory, ret;
1196 	ushort_t flt_status = 0;
1197 	char unum[UNUM_NAMLEN];
1198 
1199 	/*
1200 	 * Check for an invalid address.
1201 	 */
1202 	if (afar == (uint64_t)-1)
1203 		return (ENXIO);
1204 
1205 	if (synd == (uint64_t)-1)
1206 		synd_status = AFLT_STAT_INVALID;
1207 	else
1208 		synd_status = AFLT_STAT_VALID;
1209 
1210 	flt_in_memory = (*afsr & SFSR_MEMORY) &&
1211 		pf_is_memory(afar >> MMU_PAGESHIFT);
1212 
1213 	ret = cpu_get_mem_unum(synd_status, (ushort_t)synd, *afsr, afar,
1214 		CPU->cpu_id, flt_in_memory, flt_status, unum,
1215 		UNUM_NAMLEN, lenp);
1216 	if (ret != 0)
1217 		return (ret);
1218 
1219 	if (*lenp >= buflen)
1220 		return (ENAMETOOLONG);
1221 
1222 	(void) strncpy(buf, unum, buflen);
1223 
1224 	return (0);
1225 }
1226 
1227 /*
1228  * Routine to return memory information associated
1229  * with a physical address and syndrome.
1230  */
1231 /*ARGSUSED*/
1232 int
1233 cpu_get_mem_info(uint64_t synd, uint64_t afar,
1234     uint64_t *mem_sizep, uint64_t *seg_sizep, uint64_t *bank_sizep,
1235     int *segsp, int *banksp, int *mcidp)
1236 {
1237 	int synd_code = (int)synd;
1238 
1239 	if (afar == (uint64_t)-1)
1240 		return (ENXIO);
1241 
1242 	if (p2get_mem_info != NULL)
1243 		return ((p2get_mem_info)(synd_code, afar,
1244 			mem_sizep, seg_sizep, bank_sizep,
1245 			segsp, banksp, mcidp));
1246 	else
1247 		return (ENOTSUP);
1248 }
1249 
1250 /*
1251  * Routine to return a string identifying the physical
1252  * name associated with a cpuid.
1253  */
1254 int
1255 cpu_get_cpu_unum(int cpuid, char *buf, int buflen, int *lenp)
1256 {
1257 	int ret;
1258 	char unum[UNUM_NAMLEN];
1259 
1260 	if (&plat_get_cpu_unum) {
1261 		if ((ret = plat_get_cpu_unum(cpuid, unum, UNUM_NAMLEN, lenp))
1262 			!= 0)
1263 			return (ret);
1264 	} else {
1265 		return (ENOTSUP);
1266 	}
1267 
1268 	if (*lenp >= buflen)
1269 		return (ENAMETOOLONG);
1270 
1271 	(void) strncpy(buf, unum, *lenp);
1272 
1273 	return (0);
1274 }
1275 
1276 /*
1277  * This routine exports the name buffer size.
1278  */
1279 size_t
1280 cpu_get_name_bufsize()
1281 {
1282 	return (UNUM_NAMLEN);
1283 }
1284 
1285 /*
1286  * Flush the entire ecache by ASI_L2_CNTL.U2_FLUSH
1287  */
1288 void
1289 cpu_flush_ecache(void)
1290 {
1291 	flush_ecache(ecache_flushaddr, cpunodes[CPU->cpu_id].ecache_size,
1292 	    cpunodes[CPU->cpu_id].ecache_linesize);
1293 }
1294 
1295 static uint8_t
1296 flt_to_trap_type(struct async_flt *aflt)
1297 {
1298 	if (aflt->flt_status & OPL_ECC_ISYNC_TRAP)
1299 		return (TRAP_TYPE_ECC_I);
1300 	if (aflt->flt_status & OPL_ECC_DSYNC_TRAP)
1301 		return (TRAP_TYPE_ECC_D);
1302 	if (aflt->flt_status & OPL_ECC_URGENT_TRAP)
1303 		return (TRAP_TYPE_URGENT);
1304 	return (-1);
1305 }
1306 
1307 /*
1308  * Encode the data saved in the opl_async_flt_t struct into
1309  * the FM ereport payload.
1310  */
1311 /* ARGSUSED */
1312 static void
1313 cpu_payload_add_aflt(struct async_flt *aflt, nvlist_t *payload,
1314 		nvlist_t *resource)
1315 {
1316 	opl_async_flt_t *opl_flt = (opl_async_flt_t *)aflt;
1317 	char unum[UNUM_NAMLEN];
1318 	char sbuf[21]; /* sizeof (UINT64_MAX) + '\0' */
1319 	int len;
1320 
1321 
1322 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_SFSR) {
1323 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_SFSR,
1324 			DATA_TYPE_UINT64, aflt->flt_stat, NULL);
1325 	}
1326 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_SFAR) {
1327 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_SFAR,
1328 			DATA_TYPE_UINT64, aflt->flt_addr, NULL);
1329 	}
1330 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_UGESR) {
1331 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_UGESR,
1332 			DATA_TYPE_UINT64, aflt->flt_stat, NULL);
1333 	}
1334 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_PC) {
1335 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_PC,
1336 		    DATA_TYPE_UINT64, (uint64_t)aflt->flt_pc, NULL);
1337 	}
1338 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_TL) {
1339 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_TL,
1340 		    DATA_TYPE_UINT8, (uint8_t)aflt->flt_tl, NULL);
1341 	}
1342 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_TT) {
1343 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_TT,
1344 		    DATA_TYPE_UINT8, flt_to_trap_type(aflt), NULL);
1345 	}
1346 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_PRIV) {
1347 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_PRIV,
1348 		    DATA_TYPE_BOOLEAN_VALUE,
1349 		    (aflt->flt_priv ? B_TRUE : B_FALSE), NULL);
1350 	}
1351 	if (aflt->flt_payload & FM_EREPORT_PAYLOAD_FLAG_FLT_STATUS) {
1352 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_FLT_STATUS,
1353 			DATA_TYPE_UINT64, (uint64_t)aflt->flt_status, NULL);
1354 	}
1355 
1356 	switch (opl_flt->flt_eid_mod) {
1357 	case OPL_ERRID_CPU:
1358 		(void) snprintf(sbuf, sizeof (sbuf), "%llX",
1359 			(u_longlong_t)cpunodes[opl_flt->flt_eid_sid].device_id);
1360 		(void) fm_fmri_cpu_set(resource, FM_CPU_SCHEME_VERSION,
1361 			NULL, opl_flt->flt_eid_sid,
1362 			(uint8_t *)&cpunodes[opl_flt->flt_eid_sid].version,
1363 			sbuf);
1364 		fm_payload_set(payload,
1365 			FM_EREPORT_PAYLOAD_NAME_RESOURCE,
1366 			DATA_TYPE_NVLIST, resource, NULL);
1367 		break;
1368 
1369 	case OPL_ERRID_CHANNEL:
1370 		/*
1371 		 * No resource is created but the cpumem DE will find
1372 		 * the defective path by retreiving EID from SFSR which is
1373 		 * included in the payload.
1374 		 */
1375 		break;
1376 
1377 	case OPL_ERRID_MEM:
1378 		(void) cpu_get_mem_unum_aflt(0, aflt, unum, UNUM_NAMLEN, &len);
1379 		(void) fm_fmri_mem_set(resource, FM_MEM_SCHEME_VERSION,
1380 			NULL, unum, NULL, (uint64_t)-1);
1381 		fm_payload_set(payload, FM_EREPORT_PAYLOAD_NAME_RESOURCE,
1382 			DATA_TYPE_NVLIST, resource, NULL);
1383 		break;
1384 
1385 	case OPL_ERRID_PATH:
1386 		/*
1387 		 * No resource is created but the cpumem DE will find
1388 		 * the defective path by retreiving EID from SFSR which is
1389 		 * included in the payload.
1390 		 */
1391 		break;
1392 	}
1393 }
1394 
1395 /*
1396  * Returns whether fault address is valid for this error bit and
1397  * whether the address is "in memory" (i.e. pf_is_memory returns 1).
1398  */
1399 /*ARGSUSED*/
1400 static int
1401 cpu_flt_in_memory(opl_async_flt_t *opl_flt, uint64_t t_afsr_bit)
1402 {
1403 	struct async_flt *aflt = (struct async_flt *)opl_flt;
1404 
1405 	if (aflt->flt_status & (OPL_ECC_SYNC_TRAP)) {
1406 		return ((t_afsr_bit & SFSR_MEMORY) &&
1407 		    pf_is_memory(aflt->flt_addr >> MMU_PAGESHIFT));
1408 	}
1409 	return (0);
1410 }
1411 
1412 /*
1413  * In OPL SCF does the stick synchronization.
1414  */
1415 void
1416 sticksync_slave(void)
1417 {
1418 }
1419 
1420 /*
1421  * In OPL SCF does the stick synchronization.
1422  */
1423 void
1424 sticksync_master(void)
1425 {
1426 }
1427 
1428 /*
1429  * Cpu private unitialization.  OPL cpus do not use the private area.
1430  */
1431 void
1432 cpu_uninit_private(struct cpu *cp)
1433 {
1434 	cmp_delete_cpu(cp->cpu_id);
1435 }
1436 
1437 /*
1438  * Always flush an entire cache.
1439  */
1440 void
1441 cpu_error_ecache_flush(void)
1442 {
1443 	cpu_flush_ecache();
1444 }
1445 
1446 void
1447 cpu_ereport_post(struct async_flt *aflt)
1448 {
1449 	char *cpu_type, buf[FM_MAX_CLASS];
1450 	nv_alloc_t *nva = NULL;
1451 	nvlist_t *ereport, *detector, *resource;
1452 	errorq_elem_t *eqep;
1453 	char sbuf[21]; /* sizeof (UINT64_MAX) + '\0' */
1454 
1455 	if (aflt->flt_panic || panicstr) {
1456 		eqep = errorq_reserve(ereport_errorq);
1457 		if (eqep == NULL)
1458 			return;
1459 		ereport = errorq_elem_nvl(ereport_errorq, eqep);
1460 		nva = errorq_elem_nva(ereport_errorq, eqep);
1461 	} else {
1462 		ereport = fm_nvlist_create(nva);
1463 	}
1464 
1465 	/*
1466 	 * Create the scheme "cpu" FMRI.
1467 	 */
1468 	detector = fm_nvlist_create(nva);
1469 	resource = fm_nvlist_create(nva);
1470 	switch (cpunodes[aflt->flt_inst].implementation) {
1471 	case OLYMPUS_C_IMPL:
1472 		cpu_type = FM_EREPORT_CPU_SPARC64_VI;
1473 		break;
1474 	default:
1475 		cpu_type = FM_EREPORT_CPU_UNSUPPORTED;
1476 		break;
1477 	}
1478 	(void) snprintf(sbuf, sizeof (sbuf), "%llX",
1479 	    (u_longlong_t)cpunodes[aflt->flt_inst].device_id);
1480 	(void) fm_fmri_cpu_set(detector, FM_CPU_SCHEME_VERSION, NULL,
1481 	    aflt->flt_inst, (uint8_t *)&cpunodes[aflt->flt_inst].version,
1482 	    sbuf);
1483 
1484 	/*
1485 	 * Encode all the common data into the ereport.
1486 	 */
1487 	(void) snprintf(buf, FM_MAX_CLASS, "%s.%s.%s",
1488 	    FM_ERROR_CPU, cpu_type, aflt->flt_erpt_class);
1489 
1490 	fm_ereport_set(ereport, FM_EREPORT_VERSION, buf,
1491 	    fm_ena_generate(aflt->flt_id, FM_ENA_FMT1), detector, NULL);
1492 
1493 	/*
1494 	 * Encode the error specific data that was saved in
1495 	 * the async_flt structure into the ereport.
1496 	 */
1497 	cpu_payload_add_aflt(aflt, ereport, resource);
1498 
1499 	if (aflt->flt_panic || panicstr) {
1500 		errorq_commit(ereport_errorq, eqep, ERRORQ_SYNC);
1501 	} else {
1502 		(void) fm_ereport_post(ereport, EVCH_TRYHARD);
1503 		fm_nvlist_destroy(ereport, FM_NVA_FREE);
1504 		fm_nvlist_destroy(detector, FM_NVA_FREE);
1505 		fm_nvlist_destroy(resource, FM_NVA_FREE);
1506 	}
1507 }
1508 
1509 void
1510 cpu_run_bus_error_handlers(struct async_flt *aflt, int expected)
1511 {
1512 	int status;
1513 	ddi_fm_error_t de;
1514 
1515 	bzero(&de, sizeof (ddi_fm_error_t));
1516 
1517 	de.fme_version = DDI_FME_VERSION;
1518 	de.fme_ena = fm_ena_generate(aflt->flt_id, FM_ENA_FMT1);
1519 	de.fme_flag = expected;
1520 	de.fme_bus_specific = (void *)aflt->flt_addr;
1521 	status = ndi_fm_handler_dispatch(ddi_root_node(), NULL, &de);
1522 	if ((aflt->flt_prot == AFLT_PROT_NONE) && (status == DDI_FM_FATAL))
1523 		aflt->flt_panic = 1;
1524 }
1525 
1526 void
1527 cpu_errorq_dispatch(char *error_class, void *payload, size_t payload_sz,
1528     errorq_t *eqp, uint_t flag)
1529 {
1530 	struct async_flt *aflt = (struct async_flt *)payload;
1531 
1532 	aflt->flt_erpt_class = error_class;
1533 	errorq_dispatch(eqp, payload, payload_sz, flag);
1534 }
1535 
1536 void
1537 adjust_hw_copy_limits(int ecache_size)
1538 {
1539 	/*
1540 	 * Set hw copy limits.
1541 	 *
1542 	 * /etc/system will be parsed later and can override one or more
1543 	 * of these settings.
1544 	 *
1545 	 * At this time, ecache size seems only mildly relevant.
1546 	 * We seem to run into issues with the d-cache and stalls
1547 	 * we see on misses.
1548 	 *
1549 	 * Cycle measurement indicates that 2 byte aligned copies fare
1550 	 * little better than doing things with VIS at around 512 bytes.
1551 	 * 4 byte aligned shows promise until around 1024 bytes. 8 Byte
1552 	 * aligned is faster whenever the source and destination data
1553 	 * in cache and the total size is less than 2 Kbytes.  The 2K
1554 	 * limit seems to be driven by the 2K write cache.
1555 	 * When more than 2K of copies are done in non-VIS mode, stores
1556 	 * backup in the write cache.  In VIS mode, the write cache is
1557 	 * bypassed, allowing faster cache-line writes aligned on cache
1558 	 * boundaries.
1559 	 *
1560 	 * In addition, in non-VIS mode, there is no prefetching, so
1561 	 * for larger copies, the advantage of prefetching to avoid even
1562 	 * occasional cache misses is enough to justify using the VIS code.
1563 	 *
1564 	 * During testing, it was discovered that netbench ran 3% slower
1565 	 * when hw_copy_limit_8 was 2K or larger.  Apparently for server
1566 	 * applications, data is only used once (copied to the output
1567 	 * buffer, then copied by the network device off the system).  Using
1568 	 * the VIS copy saves more L2 cache state.  Network copies are
1569 	 * around 1.3K to 1.5K in size for historical reasons.
1570 	 *
1571 	 * Therefore, a limit of 1K bytes will be used for the 8 byte
1572 	 * aligned copy even for large caches and 8 MB ecache.  The
1573 	 * infrastructure to allow different limits for different sized
1574 	 * caches is kept to allow further tuning in later releases.
1575 	 */
1576 
1577 	if (min_ecache_size == 0 && use_hw_bcopy) {
1578 		/*
1579 		 * First time through - should be before /etc/system
1580 		 * is read.
1581 		 * Could skip the checks for zero but this lets us
1582 		 * preserve any debugger rewrites.
1583 		 */
1584 		if (hw_copy_limit_1 == 0) {
1585 			hw_copy_limit_1 = VIS_COPY_THRESHOLD;
1586 			priv_hcl_1 = hw_copy_limit_1;
1587 		}
1588 		if (hw_copy_limit_2 == 0) {
1589 			hw_copy_limit_2 = 2 * VIS_COPY_THRESHOLD;
1590 			priv_hcl_2 = hw_copy_limit_2;
1591 		}
1592 		if (hw_copy_limit_4 == 0) {
1593 			hw_copy_limit_4 = 4 * VIS_COPY_THRESHOLD;
1594 			priv_hcl_4 = hw_copy_limit_4;
1595 		}
1596 		if (hw_copy_limit_8 == 0) {
1597 			hw_copy_limit_8 = 4 * VIS_COPY_THRESHOLD;
1598 			priv_hcl_8 = hw_copy_limit_8;
1599 		}
1600 		min_ecache_size = ecache_size;
1601 	} else {
1602 		/*
1603 		 * MP initialization. Called *after* /etc/system has
1604 		 * been parsed. One CPU has already been initialized.
1605 		 * Need to cater for /etc/system having scragged one
1606 		 * of our values.
1607 		 */
1608 		if (ecache_size == min_ecache_size) {
1609 			/*
1610 			 * Same size ecache. We do nothing unless we
1611 			 * have a pessimistic ecache setting. In that
1612 			 * case we become more optimistic (if the cache is
1613 			 * large enough).
1614 			 */
1615 			if (hw_copy_limit_8 == 4 * VIS_COPY_THRESHOLD) {
1616 				/*
1617 				 * Need to adjust hw_copy_limit* from our
1618 				 * pessimistic uniprocessor value to a more
1619 				 * optimistic UP value *iff* it hasn't been
1620 				 * reset.
1621 				 */
1622 				if ((ecache_size > 1048576) &&
1623 				    (priv_hcl_8 == hw_copy_limit_8)) {
1624 					if (ecache_size <= 2097152)
1625 						hw_copy_limit_8 = 4 *
1626 						    VIS_COPY_THRESHOLD;
1627 					else if (ecache_size <= 4194304)
1628 						hw_copy_limit_8 = 4 *
1629 						    VIS_COPY_THRESHOLD;
1630 					else
1631 						hw_copy_limit_8 = 4 *
1632 						    VIS_COPY_THRESHOLD;
1633 					priv_hcl_8 = hw_copy_limit_8;
1634 				}
1635 			}
1636 		} else if (ecache_size < min_ecache_size) {
1637 			/*
1638 			 * A different ecache size. Can this even happen?
1639 			 */
1640 			if (priv_hcl_8 == hw_copy_limit_8) {
1641 				/*
1642 				 * The previous value that we set
1643 				 * is unchanged (i.e., it hasn't been
1644 				 * scragged by /etc/system). Rewrite it.
1645 				 */
1646 				if (ecache_size <= 1048576)
1647 					hw_copy_limit_8 = 8 *
1648 					    VIS_COPY_THRESHOLD;
1649 				else if (ecache_size <= 2097152)
1650 					hw_copy_limit_8 = 8 *
1651 					    VIS_COPY_THRESHOLD;
1652 				else if (ecache_size <= 4194304)
1653 					hw_copy_limit_8 = 8 *
1654 					    VIS_COPY_THRESHOLD;
1655 				else
1656 					hw_copy_limit_8 = 10 *
1657 					    VIS_COPY_THRESHOLD;
1658 				priv_hcl_8 = hw_copy_limit_8;
1659 				min_ecache_size = ecache_size;
1660 			}
1661 		}
1662 	}
1663 }
1664 
1665 #define	VIS_BLOCKSIZE		64
1666 
1667 int
1668 dtrace_blksuword32_err(uintptr_t addr, uint32_t *data)
1669 {
1670 	int ret, watched;
1671 
1672 	watched = watch_disable_addr((void *)addr, VIS_BLOCKSIZE, S_WRITE);
1673 	ret = dtrace_blksuword32(addr, data, 0);
1674 	if (watched)
1675 		watch_enable_addr((void *)addr, VIS_BLOCKSIZE, S_WRITE);
1676 
1677 	return (ret);
1678 }
1679 
1680 void
1681 opl_cpu_reg_init()
1682 {
1683 	uint64_t	this_cpu_log;
1684 
1685 	/*
1686 	 * We do not need to re-initialize cpu0 registers.
1687 	 */
1688 	if (cpu[getprocessorid()] == &cpu0)
1689 		return;
1690 
1691 	/*
1692 	 * Initialize Error log Scratch register for error handling.
1693 	 */
1694 
1695 	this_cpu_log = va_to_pa((void*)(((uint64_t)opl_err_log) +
1696 		ERRLOG_BUFSZ * (getprocessorid())));
1697 	opl_error_setup(this_cpu_log);
1698 
1699 	/*
1700 	 * Enable MMU translating multiple page sizes for
1701 	 * sITLB and sDTLB.
1702 	 */
1703 	opl_mpg_enable();
1704 }
1705 
1706 /*
1707  * Queue one event in ue_queue based on ecc_type_to_info entry.
1708  */
1709 static void
1710 cpu_queue_one_event(opl_async_flt_t *opl_flt, char *reason,
1711     ecc_type_to_info_t *eccp)
1712 {
1713 	struct async_flt *aflt = (struct async_flt *)opl_flt;
1714 
1715 	if (reason &&
1716 	    strlen(reason) + strlen(eccp->ec_reason) < MAX_REASON_STRING) {
1717 		(void) strcat(reason, eccp->ec_reason);
1718 	}
1719 
1720 	opl_flt->flt_bit = eccp->ec_afsr_bit;
1721 	opl_flt->flt_type = eccp->ec_flt_type;
1722 	aflt->flt_in_memory = cpu_flt_in_memory(opl_flt, opl_flt->flt_bit);
1723 	aflt->flt_payload = eccp->ec_err_payload;
1724 
1725 	ASSERT(aflt->flt_status & (OPL_ECC_SYNC_TRAP|OPL_ECC_URGENT_TRAP));
1726 	cpu_errorq_dispatch(eccp->ec_err_class,
1727 		(void *)opl_flt, sizeof (opl_async_flt_t),
1728 		ue_queue,
1729 		aflt->flt_panic);
1730 }
1731 
1732 /*
1733  * Queue events on async event queue one event per error bit.
1734  * Return number of events queued.
1735  */
1736 int
1737 cpu_queue_events(opl_async_flt_t *opl_flt, char *reason, uint64_t t_afsr_errs)
1738 {
1739 	struct async_flt *aflt = (struct async_flt *)opl_flt;
1740 	ecc_type_to_info_t *eccp;
1741 	int nevents = 0;
1742 
1743 	/*
1744 	 * Queue expected errors, error bit and fault type must must match
1745 	 * in the ecc_type_to_info table.
1746 	 */
1747 	for (eccp = ecc_type_to_info; t_afsr_errs != 0 && eccp->ec_desc != NULL;
1748 		eccp++) {
1749 		if ((eccp->ec_afsr_bit & t_afsr_errs) != 0 &&
1750 		    (eccp->ec_flags & aflt->flt_status) != 0) {
1751 			/*
1752 			 * UE error event can be further
1753 			 * classified/breakdown into finer granularity
1754 			 * based on the flt_eid_mod value set by HW.  We do
1755 			 * special handling here so that we can report UE
1756 			 * error in finer granularity as ue_mem,
1757 			 * ue_channel, ue_cpu or ue_path.
1758 			 */
1759 			if (eccp->ec_flt_type == OPL_CPU_SYNC_UE) {
1760 				opl_flt->flt_eid_mod =
1761 					(aflt->flt_stat & SFSR_EID_MOD)
1762 					>> SFSR_EID_MOD_SHIFT;
1763 				opl_flt->flt_eid_sid =
1764 					(aflt->flt_stat & SFSR_EID_SID)
1765 					>> SFSR_EID_SID_SHIFT;
1766 				/*
1767 				 * Need to advance eccp pointer by flt_eid_mod
1768 				 * so that we get an appropriate ecc pointer
1769 				 *
1770 				 * EID			# of advances
1771 				 * ----------------------------------
1772 				 * OPL_ERRID_MEM	0
1773 				 * OPL_ERRID_CHANNEL	1
1774 				 * OPL_ERRID_CPU	2
1775 				 * OPL_ERRID_PATH	3
1776 				 */
1777 				eccp += opl_flt->flt_eid_mod;
1778 			}
1779 			cpu_queue_one_event(opl_flt, reason, eccp);
1780 			t_afsr_errs &= ~eccp->ec_afsr_bit;
1781 			nevents++;
1782 		}
1783 	}
1784 
1785 	return (nevents);
1786 }
1787 
1788 /*
1789  * Sync. error wrapper functions.
1790  * We use these functions in order to transfer here from the
1791  * nucleus trap handler information about trap type (data or
1792  * instruction) and trap level (0 or above 0). This way we
1793  * get rid of using SFSR's reserved bits.
1794  */
1795 
1796 #define	OPL_SYNC_TL0	0
1797 #define	OPL_SYNC_TL1	1
1798 #define	OPL_ISYNC_ERR	0
1799 #define	OPL_DSYNC_ERR	1
1800 
1801 void
1802 opl_cpu_isync_tl0_error(struct regs *rp, ulong_t p_sfar, ulong_t p_sfsr)
1803 {
1804 	uint64_t t_sfar = p_sfar;
1805 	uint64_t t_sfsr = p_sfsr;
1806 
1807 	opl_cpu_sync_error(rp, t_sfar, t_sfsr,
1808 	    OPL_SYNC_TL0, OPL_ISYNC_ERR);
1809 }
1810 
1811 void
1812 opl_cpu_isync_tl1_error(struct regs *rp, ulong_t p_sfar, ulong_t p_sfsr)
1813 {
1814 	uint64_t t_sfar = p_sfar;
1815 	uint64_t t_sfsr = p_sfsr;
1816 
1817 	opl_cpu_sync_error(rp, t_sfar, t_sfsr,
1818 	    OPL_SYNC_TL1, OPL_ISYNC_ERR);
1819 }
1820 
1821 void
1822 opl_cpu_dsync_tl0_error(struct regs *rp, ulong_t p_sfar, ulong_t p_sfsr)
1823 {
1824 	uint64_t t_sfar = p_sfar;
1825 	uint64_t t_sfsr = p_sfsr;
1826 
1827 	opl_cpu_sync_error(rp, t_sfar, t_sfsr,
1828 	    OPL_SYNC_TL0, OPL_DSYNC_ERR);
1829 }
1830 
1831 void
1832 opl_cpu_dsync_tl1_error(struct regs *rp, ulong_t p_sfar, ulong_t p_sfsr)
1833 {
1834 	uint64_t t_sfar = p_sfar;
1835 	uint64_t t_sfsr = p_sfsr;
1836 
1837 	opl_cpu_sync_error(rp, t_sfar, t_sfsr,
1838 	    OPL_SYNC_TL1, OPL_DSYNC_ERR);
1839 }
1840 
1841 /*
1842  * The fj sync err handler transfers control here for UE, BERR, TO, TLB_MUL
1843  * and TLB_PRT.
1844  * This function is designed based on cpu_deferred_error().
1845  */
1846 
1847 static void
1848 opl_cpu_sync_error(struct regs *rp, ulong_t t_sfar, ulong_t t_sfsr,
1849     uint_t tl, uint_t derr)
1850 {
1851 	opl_async_flt_t opl_flt;
1852 	struct async_flt *aflt;
1853 	int trampolined = 0;
1854 	char pr_reason[MAX_REASON_STRING];
1855 	uint64_t log_sfsr;
1856 	int expected = DDI_FM_ERR_UNEXPECTED;
1857 	ddi_acc_hdl_t *hp;
1858 
1859 	/*
1860 	 * We need to look at p_flag to determine if the thread detected an
1861 	 * error while dumping core.  We can't grab p_lock here, but it's ok
1862 	 * because we just need a consistent snapshot and we know that everyone
1863 	 * else will store a consistent set of bits while holding p_lock.  We
1864 	 * don't have to worry about a race because SDOCORE is set once prior
1865 	 * to doing i/o from the process's address space and is never cleared.
1866 	 */
1867 	uint_t pflag = ttoproc(curthread)->p_flag;
1868 
1869 	pr_reason[0] = '\0';
1870 
1871 	/*
1872 	 * handle the specific error
1873 	 */
1874 	bzero(&opl_flt, sizeof (opl_async_flt_t));
1875 	aflt = (struct async_flt *)&opl_flt;
1876 	aflt->flt_id = gethrtime_waitfree();
1877 	aflt->flt_bus_id = getprocessorid();
1878 	aflt->flt_inst = CPU->cpu_id;
1879 	aflt->flt_stat = t_sfsr;
1880 	aflt->flt_addr = t_sfar;
1881 	aflt->flt_pc = (caddr_t)rp->r_pc;
1882 	aflt->flt_prot = (uchar_t)AFLT_PROT_NONE;
1883 	aflt->flt_class = (uchar_t)CPU_FAULT;
1884 	aflt->flt_priv = (uchar_t)
1885 		(tl == 1 ? 1 : ((rp->r_tstate & TSTATE_PRIV) ?  1 : 0));
1886 	aflt->flt_tl = (uchar_t)tl;
1887 	aflt->flt_panic = (uchar_t)(tl != 0 || aft_testfatal != 0 ||
1888 	    (t_sfsr & (SFSR_TLB_MUL|SFSR_TLB_PRT)) != 0);
1889 	aflt->flt_core = (pflag & SDOCORE) ? 1 : 0;
1890 	aflt->flt_status = (derr) ? OPL_ECC_DSYNC_TRAP : OPL_ECC_ISYNC_TRAP;
1891 
1892 	/*
1893 	 * If SFSR.FV is not set, both SFSR and SFAR/SFPAR values are uncertain.
1894 	 * So, clear all error bits to avoid mis-handling and force the system
1895 	 * panicked.
1896 	 * We skip all the procedures below down to the panic message call.
1897 	 */
1898 	if (!(t_sfsr & SFSR_FV)) {
1899 		opl_flt.flt_type = OPL_CPU_INV_SFSR;
1900 		aflt->flt_panic = 1;
1901 		aflt->flt_payload = FM_EREPORT_PAYLOAD_SYNC;
1902 		cpu_errorq_dispatch(FM_EREPORT_CPU_INV_SFSR,
1903 			(void *)&opl_flt, sizeof (opl_async_flt_t), ue_queue,
1904 			aflt->flt_panic);
1905 		fm_panic("%sErrors(s)", "invalid SFSR");
1906 	}
1907 
1908 	/*
1909 	 * If either UE and MK bit is off, this is not valid UE error.
1910 	 * If it is not valid UE error, clear UE & MK_UE bits to prevent
1911 	 * mis-handling below.
1912 	 * aflt->flt_stat keeps the original bits as a reference.
1913 	 */
1914 	if ((t_sfsr & (SFSR_MK_UE|SFSR_UE)) !=
1915 	    (SFSR_MK_UE|SFSR_UE)) {
1916 		t_sfsr &= ~(SFSR_MK_UE|SFSR_UE);
1917 	}
1918 
1919 	/*
1920 	 * If the trap occurred in privileged mode at TL=0, we need to check to
1921 	 * see if we were executing in the kernel under on_trap() or t_lofault
1922 	 * protection.  If so, modify the saved registers so that we return
1923 	 * from the trap to the appropriate trampoline routine.
1924 	 */
1925 	if (!aflt->flt_panic && aflt->flt_priv && tl == 0) {
1926 		if (curthread->t_ontrap != NULL) {
1927 			on_trap_data_t *otp = curthread->t_ontrap;
1928 
1929 			if (otp->ot_prot & OT_DATA_EC) {
1930 				aflt->flt_prot = (uchar_t)AFLT_PROT_EC;
1931 				otp->ot_trap |= (ushort_t)OT_DATA_EC;
1932 				rp->r_pc = otp->ot_trampoline;
1933 				rp->r_npc = rp->r_pc + 4;
1934 				trampolined = 1;
1935 			}
1936 
1937 			if ((t_sfsr & (SFSR_TO | SFSR_BERR)) &&
1938 			    (otp->ot_prot & OT_DATA_ACCESS)) {
1939 				aflt->flt_prot = (uchar_t)AFLT_PROT_ACCESS;
1940 				otp->ot_trap |= (ushort_t)OT_DATA_ACCESS;
1941 				rp->r_pc = otp->ot_trampoline;
1942 				rp->r_npc = rp->r_pc + 4;
1943 				trampolined = 1;
1944 				/*
1945 				 * for peeks and caut_gets errors are expected
1946 				 */
1947 				hp = (ddi_acc_hdl_t *)otp->ot_handle;
1948 				if (!hp)
1949 					expected = DDI_FM_ERR_PEEK;
1950 				else if (hp->ah_acc.devacc_attr_access ==
1951 				    DDI_CAUTIOUS_ACC)
1952 					expected = DDI_FM_ERR_EXPECTED;
1953 			}
1954 
1955 		} else if (curthread->t_lofault) {
1956 			aflt->flt_prot = AFLT_PROT_COPY;
1957 			rp->r_g1 = EFAULT;
1958 			rp->r_pc = curthread->t_lofault;
1959 			rp->r_npc = rp->r_pc + 4;
1960 			trampolined = 1;
1961 		}
1962 	}
1963 
1964 	/*
1965 	 * If we're in user mode or we're doing a protected copy, we either
1966 	 * want the ASTON code below to send a signal to the user process
1967 	 * or we want to panic if aft_panic is set.
1968 	 *
1969 	 * If we're in privileged mode and we're not doing a copy, then we
1970 	 * need to check if we've trampolined.  If we haven't trampolined,
1971 	 * we should panic.
1972 	 */
1973 	if (!aflt->flt_priv || aflt->flt_prot == AFLT_PROT_COPY) {
1974 		if (t_sfsr & (SFSR_ERRS & ~(SFSR_BERR | SFSR_TO)))
1975 			aflt->flt_panic |= aft_panic;
1976 	} else if (!trampolined) {
1977 		aflt->flt_panic = 1;
1978 	}
1979 
1980 	/*
1981 	 * If we've trampolined due to a privileged TO or BERR, or if an
1982 	 * unprivileged TO or BERR occurred, we don't want to enqueue an
1983 	 * event for that TO or BERR.  Queue all other events (if any) besides
1984 	 * the TO/BERR.
1985 	 */
1986 	log_sfsr = t_sfsr;
1987 	if (trampolined) {
1988 		log_sfsr &= ~(SFSR_TO | SFSR_BERR);
1989 	} else if (!aflt->flt_priv) {
1990 		/*
1991 		 * User mode, suppress messages if
1992 		 * cpu_berr_to_verbose is not set.
1993 		 */
1994 		if (!cpu_berr_to_verbose)
1995 			log_sfsr &= ~(SFSR_TO | SFSR_BERR);
1996 	}
1997 
1998 	if (((log_sfsr & SFSR_ERRS) &&
1999 		(cpu_queue_events(&opl_flt, pr_reason, t_sfsr) == 0)) ||
2000 	    ((t_sfsr & SFSR_ERRS) == 0)) {
2001 		opl_flt.flt_type = OPL_CPU_INV_SFSR;
2002 		aflt->flt_payload = FM_EREPORT_PAYLOAD_SYNC;
2003 		cpu_errorq_dispatch(FM_EREPORT_CPU_INV_SFSR,
2004 			(void *)&opl_flt, sizeof (opl_async_flt_t), ue_queue,
2005 			aflt->flt_panic);
2006 	}
2007 
2008 	if (t_sfsr & (SFSR_UE|SFSR_TO|SFSR_BERR)) {
2009 		cpu_run_bus_error_handlers(aflt, expected);
2010 	}
2011 
2012 	/*
2013 	 * Panic here if aflt->flt_panic has been set.  Enqueued errors will
2014 	 * be logged as part of the panic flow.
2015 	 */
2016 	if (aflt->flt_panic) {
2017 		if (pr_reason[0] == 0)
2018 			strcpy(pr_reason, "invalid SFSR ");
2019 
2020 		fm_panic("%sErrors(s)", pr_reason);
2021 	}
2022 
2023 	/*
2024 	 * If we queued an error and we are going to return from the trap and
2025 	 * the error was in user mode or inside of a copy routine, set AST flag
2026 	 * so the queue will be drained before returning to user mode.  The
2027 	 * AST processing will also act on our failure policy.
2028 	 */
2029 	if (!aflt->flt_priv || aflt->flt_prot == AFLT_PROT_COPY) {
2030 		int pcb_flag = 0;
2031 
2032 		if (t_sfsr & (SFSR_ERRS &
2033 			~(SFSR_BERR | SFSR_TO)))
2034 			pcb_flag |= ASYNC_HWERR;
2035 
2036 		if (t_sfsr & SFSR_BERR)
2037 			pcb_flag |= ASYNC_BERR;
2038 
2039 		if (t_sfsr & SFSR_TO)
2040 			pcb_flag |= ASYNC_BTO;
2041 
2042 		ttolwp(curthread)->lwp_pcb.pcb_flags |= pcb_flag;
2043 		aston(curthread);
2044 	}
2045 }
2046 
2047 /*ARGSUSED*/
2048 void
2049 opl_cpu_urgent_error(struct regs *rp, ulong_t p_ugesr, ulong_t tl)
2050 {
2051 	opl_async_flt_t opl_flt;
2052 	struct async_flt *aflt;
2053 	char pr_reason[MAX_REASON_STRING];
2054 
2055 	/* normalize tl */
2056 	tl = (tl >= 2 ? 1 : 0);
2057 	pr_reason[0] = '\0';
2058 
2059 	bzero(&opl_flt, sizeof (opl_async_flt_t));
2060 	aflt = (struct async_flt *)&opl_flt;
2061 	aflt->flt_id = gethrtime_waitfree();
2062 	aflt->flt_bus_id = getprocessorid();
2063 	aflt->flt_inst = CPU->cpu_id;
2064 	aflt->flt_stat = p_ugesr;
2065 	aflt->flt_pc = (caddr_t)rp->r_pc;
2066 	aflt->flt_class = (uchar_t)CPU_FAULT;
2067 	aflt->flt_tl = tl;
2068 	aflt->flt_priv = (uchar_t)
2069 		(tl == 1 ? 1 : ((rp->r_tstate & TSTATE_PRIV) ?  1 : 0));
2070 	aflt->flt_status = OPL_ECC_URGENT_TRAP;
2071 	aflt->flt_panic = 1;
2072 	/*
2073 	 * HW does not set mod/sid in case of urgent error.
2074 	 * So we have to set it here.
2075 	 */
2076 	opl_flt.flt_eid_mod = OPL_ERRID_CPU;
2077 	opl_flt.flt_eid_sid = aflt->flt_inst;
2078 
2079 	if (cpu_queue_events(&opl_flt, pr_reason, p_ugesr) == 0) {
2080 		opl_flt.flt_type = OPL_CPU_INV_UGESR;
2081 		aflt->flt_payload = FM_EREPORT_PAYLOAD_URGENT;
2082 		cpu_errorq_dispatch(FM_EREPORT_CPU_INV_URG,
2083 			(void *)&opl_flt, sizeof (opl_async_flt_t),
2084 			ue_queue, aflt->flt_panic);
2085 	}
2086 
2087 	fm_panic("Urgent Error");
2088 }
2089 
2090 /*
2091  * Initialization error counters resetting.
2092  */
2093 /* ARGSUSED */
2094 static void
2095 opl_ras_online(void *arg, cpu_t *cp, cyc_handler_t *hdlr, cyc_time_t *when)
2096 {
2097 	hdlr->cyh_func = (cyc_func_t)ras_cntr_reset;
2098 	hdlr->cyh_level = CY_LOW_LEVEL;
2099 	hdlr->cyh_arg = (void *)(uintptr_t)cp->cpu_id;
2100 
2101 	when->cyt_when = cp->cpu_id * (((hrtime_t)NANOSEC * 10)/ NCPU);
2102 	when->cyt_interval = (hrtime_t)NANOSEC * opl_async_check_interval;
2103 }
2104 
2105 void
2106 cpu_mp_init(void)
2107 {
2108 	cyc_omni_handler_t hdlr;
2109 
2110 	hdlr.cyo_online = opl_ras_online;
2111 	hdlr.cyo_offline = NULL;
2112 	hdlr.cyo_arg = NULL;
2113 	mutex_enter(&cpu_lock);
2114 	(void) cyclic_add_omni(&hdlr);
2115 	mutex_exit(&cpu_lock);
2116 }
2117 
2118 /*ARGSUSED*/
2119 void
2120 mmu_init_kernel_pgsz(struct hat *hat)
2121 {
2122 }
2123 
2124 size_t
2125 mmu_get_kernel_lpsize(size_t lpsize)
2126 {
2127 	uint_t tte;
2128 
2129 	if (lpsize == 0) {
2130 		/* no setting for segkmem_lpsize in /etc/system: use default */
2131 		return (MMU_PAGESIZE4M);
2132 	}
2133 
2134 	for (tte = TTE8K; tte <= TTE4M; tte++) {
2135 		if (lpsize == TTEBYTES(tte))
2136 			return (lpsize);
2137 	}
2138 
2139 	return (TTEBYTES(TTE8K));
2140 }
2141 
2142 /*
2143  * The following are functions that are unused in
2144  * OPL cpu module. They are defined here to resolve
2145  * dependencies in the "unix" module.
2146  * Unused functions that should never be called in
2147  * OPL are coded with ASSERT(0).
2148  */
2149 
2150 void
2151 cpu_disable_errors(void)
2152 {}
2153 
2154 void
2155 cpu_enable_errors(void)
2156 { ASSERT(0); }
2157 
2158 /*ARGSUSED*/
2159 void
2160 cpu_ce_scrub_mem_err(struct async_flt *ecc, boolean_t t)
2161 { ASSERT(0); }
2162 
2163 /*ARGSUSED*/
2164 void
2165 cpu_faulted_enter(struct cpu *cp)
2166 {}
2167 
2168 /*ARGSUSED*/
2169 void
2170 cpu_faulted_exit(struct cpu *cp)
2171 {}
2172 
2173 /*ARGSUSED*/
2174 void
2175 cpu_check_allcpus(struct async_flt *aflt)
2176 {}
2177 
2178 /*ARGSUSED*/
2179 void
2180 cpu_ce_log_err(struct async_flt *aflt, errorq_elem_t *t)
2181 { ASSERT(0); }
2182 
2183 /*ARGSUSED*/
2184 void
2185 cpu_check_ce(int flag, uint64_t pa, caddr_t va, uint_t psz)
2186 { ASSERT(0); }
2187 
2188 /*ARGSUSED*/
2189 void
2190 cpu_ce_count_unum(struct async_flt *ecc, int len, char *unum)
2191 { ASSERT(0); }
2192 
2193 /*ARGSUSED*/
2194 void
2195 cpu_busy_ecache_scrub(struct cpu *cp)
2196 {}
2197 
2198 /*ARGSUSED*/
2199 void
2200 cpu_idle_ecache_scrub(struct cpu *cp)
2201 {}
2202 
2203 /* ARGSUSED */
2204 void
2205 cpu_change_speed(uint64_t divisor, uint64_t arg2)
2206 { ASSERT(0); }
2207 
2208 void
2209 cpu_init_cache_scrub(void)
2210 {}
2211 
2212 /* ARGSUSED */
2213 int
2214 cpu_get_mem_sid(char *unum, char *buf, int buflen, int *lenp)
2215 {
2216 	if (&plat_get_mem_sid) {
2217 		return (plat_get_mem_sid(unum, buf, buflen, lenp));
2218 	} else {
2219 		return (ENOTSUP);
2220 	}
2221 }
2222 
2223 /* ARGSUSED */
2224 int
2225 cpu_get_mem_addr(char *unum, char *sid, uint64_t offset, uint64_t *addrp)
2226 {
2227 	if (&plat_get_mem_addr) {
2228 		return (plat_get_mem_addr(unum, sid, offset, addrp));
2229 	} else {
2230 		return (ENOTSUP);
2231 	}
2232 }
2233 
2234 /* ARGSUSED */
2235 int
2236 cpu_get_mem_offset(uint64_t flt_addr, uint64_t *offp)
2237 {
2238 	if (&plat_get_mem_offset) {
2239 		return (plat_get_mem_offset(flt_addr, offp));
2240 	} else {
2241 		return (ENOTSUP);
2242 	}
2243 }
2244 
2245 /*ARGSUSED*/
2246 void
2247 itlb_rd_entry(uint_t entry, tte_t *tte, uint64_t *va_tag)
2248 { ASSERT(0); }
2249 
2250 /*ARGSUSED*/
2251 void
2252 dtlb_rd_entry(uint_t entry, tte_t *tte, uint64_t *va_tag)
2253 { ASSERT(0); }
2254 
2255 /*ARGSUSED*/
2256 void
2257 read_ecc_data(struct async_flt *aflt, short verbose, short ce_err)
2258 { ASSERT(0); }
2259 
2260 /*ARGSUSED*/
2261 int
2262 ce_scrub_xdiag_recirc(struct async_flt *aflt, errorq_t *eqp,
2263     errorq_elem_t *eqep, size_t afltoffset)
2264 {
2265 	ASSERT(0);
2266 	return (0);
2267 }
2268 
2269 /*ARGSUSED*/
2270 char *
2271 flt_to_error_type(struct async_flt *aflt)
2272 {
2273 	ASSERT(0);
2274 	return (NULL);
2275 }
2276